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 <linux/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 #define FLAG_TS_PROGRESS 0x40000 /* Positive timestamp delta */ 106 107 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED) 108 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED) 109 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE|FLAG_DSACKING_ACK) 110 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED) 111 112 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH) 113 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH)) 114 115 #define REXMIT_NONE 0 /* no loss recovery to do */ 116 #define REXMIT_LOST 1 /* retransmit packets marked lost */ 117 #define REXMIT_NEW 2 /* FRTO-style transmit of unsent/new packets */ 118 119 #if IS_ENABLED(CONFIG_TLS_DEVICE) 120 static DEFINE_STATIC_KEY_DEFERRED_FALSE(clean_acked_data_enabled, HZ); 121 122 void clean_acked_data_enable(struct inet_connection_sock *icsk, 123 void (*cad)(struct sock *sk, u32 ack_seq)) 124 { 125 icsk->icsk_clean_acked = cad; 126 static_branch_deferred_inc(&clean_acked_data_enabled); 127 } 128 EXPORT_SYMBOL_GPL(clean_acked_data_enable); 129 130 void clean_acked_data_disable(struct inet_connection_sock *icsk) 131 { 132 static_branch_slow_dec_deferred(&clean_acked_data_enabled); 133 icsk->icsk_clean_acked = NULL; 134 } 135 EXPORT_SYMBOL_GPL(clean_acked_data_disable); 136 137 void clean_acked_data_flush(void) 138 { 139 static_key_deferred_flush(&clean_acked_data_enabled); 140 } 141 EXPORT_SYMBOL_GPL(clean_acked_data_flush); 142 #endif 143 144 #ifdef CONFIG_CGROUP_BPF 145 static void bpf_skops_parse_hdr(struct sock *sk, struct sk_buff *skb) 146 { 147 bool unknown_opt = tcp_sk(sk)->rx_opt.saw_unknown && 148 BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), 149 BPF_SOCK_OPS_PARSE_UNKNOWN_HDR_OPT_CB_FLAG); 150 bool parse_all_opt = BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), 151 BPF_SOCK_OPS_PARSE_ALL_HDR_OPT_CB_FLAG); 152 struct bpf_sock_ops_kern sock_ops; 153 154 if (likely(!unknown_opt && !parse_all_opt)) 155 return; 156 157 /* The skb will be handled in the 158 * bpf_skops_established() or 159 * bpf_skops_write_hdr_opt(). 160 */ 161 switch (sk->sk_state) { 162 case TCP_SYN_RECV: 163 case TCP_SYN_SENT: 164 case TCP_LISTEN: 165 return; 166 } 167 168 sock_owned_by_me(sk); 169 170 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp)); 171 sock_ops.op = BPF_SOCK_OPS_PARSE_HDR_OPT_CB; 172 sock_ops.is_fullsock = 1; 173 sock_ops.is_locked_tcp_sock = 1; 174 sock_ops.sk = sk; 175 bpf_skops_init_skb(&sock_ops, skb, tcp_hdrlen(skb)); 176 177 BPF_CGROUP_RUN_PROG_SOCK_OPS(&sock_ops); 178 } 179 180 static void bpf_skops_established(struct sock *sk, int bpf_op, 181 struct sk_buff *skb) 182 { 183 struct bpf_sock_ops_kern sock_ops; 184 185 sock_owned_by_me(sk); 186 187 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp)); 188 sock_ops.op = bpf_op; 189 sock_ops.is_fullsock = 1; 190 sock_ops.is_locked_tcp_sock = 1; 191 sock_ops.sk = sk; 192 /* sk with TCP_REPAIR_ON does not have skb in tcp_finish_connect */ 193 if (skb) 194 bpf_skops_init_skb(&sock_ops, skb, tcp_hdrlen(skb)); 195 196 BPF_CGROUP_RUN_PROG_SOCK_OPS(&sock_ops); 197 } 198 #else 199 static void bpf_skops_parse_hdr(struct sock *sk, struct sk_buff *skb) 200 { 201 } 202 203 static void bpf_skops_established(struct sock *sk, int bpf_op, 204 struct sk_buff *skb) 205 { 206 } 207 #endif 208 209 static __cold void tcp_gro_dev_warn(const struct sock *sk, const struct sk_buff *skb, 210 unsigned int len) 211 { 212 struct net_device *dev; 213 214 rcu_read_lock(); 215 dev = dev_get_by_index_rcu(sock_net(sk), skb->skb_iif); 216 if (!dev || len >= READ_ONCE(dev->mtu)) 217 pr_warn("%s: Driver has suspect GRO implementation, TCP performance may be compromised.\n", 218 dev ? dev->name : "Unknown driver"); 219 rcu_read_unlock(); 220 } 221 222 /* Adapt the MSS value used to make delayed ack decision to the 223 * real world. 224 */ 225 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb) 226 { 227 struct inet_connection_sock *icsk = inet_csk(sk); 228 const unsigned int lss = icsk->icsk_ack.last_seg_size; 229 unsigned int len; 230 231 icsk->icsk_ack.last_seg_size = 0; 232 233 /* skb->len may jitter because of SACKs, even if peer 234 * sends good full-sized frames. 235 */ 236 len = skb_shinfo(skb)->gso_size ? : skb->len; 237 if (len >= icsk->icsk_ack.rcv_mss) { 238 /* Note: divides are still a bit expensive. 239 * For the moment, only adjust scaling_ratio 240 * when we update icsk_ack.rcv_mss. 241 */ 242 if (unlikely(len != icsk->icsk_ack.rcv_mss)) { 243 u64 val = (u64)skb->len << TCP_RMEM_TO_WIN_SCALE; 244 u8 old_ratio = tcp_sk(sk)->scaling_ratio; 245 246 do_div(val, skb->truesize); 247 tcp_sk(sk)->scaling_ratio = val ? val : 1; 248 249 if (old_ratio != tcp_sk(sk)->scaling_ratio) { 250 struct tcp_sock *tp = tcp_sk(sk); 251 252 val = tcp_win_from_space(sk, sk->sk_rcvbuf); 253 tcp_set_window_clamp(sk, val); 254 255 if (tp->window_clamp < tp->rcvq_space.space) 256 tp->rcvq_space.space = tp->window_clamp; 257 } 258 } 259 icsk->icsk_ack.rcv_mss = min_t(unsigned int, len, 260 tcp_sk(sk)->advmss); 261 /* Account for possibly-removed options */ 262 DO_ONCE_LITE_IF(len > icsk->icsk_ack.rcv_mss + MAX_TCP_OPTION_SPACE, 263 tcp_gro_dev_warn, sk, skb, len); 264 /* If the skb has a len of exactly 1*MSS and has the PSH bit 265 * set then it is likely the end of an application write. So 266 * more data may not be arriving soon, and yet the data sender 267 * may be waiting for an ACK if cwnd-bound or using TX zero 268 * copy. So we set ICSK_ACK_PUSHED here so that 269 * tcp_cleanup_rbuf() will send an ACK immediately if the app 270 * reads all of the data and is not ping-pong. If len > MSS 271 * then this logic does not matter (and does not hurt) because 272 * tcp_cleanup_rbuf() will always ACK immediately if the app 273 * reads data and there is more than an MSS of unACKed data. 274 */ 275 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_PSH) 276 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED; 277 } else { 278 /* Otherwise, we make more careful check taking into account, 279 * that SACKs block is variable. 280 * 281 * "len" is invariant segment length, including TCP header. 282 */ 283 len += skb->data - skb_transport_header(skb); 284 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) || 285 /* If PSH is not set, packet should be 286 * full sized, provided peer TCP is not badly broken. 287 * This observation (if it is correct 8)) allows 288 * to handle super-low mtu links fairly. 289 */ 290 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) && 291 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) { 292 /* Subtract also invariant (if peer is RFC compliant), 293 * tcp header plus fixed timestamp option length. 294 * Resulting "len" is MSS free of SACK jitter. 295 */ 296 len -= tcp_sk(sk)->tcp_header_len; 297 icsk->icsk_ack.last_seg_size = len; 298 if (len == lss) { 299 icsk->icsk_ack.rcv_mss = len; 300 return; 301 } 302 } 303 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED) 304 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2; 305 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED; 306 } 307 } 308 309 static void tcp_incr_quickack(struct sock *sk, unsigned int max_quickacks) 310 { 311 struct inet_connection_sock *icsk = inet_csk(sk); 312 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss); 313 314 if (quickacks == 0) 315 quickacks = 2; 316 quickacks = min(quickacks, max_quickacks); 317 if (quickacks > icsk->icsk_ack.quick) 318 icsk->icsk_ack.quick = quickacks; 319 } 320 321 static void tcp_enter_quickack_mode(struct sock *sk, unsigned int max_quickacks) 322 { 323 struct inet_connection_sock *icsk = inet_csk(sk); 324 325 tcp_incr_quickack(sk, max_quickacks); 326 inet_csk_exit_pingpong_mode(sk); 327 icsk->icsk_ack.ato = TCP_ATO_MIN; 328 } 329 330 /* Send ACKs quickly, if "quick" count is not exhausted 331 * and the session is not interactive. 332 */ 333 334 static bool tcp_in_quickack_mode(struct sock *sk) 335 { 336 const struct inet_connection_sock *icsk = inet_csk(sk); 337 const struct dst_entry *dst = __sk_dst_get(sk); 338 339 return (dst && dst_metric(dst, RTAX_QUICKACK)) || 340 (icsk->icsk_ack.quick && !inet_csk_in_pingpong_mode(sk)); 341 } 342 343 static void tcp_ecn_queue_cwr(struct tcp_sock *tp) 344 { 345 if (tcp_ecn_mode_rfc3168(tp)) 346 tp->ecn_flags |= TCP_ECN_QUEUE_CWR; 347 } 348 349 static void tcp_ecn_accept_cwr(struct sock *sk, const struct sk_buff *skb) 350 { 351 if (tcp_hdr(skb)->cwr) { 352 tcp_sk(sk)->ecn_flags &= ~TCP_ECN_DEMAND_CWR; 353 354 /* If the sender is telling us it has entered CWR, then its 355 * cwnd may be very low (even just 1 packet), so we should ACK 356 * immediately. 357 */ 358 if (TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) 359 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW; 360 } 361 } 362 363 static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp) 364 { 365 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; 366 } 367 368 static void tcp_data_ecn_check(struct sock *sk, const struct sk_buff *skb) 369 { 370 struct tcp_sock *tp = tcp_sk(sk); 371 372 if (tcp_ecn_disabled(tp)) 373 return; 374 375 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) { 376 case INET_ECN_NOT_ECT: 377 /* Funny extension: if ECT is not set on a segment, 378 * and we already seen ECT on a previous segment, 379 * it is probably a retransmit. 380 */ 381 if (tp->ecn_flags & TCP_ECN_SEEN) 382 tcp_enter_quickack_mode(sk, 2); 383 break; 384 case INET_ECN_CE: 385 if (tcp_ca_needs_ecn(sk)) 386 tcp_ca_event(sk, CA_EVENT_ECN_IS_CE); 387 388 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) { 389 /* Better not delay acks, sender can have a very low cwnd */ 390 tcp_enter_quickack_mode(sk, 2); 391 tp->ecn_flags |= TCP_ECN_DEMAND_CWR; 392 } 393 tp->ecn_flags |= TCP_ECN_SEEN; 394 break; 395 default: 396 if (tcp_ca_needs_ecn(sk)) 397 tcp_ca_event(sk, CA_EVENT_ECN_NO_CE); 398 tp->ecn_flags |= TCP_ECN_SEEN; 399 break; 400 } 401 } 402 403 static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th) 404 { 405 if (tcp_ecn_mode_rfc3168(tp) && (!th->ece || th->cwr)) 406 tcp_ecn_mode_set(tp, TCP_ECN_DISABLED); 407 } 408 409 static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th) 410 { 411 if (tcp_ecn_mode_rfc3168(tp) && (!th->ece || !th->cwr)) 412 tcp_ecn_mode_set(tp, TCP_ECN_DISABLED); 413 } 414 415 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th) 416 { 417 if (th->ece && !th->syn && tcp_ecn_mode_rfc3168(tp)) 418 return true; 419 return false; 420 } 421 422 static void tcp_count_delivered_ce(struct tcp_sock *tp, u32 ecn_count) 423 { 424 tp->delivered_ce += ecn_count; 425 } 426 427 /* Updates the delivered and delivered_ce counts */ 428 static void tcp_count_delivered(struct tcp_sock *tp, u32 delivered, 429 bool ece_ack) 430 { 431 tp->delivered += delivered; 432 if (ece_ack) 433 tcp_count_delivered_ce(tp, delivered); 434 } 435 436 /* Buffer size and advertised window tuning. 437 * 438 * 1. Tuning sk->sk_sndbuf, when connection enters established state. 439 */ 440 441 static void tcp_sndbuf_expand(struct sock *sk) 442 { 443 const struct tcp_sock *tp = tcp_sk(sk); 444 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 445 int sndmem, per_mss; 446 u32 nr_segs; 447 448 /* Worst case is non GSO/TSO : each frame consumes one skb 449 * and skb->head is kmalloced using power of two area of memory 450 */ 451 per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) + 452 MAX_TCP_HEADER + 453 SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); 454 455 per_mss = roundup_pow_of_two(per_mss) + 456 SKB_DATA_ALIGN(sizeof(struct sk_buff)); 457 458 nr_segs = max_t(u32, TCP_INIT_CWND, tcp_snd_cwnd(tp)); 459 nr_segs = max_t(u32, nr_segs, tp->reordering + 1); 460 461 /* Fast Recovery (RFC 5681 3.2) : 462 * Cubic needs 1.7 factor, rounded to 2 to include 463 * extra cushion (application might react slowly to EPOLLOUT) 464 */ 465 sndmem = ca_ops->sndbuf_expand ? ca_ops->sndbuf_expand(sk) : 2; 466 sndmem *= nr_segs * per_mss; 467 468 if (sk->sk_sndbuf < sndmem) 469 WRITE_ONCE(sk->sk_sndbuf, 470 min(sndmem, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_wmem[2]))); 471 } 472 473 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh) 474 * 475 * All tcp_full_space() is split to two parts: "network" buffer, allocated 476 * forward and advertised in receiver window (tp->rcv_wnd) and 477 * "application buffer", required to isolate scheduling/application 478 * latencies from network. 479 * window_clamp is maximal advertised window. It can be less than 480 * tcp_full_space(), in this case tcp_full_space() - window_clamp 481 * is reserved for "application" buffer. The less window_clamp is 482 * the smoother our behaviour from viewpoint of network, but the lower 483 * throughput and the higher sensitivity of the connection to losses. 8) 484 * 485 * rcv_ssthresh is more strict window_clamp used at "slow start" 486 * phase to predict further behaviour of this connection. 487 * It is used for two goals: 488 * - to enforce header prediction at sender, even when application 489 * requires some significant "application buffer". It is check #1. 490 * - to prevent pruning of receive queue because of misprediction 491 * of receiver window. Check #2. 492 * 493 * The scheme does not work when sender sends good segments opening 494 * window and then starts to feed us spaghetti. But it should work 495 * in common situations. Otherwise, we have to rely on queue collapsing. 496 */ 497 498 /* Slow part of check#2. */ 499 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb, 500 unsigned int skbtruesize) 501 { 502 const struct tcp_sock *tp = tcp_sk(sk); 503 /* Optimize this! */ 504 int truesize = tcp_win_from_space(sk, skbtruesize) >> 1; 505 int window = tcp_win_from_space(sk, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2])) >> 1; 506 507 while (tp->rcv_ssthresh <= window) { 508 if (truesize <= skb->len) 509 return 2 * inet_csk(sk)->icsk_ack.rcv_mss; 510 511 truesize >>= 1; 512 window >>= 1; 513 } 514 return 0; 515 } 516 517 /* Even if skb appears to have a bad len/truesize ratio, TCP coalescing 518 * can play nice with us, as sk_buff and skb->head might be either 519 * freed or shared with up to MAX_SKB_FRAGS segments. 520 * Only give a boost to drivers using page frag(s) to hold the frame(s), 521 * and if no payload was pulled in skb->head before reaching us. 522 */ 523 static u32 truesize_adjust(bool adjust, const struct sk_buff *skb) 524 { 525 u32 truesize = skb->truesize; 526 527 if (adjust && !skb_headlen(skb)) { 528 truesize -= SKB_TRUESIZE(skb_end_offset(skb)); 529 /* paranoid check, some drivers might be buggy */ 530 if (unlikely((int)truesize < (int)skb->len)) 531 truesize = skb->truesize; 532 } 533 return truesize; 534 } 535 536 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb, 537 bool adjust) 538 { 539 struct tcp_sock *tp = tcp_sk(sk); 540 int room; 541 542 room = min_t(int, tp->window_clamp, tcp_space(sk)) - tp->rcv_ssthresh; 543 544 if (room <= 0) 545 return; 546 547 /* Check #1 */ 548 if (!tcp_under_memory_pressure(sk)) { 549 unsigned int truesize = truesize_adjust(adjust, skb); 550 int incr; 551 552 /* Check #2. Increase window, if skb with such overhead 553 * will fit to rcvbuf in future. 554 */ 555 if (tcp_win_from_space(sk, truesize) <= skb->len) 556 incr = 2 * tp->advmss; 557 else 558 incr = __tcp_grow_window(sk, skb, truesize); 559 560 if (incr) { 561 incr = max_t(int, incr, 2 * skb->len); 562 tp->rcv_ssthresh += min(room, incr); 563 inet_csk(sk)->icsk_ack.quick |= 1; 564 } 565 } else { 566 /* Under pressure: 567 * Adjust rcv_ssthresh according to reserved mem 568 */ 569 tcp_adjust_rcv_ssthresh(sk); 570 } 571 } 572 573 /* 3. Try to fixup all. It is made immediately after connection enters 574 * established state. 575 */ 576 static void tcp_init_buffer_space(struct sock *sk) 577 { 578 int tcp_app_win = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_app_win); 579 struct tcp_sock *tp = tcp_sk(sk); 580 int maxwin; 581 582 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK)) 583 tcp_sndbuf_expand(sk); 584 585 tcp_mstamp_refresh(tp); 586 tp->rcvq_space.time = tp->tcp_mstamp; 587 tp->rcvq_space.seq = tp->copied_seq; 588 589 maxwin = tcp_full_space(sk); 590 591 if (tp->window_clamp >= maxwin) { 592 WRITE_ONCE(tp->window_clamp, maxwin); 593 594 if (tcp_app_win && maxwin > 4 * tp->advmss) 595 WRITE_ONCE(tp->window_clamp, 596 max(maxwin - (maxwin >> tcp_app_win), 597 4 * tp->advmss)); 598 } 599 600 /* Force reservation of one segment. */ 601 if (tcp_app_win && 602 tp->window_clamp > 2 * tp->advmss && 603 tp->window_clamp + tp->advmss > maxwin) 604 WRITE_ONCE(tp->window_clamp, 605 max(2 * tp->advmss, maxwin - tp->advmss)); 606 607 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp); 608 tp->snd_cwnd_stamp = tcp_jiffies32; 609 tp->rcvq_space.space = min3(tp->rcv_ssthresh, tp->rcv_wnd, 610 (u32)TCP_INIT_CWND * tp->advmss); 611 } 612 613 /* 4. Recalculate window clamp after socket hit its memory bounds. */ 614 static void tcp_clamp_window(struct sock *sk) 615 { 616 struct tcp_sock *tp = tcp_sk(sk); 617 struct inet_connection_sock *icsk = inet_csk(sk); 618 struct net *net = sock_net(sk); 619 int rmem2; 620 621 icsk->icsk_ack.quick = 0; 622 rmem2 = READ_ONCE(net->ipv4.sysctl_tcp_rmem[2]); 623 624 if (sk->sk_rcvbuf < rmem2 && 625 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) && 626 !tcp_under_memory_pressure(sk) && 627 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) { 628 WRITE_ONCE(sk->sk_rcvbuf, 629 min(atomic_read(&sk->sk_rmem_alloc), rmem2)); 630 } 631 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf) 632 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss); 633 } 634 635 /* Initialize RCV_MSS value. 636 * RCV_MSS is an our guess about MSS used by the peer. 637 * We haven't any direct information about the MSS. 638 * It's better to underestimate the RCV_MSS rather than overestimate. 639 * Overestimations make us ACKing less frequently than needed. 640 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss(). 641 */ 642 void tcp_initialize_rcv_mss(struct sock *sk) 643 { 644 const struct tcp_sock *tp = tcp_sk(sk); 645 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache); 646 647 hint = min(hint, tp->rcv_wnd / 2); 648 hint = min(hint, TCP_MSS_DEFAULT); 649 hint = max(hint, TCP_MIN_MSS); 650 651 inet_csk(sk)->icsk_ack.rcv_mss = hint; 652 } 653 EXPORT_IPV6_MOD(tcp_initialize_rcv_mss); 654 655 /* Receiver "autotuning" code. 656 * 657 * The algorithm for RTT estimation w/o timestamps is based on 658 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL. 659 * <https://public.lanl.gov/radiant/pubs.html#DRS> 660 * 661 * More detail on this code can be found at 662 * <http://staff.psc.edu/jheffner/>, 663 * though this reference is out of date. A new paper 664 * is pending. 665 */ 666 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep) 667 { 668 u32 new_sample = tp->rcv_rtt_est.rtt_us; 669 long m = sample; 670 671 if (new_sample != 0) { 672 /* If we sample in larger samples in the non-timestamp 673 * case, we could grossly overestimate the RTT especially 674 * with chatty applications or bulk transfer apps which 675 * are stalled on filesystem I/O. 676 * 677 * Also, since we are only going for a minimum in the 678 * non-timestamp case, we do not smooth things out 679 * else with timestamps disabled convergence takes too 680 * long. 681 */ 682 if (!win_dep) { 683 m -= (new_sample >> 3); 684 new_sample += m; 685 } else { 686 m <<= 3; 687 if (m < new_sample) 688 new_sample = m; 689 } 690 } else { 691 /* No previous measure. */ 692 new_sample = m << 3; 693 } 694 695 tp->rcv_rtt_est.rtt_us = new_sample; 696 } 697 698 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp) 699 { 700 u32 delta_us; 701 702 if (tp->rcv_rtt_est.time == 0) 703 goto new_measure; 704 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq)) 705 return; 706 delta_us = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcv_rtt_est.time); 707 if (!delta_us) 708 delta_us = 1; 709 tcp_rcv_rtt_update(tp, delta_us, 1); 710 711 new_measure: 712 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd; 713 tp->rcv_rtt_est.time = tp->tcp_mstamp; 714 } 715 716 static s32 tcp_rtt_tsopt_us(const struct tcp_sock *tp) 717 { 718 u32 delta, delta_us; 719 720 delta = tcp_time_stamp_ts(tp) - tp->rx_opt.rcv_tsecr; 721 if (tp->tcp_usec_ts) 722 return delta; 723 724 if (likely(delta < INT_MAX / (USEC_PER_SEC / TCP_TS_HZ))) { 725 if (!delta) 726 delta = 1; 727 delta_us = delta * (USEC_PER_SEC / TCP_TS_HZ); 728 return delta_us; 729 } 730 return -1; 731 } 732 733 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, 734 const struct sk_buff *skb) 735 { 736 struct tcp_sock *tp = tcp_sk(sk); 737 738 if (tp->rx_opt.rcv_tsecr == tp->rcv_rtt_last_tsecr) 739 return; 740 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr; 741 742 if (TCP_SKB_CB(skb)->end_seq - 743 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss) { 744 s32 delta = tcp_rtt_tsopt_us(tp); 745 746 if (delta >= 0) 747 tcp_rcv_rtt_update(tp, delta, 0); 748 } 749 } 750 751 /* 752 * This function should be called every time data is copied to user space. 753 * It calculates the appropriate TCP receive buffer space. 754 */ 755 void tcp_rcv_space_adjust(struct sock *sk) 756 { 757 struct tcp_sock *tp = tcp_sk(sk); 758 u32 copied; 759 int time; 760 761 trace_tcp_rcv_space_adjust(sk); 762 763 tcp_mstamp_refresh(tp); 764 time = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcvq_space.time); 765 if (time < (tp->rcv_rtt_est.rtt_us >> 3) || tp->rcv_rtt_est.rtt_us == 0) 766 return; 767 768 /* Number of bytes copied to user in last RTT */ 769 copied = tp->copied_seq - tp->rcvq_space.seq; 770 if (copied <= tp->rcvq_space.space) 771 goto new_measure; 772 773 /* A bit of theory : 774 * copied = bytes received in previous RTT, our base window 775 * To cope with packet losses, we need a 2x factor 776 * To cope with slow start, and sender growing its cwin by 100 % 777 * every RTT, we need a 4x factor, because the ACK we are sending 778 * now is for the next RTT, not the current one : 779 * <prev RTT . ><current RTT .. ><next RTT .... > 780 */ 781 782 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf) && 783 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) { 784 u64 rcvwin, grow; 785 int rcvbuf; 786 787 /* minimal window to cope with packet losses, assuming 788 * steady state. Add some cushion because of small variations. 789 */ 790 rcvwin = ((u64)copied << 1) + 16 * tp->advmss; 791 792 /* Accommodate for sender rate increase (eg. slow start) */ 793 grow = rcvwin * (copied - tp->rcvq_space.space); 794 do_div(grow, tp->rcvq_space.space); 795 rcvwin += (grow << 1); 796 797 rcvbuf = min_t(u64, tcp_space_from_win(sk, rcvwin), 798 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2])); 799 if (rcvbuf > sk->sk_rcvbuf) { 800 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf); 801 802 /* Make the window clamp follow along. */ 803 WRITE_ONCE(tp->window_clamp, 804 tcp_win_from_space(sk, rcvbuf)); 805 } 806 } 807 tp->rcvq_space.space = copied; 808 809 new_measure: 810 tp->rcvq_space.seq = tp->copied_seq; 811 tp->rcvq_space.time = tp->tcp_mstamp; 812 } 813 814 static void tcp_save_lrcv_flowlabel(struct sock *sk, const struct sk_buff *skb) 815 { 816 #if IS_ENABLED(CONFIG_IPV6) 817 struct inet_connection_sock *icsk = inet_csk(sk); 818 819 if (skb->protocol == htons(ETH_P_IPV6)) 820 icsk->icsk_ack.lrcv_flowlabel = ntohl(ip6_flowlabel(ipv6_hdr(skb))); 821 #endif 822 } 823 824 /* There is something which you must keep in mind when you analyze the 825 * behavior of the tp->ato delayed ack timeout interval. When a 826 * connection starts up, we want to ack as quickly as possible. The 827 * problem is that "good" TCP's do slow start at the beginning of data 828 * transmission. The means that until we send the first few ACK's the 829 * sender will sit on his end and only queue most of his data, because 830 * he can only send snd_cwnd unacked packets at any given time. For 831 * each ACK we send, he increments snd_cwnd and transmits more of his 832 * queue. -DaveM 833 */ 834 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb) 835 { 836 struct tcp_sock *tp = tcp_sk(sk); 837 struct inet_connection_sock *icsk = inet_csk(sk); 838 u32 now; 839 840 inet_csk_schedule_ack(sk); 841 842 tcp_measure_rcv_mss(sk, skb); 843 844 tcp_rcv_rtt_measure(tp); 845 846 now = tcp_jiffies32; 847 848 if (!icsk->icsk_ack.ato) { 849 /* The _first_ data packet received, initialize 850 * delayed ACK engine. 851 */ 852 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS); 853 icsk->icsk_ack.ato = TCP_ATO_MIN; 854 } else { 855 int m = now - icsk->icsk_ack.lrcvtime; 856 857 if (m <= TCP_ATO_MIN / 2) { 858 /* The fastest case is the first. */ 859 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2; 860 } else if (m < icsk->icsk_ack.ato) { 861 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m; 862 if (icsk->icsk_ack.ato > icsk->icsk_rto) 863 icsk->icsk_ack.ato = icsk->icsk_rto; 864 } else if (m > icsk->icsk_rto) { 865 /* Too long gap. Apparently sender failed to 866 * restart window, so that we send ACKs quickly. 867 */ 868 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS); 869 } 870 } 871 icsk->icsk_ack.lrcvtime = now; 872 tcp_save_lrcv_flowlabel(sk, skb); 873 874 tcp_data_ecn_check(sk, skb); 875 876 if (skb->len >= 128) 877 tcp_grow_window(sk, skb, true); 878 } 879 880 /* Called to compute a smoothed rtt estimate. The data fed to this 881 * routine either comes from timestamps, or from segments that were 882 * known _not_ to have been retransmitted [see Karn/Partridge 883 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88 884 * piece by Van Jacobson. 885 * NOTE: the next three routines used to be one big routine. 886 * To save cycles in the RFC 1323 implementation it was better to break 887 * it up into three procedures. -- erics 888 */ 889 static void tcp_rtt_estimator(struct sock *sk, long mrtt_us) 890 { 891 struct tcp_sock *tp = tcp_sk(sk); 892 long m = mrtt_us; /* RTT */ 893 u32 srtt = tp->srtt_us; 894 895 /* The following amusing code comes from Jacobson's 896 * article in SIGCOMM '88. Note that rtt and mdev 897 * are scaled versions of rtt and mean deviation. 898 * This is designed to be as fast as possible 899 * m stands for "measurement". 900 * 901 * On a 1990 paper the rto value is changed to: 902 * RTO = rtt + 4 * mdev 903 * 904 * Funny. This algorithm seems to be very broken. 905 * These formulae increase RTO, when it should be decreased, increase 906 * too slowly, when it should be increased quickly, decrease too quickly 907 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely 908 * does not matter how to _calculate_ it. Seems, it was trap 909 * that VJ failed to avoid. 8) 910 */ 911 if (srtt != 0) { 912 m -= (srtt >> 3); /* m is now error in rtt est */ 913 srtt += m; /* rtt = 7/8 rtt + 1/8 new */ 914 if (m < 0) { 915 m = -m; /* m is now abs(error) */ 916 m -= (tp->mdev_us >> 2); /* similar update on mdev */ 917 /* This is similar to one of Eifel findings. 918 * Eifel blocks mdev updates when rtt decreases. 919 * This solution is a bit different: we use finer gain 920 * for mdev in this case (alpha*beta). 921 * Like Eifel it also prevents growth of rto, 922 * but also it limits too fast rto decreases, 923 * happening in pure Eifel. 924 */ 925 if (m > 0) 926 m >>= 3; 927 } else { 928 m -= (tp->mdev_us >> 2); /* similar update on mdev */ 929 } 930 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */ 931 if (tp->mdev_us > tp->mdev_max_us) { 932 tp->mdev_max_us = tp->mdev_us; 933 if (tp->mdev_max_us > tp->rttvar_us) 934 tp->rttvar_us = tp->mdev_max_us; 935 } 936 if (after(tp->snd_una, tp->rtt_seq)) { 937 if (tp->mdev_max_us < tp->rttvar_us) 938 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2; 939 tp->rtt_seq = tp->snd_nxt; 940 tp->mdev_max_us = tcp_rto_min_us(sk); 941 942 tcp_bpf_rtt(sk, mrtt_us, srtt); 943 } 944 } else { 945 /* no previous measure. */ 946 srtt = m << 3; /* take the measured time to be rtt */ 947 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */ 948 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk)); 949 tp->mdev_max_us = tp->rttvar_us; 950 tp->rtt_seq = tp->snd_nxt; 951 952 tcp_bpf_rtt(sk, mrtt_us, srtt); 953 } 954 tp->srtt_us = max(1U, srtt); 955 } 956 957 static void tcp_update_pacing_rate(struct sock *sk) 958 { 959 const struct tcp_sock *tp = tcp_sk(sk); 960 u64 rate; 961 962 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */ 963 rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3); 964 965 /* current rate is (cwnd * mss) / srtt 966 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate. 967 * In Congestion Avoidance phase, set it to 120 % the current rate. 968 * 969 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh) 970 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching 971 * end of slow start and should slow down. 972 */ 973 if (tcp_snd_cwnd(tp) < tp->snd_ssthresh / 2) 974 rate *= READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_pacing_ss_ratio); 975 else 976 rate *= READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_pacing_ca_ratio); 977 978 rate *= max(tcp_snd_cwnd(tp), tp->packets_out); 979 980 if (likely(tp->srtt_us)) 981 do_div(rate, tp->srtt_us); 982 983 /* WRITE_ONCE() is needed because sch_fq fetches sk_pacing_rate 984 * without any lock. We want to make sure compiler wont store 985 * intermediate values in this location. 986 */ 987 WRITE_ONCE(sk->sk_pacing_rate, 988 min_t(u64, rate, READ_ONCE(sk->sk_max_pacing_rate))); 989 } 990 991 /* Calculate rto without backoff. This is the second half of Van Jacobson's 992 * routine referred to above. 993 */ 994 static void tcp_set_rto(struct sock *sk) 995 { 996 const struct tcp_sock *tp = tcp_sk(sk); 997 /* Old crap is replaced with new one. 8) 998 * 999 * More seriously: 1000 * 1. If rtt variance happened to be less 50msec, it is hallucination. 1001 * It cannot be less due to utterly erratic ACK generation made 1002 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_ 1003 * to do with delayed acks, because at cwnd>2 true delack timeout 1004 * is invisible. Actually, Linux-2.4 also generates erratic 1005 * ACKs in some circumstances. 1006 */ 1007 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp); 1008 1009 /* 2. Fixups made earlier cannot be right. 1010 * If we do not estimate RTO correctly without them, 1011 * all the algo is pure shit and should be replaced 1012 * with correct one. It is exactly, which we pretend to do. 1013 */ 1014 1015 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo 1016 * guarantees that rto is higher. 1017 */ 1018 tcp_bound_rto(sk); 1019 } 1020 1021 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst) 1022 { 1023 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0); 1024 1025 if (!cwnd) 1026 cwnd = TCP_INIT_CWND; 1027 return min_t(__u32, cwnd, tp->snd_cwnd_clamp); 1028 } 1029 1030 struct tcp_sacktag_state { 1031 /* Timestamps for earliest and latest never-retransmitted segment 1032 * that was SACKed. RTO needs the earliest RTT to stay conservative, 1033 * but congestion control should still get an accurate delay signal. 1034 */ 1035 u64 first_sackt; 1036 u64 last_sackt; 1037 u32 reord; 1038 u32 sack_delivered; 1039 int flag; 1040 unsigned int mss_now; 1041 struct rate_sample *rate; 1042 }; 1043 1044 /* Take a notice that peer is sending D-SACKs. Skip update of data delivery 1045 * and spurious retransmission information if this DSACK is unlikely caused by 1046 * sender's action: 1047 * - DSACKed sequence range is larger than maximum receiver's window. 1048 * - Total no. of DSACKed segments exceed the total no. of retransmitted segs. 1049 */ 1050 static u32 tcp_dsack_seen(struct tcp_sock *tp, u32 start_seq, 1051 u32 end_seq, struct tcp_sacktag_state *state) 1052 { 1053 u32 seq_len, dup_segs = 1; 1054 1055 if (!before(start_seq, end_seq)) 1056 return 0; 1057 1058 seq_len = end_seq - start_seq; 1059 /* Dubious DSACK: DSACKed range greater than maximum advertised rwnd */ 1060 if (seq_len > tp->max_window) 1061 return 0; 1062 if (seq_len > tp->mss_cache) 1063 dup_segs = DIV_ROUND_UP(seq_len, tp->mss_cache); 1064 else if (tp->tlp_high_seq && tp->tlp_high_seq == end_seq) 1065 state->flag |= FLAG_DSACK_TLP; 1066 1067 tp->dsack_dups += dup_segs; 1068 /* Skip the DSACK if dup segs weren't retransmitted by sender */ 1069 if (tp->dsack_dups > tp->total_retrans) 1070 return 0; 1071 1072 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN; 1073 /* We increase the RACK ordering window in rounds where we receive 1074 * DSACKs that may have been due to reordering causing RACK to trigger 1075 * a spurious fast recovery. Thus RACK ignores DSACKs that happen 1076 * without having seen reordering, or that match TLP probes (TLP 1077 * is timer-driven, not triggered by RACK). 1078 */ 1079 if (tp->reord_seen && !(state->flag & FLAG_DSACK_TLP)) 1080 tp->rack.dsack_seen = 1; 1081 1082 state->flag |= FLAG_DSACKING_ACK; 1083 /* A spurious retransmission is delivered */ 1084 state->sack_delivered += dup_segs; 1085 1086 return dup_segs; 1087 } 1088 1089 /* It's reordering when higher sequence was delivered (i.e. sacked) before 1090 * some lower never-retransmitted sequence ("low_seq"). The maximum reordering 1091 * distance is approximated in full-mss packet distance ("reordering"). 1092 */ 1093 static void tcp_check_sack_reordering(struct sock *sk, const u32 low_seq, 1094 const int ts) 1095 { 1096 struct tcp_sock *tp = tcp_sk(sk); 1097 const u32 mss = tp->mss_cache; 1098 u32 fack, metric; 1099 1100 fack = tcp_highest_sack_seq(tp); 1101 if (!before(low_seq, fack)) 1102 return; 1103 1104 metric = fack - low_seq; 1105 if ((metric > tp->reordering * mss) && mss) { 1106 #if FASTRETRANS_DEBUG > 1 1107 pr_debug("Disorder%d %d %u f%u s%u rr%d\n", 1108 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state, 1109 tp->reordering, 1110 0, 1111 tp->sacked_out, 1112 tp->undo_marker ? tp->undo_retrans : 0); 1113 #endif 1114 tp->reordering = min_t(u32, (metric + mss - 1) / mss, 1115 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_max_reordering)); 1116 } 1117 1118 /* This exciting event is worth to be remembered. 8) */ 1119 tp->reord_seen++; 1120 NET_INC_STATS(sock_net(sk), 1121 ts ? LINUX_MIB_TCPTSREORDER : LINUX_MIB_TCPSACKREORDER); 1122 } 1123 1124 /* This must be called before lost_out or retrans_out are updated 1125 * on a new loss, because we want to know if all skbs previously 1126 * known to be lost have already been retransmitted, indicating 1127 * that this newly lost skb is our next skb to retransmit. 1128 */ 1129 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb) 1130 { 1131 if ((!tp->retransmit_skb_hint && tp->retrans_out >= tp->lost_out) || 1132 (tp->retransmit_skb_hint && 1133 before(TCP_SKB_CB(skb)->seq, 1134 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))) 1135 tp->retransmit_skb_hint = skb; 1136 } 1137 1138 /* Sum the number of packets on the wire we have marked as lost, and 1139 * notify the congestion control module that the given skb was marked lost. 1140 */ 1141 static void tcp_notify_skb_loss_event(struct tcp_sock *tp, const struct sk_buff *skb) 1142 { 1143 tp->lost += tcp_skb_pcount(skb); 1144 } 1145 1146 void tcp_mark_skb_lost(struct sock *sk, struct sk_buff *skb) 1147 { 1148 __u8 sacked = TCP_SKB_CB(skb)->sacked; 1149 struct tcp_sock *tp = tcp_sk(sk); 1150 1151 if (sacked & TCPCB_SACKED_ACKED) 1152 return; 1153 1154 tcp_verify_retransmit_hint(tp, skb); 1155 if (sacked & TCPCB_LOST) { 1156 if (sacked & TCPCB_SACKED_RETRANS) { 1157 /* Account for retransmits that are lost again */ 1158 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 1159 tp->retrans_out -= tcp_skb_pcount(skb); 1160 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT, 1161 tcp_skb_pcount(skb)); 1162 tcp_notify_skb_loss_event(tp, skb); 1163 } 1164 } else { 1165 tp->lost_out += tcp_skb_pcount(skb); 1166 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 1167 tcp_notify_skb_loss_event(tp, skb); 1168 } 1169 } 1170 1171 /* This procedure tags the retransmission queue when SACKs arrive. 1172 * 1173 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L). 1174 * Packets in queue with these bits set are counted in variables 1175 * sacked_out, retrans_out and lost_out, correspondingly. 1176 * 1177 * Valid combinations are: 1178 * Tag InFlight Description 1179 * 0 1 - orig segment is in flight. 1180 * S 0 - nothing flies, orig reached receiver. 1181 * L 0 - nothing flies, orig lost by net. 1182 * R 2 - both orig and retransmit are in flight. 1183 * L|R 1 - orig is lost, retransmit is in flight. 1184 * S|R 1 - orig reached receiver, retrans is still in flight. 1185 * (L|S|R is logically valid, it could occur when L|R is sacked, 1186 * but it is equivalent to plain S and code short-circuits it to S. 1187 * L|S is logically invalid, it would mean -1 packet in flight 8)) 1188 * 1189 * These 6 states form finite state machine, controlled by the following events: 1190 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue()) 1191 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue()) 1192 * 3. Loss detection event of two flavors: 1193 * A. Scoreboard estimator decided the packet is lost. 1194 * A'. Reno "three dupacks" marks head of queue lost. 1195 * B. SACK arrives sacking SND.NXT at the moment, when the 1196 * segment was retransmitted. 1197 * 4. D-SACK added new rule: D-SACK changes any tag to S. 1198 * 1199 * It is pleasant to note, that state diagram turns out to be commutative, 1200 * so that we are allowed not to be bothered by order of our actions, 1201 * when multiple events arrive simultaneously. (see the function below). 1202 * 1203 * Reordering detection. 1204 * -------------------- 1205 * Reordering metric is maximal distance, which a packet can be displaced 1206 * in packet stream. With SACKs we can estimate it: 1207 * 1208 * 1. SACK fills old hole and the corresponding segment was not 1209 * ever retransmitted -> reordering. Alas, we cannot use it 1210 * when segment was retransmitted. 1211 * 2. The last flaw is solved with D-SACK. D-SACK arrives 1212 * for retransmitted and already SACKed segment -> reordering.. 1213 * Both of these heuristics are not used in Loss state, when we cannot 1214 * account for retransmits accurately. 1215 * 1216 * SACK block validation. 1217 * ---------------------- 1218 * 1219 * SACK block range validation checks that the received SACK block fits to 1220 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT. 1221 * Note that SND.UNA is not included to the range though being valid because 1222 * it means that the receiver is rather inconsistent with itself reporting 1223 * SACK reneging when it should advance SND.UNA. Such SACK block this is 1224 * perfectly valid, however, in light of RFC2018 which explicitly states 1225 * that "SACK block MUST reflect the newest segment. Even if the newest 1226 * segment is going to be discarded ...", not that it looks very clever 1227 * in case of head skb. Due to potentional receiver driven attacks, we 1228 * choose to avoid immediate execution of a walk in write queue due to 1229 * reneging and defer head skb's loss recovery to standard loss recovery 1230 * procedure that will eventually trigger (nothing forbids us doing this). 1231 * 1232 * Implements also blockage to start_seq wrap-around. Problem lies in the 1233 * fact that though start_seq (s) is before end_seq (i.e., not reversed), 1234 * there's no guarantee that it will be before snd_nxt (n). The problem 1235 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt 1236 * wrap (s_w): 1237 * 1238 * <- outs wnd -> <- wrapzone -> 1239 * u e n u_w e_w s n_w 1240 * | | | | | | | 1241 * |<------------+------+----- TCP seqno space --------------+---------->| 1242 * ...-- <2^31 ->| |<--------... 1243 * ...---- >2^31 ------>| |<--------... 1244 * 1245 * Current code wouldn't be vulnerable but it's better still to discard such 1246 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat 1247 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in 1248 * snd_nxt wrap -> snd_una region will then become "well defined", i.e., 1249 * equal to the ideal case (infinite seqno space without wrap caused issues). 1250 * 1251 * With D-SACK the lower bound is extended to cover sequence space below 1252 * SND.UNA down to undo_marker, which is the last point of interest. Yet 1253 * again, D-SACK block must not to go across snd_una (for the same reason as 1254 * for the normal SACK blocks, explained above). But there all simplicity 1255 * ends, TCP might receive valid D-SACKs below that. As long as they reside 1256 * fully below undo_marker they do not affect behavior in anyway and can 1257 * therefore be safely ignored. In rare cases (which are more or less 1258 * theoretical ones), the D-SACK will nicely cross that boundary due to skb 1259 * fragmentation and packet reordering past skb's retransmission. To consider 1260 * them correctly, the acceptable range must be extended even more though 1261 * the exact amount is rather hard to quantify. However, tp->max_window can 1262 * be used as an exaggerated estimate. 1263 */ 1264 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack, 1265 u32 start_seq, u32 end_seq) 1266 { 1267 /* Too far in future, or reversed (interpretation is ambiguous) */ 1268 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq)) 1269 return false; 1270 1271 /* Nasty start_seq wrap-around check (see comments above) */ 1272 if (!before(start_seq, tp->snd_nxt)) 1273 return false; 1274 1275 /* In outstanding window? ...This is valid exit for D-SACKs too. 1276 * start_seq == snd_una is non-sensical (see comments above) 1277 */ 1278 if (after(start_seq, tp->snd_una)) 1279 return true; 1280 1281 if (!is_dsack || !tp->undo_marker) 1282 return false; 1283 1284 /* ...Then it's D-SACK, and must reside below snd_una completely */ 1285 if (after(end_seq, tp->snd_una)) 1286 return false; 1287 1288 if (!before(start_seq, tp->undo_marker)) 1289 return true; 1290 1291 /* Too old */ 1292 if (!after(end_seq, tp->undo_marker)) 1293 return false; 1294 1295 /* Undo_marker boundary crossing (overestimates a lot). Known already: 1296 * start_seq < undo_marker and end_seq >= undo_marker. 1297 */ 1298 return !before(start_seq, end_seq - tp->max_window); 1299 } 1300 1301 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb, 1302 struct tcp_sack_block_wire *sp, int num_sacks, 1303 u32 prior_snd_una, struct tcp_sacktag_state *state) 1304 { 1305 struct tcp_sock *tp = tcp_sk(sk); 1306 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq); 1307 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq); 1308 u32 dup_segs; 1309 1310 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) { 1311 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECV); 1312 } else if (num_sacks > 1) { 1313 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq); 1314 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq); 1315 1316 if (after(end_seq_0, end_seq_1) || before(start_seq_0, start_seq_1)) 1317 return false; 1318 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKOFORECV); 1319 } else { 1320 return false; 1321 } 1322 1323 dup_segs = tcp_dsack_seen(tp, start_seq_0, end_seq_0, state); 1324 if (!dup_segs) { /* Skip dubious DSACK */ 1325 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKIGNOREDDUBIOUS); 1326 return false; 1327 } 1328 1329 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECVSEGS, dup_segs); 1330 1331 /* D-SACK for already forgotten data... Do dumb counting. */ 1332 if (tp->undo_marker && tp->undo_retrans > 0 && 1333 !after(end_seq_0, prior_snd_una) && 1334 after(end_seq_0, tp->undo_marker)) 1335 tp->undo_retrans = max_t(int, 0, tp->undo_retrans - dup_segs); 1336 1337 return true; 1338 } 1339 1340 /* Check if skb is fully within the SACK block. In presence of GSO skbs, 1341 * the incoming SACK may not exactly match but we can find smaller MSS 1342 * aligned portion of it that matches. Therefore we might need to fragment 1343 * which may fail and creates some hassle (caller must handle error case 1344 * returns). 1345 * 1346 * FIXME: this could be merged to shift decision code 1347 */ 1348 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb, 1349 u32 start_seq, u32 end_seq) 1350 { 1351 int err; 1352 bool in_sack; 1353 unsigned int pkt_len; 1354 unsigned int mss; 1355 1356 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && 1357 !before(end_seq, TCP_SKB_CB(skb)->end_seq); 1358 1359 if (tcp_skb_pcount(skb) > 1 && !in_sack && 1360 after(TCP_SKB_CB(skb)->end_seq, start_seq)) { 1361 mss = tcp_skb_mss(skb); 1362 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); 1363 1364 if (!in_sack) { 1365 pkt_len = start_seq - TCP_SKB_CB(skb)->seq; 1366 if (pkt_len < mss) 1367 pkt_len = mss; 1368 } else { 1369 pkt_len = end_seq - TCP_SKB_CB(skb)->seq; 1370 if (pkt_len < mss) 1371 return -EINVAL; 1372 } 1373 1374 /* Round if necessary so that SACKs cover only full MSSes 1375 * and/or the remaining small portion (if present) 1376 */ 1377 if (pkt_len > mss) { 1378 unsigned int new_len = (pkt_len / mss) * mss; 1379 if (!in_sack && new_len < pkt_len) 1380 new_len += mss; 1381 pkt_len = new_len; 1382 } 1383 1384 if (pkt_len >= skb->len && !in_sack) 1385 return 0; 1386 1387 err = tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, 1388 pkt_len, mss, GFP_ATOMIC); 1389 if (err < 0) 1390 return err; 1391 } 1392 1393 return in_sack; 1394 } 1395 1396 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */ 1397 static u8 tcp_sacktag_one(struct sock *sk, 1398 struct tcp_sacktag_state *state, u8 sacked, 1399 u32 start_seq, u32 end_seq, 1400 int dup_sack, int pcount, 1401 u64 xmit_time) 1402 { 1403 struct tcp_sock *tp = tcp_sk(sk); 1404 1405 /* Account D-SACK for retransmitted packet. */ 1406 if (dup_sack && (sacked & TCPCB_RETRANS)) { 1407 if (tp->undo_marker && tp->undo_retrans > 0 && 1408 after(end_seq, tp->undo_marker)) 1409 tp->undo_retrans = max_t(int, 0, tp->undo_retrans - pcount); 1410 if ((sacked & TCPCB_SACKED_ACKED) && 1411 before(start_seq, state->reord)) 1412 state->reord = start_seq; 1413 } 1414 1415 /* Nothing to do; acked frame is about to be dropped (was ACKed). */ 1416 if (!after(end_seq, tp->snd_una)) 1417 return sacked; 1418 1419 if (!(sacked & TCPCB_SACKED_ACKED)) { 1420 tcp_rack_advance(tp, sacked, end_seq, xmit_time); 1421 1422 if (sacked & TCPCB_SACKED_RETRANS) { 1423 /* If the segment is not tagged as lost, 1424 * we do not clear RETRANS, believing 1425 * that retransmission is still in flight. 1426 */ 1427 if (sacked & TCPCB_LOST) { 1428 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS); 1429 tp->lost_out -= pcount; 1430 tp->retrans_out -= pcount; 1431 } 1432 } else { 1433 if (!(sacked & TCPCB_RETRANS)) { 1434 /* New sack for not retransmitted frame, 1435 * which was in hole. It is reordering. 1436 */ 1437 if (before(start_seq, 1438 tcp_highest_sack_seq(tp)) && 1439 before(start_seq, state->reord)) 1440 state->reord = start_seq; 1441 1442 if (!after(end_seq, tp->high_seq)) 1443 state->flag |= FLAG_ORIG_SACK_ACKED; 1444 if (state->first_sackt == 0) 1445 state->first_sackt = xmit_time; 1446 state->last_sackt = xmit_time; 1447 } 1448 1449 if (sacked & TCPCB_LOST) { 1450 sacked &= ~TCPCB_LOST; 1451 tp->lost_out -= pcount; 1452 } 1453 } 1454 1455 sacked |= TCPCB_SACKED_ACKED; 1456 state->flag |= FLAG_DATA_SACKED; 1457 tp->sacked_out += pcount; 1458 /* Out-of-order packets delivered */ 1459 state->sack_delivered += pcount; 1460 1461 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */ 1462 if (tp->lost_skb_hint && 1463 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq)) 1464 tp->lost_cnt_hint += pcount; 1465 } 1466 1467 /* D-SACK. We can detect redundant retransmission in S|R and plain R 1468 * frames and clear it. undo_retrans is decreased above, L|R frames 1469 * are accounted above as well. 1470 */ 1471 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) { 1472 sacked &= ~TCPCB_SACKED_RETRANS; 1473 tp->retrans_out -= pcount; 1474 } 1475 1476 return sacked; 1477 } 1478 1479 /* Shift newly-SACKed bytes from this skb to the immediately previous 1480 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such. 1481 */ 1482 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *prev, 1483 struct sk_buff *skb, 1484 struct tcp_sacktag_state *state, 1485 unsigned int pcount, int shifted, int mss, 1486 bool dup_sack) 1487 { 1488 struct tcp_sock *tp = tcp_sk(sk); 1489 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */ 1490 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */ 1491 1492 BUG_ON(!pcount); 1493 1494 /* Adjust counters and hints for the newly sacked sequence 1495 * range but discard the return value since prev is already 1496 * marked. We must tag the range first because the seq 1497 * advancement below implicitly advances 1498 * tcp_highest_sack_seq() when skb is highest_sack. 1499 */ 1500 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked, 1501 start_seq, end_seq, dup_sack, pcount, 1502 tcp_skb_timestamp_us(skb)); 1503 tcp_rate_skb_delivered(sk, skb, state->rate); 1504 1505 if (skb == tp->lost_skb_hint) 1506 tp->lost_cnt_hint += pcount; 1507 1508 TCP_SKB_CB(prev)->end_seq += shifted; 1509 TCP_SKB_CB(skb)->seq += shifted; 1510 1511 tcp_skb_pcount_add(prev, pcount); 1512 WARN_ON_ONCE(tcp_skb_pcount(skb) < pcount); 1513 tcp_skb_pcount_add(skb, -pcount); 1514 1515 /* When we're adding to gso_segs == 1, gso_size will be zero, 1516 * in theory this shouldn't be necessary but as long as DSACK 1517 * code can come after this skb later on it's better to keep 1518 * setting gso_size to something. 1519 */ 1520 if (!TCP_SKB_CB(prev)->tcp_gso_size) 1521 TCP_SKB_CB(prev)->tcp_gso_size = mss; 1522 1523 /* CHECKME: To clear or not to clear? Mimics normal skb currently */ 1524 if (tcp_skb_pcount(skb) <= 1) 1525 TCP_SKB_CB(skb)->tcp_gso_size = 0; 1526 1527 /* Difference in this won't matter, both ACKed by the same cumul. ACK */ 1528 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS); 1529 1530 if (skb->len > 0) { 1531 BUG_ON(!tcp_skb_pcount(skb)); 1532 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTED); 1533 return false; 1534 } 1535 1536 /* Whole SKB was eaten :-) */ 1537 1538 if (skb == tp->retransmit_skb_hint) 1539 tp->retransmit_skb_hint = prev; 1540 if (skb == tp->lost_skb_hint) { 1541 tp->lost_skb_hint = prev; 1542 tp->lost_cnt_hint -= tcp_skb_pcount(prev); 1543 } 1544 1545 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; 1546 TCP_SKB_CB(prev)->eor = TCP_SKB_CB(skb)->eor; 1547 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 1548 TCP_SKB_CB(prev)->end_seq++; 1549 1550 if (skb == tcp_highest_sack(sk)) 1551 tcp_advance_highest_sack(sk, skb); 1552 1553 tcp_skb_collapse_tstamp(prev, skb); 1554 if (unlikely(TCP_SKB_CB(prev)->tx.delivered_mstamp)) 1555 TCP_SKB_CB(prev)->tx.delivered_mstamp = 0; 1556 1557 tcp_rtx_queue_unlink_and_free(skb, sk); 1558 1559 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKMERGED); 1560 1561 return true; 1562 } 1563 1564 /* I wish gso_size would have a bit more sane initialization than 1565 * something-or-zero which complicates things 1566 */ 1567 static int tcp_skb_seglen(const struct sk_buff *skb) 1568 { 1569 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb); 1570 } 1571 1572 /* Shifting pages past head area doesn't work */ 1573 static int skb_can_shift(const struct sk_buff *skb) 1574 { 1575 return !skb_headlen(skb) && skb_is_nonlinear(skb); 1576 } 1577 1578 int tcp_skb_shift(struct sk_buff *to, struct sk_buff *from, 1579 int pcount, int shiftlen) 1580 { 1581 /* TCP min gso_size is 8 bytes (TCP_MIN_GSO_SIZE) 1582 * Since TCP_SKB_CB(skb)->tcp_gso_segs is 16 bits, we need 1583 * to make sure not storing more than 65535 * 8 bytes per skb, 1584 * even if current MSS is bigger. 1585 */ 1586 if (unlikely(to->len + shiftlen >= 65535 * TCP_MIN_GSO_SIZE)) 1587 return 0; 1588 if (unlikely(tcp_skb_pcount(to) + pcount > 65535)) 1589 return 0; 1590 return skb_shift(to, from, shiftlen); 1591 } 1592 1593 /* Try collapsing SACK blocks spanning across multiple skbs to a single 1594 * skb. 1595 */ 1596 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb, 1597 struct tcp_sacktag_state *state, 1598 u32 start_seq, u32 end_seq, 1599 bool dup_sack) 1600 { 1601 struct tcp_sock *tp = tcp_sk(sk); 1602 struct sk_buff *prev; 1603 int mss; 1604 int pcount = 0; 1605 int len; 1606 int in_sack; 1607 1608 /* Normally R but no L won't result in plain S */ 1609 if (!dup_sack && 1610 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS) 1611 goto fallback; 1612 if (!skb_can_shift(skb)) 1613 goto fallback; 1614 /* This frame is about to be dropped (was ACKed). */ 1615 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 1616 goto fallback; 1617 1618 /* Can only happen with delayed DSACK + discard craziness */ 1619 prev = skb_rb_prev(skb); 1620 if (!prev) 1621 goto fallback; 1622 1623 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) 1624 goto fallback; 1625 1626 if (!tcp_skb_can_collapse(prev, skb)) 1627 goto fallback; 1628 1629 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && 1630 !before(end_seq, TCP_SKB_CB(skb)->end_seq); 1631 1632 if (in_sack) { 1633 len = skb->len; 1634 pcount = tcp_skb_pcount(skb); 1635 mss = tcp_skb_seglen(skb); 1636 1637 /* TODO: Fix DSACKs to not fragment already SACKed and we can 1638 * drop this restriction as unnecessary 1639 */ 1640 if (mss != tcp_skb_seglen(prev)) 1641 goto fallback; 1642 } else { 1643 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq)) 1644 goto noop; 1645 /* CHECKME: This is non-MSS split case only?, this will 1646 * cause skipped skbs due to advancing loop btw, original 1647 * has that feature too 1648 */ 1649 if (tcp_skb_pcount(skb) <= 1) 1650 goto noop; 1651 1652 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); 1653 if (!in_sack) { 1654 /* TODO: head merge to next could be attempted here 1655 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)), 1656 * though it might not be worth of the additional hassle 1657 * 1658 * ...we can probably just fallback to what was done 1659 * previously. We could try merging non-SACKed ones 1660 * as well but it probably isn't going to buy off 1661 * because later SACKs might again split them, and 1662 * it would make skb timestamp tracking considerably 1663 * harder problem. 1664 */ 1665 goto fallback; 1666 } 1667 1668 len = end_seq - TCP_SKB_CB(skb)->seq; 1669 BUG_ON(len < 0); 1670 BUG_ON(len > skb->len); 1671 1672 /* MSS boundaries should be honoured or else pcount will 1673 * severely break even though it makes things bit trickier. 1674 * Optimize common case to avoid most of the divides 1675 */ 1676 mss = tcp_skb_mss(skb); 1677 1678 /* TODO: Fix DSACKs to not fragment already SACKed and we can 1679 * drop this restriction as unnecessary 1680 */ 1681 if (mss != tcp_skb_seglen(prev)) 1682 goto fallback; 1683 1684 if (len == mss) { 1685 pcount = 1; 1686 } else if (len < mss) { 1687 goto noop; 1688 } else { 1689 pcount = len / mss; 1690 len = pcount * mss; 1691 } 1692 } 1693 1694 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */ 1695 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una)) 1696 goto fallback; 1697 1698 if (!tcp_skb_shift(prev, skb, pcount, len)) 1699 goto fallback; 1700 if (!tcp_shifted_skb(sk, prev, skb, state, pcount, len, mss, dup_sack)) 1701 goto out; 1702 1703 /* Hole filled allows collapsing with the next as well, this is very 1704 * useful when hole on every nth skb pattern happens 1705 */ 1706 skb = skb_rb_next(prev); 1707 if (!skb) 1708 goto out; 1709 1710 if (!skb_can_shift(skb) || 1711 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) || 1712 (mss != tcp_skb_seglen(skb))) 1713 goto out; 1714 1715 if (!tcp_skb_can_collapse(prev, skb)) 1716 goto out; 1717 len = skb->len; 1718 pcount = tcp_skb_pcount(skb); 1719 if (tcp_skb_shift(prev, skb, pcount, len)) 1720 tcp_shifted_skb(sk, prev, skb, state, pcount, 1721 len, mss, 0); 1722 1723 out: 1724 return prev; 1725 1726 noop: 1727 return skb; 1728 1729 fallback: 1730 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK); 1731 return NULL; 1732 } 1733 1734 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk, 1735 struct tcp_sack_block *next_dup, 1736 struct tcp_sacktag_state *state, 1737 u32 start_seq, u32 end_seq, 1738 bool dup_sack_in) 1739 { 1740 struct tcp_sock *tp = tcp_sk(sk); 1741 struct sk_buff *tmp; 1742 1743 skb_rbtree_walk_from(skb) { 1744 int in_sack = 0; 1745 bool dup_sack = dup_sack_in; 1746 1747 /* queue is in-order => we can short-circuit the walk early */ 1748 if (!before(TCP_SKB_CB(skb)->seq, end_seq)) 1749 break; 1750 1751 if (next_dup && 1752 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) { 1753 in_sack = tcp_match_skb_to_sack(sk, skb, 1754 next_dup->start_seq, 1755 next_dup->end_seq); 1756 if (in_sack > 0) 1757 dup_sack = true; 1758 } 1759 1760 /* skb reference here is a bit tricky to get right, since 1761 * shifting can eat and free both this skb and the next, 1762 * so not even _safe variant of the loop is enough. 1763 */ 1764 if (in_sack <= 0) { 1765 tmp = tcp_shift_skb_data(sk, skb, state, 1766 start_seq, end_seq, dup_sack); 1767 if (tmp) { 1768 if (tmp != skb) { 1769 skb = tmp; 1770 continue; 1771 } 1772 1773 in_sack = 0; 1774 } else { 1775 in_sack = tcp_match_skb_to_sack(sk, skb, 1776 start_seq, 1777 end_seq); 1778 } 1779 } 1780 1781 if (unlikely(in_sack < 0)) 1782 break; 1783 1784 if (in_sack) { 1785 TCP_SKB_CB(skb)->sacked = 1786 tcp_sacktag_one(sk, 1787 state, 1788 TCP_SKB_CB(skb)->sacked, 1789 TCP_SKB_CB(skb)->seq, 1790 TCP_SKB_CB(skb)->end_seq, 1791 dup_sack, 1792 tcp_skb_pcount(skb), 1793 tcp_skb_timestamp_us(skb)); 1794 tcp_rate_skb_delivered(sk, skb, state->rate); 1795 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 1796 list_del_init(&skb->tcp_tsorted_anchor); 1797 1798 if (!before(TCP_SKB_CB(skb)->seq, 1799 tcp_highest_sack_seq(tp))) 1800 tcp_advance_highest_sack(sk, skb); 1801 } 1802 } 1803 return skb; 1804 } 1805 1806 static struct sk_buff *tcp_sacktag_bsearch(struct sock *sk, u32 seq) 1807 { 1808 struct rb_node *parent, **p = &sk->tcp_rtx_queue.rb_node; 1809 struct sk_buff *skb; 1810 1811 while (*p) { 1812 parent = *p; 1813 skb = rb_to_skb(parent); 1814 if (before(seq, TCP_SKB_CB(skb)->seq)) { 1815 p = &parent->rb_left; 1816 continue; 1817 } 1818 if (!before(seq, TCP_SKB_CB(skb)->end_seq)) { 1819 p = &parent->rb_right; 1820 continue; 1821 } 1822 return skb; 1823 } 1824 return NULL; 1825 } 1826 1827 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk, 1828 u32 skip_to_seq) 1829 { 1830 if (skb && after(TCP_SKB_CB(skb)->seq, skip_to_seq)) 1831 return skb; 1832 1833 return tcp_sacktag_bsearch(sk, skip_to_seq); 1834 } 1835 1836 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb, 1837 struct sock *sk, 1838 struct tcp_sack_block *next_dup, 1839 struct tcp_sacktag_state *state, 1840 u32 skip_to_seq) 1841 { 1842 if (!next_dup) 1843 return skb; 1844 1845 if (before(next_dup->start_seq, skip_to_seq)) { 1846 skb = tcp_sacktag_skip(skb, sk, next_dup->start_seq); 1847 skb = tcp_sacktag_walk(skb, sk, NULL, state, 1848 next_dup->start_seq, next_dup->end_seq, 1849 1); 1850 } 1851 1852 return skb; 1853 } 1854 1855 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache) 1856 { 1857 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache); 1858 } 1859 1860 static int 1861 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb, 1862 u32 prior_snd_una, struct tcp_sacktag_state *state) 1863 { 1864 struct tcp_sock *tp = tcp_sk(sk); 1865 const unsigned char *ptr = (skb_transport_header(ack_skb) + 1866 TCP_SKB_CB(ack_skb)->sacked); 1867 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2); 1868 struct tcp_sack_block sp[TCP_NUM_SACKS]; 1869 struct tcp_sack_block *cache; 1870 struct sk_buff *skb; 1871 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3); 1872 int used_sacks; 1873 bool found_dup_sack = false; 1874 int i, j; 1875 int first_sack_index; 1876 1877 state->flag = 0; 1878 state->reord = tp->snd_nxt; 1879 1880 if (!tp->sacked_out) 1881 tcp_highest_sack_reset(sk); 1882 1883 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire, 1884 num_sacks, prior_snd_una, state); 1885 1886 /* Eliminate too old ACKs, but take into 1887 * account more or less fresh ones, they can 1888 * contain valid SACK info. 1889 */ 1890 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window)) 1891 return 0; 1892 1893 if (!tp->packets_out) 1894 goto out; 1895 1896 used_sacks = 0; 1897 first_sack_index = 0; 1898 for (i = 0; i < num_sacks; i++) { 1899 bool dup_sack = !i && found_dup_sack; 1900 1901 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq); 1902 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq); 1903 1904 if (!tcp_is_sackblock_valid(tp, dup_sack, 1905 sp[used_sacks].start_seq, 1906 sp[used_sacks].end_seq)) { 1907 int mib_idx; 1908 1909 if (dup_sack) { 1910 if (!tp->undo_marker) 1911 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO; 1912 else 1913 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD; 1914 } else { 1915 /* Don't count olds caused by ACK reordering */ 1916 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) && 1917 !after(sp[used_sacks].end_seq, tp->snd_una)) 1918 continue; 1919 mib_idx = LINUX_MIB_TCPSACKDISCARD; 1920 } 1921 1922 NET_INC_STATS(sock_net(sk), mib_idx); 1923 if (i == 0) 1924 first_sack_index = -1; 1925 continue; 1926 } 1927 1928 /* Ignore very old stuff early */ 1929 if (!after(sp[used_sacks].end_seq, prior_snd_una)) { 1930 if (i == 0) 1931 first_sack_index = -1; 1932 continue; 1933 } 1934 1935 used_sacks++; 1936 } 1937 1938 /* order SACK blocks to allow in order walk of the retrans queue */ 1939 for (i = used_sacks - 1; i > 0; i--) { 1940 for (j = 0; j < i; j++) { 1941 if (after(sp[j].start_seq, sp[j + 1].start_seq)) { 1942 swap(sp[j], sp[j + 1]); 1943 1944 /* Track where the first SACK block goes to */ 1945 if (j == first_sack_index) 1946 first_sack_index = j + 1; 1947 } 1948 } 1949 } 1950 1951 state->mss_now = tcp_current_mss(sk); 1952 skb = NULL; 1953 i = 0; 1954 1955 if (!tp->sacked_out) { 1956 /* It's already past, so skip checking against it */ 1957 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache); 1958 } else { 1959 cache = tp->recv_sack_cache; 1960 /* Skip empty blocks in at head of the cache */ 1961 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq && 1962 !cache->end_seq) 1963 cache++; 1964 } 1965 1966 while (i < used_sacks) { 1967 u32 start_seq = sp[i].start_seq; 1968 u32 end_seq = sp[i].end_seq; 1969 bool dup_sack = (found_dup_sack && (i == first_sack_index)); 1970 struct tcp_sack_block *next_dup = NULL; 1971 1972 if (found_dup_sack && ((i + 1) == first_sack_index)) 1973 next_dup = &sp[i + 1]; 1974 1975 /* Skip too early cached blocks */ 1976 while (tcp_sack_cache_ok(tp, cache) && 1977 !before(start_seq, cache->end_seq)) 1978 cache++; 1979 1980 /* Can skip some work by looking recv_sack_cache? */ 1981 if (tcp_sack_cache_ok(tp, cache) && !dup_sack && 1982 after(end_seq, cache->start_seq)) { 1983 1984 /* Head todo? */ 1985 if (before(start_seq, cache->start_seq)) { 1986 skb = tcp_sacktag_skip(skb, sk, start_seq); 1987 skb = tcp_sacktag_walk(skb, sk, next_dup, 1988 state, 1989 start_seq, 1990 cache->start_seq, 1991 dup_sack); 1992 } 1993 1994 /* Rest of the block already fully processed? */ 1995 if (!after(end_seq, cache->end_seq)) 1996 goto advance_sp; 1997 1998 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup, 1999 state, 2000 cache->end_seq); 2001 2002 /* ...tail remains todo... */ 2003 if (tcp_highest_sack_seq(tp) == cache->end_seq) { 2004 /* ...but better entrypoint exists! */ 2005 skb = tcp_highest_sack(sk); 2006 if (!skb) 2007 break; 2008 cache++; 2009 goto walk; 2010 } 2011 2012 skb = tcp_sacktag_skip(skb, sk, cache->end_seq); 2013 /* Check overlap against next cached too (past this one already) */ 2014 cache++; 2015 continue; 2016 } 2017 2018 if (!before(start_seq, tcp_highest_sack_seq(tp))) { 2019 skb = tcp_highest_sack(sk); 2020 if (!skb) 2021 break; 2022 } 2023 skb = tcp_sacktag_skip(skb, sk, start_seq); 2024 2025 walk: 2026 skb = tcp_sacktag_walk(skb, sk, next_dup, state, 2027 start_seq, end_seq, dup_sack); 2028 2029 advance_sp: 2030 i++; 2031 } 2032 2033 /* Clear the head of the cache sack blocks so we can skip it next time */ 2034 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) { 2035 tp->recv_sack_cache[i].start_seq = 0; 2036 tp->recv_sack_cache[i].end_seq = 0; 2037 } 2038 for (j = 0; j < used_sacks; j++) 2039 tp->recv_sack_cache[i++] = sp[j]; 2040 2041 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Loss || tp->undo_marker) 2042 tcp_check_sack_reordering(sk, state->reord, 0); 2043 2044 tcp_verify_left_out(tp); 2045 out: 2046 2047 #if FASTRETRANS_DEBUG > 0 2048 WARN_ON((int)tp->sacked_out < 0); 2049 WARN_ON((int)tp->lost_out < 0); 2050 WARN_ON((int)tp->retrans_out < 0); 2051 WARN_ON((int)tcp_packets_in_flight(tp) < 0); 2052 #endif 2053 return state->flag; 2054 } 2055 2056 /* Limits sacked_out so that sum with lost_out isn't ever larger than 2057 * packets_out. Returns false if sacked_out adjustement wasn't necessary. 2058 */ 2059 static bool tcp_limit_reno_sacked(struct tcp_sock *tp) 2060 { 2061 u32 holes; 2062 2063 holes = max(tp->lost_out, 1U); 2064 holes = min(holes, tp->packets_out); 2065 2066 if ((tp->sacked_out + holes) > tp->packets_out) { 2067 tp->sacked_out = tp->packets_out - holes; 2068 return true; 2069 } 2070 return false; 2071 } 2072 2073 /* If we receive more dupacks than we expected counting segments 2074 * in assumption of absent reordering, interpret this as reordering. 2075 * The only another reason could be bug in receiver TCP. 2076 */ 2077 static void tcp_check_reno_reordering(struct sock *sk, const int addend) 2078 { 2079 struct tcp_sock *tp = tcp_sk(sk); 2080 2081 if (!tcp_limit_reno_sacked(tp)) 2082 return; 2083 2084 tp->reordering = min_t(u32, tp->packets_out + addend, 2085 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_max_reordering)); 2086 tp->reord_seen++; 2087 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRENOREORDER); 2088 } 2089 2090 /* Emulate SACKs for SACKless connection: account for a new dupack. */ 2091 2092 static void tcp_add_reno_sack(struct sock *sk, int num_dupack, bool ece_ack) 2093 { 2094 if (num_dupack) { 2095 struct tcp_sock *tp = tcp_sk(sk); 2096 u32 prior_sacked = tp->sacked_out; 2097 s32 delivered; 2098 2099 tp->sacked_out += num_dupack; 2100 tcp_check_reno_reordering(sk, 0); 2101 delivered = tp->sacked_out - prior_sacked; 2102 if (delivered > 0) 2103 tcp_count_delivered(tp, delivered, ece_ack); 2104 tcp_verify_left_out(tp); 2105 } 2106 } 2107 2108 /* Account for ACK, ACKing some data in Reno Recovery phase. */ 2109 2110 static void tcp_remove_reno_sacks(struct sock *sk, int acked, bool ece_ack) 2111 { 2112 struct tcp_sock *tp = tcp_sk(sk); 2113 2114 if (acked > 0) { 2115 /* One ACK acked hole. The rest eat duplicate ACKs. */ 2116 tcp_count_delivered(tp, max_t(int, acked - tp->sacked_out, 1), 2117 ece_ack); 2118 if (acked - 1 >= tp->sacked_out) 2119 tp->sacked_out = 0; 2120 else 2121 tp->sacked_out -= acked - 1; 2122 } 2123 tcp_check_reno_reordering(sk, acked); 2124 tcp_verify_left_out(tp); 2125 } 2126 2127 static inline void tcp_reset_reno_sack(struct tcp_sock *tp) 2128 { 2129 tp->sacked_out = 0; 2130 } 2131 2132 void tcp_clear_retrans(struct tcp_sock *tp) 2133 { 2134 tp->retrans_out = 0; 2135 tp->lost_out = 0; 2136 tp->undo_marker = 0; 2137 tp->undo_retrans = -1; 2138 tp->sacked_out = 0; 2139 tp->rto_stamp = 0; 2140 tp->total_rto = 0; 2141 tp->total_rto_recoveries = 0; 2142 tp->total_rto_time = 0; 2143 } 2144 2145 static inline void tcp_init_undo(struct tcp_sock *tp) 2146 { 2147 tp->undo_marker = tp->snd_una; 2148 2149 /* Retransmission still in flight may cause DSACKs later. */ 2150 /* First, account for regular retransmits in flight: */ 2151 tp->undo_retrans = tp->retrans_out; 2152 /* Next, account for TLP retransmits in flight: */ 2153 if (tp->tlp_high_seq && tp->tlp_retrans) 2154 tp->undo_retrans++; 2155 /* Finally, avoid 0, because undo_retrans==0 means "can undo now": */ 2156 if (!tp->undo_retrans) 2157 tp->undo_retrans = -1; 2158 } 2159 2160 static bool tcp_is_rack(const struct sock *sk) 2161 { 2162 return READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_recovery) & 2163 TCP_RACK_LOSS_DETECTION; 2164 } 2165 2166 /* If we detect SACK reneging, forget all SACK information 2167 * and reset tags completely, otherwise preserve SACKs. If receiver 2168 * dropped its ofo queue, we will know this due to reneging detection. 2169 */ 2170 static void tcp_timeout_mark_lost(struct sock *sk) 2171 { 2172 struct tcp_sock *tp = tcp_sk(sk); 2173 struct sk_buff *skb, *head; 2174 bool is_reneg; /* is receiver reneging on SACKs? */ 2175 2176 head = tcp_rtx_queue_head(sk); 2177 is_reneg = head && (TCP_SKB_CB(head)->sacked & TCPCB_SACKED_ACKED); 2178 if (is_reneg) { 2179 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSACKRENEGING); 2180 tp->sacked_out = 0; 2181 /* Mark SACK reneging until we recover from this loss event. */ 2182 tp->is_sack_reneg = 1; 2183 } else if (tcp_is_reno(tp)) { 2184 tcp_reset_reno_sack(tp); 2185 } 2186 2187 skb = head; 2188 skb_rbtree_walk_from(skb) { 2189 if (is_reneg) 2190 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED; 2191 else if (tcp_is_rack(sk) && skb != head && 2192 tcp_rack_skb_timeout(tp, skb, 0) > 0) 2193 continue; /* Don't mark recently sent ones lost yet */ 2194 tcp_mark_skb_lost(sk, skb); 2195 } 2196 tcp_verify_left_out(tp); 2197 tcp_clear_all_retrans_hints(tp); 2198 } 2199 2200 /* Enter Loss state. */ 2201 void tcp_enter_loss(struct sock *sk) 2202 { 2203 const struct inet_connection_sock *icsk = inet_csk(sk); 2204 struct tcp_sock *tp = tcp_sk(sk); 2205 struct net *net = sock_net(sk); 2206 bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery; 2207 u8 reordering; 2208 2209 tcp_timeout_mark_lost(sk); 2210 2211 /* Reduce ssthresh if it has not yet been made inside this window. */ 2212 if (icsk->icsk_ca_state <= TCP_CA_Disorder || 2213 !after(tp->high_seq, tp->snd_una) || 2214 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) { 2215 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2216 tp->prior_cwnd = tcp_snd_cwnd(tp); 2217 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 2218 tcp_ca_event(sk, CA_EVENT_LOSS); 2219 tcp_init_undo(tp); 2220 } 2221 tcp_snd_cwnd_set(tp, tcp_packets_in_flight(tp) + 1); 2222 tp->snd_cwnd_cnt = 0; 2223 tp->snd_cwnd_stamp = tcp_jiffies32; 2224 2225 /* Timeout in disordered state after receiving substantial DUPACKs 2226 * suggests that the degree of reordering is over-estimated. 2227 */ 2228 reordering = READ_ONCE(net->ipv4.sysctl_tcp_reordering); 2229 if (icsk->icsk_ca_state <= TCP_CA_Disorder && 2230 tp->sacked_out >= reordering) 2231 tp->reordering = min_t(unsigned int, tp->reordering, 2232 reordering); 2233 2234 tcp_set_ca_state(sk, TCP_CA_Loss); 2235 tp->high_seq = tp->snd_nxt; 2236 tp->tlp_high_seq = 0; 2237 tcp_ecn_queue_cwr(tp); 2238 2239 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous 2240 * loss recovery is underway except recurring timeout(s) on 2241 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing 2242 */ 2243 tp->frto = READ_ONCE(net->ipv4.sysctl_tcp_frto) && 2244 (new_recovery || icsk->icsk_retransmits) && 2245 !inet_csk(sk)->icsk_mtup.probe_size; 2246 } 2247 2248 /* If ACK arrived pointing to a remembered SACK, it means that our 2249 * remembered SACKs do not reflect real state of receiver i.e. 2250 * receiver _host_ is heavily congested (or buggy). 2251 * 2252 * To avoid big spurious retransmission bursts due to transient SACK 2253 * scoreboard oddities that look like reneging, we give the receiver a 2254 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will 2255 * restore sanity to the SACK scoreboard. If the apparent reneging 2256 * persists until this RTO then we'll clear the SACK scoreboard. 2257 */ 2258 static bool tcp_check_sack_reneging(struct sock *sk, int *ack_flag) 2259 { 2260 if (*ack_flag & FLAG_SACK_RENEGING && 2261 *ack_flag & FLAG_SND_UNA_ADVANCED) { 2262 struct tcp_sock *tp = tcp_sk(sk); 2263 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4), 2264 msecs_to_jiffies(10)); 2265 2266 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, delay, false); 2267 *ack_flag &= ~FLAG_SET_XMIT_TIMER; 2268 return true; 2269 } 2270 return false; 2271 } 2272 2273 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs 2274 * counter when SACK is enabled (without SACK, sacked_out is used for 2275 * that purpose). 2276 * 2277 * With reordering, holes may still be in flight, so RFC3517 recovery 2278 * uses pure sacked_out (total number of SACKed segments) even though 2279 * it violates the RFC that uses duplicate ACKs, often these are equal 2280 * but when e.g. out-of-window ACKs or packet duplication occurs, 2281 * they differ. Since neither occurs due to loss, TCP should really 2282 * ignore them. 2283 */ 2284 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp) 2285 { 2286 return tp->sacked_out + 1; 2287 } 2288 2289 /* Linux NewReno/SACK/ECN state machine. 2290 * -------------------------------------- 2291 * 2292 * "Open" Normal state, no dubious events, fast path. 2293 * "Disorder" In all the respects it is "Open", 2294 * but requires a bit more attention. It is entered when 2295 * we see some SACKs or dupacks. It is split of "Open" 2296 * mainly to move some processing from fast path to slow one. 2297 * "CWR" CWND was reduced due to some Congestion Notification event. 2298 * It can be ECN, ICMP source quench, local device congestion. 2299 * "Recovery" CWND was reduced, we are fast-retransmitting. 2300 * "Loss" CWND was reduced due to RTO timeout or SACK reneging. 2301 * 2302 * tcp_fastretrans_alert() is entered: 2303 * - each incoming ACK, if state is not "Open" 2304 * - when arrived ACK is unusual, namely: 2305 * * SACK 2306 * * Duplicate ACK. 2307 * * ECN ECE. 2308 * 2309 * Counting packets in flight is pretty simple. 2310 * 2311 * in_flight = packets_out - left_out + retrans_out 2312 * 2313 * packets_out is SND.NXT-SND.UNA counted in packets. 2314 * 2315 * retrans_out is number of retransmitted segments. 2316 * 2317 * left_out is number of segments left network, but not ACKed yet. 2318 * 2319 * left_out = sacked_out + lost_out 2320 * 2321 * sacked_out: Packets, which arrived to receiver out of order 2322 * and hence not ACKed. With SACKs this number is simply 2323 * amount of SACKed data. Even without SACKs 2324 * it is easy to give pretty reliable estimate of this number, 2325 * counting duplicate ACKs. 2326 * 2327 * lost_out: Packets lost by network. TCP has no explicit 2328 * "loss notification" feedback from network (for now). 2329 * It means that this number can be only _guessed_. 2330 * Actually, it is the heuristics to predict lossage that 2331 * distinguishes different algorithms. 2332 * 2333 * F.e. after RTO, when all the queue is considered as lost, 2334 * lost_out = packets_out and in_flight = retrans_out. 2335 * 2336 * Essentially, we have now a few algorithms detecting 2337 * lost packets. 2338 * 2339 * If the receiver supports SACK: 2340 * 2341 * RFC6675/3517: It is the conventional algorithm. A packet is 2342 * considered lost if the number of higher sequence packets 2343 * SACKed is greater than or equal the DUPACK thoreshold 2344 * (reordering). This is implemented in tcp_mark_head_lost and 2345 * tcp_update_scoreboard. 2346 * 2347 * RACK (draft-ietf-tcpm-rack-01): it is a newer algorithm 2348 * (2017-) that checks timing instead of counting DUPACKs. 2349 * Essentially a packet is considered lost if it's not S/ACKed 2350 * after RTT + reordering_window, where both metrics are 2351 * dynamically measured and adjusted. This is implemented in 2352 * tcp_rack_mark_lost. 2353 * 2354 * If the receiver does not support SACK: 2355 * 2356 * NewReno (RFC6582): in Recovery we assume that one segment 2357 * is lost (classic Reno). While we are in Recovery and 2358 * a partial ACK arrives, we assume that one more packet 2359 * is lost (NewReno). This heuristics are the same in NewReno 2360 * and SACK. 2361 * 2362 * Really tricky (and requiring careful tuning) part of algorithm 2363 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue(). 2364 * The first determines the moment _when_ we should reduce CWND and, 2365 * hence, slow down forward transmission. In fact, it determines the moment 2366 * when we decide that hole is caused by loss, rather than by a reorder. 2367 * 2368 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill 2369 * holes, caused by lost packets. 2370 * 2371 * And the most logically complicated part of algorithm is undo 2372 * heuristics. We detect false retransmits due to both too early 2373 * fast retransmit (reordering) and underestimated RTO, analyzing 2374 * timestamps and D-SACKs. When we detect that some segments were 2375 * retransmitted by mistake and CWND reduction was wrong, we undo 2376 * window reduction and abort recovery phase. This logic is hidden 2377 * inside several functions named tcp_try_undo_<something>. 2378 */ 2379 2380 /* This function decides, when we should leave Disordered state 2381 * and enter Recovery phase, reducing congestion window. 2382 * 2383 * Main question: may we further continue forward transmission 2384 * with the same cwnd? 2385 */ 2386 static bool tcp_time_to_recover(struct sock *sk, int flag) 2387 { 2388 struct tcp_sock *tp = tcp_sk(sk); 2389 2390 /* Trick#1: The loss is proven. */ 2391 if (tp->lost_out) 2392 return true; 2393 2394 /* Not-A-Trick#2 : Classic rule... */ 2395 if (!tcp_is_rack(sk) && tcp_dupack_heuristics(tp) > tp->reordering) 2396 return true; 2397 2398 return false; 2399 } 2400 2401 /* Detect loss in event "A" above by marking head of queue up as lost. 2402 * For RFC3517 SACK, a segment is considered lost if it 2403 * has at least tp->reordering SACKed seqments above it; "packets" refers to 2404 * the maximum SACKed segments to pass before reaching this limit. 2405 */ 2406 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head) 2407 { 2408 struct tcp_sock *tp = tcp_sk(sk); 2409 struct sk_buff *skb; 2410 int cnt; 2411 /* Use SACK to deduce losses of new sequences sent during recovery */ 2412 const u32 loss_high = tp->snd_nxt; 2413 2414 WARN_ON(packets > tp->packets_out); 2415 skb = tp->lost_skb_hint; 2416 if (skb) { 2417 /* Head already handled? */ 2418 if (mark_head && after(TCP_SKB_CB(skb)->seq, tp->snd_una)) 2419 return; 2420 cnt = tp->lost_cnt_hint; 2421 } else { 2422 skb = tcp_rtx_queue_head(sk); 2423 cnt = 0; 2424 } 2425 2426 skb_rbtree_walk_from(skb) { 2427 /* TODO: do this better */ 2428 /* this is not the most efficient way to do this... */ 2429 tp->lost_skb_hint = skb; 2430 tp->lost_cnt_hint = cnt; 2431 2432 if (after(TCP_SKB_CB(skb)->end_seq, loss_high)) 2433 break; 2434 2435 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 2436 cnt += tcp_skb_pcount(skb); 2437 2438 if (cnt > packets) 2439 break; 2440 2441 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_LOST)) 2442 tcp_mark_skb_lost(sk, skb); 2443 2444 if (mark_head) 2445 break; 2446 } 2447 tcp_verify_left_out(tp); 2448 } 2449 2450 /* Account newly detected lost packet(s) */ 2451 2452 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit) 2453 { 2454 struct tcp_sock *tp = tcp_sk(sk); 2455 2456 if (tcp_is_sack(tp)) { 2457 int sacked_upto = tp->sacked_out - tp->reordering; 2458 if (sacked_upto >= 0) 2459 tcp_mark_head_lost(sk, sacked_upto, 0); 2460 else if (fast_rexmit) 2461 tcp_mark_head_lost(sk, 1, 1); 2462 } 2463 } 2464 2465 static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when) 2466 { 2467 return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 2468 before(tp->rx_opt.rcv_tsecr, when); 2469 } 2470 2471 /* skb is spurious retransmitted if the returned timestamp echo 2472 * reply is prior to the skb transmission time 2473 */ 2474 static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp, 2475 const struct sk_buff *skb) 2476 { 2477 return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) && 2478 tcp_tsopt_ecr_before(tp, tcp_skb_timestamp_ts(tp->tcp_usec_ts, skb)); 2479 } 2480 2481 /* Nothing was retransmitted or returned timestamp is less 2482 * than timestamp of the first retransmission. 2483 */ 2484 static inline bool tcp_packet_delayed(const struct tcp_sock *tp) 2485 { 2486 const struct sock *sk = (const struct sock *)tp; 2487 2488 if (tp->retrans_stamp && 2489 tcp_tsopt_ecr_before(tp, tp->retrans_stamp)) 2490 return true; /* got echoed TS before first retransmission */ 2491 2492 /* Check if nothing was retransmitted (retrans_stamp==0), which may 2493 * happen in fast recovery due to TSQ. But we ignore zero retrans_stamp 2494 * in TCP_SYN_SENT, since when we set FLAG_SYN_ACKED we also clear 2495 * retrans_stamp even if we had retransmitted the SYN. 2496 */ 2497 if (!tp->retrans_stamp && /* no record of a retransmit/SYN? */ 2498 sk->sk_state != TCP_SYN_SENT) /* not the FLAG_SYN_ACKED case? */ 2499 return true; /* nothing was retransmitted */ 2500 2501 return false; 2502 } 2503 2504 /* Undo procedures. */ 2505 2506 /* We can clear retrans_stamp when there are no retransmissions in the 2507 * window. It would seem that it is trivially available for us in 2508 * tp->retrans_out, however, that kind of assumptions doesn't consider 2509 * what will happen if errors occur when sending retransmission for the 2510 * second time. ...It could the that such segment has only 2511 * TCPCB_EVER_RETRANS set at the present time. It seems that checking 2512 * the head skb is enough except for some reneging corner cases that 2513 * are not worth the effort. 2514 * 2515 * Main reason for all this complexity is the fact that connection dying 2516 * time now depends on the validity of the retrans_stamp, in particular, 2517 * that successive retransmissions of a segment must not advance 2518 * retrans_stamp under any conditions. 2519 */ 2520 static bool tcp_any_retrans_done(const struct sock *sk) 2521 { 2522 const struct tcp_sock *tp = tcp_sk(sk); 2523 struct sk_buff *skb; 2524 2525 if (tp->retrans_out) 2526 return true; 2527 2528 skb = tcp_rtx_queue_head(sk); 2529 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS)) 2530 return true; 2531 2532 return false; 2533 } 2534 2535 /* If loss recovery is finished and there are no retransmits out in the 2536 * network, then we clear retrans_stamp so that upon the next loss recovery 2537 * retransmits_timed_out() and timestamp-undo are using the correct value. 2538 */ 2539 static void tcp_retrans_stamp_cleanup(struct sock *sk) 2540 { 2541 if (!tcp_any_retrans_done(sk)) 2542 tcp_sk(sk)->retrans_stamp = 0; 2543 } 2544 2545 static void DBGUNDO(struct sock *sk, const char *msg) 2546 { 2547 #if FASTRETRANS_DEBUG > 1 2548 struct tcp_sock *tp = tcp_sk(sk); 2549 struct inet_sock *inet = inet_sk(sk); 2550 2551 if (sk->sk_family == AF_INET) { 2552 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n", 2553 msg, 2554 &inet->inet_daddr, ntohs(inet->inet_dport), 2555 tcp_snd_cwnd(tp), tcp_left_out(tp), 2556 tp->snd_ssthresh, tp->prior_ssthresh, 2557 tp->packets_out); 2558 } 2559 #if IS_ENABLED(CONFIG_IPV6) 2560 else if (sk->sk_family == AF_INET6) { 2561 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n", 2562 msg, 2563 &sk->sk_v6_daddr, ntohs(inet->inet_dport), 2564 tcp_snd_cwnd(tp), tcp_left_out(tp), 2565 tp->snd_ssthresh, tp->prior_ssthresh, 2566 tp->packets_out); 2567 } 2568 #endif 2569 #endif 2570 } 2571 2572 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss) 2573 { 2574 struct tcp_sock *tp = tcp_sk(sk); 2575 2576 if (unmark_loss) { 2577 struct sk_buff *skb; 2578 2579 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) { 2580 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; 2581 } 2582 tp->lost_out = 0; 2583 tcp_clear_all_retrans_hints(tp); 2584 } 2585 2586 if (tp->prior_ssthresh) { 2587 const struct inet_connection_sock *icsk = inet_csk(sk); 2588 2589 tcp_snd_cwnd_set(tp, icsk->icsk_ca_ops->undo_cwnd(sk)); 2590 2591 if (tp->prior_ssthresh > tp->snd_ssthresh) { 2592 tp->snd_ssthresh = tp->prior_ssthresh; 2593 tcp_ecn_withdraw_cwr(tp); 2594 } 2595 } 2596 tp->snd_cwnd_stamp = tcp_jiffies32; 2597 tp->undo_marker = 0; 2598 tp->rack.advanced = 1; /* Force RACK to re-exam losses */ 2599 } 2600 2601 static inline bool tcp_may_undo(const struct tcp_sock *tp) 2602 { 2603 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp)); 2604 } 2605 2606 static bool tcp_is_non_sack_preventing_reopen(struct sock *sk) 2607 { 2608 struct tcp_sock *tp = tcp_sk(sk); 2609 2610 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) { 2611 /* Hold old state until something *above* high_seq 2612 * is ACKed. For Reno it is MUST to prevent false 2613 * fast retransmits (RFC2582). SACK TCP is safe. */ 2614 if (!tcp_any_retrans_done(sk)) 2615 tp->retrans_stamp = 0; 2616 return true; 2617 } 2618 return false; 2619 } 2620 2621 /* People celebrate: "We love our President!" */ 2622 static bool tcp_try_undo_recovery(struct sock *sk) 2623 { 2624 struct tcp_sock *tp = tcp_sk(sk); 2625 2626 if (tcp_may_undo(tp)) { 2627 int mib_idx; 2628 2629 /* Happy end! We did not retransmit anything 2630 * or our original transmission succeeded. 2631 */ 2632 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans"); 2633 tcp_undo_cwnd_reduction(sk, false); 2634 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss) 2635 mib_idx = LINUX_MIB_TCPLOSSUNDO; 2636 else 2637 mib_idx = LINUX_MIB_TCPFULLUNDO; 2638 2639 NET_INC_STATS(sock_net(sk), mib_idx); 2640 } else if (tp->rack.reo_wnd_persist) { 2641 tp->rack.reo_wnd_persist--; 2642 } 2643 if (tcp_is_non_sack_preventing_reopen(sk)) 2644 return true; 2645 tcp_set_ca_state(sk, TCP_CA_Open); 2646 tp->is_sack_reneg = 0; 2647 return false; 2648 } 2649 2650 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */ 2651 static bool tcp_try_undo_dsack(struct sock *sk) 2652 { 2653 struct tcp_sock *tp = tcp_sk(sk); 2654 2655 if (tp->undo_marker && !tp->undo_retrans) { 2656 tp->rack.reo_wnd_persist = min(TCP_RACK_RECOVERY_THRESH, 2657 tp->rack.reo_wnd_persist + 1); 2658 DBGUNDO(sk, "D-SACK"); 2659 tcp_undo_cwnd_reduction(sk, false); 2660 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKUNDO); 2661 return true; 2662 } 2663 return false; 2664 } 2665 2666 /* Undo during loss recovery after partial ACK or using F-RTO. */ 2667 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo) 2668 { 2669 struct tcp_sock *tp = tcp_sk(sk); 2670 2671 if (frto_undo || tcp_may_undo(tp)) { 2672 tcp_undo_cwnd_reduction(sk, true); 2673 2674 DBGUNDO(sk, "partial loss"); 2675 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSUNDO); 2676 if (frto_undo) 2677 NET_INC_STATS(sock_net(sk), 2678 LINUX_MIB_TCPSPURIOUSRTOS); 2679 inet_csk(sk)->icsk_retransmits = 0; 2680 if (tcp_is_non_sack_preventing_reopen(sk)) 2681 return true; 2682 if (frto_undo || tcp_is_sack(tp)) { 2683 tcp_set_ca_state(sk, TCP_CA_Open); 2684 tp->is_sack_reneg = 0; 2685 } 2686 return true; 2687 } 2688 return false; 2689 } 2690 2691 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937. 2692 * It computes the number of packets to send (sndcnt) based on packets newly 2693 * delivered: 2694 * 1) If the packets in flight is larger than ssthresh, PRR spreads the 2695 * cwnd reductions across a full RTT. 2696 * 2) Otherwise PRR uses packet conservation to send as much as delivered. 2697 * But when SND_UNA is acked without further losses, 2698 * slow starts cwnd up to ssthresh to speed up the recovery. 2699 */ 2700 static void tcp_init_cwnd_reduction(struct sock *sk) 2701 { 2702 struct tcp_sock *tp = tcp_sk(sk); 2703 2704 tp->high_seq = tp->snd_nxt; 2705 tp->tlp_high_seq = 0; 2706 tp->snd_cwnd_cnt = 0; 2707 tp->prior_cwnd = tcp_snd_cwnd(tp); 2708 tp->prr_delivered = 0; 2709 tp->prr_out = 0; 2710 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk); 2711 tcp_ecn_queue_cwr(tp); 2712 } 2713 2714 void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked, int newly_lost, int flag) 2715 { 2716 struct tcp_sock *tp = tcp_sk(sk); 2717 int sndcnt = 0; 2718 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp); 2719 2720 if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd)) 2721 return; 2722 2723 trace_tcp_cwnd_reduction_tp(sk, newly_acked_sacked, newly_lost, flag); 2724 2725 tp->prr_delivered += newly_acked_sacked; 2726 if (delta < 0) { 2727 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered + 2728 tp->prior_cwnd - 1; 2729 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out; 2730 } else { 2731 sndcnt = max_t(int, tp->prr_delivered - tp->prr_out, 2732 newly_acked_sacked); 2733 if (flag & FLAG_SND_UNA_ADVANCED && !newly_lost) 2734 sndcnt++; 2735 sndcnt = min(delta, sndcnt); 2736 } 2737 /* Force a fast retransmit upon entering fast recovery */ 2738 sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1)); 2739 tcp_snd_cwnd_set(tp, tcp_packets_in_flight(tp) + sndcnt); 2740 } 2741 2742 static inline void tcp_end_cwnd_reduction(struct sock *sk) 2743 { 2744 struct tcp_sock *tp = tcp_sk(sk); 2745 2746 if (inet_csk(sk)->icsk_ca_ops->cong_control) 2747 return; 2748 2749 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */ 2750 if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH && 2751 (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR || tp->undo_marker)) { 2752 tcp_snd_cwnd_set(tp, tp->snd_ssthresh); 2753 tp->snd_cwnd_stamp = tcp_jiffies32; 2754 } 2755 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR); 2756 } 2757 2758 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */ 2759 void tcp_enter_cwr(struct sock *sk) 2760 { 2761 struct tcp_sock *tp = tcp_sk(sk); 2762 2763 tp->prior_ssthresh = 0; 2764 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) { 2765 tp->undo_marker = 0; 2766 tcp_init_cwnd_reduction(sk); 2767 tcp_set_ca_state(sk, TCP_CA_CWR); 2768 } 2769 } 2770 EXPORT_SYMBOL(tcp_enter_cwr); 2771 2772 static void tcp_try_keep_open(struct sock *sk) 2773 { 2774 struct tcp_sock *tp = tcp_sk(sk); 2775 int state = TCP_CA_Open; 2776 2777 if (tcp_left_out(tp) || tcp_any_retrans_done(sk)) 2778 state = TCP_CA_Disorder; 2779 2780 if (inet_csk(sk)->icsk_ca_state != state) { 2781 tcp_set_ca_state(sk, state); 2782 tp->high_seq = tp->snd_nxt; 2783 } 2784 } 2785 2786 static void tcp_try_to_open(struct sock *sk, int flag) 2787 { 2788 struct tcp_sock *tp = tcp_sk(sk); 2789 2790 tcp_verify_left_out(tp); 2791 2792 if (!tcp_any_retrans_done(sk)) 2793 tp->retrans_stamp = 0; 2794 2795 if (flag & FLAG_ECE) 2796 tcp_enter_cwr(sk); 2797 2798 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) { 2799 tcp_try_keep_open(sk); 2800 } 2801 } 2802 2803 static void tcp_mtup_probe_failed(struct sock *sk) 2804 { 2805 struct inet_connection_sock *icsk = inet_csk(sk); 2806 2807 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1; 2808 icsk->icsk_mtup.probe_size = 0; 2809 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPFAIL); 2810 } 2811 2812 static void tcp_mtup_probe_success(struct sock *sk) 2813 { 2814 struct tcp_sock *tp = tcp_sk(sk); 2815 struct inet_connection_sock *icsk = inet_csk(sk); 2816 u64 val; 2817 2818 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2819 2820 val = (u64)tcp_snd_cwnd(tp) * tcp_mss_to_mtu(sk, tp->mss_cache); 2821 do_div(val, icsk->icsk_mtup.probe_size); 2822 DEBUG_NET_WARN_ON_ONCE((u32)val != val); 2823 tcp_snd_cwnd_set(tp, max_t(u32, 1U, val)); 2824 2825 tp->snd_cwnd_cnt = 0; 2826 tp->snd_cwnd_stamp = tcp_jiffies32; 2827 tp->snd_ssthresh = tcp_current_ssthresh(sk); 2828 2829 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size; 2830 icsk->icsk_mtup.probe_size = 0; 2831 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 2832 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS); 2833 } 2834 2835 /* Sometimes we deduce that packets have been dropped due to reasons other than 2836 * congestion, like path MTU reductions or failed client TFO attempts. In these 2837 * cases we call this function to retransmit as many packets as cwnd allows, 2838 * without reducing cwnd. Given that retransmits will set retrans_stamp to a 2839 * non-zero value (and may do so in a later calling context due to TSQ), we 2840 * also enter CA_Loss so that we track when all retransmitted packets are ACKed 2841 * and clear retrans_stamp when that happens (to ensure later recurring RTOs 2842 * are using the correct retrans_stamp and don't declare ETIMEDOUT 2843 * prematurely). 2844 */ 2845 static void tcp_non_congestion_loss_retransmit(struct sock *sk) 2846 { 2847 const struct inet_connection_sock *icsk = inet_csk(sk); 2848 struct tcp_sock *tp = tcp_sk(sk); 2849 2850 if (icsk->icsk_ca_state != TCP_CA_Loss) { 2851 tp->high_seq = tp->snd_nxt; 2852 tp->snd_ssthresh = tcp_current_ssthresh(sk); 2853 tp->prior_ssthresh = 0; 2854 tp->undo_marker = 0; 2855 tcp_set_ca_state(sk, TCP_CA_Loss); 2856 } 2857 tcp_xmit_retransmit_queue(sk); 2858 } 2859 2860 /* Do a simple retransmit without using the backoff mechanisms in 2861 * tcp_timer. This is used for path mtu discovery. 2862 * The socket is already locked here. 2863 */ 2864 void tcp_simple_retransmit(struct sock *sk) 2865 { 2866 struct tcp_sock *tp = tcp_sk(sk); 2867 struct sk_buff *skb; 2868 int mss; 2869 2870 /* A fastopen SYN request is stored as two separate packets within 2871 * the retransmit queue, this is done by tcp_send_syn_data(). 2872 * As a result simply checking the MSS of the frames in the queue 2873 * will not work for the SYN packet. 2874 * 2875 * Us being here is an indication of a path MTU issue so we can 2876 * assume that the fastopen SYN was lost and just mark all the 2877 * frames in the retransmit queue as lost. We will use an MSS of 2878 * -1 to mark all frames as lost, otherwise compute the current MSS. 2879 */ 2880 if (tp->syn_data && sk->sk_state == TCP_SYN_SENT) 2881 mss = -1; 2882 else 2883 mss = tcp_current_mss(sk); 2884 2885 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) { 2886 if (tcp_skb_seglen(skb) > mss) 2887 tcp_mark_skb_lost(sk, skb); 2888 } 2889 2890 tcp_clear_retrans_hints_partial(tp); 2891 2892 if (!tp->lost_out) 2893 return; 2894 2895 if (tcp_is_reno(tp)) 2896 tcp_limit_reno_sacked(tp); 2897 2898 tcp_verify_left_out(tp); 2899 2900 /* Don't muck with the congestion window here. 2901 * Reason is that we do not increase amount of _data_ 2902 * in network, but units changed and effective 2903 * cwnd/ssthresh really reduced now. 2904 */ 2905 tcp_non_congestion_loss_retransmit(sk); 2906 } 2907 EXPORT_IPV6_MOD(tcp_simple_retransmit); 2908 2909 void tcp_enter_recovery(struct sock *sk, bool ece_ack) 2910 { 2911 struct tcp_sock *tp = tcp_sk(sk); 2912 int mib_idx; 2913 2914 /* Start the clock with our fast retransmit, for undo and ETIMEDOUT. */ 2915 tcp_retrans_stamp_cleanup(sk); 2916 2917 if (tcp_is_reno(tp)) 2918 mib_idx = LINUX_MIB_TCPRENORECOVERY; 2919 else 2920 mib_idx = LINUX_MIB_TCPSACKRECOVERY; 2921 2922 NET_INC_STATS(sock_net(sk), mib_idx); 2923 2924 tp->prior_ssthresh = 0; 2925 tcp_init_undo(tp); 2926 2927 if (!tcp_in_cwnd_reduction(sk)) { 2928 if (!ece_ack) 2929 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2930 tcp_init_cwnd_reduction(sk); 2931 } 2932 tcp_set_ca_state(sk, TCP_CA_Recovery); 2933 } 2934 2935 static void tcp_update_rto_time(struct tcp_sock *tp) 2936 { 2937 if (tp->rto_stamp) { 2938 tp->total_rto_time += tcp_time_stamp_ms(tp) - tp->rto_stamp; 2939 tp->rto_stamp = 0; 2940 } 2941 } 2942 2943 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are 2944 * recovered or spurious. Otherwise retransmits more on partial ACKs. 2945 */ 2946 static void tcp_process_loss(struct sock *sk, int flag, int num_dupack, 2947 int *rexmit) 2948 { 2949 struct tcp_sock *tp = tcp_sk(sk); 2950 bool recovered = !before(tp->snd_una, tp->high_seq); 2951 2952 if ((flag & FLAG_SND_UNA_ADVANCED || rcu_access_pointer(tp->fastopen_rsk)) && 2953 tcp_try_undo_loss(sk, false)) 2954 return; 2955 2956 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */ 2957 /* Step 3.b. A timeout is spurious if not all data are 2958 * lost, i.e., never-retransmitted data are (s)acked. 2959 */ 2960 if ((flag & FLAG_ORIG_SACK_ACKED) && 2961 tcp_try_undo_loss(sk, true)) 2962 return; 2963 2964 if (after(tp->snd_nxt, tp->high_seq)) { 2965 if (flag & FLAG_DATA_SACKED || num_dupack) 2966 tp->frto = 0; /* Step 3.a. loss was real */ 2967 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) { 2968 tp->high_seq = tp->snd_nxt; 2969 /* Step 2.b. Try send new data (but deferred until cwnd 2970 * is updated in tcp_ack()). Otherwise fall back to 2971 * the conventional recovery. 2972 */ 2973 if (!tcp_write_queue_empty(sk) && 2974 after(tcp_wnd_end(tp), tp->snd_nxt)) { 2975 *rexmit = REXMIT_NEW; 2976 return; 2977 } 2978 tp->frto = 0; 2979 } 2980 } 2981 2982 if (recovered) { 2983 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */ 2984 tcp_try_undo_recovery(sk); 2985 return; 2986 } 2987 if (tcp_is_reno(tp)) { 2988 /* A Reno DUPACK means new data in F-RTO step 2.b above are 2989 * delivered. Lower inflight to clock out (re)transmissions. 2990 */ 2991 if (after(tp->snd_nxt, tp->high_seq) && num_dupack) 2992 tcp_add_reno_sack(sk, num_dupack, flag & FLAG_ECE); 2993 else if (flag & FLAG_SND_UNA_ADVANCED) 2994 tcp_reset_reno_sack(tp); 2995 } 2996 *rexmit = REXMIT_LOST; 2997 } 2998 2999 static bool tcp_force_fast_retransmit(struct sock *sk) 3000 { 3001 struct tcp_sock *tp = tcp_sk(sk); 3002 3003 return after(tcp_highest_sack_seq(tp), 3004 tp->snd_una + tp->reordering * tp->mss_cache); 3005 } 3006 3007 /* Undo during fast recovery after partial ACK. */ 3008 static bool tcp_try_undo_partial(struct sock *sk, u32 prior_snd_una, 3009 bool *do_lost) 3010 { 3011 struct tcp_sock *tp = tcp_sk(sk); 3012 3013 if (tp->undo_marker && tcp_packet_delayed(tp)) { 3014 /* Plain luck! Hole if filled with delayed 3015 * packet, rather than with a retransmit. Check reordering. 3016 */ 3017 tcp_check_sack_reordering(sk, prior_snd_una, 1); 3018 3019 /* We are getting evidence that the reordering degree is higher 3020 * than we realized. If there are no retransmits out then we 3021 * can undo. Otherwise we clock out new packets but do not 3022 * mark more packets lost or retransmit more. 3023 */ 3024 if (tp->retrans_out) 3025 return true; 3026 3027 if (!tcp_any_retrans_done(sk)) 3028 tp->retrans_stamp = 0; 3029 3030 DBGUNDO(sk, "partial recovery"); 3031 tcp_undo_cwnd_reduction(sk, true); 3032 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO); 3033 tcp_try_keep_open(sk); 3034 } else { 3035 /* Partial ACK arrived. Force fast retransmit. */ 3036 *do_lost = tcp_force_fast_retransmit(sk); 3037 } 3038 return false; 3039 } 3040 3041 static void tcp_identify_packet_loss(struct sock *sk, int *ack_flag) 3042 { 3043 struct tcp_sock *tp = tcp_sk(sk); 3044 3045 if (tcp_rtx_queue_empty(sk)) 3046 return; 3047 3048 if (unlikely(tcp_is_reno(tp))) { 3049 tcp_newreno_mark_lost(sk, *ack_flag & FLAG_SND_UNA_ADVANCED); 3050 } else if (tcp_is_rack(sk)) { 3051 u32 prior_retrans = tp->retrans_out; 3052 3053 if (tcp_rack_mark_lost(sk)) 3054 *ack_flag &= ~FLAG_SET_XMIT_TIMER; 3055 if (prior_retrans > tp->retrans_out) 3056 *ack_flag |= FLAG_LOST_RETRANS; 3057 } 3058 } 3059 3060 /* Process an event, which can update packets-in-flight not trivially. 3061 * Main goal of this function is to calculate new estimate for left_out, 3062 * taking into account both packets sitting in receiver's buffer and 3063 * packets lost by network. 3064 * 3065 * Besides that it updates the congestion state when packet loss or ECN 3066 * is detected. But it does not reduce the cwnd, it is done by the 3067 * congestion control later. 3068 * 3069 * It does _not_ decide what to send, it is made in function 3070 * tcp_xmit_retransmit_queue(). 3071 */ 3072 static void tcp_fastretrans_alert(struct sock *sk, const u32 prior_snd_una, 3073 int num_dupack, int *ack_flag, int *rexmit) 3074 { 3075 struct inet_connection_sock *icsk = inet_csk(sk); 3076 struct tcp_sock *tp = tcp_sk(sk); 3077 int fast_rexmit = 0, flag = *ack_flag; 3078 bool ece_ack = flag & FLAG_ECE; 3079 bool do_lost = num_dupack || ((flag & FLAG_DATA_SACKED) && 3080 tcp_force_fast_retransmit(sk)); 3081 3082 if (!tp->packets_out && tp->sacked_out) 3083 tp->sacked_out = 0; 3084 3085 /* Now state machine starts. 3086 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */ 3087 if (ece_ack) 3088 tp->prior_ssthresh = 0; 3089 3090 /* B. In all the states check for reneging SACKs. */ 3091 if (tcp_check_sack_reneging(sk, ack_flag)) 3092 return; 3093 3094 /* C. Check consistency of the current state. */ 3095 tcp_verify_left_out(tp); 3096 3097 /* D. Check state exit conditions. State can be terminated 3098 * when high_seq is ACKed. */ 3099 if (icsk->icsk_ca_state == TCP_CA_Open) { 3100 WARN_ON(tp->retrans_out != 0 && !tp->syn_data); 3101 tp->retrans_stamp = 0; 3102 } else if (!before(tp->snd_una, tp->high_seq)) { 3103 switch (icsk->icsk_ca_state) { 3104 case TCP_CA_CWR: 3105 /* CWR is to be held something *above* high_seq 3106 * is ACKed for CWR bit to reach receiver. */ 3107 if (tp->snd_una != tp->high_seq) { 3108 tcp_end_cwnd_reduction(sk); 3109 tcp_set_ca_state(sk, TCP_CA_Open); 3110 } 3111 break; 3112 3113 case TCP_CA_Recovery: 3114 if (tcp_is_reno(tp)) 3115 tcp_reset_reno_sack(tp); 3116 if (tcp_try_undo_recovery(sk)) 3117 return; 3118 tcp_end_cwnd_reduction(sk); 3119 break; 3120 } 3121 } 3122 3123 /* E. Process state. */ 3124 switch (icsk->icsk_ca_state) { 3125 case TCP_CA_Recovery: 3126 if (!(flag & FLAG_SND_UNA_ADVANCED)) { 3127 if (tcp_is_reno(tp)) 3128 tcp_add_reno_sack(sk, num_dupack, ece_ack); 3129 } else if (tcp_try_undo_partial(sk, prior_snd_una, &do_lost)) 3130 return; 3131 3132 if (tcp_try_undo_dsack(sk)) 3133 tcp_try_to_open(sk, flag); 3134 3135 tcp_identify_packet_loss(sk, ack_flag); 3136 if (icsk->icsk_ca_state != TCP_CA_Recovery) { 3137 if (!tcp_time_to_recover(sk, flag)) 3138 return; 3139 /* Undo reverts the recovery state. If loss is evident, 3140 * starts a new recovery (e.g. reordering then loss); 3141 */ 3142 tcp_enter_recovery(sk, ece_ack); 3143 } 3144 break; 3145 case TCP_CA_Loss: 3146 tcp_process_loss(sk, flag, num_dupack, rexmit); 3147 if (icsk->icsk_ca_state != TCP_CA_Loss) 3148 tcp_update_rto_time(tp); 3149 tcp_identify_packet_loss(sk, ack_flag); 3150 if (!(icsk->icsk_ca_state == TCP_CA_Open || 3151 (*ack_flag & FLAG_LOST_RETRANS))) 3152 return; 3153 /* Change state if cwnd is undone or retransmits are lost */ 3154 fallthrough; 3155 default: 3156 if (tcp_is_reno(tp)) { 3157 if (flag & FLAG_SND_UNA_ADVANCED) 3158 tcp_reset_reno_sack(tp); 3159 tcp_add_reno_sack(sk, num_dupack, ece_ack); 3160 } 3161 3162 if (icsk->icsk_ca_state <= TCP_CA_Disorder) 3163 tcp_try_undo_dsack(sk); 3164 3165 tcp_identify_packet_loss(sk, ack_flag); 3166 if (!tcp_time_to_recover(sk, flag)) { 3167 tcp_try_to_open(sk, flag); 3168 return; 3169 } 3170 3171 /* MTU probe failure: don't reduce cwnd */ 3172 if (icsk->icsk_ca_state < TCP_CA_CWR && 3173 icsk->icsk_mtup.probe_size && 3174 tp->snd_una == tp->mtu_probe.probe_seq_start) { 3175 tcp_mtup_probe_failed(sk); 3176 /* Restores the reduction we did in tcp_mtup_probe() */ 3177 tcp_snd_cwnd_set(tp, tcp_snd_cwnd(tp) + 1); 3178 tcp_simple_retransmit(sk); 3179 return; 3180 } 3181 3182 /* Otherwise enter Recovery state */ 3183 tcp_enter_recovery(sk, ece_ack); 3184 fast_rexmit = 1; 3185 } 3186 3187 if (!tcp_is_rack(sk) && do_lost) 3188 tcp_update_scoreboard(sk, fast_rexmit); 3189 *rexmit = REXMIT_LOST; 3190 } 3191 3192 static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us, const int flag) 3193 { 3194 u32 wlen = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_rtt_wlen) * HZ; 3195 struct tcp_sock *tp = tcp_sk(sk); 3196 3197 if ((flag & FLAG_ACK_MAYBE_DELAYED) && rtt_us > tcp_min_rtt(tp)) { 3198 /* If the remote keeps returning delayed ACKs, eventually 3199 * the min filter would pick it up and overestimate the 3200 * prop. delay when it expires. Skip suspected delayed ACKs. 3201 */ 3202 return; 3203 } 3204 minmax_running_min(&tp->rtt_min, wlen, tcp_jiffies32, 3205 rtt_us ? : jiffies_to_usecs(1)); 3206 } 3207 3208 static bool tcp_ack_update_rtt(struct sock *sk, const int flag, 3209 long seq_rtt_us, long sack_rtt_us, 3210 long ca_rtt_us, struct rate_sample *rs) 3211 { 3212 const struct tcp_sock *tp = tcp_sk(sk); 3213 3214 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because 3215 * broken middle-boxes or peers may corrupt TS-ECR fields. But 3216 * Karn's algorithm forbids taking RTT if some retransmitted data 3217 * is acked (RFC6298). 3218 */ 3219 if (seq_rtt_us < 0) 3220 seq_rtt_us = sack_rtt_us; 3221 3222 /* RTTM Rule: A TSecr value received in a segment is used to 3223 * update the averaged RTT measurement only if the segment 3224 * acknowledges some new data, i.e., only if it advances the 3225 * left edge of the send window. 3226 * See draft-ietf-tcplw-high-performance-00, section 3.3. 3227 */ 3228 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && 3229 tp->rx_opt.rcv_tsecr && flag & FLAG_ACKED) 3230 seq_rtt_us = ca_rtt_us = tcp_rtt_tsopt_us(tp); 3231 3232 rs->rtt_us = ca_rtt_us; /* RTT of last (S)ACKed packet (or -1) */ 3233 if (seq_rtt_us < 0) 3234 return false; 3235 3236 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is 3237 * always taken together with ACK, SACK, or TS-opts. Any negative 3238 * values will be skipped with the seq_rtt_us < 0 check above. 3239 */ 3240 tcp_update_rtt_min(sk, ca_rtt_us, flag); 3241 tcp_rtt_estimator(sk, seq_rtt_us); 3242 tcp_set_rto(sk); 3243 3244 /* RFC6298: only reset backoff on valid RTT measurement. */ 3245 inet_csk(sk)->icsk_backoff = 0; 3246 return true; 3247 } 3248 3249 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */ 3250 void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req) 3251 { 3252 struct rate_sample rs; 3253 long rtt_us = -1L; 3254 3255 if (req && !req->num_retrans && tcp_rsk(req)->snt_synack) 3256 rtt_us = tcp_stamp_us_delta(tcp_clock_us(), tcp_rsk(req)->snt_synack); 3257 3258 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us, &rs); 3259 } 3260 3261 3262 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked) 3263 { 3264 const struct inet_connection_sock *icsk = inet_csk(sk); 3265 3266 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked); 3267 tcp_sk(sk)->snd_cwnd_stamp = tcp_jiffies32; 3268 } 3269 3270 /* Restart timer after forward progress on connection. 3271 * RFC2988 recommends to restart timer to now+rto. 3272 */ 3273 void tcp_rearm_rto(struct sock *sk) 3274 { 3275 const struct inet_connection_sock *icsk = inet_csk(sk); 3276 struct tcp_sock *tp = tcp_sk(sk); 3277 3278 /* If the retrans timer is currently being used by Fast Open 3279 * for SYN-ACK retrans purpose, stay put. 3280 */ 3281 if (rcu_access_pointer(tp->fastopen_rsk)) 3282 return; 3283 3284 if (!tp->packets_out) { 3285 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS); 3286 } else { 3287 u32 rto = inet_csk(sk)->icsk_rto; 3288 /* Offset the time elapsed after installing regular RTO */ 3289 if (icsk->icsk_pending == ICSK_TIME_REO_TIMEOUT || 3290 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) { 3291 s64 delta_us = tcp_rto_delta_us(sk); 3292 /* delta_us may not be positive if the socket is locked 3293 * when the retrans timer fires and is rescheduled. 3294 */ 3295 rto = usecs_to_jiffies(max_t(int, delta_us, 1)); 3296 } 3297 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto, true); 3298 } 3299 } 3300 3301 /* Try to schedule a loss probe; if that doesn't work, then schedule an RTO. */ 3302 static void tcp_set_xmit_timer(struct sock *sk) 3303 { 3304 if (!tcp_schedule_loss_probe(sk, true)) 3305 tcp_rearm_rto(sk); 3306 } 3307 3308 /* If we get here, the whole TSO packet has not been acked. */ 3309 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb) 3310 { 3311 struct tcp_sock *tp = tcp_sk(sk); 3312 u32 packets_acked; 3313 3314 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)); 3315 3316 packets_acked = tcp_skb_pcount(skb); 3317 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 3318 return 0; 3319 packets_acked -= tcp_skb_pcount(skb); 3320 3321 if (packets_acked) { 3322 BUG_ON(tcp_skb_pcount(skb) == 0); 3323 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)); 3324 } 3325 3326 return packets_acked; 3327 } 3328 3329 static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb, 3330 const struct sk_buff *ack_skb, u32 prior_snd_una) 3331 { 3332 const struct skb_shared_info *shinfo; 3333 3334 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */ 3335 if (likely(!TCP_SKB_CB(skb)->txstamp_ack)) 3336 return; 3337 3338 shinfo = skb_shinfo(skb); 3339 if (!before(shinfo->tskey, prior_snd_una) && 3340 before(shinfo->tskey, tcp_sk(sk)->snd_una)) { 3341 tcp_skb_tsorted_save(skb) { 3342 __skb_tstamp_tx(skb, ack_skb, NULL, sk, SCM_TSTAMP_ACK); 3343 } tcp_skb_tsorted_restore(skb); 3344 } 3345 } 3346 3347 /* Remove acknowledged frames from the retransmission queue. If our packet 3348 * is before the ack sequence we can discard it as it's confirmed to have 3349 * arrived at the other end. 3350 */ 3351 static int tcp_clean_rtx_queue(struct sock *sk, const struct sk_buff *ack_skb, 3352 u32 prior_fack, u32 prior_snd_una, 3353 struct tcp_sacktag_state *sack, bool ece_ack) 3354 { 3355 const struct inet_connection_sock *icsk = inet_csk(sk); 3356 u64 first_ackt, last_ackt; 3357 struct tcp_sock *tp = tcp_sk(sk); 3358 u32 prior_sacked = tp->sacked_out; 3359 u32 reord = tp->snd_nxt; /* lowest acked un-retx un-sacked seq */ 3360 struct sk_buff *skb, *next; 3361 bool fully_acked = true; 3362 long sack_rtt_us = -1L; 3363 long seq_rtt_us = -1L; 3364 long ca_rtt_us = -1L; 3365 u32 pkts_acked = 0; 3366 bool rtt_update; 3367 int flag = 0; 3368 3369 first_ackt = 0; 3370 3371 for (skb = skb_rb_first(&sk->tcp_rtx_queue); skb; skb = next) { 3372 struct tcp_skb_cb *scb = TCP_SKB_CB(skb); 3373 const u32 start_seq = scb->seq; 3374 u8 sacked = scb->sacked; 3375 u32 acked_pcount; 3376 3377 /* Determine how many packets and what bytes were acked, tso and else */ 3378 if (after(scb->end_seq, tp->snd_una)) { 3379 if (tcp_skb_pcount(skb) == 1 || 3380 !after(tp->snd_una, scb->seq)) 3381 break; 3382 3383 acked_pcount = tcp_tso_acked(sk, skb); 3384 if (!acked_pcount) 3385 break; 3386 fully_acked = false; 3387 } else { 3388 acked_pcount = tcp_skb_pcount(skb); 3389 } 3390 3391 if (unlikely(sacked & TCPCB_RETRANS)) { 3392 if (sacked & TCPCB_SACKED_RETRANS) 3393 tp->retrans_out -= acked_pcount; 3394 flag |= FLAG_RETRANS_DATA_ACKED; 3395 } else if (!(sacked & TCPCB_SACKED_ACKED)) { 3396 last_ackt = tcp_skb_timestamp_us(skb); 3397 WARN_ON_ONCE(last_ackt == 0); 3398 if (!first_ackt) 3399 first_ackt = last_ackt; 3400 3401 if (before(start_seq, reord)) 3402 reord = start_seq; 3403 if (!after(scb->end_seq, tp->high_seq)) 3404 flag |= FLAG_ORIG_SACK_ACKED; 3405 } 3406 3407 if (sacked & TCPCB_SACKED_ACKED) { 3408 tp->sacked_out -= acked_pcount; 3409 } else if (tcp_is_sack(tp)) { 3410 tcp_count_delivered(tp, acked_pcount, ece_ack); 3411 if (!tcp_skb_spurious_retrans(tp, skb)) 3412 tcp_rack_advance(tp, sacked, scb->end_seq, 3413 tcp_skb_timestamp_us(skb)); 3414 } 3415 if (sacked & TCPCB_LOST) 3416 tp->lost_out -= acked_pcount; 3417 3418 tp->packets_out -= acked_pcount; 3419 pkts_acked += acked_pcount; 3420 tcp_rate_skb_delivered(sk, skb, sack->rate); 3421 3422 /* Initial outgoing SYN's get put onto the write_queue 3423 * just like anything else we transmit. It is not 3424 * true data, and if we misinform our callers that 3425 * this ACK acks real data, we will erroneously exit 3426 * connection startup slow start one packet too 3427 * quickly. This is severely frowned upon behavior. 3428 */ 3429 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) { 3430 flag |= FLAG_DATA_ACKED; 3431 } else { 3432 flag |= FLAG_SYN_ACKED; 3433 tp->retrans_stamp = 0; 3434 } 3435 3436 if (!fully_acked) 3437 break; 3438 3439 tcp_ack_tstamp(sk, skb, ack_skb, prior_snd_una); 3440 3441 next = skb_rb_next(skb); 3442 if (unlikely(skb == tp->retransmit_skb_hint)) 3443 tp->retransmit_skb_hint = NULL; 3444 if (unlikely(skb == tp->lost_skb_hint)) 3445 tp->lost_skb_hint = NULL; 3446 tcp_highest_sack_replace(sk, skb, next); 3447 tcp_rtx_queue_unlink_and_free(skb, sk); 3448 } 3449 3450 if (!skb) 3451 tcp_chrono_stop(sk, TCP_CHRONO_BUSY); 3452 3453 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una))) 3454 tp->snd_up = tp->snd_una; 3455 3456 if (skb) { 3457 tcp_ack_tstamp(sk, skb, ack_skb, prior_snd_una); 3458 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 3459 flag |= FLAG_SACK_RENEGING; 3460 } 3461 3462 if (likely(first_ackt) && !(flag & FLAG_RETRANS_DATA_ACKED)) { 3463 seq_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, first_ackt); 3464 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, last_ackt); 3465 3466 if (pkts_acked == 1 && fully_acked && !prior_sacked && 3467 (tp->snd_una - prior_snd_una) < tp->mss_cache && 3468 sack->rate->prior_delivered + 1 == tp->delivered && 3469 !(flag & (FLAG_CA_ALERT | FLAG_SYN_ACKED))) { 3470 /* Conservatively mark a delayed ACK. It's typically 3471 * from a lone runt packet over the round trip to 3472 * a receiver w/o out-of-order or CE events. 3473 */ 3474 flag |= FLAG_ACK_MAYBE_DELAYED; 3475 } 3476 } 3477 if (sack->first_sackt) { 3478 sack_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->first_sackt); 3479 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->last_sackt); 3480 } 3481 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us, 3482 ca_rtt_us, sack->rate); 3483 3484 if (flag & FLAG_ACKED) { 3485 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */ 3486 if (unlikely(icsk->icsk_mtup.probe_size && 3487 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) { 3488 tcp_mtup_probe_success(sk); 3489 } 3490 3491 if (tcp_is_reno(tp)) { 3492 tcp_remove_reno_sacks(sk, pkts_acked, ece_ack); 3493 3494 /* If any of the cumulatively ACKed segments was 3495 * retransmitted, non-SACK case cannot confirm that 3496 * progress was due to original transmission due to 3497 * lack of TCPCB_SACKED_ACKED bits even if some of 3498 * the packets may have been never retransmitted. 3499 */ 3500 if (flag & FLAG_RETRANS_DATA_ACKED) 3501 flag &= ~FLAG_ORIG_SACK_ACKED; 3502 } else { 3503 int delta; 3504 3505 /* Non-retransmitted hole got filled? That's reordering */ 3506 if (before(reord, prior_fack)) 3507 tcp_check_sack_reordering(sk, reord, 0); 3508 3509 delta = prior_sacked - tp->sacked_out; 3510 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta); 3511 } 3512 } else if (skb && rtt_update && sack_rtt_us >= 0 && 3513 sack_rtt_us > tcp_stamp_us_delta(tp->tcp_mstamp, 3514 tcp_skb_timestamp_us(skb))) { 3515 /* Do not re-arm RTO if the sack RTT is measured from data sent 3516 * after when the head was last (re)transmitted. Otherwise the 3517 * timeout may continue to extend in loss recovery. 3518 */ 3519 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */ 3520 } 3521 3522 if (icsk->icsk_ca_ops->pkts_acked) { 3523 struct ack_sample sample = { .pkts_acked = pkts_acked, 3524 .rtt_us = sack->rate->rtt_us }; 3525 3526 sample.in_flight = tp->mss_cache * 3527 (tp->delivered - sack->rate->prior_delivered); 3528 icsk->icsk_ca_ops->pkts_acked(sk, &sample); 3529 } 3530 3531 #if FASTRETRANS_DEBUG > 0 3532 WARN_ON((int)tp->sacked_out < 0); 3533 WARN_ON((int)tp->lost_out < 0); 3534 WARN_ON((int)tp->retrans_out < 0); 3535 if (!tp->packets_out && tcp_is_sack(tp)) { 3536 icsk = inet_csk(sk); 3537 if (tp->lost_out) { 3538 pr_debug("Leak l=%u %d\n", 3539 tp->lost_out, icsk->icsk_ca_state); 3540 tp->lost_out = 0; 3541 } 3542 if (tp->sacked_out) { 3543 pr_debug("Leak s=%u %d\n", 3544 tp->sacked_out, icsk->icsk_ca_state); 3545 tp->sacked_out = 0; 3546 } 3547 if (tp->retrans_out) { 3548 pr_debug("Leak r=%u %d\n", 3549 tp->retrans_out, icsk->icsk_ca_state); 3550 tp->retrans_out = 0; 3551 } 3552 } 3553 #endif 3554 return flag; 3555 } 3556 3557 static void tcp_ack_probe(struct sock *sk) 3558 { 3559 struct inet_connection_sock *icsk = inet_csk(sk); 3560 struct sk_buff *head = tcp_send_head(sk); 3561 const struct tcp_sock *tp = tcp_sk(sk); 3562 3563 /* Was it a usable window open? */ 3564 if (!head) 3565 return; 3566 if (!after(TCP_SKB_CB(head)->end_seq, tcp_wnd_end(tp))) { 3567 icsk->icsk_backoff = 0; 3568 icsk->icsk_probes_tstamp = 0; 3569 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0); 3570 /* Socket must be waked up by subsequent tcp_data_snd_check(). 3571 * This function is not for random using! 3572 */ 3573 } else { 3574 unsigned long when = tcp_probe0_when(sk, tcp_rto_max(sk)); 3575 3576 when = tcp_clamp_probe0_to_user_timeout(sk, when); 3577 tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, when, true); 3578 } 3579 } 3580 3581 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag) 3582 { 3583 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) || 3584 inet_csk(sk)->icsk_ca_state != TCP_CA_Open; 3585 } 3586 3587 /* Decide wheather to run the increase function of congestion control. */ 3588 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag) 3589 { 3590 /* If reordering is high then always grow cwnd whenever data is 3591 * delivered regardless of its ordering. Otherwise stay conservative 3592 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/ 3593 * new SACK or ECE mark may first advance cwnd here and later reduce 3594 * cwnd in tcp_fastretrans_alert() based on more states. 3595 */ 3596 if (tcp_sk(sk)->reordering > 3597 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_reordering)) 3598 return flag & FLAG_FORWARD_PROGRESS; 3599 3600 return flag & FLAG_DATA_ACKED; 3601 } 3602 3603 /* The "ultimate" congestion control function that aims to replace the rigid 3604 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction). 3605 * It's called toward the end of processing an ACK with precise rate 3606 * information. All transmission or retransmission are delayed afterwards. 3607 */ 3608 static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked, 3609 int flag, const struct rate_sample *rs) 3610 { 3611 const struct inet_connection_sock *icsk = inet_csk(sk); 3612 3613 if (icsk->icsk_ca_ops->cong_control) { 3614 icsk->icsk_ca_ops->cong_control(sk, ack, flag, rs); 3615 return; 3616 } 3617 3618 if (tcp_in_cwnd_reduction(sk)) { 3619 /* Reduce cwnd if state mandates */ 3620 tcp_cwnd_reduction(sk, acked_sacked, rs->losses, flag); 3621 } else if (tcp_may_raise_cwnd(sk, flag)) { 3622 /* Advance cwnd if state allows */ 3623 tcp_cong_avoid(sk, ack, acked_sacked); 3624 } 3625 tcp_update_pacing_rate(sk); 3626 } 3627 3628 /* Check that window update is acceptable. 3629 * The function assumes that snd_una<=ack<=snd_next. 3630 */ 3631 static inline bool tcp_may_update_window(const struct tcp_sock *tp, 3632 const u32 ack, const u32 ack_seq, 3633 const u32 nwin) 3634 { 3635 return after(ack, tp->snd_una) || 3636 after(ack_seq, tp->snd_wl1) || 3637 (ack_seq == tp->snd_wl1 && (nwin > tp->snd_wnd || !nwin)); 3638 } 3639 3640 static void tcp_snd_sne_update(struct tcp_sock *tp, u32 ack) 3641 { 3642 #ifdef CONFIG_TCP_AO 3643 struct tcp_ao_info *ao; 3644 3645 if (!static_branch_unlikely(&tcp_ao_needed.key)) 3646 return; 3647 3648 ao = rcu_dereference_protected(tp->ao_info, 3649 lockdep_sock_is_held((struct sock *)tp)); 3650 if (ao && ack < tp->snd_una) { 3651 ao->snd_sne++; 3652 trace_tcp_ao_snd_sne_update((struct sock *)tp, ao->snd_sne); 3653 } 3654 #endif 3655 } 3656 3657 /* If we update tp->snd_una, also update tp->bytes_acked */ 3658 static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack) 3659 { 3660 u32 delta = ack - tp->snd_una; 3661 3662 sock_owned_by_me((struct sock *)tp); 3663 tp->bytes_acked += delta; 3664 tcp_snd_sne_update(tp, ack); 3665 tp->snd_una = ack; 3666 } 3667 3668 static void tcp_rcv_sne_update(struct tcp_sock *tp, u32 seq) 3669 { 3670 #ifdef CONFIG_TCP_AO 3671 struct tcp_ao_info *ao; 3672 3673 if (!static_branch_unlikely(&tcp_ao_needed.key)) 3674 return; 3675 3676 ao = rcu_dereference_protected(tp->ao_info, 3677 lockdep_sock_is_held((struct sock *)tp)); 3678 if (ao && seq < tp->rcv_nxt) { 3679 ao->rcv_sne++; 3680 trace_tcp_ao_rcv_sne_update((struct sock *)tp, ao->rcv_sne); 3681 } 3682 #endif 3683 } 3684 3685 /* If we update tp->rcv_nxt, also update tp->bytes_received */ 3686 static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq) 3687 { 3688 u32 delta = seq - tp->rcv_nxt; 3689 3690 sock_owned_by_me((struct sock *)tp); 3691 tp->bytes_received += delta; 3692 tcp_rcv_sne_update(tp, seq); 3693 WRITE_ONCE(tp->rcv_nxt, seq); 3694 } 3695 3696 /* Update our send window. 3697 * 3698 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2 3699 * and in FreeBSD. NetBSD's one is even worse.) is wrong. 3700 */ 3701 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack, 3702 u32 ack_seq) 3703 { 3704 struct tcp_sock *tp = tcp_sk(sk); 3705 int flag = 0; 3706 u32 nwin = ntohs(tcp_hdr(skb)->window); 3707 3708 if (likely(!tcp_hdr(skb)->syn)) 3709 nwin <<= tp->rx_opt.snd_wscale; 3710 3711 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) { 3712 flag |= FLAG_WIN_UPDATE; 3713 tcp_update_wl(tp, ack_seq); 3714 3715 if (tp->snd_wnd != nwin) { 3716 tp->snd_wnd = nwin; 3717 3718 /* Note, it is the only place, where 3719 * fast path is recovered for sending TCP. 3720 */ 3721 tp->pred_flags = 0; 3722 tcp_fast_path_check(sk); 3723 3724 if (!tcp_write_queue_empty(sk)) 3725 tcp_slow_start_after_idle_check(sk); 3726 3727 if (nwin > tp->max_window) { 3728 tp->max_window = nwin; 3729 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie); 3730 } 3731 } 3732 } 3733 3734 tcp_snd_una_update(tp, ack); 3735 3736 return flag; 3737 } 3738 3739 static bool __tcp_oow_rate_limited(struct net *net, int mib_idx, 3740 u32 *last_oow_ack_time) 3741 { 3742 /* Paired with the WRITE_ONCE() in this function. */ 3743 u32 val = READ_ONCE(*last_oow_ack_time); 3744 3745 if (val) { 3746 s32 elapsed = (s32)(tcp_jiffies32 - val); 3747 3748 if (0 <= elapsed && 3749 elapsed < READ_ONCE(net->ipv4.sysctl_tcp_invalid_ratelimit)) { 3750 NET_INC_STATS(net, mib_idx); 3751 return true; /* rate-limited: don't send yet! */ 3752 } 3753 } 3754 3755 /* Paired with the prior READ_ONCE() and with itself, 3756 * as we might be lockless. 3757 */ 3758 WRITE_ONCE(*last_oow_ack_time, tcp_jiffies32); 3759 3760 return false; /* not rate-limited: go ahead, send dupack now! */ 3761 } 3762 3763 /* Return true if we're currently rate-limiting out-of-window ACKs and 3764 * thus shouldn't send a dupack right now. We rate-limit dupacks in 3765 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS 3766 * attacks that send repeated SYNs or ACKs for the same connection. To 3767 * do this, we do not send a duplicate SYNACK or ACK if the remote 3768 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate. 3769 */ 3770 bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb, 3771 int mib_idx, u32 *last_oow_ack_time) 3772 { 3773 /* Data packets without SYNs are not likely part of an ACK loop. */ 3774 if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) && 3775 !tcp_hdr(skb)->syn) 3776 return false; 3777 3778 return __tcp_oow_rate_limited(net, mib_idx, last_oow_ack_time); 3779 } 3780 3781 /* RFC 5961 7 [ACK Throttling] */ 3782 static void tcp_send_challenge_ack(struct sock *sk) 3783 { 3784 struct tcp_sock *tp = tcp_sk(sk); 3785 struct net *net = sock_net(sk); 3786 u32 count, now, ack_limit; 3787 3788 /* First check our per-socket dupack rate limit. */ 3789 if (__tcp_oow_rate_limited(net, 3790 LINUX_MIB_TCPACKSKIPPEDCHALLENGE, 3791 &tp->last_oow_ack_time)) 3792 return; 3793 3794 ack_limit = READ_ONCE(net->ipv4.sysctl_tcp_challenge_ack_limit); 3795 if (ack_limit == INT_MAX) 3796 goto send_ack; 3797 3798 /* Then check host-wide RFC 5961 rate limit. */ 3799 now = jiffies / HZ; 3800 if (now != READ_ONCE(net->ipv4.tcp_challenge_timestamp)) { 3801 u32 half = (ack_limit + 1) >> 1; 3802 3803 WRITE_ONCE(net->ipv4.tcp_challenge_timestamp, now); 3804 WRITE_ONCE(net->ipv4.tcp_challenge_count, 3805 get_random_u32_inclusive(half, ack_limit + half - 1)); 3806 } 3807 count = READ_ONCE(net->ipv4.tcp_challenge_count); 3808 if (count > 0) { 3809 WRITE_ONCE(net->ipv4.tcp_challenge_count, count - 1); 3810 send_ack: 3811 NET_INC_STATS(net, LINUX_MIB_TCPCHALLENGEACK); 3812 tcp_send_ack(sk); 3813 } 3814 } 3815 3816 static void tcp_store_ts_recent(struct tcp_sock *tp) 3817 { 3818 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval; 3819 tp->rx_opt.ts_recent_stamp = ktime_get_seconds(); 3820 } 3821 3822 static int __tcp_replace_ts_recent(struct tcp_sock *tp, s32 tstamp_delta) 3823 { 3824 tcp_store_ts_recent(tp); 3825 return tstamp_delta > 0 ? FLAG_TS_PROGRESS : 0; 3826 } 3827 3828 static int tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq) 3829 { 3830 s32 delta; 3831 3832 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) { 3833 /* PAWS bug workaround wrt. ACK frames, the PAWS discard 3834 * extra check below makes sure this can only happen 3835 * for pure ACK frames. -DaveM 3836 * 3837 * Not only, also it occurs for expired timestamps. 3838 */ 3839 3840 if (tcp_paws_check(&tp->rx_opt, 0)) { 3841 delta = tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent; 3842 return __tcp_replace_ts_recent(tp, delta); 3843 } 3844 } 3845 3846 return 0; 3847 } 3848 3849 /* This routine deals with acks during a TLP episode and ends an episode by 3850 * resetting tlp_high_seq. Ref: TLP algorithm in draft-ietf-tcpm-rack 3851 */ 3852 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag) 3853 { 3854 struct tcp_sock *tp = tcp_sk(sk); 3855 3856 if (before(ack, tp->tlp_high_seq)) 3857 return; 3858 3859 if (!tp->tlp_retrans) { 3860 /* TLP of new data has been acknowledged */ 3861 tp->tlp_high_seq = 0; 3862 } else if (flag & FLAG_DSACK_TLP) { 3863 /* This DSACK means original and TLP probe arrived; no loss */ 3864 tp->tlp_high_seq = 0; 3865 } else if (after(ack, tp->tlp_high_seq)) { 3866 /* ACK advances: there was a loss, so reduce cwnd. Reset 3867 * tlp_high_seq in tcp_init_cwnd_reduction() 3868 */ 3869 tcp_init_cwnd_reduction(sk); 3870 tcp_set_ca_state(sk, TCP_CA_CWR); 3871 tcp_end_cwnd_reduction(sk); 3872 tcp_try_keep_open(sk); 3873 NET_INC_STATS(sock_net(sk), 3874 LINUX_MIB_TCPLOSSPROBERECOVERY); 3875 } else if (!(flag & (FLAG_SND_UNA_ADVANCED | 3876 FLAG_NOT_DUP | FLAG_DATA_SACKED))) { 3877 /* Pure dupack: original and TLP probe arrived; no loss */ 3878 tp->tlp_high_seq = 0; 3879 } 3880 } 3881 3882 static void tcp_in_ack_event(struct sock *sk, int flag) 3883 { 3884 const struct inet_connection_sock *icsk = inet_csk(sk); 3885 3886 if (icsk->icsk_ca_ops->in_ack_event) { 3887 u32 ack_ev_flags = 0; 3888 3889 if (flag & FLAG_WIN_UPDATE) 3890 ack_ev_flags |= CA_ACK_WIN_UPDATE; 3891 if (flag & FLAG_SLOWPATH) { 3892 ack_ev_flags |= CA_ACK_SLOWPATH; 3893 if (flag & FLAG_ECE) 3894 ack_ev_flags |= CA_ACK_ECE; 3895 } 3896 3897 icsk->icsk_ca_ops->in_ack_event(sk, ack_ev_flags); 3898 } 3899 } 3900 3901 /* Congestion control has updated the cwnd already. So if we're in 3902 * loss recovery then now we do any new sends (for FRTO) or 3903 * retransmits (for CA_Loss or CA_recovery) that make sense. 3904 */ 3905 static void tcp_xmit_recovery(struct sock *sk, int rexmit) 3906 { 3907 struct tcp_sock *tp = tcp_sk(sk); 3908 3909 if (rexmit == REXMIT_NONE || sk->sk_state == TCP_SYN_SENT) 3910 return; 3911 3912 if (unlikely(rexmit == REXMIT_NEW)) { 3913 __tcp_push_pending_frames(sk, tcp_current_mss(sk), 3914 TCP_NAGLE_OFF); 3915 if (after(tp->snd_nxt, tp->high_seq)) 3916 return; 3917 tp->frto = 0; 3918 } 3919 tcp_xmit_retransmit_queue(sk); 3920 } 3921 3922 /* Returns the number of packets newly acked or sacked by the current ACK */ 3923 static u32 tcp_newly_delivered(struct sock *sk, u32 prior_delivered, int flag) 3924 { 3925 const struct net *net = sock_net(sk); 3926 struct tcp_sock *tp = tcp_sk(sk); 3927 u32 delivered; 3928 3929 delivered = tp->delivered - prior_delivered; 3930 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVERED, delivered); 3931 if (flag & FLAG_ECE) 3932 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVEREDCE, delivered); 3933 3934 return delivered; 3935 } 3936 3937 /* This routine deals with incoming acks, but not outgoing ones. */ 3938 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag) 3939 { 3940 struct inet_connection_sock *icsk = inet_csk(sk); 3941 struct tcp_sock *tp = tcp_sk(sk); 3942 struct tcp_sacktag_state sack_state; 3943 struct rate_sample rs = { .prior_delivered = 0 }; 3944 u32 prior_snd_una = tp->snd_una; 3945 bool is_sack_reneg = tp->is_sack_reneg; 3946 u32 ack_seq = TCP_SKB_CB(skb)->seq; 3947 u32 ack = TCP_SKB_CB(skb)->ack_seq; 3948 int num_dupack = 0; 3949 int prior_packets = tp->packets_out; 3950 u32 delivered = tp->delivered; 3951 u32 lost = tp->lost; 3952 int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */ 3953 u32 prior_fack; 3954 3955 sack_state.first_sackt = 0; 3956 sack_state.rate = &rs; 3957 sack_state.sack_delivered = 0; 3958 3959 /* We very likely will need to access rtx queue. */ 3960 prefetch(sk->tcp_rtx_queue.rb_node); 3961 3962 /* If the ack is older than previous acks 3963 * then we can probably ignore it. 3964 */ 3965 if (before(ack, prior_snd_una)) { 3966 u32 max_window; 3967 3968 /* do not accept ACK for bytes we never sent. */ 3969 max_window = min_t(u64, tp->max_window, tp->bytes_acked); 3970 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */ 3971 if (before(ack, prior_snd_una - max_window)) { 3972 if (!(flag & FLAG_NO_CHALLENGE_ACK)) 3973 tcp_send_challenge_ack(sk); 3974 return -SKB_DROP_REASON_TCP_TOO_OLD_ACK; 3975 } 3976 goto old_ack; 3977 } 3978 3979 /* If the ack includes data we haven't sent yet, discard 3980 * this segment (RFC793 Section 3.9). 3981 */ 3982 if (after(ack, tp->snd_nxt)) 3983 return -SKB_DROP_REASON_TCP_ACK_UNSENT_DATA; 3984 3985 if (after(ack, prior_snd_una)) { 3986 flag |= FLAG_SND_UNA_ADVANCED; 3987 icsk->icsk_retransmits = 0; 3988 3989 #if IS_ENABLED(CONFIG_TLS_DEVICE) 3990 if (static_branch_unlikely(&clean_acked_data_enabled.key)) 3991 if (icsk->icsk_clean_acked) 3992 icsk->icsk_clean_acked(sk, ack); 3993 #endif 3994 } 3995 3996 prior_fack = tcp_is_sack(tp) ? tcp_highest_sack_seq(tp) : tp->snd_una; 3997 rs.prior_in_flight = tcp_packets_in_flight(tp); 3998 3999 /* ts_recent update must be made after we are sure that the packet 4000 * is in window. 4001 */ 4002 if (flag & FLAG_UPDATE_TS_RECENT) 4003 flag |= tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq); 4004 4005 if ((flag & (FLAG_SLOWPATH | FLAG_SND_UNA_ADVANCED)) == 4006 FLAG_SND_UNA_ADVANCED) { 4007 /* Window is constant, pure forward advance. 4008 * No more checks are required. 4009 * Note, we use the fact that SND.UNA>=SND.WL2. 4010 */ 4011 tcp_update_wl(tp, ack_seq); 4012 tcp_snd_una_update(tp, ack); 4013 flag |= FLAG_WIN_UPDATE; 4014 4015 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPACKS); 4016 } else { 4017 if (ack_seq != TCP_SKB_CB(skb)->end_seq) 4018 flag |= FLAG_DATA; 4019 else 4020 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPUREACKS); 4021 4022 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq); 4023 4024 if (TCP_SKB_CB(skb)->sacked) 4025 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una, 4026 &sack_state); 4027 4028 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) 4029 flag |= FLAG_ECE; 4030 4031 if (sack_state.sack_delivered) 4032 tcp_count_delivered(tp, sack_state.sack_delivered, 4033 flag & FLAG_ECE); 4034 } 4035 4036 /* This is a deviation from RFC3168 since it states that: 4037 * "When the TCP data sender is ready to set the CWR bit after reducing 4038 * the congestion window, it SHOULD set the CWR bit only on the first 4039 * new data packet that it transmits." 4040 * We accept CWR on pure ACKs to be more robust 4041 * with widely-deployed TCP implementations that do this. 4042 */ 4043 tcp_ecn_accept_cwr(sk, skb); 4044 4045 /* We passed data and got it acked, remove any soft error 4046 * log. Something worked... 4047 */ 4048 WRITE_ONCE(sk->sk_err_soft, 0); 4049 icsk->icsk_probes_out = 0; 4050 tp->rcv_tstamp = tcp_jiffies32; 4051 if (!prior_packets) 4052 goto no_queue; 4053 4054 /* See if we can take anything off of the retransmit queue. */ 4055 flag |= tcp_clean_rtx_queue(sk, skb, prior_fack, prior_snd_una, 4056 &sack_state, flag & FLAG_ECE); 4057 4058 tcp_rack_update_reo_wnd(sk, &rs); 4059 4060 tcp_in_ack_event(sk, flag); 4061 4062 if (tp->tlp_high_seq) 4063 tcp_process_tlp_ack(sk, ack, flag); 4064 4065 if (tcp_ack_is_dubious(sk, flag)) { 4066 if (!(flag & (FLAG_SND_UNA_ADVANCED | 4067 FLAG_NOT_DUP | FLAG_DSACKING_ACK))) { 4068 num_dupack = 1; 4069 /* Consider if pure acks were aggregated in tcp_add_backlog() */ 4070 if (!(flag & FLAG_DATA)) 4071 num_dupack = max_t(u16, 1, skb_shinfo(skb)->gso_segs); 4072 } 4073 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag, 4074 &rexmit); 4075 } 4076 4077 /* If needed, reset TLP/RTO timer when RACK doesn't set. */ 4078 if (flag & FLAG_SET_XMIT_TIMER) 4079 tcp_set_xmit_timer(sk); 4080 4081 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) 4082 sk_dst_confirm(sk); 4083 4084 delivered = tcp_newly_delivered(sk, delivered, flag); 4085 lost = tp->lost - lost; /* freshly marked lost */ 4086 rs.is_ack_delayed = !!(flag & FLAG_ACK_MAYBE_DELAYED); 4087 tcp_rate_gen(sk, delivered, lost, is_sack_reneg, sack_state.rate); 4088 tcp_cong_control(sk, ack, delivered, flag, sack_state.rate); 4089 tcp_xmit_recovery(sk, rexmit); 4090 return 1; 4091 4092 no_queue: 4093 tcp_in_ack_event(sk, flag); 4094 /* If data was DSACKed, see if we can undo a cwnd reduction. */ 4095 if (flag & FLAG_DSACKING_ACK) { 4096 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag, 4097 &rexmit); 4098 tcp_newly_delivered(sk, delivered, flag); 4099 } 4100 /* If this ack opens up a zero window, clear backoff. It was 4101 * being used to time the probes, and is probably far higher than 4102 * it needs to be for normal retransmission. 4103 */ 4104 tcp_ack_probe(sk); 4105 4106 if (tp->tlp_high_seq) 4107 tcp_process_tlp_ack(sk, ack, flag); 4108 return 1; 4109 4110 old_ack: 4111 /* If data was SACKed, tag it and see if we should send more data. 4112 * If data was DSACKed, see if we can undo a cwnd reduction. 4113 */ 4114 if (TCP_SKB_CB(skb)->sacked) { 4115 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una, 4116 &sack_state); 4117 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag, 4118 &rexmit); 4119 tcp_newly_delivered(sk, delivered, flag); 4120 tcp_xmit_recovery(sk, rexmit); 4121 } 4122 4123 return 0; 4124 } 4125 4126 static void tcp_parse_fastopen_option(int len, const unsigned char *cookie, 4127 bool syn, struct tcp_fastopen_cookie *foc, 4128 bool exp_opt) 4129 { 4130 /* Valid only in SYN or SYN-ACK with an even length. */ 4131 if (!foc || !syn || len < 0 || (len & 1)) 4132 return; 4133 4134 if (len >= TCP_FASTOPEN_COOKIE_MIN && 4135 len <= TCP_FASTOPEN_COOKIE_MAX) 4136 memcpy(foc->val, cookie, len); 4137 else if (len != 0) 4138 len = -1; 4139 foc->len = len; 4140 foc->exp = exp_opt; 4141 } 4142 4143 static bool smc_parse_options(const struct tcphdr *th, 4144 struct tcp_options_received *opt_rx, 4145 const unsigned char *ptr, 4146 int opsize) 4147 { 4148 #if IS_ENABLED(CONFIG_SMC) 4149 if (static_branch_unlikely(&tcp_have_smc)) { 4150 if (th->syn && !(opsize & 1) && 4151 opsize >= TCPOLEN_EXP_SMC_BASE && 4152 get_unaligned_be32(ptr) == TCPOPT_SMC_MAGIC) { 4153 opt_rx->smc_ok = 1; 4154 return true; 4155 } 4156 } 4157 #endif 4158 return false; 4159 } 4160 4161 /* Try to parse the MSS option from the TCP header. Return 0 on failure, clamped 4162 * value on success. 4163 */ 4164 u16 tcp_parse_mss_option(const struct tcphdr *th, u16 user_mss) 4165 { 4166 const unsigned char *ptr = (const unsigned char *)(th + 1); 4167 int length = (th->doff * 4) - sizeof(struct tcphdr); 4168 u16 mss = 0; 4169 4170 while (length > 0) { 4171 int opcode = *ptr++; 4172 int opsize; 4173 4174 switch (opcode) { 4175 case TCPOPT_EOL: 4176 return mss; 4177 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */ 4178 length--; 4179 continue; 4180 default: 4181 if (length < 2) 4182 return mss; 4183 opsize = *ptr++; 4184 if (opsize < 2) /* "silly options" */ 4185 return mss; 4186 if (opsize > length) 4187 return mss; /* fail on partial options */ 4188 if (opcode == TCPOPT_MSS && opsize == TCPOLEN_MSS) { 4189 u16 in_mss = get_unaligned_be16(ptr); 4190 4191 if (in_mss) { 4192 if (user_mss && user_mss < in_mss) 4193 in_mss = user_mss; 4194 mss = in_mss; 4195 } 4196 } 4197 ptr += opsize - 2; 4198 length -= opsize; 4199 } 4200 } 4201 return mss; 4202 } 4203 4204 /* Look for tcp options. Normally only called on SYN and SYNACK packets. 4205 * But, this can also be called on packets in the established flow when 4206 * the fast version below fails. 4207 */ 4208 void tcp_parse_options(const struct net *net, 4209 const struct sk_buff *skb, 4210 struct tcp_options_received *opt_rx, int estab, 4211 struct tcp_fastopen_cookie *foc) 4212 { 4213 const unsigned char *ptr; 4214 const struct tcphdr *th = tcp_hdr(skb); 4215 int length = (th->doff * 4) - sizeof(struct tcphdr); 4216 4217 ptr = (const unsigned char *)(th + 1); 4218 opt_rx->saw_tstamp = 0; 4219 opt_rx->saw_unknown = 0; 4220 4221 while (length > 0) { 4222 int opcode = *ptr++; 4223 int opsize; 4224 4225 switch (opcode) { 4226 case TCPOPT_EOL: 4227 return; 4228 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */ 4229 length--; 4230 continue; 4231 default: 4232 if (length < 2) 4233 return; 4234 opsize = *ptr++; 4235 if (opsize < 2) /* "silly options" */ 4236 return; 4237 if (opsize > length) 4238 return; /* don't parse partial options */ 4239 switch (opcode) { 4240 case TCPOPT_MSS: 4241 if (opsize == TCPOLEN_MSS && th->syn && !estab) { 4242 u16 in_mss = get_unaligned_be16(ptr); 4243 if (in_mss) { 4244 if (opt_rx->user_mss && 4245 opt_rx->user_mss < in_mss) 4246 in_mss = opt_rx->user_mss; 4247 opt_rx->mss_clamp = in_mss; 4248 } 4249 } 4250 break; 4251 case TCPOPT_WINDOW: 4252 if (opsize == TCPOLEN_WINDOW && th->syn && 4253 !estab && READ_ONCE(net->ipv4.sysctl_tcp_window_scaling)) { 4254 __u8 snd_wscale = *(__u8 *)ptr; 4255 opt_rx->wscale_ok = 1; 4256 if (snd_wscale > TCP_MAX_WSCALE) { 4257 net_info_ratelimited("%s: Illegal window scaling value %d > %u received\n", 4258 __func__, 4259 snd_wscale, 4260 TCP_MAX_WSCALE); 4261 snd_wscale = TCP_MAX_WSCALE; 4262 } 4263 opt_rx->snd_wscale = snd_wscale; 4264 } 4265 break; 4266 case TCPOPT_TIMESTAMP: 4267 if ((opsize == TCPOLEN_TIMESTAMP) && 4268 ((estab && opt_rx->tstamp_ok) || 4269 (!estab && READ_ONCE(net->ipv4.sysctl_tcp_timestamps)))) { 4270 opt_rx->saw_tstamp = 1; 4271 opt_rx->rcv_tsval = get_unaligned_be32(ptr); 4272 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4); 4273 } 4274 break; 4275 case TCPOPT_SACK_PERM: 4276 if (opsize == TCPOLEN_SACK_PERM && th->syn && 4277 !estab && READ_ONCE(net->ipv4.sysctl_tcp_sack)) { 4278 opt_rx->sack_ok = TCP_SACK_SEEN; 4279 tcp_sack_reset(opt_rx); 4280 } 4281 break; 4282 4283 case TCPOPT_SACK: 4284 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) && 4285 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) && 4286 opt_rx->sack_ok) { 4287 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th; 4288 } 4289 break; 4290 #ifdef CONFIG_TCP_MD5SIG 4291 case TCPOPT_MD5SIG: 4292 /* The MD5 Hash has already been 4293 * checked (see tcp_v{4,6}_rcv()). 4294 */ 4295 break; 4296 #endif 4297 #ifdef CONFIG_TCP_AO 4298 case TCPOPT_AO: 4299 /* TCP AO has already been checked 4300 * (see tcp_inbound_ao_hash()). 4301 */ 4302 break; 4303 #endif 4304 case TCPOPT_FASTOPEN: 4305 tcp_parse_fastopen_option( 4306 opsize - TCPOLEN_FASTOPEN_BASE, 4307 ptr, th->syn, foc, false); 4308 break; 4309 4310 case TCPOPT_EXP: 4311 /* Fast Open option shares code 254 using a 4312 * 16 bits magic number. 4313 */ 4314 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE && 4315 get_unaligned_be16(ptr) == 4316 TCPOPT_FASTOPEN_MAGIC) { 4317 tcp_parse_fastopen_option(opsize - 4318 TCPOLEN_EXP_FASTOPEN_BASE, 4319 ptr + 2, th->syn, foc, true); 4320 break; 4321 } 4322 4323 if (smc_parse_options(th, opt_rx, ptr, opsize)) 4324 break; 4325 4326 opt_rx->saw_unknown = 1; 4327 break; 4328 4329 default: 4330 opt_rx->saw_unknown = 1; 4331 } 4332 ptr += opsize-2; 4333 length -= opsize; 4334 } 4335 } 4336 } 4337 EXPORT_SYMBOL(tcp_parse_options); 4338 4339 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th) 4340 { 4341 const __be32 *ptr = (const __be32 *)(th + 1); 4342 4343 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) 4344 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) { 4345 tp->rx_opt.saw_tstamp = 1; 4346 ++ptr; 4347 tp->rx_opt.rcv_tsval = ntohl(*ptr); 4348 ++ptr; 4349 if (*ptr) 4350 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset; 4351 else 4352 tp->rx_opt.rcv_tsecr = 0; 4353 return true; 4354 } 4355 return false; 4356 } 4357 4358 /* Fast parse options. This hopes to only see timestamps. 4359 * If it is wrong it falls back on tcp_parse_options(). 4360 */ 4361 static bool tcp_fast_parse_options(const struct net *net, 4362 const struct sk_buff *skb, 4363 const struct tcphdr *th, struct tcp_sock *tp) 4364 { 4365 /* In the spirit of fast parsing, compare doff directly to constant 4366 * values. Because equality is used, short doff can be ignored here. 4367 */ 4368 if (th->doff == (sizeof(*th) / 4)) { 4369 tp->rx_opt.saw_tstamp = 0; 4370 return false; 4371 } else if (tp->rx_opt.tstamp_ok && 4372 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) { 4373 if (tcp_parse_aligned_timestamp(tp, th)) 4374 return true; 4375 } 4376 4377 tcp_parse_options(net, skb, &tp->rx_opt, 1, NULL); 4378 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr) 4379 tp->rx_opt.rcv_tsecr -= tp->tsoffset; 4380 4381 return true; 4382 } 4383 4384 #if defined(CONFIG_TCP_MD5SIG) || defined(CONFIG_TCP_AO) 4385 /* 4386 * Parse Signature options 4387 */ 4388 int tcp_do_parse_auth_options(const struct tcphdr *th, 4389 const u8 **md5_hash, const u8 **ao_hash) 4390 { 4391 int length = (th->doff << 2) - sizeof(*th); 4392 const u8 *ptr = (const u8 *)(th + 1); 4393 unsigned int minlen = TCPOLEN_MD5SIG; 4394 4395 if (IS_ENABLED(CONFIG_TCP_AO)) 4396 minlen = sizeof(struct tcp_ao_hdr) + 1; 4397 4398 *md5_hash = NULL; 4399 *ao_hash = NULL; 4400 4401 /* If not enough data remaining, we can short cut */ 4402 while (length >= minlen) { 4403 int opcode = *ptr++; 4404 int opsize; 4405 4406 switch (opcode) { 4407 case TCPOPT_EOL: 4408 return 0; 4409 case TCPOPT_NOP: 4410 length--; 4411 continue; 4412 default: 4413 opsize = *ptr++; 4414 if (opsize < 2 || opsize > length) 4415 return -EINVAL; 4416 if (opcode == TCPOPT_MD5SIG) { 4417 if (opsize != TCPOLEN_MD5SIG) 4418 return -EINVAL; 4419 if (unlikely(*md5_hash || *ao_hash)) 4420 return -EEXIST; 4421 *md5_hash = ptr; 4422 } else if (opcode == TCPOPT_AO) { 4423 if (opsize <= sizeof(struct tcp_ao_hdr)) 4424 return -EINVAL; 4425 if (unlikely(*md5_hash || *ao_hash)) 4426 return -EEXIST; 4427 *ao_hash = ptr; 4428 } 4429 } 4430 ptr += opsize - 2; 4431 length -= opsize; 4432 } 4433 return 0; 4434 } 4435 EXPORT_SYMBOL(tcp_do_parse_auth_options); 4436 #endif 4437 4438 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM 4439 * 4440 * It is not fatal. If this ACK does _not_ change critical state (seqs, window) 4441 * it can pass through stack. So, the following predicate verifies that 4442 * this segment is not used for anything but congestion avoidance or 4443 * fast retransmit. Moreover, we even are able to eliminate most of such 4444 * second order effects, if we apply some small "replay" window (~RTO) 4445 * to timestamp space. 4446 * 4447 * All these measures still do not guarantee that we reject wrapped ACKs 4448 * on networks with high bandwidth, when sequence space is recycled fastly, 4449 * but it guarantees that such events will be very rare and do not affect 4450 * connection seriously. This doesn't look nice, but alas, PAWS is really 4451 * buggy extension. 4452 * 4453 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC 4454 * states that events when retransmit arrives after original data are rare. 4455 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is 4456 * the biggest problem on large power networks even with minor reordering. 4457 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe 4458 * up to bandwidth of 18Gigabit/sec. 8) ] 4459 */ 4460 4461 /* Estimates max number of increments of remote peer TSval in 4462 * a replay window (based on our current RTO estimation). 4463 */ 4464 static u32 tcp_tsval_replay(const struct sock *sk) 4465 { 4466 /* If we use usec TS resolution, 4467 * then expect the remote peer to use the same resolution. 4468 */ 4469 if (tcp_sk(sk)->tcp_usec_ts) 4470 return inet_csk(sk)->icsk_rto * (USEC_PER_SEC / HZ); 4471 4472 /* RFC 7323 recommends a TSval clock between 1ms and 1sec. 4473 * We know that some OS (including old linux) can use 1200 Hz. 4474 */ 4475 return inet_csk(sk)->icsk_rto * 1200 / HZ; 4476 } 4477 4478 static enum skb_drop_reason tcp_disordered_ack_check(const struct sock *sk, 4479 const struct sk_buff *skb) 4480 { 4481 const struct tcp_sock *tp = tcp_sk(sk); 4482 const struct tcphdr *th = tcp_hdr(skb); 4483 SKB_DR_INIT(reason, TCP_RFC7323_PAWS); 4484 u32 ack = TCP_SKB_CB(skb)->ack_seq; 4485 u32 seq = TCP_SKB_CB(skb)->seq; 4486 4487 /* 1. Is this not a pure ACK ? */ 4488 if (!th->ack || seq != TCP_SKB_CB(skb)->end_seq) 4489 return reason; 4490 4491 /* 2. Is its sequence not the expected one ? */ 4492 if (seq != tp->rcv_nxt) 4493 return before(seq, tp->rcv_nxt) ? 4494 SKB_DROP_REASON_TCP_RFC7323_PAWS_ACK : 4495 reason; 4496 4497 /* 3. Is this not a duplicate ACK ? */ 4498 if (ack != tp->snd_una) 4499 return reason; 4500 4501 /* 4. Is this updating the window ? */ 4502 if (tcp_may_update_window(tp, ack, seq, ntohs(th->window) << 4503 tp->rx_opt.snd_wscale)) 4504 return reason; 4505 4506 /* 5. Is this not in the replay window ? */ 4507 if ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > 4508 tcp_tsval_replay(sk)) 4509 return reason; 4510 4511 return 0; 4512 } 4513 4514 /* Check segment sequence number for validity. 4515 * 4516 * Segment controls are considered valid, if the segment 4517 * fits to the window after truncation to the window. Acceptability 4518 * of data (and SYN, FIN, of course) is checked separately. 4519 * See tcp_data_queue(), for example. 4520 * 4521 * Also, controls (RST is main one) are accepted using RCV.WUP instead 4522 * of RCV.NXT. Peer still did not advance his SND.UNA when we 4523 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP. 4524 * (borrowed from freebsd) 4525 */ 4526 4527 static enum skb_drop_reason tcp_sequence(const struct tcp_sock *tp, 4528 u32 seq, u32 end_seq) 4529 { 4530 if (before(end_seq, tp->rcv_wup)) 4531 return SKB_DROP_REASON_TCP_OLD_SEQUENCE; 4532 4533 if (after(seq, tp->rcv_nxt + tcp_receive_window(tp))) 4534 return SKB_DROP_REASON_TCP_INVALID_SEQUENCE; 4535 4536 return SKB_NOT_DROPPED_YET; 4537 } 4538 4539 4540 void tcp_done_with_error(struct sock *sk, int err) 4541 { 4542 /* This barrier is coupled with smp_rmb() in tcp_poll() */ 4543 WRITE_ONCE(sk->sk_err, err); 4544 smp_wmb(); 4545 4546 tcp_write_queue_purge(sk); 4547 tcp_done(sk); 4548 4549 if (!sock_flag(sk, SOCK_DEAD)) 4550 sk_error_report(sk); 4551 } 4552 EXPORT_IPV6_MOD(tcp_done_with_error); 4553 4554 /* When we get a reset we do this. */ 4555 void tcp_reset(struct sock *sk, struct sk_buff *skb) 4556 { 4557 int err; 4558 4559 trace_tcp_receive_reset(sk); 4560 4561 /* mptcp can't tell us to ignore reset pkts, 4562 * so just ignore the return value of mptcp_incoming_options(). 4563 */ 4564 if (sk_is_mptcp(sk)) 4565 mptcp_incoming_options(sk, skb); 4566 4567 /* We want the right error as BSD sees it (and indeed as we do). */ 4568 switch (sk->sk_state) { 4569 case TCP_SYN_SENT: 4570 err = ECONNREFUSED; 4571 break; 4572 case TCP_CLOSE_WAIT: 4573 err = EPIPE; 4574 break; 4575 case TCP_CLOSE: 4576 return; 4577 default: 4578 err = ECONNRESET; 4579 } 4580 tcp_done_with_error(sk, err); 4581 } 4582 4583 /* 4584 * Process the FIN bit. This now behaves as it is supposed to work 4585 * and the FIN takes effect when it is validly part of sequence 4586 * space. Not before when we get holes. 4587 * 4588 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT 4589 * (and thence onto LAST-ACK and finally, CLOSE, we never enter 4590 * TIME-WAIT) 4591 * 4592 * If we are in FINWAIT-1, a received FIN indicates simultaneous 4593 * close and we go into CLOSING (and later onto TIME-WAIT) 4594 * 4595 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT. 4596 */ 4597 void tcp_fin(struct sock *sk) 4598 { 4599 struct tcp_sock *tp = tcp_sk(sk); 4600 4601 inet_csk_schedule_ack(sk); 4602 4603 WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | RCV_SHUTDOWN); 4604 sock_set_flag(sk, SOCK_DONE); 4605 4606 switch (sk->sk_state) { 4607 case TCP_SYN_RECV: 4608 case TCP_ESTABLISHED: 4609 /* Move to CLOSE_WAIT */ 4610 tcp_set_state(sk, TCP_CLOSE_WAIT); 4611 inet_csk_enter_pingpong_mode(sk); 4612 break; 4613 4614 case TCP_CLOSE_WAIT: 4615 case TCP_CLOSING: 4616 /* Received a retransmission of the FIN, do 4617 * nothing. 4618 */ 4619 break; 4620 case TCP_LAST_ACK: 4621 /* RFC793: Remain in the LAST-ACK state. */ 4622 break; 4623 4624 case TCP_FIN_WAIT1: 4625 /* This case occurs when a simultaneous close 4626 * happens, we must ack the received FIN and 4627 * enter the CLOSING state. 4628 */ 4629 tcp_send_ack(sk); 4630 tcp_set_state(sk, TCP_CLOSING); 4631 break; 4632 case TCP_FIN_WAIT2: 4633 /* Received a FIN -- send ACK and enter TIME_WAIT. */ 4634 tcp_send_ack(sk); 4635 tcp_time_wait(sk, TCP_TIME_WAIT, 0); 4636 break; 4637 default: 4638 /* Only TCP_LISTEN and TCP_CLOSE are left, in these 4639 * cases we should never reach this piece of code. 4640 */ 4641 pr_err("%s: Impossible, sk->sk_state=%d\n", 4642 __func__, sk->sk_state); 4643 break; 4644 } 4645 4646 /* It _is_ possible, that we have something out-of-order _after_ FIN. 4647 * Probably, we should reset in this case. For now drop them. 4648 */ 4649 skb_rbtree_purge(&tp->out_of_order_queue); 4650 if (tcp_is_sack(tp)) 4651 tcp_sack_reset(&tp->rx_opt); 4652 4653 if (!sock_flag(sk, SOCK_DEAD)) { 4654 sk->sk_state_change(sk); 4655 4656 /* Do not send POLL_HUP for half duplex close. */ 4657 if (sk->sk_shutdown == SHUTDOWN_MASK || 4658 sk->sk_state == TCP_CLOSE) 4659 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP); 4660 else 4661 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN); 4662 } 4663 } 4664 4665 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, 4666 u32 end_seq) 4667 { 4668 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) { 4669 if (before(seq, sp->start_seq)) 4670 sp->start_seq = seq; 4671 if (after(end_seq, sp->end_seq)) 4672 sp->end_seq = end_seq; 4673 return true; 4674 } 4675 return false; 4676 } 4677 4678 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq) 4679 { 4680 struct tcp_sock *tp = tcp_sk(sk); 4681 4682 if (tcp_is_sack(tp) && READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_dsack)) { 4683 int mib_idx; 4684 4685 if (before(seq, tp->rcv_nxt)) 4686 mib_idx = LINUX_MIB_TCPDSACKOLDSENT; 4687 else 4688 mib_idx = LINUX_MIB_TCPDSACKOFOSENT; 4689 4690 NET_INC_STATS(sock_net(sk), mib_idx); 4691 4692 tp->rx_opt.dsack = 1; 4693 tp->duplicate_sack[0].start_seq = seq; 4694 tp->duplicate_sack[0].end_seq = end_seq; 4695 } 4696 } 4697 4698 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq) 4699 { 4700 struct tcp_sock *tp = tcp_sk(sk); 4701 4702 if (!tp->rx_opt.dsack) 4703 tcp_dsack_set(sk, seq, end_seq); 4704 else 4705 tcp_sack_extend(tp->duplicate_sack, seq, end_seq); 4706 } 4707 4708 static void tcp_rcv_spurious_retrans(struct sock *sk, const struct sk_buff *skb) 4709 { 4710 /* When the ACK path fails or drops most ACKs, the sender would 4711 * timeout and spuriously retransmit the same segment repeatedly. 4712 * If it seems our ACKs are not reaching the other side, 4713 * based on receiving a duplicate data segment with new flowlabel 4714 * (suggesting the sender suffered an RTO), and we are not already 4715 * repathing due to our own RTO, then rehash the socket to repath our 4716 * packets. 4717 */ 4718 #if IS_ENABLED(CONFIG_IPV6) 4719 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Loss && 4720 skb->protocol == htons(ETH_P_IPV6) && 4721 (tcp_sk(sk)->inet_conn.icsk_ack.lrcv_flowlabel != 4722 ntohl(ip6_flowlabel(ipv6_hdr(skb)))) && 4723 sk_rethink_txhash(sk)) 4724 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDUPLICATEDATAREHASH); 4725 4726 /* Save last flowlabel after a spurious retrans. */ 4727 tcp_save_lrcv_flowlabel(sk, skb); 4728 #endif 4729 } 4730 4731 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb) 4732 { 4733 struct tcp_sock *tp = tcp_sk(sk); 4734 4735 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 4736 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 4737 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST); 4738 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS); 4739 4740 if (tcp_is_sack(tp) && READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_dsack)) { 4741 u32 end_seq = TCP_SKB_CB(skb)->end_seq; 4742 4743 tcp_rcv_spurious_retrans(sk, skb); 4744 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) 4745 end_seq = tp->rcv_nxt; 4746 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq); 4747 } 4748 } 4749 4750 tcp_send_ack(sk); 4751 } 4752 4753 /* These routines update the SACK block as out-of-order packets arrive or 4754 * in-order packets close up the sequence space. 4755 */ 4756 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp) 4757 { 4758 int this_sack; 4759 struct tcp_sack_block *sp = &tp->selective_acks[0]; 4760 struct tcp_sack_block *swalk = sp + 1; 4761 4762 /* See if the recent change to the first SACK eats into 4763 * or hits the sequence space of other SACK blocks, if so coalesce. 4764 */ 4765 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) { 4766 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) { 4767 int i; 4768 4769 /* Zap SWALK, by moving every further SACK up by one slot. 4770 * Decrease num_sacks. 4771 */ 4772 tp->rx_opt.num_sacks--; 4773 for (i = this_sack; i < tp->rx_opt.num_sacks; i++) 4774 sp[i] = sp[i + 1]; 4775 continue; 4776 } 4777 this_sack++; 4778 swalk++; 4779 } 4780 } 4781 4782 void tcp_sack_compress_send_ack(struct sock *sk) 4783 { 4784 struct tcp_sock *tp = tcp_sk(sk); 4785 4786 if (!tp->compressed_ack) 4787 return; 4788 4789 if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1) 4790 __sock_put(sk); 4791 4792 /* Since we have to send one ack finally, 4793 * substract one from tp->compressed_ack to keep 4794 * LINUX_MIB_TCPACKCOMPRESSED accurate. 4795 */ 4796 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED, 4797 tp->compressed_ack - 1); 4798 4799 tp->compressed_ack = 0; 4800 tcp_send_ack(sk); 4801 } 4802 4803 /* Reasonable amount of sack blocks included in TCP SACK option 4804 * The max is 4, but this becomes 3 if TCP timestamps are there. 4805 * Given that SACK packets might be lost, be conservative and use 2. 4806 */ 4807 #define TCP_SACK_BLOCKS_EXPECTED 2 4808 4809 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq) 4810 { 4811 struct tcp_sock *tp = tcp_sk(sk); 4812 struct tcp_sack_block *sp = &tp->selective_acks[0]; 4813 int cur_sacks = tp->rx_opt.num_sacks; 4814 int this_sack; 4815 4816 if (!cur_sacks) 4817 goto new_sack; 4818 4819 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) { 4820 if (tcp_sack_extend(sp, seq, end_seq)) { 4821 if (this_sack >= TCP_SACK_BLOCKS_EXPECTED) 4822 tcp_sack_compress_send_ack(sk); 4823 /* Rotate this_sack to the first one. */ 4824 for (; this_sack > 0; this_sack--, sp--) 4825 swap(*sp, *(sp - 1)); 4826 if (cur_sacks > 1) 4827 tcp_sack_maybe_coalesce(tp); 4828 return; 4829 } 4830 } 4831 4832 if (this_sack >= TCP_SACK_BLOCKS_EXPECTED) 4833 tcp_sack_compress_send_ack(sk); 4834 4835 /* Could not find an adjacent existing SACK, build a new one, 4836 * put it at the front, and shift everyone else down. We 4837 * always know there is at least one SACK present already here. 4838 * 4839 * If the sack array is full, forget about the last one. 4840 */ 4841 if (this_sack >= TCP_NUM_SACKS) { 4842 this_sack--; 4843 tp->rx_opt.num_sacks--; 4844 sp--; 4845 } 4846 for (; this_sack > 0; this_sack--, sp--) 4847 *sp = *(sp - 1); 4848 4849 new_sack: 4850 /* Build the new head SACK, and we're done. */ 4851 sp->start_seq = seq; 4852 sp->end_seq = end_seq; 4853 tp->rx_opt.num_sacks++; 4854 } 4855 4856 /* RCV.NXT advances, some SACKs should be eaten. */ 4857 4858 static void tcp_sack_remove(struct tcp_sock *tp) 4859 { 4860 struct tcp_sack_block *sp = &tp->selective_acks[0]; 4861 int num_sacks = tp->rx_opt.num_sacks; 4862 int this_sack; 4863 4864 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */ 4865 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) { 4866 tp->rx_opt.num_sacks = 0; 4867 return; 4868 } 4869 4870 for (this_sack = 0; this_sack < num_sacks;) { 4871 /* Check if the start of the sack is covered by RCV.NXT. */ 4872 if (!before(tp->rcv_nxt, sp->start_seq)) { 4873 int i; 4874 4875 /* RCV.NXT must cover all the block! */ 4876 WARN_ON(before(tp->rcv_nxt, sp->end_seq)); 4877 4878 /* Zap this SACK, by moving forward any other SACKS. */ 4879 for (i = this_sack+1; i < num_sacks; i++) 4880 tp->selective_acks[i-1] = tp->selective_acks[i]; 4881 num_sacks--; 4882 continue; 4883 } 4884 this_sack++; 4885 sp++; 4886 } 4887 tp->rx_opt.num_sacks = num_sacks; 4888 } 4889 4890 /** 4891 * tcp_try_coalesce - try to merge skb to prior one 4892 * @sk: socket 4893 * @to: prior buffer 4894 * @from: buffer to add in queue 4895 * @fragstolen: pointer to boolean 4896 * 4897 * Before queueing skb @from after @to, try to merge them 4898 * to reduce overall memory use and queue lengths, if cost is small. 4899 * Packets in ofo or receive queues can stay a long time. 4900 * Better try to coalesce them right now to avoid future collapses. 4901 * Returns true if caller should free @from instead of queueing it 4902 */ 4903 static bool tcp_try_coalesce(struct sock *sk, 4904 struct sk_buff *to, 4905 struct sk_buff *from, 4906 bool *fragstolen) 4907 { 4908 int delta; 4909 4910 *fragstolen = false; 4911 4912 /* Its possible this segment overlaps with prior segment in queue */ 4913 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq) 4914 return false; 4915 4916 if (!tcp_skb_can_collapse_rx(to, from)) 4917 return false; 4918 4919 if (!skb_try_coalesce(to, from, fragstolen, &delta)) 4920 return false; 4921 4922 atomic_add(delta, &sk->sk_rmem_alloc); 4923 sk_mem_charge(sk, delta); 4924 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE); 4925 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq; 4926 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq; 4927 TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags; 4928 4929 if (TCP_SKB_CB(from)->has_rxtstamp) { 4930 TCP_SKB_CB(to)->has_rxtstamp = true; 4931 to->tstamp = from->tstamp; 4932 skb_hwtstamps(to)->hwtstamp = skb_hwtstamps(from)->hwtstamp; 4933 } 4934 4935 return true; 4936 } 4937 4938 static bool tcp_ooo_try_coalesce(struct sock *sk, 4939 struct sk_buff *to, 4940 struct sk_buff *from, 4941 bool *fragstolen) 4942 { 4943 bool res = tcp_try_coalesce(sk, to, from, fragstolen); 4944 4945 /* In case tcp_drop_reason() is called later, update to->gso_segs */ 4946 if (res) { 4947 u32 gso_segs = max_t(u16, 1, skb_shinfo(to)->gso_segs) + 4948 max_t(u16, 1, skb_shinfo(from)->gso_segs); 4949 4950 skb_shinfo(to)->gso_segs = min_t(u32, gso_segs, 0xFFFF); 4951 } 4952 return res; 4953 } 4954 4955 noinline_for_tracing static void 4956 tcp_drop_reason(struct sock *sk, struct sk_buff *skb, enum skb_drop_reason reason) 4957 { 4958 sk_drops_add(sk, skb); 4959 sk_skb_reason_drop(sk, skb, reason); 4960 } 4961 4962 /* This one checks to see if we can put data from the 4963 * out_of_order queue into the receive_queue. 4964 */ 4965 static void tcp_ofo_queue(struct sock *sk) 4966 { 4967 struct tcp_sock *tp = tcp_sk(sk); 4968 __u32 dsack_high = tp->rcv_nxt; 4969 bool fin, fragstolen, eaten; 4970 struct sk_buff *skb, *tail; 4971 struct rb_node *p; 4972 4973 p = rb_first(&tp->out_of_order_queue); 4974 while (p) { 4975 skb = rb_to_skb(p); 4976 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) 4977 break; 4978 4979 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) { 4980 __u32 dsack = dsack_high; 4981 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high)) 4982 dsack_high = TCP_SKB_CB(skb)->end_seq; 4983 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack); 4984 } 4985 p = rb_next(p); 4986 rb_erase(&skb->rbnode, &tp->out_of_order_queue); 4987 4988 if (unlikely(!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))) { 4989 tcp_drop_reason(sk, skb, SKB_DROP_REASON_TCP_OFO_DROP); 4990 continue; 4991 } 4992 4993 tail = skb_peek_tail(&sk->sk_receive_queue); 4994 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen); 4995 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq); 4996 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN; 4997 if (!eaten) 4998 tcp_add_receive_queue(sk, skb); 4999 else 5000 kfree_skb_partial(skb, fragstolen); 5001 5002 if (unlikely(fin)) { 5003 tcp_fin(sk); 5004 /* tcp_fin() purges tp->out_of_order_queue, 5005 * so we must end this loop right now. 5006 */ 5007 break; 5008 } 5009 } 5010 } 5011 5012 static bool tcp_prune_ofo_queue(struct sock *sk, const struct sk_buff *in_skb); 5013 static int tcp_prune_queue(struct sock *sk, const struct sk_buff *in_skb); 5014 5015 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb, 5016 unsigned int size) 5017 { 5018 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf || 5019 !sk_rmem_schedule(sk, skb, size)) { 5020 5021 if (tcp_prune_queue(sk, skb) < 0) 5022 return -1; 5023 5024 while (!sk_rmem_schedule(sk, skb, size)) { 5025 if (!tcp_prune_ofo_queue(sk, skb)) 5026 return -1; 5027 } 5028 } 5029 return 0; 5030 } 5031 5032 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb) 5033 { 5034 struct tcp_sock *tp = tcp_sk(sk); 5035 struct rb_node **p, *parent; 5036 struct sk_buff *skb1; 5037 u32 seq, end_seq; 5038 bool fragstolen; 5039 5040 tcp_save_lrcv_flowlabel(sk, skb); 5041 tcp_data_ecn_check(sk, skb); 5042 5043 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) { 5044 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFODROP); 5045 sk->sk_data_ready(sk); 5046 tcp_drop_reason(sk, skb, SKB_DROP_REASON_PROTO_MEM); 5047 return; 5048 } 5049 5050 /* Disable header prediction. */ 5051 tp->pred_flags = 0; 5052 inet_csk_schedule_ack(sk); 5053 5054 tp->rcv_ooopack += max_t(u16, 1, skb_shinfo(skb)->gso_segs); 5055 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOQUEUE); 5056 seq = TCP_SKB_CB(skb)->seq; 5057 end_seq = TCP_SKB_CB(skb)->end_seq; 5058 5059 p = &tp->out_of_order_queue.rb_node; 5060 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) { 5061 /* Initial out of order segment, build 1 SACK. */ 5062 if (tcp_is_sack(tp)) { 5063 tp->rx_opt.num_sacks = 1; 5064 tp->selective_acks[0].start_seq = seq; 5065 tp->selective_acks[0].end_seq = end_seq; 5066 } 5067 rb_link_node(&skb->rbnode, NULL, p); 5068 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue); 5069 tp->ooo_last_skb = skb; 5070 goto end; 5071 } 5072 5073 /* In the typical case, we are adding an skb to the end of the list. 5074 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup. 5075 */ 5076 if (tcp_ooo_try_coalesce(sk, tp->ooo_last_skb, 5077 skb, &fragstolen)) { 5078 coalesce_done: 5079 /* For non sack flows, do not grow window to force DUPACK 5080 * and trigger fast retransmit. 5081 */ 5082 if (tcp_is_sack(tp)) 5083 tcp_grow_window(sk, skb, true); 5084 kfree_skb_partial(skb, fragstolen); 5085 skb = NULL; 5086 goto add_sack; 5087 } 5088 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */ 5089 if (!before(seq, TCP_SKB_CB(tp->ooo_last_skb)->end_seq)) { 5090 parent = &tp->ooo_last_skb->rbnode; 5091 p = &parent->rb_right; 5092 goto insert; 5093 } 5094 5095 /* Find place to insert this segment. Handle overlaps on the way. */ 5096 parent = NULL; 5097 while (*p) { 5098 parent = *p; 5099 skb1 = rb_to_skb(parent); 5100 if (before(seq, TCP_SKB_CB(skb1)->seq)) { 5101 p = &parent->rb_left; 5102 continue; 5103 } 5104 if (before(seq, TCP_SKB_CB(skb1)->end_seq)) { 5105 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) { 5106 /* All the bits are present. Drop. */ 5107 NET_INC_STATS(sock_net(sk), 5108 LINUX_MIB_TCPOFOMERGE); 5109 tcp_drop_reason(sk, skb, 5110 SKB_DROP_REASON_TCP_OFOMERGE); 5111 skb = NULL; 5112 tcp_dsack_set(sk, seq, end_seq); 5113 goto add_sack; 5114 } 5115 if (after(seq, TCP_SKB_CB(skb1)->seq)) { 5116 /* Partial overlap. */ 5117 tcp_dsack_set(sk, seq, TCP_SKB_CB(skb1)->end_seq); 5118 } else { 5119 /* skb's seq == skb1's seq and skb covers skb1. 5120 * Replace skb1 with skb. 5121 */ 5122 rb_replace_node(&skb1->rbnode, &skb->rbnode, 5123 &tp->out_of_order_queue); 5124 tcp_dsack_extend(sk, 5125 TCP_SKB_CB(skb1)->seq, 5126 TCP_SKB_CB(skb1)->end_seq); 5127 NET_INC_STATS(sock_net(sk), 5128 LINUX_MIB_TCPOFOMERGE); 5129 tcp_drop_reason(sk, skb1, 5130 SKB_DROP_REASON_TCP_OFOMERGE); 5131 goto merge_right; 5132 } 5133 } else if (tcp_ooo_try_coalesce(sk, skb1, 5134 skb, &fragstolen)) { 5135 goto coalesce_done; 5136 } 5137 p = &parent->rb_right; 5138 } 5139 insert: 5140 /* Insert segment into RB tree. */ 5141 rb_link_node(&skb->rbnode, parent, p); 5142 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue); 5143 5144 merge_right: 5145 /* Remove other segments covered by skb. */ 5146 while ((skb1 = skb_rb_next(skb)) != NULL) { 5147 if (!after(end_seq, TCP_SKB_CB(skb1)->seq)) 5148 break; 5149 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) { 5150 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq, 5151 end_seq); 5152 break; 5153 } 5154 rb_erase(&skb1->rbnode, &tp->out_of_order_queue); 5155 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq, 5156 TCP_SKB_CB(skb1)->end_seq); 5157 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOMERGE); 5158 tcp_drop_reason(sk, skb1, SKB_DROP_REASON_TCP_OFOMERGE); 5159 } 5160 /* If there is no skb after us, we are the last_skb ! */ 5161 if (!skb1) 5162 tp->ooo_last_skb = skb; 5163 5164 add_sack: 5165 if (tcp_is_sack(tp)) 5166 tcp_sack_new_ofo_skb(sk, seq, end_seq); 5167 end: 5168 if (skb) { 5169 /* For non sack flows, do not grow window to force DUPACK 5170 * and trigger fast retransmit. 5171 */ 5172 if (tcp_is_sack(tp)) 5173 tcp_grow_window(sk, skb, false); 5174 skb_condense(skb); 5175 skb_set_owner_r(skb, sk); 5176 } 5177 } 5178 5179 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, 5180 bool *fragstolen) 5181 { 5182 int eaten; 5183 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue); 5184 5185 eaten = (tail && 5186 tcp_try_coalesce(sk, tail, 5187 skb, fragstolen)) ? 1 : 0; 5188 tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq); 5189 if (!eaten) { 5190 tcp_add_receive_queue(sk, skb); 5191 skb_set_owner_r(skb, sk); 5192 } 5193 return eaten; 5194 } 5195 5196 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size) 5197 { 5198 struct sk_buff *skb; 5199 int err = -ENOMEM; 5200 int data_len = 0; 5201 bool fragstolen; 5202 5203 if (size == 0) 5204 return 0; 5205 5206 if (size > PAGE_SIZE) { 5207 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS); 5208 5209 data_len = npages << PAGE_SHIFT; 5210 size = data_len + (size & ~PAGE_MASK); 5211 } 5212 skb = alloc_skb_with_frags(size - data_len, data_len, 5213 PAGE_ALLOC_COSTLY_ORDER, 5214 &err, sk->sk_allocation); 5215 if (!skb) 5216 goto err; 5217 5218 skb_put(skb, size - data_len); 5219 skb->data_len = data_len; 5220 skb->len = size; 5221 5222 if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) { 5223 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP); 5224 goto err_free; 5225 } 5226 5227 err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size); 5228 if (err) 5229 goto err_free; 5230 5231 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt; 5232 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size; 5233 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1; 5234 5235 if (tcp_queue_rcv(sk, skb, &fragstolen)) { 5236 WARN_ON_ONCE(fragstolen); /* should not happen */ 5237 __kfree_skb(skb); 5238 } 5239 return size; 5240 5241 err_free: 5242 kfree_skb(skb); 5243 err: 5244 return err; 5245 5246 } 5247 5248 void tcp_data_ready(struct sock *sk) 5249 { 5250 if (tcp_epollin_ready(sk, sk->sk_rcvlowat) || sock_flag(sk, SOCK_DONE)) 5251 sk->sk_data_ready(sk); 5252 } 5253 5254 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb) 5255 { 5256 struct tcp_sock *tp = tcp_sk(sk); 5257 enum skb_drop_reason reason; 5258 bool fragstolen; 5259 int eaten; 5260 5261 /* If a subflow has been reset, the packet should not continue 5262 * to be processed, drop the packet. 5263 */ 5264 if (sk_is_mptcp(sk) && !mptcp_incoming_options(sk, skb)) { 5265 __kfree_skb(skb); 5266 return; 5267 } 5268 5269 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) { 5270 __kfree_skb(skb); 5271 return; 5272 } 5273 tcp_cleanup_skb(skb); 5274 __skb_pull(skb, tcp_hdr(skb)->doff * 4); 5275 5276 reason = SKB_DROP_REASON_NOT_SPECIFIED; 5277 tp->rx_opt.dsack = 0; 5278 5279 /* Queue data for delivery to the user. 5280 * Packets in sequence go to the receive queue. 5281 * Out of sequence packets to the out_of_order_queue. 5282 */ 5283 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) { 5284 if (tcp_receive_window(tp) == 0) { 5285 /* Some stacks are known to send bare FIN packets 5286 * in a loop even if we send RWIN 0 in our ACK. 5287 * Accepting this FIN does not hurt memory pressure 5288 * because the FIN flag will simply be merged to the 5289 * receive queue tail skb in most cases. 5290 */ 5291 if (!skb->len && 5292 (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) 5293 goto queue_and_out; 5294 5295 reason = SKB_DROP_REASON_TCP_ZEROWINDOW; 5296 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP); 5297 goto out_of_window; 5298 } 5299 5300 /* Ok. In sequence. In window. */ 5301 queue_and_out: 5302 if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) { 5303 /* TODO: maybe ratelimit these WIN 0 ACK ? */ 5304 inet_csk(sk)->icsk_ack.pending |= 5305 (ICSK_ACK_NOMEM | ICSK_ACK_NOW); 5306 inet_csk_schedule_ack(sk); 5307 sk->sk_data_ready(sk); 5308 5309 if (skb_queue_len(&sk->sk_receive_queue) && skb->len) { 5310 reason = SKB_DROP_REASON_PROTO_MEM; 5311 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP); 5312 goto drop; 5313 } 5314 sk_forced_mem_schedule(sk, skb->truesize); 5315 } 5316 5317 eaten = tcp_queue_rcv(sk, skb, &fragstolen); 5318 if (skb->len) 5319 tcp_event_data_recv(sk, skb); 5320 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 5321 tcp_fin(sk); 5322 5323 if (!RB_EMPTY_ROOT(&tp->out_of_order_queue)) { 5324 tcp_ofo_queue(sk); 5325 5326 /* RFC5681. 4.2. SHOULD send immediate ACK, when 5327 * gap in queue is filled. 5328 */ 5329 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) 5330 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW; 5331 } 5332 5333 if (tp->rx_opt.num_sacks) 5334 tcp_sack_remove(tp); 5335 5336 tcp_fast_path_check(sk); 5337 5338 if (eaten > 0) 5339 kfree_skb_partial(skb, fragstolen); 5340 if (!sock_flag(sk, SOCK_DEAD)) 5341 tcp_data_ready(sk); 5342 return; 5343 } 5344 5345 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) { 5346 tcp_rcv_spurious_retrans(sk, skb); 5347 /* A retransmit, 2nd most common case. Force an immediate ack. */ 5348 reason = SKB_DROP_REASON_TCP_OLD_DATA; 5349 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST); 5350 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); 5351 5352 out_of_window: 5353 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS); 5354 inet_csk_schedule_ack(sk); 5355 drop: 5356 tcp_drop_reason(sk, skb, reason); 5357 return; 5358 } 5359 5360 /* Out of window. F.e. zero window probe. */ 5361 if (!before(TCP_SKB_CB(skb)->seq, 5362 tp->rcv_nxt + tcp_receive_window(tp))) { 5363 reason = SKB_DROP_REASON_TCP_OVERWINDOW; 5364 goto out_of_window; 5365 } 5366 5367 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 5368 /* Partial packet, seq < rcv_next < end_seq */ 5369 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt); 5370 5371 /* If window is closed, drop tail of packet. But after 5372 * remembering D-SACK for its head made in previous line. 5373 */ 5374 if (!tcp_receive_window(tp)) { 5375 reason = SKB_DROP_REASON_TCP_ZEROWINDOW; 5376 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP); 5377 goto out_of_window; 5378 } 5379 goto queue_and_out; 5380 } 5381 5382 tcp_data_queue_ofo(sk, skb); 5383 } 5384 5385 static struct sk_buff *tcp_skb_next(struct sk_buff *skb, struct sk_buff_head *list) 5386 { 5387 if (list) 5388 return !skb_queue_is_last(list, skb) ? skb->next : NULL; 5389 5390 return skb_rb_next(skb); 5391 } 5392 5393 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb, 5394 struct sk_buff_head *list, 5395 struct rb_root *root) 5396 { 5397 struct sk_buff *next = tcp_skb_next(skb, list); 5398 5399 if (list) 5400 __skb_unlink(skb, list); 5401 else 5402 rb_erase(&skb->rbnode, root); 5403 5404 __kfree_skb(skb); 5405 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED); 5406 5407 return next; 5408 } 5409 5410 /* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */ 5411 void tcp_rbtree_insert(struct rb_root *root, struct sk_buff *skb) 5412 { 5413 struct rb_node **p = &root->rb_node; 5414 struct rb_node *parent = NULL; 5415 struct sk_buff *skb1; 5416 5417 while (*p) { 5418 parent = *p; 5419 skb1 = rb_to_skb(parent); 5420 if (before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb1)->seq)) 5421 p = &parent->rb_left; 5422 else 5423 p = &parent->rb_right; 5424 } 5425 rb_link_node(&skb->rbnode, parent, p); 5426 rb_insert_color(&skb->rbnode, root); 5427 } 5428 5429 /* Collapse contiguous sequence of skbs head..tail with 5430 * sequence numbers start..end. 5431 * 5432 * If tail is NULL, this means until the end of the queue. 5433 * 5434 * Segments with FIN/SYN are not collapsed (only because this 5435 * simplifies code) 5436 */ 5437 static void 5438 tcp_collapse(struct sock *sk, struct sk_buff_head *list, struct rb_root *root, 5439 struct sk_buff *head, struct sk_buff *tail, u32 start, u32 end) 5440 { 5441 struct sk_buff *skb = head, *n; 5442 struct sk_buff_head tmp; 5443 bool end_of_skbs; 5444 5445 /* First, check that queue is collapsible and find 5446 * the point where collapsing can be useful. 5447 */ 5448 restart: 5449 for (end_of_skbs = true; skb != NULL && skb != tail; skb = n) { 5450 n = tcp_skb_next(skb, list); 5451 5452 if (!skb_frags_readable(skb)) 5453 goto skip_this; 5454 5455 /* No new bits? It is possible on ofo queue. */ 5456 if (!before(start, TCP_SKB_CB(skb)->end_seq)) { 5457 skb = tcp_collapse_one(sk, skb, list, root); 5458 if (!skb) 5459 break; 5460 goto restart; 5461 } 5462 5463 /* The first skb to collapse is: 5464 * - not SYN/FIN and 5465 * - bloated or contains data before "start" or 5466 * overlaps to the next one and mptcp allow collapsing. 5467 */ 5468 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) && 5469 (tcp_win_from_space(sk, skb->truesize) > skb->len || 5470 before(TCP_SKB_CB(skb)->seq, start))) { 5471 end_of_skbs = false; 5472 break; 5473 } 5474 5475 if (n && n != tail && skb_frags_readable(n) && 5476 tcp_skb_can_collapse_rx(skb, n) && 5477 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(n)->seq) { 5478 end_of_skbs = false; 5479 break; 5480 } 5481 5482 skip_this: 5483 /* Decided to skip this, advance start seq. */ 5484 start = TCP_SKB_CB(skb)->end_seq; 5485 } 5486 if (end_of_skbs || 5487 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) || 5488 !skb_frags_readable(skb)) 5489 return; 5490 5491 __skb_queue_head_init(&tmp); 5492 5493 while (before(start, end)) { 5494 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start); 5495 struct sk_buff *nskb; 5496 5497 nskb = alloc_skb(copy, GFP_ATOMIC); 5498 if (!nskb) 5499 break; 5500 5501 memcpy(nskb->cb, skb->cb, sizeof(skb->cb)); 5502 skb_copy_decrypted(nskb, skb); 5503 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start; 5504 if (list) 5505 __skb_queue_before(list, skb, nskb); 5506 else 5507 __skb_queue_tail(&tmp, nskb); /* defer rbtree insertion */ 5508 skb_set_owner_r(nskb, sk); 5509 mptcp_skb_ext_move(nskb, skb); 5510 5511 /* Copy data, releasing collapsed skbs. */ 5512 while (copy > 0) { 5513 int offset = start - TCP_SKB_CB(skb)->seq; 5514 int size = TCP_SKB_CB(skb)->end_seq - start; 5515 5516 BUG_ON(offset < 0); 5517 if (size > 0) { 5518 size = min(copy, size); 5519 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size)) 5520 BUG(); 5521 TCP_SKB_CB(nskb)->end_seq += size; 5522 copy -= size; 5523 start += size; 5524 } 5525 if (!before(start, TCP_SKB_CB(skb)->end_seq)) { 5526 skb = tcp_collapse_one(sk, skb, list, root); 5527 if (!skb || 5528 skb == tail || 5529 !tcp_skb_can_collapse_rx(nskb, skb) || 5530 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) || 5531 !skb_frags_readable(skb)) 5532 goto end; 5533 } 5534 } 5535 } 5536 end: 5537 skb_queue_walk_safe(&tmp, skb, n) 5538 tcp_rbtree_insert(root, skb); 5539 } 5540 5541 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs 5542 * and tcp_collapse() them until all the queue is collapsed. 5543 */ 5544 static void tcp_collapse_ofo_queue(struct sock *sk) 5545 { 5546 struct tcp_sock *tp = tcp_sk(sk); 5547 u32 range_truesize, sum_tiny = 0; 5548 struct sk_buff *skb, *head; 5549 u32 start, end; 5550 5551 skb = skb_rb_first(&tp->out_of_order_queue); 5552 new_range: 5553 if (!skb) { 5554 tp->ooo_last_skb = skb_rb_last(&tp->out_of_order_queue); 5555 return; 5556 } 5557 start = TCP_SKB_CB(skb)->seq; 5558 end = TCP_SKB_CB(skb)->end_seq; 5559 range_truesize = skb->truesize; 5560 5561 for (head = skb;;) { 5562 skb = skb_rb_next(skb); 5563 5564 /* Range is terminated when we see a gap or when 5565 * we are at the queue end. 5566 */ 5567 if (!skb || 5568 after(TCP_SKB_CB(skb)->seq, end) || 5569 before(TCP_SKB_CB(skb)->end_seq, start)) { 5570 /* Do not attempt collapsing tiny skbs */ 5571 if (range_truesize != head->truesize || 5572 end - start >= SKB_WITH_OVERHEAD(PAGE_SIZE)) { 5573 tcp_collapse(sk, NULL, &tp->out_of_order_queue, 5574 head, skb, start, end); 5575 } else { 5576 sum_tiny += range_truesize; 5577 if (sum_tiny > sk->sk_rcvbuf >> 3) 5578 return; 5579 } 5580 goto new_range; 5581 } 5582 5583 range_truesize += skb->truesize; 5584 if (unlikely(before(TCP_SKB_CB(skb)->seq, start))) 5585 start = TCP_SKB_CB(skb)->seq; 5586 if (after(TCP_SKB_CB(skb)->end_seq, end)) 5587 end = TCP_SKB_CB(skb)->end_seq; 5588 } 5589 } 5590 5591 /* 5592 * Clean the out-of-order queue to make room. 5593 * We drop high sequences packets to : 5594 * 1) Let a chance for holes to be filled. 5595 * This means we do not drop packets from ooo queue if their sequence 5596 * is before incoming packet sequence. 5597 * 2) not add too big latencies if thousands of packets sit there. 5598 * (But if application shrinks SO_RCVBUF, we could still end up 5599 * freeing whole queue here) 5600 * 3) Drop at least 12.5 % of sk_rcvbuf to avoid malicious attacks. 5601 * 5602 * Return true if queue has shrunk. 5603 */ 5604 static bool tcp_prune_ofo_queue(struct sock *sk, const struct sk_buff *in_skb) 5605 { 5606 struct tcp_sock *tp = tcp_sk(sk); 5607 struct rb_node *node, *prev; 5608 bool pruned = false; 5609 int goal; 5610 5611 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) 5612 return false; 5613 5614 goal = sk->sk_rcvbuf >> 3; 5615 node = &tp->ooo_last_skb->rbnode; 5616 5617 do { 5618 struct sk_buff *skb = rb_to_skb(node); 5619 5620 /* If incoming skb would land last in ofo queue, stop pruning. */ 5621 if (after(TCP_SKB_CB(in_skb)->seq, TCP_SKB_CB(skb)->seq)) 5622 break; 5623 pruned = true; 5624 prev = rb_prev(node); 5625 rb_erase(node, &tp->out_of_order_queue); 5626 goal -= skb->truesize; 5627 tcp_drop_reason(sk, skb, SKB_DROP_REASON_TCP_OFO_QUEUE_PRUNE); 5628 tp->ooo_last_skb = rb_to_skb(prev); 5629 if (!prev || goal <= 0) { 5630 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf && 5631 !tcp_under_memory_pressure(sk)) 5632 break; 5633 goal = sk->sk_rcvbuf >> 3; 5634 } 5635 node = prev; 5636 } while (node); 5637 5638 if (pruned) { 5639 NET_INC_STATS(sock_net(sk), LINUX_MIB_OFOPRUNED); 5640 /* Reset SACK state. A conforming SACK implementation will 5641 * do the same at a timeout based retransmit. When a connection 5642 * is in a sad state like this, we care only about integrity 5643 * of the connection not performance. 5644 */ 5645 if (tp->rx_opt.sack_ok) 5646 tcp_sack_reset(&tp->rx_opt); 5647 } 5648 return pruned; 5649 } 5650 5651 /* Reduce allocated memory if we can, trying to get 5652 * the socket within its memory limits again. 5653 * 5654 * Return less than zero if we should start dropping frames 5655 * until the socket owning process reads some of the data 5656 * to stabilize the situation. 5657 */ 5658 static int tcp_prune_queue(struct sock *sk, const struct sk_buff *in_skb) 5659 { 5660 struct tcp_sock *tp = tcp_sk(sk); 5661 5662 NET_INC_STATS(sock_net(sk), LINUX_MIB_PRUNECALLED); 5663 5664 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf) 5665 tcp_clamp_window(sk); 5666 else if (tcp_under_memory_pressure(sk)) 5667 tcp_adjust_rcv_ssthresh(sk); 5668 5669 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) 5670 return 0; 5671 5672 tcp_collapse_ofo_queue(sk); 5673 if (!skb_queue_empty(&sk->sk_receive_queue)) 5674 tcp_collapse(sk, &sk->sk_receive_queue, NULL, 5675 skb_peek(&sk->sk_receive_queue), 5676 NULL, 5677 tp->copied_seq, tp->rcv_nxt); 5678 5679 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) 5680 return 0; 5681 5682 /* Collapsing did not help, destructive actions follow. 5683 * This must not ever occur. */ 5684 5685 tcp_prune_ofo_queue(sk, in_skb); 5686 5687 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) 5688 return 0; 5689 5690 /* If we are really being abused, tell the caller to silently 5691 * drop receive data on the floor. It will get retransmitted 5692 * and hopefully then we'll have sufficient space. 5693 */ 5694 NET_INC_STATS(sock_net(sk), LINUX_MIB_RCVPRUNED); 5695 5696 /* Massive buffer overcommit. */ 5697 tp->pred_flags = 0; 5698 return -1; 5699 } 5700 5701 static bool tcp_should_expand_sndbuf(struct sock *sk) 5702 { 5703 const struct tcp_sock *tp = tcp_sk(sk); 5704 5705 /* If the user specified a specific send buffer setting, do 5706 * not modify it. 5707 */ 5708 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK) 5709 return false; 5710 5711 /* If we are under global TCP memory pressure, do not expand. */ 5712 if (tcp_under_memory_pressure(sk)) { 5713 int unused_mem = sk_unused_reserved_mem(sk); 5714 5715 /* Adjust sndbuf according to reserved mem. But make sure 5716 * it never goes below SOCK_MIN_SNDBUF. 5717 * See sk_stream_moderate_sndbuf() for more details. 5718 */ 5719 if (unused_mem > SOCK_MIN_SNDBUF) 5720 WRITE_ONCE(sk->sk_sndbuf, unused_mem); 5721 5722 return false; 5723 } 5724 5725 /* If we are under soft global TCP memory pressure, do not expand. */ 5726 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0)) 5727 return false; 5728 5729 /* If we filled the congestion window, do not expand. */ 5730 if (tcp_packets_in_flight(tp) >= tcp_snd_cwnd(tp)) 5731 return false; 5732 5733 return true; 5734 } 5735 5736 static void tcp_new_space(struct sock *sk) 5737 { 5738 struct tcp_sock *tp = tcp_sk(sk); 5739 5740 if (tcp_should_expand_sndbuf(sk)) { 5741 tcp_sndbuf_expand(sk); 5742 tp->snd_cwnd_stamp = tcp_jiffies32; 5743 } 5744 5745 INDIRECT_CALL_1(sk->sk_write_space, sk_stream_write_space, sk); 5746 } 5747 5748 /* Caller made space either from: 5749 * 1) Freeing skbs in rtx queues (after tp->snd_una has advanced) 5750 * 2) Sent skbs from output queue (and thus advancing tp->snd_nxt) 5751 * 5752 * We might be able to generate EPOLLOUT to the application if: 5753 * 1) Space consumed in output/rtx queues is below sk->sk_sndbuf/2 5754 * 2) notsent amount (tp->write_seq - tp->snd_nxt) became 5755 * small enough that tcp_stream_memory_free() decides it 5756 * is time to generate EPOLLOUT. 5757 */ 5758 void tcp_check_space(struct sock *sk) 5759 { 5760 /* pairs with tcp_poll() */ 5761 smp_mb(); 5762 if (sk->sk_socket && 5763 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { 5764 tcp_new_space(sk); 5765 if (!test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) 5766 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED); 5767 } 5768 } 5769 5770 static inline void tcp_data_snd_check(struct sock *sk) 5771 { 5772 tcp_push_pending_frames(sk); 5773 tcp_check_space(sk); 5774 } 5775 5776 /* 5777 * Check if sending an ack is needed. 5778 */ 5779 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible) 5780 { 5781 struct tcp_sock *tp = tcp_sk(sk); 5782 unsigned long rtt, delay; 5783 5784 /* More than one full frame received... */ 5785 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss && 5786 /* ... and right edge of window advances far enough. 5787 * (tcp_recvmsg() will send ACK otherwise). 5788 * If application uses SO_RCVLOWAT, we want send ack now if 5789 * we have not received enough bytes to satisfy the condition. 5790 */ 5791 (tp->rcv_nxt - tp->copied_seq < sk->sk_rcvlowat || 5792 __tcp_select_window(sk) >= tp->rcv_wnd)) || 5793 /* We ACK each frame or... */ 5794 tcp_in_quickack_mode(sk) || 5795 /* Protocol state mandates a one-time immediate ACK */ 5796 inet_csk(sk)->icsk_ack.pending & ICSK_ACK_NOW) { 5797 /* If we are running from __release_sock() in user context, 5798 * Defer the ack until tcp_release_cb(). 5799 */ 5800 if (sock_owned_by_user_nocheck(sk) && 5801 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_backlog_ack_defer)) { 5802 set_bit(TCP_ACK_DEFERRED, &sk->sk_tsq_flags); 5803 return; 5804 } 5805 send_now: 5806 tcp_send_ack(sk); 5807 return; 5808 } 5809 5810 if (!ofo_possible || RB_EMPTY_ROOT(&tp->out_of_order_queue)) { 5811 tcp_send_delayed_ack(sk); 5812 return; 5813 } 5814 5815 if (!tcp_is_sack(tp) || 5816 tp->compressed_ack >= READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_nr)) 5817 goto send_now; 5818 5819 if (tp->compressed_ack_rcv_nxt != tp->rcv_nxt) { 5820 tp->compressed_ack_rcv_nxt = tp->rcv_nxt; 5821 tp->dup_ack_counter = 0; 5822 } 5823 if (tp->dup_ack_counter < TCP_FASTRETRANS_THRESH) { 5824 tp->dup_ack_counter++; 5825 goto send_now; 5826 } 5827 tp->compressed_ack++; 5828 if (hrtimer_is_queued(&tp->compressed_ack_timer)) 5829 return; 5830 5831 /* compress ack timer : 5 % of rtt, but no more than tcp_comp_sack_delay_ns */ 5832 5833 rtt = tp->rcv_rtt_est.rtt_us; 5834 if (tp->srtt_us && tp->srtt_us < rtt) 5835 rtt = tp->srtt_us; 5836 5837 delay = min_t(unsigned long, 5838 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_delay_ns), 5839 rtt * (NSEC_PER_USEC >> 3)/20); 5840 sock_hold(sk); 5841 hrtimer_start_range_ns(&tp->compressed_ack_timer, ns_to_ktime(delay), 5842 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_slack_ns), 5843 HRTIMER_MODE_REL_PINNED_SOFT); 5844 } 5845 5846 static inline void tcp_ack_snd_check(struct sock *sk) 5847 { 5848 if (!inet_csk_ack_scheduled(sk)) { 5849 /* We sent a data segment already. */ 5850 return; 5851 } 5852 __tcp_ack_snd_check(sk, 1); 5853 } 5854 5855 /* 5856 * This routine is only called when we have urgent data 5857 * signaled. Its the 'slow' part of tcp_urg. It could be 5858 * moved inline now as tcp_urg is only called from one 5859 * place. We handle URGent data wrong. We have to - as 5860 * BSD still doesn't use the correction from RFC961. 5861 * For 1003.1g we should support a new option TCP_STDURG to permit 5862 * either form (or just set the sysctl tcp_stdurg). 5863 */ 5864 5865 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th) 5866 { 5867 struct tcp_sock *tp = tcp_sk(sk); 5868 u32 ptr = ntohs(th->urg_ptr); 5869 5870 if (ptr && !READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_stdurg)) 5871 ptr--; 5872 ptr += ntohl(th->seq); 5873 5874 /* Ignore urgent data that we've already seen and read. */ 5875 if (after(tp->copied_seq, ptr)) 5876 return; 5877 5878 /* Do not replay urg ptr. 5879 * 5880 * NOTE: interesting situation not covered by specs. 5881 * Misbehaving sender may send urg ptr, pointing to segment, 5882 * which we already have in ofo queue. We are not able to fetch 5883 * such data and will stay in TCP_URG_NOTYET until will be eaten 5884 * by recvmsg(). Seems, we are not obliged to handle such wicked 5885 * situations. But it is worth to think about possibility of some 5886 * DoSes using some hypothetical application level deadlock. 5887 */ 5888 if (before(ptr, tp->rcv_nxt)) 5889 return; 5890 5891 /* Do we already have a newer (or duplicate) urgent pointer? */ 5892 if (tp->urg_data && !after(ptr, tp->urg_seq)) 5893 return; 5894 5895 /* Tell the world about our new urgent pointer. */ 5896 sk_send_sigurg(sk); 5897 5898 /* We may be adding urgent data when the last byte read was 5899 * urgent. To do this requires some care. We cannot just ignore 5900 * tp->copied_seq since we would read the last urgent byte again 5901 * as data, nor can we alter copied_seq until this data arrives 5902 * or we break the semantics of SIOCATMARK (and thus sockatmark()) 5903 * 5904 * NOTE. Double Dutch. Rendering to plain English: author of comment 5905 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB); 5906 * and expect that both A and B disappear from stream. This is _wrong_. 5907 * Though this happens in BSD with high probability, this is occasional. 5908 * Any application relying on this is buggy. Note also, that fix "works" 5909 * only in this artificial test. Insert some normal data between A and B and we will 5910 * decline of BSD again. Verdict: it is better to remove to trap 5911 * buggy users. 5912 */ 5913 if (tp->urg_seq == tp->copied_seq && tp->urg_data && 5914 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) { 5915 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue); 5916 tp->copied_seq++; 5917 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) { 5918 __skb_unlink(skb, &sk->sk_receive_queue); 5919 __kfree_skb(skb); 5920 } 5921 } 5922 5923 WRITE_ONCE(tp->urg_data, TCP_URG_NOTYET); 5924 WRITE_ONCE(tp->urg_seq, ptr); 5925 5926 /* Disable header prediction. */ 5927 tp->pred_flags = 0; 5928 } 5929 5930 /* This is the 'fast' part of urgent handling. */ 5931 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th) 5932 { 5933 struct tcp_sock *tp = tcp_sk(sk); 5934 5935 /* Check if we get a new urgent pointer - normally not. */ 5936 if (unlikely(th->urg)) 5937 tcp_check_urg(sk, th); 5938 5939 /* Do we wait for any urgent data? - normally not... */ 5940 if (unlikely(tp->urg_data == TCP_URG_NOTYET)) { 5941 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) - 5942 th->syn; 5943 5944 /* Is the urgent pointer pointing into this packet? */ 5945 if (ptr < skb->len) { 5946 u8 tmp; 5947 if (skb_copy_bits(skb, ptr, &tmp, 1)) 5948 BUG(); 5949 WRITE_ONCE(tp->urg_data, TCP_URG_VALID | tmp); 5950 if (!sock_flag(sk, SOCK_DEAD)) 5951 sk->sk_data_ready(sk); 5952 } 5953 } 5954 } 5955 5956 /* Accept RST for rcv_nxt - 1 after a FIN. 5957 * When tcp connections are abruptly terminated from Mac OSX (via ^C), a 5958 * FIN is sent followed by a RST packet. The RST is sent with the same 5959 * sequence number as the FIN, and thus according to RFC 5961 a challenge 5960 * ACK should be sent. However, Mac OSX rate limits replies to challenge 5961 * ACKs on the closed socket. In addition middleboxes can drop either the 5962 * challenge ACK or a subsequent RST. 5963 */ 5964 static bool tcp_reset_check(const struct sock *sk, const struct sk_buff *skb) 5965 { 5966 const struct tcp_sock *tp = tcp_sk(sk); 5967 5968 return unlikely(TCP_SKB_CB(skb)->seq == (tp->rcv_nxt - 1) && 5969 (1 << sk->sk_state) & (TCPF_CLOSE_WAIT | TCPF_LAST_ACK | 5970 TCPF_CLOSING)); 5971 } 5972 5973 /* Does PAWS and seqno based validation of an incoming segment, flags will 5974 * play significant role here. 5975 */ 5976 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb, 5977 const struct tcphdr *th, int syn_inerr) 5978 { 5979 struct tcp_sock *tp = tcp_sk(sk); 5980 SKB_DR(reason); 5981 5982 /* RFC1323: H1. Apply PAWS check first. */ 5983 if (!tcp_fast_parse_options(sock_net(sk), skb, th, tp) || 5984 !tp->rx_opt.saw_tstamp || 5985 tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW)) 5986 goto step1; 5987 5988 reason = tcp_disordered_ack_check(sk, skb); 5989 if (!reason) 5990 goto step1; 5991 /* Reset is accepted even if it did not pass PAWS. */ 5992 if (th->rst) 5993 goto step1; 5994 if (unlikely(th->syn)) 5995 goto syn_challenge; 5996 5997 /* Old ACK are common, increment PAWS_OLD_ACK 5998 * and do not send a dupack. 5999 */ 6000 if (reason == SKB_DROP_REASON_TCP_RFC7323_PAWS_ACK) { 6001 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWS_OLD_ACK); 6002 goto discard; 6003 } 6004 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED); 6005 if (!tcp_oow_rate_limited(sock_net(sk), skb, 6006 LINUX_MIB_TCPACKSKIPPEDPAWS, 6007 &tp->last_oow_ack_time)) 6008 tcp_send_dupack(sk, skb); 6009 goto discard; 6010 6011 step1: 6012 /* Step 1: check sequence number */ 6013 reason = tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); 6014 if (reason) { 6015 /* RFC793, page 37: "In all states except SYN-SENT, all reset 6016 * (RST) segments are validated by checking their SEQ-fields." 6017 * And page 69: "If an incoming segment is not acceptable, 6018 * an acknowledgment should be sent in reply (unless the RST 6019 * bit is set, if so drop the segment and return)". 6020 */ 6021 if (!th->rst) { 6022 if (th->syn) 6023 goto syn_challenge; 6024 if (!tcp_oow_rate_limited(sock_net(sk), skb, 6025 LINUX_MIB_TCPACKSKIPPEDSEQ, 6026 &tp->last_oow_ack_time)) 6027 tcp_send_dupack(sk, skb); 6028 } else if (tcp_reset_check(sk, skb)) { 6029 goto reset; 6030 } 6031 goto discard; 6032 } 6033 6034 /* Step 2: check RST bit */ 6035 if (th->rst) { 6036 /* RFC 5961 3.2 (extend to match against (RCV.NXT - 1) after a 6037 * FIN and SACK too if available): 6038 * If seq num matches RCV.NXT or (RCV.NXT - 1) after a FIN, or 6039 * the right-most SACK block, 6040 * then 6041 * RESET the connection 6042 * else 6043 * Send a challenge ACK 6044 */ 6045 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt || 6046 tcp_reset_check(sk, skb)) 6047 goto reset; 6048 6049 if (tcp_is_sack(tp) && tp->rx_opt.num_sacks > 0) { 6050 struct tcp_sack_block *sp = &tp->selective_acks[0]; 6051 int max_sack = sp[0].end_seq; 6052 int this_sack; 6053 6054 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks; 6055 ++this_sack) { 6056 max_sack = after(sp[this_sack].end_seq, 6057 max_sack) ? 6058 sp[this_sack].end_seq : max_sack; 6059 } 6060 6061 if (TCP_SKB_CB(skb)->seq == max_sack) 6062 goto reset; 6063 } 6064 6065 /* Disable TFO if RST is out-of-order 6066 * and no data has been received 6067 * for current active TFO socket 6068 */ 6069 if (tp->syn_fastopen && !tp->data_segs_in && 6070 sk->sk_state == TCP_ESTABLISHED) 6071 tcp_fastopen_active_disable(sk); 6072 tcp_send_challenge_ack(sk); 6073 SKB_DR_SET(reason, TCP_RESET); 6074 goto discard; 6075 } 6076 6077 /* step 3: check security and precedence [ignored] */ 6078 6079 /* step 4: Check for a SYN 6080 * RFC 5961 4.2 : Send a challenge ack 6081 */ 6082 if (th->syn) { 6083 if (sk->sk_state == TCP_SYN_RECV && sk->sk_socket && th->ack && 6084 TCP_SKB_CB(skb)->seq + 1 == TCP_SKB_CB(skb)->end_seq && 6085 TCP_SKB_CB(skb)->seq + 1 == tp->rcv_nxt && 6086 TCP_SKB_CB(skb)->ack_seq == tp->snd_nxt) 6087 goto pass; 6088 syn_challenge: 6089 if (syn_inerr) 6090 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS); 6091 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE); 6092 tcp_send_challenge_ack(sk); 6093 SKB_DR_SET(reason, TCP_INVALID_SYN); 6094 goto discard; 6095 } 6096 6097 pass: 6098 bpf_skops_parse_hdr(sk, skb); 6099 6100 return true; 6101 6102 discard: 6103 tcp_drop_reason(sk, skb, reason); 6104 return false; 6105 6106 reset: 6107 tcp_reset(sk, skb); 6108 __kfree_skb(skb); 6109 return false; 6110 } 6111 6112 /* 6113 * TCP receive function for the ESTABLISHED state. 6114 * 6115 * It is split into a fast path and a slow path. The fast path is 6116 * disabled when: 6117 * - A zero window was announced from us - zero window probing 6118 * is only handled properly in the slow path. 6119 * - Out of order segments arrived. 6120 * - Urgent data is expected. 6121 * - There is no buffer space left 6122 * - Unexpected TCP flags/window values/header lengths are received 6123 * (detected by checking the TCP header against pred_flags) 6124 * - Data is sent in both directions. Fast path only supports pure senders 6125 * or pure receivers (this means either the sequence number or the ack 6126 * value must stay constant) 6127 * - Unexpected TCP option. 6128 * 6129 * When these conditions are not satisfied it drops into a standard 6130 * receive procedure patterned after RFC793 to handle all cases. 6131 * The first three cases are guaranteed by proper pred_flags setting, 6132 * the rest is checked inline. Fast processing is turned on in 6133 * tcp_data_queue when everything is OK. 6134 */ 6135 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb) 6136 { 6137 enum skb_drop_reason reason = SKB_DROP_REASON_NOT_SPECIFIED; 6138 const struct tcphdr *th = (const struct tcphdr *)skb->data; 6139 struct tcp_sock *tp = tcp_sk(sk); 6140 unsigned int len = skb->len; 6141 6142 /* TCP congestion window tracking */ 6143 trace_tcp_probe(sk, skb); 6144 6145 tcp_mstamp_refresh(tp); 6146 if (unlikely(!rcu_access_pointer(sk->sk_rx_dst))) 6147 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb); 6148 /* 6149 * Header prediction. 6150 * The code loosely follows the one in the famous 6151 * "30 instruction TCP receive" Van Jacobson mail. 6152 * 6153 * Van's trick is to deposit buffers into socket queue 6154 * on a device interrupt, to call tcp_recv function 6155 * on the receive process context and checksum and copy 6156 * the buffer to user space. smart... 6157 * 6158 * Our current scheme is not silly either but we take the 6159 * extra cost of the net_bh soft interrupt processing... 6160 * We do checksum and copy also but from device to kernel. 6161 */ 6162 6163 tp->rx_opt.saw_tstamp = 0; 6164 6165 /* pred_flags is 0xS?10 << 16 + snd_wnd 6166 * if header_prediction is to be made 6167 * 'S' will always be tp->tcp_header_len >> 2 6168 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to 6169 * turn it off (when there are holes in the receive 6170 * space for instance) 6171 * PSH flag is ignored. 6172 */ 6173 6174 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags && 6175 TCP_SKB_CB(skb)->seq == tp->rcv_nxt && 6176 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) { 6177 int tcp_header_len = tp->tcp_header_len; 6178 s32 delta = 0; 6179 int flag = 0; 6180 6181 /* Timestamp header prediction: tcp_header_len 6182 * is automatically equal to th->doff*4 due to pred_flags 6183 * match. 6184 */ 6185 6186 /* Check timestamp */ 6187 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) { 6188 /* No? Slow path! */ 6189 if (!tcp_parse_aligned_timestamp(tp, th)) 6190 goto slow_path; 6191 6192 delta = tp->rx_opt.rcv_tsval - 6193 tp->rx_opt.ts_recent; 6194 /* If PAWS failed, check it more carefully in slow path */ 6195 if (delta < 0) 6196 goto slow_path; 6197 6198 /* DO NOT update ts_recent here, if checksum fails 6199 * and timestamp was corrupted part, it will result 6200 * in a hung connection since we will drop all 6201 * future packets due to the PAWS test. 6202 */ 6203 } 6204 6205 if (len <= tcp_header_len) { 6206 /* Bulk data transfer: sender */ 6207 if (len == tcp_header_len) { 6208 /* Predicted packet is in window by definition. 6209 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 6210 * Hence, check seq<=rcv_wup reduces to: 6211 */ 6212 if (tcp_header_len == 6213 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && 6214 tp->rcv_nxt == tp->rcv_wup) 6215 flag |= __tcp_replace_ts_recent(tp, 6216 delta); 6217 6218 /* We know that such packets are checksummed 6219 * on entry. 6220 */ 6221 tcp_ack(sk, skb, flag); 6222 __kfree_skb(skb); 6223 tcp_data_snd_check(sk); 6224 /* When receiving pure ack in fast path, update 6225 * last ts ecr directly instead of calling 6226 * tcp_rcv_rtt_measure_ts() 6227 */ 6228 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr; 6229 return; 6230 } else { /* Header too small */ 6231 reason = SKB_DROP_REASON_PKT_TOO_SMALL; 6232 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS); 6233 goto discard; 6234 } 6235 } else { 6236 int eaten = 0; 6237 bool fragstolen = false; 6238 6239 if (tcp_checksum_complete(skb)) 6240 goto csum_error; 6241 6242 if ((int)skb->truesize > sk->sk_forward_alloc) 6243 goto step5; 6244 6245 /* Predicted packet is in window by definition. 6246 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 6247 * Hence, check seq<=rcv_wup reduces to: 6248 */ 6249 if (tcp_header_len == 6250 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && 6251 tp->rcv_nxt == tp->rcv_wup) 6252 flag |= __tcp_replace_ts_recent(tp, 6253 delta); 6254 6255 tcp_rcv_rtt_measure_ts(sk, skb); 6256 6257 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPHITS); 6258 6259 /* Bulk data transfer: receiver */ 6260 tcp_cleanup_skb(skb); 6261 __skb_pull(skb, tcp_header_len); 6262 eaten = tcp_queue_rcv(sk, skb, &fragstolen); 6263 6264 tcp_event_data_recv(sk, skb); 6265 6266 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) { 6267 /* Well, only one small jumplet in fast path... */ 6268 tcp_ack(sk, skb, flag | FLAG_DATA); 6269 tcp_data_snd_check(sk); 6270 if (!inet_csk_ack_scheduled(sk)) 6271 goto no_ack; 6272 } else { 6273 tcp_update_wl(tp, TCP_SKB_CB(skb)->seq); 6274 } 6275 6276 __tcp_ack_snd_check(sk, 0); 6277 no_ack: 6278 if (eaten) 6279 kfree_skb_partial(skb, fragstolen); 6280 tcp_data_ready(sk); 6281 return; 6282 } 6283 } 6284 6285 slow_path: 6286 if (len < (th->doff << 2) || tcp_checksum_complete(skb)) 6287 goto csum_error; 6288 6289 if (!th->ack && !th->rst && !th->syn) { 6290 reason = SKB_DROP_REASON_TCP_FLAGS; 6291 goto discard; 6292 } 6293 6294 /* 6295 * Standard slow path. 6296 */ 6297 6298 if (!tcp_validate_incoming(sk, skb, th, 1)) 6299 return; 6300 6301 step5: 6302 reason = tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT); 6303 if ((int)reason < 0) { 6304 reason = -reason; 6305 goto discard; 6306 } 6307 tcp_rcv_rtt_measure_ts(sk, skb); 6308 6309 /* Process urgent data. */ 6310 tcp_urg(sk, skb, th); 6311 6312 /* step 7: process the segment text */ 6313 tcp_data_queue(sk, skb); 6314 6315 tcp_data_snd_check(sk); 6316 tcp_ack_snd_check(sk); 6317 return; 6318 6319 csum_error: 6320 reason = SKB_DROP_REASON_TCP_CSUM; 6321 trace_tcp_bad_csum(skb); 6322 TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS); 6323 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS); 6324 6325 discard: 6326 tcp_drop_reason(sk, skb, reason); 6327 } 6328 EXPORT_IPV6_MOD(tcp_rcv_established); 6329 6330 void tcp_init_transfer(struct sock *sk, int bpf_op, struct sk_buff *skb) 6331 { 6332 struct inet_connection_sock *icsk = inet_csk(sk); 6333 struct tcp_sock *tp = tcp_sk(sk); 6334 6335 tcp_mtup_init(sk); 6336 icsk->icsk_af_ops->rebuild_header(sk); 6337 tcp_init_metrics(sk); 6338 6339 /* Initialize the congestion window to start the transfer. 6340 * Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been 6341 * retransmitted. In light of RFC6298 more aggressive 1sec 6342 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK 6343 * retransmission has occurred. 6344 */ 6345 if (tp->total_retrans > 1 && tp->undo_marker) 6346 tcp_snd_cwnd_set(tp, 1); 6347 else 6348 tcp_snd_cwnd_set(tp, tcp_init_cwnd(tp, __sk_dst_get(sk))); 6349 tp->snd_cwnd_stamp = tcp_jiffies32; 6350 6351 bpf_skops_established(sk, bpf_op, skb); 6352 /* Initialize congestion control unless BPF initialized it already: */ 6353 if (!icsk->icsk_ca_initialized) 6354 tcp_init_congestion_control(sk); 6355 tcp_init_buffer_space(sk); 6356 } 6357 6358 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb) 6359 { 6360 struct tcp_sock *tp = tcp_sk(sk); 6361 struct inet_connection_sock *icsk = inet_csk(sk); 6362 6363 tcp_ao_finish_connect(sk, skb); 6364 tcp_set_state(sk, TCP_ESTABLISHED); 6365 icsk->icsk_ack.lrcvtime = tcp_jiffies32; 6366 6367 if (skb) { 6368 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb); 6369 security_inet_conn_established(sk, skb); 6370 sk_mark_napi_id(sk, skb); 6371 } 6372 6373 tcp_init_transfer(sk, BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB, skb); 6374 6375 /* Prevent spurious tcp_cwnd_restart() on first data 6376 * packet. 6377 */ 6378 tp->lsndtime = tcp_jiffies32; 6379 6380 if (sock_flag(sk, SOCK_KEEPOPEN)) 6381 tcp_reset_keepalive_timer(sk, keepalive_time_when(tp)); 6382 6383 if (!tp->rx_opt.snd_wscale) 6384 __tcp_fast_path_on(tp, tp->snd_wnd); 6385 else 6386 tp->pred_flags = 0; 6387 } 6388 6389 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack, 6390 struct tcp_fastopen_cookie *cookie) 6391 { 6392 struct tcp_sock *tp = tcp_sk(sk); 6393 struct sk_buff *data = tp->syn_data ? tcp_rtx_queue_head(sk) : NULL; 6394 u16 mss = tp->rx_opt.mss_clamp, try_exp = 0; 6395 bool syn_drop = false; 6396 6397 if (mss == tp->rx_opt.user_mss) { 6398 struct tcp_options_received opt; 6399 6400 /* Get original SYNACK MSS value if user MSS sets mss_clamp */ 6401 tcp_clear_options(&opt); 6402 opt.user_mss = opt.mss_clamp = 0; 6403 tcp_parse_options(sock_net(sk), synack, &opt, 0, NULL); 6404 mss = opt.mss_clamp; 6405 } 6406 6407 if (!tp->syn_fastopen) { 6408 /* Ignore an unsolicited cookie */ 6409 cookie->len = -1; 6410 } else if (tp->total_retrans) { 6411 /* SYN timed out and the SYN-ACK neither has a cookie nor 6412 * acknowledges data. Presumably the remote received only 6413 * the retransmitted (regular) SYNs: either the original 6414 * SYN-data or the corresponding SYN-ACK was dropped. 6415 */ 6416 syn_drop = (cookie->len < 0 && data); 6417 } else if (cookie->len < 0 && !tp->syn_data) { 6418 /* We requested a cookie but didn't get it. If we did not use 6419 * the (old) exp opt format then try so next time (try_exp=1). 6420 * Otherwise we go back to use the RFC7413 opt (try_exp=2). 6421 */ 6422 try_exp = tp->syn_fastopen_exp ? 2 : 1; 6423 } 6424 6425 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp); 6426 6427 if (data) { /* Retransmit unacked data in SYN */ 6428 if (tp->total_retrans) 6429 tp->fastopen_client_fail = TFO_SYN_RETRANSMITTED; 6430 else 6431 tp->fastopen_client_fail = TFO_DATA_NOT_ACKED; 6432 skb_rbtree_walk_from(data) 6433 tcp_mark_skb_lost(sk, data); 6434 tcp_non_congestion_loss_retransmit(sk); 6435 NET_INC_STATS(sock_net(sk), 6436 LINUX_MIB_TCPFASTOPENACTIVEFAIL); 6437 return true; 6438 } 6439 tp->syn_data_acked = tp->syn_data; 6440 if (tp->syn_data_acked) { 6441 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE); 6442 /* SYN-data is counted as two separate packets in tcp_ack() */ 6443 if (tp->delivered > 1) 6444 --tp->delivered; 6445 } 6446 6447 tcp_fastopen_add_skb(sk, synack); 6448 6449 return false; 6450 } 6451 6452 static void smc_check_reset_syn(struct tcp_sock *tp) 6453 { 6454 #if IS_ENABLED(CONFIG_SMC) 6455 if (static_branch_unlikely(&tcp_have_smc)) { 6456 if (tp->syn_smc && !tp->rx_opt.smc_ok) 6457 tp->syn_smc = 0; 6458 } 6459 #endif 6460 } 6461 6462 static void tcp_try_undo_spurious_syn(struct sock *sk) 6463 { 6464 struct tcp_sock *tp = tcp_sk(sk); 6465 u32 syn_stamp; 6466 6467 /* undo_marker is set when SYN or SYNACK times out. The timeout is 6468 * spurious if the ACK's timestamp option echo value matches the 6469 * original SYN timestamp. 6470 */ 6471 syn_stamp = tp->retrans_stamp; 6472 if (tp->undo_marker && syn_stamp && tp->rx_opt.saw_tstamp && 6473 syn_stamp == tp->rx_opt.rcv_tsecr) 6474 tp->undo_marker = 0; 6475 } 6476 6477 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb, 6478 const struct tcphdr *th) 6479 { 6480 struct inet_connection_sock *icsk = inet_csk(sk); 6481 struct tcp_sock *tp = tcp_sk(sk); 6482 struct tcp_fastopen_cookie foc = { .len = -1 }; 6483 int saved_clamp = tp->rx_opt.mss_clamp; 6484 bool fastopen_fail; 6485 SKB_DR(reason); 6486 6487 tcp_parse_options(sock_net(sk), skb, &tp->rx_opt, 0, &foc); 6488 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr) 6489 tp->rx_opt.rcv_tsecr -= tp->tsoffset; 6490 6491 if (th->ack) { 6492 /* rfc793: 6493 * "If the state is SYN-SENT then 6494 * first check the ACK bit 6495 * If the ACK bit is set 6496 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send 6497 * a reset (unless the RST bit is set, if so drop 6498 * the segment and return)" 6499 */ 6500 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) || 6501 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) { 6502 /* Previous FIN/ACK or RST/ACK might be ignored. */ 6503 if (icsk->icsk_retransmits == 0) 6504 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 6505 TCP_TIMEOUT_MIN, false); 6506 SKB_DR_SET(reason, TCP_INVALID_ACK_SEQUENCE); 6507 goto reset_and_undo; 6508 } 6509 6510 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 6511 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp, 6512 tcp_time_stamp_ts(tp))) { 6513 NET_INC_STATS(sock_net(sk), 6514 LINUX_MIB_PAWSACTIVEREJECTED); 6515 SKB_DR_SET(reason, TCP_RFC7323_PAWS); 6516 goto reset_and_undo; 6517 } 6518 6519 /* Now ACK is acceptable. 6520 * 6521 * "If the RST bit is set 6522 * If the ACK was acceptable then signal the user "error: 6523 * connection reset", drop the segment, enter CLOSED state, 6524 * delete TCB, and return." 6525 */ 6526 6527 if (th->rst) { 6528 tcp_reset(sk, skb); 6529 consume: 6530 __kfree_skb(skb); 6531 return 0; 6532 } 6533 6534 /* rfc793: 6535 * "fifth, if neither of the SYN or RST bits is set then 6536 * drop the segment and return." 6537 * 6538 * See note below! 6539 * --ANK(990513) 6540 */ 6541 if (!th->syn) { 6542 SKB_DR_SET(reason, TCP_FLAGS); 6543 goto discard_and_undo; 6544 } 6545 /* rfc793: 6546 * "If the SYN bit is on ... 6547 * are acceptable then ... 6548 * (our SYN has been ACKed), change the connection 6549 * state to ESTABLISHED..." 6550 */ 6551 6552 tcp_ecn_rcv_synack(tp, th); 6553 6554 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq); 6555 tcp_try_undo_spurious_syn(sk); 6556 tcp_ack(sk, skb, FLAG_SLOWPATH); 6557 6558 /* Ok.. it's good. Set up sequence numbers and 6559 * move to established. 6560 */ 6561 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1); 6562 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; 6563 6564 /* RFC1323: The window in SYN & SYN/ACK segments is 6565 * never scaled. 6566 */ 6567 tp->snd_wnd = ntohs(th->window); 6568 6569 if (!tp->rx_opt.wscale_ok) { 6570 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0; 6571 WRITE_ONCE(tp->window_clamp, 6572 min(tp->window_clamp, 65535U)); 6573 } 6574 6575 if (tp->rx_opt.saw_tstamp) { 6576 tp->rx_opt.tstamp_ok = 1; 6577 tp->tcp_header_len = 6578 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 6579 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; 6580 tcp_store_ts_recent(tp); 6581 } else { 6582 tp->tcp_header_len = sizeof(struct tcphdr); 6583 } 6584 6585 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 6586 tcp_initialize_rcv_mss(sk); 6587 6588 /* Remember, tcp_poll() does not lock socket! 6589 * Change state from SYN-SENT only after copied_seq 6590 * is initialized. */ 6591 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt); 6592 6593 smc_check_reset_syn(tp); 6594 6595 smp_mb(); 6596 6597 tcp_finish_connect(sk, skb); 6598 6599 fastopen_fail = (tp->syn_fastopen || tp->syn_data) && 6600 tcp_rcv_fastopen_synack(sk, skb, &foc); 6601 6602 if (!sock_flag(sk, SOCK_DEAD)) { 6603 sk->sk_state_change(sk); 6604 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT); 6605 } 6606 if (fastopen_fail) 6607 return -1; 6608 if (sk->sk_write_pending || 6609 READ_ONCE(icsk->icsk_accept_queue.rskq_defer_accept) || 6610 inet_csk_in_pingpong_mode(sk)) { 6611 /* Save one ACK. Data will be ready after 6612 * several ticks, if write_pending is set. 6613 * 6614 * It may be deleted, but with this feature tcpdumps 6615 * look so _wonderfully_ clever, that I was not able 6616 * to stand against the temptation 8) --ANK 6617 */ 6618 inet_csk_schedule_ack(sk); 6619 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS); 6620 tcp_reset_xmit_timer(sk, ICSK_TIME_DACK, 6621 TCP_DELACK_MAX, false); 6622 goto consume; 6623 } 6624 tcp_send_ack(sk); 6625 return -1; 6626 } 6627 6628 /* No ACK in the segment */ 6629 6630 if (th->rst) { 6631 /* rfc793: 6632 * "If the RST bit is set 6633 * 6634 * Otherwise (no ACK) drop the segment and return." 6635 */ 6636 SKB_DR_SET(reason, TCP_RESET); 6637 goto discard_and_undo; 6638 } 6639 6640 /* PAWS check. */ 6641 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && 6642 tcp_paws_reject(&tp->rx_opt, 0)) { 6643 SKB_DR_SET(reason, TCP_RFC7323_PAWS); 6644 goto discard_and_undo; 6645 } 6646 if (th->syn) { 6647 /* We see SYN without ACK. It is attempt of 6648 * simultaneous connect with crossed SYNs. 6649 * Particularly, it can be connect to self. 6650 */ 6651 #ifdef CONFIG_TCP_AO 6652 struct tcp_ao_info *ao; 6653 6654 ao = rcu_dereference_protected(tp->ao_info, 6655 lockdep_sock_is_held(sk)); 6656 if (ao) { 6657 WRITE_ONCE(ao->risn, th->seq); 6658 ao->rcv_sne = 0; 6659 } 6660 #endif 6661 tcp_set_state(sk, TCP_SYN_RECV); 6662 6663 if (tp->rx_opt.saw_tstamp) { 6664 tp->rx_opt.tstamp_ok = 1; 6665 tcp_store_ts_recent(tp); 6666 tp->tcp_header_len = 6667 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 6668 } else { 6669 tp->tcp_header_len = sizeof(struct tcphdr); 6670 } 6671 6672 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1); 6673 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt); 6674 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; 6675 6676 /* RFC1323: The window in SYN & SYN/ACK segments is 6677 * never scaled. 6678 */ 6679 tp->snd_wnd = ntohs(th->window); 6680 tp->snd_wl1 = TCP_SKB_CB(skb)->seq; 6681 tp->max_window = tp->snd_wnd; 6682 6683 tcp_ecn_rcv_syn(tp, th); 6684 6685 tcp_mtup_init(sk); 6686 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 6687 tcp_initialize_rcv_mss(sk); 6688 6689 tcp_send_synack(sk); 6690 #if 0 6691 /* Note, we could accept data and URG from this segment. 6692 * There are no obstacles to make this (except that we must 6693 * either change tcp_recvmsg() to prevent it from returning data 6694 * before 3WHS completes per RFC793, or employ TCP Fast Open). 6695 * 6696 * However, if we ignore data in ACKless segments sometimes, 6697 * we have no reasons to accept it sometimes. 6698 * Also, seems the code doing it in step6 of tcp_rcv_state_process 6699 * is not flawless. So, discard packet for sanity. 6700 * Uncomment this return to process the data. 6701 */ 6702 return -1; 6703 #else 6704 goto consume; 6705 #endif 6706 } 6707 /* "fifth, if neither of the SYN or RST bits is set then 6708 * drop the segment and return." 6709 */ 6710 6711 discard_and_undo: 6712 tcp_clear_options(&tp->rx_opt); 6713 tp->rx_opt.mss_clamp = saved_clamp; 6714 tcp_drop_reason(sk, skb, reason); 6715 return 0; 6716 6717 reset_and_undo: 6718 tcp_clear_options(&tp->rx_opt); 6719 tp->rx_opt.mss_clamp = saved_clamp; 6720 /* we can reuse/return @reason to its caller to handle the exception */ 6721 return reason; 6722 } 6723 6724 static void tcp_rcv_synrecv_state_fastopen(struct sock *sk) 6725 { 6726 struct tcp_sock *tp = tcp_sk(sk); 6727 struct request_sock *req; 6728 6729 /* If we are still handling the SYNACK RTO, see if timestamp ECR allows 6730 * undo. If peer SACKs triggered fast recovery, we can't undo here. 6731 */ 6732 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss && !tp->packets_out) 6733 tcp_try_undo_recovery(sk); 6734 6735 tcp_update_rto_time(tp); 6736 inet_csk(sk)->icsk_retransmits = 0; 6737 /* In tcp_fastopen_synack_timer() on the first SYNACK RTO we set 6738 * retrans_stamp but don't enter CA_Loss, so in case that happened we 6739 * need to zero retrans_stamp here to prevent spurious 6740 * retransmits_timed_out(). However, if the ACK of our SYNACK caused us 6741 * to enter CA_Recovery then we need to leave retrans_stamp as it was 6742 * set entering CA_Recovery, for correct retransmits_timed_out() and 6743 * undo behavior. 6744 */ 6745 tcp_retrans_stamp_cleanup(sk); 6746 6747 /* Once we leave TCP_SYN_RECV or TCP_FIN_WAIT_1, 6748 * we no longer need req so release it. 6749 */ 6750 req = rcu_dereference_protected(tp->fastopen_rsk, 6751 lockdep_sock_is_held(sk)); 6752 reqsk_fastopen_remove(sk, req, false); 6753 6754 /* Re-arm the timer because data may have been sent out. 6755 * This is similar to the regular data transmission case 6756 * when new data has just been ack'ed. 6757 * 6758 * (TFO) - we could try to be more aggressive and 6759 * retransmitting any data sooner based on when they 6760 * are sent out. 6761 */ 6762 tcp_rearm_rto(sk); 6763 } 6764 6765 /* 6766 * This function implements the receiving procedure of RFC 793 for 6767 * all states except ESTABLISHED and TIME_WAIT. 6768 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be 6769 * address independent. 6770 */ 6771 6772 enum skb_drop_reason 6773 tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb) 6774 { 6775 struct tcp_sock *tp = tcp_sk(sk); 6776 struct inet_connection_sock *icsk = inet_csk(sk); 6777 const struct tcphdr *th = tcp_hdr(skb); 6778 struct request_sock *req; 6779 int queued = 0; 6780 SKB_DR(reason); 6781 6782 switch (sk->sk_state) { 6783 case TCP_CLOSE: 6784 SKB_DR_SET(reason, TCP_CLOSE); 6785 goto discard; 6786 6787 case TCP_LISTEN: 6788 if (th->ack) 6789 return SKB_DROP_REASON_TCP_FLAGS; 6790 6791 if (th->rst) { 6792 SKB_DR_SET(reason, TCP_RESET); 6793 goto discard; 6794 } 6795 if (th->syn) { 6796 if (th->fin) { 6797 SKB_DR_SET(reason, TCP_FLAGS); 6798 goto discard; 6799 } 6800 /* It is possible that we process SYN packets from backlog, 6801 * so we need to make sure to disable BH and RCU right there. 6802 */ 6803 rcu_read_lock(); 6804 local_bh_disable(); 6805 icsk->icsk_af_ops->conn_request(sk, skb); 6806 local_bh_enable(); 6807 rcu_read_unlock(); 6808 6809 consume_skb(skb); 6810 return 0; 6811 } 6812 SKB_DR_SET(reason, TCP_FLAGS); 6813 goto discard; 6814 6815 case TCP_SYN_SENT: 6816 tp->rx_opt.saw_tstamp = 0; 6817 tcp_mstamp_refresh(tp); 6818 queued = tcp_rcv_synsent_state_process(sk, skb, th); 6819 if (queued >= 0) 6820 return queued; 6821 6822 /* Do step6 onward by hand. */ 6823 tcp_urg(sk, skb, th); 6824 __kfree_skb(skb); 6825 tcp_data_snd_check(sk); 6826 return 0; 6827 } 6828 6829 tcp_mstamp_refresh(tp); 6830 tp->rx_opt.saw_tstamp = 0; 6831 req = rcu_dereference_protected(tp->fastopen_rsk, 6832 lockdep_sock_is_held(sk)); 6833 if (req) { 6834 bool req_stolen; 6835 6836 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV && 6837 sk->sk_state != TCP_FIN_WAIT1); 6838 6839 SKB_DR_SET(reason, TCP_FASTOPEN); 6840 if (!tcp_check_req(sk, skb, req, true, &req_stolen, &reason)) 6841 goto discard; 6842 } 6843 6844 if (!th->ack && !th->rst && !th->syn) { 6845 SKB_DR_SET(reason, TCP_FLAGS); 6846 goto discard; 6847 } 6848 if (!tcp_validate_incoming(sk, skb, th, 0)) 6849 return 0; 6850 6851 /* step 5: check the ACK field */ 6852 reason = tcp_ack(sk, skb, FLAG_SLOWPATH | 6853 FLAG_UPDATE_TS_RECENT | 6854 FLAG_NO_CHALLENGE_ACK); 6855 6856 if ((int)reason <= 0) { 6857 if (sk->sk_state == TCP_SYN_RECV) { 6858 /* send one RST */ 6859 if (!reason) 6860 return SKB_DROP_REASON_TCP_OLD_ACK; 6861 return -reason; 6862 } 6863 /* accept old ack during closing */ 6864 if ((int)reason < 0) { 6865 tcp_send_challenge_ack(sk); 6866 reason = -reason; 6867 goto discard; 6868 } 6869 } 6870 SKB_DR_SET(reason, NOT_SPECIFIED); 6871 switch (sk->sk_state) { 6872 case TCP_SYN_RECV: 6873 tp->delivered++; /* SYN-ACK delivery isn't tracked in tcp_ack */ 6874 if (!tp->srtt_us) 6875 tcp_synack_rtt_meas(sk, req); 6876 6877 if (req) { 6878 tcp_rcv_synrecv_state_fastopen(sk); 6879 } else { 6880 tcp_try_undo_spurious_syn(sk); 6881 tp->retrans_stamp = 0; 6882 tcp_init_transfer(sk, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB, 6883 skb); 6884 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt); 6885 } 6886 tcp_ao_established(sk); 6887 smp_mb(); 6888 tcp_set_state(sk, TCP_ESTABLISHED); 6889 sk->sk_state_change(sk); 6890 6891 /* Note, that this wakeup is only for marginal crossed SYN case. 6892 * Passively open sockets are not waked up, because 6893 * sk->sk_sleep == NULL and sk->sk_socket == NULL. 6894 */ 6895 if (sk->sk_socket) 6896 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT); 6897 6898 tp->snd_una = TCP_SKB_CB(skb)->ack_seq; 6899 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale; 6900 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq); 6901 6902 if (tp->rx_opt.tstamp_ok) 6903 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; 6904 6905 if (!inet_csk(sk)->icsk_ca_ops->cong_control) 6906 tcp_update_pacing_rate(sk); 6907 6908 /* Prevent spurious tcp_cwnd_restart() on first data packet */ 6909 tp->lsndtime = tcp_jiffies32; 6910 6911 tcp_initialize_rcv_mss(sk); 6912 tcp_fast_path_on(tp); 6913 if (sk->sk_shutdown & SEND_SHUTDOWN) 6914 tcp_shutdown(sk, SEND_SHUTDOWN); 6915 break; 6916 6917 case TCP_FIN_WAIT1: { 6918 int tmo; 6919 6920 if (req) 6921 tcp_rcv_synrecv_state_fastopen(sk); 6922 6923 if (tp->snd_una != tp->write_seq) 6924 break; 6925 6926 tcp_set_state(sk, TCP_FIN_WAIT2); 6927 WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | SEND_SHUTDOWN); 6928 6929 sk_dst_confirm(sk); 6930 6931 if (!sock_flag(sk, SOCK_DEAD)) { 6932 /* Wake up lingering close() */ 6933 sk->sk_state_change(sk); 6934 break; 6935 } 6936 6937 if (READ_ONCE(tp->linger2) < 0) { 6938 tcp_done(sk); 6939 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA); 6940 return SKB_DROP_REASON_TCP_ABORT_ON_DATA; 6941 } 6942 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 6943 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) { 6944 /* Receive out of order FIN after close() */ 6945 if (tp->syn_fastopen && th->fin) 6946 tcp_fastopen_active_disable(sk); 6947 tcp_done(sk); 6948 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA); 6949 return SKB_DROP_REASON_TCP_ABORT_ON_DATA; 6950 } 6951 6952 tmo = tcp_fin_time(sk); 6953 if (tmo > TCP_TIMEWAIT_LEN) { 6954 tcp_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN); 6955 } else if (th->fin || sock_owned_by_user(sk)) { 6956 /* Bad case. We could lose such FIN otherwise. 6957 * It is not a big problem, but it looks confusing 6958 * and not so rare event. We still can lose it now, 6959 * if it spins in bh_lock_sock(), but it is really 6960 * marginal case. 6961 */ 6962 tcp_reset_keepalive_timer(sk, tmo); 6963 } else { 6964 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo); 6965 goto consume; 6966 } 6967 break; 6968 } 6969 6970 case TCP_CLOSING: 6971 if (tp->snd_una == tp->write_seq) { 6972 tcp_time_wait(sk, TCP_TIME_WAIT, 0); 6973 goto consume; 6974 } 6975 break; 6976 6977 case TCP_LAST_ACK: 6978 if (tp->snd_una == tp->write_seq) { 6979 tcp_update_metrics(sk); 6980 tcp_done(sk); 6981 goto consume; 6982 } 6983 break; 6984 } 6985 6986 /* step 6: check the URG bit */ 6987 tcp_urg(sk, skb, th); 6988 6989 /* step 7: process the segment text */ 6990 switch (sk->sk_state) { 6991 case TCP_CLOSE_WAIT: 6992 case TCP_CLOSING: 6993 case TCP_LAST_ACK: 6994 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 6995 /* If a subflow has been reset, the packet should not 6996 * continue to be processed, drop the packet. 6997 */ 6998 if (sk_is_mptcp(sk) && !mptcp_incoming_options(sk, skb)) 6999 goto discard; 7000 break; 7001 } 7002 fallthrough; 7003 case TCP_FIN_WAIT1: 7004 case TCP_FIN_WAIT2: 7005 /* RFC 793 says to queue data in these states, 7006 * RFC 1122 says we MUST send a reset. 7007 * BSD 4.4 also does reset. 7008 */ 7009 if (sk->sk_shutdown & RCV_SHUTDOWN) { 7010 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 7011 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) { 7012 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA); 7013 tcp_reset(sk, skb); 7014 return SKB_DROP_REASON_TCP_ABORT_ON_DATA; 7015 } 7016 } 7017 fallthrough; 7018 case TCP_ESTABLISHED: 7019 tcp_data_queue(sk, skb); 7020 queued = 1; 7021 break; 7022 } 7023 7024 /* tcp_data could move socket to TIME-WAIT */ 7025 if (sk->sk_state != TCP_CLOSE) { 7026 tcp_data_snd_check(sk); 7027 tcp_ack_snd_check(sk); 7028 } 7029 7030 if (!queued) { 7031 discard: 7032 tcp_drop_reason(sk, skb, reason); 7033 } 7034 return 0; 7035 7036 consume: 7037 __kfree_skb(skb); 7038 return 0; 7039 } 7040 EXPORT_IPV6_MOD(tcp_rcv_state_process); 7041 7042 static inline void pr_drop_req(struct request_sock *req, __u16 port, int family) 7043 { 7044 struct inet_request_sock *ireq = inet_rsk(req); 7045 7046 if (family == AF_INET) 7047 net_dbg_ratelimited("drop open request from %pI4/%u\n", 7048 &ireq->ir_rmt_addr, port); 7049 #if IS_ENABLED(CONFIG_IPV6) 7050 else if (family == AF_INET6) 7051 net_dbg_ratelimited("drop open request from %pI6/%u\n", 7052 &ireq->ir_v6_rmt_addr, port); 7053 #endif 7054 } 7055 7056 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set 7057 * 7058 * If we receive a SYN packet with these bits set, it means a 7059 * network is playing bad games with TOS bits. In order to 7060 * avoid possible false congestion notifications, we disable 7061 * TCP ECN negotiation. 7062 * 7063 * Exception: tcp_ca wants ECN. This is required for DCTCP 7064 * congestion control: Linux DCTCP asserts ECT on all packets, 7065 * including SYN, which is most optimal solution; however, 7066 * others, such as FreeBSD do not. 7067 * 7068 * Exception: At least one of the reserved bits of the TCP header (th->res1) is 7069 * set, indicating the use of a future TCP extension (such as AccECN). See 7070 * RFC8311 §4.3 which updates RFC3168 to allow the development of such 7071 * extensions. 7072 */ 7073 static void tcp_ecn_create_request(struct request_sock *req, 7074 const struct sk_buff *skb, 7075 const struct sock *listen_sk, 7076 const struct dst_entry *dst) 7077 { 7078 const struct tcphdr *th = tcp_hdr(skb); 7079 const struct net *net = sock_net(listen_sk); 7080 bool th_ecn = th->ece && th->cwr; 7081 bool ect, ecn_ok; 7082 u32 ecn_ok_dst; 7083 7084 if (!th_ecn) 7085 return; 7086 7087 ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield); 7088 ecn_ok_dst = dst_feature(dst, DST_FEATURE_ECN_MASK); 7089 ecn_ok = READ_ONCE(net->ipv4.sysctl_tcp_ecn) || ecn_ok_dst; 7090 7091 if (((!ect || th->res1) && ecn_ok) || tcp_ca_needs_ecn(listen_sk) || 7092 (ecn_ok_dst & DST_FEATURE_ECN_CA) || 7093 tcp_bpf_ca_needs_ecn((struct sock *)req)) 7094 inet_rsk(req)->ecn_ok = 1; 7095 } 7096 7097 static void tcp_openreq_init(struct request_sock *req, 7098 const struct tcp_options_received *rx_opt, 7099 struct sk_buff *skb, const struct sock *sk) 7100 { 7101 struct inet_request_sock *ireq = inet_rsk(req); 7102 7103 req->rsk_rcv_wnd = 0; /* So that tcp_send_synack() knows! */ 7104 tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq; 7105 tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; 7106 tcp_rsk(req)->snt_synack = 0; 7107 tcp_rsk(req)->snt_tsval_first = 0; 7108 tcp_rsk(req)->last_oow_ack_time = 0; 7109 req->mss = rx_opt->mss_clamp; 7110 req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0; 7111 ireq->tstamp_ok = rx_opt->tstamp_ok; 7112 ireq->sack_ok = rx_opt->sack_ok; 7113 ireq->snd_wscale = rx_opt->snd_wscale; 7114 ireq->wscale_ok = rx_opt->wscale_ok; 7115 ireq->acked = 0; 7116 ireq->ecn_ok = 0; 7117 ireq->ir_rmt_port = tcp_hdr(skb)->source; 7118 ireq->ir_num = ntohs(tcp_hdr(skb)->dest); 7119 ireq->ir_mark = inet_request_mark(sk, skb); 7120 #if IS_ENABLED(CONFIG_SMC) 7121 ireq->smc_ok = rx_opt->smc_ok && !(tcp_sk(sk)->smc_hs_congested && 7122 tcp_sk(sk)->smc_hs_congested(sk)); 7123 #endif 7124 } 7125 7126 /* 7127 * Return true if a syncookie should be sent 7128 */ 7129 static bool tcp_syn_flood_action(struct sock *sk, const char *proto) 7130 { 7131 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue; 7132 const char *msg = "Dropping request"; 7133 struct net *net = sock_net(sk); 7134 bool want_cookie = false; 7135 u8 syncookies; 7136 7137 syncookies = READ_ONCE(net->ipv4.sysctl_tcp_syncookies); 7138 7139 #ifdef CONFIG_SYN_COOKIES 7140 if (syncookies) { 7141 msg = "Sending cookies"; 7142 want_cookie = true; 7143 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES); 7144 } else 7145 #endif 7146 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP); 7147 7148 if (!READ_ONCE(queue->synflood_warned) && syncookies != 2 && 7149 xchg(&queue->synflood_warned, 1) == 0) { 7150 if (IS_ENABLED(CONFIG_IPV6) && sk->sk_family == AF_INET6) { 7151 net_info_ratelimited("%s: Possible SYN flooding on port [%pI6c]:%u. %s.\n", 7152 proto, inet6_rcv_saddr(sk), 7153 sk->sk_num, msg); 7154 } else { 7155 net_info_ratelimited("%s: Possible SYN flooding on port %pI4:%u. %s.\n", 7156 proto, &sk->sk_rcv_saddr, 7157 sk->sk_num, msg); 7158 } 7159 } 7160 7161 return want_cookie; 7162 } 7163 7164 static void tcp_reqsk_record_syn(const struct sock *sk, 7165 struct request_sock *req, 7166 const struct sk_buff *skb) 7167 { 7168 if (tcp_sk(sk)->save_syn) { 7169 u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb); 7170 struct saved_syn *saved_syn; 7171 u32 mac_hdrlen; 7172 void *base; 7173 7174 if (tcp_sk(sk)->save_syn == 2) { /* Save full header. */ 7175 base = skb_mac_header(skb); 7176 mac_hdrlen = skb_mac_header_len(skb); 7177 len += mac_hdrlen; 7178 } else { 7179 base = skb_network_header(skb); 7180 mac_hdrlen = 0; 7181 } 7182 7183 saved_syn = kmalloc(struct_size(saved_syn, data, len), 7184 GFP_ATOMIC); 7185 if (saved_syn) { 7186 saved_syn->mac_hdrlen = mac_hdrlen; 7187 saved_syn->network_hdrlen = skb_network_header_len(skb); 7188 saved_syn->tcp_hdrlen = tcp_hdrlen(skb); 7189 memcpy(saved_syn->data, base, len); 7190 req->saved_syn = saved_syn; 7191 } 7192 } 7193 } 7194 7195 /* If a SYN cookie is required and supported, returns a clamped MSS value to be 7196 * used for SYN cookie generation. 7197 */ 7198 u16 tcp_get_syncookie_mss(struct request_sock_ops *rsk_ops, 7199 const struct tcp_request_sock_ops *af_ops, 7200 struct sock *sk, struct tcphdr *th) 7201 { 7202 struct tcp_sock *tp = tcp_sk(sk); 7203 u16 mss; 7204 7205 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_syncookies) != 2 && 7206 !inet_csk_reqsk_queue_is_full(sk)) 7207 return 0; 7208 7209 if (!tcp_syn_flood_action(sk, rsk_ops->slab_name)) 7210 return 0; 7211 7212 if (sk_acceptq_is_full(sk)) { 7213 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS); 7214 return 0; 7215 } 7216 7217 mss = tcp_parse_mss_option(th, tp->rx_opt.user_mss); 7218 if (!mss) 7219 mss = af_ops->mss_clamp; 7220 7221 return mss; 7222 } 7223 EXPORT_IPV6_MOD_GPL(tcp_get_syncookie_mss); 7224 7225 int tcp_conn_request(struct request_sock_ops *rsk_ops, 7226 const struct tcp_request_sock_ops *af_ops, 7227 struct sock *sk, struct sk_buff *skb) 7228 { 7229 struct tcp_fastopen_cookie foc = { .len = -1 }; 7230 struct tcp_options_received tmp_opt; 7231 struct tcp_sock *tp = tcp_sk(sk); 7232 struct net *net = sock_net(sk); 7233 struct sock *fastopen_sk = NULL; 7234 struct request_sock *req; 7235 bool want_cookie = false; 7236 struct dst_entry *dst; 7237 struct flowi fl; 7238 u8 syncookies; 7239 u32 isn; 7240 7241 #ifdef CONFIG_TCP_AO 7242 const struct tcp_ao_hdr *aoh; 7243 #endif 7244 7245 isn = __this_cpu_read(tcp_tw_isn); 7246 if (isn) { 7247 /* TW buckets are converted to open requests without 7248 * limitations, they conserve resources and peer is 7249 * evidently real one. 7250 */ 7251 __this_cpu_write(tcp_tw_isn, 0); 7252 } else { 7253 syncookies = READ_ONCE(net->ipv4.sysctl_tcp_syncookies); 7254 7255 if (syncookies == 2 || inet_csk_reqsk_queue_is_full(sk)) { 7256 want_cookie = tcp_syn_flood_action(sk, 7257 rsk_ops->slab_name); 7258 if (!want_cookie) 7259 goto drop; 7260 } 7261 } 7262 7263 if (sk_acceptq_is_full(sk)) { 7264 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS); 7265 goto drop; 7266 } 7267 7268 req = inet_reqsk_alloc(rsk_ops, sk, !want_cookie); 7269 if (!req) 7270 goto drop; 7271 7272 req->syncookie = want_cookie; 7273 tcp_rsk(req)->af_specific = af_ops; 7274 tcp_rsk(req)->ts_off = 0; 7275 tcp_rsk(req)->req_usec_ts = false; 7276 #if IS_ENABLED(CONFIG_MPTCP) 7277 tcp_rsk(req)->is_mptcp = 0; 7278 #endif 7279 7280 tcp_clear_options(&tmp_opt); 7281 tmp_opt.mss_clamp = af_ops->mss_clamp; 7282 tmp_opt.user_mss = tp->rx_opt.user_mss; 7283 tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0, 7284 want_cookie ? NULL : &foc); 7285 7286 if (want_cookie && !tmp_opt.saw_tstamp) 7287 tcp_clear_options(&tmp_opt); 7288 7289 if (IS_ENABLED(CONFIG_SMC) && want_cookie) 7290 tmp_opt.smc_ok = 0; 7291 7292 tmp_opt.tstamp_ok = tmp_opt.saw_tstamp; 7293 tcp_openreq_init(req, &tmp_opt, skb, sk); 7294 inet_rsk(req)->no_srccheck = inet_test_bit(TRANSPARENT, sk); 7295 7296 /* Note: tcp_v6_init_req() might override ir_iif for link locals */ 7297 inet_rsk(req)->ir_iif = inet_request_bound_dev_if(sk, skb); 7298 7299 dst = af_ops->route_req(sk, skb, &fl, req, isn); 7300 if (!dst) 7301 goto drop_and_free; 7302 7303 if (tmp_opt.tstamp_ok) { 7304 tcp_rsk(req)->req_usec_ts = dst_tcp_usec_ts(dst); 7305 tcp_rsk(req)->ts_off = af_ops->init_ts_off(net, skb); 7306 } 7307 if (!want_cookie && !isn) { 7308 int max_syn_backlog = READ_ONCE(net->ipv4.sysctl_max_syn_backlog); 7309 7310 /* Kill the following clause, if you dislike this way. */ 7311 if (!syncookies && 7312 (max_syn_backlog - inet_csk_reqsk_queue_len(sk) < 7313 (max_syn_backlog >> 2)) && 7314 !tcp_peer_is_proven(req, dst)) { 7315 /* Without syncookies last quarter of 7316 * backlog is filled with destinations, 7317 * proven to be alive. 7318 * It means that we continue to communicate 7319 * to destinations, already remembered 7320 * to the moment of synflood. 7321 */ 7322 pr_drop_req(req, ntohs(tcp_hdr(skb)->source), 7323 rsk_ops->family); 7324 goto drop_and_release; 7325 } 7326 7327 isn = af_ops->init_seq(skb); 7328 } 7329 7330 tcp_ecn_create_request(req, skb, sk, dst); 7331 7332 if (want_cookie) { 7333 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss); 7334 if (!tmp_opt.tstamp_ok) 7335 inet_rsk(req)->ecn_ok = 0; 7336 } 7337 7338 #ifdef CONFIG_TCP_AO 7339 if (tcp_parse_auth_options(tcp_hdr(skb), NULL, &aoh)) 7340 goto drop_and_release; /* Invalid TCP options */ 7341 if (aoh) { 7342 tcp_rsk(req)->used_tcp_ao = true; 7343 tcp_rsk(req)->ao_rcv_next = aoh->keyid; 7344 tcp_rsk(req)->ao_keyid = aoh->rnext_keyid; 7345 7346 } else { 7347 tcp_rsk(req)->used_tcp_ao = false; 7348 } 7349 #endif 7350 tcp_rsk(req)->snt_isn = isn; 7351 tcp_rsk(req)->txhash = net_tx_rndhash(); 7352 tcp_rsk(req)->syn_tos = TCP_SKB_CB(skb)->ip_dsfield; 7353 tcp_openreq_init_rwin(req, sk, dst); 7354 sk_rx_queue_set(req_to_sk(req), skb); 7355 if (!want_cookie) { 7356 tcp_reqsk_record_syn(sk, req, skb); 7357 fastopen_sk = tcp_try_fastopen(sk, skb, req, &foc, dst); 7358 } 7359 if (fastopen_sk) { 7360 af_ops->send_synack(fastopen_sk, dst, &fl, req, 7361 &foc, TCP_SYNACK_FASTOPEN, skb); 7362 /* Add the child socket directly into the accept queue */ 7363 if (!inet_csk_reqsk_queue_add(sk, req, fastopen_sk)) { 7364 reqsk_fastopen_remove(fastopen_sk, req, false); 7365 bh_unlock_sock(fastopen_sk); 7366 sock_put(fastopen_sk); 7367 goto drop_and_free; 7368 } 7369 sk->sk_data_ready(sk); 7370 bh_unlock_sock(fastopen_sk); 7371 sock_put(fastopen_sk); 7372 } else { 7373 tcp_rsk(req)->tfo_listener = false; 7374 if (!want_cookie) { 7375 req->timeout = tcp_timeout_init((struct sock *)req); 7376 if (unlikely(!inet_csk_reqsk_queue_hash_add(sk, req, 7377 req->timeout))) { 7378 reqsk_free(req); 7379 dst_release(dst); 7380 return 0; 7381 } 7382 7383 } 7384 af_ops->send_synack(sk, dst, &fl, req, &foc, 7385 !want_cookie ? TCP_SYNACK_NORMAL : 7386 TCP_SYNACK_COOKIE, 7387 skb); 7388 if (want_cookie) { 7389 reqsk_free(req); 7390 return 0; 7391 } 7392 } 7393 reqsk_put(req); 7394 return 0; 7395 7396 drop_and_release: 7397 dst_release(dst); 7398 drop_and_free: 7399 __reqsk_free(req); 7400 drop: 7401 tcp_listendrop(sk); 7402 return 0; 7403 } 7404 EXPORT_IPV6_MOD(tcp_conn_request); 7405