tcp_input.c revision 9ead9a1d385ae2c52a6dcf2828d84ce66be04fc2
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; 89int sysctl_tcp_nometrics_save __read_mostly; 90 91int sysctl_tcp_moderate_rcvbuf __read_mostly = 1; 92int sysctl_tcp_abc __read_mostly; 93 94#define FLAG_DATA 0x01 /* Incoming frame contained data. */ 95#define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */ 96#define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */ 97#define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */ 98#define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */ 99#define FLAG_DATA_SACKED 0x20 /* New SACK. */ 100#define FLAG_ECE 0x40 /* ECE in this ACK */ 101#define FLAG_DATA_LOST 0x80 /* SACK detected data lossage. */ 102#define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/ 103 104#define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED) 105#define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED) 106#define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE) 107#define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED) 108 109#define IsReno(tp) ((tp)->rx_opt.sack_ok == 0) 110#define IsFack(tp) ((tp)->rx_opt.sack_ok & 2) 111#define IsDSack(tp) ((tp)->rx_opt.sack_ok & 4) 112 113#define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH) 114 115/* Adapt the MSS value used to make delayed ack decision to the 116 * real world. 117 */ 118static void tcp_measure_rcv_mss(struct sock *sk, 119 const struct sk_buff *skb) 120{ 121 struct inet_connection_sock *icsk = inet_csk(sk); 122 const unsigned int lss = icsk->icsk_ack.last_seg_size; 123 unsigned int len; 124 125 icsk->icsk_ack.last_seg_size = 0; 126 127 /* skb->len may jitter because of SACKs, even if peer 128 * sends good full-sized frames. 129 */ 130 len = skb_shinfo(skb)->gso_size ?: skb->len; 131 if (len >= icsk->icsk_ack.rcv_mss) { 132 icsk->icsk_ack.rcv_mss = len; 133 } else { 134 /* Otherwise, we make more careful check taking into account, 135 * that SACKs block is variable. 136 * 137 * "len" is invariant segment length, including TCP header. 138 */ 139 len += skb->data - skb->h.raw; 140 if (len >= TCP_MIN_RCVMSS + sizeof(struct tcphdr) || 141 /* If PSH is not set, packet should be 142 * full sized, provided peer TCP is not badly broken. 143 * This observation (if it is correct 8)) allows 144 * to handle super-low mtu links fairly. 145 */ 146 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) && 147 !(tcp_flag_word(skb->h.th)&TCP_REMNANT))) { 148 /* Subtract also invariant (if peer is RFC compliant), 149 * tcp header plus fixed timestamp option length. 150 * Resulting "len" is MSS free of SACK jitter. 151 */ 152 len -= tcp_sk(sk)->tcp_header_len; 153 icsk->icsk_ack.last_seg_size = len; 154 if (len == lss) { 155 icsk->icsk_ack.rcv_mss = len; 156 return; 157 } 158 } 159 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED) 160 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2; 161 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED; 162 } 163} 164 165static void tcp_incr_quickack(struct sock *sk) 166{ 167 struct inet_connection_sock *icsk = inet_csk(sk); 168 unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss); 169 170 if (quickacks==0) 171 quickacks=2; 172 if (quickacks > icsk->icsk_ack.quick) 173 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS); 174} 175 176void tcp_enter_quickack_mode(struct sock *sk) 177{ 178 struct inet_connection_sock *icsk = inet_csk(sk); 179 tcp_incr_quickack(sk); 180 icsk->icsk_ack.pingpong = 0; 181 icsk->icsk_ack.ato = TCP_ATO_MIN; 182} 183 184/* Send ACKs quickly, if "quick" count is not exhausted 185 * and the session is not interactive. 186 */ 187 188static inline int tcp_in_quickack_mode(const struct sock *sk) 189{ 190 const struct inet_connection_sock *icsk = inet_csk(sk); 191 return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong; 192} 193 194/* Buffer size and advertised window tuning. 195 * 196 * 1. Tuning sk->sk_sndbuf, when connection enters established state. 197 */ 198 199static void tcp_fixup_sndbuf(struct sock *sk) 200{ 201 int sndmem = tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER + 16 + 202 sizeof(struct sk_buff); 203 204 if (sk->sk_sndbuf < 3 * sndmem) 205 sk->sk_sndbuf = min(3 * sndmem, sysctl_tcp_wmem[2]); 206} 207 208/* 2. Tuning advertised window (window_clamp, rcv_ssthresh) 209 * 210 * All tcp_full_space() is split to two parts: "network" buffer, allocated 211 * forward and advertised in receiver window (tp->rcv_wnd) and 212 * "application buffer", required to isolate scheduling/application 213 * latencies from network. 214 * window_clamp is maximal advertised window. It can be less than 215 * tcp_full_space(), in this case tcp_full_space() - window_clamp 216 * is reserved for "application" buffer. The less window_clamp is 217 * the smoother our behaviour from viewpoint of network, but the lower 218 * throughput and the higher sensitivity of the connection to losses. 8) 219 * 220 * rcv_ssthresh is more strict window_clamp used at "slow start" 221 * phase to predict further behaviour of this connection. 222 * It is used for two goals: 223 * - to enforce header prediction at sender, even when application 224 * requires some significant "application buffer". It is check #1. 225 * - to prevent pruning of receive queue because of misprediction 226 * of receiver window. Check #2. 227 * 228 * The scheme does not work when sender sends good segments opening 229 * window and then starts to feed us spaghetti. But it should work 230 * in common situations. Otherwise, we have to rely on queue collapsing. 231 */ 232 233/* Slow part of check#2. */ 234static int __tcp_grow_window(const struct sock *sk, struct tcp_sock *tp, 235 const struct sk_buff *skb) 236{ 237 /* Optimize this! */ 238 int truesize = tcp_win_from_space(skb->truesize)/2; 239 int window = tcp_win_from_space(sysctl_tcp_rmem[2])/2; 240 241 while (tp->rcv_ssthresh <= window) { 242 if (truesize <= skb->len) 243 return 2 * inet_csk(sk)->icsk_ack.rcv_mss; 244 245 truesize >>= 1; 246 window >>= 1; 247 } 248 return 0; 249} 250 251static void tcp_grow_window(struct sock *sk, struct tcp_sock *tp, 252 struct sk_buff *skb) 253{ 254 /* Check #1 */ 255 if (tp->rcv_ssthresh < tp->window_clamp && 256 (int)tp->rcv_ssthresh < tcp_space(sk) && 257 !tcp_memory_pressure) { 258 int incr; 259 260 /* Check #2. Increase window, if skb with such overhead 261 * will fit to rcvbuf in future. 262 */ 263 if (tcp_win_from_space(skb->truesize) <= skb->len) 264 incr = 2*tp->advmss; 265 else 266 incr = __tcp_grow_window(sk, tp, skb); 267 268 if (incr) { 269 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr, tp->window_clamp); 270 inet_csk(sk)->icsk_ack.quick |= 1; 271 } 272 } 273} 274 275/* 3. Tuning rcvbuf, when connection enters established state. */ 276 277static void tcp_fixup_rcvbuf(struct sock *sk) 278{ 279 struct tcp_sock *tp = tcp_sk(sk); 280 int rcvmem = tp->advmss + MAX_TCP_HEADER + 16 + sizeof(struct sk_buff); 281 282 /* Try to select rcvbuf so that 4 mss-sized segments 283 * will fit to window and corresponding skbs will fit to our rcvbuf. 284 * (was 3; 4 is minimum to allow fast retransmit to work.) 285 */ 286 while (tcp_win_from_space(rcvmem) < tp->advmss) 287 rcvmem += 128; 288 if (sk->sk_rcvbuf < 4 * rcvmem) 289 sk->sk_rcvbuf = min(4 * rcvmem, sysctl_tcp_rmem[2]); 290} 291 292/* 4. Try to fixup all. It is made immediately after connection enters 293 * established state. 294 */ 295static void tcp_init_buffer_space(struct sock *sk) 296{ 297 struct tcp_sock *tp = tcp_sk(sk); 298 int maxwin; 299 300 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) 301 tcp_fixup_rcvbuf(sk); 302 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK)) 303 tcp_fixup_sndbuf(sk); 304 305 tp->rcvq_space.space = tp->rcv_wnd; 306 307 maxwin = tcp_full_space(sk); 308 309 if (tp->window_clamp >= maxwin) { 310 tp->window_clamp = maxwin; 311 312 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss) 313 tp->window_clamp = max(maxwin - 314 (maxwin >> sysctl_tcp_app_win), 315 4 * tp->advmss); 316 } 317 318 /* Force reservation of one segment. */ 319 if (sysctl_tcp_app_win && 320 tp->window_clamp > 2 * tp->advmss && 321 tp->window_clamp + tp->advmss > maxwin) 322 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss); 323 324 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp); 325 tp->snd_cwnd_stamp = tcp_time_stamp; 326} 327 328/* 5. Recalculate window clamp after socket hit its memory bounds. */ 329static void tcp_clamp_window(struct sock *sk, struct tcp_sock *tp) 330{ 331 struct inet_connection_sock *icsk = inet_csk(sk); 332 333 icsk->icsk_ack.quick = 0; 334 335 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] && 336 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) && 337 !tcp_memory_pressure && 338 atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) { 339 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc), 340 sysctl_tcp_rmem[2]); 341 } 342 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf) 343 tp->rcv_ssthresh = min(tp->window_clamp, 2U*tp->advmss); 344} 345 346 347/* Initialize RCV_MSS value. 348 * RCV_MSS is an our guess about MSS used by the peer. 349 * We haven't any direct information about the MSS. 350 * It's better to underestimate the RCV_MSS rather than overestimate. 351 * Overestimations make us ACKing less frequently than needed. 352 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss(). 353 */ 354void tcp_initialize_rcv_mss(struct sock *sk) 355{ 356 struct tcp_sock *tp = tcp_sk(sk); 357 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache); 358 359 hint = min(hint, tp->rcv_wnd/2); 360 hint = min(hint, TCP_MIN_RCVMSS); 361 hint = max(hint, TCP_MIN_MSS); 362 363 inet_csk(sk)->icsk_ack.rcv_mss = hint; 364} 365 366/* Receiver "autotuning" code. 367 * 368 * The algorithm for RTT estimation w/o timestamps is based on 369 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL. 370 * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps> 371 * 372 * More detail on this code can be found at 373 * <http://www.psc.edu/~jheffner/senior_thesis.ps>, 374 * though this reference is out of date. A new paper 375 * is pending. 376 */ 377static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep) 378{ 379 u32 new_sample = tp->rcv_rtt_est.rtt; 380 long m = sample; 381 382 if (m == 0) 383 m = 1; 384 385 if (new_sample != 0) { 386 /* If we sample in larger samples in the non-timestamp 387 * case, we could grossly overestimate the RTT especially 388 * with chatty applications or bulk transfer apps which 389 * are stalled on filesystem I/O. 390 * 391 * Also, since we are only going for a minimum in the 392 * non-timestamp case, we do not smooth things out 393 * else with timestamps disabled convergence takes too 394 * long. 395 */ 396 if (!win_dep) { 397 m -= (new_sample >> 3); 398 new_sample += m; 399 } else if (m < new_sample) 400 new_sample = m << 3; 401 } else { 402 /* No previous measure. */ 403 new_sample = m << 3; 404 } 405 406 if (tp->rcv_rtt_est.rtt != new_sample) 407 tp->rcv_rtt_est.rtt = new_sample; 408} 409 410static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp) 411{ 412 if (tp->rcv_rtt_est.time == 0) 413 goto new_measure; 414 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq)) 415 return; 416 tcp_rcv_rtt_update(tp, 417 jiffies - tp->rcv_rtt_est.time, 418 1); 419 420new_measure: 421 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd; 422 tp->rcv_rtt_est.time = tcp_time_stamp; 423} 424 425static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, const struct sk_buff *skb) 426{ 427 struct tcp_sock *tp = tcp_sk(sk); 428 if (tp->rx_opt.rcv_tsecr && 429 (TCP_SKB_CB(skb)->end_seq - 430 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss)) 431 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0); 432} 433 434/* 435 * This function should be called every time data is copied to user space. 436 * It calculates the appropriate TCP receive buffer space. 437 */ 438void tcp_rcv_space_adjust(struct sock *sk) 439{ 440 struct tcp_sock *tp = tcp_sk(sk); 441 int time; 442 int space; 443 444 if (tp->rcvq_space.time == 0) 445 goto new_measure; 446 447 time = tcp_time_stamp - tp->rcvq_space.time; 448 if (time < (tp->rcv_rtt_est.rtt >> 3) || 449 tp->rcv_rtt_est.rtt == 0) 450 return; 451 452 space = 2 * (tp->copied_seq - tp->rcvq_space.seq); 453 454 space = max(tp->rcvq_space.space, space); 455 456 if (tp->rcvq_space.space != space) { 457 int rcvmem; 458 459 tp->rcvq_space.space = space; 460 461 if (sysctl_tcp_moderate_rcvbuf && 462 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) { 463 int new_clamp = space; 464 465 /* Receive space grows, normalize in order to 466 * take into account packet headers and sk_buff 467 * structure overhead. 468 */ 469 space /= tp->advmss; 470 if (!space) 471 space = 1; 472 rcvmem = (tp->advmss + MAX_TCP_HEADER + 473 16 + sizeof(struct sk_buff)); 474 while (tcp_win_from_space(rcvmem) < tp->advmss) 475 rcvmem += 128; 476 space *= rcvmem; 477 space = min(space, sysctl_tcp_rmem[2]); 478 if (space > sk->sk_rcvbuf) { 479 sk->sk_rcvbuf = space; 480 481 /* Make the window clamp follow along. */ 482 tp->window_clamp = new_clamp; 483 } 484 } 485 } 486 487new_measure: 488 tp->rcvq_space.seq = tp->copied_seq; 489 tp->rcvq_space.time = tcp_time_stamp; 490} 491 492/* There is something which you must keep in mind when you analyze the 493 * behavior of the tp->ato delayed ack timeout interval. When a 494 * connection starts up, we want to ack as quickly as possible. The 495 * problem is that "good" TCP's do slow start at the beginning of data 496 * transmission. The means that until we send the first few ACK's the 497 * sender will sit on his end and only queue most of his data, because 498 * he can only send snd_cwnd unacked packets at any given time. For 499 * each ACK we send, he increments snd_cwnd and transmits more of his 500 * queue. -DaveM 501 */ 502static void tcp_event_data_recv(struct sock *sk, struct tcp_sock *tp, struct sk_buff *skb) 503{ 504 struct inet_connection_sock *icsk = inet_csk(sk); 505 u32 now; 506 507 inet_csk_schedule_ack(sk); 508 509 tcp_measure_rcv_mss(sk, skb); 510 511 tcp_rcv_rtt_measure(tp); 512 513 now = tcp_time_stamp; 514 515 if (!icsk->icsk_ack.ato) { 516 /* The _first_ data packet received, initialize 517 * delayed ACK engine. 518 */ 519 tcp_incr_quickack(sk); 520 icsk->icsk_ack.ato = TCP_ATO_MIN; 521 } else { 522 int m = now - icsk->icsk_ack.lrcvtime; 523 524 if (m <= TCP_ATO_MIN/2) { 525 /* The fastest case is the first. */ 526 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2; 527 } else if (m < icsk->icsk_ack.ato) { 528 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m; 529 if (icsk->icsk_ack.ato > icsk->icsk_rto) 530 icsk->icsk_ack.ato = icsk->icsk_rto; 531 } else if (m > icsk->icsk_rto) { 532 /* Too long gap. Apparently sender failed to 533 * restart window, so that we send ACKs quickly. 534 */ 535 tcp_incr_quickack(sk); 536 sk_stream_mem_reclaim(sk); 537 } 538 } 539 icsk->icsk_ack.lrcvtime = now; 540 541 TCP_ECN_check_ce(tp, skb); 542 543 if (skb->len >= 128) 544 tcp_grow_window(sk, tp, skb); 545} 546 547/* Called to compute a smoothed rtt estimate. The data fed to this 548 * routine either comes from timestamps, or from segments that were 549 * known _not_ to have been retransmitted [see Karn/Partridge 550 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88 551 * piece by Van Jacobson. 552 * NOTE: the next three routines used to be one big routine. 553 * To save cycles in the RFC 1323 implementation it was better to break 554 * it up into three procedures. -- erics 555 */ 556static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt) 557{ 558 struct tcp_sock *tp = tcp_sk(sk); 559 long m = mrtt; /* RTT */ 560 561 /* The following amusing code comes from Jacobson's 562 * article in SIGCOMM '88. Note that rtt and mdev 563 * are scaled versions of rtt and mean deviation. 564 * This is designed to be as fast as possible 565 * m stands for "measurement". 566 * 567 * On a 1990 paper the rto value is changed to: 568 * RTO = rtt + 4 * mdev 569 * 570 * Funny. This algorithm seems to be very broken. 571 * These formulae increase RTO, when it should be decreased, increase 572 * too slowly, when it should be increased quickly, decrease too quickly 573 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely 574 * does not matter how to _calculate_ it. Seems, it was trap 575 * that VJ failed to avoid. 8) 576 */ 577 if(m == 0) 578 m = 1; 579 if (tp->srtt != 0) { 580 m -= (tp->srtt >> 3); /* m is now error in rtt est */ 581 tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */ 582 if (m < 0) { 583 m = -m; /* m is now abs(error) */ 584 m -= (tp->mdev >> 2); /* similar update on mdev */ 585 /* This is similar to one of Eifel findings. 586 * Eifel blocks mdev updates when rtt decreases. 587 * This solution is a bit different: we use finer gain 588 * for mdev in this case (alpha*beta). 589 * Like Eifel it also prevents growth of rto, 590 * but also it limits too fast rto decreases, 591 * happening in pure Eifel. 592 */ 593 if (m > 0) 594 m >>= 3; 595 } else { 596 m -= (tp->mdev >> 2); /* similar update on mdev */ 597 } 598 tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */ 599 if (tp->mdev > tp->mdev_max) { 600 tp->mdev_max = tp->mdev; 601 if (tp->mdev_max > tp->rttvar) 602 tp->rttvar = tp->mdev_max; 603 } 604 if (after(tp->snd_una, tp->rtt_seq)) { 605 if (tp->mdev_max < tp->rttvar) 606 tp->rttvar -= (tp->rttvar-tp->mdev_max)>>2; 607 tp->rtt_seq = tp->snd_nxt; 608 tp->mdev_max = TCP_RTO_MIN; 609 } 610 } else { 611 /* no previous measure. */ 612 tp->srtt = m<<3; /* take the measured time to be rtt */ 613 tp->mdev = m<<1; /* make sure rto = 3*rtt */ 614 tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN); 615 tp->rtt_seq = tp->snd_nxt; 616 } 617} 618 619/* Calculate rto without backoff. This is the second half of Van Jacobson's 620 * routine referred to above. 621 */ 622static inline void tcp_set_rto(struct sock *sk) 623{ 624 const struct tcp_sock *tp = tcp_sk(sk); 625 /* Old crap is replaced with new one. 8) 626 * 627 * More seriously: 628 * 1. If rtt variance happened to be less 50msec, it is hallucination. 629 * It cannot be less due to utterly erratic ACK generation made 630 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_ 631 * to do with delayed acks, because at cwnd>2 true delack timeout 632 * is invisible. Actually, Linux-2.4 also generates erratic 633 * ACKs in some circumstances. 634 */ 635 inet_csk(sk)->icsk_rto = (tp->srtt >> 3) + tp->rttvar; 636 637 /* 2. Fixups made earlier cannot be right. 638 * If we do not estimate RTO correctly without them, 639 * all the algo is pure shit and should be replaced 640 * with correct one. It is exactly, which we pretend to do. 641 */ 642} 643 644/* NOTE: clamping at TCP_RTO_MIN is not required, current algo 645 * guarantees that rto is higher. 646 */ 647static inline void tcp_bound_rto(struct sock *sk) 648{ 649 if (inet_csk(sk)->icsk_rto > TCP_RTO_MAX) 650 inet_csk(sk)->icsk_rto = TCP_RTO_MAX; 651} 652 653/* Save metrics learned by this TCP session. 654 This function is called only, when TCP finishes successfully 655 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE. 656 */ 657void tcp_update_metrics(struct sock *sk) 658{ 659 struct tcp_sock *tp = tcp_sk(sk); 660 struct dst_entry *dst = __sk_dst_get(sk); 661 662 if (sysctl_tcp_nometrics_save) 663 return; 664 665 dst_confirm(dst); 666 667 if (dst && (dst->flags&DST_HOST)) { 668 const struct inet_connection_sock *icsk = inet_csk(sk); 669 int m; 670 671 if (icsk->icsk_backoff || !tp->srtt) { 672 /* This session failed to estimate rtt. Why? 673 * Probably, no packets returned in time. 674 * Reset our results. 675 */ 676 if (!(dst_metric_locked(dst, RTAX_RTT))) 677 dst->metrics[RTAX_RTT-1] = 0; 678 return; 679 } 680 681 m = dst_metric(dst, RTAX_RTT) - tp->srtt; 682 683 /* If newly calculated rtt larger than stored one, 684 * store new one. Otherwise, use EWMA. Remember, 685 * rtt overestimation is always better than underestimation. 686 */ 687 if (!(dst_metric_locked(dst, RTAX_RTT))) { 688 if (m <= 0) 689 dst->metrics[RTAX_RTT-1] = tp->srtt; 690 else 691 dst->metrics[RTAX_RTT-1] -= (m>>3); 692 } 693 694 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) { 695 if (m < 0) 696 m = -m; 697 698 /* Scale deviation to rttvar fixed point */ 699 m >>= 1; 700 if (m < tp->mdev) 701 m = tp->mdev; 702 703 if (m >= dst_metric(dst, RTAX_RTTVAR)) 704 dst->metrics[RTAX_RTTVAR-1] = m; 705 else 706 dst->metrics[RTAX_RTTVAR-1] -= 707 (dst->metrics[RTAX_RTTVAR-1] - m)>>2; 708 } 709 710 if (tp->snd_ssthresh >= 0xFFFF) { 711 /* Slow start still did not finish. */ 712 if (dst_metric(dst, RTAX_SSTHRESH) && 713 !dst_metric_locked(dst, RTAX_SSTHRESH) && 714 (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH)) 715 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1; 716 if (!dst_metric_locked(dst, RTAX_CWND) && 717 tp->snd_cwnd > dst_metric(dst, RTAX_CWND)) 718 dst->metrics[RTAX_CWND-1] = tp->snd_cwnd; 719 } else if (tp->snd_cwnd > tp->snd_ssthresh && 720 icsk->icsk_ca_state == TCP_CA_Open) { 721 /* Cong. avoidance phase, cwnd is reliable. */ 722 if (!dst_metric_locked(dst, RTAX_SSTHRESH)) 723 dst->metrics[RTAX_SSTHRESH-1] = 724 max(tp->snd_cwnd >> 1, tp->snd_ssthresh); 725 if (!dst_metric_locked(dst, RTAX_CWND)) 726 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_cwnd) >> 1; 727 } else { 728 /* Else slow start did not finish, cwnd is non-sense, 729 ssthresh may be also invalid. 730 */ 731 if (!dst_metric_locked(dst, RTAX_CWND)) 732 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_ssthresh) >> 1; 733 if (dst->metrics[RTAX_SSTHRESH-1] && 734 !dst_metric_locked(dst, RTAX_SSTHRESH) && 735 tp->snd_ssthresh > dst->metrics[RTAX_SSTHRESH-1]) 736 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh; 737 } 738 739 if (!dst_metric_locked(dst, RTAX_REORDERING)) { 740 if (dst->metrics[RTAX_REORDERING-1] < tp->reordering && 741 tp->reordering != sysctl_tcp_reordering) 742 dst->metrics[RTAX_REORDERING-1] = tp->reordering; 743 } 744 } 745} 746 747/* Numbers are taken from RFC2414. */ 748__u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst) 749{ 750 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0); 751 752 if (!cwnd) { 753 if (tp->mss_cache > 1460) 754 cwnd = 2; 755 else 756 cwnd = (tp->mss_cache > 1095) ? 3 : 4; 757 } 758 return min_t(__u32, cwnd, tp->snd_cwnd_clamp); 759} 760 761/* Set slow start threshold and cwnd not falling to slow start */ 762void tcp_enter_cwr(struct sock *sk) 763{ 764 struct tcp_sock *tp = tcp_sk(sk); 765 766 tp->prior_ssthresh = 0; 767 tp->bytes_acked = 0; 768 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) { 769 tp->undo_marker = 0; 770 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk); 771 tp->snd_cwnd = min(tp->snd_cwnd, 772 tcp_packets_in_flight(tp) + 1U); 773 tp->snd_cwnd_cnt = 0; 774 tp->high_seq = tp->snd_nxt; 775 tp->snd_cwnd_stamp = tcp_time_stamp; 776 TCP_ECN_queue_cwr(tp); 777 778 tcp_set_ca_state(sk, TCP_CA_CWR); 779 } 780} 781 782/* Initialize metrics on socket. */ 783 784static void tcp_init_metrics(struct sock *sk) 785{ 786 struct tcp_sock *tp = tcp_sk(sk); 787 struct dst_entry *dst = __sk_dst_get(sk); 788 789 if (dst == NULL) 790 goto reset; 791 792 dst_confirm(dst); 793 794 if (dst_metric_locked(dst, RTAX_CWND)) 795 tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND); 796 if (dst_metric(dst, RTAX_SSTHRESH)) { 797 tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH); 798 if (tp->snd_ssthresh > tp->snd_cwnd_clamp) 799 tp->snd_ssthresh = tp->snd_cwnd_clamp; 800 } 801 if (dst_metric(dst, RTAX_REORDERING) && 802 tp->reordering != dst_metric(dst, RTAX_REORDERING)) { 803 tp->rx_opt.sack_ok &= ~2; 804 tp->reordering = dst_metric(dst, RTAX_REORDERING); 805 } 806 807 if (dst_metric(dst, RTAX_RTT) == 0) 808 goto reset; 809 810 if (!tp->srtt && dst_metric(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3)) 811 goto reset; 812 813 /* Initial rtt is determined from SYN,SYN-ACK. 814 * The segment is small and rtt may appear much 815 * less than real one. Use per-dst memory 816 * to make it more realistic. 817 * 818 * A bit of theory. RTT is time passed after "normal" sized packet 819 * is sent until it is ACKed. In normal circumstances sending small 820 * packets force peer to delay ACKs and calculation is correct too. 821 * The algorithm is adaptive and, provided we follow specs, it 822 * NEVER underestimate RTT. BUT! If peer tries to make some clever 823 * tricks sort of "quick acks" for time long enough to decrease RTT 824 * to low value, and then abruptly stops to do it and starts to delay 825 * ACKs, wait for troubles. 826 */ 827 if (dst_metric(dst, RTAX_RTT) > tp->srtt) { 828 tp->srtt = dst_metric(dst, RTAX_RTT); 829 tp->rtt_seq = tp->snd_nxt; 830 } 831 if (dst_metric(dst, RTAX_RTTVAR) > tp->mdev) { 832 tp->mdev = dst_metric(dst, RTAX_RTTVAR); 833 tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN); 834 } 835 tcp_set_rto(sk); 836 tcp_bound_rto(sk); 837 if (inet_csk(sk)->icsk_rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp) 838 goto reset; 839 tp->snd_cwnd = tcp_init_cwnd(tp, dst); 840 tp->snd_cwnd_stamp = tcp_time_stamp; 841 return; 842 843reset: 844 /* Play conservative. If timestamps are not 845 * supported, TCP will fail to recalculate correct 846 * rtt, if initial rto is too small. FORGET ALL AND RESET! 847 */ 848 if (!tp->rx_opt.saw_tstamp && tp->srtt) { 849 tp->srtt = 0; 850 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT; 851 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; 852 } 853} 854 855static void tcp_update_reordering(struct sock *sk, const int metric, 856 const int ts) 857{ 858 struct tcp_sock *tp = tcp_sk(sk); 859 if (metric > tp->reordering) { 860 tp->reordering = min(TCP_MAX_REORDERING, metric); 861 862 /* This exciting event is worth to be remembered. 8) */ 863 if (ts) 864 NET_INC_STATS_BH(LINUX_MIB_TCPTSREORDER); 865 else if (IsReno(tp)) 866 NET_INC_STATS_BH(LINUX_MIB_TCPRENOREORDER); 867 else if (IsFack(tp)) 868 NET_INC_STATS_BH(LINUX_MIB_TCPFACKREORDER); 869 else 870 NET_INC_STATS_BH(LINUX_MIB_TCPSACKREORDER); 871#if FASTRETRANS_DEBUG > 1 872 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n", 873 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state, 874 tp->reordering, 875 tp->fackets_out, 876 tp->sacked_out, 877 tp->undo_marker ? tp->undo_retrans : 0); 878#endif 879 /* Disable FACK yet. */ 880 tp->rx_opt.sack_ok &= ~2; 881 } 882} 883 884/* This procedure tags the retransmission queue when SACKs arrive. 885 * 886 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L). 887 * Packets in queue with these bits set are counted in variables 888 * sacked_out, retrans_out and lost_out, correspondingly. 889 * 890 * Valid combinations are: 891 * Tag InFlight Description 892 * 0 1 - orig segment is in flight. 893 * S 0 - nothing flies, orig reached receiver. 894 * L 0 - nothing flies, orig lost by net. 895 * R 2 - both orig and retransmit are in flight. 896 * L|R 1 - orig is lost, retransmit is in flight. 897 * S|R 1 - orig reached receiver, retrans is still in flight. 898 * (L|S|R is logically valid, it could occur when L|R is sacked, 899 * but it is equivalent to plain S and code short-curcuits it to S. 900 * L|S is logically invalid, it would mean -1 packet in flight 8)) 901 * 902 * These 6 states form finite state machine, controlled by the following events: 903 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue()) 904 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue()) 905 * 3. Loss detection event of one of three flavors: 906 * A. Scoreboard estimator decided the packet is lost. 907 * A'. Reno "three dupacks" marks head of queue lost. 908 * A''. Its FACK modfication, head until snd.fack is lost. 909 * B. SACK arrives sacking data transmitted after never retransmitted 910 * hole was sent out. 911 * C. SACK arrives sacking SND.NXT at the moment, when the 912 * segment was retransmitted. 913 * 4. D-SACK added new rule: D-SACK changes any tag to S. 914 * 915 * It is pleasant to note, that state diagram turns out to be commutative, 916 * so that we are allowed not to be bothered by order of our actions, 917 * when multiple events arrive simultaneously. (see the function below). 918 * 919 * Reordering detection. 920 * -------------------- 921 * Reordering metric is maximal distance, which a packet can be displaced 922 * in packet stream. With SACKs we can estimate it: 923 * 924 * 1. SACK fills old hole and the corresponding segment was not 925 * ever retransmitted -> reordering. Alas, we cannot use it 926 * when segment was retransmitted. 927 * 2. The last flaw is solved with D-SACK. D-SACK arrives 928 * for retransmitted and already SACKed segment -> reordering.. 929 * Both of these heuristics are not used in Loss state, when we cannot 930 * account for retransmits accurately. 931 */ 932static int 933tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, u32 prior_snd_una) 934{ 935 const struct inet_connection_sock *icsk = inet_csk(sk); 936 struct tcp_sock *tp = tcp_sk(sk); 937 unsigned char *ptr = ack_skb->h.raw + TCP_SKB_CB(ack_skb)->sacked; 938 struct tcp_sack_block_wire *sp = (struct tcp_sack_block_wire *)(ptr+2); 939 struct sk_buff *cached_skb; 940 int num_sacks = (ptr[1] - TCPOLEN_SACK_BASE)>>3; 941 int reord = tp->packets_out; 942 int prior_fackets; 943 u32 lost_retrans = 0; 944 int flag = 0; 945 int dup_sack = 0; 946 int cached_fack_count; 947 int i; 948 int first_sack_index; 949 950 if (!tp->sacked_out) 951 tp->fackets_out = 0; 952 prior_fackets = tp->fackets_out; 953 954 /* Check for D-SACK. */ 955 if (before(ntohl(sp[0].start_seq), TCP_SKB_CB(ack_skb)->ack_seq)) { 956 dup_sack = 1; 957 tp->rx_opt.sack_ok |= 4; 958 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV); 959 } else if (num_sacks > 1 && 960 !after(ntohl(sp[0].end_seq), ntohl(sp[1].end_seq)) && 961 !before(ntohl(sp[0].start_seq), ntohl(sp[1].start_seq))) { 962 dup_sack = 1; 963 tp->rx_opt.sack_ok |= 4; 964 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV); 965 } 966 967 /* D-SACK for already forgotten data... 968 * Do dumb counting. */ 969 if (dup_sack && 970 !after(ntohl(sp[0].end_seq), prior_snd_una) && 971 after(ntohl(sp[0].end_seq), tp->undo_marker)) 972 tp->undo_retrans--; 973 974 /* Eliminate too old ACKs, but take into 975 * account more or less fresh ones, they can 976 * contain valid SACK info. 977 */ 978 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window)) 979 return 0; 980 981 /* SACK fastpath: 982 * if the only SACK change is the increase of the end_seq of 983 * the first block then only apply that SACK block 984 * and use retrans queue hinting otherwise slowpath */ 985 flag = 1; 986 for (i = 0; i < num_sacks; i++) { 987 __be32 start_seq = sp[i].start_seq; 988 __be32 end_seq = sp[i].end_seq; 989 990 if (i == 0) { 991 if (tp->recv_sack_cache[i].start_seq != start_seq) 992 flag = 0; 993 } else { 994 if ((tp->recv_sack_cache[i].start_seq != start_seq) || 995 (tp->recv_sack_cache[i].end_seq != end_seq)) 996 flag = 0; 997 } 998 tp->recv_sack_cache[i].start_seq = start_seq; 999 tp->recv_sack_cache[i].end_seq = end_seq; 1000 } 1001 /* Clear the rest of the cache sack blocks so they won't match mistakenly. */ 1002 for (; i < ARRAY_SIZE(tp->recv_sack_cache); i++) { 1003 tp->recv_sack_cache[i].start_seq = 0; 1004 tp->recv_sack_cache[i].end_seq = 0; 1005 } 1006 1007 first_sack_index = 0; 1008 if (flag) 1009 num_sacks = 1; 1010 else { 1011 int j; 1012 tp->fastpath_skb_hint = NULL; 1013 1014 /* order SACK blocks to allow in order walk of the retrans queue */ 1015 for (i = num_sacks-1; i > 0; i--) { 1016 for (j = 0; j < i; j++){ 1017 if (after(ntohl(sp[j].start_seq), 1018 ntohl(sp[j+1].start_seq))){ 1019 struct tcp_sack_block_wire tmp; 1020 1021 tmp = sp[j]; 1022 sp[j] = sp[j+1]; 1023 sp[j+1] = tmp; 1024 1025 /* Track where the first SACK block goes to */ 1026 if (j == first_sack_index) 1027 first_sack_index = j+1; 1028 } 1029 1030 } 1031 } 1032 } 1033 1034 /* clear flag as used for different purpose in following code */ 1035 flag = 0; 1036 1037 /* Use SACK fastpath hint if valid */ 1038 cached_skb = tp->fastpath_skb_hint; 1039 cached_fack_count = tp->fastpath_cnt_hint; 1040 if (!cached_skb) { 1041 cached_skb = sk->sk_write_queue.next; 1042 cached_fack_count = 0; 1043 } 1044 1045 for (i=0; i<num_sacks; i++, sp++) { 1046 struct sk_buff *skb; 1047 __u32 start_seq = ntohl(sp->start_seq); 1048 __u32 end_seq = ntohl(sp->end_seq); 1049 int fack_count; 1050 1051 skb = cached_skb; 1052 fack_count = cached_fack_count; 1053 1054 /* Event "B" in the comment above. */ 1055 if (after(end_seq, tp->high_seq)) 1056 flag |= FLAG_DATA_LOST; 1057 1058 sk_stream_for_retrans_queue_from(skb, sk) { 1059 int in_sack, pcount; 1060 u8 sacked; 1061 1062 cached_skb = skb; 1063 cached_fack_count = fack_count; 1064 if (i == first_sack_index) { 1065 tp->fastpath_skb_hint = skb; 1066 tp->fastpath_cnt_hint = fack_count; 1067 } 1068 1069 /* The retransmission queue is always in order, so 1070 * we can short-circuit the walk early. 1071 */ 1072 if (!before(TCP_SKB_CB(skb)->seq, end_seq)) 1073 break; 1074 1075 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && 1076 !before(end_seq, TCP_SKB_CB(skb)->end_seq); 1077 1078 pcount = tcp_skb_pcount(skb); 1079 1080 if (pcount > 1 && !in_sack && 1081 after(TCP_SKB_CB(skb)->end_seq, start_seq)) { 1082 unsigned int pkt_len; 1083 1084 in_sack = !after(start_seq, 1085 TCP_SKB_CB(skb)->seq); 1086 1087 if (!in_sack) 1088 pkt_len = (start_seq - 1089 TCP_SKB_CB(skb)->seq); 1090 else 1091 pkt_len = (end_seq - 1092 TCP_SKB_CB(skb)->seq); 1093 if (tcp_fragment(sk, skb, pkt_len, skb_shinfo(skb)->gso_size)) 1094 break; 1095 pcount = tcp_skb_pcount(skb); 1096 } 1097 1098 fack_count += pcount; 1099 1100 sacked = TCP_SKB_CB(skb)->sacked; 1101 1102 /* Account D-SACK for retransmitted packet. */ 1103 if ((dup_sack && in_sack) && 1104 (sacked & TCPCB_RETRANS) && 1105 after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker)) 1106 tp->undo_retrans--; 1107 1108 /* The frame is ACKed. */ 1109 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) { 1110 if (sacked&TCPCB_RETRANS) { 1111 if ((dup_sack && in_sack) && 1112 (sacked&TCPCB_SACKED_ACKED)) 1113 reord = min(fack_count, reord); 1114 } else { 1115 /* If it was in a hole, we detected reordering. */ 1116 if (fack_count < prior_fackets && 1117 !(sacked&TCPCB_SACKED_ACKED)) 1118 reord = min(fack_count, reord); 1119 } 1120 1121 /* Nothing to do; acked frame is about to be dropped. */ 1122 continue; 1123 } 1124 1125 if ((sacked&TCPCB_SACKED_RETRANS) && 1126 after(end_seq, TCP_SKB_CB(skb)->ack_seq) && 1127 (!lost_retrans || after(end_seq, lost_retrans))) 1128 lost_retrans = end_seq; 1129 1130 if (!in_sack) 1131 continue; 1132 1133 if (!(sacked&TCPCB_SACKED_ACKED)) { 1134 if (sacked & TCPCB_SACKED_RETRANS) { 1135 /* If the segment is not tagged as lost, 1136 * we do not clear RETRANS, believing 1137 * that retransmission is still in flight. 1138 */ 1139 if (sacked & TCPCB_LOST) { 1140 TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS); 1141 tp->lost_out -= tcp_skb_pcount(skb); 1142 tp->retrans_out -= tcp_skb_pcount(skb); 1143 1144 /* clear lost hint */ 1145 tp->retransmit_skb_hint = NULL; 1146 } 1147 } else { 1148 /* New sack for not retransmitted frame, 1149 * which was in hole. It is reordering. 1150 */ 1151 if (!(sacked & TCPCB_RETRANS) && 1152 fack_count < prior_fackets) 1153 reord = min(fack_count, reord); 1154 1155 if (sacked & TCPCB_LOST) { 1156 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; 1157 tp->lost_out -= tcp_skb_pcount(skb); 1158 1159 /* clear lost hint */ 1160 tp->retransmit_skb_hint = NULL; 1161 } 1162 } 1163 1164 TCP_SKB_CB(skb)->sacked |= TCPCB_SACKED_ACKED; 1165 flag |= FLAG_DATA_SACKED; 1166 tp->sacked_out += tcp_skb_pcount(skb); 1167 1168 if (fack_count > tp->fackets_out) 1169 tp->fackets_out = fack_count; 1170 } else { 1171 if (dup_sack && (sacked&TCPCB_RETRANS)) 1172 reord = min(fack_count, reord); 1173 } 1174 1175 /* D-SACK. We can detect redundant retransmission 1176 * in S|R and plain R frames and clear it. 1177 * undo_retrans is decreased above, L|R frames 1178 * are accounted above as well. 1179 */ 1180 if (dup_sack && 1181 (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS)) { 1182 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 1183 tp->retrans_out -= tcp_skb_pcount(skb); 1184 tp->retransmit_skb_hint = NULL; 1185 } 1186 } 1187 } 1188 1189 /* Check for lost retransmit. This superb idea is 1190 * borrowed from "ratehalving". Event "C". 1191 * Later note: FACK people cheated me again 8), 1192 * we have to account for reordering! Ugly, 1193 * but should help. 1194 */ 1195 if (lost_retrans && icsk->icsk_ca_state == TCP_CA_Recovery) { 1196 struct sk_buff *skb; 1197 1198 sk_stream_for_retrans_queue(skb, sk) { 1199 if (after(TCP_SKB_CB(skb)->seq, lost_retrans)) 1200 break; 1201 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 1202 continue; 1203 if ((TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) && 1204 after(lost_retrans, TCP_SKB_CB(skb)->ack_seq) && 1205 (IsFack(tp) || 1206 !before(lost_retrans, 1207 TCP_SKB_CB(skb)->ack_seq + tp->reordering * 1208 tp->mss_cache))) { 1209 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 1210 tp->retrans_out -= tcp_skb_pcount(skb); 1211 1212 /* clear lost hint */ 1213 tp->retransmit_skb_hint = NULL; 1214 1215 if (!(TCP_SKB_CB(skb)->sacked&(TCPCB_LOST|TCPCB_SACKED_ACKED))) { 1216 tp->lost_out += tcp_skb_pcount(skb); 1217 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 1218 flag |= FLAG_DATA_SACKED; 1219 NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT); 1220 } 1221 } 1222 } 1223 } 1224 1225 tp->left_out = tp->sacked_out + tp->lost_out; 1226 1227 if ((reord < tp->fackets_out) && icsk->icsk_ca_state != TCP_CA_Loss) 1228 tcp_update_reordering(sk, ((tp->fackets_out + 1) - reord), 0); 1229 1230#if FASTRETRANS_DEBUG > 0 1231 BUG_TRAP((int)tp->sacked_out >= 0); 1232 BUG_TRAP((int)tp->lost_out >= 0); 1233 BUG_TRAP((int)tp->retrans_out >= 0); 1234 BUG_TRAP((int)tcp_packets_in_flight(tp) >= 0); 1235#endif 1236 return flag; 1237} 1238 1239/* RTO occurred, but do not yet enter loss state. Instead, transmit two new 1240 * segments to see from the next ACKs whether any data was really missing. 1241 * If the RTO was spurious, new ACKs should arrive. 1242 */ 1243void tcp_enter_frto(struct sock *sk) 1244{ 1245 const struct inet_connection_sock *icsk = inet_csk(sk); 1246 struct tcp_sock *tp = tcp_sk(sk); 1247 struct sk_buff *skb; 1248 1249 tp->frto_counter = 1; 1250 1251 if (icsk->icsk_ca_state <= TCP_CA_Disorder || 1252 tp->snd_una == tp->high_seq || 1253 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) { 1254 tp->prior_ssthresh = tcp_current_ssthresh(sk); 1255 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 1256 tcp_ca_event(sk, CA_EVENT_FRTO); 1257 } 1258 1259 /* Have to clear retransmission markers here to keep the bookkeeping 1260 * in shape, even though we are not yet in Loss state. 1261 * If something was really lost, it is eventually caught up 1262 * in tcp_enter_frto_loss. 1263 */ 1264 tp->retrans_out = 0; 1265 tp->undo_marker = tp->snd_una; 1266 tp->undo_retrans = 0; 1267 1268 sk_stream_for_retrans_queue(skb, sk) { 1269 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 1270 } 1271 tcp_sync_left_out(tp); 1272 1273 tcp_set_ca_state(sk, TCP_CA_Open); 1274 tp->frto_highmark = tp->snd_nxt; 1275} 1276 1277/* Enter Loss state after F-RTO was applied. Dupack arrived after RTO, 1278 * which indicates that we should follow the traditional RTO recovery, 1279 * i.e. mark everything lost and do go-back-N retransmission. 1280 */ 1281static void tcp_enter_frto_loss(struct sock *sk) 1282{ 1283 struct tcp_sock *tp = tcp_sk(sk); 1284 struct sk_buff *skb; 1285 int cnt = 0; 1286 1287 tp->sacked_out = 0; 1288 tp->lost_out = 0; 1289 tp->fackets_out = 0; 1290 1291 sk_stream_for_retrans_queue(skb, sk) { 1292 cnt += tcp_skb_pcount(skb); 1293 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; 1294 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) { 1295 1296 /* Do not mark those segments lost that were 1297 * forward transmitted after RTO 1298 */ 1299 if (!after(TCP_SKB_CB(skb)->end_seq, 1300 tp->frto_highmark)) { 1301 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 1302 tp->lost_out += tcp_skb_pcount(skb); 1303 } 1304 } else { 1305 tp->sacked_out += tcp_skb_pcount(skb); 1306 tp->fackets_out = cnt; 1307 } 1308 } 1309 tcp_sync_left_out(tp); 1310 1311 tp->snd_cwnd = tp->frto_counter + tcp_packets_in_flight(tp)+1; 1312 tp->snd_cwnd_cnt = 0; 1313 tp->snd_cwnd_stamp = tcp_time_stamp; 1314 tp->undo_marker = 0; 1315 tp->frto_counter = 0; 1316 1317 tp->reordering = min_t(unsigned int, tp->reordering, 1318 sysctl_tcp_reordering); 1319 tcp_set_ca_state(sk, TCP_CA_Loss); 1320 tp->high_seq = tp->frto_highmark; 1321 TCP_ECN_queue_cwr(tp); 1322 1323 clear_all_retrans_hints(tp); 1324} 1325 1326void tcp_clear_retrans(struct tcp_sock *tp) 1327{ 1328 tp->left_out = 0; 1329 tp->retrans_out = 0; 1330 1331 tp->fackets_out = 0; 1332 tp->sacked_out = 0; 1333 tp->lost_out = 0; 1334 1335 tp->undo_marker = 0; 1336 tp->undo_retrans = 0; 1337} 1338 1339/* Enter Loss state. If "how" is not zero, forget all SACK information 1340 * and reset tags completely, otherwise preserve SACKs. If receiver 1341 * dropped its ofo queue, we will know this due to reneging detection. 1342 */ 1343void tcp_enter_loss(struct sock *sk, int how) 1344{ 1345 const struct inet_connection_sock *icsk = inet_csk(sk); 1346 struct tcp_sock *tp = tcp_sk(sk); 1347 struct sk_buff *skb; 1348 int cnt = 0; 1349 1350 /* Reduce ssthresh if it has not yet been made inside this window. */ 1351 if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq || 1352 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) { 1353 tp->prior_ssthresh = tcp_current_ssthresh(sk); 1354 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 1355 tcp_ca_event(sk, CA_EVENT_LOSS); 1356 } 1357 tp->snd_cwnd = 1; 1358 tp->snd_cwnd_cnt = 0; 1359 tp->snd_cwnd_stamp = tcp_time_stamp; 1360 1361 tp->bytes_acked = 0; 1362 tcp_clear_retrans(tp); 1363 1364 /* Push undo marker, if it was plain RTO and nothing 1365 * was retransmitted. */ 1366 if (!how) 1367 tp->undo_marker = tp->snd_una; 1368 1369 sk_stream_for_retrans_queue(skb, sk) { 1370 cnt += tcp_skb_pcount(skb); 1371 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS) 1372 tp->undo_marker = 0; 1373 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED; 1374 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) { 1375 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED; 1376 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 1377 tp->lost_out += tcp_skb_pcount(skb); 1378 } else { 1379 tp->sacked_out += tcp_skb_pcount(skb); 1380 tp->fackets_out = cnt; 1381 } 1382 } 1383 tcp_sync_left_out(tp); 1384 1385 tp->reordering = min_t(unsigned int, tp->reordering, 1386 sysctl_tcp_reordering); 1387 tcp_set_ca_state(sk, TCP_CA_Loss); 1388 tp->high_seq = tp->snd_nxt; 1389 TCP_ECN_queue_cwr(tp); 1390 1391 clear_all_retrans_hints(tp); 1392} 1393 1394static int tcp_check_sack_reneging(struct sock *sk) 1395{ 1396 struct sk_buff *skb; 1397 1398 /* If ACK arrived pointing to a remembered SACK, 1399 * it means that our remembered SACKs do not reflect 1400 * real state of receiver i.e. 1401 * receiver _host_ is heavily congested (or buggy). 1402 * Do processing similar to RTO timeout. 1403 */ 1404 if ((skb = skb_peek(&sk->sk_write_queue)) != NULL && 1405 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) { 1406 struct inet_connection_sock *icsk = inet_csk(sk); 1407 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING); 1408 1409 tcp_enter_loss(sk, 1); 1410 icsk->icsk_retransmits++; 1411 tcp_retransmit_skb(sk, skb_peek(&sk->sk_write_queue)); 1412 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 1413 icsk->icsk_rto, TCP_RTO_MAX); 1414 return 1; 1415 } 1416 return 0; 1417} 1418 1419static inline int tcp_fackets_out(struct tcp_sock *tp) 1420{ 1421 return IsReno(tp) ? tp->sacked_out+1 : tp->fackets_out; 1422} 1423 1424static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb) 1425{ 1426 return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto); 1427} 1428 1429static inline int tcp_head_timedout(struct sock *sk, struct tcp_sock *tp) 1430{ 1431 return tp->packets_out && 1432 tcp_skb_timedout(sk, skb_peek(&sk->sk_write_queue)); 1433} 1434 1435/* Linux NewReno/SACK/FACK/ECN state machine. 1436 * -------------------------------------- 1437 * 1438 * "Open" Normal state, no dubious events, fast path. 1439 * "Disorder" In all the respects it is "Open", 1440 * but requires a bit more attention. It is entered when 1441 * we see some SACKs or dupacks. It is split of "Open" 1442 * mainly to move some processing from fast path to slow one. 1443 * "CWR" CWND was reduced due to some Congestion Notification event. 1444 * It can be ECN, ICMP source quench, local device congestion. 1445 * "Recovery" CWND was reduced, we are fast-retransmitting. 1446 * "Loss" CWND was reduced due to RTO timeout or SACK reneging. 1447 * 1448 * tcp_fastretrans_alert() is entered: 1449 * - each incoming ACK, if state is not "Open" 1450 * - when arrived ACK is unusual, namely: 1451 * * SACK 1452 * * Duplicate ACK. 1453 * * ECN ECE. 1454 * 1455 * Counting packets in flight is pretty simple. 1456 * 1457 * in_flight = packets_out - left_out + retrans_out 1458 * 1459 * packets_out is SND.NXT-SND.UNA counted in packets. 1460 * 1461 * retrans_out is number of retransmitted segments. 1462 * 1463 * left_out is number of segments left network, but not ACKed yet. 1464 * 1465 * left_out = sacked_out + lost_out 1466 * 1467 * sacked_out: Packets, which arrived to receiver out of order 1468 * and hence not ACKed. With SACKs this number is simply 1469 * amount of SACKed data. Even without SACKs 1470 * it is easy to give pretty reliable estimate of this number, 1471 * counting duplicate ACKs. 1472 * 1473 * lost_out: Packets lost by network. TCP has no explicit 1474 * "loss notification" feedback from network (for now). 1475 * It means that this number can be only _guessed_. 1476 * Actually, it is the heuristics to predict lossage that 1477 * distinguishes different algorithms. 1478 * 1479 * F.e. after RTO, when all the queue is considered as lost, 1480 * lost_out = packets_out and in_flight = retrans_out. 1481 * 1482 * Essentially, we have now two algorithms counting 1483 * lost packets. 1484 * 1485 * FACK: It is the simplest heuristics. As soon as we decided 1486 * that something is lost, we decide that _all_ not SACKed 1487 * packets until the most forward SACK are lost. I.e. 1488 * lost_out = fackets_out - sacked_out and left_out = fackets_out. 1489 * It is absolutely correct estimate, if network does not reorder 1490 * packets. And it loses any connection to reality when reordering 1491 * takes place. We use FACK by default until reordering 1492 * is suspected on the path to this destination. 1493 * 1494 * NewReno: when Recovery is entered, we assume that one segment 1495 * is lost (classic Reno). While we are in Recovery and 1496 * a partial ACK arrives, we assume that one more packet 1497 * is lost (NewReno). This heuristics are the same in NewReno 1498 * and SACK. 1499 * 1500 * Imagine, that's all! Forget about all this shamanism about CWND inflation 1501 * deflation etc. CWND is real congestion window, never inflated, changes 1502 * only according to classic VJ rules. 1503 * 1504 * Really tricky (and requiring careful tuning) part of algorithm 1505 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue(). 1506 * The first determines the moment _when_ we should reduce CWND and, 1507 * hence, slow down forward transmission. In fact, it determines the moment 1508 * when we decide that hole is caused by loss, rather than by a reorder. 1509 * 1510 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill 1511 * holes, caused by lost packets. 1512 * 1513 * And the most logically complicated part of algorithm is undo 1514 * heuristics. We detect false retransmits due to both too early 1515 * fast retransmit (reordering) and underestimated RTO, analyzing 1516 * timestamps and D-SACKs. When we detect that some segments were 1517 * retransmitted by mistake and CWND reduction was wrong, we undo 1518 * window reduction and abort recovery phase. This logic is hidden 1519 * inside several functions named tcp_try_undo_<something>. 1520 */ 1521 1522/* This function decides, when we should leave Disordered state 1523 * and enter Recovery phase, reducing congestion window. 1524 * 1525 * Main question: may we further continue forward transmission 1526 * with the same cwnd? 1527 */ 1528static int tcp_time_to_recover(struct sock *sk, struct tcp_sock *tp) 1529{ 1530 __u32 packets_out; 1531 1532 /* Trick#1: The loss is proven. */ 1533 if (tp->lost_out) 1534 return 1; 1535 1536 /* Not-A-Trick#2 : Classic rule... */ 1537 if (tcp_fackets_out(tp) > tp->reordering) 1538 return 1; 1539 1540 /* Trick#3 : when we use RFC2988 timer restart, fast 1541 * retransmit can be triggered by timeout of queue head. 1542 */ 1543 if (tcp_head_timedout(sk, tp)) 1544 return 1; 1545 1546 /* Trick#4: It is still not OK... But will it be useful to delay 1547 * recovery more? 1548 */ 1549 packets_out = tp->packets_out; 1550 if (packets_out <= tp->reordering && 1551 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) && 1552 !tcp_may_send_now(sk, tp)) { 1553 /* We have nothing to send. This connection is limited 1554 * either by receiver window or by application. 1555 */ 1556 return 1; 1557 } 1558 1559 return 0; 1560} 1561 1562/* If we receive more dupacks than we expected counting segments 1563 * in assumption of absent reordering, interpret this as reordering. 1564 * The only another reason could be bug in receiver TCP. 1565 */ 1566static void tcp_check_reno_reordering(struct sock *sk, const int addend) 1567{ 1568 struct tcp_sock *tp = tcp_sk(sk); 1569 u32 holes; 1570 1571 holes = max(tp->lost_out, 1U); 1572 holes = min(holes, tp->packets_out); 1573 1574 if ((tp->sacked_out + holes) > tp->packets_out) { 1575 tp->sacked_out = tp->packets_out - holes; 1576 tcp_update_reordering(sk, tp->packets_out + addend, 0); 1577 } 1578} 1579 1580/* Emulate SACKs for SACKless connection: account for a new dupack. */ 1581 1582static void tcp_add_reno_sack(struct sock *sk) 1583{ 1584 struct tcp_sock *tp = tcp_sk(sk); 1585 tp->sacked_out++; 1586 tcp_check_reno_reordering(sk, 0); 1587 tcp_sync_left_out(tp); 1588} 1589 1590/* Account for ACK, ACKing some data in Reno Recovery phase. */ 1591 1592static void tcp_remove_reno_sacks(struct sock *sk, struct tcp_sock *tp, int acked) 1593{ 1594 if (acked > 0) { 1595 /* One ACK acked hole. The rest eat duplicate ACKs. */ 1596 if (acked-1 >= tp->sacked_out) 1597 tp->sacked_out = 0; 1598 else 1599 tp->sacked_out -= acked-1; 1600 } 1601 tcp_check_reno_reordering(sk, acked); 1602 tcp_sync_left_out(tp); 1603} 1604 1605static inline void tcp_reset_reno_sack(struct tcp_sock *tp) 1606{ 1607 tp->sacked_out = 0; 1608 tp->left_out = tp->lost_out; 1609} 1610 1611/* Mark head of queue up as lost. */ 1612static void tcp_mark_head_lost(struct sock *sk, struct tcp_sock *tp, 1613 int packets, u32 high_seq) 1614{ 1615 struct sk_buff *skb; 1616 int cnt; 1617 1618 BUG_TRAP(packets <= tp->packets_out); 1619 if (tp->lost_skb_hint) { 1620 skb = tp->lost_skb_hint; 1621 cnt = tp->lost_cnt_hint; 1622 } else { 1623 skb = sk->sk_write_queue.next; 1624 cnt = 0; 1625 } 1626 1627 sk_stream_for_retrans_queue_from(skb, sk) { 1628 /* TODO: do this better */ 1629 /* this is not the most efficient way to do this... */ 1630 tp->lost_skb_hint = skb; 1631 tp->lost_cnt_hint = cnt; 1632 cnt += tcp_skb_pcount(skb); 1633 if (cnt > packets || after(TCP_SKB_CB(skb)->end_seq, high_seq)) 1634 break; 1635 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) { 1636 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 1637 tp->lost_out += tcp_skb_pcount(skb); 1638 1639 /* clear xmit_retransmit_queue hints 1640 * if this is beyond hint */ 1641 if(tp->retransmit_skb_hint != NULL && 1642 before(TCP_SKB_CB(skb)->seq, 1643 TCP_SKB_CB(tp->retransmit_skb_hint)->seq)) { 1644 1645 tp->retransmit_skb_hint = NULL; 1646 } 1647 } 1648 } 1649 tcp_sync_left_out(tp); 1650} 1651 1652/* Account newly detected lost packet(s) */ 1653 1654static void tcp_update_scoreboard(struct sock *sk, struct tcp_sock *tp) 1655{ 1656 if (IsFack(tp)) { 1657 int lost = tp->fackets_out - tp->reordering; 1658 if (lost <= 0) 1659 lost = 1; 1660 tcp_mark_head_lost(sk, tp, lost, tp->high_seq); 1661 } else { 1662 tcp_mark_head_lost(sk, tp, 1, tp->high_seq); 1663 } 1664 1665 /* New heuristics: it is possible only after we switched 1666 * to restart timer each time when something is ACKed. 1667 * Hence, we can detect timed out packets during fast 1668 * retransmit without falling to slow start. 1669 */ 1670 if (!IsReno(tp) && tcp_head_timedout(sk, tp)) { 1671 struct sk_buff *skb; 1672 1673 skb = tp->scoreboard_skb_hint ? tp->scoreboard_skb_hint 1674 : sk->sk_write_queue.next; 1675 1676 sk_stream_for_retrans_queue_from(skb, sk) { 1677 if (!tcp_skb_timedout(sk, skb)) 1678 break; 1679 1680 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) { 1681 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 1682 tp->lost_out += tcp_skb_pcount(skb); 1683 1684 /* clear xmit_retrans hint */ 1685 if (tp->retransmit_skb_hint && 1686 before(TCP_SKB_CB(skb)->seq, 1687 TCP_SKB_CB(tp->retransmit_skb_hint)->seq)) 1688 1689 tp->retransmit_skb_hint = NULL; 1690 } 1691 } 1692 1693 tp->scoreboard_skb_hint = skb; 1694 1695 tcp_sync_left_out(tp); 1696 } 1697} 1698 1699/* CWND moderation, preventing bursts due to too big ACKs 1700 * in dubious situations. 1701 */ 1702static inline void tcp_moderate_cwnd(struct tcp_sock *tp) 1703{ 1704 tp->snd_cwnd = min(tp->snd_cwnd, 1705 tcp_packets_in_flight(tp)+tcp_max_burst(tp)); 1706 tp->snd_cwnd_stamp = tcp_time_stamp; 1707} 1708 1709/* Lower bound on congestion window is slow start threshold 1710 * unless congestion avoidance choice decides to overide it. 1711 */ 1712static inline u32 tcp_cwnd_min(const struct sock *sk) 1713{ 1714 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 1715 1716 return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh; 1717} 1718 1719/* Decrease cwnd each second ack. */ 1720static void tcp_cwnd_down(struct sock *sk) 1721{ 1722 struct tcp_sock *tp = tcp_sk(sk); 1723 int decr = tp->snd_cwnd_cnt + 1; 1724 1725 tp->snd_cwnd_cnt = decr&1; 1726 decr >>= 1; 1727 1728 if (decr && tp->snd_cwnd > tcp_cwnd_min(sk)) 1729 tp->snd_cwnd -= decr; 1730 1731 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp)+1); 1732 tp->snd_cwnd_stamp = tcp_time_stamp; 1733} 1734 1735/* Nothing was retransmitted or returned timestamp is less 1736 * than timestamp of the first retransmission. 1737 */ 1738static inline int tcp_packet_delayed(struct tcp_sock *tp) 1739{ 1740 return !tp->retrans_stamp || 1741 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 1742 (__s32)(tp->rx_opt.rcv_tsecr - tp->retrans_stamp) < 0); 1743} 1744 1745/* Undo procedures. */ 1746 1747#if FASTRETRANS_DEBUG > 1 1748static void DBGUNDO(struct sock *sk, struct tcp_sock *tp, const char *msg) 1749{ 1750 struct inet_sock *inet = inet_sk(sk); 1751 printk(KERN_DEBUG "Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n", 1752 msg, 1753 NIPQUAD(inet->daddr), ntohs(inet->dport), 1754 tp->snd_cwnd, tp->left_out, 1755 tp->snd_ssthresh, tp->prior_ssthresh, 1756 tp->packets_out); 1757} 1758#else 1759#define DBGUNDO(x...) do { } while (0) 1760#endif 1761 1762static void tcp_undo_cwr(struct sock *sk, const int undo) 1763{ 1764 struct tcp_sock *tp = tcp_sk(sk); 1765 1766 if (tp->prior_ssthresh) { 1767 const struct inet_connection_sock *icsk = inet_csk(sk); 1768 1769 if (icsk->icsk_ca_ops->undo_cwnd) 1770 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk); 1771 else 1772 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh<<1); 1773 1774 if (undo && tp->prior_ssthresh > tp->snd_ssthresh) { 1775 tp->snd_ssthresh = tp->prior_ssthresh; 1776 TCP_ECN_withdraw_cwr(tp); 1777 } 1778 } else { 1779 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh); 1780 } 1781 tcp_moderate_cwnd(tp); 1782 tp->snd_cwnd_stamp = tcp_time_stamp; 1783 1784 /* There is something screwy going on with the retrans hints after 1785 an undo */ 1786 clear_all_retrans_hints(tp); 1787} 1788 1789static inline int tcp_may_undo(struct tcp_sock *tp) 1790{ 1791 return tp->undo_marker && 1792 (!tp->undo_retrans || tcp_packet_delayed(tp)); 1793} 1794 1795/* People celebrate: "We love our President!" */ 1796static int tcp_try_undo_recovery(struct sock *sk, struct tcp_sock *tp) 1797{ 1798 if (tcp_may_undo(tp)) { 1799 /* Happy end! We did not retransmit anything 1800 * or our original transmission succeeded. 1801 */ 1802 DBGUNDO(sk, tp, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans"); 1803 tcp_undo_cwr(sk, 1); 1804 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss) 1805 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO); 1806 else 1807 NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO); 1808 tp->undo_marker = 0; 1809 } 1810 if (tp->snd_una == tp->high_seq && IsReno(tp)) { 1811 /* Hold old state until something *above* high_seq 1812 * is ACKed. For Reno it is MUST to prevent false 1813 * fast retransmits (RFC2582). SACK TCP is safe. */ 1814 tcp_moderate_cwnd(tp); 1815 return 1; 1816 } 1817 tcp_set_ca_state(sk, TCP_CA_Open); 1818 return 0; 1819} 1820 1821/* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */ 1822static void tcp_try_undo_dsack(struct sock *sk, struct tcp_sock *tp) 1823{ 1824 if (tp->undo_marker && !tp->undo_retrans) { 1825 DBGUNDO(sk, tp, "D-SACK"); 1826 tcp_undo_cwr(sk, 1); 1827 tp->undo_marker = 0; 1828 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO); 1829 } 1830} 1831 1832/* Undo during fast recovery after partial ACK. */ 1833 1834static int tcp_try_undo_partial(struct sock *sk, struct tcp_sock *tp, 1835 int acked) 1836{ 1837 /* Partial ACK arrived. Force Hoe's retransmit. */ 1838 int failed = IsReno(tp) || tp->fackets_out>tp->reordering; 1839 1840 if (tcp_may_undo(tp)) { 1841 /* Plain luck! Hole if filled with delayed 1842 * packet, rather than with a retransmit. 1843 */ 1844 if (tp->retrans_out == 0) 1845 tp->retrans_stamp = 0; 1846 1847 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1); 1848 1849 DBGUNDO(sk, tp, "Hoe"); 1850 tcp_undo_cwr(sk, 0); 1851 NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO); 1852 1853 /* So... Do not make Hoe's retransmit yet. 1854 * If the first packet was delayed, the rest 1855 * ones are most probably delayed as well. 1856 */ 1857 failed = 0; 1858 } 1859 return failed; 1860} 1861 1862/* Undo during loss recovery after partial ACK. */ 1863static int tcp_try_undo_loss(struct sock *sk, struct tcp_sock *tp) 1864{ 1865 if (tcp_may_undo(tp)) { 1866 struct sk_buff *skb; 1867 sk_stream_for_retrans_queue(skb, sk) { 1868 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; 1869 } 1870 1871 clear_all_retrans_hints(tp); 1872 1873 DBGUNDO(sk, tp, "partial loss"); 1874 tp->lost_out = 0; 1875 tp->left_out = tp->sacked_out; 1876 tcp_undo_cwr(sk, 1); 1877 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO); 1878 inet_csk(sk)->icsk_retransmits = 0; 1879 tp->undo_marker = 0; 1880 if (!IsReno(tp)) 1881 tcp_set_ca_state(sk, TCP_CA_Open); 1882 return 1; 1883 } 1884 return 0; 1885} 1886 1887static inline void tcp_complete_cwr(struct sock *sk) 1888{ 1889 struct tcp_sock *tp = tcp_sk(sk); 1890 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh); 1891 tp->snd_cwnd_stamp = tcp_time_stamp; 1892 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR); 1893} 1894 1895static void tcp_try_to_open(struct sock *sk, struct tcp_sock *tp, int flag) 1896{ 1897 tp->left_out = tp->sacked_out; 1898 1899 if (tp->retrans_out == 0) 1900 tp->retrans_stamp = 0; 1901 1902 if (flag&FLAG_ECE) 1903 tcp_enter_cwr(sk); 1904 1905 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) { 1906 int state = TCP_CA_Open; 1907 1908 if (tp->left_out || tp->retrans_out || tp->undo_marker) 1909 state = TCP_CA_Disorder; 1910 1911 if (inet_csk(sk)->icsk_ca_state != state) { 1912 tcp_set_ca_state(sk, state); 1913 tp->high_seq = tp->snd_nxt; 1914 } 1915 tcp_moderate_cwnd(tp); 1916 } else { 1917 tcp_cwnd_down(sk); 1918 } 1919} 1920 1921static void tcp_mtup_probe_failed(struct sock *sk) 1922{ 1923 struct inet_connection_sock *icsk = inet_csk(sk); 1924 1925 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1; 1926 icsk->icsk_mtup.probe_size = 0; 1927} 1928 1929static void tcp_mtup_probe_success(struct sock *sk, struct sk_buff *skb) 1930{ 1931 struct tcp_sock *tp = tcp_sk(sk); 1932 struct inet_connection_sock *icsk = inet_csk(sk); 1933 1934 /* FIXME: breaks with very large cwnd */ 1935 tp->prior_ssthresh = tcp_current_ssthresh(sk); 1936 tp->snd_cwnd = tp->snd_cwnd * 1937 tcp_mss_to_mtu(sk, tp->mss_cache) / 1938 icsk->icsk_mtup.probe_size; 1939 tp->snd_cwnd_cnt = 0; 1940 tp->snd_cwnd_stamp = tcp_time_stamp; 1941 tp->rcv_ssthresh = tcp_current_ssthresh(sk); 1942 1943 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size; 1944 icsk->icsk_mtup.probe_size = 0; 1945 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 1946} 1947 1948 1949/* Process an event, which can update packets-in-flight not trivially. 1950 * Main goal of this function is to calculate new estimate for left_out, 1951 * taking into account both packets sitting in receiver's buffer and 1952 * packets lost by network. 1953 * 1954 * Besides that it does CWND reduction, when packet loss is detected 1955 * and changes state of machine. 1956 * 1957 * It does _not_ decide what to send, it is made in function 1958 * tcp_xmit_retransmit_queue(). 1959 */ 1960static void 1961tcp_fastretrans_alert(struct sock *sk, u32 prior_snd_una, 1962 int prior_packets, int flag) 1963{ 1964 struct inet_connection_sock *icsk = inet_csk(sk); 1965 struct tcp_sock *tp = tcp_sk(sk); 1966 int is_dupack = (tp->snd_una == prior_snd_una && !(flag&FLAG_NOT_DUP)); 1967 1968 /* Some technical things: 1969 * 1. Reno does not count dupacks (sacked_out) automatically. */ 1970 if (!tp->packets_out) 1971 tp->sacked_out = 0; 1972 /* 2. SACK counts snd_fack in packets inaccurately. */ 1973 if (tp->sacked_out == 0) 1974 tp->fackets_out = 0; 1975 1976 /* Now state machine starts. 1977 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */ 1978 if (flag&FLAG_ECE) 1979 tp->prior_ssthresh = 0; 1980 1981 /* B. In all the states check for reneging SACKs. */ 1982 if (tp->sacked_out && tcp_check_sack_reneging(sk)) 1983 return; 1984 1985 /* C. Process data loss notification, provided it is valid. */ 1986 if ((flag&FLAG_DATA_LOST) && 1987 before(tp->snd_una, tp->high_seq) && 1988 icsk->icsk_ca_state != TCP_CA_Open && 1989 tp->fackets_out > tp->reordering) { 1990 tcp_mark_head_lost(sk, tp, tp->fackets_out-tp->reordering, tp->high_seq); 1991 NET_INC_STATS_BH(LINUX_MIB_TCPLOSS); 1992 } 1993 1994 /* D. Synchronize left_out to current state. */ 1995 tcp_sync_left_out(tp); 1996 1997 /* E. Check state exit conditions. State can be terminated 1998 * when high_seq is ACKed. */ 1999 if (icsk->icsk_ca_state == TCP_CA_Open) { 2000 if (!sysctl_tcp_frto) 2001 BUG_TRAP(tp->retrans_out == 0); 2002 tp->retrans_stamp = 0; 2003 } else if (!before(tp->snd_una, tp->high_seq)) { 2004 switch (icsk->icsk_ca_state) { 2005 case TCP_CA_Loss: 2006 icsk->icsk_retransmits = 0; 2007 if (tcp_try_undo_recovery(sk, tp)) 2008 return; 2009 break; 2010 2011 case TCP_CA_CWR: 2012 /* CWR is to be held something *above* high_seq 2013 * is ACKed for CWR bit to reach receiver. */ 2014 if (tp->snd_una != tp->high_seq) { 2015 tcp_complete_cwr(sk); 2016 tcp_set_ca_state(sk, TCP_CA_Open); 2017 } 2018 break; 2019 2020 case TCP_CA_Disorder: 2021 tcp_try_undo_dsack(sk, tp); 2022 if (!tp->undo_marker || 2023 /* For SACK case do not Open to allow to undo 2024 * catching for all duplicate ACKs. */ 2025 IsReno(tp) || tp->snd_una != tp->high_seq) { 2026 tp->undo_marker = 0; 2027 tcp_set_ca_state(sk, TCP_CA_Open); 2028 } 2029 break; 2030 2031 case TCP_CA_Recovery: 2032 if (IsReno(tp)) 2033 tcp_reset_reno_sack(tp); 2034 if (tcp_try_undo_recovery(sk, tp)) 2035 return; 2036 tcp_complete_cwr(sk); 2037 break; 2038 } 2039 } 2040 2041 /* F. Process state. */ 2042 switch (icsk->icsk_ca_state) { 2043 case TCP_CA_Recovery: 2044 if (prior_snd_una == tp->snd_una) { 2045 if (IsReno(tp) && is_dupack) 2046 tcp_add_reno_sack(sk); 2047 } else { 2048 int acked = prior_packets - tp->packets_out; 2049 if (IsReno(tp)) 2050 tcp_remove_reno_sacks(sk, tp, acked); 2051 is_dupack = tcp_try_undo_partial(sk, tp, acked); 2052 } 2053 break; 2054 case TCP_CA_Loss: 2055 if (flag&FLAG_DATA_ACKED) 2056 icsk->icsk_retransmits = 0; 2057 if (!tcp_try_undo_loss(sk, tp)) { 2058 tcp_moderate_cwnd(tp); 2059 tcp_xmit_retransmit_queue(sk); 2060 return; 2061 } 2062 if (icsk->icsk_ca_state != TCP_CA_Open) 2063 return; 2064 /* Loss is undone; fall through to processing in Open state. */ 2065 default: 2066 if (IsReno(tp)) { 2067 if (tp->snd_una != prior_snd_una) 2068 tcp_reset_reno_sack(tp); 2069 if (is_dupack) 2070 tcp_add_reno_sack(sk); 2071 } 2072 2073 if (icsk->icsk_ca_state == TCP_CA_Disorder) 2074 tcp_try_undo_dsack(sk, tp); 2075 2076 if (!tcp_time_to_recover(sk, tp)) { 2077 tcp_try_to_open(sk, tp, flag); 2078 return; 2079 } 2080 2081 /* MTU probe failure: don't reduce cwnd */ 2082 if (icsk->icsk_ca_state < TCP_CA_CWR && 2083 icsk->icsk_mtup.probe_size && 2084 tp->snd_una == tp->mtu_probe.probe_seq_start) { 2085 tcp_mtup_probe_failed(sk); 2086 /* Restores the reduction we did in tcp_mtup_probe() */ 2087 tp->snd_cwnd++; 2088 tcp_simple_retransmit(sk); 2089 return; 2090 } 2091 2092 /* Otherwise enter Recovery state */ 2093 2094 if (IsReno(tp)) 2095 NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY); 2096 else 2097 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY); 2098 2099 tp->high_seq = tp->snd_nxt; 2100 tp->prior_ssthresh = 0; 2101 tp->undo_marker = tp->snd_una; 2102 tp->undo_retrans = tp->retrans_out; 2103 2104 if (icsk->icsk_ca_state < TCP_CA_CWR) { 2105 if (!(flag&FLAG_ECE)) 2106 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2107 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 2108 TCP_ECN_queue_cwr(tp); 2109 } 2110 2111 tp->bytes_acked = 0; 2112 tp->snd_cwnd_cnt = 0; 2113 tcp_set_ca_state(sk, TCP_CA_Recovery); 2114 } 2115 2116 if (is_dupack || tcp_head_timedout(sk, tp)) 2117 tcp_update_scoreboard(sk, tp); 2118 tcp_cwnd_down(sk); 2119 tcp_xmit_retransmit_queue(sk); 2120} 2121 2122/* Read draft-ietf-tcplw-high-performance before mucking 2123 * with this code. (Supersedes RFC1323) 2124 */ 2125static void tcp_ack_saw_tstamp(struct sock *sk, int flag) 2126{ 2127 /* RTTM Rule: A TSecr value received in a segment is used to 2128 * update the averaged RTT measurement only if the segment 2129 * acknowledges some new data, i.e., only if it advances the 2130 * left edge of the send window. 2131 * 2132 * See draft-ietf-tcplw-high-performance-00, section 3.3. 2133 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru> 2134 * 2135 * Changed: reset backoff as soon as we see the first valid sample. 2136 * If we do not, we get strongly overestimated rto. With timestamps 2137 * samples are accepted even from very old segments: f.e., when rtt=1 2138 * increases to 8, we retransmit 5 times and after 8 seconds delayed 2139 * answer arrives rto becomes 120 seconds! If at least one of segments 2140 * in window is lost... Voila. --ANK (010210) 2141 */ 2142 struct tcp_sock *tp = tcp_sk(sk); 2143 const __u32 seq_rtt = tcp_time_stamp - tp->rx_opt.rcv_tsecr; 2144 tcp_rtt_estimator(sk, seq_rtt); 2145 tcp_set_rto(sk); 2146 inet_csk(sk)->icsk_backoff = 0; 2147 tcp_bound_rto(sk); 2148} 2149 2150static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag) 2151{ 2152 /* We don't have a timestamp. Can only use 2153 * packets that are not retransmitted to determine 2154 * rtt estimates. Also, we must not reset the 2155 * backoff for rto until we get a non-retransmitted 2156 * packet. This allows us to deal with a situation 2157 * where the network delay has increased suddenly. 2158 * I.e. Karn's algorithm. (SIGCOMM '87, p5.) 2159 */ 2160 2161 if (flag & FLAG_RETRANS_DATA_ACKED) 2162 return; 2163 2164 tcp_rtt_estimator(sk, seq_rtt); 2165 tcp_set_rto(sk); 2166 inet_csk(sk)->icsk_backoff = 0; 2167 tcp_bound_rto(sk); 2168} 2169 2170static inline void tcp_ack_update_rtt(struct sock *sk, const int flag, 2171 const s32 seq_rtt) 2172{ 2173 const struct tcp_sock *tp = tcp_sk(sk); 2174 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */ 2175 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr) 2176 tcp_ack_saw_tstamp(sk, flag); 2177 else if (seq_rtt >= 0) 2178 tcp_ack_no_tstamp(sk, seq_rtt, flag); 2179} 2180 2181static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 rtt, 2182 u32 in_flight, int good) 2183{ 2184 const struct inet_connection_sock *icsk = inet_csk(sk); 2185 icsk->icsk_ca_ops->cong_avoid(sk, ack, rtt, in_flight, good); 2186 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp; 2187} 2188 2189/* Restart timer after forward progress on connection. 2190 * RFC2988 recommends to restart timer to now+rto. 2191 */ 2192 2193static void tcp_ack_packets_out(struct sock *sk, struct tcp_sock *tp) 2194{ 2195 if (!tp->packets_out) { 2196 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS); 2197 } else { 2198 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 2199 } 2200} 2201 2202static int tcp_tso_acked(struct sock *sk, struct sk_buff *skb, 2203 __u32 now, __s32 *seq_rtt) 2204{ 2205 struct tcp_sock *tp = tcp_sk(sk); 2206 struct tcp_skb_cb *scb = TCP_SKB_CB(skb); 2207 __u32 seq = tp->snd_una; 2208 __u32 packets_acked; 2209 int acked = 0; 2210 2211 /* If we get here, the whole TSO packet has not been 2212 * acked. 2213 */ 2214 BUG_ON(!after(scb->end_seq, seq)); 2215 2216 packets_acked = tcp_skb_pcount(skb); 2217 if (tcp_trim_head(sk, skb, seq - scb->seq)) 2218 return 0; 2219 packets_acked -= tcp_skb_pcount(skb); 2220 2221 if (packets_acked) { 2222 __u8 sacked = scb->sacked; 2223 2224 acked |= FLAG_DATA_ACKED; 2225 if (sacked) { 2226 if (sacked & TCPCB_RETRANS) { 2227 if (sacked & TCPCB_SACKED_RETRANS) 2228 tp->retrans_out -= packets_acked; 2229 acked |= FLAG_RETRANS_DATA_ACKED; 2230 *seq_rtt = -1; 2231 } else if (*seq_rtt < 0) 2232 *seq_rtt = now - scb->when; 2233 if (sacked & TCPCB_SACKED_ACKED) 2234 tp->sacked_out -= packets_acked; 2235 if (sacked & TCPCB_LOST) 2236 tp->lost_out -= packets_acked; 2237 if (sacked & TCPCB_URG) { 2238 if (tp->urg_mode && 2239 !before(seq, tp->snd_up)) 2240 tp->urg_mode = 0; 2241 } 2242 } else if (*seq_rtt < 0) 2243 *seq_rtt = now - scb->when; 2244 2245 if (tp->fackets_out) { 2246 __u32 dval = min(tp->fackets_out, packets_acked); 2247 tp->fackets_out -= dval; 2248 } 2249 tp->packets_out -= packets_acked; 2250 2251 BUG_ON(tcp_skb_pcount(skb) == 0); 2252 BUG_ON(!before(scb->seq, scb->end_seq)); 2253 } 2254 2255 return acked; 2256} 2257 2258static u32 tcp_usrtt(struct timeval *tv) 2259{ 2260 struct timeval now; 2261 2262 do_gettimeofday(&now); 2263 return (now.tv_sec - tv->tv_sec) * 1000000 + (now.tv_usec - tv->tv_usec); 2264} 2265 2266/* Remove acknowledged frames from the retransmission queue. */ 2267static int tcp_clean_rtx_queue(struct sock *sk, __s32 *seq_rtt_p) 2268{ 2269 struct tcp_sock *tp = tcp_sk(sk); 2270 const struct inet_connection_sock *icsk = inet_csk(sk); 2271 struct sk_buff *skb; 2272 __u32 now = tcp_time_stamp; 2273 int acked = 0; 2274 __s32 seq_rtt = -1; 2275 u32 pkts_acked = 0; 2276 void (*rtt_sample)(struct sock *sk, u32 usrtt) 2277 = icsk->icsk_ca_ops->rtt_sample; 2278 struct timeval tv = { .tv_sec = 0, .tv_usec = 0 }; 2279 2280 while ((skb = skb_peek(&sk->sk_write_queue)) && 2281 skb != sk->sk_send_head) { 2282 struct tcp_skb_cb *scb = TCP_SKB_CB(skb); 2283 __u8 sacked = scb->sacked; 2284 2285 /* If our packet is before the ack sequence we can 2286 * discard it as it's confirmed to have arrived at 2287 * the other end. 2288 */ 2289 if (after(scb->end_seq, tp->snd_una)) { 2290 if (tcp_skb_pcount(skb) > 1 && 2291 after(tp->snd_una, scb->seq)) 2292 acked |= tcp_tso_acked(sk, skb, 2293 now, &seq_rtt); 2294 break; 2295 } 2296 2297 /* Initial outgoing SYN's get put onto the write_queue 2298 * just like anything else we transmit. It is not 2299 * true data, and if we misinform our callers that 2300 * this ACK acks real data, we will erroneously exit 2301 * connection startup slow start one packet too 2302 * quickly. This is severely frowned upon behavior. 2303 */ 2304 if (!(scb->flags & TCPCB_FLAG_SYN)) { 2305 acked |= FLAG_DATA_ACKED; 2306 ++pkts_acked; 2307 } else { 2308 acked |= FLAG_SYN_ACKED; 2309 tp->retrans_stamp = 0; 2310 } 2311 2312 /* MTU probing checks */ 2313 if (icsk->icsk_mtup.probe_size) { 2314 if (!after(tp->mtu_probe.probe_seq_end, TCP_SKB_CB(skb)->end_seq)) { 2315 tcp_mtup_probe_success(sk, skb); 2316 } 2317 } 2318 2319 if (sacked) { 2320 if (sacked & TCPCB_RETRANS) { 2321 if(sacked & TCPCB_SACKED_RETRANS) 2322 tp->retrans_out -= tcp_skb_pcount(skb); 2323 acked |= FLAG_RETRANS_DATA_ACKED; 2324 seq_rtt = -1; 2325 } else if (seq_rtt < 0) { 2326 seq_rtt = now - scb->when; 2327 skb_get_timestamp(skb, &tv); 2328 } 2329 if (sacked & TCPCB_SACKED_ACKED) 2330 tp->sacked_out -= tcp_skb_pcount(skb); 2331 if (sacked & TCPCB_LOST) 2332 tp->lost_out -= tcp_skb_pcount(skb); 2333 if (sacked & TCPCB_URG) { 2334 if (tp->urg_mode && 2335 !before(scb->end_seq, tp->snd_up)) 2336 tp->urg_mode = 0; 2337 } 2338 } else if (seq_rtt < 0) { 2339 seq_rtt = now - scb->when; 2340 skb_get_timestamp(skb, &tv); 2341 } 2342 tcp_dec_pcount_approx(&tp->fackets_out, skb); 2343 tcp_packets_out_dec(tp, skb); 2344 __skb_unlink(skb, &sk->sk_write_queue); 2345 sk_stream_free_skb(sk, skb); 2346 clear_all_retrans_hints(tp); 2347 } 2348 2349 if (acked&FLAG_ACKED) { 2350 tcp_ack_update_rtt(sk, acked, seq_rtt); 2351 tcp_ack_packets_out(sk, tp); 2352 if (rtt_sample && !(acked & FLAG_RETRANS_DATA_ACKED)) 2353 (*rtt_sample)(sk, tcp_usrtt(&tv)); 2354 2355 if (icsk->icsk_ca_ops->pkts_acked) 2356 icsk->icsk_ca_ops->pkts_acked(sk, pkts_acked); 2357 } 2358 2359#if FASTRETRANS_DEBUG > 0 2360 BUG_TRAP((int)tp->sacked_out >= 0); 2361 BUG_TRAP((int)tp->lost_out >= 0); 2362 BUG_TRAP((int)tp->retrans_out >= 0); 2363 if (!tp->packets_out && tp->rx_opt.sack_ok) { 2364 const struct inet_connection_sock *icsk = inet_csk(sk); 2365 if (tp->lost_out) { 2366 printk(KERN_DEBUG "Leak l=%u %d\n", 2367 tp->lost_out, icsk->icsk_ca_state); 2368 tp->lost_out = 0; 2369 } 2370 if (tp->sacked_out) { 2371 printk(KERN_DEBUG "Leak s=%u %d\n", 2372 tp->sacked_out, icsk->icsk_ca_state); 2373 tp->sacked_out = 0; 2374 } 2375 if (tp->retrans_out) { 2376 printk(KERN_DEBUG "Leak r=%u %d\n", 2377 tp->retrans_out, icsk->icsk_ca_state); 2378 tp->retrans_out = 0; 2379 } 2380 } 2381#endif 2382 *seq_rtt_p = seq_rtt; 2383 return acked; 2384} 2385 2386static void tcp_ack_probe(struct sock *sk) 2387{ 2388 const struct tcp_sock *tp = tcp_sk(sk); 2389 struct inet_connection_sock *icsk = inet_csk(sk); 2390 2391 /* Was it a usable window open? */ 2392 2393 if (!after(TCP_SKB_CB(sk->sk_send_head)->end_seq, 2394 tp->snd_una + tp->snd_wnd)) { 2395 icsk->icsk_backoff = 0; 2396 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0); 2397 /* Socket must be waked up by subsequent tcp_data_snd_check(). 2398 * This function is not for random using! 2399 */ 2400 } else { 2401 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 2402 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), 2403 TCP_RTO_MAX); 2404 } 2405} 2406 2407static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag) 2408{ 2409 return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) || 2410 inet_csk(sk)->icsk_ca_state != TCP_CA_Open); 2411} 2412 2413static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag) 2414{ 2415 const struct tcp_sock *tp = tcp_sk(sk); 2416 return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) && 2417 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR)); 2418} 2419 2420/* Check that window update is acceptable. 2421 * The function assumes that snd_una<=ack<=snd_next. 2422 */ 2423static inline int tcp_may_update_window(const struct tcp_sock *tp, const u32 ack, 2424 const u32 ack_seq, const u32 nwin) 2425{ 2426 return (after(ack, tp->snd_una) || 2427 after(ack_seq, tp->snd_wl1) || 2428 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd)); 2429} 2430 2431/* Update our send window. 2432 * 2433 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2 2434 * and in FreeBSD. NetBSD's one is even worse.) is wrong. 2435 */ 2436static int tcp_ack_update_window(struct sock *sk, struct tcp_sock *tp, 2437 struct sk_buff *skb, u32 ack, u32 ack_seq) 2438{ 2439 int flag = 0; 2440 u32 nwin = ntohs(skb->h.th->window); 2441 2442 if (likely(!skb->h.th->syn)) 2443 nwin <<= tp->rx_opt.snd_wscale; 2444 2445 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) { 2446 flag |= FLAG_WIN_UPDATE; 2447 tcp_update_wl(tp, ack, ack_seq); 2448 2449 if (tp->snd_wnd != nwin) { 2450 tp->snd_wnd = nwin; 2451 2452 /* Note, it is the only place, where 2453 * fast path is recovered for sending TCP. 2454 */ 2455 tp->pred_flags = 0; 2456 tcp_fast_path_check(sk, tp); 2457 2458 if (nwin > tp->max_window) { 2459 tp->max_window = nwin; 2460 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie); 2461 } 2462 } 2463 } 2464 2465 tp->snd_una = ack; 2466 2467 return flag; 2468} 2469 2470/* A very conservative spurious RTO response algorithm: reduce cwnd and 2471 * continue in congestion avoidance. 2472 */ 2473static void tcp_conservative_spur_to_response(struct tcp_sock *tp) 2474{ 2475 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh); 2476 tcp_moderate_cwnd(tp); 2477} 2478 2479static void tcp_process_frto(struct sock *sk, u32 prior_snd_una) 2480{ 2481 struct tcp_sock *tp = tcp_sk(sk); 2482 2483 tcp_sync_left_out(tp); 2484 2485 if (tp->snd_una == prior_snd_una || 2486 !before(tp->snd_una, tp->frto_highmark)) { 2487 /* RTO was caused by loss, start retransmitting in 2488 * go-back-N slow start 2489 */ 2490 tcp_enter_frto_loss(sk); 2491 return; 2492 } 2493 2494 if (tp->frto_counter == 1) { 2495 /* First ACK after RTO advances the window: allow two new 2496 * segments out. 2497 */ 2498 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2; 2499 } else { 2500 tcp_conservative_spur_to_response(tp); 2501 } 2502 2503 /* F-RTO affects on two new ACKs following RTO. 2504 * At latest on third ACK the TCP behavior is back to normal. 2505 */ 2506 tp->frto_counter = (tp->frto_counter + 1) % 3; 2507} 2508 2509/* This routine deals with incoming acks, but not outgoing ones. */ 2510static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag) 2511{ 2512 struct inet_connection_sock *icsk = inet_csk(sk); 2513 struct tcp_sock *tp = tcp_sk(sk); 2514 u32 prior_snd_una = tp->snd_una; 2515 u32 ack_seq = TCP_SKB_CB(skb)->seq; 2516 u32 ack = TCP_SKB_CB(skb)->ack_seq; 2517 u32 prior_in_flight; 2518 s32 seq_rtt; 2519 int prior_packets; 2520 2521 /* If the ack is newer than sent or older than previous acks 2522 * then we can probably ignore it. 2523 */ 2524 if (after(ack, tp->snd_nxt)) 2525 goto uninteresting_ack; 2526 2527 if (before(ack, prior_snd_una)) 2528 goto old_ack; 2529 2530 if (sysctl_tcp_abc) { 2531 if (icsk->icsk_ca_state < TCP_CA_CWR) 2532 tp->bytes_acked += ack - prior_snd_una; 2533 else if (icsk->icsk_ca_state == TCP_CA_Loss) 2534 /* we assume just one segment left network */ 2535 tp->bytes_acked += min(ack - prior_snd_una, tp->mss_cache); 2536 } 2537 2538 if (!(flag&FLAG_SLOWPATH) && after(ack, prior_snd_una)) { 2539 /* Window is constant, pure forward advance. 2540 * No more checks are required. 2541 * Note, we use the fact that SND.UNA>=SND.WL2. 2542 */ 2543 tcp_update_wl(tp, ack, ack_seq); 2544 tp->snd_una = ack; 2545 flag |= FLAG_WIN_UPDATE; 2546 2547 tcp_ca_event(sk, CA_EVENT_FAST_ACK); 2548 2549 NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS); 2550 } else { 2551 if (ack_seq != TCP_SKB_CB(skb)->end_seq) 2552 flag |= FLAG_DATA; 2553 else 2554 NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS); 2555 2556 flag |= tcp_ack_update_window(sk, tp, skb, ack, ack_seq); 2557 2558 if (TCP_SKB_CB(skb)->sacked) 2559 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una); 2560 2561 if (TCP_ECN_rcv_ecn_echo(tp, skb->h.th)) 2562 flag |= FLAG_ECE; 2563 2564 tcp_ca_event(sk, CA_EVENT_SLOW_ACK); 2565 } 2566 2567 /* We passed data and got it acked, remove any soft error 2568 * log. Something worked... 2569 */ 2570 sk->sk_err_soft = 0; 2571 tp->rcv_tstamp = tcp_time_stamp; 2572 prior_packets = tp->packets_out; 2573 if (!prior_packets) 2574 goto no_queue; 2575 2576 prior_in_flight = tcp_packets_in_flight(tp); 2577 2578 /* See if we can take anything off of the retransmit queue. */ 2579 flag |= tcp_clean_rtx_queue(sk, &seq_rtt); 2580 2581 if (tp->frto_counter) 2582 tcp_process_frto(sk, prior_snd_una); 2583 2584 if (tcp_ack_is_dubious(sk, flag)) { 2585 /* Advance CWND, if state allows this. */ 2586 if ((flag & FLAG_DATA_ACKED) && tcp_may_raise_cwnd(sk, flag)) 2587 tcp_cong_avoid(sk, ack, seq_rtt, prior_in_flight, 0); 2588 tcp_fastretrans_alert(sk, prior_snd_una, prior_packets, flag); 2589 } else { 2590 if ((flag & FLAG_DATA_ACKED)) 2591 tcp_cong_avoid(sk, ack, seq_rtt, prior_in_flight, 1); 2592 } 2593 2594 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag&FLAG_NOT_DUP)) 2595 dst_confirm(sk->sk_dst_cache); 2596 2597 return 1; 2598 2599no_queue: 2600 icsk->icsk_probes_out = 0; 2601 2602 /* If this ack opens up a zero window, clear backoff. It was 2603 * being used to time the probes, and is probably far higher than 2604 * it needs to be for normal retransmission. 2605 */ 2606 if (sk->sk_send_head) 2607 tcp_ack_probe(sk); 2608 return 1; 2609 2610old_ack: 2611 if (TCP_SKB_CB(skb)->sacked) 2612 tcp_sacktag_write_queue(sk, skb, prior_snd_una); 2613 2614uninteresting_ack: 2615 SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt); 2616 return 0; 2617} 2618 2619 2620/* Look for tcp options. Normally only called on SYN and SYNACK packets. 2621 * But, this can also be called on packets in the established flow when 2622 * the fast version below fails. 2623 */ 2624void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx, int estab) 2625{ 2626 unsigned char *ptr; 2627 struct tcphdr *th = skb->h.th; 2628 int length=(th->doff*4)-sizeof(struct tcphdr); 2629 2630 ptr = (unsigned char *)(th + 1); 2631 opt_rx->saw_tstamp = 0; 2632 2633 while(length>0) { 2634 int opcode=*ptr++; 2635 int opsize; 2636 2637 switch (opcode) { 2638 case TCPOPT_EOL: 2639 return; 2640 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */ 2641 length--; 2642 continue; 2643 default: 2644 opsize=*ptr++; 2645 if (opsize < 2) /* "silly options" */ 2646 return; 2647 if (opsize > length) 2648 return; /* don't parse partial options */ 2649 switch(opcode) { 2650 case TCPOPT_MSS: 2651 if(opsize==TCPOLEN_MSS && th->syn && !estab) { 2652 u16 in_mss = ntohs(get_unaligned((__be16 *)ptr)); 2653 if (in_mss) { 2654 if (opt_rx->user_mss && opt_rx->user_mss < in_mss) 2655 in_mss = opt_rx->user_mss; 2656 opt_rx->mss_clamp = in_mss; 2657 } 2658 } 2659 break; 2660 case TCPOPT_WINDOW: 2661 if(opsize==TCPOLEN_WINDOW && th->syn && !estab) 2662 if (sysctl_tcp_window_scaling) { 2663 __u8 snd_wscale = *(__u8 *) ptr; 2664 opt_rx->wscale_ok = 1; 2665 if (snd_wscale > 14) { 2666 if(net_ratelimit()) 2667 printk(KERN_INFO "tcp_parse_options: Illegal window " 2668 "scaling value %d >14 received.\n", 2669 snd_wscale); 2670 snd_wscale = 14; 2671 } 2672 opt_rx->snd_wscale = snd_wscale; 2673 } 2674 break; 2675 case TCPOPT_TIMESTAMP: 2676 if(opsize==TCPOLEN_TIMESTAMP) { 2677 if ((estab && opt_rx->tstamp_ok) || 2678 (!estab && sysctl_tcp_timestamps)) { 2679 opt_rx->saw_tstamp = 1; 2680 opt_rx->rcv_tsval = ntohl(get_unaligned((__be32 *)ptr)); 2681 opt_rx->rcv_tsecr = ntohl(get_unaligned((__be32 *)(ptr+4))); 2682 } 2683 } 2684 break; 2685 case TCPOPT_SACK_PERM: 2686 if(opsize==TCPOLEN_SACK_PERM && th->syn && !estab) { 2687 if (sysctl_tcp_sack) { 2688 opt_rx->sack_ok = 1; 2689 tcp_sack_reset(opt_rx); 2690 } 2691 } 2692 break; 2693 2694 case TCPOPT_SACK: 2695 if((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) && 2696 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) && 2697 opt_rx->sack_ok) { 2698 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th; 2699 } 2700#ifdef CONFIG_TCP_MD5SIG 2701 case TCPOPT_MD5SIG: 2702 /* 2703 * The MD5 Hash has already been 2704 * checked (see tcp_v{4,6}_do_rcv()). 2705 */ 2706 break; 2707#endif 2708 }; 2709 ptr+=opsize-2; 2710 length-=opsize; 2711 }; 2712 } 2713} 2714 2715/* Fast parse options. This hopes to only see timestamps. 2716 * If it is wrong it falls back on tcp_parse_options(). 2717 */ 2718static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th, 2719 struct tcp_sock *tp) 2720{ 2721 if (th->doff == sizeof(struct tcphdr)>>2) { 2722 tp->rx_opt.saw_tstamp = 0; 2723 return 0; 2724 } else if (tp->rx_opt.tstamp_ok && 2725 th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) { 2726 __be32 *ptr = (__be32 *)(th + 1); 2727 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) 2728 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) { 2729 tp->rx_opt.saw_tstamp = 1; 2730 ++ptr; 2731 tp->rx_opt.rcv_tsval = ntohl(*ptr); 2732 ++ptr; 2733 tp->rx_opt.rcv_tsecr = ntohl(*ptr); 2734 return 1; 2735 } 2736 } 2737 tcp_parse_options(skb, &tp->rx_opt, 1); 2738 return 1; 2739} 2740 2741static inline void tcp_store_ts_recent(struct tcp_sock *tp) 2742{ 2743 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval; 2744 tp->rx_opt.ts_recent_stamp = xtime.tv_sec; 2745} 2746 2747static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq) 2748{ 2749 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) { 2750 /* PAWS bug workaround wrt. ACK frames, the PAWS discard 2751 * extra check below makes sure this can only happen 2752 * for pure ACK frames. -DaveM 2753 * 2754 * Not only, also it occurs for expired timestamps. 2755 */ 2756 2757 if((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) >= 0 || 2758 xtime.tv_sec >= tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS) 2759 tcp_store_ts_recent(tp); 2760 } 2761} 2762 2763/* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM 2764 * 2765 * It is not fatal. If this ACK does _not_ change critical state (seqs, window) 2766 * it can pass through stack. So, the following predicate verifies that 2767 * this segment is not used for anything but congestion avoidance or 2768 * fast retransmit. Moreover, we even are able to eliminate most of such 2769 * second order effects, if we apply some small "replay" window (~RTO) 2770 * to timestamp space. 2771 * 2772 * All these measures still do not guarantee that we reject wrapped ACKs 2773 * on networks with high bandwidth, when sequence space is recycled fastly, 2774 * but it guarantees that such events will be very rare and do not affect 2775 * connection seriously. This doesn't look nice, but alas, PAWS is really 2776 * buggy extension. 2777 * 2778 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC 2779 * states that events when retransmit arrives after original data are rare. 2780 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is 2781 * the biggest problem on large power networks even with minor reordering. 2782 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe 2783 * up to bandwidth of 18Gigabit/sec. 8) ] 2784 */ 2785 2786static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb) 2787{ 2788 struct tcp_sock *tp = tcp_sk(sk); 2789 struct tcphdr *th = skb->h.th; 2790 u32 seq = TCP_SKB_CB(skb)->seq; 2791 u32 ack = TCP_SKB_CB(skb)->ack_seq; 2792 2793 return (/* 1. Pure ACK with correct sequence number. */ 2794 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) && 2795 2796 /* 2. ... and duplicate ACK. */ 2797 ack == tp->snd_una && 2798 2799 /* 3. ... and does not update window. */ 2800 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) && 2801 2802 /* 4. ... and sits in replay window. */ 2803 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ); 2804} 2805 2806static inline int tcp_paws_discard(const struct sock *sk, const struct sk_buff *skb) 2807{ 2808 const struct tcp_sock *tp = tcp_sk(sk); 2809 return ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > TCP_PAWS_WINDOW && 2810 xtime.tv_sec < tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS && 2811 !tcp_disordered_ack(sk, skb)); 2812} 2813 2814/* Check segment sequence number for validity. 2815 * 2816 * Segment controls are considered valid, if the segment 2817 * fits to the window after truncation to the window. Acceptability 2818 * of data (and SYN, FIN, of course) is checked separately. 2819 * See tcp_data_queue(), for example. 2820 * 2821 * Also, controls (RST is main one) are accepted using RCV.WUP instead 2822 * of RCV.NXT. Peer still did not advance his SND.UNA when we 2823 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP. 2824 * (borrowed from freebsd) 2825 */ 2826 2827static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq) 2828{ 2829 return !before(end_seq, tp->rcv_wup) && 2830 !after(seq, tp->rcv_nxt + tcp_receive_window(tp)); 2831} 2832 2833/* When we get a reset we do this. */ 2834static void tcp_reset(struct sock *sk) 2835{ 2836 /* We want the right error as BSD sees it (and indeed as we do). */ 2837 switch (sk->sk_state) { 2838 case TCP_SYN_SENT: 2839 sk->sk_err = ECONNREFUSED; 2840 break; 2841 case TCP_CLOSE_WAIT: 2842 sk->sk_err = EPIPE; 2843 break; 2844 case TCP_CLOSE: 2845 return; 2846 default: 2847 sk->sk_err = ECONNRESET; 2848 } 2849 2850 if (!sock_flag(sk, SOCK_DEAD)) 2851 sk->sk_error_report(sk); 2852 2853 tcp_done(sk); 2854} 2855 2856/* 2857 * Process the FIN bit. This now behaves as it is supposed to work 2858 * and the FIN takes effect when it is validly part of sequence 2859 * space. Not before when we get holes. 2860 * 2861 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT 2862 * (and thence onto LAST-ACK and finally, CLOSE, we never enter 2863 * TIME-WAIT) 2864 * 2865 * If we are in FINWAIT-1, a received FIN indicates simultaneous 2866 * close and we go into CLOSING (and later onto TIME-WAIT) 2867 * 2868 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT. 2869 */ 2870static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th) 2871{ 2872 struct tcp_sock *tp = tcp_sk(sk); 2873 2874 inet_csk_schedule_ack(sk); 2875 2876 sk->sk_shutdown |= RCV_SHUTDOWN; 2877 sock_set_flag(sk, SOCK_DONE); 2878 2879 switch (sk->sk_state) { 2880 case TCP_SYN_RECV: 2881 case TCP_ESTABLISHED: 2882 /* Move to CLOSE_WAIT */ 2883 tcp_set_state(sk, TCP_CLOSE_WAIT); 2884 inet_csk(sk)->icsk_ack.pingpong = 1; 2885 break; 2886 2887 case TCP_CLOSE_WAIT: 2888 case TCP_CLOSING: 2889 /* Received a retransmission of the FIN, do 2890 * nothing. 2891 */ 2892 break; 2893 case TCP_LAST_ACK: 2894 /* RFC793: Remain in the LAST-ACK state. */ 2895 break; 2896 2897 case TCP_FIN_WAIT1: 2898 /* This case occurs when a simultaneous close 2899 * happens, we must ack the received FIN and 2900 * enter the CLOSING state. 2901 */ 2902 tcp_send_ack(sk); 2903 tcp_set_state(sk, TCP_CLOSING); 2904 break; 2905 case TCP_FIN_WAIT2: 2906 /* Received a FIN -- send ACK and enter TIME_WAIT. */ 2907 tcp_send_ack(sk); 2908 tcp_time_wait(sk, TCP_TIME_WAIT, 0); 2909 break; 2910 default: 2911 /* Only TCP_LISTEN and TCP_CLOSE are left, in these 2912 * cases we should never reach this piece of code. 2913 */ 2914 printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n", 2915 __FUNCTION__, sk->sk_state); 2916 break; 2917 }; 2918 2919 /* It _is_ possible, that we have something out-of-order _after_ FIN. 2920 * Probably, we should reset in this case. For now drop them. 2921 */ 2922 __skb_queue_purge(&tp->out_of_order_queue); 2923 if (tp->rx_opt.sack_ok) 2924 tcp_sack_reset(&tp->rx_opt); 2925 sk_stream_mem_reclaim(sk); 2926 2927 if (!sock_flag(sk, SOCK_DEAD)) { 2928 sk->sk_state_change(sk); 2929 2930 /* Do not send POLL_HUP for half duplex close. */ 2931 if (sk->sk_shutdown == SHUTDOWN_MASK || 2932 sk->sk_state == TCP_CLOSE) 2933 sk_wake_async(sk, 1, POLL_HUP); 2934 else 2935 sk_wake_async(sk, 1, POLL_IN); 2936 } 2937} 2938 2939static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, u32 end_seq) 2940{ 2941 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) { 2942 if (before(seq, sp->start_seq)) 2943 sp->start_seq = seq; 2944 if (after(end_seq, sp->end_seq)) 2945 sp->end_seq = end_seq; 2946 return 1; 2947 } 2948 return 0; 2949} 2950 2951static void tcp_dsack_set(struct tcp_sock *tp, u32 seq, u32 end_seq) 2952{ 2953 if (tp->rx_opt.sack_ok && sysctl_tcp_dsack) { 2954 if (before(seq, tp->rcv_nxt)) 2955 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT); 2956 else 2957 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT); 2958 2959 tp->rx_opt.dsack = 1; 2960 tp->duplicate_sack[0].start_seq = seq; 2961 tp->duplicate_sack[0].end_seq = end_seq; 2962 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + 1, 4 - tp->rx_opt.tstamp_ok); 2963 } 2964} 2965 2966static void tcp_dsack_extend(struct tcp_sock *tp, u32 seq, u32 end_seq) 2967{ 2968 if (!tp->rx_opt.dsack) 2969 tcp_dsack_set(tp, seq, end_seq); 2970 else 2971 tcp_sack_extend(tp->duplicate_sack, seq, end_seq); 2972} 2973 2974static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb) 2975{ 2976 struct tcp_sock *tp = tcp_sk(sk); 2977 2978 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 2979 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 2980 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST); 2981 tcp_enter_quickack_mode(sk); 2982 2983 if (tp->rx_opt.sack_ok && sysctl_tcp_dsack) { 2984 u32 end_seq = TCP_SKB_CB(skb)->end_seq; 2985 2986 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) 2987 end_seq = tp->rcv_nxt; 2988 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, end_seq); 2989 } 2990 } 2991 2992 tcp_send_ack(sk); 2993} 2994 2995/* These routines update the SACK block as out-of-order packets arrive or 2996 * in-order packets close up the sequence space. 2997 */ 2998static void tcp_sack_maybe_coalesce(struct tcp_sock *tp) 2999{ 3000 int this_sack; 3001 struct tcp_sack_block *sp = &tp->selective_acks[0]; 3002 struct tcp_sack_block *swalk = sp+1; 3003 3004 /* See if the recent change to the first SACK eats into 3005 * or hits the sequence space of other SACK blocks, if so coalesce. 3006 */ 3007 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks; ) { 3008 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) { 3009 int i; 3010 3011 /* Zap SWALK, by moving every further SACK up by one slot. 3012 * Decrease num_sacks. 3013 */ 3014 tp->rx_opt.num_sacks--; 3015 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok); 3016 for(i=this_sack; i < tp->rx_opt.num_sacks; i++) 3017 sp[i] = sp[i+1]; 3018 continue; 3019 } 3020 this_sack++, swalk++; 3021 } 3022} 3023 3024static inline void tcp_sack_swap(struct tcp_sack_block *sack1, struct tcp_sack_block *sack2) 3025{ 3026 __u32 tmp; 3027 3028 tmp = sack1->start_seq; 3029 sack1->start_seq = sack2->start_seq; 3030 sack2->start_seq = tmp; 3031 3032 tmp = sack1->end_seq; 3033 sack1->end_seq = sack2->end_seq; 3034 sack2->end_seq = tmp; 3035} 3036 3037static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq) 3038{ 3039 struct tcp_sock *tp = tcp_sk(sk); 3040 struct tcp_sack_block *sp = &tp->selective_acks[0]; 3041 int cur_sacks = tp->rx_opt.num_sacks; 3042 int this_sack; 3043 3044 if (!cur_sacks) 3045 goto new_sack; 3046 3047 for (this_sack=0; this_sack<cur_sacks; this_sack++, sp++) { 3048 if (tcp_sack_extend(sp, seq, end_seq)) { 3049 /* Rotate this_sack to the first one. */ 3050 for (; this_sack>0; this_sack--, sp--) 3051 tcp_sack_swap(sp, sp-1); 3052 if (cur_sacks > 1) 3053 tcp_sack_maybe_coalesce(tp); 3054 return; 3055 } 3056 } 3057 3058 /* Could not find an adjacent existing SACK, build a new one, 3059 * put it at the front, and shift everyone else down. We 3060 * always know there is at least one SACK present already here. 3061 * 3062 * If the sack array is full, forget about the last one. 3063 */ 3064 if (this_sack >= 4) { 3065 this_sack--; 3066 tp->rx_opt.num_sacks--; 3067 sp--; 3068 } 3069 for(; this_sack > 0; this_sack--, sp--) 3070 *sp = *(sp-1); 3071 3072new_sack: 3073 /* Build the new head SACK, and we're done. */ 3074 sp->start_seq = seq; 3075 sp->end_seq = end_seq; 3076 tp->rx_opt.num_sacks++; 3077 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok); 3078} 3079 3080/* RCV.NXT advances, some SACKs should be eaten. */ 3081 3082static void tcp_sack_remove(struct tcp_sock *tp) 3083{ 3084 struct tcp_sack_block *sp = &tp->selective_acks[0]; 3085 int num_sacks = tp->rx_opt.num_sacks; 3086 int this_sack; 3087 3088 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */ 3089 if (skb_queue_empty(&tp->out_of_order_queue)) { 3090 tp->rx_opt.num_sacks = 0; 3091 tp->rx_opt.eff_sacks = tp->rx_opt.dsack; 3092 return; 3093 } 3094 3095 for(this_sack = 0; this_sack < num_sacks; ) { 3096 /* Check if the start of the sack is covered by RCV.NXT. */ 3097 if (!before(tp->rcv_nxt, sp->start_seq)) { 3098 int i; 3099 3100 /* RCV.NXT must cover all the block! */ 3101 BUG_TRAP(!before(tp->rcv_nxt, sp->end_seq)); 3102 3103 /* Zap this SACK, by moving forward any other SACKS. */ 3104 for (i=this_sack+1; i < num_sacks; i++) 3105 tp->selective_acks[i-1] = tp->selective_acks[i]; 3106 num_sacks--; 3107 continue; 3108 } 3109 this_sack++; 3110 sp++; 3111 } 3112 if (num_sacks != tp->rx_opt.num_sacks) { 3113 tp->rx_opt.num_sacks = num_sacks; 3114 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok); 3115 } 3116} 3117 3118/* This one checks to see if we can put data from the 3119 * out_of_order queue into the receive_queue. 3120 */ 3121static void tcp_ofo_queue(struct sock *sk) 3122{ 3123 struct tcp_sock *tp = tcp_sk(sk); 3124 __u32 dsack_high = tp->rcv_nxt; 3125 struct sk_buff *skb; 3126 3127 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) { 3128 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) 3129 break; 3130 3131 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) { 3132 __u32 dsack = dsack_high; 3133 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high)) 3134 dsack_high = TCP_SKB_CB(skb)->end_seq; 3135 tcp_dsack_extend(tp, TCP_SKB_CB(skb)->seq, dsack); 3136 } 3137 3138 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) { 3139 SOCK_DEBUG(sk, "ofo packet was already received \n"); 3140 __skb_unlink(skb, &tp->out_of_order_queue); 3141 __kfree_skb(skb); 3142 continue; 3143 } 3144 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n", 3145 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, 3146 TCP_SKB_CB(skb)->end_seq); 3147 3148 __skb_unlink(skb, &tp->out_of_order_queue); 3149 __skb_queue_tail(&sk->sk_receive_queue, skb); 3150 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; 3151 if(skb->h.th->fin) 3152 tcp_fin(skb, sk, skb->h.th); 3153 } 3154} 3155 3156static int tcp_prune_queue(struct sock *sk); 3157 3158static void tcp_data_queue(struct sock *sk, struct sk_buff *skb) 3159{ 3160 struct tcphdr *th = skb->h.th; 3161 struct tcp_sock *tp = tcp_sk(sk); 3162 int eaten = -1; 3163 3164 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) 3165 goto drop; 3166 3167 __skb_pull(skb, th->doff*4); 3168 3169 TCP_ECN_accept_cwr(tp, skb); 3170 3171 if (tp->rx_opt.dsack) { 3172 tp->rx_opt.dsack = 0; 3173 tp->rx_opt.eff_sacks = min_t(unsigned int, tp->rx_opt.num_sacks, 3174 4 - tp->rx_opt.tstamp_ok); 3175 } 3176 3177 /* Queue data for delivery to the user. 3178 * Packets in sequence go to the receive queue. 3179 * Out of sequence packets to the out_of_order_queue. 3180 */ 3181 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) { 3182 if (tcp_receive_window(tp) == 0) 3183 goto out_of_window; 3184 3185 /* Ok. In sequence. In window. */ 3186 if (tp->ucopy.task == current && 3187 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len && 3188 sock_owned_by_user(sk) && !tp->urg_data) { 3189 int chunk = min_t(unsigned int, skb->len, 3190 tp->ucopy.len); 3191 3192 __set_current_state(TASK_RUNNING); 3193 3194 local_bh_enable(); 3195 if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) { 3196 tp->ucopy.len -= chunk; 3197 tp->copied_seq += chunk; 3198 eaten = (chunk == skb->len && !th->fin); 3199 tcp_rcv_space_adjust(sk); 3200 } 3201 local_bh_disable(); 3202 } 3203 3204 if (eaten <= 0) { 3205queue_and_out: 3206 if (eaten < 0 && 3207 (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf || 3208 !sk_stream_rmem_schedule(sk, skb))) { 3209 if (tcp_prune_queue(sk) < 0 || 3210 !sk_stream_rmem_schedule(sk, skb)) 3211 goto drop; 3212 } 3213 sk_stream_set_owner_r(skb, sk); 3214 __skb_queue_tail(&sk->sk_receive_queue, skb); 3215 } 3216 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; 3217 if(skb->len) 3218 tcp_event_data_recv(sk, tp, skb); 3219 if(th->fin) 3220 tcp_fin(skb, sk, th); 3221 3222 if (!skb_queue_empty(&tp->out_of_order_queue)) { 3223 tcp_ofo_queue(sk); 3224 3225 /* RFC2581. 4.2. SHOULD send immediate ACK, when 3226 * gap in queue is filled. 3227 */ 3228 if (skb_queue_empty(&tp->out_of_order_queue)) 3229 inet_csk(sk)->icsk_ack.pingpong = 0; 3230 } 3231 3232 if (tp->rx_opt.num_sacks) 3233 tcp_sack_remove(tp); 3234 3235 tcp_fast_path_check(sk, tp); 3236 3237 if (eaten > 0) 3238 __kfree_skb(skb); 3239 else if (!sock_flag(sk, SOCK_DEAD)) 3240 sk->sk_data_ready(sk, 0); 3241 return; 3242 } 3243 3244 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) { 3245 /* A retransmit, 2nd most common case. Force an immediate ack. */ 3246 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST); 3247 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); 3248 3249out_of_window: 3250 tcp_enter_quickack_mode(sk); 3251 inet_csk_schedule_ack(sk); 3252drop: 3253 __kfree_skb(skb); 3254 return; 3255 } 3256 3257 /* Out of window. F.e. zero window probe. */ 3258 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp))) 3259 goto out_of_window; 3260 3261 tcp_enter_quickack_mode(sk); 3262 3263 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 3264 /* Partial packet, seq < rcv_next < end_seq */ 3265 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n", 3266 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, 3267 TCP_SKB_CB(skb)->end_seq); 3268 3269 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, tp->rcv_nxt); 3270 3271 /* If window is closed, drop tail of packet. But after 3272 * remembering D-SACK for its head made in previous line. 3273 */ 3274 if (!tcp_receive_window(tp)) 3275 goto out_of_window; 3276 goto queue_and_out; 3277 } 3278 3279 TCP_ECN_check_ce(tp, skb); 3280 3281 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf || 3282 !sk_stream_rmem_schedule(sk, skb)) { 3283 if (tcp_prune_queue(sk) < 0 || 3284 !sk_stream_rmem_schedule(sk, skb)) 3285 goto drop; 3286 } 3287 3288 /* Disable header prediction. */ 3289 tp->pred_flags = 0; 3290 inet_csk_schedule_ack(sk); 3291 3292 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n", 3293 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); 3294 3295 sk_stream_set_owner_r(skb, sk); 3296 3297 if (!skb_peek(&tp->out_of_order_queue)) { 3298 /* Initial out of order segment, build 1 SACK. */ 3299 if (tp->rx_opt.sack_ok) { 3300 tp->rx_opt.num_sacks = 1; 3301 tp->rx_opt.dsack = 0; 3302 tp->rx_opt.eff_sacks = 1; 3303 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq; 3304 tp->selective_acks[0].end_seq = 3305 TCP_SKB_CB(skb)->end_seq; 3306 } 3307 __skb_queue_head(&tp->out_of_order_queue,skb); 3308 } else { 3309 struct sk_buff *skb1 = tp->out_of_order_queue.prev; 3310 u32 seq = TCP_SKB_CB(skb)->seq; 3311 u32 end_seq = TCP_SKB_CB(skb)->end_seq; 3312 3313 if (seq == TCP_SKB_CB(skb1)->end_seq) { 3314 __skb_append(skb1, skb, &tp->out_of_order_queue); 3315 3316 if (!tp->rx_opt.num_sacks || 3317 tp->selective_acks[0].end_seq != seq) 3318 goto add_sack; 3319 3320 /* Common case: data arrive in order after hole. */ 3321 tp->selective_acks[0].end_seq = end_seq; 3322 return; 3323 } 3324 3325 /* Find place to insert this segment. */ 3326 do { 3327 if (!after(TCP_SKB_CB(skb1)->seq, seq)) 3328 break; 3329 } while ((skb1 = skb1->prev) != 3330 (struct sk_buff*)&tp->out_of_order_queue); 3331 3332 /* Do skb overlap to previous one? */ 3333 if (skb1 != (struct sk_buff*)&tp->out_of_order_queue && 3334 before(seq, TCP_SKB_CB(skb1)->end_seq)) { 3335 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) { 3336 /* All the bits are present. Drop. */ 3337 __kfree_skb(skb); 3338 tcp_dsack_set(tp, seq, end_seq); 3339 goto add_sack; 3340 } 3341 if (after(seq, TCP_SKB_CB(skb1)->seq)) { 3342 /* Partial overlap. */ 3343 tcp_dsack_set(tp, seq, TCP_SKB_CB(skb1)->end_seq); 3344 } else { 3345 skb1 = skb1->prev; 3346 } 3347 } 3348 __skb_insert(skb, skb1, skb1->next, &tp->out_of_order_queue); 3349 3350 /* And clean segments covered by new one as whole. */ 3351 while ((skb1 = skb->next) != 3352 (struct sk_buff*)&tp->out_of_order_queue && 3353 after(end_seq, TCP_SKB_CB(skb1)->seq)) { 3354 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) { 3355 tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, end_seq); 3356 break; 3357 } 3358 __skb_unlink(skb1, &tp->out_of_order_queue); 3359 tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, TCP_SKB_CB(skb1)->end_seq); 3360 __kfree_skb(skb1); 3361 } 3362 3363add_sack: 3364 if (tp->rx_opt.sack_ok) 3365 tcp_sack_new_ofo_skb(sk, seq, end_seq); 3366 } 3367} 3368 3369/* Collapse contiguous sequence of skbs head..tail with 3370 * sequence numbers start..end. 3371 * Segments with FIN/SYN are not collapsed (only because this 3372 * simplifies code) 3373 */ 3374static void 3375tcp_collapse(struct sock *sk, struct sk_buff_head *list, 3376 struct sk_buff *head, struct sk_buff *tail, 3377 u32 start, u32 end) 3378{ 3379 struct sk_buff *skb; 3380 3381 /* First, check that queue is collapsible and find 3382 * the point where collapsing can be useful. */ 3383 for (skb = head; skb != tail; ) { 3384 /* No new bits? It is possible on ofo queue. */ 3385 if (!before(start, TCP_SKB_CB(skb)->end_seq)) { 3386 struct sk_buff *next = skb->next; 3387 __skb_unlink(skb, list); 3388 __kfree_skb(skb); 3389 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED); 3390 skb = next; 3391 continue; 3392 } 3393 3394 /* The first skb to collapse is: 3395 * - not SYN/FIN and 3396 * - bloated or contains data before "start" or 3397 * overlaps to the next one. 3398 */ 3399 if (!skb->h.th->syn && !skb->h.th->fin && 3400 (tcp_win_from_space(skb->truesize) > skb->len || 3401 before(TCP_SKB_CB(skb)->seq, start) || 3402 (skb->next != tail && 3403 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq))) 3404 break; 3405 3406 /* Decided to skip this, advance start seq. */ 3407 start = TCP_SKB_CB(skb)->end_seq; 3408 skb = skb->next; 3409 } 3410 if (skb == tail || skb->h.th->syn || skb->h.th->fin) 3411 return; 3412 3413 while (before(start, end)) { 3414 struct sk_buff *nskb; 3415 int header = skb_headroom(skb); 3416 int copy = SKB_MAX_ORDER(header, 0); 3417 3418 /* Too big header? This can happen with IPv6. */ 3419 if (copy < 0) 3420 return; 3421 if (end-start < copy) 3422 copy = end-start; 3423 nskb = alloc_skb(copy+header, GFP_ATOMIC); 3424 if (!nskb) 3425 return; 3426 skb_reserve(nskb, header); 3427 memcpy(nskb->head, skb->head, header); 3428 nskb->nh.raw = nskb->head + (skb->nh.raw-skb->head); 3429 nskb->h.raw = nskb->head + (skb->h.raw-skb->head); 3430 nskb->mac.raw = nskb->head + (skb->mac.raw-skb->head); 3431 memcpy(nskb->cb, skb->cb, sizeof(skb->cb)); 3432 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start; 3433 __skb_insert(nskb, skb->prev, skb, list); 3434 sk_stream_set_owner_r(nskb, sk); 3435 3436 /* Copy data, releasing collapsed skbs. */ 3437 while (copy > 0) { 3438 int offset = start - TCP_SKB_CB(skb)->seq; 3439 int size = TCP_SKB_CB(skb)->end_seq - start; 3440 3441 BUG_ON(offset < 0); 3442 if (size > 0) { 3443 size = min(copy, size); 3444 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size)) 3445 BUG(); 3446 TCP_SKB_CB(nskb)->end_seq += size; 3447 copy -= size; 3448 start += size; 3449 } 3450 if (!before(start, TCP_SKB_CB(skb)->end_seq)) { 3451 struct sk_buff *next = skb->next; 3452 __skb_unlink(skb, list); 3453 __kfree_skb(skb); 3454 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED); 3455 skb = next; 3456 if (skb == tail || skb->h.th->syn || skb->h.th->fin) 3457 return; 3458 } 3459 } 3460 } 3461} 3462 3463/* Collapse ofo queue. Algorithm: select contiguous sequence of skbs 3464 * and tcp_collapse() them until all the queue is collapsed. 3465 */ 3466static void tcp_collapse_ofo_queue(struct sock *sk) 3467{ 3468 struct tcp_sock *tp = tcp_sk(sk); 3469 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue); 3470 struct sk_buff *head; 3471 u32 start, end; 3472 3473 if (skb == NULL) 3474 return; 3475 3476 start = TCP_SKB_CB(skb)->seq; 3477 end = TCP_SKB_CB(skb)->end_seq; 3478 head = skb; 3479 3480 for (;;) { 3481 skb = skb->next; 3482 3483 /* Segment is terminated when we see gap or when 3484 * we are at the end of all the queue. */ 3485 if (skb == (struct sk_buff *)&tp->out_of_order_queue || 3486 after(TCP_SKB_CB(skb)->seq, end) || 3487 before(TCP_SKB_CB(skb)->end_seq, start)) { 3488 tcp_collapse(sk, &tp->out_of_order_queue, 3489 head, skb, start, end); 3490 head = skb; 3491 if (skb == (struct sk_buff *)&tp->out_of_order_queue) 3492 break; 3493 /* Start new segment */ 3494 start = TCP_SKB_CB(skb)->seq; 3495 end = TCP_SKB_CB(skb)->end_seq; 3496 } else { 3497 if (before(TCP_SKB_CB(skb)->seq, start)) 3498 start = TCP_SKB_CB(skb)->seq; 3499 if (after(TCP_SKB_CB(skb)->end_seq, end)) 3500 end = TCP_SKB_CB(skb)->end_seq; 3501 } 3502 } 3503} 3504 3505/* Reduce allocated memory if we can, trying to get 3506 * the socket within its memory limits again. 3507 * 3508 * Return less than zero if we should start dropping frames 3509 * until the socket owning process reads some of the data 3510 * to stabilize the situation. 3511 */ 3512static int tcp_prune_queue(struct sock *sk) 3513{ 3514 struct tcp_sock *tp = tcp_sk(sk); 3515 3516 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq); 3517 3518 NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED); 3519 3520 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf) 3521 tcp_clamp_window(sk, tp); 3522 else if (tcp_memory_pressure) 3523 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss); 3524 3525 tcp_collapse_ofo_queue(sk); 3526 tcp_collapse(sk, &sk->sk_receive_queue, 3527 sk->sk_receive_queue.next, 3528 (struct sk_buff*)&sk->sk_receive_queue, 3529 tp->copied_seq, tp->rcv_nxt); 3530 sk_stream_mem_reclaim(sk); 3531 3532 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) 3533 return 0; 3534 3535 /* Collapsing did not help, destructive actions follow. 3536 * This must not ever occur. */ 3537 3538 /* First, purge the out_of_order queue. */ 3539 if (!skb_queue_empty(&tp->out_of_order_queue)) { 3540 NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED); 3541 __skb_queue_purge(&tp->out_of_order_queue); 3542 3543 /* Reset SACK state. A conforming SACK implementation will 3544 * do the same at a timeout based retransmit. When a connection 3545 * is in a sad state like this, we care only about integrity 3546 * of the connection not performance. 3547 */ 3548 if (tp->rx_opt.sack_ok) 3549 tcp_sack_reset(&tp->rx_opt); 3550 sk_stream_mem_reclaim(sk); 3551 } 3552 3553 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) 3554 return 0; 3555 3556 /* If we are really being abused, tell the caller to silently 3557 * drop receive data on the floor. It will get retransmitted 3558 * and hopefully then we'll have sufficient space. 3559 */ 3560 NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED); 3561 3562 /* Massive buffer overcommit. */ 3563 tp->pred_flags = 0; 3564 return -1; 3565} 3566 3567 3568/* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. 3569 * As additional protections, we do not touch cwnd in retransmission phases, 3570 * and if application hit its sndbuf limit recently. 3571 */ 3572void tcp_cwnd_application_limited(struct sock *sk) 3573{ 3574 struct tcp_sock *tp = tcp_sk(sk); 3575 3576 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && 3577 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { 3578 /* Limited by application or receiver window. */ 3579 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); 3580 u32 win_used = max(tp->snd_cwnd_used, init_win); 3581 if (win_used < tp->snd_cwnd) { 3582 tp->snd_ssthresh = tcp_current_ssthresh(sk); 3583 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1; 3584 } 3585 tp->snd_cwnd_used = 0; 3586 } 3587 tp->snd_cwnd_stamp = tcp_time_stamp; 3588} 3589 3590static int tcp_should_expand_sndbuf(struct sock *sk, struct tcp_sock *tp) 3591{ 3592 /* If the user specified a specific send buffer setting, do 3593 * not modify it. 3594 */ 3595 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK) 3596 return 0; 3597 3598 /* If we are under global TCP memory pressure, do not expand. */ 3599 if (tcp_memory_pressure) 3600 return 0; 3601 3602 /* If we are under soft global TCP memory pressure, do not expand. */ 3603 if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0]) 3604 return 0; 3605 3606 /* If we filled the congestion window, do not expand. */ 3607 if (tp->packets_out >= tp->snd_cwnd) 3608 return 0; 3609 3610 return 1; 3611} 3612 3613/* When incoming ACK allowed to free some skb from write_queue, 3614 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket 3615 * on the exit from tcp input handler. 3616 * 3617 * PROBLEM: sndbuf expansion does not work well with largesend. 3618 */ 3619static void tcp_new_space(struct sock *sk) 3620{ 3621 struct tcp_sock *tp = tcp_sk(sk); 3622 3623 if (tcp_should_expand_sndbuf(sk, tp)) { 3624 int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) + 3625 MAX_TCP_HEADER + 16 + sizeof(struct sk_buff), 3626 demanded = max_t(unsigned int, tp->snd_cwnd, 3627 tp->reordering + 1); 3628 sndmem *= 2*demanded; 3629 if (sndmem > sk->sk_sndbuf) 3630 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]); 3631 tp->snd_cwnd_stamp = tcp_time_stamp; 3632 } 3633 3634 sk->sk_write_space(sk); 3635} 3636 3637static void tcp_check_space(struct sock *sk) 3638{ 3639 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) { 3640 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK); 3641 if (sk->sk_socket && 3642 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) 3643 tcp_new_space(sk); 3644 } 3645} 3646 3647static inline void tcp_data_snd_check(struct sock *sk, struct tcp_sock *tp) 3648{ 3649 tcp_push_pending_frames(sk, tp); 3650 tcp_check_space(sk); 3651} 3652 3653/* 3654 * Check if sending an ack is needed. 3655 */ 3656static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible) 3657{ 3658 struct tcp_sock *tp = tcp_sk(sk); 3659 3660 /* More than one full frame received... */ 3661 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss 3662 /* ... and right edge of window advances far enough. 3663 * (tcp_recvmsg() will send ACK otherwise). Or... 3664 */ 3665 && __tcp_select_window(sk) >= tp->rcv_wnd) || 3666 /* We ACK each frame or... */ 3667 tcp_in_quickack_mode(sk) || 3668 /* We have out of order data. */ 3669 (ofo_possible && 3670 skb_peek(&tp->out_of_order_queue))) { 3671 /* Then ack it now */ 3672 tcp_send_ack(sk); 3673 } else { 3674 /* Else, send delayed ack. */ 3675 tcp_send_delayed_ack(sk); 3676 } 3677} 3678 3679static inline void tcp_ack_snd_check(struct sock *sk) 3680{ 3681 if (!inet_csk_ack_scheduled(sk)) { 3682 /* We sent a data segment already. */ 3683 return; 3684 } 3685 __tcp_ack_snd_check(sk, 1); 3686} 3687 3688/* 3689 * This routine is only called when we have urgent data 3690 * signaled. Its the 'slow' part of tcp_urg. It could be 3691 * moved inline now as tcp_urg is only called from one 3692 * place. We handle URGent data wrong. We have to - as 3693 * BSD still doesn't use the correction from RFC961. 3694 * For 1003.1g we should support a new option TCP_STDURG to permit 3695 * either form (or just set the sysctl tcp_stdurg). 3696 */ 3697 3698static void tcp_check_urg(struct sock * sk, struct tcphdr * th) 3699{ 3700 struct tcp_sock *tp = tcp_sk(sk); 3701 u32 ptr = ntohs(th->urg_ptr); 3702 3703 if (ptr && !sysctl_tcp_stdurg) 3704 ptr--; 3705 ptr += ntohl(th->seq); 3706 3707 /* Ignore urgent data that we've already seen and read. */ 3708 if (after(tp->copied_seq, ptr)) 3709 return; 3710 3711 /* Do not replay urg ptr. 3712 * 3713 * NOTE: interesting situation not covered by specs. 3714 * Misbehaving sender may send urg ptr, pointing to segment, 3715 * which we already have in ofo queue. We are not able to fetch 3716 * such data and will stay in TCP_URG_NOTYET until will be eaten 3717 * by recvmsg(). Seems, we are not obliged to handle such wicked 3718 * situations. But it is worth to think about possibility of some 3719 * DoSes using some hypothetical application level deadlock. 3720 */ 3721 if (before(ptr, tp->rcv_nxt)) 3722 return; 3723 3724 /* Do we already have a newer (or duplicate) urgent pointer? */ 3725 if (tp->urg_data && !after(ptr, tp->urg_seq)) 3726 return; 3727 3728 /* Tell the world about our new urgent pointer. */ 3729 sk_send_sigurg(sk); 3730 3731 /* We may be adding urgent data when the last byte read was 3732 * urgent. To do this requires some care. We cannot just ignore 3733 * tp->copied_seq since we would read the last urgent byte again 3734 * as data, nor can we alter copied_seq until this data arrives 3735 * or we break the semantics of SIOCATMARK (and thus sockatmark()) 3736 * 3737 * NOTE. Double Dutch. Rendering to plain English: author of comment 3738 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB); 3739 * and expect that both A and B disappear from stream. This is _wrong_. 3740 * Though this happens in BSD with high probability, this is occasional. 3741 * Any application relying on this is buggy. Note also, that fix "works" 3742 * only in this artificial test. Insert some normal data between A and B and we will 3743 * decline of BSD again. Verdict: it is better to remove to trap 3744 * buggy users. 3745 */ 3746 if (tp->urg_seq == tp->copied_seq && tp->urg_data && 3747 !sock_flag(sk, SOCK_URGINLINE) && 3748 tp->copied_seq != tp->rcv_nxt) { 3749 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue); 3750 tp->copied_seq++; 3751 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) { 3752 __skb_unlink(skb, &sk->sk_receive_queue); 3753 __kfree_skb(skb); 3754 } 3755 } 3756 3757 tp->urg_data = TCP_URG_NOTYET; 3758 tp->urg_seq = ptr; 3759 3760 /* Disable header prediction. */ 3761 tp->pred_flags = 0; 3762} 3763 3764/* This is the 'fast' part of urgent handling. */ 3765static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th) 3766{ 3767 struct tcp_sock *tp = tcp_sk(sk); 3768 3769 /* Check if we get a new urgent pointer - normally not. */ 3770 if (th->urg) 3771 tcp_check_urg(sk,th); 3772 3773 /* Do we wait for any urgent data? - normally not... */ 3774 if (tp->urg_data == TCP_URG_NOTYET) { 3775 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) - 3776 th->syn; 3777 3778 /* Is the urgent pointer pointing into this packet? */ 3779 if (ptr < skb->len) { 3780 u8 tmp; 3781 if (skb_copy_bits(skb, ptr, &tmp, 1)) 3782 BUG(); 3783 tp->urg_data = TCP_URG_VALID | tmp; 3784 if (!sock_flag(sk, SOCK_DEAD)) 3785 sk->sk_data_ready(sk, 0); 3786 } 3787 } 3788} 3789 3790static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen) 3791{ 3792 struct tcp_sock *tp = tcp_sk(sk); 3793 int chunk = skb->len - hlen; 3794 int err; 3795 3796 local_bh_enable(); 3797 if (skb->ip_summed==CHECKSUM_UNNECESSARY) 3798 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk); 3799 else 3800 err = skb_copy_and_csum_datagram_iovec(skb, hlen, 3801 tp->ucopy.iov); 3802 3803 if (!err) { 3804 tp->ucopy.len -= chunk; 3805 tp->copied_seq += chunk; 3806 tcp_rcv_space_adjust(sk); 3807 } 3808 3809 local_bh_disable(); 3810 return err; 3811} 3812 3813static __sum16 __tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb) 3814{ 3815 __sum16 result; 3816 3817 if (sock_owned_by_user(sk)) { 3818 local_bh_enable(); 3819 result = __tcp_checksum_complete(skb); 3820 local_bh_disable(); 3821 } else { 3822 result = __tcp_checksum_complete(skb); 3823 } 3824 return result; 3825} 3826 3827static inline int tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb) 3828{ 3829 return skb->ip_summed != CHECKSUM_UNNECESSARY && 3830 __tcp_checksum_complete_user(sk, skb); 3831} 3832 3833#ifdef CONFIG_NET_DMA 3834static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb, int hlen) 3835{ 3836 struct tcp_sock *tp = tcp_sk(sk); 3837 int chunk = skb->len - hlen; 3838 int dma_cookie; 3839 int copied_early = 0; 3840 3841 if (tp->ucopy.wakeup) 3842 return 0; 3843 3844 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list) 3845 tp->ucopy.dma_chan = get_softnet_dma(); 3846 3847 if (tp->ucopy.dma_chan && skb->ip_summed == CHECKSUM_UNNECESSARY) { 3848 3849 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan, 3850 skb, hlen, tp->ucopy.iov, chunk, tp->ucopy.pinned_list); 3851 3852 if (dma_cookie < 0) 3853 goto out; 3854 3855 tp->ucopy.dma_cookie = dma_cookie; 3856 copied_early = 1; 3857 3858 tp->ucopy.len -= chunk; 3859 tp->copied_seq += chunk; 3860 tcp_rcv_space_adjust(sk); 3861 3862 if ((tp->ucopy.len == 0) || 3863 (tcp_flag_word(skb->h.th) & TCP_FLAG_PSH) || 3864 (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) { 3865 tp->ucopy.wakeup = 1; 3866 sk->sk_data_ready(sk, 0); 3867 } 3868 } else if (chunk > 0) { 3869 tp->ucopy.wakeup = 1; 3870 sk->sk_data_ready(sk, 0); 3871 } 3872out: 3873 return copied_early; 3874} 3875#endif /* CONFIG_NET_DMA */ 3876 3877/* 3878 * TCP receive function for the ESTABLISHED state. 3879 * 3880 * It is split into a fast path and a slow path. The fast path is 3881 * disabled when: 3882 * - A zero window was announced from us - zero window probing 3883 * is only handled properly in the slow path. 3884 * - Out of order segments arrived. 3885 * - Urgent data is expected. 3886 * - There is no buffer space left 3887 * - Unexpected TCP flags/window values/header lengths are received 3888 * (detected by checking the TCP header against pred_flags) 3889 * - Data is sent in both directions. Fast path only supports pure senders 3890 * or pure receivers (this means either the sequence number or the ack 3891 * value must stay constant) 3892 * - Unexpected TCP option. 3893 * 3894 * When these conditions are not satisfied it drops into a standard 3895 * receive procedure patterned after RFC793 to handle all cases. 3896 * The first three cases are guaranteed by proper pred_flags setting, 3897 * the rest is checked inline. Fast processing is turned on in 3898 * tcp_data_queue when everything is OK. 3899 */ 3900int tcp_rcv_established(struct sock *sk, struct sk_buff *skb, 3901 struct tcphdr *th, unsigned len) 3902{ 3903 struct tcp_sock *tp = tcp_sk(sk); 3904 3905 /* 3906 * Header prediction. 3907 * The code loosely follows the one in the famous 3908 * "30 instruction TCP receive" Van Jacobson mail. 3909 * 3910 * Van's trick is to deposit buffers into socket queue 3911 * on a device interrupt, to call tcp_recv function 3912 * on the receive process context and checksum and copy 3913 * the buffer to user space. smart... 3914 * 3915 * Our current scheme is not silly either but we take the 3916 * extra cost of the net_bh soft interrupt processing... 3917 * We do checksum and copy also but from device to kernel. 3918 */ 3919 3920 tp->rx_opt.saw_tstamp = 0; 3921 3922 /* pred_flags is 0xS?10 << 16 + snd_wnd 3923 * if header_prediction is to be made 3924 * 'S' will always be tp->tcp_header_len >> 2 3925 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to 3926 * turn it off (when there are holes in the receive 3927 * space for instance) 3928 * PSH flag is ignored. 3929 */ 3930 3931 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags && 3932 TCP_SKB_CB(skb)->seq == tp->rcv_nxt) { 3933 int tcp_header_len = tp->tcp_header_len; 3934 3935 /* Timestamp header prediction: tcp_header_len 3936 * is automatically equal to th->doff*4 due to pred_flags 3937 * match. 3938 */ 3939 3940 /* Check timestamp */ 3941 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) { 3942 __be32 *ptr = (__be32 *)(th + 1); 3943 3944 /* No? Slow path! */ 3945 if (*ptr != htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) 3946 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) 3947 goto slow_path; 3948 3949 tp->rx_opt.saw_tstamp = 1; 3950 ++ptr; 3951 tp->rx_opt.rcv_tsval = ntohl(*ptr); 3952 ++ptr; 3953 tp->rx_opt.rcv_tsecr = ntohl(*ptr); 3954 3955 /* If PAWS failed, check it more carefully in slow path */ 3956 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0) 3957 goto slow_path; 3958 3959 /* DO NOT update ts_recent here, if checksum fails 3960 * and timestamp was corrupted part, it will result 3961 * in a hung connection since we will drop all 3962 * future packets due to the PAWS test. 3963 */ 3964 } 3965 3966 if (len <= tcp_header_len) { 3967 /* Bulk data transfer: sender */ 3968 if (len == tcp_header_len) { 3969 /* Predicted packet is in window by definition. 3970 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 3971 * Hence, check seq<=rcv_wup reduces to: 3972 */ 3973 if (tcp_header_len == 3974 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && 3975 tp->rcv_nxt == tp->rcv_wup) 3976 tcp_store_ts_recent(tp); 3977 3978 /* We know that such packets are checksummed 3979 * on entry. 3980 */ 3981 tcp_ack(sk, skb, 0); 3982 __kfree_skb(skb); 3983 tcp_data_snd_check(sk, tp); 3984 return 0; 3985 } else { /* Header too small */ 3986 TCP_INC_STATS_BH(TCP_MIB_INERRS); 3987 goto discard; 3988 } 3989 } else { 3990 int eaten = 0; 3991 int copied_early = 0; 3992 3993 if (tp->copied_seq == tp->rcv_nxt && 3994 len - tcp_header_len <= tp->ucopy.len) { 3995#ifdef CONFIG_NET_DMA 3996 if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) { 3997 copied_early = 1; 3998 eaten = 1; 3999 } 4000#endif 4001 if (tp->ucopy.task == current && sock_owned_by_user(sk) && !copied_early) { 4002 __set_current_state(TASK_RUNNING); 4003 4004 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len)) 4005 eaten = 1; 4006 } 4007 if (eaten) { 4008 /* Predicted packet is in window by definition. 4009 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 4010 * Hence, check seq<=rcv_wup reduces to: 4011 */ 4012 if (tcp_header_len == 4013 (sizeof(struct tcphdr) + 4014 TCPOLEN_TSTAMP_ALIGNED) && 4015 tp->rcv_nxt == tp->rcv_wup) 4016 tcp_store_ts_recent(tp); 4017 4018 tcp_rcv_rtt_measure_ts(sk, skb); 4019 4020 __skb_pull(skb, tcp_header_len); 4021 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; 4022 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER); 4023 } 4024 if (copied_early) 4025 tcp_cleanup_rbuf(sk, skb->len); 4026 } 4027 if (!eaten) { 4028 if (tcp_checksum_complete_user(sk, skb)) 4029 goto csum_error; 4030 4031 /* Predicted packet is in window by definition. 4032 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 4033 * Hence, check seq<=rcv_wup reduces to: 4034 */ 4035 if (tcp_header_len == 4036 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && 4037 tp->rcv_nxt == tp->rcv_wup) 4038 tcp_store_ts_recent(tp); 4039 4040 tcp_rcv_rtt_measure_ts(sk, skb); 4041 4042 if ((int)skb->truesize > sk->sk_forward_alloc) 4043 goto step5; 4044 4045 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS); 4046 4047 /* Bulk data transfer: receiver */ 4048 __skb_pull(skb,tcp_header_len); 4049 __skb_queue_tail(&sk->sk_receive_queue, skb); 4050 sk_stream_set_owner_r(skb, sk); 4051 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; 4052 } 4053 4054 tcp_event_data_recv(sk, tp, skb); 4055 4056 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) { 4057 /* Well, only one small jumplet in fast path... */ 4058 tcp_ack(sk, skb, FLAG_DATA); 4059 tcp_data_snd_check(sk, tp); 4060 if (!inet_csk_ack_scheduled(sk)) 4061 goto no_ack; 4062 } 4063 4064 __tcp_ack_snd_check(sk, 0); 4065no_ack: 4066#ifdef CONFIG_NET_DMA 4067 if (copied_early) 4068 __skb_queue_tail(&sk->sk_async_wait_queue, skb); 4069 else 4070#endif 4071 if (eaten) 4072 __kfree_skb(skb); 4073 else 4074 sk->sk_data_ready(sk, 0); 4075 return 0; 4076 } 4077 } 4078 4079slow_path: 4080 if (len < (th->doff<<2) || tcp_checksum_complete_user(sk, skb)) 4081 goto csum_error; 4082 4083 /* 4084 * RFC1323: H1. Apply PAWS check first. 4085 */ 4086 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp && 4087 tcp_paws_discard(sk, skb)) { 4088 if (!th->rst) { 4089 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED); 4090 tcp_send_dupack(sk, skb); 4091 goto discard; 4092 } 4093 /* Resets are accepted even if PAWS failed. 4094 4095 ts_recent update must be made after we are sure 4096 that the packet is in window. 4097 */ 4098 } 4099 4100 /* 4101 * Standard slow path. 4102 */ 4103 4104 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) { 4105 /* RFC793, page 37: "In all states except SYN-SENT, all reset 4106 * (RST) segments are validated by checking their SEQ-fields." 4107 * And page 69: "If an incoming segment is not acceptable, 4108 * an acknowledgment should be sent in reply (unless the RST bit 4109 * is set, if so drop the segment and return)". 4110 */ 4111 if (!th->rst) 4112 tcp_send_dupack(sk, skb); 4113 goto discard; 4114 } 4115 4116 if(th->rst) { 4117 tcp_reset(sk); 4118 goto discard; 4119 } 4120 4121 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq); 4122 4123 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 4124 TCP_INC_STATS_BH(TCP_MIB_INERRS); 4125 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN); 4126 tcp_reset(sk); 4127 return 1; 4128 } 4129 4130step5: 4131 if(th->ack) 4132 tcp_ack(sk, skb, FLAG_SLOWPATH); 4133 4134 tcp_rcv_rtt_measure_ts(sk, skb); 4135 4136 /* Process urgent data. */ 4137 tcp_urg(sk, skb, th); 4138 4139 /* step 7: process the segment text */ 4140 tcp_data_queue(sk, skb); 4141 4142 tcp_data_snd_check(sk, tp); 4143 tcp_ack_snd_check(sk); 4144 return 0; 4145 4146csum_error: 4147 TCP_INC_STATS_BH(TCP_MIB_INERRS); 4148 4149discard: 4150 __kfree_skb(skb); 4151 return 0; 4152} 4153 4154static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb, 4155 struct tcphdr *th, unsigned len) 4156{ 4157 struct tcp_sock *tp = tcp_sk(sk); 4158 struct inet_connection_sock *icsk = inet_csk(sk); 4159 int saved_clamp = tp->rx_opt.mss_clamp; 4160 4161 tcp_parse_options(skb, &tp->rx_opt, 0); 4162 4163 if (th->ack) { 4164 /* rfc793: 4165 * "If the state is SYN-SENT then 4166 * first check the ACK bit 4167 * If the ACK bit is set 4168 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send 4169 * a reset (unless the RST bit is set, if so drop 4170 * the segment and return)" 4171 * 4172 * We do not send data with SYN, so that RFC-correct 4173 * test reduces to: 4174 */ 4175 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt) 4176 goto reset_and_undo; 4177 4178 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 4179 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp, 4180 tcp_time_stamp)) { 4181 NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED); 4182 goto reset_and_undo; 4183 } 4184 4185 /* Now ACK is acceptable. 4186 * 4187 * "If the RST bit is set 4188 * If the ACK was acceptable then signal the user "error: 4189 * connection reset", drop the segment, enter CLOSED state, 4190 * delete TCB, and return." 4191 */ 4192 4193 if (th->rst) { 4194 tcp_reset(sk); 4195 goto discard; 4196 } 4197 4198 /* rfc793: 4199 * "fifth, if neither of the SYN or RST bits is set then 4200 * drop the segment and return." 4201 * 4202 * See note below! 4203 * --ANK(990513) 4204 */ 4205 if (!th->syn) 4206 goto discard_and_undo; 4207 4208 /* rfc793: 4209 * "If the SYN bit is on ... 4210 * are acceptable then ... 4211 * (our SYN has been ACKed), change the connection 4212 * state to ESTABLISHED..." 4213 */ 4214 4215 TCP_ECN_rcv_synack(tp, th); 4216 4217 tp->snd_wl1 = TCP_SKB_CB(skb)->seq; 4218 tcp_ack(sk, skb, FLAG_SLOWPATH); 4219 4220 /* Ok.. it's good. Set up sequence numbers and 4221 * move to established. 4222 */ 4223 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; 4224 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; 4225 4226 /* RFC1323: The window in SYN & SYN/ACK segments is 4227 * never scaled. 4228 */ 4229 tp->snd_wnd = ntohs(th->window); 4230 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(skb)->seq); 4231 4232 if (!tp->rx_opt.wscale_ok) { 4233 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0; 4234 tp->window_clamp = min(tp->window_clamp, 65535U); 4235 } 4236 4237 if (tp->rx_opt.saw_tstamp) { 4238 tp->rx_opt.tstamp_ok = 1; 4239 tp->tcp_header_len = 4240 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 4241 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; 4242 tcp_store_ts_recent(tp); 4243 } else { 4244 tp->tcp_header_len = sizeof(struct tcphdr); 4245 } 4246 4247 if (tp->rx_opt.sack_ok && sysctl_tcp_fack) 4248 tp->rx_opt.sack_ok |= 2; 4249 4250 tcp_mtup_init(sk); 4251 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 4252 tcp_initialize_rcv_mss(sk); 4253 4254 /* Remember, tcp_poll() does not lock socket! 4255 * Change state from SYN-SENT only after copied_seq 4256 * is initialized. */ 4257 tp->copied_seq = tp->rcv_nxt; 4258 smp_mb(); 4259 tcp_set_state(sk, TCP_ESTABLISHED); 4260 4261 security_inet_conn_established(sk, skb); 4262 4263 /* Make sure socket is routed, for correct metrics. */ 4264 icsk->icsk_af_ops->rebuild_header(sk); 4265 4266 tcp_init_metrics(sk); 4267 4268 tcp_init_congestion_control(sk); 4269 4270 /* Prevent spurious tcp_cwnd_restart() on first data 4271 * packet. 4272 */ 4273 tp->lsndtime = tcp_time_stamp; 4274 4275 tcp_init_buffer_space(sk); 4276 4277 if (sock_flag(sk, SOCK_KEEPOPEN)) 4278 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp)); 4279 4280 if (!tp->rx_opt.snd_wscale) 4281 __tcp_fast_path_on(tp, tp->snd_wnd); 4282 else 4283 tp->pred_flags = 0; 4284 4285 if (!sock_flag(sk, SOCK_DEAD)) { 4286 sk->sk_state_change(sk); 4287 sk_wake_async(sk, 0, POLL_OUT); 4288 } 4289 4290 if (sk->sk_write_pending || 4291 icsk->icsk_accept_queue.rskq_defer_accept || 4292 icsk->icsk_ack.pingpong) { 4293 /* Save one ACK. Data will be ready after 4294 * several ticks, if write_pending is set. 4295 * 4296 * It may be deleted, but with this feature tcpdumps 4297 * look so _wonderfully_ clever, that I was not able 4298 * to stand against the temptation 8) --ANK 4299 */ 4300 inet_csk_schedule_ack(sk); 4301 icsk->icsk_ack.lrcvtime = tcp_time_stamp; 4302 icsk->icsk_ack.ato = TCP_ATO_MIN; 4303 tcp_incr_quickack(sk); 4304 tcp_enter_quickack_mode(sk); 4305 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 4306 TCP_DELACK_MAX, TCP_RTO_MAX); 4307 4308discard: 4309 __kfree_skb(skb); 4310 return 0; 4311 } else { 4312 tcp_send_ack(sk); 4313 } 4314 return -1; 4315 } 4316 4317 /* No ACK in the segment */ 4318 4319 if (th->rst) { 4320 /* rfc793: 4321 * "If the RST bit is set 4322 * 4323 * Otherwise (no ACK) drop the segment and return." 4324 */ 4325 4326 goto discard_and_undo; 4327 } 4328 4329 /* PAWS check. */ 4330 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && tcp_paws_check(&tp->rx_opt, 0)) 4331 goto discard_and_undo; 4332 4333 if (th->syn) { 4334 /* We see SYN without ACK. It is attempt of 4335 * simultaneous connect with crossed SYNs. 4336 * Particularly, it can be connect to self. 4337 */ 4338 tcp_set_state(sk, TCP_SYN_RECV); 4339 4340 if (tp->rx_opt.saw_tstamp) { 4341 tp->rx_opt.tstamp_ok = 1; 4342 tcp_store_ts_recent(tp); 4343 tp->tcp_header_len = 4344 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 4345 } else { 4346 tp->tcp_header_len = sizeof(struct tcphdr); 4347 } 4348 4349 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; 4350 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; 4351 4352 /* RFC1323: The window in SYN & SYN/ACK segments is 4353 * never scaled. 4354 */ 4355 tp->snd_wnd = ntohs(th->window); 4356 tp->snd_wl1 = TCP_SKB_CB(skb)->seq; 4357 tp->max_window = tp->snd_wnd; 4358 4359 TCP_ECN_rcv_syn(tp, th); 4360 4361 tcp_mtup_init(sk); 4362 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 4363 tcp_initialize_rcv_mss(sk); 4364 4365 4366 tcp_send_synack(sk); 4367#if 0 4368 /* Note, we could accept data and URG from this segment. 4369 * There are no obstacles to make this. 4370 * 4371 * However, if we ignore data in ACKless segments sometimes, 4372 * we have no reasons to accept it sometimes. 4373 * Also, seems the code doing it in step6 of tcp_rcv_state_process 4374 * is not flawless. So, discard packet for sanity. 4375 * Uncomment this return to process the data. 4376 */ 4377 return -1; 4378#else 4379 goto discard; 4380#endif 4381 } 4382 /* "fifth, if neither of the SYN or RST bits is set then 4383 * drop the segment and return." 4384 */ 4385 4386discard_and_undo: 4387 tcp_clear_options(&tp->rx_opt); 4388 tp->rx_opt.mss_clamp = saved_clamp; 4389 goto discard; 4390 4391reset_and_undo: 4392 tcp_clear_options(&tp->rx_opt); 4393 tp->rx_opt.mss_clamp = saved_clamp; 4394 return 1; 4395} 4396 4397 4398/* 4399 * This function implements the receiving procedure of RFC 793 for 4400 * all states except ESTABLISHED and TIME_WAIT. 4401 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be 4402 * address independent. 4403 */ 4404 4405int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb, 4406 struct tcphdr *th, unsigned len) 4407{ 4408 struct tcp_sock *tp = tcp_sk(sk); 4409 struct inet_connection_sock *icsk = inet_csk(sk); 4410 int queued = 0; 4411 4412 tp->rx_opt.saw_tstamp = 0; 4413 4414 switch (sk->sk_state) { 4415 case TCP_CLOSE: 4416 goto discard; 4417 4418 case TCP_LISTEN: 4419 if(th->ack) 4420 return 1; 4421 4422 if(th->rst) 4423 goto discard; 4424 4425 if(th->syn) { 4426 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0) 4427 return 1; 4428 4429 /* Now we have several options: In theory there is 4430 * nothing else in the frame. KA9Q has an option to 4431 * send data with the syn, BSD accepts data with the 4432 * syn up to the [to be] advertised window and 4433 * Solaris 2.1 gives you a protocol error. For now 4434 * we just ignore it, that fits the spec precisely 4435 * and avoids incompatibilities. It would be nice in 4436 * future to drop through and process the data. 4437 * 4438 * Now that TTCP is starting to be used we ought to 4439 * queue this data. 4440 * But, this leaves one open to an easy denial of 4441 * service attack, and SYN cookies can't defend 4442 * against this problem. So, we drop the data 4443 * in the interest of security over speed unless 4444 * it's still in use. 4445 */ 4446 kfree_skb(skb); 4447 return 0; 4448 } 4449 goto discard; 4450 4451 case TCP_SYN_SENT: 4452 queued = tcp_rcv_synsent_state_process(sk, skb, th, len); 4453 if (queued >= 0) 4454 return queued; 4455 4456 /* Do step6 onward by hand. */ 4457 tcp_urg(sk, skb, th); 4458 __kfree_skb(skb); 4459 tcp_data_snd_check(sk, tp); 4460 return 0; 4461 } 4462 4463 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp && 4464 tcp_paws_discard(sk, skb)) { 4465 if (!th->rst) { 4466 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED); 4467 tcp_send_dupack(sk, skb); 4468 goto discard; 4469 } 4470 /* Reset is accepted even if it did not pass PAWS. */ 4471 } 4472 4473 /* step 1: check sequence number */ 4474 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) { 4475 if (!th->rst) 4476 tcp_send_dupack(sk, skb); 4477 goto discard; 4478 } 4479 4480 /* step 2: check RST bit */ 4481 if(th->rst) { 4482 tcp_reset(sk); 4483 goto discard; 4484 } 4485 4486 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq); 4487 4488 /* step 3: check security and precedence [ignored] */ 4489 4490 /* step 4: 4491 * 4492 * Check for a SYN in window. 4493 */ 4494 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 4495 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN); 4496 tcp_reset(sk); 4497 return 1; 4498 } 4499 4500 /* step 5: check the ACK field */ 4501 if (th->ack) { 4502 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH); 4503 4504 switch(sk->sk_state) { 4505 case TCP_SYN_RECV: 4506 if (acceptable) { 4507 tp->copied_seq = tp->rcv_nxt; 4508 smp_mb(); 4509 tcp_set_state(sk, TCP_ESTABLISHED); 4510 sk->sk_state_change(sk); 4511 4512 /* Note, that this wakeup is only for marginal 4513 * crossed SYN case. Passively open sockets 4514 * are not waked up, because sk->sk_sleep == 4515 * NULL and sk->sk_socket == NULL. 4516 */ 4517 if (sk->sk_socket) { 4518 sk_wake_async(sk,0,POLL_OUT); 4519 } 4520 4521 tp->snd_una = TCP_SKB_CB(skb)->ack_seq; 4522 tp->snd_wnd = ntohs(th->window) << 4523 tp->rx_opt.snd_wscale; 4524 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, 4525 TCP_SKB_CB(skb)->seq); 4526 4527 /* tcp_ack considers this ACK as duplicate 4528 * and does not calculate rtt. 4529 * Fix it at least with timestamps. 4530 */ 4531 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 4532 !tp->srtt) 4533 tcp_ack_saw_tstamp(sk, 0); 4534 4535 if (tp->rx_opt.tstamp_ok) 4536 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; 4537 4538 /* Make sure socket is routed, for 4539 * correct metrics. 4540 */ 4541 icsk->icsk_af_ops->rebuild_header(sk); 4542 4543 tcp_init_metrics(sk); 4544 4545 tcp_init_congestion_control(sk); 4546 4547 /* Prevent spurious tcp_cwnd_restart() on 4548 * first data packet. 4549 */ 4550 tp->lsndtime = tcp_time_stamp; 4551 4552 tcp_mtup_init(sk); 4553 tcp_initialize_rcv_mss(sk); 4554 tcp_init_buffer_space(sk); 4555 tcp_fast_path_on(tp); 4556 } else { 4557 return 1; 4558 } 4559 break; 4560 4561 case TCP_FIN_WAIT1: 4562 if (tp->snd_una == tp->write_seq) { 4563 tcp_set_state(sk, TCP_FIN_WAIT2); 4564 sk->sk_shutdown |= SEND_SHUTDOWN; 4565 dst_confirm(sk->sk_dst_cache); 4566 4567 if (!sock_flag(sk, SOCK_DEAD)) 4568 /* Wake up lingering close() */ 4569 sk->sk_state_change(sk); 4570 else { 4571 int tmo; 4572 4573 if (tp->linger2 < 0 || 4574 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 4575 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) { 4576 tcp_done(sk); 4577 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA); 4578 return 1; 4579 } 4580 4581 tmo = tcp_fin_time(sk); 4582 if (tmo > TCP_TIMEWAIT_LEN) { 4583 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN); 4584 } else if (th->fin || sock_owned_by_user(sk)) { 4585 /* Bad case. We could lose such FIN otherwise. 4586 * It is not a big problem, but it looks confusing 4587 * and not so rare event. We still can lose it now, 4588 * if it spins in bh_lock_sock(), but it is really 4589 * marginal case. 4590 */ 4591 inet_csk_reset_keepalive_timer(sk, tmo); 4592 } else { 4593 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo); 4594 goto discard; 4595 } 4596 } 4597 } 4598 break; 4599 4600 case TCP_CLOSING: 4601 if (tp->snd_una == tp->write_seq) { 4602 tcp_time_wait(sk, TCP_TIME_WAIT, 0); 4603 goto discard; 4604 } 4605 break; 4606 4607 case TCP_LAST_ACK: 4608 if (tp->snd_una == tp->write_seq) { 4609 tcp_update_metrics(sk); 4610 tcp_done(sk); 4611 goto discard; 4612 } 4613 break; 4614 } 4615 } else 4616 goto discard; 4617 4618 /* step 6: check the URG bit */ 4619 tcp_urg(sk, skb, th); 4620 4621 /* step 7: process the segment text */ 4622 switch (sk->sk_state) { 4623 case TCP_CLOSE_WAIT: 4624 case TCP_CLOSING: 4625 case TCP_LAST_ACK: 4626 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) 4627 break; 4628 case TCP_FIN_WAIT1: 4629 case TCP_FIN_WAIT2: 4630 /* RFC 793 says to queue data in these states, 4631 * RFC 1122 says we MUST send a reset. 4632 * BSD 4.4 also does reset. 4633 */ 4634 if (sk->sk_shutdown & RCV_SHUTDOWN) { 4635 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 4636 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) { 4637 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA); 4638 tcp_reset(sk); 4639 return 1; 4640 } 4641 } 4642 /* Fall through */ 4643 case TCP_ESTABLISHED: 4644 tcp_data_queue(sk, skb); 4645 queued = 1; 4646 break; 4647 } 4648 4649 /* tcp_data could move socket to TIME-WAIT */ 4650 if (sk->sk_state != TCP_CLOSE) { 4651 tcp_data_snd_check(sk, tp); 4652 tcp_ack_snd_check(sk); 4653 } 4654 4655 if (!queued) { 4656discard: 4657 __kfree_skb(skb); 4658 } 4659 return 0; 4660} 4661 4662EXPORT_SYMBOL(sysctl_tcp_ecn); 4663EXPORT_SYMBOL(sysctl_tcp_reordering); 4664EXPORT_SYMBOL(tcp_parse_options); 4665EXPORT_SYMBOL(tcp_rcv_established); 4666EXPORT_SYMBOL(tcp_rcv_state_process); 4667EXPORT_SYMBOL(tcp_initialize_rcv_mss); 4668