1 /* 2 * INET An implementation of the TCP/IP protocol suite for the LINUX 3 * operating system. INET is implemented using the BSD Socket 4 * interface as the means of communication with the user level. 5 * 6 * Implementation of the Transmission Control Protocol(TCP). 7 * 8 * Authors: Ross Biro 9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> 10 * Mark Evans, <evansmp@uhura.aston.ac.uk> 11 * Corey Minyard <wf-rch!minyard@relay.EU.net> 12 * Florian La Roche, <flla@stud.uni-sb.de> 13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu> 14 * Linus Torvalds, <torvalds@cs.helsinki.fi> 15 * Alan Cox, <gw4pts@gw4pts.ampr.org> 16 * Matthew Dillon, <dillon@apollo.west.oic.com> 17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no> 18 * Jorge Cwik, <jorge@laser.satlink.net> 19 */ 20 21 /* 22 * Changes: 23 * Pedro Roque : Fast Retransmit/Recovery. 24 * Two receive queues. 25 * Retransmit queue handled by TCP. 26 * Better retransmit timer handling. 27 * New congestion avoidance. 28 * Header prediction. 29 * Variable renaming. 30 * 31 * Eric : Fast Retransmit. 32 * Randy Scott : MSS option defines. 33 * Eric Schenk : Fixes to slow start algorithm. 34 * Eric Schenk : Yet another double ACK bug. 35 * Eric Schenk : Delayed ACK bug fixes. 36 * Eric Schenk : Floyd style fast retrans war avoidance. 37 * David S. Miller : Don't allow zero congestion window. 38 * Eric Schenk : Fix retransmitter so that it sends 39 * next packet on ack of previous packet. 40 * Andi Kleen : Moved open_request checking here 41 * and process RSTs for open_requests. 42 * Andi Kleen : Better prune_queue, and other fixes. 43 * Andrey Savochkin: Fix RTT measurements in the presence of 44 * timestamps. 45 * Andrey Savochkin: Check sequence numbers correctly when 46 * removing SACKs due to in sequence incoming 47 * data segments. 48 * Andi Kleen: Make sure we never ack data there is not 49 * enough room for. Also make this condition 50 * a fatal error if it might still happen. 51 * Andi Kleen: Add tcp_measure_rcv_mss to make 52 * connections with MSS<min(MTU,ann. MSS) 53 * work without delayed acks. 54 * Andi Kleen: Process packets with PSH set in the 55 * fast path. 56 * J Hadi Salim: ECN support 57 * Andrei Gurtov, 58 * Pasi Sarolahti, 59 * Panu Kuhlberg: Experimental audit of TCP (re)transmission 60 * engine. Lots of bugs are found. 61 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs 62 */ 63 64 #define pr_fmt(fmt) "TCP: " fmt 65 66 #include <linux/mm.h> 67 #include <linux/slab.h> 68 #include <linux/module.h> 69 #include <linux/sysctl.h> 70 #include <linux/kernel.h> 71 #include <linux/prefetch.h> 72 #include <net/dst.h> 73 #include <net/tcp.h> 74 #include <net/inet_common.h> 75 #include <linux/ipsec.h> 76 #include <asm/unaligned.h> 77 #include <linux/errqueue.h> 78 79 int sysctl_tcp_timestamps __read_mostly = 1; 80 int sysctl_tcp_window_scaling __read_mostly = 1; 81 int sysctl_tcp_sack __read_mostly = 1; 82 int sysctl_tcp_fack __read_mostly = 1; 83 int sysctl_tcp_max_reordering __read_mostly = 300; 84 int sysctl_tcp_dsack __read_mostly = 1; 85 int sysctl_tcp_app_win __read_mostly = 31; 86 int sysctl_tcp_adv_win_scale __read_mostly = 1; 87 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale); 88 89 /* rfc5961 challenge ack rate limiting */ 90 int sysctl_tcp_challenge_ack_limit = 100; 91 92 int sysctl_tcp_stdurg __read_mostly; 93 int sysctl_tcp_rfc1337 __read_mostly; 94 int sysctl_tcp_max_orphans __read_mostly = NR_FILE; 95 int sysctl_tcp_frto __read_mostly = 2; 96 int sysctl_tcp_min_rtt_wlen __read_mostly = 300; 97 98 int sysctl_tcp_thin_dupack __read_mostly; 99 100 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1; 101 int sysctl_tcp_early_retrans __read_mostly = 3; 102 int sysctl_tcp_invalid_ratelimit __read_mostly = HZ/2; 103 104 #define FLAG_DATA 0x01 /* Incoming frame contained data. */ 105 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */ 106 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */ 107 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */ 108 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */ 109 #define FLAG_DATA_SACKED 0x20 /* New SACK. */ 110 #define FLAG_ECE 0x40 /* ECE in this ACK */ 111 #define FLAG_LOST_RETRANS 0x80 /* This ACK marks some retransmission lost */ 112 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/ 113 #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */ 114 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */ 115 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */ 116 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */ 117 #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */ 118 119 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED) 120 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED) 121 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE) 122 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED) 123 124 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH) 125 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH)) 126 127 #define REXMIT_NONE 0 /* no loss recovery to do */ 128 #define REXMIT_LOST 1 /* retransmit packets marked lost */ 129 #define REXMIT_NEW 2 /* FRTO-style transmit of unsent/new packets */ 130 131 /* Adapt the MSS value used to make delayed ack decision to the 132 * real world. 133 */ 134 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb) 135 { 136 struct inet_connection_sock *icsk = inet_csk(sk); 137 const unsigned int lss = icsk->icsk_ack.last_seg_size; 138 unsigned int len; 139 140 icsk->icsk_ack.last_seg_size = 0; 141 142 /* skb->len may jitter because of SACKs, even if peer 143 * sends good full-sized frames. 144 */ 145 len = skb_shinfo(skb)->gso_size ? : skb->len; 146 if (len >= icsk->icsk_ack.rcv_mss) { 147 icsk->icsk_ack.rcv_mss = len; 148 } else { 149 /* Otherwise, we make more careful check taking into account, 150 * that SACKs block is variable. 151 * 152 * "len" is invariant segment length, including TCP header. 153 */ 154 len += skb->data - skb_transport_header(skb); 155 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) || 156 /* If PSH is not set, packet should be 157 * full sized, provided peer TCP is not badly broken. 158 * This observation (if it is correct 8)) allows 159 * to handle super-low mtu links fairly. 160 */ 161 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) && 162 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) { 163 /* Subtract also invariant (if peer is RFC compliant), 164 * tcp header plus fixed timestamp option length. 165 * Resulting "len" is MSS free of SACK jitter. 166 */ 167 len -= tcp_sk(sk)->tcp_header_len; 168 icsk->icsk_ack.last_seg_size = len; 169 if (len == lss) { 170 icsk->icsk_ack.rcv_mss = len; 171 return; 172 } 173 } 174 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED) 175 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2; 176 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED; 177 } 178 } 179 180 static void tcp_incr_quickack(struct sock *sk) 181 { 182 struct inet_connection_sock *icsk = inet_csk(sk); 183 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss); 184 185 if (quickacks == 0) 186 quickacks = 2; 187 if (quickacks > icsk->icsk_ack.quick) 188 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS); 189 } 190 191 static void tcp_enter_quickack_mode(struct sock *sk) 192 { 193 struct inet_connection_sock *icsk = inet_csk(sk); 194 tcp_incr_quickack(sk); 195 icsk->icsk_ack.pingpong = 0; 196 icsk->icsk_ack.ato = TCP_ATO_MIN; 197 } 198 199 /* Send ACKs quickly, if "quick" count is not exhausted 200 * and the session is not interactive. 201 */ 202 203 static bool tcp_in_quickack_mode(struct sock *sk) 204 { 205 const struct inet_connection_sock *icsk = inet_csk(sk); 206 const struct dst_entry *dst = __sk_dst_get(sk); 207 208 return (dst && dst_metric(dst, RTAX_QUICKACK)) || 209 (icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong); 210 } 211 212 static void tcp_ecn_queue_cwr(struct tcp_sock *tp) 213 { 214 if (tp->ecn_flags & TCP_ECN_OK) 215 tp->ecn_flags |= TCP_ECN_QUEUE_CWR; 216 } 217 218 static void tcp_ecn_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb) 219 { 220 if (tcp_hdr(skb)->cwr) 221 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR; 222 } 223 224 static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp) 225 { 226 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR; 227 } 228 229 static void __tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb) 230 { 231 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) { 232 case INET_ECN_NOT_ECT: 233 /* Funny extension: if ECT is not set on a segment, 234 * and we already seen ECT on a previous segment, 235 * it is probably a retransmit. 236 */ 237 if (tp->ecn_flags & TCP_ECN_SEEN) 238 tcp_enter_quickack_mode((struct sock *)tp); 239 break; 240 case INET_ECN_CE: 241 if (tcp_ca_needs_ecn((struct sock *)tp)) 242 tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_IS_CE); 243 244 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) { 245 /* Better not delay acks, sender can have a very low cwnd */ 246 tcp_enter_quickack_mode((struct sock *)tp); 247 tp->ecn_flags |= TCP_ECN_DEMAND_CWR; 248 } 249 tp->ecn_flags |= TCP_ECN_SEEN; 250 break; 251 default: 252 if (tcp_ca_needs_ecn((struct sock *)tp)) 253 tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_NO_CE); 254 tp->ecn_flags |= TCP_ECN_SEEN; 255 break; 256 } 257 } 258 259 static void tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb) 260 { 261 if (tp->ecn_flags & TCP_ECN_OK) 262 __tcp_ecn_check_ce(tp, skb); 263 } 264 265 static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th) 266 { 267 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr)) 268 tp->ecn_flags &= ~TCP_ECN_OK; 269 } 270 271 static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th) 272 { 273 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr)) 274 tp->ecn_flags &= ~TCP_ECN_OK; 275 } 276 277 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th) 278 { 279 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK)) 280 return true; 281 return false; 282 } 283 284 /* Buffer size and advertised window tuning. 285 * 286 * 1. Tuning sk->sk_sndbuf, when connection enters established state. 287 */ 288 289 static void tcp_sndbuf_expand(struct sock *sk) 290 { 291 const struct tcp_sock *tp = tcp_sk(sk); 292 int sndmem, per_mss; 293 u32 nr_segs; 294 295 /* Worst case is non GSO/TSO : each frame consumes one skb 296 * and skb->head is kmalloced using power of two area of memory 297 */ 298 per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) + 299 MAX_TCP_HEADER + 300 SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); 301 302 per_mss = roundup_pow_of_two(per_mss) + 303 SKB_DATA_ALIGN(sizeof(struct sk_buff)); 304 305 nr_segs = max_t(u32, TCP_INIT_CWND, tp->snd_cwnd); 306 nr_segs = max_t(u32, nr_segs, tp->reordering + 1); 307 308 /* Fast Recovery (RFC 5681 3.2) : 309 * Cubic needs 1.7 factor, rounded to 2 to include 310 * extra cushion (application might react slowly to POLLOUT) 311 */ 312 sndmem = 2 * nr_segs * per_mss; 313 314 if (sk->sk_sndbuf < sndmem) 315 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]); 316 } 317 318 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh) 319 * 320 * All tcp_full_space() is split to two parts: "network" buffer, allocated 321 * forward and advertised in receiver window (tp->rcv_wnd) and 322 * "application buffer", required to isolate scheduling/application 323 * latencies from network. 324 * window_clamp is maximal advertised window. It can be less than 325 * tcp_full_space(), in this case tcp_full_space() - window_clamp 326 * is reserved for "application" buffer. The less window_clamp is 327 * the smoother our behaviour from viewpoint of network, but the lower 328 * throughput and the higher sensitivity of the connection to losses. 8) 329 * 330 * rcv_ssthresh is more strict window_clamp used at "slow start" 331 * phase to predict further behaviour of this connection. 332 * It is used for two goals: 333 * - to enforce header prediction at sender, even when application 334 * requires some significant "application buffer". It is check #1. 335 * - to prevent pruning of receive queue because of misprediction 336 * of receiver window. Check #2. 337 * 338 * The scheme does not work when sender sends good segments opening 339 * window and then starts to feed us spaghetti. But it should work 340 * in common situations. Otherwise, we have to rely on queue collapsing. 341 */ 342 343 /* Slow part of check#2. */ 344 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb) 345 { 346 struct tcp_sock *tp = tcp_sk(sk); 347 /* Optimize this! */ 348 int truesize = tcp_win_from_space(skb->truesize) >> 1; 349 int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1; 350 351 while (tp->rcv_ssthresh <= window) { 352 if (truesize <= skb->len) 353 return 2 * inet_csk(sk)->icsk_ack.rcv_mss; 354 355 truesize >>= 1; 356 window >>= 1; 357 } 358 return 0; 359 } 360 361 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb) 362 { 363 struct tcp_sock *tp = tcp_sk(sk); 364 365 /* Check #1 */ 366 if (tp->rcv_ssthresh < tp->window_clamp && 367 (int)tp->rcv_ssthresh < tcp_space(sk) && 368 !tcp_under_memory_pressure(sk)) { 369 int incr; 370 371 /* Check #2. Increase window, if skb with such overhead 372 * will fit to rcvbuf in future. 373 */ 374 if (tcp_win_from_space(skb->truesize) <= skb->len) 375 incr = 2 * tp->advmss; 376 else 377 incr = __tcp_grow_window(sk, skb); 378 379 if (incr) { 380 incr = max_t(int, incr, 2 * skb->len); 381 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr, 382 tp->window_clamp); 383 inet_csk(sk)->icsk_ack.quick |= 1; 384 } 385 } 386 } 387 388 /* 3. Tuning rcvbuf, when connection enters established state. */ 389 static void tcp_fixup_rcvbuf(struct sock *sk) 390 { 391 u32 mss = tcp_sk(sk)->advmss; 392 int rcvmem; 393 394 rcvmem = 2 * SKB_TRUESIZE(mss + MAX_TCP_HEADER) * 395 tcp_default_init_rwnd(mss); 396 397 /* Dynamic Right Sizing (DRS) has 2 to 3 RTT latency 398 * Allow enough cushion so that sender is not limited by our window 399 */ 400 if (sysctl_tcp_moderate_rcvbuf) 401 rcvmem <<= 2; 402 403 if (sk->sk_rcvbuf < rcvmem) 404 sk->sk_rcvbuf = min(rcvmem, sysctl_tcp_rmem[2]); 405 } 406 407 /* 4. Try to fixup all. It is made immediately after connection enters 408 * established state. 409 */ 410 void tcp_init_buffer_space(struct sock *sk) 411 { 412 struct tcp_sock *tp = tcp_sk(sk); 413 int maxwin; 414 415 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) 416 tcp_fixup_rcvbuf(sk); 417 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK)) 418 tcp_sndbuf_expand(sk); 419 420 tp->rcvq_space.space = tp->rcv_wnd; 421 tp->rcvq_space.time = tcp_time_stamp; 422 tp->rcvq_space.seq = tp->copied_seq; 423 424 maxwin = tcp_full_space(sk); 425 426 if (tp->window_clamp >= maxwin) { 427 tp->window_clamp = maxwin; 428 429 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss) 430 tp->window_clamp = max(maxwin - 431 (maxwin >> sysctl_tcp_app_win), 432 4 * tp->advmss); 433 } 434 435 /* Force reservation of one segment. */ 436 if (sysctl_tcp_app_win && 437 tp->window_clamp > 2 * tp->advmss && 438 tp->window_clamp + tp->advmss > maxwin) 439 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss); 440 441 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp); 442 tp->snd_cwnd_stamp = tcp_time_stamp; 443 } 444 445 /* 5. Recalculate window clamp after socket hit its memory bounds. */ 446 static void tcp_clamp_window(struct sock *sk) 447 { 448 struct tcp_sock *tp = tcp_sk(sk); 449 struct inet_connection_sock *icsk = inet_csk(sk); 450 451 icsk->icsk_ack.quick = 0; 452 453 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] && 454 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) && 455 !tcp_under_memory_pressure(sk) && 456 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) { 457 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc), 458 sysctl_tcp_rmem[2]); 459 } 460 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf) 461 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss); 462 } 463 464 /* Initialize RCV_MSS value. 465 * RCV_MSS is an our guess about MSS used by the peer. 466 * We haven't any direct information about the MSS. 467 * It's better to underestimate the RCV_MSS rather than overestimate. 468 * Overestimations make us ACKing less frequently than needed. 469 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss(). 470 */ 471 void tcp_initialize_rcv_mss(struct sock *sk) 472 { 473 const struct tcp_sock *tp = tcp_sk(sk); 474 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache); 475 476 hint = min(hint, tp->rcv_wnd / 2); 477 hint = min(hint, TCP_MSS_DEFAULT); 478 hint = max(hint, TCP_MIN_MSS); 479 480 inet_csk(sk)->icsk_ack.rcv_mss = hint; 481 } 482 EXPORT_SYMBOL(tcp_initialize_rcv_mss); 483 484 /* Receiver "autotuning" code. 485 * 486 * The algorithm for RTT estimation w/o timestamps is based on 487 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL. 488 * <http://public.lanl.gov/radiant/pubs.html#DRS> 489 * 490 * More detail on this code can be found at 491 * <http://staff.psc.edu/jheffner/>, 492 * though this reference is out of date. A new paper 493 * is pending. 494 */ 495 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep) 496 { 497 u32 new_sample = tp->rcv_rtt_est.rtt; 498 long m = sample; 499 500 if (m == 0) 501 m = 1; 502 503 if (new_sample != 0) { 504 /* If we sample in larger samples in the non-timestamp 505 * case, we could grossly overestimate the RTT especially 506 * with chatty applications or bulk transfer apps which 507 * are stalled on filesystem I/O. 508 * 509 * Also, since we are only going for a minimum in the 510 * non-timestamp case, we do not smooth things out 511 * else with timestamps disabled convergence takes too 512 * long. 513 */ 514 if (!win_dep) { 515 m -= (new_sample >> 3); 516 new_sample += m; 517 } else { 518 m <<= 3; 519 if (m < new_sample) 520 new_sample = m; 521 } 522 } else { 523 /* No previous measure. */ 524 new_sample = m << 3; 525 } 526 527 if (tp->rcv_rtt_est.rtt != new_sample) 528 tp->rcv_rtt_est.rtt = new_sample; 529 } 530 531 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp) 532 { 533 if (tp->rcv_rtt_est.time == 0) 534 goto new_measure; 535 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq)) 536 return; 537 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rcv_rtt_est.time, 1); 538 539 new_measure: 540 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd; 541 tp->rcv_rtt_est.time = tcp_time_stamp; 542 } 543 544 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, 545 const struct sk_buff *skb) 546 { 547 struct tcp_sock *tp = tcp_sk(sk); 548 if (tp->rx_opt.rcv_tsecr && 549 (TCP_SKB_CB(skb)->end_seq - 550 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss)) 551 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0); 552 } 553 554 /* 555 * This function should be called every time data is copied to user space. 556 * It calculates the appropriate TCP receive buffer space. 557 */ 558 void tcp_rcv_space_adjust(struct sock *sk) 559 { 560 struct tcp_sock *tp = tcp_sk(sk); 561 int time; 562 int copied; 563 564 time = tcp_time_stamp - tp->rcvq_space.time; 565 if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0) 566 return; 567 568 /* Number of bytes copied to user in last RTT */ 569 copied = tp->copied_seq - tp->rcvq_space.seq; 570 if (copied <= tp->rcvq_space.space) 571 goto new_measure; 572 573 /* A bit of theory : 574 * copied = bytes received in previous RTT, our base window 575 * To cope with packet losses, we need a 2x factor 576 * To cope with slow start, and sender growing its cwin by 100 % 577 * every RTT, we need a 4x factor, because the ACK we are sending 578 * now is for the next RTT, not the current one : 579 * <prev RTT . ><current RTT .. ><next RTT .... > 580 */ 581 582 if (sysctl_tcp_moderate_rcvbuf && 583 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) { 584 int rcvwin, rcvmem, rcvbuf; 585 586 /* minimal window to cope with packet losses, assuming 587 * steady state. Add some cushion because of small variations. 588 */ 589 rcvwin = (copied << 1) + 16 * tp->advmss; 590 591 /* If rate increased by 25%, 592 * assume slow start, rcvwin = 3 * copied 593 * If rate increased by 50%, 594 * assume sender can use 2x growth, rcvwin = 4 * copied 595 */ 596 if (copied >= 597 tp->rcvq_space.space + (tp->rcvq_space.space >> 2)) { 598 if (copied >= 599 tp->rcvq_space.space + (tp->rcvq_space.space >> 1)) 600 rcvwin <<= 1; 601 else 602 rcvwin += (rcvwin >> 1); 603 } 604 605 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER); 606 while (tcp_win_from_space(rcvmem) < tp->advmss) 607 rcvmem += 128; 608 609 rcvbuf = min(rcvwin / tp->advmss * rcvmem, sysctl_tcp_rmem[2]); 610 if (rcvbuf > sk->sk_rcvbuf) { 611 sk->sk_rcvbuf = rcvbuf; 612 613 /* Make the window clamp follow along. */ 614 tp->window_clamp = rcvwin; 615 } 616 } 617 tp->rcvq_space.space = copied; 618 619 new_measure: 620 tp->rcvq_space.seq = tp->copied_seq; 621 tp->rcvq_space.time = tcp_time_stamp; 622 } 623 624 /* There is something which you must keep in mind when you analyze the 625 * behavior of the tp->ato delayed ack timeout interval. When a 626 * connection starts up, we want to ack as quickly as possible. The 627 * problem is that "good" TCP's do slow start at the beginning of data 628 * transmission. The means that until we send the first few ACK's the 629 * sender will sit on his end and only queue most of his data, because 630 * he can only send snd_cwnd unacked packets at any given time. For 631 * each ACK we send, he increments snd_cwnd and transmits more of his 632 * queue. -DaveM 633 */ 634 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb) 635 { 636 struct tcp_sock *tp = tcp_sk(sk); 637 struct inet_connection_sock *icsk = inet_csk(sk); 638 u32 now; 639 640 inet_csk_schedule_ack(sk); 641 642 tcp_measure_rcv_mss(sk, skb); 643 644 tcp_rcv_rtt_measure(tp); 645 646 now = tcp_time_stamp; 647 648 if (!icsk->icsk_ack.ato) { 649 /* The _first_ data packet received, initialize 650 * delayed ACK engine. 651 */ 652 tcp_incr_quickack(sk); 653 icsk->icsk_ack.ato = TCP_ATO_MIN; 654 } else { 655 int m = now - icsk->icsk_ack.lrcvtime; 656 657 if (m <= TCP_ATO_MIN / 2) { 658 /* The fastest case is the first. */ 659 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2; 660 } else if (m < icsk->icsk_ack.ato) { 661 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m; 662 if (icsk->icsk_ack.ato > icsk->icsk_rto) 663 icsk->icsk_ack.ato = icsk->icsk_rto; 664 } else if (m > icsk->icsk_rto) { 665 /* Too long gap. Apparently sender failed to 666 * restart window, so that we send ACKs quickly. 667 */ 668 tcp_incr_quickack(sk); 669 sk_mem_reclaim(sk); 670 } 671 } 672 icsk->icsk_ack.lrcvtime = now; 673 674 tcp_ecn_check_ce(tp, skb); 675 676 if (skb->len >= 128) 677 tcp_grow_window(sk, skb); 678 } 679 680 /* Called to compute a smoothed rtt estimate. The data fed to this 681 * routine either comes from timestamps, or from segments that were 682 * known _not_ to have been retransmitted [see Karn/Partridge 683 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88 684 * piece by Van Jacobson. 685 * NOTE: the next three routines used to be one big routine. 686 * To save cycles in the RFC 1323 implementation it was better to break 687 * it up into three procedures. -- erics 688 */ 689 static void tcp_rtt_estimator(struct sock *sk, long mrtt_us) 690 { 691 struct tcp_sock *tp = tcp_sk(sk); 692 long m = mrtt_us; /* RTT */ 693 u32 srtt = tp->srtt_us; 694 695 /* The following amusing code comes from Jacobson's 696 * article in SIGCOMM '88. Note that rtt and mdev 697 * are scaled versions of rtt and mean deviation. 698 * This is designed to be as fast as possible 699 * m stands for "measurement". 700 * 701 * On a 1990 paper the rto value is changed to: 702 * RTO = rtt + 4 * mdev 703 * 704 * Funny. This algorithm seems to be very broken. 705 * These formulae increase RTO, when it should be decreased, increase 706 * too slowly, when it should be increased quickly, decrease too quickly 707 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely 708 * does not matter how to _calculate_ it. Seems, it was trap 709 * that VJ failed to avoid. 8) 710 */ 711 if (srtt != 0) { 712 m -= (srtt >> 3); /* m is now error in rtt est */ 713 srtt += m; /* rtt = 7/8 rtt + 1/8 new */ 714 if (m < 0) { 715 m = -m; /* m is now abs(error) */ 716 m -= (tp->mdev_us >> 2); /* similar update on mdev */ 717 /* This is similar to one of Eifel findings. 718 * Eifel blocks mdev updates when rtt decreases. 719 * This solution is a bit different: we use finer gain 720 * for mdev in this case (alpha*beta). 721 * Like Eifel it also prevents growth of rto, 722 * but also it limits too fast rto decreases, 723 * happening in pure Eifel. 724 */ 725 if (m > 0) 726 m >>= 3; 727 } else { 728 m -= (tp->mdev_us >> 2); /* similar update on mdev */ 729 } 730 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */ 731 if (tp->mdev_us > tp->mdev_max_us) { 732 tp->mdev_max_us = tp->mdev_us; 733 if (tp->mdev_max_us > tp->rttvar_us) 734 tp->rttvar_us = tp->mdev_max_us; 735 } 736 if (after(tp->snd_una, tp->rtt_seq)) { 737 if (tp->mdev_max_us < tp->rttvar_us) 738 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2; 739 tp->rtt_seq = tp->snd_nxt; 740 tp->mdev_max_us = tcp_rto_min_us(sk); 741 } 742 } else { 743 /* no previous measure. */ 744 srtt = m << 3; /* take the measured time to be rtt */ 745 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */ 746 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk)); 747 tp->mdev_max_us = tp->rttvar_us; 748 tp->rtt_seq = tp->snd_nxt; 749 } 750 tp->srtt_us = max(1U, srtt); 751 } 752 753 /* Set the sk_pacing_rate to allow proper sizing of TSO packets. 754 * Note: TCP stack does not yet implement pacing. 755 * FQ packet scheduler can be used to implement cheap but effective 756 * TCP pacing, to smooth the burst on large writes when packets 757 * in flight is significantly lower than cwnd (or rwin) 758 */ 759 int sysctl_tcp_pacing_ss_ratio __read_mostly = 200; 760 int sysctl_tcp_pacing_ca_ratio __read_mostly = 120; 761 762 static void tcp_update_pacing_rate(struct sock *sk) 763 { 764 const struct tcp_sock *tp = tcp_sk(sk); 765 u64 rate; 766 767 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */ 768 rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3); 769 770 /* current rate is (cwnd * mss) / srtt 771 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate. 772 * In Congestion Avoidance phase, set it to 120 % the current rate. 773 * 774 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh) 775 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching 776 * end of slow start and should slow down. 777 */ 778 if (tp->snd_cwnd < tp->snd_ssthresh / 2) 779 rate *= sysctl_tcp_pacing_ss_ratio; 780 else 781 rate *= sysctl_tcp_pacing_ca_ratio; 782 783 rate *= max(tp->snd_cwnd, tp->packets_out); 784 785 if (likely(tp->srtt_us)) 786 do_div(rate, tp->srtt_us); 787 788 /* ACCESS_ONCE() is needed because sch_fq fetches sk_pacing_rate 789 * without any lock. We want to make sure compiler wont store 790 * intermediate values in this location. 791 */ 792 ACCESS_ONCE(sk->sk_pacing_rate) = min_t(u64, rate, 793 sk->sk_max_pacing_rate); 794 } 795 796 /* Calculate rto without backoff. This is the second half of Van Jacobson's 797 * routine referred to above. 798 */ 799 static void tcp_set_rto(struct sock *sk) 800 { 801 const struct tcp_sock *tp = tcp_sk(sk); 802 /* Old crap is replaced with new one. 8) 803 * 804 * More seriously: 805 * 1. If rtt variance happened to be less 50msec, it is hallucination. 806 * It cannot be less due to utterly erratic ACK generation made 807 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_ 808 * to do with delayed acks, because at cwnd>2 true delack timeout 809 * is invisible. Actually, Linux-2.4 also generates erratic 810 * ACKs in some circumstances. 811 */ 812 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp); 813 814 /* 2. Fixups made earlier cannot be right. 815 * If we do not estimate RTO correctly without them, 816 * all the algo is pure shit and should be replaced 817 * with correct one. It is exactly, which we pretend to do. 818 */ 819 820 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo 821 * guarantees that rto is higher. 822 */ 823 tcp_bound_rto(sk); 824 } 825 826 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst) 827 { 828 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0); 829 830 if (!cwnd) 831 cwnd = TCP_INIT_CWND; 832 return min_t(__u32, cwnd, tp->snd_cwnd_clamp); 833 } 834 835 /* 836 * Packet counting of FACK is based on in-order assumptions, therefore TCP 837 * disables it when reordering is detected 838 */ 839 void tcp_disable_fack(struct tcp_sock *tp) 840 { 841 /* RFC3517 uses different metric in lost marker => reset on change */ 842 if (tcp_is_fack(tp)) 843 tp->lost_skb_hint = NULL; 844 tp->rx_opt.sack_ok &= ~TCP_FACK_ENABLED; 845 } 846 847 /* Take a notice that peer is sending D-SACKs */ 848 static void tcp_dsack_seen(struct tcp_sock *tp) 849 { 850 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN; 851 } 852 853 static void tcp_update_reordering(struct sock *sk, const int metric, 854 const int ts) 855 { 856 struct tcp_sock *tp = tcp_sk(sk); 857 if (metric > tp->reordering) { 858 int mib_idx; 859 860 tp->reordering = min(sysctl_tcp_max_reordering, metric); 861 862 /* This exciting event is worth to be remembered. 8) */ 863 if (ts) 864 mib_idx = LINUX_MIB_TCPTSREORDER; 865 else if (tcp_is_reno(tp)) 866 mib_idx = LINUX_MIB_TCPRENOREORDER; 867 else if (tcp_is_fack(tp)) 868 mib_idx = LINUX_MIB_TCPFACKREORDER; 869 else 870 mib_idx = LINUX_MIB_TCPSACKREORDER; 871 872 NET_INC_STATS_BH(sock_net(sk), mib_idx); 873 #if FASTRETRANS_DEBUG > 1 874 pr_debug("Disorder%d %d %u f%u s%u rr%d\n", 875 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state, 876 tp->reordering, 877 tp->fackets_out, 878 tp->sacked_out, 879 tp->undo_marker ? tp->undo_retrans : 0); 880 #endif 881 tcp_disable_fack(tp); 882 } 883 884 if (metric > 0) 885 tcp_disable_early_retrans(tp); 886 tp->rack.reord = 1; 887 } 888 889 /* This must be called before lost_out is incremented */ 890 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb) 891 { 892 if (!tp->retransmit_skb_hint || 893 before(TCP_SKB_CB(skb)->seq, 894 TCP_SKB_CB(tp->retransmit_skb_hint)->seq)) 895 tp->retransmit_skb_hint = skb; 896 897 if (!tp->lost_out || 898 after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high)) 899 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq; 900 } 901 902 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb) 903 { 904 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) { 905 tcp_verify_retransmit_hint(tp, skb); 906 907 tp->lost_out += tcp_skb_pcount(skb); 908 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 909 } 910 } 911 912 void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp, struct sk_buff *skb) 913 { 914 tcp_verify_retransmit_hint(tp, skb); 915 916 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) { 917 tp->lost_out += tcp_skb_pcount(skb); 918 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 919 } 920 } 921 922 /* This procedure tags the retransmission queue when SACKs arrive. 923 * 924 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L). 925 * Packets in queue with these bits set are counted in variables 926 * sacked_out, retrans_out and lost_out, correspondingly. 927 * 928 * Valid combinations are: 929 * Tag InFlight Description 930 * 0 1 - orig segment is in flight. 931 * S 0 - nothing flies, orig reached receiver. 932 * L 0 - nothing flies, orig lost by net. 933 * R 2 - both orig and retransmit are in flight. 934 * L|R 1 - orig is lost, retransmit is in flight. 935 * S|R 1 - orig reached receiver, retrans is still in flight. 936 * (L|S|R is logically valid, it could occur when L|R is sacked, 937 * but it is equivalent to plain S and code short-curcuits it to S. 938 * L|S is logically invalid, it would mean -1 packet in flight 8)) 939 * 940 * These 6 states form finite state machine, controlled by the following events: 941 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue()) 942 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue()) 943 * 3. Loss detection event of two flavors: 944 * A. Scoreboard estimator decided the packet is lost. 945 * A'. Reno "three dupacks" marks head of queue lost. 946 * A''. Its FACK modification, head until snd.fack is lost. 947 * B. SACK arrives sacking SND.NXT at the moment, when the 948 * segment was retransmitted. 949 * 4. D-SACK added new rule: D-SACK changes any tag to S. 950 * 951 * It is pleasant to note, that state diagram turns out to be commutative, 952 * so that we are allowed not to be bothered by order of our actions, 953 * when multiple events arrive simultaneously. (see the function below). 954 * 955 * Reordering detection. 956 * -------------------- 957 * Reordering metric is maximal distance, which a packet can be displaced 958 * in packet stream. With SACKs we can estimate it: 959 * 960 * 1. SACK fills old hole and the corresponding segment was not 961 * ever retransmitted -> reordering. Alas, we cannot use it 962 * when segment was retransmitted. 963 * 2. The last flaw is solved with D-SACK. D-SACK arrives 964 * for retransmitted and already SACKed segment -> reordering.. 965 * Both of these heuristics are not used in Loss state, when we cannot 966 * account for retransmits accurately. 967 * 968 * SACK block validation. 969 * ---------------------- 970 * 971 * SACK block range validation checks that the received SACK block fits to 972 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT. 973 * Note that SND.UNA is not included to the range though being valid because 974 * it means that the receiver is rather inconsistent with itself reporting 975 * SACK reneging when it should advance SND.UNA. Such SACK block this is 976 * perfectly valid, however, in light of RFC2018 which explicitly states 977 * that "SACK block MUST reflect the newest segment. Even if the newest 978 * segment is going to be discarded ...", not that it looks very clever 979 * in case of head skb. Due to potentional receiver driven attacks, we 980 * choose to avoid immediate execution of a walk in write queue due to 981 * reneging and defer head skb's loss recovery to standard loss recovery 982 * procedure that will eventually trigger (nothing forbids us doing this). 983 * 984 * Implements also blockage to start_seq wrap-around. Problem lies in the 985 * fact that though start_seq (s) is before end_seq (i.e., not reversed), 986 * there's no guarantee that it will be before snd_nxt (n). The problem 987 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt 988 * wrap (s_w): 989 * 990 * <- outs wnd -> <- wrapzone -> 991 * u e n u_w e_w s n_w 992 * | | | | | | | 993 * |<------------+------+----- TCP seqno space --------------+---------->| 994 * ...-- <2^31 ->| |<--------... 995 * ...---- >2^31 ------>| |<--------... 996 * 997 * Current code wouldn't be vulnerable but it's better still to discard such 998 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat 999 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in 1000 * snd_nxt wrap -> snd_una region will then become "well defined", i.e., 1001 * equal to the ideal case (infinite seqno space without wrap caused issues). 1002 * 1003 * With D-SACK the lower bound is extended to cover sequence space below 1004 * SND.UNA down to undo_marker, which is the last point of interest. Yet 1005 * again, D-SACK block must not to go across snd_una (for the same reason as 1006 * for the normal SACK blocks, explained above). But there all simplicity 1007 * ends, TCP might receive valid D-SACKs below that. As long as they reside 1008 * fully below undo_marker they do not affect behavior in anyway and can 1009 * therefore be safely ignored. In rare cases (which are more or less 1010 * theoretical ones), the D-SACK will nicely cross that boundary due to skb 1011 * fragmentation and packet reordering past skb's retransmission. To consider 1012 * them correctly, the acceptable range must be extended even more though 1013 * the exact amount is rather hard to quantify. However, tp->max_window can 1014 * be used as an exaggerated estimate. 1015 */ 1016 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack, 1017 u32 start_seq, u32 end_seq) 1018 { 1019 /* Too far in future, or reversed (interpretation is ambiguous) */ 1020 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq)) 1021 return false; 1022 1023 /* Nasty start_seq wrap-around check (see comments above) */ 1024 if (!before(start_seq, tp->snd_nxt)) 1025 return false; 1026 1027 /* In outstanding window? ...This is valid exit for D-SACKs too. 1028 * start_seq == snd_una is non-sensical (see comments above) 1029 */ 1030 if (after(start_seq, tp->snd_una)) 1031 return true; 1032 1033 if (!is_dsack || !tp->undo_marker) 1034 return false; 1035 1036 /* ...Then it's D-SACK, and must reside below snd_una completely */ 1037 if (after(end_seq, tp->snd_una)) 1038 return false; 1039 1040 if (!before(start_seq, tp->undo_marker)) 1041 return true; 1042 1043 /* Too old */ 1044 if (!after(end_seq, tp->undo_marker)) 1045 return false; 1046 1047 /* Undo_marker boundary crossing (overestimates a lot). Known already: 1048 * start_seq < undo_marker and end_seq >= undo_marker. 1049 */ 1050 return !before(start_seq, end_seq - tp->max_window); 1051 } 1052 1053 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb, 1054 struct tcp_sack_block_wire *sp, int num_sacks, 1055 u32 prior_snd_una) 1056 { 1057 struct tcp_sock *tp = tcp_sk(sk); 1058 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq); 1059 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq); 1060 bool dup_sack = false; 1061 1062 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) { 1063 dup_sack = true; 1064 tcp_dsack_seen(tp); 1065 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV); 1066 } else if (num_sacks > 1) { 1067 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq); 1068 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq); 1069 1070 if (!after(end_seq_0, end_seq_1) && 1071 !before(start_seq_0, start_seq_1)) { 1072 dup_sack = true; 1073 tcp_dsack_seen(tp); 1074 NET_INC_STATS_BH(sock_net(sk), 1075 LINUX_MIB_TCPDSACKOFORECV); 1076 } 1077 } 1078 1079 /* D-SACK for already forgotten data... Do dumb counting. */ 1080 if (dup_sack && tp->undo_marker && tp->undo_retrans > 0 && 1081 !after(end_seq_0, prior_snd_una) && 1082 after(end_seq_0, tp->undo_marker)) 1083 tp->undo_retrans--; 1084 1085 return dup_sack; 1086 } 1087 1088 struct tcp_sacktag_state { 1089 int reord; 1090 int fack_count; 1091 /* Timestamps for earliest and latest never-retransmitted segment 1092 * that was SACKed. RTO needs the earliest RTT to stay conservative, 1093 * but congestion control should still get an accurate delay signal. 1094 */ 1095 struct skb_mstamp first_sackt; 1096 struct skb_mstamp last_sackt; 1097 int flag; 1098 }; 1099 1100 /* Check if skb is fully within the SACK block. In presence of GSO skbs, 1101 * the incoming SACK may not exactly match but we can find smaller MSS 1102 * aligned portion of it that matches. Therefore we might need to fragment 1103 * which may fail and creates some hassle (caller must handle error case 1104 * returns). 1105 * 1106 * FIXME: this could be merged to shift decision code 1107 */ 1108 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb, 1109 u32 start_seq, u32 end_seq) 1110 { 1111 int err; 1112 bool in_sack; 1113 unsigned int pkt_len; 1114 unsigned int mss; 1115 1116 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && 1117 !before(end_seq, TCP_SKB_CB(skb)->end_seq); 1118 1119 if (tcp_skb_pcount(skb) > 1 && !in_sack && 1120 after(TCP_SKB_CB(skb)->end_seq, start_seq)) { 1121 mss = tcp_skb_mss(skb); 1122 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); 1123 1124 if (!in_sack) { 1125 pkt_len = start_seq - TCP_SKB_CB(skb)->seq; 1126 if (pkt_len < mss) 1127 pkt_len = mss; 1128 } else { 1129 pkt_len = end_seq - TCP_SKB_CB(skb)->seq; 1130 if (pkt_len < mss) 1131 return -EINVAL; 1132 } 1133 1134 /* Round if necessary so that SACKs cover only full MSSes 1135 * and/or the remaining small portion (if present) 1136 */ 1137 if (pkt_len > mss) { 1138 unsigned int new_len = (pkt_len / mss) * mss; 1139 if (!in_sack && new_len < pkt_len) { 1140 new_len += mss; 1141 if (new_len >= skb->len) 1142 return 0; 1143 } 1144 pkt_len = new_len; 1145 } 1146 err = tcp_fragment(sk, skb, pkt_len, mss, GFP_ATOMIC); 1147 if (err < 0) 1148 return err; 1149 } 1150 1151 return in_sack; 1152 } 1153 1154 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */ 1155 static u8 tcp_sacktag_one(struct sock *sk, 1156 struct tcp_sacktag_state *state, u8 sacked, 1157 u32 start_seq, u32 end_seq, 1158 int dup_sack, int pcount, 1159 const struct skb_mstamp *xmit_time) 1160 { 1161 struct tcp_sock *tp = tcp_sk(sk); 1162 int fack_count = state->fack_count; 1163 1164 /* Account D-SACK for retransmitted packet. */ 1165 if (dup_sack && (sacked & TCPCB_RETRANS)) { 1166 if (tp->undo_marker && tp->undo_retrans > 0 && 1167 after(end_seq, tp->undo_marker)) 1168 tp->undo_retrans--; 1169 if (sacked & TCPCB_SACKED_ACKED) 1170 state->reord = min(fack_count, state->reord); 1171 } 1172 1173 /* Nothing to do; acked frame is about to be dropped (was ACKed). */ 1174 if (!after(end_seq, tp->snd_una)) 1175 return sacked; 1176 1177 if (!(sacked & TCPCB_SACKED_ACKED)) { 1178 tcp_rack_advance(tp, xmit_time, sacked); 1179 1180 if (sacked & TCPCB_SACKED_RETRANS) { 1181 /* If the segment is not tagged as lost, 1182 * we do not clear RETRANS, believing 1183 * that retransmission is still in flight. 1184 */ 1185 if (sacked & TCPCB_LOST) { 1186 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS); 1187 tp->lost_out -= pcount; 1188 tp->retrans_out -= pcount; 1189 } 1190 } else { 1191 if (!(sacked & TCPCB_RETRANS)) { 1192 /* New sack for not retransmitted frame, 1193 * which was in hole. It is reordering. 1194 */ 1195 if (before(start_seq, 1196 tcp_highest_sack_seq(tp))) 1197 state->reord = min(fack_count, 1198 state->reord); 1199 if (!after(end_seq, tp->high_seq)) 1200 state->flag |= FLAG_ORIG_SACK_ACKED; 1201 if (state->first_sackt.v64 == 0) 1202 state->first_sackt = *xmit_time; 1203 state->last_sackt = *xmit_time; 1204 } 1205 1206 if (sacked & TCPCB_LOST) { 1207 sacked &= ~TCPCB_LOST; 1208 tp->lost_out -= pcount; 1209 } 1210 } 1211 1212 sacked |= TCPCB_SACKED_ACKED; 1213 state->flag |= FLAG_DATA_SACKED; 1214 tp->sacked_out += pcount; 1215 tp->delivered += pcount; /* Out-of-order packets delivered */ 1216 1217 fack_count += pcount; 1218 1219 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */ 1220 if (!tcp_is_fack(tp) && tp->lost_skb_hint && 1221 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq)) 1222 tp->lost_cnt_hint += pcount; 1223 1224 if (fack_count > tp->fackets_out) 1225 tp->fackets_out = fack_count; 1226 } 1227 1228 /* D-SACK. We can detect redundant retransmission in S|R and plain R 1229 * frames and clear it. undo_retrans is decreased above, L|R frames 1230 * are accounted above as well. 1231 */ 1232 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) { 1233 sacked &= ~TCPCB_SACKED_RETRANS; 1234 tp->retrans_out -= pcount; 1235 } 1236 1237 return sacked; 1238 } 1239 1240 /* Shift newly-SACKed bytes from this skb to the immediately previous 1241 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such. 1242 */ 1243 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *skb, 1244 struct tcp_sacktag_state *state, 1245 unsigned int pcount, int shifted, int mss, 1246 bool dup_sack) 1247 { 1248 struct tcp_sock *tp = tcp_sk(sk); 1249 struct sk_buff *prev = tcp_write_queue_prev(sk, skb); 1250 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */ 1251 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */ 1252 1253 BUG_ON(!pcount); 1254 1255 /* Adjust counters and hints for the newly sacked sequence 1256 * range but discard the return value since prev is already 1257 * marked. We must tag the range first because the seq 1258 * advancement below implicitly advances 1259 * tcp_highest_sack_seq() when skb is highest_sack. 1260 */ 1261 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked, 1262 start_seq, end_seq, dup_sack, pcount, 1263 &skb->skb_mstamp); 1264 1265 if (skb == tp->lost_skb_hint) 1266 tp->lost_cnt_hint += pcount; 1267 1268 TCP_SKB_CB(prev)->end_seq += shifted; 1269 TCP_SKB_CB(skb)->seq += shifted; 1270 1271 tcp_skb_pcount_add(prev, pcount); 1272 BUG_ON(tcp_skb_pcount(skb) < pcount); 1273 tcp_skb_pcount_add(skb, -pcount); 1274 1275 /* When we're adding to gso_segs == 1, gso_size will be zero, 1276 * in theory this shouldn't be necessary but as long as DSACK 1277 * code can come after this skb later on it's better to keep 1278 * setting gso_size to something. 1279 */ 1280 if (!TCP_SKB_CB(prev)->tcp_gso_size) 1281 TCP_SKB_CB(prev)->tcp_gso_size = mss; 1282 1283 /* CHECKME: To clear or not to clear? Mimics normal skb currently */ 1284 if (tcp_skb_pcount(skb) <= 1) 1285 TCP_SKB_CB(skb)->tcp_gso_size = 0; 1286 1287 /* Difference in this won't matter, both ACKed by the same cumul. ACK */ 1288 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS); 1289 1290 if (skb->len > 0) { 1291 BUG_ON(!tcp_skb_pcount(skb)); 1292 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED); 1293 return false; 1294 } 1295 1296 /* Whole SKB was eaten :-) */ 1297 1298 if (skb == tp->retransmit_skb_hint) 1299 tp->retransmit_skb_hint = prev; 1300 if (skb == tp->lost_skb_hint) { 1301 tp->lost_skb_hint = prev; 1302 tp->lost_cnt_hint -= tcp_skb_pcount(prev); 1303 } 1304 1305 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; 1306 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 1307 TCP_SKB_CB(prev)->end_seq++; 1308 1309 if (skb == tcp_highest_sack(sk)) 1310 tcp_advance_highest_sack(sk, skb); 1311 1312 tcp_skb_collapse_tstamp(prev, skb); 1313 tcp_unlink_write_queue(skb, sk); 1314 sk_wmem_free_skb(sk, skb); 1315 1316 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED); 1317 1318 return true; 1319 } 1320 1321 /* I wish gso_size would have a bit more sane initialization than 1322 * something-or-zero which complicates things 1323 */ 1324 static int tcp_skb_seglen(const struct sk_buff *skb) 1325 { 1326 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb); 1327 } 1328 1329 /* Shifting pages past head area doesn't work */ 1330 static int skb_can_shift(const struct sk_buff *skb) 1331 { 1332 return !skb_headlen(skb) && skb_is_nonlinear(skb); 1333 } 1334 1335 /* Try collapsing SACK blocks spanning across multiple skbs to a single 1336 * skb. 1337 */ 1338 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb, 1339 struct tcp_sacktag_state *state, 1340 u32 start_seq, u32 end_seq, 1341 bool dup_sack) 1342 { 1343 struct tcp_sock *tp = tcp_sk(sk); 1344 struct sk_buff *prev; 1345 int mss; 1346 int pcount = 0; 1347 int len; 1348 int in_sack; 1349 1350 if (!sk_can_gso(sk)) 1351 goto fallback; 1352 1353 /* Normally R but no L won't result in plain S */ 1354 if (!dup_sack && 1355 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS) 1356 goto fallback; 1357 if (!skb_can_shift(skb)) 1358 goto fallback; 1359 /* This frame is about to be dropped (was ACKed). */ 1360 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 1361 goto fallback; 1362 1363 /* Can only happen with delayed DSACK + discard craziness */ 1364 if (unlikely(skb == tcp_write_queue_head(sk))) 1365 goto fallback; 1366 prev = tcp_write_queue_prev(sk, skb); 1367 1368 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) 1369 goto fallback; 1370 1371 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && 1372 !before(end_seq, TCP_SKB_CB(skb)->end_seq); 1373 1374 if (in_sack) { 1375 len = skb->len; 1376 pcount = tcp_skb_pcount(skb); 1377 mss = tcp_skb_seglen(skb); 1378 1379 /* TODO: Fix DSACKs to not fragment already SACKed and we can 1380 * drop this restriction as unnecessary 1381 */ 1382 if (mss != tcp_skb_seglen(prev)) 1383 goto fallback; 1384 } else { 1385 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq)) 1386 goto noop; 1387 /* CHECKME: This is non-MSS split case only?, this will 1388 * cause skipped skbs due to advancing loop btw, original 1389 * has that feature too 1390 */ 1391 if (tcp_skb_pcount(skb) <= 1) 1392 goto noop; 1393 1394 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); 1395 if (!in_sack) { 1396 /* TODO: head merge to next could be attempted here 1397 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)), 1398 * though it might not be worth of the additional hassle 1399 * 1400 * ...we can probably just fallback to what was done 1401 * previously. We could try merging non-SACKed ones 1402 * as well but it probably isn't going to buy off 1403 * because later SACKs might again split them, and 1404 * it would make skb timestamp tracking considerably 1405 * harder problem. 1406 */ 1407 goto fallback; 1408 } 1409 1410 len = end_seq - TCP_SKB_CB(skb)->seq; 1411 BUG_ON(len < 0); 1412 BUG_ON(len > skb->len); 1413 1414 /* MSS boundaries should be honoured or else pcount will 1415 * severely break even though it makes things bit trickier. 1416 * Optimize common case to avoid most of the divides 1417 */ 1418 mss = tcp_skb_mss(skb); 1419 1420 /* TODO: Fix DSACKs to not fragment already SACKed and we can 1421 * drop this restriction as unnecessary 1422 */ 1423 if (mss != tcp_skb_seglen(prev)) 1424 goto fallback; 1425 1426 if (len == mss) { 1427 pcount = 1; 1428 } else if (len < mss) { 1429 goto noop; 1430 } else { 1431 pcount = len / mss; 1432 len = pcount * mss; 1433 } 1434 } 1435 1436 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */ 1437 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una)) 1438 goto fallback; 1439 1440 if (!skb_shift(prev, skb, len)) 1441 goto fallback; 1442 if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack)) 1443 goto out; 1444 1445 /* Hole filled allows collapsing with the next as well, this is very 1446 * useful when hole on every nth skb pattern happens 1447 */ 1448 if (prev == tcp_write_queue_tail(sk)) 1449 goto out; 1450 skb = tcp_write_queue_next(sk, prev); 1451 1452 if (!skb_can_shift(skb) || 1453 (skb == tcp_send_head(sk)) || 1454 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) || 1455 (mss != tcp_skb_seglen(skb))) 1456 goto out; 1457 1458 len = skb->len; 1459 if (skb_shift(prev, skb, len)) { 1460 pcount += tcp_skb_pcount(skb); 1461 tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0); 1462 } 1463 1464 out: 1465 state->fack_count += pcount; 1466 return prev; 1467 1468 noop: 1469 return skb; 1470 1471 fallback: 1472 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK); 1473 return NULL; 1474 } 1475 1476 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk, 1477 struct tcp_sack_block *next_dup, 1478 struct tcp_sacktag_state *state, 1479 u32 start_seq, u32 end_seq, 1480 bool dup_sack_in) 1481 { 1482 struct tcp_sock *tp = tcp_sk(sk); 1483 struct sk_buff *tmp; 1484 1485 tcp_for_write_queue_from(skb, sk) { 1486 int in_sack = 0; 1487 bool dup_sack = dup_sack_in; 1488 1489 if (skb == tcp_send_head(sk)) 1490 break; 1491 1492 /* queue is in-order => we can short-circuit the walk early */ 1493 if (!before(TCP_SKB_CB(skb)->seq, end_seq)) 1494 break; 1495 1496 if (next_dup && 1497 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) { 1498 in_sack = tcp_match_skb_to_sack(sk, skb, 1499 next_dup->start_seq, 1500 next_dup->end_seq); 1501 if (in_sack > 0) 1502 dup_sack = true; 1503 } 1504 1505 /* skb reference here is a bit tricky to get right, since 1506 * shifting can eat and free both this skb and the next, 1507 * so not even _safe variant of the loop is enough. 1508 */ 1509 if (in_sack <= 0) { 1510 tmp = tcp_shift_skb_data(sk, skb, state, 1511 start_seq, end_seq, dup_sack); 1512 if (tmp) { 1513 if (tmp != skb) { 1514 skb = tmp; 1515 continue; 1516 } 1517 1518 in_sack = 0; 1519 } else { 1520 in_sack = tcp_match_skb_to_sack(sk, skb, 1521 start_seq, 1522 end_seq); 1523 } 1524 } 1525 1526 if (unlikely(in_sack < 0)) 1527 break; 1528 1529 if (in_sack) { 1530 TCP_SKB_CB(skb)->sacked = 1531 tcp_sacktag_one(sk, 1532 state, 1533 TCP_SKB_CB(skb)->sacked, 1534 TCP_SKB_CB(skb)->seq, 1535 TCP_SKB_CB(skb)->end_seq, 1536 dup_sack, 1537 tcp_skb_pcount(skb), 1538 &skb->skb_mstamp); 1539 1540 if (!before(TCP_SKB_CB(skb)->seq, 1541 tcp_highest_sack_seq(tp))) 1542 tcp_advance_highest_sack(sk, skb); 1543 } 1544 1545 state->fack_count += tcp_skb_pcount(skb); 1546 } 1547 return skb; 1548 } 1549 1550 /* Avoid all extra work that is being done by sacktag while walking in 1551 * a normal way 1552 */ 1553 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk, 1554 struct tcp_sacktag_state *state, 1555 u32 skip_to_seq) 1556 { 1557 tcp_for_write_queue_from(skb, sk) { 1558 if (skb == tcp_send_head(sk)) 1559 break; 1560 1561 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq)) 1562 break; 1563 1564 state->fack_count += tcp_skb_pcount(skb); 1565 } 1566 return skb; 1567 } 1568 1569 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb, 1570 struct sock *sk, 1571 struct tcp_sack_block *next_dup, 1572 struct tcp_sacktag_state *state, 1573 u32 skip_to_seq) 1574 { 1575 if (!next_dup) 1576 return skb; 1577 1578 if (before(next_dup->start_seq, skip_to_seq)) { 1579 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq); 1580 skb = tcp_sacktag_walk(skb, sk, NULL, state, 1581 next_dup->start_seq, next_dup->end_seq, 1582 1); 1583 } 1584 1585 return skb; 1586 } 1587 1588 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache) 1589 { 1590 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache); 1591 } 1592 1593 static int 1594 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb, 1595 u32 prior_snd_una, struct tcp_sacktag_state *state) 1596 { 1597 struct tcp_sock *tp = tcp_sk(sk); 1598 const unsigned char *ptr = (skb_transport_header(ack_skb) + 1599 TCP_SKB_CB(ack_skb)->sacked); 1600 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2); 1601 struct tcp_sack_block sp[TCP_NUM_SACKS]; 1602 struct tcp_sack_block *cache; 1603 struct sk_buff *skb; 1604 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3); 1605 int used_sacks; 1606 bool found_dup_sack = false; 1607 int i, j; 1608 int first_sack_index; 1609 1610 state->flag = 0; 1611 state->reord = tp->packets_out; 1612 1613 if (!tp->sacked_out) { 1614 if (WARN_ON(tp->fackets_out)) 1615 tp->fackets_out = 0; 1616 tcp_highest_sack_reset(sk); 1617 } 1618 1619 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire, 1620 num_sacks, prior_snd_una); 1621 if (found_dup_sack) 1622 state->flag |= FLAG_DSACKING_ACK; 1623 1624 /* Eliminate too old ACKs, but take into 1625 * account more or less fresh ones, they can 1626 * contain valid SACK info. 1627 */ 1628 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window)) 1629 return 0; 1630 1631 if (!tp->packets_out) 1632 goto out; 1633 1634 used_sacks = 0; 1635 first_sack_index = 0; 1636 for (i = 0; i < num_sacks; i++) { 1637 bool dup_sack = !i && found_dup_sack; 1638 1639 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq); 1640 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq); 1641 1642 if (!tcp_is_sackblock_valid(tp, dup_sack, 1643 sp[used_sacks].start_seq, 1644 sp[used_sacks].end_seq)) { 1645 int mib_idx; 1646 1647 if (dup_sack) { 1648 if (!tp->undo_marker) 1649 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO; 1650 else 1651 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD; 1652 } else { 1653 /* Don't count olds caused by ACK reordering */ 1654 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) && 1655 !after(sp[used_sacks].end_seq, tp->snd_una)) 1656 continue; 1657 mib_idx = LINUX_MIB_TCPSACKDISCARD; 1658 } 1659 1660 NET_INC_STATS_BH(sock_net(sk), mib_idx); 1661 if (i == 0) 1662 first_sack_index = -1; 1663 continue; 1664 } 1665 1666 /* Ignore very old stuff early */ 1667 if (!after(sp[used_sacks].end_seq, prior_snd_una)) 1668 continue; 1669 1670 used_sacks++; 1671 } 1672 1673 /* order SACK blocks to allow in order walk of the retrans queue */ 1674 for (i = used_sacks - 1; i > 0; i--) { 1675 for (j = 0; j < i; j++) { 1676 if (after(sp[j].start_seq, sp[j + 1].start_seq)) { 1677 swap(sp[j], sp[j + 1]); 1678 1679 /* Track where the first SACK block goes to */ 1680 if (j == first_sack_index) 1681 first_sack_index = j + 1; 1682 } 1683 } 1684 } 1685 1686 skb = tcp_write_queue_head(sk); 1687 state->fack_count = 0; 1688 i = 0; 1689 1690 if (!tp->sacked_out) { 1691 /* It's already past, so skip checking against it */ 1692 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache); 1693 } else { 1694 cache = tp->recv_sack_cache; 1695 /* Skip empty blocks in at head of the cache */ 1696 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq && 1697 !cache->end_seq) 1698 cache++; 1699 } 1700 1701 while (i < used_sacks) { 1702 u32 start_seq = sp[i].start_seq; 1703 u32 end_seq = sp[i].end_seq; 1704 bool dup_sack = (found_dup_sack && (i == first_sack_index)); 1705 struct tcp_sack_block *next_dup = NULL; 1706 1707 if (found_dup_sack && ((i + 1) == first_sack_index)) 1708 next_dup = &sp[i + 1]; 1709 1710 /* Skip too early cached blocks */ 1711 while (tcp_sack_cache_ok(tp, cache) && 1712 !before(start_seq, cache->end_seq)) 1713 cache++; 1714 1715 /* Can skip some work by looking recv_sack_cache? */ 1716 if (tcp_sack_cache_ok(tp, cache) && !dup_sack && 1717 after(end_seq, cache->start_seq)) { 1718 1719 /* Head todo? */ 1720 if (before(start_seq, cache->start_seq)) { 1721 skb = tcp_sacktag_skip(skb, sk, state, 1722 start_seq); 1723 skb = tcp_sacktag_walk(skb, sk, next_dup, 1724 state, 1725 start_seq, 1726 cache->start_seq, 1727 dup_sack); 1728 } 1729 1730 /* Rest of the block already fully processed? */ 1731 if (!after(end_seq, cache->end_seq)) 1732 goto advance_sp; 1733 1734 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup, 1735 state, 1736 cache->end_seq); 1737 1738 /* ...tail remains todo... */ 1739 if (tcp_highest_sack_seq(tp) == cache->end_seq) { 1740 /* ...but better entrypoint exists! */ 1741 skb = tcp_highest_sack(sk); 1742 if (!skb) 1743 break; 1744 state->fack_count = tp->fackets_out; 1745 cache++; 1746 goto walk; 1747 } 1748 1749 skb = tcp_sacktag_skip(skb, sk, state, cache->end_seq); 1750 /* Check overlap against next cached too (past this one already) */ 1751 cache++; 1752 continue; 1753 } 1754 1755 if (!before(start_seq, tcp_highest_sack_seq(tp))) { 1756 skb = tcp_highest_sack(sk); 1757 if (!skb) 1758 break; 1759 state->fack_count = tp->fackets_out; 1760 } 1761 skb = tcp_sacktag_skip(skb, sk, state, start_seq); 1762 1763 walk: 1764 skb = tcp_sacktag_walk(skb, sk, next_dup, state, 1765 start_seq, end_seq, dup_sack); 1766 1767 advance_sp: 1768 i++; 1769 } 1770 1771 /* Clear the head of the cache sack blocks so we can skip it next time */ 1772 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) { 1773 tp->recv_sack_cache[i].start_seq = 0; 1774 tp->recv_sack_cache[i].end_seq = 0; 1775 } 1776 for (j = 0; j < used_sacks; j++) 1777 tp->recv_sack_cache[i++] = sp[j]; 1778 1779 if ((state->reord < tp->fackets_out) && 1780 ((inet_csk(sk)->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker)) 1781 tcp_update_reordering(sk, tp->fackets_out - state->reord, 0); 1782 1783 tcp_verify_left_out(tp); 1784 out: 1785 1786 #if FASTRETRANS_DEBUG > 0 1787 WARN_ON((int)tp->sacked_out < 0); 1788 WARN_ON((int)tp->lost_out < 0); 1789 WARN_ON((int)tp->retrans_out < 0); 1790 WARN_ON((int)tcp_packets_in_flight(tp) < 0); 1791 #endif 1792 return state->flag; 1793 } 1794 1795 /* Limits sacked_out so that sum with lost_out isn't ever larger than 1796 * packets_out. Returns false if sacked_out adjustement wasn't necessary. 1797 */ 1798 static bool tcp_limit_reno_sacked(struct tcp_sock *tp) 1799 { 1800 u32 holes; 1801 1802 holes = max(tp->lost_out, 1U); 1803 holes = min(holes, tp->packets_out); 1804 1805 if ((tp->sacked_out + holes) > tp->packets_out) { 1806 tp->sacked_out = tp->packets_out - holes; 1807 return true; 1808 } 1809 return false; 1810 } 1811 1812 /* If we receive more dupacks than we expected counting segments 1813 * in assumption of absent reordering, interpret this as reordering. 1814 * The only another reason could be bug in receiver TCP. 1815 */ 1816 static void tcp_check_reno_reordering(struct sock *sk, const int addend) 1817 { 1818 struct tcp_sock *tp = tcp_sk(sk); 1819 if (tcp_limit_reno_sacked(tp)) 1820 tcp_update_reordering(sk, tp->packets_out + addend, 0); 1821 } 1822 1823 /* Emulate SACKs for SACKless connection: account for a new dupack. */ 1824 1825 static void tcp_add_reno_sack(struct sock *sk) 1826 { 1827 struct tcp_sock *tp = tcp_sk(sk); 1828 u32 prior_sacked = tp->sacked_out; 1829 1830 tp->sacked_out++; 1831 tcp_check_reno_reordering(sk, 0); 1832 if (tp->sacked_out > prior_sacked) 1833 tp->delivered++; /* Some out-of-order packet is delivered */ 1834 tcp_verify_left_out(tp); 1835 } 1836 1837 /* Account for ACK, ACKing some data in Reno Recovery phase. */ 1838 1839 static void tcp_remove_reno_sacks(struct sock *sk, int acked) 1840 { 1841 struct tcp_sock *tp = tcp_sk(sk); 1842 1843 if (acked > 0) { 1844 /* One ACK acked hole. The rest eat duplicate ACKs. */ 1845 tp->delivered += max_t(int, acked - tp->sacked_out, 1); 1846 if (acked - 1 >= tp->sacked_out) 1847 tp->sacked_out = 0; 1848 else 1849 tp->sacked_out -= acked - 1; 1850 } 1851 tcp_check_reno_reordering(sk, acked); 1852 tcp_verify_left_out(tp); 1853 } 1854 1855 static inline void tcp_reset_reno_sack(struct tcp_sock *tp) 1856 { 1857 tp->sacked_out = 0; 1858 } 1859 1860 void tcp_clear_retrans(struct tcp_sock *tp) 1861 { 1862 tp->retrans_out = 0; 1863 tp->lost_out = 0; 1864 tp->undo_marker = 0; 1865 tp->undo_retrans = -1; 1866 tp->fackets_out = 0; 1867 tp->sacked_out = 0; 1868 } 1869 1870 static inline void tcp_init_undo(struct tcp_sock *tp) 1871 { 1872 tp->undo_marker = tp->snd_una; 1873 /* Retransmission still in flight may cause DSACKs later. */ 1874 tp->undo_retrans = tp->retrans_out ? : -1; 1875 } 1876 1877 /* Enter Loss state. If we detect SACK reneging, forget all SACK information 1878 * and reset tags completely, otherwise preserve SACKs. If receiver 1879 * dropped its ofo queue, we will know this due to reneging detection. 1880 */ 1881 void tcp_enter_loss(struct sock *sk) 1882 { 1883 const struct inet_connection_sock *icsk = inet_csk(sk); 1884 struct tcp_sock *tp = tcp_sk(sk); 1885 struct net *net = sock_net(sk); 1886 struct sk_buff *skb; 1887 bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery; 1888 bool is_reneg; /* is receiver reneging on SACKs? */ 1889 1890 /* Reduce ssthresh if it has not yet been made inside this window. */ 1891 if (icsk->icsk_ca_state <= TCP_CA_Disorder || 1892 !after(tp->high_seq, tp->snd_una) || 1893 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) { 1894 tp->prior_ssthresh = tcp_current_ssthresh(sk); 1895 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 1896 tcp_ca_event(sk, CA_EVENT_LOSS); 1897 tcp_init_undo(tp); 1898 } 1899 tp->snd_cwnd = 1; 1900 tp->snd_cwnd_cnt = 0; 1901 tp->snd_cwnd_stamp = tcp_time_stamp; 1902 1903 tp->retrans_out = 0; 1904 tp->lost_out = 0; 1905 1906 if (tcp_is_reno(tp)) 1907 tcp_reset_reno_sack(tp); 1908 1909 skb = tcp_write_queue_head(sk); 1910 is_reneg = skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED); 1911 if (is_reneg) { 1912 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING); 1913 tp->sacked_out = 0; 1914 tp->fackets_out = 0; 1915 } 1916 tcp_clear_all_retrans_hints(tp); 1917 1918 tcp_for_write_queue(skb, sk) { 1919 if (skb == tcp_send_head(sk)) 1920 break; 1921 1922 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED; 1923 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || is_reneg) { 1924 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED; 1925 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 1926 tp->lost_out += tcp_skb_pcount(skb); 1927 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq; 1928 } 1929 } 1930 tcp_verify_left_out(tp); 1931 1932 /* Timeout in disordered state after receiving substantial DUPACKs 1933 * suggests that the degree of reordering is over-estimated. 1934 */ 1935 if (icsk->icsk_ca_state <= TCP_CA_Disorder && 1936 tp->sacked_out >= net->ipv4.sysctl_tcp_reordering) 1937 tp->reordering = min_t(unsigned int, tp->reordering, 1938 net->ipv4.sysctl_tcp_reordering); 1939 tcp_set_ca_state(sk, TCP_CA_Loss); 1940 tp->high_seq = tp->snd_nxt; 1941 tcp_ecn_queue_cwr(tp); 1942 1943 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous 1944 * loss recovery is underway except recurring timeout(s) on 1945 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing 1946 */ 1947 tp->frto = sysctl_tcp_frto && 1948 (new_recovery || icsk->icsk_retransmits) && 1949 !inet_csk(sk)->icsk_mtup.probe_size; 1950 } 1951 1952 /* If ACK arrived pointing to a remembered SACK, it means that our 1953 * remembered SACKs do not reflect real state of receiver i.e. 1954 * receiver _host_ is heavily congested (or buggy). 1955 * 1956 * To avoid big spurious retransmission bursts due to transient SACK 1957 * scoreboard oddities that look like reneging, we give the receiver a 1958 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will 1959 * restore sanity to the SACK scoreboard. If the apparent reneging 1960 * persists until this RTO then we'll clear the SACK scoreboard. 1961 */ 1962 static bool tcp_check_sack_reneging(struct sock *sk, int flag) 1963 { 1964 if (flag & FLAG_SACK_RENEGING) { 1965 struct tcp_sock *tp = tcp_sk(sk); 1966 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4), 1967 msecs_to_jiffies(10)); 1968 1969 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 1970 delay, TCP_RTO_MAX); 1971 return true; 1972 } 1973 return false; 1974 } 1975 1976 static inline int tcp_fackets_out(const struct tcp_sock *tp) 1977 { 1978 return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out; 1979 } 1980 1981 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs 1982 * counter when SACK is enabled (without SACK, sacked_out is used for 1983 * that purpose). 1984 * 1985 * Instead, with FACK TCP uses fackets_out that includes both SACKed 1986 * segments up to the highest received SACK block so far and holes in 1987 * between them. 1988 * 1989 * With reordering, holes may still be in flight, so RFC3517 recovery 1990 * uses pure sacked_out (total number of SACKed segments) even though 1991 * it violates the RFC that uses duplicate ACKs, often these are equal 1992 * but when e.g. out-of-window ACKs or packet duplication occurs, 1993 * they differ. Since neither occurs due to loss, TCP should really 1994 * ignore them. 1995 */ 1996 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp) 1997 { 1998 return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1; 1999 } 2000 2001 static bool tcp_pause_early_retransmit(struct sock *sk, int flag) 2002 { 2003 struct tcp_sock *tp = tcp_sk(sk); 2004 unsigned long delay; 2005 2006 /* Delay early retransmit and entering fast recovery for 2007 * max(RTT/4, 2msec) unless ack has ECE mark, no RTT samples 2008 * available, or RTO is scheduled to fire first. 2009 */ 2010 if (sysctl_tcp_early_retrans < 2 || sysctl_tcp_early_retrans > 3 || 2011 (flag & FLAG_ECE) || !tp->srtt_us) 2012 return false; 2013 2014 delay = max(usecs_to_jiffies(tp->srtt_us >> 5), 2015 msecs_to_jiffies(2)); 2016 2017 if (!time_after(inet_csk(sk)->icsk_timeout, (jiffies + delay))) 2018 return false; 2019 2020 inet_csk_reset_xmit_timer(sk, ICSK_TIME_EARLY_RETRANS, delay, 2021 TCP_RTO_MAX); 2022 return true; 2023 } 2024 2025 /* Linux NewReno/SACK/FACK/ECN state machine. 2026 * -------------------------------------- 2027 * 2028 * "Open" Normal state, no dubious events, fast path. 2029 * "Disorder" In all the respects it is "Open", 2030 * but requires a bit more attention. It is entered when 2031 * we see some SACKs or dupacks. It is split of "Open" 2032 * mainly to move some processing from fast path to slow one. 2033 * "CWR" CWND was reduced due to some Congestion Notification event. 2034 * It can be ECN, ICMP source quench, local device congestion. 2035 * "Recovery" CWND was reduced, we are fast-retransmitting. 2036 * "Loss" CWND was reduced due to RTO timeout or SACK reneging. 2037 * 2038 * tcp_fastretrans_alert() is entered: 2039 * - each incoming ACK, if state is not "Open" 2040 * - when arrived ACK is unusual, namely: 2041 * * SACK 2042 * * Duplicate ACK. 2043 * * ECN ECE. 2044 * 2045 * Counting packets in flight is pretty simple. 2046 * 2047 * in_flight = packets_out - left_out + retrans_out 2048 * 2049 * packets_out is SND.NXT-SND.UNA counted in packets. 2050 * 2051 * retrans_out is number of retransmitted segments. 2052 * 2053 * left_out is number of segments left network, but not ACKed yet. 2054 * 2055 * left_out = sacked_out + lost_out 2056 * 2057 * sacked_out: Packets, which arrived to receiver out of order 2058 * and hence not ACKed. With SACKs this number is simply 2059 * amount of SACKed data. Even without SACKs 2060 * it is easy to give pretty reliable estimate of this number, 2061 * counting duplicate ACKs. 2062 * 2063 * lost_out: Packets lost by network. TCP has no explicit 2064 * "loss notification" feedback from network (for now). 2065 * It means that this number can be only _guessed_. 2066 * Actually, it is the heuristics to predict lossage that 2067 * distinguishes different algorithms. 2068 * 2069 * F.e. after RTO, when all the queue is considered as lost, 2070 * lost_out = packets_out and in_flight = retrans_out. 2071 * 2072 * Essentially, we have now two algorithms counting 2073 * lost packets. 2074 * 2075 * FACK: It is the simplest heuristics. As soon as we decided 2076 * that something is lost, we decide that _all_ not SACKed 2077 * packets until the most forward SACK are lost. I.e. 2078 * lost_out = fackets_out - sacked_out and left_out = fackets_out. 2079 * It is absolutely correct estimate, if network does not reorder 2080 * packets. And it loses any connection to reality when reordering 2081 * takes place. We use FACK by default until reordering 2082 * is suspected on the path to this destination. 2083 * 2084 * NewReno: when Recovery is entered, we assume that one segment 2085 * is lost (classic Reno). While we are in Recovery and 2086 * a partial ACK arrives, we assume that one more packet 2087 * is lost (NewReno). This heuristics are the same in NewReno 2088 * and SACK. 2089 * 2090 * Imagine, that's all! Forget about all this shamanism about CWND inflation 2091 * deflation etc. CWND is real congestion window, never inflated, changes 2092 * only according to classic VJ rules. 2093 * 2094 * Really tricky (and requiring careful tuning) part of algorithm 2095 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue(). 2096 * The first determines the moment _when_ we should reduce CWND and, 2097 * hence, slow down forward transmission. In fact, it determines the moment 2098 * when we decide that hole is caused by loss, rather than by a reorder. 2099 * 2100 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill 2101 * holes, caused by lost packets. 2102 * 2103 * And the most logically complicated part of algorithm is undo 2104 * heuristics. We detect false retransmits due to both too early 2105 * fast retransmit (reordering) and underestimated RTO, analyzing 2106 * timestamps and D-SACKs. When we detect that some segments were 2107 * retransmitted by mistake and CWND reduction was wrong, we undo 2108 * window reduction and abort recovery phase. This logic is hidden 2109 * inside several functions named tcp_try_undo_<something>. 2110 */ 2111 2112 /* This function decides, when we should leave Disordered state 2113 * and enter Recovery phase, reducing congestion window. 2114 * 2115 * Main question: may we further continue forward transmission 2116 * with the same cwnd? 2117 */ 2118 static bool tcp_time_to_recover(struct sock *sk, int flag) 2119 { 2120 struct tcp_sock *tp = tcp_sk(sk); 2121 __u32 packets_out; 2122 int tcp_reordering = sock_net(sk)->ipv4.sysctl_tcp_reordering; 2123 2124 /* Trick#1: The loss is proven. */ 2125 if (tp->lost_out) 2126 return true; 2127 2128 /* Not-A-Trick#2 : Classic rule... */ 2129 if (tcp_dupack_heuristics(tp) > tp->reordering) 2130 return true; 2131 2132 /* Trick#4: It is still not OK... But will it be useful to delay 2133 * recovery more? 2134 */ 2135 packets_out = tp->packets_out; 2136 if (packets_out <= tp->reordering && 2137 tp->sacked_out >= max_t(__u32, packets_out/2, tcp_reordering) && 2138 !tcp_may_send_now(sk)) { 2139 /* We have nothing to send. This connection is limited 2140 * either by receiver window or by application. 2141 */ 2142 return true; 2143 } 2144 2145 /* If a thin stream is detected, retransmit after first 2146 * received dupack. Employ only if SACK is supported in order 2147 * to avoid possible corner-case series of spurious retransmissions 2148 * Use only if there are no unsent data. 2149 */ 2150 if ((tp->thin_dupack || sysctl_tcp_thin_dupack) && 2151 tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 && 2152 tcp_is_sack(tp) && !tcp_send_head(sk)) 2153 return true; 2154 2155 /* Trick#6: TCP early retransmit, per RFC5827. To avoid spurious 2156 * retransmissions due to small network reorderings, we implement 2157 * Mitigation A.3 in the RFC and delay the retransmission for a short 2158 * interval if appropriate. 2159 */ 2160 if (tp->do_early_retrans && !tp->retrans_out && tp->sacked_out && 2161 (tp->packets_out >= (tp->sacked_out + 1) && tp->packets_out < 4) && 2162 !tcp_may_send_now(sk)) 2163 return !tcp_pause_early_retransmit(sk, flag); 2164 2165 return false; 2166 } 2167 2168 /* Detect loss in event "A" above by marking head of queue up as lost. 2169 * For FACK or non-SACK(Reno) senders, the first "packets" number of segments 2170 * are considered lost. For RFC3517 SACK, a segment is considered lost if it 2171 * has at least tp->reordering SACKed seqments above it; "packets" refers to 2172 * the maximum SACKed segments to pass before reaching this limit. 2173 */ 2174 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head) 2175 { 2176 struct tcp_sock *tp = tcp_sk(sk); 2177 struct sk_buff *skb; 2178 int cnt, oldcnt, lost; 2179 unsigned int mss; 2180 /* Use SACK to deduce losses of new sequences sent during recovery */ 2181 const u32 loss_high = tcp_is_sack(tp) ? tp->snd_nxt : tp->high_seq; 2182 2183 WARN_ON(packets > tp->packets_out); 2184 if (tp->lost_skb_hint) { 2185 skb = tp->lost_skb_hint; 2186 cnt = tp->lost_cnt_hint; 2187 /* Head already handled? */ 2188 if (mark_head && skb != tcp_write_queue_head(sk)) 2189 return; 2190 } else { 2191 skb = tcp_write_queue_head(sk); 2192 cnt = 0; 2193 } 2194 2195 tcp_for_write_queue_from(skb, sk) { 2196 if (skb == tcp_send_head(sk)) 2197 break; 2198 /* TODO: do this better */ 2199 /* this is not the most efficient way to do this... */ 2200 tp->lost_skb_hint = skb; 2201 tp->lost_cnt_hint = cnt; 2202 2203 if (after(TCP_SKB_CB(skb)->end_seq, loss_high)) 2204 break; 2205 2206 oldcnt = cnt; 2207 if (tcp_is_fack(tp) || tcp_is_reno(tp) || 2208 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) 2209 cnt += tcp_skb_pcount(skb); 2210 2211 if (cnt > packets) { 2212 if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) || 2213 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) || 2214 (oldcnt >= packets)) 2215 break; 2216 2217 mss = tcp_skb_mss(skb); 2218 /* If needed, chop off the prefix to mark as lost. */ 2219 lost = (packets - oldcnt) * mss; 2220 if (lost < skb->len && 2221 tcp_fragment(sk, skb, lost, mss, GFP_ATOMIC) < 0) 2222 break; 2223 cnt = packets; 2224 } 2225 2226 tcp_skb_mark_lost(tp, skb); 2227 2228 if (mark_head) 2229 break; 2230 } 2231 tcp_verify_left_out(tp); 2232 } 2233 2234 /* Account newly detected lost packet(s) */ 2235 2236 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit) 2237 { 2238 struct tcp_sock *tp = tcp_sk(sk); 2239 2240 if (tcp_is_reno(tp)) { 2241 tcp_mark_head_lost(sk, 1, 1); 2242 } else if (tcp_is_fack(tp)) { 2243 int lost = tp->fackets_out - tp->reordering; 2244 if (lost <= 0) 2245 lost = 1; 2246 tcp_mark_head_lost(sk, lost, 0); 2247 } else { 2248 int sacked_upto = tp->sacked_out - tp->reordering; 2249 if (sacked_upto >= 0) 2250 tcp_mark_head_lost(sk, sacked_upto, 0); 2251 else if (fast_rexmit) 2252 tcp_mark_head_lost(sk, 1, 1); 2253 } 2254 } 2255 2256 /* CWND moderation, preventing bursts due to too big ACKs 2257 * in dubious situations. 2258 */ 2259 static inline void tcp_moderate_cwnd(struct tcp_sock *tp) 2260 { 2261 tp->snd_cwnd = min(tp->snd_cwnd, 2262 tcp_packets_in_flight(tp) + tcp_max_burst(tp)); 2263 tp->snd_cwnd_stamp = tcp_time_stamp; 2264 } 2265 2266 static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when) 2267 { 2268 return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 2269 before(tp->rx_opt.rcv_tsecr, when); 2270 } 2271 2272 /* skb is spurious retransmitted if the returned timestamp echo 2273 * reply is prior to the skb transmission time 2274 */ 2275 static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp, 2276 const struct sk_buff *skb) 2277 { 2278 return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) && 2279 tcp_tsopt_ecr_before(tp, tcp_skb_timestamp(skb)); 2280 } 2281 2282 /* Nothing was retransmitted or returned timestamp is less 2283 * than timestamp of the first retransmission. 2284 */ 2285 static inline bool tcp_packet_delayed(const struct tcp_sock *tp) 2286 { 2287 return !tp->retrans_stamp || 2288 tcp_tsopt_ecr_before(tp, tp->retrans_stamp); 2289 } 2290 2291 /* Undo procedures. */ 2292 2293 /* We can clear retrans_stamp when there are no retransmissions in the 2294 * window. It would seem that it is trivially available for us in 2295 * tp->retrans_out, however, that kind of assumptions doesn't consider 2296 * what will happen if errors occur when sending retransmission for the 2297 * second time. ...It could the that such segment has only 2298 * TCPCB_EVER_RETRANS set at the present time. It seems that checking 2299 * the head skb is enough except for some reneging corner cases that 2300 * are not worth the effort. 2301 * 2302 * Main reason for all this complexity is the fact that connection dying 2303 * time now depends on the validity of the retrans_stamp, in particular, 2304 * that successive retransmissions of a segment must not advance 2305 * retrans_stamp under any conditions. 2306 */ 2307 static bool tcp_any_retrans_done(const struct sock *sk) 2308 { 2309 const struct tcp_sock *tp = tcp_sk(sk); 2310 struct sk_buff *skb; 2311 2312 if (tp->retrans_out) 2313 return true; 2314 2315 skb = tcp_write_queue_head(sk); 2316 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS)) 2317 return true; 2318 2319 return false; 2320 } 2321 2322 #if FASTRETRANS_DEBUG > 1 2323 static void DBGUNDO(struct sock *sk, const char *msg) 2324 { 2325 struct tcp_sock *tp = tcp_sk(sk); 2326 struct inet_sock *inet = inet_sk(sk); 2327 2328 if (sk->sk_family == AF_INET) { 2329 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n", 2330 msg, 2331 &inet->inet_daddr, ntohs(inet->inet_dport), 2332 tp->snd_cwnd, tcp_left_out(tp), 2333 tp->snd_ssthresh, tp->prior_ssthresh, 2334 tp->packets_out); 2335 } 2336 #if IS_ENABLED(CONFIG_IPV6) 2337 else if (sk->sk_family == AF_INET6) { 2338 struct ipv6_pinfo *np = inet6_sk(sk); 2339 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n", 2340 msg, 2341 &np->daddr, ntohs(inet->inet_dport), 2342 tp->snd_cwnd, tcp_left_out(tp), 2343 tp->snd_ssthresh, tp->prior_ssthresh, 2344 tp->packets_out); 2345 } 2346 #endif 2347 } 2348 #else 2349 #define DBGUNDO(x...) do { } while (0) 2350 #endif 2351 2352 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss) 2353 { 2354 struct tcp_sock *tp = tcp_sk(sk); 2355 2356 if (unmark_loss) { 2357 struct sk_buff *skb; 2358 2359 tcp_for_write_queue(skb, sk) { 2360 if (skb == tcp_send_head(sk)) 2361 break; 2362 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; 2363 } 2364 tp->lost_out = 0; 2365 tcp_clear_all_retrans_hints(tp); 2366 } 2367 2368 if (tp->prior_ssthresh) { 2369 const struct inet_connection_sock *icsk = inet_csk(sk); 2370 2371 if (icsk->icsk_ca_ops->undo_cwnd) 2372 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk); 2373 else 2374 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1); 2375 2376 if (tp->prior_ssthresh > tp->snd_ssthresh) { 2377 tp->snd_ssthresh = tp->prior_ssthresh; 2378 tcp_ecn_withdraw_cwr(tp); 2379 } 2380 } 2381 tp->snd_cwnd_stamp = tcp_time_stamp; 2382 tp->undo_marker = 0; 2383 } 2384 2385 static inline bool tcp_may_undo(const struct tcp_sock *tp) 2386 { 2387 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp)); 2388 } 2389 2390 /* People celebrate: "We love our President!" */ 2391 static bool tcp_try_undo_recovery(struct sock *sk) 2392 { 2393 struct tcp_sock *tp = tcp_sk(sk); 2394 2395 if (tcp_may_undo(tp)) { 2396 int mib_idx; 2397 2398 /* Happy end! We did not retransmit anything 2399 * or our original transmission succeeded. 2400 */ 2401 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans"); 2402 tcp_undo_cwnd_reduction(sk, false); 2403 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss) 2404 mib_idx = LINUX_MIB_TCPLOSSUNDO; 2405 else 2406 mib_idx = LINUX_MIB_TCPFULLUNDO; 2407 2408 NET_INC_STATS_BH(sock_net(sk), mib_idx); 2409 } 2410 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) { 2411 /* Hold old state until something *above* high_seq 2412 * is ACKed. For Reno it is MUST to prevent false 2413 * fast retransmits (RFC2582). SACK TCP is safe. */ 2414 tcp_moderate_cwnd(tp); 2415 if (!tcp_any_retrans_done(sk)) 2416 tp->retrans_stamp = 0; 2417 return true; 2418 } 2419 tcp_set_ca_state(sk, TCP_CA_Open); 2420 return false; 2421 } 2422 2423 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */ 2424 static bool tcp_try_undo_dsack(struct sock *sk) 2425 { 2426 struct tcp_sock *tp = tcp_sk(sk); 2427 2428 if (tp->undo_marker && !tp->undo_retrans) { 2429 DBGUNDO(sk, "D-SACK"); 2430 tcp_undo_cwnd_reduction(sk, false); 2431 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO); 2432 return true; 2433 } 2434 return false; 2435 } 2436 2437 /* Undo during loss recovery after partial ACK or using F-RTO. */ 2438 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo) 2439 { 2440 struct tcp_sock *tp = tcp_sk(sk); 2441 2442 if (frto_undo || tcp_may_undo(tp)) { 2443 tcp_undo_cwnd_reduction(sk, true); 2444 2445 DBGUNDO(sk, "partial loss"); 2446 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO); 2447 if (frto_undo) 2448 NET_INC_STATS_BH(sock_net(sk), 2449 LINUX_MIB_TCPSPURIOUSRTOS); 2450 inet_csk(sk)->icsk_retransmits = 0; 2451 if (frto_undo || tcp_is_sack(tp)) 2452 tcp_set_ca_state(sk, TCP_CA_Open); 2453 return true; 2454 } 2455 return false; 2456 } 2457 2458 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937. 2459 * It computes the number of packets to send (sndcnt) based on packets newly 2460 * delivered: 2461 * 1) If the packets in flight is larger than ssthresh, PRR spreads the 2462 * cwnd reductions across a full RTT. 2463 * 2) Otherwise PRR uses packet conservation to send as much as delivered. 2464 * But when the retransmits are acked without further losses, PRR 2465 * slow starts cwnd up to ssthresh to speed up the recovery. 2466 */ 2467 static void tcp_init_cwnd_reduction(struct sock *sk) 2468 { 2469 struct tcp_sock *tp = tcp_sk(sk); 2470 2471 tp->high_seq = tp->snd_nxt; 2472 tp->tlp_high_seq = 0; 2473 tp->snd_cwnd_cnt = 0; 2474 tp->prior_cwnd = tp->snd_cwnd; 2475 tp->prr_delivered = 0; 2476 tp->prr_out = 0; 2477 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk); 2478 tcp_ecn_queue_cwr(tp); 2479 } 2480 2481 static void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked, 2482 int flag) 2483 { 2484 struct tcp_sock *tp = tcp_sk(sk); 2485 int sndcnt = 0; 2486 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp); 2487 2488 if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd)) 2489 return; 2490 2491 tp->prr_delivered += newly_acked_sacked; 2492 if (delta < 0) { 2493 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered + 2494 tp->prior_cwnd - 1; 2495 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out; 2496 } else if ((flag & FLAG_RETRANS_DATA_ACKED) && 2497 !(flag & FLAG_LOST_RETRANS)) { 2498 sndcnt = min_t(int, delta, 2499 max_t(int, tp->prr_delivered - tp->prr_out, 2500 newly_acked_sacked) + 1); 2501 } else { 2502 sndcnt = min(delta, newly_acked_sacked); 2503 } 2504 /* Force a fast retransmit upon entering fast recovery */ 2505 sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1)); 2506 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt; 2507 } 2508 2509 static inline void tcp_end_cwnd_reduction(struct sock *sk) 2510 { 2511 struct tcp_sock *tp = tcp_sk(sk); 2512 2513 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */ 2514 if (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR || 2515 (tp->undo_marker && tp->snd_ssthresh < TCP_INFINITE_SSTHRESH)) { 2516 tp->snd_cwnd = tp->snd_ssthresh; 2517 tp->snd_cwnd_stamp = tcp_time_stamp; 2518 } 2519 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR); 2520 } 2521 2522 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */ 2523 void tcp_enter_cwr(struct sock *sk) 2524 { 2525 struct tcp_sock *tp = tcp_sk(sk); 2526 2527 tp->prior_ssthresh = 0; 2528 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) { 2529 tp->undo_marker = 0; 2530 tcp_init_cwnd_reduction(sk); 2531 tcp_set_ca_state(sk, TCP_CA_CWR); 2532 } 2533 } 2534 EXPORT_SYMBOL(tcp_enter_cwr); 2535 2536 static void tcp_try_keep_open(struct sock *sk) 2537 { 2538 struct tcp_sock *tp = tcp_sk(sk); 2539 int state = TCP_CA_Open; 2540 2541 if (tcp_left_out(tp) || tcp_any_retrans_done(sk)) 2542 state = TCP_CA_Disorder; 2543 2544 if (inet_csk(sk)->icsk_ca_state != state) { 2545 tcp_set_ca_state(sk, state); 2546 tp->high_seq = tp->snd_nxt; 2547 } 2548 } 2549 2550 static void tcp_try_to_open(struct sock *sk, int flag) 2551 { 2552 struct tcp_sock *tp = tcp_sk(sk); 2553 2554 tcp_verify_left_out(tp); 2555 2556 if (!tcp_any_retrans_done(sk)) 2557 tp->retrans_stamp = 0; 2558 2559 if (flag & FLAG_ECE) 2560 tcp_enter_cwr(sk); 2561 2562 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) { 2563 tcp_try_keep_open(sk); 2564 } 2565 } 2566 2567 static void tcp_mtup_probe_failed(struct sock *sk) 2568 { 2569 struct inet_connection_sock *icsk = inet_csk(sk); 2570 2571 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1; 2572 icsk->icsk_mtup.probe_size = 0; 2573 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPMTUPFAIL); 2574 } 2575 2576 static void tcp_mtup_probe_success(struct sock *sk) 2577 { 2578 struct tcp_sock *tp = tcp_sk(sk); 2579 struct inet_connection_sock *icsk = inet_csk(sk); 2580 2581 /* FIXME: breaks with very large cwnd */ 2582 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2583 tp->snd_cwnd = tp->snd_cwnd * 2584 tcp_mss_to_mtu(sk, tp->mss_cache) / 2585 icsk->icsk_mtup.probe_size; 2586 tp->snd_cwnd_cnt = 0; 2587 tp->snd_cwnd_stamp = tcp_time_stamp; 2588 tp->snd_ssthresh = tcp_current_ssthresh(sk); 2589 2590 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size; 2591 icsk->icsk_mtup.probe_size = 0; 2592 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 2593 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS); 2594 } 2595 2596 /* Do a simple retransmit without using the backoff mechanisms in 2597 * tcp_timer. This is used for path mtu discovery. 2598 * The socket is already locked here. 2599 */ 2600 void tcp_simple_retransmit(struct sock *sk) 2601 { 2602 const struct inet_connection_sock *icsk = inet_csk(sk); 2603 struct tcp_sock *tp = tcp_sk(sk); 2604 struct sk_buff *skb; 2605 unsigned int mss = tcp_current_mss(sk); 2606 u32 prior_lost = tp->lost_out; 2607 2608 tcp_for_write_queue(skb, sk) { 2609 if (skb == tcp_send_head(sk)) 2610 break; 2611 if (tcp_skb_seglen(skb) > mss && 2612 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) { 2613 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 2614 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 2615 tp->retrans_out -= tcp_skb_pcount(skb); 2616 } 2617 tcp_skb_mark_lost_uncond_verify(tp, skb); 2618 } 2619 } 2620 2621 tcp_clear_retrans_hints_partial(tp); 2622 2623 if (prior_lost == tp->lost_out) 2624 return; 2625 2626 if (tcp_is_reno(tp)) 2627 tcp_limit_reno_sacked(tp); 2628 2629 tcp_verify_left_out(tp); 2630 2631 /* Don't muck with the congestion window here. 2632 * Reason is that we do not increase amount of _data_ 2633 * in network, but units changed and effective 2634 * cwnd/ssthresh really reduced now. 2635 */ 2636 if (icsk->icsk_ca_state != TCP_CA_Loss) { 2637 tp->high_seq = tp->snd_nxt; 2638 tp->snd_ssthresh = tcp_current_ssthresh(sk); 2639 tp->prior_ssthresh = 0; 2640 tp->undo_marker = 0; 2641 tcp_set_ca_state(sk, TCP_CA_Loss); 2642 } 2643 tcp_xmit_retransmit_queue(sk); 2644 } 2645 EXPORT_SYMBOL(tcp_simple_retransmit); 2646 2647 static void tcp_enter_recovery(struct sock *sk, bool ece_ack) 2648 { 2649 struct tcp_sock *tp = tcp_sk(sk); 2650 int mib_idx; 2651 2652 if (tcp_is_reno(tp)) 2653 mib_idx = LINUX_MIB_TCPRENORECOVERY; 2654 else 2655 mib_idx = LINUX_MIB_TCPSACKRECOVERY; 2656 2657 NET_INC_STATS_BH(sock_net(sk), mib_idx); 2658 2659 tp->prior_ssthresh = 0; 2660 tcp_init_undo(tp); 2661 2662 if (!tcp_in_cwnd_reduction(sk)) { 2663 if (!ece_ack) 2664 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2665 tcp_init_cwnd_reduction(sk); 2666 } 2667 tcp_set_ca_state(sk, TCP_CA_Recovery); 2668 } 2669 2670 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are 2671 * recovered or spurious. Otherwise retransmits more on partial ACKs. 2672 */ 2673 static void tcp_process_loss(struct sock *sk, int flag, bool is_dupack, 2674 int *rexmit) 2675 { 2676 struct tcp_sock *tp = tcp_sk(sk); 2677 bool recovered = !before(tp->snd_una, tp->high_seq); 2678 2679 if ((flag & FLAG_SND_UNA_ADVANCED) && 2680 tcp_try_undo_loss(sk, false)) 2681 return; 2682 2683 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */ 2684 /* Step 3.b. A timeout is spurious if not all data are 2685 * lost, i.e., never-retransmitted data are (s)acked. 2686 */ 2687 if ((flag & FLAG_ORIG_SACK_ACKED) && 2688 tcp_try_undo_loss(sk, true)) 2689 return; 2690 2691 if (after(tp->snd_nxt, tp->high_seq)) { 2692 if (flag & FLAG_DATA_SACKED || is_dupack) 2693 tp->frto = 0; /* Step 3.a. loss was real */ 2694 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) { 2695 tp->high_seq = tp->snd_nxt; 2696 /* Step 2.b. Try send new data (but deferred until cwnd 2697 * is updated in tcp_ack()). Otherwise fall back to 2698 * the conventional recovery. 2699 */ 2700 if (tcp_send_head(sk) && 2701 after(tcp_wnd_end(tp), tp->snd_nxt)) { 2702 *rexmit = REXMIT_NEW; 2703 return; 2704 } 2705 tp->frto = 0; 2706 } 2707 } 2708 2709 if (recovered) { 2710 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */ 2711 tcp_try_undo_recovery(sk); 2712 return; 2713 } 2714 if (tcp_is_reno(tp)) { 2715 /* A Reno DUPACK means new data in F-RTO step 2.b above are 2716 * delivered. Lower inflight to clock out (re)tranmissions. 2717 */ 2718 if (after(tp->snd_nxt, tp->high_seq) && is_dupack) 2719 tcp_add_reno_sack(sk); 2720 else if (flag & FLAG_SND_UNA_ADVANCED) 2721 tcp_reset_reno_sack(tp); 2722 } 2723 *rexmit = REXMIT_LOST; 2724 } 2725 2726 /* Undo during fast recovery after partial ACK. */ 2727 static bool tcp_try_undo_partial(struct sock *sk, const int acked) 2728 { 2729 struct tcp_sock *tp = tcp_sk(sk); 2730 2731 if (tp->undo_marker && tcp_packet_delayed(tp)) { 2732 /* Plain luck! Hole if filled with delayed 2733 * packet, rather than with a retransmit. 2734 */ 2735 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1); 2736 2737 /* We are getting evidence that the reordering degree is higher 2738 * than we realized. If there are no retransmits out then we 2739 * can undo. Otherwise we clock out new packets but do not 2740 * mark more packets lost or retransmit more. 2741 */ 2742 if (tp->retrans_out) 2743 return true; 2744 2745 if (!tcp_any_retrans_done(sk)) 2746 tp->retrans_stamp = 0; 2747 2748 DBGUNDO(sk, "partial recovery"); 2749 tcp_undo_cwnd_reduction(sk, true); 2750 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO); 2751 tcp_try_keep_open(sk); 2752 return true; 2753 } 2754 return false; 2755 } 2756 2757 /* Process an event, which can update packets-in-flight not trivially. 2758 * Main goal of this function is to calculate new estimate for left_out, 2759 * taking into account both packets sitting in receiver's buffer and 2760 * packets lost by network. 2761 * 2762 * Besides that it updates the congestion state when packet loss or ECN 2763 * is detected. But it does not reduce the cwnd, it is done by the 2764 * congestion control later. 2765 * 2766 * It does _not_ decide what to send, it is made in function 2767 * tcp_xmit_retransmit_queue(). 2768 */ 2769 static void tcp_fastretrans_alert(struct sock *sk, const int acked, 2770 bool is_dupack, int *ack_flag, int *rexmit) 2771 { 2772 struct inet_connection_sock *icsk = inet_csk(sk); 2773 struct tcp_sock *tp = tcp_sk(sk); 2774 int fast_rexmit = 0, flag = *ack_flag; 2775 bool do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) && 2776 (tcp_fackets_out(tp) > tp->reordering)); 2777 2778 if (WARN_ON(!tp->packets_out && tp->sacked_out)) 2779 tp->sacked_out = 0; 2780 if (WARN_ON(!tp->sacked_out && tp->fackets_out)) 2781 tp->fackets_out = 0; 2782 2783 /* Now state machine starts. 2784 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */ 2785 if (flag & FLAG_ECE) 2786 tp->prior_ssthresh = 0; 2787 2788 /* B. In all the states check for reneging SACKs. */ 2789 if (tcp_check_sack_reneging(sk, flag)) 2790 return; 2791 2792 /* C. Check consistency of the current state. */ 2793 tcp_verify_left_out(tp); 2794 2795 /* D. Check state exit conditions. State can be terminated 2796 * when high_seq is ACKed. */ 2797 if (icsk->icsk_ca_state == TCP_CA_Open) { 2798 WARN_ON(tp->retrans_out != 0); 2799 tp->retrans_stamp = 0; 2800 } else if (!before(tp->snd_una, tp->high_seq)) { 2801 switch (icsk->icsk_ca_state) { 2802 case TCP_CA_CWR: 2803 /* CWR is to be held something *above* high_seq 2804 * is ACKed for CWR bit to reach receiver. */ 2805 if (tp->snd_una != tp->high_seq) { 2806 tcp_end_cwnd_reduction(sk); 2807 tcp_set_ca_state(sk, TCP_CA_Open); 2808 } 2809 break; 2810 2811 case TCP_CA_Recovery: 2812 if (tcp_is_reno(tp)) 2813 tcp_reset_reno_sack(tp); 2814 if (tcp_try_undo_recovery(sk)) 2815 return; 2816 tcp_end_cwnd_reduction(sk); 2817 break; 2818 } 2819 } 2820 2821 /* Use RACK to detect loss */ 2822 if (sysctl_tcp_recovery & TCP_RACK_LOST_RETRANS && 2823 tcp_rack_mark_lost(sk)) { 2824 flag |= FLAG_LOST_RETRANS; 2825 *ack_flag |= FLAG_LOST_RETRANS; 2826 } 2827 2828 /* E. Process state. */ 2829 switch (icsk->icsk_ca_state) { 2830 case TCP_CA_Recovery: 2831 if (!(flag & FLAG_SND_UNA_ADVANCED)) { 2832 if (tcp_is_reno(tp) && is_dupack) 2833 tcp_add_reno_sack(sk); 2834 } else { 2835 if (tcp_try_undo_partial(sk, acked)) 2836 return; 2837 /* Partial ACK arrived. Force fast retransmit. */ 2838 do_lost = tcp_is_reno(tp) || 2839 tcp_fackets_out(tp) > tp->reordering; 2840 } 2841 if (tcp_try_undo_dsack(sk)) { 2842 tcp_try_keep_open(sk); 2843 return; 2844 } 2845 break; 2846 case TCP_CA_Loss: 2847 tcp_process_loss(sk, flag, is_dupack, rexmit); 2848 if (icsk->icsk_ca_state != TCP_CA_Open && 2849 !(flag & FLAG_LOST_RETRANS)) 2850 return; 2851 /* Change state if cwnd is undone or retransmits are lost */ 2852 default: 2853 if (tcp_is_reno(tp)) { 2854 if (flag & FLAG_SND_UNA_ADVANCED) 2855 tcp_reset_reno_sack(tp); 2856 if (is_dupack) 2857 tcp_add_reno_sack(sk); 2858 } 2859 2860 if (icsk->icsk_ca_state <= TCP_CA_Disorder) 2861 tcp_try_undo_dsack(sk); 2862 2863 if (!tcp_time_to_recover(sk, flag)) { 2864 tcp_try_to_open(sk, flag); 2865 return; 2866 } 2867 2868 /* MTU probe failure: don't reduce cwnd */ 2869 if (icsk->icsk_ca_state < TCP_CA_CWR && 2870 icsk->icsk_mtup.probe_size && 2871 tp->snd_una == tp->mtu_probe.probe_seq_start) { 2872 tcp_mtup_probe_failed(sk); 2873 /* Restores the reduction we did in tcp_mtup_probe() */ 2874 tp->snd_cwnd++; 2875 tcp_simple_retransmit(sk); 2876 return; 2877 } 2878 2879 /* Otherwise enter Recovery state */ 2880 tcp_enter_recovery(sk, (flag & FLAG_ECE)); 2881 fast_rexmit = 1; 2882 } 2883 2884 if (do_lost) 2885 tcp_update_scoreboard(sk, fast_rexmit); 2886 *rexmit = REXMIT_LOST; 2887 } 2888 2889 /* Kathleen Nichols' algorithm for tracking the minimum value of 2890 * a data stream over some fixed time interval. (E.g., the minimum 2891 * RTT over the past five minutes.) It uses constant space and constant 2892 * time per update yet almost always delivers the same minimum as an 2893 * implementation that has to keep all the data in the window. 2894 * 2895 * The algorithm keeps track of the best, 2nd best & 3rd best min 2896 * values, maintaining an invariant that the measurement time of the 2897 * n'th best >= n-1'th best. It also makes sure that the three values 2898 * are widely separated in the time window since that bounds the worse 2899 * case error when that data is monotonically increasing over the window. 2900 * 2901 * Upon getting a new min, we can forget everything earlier because it 2902 * has no value - the new min is <= everything else in the window by 2903 * definition and it's the most recent. So we restart fresh on every new min 2904 * and overwrites 2nd & 3rd choices. The same property holds for 2nd & 3rd 2905 * best. 2906 */ 2907 static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us) 2908 { 2909 const u32 now = tcp_time_stamp, wlen = sysctl_tcp_min_rtt_wlen * HZ; 2910 struct rtt_meas *m = tcp_sk(sk)->rtt_min; 2911 struct rtt_meas rttm = { 2912 .rtt = likely(rtt_us) ? rtt_us : jiffies_to_usecs(1), 2913 .ts = now, 2914 }; 2915 u32 elapsed; 2916 2917 /* Check if the new measurement updates the 1st, 2nd, or 3rd choices */ 2918 if (unlikely(rttm.rtt <= m[0].rtt)) 2919 m[0] = m[1] = m[2] = rttm; 2920 else if (rttm.rtt <= m[1].rtt) 2921 m[1] = m[2] = rttm; 2922 else if (rttm.rtt <= m[2].rtt) 2923 m[2] = rttm; 2924 2925 elapsed = now - m[0].ts; 2926 if (unlikely(elapsed > wlen)) { 2927 /* Passed entire window without a new min so make 2nd choice 2928 * the new min & 3rd choice the new 2nd. So forth and so on. 2929 */ 2930 m[0] = m[1]; 2931 m[1] = m[2]; 2932 m[2] = rttm; 2933 if (now - m[0].ts > wlen) { 2934 m[0] = m[1]; 2935 m[1] = rttm; 2936 if (now - m[0].ts > wlen) 2937 m[0] = rttm; 2938 } 2939 } else if (m[1].ts == m[0].ts && elapsed > wlen / 4) { 2940 /* Passed a quarter of the window without a new min so 2941 * take 2nd choice from the 2nd quarter of the window. 2942 */ 2943 m[2] = m[1] = rttm; 2944 } else if (m[2].ts == m[1].ts && elapsed > wlen / 2) { 2945 /* Passed half the window without a new min so take the 3rd 2946 * choice from the last half of the window. 2947 */ 2948 m[2] = rttm; 2949 } 2950 } 2951 2952 static inline bool tcp_ack_update_rtt(struct sock *sk, const int flag, 2953 long seq_rtt_us, long sack_rtt_us, 2954 long ca_rtt_us) 2955 { 2956 const struct tcp_sock *tp = tcp_sk(sk); 2957 2958 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because 2959 * broken middle-boxes or peers may corrupt TS-ECR fields. But 2960 * Karn's algorithm forbids taking RTT if some retransmitted data 2961 * is acked (RFC6298). 2962 */ 2963 if (seq_rtt_us < 0) 2964 seq_rtt_us = sack_rtt_us; 2965 2966 /* RTTM Rule: A TSecr value received in a segment is used to 2967 * update the averaged RTT measurement only if the segment 2968 * acknowledges some new data, i.e., only if it advances the 2969 * left edge of the send window. 2970 * See draft-ietf-tcplw-high-performance-00, section 3.3. 2971 */ 2972 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 2973 flag & FLAG_ACKED) 2974 seq_rtt_us = ca_rtt_us = jiffies_to_usecs(tcp_time_stamp - 2975 tp->rx_opt.rcv_tsecr); 2976 if (seq_rtt_us < 0) 2977 return false; 2978 2979 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is 2980 * always taken together with ACK, SACK, or TS-opts. Any negative 2981 * values will be skipped with the seq_rtt_us < 0 check above. 2982 */ 2983 tcp_update_rtt_min(sk, ca_rtt_us); 2984 tcp_rtt_estimator(sk, seq_rtt_us); 2985 tcp_set_rto(sk); 2986 2987 /* RFC6298: only reset backoff on valid RTT measurement. */ 2988 inet_csk(sk)->icsk_backoff = 0; 2989 return true; 2990 } 2991 2992 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */ 2993 void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req) 2994 { 2995 long rtt_us = -1L; 2996 2997 if (req && !req->num_retrans && tcp_rsk(req)->snt_synack.v64) { 2998 struct skb_mstamp now; 2999 3000 skb_mstamp_get(&now); 3001 rtt_us = skb_mstamp_us_delta(&now, &tcp_rsk(req)->snt_synack); 3002 } 3003 3004 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us); 3005 } 3006 3007 3008 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked) 3009 { 3010 const struct inet_connection_sock *icsk = inet_csk(sk); 3011 3012 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked); 3013 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp; 3014 } 3015 3016 /* Restart timer after forward progress on connection. 3017 * RFC2988 recommends to restart timer to now+rto. 3018 */ 3019 void tcp_rearm_rto(struct sock *sk) 3020 { 3021 const struct inet_connection_sock *icsk = inet_csk(sk); 3022 struct tcp_sock *tp = tcp_sk(sk); 3023 3024 /* If the retrans timer is currently being used by Fast Open 3025 * for SYN-ACK retrans purpose, stay put. 3026 */ 3027 if (tp->fastopen_rsk) 3028 return; 3029 3030 if (!tp->packets_out) { 3031 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS); 3032 } else { 3033 u32 rto = inet_csk(sk)->icsk_rto; 3034 /* Offset the time elapsed after installing regular RTO */ 3035 if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS || 3036 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) { 3037 struct sk_buff *skb = tcp_write_queue_head(sk); 3038 const u32 rto_time_stamp = 3039 tcp_skb_timestamp(skb) + rto; 3040 s32 delta = (s32)(rto_time_stamp - tcp_time_stamp); 3041 /* delta may not be positive if the socket is locked 3042 * when the retrans timer fires and is rescheduled. 3043 */ 3044 if (delta > 0) 3045 rto = delta; 3046 } 3047 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto, 3048 TCP_RTO_MAX); 3049 } 3050 } 3051 3052 /* This function is called when the delayed ER timer fires. TCP enters 3053 * fast recovery and performs fast-retransmit. 3054 */ 3055 void tcp_resume_early_retransmit(struct sock *sk) 3056 { 3057 struct tcp_sock *tp = tcp_sk(sk); 3058 3059 tcp_rearm_rto(sk); 3060 3061 /* Stop if ER is disabled after the delayed ER timer is scheduled */ 3062 if (!tp->do_early_retrans) 3063 return; 3064 3065 tcp_enter_recovery(sk, false); 3066 tcp_update_scoreboard(sk, 1); 3067 tcp_xmit_retransmit_queue(sk); 3068 } 3069 3070 /* If we get here, the whole TSO packet has not been acked. */ 3071 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb) 3072 { 3073 struct tcp_sock *tp = tcp_sk(sk); 3074 u32 packets_acked; 3075 3076 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)); 3077 3078 packets_acked = tcp_skb_pcount(skb); 3079 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 3080 return 0; 3081 packets_acked -= tcp_skb_pcount(skb); 3082 3083 if (packets_acked) { 3084 BUG_ON(tcp_skb_pcount(skb) == 0); 3085 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)); 3086 } 3087 3088 return packets_acked; 3089 } 3090 3091 static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb, 3092 u32 prior_snd_una) 3093 { 3094 const struct skb_shared_info *shinfo; 3095 3096 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */ 3097 if (likely(!(sk->sk_tsflags & SOF_TIMESTAMPING_TX_ACK))) 3098 return; 3099 3100 shinfo = skb_shinfo(skb); 3101 if ((shinfo->tx_flags & SKBTX_ACK_TSTAMP) && 3102 !before(shinfo->tskey, prior_snd_una) && 3103 before(shinfo->tskey, tcp_sk(sk)->snd_una)) 3104 __skb_tstamp_tx(skb, NULL, sk, SCM_TSTAMP_ACK); 3105 } 3106 3107 /* Remove acknowledged frames from the retransmission queue. If our packet 3108 * is before the ack sequence we can discard it as it's confirmed to have 3109 * arrived at the other end. 3110 */ 3111 static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets, 3112 u32 prior_snd_una, int *acked, 3113 struct tcp_sacktag_state *sack) 3114 { 3115 const struct inet_connection_sock *icsk = inet_csk(sk); 3116 struct skb_mstamp first_ackt, last_ackt, now; 3117 struct tcp_sock *tp = tcp_sk(sk); 3118 u32 prior_sacked = tp->sacked_out; 3119 u32 reord = tp->packets_out; 3120 bool fully_acked = true; 3121 long sack_rtt_us = -1L; 3122 long seq_rtt_us = -1L; 3123 long ca_rtt_us = -1L; 3124 struct sk_buff *skb; 3125 u32 pkts_acked = 0; 3126 bool rtt_update; 3127 int flag = 0; 3128 3129 first_ackt.v64 = 0; 3130 3131 while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) { 3132 struct tcp_skb_cb *scb = TCP_SKB_CB(skb); 3133 u8 sacked = scb->sacked; 3134 u32 acked_pcount; 3135 3136 tcp_ack_tstamp(sk, skb, prior_snd_una); 3137 3138 /* Determine how many packets and what bytes were acked, tso and else */ 3139 if (after(scb->end_seq, tp->snd_una)) { 3140 if (tcp_skb_pcount(skb) == 1 || 3141 !after(tp->snd_una, scb->seq)) 3142 break; 3143 3144 acked_pcount = tcp_tso_acked(sk, skb); 3145 if (!acked_pcount) 3146 break; 3147 3148 fully_acked = false; 3149 } else { 3150 /* Speedup tcp_unlink_write_queue() and next loop */ 3151 prefetchw(skb->next); 3152 acked_pcount = tcp_skb_pcount(skb); 3153 } 3154 3155 if (unlikely(sacked & TCPCB_RETRANS)) { 3156 if (sacked & TCPCB_SACKED_RETRANS) 3157 tp->retrans_out -= acked_pcount; 3158 flag |= FLAG_RETRANS_DATA_ACKED; 3159 } else if (!(sacked & TCPCB_SACKED_ACKED)) { 3160 last_ackt = skb->skb_mstamp; 3161 WARN_ON_ONCE(last_ackt.v64 == 0); 3162 if (!first_ackt.v64) 3163 first_ackt = last_ackt; 3164 3165 reord = min(pkts_acked, reord); 3166 if (!after(scb->end_seq, tp->high_seq)) 3167 flag |= FLAG_ORIG_SACK_ACKED; 3168 } 3169 3170 if (sacked & TCPCB_SACKED_ACKED) { 3171 tp->sacked_out -= acked_pcount; 3172 } else if (tcp_is_sack(tp)) { 3173 tp->delivered += acked_pcount; 3174 if (!tcp_skb_spurious_retrans(tp, skb)) 3175 tcp_rack_advance(tp, &skb->skb_mstamp, sacked); 3176 } 3177 if (sacked & TCPCB_LOST) 3178 tp->lost_out -= acked_pcount; 3179 3180 tp->packets_out -= acked_pcount; 3181 pkts_acked += acked_pcount; 3182 3183 /* Initial outgoing SYN's get put onto the write_queue 3184 * just like anything else we transmit. It is not 3185 * true data, and if we misinform our callers that 3186 * this ACK acks real data, we will erroneously exit 3187 * connection startup slow start one packet too 3188 * quickly. This is severely frowned upon behavior. 3189 */ 3190 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) { 3191 flag |= FLAG_DATA_ACKED; 3192 } else { 3193 flag |= FLAG_SYN_ACKED; 3194 tp->retrans_stamp = 0; 3195 } 3196 3197 if (!fully_acked) 3198 break; 3199 3200 tcp_unlink_write_queue(skb, sk); 3201 sk_wmem_free_skb(sk, skb); 3202 if (unlikely(skb == tp->retransmit_skb_hint)) 3203 tp->retransmit_skb_hint = NULL; 3204 if (unlikely(skb == tp->lost_skb_hint)) 3205 tp->lost_skb_hint = NULL; 3206 } 3207 3208 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una))) 3209 tp->snd_up = tp->snd_una; 3210 3211 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) 3212 flag |= FLAG_SACK_RENEGING; 3213 3214 skb_mstamp_get(&now); 3215 if (likely(first_ackt.v64) && !(flag & FLAG_RETRANS_DATA_ACKED)) { 3216 seq_rtt_us = skb_mstamp_us_delta(&now, &first_ackt); 3217 ca_rtt_us = skb_mstamp_us_delta(&now, &last_ackt); 3218 } 3219 if (sack->first_sackt.v64) { 3220 sack_rtt_us = skb_mstamp_us_delta(&now, &sack->first_sackt); 3221 ca_rtt_us = skb_mstamp_us_delta(&now, &sack->last_sackt); 3222 } 3223 3224 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us, 3225 ca_rtt_us); 3226 3227 if (flag & FLAG_ACKED) { 3228 tcp_rearm_rto(sk); 3229 if (unlikely(icsk->icsk_mtup.probe_size && 3230 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) { 3231 tcp_mtup_probe_success(sk); 3232 } 3233 3234 if (tcp_is_reno(tp)) { 3235 tcp_remove_reno_sacks(sk, pkts_acked); 3236 } else { 3237 int delta; 3238 3239 /* Non-retransmitted hole got filled? That's reordering */ 3240 if (reord < prior_fackets) 3241 tcp_update_reordering(sk, tp->fackets_out - reord, 0); 3242 3243 delta = tcp_is_fack(tp) ? pkts_acked : 3244 prior_sacked - tp->sacked_out; 3245 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta); 3246 } 3247 3248 tp->fackets_out -= min(pkts_acked, tp->fackets_out); 3249 3250 } else if (skb && rtt_update && sack_rtt_us >= 0 && 3251 sack_rtt_us > skb_mstamp_us_delta(&now, &skb->skb_mstamp)) { 3252 /* Do not re-arm RTO if the sack RTT is measured from data sent 3253 * after when the head was last (re)transmitted. Otherwise the 3254 * timeout may continue to extend in loss recovery. 3255 */ 3256 tcp_rearm_rto(sk); 3257 } 3258 3259 if (icsk->icsk_ca_ops->pkts_acked) 3260 icsk->icsk_ca_ops->pkts_acked(sk, pkts_acked, ca_rtt_us); 3261 3262 #if FASTRETRANS_DEBUG > 0 3263 WARN_ON((int)tp->sacked_out < 0); 3264 WARN_ON((int)tp->lost_out < 0); 3265 WARN_ON((int)tp->retrans_out < 0); 3266 if (!tp->packets_out && tcp_is_sack(tp)) { 3267 icsk = inet_csk(sk); 3268 if (tp->lost_out) { 3269 pr_debug("Leak l=%u %d\n", 3270 tp->lost_out, icsk->icsk_ca_state); 3271 tp->lost_out = 0; 3272 } 3273 if (tp->sacked_out) { 3274 pr_debug("Leak s=%u %d\n", 3275 tp->sacked_out, icsk->icsk_ca_state); 3276 tp->sacked_out = 0; 3277 } 3278 if (tp->retrans_out) { 3279 pr_debug("Leak r=%u %d\n", 3280 tp->retrans_out, icsk->icsk_ca_state); 3281 tp->retrans_out = 0; 3282 } 3283 } 3284 #endif 3285 *acked = pkts_acked; 3286 return flag; 3287 } 3288 3289 static void tcp_ack_probe(struct sock *sk) 3290 { 3291 const struct tcp_sock *tp = tcp_sk(sk); 3292 struct inet_connection_sock *icsk = inet_csk(sk); 3293 3294 /* Was it a usable window open? */ 3295 3296 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) { 3297 icsk->icsk_backoff = 0; 3298 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0); 3299 /* Socket must be waked up by subsequent tcp_data_snd_check(). 3300 * This function is not for random using! 3301 */ 3302 } else { 3303 unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX); 3304 3305 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 3306 when, TCP_RTO_MAX); 3307 } 3308 } 3309 3310 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag) 3311 { 3312 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) || 3313 inet_csk(sk)->icsk_ca_state != TCP_CA_Open; 3314 } 3315 3316 /* Decide wheather to run the increase function of congestion control. */ 3317 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag) 3318 { 3319 /* If reordering is high then always grow cwnd whenever data is 3320 * delivered regardless of its ordering. Otherwise stay conservative 3321 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/ 3322 * new SACK or ECE mark may first advance cwnd here and later reduce 3323 * cwnd in tcp_fastretrans_alert() based on more states. 3324 */ 3325 if (tcp_sk(sk)->reordering > sock_net(sk)->ipv4.sysctl_tcp_reordering) 3326 return flag & FLAG_FORWARD_PROGRESS; 3327 3328 return flag & FLAG_DATA_ACKED; 3329 } 3330 3331 /* The "ultimate" congestion control function that aims to replace the rigid 3332 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction). 3333 * It's called toward the end of processing an ACK with precise rate 3334 * information. All transmission or retransmission are delayed afterwards. 3335 */ 3336 static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked, 3337 int flag) 3338 { 3339 if (tcp_in_cwnd_reduction(sk)) { 3340 /* Reduce cwnd if state mandates */ 3341 tcp_cwnd_reduction(sk, acked_sacked, flag); 3342 } else if (tcp_may_raise_cwnd(sk, flag)) { 3343 /* Advance cwnd if state allows */ 3344 tcp_cong_avoid(sk, ack, acked_sacked); 3345 } 3346 tcp_update_pacing_rate(sk); 3347 } 3348 3349 /* Check that window update is acceptable. 3350 * The function assumes that snd_una<=ack<=snd_next. 3351 */ 3352 static inline bool tcp_may_update_window(const struct tcp_sock *tp, 3353 const u32 ack, const u32 ack_seq, 3354 const u32 nwin) 3355 { 3356 return after(ack, tp->snd_una) || 3357 after(ack_seq, tp->snd_wl1) || 3358 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd); 3359 } 3360 3361 /* If we update tp->snd_una, also update tp->bytes_acked */ 3362 static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack) 3363 { 3364 u32 delta = ack - tp->snd_una; 3365 3366 u64_stats_update_begin(&tp->syncp); 3367 tp->bytes_acked += delta; 3368 u64_stats_update_end(&tp->syncp); 3369 tp->snd_una = ack; 3370 } 3371 3372 /* If we update tp->rcv_nxt, also update tp->bytes_received */ 3373 static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq) 3374 { 3375 u32 delta = seq - tp->rcv_nxt; 3376 3377 u64_stats_update_begin(&tp->syncp); 3378 tp->bytes_received += delta; 3379 u64_stats_update_end(&tp->syncp); 3380 tp->rcv_nxt = seq; 3381 } 3382 3383 /* Update our send window. 3384 * 3385 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2 3386 * and in FreeBSD. NetBSD's one is even worse.) is wrong. 3387 */ 3388 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack, 3389 u32 ack_seq) 3390 { 3391 struct tcp_sock *tp = tcp_sk(sk); 3392 int flag = 0; 3393 u32 nwin = ntohs(tcp_hdr(skb)->window); 3394 3395 if (likely(!tcp_hdr(skb)->syn)) 3396 nwin <<= tp->rx_opt.snd_wscale; 3397 3398 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) { 3399 flag |= FLAG_WIN_UPDATE; 3400 tcp_update_wl(tp, ack_seq); 3401 3402 if (tp->snd_wnd != nwin) { 3403 tp->snd_wnd = nwin; 3404 3405 /* Note, it is the only place, where 3406 * fast path is recovered for sending TCP. 3407 */ 3408 tp->pred_flags = 0; 3409 tcp_fast_path_check(sk); 3410 3411 if (tcp_send_head(sk)) 3412 tcp_slow_start_after_idle_check(sk); 3413 3414 if (nwin > tp->max_window) { 3415 tp->max_window = nwin; 3416 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie); 3417 } 3418 } 3419 } 3420 3421 tcp_snd_una_update(tp, ack); 3422 3423 return flag; 3424 } 3425 3426 /* Return true if we're currently rate-limiting out-of-window ACKs and 3427 * thus shouldn't send a dupack right now. We rate-limit dupacks in 3428 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS 3429 * attacks that send repeated SYNs or ACKs for the same connection. To 3430 * do this, we do not send a duplicate SYNACK or ACK if the remote 3431 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate. 3432 */ 3433 bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb, 3434 int mib_idx, u32 *last_oow_ack_time) 3435 { 3436 /* Data packets without SYNs are not likely part of an ACK loop. */ 3437 if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) && 3438 !tcp_hdr(skb)->syn) 3439 goto not_rate_limited; 3440 3441 if (*last_oow_ack_time) { 3442 s32 elapsed = (s32)(tcp_time_stamp - *last_oow_ack_time); 3443 3444 if (0 <= elapsed && elapsed < sysctl_tcp_invalid_ratelimit) { 3445 NET_INC_STATS_BH(net, mib_idx); 3446 return true; /* rate-limited: don't send yet! */ 3447 } 3448 } 3449 3450 *last_oow_ack_time = tcp_time_stamp; 3451 3452 not_rate_limited: 3453 return false; /* not rate-limited: go ahead, send dupack now! */ 3454 } 3455 3456 /* RFC 5961 7 [ACK Throttling] */ 3457 static void tcp_send_challenge_ack(struct sock *sk, const struct sk_buff *skb) 3458 { 3459 /* unprotected vars, we dont care of overwrites */ 3460 static u32 challenge_timestamp; 3461 static unsigned int challenge_count; 3462 struct tcp_sock *tp = tcp_sk(sk); 3463 u32 now; 3464 3465 /* First check our per-socket dupack rate limit. */ 3466 if (tcp_oow_rate_limited(sock_net(sk), skb, 3467 LINUX_MIB_TCPACKSKIPPEDCHALLENGE, 3468 &tp->last_oow_ack_time)) 3469 return; 3470 3471 /* Then check the check host-wide RFC 5961 rate limit. */ 3472 now = jiffies / HZ; 3473 if (now != challenge_timestamp) { 3474 challenge_timestamp = now; 3475 challenge_count = 0; 3476 } 3477 if (++challenge_count <= sysctl_tcp_challenge_ack_limit) { 3478 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPCHALLENGEACK); 3479 tcp_send_ack(sk); 3480 } 3481 } 3482 3483 static void tcp_store_ts_recent(struct tcp_sock *tp) 3484 { 3485 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval; 3486 tp->rx_opt.ts_recent_stamp = get_seconds(); 3487 } 3488 3489 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq) 3490 { 3491 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) { 3492 /* PAWS bug workaround wrt. ACK frames, the PAWS discard 3493 * extra check below makes sure this can only happen 3494 * for pure ACK frames. -DaveM 3495 * 3496 * Not only, also it occurs for expired timestamps. 3497 */ 3498 3499 if (tcp_paws_check(&tp->rx_opt, 0)) 3500 tcp_store_ts_recent(tp); 3501 } 3502 } 3503 3504 /* This routine deals with acks during a TLP episode. 3505 * We mark the end of a TLP episode on receiving TLP dupack or when 3506 * ack is after tlp_high_seq. 3507 * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe. 3508 */ 3509 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag) 3510 { 3511 struct tcp_sock *tp = tcp_sk(sk); 3512 3513 if (before(ack, tp->tlp_high_seq)) 3514 return; 3515 3516 if (flag & FLAG_DSACKING_ACK) { 3517 /* This DSACK means original and TLP probe arrived; no loss */ 3518 tp->tlp_high_seq = 0; 3519 } else if (after(ack, tp->tlp_high_seq)) { 3520 /* ACK advances: there was a loss, so reduce cwnd. Reset 3521 * tlp_high_seq in tcp_init_cwnd_reduction() 3522 */ 3523 tcp_init_cwnd_reduction(sk); 3524 tcp_set_ca_state(sk, TCP_CA_CWR); 3525 tcp_end_cwnd_reduction(sk); 3526 tcp_try_keep_open(sk); 3527 NET_INC_STATS_BH(sock_net(sk), 3528 LINUX_MIB_TCPLOSSPROBERECOVERY); 3529 } else if (!(flag & (FLAG_SND_UNA_ADVANCED | 3530 FLAG_NOT_DUP | FLAG_DATA_SACKED))) { 3531 /* Pure dupack: original and TLP probe arrived; no loss */ 3532 tp->tlp_high_seq = 0; 3533 } 3534 } 3535 3536 static inline void tcp_in_ack_event(struct sock *sk, u32 flags) 3537 { 3538 const struct inet_connection_sock *icsk = inet_csk(sk); 3539 3540 if (icsk->icsk_ca_ops->in_ack_event) 3541 icsk->icsk_ca_ops->in_ack_event(sk, flags); 3542 } 3543 3544 /* Congestion control has updated the cwnd already. So if we're in 3545 * loss recovery then now we do any new sends (for FRTO) or 3546 * retransmits (for CA_Loss or CA_recovery) that make sense. 3547 */ 3548 static void tcp_xmit_recovery(struct sock *sk, int rexmit) 3549 { 3550 struct tcp_sock *tp = tcp_sk(sk); 3551 3552 if (rexmit == REXMIT_NONE) 3553 return; 3554 3555 if (unlikely(rexmit == 2)) { 3556 __tcp_push_pending_frames(sk, tcp_current_mss(sk), 3557 TCP_NAGLE_OFF); 3558 if (after(tp->snd_nxt, tp->high_seq)) 3559 return; 3560 tp->frto = 0; 3561 } 3562 tcp_xmit_retransmit_queue(sk); 3563 } 3564 3565 /* This routine deals with incoming acks, but not outgoing ones. */ 3566 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag) 3567 { 3568 struct inet_connection_sock *icsk = inet_csk(sk); 3569 struct tcp_sock *tp = tcp_sk(sk); 3570 struct tcp_sacktag_state sack_state; 3571 u32 prior_snd_una = tp->snd_una; 3572 u32 ack_seq = TCP_SKB_CB(skb)->seq; 3573 u32 ack = TCP_SKB_CB(skb)->ack_seq; 3574 bool is_dupack = false; 3575 u32 prior_fackets; 3576 int prior_packets = tp->packets_out; 3577 u32 prior_delivered = tp->delivered; 3578 int acked = 0; /* Number of packets newly acked */ 3579 int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */ 3580 3581 sack_state.first_sackt.v64 = 0; 3582 3583 /* We very likely will need to access write queue head. */ 3584 prefetchw(sk->sk_write_queue.next); 3585 3586 /* If the ack is older than previous acks 3587 * then we can probably ignore it. 3588 */ 3589 if (before(ack, prior_snd_una)) { 3590 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */ 3591 if (before(ack, prior_snd_una - tp->max_window)) { 3592 tcp_send_challenge_ack(sk, skb); 3593 return -1; 3594 } 3595 goto old_ack; 3596 } 3597 3598 /* If the ack includes data we haven't sent yet, discard 3599 * this segment (RFC793 Section 3.9). 3600 */ 3601 if (after(ack, tp->snd_nxt)) 3602 goto invalid_ack; 3603 3604 if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS || 3605 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) 3606 tcp_rearm_rto(sk); 3607 3608 if (after(ack, prior_snd_una)) { 3609 flag |= FLAG_SND_UNA_ADVANCED; 3610 icsk->icsk_retransmits = 0; 3611 } 3612 3613 prior_fackets = tp->fackets_out; 3614 3615 /* ts_recent update must be made after we are sure that the packet 3616 * is in window. 3617 */ 3618 if (flag & FLAG_UPDATE_TS_RECENT) 3619 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq); 3620 3621 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) { 3622 /* Window is constant, pure forward advance. 3623 * No more checks are required. 3624 * Note, we use the fact that SND.UNA>=SND.WL2. 3625 */ 3626 tcp_update_wl(tp, ack_seq); 3627 tcp_snd_una_update(tp, ack); 3628 flag |= FLAG_WIN_UPDATE; 3629 3630 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE); 3631 3632 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS); 3633 } else { 3634 u32 ack_ev_flags = CA_ACK_SLOWPATH; 3635 3636 if (ack_seq != TCP_SKB_CB(skb)->end_seq) 3637 flag |= FLAG_DATA; 3638 else 3639 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS); 3640 3641 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq); 3642 3643 if (TCP_SKB_CB(skb)->sacked) 3644 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una, 3645 &sack_state); 3646 3647 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) { 3648 flag |= FLAG_ECE; 3649 ack_ev_flags |= CA_ACK_ECE; 3650 } 3651 3652 if (flag & FLAG_WIN_UPDATE) 3653 ack_ev_flags |= CA_ACK_WIN_UPDATE; 3654 3655 tcp_in_ack_event(sk, ack_ev_flags); 3656 } 3657 3658 /* We passed data and got it acked, remove any soft error 3659 * log. Something worked... 3660 */ 3661 sk->sk_err_soft = 0; 3662 icsk->icsk_probes_out = 0; 3663 tp->rcv_tstamp = tcp_time_stamp; 3664 if (!prior_packets) 3665 goto no_queue; 3666 3667 /* See if we can take anything off of the retransmit queue. */ 3668 flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una, &acked, 3669 &sack_state); 3670 3671 if (tcp_ack_is_dubious(sk, flag)) { 3672 is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP)); 3673 tcp_fastretrans_alert(sk, acked, is_dupack, &flag, &rexmit); 3674 } 3675 if (tp->tlp_high_seq) 3676 tcp_process_tlp_ack(sk, ack, flag); 3677 3678 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) { 3679 struct dst_entry *dst = __sk_dst_get(sk); 3680 if (dst) 3681 dst_confirm(dst); 3682 } 3683 3684 if (icsk->icsk_pending == ICSK_TIME_RETRANS) 3685 tcp_schedule_loss_probe(sk); 3686 tcp_cong_control(sk, ack, tp->delivered - prior_delivered, flag); 3687 tcp_xmit_recovery(sk, rexmit); 3688 return 1; 3689 3690 no_queue: 3691 /* If data was DSACKed, see if we can undo a cwnd reduction. */ 3692 if (flag & FLAG_DSACKING_ACK) 3693 tcp_fastretrans_alert(sk, acked, is_dupack, &flag, &rexmit); 3694 /* If this ack opens up a zero window, clear backoff. It was 3695 * being used to time the probes, and is probably far higher than 3696 * it needs to be for normal retransmission. 3697 */ 3698 if (tcp_send_head(sk)) 3699 tcp_ack_probe(sk); 3700 3701 if (tp->tlp_high_seq) 3702 tcp_process_tlp_ack(sk, ack, flag); 3703 return 1; 3704 3705 invalid_ack: 3706 SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt); 3707 return -1; 3708 3709 old_ack: 3710 /* If data was SACKed, tag it and see if we should send more data. 3711 * If data was DSACKed, see if we can undo a cwnd reduction. 3712 */ 3713 if (TCP_SKB_CB(skb)->sacked) { 3714 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una, 3715 &sack_state); 3716 tcp_fastretrans_alert(sk, acked, is_dupack, &flag, &rexmit); 3717 tcp_xmit_recovery(sk, rexmit); 3718 } 3719 3720 SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt); 3721 return 0; 3722 } 3723 3724 static void tcp_parse_fastopen_option(int len, const unsigned char *cookie, 3725 bool syn, struct tcp_fastopen_cookie *foc, 3726 bool exp_opt) 3727 { 3728 /* Valid only in SYN or SYN-ACK with an even length. */ 3729 if (!foc || !syn || len < 0 || (len & 1)) 3730 return; 3731 3732 if (len >= TCP_FASTOPEN_COOKIE_MIN && 3733 len <= TCP_FASTOPEN_COOKIE_MAX) 3734 memcpy(foc->val, cookie, len); 3735 else if (len != 0) 3736 len = -1; 3737 foc->len = len; 3738 foc->exp = exp_opt; 3739 } 3740 3741 /* Look for tcp options. Normally only called on SYN and SYNACK packets. 3742 * But, this can also be called on packets in the established flow when 3743 * the fast version below fails. 3744 */ 3745 void tcp_parse_options(const struct sk_buff *skb, 3746 struct tcp_options_received *opt_rx, int estab, 3747 struct tcp_fastopen_cookie *foc) 3748 { 3749 const unsigned char *ptr; 3750 const struct tcphdr *th = tcp_hdr(skb); 3751 int length = (th->doff * 4) - sizeof(struct tcphdr); 3752 3753 ptr = (const unsigned char *)(th + 1); 3754 opt_rx->saw_tstamp = 0; 3755 3756 while (length > 0) { 3757 int opcode = *ptr++; 3758 int opsize; 3759 3760 switch (opcode) { 3761 case TCPOPT_EOL: 3762 return; 3763 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */ 3764 length--; 3765 continue; 3766 default: 3767 opsize = *ptr++; 3768 if (opsize < 2) /* "silly options" */ 3769 return; 3770 if (opsize > length) 3771 return; /* don't parse partial options */ 3772 switch (opcode) { 3773 case TCPOPT_MSS: 3774 if (opsize == TCPOLEN_MSS && th->syn && !estab) { 3775 u16 in_mss = get_unaligned_be16(ptr); 3776 if (in_mss) { 3777 if (opt_rx->user_mss && 3778 opt_rx->user_mss < in_mss) 3779 in_mss = opt_rx->user_mss; 3780 opt_rx->mss_clamp = in_mss; 3781 } 3782 } 3783 break; 3784 case TCPOPT_WINDOW: 3785 if (opsize == TCPOLEN_WINDOW && th->syn && 3786 !estab && sysctl_tcp_window_scaling) { 3787 __u8 snd_wscale = *(__u8 *)ptr; 3788 opt_rx->wscale_ok = 1; 3789 if (snd_wscale > 14) { 3790 net_info_ratelimited("%s: Illegal window scaling value %d >14 received\n", 3791 __func__, 3792 snd_wscale); 3793 snd_wscale = 14; 3794 } 3795 opt_rx->snd_wscale = snd_wscale; 3796 } 3797 break; 3798 case TCPOPT_TIMESTAMP: 3799 if ((opsize == TCPOLEN_TIMESTAMP) && 3800 ((estab && opt_rx->tstamp_ok) || 3801 (!estab && sysctl_tcp_timestamps))) { 3802 opt_rx->saw_tstamp = 1; 3803 opt_rx->rcv_tsval = get_unaligned_be32(ptr); 3804 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4); 3805 } 3806 break; 3807 case TCPOPT_SACK_PERM: 3808 if (opsize == TCPOLEN_SACK_PERM && th->syn && 3809 !estab && sysctl_tcp_sack) { 3810 opt_rx->sack_ok = TCP_SACK_SEEN; 3811 tcp_sack_reset(opt_rx); 3812 } 3813 break; 3814 3815 case TCPOPT_SACK: 3816 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) && 3817 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) && 3818 opt_rx->sack_ok) { 3819 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th; 3820 } 3821 break; 3822 #ifdef CONFIG_TCP_MD5SIG 3823 case TCPOPT_MD5SIG: 3824 /* 3825 * The MD5 Hash has already been 3826 * checked (see tcp_v{4,6}_do_rcv()). 3827 */ 3828 break; 3829 #endif 3830 case TCPOPT_FASTOPEN: 3831 tcp_parse_fastopen_option( 3832 opsize - TCPOLEN_FASTOPEN_BASE, 3833 ptr, th->syn, foc, false); 3834 break; 3835 3836 case TCPOPT_EXP: 3837 /* Fast Open option shares code 254 using a 3838 * 16 bits magic number. 3839 */ 3840 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE && 3841 get_unaligned_be16(ptr) == 3842 TCPOPT_FASTOPEN_MAGIC) 3843 tcp_parse_fastopen_option(opsize - 3844 TCPOLEN_EXP_FASTOPEN_BASE, 3845 ptr + 2, th->syn, foc, true); 3846 break; 3847 3848 } 3849 ptr += opsize-2; 3850 length -= opsize; 3851 } 3852 } 3853 } 3854 EXPORT_SYMBOL(tcp_parse_options); 3855 3856 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th) 3857 { 3858 const __be32 *ptr = (const __be32 *)(th + 1); 3859 3860 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) 3861 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) { 3862 tp->rx_opt.saw_tstamp = 1; 3863 ++ptr; 3864 tp->rx_opt.rcv_tsval = ntohl(*ptr); 3865 ++ptr; 3866 if (*ptr) 3867 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset; 3868 else 3869 tp->rx_opt.rcv_tsecr = 0; 3870 return true; 3871 } 3872 return false; 3873 } 3874 3875 /* Fast parse options. This hopes to only see timestamps. 3876 * If it is wrong it falls back on tcp_parse_options(). 3877 */ 3878 static bool tcp_fast_parse_options(const struct sk_buff *skb, 3879 const struct tcphdr *th, struct tcp_sock *tp) 3880 { 3881 /* In the spirit of fast parsing, compare doff directly to constant 3882 * values. Because equality is used, short doff can be ignored here. 3883 */ 3884 if (th->doff == (sizeof(*th) / 4)) { 3885 tp->rx_opt.saw_tstamp = 0; 3886 return false; 3887 } else if (tp->rx_opt.tstamp_ok && 3888 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) { 3889 if (tcp_parse_aligned_timestamp(tp, th)) 3890 return true; 3891 } 3892 3893 tcp_parse_options(skb, &tp->rx_opt, 1, NULL); 3894 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr) 3895 tp->rx_opt.rcv_tsecr -= tp->tsoffset; 3896 3897 return true; 3898 } 3899 3900 #ifdef CONFIG_TCP_MD5SIG 3901 /* 3902 * Parse MD5 Signature option 3903 */ 3904 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th) 3905 { 3906 int length = (th->doff << 2) - sizeof(*th); 3907 const u8 *ptr = (const u8 *)(th + 1); 3908 3909 /* If the TCP option is too short, we can short cut */ 3910 if (length < TCPOLEN_MD5SIG) 3911 return NULL; 3912 3913 while (length > 0) { 3914 int opcode = *ptr++; 3915 int opsize; 3916 3917 switch (opcode) { 3918 case TCPOPT_EOL: 3919 return NULL; 3920 case TCPOPT_NOP: 3921 length--; 3922 continue; 3923 default: 3924 opsize = *ptr++; 3925 if (opsize < 2 || opsize > length) 3926 return NULL; 3927 if (opcode == TCPOPT_MD5SIG) 3928 return opsize == TCPOLEN_MD5SIG ? ptr : NULL; 3929 } 3930 ptr += opsize - 2; 3931 length -= opsize; 3932 } 3933 return NULL; 3934 } 3935 EXPORT_SYMBOL(tcp_parse_md5sig_option); 3936 #endif 3937 3938 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM 3939 * 3940 * It is not fatal. If this ACK does _not_ change critical state (seqs, window) 3941 * it can pass through stack. So, the following predicate verifies that 3942 * this segment is not used for anything but congestion avoidance or 3943 * fast retransmit. Moreover, we even are able to eliminate most of such 3944 * second order effects, if we apply some small "replay" window (~RTO) 3945 * to timestamp space. 3946 * 3947 * All these measures still do not guarantee that we reject wrapped ACKs 3948 * on networks with high bandwidth, when sequence space is recycled fastly, 3949 * but it guarantees that such events will be very rare and do not affect 3950 * connection seriously. This doesn't look nice, but alas, PAWS is really 3951 * buggy extension. 3952 * 3953 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC 3954 * states that events when retransmit arrives after original data are rare. 3955 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is 3956 * the biggest problem on large power networks even with minor reordering. 3957 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe 3958 * up to bandwidth of 18Gigabit/sec. 8) ] 3959 */ 3960 3961 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb) 3962 { 3963 const struct tcp_sock *tp = tcp_sk(sk); 3964 const struct tcphdr *th = tcp_hdr(skb); 3965 u32 seq = TCP_SKB_CB(skb)->seq; 3966 u32 ack = TCP_SKB_CB(skb)->ack_seq; 3967 3968 return (/* 1. Pure ACK with correct sequence number. */ 3969 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) && 3970 3971 /* 2. ... and duplicate ACK. */ 3972 ack == tp->snd_una && 3973 3974 /* 3. ... and does not update window. */ 3975 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) && 3976 3977 /* 4. ... and sits in replay window. */ 3978 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ); 3979 } 3980 3981 static inline bool tcp_paws_discard(const struct sock *sk, 3982 const struct sk_buff *skb) 3983 { 3984 const struct tcp_sock *tp = tcp_sk(sk); 3985 3986 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) && 3987 !tcp_disordered_ack(sk, skb); 3988 } 3989 3990 /* Check segment sequence number for validity. 3991 * 3992 * Segment controls are considered valid, if the segment 3993 * fits to the window after truncation to the window. Acceptability 3994 * of data (and SYN, FIN, of course) is checked separately. 3995 * See tcp_data_queue(), for example. 3996 * 3997 * Also, controls (RST is main one) are accepted using RCV.WUP instead 3998 * of RCV.NXT. Peer still did not advance his SND.UNA when we 3999 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP. 4000 * (borrowed from freebsd) 4001 */ 4002 4003 static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq) 4004 { 4005 return !before(end_seq, tp->rcv_wup) && 4006 !after(seq, tp->rcv_nxt + tcp_receive_window(tp)); 4007 } 4008 4009 /* When we get a reset we do this. */ 4010 void tcp_reset(struct sock *sk) 4011 { 4012 /* We want the right error as BSD sees it (and indeed as we do). */ 4013 switch (sk->sk_state) { 4014 case TCP_SYN_SENT: 4015 sk->sk_err = ECONNREFUSED; 4016 break; 4017 case TCP_CLOSE_WAIT: 4018 sk->sk_err = EPIPE; 4019 break; 4020 case TCP_CLOSE: 4021 return; 4022 default: 4023 sk->sk_err = ECONNRESET; 4024 } 4025 /* This barrier is coupled with smp_rmb() in tcp_poll() */ 4026 smp_wmb(); 4027 4028 if (!sock_flag(sk, SOCK_DEAD)) 4029 sk->sk_error_report(sk); 4030 4031 tcp_done(sk); 4032 } 4033 4034 /* 4035 * Process the FIN bit. This now behaves as it is supposed to work 4036 * and the FIN takes effect when it is validly part of sequence 4037 * space. Not before when we get holes. 4038 * 4039 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT 4040 * (and thence onto LAST-ACK and finally, CLOSE, we never enter 4041 * TIME-WAIT) 4042 * 4043 * If we are in FINWAIT-1, a received FIN indicates simultaneous 4044 * close and we go into CLOSING (and later onto TIME-WAIT) 4045 * 4046 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT. 4047 */ 4048 void tcp_fin(struct sock *sk) 4049 { 4050 struct tcp_sock *tp = tcp_sk(sk); 4051 4052 inet_csk_schedule_ack(sk); 4053 4054 sk->sk_shutdown |= RCV_SHUTDOWN; 4055 sock_set_flag(sk, SOCK_DONE); 4056 4057 switch (sk->sk_state) { 4058 case TCP_SYN_RECV: 4059 case TCP_ESTABLISHED: 4060 /* Move to CLOSE_WAIT */ 4061 tcp_set_state(sk, TCP_CLOSE_WAIT); 4062 inet_csk(sk)->icsk_ack.pingpong = 1; 4063 break; 4064 4065 case TCP_CLOSE_WAIT: 4066 case TCP_CLOSING: 4067 /* Received a retransmission of the FIN, do 4068 * nothing. 4069 */ 4070 break; 4071 case TCP_LAST_ACK: 4072 /* RFC793: Remain in the LAST-ACK state. */ 4073 break; 4074 4075 case TCP_FIN_WAIT1: 4076 /* This case occurs when a simultaneous close 4077 * happens, we must ack the received FIN and 4078 * enter the CLOSING state. 4079 */ 4080 tcp_send_ack(sk); 4081 tcp_set_state(sk, TCP_CLOSING); 4082 break; 4083 case TCP_FIN_WAIT2: 4084 /* Received a FIN -- send ACK and enter TIME_WAIT. */ 4085 tcp_send_ack(sk); 4086 tcp_time_wait(sk, TCP_TIME_WAIT, 0); 4087 break; 4088 default: 4089 /* Only TCP_LISTEN and TCP_CLOSE are left, in these 4090 * cases we should never reach this piece of code. 4091 */ 4092 pr_err("%s: Impossible, sk->sk_state=%d\n", 4093 __func__, sk->sk_state); 4094 break; 4095 } 4096 4097 /* It _is_ possible, that we have something out-of-order _after_ FIN. 4098 * Probably, we should reset in this case. For now drop them. 4099 */ 4100 __skb_queue_purge(&tp->out_of_order_queue); 4101 if (tcp_is_sack(tp)) 4102 tcp_sack_reset(&tp->rx_opt); 4103 sk_mem_reclaim(sk); 4104 4105 if (!sock_flag(sk, SOCK_DEAD)) { 4106 sk->sk_state_change(sk); 4107 4108 /* Do not send POLL_HUP for half duplex close. */ 4109 if (sk->sk_shutdown == SHUTDOWN_MASK || 4110 sk->sk_state == TCP_CLOSE) 4111 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP); 4112 else 4113 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN); 4114 } 4115 } 4116 4117 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, 4118 u32 end_seq) 4119 { 4120 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) { 4121 if (before(seq, sp->start_seq)) 4122 sp->start_seq = seq; 4123 if (after(end_seq, sp->end_seq)) 4124 sp->end_seq = end_seq; 4125 return true; 4126 } 4127 return false; 4128 } 4129 4130 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq) 4131 { 4132 struct tcp_sock *tp = tcp_sk(sk); 4133 4134 if (tcp_is_sack(tp) && sysctl_tcp_dsack) { 4135 int mib_idx; 4136 4137 if (before(seq, tp->rcv_nxt)) 4138 mib_idx = LINUX_MIB_TCPDSACKOLDSENT; 4139 else 4140 mib_idx = LINUX_MIB_TCPDSACKOFOSENT; 4141 4142 NET_INC_STATS_BH(sock_net(sk), mib_idx); 4143 4144 tp->rx_opt.dsack = 1; 4145 tp->duplicate_sack[0].start_seq = seq; 4146 tp->duplicate_sack[0].end_seq = end_seq; 4147 } 4148 } 4149 4150 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq) 4151 { 4152 struct tcp_sock *tp = tcp_sk(sk); 4153 4154 if (!tp->rx_opt.dsack) 4155 tcp_dsack_set(sk, seq, end_seq); 4156 else 4157 tcp_sack_extend(tp->duplicate_sack, seq, end_seq); 4158 } 4159 4160 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb) 4161 { 4162 struct tcp_sock *tp = tcp_sk(sk); 4163 4164 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 4165 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 4166 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST); 4167 tcp_enter_quickack_mode(sk); 4168 4169 if (tcp_is_sack(tp) && sysctl_tcp_dsack) { 4170 u32 end_seq = TCP_SKB_CB(skb)->end_seq; 4171 4172 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) 4173 end_seq = tp->rcv_nxt; 4174 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq); 4175 } 4176 } 4177 4178 tcp_send_ack(sk); 4179 } 4180 4181 /* These routines update the SACK block as out-of-order packets arrive or 4182 * in-order packets close up the sequence space. 4183 */ 4184 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp) 4185 { 4186 int this_sack; 4187 struct tcp_sack_block *sp = &tp->selective_acks[0]; 4188 struct tcp_sack_block *swalk = sp + 1; 4189 4190 /* See if the recent change to the first SACK eats into 4191 * or hits the sequence space of other SACK blocks, if so coalesce. 4192 */ 4193 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) { 4194 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) { 4195 int i; 4196 4197 /* Zap SWALK, by moving every further SACK up by one slot. 4198 * Decrease num_sacks. 4199 */ 4200 tp->rx_opt.num_sacks--; 4201 for (i = this_sack; i < tp->rx_opt.num_sacks; i++) 4202 sp[i] = sp[i + 1]; 4203 continue; 4204 } 4205 this_sack++, swalk++; 4206 } 4207 } 4208 4209 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq) 4210 { 4211 struct tcp_sock *tp = tcp_sk(sk); 4212 struct tcp_sack_block *sp = &tp->selective_acks[0]; 4213 int cur_sacks = tp->rx_opt.num_sacks; 4214 int this_sack; 4215 4216 if (!cur_sacks) 4217 goto new_sack; 4218 4219 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) { 4220 if (tcp_sack_extend(sp, seq, end_seq)) { 4221 /* Rotate this_sack to the first one. */ 4222 for (; this_sack > 0; this_sack--, sp--) 4223 swap(*sp, *(sp - 1)); 4224 if (cur_sacks > 1) 4225 tcp_sack_maybe_coalesce(tp); 4226 return; 4227 } 4228 } 4229 4230 /* Could not find an adjacent existing SACK, build a new one, 4231 * put it at the front, and shift everyone else down. We 4232 * always know there is at least one SACK present already here. 4233 * 4234 * If the sack array is full, forget about the last one. 4235 */ 4236 if (this_sack >= TCP_NUM_SACKS) { 4237 this_sack--; 4238 tp->rx_opt.num_sacks--; 4239 sp--; 4240 } 4241 for (; this_sack > 0; this_sack--, sp--) 4242 *sp = *(sp - 1); 4243 4244 new_sack: 4245 /* Build the new head SACK, and we're done. */ 4246 sp->start_seq = seq; 4247 sp->end_seq = end_seq; 4248 tp->rx_opt.num_sacks++; 4249 } 4250 4251 /* RCV.NXT advances, some SACKs should be eaten. */ 4252 4253 static void tcp_sack_remove(struct tcp_sock *tp) 4254 { 4255 struct tcp_sack_block *sp = &tp->selective_acks[0]; 4256 int num_sacks = tp->rx_opt.num_sacks; 4257 int this_sack; 4258 4259 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */ 4260 if (skb_queue_empty(&tp->out_of_order_queue)) { 4261 tp->rx_opt.num_sacks = 0; 4262 return; 4263 } 4264 4265 for (this_sack = 0; this_sack < num_sacks;) { 4266 /* Check if the start of the sack is covered by RCV.NXT. */ 4267 if (!before(tp->rcv_nxt, sp->start_seq)) { 4268 int i; 4269 4270 /* RCV.NXT must cover all the block! */ 4271 WARN_ON(before(tp->rcv_nxt, sp->end_seq)); 4272 4273 /* Zap this SACK, by moving forward any other SACKS. */ 4274 for (i = this_sack+1; i < num_sacks; i++) 4275 tp->selective_acks[i-1] = tp->selective_acks[i]; 4276 num_sacks--; 4277 continue; 4278 } 4279 this_sack++; 4280 sp++; 4281 } 4282 tp->rx_opt.num_sacks = num_sacks; 4283 } 4284 4285 /** 4286 * tcp_try_coalesce - try to merge skb to prior one 4287 * @sk: socket 4288 * @to: prior buffer 4289 * @from: buffer to add in queue 4290 * @fragstolen: pointer to boolean 4291 * 4292 * Before queueing skb @from after @to, try to merge them 4293 * to reduce overall memory use and queue lengths, if cost is small. 4294 * Packets in ofo or receive queues can stay a long time. 4295 * Better try to coalesce them right now to avoid future collapses. 4296 * Returns true if caller should free @from instead of queueing it 4297 */ 4298 static bool tcp_try_coalesce(struct sock *sk, 4299 struct sk_buff *to, 4300 struct sk_buff *from, 4301 bool *fragstolen) 4302 { 4303 int delta; 4304 4305 *fragstolen = false; 4306 4307 /* Its possible this segment overlaps with prior segment in queue */ 4308 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq) 4309 return false; 4310 4311 if (!skb_try_coalesce(to, from, fragstolen, &delta)) 4312 return false; 4313 4314 atomic_add(delta, &sk->sk_rmem_alloc); 4315 sk_mem_charge(sk, delta); 4316 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE); 4317 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq; 4318 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq; 4319 TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags; 4320 return true; 4321 } 4322 4323 /* This one checks to see if we can put data from the 4324 * out_of_order queue into the receive_queue. 4325 */ 4326 static void tcp_ofo_queue(struct sock *sk) 4327 { 4328 struct tcp_sock *tp = tcp_sk(sk); 4329 __u32 dsack_high = tp->rcv_nxt; 4330 struct sk_buff *skb, *tail; 4331 bool fragstolen, eaten; 4332 4333 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) { 4334 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) 4335 break; 4336 4337 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) { 4338 __u32 dsack = dsack_high; 4339 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high)) 4340 dsack_high = TCP_SKB_CB(skb)->end_seq; 4341 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack); 4342 } 4343 4344 __skb_unlink(skb, &tp->out_of_order_queue); 4345 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) { 4346 SOCK_DEBUG(sk, "ofo packet was already received\n"); 4347 __kfree_skb(skb); 4348 continue; 4349 } 4350 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n", 4351 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, 4352 TCP_SKB_CB(skb)->end_seq); 4353 4354 tail = skb_peek_tail(&sk->sk_receive_queue); 4355 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen); 4356 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq); 4357 if (!eaten) 4358 __skb_queue_tail(&sk->sk_receive_queue, skb); 4359 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 4360 tcp_fin(sk); 4361 if (eaten) 4362 kfree_skb_partial(skb, fragstolen); 4363 } 4364 } 4365 4366 static bool tcp_prune_ofo_queue(struct sock *sk); 4367 static int tcp_prune_queue(struct sock *sk); 4368 4369 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb, 4370 unsigned int size) 4371 { 4372 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf || 4373 !sk_rmem_schedule(sk, skb, size)) { 4374 4375 if (tcp_prune_queue(sk) < 0) 4376 return -1; 4377 4378 if (!sk_rmem_schedule(sk, skb, size)) { 4379 if (!tcp_prune_ofo_queue(sk)) 4380 return -1; 4381 4382 if (!sk_rmem_schedule(sk, skb, size)) 4383 return -1; 4384 } 4385 } 4386 return 0; 4387 } 4388 4389 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb) 4390 { 4391 struct tcp_sock *tp = tcp_sk(sk); 4392 struct sk_buff *skb1; 4393 u32 seq, end_seq; 4394 4395 tcp_ecn_check_ce(tp, skb); 4396 4397 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) { 4398 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFODROP); 4399 __kfree_skb(skb); 4400 return; 4401 } 4402 4403 /* Disable header prediction. */ 4404 tp->pred_flags = 0; 4405 inet_csk_schedule_ack(sk); 4406 4407 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOQUEUE); 4408 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n", 4409 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); 4410 4411 skb1 = skb_peek_tail(&tp->out_of_order_queue); 4412 if (!skb1) { 4413 /* Initial out of order segment, build 1 SACK. */ 4414 if (tcp_is_sack(tp)) { 4415 tp->rx_opt.num_sacks = 1; 4416 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq; 4417 tp->selective_acks[0].end_seq = 4418 TCP_SKB_CB(skb)->end_seq; 4419 } 4420 __skb_queue_head(&tp->out_of_order_queue, skb); 4421 goto end; 4422 } 4423 4424 seq = TCP_SKB_CB(skb)->seq; 4425 end_seq = TCP_SKB_CB(skb)->end_seq; 4426 4427 if (seq == TCP_SKB_CB(skb1)->end_seq) { 4428 bool fragstolen; 4429 4430 if (!tcp_try_coalesce(sk, skb1, skb, &fragstolen)) { 4431 __skb_queue_after(&tp->out_of_order_queue, skb1, skb); 4432 } else { 4433 tcp_grow_window(sk, skb); 4434 kfree_skb_partial(skb, fragstolen); 4435 skb = NULL; 4436 } 4437 4438 if (!tp->rx_opt.num_sacks || 4439 tp->selective_acks[0].end_seq != seq) 4440 goto add_sack; 4441 4442 /* Common case: data arrive in order after hole. */ 4443 tp->selective_acks[0].end_seq = end_seq; 4444 goto end; 4445 } 4446 4447 /* Find place to insert this segment. */ 4448 while (1) { 4449 if (!after(TCP_SKB_CB(skb1)->seq, seq)) 4450 break; 4451 if (skb_queue_is_first(&tp->out_of_order_queue, skb1)) { 4452 skb1 = NULL; 4453 break; 4454 } 4455 skb1 = skb_queue_prev(&tp->out_of_order_queue, skb1); 4456 } 4457 4458 /* Do skb overlap to previous one? */ 4459 if (skb1 && before(seq, TCP_SKB_CB(skb1)->end_seq)) { 4460 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) { 4461 /* All the bits are present. Drop. */ 4462 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE); 4463 __kfree_skb(skb); 4464 skb = NULL; 4465 tcp_dsack_set(sk, seq, end_seq); 4466 goto add_sack; 4467 } 4468 if (after(seq, TCP_SKB_CB(skb1)->seq)) { 4469 /* Partial overlap. */ 4470 tcp_dsack_set(sk, seq, 4471 TCP_SKB_CB(skb1)->end_seq); 4472 } else { 4473 if (skb_queue_is_first(&tp->out_of_order_queue, 4474 skb1)) 4475 skb1 = NULL; 4476 else 4477 skb1 = skb_queue_prev( 4478 &tp->out_of_order_queue, 4479 skb1); 4480 } 4481 } 4482 if (!skb1) 4483 __skb_queue_head(&tp->out_of_order_queue, skb); 4484 else 4485 __skb_queue_after(&tp->out_of_order_queue, skb1, skb); 4486 4487 /* And clean segments covered by new one as whole. */ 4488 while (!skb_queue_is_last(&tp->out_of_order_queue, skb)) { 4489 skb1 = skb_queue_next(&tp->out_of_order_queue, skb); 4490 4491 if (!after(end_seq, TCP_SKB_CB(skb1)->seq)) 4492 break; 4493 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) { 4494 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq, 4495 end_seq); 4496 break; 4497 } 4498 __skb_unlink(skb1, &tp->out_of_order_queue); 4499 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq, 4500 TCP_SKB_CB(skb1)->end_seq); 4501 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE); 4502 __kfree_skb(skb1); 4503 } 4504 4505 add_sack: 4506 if (tcp_is_sack(tp)) 4507 tcp_sack_new_ofo_skb(sk, seq, end_seq); 4508 end: 4509 if (skb) { 4510 tcp_grow_window(sk, skb); 4511 skb_set_owner_r(skb, sk); 4512 } 4513 } 4514 4515 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, int hdrlen, 4516 bool *fragstolen) 4517 { 4518 int eaten; 4519 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue); 4520 4521 __skb_pull(skb, hdrlen); 4522 eaten = (tail && 4523 tcp_try_coalesce(sk, tail, skb, fragstolen)) ? 1 : 0; 4524 tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq); 4525 if (!eaten) { 4526 __skb_queue_tail(&sk->sk_receive_queue, skb); 4527 skb_set_owner_r(skb, sk); 4528 } 4529 return eaten; 4530 } 4531 4532 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size) 4533 { 4534 struct sk_buff *skb; 4535 int err = -ENOMEM; 4536 int data_len = 0; 4537 bool fragstolen; 4538 4539 if (size == 0) 4540 return 0; 4541 4542 if (size > PAGE_SIZE) { 4543 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS); 4544 4545 data_len = npages << PAGE_SHIFT; 4546 size = data_len + (size & ~PAGE_MASK); 4547 } 4548 skb = alloc_skb_with_frags(size - data_len, data_len, 4549 PAGE_ALLOC_COSTLY_ORDER, 4550 &err, sk->sk_allocation); 4551 if (!skb) 4552 goto err; 4553 4554 skb_put(skb, size - data_len); 4555 skb->data_len = data_len; 4556 skb->len = size; 4557 4558 if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) 4559 goto err_free; 4560 4561 err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size); 4562 if (err) 4563 goto err_free; 4564 4565 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt; 4566 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size; 4567 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1; 4568 4569 if (tcp_queue_rcv(sk, skb, 0, &fragstolen)) { 4570 WARN_ON_ONCE(fragstolen); /* should not happen */ 4571 __kfree_skb(skb); 4572 } 4573 return size; 4574 4575 err_free: 4576 kfree_skb(skb); 4577 err: 4578 return err; 4579 4580 } 4581 4582 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb) 4583 { 4584 struct tcp_sock *tp = tcp_sk(sk); 4585 int eaten = -1; 4586 bool fragstolen = false; 4587 4588 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) 4589 goto drop; 4590 4591 skb_dst_drop(skb); 4592 __skb_pull(skb, tcp_hdr(skb)->doff * 4); 4593 4594 tcp_ecn_accept_cwr(tp, skb); 4595 4596 tp->rx_opt.dsack = 0; 4597 4598 /* Queue data for delivery to the user. 4599 * Packets in sequence go to the receive queue. 4600 * Out of sequence packets to the out_of_order_queue. 4601 */ 4602 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) { 4603 if (tcp_receive_window(tp) == 0) 4604 goto out_of_window; 4605 4606 /* Ok. In sequence. In window. */ 4607 if (tp->ucopy.task == current && 4608 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len && 4609 sock_owned_by_user(sk) && !tp->urg_data) { 4610 int chunk = min_t(unsigned int, skb->len, 4611 tp->ucopy.len); 4612 4613 __set_current_state(TASK_RUNNING); 4614 4615 local_bh_enable(); 4616 if (!skb_copy_datagram_msg(skb, 0, tp->ucopy.msg, chunk)) { 4617 tp->ucopy.len -= chunk; 4618 tp->copied_seq += chunk; 4619 eaten = (chunk == skb->len); 4620 tcp_rcv_space_adjust(sk); 4621 } 4622 local_bh_disable(); 4623 } 4624 4625 if (eaten <= 0) { 4626 queue_and_out: 4627 if (eaten < 0) { 4628 if (skb_queue_len(&sk->sk_receive_queue) == 0) 4629 sk_forced_mem_schedule(sk, skb->truesize); 4630 else if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) 4631 goto drop; 4632 } 4633 eaten = tcp_queue_rcv(sk, skb, 0, &fragstolen); 4634 } 4635 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq); 4636 if (skb->len) 4637 tcp_event_data_recv(sk, skb); 4638 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 4639 tcp_fin(sk); 4640 4641 if (!skb_queue_empty(&tp->out_of_order_queue)) { 4642 tcp_ofo_queue(sk); 4643 4644 /* RFC2581. 4.2. SHOULD send immediate ACK, when 4645 * gap in queue is filled. 4646 */ 4647 if (skb_queue_empty(&tp->out_of_order_queue)) 4648 inet_csk(sk)->icsk_ack.pingpong = 0; 4649 } 4650 4651 if (tp->rx_opt.num_sacks) 4652 tcp_sack_remove(tp); 4653 4654 tcp_fast_path_check(sk); 4655 4656 if (eaten > 0) 4657 kfree_skb_partial(skb, fragstolen); 4658 if (!sock_flag(sk, SOCK_DEAD)) 4659 sk->sk_data_ready(sk); 4660 return; 4661 } 4662 4663 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) { 4664 /* A retransmit, 2nd most common case. Force an immediate ack. */ 4665 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST); 4666 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); 4667 4668 out_of_window: 4669 tcp_enter_quickack_mode(sk); 4670 inet_csk_schedule_ack(sk); 4671 drop: 4672 __kfree_skb(skb); 4673 return; 4674 } 4675 4676 /* Out of window. F.e. zero window probe. */ 4677 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp))) 4678 goto out_of_window; 4679 4680 tcp_enter_quickack_mode(sk); 4681 4682 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 4683 /* Partial packet, seq < rcv_next < end_seq */ 4684 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n", 4685 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, 4686 TCP_SKB_CB(skb)->end_seq); 4687 4688 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt); 4689 4690 /* If window is closed, drop tail of packet. But after 4691 * remembering D-SACK for its head made in previous line. 4692 */ 4693 if (!tcp_receive_window(tp)) 4694 goto out_of_window; 4695 goto queue_and_out; 4696 } 4697 4698 tcp_data_queue_ofo(sk, skb); 4699 } 4700 4701 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb, 4702 struct sk_buff_head *list) 4703 { 4704 struct sk_buff *next = NULL; 4705 4706 if (!skb_queue_is_last(list, skb)) 4707 next = skb_queue_next(list, skb); 4708 4709 __skb_unlink(skb, list); 4710 __kfree_skb(skb); 4711 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED); 4712 4713 return next; 4714 } 4715 4716 /* Collapse contiguous sequence of skbs head..tail with 4717 * sequence numbers start..end. 4718 * 4719 * If tail is NULL, this means until the end of the list. 4720 * 4721 * Segments with FIN/SYN are not collapsed (only because this 4722 * simplifies code) 4723 */ 4724 static void 4725 tcp_collapse(struct sock *sk, struct sk_buff_head *list, 4726 struct sk_buff *head, struct sk_buff *tail, 4727 u32 start, u32 end) 4728 { 4729 struct sk_buff *skb, *n; 4730 bool end_of_skbs; 4731 4732 /* First, check that queue is collapsible and find 4733 * the point where collapsing can be useful. */ 4734 skb = head; 4735 restart: 4736 end_of_skbs = true; 4737 skb_queue_walk_from_safe(list, skb, n) { 4738 if (skb == tail) 4739 break; 4740 /* No new bits? It is possible on ofo queue. */ 4741 if (!before(start, TCP_SKB_CB(skb)->end_seq)) { 4742 skb = tcp_collapse_one(sk, skb, list); 4743 if (!skb) 4744 break; 4745 goto restart; 4746 } 4747 4748 /* The first skb to collapse is: 4749 * - not SYN/FIN and 4750 * - bloated or contains data before "start" or 4751 * overlaps to the next one. 4752 */ 4753 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) && 4754 (tcp_win_from_space(skb->truesize) > skb->len || 4755 before(TCP_SKB_CB(skb)->seq, start))) { 4756 end_of_skbs = false; 4757 break; 4758 } 4759 4760 if (!skb_queue_is_last(list, skb)) { 4761 struct sk_buff *next = skb_queue_next(list, skb); 4762 if (next != tail && 4763 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(next)->seq) { 4764 end_of_skbs = false; 4765 break; 4766 } 4767 } 4768 4769 /* Decided to skip this, advance start seq. */ 4770 start = TCP_SKB_CB(skb)->end_seq; 4771 } 4772 if (end_of_skbs || 4773 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN))) 4774 return; 4775 4776 while (before(start, end)) { 4777 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start); 4778 struct sk_buff *nskb; 4779 4780 nskb = alloc_skb(copy, GFP_ATOMIC); 4781 if (!nskb) 4782 return; 4783 4784 memcpy(nskb->cb, skb->cb, sizeof(skb->cb)); 4785 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start; 4786 __skb_queue_before(list, skb, nskb); 4787 skb_set_owner_r(nskb, sk); 4788 4789 /* Copy data, releasing collapsed skbs. */ 4790 while (copy > 0) { 4791 int offset = start - TCP_SKB_CB(skb)->seq; 4792 int size = TCP_SKB_CB(skb)->end_seq - start; 4793 4794 BUG_ON(offset < 0); 4795 if (size > 0) { 4796 size = min(copy, size); 4797 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size)) 4798 BUG(); 4799 TCP_SKB_CB(nskb)->end_seq += size; 4800 copy -= size; 4801 start += size; 4802 } 4803 if (!before(start, TCP_SKB_CB(skb)->end_seq)) { 4804 skb = tcp_collapse_one(sk, skb, list); 4805 if (!skb || 4806 skb == tail || 4807 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN))) 4808 return; 4809 } 4810 } 4811 } 4812 } 4813 4814 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs 4815 * and tcp_collapse() them until all the queue is collapsed. 4816 */ 4817 static void tcp_collapse_ofo_queue(struct sock *sk) 4818 { 4819 struct tcp_sock *tp = tcp_sk(sk); 4820 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue); 4821 struct sk_buff *head; 4822 u32 start, end; 4823 4824 if (!skb) 4825 return; 4826 4827 start = TCP_SKB_CB(skb)->seq; 4828 end = TCP_SKB_CB(skb)->end_seq; 4829 head = skb; 4830 4831 for (;;) { 4832 struct sk_buff *next = NULL; 4833 4834 if (!skb_queue_is_last(&tp->out_of_order_queue, skb)) 4835 next = skb_queue_next(&tp->out_of_order_queue, skb); 4836 skb = next; 4837 4838 /* Segment is terminated when we see gap or when 4839 * we are at the end of all the queue. */ 4840 if (!skb || 4841 after(TCP_SKB_CB(skb)->seq, end) || 4842 before(TCP_SKB_CB(skb)->end_seq, start)) { 4843 tcp_collapse(sk, &tp->out_of_order_queue, 4844 head, skb, start, end); 4845 head = skb; 4846 if (!skb) 4847 break; 4848 /* Start new segment */ 4849 start = TCP_SKB_CB(skb)->seq; 4850 end = TCP_SKB_CB(skb)->end_seq; 4851 } else { 4852 if (before(TCP_SKB_CB(skb)->seq, start)) 4853 start = TCP_SKB_CB(skb)->seq; 4854 if (after(TCP_SKB_CB(skb)->end_seq, end)) 4855 end = TCP_SKB_CB(skb)->end_seq; 4856 } 4857 } 4858 } 4859 4860 /* 4861 * Purge the out-of-order queue. 4862 * Return true if queue was pruned. 4863 */ 4864 static bool tcp_prune_ofo_queue(struct sock *sk) 4865 { 4866 struct tcp_sock *tp = tcp_sk(sk); 4867 bool res = false; 4868 4869 if (!skb_queue_empty(&tp->out_of_order_queue)) { 4870 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED); 4871 __skb_queue_purge(&tp->out_of_order_queue); 4872 4873 /* Reset SACK state. A conforming SACK implementation will 4874 * do the same at a timeout based retransmit. When a connection 4875 * is in a sad state like this, we care only about integrity 4876 * of the connection not performance. 4877 */ 4878 if (tp->rx_opt.sack_ok) 4879 tcp_sack_reset(&tp->rx_opt); 4880 sk_mem_reclaim(sk); 4881 res = true; 4882 } 4883 return res; 4884 } 4885 4886 /* Reduce allocated memory if we can, trying to get 4887 * the socket within its memory limits again. 4888 * 4889 * Return less than zero if we should start dropping frames 4890 * until the socket owning process reads some of the data 4891 * to stabilize the situation. 4892 */ 4893 static int tcp_prune_queue(struct sock *sk) 4894 { 4895 struct tcp_sock *tp = tcp_sk(sk); 4896 4897 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq); 4898 4899 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED); 4900 4901 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf) 4902 tcp_clamp_window(sk); 4903 else if (tcp_under_memory_pressure(sk)) 4904 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss); 4905 4906 tcp_collapse_ofo_queue(sk); 4907 if (!skb_queue_empty(&sk->sk_receive_queue)) 4908 tcp_collapse(sk, &sk->sk_receive_queue, 4909 skb_peek(&sk->sk_receive_queue), 4910 NULL, 4911 tp->copied_seq, tp->rcv_nxt); 4912 sk_mem_reclaim(sk); 4913 4914 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) 4915 return 0; 4916 4917 /* Collapsing did not help, destructive actions follow. 4918 * This must not ever occur. */ 4919 4920 tcp_prune_ofo_queue(sk); 4921 4922 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) 4923 return 0; 4924 4925 /* If we are really being abused, tell the caller to silently 4926 * drop receive data on the floor. It will get retransmitted 4927 * and hopefully then we'll have sufficient space. 4928 */ 4929 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED); 4930 4931 /* Massive buffer overcommit. */ 4932 tp->pred_flags = 0; 4933 return -1; 4934 } 4935 4936 static bool tcp_should_expand_sndbuf(const struct sock *sk) 4937 { 4938 const struct tcp_sock *tp = tcp_sk(sk); 4939 4940 /* If the user specified a specific send buffer setting, do 4941 * not modify it. 4942 */ 4943 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK) 4944 return false; 4945 4946 /* If we are under global TCP memory pressure, do not expand. */ 4947 if (tcp_under_memory_pressure(sk)) 4948 return false; 4949 4950 /* If we are under soft global TCP memory pressure, do not expand. */ 4951 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0)) 4952 return false; 4953 4954 /* If we filled the congestion window, do not expand. */ 4955 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) 4956 return false; 4957 4958 return true; 4959 } 4960 4961 /* When incoming ACK allowed to free some skb from write_queue, 4962 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket 4963 * on the exit from tcp input handler. 4964 * 4965 * PROBLEM: sndbuf expansion does not work well with largesend. 4966 */ 4967 static void tcp_new_space(struct sock *sk) 4968 { 4969 struct tcp_sock *tp = tcp_sk(sk); 4970 4971 if (tcp_should_expand_sndbuf(sk)) { 4972 tcp_sndbuf_expand(sk); 4973 tp->snd_cwnd_stamp = tcp_time_stamp; 4974 } 4975 4976 sk->sk_write_space(sk); 4977 } 4978 4979 static void tcp_check_space(struct sock *sk) 4980 { 4981 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) { 4982 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK); 4983 /* pairs with tcp_poll() */ 4984 smp_mb__after_atomic(); 4985 if (sk->sk_socket && 4986 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) 4987 tcp_new_space(sk); 4988 } 4989 } 4990 4991 static inline void tcp_data_snd_check(struct sock *sk) 4992 { 4993 tcp_push_pending_frames(sk); 4994 tcp_check_space(sk); 4995 } 4996 4997 /* 4998 * Check if sending an ack is needed. 4999 */ 5000 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible) 5001 { 5002 struct tcp_sock *tp = tcp_sk(sk); 5003 5004 /* More than one full frame received... */ 5005 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss && 5006 /* ... and right edge of window advances far enough. 5007 * (tcp_recvmsg() will send ACK otherwise). Or... 5008 */ 5009 __tcp_select_window(sk) >= tp->rcv_wnd) || 5010 /* We ACK each frame or... */ 5011 tcp_in_quickack_mode(sk) || 5012 /* We have out of order data. */ 5013 (ofo_possible && skb_peek(&tp->out_of_order_queue))) { 5014 /* Then ack it now */ 5015 tcp_send_ack(sk); 5016 } else { 5017 /* Else, send delayed ack. */ 5018 tcp_send_delayed_ack(sk); 5019 } 5020 } 5021 5022 static inline void tcp_ack_snd_check(struct sock *sk) 5023 { 5024 if (!inet_csk_ack_scheduled(sk)) { 5025 /* We sent a data segment already. */ 5026 return; 5027 } 5028 __tcp_ack_snd_check(sk, 1); 5029 } 5030 5031 /* 5032 * This routine is only called when we have urgent data 5033 * signaled. Its the 'slow' part of tcp_urg. It could be 5034 * moved inline now as tcp_urg is only called from one 5035 * place. We handle URGent data wrong. We have to - as 5036 * BSD still doesn't use the correction from RFC961. 5037 * For 1003.1g we should support a new option TCP_STDURG to permit 5038 * either form (or just set the sysctl tcp_stdurg). 5039 */ 5040 5041 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th) 5042 { 5043 struct tcp_sock *tp = tcp_sk(sk); 5044 u32 ptr = ntohs(th->urg_ptr); 5045 5046 if (ptr && !sysctl_tcp_stdurg) 5047 ptr--; 5048 ptr += ntohl(th->seq); 5049 5050 /* Ignore urgent data that we've already seen and read. */ 5051 if (after(tp->copied_seq, ptr)) 5052 return; 5053 5054 /* Do not replay urg ptr. 5055 * 5056 * NOTE: interesting situation not covered by specs. 5057 * Misbehaving sender may send urg ptr, pointing to segment, 5058 * which we already have in ofo queue. We are not able to fetch 5059 * such data and will stay in TCP_URG_NOTYET until will be eaten 5060 * by recvmsg(). Seems, we are not obliged to handle such wicked 5061 * situations. But it is worth to think about possibility of some 5062 * DoSes using some hypothetical application level deadlock. 5063 */ 5064 if (before(ptr, tp->rcv_nxt)) 5065 return; 5066 5067 /* Do we already have a newer (or duplicate) urgent pointer? */ 5068 if (tp->urg_data && !after(ptr, tp->urg_seq)) 5069 return; 5070 5071 /* Tell the world about our new urgent pointer. */ 5072 sk_send_sigurg(sk); 5073 5074 /* We may be adding urgent data when the last byte read was 5075 * urgent. To do this requires some care. We cannot just ignore 5076 * tp->copied_seq since we would read the last urgent byte again 5077 * as data, nor can we alter copied_seq until this data arrives 5078 * or we break the semantics of SIOCATMARK (and thus sockatmark()) 5079 * 5080 * NOTE. Double Dutch. Rendering to plain English: author of comment 5081 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB); 5082 * and expect that both A and B disappear from stream. This is _wrong_. 5083 * Though this happens in BSD with high probability, this is occasional. 5084 * Any application relying on this is buggy. Note also, that fix "works" 5085 * only in this artificial test. Insert some normal data between A and B and we will 5086 * decline of BSD again. Verdict: it is better to remove to trap 5087 * buggy users. 5088 */ 5089 if (tp->urg_seq == tp->copied_seq && tp->urg_data && 5090 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) { 5091 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue); 5092 tp->copied_seq++; 5093 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) { 5094 __skb_unlink(skb, &sk->sk_receive_queue); 5095 __kfree_skb(skb); 5096 } 5097 } 5098 5099 tp->urg_data = TCP_URG_NOTYET; 5100 tp->urg_seq = ptr; 5101 5102 /* Disable header prediction. */ 5103 tp->pred_flags = 0; 5104 } 5105 5106 /* This is the 'fast' part of urgent handling. */ 5107 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th) 5108 { 5109 struct tcp_sock *tp = tcp_sk(sk); 5110 5111 /* Check if we get a new urgent pointer - normally not. */ 5112 if (th->urg) 5113 tcp_check_urg(sk, th); 5114 5115 /* Do we wait for any urgent data? - normally not... */ 5116 if (tp->urg_data == TCP_URG_NOTYET) { 5117 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) - 5118 th->syn; 5119 5120 /* Is the urgent pointer pointing into this packet? */ 5121 if (ptr < skb->len) { 5122 u8 tmp; 5123 if (skb_copy_bits(skb, ptr, &tmp, 1)) 5124 BUG(); 5125 tp->urg_data = TCP_URG_VALID | tmp; 5126 if (!sock_flag(sk, SOCK_DEAD)) 5127 sk->sk_data_ready(sk); 5128 } 5129 } 5130 } 5131 5132 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen) 5133 { 5134 struct tcp_sock *tp = tcp_sk(sk); 5135 int chunk = skb->len - hlen; 5136 int err; 5137 5138 local_bh_enable(); 5139 if (skb_csum_unnecessary(skb)) 5140 err = skb_copy_datagram_msg(skb, hlen, tp->ucopy.msg, chunk); 5141 else 5142 err = skb_copy_and_csum_datagram_msg(skb, hlen, tp->ucopy.msg); 5143 5144 if (!err) { 5145 tp->ucopy.len -= chunk; 5146 tp->copied_seq += chunk; 5147 tcp_rcv_space_adjust(sk); 5148 } 5149 5150 local_bh_disable(); 5151 return err; 5152 } 5153 5154 static __sum16 __tcp_checksum_complete_user(struct sock *sk, 5155 struct sk_buff *skb) 5156 { 5157 __sum16 result; 5158 5159 if (sock_owned_by_user(sk)) { 5160 local_bh_enable(); 5161 result = __tcp_checksum_complete(skb); 5162 local_bh_disable(); 5163 } else { 5164 result = __tcp_checksum_complete(skb); 5165 } 5166 return result; 5167 } 5168 5169 static inline bool tcp_checksum_complete_user(struct sock *sk, 5170 struct sk_buff *skb) 5171 { 5172 return !skb_csum_unnecessary(skb) && 5173 __tcp_checksum_complete_user(sk, skb); 5174 } 5175 5176 /* Does PAWS and seqno based validation of an incoming segment, flags will 5177 * play significant role here. 5178 */ 5179 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb, 5180 const struct tcphdr *th, int syn_inerr) 5181 { 5182 struct tcp_sock *tp = tcp_sk(sk); 5183 5184 /* RFC1323: H1. Apply PAWS check first. */ 5185 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp && 5186 tcp_paws_discard(sk, skb)) { 5187 if (!th->rst) { 5188 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED); 5189 if (!tcp_oow_rate_limited(sock_net(sk), skb, 5190 LINUX_MIB_TCPACKSKIPPEDPAWS, 5191 &tp->last_oow_ack_time)) 5192 tcp_send_dupack(sk, skb); 5193 goto discard; 5194 } 5195 /* Reset is accepted even if it did not pass PAWS. */ 5196 } 5197 5198 /* Step 1: check sequence number */ 5199 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) { 5200 /* RFC793, page 37: "In all states except SYN-SENT, all reset 5201 * (RST) segments are validated by checking their SEQ-fields." 5202 * And page 69: "If an incoming segment is not acceptable, 5203 * an acknowledgment should be sent in reply (unless the RST 5204 * bit is set, if so drop the segment and return)". 5205 */ 5206 if (!th->rst) { 5207 if (th->syn) 5208 goto syn_challenge; 5209 if (!tcp_oow_rate_limited(sock_net(sk), skb, 5210 LINUX_MIB_TCPACKSKIPPEDSEQ, 5211 &tp->last_oow_ack_time)) 5212 tcp_send_dupack(sk, skb); 5213 } 5214 goto discard; 5215 } 5216 5217 /* Step 2: check RST bit */ 5218 if (th->rst) { 5219 /* RFC 5961 3.2 : 5220 * If sequence number exactly matches RCV.NXT, then 5221 * RESET the connection 5222 * else 5223 * Send a challenge ACK 5224 */ 5225 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) 5226 tcp_reset(sk); 5227 else 5228 tcp_send_challenge_ack(sk, skb); 5229 goto discard; 5230 } 5231 5232 /* step 3: check security and precedence [ignored] */ 5233 5234 /* step 4: Check for a SYN 5235 * RFC 5961 4.2 : Send a challenge ack 5236 */ 5237 if (th->syn) { 5238 syn_challenge: 5239 if (syn_inerr) 5240 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS); 5241 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE); 5242 tcp_send_challenge_ack(sk, skb); 5243 goto discard; 5244 } 5245 5246 return true; 5247 5248 discard: 5249 __kfree_skb(skb); 5250 return false; 5251 } 5252 5253 /* 5254 * TCP receive function for the ESTABLISHED state. 5255 * 5256 * It is split into a fast path and a slow path. The fast path is 5257 * disabled when: 5258 * - A zero window was announced from us - zero window probing 5259 * is only handled properly in the slow path. 5260 * - Out of order segments arrived. 5261 * - Urgent data is expected. 5262 * - There is no buffer space left 5263 * - Unexpected TCP flags/window values/header lengths are received 5264 * (detected by checking the TCP header against pred_flags) 5265 * - Data is sent in both directions. Fast path only supports pure senders 5266 * or pure receivers (this means either the sequence number or the ack 5267 * value must stay constant) 5268 * - Unexpected TCP option. 5269 * 5270 * When these conditions are not satisfied it drops into a standard 5271 * receive procedure patterned after RFC793 to handle all cases. 5272 * The first three cases are guaranteed by proper pred_flags setting, 5273 * the rest is checked inline. Fast processing is turned on in 5274 * tcp_data_queue when everything is OK. 5275 */ 5276 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb, 5277 const struct tcphdr *th, unsigned int len) 5278 { 5279 struct tcp_sock *tp = tcp_sk(sk); 5280 5281 if (unlikely(!sk->sk_rx_dst)) 5282 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb); 5283 /* 5284 * Header prediction. 5285 * The code loosely follows the one in the famous 5286 * "30 instruction TCP receive" Van Jacobson mail. 5287 * 5288 * Van's trick is to deposit buffers into socket queue 5289 * on a device interrupt, to call tcp_recv function 5290 * on the receive process context and checksum and copy 5291 * the buffer to user space. smart... 5292 * 5293 * Our current scheme is not silly either but we take the 5294 * extra cost of the net_bh soft interrupt processing... 5295 * We do checksum and copy also but from device to kernel. 5296 */ 5297 5298 tp->rx_opt.saw_tstamp = 0; 5299 5300 /* pred_flags is 0xS?10 << 16 + snd_wnd 5301 * if header_prediction is to be made 5302 * 'S' will always be tp->tcp_header_len >> 2 5303 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to 5304 * turn it off (when there are holes in the receive 5305 * space for instance) 5306 * PSH flag is ignored. 5307 */ 5308 5309 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags && 5310 TCP_SKB_CB(skb)->seq == tp->rcv_nxt && 5311 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) { 5312 int tcp_header_len = tp->tcp_header_len; 5313 5314 /* Timestamp header prediction: tcp_header_len 5315 * is automatically equal to th->doff*4 due to pred_flags 5316 * match. 5317 */ 5318 5319 /* Check timestamp */ 5320 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) { 5321 /* No? Slow path! */ 5322 if (!tcp_parse_aligned_timestamp(tp, th)) 5323 goto slow_path; 5324 5325 /* If PAWS failed, check it more carefully in slow path */ 5326 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0) 5327 goto slow_path; 5328 5329 /* DO NOT update ts_recent here, if checksum fails 5330 * and timestamp was corrupted part, it will result 5331 * in a hung connection since we will drop all 5332 * future packets due to the PAWS test. 5333 */ 5334 } 5335 5336 if (len <= tcp_header_len) { 5337 /* Bulk data transfer: sender */ 5338 if (len == tcp_header_len) { 5339 /* Predicted packet is in window by definition. 5340 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 5341 * Hence, check seq<=rcv_wup reduces to: 5342 */ 5343 if (tcp_header_len == 5344 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && 5345 tp->rcv_nxt == tp->rcv_wup) 5346 tcp_store_ts_recent(tp); 5347 5348 /* We know that such packets are checksummed 5349 * on entry. 5350 */ 5351 tcp_ack(sk, skb, 0); 5352 __kfree_skb(skb); 5353 tcp_data_snd_check(sk); 5354 return; 5355 } else { /* Header too small */ 5356 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS); 5357 goto discard; 5358 } 5359 } else { 5360 int eaten = 0; 5361 bool fragstolen = false; 5362 5363 if (tp->ucopy.task == current && 5364 tp->copied_seq == tp->rcv_nxt && 5365 len - tcp_header_len <= tp->ucopy.len && 5366 sock_owned_by_user(sk)) { 5367 __set_current_state(TASK_RUNNING); 5368 5369 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len)) { 5370 /* Predicted packet is in window by definition. 5371 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 5372 * Hence, check seq<=rcv_wup reduces to: 5373 */ 5374 if (tcp_header_len == 5375 (sizeof(struct tcphdr) + 5376 TCPOLEN_TSTAMP_ALIGNED) && 5377 tp->rcv_nxt == tp->rcv_wup) 5378 tcp_store_ts_recent(tp); 5379 5380 tcp_rcv_rtt_measure_ts(sk, skb); 5381 5382 __skb_pull(skb, tcp_header_len); 5383 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq); 5384 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER); 5385 eaten = 1; 5386 } 5387 } 5388 if (!eaten) { 5389 if (tcp_checksum_complete_user(sk, skb)) 5390 goto csum_error; 5391 5392 if ((int)skb->truesize > sk->sk_forward_alloc) 5393 goto step5; 5394 5395 /* Predicted packet is in window by definition. 5396 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 5397 * Hence, check seq<=rcv_wup reduces to: 5398 */ 5399 if (tcp_header_len == 5400 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && 5401 tp->rcv_nxt == tp->rcv_wup) 5402 tcp_store_ts_recent(tp); 5403 5404 tcp_rcv_rtt_measure_ts(sk, skb); 5405 5406 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS); 5407 5408 /* Bulk data transfer: receiver */ 5409 eaten = tcp_queue_rcv(sk, skb, tcp_header_len, 5410 &fragstolen); 5411 } 5412 5413 tcp_event_data_recv(sk, skb); 5414 5415 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) { 5416 /* Well, only one small jumplet in fast path... */ 5417 tcp_ack(sk, skb, FLAG_DATA); 5418 tcp_data_snd_check(sk); 5419 if (!inet_csk_ack_scheduled(sk)) 5420 goto no_ack; 5421 } 5422 5423 __tcp_ack_snd_check(sk, 0); 5424 no_ack: 5425 if (eaten) 5426 kfree_skb_partial(skb, fragstolen); 5427 sk->sk_data_ready(sk); 5428 return; 5429 } 5430 } 5431 5432 slow_path: 5433 if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb)) 5434 goto csum_error; 5435 5436 if (!th->ack && !th->rst && !th->syn) 5437 goto discard; 5438 5439 /* 5440 * Standard slow path. 5441 */ 5442 5443 if (!tcp_validate_incoming(sk, skb, th, 1)) 5444 return; 5445 5446 step5: 5447 if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0) 5448 goto discard; 5449 5450 tcp_rcv_rtt_measure_ts(sk, skb); 5451 5452 /* Process urgent data. */ 5453 tcp_urg(sk, skb, th); 5454 5455 /* step 7: process the segment text */ 5456 tcp_data_queue(sk, skb); 5457 5458 tcp_data_snd_check(sk); 5459 tcp_ack_snd_check(sk); 5460 return; 5461 5462 csum_error: 5463 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_CSUMERRORS); 5464 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS); 5465 5466 discard: 5467 __kfree_skb(skb); 5468 } 5469 EXPORT_SYMBOL(tcp_rcv_established); 5470 5471 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb) 5472 { 5473 struct tcp_sock *tp = tcp_sk(sk); 5474 struct inet_connection_sock *icsk = inet_csk(sk); 5475 5476 tcp_set_state(sk, TCP_ESTABLISHED); 5477 5478 if (skb) { 5479 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb); 5480 security_inet_conn_established(sk, skb); 5481 } 5482 5483 /* Make sure socket is routed, for correct metrics. */ 5484 icsk->icsk_af_ops->rebuild_header(sk); 5485 5486 tcp_init_metrics(sk); 5487 5488 tcp_init_congestion_control(sk); 5489 5490 /* Prevent spurious tcp_cwnd_restart() on first data 5491 * packet. 5492 */ 5493 tp->lsndtime = tcp_time_stamp; 5494 5495 tcp_init_buffer_space(sk); 5496 5497 if (sock_flag(sk, SOCK_KEEPOPEN)) 5498 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp)); 5499 5500 if (!tp->rx_opt.snd_wscale) 5501 __tcp_fast_path_on(tp, tp->snd_wnd); 5502 else 5503 tp->pred_flags = 0; 5504 5505 if (!sock_flag(sk, SOCK_DEAD)) { 5506 sk->sk_state_change(sk); 5507 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT); 5508 } 5509 } 5510 5511 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack, 5512 struct tcp_fastopen_cookie *cookie) 5513 { 5514 struct tcp_sock *tp = tcp_sk(sk); 5515 struct sk_buff *data = tp->syn_data ? tcp_write_queue_head(sk) : NULL; 5516 u16 mss = tp->rx_opt.mss_clamp, try_exp = 0; 5517 bool syn_drop = false; 5518 5519 if (mss == tp->rx_opt.user_mss) { 5520 struct tcp_options_received opt; 5521 5522 /* Get original SYNACK MSS value if user MSS sets mss_clamp */ 5523 tcp_clear_options(&opt); 5524 opt.user_mss = opt.mss_clamp = 0; 5525 tcp_parse_options(synack, &opt, 0, NULL); 5526 mss = opt.mss_clamp; 5527 } 5528 5529 if (!tp->syn_fastopen) { 5530 /* Ignore an unsolicited cookie */ 5531 cookie->len = -1; 5532 } else if (tp->total_retrans) { 5533 /* SYN timed out and the SYN-ACK neither has a cookie nor 5534 * acknowledges data. Presumably the remote received only 5535 * the retransmitted (regular) SYNs: either the original 5536 * SYN-data or the corresponding SYN-ACK was dropped. 5537 */ 5538 syn_drop = (cookie->len < 0 && data); 5539 } else if (cookie->len < 0 && !tp->syn_data) { 5540 /* We requested a cookie but didn't get it. If we did not use 5541 * the (old) exp opt format then try so next time (try_exp=1). 5542 * Otherwise we go back to use the RFC7413 opt (try_exp=2). 5543 */ 5544 try_exp = tp->syn_fastopen_exp ? 2 : 1; 5545 } 5546 5547 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp); 5548 5549 if (data) { /* Retransmit unacked data in SYN */ 5550 tcp_for_write_queue_from(data, sk) { 5551 if (data == tcp_send_head(sk) || 5552 __tcp_retransmit_skb(sk, data)) 5553 break; 5554 } 5555 tcp_rearm_rto(sk); 5556 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVEFAIL); 5557 return true; 5558 } 5559 tp->syn_data_acked = tp->syn_data; 5560 if (tp->syn_data_acked) 5561 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE); 5562 5563 tcp_fastopen_add_skb(sk, synack); 5564 5565 return false; 5566 } 5567 5568 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb, 5569 const struct tcphdr *th) 5570 { 5571 struct inet_connection_sock *icsk = inet_csk(sk); 5572 struct tcp_sock *tp = tcp_sk(sk); 5573 struct tcp_fastopen_cookie foc = { .len = -1 }; 5574 int saved_clamp = tp->rx_opt.mss_clamp; 5575 5576 tcp_parse_options(skb, &tp->rx_opt, 0, &foc); 5577 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr) 5578 tp->rx_opt.rcv_tsecr -= tp->tsoffset; 5579 5580 if (th->ack) { 5581 /* rfc793: 5582 * "If the state is SYN-SENT then 5583 * first check the ACK bit 5584 * If the ACK bit is set 5585 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send 5586 * a reset (unless the RST bit is set, if so drop 5587 * the segment and return)" 5588 */ 5589 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) || 5590 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) 5591 goto reset_and_undo; 5592 5593 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 5594 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp, 5595 tcp_time_stamp)) { 5596 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED); 5597 goto reset_and_undo; 5598 } 5599 5600 /* Now ACK is acceptable. 5601 * 5602 * "If the RST bit is set 5603 * If the ACK was acceptable then signal the user "error: 5604 * connection reset", drop the segment, enter CLOSED state, 5605 * delete TCB, and return." 5606 */ 5607 5608 if (th->rst) { 5609 tcp_reset(sk); 5610 goto discard; 5611 } 5612 5613 /* rfc793: 5614 * "fifth, if neither of the SYN or RST bits is set then 5615 * drop the segment and return." 5616 * 5617 * See note below! 5618 * --ANK(990513) 5619 */ 5620 if (!th->syn) 5621 goto discard_and_undo; 5622 5623 /* rfc793: 5624 * "If the SYN bit is on ... 5625 * are acceptable then ... 5626 * (our SYN has been ACKed), change the connection 5627 * state to ESTABLISHED..." 5628 */ 5629 5630 tcp_ecn_rcv_synack(tp, th); 5631 5632 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq); 5633 tcp_ack(sk, skb, FLAG_SLOWPATH); 5634 5635 /* Ok.. it's good. Set up sequence numbers and 5636 * move to established. 5637 */ 5638 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; 5639 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; 5640 5641 /* RFC1323: The window in SYN & SYN/ACK segments is 5642 * never scaled. 5643 */ 5644 tp->snd_wnd = ntohs(th->window); 5645 5646 if (!tp->rx_opt.wscale_ok) { 5647 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0; 5648 tp->window_clamp = min(tp->window_clamp, 65535U); 5649 } 5650 5651 if (tp->rx_opt.saw_tstamp) { 5652 tp->rx_opt.tstamp_ok = 1; 5653 tp->tcp_header_len = 5654 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 5655 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; 5656 tcp_store_ts_recent(tp); 5657 } else { 5658 tp->tcp_header_len = sizeof(struct tcphdr); 5659 } 5660 5661 if (tcp_is_sack(tp) && sysctl_tcp_fack) 5662 tcp_enable_fack(tp); 5663 5664 tcp_mtup_init(sk); 5665 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 5666 tcp_initialize_rcv_mss(sk); 5667 5668 /* Remember, tcp_poll() does not lock socket! 5669 * Change state from SYN-SENT only after copied_seq 5670 * is initialized. */ 5671 tp->copied_seq = tp->rcv_nxt; 5672 5673 smp_mb(); 5674 5675 tcp_finish_connect(sk, skb); 5676 5677 if ((tp->syn_fastopen || tp->syn_data) && 5678 tcp_rcv_fastopen_synack(sk, skb, &foc)) 5679 return -1; 5680 5681 if (sk->sk_write_pending || 5682 icsk->icsk_accept_queue.rskq_defer_accept || 5683 icsk->icsk_ack.pingpong) { 5684 /* Save one ACK. Data will be ready after 5685 * several ticks, if write_pending is set. 5686 * 5687 * It may be deleted, but with this feature tcpdumps 5688 * look so _wonderfully_ clever, that I was not able 5689 * to stand against the temptation 8) --ANK 5690 */ 5691 inet_csk_schedule_ack(sk); 5692 icsk->icsk_ack.lrcvtime = tcp_time_stamp; 5693 tcp_enter_quickack_mode(sk); 5694 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 5695 TCP_DELACK_MAX, TCP_RTO_MAX); 5696 5697 discard: 5698 __kfree_skb(skb); 5699 return 0; 5700 } else { 5701 tcp_send_ack(sk); 5702 } 5703 return -1; 5704 } 5705 5706 /* No ACK in the segment */ 5707 5708 if (th->rst) { 5709 /* rfc793: 5710 * "If the RST bit is set 5711 * 5712 * Otherwise (no ACK) drop the segment and return." 5713 */ 5714 5715 goto discard_and_undo; 5716 } 5717 5718 /* PAWS check. */ 5719 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && 5720 tcp_paws_reject(&tp->rx_opt, 0)) 5721 goto discard_and_undo; 5722 5723 if (th->syn) { 5724 /* We see SYN without ACK. It is attempt of 5725 * simultaneous connect with crossed SYNs. 5726 * Particularly, it can be connect to self. 5727 */ 5728 tcp_set_state(sk, TCP_SYN_RECV); 5729 5730 if (tp->rx_opt.saw_tstamp) { 5731 tp->rx_opt.tstamp_ok = 1; 5732 tcp_store_ts_recent(tp); 5733 tp->tcp_header_len = 5734 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 5735 } else { 5736 tp->tcp_header_len = sizeof(struct tcphdr); 5737 } 5738 5739 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; 5740 tp->copied_seq = tp->rcv_nxt; 5741 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; 5742 5743 /* RFC1323: The window in SYN & SYN/ACK segments is 5744 * never scaled. 5745 */ 5746 tp->snd_wnd = ntohs(th->window); 5747 tp->snd_wl1 = TCP_SKB_CB(skb)->seq; 5748 tp->max_window = tp->snd_wnd; 5749 5750 tcp_ecn_rcv_syn(tp, th); 5751 5752 tcp_mtup_init(sk); 5753 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 5754 tcp_initialize_rcv_mss(sk); 5755 5756 tcp_send_synack(sk); 5757 #if 0 5758 /* Note, we could accept data and URG from this segment. 5759 * There are no obstacles to make this (except that we must 5760 * either change tcp_recvmsg() to prevent it from returning data 5761 * before 3WHS completes per RFC793, or employ TCP Fast Open). 5762 * 5763 * However, if we ignore data in ACKless segments sometimes, 5764 * we have no reasons to accept it sometimes. 5765 * Also, seems the code doing it in step6 of tcp_rcv_state_process 5766 * is not flawless. So, discard packet for sanity. 5767 * Uncomment this return to process the data. 5768 */ 5769 return -1; 5770 #else 5771 goto discard; 5772 #endif 5773 } 5774 /* "fifth, if neither of the SYN or RST bits is set then 5775 * drop the segment and return." 5776 */ 5777 5778 discard_and_undo: 5779 tcp_clear_options(&tp->rx_opt); 5780 tp->rx_opt.mss_clamp = saved_clamp; 5781 goto discard; 5782 5783 reset_and_undo: 5784 tcp_clear_options(&tp->rx_opt); 5785 tp->rx_opt.mss_clamp = saved_clamp; 5786 return 1; 5787 } 5788 5789 /* 5790 * This function implements the receiving procedure of RFC 793 for 5791 * all states except ESTABLISHED and TIME_WAIT. 5792 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be 5793 * address independent. 5794 */ 5795 5796 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb) 5797 { 5798 struct tcp_sock *tp = tcp_sk(sk); 5799 struct inet_connection_sock *icsk = inet_csk(sk); 5800 const struct tcphdr *th = tcp_hdr(skb); 5801 struct request_sock *req; 5802 int queued = 0; 5803 bool acceptable; 5804 5805 tp->rx_opt.saw_tstamp = 0; 5806 5807 switch (sk->sk_state) { 5808 case TCP_CLOSE: 5809 goto discard; 5810 5811 case TCP_LISTEN: 5812 if (th->ack) 5813 return 1; 5814 5815 if (th->rst) 5816 goto discard; 5817 5818 if (th->syn) { 5819 if (th->fin) 5820 goto discard; 5821 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0) 5822 return 1; 5823 5824 /* Now we have several options: In theory there is 5825 * nothing else in the frame. KA9Q has an option to 5826 * send data with the syn, BSD accepts data with the 5827 * syn up to the [to be] advertised window and 5828 * Solaris 2.1 gives you a protocol error. For now 5829 * we just ignore it, that fits the spec precisely 5830 * and avoids incompatibilities. It would be nice in 5831 * future to drop through and process the data. 5832 * 5833 * Now that TTCP is starting to be used we ought to 5834 * queue this data. 5835 * But, this leaves one open to an easy denial of 5836 * service attack, and SYN cookies can't defend 5837 * against this problem. So, we drop the data 5838 * in the interest of security over speed unless 5839 * it's still in use. 5840 */ 5841 kfree_skb(skb); 5842 return 0; 5843 } 5844 goto discard; 5845 5846 case TCP_SYN_SENT: 5847 queued = tcp_rcv_synsent_state_process(sk, skb, th); 5848 if (queued >= 0) 5849 return queued; 5850 5851 /* Do step6 onward by hand. */ 5852 tcp_urg(sk, skb, th); 5853 __kfree_skb(skb); 5854 tcp_data_snd_check(sk); 5855 return 0; 5856 } 5857 5858 req = tp->fastopen_rsk; 5859 if (req) { 5860 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV && 5861 sk->sk_state != TCP_FIN_WAIT1); 5862 5863 if (!tcp_check_req(sk, skb, req, true)) 5864 goto discard; 5865 } 5866 5867 if (!th->ack && !th->rst && !th->syn) 5868 goto discard; 5869 5870 if (!tcp_validate_incoming(sk, skb, th, 0)) 5871 return 0; 5872 5873 /* step 5: check the ACK field */ 5874 acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH | 5875 FLAG_UPDATE_TS_RECENT) > 0; 5876 5877 switch (sk->sk_state) { 5878 case TCP_SYN_RECV: 5879 if (!acceptable) 5880 return 1; 5881 5882 if (!tp->srtt_us) 5883 tcp_synack_rtt_meas(sk, req); 5884 5885 /* Once we leave TCP_SYN_RECV, we no longer need req 5886 * so release it. 5887 */ 5888 if (req) { 5889 tp->total_retrans = req->num_retrans; 5890 reqsk_fastopen_remove(sk, req, false); 5891 } else { 5892 /* Make sure socket is routed, for correct metrics. */ 5893 icsk->icsk_af_ops->rebuild_header(sk); 5894 tcp_init_congestion_control(sk); 5895 5896 tcp_mtup_init(sk); 5897 tp->copied_seq = tp->rcv_nxt; 5898 tcp_init_buffer_space(sk); 5899 } 5900 smp_mb(); 5901 tcp_set_state(sk, TCP_ESTABLISHED); 5902 sk->sk_state_change(sk); 5903 5904 /* Note, that this wakeup is only for marginal crossed SYN case. 5905 * Passively open sockets are not waked up, because 5906 * sk->sk_sleep == NULL and sk->sk_socket == NULL. 5907 */ 5908 if (sk->sk_socket) 5909 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT); 5910 5911 tp->snd_una = TCP_SKB_CB(skb)->ack_seq; 5912 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale; 5913 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq); 5914 5915 if (tp->rx_opt.tstamp_ok) 5916 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; 5917 5918 if (req) { 5919 /* Re-arm the timer because data may have been sent out. 5920 * This is similar to the regular data transmission case 5921 * when new data has just been ack'ed. 5922 * 5923 * (TFO) - we could try to be more aggressive and 5924 * retransmitting any data sooner based on when they 5925 * are sent out. 5926 */ 5927 tcp_rearm_rto(sk); 5928 } else 5929 tcp_init_metrics(sk); 5930 5931 tcp_update_pacing_rate(sk); 5932 5933 /* Prevent spurious tcp_cwnd_restart() on first data packet */ 5934 tp->lsndtime = tcp_time_stamp; 5935 5936 tcp_initialize_rcv_mss(sk); 5937 tcp_fast_path_on(tp); 5938 break; 5939 5940 case TCP_FIN_WAIT1: { 5941 struct dst_entry *dst; 5942 int tmo; 5943 5944 /* If we enter the TCP_FIN_WAIT1 state and we are a 5945 * Fast Open socket and this is the first acceptable 5946 * ACK we have received, this would have acknowledged 5947 * our SYNACK so stop the SYNACK timer. 5948 */ 5949 if (req) { 5950 /* Return RST if ack_seq is invalid. 5951 * Note that RFC793 only says to generate a 5952 * DUPACK for it but for TCP Fast Open it seems 5953 * better to treat this case like TCP_SYN_RECV 5954 * above. 5955 */ 5956 if (!acceptable) 5957 return 1; 5958 /* We no longer need the request sock. */ 5959 reqsk_fastopen_remove(sk, req, false); 5960 tcp_rearm_rto(sk); 5961 } 5962 if (tp->snd_una != tp->write_seq) 5963 break; 5964 5965 tcp_set_state(sk, TCP_FIN_WAIT2); 5966 sk->sk_shutdown |= SEND_SHUTDOWN; 5967 5968 dst = __sk_dst_get(sk); 5969 if (dst) 5970 dst_confirm(dst); 5971 5972 if (!sock_flag(sk, SOCK_DEAD)) { 5973 /* Wake up lingering close() */ 5974 sk->sk_state_change(sk); 5975 break; 5976 } 5977 5978 if (tp->linger2 < 0 || 5979 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 5980 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) { 5981 tcp_done(sk); 5982 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA); 5983 return 1; 5984 } 5985 5986 tmo = tcp_fin_time(sk); 5987 if (tmo > TCP_TIMEWAIT_LEN) { 5988 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN); 5989 } else if (th->fin || sock_owned_by_user(sk)) { 5990 /* Bad case. We could lose such FIN otherwise. 5991 * It is not a big problem, but it looks confusing 5992 * and not so rare event. We still can lose it now, 5993 * if it spins in bh_lock_sock(), but it is really 5994 * marginal case. 5995 */ 5996 inet_csk_reset_keepalive_timer(sk, tmo); 5997 } else { 5998 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo); 5999 goto discard; 6000 } 6001 break; 6002 } 6003 6004 case TCP_CLOSING: 6005 if (tp->snd_una == tp->write_seq) { 6006 tcp_time_wait(sk, TCP_TIME_WAIT, 0); 6007 goto discard; 6008 } 6009 break; 6010 6011 case TCP_LAST_ACK: 6012 if (tp->snd_una == tp->write_seq) { 6013 tcp_update_metrics(sk); 6014 tcp_done(sk); 6015 goto discard; 6016 } 6017 break; 6018 } 6019 6020 /* step 6: check the URG bit */ 6021 tcp_urg(sk, skb, th); 6022 6023 /* step 7: process the segment text */ 6024 switch (sk->sk_state) { 6025 case TCP_CLOSE_WAIT: 6026 case TCP_CLOSING: 6027 case TCP_LAST_ACK: 6028 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) 6029 break; 6030 case TCP_FIN_WAIT1: 6031 case TCP_FIN_WAIT2: 6032 /* RFC 793 says to queue data in these states, 6033 * RFC 1122 says we MUST send a reset. 6034 * BSD 4.4 also does reset. 6035 */ 6036 if (sk->sk_shutdown & RCV_SHUTDOWN) { 6037 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 6038 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) { 6039 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA); 6040 tcp_reset(sk); 6041 return 1; 6042 } 6043 } 6044 /* Fall through */ 6045 case TCP_ESTABLISHED: 6046 tcp_data_queue(sk, skb); 6047 queued = 1; 6048 break; 6049 } 6050 6051 /* tcp_data could move socket to TIME-WAIT */ 6052 if (sk->sk_state != TCP_CLOSE) { 6053 tcp_data_snd_check(sk); 6054 tcp_ack_snd_check(sk); 6055 } 6056 6057 if (!queued) { 6058 discard: 6059 __kfree_skb(skb); 6060 } 6061 return 0; 6062 } 6063 EXPORT_SYMBOL(tcp_rcv_state_process); 6064 6065 static inline void pr_drop_req(struct request_sock *req, __u16 port, int family) 6066 { 6067 struct inet_request_sock *ireq = inet_rsk(req); 6068 6069 if (family == AF_INET) 6070 net_dbg_ratelimited("drop open request from %pI4/%u\n", 6071 &ireq->ir_rmt_addr, port); 6072 #if IS_ENABLED(CONFIG_IPV6) 6073 else if (family == AF_INET6) 6074 net_dbg_ratelimited("drop open request from %pI6/%u\n", 6075 &ireq->ir_v6_rmt_addr, port); 6076 #endif 6077 } 6078 6079 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set 6080 * 6081 * If we receive a SYN packet with these bits set, it means a 6082 * network is playing bad games with TOS bits. In order to 6083 * avoid possible false congestion notifications, we disable 6084 * TCP ECN negotiation. 6085 * 6086 * Exception: tcp_ca wants ECN. This is required for DCTCP 6087 * congestion control: Linux DCTCP asserts ECT on all packets, 6088 * including SYN, which is most optimal solution; however, 6089 * others, such as FreeBSD do not. 6090 */ 6091 static void tcp_ecn_create_request(struct request_sock *req, 6092 const struct sk_buff *skb, 6093 const struct sock *listen_sk, 6094 const struct dst_entry *dst) 6095 { 6096 const struct tcphdr *th = tcp_hdr(skb); 6097 const struct net *net = sock_net(listen_sk); 6098 bool th_ecn = th->ece && th->cwr; 6099 bool ect, ecn_ok; 6100 u32 ecn_ok_dst; 6101 6102 if (!th_ecn) 6103 return; 6104 6105 ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield); 6106 ecn_ok_dst = dst_feature(dst, DST_FEATURE_ECN_MASK); 6107 ecn_ok = net->ipv4.sysctl_tcp_ecn || ecn_ok_dst; 6108 6109 if ((!ect && ecn_ok) || tcp_ca_needs_ecn(listen_sk) || 6110 (ecn_ok_dst & DST_FEATURE_ECN_CA)) 6111 inet_rsk(req)->ecn_ok = 1; 6112 } 6113 6114 static void tcp_openreq_init(struct request_sock *req, 6115 const struct tcp_options_received *rx_opt, 6116 struct sk_buff *skb, const struct sock *sk) 6117 { 6118 struct inet_request_sock *ireq = inet_rsk(req); 6119 6120 req->rsk_rcv_wnd = 0; /* So that tcp_send_synack() knows! */ 6121 req->cookie_ts = 0; 6122 tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq; 6123 tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; 6124 skb_mstamp_get(&tcp_rsk(req)->snt_synack); 6125 tcp_rsk(req)->last_oow_ack_time = 0; 6126 req->mss = rx_opt->mss_clamp; 6127 req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0; 6128 ireq->tstamp_ok = rx_opt->tstamp_ok; 6129 ireq->sack_ok = rx_opt->sack_ok; 6130 ireq->snd_wscale = rx_opt->snd_wscale; 6131 ireq->wscale_ok = rx_opt->wscale_ok; 6132 ireq->acked = 0; 6133 ireq->ecn_ok = 0; 6134 ireq->ir_rmt_port = tcp_hdr(skb)->source; 6135 ireq->ir_num = ntohs(tcp_hdr(skb)->dest); 6136 ireq->ir_mark = inet_request_mark(sk, skb); 6137 } 6138 6139 struct request_sock *inet_reqsk_alloc(const struct request_sock_ops *ops, 6140 struct sock *sk_listener, 6141 bool attach_listener) 6142 { 6143 struct request_sock *req = reqsk_alloc(ops, sk_listener, 6144 attach_listener); 6145 6146 if (req) { 6147 struct inet_request_sock *ireq = inet_rsk(req); 6148 6149 kmemcheck_annotate_bitfield(ireq, flags); 6150 ireq->opt = NULL; 6151 atomic64_set(&ireq->ir_cookie, 0); 6152 ireq->ireq_state = TCP_NEW_SYN_RECV; 6153 write_pnet(&ireq->ireq_net, sock_net(sk_listener)); 6154 ireq->ireq_family = sk_listener->sk_family; 6155 } 6156 6157 return req; 6158 } 6159 EXPORT_SYMBOL(inet_reqsk_alloc); 6160 6161 /* 6162 * Return true if a syncookie should be sent 6163 */ 6164 static bool tcp_syn_flood_action(const struct sock *sk, 6165 const struct sk_buff *skb, 6166 const char *proto) 6167 { 6168 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue; 6169 const char *msg = "Dropping request"; 6170 bool want_cookie = false; 6171 struct net *net = sock_net(sk); 6172 6173 #ifdef CONFIG_SYN_COOKIES 6174 if (net->ipv4.sysctl_tcp_syncookies) { 6175 msg = "Sending cookies"; 6176 want_cookie = true; 6177 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES); 6178 } else 6179 #endif 6180 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP); 6181 6182 if (!queue->synflood_warned && 6183 net->ipv4.sysctl_tcp_syncookies != 2 && 6184 xchg(&queue->synflood_warned, 1) == 0) 6185 pr_info("%s: Possible SYN flooding on port %d. %s. Check SNMP counters.\n", 6186 proto, ntohs(tcp_hdr(skb)->dest), msg); 6187 6188 return want_cookie; 6189 } 6190 6191 static void tcp_reqsk_record_syn(const struct sock *sk, 6192 struct request_sock *req, 6193 const struct sk_buff *skb) 6194 { 6195 if (tcp_sk(sk)->save_syn) { 6196 u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb); 6197 u32 *copy; 6198 6199 copy = kmalloc(len + sizeof(u32), GFP_ATOMIC); 6200 if (copy) { 6201 copy[0] = len; 6202 memcpy(©[1], skb_network_header(skb), len); 6203 req->saved_syn = copy; 6204 } 6205 } 6206 } 6207 6208 int tcp_conn_request(struct request_sock_ops *rsk_ops, 6209 const struct tcp_request_sock_ops *af_ops, 6210 struct sock *sk, struct sk_buff *skb) 6211 { 6212 struct tcp_fastopen_cookie foc = { .len = -1 }; 6213 __u32 isn = TCP_SKB_CB(skb)->tcp_tw_isn; 6214 struct tcp_options_received tmp_opt; 6215 struct tcp_sock *tp = tcp_sk(sk); 6216 struct net *net = sock_net(sk); 6217 struct sock *fastopen_sk = NULL; 6218 struct dst_entry *dst = NULL; 6219 struct request_sock *req; 6220 bool want_cookie = false; 6221 struct flowi fl; 6222 6223 /* TW buckets are converted to open requests without 6224 * limitations, they conserve resources and peer is 6225 * evidently real one. 6226 */ 6227 if ((net->ipv4.sysctl_tcp_syncookies == 2 || 6228 inet_csk_reqsk_queue_is_full(sk)) && !isn) { 6229 want_cookie = tcp_syn_flood_action(sk, skb, rsk_ops->slab_name); 6230 if (!want_cookie) 6231 goto drop; 6232 } 6233 6234 6235 /* Accept backlog is full. If we have already queued enough 6236 * of warm entries in syn queue, drop request. It is better than 6237 * clogging syn queue with openreqs with exponentially increasing 6238 * timeout. 6239 */ 6240 if (sk_acceptq_is_full(sk) && inet_csk_reqsk_queue_young(sk) > 1) { 6241 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS); 6242 goto drop; 6243 } 6244 6245 req = inet_reqsk_alloc(rsk_ops, sk, !want_cookie); 6246 if (!req) 6247 goto drop; 6248 6249 tcp_rsk(req)->af_specific = af_ops; 6250 6251 tcp_clear_options(&tmp_opt); 6252 tmp_opt.mss_clamp = af_ops->mss_clamp; 6253 tmp_opt.user_mss = tp->rx_opt.user_mss; 6254 tcp_parse_options(skb, &tmp_opt, 0, want_cookie ? NULL : &foc); 6255 6256 if (want_cookie && !tmp_opt.saw_tstamp) 6257 tcp_clear_options(&tmp_opt); 6258 6259 tmp_opt.tstamp_ok = tmp_opt.saw_tstamp; 6260 tcp_openreq_init(req, &tmp_opt, skb, sk); 6261 6262 /* Note: tcp_v6_init_req() might override ir_iif for link locals */ 6263 inet_rsk(req)->ir_iif = inet_request_bound_dev_if(sk, skb); 6264 6265 af_ops->init_req(req, sk, skb); 6266 6267 if (security_inet_conn_request(sk, skb, req)) 6268 goto drop_and_free; 6269 6270 if (!want_cookie && !isn) { 6271 /* VJ's idea. We save last timestamp seen 6272 * from the destination in peer table, when entering 6273 * state TIME-WAIT, and check against it before 6274 * accepting new connection request. 6275 * 6276 * If "isn" is not zero, this request hit alive 6277 * timewait bucket, so that all the necessary checks 6278 * are made in the function processing timewait state. 6279 */ 6280 if (tcp_death_row.sysctl_tw_recycle) { 6281 bool strict; 6282 6283 dst = af_ops->route_req(sk, &fl, req, &strict); 6284 6285 if (dst && strict && 6286 !tcp_peer_is_proven(req, dst, true, 6287 tmp_opt.saw_tstamp)) { 6288 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSPASSIVEREJECTED); 6289 goto drop_and_release; 6290 } 6291 } 6292 /* Kill the following clause, if you dislike this way. */ 6293 else if (!net->ipv4.sysctl_tcp_syncookies && 6294 (sysctl_max_syn_backlog - inet_csk_reqsk_queue_len(sk) < 6295 (sysctl_max_syn_backlog >> 2)) && 6296 !tcp_peer_is_proven(req, dst, false, 6297 tmp_opt.saw_tstamp)) { 6298 /* Without syncookies last quarter of 6299 * backlog is filled with destinations, 6300 * proven to be alive. 6301 * It means that we continue to communicate 6302 * to destinations, already remembered 6303 * to the moment of synflood. 6304 */ 6305 pr_drop_req(req, ntohs(tcp_hdr(skb)->source), 6306 rsk_ops->family); 6307 goto drop_and_release; 6308 } 6309 6310 isn = af_ops->init_seq(skb); 6311 } 6312 if (!dst) { 6313 dst = af_ops->route_req(sk, &fl, req, NULL); 6314 if (!dst) 6315 goto drop_and_free; 6316 } 6317 6318 tcp_ecn_create_request(req, skb, sk, dst); 6319 6320 if (want_cookie) { 6321 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss); 6322 req->cookie_ts = tmp_opt.tstamp_ok; 6323 if (!tmp_opt.tstamp_ok) 6324 inet_rsk(req)->ecn_ok = 0; 6325 } 6326 6327 tcp_rsk(req)->snt_isn = isn; 6328 tcp_rsk(req)->txhash = net_tx_rndhash(); 6329 tcp_openreq_init_rwin(req, sk, dst); 6330 if (!want_cookie) { 6331 tcp_reqsk_record_syn(sk, req, skb); 6332 fastopen_sk = tcp_try_fastopen(sk, skb, req, &foc, dst); 6333 } 6334 if (fastopen_sk) { 6335 af_ops->send_synack(fastopen_sk, dst, &fl, req, 6336 &foc, false); 6337 /* Add the child socket directly into the accept queue */ 6338 inet_csk_reqsk_queue_add(sk, req, fastopen_sk); 6339 sk->sk_data_ready(sk); 6340 bh_unlock_sock(fastopen_sk); 6341 sock_put(fastopen_sk); 6342 } else { 6343 tcp_rsk(req)->tfo_listener = false; 6344 if (!want_cookie) 6345 inet_csk_reqsk_queue_hash_add(sk, req, TCP_TIMEOUT_INIT); 6346 af_ops->send_synack(sk, dst, &fl, req, 6347 &foc, !want_cookie); 6348 if (want_cookie) 6349 goto drop_and_free; 6350 } 6351 reqsk_put(req); 6352 return 0; 6353 6354 drop_and_release: 6355 dst_release(dst); 6356 drop_and_free: 6357 reqsk_free(req); 6358 drop: 6359 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_LISTENDROPS); 6360 return 0; 6361 } 6362 EXPORT_SYMBOL(tcp_conn_request); 6363