tcp_input.c revision 8dd71c5d28cd88d4400e7f474986e799e39aff37
1/* 2 * INET An implementation of the TCP/IP protocol suite for the LINUX 3 * operating system. INET is implemented using the BSD Socket 4 * interface as the means of communication with the user level. 5 * 6 * Implementation of the Transmission Control Protocol(TCP). 7 * 8 * Version: $Id: tcp_input.c,v 1.243 2002/02/01 22:01:04 davem Exp $ 9 * 10 * Authors: Ross Biro 11 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> 12 * Mark Evans, <evansmp@uhura.aston.ac.uk> 13 * Corey Minyard <wf-rch!minyard@relay.EU.net> 14 * Florian La Roche, <flla@stud.uni-sb.de> 15 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu> 16 * Linus Torvalds, <torvalds@cs.helsinki.fi> 17 * Alan Cox, <gw4pts@gw4pts.ampr.org> 18 * Matthew Dillon, <dillon@apollo.west.oic.com> 19 * Arnt Gulbrandsen, <agulbra@nvg.unit.no> 20 * Jorge Cwik, <jorge@laser.satlink.net> 21 */ 22 23/* 24 * Changes: 25 * Pedro Roque : Fast Retransmit/Recovery. 26 * Two receive queues. 27 * Retransmit queue handled by TCP. 28 * Better retransmit timer handling. 29 * New congestion avoidance. 30 * Header prediction. 31 * Variable renaming. 32 * 33 * Eric : Fast Retransmit. 34 * Randy Scott : MSS option defines. 35 * Eric Schenk : Fixes to slow start algorithm. 36 * Eric Schenk : Yet another double ACK bug. 37 * Eric Schenk : Delayed ACK bug fixes. 38 * Eric Schenk : Floyd style fast retrans war avoidance. 39 * David S. Miller : Don't allow zero congestion window. 40 * Eric Schenk : Fix retransmitter so that it sends 41 * next packet on ack of previous packet. 42 * Andi Kleen : Moved open_request checking here 43 * and process RSTs for open_requests. 44 * Andi Kleen : Better prune_queue, and other fixes. 45 * Andrey Savochkin: Fix RTT measurements in the presence of 46 * timestamps. 47 * Andrey Savochkin: Check sequence numbers correctly when 48 * removing SACKs due to in sequence incoming 49 * data segments. 50 * Andi Kleen: Make sure we never ack data there is not 51 * enough room for. Also make this condition 52 * a fatal error if it might still happen. 53 * Andi Kleen: Add tcp_measure_rcv_mss to make 54 * connections with MSS<min(MTU,ann. MSS) 55 * work without delayed acks. 56 * Andi Kleen: Process packets with PSH set in the 57 * fast path. 58 * J Hadi Salim: ECN support 59 * Andrei Gurtov, 60 * Pasi Sarolahti, 61 * Panu Kuhlberg: Experimental audit of TCP (re)transmission 62 * engine. Lots of bugs are found. 63 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs 64 */ 65 66#include <linux/mm.h> 67#include <linux/module.h> 68#include <linux/sysctl.h> 69#include <net/tcp.h> 70#include <net/inet_common.h> 71#include <linux/ipsec.h> 72#include <asm/unaligned.h> 73#include <net/netdma.h> 74 75int sysctl_tcp_timestamps __read_mostly = 1; 76int sysctl_tcp_window_scaling __read_mostly = 1; 77int sysctl_tcp_sack __read_mostly = 1; 78int sysctl_tcp_fack __read_mostly = 1; 79int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH; 80int sysctl_tcp_ecn __read_mostly; 81int sysctl_tcp_dsack __read_mostly = 1; 82int sysctl_tcp_app_win __read_mostly = 31; 83int sysctl_tcp_adv_win_scale __read_mostly = 2; 84 85int sysctl_tcp_stdurg __read_mostly; 86int sysctl_tcp_rfc1337 __read_mostly; 87int sysctl_tcp_max_orphans __read_mostly = NR_FILE; 88int sysctl_tcp_frto __read_mostly = 2; 89int sysctl_tcp_frto_response __read_mostly; 90int sysctl_tcp_nometrics_save __read_mostly; 91 92int sysctl_tcp_moderate_rcvbuf __read_mostly = 1; 93int sysctl_tcp_abc __read_mostly; 94 95#define FLAG_DATA 0x01 /* Incoming frame contained data. */ 96#define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */ 97#define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */ 98#define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */ 99#define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */ 100#define FLAG_DATA_SACKED 0x20 /* New SACK. */ 101#define FLAG_ECE 0x40 /* ECE in this ACK */ 102#define FLAG_DATA_LOST 0x80 /* SACK detected data lossage. */ 103#define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/ 104#define FLAG_ONLY_ORIG_SACKED 0x200 /* SACKs only non-rexmit sent before RTO */ 105#define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */ 106#define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */ 107#define FLAG_NONHEAD_RETRANS_ACKED 0x1000 /* Non-head rexmitted data was ACKed */ 108 109#define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED) 110#define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED) 111#define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE) 112#define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED) 113#define FLAG_ANY_PROGRESS (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED) 114 115#define IsSackFrto() (sysctl_tcp_frto == 0x2) 116 117#define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH) 118#define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH)) 119 120/* Adapt the MSS value used to make delayed ack decision to the 121 * real world. 122 */ 123static void tcp_measure_rcv_mss(struct sock *sk, 124 const struct sk_buff *skb) 125{ 126 struct inet_connection_sock *icsk = inet_csk(sk); 127 const unsigned int lss = icsk->icsk_ack.last_seg_size; 128 unsigned int len; 129 130 icsk->icsk_ack.last_seg_size = 0; 131 132 /* skb->len may jitter because of SACKs, even if peer 133 * sends good full-sized frames. 134 */ 135 len = skb_shinfo(skb)->gso_size ?: skb->len; 136 if (len >= icsk->icsk_ack.rcv_mss) { 137 icsk->icsk_ack.rcv_mss = len; 138 } else { 139 /* Otherwise, we make more careful check taking into account, 140 * that SACKs block is variable. 141 * 142 * "len" is invariant segment length, including TCP header. 143 */ 144 len += skb->data - skb_transport_header(skb); 145 if (len >= TCP_MIN_RCVMSS + sizeof(struct tcphdr) || 146 /* If PSH is not set, packet should be 147 * full sized, provided peer TCP is not badly broken. 148 * This observation (if it is correct 8)) allows 149 * to handle super-low mtu links fairly. 150 */ 151 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) && 152 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) { 153 /* Subtract also invariant (if peer is RFC compliant), 154 * tcp header plus fixed timestamp option length. 155 * Resulting "len" is MSS free of SACK jitter. 156 */ 157 len -= tcp_sk(sk)->tcp_header_len; 158 icsk->icsk_ack.last_seg_size = len; 159 if (len == lss) { 160 icsk->icsk_ack.rcv_mss = len; 161 return; 162 } 163 } 164 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED) 165 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2; 166 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED; 167 } 168} 169 170static void tcp_incr_quickack(struct sock *sk) 171{ 172 struct inet_connection_sock *icsk = inet_csk(sk); 173 unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss); 174 175 if (quickacks==0) 176 quickacks=2; 177 if (quickacks > icsk->icsk_ack.quick) 178 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS); 179} 180 181void tcp_enter_quickack_mode(struct sock *sk) 182{ 183 struct inet_connection_sock *icsk = inet_csk(sk); 184 tcp_incr_quickack(sk); 185 icsk->icsk_ack.pingpong = 0; 186 icsk->icsk_ack.ato = TCP_ATO_MIN; 187} 188 189/* Send ACKs quickly, if "quick" count is not exhausted 190 * and the session is not interactive. 191 */ 192 193static inline int tcp_in_quickack_mode(const struct sock *sk) 194{ 195 const struct inet_connection_sock *icsk = inet_csk(sk); 196 return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong; 197} 198 199static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp) 200{ 201 if (tp->ecn_flags&TCP_ECN_OK) 202 tp->ecn_flags |= TCP_ECN_QUEUE_CWR; 203} 204 205static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, struct sk_buff *skb) 206{ 207 if (tcp_hdr(skb)->cwr) 208 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR; 209} 210 211static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp) 212{ 213 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR; 214} 215 216static inline void TCP_ECN_check_ce(struct tcp_sock *tp, struct sk_buff *skb) 217{ 218 if (tp->ecn_flags&TCP_ECN_OK) { 219 if (INET_ECN_is_ce(TCP_SKB_CB(skb)->flags)) 220 tp->ecn_flags |= TCP_ECN_DEMAND_CWR; 221 /* Funny extension: if ECT is not set on a segment, 222 * it is surely retransmit. It is not in ECN RFC, 223 * but Linux follows this rule. */ 224 else if (INET_ECN_is_not_ect((TCP_SKB_CB(skb)->flags))) 225 tcp_enter_quickack_mode((struct sock *)tp); 226 } 227} 228 229static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, struct tcphdr *th) 230{ 231 if ((tp->ecn_flags&TCP_ECN_OK) && (!th->ece || th->cwr)) 232 tp->ecn_flags &= ~TCP_ECN_OK; 233} 234 235static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, struct tcphdr *th) 236{ 237 if ((tp->ecn_flags&TCP_ECN_OK) && (!th->ece || !th->cwr)) 238 tp->ecn_flags &= ~TCP_ECN_OK; 239} 240 241static inline int TCP_ECN_rcv_ecn_echo(struct tcp_sock *tp, struct tcphdr *th) 242{ 243 if (th->ece && !th->syn && (tp->ecn_flags&TCP_ECN_OK)) 244 return 1; 245 return 0; 246} 247 248/* Buffer size and advertised window tuning. 249 * 250 * 1. Tuning sk->sk_sndbuf, when connection enters established state. 251 */ 252 253static void tcp_fixup_sndbuf(struct sock *sk) 254{ 255 int sndmem = tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER + 16 + 256 sizeof(struct sk_buff); 257 258 if (sk->sk_sndbuf < 3 * sndmem) 259 sk->sk_sndbuf = min(3 * sndmem, sysctl_tcp_wmem[2]); 260} 261 262/* 2. Tuning advertised window (window_clamp, rcv_ssthresh) 263 * 264 * All tcp_full_space() is split to two parts: "network" buffer, allocated 265 * forward and advertised in receiver window (tp->rcv_wnd) and 266 * "application buffer", required to isolate scheduling/application 267 * latencies from network. 268 * window_clamp is maximal advertised window. It can be less than 269 * tcp_full_space(), in this case tcp_full_space() - window_clamp 270 * is reserved for "application" buffer. The less window_clamp is 271 * the smoother our behaviour from viewpoint of network, but the lower 272 * throughput and the higher sensitivity of the connection to losses. 8) 273 * 274 * rcv_ssthresh is more strict window_clamp used at "slow start" 275 * phase to predict further behaviour of this connection. 276 * It is used for two goals: 277 * - to enforce header prediction at sender, even when application 278 * requires some significant "application buffer". It is check #1. 279 * - to prevent pruning of receive queue because of misprediction 280 * of receiver window. Check #2. 281 * 282 * The scheme does not work when sender sends good segments opening 283 * window and then starts to feed us spaghetti. But it should work 284 * in common situations. Otherwise, we have to rely on queue collapsing. 285 */ 286 287/* Slow part of check#2. */ 288static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb) 289{ 290 struct tcp_sock *tp = tcp_sk(sk); 291 /* Optimize this! */ 292 int truesize = tcp_win_from_space(skb->truesize)/2; 293 int window = tcp_win_from_space(sysctl_tcp_rmem[2])/2; 294 295 while (tp->rcv_ssthresh <= window) { 296 if (truesize <= skb->len) 297 return 2 * inet_csk(sk)->icsk_ack.rcv_mss; 298 299 truesize >>= 1; 300 window >>= 1; 301 } 302 return 0; 303} 304 305static void tcp_grow_window(struct sock *sk, 306 struct sk_buff *skb) 307{ 308 struct tcp_sock *tp = tcp_sk(sk); 309 310 /* Check #1 */ 311 if (tp->rcv_ssthresh < tp->window_clamp && 312 (int)tp->rcv_ssthresh < tcp_space(sk) && 313 !tcp_memory_pressure) { 314 int incr; 315 316 /* Check #2. Increase window, if skb with such overhead 317 * will fit to rcvbuf in future. 318 */ 319 if (tcp_win_from_space(skb->truesize) <= skb->len) 320 incr = 2*tp->advmss; 321 else 322 incr = __tcp_grow_window(sk, skb); 323 324 if (incr) { 325 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr, tp->window_clamp); 326 inet_csk(sk)->icsk_ack.quick |= 1; 327 } 328 } 329} 330 331/* 3. Tuning rcvbuf, when connection enters established state. */ 332 333static void tcp_fixup_rcvbuf(struct sock *sk) 334{ 335 struct tcp_sock *tp = tcp_sk(sk); 336 int rcvmem = tp->advmss + MAX_TCP_HEADER + 16 + sizeof(struct sk_buff); 337 338 /* Try to select rcvbuf so that 4 mss-sized segments 339 * will fit to window and corresponding skbs will fit to our rcvbuf. 340 * (was 3; 4 is minimum to allow fast retransmit to work.) 341 */ 342 while (tcp_win_from_space(rcvmem) < tp->advmss) 343 rcvmem += 128; 344 if (sk->sk_rcvbuf < 4 * rcvmem) 345 sk->sk_rcvbuf = min(4 * rcvmem, sysctl_tcp_rmem[2]); 346} 347 348/* 4. Try to fixup all. It is made immediately after connection enters 349 * established state. 350 */ 351static void tcp_init_buffer_space(struct sock *sk) 352{ 353 struct tcp_sock *tp = tcp_sk(sk); 354 int maxwin; 355 356 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) 357 tcp_fixup_rcvbuf(sk); 358 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK)) 359 tcp_fixup_sndbuf(sk); 360 361 tp->rcvq_space.space = tp->rcv_wnd; 362 363 maxwin = tcp_full_space(sk); 364 365 if (tp->window_clamp >= maxwin) { 366 tp->window_clamp = maxwin; 367 368 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss) 369 tp->window_clamp = max(maxwin - 370 (maxwin >> sysctl_tcp_app_win), 371 4 * tp->advmss); 372 } 373 374 /* Force reservation of one segment. */ 375 if (sysctl_tcp_app_win && 376 tp->window_clamp > 2 * tp->advmss && 377 tp->window_clamp + tp->advmss > maxwin) 378 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss); 379 380 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp); 381 tp->snd_cwnd_stamp = tcp_time_stamp; 382} 383 384/* 5. Recalculate window clamp after socket hit its memory bounds. */ 385static void tcp_clamp_window(struct sock *sk) 386{ 387 struct tcp_sock *tp = tcp_sk(sk); 388 struct inet_connection_sock *icsk = inet_csk(sk); 389 390 icsk->icsk_ack.quick = 0; 391 392 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] && 393 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) && 394 !tcp_memory_pressure && 395 atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) { 396 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc), 397 sysctl_tcp_rmem[2]); 398 } 399 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf) 400 tp->rcv_ssthresh = min(tp->window_clamp, 2U*tp->advmss); 401} 402 403 404/* Initialize RCV_MSS value. 405 * RCV_MSS is an our guess about MSS used by the peer. 406 * We haven't any direct information about the MSS. 407 * It's better to underestimate the RCV_MSS rather than overestimate. 408 * Overestimations make us ACKing less frequently than needed. 409 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss(). 410 */ 411void tcp_initialize_rcv_mss(struct sock *sk) 412{ 413 struct tcp_sock *tp = tcp_sk(sk); 414 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache); 415 416 hint = min(hint, tp->rcv_wnd/2); 417 hint = min(hint, TCP_MIN_RCVMSS); 418 hint = max(hint, TCP_MIN_MSS); 419 420 inet_csk(sk)->icsk_ack.rcv_mss = hint; 421} 422 423/* Receiver "autotuning" code. 424 * 425 * The algorithm for RTT estimation w/o timestamps is based on 426 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL. 427 * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps> 428 * 429 * More detail on this code can be found at 430 * <http://www.psc.edu/~jheffner/senior_thesis.ps>, 431 * though this reference is out of date. A new paper 432 * is pending. 433 */ 434static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep) 435{ 436 u32 new_sample = tp->rcv_rtt_est.rtt; 437 long m = sample; 438 439 if (m == 0) 440 m = 1; 441 442 if (new_sample != 0) { 443 /* If we sample in larger samples in the non-timestamp 444 * case, we could grossly overestimate the RTT especially 445 * with chatty applications or bulk transfer apps which 446 * are stalled on filesystem I/O. 447 * 448 * Also, since we are only going for a minimum in the 449 * non-timestamp case, we do not smooth things out 450 * else with timestamps disabled convergence takes too 451 * long. 452 */ 453 if (!win_dep) { 454 m -= (new_sample >> 3); 455 new_sample += m; 456 } else if (m < new_sample) 457 new_sample = m << 3; 458 } else { 459 /* No previous measure. */ 460 new_sample = m << 3; 461 } 462 463 if (tp->rcv_rtt_est.rtt != new_sample) 464 tp->rcv_rtt_est.rtt = new_sample; 465} 466 467static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp) 468{ 469 if (tp->rcv_rtt_est.time == 0) 470 goto new_measure; 471 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq)) 472 return; 473 tcp_rcv_rtt_update(tp, 474 jiffies - tp->rcv_rtt_est.time, 475 1); 476 477new_measure: 478 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd; 479 tp->rcv_rtt_est.time = tcp_time_stamp; 480} 481 482static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, const struct sk_buff *skb) 483{ 484 struct tcp_sock *tp = tcp_sk(sk); 485 if (tp->rx_opt.rcv_tsecr && 486 (TCP_SKB_CB(skb)->end_seq - 487 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss)) 488 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0); 489} 490 491/* 492 * This function should be called every time data is copied to user space. 493 * It calculates the appropriate TCP receive buffer space. 494 */ 495void tcp_rcv_space_adjust(struct sock *sk) 496{ 497 struct tcp_sock *tp = tcp_sk(sk); 498 int time; 499 int space; 500 501 if (tp->rcvq_space.time == 0) 502 goto new_measure; 503 504 time = tcp_time_stamp - tp->rcvq_space.time; 505 if (time < (tp->rcv_rtt_est.rtt >> 3) || 506 tp->rcv_rtt_est.rtt == 0) 507 return; 508 509 space = 2 * (tp->copied_seq - tp->rcvq_space.seq); 510 511 space = max(tp->rcvq_space.space, space); 512 513 if (tp->rcvq_space.space != space) { 514 int rcvmem; 515 516 tp->rcvq_space.space = space; 517 518 if (sysctl_tcp_moderate_rcvbuf && 519 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) { 520 int new_clamp = space; 521 522 /* Receive space grows, normalize in order to 523 * take into account packet headers and sk_buff 524 * structure overhead. 525 */ 526 space /= tp->advmss; 527 if (!space) 528 space = 1; 529 rcvmem = (tp->advmss + MAX_TCP_HEADER + 530 16 + sizeof(struct sk_buff)); 531 while (tcp_win_from_space(rcvmem) < tp->advmss) 532 rcvmem += 128; 533 space *= rcvmem; 534 space = min(space, sysctl_tcp_rmem[2]); 535 if (space > sk->sk_rcvbuf) { 536 sk->sk_rcvbuf = space; 537 538 /* Make the window clamp follow along. */ 539 tp->window_clamp = new_clamp; 540 } 541 } 542 } 543 544new_measure: 545 tp->rcvq_space.seq = tp->copied_seq; 546 tp->rcvq_space.time = tcp_time_stamp; 547} 548 549/* There is something which you must keep in mind when you analyze the 550 * behavior of the tp->ato delayed ack timeout interval. When a 551 * connection starts up, we want to ack as quickly as possible. The 552 * problem is that "good" TCP's do slow start at the beginning of data 553 * transmission. The means that until we send the first few ACK's the 554 * sender will sit on his end and only queue most of his data, because 555 * he can only send snd_cwnd unacked packets at any given time. For 556 * each ACK we send, he increments snd_cwnd and transmits more of his 557 * queue. -DaveM 558 */ 559static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb) 560{ 561 struct tcp_sock *tp = tcp_sk(sk); 562 struct inet_connection_sock *icsk = inet_csk(sk); 563 u32 now; 564 565 inet_csk_schedule_ack(sk); 566 567 tcp_measure_rcv_mss(sk, skb); 568 569 tcp_rcv_rtt_measure(tp); 570 571 now = tcp_time_stamp; 572 573 if (!icsk->icsk_ack.ato) { 574 /* The _first_ data packet received, initialize 575 * delayed ACK engine. 576 */ 577 tcp_incr_quickack(sk); 578 icsk->icsk_ack.ato = TCP_ATO_MIN; 579 } else { 580 int m = now - icsk->icsk_ack.lrcvtime; 581 582 if (m <= TCP_ATO_MIN/2) { 583 /* The fastest case is the first. */ 584 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2; 585 } else if (m < icsk->icsk_ack.ato) { 586 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m; 587 if (icsk->icsk_ack.ato > icsk->icsk_rto) 588 icsk->icsk_ack.ato = icsk->icsk_rto; 589 } else if (m > icsk->icsk_rto) { 590 /* Too long gap. Apparently sender failed to 591 * restart window, so that we send ACKs quickly. 592 */ 593 tcp_incr_quickack(sk); 594 sk_stream_mem_reclaim(sk); 595 } 596 } 597 icsk->icsk_ack.lrcvtime = now; 598 599 TCP_ECN_check_ce(tp, skb); 600 601 if (skb->len >= 128) 602 tcp_grow_window(sk, skb); 603} 604 605static u32 tcp_rto_min(struct sock *sk) 606{ 607 struct dst_entry *dst = __sk_dst_get(sk); 608 u32 rto_min = TCP_RTO_MIN; 609 610 if (dst && dst_metric_locked(dst, RTAX_RTO_MIN)) 611 rto_min = dst->metrics[RTAX_RTO_MIN-1]; 612 return rto_min; 613} 614 615/* Called to compute a smoothed rtt estimate. The data fed to this 616 * routine either comes from timestamps, or from segments that were 617 * known _not_ to have been retransmitted [see Karn/Partridge 618 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88 619 * piece by Van Jacobson. 620 * NOTE: the next three routines used to be one big routine. 621 * To save cycles in the RFC 1323 implementation it was better to break 622 * it up into three procedures. -- erics 623 */ 624static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt) 625{ 626 struct tcp_sock *tp = tcp_sk(sk); 627 long m = mrtt; /* RTT */ 628 629 /* The following amusing code comes from Jacobson's 630 * article in SIGCOMM '88. Note that rtt and mdev 631 * are scaled versions of rtt and mean deviation. 632 * This is designed to be as fast as possible 633 * m stands for "measurement". 634 * 635 * On a 1990 paper the rto value is changed to: 636 * RTO = rtt + 4 * mdev 637 * 638 * Funny. This algorithm seems to be very broken. 639 * These formulae increase RTO, when it should be decreased, increase 640 * too slowly, when it should be increased quickly, decrease too quickly 641 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely 642 * does not matter how to _calculate_ it. Seems, it was trap 643 * that VJ failed to avoid. 8) 644 */ 645 if (m == 0) 646 m = 1; 647 if (tp->srtt != 0) { 648 m -= (tp->srtt >> 3); /* m is now error in rtt est */ 649 tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */ 650 if (m < 0) { 651 m = -m; /* m is now abs(error) */ 652 m -= (tp->mdev >> 2); /* similar update on mdev */ 653 /* This is similar to one of Eifel findings. 654 * Eifel blocks mdev updates when rtt decreases. 655 * This solution is a bit different: we use finer gain 656 * for mdev in this case (alpha*beta). 657 * Like Eifel it also prevents growth of rto, 658 * but also it limits too fast rto decreases, 659 * happening in pure Eifel. 660 */ 661 if (m > 0) 662 m >>= 3; 663 } else { 664 m -= (tp->mdev >> 2); /* similar update on mdev */ 665 } 666 tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */ 667 if (tp->mdev > tp->mdev_max) { 668 tp->mdev_max = tp->mdev; 669 if (tp->mdev_max > tp->rttvar) 670 tp->rttvar = tp->mdev_max; 671 } 672 if (after(tp->snd_una, tp->rtt_seq)) { 673 if (tp->mdev_max < tp->rttvar) 674 tp->rttvar -= (tp->rttvar-tp->mdev_max)>>2; 675 tp->rtt_seq = tp->snd_nxt; 676 tp->mdev_max = tcp_rto_min(sk); 677 } 678 } else { 679 /* no previous measure. */ 680 tp->srtt = m<<3; /* take the measured time to be rtt */ 681 tp->mdev = m<<1; /* make sure rto = 3*rtt */ 682 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk)); 683 tp->rtt_seq = tp->snd_nxt; 684 } 685} 686 687/* Calculate rto without backoff. This is the second half of Van Jacobson's 688 * routine referred to above. 689 */ 690static inline void tcp_set_rto(struct sock *sk) 691{ 692 const struct tcp_sock *tp = tcp_sk(sk); 693 /* Old crap is replaced with new one. 8) 694 * 695 * More seriously: 696 * 1. If rtt variance happened to be less 50msec, it is hallucination. 697 * It cannot be less due to utterly erratic ACK generation made 698 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_ 699 * to do with delayed acks, because at cwnd>2 true delack timeout 700 * is invisible. Actually, Linux-2.4 also generates erratic 701 * ACKs in some circumstances. 702 */ 703 inet_csk(sk)->icsk_rto = (tp->srtt >> 3) + tp->rttvar; 704 705 /* 2. Fixups made earlier cannot be right. 706 * If we do not estimate RTO correctly without them, 707 * all the algo is pure shit and should be replaced 708 * with correct one. It is exactly, which we pretend to do. 709 */ 710} 711 712/* NOTE: clamping at TCP_RTO_MIN is not required, current algo 713 * guarantees that rto is higher. 714 */ 715static inline void tcp_bound_rto(struct sock *sk) 716{ 717 if (inet_csk(sk)->icsk_rto > TCP_RTO_MAX) 718 inet_csk(sk)->icsk_rto = TCP_RTO_MAX; 719} 720 721/* Save metrics learned by this TCP session. 722 This function is called only, when TCP finishes successfully 723 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE. 724 */ 725void tcp_update_metrics(struct sock *sk) 726{ 727 struct tcp_sock *tp = tcp_sk(sk); 728 struct dst_entry *dst = __sk_dst_get(sk); 729 730 if (sysctl_tcp_nometrics_save) 731 return; 732 733 dst_confirm(dst); 734 735 if (dst && (dst->flags&DST_HOST)) { 736 const struct inet_connection_sock *icsk = inet_csk(sk); 737 int m; 738 739 if (icsk->icsk_backoff || !tp->srtt) { 740 /* This session failed to estimate rtt. Why? 741 * Probably, no packets returned in time. 742 * Reset our results. 743 */ 744 if (!(dst_metric_locked(dst, RTAX_RTT))) 745 dst->metrics[RTAX_RTT-1] = 0; 746 return; 747 } 748 749 m = dst_metric(dst, RTAX_RTT) - tp->srtt; 750 751 /* If newly calculated rtt larger than stored one, 752 * store new one. Otherwise, use EWMA. Remember, 753 * rtt overestimation is always better than underestimation. 754 */ 755 if (!(dst_metric_locked(dst, RTAX_RTT))) { 756 if (m <= 0) 757 dst->metrics[RTAX_RTT-1] = tp->srtt; 758 else 759 dst->metrics[RTAX_RTT-1] -= (m>>3); 760 } 761 762 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) { 763 if (m < 0) 764 m = -m; 765 766 /* Scale deviation to rttvar fixed point */ 767 m >>= 1; 768 if (m < tp->mdev) 769 m = tp->mdev; 770 771 if (m >= dst_metric(dst, RTAX_RTTVAR)) 772 dst->metrics[RTAX_RTTVAR-1] = m; 773 else 774 dst->metrics[RTAX_RTTVAR-1] -= 775 (dst->metrics[RTAX_RTTVAR-1] - m)>>2; 776 } 777 778 if (tp->snd_ssthresh >= 0xFFFF) { 779 /* Slow start still did not finish. */ 780 if (dst_metric(dst, RTAX_SSTHRESH) && 781 !dst_metric_locked(dst, RTAX_SSTHRESH) && 782 (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH)) 783 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1; 784 if (!dst_metric_locked(dst, RTAX_CWND) && 785 tp->snd_cwnd > dst_metric(dst, RTAX_CWND)) 786 dst->metrics[RTAX_CWND-1] = tp->snd_cwnd; 787 } else if (tp->snd_cwnd > tp->snd_ssthresh && 788 icsk->icsk_ca_state == TCP_CA_Open) { 789 /* Cong. avoidance phase, cwnd is reliable. */ 790 if (!dst_metric_locked(dst, RTAX_SSTHRESH)) 791 dst->metrics[RTAX_SSTHRESH-1] = 792 max(tp->snd_cwnd >> 1, tp->snd_ssthresh); 793 if (!dst_metric_locked(dst, RTAX_CWND)) 794 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_cwnd) >> 1; 795 } else { 796 /* Else slow start did not finish, cwnd is non-sense, 797 ssthresh may be also invalid. 798 */ 799 if (!dst_metric_locked(dst, RTAX_CWND)) 800 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_ssthresh) >> 1; 801 if (dst->metrics[RTAX_SSTHRESH-1] && 802 !dst_metric_locked(dst, RTAX_SSTHRESH) && 803 tp->snd_ssthresh > dst->metrics[RTAX_SSTHRESH-1]) 804 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh; 805 } 806 807 if (!dst_metric_locked(dst, RTAX_REORDERING)) { 808 if (dst->metrics[RTAX_REORDERING-1] < tp->reordering && 809 tp->reordering != sysctl_tcp_reordering) 810 dst->metrics[RTAX_REORDERING-1] = tp->reordering; 811 } 812 } 813} 814 815/* Numbers are taken from RFC3390. 816 * 817 * John Heffner states: 818 * 819 * The RFC specifies a window of no more than 4380 bytes 820 * unless 2*MSS > 4380. Reading the pseudocode in the RFC 821 * is a bit misleading because they use a clamp at 4380 bytes 822 * rather than use a multiplier in the relevant range. 823 */ 824__u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst) 825{ 826 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0); 827 828 if (!cwnd) { 829 if (tp->mss_cache > 1460) 830 cwnd = 2; 831 else 832 cwnd = (tp->mss_cache > 1095) ? 3 : 4; 833 } 834 return min_t(__u32, cwnd, tp->snd_cwnd_clamp); 835} 836 837/* Set slow start threshold and cwnd not falling to slow start */ 838void tcp_enter_cwr(struct sock *sk, const int set_ssthresh) 839{ 840 struct tcp_sock *tp = tcp_sk(sk); 841 const struct inet_connection_sock *icsk = inet_csk(sk); 842 843 tp->prior_ssthresh = 0; 844 tp->bytes_acked = 0; 845 if (icsk->icsk_ca_state < TCP_CA_CWR) { 846 tp->undo_marker = 0; 847 if (set_ssthresh) 848 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 849 tp->snd_cwnd = min(tp->snd_cwnd, 850 tcp_packets_in_flight(tp) + 1U); 851 tp->snd_cwnd_cnt = 0; 852 tp->high_seq = tp->snd_nxt; 853 tp->snd_cwnd_stamp = tcp_time_stamp; 854 TCP_ECN_queue_cwr(tp); 855 856 tcp_set_ca_state(sk, TCP_CA_CWR); 857 } 858} 859 860/* 861 * Packet counting of FACK is based on in-order assumptions, therefore TCP 862 * disables it when reordering is detected 863 */ 864static void tcp_disable_fack(struct tcp_sock *tp) 865{ 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(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(dst, RTAX_RTT) > tp->srtt) { 921 tp->srtt = dst_metric(dst, RTAX_RTT); 922 tp->rtt_seq = tp->snd_nxt; 923 } 924 if (dst_metric(dst, RTAX_RTTVAR) > tp->mdev) { 925 tp->mdev = dst_metric(dst, RTAX_RTTVAR); 926 tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN); 927 } 928 tcp_set_rto(sk); 929 tcp_bound_rto(sk); 930 if (inet_csk(sk)->icsk_rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp) 931 goto reset; 932 tp->snd_cwnd = tcp_init_cwnd(tp, dst); 933 tp->snd_cwnd_stamp = tcp_time_stamp; 934 return; 935 936reset: 937 /* Play conservative. If timestamps are not 938 * supported, TCP will fail to recalculate correct 939 * rtt, if initial rto is too small. FORGET ALL AND RESET! 940 */ 941 if (!tp->rx_opt.saw_tstamp && tp->srtt) { 942 tp->srtt = 0; 943 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT; 944 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; 945 } 946} 947 948static void tcp_update_reordering(struct sock *sk, const int metric, 949 const int ts) 950{ 951 struct tcp_sock *tp = tcp_sk(sk); 952 if (metric > tp->reordering) { 953 tp->reordering = min(TCP_MAX_REORDERING, metric); 954 955 /* This exciting event is worth to be remembered. 8) */ 956 if (ts) 957 NET_INC_STATS_BH(LINUX_MIB_TCPTSREORDER); 958 else if (tcp_is_reno(tp)) 959 NET_INC_STATS_BH(LINUX_MIB_TCPRENOREORDER); 960 else if (tcp_is_fack(tp)) 961 NET_INC_STATS_BH(LINUX_MIB_TCPFACKREORDER); 962 else 963 NET_INC_STATS_BH(LINUX_MIB_TCPSACKREORDER); 964#if FASTRETRANS_DEBUG > 1 965 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n", 966 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state, 967 tp->reordering, 968 tp->fackets_out, 969 tp->sacked_out, 970 tp->undo_marker ? tp->undo_retrans : 0); 971#endif 972 tcp_disable_fack(tp); 973 } 974} 975 976/* This procedure tags the retransmission queue when SACKs arrive. 977 * 978 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L). 979 * Packets in queue with these bits set are counted in variables 980 * sacked_out, retrans_out and lost_out, correspondingly. 981 * 982 * Valid combinations are: 983 * Tag InFlight Description 984 * 0 1 - orig segment is in flight. 985 * S 0 - nothing flies, orig reached receiver. 986 * L 0 - nothing flies, orig lost by net. 987 * R 2 - both orig and retransmit are in flight. 988 * L|R 1 - orig is lost, retransmit is in flight. 989 * S|R 1 - orig reached receiver, retrans is still in flight. 990 * (L|S|R is logically valid, it could occur when L|R is sacked, 991 * but it is equivalent to plain S and code short-curcuits it to S. 992 * L|S is logically invalid, it would mean -1 packet in flight 8)) 993 * 994 * These 6 states form finite state machine, controlled by the following events: 995 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue()) 996 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue()) 997 * 3. Loss detection event of one of three flavors: 998 * A. Scoreboard estimator decided the packet is lost. 999 * A'. Reno "three dupacks" marks head of queue lost. 1000 * A''. Its FACK modfication, head until snd.fack is lost. 1001 * B. SACK arrives sacking data transmitted after never retransmitted 1002 * hole was sent out. 1003 * C. SACK arrives sacking SND.NXT at the moment, when the 1004 * segment was retransmitted. 1005 * 4. D-SACK added new rule: D-SACK changes any tag to S. 1006 * 1007 * It is pleasant to note, that state diagram turns out to be commutative, 1008 * so that we are allowed not to be bothered by order of our actions, 1009 * when multiple events arrive simultaneously. (see the function below). 1010 * 1011 * Reordering detection. 1012 * -------------------- 1013 * Reordering metric is maximal distance, which a packet can be displaced 1014 * in packet stream. With SACKs we can estimate it: 1015 * 1016 * 1. SACK fills old hole and the corresponding segment was not 1017 * ever retransmitted -> reordering. Alas, we cannot use it 1018 * when segment was retransmitted. 1019 * 2. The last flaw is solved with D-SACK. D-SACK arrives 1020 * for retransmitted and already SACKed segment -> reordering.. 1021 * Both of these heuristics are not used in Loss state, when we cannot 1022 * account for retransmits accurately. 1023 * 1024 * SACK block validation. 1025 * ---------------------- 1026 * 1027 * SACK block range validation checks that the received SACK block fits to 1028 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT. 1029 * Note that SND.UNA is not included to the range though being valid because 1030 * it means that the receiver is rather inconsistent with itself reporting 1031 * SACK reneging when it should advance SND.UNA. Such SACK block this is 1032 * perfectly valid, however, in light of RFC2018 which explicitly states 1033 * that "SACK block MUST reflect the newest segment. Even if the newest 1034 * segment is going to be discarded ...", not that it looks very clever 1035 * in case of head skb. Due to potentional receiver driven attacks, we 1036 * choose to avoid immediate execution of a walk in write queue due to 1037 * reneging and defer head skb's loss recovery to standard loss recovery 1038 * procedure that will eventually trigger (nothing forbids us doing this). 1039 * 1040 * Implements also blockage to start_seq wrap-around. Problem lies in the 1041 * fact that though start_seq (s) is before end_seq (i.e., not reversed), 1042 * there's no guarantee that it will be before snd_nxt (n). The problem 1043 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt 1044 * wrap (s_w): 1045 * 1046 * <- outs wnd -> <- wrapzone -> 1047 * u e n u_w e_w s n_w 1048 * | | | | | | | 1049 * |<------------+------+----- TCP seqno space --------------+---------->| 1050 * ...-- <2^31 ->| |<--------... 1051 * ...---- >2^31 ------>| |<--------... 1052 * 1053 * Current code wouldn't be vulnerable but it's better still to discard such 1054 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat 1055 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in 1056 * snd_nxt wrap -> snd_una region will then become "well defined", i.e., 1057 * equal to the ideal case (infinite seqno space without wrap caused issues). 1058 * 1059 * With D-SACK the lower bound is extended to cover sequence space below 1060 * SND.UNA down to undo_marker, which is the last point of interest. Yet 1061 * again, D-SACK block must not to go across snd_una (for the same reason as 1062 * for the normal SACK blocks, explained above). But there all simplicity 1063 * ends, TCP might receive valid D-SACKs below that. As long as they reside 1064 * fully below undo_marker they do not affect behavior in anyway and can 1065 * therefore be safely ignored. In rare cases (which are more or less 1066 * theoretical ones), the D-SACK will nicely cross that boundary due to skb 1067 * fragmentation and packet reordering past skb's retransmission. To consider 1068 * them correctly, the acceptable range must be extended even more though 1069 * the exact amount is rather hard to quantify. However, tp->max_window can 1070 * be used as an exaggerated estimate. 1071 */ 1072static int tcp_is_sackblock_valid(struct tcp_sock *tp, int is_dsack, 1073 u32 start_seq, u32 end_seq) 1074{ 1075 /* Too far in future, or reversed (interpretation is ambiguous) */ 1076 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq)) 1077 return 0; 1078 1079 /* Nasty start_seq wrap-around check (see comments above) */ 1080 if (!before(start_seq, tp->snd_nxt)) 1081 return 0; 1082 1083 /* In outstanding window? ...This is valid exit for D-SACKs too. 1084 * start_seq == snd_una is non-sensical (see comments above) 1085 */ 1086 if (after(start_seq, tp->snd_una)) 1087 return 1; 1088 1089 if (!is_dsack || !tp->undo_marker) 1090 return 0; 1091 1092 /* ...Then it's D-SACK, and must reside below snd_una completely */ 1093 if (!after(end_seq, tp->snd_una)) 1094 return 0; 1095 1096 if (!before(start_seq, tp->undo_marker)) 1097 return 1; 1098 1099 /* Too old */ 1100 if (!after(end_seq, tp->undo_marker)) 1101 return 0; 1102 1103 /* Undo_marker boundary crossing (overestimates a lot). Known already: 1104 * start_seq < undo_marker and end_seq >= undo_marker. 1105 */ 1106 return !before(start_seq, end_seq - tp->max_window); 1107} 1108 1109/* Check for lost retransmit. This superb idea is borrowed from "ratehalving". 1110 * Event "C". Later note: FACK people cheated me again 8), we have to account 1111 * for reordering! Ugly, but should help. 1112 * 1113 * Search retransmitted skbs from write_queue that were sent when snd_nxt was 1114 * less than what is now known to be received by the other end (derived from 1115 * SACK blocks by the caller). Also calculate the lowest snd_nxt among the 1116 * remaining retransmitted skbs to avoid some costly processing per ACKs. 1117 */ 1118static int tcp_mark_lost_retrans(struct sock *sk, u32 received_upto) 1119{ 1120 struct tcp_sock *tp = tcp_sk(sk); 1121 struct sk_buff *skb; 1122 int flag = 0; 1123 int cnt = 0; 1124 u32 new_low_seq = tp->snd_nxt; 1125 1126 tcp_for_write_queue(skb, sk) { 1127 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq; 1128 1129 if (skb == tcp_send_head(sk)) 1130 break; 1131 if (cnt == tp->retrans_out) 1132 break; 1133 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 1134 continue; 1135 1136 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)) 1137 continue; 1138 1139 if (after(received_upto, ack_seq) && 1140 (tcp_is_fack(tp) || 1141 !before(received_upto, 1142 ack_seq + tp->reordering * tp->mss_cache))) { 1143 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 1144 tp->retrans_out -= tcp_skb_pcount(skb); 1145 1146 /* clear lost hint */ 1147 tp->retransmit_skb_hint = NULL; 1148 1149 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) { 1150 tp->lost_out += tcp_skb_pcount(skb); 1151 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 1152 flag |= FLAG_DATA_SACKED; 1153 NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT); 1154 } 1155 } else { 1156 if (before(ack_seq, new_low_seq)) 1157 new_low_seq = ack_seq; 1158 cnt += tcp_skb_pcount(skb); 1159 } 1160 } 1161 1162 if (tp->retrans_out) 1163 tp->lost_retrans_low = new_low_seq; 1164 1165 return flag; 1166} 1167 1168static int tcp_check_dsack(struct tcp_sock *tp, struct sk_buff *ack_skb, 1169 struct tcp_sack_block_wire *sp, int num_sacks, 1170 u32 prior_snd_una) 1171{ 1172 u32 start_seq_0 = ntohl(get_unaligned(&sp[0].start_seq)); 1173 u32 end_seq_0 = ntohl(get_unaligned(&sp[0].end_seq)); 1174 int dup_sack = 0; 1175 1176 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) { 1177 dup_sack = 1; 1178 tcp_dsack_seen(tp); 1179 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV); 1180 } else if (num_sacks > 1) { 1181 u32 end_seq_1 = ntohl(get_unaligned(&sp[1].end_seq)); 1182 u32 start_seq_1 = ntohl(get_unaligned(&sp[1].start_seq)); 1183 1184 if (!after(end_seq_0, end_seq_1) && 1185 !before(start_seq_0, start_seq_1)) { 1186 dup_sack = 1; 1187 tcp_dsack_seen(tp); 1188 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV); 1189 } 1190 } 1191 1192 /* D-SACK for already forgotten data... Do dumb counting. */ 1193 if (dup_sack && 1194 !after(end_seq_0, prior_snd_una) && 1195 after(end_seq_0, tp->undo_marker)) 1196 tp->undo_retrans--; 1197 1198 return dup_sack; 1199} 1200 1201/* Check if skb is fully within the SACK block. In presence of GSO skbs, 1202 * the incoming SACK may not exactly match but we can find smaller MSS 1203 * aligned portion of it that matches. Therefore we might need to fragment 1204 * which may fail and creates some hassle (caller must handle error case 1205 * returns). 1206 */ 1207static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb, 1208 u32 start_seq, u32 end_seq) 1209{ 1210 int in_sack, err; 1211 unsigned int pkt_len; 1212 1213 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && 1214 !before(end_seq, TCP_SKB_CB(skb)->end_seq); 1215 1216 if (tcp_skb_pcount(skb) > 1 && !in_sack && 1217 after(TCP_SKB_CB(skb)->end_seq, start_seq)) { 1218 1219 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); 1220 1221 if (!in_sack) 1222 pkt_len = start_seq - TCP_SKB_CB(skb)->seq; 1223 else 1224 pkt_len = end_seq - TCP_SKB_CB(skb)->seq; 1225 err = tcp_fragment(sk, skb, pkt_len, skb_shinfo(skb)->gso_size); 1226 if (err < 0) 1227 return err; 1228 } 1229 1230 return in_sack; 1231} 1232 1233static int 1234tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, u32 prior_snd_una) 1235{ 1236 const struct inet_connection_sock *icsk = inet_csk(sk); 1237 struct tcp_sock *tp = tcp_sk(sk); 1238 unsigned char *ptr = (skb_transport_header(ack_skb) + 1239 TCP_SKB_CB(ack_skb)->sacked); 1240 struct tcp_sack_block_wire *sp = (struct tcp_sack_block_wire *)(ptr+2); 1241 struct sk_buff *cached_skb; 1242 int num_sacks = (ptr[1] - TCPOLEN_SACK_BASE)>>3; 1243 int reord = tp->packets_out; 1244 int prior_fackets; 1245 u32 highest_sack_end_seq = tp->lost_retrans_low; 1246 int flag = 0; 1247 int found_dup_sack = 0; 1248 int cached_fack_count; 1249 int i; 1250 int first_sack_index; 1251 int force_one_sack; 1252 1253 if (!tp->sacked_out) { 1254 if (WARN_ON(tp->fackets_out)) 1255 tp->fackets_out = 0; 1256 tp->highest_sack = tp->snd_una; 1257 } 1258 prior_fackets = tp->fackets_out; 1259 1260 found_dup_sack = tcp_check_dsack(tp, ack_skb, sp, 1261 num_sacks, prior_snd_una); 1262 if (found_dup_sack) 1263 flag |= FLAG_DSACKING_ACK; 1264 1265 /* Eliminate too old ACKs, but take into 1266 * account more or less fresh ones, they can 1267 * contain valid SACK info. 1268 */ 1269 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window)) 1270 return 0; 1271 1272 /* SACK fastpath: 1273 * if the only SACK change is the increase of the end_seq of 1274 * the first block then only apply that SACK block 1275 * and use retrans queue hinting otherwise slowpath */ 1276 force_one_sack = 1; 1277 for (i = 0; i < num_sacks; i++) { 1278 __be32 start_seq = sp[i].start_seq; 1279 __be32 end_seq = sp[i].end_seq; 1280 1281 if (i == 0) { 1282 if (tp->recv_sack_cache[i].start_seq != start_seq) 1283 force_one_sack = 0; 1284 } else { 1285 if ((tp->recv_sack_cache[i].start_seq != start_seq) || 1286 (tp->recv_sack_cache[i].end_seq != end_seq)) 1287 force_one_sack = 0; 1288 } 1289 tp->recv_sack_cache[i].start_seq = start_seq; 1290 tp->recv_sack_cache[i].end_seq = end_seq; 1291 } 1292 /* Clear the rest of the cache sack blocks so they won't match mistakenly. */ 1293 for (; i < ARRAY_SIZE(tp->recv_sack_cache); i++) { 1294 tp->recv_sack_cache[i].start_seq = 0; 1295 tp->recv_sack_cache[i].end_seq = 0; 1296 } 1297 1298 first_sack_index = 0; 1299 if (force_one_sack) 1300 num_sacks = 1; 1301 else { 1302 int j; 1303 tp->fastpath_skb_hint = NULL; 1304 1305 /* order SACK blocks to allow in order walk of the retrans queue */ 1306 for (i = num_sacks-1; i > 0; i--) { 1307 for (j = 0; j < i; j++){ 1308 if (after(ntohl(sp[j].start_seq), 1309 ntohl(sp[j+1].start_seq))){ 1310 struct tcp_sack_block_wire tmp; 1311 1312 tmp = sp[j]; 1313 sp[j] = sp[j+1]; 1314 sp[j+1] = tmp; 1315 1316 /* Track where the first SACK block goes to */ 1317 if (j == first_sack_index) 1318 first_sack_index = j+1; 1319 } 1320 1321 } 1322 } 1323 } 1324 1325 /* Use SACK fastpath hint if valid */ 1326 cached_skb = tp->fastpath_skb_hint; 1327 cached_fack_count = tp->fastpath_cnt_hint; 1328 if (!cached_skb) { 1329 cached_skb = tcp_write_queue_head(sk); 1330 cached_fack_count = 0; 1331 } 1332 1333 for (i = 0; i < num_sacks; i++) { 1334 struct sk_buff *skb; 1335 __u32 start_seq = ntohl(sp->start_seq); 1336 __u32 end_seq = ntohl(sp->end_seq); 1337 int fack_count; 1338 int dup_sack = (found_dup_sack && (i == first_sack_index)); 1339 int next_dup = (found_dup_sack && (i+1 == first_sack_index)); 1340 1341 sp++; 1342 1343 if (!tcp_is_sackblock_valid(tp, dup_sack, start_seq, end_seq)) { 1344 if (dup_sack) { 1345 if (!tp->undo_marker) 1346 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKIGNOREDNOUNDO); 1347 else 1348 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKIGNOREDOLD); 1349 } else { 1350 /* Don't count olds caused by ACK reordering */ 1351 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) && 1352 !after(end_seq, tp->snd_una)) 1353 continue; 1354 NET_INC_STATS_BH(LINUX_MIB_TCPSACKDISCARD); 1355 } 1356 continue; 1357 } 1358 1359 skb = cached_skb; 1360 fack_count = cached_fack_count; 1361 1362 /* Event "B" in the comment above. */ 1363 if (after(end_seq, tp->high_seq)) 1364 flag |= FLAG_DATA_LOST; 1365 1366 tcp_for_write_queue_from(skb, sk) { 1367 int in_sack = 0; 1368 u8 sacked; 1369 1370 if (skb == tcp_send_head(sk)) 1371 break; 1372 1373 cached_skb = skb; 1374 cached_fack_count = fack_count; 1375 if (i == first_sack_index) { 1376 tp->fastpath_skb_hint = skb; 1377 tp->fastpath_cnt_hint = fack_count; 1378 } 1379 1380 /* The retransmission queue is always in order, so 1381 * we can short-circuit the walk early. 1382 */ 1383 if (!before(TCP_SKB_CB(skb)->seq, end_seq)) 1384 break; 1385 1386 dup_sack = (found_dup_sack && (i == first_sack_index)); 1387 1388 /* Due to sorting DSACK may reside within this SACK block! */ 1389 if (next_dup) { 1390 u32 dup_start = ntohl(sp->start_seq); 1391 u32 dup_end = ntohl(sp->end_seq); 1392 1393 if (before(TCP_SKB_CB(skb)->seq, dup_end)) { 1394 in_sack = tcp_match_skb_to_sack(sk, skb, dup_start, dup_end); 1395 if (in_sack > 0) 1396 dup_sack = 1; 1397 } 1398 } 1399 1400 /* DSACK info lost if out-of-mem, try SACK still */ 1401 if (in_sack <= 0) 1402 in_sack = tcp_match_skb_to_sack(sk, skb, start_seq, end_seq); 1403 if (in_sack < 0) 1404 break; 1405 1406 sacked = TCP_SKB_CB(skb)->sacked; 1407 1408 /* Account D-SACK for retransmitted packet. */ 1409 if ((dup_sack && in_sack) && 1410 (sacked & TCPCB_RETRANS) && 1411 after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker)) 1412 tp->undo_retrans--; 1413 1414 /* The frame is ACKed. */ 1415 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) { 1416 if (sacked&TCPCB_RETRANS) { 1417 if ((dup_sack && in_sack) && 1418 (sacked&TCPCB_SACKED_ACKED)) 1419 reord = min(fack_count, reord); 1420 } else { 1421 /* If it was in a hole, we detected reordering. */ 1422 if (fack_count < prior_fackets && 1423 !(sacked&TCPCB_SACKED_ACKED)) 1424 reord = min(fack_count, reord); 1425 } 1426 1427 /* Nothing to do; acked frame is about to be dropped. */ 1428 fack_count += tcp_skb_pcount(skb); 1429 continue; 1430 } 1431 1432 if (!in_sack) { 1433 fack_count += tcp_skb_pcount(skb); 1434 continue; 1435 } 1436 1437 if (!(sacked&TCPCB_SACKED_ACKED)) { 1438 if (sacked & TCPCB_SACKED_RETRANS) { 1439 /* If the segment is not tagged as lost, 1440 * we do not clear RETRANS, believing 1441 * that retransmission is still in flight. 1442 */ 1443 if (sacked & TCPCB_LOST) { 1444 TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS); 1445 tp->lost_out -= tcp_skb_pcount(skb); 1446 tp->retrans_out -= tcp_skb_pcount(skb); 1447 1448 /* clear lost hint */ 1449 tp->retransmit_skb_hint = NULL; 1450 } 1451 } else { 1452 /* New sack for not retransmitted frame, 1453 * which was in hole. It is reordering. 1454 */ 1455 if (!(sacked & TCPCB_RETRANS) && 1456 fack_count < prior_fackets) 1457 reord = min(fack_count, reord); 1458 1459 if (sacked & TCPCB_LOST) { 1460 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; 1461 tp->lost_out -= tcp_skb_pcount(skb); 1462 1463 /* clear lost hint */ 1464 tp->retransmit_skb_hint = NULL; 1465 } 1466 /* SACK enhanced F-RTO detection. 1467 * Set flag if and only if non-rexmitted 1468 * segments below frto_highmark are 1469 * SACKed (RFC4138; Appendix B). 1470 * Clearing correct due to in-order walk 1471 */ 1472 if (after(end_seq, tp->frto_highmark)) { 1473 flag &= ~FLAG_ONLY_ORIG_SACKED; 1474 } else { 1475 if (!(sacked & TCPCB_RETRANS)) 1476 flag |= FLAG_ONLY_ORIG_SACKED; 1477 } 1478 } 1479 1480 TCP_SKB_CB(skb)->sacked |= TCPCB_SACKED_ACKED; 1481 flag |= FLAG_DATA_SACKED; 1482 tp->sacked_out += tcp_skb_pcount(skb); 1483 1484 fack_count += tcp_skb_pcount(skb); 1485 if (fack_count > tp->fackets_out) 1486 tp->fackets_out = fack_count; 1487 1488 if (after(TCP_SKB_CB(skb)->seq, tp->highest_sack)) { 1489 tp->highest_sack = TCP_SKB_CB(skb)->seq; 1490 highest_sack_end_seq = TCP_SKB_CB(skb)->end_seq; 1491 } 1492 } else { 1493 if (dup_sack && (sacked&TCPCB_RETRANS)) 1494 reord = min(fack_count, reord); 1495 1496 fack_count += tcp_skb_pcount(skb); 1497 } 1498 1499 /* D-SACK. We can detect redundant retransmission 1500 * in S|R and plain R frames and clear it. 1501 * undo_retrans is decreased above, L|R frames 1502 * are accounted above as well. 1503 */ 1504 if (dup_sack && 1505 (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS)) { 1506 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 1507 tp->retrans_out -= tcp_skb_pcount(skb); 1508 tp->retransmit_skb_hint = NULL; 1509 } 1510 } 1511 } 1512 1513 if (tp->retrans_out && 1514 after(highest_sack_end_seq, tp->lost_retrans_low) && 1515 icsk->icsk_ca_state == TCP_CA_Recovery) 1516 flag |= tcp_mark_lost_retrans(sk, highest_sack_end_seq); 1517 1518 tcp_verify_left_out(tp); 1519 1520 if ((reord < tp->fackets_out) && icsk->icsk_ca_state != TCP_CA_Loss && 1521 (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark))) 1522 tcp_update_reordering(sk, tp->fackets_out - reord, 0); 1523 1524#if FASTRETRANS_DEBUG > 0 1525 BUG_TRAP((int)tp->sacked_out >= 0); 1526 BUG_TRAP((int)tp->lost_out >= 0); 1527 BUG_TRAP((int)tp->retrans_out >= 0); 1528 BUG_TRAP((int)tcp_packets_in_flight(tp) >= 0); 1529#endif 1530 return flag; 1531} 1532 1533/* If we receive more dupacks than we expected counting segments 1534 * in assumption of absent reordering, interpret this as reordering. 1535 * The only another reason could be bug in receiver TCP. 1536 */ 1537static void tcp_check_reno_reordering(struct sock *sk, const int addend) 1538{ 1539 struct tcp_sock *tp = tcp_sk(sk); 1540 u32 holes; 1541 1542 holes = max(tp->lost_out, 1U); 1543 holes = min(holes, tp->packets_out); 1544 1545 if ((tp->sacked_out + holes) > tp->packets_out) { 1546 tp->sacked_out = tp->packets_out - holes; 1547 tcp_update_reordering(sk, tp->packets_out + addend, 0); 1548 } 1549} 1550 1551/* Emulate SACKs for SACKless connection: account for a new dupack. */ 1552 1553static void tcp_add_reno_sack(struct sock *sk) 1554{ 1555 struct tcp_sock *tp = tcp_sk(sk); 1556 tp->sacked_out++; 1557 tcp_check_reno_reordering(sk, 0); 1558 tcp_verify_left_out(tp); 1559} 1560 1561/* Account for ACK, ACKing some data in Reno Recovery phase. */ 1562 1563static void tcp_remove_reno_sacks(struct sock *sk, int acked) 1564{ 1565 struct tcp_sock *tp = tcp_sk(sk); 1566 1567 if (acked > 0) { 1568 /* One ACK acked hole. The rest eat duplicate ACKs. */ 1569 if (acked-1 >= tp->sacked_out) 1570 tp->sacked_out = 0; 1571 else 1572 tp->sacked_out -= acked-1; 1573 } 1574 tcp_check_reno_reordering(sk, acked); 1575 tcp_verify_left_out(tp); 1576} 1577 1578static inline void tcp_reset_reno_sack(struct tcp_sock *tp) 1579{ 1580 tp->sacked_out = 0; 1581} 1582 1583/* F-RTO can only be used if TCP has never retransmitted anything other than 1584 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here) 1585 */ 1586int tcp_use_frto(struct sock *sk) 1587{ 1588 const struct tcp_sock *tp = tcp_sk(sk); 1589 struct sk_buff *skb; 1590 1591 if (!sysctl_tcp_frto) 1592 return 0; 1593 1594 if (IsSackFrto()) 1595 return 1; 1596 1597 /* Avoid expensive walking of rexmit queue if possible */ 1598 if (tp->retrans_out > 1) 1599 return 0; 1600 1601 skb = tcp_write_queue_head(sk); 1602 skb = tcp_write_queue_next(sk, skb); /* Skips head */ 1603 tcp_for_write_queue_from(skb, sk) { 1604 if (skb == tcp_send_head(sk)) 1605 break; 1606 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS) 1607 return 0; 1608 /* Short-circuit when first non-SACKed skb has been checked */ 1609 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) 1610 break; 1611 } 1612 return 1; 1613} 1614 1615/* RTO occurred, but do not yet enter Loss state. Instead, defer RTO 1616 * recovery a bit and use heuristics in tcp_process_frto() to detect if 1617 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to 1618 * keep retrans_out counting accurate (with SACK F-RTO, other than head 1619 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS 1620 * bits are handled if the Loss state is really to be entered (in 1621 * tcp_enter_frto_loss). 1622 * 1623 * Do like tcp_enter_loss() would; when RTO expires the second time it 1624 * does: 1625 * "Reduce ssthresh if it has not yet been made inside this window." 1626 */ 1627void tcp_enter_frto(struct sock *sk) 1628{ 1629 const struct inet_connection_sock *icsk = inet_csk(sk); 1630 struct tcp_sock *tp = tcp_sk(sk); 1631 struct sk_buff *skb; 1632 1633 if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) || 1634 tp->snd_una == tp->high_seq || 1635 ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) && 1636 !icsk->icsk_retransmits)) { 1637 tp->prior_ssthresh = tcp_current_ssthresh(sk); 1638 /* Our state is too optimistic in ssthresh() call because cwnd 1639 * is not reduced until tcp_enter_frto_loss() when previous F-RTO 1640 * recovery has not yet completed. Pattern would be this: RTO, 1641 * Cumulative ACK, RTO (2xRTO for the same segment does not end 1642 * up here twice). 1643 * RFC4138 should be more specific on what to do, even though 1644 * RTO is quite unlikely to occur after the first Cumulative ACK 1645 * due to back-off and complexity of triggering events ... 1646 */ 1647 if (tp->frto_counter) { 1648 u32 stored_cwnd; 1649 stored_cwnd = tp->snd_cwnd; 1650 tp->snd_cwnd = 2; 1651 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 1652 tp->snd_cwnd = stored_cwnd; 1653 } else { 1654 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 1655 } 1656 /* ... in theory, cong.control module could do "any tricks" in 1657 * ssthresh(), which means that ca_state, lost bits and lost_out 1658 * counter would have to be faked before the call occurs. We 1659 * consider that too expensive, unlikely and hacky, so modules 1660 * using these in ssthresh() must deal these incompatibility 1661 * issues if they receives CA_EVENT_FRTO and frto_counter != 0 1662 */ 1663 tcp_ca_event(sk, CA_EVENT_FRTO); 1664 } 1665 1666 tp->undo_marker = tp->snd_una; 1667 tp->undo_retrans = 0; 1668 1669 skb = tcp_write_queue_head(sk); 1670 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) 1671 tp->undo_marker = 0; 1672 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 1673 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 1674 tp->retrans_out -= tcp_skb_pcount(skb); 1675 } 1676 tcp_verify_left_out(tp); 1677 1678 /* Earlier loss recovery underway (see RFC4138; Appendix B). 1679 * The last condition is necessary at least in tp->frto_counter case. 1680 */ 1681 if (IsSackFrto() && (tp->frto_counter || 1682 ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) && 1683 after(tp->high_seq, tp->snd_una)) { 1684 tp->frto_highmark = tp->high_seq; 1685 } else { 1686 tp->frto_highmark = tp->snd_nxt; 1687 } 1688 tcp_set_ca_state(sk, TCP_CA_Disorder); 1689 tp->high_seq = tp->snd_nxt; 1690 tp->frto_counter = 1; 1691} 1692 1693/* Enter Loss state after F-RTO was applied. Dupack arrived after RTO, 1694 * which indicates that we should follow the traditional RTO recovery, 1695 * i.e. mark everything lost and do go-back-N retransmission. 1696 */ 1697static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag) 1698{ 1699 struct tcp_sock *tp = tcp_sk(sk); 1700 struct sk_buff *skb; 1701 1702 tp->lost_out = 0; 1703 tp->retrans_out = 0; 1704 if (tcp_is_reno(tp)) 1705 tcp_reset_reno_sack(tp); 1706 1707 tcp_for_write_queue(skb, sk) { 1708 if (skb == tcp_send_head(sk)) 1709 break; 1710 /* 1711 * Count the retransmission made on RTO correctly (only when 1712 * waiting for the first ACK and did not get it)... 1713 */ 1714 if ((tp->frto_counter == 1) && !(flag&FLAG_DATA_ACKED)) { 1715 /* For some reason this R-bit might get cleared? */ 1716 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 1717 tp->retrans_out += tcp_skb_pcount(skb); 1718 /* ...enter this if branch just for the first segment */ 1719 flag |= FLAG_DATA_ACKED; 1720 } else { 1721 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) 1722 tp->undo_marker = 0; 1723 TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS); 1724 } 1725 1726 /* Don't lost mark skbs that were fwd transmitted after RTO */ 1727 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) && 1728 !after(TCP_SKB_CB(skb)->end_seq, tp->frto_highmark)) { 1729 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 1730 tp->lost_out += tcp_skb_pcount(skb); 1731 } 1732 } 1733 tcp_verify_left_out(tp); 1734 1735 tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments; 1736 tp->snd_cwnd_cnt = 0; 1737 tp->snd_cwnd_stamp = tcp_time_stamp; 1738 tp->frto_counter = 0; 1739 tp->bytes_acked = 0; 1740 1741 tp->reordering = min_t(unsigned int, tp->reordering, 1742 sysctl_tcp_reordering); 1743 tcp_set_ca_state(sk, TCP_CA_Loss); 1744 tp->high_seq = tp->frto_highmark; 1745 TCP_ECN_queue_cwr(tp); 1746 1747 tcp_clear_retrans_hints_partial(tp); 1748} 1749 1750static void tcp_clear_retrans_partial(struct tcp_sock *tp) 1751{ 1752 tp->retrans_out = 0; 1753 tp->lost_out = 0; 1754 1755 tp->undo_marker = 0; 1756 tp->undo_retrans = 0; 1757} 1758 1759void tcp_clear_retrans(struct tcp_sock *tp) 1760{ 1761 tcp_clear_retrans_partial(tp); 1762 1763 tp->fackets_out = 0; 1764 tp->sacked_out = 0; 1765} 1766 1767/* Enter Loss state. If "how" is not zero, forget all SACK information 1768 * and reset tags completely, otherwise preserve SACKs. If receiver 1769 * dropped its ofo queue, we will know this due to reneging detection. 1770 */ 1771void tcp_enter_loss(struct sock *sk, int how) 1772{ 1773 const struct inet_connection_sock *icsk = inet_csk(sk); 1774 struct tcp_sock *tp = tcp_sk(sk); 1775 struct sk_buff *skb; 1776 1777 /* Reduce ssthresh if it has not yet been made inside this window. */ 1778 if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq || 1779 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) { 1780 tp->prior_ssthresh = tcp_current_ssthresh(sk); 1781 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 1782 tcp_ca_event(sk, CA_EVENT_LOSS); 1783 } 1784 tp->snd_cwnd = 1; 1785 tp->snd_cwnd_cnt = 0; 1786 tp->snd_cwnd_stamp = tcp_time_stamp; 1787 1788 tp->bytes_acked = 0; 1789 tcp_clear_retrans_partial(tp); 1790 1791 if (tcp_is_reno(tp)) 1792 tcp_reset_reno_sack(tp); 1793 1794 if (!how) { 1795 /* Push undo marker, if it was plain RTO and nothing 1796 * was retransmitted. */ 1797 tp->undo_marker = tp->snd_una; 1798 tcp_clear_retrans_hints_partial(tp); 1799 } else { 1800 tp->sacked_out = 0; 1801 tp->fackets_out = 0; 1802 tcp_clear_all_retrans_hints(tp); 1803 } 1804 1805 tcp_for_write_queue(skb, sk) { 1806 if (skb == tcp_send_head(sk)) 1807 break; 1808 1809 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS) 1810 tp->undo_marker = 0; 1811 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED; 1812 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) { 1813 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED; 1814 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 1815 tp->lost_out += tcp_skb_pcount(skb); 1816 } 1817 } 1818 tcp_verify_left_out(tp); 1819 1820 tp->reordering = min_t(unsigned int, tp->reordering, 1821 sysctl_tcp_reordering); 1822 tcp_set_ca_state(sk, TCP_CA_Loss); 1823 tp->high_seq = tp->snd_nxt; 1824 TCP_ECN_queue_cwr(tp); 1825 /* Abort F-RTO algorithm if one is in progress */ 1826 tp->frto_counter = 0; 1827} 1828 1829static int tcp_check_sack_reneging(struct sock *sk) 1830{ 1831 struct sk_buff *skb; 1832 1833 /* If ACK arrived pointing to a remembered SACK, 1834 * it means that our remembered SACKs do not reflect 1835 * real state of receiver i.e. 1836 * receiver _host_ is heavily congested (or buggy). 1837 * Do processing similar to RTO timeout. 1838 */ 1839 if ((skb = tcp_write_queue_head(sk)) != NULL && 1840 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) { 1841 struct inet_connection_sock *icsk = inet_csk(sk); 1842 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING); 1843 1844 tcp_enter_loss(sk, 1); 1845 icsk->icsk_retransmits++; 1846 tcp_retransmit_skb(sk, tcp_write_queue_head(sk)); 1847 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 1848 icsk->icsk_rto, TCP_RTO_MAX); 1849 return 1; 1850 } 1851 return 0; 1852} 1853 1854static inline int tcp_fackets_out(struct tcp_sock *tp) 1855{ 1856 return tcp_is_reno(tp) ? tp->sacked_out+1 : tp->fackets_out; 1857} 1858 1859static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb) 1860{ 1861 return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto); 1862} 1863 1864static inline int tcp_head_timedout(struct sock *sk) 1865{ 1866 struct tcp_sock *tp = tcp_sk(sk); 1867 1868 return tp->packets_out && 1869 tcp_skb_timedout(sk, tcp_write_queue_head(sk)); 1870} 1871 1872/* Linux NewReno/SACK/FACK/ECN state machine. 1873 * -------------------------------------- 1874 * 1875 * "Open" Normal state, no dubious events, fast path. 1876 * "Disorder" In all the respects it is "Open", 1877 * but requires a bit more attention. It is entered when 1878 * we see some SACKs or dupacks. It is split of "Open" 1879 * mainly to move some processing from fast path to slow one. 1880 * "CWR" CWND was reduced due to some Congestion Notification event. 1881 * It can be ECN, ICMP source quench, local device congestion. 1882 * "Recovery" CWND was reduced, we are fast-retransmitting. 1883 * "Loss" CWND was reduced due to RTO timeout or SACK reneging. 1884 * 1885 * tcp_fastretrans_alert() is entered: 1886 * - each incoming ACK, if state is not "Open" 1887 * - when arrived ACK is unusual, namely: 1888 * * SACK 1889 * * Duplicate ACK. 1890 * * ECN ECE. 1891 * 1892 * Counting packets in flight is pretty simple. 1893 * 1894 * in_flight = packets_out - left_out + retrans_out 1895 * 1896 * packets_out is SND.NXT-SND.UNA counted in packets. 1897 * 1898 * retrans_out is number of retransmitted segments. 1899 * 1900 * left_out is number of segments left network, but not ACKed yet. 1901 * 1902 * left_out = sacked_out + lost_out 1903 * 1904 * sacked_out: Packets, which arrived to receiver out of order 1905 * and hence not ACKed. With SACKs this number is simply 1906 * amount of SACKed data. Even without SACKs 1907 * it is easy to give pretty reliable estimate of this number, 1908 * counting duplicate ACKs. 1909 * 1910 * lost_out: Packets lost by network. TCP has no explicit 1911 * "loss notification" feedback from network (for now). 1912 * It means that this number can be only _guessed_. 1913 * Actually, it is the heuristics to predict lossage that 1914 * distinguishes different algorithms. 1915 * 1916 * F.e. after RTO, when all the queue is considered as lost, 1917 * lost_out = packets_out and in_flight = retrans_out. 1918 * 1919 * Essentially, we have now two algorithms counting 1920 * lost packets. 1921 * 1922 * FACK: It is the simplest heuristics. As soon as we decided 1923 * that something is lost, we decide that _all_ not SACKed 1924 * packets until the most forward SACK are lost. I.e. 1925 * lost_out = fackets_out - sacked_out and left_out = fackets_out. 1926 * It is absolutely correct estimate, if network does not reorder 1927 * packets. And it loses any connection to reality when reordering 1928 * takes place. We use FACK by default until reordering 1929 * is suspected on the path to this destination. 1930 * 1931 * NewReno: when Recovery is entered, we assume that one segment 1932 * is lost (classic Reno). While we are in Recovery and 1933 * a partial ACK arrives, we assume that one more packet 1934 * is lost (NewReno). This heuristics are the same in NewReno 1935 * and SACK. 1936 * 1937 * Imagine, that's all! Forget about all this shamanism about CWND inflation 1938 * deflation etc. CWND is real congestion window, never inflated, changes 1939 * only according to classic VJ rules. 1940 * 1941 * Really tricky (and requiring careful tuning) part of algorithm 1942 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue(). 1943 * The first determines the moment _when_ we should reduce CWND and, 1944 * hence, slow down forward transmission. In fact, it determines the moment 1945 * when we decide that hole is caused by loss, rather than by a reorder. 1946 * 1947 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill 1948 * holes, caused by lost packets. 1949 * 1950 * And the most logically complicated part of algorithm is undo 1951 * heuristics. We detect false retransmits due to both too early 1952 * fast retransmit (reordering) and underestimated RTO, analyzing 1953 * timestamps and D-SACKs. When we detect that some segments were 1954 * retransmitted by mistake and CWND reduction was wrong, we undo 1955 * window reduction and abort recovery phase. This logic is hidden 1956 * inside several functions named tcp_try_undo_<something>. 1957 */ 1958 1959/* This function decides, when we should leave Disordered state 1960 * and enter Recovery phase, reducing congestion window. 1961 * 1962 * Main question: may we further continue forward transmission 1963 * with the same cwnd? 1964 */ 1965static int tcp_time_to_recover(struct sock *sk) 1966{ 1967 struct tcp_sock *tp = tcp_sk(sk); 1968 __u32 packets_out; 1969 1970 /* Do not perform any recovery during F-RTO algorithm */ 1971 if (tp->frto_counter) 1972 return 0; 1973 1974 /* Trick#1: The loss is proven. */ 1975 if (tp->lost_out) 1976 return 1; 1977 1978 /* Not-A-Trick#2 : Classic rule... */ 1979 if (tcp_fackets_out(tp) > tp->reordering) 1980 return 1; 1981 1982 /* Trick#3 : when we use RFC2988 timer restart, fast 1983 * retransmit can be triggered by timeout of queue head. 1984 */ 1985 if (tcp_head_timedout(sk)) 1986 return 1; 1987 1988 /* Trick#4: It is still not OK... But will it be useful to delay 1989 * recovery more? 1990 */ 1991 packets_out = tp->packets_out; 1992 if (packets_out <= tp->reordering && 1993 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) && 1994 !tcp_may_send_now(sk)) { 1995 /* We have nothing to send. This connection is limited 1996 * either by receiver window or by application. 1997 */ 1998 return 1; 1999 } 2000 2001 return 0; 2002} 2003 2004/* RFC: This is from the original, I doubt that this is necessary at all: 2005 * clear xmit_retrans hint if seq of this skb is beyond hint. How could we 2006 * retransmitted past LOST markings in the first place? I'm not fully sure 2007 * about undo and end of connection cases, which can cause R without L? 2008 */ 2009static void tcp_verify_retransmit_hint(struct tcp_sock *tp, 2010 struct sk_buff *skb) 2011{ 2012 if ((tp->retransmit_skb_hint != NULL) && 2013 before(TCP_SKB_CB(skb)->seq, 2014 TCP_SKB_CB(tp->retransmit_skb_hint)->seq)) 2015 tp->retransmit_skb_hint = NULL; 2016} 2017 2018/* Mark head of queue up as lost. */ 2019static void tcp_mark_head_lost(struct sock *sk, int packets) 2020{ 2021 struct tcp_sock *tp = tcp_sk(sk); 2022 struct sk_buff *skb; 2023 int cnt; 2024 2025 BUG_TRAP(packets <= tp->packets_out); 2026 if (tp->lost_skb_hint) { 2027 skb = tp->lost_skb_hint; 2028 cnt = tp->lost_cnt_hint; 2029 } else { 2030 skb = tcp_write_queue_head(sk); 2031 cnt = 0; 2032 } 2033 2034 tcp_for_write_queue_from(skb, sk) { 2035 if (skb == tcp_send_head(sk)) 2036 break; 2037 /* TODO: do this better */ 2038 /* this is not the most efficient way to do this... */ 2039 tp->lost_skb_hint = skb; 2040 tp->lost_cnt_hint = cnt; 2041 cnt += tcp_skb_pcount(skb); 2042 if (cnt > packets || after(TCP_SKB_CB(skb)->end_seq, tp->high_seq)) 2043 break; 2044 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_SACKED_ACKED|TCPCB_LOST))) { 2045 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 2046 tp->lost_out += tcp_skb_pcount(skb); 2047 tcp_verify_retransmit_hint(tp, skb); 2048 } 2049 } 2050 tcp_verify_left_out(tp); 2051} 2052 2053/* Account newly detected lost packet(s) */ 2054 2055static void tcp_update_scoreboard(struct sock *sk) 2056{ 2057 struct tcp_sock *tp = tcp_sk(sk); 2058 2059 if (tcp_is_fack(tp)) { 2060 int lost = tp->fackets_out - tp->reordering; 2061 if (lost <= 0) 2062 lost = 1; 2063 tcp_mark_head_lost(sk, lost); 2064 } else { 2065 tcp_mark_head_lost(sk, 1); 2066 } 2067 2068 /* New heuristics: it is possible only after we switched 2069 * to restart timer each time when something is ACKed. 2070 * Hence, we can detect timed out packets during fast 2071 * retransmit without falling to slow start. 2072 */ 2073 if (!tcp_is_reno(tp) && tcp_head_timedout(sk)) { 2074 struct sk_buff *skb; 2075 2076 skb = tp->scoreboard_skb_hint ? tp->scoreboard_skb_hint 2077 : tcp_write_queue_head(sk); 2078 2079 tcp_for_write_queue_from(skb, sk) { 2080 if (skb == tcp_send_head(sk)) 2081 break; 2082 if (!tcp_skb_timedout(sk, skb)) 2083 break; 2084 2085 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_SACKED_ACKED|TCPCB_LOST))) { 2086 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 2087 tp->lost_out += tcp_skb_pcount(skb); 2088 tcp_verify_retransmit_hint(tp, skb); 2089 } 2090 } 2091 2092 tp->scoreboard_skb_hint = skb; 2093 2094 tcp_verify_left_out(tp); 2095 } 2096} 2097 2098/* CWND moderation, preventing bursts due to too big ACKs 2099 * in dubious situations. 2100 */ 2101static inline void tcp_moderate_cwnd(struct tcp_sock *tp) 2102{ 2103 tp->snd_cwnd = min(tp->snd_cwnd, 2104 tcp_packets_in_flight(tp)+tcp_max_burst(tp)); 2105 tp->snd_cwnd_stamp = tcp_time_stamp; 2106} 2107 2108/* Lower bound on congestion window is slow start threshold 2109 * unless congestion avoidance choice decides to overide it. 2110 */ 2111static inline u32 tcp_cwnd_min(const struct sock *sk) 2112{ 2113 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 2114 2115 return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh; 2116} 2117 2118/* Decrease cwnd each second ack. */ 2119static void tcp_cwnd_down(struct sock *sk, int flag) 2120{ 2121 struct tcp_sock *tp = tcp_sk(sk); 2122 int decr = tp->snd_cwnd_cnt + 1; 2123 2124 if ((flag&(FLAG_ANY_PROGRESS|FLAG_DSACKING_ACK)) || 2125 (tcp_is_reno(tp) && !(flag&FLAG_NOT_DUP))) { 2126 tp->snd_cwnd_cnt = decr&1; 2127 decr >>= 1; 2128 2129 if (decr && tp->snd_cwnd > tcp_cwnd_min(sk)) 2130 tp->snd_cwnd -= decr; 2131 2132 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp)+1); 2133 tp->snd_cwnd_stamp = tcp_time_stamp; 2134 } 2135} 2136 2137/* Nothing was retransmitted or returned timestamp is less 2138 * than timestamp of the first retransmission. 2139 */ 2140static inline int tcp_packet_delayed(struct tcp_sock *tp) 2141{ 2142 return !tp->retrans_stamp || 2143 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 2144 (__s32)(tp->rx_opt.rcv_tsecr - tp->retrans_stamp) < 0); 2145} 2146 2147/* Undo procedures. */ 2148 2149#if FASTRETRANS_DEBUG > 1 2150static void DBGUNDO(struct sock *sk, const char *msg) 2151{ 2152 struct tcp_sock *tp = tcp_sk(sk); 2153 struct inet_sock *inet = inet_sk(sk); 2154 2155 printk(KERN_DEBUG "Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n", 2156 msg, 2157 NIPQUAD(inet->daddr), ntohs(inet->dport), 2158 tp->snd_cwnd, tcp_left_out(tp), 2159 tp->snd_ssthresh, tp->prior_ssthresh, 2160 tp->packets_out); 2161} 2162#else 2163#define DBGUNDO(x...) do { } while (0) 2164#endif 2165 2166static void tcp_undo_cwr(struct sock *sk, const int undo) 2167{ 2168 struct tcp_sock *tp = tcp_sk(sk); 2169 2170 if (tp->prior_ssthresh) { 2171 const struct inet_connection_sock *icsk = inet_csk(sk); 2172 2173 if (icsk->icsk_ca_ops->undo_cwnd) 2174 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk); 2175 else 2176 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh<<1); 2177 2178 if (undo && tp->prior_ssthresh > tp->snd_ssthresh) { 2179 tp->snd_ssthresh = tp->prior_ssthresh; 2180 TCP_ECN_withdraw_cwr(tp); 2181 } 2182 } else { 2183 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh); 2184 } 2185 tcp_moderate_cwnd(tp); 2186 tp->snd_cwnd_stamp = tcp_time_stamp; 2187 2188 /* There is something screwy going on with the retrans hints after 2189 an undo */ 2190 tcp_clear_all_retrans_hints(tp); 2191} 2192 2193static inline int tcp_may_undo(struct tcp_sock *tp) 2194{ 2195 return tp->undo_marker && 2196 (!tp->undo_retrans || tcp_packet_delayed(tp)); 2197} 2198 2199/* People celebrate: "We love our President!" */ 2200static int tcp_try_undo_recovery(struct sock *sk) 2201{ 2202 struct tcp_sock *tp = tcp_sk(sk); 2203 2204 if (tcp_may_undo(tp)) { 2205 /* Happy end! We did not retransmit anything 2206 * or our original transmission succeeded. 2207 */ 2208 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans"); 2209 tcp_undo_cwr(sk, 1); 2210 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss) 2211 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO); 2212 else 2213 NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO); 2214 tp->undo_marker = 0; 2215 } 2216 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) { 2217 /* Hold old state until something *above* high_seq 2218 * is ACKed. For Reno it is MUST to prevent false 2219 * fast retransmits (RFC2582). SACK TCP is safe. */ 2220 tcp_moderate_cwnd(tp); 2221 return 1; 2222 } 2223 tcp_set_ca_state(sk, TCP_CA_Open); 2224 return 0; 2225} 2226 2227/* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */ 2228static void tcp_try_undo_dsack(struct sock *sk) 2229{ 2230 struct tcp_sock *tp = tcp_sk(sk); 2231 2232 if (tp->undo_marker && !tp->undo_retrans) { 2233 DBGUNDO(sk, "D-SACK"); 2234 tcp_undo_cwr(sk, 1); 2235 tp->undo_marker = 0; 2236 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO); 2237 } 2238} 2239 2240/* Undo during fast recovery after partial ACK. */ 2241 2242static int tcp_try_undo_partial(struct sock *sk, int acked) 2243{ 2244 struct tcp_sock *tp = tcp_sk(sk); 2245 /* Partial ACK arrived. Force Hoe's retransmit. */ 2246 int failed = tcp_is_reno(tp) || tp->fackets_out>tp->reordering; 2247 2248 if (tcp_may_undo(tp)) { 2249 /* Plain luck! Hole if filled with delayed 2250 * packet, rather than with a retransmit. 2251 */ 2252 if (tp->retrans_out == 0) 2253 tp->retrans_stamp = 0; 2254 2255 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1); 2256 2257 DBGUNDO(sk, "Hoe"); 2258 tcp_undo_cwr(sk, 0); 2259 NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO); 2260 2261 /* So... Do not make Hoe's retransmit yet. 2262 * If the first packet was delayed, the rest 2263 * ones are most probably delayed as well. 2264 */ 2265 failed = 0; 2266 } 2267 return failed; 2268} 2269 2270/* Undo during loss recovery after partial ACK. */ 2271static int tcp_try_undo_loss(struct sock *sk) 2272{ 2273 struct tcp_sock *tp = tcp_sk(sk); 2274 2275 if (tcp_may_undo(tp)) { 2276 struct sk_buff *skb; 2277 tcp_for_write_queue(skb, sk) { 2278 if (skb == tcp_send_head(sk)) 2279 break; 2280 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; 2281 } 2282 2283 tcp_clear_all_retrans_hints(tp); 2284 2285 DBGUNDO(sk, "partial loss"); 2286 tp->lost_out = 0; 2287 tcp_undo_cwr(sk, 1); 2288 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO); 2289 inet_csk(sk)->icsk_retransmits = 0; 2290 tp->undo_marker = 0; 2291 if (tcp_is_sack(tp)) 2292 tcp_set_ca_state(sk, TCP_CA_Open); 2293 return 1; 2294 } 2295 return 0; 2296} 2297 2298static inline void tcp_complete_cwr(struct sock *sk) 2299{ 2300 struct tcp_sock *tp = tcp_sk(sk); 2301 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh); 2302 tp->snd_cwnd_stamp = tcp_time_stamp; 2303 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR); 2304} 2305 2306static void tcp_try_to_open(struct sock *sk, int flag) 2307{ 2308 struct tcp_sock *tp = tcp_sk(sk); 2309 2310 tcp_verify_left_out(tp); 2311 2312 if (tp->retrans_out == 0) 2313 tp->retrans_stamp = 0; 2314 2315 if (flag&FLAG_ECE) 2316 tcp_enter_cwr(sk, 1); 2317 2318 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) { 2319 int state = TCP_CA_Open; 2320 2321 if (tcp_left_out(tp) || tp->retrans_out || tp->undo_marker) 2322 state = TCP_CA_Disorder; 2323 2324 if (inet_csk(sk)->icsk_ca_state != state) { 2325 tcp_set_ca_state(sk, state); 2326 tp->high_seq = tp->snd_nxt; 2327 } 2328 tcp_moderate_cwnd(tp); 2329 } else { 2330 tcp_cwnd_down(sk, flag); 2331 } 2332} 2333 2334static void tcp_mtup_probe_failed(struct sock *sk) 2335{ 2336 struct inet_connection_sock *icsk = inet_csk(sk); 2337 2338 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1; 2339 icsk->icsk_mtup.probe_size = 0; 2340} 2341 2342static void tcp_mtup_probe_success(struct sock *sk, struct sk_buff *skb) 2343{ 2344 struct tcp_sock *tp = tcp_sk(sk); 2345 struct inet_connection_sock *icsk = inet_csk(sk); 2346 2347 /* FIXME: breaks with very large cwnd */ 2348 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2349 tp->snd_cwnd = tp->snd_cwnd * 2350 tcp_mss_to_mtu(sk, tp->mss_cache) / 2351 icsk->icsk_mtup.probe_size; 2352 tp->snd_cwnd_cnt = 0; 2353 tp->snd_cwnd_stamp = tcp_time_stamp; 2354 tp->rcv_ssthresh = tcp_current_ssthresh(sk); 2355 2356 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size; 2357 icsk->icsk_mtup.probe_size = 0; 2358 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 2359} 2360 2361 2362/* Process an event, which can update packets-in-flight not trivially. 2363 * Main goal of this function is to calculate new estimate for left_out, 2364 * taking into account both packets sitting in receiver's buffer and 2365 * packets lost by network. 2366 * 2367 * Besides that it does CWND reduction, when packet loss is detected 2368 * and changes state of machine. 2369 * 2370 * It does _not_ decide what to send, it is made in function 2371 * tcp_xmit_retransmit_queue(). 2372 */ 2373static void 2374tcp_fastretrans_alert(struct sock *sk, int pkts_acked, int flag) 2375{ 2376 struct inet_connection_sock *icsk = inet_csk(sk); 2377 struct tcp_sock *tp = tcp_sk(sk); 2378 int is_dupack = !(flag&(FLAG_SND_UNA_ADVANCED|FLAG_NOT_DUP)); 2379 int do_lost = is_dupack || ((flag&FLAG_DATA_SACKED) && 2380 (tp->fackets_out > tp->reordering)); 2381 2382 /* Some technical things: 2383 * 1. Reno does not count dupacks (sacked_out) automatically. */ 2384 if (!tp->packets_out) 2385 tp->sacked_out = 0; 2386 2387 if (WARN_ON(!tp->sacked_out && tp->fackets_out)) 2388 tp->fackets_out = 0; 2389 2390 /* Now state machine starts. 2391 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */ 2392 if (flag&FLAG_ECE) 2393 tp->prior_ssthresh = 0; 2394 2395 /* B. In all the states check for reneging SACKs. */ 2396 if (tp->sacked_out && tcp_check_sack_reneging(sk)) 2397 return; 2398 2399 /* C. Process data loss notification, provided it is valid. */ 2400 if ((flag&FLAG_DATA_LOST) && 2401 before(tp->snd_una, tp->high_seq) && 2402 icsk->icsk_ca_state != TCP_CA_Open && 2403 tp->fackets_out > tp->reordering) { 2404 tcp_mark_head_lost(sk, tp->fackets_out - tp->reordering); 2405 NET_INC_STATS_BH(LINUX_MIB_TCPLOSS); 2406 } 2407 2408 /* D. Check consistency of the current state. */ 2409 tcp_verify_left_out(tp); 2410 2411 /* E. Check state exit conditions. State can be terminated 2412 * when high_seq is ACKed. */ 2413 if (icsk->icsk_ca_state == TCP_CA_Open) { 2414 BUG_TRAP(tp->retrans_out == 0); 2415 tp->retrans_stamp = 0; 2416 } else if (!before(tp->snd_una, tp->high_seq)) { 2417 switch (icsk->icsk_ca_state) { 2418 case TCP_CA_Loss: 2419 icsk->icsk_retransmits = 0; 2420 if (tcp_try_undo_recovery(sk)) 2421 return; 2422 break; 2423 2424 case TCP_CA_CWR: 2425 /* CWR is to be held something *above* high_seq 2426 * is ACKed for CWR bit to reach receiver. */ 2427 if (tp->snd_una != tp->high_seq) { 2428 tcp_complete_cwr(sk); 2429 tcp_set_ca_state(sk, TCP_CA_Open); 2430 } 2431 break; 2432 2433 case TCP_CA_Disorder: 2434 tcp_try_undo_dsack(sk); 2435 if (!tp->undo_marker || 2436 /* For SACK case do not Open to allow to undo 2437 * catching for all duplicate ACKs. */ 2438 tcp_is_reno(tp) || tp->snd_una != tp->high_seq) { 2439 tp->undo_marker = 0; 2440 tcp_set_ca_state(sk, TCP_CA_Open); 2441 } 2442 break; 2443 2444 case TCP_CA_Recovery: 2445 if (tcp_is_reno(tp)) 2446 tcp_reset_reno_sack(tp); 2447 if (tcp_try_undo_recovery(sk)) 2448 return; 2449 tcp_complete_cwr(sk); 2450 break; 2451 } 2452 } 2453 2454 /* F. Process state. */ 2455 switch (icsk->icsk_ca_state) { 2456 case TCP_CA_Recovery: 2457 if (!(flag & FLAG_SND_UNA_ADVANCED)) { 2458 if (tcp_is_reno(tp) && is_dupack) 2459 tcp_add_reno_sack(sk); 2460 } else 2461 do_lost = tcp_try_undo_partial(sk, pkts_acked); 2462 break; 2463 case TCP_CA_Loss: 2464 if (flag&FLAG_DATA_ACKED) 2465 icsk->icsk_retransmits = 0; 2466 if (!tcp_try_undo_loss(sk)) { 2467 tcp_moderate_cwnd(tp); 2468 tcp_xmit_retransmit_queue(sk); 2469 return; 2470 } 2471 if (icsk->icsk_ca_state != TCP_CA_Open) 2472 return; 2473 /* Loss is undone; fall through to processing in Open state. */ 2474 default: 2475 if (tcp_is_reno(tp)) { 2476 if (flag & FLAG_SND_UNA_ADVANCED) 2477 tcp_reset_reno_sack(tp); 2478 if (is_dupack) 2479 tcp_add_reno_sack(sk); 2480 } 2481 2482 if (icsk->icsk_ca_state == TCP_CA_Disorder) 2483 tcp_try_undo_dsack(sk); 2484 2485 if (!tcp_time_to_recover(sk)) { 2486 tcp_try_to_open(sk, flag); 2487 return; 2488 } 2489 2490 /* MTU probe failure: don't reduce cwnd */ 2491 if (icsk->icsk_ca_state < TCP_CA_CWR && 2492 icsk->icsk_mtup.probe_size && 2493 tp->snd_una == tp->mtu_probe.probe_seq_start) { 2494 tcp_mtup_probe_failed(sk); 2495 /* Restores the reduction we did in tcp_mtup_probe() */ 2496 tp->snd_cwnd++; 2497 tcp_simple_retransmit(sk); 2498 return; 2499 } 2500 2501 /* Otherwise enter Recovery state */ 2502 2503 if (tcp_is_reno(tp)) 2504 NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY); 2505 else 2506 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY); 2507 2508 tp->high_seq = tp->snd_nxt; 2509 tp->prior_ssthresh = 0; 2510 tp->undo_marker = tp->snd_una; 2511 tp->undo_retrans = tp->retrans_out; 2512 2513 if (icsk->icsk_ca_state < TCP_CA_CWR) { 2514 if (!(flag&FLAG_ECE)) 2515 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2516 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 2517 TCP_ECN_queue_cwr(tp); 2518 } 2519 2520 tp->bytes_acked = 0; 2521 tp->snd_cwnd_cnt = 0; 2522 tcp_set_ca_state(sk, TCP_CA_Recovery); 2523 } 2524 2525 if (do_lost || tcp_head_timedout(sk)) 2526 tcp_update_scoreboard(sk); 2527 tcp_cwnd_down(sk, flag); 2528 tcp_xmit_retransmit_queue(sk); 2529} 2530 2531/* Read draft-ietf-tcplw-high-performance before mucking 2532 * with this code. (Supersedes RFC1323) 2533 */ 2534static void tcp_ack_saw_tstamp(struct sock *sk, int flag) 2535{ 2536 /* RTTM Rule: A TSecr value received in a segment is used to 2537 * update the averaged RTT measurement only if the segment 2538 * acknowledges some new data, i.e., only if it advances the 2539 * left edge of the send window. 2540 * 2541 * See draft-ietf-tcplw-high-performance-00, section 3.3. 2542 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru> 2543 * 2544 * Changed: reset backoff as soon as we see the first valid sample. 2545 * If we do not, we get strongly overestimated rto. With timestamps 2546 * samples are accepted even from very old segments: f.e., when rtt=1 2547 * increases to 8, we retransmit 5 times and after 8 seconds delayed 2548 * answer arrives rto becomes 120 seconds! If at least one of segments 2549 * in window is lost... Voila. --ANK (010210) 2550 */ 2551 struct tcp_sock *tp = tcp_sk(sk); 2552 const __u32 seq_rtt = tcp_time_stamp - tp->rx_opt.rcv_tsecr; 2553 tcp_rtt_estimator(sk, seq_rtt); 2554 tcp_set_rto(sk); 2555 inet_csk(sk)->icsk_backoff = 0; 2556 tcp_bound_rto(sk); 2557} 2558 2559static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag) 2560{ 2561 /* We don't have a timestamp. Can only use 2562 * packets that are not retransmitted to determine 2563 * rtt estimates. Also, we must not reset the 2564 * backoff for rto until we get a non-retransmitted 2565 * packet. This allows us to deal with a situation 2566 * where the network delay has increased suddenly. 2567 * I.e. Karn's algorithm. (SIGCOMM '87, p5.) 2568 */ 2569 2570 if (flag & FLAG_RETRANS_DATA_ACKED) 2571 return; 2572 2573 tcp_rtt_estimator(sk, seq_rtt); 2574 tcp_set_rto(sk); 2575 inet_csk(sk)->icsk_backoff = 0; 2576 tcp_bound_rto(sk); 2577} 2578 2579static inline void tcp_ack_update_rtt(struct sock *sk, const int flag, 2580 const s32 seq_rtt) 2581{ 2582 const struct tcp_sock *tp = tcp_sk(sk); 2583 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */ 2584 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr) 2585 tcp_ack_saw_tstamp(sk, flag); 2586 else if (seq_rtt >= 0) 2587 tcp_ack_no_tstamp(sk, seq_rtt, flag); 2588} 2589 2590static void tcp_cong_avoid(struct sock *sk, u32 ack, 2591 u32 in_flight, int good) 2592{ 2593 const struct inet_connection_sock *icsk = inet_csk(sk); 2594 icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight, good); 2595 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp; 2596} 2597 2598/* Restart timer after forward progress on connection. 2599 * RFC2988 recommends to restart timer to now+rto. 2600 */ 2601static void tcp_rearm_rto(struct sock *sk) 2602{ 2603 struct tcp_sock *tp = tcp_sk(sk); 2604 2605 if (!tp->packets_out) { 2606 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS); 2607 } else { 2608 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 2609 } 2610} 2611 2612/* If we get here, the whole TSO packet has not been acked. */ 2613static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb) 2614{ 2615 struct tcp_sock *tp = tcp_sk(sk); 2616 u32 packets_acked; 2617 2618 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)); 2619 2620 packets_acked = tcp_skb_pcount(skb); 2621 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 2622 return 0; 2623 packets_acked -= tcp_skb_pcount(skb); 2624 2625 if (packets_acked) { 2626 BUG_ON(tcp_skb_pcount(skb) == 0); 2627 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)); 2628 } 2629 2630 return packets_acked; 2631} 2632 2633/* Remove acknowledged frames from the retransmission queue. If our packet 2634 * is before the ack sequence we can discard it as it's confirmed to have 2635 * arrived at the other end. 2636 */ 2637static int tcp_clean_rtx_queue(struct sock *sk, s32 *seq_rtt_p) 2638{ 2639 struct tcp_sock *tp = tcp_sk(sk); 2640 const struct inet_connection_sock *icsk = inet_csk(sk); 2641 struct sk_buff *skb; 2642 u32 now = tcp_time_stamp; 2643 int fully_acked = 1; 2644 int flag = 0; 2645 int prior_packets = tp->packets_out; 2646 s32 seq_rtt = -1; 2647 ktime_t last_ackt = net_invalid_timestamp(); 2648 2649 while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) { 2650 struct tcp_skb_cb *scb = TCP_SKB_CB(skb); 2651 u32 end_seq; 2652 u32 packets_acked; 2653 u8 sacked = scb->sacked; 2654 2655 if (after(scb->end_seq, tp->snd_una)) { 2656 if (tcp_skb_pcount(skb) == 1 || 2657 !after(tp->snd_una, scb->seq)) 2658 break; 2659 2660 packets_acked = tcp_tso_acked(sk, skb); 2661 if (!packets_acked) 2662 break; 2663 2664 fully_acked = 0; 2665 end_seq = tp->snd_una; 2666 } else { 2667 packets_acked = tcp_skb_pcount(skb); 2668 end_seq = scb->end_seq; 2669 } 2670 2671 /* MTU probing checks */ 2672 if (fully_acked && icsk->icsk_mtup.probe_size && 2673 !after(tp->mtu_probe.probe_seq_end, scb->end_seq)) { 2674 tcp_mtup_probe_success(sk, skb); 2675 } 2676 2677 if (sacked) { 2678 if (sacked & TCPCB_RETRANS) { 2679 if (sacked & TCPCB_SACKED_RETRANS) 2680 tp->retrans_out -= packets_acked; 2681 flag |= FLAG_RETRANS_DATA_ACKED; 2682 seq_rtt = -1; 2683 if ((flag & FLAG_DATA_ACKED) || 2684 (packets_acked > 1)) 2685 flag |= FLAG_NONHEAD_RETRANS_ACKED; 2686 } else if (seq_rtt < 0) { 2687 seq_rtt = now - scb->when; 2688 if (fully_acked) 2689 last_ackt = skb->tstamp; 2690 } 2691 2692 if (sacked & TCPCB_SACKED_ACKED) 2693 tp->sacked_out -= packets_acked; 2694 if (sacked & TCPCB_LOST) 2695 tp->lost_out -= packets_acked; 2696 2697 if ((sacked & TCPCB_URG) && tp->urg_mode && 2698 !before(end_seq, tp->snd_up)) 2699 tp->urg_mode = 0; 2700 } else if (seq_rtt < 0) { 2701 seq_rtt = now - scb->when; 2702 if (fully_acked) 2703 last_ackt = skb->tstamp; 2704 } 2705 tp->packets_out -= packets_acked; 2706 2707 /* Initial outgoing SYN's get put onto the write_queue 2708 * just like anything else we transmit. It is not 2709 * true data, and if we misinform our callers that 2710 * this ACK acks real data, we will erroneously exit 2711 * connection startup slow start one packet too 2712 * quickly. This is severely frowned upon behavior. 2713 */ 2714 if (!(scb->flags & TCPCB_FLAG_SYN)) { 2715 flag |= FLAG_DATA_ACKED; 2716 } else { 2717 flag |= FLAG_SYN_ACKED; 2718 tp->retrans_stamp = 0; 2719 } 2720 2721 if (!fully_acked) 2722 break; 2723 2724 tcp_unlink_write_queue(skb, sk); 2725 sk_stream_free_skb(sk, skb); 2726 tcp_clear_all_retrans_hints(tp); 2727 } 2728 2729 if (flag & FLAG_ACKED) { 2730 u32 pkts_acked = prior_packets - tp->packets_out; 2731 const struct tcp_congestion_ops *ca_ops 2732 = inet_csk(sk)->icsk_ca_ops; 2733 2734 tcp_ack_update_rtt(sk, flag, seq_rtt); 2735 tcp_rearm_rto(sk); 2736 2737 tp->fackets_out -= min(pkts_acked, tp->fackets_out); 2738 /* hint's skb might be NULL but we don't need to care */ 2739 tp->fastpath_cnt_hint -= min_t(u32, pkts_acked, 2740 tp->fastpath_cnt_hint); 2741 if (tcp_is_reno(tp)) 2742 tcp_remove_reno_sacks(sk, pkts_acked); 2743 2744 if (ca_ops->pkts_acked) { 2745 s32 rtt_us = -1; 2746 2747 /* Is the ACK triggering packet unambiguous? */ 2748 if (!(flag & FLAG_RETRANS_DATA_ACKED)) { 2749 /* High resolution needed and available? */ 2750 if (ca_ops->flags & TCP_CONG_RTT_STAMP && 2751 !ktime_equal(last_ackt, 2752 net_invalid_timestamp())) 2753 rtt_us = ktime_us_delta(ktime_get_real(), 2754 last_ackt); 2755 else if (seq_rtt > 0) 2756 rtt_us = jiffies_to_usecs(seq_rtt); 2757 } 2758 2759 ca_ops->pkts_acked(sk, pkts_acked, rtt_us); 2760 } 2761 } 2762 2763#if FASTRETRANS_DEBUG > 0 2764 BUG_TRAP((int)tp->sacked_out >= 0); 2765 BUG_TRAP((int)tp->lost_out >= 0); 2766 BUG_TRAP((int)tp->retrans_out >= 0); 2767 if (!tp->packets_out && tcp_is_sack(tp)) { 2768 icsk = inet_csk(sk); 2769 if (tp->lost_out) { 2770 printk(KERN_DEBUG "Leak l=%u %d\n", 2771 tp->lost_out, icsk->icsk_ca_state); 2772 tp->lost_out = 0; 2773 } 2774 if (tp->sacked_out) { 2775 printk(KERN_DEBUG "Leak s=%u %d\n", 2776 tp->sacked_out, icsk->icsk_ca_state); 2777 tp->sacked_out = 0; 2778 } 2779 if (tp->retrans_out) { 2780 printk(KERN_DEBUG "Leak r=%u %d\n", 2781 tp->retrans_out, icsk->icsk_ca_state); 2782 tp->retrans_out = 0; 2783 } 2784 } 2785#endif 2786 *seq_rtt_p = seq_rtt; 2787 return flag; 2788} 2789 2790static void tcp_ack_probe(struct sock *sk) 2791{ 2792 const struct tcp_sock *tp = tcp_sk(sk); 2793 struct inet_connection_sock *icsk = inet_csk(sk); 2794 2795 /* Was it a usable window open? */ 2796 2797 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, 2798 tp->snd_una + tp->snd_wnd)) { 2799 icsk->icsk_backoff = 0; 2800 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0); 2801 /* Socket must be waked up by subsequent tcp_data_snd_check(). 2802 * This function is not for random using! 2803 */ 2804 } else { 2805 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 2806 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), 2807 TCP_RTO_MAX); 2808 } 2809} 2810 2811static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag) 2812{ 2813 return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) || 2814 inet_csk(sk)->icsk_ca_state != TCP_CA_Open); 2815} 2816 2817static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag) 2818{ 2819 const struct tcp_sock *tp = tcp_sk(sk); 2820 return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) && 2821 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR)); 2822} 2823 2824/* Check that window update is acceptable. 2825 * The function assumes that snd_una<=ack<=snd_next. 2826 */ 2827static inline int tcp_may_update_window(const struct tcp_sock *tp, const u32 ack, 2828 const u32 ack_seq, const u32 nwin) 2829{ 2830 return (after(ack, tp->snd_una) || 2831 after(ack_seq, tp->snd_wl1) || 2832 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd)); 2833} 2834 2835/* Update our send window. 2836 * 2837 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2 2838 * and in FreeBSD. NetBSD's one is even worse.) is wrong. 2839 */ 2840static int tcp_ack_update_window(struct sock *sk, struct sk_buff *skb, u32 ack, 2841 u32 ack_seq) 2842{ 2843 struct tcp_sock *tp = tcp_sk(sk); 2844 int flag = 0; 2845 u32 nwin = ntohs(tcp_hdr(skb)->window); 2846 2847 if (likely(!tcp_hdr(skb)->syn)) 2848 nwin <<= tp->rx_opt.snd_wscale; 2849 2850 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) { 2851 flag |= FLAG_WIN_UPDATE; 2852 tcp_update_wl(tp, ack, ack_seq); 2853 2854 if (tp->snd_wnd != nwin) { 2855 tp->snd_wnd = nwin; 2856 2857 /* Note, it is the only place, where 2858 * fast path is recovered for sending TCP. 2859 */ 2860 tp->pred_flags = 0; 2861 tcp_fast_path_check(sk); 2862 2863 if (nwin > tp->max_window) { 2864 tp->max_window = nwin; 2865 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie); 2866 } 2867 } 2868 } 2869 2870 tp->snd_una = ack; 2871 2872 return flag; 2873} 2874 2875/* A very conservative spurious RTO response algorithm: reduce cwnd and 2876 * continue in congestion avoidance. 2877 */ 2878static void tcp_conservative_spur_to_response(struct tcp_sock *tp) 2879{ 2880 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh); 2881 tp->snd_cwnd_cnt = 0; 2882 tp->bytes_acked = 0; 2883 TCP_ECN_queue_cwr(tp); 2884 tcp_moderate_cwnd(tp); 2885} 2886 2887/* A conservative spurious RTO response algorithm: reduce cwnd using 2888 * rate halving and continue in congestion avoidance. 2889 */ 2890static void tcp_ratehalving_spur_to_response(struct sock *sk) 2891{ 2892 tcp_enter_cwr(sk, 0); 2893} 2894 2895static void tcp_undo_spur_to_response(struct sock *sk, int flag) 2896{ 2897 if (flag&FLAG_ECE) 2898 tcp_ratehalving_spur_to_response(sk); 2899 else 2900 tcp_undo_cwr(sk, 1); 2901} 2902 2903/* F-RTO spurious RTO detection algorithm (RFC4138) 2904 * 2905 * F-RTO affects during two new ACKs following RTO (well, almost, see inline 2906 * comments). State (ACK number) is kept in frto_counter. When ACK advances 2907 * window (but not to or beyond highest sequence sent before RTO): 2908 * On First ACK, send two new segments out. 2909 * On Second ACK, RTO was likely spurious. Do spurious response (response 2910 * algorithm is not part of the F-RTO detection algorithm 2911 * given in RFC4138 but can be selected separately). 2912 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss 2913 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding 2914 * of Nagle, this is done using frto_counter states 2 and 3, when a new data 2915 * segment of any size sent during F-RTO, state 2 is upgraded to 3. 2916 * 2917 * Rationale: if the RTO was spurious, new ACKs should arrive from the 2918 * original window even after we transmit two new data segments. 2919 * 2920 * SACK version: 2921 * on first step, wait until first cumulative ACK arrives, then move to 2922 * the second step. In second step, the next ACK decides. 2923 * 2924 * F-RTO is implemented (mainly) in four functions: 2925 * - tcp_use_frto() is used to determine if TCP is can use F-RTO 2926 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is 2927 * called when tcp_use_frto() showed green light 2928 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm 2929 * - tcp_enter_frto_loss() is called if there is not enough evidence 2930 * to prove that the RTO is indeed spurious. It transfers the control 2931 * from F-RTO to the conventional RTO recovery 2932 */ 2933static int tcp_process_frto(struct sock *sk, int flag) 2934{ 2935 struct tcp_sock *tp = tcp_sk(sk); 2936 2937 tcp_verify_left_out(tp); 2938 2939 /* Duplicate the behavior from Loss state (fastretrans_alert) */ 2940 if (flag&FLAG_DATA_ACKED) 2941 inet_csk(sk)->icsk_retransmits = 0; 2942 2943 if ((flag & FLAG_NONHEAD_RETRANS_ACKED) || 2944 ((tp->frto_counter >= 2) && (flag & FLAG_RETRANS_DATA_ACKED))) 2945 tp->undo_marker = 0; 2946 2947 if (!before(tp->snd_una, tp->frto_highmark)) { 2948 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag); 2949 return 1; 2950 } 2951 2952 if (!IsSackFrto() || tcp_is_reno(tp)) { 2953 /* RFC4138 shortcoming in step 2; should also have case c): 2954 * ACK isn't duplicate nor advances window, e.g., opposite dir 2955 * data, winupdate 2956 */ 2957 if (!(flag&FLAG_ANY_PROGRESS) && (flag&FLAG_NOT_DUP)) 2958 return 1; 2959 2960 if (!(flag&FLAG_DATA_ACKED)) { 2961 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3), 2962 flag); 2963 return 1; 2964 } 2965 } else { 2966 if (!(flag&FLAG_DATA_ACKED) && (tp->frto_counter == 1)) { 2967 /* Prevent sending of new data. */ 2968 tp->snd_cwnd = min(tp->snd_cwnd, 2969 tcp_packets_in_flight(tp)); 2970 return 1; 2971 } 2972 2973 if ((tp->frto_counter >= 2) && 2974 (!(flag&FLAG_FORWARD_PROGRESS) || 2975 ((flag&FLAG_DATA_SACKED) && !(flag&FLAG_ONLY_ORIG_SACKED)))) { 2976 /* RFC4138 shortcoming (see comment above) */ 2977 if (!(flag&FLAG_FORWARD_PROGRESS) && (flag&FLAG_NOT_DUP)) 2978 return 1; 2979 2980 tcp_enter_frto_loss(sk, 3, flag); 2981 return 1; 2982 } 2983 } 2984 2985 if (tp->frto_counter == 1) { 2986 /* Sending of the next skb must be allowed or no F-RTO */ 2987 if (!tcp_send_head(sk) || 2988 after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, 2989 tp->snd_una + tp->snd_wnd)) { 2990 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), 2991 flag); 2992 return 1; 2993 } 2994 2995 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2; 2996 tp->frto_counter = 2; 2997 return 1; 2998 } else { 2999 switch (sysctl_tcp_frto_response) { 3000 case 2: 3001 tcp_undo_spur_to_response(sk, flag); 3002 break; 3003 case 1: 3004 tcp_conservative_spur_to_response(tp); 3005 break; 3006 default: 3007 tcp_ratehalving_spur_to_response(sk); 3008 break; 3009 } 3010 tp->frto_counter = 0; 3011 tp->undo_marker = 0; 3012 NET_INC_STATS_BH(LINUX_MIB_TCPSPURIOUSRTOS); 3013 } 3014 return 0; 3015} 3016 3017/* This routine deals with incoming acks, but not outgoing ones. */ 3018static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag) 3019{ 3020 struct inet_connection_sock *icsk = inet_csk(sk); 3021 struct tcp_sock *tp = tcp_sk(sk); 3022 u32 prior_snd_una = tp->snd_una; 3023 u32 ack_seq = TCP_SKB_CB(skb)->seq; 3024 u32 ack = TCP_SKB_CB(skb)->ack_seq; 3025 u32 prior_in_flight; 3026 s32 seq_rtt; 3027 int prior_packets; 3028 int frto_cwnd = 0; 3029 3030 /* If the ack is newer than sent or older than previous acks 3031 * then we can probably ignore it. 3032 */ 3033 if (after(ack, tp->snd_nxt)) 3034 goto uninteresting_ack; 3035 3036 if (before(ack, prior_snd_una)) 3037 goto old_ack; 3038 3039 if (after(ack, prior_snd_una)) 3040 flag |= FLAG_SND_UNA_ADVANCED; 3041 3042 if (sysctl_tcp_abc) { 3043 if (icsk->icsk_ca_state < TCP_CA_CWR) 3044 tp->bytes_acked += ack - prior_snd_una; 3045 else if (icsk->icsk_ca_state == TCP_CA_Loss) 3046 /* we assume just one segment left network */ 3047 tp->bytes_acked += min(ack - prior_snd_una, tp->mss_cache); 3048 } 3049 3050 if (!(flag&FLAG_SLOWPATH) && after(ack, prior_snd_una)) { 3051 /* Window is constant, pure forward advance. 3052 * No more checks are required. 3053 * Note, we use the fact that SND.UNA>=SND.WL2. 3054 */ 3055 tcp_update_wl(tp, ack, ack_seq); 3056 tp->snd_una = ack; 3057 flag |= FLAG_WIN_UPDATE; 3058 3059 tcp_ca_event(sk, CA_EVENT_FAST_ACK); 3060 3061 NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS); 3062 } else { 3063 if (ack_seq != TCP_SKB_CB(skb)->end_seq) 3064 flag |= FLAG_DATA; 3065 else 3066 NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS); 3067 3068 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq); 3069 3070 if (TCP_SKB_CB(skb)->sacked) 3071 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una); 3072 3073 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb))) 3074 flag |= FLAG_ECE; 3075 3076 tcp_ca_event(sk, CA_EVENT_SLOW_ACK); 3077 } 3078 3079 /* We passed data and got it acked, remove any soft error 3080 * log. Something worked... 3081 */ 3082 sk->sk_err_soft = 0; 3083 tp->rcv_tstamp = tcp_time_stamp; 3084 prior_packets = tp->packets_out; 3085 if (!prior_packets) 3086 goto no_queue; 3087 3088 prior_in_flight = tcp_packets_in_flight(tp); 3089 3090 /* See if we can take anything off of the retransmit queue. */ 3091 flag |= tcp_clean_rtx_queue(sk, &seq_rtt); 3092 3093 /* Guarantee sacktag reordering detection against wrap-arounds */ 3094 if (before(tp->frto_highmark, tp->snd_una)) 3095 tp->frto_highmark = 0; 3096 if (tp->frto_counter) 3097 frto_cwnd = tcp_process_frto(sk, flag); 3098 3099 if (tcp_ack_is_dubious(sk, flag)) { 3100 /* Advance CWND, if state allows this. */ 3101 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd && 3102 tcp_may_raise_cwnd(sk, flag)) 3103 tcp_cong_avoid(sk, ack, prior_in_flight, 0); 3104 tcp_fastretrans_alert(sk, prior_packets - tp->packets_out, flag); 3105 } else { 3106 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd) 3107 tcp_cong_avoid(sk, ack, prior_in_flight, 1); 3108 } 3109 3110 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag&FLAG_NOT_DUP)) 3111 dst_confirm(sk->sk_dst_cache); 3112 3113 return 1; 3114 3115no_queue: 3116 icsk->icsk_probes_out = 0; 3117 3118 /* If this ack opens up a zero window, clear backoff. It was 3119 * being used to time the probes, and is probably far higher than 3120 * it needs to be for normal retransmission. 3121 */ 3122 if (tcp_send_head(sk)) 3123 tcp_ack_probe(sk); 3124 return 1; 3125 3126old_ack: 3127 if (TCP_SKB_CB(skb)->sacked) 3128 tcp_sacktag_write_queue(sk, skb, prior_snd_una); 3129 3130uninteresting_ack: 3131 SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt); 3132 return 0; 3133} 3134 3135 3136/* Look for tcp options. Normally only called on SYN and SYNACK packets. 3137 * But, this can also be called on packets in the established flow when 3138 * the fast version below fails. 3139 */ 3140void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx, int estab) 3141{ 3142 unsigned char *ptr; 3143 struct tcphdr *th = tcp_hdr(skb); 3144 int length=(th->doff*4)-sizeof(struct tcphdr); 3145 3146 ptr = (unsigned char *)(th + 1); 3147 opt_rx->saw_tstamp = 0; 3148 3149 while (length > 0) { 3150 int opcode=*ptr++; 3151 int opsize; 3152 3153 switch (opcode) { 3154 case TCPOPT_EOL: 3155 return; 3156 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */ 3157 length--; 3158 continue; 3159 default: 3160 opsize=*ptr++; 3161 if (opsize < 2) /* "silly options" */ 3162 return; 3163 if (opsize > length) 3164 return; /* don't parse partial options */ 3165 switch (opcode) { 3166 case TCPOPT_MSS: 3167 if (opsize==TCPOLEN_MSS && th->syn && !estab) { 3168 u16 in_mss = ntohs(get_unaligned((__be16 *)ptr)); 3169 if (in_mss) { 3170 if (opt_rx->user_mss && opt_rx->user_mss < in_mss) 3171 in_mss = opt_rx->user_mss; 3172 opt_rx->mss_clamp = in_mss; 3173 } 3174 } 3175 break; 3176 case TCPOPT_WINDOW: 3177 if (opsize==TCPOLEN_WINDOW && th->syn && !estab) 3178 if (sysctl_tcp_window_scaling) { 3179 __u8 snd_wscale = *(__u8 *) ptr; 3180 opt_rx->wscale_ok = 1; 3181 if (snd_wscale > 14) { 3182 if (net_ratelimit()) 3183 printk(KERN_INFO "tcp_parse_options: Illegal window " 3184 "scaling value %d >14 received.\n", 3185 snd_wscale); 3186 snd_wscale = 14; 3187 } 3188 opt_rx->snd_wscale = snd_wscale; 3189 } 3190 break; 3191 case TCPOPT_TIMESTAMP: 3192 if (opsize==TCPOLEN_TIMESTAMP) { 3193 if ((estab && opt_rx->tstamp_ok) || 3194 (!estab && sysctl_tcp_timestamps)) { 3195 opt_rx->saw_tstamp = 1; 3196 opt_rx->rcv_tsval = ntohl(get_unaligned((__be32 *)ptr)); 3197 opt_rx->rcv_tsecr = ntohl(get_unaligned((__be32 *)(ptr+4))); 3198 } 3199 } 3200 break; 3201 case TCPOPT_SACK_PERM: 3202 if (opsize==TCPOLEN_SACK_PERM && th->syn && !estab) { 3203 if (sysctl_tcp_sack) { 3204 opt_rx->sack_ok = 1; 3205 tcp_sack_reset(opt_rx); 3206 } 3207 } 3208 break; 3209 3210 case TCPOPT_SACK: 3211 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) && 3212 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) && 3213 opt_rx->sack_ok) { 3214 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th; 3215 } 3216 break; 3217#ifdef CONFIG_TCP_MD5SIG 3218 case TCPOPT_MD5SIG: 3219 /* 3220 * The MD5 Hash has already been 3221 * checked (see tcp_v{4,6}_do_rcv()). 3222 */ 3223 break; 3224#endif 3225 } 3226 3227 ptr+=opsize-2; 3228 length-=opsize; 3229 } 3230 } 3231} 3232 3233/* Fast parse options. This hopes to only see timestamps. 3234 * If it is wrong it falls back on tcp_parse_options(). 3235 */ 3236static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th, 3237 struct tcp_sock *tp) 3238{ 3239 if (th->doff == sizeof(struct tcphdr)>>2) { 3240 tp->rx_opt.saw_tstamp = 0; 3241 return 0; 3242 } else if (tp->rx_opt.tstamp_ok && 3243 th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) { 3244 __be32 *ptr = (__be32 *)(th + 1); 3245 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) 3246 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) { 3247 tp->rx_opt.saw_tstamp = 1; 3248 ++ptr; 3249 tp->rx_opt.rcv_tsval = ntohl(*ptr); 3250 ++ptr; 3251 tp->rx_opt.rcv_tsecr = ntohl(*ptr); 3252 return 1; 3253 } 3254 } 3255 tcp_parse_options(skb, &tp->rx_opt, 1); 3256 return 1; 3257} 3258 3259static inline void tcp_store_ts_recent(struct tcp_sock *tp) 3260{ 3261 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval; 3262 tp->rx_opt.ts_recent_stamp = get_seconds(); 3263} 3264 3265static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq) 3266{ 3267 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) { 3268 /* PAWS bug workaround wrt. ACK frames, the PAWS discard 3269 * extra check below makes sure this can only happen 3270 * for pure ACK frames. -DaveM 3271 * 3272 * Not only, also it occurs for expired timestamps. 3273 */ 3274 3275 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) >= 0 || 3276 get_seconds() >= tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS) 3277 tcp_store_ts_recent(tp); 3278 } 3279} 3280 3281/* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM 3282 * 3283 * It is not fatal. If this ACK does _not_ change critical state (seqs, window) 3284 * it can pass through stack. So, the following predicate verifies that 3285 * this segment is not used for anything but congestion avoidance or 3286 * fast retransmit. Moreover, we even are able to eliminate most of such 3287 * second order effects, if we apply some small "replay" window (~RTO) 3288 * to timestamp space. 3289 * 3290 * All these measures still do not guarantee that we reject wrapped ACKs 3291 * on networks with high bandwidth, when sequence space is recycled fastly, 3292 * but it guarantees that such events will be very rare and do not affect 3293 * connection seriously. This doesn't look nice, but alas, PAWS is really 3294 * buggy extension. 3295 * 3296 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC 3297 * states that events when retransmit arrives after original data are rare. 3298 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is 3299 * the biggest problem on large power networks even with minor reordering. 3300 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe 3301 * up to bandwidth of 18Gigabit/sec. 8) ] 3302 */ 3303 3304static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb) 3305{ 3306 struct tcp_sock *tp = tcp_sk(sk); 3307 struct tcphdr *th = tcp_hdr(skb); 3308 u32 seq = TCP_SKB_CB(skb)->seq; 3309 u32 ack = TCP_SKB_CB(skb)->ack_seq; 3310 3311 return (/* 1. Pure ACK with correct sequence number. */ 3312 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) && 3313 3314 /* 2. ... and duplicate ACK. */ 3315 ack == tp->snd_una && 3316 3317 /* 3. ... and does not update window. */ 3318 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) && 3319 3320 /* 4. ... and sits in replay window. */ 3321 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ); 3322} 3323 3324static inline int tcp_paws_discard(const struct sock *sk, const struct sk_buff *skb) 3325{ 3326 const struct tcp_sock *tp = tcp_sk(sk); 3327 return ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > TCP_PAWS_WINDOW && 3328 get_seconds() < tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS && 3329 !tcp_disordered_ack(sk, skb)); 3330} 3331 3332/* Check segment sequence number for validity. 3333 * 3334 * Segment controls are considered valid, if the segment 3335 * fits to the window after truncation to the window. Acceptability 3336 * of data (and SYN, FIN, of course) is checked separately. 3337 * See tcp_data_queue(), for example. 3338 * 3339 * Also, controls (RST is main one) are accepted using RCV.WUP instead 3340 * of RCV.NXT. Peer still did not advance his SND.UNA when we 3341 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP. 3342 * (borrowed from freebsd) 3343 */ 3344 3345static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq) 3346{ 3347 return !before(end_seq, tp->rcv_wup) && 3348 !after(seq, tp->rcv_nxt + tcp_receive_window(tp)); 3349} 3350 3351/* When we get a reset we do this. */ 3352static void tcp_reset(struct sock *sk) 3353{ 3354 /* We want the right error as BSD sees it (and indeed as we do). */ 3355 switch (sk->sk_state) { 3356 case TCP_SYN_SENT: 3357 sk->sk_err = ECONNREFUSED; 3358 break; 3359 case TCP_CLOSE_WAIT: 3360 sk->sk_err = EPIPE; 3361 break; 3362 case TCP_CLOSE: 3363 return; 3364 default: 3365 sk->sk_err = ECONNRESET; 3366 } 3367 3368 if (!sock_flag(sk, SOCK_DEAD)) 3369 sk->sk_error_report(sk); 3370 3371 tcp_done(sk); 3372} 3373 3374/* 3375 * Process the FIN bit. This now behaves as it is supposed to work 3376 * and the FIN takes effect when it is validly part of sequence 3377 * space. Not before when we get holes. 3378 * 3379 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT 3380 * (and thence onto LAST-ACK and finally, CLOSE, we never enter 3381 * TIME-WAIT) 3382 * 3383 * If we are in FINWAIT-1, a received FIN indicates simultaneous 3384 * close and we go into CLOSING (and later onto TIME-WAIT) 3385 * 3386 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT. 3387 */ 3388static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th) 3389{ 3390 struct tcp_sock *tp = tcp_sk(sk); 3391 3392 inet_csk_schedule_ack(sk); 3393 3394 sk->sk_shutdown |= RCV_SHUTDOWN; 3395 sock_set_flag(sk, SOCK_DONE); 3396 3397 switch (sk->sk_state) { 3398 case TCP_SYN_RECV: 3399 case TCP_ESTABLISHED: 3400 /* Move to CLOSE_WAIT */ 3401 tcp_set_state(sk, TCP_CLOSE_WAIT); 3402 inet_csk(sk)->icsk_ack.pingpong = 1; 3403 break; 3404 3405 case TCP_CLOSE_WAIT: 3406 case TCP_CLOSING: 3407 /* Received a retransmission of the FIN, do 3408 * nothing. 3409 */ 3410 break; 3411 case TCP_LAST_ACK: 3412 /* RFC793: Remain in the LAST-ACK state. */ 3413 break; 3414 3415 case TCP_FIN_WAIT1: 3416 /* This case occurs when a simultaneous close 3417 * happens, we must ack the received FIN and 3418 * enter the CLOSING state. 3419 */ 3420 tcp_send_ack(sk); 3421 tcp_set_state(sk, TCP_CLOSING); 3422 break; 3423 case TCP_FIN_WAIT2: 3424 /* Received a FIN -- send ACK and enter TIME_WAIT. */ 3425 tcp_send_ack(sk); 3426 tcp_time_wait(sk, TCP_TIME_WAIT, 0); 3427 break; 3428 default: 3429 /* Only TCP_LISTEN and TCP_CLOSE are left, in these 3430 * cases we should never reach this piece of code. 3431 */ 3432 printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n", 3433 __FUNCTION__, sk->sk_state); 3434 break; 3435 } 3436 3437 /* It _is_ possible, that we have something out-of-order _after_ FIN. 3438 * Probably, we should reset in this case. For now drop them. 3439 */ 3440 __skb_queue_purge(&tp->out_of_order_queue); 3441 if (tcp_is_sack(tp)) 3442 tcp_sack_reset(&tp->rx_opt); 3443 sk_stream_mem_reclaim(sk); 3444 3445 if (!sock_flag(sk, SOCK_DEAD)) { 3446 sk->sk_state_change(sk); 3447 3448 /* Do not send POLL_HUP for half duplex close. */ 3449 if (sk->sk_shutdown == SHUTDOWN_MASK || 3450 sk->sk_state == TCP_CLOSE) 3451 sk_wake_async(sk, 1, POLL_HUP); 3452 else 3453 sk_wake_async(sk, 1, POLL_IN); 3454 } 3455} 3456 3457static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, u32 end_seq) 3458{ 3459 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) { 3460 if (before(seq, sp->start_seq)) 3461 sp->start_seq = seq; 3462 if (after(end_seq, sp->end_seq)) 3463 sp->end_seq = end_seq; 3464 return 1; 3465 } 3466 return 0; 3467} 3468 3469static void tcp_dsack_set(struct tcp_sock *tp, u32 seq, u32 end_seq) 3470{ 3471 if (tcp_is_sack(tp) && sysctl_tcp_dsack) { 3472 if (before(seq, tp->rcv_nxt)) 3473 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT); 3474 else 3475 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT); 3476 3477 tp->rx_opt.dsack = 1; 3478 tp->duplicate_sack[0].start_seq = seq; 3479 tp->duplicate_sack[0].end_seq = end_seq; 3480 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + 1, 4 - tp->rx_opt.tstamp_ok); 3481 } 3482} 3483 3484static void tcp_dsack_extend(struct tcp_sock *tp, u32 seq, u32 end_seq) 3485{ 3486 if (!tp->rx_opt.dsack) 3487 tcp_dsack_set(tp, seq, end_seq); 3488 else 3489 tcp_sack_extend(tp->duplicate_sack, seq, end_seq); 3490} 3491 3492static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb) 3493{ 3494 struct tcp_sock *tp = tcp_sk(sk); 3495 3496 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 3497 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 3498 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST); 3499 tcp_enter_quickack_mode(sk); 3500 3501 if (tcp_is_sack(tp) && sysctl_tcp_dsack) { 3502 u32 end_seq = TCP_SKB_CB(skb)->end_seq; 3503 3504 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) 3505 end_seq = tp->rcv_nxt; 3506 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, end_seq); 3507 } 3508 } 3509 3510 tcp_send_ack(sk); 3511} 3512 3513/* These routines update the SACK block as out-of-order packets arrive or 3514 * in-order packets close up the sequence space. 3515 */ 3516static void tcp_sack_maybe_coalesce(struct tcp_sock *tp) 3517{ 3518 int this_sack; 3519 struct tcp_sack_block *sp = &tp->selective_acks[0]; 3520 struct tcp_sack_block *swalk = sp+1; 3521 3522 /* See if the recent change to the first SACK eats into 3523 * or hits the sequence space of other SACK blocks, if so coalesce. 3524 */ 3525 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks; ) { 3526 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) { 3527 int i; 3528 3529 /* Zap SWALK, by moving every further SACK up by one slot. 3530 * Decrease num_sacks. 3531 */ 3532 tp->rx_opt.num_sacks--; 3533 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok); 3534 for (i=this_sack; i < tp->rx_opt.num_sacks; i++) 3535 sp[i] = sp[i+1]; 3536 continue; 3537 } 3538 this_sack++, swalk++; 3539 } 3540} 3541 3542static inline void tcp_sack_swap(struct tcp_sack_block *sack1, struct tcp_sack_block *sack2) 3543{ 3544 __u32 tmp; 3545 3546 tmp = sack1->start_seq; 3547 sack1->start_seq = sack2->start_seq; 3548 sack2->start_seq = tmp; 3549 3550 tmp = sack1->end_seq; 3551 sack1->end_seq = sack2->end_seq; 3552 sack2->end_seq = tmp; 3553} 3554 3555static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq) 3556{ 3557 struct tcp_sock *tp = tcp_sk(sk); 3558 struct tcp_sack_block *sp = &tp->selective_acks[0]; 3559 int cur_sacks = tp->rx_opt.num_sacks; 3560 int this_sack; 3561 3562 if (!cur_sacks) 3563 goto new_sack; 3564 3565 for (this_sack=0; this_sack<cur_sacks; this_sack++, sp++) { 3566 if (tcp_sack_extend(sp, seq, end_seq)) { 3567 /* Rotate this_sack to the first one. */ 3568 for (; this_sack>0; this_sack--, sp--) 3569 tcp_sack_swap(sp, sp-1); 3570 if (cur_sacks > 1) 3571 tcp_sack_maybe_coalesce(tp); 3572 return; 3573 } 3574 } 3575 3576 /* Could not find an adjacent existing SACK, build a new one, 3577 * put it at the front, and shift everyone else down. We 3578 * always know there is at least one SACK present already here. 3579 * 3580 * If the sack array is full, forget about the last one. 3581 */ 3582 if (this_sack >= 4) { 3583 this_sack--; 3584 tp->rx_opt.num_sacks--; 3585 sp--; 3586 } 3587 for (; this_sack > 0; this_sack--, sp--) 3588 *sp = *(sp-1); 3589 3590new_sack: 3591 /* Build the new head SACK, and we're done. */ 3592 sp->start_seq = seq; 3593 sp->end_seq = end_seq; 3594 tp->rx_opt.num_sacks++; 3595 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok); 3596} 3597 3598/* RCV.NXT advances, some SACKs should be eaten. */ 3599 3600static void tcp_sack_remove(struct tcp_sock *tp) 3601{ 3602 struct tcp_sack_block *sp = &tp->selective_acks[0]; 3603 int num_sacks = tp->rx_opt.num_sacks; 3604 int this_sack; 3605 3606 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */ 3607 if (skb_queue_empty(&tp->out_of_order_queue)) { 3608 tp->rx_opt.num_sacks = 0; 3609 tp->rx_opt.eff_sacks = tp->rx_opt.dsack; 3610 return; 3611 } 3612 3613 for (this_sack = 0; this_sack < num_sacks; ) { 3614 /* Check if the start of the sack is covered by RCV.NXT. */ 3615 if (!before(tp->rcv_nxt, sp->start_seq)) { 3616 int i; 3617 3618 /* RCV.NXT must cover all the block! */ 3619 BUG_TRAP(!before(tp->rcv_nxt, sp->end_seq)); 3620 3621 /* Zap this SACK, by moving forward any other SACKS. */ 3622 for (i=this_sack+1; i < num_sacks; i++) 3623 tp->selective_acks[i-1] = tp->selective_acks[i]; 3624 num_sacks--; 3625 continue; 3626 } 3627 this_sack++; 3628 sp++; 3629 } 3630 if (num_sacks != tp->rx_opt.num_sacks) { 3631 tp->rx_opt.num_sacks = num_sacks; 3632 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok); 3633 } 3634} 3635 3636/* This one checks to see if we can put data from the 3637 * out_of_order queue into the receive_queue. 3638 */ 3639static void tcp_ofo_queue(struct sock *sk) 3640{ 3641 struct tcp_sock *tp = tcp_sk(sk); 3642 __u32 dsack_high = tp->rcv_nxt; 3643 struct sk_buff *skb; 3644 3645 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) { 3646 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) 3647 break; 3648 3649 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) { 3650 __u32 dsack = dsack_high; 3651 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high)) 3652 dsack_high = TCP_SKB_CB(skb)->end_seq; 3653 tcp_dsack_extend(tp, TCP_SKB_CB(skb)->seq, dsack); 3654 } 3655 3656 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) { 3657 SOCK_DEBUG(sk, "ofo packet was already received \n"); 3658 __skb_unlink(skb, &tp->out_of_order_queue); 3659 __kfree_skb(skb); 3660 continue; 3661 } 3662 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n", 3663 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, 3664 TCP_SKB_CB(skb)->end_seq); 3665 3666 __skb_unlink(skb, &tp->out_of_order_queue); 3667 __skb_queue_tail(&sk->sk_receive_queue, skb); 3668 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; 3669 if (tcp_hdr(skb)->fin) 3670 tcp_fin(skb, sk, tcp_hdr(skb)); 3671 } 3672} 3673 3674static int tcp_prune_queue(struct sock *sk); 3675 3676static void tcp_data_queue(struct sock *sk, struct sk_buff *skb) 3677{ 3678 struct tcphdr *th = tcp_hdr(skb); 3679 struct tcp_sock *tp = tcp_sk(sk); 3680 int eaten = -1; 3681 3682 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) 3683 goto drop; 3684 3685 __skb_pull(skb, th->doff*4); 3686 3687 TCP_ECN_accept_cwr(tp, skb); 3688 3689 if (tp->rx_opt.dsack) { 3690 tp->rx_opt.dsack = 0; 3691 tp->rx_opt.eff_sacks = min_t(unsigned int, tp->rx_opt.num_sacks, 3692 4 - tp->rx_opt.tstamp_ok); 3693 } 3694 3695 /* Queue data for delivery to the user. 3696 * Packets in sequence go to the receive queue. 3697 * Out of sequence packets to the out_of_order_queue. 3698 */ 3699 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) { 3700 if (tcp_receive_window(tp) == 0) 3701 goto out_of_window; 3702 3703 /* Ok. In sequence. In window. */ 3704 if (tp->ucopy.task == current && 3705 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len && 3706 sock_owned_by_user(sk) && !tp->urg_data) { 3707 int chunk = min_t(unsigned int, skb->len, 3708 tp->ucopy.len); 3709 3710 __set_current_state(TASK_RUNNING); 3711 3712 local_bh_enable(); 3713 if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) { 3714 tp->ucopy.len -= chunk; 3715 tp->copied_seq += chunk; 3716 eaten = (chunk == skb->len && !th->fin); 3717 tcp_rcv_space_adjust(sk); 3718 } 3719 local_bh_disable(); 3720 } 3721 3722 if (eaten <= 0) { 3723queue_and_out: 3724 if (eaten < 0 && 3725 (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf || 3726 !sk_stream_rmem_schedule(sk, skb))) { 3727 if (tcp_prune_queue(sk) < 0 || 3728 !sk_stream_rmem_schedule(sk, skb)) 3729 goto drop; 3730 } 3731 sk_stream_set_owner_r(skb, sk); 3732 __skb_queue_tail(&sk->sk_receive_queue, skb); 3733 } 3734 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; 3735 if (skb->len) 3736 tcp_event_data_recv(sk, skb); 3737 if (th->fin) 3738 tcp_fin(skb, sk, th); 3739 3740 if (!skb_queue_empty(&tp->out_of_order_queue)) { 3741 tcp_ofo_queue(sk); 3742 3743 /* RFC2581. 4.2. SHOULD send immediate ACK, when 3744 * gap in queue is filled. 3745 */ 3746 if (skb_queue_empty(&tp->out_of_order_queue)) 3747 inet_csk(sk)->icsk_ack.pingpong = 0; 3748 } 3749 3750 if (tp->rx_opt.num_sacks) 3751 tcp_sack_remove(tp); 3752 3753 tcp_fast_path_check(sk); 3754 3755 if (eaten > 0) 3756 __kfree_skb(skb); 3757 else if (!sock_flag(sk, SOCK_DEAD)) 3758 sk->sk_data_ready(sk, 0); 3759 return; 3760 } 3761 3762 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) { 3763 /* A retransmit, 2nd most common case. Force an immediate ack. */ 3764 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST); 3765 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); 3766 3767out_of_window: 3768 tcp_enter_quickack_mode(sk); 3769 inet_csk_schedule_ack(sk); 3770drop: 3771 __kfree_skb(skb); 3772 return; 3773 } 3774 3775 /* Out of window. F.e. zero window probe. */ 3776 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp))) 3777 goto out_of_window; 3778 3779 tcp_enter_quickack_mode(sk); 3780 3781 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 3782 /* Partial packet, seq < rcv_next < end_seq */ 3783 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n", 3784 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, 3785 TCP_SKB_CB(skb)->end_seq); 3786 3787 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, tp->rcv_nxt); 3788 3789 /* If window is closed, drop tail of packet. But after 3790 * remembering D-SACK for its head made in previous line. 3791 */ 3792 if (!tcp_receive_window(tp)) 3793 goto out_of_window; 3794 goto queue_and_out; 3795 } 3796 3797 TCP_ECN_check_ce(tp, skb); 3798 3799 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf || 3800 !sk_stream_rmem_schedule(sk, skb)) { 3801 if (tcp_prune_queue(sk) < 0 || 3802 !sk_stream_rmem_schedule(sk, skb)) 3803 goto drop; 3804 } 3805 3806 /* Disable header prediction. */ 3807 tp->pred_flags = 0; 3808 inet_csk_schedule_ack(sk); 3809 3810 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n", 3811 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); 3812 3813 sk_stream_set_owner_r(skb, sk); 3814 3815 if (!skb_peek(&tp->out_of_order_queue)) { 3816 /* Initial out of order segment, build 1 SACK. */ 3817 if (tcp_is_sack(tp)) { 3818 tp->rx_opt.num_sacks = 1; 3819 tp->rx_opt.dsack = 0; 3820 tp->rx_opt.eff_sacks = 1; 3821 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq; 3822 tp->selective_acks[0].end_seq = 3823 TCP_SKB_CB(skb)->end_seq; 3824 } 3825 __skb_queue_head(&tp->out_of_order_queue,skb); 3826 } else { 3827 struct sk_buff *skb1 = tp->out_of_order_queue.prev; 3828 u32 seq = TCP_SKB_CB(skb)->seq; 3829 u32 end_seq = TCP_SKB_CB(skb)->end_seq; 3830 3831 if (seq == TCP_SKB_CB(skb1)->end_seq) { 3832 __skb_append(skb1, skb, &tp->out_of_order_queue); 3833 3834 if (!tp->rx_opt.num_sacks || 3835 tp->selective_acks[0].end_seq != seq) 3836 goto add_sack; 3837 3838 /* Common case: data arrive in order after hole. */ 3839 tp->selective_acks[0].end_seq = end_seq; 3840 return; 3841 } 3842 3843 /* Find place to insert this segment. */ 3844 do { 3845 if (!after(TCP_SKB_CB(skb1)->seq, seq)) 3846 break; 3847 } while ((skb1 = skb1->prev) != 3848 (struct sk_buff*)&tp->out_of_order_queue); 3849 3850 /* Do skb overlap to previous one? */ 3851 if (skb1 != (struct sk_buff*)&tp->out_of_order_queue && 3852 before(seq, TCP_SKB_CB(skb1)->end_seq)) { 3853 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) { 3854 /* All the bits are present. Drop. */ 3855 __kfree_skb(skb); 3856 tcp_dsack_set(tp, seq, end_seq); 3857 goto add_sack; 3858 } 3859 if (after(seq, TCP_SKB_CB(skb1)->seq)) { 3860 /* Partial overlap. */ 3861 tcp_dsack_set(tp, seq, TCP_SKB_CB(skb1)->end_seq); 3862 } else { 3863 skb1 = skb1->prev; 3864 } 3865 } 3866 __skb_insert(skb, skb1, skb1->next, &tp->out_of_order_queue); 3867 3868 /* And clean segments covered by new one as whole. */ 3869 while ((skb1 = skb->next) != 3870 (struct sk_buff*)&tp->out_of_order_queue && 3871 after(end_seq, TCP_SKB_CB(skb1)->seq)) { 3872 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) { 3873 tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, end_seq); 3874 break; 3875 } 3876 __skb_unlink(skb1, &tp->out_of_order_queue); 3877 tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, TCP_SKB_CB(skb1)->end_seq); 3878 __kfree_skb(skb1); 3879 } 3880 3881add_sack: 3882 if (tcp_is_sack(tp)) 3883 tcp_sack_new_ofo_skb(sk, seq, end_seq); 3884 } 3885} 3886 3887/* Collapse contiguous sequence of skbs head..tail with 3888 * sequence numbers start..end. 3889 * Segments with FIN/SYN are not collapsed (only because this 3890 * simplifies code) 3891 */ 3892static void 3893tcp_collapse(struct sock *sk, struct sk_buff_head *list, 3894 struct sk_buff *head, struct sk_buff *tail, 3895 u32 start, u32 end) 3896{ 3897 struct sk_buff *skb; 3898 3899 /* First, check that queue is collapsible and find 3900 * the point where collapsing can be useful. */ 3901 for (skb = head; skb != tail; ) { 3902 /* No new bits? It is possible on ofo queue. */ 3903 if (!before(start, TCP_SKB_CB(skb)->end_seq)) { 3904 struct sk_buff *next = skb->next; 3905 __skb_unlink(skb, list); 3906 __kfree_skb(skb); 3907 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED); 3908 skb = next; 3909 continue; 3910 } 3911 3912 /* The first skb to collapse is: 3913 * - not SYN/FIN and 3914 * - bloated or contains data before "start" or 3915 * overlaps to the next one. 3916 */ 3917 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin && 3918 (tcp_win_from_space(skb->truesize) > skb->len || 3919 before(TCP_SKB_CB(skb)->seq, start) || 3920 (skb->next != tail && 3921 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq))) 3922 break; 3923 3924 /* Decided to skip this, advance start seq. */ 3925 start = TCP_SKB_CB(skb)->end_seq; 3926 skb = skb->next; 3927 } 3928 if (skb == tail || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin) 3929 return; 3930 3931 while (before(start, end)) { 3932 struct sk_buff *nskb; 3933 unsigned int header = skb_headroom(skb); 3934 int copy = SKB_MAX_ORDER(header, 0); 3935 3936 /* Too big header? This can happen with IPv6. */ 3937 if (copy < 0) 3938 return; 3939 if (end-start < copy) 3940 copy = end-start; 3941 nskb = alloc_skb(copy+header, GFP_ATOMIC); 3942 if (!nskb) 3943 return; 3944 3945 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head); 3946 skb_set_network_header(nskb, (skb_network_header(skb) - 3947 skb->head)); 3948 skb_set_transport_header(nskb, (skb_transport_header(skb) - 3949 skb->head)); 3950 skb_reserve(nskb, header); 3951 memcpy(nskb->head, skb->head, header); 3952 memcpy(nskb->cb, skb->cb, sizeof(skb->cb)); 3953 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start; 3954 __skb_insert(nskb, skb->prev, skb, list); 3955 sk_stream_set_owner_r(nskb, sk); 3956 3957 /* Copy data, releasing collapsed skbs. */ 3958 while (copy > 0) { 3959 int offset = start - TCP_SKB_CB(skb)->seq; 3960 int size = TCP_SKB_CB(skb)->end_seq - start; 3961 3962 BUG_ON(offset < 0); 3963 if (size > 0) { 3964 size = min(copy, size); 3965 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size)) 3966 BUG(); 3967 TCP_SKB_CB(nskb)->end_seq += size; 3968 copy -= size; 3969 start += size; 3970 } 3971 if (!before(start, TCP_SKB_CB(skb)->end_seq)) { 3972 struct sk_buff *next = skb->next; 3973 __skb_unlink(skb, list); 3974 __kfree_skb(skb); 3975 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED); 3976 skb = next; 3977 if (skb == tail || 3978 tcp_hdr(skb)->syn || 3979 tcp_hdr(skb)->fin) 3980 return; 3981 } 3982 } 3983 } 3984} 3985 3986/* Collapse ofo queue. Algorithm: select contiguous sequence of skbs 3987 * and tcp_collapse() them until all the queue is collapsed. 3988 */ 3989static void tcp_collapse_ofo_queue(struct sock *sk) 3990{ 3991 struct tcp_sock *tp = tcp_sk(sk); 3992 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue); 3993 struct sk_buff *head; 3994 u32 start, end; 3995 3996 if (skb == NULL) 3997 return; 3998 3999 start = TCP_SKB_CB(skb)->seq; 4000 end = TCP_SKB_CB(skb)->end_seq; 4001 head = skb; 4002 4003 for (;;) { 4004 skb = skb->next; 4005 4006 /* Segment is terminated when we see gap or when 4007 * we are at the end of all the queue. */ 4008 if (skb == (struct sk_buff *)&tp->out_of_order_queue || 4009 after(TCP_SKB_CB(skb)->seq, end) || 4010 before(TCP_SKB_CB(skb)->end_seq, start)) { 4011 tcp_collapse(sk, &tp->out_of_order_queue, 4012 head, skb, start, end); 4013 head = skb; 4014 if (skb == (struct sk_buff *)&tp->out_of_order_queue) 4015 break; 4016 /* Start new segment */ 4017 start = TCP_SKB_CB(skb)->seq; 4018 end = TCP_SKB_CB(skb)->end_seq; 4019 } else { 4020 if (before(TCP_SKB_CB(skb)->seq, start)) 4021 start = TCP_SKB_CB(skb)->seq; 4022 if (after(TCP_SKB_CB(skb)->end_seq, end)) 4023 end = TCP_SKB_CB(skb)->end_seq; 4024 } 4025 } 4026} 4027 4028/* Reduce allocated memory if we can, trying to get 4029 * the socket within its memory limits again. 4030 * 4031 * Return less than zero if we should start dropping frames 4032 * until the socket owning process reads some of the data 4033 * to stabilize the situation. 4034 */ 4035static int tcp_prune_queue(struct sock *sk) 4036{ 4037 struct tcp_sock *tp = tcp_sk(sk); 4038 4039 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq); 4040 4041 NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED); 4042 4043 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf) 4044 tcp_clamp_window(sk); 4045 else if (tcp_memory_pressure) 4046 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss); 4047 4048 tcp_collapse_ofo_queue(sk); 4049 tcp_collapse(sk, &sk->sk_receive_queue, 4050 sk->sk_receive_queue.next, 4051 (struct sk_buff*)&sk->sk_receive_queue, 4052 tp->copied_seq, tp->rcv_nxt); 4053 sk_stream_mem_reclaim(sk); 4054 4055 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) 4056 return 0; 4057 4058 /* Collapsing did not help, destructive actions follow. 4059 * This must not ever occur. */ 4060 4061 /* First, purge the out_of_order queue. */ 4062 if (!skb_queue_empty(&tp->out_of_order_queue)) { 4063 NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED); 4064 __skb_queue_purge(&tp->out_of_order_queue); 4065 4066 /* Reset SACK state. A conforming SACK implementation will 4067 * do the same at a timeout based retransmit. When a connection 4068 * is in a sad state like this, we care only about integrity 4069 * of the connection not performance. 4070 */ 4071 if (tcp_is_sack(tp)) 4072 tcp_sack_reset(&tp->rx_opt); 4073 sk_stream_mem_reclaim(sk); 4074 } 4075 4076 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) 4077 return 0; 4078 4079 /* If we are really being abused, tell the caller to silently 4080 * drop receive data on the floor. It will get retransmitted 4081 * and hopefully then we'll have sufficient space. 4082 */ 4083 NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED); 4084 4085 /* Massive buffer overcommit. */ 4086 tp->pred_flags = 0; 4087 return -1; 4088} 4089 4090 4091/* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. 4092 * As additional protections, we do not touch cwnd in retransmission phases, 4093 * and if application hit its sndbuf limit recently. 4094 */ 4095void tcp_cwnd_application_limited(struct sock *sk) 4096{ 4097 struct tcp_sock *tp = tcp_sk(sk); 4098 4099 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && 4100 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { 4101 /* Limited by application or receiver window. */ 4102 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); 4103 u32 win_used = max(tp->snd_cwnd_used, init_win); 4104 if (win_used < tp->snd_cwnd) { 4105 tp->snd_ssthresh = tcp_current_ssthresh(sk); 4106 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1; 4107 } 4108 tp->snd_cwnd_used = 0; 4109 } 4110 tp->snd_cwnd_stamp = tcp_time_stamp; 4111} 4112 4113static int tcp_should_expand_sndbuf(struct sock *sk) 4114{ 4115 struct tcp_sock *tp = tcp_sk(sk); 4116 4117 /* If the user specified a specific send buffer setting, do 4118 * not modify it. 4119 */ 4120 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK) 4121 return 0; 4122 4123 /* If we are under global TCP memory pressure, do not expand. */ 4124 if (tcp_memory_pressure) 4125 return 0; 4126 4127 /* If we are under soft global TCP memory pressure, do not expand. */ 4128 if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0]) 4129 return 0; 4130 4131 /* If we filled the congestion window, do not expand. */ 4132 if (tp->packets_out >= tp->snd_cwnd) 4133 return 0; 4134 4135 return 1; 4136} 4137 4138/* When incoming ACK allowed to free some skb from write_queue, 4139 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket 4140 * on the exit from tcp input handler. 4141 * 4142 * PROBLEM: sndbuf expansion does not work well with largesend. 4143 */ 4144static void tcp_new_space(struct sock *sk) 4145{ 4146 struct tcp_sock *tp = tcp_sk(sk); 4147 4148 if (tcp_should_expand_sndbuf(sk)) { 4149 int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) + 4150 MAX_TCP_HEADER + 16 + sizeof(struct sk_buff), 4151 demanded = max_t(unsigned int, tp->snd_cwnd, 4152 tp->reordering + 1); 4153 sndmem *= 2*demanded; 4154 if (sndmem > sk->sk_sndbuf) 4155 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]); 4156 tp->snd_cwnd_stamp = tcp_time_stamp; 4157 } 4158 4159 sk->sk_write_space(sk); 4160} 4161 4162static void tcp_check_space(struct sock *sk) 4163{ 4164 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) { 4165 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK); 4166 if (sk->sk_socket && 4167 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) 4168 tcp_new_space(sk); 4169 } 4170} 4171 4172static inline void tcp_data_snd_check(struct sock *sk) 4173{ 4174 tcp_push_pending_frames(sk); 4175 tcp_check_space(sk); 4176} 4177 4178/* 4179 * Check if sending an ack is needed. 4180 */ 4181static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible) 4182{ 4183 struct tcp_sock *tp = tcp_sk(sk); 4184 4185 /* More than one full frame received... */ 4186 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss 4187 /* ... and right edge of window advances far enough. 4188 * (tcp_recvmsg() will send ACK otherwise). Or... 4189 */ 4190 && __tcp_select_window(sk) >= tp->rcv_wnd) || 4191 /* We ACK each frame or... */ 4192 tcp_in_quickack_mode(sk) || 4193 /* We have out of order data. */ 4194 (ofo_possible && 4195 skb_peek(&tp->out_of_order_queue))) { 4196 /* Then ack it now */ 4197 tcp_send_ack(sk); 4198 } else { 4199 /* Else, send delayed ack. */ 4200 tcp_send_delayed_ack(sk); 4201 } 4202} 4203 4204static inline void tcp_ack_snd_check(struct sock *sk) 4205{ 4206 if (!inet_csk_ack_scheduled(sk)) { 4207 /* We sent a data segment already. */ 4208 return; 4209 } 4210 __tcp_ack_snd_check(sk, 1); 4211} 4212 4213/* 4214 * This routine is only called when we have urgent data 4215 * signaled. Its the 'slow' part of tcp_urg. It could be 4216 * moved inline now as tcp_urg is only called from one 4217 * place. We handle URGent data wrong. We have to - as 4218 * BSD still doesn't use the correction from RFC961. 4219 * For 1003.1g we should support a new option TCP_STDURG to permit 4220 * either form (or just set the sysctl tcp_stdurg). 4221 */ 4222 4223static void tcp_check_urg(struct sock * sk, struct tcphdr * th) 4224{ 4225 struct tcp_sock *tp = tcp_sk(sk); 4226 u32 ptr = ntohs(th->urg_ptr); 4227 4228 if (ptr && !sysctl_tcp_stdurg) 4229 ptr--; 4230 ptr += ntohl(th->seq); 4231 4232 /* Ignore urgent data that we've already seen and read. */ 4233 if (after(tp->copied_seq, ptr)) 4234 return; 4235 4236 /* Do not replay urg ptr. 4237 * 4238 * NOTE: interesting situation not covered by specs. 4239 * Misbehaving sender may send urg ptr, pointing to segment, 4240 * which we already have in ofo queue. We are not able to fetch 4241 * such data and will stay in TCP_URG_NOTYET until will be eaten 4242 * by recvmsg(). Seems, we are not obliged to handle such wicked 4243 * situations. But it is worth to think about possibility of some 4244 * DoSes using some hypothetical application level deadlock. 4245 */ 4246 if (before(ptr, tp->rcv_nxt)) 4247 return; 4248 4249 /* Do we already have a newer (or duplicate) urgent pointer? */ 4250 if (tp->urg_data && !after(ptr, tp->urg_seq)) 4251 return; 4252 4253 /* Tell the world about our new urgent pointer. */ 4254 sk_send_sigurg(sk); 4255 4256 /* We may be adding urgent data when the last byte read was 4257 * urgent. To do this requires some care. We cannot just ignore 4258 * tp->copied_seq since we would read the last urgent byte again 4259 * as data, nor can we alter copied_seq until this data arrives 4260 * or we break the semantics of SIOCATMARK (and thus sockatmark()) 4261 * 4262 * NOTE. Double Dutch. Rendering to plain English: author of comment 4263 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB); 4264 * and expect that both A and B disappear from stream. This is _wrong_. 4265 * Though this happens in BSD with high probability, this is occasional. 4266 * Any application relying on this is buggy. Note also, that fix "works" 4267 * only in this artificial test. Insert some normal data between A and B and we will 4268 * decline of BSD again. Verdict: it is better to remove to trap 4269 * buggy users. 4270 */ 4271 if (tp->urg_seq == tp->copied_seq && tp->urg_data && 4272 !sock_flag(sk, SOCK_URGINLINE) && 4273 tp->copied_seq != tp->rcv_nxt) { 4274 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue); 4275 tp->copied_seq++; 4276 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) { 4277 __skb_unlink(skb, &sk->sk_receive_queue); 4278 __kfree_skb(skb); 4279 } 4280 } 4281 4282 tp->urg_data = TCP_URG_NOTYET; 4283 tp->urg_seq = ptr; 4284 4285 /* Disable header prediction. */ 4286 tp->pred_flags = 0; 4287} 4288 4289/* This is the 'fast' part of urgent handling. */ 4290static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th) 4291{ 4292 struct tcp_sock *tp = tcp_sk(sk); 4293 4294 /* Check if we get a new urgent pointer - normally not. */ 4295 if (th->urg) 4296 tcp_check_urg(sk,th); 4297 4298 /* Do we wait for any urgent data? - normally not... */ 4299 if (tp->urg_data == TCP_URG_NOTYET) { 4300 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) - 4301 th->syn; 4302 4303 /* Is the urgent pointer pointing into this packet? */ 4304 if (ptr < skb->len) { 4305 u8 tmp; 4306 if (skb_copy_bits(skb, ptr, &tmp, 1)) 4307 BUG(); 4308 tp->urg_data = TCP_URG_VALID | tmp; 4309 if (!sock_flag(sk, SOCK_DEAD)) 4310 sk->sk_data_ready(sk, 0); 4311 } 4312 } 4313} 4314 4315static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen) 4316{ 4317 struct tcp_sock *tp = tcp_sk(sk); 4318 int chunk = skb->len - hlen; 4319 int err; 4320 4321 local_bh_enable(); 4322 if (skb_csum_unnecessary(skb)) 4323 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk); 4324 else 4325 err = skb_copy_and_csum_datagram_iovec(skb, hlen, 4326 tp->ucopy.iov); 4327 4328 if (!err) { 4329 tp->ucopy.len -= chunk; 4330 tp->copied_seq += chunk; 4331 tcp_rcv_space_adjust(sk); 4332 } 4333 4334 local_bh_disable(); 4335 return err; 4336} 4337 4338static __sum16 __tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb) 4339{ 4340 __sum16 result; 4341 4342 if (sock_owned_by_user(sk)) { 4343 local_bh_enable(); 4344 result = __tcp_checksum_complete(skb); 4345 local_bh_disable(); 4346 } else { 4347 result = __tcp_checksum_complete(skb); 4348 } 4349 return result; 4350} 4351 4352static inline int tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb) 4353{ 4354 return !skb_csum_unnecessary(skb) && 4355 __tcp_checksum_complete_user(sk, skb); 4356} 4357 4358#ifdef CONFIG_NET_DMA 4359static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb, int hlen) 4360{ 4361 struct tcp_sock *tp = tcp_sk(sk); 4362 int chunk = skb->len - hlen; 4363 int dma_cookie; 4364 int copied_early = 0; 4365 4366 if (tp->ucopy.wakeup) 4367 return 0; 4368 4369 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list) 4370 tp->ucopy.dma_chan = get_softnet_dma(); 4371 4372 if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) { 4373 4374 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan, 4375 skb, hlen, tp->ucopy.iov, chunk, tp->ucopy.pinned_list); 4376 4377 if (dma_cookie < 0) 4378 goto out; 4379 4380 tp->ucopy.dma_cookie = dma_cookie; 4381 copied_early = 1; 4382 4383 tp->ucopy.len -= chunk; 4384 tp->copied_seq += chunk; 4385 tcp_rcv_space_adjust(sk); 4386 4387 if ((tp->ucopy.len == 0) || 4388 (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) || 4389 (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) { 4390 tp->ucopy.wakeup = 1; 4391 sk->sk_data_ready(sk, 0); 4392 } 4393 } else if (chunk > 0) { 4394 tp->ucopy.wakeup = 1; 4395 sk->sk_data_ready(sk, 0); 4396 } 4397out: 4398 return copied_early; 4399} 4400#endif /* CONFIG_NET_DMA */ 4401 4402/* 4403 * TCP receive function for the ESTABLISHED state. 4404 * 4405 * It is split into a fast path and a slow path. The fast path is 4406 * disabled when: 4407 * - A zero window was announced from us - zero window probing 4408 * is only handled properly in the slow path. 4409 * - Out of order segments arrived. 4410 * - Urgent data is expected. 4411 * - There is no buffer space left 4412 * - Unexpected TCP flags/window values/header lengths are received 4413 * (detected by checking the TCP header against pred_flags) 4414 * - Data is sent in both directions. Fast path only supports pure senders 4415 * or pure receivers (this means either the sequence number or the ack 4416 * value must stay constant) 4417 * - Unexpected TCP option. 4418 * 4419 * When these conditions are not satisfied it drops into a standard 4420 * receive procedure patterned after RFC793 to handle all cases. 4421 * The first three cases are guaranteed by proper pred_flags setting, 4422 * the rest is checked inline. Fast processing is turned on in 4423 * tcp_data_queue when everything is OK. 4424 */ 4425int tcp_rcv_established(struct sock *sk, struct sk_buff *skb, 4426 struct tcphdr *th, unsigned len) 4427{ 4428 struct tcp_sock *tp = tcp_sk(sk); 4429 4430 /* 4431 * Header prediction. 4432 * The code loosely follows the one in the famous 4433 * "30 instruction TCP receive" Van Jacobson mail. 4434 * 4435 * Van's trick is to deposit buffers into socket queue 4436 * on a device interrupt, to call tcp_recv function 4437 * on the receive process context and checksum and copy 4438 * the buffer to user space. smart... 4439 * 4440 * Our current scheme is not silly either but we take the 4441 * extra cost of the net_bh soft interrupt processing... 4442 * We do checksum and copy also but from device to kernel. 4443 */ 4444 4445 tp->rx_opt.saw_tstamp = 0; 4446 4447 /* pred_flags is 0xS?10 << 16 + snd_wnd 4448 * if header_prediction is to be made 4449 * 'S' will always be tp->tcp_header_len >> 2 4450 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to 4451 * turn it off (when there are holes in the receive 4452 * space for instance) 4453 * PSH flag is ignored. 4454 */ 4455 4456 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags && 4457 TCP_SKB_CB(skb)->seq == tp->rcv_nxt) { 4458 int tcp_header_len = tp->tcp_header_len; 4459 4460 /* Timestamp header prediction: tcp_header_len 4461 * is automatically equal to th->doff*4 due to pred_flags 4462 * match. 4463 */ 4464 4465 /* Check timestamp */ 4466 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) { 4467 __be32 *ptr = (__be32 *)(th + 1); 4468 4469 /* No? Slow path! */ 4470 if (*ptr != htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) 4471 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) 4472 goto slow_path; 4473 4474 tp->rx_opt.saw_tstamp = 1; 4475 ++ptr; 4476 tp->rx_opt.rcv_tsval = ntohl(*ptr); 4477 ++ptr; 4478 tp->rx_opt.rcv_tsecr = ntohl(*ptr); 4479 4480 /* If PAWS failed, check it more carefully in slow path */ 4481 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0) 4482 goto slow_path; 4483 4484 /* DO NOT update ts_recent here, if checksum fails 4485 * and timestamp was corrupted part, it will result 4486 * in a hung connection since we will drop all 4487 * future packets due to the PAWS test. 4488 */ 4489 } 4490 4491 if (len <= tcp_header_len) { 4492 /* Bulk data transfer: sender */ 4493 if (len == tcp_header_len) { 4494 /* Predicted packet is in window by definition. 4495 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 4496 * Hence, check seq<=rcv_wup reduces to: 4497 */ 4498 if (tcp_header_len == 4499 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && 4500 tp->rcv_nxt == tp->rcv_wup) 4501 tcp_store_ts_recent(tp); 4502 4503 /* We know that such packets are checksummed 4504 * on entry. 4505 */ 4506 tcp_ack(sk, skb, 0); 4507 __kfree_skb(skb); 4508 tcp_data_snd_check(sk); 4509 return 0; 4510 } else { /* Header too small */ 4511 TCP_INC_STATS_BH(TCP_MIB_INERRS); 4512 goto discard; 4513 } 4514 } else { 4515 int eaten = 0; 4516 int copied_early = 0; 4517 4518 if (tp->copied_seq == tp->rcv_nxt && 4519 len - tcp_header_len <= tp->ucopy.len) { 4520#ifdef CONFIG_NET_DMA 4521 if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) { 4522 copied_early = 1; 4523 eaten = 1; 4524 } 4525#endif 4526 if (tp->ucopy.task == current && sock_owned_by_user(sk) && !copied_early) { 4527 __set_current_state(TASK_RUNNING); 4528 4529 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len)) 4530 eaten = 1; 4531 } 4532 if (eaten) { 4533 /* Predicted packet is in window by definition. 4534 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 4535 * Hence, check seq<=rcv_wup reduces to: 4536 */ 4537 if (tcp_header_len == 4538 (sizeof(struct tcphdr) + 4539 TCPOLEN_TSTAMP_ALIGNED) && 4540 tp->rcv_nxt == tp->rcv_wup) 4541 tcp_store_ts_recent(tp); 4542 4543 tcp_rcv_rtt_measure_ts(sk, skb); 4544 4545 __skb_pull(skb, tcp_header_len); 4546 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; 4547 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER); 4548 } 4549 if (copied_early) 4550 tcp_cleanup_rbuf(sk, skb->len); 4551 } 4552 if (!eaten) { 4553 if (tcp_checksum_complete_user(sk, skb)) 4554 goto csum_error; 4555 4556 /* Predicted packet is in window by definition. 4557 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 4558 * Hence, check seq<=rcv_wup reduces to: 4559 */ 4560 if (tcp_header_len == 4561 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && 4562 tp->rcv_nxt == tp->rcv_wup) 4563 tcp_store_ts_recent(tp); 4564 4565 tcp_rcv_rtt_measure_ts(sk, skb); 4566 4567 if ((int)skb->truesize > sk->sk_forward_alloc) 4568 goto step5; 4569 4570 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS); 4571 4572 /* Bulk data transfer: receiver */ 4573 __skb_pull(skb,tcp_header_len); 4574 __skb_queue_tail(&sk->sk_receive_queue, skb); 4575 sk_stream_set_owner_r(skb, sk); 4576 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; 4577 } 4578 4579 tcp_event_data_recv(sk, skb); 4580 4581 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) { 4582 /* Well, only one small jumplet in fast path... */ 4583 tcp_ack(sk, skb, FLAG_DATA); 4584 tcp_data_snd_check(sk); 4585 if (!inet_csk_ack_scheduled(sk)) 4586 goto no_ack; 4587 } 4588 4589 __tcp_ack_snd_check(sk, 0); 4590no_ack: 4591#ifdef CONFIG_NET_DMA 4592 if (copied_early) 4593 __skb_queue_tail(&sk->sk_async_wait_queue, skb); 4594 else 4595#endif 4596 if (eaten) 4597 __kfree_skb(skb); 4598 else 4599 sk->sk_data_ready(sk, 0); 4600 return 0; 4601 } 4602 } 4603 4604slow_path: 4605 if (len < (th->doff<<2) || tcp_checksum_complete_user(sk, skb)) 4606 goto csum_error; 4607 4608 /* 4609 * RFC1323: H1. Apply PAWS check first. 4610 */ 4611 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp && 4612 tcp_paws_discard(sk, skb)) { 4613 if (!th->rst) { 4614 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED); 4615 tcp_send_dupack(sk, skb); 4616 goto discard; 4617 } 4618 /* Resets are accepted even if PAWS failed. 4619 4620 ts_recent update must be made after we are sure 4621 that the packet is in window. 4622 */ 4623 } 4624 4625 /* 4626 * Standard slow path. 4627 */ 4628 4629 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) { 4630 /* RFC793, page 37: "In all states except SYN-SENT, all reset 4631 * (RST) segments are validated by checking their SEQ-fields." 4632 * And page 69: "If an incoming segment is not acceptable, 4633 * an acknowledgment should be sent in reply (unless the RST bit 4634 * is set, if so drop the segment and return)". 4635 */ 4636 if (!th->rst) 4637 tcp_send_dupack(sk, skb); 4638 goto discard; 4639 } 4640 4641 if (th->rst) { 4642 tcp_reset(sk); 4643 goto discard; 4644 } 4645 4646 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq); 4647 4648 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 4649 TCP_INC_STATS_BH(TCP_MIB_INERRS); 4650 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN); 4651 tcp_reset(sk); 4652 return 1; 4653 } 4654 4655step5: 4656 if (th->ack) 4657 tcp_ack(sk, skb, FLAG_SLOWPATH); 4658 4659 tcp_rcv_rtt_measure_ts(sk, skb); 4660 4661 /* Process urgent data. */ 4662 tcp_urg(sk, skb, th); 4663 4664 /* step 7: process the segment text */ 4665 tcp_data_queue(sk, skb); 4666 4667 tcp_data_snd_check(sk); 4668 tcp_ack_snd_check(sk); 4669 return 0; 4670 4671csum_error: 4672 TCP_INC_STATS_BH(TCP_MIB_INERRS); 4673 4674discard: 4675 __kfree_skb(skb); 4676 return 0; 4677} 4678 4679static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb, 4680 struct tcphdr *th, unsigned len) 4681{ 4682 struct tcp_sock *tp = tcp_sk(sk); 4683 struct inet_connection_sock *icsk = inet_csk(sk); 4684 int saved_clamp = tp->rx_opt.mss_clamp; 4685 4686 tcp_parse_options(skb, &tp->rx_opt, 0); 4687 4688 if (th->ack) { 4689 /* rfc793: 4690 * "If the state is SYN-SENT then 4691 * first check the ACK bit 4692 * If the ACK bit is set 4693 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send 4694 * a reset (unless the RST bit is set, if so drop 4695 * the segment and return)" 4696 * 4697 * We do not send data with SYN, so that RFC-correct 4698 * test reduces to: 4699 */ 4700 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt) 4701 goto reset_and_undo; 4702 4703 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 4704 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp, 4705 tcp_time_stamp)) { 4706 NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED); 4707 goto reset_and_undo; 4708 } 4709 4710 /* Now ACK is acceptable. 4711 * 4712 * "If the RST bit is set 4713 * If the ACK was acceptable then signal the user "error: 4714 * connection reset", drop the segment, enter CLOSED state, 4715 * delete TCB, and return." 4716 */ 4717 4718 if (th->rst) { 4719 tcp_reset(sk); 4720 goto discard; 4721 } 4722 4723 /* rfc793: 4724 * "fifth, if neither of the SYN or RST bits is set then 4725 * drop the segment and return." 4726 * 4727 * See note below! 4728 * --ANK(990513) 4729 */ 4730 if (!th->syn) 4731 goto discard_and_undo; 4732 4733 /* rfc793: 4734 * "If the SYN bit is on ... 4735 * are acceptable then ... 4736 * (our SYN has been ACKed), change the connection 4737 * state to ESTABLISHED..." 4738 */ 4739 4740 TCP_ECN_rcv_synack(tp, th); 4741 4742 tp->snd_wl1 = TCP_SKB_CB(skb)->seq; 4743 tcp_ack(sk, skb, FLAG_SLOWPATH); 4744 4745 /* Ok.. it's good. Set up sequence numbers and 4746 * move to established. 4747 */ 4748 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; 4749 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; 4750 4751 /* RFC1323: The window in SYN & SYN/ACK segments is 4752 * never scaled. 4753 */ 4754 tp->snd_wnd = ntohs(th->window); 4755 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(skb)->seq); 4756 4757 if (!tp->rx_opt.wscale_ok) { 4758 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0; 4759 tp->window_clamp = min(tp->window_clamp, 65535U); 4760 } 4761 4762 if (tp->rx_opt.saw_tstamp) { 4763 tp->rx_opt.tstamp_ok = 1; 4764 tp->tcp_header_len = 4765 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 4766 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; 4767 tcp_store_ts_recent(tp); 4768 } else { 4769 tp->tcp_header_len = sizeof(struct tcphdr); 4770 } 4771 4772 if (tcp_is_sack(tp) && sysctl_tcp_fack) 4773 tcp_enable_fack(tp); 4774 4775 tcp_mtup_init(sk); 4776 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 4777 tcp_initialize_rcv_mss(sk); 4778 4779 /* Remember, tcp_poll() does not lock socket! 4780 * Change state from SYN-SENT only after copied_seq 4781 * is initialized. */ 4782 tp->copied_seq = tp->rcv_nxt; 4783 smp_mb(); 4784 tcp_set_state(sk, TCP_ESTABLISHED); 4785 4786 security_inet_conn_established(sk, skb); 4787 4788 /* Make sure socket is routed, for correct metrics. */ 4789 icsk->icsk_af_ops->rebuild_header(sk); 4790 4791 tcp_init_metrics(sk); 4792 4793 tcp_init_congestion_control(sk); 4794 4795 /* Prevent spurious tcp_cwnd_restart() on first data 4796 * packet. 4797 */ 4798 tp->lsndtime = tcp_time_stamp; 4799 4800 tcp_init_buffer_space(sk); 4801 4802 if (sock_flag(sk, SOCK_KEEPOPEN)) 4803 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp)); 4804 4805 if (!tp->rx_opt.snd_wscale) 4806 __tcp_fast_path_on(tp, tp->snd_wnd); 4807 else 4808 tp->pred_flags = 0; 4809 4810 if (!sock_flag(sk, SOCK_DEAD)) { 4811 sk->sk_state_change(sk); 4812 sk_wake_async(sk, 0, POLL_OUT); 4813 } 4814 4815 if (sk->sk_write_pending || 4816 icsk->icsk_accept_queue.rskq_defer_accept || 4817 icsk->icsk_ack.pingpong) { 4818 /* Save one ACK. Data will be ready after 4819 * several ticks, if write_pending is set. 4820 * 4821 * It may be deleted, but with this feature tcpdumps 4822 * look so _wonderfully_ clever, that I was not able 4823 * to stand against the temptation 8) --ANK 4824 */ 4825 inet_csk_schedule_ack(sk); 4826 icsk->icsk_ack.lrcvtime = tcp_time_stamp; 4827 icsk->icsk_ack.ato = TCP_ATO_MIN; 4828 tcp_incr_quickack(sk); 4829 tcp_enter_quickack_mode(sk); 4830 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 4831 TCP_DELACK_MAX, TCP_RTO_MAX); 4832 4833discard: 4834 __kfree_skb(skb); 4835 return 0; 4836 } else { 4837 tcp_send_ack(sk); 4838 } 4839 return -1; 4840 } 4841 4842 /* No ACK in the segment */ 4843 4844 if (th->rst) { 4845 /* rfc793: 4846 * "If the RST bit is set 4847 * 4848 * Otherwise (no ACK) drop the segment and return." 4849 */ 4850 4851 goto discard_and_undo; 4852 } 4853 4854 /* PAWS check. */ 4855 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && tcp_paws_check(&tp->rx_opt, 0)) 4856 goto discard_and_undo; 4857 4858 if (th->syn) { 4859 /* We see SYN without ACK. It is attempt of 4860 * simultaneous connect with crossed SYNs. 4861 * Particularly, it can be connect to self. 4862 */ 4863 tcp_set_state(sk, TCP_SYN_RECV); 4864 4865 if (tp->rx_opt.saw_tstamp) { 4866 tp->rx_opt.tstamp_ok = 1; 4867 tcp_store_ts_recent(tp); 4868 tp->tcp_header_len = 4869 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 4870 } else { 4871 tp->tcp_header_len = sizeof(struct tcphdr); 4872 } 4873 4874 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; 4875 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; 4876 4877 /* RFC1323: The window in SYN & SYN/ACK segments is 4878 * never scaled. 4879 */ 4880 tp->snd_wnd = ntohs(th->window); 4881 tp->snd_wl1 = TCP_SKB_CB(skb)->seq; 4882 tp->max_window = tp->snd_wnd; 4883 4884 TCP_ECN_rcv_syn(tp, th); 4885 4886 tcp_mtup_init(sk); 4887 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 4888 tcp_initialize_rcv_mss(sk); 4889 4890 4891 tcp_send_synack(sk); 4892#if 0 4893 /* Note, we could accept data and URG from this segment. 4894 * There are no obstacles to make this. 4895 * 4896 * However, if we ignore data in ACKless segments sometimes, 4897 * we have no reasons to accept it sometimes. 4898 * Also, seems the code doing it in step6 of tcp_rcv_state_process 4899 * is not flawless. So, discard packet for sanity. 4900 * Uncomment this return to process the data. 4901 */ 4902 return -1; 4903#else 4904 goto discard; 4905#endif 4906 } 4907 /* "fifth, if neither of the SYN or RST bits is set then 4908 * drop the segment and return." 4909 */ 4910 4911discard_and_undo: 4912 tcp_clear_options(&tp->rx_opt); 4913 tp->rx_opt.mss_clamp = saved_clamp; 4914 goto discard; 4915 4916reset_and_undo: 4917 tcp_clear_options(&tp->rx_opt); 4918 tp->rx_opt.mss_clamp = saved_clamp; 4919 return 1; 4920} 4921 4922 4923/* 4924 * This function implements the receiving procedure of RFC 793 for 4925 * all states except ESTABLISHED and TIME_WAIT. 4926 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be 4927 * address independent. 4928 */ 4929 4930int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb, 4931 struct tcphdr *th, unsigned len) 4932{ 4933 struct tcp_sock *tp = tcp_sk(sk); 4934 struct inet_connection_sock *icsk = inet_csk(sk); 4935 int queued = 0; 4936 4937 tp->rx_opt.saw_tstamp = 0; 4938 4939 switch (sk->sk_state) { 4940 case TCP_CLOSE: 4941 goto discard; 4942 4943 case TCP_LISTEN: 4944 if (th->ack) 4945 return 1; 4946 4947 if (th->rst) 4948 goto discard; 4949 4950 if (th->syn) { 4951 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0) 4952 return 1; 4953 4954 /* Now we have several options: In theory there is 4955 * nothing else in the frame. KA9Q has an option to 4956 * send data with the syn, BSD accepts data with the 4957 * syn up to the [to be] advertised window and 4958 * Solaris 2.1 gives you a protocol error. For now 4959 * we just ignore it, that fits the spec precisely 4960 * and avoids incompatibilities. It would be nice in 4961 * future to drop through and process the data. 4962 * 4963 * Now that TTCP is starting to be used we ought to 4964 * queue this data. 4965 * But, this leaves one open to an easy denial of 4966 * service attack, and SYN cookies can't defend 4967 * against this problem. So, we drop the data 4968 * in the interest of security over speed unless 4969 * it's still in use. 4970 */ 4971 kfree_skb(skb); 4972 return 0; 4973 } 4974 goto discard; 4975 4976 case TCP_SYN_SENT: 4977 queued = tcp_rcv_synsent_state_process(sk, skb, th, len); 4978 if (queued >= 0) 4979 return queued; 4980 4981 /* Do step6 onward by hand. */ 4982 tcp_urg(sk, skb, th); 4983 __kfree_skb(skb); 4984 tcp_data_snd_check(sk); 4985 return 0; 4986 } 4987 4988 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp && 4989 tcp_paws_discard(sk, skb)) { 4990 if (!th->rst) { 4991 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED); 4992 tcp_send_dupack(sk, skb); 4993 goto discard; 4994 } 4995 /* Reset is accepted even if it did not pass PAWS. */ 4996 } 4997 4998 /* step 1: check sequence number */ 4999 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) { 5000 if (!th->rst) 5001 tcp_send_dupack(sk, skb); 5002 goto discard; 5003 } 5004 5005 /* step 2: check RST bit */ 5006 if (th->rst) { 5007 tcp_reset(sk); 5008 goto discard; 5009 } 5010 5011 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq); 5012 5013 /* step 3: check security and precedence [ignored] */ 5014 5015 /* step 4: 5016 * 5017 * Check for a SYN in window. 5018 */ 5019 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 5020 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN); 5021 tcp_reset(sk); 5022 return 1; 5023 } 5024 5025 /* step 5: check the ACK field */ 5026 if (th->ack) { 5027 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH); 5028 5029 switch (sk->sk_state) { 5030 case TCP_SYN_RECV: 5031 if (acceptable) { 5032 tp->copied_seq = tp->rcv_nxt; 5033 smp_mb(); 5034 tcp_set_state(sk, TCP_ESTABLISHED); 5035 sk->sk_state_change(sk); 5036 5037 /* Note, that this wakeup is only for marginal 5038 * crossed SYN case. Passively open sockets 5039 * are not waked up, because sk->sk_sleep == 5040 * NULL and sk->sk_socket == NULL. 5041 */ 5042 if (sk->sk_socket) { 5043 sk_wake_async(sk,0,POLL_OUT); 5044 } 5045 5046 tp->snd_una = TCP_SKB_CB(skb)->ack_seq; 5047 tp->snd_wnd = ntohs(th->window) << 5048 tp->rx_opt.snd_wscale; 5049 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, 5050 TCP_SKB_CB(skb)->seq); 5051 5052 /* tcp_ack considers this ACK as duplicate 5053 * and does not calculate rtt. 5054 * Fix it at least with timestamps. 5055 */ 5056 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 5057 !tp->srtt) 5058 tcp_ack_saw_tstamp(sk, 0); 5059 5060 if (tp->rx_opt.tstamp_ok) 5061 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; 5062 5063 /* Make sure socket is routed, for 5064 * correct metrics. 5065 */ 5066 icsk->icsk_af_ops->rebuild_header(sk); 5067 5068 tcp_init_metrics(sk); 5069 5070 tcp_init_congestion_control(sk); 5071 5072 /* Prevent spurious tcp_cwnd_restart() on 5073 * first data packet. 5074 */ 5075 tp->lsndtime = tcp_time_stamp; 5076 5077 tcp_mtup_init(sk); 5078 tcp_initialize_rcv_mss(sk); 5079 tcp_init_buffer_space(sk); 5080 tcp_fast_path_on(tp); 5081 } else { 5082 return 1; 5083 } 5084 break; 5085 5086 case TCP_FIN_WAIT1: 5087 if (tp->snd_una == tp->write_seq) { 5088 tcp_set_state(sk, TCP_FIN_WAIT2); 5089 sk->sk_shutdown |= SEND_SHUTDOWN; 5090 dst_confirm(sk->sk_dst_cache); 5091 5092 if (!sock_flag(sk, SOCK_DEAD)) 5093 /* Wake up lingering close() */ 5094 sk->sk_state_change(sk); 5095 else { 5096 int tmo; 5097 5098 if (tp->linger2 < 0 || 5099 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 5100 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) { 5101 tcp_done(sk); 5102 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA); 5103 return 1; 5104 } 5105 5106 tmo = tcp_fin_time(sk); 5107 if (tmo > TCP_TIMEWAIT_LEN) { 5108 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN); 5109 } else if (th->fin || sock_owned_by_user(sk)) { 5110 /* Bad case. We could lose such FIN otherwise. 5111 * It is not a big problem, but it looks confusing 5112 * and not so rare event. We still can lose it now, 5113 * if it spins in bh_lock_sock(), but it is really 5114 * marginal case. 5115 */ 5116 inet_csk_reset_keepalive_timer(sk, tmo); 5117 } else { 5118 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo); 5119 goto discard; 5120 } 5121 } 5122 } 5123 break; 5124 5125 case TCP_CLOSING: 5126 if (tp->snd_una == tp->write_seq) { 5127 tcp_time_wait(sk, TCP_TIME_WAIT, 0); 5128 goto discard; 5129 } 5130 break; 5131 5132 case TCP_LAST_ACK: 5133 if (tp->snd_una == tp->write_seq) { 5134 tcp_update_metrics(sk); 5135 tcp_done(sk); 5136 goto discard; 5137 } 5138 break; 5139 } 5140 } else 5141 goto discard; 5142 5143 /* step 6: check the URG bit */ 5144 tcp_urg(sk, skb, th); 5145 5146 /* step 7: process the segment text */ 5147 switch (sk->sk_state) { 5148 case TCP_CLOSE_WAIT: 5149 case TCP_CLOSING: 5150 case TCP_LAST_ACK: 5151 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) 5152 break; 5153 case TCP_FIN_WAIT1: 5154 case TCP_FIN_WAIT2: 5155 /* RFC 793 says to queue data in these states, 5156 * RFC 1122 says we MUST send a reset. 5157 * BSD 4.4 also does reset. 5158 */ 5159 if (sk->sk_shutdown & RCV_SHUTDOWN) { 5160 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 5161 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) { 5162 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA); 5163 tcp_reset(sk); 5164 return 1; 5165 } 5166 } 5167 /* Fall through */ 5168 case TCP_ESTABLISHED: 5169 tcp_data_queue(sk, skb); 5170 queued = 1; 5171 break; 5172 } 5173 5174 /* tcp_data could move socket to TIME-WAIT */ 5175 if (sk->sk_state != TCP_CLOSE) { 5176 tcp_data_snd_check(sk); 5177 tcp_ack_snd_check(sk); 5178 } 5179 5180 if (!queued) { 5181discard: 5182 __kfree_skb(skb); 5183 } 5184 return 0; 5185} 5186 5187EXPORT_SYMBOL(sysctl_tcp_ecn); 5188EXPORT_SYMBOL(sysctl_tcp_reordering); 5189EXPORT_SYMBOL(tcp_parse_options); 5190EXPORT_SYMBOL(tcp_rcv_established); 5191EXPORT_SYMBOL(tcp_rcv_state_process); 5192EXPORT_SYMBOL(tcp_initialize_rcv_mss); 5193