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 */ 105 106 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED) 107 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED) 108 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE|FLAG_DSACKING_ACK) 109 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED) 110 111 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH) 112 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH)) 113 114 #define REXMIT_NONE 0 /* no loss recovery to do */ 115 #define REXMIT_LOST 1 /* retransmit packets marked lost */ 116 #define REXMIT_NEW 2 /* FRTO-style transmit of unsent/new packets */ 117 118 #if IS_ENABLED(CONFIG_TLS_DEVICE) 119 static DEFINE_STATIC_KEY_DEFERRED_FALSE(clean_acked_data_enabled, HZ); 120 121 void clean_acked_data_enable(struct inet_connection_sock *icsk, 122 void (*cad)(struct sock *sk, u32 ack_seq)) 123 { 124 icsk->icsk_clean_acked = cad; 125 static_branch_deferred_inc(&clean_acked_data_enabled); 126 } 127 EXPORT_SYMBOL_GPL(clean_acked_data_enable); 128 129 void clean_acked_data_disable(struct inet_connection_sock *icsk) 130 { 131 static_branch_slow_dec_deferred(&clean_acked_data_enabled); 132 icsk->icsk_clean_acked = NULL; 133 } 134 EXPORT_SYMBOL_GPL(clean_acked_data_disable); 135 136 void clean_acked_data_flush(void) 137 { 138 static_key_deferred_flush(&clean_acked_data_enabled); 139 } 140 EXPORT_SYMBOL_GPL(clean_acked_data_flush); 141 #endif 142 143 #ifdef CONFIG_CGROUP_BPF 144 static void bpf_skops_parse_hdr(struct sock *sk, struct sk_buff *skb) 145 { 146 bool unknown_opt = tcp_sk(sk)->rx_opt.saw_unknown && 147 BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), 148 BPF_SOCK_OPS_PARSE_UNKNOWN_HDR_OPT_CB_FLAG); 149 bool parse_all_opt = BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), 150 BPF_SOCK_OPS_PARSE_ALL_HDR_OPT_CB_FLAG); 151 struct bpf_sock_ops_kern sock_ops; 152 153 if (likely(!unknown_opt && !parse_all_opt)) 154 return; 155 156 /* The skb will be handled in the 157 * bpf_skops_established() or 158 * bpf_skops_write_hdr_opt(). 159 */ 160 switch (sk->sk_state) { 161 case TCP_SYN_RECV: 162 case TCP_SYN_SENT: 163 case TCP_LISTEN: 164 return; 165 } 166 167 sock_owned_by_me(sk); 168 169 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp)); 170 sock_ops.op = BPF_SOCK_OPS_PARSE_HDR_OPT_CB; 171 sock_ops.is_fullsock = 1; 172 sock_ops.sk = sk; 173 bpf_skops_init_skb(&sock_ops, skb, tcp_hdrlen(skb)); 174 175 BPF_CGROUP_RUN_PROG_SOCK_OPS(&sock_ops); 176 } 177 178 static void bpf_skops_established(struct sock *sk, int bpf_op, 179 struct sk_buff *skb) 180 { 181 struct bpf_sock_ops_kern sock_ops; 182 183 sock_owned_by_me(sk); 184 185 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp)); 186 sock_ops.op = bpf_op; 187 sock_ops.is_fullsock = 1; 188 sock_ops.sk = sk; 189 /* sk with TCP_REPAIR_ON does not have skb in tcp_finish_connect */ 190 if (skb) 191 bpf_skops_init_skb(&sock_ops, skb, tcp_hdrlen(skb)); 192 193 BPF_CGROUP_RUN_PROG_SOCK_OPS(&sock_ops); 194 } 195 #else 196 static void bpf_skops_parse_hdr(struct sock *sk, struct sk_buff *skb) 197 { 198 } 199 200 static void bpf_skops_established(struct sock *sk, int bpf_op, 201 struct sk_buff *skb) 202 { 203 } 204 #endif 205 206 static __cold void tcp_gro_dev_warn(const struct sock *sk, const struct sk_buff *skb, 207 unsigned int len) 208 { 209 struct net_device *dev; 210 211 rcu_read_lock(); 212 dev = dev_get_by_index_rcu(sock_net(sk), skb->skb_iif); 213 if (!dev || len >= READ_ONCE(dev->mtu)) 214 pr_warn("%s: Driver has suspect GRO implementation, TCP performance may be compromised.\n", 215 dev ? dev->name : "Unknown driver"); 216 rcu_read_unlock(); 217 } 218 219 /* Adapt the MSS value used to make delayed ack decision to the 220 * real world. 221 */ 222 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb) 223 { 224 struct inet_connection_sock *icsk = inet_csk(sk); 225 const unsigned int lss = icsk->icsk_ack.last_seg_size; 226 unsigned int len; 227 228 icsk->icsk_ack.last_seg_size = 0; 229 230 /* skb->len may jitter because of SACKs, even if peer 231 * sends good full-sized frames. 232 */ 233 len = skb_shinfo(skb)->gso_size ? : skb->len; 234 if (len >= icsk->icsk_ack.rcv_mss) { 235 /* Note: divides are still a bit expensive. 236 * For the moment, only adjust scaling_ratio 237 * when we update icsk_ack.rcv_mss. 238 */ 239 if (unlikely(len != icsk->icsk_ack.rcv_mss)) { 240 u64 val = (u64)skb->len << TCP_RMEM_TO_WIN_SCALE; 241 u8 old_ratio = tcp_sk(sk)->scaling_ratio; 242 243 do_div(val, skb->truesize); 244 tcp_sk(sk)->scaling_ratio = val ? val : 1; 245 246 if (old_ratio != tcp_sk(sk)->scaling_ratio) { 247 struct tcp_sock *tp = tcp_sk(sk); 248 249 val = tcp_win_from_space(sk, sk->sk_rcvbuf); 250 tcp_set_window_clamp(sk, val); 251 252 if (tp->window_clamp < tp->rcvq_space.space) 253 tp->rcvq_space.space = tp->window_clamp; 254 } 255 } 256 icsk->icsk_ack.rcv_mss = min_t(unsigned int, len, 257 tcp_sk(sk)->advmss); 258 /* Account for possibly-removed options */ 259 DO_ONCE_LITE_IF(len > icsk->icsk_ack.rcv_mss + MAX_TCP_OPTION_SPACE, 260 tcp_gro_dev_warn, sk, skb, len); 261 /* If the skb has a len of exactly 1*MSS and has the PSH bit 262 * set then it is likely the end of an application write. So 263 * more data may not be arriving soon, and yet the data sender 264 * may be waiting for an ACK if cwnd-bound or using TX zero 265 * copy. So we set ICSK_ACK_PUSHED here so that 266 * tcp_cleanup_rbuf() will send an ACK immediately if the app 267 * reads all of the data and is not ping-pong. If len > MSS 268 * then this logic does not matter (and does not hurt) because 269 * tcp_cleanup_rbuf() will always ACK immediately if the app 270 * reads data and there is more than an MSS of unACKed data. 271 */ 272 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_PSH) 273 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED; 274 } else { 275 /* Otherwise, we make more careful check taking into account, 276 * that SACKs block is variable. 277 * 278 * "len" is invariant segment length, including TCP header. 279 */ 280 len += skb->data - skb_transport_header(skb); 281 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) || 282 /* If PSH is not set, packet should be 283 * full sized, provided peer TCP is not badly broken. 284 * This observation (if it is correct 8)) allows 285 * to handle super-low mtu links fairly. 286 */ 287 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) && 288 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) { 289 /* Subtract also invariant (if peer is RFC compliant), 290 * tcp header plus fixed timestamp option length. 291 * Resulting "len" is MSS free of SACK jitter. 292 */ 293 len -= tcp_sk(sk)->tcp_header_len; 294 icsk->icsk_ack.last_seg_size = len; 295 if (len == lss) { 296 icsk->icsk_ack.rcv_mss = len; 297 return; 298 } 299 } 300 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED) 301 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2; 302 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED; 303 } 304 } 305 306 static void tcp_incr_quickack(struct sock *sk, unsigned int max_quickacks) 307 { 308 struct inet_connection_sock *icsk = inet_csk(sk); 309 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss); 310 311 if (quickacks == 0) 312 quickacks = 2; 313 quickacks = min(quickacks, max_quickacks); 314 if (quickacks > icsk->icsk_ack.quick) 315 icsk->icsk_ack.quick = quickacks; 316 } 317 318 static void tcp_enter_quickack_mode(struct sock *sk, unsigned int max_quickacks) 319 { 320 struct inet_connection_sock *icsk = inet_csk(sk); 321 322 tcp_incr_quickack(sk, max_quickacks); 323 inet_csk_exit_pingpong_mode(sk); 324 icsk->icsk_ack.ato = TCP_ATO_MIN; 325 } 326 327 /* Send ACKs quickly, if "quick" count is not exhausted 328 * and the session is not interactive. 329 */ 330 331 static bool tcp_in_quickack_mode(struct sock *sk) 332 { 333 const struct inet_connection_sock *icsk = inet_csk(sk); 334 const struct dst_entry *dst = __sk_dst_get(sk); 335 336 return (dst && dst_metric(dst, RTAX_QUICKACK)) || 337 (icsk->icsk_ack.quick && !inet_csk_in_pingpong_mode(sk)); 338 } 339 340 static void tcp_ecn_queue_cwr(struct tcp_sock *tp) 341 { 342 if (tp->ecn_flags & TCP_ECN_OK) 343 tp->ecn_flags |= TCP_ECN_QUEUE_CWR; 344 } 345 346 static void tcp_ecn_accept_cwr(struct sock *sk, const struct sk_buff *skb) 347 { 348 if (tcp_hdr(skb)->cwr) { 349 tcp_sk(sk)->ecn_flags &= ~TCP_ECN_DEMAND_CWR; 350 351 /* If the sender is telling us it has entered CWR, then its 352 * cwnd may be very low (even just 1 packet), so we should ACK 353 * immediately. 354 */ 355 if (TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) 356 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW; 357 } 358 } 359 360 static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp) 361 { 362 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; 363 } 364 365 static void __tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb) 366 { 367 struct tcp_sock *tp = tcp_sk(sk); 368 369 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) { 370 case INET_ECN_NOT_ECT: 371 /* Funny extension: if ECT is not set on a segment, 372 * and we already seen ECT on a previous segment, 373 * it is probably a retransmit. 374 */ 375 if (tp->ecn_flags & TCP_ECN_SEEN) 376 tcp_enter_quickack_mode(sk, 2); 377 break; 378 case INET_ECN_CE: 379 if (tcp_ca_needs_ecn(sk)) 380 tcp_ca_event(sk, CA_EVENT_ECN_IS_CE); 381 382 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) { 383 /* Better not delay acks, sender can have a very low cwnd */ 384 tcp_enter_quickack_mode(sk, 2); 385 tp->ecn_flags |= TCP_ECN_DEMAND_CWR; 386 } 387 tp->ecn_flags |= TCP_ECN_SEEN; 388 break; 389 default: 390 if (tcp_ca_needs_ecn(sk)) 391 tcp_ca_event(sk, CA_EVENT_ECN_NO_CE); 392 tp->ecn_flags |= TCP_ECN_SEEN; 393 break; 394 } 395 } 396 397 static void tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb) 398 { 399 if (tcp_sk(sk)->ecn_flags & TCP_ECN_OK) 400 __tcp_ecn_check_ce(sk, skb); 401 } 402 403 static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th) 404 { 405 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr)) 406 tp->ecn_flags &= ~TCP_ECN_OK; 407 } 408 409 static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th) 410 { 411 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr)) 412 tp->ecn_flags &= ~TCP_ECN_OK; 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 && (tp->ecn_flags & TCP_ECN_OK)) 418 return true; 419 return false; 420 } 421 422 /* Buffer size and advertised window tuning. 423 * 424 * 1. Tuning sk->sk_sndbuf, when connection enters established state. 425 */ 426 427 static void tcp_sndbuf_expand(struct sock *sk) 428 { 429 const struct tcp_sock *tp = tcp_sk(sk); 430 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 431 int sndmem, per_mss; 432 u32 nr_segs; 433 434 /* Worst case is non GSO/TSO : each frame consumes one skb 435 * and skb->head is kmalloced using power of two area of memory 436 */ 437 per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) + 438 MAX_TCP_HEADER + 439 SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); 440 441 per_mss = roundup_pow_of_two(per_mss) + 442 SKB_DATA_ALIGN(sizeof(struct sk_buff)); 443 444 nr_segs = max_t(u32, TCP_INIT_CWND, tcp_snd_cwnd(tp)); 445 nr_segs = max_t(u32, nr_segs, tp->reordering + 1); 446 447 /* Fast Recovery (RFC 5681 3.2) : 448 * Cubic needs 1.7 factor, rounded to 2 to include 449 * extra cushion (application might react slowly to EPOLLOUT) 450 */ 451 sndmem = ca_ops->sndbuf_expand ? ca_ops->sndbuf_expand(sk) : 2; 452 sndmem *= nr_segs * per_mss; 453 454 if (sk->sk_sndbuf < sndmem) 455 WRITE_ONCE(sk->sk_sndbuf, 456 min(sndmem, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_wmem[2]))); 457 } 458 459 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh) 460 * 461 * All tcp_full_space() is split to two parts: "network" buffer, allocated 462 * forward and advertised in receiver window (tp->rcv_wnd) and 463 * "application buffer", required to isolate scheduling/application 464 * latencies from network. 465 * window_clamp is maximal advertised window. It can be less than 466 * tcp_full_space(), in this case tcp_full_space() - window_clamp 467 * is reserved for "application" buffer. The less window_clamp is 468 * the smoother our behaviour from viewpoint of network, but the lower 469 * throughput and the higher sensitivity of the connection to losses. 8) 470 * 471 * rcv_ssthresh is more strict window_clamp used at "slow start" 472 * phase to predict further behaviour of this connection. 473 * It is used for two goals: 474 * - to enforce header prediction at sender, even when application 475 * requires some significant "application buffer". It is check #1. 476 * - to prevent pruning of receive queue because of misprediction 477 * of receiver window. Check #2. 478 * 479 * The scheme does not work when sender sends good segments opening 480 * window and then starts to feed us spaghetti. But it should work 481 * in common situations. Otherwise, we have to rely on queue collapsing. 482 */ 483 484 /* Slow part of check#2. */ 485 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb, 486 unsigned int skbtruesize) 487 { 488 const struct tcp_sock *tp = tcp_sk(sk); 489 /* Optimize this! */ 490 int truesize = tcp_win_from_space(sk, skbtruesize) >> 1; 491 int window = tcp_win_from_space(sk, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2])) >> 1; 492 493 while (tp->rcv_ssthresh <= window) { 494 if (truesize <= skb->len) 495 return 2 * inet_csk(sk)->icsk_ack.rcv_mss; 496 497 truesize >>= 1; 498 window >>= 1; 499 } 500 return 0; 501 } 502 503 /* Even if skb appears to have a bad len/truesize ratio, TCP coalescing 504 * can play nice with us, as sk_buff and skb->head might be either 505 * freed or shared with up to MAX_SKB_FRAGS segments. 506 * Only give a boost to drivers using page frag(s) to hold the frame(s), 507 * and if no payload was pulled in skb->head before reaching us. 508 */ 509 static u32 truesize_adjust(bool adjust, const struct sk_buff *skb) 510 { 511 u32 truesize = skb->truesize; 512 513 if (adjust && !skb_headlen(skb)) { 514 truesize -= SKB_TRUESIZE(skb_end_offset(skb)); 515 /* paranoid check, some drivers might be buggy */ 516 if (unlikely((int)truesize < (int)skb->len)) 517 truesize = skb->truesize; 518 } 519 return truesize; 520 } 521 522 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb, 523 bool adjust) 524 { 525 struct tcp_sock *tp = tcp_sk(sk); 526 int room; 527 528 room = min_t(int, tp->window_clamp, tcp_space(sk)) - tp->rcv_ssthresh; 529 530 if (room <= 0) 531 return; 532 533 /* Check #1 */ 534 if (!tcp_under_memory_pressure(sk)) { 535 unsigned int truesize = truesize_adjust(adjust, skb); 536 int incr; 537 538 /* Check #2. Increase window, if skb with such overhead 539 * will fit to rcvbuf in future. 540 */ 541 if (tcp_win_from_space(sk, truesize) <= skb->len) 542 incr = 2 * tp->advmss; 543 else 544 incr = __tcp_grow_window(sk, skb, truesize); 545 546 if (incr) { 547 incr = max_t(int, incr, 2 * skb->len); 548 tp->rcv_ssthresh += min(room, incr); 549 inet_csk(sk)->icsk_ack.quick |= 1; 550 } 551 } else { 552 /* Under pressure: 553 * Adjust rcv_ssthresh according to reserved mem 554 */ 555 tcp_adjust_rcv_ssthresh(sk); 556 } 557 } 558 559 /* 3. Try to fixup all. It is made immediately after connection enters 560 * established state. 561 */ 562 static void tcp_init_buffer_space(struct sock *sk) 563 { 564 int tcp_app_win = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_app_win); 565 struct tcp_sock *tp = tcp_sk(sk); 566 int maxwin; 567 568 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK)) 569 tcp_sndbuf_expand(sk); 570 571 tcp_mstamp_refresh(tp); 572 tp->rcvq_space.time = tp->tcp_mstamp; 573 tp->rcvq_space.seq = tp->copied_seq; 574 575 maxwin = tcp_full_space(sk); 576 577 if (tp->window_clamp >= maxwin) { 578 WRITE_ONCE(tp->window_clamp, maxwin); 579 580 if (tcp_app_win && maxwin > 4 * tp->advmss) 581 WRITE_ONCE(tp->window_clamp, 582 max(maxwin - (maxwin >> tcp_app_win), 583 4 * tp->advmss)); 584 } 585 586 /* Force reservation of one segment. */ 587 if (tcp_app_win && 588 tp->window_clamp > 2 * tp->advmss && 589 tp->window_clamp + tp->advmss > maxwin) 590 WRITE_ONCE(tp->window_clamp, 591 max(2 * tp->advmss, maxwin - tp->advmss)); 592 593 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp); 594 tp->snd_cwnd_stamp = tcp_jiffies32; 595 tp->rcvq_space.space = min3(tp->rcv_ssthresh, tp->rcv_wnd, 596 (u32)TCP_INIT_CWND * tp->advmss); 597 } 598 599 /* 4. Recalculate window clamp after socket hit its memory bounds. */ 600 static void tcp_clamp_window(struct sock *sk) 601 { 602 struct tcp_sock *tp = tcp_sk(sk); 603 struct inet_connection_sock *icsk = inet_csk(sk); 604 struct net *net = sock_net(sk); 605 int rmem2; 606 607 icsk->icsk_ack.quick = 0; 608 rmem2 = READ_ONCE(net->ipv4.sysctl_tcp_rmem[2]); 609 610 if (sk->sk_rcvbuf < rmem2 && 611 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) && 612 !tcp_under_memory_pressure(sk) && 613 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) { 614 WRITE_ONCE(sk->sk_rcvbuf, 615 min(atomic_read(&sk->sk_rmem_alloc), rmem2)); 616 } 617 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf) 618 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss); 619 } 620 621 /* Initialize RCV_MSS value. 622 * RCV_MSS is an our guess about MSS used by the peer. 623 * We haven't any direct information about the MSS. 624 * It's better to underestimate the RCV_MSS rather than overestimate. 625 * Overestimations make us ACKing less frequently than needed. 626 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss(). 627 */ 628 void tcp_initialize_rcv_mss(struct sock *sk) 629 { 630 const struct tcp_sock *tp = tcp_sk(sk); 631 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache); 632 633 hint = min(hint, tp->rcv_wnd / 2); 634 hint = min(hint, TCP_MSS_DEFAULT); 635 hint = max(hint, TCP_MIN_MSS); 636 637 inet_csk(sk)->icsk_ack.rcv_mss = hint; 638 } 639 EXPORT_SYMBOL(tcp_initialize_rcv_mss); 640 641 /* Receiver "autotuning" code. 642 * 643 * The algorithm for RTT estimation w/o timestamps is based on 644 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL. 645 * <https://public.lanl.gov/radiant/pubs.html#DRS> 646 * 647 * More detail on this code can be found at 648 * <http://staff.psc.edu/jheffner/>, 649 * though this reference is out of date. A new paper 650 * is pending. 651 */ 652 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep) 653 { 654 u32 new_sample = tp->rcv_rtt_est.rtt_us; 655 long m = sample; 656 657 if (new_sample != 0) { 658 /* If we sample in larger samples in the non-timestamp 659 * case, we could grossly overestimate the RTT especially 660 * with chatty applications or bulk transfer apps which 661 * are stalled on filesystem I/O. 662 * 663 * Also, since we are only going for a minimum in the 664 * non-timestamp case, we do not smooth things out 665 * else with timestamps disabled convergence takes too 666 * long. 667 */ 668 if (!win_dep) { 669 m -= (new_sample >> 3); 670 new_sample += m; 671 } else { 672 m <<= 3; 673 if (m < new_sample) 674 new_sample = m; 675 } 676 } else { 677 /* No previous measure. */ 678 new_sample = m << 3; 679 } 680 681 tp->rcv_rtt_est.rtt_us = new_sample; 682 } 683 684 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp) 685 { 686 u32 delta_us; 687 688 if (tp->rcv_rtt_est.time == 0) 689 goto new_measure; 690 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq)) 691 return; 692 delta_us = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcv_rtt_est.time); 693 if (!delta_us) 694 delta_us = 1; 695 tcp_rcv_rtt_update(tp, delta_us, 1); 696 697 new_measure: 698 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd; 699 tp->rcv_rtt_est.time = tp->tcp_mstamp; 700 } 701 702 static s32 tcp_rtt_tsopt_us(const struct tcp_sock *tp) 703 { 704 u32 delta, delta_us; 705 706 delta = tcp_time_stamp_ts(tp) - tp->rx_opt.rcv_tsecr; 707 if (tp->tcp_usec_ts) 708 return delta; 709 710 if (likely(delta < INT_MAX / (USEC_PER_SEC / TCP_TS_HZ))) { 711 if (!delta) 712 delta = 1; 713 delta_us = delta * (USEC_PER_SEC / TCP_TS_HZ); 714 return delta_us; 715 } 716 return -1; 717 } 718 719 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, 720 const struct sk_buff *skb) 721 { 722 struct tcp_sock *tp = tcp_sk(sk); 723 724 if (tp->rx_opt.rcv_tsecr == tp->rcv_rtt_last_tsecr) 725 return; 726 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr; 727 728 if (TCP_SKB_CB(skb)->end_seq - 729 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss) { 730 s32 delta = tcp_rtt_tsopt_us(tp); 731 732 if (delta >= 0) 733 tcp_rcv_rtt_update(tp, delta, 0); 734 } 735 } 736 737 /* 738 * This function should be called every time data is copied to user space. 739 * It calculates the appropriate TCP receive buffer space. 740 */ 741 void tcp_rcv_space_adjust(struct sock *sk) 742 { 743 struct tcp_sock *tp = tcp_sk(sk); 744 u32 copied; 745 int time; 746 747 trace_tcp_rcv_space_adjust(sk); 748 749 tcp_mstamp_refresh(tp); 750 time = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcvq_space.time); 751 if (time < (tp->rcv_rtt_est.rtt_us >> 3) || tp->rcv_rtt_est.rtt_us == 0) 752 return; 753 754 /* Number of bytes copied to user in last RTT */ 755 copied = tp->copied_seq - tp->rcvq_space.seq; 756 if (copied <= tp->rcvq_space.space) 757 goto new_measure; 758 759 /* A bit of theory : 760 * copied = bytes received in previous RTT, our base window 761 * To cope with packet losses, we need a 2x factor 762 * To cope with slow start, and sender growing its cwin by 100 % 763 * every RTT, we need a 4x factor, because the ACK we are sending 764 * now is for the next RTT, not the current one : 765 * <prev RTT . ><current RTT .. ><next RTT .... > 766 */ 767 768 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf) && 769 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) { 770 u64 rcvwin, grow; 771 int rcvbuf; 772 773 /* minimal window to cope with packet losses, assuming 774 * steady state. Add some cushion because of small variations. 775 */ 776 rcvwin = ((u64)copied << 1) + 16 * tp->advmss; 777 778 /* Accommodate for sender rate increase (eg. slow start) */ 779 grow = rcvwin * (copied - tp->rcvq_space.space); 780 do_div(grow, tp->rcvq_space.space); 781 rcvwin += (grow << 1); 782 783 rcvbuf = min_t(u64, tcp_space_from_win(sk, rcvwin), 784 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2])); 785 if (rcvbuf > sk->sk_rcvbuf) { 786 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf); 787 788 /* Make the window clamp follow along. */ 789 WRITE_ONCE(tp->window_clamp, 790 tcp_win_from_space(sk, rcvbuf)); 791 } 792 } 793 tp->rcvq_space.space = copied; 794 795 new_measure: 796 tp->rcvq_space.seq = tp->copied_seq; 797 tp->rcvq_space.time = tp->tcp_mstamp; 798 } 799 800 static void tcp_save_lrcv_flowlabel(struct sock *sk, const struct sk_buff *skb) 801 { 802 #if IS_ENABLED(CONFIG_IPV6) 803 struct inet_connection_sock *icsk = inet_csk(sk); 804 805 if (skb->protocol == htons(ETH_P_IPV6)) 806 icsk->icsk_ack.lrcv_flowlabel = ntohl(ip6_flowlabel(ipv6_hdr(skb))); 807 #endif 808 } 809 810 /* There is something which you must keep in mind when you analyze the 811 * behavior of the tp->ato delayed ack timeout interval. When a 812 * connection starts up, we want to ack as quickly as possible. The 813 * problem is that "good" TCP's do slow start at the beginning of data 814 * transmission. The means that until we send the first few ACK's the 815 * sender will sit on his end and only queue most of his data, because 816 * he can only send snd_cwnd unacked packets at any given time. For 817 * each ACK we send, he increments snd_cwnd and transmits more of his 818 * queue. -DaveM 819 */ 820 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb) 821 { 822 struct tcp_sock *tp = tcp_sk(sk); 823 struct inet_connection_sock *icsk = inet_csk(sk); 824 u32 now; 825 826 inet_csk_schedule_ack(sk); 827 828 tcp_measure_rcv_mss(sk, skb); 829 830 tcp_rcv_rtt_measure(tp); 831 832 now = tcp_jiffies32; 833 834 if (!icsk->icsk_ack.ato) { 835 /* The _first_ data packet received, initialize 836 * delayed ACK engine. 837 */ 838 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS); 839 icsk->icsk_ack.ato = TCP_ATO_MIN; 840 } else { 841 int m = now - icsk->icsk_ack.lrcvtime; 842 843 if (m <= TCP_ATO_MIN / 2) { 844 /* The fastest case is the first. */ 845 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2; 846 } else if (m < icsk->icsk_ack.ato) { 847 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m; 848 if (icsk->icsk_ack.ato > icsk->icsk_rto) 849 icsk->icsk_ack.ato = icsk->icsk_rto; 850 } else if (m > icsk->icsk_rto) { 851 /* Too long gap. Apparently sender failed to 852 * restart window, so that we send ACKs quickly. 853 */ 854 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS); 855 } 856 } 857 icsk->icsk_ack.lrcvtime = now; 858 tcp_save_lrcv_flowlabel(sk, skb); 859 860 tcp_ecn_check_ce(sk, skb); 861 862 if (skb->len >= 128) 863 tcp_grow_window(sk, skb, true); 864 } 865 866 /* Called to compute a smoothed rtt estimate. The data fed to this 867 * routine either comes from timestamps, or from segments that were 868 * known _not_ to have been retransmitted [see Karn/Partridge 869 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88 870 * piece by Van Jacobson. 871 * NOTE: the next three routines used to be one big routine. 872 * To save cycles in the RFC 1323 implementation it was better to break 873 * it up into three procedures. -- erics 874 */ 875 static void tcp_rtt_estimator(struct sock *sk, long mrtt_us) 876 { 877 struct tcp_sock *tp = tcp_sk(sk); 878 long m = mrtt_us; /* RTT */ 879 u32 srtt = tp->srtt_us; 880 881 /* The following amusing code comes from Jacobson's 882 * article in SIGCOMM '88. Note that rtt and mdev 883 * are scaled versions of rtt and mean deviation. 884 * This is designed to be as fast as possible 885 * m stands for "measurement". 886 * 887 * On a 1990 paper the rto value is changed to: 888 * RTO = rtt + 4 * mdev 889 * 890 * Funny. This algorithm seems to be very broken. 891 * These formulae increase RTO, when it should be decreased, increase 892 * too slowly, when it should be increased quickly, decrease too quickly 893 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely 894 * does not matter how to _calculate_ it. Seems, it was trap 895 * that VJ failed to avoid. 8) 896 */ 897 if (srtt != 0) { 898 m -= (srtt >> 3); /* m is now error in rtt est */ 899 srtt += m; /* rtt = 7/8 rtt + 1/8 new */ 900 if (m < 0) { 901 m = -m; /* m is now abs(error) */ 902 m -= (tp->mdev_us >> 2); /* similar update on mdev */ 903 /* This is similar to one of Eifel findings. 904 * Eifel blocks mdev updates when rtt decreases. 905 * This solution is a bit different: we use finer gain 906 * for mdev in this case (alpha*beta). 907 * Like Eifel it also prevents growth of rto, 908 * but also it limits too fast rto decreases, 909 * happening in pure Eifel. 910 */ 911 if (m > 0) 912 m >>= 3; 913 } else { 914 m -= (tp->mdev_us >> 2); /* similar update on mdev */ 915 } 916 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */ 917 if (tp->mdev_us > tp->mdev_max_us) { 918 tp->mdev_max_us = tp->mdev_us; 919 if (tp->mdev_max_us > tp->rttvar_us) 920 tp->rttvar_us = tp->mdev_max_us; 921 } 922 if (after(tp->snd_una, tp->rtt_seq)) { 923 if (tp->mdev_max_us < tp->rttvar_us) 924 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2; 925 tp->rtt_seq = tp->snd_nxt; 926 tp->mdev_max_us = tcp_rto_min_us(sk); 927 928 tcp_bpf_rtt(sk, mrtt_us, srtt); 929 } 930 } else { 931 /* no previous measure. */ 932 srtt = m << 3; /* take the measured time to be rtt */ 933 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */ 934 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk)); 935 tp->mdev_max_us = tp->rttvar_us; 936 tp->rtt_seq = tp->snd_nxt; 937 938 tcp_bpf_rtt(sk, mrtt_us, srtt); 939 } 940 tp->srtt_us = max(1U, srtt); 941 } 942 943 static void tcp_update_pacing_rate(struct sock *sk) 944 { 945 const struct tcp_sock *tp = tcp_sk(sk); 946 u64 rate; 947 948 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */ 949 rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3); 950 951 /* current rate is (cwnd * mss) / srtt 952 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate. 953 * In Congestion Avoidance phase, set it to 120 % the current rate. 954 * 955 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh) 956 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching 957 * end of slow start and should slow down. 958 */ 959 if (tcp_snd_cwnd(tp) < tp->snd_ssthresh / 2) 960 rate *= READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_pacing_ss_ratio); 961 else 962 rate *= READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_pacing_ca_ratio); 963 964 rate *= max(tcp_snd_cwnd(tp), tp->packets_out); 965 966 if (likely(tp->srtt_us)) 967 do_div(rate, tp->srtt_us); 968 969 /* WRITE_ONCE() is needed because sch_fq fetches sk_pacing_rate 970 * without any lock. We want to make sure compiler wont store 971 * intermediate values in this location. 972 */ 973 WRITE_ONCE(sk->sk_pacing_rate, 974 min_t(u64, rate, READ_ONCE(sk->sk_max_pacing_rate))); 975 } 976 977 /* Calculate rto without backoff. This is the second half of Van Jacobson's 978 * routine referred to above. 979 */ 980 static void tcp_set_rto(struct sock *sk) 981 { 982 const struct tcp_sock *tp = tcp_sk(sk); 983 /* Old crap is replaced with new one. 8) 984 * 985 * More seriously: 986 * 1. If rtt variance happened to be less 50msec, it is hallucination. 987 * It cannot be less due to utterly erratic ACK generation made 988 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_ 989 * to do with delayed acks, because at cwnd>2 true delack timeout 990 * is invisible. Actually, Linux-2.4 also generates erratic 991 * ACKs in some circumstances. 992 */ 993 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp); 994 995 /* 2. Fixups made earlier cannot be right. 996 * If we do not estimate RTO correctly without them, 997 * all the algo is pure shit and should be replaced 998 * with correct one. It is exactly, which we pretend to do. 999 */ 1000 1001 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo 1002 * guarantees that rto is higher. 1003 */ 1004 tcp_bound_rto(sk); 1005 } 1006 1007 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst) 1008 { 1009 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0); 1010 1011 if (!cwnd) 1012 cwnd = TCP_INIT_CWND; 1013 return min_t(__u32, cwnd, tp->snd_cwnd_clamp); 1014 } 1015 1016 struct tcp_sacktag_state { 1017 /* Timestamps for earliest and latest never-retransmitted segment 1018 * that was SACKed. RTO needs the earliest RTT to stay conservative, 1019 * but congestion control should still get an accurate delay signal. 1020 */ 1021 u64 first_sackt; 1022 u64 last_sackt; 1023 u32 reord; 1024 u32 sack_delivered; 1025 int flag; 1026 unsigned int mss_now; 1027 struct rate_sample *rate; 1028 }; 1029 1030 /* Take a notice that peer is sending D-SACKs. Skip update of data delivery 1031 * and spurious retransmission information if this DSACK is unlikely caused by 1032 * sender's action: 1033 * - DSACKed sequence range is larger than maximum receiver's window. 1034 * - Total no. of DSACKed segments exceed the total no. of retransmitted segs. 1035 */ 1036 static u32 tcp_dsack_seen(struct tcp_sock *tp, u32 start_seq, 1037 u32 end_seq, struct tcp_sacktag_state *state) 1038 { 1039 u32 seq_len, dup_segs = 1; 1040 1041 if (!before(start_seq, end_seq)) 1042 return 0; 1043 1044 seq_len = end_seq - start_seq; 1045 /* Dubious DSACK: DSACKed range greater than maximum advertised rwnd */ 1046 if (seq_len > tp->max_window) 1047 return 0; 1048 if (seq_len > tp->mss_cache) 1049 dup_segs = DIV_ROUND_UP(seq_len, tp->mss_cache); 1050 else if (tp->tlp_high_seq && tp->tlp_high_seq == end_seq) 1051 state->flag |= FLAG_DSACK_TLP; 1052 1053 tp->dsack_dups += dup_segs; 1054 /* Skip the DSACK if dup segs weren't retransmitted by sender */ 1055 if (tp->dsack_dups > tp->total_retrans) 1056 return 0; 1057 1058 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN; 1059 /* We increase the RACK ordering window in rounds where we receive 1060 * DSACKs that may have been due to reordering causing RACK to trigger 1061 * a spurious fast recovery. Thus RACK ignores DSACKs that happen 1062 * without having seen reordering, or that match TLP probes (TLP 1063 * is timer-driven, not triggered by RACK). 1064 */ 1065 if (tp->reord_seen && !(state->flag & FLAG_DSACK_TLP)) 1066 tp->rack.dsack_seen = 1; 1067 1068 state->flag |= FLAG_DSACKING_ACK; 1069 /* A spurious retransmission is delivered */ 1070 state->sack_delivered += dup_segs; 1071 1072 return dup_segs; 1073 } 1074 1075 /* It's reordering when higher sequence was delivered (i.e. sacked) before 1076 * some lower never-retransmitted sequence ("low_seq"). The maximum reordering 1077 * distance is approximated in full-mss packet distance ("reordering"). 1078 */ 1079 static void tcp_check_sack_reordering(struct sock *sk, const u32 low_seq, 1080 const int ts) 1081 { 1082 struct tcp_sock *tp = tcp_sk(sk); 1083 const u32 mss = tp->mss_cache; 1084 u32 fack, metric; 1085 1086 fack = tcp_highest_sack_seq(tp); 1087 if (!before(low_seq, fack)) 1088 return; 1089 1090 metric = fack - low_seq; 1091 if ((metric > tp->reordering * mss) && mss) { 1092 #if FASTRETRANS_DEBUG > 1 1093 pr_debug("Disorder%d %d %u f%u s%u rr%d\n", 1094 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state, 1095 tp->reordering, 1096 0, 1097 tp->sacked_out, 1098 tp->undo_marker ? tp->undo_retrans : 0); 1099 #endif 1100 tp->reordering = min_t(u32, (metric + mss - 1) / mss, 1101 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_max_reordering)); 1102 } 1103 1104 /* This exciting event is worth to be remembered. 8) */ 1105 tp->reord_seen++; 1106 NET_INC_STATS(sock_net(sk), 1107 ts ? LINUX_MIB_TCPTSREORDER : LINUX_MIB_TCPSACKREORDER); 1108 } 1109 1110 /* This must be called before lost_out or retrans_out are updated 1111 * on a new loss, because we want to know if all skbs previously 1112 * known to be lost have already been retransmitted, indicating 1113 * that this newly lost skb is our next skb to retransmit. 1114 */ 1115 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb) 1116 { 1117 if ((!tp->retransmit_skb_hint && tp->retrans_out >= tp->lost_out) || 1118 (tp->retransmit_skb_hint && 1119 before(TCP_SKB_CB(skb)->seq, 1120 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))) 1121 tp->retransmit_skb_hint = skb; 1122 } 1123 1124 /* Sum the number of packets on the wire we have marked as lost, and 1125 * notify the congestion control module that the given skb was marked lost. 1126 */ 1127 static void tcp_notify_skb_loss_event(struct tcp_sock *tp, const struct sk_buff *skb) 1128 { 1129 tp->lost += tcp_skb_pcount(skb); 1130 } 1131 1132 void tcp_mark_skb_lost(struct sock *sk, struct sk_buff *skb) 1133 { 1134 __u8 sacked = TCP_SKB_CB(skb)->sacked; 1135 struct tcp_sock *tp = tcp_sk(sk); 1136 1137 if (sacked & TCPCB_SACKED_ACKED) 1138 return; 1139 1140 tcp_verify_retransmit_hint(tp, skb); 1141 if (sacked & TCPCB_LOST) { 1142 if (sacked & TCPCB_SACKED_RETRANS) { 1143 /* Account for retransmits that are lost again */ 1144 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 1145 tp->retrans_out -= tcp_skb_pcount(skb); 1146 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT, 1147 tcp_skb_pcount(skb)); 1148 tcp_notify_skb_loss_event(tp, skb); 1149 } 1150 } else { 1151 tp->lost_out += tcp_skb_pcount(skb); 1152 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 1153 tcp_notify_skb_loss_event(tp, skb); 1154 } 1155 } 1156 1157 /* Updates the delivered and delivered_ce counts */ 1158 static void tcp_count_delivered(struct tcp_sock *tp, u32 delivered, 1159 bool ece_ack) 1160 { 1161 tp->delivered += delivered; 1162 if (ece_ack) 1163 tp->delivered_ce += delivered; 1164 } 1165 1166 /* This procedure tags the retransmission queue when SACKs arrive. 1167 * 1168 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L). 1169 * Packets in queue with these bits set are counted in variables 1170 * sacked_out, retrans_out and lost_out, correspondingly. 1171 * 1172 * Valid combinations are: 1173 * Tag InFlight Description 1174 * 0 1 - orig segment is in flight. 1175 * S 0 - nothing flies, orig reached receiver. 1176 * L 0 - nothing flies, orig lost by net. 1177 * R 2 - both orig and retransmit are in flight. 1178 * L|R 1 - orig is lost, retransmit is in flight. 1179 * S|R 1 - orig reached receiver, retrans is still in flight. 1180 * (L|S|R is logically valid, it could occur when L|R is sacked, 1181 * but it is equivalent to plain S and code short-circuits it to S. 1182 * L|S is logically invalid, it would mean -1 packet in flight 8)) 1183 * 1184 * These 6 states form finite state machine, controlled by the following events: 1185 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue()) 1186 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue()) 1187 * 3. Loss detection event of two flavors: 1188 * A. Scoreboard estimator decided the packet is lost. 1189 * A'. Reno "three dupacks" marks head of queue lost. 1190 * B. SACK arrives sacking SND.NXT at the moment, when the 1191 * segment was retransmitted. 1192 * 4. D-SACK added new rule: D-SACK changes any tag to S. 1193 * 1194 * It is pleasant to note, that state diagram turns out to be commutative, 1195 * so that we are allowed not to be bothered by order of our actions, 1196 * when multiple events arrive simultaneously. (see the function below). 1197 * 1198 * Reordering detection. 1199 * -------------------- 1200 * Reordering metric is maximal distance, which a packet can be displaced 1201 * in packet stream. With SACKs we can estimate it: 1202 * 1203 * 1. SACK fills old hole and the corresponding segment was not 1204 * ever retransmitted -> reordering. Alas, we cannot use it 1205 * when segment was retransmitted. 1206 * 2. The last flaw is solved with D-SACK. D-SACK arrives 1207 * for retransmitted and already SACKed segment -> reordering.. 1208 * Both of these heuristics are not used in Loss state, when we cannot 1209 * account for retransmits accurately. 1210 * 1211 * SACK block validation. 1212 * ---------------------- 1213 * 1214 * SACK block range validation checks that the received SACK block fits to 1215 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT. 1216 * Note that SND.UNA is not included to the range though being valid because 1217 * it means that the receiver is rather inconsistent with itself reporting 1218 * SACK reneging when it should advance SND.UNA. Such SACK block this is 1219 * perfectly valid, however, in light of RFC2018 which explicitly states 1220 * that "SACK block MUST reflect the newest segment. Even if the newest 1221 * segment is going to be discarded ...", not that it looks very clever 1222 * in case of head skb. Due to potentional receiver driven attacks, we 1223 * choose to avoid immediate execution of a walk in write queue due to 1224 * reneging and defer head skb's loss recovery to standard loss recovery 1225 * procedure that will eventually trigger (nothing forbids us doing this). 1226 * 1227 * Implements also blockage to start_seq wrap-around. Problem lies in the 1228 * fact that though start_seq (s) is before end_seq (i.e., not reversed), 1229 * there's no guarantee that it will be before snd_nxt (n). The problem 1230 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt 1231 * wrap (s_w): 1232 * 1233 * <- outs wnd -> <- wrapzone -> 1234 * u e n u_w e_w s n_w 1235 * | | | | | | | 1236 * |<------------+------+----- TCP seqno space --------------+---------->| 1237 * ...-- <2^31 ->| |<--------... 1238 * ...---- >2^31 ------>| |<--------... 1239 * 1240 * Current code wouldn't be vulnerable but it's better still to discard such 1241 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat 1242 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in 1243 * snd_nxt wrap -> snd_una region will then become "well defined", i.e., 1244 * equal to the ideal case (infinite seqno space without wrap caused issues). 1245 * 1246 * With D-SACK the lower bound is extended to cover sequence space below 1247 * SND.UNA down to undo_marker, which is the last point of interest. Yet 1248 * again, D-SACK block must not to go across snd_una (for the same reason as 1249 * for the normal SACK blocks, explained above). But there all simplicity 1250 * ends, TCP might receive valid D-SACKs below that. As long as they reside 1251 * fully below undo_marker they do not affect behavior in anyway and can 1252 * therefore be safely ignored. In rare cases (which are more or less 1253 * theoretical ones), the D-SACK will nicely cross that boundary due to skb 1254 * fragmentation and packet reordering past skb's retransmission. To consider 1255 * them correctly, the acceptable range must be extended even more though 1256 * the exact amount is rather hard to quantify. However, tp->max_window can 1257 * be used as an exaggerated estimate. 1258 */ 1259 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack, 1260 u32 start_seq, u32 end_seq) 1261 { 1262 /* Too far in future, or reversed (interpretation is ambiguous) */ 1263 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq)) 1264 return false; 1265 1266 /* Nasty start_seq wrap-around check (see comments above) */ 1267 if (!before(start_seq, tp->snd_nxt)) 1268 return false; 1269 1270 /* In outstanding window? ...This is valid exit for D-SACKs too. 1271 * start_seq == snd_una is non-sensical (see comments above) 1272 */ 1273 if (after(start_seq, tp->snd_una)) 1274 return true; 1275 1276 if (!is_dsack || !tp->undo_marker) 1277 return false; 1278 1279 /* ...Then it's D-SACK, and must reside below snd_una completely */ 1280 if (after(end_seq, tp->snd_una)) 1281 return false; 1282 1283 if (!before(start_seq, tp->undo_marker)) 1284 return true; 1285 1286 /* Too old */ 1287 if (!after(end_seq, tp->undo_marker)) 1288 return false; 1289 1290 /* Undo_marker boundary crossing (overestimates a lot). Known already: 1291 * start_seq < undo_marker and end_seq >= undo_marker. 1292 */ 1293 return !before(start_seq, end_seq - tp->max_window); 1294 } 1295 1296 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb, 1297 struct tcp_sack_block_wire *sp, int num_sacks, 1298 u32 prior_snd_una, struct tcp_sacktag_state *state) 1299 { 1300 struct tcp_sock *tp = tcp_sk(sk); 1301 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq); 1302 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq); 1303 u32 dup_segs; 1304 1305 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) { 1306 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECV); 1307 } else if (num_sacks > 1) { 1308 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq); 1309 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq); 1310 1311 if (after(end_seq_0, end_seq_1) || before(start_seq_0, start_seq_1)) 1312 return false; 1313 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKOFORECV); 1314 } else { 1315 return false; 1316 } 1317 1318 dup_segs = tcp_dsack_seen(tp, start_seq_0, end_seq_0, state); 1319 if (!dup_segs) { /* Skip dubious DSACK */ 1320 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKIGNOREDDUBIOUS); 1321 return false; 1322 } 1323 1324 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECVSEGS, dup_segs); 1325 1326 /* D-SACK for already forgotten data... Do dumb counting. */ 1327 if (tp->undo_marker && tp->undo_retrans > 0 && 1328 !after(end_seq_0, prior_snd_una) && 1329 after(end_seq_0, tp->undo_marker)) 1330 tp->undo_retrans = max_t(int, 0, tp->undo_retrans - dup_segs); 1331 1332 return true; 1333 } 1334 1335 /* Check if skb is fully within the SACK block. In presence of GSO skbs, 1336 * the incoming SACK may not exactly match but we can find smaller MSS 1337 * aligned portion of it that matches. Therefore we might need to fragment 1338 * which may fail and creates some hassle (caller must handle error case 1339 * returns). 1340 * 1341 * FIXME: this could be merged to shift decision code 1342 */ 1343 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb, 1344 u32 start_seq, u32 end_seq) 1345 { 1346 int err; 1347 bool in_sack; 1348 unsigned int pkt_len; 1349 unsigned int mss; 1350 1351 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && 1352 !before(end_seq, TCP_SKB_CB(skb)->end_seq); 1353 1354 if (tcp_skb_pcount(skb) > 1 && !in_sack && 1355 after(TCP_SKB_CB(skb)->end_seq, start_seq)) { 1356 mss = tcp_skb_mss(skb); 1357 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); 1358 1359 if (!in_sack) { 1360 pkt_len = start_seq - TCP_SKB_CB(skb)->seq; 1361 if (pkt_len < mss) 1362 pkt_len = mss; 1363 } else { 1364 pkt_len = end_seq - TCP_SKB_CB(skb)->seq; 1365 if (pkt_len < mss) 1366 return -EINVAL; 1367 } 1368 1369 /* Round if necessary so that SACKs cover only full MSSes 1370 * and/or the remaining small portion (if present) 1371 */ 1372 if (pkt_len > mss) { 1373 unsigned int new_len = (pkt_len / mss) * mss; 1374 if (!in_sack && new_len < pkt_len) 1375 new_len += mss; 1376 pkt_len = new_len; 1377 } 1378 1379 if (pkt_len >= skb->len && !in_sack) 1380 return 0; 1381 1382 err = tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, 1383 pkt_len, mss, GFP_ATOMIC); 1384 if (err < 0) 1385 return err; 1386 } 1387 1388 return in_sack; 1389 } 1390 1391 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */ 1392 static u8 tcp_sacktag_one(struct sock *sk, 1393 struct tcp_sacktag_state *state, u8 sacked, 1394 u32 start_seq, u32 end_seq, 1395 int dup_sack, int pcount, 1396 u64 xmit_time) 1397 { 1398 struct tcp_sock *tp = tcp_sk(sk); 1399 1400 /* Account D-SACK for retransmitted packet. */ 1401 if (dup_sack && (sacked & TCPCB_RETRANS)) { 1402 if (tp->undo_marker && tp->undo_retrans > 0 && 1403 after(end_seq, tp->undo_marker)) 1404 tp->undo_retrans = max_t(int, 0, tp->undo_retrans - pcount); 1405 if ((sacked & TCPCB_SACKED_ACKED) && 1406 before(start_seq, state->reord)) 1407 state->reord = start_seq; 1408 } 1409 1410 /* Nothing to do; acked frame is about to be dropped (was ACKed). */ 1411 if (!after(end_seq, tp->snd_una)) 1412 return sacked; 1413 1414 if (!(sacked & TCPCB_SACKED_ACKED)) { 1415 tcp_rack_advance(tp, sacked, end_seq, xmit_time); 1416 1417 if (sacked & TCPCB_SACKED_RETRANS) { 1418 /* If the segment is not tagged as lost, 1419 * we do not clear RETRANS, believing 1420 * that retransmission is still in flight. 1421 */ 1422 if (sacked & TCPCB_LOST) { 1423 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS); 1424 tp->lost_out -= pcount; 1425 tp->retrans_out -= pcount; 1426 } 1427 } else { 1428 if (!(sacked & TCPCB_RETRANS)) { 1429 /* New sack for not retransmitted frame, 1430 * which was in hole. It is reordering. 1431 */ 1432 if (before(start_seq, 1433 tcp_highest_sack_seq(tp)) && 1434 before(start_seq, state->reord)) 1435 state->reord = start_seq; 1436 1437 if (!after(end_seq, tp->high_seq)) 1438 state->flag |= FLAG_ORIG_SACK_ACKED; 1439 if (state->first_sackt == 0) 1440 state->first_sackt = xmit_time; 1441 state->last_sackt = xmit_time; 1442 } 1443 1444 if (sacked & TCPCB_LOST) { 1445 sacked &= ~TCPCB_LOST; 1446 tp->lost_out -= pcount; 1447 } 1448 } 1449 1450 sacked |= TCPCB_SACKED_ACKED; 1451 state->flag |= FLAG_DATA_SACKED; 1452 tp->sacked_out += pcount; 1453 /* Out-of-order packets delivered */ 1454 state->sack_delivered += pcount; 1455 1456 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */ 1457 if (tp->lost_skb_hint && 1458 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq)) 1459 tp->lost_cnt_hint += pcount; 1460 } 1461 1462 /* D-SACK. We can detect redundant retransmission in S|R and plain R 1463 * frames and clear it. undo_retrans is decreased above, L|R frames 1464 * are accounted above as well. 1465 */ 1466 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) { 1467 sacked &= ~TCPCB_SACKED_RETRANS; 1468 tp->retrans_out -= pcount; 1469 } 1470 1471 return sacked; 1472 } 1473 1474 /* Shift newly-SACKed bytes from this skb to the immediately previous 1475 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such. 1476 */ 1477 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *prev, 1478 struct sk_buff *skb, 1479 struct tcp_sacktag_state *state, 1480 unsigned int pcount, int shifted, int mss, 1481 bool dup_sack) 1482 { 1483 struct tcp_sock *tp = tcp_sk(sk); 1484 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */ 1485 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */ 1486 1487 BUG_ON(!pcount); 1488 1489 /* Adjust counters and hints for the newly sacked sequence 1490 * range but discard the return value since prev is already 1491 * marked. We must tag the range first because the seq 1492 * advancement below implicitly advances 1493 * tcp_highest_sack_seq() when skb is highest_sack. 1494 */ 1495 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked, 1496 start_seq, end_seq, dup_sack, pcount, 1497 tcp_skb_timestamp_us(skb)); 1498 tcp_rate_skb_delivered(sk, skb, state->rate); 1499 1500 if (skb == tp->lost_skb_hint) 1501 tp->lost_cnt_hint += pcount; 1502 1503 TCP_SKB_CB(prev)->end_seq += shifted; 1504 TCP_SKB_CB(skb)->seq += shifted; 1505 1506 tcp_skb_pcount_add(prev, pcount); 1507 WARN_ON_ONCE(tcp_skb_pcount(skb) < pcount); 1508 tcp_skb_pcount_add(skb, -pcount); 1509 1510 /* When we're adding to gso_segs == 1, gso_size will be zero, 1511 * in theory this shouldn't be necessary but as long as DSACK 1512 * code can come after this skb later on it's better to keep 1513 * setting gso_size to something. 1514 */ 1515 if (!TCP_SKB_CB(prev)->tcp_gso_size) 1516 TCP_SKB_CB(prev)->tcp_gso_size = mss; 1517 1518 /* CHECKME: To clear or not to clear? Mimics normal skb currently */ 1519 if (tcp_skb_pcount(skb) <= 1) 1520 TCP_SKB_CB(skb)->tcp_gso_size = 0; 1521 1522 /* Difference in this won't matter, both ACKed by the same cumul. ACK */ 1523 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS); 1524 1525 if (skb->len > 0) { 1526 BUG_ON(!tcp_skb_pcount(skb)); 1527 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTED); 1528 return false; 1529 } 1530 1531 /* Whole SKB was eaten :-) */ 1532 1533 if (skb == tp->retransmit_skb_hint) 1534 tp->retransmit_skb_hint = prev; 1535 if (skb == tp->lost_skb_hint) { 1536 tp->lost_skb_hint = prev; 1537 tp->lost_cnt_hint -= tcp_skb_pcount(prev); 1538 } 1539 1540 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; 1541 TCP_SKB_CB(prev)->eor = TCP_SKB_CB(skb)->eor; 1542 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 1543 TCP_SKB_CB(prev)->end_seq++; 1544 1545 if (skb == tcp_highest_sack(sk)) 1546 tcp_advance_highest_sack(sk, skb); 1547 1548 tcp_skb_collapse_tstamp(prev, skb); 1549 if (unlikely(TCP_SKB_CB(prev)->tx.delivered_mstamp)) 1550 TCP_SKB_CB(prev)->tx.delivered_mstamp = 0; 1551 1552 tcp_rtx_queue_unlink_and_free(skb, sk); 1553 1554 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKMERGED); 1555 1556 return true; 1557 } 1558 1559 /* I wish gso_size would have a bit more sane initialization than 1560 * something-or-zero which complicates things 1561 */ 1562 static int tcp_skb_seglen(const struct sk_buff *skb) 1563 { 1564 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb); 1565 } 1566 1567 /* Shifting pages past head area doesn't work */ 1568 static int skb_can_shift(const struct sk_buff *skb) 1569 { 1570 return !skb_headlen(skb) && skb_is_nonlinear(skb); 1571 } 1572 1573 int tcp_skb_shift(struct sk_buff *to, struct sk_buff *from, 1574 int pcount, int shiftlen) 1575 { 1576 /* TCP min gso_size is 8 bytes (TCP_MIN_GSO_SIZE) 1577 * Since TCP_SKB_CB(skb)->tcp_gso_segs is 16 bits, we need 1578 * to make sure not storing more than 65535 * 8 bytes per skb, 1579 * even if current MSS is bigger. 1580 */ 1581 if (unlikely(to->len + shiftlen >= 65535 * TCP_MIN_GSO_SIZE)) 1582 return 0; 1583 if (unlikely(tcp_skb_pcount(to) + pcount > 65535)) 1584 return 0; 1585 return skb_shift(to, from, shiftlen); 1586 } 1587 1588 /* Try collapsing SACK blocks spanning across multiple skbs to a single 1589 * skb. 1590 */ 1591 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb, 1592 struct tcp_sacktag_state *state, 1593 u32 start_seq, u32 end_seq, 1594 bool dup_sack) 1595 { 1596 struct tcp_sock *tp = tcp_sk(sk); 1597 struct sk_buff *prev; 1598 int mss; 1599 int pcount = 0; 1600 int len; 1601 int in_sack; 1602 1603 /* Normally R but no L won't result in plain S */ 1604 if (!dup_sack && 1605 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS) 1606 goto fallback; 1607 if (!skb_can_shift(skb)) 1608 goto fallback; 1609 /* This frame is about to be dropped (was ACKed). */ 1610 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 1611 goto fallback; 1612 1613 /* Can only happen with delayed DSACK + discard craziness */ 1614 prev = skb_rb_prev(skb); 1615 if (!prev) 1616 goto fallback; 1617 1618 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) 1619 goto fallback; 1620 1621 if (!tcp_skb_can_collapse(prev, skb)) 1622 goto fallback; 1623 1624 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && 1625 !before(end_seq, TCP_SKB_CB(skb)->end_seq); 1626 1627 if (in_sack) { 1628 len = skb->len; 1629 pcount = tcp_skb_pcount(skb); 1630 mss = tcp_skb_seglen(skb); 1631 1632 /* TODO: Fix DSACKs to not fragment already SACKed and we can 1633 * drop this restriction as unnecessary 1634 */ 1635 if (mss != tcp_skb_seglen(prev)) 1636 goto fallback; 1637 } else { 1638 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq)) 1639 goto noop; 1640 /* CHECKME: This is non-MSS split case only?, this will 1641 * cause skipped skbs due to advancing loop btw, original 1642 * has that feature too 1643 */ 1644 if (tcp_skb_pcount(skb) <= 1) 1645 goto noop; 1646 1647 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); 1648 if (!in_sack) { 1649 /* TODO: head merge to next could be attempted here 1650 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)), 1651 * though it might not be worth of the additional hassle 1652 * 1653 * ...we can probably just fallback to what was done 1654 * previously. We could try merging non-SACKed ones 1655 * as well but it probably isn't going to buy off 1656 * because later SACKs might again split them, and 1657 * it would make skb timestamp tracking considerably 1658 * harder problem. 1659 */ 1660 goto fallback; 1661 } 1662 1663 len = end_seq - TCP_SKB_CB(skb)->seq; 1664 BUG_ON(len < 0); 1665 BUG_ON(len > skb->len); 1666 1667 /* MSS boundaries should be honoured or else pcount will 1668 * severely break even though it makes things bit trickier. 1669 * Optimize common case to avoid most of the divides 1670 */ 1671 mss = tcp_skb_mss(skb); 1672 1673 /* TODO: Fix DSACKs to not fragment already SACKed and we can 1674 * drop this restriction as unnecessary 1675 */ 1676 if (mss != tcp_skb_seglen(prev)) 1677 goto fallback; 1678 1679 if (len == mss) { 1680 pcount = 1; 1681 } else if (len < mss) { 1682 goto noop; 1683 } else { 1684 pcount = len / mss; 1685 len = pcount * mss; 1686 } 1687 } 1688 1689 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */ 1690 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una)) 1691 goto fallback; 1692 1693 if (!tcp_skb_shift(prev, skb, pcount, len)) 1694 goto fallback; 1695 if (!tcp_shifted_skb(sk, prev, skb, state, pcount, len, mss, dup_sack)) 1696 goto out; 1697 1698 /* Hole filled allows collapsing with the next as well, this is very 1699 * useful when hole on every nth skb pattern happens 1700 */ 1701 skb = skb_rb_next(prev); 1702 if (!skb) 1703 goto out; 1704 1705 if (!skb_can_shift(skb) || 1706 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) || 1707 (mss != tcp_skb_seglen(skb))) 1708 goto out; 1709 1710 if (!tcp_skb_can_collapse(prev, skb)) 1711 goto out; 1712 len = skb->len; 1713 pcount = tcp_skb_pcount(skb); 1714 if (tcp_skb_shift(prev, skb, pcount, len)) 1715 tcp_shifted_skb(sk, prev, skb, state, pcount, 1716 len, mss, 0); 1717 1718 out: 1719 return prev; 1720 1721 noop: 1722 return skb; 1723 1724 fallback: 1725 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK); 1726 return NULL; 1727 } 1728 1729 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk, 1730 struct tcp_sack_block *next_dup, 1731 struct tcp_sacktag_state *state, 1732 u32 start_seq, u32 end_seq, 1733 bool dup_sack_in) 1734 { 1735 struct tcp_sock *tp = tcp_sk(sk); 1736 struct sk_buff *tmp; 1737 1738 skb_rbtree_walk_from(skb) { 1739 int in_sack = 0; 1740 bool dup_sack = dup_sack_in; 1741 1742 /* queue is in-order => we can short-circuit the walk early */ 1743 if (!before(TCP_SKB_CB(skb)->seq, end_seq)) 1744 break; 1745 1746 if (next_dup && 1747 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) { 1748 in_sack = tcp_match_skb_to_sack(sk, skb, 1749 next_dup->start_seq, 1750 next_dup->end_seq); 1751 if (in_sack > 0) 1752 dup_sack = true; 1753 } 1754 1755 /* skb reference here is a bit tricky to get right, since 1756 * shifting can eat and free both this skb and the next, 1757 * so not even _safe variant of the loop is enough. 1758 */ 1759 if (in_sack <= 0) { 1760 tmp = tcp_shift_skb_data(sk, skb, state, 1761 start_seq, end_seq, dup_sack); 1762 if (tmp) { 1763 if (tmp != skb) { 1764 skb = tmp; 1765 continue; 1766 } 1767 1768 in_sack = 0; 1769 } else { 1770 in_sack = tcp_match_skb_to_sack(sk, skb, 1771 start_seq, 1772 end_seq); 1773 } 1774 } 1775 1776 if (unlikely(in_sack < 0)) 1777 break; 1778 1779 if (in_sack) { 1780 TCP_SKB_CB(skb)->sacked = 1781 tcp_sacktag_one(sk, 1782 state, 1783 TCP_SKB_CB(skb)->sacked, 1784 TCP_SKB_CB(skb)->seq, 1785 TCP_SKB_CB(skb)->end_seq, 1786 dup_sack, 1787 tcp_skb_pcount(skb), 1788 tcp_skb_timestamp_us(skb)); 1789 tcp_rate_skb_delivered(sk, skb, state->rate); 1790 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 1791 list_del_init(&skb->tcp_tsorted_anchor); 1792 1793 if (!before(TCP_SKB_CB(skb)->seq, 1794 tcp_highest_sack_seq(tp))) 1795 tcp_advance_highest_sack(sk, skb); 1796 } 1797 } 1798 return skb; 1799 } 1800 1801 static struct sk_buff *tcp_sacktag_bsearch(struct sock *sk, u32 seq) 1802 { 1803 struct rb_node *parent, **p = &sk->tcp_rtx_queue.rb_node; 1804 struct sk_buff *skb; 1805 1806 while (*p) { 1807 parent = *p; 1808 skb = rb_to_skb(parent); 1809 if (before(seq, TCP_SKB_CB(skb)->seq)) { 1810 p = &parent->rb_left; 1811 continue; 1812 } 1813 if (!before(seq, TCP_SKB_CB(skb)->end_seq)) { 1814 p = &parent->rb_right; 1815 continue; 1816 } 1817 return skb; 1818 } 1819 return NULL; 1820 } 1821 1822 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk, 1823 u32 skip_to_seq) 1824 { 1825 if (skb && after(TCP_SKB_CB(skb)->seq, skip_to_seq)) 1826 return skb; 1827 1828 return tcp_sacktag_bsearch(sk, skip_to_seq); 1829 } 1830 1831 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb, 1832 struct sock *sk, 1833 struct tcp_sack_block *next_dup, 1834 struct tcp_sacktag_state *state, 1835 u32 skip_to_seq) 1836 { 1837 if (!next_dup) 1838 return skb; 1839 1840 if (before(next_dup->start_seq, skip_to_seq)) { 1841 skb = tcp_sacktag_skip(skb, sk, next_dup->start_seq); 1842 skb = tcp_sacktag_walk(skb, sk, NULL, state, 1843 next_dup->start_seq, next_dup->end_seq, 1844 1); 1845 } 1846 1847 return skb; 1848 } 1849 1850 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache) 1851 { 1852 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache); 1853 } 1854 1855 static int 1856 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb, 1857 u32 prior_snd_una, struct tcp_sacktag_state *state) 1858 { 1859 struct tcp_sock *tp = tcp_sk(sk); 1860 const unsigned char *ptr = (skb_transport_header(ack_skb) + 1861 TCP_SKB_CB(ack_skb)->sacked); 1862 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2); 1863 struct tcp_sack_block sp[TCP_NUM_SACKS]; 1864 struct tcp_sack_block *cache; 1865 struct sk_buff *skb; 1866 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3); 1867 int used_sacks; 1868 bool found_dup_sack = false; 1869 int i, j; 1870 int first_sack_index; 1871 1872 state->flag = 0; 1873 state->reord = tp->snd_nxt; 1874 1875 if (!tp->sacked_out) 1876 tcp_highest_sack_reset(sk); 1877 1878 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire, 1879 num_sacks, prior_snd_una, state); 1880 1881 /* Eliminate too old ACKs, but take into 1882 * account more or less fresh ones, they can 1883 * contain valid SACK info. 1884 */ 1885 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window)) 1886 return 0; 1887 1888 if (!tp->packets_out) 1889 goto out; 1890 1891 used_sacks = 0; 1892 first_sack_index = 0; 1893 for (i = 0; i < num_sacks; i++) { 1894 bool dup_sack = !i && found_dup_sack; 1895 1896 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq); 1897 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq); 1898 1899 if (!tcp_is_sackblock_valid(tp, dup_sack, 1900 sp[used_sacks].start_seq, 1901 sp[used_sacks].end_seq)) { 1902 int mib_idx; 1903 1904 if (dup_sack) { 1905 if (!tp->undo_marker) 1906 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO; 1907 else 1908 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD; 1909 } else { 1910 /* Don't count olds caused by ACK reordering */ 1911 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) && 1912 !after(sp[used_sacks].end_seq, tp->snd_una)) 1913 continue; 1914 mib_idx = LINUX_MIB_TCPSACKDISCARD; 1915 } 1916 1917 NET_INC_STATS(sock_net(sk), mib_idx); 1918 if (i == 0) 1919 first_sack_index = -1; 1920 continue; 1921 } 1922 1923 /* Ignore very old stuff early */ 1924 if (!after(sp[used_sacks].end_seq, prior_snd_una)) { 1925 if (i == 0) 1926 first_sack_index = -1; 1927 continue; 1928 } 1929 1930 used_sacks++; 1931 } 1932 1933 /* order SACK blocks to allow in order walk of the retrans queue */ 1934 for (i = used_sacks - 1; i > 0; i--) { 1935 for (j = 0; j < i; j++) { 1936 if (after(sp[j].start_seq, sp[j + 1].start_seq)) { 1937 swap(sp[j], sp[j + 1]); 1938 1939 /* Track where the first SACK block goes to */ 1940 if (j == first_sack_index) 1941 first_sack_index = j + 1; 1942 } 1943 } 1944 } 1945 1946 state->mss_now = tcp_current_mss(sk); 1947 skb = NULL; 1948 i = 0; 1949 1950 if (!tp->sacked_out) { 1951 /* It's already past, so skip checking against it */ 1952 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache); 1953 } else { 1954 cache = tp->recv_sack_cache; 1955 /* Skip empty blocks in at head of the cache */ 1956 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq && 1957 !cache->end_seq) 1958 cache++; 1959 } 1960 1961 while (i < used_sacks) { 1962 u32 start_seq = sp[i].start_seq; 1963 u32 end_seq = sp[i].end_seq; 1964 bool dup_sack = (found_dup_sack && (i == first_sack_index)); 1965 struct tcp_sack_block *next_dup = NULL; 1966 1967 if (found_dup_sack && ((i + 1) == first_sack_index)) 1968 next_dup = &sp[i + 1]; 1969 1970 /* Skip too early cached blocks */ 1971 while (tcp_sack_cache_ok(tp, cache) && 1972 !before(start_seq, cache->end_seq)) 1973 cache++; 1974 1975 /* Can skip some work by looking recv_sack_cache? */ 1976 if (tcp_sack_cache_ok(tp, cache) && !dup_sack && 1977 after(end_seq, cache->start_seq)) { 1978 1979 /* Head todo? */ 1980 if (before(start_seq, cache->start_seq)) { 1981 skb = tcp_sacktag_skip(skb, sk, start_seq); 1982 skb = tcp_sacktag_walk(skb, sk, next_dup, 1983 state, 1984 start_seq, 1985 cache->start_seq, 1986 dup_sack); 1987 } 1988 1989 /* Rest of the block already fully processed? */ 1990 if (!after(end_seq, cache->end_seq)) 1991 goto advance_sp; 1992 1993 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup, 1994 state, 1995 cache->end_seq); 1996 1997 /* ...tail remains todo... */ 1998 if (tcp_highest_sack_seq(tp) == cache->end_seq) { 1999 /* ...but better entrypoint exists! */ 2000 skb = tcp_highest_sack(sk); 2001 if (!skb) 2002 break; 2003 cache++; 2004 goto walk; 2005 } 2006 2007 skb = tcp_sacktag_skip(skb, sk, cache->end_seq); 2008 /* Check overlap against next cached too (past this one already) */ 2009 cache++; 2010 continue; 2011 } 2012 2013 if (!before(start_seq, tcp_highest_sack_seq(tp))) { 2014 skb = tcp_highest_sack(sk); 2015 if (!skb) 2016 break; 2017 } 2018 skb = tcp_sacktag_skip(skb, sk, start_seq); 2019 2020 walk: 2021 skb = tcp_sacktag_walk(skb, sk, next_dup, state, 2022 start_seq, end_seq, dup_sack); 2023 2024 advance_sp: 2025 i++; 2026 } 2027 2028 /* Clear the head of the cache sack blocks so we can skip it next time */ 2029 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) { 2030 tp->recv_sack_cache[i].start_seq = 0; 2031 tp->recv_sack_cache[i].end_seq = 0; 2032 } 2033 for (j = 0; j < used_sacks; j++) 2034 tp->recv_sack_cache[i++] = sp[j]; 2035 2036 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Loss || tp->undo_marker) 2037 tcp_check_sack_reordering(sk, state->reord, 0); 2038 2039 tcp_verify_left_out(tp); 2040 out: 2041 2042 #if FASTRETRANS_DEBUG > 0 2043 WARN_ON((int)tp->sacked_out < 0); 2044 WARN_ON((int)tp->lost_out < 0); 2045 WARN_ON((int)tp->retrans_out < 0); 2046 WARN_ON((int)tcp_packets_in_flight(tp) < 0); 2047 #endif 2048 return state->flag; 2049 } 2050 2051 /* Limits sacked_out so that sum with lost_out isn't ever larger than 2052 * packets_out. Returns false if sacked_out adjustement wasn't necessary. 2053 */ 2054 static bool tcp_limit_reno_sacked(struct tcp_sock *tp) 2055 { 2056 u32 holes; 2057 2058 holes = max(tp->lost_out, 1U); 2059 holes = min(holes, tp->packets_out); 2060 2061 if ((tp->sacked_out + holes) > tp->packets_out) { 2062 tp->sacked_out = tp->packets_out - holes; 2063 return true; 2064 } 2065 return false; 2066 } 2067 2068 /* If we receive more dupacks than we expected counting segments 2069 * in assumption of absent reordering, interpret this as reordering. 2070 * The only another reason could be bug in receiver TCP. 2071 */ 2072 static void tcp_check_reno_reordering(struct sock *sk, const int addend) 2073 { 2074 struct tcp_sock *tp = tcp_sk(sk); 2075 2076 if (!tcp_limit_reno_sacked(tp)) 2077 return; 2078 2079 tp->reordering = min_t(u32, tp->packets_out + addend, 2080 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_max_reordering)); 2081 tp->reord_seen++; 2082 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRENOREORDER); 2083 } 2084 2085 /* Emulate SACKs for SACKless connection: account for a new dupack. */ 2086 2087 static void tcp_add_reno_sack(struct sock *sk, int num_dupack, bool ece_ack) 2088 { 2089 if (num_dupack) { 2090 struct tcp_sock *tp = tcp_sk(sk); 2091 u32 prior_sacked = tp->sacked_out; 2092 s32 delivered; 2093 2094 tp->sacked_out += num_dupack; 2095 tcp_check_reno_reordering(sk, 0); 2096 delivered = tp->sacked_out - prior_sacked; 2097 if (delivered > 0) 2098 tcp_count_delivered(tp, delivered, ece_ack); 2099 tcp_verify_left_out(tp); 2100 } 2101 } 2102 2103 /* Account for ACK, ACKing some data in Reno Recovery phase. */ 2104 2105 static void tcp_remove_reno_sacks(struct sock *sk, int acked, bool ece_ack) 2106 { 2107 struct tcp_sock *tp = tcp_sk(sk); 2108 2109 if (acked > 0) { 2110 /* One ACK acked hole. The rest eat duplicate ACKs. */ 2111 tcp_count_delivered(tp, max_t(int, acked - tp->sacked_out, 1), 2112 ece_ack); 2113 if (acked - 1 >= tp->sacked_out) 2114 tp->sacked_out = 0; 2115 else 2116 tp->sacked_out -= acked - 1; 2117 } 2118 tcp_check_reno_reordering(sk, acked); 2119 tcp_verify_left_out(tp); 2120 } 2121 2122 static inline void tcp_reset_reno_sack(struct tcp_sock *tp) 2123 { 2124 tp->sacked_out = 0; 2125 } 2126 2127 void tcp_clear_retrans(struct tcp_sock *tp) 2128 { 2129 tp->retrans_out = 0; 2130 tp->lost_out = 0; 2131 tp->undo_marker = 0; 2132 tp->undo_retrans = -1; 2133 tp->sacked_out = 0; 2134 tp->rto_stamp = 0; 2135 tp->total_rto = 0; 2136 tp->total_rto_recoveries = 0; 2137 tp->total_rto_time = 0; 2138 } 2139 2140 static inline void tcp_init_undo(struct tcp_sock *tp) 2141 { 2142 tp->undo_marker = tp->snd_una; 2143 2144 /* Retransmission still in flight may cause DSACKs later. */ 2145 /* First, account for regular retransmits in flight: */ 2146 tp->undo_retrans = tp->retrans_out; 2147 /* Next, account for TLP retransmits in flight: */ 2148 if (tp->tlp_high_seq && tp->tlp_retrans) 2149 tp->undo_retrans++; 2150 /* Finally, avoid 0, because undo_retrans==0 means "can undo now": */ 2151 if (!tp->undo_retrans) 2152 tp->undo_retrans = -1; 2153 } 2154 2155 static bool tcp_is_rack(const struct sock *sk) 2156 { 2157 return READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_recovery) & 2158 TCP_RACK_LOSS_DETECTION; 2159 } 2160 2161 /* If we detect SACK reneging, forget all SACK information 2162 * and reset tags completely, otherwise preserve SACKs. If receiver 2163 * dropped its ofo queue, we will know this due to reneging detection. 2164 */ 2165 static void tcp_timeout_mark_lost(struct sock *sk) 2166 { 2167 struct tcp_sock *tp = tcp_sk(sk); 2168 struct sk_buff *skb, *head; 2169 bool is_reneg; /* is receiver reneging on SACKs? */ 2170 2171 head = tcp_rtx_queue_head(sk); 2172 is_reneg = head && (TCP_SKB_CB(head)->sacked & TCPCB_SACKED_ACKED); 2173 if (is_reneg) { 2174 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSACKRENEGING); 2175 tp->sacked_out = 0; 2176 /* Mark SACK reneging until we recover from this loss event. */ 2177 tp->is_sack_reneg = 1; 2178 } else if (tcp_is_reno(tp)) { 2179 tcp_reset_reno_sack(tp); 2180 } 2181 2182 skb = head; 2183 skb_rbtree_walk_from(skb) { 2184 if (is_reneg) 2185 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED; 2186 else if (tcp_is_rack(sk) && skb != head && 2187 tcp_rack_skb_timeout(tp, skb, 0) > 0) 2188 continue; /* Don't mark recently sent ones lost yet */ 2189 tcp_mark_skb_lost(sk, skb); 2190 } 2191 tcp_verify_left_out(tp); 2192 tcp_clear_all_retrans_hints(tp); 2193 } 2194 2195 /* Enter Loss state. */ 2196 void tcp_enter_loss(struct sock *sk) 2197 { 2198 const struct inet_connection_sock *icsk = inet_csk(sk); 2199 struct tcp_sock *tp = tcp_sk(sk); 2200 struct net *net = sock_net(sk); 2201 bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery; 2202 u8 reordering; 2203 2204 tcp_timeout_mark_lost(sk); 2205 2206 /* Reduce ssthresh if it has not yet been made inside this window. */ 2207 if (icsk->icsk_ca_state <= TCP_CA_Disorder || 2208 !after(tp->high_seq, tp->snd_una) || 2209 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) { 2210 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2211 tp->prior_cwnd = tcp_snd_cwnd(tp); 2212 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 2213 tcp_ca_event(sk, CA_EVENT_LOSS); 2214 tcp_init_undo(tp); 2215 } 2216 tcp_snd_cwnd_set(tp, tcp_packets_in_flight(tp) + 1); 2217 tp->snd_cwnd_cnt = 0; 2218 tp->snd_cwnd_stamp = tcp_jiffies32; 2219 2220 /* Timeout in disordered state after receiving substantial DUPACKs 2221 * suggests that the degree of reordering is over-estimated. 2222 */ 2223 reordering = READ_ONCE(net->ipv4.sysctl_tcp_reordering); 2224 if (icsk->icsk_ca_state <= TCP_CA_Disorder && 2225 tp->sacked_out >= reordering) 2226 tp->reordering = min_t(unsigned int, tp->reordering, 2227 reordering); 2228 2229 tcp_set_ca_state(sk, TCP_CA_Loss); 2230 tp->high_seq = tp->snd_nxt; 2231 tp->tlp_high_seq = 0; 2232 tcp_ecn_queue_cwr(tp); 2233 2234 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous 2235 * loss recovery is underway except recurring timeout(s) on 2236 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing 2237 */ 2238 tp->frto = READ_ONCE(net->ipv4.sysctl_tcp_frto) && 2239 (new_recovery || icsk->icsk_retransmits) && 2240 !inet_csk(sk)->icsk_mtup.probe_size; 2241 } 2242 2243 /* If ACK arrived pointing to a remembered SACK, it means that our 2244 * remembered SACKs do not reflect real state of receiver i.e. 2245 * receiver _host_ is heavily congested (or buggy). 2246 * 2247 * To avoid big spurious retransmission bursts due to transient SACK 2248 * scoreboard oddities that look like reneging, we give the receiver a 2249 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will 2250 * restore sanity to the SACK scoreboard. If the apparent reneging 2251 * persists until this RTO then we'll clear the SACK scoreboard. 2252 */ 2253 static bool tcp_check_sack_reneging(struct sock *sk, int *ack_flag) 2254 { 2255 if (*ack_flag & FLAG_SACK_RENEGING && 2256 *ack_flag & FLAG_SND_UNA_ADVANCED) { 2257 struct tcp_sock *tp = tcp_sk(sk); 2258 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4), 2259 msecs_to_jiffies(10)); 2260 2261 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 2262 delay, TCP_RTO_MAX); 2263 *ack_flag &= ~FLAG_SET_XMIT_TIMER; 2264 return true; 2265 } 2266 return false; 2267 } 2268 2269 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs 2270 * counter when SACK is enabled (without SACK, sacked_out is used for 2271 * that purpose). 2272 * 2273 * With reordering, holes may still be in flight, so RFC3517 recovery 2274 * uses pure sacked_out (total number of SACKed segments) even though 2275 * it violates the RFC that uses duplicate ACKs, often these are equal 2276 * but when e.g. out-of-window ACKs or packet duplication occurs, 2277 * they differ. Since neither occurs due to loss, TCP should really 2278 * ignore them. 2279 */ 2280 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp) 2281 { 2282 return tp->sacked_out + 1; 2283 } 2284 2285 /* Linux NewReno/SACK/ECN state machine. 2286 * -------------------------------------- 2287 * 2288 * "Open" Normal state, no dubious events, fast path. 2289 * "Disorder" In all the respects it is "Open", 2290 * but requires a bit more attention. It is entered when 2291 * we see some SACKs or dupacks. It is split of "Open" 2292 * mainly to move some processing from fast path to slow one. 2293 * "CWR" CWND was reduced due to some Congestion Notification event. 2294 * It can be ECN, ICMP source quench, local device congestion. 2295 * "Recovery" CWND was reduced, we are fast-retransmitting. 2296 * "Loss" CWND was reduced due to RTO timeout or SACK reneging. 2297 * 2298 * tcp_fastretrans_alert() is entered: 2299 * - each incoming ACK, if state is not "Open" 2300 * - when arrived ACK is unusual, namely: 2301 * * SACK 2302 * * Duplicate ACK. 2303 * * ECN ECE. 2304 * 2305 * Counting packets in flight is pretty simple. 2306 * 2307 * in_flight = packets_out - left_out + retrans_out 2308 * 2309 * packets_out is SND.NXT-SND.UNA counted in packets. 2310 * 2311 * retrans_out is number of retransmitted segments. 2312 * 2313 * left_out is number of segments left network, but not ACKed yet. 2314 * 2315 * left_out = sacked_out + lost_out 2316 * 2317 * sacked_out: Packets, which arrived to receiver out of order 2318 * and hence not ACKed. With SACKs this number is simply 2319 * amount of SACKed data. Even without SACKs 2320 * it is easy to give pretty reliable estimate of this number, 2321 * counting duplicate ACKs. 2322 * 2323 * lost_out: Packets lost by network. TCP has no explicit 2324 * "loss notification" feedback from network (for now). 2325 * It means that this number can be only _guessed_. 2326 * Actually, it is the heuristics to predict lossage that 2327 * distinguishes different algorithms. 2328 * 2329 * F.e. after RTO, when all the queue is considered as lost, 2330 * lost_out = packets_out and in_flight = retrans_out. 2331 * 2332 * Essentially, we have now a few algorithms detecting 2333 * lost packets. 2334 * 2335 * If the receiver supports SACK: 2336 * 2337 * RFC6675/3517: It is the conventional algorithm. A packet is 2338 * considered lost if the number of higher sequence packets 2339 * SACKed is greater than or equal the DUPACK thoreshold 2340 * (reordering). This is implemented in tcp_mark_head_lost and 2341 * tcp_update_scoreboard. 2342 * 2343 * RACK (draft-ietf-tcpm-rack-01): it is a newer algorithm 2344 * (2017-) that checks timing instead of counting DUPACKs. 2345 * Essentially a packet is considered lost if it's not S/ACKed 2346 * after RTT + reordering_window, where both metrics are 2347 * dynamically measured and adjusted. This is implemented in 2348 * tcp_rack_mark_lost. 2349 * 2350 * If the receiver does not support SACK: 2351 * 2352 * NewReno (RFC6582): in Recovery we assume that one segment 2353 * is lost (classic Reno). While we are in Recovery and 2354 * a partial ACK arrives, we assume that one more packet 2355 * is lost (NewReno). This heuristics are the same in NewReno 2356 * and SACK. 2357 * 2358 * Really tricky (and requiring careful tuning) part of algorithm 2359 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue(). 2360 * The first determines the moment _when_ we should reduce CWND and, 2361 * hence, slow down forward transmission. In fact, it determines the moment 2362 * when we decide that hole is caused by loss, rather than by a reorder. 2363 * 2364 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill 2365 * holes, caused by lost packets. 2366 * 2367 * And the most logically complicated part of algorithm is undo 2368 * heuristics. We detect false retransmits due to both too early 2369 * fast retransmit (reordering) and underestimated RTO, analyzing 2370 * timestamps and D-SACKs. When we detect that some segments were 2371 * retransmitted by mistake and CWND reduction was wrong, we undo 2372 * window reduction and abort recovery phase. This logic is hidden 2373 * inside several functions named tcp_try_undo_<something>. 2374 */ 2375 2376 /* This function decides, when we should leave Disordered state 2377 * and enter Recovery phase, reducing congestion window. 2378 * 2379 * Main question: may we further continue forward transmission 2380 * with the same cwnd? 2381 */ 2382 static bool tcp_time_to_recover(struct sock *sk, int flag) 2383 { 2384 struct tcp_sock *tp = tcp_sk(sk); 2385 2386 /* Trick#1: The loss is proven. */ 2387 if (tp->lost_out) 2388 return true; 2389 2390 /* Not-A-Trick#2 : Classic rule... */ 2391 if (!tcp_is_rack(sk) && tcp_dupack_heuristics(tp) > tp->reordering) 2392 return true; 2393 2394 return false; 2395 } 2396 2397 /* Detect loss in event "A" above by marking head of queue up as lost. 2398 * For RFC3517 SACK, a segment is considered lost if it 2399 * has at least tp->reordering SACKed seqments above it; "packets" refers to 2400 * the maximum SACKed segments to pass before reaching this limit. 2401 */ 2402 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head) 2403 { 2404 struct tcp_sock *tp = tcp_sk(sk); 2405 struct sk_buff *skb; 2406 int cnt; 2407 /* Use SACK to deduce losses of new sequences sent during recovery */ 2408 const u32 loss_high = tp->snd_nxt; 2409 2410 WARN_ON(packets > tp->packets_out); 2411 skb = tp->lost_skb_hint; 2412 if (skb) { 2413 /* Head already handled? */ 2414 if (mark_head && after(TCP_SKB_CB(skb)->seq, tp->snd_una)) 2415 return; 2416 cnt = tp->lost_cnt_hint; 2417 } else { 2418 skb = tcp_rtx_queue_head(sk); 2419 cnt = 0; 2420 } 2421 2422 skb_rbtree_walk_from(skb) { 2423 /* TODO: do this better */ 2424 /* this is not the most efficient way to do this... */ 2425 tp->lost_skb_hint = skb; 2426 tp->lost_cnt_hint = cnt; 2427 2428 if (after(TCP_SKB_CB(skb)->end_seq, loss_high)) 2429 break; 2430 2431 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 2432 cnt += tcp_skb_pcount(skb); 2433 2434 if (cnt > packets) 2435 break; 2436 2437 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_LOST)) 2438 tcp_mark_skb_lost(sk, skb); 2439 2440 if (mark_head) 2441 break; 2442 } 2443 tcp_verify_left_out(tp); 2444 } 2445 2446 /* Account newly detected lost packet(s) */ 2447 2448 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit) 2449 { 2450 struct tcp_sock *tp = tcp_sk(sk); 2451 2452 if (tcp_is_sack(tp)) { 2453 int sacked_upto = tp->sacked_out - tp->reordering; 2454 if (sacked_upto >= 0) 2455 tcp_mark_head_lost(sk, sacked_upto, 0); 2456 else if (fast_rexmit) 2457 tcp_mark_head_lost(sk, 1, 1); 2458 } 2459 } 2460 2461 static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when) 2462 { 2463 return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 2464 before(tp->rx_opt.rcv_tsecr, when); 2465 } 2466 2467 /* skb is spurious retransmitted if the returned timestamp echo 2468 * reply is prior to the skb transmission time 2469 */ 2470 static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp, 2471 const struct sk_buff *skb) 2472 { 2473 return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) && 2474 tcp_tsopt_ecr_before(tp, tcp_skb_timestamp_ts(tp->tcp_usec_ts, skb)); 2475 } 2476 2477 /* Nothing was retransmitted or returned timestamp is less 2478 * than timestamp of the first retransmission. 2479 */ 2480 static inline bool tcp_packet_delayed(const struct tcp_sock *tp) 2481 { 2482 const struct sock *sk = (const struct sock *)tp; 2483 2484 if (tp->retrans_stamp && 2485 tcp_tsopt_ecr_before(tp, tp->retrans_stamp)) 2486 return true; /* got echoed TS before first retransmission */ 2487 2488 /* Check if nothing was retransmitted (retrans_stamp==0), which may 2489 * happen in fast recovery due to TSQ. But we ignore zero retrans_stamp 2490 * in TCP_SYN_SENT, since when we set FLAG_SYN_ACKED we also clear 2491 * retrans_stamp even if we had retransmitted the SYN. 2492 */ 2493 if (!tp->retrans_stamp && /* no record of a retransmit/SYN? */ 2494 sk->sk_state != TCP_SYN_SENT) /* not the FLAG_SYN_ACKED case? */ 2495 return true; /* nothing was retransmitted */ 2496 2497 return false; 2498 } 2499 2500 /* Undo procedures. */ 2501 2502 /* We can clear retrans_stamp when there are no retransmissions in the 2503 * window. It would seem that it is trivially available for us in 2504 * tp->retrans_out, however, that kind of assumptions doesn't consider 2505 * what will happen if errors occur when sending retransmission for the 2506 * second time. ...It could the that such segment has only 2507 * TCPCB_EVER_RETRANS set at the present time. It seems that checking 2508 * the head skb is enough except for some reneging corner cases that 2509 * are not worth the effort. 2510 * 2511 * Main reason for all this complexity is the fact that connection dying 2512 * time now depends on the validity of the retrans_stamp, in particular, 2513 * that successive retransmissions of a segment must not advance 2514 * retrans_stamp under any conditions. 2515 */ 2516 static bool tcp_any_retrans_done(const struct sock *sk) 2517 { 2518 const struct tcp_sock *tp = tcp_sk(sk); 2519 struct sk_buff *skb; 2520 2521 if (tp->retrans_out) 2522 return true; 2523 2524 skb = tcp_rtx_queue_head(sk); 2525 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS)) 2526 return true; 2527 2528 return false; 2529 } 2530 2531 /* If loss recovery is finished and there are no retransmits out in the 2532 * network, then we clear retrans_stamp so that upon the next loss recovery 2533 * retransmits_timed_out() and timestamp-undo are using the correct value. 2534 */ 2535 static void tcp_retrans_stamp_cleanup(struct sock *sk) 2536 { 2537 if (!tcp_any_retrans_done(sk)) 2538 tcp_sk(sk)->retrans_stamp = 0; 2539 } 2540 2541 static void DBGUNDO(struct sock *sk, const char *msg) 2542 { 2543 #if FASTRETRANS_DEBUG > 1 2544 struct tcp_sock *tp = tcp_sk(sk); 2545 struct inet_sock *inet = inet_sk(sk); 2546 2547 if (sk->sk_family == AF_INET) { 2548 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n", 2549 msg, 2550 &inet->inet_daddr, ntohs(inet->inet_dport), 2551 tcp_snd_cwnd(tp), tcp_left_out(tp), 2552 tp->snd_ssthresh, tp->prior_ssthresh, 2553 tp->packets_out); 2554 } 2555 #if IS_ENABLED(CONFIG_IPV6) 2556 else if (sk->sk_family == AF_INET6) { 2557 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n", 2558 msg, 2559 &sk->sk_v6_daddr, ntohs(inet->inet_dport), 2560 tcp_snd_cwnd(tp), tcp_left_out(tp), 2561 tp->snd_ssthresh, tp->prior_ssthresh, 2562 tp->packets_out); 2563 } 2564 #endif 2565 #endif 2566 } 2567 2568 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss) 2569 { 2570 struct tcp_sock *tp = tcp_sk(sk); 2571 2572 if (unmark_loss) { 2573 struct sk_buff *skb; 2574 2575 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) { 2576 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; 2577 } 2578 tp->lost_out = 0; 2579 tcp_clear_all_retrans_hints(tp); 2580 } 2581 2582 if (tp->prior_ssthresh) { 2583 const struct inet_connection_sock *icsk = inet_csk(sk); 2584 2585 tcp_snd_cwnd_set(tp, icsk->icsk_ca_ops->undo_cwnd(sk)); 2586 2587 if (tp->prior_ssthresh > tp->snd_ssthresh) { 2588 tp->snd_ssthresh = tp->prior_ssthresh; 2589 tcp_ecn_withdraw_cwr(tp); 2590 } 2591 } 2592 tp->snd_cwnd_stamp = tcp_jiffies32; 2593 tp->undo_marker = 0; 2594 tp->rack.advanced = 1; /* Force RACK to re-exam losses */ 2595 } 2596 2597 static inline bool tcp_may_undo(const struct tcp_sock *tp) 2598 { 2599 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp)); 2600 } 2601 2602 static bool tcp_is_non_sack_preventing_reopen(struct sock *sk) 2603 { 2604 struct tcp_sock *tp = tcp_sk(sk); 2605 2606 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) { 2607 /* Hold old state until something *above* high_seq 2608 * is ACKed. For Reno it is MUST to prevent false 2609 * fast retransmits (RFC2582). SACK TCP is safe. */ 2610 if (!tcp_any_retrans_done(sk)) 2611 tp->retrans_stamp = 0; 2612 return true; 2613 } 2614 return false; 2615 } 2616 2617 /* People celebrate: "We love our President!" */ 2618 static bool tcp_try_undo_recovery(struct sock *sk) 2619 { 2620 struct tcp_sock *tp = tcp_sk(sk); 2621 2622 if (tcp_may_undo(tp)) { 2623 int mib_idx; 2624 2625 /* Happy end! We did not retransmit anything 2626 * or our original transmission succeeded. 2627 */ 2628 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans"); 2629 tcp_undo_cwnd_reduction(sk, false); 2630 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss) 2631 mib_idx = LINUX_MIB_TCPLOSSUNDO; 2632 else 2633 mib_idx = LINUX_MIB_TCPFULLUNDO; 2634 2635 NET_INC_STATS(sock_net(sk), mib_idx); 2636 } else if (tp->rack.reo_wnd_persist) { 2637 tp->rack.reo_wnd_persist--; 2638 } 2639 if (tcp_is_non_sack_preventing_reopen(sk)) 2640 return true; 2641 tcp_set_ca_state(sk, TCP_CA_Open); 2642 tp->is_sack_reneg = 0; 2643 return false; 2644 } 2645 2646 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */ 2647 static bool tcp_try_undo_dsack(struct sock *sk) 2648 { 2649 struct tcp_sock *tp = tcp_sk(sk); 2650 2651 if (tp->undo_marker && !tp->undo_retrans) { 2652 tp->rack.reo_wnd_persist = min(TCP_RACK_RECOVERY_THRESH, 2653 tp->rack.reo_wnd_persist + 1); 2654 DBGUNDO(sk, "D-SACK"); 2655 tcp_undo_cwnd_reduction(sk, false); 2656 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKUNDO); 2657 return true; 2658 } 2659 return false; 2660 } 2661 2662 /* Undo during loss recovery after partial ACK or using F-RTO. */ 2663 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo) 2664 { 2665 struct tcp_sock *tp = tcp_sk(sk); 2666 2667 if (frto_undo || tcp_may_undo(tp)) { 2668 tcp_undo_cwnd_reduction(sk, true); 2669 2670 DBGUNDO(sk, "partial loss"); 2671 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSUNDO); 2672 if (frto_undo) 2673 NET_INC_STATS(sock_net(sk), 2674 LINUX_MIB_TCPSPURIOUSRTOS); 2675 inet_csk(sk)->icsk_retransmits = 0; 2676 if (tcp_is_non_sack_preventing_reopen(sk)) 2677 return true; 2678 if (frto_undo || tcp_is_sack(tp)) { 2679 tcp_set_ca_state(sk, TCP_CA_Open); 2680 tp->is_sack_reneg = 0; 2681 } 2682 return true; 2683 } 2684 return false; 2685 } 2686 2687 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937. 2688 * It computes the number of packets to send (sndcnt) based on packets newly 2689 * delivered: 2690 * 1) If the packets in flight is larger than ssthresh, PRR spreads the 2691 * cwnd reductions across a full RTT. 2692 * 2) Otherwise PRR uses packet conservation to send as much as delivered. 2693 * But when SND_UNA is acked without further losses, 2694 * slow starts cwnd up to ssthresh to speed up the recovery. 2695 */ 2696 static void tcp_init_cwnd_reduction(struct sock *sk) 2697 { 2698 struct tcp_sock *tp = tcp_sk(sk); 2699 2700 tp->high_seq = tp->snd_nxt; 2701 tp->tlp_high_seq = 0; 2702 tp->snd_cwnd_cnt = 0; 2703 tp->prior_cwnd = tcp_snd_cwnd(tp); 2704 tp->prr_delivered = 0; 2705 tp->prr_out = 0; 2706 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk); 2707 tcp_ecn_queue_cwr(tp); 2708 } 2709 2710 void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked, int newly_lost, int flag) 2711 { 2712 struct tcp_sock *tp = tcp_sk(sk); 2713 int sndcnt = 0; 2714 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp); 2715 2716 if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd)) 2717 return; 2718 2719 tp->prr_delivered += newly_acked_sacked; 2720 if (delta < 0) { 2721 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered + 2722 tp->prior_cwnd - 1; 2723 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out; 2724 } else { 2725 sndcnt = max_t(int, tp->prr_delivered - tp->prr_out, 2726 newly_acked_sacked); 2727 if (flag & FLAG_SND_UNA_ADVANCED && !newly_lost) 2728 sndcnt++; 2729 sndcnt = min(delta, sndcnt); 2730 } 2731 /* Force a fast retransmit upon entering fast recovery */ 2732 sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1)); 2733 tcp_snd_cwnd_set(tp, tcp_packets_in_flight(tp) + sndcnt); 2734 } 2735 2736 static inline void tcp_end_cwnd_reduction(struct sock *sk) 2737 { 2738 struct tcp_sock *tp = tcp_sk(sk); 2739 2740 if (inet_csk(sk)->icsk_ca_ops->cong_control) 2741 return; 2742 2743 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */ 2744 if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH && 2745 (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR || tp->undo_marker)) { 2746 tcp_snd_cwnd_set(tp, tp->snd_ssthresh); 2747 tp->snd_cwnd_stamp = tcp_jiffies32; 2748 } 2749 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR); 2750 } 2751 2752 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */ 2753 void tcp_enter_cwr(struct sock *sk) 2754 { 2755 struct tcp_sock *tp = tcp_sk(sk); 2756 2757 tp->prior_ssthresh = 0; 2758 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) { 2759 tp->undo_marker = 0; 2760 tcp_init_cwnd_reduction(sk); 2761 tcp_set_ca_state(sk, TCP_CA_CWR); 2762 } 2763 } 2764 EXPORT_SYMBOL(tcp_enter_cwr); 2765 2766 static void tcp_try_keep_open(struct sock *sk) 2767 { 2768 struct tcp_sock *tp = tcp_sk(sk); 2769 int state = TCP_CA_Open; 2770 2771 if (tcp_left_out(tp) || tcp_any_retrans_done(sk)) 2772 state = TCP_CA_Disorder; 2773 2774 if (inet_csk(sk)->icsk_ca_state != state) { 2775 tcp_set_ca_state(sk, state); 2776 tp->high_seq = tp->snd_nxt; 2777 } 2778 } 2779 2780 static void tcp_try_to_open(struct sock *sk, int flag) 2781 { 2782 struct tcp_sock *tp = tcp_sk(sk); 2783 2784 tcp_verify_left_out(tp); 2785 2786 if (!tcp_any_retrans_done(sk)) 2787 tp->retrans_stamp = 0; 2788 2789 if (flag & FLAG_ECE) 2790 tcp_enter_cwr(sk); 2791 2792 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) { 2793 tcp_try_keep_open(sk); 2794 } 2795 } 2796 2797 static void tcp_mtup_probe_failed(struct sock *sk) 2798 { 2799 struct inet_connection_sock *icsk = inet_csk(sk); 2800 2801 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1; 2802 icsk->icsk_mtup.probe_size = 0; 2803 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPFAIL); 2804 } 2805 2806 static void tcp_mtup_probe_success(struct sock *sk) 2807 { 2808 struct tcp_sock *tp = tcp_sk(sk); 2809 struct inet_connection_sock *icsk = inet_csk(sk); 2810 u64 val; 2811 2812 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2813 2814 val = (u64)tcp_snd_cwnd(tp) * tcp_mss_to_mtu(sk, tp->mss_cache); 2815 do_div(val, icsk->icsk_mtup.probe_size); 2816 DEBUG_NET_WARN_ON_ONCE((u32)val != val); 2817 tcp_snd_cwnd_set(tp, max_t(u32, 1U, val)); 2818 2819 tp->snd_cwnd_cnt = 0; 2820 tp->snd_cwnd_stamp = tcp_jiffies32; 2821 tp->snd_ssthresh = tcp_current_ssthresh(sk); 2822 2823 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size; 2824 icsk->icsk_mtup.probe_size = 0; 2825 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 2826 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS); 2827 } 2828 2829 /* Sometimes we deduce that packets have been dropped due to reasons other than 2830 * congestion, like path MTU reductions or failed client TFO attempts. In these 2831 * cases we call this function to retransmit as many packets as cwnd allows, 2832 * without reducing cwnd. Given that retransmits will set retrans_stamp to a 2833 * non-zero value (and may do so in a later calling context due to TSQ), we 2834 * also enter CA_Loss so that we track when all retransmitted packets are ACKed 2835 * and clear retrans_stamp when that happens (to ensure later recurring RTOs 2836 * are using the correct retrans_stamp and don't declare ETIMEDOUT 2837 * prematurely). 2838 */ 2839 static void tcp_non_congestion_loss_retransmit(struct sock *sk) 2840 { 2841 const struct inet_connection_sock *icsk = inet_csk(sk); 2842 struct tcp_sock *tp = tcp_sk(sk); 2843 2844 if (icsk->icsk_ca_state != TCP_CA_Loss) { 2845 tp->high_seq = tp->snd_nxt; 2846 tp->snd_ssthresh = tcp_current_ssthresh(sk); 2847 tp->prior_ssthresh = 0; 2848 tp->undo_marker = 0; 2849 tcp_set_ca_state(sk, TCP_CA_Loss); 2850 } 2851 tcp_xmit_retransmit_queue(sk); 2852 } 2853 2854 /* Do a simple retransmit without using the backoff mechanisms in 2855 * tcp_timer. This is used for path mtu discovery. 2856 * The socket is already locked here. 2857 */ 2858 void tcp_simple_retransmit(struct sock *sk) 2859 { 2860 struct tcp_sock *tp = tcp_sk(sk); 2861 struct sk_buff *skb; 2862 int mss; 2863 2864 /* A fastopen SYN request is stored as two separate packets within 2865 * the retransmit queue, this is done by tcp_send_syn_data(). 2866 * As a result simply checking the MSS of the frames in the queue 2867 * will not work for the SYN packet. 2868 * 2869 * Us being here is an indication of a path MTU issue so we can 2870 * assume that the fastopen SYN was lost and just mark all the 2871 * frames in the retransmit queue as lost. We will use an MSS of 2872 * -1 to mark all frames as lost, otherwise compute the current MSS. 2873 */ 2874 if (tp->syn_data && sk->sk_state == TCP_SYN_SENT) 2875 mss = -1; 2876 else 2877 mss = tcp_current_mss(sk); 2878 2879 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) { 2880 if (tcp_skb_seglen(skb) > mss) 2881 tcp_mark_skb_lost(sk, skb); 2882 } 2883 2884 tcp_clear_retrans_hints_partial(tp); 2885 2886 if (!tp->lost_out) 2887 return; 2888 2889 if (tcp_is_reno(tp)) 2890 tcp_limit_reno_sacked(tp); 2891 2892 tcp_verify_left_out(tp); 2893 2894 /* Don't muck with the congestion window here. 2895 * Reason is that we do not increase amount of _data_ 2896 * in network, but units changed and effective 2897 * cwnd/ssthresh really reduced now. 2898 */ 2899 tcp_non_congestion_loss_retransmit(sk); 2900 } 2901 EXPORT_SYMBOL(tcp_simple_retransmit); 2902 2903 void tcp_enter_recovery(struct sock *sk, bool ece_ack) 2904 { 2905 struct tcp_sock *tp = tcp_sk(sk); 2906 int mib_idx; 2907 2908 /* Start the clock with our fast retransmit, for undo and ETIMEDOUT. */ 2909 tcp_retrans_stamp_cleanup(sk); 2910 2911 if (tcp_is_reno(tp)) 2912 mib_idx = LINUX_MIB_TCPRENORECOVERY; 2913 else 2914 mib_idx = LINUX_MIB_TCPSACKRECOVERY; 2915 2916 NET_INC_STATS(sock_net(sk), mib_idx); 2917 2918 tp->prior_ssthresh = 0; 2919 tcp_init_undo(tp); 2920 2921 if (!tcp_in_cwnd_reduction(sk)) { 2922 if (!ece_ack) 2923 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2924 tcp_init_cwnd_reduction(sk); 2925 } 2926 tcp_set_ca_state(sk, TCP_CA_Recovery); 2927 } 2928 2929 static void tcp_update_rto_time(struct tcp_sock *tp) 2930 { 2931 if (tp->rto_stamp) { 2932 tp->total_rto_time += tcp_time_stamp_ms(tp) - tp->rto_stamp; 2933 tp->rto_stamp = 0; 2934 } 2935 } 2936 2937 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are 2938 * recovered or spurious. Otherwise retransmits more on partial ACKs. 2939 */ 2940 static void tcp_process_loss(struct sock *sk, int flag, int num_dupack, 2941 int *rexmit) 2942 { 2943 struct tcp_sock *tp = tcp_sk(sk); 2944 bool recovered = !before(tp->snd_una, tp->high_seq); 2945 2946 if ((flag & FLAG_SND_UNA_ADVANCED || rcu_access_pointer(tp->fastopen_rsk)) && 2947 tcp_try_undo_loss(sk, false)) 2948 return; 2949 2950 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */ 2951 /* Step 3.b. A timeout is spurious if not all data are 2952 * lost, i.e., never-retransmitted data are (s)acked. 2953 */ 2954 if ((flag & FLAG_ORIG_SACK_ACKED) && 2955 tcp_try_undo_loss(sk, true)) 2956 return; 2957 2958 if (after(tp->snd_nxt, tp->high_seq)) { 2959 if (flag & FLAG_DATA_SACKED || num_dupack) 2960 tp->frto = 0; /* Step 3.a. loss was real */ 2961 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) { 2962 tp->high_seq = tp->snd_nxt; 2963 /* Step 2.b. Try send new data (but deferred until cwnd 2964 * is updated in tcp_ack()). Otherwise fall back to 2965 * the conventional recovery. 2966 */ 2967 if (!tcp_write_queue_empty(sk) && 2968 after(tcp_wnd_end(tp), tp->snd_nxt)) { 2969 *rexmit = REXMIT_NEW; 2970 return; 2971 } 2972 tp->frto = 0; 2973 } 2974 } 2975 2976 if (recovered) { 2977 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */ 2978 tcp_try_undo_recovery(sk); 2979 return; 2980 } 2981 if (tcp_is_reno(tp)) { 2982 /* A Reno DUPACK means new data in F-RTO step 2.b above are 2983 * delivered. Lower inflight to clock out (re)transmissions. 2984 */ 2985 if (after(tp->snd_nxt, tp->high_seq) && num_dupack) 2986 tcp_add_reno_sack(sk, num_dupack, flag & FLAG_ECE); 2987 else if (flag & FLAG_SND_UNA_ADVANCED) 2988 tcp_reset_reno_sack(tp); 2989 } 2990 *rexmit = REXMIT_LOST; 2991 } 2992 2993 static bool tcp_force_fast_retransmit(struct sock *sk) 2994 { 2995 struct tcp_sock *tp = tcp_sk(sk); 2996 2997 return after(tcp_highest_sack_seq(tp), 2998 tp->snd_una + tp->reordering * tp->mss_cache); 2999 } 3000 3001 /* Undo during fast recovery after partial ACK. */ 3002 static bool tcp_try_undo_partial(struct sock *sk, u32 prior_snd_una, 3003 bool *do_lost) 3004 { 3005 struct tcp_sock *tp = tcp_sk(sk); 3006 3007 if (tp->undo_marker && tcp_packet_delayed(tp)) { 3008 /* Plain luck! Hole if filled with delayed 3009 * packet, rather than with a retransmit. Check reordering. 3010 */ 3011 tcp_check_sack_reordering(sk, prior_snd_una, 1); 3012 3013 /* We are getting evidence that the reordering degree is higher 3014 * than we realized. If there are no retransmits out then we 3015 * can undo. Otherwise we clock out new packets but do not 3016 * mark more packets lost or retransmit more. 3017 */ 3018 if (tp->retrans_out) 3019 return true; 3020 3021 if (!tcp_any_retrans_done(sk)) 3022 tp->retrans_stamp = 0; 3023 3024 DBGUNDO(sk, "partial recovery"); 3025 tcp_undo_cwnd_reduction(sk, true); 3026 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO); 3027 tcp_try_keep_open(sk); 3028 } else { 3029 /* Partial ACK arrived. Force fast retransmit. */ 3030 *do_lost = tcp_force_fast_retransmit(sk); 3031 } 3032 return false; 3033 } 3034 3035 static void tcp_identify_packet_loss(struct sock *sk, int *ack_flag) 3036 { 3037 struct tcp_sock *tp = tcp_sk(sk); 3038 3039 if (tcp_rtx_queue_empty(sk)) 3040 return; 3041 3042 if (unlikely(tcp_is_reno(tp))) { 3043 tcp_newreno_mark_lost(sk, *ack_flag & FLAG_SND_UNA_ADVANCED); 3044 } else if (tcp_is_rack(sk)) { 3045 u32 prior_retrans = tp->retrans_out; 3046 3047 if (tcp_rack_mark_lost(sk)) 3048 *ack_flag &= ~FLAG_SET_XMIT_TIMER; 3049 if (prior_retrans > tp->retrans_out) 3050 *ack_flag |= FLAG_LOST_RETRANS; 3051 } 3052 } 3053 3054 /* Process an event, which can update packets-in-flight not trivially. 3055 * Main goal of this function is to calculate new estimate for left_out, 3056 * taking into account both packets sitting in receiver's buffer and 3057 * packets lost by network. 3058 * 3059 * Besides that it updates the congestion state when packet loss or ECN 3060 * is detected. But it does not reduce the cwnd, it is done by the 3061 * congestion control later. 3062 * 3063 * It does _not_ decide what to send, it is made in function 3064 * tcp_xmit_retransmit_queue(). 3065 */ 3066 static void tcp_fastretrans_alert(struct sock *sk, const u32 prior_snd_una, 3067 int num_dupack, int *ack_flag, int *rexmit) 3068 { 3069 struct inet_connection_sock *icsk = inet_csk(sk); 3070 struct tcp_sock *tp = tcp_sk(sk); 3071 int fast_rexmit = 0, flag = *ack_flag; 3072 bool ece_ack = flag & FLAG_ECE; 3073 bool do_lost = num_dupack || ((flag & FLAG_DATA_SACKED) && 3074 tcp_force_fast_retransmit(sk)); 3075 3076 if (!tp->packets_out && tp->sacked_out) 3077 tp->sacked_out = 0; 3078 3079 /* Now state machine starts. 3080 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */ 3081 if (ece_ack) 3082 tp->prior_ssthresh = 0; 3083 3084 /* B. In all the states check for reneging SACKs. */ 3085 if (tcp_check_sack_reneging(sk, ack_flag)) 3086 return; 3087 3088 /* C. Check consistency of the current state. */ 3089 tcp_verify_left_out(tp); 3090 3091 /* D. Check state exit conditions. State can be terminated 3092 * when high_seq is ACKed. */ 3093 if (icsk->icsk_ca_state == TCP_CA_Open) { 3094 WARN_ON(tp->retrans_out != 0 && !tp->syn_data); 3095 tp->retrans_stamp = 0; 3096 } else if (!before(tp->snd_una, tp->high_seq)) { 3097 switch (icsk->icsk_ca_state) { 3098 case TCP_CA_CWR: 3099 /* CWR is to be held something *above* high_seq 3100 * is ACKed for CWR bit to reach receiver. */ 3101 if (tp->snd_una != tp->high_seq) { 3102 tcp_end_cwnd_reduction(sk); 3103 tcp_set_ca_state(sk, TCP_CA_Open); 3104 } 3105 break; 3106 3107 case TCP_CA_Recovery: 3108 if (tcp_is_reno(tp)) 3109 tcp_reset_reno_sack(tp); 3110 if (tcp_try_undo_recovery(sk)) 3111 return; 3112 tcp_end_cwnd_reduction(sk); 3113 break; 3114 } 3115 } 3116 3117 /* E. Process state. */ 3118 switch (icsk->icsk_ca_state) { 3119 case TCP_CA_Recovery: 3120 if (!(flag & FLAG_SND_UNA_ADVANCED)) { 3121 if (tcp_is_reno(tp)) 3122 tcp_add_reno_sack(sk, num_dupack, ece_ack); 3123 } else if (tcp_try_undo_partial(sk, prior_snd_una, &do_lost)) 3124 return; 3125 3126 if (tcp_try_undo_dsack(sk)) 3127 tcp_try_to_open(sk, flag); 3128 3129 tcp_identify_packet_loss(sk, ack_flag); 3130 if (icsk->icsk_ca_state != TCP_CA_Recovery) { 3131 if (!tcp_time_to_recover(sk, flag)) 3132 return; 3133 /* Undo reverts the recovery state. If loss is evident, 3134 * starts a new recovery (e.g. reordering then loss); 3135 */ 3136 tcp_enter_recovery(sk, ece_ack); 3137 } 3138 break; 3139 case TCP_CA_Loss: 3140 tcp_process_loss(sk, flag, num_dupack, rexmit); 3141 if (icsk->icsk_ca_state != TCP_CA_Loss) 3142 tcp_update_rto_time(tp); 3143 tcp_identify_packet_loss(sk, ack_flag); 3144 if (!(icsk->icsk_ca_state == TCP_CA_Open || 3145 (*ack_flag & FLAG_LOST_RETRANS))) 3146 return; 3147 /* Change state if cwnd is undone or retransmits are lost */ 3148 fallthrough; 3149 default: 3150 if (tcp_is_reno(tp)) { 3151 if (flag & FLAG_SND_UNA_ADVANCED) 3152 tcp_reset_reno_sack(tp); 3153 tcp_add_reno_sack(sk, num_dupack, ece_ack); 3154 } 3155 3156 if (icsk->icsk_ca_state <= TCP_CA_Disorder) 3157 tcp_try_undo_dsack(sk); 3158 3159 tcp_identify_packet_loss(sk, ack_flag); 3160 if (!tcp_time_to_recover(sk, flag)) { 3161 tcp_try_to_open(sk, flag); 3162 return; 3163 } 3164 3165 /* MTU probe failure: don't reduce cwnd */ 3166 if (icsk->icsk_ca_state < TCP_CA_CWR && 3167 icsk->icsk_mtup.probe_size && 3168 tp->snd_una == tp->mtu_probe.probe_seq_start) { 3169 tcp_mtup_probe_failed(sk); 3170 /* Restores the reduction we did in tcp_mtup_probe() */ 3171 tcp_snd_cwnd_set(tp, tcp_snd_cwnd(tp) + 1); 3172 tcp_simple_retransmit(sk); 3173 return; 3174 } 3175 3176 /* Otherwise enter Recovery state */ 3177 tcp_enter_recovery(sk, ece_ack); 3178 fast_rexmit = 1; 3179 } 3180 3181 if (!tcp_is_rack(sk) && do_lost) 3182 tcp_update_scoreboard(sk, fast_rexmit); 3183 *rexmit = REXMIT_LOST; 3184 } 3185 3186 static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us, const int flag) 3187 { 3188 u32 wlen = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_rtt_wlen) * HZ; 3189 struct tcp_sock *tp = tcp_sk(sk); 3190 3191 if ((flag & FLAG_ACK_MAYBE_DELAYED) && rtt_us > tcp_min_rtt(tp)) { 3192 /* If the remote keeps returning delayed ACKs, eventually 3193 * the min filter would pick it up and overestimate the 3194 * prop. delay when it expires. Skip suspected delayed ACKs. 3195 */ 3196 return; 3197 } 3198 minmax_running_min(&tp->rtt_min, wlen, tcp_jiffies32, 3199 rtt_us ? : jiffies_to_usecs(1)); 3200 } 3201 3202 static bool tcp_ack_update_rtt(struct sock *sk, const int flag, 3203 long seq_rtt_us, long sack_rtt_us, 3204 long ca_rtt_us, struct rate_sample *rs) 3205 { 3206 const struct tcp_sock *tp = tcp_sk(sk); 3207 3208 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because 3209 * broken middle-boxes or peers may corrupt TS-ECR fields. But 3210 * Karn's algorithm forbids taking RTT if some retransmitted data 3211 * is acked (RFC6298). 3212 */ 3213 if (seq_rtt_us < 0) 3214 seq_rtt_us = sack_rtt_us; 3215 3216 /* RTTM Rule: A TSecr value received in a segment is used to 3217 * update the averaged RTT measurement only if the segment 3218 * acknowledges some new data, i.e., only if it advances the 3219 * left edge of the send window. 3220 * See draft-ietf-tcplw-high-performance-00, section 3.3. 3221 */ 3222 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && 3223 tp->rx_opt.rcv_tsecr && flag & FLAG_ACKED) 3224 seq_rtt_us = ca_rtt_us = tcp_rtt_tsopt_us(tp); 3225 3226 rs->rtt_us = ca_rtt_us; /* RTT of last (S)ACKed packet (or -1) */ 3227 if (seq_rtt_us < 0) 3228 return false; 3229 3230 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is 3231 * always taken together with ACK, SACK, or TS-opts. Any negative 3232 * values will be skipped with the seq_rtt_us < 0 check above. 3233 */ 3234 tcp_update_rtt_min(sk, ca_rtt_us, flag); 3235 tcp_rtt_estimator(sk, seq_rtt_us); 3236 tcp_set_rto(sk); 3237 3238 /* RFC6298: only reset backoff on valid RTT measurement. */ 3239 inet_csk(sk)->icsk_backoff = 0; 3240 return true; 3241 } 3242 3243 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */ 3244 void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req) 3245 { 3246 struct rate_sample rs; 3247 long rtt_us = -1L; 3248 3249 if (req && !req->num_retrans && tcp_rsk(req)->snt_synack) 3250 rtt_us = tcp_stamp_us_delta(tcp_clock_us(), tcp_rsk(req)->snt_synack); 3251 3252 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us, &rs); 3253 } 3254 3255 3256 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked) 3257 { 3258 const struct inet_connection_sock *icsk = inet_csk(sk); 3259 3260 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked); 3261 tcp_sk(sk)->snd_cwnd_stamp = tcp_jiffies32; 3262 } 3263 3264 /* Restart timer after forward progress on connection. 3265 * RFC2988 recommends to restart timer to now+rto. 3266 */ 3267 void tcp_rearm_rto(struct sock *sk) 3268 { 3269 const struct inet_connection_sock *icsk = inet_csk(sk); 3270 struct tcp_sock *tp = tcp_sk(sk); 3271 3272 /* If the retrans timer is currently being used by Fast Open 3273 * for SYN-ACK retrans purpose, stay put. 3274 */ 3275 if (rcu_access_pointer(tp->fastopen_rsk)) 3276 return; 3277 3278 if (!tp->packets_out) { 3279 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS); 3280 } else { 3281 u32 rto = inet_csk(sk)->icsk_rto; 3282 /* Offset the time elapsed after installing regular RTO */ 3283 if (icsk->icsk_pending == ICSK_TIME_REO_TIMEOUT || 3284 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) { 3285 s64 delta_us = tcp_rto_delta_us(sk); 3286 /* delta_us may not be positive if the socket is locked 3287 * when the retrans timer fires and is rescheduled. 3288 */ 3289 rto = usecs_to_jiffies(max_t(int, delta_us, 1)); 3290 } 3291 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto, 3292 TCP_RTO_MAX); 3293 } 3294 } 3295 3296 /* Try to schedule a loss probe; if that doesn't work, then schedule an RTO. */ 3297 static void tcp_set_xmit_timer(struct sock *sk) 3298 { 3299 if (!tcp_schedule_loss_probe(sk, true)) 3300 tcp_rearm_rto(sk); 3301 } 3302 3303 /* If we get here, the whole TSO packet has not been acked. */ 3304 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb) 3305 { 3306 struct tcp_sock *tp = tcp_sk(sk); 3307 u32 packets_acked; 3308 3309 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)); 3310 3311 packets_acked = tcp_skb_pcount(skb); 3312 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 3313 return 0; 3314 packets_acked -= tcp_skb_pcount(skb); 3315 3316 if (packets_acked) { 3317 BUG_ON(tcp_skb_pcount(skb) == 0); 3318 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)); 3319 } 3320 3321 return packets_acked; 3322 } 3323 3324 static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb, 3325 const struct sk_buff *ack_skb, u32 prior_snd_una) 3326 { 3327 const struct skb_shared_info *shinfo; 3328 3329 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */ 3330 if (likely(!TCP_SKB_CB(skb)->txstamp_ack)) 3331 return; 3332 3333 shinfo = skb_shinfo(skb); 3334 if (!before(shinfo->tskey, prior_snd_una) && 3335 before(shinfo->tskey, tcp_sk(sk)->snd_una)) { 3336 tcp_skb_tsorted_save(skb) { 3337 __skb_tstamp_tx(skb, ack_skb, NULL, sk, SCM_TSTAMP_ACK); 3338 } tcp_skb_tsorted_restore(skb); 3339 } 3340 } 3341 3342 /* Remove acknowledged frames from the retransmission queue. If our packet 3343 * is before the ack sequence we can discard it as it's confirmed to have 3344 * arrived at the other end. 3345 */ 3346 static int tcp_clean_rtx_queue(struct sock *sk, const struct sk_buff *ack_skb, 3347 u32 prior_fack, u32 prior_snd_una, 3348 struct tcp_sacktag_state *sack, bool ece_ack) 3349 { 3350 const struct inet_connection_sock *icsk = inet_csk(sk); 3351 u64 first_ackt, last_ackt; 3352 struct tcp_sock *tp = tcp_sk(sk); 3353 u32 prior_sacked = tp->sacked_out; 3354 u32 reord = tp->snd_nxt; /* lowest acked un-retx un-sacked seq */ 3355 struct sk_buff *skb, *next; 3356 bool fully_acked = true; 3357 long sack_rtt_us = -1L; 3358 long seq_rtt_us = -1L; 3359 long ca_rtt_us = -1L; 3360 u32 pkts_acked = 0; 3361 bool rtt_update; 3362 int flag = 0; 3363 3364 first_ackt = 0; 3365 3366 for (skb = skb_rb_first(&sk->tcp_rtx_queue); skb; skb = next) { 3367 struct tcp_skb_cb *scb = TCP_SKB_CB(skb); 3368 const u32 start_seq = scb->seq; 3369 u8 sacked = scb->sacked; 3370 u32 acked_pcount; 3371 3372 /* Determine how many packets and what bytes were acked, tso and else */ 3373 if (after(scb->end_seq, tp->snd_una)) { 3374 if (tcp_skb_pcount(skb) == 1 || 3375 !after(tp->snd_una, scb->seq)) 3376 break; 3377 3378 acked_pcount = tcp_tso_acked(sk, skb); 3379 if (!acked_pcount) 3380 break; 3381 fully_acked = false; 3382 } else { 3383 acked_pcount = tcp_skb_pcount(skb); 3384 } 3385 3386 if (unlikely(sacked & TCPCB_RETRANS)) { 3387 if (sacked & TCPCB_SACKED_RETRANS) 3388 tp->retrans_out -= acked_pcount; 3389 flag |= FLAG_RETRANS_DATA_ACKED; 3390 } else if (!(sacked & TCPCB_SACKED_ACKED)) { 3391 last_ackt = tcp_skb_timestamp_us(skb); 3392 WARN_ON_ONCE(last_ackt == 0); 3393 if (!first_ackt) 3394 first_ackt = last_ackt; 3395 3396 if (before(start_seq, reord)) 3397 reord = start_seq; 3398 if (!after(scb->end_seq, tp->high_seq)) 3399 flag |= FLAG_ORIG_SACK_ACKED; 3400 } 3401 3402 if (sacked & TCPCB_SACKED_ACKED) { 3403 tp->sacked_out -= acked_pcount; 3404 } else if (tcp_is_sack(tp)) { 3405 tcp_count_delivered(tp, acked_pcount, ece_ack); 3406 if (!tcp_skb_spurious_retrans(tp, skb)) 3407 tcp_rack_advance(tp, sacked, scb->end_seq, 3408 tcp_skb_timestamp_us(skb)); 3409 } 3410 if (sacked & TCPCB_LOST) 3411 tp->lost_out -= acked_pcount; 3412 3413 tp->packets_out -= acked_pcount; 3414 pkts_acked += acked_pcount; 3415 tcp_rate_skb_delivered(sk, skb, sack->rate); 3416 3417 /* Initial outgoing SYN's get put onto the write_queue 3418 * just like anything else we transmit. It is not 3419 * true data, and if we misinform our callers that 3420 * this ACK acks real data, we will erroneously exit 3421 * connection startup slow start one packet too 3422 * quickly. This is severely frowned upon behavior. 3423 */ 3424 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) { 3425 flag |= FLAG_DATA_ACKED; 3426 } else { 3427 flag |= FLAG_SYN_ACKED; 3428 tp->retrans_stamp = 0; 3429 } 3430 3431 if (!fully_acked) 3432 break; 3433 3434 tcp_ack_tstamp(sk, skb, ack_skb, prior_snd_una); 3435 3436 next = skb_rb_next(skb); 3437 if (unlikely(skb == tp->retransmit_skb_hint)) 3438 tp->retransmit_skb_hint = NULL; 3439 if (unlikely(skb == tp->lost_skb_hint)) 3440 tp->lost_skb_hint = NULL; 3441 tcp_highest_sack_replace(sk, skb, next); 3442 tcp_rtx_queue_unlink_and_free(skb, sk); 3443 } 3444 3445 if (!skb) 3446 tcp_chrono_stop(sk, TCP_CHRONO_BUSY); 3447 3448 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una))) 3449 tp->snd_up = tp->snd_una; 3450 3451 if (skb) { 3452 tcp_ack_tstamp(sk, skb, ack_skb, prior_snd_una); 3453 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 3454 flag |= FLAG_SACK_RENEGING; 3455 } 3456 3457 if (likely(first_ackt) && !(flag & FLAG_RETRANS_DATA_ACKED)) { 3458 seq_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, first_ackt); 3459 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, last_ackt); 3460 3461 if (pkts_acked == 1 && fully_acked && !prior_sacked && 3462 (tp->snd_una - prior_snd_una) < tp->mss_cache && 3463 sack->rate->prior_delivered + 1 == tp->delivered && 3464 !(flag & (FLAG_CA_ALERT | FLAG_SYN_ACKED))) { 3465 /* Conservatively mark a delayed ACK. It's typically 3466 * from a lone runt packet over the round trip to 3467 * a receiver w/o out-of-order or CE events. 3468 */ 3469 flag |= FLAG_ACK_MAYBE_DELAYED; 3470 } 3471 } 3472 if (sack->first_sackt) { 3473 sack_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->first_sackt); 3474 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->last_sackt); 3475 } 3476 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us, 3477 ca_rtt_us, sack->rate); 3478 3479 if (flag & FLAG_ACKED) { 3480 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */ 3481 if (unlikely(icsk->icsk_mtup.probe_size && 3482 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) { 3483 tcp_mtup_probe_success(sk); 3484 } 3485 3486 if (tcp_is_reno(tp)) { 3487 tcp_remove_reno_sacks(sk, pkts_acked, ece_ack); 3488 3489 /* If any of the cumulatively ACKed segments was 3490 * retransmitted, non-SACK case cannot confirm that 3491 * progress was due to original transmission due to 3492 * lack of TCPCB_SACKED_ACKED bits even if some of 3493 * the packets may have been never retransmitted. 3494 */ 3495 if (flag & FLAG_RETRANS_DATA_ACKED) 3496 flag &= ~FLAG_ORIG_SACK_ACKED; 3497 } else { 3498 int delta; 3499 3500 /* Non-retransmitted hole got filled? That's reordering */ 3501 if (before(reord, prior_fack)) 3502 tcp_check_sack_reordering(sk, reord, 0); 3503 3504 delta = prior_sacked - tp->sacked_out; 3505 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta); 3506 } 3507 } else if (skb && rtt_update && sack_rtt_us >= 0 && 3508 sack_rtt_us > tcp_stamp_us_delta(tp->tcp_mstamp, 3509 tcp_skb_timestamp_us(skb))) { 3510 /* Do not re-arm RTO if the sack RTT is measured from data sent 3511 * after when the head was last (re)transmitted. Otherwise the 3512 * timeout may continue to extend in loss recovery. 3513 */ 3514 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */ 3515 } 3516 3517 if (icsk->icsk_ca_ops->pkts_acked) { 3518 struct ack_sample sample = { .pkts_acked = pkts_acked, 3519 .rtt_us = sack->rate->rtt_us }; 3520 3521 sample.in_flight = tp->mss_cache * 3522 (tp->delivered - sack->rate->prior_delivered); 3523 icsk->icsk_ca_ops->pkts_acked(sk, &sample); 3524 } 3525 3526 #if FASTRETRANS_DEBUG > 0 3527 WARN_ON((int)tp->sacked_out < 0); 3528 WARN_ON((int)tp->lost_out < 0); 3529 WARN_ON((int)tp->retrans_out < 0); 3530 if (!tp->packets_out && tcp_is_sack(tp)) { 3531 icsk = inet_csk(sk); 3532 if (tp->lost_out) { 3533 pr_debug("Leak l=%u %d\n", 3534 tp->lost_out, icsk->icsk_ca_state); 3535 tp->lost_out = 0; 3536 } 3537 if (tp->sacked_out) { 3538 pr_debug("Leak s=%u %d\n", 3539 tp->sacked_out, icsk->icsk_ca_state); 3540 tp->sacked_out = 0; 3541 } 3542 if (tp->retrans_out) { 3543 pr_debug("Leak r=%u %d\n", 3544 tp->retrans_out, icsk->icsk_ca_state); 3545 tp->retrans_out = 0; 3546 } 3547 } 3548 #endif 3549 return flag; 3550 } 3551 3552 static void tcp_ack_probe(struct sock *sk) 3553 { 3554 struct inet_connection_sock *icsk = inet_csk(sk); 3555 struct sk_buff *head = tcp_send_head(sk); 3556 const struct tcp_sock *tp = tcp_sk(sk); 3557 3558 /* Was it a usable window open? */ 3559 if (!head) 3560 return; 3561 if (!after(TCP_SKB_CB(head)->end_seq, tcp_wnd_end(tp))) { 3562 icsk->icsk_backoff = 0; 3563 icsk->icsk_probes_tstamp = 0; 3564 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0); 3565 /* Socket must be waked up by subsequent tcp_data_snd_check(). 3566 * This function is not for random using! 3567 */ 3568 } else { 3569 unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX); 3570 3571 when = tcp_clamp_probe0_to_user_timeout(sk, when); 3572 tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, when, TCP_RTO_MAX); 3573 } 3574 } 3575 3576 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag) 3577 { 3578 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) || 3579 inet_csk(sk)->icsk_ca_state != TCP_CA_Open; 3580 } 3581 3582 /* Decide wheather to run the increase function of congestion control. */ 3583 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag) 3584 { 3585 /* If reordering is high then always grow cwnd whenever data is 3586 * delivered regardless of its ordering. Otherwise stay conservative 3587 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/ 3588 * new SACK or ECE mark may first advance cwnd here and later reduce 3589 * cwnd in tcp_fastretrans_alert() based on more states. 3590 */ 3591 if (tcp_sk(sk)->reordering > 3592 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_reordering)) 3593 return flag & FLAG_FORWARD_PROGRESS; 3594 3595 return flag & FLAG_DATA_ACKED; 3596 } 3597 3598 /* The "ultimate" congestion control function that aims to replace the rigid 3599 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction). 3600 * It's called toward the end of processing an ACK with precise rate 3601 * information. All transmission or retransmission are delayed afterwards. 3602 */ 3603 static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked, 3604 int flag, const struct rate_sample *rs) 3605 { 3606 const struct inet_connection_sock *icsk = inet_csk(sk); 3607 3608 if (icsk->icsk_ca_ops->cong_control) { 3609 icsk->icsk_ca_ops->cong_control(sk, ack, flag, rs); 3610 return; 3611 } 3612 3613 if (tcp_in_cwnd_reduction(sk)) { 3614 /* Reduce cwnd if state mandates */ 3615 tcp_cwnd_reduction(sk, acked_sacked, rs->losses, flag); 3616 } else if (tcp_may_raise_cwnd(sk, flag)) { 3617 /* Advance cwnd if state allows */ 3618 tcp_cong_avoid(sk, ack, acked_sacked); 3619 } 3620 tcp_update_pacing_rate(sk); 3621 } 3622 3623 /* Check that window update is acceptable. 3624 * The function assumes that snd_una<=ack<=snd_next. 3625 */ 3626 static inline bool tcp_may_update_window(const struct tcp_sock *tp, 3627 const u32 ack, const u32 ack_seq, 3628 const u32 nwin) 3629 { 3630 return after(ack, tp->snd_una) || 3631 after(ack_seq, tp->snd_wl1) || 3632 (ack_seq == tp->snd_wl1 && (nwin > tp->snd_wnd || !nwin)); 3633 } 3634 3635 static void tcp_snd_sne_update(struct tcp_sock *tp, u32 ack) 3636 { 3637 #ifdef CONFIG_TCP_AO 3638 struct tcp_ao_info *ao; 3639 3640 if (!static_branch_unlikely(&tcp_ao_needed.key)) 3641 return; 3642 3643 ao = rcu_dereference_protected(tp->ao_info, 3644 lockdep_sock_is_held((struct sock *)tp)); 3645 if (ao && ack < tp->snd_una) { 3646 ao->snd_sne++; 3647 trace_tcp_ao_snd_sne_update((struct sock *)tp, ao->snd_sne); 3648 } 3649 #endif 3650 } 3651 3652 /* If we update tp->snd_una, also update tp->bytes_acked */ 3653 static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack) 3654 { 3655 u32 delta = ack - tp->snd_una; 3656 3657 sock_owned_by_me((struct sock *)tp); 3658 tp->bytes_acked += delta; 3659 tcp_snd_sne_update(tp, ack); 3660 tp->snd_una = ack; 3661 } 3662 3663 static void tcp_rcv_sne_update(struct tcp_sock *tp, u32 seq) 3664 { 3665 #ifdef CONFIG_TCP_AO 3666 struct tcp_ao_info *ao; 3667 3668 if (!static_branch_unlikely(&tcp_ao_needed.key)) 3669 return; 3670 3671 ao = rcu_dereference_protected(tp->ao_info, 3672 lockdep_sock_is_held((struct sock *)tp)); 3673 if (ao && seq < tp->rcv_nxt) { 3674 ao->rcv_sne++; 3675 trace_tcp_ao_rcv_sne_update((struct sock *)tp, ao->rcv_sne); 3676 } 3677 #endif 3678 } 3679 3680 /* If we update tp->rcv_nxt, also update tp->bytes_received */ 3681 static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq) 3682 { 3683 u32 delta = seq - tp->rcv_nxt; 3684 3685 sock_owned_by_me((struct sock *)tp); 3686 tp->bytes_received += delta; 3687 tcp_rcv_sne_update(tp, seq); 3688 WRITE_ONCE(tp->rcv_nxt, seq); 3689 } 3690 3691 /* Update our send window. 3692 * 3693 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2 3694 * and in FreeBSD. NetBSD's one is even worse.) is wrong. 3695 */ 3696 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack, 3697 u32 ack_seq) 3698 { 3699 struct tcp_sock *tp = tcp_sk(sk); 3700 int flag = 0; 3701 u32 nwin = ntohs(tcp_hdr(skb)->window); 3702 3703 if (likely(!tcp_hdr(skb)->syn)) 3704 nwin <<= tp->rx_opt.snd_wscale; 3705 3706 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) { 3707 flag |= FLAG_WIN_UPDATE; 3708 tcp_update_wl(tp, ack_seq); 3709 3710 if (tp->snd_wnd != nwin) { 3711 tp->snd_wnd = nwin; 3712 3713 /* Note, it is the only place, where 3714 * fast path is recovered for sending TCP. 3715 */ 3716 tp->pred_flags = 0; 3717 tcp_fast_path_check(sk); 3718 3719 if (!tcp_write_queue_empty(sk)) 3720 tcp_slow_start_after_idle_check(sk); 3721 3722 if (nwin > tp->max_window) { 3723 tp->max_window = nwin; 3724 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie); 3725 } 3726 } 3727 } 3728 3729 tcp_snd_una_update(tp, ack); 3730 3731 return flag; 3732 } 3733 3734 static bool __tcp_oow_rate_limited(struct net *net, int mib_idx, 3735 u32 *last_oow_ack_time) 3736 { 3737 /* Paired with the WRITE_ONCE() in this function. */ 3738 u32 val = READ_ONCE(*last_oow_ack_time); 3739 3740 if (val) { 3741 s32 elapsed = (s32)(tcp_jiffies32 - val); 3742 3743 if (0 <= elapsed && 3744 elapsed < READ_ONCE(net->ipv4.sysctl_tcp_invalid_ratelimit)) { 3745 NET_INC_STATS(net, mib_idx); 3746 return true; /* rate-limited: don't send yet! */ 3747 } 3748 } 3749 3750 /* Paired with the prior READ_ONCE() and with itself, 3751 * as we might be lockless. 3752 */ 3753 WRITE_ONCE(*last_oow_ack_time, tcp_jiffies32); 3754 3755 return false; /* not rate-limited: go ahead, send dupack now! */ 3756 } 3757 3758 /* Return true if we're currently rate-limiting out-of-window ACKs and 3759 * thus shouldn't send a dupack right now. We rate-limit dupacks in 3760 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS 3761 * attacks that send repeated SYNs or ACKs for the same connection. To 3762 * do this, we do not send a duplicate SYNACK or ACK if the remote 3763 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate. 3764 */ 3765 bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb, 3766 int mib_idx, u32 *last_oow_ack_time) 3767 { 3768 /* Data packets without SYNs are not likely part of an ACK loop. */ 3769 if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) && 3770 !tcp_hdr(skb)->syn) 3771 return false; 3772 3773 return __tcp_oow_rate_limited(net, mib_idx, last_oow_ack_time); 3774 } 3775 3776 /* RFC 5961 7 [ACK Throttling] */ 3777 static void tcp_send_challenge_ack(struct sock *sk) 3778 { 3779 struct tcp_sock *tp = tcp_sk(sk); 3780 struct net *net = sock_net(sk); 3781 u32 count, now, ack_limit; 3782 3783 /* First check our per-socket dupack rate limit. */ 3784 if (__tcp_oow_rate_limited(net, 3785 LINUX_MIB_TCPACKSKIPPEDCHALLENGE, 3786 &tp->last_oow_ack_time)) 3787 return; 3788 3789 ack_limit = READ_ONCE(net->ipv4.sysctl_tcp_challenge_ack_limit); 3790 if (ack_limit == INT_MAX) 3791 goto send_ack; 3792 3793 /* Then check host-wide RFC 5961 rate limit. */ 3794 now = jiffies / HZ; 3795 if (now != READ_ONCE(net->ipv4.tcp_challenge_timestamp)) { 3796 u32 half = (ack_limit + 1) >> 1; 3797 3798 WRITE_ONCE(net->ipv4.tcp_challenge_timestamp, now); 3799 WRITE_ONCE(net->ipv4.tcp_challenge_count, 3800 get_random_u32_inclusive(half, ack_limit + half - 1)); 3801 } 3802 count = READ_ONCE(net->ipv4.tcp_challenge_count); 3803 if (count > 0) { 3804 WRITE_ONCE(net->ipv4.tcp_challenge_count, count - 1); 3805 send_ack: 3806 NET_INC_STATS(net, LINUX_MIB_TCPCHALLENGEACK); 3807 tcp_send_ack(sk); 3808 } 3809 } 3810 3811 static void tcp_store_ts_recent(struct tcp_sock *tp) 3812 { 3813 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval; 3814 tp->rx_opt.ts_recent_stamp = ktime_get_seconds(); 3815 } 3816 3817 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq) 3818 { 3819 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) { 3820 /* PAWS bug workaround wrt. ACK frames, the PAWS discard 3821 * extra check below makes sure this can only happen 3822 * for pure ACK frames. -DaveM 3823 * 3824 * Not only, also it occurs for expired timestamps. 3825 */ 3826 3827 if (tcp_paws_check(&tp->rx_opt, 0)) 3828 tcp_store_ts_recent(tp); 3829 } 3830 } 3831 3832 /* This routine deals with acks during a TLP episode and ends an episode by 3833 * resetting tlp_high_seq. Ref: TLP algorithm in draft-ietf-tcpm-rack 3834 */ 3835 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag) 3836 { 3837 struct tcp_sock *tp = tcp_sk(sk); 3838 3839 if (before(ack, tp->tlp_high_seq)) 3840 return; 3841 3842 if (!tp->tlp_retrans) { 3843 /* TLP of new data has been acknowledged */ 3844 tp->tlp_high_seq = 0; 3845 } else if (flag & FLAG_DSACK_TLP) { 3846 /* This DSACK means original and TLP probe arrived; no loss */ 3847 tp->tlp_high_seq = 0; 3848 } else if (after(ack, tp->tlp_high_seq)) { 3849 /* ACK advances: there was a loss, so reduce cwnd. Reset 3850 * tlp_high_seq in tcp_init_cwnd_reduction() 3851 */ 3852 tcp_init_cwnd_reduction(sk); 3853 tcp_set_ca_state(sk, TCP_CA_CWR); 3854 tcp_end_cwnd_reduction(sk); 3855 tcp_try_keep_open(sk); 3856 NET_INC_STATS(sock_net(sk), 3857 LINUX_MIB_TCPLOSSPROBERECOVERY); 3858 } else if (!(flag & (FLAG_SND_UNA_ADVANCED | 3859 FLAG_NOT_DUP | FLAG_DATA_SACKED))) { 3860 /* Pure dupack: original and TLP probe arrived; no loss */ 3861 tp->tlp_high_seq = 0; 3862 } 3863 } 3864 3865 static inline void tcp_in_ack_event(struct sock *sk, u32 flags) 3866 { 3867 const struct inet_connection_sock *icsk = inet_csk(sk); 3868 3869 if (icsk->icsk_ca_ops->in_ack_event) 3870 icsk->icsk_ca_ops->in_ack_event(sk, flags); 3871 } 3872 3873 /* Congestion control has updated the cwnd already. So if we're in 3874 * loss recovery then now we do any new sends (for FRTO) or 3875 * retransmits (for CA_Loss or CA_recovery) that make sense. 3876 */ 3877 static void tcp_xmit_recovery(struct sock *sk, int rexmit) 3878 { 3879 struct tcp_sock *tp = tcp_sk(sk); 3880 3881 if (rexmit == REXMIT_NONE || sk->sk_state == TCP_SYN_SENT) 3882 return; 3883 3884 if (unlikely(rexmit == REXMIT_NEW)) { 3885 __tcp_push_pending_frames(sk, tcp_current_mss(sk), 3886 TCP_NAGLE_OFF); 3887 if (after(tp->snd_nxt, tp->high_seq)) 3888 return; 3889 tp->frto = 0; 3890 } 3891 tcp_xmit_retransmit_queue(sk); 3892 } 3893 3894 /* Returns the number of packets newly acked or sacked by the current ACK */ 3895 static u32 tcp_newly_delivered(struct sock *sk, u32 prior_delivered, int flag) 3896 { 3897 const struct net *net = sock_net(sk); 3898 struct tcp_sock *tp = tcp_sk(sk); 3899 u32 delivered; 3900 3901 delivered = tp->delivered - prior_delivered; 3902 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVERED, delivered); 3903 if (flag & FLAG_ECE) 3904 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVEREDCE, delivered); 3905 3906 return delivered; 3907 } 3908 3909 /* This routine deals with incoming acks, but not outgoing ones. */ 3910 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag) 3911 { 3912 struct inet_connection_sock *icsk = inet_csk(sk); 3913 struct tcp_sock *tp = tcp_sk(sk); 3914 struct tcp_sacktag_state sack_state; 3915 struct rate_sample rs = { .prior_delivered = 0 }; 3916 u32 prior_snd_una = tp->snd_una; 3917 bool is_sack_reneg = tp->is_sack_reneg; 3918 u32 ack_seq = TCP_SKB_CB(skb)->seq; 3919 u32 ack = TCP_SKB_CB(skb)->ack_seq; 3920 int num_dupack = 0; 3921 int prior_packets = tp->packets_out; 3922 u32 delivered = tp->delivered; 3923 u32 lost = tp->lost; 3924 int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */ 3925 u32 prior_fack; 3926 3927 sack_state.first_sackt = 0; 3928 sack_state.rate = &rs; 3929 sack_state.sack_delivered = 0; 3930 3931 /* We very likely will need to access rtx queue. */ 3932 prefetch(sk->tcp_rtx_queue.rb_node); 3933 3934 /* If the ack is older than previous acks 3935 * then we can probably ignore it. 3936 */ 3937 if (before(ack, prior_snd_una)) { 3938 u32 max_window; 3939 3940 /* do not accept ACK for bytes we never sent. */ 3941 max_window = min_t(u64, tp->max_window, tp->bytes_acked); 3942 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */ 3943 if (before(ack, prior_snd_una - max_window)) { 3944 if (!(flag & FLAG_NO_CHALLENGE_ACK)) 3945 tcp_send_challenge_ack(sk); 3946 return -SKB_DROP_REASON_TCP_TOO_OLD_ACK; 3947 } 3948 goto old_ack; 3949 } 3950 3951 /* If the ack includes data we haven't sent yet, discard 3952 * this segment (RFC793 Section 3.9). 3953 */ 3954 if (after(ack, tp->snd_nxt)) 3955 return -SKB_DROP_REASON_TCP_ACK_UNSENT_DATA; 3956 3957 if (after(ack, prior_snd_una)) { 3958 flag |= FLAG_SND_UNA_ADVANCED; 3959 icsk->icsk_retransmits = 0; 3960 3961 #if IS_ENABLED(CONFIG_TLS_DEVICE) 3962 if (static_branch_unlikely(&clean_acked_data_enabled.key)) 3963 if (icsk->icsk_clean_acked) 3964 icsk->icsk_clean_acked(sk, ack); 3965 #endif 3966 } 3967 3968 prior_fack = tcp_is_sack(tp) ? tcp_highest_sack_seq(tp) : tp->snd_una; 3969 rs.prior_in_flight = tcp_packets_in_flight(tp); 3970 3971 /* ts_recent update must be made after we are sure that the packet 3972 * is in window. 3973 */ 3974 if (flag & FLAG_UPDATE_TS_RECENT) 3975 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq); 3976 3977 if ((flag & (FLAG_SLOWPATH | FLAG_SND_UNA_ADVANCED)) == 3978 FLAG_SND_UNA_ADVANCED) { 3979 /* Window is constant, pure forward advance. 3980 * No more checks are required. 3981 * Note, we use the fact that SND.UNA>=SND.WL2. 3982 */ 3983 tcp_update_wl(tp, ack_seq); 3984 tcp_snd_una_update(tp, ack); 3985 flag |= FLAG_WIN_UPDATE; 3986 3987 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE); 3988 3989 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPACKS); 3990 } else { 3991 u32 ack_ev_flags = CA_ACK_SLOWPATH; 3992 3993 if (ack_seq != TCP_SKB_CB(skb)->end_seq) 3994 flag |= FLAG_DATA; 3995 else 3996 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPUREACKS); 3997 3998 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq); 3999 4000 if (TCP_SKB_CB(skb)->sacked) 4001 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una, 4002 &sack_state); 4003 4004 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) { 4005 flag |= FLAG_ECE; 4006 ack_ev_flags |= CA_ACK_ECE; 4007 } 4008 4009 if (sack_state.sack_delivered) 4010 tcp_count_delivered(tp, sack_state.sack_delivered, 4011 flag & FLAG_ECE); 4012 4013 if (flag & FLAG_WIN_UPDATE) 4014 ack_ev_flags |= CA_ACK_WIN_UPDATE; 4015 4016 tcp_in_ack_event(sk, ack_ev_flags); 4017 } 4018 4019 /* This is a deviation from RFC3168 since it states that: 4020 * "When the TCP data sender is ready to set the CWR bit after reducing 4021 * the congestion window, it SHOULD set the CWR bit only on the first 4022 * new data packet that it transmits." 4023 * We accept CWR on pure ACKs to be more robust 4024 * with widely-deployed TCP implementations that do this. 4025 */ 4026 tcp_ecn_accept_cwr(sk, skb); 4027 4028 /* We passed data and got it acked, remove any soft error 4029 * log. Something worked... 4030 */ 4031 WRITE_ONCE(sk->sk_err_soft, 0); 4032 icsk->icsk_probes_out = 0; 4033 tp->rcv_tstamp = tcp_jiffies32; 4034 if (!prior_packets) 4035 goto no_queue; 4036 4037 /* See if we can take anything off of the retransmit queue. */ 4038 flag |= tcp_clean_rtx_queue(sk, skb, prior_fack, prior_snd_una, 4039 &sack_state, flag & FLAG_ECE); 4040 4041 tcp_rack_update_reo_wnd(sk, &rs); 4042 4043 if (tp->tlp_high_seq) 4044 tcp_process_tlp_ack(sk, ack, flag); 4045 4046 if (tcp_ack_is_dubious(sk, flag)) { 4047 if (!(flag & (FLAG_SND_UNA_ADVANCED | 4048 FLAG_NOT_DUP | FLAG_DSACKING_ACK))) { 4049 num_dupack = 1; 4050 /* Consider if pure acks were aggregated in tcp_add_backlog() */ 4051 if (!(flag & FLAG_DATA)) 4052 num_dupack = max_t(u16, 1, skb_shinfo(skb)->gso_segs); 4053 } 4054 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag, 4055 &rexmit); 4056 } 4057 4058 /* If needed, reset TLP/RTO timer when RACK doesn't set. */ 4059 if (flag & FLAG_SET_XMIT_TIMER) 4060 tcp_set_xmit_timer(sk); 4061 4062 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) 4063 sk_dst_confirm(sk); 4064 4065 delivered = tcp_newly_delivered(sk, delivered, flag); 4066 lost = tp->lost - lost; /* freshly marked lost */ 4067 rs.is_ack_delayed = !!(flag & FLAG_ACK_MAYBE_DELAYED); 4068 tcp_rate_gen(sk, delivered, lost, is_sack_reneg, sack_state.rate); 4069 tcp_cong_control(sk, ack, delivered, flag, sack_state.rate); 4070 tcp_xmit_recovery(sk, rexmit); 4071 return 1; 4072 4073 no_queue: 4074 /* If data was DSACKed, see if we can undo a cwnd reduction. */ 4075 if (flag & FLAG_DSACKING_ACK) { 4076 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag, 4077 &rexmit); 4078 tcp_newly_delivered(sk, delivered, flag); 4079 } 4080 /* If this ack opens up a zero window, clear backoff. It was 4081 * being used to time the probes, and is probably far higher than 4082 * it needs to be for normal retransmission. 4083 */ 4084 tcp_ack_probe(sk); 4085 4086 if (tp->tlp_high_seq) 4087 tcp_process_tlp_ack(sk, ack, flag); 4088 return 1; 4089 4090 old_ack: 4091 /* If data was SACKed, tag it and see if we should send more data. 4092 * If data was DSACKed, see if we can undo a cwnd reduction. 4093 */ 4094 if (TCP_SKB_CB(skb)->sacked) { 4095 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una, 4096 &sack_state); 4097 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag, 4098 &rexmit); 4099 tcp_newly_delivered(sk, delivered, flag); 4100 tcp_xmit_recovery(sk, rexmit); 4101 } 4102 4103 return 0; 4104 } 4105 4106 static void tcp_parse_fastopen_option(int len, const unsigned char *cookie, 4107 bool syn, struct tcp_fastopen_cookie *foc, 4108 bool exp_opt) 4109 { 4110 /* Valid only in SYN or SYN-ACK with an even length. */ 4111 if (!foc || !syn || len < 0 || (len & 1)) 4112 return; 4113 4114 if (len >= TCP_FASTOPEN_COOKIE_MIN && 4115 len <= TCP_FASTOPEN_COOKIE_MAX) 4116 memcpy(foc->val, cookie, len); 4117 else if (len != 0) 4118 len = -1; 4119 foc->len = len; 4120 foc->exp = exp_opt; 4121 } 4122 4123 static bool smc_parse_options(const struct tcphdr *th, 4124 struct tcp_options_received *opt_rx, 4125 const unsigned char *ptr, 4126 int opsize) 4127 { 4128 #if IS_ENABLED(CONFIG_SMC) 4129 if (static_branch_unlikely(&tcp_have_smc)) { 4130 if (th->syn && !(opsize & 1) && 4131 opsize >= TCPOLEN_EXP_SMC_BASE && 4132 get_unaligned_be32(ptr) == TCPOPT_SMC_MAGIC) { 4133 opt_rx->smc_ok = 1; 4134 return true; 4135 } 4136 } 4137 #endif 4138 return false; 4139 } 4140 4141 /* Try to parse the MSS option from the TCP header. Return 0 on failure, clamped 4142 * value on success. 4143 */ 4144 u16 tcp_parse_mss_option(const struct tcphdr *th, u16 user_mss) 4145 { 4146 const unsigned char *ptr = (const unsigned char *)(th + 1); 4147 int length = (th->doff * 4) - sizeof(struct tcphdr); 4148 u16 mss = 0; 4149 4150 while (length > 0) { 4151 int opcode = *ptr++; 4152 int opsize; 4153 4154 switch (opcode) { 4155 case TCPOPT_EOL: 4156 return mss; 4157 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */ 4158 length--; 4159 continue; 4160 default: 4161 if (length < 2) 4162 return mss; 4163 opsize = *ptr++; 4164 if (opsize < 2) /* "silly options" */ 4165 return mss; 4166 if (opsize > length) 4167 return mss; /* fail on partial options */ 4168 if (opcode == TCPOPT_MSS && opsize == TCPOLEN_MSS) { 4169 u16 in_mss = get_unaligned_be16(ptr); 4170 4171 if (in_mss) { 4172 if (user_mss && user_mss < in_mss) 4173 in_mss = user_mss; 4174 mss = in_mss; 4175 } 4176 } 4177 ptr += opsize - 2; 4178 length -= opsize; 4179 } 4180 } 4181 return mss; 4182 } 4183 EXPORT_SYMBOL_GPL(tcp_parse_mss_option); 4184 4185 /* Look for tcp options. Normally only called on SYN and SYNACK packets. 4186 * But, this can also be called on packets in the established flow when 4187 * the fast version below fails. 4188 */ 4189 void tcp_parse_options(const struct net *net, 4190 const struct sk_buff *skb, 4191 struct tcp_options_received *opt_rx, int estab, 4192 struct tcp_fastopen_cookie *foc) 4193 { 4194 const unsigned char *ptr; 4195 const struct tcphdr *th = tcp_hdr(skb); 4196 int length = (th->doff * 4) - sizeof(struct tcphdr); 4197 4198 ptr = (const unsigned char *)(th + 1); 4199 opt_rx->saw_tstamp = 0; 4200 opt_rx->saw_unknown = 0; 4201 4202 while (length > 0) { 4203 int opcode = *ptr++; 4204 int opsize; 4205 4206 switch (opcode) { 4207 case TCPOPT_EOL: 4208 return; 4209 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */ 4210 length--; 4211 continue; 4212 default: 4213 if (length < 2) 4214 return; 4215 opsize = *ptr++; 4216 if (opsize < 2) /* "silly options" */ 4217 return; 4218 if (opsize > length) 4219 return; /* don't parse partial options */ 4220 switch (opcode) { 4221 case TCPOPT_MSS: 4222 if (opsize == TCPOLEN_MSS && th->syn && !estab) { 4223 u16 in_mss = get_unaligned_be16(ptr); 4224 if (in_mss) { 4225 if (opt_rx->user_mss && 4226 opt_rx->user_mss < in_mss) 4227 in_mss = opt_rx->user_mss; 4228 opt_rx->mss_clamp = in_mss; 4229 } 4230 } 4231 break; 4232 case TCPOPT_WINDOW: 4233 if (opsize == TCPOLEN_WINDOW && th->syn && 4234 !estab && READ_ONCE(net->ipv4.sysctl_tcp_window_scaling)) { 4235 __u8 snd_wscale = *(__u8 *)ptr; 4236 opt_rx->wscale_ok = 1; 4237 if (snd_wscale > TCP_MAX_WSCALE) { 4238 net_info_ratelimited("%s: Illegal window scaling value %d > %u received\n", 4239 __func__, 4240 snd_wscale, 4241 TCP_MAX_WSCALE); 4242 snd_wscale = TCP_MAX_WSCALE; 4243 } 4244 opt_rx->snd_wscale = snd_wscale; 4245 } 4246 break; 4247 case TCPOPT_TIMESTAMP: 4248 if ((opsize == TCPOLEN_TIMESTAMP) && 4249 ((estab && opt_rx->tstamp_ok) || 4250 (!estab && READ_ONCE(net->ipv4.sysctl_tcp_timestamps)))) { 4251 opt_rx->saw_tstamp = 1; 4252 opt_rx->rcv_tsval = get_unaligned_be32(ptr); 4253 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4); 4254 } 4255 break; 4256 case TCPOPT_SACK_PERM: 4257 if (opsize == TCPOLEN_SACK_PERM && th->syn && 4258 !estab && READ_ONCE(net->ipv4.sysctl_tcp_sack)) { 4259 opt_rx->sack_ok = TCP_SACK_SEEN; 4260 tcp_sack_reset(opt_rx); 4261 } 4262 break; 4263 4264 case TCPOPT_SACK: 4265 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) && 4266 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) && 4267 opt_rx->sack_ok) { 4268 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th; 4269 } 4270 break; 4271 #ifdef CONFIG_TCP_MD5SIG 4272 case TCPOPT_MD5SIG: 4273 /* The MD5 Hash has already been 4274 * checked (see tcp_v{4,6}_rcv()). 4275 */ 4276 break; 4277 #endif 4278 #ifdef CONFIG_TCP_AO 4279 case TCPOPT_AO: 4280 /* TCP AO has already been checked 4281 * (see tcp_inbound_ao_hash()). 4282 */ 4283 break; 4284 #endif 4285 case TCPOPT_FASTOPEN: 4286 tcp_parse_fastopen_option( 4287 opsize - TCPOLEN_FASTOPEN_BASE, 4288 ptr, th->syn, foc, false); 4289 break; 4290 4291 case TCPOPT_EXP: 4292 /* Fast Open option shares code 254 using a 4293 * 16 bits magic number. 4294 */ 4295 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE && 4296 get_unaligned_be16(ptr) == 4297 TCPOPT_FASTOPEN_MAGIC) { 4298 tcp_parse_fastopen_option(opsize - 4299 TCPOLEN_EXP_FASTOPEN_BASE, 4300 ptr + 2, th->syn, foc, true); 4301 break; 4302 } 4303 4304 if (smc_parse_options(th, opt_rx, ptr, opsize)) 4305 break; 4306 4307 opt_rx->saw_unknown = 1; 4308 break; 4309 4310 default: 4311 opt_rx->saw_unknown = 1; 4312 } 4313 ptr += opsize-2; 4314 length -= opsize; 4315 } 4316 } 4317 } 4318 EXPORT_SYMBOL(tcp_parse_options); 4319 4320 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th) 4321 { 4322 const __be32 *ptr = (const __be32 *)(th + 1); 4323 4324 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) 4325 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) { 4326 tp->rx_opt.saw_tstamp = 1; 4327 ++ptr; 4328 tp->rx_opt.rcv_tsval = ntohl(*ptr); 4329 ++ptr; 4330 if (*ptr) 4331 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset; 4332 else 4333 tp->rx_opt.rcv_tsecr = 0; 4334 return true; 4335 } 4336 return false; 4337 } 4338 4339 /* Fast parse options. This hopes to only see timestamps. 4340 * If it is wrong it falls back on tcp_parse_options(). 4341 */ 4342 static bool tcp_fast_parse_options(const struct net *net, 4343 const struct sk_buff *skb, 4344 const struct tcphdr *th, struct tcp_sock *tp) 4345 { 4346 /* In the spirit of fast parsing, compare doff directly to constant 4347 * values. Because equality is used, short doff can be ignored here. 4348 */ 4349 if (th->doff == (sizeof(*th) / 4)) { 4350 tp->rx_opt.saw_tstamp = 0; 4351 return false; 4352 } else if (tp->rx_opt.tstamp_ok && 4353 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) { 4354 if (tcp_parse_aligned_timestamp(tp, th)) 4355 return true; 4356 } 4357 4358 tcp_parse_options(net, skb, &tp->rx_opt, 1, NULL); 4359 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr) 4360 tp->rx_opt.rcv_tsecr -= tp->tsoffset; 4361 4362 return true; 4363 } 4364 4365 #if defined(CONFIG_TCP_MD5SIG) || defined(CONFIG_TCP_AO) 4366 /* 4367 * Parse Signature options 4368 */ 4369 int tcp_do_parse_auth_options(const struct tcphdr *th, 4370 const u8 **md5_hash, const u8 **ao_hash) 4371 { 4372 int length = (th->doff << 2) - sizeof(*th); 4373 const u8 *ptr = (const u8 *)(th + 1); 4374 unsigned int minlen = TCPOLEN_MD5SIG; 4375 4376 if (IS_ENABLED(CONFIG_TCP_AO)) 4377 minlen = sizeof(struct tcp_ao_hdr) + 1; 4378 4379 *md5_hash = NULL; 4380 *ao_hash = NULL; 4381 4382 /* If not enough data remaining, we can short cut */ 4383 while (length >= minlen) { 4384 int opcode = *ptr++; 4385 int opsize; 4386 4387 switch (opcode) { 4388 case TCPOPT_EOL: 4389 return 0; 4390 case TCPOPT_NOP: 4391 length--; 4392 continue; 4393 default: 4394 opsize = *ptr++; 4395 if (opsize < 2 || opsize > length) 4396 return -EINVAL; 4397 if (opcode == TCPOPT_MD5SIG) { 4398 if (opsize != TCPOLEN_MD5SIG) 4399 return -EINVAL; 4400 if (unlikely(*md5_hash || *ao_hash)) 4401 return -EEXIST; 4402 *md5_hash = ptr; 4403 } else if (opcode == TCPOPT_AO) { 4404 if (opsize <= sizeof(struct tcp_ao_hdr)) 4405 return -EINVAL; 4406 if (unlikely(*md5_hash || *ao_hash)) 4407 return -EEXIST; 4408 *ao_hash = ptr; 4409 } 4410 } 4411 ptr += opsize - 2; 4412 length -= opsize; 4413 } 4414 return 0; 4415 } 4416 EXPORT_SYMBOL(tcp_do_parse_auth_options); 4417 #endif 4418 4419 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM 4420 * 4421 * It is not fatal. If this ACK does _not_ change critical state (seqs, window) 4422 * it can pass through stack. So, the following predicate verifies that 4423 * this segment is not used for anything but congestion avoidance or 4424 * fast retransmit. Moreover, we even are able to eliminate most of such 4425 * second order effects, if we apply some small "replay" window (~RTO) 4426 * to timestamp space. 4427 * 4428 * All these measures still do not guarantee that we reject wrapped ACKs 4429 * on networks with high bandwidth, when sequence space is recycled fastly, 4430 * but it guarantees that such events will be very rare and do not affect 4431 * connection seriously. This doesn't look nice, but alas, PAWS is really 4432 * buggy extension. 4433 * 4434 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC 4435 * states that events when retransmit arrives after original data are rare. 4436 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is 4437 * the biggest problem on large power networks even with minor reordering. 4438 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe 4439 * up to bandwidth of 18Gigabit/sec. 8) ] 4440 */ 4441 4442 /* Estimates max number of increments of remote peer TSval in 4443 * a replay window (based on our current RTO estimation). 4444 */ 4445 static u32 tcp_tsval_replay(const struct sock *sk) 4446 { 4447 /* If we use usec TS resolution, 4448 * then expect the remote peer to use the same resolution. 4449 */ 4450 if (tcp_sk(sk)->tcp_usec_ts) 4451 return inet_csk(sk)->icsk_rto * (USEC_PER_SEC / HZ); 4452 4453 /* RFC 7323 recommends a TSval clock between 1ms and 1sec. 4454 * We know that some OS (including old linux) can use 1200 Hz. 4455 */ 4456 return inet_csk(sk)->icsk_rto * 1200 / HZ; 4457 } 4458 4459 static enum skb_drop_reason tcp_disordered_ack_check(const struct sock *sk, 4460 const struct sk_buff *skb) 4461 { 4462 const struct tcp_sock *tp = tcp_sk(sk); 4463 const struct tcphdr *th = tcp_hdr(skb); 4464 SKB_DR_INIT(reason, TCP_RFC7323_PAWS); 4465 u32 ack = TCP_SKB_CB(skb)->ack_seq; 4466 u32 seq = TCP_SKB_CB(skb)->seq; 4467 4468 /* 1. Is this not a pure ACK ? */ 4469 if (!th->ack || seq != TCP_SKB_CB(skb)->end_seq) 4470 return reason; 4471 4472 /* 2. Is its sequence not the expected one ? */ 4473 if (seq != tp->rcv_nxt) 4474 return before(seq, tp->rcv_nxt) ? 4475 SKB_DROP_REASON_TCP_RFC7323_PAWS_ACK : 4476 reason; 4477 4478 /* 3. Is this not a duplicate ACK ? */ 4479 if (ack != tp->snd_una) 4480 return reason; 4481 4482 /* 4. Is this updating the window ? */ 4483 if (tcp_may_update_window(tp, ack, seq, ntohs(th->window) << 4484 tp->rx_opt.snd_wscale)) 4485 return reason; 4486 4487 /* 5. Is this not in the replay window ? */ 4488 if ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > 4489 tcp_tsval_replay(sk)) 4490 return reason; 4491 4492 return 0; 4493 } 4494 4495 /* Check segment sequence number for validity. 4496 * 4497 * Segment controls are considered valid, if the segment 4498 * fits to the window after truncation to the window. Acceptability 4499 * of data (and SYN, FIN, of course) is checked separately. 4500 * See tcp_data_queue(), for example. 4501 * 4502 * Also, controls (RST is main one) are accepted using RCV.WUP instead 4503 * of RCV.NXT. Peer still did not advance his SND.UNA when we 4504 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP. 4505 * (borrowed from freebsd) 4506 */ 4507 4508 static enum skb_drop_reason tcp_sequence(const struct tcp_sock *tp, 4509 u32 seq, u32 end_seq) 4510 { 4511 if (before(end_seq, tp->rcv_wup)) 4512 return SKB_DROP_REASON_TCP_OLD_SEQUENCE; 4513 4514 if (after(seq, tp->rcv_nxt + tcp_receive_window(tp))) 4515 return SKB_DROP_REASON_TCP_INVALID_SEQUENCE; 4516 4517 return SKB_NOT_DROPPED_YET; 4518 } 4519 4520 4521 void tcp_done_with_error(struct sock *sk, int err) 4522 { 4523 /* This barrier is coupled with smp_rmb() in tcp_poll() */ 4524 WRITE_ONCE(sk->sk_err, err); 4525 smp_wmb(); 4526 4527 tcp_write_queue_purge(sk); 4528 tcp_done(sk); 4529 4530 if (!sock_flag(sk, SOCK_DEAD)) 4531 sk_error_report(sk); 4532 } 4533 EXPORT_SYMBOL(tcp_done_with_error); 4534 4535 /* When we get a reset we do this. */ 4536 void tcp_reset(struct sock *sk, struct sk_buff *skb) 4537 { 4538 int err; 4539 4540 trace_tcp_receive_reset(sk); 4541 4542 /* mptcp can't tell us to ignore reset pkts, 4543 * so just ignore the return value of mptcp_incoming_options(). 4544 */ 4545 if (sk_is_mptcp(sk)) 4546 mptcp_incoming_options(sk, skb); 4547 4548 /* We want the right error as BSD sees it (and indeed as we do). */ 4549 switch (sk->sk_state) { 4550 case TCP_SYN_SENT: 4551 err = ECONNREFUSED; 4552 break; 4553 case TCP_CLOSE_WAIT: 4554 err = EPIPE; 4555 break; 4556 case TCP_CLOSE: 4557 return; 4558 default: 4559 err = ECONNRESET; 4560 } 4561 tcp_done_with_error(sk, err); 4562 } 4563 4564 /* 4565 * Process the FIN bit. This now behaves as it is supposed to work 4566 * and the FIN takes effect when it is validly part of sequence 4567 * space. Not before when we get holes. 4568 * 4569 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT 4570 * (and thence onto LAST-ACK and finally, CLOSE, we never enter 4571 * TIME-WAIT) 4572 * 4573 * If we are in FINWAIT-1, a received FIN indicates simultaneous 4574 * close and we go into CLOSING (and later onto TIME-WAIT) 4575 * 4576 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT. 4577 */ 4578 void tcp_fin(struct sock *sk) 4579 { 4580 struct tcp_sock *tp = tcp_sk(sk); 4581 4582 inet_csk_schedule_ack(sk); 4583 4584 WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | RCV_SHUTDOWN); 4585 sock_set_flag(sk, SOCK_DONE); 4586 4587 switch (sk->sk_state) { 4588 case TCP_SYN_RECV: 4589 case TCP_ESTABLISHED: 4590 /* Move to CLOSE_WAIT */ 4591 tcp_set_state(sk, TCP_CLOSE_WAIT); 4592 inet_csk_enter_pingpong_mode(sk); 4593 break; 4594 4595 case TCP_CLOSE_WAIT: 4596 case TCP_CLOSING: 4597 /* Received a retransmission of the FIN, do 4598 * nothing. 4599 */ 4600 break; 4601 case TCP_LAST_ACK: 4602 /* RFC793: Remain in the LAST-ACK state. */ 4603 break; 4604 4605 case TCP_FIN_WAIT1: 4606 /* This case occurs when a simultaneous close 4607 * happens, we must ack the received FIN and 4608 * enter the CLOSING state. 4609 */ 4610 tcp_send_ack(sk); 4611 tcp_set_state(sk, TCP_CLOSING); 4612 break; 4613 case TCP_FIN_WAIT2: 4614 /* Received a FIN -- send ACK and enter TIME_WAIT. */ 4615 tcp_send_ack(sk); 4616 tcp_time_wait(sk, TCP_TIME_WAIT, 0); 4617 break; 4618 default: 4619 /* Only TCP_LISTEN and TCP_CLOSE are left, in these 4620 * cases we should never reach this piece of code. 4621 */ 4622 pr_err("%s: Impossible, sk->sk_state=%d\n", 4623 __func__, sk->sk_state); 4624 break; 4625 } 4626 4627 /* It _is_ possible, that we have something out-of-order _after_ FIN. 4628 * Probably, we should reset in this case. For now drop them. 4629 */ 4630 skb_rbtree_purge(&tp->out_of_order_queue); 4631 if (tcp_is_sack(tp)) 4632 tcp_sack_reset(&tp->rx_opt); 4633 4634 if (!sock_flag(sk, SOCK_DEAD)) { 4635 sk->sk_state_change(sk); 4636 4637 /* Do not send POLL_HUP for half duplex close. */ 4638 if (sk->sk_shutdown == SHUTDOWN_MASK || 4639 sk->sk_state == TCP_CLOSE) 4640 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP); 4641 else 4642 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN); 4643 } 4644 } 4645 4646 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, 4647 u32 end_seq) 4648 { 4649 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) { 4650 if (before(seq, sp->start_seq)) 4651 sp->start_seq = seq; 4652 if (after(end_seq, sp->end_seq)) 4653 sp->end_seq = end_seq; 4654 return true; 4655 } 4656 return false; 4657 } 4658 4659 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq) 4660 { 4661 struct tcp_sock *tp = tcp_sk(sk); 4662 4663 if (tcp_is_sack(tp) && READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_dsack)) { 4664 int mib_idx; 4665 4666 if (before(seq, tp->rcv_nxt)) 4667 mib_idx = LINUX_MIB_TCPDSACKOLDSENT; 4668 else 4669 mib_idx = LINUX_MIB_TCPDSACKOFOSENT; 4670 4671 NET_INC_STATS(sock_net(sk), mib_idx); 4672 4673 tp->rx_opt.dsack = 1; 4674 tp->duplicate_sack[0].start_seq = seq; 4675 tp->duplicate_sack[0].end_seq = end_seq; 4676 } 4677 } 4678 4679 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq) 4680 { 4681 struct tcp_sock *tp = tcp_sk(sk); 4682 4683 if (!tp->rx_opt.dsack) 4684 tcp_dsack_set(sk, seq, end_seq); 4685 else 4686 tcp_sack_extend(tp->duplicate_sack, seq, end_seq); 4687 } 4688 4689 static void tcp_rcv_spurious_retrans(struct sock *sk, const struct sk_buff *skb) 4690 { 4691 /* When the ACK path fails or drops most ACKs, the sender would 4692 * timeout and spuriously retransmit the same segment repeatedly. 4693 * If it seems our ACKs are not reaching the other side, 4694 * based on receiving a duplicate data segment with new flowlabel 4695 * (suggesting the sender suffered an RTO), and we are not already 4696 * repathing due to our own RTO, then rehash the socket to repath our 4697 * packets. 4698 */ 4699 #if IS_ENABLED(CONFIG_IPV6) 4700 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Loss && 4701 skb->protocol == htons(ETH_P_IPV6) && 4702 (tcp_sk(sk)->inet_conn.icsk_ack.lrcv_flowlabel != 4703 ntohl(ip6_flowlabel(ipv6_hdr(skb)))) && 4704 sk_rethink_txhash(sk)) 4705 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDUPLICATEDATAREHASH); 4706 4707 /* Save last flowlabel after a spurious retrans. */ 4708 tcp_save_lrcv_flowlabel(sk, skb); 4709 #endif 4710 } 4711 4712 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb) 4713 { 4714 struct tcp_sock *tp = tcp_sk(sk); 4715 4716 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 4717 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 4718 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST); 4719 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS); 4720 4721 if (tcp_is_sack(tp) && READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_dsack)) { 4722 u32 end_seq = TCP_SKB_CB(skb)->end_seq; 4723 4724 tcp_rcv_spurious_retrans(sk, skb); 4725 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) 4726 end_seq = tp->rcv_nxt; 4727 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq); 4728 } 4729 } 4730 4731 tcp_send_ack(sk); 4732 } 4733 4734 /* These routines update the SACK block as out-of-order packets arrive or 4735 * in-order packets close up the sequence space. 4736 */ 4737 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp) 4738 { 4739 int this_sack; 4740 struct tcp_sack_block *sp = &tp->selective_acks[0]; 4741 struct tcp_sack_block *swalk = sp + 1; 4742 4743 /* See if the recent change to the first SACK eats into 4744 * or hits the sequence space of other SACK blocks, if so coalesce. 4745 */ 4746 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) { 4747 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) { 4748 int i; 4749 4750 /* Zap SWALK, by moving every further SACK up by one slot. 4751 * Decrease num_sacks. 4752 */ 4753 tp->rx_opt.num_sacks--; 4754 for (i = this_sack; i < tp->rx_opt.num_sacks; i++) 4755 sp[i] = sp[i + 1]; 4756 continue; 4757 } 4758 this_sack++; 4759 swalk++; 4760 } 4761 } 4762 4763 void tcp_sack_compress_send_ack(struct sock *sk) 4764 { 4765 struct tcp_sock *tp = tcp_sk(sk); 4766 4767 if (!tp->compressed_ack) 4768 return; 4769 4770 if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1) 4771 __sock_put(sk); 4772 4773 /* Since we have to send one ack finally, 4774 * substract one from tp->compressed_ack to keep 4775 * LINUX_MIB_TCPACKCOMPRESSED accurate. 4776 */ 4777 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED, 4778 tp->compressed_ack - 1); 4779 4780 tp->compressed_ack = 0; 4781 tcp_send_ack(sk); 4782 } 4783 4784 /* Reasonable amount of sack blocks included in TCP SACK option 4785 * The max is 4, but this becomes 3 if TCP timestamps are there. 4786 * Given that SACK packets might be lost, be conservative and use 2. 4787 */ 4788 #define TCP_SACK_BLOCKS_EXPECTED 2 4789 4790 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq) 4791 { 4792 struct tcp_sock *tp = tcp_sk(sk); 4793 struct tcp_sack_block *sp = &tp->selective_acks[0]; 4794 int cur_sacks = tp->rx_opt.num_sacks; 4795 int this_sack; 4796 4797 if (!cur_sacks) 4798 goto new_sack; 4799 4800 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) { 4801 if (tcp_sack_extend(sp, seq, end_seq)) { 4802 if (this_sack >= TCP_SACK_BLOCKS_EXPECTED) 4803 tcp_sack_compress_send_ack(sk); 4804 /* Rotate this_sack to the first one. */ 4805 for (; this_sack > 0; this_sack--, sp--) 4806 swap(*sp, *(sp - 1)); 4807 if (cur_sacks > 1) 4808 tcp_sack_maybe_coalesce(tp); 4809 return; 4810 } 4811 } 4812 4813 if (this_sack >= TCP_SACK_BLOCKS_EXPECTED) 4814 tcp_sack_compress_send_ack(sk); 4815 4816 /* Could not find an adjacent existing SACK, build a new one, 4817 * put it at the front, and shift everyone else down. We 4818 * always know there is at least one SACK present already here. 4819 * 4820 * If the sack array is full, forget about the last one. 4821 */ 4822 if (this_sack >= TCP_NUM_SACKS) { 4823 this_sack--; 4824 tp->rx_opt.num_sacks--; 4825 sp--; 4826 } 4827 for (; this_sack > 0; this_sack--, sp--) 4828 *sp = *(sp - 1); 4829 4830 new_sack: 4831 /* Build the new head SACK, and we're done. */ 4832 sp->start_seq = seq; 4833 sp->end_seq = end_seq; 4834 tp->rx_opt.num_sacks++; 4835 } 4836 4837 /* RCV.NXT advances, some SACKs should be eaten. */ 4838 4839 static void tcp_sack_remove(struct tcp_sock *tp) 4840 { 4841 struct tcp_sack_block *sp = &tp->selective_acks[0]; 4842 int num_sacks = tp->rx_opt.num_sacks; 4843 int this_sack; 4844 4845 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */ 4846 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) { 4847 tp->rx_opt.num_sacks = 0; 4848 return; 4849 } 4850 4851 for (this_sack = 0; this_sack < num_sacks;) { 4852 /* Check if the start of the sack is covered by RCV.NXT. */ 4853 if (!before(tp->rcv_nxt, sp->start_seq)) { 4854 int i; 4855 4856 /* RCV.NXT must cover all the block! */ 4857 WARN_ON(before(tp->rcv_nxt, sp->end_seq)); 4858 4859 /* Zap this SACK, by moving forward any other SACKS. */ 4860 for (i = this_sack+1; i < num_sacks; i++) 4861 tp->selective_acks[i-1] = tp->selective_acks[i]; 4862 num_sacks--; 4863 continue; 4864 } 4865 this_sack++; 4866 sp++; 4867 } 4868 tp->rx_opt.num_sacks = num_sacks; 4869 } 4870 4871 /** 4872 * tcp_try_coalesce - try to merge skb to prior one 4873 * @sk: socket 4874 * @to: prior buffer 4875 * @from: buffer to add in queue 4876 * @fragstolen: pointer to boolean 4877 * 4878 * Before queueing skb @from after @to, try to merge them 4879 * to reduce overall memory use and queue lengths, if cost is small. 4880 * Packets in ofo or receive queues can stay a long time. 4881 * Better try to coalesce them right now to avoid future collapses. 4882 * Returns true if caller should free @from instead of queueing it 4883 */ 4884 static bool tcp_try_coalesce(struct sock *sk, 4885 struct sk_buff *to, 4886 struct sk_buff *from, 4887 bool *fragstolen) 4888 { 4889 int delta; 4890 4891 *fragstolen = false; 4892 4893 /* Its possible this segment overlaps with prior segment in queue */ 4894 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq) 4895 return false; 4896 4897 if (!tcp_skb_can_collapse_rx(to, from)) 4898 return false; 4899 4900 if (!skb_try_coalesce(to, from, fragstolen, &delta)) 4901 return false; 4902 4903 atomic_add(delta, &sk->sk_rmem_alloc); 4904 sk_mem_charge(sk, delta); 4905 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE); 4906 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq; 4907 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq; 4908 TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags; 4909 4910 if (TCP_SKB_CB(from)->has_rxtstamp) { 4911 TCP_SKB_CB(to)->has_rxtstamp = true; 4912 to->tstamp = from->tstamp; 4913 skb_hwtstamps(to)->hwtstamp = skb_hwtstamps(from)->hwtstamp; 4914 } 4915 4916 return true; 4917 } 4918 4919 static bool tcp_ooo_try_coalesce(struct sock *sk, 4920 struct sk_buff *to, 4921 struct sk_buff *from, 4922 bool *fragstolen) 4923 { 4924 bool res = tcp_try_coalesce(sk, to, from, fragstolen); 4925 4926 /* In case tcp_drop_reason() is called later, update to->gso_segs */ 4927 if (res) { 4928 u32 gso_segs = max_t(u16, 1, skb_shinfo(to)->gso_segs) + 4929 max_t(u16, 1, skb_shinfo(from)->gso_segs); 4930 4931 skb_shinfo(to)->gso_segs = min_t(u32, gso_segs, 0xFFFF); 4932 } 4933 return res; 4934 } 4935 4936 noinline_for_tracing static void 4937 tcp_drop_reason(struct sock *sk, struct sk_buff *skb, enum skb_drop_reason reason) 4938 { 4939 sk_drops_add(sk, skb); 4940 sk_skb_reason_drop(sk, skb, reason); 4941 } 4942 4943 /* This one checks to see if we can put data from the 4944 * out_of_order queue into the receive_queue. 4945 */ 4946 static void tcp_ofo_queue(struct sock *sk) 4947 { 4948 struct tcp_sock *tp = tcp_sk(sk); 4949 __u32 dsack_high = tp->rcv_nxt; 4950 bool fin, fragstolen, eaten; 4951 struct sk_buff *skb, *tail; 4952 struct rb_node *p; 4953 4954 p = rb_first(&tp->out_of_order_queue); 4955 while (p) { 4956 skb = rb_to_skb(p); 4957 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) 4958 break; 4959 4960 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) { 4961 __u32 dsack = dsack_high; 4962 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high)) 4963 dsack_high = TCP_SKB_CB(skb)->end_seq; 4964 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack); 4965 } 4966 p = rb_next(p); 4967 rb_erase(&skb->rbnode, &tp->out_of_order_queue); 4968 4969 if (unlikely(!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))) { 4970 tcp_drop_reason(sk, skb, SKB_DROP_REASON_TCP_OFO_DROP); 4971 continue; 4972 } 4973 4974 tail = skb_peek_tail(&sk->sk_receive_queue); 4975 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen); 4976 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq); 4977 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN; 4978 if (!eaten) 4979 tcp_add_receive_queue(sk, skb); 4980 else 4981 kfree_skb_partial(skb, fragstolen); 4982 4983 if (unlikely(fin)) { 4984 tcp_fin(sk); 4985 /* tcp_fin() purges tp->out_of_order_queue, 4986 * so we must end this loop right now. 4987 */ 4988 break; 4989 } 4990 } 4991 } 4992 4993 static bool tcp_prune_ofo_queue(struct sock *sk, const struct sk_buff *in_skb); 4994 static int tcp_prune_queue(struct sock *sk, const struct sk_buff *in_skb); 4995 4996 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb, 4997 unsigned int size) 4998 { 4999 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf || 5000 !sk_rmem_schedule(sk, skb, size)) { 5001 5002 if (tcp_prune_queue(sk, skb) < 0) 5003 return -1; 5004 5005 while (!sk_rmem_schedule(sk, skb, size)) { 5006 if (!tcp_prune_ofo_queue(sk, skb)) 5007 return -1; 5008 } 5009 } 5010 return 0; 5011 } 5012 5013 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb) 5014 { 5015 struct tcp_sock *tp = tcp_sk(sk); 5016 struct rb_node **p, *parent; 5017 struct sk_buff *skb1; 5018 u32 seq, end_seq; 5019 bool fragstolen; 5020 5021 tcp_save_lrcv_flowlabel(sk, skb); 5022 tcp_ecn_check_ce(sk, skb); 5023 5024 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) { 5025 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFODROP); 5026 sk->sk_data_ready(sk); 5027 tcp_drop_reason(sk, skb, SKB_DROP_REASON_PROTO_MEM); 5028 return; 5029 } 5030 5031 /* Disable header prediction. */ 5032 tp->pred_flags = 0; 5033 inet_csk_schedule_ack(sk); 5034 5035 tp->rcv_ooopack += max_t(u16, 1, skb_shinfo(skb)->gso_segs); 5036 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOQUEUE); 5037 seq = TCP_SKB_CB(skb)->seq; 5038 end_seq = TCP_SKB_CB(skb)->end_seq; 5039 5040 p = &tp->out_of_order_queue.rb_node; 5041 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) { 5042 /* Initial out of order segment, build 1 SACK. */ 5043 if (tcp_is_sack(tp)) { 5044 tp->rx_opt.num_sacks = 1; 5045 tp->selective_acks[0].start_seq = seq; 5046 tp->selective_acks[0].end_seq = end_seq; 5047 } 5048 rb_link_node(&skb->rbnode, NULL, p); 5049 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue); 5050 tp->ooo_last_skb = skb; 5051 goto end; 5052 } 5053 5054 /* In the typical case, we are adding an skb to the end of the list. 5055 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup. 5056 */ 5057 if (tcp_ooo_try_coalesce(sk, tp->ooo_last_skb, 5058 skb, &fragstolen)) { 5059 coalesce_done: 5060 /* For non sack flows, do not grow window to force DUPACK 5061 * and trigger fast retransmit. 5062 */ 5063 if (tcp_is_sack(tp)) 5064 tcp_grow_window(sk, skb, true); 5065 kfree_skb_partial(skb, fragstolen); 5066 skb = NULL; 5067 goto add_sack; 5068 } 5069 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */ 5070 if (!before(seq, TCP_SKB_CB(tp->ooo_last_skb)->end_seq)) { 5071 parent = &tp->ooo_last_skb->rbnode; 5072 p = &parent->rb_right; 5073 goto insert; 5074 } 5075 5076 /* Find place to insert this segment. Handle overlaps on the way. */ 5077 parent = NULL; 5078 while (*p) { 5079 parent = *p; 5080 skb1 = rb_to_skb(parent); 5081 if (before(seq, TCP_SKB_CB(skb1)->seq)) { 5082 p = &parent->rb_left; 5083 continue; 5084 } 5085 if (before(seq, TCP_SKB_CB(skb1)->end_seq)) { 5086 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) { 5087 /* All the bits are present. Drop. */ 5088 NET_INC_STATS(sock_net(sk), 5089 LINUX_MIB_TCPOFOMERGE); 5090 tcp_drop_reason(sk, skb, 5091 SKB_DROP_REASON_TCP_OFOMERGE); 5092 skb = NULL; 5093 tcp_dsack_set(sk, seq, end_seq); 5094 goto add_sack; 5095 } 5096 if (after(seq, TCP_SKB_CB(skb1)->seq)) { 5097 /* Partial overlap. */ 5098 tcp_dsack_set(sk, seq, TCP_SKB_CB(skb1)->end_seq); 5099 } else { 5100 /* skb's seq == skb1's seq and skb covers skb1. 5101 * Replace skb1 with skb. 5102 */ 5103 rb_replace_node(&skb1->rbnode, &skb->rbnode, 5104 &tp->out_of_order_queue); 5105 tcp_dsack_extend(sk, 5106 TCP_SKB_CB(skb1)->seq, 5107 TCP_SKB_CB(skb1)->end_seq); 5108 NET_INC_STATS(sock_net(sk), 5109 LINUX_MIB_TCPOFOMERGE); 5110 tcp_drop_reason(sk, skb1, 5111 SKB_DROP_REASON_TCP_OFOMERGE); 5112 goto merge_right; 5113 } 5114 } else if (tcp_ooo_try_coalesce(sk, skb1, 5115 skb, &fragstolen)) { 5116 goto coalesce_done; 5117 } 5118 p = &parent->rb_right; 5119 } 5120 insert: 5121 /* Insert segment into RB tree. */ 5122 rb_link_node(&skb->rbnode, parent, p); 5123 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue); 5124 5125 merge_right: 5126 /* Remove other segments covered by skb. */ 5127 while ((skb1 = skb_rb_next(skb)) != NULL) { 5128 if (!after(end_seq, TCP_SKB_CB(skb1)->seq)) 5129 break; 5130 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) { 5131 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq, 5132 end_seq); 5133 break; 5134 } 5135 rb_erase(&skb1->rbnode, &tp->out_of_order_queue); 5136 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq, 5137 TCP_SKB_CB(skb1)->end_seq); 5138 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOMERGE); 5139 tcp_drop_reason(sk, skb1, SKB_DROP_REASON_TCP_OFOMERGE); 5140 } 5141 /* If there is no skb after us, we are the last_skb ! */ 5142 if (!skb1) 5143 tp->ooo_last_skb = skb; 5144 5145 add_sack: 5146 if (tcp_is_sack(tp)) 5147 tcp_sack_new_ofo_skb(sk, seq, end_seq); 5148 end: 5149 if (skb) { 5150 /* For non sack flows, do not grow window to force DUPACK 5151 * and trigger fast retransmit. 5152 */ 5153 if (tcp_is_sack(tp)) 5154 tcp_grow_window(sk, skb, false); 5155 skb_condense(skb); 5156 skb_set_owner_r(skb, sk); 5157 } 5158 } 5159 5160 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, 5161 bool *fragstolen) 5162 { 5163 int eaten; 5164 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue); 5165 5166 eaten = (tail && 5167 tcp_try_coalesce(sk, tail, 5168 skb, fragstolen)) ? 1 : 0; 5169 tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq); 5170 if (!eaten) { 5171 tcp_add_receive_queue(sk, skb); 5172 skb_set_owner_r(skb, sk); 5173 } 5174 return eaten; 5175 } 5176 5177 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size) 5178 { 5179 struct sk_buff *skb; 5180 int err = -ENOMEM; 5181 int data_len = 0; 5182 bool fragstolen; 5183 5184 if (size == 0) 5185 return 0; 5186 5187 if (size > PAGE_SIZE) { 5188 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS); 5189 5190 data_len = npages << PAGE_SHIFT; 5191 size = data_len + (size & ~PAGE_MASK); 5192 } 5193 skb = alloc_skb_with_frags(size - data_len, data_len, 5194 PAGE_ALLOC_COSTLY_ORDER, 5195 &err, sk->sk_allocation); 5196 if (!skb) 5197 goto err; 5198 5199 skb_put(skb, size - data_len); 5200 skb->data_len = data_len; 5201 skb->len = size; 5202 5203 if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) { 5204 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP); 5205 goto err_free; 5206 } 5207 5208 err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size); 5209 if (err) 5210 goto err_free; 5211 5212 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt; 5213 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size; 5214 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1; 5215 5216 if (tcp_queue_rcv(sk, skb, &fragstolen)) { 5217 WARN_ON_ONCE(fragstolen); /* should not happen */ 5218 __kfree_skb(skb); 5219 } 5220 return size; 5221 5222 err_free: 5223 kfree_skb(skb); 5224 err: 5225 return err; 5226 5227 } 5228 5229 void tcp_data_ready(struct sock *sk) 5230 { 5231 if (tcp_epollin_ready(sk, sk->sk_rcvlowat) || sock_flag(sk, SOCK_DONE)) 5232 sk->sk_data_ready(sk); 5233 } 5234 5235 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb) 5236 { 5237 struct tcp_sock *tp = tcp_sk(sk); 5238 enum skb_drop_reason reason; 5239 bool fragstolen; 5240 int eaten; 5241 5242 /* If a subflow has been reset, the packet should not continue 5243 * to be processed, drop the packet. 5244 */ 5245 if (sk_is_mptcp(sk) && !mptcp_incoming_options(sk, skb)) { 5246 __kfree_skb(skb); 5247 return; 5248 } 5249 5250 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) { 5251 __kfree_skb(skb); 5252 return; 5253 } 5254 tcp_cleanup_skb(skb); 5255 __skb_pull(skb, tcp_hdr(skb)->doff * 4); 5256 5257 reason = SKB_DROP_REASON_NOT_SPECIFIED; 5258 tp->rx_opt.dsack = 0; 5259 5260 /* Queue data for delivery to the user. 5261 * Packets in sequence go to the receive queue. 5262 * Out of sequence packets to the out_of_order_queue. 5263 */ 5264 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) { 5265 if (tcp_receive_window(tp) == 0) { 5266 /* Some stacks are known to send bare FIN packets 5267 * in a loop even if we send RWIN 0 in our ACK. 5268 * Accepting this FIN does not hurt memory pressure 5269 * because the FIN flag will simply be merged to the 5270 * receive queue tail skb in most cases. 5271 */ 5272 if (!skb->len && 5273 (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) 5274 goto queue_and_out; 5275 5276 reason = SKB_DROP_REASON_TCP_ZEROWINDOW; 5277 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP); 5278 goto out_of_window; 5279 } 5280 5281 /* Ok. In sequence. In window. */ 5282 queue_and_out: 5283 if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) { 5284 /* TODO: maybe ratelimit these WIN 0 ACK ? */ 5285 inet_csk(sk)->icsk_ack.pending |= 5286 (ICSK_ACK_NOMEM | ICSK_ACK_NOW); 5287 inet_csk_schedule_ack(sk); 5288 sk->sk_data_ready(sk); 5289 5290 if (skb_queue_len(&sk->sk_receive_queue) && skb->len) { 5291 reason = SKB_DROP_REASON_PROTO_MEM; 5292 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP); 5293 goto drop; 5294 } 5295 sk_forced_mem_schedule(sk, skb->truesize); 5296 } 5297 5298 eaten = tcp_queue_rcv(sk, skb, &fragstolen); 5299 if (skb->len) 5300 tcp_event_data_recv(sk, skb); 5301 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 5302 tcp_fin(sk); 5303 5304 if (!RB_EMPTY_ROOT(&tp->out_of_order_queue)) { 5305 tcp_ofo_queue(sk); 5306 5307 /* RFC5681. 4.2. SHOULD send immediate ACK, when 5308 * gap in queue is filled. 5309 */ 5310 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) 5311 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW; 5312 } 5313 5314 if (tp->rx_opt.num_sacks) 5315 tcp_sack_remove(tp); 5316 5317 tcp_fast_path_check(sk); 5318 5319 if (eaten > 0) 5320 kfree_skb_partial(skb, fragstolen); 5321 if (!sock_flag(sk, SOCK_DEAD)) 5322 tcp_data_ready(sk); 5323 return; 5324 } 5325 5326 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) { 5327 tcp_rcv_spurious_retrans(sk, skb); 5328 /* A retransmit, 2nd most common case. Force an immediate ack. */ 5329 reason = SKB_DROP_REASON_TCP_OLD_DATA; 5330 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST); 5331 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); 5332 5333 out_of_window: 5334 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS); 5335 inet_csk_schedule_ack(sk); 5336 drop: 5337 tcp_drop_reason(sk, skb, reason); 5338 return; 5339 } 5340 5341 /* Out of window. F.e. zero window probe. */ 5342 if (!before(TCP_SKB_CB(skb)->seq, 5343 tp->rcv_nxt + tcp_receive_window(tp))) { 5344 reason = SKB_DROP_REASON_TCP_OVERWINDOW; 5345 goto out_of_window; 5346 } 5347 5348 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 5349 /* Partial packet, seq < rcv_next < end_seq */ 5350 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt); 5351 5352 /* If window is closed, drop tail of packet. But after 5353 * remembering D-SACK for its head made in previous line. 5354 */ 5355 if (!tcp_receive_window(tp)) { 5356 reason = SKB_DROP_REASON_TCP_ZEROWINDOW; 5357 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP); 5358 goto out_of_window; 5359 } 5360 goto queue_and_out; 5361 } 5362 5363 tcp_data_queue_ofo(sk, skb); 5364 } 5365 5366 static struct sk_buff *tcp_skb_next(struct sk_buff *skb, struct sk_buff_head *list) 5367 { 5368 if (list) 5369 return !skb_queue_is_last(list, skb) ? skb->next : NULL; 5370 5371 return skb_rb_next(skb); 5372 } 5373 5374 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb, 5375 struct sk_buff_head *list, 5376 struct rb_root *root) 5377 { 5378 struct sk_buff *next = tcp_skb_next(skb, list); 5379 5380 if (list) 5381 __skb_unlink(skb, list); 5382 else 5383 rb_erase(&skb->rbnode, root); 5384 5385 __kfree_skb(skb); 5386 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED); 5387 5388 return next; 5389 } 5390 5391 /* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */ 5392 void tcp_rbtree_insert(struct rb_root *root, struct sk_buff *skb) 5393 { 5394 struct rb_node **p = &root->rb_node; 5395 struct rb_node *parent = NULL; 5396 struct sk_buff *skb1; 5397 5398 while (*p) { 5399 parent = *p; 5400 skb1 = rb_to_skb(parent); 5401 if (before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb1)->seq)) 5402 p = &parent->rb_left; 5403 else 5404 p = &parent->rb_right; 5405 } 5406 rb_link_node(&skb->rbnode, parent, p); 5407 rb_insert_color(&skb->rbnode, root); 5408 } 5409 5410 /* Collapse contiguous sequence of skbs head..tail with 5411 * sequence numbers start..end. 5412 * 5413 * If tail is NULL, this means until the end of the queue. 5414 * 5415 * Segments with FIN/SYN are not collapsed (only because this 5416 * simplifies code) 5417 */ 5418 static void 5419 tcp_collapse(struct sock *sk, struct sk_buff_head *list, struct rb_root *root, 5420 struct sk_buff *head, struct sk_buff *tail, u32 start, u32 end) 5421 { 5422 struct sk_buff *skb = head, *n; 5423 struct sk_buff_head tmp; 5424 bool end_of_skbs; 5425 5426 /* First, check that queue is collapsible and find 5427 * the point where collapsing can be useful. 5428 */ 5429 restart: 5430 for (end_of_skbs = true; skb != NULL && skb != tail; skb = n) { 5431 n = tcp_skb_next(skb, list); 5432 5433 if (!skb_frags_readable(skb)) 5434 goto skip_this; 5435 5436 /* No new bits? It is possible on ofo queue. */ 5437 if (!before(start, TCP_SKB_CB(skb)->end_seq)) { 5438 skb = tcp_collapse_one(sk, skb, list, root); 5439 if (!skb) 5440 break; 5441 goto restart; 5442 } 5443 5444 /* The first skb to collapse is: 5445 * - not SYN/FIN and 5446 * - bloated or contains data before "start" or 5447 * overlaps to the next one and mptcp allow collapsing. 5448 */ 5449 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) && 5450 (tcp_win_from_space(sk, skb->truesize) > skb->len || 5451 before(TCP_SKB_CB(skb)->seq, start))) { 5452 end_of_skbs = false; 5453 break; 5454 } 5455 5456 if (n && n != tail && skb_frags_readable(n) && 5457 tcp_skb_can_collapse_rx(skb, n) && 5458 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(n)->seq) { 5459 end_of_skbs = false; 5460 break; 5461 } 5462 5463 skip_this: 5464 /* Decided to skip this, advance start seq. */ 5465 start = TCP_SKB_CB(skb)->end_seq; 5466 } 5467 if (end_of_skbs || 5468 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) || 5469 !skb_frags_readable(skb)) 5470 return; 5471 5472 __skb_queue_head_init(&tmp); 5473 5474 while (before(start, end)) { 5475 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start); 5476 struct sk_buff *nskb; 5477 5478 nskb = alloc_skb(copy, GFP_ATOMIC); 5479 if (!nskb) 5480 break; 5481 5482 memcpy(nskb->cb, skb->cb, sizeof(skb->cb)); 5483 skb_copy_decrypted(nskb, skb); 5484 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start; 5485 if (list) 5486 __skb_queue_before(list, skb, nskb); 5487 else 5488 __skb_queue_tail(&tmp, nskb); /* defer rbtree insertion */ 5489 skb_set_owner_r(nskb, sk); 5490 mptcp_skb_ext_move(nskb, skb); 5491 5492 /* Copy data, releasing collapsed skbs. */ 5493 while (copy > 0) { 5494 int offset = start - TCP_SKB_CB(skb)->seq; 5495 int size = TCP_SKB_CB(skb)->end_seq - start; 5496 5497 BUG_ON(offset < 0); 5498 if (size > 0) { 5499 size = min(copy, size); 5500 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size)) 5501 BUG(); 5502 TCP_SKB_CB(nskb)->end_seq += size; 5503 copy -= size; 5504 start += size; 5505 } 5506 if (!before(start, TCP_SKB_CB(skb)->end_seq)) { 5507 skb = tcp_collapse_one(sk, skb, list, root); 5508 if (!skb || 5509 skb == tail || 5510 !tcp_skb_can_collapse_rx(nskb, skb) || 5511 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) || 5512 !skb_frags_readable(skb)) 5513 goto end; 5514 } 5515 } 5516 } 5517 end: 5518 skb_queue_walk_safe(&tmp, skb, n) 5519 tcp_rbtree_insert(root, skb); 5520 } 5521 5522 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs 5523 * and tcp_collapse() them until all the queue is collapsed. 5524 */ 5525 static void tcp_collapse_ofo_queue(struct sock *sk) 5526 { 5527 struct tcp_sock *tp = tcp_sk(sk); 5528 u32 range_truesize, sum_tiny = 0; 5529 struct sk_buff *skb, *head; 5530 u32 start, end; 5531 5532 skb = skb_rb_first(&tp->out_of_order_queue); 5533 new_range: 5534 if (!skb) { 5535 tp->ooo_last_skb = skb_rb_last(&tp->out_of_order_queue); 5536 return; 5537 } 5538 start = TCP_SKB_CB(skb)->seq; 5539 end = TCP_SKB_CB(skb)->end_seq; 5540 range_truesize = skb->truesize; 5541 5542 for (head = skb;;) { 5543 skb = skb_rb_next(skb); 5544 5545 /* Range is terminated when we see a gap or when 5546 * we are at the queue end. 5547 */ 5548 if (!skb || 5549 after(TCP_SKB_CB(skb)->seq, end) || 5550 before(TCP_SKB_CB(skb)->end_seq, start)) { 5551 /* Do not attempt collapsing tiny skbs */ 5552 if (range_truesize != head->truesize || 5553 end - start >= SKB_WITH_OVERHEAD(PAGE_SIZE)) { 5554 tcp_collapse(sk, NULL, &tp->out_of_order_queue, 5555 head, skb, start, end); 5556 } else { 5557 sum_tiny += range_truesize; 5558 if (sum_tiny > sk->sk_rcvbuf >> 3) 5559 return; 5560 } 5561 goto new_range; 5562 } 5563 5564 range_truesize += skb->truesize; 5565 if (unlikely(before(TCP_SKB_CB(skb)->seq, start))) 5566 start = TCP_SKB_CB(skb)->seq; 5567 if (after(TCP_SKB_CB(skb)->end_seq, end)) 5568 end = TCP_SKB_CB(skb)->end_seq; 5569 } 5570 } 5571 5572 /* 5573 * Clean the out-of-order queue to make room. 5574 * We drop high sequences packets to : 5575 * 1) Let a chance for holes to be filled. 5576 * This means we do not drop packets from ooo queue if their sequence 5577 * is before incoming packet sequence. 5578 * 2) not add too big latencies if thousands of packets sit there. 5579 * (But if application shrinks SO_RCVBUF, we could still end up 5580 * freeing whole queue here) 5581 * 3) Drop at least 12.5 % of sk_rcvbuf to avoid malicious attacks. 5582 * 5583 * Return true if queue has shrunk. 5584 */ 5585 static bool tcp_prune_ofo_queue(struct sock *sk, const struct sk_buff *in_skb) 5586 { 5587 struct tcp_sock *tp = tcp_sk(sk); 5588 struct rb_node *node, *prev; 5589 bool pruned = false; 5590 int goal; 5591 5592 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) 5593 return false; 5594 5595 goal = sk->sk_rcvbuf >> 3; 5596 node = &tp->ooo_last_skb->rbnode; 5597 5598 do { 5599 struct sk_buff *skb = rb_to_skb(node); 5600 5601 /* If incoming skb would land last in ofo queue, stop pruning. */ 5602 if (after(TCP_SKB_CB(in_skb)->seq, TCP_SKB_CB(skb)->seq)) 5603 break; 5604 pruned = true; 5605 prev = rb_prev(node); 5606 rb_erase(node, &tp->out_of_order_queue); 5607 goal -= skb->truesize; 5608 tcp_drop_reason(sk, skb, SKB_DROP_REASON_TCP_OFO_QUEUE_PRUNE); 5609 tp->ooo_last_skb = rb_to_skb(prev); 5610 if (!prev || goal <= 0) { 5611 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf && 5612 !tcp_under_memory_pressure(sk)) 5613 break; 5614 goal = sk->sk_rcvbuf >> 3; 5615 } 5616 node = prev; 5617 } while (node); 5618 5619 if (pruned) { 5620 NET_INC_STATS(sock_net(sk), LINUX_MIB_OFOPRUNED); 5621 /* Reset SACK state. A conforming SACK implementation will 5622 * do the same at a timeout based retransmit. When a connection 5623 * is in a sad state like this, we care only about integrity 5624 * of the connection not performance. 5625 */ 5626 if (tp->rx_opt.sack_ok) 5627 tcp_sack_reset(&tp->rx_opt); 5628 } 5629 return pruned; 5630 } 5631 5632 /* Reduce allocated memory if we can, trying to get 5633 * the socket within its memory limits again. 5634 * 5635 * Return less than zero if we should start dropping frames 5636 * until the socket owning process reads some of the data 5637 * to stabilize the situation. 5638 */ 5639 static int tcp_prune_queue(struct sock *sk, const struct sk_buff *in_skb) 5640 { 5641 struct tcp_sock *tp = tcp_sk(sk); 5642 5643 NET_INC_STATS(sock_net(sk), LINUX_MIB_PRUNECALLED); 5644 5645 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf) 5646 tcp_clamp_window(sk); 5647 else if (tcp_under_memory_pressure(sk)) 5648 tcp_adjust_rcv_ssthresh(sk); 5649 5650 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) 5651 return 0; 5652 5653 tcp_collapse_ofo_queue(sk); 5654 if (!skb_queue_empty(&sk->sk_receive_queue)) 5655 tcp_collapse(sk, &sk->sk_receive_queue, NULL, 5656 skb_peek(&sk->sk_receive_queue), 5657 NULL, 5658 tp->copied_seq, tp->rcv_nxt); 5659 5660 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) 5661 return 0; 5662 5663 /* Collapsing did not help, destructive actions follow. 5664 * This must not ever occur. */ 5665 5666 tcp_prune_ofo_queue(sk, in_skb); 5667 5668 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) 5669 return 0; 5670 5671 /* If we are really being abused, tell the caller to silently 5672 * drop receive data on the floor. It will get retransmitted 5673 * and hopefully then we'll have sufficient space. 5674 */ 5675 NET_INC_STATS(sock_net(sk), LINUX_MIB_RCVPRUNED); 5676 5677 /* Massive buffer overcommit. */ 5678 tp->pred_flags = 0; 5679 return -1; 5680 } 5681 5682 static bool tcp_should_expand_sndbuf(struct sock *sk) 5683 { 5684 const struct tcp_sock *tp = tcp_sk(sk); 5685 5686 /* If the user specified a specific send buffer setting, do 5687 * not modify it. 5688 */ 5689 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK) 5690 return false; 5691 5692 /* If we are under global TCP memory pressure, do not expand. */ 5693 if (tcp_under_memory_pressure(sk)) { 5694 int unused_mem = sk_unused_reserved_mem(sk); 5695 5696 /* Adjust sndbuf according to reserved mem. But make sure 5697 * it never goes below SOCK_MIN_SNDBUF. 5698 * See sk_stream_moderate_sndbuf() for more details. 5699 */ 5700 if (unused_mem > SOCK_MIN_SNDBUF) 5701 WRITE_ONCE(sk->sk_sndbuf, unused_mem); 5702 5703 return false; 5704 } 5705 5706 /* If we are under soft global TCP memory pressure, do not expand. */ 5707 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0)) 5708 return false; 5709 5710 /* If we filled the congestion window, do not expand. */ 5711 if (tcp_packets_in_flight(tp) >= tcp_snd_cwnd(tp)) 5712 return false; 5713 5714 return true; 5715 } 5716 5717 static void tcp_new_space(struct sock *sk) 5718 { 5719 struct tcp_sock *tp = tcp_sk(sk); 5720 5721 if (tcp_should_expand_sndbuf(sk)) { 5722 tcp_sndbuf_expand(sk); 5723 tp->snd_cwnd_stamp = tcp_jiffies32; 5724 } 5725 5726 INDIRECT_CALL_1(sk->sk_write_space, sk_stream_write_space, sk); 5727 } 5728 5729 /* Caller made space either from: 5730 * 1) Freeing skbs in rtx queues (after tp->snd_una has advanced) 5731 * 2) Sent skbs from output queue (and thus advancing tp->snd_nxt) 5732 * 5733 * We might be able to generate EPOLLOUT to the application if: 5734 * 1) Space consumed in output/rtx queues is below sk->sk_sndbuf/2 5735 * 2) notsent amount (tp->write_seq - tp->snd_nxt) became 5736 * small enough that tcp_stream_memory_free() decides it 5737 * is time to generate EPOLLOUT. 5738 */ 5739 void tcp_check_space(struct sock *sk) 5740 { 5741 /* pairs with tcp_poll() */ 5742 smp_mb(); 5743 if (sk->sk_socket && 5744 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { 5745 tcp_new_space(sk); 5746 if (!test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) 5747 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED); 5748 } 5749 } 5750 5751 static inline void tcp_data_snd_check(struct sock *sk) 5752 { 5753 tcp_push_pending_frames(sk); 5754 tcp_check_space(sk); 5755 } 5756 5757 /* 5758 * Check if sending an ack is needed. 5759 */ 5760 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible) 5761 { 5762 struct tcp_sock *tp = tcp_sk(sk); 5763 unsigned long rtt, delay; 5764 5765 /* More than one full frame received... */ 5766 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss && 5767 /* ... and right edge of window advances far enough. 5768 * (tcp_recvmsg() will send ACK otherwise). 5769 * If application uses SO_RCVLOWAT, we want send ack now if 5770 * we have not received enough bytes to satisfy the condition. 5771 */ 5772 (tp->rcv_nxt - tp->copied_seq < sk->sk_rcvlowat || 5773 __tcp_select_window(sk) >= tp->rcv_wnd)) || 5774 /* We ACK each frame or... */ 5775 tcp_in_quickack_mode(sk) || 5776 /* Protocol state mandates a one-time immediate ACK */ 5777 inet_csk(sk)->icsk_ack.pending & ICSK_ACK_NOW) { 5778 /* If we are running from __release_sock() in user context, 5779 * Defer the ack until tcp_release_cb(). 5780 */ 5781 if (sock_owned_by_user_nocheck(sk) && 5782 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_backlog_ack_defer)) { 5783 set_bit(TCP_ACK_DEFERRED, &sk->sk_tsq_flags); 5784 return; 5785 } 5786 send_now: 5787 tcp_send_ack(sk); 5788 return; 5789 } 5790 5791 if (!ofo_possible || RB_EMPTY_ROOT(&tp->out_of_order_queue)) { 5792 tcp_send_delayed_ack(sk); 5793 return; 5794 } 5795 5796 if (!tcp_is_sack(tp) || 5797 tp->compressed_ack >= READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_nr)) 5798 goto send_now; 5799 5800 if (tp->compressed_ack_rcv_nxt != tp->rcv_nxt) { 5801 tp->compressed_ack_rcv_nxt = tp->rcv_nxt; 5802 tp->dup_ack_counter = 0; 5803 } 5804 if (tp->dup_ack_counter < TCP_FASTRETRANS_THRESH) { 5805 tp->dup_ack_counter++; 5806 goto send_now; 5807 } 5808 tp->compressed_ack++; 5809 if (hrtimer_is_queued(&tp->compressed_ack_timer)) 5810 return; 5811 5812 /* compress ack timer : 5 % of rtt, but no more than tcp_comp_sack_delay_ns */ 5813 5814 rtt = tp->rcv_rtt_est.rtt_us; 5815 if (tp->srtt_us && tp->srtt_us < rtt) 5816 rtt = tp->srtt_us; 5817 5818 delay = min_t(unsigned long, 5819 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_delay_ns), 5820 rtt * (NSEC_PER_USEC >> 3)/20); 5821 sock_hold(sk); 5822 hrtimer_start_range_ns(&tp->compressed_ack_timer, ns_to_ktime(delay), 5823 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_slack_ns), 5824 HRTIMER_MODE_REL_PINNED_SOFT); 5825 } 5826 5827 static inline void tcp_ack_snd_check(struct sock *sk) 5828 { 5829 if (!inet_csk_ack_scheduled(sk)) { 5830 /* We sent a data segment already. */ 5831 return; 5832 } 5833 __tcp_ack_snd_check(sk, 1); 5834 } 5835 5836 /* 5837 * This routine is only called when we have urgent data 5838 * signaled. Its the 'slow' part of tcp_urg. It could be 5839 * moved inline now as tcp_urg is only called from one 5840 * place. We handle URGent data wrong. We have to - as 5841 * BSD still doesn't use the correction from RFC961. 5842 * For 1003.1g we should support a new option TCP_STDURG to permit 5843 * either form (or just set the sysctl tcp_stdurg). 5844 */ 5845 5846 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th) 5847 { 5848 struct tcp_sock *tp = tcp_sk(sk); 5849 u32 ptr = ntohs(th->urg_ptr); 5850 5851 if (ptr && !READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_stdurg)) 5852 ptr--; 5853 ptr += ntohl(th->seq); 5854 5855 /* Ignore urgent data that we've already seen and read. */ 5856 if (after(tp->copied_seq, ptr)) 5857 return; 5858 5859 /* Do not replay urg ptr. 5860 * 5861 * NOTE: interesting situation not covered by specs. 5862 * Misbehaving sender may send urg ptr, pointing to segment, 5863 * which we already have in ofo queue. We are not able to fetch 5864 * such data and will stay in TCP_URG_NOTYET until will be eaten 5865 * by recvmsg(). Seems, we are not obliged to handle such wicked 5866 * situations. But it is worth to think about possibility of some 5867 * DoSes using some hypothetical application level deadlock. 5868 */ 5869 if (before(ptr, tp->rcv_nxt)) 5870 return; 5871 5872 /* Do we already have a newer (or duplicate) urgent pointer? */ 5873 if (tp->urg_data && !after(ptr, tp->urg_seq)) 5874 return; 5875 5876 /* Tell the world about our new urgent pointer. */ 5877 sk_send_sigurg(sk); 5878 5879 /* We may be adding urgent data when the last byte read was 5880 * urgent. To do this requires some care. We cannot just ignore 5881 * tp->copied_seq since we would read the last urgent byte again 5882 * as data, nor can we alter copied_seq until this data arrives 5883 * or we break the semantics of SIOCATMARK (and thus sockatmark()) 5884 * 5885 * NOTE. Double Dutch. Rendering to plain English: author of comment 5886 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB); 5887 * and expect that both A and B disappear from stream. This is _wrong_. 5888 * Though this happens in BSD with high probability, this is occasional. 5889 * Any application relying on this is buggy. Note also, that fix "works" 5890 * only in this artificial test. Insert some normal data between A and B and we will 5891 * decline of BSD again. Verdict: it is better to remove to trap 5892 * buggy users. 5893 */ 5894 if (tp->urg_seq == tp->copied_seq && tp->urg_data && 5895 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) { 5896 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue); 5897 tp->copied_seq++; 5898 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) { 5899 __skb_unlink(skb, &sk->sk_receive_queue); 5900 __kfree_skb(skb); 5901 } 5902 } 5903 5904 WRITE_ONCE(tp->urg_data, TCP_URG_NOTYET); 5905 WRITE_ONCE(tp->urg_seq, ptr); 5906 5907 /* Disable header prediction. */ 5908 tp->pred_flags = 0; 5909 } 5910 5911 /* This is the 'fast' part of urgent handling. */ 5912 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th) 5913 { 5914 struct tcp_sock *tp = tcp_sk(sk); 5915 5916 /* Check if we get a new urgent pointer - normally not. */ 5917 if (unlikely(th->urg)) 5918 tcp_check_urg(sk, th); 5919 5920 /* Do we wait for any urgent data? - normally not... */ 5921 if (unlikely(tp->urg_data == TCP_URG_NOTYET)) { 5922 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) - 5923 th->syn; 5924 5925 /* Is the urgent pointer pointing into this packet? */ 5926 if (ptr < skb->len) { 5927 u8 tmp; 5928 if (skb_copy_bits(skb, ptr, &tmp, 1)) 5929 BUG(); 5930 WRITE_ONCE(tp->urg_data, TCP_URG_VALID | tmp); 5931 if (!sock_flag(sk, SOCK_DEAD)) 5932 sk->sk_data_ready(sk); 5933 } 5934 } 5935 } 5936 5937 /* Accept RST for rcv_nxt - 1 after a FIN. 5938 * When tcp connections are abruptly terminated from Mac OSX (via ^C), a 5939 * FIN is sent followed by a RST packet. The RST is sent with the same 5940 * sequence number as the FIN, and thus according to RFC 5961 a challenge 5941 * ACK should be sent. However, Mac OSX rate limits replies to challenge 5942 * ACKs on the closed socket. In addition middleboxes can drop either the 5943 * challenge ACK or a subsequent RST. 5944 */ 5945 static bool tcp_reset_check(const struct sock *sk, const struct sk_buff *skb) 5946 { 5947 const struct tcp_sock *tp = tcp_sk(sk); 5948 5949 return unlikely(TCP_SKB_CB(skb)->seq == (tp->rcv_nxt - 1) && 5950 (1 << sk->sk_state) & (TCPF_CLOSE_WAIT | TCPF_LAST_ACK | 5951 TCPF_CLOSING)); 5952 } 5953 5954 /* Does PAWS and seqno based validation of an incoming segment, flags will 5955 * play significant role here. 5956 */ 5957 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb, 5958 const struct tcphdr *th, int syn_inerr) 5959 { 5960 struct tcp_sock *tp = tcp_sk(sk); 5961 SKB_DR(reason); 5962 5963 /* RFC1323: H1. Apply PAWS check first. */ 5964 if (!tcp_fast_parse_options(sock_net(sk), skb, th, tp) || 5965 !tp->rx_opt.saw_tstamp || 5966 tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW)) 5967 goto step1; 5968 5969 reason = tcp_disordered_ack_check(sk, skb); 5970 if (!reason) 5971 goto step1; 5972 /* Reset is accepted even if it did not pass PAWS. */ 5973 if (th->rst) 5974 goto step1; 5975 if (unlikely(th->syn)) 5976 goto syn_challenge; 5977 5978 /* Old ACK are common, increment PAWS_OLD_ACK 5979 * and do not send a dupack. 5980 */ 5981 if (reason == SKB_DROP_REASON_TCP_RFC7323_PAWS_ACK) { 5982 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWS_OLD_ACK); 5983 goto discard; 5984 } 5985 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED); 5986 if (!tcp_oow_rate_limited(sock_net(sk), skb, 5987 LINUX_MIB_TCPACKSKIPPEDPAWS, 5988 &tp->last_oow_ack_time)) 5989 tcp_send_dupack(sk, skb); 5990 goto discard; 5991 5992 step1: 5993 /* Step 1: check sequence number */ 5994 reason = tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); 5995 if (reason) { 5996 /* RFC793, page 37: "In all states except SYN-SENT, all reset 5997 * (RST) segments are validated by checking their SEQ-fields." 5998 * And page 69: "If an incoming segment is not acceptable, 5999 * an acknowledgment should be sent in reply (unless the RST 6000 * bit is set, if so drop the segment and return)". 6001 */ 6002 if (!th->rst) { 6003 if (th->syn) 6004 goto syn_challenge; 6005 if (!tcp_oow_rate_limited(sock_net(sk), skb, 6006 LINUX_MIB_TCPACKSKIPPEDSEQ, 6007 &tp->last_oow_ack_time)) 6008 tcp_send_dupack(sk, skb); 6009 } else if (tcp_reset_check(sk, skb)) { 6010 goto reset; 6011 } 6012 goto discard; 6013 } 6014 6015 /* Step 2: check RST bit */ 6016 if (th->rst) { 6017 /* RFC 5961 3.2 (extend to match against (RCV.NXT - 1) after a 6018 * FIN and SACK too if available): 6019 * If seq num matches RCV.NXT or (RCV.NXT - 1) after a FIN, or 6020 * the right-most SACK block, 6021 * then 6022 * RESET the connection 6023 * else 6024 * Send a challenge ACK 6025 */ 6026 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt || 6027 tcp_reset_check(sk, skb)) 6028 goto reset; 6029 6030 if (tcp_is_sack(tp) && tp->rx_opt.num_sacks > 0) { 6031 struct tcp_sack_block *sp = &tp->selective_acks[0]; 6032 int max_sack = sp[0].end_seq; 6033 int this_sack; 6034 6035 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks; 6036 ++this_sack) { 6037 max_sack = after(sp[this_sack].end_seq, 6038 max_sack) ? 6039 sp[this_sack].end_seq : max_sack; 6040 } 6041 6042 if (TCP_SKB_CB(skb)->seq == max_sack) 6043 goto reset; 6044 } 6045 6046 /* Disable TFO if RST is out-of-order 6047 * and no data has been received 6048 * for current active TFO socket 6049 */ 6050 if (tp->syn_fastopen && !tp->data_segs_in && 6051 sk->sk_state == TCP_ESTABLISHED) 6052 tcp_fastopen_active_disable(sk); 6053 tcp_send_challenge_ack(sk); 6054 SKB_DR_SET(reason, TCP_RESET); 6055 goto discard; 6056 } 6057 6058 /* step 3: check security and precedence [ignored] */ 6059 6060 /* step 4: Check for a SYN 6061 * RFC 5961 4.2 : Send a challenge ack 6062 */ 6063 if (th->syn) { 6064 if (sk->sk_state == TCP_SYN_RECV && sk->sk_socket && th->ack && 6065 TCP_SKB_CB(skb)->seq + 1 == TCP_SKB_CB(skb)->end_seq && 6066 TCP_SKB_CB(skb)->seq + 1 == tp->rcv_nxt && 6067 TCP_SKB_CB(skb)->ack_seq == tp->snd_nxt) 6068 goto pass; 6069 syn_challenge: 6070 if (syn_inerr) 6071 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS); 6072 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE); 6073 tcp_send_challenge_ack(sk); 6074 SKB_DR_SET(reason, TCP_INVALID_SYN); 6075 goto discard; 6076 } 6077 6078 pass: 6079 bpf_skops_parse_hdr(sk, skb); 6080 6081 return true; 6082 6083 discard: 6084 tcp_drop_reason(sk, skb, reason); 6085 return false; 6086 6087 reset: 6088 tcp_reset(sk, skb); 6089 __kfree_skb(skb); 6090 return false; 6091 } 6092 6093 /* 6094 * TCP receive function for the ESTABLISHED state. 6095 * 6096 * It is split into a fast path and a slow path. The fast path is 6097 * disabled when: 6098 * - A zero window was announced from us - zero window probing 6099 * is only handled properly in the slow path. 6100 * - Out of order segments arrived. 6101 * - Urgent data is expected. 6102 * - There is no buffer space left 6103 * - Unexpected TCP flags/window values/header lengths are received 6104 * (detected by checking the TCP header against pred_flags) 6105 * - Data is sent in both directions. Fast path only supports pure senders 6106 * or pure receivers (this means either the sequence number or the ack 6107 * value must stay constant) 6108 * - Unexpected TCP option. 6109 * 6110 * When these conditions are not satisfied it drops into a standard 6111 * receive procedure patterned after RFC793 to handle all cases. 6112 * The first three cases are guaranteed by proper pred_flags setting, 6113 * the rest is checked inline. Fast processing is turned on in 6114 * tcp_data_queue when everything is OK. 6115 */ 6116 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb) 6117 { 6118 enum skb_drop_reason reason = SKB_DROP_REASON_NOT_SPECIFIED; 6119 const struct tcphdr *th = (const struct tcphdr *)skb->data; 6120 struct tcp_sock *tp = tcp_sk(sk); 6121 unsigned int len = skb->len; 6122 6123 /* TCP congestion window tracking */ 6124 trace_tcp_probe(sk, skb); 6125 6126 tcp_mstamp_refresh(tp); 6127 if (unlikely(!rcu_access_pointer(sk->sk_rx_dst))) 6128 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb); 6129 /* 6130 * Header prediction. 6131 * The code loosely follows the one in the famous 6132 * "30 instruction TCP receive" Van Jacobson mail. 6133 * 6134 * Van's trick is to deposit buffers into socket queue 6135 * on a device interrupt, to call tcp_recv function 6136 * on the receive process context and checksum and copy 6137 * the buffer to user space. smart... 6138 * 6139 * Our current scheme is not silly either but we take the 6140 * extra cost of the net_bh soft interrupt processing... 6141 * We do checksum and copy also but from device to kernel. 6142 */ 6143 6144 tp->rx_opt.saw_tstamp = 0; 6145 6146 /* pred_flags is 0xS?10 << 16 + snd_wnd 6147 * if header_prediction is to be made 6148 * 'S' will always be tp->tcp_header_len >> 2 6149 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to 6150 * turn it off (when there are holes in the receive 6151 * space for instance) 6152 * PSH flag is ignored. 6153 */ 6154 6155 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags && 6156 TCP_SKB_CB(skb)->seq == tp->rcv_nxt && 6157 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) { 6158 int tcp_header_len = tp->tcp_header_len; 6159 6160 /* Timestamp header prediction: tcp_header_len 6161 * is automatically equal to th->doff*4 due to pred_flags 6162 * match. 6163 */ 6164 6165 /* Check timestamp */ 6166 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) { 6167 /* No? Slow path! */ 6168 if (!tcp_parse_aligned_timestamp(tp, th)) 6169 goto slow_path; 6170 6171 /* If PAWS failed, check it more carefully in slow path */ 6172 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0) 6173 goto slow_path; 6174 6175 /* DO NOT update ts_recent here, if checksum fails 6176 * and timestamp was corrupted part, it will result 6177 * in a hung connection since we will drop all 6178 * future packets due to the PAWS test. 6179 */ 6180 } 6181 6182 if (len <= tcp_header_len) { 6183 /* Bulk data transfer: sender */ 6184 if (len == tcp_header_len) { 6185 /* Predicted packet is in window by definition. 6186 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 6187 * Hence, check seq<=rcv_wup reduces to: 6188 */ 6189 if (tcp_header_len == 6190 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && 6191 tp->rcv_nxt == tp->rcv_wup) 6192 tcp_store_ts_recent(tp); 6193 6194 /* We know that such packets are checksummed 6195 * on entry. 6196 */ 6197 tcp_ack(sk, skb, 0); 6198 __kfree_skb(skb); 6199 tcp_data_snd_check(sk); 6200 /* When receiving pure ack in fast path, update 6201 * last ts ecr directly instead of calling 6202 * tcp_rcv_rtt_measure_ts() 6203 */ 6204 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr; 6205 return; 6206 } else { /* Header too small */ 6207 reason = SKB_DROP_REASON_PKT_TOO_SMALL; 6208 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS); 6209 goto discard; 6210 } 6211 } else { 6212 int eaten = 0; 6213 bool fragstolen = false; 6214 6215 if (tcp_checksum_complete(skb)) 6216 goto csum_error; 6217 6218 if ((int)skb->truesize > sk->sk_forward_alloc) 6219 goto step5; 6220 6221 /* Predicted packet is in window by definition. 6222 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 6223 * Hence, check seq<=rcv_wup reduces to: 6224 */ 6225 if (tcp_header_len == 6226 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && 6227 tp->rcv_nxt == tp->rcv_wup) 6228 tcp_store_ts_recent(tp); 6229 6230 tcp_rcv_rtt_measure_ts(sk, skb); 6231 6232 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPHITS); 6233 6234 /* Bulk data transfer: receiver */ 6235 tcp_cleanup_skb(skb); 6236 __skb_pull(skb, tcp_header_len); 6237 eaten = tcp_queue_rcv(sk, skb, &fragstolen); 6238 6239 tcp_event_data_recv(sk, skb); 6240 6241 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) { 6242 /* Well, only one small jumplet in fast path... */ 6243 tcp_ack(sk, skb, FLAG_DATA); 6244 tcp_data_snd_check(sk); 6245 if (!inet_csk_ack_scheduled(sk)) 6246 goto no_ack; 6247 } else { 6248 tcp_update_wl(tp, TCP_SKB_CB(skb)->seq); 6249 } 6250 6251 __tcp_ack_snd_check(sk, 0); 6252 no_ack: 6253 if (eaten) 6254 kfree_skb_partial(skb, fragstolen); 6255 tcp_data_ready(sk); 6256 return; 6257 } 6258 } 6259 6260 slow_path: 6261 if (len < (th->doff << 2) || tcp_checksum_complete(skb)) 6262 goto csum_error; 6263 6264 if (!th->ack && !th->rst && !th->syn) { 6265 reason = SKB_DROP_REASON_TCP_FLAGS; 6266 goto discard; 6267 } 6268 6269 /* 6270 * Standard slow path. 6271 */ 6272 6273 if (!tcp_validate_incoming(sk, skb, th, 1)) 6274 return; 6275 6276 step5: 6277 reason = tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT); 6278 if ((int)reason < 0) { 6279 reason = -reason; 6280 goto discard; 6281 } 6282 tcp_rcv_rtt_measure_ts(sk, skb); 6283 6284 /* Process urgent data. */ 6285 tcp_urg(sk, skb, th); 6286 6287 /* step 7: process the segment text */ 6288 tcp_data_queue(sk, skb); 6289 6290 tcp_data_snd_check(sk); 6291 tcp_ack_snd_check(sk); 6292 return; 6293 6294 csum_error: 6295 reason = SKB_DROP_REASON_TCP_CSUM; 6296 trace_tcp_bad_csum(skb); 6297 TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS); 6298 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS); 6299 6300 discard: 6301 tcp_drop_reason(sk, skb, reason); 6302 } 6303 EXPORT_SYMBOL(tcp_rcv_established); 6304 6305 void tcp_init_transfer(struct sock *sk, int bpf_op, struct sk_buff *skb) 6306 { 6307 struct inet_connection_sock *icsk = inet_csk(sk); 6308 struct tcp_sock *tp = tcp_sk(sk); 6309 6310 tcp_mtup_init(sk); 6311 icsk->icsk_af_ops->rebuild_header(sk); 6312 tcp_init_metrics(sk); 6313 6314 /* Initialize the congestion window to start the transfer. 6315 * Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been 6316 * retransmitted. In light of RFC6298 more aggressive 1sec 6317 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK 6318 * retransmission has occurred. 6319 */ 6320 if (tp->total_retrans > 1 && tp->undo_marker) 6321 tcp_snd_cwnd_set(tp, 1); 6322 else 6323 tcp_snd_cwnd_set(tp, tcp_init_cwnd(tp, __sk_dst_get(sk))); 6324 tp->snd_cwnd_stamp = tcp_jiffies32; 6325 6326 bpf_skops_established(sk, bpf_op, skb); 6327 /* Initialize congestion control unless BPF initialized it already: */ 6328 if (!icsk->icsk_ca_initialized) 6329 tcp_init_congestion_control(sk); 6330 tcp_init_buffer_space(sk); 6331 } 6332 6333 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb) 6334 { 6335 struct tcp_sock *tp = tcp_sk(sk); 6336 struct inet_connection_sock *icsk = inet_csk(sk); 6337 6338 tcp_ao_finish_connect(sk, skb); 6339 tcp_set_state(sk, TCP_ESTABLISHED); 6340 icsk->icsk_ack.lrcvtime = tcp_jiffies32; 6341 6342 if (skb) { 6343 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb); 6344 security_inet_conn_established(sk, skb); 6345 sk_mark_napi_id(sk, skb); 6346 } 6347 6348 tcp_init_transfer(sk, BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB, skb); 6349 6350 /* Prevent spurious tcp_cwnd_restart() on first data 6351 * packet. 6352 */ 6353 tp->lsndtime = tcp_jiffies32; 6354 6355 if (sock_flag(sk, SOCK_KEEPOPEN)) 6356 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp)); 6357 6358 if (!tp->rx_opt.snd_wscale) 6359 __tcp_fast_path_on(tp, tp->snd_wnd); 6360 else 6361 tp->pred_flags = 0; 6362 } 6363 6364 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack, 6365 struct tcp_fastopen_cookie *cookie) 6366 { 6367 struct tcp_sock *tp = tcp_sk(sk); 6368 struct sk_buff *data = tp->syn_data ? tcp_rtx_queue_head(sk) : NULL; 6369 u16 mss = tp->rx_opt.mss_clamp, try_exp = 0; 6370 bool syn_drop = false; 6371 6372 if (mss == tp->rx_opt.user_mss) { 6373 struct tcp_options_received opt; 6374 6375 /* Get original SYNACK MSS value if user MSS sets mss_clamp */ 6376 tcp_clear_options(&opt); 6377 opt.user_mss = opt.mss_clamp = 0; 6378 tcp_parse_options(sock_net(sk), synack, &opt, 0, NULL); 6379 mss = opt.mss_clamp; 6380 } 6381 6382 if (!tp->syn_fastopen) { 6383 /* Ignore an unsolicited cookie */ 6384 cookie->len = -1; 6385 } else if (tp->total_retrans) { 6386 /* SYN timed out and the SYN-ACK neither has a cookie nor 6387 * acknowledges data. Presumably the remote received only 6388 * the retransmitted (regular) SYNs: either the original 6389 * SYN-data or the corresponding SYN-ACK was dropped. 6390 */ 6391 syn_drop = (cookie->len < 0 && data); 6392 } else if (cookie->len < 0 && !tp->syn_data) { 6393 /* We requested a cookie but didn't get it. If we did not use 6394 * the (old) exp opt format then try so next time (try_exp=1). 6395 * Otherwise we go back to use the RFC7413 opt (try_exp=2). 6396 */ 6397 try_exp = tp->syn_fastopen_exp ? 2 : 1; 6398 } 6399 6400 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp); 6401 6402 if (data) { /* Retransmit unacked data in SYN */ 6403 if (tp->total_retrans) 6404 tp->fastopen_client_fail = TFO_SYN_RETRANSMITTED; 6405 else 6406 tp->fastopen_client_fail = TFO_DATA_NOT_ACKED; 6407 skb_rbtree_walk_from(data) 6408 tcp_mark_skb_lost(sk, data); 6409 tcp_non_congestion_loss_retransmit(sk); 6410 NET_INC_STATS(sock_net(sk), 6411 LINUX_MIB_TCPFASTOPENACTIVEFAIL); 6412 return true; 6413 } 6414 tp->syn_data_acked = tp->syn_data; 6415 if (tp->syn_data_acked) { 6416 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE); 6417 /* SYN-data is counted as two separate packets in tcp_ack() */ 6418 if (tp->delivered > 1) 6419 --tp->delivered; 6420 } 6421 6422 tcp_fastopen_add_skb(sk, synack); 6423 6424 return false; 6425 } 6426 6427 static void smc_check_reset_syn(struct tcp_sock *tp) 6428 { 6429 #if IS_ENABLED(CONFIG_SMC) 6430 if (static_branch_unlikely(&tcp_have_smc)) { 6431 if (tp->syn_smc && !tp->rx_opt.smc_ok) 6432 tp->syn_smc = 0; 6433 } 6434 #endif 6435 } 6436 6437 static void tcp_try_undo_spurious_syn(struct sock *sk) 6438 { 6439 struct tcp_sock *tp = tcp_sk(sk); 6440 u32 syn_stamp; 6441 6442 /* undo_marker is set when SYN or SYNACK times out. The timeout is 6443 * spurious if the ACK's timestamp option echo value matches the 6444 * original SYN timestamp. 6445 */ 6446 syn_stamp = tp->retrans_stamp; 6447 if (tp->undo_marker && syn_stamp && tp->rx_opt.saw_tstamp && 6448 syn_stamp == tp->rx_opt.rcv_tsecr) 6449 tp->undo_marker = 0; 6450 } 6451 6452 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb, 6453 const struct tcphdr *th) 6454 { 6455 struct inet_connection_sock *icsk = inet_csk(sk); 6456 struct tcp_sock *tp = tcp_sk(sk); 6457 struct tcp_fastopen_cookie foc = { .len = -1 }; 6458 int saved_clamp = tp->rx_opt.mss_clamp; 6459 bool fastopen_fail; 6460 SKB_DR(reason); 6461 6462 tcp_parse_options(sock_net(sk), skb, &tp->rx_opt, 0, &foc); 6463 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr) 6464 tp->rx_opt.rcv_tsecr -= tp->tsoffset; 6465 6466 if (th->ack) { 6467 /* rfc793: 6468 * "If the state is SYN-SENT then 6469 * first check the ACK bit 6470 * If the ACK bit is set 6471 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send 6472 * a reset (unless the RST bit is set, if so drop 6473 * the segment and return)" 6474 */ 6475 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) || 6476 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) { 6477 /* Previous FIN/ACK or RST/ACK might be ignored. */ 6478 if (icsk->icsk_retransmits == 0) 6479 inet_csk_reset_xmit_timer(sk, 6480 ICSK_TIME_RETRANS, 6481 TCP_TIMEOUT_MIN, TCP_RTO_MAX); 6482 SKB_DR_SET(reason, TCP_INVALID_ACK_SEQUENCE); 6483 goto reset_and_undo; 6484 } 6485 6486 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 6487 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp, 6488 tcp_time_stamp_ts(tp))) { 6489 NET_INC_STATS(sock_net(sk), 6490 LINUX_MIB_PAWSACTIVEREJECTED); 6491 SKB_DR_SET(reason, TCP_RFC7323_PAWS); 6492 goto reset_and_undo; 6493 } 6494 6495 /* Now ACK is acceptable. 6496 * 6497 * "If the RST bit is set 6498 * If the ACK was acceptable then signal the user "error: 6499 * connection reset", drop the segment, enter CLOSED state, 6500 * delete TCB, and return." 6501 */ 6502 6503 if (th->rst) { 6504 tcp_reset(sk, skb); 6505 consume: 6506 __kfree_skb(skb); 6507 return 0; 6508 } 6509 6510 /* rfc793: 6511 * "fifth, if neither of the SYN or RST bits is set then 6512 * drop the segment and return." 6513 * 6514 * See note below! 6515 * --ANK(990513) 6516 */ 6517 if (!th->syn) { 6518 SKB_DR_SET(reason, TCP_FLAGS); 6519 goto discard_and_undo; 6520 } 6521 /* rfc793: 6522 * "If the SYN bit is on ... 6523 * are acceptable then ... 6524 * (our SYN has been ACKed), change the connection 6525 * state to ESTABLISHED..." 6526 */ 6527 6528 tcp_ecn_rcv_synack(tp, th); 6529 6530 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq); 6531 tcp_try_undo_spurious_syn(sk); 6532 tcp_ack(sk, skb, FLAG_SLOWPATH); 6533 6534 /* Ok.. it's good. Set up sequence numbers and 6535 * move to established. 6536 */ 6537 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1); 6538 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; 6539 6540 /* RFC1323: The window in SYN & SYN/ACK segments is 6541 * never scaled. 6542 */ 6543 tp->snd_wnd = ntohs(th->window); 6544 6545 if (!tp->rx_opt.wscale_ok) { 6546 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0; 6547 WRITE_ONCE(tp->window_clamp, 6548 min(tp->window_clamp, 65535U)); 6549 } 6550 6551 if (tp->rx_opt.saw_tstamp) { 6552 tp->rx_opt.tstamp_ok = 1; 6553 tp->tcp_header_len = 6554 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 6555 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; 6556 tcp_store_ts_recent(tp); 6557 } else { 6558 tp->tcp_header_len = sizeof(struct tcphdr); 6559 } 6560 6561 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 6562 tcp_initialize_rcv_mss(sk); 6563 6564 /* Remember, tcp_poll() does not lock socket! 6565 * Change state from SYN-SENT only after copied_seq 6566 * is initialized. */ 6567 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt); 6568 6569 smc_check_reset_syn(tp); 6570 6571 smp_mb(); 6572 6573 tcp_finish_connect(sk, skb); 6574 6575 fastopen_fail = (tp->syn_fastopen || tp->syn_data) && 6576 tcp_rcv_fastopen_synack(sk, skb, &foc); 6577 6578 if (!sock_flag(sk, SOCK_DEAD)) { 6579 sk->sk_state_change(sk); 6580 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT); 6581 } 6582 if (fastopen_fail) 6583 return -1; 6584 if (sk->sk_write_pending || 6585 READ_ONCE(icsk->icsk_accept_queue.rskq_defer_accept) || 6586 inet_csk_in_pingpong_mode(sk)) { 6587 /* Save one ACK. Data will be ready after 6588 * several ticks, if write_pending is set. 6589 * 6590 * It may be deleted, but with this feature tcpdumps 6591 * look so _wonderfully_ clever, that I was not able 6592 * to stand against the temptation 8) --ANK 6593 */ 6594 inet_csk_schedule_ack(sk); 6595 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS); 6596 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 6597 TCP_DELACK_MAX, TCP_RTO_MAX); 6598 goto consume; 6599 } 6600 tcp_send_ack(sk); 6601 return -1; 6602 } 6603 6604 /* No ACK in the segment */ 6605 6606 if (th->rst) { 6607 /* rfc793: 6608 * "If the RST bit is set 6609 * 6610 * Otherwise (no ACK) drop the segment and return." 6611 */ 6612 SKB_DR_SET(reason, TCP_RESET); 6613 goto discard_and_undo; 6614 } 6615 6616 /* PAWS check. */ 6617 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && 6618 tcp_paws_reject(&tp->rx_opt, 0)) { 6619 SKB_DR_SET(reason, TCP_RFC7323_PAWS); 6620 goto discard_and_undo; 6621 } 6622 if (th->syn) { 6623 /* We see SYN without ACK. It is attempt of 6624 * simultaneous connect with crossed SYNs. 6625 * Particularly, it can be connect to self. 6626 */ 6627 #ifdef CONFIG_TCP_AO 6628 struct tcp_ao_info *ao; 6629 6630 ao = rcu_dereference_protected(tp->ao_info, 6631 lockdep_sock_is_held(sk)); 6632 if (ao) { 6633 WRITE_ONCE(ao->risn, th->seq); 6634 ao->rcv_sne = 0; 6635 } 6636 #endif 6637 tcp_set_state(sk, TCP_SYN_RECV); 6638 6639 if (tp->rx_opt.saw_tstamp) { 6640 tp->rx_opt.tstamp_ok = 1; 6641 tcp_store_ts_recent(tp); 6642 tp->tcp_header_len = 6643 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 6644 } else { 6645 tp->tcp_header_len = sizeof(struct tcphdr); 6646 } 6647 6648 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1); 6649 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt); 6650 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; 6651 6652 /* RFC1323: The window in SYN & SYN/ACK segments is 6653 * never scaled. 6654 */ 6655 tp->snd_wnd = ntohs(th->window); 6656 tp->snd_wl1 = TCP_SKB_CB(skb)->seq; 6657 tp->max_window = tp->snd_wnd; 6658 6659 tcp_ecn_rcv_syn(tp, th); 6660 6661 tcp_mtup_init(sk); 6662 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 6663 tcp_initialize_rcv_mss(sk); 6664 6665 tcp_send_synack(sk); 6666 #if 0 6667 /* Note, we could accept data and URG from this segment. 6668 * There are no obstacles to make this (except that we must 6669 * either change tcp_recvmsg() to prevent it from returning data 6670 * before 3WHS completes per RFC793, or employ TCP Fast Open). 6671 * 6672 * However, if we ignore data in ACKless segments sometimes, 6673 * we have no reasons to accept it sometimes. 6674 * Also, seems the code doing it in step6 of tcp_rcv_state_process 6675 * is not flawless. So, discard packet for sanity. 6676 * Uncomment this return to process the data. 6677 */ 6678 return -1; 6679 #else 6680 goto consume; 6681 #endif 6682 } 6683 /* "fifth, if neither of the SYN or RST bits is set then 6684 * drop the segment and return." 6685 */ 6686 6687 discard_and_undo: 6688 tcp_clear_options(&tp->rx_opt); 6689 tp->rx_opt.mss_clamp = saved_clamp; 6690 tcp_drop_reason(sk, skb, reason); 6691 return 0; 6692 6693 reset_and_undo: 6694 tcp_clear_options(&tp->rx_opt); 6695 tp->rx_opt.mss_clamp = saved_clamp; 6696 /* we can reuse/return @reason to its caller to handle the exception */ 6697 return reason; 6698 } 6699 6700 static void tcp_rcv_synrecv_state_fastopen(struct sock *sk) 6701 { 6702 struct tcp_sock *tp = tcp_sk(sk); 6703 struct request_sock *req; 6704 6705 /* If we are still handling the SYNACK RTO, see if timestamp ECR allows 6706 * undo. If peer SACKs triggered fast recovery, we can't undo here. 6707 */ 6708 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss && !tp->packets_out) 6709 tcp_try_undo_recovery(sk); 6710 6711 tcp_update_rto_time(tp); 6712 inet_csk(sk)->icsk_retransmits = 0; 6713 /* In tcp_fastopen_synack_timer() on the first SYNACK RTO we set 6714 * retrans_stamp but don't enter CA_Loss, so in case that happened we 6715 * need to zero retrans_stamp here to prevent spurious 6716 * retransmits_timed_out(). However, if the ACK of our SYNACK caused us 6717 * to enter CA_Recovery then we need to leave retrans_stamp as it was 6718 * set entering CA_Recovery, for correct retransmits_timed_out() and 6719 * undo behavior. 6720 */ 6721 tcp_retrans_stamp_cleanup(sk); 6722 6723 /* Once we leave TCP_SYN_RECV or TCP_FIN_WAIT_1, 6724 * we no longer need req so release it. 6725 */ 6726 req = rcu_dereference_protected(tp->fastopen_rsk, 6727 lockdep_sock_is_held(sk)); 6728 reqsk_fastopen_remove(sk, req, false); 6729 6730 /* Re-arm the timer because data may have been sent out. 6731 * This is similar to the regular data transmission case 6732 * when new data has just been ack'ed. 6733 * 6734 * (TFO) - we could try to be more aggressive and 6735 * retransmitting any data sooner based on when they 6736 * are sent out. 6737 */ 6738 tcp_rearm_rto(sk); 6739 } 6740 6741 /* 6742 * This function implements the receiving procedure of RFC 793 for 6743 * all states except ESTABLISHED and TIME_WAIT. 6744 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be 6745 * address independent. 6746 */ 6747 6748 enum skb_drop_reason 6749 tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb) 6750 { 6751 struct tcp_sock *tp = tcp_sk(sk); 6752 struct inet_connection_sock *icsk = inet_csk(sk); 6753 const struct tcphdr *th = tcp_hdr(skb); 6754 struct request_sock *req; 6755 int queued = 0; 6756 SKB_DR(reason); 6757 6758 switch (sk->sk_state) { 6759 case TCP_CLOSE: 6760 SKB_DR_SET(reason, TCP_CLOSE); 6761 goto discard; 6762 6763 case TCP_LISTEN: 6764 if (th->ack) 6765 return SKB_DROP_REASON_TCP_FLAGS; 6766 6767 if (th->rst) { 6768 SKB_DR_SET(reason, TCP_RESET); 6769 goto discard; 6770 } 6771 if (th->syn) { 6772 if (th->fin) { 6773 SKB_DR_SET(reason, TCP_FLAGS); 6774 goto discard; 6775 } 6776 /* It is possible that we process SYN packets from backlog, 6777 * so we need to make sure to disable BH and RCU right there. 6778 */ 6779 rcu_read_lock(); 6780 local_bh_disable(); 6781 icsk->icsk_af_ops->conn_request(sk, skb); 6782 local_bh_enable(); 6783 rcu_read_unlock(); 6784 6785 consume_skb(skb); 6786 return 0; 6787 } 6788 SKB_DR_SET(reason, TCP_FLAGS); 6789 goto discard; 6790 6791 case TCP_SYN_SENT: 6792 tp->rx_opt.saw_tstamp = 0; 6793 tcp_mstamp_refresh(tp); 6794 queued = tcp_rcv_synsent_state_process(sk, skb, th); 6795 if (queued >= 0) 6796 return queued; 6797 6798 /* Do step6 onward by hand. */ 6799 tcp_urg(sk, skb, th); 6800 __kfree_skb(skb); 6801 tcp_data_snd_check(sk); 6802 return 0; 6803 } 6804 6805 tcp_mstamp_refresh(tp); 6806 tp->rx_opt.saw_tstamp = 0; 6807 req = rcu_dereference_protected(tp->fastopen_rsk, 6808 lockdep_sock_is_held(sk)); 6809 if (req) { 6810 bool req_stolen; 6811 6812 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV && 6813 sk->sk_state != TCP_FIN_WAIT1); 6814 6815 if (!tcp_check_req(sk, skb, req, true, &req_stolen)) { 6816 SKB_DR_SET(reason, TCP_FASTOPEN); 6817 goto discard; 6818 } 6819 } 6820 6821 if (!th->ack && !th->rst && !th->syn) { 6822 SKB_DR_SET(reason, TCP_FLAGS); 6823 goto discard; 6824 } 6825 if (!tcp_validate_incoming(sk, skb, th, 0)) 6826 return 0; 6827 6828 /* step 5: check the ACK field */ 6829 reason = tcp_ack(sk, skb, FLAG_SLOWPATH | 6830 FLAG_UPDATE_TS_RECENT | 6831 FLAG_NO_CHALLENGE_ACK); 6832 6833 if ((int)reason <= 0) { 6834 if (sk->sk_state == TCP_SYN_RECV) { 6835 /* send one RST */ 6836 if (!reason) 6837 return SKB_DROP_REASON_TCP_OLD_ACK; 6838 return -reason; 6839 } 6840 /* accept old ack during closing */ 6841 if ((int)reason < 0) { 6842 tcp_send_challenge_ack(sk); 6843 reason = -reason; 6844 goto discard; 6845 } 6846 } 6847 SKB_DR_SET(reason, NOT_SPECIFIED); 6848 switch (sk->sk_state) { 6849 case TCP_SYN_RECV: 6850 tp->delivered++; /* SYN-ACK delivery isn't tracked in tcp_ack */ 6851 if (!tp->srtt_us) 6852 tcp_synack_rtt_meas(sk, req); 6853 6854 if (req) { 6855 tcp_rcv_synrecv_state_fastopen(sk); 6856 } else { 6857 tcp_try_undo_spurious_syn(sk); 6858 tp->retrans_stamp = 0; 6859 tcp_init_transfer(sk, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB, 6860 skb); 6861 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt); 6862 } 6863 tcp_ao_established(sk); 6864 smp_mb(); 6865 tcp_set_state(sk, TCP_ESTABLISHED); 6866 sk->sk_state_change(sk); 6867 6868 /* Note, that this wakeup is only for marginal crossed SYN case. 6869 * Passively open sockets are not waked up, because 6870 * sk->sk_sleep == NULL and sk->sk_socket == NULL. 6871 */ 6872 if (sk->sk_socket) 6873 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT); 6874 6875 tp->snd_una = TCP_SKB_CB(skb)->ack_seq; 6876 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale; 6877 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq); 6878 6879 if (tp->rx_opt.tstamp_ok) 6880 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; 6881 6882 if (!inet_csk(sk)->icsk_ca_ops->cong_control) 6883 tcp_update_pacing_rate(sk); 6884 6885 /* Prevent spurious tcp_cwnd_restart() on first data packet */ 6886 tp->lsndtime = tcp_jiffies32; 6887 6888 tcp_initialize_rcv_mss(sk); 6889 tcp_fast_path_on(tp); 6890 if (sk->sk_shutdown & SEND_SHUTDOWN) 6891 tcp_shutdown(sk, SEND_SHUTDOWN); 6892 break; 6893 6894 case TCP_FIN_WAIT1: { 6895 int tmo; 6896 6897 if (req) 6898 tcp_rcv_synrecv_state_fastopen(sk); 6899 6900 if (tp->snd_una != tp->write_seq) 6901 break; 6902 6903 tcp_set_state(sk, TCP_FIN_WAIT2); 6904 WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | SEND_SHUTDOWN); 6905 6906 sk_dst_confirm(sk); 6907 6908 if (!sock_flag(sk, SOCK_DEAD)) { 6909 /* Wake up lingering close() */ 6910 sk->sk_state_change(sk); 6911 break; 6912 } 6913 6914 if (READ_ONCE(tp->linger2) < 0) { 6915 tcp_done(sk); 6916 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA); 6917 return SKB_DROP_REASON_TCP_ABORT_ON_DATA; 6918 } 6919 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 6920 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) { 6921 /* Receive out of order FIN after close() */ 6922 if (tp->syn_fastopen && th->fin) 6923 tcp_fastopen_active_disable(sk); 6924 tcp_done(sk); 6925 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA); 6926 return SKB_DROP_REASON_TCP_ABORT_ON_DATA; 6927 } 6928 6929 tmo = tcp_fin_time(sk); 6930 if (tmo > TCP_TIMEWAIT_LEN) { 6931 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN); 6932 } else if (th->fin || sock_owned_by_user(sk)) { 6933 /* Bad case. We could lose such FIN otherwise. 6934 * It is not a big problem, but it looks confusing 6935 * and not so rare event. We still can lose it now, 6936 * if it spins in bh_lock_sock(), but it is really 6937 * marginal case. 6938 */ 6939 inet_csk_reset_keepalive_timer(sk, tmo); 6940 } else { 6941 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo); 6942 goto consume; 6943 } 6944 break; 6945 } 6946 6947 case TCP_CLOSING: 6948 if (tp->snd_una == tp->write_seq) { 6949 tcp_time_wait(sk, TCP_TIME_WAIT, 0); 6950 goto consume; 6951 } 6952 break; 6953 6954 case TCP_LAST_ACK: 6955 if (tp->snd_una == tp->write_seq) { 6956 tcp_update_metrics(sk); 6957 tcp_done(sk); 6958 goto consume; 6959 } 6960 break; 6961 } 6962 6963 /* step 6: check the URG bit */ 6964 tcp_urg(sk, skb, th); 6965 6966 /* step 7: process the segment text */ 6967 switch (sk->sk_state) { 6968 case TCP_CLOSE_WAIT: 6969 case TCP_CLOSING: 6970 case TCP_LAST_ACK: 6971 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 6972 /* If a subflow has been reset, the packet should not 6973 * continue to be processed, drop the packet. 6974 */ 6975 if (sk_is_mptcp(sk) && !mptcp_incoming_options(sk, skb)) 6976 goto discard; 6977 break; 6978 } 6979 fallthrough; 6980 case TCP_FIN_WAIT1: 6981 case TCP_FIN_WAIT2: 6982 /* RFC 793 says to queue data in these states, 6983 * RFC 1122 says we MUST send a reset. 6984 * BSD 4.4 also does reset. 6985 */ 6986 if (sk->sk_shutdown & RCV_SHUTDOWN) { 6987 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 6988 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) { 6989 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA); 6990 tcp_reset(sk, skb); 6991 return SKB_DROP_REASON_TCP_ABORT_ON_DATA; 6992 } 6993 } 6994 fallthrough; 6995 case TCP_ESTABLISHED: 6996 tcp_data_queue(sk, skb); 6997 queued = 1; 6998 break; 6999 } 7000 7001 /* tcp_data could move socket to TIME-WAIT */ 7002 if (sk->sk_state != TCP_CLOSE) { 7003 tcp_data_snd_check(sk); 7004 tcp_ack_snd_check(sk); 7005 } 7006 7007 if (!queued) { 7008 discard: 7009 tcp_drop_reason(sk, skb, reason); 7010 } 7011 return 0; 7012 7013 consume: 7014 __kfree_skb(skb); 7015 return 0; 7016 } 7017 EXPORT_SYMBOL(tcp_rcv_state_process); 7018 7019 static inline void pr_drop_req(struct request_sock *req, __u16 port, int family) 7020 { 7021 struct inet_request_sock *ireq = inet_rsk(req); 7022 7023 if (family == AF_INET) 7024 net_dbg_ratelimited("drop open request from %pI4/%u\n", 7025 &ireq->ir_rmt_addr, port); 7026 #if IS_ENABLED(CONFIG_IPV6) 7027 else if (family == AF_INET6) 7028 net_dbg_ratelimited("drop open request from %pI6/%u\n", 7029 &ireq->ir_v6_rmt_addr, port); 7030 #endif 7031 } 7032 7033 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set 7034 * 7035 * If we receive a SYN packet with these bits set, it means a 7036 * network is playing bad games with TOS bits. In order to 7037 * avoid possible false congestion notifications, we disable 7038 * TCP ECN negotiation. 7039 * 7040 * Exception: tcp_ca wants ECN. This is required for DCTCP 7041 * congestion control: Linux DCTCP asserts ECT on all packets, 7042 * including SYN, which is most optimal solution; however, 7043 * others, such as FreeBSD do not. 7044 * 7045 * Exception: At least one of the reserved bits of the TCP header (th->res1) is 7046 * set, indicating the use of a future TCP extension (such as AccECN). See 7047 * RFC8311 §4.3 which updates RFC3168 to allow the development of such 7048 * extensions. 7049 */ 7050 static void tcp_ecn_create_request(struct request_sock *req, 7051 const struct sk_buff *skb, 7052 const struct sock *listen_sk, 7053 const struct dst_entry *dst) 7054 { 7055 const struct tcphdr *th = tcp_hdr(skb); 7056 const struct net *net = sock_net(listen_sk); 7057 bool th_ecn = th->ece && th->cwr; 7058 bool ect, ecn_ok; 7059 u32 ecn_ok_dst; 7060 7061 if (!th_ecn) 7062 return; 7063 7064 ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield); 7065 ecn_ok_dst = dst_feature(dst, DST_FEATURE_ECN_MASK); 7066 ecn_ok = READ_ONCE(net->ipv4.sysctl_tcp_ecn) || ecn_ok_dst; 7067 7068 if (((!ect || th->res1) && ecn_ok) || tcp_ca_needs_ecn(listen_sk) || 7069 (ecn_ok_dst & DST_FEATURE_ECN_CA) || 7070 tcp_bpf_ca_needs_ecn((struct sock *)req)) 7071 inet_rsk(req)->ecn_ok = 1; 7072 } 7073 7074 static void tcp_openreq_init(struct request_sock *req, 7075 const struct tcp_options_received *rx_opt, 7076 struct sk_buff *skb, const struct sock *sk) 7077 { 7078 struct inet_request_sock *ireq = inet_rsk(req); 7079 7080 req->rsk_rcv_wnd = 0; /* So that tcp_send_synack() knows! */ 7081 tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq; 7082 tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; 7083 tcp_rsk(req)->snt_synack = 0; 7084 tcp_rsk(req)->last_oow_ack_time = 0; 7085 req->mss = rx_opt->mss_clamp; 7086 req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0; 7087 ireq->tstamp_ok = rx_opt->tstamp_ok; 7088 ireq->sack_ok = rx_opt->sack_ok; 7089 ireq->snd_wscale = rx_opt->snd_wscale; 7090 ireq->wscale_ok = rx_opt->wscale_ok; 7091 ireq->acked = 0; 7092 ireq->ecn_ok = 0; 7093 ireq->ir_rmt_port = tcp_hdr(skb)->source; 7094 ireq->ir_num = ntohs(tcp_hdr(skb)->dest); 7095 ireq->ir_mark = inet_request_mark(sk, skb); 7096 #if IS_ENABLED(CONFIG_SMC) 7097 ireq->smc_ok = rx_opt->smc_ok && !(tcp_sk(sk)->smc_hs_congested && 7098 tcp_sk(sk)->smc_hs_congested(sk)); 7099 #endif 7100 } 7101 7102 /* 7103 * Return true if a syncookie should be sent 7104 */ 7105 static bool tcp_syn_flood_action(struct sock *sk, const char *proto) 7106 { 7107 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue; 7108 const char *msg = "Dropping request"; 7109 struct net *net = sock_net(sk); 7110 bool want_cookie = false; 7111 u8 syncookies; 7112 7113 syncookies = READ_ONCE(net->ipv4.sysctl_tcp_syncookies); 7114 7115 #ifdef CONFIG_SYN_COOKIES 7116 if (syncookies) { 7117 msg = "Sending cookies"; 7118 want_cookie = true; 7119 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES); 7120 } else 7121 #endif 7122 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP); 7123 7124 if (!READ_ONCE(queue->synflood_warned) && syncookies != 2 && 7125 xchg(&queue->synflood_warned, 1) == 0) { 7126 if (IS_ENABLED(CONFIG_IPV6) && sk->sk_family == AF_INET6) { 7127 net_info_ratelimited("%s: Possible SYN flooding on port [%pI6c]:%u. %s.\n", 7128 proto, inet6_rcv_saddr(sk), 7129 sk->sk_num, msg); 7130 } else { 7131 net_info_ratelimited("%s: Possible SYN flooding on port %pI4:%u. %s.\n", 7132 proto, &sk->sk_rcv_saddr, 7133 sk->sk_num, msg); 7134 } 7135 } 7136 7137 return want_cookie; 7138 } 7139 7140 static void tcp_reqsk_record_syn(const struct sock *sk, 7141 struct request_sock *req, 7142 const struct sk_buff *skb) 7143 { 7144 if (tcp_sk(sk)->save_syn) { 7145 u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb); 7146 struct saved_syn *saved_syn; 7147 u32 mac_hdrlen; 7148 void *base; 7149 7150 if (tcp_sk(sk)->save_syn == 2) { /* Save full header. */ 7151 base = skb_mac_header(skb); 7152 mac_hdrlen = skb_mac_header_len(skb); 7153 len += mac_hdrlen; 7154 } else { 7155 base = skb_network_header(skb); 7156 mac_hdrlen = 0; 7157 } 7158 7159 saved_syn = kmalloc(struct_size(saved_syn, data, len), 7160 GFP_ATOMIC); 7161 if (saved_syn) { 7162 saved_syn->mac_hdrlen = mac_hdrlen; 7163 saved_syn->network_hdrlen = skb_network_header_len(skb); 7164 saved_syn->tcp_hdrlen = tcp_hdrlen(skb); 7165 memcpy(saved_syn->data, base, len); 7166 req->saved_syn = saved_syn; 7167 } 7168 } 7169 } 7170 7171 /* If a SYN cookie is required and supported, returns a clamped MSS value to be 7172 * used for SYN cookie generation. 7173 */ 7174 u16 tcp_get_syncookie_mss(struct request_sock_ops *rsk_ops, 7175 const struct tcp_request_sock_ops *af_ops, 7176 struct sock *sk, struct tcphdr *th) 7177 { 7178 struct tcp_sock *tp = tcp_sk(sk); 7179 u16 mss; 7180 7181 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_syncookies) != 2 && 7182 !inet_csk_reqsk_queue_is_full(sk)) 7183 return 0; 7184 7185 if (!tcp_syn_flood_action(sk, rsk_ops->slab_name)) 7186 return 0; 7187 7188 if (sk_acceptq_is_full(sk)) { 7189 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS); 7190 return 0; 7191 } 7192 7193 mss = tcp_parse_mss_option(th, tp->rx_opt.user_mss); 7194 if (!mss) 7195 mss = af_ops->mss_clamp; 7196 7197 return mss; 7198 } 7199 EXPORT_SYMBOL_GPL(tcp_get_syncookie_mss); 7200 7201 int tcp_conn_request(struct request_sock_ops *rsk_ops, 7202 const struct tcp_request_sock_ops *af_ops, 7203 struct sock *sk, struct sk_buff *skb) 7204 { 7205 struct tcp_fastopen_cookie foc = { .len = -1 }; 7206 struct tcp_options_received tmp_opt; 7207 struct tcp_sock *tp = tcp_sk(sk); 7208 struct net *net = sock_net(sk); 7209 struct sock *fastopen_sk = NULL; 7210 struct request_sock *req; 7211 bool want_cookie = false; 7212 struct dst_entry *dst; 7213 struct flowi fl; 7214 u8 syncookies; 7215 u32 isn; 7216 7217 #ifdef CONFIG_TCP_AO 7218 const struct tcp_ao_hdr *aoh; 7219 #endif 7220 7221 isn = __this_cpu_read(tcp_tw_isn); 7222 if (isn) { 7223 /* TW buckets are converted to open requests without 7224 * limitations, they conserve resources and peer is 7225 * evidently real one. 7226 */ 7227 __this_cpu_write(tcp_tw_isn, 0); 7228 } else { 7229 syncookies = READ_ONCE(net->ipv4.sysctl_tcp_syncookies); 7230 7231 if (syncookies == 2 || inet_csk_reqsk_queue_is_full(sk)) { 7232 want_cookie = tcp_syn_flood_action(sk, 7233 rsk_ops->slab_name); 7234 if (!want_cookie) 7235 goto drop; 7236 } 7237 } 7238 7239 if (sk_acceptq_is_full(sk)) { 7240 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS); 7241 goto drop; 7242 } 7243 7244 req = inet_reqsk_alloc(rsk_ops, sk, !want_cookie); 7245 if (!req) 7246 goto drop; 7247 7248 req->syncookie = want_cookie; 7249 tcp_rsk(req)->af_specific = af_ops; 7250 tcp_rsk(req)->ts_off = 0; 7251 tcp_rsk(req)->req_usec_ts = false; 7252 #if IS_ENABLED(CONFIG_MPTCP) 7253 tcp_rsk(req)->is_mptcp = 0; 7254 #endif 7255 7256 tcp_clear_options(&tmp_opt); 7257 tmp_opt.mss_clamp = af_ops->mss_clamp; 7258 tmp_opt.user_mss = tp->rx_opt.user_mss; 7259 tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0, 7260 want_cookie ? NULL : &foc); 7261 7262 if (want_cookie && !tmp_opt.saw_tstamp) 7263 tcp_clear_options(&tmp_opt); 7264 7265 if (IS_ENABLED(CONFIG_SMC) && want_cookie) 7266 tmp_opt.smc_ok = 0; 7267 7268 tmp_opt.tstamp_ok = tmp_opt.saw_tstamp; 7269 tcp_openreq_init(req, &tmp_opt, skb, sk); 7270 inet_rsk(req)->no_srccheck = inet_test_bit(TRANSPARENT, sk); 7271 7272 /* Note: tcp_v6_init_req() might override ir_iif for link locals */ 7273 inet_rsk(req)->ir_iif = inet_request_bound_dev_if(sk, skb); 7274 7275 dst = af_ops->route_req(sk, skb, &fl, req, isn); 7276 if (!dst) 7277 goto drop_and_free; 7278 7279 if (tmp_opt.tstamp_ok) { 7280 tcp_rsk(req)->req_usec_ts = dst_tcp_usec_ts(dst); 7281 tcp_rsk(req)->ts_off = af_ops->init_ts_off(net, skb); 7282 } 7283 if (!want_cookie && !isn) { 7284 int max_syn_backlog = READ_ONCE(net->ipv4.sysctl_max_syn_backlog); 7285 7286 /* Kill the following clause, if you dislike this way. */ 7287 if (!syncookies && 7288 (max_syn_backlog - inet_csk_reqsk_queue_len(sk) < 7289 (max_syn_backlog >> 2)) && 7290 !tcp_peer_is_proven(req, dst)) { 7291 /* Without syncookies last quarter of 7292 * backlog is filled with destinations, 7293 * proven to be alive. 7294 * It means that we continue to communicate 7295 * to destinations, already remembered 7296 * to the moment of synflood. 7297 */ 7298 pr_drop_req(req, ntohs(tcp_hdr(skb)->source), 7299 rsk_ops->family); 7300 goto drop_and_release; 7301 } 7302 7303 isn = af_ops->init_seq(skb); 7304 } 7305 7306 tcp_ecn_create_request(req, skb, sk, dst); 7307 7308 if (want_cookie) { 7309 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss); 7310 if (!tmp_opt.tstamp_ok) 7311 inet_rsk(req)->ecn_ok = 0; 7312 } 7313 7314 #ifdef CONFIG_TCP_AO 7315 if (tcp_parse_auth_options(tcp_hdr(skb), NULL, &aoh)) 7316 goto drop_and_release; /* Invalid TCP options */ 7317 if (aoh) { 7318 tcp_rsk(req)->used_tcp_ao = true; 7319 tcp_rsk(req)->ao_rcv_next = aoh->keyid; 7320 tcp_rsk(req)->ao_keyid = aoh->rnext_keyid; 7321 7322 } else { 7323 tcp_rsk(req)->used_tcp_ao = false; 7324 } 7325 #endif 7326 tcp_rsk(req)->snt_isn = isn; 7327 tcp_rsk(req)->txhash = net_tx_rndhash(); 7328 tcp_rsk(req)->syn_tos = TCP_SKB_CB(skb)->ip_dsfield; 7329 tcp_openreq_init_rwin(req, sk, dst); 7330 sk_rx_queue_set(req_to_sk(req), skb); 7331 if (!want_cookie) { 7332 tcp_reqsk_record_syn(sk, req, skb); 7333 fastopen_sk = tcp_try_fastopen(sk, skb, req, &foc, dst); 7334 } 7335 if (fastopen_sk) { 7336 af_ops->send_synack(fastopen_sk, dst, &fl, req, 7337 &foc, TCP_SYNACK_FASTOPEN, skb); 7338 /* Add the child socket directly into the accept queue */ 7339 if (!inet_csk_reqsk_queue_add(sk, req, fastopen_sk)) { 7340 reqsk_fastopen_remove(fastopen_sk, req, false); 7341 bh_unlock_sock(fastopen_sk); 7342 sock_put(fastopen_sk); 7343 goto drop_and_free; 7344 } 7345 sk->sk_data_ready(sk); 7346 bh_unlock_sock(fastopen_sk); 7347 sock_put(fastopen_sk); 7348 } else { 7349 tcp_rsk(req)->tfo_listener = false; 7350 if (!want_cookie) { 7351 req->timeout = tcp_timeout_init((struct sock *)req); 7352 if (unlikely(!inet_csk_reqsk_queue_hash_add(sk, req, 7353 req->timeout))) { 7354 reqsk_free(req); 7355 dst_release(dst); 7356 return 0; 7357 } 7358 7359 } 7360 af_ops->send_synack(sk, dst, &fl, req, &foc, 7361 !want_cookie ? TCP_SYNACK_NORMAL : 7362 TCP_SYNACK_COOKIE, 7363 skb); 7364 if (want_cookie) { 7365 reqsk_free(req); 7366 return 0; 7367 } 7368 } 7369 reqsk_put(req); 7370 return 0; 7371 7372 drop_and_release: 7373 dst_release(dst); 7374 drop_and_free: 7375 __reqsk_free(req); 7376 drop: 7377 tcp_listendrop(sk); 7378 return 0; 7379 } 7380 EXPORT_SYMBOL(tcp_conn_request); 7381