1/* Modified by Broadcom Corp. Portions Copyright (c) Broadcom Corp, 2012. */ 2/* 3 * INET An implementation of the TCP/IP protocol suite for the LINUX 4 * operating system. INET is implemented using the BSD Socket 5 * interface as the means of communication with the user level. 6 * 7 * Implementation of the Transmission Control Protocol(TCP). 8 * 9 * Authors: Ross Biro 10 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> 11 * Mark Evans, <evansmp@uhura.aston.ac.uk> 12 * Corey Minyard <wf-rch!minyard@relay.EU.net> 13 * Florian La Roche, <flla@stud.uni-sb.de> 14 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu> 15 * Linus Torvalds, <torvalds@cs.helsinki.fi> 16 * Alan Cox, <gw4pts@gw4pts.ampr.org> 17 * Matthew Dillon, <dillon@apollo.west.oic.com> 18 * Arnt Gulbrandsen, <agulbra@nvg.unit.no> 19 * Jorge Cwik, <jorge@laser.satlink.net> 20 */ 21 22/* 23 * Changes: 24 * Pedro Roque : Fast Retransmit/Recovery. 25 * Two receive queues. 26 * Retransmit queue handled by TCP. 27 * Better retransmit timer handling. 28 * New congestion avoidance. 29 * Header prediction. 30 * Variable renaming. 31 * 32 * Eric : Fast Retransmit. 33 * Randy Scott : MSS option defines. 34 * Eric Schenk : Fixes to slow start algorithm. 35 * Eric Schenk : Yet another double ACK bug. 36 * Eric Schenk : Delayed ACK bug fixes. 37 * Eric Schenk : Floyd style fast retrans war avoidance. 38 * David S. Miller : Don't allow zero congestion window. 39 * Eric Schenk : Fix retransmitter so that it sends 40 * next packet on ack of previous packet. 41 * Andi Kleen : Moved open_request checking here 42 * and process RSTs for open_requests. 43 * Andi Kleen : Better prune_queue, and other fixes. 44 * Andrey Savochkin: Fix RTT measurements in the presence of 45 * timestamps. 46 * Andrey Savochkin: Check sequence numbers correctly when 47 * removing SACKs due to in sequence incoming 48 * data segments. 49 * Andi Kleen: Make sure we never ack data there is not 50 * enough room for. Also make this condition 51 * a fatal error if it might still happen. 52 * Andi Kleen: Add tcp_measure_rcv_mss to make 53 * connections with MSS<min(MTU,ann. MSS) 54 * work without delayed acks. 55 * Andi Kleen: Process packets with PSH set in the 56 * fast path. 57 * J Hadi Salim: ECN support 58 * Andrei Gurtov, 59 * Pasi Sarolahti, 60 * Panu Kuhlberg: Experimental audit of TCP (re)transmission 61 * engine. Lots of bugs are found. 62 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs 63 */ 64 65#include <linux/mm.h> 66#include <linux/slab.h> 67#include <linux/module.h> 68#include <linux/sysctl.h> 69#include <linux/kernel.h> 70#include <net/dst.h> 71#include <net/tcp.h> 72#include <net/inet_common.h> 73#include <linux/ipsec.h> 74#include <asm/unaligned.h> 75#include <net/netdma.h> 76 77#include <typedefs.h> 78#include <bcmdefs.h> 79 80int sysctl_tcp_timestamps __read_mostly = 1; 81int sysctl_tcp_window_scaling __read_mostly = 1; 82int sysctl_tcp_sack __read_mostly = 1; 83int sysctl_tcp_fack __read_mostly = 1; 84int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH; 85EXPORT_SYMBOL(sysctl_tcp_reordering); 86int sysctl_tcp_ecn __read_mostly = 2; 87EXPORT_SYMBOL(sysctl_tcp_ecn); 88int sysctl_tcp_dsack __read_mostly = 1; 89int sysctl_tcp_app_win __read_mostly = 31; 90int sysctl_tcp_adv_win_scale __read_mostly = 2; 91EXPORT_SYMBOL(sysctl_tcp_adv_win_scale); 92 93int sysctl_tcp_stdurg __read_mostly; 94int sysctl_tcp_rfc1337 __read_mostly; 95int sysctl_tcp_max_orphans __read_mostly = NR_FILE; 96int sysctl_tcp_frto __read_mostly = 2; 97int sysctl_tcp_frto_response __read_mostly; 98int sysctl_tcp_nometrics_save __read_mostly; 99 100int sysctl_tcp_thin_dupack __read_mostly; 101 102int sysctl_tcp_moderate_rcvbuf __read_mostly = 1; 103int sysctl_tcp_abc __read_mostly; 104 105#define FLAG_DATA 0x01 /* Incoming frame contained data. */ 106#define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */ 107#define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */ 108#define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */ 109#define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */ 110#define FLAG_DATA_SACKED 0x20 /* New SACK. */ 111#define FLAG_ECE 0x40 /* ECE in this ACK */ 112#define FLAG_DATA_LOST 0x80 /* SACK detected data lossage. */ 113#define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/ 114#define FLAG_ONLY_ORIG_SACKED 0x200 /* SACKs only non-rexmit sent before RTO */ 115#define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */ 116#define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */ 117#define FLAG_NONHEAD_RETRANS_ACKED 0x1000 /* Non-head rexmitted data was ACKed */ 118#define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */ 119 120#define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED) 121#define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED) 122#define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE) 123#define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED) 124#define FLAG_ANY_PROGRESS (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED) 125 126#define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH) 127#define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH)) 128 129/* Adapt the MSS value used to make delayed ack decision to the 130 * real world. 131 */ 132static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb) 133{ 134 struct inet_connection_sock *icsk = inet_csk(sk); 135 const unsigned int lss = icsk->icsk_ack.last_seg_size; 136 unsigned int len; 137 138 icsk->icsk_ack.last_seg_size = 0; 139 140 /* skb->len may jitter because of SACKs, even if peer 141 * sends good full-sized frames. 142 */ 143 len = skb_shinfo(skb)->gso_size ? : skb->len; 144 if (len >= icsk->icsk_ack.rcv_mss) { 145 icsk->icsk_ack.rcv_mss = len; 146 } else { 147 /* Otherwise, we make more careful check taking into account, 148 * that SACKs block is variable. 149 * 150 * "len" is invariant segment length, including TCP header. 151 */ 152 len += skb->data - skb_transport_header(skb); 153 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) || 154 /* If PSH is not set, packet should be 155 * full sized, provided peer TCP is not badly broken. 156 * This observation (if it is correct 8)) allows 157 * to handle super-low mtu links fairly. 158 */ 159 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) && 160 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) { 161 /* Subtract also invariant (if peer is RFC compliant), 162 * tcp header plus fixed timestamp option length. 163 * Resulting "len" is MSS free of SACK jitter. 164 */ 165 len -= tcp_sk(sk)->tcp_header_len; 166 icsk->icsk_ack.last_seg_size = len; 167 if (len == lss) { 168 icsk->icsk_ack.rcv_mss = len; 169 return; 170 } 171 } 172 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED) 173 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2; 174 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED; 175 } 176} 177 178static void tcp_incr_quickack(struct sock *sk) 179{ 180 struct inet_connection_sock *icsk = inet_csk(sk); 181 unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss); 182 183 if (quickacks == 0) 184 quickacks = 2; 185 if (quickacks > icsk->icsk_ack.quick) 186 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS); 187} 188 189void tcp_enter_quickack_mode(struct sock *sk) 190{ 191 struct inet_connection_sock *icsk = inet_csk(sk); 192 tcp_incr_quickack(sk); 193 icsk->icsk_ack.pingpong = 0; 194 icsk->icsk_ack.ato = TCP_ATO_MIN; 195} 196 197/* Send ACKs quickly, if "quick" count is not exhausted 198 * and the session is not interactive. 199 */ 200 201static inline int tcp_in_quickack_mode(const struct sock *sk) 202{ 203 const struct inet_connection_sock *icsk = inet_csk(sk); 204 return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong; 205} 206 207static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp) 208{ 209 if (tp->ecn_flags & TCP_ECN_OK) 210 tp->ecn_flags |= TCP_ECN_QUEUE_CWR; 211} 212 213static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, struct sk_buff *skb) 214{ 215 if (tcp_hdr(skb)->cwr) 216 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR; 217} 218 219static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp) 220{ 221 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR; 222} 223 224static inline void TCP_ECN_check_ce(struct tcp_sock *tp, struct sk_buff *skb) 225{ 226 if (tp->ecn_flags & TCP_ECN_OK) { 227 if (INET_ECN_is_ce(TCP_SKB_CB(skb)->flags)) 228 tp->ecn_flags |= TCP_ECN_DEMAND_CWR; 229 /* Funny extension: if ECT is not set on a segment, 230 * it is surely retransmit. It is not in ECN RFC, 231 * but Linux follows this rule. */ 232 else if (INET_ECN_is_not_ect((TCP_SKB_CB(skb)->flags))) 233 tcp_enter_quickack_mode((struct sock *)tp); 234 } 235} 236 237static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, struct tcphdr *th) 238{ 239 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr)) 240 tp->ecn_flags &= ~TCP_ECN_OK; 241} 242 243static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, struct tcphdr *th) 244{ 245 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr)) 246 tp->ecn_flags &= ~TCP_ECN_OK; 247} 248 249static inline int TCP_ECN_rcv_ecn_echo(struct tcp_sock *tp, struct tcphdr *th) 250{ 251 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK)) 252 return 1; 253 return 0; 254} 255 256/* Buffer size and advertised window tuning. 257 * 258 * 1. Tuning sk->sk_sndbuf, when connection enters established state. 259 */ 260 261static void tcp_fixup_sndbuf(struct sock *sk) 262{ 263 int sndmem = tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER + 16 + 264 sizeof(struct sk_buff); 265 266 if (sk->sk_sndbuf < 3 * sndmem) 267 sk->sk_sndbuf = min(3 * sndmem, sysctl_tcp_wmem[2]); 268} 269 270/* 2. Tuning advertised window (window_clamp, rcv_ssthresh) 271 * 272 * All tcp_full_space() is split to two parts: "network" buffer, allocated 273 * forward and advertised in receiver window (tp->rcv_wnd) and 274 * "application buffer", required to isolate scheduling/application 275 * latencies from network. 276 * window_clamp is maximal advertised window. It can be less than 277 * tcp_full_space(), in this case tcp_full_space() - window_clamp 278 * is reserved for "application" buffer. The less window_clamp is 279 * the smoother our behaviour from viewpoint of network, but the lower 280 * throughput and the higher sensitivity of the connection to losses. 8) 281 * 282 * rcv_ssthresh is more strict window_clamp used at "slow start" 283 * phase to predict further behaviour of this connection. 284 * It is used for two goals: 285 * - to enforce header prediction at sender, even when application 286 * requires some significant "application buffer". It is check #1. 287 * - to prevent pruning of receive queue because of misprediction 288 * of receiver window. Check #2. 289 * 290 * The scheme does not work when sender sends good segments opening 291 * window and then starts to feed us spaghetti. But it should work 292 * in common situations. Otherwise, we have to rely on queue collapsing. 293 */ 294 295/* Slow part of check#2. */ 296static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb) 297{ 298 struct tcp_sock *tp = tcp_sk(sk); 299 /* Optimize this! */ 300 int truesize = tcp_win_from_space(skb->truesize) >> 1; 301 int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1; 302 303 while (tp->rcv_ssthresh <= window) { 304 if (truesize <= skb->len) 305 return 2 * inet_csk(sk)->icsk_ack.rcv_mss; 306 307 truesize >>= 1; 308 window >>= 1; 309 } 310 return 0; 311} 312 313static void tcp_grow_window(struct sock *sk, struct sk_buff *skb) 314{ 315 struct tcp_sock *tp = tcp_sk(sk); 316 317 /* Check #1 */ 318 if (tp->rcv_ssthresh < tp->window_clamp && 319 (int)tp->rcv_ssthresh < tcp_space(sk) && 320 !tcp_memory_pressure) { 321 int incr; 322 323 /* Check #2. Increase window, if skb with such overhead 324 * will fit to rcvbuf in future. 325 */ 326 if (tcp_win_from_space(skb->truesize) <= skb->len) 327 incr = 2 * tp->advmss; 328 else 329 incr = __tcp_grow_window(sk, skb); 330 331 if (incr) { 332 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr, 333 tp->window_clamp); 334 inet_csk(sk)->icsk_ack.quick |= 1; 335 } 336 } 337} 338 339/* 3. Tuning rcvbuf, when connection enters established state. */ 340 341static void tcp_fixup_rcvbuf(struct sock *sk) 342{ 343 struct tcp_sock *tp = tcp_sk(sk); 344 int rcvmem = tp->advmss + MAX_TCP_HEADER + 16 + sizeof(struct sk_buff); 345 346 /* Try to select rcvbuf so that 4 mss-sized segments 347 * will fit to window and corresponding skbs will fit to our rcvbuf. 348 * (was 3; 4 is minimum to allow fast retransmit to work.) 349 */ 350 while (tcp_win_from_space(rcvmem) < tp->advmss) 351 rcvmem += 128; 352 if (sk->sk_rcvbuf < 4 * rcvmem) 353 sk->sk_rcvbuf = min(4 * rcvmem, sysctl_tcp_rmem[2]); 354} 355 356/* 4. Try to fixup all. It is made immediately after connection enters 357 * established state. 358 */ 359static void tcp_init_buffer_space(struct sock *sk) 360{ 361 struct tcp_sock *tp = tcp_sk(sk); 362 int maxwin; 363 364 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) 365 tcp_fixup_rcvbuf(sk); 366 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK)) 367 tcp_fixup_sndbuf(sk); 368 369 tp->rcvq_space.space = tp->rcv_wnd; 370 371 maxwin = tcp_full_space(sk); 372 373 if (tp->window_clamp >= maxwin) { 374 tp->window_clamp = maxwin; 375 376 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss) 377 tp->window_clamp = max(maxwin - 378 (maxwin >> sysctl_tcp_app_win), 379 4 * tp->advmss); 380 } 381 382 /* Force reservation of one segment. */ 383 if (sysctl_tcp_app_win && 384 tp->window_clamp > 2 * tp->advmss && 385 tp->window_clamp + tp->advmss > maxwin) 386 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss); 387 388 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp); 389 tp->snd_cwnd_stamp = tcp_time_stamp; 390} 391 392/* 5. Recalculate window clamp after socket hit its memory bounds. */ 393static void tcp_clamp_window(struct sock *sk) 394{ 395 struct tcp_sock *tp = tcp_sk(sk); 396 struct inet_connection_sock *icsk = inet_csk(sk); 397 398 icsk->icsk_ack.quick = 0; 399 400 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] && 401 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) && 402 !tcp_memory_pressure && 403 atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) { 404 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc), 405 sysctl_tcp_rmem[2]); 406 } 407 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf) 408 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss); 409} 410 411/* Initialize RCV_MSS value. 412 * RCV_MSS is an our guess about MSS used by the peer. 413 * We haven't any direct information about the MSS. 414 * It's better to underestimate the RCV_MSS rather than overestimate. 415 * Overestimations make us ACKing less frequently than needed. 416 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss(). 417 */ 418void tcp_initialize_rcv_mss(struct sock *sk) 419{ 420 struct tcp_sock *tp = tcp_sk(sk); 421 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache); 422 423 hint = min(hint, tp->rcv_wnd / 2); 424 hint = min(hint, TCP_MSS_DEFAULT); 425 hint = max(hint, TCP_MIN_MSS); 426 427 inet_csk(sk)->icsk_ack.rcv_mss = hint; 428} 429EXPORT_SYMBOL(tcp_initialize_rcv_mss); 430 431/* Receiver "autotuning" code. 432 * 433 * The algorithm for RTT estimation w/o timestamps is based on 434 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL. 435 * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps> 436 * 437 * More detail on this code can be found at 438 * <http://www.psc.edu/~jheffner/senior_thesis.ps>, 439 * though this reference is out of date. A new paper 440 * is pending. 441 */ 442static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep) 443{ 444 u32 new_sample = tp->rcv_rtt_est.rtt; 445 long m = sample; 446 447 if (m == 0) 448 m = 1; 449 450 if (new_sample != 0) { 451 /* If we sample in larger samples in the non-timestamp 452 * case, we could grossly overestimate the RTT especially 453 * with chatty applications or bulk transfer apps which 454 * are stalled on filesystem I/O. 455 * 456 * Also, since we are only going for a minimum in the 457 * non-timestamp case, we do not smooth things out 458 * else with timestamps disabled convergence takes too 459 * long. 460 */ 461 if (!win_dep) { 462 m -= (new_sample >> 3); 463 new_sample += m; 464 } else if (m < new_sample) 465 new_sample = m << 3; 466 } else { 467 /* No previous measure. */ 468 new_sample = m << 3; 469 } 470 471 if (tp->rcv_rtt_est.rtt != new_sample) 472 tp->rcv_rtt_est.rtt = new_sample; 473} 474 475static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp) 476{ 477 if (tp->rcv_rtt_est.time == 0) 478 goto new_measure; 479 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq)) 480 return; 481 tcp_rcv_rtt_update(tp, jiffies - tp->rcv_rtt_est.time, 1); 482 483new_measure: 484 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd; 485 tp->rcv_rtt_est.time = tcp_time_stamp; 486} 487 488static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, 489 const struct sk_buff *skb) 490{ 491 struct tcp_sock *tp = tcp_sk(sk); 492 if (tp->rx_opt.rcv_tsecr && 493 (TCP_SKB_CB(skb)->end_seq - 494 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss)) 495 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0); 496} 497 498/* 499 * This function should be called every time data is copied to user space. 500 * It calculates the appropriate TCP receive buffer space. 501 */ 502void tcp_rcv_space_adjust(struct sock *sk) 503{ 504 struct tcp_sock *tp = tcp_sk(sk); 505 int time; 506 int space; 507 508 if (tp->rcvq_space.time == 0) 509 goto new_measure; 510 511 time = tcp_time_stamp - tp->rcvq_space.time; 512 if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0) 513 return; 514 515 space = 2 * (tp->copied_seq - tp->rcvq_space.seq); 516 517 space = max(tp->rcvq_space.space, space); 518 519 if (tp->rcvq_space.space != space) { 520 int rcvmem; 521 522 tp->rcvq_space.space = space; 523 524 if (sysctl_tcp_moderate_rcvbuf && 525 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) { 526 int new_clamp = space; 527 528 /* Receive space grows, normalize in order to 529 * take into account packet headers and sk_buff 530 * structure overhead. 531 */ 532 space /= tp->advmss; 533 if (!space) 534 space = 1; 535 rcvmem = (tp->advmss + MAX_TCP_HEADER + 536 16 + sizeof(struct sk_buff)); 537 while (tcp_win_from_space(rcvmem) < tp->advmss) 538 rcvmem += 128; 539 space *= rcvmem; 540 space = min(space, sysctl_tcp_rmem[2]); 541 if (space > sk->sk_rcvbuf) { 542 sk->sk_rcvbuf = space; 543 544 /* Make the window clamp follow along. */ 545 tp->window_clamp = new_clamp; 546 } 547 } 548 } 549 550new_measure: 551 tp->rcvq_space.seq = tp->copied_seq; 552 tp->rcvq_space.time = tcp_time_stamp; 553} 554 555/* There is something which you must keep in mind when you analyze the 556 * behavior of the tp->ato delayed ack timeout interval. When a 557 * connection starts up, we want to ack as quickly as possible. The 558 * problem is that "good" TCP's do slow start at the beginning of data 559 * transmission. The means that until we send the first few ACK's the 560 * sender will sit on his end and only queue most of his data, because 561 * he can only send snd_cwnd unacked packets at any given time. For 562 * each ACK we send, he increments snd_cwnd and transmits more of his 563 * queue. -DaveM 564 */ 565static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb) 566{ 567 struct tcp_sock *tp = tcp_sk(sk); 568 struct inet_connection_sock *icsk = inet_csk(sk); 569 u32 now; 570 571 inet_csk_schedule_ack(sk); 572 573 tcp_measure_rcv_mss(sk, skb); 574 575 tcp_rcv_rtt_measure(tp); 576 577 now = tcp_time_stamp; 578 579 if (!icsk->icsk_ack.ato) { 580 /* The _first_ data packet received, initialize 581 * delayed ACK engine. 582 */ 583 tcp_incr_quickack(sk); 584 icsk->icsk_ack.ato = TCP_ATO_MIN; 585 } else { 586 int m = now - icsk->icsk_ack.lrcvtime; 587 588 if (m <= TCP_ATO_MIN / 2) { 589 /* The fastest case is the first. */ 590 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2; 591 } else if (m < icsk->icsk_ack.ato) { 592 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m; 593 if (icsk->icsk_ack.ato > icsk->icsk_rto) 594 icsk->icsk_ack.ato = icsk->icsk_rto; 595 } else if (m > icsk->icsk_rto) { 596 /* Too long gap. Apparently sender failed to 597 * restart window, so that we send ACKs quickly. 598 */ 599 tcp_incr_quickack(sk); 600 sk_mem_reclaim(sk); 601 } 602 } 603 icsk->icsk_ack.lrcvtime = now; 604 605 TCP_ECN_check_ce(tp, skb); 606 607 if (skb->len >= 128) 608 tcp_grow_window(sk, skb); 609} 610 611/* Called to compute a smoothed rtt estimate. The data fed to this 612 * routine either comes from timestamps, or from segments that were 613 * known _not_ to have been retransmitted [see Karn/Partridge 614 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88 615 * piece by Van Jacobson. 616 * NOTE: the next three routines used to be one big routine. 617 * To save cycles in the RFC 1323 implementation it was better to break 618 * it up into three procedures. -- erics 619 */ 620static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt) 621{ 622 struct tcp_sock *tp = tcp_sk(sk); 623 long m = mrtt; /* RTT */ 624 625 /* The following amusing code comes from Jacobson's 626 * article in SIGCOMM '88. Note that rtt and mdev 627 * are scaled versions of rtt and mean deviation. 628 * This is designed to be as fast as possible 629 * m stands for "measurement". 630 * 631 * On a 1990 paper the rto value is changed to: 632 * RTO = rtt + 4 * mdev 633 * 634 * Funny. This algorithm seems to be very broken. 635 * These formulae increase RTO, when it should be decreased, increase 636 * too slowly, when it should be increased quickly, decrease too quickly 637 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely 638 * does not matter how to _calculate_ it. Seems, it was trap 639 * that VJ failed to avoid. 8) 640 */ 641 if (m == 0) 642 m = 1; 643 if (tp->srtt != 0) { 644 m -= (tp->srtt >> 3); /* m is now error in rtt est */ 645 tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */ 646 if (m < 0) { 647 m = -m; /* m is now abs(error) */ 648 m -= (tp->mdev >> 2); /* similar update on mdev */ 649 /* This is similar to one of Eifel findings. 650 * Eifel blocks mdev updates when rtt decreases. 651 * This solution is a bit different: we use finer gain 652 * for mdev in this case (alpha*beta). 653 * Like Eifel it also prevents growth of rto, 654 * but also it limits too fast rto decreases, 655 * happening in pure Eifel. 656 */ 657 if (m > 0) 658 m >>= 3; 659 } else { 660 m -= (tp->mdev >> 2); /* similar update on mdev */ 661 } 662 tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */ 663 if (tp->mdev > tp->mdev_max) { 664 tp->mdev_max = tp->mdev; 665 if (tp->mdev_max > tp->rttvar) 666 tp->rttvar = tp->mdev_max; 667 } 668 if (after(tp->snd_una, tp->rtt_seq)) { 669 if (tp->mdev_max < tp->rttvar) 670 tp->rttvar -= (tp->rttvar - tp->mdev_max) >> 2; 671 tp->rtt_seq = tp->snd_nxt; 672 tp->mdev_max = tcp_rto_min(sk); 673 } 674 } else { 675 /* no previous measure. */ 676 tp->srtt = m << 3; /* take the measured time to be rtt */ 677 tp->mdev = m << 1; /* make sure rto = 3*rtt */ 678 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk)); 679 tp->rtt_seq = tp->snd_nxt; 680 } 681} 682 683/* Calculate rto without backoff. This is the second half of Van Jacobson's 684 * routine referred to above. 685 */ 686static inline void tcp_set_rto(struct sock *sk) 687{ 688 const struct tcp_sock *tp = tcp_sk(sk); 689 /* Old crap is replaced with new one. 8) 690 * 691 * More seriously: 692 * 1. If rtt variance happened to be less 50msec, it is hallucination. 693 * It cannot be less due to utterly erratic ACK generation made 694 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_ 695 * to do with delayed acks, because at cwnd>2 true delack timeout 696 * is invisible. Actually, Linux-2.4 also generates erratic 697 * ACKs in some circumstances. 698 */ 699 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp); 700 701 /* 2. Fixups made earlier cannot be right. 702 * If we do not estimate RTO correctly without them, 703 * all the algo is pure shit and should be replaced 704 * with correct one. It is exactly, which we pretend to do. 705 */ 706 707 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo 708 * guarantees that rto is higher. 709 */ 710 tcp_bound_rto(sk); 711} 712 713/* Save metrics learned by this TCP session. 714 This function is called only, when TCP finishes successfully 715 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE. 716 */ 717void tcp_update_metrics(struct sock *sk) 718{ 719 struct tcp_sock *tp = tcp_sk(sk); 720 struct dst_entry *dst = __sk_dst_get(sk); 721 722 if (sysctl_tcp_nometrics_save) 723 return; 724 725 dst_confirm(dst); 726 727 if (dst && (dst->flags & DST_HOST)) { 728 const struct inet_connection_sock *icsk = inet_csk(sk); 729 int m; 730 unsigned long rtt; 731 732 if (icsk->icsk_backoff || !tp->srtt) { 733 /* This session failed to estimate rtt. Why? 734 * Probably, no packets returned in time. 735 * Reset our results. 736 */ 737 if (!(dst_metric_locked(dst, RTAX_RTT))) 738 dst->metrics[RTAX_RTT - 1] = 0; 739 return; 740 } 741 742 rtt = dst_metric_rtt(dst, RTAX_RTT); 743 m = rtt - tp->srtt; 744 745 /* If newly calculated rtt larger than stored one, 746 * store new one. Otherwise, use EWMA. Remember, 747 * rtt overestimation is always better than underestimation. 748 */ 749 if (!(dst_metric_locked(dst, RTAX_RTT))) { 750 if (m <= 0) 751 set_dst_metric_rtt(dst, RTAX_RTT, tp->srtt); 752 else 753 set_dst_metric_rtt(dst, RTAX_RTT, rtt - (m >> 3)); 754 } 755 756 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) { 757 unsigned long var; 758 if (m < 0) 759 m = -m; 760 761 /* Scale deviation to rttvar fixed point */ 762 m >>= 1; 763 if (m < tp->mdev) 764 m = tp->mdev; 765 766 var = dst_metric_rtt(dst, RTAX_RTTVAR); 767 if (m >= var) 768 var = m; 769 else 770 var -= (var - m) >> 2; 771 772 set_dst_metric_rtt(dst, RTAX_RTTVAR, var); 773 } 774 775 if (tcp_in_initial_slowstart(tp)) { 776 /* Slow start still did not finish. */ 777 if (dst_metric(dst, RTAX_SSTHRESH) && 778 !dst_metric_locked(dst, RTAX_SSTHRESH) && 779 (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH)) 780 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1; 781 if (!dst_metric_locked(dst, RTAX_CWND) && 782 tp->snd_cwnd > dst_metric(dst, RTAX_CWND)) 783 dst->metrics[RTAX_CWND - 1] = tp->snd_cwnd; 784 } else if (tp->snd_cwnd > tp->snd_ssthresh && 785 icsk->icsk_ca_state == TCP_CA_Open) { 786 /* Cong. avoidance phase, cwnd is reliable. */ 787 if (!dst_metric_locked(dst, RTAX_SSTHRESH)) 788 dst->metrics[RTAX_SSTHRESH-1] = 789 max(tp->snd_cwnd >> 1, tp->snd_ssthresh); 790 if (!dst_metric_locked(dst, RTAX_CWND)) 791 dst->metrics[RTAX_CWND-1] = (dst_metric(dst, RTAX_CWND) + tp->snd_cwnd) >> 1; 792 } else { 793 /* Else slow start did not finish, cwnd is non-sense, 794 ssthresh may be also invalid. 795 */ 796 if (!dst_metric_locked(dst, RTAX_CWND)) 797 dst->metrics[RTAX_CWND-1] = (dst_metric(dst, RTAX_CWND) + tp->snd_ssthresh) >> 1; 798 if (dst_metric(dst, RTAX_SSTHRESH) && 799 !dst_metric_locked(dst, RTAX_SSTHRESH) && 800 tp->snd_ssthresh > dst_metric(dst, RTAX_SSTHRESH)) 801 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh; 802 } 803 804 if (!dst_metric_locked(dst, RTAX_REORDERING)) { 805 if (dst_metric(dst, RTAX_REORDERING) < tp->reordering && 806 tp->reordering != sysctl_tcp_reordering) 807 dst->metrics[RTAX_REORDERING-1] = tp->reordering; 808 } 809 } 810} 811 812/* Numbers are taken from RFC3390. 813 * 814 * John Heffner states: 815 * 816 * The RFC specifies a window of no more than 4380 bytes 817 * unless 2*MSS > 4380. Reading the pseudocode in the RFC 818 * is a bit misleading because they use a clamp at 4380 bytes 819 * rather than use a multiplier in the relevant range. 820 */ 821__u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst) 822{ 823 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0); 824 825 if (!cwnd) { 826 if (tp->mss_cache > 1460) 827 cwnd = 2; 828 else 829 cwnd = (tp->mss_cache > 1095) ? 3 : 4; 830 } 831 return min_t(__u32, cwnd, tp->snd_cwnd_clamp); 832} 833 834/* Set slow start threshold and cwnd not falling to slow start */ 835void tcp_enter_cwr(struct sock *sk, const int set_ssthresh) 836{ 837 struct tcp_sock *tp = tcp_sk(sk); 838 const struct inet_connection_sock *icsk = inet_csk(sk); 839 840 tp->prior_ssthresh = 0; 841 tp->bytes_acked = 0; 842 if (icsk->icsk_ca_state < TCP_CA_CWR) { 843 tp->undo_marker = 0; 844 if (set_ssthresh) 845 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 846 tp->snd_cwnd = min(tp->snd_cwnd, 847 tcp_packets_in_flight(tp) + 1U); 848 tp->snd_cwnd_cnt = 0; 849 tp->high_seq = tp->snd_nxt; 850 tp->snd_cwnd_stamp = tcp_time_stamp; 851 TCP_ECN_queue_cwr(tp); 852 853 tcp_set_ca_state(sk, TCP_CA_CWR); 854 } 855} 856 857/* 858 * Packet counting of FACK is based on in-order assumptions, therefore TCP 859 * disables it when reordering is detected 860 */ 861static void tcp_disable_fack(struct tcp_sock *tp) 862{ 863 /* RFC3517 uses different metric in lost marker => reset on change */ 864 if (tcp_is_fack(tp)) 865 tp->lost_skb_hint = NULL; 866 tp->rx_opt.sack_ok &= ~2; 867} 868 869/* Take a notice that peer is sending D-SACKs */ 870static void tcp_dsack_seen(struct tcp_sock *tp) 871{ 872 tp->rx_opt.sack_ok |= 4; 873} 874 875/* Initialize metrics on socket. */ 876 877static void tcp_init_metrics(struct sock *sk) 878{ 879 struct tcp_sock *tp = tcp_sk(sk); 880 struct dst_entry *dst = __sk_dst_get(sk); 881 882 if (dst == NULL) 883 goto reset; 884 885 dst_confirm(dst); 886 887 if (dst_metric_locked(dst, RTAX_CWND)) 888 tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND); 889 if (dst_metric(dst, RTAX_SSTHRESH)) { 890 tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH); 891 if (tp->snd_ssthresh > tp->snd_cwnd_clamp) 892 tp->snd_ssthresh = tp->snd_cwnd_clamp; 893 } 894 if (dst_metric(dst, RTAX_REORDERING) && 895 tp->reordering != dst_metric(dst, RTAX_REORDERING)) { 896 tcp_disable_fack(tp); 897 tp->reordering = dst_metric(dst, RTAX_REORDERING); 898 } 899 900 if (dst_metric(dst, RTAX_RTT) == 0) 901 goto reset; 902 903 if (!tp->srtt && dst_metric_rtt(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3)) 904 goto reset; 905 906 /* Initial rtt is determined from SYN,SYN-ACK. 907 * The segment is small and rtt may appear much 908 * less than real one. Use per-dst memory 909 * to make it more realistic. 910 * 911 * A bit of theory. RTT is time passed after "normal" sized packet 912 * is sent until it is ACKed. In normal circumstances sending small 913 * packets force peer to delay ACKs and calculation is correct too. 914 * The algorithm is adaptive and, provided we follow specs, it 915 * NEVER underestimate RTT. BUT! If peer tries to make some clever 916 * tricks sort of "quick acks" for time long enough to decrease RTT 917 * to low value, and then abruptly stops to do it and starts to delay 918 * ACKs, wait for troubles. 919 */ 920 if (dst_metric_rtt(dst, RTAX_RTT) > tp->srtt) { 921 tp->srtt = dst_metric_rtt(dst, RTAX_RTT); 922 tp->rtt_seq = tp->snd_nxt; 923 } 924 if (dst_metric_rtt(dst, RTAX_RTTVAR) > tp->mdev) { 925 tp->mdev = dst_metric_rtt(dst, RTAX_RTTVAR); 926 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk)); 927 } 928 tcp_set_rto(sk); 929 if (inet_csk(sk)->icsk_rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp) 930 goto reset; 931 932cwnd: 933 tp->snd_cwnd = tcp_init_cwnd(tp, dst); 934 tp->snd_cwnd_stamp = tcp_time_stamp; 935 return; 936 937reset: 938 /* Play conservative. If timestamps are not 939 * supported, TCP will fail to recalculate correct 940 * rtt, if initial rto is too small. FORGET ALL AND RESET! 941 */ 942 if (!tp->rx_opt.saw_tstamp && tp->srtt) { 943 tp->srtt = 0; 944 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT; 945 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; 946 } 947 goto cwnd; 948} 949 950static void tcp_update_reordering(struct sock *sk, const int metric, 951 const int ts) 952{ 953 struct tcp_sock *tp = tcp_sk(sk); 954 if (metric > tp->reordering) { 955 int mib_idx; 956 957 tp->reordering = min(TCP_MAX_REORDERING, metric); 958 959 /* This exciting event is worth to be remembered. 8) */ 960 if (ts) 961 mib_idx = LINUX_MIB_TCPTSREORDER; 962 else if (tcp_is_reno(tp)) 963 mib_idx = LINUX_MIB_TCPRENOREORDER; 964 else if (tcp_is_fack(tp)) 965 mib_idx = LINUX_MIB_TCPFACKREORDER; 966 else 967 mib_idx = LINUX_MIB_TCPSACKREORDER; 968 969 NET_INC_STATS_BH(sock_net(sk), mib_idx); 970#if FASTRETRANS_DEBUG > 1 971 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n", 972 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state, 973 tp->reordering, 974 tp->fackets_out, 975 tp->sacked_out, 976 tp->undo_marker ? tp->undo_retrans : 0); 977#endif 978 tcp_disable_fack(tp); 979 } 980} 981 982/* This must be called before lost_out is incremented */ 983static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb) 984{ 985 if ((tp->retransmit_skb_hint == NULL) || 986 before(TCP_SKB_CB(skb)->seq, 987 TCP_SKB_CB(tp->retransmit_skb_hint)->seq)) 988 tp->retransmit_skb_hint = skb; 989 990 if (!tp->lost_out || 991 after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high)) 992 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq; 993} 994 995static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb) 996{ 997 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) { 998 tcp_verify_retransmit_hint(tp, skb); 999 1000 tp->lost_out += tcp_skb_pcount(skb); 1001 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 1002 } 1003} 1004 1005static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp, 1006 struct sk_buff *skb) 1007{ 1008 tcp_verify_retransmit_hint(tp, skb); 1009 1010 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) { 1011 tp->lost_out += tcp_skb_pcount(skb); 1012 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 1013 } 1014} 1015 1016/* This procedure tags the retransmission queue when SACKs arrive. 1017 * 1018 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L). 1019 * Packets in queue with these bits set are counted in variables 1020 * sacked_out, retrans_out and lost_out, correspondingly. 1021 * 1022 * Valid combinations are: 1023 * Tag InFlight Description 1024 * 0 1 - orig segment is in flight. 1025 * S 0 - nothing flies, orig reached receiver. 1026 * L 0 - nothing flies, orig lost by net. 1027 * R 2 - both orig and retransmit are in flight. 1028 * L|R 1 - orig is lost, retransmit is in flight. 1029 * S|R 1 - orig reached receiver, retrans is still in flight. 1030 * (L|S|R is logically valid, it could occur when L|R is sacked, 1031 * but it is equivalent to plain S and code short-curcuits it to S. 1032 * L|S is logically invalid, it would mean -1 packet in flight 8)) 1033 * 1034 * These 6 states form finite state machine, controlled by the following events: 1035 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue()) 1036 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue()) 1037 * 3. Loss detection event of one of three flavors: 1038 * A. Scoreboard estimator decided the packet is lost. 1039 * A'. Reno "three dupacks" marks head of queue lost. 1040 * A''. Its FACK modfication, head until snd.fack is lost. 1041 * B. SACK arrives sacking data transmitted after never retransmitted 1042 * hole was sent out. 1043 * C. SACK arrives sacking SND.NXT at the moment, when the 1044 * segment was retransmitted. 1045 * 4. D-SACK added new rule: D-SACK changes any tag to S. 1046 * 1047 * It is pleasant to note, that state diagram turns out to be commutative, 1048 * so that we are allowed not to be bothered by order of our actions, 1049 * when multiple events arrive simultaneously. (see the function below). 1050 * 1051 * Reordering detection. 1052 * -------------------- 1053 * Reordering metric is maximal distance, which a packet can be displaced 1054 * in packet stream. With SACKs we can estimate it: 1055 * 1056 * 1. SACK fills old hole and the corresponding segment was not 1057 * ever retransmitted -> reordering. Alas, we cannot use it 1058 * when segment was retransmitted. 1059 * 2. The last flaw is solved with D-SACK. D-SACK arrives 1060 * for retransmitted and already SACKed segment -> reordering.. 1061 * Both of these heuristics are not used in Loss state, when we cannot 1062 * account for retransmits accurately. 1063 * 1064 * SACK block validation. 1065 * ---------------------- 1066 * 1067 * SACK block range validation checks that the received SACK block fits to 1068 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT. 1069 * Note that SND.UNA is not included to the range though being valid because 1070 * it means that the receiver is rather inconsistent with itself reporting 1071 * SACK reneging when it should advance SND.UNA. Such SACK block this is 1072 * perfectly valid, however, in light of RFC2018 which explicitly states 1073 * that "SACK block MUST reflect the newest segment. Even if the newest 1074 * segment is going to be discarded ...", not that it looks very clever 1075 * in case of head skb. Due to potentional receiver driven attacks, we 1076 * choose to avoid immediate execution of a walk in write queue due to 1077 * reneging and defer head skb's loss recovery to standard loss recovery 1078 * procedure that will eventually trigger (nothing forbids us doing this). 1079 * 1080 * Implements also blockage to start_seq wrap-around. Problem lies in the 1081 * fact that though start_seq (s) is before end_seq (i.e., not reversed), 1082 * there's no guarantee that it will be before snd_nxt (n). The problem 1083 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt 1084 * wrap (s_w): 1085 * 1086 * <- outs wnd -> <- wrapzone -> 1087 * u e n u_w e_w s n_w 1088 * | | | | | | | 1089 * |<------------+------+----- TCP seqno space --------------+---------->| 1090 * ...-- <2^31 ->| |<--------... 1091 * ...---- >2^31 ------>| |<--------... 1092 * 1093 * Current code wouldn't be vulnerable but it's better still to discard such 1094 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat 1095 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in 1096 * snd_nxt wrap -> snd_una region will then become "well defined", i.e., 1097 * equal to the ideal case (infinite seqno space without wrap caused issues). 1098 * 1099 * With D-SACK the lower bound is extended to cover sequence space below 1100 * SND.UNA down to undo_marker, which is the last point of interest. Yet 1101 * again, D-SACK block must not to go across snd_una (for the same reason as 1102 * for the normal SACK blocks, explained above). But there all simplicity 1103 * ends, TCP might receive valid D-SACKs below that. As long as they reside 1104 * fully below undo_marker they do not affect behavior in anyway and can 1105 * therefore be safely ignored. In rare cases (which are more or less 1106 * theoretical ones), the D-SACK will nicely cross that boundary due to skb 1107 * fragmentation and packet reordering past skb's retransmission. To consider 1108 * them correctly, the acceptable range must be extended even more though 1109 * the exact amount is rather hard to quantify. However, tp->max_window can 1110 * be used as an exaggerated estimate. 1111 */ 1112static int tcp_is_sackblock_valid(struct tcp_sock *tp, int is_dsack, 1113 u32 start_seq, u32 end_seq) 1114{ 1115 /* Too far in future, or reversed (interpretation is ambiguous) */ 1116 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq)) 1117 return 0; 1118 1119 /* Nasty start_seq wrap-around check (see comments above) */ 1120 if (!before(start_seq, tp->snd_nxt)) 1121 return 0; 1122 1123 /* In outstanding window? ...This is valid exit for D-SACKs too. 1124 * start_seq == snd_una is non-sensical (see comments above) 1125 */ 1126 if (after(start_seq, tp->snd_una)) 1127 return 1; 1128 1129 if (!is_dsack || !tp->undo_marker) 1130 return 0; 1131 1132 /* ...Then it's D-SACK, and must reside below snd_una completely */ 1133 if (!after(end_seq, tp->snd_una)) 1134 return 0; 1135 1136 if (!before(start_seq, tp->undo_marker)) 1137 return 1; 1138 1139 /* Too old */ 1140 if (!after(end_seq, tp->undo_marker)) 1141 return 0; 1142 1143 /* Undo_marker boundary crossing (overestimates a lot). Known already: 1144 * start_seq < undo_marker and end_seq >= undo_marker. 1145 */ 1146 return !before(start_seq, end_seq - tp->max_window); 1147} 1148 1149/* Check for lost retransmit. This superb idea is borrowed from "ratehalving". 1150 * Event "C". Later note: FACK people cheated me again 8), we have to account 1151 * for reordering! Ugly, but should help. 1152 * 1153 * Search retransmitted skbs from write_queue that were sent when snd_nxt was 1154 * less than what is now known to be received by the other end (derived from 1155 * highest SACK block). Also calculate the lowest snd_nxt among the remaining 1156 * retransmitted skbs to avoid some costly processing per ACKs. 1157 */ 1158static void tcp_mark_lost_retrans(struct sock *sk) 1159{ 1160 const struct inet_connection_sock *icsk = inet_csk(sk); 1161 struct tcp_sock *tp = tcp_sk(sk); 1162 struct sk_buff *skb; 1163 int cnt = 0; 1164 u32 new_low_seq = tp->snd_nxt; 1165 u32 received_upto = tcp_highest_sack_seq(tp); 1166 1167 if (!tcp_is_fack(tp) || !tp->retrans_out || 1168 !after(received_upto, tp->lost_retrans_low) || 1169 icsk->icsk_ca_state != TCP_CA_Recovery) 1170 return; 1171 1172 tcp_for_write_queue(skb, sk) { 1173 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq; 1174 1175 if (skb == tcp_send_head(sk)) 1176 break; 1177 if (cnt == tp->retrans_out) 1178 break; 1179 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 1180 continue; 1181 1182 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)) 1183 continue; 1184 1185 /* TODO: We would like to get rid of tcp_is_fack(tp) only 1186 * constraint here (see above) but figuring out that at 1187 * least tp->reordering SACK blocks reside between ack_seq 1188 * and received_upto is not easy task to do cheaply with 1189 * the available datastructures. 1190 * 1191 * Whether FACK should check here for tp->reordering segs 1192 * in-between one could argue for either way (it would be 1193 * rather simple to implement as we could count fack_count 1194 * during the walk and do tp->fackets_out - fack_count). 1195 */ 1196 if (after(received_upto, ack_seq)) { 1197 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 1198 tp->retrans_out -= tcp_skb_pcount(skb); 1199 1200 tcp_skb_mark_lost_uncond_verify(tp, skb); 1201 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT); 1202 } else { 1203 if (before(ack_seq, new_low_seq)) 1204 new_low_seq = ack_seq; 1205 cnt += tcp_skb_pcount(skb); 1206 } 1207 } 1208 1209 if (tp->retrans_out) 1210 tp->lost_retrans_low = new_low_seq; 1211} 1212 1213static int tcp_check_dsack(struct sock *sk, struct sk_buff *ack_skb, 1214 struct tcp_sack_block_wire *sp, int num_sacks, 1215 u32 prior_snd_una) 1216{ 1217 struct tcp_sock *tp = tcp_sk(sk); 1218 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq); 1219 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq); 1220 int dup_sack = 0; 1221 1222 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) { 1223 dup_sack = 1; 1224 tcp_dsack_seen(tp); 1225 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV); 1226 } else if (num_sacks > 1) { 1227 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq); 1228 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq); 1229 1230 if (!after(end_seq_0, end_seq_1) && 1231 !before(start_seq_0, start_seq_1)) { 1232 dup_sack = 1; 1233 tcp_dsack_seen(tp); 1234 NET_INC_STATS_BH(sock_net(sk), 1235 LINUX_MIB_TCPDSACKOFORECV); 1236 } 1237 } 1238 1239 /* D-SACK for already forgotten data... Do dumb counting. */ 1240 if (dup_sack && 1241 !after(end_seq_0, prior_snd_una) && 1242 after(end_seq_0, tp->undo_marker)) 1243 tp->undo_retrans--; 1244 1245 return dup_sack; 1246} 1247 1248struct tcp_sacktag_state { 1249 int reord; 1250 int fack_count; 1251 int flag; 1252}; 1253 1254static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb, 1255 u32 start_seq, u32 end_seq) 1256{ 1257 int in_sack, err; 1258 unsigned int pkt_len; 1259 unsigned int mss; 1260 1261 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && 1262 !before(end_seq, TCP_SKB_CB(skb)->end_seq); 1263 1264 if (tcp_skb_pcount(skb) > 1 && !in_sack && 1265 after(TCP_SKB_CB(skb)->end_seq, start_seq)) { 1266 mss = tcp_skb_mss(skb); 1267 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); 1268 1269 if (!in_sack) { 1270 pkt_len = start_seq - TCP_SKB_CB(skb)->seq; 1271 if (pkt_len < mss) 1272 pkt_len = mss; 1273 } else { 1274 pkt_len = end_seq - TCP_SKB_CB(skb)->seq; 1275 if (pkt_len < mss) 1276 return -EINVAL; 1277 } 1278 1279 /* Round if necessary so that SACKs cover only full MSSes 1280 * and/or the remaining small portion (if present) 1281 */ 1282 if (pkt_len > mss) { 1283 unsigned int new_len = (pkt_len / mss) * mss; 1284 if (!in_sack && new_len < pkt_len) { 1285 new_len += mss; 1286 if (new_len > skb->len) 1287 return 0; 1288 } 1289 pkt_len = new_len; 1290 } 1291 err = tcp_fragment(sk, skb, pkt_len, mss); 1292 if (err < 0) 1293 return err; 1294 } 1295 1296 return in_sack; 1297} 1298 1299static u8 tcp_sacktag_one(struct sk_buff *skb, struct sock *sk, 1300 struct tcp_sacktag_state *state, 1301 int dup_sack, int pcount) 1302{ 1303 struct tcp_sock *tp = tcp_sk(sk); 1304 u8 sacked = TCP_SKB_CB(skb)->sacked; 1305 int fack_count = state->fack_count; 1306 1307 /* Account D-SACK for retransmitted packet. */ 1308 if (dup_sack && (sacked & TCPCB_RETRANS)) { 1309 if (after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker)) 1310 tp->undo_retrans--; 1311 if (sacked & TCPCB_SACKED_ACKED) 1312 state->reord = min(fack_count, state->reord); 1313 } 1314 1315 /* Nothing to do; acked frame is about to be dropped (was ACKed). */ 1316 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 1317 return sacked; 1318 1319 if (!(sacked & TCPCB_SACKED_ACKED)) { 1320 if (sacked & TCPCB_SACKED_RETRANS) { 1321 /* If the segment is not tagged as lost, 1322 * we do not clear RETRANS, believing 1323 * that retransmission is still in flight. 1324 */ 1325 if (sacked & TCPCB_LOST) { 1326 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS); 1327 tp->lost_out -= pcount; 1328 tp->retrans_out -= pcount; 1329 } 1330 } else { 1331 if (!(sacked & TCPCB_RETRANS)) { 1332 /* New sack for not retransmitted frame, 1333 * which was in hole. It is reordering. 1334 */ 1335 if (before(TCP_SKB_CB(skb)->seq, 1336 tcp_highest_sack_seq(tp))) 1337 state->reord = min(fack_count, 1338 state->reord); 1339 1340 /* SACK enhanced F-RTO (RFC4138; Appendix B) */ 1341 if (!after(TCP_SKB_CB(skb)->end_seq, tp->frto_highmark)) 1342 state->flag |= FLAG_ONLY_ORIG_SACKED; 1343 } 1344 1345 if (sacked & TCPCB_LOST) { 1346 sacked &= ~TCPCB_LOST; 1347 tp->lost_out -= pcount; 1348 } 1349 } 1350 1351 sacked |= TCPCB_SACKED_ACKED; 1352 state->flag |= FLAG_DATA_SACKED; 1353 tp->sacked_out += pcount; 1354 1355 fack_count += pcount; 1356 1357 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */ 1358 if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) && 1359 before(TCP_SKB_CB(skb)->seq, 1360 TCP_SKB_CB(tp->lost_skb_hint)->seq)) 1361 tp->lost_cnt_hint += pcount; 1362 1363 if (fack_count > tp->fackets_out) 1364 tp->fackets_out = fack_count; 1365 } 1366 1367 /* D-SACK. We can detect redundant retransmission in S|R and plain R 1368 * frames and clear it. undo_retrans is decreased above, L|R frames 1369 * are accounted above as well. 1370 */ 1371 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) { 1372 sacked &= ~TCPCB_SACKED_RETRANS; 1373 tp->retrans_out -= pcount; 1374 } 1375 1376 return sacked; 1377} 1378 1379static int tcp_shifted_skb(struct sock *sk, struct sk_buff *skb, 1380 struct tcp_sacktag_state *state, 1381 unsigned int pcount, int shifted, int mss, 1382 int dup_sack) 1383{ 1384 struct tcp_sock *tp = tcp_sk(sk); 1385 struct sk_buff *prev = tcp_write_queue_prev(sk, skb); 1386 1387 BUG_ON(!pcount); 1388 1389 /* Tweak before seqno plays */ 1390 if (!tcp_is_fack(tp) && tcp_is_sack(tp) && tp->lost_skb_hint && 1391 !before(TCP_SKB_CB(tp->lost_skb_hint)->seq, TCP_SKB_CB(skb)->seq)) 1392 tp->lost_cnt_hint += pcount; 1393 1394 TCP_SKB_CB(prev)->end_seq += shifted; 1395 TCP_SKB_CB(skb)->seq += shifted; 1396 1397 skb_shinfo(prev)->gso_segs += pcount; 1398 BUG_ON(skb_shinfo(skb)->gso_segs < pcount); 1399 skb_shinfo(skb)->gso_segs -= pcount; 1400 1401 /* When we're adding to gso_segs == 1, gso_size will be zero, 1402 * in theory this shouldn't be necessary but as long as DSACK 1403 * code can come after this skb later on it's better to keep 1404 * setting gso_size to something. 1405 */ 1406 if (!skb_shinfo(prev)->gso_size) { 1407 skb_shinfo(prev)->gso_size = mss; 1408 skb_shinfo(prev)->gso_type = sk->sk_gso_type; 1409 } 1410 1411 /* CHECKME: To clear or not to clear? Mimics normal skb currently */ 1412 if (skb_shinfo(skb)->gso_segs <= 1) { 1413 skb_shinfo(skb)->gso_size = 0; 1414 skb_shinfo(skb)->gso_type = 0; 1415 } 1416 1417 /* We discard results */ 1418 tcp_sacktag_one(skb, sk, state, dup_sack, pcount); 1419 1420 /* Difference in this won't matter, both ACKed by the same cumul. ACK */ 1421 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS); 1422 1423 if (skb->len > 0) { 1424 BUG_ON(!tcp_skb_pcount(skb)); 1425 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED); 1426 return 0; 1427 } 1428 1429 /* Whole SKB was eaten :-) */ 1430 1431 if (skb == tp->retransmit_skb_hint) 1432 tp->retransmit_skb_hint = prev; 1433 if (skb == tp->scoreboard_skb_hint) 1434 tp->scoreboard_skb_hint = prev; 1435 if (skb == tp->lost_skb_hint) { 1436 tp->lost_skb_hint = prev; 1437 tp->lost_cnt_hint -= tcp_skb_pcount(prev); 1438 } 1439 1440 TCP_SKB_CB(skb)->flags |= TCP_SKB_CB(prev)->flags; 1441 if (skb == tcp_highest_sack(sk)) 1442 tcp_advance_highest_sack(sk, skb); 1443 1444 tcp_unlink_write_queue(skb, sk); 1445 sk_wmem_free_skb(sk, skb); 1446 1447 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED); 1448 1449 return 1; 1450} 1451 1452/* I wish gso_size would have a bit more sane initialization than 1453 * something-or-zero which complicates things 1454 */ 1455static int tcp_skb_seglen(struct sk_buff *skb) 1456{ 1457 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb); 1458} 1459 1460/* Shifting pages past head area doesn't work */ 1461static int skb_can_shift(struct sk_buff *skb) 1462{ 1463 return !skb_headlen(skb) && skb_is_nonlinear(skb); 1464} 1465 1466/* Try collapsing SACK blocks spanning across multiple skbs to a single 1467 * skb. 1468 */ 1469static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb, 1470 struct tcp_sacktag_state *state, 1471 u32 start_seq, u32 end_seq, 1472 int dup_sack) 1473{ 1474 struct tcp_sock *tp = tcp_sk(sk); 1475 struct sk_buff *prev; 1476 int mss; 1477 int pcount = 0; 1478 int len; 1479 int in_sack; 1480 1481 if (!sk_can_gso(sk)) 1482 goto fallback; 1483 1484 /* Normally R but no L won't result in plain S */ 1485 if (!dup_sack && 1486 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS) 1487 goto fallback; 1488 if (!skb_can_shift(skb)) 1489 goto fallback; 1490 /* This frame is about to be dropped (was ACKed). */ 1491 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 1492 goto fallback; 1493 1494 /* Can only happen with delayed DSACK + discard craziness */ 1495 if (unlikely(skb == tcp_write_queue_head(sk))) 1496 goto fallback; 1497 prev = tcp_write_queue_prev(sk, skb); 1498 1499 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) 1500 goto fallback; 1501 1502 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && 1503 !before(end_seq, TCP_SKB_CB(skb)->end_seq); 1504 1505 if (in_sack) { 1506 len = skb->len; 1507 pcount = tcp_skb_pcount(skb); 1508 mss = tcp_skb_seglen(skb); 1509 1510 /* TODO: Fix DSACKs to not fragment already SACKed and we can 1511 * drop this restriction as unnecessary 1512 */ 1513 if (mss != tcp_skb_seglen(prev)) 1514 goto fallback; 1515 } else { 1516 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq)) 1517 goto noop; 1518 /* CHECKME: This is non-MSS split case only?, this will 1519 * cause skipped skbs due to advancing loop btw, original 1520 * has that feature too 1521 */ 1522 if (tcp_skb_pcount(skb) <= 1) 1523 goto noop; 1524 1525 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); 1526 if (!in_sack) { 1527 /* TODO: head merge to next could be attempted here 1528 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)), 1529 * though it might not be worth of the additional hassle 1530 * 1531 * ...we can probably just fallback to what was done 1532 * previously. We could try merging non-SACKed ones 1533 * as well but it probably isn't going to buy off 1534 * because later SACKs might again split them, and 1535 * it would make skb timestamp tracking considerably 1536 * harder problem. 1537 */ 1538 goto fallback; 1539 } 1540 1541 len = end_seq - TCP_SKB_CB(skb)->seq; 1542 BUG_ON(len < 0); 1543 BUG_ON(len > skb->len); 1544 1545 /* MSS boundaries should be honoured or else pcount will 1546 * severely break even though it makes things bit trickier. 1547 * Optimize common case to avoid most of the divides 1548 */ 1549 mss = tcp_skb_mss(skb); 1550 1551 /* TODO: Fix DSACKs to not fragment already SACKed and we can 1552 * drop this restriction as unnecessary 1553 */ 1554 if (mss != tcp_skb_seglen(prev)) 1555 goto fallback; 1556 1557 if (len == mss) { 1558 pcount = 1; 1559 } else if (len < mss) { 1560 goto noop; 1561 } else { 1562 pcount = len / mss; 1563 len = pcount * mss; 1564 } 1565 } 1566 1567 if (!skb_shift(prev, skb, len)) 1568 goto fallback; 1569 if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack)) 1570 goto out; 1571 1572 /* Hole filled allows collapsing with the next as well, this is very 1573 * useful when hole on every nth skb pattern happens 1574 */ 1575 if (prev == tcp_write_queue_tail(sk)) 1576 goto out; 1577 skb = tcp_write_queue_next(sk, prev); 1578 1579 if (!skb_can_shift(skb) || 1580 (skb == tcp_send_head(sk)) || 1581 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) || 1582 (mss != tcp_skb_seglen(skb))) 1583 goto out; 1584 1585 len = skb->len; 1586 if (skb_shift(prev, skb, len)) { 1587 pcount += tcp_skb_pcount(skb); 1588 tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0); 1589 } 1590 1591out: 1592 state->fack_count += pcount; 1593 return prev; 1594 1595noop: 1596 return skb; 1597 1598fallback: 1599 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK); 1600 return NULL; 1601} 1602 1603static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk, 1604 struct tcp_sack_block *next_dup, 1605 struct tcp_sacktag_state *state, 1606 u32 start_seq, u32 end_seq, 1607 int dup_sack_in) 1608{ 1609 struct tcp_sock *tp = tcp_sk(sk); 1610 struct sk_buff *tmp; 1611 1612 tcp_for_write_queue_from(skb, sk) { 1613 int in_sack = 0; 1614 int dup_sack = dup_sack_in; 1615 1616 if (skb == tcp_send_head(sk)) 1617 break; 1618 1619 /* queue is in-order => we can short-circuit the walk early */ 1620 if (!before(TCP_SKB_CB(skb)->seq, end_seq)) 1621 break; 1622 1623 if ((next_dup != NULL) && 1624 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) { 1625 in_sack = tcp_match_skb_to_sack(sk, skb, 1626 next_dup->start_seq, 1627 next_dup->end_seq); 1628 if (in_sack > 0) 1629 dup_sack = 1; 1630 } 1631 1632 /* skb reference here is a bit tricky to get right, since 1633 * shifting can eat and free both this skb and the next, 1634 * so not even _safe variant of the loop is enough. 1635 */ 1636 if (in_sack <= 0) { 1637 tmp = tcp_shift_skb_data(sk, skb, state, 1638 start_seq, end_seq, dup_sack); 1639 if (tmp != NULL) { 1640 if (tmp != skb) { 1641 skb = tmp; 1642 continue; 1643 } 1644 1645 in_sack = 0; 1646 } else { 1647 in_sack = tcp_match_skb_to_sack(sk, skb, 1648 start_seq, 1649 end_seq); 1650 } 1651 } 1652 1653 if (unlikely(in_sack < 0)) 1654 break; 1655 1656 if (in_sack) { 1657 TCP_SKB_CB(skb)->sacked = tcp_sacktag_one(skb, sk, 1658 state, 1659 dup_sack, 1660 tcp_skb_pcount(skb)); 1661 1662 if (!before(TCP_SKB_CB(skb)->seq, 1663 tcp_highest_sack_seq(tp))) 1664 tcp_advance_highest_sack(sk, skb); 1665 } 1666 1667 state->fack_count += tcp_skb_pcount(skb); 1668 } 1669 return skb; 1670} 1671 1672/* Avoid all extra work that is being done by sacktag while walking in 1673 * a normal way 1674 */ 1675static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk, 1676 struct tcp_sacktag_state *state, 1677 u32 skip_to_seq) 1678{ 1679 tcp_for_write_queue_from(skb, sk) { 1680 if (skb == tcp_send_head(sk)) 1681 break; 1682 1683 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq)) 1684 break; 1685 1686 state->fack_count += tcp_skb_pcount(skb); 1687 } 1688 return skb; 1689} 1690 1691static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb, 1692 struct sock *sk, 1693 struct tcp_sack_block *next_dup, 1694 struct tcp_sacktag_state *state, 1695 u32 skip_to_seq) 1696{ 1697 if (next_dup == NULL) 1698 return skb; 1699 1700 if (before(next_dup->start_seq, skip_to_seq)) { 1701 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq); 1702 skb = tcp_sacktag_walk(skb, sk, NULL, state, 1703 next_dup->start_seq, next_dup->end_seq, 1704 1); 1705 } 1706 1707 return skb; 1708} 1709 1710static int tcp_sack_cache_ok(struct tcp_sock *tp, struct tcp_sack_block *cache) 1711{ 1712 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache); 1713} 1714 1715static int 1716tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, 1717 u32 prior_snd_una) 1718{ 1719 const struct inet_connection_sock *icsk = inet_csk(sk); 1720 struct tcp_sock *tp = tcp_sk(sk); 1721 unsigned char *ptr = (skb_transport_header(ack_skb) + 1722 TCP_SKB_CB(ack_skb)->sacked); 1723 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2); 1724 struct tcp_sack_block sp[TCP_NUM_SACKS]; 1725 struct tcp_sack_block *cache; 1726 struct tcp_sacktag_state state; 1727 struct sk_buff *skb; 1728 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3); 1729 int used_sacks; 1730 int found_dup_sack = 0; 1731 int i, j; 1732 int first_sack_index; 1733 1734 state.flag = 0; 1735 state.reord = tp->packets_out; 1736 1737 if (!tp->sacked_out) { 1738 if (WARN_ON(tp->fackets_out)) 1739 tp->fackets_out = 0; 1740 tcp_highest_sack_reset(sk); 1741 } 1742 1743 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire, 1744 num_sacks, prior_snd_una); 1745 if (found_dup_sack) 1746 state.flag |= FLAG_DSACKING_ACK; 1747 1748 /* Eliminate too old ACKs, but take into 1749 * account more or less fresh ones, they can 1750 * contain valid SACK info. 1751 */ 1752 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window)) 1753 return 0; 1754 1755 if (!tp->packets_out) 1756 goto out; 1757 1758 used_sacks = 0; 1759 first_sack_index = 0; 1760 for (i = 0; i < num_sacks; i++) { 1761 int dup_sack = !i && found_dup_sack; 1762 1763 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq); 1764 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq); 1765 1766 if (!tcp_is_sackblock_valid(tp, dup_sack, 1767 sp[used_sacks].start_seq, 1768 sp[used_sacks].end_seq)) { 1769 int mib_idx; 1770 1771 if (dup_sack) { 1772 if (!tp->undo_marker) 1773 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO; 1774 else 1775 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD; 1776 } else { 1777 /* Don't count olds caused by ACK reordering */ 1778 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) && 1779 !after(sp[used_sacks].end_seq, tp->snd_una)) 1780 continue; 1781 mib_idx = LINUX_MIB_TCPSACKDISCARD; 1782 } 1783 1784 NET_INC_STATS_BH(sock_net(sk), mib_idx); 1785 if (i == 0) 1786 first_sack_index = -1; 1787 continue; 1788 } 1789 1790 /* Ignore very old stuff early */ 1791 if (!after(sp[used_sacks].end_seq, prior_snd_una)) 1792 continue; 1793 1794 used_sacks++; 1795 } 1796 1797 /* order SACK blocks to allow in order walk of the retrans queue */ 1798 for (i = used_sacks - 1; i > 0; i--) { 1799 for (j = 0; j < i; j++) { 1800 if (after(sp[j].start_seq, sp[j + 1].start_seq)) { 1801 swap(sp[j], sp[j + 1]); 1802 1803 /* Track where the first SACK block goes to */ 1804 if (j == first_sack_index) 1805 first_sack_index = j + 1; 1806 } 1807 } 1808 } 1809 1810 skb = tcp_write_queue_head(sk); 1811 state.fack_count = 0; 1812 i = 0; 1813 1814 if (!tp->sacked_out) { 1815 /* It's already past, so skip checking against it */ 1816 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache); 1817 } else { 1818 cache = tp->recv_sack_cache; 1819 /* Skip empty blocks in at head of the cache */ 1820 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq && 1821 !cache->end_seq) 1822 cache++; 1823 } 1824 1825 while (i < used_sacks) { 1826 u32 start_seq = sp[i].start_seq; 1827 u32 end_seq = sp[i].end_seq; 1828 int dup_sack = (found_dup_sack && (i == first_sack_index)); 1829 struct tcp_sack_block *next_dup = NULL; 1830 1831 if (found_dup_sack && ((i + 1) == first_sack_index)) 1832 next_dup = &sp[i + 1]; 1833 1834 /* Event "B" in the comment above. */ 1835 if (after(end_seq, tp->high_seq)) 1836 state.flag |= FLAG_DATA_LOST; 1837 1838 /* Skip too early cached blocks */ 1839 while (tcp_sack_cache_ok(tp, cache) && 1840 !before(start_seq, cache->end_seq)) 1841 cache++; 1842 1843 /* Can skip some work by looking recv_sack_cache? */ 1844 if (tcp_sack_cache_ok(tp, cache) && !dup_sack && 1845 after(end_seq, cache->start_seq)) { 1846 1847 /* Head todo? */ 1848 if (before(start_seq, cache->start_seq)) { 1849 skb = tcp_sacktag_skip(skb, sk, &state, 1850 start_seq); 1851 skb = tcp_sacktag_walk(skb, sk, next_dup, 1852 &state, 1853 start_seq, 1854 cache->start_seq, 1855 dup_sack); 1856 } 1857 1858 /* Rest of the block already fully processed? */ 1859 if (!after(end_seq, cache->end_seq)) 1860 goto advance_sp; 1861 1862 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup, 1863 &state, 1864 cache->end_seq); 1865 1866 /* ...tail remains todo... */ 1867 if (tcp_highest_sack_seq(tp) == cache->end_seq) { 1868 /* ...but better entrypoint exists! */ 1869 skb = tcp_highest_sack(sk); 1870 if (skb == NULL) 1871 break; 1872 state.fack_count = tp->fackets_out; 1873 cache++; 1874 goto walk; 1875 } 1876 1877 skb = tcp_sacktag_skip(skb, sk, &state, cache->end_seq); 1878 /* Check overlap against next cached too (past this one already) */ 1879 cache++; 1880 continue; 1881 } 1882 1883 if (!before(start_seq, tcp_highest_sack_seq(tp))) { 1884 skb = tcp_highest_sack(sk); 1885 if (skb == NULL) 1886 break; 1887 state.fack_count = tp->fackets_out; 1888 } 1889 skb = tcp_sacktag_skip(skb, sk, &state, start_seq); 1890 1891walk: 1892 skb = tcp_sacktag_walk(skb, sk, next_dup, &state, 1893 start_seq, end_seq, dup_sack); 1894 1895advance_sp: 1896 /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct 1897 * due to in-order walk 1898 */ 1899 if (after(end_seq, tp->frto_highmark)) 1900 state.flag &= ~FLAG_ONLY_ORIG_SACKED; 1901 1902 i++; 1903 } 1904 1905 /* Clear the head of the cache sack blocks so we can skip it next time */ 1906 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) { 1907 tp->recv_sack_cache[i].start_seq = 0; 1908 tp->recv_sack_cache[i].end_seq = 0; 1909 } 1910 for (j = 0; j < used_sacks; j++) 1911 tp->recv_sack_cache[i++] = sp[j]; 1912 1913 tcp_mark_lost_retrans(sk); 1914 1915 tcp_verify_left_out(tp); 1916 1917 if ((state.reord < tp->fackets_out) && 1918 ((icsk->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker) && 1919 (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark))) 1920 tcp_update_reordering(sk, tp->fackets_out - state.reord, 0); 1921 1922out: 1923 1924#if FASTRETRANS_DEBUG > 0 1925 WARN_ON((int)tp->sacked_out < 0); 1926 WARN_ON((int)tp->lost_out < 0); 1927 WARN_ON((int)tp->retrans_out < 0); 1928 WARN_ON((int)tcp_packets_in_flight(tp) < 0); 1929#endif 1930 return state.flag; 1931} 1932 1933/* Limits sacked_out so that sum with lost_out isn't ever larger than 1934 * packets_out. Returns zero if sacked_out adjustement wasn't necessary. 1935 */ 1936static int tcp_limit_reno_sacked(struct tcp_sock *tp) 1937{ 1938 u32 holes; 1939 1940 holes = max(tp->lost_out, 1U); 1941 holes = min(holes, tp->packets_out); 1942 1943 if ((tp->sacked_out + holes) > tp->packets_out) { 1944 tp->sacked_out = tp->packets_out - holes; 1945 return 1; 1946 } 1947 return 0; 1948} 1949 1950/* If we receive more dupacks than we expected counting segments 1951 * in assumption of absent reordering, interpret this as reordering. 1952 * The only another reason could be bug in receiver TCP. 1953 */ 1954static void tcp_check_reno_reordering(struct sock *sk, const int addend) 1955{ 1956 struct tcp_sock *tp = tcp_sk(sk); 1957 if (tcp_limit_reno_sacked(tp)) 1958 tcp_update_reordering(sk, tp->packets_out + addend, 0); 1959} 1960 1961/* Emulate SACKs for SACKless connection: account for a new dupack. */ 1962 1963static void tcp_add_reno_sack(struct sock *sk) 1964{ 1965 struct tcp_sock *tp = tcp_sk(sk); 1966 tp->sacked_out++; 1967 tcp_check_reno_reordering(sk, 0); 1968 tcp_verify_left_out(tp); 1969} 1970 1971/* Account for ACK, ACKing some data in Reno Recovery phase. */ 1972 1973static void tcp_remove_reno_sacks(struct sock *sk, int acked) 1974{ 1975 struct tcp_sock *tp = tcp_sk(sk); 1976 1977 if (acked > 0) { 1978 /* One ACK acked hole. The rest eat duplicate ACKs. */ 1979 if (acked - 1 >= tp->sacked_out) 1980 tp->sacked_out = 0; 1981 else 1982 tp->sacked_out -= acked - 1; 1983 } 1984 tcp_check_reno_reordering(sk, acked); 1985 tcp_verify_left_out(tp); 1986} 1987 1988static inline void tcp_reset_reno_sack(struct tcp_sock *tp) 1989{ 1990 tp->sacked_out = 0; 1991} 1992 1993static int tcp_is_sackfrto(const struct tcp_sock *tp) 1994{ 1995 return (sysctl_tcp_frto == 0x2) && !tcp_is_reno(tp); 1996} 1997 1998/* F-RTO can only be used if TCP has never retransmitted anything other than 1999 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here) 2000 */ 2001int tcp_use_frto(struct sock *sk) 2002{ 2003 const struct tcp_sock *tp = tcp_sk(sk); 2004 const struct inet_connection_sock *icsk = inet_csk(sk); 2005 struct sk_buff *skb; 2006 2007 if (!sysctl_tcp_frto) 2008 return 0; 2009 2010 /* MTU probe and F-RTO won't really play nicely along currently */ 2011 if (icsk->icsk_mtup.probe_size) 2012 return 0; 2013 2014 if (tcp_is_sackfrto(tp)) 2015 return 1; 2016 2017 /* Avoid expensive walking of rexmit queue if possible */ 2018 if (tp->retrans_out > 1) 2019 return 0; 2020 2021 skb = tcp_write_queue_head(sk); 2022 if (tcp_skb_is_last(sk, skb)) 2023 return 1; 2024 skb = tcp_write_queue_next(sk, skb); /* Skips head */ 2025 tcp_for_write_queue_from(skb, sk) { 2026 if (skb == tcp_send_head(sk)) 2027 break; 2028 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) 2029 return 0; 2030 /* Short-circuit when first non-SACKed skb has been checked */ 2031 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) 2032 break; 2033 } 2034 return 1; 2035} 2036 2037/* RTO occurred, but do not yet enter Loss state. Instead, defer RTO 2038 * recovery a bit and use heuristics in tcp_process_frto() to detect if 2039 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to 2040 * keep retrans_out counting accurate (with SACK F-RTO, other than head 2041 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS 2042 * bits are handled if the Loss state is really to be entered (in 2043 * tcp_enter_frto_loss). 2044 * 2045 * Do like tcp_enter_loss() would; when RTO expires the second time it 2046 * does: 2047 * "Reduce ssthresh if it has not yet been made inside this window." 2048 */ 2049void tcp_enter_frto(struct sock *sk) 2050{ 2051 const struct inet_connection_sock *icsk = inet_csk(sk); 2052 struct tcp_sock *tp = tcp_sk(sk); 2053 struct sk_buff *skb; 2054 2055 if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) || 2056 tp->snd_una == tp->high_seq || 2057 ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) && 2058 !icsk->icsk_retransmits)) { 2059 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2060 /* Our state is too optimistic in ssthresh() call because cwnd 2061 * is not reduced until tcp_enter_frto_loss() when previous F-RTO 2062 * recovery has not yet completed. Pattern would be this: RTO, 2063 * Cumulative ACK, RTO (2xRTO for the same segment does not end 2064 * up here twice). 2065 * RFC4138 should be more specific on what to do, even though 2066 * RTO is quite unlikely to occur after the first Cumulative ACK 2067 * due to back-off and complexity of triggering events ... 2068 */ 2069 if (tp->frto_counter) { 2070 u32 stored_cwnd; 2071 stored_cwnd = tp->snd_cwnd; 2072 tp->snd_cwnd = 2; 2073 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 2074 tp->snd_cwnd = stored_cwnd; 2075 } else { 2076 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 2077 } 2078 /* ... in theory, cong.control module could do "any tricks" in 2079 * ssthresh(), which means that ca_state, lost bits and lost_out 2080 * counter would have to be faked before the call occurs. We 2081 * consider that too expensive, unlikely and hacky, so modules 2082 * using these in ssthresh() must deal these incompatibility 2083 * issues if they receives CA_EVENT_FRTO and frto_counter != 0 2084 */ 2085 tcp_ca_event(sk, CA_EVENT_FRTO); 2086 } 2087 2088 tp->undo_marker = tp->snd_una; 2089 tp->undo_retrans = 0; 2090 2091 skb = tcp_write_queue_head(sk); 2092 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) 2093 tp->undo_marker = 0; 2094 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 2095 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 2096 tp->retrans_out -= tcp_skb_pcount(skb); 2097 } 2098 tcp_verify_left_out(tp); 2099 2100 /* Too bad if TCP was application limited */ 2101 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1); 2102 2103 /* Earlier loss recovery underway (see RFC4138; Appendix B). 2104 * The last condition is necessary at least in tp->frto_counter case. 2105 */ 2106 if (tcp_is_sackfrto(tp) && (tp->frto_counter || 2107 ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) && 2108 after(tp->high_seq, tp->snd_una)) { 2109 tp->frto_highmark = tp->high_seq; 2110 } else { 2111 tp->frto_highmark = tp->snd_nxt; 2112 } 2113 tcp_set_ca_state(sk, TCP_CA_Disorder); 2114 tp->high_seq = tp->snd_nxt; 2115 tp->frto_counter = 1; 2116} 2117 2118/* Enter Loss state after F-RTO was applied. Dupack arrived after RTO, 2119 * which indicates that we should follow the traditional RTO recovery, 2120 * i.e. mark everything lost and do go-back-N retransmission. 2121 */ 2122static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag) 2123{ 2124 struct tcp_sock *tp = tcp_sk(sk); 2125 struct sk_buff *skb; 2126 2127 tp->lost_out = 0; 2128 tp->retrans_out = 0; 2129 if (tcp_is_reno(tp)) 2130 tcp_reset_reno_sack(tp); 2131 2132 tcp_for_write_queue(skb, sk) { 2133 if (skb == tcp_send_head(sk)) 2134 break; 2135 2136 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; 2137 /* 2138 * Count the retransmission made on RTO correctly (only when 2139 * waiting for the first ACK and did not get it)... 2140 */ 2141 if ((tp->frto_counter == 1) && !(flag & FLAG_DATA_ACKED)) { 2142 /* For some reason this R-bit might get cleared? */ 2143 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 2144 tp->retrans_out += tcp_skb_pcount(skb); 2145 /* ...enter this if branch just for the first segment */ 2146 flag |= FLAG_DATA_ACKED; 2147 } else { 2148 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) 2149 tp->undo_marker = 0; 2150 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 2151 } 2152 2153 /* Marking forward transmissions that were made after RTO lost 2154 * can cause unnecessary retransmissions in some scenarios, 2155 * SACK blocks will mitigate that in some but not in all cases. 2156 * We used to not mark them but it was causing break-ups with 2157 * receivers that do only in-order receival. 2158 * 2159 * TODO: we could detect presence of such receiver and select 2160 * different behavior per flow. 2161 */ 2162 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) { 2163 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 2164 tp->lost_out += tcp_skb_pcount(skb); 2165 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq; 2166 } 2167 } 2168 tcp_verify_left_out(tp); 2169 2170 tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments; 2171 tp->snd_cwnd_cnt = 0; 2172 tp->snd_cwnd_stamp = tcp_time_stamp; 2173 tp->frto_counter = 0; 2174 tp->bytes_acked = 0; 2175 2176 tp->reordering = min_t(unsigned int, tp->reordering, 2177 sysctl_tcp_reordering); 2178 tcp_set_ca_state(sk, TCP_CA_Loss); 2179 tp->high_seq = tp->snd_nxt; 2180 TCP_ECN_queue_cwr(tp); 2181 2182 tcp_clear_all_retrans_hints(tp); 2183} 2184 2185static void tcp_clear_retrans_partial(struct tcp_sock *tp) 2186{ 2187 tp->retrans_out = 0; 2188 tp->lost_out = 0; 2189 2190 tp->undo_marker = 0; 2191 tp->undo_retrans = 0; 2192} 2193 2194void tcp_clear_retrans(struct tcp_sock *tp) 2195{ 2196 tcp_clear_retrans_partial(tp); 2197 2198 tp->fackets_out = 0; 2199 tp->sacked_out = 0; 2200} 2201 2202/* Enter Loss state. If "how" is not zero, forget all SACK information 2203 * and reset tags completely, otherwise preserve SACKs. If receiver 2204 * dropped its ofo queue, we will know this due to reneging detection. 2205 */ 2206void tcp_enter_loss(struct sock *sk, int how) 2207{ 2208 const struct inet_connection_sock *icsk = inet_csk(sk); 2209 struct tcp_sock *tp = tcp_sk(sk); 2210 struct sk_buff *skb; 2211 2212 /* Reduce ssthresh if it has not yet been made inside this window. */ 2213 if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq || 2214 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) { 2215 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2216 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 2217 tcp_ca_event(sk, CA_EVENT_LOSS); 2218 } 2219 tp->snd_cwnd = 1; 2220 tp->snd_cwnd_cnt = 0; 2221 tp->snd_cwnd_stamp = tcp_time_stamp; 2222 2223 tp->bytes_acked = 0; 2224 tcp_clear_retrans_partial(tp); 2225 2226 if (tcp_is_reno(tp)) 2227 tcp_reset_reno_sack(tp); 2228 2229 if (!how) { 2230 /* Push undo marker, if it was plain RTO and nothing 2231 * was retransmitted. */ 2232 tp->undo_marker = tp->snd_una; 2233 } else { 2234 tp->sacked_out = 0; 2235 tp->fackets_out = 0; 2236 } 2237 tcp_clear_all_retrans_hints(tp); 2238 2239 tcp_for_write_queue(skb, sk) { 2240 if (skb == tcp_send_head(sk)) 2241 break; 2242 2243 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) 2244 tp->undo_marker = 0; 2245 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED; 2246 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) { 2247 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED; 2248 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 2249 tp->lost_out += tcp_skb_pcount(skb); 2250 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq; 2251 } 2252 } 2253 tcp_verify_left_out(tp); 2254 2255 tp->reordering = min_t(unsigned int, tp->reordering, 2256 sysctl_tcp_reordering); 2257 tcp_set_ca_state(sk, TCP_CA_Loss); 2258 tp->high_seq = tp->snd_nxt; 2259 TCP_ECN_queue_cwr(tp); 2260 /* Abort F-RTO algorithm if one is in progress */ 2261 tp->frto_counter = 0; 2262} 2263 2264/* If ACK arrived pointing to a remembered SACK, it means that our 2265 * remembered SACKs do not reflect real state of receiver i.e. 2266 * receiver _host_ is heavily congested (or buggy). 2267 * 2268 * Do processing similar to RTO timeout. 2269 */ 2270static int tcp_check_sack_reneging(struct sock *sk, int flag) 2271{ 2272 if (flag & FLAG_SACK_RENEGING) { 2273 struct inet_connection_sock *icsk = inet_csk(sk); 2274 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING); 2275 2276 tcp_enter_loss(sk, 1); 2277 icsk->icsk_retransmits++; 2278 tcp_retransmit_skb(sk, tcp_write_queue_head(sk)); 2279 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 2280 icsk->icsk_rto, TCP_RTO_MAX); 2281 return 1; 2282 } 2283 return 0; 2284} 2285 2286static inline int tcp_fackets_out(struct tcp_sock *tp) 2287{ 2288 return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out; 2289} 2290 2291/* Heurestics to calculate number of duplicate ACKs. There's no dupACKs 2292 * counter when SACK is enabled (without SACK, sacked_out is used for 2293 * that purpose). 2294 * 2295 * Instead, with FACK TCP uses fackets_out that includes both SACKed 2296 * segments up to the highest received SACK block so far and holes in 2297 * between them. 2298 * 2299 * With reordering, holes may still be in flight, so RFC3517 recovery 2300 * uses pure sacked_out (total number of SACKed segments) even though 2301 * it violates the RFC that uses duplicate ACKs, often these are equal 2302 * but when e.g. out-of-window ACKs or packet duplication occurs, 2303 * they differ. Since neither occurs due to loss, TCP should really 2304 * ignore them. 2305 */ 2306static inline int tcp_dupack_heuristics(struct tcp_sock *tp) 2307{ 2308 return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1; 2309} 2310 2311static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb) 2312{ 2313 return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto); 2314} 2315 2316static inline int tcp_head_timedout(struct sock *sk) 2317{ 2318 struct tcp_sock *tp = tcp_sk(sk); 2319 2320 return tp->packets_out && 2321 tcp_skb_timedout(sk, tcp_write_queue_head(sk)); 2322} 2323 2324/* Linux NewReno/SACK/FACK/ECN state machine. 2325 * -------------------------------------- 2326 * 2327 * "Open" Normal state, no dubious events, fast path. 2328 * "Disorder" In all the respects it is "Open", 2329 * but requires a bit more attention. It is entered when 2330 * we see some SACKs or dupacks. It is split of "Open" 2331 * mainly to move some processing from fast path to slow one. 2332 * "CWR" CWND was reduced due to some Congestion Notification event. 2333 * It can be ECN, ICMP source quench, local device congestion. 2334 * "Recovery" CWND was reduced, we are fast-retransmitting. 2335 * "Loss" CWND was reduced due to RTO timeout or SACK reneging. 2336 * 2337 * tcp_fastretrans_alert() is entered: 2338 * - each incoming ACK, if state is not "Open" 2339 * - when arrived ACK is unusual, namely: 2340 * * SACK 2341 * * Duplicate ACK. 2342 * * ECN ECE. 2343 * 2344 * Counting packets in flight is pretty simple. 2345 * 2346 * in_flight = packets_out - left_out + retrans_out 2347 * 2348 * packets_out is SND.NXT-SND.UNA counted in packets. 2349 * 2350 * retrans_out is number of retransmitted segments. 2351 * 2352 * left_out is number of segments left network, but not ACKed yet. 2353 * 2354 * left_out = sacked_out + lost_out 2355 * 2356 * sacked_out: Packets, which arrived to receiver out of order 2357 * and hence not ACKed. With SACKs this number is simply 2358 * amount of SACKed data. Even without SACKs 2359 * it is easy to give pretty reliable estimate of this number, 2360 * counting duplicate ACKs. 2361 * 2362 * lost_out: Packets lost by network. TCP has no explicit 2363 * "loss notification" feedback from network (for now). 2364 * It means that this number can be only _guessed_. 2365 * Actually, it is the heuristics to predict lossage that 2366 * distinguishes different algorithms. 2367 * 2368 * F.e. after RTO, when all the queue is considered as lost, 2369 * lost_out = packets_out and in_flight = retrans_out. 2370 * 2371 * Essentially, we have now two algorithms counting 2372 * lost packets. 2373 * 2374 * FACK: It is the simplest heuristics. As soon as we decided 2375 * that something is lost, we decide that _all_ not SACKed 2376 * packets until the most forward SACK are lost. I.e. 2377 * lost_out = fackets_out - sacked_out and left_out = fackets_out. 2378 * It is absolutely correct estimate, if network does not reorder 2379 * packets. And it loses any connection to reality when reordering 2380 * takes place. We use FACK by default until reordering 2381 * is suspected on the path to this destination. 2382 * 2383 * NewReno: when Recovery is entered, we assume that one segment 2384 * is lost (classic Reno). While we are in Recovery and 2385 * a partial ACK arrives, we assume that one more packet 2386 * is lost (NewReno). This heuristics are the same in NewReno 2387 * and SACK. 2388 * 2389 * Imagine, that's all! Forget about all this shamanism about CWND inflation 2390 * deflation etc. CWND is real congestion window, never inflated, changes 2391 * only according to classic VJ rules. 2392 * 2393 * Really tricky (and requiring careful tuning) part of algorithm 2394 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue(). 2395 * The first determines the moment _when_ we should reduce CWND and, 2396 * hence, slow down forward transmission. In fact, it determines the moment 2397 * when we decide that hole is caused by loss, rather than by a reorder. 2398 * 2399 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill 2400 * holes, caused by lost packets. 2401 * 2402 * And the most logically complicated part of algorithm is undo 2403 * heuristics. We detect false retransmits due to both too early 2404 * fast retransmit (reordering) and underestimated RTO, analyzing 2405 * timestamps and D-SACKs. When we detect that some segments were 2406 * retransmitted by mistake and CWND reduction was wrong, we undo 2407 * window reduction and abort recovery phase. This logic is hidden 2408 * inside several functions named tcp_try_undo_<something>. 2409 */ 2410 2411/* This function decides, when we should leave Disordered state 2412 * and enter Recovery phase, reducing congestion window. 2413 * 2414 * Main question: may we further continue forward transmission 2415 * with the same cwnd? 2416 */ 2417static int tcp_time_to_recover(struct sock *sk) 2418{ 2419 struct tcp_sock *tp = tcp_sk(sk); 2420 __u32 packets_out; 2421 2422 /* Do not perform any recovery during F-RTO algorithm */ 2423 if (tp->frto_counter) 2424 return 0; 2425 2426 /* Trick#1: The loss is proven. */ 2427 if (tp->lost_out) 2428 return 1; 2429 2430 /* Not-A-Trick#2 : Classic rule... */ 2431 if (tcp_dupack_heuristics(tp) > tp->reordering) 2432 return 1; 2433 2434 /* Trick#3 : when we use RFC2988 timer restart, fast 2435 * retransmit can be triggered by timeout of queue head. 2436 */ 2437 if (tcp_is_fack(tp) && tcp_head_timedout(sk)) 2438 return 1; 2439 2440 /* Trick#4: It is still not OK... But will it be useful to delay 2441 * recovery more? 2442 */ 2443 packets_out = tp->packets_out; 2444 if (packets_out <= tp->reordering && 2445 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) && 2446 !tcp_may_send_now(sk)) { 2447 /* We have nothing to send. This connection is limited 2448 * either by receiver window or by application. 2449 */ 2450 return 1; 2451 } 2452 2453 /* If a thin stream is detected, retransmit after first 2454 * received dupack. Employ only if SACK is supported in order 2455 * to avoid possible corner-case series of spurious retransmissions 2456 * Use only if there are no unsent data. 2457 */ 2458 if ((tp->thin_dupack || sysctl_tcp_thin_dupack) && 2459 tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 && 2460 tcp_is_sack(tp) && !tcp_send_head(sk)) 2461 return 1; 2462 2463 return 0; 2464} 2465 2466/* New heuristics: it is possible only after we switched to restart timer 2467 * each time when something is ACKed. Hence, we can detect timed out packets 2468 * during fast retransmit without falling to slow start. 2469 * 2470 * Usefulness of this as is very questionable, since we should know which of 2471 * the segments is the next to timeout which is relatively expensive to find 2472 * in general case unless we add some data structure just for that. The 2473 * current approach certainly won't find the right one too often and when it 2474 * finally does find _something_ it usually marks large part of the window 2475 * right away (because a retransmission with a larger timestamp blocks the 2476 * loop from advancing). -ij 2477 */ 2478static void tcp_timeout_skbs(struct sock *sk) 2479{ 2480 struct tcp_sock *tp = tcp_sk(sk); 2481 struct sk_buff *skb; 2482 2483 if (!tcp_is_fack(tp) || !tcp_head_timedout(sk)) 2484 return; 2485 2486 skb = tp->scoreboard_skb_hint; 2487 if (tp->scoreboard_skb_hint == NULL) 2488 skb = tcp_write_queue_head(sk); 2489 2490 tcp_for_write_queue_from(skb, sk) { 2491 if (skb == tcp_send_head(sk)) 2492 break; 2493 if (!tcp_skb_timedout(sk, skb)) 2494 break; 2495 2496 tcp_skb_mark_lost(tp, skb); 2497 } 2498 2499 tp->scoreboard_skb_hint = skb; 2500 2501 tcp_verify_left_out(tp); 2502} 2503 2504/* Mark head of queue up as lost. With RFC3517 SACK, the packets is 2505 * is against sacked "cnt", otherwise it's against facked "cnt" 2506 */ 2507static void tcp_mark_head_lost(struct sock *sk, int packets) 2508{ 2509 struct tcp_sock *tp = tcp_sk(sk); 2510 struct sk_buff *skb; 2511 int cnt, oldcnt; 2512 int err; 2513 unsigned int mss; 2514 2515 if (packets == 0) 2516 return; 2517 2518 WARN_ON(packets > tp->packets_out); 2519 if (tp->lost_skb_hint) { 2520 skb = tp->lost_skb_hint; 2521 cnt = tp->lost_cnt_hint; 2522 } else { 2523 skb = tcp_write_queue_head(sk); 2524 cnt = 0; 2525 } 2526 2527 tcp_for_write_queue_from(skb, sk) { 2528 if (skb == tcp_send_head(sk)) 2529 break; 2530 /* TODO: do this better */ 2531 /* this is not the most efficient way to do this... */ 2532 tp->lost_skb_hint = skb; 2533 tp->lost_cnt_hint = cnt; 2534 2535 if (after(TCP_SKB_CB(skb)->end_seq, tp->high_seq)) 2536 break; 2537 2538 oldcnt = cnt; 2539 if (tcp_is_fack(tp) || tcp_is_reno(tp) || 2540 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) 2541 cnt += tcp_skb_pcount(skb); 2542 2543 if (cnt > packets) { 2544 if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) || 2545 (oldcnt >= packets)) 2546 break; 2547 2548 mss = skb_shinfo(skb)->gso_size; 2549 err = tcp_fragment(sk, skb, (packets - oldcnt) * mss, mss); 2550 if (err < 0) 2551 break; 2552 cnt = packets; 2553 } 2554 2555 tcp_skb_mark_lost(tp, skb); 2556 } 2557 tcp_verify_left_out(tp); 2558} 2559 2560/* Account newly detected lost packet(s) */ 2561 2562static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit) 2563{ 2564 struct tcp_sock *tp = tcp_sk(sk); 2565 2566 if (tcp_is_reno(tp)) { 2567 tcp_mark_head_lost(sk, 1); 2568 } else if (tcp_is_fack(tp)) { 2569 int lost = tp->fackets_out - tp->reordering; 2570 if (lost <= 0) 2571 lost = 1; 2572 tcp_mark_head_lost(sk, lost); 2573 } else { 2574 int sacked_upto = tp->sacked_out - tp->reordering; 2575 if (sacked_upto < fast_rexmit) 2576 sacked_upto = fast_rexmit; 2577 tcp_mark_head_lost(sk, sacked_upto); 2578 } 2579 2580 tcp_timeout_skbs(sk); 2581} 2582 2583/* CWND moderation, preventing bursts due to too big ACKs 2584 * in dubious situations. 2585 */ 2586static inline void tcp_moderate_cwnd(struct tcp_sock *tp) 2587{ 2588 tp->snd_cwnd = min(tp->snd_cwnd, 2589 tcp_packets_in_flight(tp) + tcp_max_burst(tp)); 2590 tp->snd_cwnd_stamp = tcp_time_stamp; 2591} 2592 2593/* Lower bound on congestion window is slow start threshold 2594 * unless congestion avoidance choice decides to overide it. 2595 */ 2596static inline u32 tcp_cwnd_min(const struct sock *sk) 2597{ 2598 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 2599 2600 return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh; 2601} 2602 2603/* Decrease cwnd each second ack. */ 2604static void tcp_cwnd_down(struct sock *sk, int flag) 2605{ 2606 struct tcp_sock *tp = tcp_sk(sk); 2607 int decr = tp->snd_cwnd_cnt + 1; 2608 2609 if ((flag & (FLAG_ANY_PROGRESS | FLAG_DSACKING_ACK)) || 2610 (tcp_is_reno(tp) && !(flag & FLAG_NOT_DUP))) { 2611 tp->snd_cwnd_cnt = decr & 1; 2612 decr >>= 1; 2613 2614 if (decr && tp->snd_cwnd > tcp_cwnd_min(sk)) 2615 tp->snd_cwnd -= decr; 2616 2617 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1); 2618 tp->snd_cwnd_stamp = tcp_time_stamp; 2619 } 2620} 2621 2622/* Nothing was retransmitted or returned timestamp is less 2623 * than timestamp of the first retransmission. 2624 */ 2625static inline int tcp_packet_delayed(struct tcp_sock *tp) 2626{ 2627 return !tp->retrans_stamp || 2628 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 2629 before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp)); 2630} 2631 2632/* Undo procedures. */ 2633 2634#if FASTRETRANS_DEBUG > 1 2635static void DBGUNDO(struct sock *sk, const char *msg) 2636{ 2637 struct tcp_sock *tp = tcp_sk(sk); 2638 struct inet_sock *inet = inet_sk(sk); 2639 2640 if (sk->sk_family == AF_INET) { 2641 printk(KERN_DEBUG "Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n", 2642 msg, 2643 &inet->inet_daddr, ntohs(inet->inet_dport), 2644 tp->snd_cwnd, tcp_left_out(tp), 2645 tp->snd_ssthresh, tp->prior_ssthresh, 2646 tp->packets_out); 2647 } 2648#if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE) 2649 else if (sk->sk_family == AF_INET6) { 2650 struct ipv6_pinfo *np = inet6_sk(sk); 2651 printk(KERN_DEBUG "Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n", 2652 msg, 2653 &np->daddr, ntohs(inet->inet_dport), 2654 tp->snd_cwnd, tcp_left_out(tp), 2655 tp->snd_ssthresh, tp->prior_ssthresh, 2656 tp->packets_out); 2657 } 2658#endif 2659} 2660#else 2661#define DBGUNDO(x...) do { } while (0) 2662#endif 2663 2664static void tcp_undo_cwr(struct sock *sk, const int undo) 2665{ 2666 struct tcp_sock *tp = tcp_sk(sk); 2667 2668 if (tp->prior_ssthresh) { 2669 const struct inet_connection_sock *icsk = inet_csk(sk); 2670 2671 if (icsk->icsk_ca_ops->undo_cwnd) 2672 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk); 2673 else 2674 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1); 2675 2676 if (undo && tp->prior_ssthresh > tp->snd_ssthresh) { 2677 tp->snd_ssthresh = tp->prior_ssthresh; 2678 TCP_ECN_withdraw_cwr(tp); 2679 } 2680 } else { 2681 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh); 2682 } 2683 tcp_moderate_cwnd(tp); 2684 tp->snd_cwnd_stamp = tcp_time_stamp; 2685} 2686 2687static inline int tcp_may_undo(struct tcp_sock *tp) 2688{ 2689 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp)); 2690} 2691 2692/* People celebrate: "We love our President!" */ 2693static int tcp_try_undo_recovery(struct sock *sk) 2694{ 2695 struct tcp_sock *tp = tcp_sk(sk); 2696 2697 if (tcp_may_undo(tp)) { 2698 int mib_idx; 2699 2700 /* Happy end! We did not retransmit anything 2701 * or our original transmission succeeded. 2702 */ 2703 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans"); 2704 tcp_undo_cwr(sk, 1); 2705 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss) 2706 mib_idx = LINUX_MIB_TCPLOSSUNDO; 2707 else 2708 mib_idx = LINUX_MIB_TCPFULLUNDO; 2709 2710 NET_INC_STATS_BH(sock_net(sk), mib_idx); 2711 tp->undo_marker = 0; 2712 } 2713 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) { 2714 /* Hold old state until something *above* high_seq 2715 * is ACKed. For Reno it is MUST to prevent false 2716 * fast retransmits (RFC2582). SACK TCP is safe. */ 2717 tcp_moderate_cwnd(tp); 2718 return 1; 2719 } 2720 tcp_set_ca_state(sk, TCP_CA_Open); 2721 return 0; 2722} 2723 2724/* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */ 2725static void tcp_try_undo_dsack(struct sock *sk) 2726{ 2727 struct tcp_sock *tp = tcp_sk(sk); 2728 2729 if (tp->undo_marker && !tp->undo_retrans) { 2730 DBGUNDO(sk, "D-SACK"); 2731 tcp_undo_cwr(sk, 1); 2732 tp->undo_marker = 0; 2733 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO); 2734 } 2735} 2736 2737/* We can clear retrans_stamp when there are no retransmissions in the 2738 * window. It would seem that it is trivially available for us in 2739 * tp->retrans_out, however, that kind of assumptions doesn't consider 2740 * what will happen if errors occur when sending retransmission for the 2741 * second time. ...It could the that such segment has only 2742 * TCPCB_EVER_RETRANS set at the present time. It seems that checking 2743 * the head skb is enough except for some reneging corner cases that 2744 * are not worth the effort. 2745 * 2746 * Main reason for all this complexity is the fact that connection dying 2747 * time now depends on the validity of the retrans_stamp, in particular, 2748 * that successive retransmissions of a segment must not advance 2749 * retrans_stamp under any conditions. 2750 */ 2751static int tcp_any_retrans_done(struct sock *sk) 2752{ 2753 struct tcp_sock *tp = tcp_sk(sk); 2754 struct sk_buff *skb; 2755 2756 if (tp->retrans_out) 2757 return 1; 2758 2759 skb = tcp_write_queue_head(sk); 2760 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS)) 2761 return 1; 2762 2763 return 0; 2764} 2765 2766/* Undo during fast recovery after partial ACK. */ 2767 2768static int tcp_try_undo_partial(struct sock *sk, int acked) 2769{ 2770 struct tcp_sock *tp = tcp_sk(sk); 2771 /* Partial ACK arrived. Force Hoe's retransmit. */ 2772 int failed = tcp_is_reno(tp) || (tcp_fackets_out(tp) > tp->reordering); 2773 2774 if (tcp_may_undo(tp)) { 2775 /* Plain luck! Hole if filled with delayed 2776 * packet, rather than with a retransmit. 2777 */ 2778 if (!tcp_any_retrans_done(sk)) 2779 tp->retrans_stamp = 0; 2780 2781 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1); 2782 2783 DBGUNDO(sk, "Hoe"); 2784 tcp_undo_cwr(sk, 0); 2785 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO); 2786 2787 /* So... Do not make Hoe's retransmit yet. 2788 * If the first packet was delayed, the rest 2789 * ones are most probably delayed as well. 2790 */ 2791 failed = 0; 2792 } 2793 return failed; 2794} 2795 2796/* Undo during loss recovery after partial ACK. */ 2797static int tcp_try_undo_loss(struct sock *sk) 2798{ 2799 struct tcp_sock *tp = tcp_sk(sk); 2800 2801 if (tcp_may_undo(tp)) { 2802 struct sk_buff *skb; 2803 tcp_for_write_queue(skb, sk) { 2804 if (skb == tcp_send_head(sk)) 2805 break; 2806 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; 2807 } 2808 2809 tcp_clear_all_retrans_hints(tp); 2810 2811 DBGUNDO(sk, "partial loss"); 2812 tp->lost_out = 0; 2813 tcp_undo_cwr(sk, 1); 2814 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO); 2815 inet_csk(sk)->icsk_retransmits = 0; 2816 tp->undo_marker = 0; 2817 if (tcp_is_sack(tp)) 2818 tcp_set_ca_state(sk, TCP_CA_Open); 2819 return 1; 2820 } 2821 return 0; 2822} 2823 2824static inline void tcp_complete_cwr(struct sock *sk) 2825{ 2826 struct tcp_sock *tp = tcp_sk(sk); 2827 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh); 2828 tp->snd_cwnd_stamp = tcp_time_stamp; 2829 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR); 2830} 2831 2832static void tcp_try_keep_open(struct sock *sk) 2833{ 2834 struct tcp_sock *tp = tcp_sk(sk); 2835 int state = TCP_CA_Open; 2836 2837 if (tcp_left_out(tp) || tcp_any_retrans_done(sk) || tp->undo_marker) 2838 state = TCP_CA_Disorder; 2839 2840 if (inet_csk(sk)->icsk_ca_state != state) { 2841 tcp_set_ca_state(sk, state); 2842 tp->high_seq = tp->snd_nxt; 2843 } 2844} 2845 2846static void tcp_try_to_open(struct sock *sk, int flag) 2847{ 2848 struct tcp_sock *tp = tcp_sk(sk); 2849 2850 tcp_verify_left_out(tp); 2851 2852 if (!tp->frto_counter && !tcp_any_retrans_done(sk)) 2853 tp->retrans_stamp = 0; 2854 2855 if (flag & FLAG_ECE) 2856 tcp_enter_cwr(sk, 1); 2857 2858 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) { 2859 tcp_try_keep_open(sk); 2860 tcp_moderate_cwnd(tp); 2861 } else { 2862 tcp_cwnd_down(sk, flag); 2863 } 2864} 2865 2866static void tcp_mtup_probe_failed(struct sock *sk) 2867{ 2868 struct inet_connection_sock *icsk = inet_csk(sk); 2869 2870 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1; 2871 icsk->icsk_mtup.probe_size = 0; 2872} 2873 2874static void tcp_mtup_probe_success(struct sock *sk) 2875{ 2876 struct tcp_sock *tp = tcp_sk(sk); 2877 struct inet_connection_sock *icsk = inet_csk(sk); 2878 2879 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2880 tp->snd_cwnd = tp->snd_cwnd * 2881 tcp_mss_to_mtu(sk, tp->mss_cache) / 2882 icsk->icsk_mtup.probe_size; 2883 tp->snd_cwnd_cnt = 0; 2884 tp->snd_cwnd_stamp = tcp_time_stamp; 2885 tp->rcv_ssthresh = tcp_current_ssthresh(sk); 2886 2887 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size; 2888 icsk->icsk_mtup.probe_size = 0; 2889 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 2890} 2891 2892/* Do a simple retransmit without using the backoff mechanisms in 2893 * tcp_timer. This is used for path mtu discovery. 2894 * The socket is already locked here. 2895 */ 2896void tcp_simple_retransmit(struct sock *sk) 2897{ 2898 const struct inet_connection_sock *icsk = inet_csk(sk); 2899 struct tcp_sock *tp = tcp_sk(sk); 2900 struct sk_buff *skb; 2901 unsigned int mss = tcp_current_mss(sk); 2902 u32 prior_lost = tp->lost_out; 2903 2904 tcp_for_write_queue(skb, sk) { 2905 if (skb == tcp_send_head(sk)) 2906 break; 2907 if (tcp_skb_seglen(skb) > mss && 2908 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) { 2909 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 2910 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 2911 tp->retrans_out -= tcp_skb_pcount(skb); 2912 } 2913 tcp_skb_mark_lost_uncond_verify(tp, skb); 2914 } 2915 } 2916 2917 tcp_clear_retrans_hints_partial(tp); 2918 2919 if (prior_lost == tp->lost_out) 2920 return; 2921 2922 if (tcp_is_reno(tp)) 2923 tcp_limit_reno_sacked(tp); 2924 2925 tcp_verify_left_out(tp); 2926 2927 /* Don't muck with the congestion window here. 2928 * Reason is that we do not increase amount of _data_ 2929 * in network, but units changed and effective 2930 * cwnd/ssthresh really reduced now. 2931 */ 2932 if (icsk->icsk_ca_state != TCP_CA_Loss) { 2933 tp->high_seq = tp->snd_nxt; 2934 tp->snd_ssthresh = tcp_current_ssthresh(sk); 2935 tp->prior_ssthresh = 0; 2936 tp->undo_marker = 0; 2937 tcp_set_ca_state(sk, TCP_CA_Loss); 2938 } 2939 tcp_xmit_retransmit_queue(sk); 2940} 2941EXPORT_SYMBOL(tcp_simple_retransmit); 2942 2943/* Process an event, which can update packets-in-flight not trivially. 2944 * Main goal of this function is to calculate new estimate for left_out, 2945 * taking into account both packets sitting in receiver's buffer and 2946 * packets lost by network. 2947 * 2948 * Besides that it does CWND reduction, when packet loss is detected 2949 * and changes state of machine. 2950 * 2951 * It does _not_ decide what to send, it is made in function 2952 * tcp_xmit_retransmit_queue(). 2953 */ 2954static void tcp_fastretrans_alert(struct sock *sk, int pkts_acked, int flag) 2955{ 2956 struct inet_connection_sock *icsk = inet_csk(sk); 2957 struct tcp_sock *tp = tcp_sk(sk); 2958 int is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP)); 2959 int do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) && 2960 (tcp_fackets_out(tp) > tp->reordering)); 2961 int fast_rexmit = 0, mib_idx; 2962 2963 if (WARN_ON(!tp->packets_out && tp->sacked_out)) 2964 tp->sacked_out = 0; 2965 if (WARN_ON(!tp->sacked_out && tp->fackets_out)) 2966 tp->fackets_out = 0; 2967 2968 /* Now state machine starts. 2969 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */ 2970 if (flag & FLAG_ECE) 2971 tp->prior_ssthresh = 0; 2972 2973 /* B. In all the states check for reneging SACKs. */ 2974 if (tcp_check_sack_reneging(sk, flag)) 2975 return; 2976 2977 /* C. Process data loss notification, provided it is valid. */ 2978 if (tcp_is_fack(tp) && (flag & FLAG_DATA_LOST) && 2979 before(tp->snd_una, tp->high_seq) && 2980 icsk->icsk_ca_state != TCP_CA_Open && 2981 tp->fackets_out > tp->reordering) { 2982 tcp_mark_head_lost(sk, tp->fackets_out - tp->reordering); 2983 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSS); 2984 } 2985 2986 /* D. Check consistency of the current state. */ 2987 tcp_verify_left_out(tp); 2988 2989 /* E. Check state exit conditions. State can be terminated 2990 * when high_seq is ACKed. */ 2991 if (icsk->icsk_ca_state == TCP_CA_Open) { 2992 WARN_ON(tp->retrans_out != 0); 2993 tp->retrans_stamp = 0; 2994 } else if (!before(tp->snd_una, tp->high_seq)) { 2995 switch (icsk->icsk_ca_state) { 2996 case TCP_CA_Loss: 2997 icsk->icsk_retransmits = 0; 2998 if (tcp_try_undo_recovery(sk)) 2999 return; 3000 break; 3001 3002 case TCP_CA_CWR: 3003 /* CWR is to be held something *above* high_seq 3004 * is ACKed for CWR bit to reach receiver. */ 3005 if (tp->snd_una != tp->high_seq) { 3006 tcp_complete_cwr(sk); 3007 tcp_set_ca_state(sk, TCP_CA_Open); 3008 } 3009 break; 3010 3011 case TCP_CA_Disorder: 3012 tcp_try_undo_dsack(sk); 3013 if (!tp->undo_marker || 3014 /* For SACK case do not Open to allow to undo 3015 * catching for all duplicate ACKs. */ 3016 tcp_is_reno(tp) || tp->snd_una != tp->high_seq) { 3017 tp->undo_marker = 0; 3018 tcp_set_ca_state(sk, TCP_CA_Open); 3019 } 3020 break; 3021 3022 case TCP_CA_Recovery: 3023 if (tcp_is_reno(tp)) 3024 tcp_reset_reno_sack(tp); 3025 if (tcp_try_undo_recovery(sk)) 3026 return; 3027 tcp_complete_cwr(sk); 3028 break; 3029 } 3030 } 3031 3032 /* F. Process state. */ 3033 switch (icsk->icsk_ca_state) { 3034 case TCP_CA_Recovery: 3035 if (!(flag & FLAG_SND_UNA_ADVANCED)) { 3036 if (tcp_is_reno(tp) && is_dupack) 3037 tcp_add_reno_sack(sk); 3038 } else 3039 do_lost = tcp_try_undo_partial(sk, pkts_acked); 3040 break; 3041 case TCP_CA_Loss: 3042 if (flag & FLAG_DATA_ACKED) 3043 icsk->icsk_retransmits = 0; 3044 if (tcp_is_reno(tp) && flag & FLAG_SND_UNA_ADVANCED) 3045 tcp_reset_reno_sack(tp); 3046 if (!tcp_try_undo_loss(sk)) { 3047 tcp_moderate_cwnd(tp); 3048 tcp_xmit_retransmit_queue(sk); 3049 return; 3050 } 3051 if (icsk->icsk_ca_state != TCP_CA_Open) 3052 return; 3053 /* Loss is undone; fall through to processing in Open state. */ 3054 default: 3055 if (tcp_is_reno(tp)) { 3056 if (flag & FLAG_SND_UNA_ADVANCED) 3057 tcp_reset_reno_sack(tp); 3058 if (is_dupack) 3059 tcp_add_reno_sack(sk); 3060 } 3061 3062 if (icsk->icsk_ca_state == TCP_CA_Disorder) 3063 tcp_try_undo_dsack(sk); 3064 3065 if (!tcp_time_to_recover(sk)) { 3066 tcp_try_to_open(sk, flag); 3067 return; 3068 } 3069 3070 /* MTU probe failure: don't reduce cwnd */ 3071 if (icsk->icsk_ca_state < TCP_CA_CWR && 3072 icsk->icsk_mtup.probe_size && 3073 tp->snd_una == tp->mtu_probe.probe_seq_start) { 3074 tcp_mtup_probe_failed(sk); 3075 /* Restores the reduction we did in tcp_mtup_probe() */ 3076 tp->snd_cwnd++; 3077 tcp_simple_retransmit(sk); 3078 return; 3079 } 3080 3081 /* Otherwise enter Recovery state */ 3082 3083 if (tcp_is_reno(tp)) 3084 mib_idx = LINUX_MIB_TCPRENORECOVERY; 3085 else 3086 mib_idx = LINUX_MIB_TCPSACKRECOVERY; 3087 3088 NET_INC_STATS_BH(sock_net(sk), mib_idx); 3089 3090 tp->high_seq = tp->snd_nxt; 3091 tp->prior_ssthresh = 0; 3092 tp->undo_marker = tp->snd_una; 3093 tp->undo_retrans = tp->retrans_out; 3094 3095 if (icsk->icsk_ca_state < TCP_CA_CWR) { 3096 if (!(flag & FLAG_ECE)) 3097 tp->prior_ssthresh = tcp_current_ssthresh(sk); 3098 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 3099 TCP_ECN_queue_cwr(tp); 3100 } 3101 3102 tp->bytes_acked = 0; 3103 tp->snd_cwnd_cnt = 0; 3104 tcp_set_ca_state(sk, TCP_CA_Recovery); 3105 fast_rexmit = 1; 3106 } 3107 3108 if (do_lost || (tcp_is_fack(tp) && tcp_head_timedout(sk))) 3109 tcp_update_scoreboard(sk, fast_rexmit); 3110 tcp_cwnd_down(sk, flag); 3111 tcp_xmit_retransmit_queue(sk); 3112} 3113 3114static void tcp_valid_rtt_meas(struct sock *sk, u32 seq_rtt) 3115{ 3116 tcp_rtt_estimator(sk, seq_rtt); 3117 tcp_set_rto(sk); 3118 inet_csk(sk)->icsk_backoff = 0; 3119} 3120 3121/* Read draft-ietf-tcplw-high-performance before mucking 3122 * with this code. (Supersedes RFC1323) 3123 */ 3124static void tcp_ack_saw_tstamp(struct sock *sk, int flag) 3125{ 3126 /* RTTM Rule: A TSecr value received in a segment is used to 3127 * update the averaged RTT measurement only if the segment 3128 * acknowledges some new data, i.e., only if it advances the 3129 * left edge of the send window. 3130 * 3131 * See draft-ietf-tcplw-high-performance-00, section 3.3. 3132 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru> 3133 * 3134 * Changed: reset backoff as soon as we see the first valid sample. 3135 * If we do not, we get strongly overestimated rto. With timestamps 3136 * samples are accepted even from very old segments: f.e., when rtt=1 3137 * increases to 8, we retransmit 5 times and after 8 seconds delayed 3138 * answer arrives rto becomes 120 seconds! If at least one of segments 3139 * in window is lost... Voila. --ANK (010210) 3140 */ 3141 struct tcp_sock *tp = tcp_sk(sk); 3142 3143 tcp_valid_rtt_meas(sk, tcp_time_stamp - tp->rx_opt.rcv_tsecr); 3144} 3145 3146static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag) 3147{ 3148 /* We don't have a timestamp. Can only use 3149 * packets that are not retransmitted to determine 3150 * rtt estimates. Also, we must not reset the 3151 * backoff for rto until we get a non-retransmitted 3152 * packet. This allows us to deal with a situation 3153 * where the network delay has increased suddenly. 3154 * I.e. Karn's algorithm. (SIGCOMM '87, p5.) 3155 */ 3156 3157 if (flag & FLAG_RETRANS_DATA_ACKED) 3158 return; 3159 3160 tcp_valid_rtt_meas(sk, seq_rtt); 3161} 3162 3163static inline void tcp_ack_update_rtt(struct sock *sk, const int flag, 3164 const s32 seq_rtt) 3165{ 3166 const struct tcp_sock *tp = tcp_sk(sk); 3167 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */ 3168 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr) 3169 tcp_ack_saw_tstamp(sk, flag); 3170 else if (seq_rtt >= 0) 3171 tcp_ack_no_tstamp(sk, seq_rtt, flag); 3172} 3173 3174static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight) 3175{ 3176 const struct inet_connection_sock *icsk = inet_csk(sk); 3177 icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight); 3178 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp; 3179} 3180 3181/* Restart timer after forward progress on connection. 3182 * RFC2988 recommends to restart timer to now+rto. 3183 */ 3184static void tcp_rearm_rto(struct sock *sk) 3185{ 3186 struct tcp_sock *tp = tcp_sk(sk); 3187 3188 if (!tp->packets_out) { 3189 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS); 3190 } else { 3191 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 3192 inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 3193 } 3194} 3195 3196/* If we get here, the whole TSO packet has not been acked. */ 3197static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb) 3198{ 3199 struct tcp_sock *tp = tcp_sk(sk); 3200 u32 packets_acked; 3201 3202 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)); 3203 3204 packets_acked = tcp_skb_pcount(skb); 3205 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 3206 return 0; 3207 packets_acked -= tcp_skb_pcount(skb); 3208 3209 if (packets_acked) { 3210 BUG_ON(tcp_skb_pcount(skb) == 0); 3211 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)); 3212 } 3213 3214 return packets_acked; 3215} 3216 3217/* Remove acknowledged frames from the retransmission queue. If our packet 3218 * is before the ack sequence we can discard it as it's confirmed to have 3219 * arrived at the other end. 3220 */ 3221static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets, 3222 u32 prior_snd_una) 3223{ 3224 struct tcp_sock *tp = tcp_sk(sk); 3225 const struct inet_connection_sock *icsk = inet_csk(sk); 3226 struct sk_buff *skb; 3227 u32 now = tcp_time_stamp; 3228 int fully_acked = 1; 3229 int flag = 0; 3230 u32 pkts_acked = 0; 3231 u32 reord = tp->packets_out; 3232 u32 prior_sacked = tp->sacked_out; 3233 s32 seq_rtt = -1; 3234 s32 ca_seq_rtt = -1; 3235 ktime_t last_ackt = net_invalid_timestamp(); 3236 3237 while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) { 3238 struct tcp_skb_cb *scb = TCP_SKB_CB(skb); 3239 u32 acked_pcount; 3240 u8 sacked = scb->sacked; 3241 3242 /* Determine how many packets and what bytes were acked, tso and else */ 3243 if (after(scb->end_seq, tp->snd_una)) { 3244 if (tcp_skb_pcount(skb) == 1 || 3245 !after(tp->snd_una, scb->seq)) 3246 break; 3247 3248 acked_pcount = tcp_tso_acked(sk, skb); 3249 if (!acked_pcount) 3250 break; 3251 3252 fully_acked = 0; 3253 } else { 3254 acked_pcount = tcp_skb_pcount(skb); 3255 } 3256 3257 if (sacked & TCPCB_RETRANS) { 3258 if (sacked & TCPCB_SACKED_RETRANS) 3259 tp->retrans_out -= acked_pcount; 3260 flag |= FLAG_RETRANS_DATA_ACKED; 3261 ca_seq_rtt = -1; 3262 seq_rtt = -1; 3263 if ((flag & FLAG_DATA_ACKED) || (acked_pcount > 1)) 3264 flag |= FLAG_NONHEAD_RETRANS_ACKED; 3265 } else { 3266 ca_seq_rtt = now - scb->when; 3267 last_ackt = skb->tstamp; 3268 if (seq_rtt < 0) { 3269 seq_rtt = ca_seq_rtt; 3270 } 3271 if (!(sacked & TCPCB_SACKED_ACKED)) 3272 reord = min(pkts_acked, reord); 3273 } 3274 3275 if (sacked & TCPCB_SACKED_ACKED) 3276 tp->sacked_out -= acked_pcount; 3277 if (sacked & TCPCB_LOST) 3278 tp->lost_out -= acked_pcount; 3279 3280 tp->packets_out -= acked_pcount; 3281 pkts_acked += acked_pcount; 3282 3283 /* Initial outgoing SYN's get put onto the write_queue 3284 * just like anything else we transmit. It is not 3285 * true data, and if we misinform our callers that 3286 * this ACK acks real data, we will erroneously exit 3287 * connection startup slow start one packet too 3288 * quickly. This is severely frowned upon behavior. 3289 */ 3290 if (!(scb->flags & TCPHDR_SYN)) { 3291 flag |= FLAG_DATA_ACKED; 3292 } else { 3293 flag |= FLAG_SYN_ACKED; 3294 tp->retrans_stamp = 0; 3295 } 3296 3297 if (!fully_acked) 3298 break; 3299 3300 tcp_unlink_write_queue(skb, sk); 3301 sk_wmem_free_skb(sk, skb); 3302 tp->scoreboard_skb_hint = NULL; 3303 if (skb == tp->retransmit_skb_hint) 3304 tp->retransmit_skb_hint = NULL; 3305 if (skb == tp->lost_skb_hint) 3306 tp->lost_skb_hint = NULL; 3307 } 3308 3309 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una))) 3310 tp->snd_up = tp->snd_una; 3311 3312 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) 3313 flag |= FLAG_SACK_RENEGING; 3314 3315 if (flag & FLAG_ACKED) { 3316 const struct tcp_congestion_ops *ca_ops 3317 = inet_csk(sk)->icsk_ca_ops; 3318 3319 if (unlikely(icsk->icsk_mtup.probe_size && 3320 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) { 3321 tcp_mtup_probe_success(sk); 3322 } 3323 3324 tcp_ack_update_rtt(sk, flag, seq_rtt); 3325 tcp_rearm_rto(sk); 3326 3327 if (tcp_is_reno(tp)) { 3328 tcp_remove_reno_sacks(sk, pkts_acked); 3329 } else { 3330 int delta; 3331 3332 /* Non-retransmitted hole got filled? That's reordering */ 3333 if (reord < prior_fackets) 3334 tcp_update_reordering(sk, tp->fackets_out - reord, 0); 3335 3336 delta = tcp_is_fack(tp) ? pkts_acked : 3337 prior_sacked - tp->sacked_out; 3338 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta); 3339 } 3340 3341 tp->fackets_out -= min(pkts_acked, tp->fackets_out); 3342 3343 if (ca_ops->pkts_acked) { 3344 s32 rtt_us = -1; 3345 3346 /* Is the ACK triggering packet unambiguous? */ 3347 if (!(flag & FLAG_RETRANS_DATA_ACKED)) { 3348 /* High resolution needed and available? */ 3349 if (ca_ops->flags & TCP_CONG_RTT_STAMP && 3350 !ktime_equal(last_ackt, 3351 net_invalid_timestamp())) 3352 rtt_us = ktime_us_delta(ktime_get_real(), 3353 last_ackt); 3354 else if (ca_seq_rtt > 0) 3355 rtt_us = jiffies_to_usecs(ca_seq_rtt); 3356 } 3357 3358 ca_ops->pkts_acked(sk, pkts_acked, rtt_us); 3359 } 3360 } 3361 3362#if FASTRETRANS_DEBUG > 0 3363 WARN_ON((int)tp->sacked_out < 0); 3364 WARN_ON((int)tp->lost_out < 0); 3365 WARN_ON((int)tp->retrans_out < 0); 3366 if (!tp->packets_out && tcp_is_sack(tp)) { 3367 icsk = inet_csk(sk); 3368 if (tp->lost_out) { 3369 printk(KERN_DEBUG "Leak l=%u %d\n", 3370 tp->lost_out, icsk->icsk_ca_state); 3371 tp->lost_out = 0; 3372 } 3373 if (tp->sacked_out) { 3374 printk(KERN_DEBUG "Leak s=%u %d\n", 3375 tp->sacked_out, icsk->icsk_ca_state); 3376 tp->sacked_out = 0; 3377 } 3378 if (tp->retrans_out) { 3379 printk(KERN_DEBUG "Leak r=%u %d\n", 3380 tp->retrans_out, icsk->icsk_ca_state); 3381 tp->retrans_out = 0; 3382 } 3383 } 3384#endif 3385 return flag; 3386} 3387 3388static void tcp_ack_probe(struct sock *sk) 3389{ 3390 const struct tcp_sock *tp = tcp_sk(sk); 3391 struct inet_connection_sock *icsk = inet_csk(sk); 3392 3393 /* Was it a usable window open? */ 3394 3395 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) { 3396 icsk->icsk_backoff = 0; 3397 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0); 3398 /* Socket must be waked up by subsequent tcp_data_snd_check(). 3399 * This function is not for random using! 3400 */ 3401 } else { 3402 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 3403 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), 3404 TCP_RTO_MAX); 3405 } 3406} 3407 3408static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag) 3409{ 3410 return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) || 3411 inet_csk(sk)->icsk_ca_state != TCP_CA_Open); 3412} 3413 3414static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag) 3415{ 3416 const struct tcp_sock *tp = tcp_sk(sk); 3417 return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) && 3418 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR)); 3419} 3420 3421/* Check that window update is acceptable. 3422 * The function assumes that snd_una<=ack<=snd_next. 3423 */ 3424static inline int tcp_may_update_window(const struct tcp_sock *tp, 3425 const u32 ack, const u32 ack_seq, 3426 const u32 nwin) 3427{ 3428 return (after(ack, tp->snd_una) || 3429 after(ack_seq, tp->snd_wl1) || 3430 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd)); 3431} 3432 3433/* Update our send window. 3434 * 3435 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2 3436 * and in FreeBSD. NetBSD's one is even worse.) is wrong. 3437 */ 3438static int tcp_ack_update_window(struct sock *sk, struct sk_buff *skb, u32 ack, 3439 u32 ack_seq) 3440{ 3441 struct tcp_sock *tp = tcp_sk(sk); 3442 int flag = 0; 3443 u32 nwin = ntohs(tcp_hdr(skb)->window); 3444 3445 if (likely(!tcp_hdr(skb)->syn)) 3446 nwin <<= tp->rx_opt.snd_wscale; 3447 3448 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) { 3449 flag |= FLAG_WIN_UPDATE; 3450 tcp_update_wl(tp, ack_seq); 3451 3452 if (tp->snd_wnd != nwin) { 3453 tp->snd_wnd = nwin; 3454 3455 /* Note, it is the only place, where 3456 * fast path is recovered for sending TCP. 3457 */ 3458 tp->pred_flags = 0; 3459 tcp_fast_path_check(sk); 3460 3461 if (nwin > tp->max_window) { 3462 tp->max_window = nwin; 3463 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie); 3464 } 3465 } 3466 } 3467 3468 tp->snd_una = ack; 3469 3470 return flag; 3471} 3472 3473/* A very conservative spurious RTO response algorithm: reduce cwnd and 3474 * continue in congestion avoidance. 3475 */ 3476static void tcp_conservative_spur_to_response(struct tcp_sock *tp) 3477{ 3478 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh); 3479 tp->snd_cwnd_cnt = 0; 3480 tp->bytes_acked = 0; 3481 TCP_ECN_queue_cwr(tp); 3482 tcp_moderate_cwnd(tp); 3483} 3484 3485/* A conservative spurious RTO response algorithm: reduce cwnd using 3486 * rate halving and continue in congestion avoidance. 3487 */ 3488static void tcp_ratehalving_spur_to_response(struct sock *sk) 3489{ 3490 tcp_enter_cwr(sk, 0); 3491} 3492 3493static void tcp_undo_spur_to_response(struct sock *sk, int flag) 3494{ 3495 if (flag & FLAG_ECE) 3496 tcp_ratehalving_spur_to_response(sk); 3497 else 3498 tcp_undo_cwr(sk, 1); 3499} 3500 3501/* F-RTO spurious RTO detection algorithm (RFC4138) 3502 * 3503 * F-RTO affects during two new ACKs following RTO (well, almost, see inline 3504 * comments). State (ACK number) is kept in frto_counter. When ACK advances 3505 * window (but not to or beyond highest sequence sent before RTO): 3506 * On First ACK, send two new segments out. 3507 * On Second ACK, RTO was likely spurious. Do spurious response (response 3508 * algorithm is not part of the F-RTO detection algorithm 3509 * given in RFC4138 but can be selected separately). 3510 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss 3511 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding 3512 * of Nagle, this is done using frto_counter states 2 and 3, when a new data 3513 * segment of any size sent during F-RTO, state 2 is upgraded to 3. 3514 * 3515 * Rationale: if the RTO was spurious, new ACKs should arrive from the 3516 * original window even after we transmit two new data segments. 3517 * 3518 * SACK version: 3519 * on first step, wait until first cumulative ACK arrives, then move to 3520 * the second step. In second step, the next ACK decides. 3521 * 3522 * F-RTO is implemented (mainly) in four functions: 3523 * - tcp_use_frto() is used to determine if TCP is can use F-RTO 3524 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is 3525 * called when tcp_use_frto() showed green light 3526 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm 3527 * - tcp_enter_frto_loss() is called if there is not enough evidence 3528 * to prove that the RTO is indeed spurious. It transfers the control 3529 * from F-RTO to the conventional RTO recovery 3530 */ 3531static int tcp_process_frto(struct sock *sk, int flag) 3532{ 3533 struct tcp_sock *tp = tcp_sk(sk); 3534 3535 tcp_verify_left_out(tp); 3536 3537 /* Duplicate the behavior from Loss state (fastretrans_alert) */ 3538 if (flag & FLAG_DATA_ACKED) 3539 inet_csk(sk)->icsk_retransmits = 0; 3540 3541 if ((flag & FLAG_NONHEAD_RETRANS_ACKED) || 3542 ((tp->frto_counter >= 2) && (flag & FLAG_RETRANS_DATA_ACKED))) 3543 tp->undo_marker = 0; 3544 3545 if (!before(tp->snd_una, tp->frto_highmark)) { 3546 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag); 3547 return 1; 3548 } 3549 3550 if (!tcp_is_sackfrto(tp)) { 3551 /* RFC4138 shortcoming in step 2; should also have case c): 3552 * ACK isn't duplicate nor advances window, e.g., opposite dir 3553 * data, winupdate 3554 */ 3555 if (!(flag & FLAG_ANY_PROGRESS) && (flag & FLAG_NOT_DUP)) 3556 return 1; 3557 3558 if (!(flag & FLAG_DATA_ACKED)) { 3559 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3), 3560 flag); 3561 return 1; 3562 } 3563 } else { 3564 if (!(flag & FLAG_DATA_ACKED) && (tp->frto_counter == 1)) { 3565 /* Prevent sending of new data. */ 3566 tp->snd_cwnd = min(tp->snd_cwnd, 3567 tcp_packets_in_flight(tp)); 3568 return 1; 3569 } 3570 3571 if ((tp->frto_counter >= 2) && 3572 (!(flag & FLAG_FORWARD_PROGRESS) || 3573 ((flag & FLAG_DATA_SACKED) && 3574 !(flag & FLAG_ONLY_ORIG_SACKED)))) { 3575 /* RFC4138 shortcoming (see comment above) */ 3576 if (!(flag & FLAG_FORWARD_PROGRESS) && 3577 (flag & FLAG_NOT_DUP)) 3578 return 1; 3579 3580 tcp_enter_frto_loss(sk, 3, flag); 3581 return 1; 3582 } 3583 } 3584 3585 if (tp->frto_counter == 1) { 3586 /* tcp_may_send_now needs to see updated state */ 3587 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2; 3588 tp->frto_counter = 2; 3589 3590 if (!tcp_may_send_now(sk)) 3591 tcp_enter_frto_loss(sk, 2, flag); 3592 3593 return 1; 3594 } else { 3595 switch (sysctl_tcp_frto_response) { 3596 case 2: 3597 tcp_undo_spur_to_response(sk, flag); 3598 break; 3599 case 1: 3600 tcp_conservative_spur_to_response(tp); 3601 break; 3602 default: 3603 tcp_ratehalving_spur_to_response(sk); 3604 break; 3605 } 3606 tp->frto_counter = 0; 3607 tp->undo_marker = 0; 3608 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSPURIOUSRTOS); 3609 } 3610 return 0; 3611} 3612 3613/* This routine deals with incoming acks, but not outgoing ones. */ 3614static int BCMFASTPATH_HOST tcp_ack(struct sock *sk, struct sk_buff *skb, int flag) 3615{ 3616 struct inet_connection_sock *icsk = inet_csk(sk); 3617 struct tcp_sock *tp = tcp_sk(sk); 3618 u32 prior_snd_una = tp->snd_una; 3619 u32 ack_seq = TCP_SKB_CB(skb)->seq; 3620 u32 ack = TCP_SKB_CB(skb)->ack_seq; 3621 u32 prior_in_flight; 3622 u32 prior_fackets; 3623 int prior_packets; 3624 int frto_cwnd = 0; 3625 3626 /* If the ack is older than previous acks 3627 * then we can probably ignore it. 3628 */ 3629 if (before(ack, prior_snd_una)) 3630 goto old_ack; 3631 3632 /* If the ack includes data we haven't sent yet, discard 3633 * this segment (RFC793 Section 3.9). 3634 */ 3635 if (after(ack, tp->snd_nxt)) 3636 goto invalid_ack; 3637 3638 if (after(ack, prior_snd_una)) 3639 flag |= FLAG_SND_UNA_ADVANCED; 3640 3641 if (sysctl_tcp_abc) { 3642 if (icsk->icsk_ca_state < TCP_CA_CWR) 3643 tp->bytes_acked += ack - prior_snd_una; 3644 else if (icsk->icsk_ca_state == TCP_CA_Loss) 3645 /* we assume just one segment left network */ 3646 tp->bytes_acked += min(ack - prior_snd_una, 3647 tp->mss_cache); 3648 } 3649 3650 prior_fackets = tp->fackets_out; 3651 prior_in_flight = tcp_packets_in_flight(tp); 3652 3653 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) { 3654 /* Window is constant, pure forward advance. 3655 * No more checks are required. 3656 * Note, we use the fact that SND.UNA>=SND.WL2. 3657 */ 3658 tcp_update_wl(tp, ack_seq); 3659 tp->snd_una = ack; 3660 flag |= FLAG_WIN_UPDATE; 3661 3662 tcp_ca_event(sk, CA_EVENT_FAST_ACK); 3663 3664 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS); 3665 } else { 3666 if (ack_seq != TCP_SKB_CB(skb)->end_seq) 3667 flag |= FLAG_DATA; 3668 else 3669 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS); 3670 3671 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq); 3672 3673 if (TCP_SKB_CB(skb)->sacked) 3674 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una); 3675 3676 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb))) 3677 flag |= FLAG_ECE; 3678 3679 tcp_ca_event(sk, CA_EVENT_SLOW_ACK); 3680 } 3681 3682 /* We passed data and got it acked, remove any soft error 3683 * log. Something worked... 3684 */ 3685 sk->sk_err_soft = 0; 3686 icsk->icsk_probes_out = 0; 3687 tp->rcv_tstamp = tcp_time_stamp; 3688 prior_packets = tp->packets_out; 3689 if (!prior_packets) 3690 goto no_queue; 3691 3692 /* See if we can take anything off of the retransmit queue. */ 3693 flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una); 3694 3695 if (tp->frto_counter) 3696 frto_cwnd = tcp_process_frto(sk, flag); 3697 /* Guarantee sacktag reordering detection against wrap-arounds */ 3698 if (before(tp->frto_highmark, tp->snd_una)) 3699 tp->frto_highmark = 0; 3700 3701 if (tcp_ack_is_dubious(sk, flag)) { 3702 /* Advance CWND, if state allows this. */ 3703 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd && 3704 tcp_may_raise_cwnd(sk, flag)) 3705 tcp_cong_avoid(sk, ack, prior_in_flight); 3706 tcp_fastretrans_alert(sk, prior_packets - tp->packets_out, 3707 flag); 3708 } else { 3709 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd) 3710 tcp_cong_avoid(sk, ack, prior_in_flight); 3711 } 3712 3713 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) 3714 dst_confirm(__sk_dst_get(sk)); 3715 3716 return 1; 3717 3718no_queue: 3719 /* If this ack opens up a zero window, clear backoff. It was 3720 * being used to time the probes, and is probably far higher than 3721 * it needs to be for normal retransmission. 3722 */ 3723 if (tcp_send_head(sk)) 3724 tcp_ack_probe(sk); 3725 return 1; 3726 3727invalid_ack: 3728 SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt); 3729 return -1; 3730 3731old_ack: 3732 if (TCP_SKB_CB(skb)->sacked) { 3733 tcp_sacktag_write_queue(sk, skb, prior_snd_una); 3734 if (icsk->icsk_ca_state == TCP_CA_Open) 3735 tcp_try_keep_open(sk); 3736 } 3737 3738 SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt); 3739 return 0; 3740} 3741 3742/* Look for tcp options. Normally only called on SYN and SYNACK packets. 3743 * But, this can also be called on packets in the established flow when 3744 * the fast version below fails. 3745 */ 3746void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx, 3747 u8 **hvpp, int estab) 3748{ 3749 unsigned char *ptr; 3750 struct tcphdr *th = tcp_hdr(skb); 3751 int length = (th->doff * 4) - sizeof(struct tcphdr); 3752 3753 ptr = (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 if (net_ratelimit()) 3791 printk(KERN_INFO "tcp_parse_options: Illegal window " 3792 "scaling value %d >14 received.\n", 3793 snd_wscale); 3794 snd_wscale = 14; 3795 } 3796 opt_rx->snd_wscale = snd_wscale; 3797 } 3798 break; 3799 case TCPOPT_TIMESTAMP: 3800 if ((opsize == TCPOLEN_TIMESTAMP) && 3801 ((estab && opt_rx->tstamp_ok) || 3802 (!estab && sysctl_tcp_timestamps))) { 3803 opt_rx->saw_tstamp = 1; 3804 opt_rx->rcv_tsval = get_unaligned_be32(ptr); 3805 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4); 3806 } 3807 break; 3808 case TCPOPT_SACK_PERM: 3809 if (opsize == TCPOLEN_SACK_PERM && th->syn && 3810 !estab && sysctl_tcp_sack) { 3811 opt_rx->sack_ok = 1; 3812 tcp_sack_reset(opt_rx); 3813 } 3814 break; 3815 3816 case TCPOPT_SACK: 3817 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) && 3818 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) && 3819 opt_rx->sack_ok) { 3820 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th; 3821 } 3822 break; 3823#ifdef CONFIG_TCP_MD5SIG 3824 case TCPOPT_MD5SIG: 3825 /* 3826 * The MD5 Hash has already been 3827 * checked (see tcp_v{4,6}_do_rcv()). 3828 */ 3829 break; 3830#endif 3831 case TCPOPT_COOKIE: 3832 /* This option is variable length. 3833 */ 3834 switch (opsize) { 3835 case TCPOLEN_COOKIE_BASE: 3836 /* not yet implemented */ 3837 break; 3838 case TCPOLEN_COOKIE_PAIR: 3839 /* not yet implemented */ 3840 break; 3841 case TCPOLEN_COOKIE_MIN+0: 3842 case TCPOLEN_COOKIE_MIN+2: 3843 case TCPOLEN_COOKIE_MIN+4: 3844 case TCPOLEN_COOKIE_MIN+6: 3845 case TCPOLEN_COOKIE_MAX: 3846 /* 16-bit multiple */ 3847 opt_rx->cookie_plus = opsize; 3848 *hvpp = ptr; 3849 break; 3850 default: 3851 /* ignore option */ 3852 break; 3853 } 3854 break; 3855 } 3856 3857 ptr += opsize-2; 3858 length -= opsize; 3859 } 3860 } 3861} 3862EXPORT_SYMBOL(tcp_parse_options); 3863 3864static int tcp_parse_aligned_timestamp(struct tcp_sock *tp, struct tcphdr *th) 3865{ 3866 __be32 *ptr = (__be32 *)(th + 1); 3867 3868 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) 3869 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) { 3870 tp->rx_opt.saw_tstamp = 1; 3871 ++ptr; 3872 tp->rx_opt.rcv_tsval = ntohl(*ptr); 3873 ++ptr; 3874 tp->rx_opt.rcv_tsecr = ntohl(*ptr); 3875 return 1; 3876 } 3877 return 0; 3878} 3879 3880/* Fast parse options. This hopes to only see timestamps. 3881 * If it is wrong it falls back on tcp_parse_options(). 3882 */ 3883static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th, 3884 struct tcp_sock *tp, u8 **hvpp) 3885{ 3886 /* In the spirit of fast parsing, compare doff directly to constant 3887 * values. Because equality is used, short doff can be ignored here. 3888 */ 3889 if (th->doff == (sizeof(*th) / 4)) { 3890 tp->rx_opt.saw_tstamp = 0; 3891 return 0; 3892 } else if (tp->rx_opt.tstamp_ok && 3893 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) { 3894 if (tcp_parse_aligned_timestamp(tp, th)) 3895 return 1; 3896 } 3897 tcp_parse_options(skb, &tp->rx_opt, hvpp, 1); 3898 return 1; 3899} 3900 3901#ifdef CONFIG_TCP_MD5SIG 3902/* 3903 * Parse MD5 Signature option 3904 */ 3905u8 *tcp_parse_md5sig_option(struct tcphdr *th) 3906{ 3907 int length = (th->doff << 2) - sizeof (*th); 3908 u8 *ptr = (u8*)(th + 1); 3909 3910 /* If the TCP option is too short, we can short cut */ 3911 if (length < TCPOLEN_MD5SIG) 3912 return NULL; 3913 3914 while (length > 0) { 3915 int opcode = *ptr++; 3916 int opsize; 3917 3918 switch(opcode) { 3919 case TCPOPT_EOL: 3920 return NULL; 3921 case TCPOPT_NOP: 3922 length--; 3923 continue; 3924 default: 3925 opsize = *ptr++; 3926 if (opsize < 2 || opsize > length) 3927 return NULL; 3928 if (opcode == TCPOPT_MD5SIG) 3929 return opsize == TCPOLEN_MD5SIG ? ptr : NULL; 3930 } 3931 ptr += opsize - 2; 3932 length -= opsize; 3933 } 3934 return NULL; 3935} 3936EXPORT_SYMBOL(tcp_parse_md5sig_option); 3937#endif 3938 3939static inline void tcp_store_ts_recent(struct tcp_sock *tp) 3940{ 3941 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval; 3942 tp->rx_opt.ts_recent_stamp = get_seconds(); 3943} 3944 3945static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq) 3946{ 3947 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) { 3948 3949 if (tcp_paws_check(&tp->rx_opt, 0)) 3950 tcp_store_ts_recent(tp); 3951 } 3952} 3953 3954/* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM 3955 * 3956 * It is not fatal. If this ACK does _not_ change critical state (seqs, window) 3957 * it can pass through stack. So, the following predicate verifies that 3958 * this segment is not used for anything but congestion avoidance or 3959 * fast retransmit. Moreover, we even are able to eliminate most of such 3960 * second order effects, if we apply some small "replay" window (~RTO) 3961 * to timestamp space. 3962 * 3963 * All these measures still do not guarantee that we reject wrapped ACKs 3964 * on networks with high bandwidth, when sequence space is recycled fastly, 3965 * but it guarantees that such events will be very rare and do not affect 3966 * connection seriously. This doesn't look nice, but alas, PAWS is really 3967 * buggy extension. 3968 * 3969 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC 3970 * states that events when retransmit arrives after original data are rare. 3971 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is 3972 * the biggest problem on large power networks even with minor reordering. 3973 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe 3974 * up to bandwidth of 18Gigabit/sec. 8) ] 3975 */ 3976 3977static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb) 3978{ 3979 struct tcp_sock *tp = tcp_sk(sk); 3980 struct tcphdr *th = tcp_hdr(skb); 3981 u32 seq = TCP_SKB_CB(skb)->seq; 3982 u32 ack = TCP_SKB_CB(skb)->ack_seq; 3983 3984 return (/* 1. Pure ACK with correct sequence number. */ 3985 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) && 3986 3987 /* 2. ... and duplicate ACK. */ 3988 ack == tp->snd_una && 3989 3990 /* 3. ... and does not update window. */ 3991 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) && 3992 3993 /* 4. ... and sits in replay window. */ 3994 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ); 3995} 3996 3997static inline int tcp_paws_discard(const struct sock *sk, 3998 const struct sk_buff *skb) 3999{ 4000 const struct tcp_sock *tp = tcp_sk(sk); 4001 4002 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) && 4003 !tcp_disordered_ack(sk, skb); 4004} 4005 4006/* Check segment sequence number for validity. 4007 * 4008 * Segment controls are considered valid, if the segment 4009 * fits to the window after truncation to the window. Acceptability 4010 * of data (and SYN, FIN, of course) is checked separately. 4011 * See tcp_data_queue(), for example. 4012 * 4013 * Also, controls (RST is main one) are accepted using RCV.WUP instead 4014 * of RCV.NXT. Peer still did not advance his SND.UNA when we 4015 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP. 4016 * (borrowed from freebsd) 4017 */ 4018 4019static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq) 4020{ 4021 return !before(end_seq, tp->rcv_wup) && 4022 !after(seq, tp->rcv_nxt + tcp_receive_window(tp)); 4023} 4024 4025/* When we get a reset we do this. */ 4026static void tcp_reset(struct sock *sk) 4027{ 4028 /* We want the right error as BSD sees it (and indeed as we do). */ 4029 switch (sk->sk_state) { 4030 case TCP_SYN_SENT: 4031 sk->sk_err = ECONNREFUSED; 4032 break; 4033 case TCP_CLOSE_WAIT: 4034 sk->sk_err = EPIPE; 4035 break; 4036 case TCP_CLOSE: 4037 return; 4038 default: 4039 sk->sk_err = ECONNRESET; 4040 } 4041 /* This barrier is coupled with smp_rmb() in tcp_poll() */ 4042 smp_wmb(); 4043 4044 if (!sock_flag(sk, SOCK_DEAD)) 4045 sk->sk_error_report(sk); 4046 4047 tcp_done(sk); 4048} 4049 4050/* 4051 * Process the FIN bit. This now behaves as it is supposed to work 4052 * and the FIN takes effect when it is validly part of sequence 4053 * space. Not before when we get holes. 4054 * 4055 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT 4056 * (and thence onto LAST-ACK and finally, CLOSE, we never enter 4057 * TIME-WAIT) 4058 * 4059 * If we are in FINWAIT-1, a received FIN indicates simultaneous 4060 * close and we go into CLOSING (and later onto TIME-WAIT) 4061 * 4062 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT. 4063 */ 4064static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th) 4065{ 4066 struct tcp_sock *tp = tcp_sk(sk); 4067 4068 inet_csk_schedule_ack(sk); 4069 4070 sk->sk_shutdown |= RCV_SHUTDOWN; 4071 sock_set_flag(sk, SOCK_DONE); 4072 4073 switch (sk->sk_state) { 4074 case TCP_SYN_RECV: 4075 case TCP_ESTABLISHED: 4076 /* Move to CLOSE_WAIT */ 4077 tcp_set_state(sk, TCP_CLOSE_WAIT); 4078 inet_csk(sk)->icsk_ack.pingpong = 1; 4079 break; 4080 4081 case TCP_CLOSE_WAIT: 4082 case TCP_CLOSING: 4083 /* Received a retransmission of the FIN, do 4084 * nothing. 4085 */ 4086 break; 4087 case TCP_LAST_ACK: 4088 /* RFC793: Remain in the LAST-ACK state. */ 4089 break; 4090 4091 case TCP_FIN_WAIT1: 4092 /* This case occurs when a simultaneous close 4093 * happens, we must ack the received FIN and 4094 * enter the CLOSING state. 4095 */ 4096 tcp_send_ack(sk); 4097 tcp_set_state(sk, TCP_CLOSING); 4098 break; 4099 case TCP_FIN_WAIT2: 4100 /* Received a FIN -- send ACK and enter TIME_WAIT. */ 4101 tcp_send_ack(sk); 4102 tcp_time_wait(sk, TCP_TIME_WAIT, 0); 4103 break; 4104 default: 4105 /* Only TCP_LISTEN and TCP_CLOSE are left, in these 4106 * cases we should never reach this piece of code. 4107 */ 4108 printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n", 4109 __func__, sk->sk_state); 4110 break; 4111 } 4112 4113 /* It _is_ possible, that we have something out-of-order _after_ FIN. 4114 * Probably, we should reset in this case. For now drop them. 4115 */ 4116 __skb_queue_purge(&tp->out_of_order_queue); 4117 if (tcp_is_sack(tp)) 4118 tcp_sack_reset(&tp->rx_opt); 4119 sk_mem_reclaim(sk); 4120 4121 if (!sock_flag(sk, SOCK_DEAD)) { 4122 sk->sk_state_change(sk); 4123 4124 /* Do not send POLL_HUP for half duplex close. */ 4125 if (sk->sk_shutdown == SHUTDOWN_MASK || 4126 sk->sk_state == TCP_CLOSE) 4127 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP); 4128 else 4129 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN); 4130 } 4131} 4132 4133static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, 4134 u32 end_seq) 4135{ 4136 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) { 4137 if (before(seq, sp->start_seq)) 4138 sp->start_seq = seq; 4139 if (after(end_seq, sp->end_seq)) 4140 sp->end_seq = end_seq; 4141 return 1; 4142 } 4143 return 0; 4144} 4145 4146static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq) 4147{ 4148 struct tcp_sock *tp = tcp_sk(sk); 4149 4150 if (tcp_is_sack(tp) && sysctl_tcp_dsack) { 4151 int mib_idx; 4152 4153 if (before(seq, tp->rcv_nxt)) 4154 mib_idx = LINUX_MIB_TCPDSACKOLDSENT; 4155 else 4156 mib_idx = LINUX_MIB_TCPDSACKOFOSENT; 4157 4158 NET_INC_STATS_BH(sock_net(sk), mib_idx); 4159 4160 tp->rx_opt.dsack = 1; 4161 tp->duplicate_sack[0].start_seq = seq; 4162 tp->duplicate_sack[0].end_seq = end_seq; 4163 } 4164} 4165 4166static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq) 4167{ 4168 struct tcp_sock *tp = tcp_sk(sk); 4169 4170 if (!tp->rx_opt.dsack) 4171 tcp_dsack_set(sk, seq, end_seq); 4172 else 4173 tcp_sack_extend(tp->duplicate_sack, seq, end_seq); 4174} 4175 4176static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb) 4177{ 4178 struct tcp_sock *tp = tcp_sk(sk); 4179 4180 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 4181 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 4182 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST); 4183 tcp_enter_quickack_mode(sk); 4184 4185 if (tcp_is_sack(tp) && sysctl_tcp_dsack) { 4186 u32 end_seq = TCP_SKB_CB(skb)->end_seq; 4187 4188 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) 4189 end_seq = tp->rcv_nxt; 4190 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq); 4191 } 4192 } 4193 4194 tcp_send_ack(sk); 4195} 4196 4197/* These routines update the SACK block as out-of-order packets arrive or 4198 * in-order packets close up the sequence space. 4199 */ 4200static void tcp_sack_maybe_coalesce(struct tcp_sock *tp) 4201{ 4202 int this_sack; 4203 struct tcp_sack_block *sp = &tp->selective_acks[0]; 4204 struct tcp_sack_block *swalk = sp + 1; 4205 4206 /* See if the recent change to the first SACK eats into 4207 * or hits the sequence space of other SACK blocks, if so coalesce. 4208 */ 4209 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) { 4210 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) { 4211 int i; 4212 4213 /* Zap SWALK, by moving every further SACK up by one slot. 4214 * Decrease num_sacks. 4215 */ 4216 tp->rx_opt.num_sacks--; 4217 for (i = this_sack; i < tp->rx_opt.num_sacks; i++) 4218 sp[i] = sp[i + 1]; 4219 continue; 4220 } 4221 this_sack++, swalk++; 4222 } 4223} 4224 4225static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq) 4226{ 4227 struct tcp_sock *tp = tcp_sk(sk); 4228 struct tcp_sack_block *sp = &tp->selective_acks[0]; 4229 int cur_sacks = tp->rx_opt.num_sacks; 4230 int this_sack; 4231 4232 if (!cur_sacks) 4233 goto new_sack; 4234 4235 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) { 4236 if (tcp_sack_extend(sp, seq, end_seq)) { 4237 /* Rotate this_sack to the first one. */ 4238 for (; this_sack > 0; this_sack--, sp--) 4239 swap(*sp, *(sp - 1)); 4240 if (cur_sacks > 1) 4241 tcp_sack_maybe_coalesce(tp); 4242 return; 4243 } 4244 } 4245 4246 /* Could not find an adjacent existing SACK, build a new one, 4247 * put it at the front, and shift everyone else down. We 4248 * always know there is at least one SACK present already here. 4249 * 4250 * If the sack array is full, forget about the last one. 4251 */ 4252 if (this_sack >= TCP_NUM_SACKS) { 4253 this_sack--; 4254 tp->rx_opt.num_sacks--; 4255 sp--; 4256 } 4257 for (; this_sack > 0; this_sack--, sp--) 4258 *sp = *(sp - 1); 4259 4260new_sack: 4261 /* Build the new head SACK, and we're done. */ 4262 sp->start_seq = seq; 4263 sp->end_seq = end_seq; 4264 tp->rx_opt.num_sacks++; 4265} 4266 4267/* RCV.NXT advances, some SACKs should be eaten. */ 4268 4269static void tcp_sack_remove(struct tcp_sock *tp) 4270{ 4271 struct tcp_sack_block *sp = &tp->selective_acks[0]; 4272 int num_sacks = tp->rx_opt.num_sacks; 4273 int this_sack; 4274 4275 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */ 4276 if (skb_queue_empty(&tp->out_of_order_queue)) { 4277 tp->rx_opt.num_sacks = 0; 4278 return; 4279 } 4280 4281 for (this_sack = 0; this_sack < num_sacks;) { 4282 /* Check if the start of the sack is covered by RCV.NXT. */ 4283 if (!before(tp->rcv_nxt, sp->start_seq)) { 4284 int i; 4285 4286 /* RCV.NXT must cover all the block! */ 4287 WARN_ON(before(tp->rcv_nxt, sp->end_seq)); 4288 4289 /* Zap this SACK, by moving forward any other SACKS. */ 4290 for (i=this_sack+1; i < num_sacks; i++) 4291 tp->selective_acks[i-1] = tp->selective_acks[i]; 4292 num_sacks--; 4293 continue; 4294 } 4295 this_sack++; 4296 sp++; 4297 } 4298 tp->rx_opt.num_sacks = num_sacks; 4299} 4300 4301/* This one checks to see if we can put data from the 4302 * out_of_order queue into the receive_queue. 4303 */ 4304static void tcp_ofo_queue(struct sock *sk) 4305{ 4306 struct tcp_sock *tp = tcp_sk(sk); 4307 __u32 dsack_high = tp->rcv_nxt; 4308 struct sk_buff *skb; 4309 4310 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) { 4311 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) 4312 break; 4313 4314 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) { 4315 __u32 dsack = dsack_high; 4316 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high)) 4317 dsack_high = TCP_SKB_CB(skb)->end_seq; 4318 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack); 4319 } 4320 4321 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) { 4322 SOCK_DEBUG(sk, "ofo packet was already received\n"); 4323 __skb_unlink(skb, &tp->out_of_order_queue); 4324 __kfree_skb(skb); 4325 continue; 4326 } 4327 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n", 4328 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, 4329 TCP_SKB_CB(skb)->end_seq); 4330 4331 __skb_unlink(skb, &tp->out_of_order_queue); 4332 __skb_queue_tail(&sk->sk_receive_queue, skb); 4333 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; 4334 if (tcp_hdr(skb)->fin) 4335 tcp_fin(skb, sk, tcp_hdr(skb)); 4336 } 4337} 4338 4339static int tcp_prune_ofo_queue(struct sock *sk); 4340static int tcp_prune_queue(struct sock *sk); 4341 4342static inline int tcp_try_rmem_schedule(struct sock *sk, unsigned int size) 4343{ 4344 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf || 4345 !sk_rmem_schedule(sk, size)) { 4346 4347 if (tcp_prune_queue(sk) < 0) 4348 return -1; 4349 4350 if (!sk_rmem_schedule(sk, size)) { 4351 if (!tcp_prune_ofo_queue(sk)) 4352 return -1; 4353 4354 if (!sk_rmem_schedule(sk, size)) 4355 return -1; 4356 } 4357 } 4358 return 0; 4359} 4360 4361static void tcp_data_queue(struct sock *sk, struct sk_buff *skb) 4362{ 4363 struct tcphdr *th = tcp_hdr(skb); 4364 struct tcp_sock *tp = tcp_sk(sk); 4365 int eaten = -1; 4366 4367 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) 4368 goto drop; 4369 4370 skb_dst_drop(skb); 4371 __skb_pull(skb, th->doff * 4); 4372 4373 TCP_ECN_accept_cwr(tp, skb); 4374 4375 tp->rx_opt.dsack = 0; 4376 4377 /* Queue data for delivery to the user. 4378 * Packets in sequence go to the receive queue. 4379 * Out of sequence packets to the out_of_order_queue. 4380 */ 4381 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) { 4382 if (tcp_receive_window(tp) == 0) 4383 goto out_of_window; 4384 4385 /* Ok. In sequence. In window. */ 4386 if (tp->ucopy.task == current && 4387 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len && 4388 sock_owned_by_user(sk) && !tp->urg_data) { 4389 int chunk = min_t(unsigned int, skb->len, 4390 tp->ucopy.len); 4391 4392 __set_current_state(TASK_RUNNING); 4393 4394 local_bh_enable(); 4395 if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) { 4396 tp->ucopy.len -= chunk; 4397 tp->copied_seq += chunk; 4398 eaten = (chunk == skb->len && !th->fin); 4399 tcp_rcv_space_adjust(sk); 4400 } 4401 local_bh_disable(); 4402 } 4403 4404 if (eaten <= 0) { 4405queue_and_out: 4406 if (eaten < 0 && 4407 tcp_try_rmem_schedule(sk, skb->truesize)) 4408 goto drop; 4409 4410 skb_set_owner_r(skb, sk); 4411 __skb_queue_tail(&sk->sk_receive_queue, skb); 4412 } 4413 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; 4414 if (skb->len) 4415 tcp_event_data_recv(sk, skb); 4416 if (th->fin) 4417 tcp_fin(skb, sk, th); 4418 4419 if (!skb_queue_empty(&tp->out_of_order_queue)) { 4420 tcp_ofo_queue(sk); 4421 4422 /* RFC2581. 4.2. SHOULD send immediate ACK, when 4423 * gap in queue is filled. 4424 */ 4425 if (skb_queue_empty(&tp->out_of_order_queue)) 4426 inet_csk(sk)->icsk_ack.pingpong = 0; 4427 } 4428 4429 if (tp->rx_opt.num_sacks) 4430 tcp_sack_remove(tp); 4431 4432 tcp_fast_path_check(sk); 4433 4434 if (eaten > 0) 4435 __kfree_skb(skb); 4436 else if (!sock_flag(sk, SOCK_DEAD)) 4437 sk->sk_data_ready(sk, 0); 4438 return; 4439 } 4440 4441 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) { 4442 /* A retransmit, 2nd most common case. Force an immediate ack. */ 4443 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST); 4444 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); 4445 4446out_of_window: 4447 tcp_enter_quickack_mode(sk); 4448 inet_csk_schedule_ack(sk); 4449drop: 4450 __kfree_skb(skb); 4451 return; 4452 } 4453 4454 /* Out of window. F.e. zero window probe. */ 4455 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp))) 4456 goto out_of_window; 4457 4458 tcp_enter_quickack_mode(sk); 4459 4460 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 4461 /* Partial packet, seq < rcv_next < end_seq */ 4462 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n", 4463 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, 4464 TCP_SKB_CB(skb)->end_seq); 4465 4466 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt); 4467 4468 /* If window is closed, drop tail of packet. But after 4469 * remembering D-SACK for its head made in previous line. 4470 */ 4471 if (!tcp_receive_window(tp)) 4472 goto out_of_window; 4473 goto queue_and_out; 4474 } 4475 4476 TCP_ECN_check_ce(tp, skb); 4477 4478 if (tcp_try_rmem_schedule(sk, skb->truesize)) 4479 goto drop; 4480 4481 /* Disable header prediction. */ 4482 tp->pred_flags = 0; 4483 inet_csk_schedule_ack(sk); 4484 4485 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n", 4486 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); 4487 4488 skb_set_owner_r(skb, sk); 4489 4490 if (!skb_peek(&tp->out_of_order_queue)) { 4491 /* Initial out of order segment, build 1 SACK. */ 4492 if (tcp_is_sack(tp)) { 4493 tp->rx_opt.num_sacks = 1; 4494 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq; 4495 tp->selective_acks[0].end_seq = 4496 TCP_SKB_CB(skb)->end_seq; 4497 } 4498 __skb_queue_head(&tp->out_of_order_queue, skb); 4499 } else { 4500 struct sk_buff *skb1 = skb_peek_tail(&tp->out_of_order_queue); 4501 u32 seq = TCP_SKB_CB(skb)->seq; 4502 u32 end_seq = TCP_SKB_CB(skb)->end_seq; 4503 4504 if (seq == TCP_SKB_CB(skb1)->end_seq) { 4505 __skb_queue_after(&tp->out_of_order_queue, skb1, skb); 4506 4507 if (!tp->rx_opt.num_sacks || 4508 tp->selective_acks[0].end_seq != seq) 4509 goto add_sack; 4510 4511 /* Common case: data arrive in order after hole. */ 4512 tp->selective_acks[0].end_seq = end_seq; 4513 return; 4514 } 4515 4516 /* Find place to insert this segment. */ 4517 while (1) { 4518 if (!after(TCP_SKB_CB(skb1)->seq, seq)) 4519 break; 4520 if (skb_queue_is_first(&tp->out_of_order_queue, skb1)) { 4521 skb1 = NULL; 4522 break; 4523 } 4524 skb1 = skb_queue_prev(&tp->out_of_order_queue, skb1); 4525 } 4526 4527 /* Do skb overlap to previous one? */ 4528 if (skb1 && before(seq, TCP_SKB_CB(skb1)->end_seq)) { 4529 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) { 4530 /* All the bits are present. Drop. */ 4531 __kfree_skb(skb); 4532 tcp_dsack_set(sk, seq, end_seq); 4533 goto add_sack; 4534 } 4535 if (after(seq, TCP_SKB_CB(skb1)->seq)) { 4536 /* Partial overlap. */ 4537 tcp_dsack_set(sk, seq, 4538 TCP_SKB_CB(skb1)->end_seq); 4539 } else { 4540 if (skb_queue_is_first(&tp->out_of_order_queue, 4541 skb1)) 4542 skb1 = NULL; 4543 else 4544 skb1 = skb_queue_prev( 4545 &tp->out_of_order_queue, 4546 skb1); 4547 } 4548 } 4549 if (!skb1) 4550 __skb_queue_head(&tp->out_of_order_queue, skb); 4551 else 4552 __skb_queue_after(&tp->out_of_order_queue, skb1, skb); 4553 4554 /* And clean segments covered by new one as whole. */ 4555 while (!skb_queue_is_last(&tp->out_of_order_queue, skb)) { 4556 skb1 = skb_queue_next(&tp->out_of_order_queue, skb); 4557 4558 if (!after(end_seq, TCP_SKB_CB(skb1)->seq)) 4559 break; 4560 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) { 4561 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq, 4562 end_seq); 4563 break; 4564 } 4565 __skb_unlink(skb1, &tp->out_of_order_queue); 4566 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq, 4567 TCP_SKB_CB(skb1)->end_seq); 4568 __kfree_skb(skb1); 4569 } 4570 4571add_sack: 4572 if (tcp_is_sack(tp)) 4573 tcp_sack_new_ofo_skb(sk, seq, end_seq); 4574 } 4575} 4576 4577static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb, 4578 struct sk_buff_head *list) 4579{ 4580 struct sk_buff *next = NULL; 4581 4582 if (!skb_queue_is_last(list, skb)) 4583 next = skb_queue_next(list, skb); 4584 4585 __skb_unlink(skb, list); 4586 __kfree_skb(skb); 4587 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED); 4588 4589 return next; 4590} 4591 4592/* Collapse contiguous sequence of skbs head..tail with 4593 * sequence numbers start..end. 4594 * 4595 * If tail is NULL, this means until the end of the list. 4596 * 4597 * Segments with FIN/SYN are not collapsed (only because this 4598 * simplifies code) 4599 */ 4600static void 4601tcp_collapse(struct sock *sk, struct sk_buff_head *list, 4602 struct sk_buff *head, struct sk_buff *tail, 4603 u32 start, u32 end) 4604{ 4605 struct sk_buff *skb, *n; 4606 bool end_of_skbs; 4607 4608 /* First, check that queue is collapsible and find 4609 * the point where collapsing can be useful. */ 4610 skb = head; 4611restart: 4612 end_of_skbs = true; 4613 skb_queue_walk_from_safe(list, skb, n) { 4614 if (skb == tail) 4615 break; 4616 /* No new bits? It is possible on ofo queue. */ 4617 if (!before(start, TCP_SKB_CB(skb)->end_seq)) { 4618 skb = tcp_collapse_one(sk, skb, list); 4619 if (!skb) 4620 break; 4621 goto restart; 4622 } 4623 4624 /* The first skb to collapse is: 4625 * - not SYN/FIN and 4626 * - bloated or contains data before "start" or 4627 * overlaps to the next one. 4628 */ 4629 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin && 4630 (tcp_win_from_space(skb->truesize) > skb->len || 4631 before(TCP_SKB_CB(skb)->seq, start))) { 4632 end_of_skbs = false; 4633 break; 4634 } 4635 4636 if (!skb_queue_is_last(list, skb)) { 4637 struct sk_buff *next = skb_queue_next(list, skb); 4638 if (next != tail && 4639 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(next)->seq) { 4640 end_of_skbs = false; 4641 break; 4642 } 4643 } 4644 4645 /* Decided to skip this, advance start seq. */ 4646 start = TCP_SKB_CB(skb)->end_seq; 4647 } 4648 if (end_of_skbs || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin) 4649 return; 4650 4651 while (before(start, end)) { 4652 struct sk_buff *nskb; 4653 unsigned int header = skb_headroom(skb); 4654 int copy = SKB_MAX_ORDER(header, 0); 4655 4656 /* Too big header? This can happen with IPv6. */ 4657 if (copy < 0) 4658 return; 4659 if (end - start < copy) 4660 copy = end - start; 4661 nskb = alloc_skb(copy + header, GFP_ATOMIC); 4662 if (!nskb) 4663 return; 4664 4665 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head); 4666 skb_set_network_header(nskb, (skb_network_header(skb) - 4667 skb->head)); 4668 skb_set_transport_header(nskb, (skb_transport_header(skb) - 4669 skb->head)); 4670 skb_reserve(nskb, header); 4671 memcpy(nskb->head, skb->head, header); 4672 memcpy(nskb->cb, skb->cb, sizeof(skb->cb)); 4673 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start; 4674 __skb_queue_before(list, skb, nskb); 4675 skb_set_owner_r(nskb, sk); 4676 4677 /* Copy data, releasing collapsed skbs. */ 4678 while (copy > 0) { 4679 int offset = start - TCP_SKB_CB(skb)->seq; 4680 int size = TCP_SKB_CB(skb)->end_seq - start; 4681 4682 BUG_ON(offset < 0); 4683 if (size > 0) { 4684 size = min(copy, size); 4685 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size)) 4686 BUG(); 4687 TCP_SKB_CB(nskb)->end_seq += size; 4688 copy -= size; 4689 start += size; 4690 } 4691 if (!before(start, TCP_SKB_CB(skb)->end_seq)) { 4692 skb = tcp_collapse_one(sk, skb, list); 4693 if (!skb || 4694 skb == tail || 4695 tcp_hdr(skb)->syn || 4696 tcp_hdr(skb)->fin) 4697 return; 4698 } 4699 } 4700 } 4701} 4702 4703/* Collapse ofo queue. Algorithm: select contiguous sequence of skbs 4704 * and tcp_collapse() them until all the queue is collapsed. 4705 */ 4706static void tcp_collapse_ofo_queue(struct sock *sk) 4707{ 4708 struct tcp_sock *tp = tcp_sk(sk); 4709 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue); 4710 struct sk_buff *head; 4711 u32 start, end; 4712 4713 if (skb == NULL) 4714 return; 4715 4716 start = TCP_SKB_CB(skb)->seq; 4717 end = TCP_SKB_CB(skb)->end_seq; 4718 head = skb; 4719 4720 for (;;) { 4721 struct sk_buff *next = NULL; 4722 4723 if (!skb_queue_is_last(&tp->out_of_order_queue, skb)) 4724 next = skb_queue_next(&tp->out_of_order_queue, skb); 4725 skb = next; 4726 4727 /* Segment is terminated when we see gap or when 4728 * we are at the end of all the queue. */ 4729 if (!skb || 4730 after(TCP_SKB_CB(skb)->seq, end) || 4731 before(TCP_SKB_CB(skb)->end_seq, start)) { 4732 tcp_collapse(sk, &tp->out_of_order_queue, 4733 head, skb, start, end); 4734 head = skb; 4735 if (!skb) 4736 break; 4737 /* Start new segment */ 4738 start = TCP_SKB_CB(skb)->seq; 4739 end = TCP_SKB_CB(skb)->end_seq; 4740 } else { 4741 if (before(TCP_SKB_CB(skb)->seq, start)) 4742 start = TCP_SKB_CB(skb)->seq; 4743 if (after(TCP_SKB_CB(skb)->end_seq, end)) 4744 end = TCP_SKB_CB(skb)->end_seq; 4745 } 4746 } 4747} 4748 4749/* 4750 * Purge the out-of-order queue. 4751 * Return true if queue was pruned. 4752 */ 4753static int tcp_prune_ofo_queue(struct sock *sk) 4754{ 4755 struct tcp_sock *tp = tcp_sk(sk); 4756 int res = 0; 4757 4758 if (!skb_queue_empty(&tp->out_of_order_queue)) { 4759 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED); 4760 __skb_queue_purge(&tp->out_of_order_queue); 4761 4762 /* Reset SACK state. A conforming SACK implementation will 4763 * do the same at a timeout based retransmit. When a connection 4764 * is in a sad state like this, we care only about integrity 4765 * of the connection not performance. 4766 */ 4767 if (tp->rx_opt.sack_ok) 4768 tcp_sack_reset(&tp->rx_opt); 4769 sk_mem_reclaim(sk); 4770 res = 1; 4771 } 4772 return res; 4773} 4774 4775/* Reduce allocated memory if we can, trying to get 4776 * the socket within its memory limits again. 4777 * 4778 * Return less than zero if we should start dropping frames 4779 * until the socket owning process reads some of the data 4780 * to stabilize the situation. 4781 */ 4782static int tcp_prune_queue(struct sock *sk) 4783{ 4784 struct tcp_sock *tp = tcp_sk(sk); 4785 4786 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq); 4787 4788 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED); 4789 4790 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf) 4791 tcp_clamp_window(sk); 4792 else if (tcp_memory_pressure) 4793 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss); 4794 4795 tcp_collapse_ofo_queue(sk); 4796 if (!skb_queue_empty(&sk->sk_receive_queue)) 4797 tcp_collapse(sk, &sk->sk_receive_queue, 4798 skb_peek(&sk->sk_receive_queue), 4799 NULL, 4800 tp->copied_seq, tp->rcv_nxt); 4801 sk_mem_reclaim(sk); 4802 4803 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) 4804 return 0; 4805 4806 /* Collapsing did not help, destructive actions follow. 4807 * This must not ever occur. */ 4808 4809 tcp_prune_ofo_queue(sk); 4810 4811 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) 4812 return 0; 4813 4814 /* If we are really being abused, tell the caller to silently 4815 * drop receive data on the floor. It will get retransmitted 4816 * and hopefully then we'll have sufficient space. 4817 */ 4818 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED); 4819 4820 /* Massive buffer overcommit. */ 4821 tp->pred_flags = 0; 4822 return -1; 4823} 4824 4825/* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. 4826 * As additional protections, we do not touch cwnd in retransmission phases, 4827 * and if application hit its sndbuf limit recently. 4828 */ 4829void tcp_cwnd_application_limited(struct sock *sk) 4830{ 4831 struct tcp_sock *tp = tcp_sk(sk); 4832 4833 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && 4834 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { 4835 /* Limited by application or receiver window. */ 4836 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); 4837 u32 win_used = max(tp->snd_cwnd_used, init_win); 4838 if (win_used < tp->snd_cwnd) { 4839 tp->snd_ssthresh = tcp_current_ssthresh(sk); 4840 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1; 4841 } 4842 tp->snd_cwnd_used = 0; 4843 } 4844 tp->snd_cwnd_stamp = tcp_time_stamp; 4845} 4846 4847static int tcp_should_expand_sndbuf(struct sock *sk) 4848{ 4849 struct tcp_sock *tp = tcp_sk(sk); 4850 4851 /* If the user specified a specific send buffer setting, do 4852 * not modify it. 4853 */ 4854 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK) 4855 return 0; 4856 4857 /* If we are under global TCP memory pressure, do not expand. */ 4858 if (tcp_memory_pressure) 4859 return 0; 4860 4861 /* If we are under soft global TCP memory pressure, do not expand. */ 4862 if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0]) 4863 return 0; 4864 4865 /* If we filled the congestion window, do not expand. */ 4866 if (tp->packets_out >= tp->snd_cwnd) 4867 return 0; 4868 4869 return 1; 4870} 4871 4872/* When incoming ACK allowed to free some skb from write_queue, 4873 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket 4874 * on the exit from tcp input handler. 4875 * 4876 * PROBLEM: sndbuf expansion does not work well with largesend. 4877 */ 4878static void tcp_new_space(struct sock *sk) 4879{ 4880 struct tcp_sock *tp = tcp_sk(sk); 4881 4882 if (tcp_should_expand_sndbuf(sk)) { 4883 int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) + 4884 MAX_TCP_HEADER + 16 + sizeof(struct sk_buff); 4885 int demanded = max_t(unsigned int, tp->snd_cwnd, 4886 tp->reordering + 1); 4887 sndmem *= 2 * demanded; 4888 if (sndmem > sk->sk_sndbuf) 4889 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]); 4890 tp->snd_cwnd_stamp = tcp_time_stamp; 4891 } 4892 4893 sk->sk_write_space(sk); 4894} 4895 4896static void tcp_check_space(struct sock *sk) 4897{ 4898 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) { 4899 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK); 4900 if (sk->sk_socket && 4901 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) 4902 tcp_new_space(sk); 4903 } 4904} 4905 4906static inline void tcp_data_snd_check(struct sock *sk) 4907{ 4908 tcp_push_pending_frames(sk); 4909 tcp_check_space(sk); 4910} 4911 4912/* 4913 * Check if sending an ack is needed. 4914 */ 4915static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible) 4916{ 4917 struct tcp_sock *tp = tcp_sk(sk); 4918 4919 /* More than one full frame received... */ 4920 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss && 4921 /* ... and right edge of window advances far enough. 4922 * (tcp_recvmsg() will send ACK otherwise). Or... 4923 */ 4924 __tcp_select_window(sk) >= tp->rcv_wnd) || 4925 /* We ACK each frame or... */ 4926 tcp_in_quickack_mode(sk) || 4927 /* We have out of order data. */ 4928 (ofo_possible && skb_peek(&tp->out_of_order_queue))) { 4929 /* Then ack it now */ 4930 tcp_send_ack(sk); 4931 } else { 4932 /* Else, send delayed ack. */ 4933 tcp_send_delayed_ack(sk); 4934 } 4935} 4936 4937static inline void tcp_ack_snd_check(struct sock *sk) 4938{ 4939 if (!inet_csk_ack_scheduled(sk)) { 4940 /* We sent a data segment already. */ 4941 return; 4942 } 4943 __tcp_ack_snd_check(sk, 1); 4944} 4945 4946/* 4947 * This routine is only called when we have urgent data 4948 * signaled. Its the 'slow' part of tcp_urg. It could be 4949 * moved inline now as tcp_urg is only called from one 4950 * place. We handle URGent data wrong. We have to - as 4951 * BSD still doesn't use the correction from RFC961. 4952 * For 1003.1g we should support a new option TCP_STDURG to permit 4953 * either form (or just set the sysctl tcp_stdurg). 4954 */ 4955 4956static void tcp_check_urg(struct sock *sk, struct tcphdr *th) 4957{ 4958 struct tcp_sock *tp = tcp_sk(sk); 4959 u32 ptr = ntohs(th->urg_ptr); 4960 4961 if (ptr && !sysctl_tcp_stdurg) 4962 ptr--; 4963 ptr += ntohl(th->seq); 4964 4965 /* Ignore urgent data that we've already seen and read. */ 4966 if (after(tp->copied_seq, ptr)) 4967 return; 4968 4969 /* Do not replay urg ptr. 4970 * 4971 * NOTE: interesting situation not covered by specs. 4972 * Misbehaving sender may send urg ptr, pointing to segment, 4973 * which we already have in ofo queue. We are not able to fetch 4974 * such data and will stay in TCP_URG_NOTYET until will be eaten 4975 * by recvmsg(). Seems, we are not obliged to handle such wicked 4976 * situations. But it is worth to think about possibility of some 4977 * DoSes using some hypothetical application level deadlock. 4978 */ 4979 if (before(ptr, tp->rcv_nxt)) 4980 return; 4981 4982 /* Do we already have a newer (or duplicate) urgent pointer? */ 4983 if (tp->urg_data && !after(ptr, tp->urg_seq)) 4984 return; 4985 4986 /* Tell the world about our new urgent pointer. */ 4987 sk_send_sigurg(sk); 4988 4989 /* We may be adding urgent data when the last byte read was 4990 * urgent. To do this requires some care. We cannot just ignore 4991 * tp->copied_seq since we would read the last urgent byte again 4992 * as data, nor can we alter copied_seq until this data arrives 4993 * or we break the semantics of SIOCATMARK (and thus sockatmark()) 4994 * 4995 * NOTE. Double Dutch. Rendering to plain English: author of comment 4996 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB); 4997 * and expect that both A and B disappear from stream. This is _wrong_. 4998 * Though this happens in BSD with high probability, this is occasional. 4999 * Any application relying on this is buggy. Note also, that fix "works" 5000 * only in this artificial test. Insert some normal data between A and B and we will 5001 * decline of BSD again. Verdict: it is better to remove to trap 5002 * buggy users. 5003 */ 5004 if (tp->urg_seq == tp->copied_seq && tp->urg_data && 5005 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) { 5006 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue); 5007 tp->copied_seq++; 5008 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) { 5009 __skb_unlink(skb, &sk->sk_receive_queue); 5010 __kfree_skb(skb); 5011 } 5012 } 5013 5014 tp->urg_data = TCP_URG_NOTYET; 5015 tp->urg_seq = ptr; 5016 5017 /* Disable header prediction. */ 5018 tp->pred_flags = 0; 5019} 5020 5021/* This is the 'fast' part of urgent handling. */ 5022static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th) 5023{ 5024 struct tcp_sock *tp = tcp_sk(sk); 5025 5026 /* Check if we get a new urgent pointer - normally not. */ 5027 if (th->urg) 5028 tcp_check_urg(sk, th); 5029 5030 /* Do we wait for any urgent data? - normally not... */ 5031 if (tp->urg_data == TCP_URG_NOTYET) { 5032 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) - 5033 th->syn; 5034 5035 /* Is the urgent pointer pointing into this packet? */ 5036 if (ptr < skb->len) { 5037 u8 tmp; 5038 if (skb_copy_bits(skb, ptr, &tmp, 1)) 5039 BUG(); 5040 tp->urg_data = TCP_URG_VALID | tmp; 5041 if (!sock_flag(sk, SOCK_DEAD)) 5042 sk->sk_data_ready(sk, 0); 5043 } 5044 } 5045} 5046 5047static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen) 5048{ 5049 struct tcp_sock *tp = tcp_sk(sk); 5050 int chunk = skb->len - hlen; 5051 int err; 5052 5053 local_bh_enable(); 5054 if (skb_csum_unnecessary(skb)) 5055 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk); 5056 else 5057 err = skb_copy_and_csum_datagram_iovec(skb, hlen, 5058 tp->ucopy.iov); 5059 5060 if (!err) { 5061 tp->ucopy.len -= chunk; 5062 tp->copied_seq += chunk; 5063 tcp_rcv_space_adjust(sk); 5064 } 5065 5066 local_bh_disable(); 5067 return err; 5068} 5069 5070static __sum16 __tcp_checksum_complete_user(struct sock *sk, 5071 struct sk_buff *skb) 5072{ 5073 __sum16 result; 5074 5075 if (sock_owned_by_user(sk)) { 5076 local_bh_enable(); 5077 result = __tcp_checksum_complete(skb); 5078 local_bh_disable(); 5079 } else { 5080 result = __tcp_checksum_complete(skb); 5081 } 5082 return result; 5083} 5084 5085static inline int tcp_checksum_complete_user(struct sock *sk, 5086 struct sk_buff *skb) 5087{ 5088 return !skb_csum_unnecessary(skb) && 5089 __tcp_checksum_complete_user(sk, skb); 5090} 5091 5092#ifdef CONFIG_NET_DMA 5093static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb, 5094 int hlen) 5095{ 5096 struct tcp_sock *tp = tcp_sk(sk); 5097 int chunk = skb->len - hlen; 5098 int dma_cookie; 5099 int copied_early = 0; 5100 5101 if (tp->ucopy.wakeup) 5102 return 0; 5103 5104 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list) 5105 tp->ucopy.dma_chan = dma_find_channel(DMA_MEMCPY); 5106 5107 if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) { 5108 5109 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan, 5110 skb, hlen, 5111 tp->ucopy.iov, chunk, 5112 tp->ucopy.pinned_list); 5113 5114 if (dma_cookie < 0) 5115 goto out; 5116 5117 tp->ucopy.dma_cookie = dma_cookie; 5118 copied_early = 1; 5119 5120 tp->ucopy.len -= chunk; 5121 tp->copied_seq += chunk; 5122 tcp_rcv_space_adjust(sk); 5123 5124 if ((tp->ucopy.len == 0) || 5125 (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) || 5126 (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) { 5127 tp->ucopy.wakeup = 1; 5128 sk->sk_data_ready(sk, 0); 5129 } 5130 } else if (chunk > 0) { 5131 tp->ucopy.wakeup = 1; 5132 sk->sk_data_ready(sk, 0); 5133 } 5134out: 5135 return copied_early; 5136} 5137#endif /* CONFIG_NET_DMA */ 5138 5139/* Does PAWS and seqno based validation of an incoming segment, flags will 5140 * play significant role here. 5141 */ 5142static int tcp_validate_incoming(struct sock *sk, struct sk_buff *skb, 5143 struct tcphdr *th, int syn_inerr) 5144{ 5145 u8 *hash_location; 5146 struct tcp_sock *tp = tcp_sk(sk); 5147 5148 /* RFC1323: H1. Apply PAWS check first. */ 5149 if (tcp_fast_parse_options(skb, th, tp, &hash_location) && 5150 tp->rx_opt.saw_tstamp && 5151 tcp_paws_discard(sk, skb)) { 5152 if (!th->rst) { 5153 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED); 5154 tcp_send_dupack(sk, skb); 5155 goto discard; 5156 } 5157 /* Reset is accepted even if it did not pass PAWS. */ 5158 } 5159 5160 /* Step 1: check sequence number */ 5161 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) { 5162 /* RFC793, page 37: "In all states except SYN-SENT, all reset 5163 * (RST) segments are validated by checking their SEQ-fields." 5164 * And page 69: "If an incoming segment is not acceptable, 5165 * an acknowledgment should be sent in reply (unless the RST 5166 * bit is set, if so drop the segment and return)". 5167 */ 5168 if (!th->rst) 5169 tcp_send_dupack(sk, skb); 5170 goto discard; 5171 } 5172 5173 /* Step 2: check RST bit */ 5174 if (th->rst) { 5175 tcp_reset(sk); 5176 goto discard; 5177 } 5178 5179 /* ts_recent update must be made after we are sure that the packet 5180 * is in window. 5181 */ 5182 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq); 5183 5184 /* step 3: check security and precedence [ignored] */ 5185 5186 /* step 4: Check for a SYN in window. */ 5187 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 5188 if (syn_inerr) 5189 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS); 5190 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONSYN); 5191 tcp_reset(sk); 5192 return -1; 5193 } 5194 5195 return 1; 5196 5197discard: 5198 __kfree_skb(skb); 5199 return 0; 5200} 5201 5202/* 5203 * TCP receive function for the ESTABLISHED state. 5204 * 5205 * It is split into a fast path and a slow path. The fast path is 5206 * disabled when: 5207 * - A zero window was announced from us - zero window probing 5208 * is only handled properly in the slow path. 5209 * - Out of order segments arrived. 5210 * - Urgent data is expected. 5211 * - There is no buffer space left 5212 * - Unexpected TCP flags/window values/header lengths are received 5213 * (detected by checking the TCP header against pred_flags) 5214 * - Data is sent in both directions. Fast path only supports pure senders 5215 * or pure receivers (this means either the sequence number or the ack 5216 * value must stay constant) 5217 * - Unexpected TCP option. 5218 * 5219 * When these conditions are not satisfied it drops into a standard 5220 * receive procedure patterned after RFC793 to handle all cases. 5221 * The first three cases are guaranteed by proper pred_flags setting, 5222 * the rest is checked inline. Fast processing is turned on in 5223 * tcp_data_queue when everything is OK. 5224 */ 5225int BCMFASTPATH_HOST tcp_rcv_established(struct sock *sk, struct sk_buff *skb, 5226 struct tcphdr *th, unsigned len) 5227{ 5228 struct tcp_sock *tp = tcp_sk(sk); 5229 int res; 5230 5231 /* 5232 * Header prediction. 5233 * The code loosely follows the one in the famous 5234 * "30 instruction TCP receive" Van Jacobson mail. 5235 * 5236 * Van's trick is to deposit buffers into socket queue 5237 * on a device interrupt, to call tcp_recv function 5238 * on the receive process context and checksum and copy 5239 * the buffer to user space. smart... 5240 * 5241 * Our current scheme is not silly either but we take the 5242 * extra cost of the net_bh soft interrupt processing... 5243 * We do checksum and copy also but from device to kernel. 5244 */ 5245 5246 tp->rx_opt.saw_tstamp = 0; 5247 5248 /* pred_flags is 0xS?10 << 16 + snd_wnd 5249 * if header_prediction is to be made 5250 * 'S' will always be tp->tcp_header_len >> 2 5251 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to 5252 * turn it off (when there are holes in the receive 5253 * space for instance) 5254 * PSH flag is ignored. 5255 */ 5256 5257 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags && 5258 TCP_SKB_CB(skb)->seq == tp->rcv_nxt && 5259 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) { 5260 int tcp_header_len = tp->tcp_header_len; 5261 5262 /* Timestamp header prediction: tcp_header_len 5263 * is automatically equal to th->doff*4 due to pred_flags 5264 * match. 5265 */ 5266 5267 /* Check timestamp */ 5268 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) { 5269 /* No? Slow path! */ 5270 if (!tcp_parse_aligned_timestamp(tp, th)) 5271 goto slow_path; 5272 5273 /* If PAWS failed, check it more carefully in slow path */ 5274 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0) 5275 goto slow_path; 5276 5277 /* DO NOT update ts_recent here, if checksum fails 5278 * and timestamp was corrupted part, it will result 5279 * in a hung connection since we will drop all 5280 * future packets due to the PAWS test. 5281 */ 5282 } 5283 5284 if (len <= tcp_header_len) { 5285 /* Bulk data transfer: sender */ 5286 if (len == tcp_header_len) { 5287 /* Predicted packet is in window by definition. 5288 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 5289 * Hence, check seq<=rcv_wup reduces to: 5290 */ 5291 if (tcp_header_len == 5292 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && 5293 tp->rcv_nxt == tp->rcv_wup) 5294 tcp_store_ts_recent(tp); 5295 5296 /* We know that such packets are checksummed 5297 * on entry. 5298 */ 5299 tcp_ack(sk, skb, 0); 5300 __kfree_skb(skb); 5301 tcp_data_snd_check(sk); 5302 return 0; 5303 } else { /* Header too small */ 5304 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS); 5305 goto discard; 5306 } 5307 } else { 5308 int eaten = 0; 5309 int copied_early = 0; 5310 5311 if (tp->copied_seq == tp->rcv_nxt && 5312 len - tcp_header_len <= tp->ucopy.len) { 5313#ifdef CONFIG_NET_DMA 5314 if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) { 5315 copied_early = 1; 5316 eaten = 1; 5317 } 5318#endif 5319 if (tp->ucopy.task == current && 5320 sock_owned_by_user(sk) && !copied_early) { 5321 __set_current_state(TASK_RUNNING); 5322 5323 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len)) 5324 eaten = 1; 5325 } 5326 if (eaten) { 5327 /* Predicted packet is in window by definition. 5328 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 5329 * Hence, check seq<=rcv_wup reduces to: 5330 */ 5331 if (tcp_header_len == 5332 (sizeof(struct tcphdr) + 5333 TCPOLEN_TSTAMP_ALIGNED) && 5334 tp->rcv_nxt == tp->rcv_wup) 5335 tcp_store_ts_recent(tp); 5336 5337 tcp_rcv_rtt_measure_ts(sk, skb); 5338 5339 __skb_pull(skb, tcp_header_len); 5340 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; 5341 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER); 5342 } 5343 if (copied_early) 5344 tcp_cleanup_rbuf(sk, skb->len); 5345 } 5346 if (!eaten) { 5347 if (tcp_checksum_complete_user(sk, skb)) 5348 goto csum_error; 5349 5350 /* Predicted packet is in window by definition. 5351 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 5352 * Hence, check seq<=rcv_wup reduces to: 5353 */ 5354 if (tcp_header_len == 5355 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && 5356 tp->rcv_nxt == tp->rcv_wup) 5357 tcp_store_ts_recent(tp); 5358 5359 tcp_rcv_rtt_measure_ts(sk, skb); 5360 5361 if ((int)skb->truesize > sk->sk_forward_alloc) 5362 goto step5; 5363 5364 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS); 5365 5366 /* Bulk data transfer: receiver */ 5367 __skb_pull(skb, tcp_header_len); 5368 __skb_queue_tail(&sk->sk_receive_queue, skb); 5369 skb_set_owner_r(skb, sk); 5370 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; 5371 } 5372 5373 tcp_event_data_recv(sk, skb); 5374 5375 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) { 5376 /* Well, only one small jumplet in fast path... */ 5377 tcp_ack(sk, skb, FLAG_DATA); 5378 tcp_data_snd_check(sk); 5379 if (!inet_csk_ack_scheduled(sk)) 5380 goto no_ack; 5381 } 5382 5383 if (!copied_early || tp->rcv_nxt != tp->rcv_wup) 5384 __tcp_ack_snd_check(sk, 0); 5385no_ack: 5386#ifdef CONFIG_NET_DMA 5387 if (copied_early) 5388 __skb_queue_tail(&sk->sk_async_wait_queue, skb); 5389 else 5390#endif 5391 if (eaten) 5392 __kfree_skb(skb); 5393 else 5394 sk->sk_data_ready(sk, 0); 5395 return 0; 5396 } 5397 } 5398 5399slow_path: 5400 if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb)) 5401 goto csum_error; 5402 5403 /* 5404 * Standard slow path. 5405 */ 5406 5407 res = tcp_validate_incoming(sk, skb, th, 1); 5408 if (res <= 0) 5409 return -res; 5410 5411step5: 5412 if (th->ack && tcp_ack(sk, skb, FLAG_SLOWPATH) < 0) 5413 goto discard; 5414 5415 tcp_rcv_rtt_measure_ts(sk, skb); 5416 5417 /* Process urgent data. */ 5418 tcp_urg(sk, skb, th); 5419 5420 /* step 7: process the segment text */ 5421 tcp_data_queue(sk, skb); 5422 5423 tcp_data_snd_check(sk); 5424 tcp_ack_snd_check(sk); 5425 return 0; 5426 5427csum_error: 5428 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS); 5429 5430discard: 5431 __kfree_skb(skb); 5432 return 0; 5433} 5434EXPORT_SYMBOL(tcp_rcv_established); 5435 5436static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb, 5437 struct tcphdr *th, unsigned len) 5438{ 5439 u8 *hash_location; 5440 struct inet_connection_sock *icsk = inet_csk(sk); 5441 struct tcp_sock *tp = tcp_sk(sk); 5442 struct tcp_cookie_values *cvp = tp->cookie_values; 5443 int saved_clamp = tp->rx_opt.mss_clamp; 5444 5445 tcp_parse_options(skb, &tp->rx_opt, &hash_location, 0); 5446 5447 if (th->ack) { 5448 /* rfc793: 5449 * "If the state is SYN-SENT then 5450 * first check the ACK bit 5451 * If the ACK bit is set 5452 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send 5453 * a reset (unless the RST bit is set, if so drop 5454 * the segment and return)" 5455 * 5456 * We do not send data with SYN, so that RFC-correct 5457 * test reduces to: 5458 */ 5459 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt) 5460 goto reset_and_undo; 5461 5462 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 5463 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp, 5464 tcp_time_stamp)) { 5465 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED); 5466 goto reset_and_undo; 5467 } 5468 5469 /* Now ACK is acceptable. 5470 * 5471 * "If the RST bit is set 5472 * If the ACK was acceptable then signal the user "error: 5473 * connection reset", drop the segment, enter CLOSED state, 5474 * delete TCB, and return." 5475 */ 5476 5477 if (th->rst) { 5478 tcp_reset(sk); 5479 goto discard; 5480 } 5481 5482 /* rfc793: 5483 * "fifth, if neither of the SYN or RST bits is set then 5484 * drop the segment and return." 5485 * 5486 * See note below! 5487 * --ANK(990513) 5488 */ 5489 if (!th->syn) 5490 goto discard_and_undo; 5491 5492 /* rfc793: 5493 * "If the SYN bit is on ... 5494 * are acceptable then ... 5495 * (our SYN has been ACKed), change the connection 5496 * state to ESTABLISHED..." 5497 */ 5498 5499 TCP_ECN_rcv_synack(tp, th); 5500 5501 tp->snd_wl1 = TCP_SKB_CB(skb)->seq; 5502 tcp_ack(sk, skb, FLAG_SLOWPATH); 5503 5504 /* Ok.. it's good. Set up sequence numbers and 5505 * move to established. 5506 */ 5507 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; 5508 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; 5509 5510 /* RFC1323: The window in SYN & SYN/ACK segments is 5511 * never scaled. 5512 */ 5513 tp->snd_wnd = ntohs(th->window); 5514 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq); 5515 5516 if (!tp->rx_opt.wscale_ok) { 5517 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0; 5518 tp->window_clamp = min(tp->window_clamp, 65535U); 5519 } 5520 5521 if (tp->rx_opt.saw_tstamp) { 5522 tp->rx_opt.tstamp_ok = 1; 5523 tp->tcp_header_len = 5524 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 5525 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; 5526 tcp_store_ts_recent(tp); 5527 } else { 5528 tp->tcp_header_len = sizeof(struct tcphdr); 5529 } 5530 5531 if (tcp_is_sack(tp) && sysctl_tcp_fack) 5532 tcp_enable_fack(tp); 5533 5534 tcp_mtup_init(sk); 5535 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 5536 tcp_initialize_rcv_mss(sk); 5537 5538 /* Remember, tcp_poll() does not lock socket! 5539 * Change state from SYN-SENT only after copied_seq 5540 * is initialized. */ 5541 tp->copied_seq = tp->rcv_nxt; 5542 5543 if (cvp != NULL && 5544 cvp->cookie_pair_size > 0 && 5545 tp->rx_opt.cookie_plus > 0) { 5546 int cookie_size = tp->rx_opt.cookie_plus 5547 - TCPOLEN_COOKIE_BASE; 5548 int cookie_pair_size = cookie_size 5549 + cvp->cookie_desired; 5550 5551 /* A cookie extension option was sent and returned. 5552 * Note that each incoming SYNACK replaces the 5553 * Responder cookie. The initial exchange is most 5554 * fragile, as protection against spoofing relies 5555 * entirely upon the sequence and timestamp (above). 5556 * This replacement strategy allows the correct pair to 5557 * pass through, while any others will be filtered via 5558 * Responder verification later. 5559 */ 5560 if (sizeof(cvp->cookie_pair) >= cookie_pair_size) { 5561 memcpy(&cvp->cookie_pair[cvp->cookie_desired], 5562 hash_location, cookie_size); 5563 cvp->cookie_pair_size = cookie_pair_size; 5564 } 5565 } 5566 5567 smp_mb(); 5568 tcp_set_state(sk, TCP_ESTABLISHED); 5569 5570 security_inet_conn_established(sk, skb); 5571 5572 /* Make sure socket is routed, for correct metrics. */ 5573 icsk->icsk_af_ops->rebuild_header(sk); 5574 5575 tcp_init_metrics(sk); 5576 5577 tcp_init_congestion_control(sk); 5578 5579 /* Prevent spurious tcp_cwnd_restart() on first data 5580 * packet. 5581 */ 5582 tp->lsndtime = tcp_time_stamp; 5583 5584 tcp_init_buffer_space(sk); 5585 5586 if (sock_flag(sk, SOCK_KEEPOPEN)) 5587 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp)); 5588 5589 if (!tp->rx_opt.snd_wscale) 5590 __tcp_fast_path_on(tp, tp->snd_wnd); 5591 else 5592 tp->pred_flags = 0; 5593 5594 if (!sock_flag(sk, SOCK_DEAD)) { 5595 sk->sk_state_change(sk); 5596 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT); 5597 } 5598 5599 if (sk->sk_write_pending || 5600 icsk->icsk_accept_queue.rskq_defer_accept || 5601 icsk->icsk_ack.pingpong) { 5602 /* Save one ACK. Data will be ready after 5603 * several ticks, if write_pending is set. 5604 * 5605 * It may be deleted, but with this feature tcpdumps 5606 * look so _wonderfully_ clever, that I was not able 5607 * to stand against the temptation 8) --ANK 5608 */ 5609 inet_csk_schedule_ack(sk); 5610 icsk->icsk_ack.lrcvtime = tcp_time_stamp; 5611 icsk->icsk_ack.ato = TCP_ATO_MIN; 5612 tcp_incr_quickack(sk); 5613 tcp_enter_quickack_mode(sk); 5614 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 5615 TCP_DELACK_MAX, TCP_RTO_MAX); 5616 5617discard: 5618 __kfree_skb(skb); 5619 return 0; 5620 } else { 5621 tcp_send_ack(sk); 5622 } 5623 return -1; 5624 } 5625 5626 /* No ACK in the segment */ 5627 5628 if (th->rst) { 5629 /* rfc793: 5630 * "If the RST bit is set 5631 * 5632 * Otherwise (no ACK) drop the segment and return." 5633 */ 5634 5635 goto discard_and_undo; 5636 } 5637 5638 /* PAWS check. */ 5639 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && 5640 tcp_paws_reject(&tp->rx_opt, 0)) 5641 goto discard_and_undo; 5642 5643 if (th->syn) { 5644 /* We see SYN without ACK. It is attempt of 5645 * simultaneous connect with crossed SYNs. 5646 * Particularly, it can be connect to self. 5647 */ 5648 tcp_set_state(sk, TCP_SYN_RECV); 5649 5650 if (tp->rx_opt.saw_tstamp) { 5651 tp->rx_opt.tstamp_ok = 1; 5652 tcp_store_ts_recent(tp); 5653 tp->tcp_header_len = 5654 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 5655 } else { 5656 tp->tcp_header_len = sizeof(struct tcphdr); 5657 } 5658 5659 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; 5660 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; 5661 5662 /* RFC1323: The window in SYN & SYN/ACK segments is 5663 * never scaled. 5664 */ 5665 tp->snd_wnd = ntohs(th->window); 5666 tp->snd_wl1 = TCP_SKB_CB(skb)->seq; 5667 tp->max_window = tp->snd_wnd; 5668 5669 TCP_ECN_rcv_syn(tp, th); 5670 5671 tcp_mtup_init(sk); 5672 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 5673 tcp_initialize_rcv_mss(sk); 5674 5675 tcp_send_synack(sk); 5676 goto discard; 5677 } 5678 /* "fifth, if neither of the SYN or RST bits is set then 5679 * drop the segment and return." 5680 */ 5681 5682discard_and_undo: 5683 tcp_clear_options(&tp->rx_opt); 5684 tp->rx_opt.mss_clamp = saved_clamp; 5685 goto discard; 5686 5687reset_and_undo: 5688 tcp_clear_options(&tp->rx_opt); 5689 tp->rx_opt.mss_clamp = saved_clamp; 5690 return 1; 5691} 5692 5693/* 5694 * This function implements the receiving procedure of RFC 793 for 5695 * all states except ESTABLISHED and TIME_WAIT. 5696 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be 5697 * address independent. 5698 */ 5699 5700int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb, 5701 struct tcphdr *th, unsigned len) 5702{ 5703 struct tcp_sock *tp = tcp_sk(sk); 5704 struct inet_connection_sock *icsk = inet_csk(sk); 5705 int queued = 0; 5706 int res; 5707 5708 tp->rx_opt.saw_tstamp = 0; 5709 5710 switch (sk->sk_state) { 5711 case TCP_CLOSE: 5712 goto discard; 5713 5714 case TCP_LISTEN: 5715 if (th->ack) 5716 return 1; 5717 5718 if (th->rst) 5719 goto discard; 5720 5721 if (th->syn) { 5722 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0) 5723 return 1; 5724 5725 /* Now we have several options: In theory there is 5726 * nothing else in the frame. KA9Q has an option to 5727 * send data with the syn, BSD accepts data with the 5728 * syn up to the [to be] advertised window and 5729 * Solaris 2.1 gives you a protocol error. For now 5730 * we just ignore it, that fits the spec precisely 5731 * and avoids incompatibilities. It would be nice in 5732 * future to drop through and process the data. 5733 * 5734 * Now that TTCP is starting to be used we ought to 5735 * queue this data. 5736 * But, this leaves one open to an easy denial of 5737 * service attack, and SYN cookies can't defend 5738 * against this problem. So, we drop the data 5739 * in the interest of security over speed unless 5740 * it's still in use. 5741 */ 5742 kfree_skb(skb); 5743 return 0; 5744 } 5745 goto discard; 5746 5747 case TCP_SYN_SENT: 5748 queued = tcp_rcv_synsent_state_process(sk, skb, th, len); 5749 if (queued >= 0) 5750 return queued; 5751 5752 /* Do step6 onward by hand. */ 5753 tcp_urg(sk, skb, th); 5754 __kfree_skb(skb); 5755 tcp_data_snd_check(sk); 5756 return 0; 5757 } 5758 5759 res = tcp_validate_incoming(sk, skb, th, 0); 5760 if (res <= 0) 5761 return -res; 5762 5763 /* step 5: check the ACK field */ 5764 if (th->ack) { 5765 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH) > 0; 5766 5767 switch (sk->sk_state) { 5768 case TCP_SYN_RECV: 5769 if (acceptable) { 5770 tp->copied_seq = tp->rcv_nxt; 5771 smp_mb(); 5772 tcp_set_state(sk, TCP_ESTABLISHED); 5773 sk->sk_state_change(sk); 5774 5775 /* Note, that this wakeup is only for marginal 5776 * crossed SYN case. Passively open sockets 5777 * are not waked up, because sk->sk_sleep == 5778 * NULL and sk->sk_socket == NULL. 5779 */ 5780 if (sk->sk_socket) 5781 sk_wake_async(sk, 5782 SOCK_WAKE_IO, POLL_OUT); 5783 5784 tp->snd_una = TCP_SKB_CB(skb)->ack_seq; 5785 tp->snd_wnd = ntohs(th->window) << 5786 tp->rx_opt.snd_wscale; 5787 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq); 5788 5789 /* tcp_ack considers this ACK as duplicate 5790 * and does not calculate rtt. 5791 * Force it here. 5792 */ 5793 tcp_ack_update_rtt(sk, 0, 0); 5794 5795 if (tp->rx_opt.tstamp_ok) 5796 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; 5797 5798 /* Make sure socket is routed, for 5799 * correct metrics. 5800 */ 5801 icsk->icsk_af_ops->rebuild_header(sk); 5802 5803 tcp_init_metrics(sk); 5804 5805 tcp_init_congestion_control(sk); 5806 5807 /* Prevent spurious tcp_cwnd_restart() on 5808 * first data packet. 5809 */ 5810 tp->lsndtime = tcp_time_stamp; 5811 5812 tcp_mtup_init(sk); 5813 tcp_initialize_rcv_mss(sk); 5814 tcp_init_buffer_space(sk); 5815 tcp_fast_path_on(tp); 5816 } else { 5817 return 1; 5818 } 5819 break; 5820 5821 case TCP_FIN_WAIT1: 5822 if (tp->snd_una == tp->write_seq) { 5823 tcp_set_state(sk, TCP_FIN_WAIT2); 5824 sk->sk_shutdown |= SEND_SHUTDOWN; 5825 dst_confirm(__sk_dst_get(sk)); 5826 5827 if (!sock_flag(sk, SOCK_DEAD)) 5828 /* Wake up lingering close() */ 5829 sk->sk_state_change(sk); 5830 else { 5831 int tmo; 5832 5833 if (tp->linger2 < 0 || 5834 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 5835 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) { 5836 tcp_done(sk); 5837 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA); 5838 return 1; 5839 } 5840 5841 tmo = tcp_fin_time(sk); 5842 if (tmo > TCP_TIMEWAIT_LEN) { 5843 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN); 5844 } else if (th->fin || sock_owned_by_user(sk)) { 5845 /* Bad case. We could lose such FIN otherwise. 5846 * It is not a big problem, but it looks confusing 5847 * and not so rare event. We still can lose it now, 5848 * if it spins in bh_lock_sock(), but it is really 5849 * marginal case. 5850 */ 5851 inet_csk_reset_keepalive_timer(sk, tmo); 5852 } else { 5853 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo); 5854 goto discard; 5855 } 5856 } 5857 } 5858 break; 5859 5860 case TCP_CLOSING: 5861 if (tp->snd_una == tp->write_seq) { 5862 tcp_time_wait(sk, TCP_TIME_WAIT, 0); 5863 goto discard; 5864 } 5865 break; 5866 5867 case TCP_LAST_ACK: 5868 if (tp->snd_una == tp->write_seq) { 5869 tcp_update_metrics(sk); 5870 tcp_done(sk); 5871 goto discard; 5872 } 5873 break; 5874 } 5875 } else 5876 goto discard; 5877 5878 /* step 6: check the URG bit */ 5879 tcp_urg(sk, skb, th); 5880 5881 /* step 7: process the segment text */ 5882 switch (sk->sk_state) { 5883 case TCP_CLOSE_WAIT: 5884 case TCP_CLOSING: 5885 case TCP_LAST_ACK: 5886 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) 5887 break; 5888 case TCP_FIN_WAIT1: 5889 case TCP_FIN_WAIT2: 5890 /* RFC 793 says to queue data in these states, 5891 * RFC 1122 says we MUST send a reset. 5892 * BSD 4.4 also does reset. 5893 */ 5894 if (sk->sk_shutdown & RCV_SHUTDOWN) { 5895 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 5896 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) { 5897 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA); 5898 tcp_reset(sk); 5899 return 1; 5900 } 5901 } 5902 /* Fall through */ 5903 case TCP_ESTABLISHED: 5904 tcp_data_queue(sk, skb); 5905 queued = 1; 5906 break; 5907 } 5908 5909 /* tcp_data could move socket to TIME-WAIT */ 5910 if (sk->sk_state != TCP_CLOSE) { 5911 tcp_data_snd_check(sk); 5912 tcp_ack_snd_check(sk); 5913 } 5914 5915 if (!queued) { 5916discard: 5917 __kfree_skb(skb); 5918 } 5919 return 0; 5920} 5921EXPORT_SYMBOL(tcp_rcv_state_process); 5922