tcp_input.c revision 4dc2665e3634d720a62bd27128fc8781fcdad2dc
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#define FLAG_ONLY_ORIG_SACKED 0x200 /* SACKs only non-rexmit sent before RTO */ 104 105#define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED) 106#define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED) 107#define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE) 108#define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED) 109 110#define IsReno(tp) ((tp)->rx_opt.sack_ok == 0) 111#define IsFack(tp) ((tp)->rx_opt.sack_ok & 2) 112#define IsDSack(tp) ((tp)->rx_opt.sack_ok & 4) 113 114#define IsSackFrto() (sysctl_tcp_frto == 0x2) 115 116#define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH) 117 118/* Adapt the MSS value used to make delayed ack decision to the 119 * real world. 120 */ 121static void tcp_measure_rcv_mss(struct sock *sk, 122 const struct sk_buff *skb) 123{ 124 struct inet_connection_sock *icsk = inet_csk(sk); 125 const unsigned int lss = icsk->icsk_ack.last_seg_size; 126 unsigned int len; 127 128 icsk->icsk_ack.last_seg_size = 0; 129 130 /* skb->len may jitter because of SACKs, even if peer 131 * sends good full-sized frames. 132 */ 133 len = skb_shinfo(skb)->gso_size ?: skb->len; 134 if (len >= icsk->icsk_ack.rcv_mss) { 135 icsk->icsk_ack.rcv_mss = len; 136 } else { 137 /* Otherwise, we make more careful check taking into account, 138 * that SACKs block is variable. 139 * 140 * "len" is invariant segment length, including TCP header. 141 */ 142 len += skb->data - skb->h.raw; 143 if (len >= TCP_MIN_RCVMSS + sizeof(struct tcphdr) || 144 /* If PSH is not set, packet should be 145 * full sized, provided peer TCP is not badly broken. 146 * This observation (if it is correct 8)) allows 147 * to handle super-low mtu links fairly. 148 */ 149 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) && 150 !(tcp_flag_word(skb->h.th)&TCP_REMNANT))) { 151 /* Subtract also invariant (if peer is RFC compliant), 152 * tcp header plus fixed timestamp option length. 153 * Resulting "len" is MSS free of SACK jitter. 154 */ 155 len -= tcp_sk(sk)->tcp_header_len; 156 icsk->icsk_ack.last_seg_size = len; 157 if (len == lss) { 158 icsk->icsk_ack.rcv_mss = len; 159 return; 160 } 161 } 162 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED) 163 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2; 164 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED; 165 } 166} 167 168static void tcp_incr_quickack(struct sock *sk) 169{ 170 struct inet_connection_sock *icsk = inet_csk(sk); 171 unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss); 172 173 if (quickacks==0) 174 quickacks=2; 175 if (quickacks > icsk->icsk_ack.quick) 176 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS); 177} 178 179void tcp_enter_quickack_mode(struct sock *sk) 180{ 181 struct inet_connection_sock *icsk = inet_csk(sk); 182 tcp_incr_quickack(sk); 183 icsk->icsk_ack.pingpong = 0; 184 icsk->icsk_ack.ato = TCP_ATO_MIN; 185} 186 187/* Send ACKs quickly, if "quick" count is not exhausted 188 * and the session is not interactive. 189 */ 190 191static inline int tcp_in_quickack_mode(const struct sock *sk) 192{ 193 const struct inet_connection_sock *icsk = inet_csk(sk); 194 return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong; 195} 196 197/* Buffer size and advertised window tuning. 198 * 199 * 1. Tuning sk->sk_sndbuf, when connection enters established state. 200 */ 201 202static void tcp_fixup_sndbuf(struct sock *sk) 203{ 204 int sndmem = tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER + 16 + 205 sizeof(struct sk_buff); 206 207 if (sk->sk_sndbuf < 3 * sndmem) 208 sk->sk_sndbuf = min(3 * sndmem, sysctl_tcp_wmem[2]); 209} 210 211/* 2. Tuning advertised window (window_clamp, rcv_ssthresh) 212 * 213 * All tcp_full_space() is split to two parts: "network" buffer, allocated 214 * forward and advertised in receiver window (tp->rcv_wnd) and 215 * "application buffer", required to isolate scheduling/application 216 * latencies from network. 217 * window_clamp is maximal advertised window. It can be less than 218 * tcp_full_space(), in this case tcp_full_space() - window_clamp 219 * is reserved for "application" buffer. The less window_clamp is 220 * the smoother our behaviour from viewpoint of network, but the lower 221 * throughput and the higher sensitivity of the connection to losses. 8) 222 * 223 * rcv_ssthresh is more strict window_clamp used at "slow start" 224 * phase to predict further behaviour of this connection. 225 * It is used for two goals: 226 * - to enforce header prediction at sender, even when application 227 * requires some significant "application buffer". It is check #1. 228 * - to prevent pruning of receive queue because of misprediction 229 * of receiver window. Check #2. 230 * 231 * The scheme does not work when sender sends good segments opening 232 * window and then starts to feed us spaghetti. But it should work 233 * in common situations. Otherwise, we have to rely on queue collapsing. 234 */ 235 236/* Slow part of check#2. */ 237static int __tcp_grow_window(const struct sock *sk, struct tcp_sock *tp, 238 const struct sk_buff *skb) 239{ 240 /* Optimize this! */ 241 int truesize = tcp_win_from_space(skb->truesize)/2; 242 int window = tcp_win_from_space(sysctl_tcp_rmem[2])/2; 243 244 while (tp->rcv_ssthresh <= window) { 245 if (truesize <= skb->len) 246 return 2 * inet_csk(sk)->icsk_ack.rcv_mss; 247 248 truesize >>= 1; 249 window >>= 1; 250 } 251 return 0; 252} 253 254static void tcp_grow_window(struct sock *sk, struct tcp_sock *tp, 255 struct sk_buff *skb) 256{ 257 /* Check #1 */ 258 if (tp->rcv_ssthresh < tp->window_clamp && 259 (int)tp->rcv_ssthresh < tcp_space(sk) && 260 !tcp_memory_pressure) { 261 int incr; 262 263 /* Check #2. Increase window, if skb with such overhead 264 * will fit to rcvbuf in future. 265 */ 266 if (tcp_win_from_space(skb->truesize) <= skb->len) 267 incr = 2*tp->advmss; 268 else 269 incr = __tcp_grow_window(sk, tp, skb); 270 271 if (incr) { 272 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr, tp->window_clamp); 273 inet_csk(sk)->icsk_ack.quick |= 1; 274 } 275 } 276} 277 278/* 3. Tuning rcvbuf, when connection enters established state. */ 279 280static void tcp_fixup_rcvbuf(struct sock *sk) 281{ 282 struct tcp_sock *tp = tcp_sk(sk); 283 int rcvmem = tp->advmss + MAX_TCP_HEADER + 16 + sizeof(struct sk_buff); 284 285 /* Try to select rcvbuf so that 4 mss-sized segments 286 * will fit to window and corresponding skbs will fit to our rcvbuf. 287 * (was 3; 4 is minimum to allow fast retransmit to work.) 288 */ 289 while (tcp_win_from_space(rcvmem) < tp->advmss) 290 rcvmem += 128; 291 if (sk->sk_rcvbuf < 4 * rcvmem) 292 sk->sk_rcvbuf = min(4 * rcvmem, sysctl_tcp_rmem[2]); 293} 294 295/* 4. Try to fixup all. It is made immediately after connection enters 296 * established state. 297 */ 298static void tcp_init_buffer_space(struct sock *sk) 299{ 300 struct tcp_sock *tp = tcp_sk(sk); 301 int maxwin; 302 303 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) 304 tcp_fixup_rcvbuf(sk); 305 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK)) 306 tcp_fixup_sndbuf(sk); 307 308 tp->rcvq_space.space = tp->rcv_wnd; 309 310 maxwin = tcp_full_space(sk); 311 312 if (tp->window_clamp >= maxwin) { 313 tp->window_clamp = maxwin; 314 315 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss) 316 tp->window_clamp = max(maxwin - 317 (maxwin >> sysctl_tcp_app_win), 318 4 * tp->advmss); 319 } 320 321 /* Force reservation of one segment. */ 322 if (sysctl_tcp_app_win && 323 tp->window_clamp > 2 * tp->advmss && 324 tp->window_clamp + tp->advmss > maxwin) 325 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss); 326 327 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp); 328 tp->snd_cwnd_stamp = tcp_time_stamp; 329} 330 331/* 5. Recalculate window clamp after socket hit its memory bounds. */ 332static void tcp_clamp_window(struct sock *sk, struct tcp_sock *tp) 333{ 334 struct inet_connection_sock *icsk = inet_csk(sk); 335 336 icsk->icsk_ack.quick = 0; 337 338 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] && 339 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) && 340 !tcp_memory_pressure && 341 atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) { 342 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc), 343 sysctl_tcp_rmem[2]); 344 } 345 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf) 346 tp->rcv_ssthresh = min(tp->window_clamp, 2U*tp->advmss); 347} 348 349 350/* Initialize RCV_MSS value. 351 * RCV_MSS is an our guess about MSS used by the peer. 352 * We haven't any direct information about the MSS. 353 * It's better to underestimate the RCV_MSS rather than overestimate. 354 * Overestimations make us ACKing less frequently than needed. 355 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss(). 356 */ 357void tcp_initialize_rcv_mss(struct sock *sk) 358{ 359 struct tcp_sock *tp = tcp_sk(sk); 360 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache); 361 362 hint = min(hint, tp->rcv_wnd/2); 363 hint = min(hint, TCP_MIN_RCVMSS); 364 hint = max(hint, TCP_MIN_MSS); 365 366 inet_csk(sk)->icsk_ack.rcv_mss = hint; 367} 368 369/* Receiver "autotuning" code. 370 * 371 * The algorithm for RTT estimation w/o timestamps is based on 372 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL. 373 * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps> 374 * 375 * More detail on this code can be found at 376 * <http://www.psc.edu/~jheffner/senior_thesis.ps>, 377 * though this reference is out of date. A new paper 378 * is pending. 379 */ 380static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep) 381{ 382 u32 new_sample = tp->rcv_rtt_est.rtt; 383 long m = sample; 384 385 if (m == 0) 386 m = 1; 387 388 if (new_sample != 0) { 389 /* If we sample in larger samples in the non-timestamp 390 * case, we could grossly overestimate the RTT especially 391 * with chatty applications or bulk transfer apps which 392 * are stalled on filesystem I/O. 393 * 394 * Also, since we are only going for a minimum in the 395 * non-timestamp case, we do not smooth things out 396 * else with timestamps disabled convergence takes too 397 * long. 398 */ 399 if (!win_dep) { 400 m -= (new_sample >> 3); 401 new_sample += m; 402 } else if (m < new_sample) 403 new_sample = m << 3; 404 } else { 405 /* No previous measure. */ 406 new_sample = m << 3; 407 } 408 409 if (tp->rcv_rtt_est.rtt != new_sample) 410 tp->rcv_rtt_est.rtt = new_sample; 411} 412 413static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp) 414{ 415 if (tp->rcv_rtt_est.time == 0) 416 goto new_measure; 417 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq)) 418 return; 419 tcp_rcv_rtt_update(tp, 420 jiffies - tp->rcv_rtt_est.time, 421 1); 422 423new_measure: 424 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd; 425 tp->rcv_rtt_est.time = tcp_time_stamp; 426} 427 428static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, const struct sk_buff *skb) 429{ 430 struct tcp_sock *tp = tcp_sk(sk); 431 if (tp->rx_opt.rcv_tsecr && 432 (TCP_SKB_CB(skb)->end_seq - 433 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss)) 434 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0); 435} 436 437/* 438 * This function should be called every time data is copied to user space. 439 * It calculates the appropriate TCP receive buffer space. 440 */ 441void tcp_rcv_space_adjust(struct sock *sk) 442{ 443 struct tcp_sock *tp = tcp_sk(sk); 444 int time; 445 int space; 446 447 if (tp->rcvq_space.time == 0) 448 goto new_measure; 449 450 time = tcp_time_stamp - tp->rcvq_space.time; 451 if (time < (tp->rcv_rtt_est.rtt >> 3) || 452 tp->rcv_rtt_est.rtt == 0) 453 return; 454 455 space = 2 * (tp->copied_seq - tp->rcvq_space.seq); 456 457 space = max(tp->rcvq_space.space, space); 458 459 if (tp->rcvq_space.space != space) { 460 int rcvmem; 461 462 tp->rcvq_space.space = space; 463 464 if (sysctl_tcp_moderate_rcvbuf && 465 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) { 466 int new_clamp = space; 467 468 /* Receive space grows, normalize in order to 469 * take into account packet headers and sk_buff 470 * structure overhead. 471 */ 472 space /= tp->advmss; 473 if (!space) 474 space = 1; 475 rcvmem = (tp->advmss + MAX_TCP_HEADER + 476 16 + sizeof(struct sk_buff)); 477 while (tcp_win_from_space(rcvmem) < tp->advmss) 478 rcvmem += 128; 479 space *= rcvmem; 480 space = min(space, sysctl_tcp_rmem[2]); 481 if (space > sk->sk_rcvbuf) { 482 sk->sk_rcvbuf = space; 483 484 /* Make the window clamp follow along. */ 485 tp->window_clamp = new_clamp; 486 } 487 } 488 } 489 490new_measure: 491 tp->rcvq_space.seq = tp->copied_seq; 492 tp->rcvq_space.time = tcp_time_stamp; 493} 494 495/* There is something which you must keep in mind when you analyze the 496 * behavior of the tp->ato delayed ack timeout interval. When a 497 * connection starts up, we want to ack as quickly as possible. The 498 * problem is that "good" TCP's do slow start at the beginning of data 499 * transmission. The means that until we send the first few ACK's the 500 * sender will sit on his end and only queue most of his data, because 501 * he can only send snd_cwnd unacked packets at any given time. For 502 * each ACK we send, he increments snd_cwnd and transmits more of his 503 * queue. -DaveM 504 */ 505static void tcp_event_data_recv(struct sock *sk, struct tcp_sock *tp, struct sk_buff *skb) 506{ 507 struct inet_connection_sock *icsk = inet_csk(sk); 508 u32 now; 509 510 inet_csk_schedule_ack(sk); 511 512 tcp_measure_rcv_mss(sk, skb); 513 514 tcp_rcv_rtt_measure(tp); 515 516 now = tcp_time_stamp; 517 518 if (!icsk->icsk_ack.ato) { 519 /* The _first_ data packet received, initialize 520 * delayed ACK engine. 521 */ 522 tcp_incr_quickack(sk); 523 icsk->icsk_ack.ato = TCP_ATO_MIN; 524 } else { 525 int m = now - icsk->icsk_ack.lrcvtime; 526 527 if (m <= TCP_ATO_MIN/2) { 528 /* The fastest case is the first. */ 529 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2; 530 } else if (m < icsk->icsk_ack.ato) { 531 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m; 532 if (icsk->icsk_ack.ato > icsk->icsk_rto) 533 icsk->icsk_ack.ato = icsk->icsk_rto; 534 } else if (m > icsk->icsk_rto) { 535 /* Too long gap. Apparently sender failed to 536 * restart window, so that we send ACKs quickly. 537 */ 538 tcp_incr_quickack(sk); 539 sk_stream_mem_reclaim(sk); 540 } 541 } 542 icsk->icsk_ack.lrcvtime = now; 543 544 TCP_ECN_check_ce(tp, skb); 545 546 if (skb->len >= 128) 547 tcp_grow_window(sk, tp, skb); 548} 549 550/* Called to compute a smoothed rtt estimate. The data fed to this 551 * routine either comes from timestamps, or from segments that were 552 * known _not_ to have been retransmitted [see Karn/Partridge 553 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88 554 * piece by Van Jacobson. 555 * NOTE: the next three routines used to be one big routine. 556 * To save cycles in the RFC 1323 implementation it was better to break 557 * it up into three procedures. -- erics 558 */ 559static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt) 560{ 561 struct tcp_sock *tp = tcp_sk(sk); 562 long m = mrtt; /* RTT */ 563 564 /* The following amusing code comes from Jacobson's 565 * article in SIGCOMM '88. Note that rtt and mdev 566 * are scaled versions of rtt and mean deviation. 567 * This is designed to be as fast as possible 568 * m stands for "measurement". 569 * 570 * On a 1990 paper the rto value is changed to: 571 * RTO = rtt + 4 * mdev 572 * 573 * Funny. This algorithm seems to be very broken. 574 * These formulae increase RTO, when it should be decreased, increase 575 * too slowly, when it should be increased quickly, decrease too quickly 576 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely 577 * does not matter how to _calculate_ it. Seems, it was trap 578 * that VJ failed to avoid. 8) 579 */ 580 if(m == 0) 581 m = 1; 582 if (tp->srtt != 0) { 583 m -= (tp->srtt >> 3); /* m is now error in rtt est */ 584 tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */ 585 if (m < 0) { 586 m = -m; /* m is now abs(error) */ 587 m -= (tp->mdev >> 2); /* similar update on mdev */ 588 /* This is similar to one of Eifel findings. 589 * Eifel blocks mdev updates when rtt decreases. 590 * This solution is a bit different: we use finer gain 591 * for mdev in this case (alpha*beta). 592 * Like Eifel it also prevents growth of rto, 593 * but also it limits too fast rto decreases, 594 * happening in pure Eifel. 595 */ 596 if (m > 0) 597 m >>= 3; 598 } else { 599 m -= (tp->mdev >> 2); /* similar update on mdev */ 600 } 601 tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */ 602 if (tp->mdev > tp->mdev_max) { 603 tp->mdev_max = tp->mdev; 604 if (tp->mdev_max > tp->rttvar) 605 tp->rttvar = tp->mdev_max; 606 } 607 if (after(tp->snd_una, tp->rtt_seq)) { 608 if (tp->mdev_max < tp->rttvar) 609 tp->rttvar -= (tp->rttvar-tp->mdev_max)>>2; 610 tp->rtt_seq = tp->snd_nxt; 611 tp->mdev_max = TCP_RTO_MIN; 612 } 613 } else { 614 /* no previous measure. */ 615 tp->srtt = m<<3; /* take the measured time to be rtt */ 616 tp->mdev = m<<1; /* make sure rto = 3*rtt */ 617 tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN); 618 tp->rtt_seq = tp->snd_nxt; 619 } 620} 621 622/* Calculate rto without backoff. This is the second half of Van Jacobson's 623 * routine referred to above. 624 */ 625static inline void tcp_set_rto(struct sock *sk) 626{ 627 const struct tcp_sock *tp = tcp_sk(sk); 628 /* Old crap is replaced with new one. 8) 629 * 630 * More seriously: 631 * 1. If rtt variance happened to be less 50msec, it is hallucination. 632 * It cannot be less due to utterly erratic ACK generation made 633 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_ 634 * to do with delayed acks, because at cwnd>2 true delack timeout 635 * is invisible. Actually, Linux-2.4 also generates erratic 636 * ACKs in some circumstances. 637 */ 638 inet_csk(sk)->icsk_rto = (tp->srtt >> 3) + tp->rttvar; 639 640 /* 2. Fixups made earlier cannot be right. 641 * If we do not estimate RTO correctly without them, 642 * all the algo is pure shit and should be replaced 643 * with correct one. It is exactly, which we pretend to do. 644 */ 645} 646 647/* NOTE: clamping at TCP_RTO_MIN is not required, current algo 648 * guarantees that rto is higher. 649 */ 650static inline void tcp_bound_rto(struct sock *sk) 651{ 652 if (inet_csk(sk)->icsk_rto > TCP_RTO_MAX) 653 inet_csk(sk)->icsk_rto = TCP_RTO_MAX; 654} 655 656/* Save metrics learned by this TCP session. 657 This function is called only, when TCP finishes successfully 658 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE. 659 */ 660void tcp_update_metrics(struct sock *sk) 661{ 662 struct tcp_sock *tp = tcp_sk(sk); 663 struct dst_entry *dst = __sk_dst_get(sk); 664 665 if (sysctl_tcp_nometrics_save) 666 return; 667 668 dst_confirm(dst); 669 670 if (dst && (dst->flags&DST_HOST)) { 671 const struct inet_connection_sock *icsk = inet_csk(sk); 672 int m; 673 674 if (icsk->icsk_backoff || !tp->srtt) { 675 /* This session failed to estimate rtt. Why? 676 * Probably, no packets returned in time. 677 * Reset our results. 678 */ 679 if (!(dst_metric_locked(dst, RTAX_RTT))) 680 dst->metrics[RTAX_RTT-1] = 0; 681 return; 682 } 683 684 m = dst_metric(dst, RTAX_RTT) - tp->srtt; 685 686 /* If newly calculated rtt larger than stored one, 687 * store new one. Otherwise, use EWMA. Remember, 688 * rtt overestimation is always better than underestimation. 689 */ 690 if (!(dst_metric_locked(dst, RTAX_RTT))) { 691 if (m <= 0) 692 dst->metrics[RTAX_RTT-1] = tp->srtt; 693 else 694 dst->metrics[RTAX_RTT-1] -= (m>>3); 695 } 696 697 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) { 698 if (m < 0) 699 m = -m; 700 701 /* Scale deviation to rttvar fixed point */ 702 m >>= 1; 703 if (m < tp->mdev) 704 m = tp->mdev; 705 706 if (m >= dst_metric(dst, RTAX_RTTVAR)) 707 dst->metrics[RTAX_RTTVAR-1] = m; 708 else 709 dst->metrics[RTAX_RTTVAR-1] -= 710 (dst->metrics[RTAX_RTTVAR-1] - m)>>2; 711 } 712 713 if (tp->snd_ssthresh >= 0xFFFF) { 714 /* Slow start still did not finish. */ 715 if (dst_metric(dst, RTAX_SSTHRESH) && 716 !dst_metric_locked(dst, RTAX_SSTHRESH) && 717 (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH)) 718 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1; 719 if (!dst_metric_locked(dst, RTAX_CWND) && 720 tp->snd_cwnd > dst_metric(dst, RTAX_CWND)) 721 dst->metrics[RTAX_CWND-1] = tp->snd_cwnd; 722 } else if (tp->snd_cwnd > tp->snd_ssthresh && 723 icsk->icsk_ca_state == TCP_CA_Open) { 724 /* Cong. avoidance phase, cwnd is reliable. */ 725 if (!dst_metric_locked(dst, RTAX_SSTHRESH)) 726 dst->metrics[RTAX_SSTHRESH-1] = 727 max(tp->snd_cwnd >> 1, tp->snd_ssthresh); 728 if (!dst_metric_locked(dst, RTAX_CWND)) 729 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_cwnd) >> 1; 730 } else { 731 /* Else slow start did not finish, cwnd is non-sense, 732 ssthresh may be also invalid. 733 */ 734 if (!dst_metric_locked(dst, RTAX_CWND)) 735 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_ssthresh) >> 1; 736 if (dst->metrics[RTAX_SSTHRESH-1] && 737 !dst_metric_locked(dst, RTAX_SSTHRESH) && 738 tp->snd_ssthresh > dst->metrics[RTAX_SSTHRESH-1]) 739 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh; 740 } 741 742 if (!dst_metric_locked(dst, RTAX_REORDERING)) { 743 if (dst->metrics[RTAX_REORDERING-1] < tp->reordering && 744 tp->reordering != sysctl_tcp_reordering) 745 dst->metrics[RTAX_REORDERING-1] = tp->reordering; 746 } 747 } 748} 749 750/* Numbers are taken from RFC2414. */ 751__u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst) 752{ 753 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0); 754 755 if (!cwnd) { 756 if (tp->mss_cache > 1460) 757 cwnd = 2; 758 else 759 cwnd = (tp->mss_cache > 1095) ? 3 : 4; 760 } 761 return min_t(__u32, cwnd, tp->snd_cwnd_clamp); 762} 763 764/* Set slow start threshold and cwnd not falling to slow start */ 765void tcp_enter_cwr(struct sock *sk) 766{ 767 struct tcp_sock *tp = tcp_sk(sk); 768 769 tp->prior_ssthresh = 0; 770 tp->bytes_acked = 0; 771 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) { 772 tp->undo_marker = 0; 773 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk); 774 tp->snd_cwnd = min(tp->snd_cwnd, 775 tcp_packets_in_flight(tp) + 1U); 776 tp->snd_cwnd_cnt = 0; 777 tp->high_seq = tp->snd_nxt; 778 tp->snd_cwnd_stamp = tcp_time_stamp; 779 TCP_ECN_queue_cwr(tp); 780 781 tcp_set_ca_state(sk, TCP_CA_CWR); 782 } 783} 784 785/* Initialize metrics on socket. */ 786 787static void tcp_init_metrics(struct sock *sk) 788{ 789 struct tcp_sock *tp = tcp_sk(sk); 790 struct dst_entry *dst = __sk_dst_get(sk); 791 792 if (dst == NULL) 793 goto reset; 794 795 dst_confirm(dst); 796 797 if (dst_metric_locked(dst, RTAX_CWND)) 798 tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND); 799 if (dst_metric(dst, RTAX_SSTHRESH)) { 800 tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH); 801 if (tp->snd_ssthresh > tp->snd_cwnd_clamp) 802 tp->snd_ssthresh = tp->snd_cwnd_clamp; 803 } 804 if (dst_metric(dst, RTAX_REORDERING) && 805 tp->reordering != dst_metric(dst, RTAX_REORDERING)) { 806 tp->rx_opt.sack_ok &= ~2; 807 tp->reordering = dst_metric(dst, RTAX_REORDERING); 808 } 809 810 if (dst_metric(dst, RTAX_RTT) == 0) 811 goto reset; 812 813 if (!tp->srtt && dst_metric(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3)) 814 goto reset; 815 816 /* Initial rtt is determined from SYN,SYN-ACK. 817 * The segment is small and rtt may appear much 818 * less than real one. Use per-dst memory 819 * to make it more realistic. 820 * 821 * A bit of theory. RTT is time passed after "normal" sized packet 822 * is sent until it is ACKed. In normal circumstances sending small 823 * packets force peer to delay ACKs and calculation is correct too. 824 * The algorithm is adaptive and, provided we follow specs, it 825 * NEVER underestimate RTT. BUT! If peer tries to make some clever 826 * tricks sort of "quick acks" for time long enough to decrease RTT 827 * to low value, and then abruptly stops to do it and starts to delay 828 * ACKs, wait for troubles. 829 */ 830 if (dst_metric(dst, RTAX_RTT) > tp->srtt) { 831 tp->srtt = dst_metric(dst, RTAX_RTT); 832 tp->rtt_seq = tp->snd_nxt; 833 } 834 if (dst_metric(dst, RTAX_RTTVAR) > tp->mdev) { 835 tp->mdev = dst_metric(dst, RTAX_RTTVAR); 836 tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN); 837 } 838 tcp_set_rto(sk); 839 tcp_bound_rto(sk); 840 if (inet_csk(sk)->icsk_rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp) 841 goto reset; 842 tp->snd_cwnd = tcp_init_cwnd(tp, dst); 843 tp->snd_cwnd_stamp = tcp_time_stamp; 844 return; 845 846reset: 847 /* Play conservative. If timestamps are not 848 * supported, TCP will fail to recalculate correct 849 * rtt, if initial rto is too small. FORGET ALL AND RESET! 850 */ 851 if (!tp->rx_opt.saw_tstamp && tp->srtt) { 852 tp->srtt = 0; 853 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT; 854 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; 855 } 856} 857 858static void tcp_update_reordering(struct sock *sk, const int metric, 859 const int ts) 860{ 861 struct tcp_sock *tp = tcp_sk(sk); 862 if (metric > tp->reordering) { 863 tp->reordering = min(TCP_MAX_REORDERING, metric); 864 865 /* This exciting event is worth to be remembered. 8) */ 866 if (ts) 867 NET_INC_STATS_BH(LINUX_MIB_TCPTSREORDER); 868 else if (IsReno(tp)) 869 NET_INC_STATS_BH(LINUX_MIB_TCPRENOREORDER); 870 else if (IsFack(tp)) 871 NET_INC_STATS_BH(LINUX_MIB_TCPFACKREORDER); 872 else 873 NET_INC_STATS_BH(LINUX_MIB_TCPSACKREORDER); 874#if FASTRETRANS_DEBUG > 1 875 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n", 876 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state, 877 tp->reordering, 878 tp->fackets_out, 879 tp->sacked_out, 880 tp->undo_marker ? tp->undo_retrans : 0); 881#endif 882 /* Disable FACK yet. */ 883 tp->rx_opt.sack_ok &= ~2; 884 } 885} 886 887/* This procedure tags the retransmission queue when SACKs arrive. 888 * 889 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L). 890 * Packets in queue with these bits set are counted in variables 891 * sacked_out, retrans_out and lost_out, correspondingly. 892 * 893 * Valid combinations are: 894 * Tag InFlight Description 895 * 0 1 - orig segment is in flight. 896 * S 0 - nothing flies, orig reached receiver. 897 * L 0 - nothing flies, orig lost by net. 898 * R 2 - both orig and retransmit are in flight. 899 * L|R 1 - orig is lost, retransmit is in flight. 900 * S|R 1 - orig reached receiver, retrans is still in flight. 901 * (L|S|R is logically valid, it could occur when L|R is sacked, 902 * but it is equivalent to plain S and code short-curcuits it to S. 903 * L|S is logically invalid, it would mean -1 packet in flight 8)) 904 * 905 * These 6 states form finite state machine, controlled by the following events: 906 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue()) 907 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue()) 908 * 3. Loss detection event of one of three flavors: 909 * A. Scoreboard estimator decided the packet is lost. 910 * A'. Reno "three dupacks" marks head of queue lost. 911 * A''. Its FACK modfication, head until snd.fack is lost. 912 * B. SACK arrives sacking data transmitted after never retransmitted 913 * hole was sent out. 914 * C. SACK arrives sacking SND.NXT at the moment, when the 915 * segment was retransmitted. 916 * 4. D-SACK added new rule: D-SACK changes any tag to S. 917 * 918 * It is pleasant to note, that state diagram turns out to be commutative, 919 * so that we are allowed not to be bothered by order of our actions, 920 * when multiple events arrive simultaneously. (see the function below). 921 * 922 * Reordering detection. 923 * -------------------- 924 * Reordering metric is maximal distance, which a packet can be displaced 925 * in packet stream. With SACKs we can estimate it: 926 * 927 * 1. SACK fills old hole and the corresponding segment was not 928 * ever retransmitted -> reordering. Alas, we cannot use it 929 * when segment was retransmitted. 930 * 2. The last flaw is solved with D-SACK. D-SACK arrives 931 * for retransmitted and already SACKed segment -> reordering.. 932 * Both of these heuristics are not used in Loss state, when we cannot 933 * account for retransmits accurately. 934 */ 935static int 936tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, u32 prior_snd_una) 937{ 938 const struct inet_connection_sock *icsk = inet_csk(sk); 939 struct tcp_sock *tp = tcp_sk(sk); 940 unsigned char *ptr = ack_skb->h.raw + TCP_SKB_CB(ack_skb)->sacked; 941 struct tcp_sack_block_wire *sp = (struct tcp_sack_block_wire *)(ptr+2); 942 struct sk_buff *cached_skb; 943 int num_sacks = (ptr[1] - TCPOLEN_SACK_BASE)>>3; 944 int reord = tp->packets_out; 945 int prior_fackets; 946 u32 lost_retrans = 0; 947 int flag = 0; 948 int dup_sack = 0; 949 int cached_fack_count; 950 int i; 951 int first_sack_index; 952 953 if (!tp->sacked_out) 954 tp->fackets_out = 0; 955 prior_fackets = tp->fackets_out; 956 957 /* Check for D-SACK. */ 958 if (before(ntohl(sp[0].start_seq), TCP_SKB_CB(ack_skb)->ack_seq)) { 959 dup_sack = 1; 960 tp->rx_opt.sack_ok |= 4; 961 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV); 962 } else if (num_sacks > 1 && 963 !after(ntohl(sp[0].end_seq), ntohl(sp[1].end_seq)) && 964 !before(ntohl(sp[0].start_seq), ntohl(sp[1].start_seq))) { 965 dup_sack = 1; 966 tp->rx_opt.sack_ok |= 4; 967 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV); 968 } 969 970 /* D-SACK for already forgotten data... 971 * Do dumb counting. */ 972 if (dup_sack && 973 !after(ntohl(sp[0].end_seq), prior_snd_una) && 974 after(ntohl(sp[0].end_seq), tp->undo_marker)) 975 tp->undo_retrans--; 976 977 /* Eliminate too old ACKs, but take into 978 * account more or less fresh ones, they can 979 * contain valid SACK info. 980 */ 981 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window)) 982 return 0; 983 984 /* SACK fastpath: 985 * if the only SACK change is the increase of the end_seq of 986 * the first block then only apply that SACK block 987 * and use retrans queue hinting otherwise slowpath */ 988 flag = 1; 989 for (i = 0; i < num_sacks; i++) { 990 __be32 start_seq = sp[i].start_seq; 991 __be32 end_seq = sp[i].end_seq; 992 993 if (i == 0) { 994 if (tp->recv_sack_cache[i].start_seq != start_seq) 995 flag = 0; 996 } else { 997 if ((tp->recv_sack_cache[i].start_seq != start_seq) || 998 (tp->recv_sack_cache[i].end_seq != end_seq)) 999 flag = 0; 1000 } 1001 tp->recv_sack_cache[i].start_seq = start_seq; 1002 tp->recv_sack_cache[i].end_seq = end_seq; 1003 } 1004 /* Clear the rest of the cache sack blocks so they won't match mistakenly. */ 1005 for (; i < ARRAY_SIZE(tp->recv_sack_cache); i++) { 1006 tp->recv_sack_cache[i].start_seq = 0; 1007 tp->recv_sack_cache[i].end_seq = 0; 1008 } 1009 1010 first_sack_index = 0; 1011 if (flag) 1012 num_sacks = 1; 1013 else { 1014 int j; 1015 tp->fastpath_skb_hint = NULL; 1016 1017 /* order SACK blocks to allow in order walk of the retrans queue */ 1018 for (i = num_sacks-1; i > 0; i--) { 1019 for (j = 0; j < i; j++){ 1020 if (after(ntohl(sp[j].start_seq), 1021 ntohl(sp[j+1].start_seq))){ 1022 struct tcp_sack_block_wire tmp; 1023 1024 tmp = sp[j]; 1025 sp[j] = sp[j+1]; 1026 sp[j+1] = tmp; 1027 1028 /* Track where the first SACK block goes to */ 1029 if (j == first_sack_index) 1030 first_sack_index = j+1; 1031 } 1032 1033 } 1034 } 1035 } 1036 1037 /* clear flag as used for different purpose in following code */ 1038 flag = 0; 1039 1040 /* Use SACK fastpath hint if valid */ 1041 cached_skb = tp->fastpath_skb_hint; 1042 cached_fack_count = tp->fastpath_cnt_hint; 1043 if (!cached_skb) { 1044 cached_skb = sk->sk_write_queue.next; 1045 cached_fack_count = 0; 1046 } 1047 1048 for (i=0; i<num_sacks; i++, sp++) { 1049 struct sk_buff *skb; 1050 __u32 start_seq = ntohl(sp->start_seq); 1051 __u32 end_seq = ntohl(sp->end_seq); 1052 int fack_count; 1053 1054 skb = cached_skb; 1055 fack_count = cached_fack_count; 1056 1057 /* Event "B" in the comment above. */ 1058 if (after(end_seq, tp->high_seq)) 1059 flag |= FLAG_DATA_LOST; 1060 1061 sk_stream_for_retrans_queue_from(skb, sk) { 1062 int in_sack, pcount; 1063 u8 sacked; 1064 1065 cached_skb = skb; 1066 cached_fack_count = fack_count; 1067 if (i == first_sack_index) { 1068 tp->fastpath_skb_hint = skb; 1069 tp->fastpath_cnt_hint = fack_count; 1070 } 1071 1072 /* The retransmission queue is always in order, so 1073 * we can short-circuit the walk early. 1074 */ 1075 if (!before(TCP_SKB_CB(skb)->seq, end_seq)) 1076 break; 1077 1078 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && 1079 !before(end_seq, TCP_SKB_CB(skb)->end_seq); 1080 1081 pcount = tcp_skb_pcount(skb); 1082 1083 if (pcount > 1 && !in_sack && 1084 after(TCP_SKB_CB(skb)->end_seq, start_seq)) { 1085 unsigned int pkt_len; 1086 1087 in_sack = !after(start_seq, 1088 TCP_SKB_CB(skb)->seq); 1089 1090 if (!in_sack) 1091 pkt_len = (start_seq - 1092 TCP_SKB_CB(skb)->seq); 1093 else 1094 pkt_len = (end_seq - 1095 TCP_SKB_CB(skb)->seq); 1096 if (tcp_fragment(sk, skb, pkt_len, skb_shinfo(skb)->gso_size)) 1097 break; 1098 pcount = tcp_skb_pcount(skb); 1099 } 1100 1101 fack_count += pcount; 1102 1103 sacked = TCP_SKB_CB(skb)->sacked; 1104 1105 /* Account D-SACK for retransmitted packet. */ 1106 if ((dup_sack && in_sack) && 1107 (sacked & TCPCB_RETRANS) && 1108 after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker)) 1109 tp->undo_retrans--; 1110 1111 /* The frame is ACKed. */ 1112 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) { 1113 if (sacked&TCPCB_RETRANS) { 1114 if ((dup_sack && in_sack) && 1115 (sacked&TCPCB_SACKED_ACKED)) 1116 reord = min(fack_count, reord); 1117 } else { 1118 /* If it was in a hole, we detected reordering. */ 1119 if (fack_count < prior_fackets && 1120 !(sacked&TCPCB_SACKED_ACKED)) 1121 reord = min(fack_count, reord); 1122 } 1123 1124 /* Nothing to do; acked frame is about to be dropped. */ 1125 continue; 1126 } 1127 1128 if ((sacked&TCPCB_SACKED_RETRANS) && 1129 after(end_seq, TCP_SKB_CB(skb)->ack_seq) && 1130 (!lost_retrans || after(end_seq, lost_retrans))) 1131 lost_retrans = end_seq; 1132 1133 if (!in_sack) 1134 continue; 1135 1136 if (!(sacked&TCPCB_SACKED_ACKED)) { 1137 if (sacked & TCPCB_SACKED_RETRANS) { 1138 /* If the segment is not tagged as lost, 1139 * we do not clear RETRANS, believing 1140 * that retransmission is still in flight. 1141 */ 1142 if (sacked & TCPCB_LOST) { 1143 TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS); 1144 tp->lost_out -= tcp_skb_pcount(skb); 1145 tp->retrans_out -= tcp_skb_pcount(skb); 1146 1147 /* clear lost hint */ 1148 tp->retransmit_skb_hint = NULL; 1149 } 1150 } else { 1151 /* New sack for not retransmitted frame, 1152 * which was in hole. It is reordering. 1153 */ 1154 if (!(sacked & TCPCB_RETRANS) && 1155 fack_count < prior_fackets) 1156 reord = min(fack_count, reord); 1157 1158 if (sacked & TCPCB_LOST) { 1159 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; 1160 tp->lost_out -= tcp_skb_pcount(skb); 1161 1162 /* clear lost hint */ 1163 tp->retransmit_skb_hint = NULL; 1164 } 1165 /* SACK enhanced F-RTO detection. 1166 * Set flag if and only if non-rexmitted 1167 * segments below frto_highmark are 1168 * SACKed (RFC4138; Appendix B). 1169 * Clearing correct due to in-order walk 1170 */ 1171 if (after(end_seq, tp->frto_highmark)) { 1172 flag &= ~FLAG_ONLY_ORIG_SACKED; 1173 } else { 1174 if (!(sacked & TCPCB_RETRANS)) 1175 flag |= FLAG_ONLY_ORIG_SACKED; 1176 } 1177 } 1178 1179 TCP_SKB_CB(skb)->sacked |= TCPCB_SACKED_ACKED; 1180 flag |= FLAG_DATA_SACKED; 1181 tp->sacked_out += tcp_skb_pcount(skb); 1182 1183 if (fack_count > tp->fackets_out) 1184 tp->fackets_out = fack_count; 1185 } else { 1186 if (dup_sack && (sacked&TCPCB_RETRANS)) 1187 reord = min(fack_count, reord); 1188 } 1189 1190 /* D-SACK. We can detect redundant retransmission 1191 * in S|R and plain R frames and clear it. 1192 * undo_retrans is decreased above, L|R frames 1193 * are accounted above as well. 1194 */ 1195 if (dup_sack && 1196 (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS)) { 1197 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 1198 tp->retrans_out -= tcp_skb_pcount(skb); 1199 tp->retransmit_skb_hint = NULL; 1200 } 1201 } 1202 } 1203 1204 /* Check for lost retransmit. This superb idea is 1205 * borrowed from "ratehalving". Event "C". 1206 * Later note: FACK people cheated me again 8), 1207 * we have to account for reordering! Ugly, 1208 * but should help. 1209 */ 1210 if (lost_retrans && icsk->icsk_ca_state == TCP_CA_Recovery) { 1211 struct sk_buff *skb; 1212 1213 sk_stream_for_retrans_queue(skb, sk) { 1214 if (after(TCP_SKB_CB(skb)->seq, lost_retrans)) 1215 break; 1216 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 1217 continue; 1218 if ((TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) && 1219 after(lost_retrans, TCP_SKB_CB(skb)->ack_seq) && 1220 (IsFack(tp) || 1221 !before(lost_retrans, 1222 TCP_SKB_CB(skb)->ack_seq + tp->reordering * 1223 tp->mss_cache))) { 1224 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 1225 tp->retrans_out -= tcp_skb_pcount(skb); 1226 1227 /* clear lost hint */ 1228 tp->retransmit_skb_hint = NULL; 1229 1230 if (!(TCP_SKB_CB(skb)->sacked&(TCPCB_LOST|TCPCB_SACKED_ACKED))) { 1231 tp->lost_out += tcp_skb_pcount(skb); 1232 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 1233 flag |= FLAG_DATA_SACKED; 1234 NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT); 1235 } 1236 } 1237 } 1238 } 1239 1240 tp->left_out = tp->sacked_out + tp->lost_out; 1241 1242 if ((reord < tp->fackets_out) && icsk->icsk_ca_state != TCP_CA_Loss && 1243 (tp->frto_highmark && after(tp->snd_una, tp->frto_highmark))) 1244 tcp_update_reordering(sk, ((tp->fackets_out + 1) - reord), 0); 1245 1246#if FASTRETRANS_DEBUG > 0 1247 BUG_TRAP((int)tp->sacked_out >= 0); 1248 BUG_TRAP((int)tp->lost_out >= 0); 1249 BUG_TRAP((int)tp->retrans_out >= 0); 1250 BUG_TRAP((int)tcp_packets_in_flight(tp) >= 0); 1251#endif 1252 return flag; 1253} 1254 1255/* F-RTO can only be used if these conditions are satisfied: 1256 * - there must be some unsent new data 1257 * - the advertised window should allow sending it 1258 * - TCP has never retransmitted anything other than head (SACK enhanced 1259 * variant from Appendix B of RFC4138 is more robust here) 1260 */ 1261int tcp_use_frto(struct sock *sk) 1262{ 1263 const struct tcp_sock *tp = tcp_sk(sk); 1264 struct sk_buff *skb; 1265 1266 if (!sysctl_tcp_frto || !sk->sk_send_head || 1267 after(TCP_SKB_CB(sk->sk_send_head)->end_seq, 1268 tp->snd_una + tp->snd_wnd)) 1269 return 0; 1270 1271 if (IsSackFrto()) 1272 return 1; 1273 1274 /* Avoid expensive walking of rexmit queue if possible */ 1275 if (tp->retrans_out > 1) 1276 return 0; 1277 1278 skb = skb_peek(&sk->sk_write_queue)->next; /* Skips head */ 1279 sk_stream_for_retrans_queue_from(skb, sk) { 1280 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS) 1281 return 0; 1282 /* Short-circuit when first non-SACKed skb has been checked */ 1283 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) 1284 break; 1285 } 1286 return 1; 1287} 1288 1289/* RTO occurred, but do not yet enter Loss state. Instead, defer RTO 1290 * recovery a bit and use heuristics in tcp_process_frto() to detect if 1291 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to 1292 * keep retrans_out counting accurate (with SACK F-RTO, other than head 1293 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS 1294 * bits are handled if the Loss state is really to be entered (in 1295 * tcp_enter_frto_loss). 1296 * 1297 * Do like tcp_enter_loss() would; when RTO expires the second time it 1298 * does: 1299 * "Reduce ssthresh if it has not yet been made inside this window." 1300 */ 1301void tcp_enter_frto(struct sock *sk) 1302{ 1303 const struct inet_connection_sock *icsk = inet_csk(sk); 1304 struct tcp_sock *tp = tcp_sk(sk); 1305 struct sk_buff *skb; 1306 1307 if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) || 1308 tp->snd_una == tp->high_seq || 1309 ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) && 1310 !icsk->icsk_retransmits)) { 1311 tp->prior_ssthresh = tcp_current_ssthresh(sk); 1312 /* Our state is too optimistic in ssthresh() call because cwnd 1313 * is not reduced until tcp_enter_frto_loss() when previous FRTO 1314 * recovery has not yet completed. Pattern would be this: RTO, 1315 * Cumulative ACK, RTO (2xRTO for the same segment does not end 1316 * up here twice). 1317 * RFC4138 should be more specific on what to do, even though 1318 * RTO is quite unlikely to occur after the first Cumulative ACK 1319 * due to back-off and complexity of triggering events ... 1320 */ 1321 if (tp->frto_counter) { 1322 u32 stored_cwnd; 1323 stored_cwnd = tp->snd_cwnd; 1324 tp->snd_cwnd = 2; 1325 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 1326 tp->snd_cwnd = stored_cwnd; 1327 } else { 1328 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 1329 } 1330 /* ... in theory, cong.control module could do "any tricks" in 1331 * ssthresh(), which means that ca_state, lost bits and lost_out 1332 * counter would have to be faked before the call occurs. We 1333 * consider that too expensive, unlikely and hacky, so modules 1334 * using these in ssthresh() must deal these incompatibility 1335 * issues if they receives CA_EVENT_FRTO and frto_counter != 0 1336 */ 1337 tcp_ca_event(sk, CA_EVENT_FRTO); 1338 } 1339 1340 tp->undo_marker = tp->snd_una; 1341 tp->undo_retrans = 0; 1342 1343 skb = skb_peek(&sk->sk_write_queue); 1344 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 1345 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 1346 tp->retrans_out -= tcp_skb_pcount(skb); 1347 } 1348 tcp_sync_left_out(tp); 1349 1350 /* Earlier loss recovery underway (see RFC4138; Appendix B). 1351 * The last condition is necessary at least in tp->frto_counter case. 1352 */ 1353 if (IsSackFrto() && (tp->frto_counter || 1354 ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) && 1355 after(tp->high_seq, tp->snd_una)) { 1356 tp->frto_highmark = tp->high_seq; 1357 } else { 1358 tp->frto_highmark = tp->snd_nxt; 1359 } 1360 tcp_set_ca_state(sk, TCP_CA_Disorder); 1361 tp->high_seq = tp->snd_nxt; 1362 tp->frto_counter = 1; 1363} 1364 1365/* Enter Loss state after F-RTO was applied. Dupack arrived after RTO, 1366 * which indicates that we should follow the traditional RTO recovery, 1367 * i.e. mark everything lost and do go-back-N retransmission. 1368 */ 1369static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag) 1370{ 1371 struct tcp_sock *tp = tcp_sk(sk); 1372 struct sk_buff *skb; 1373 int cnt = 0; 1374 1375 tp->sacked_out = 0; 1376 tp->lost_out = 0; 1377 tp->fackets_out = 0; 1378 tp->retrans_out = 0; 1379 1380 sk_stream_for_retrans_queue(skb, sk) { 1381 cnt += tcp_skb_pcount(skb); 1382 /* 1383 * Count the retransmission made on RTO correctly (only when 1384 * waiting for the first ACK and did not get it)... 1385 */ 1386 if ((tp->frto_counter == 1) && !(flag&FLAG_DATA_ACKED)) { 1387 tp->retrans_out += tcp_skb_pcount(skb); 1388 /* ...enter this if branch just for the first segment */ 1389 flag |= FLAG_DATA_ACKED; 1390 } else { 1391 TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS); 1392 } 1393 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) { 1394 1395 /* Do not mark those segments lost that were 1396 * forward transmitted after RTO 1397 */ 1398 if (!after(TCP_SKB_CB(skb)->end_seq, 1399 tp->frto_highmark)) { 1400 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 1401 tp->lost_out += tcp_skb_pcount(skb); 1402 } 1403 } else { 1404 tp->sacked_out += tcp_skb_pcount(skb); 1405 tp->fackets_out = cnt; 1406 } 1407 } 1408 tcp_sync_left_out(tp); 1409 1410 tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments; 1411 tp->snd_cwnd_cnt = 0; 1412 tp->snd_cwnd_stamp = tcp_time_stamp; 1413 tp->undo_marker = 0; 1414 tp->frto_counter = 0; 1415 1416 tp->reordering = min_t(unsigned int, tp->reordering, 1417 sysctl_tcp_reordering); 1418 tcp_set_ca_state(sk, TCP_CA_Loss); 1419 tp->high_seq = tp->frto_highmark; 1420 TCP_ECN_queue_cwr(tp); 1421 1422 clear_all_retrans_hints(tp); 1423} 1424 1425void tcp_clear_retrans(struct tcp_sock *tp) 1426{ 1427 tp->left_out = 0; 1428 tp->retrans_out = 0; 1429 1430 tp->fackets_out = 0; 1431 tp->sacked_out = 0; 1432 tp->lost_out = 0; 1433 1434 tp->undo_marker = 0; 1435 tp->undo_retrans = 0; 1436} 1437 1438/* Enter Loss state. If "how" is not zero, forget all SACK information 1439 * and reset tags completely, otherwise preserve SACKs. If receiver 1440 * dropped its ofo queue, we will know this due to reneging detection. 1441 */ 1442void tcp_enter_loss(struct sock *sk, int how) 1443{ 1444 const struct inet_connection_sock *icsk = inet_csk(sk); 1445 struct tcp_sock *tp = tcp_sk(sk); 1446 struct sk_buff *skb; 1447 int cnt = 0; 1448 1449 /* Reduce ssthresh if it has not yet been made inside this window. */ 1450 if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq || 1451 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) { 1452 tp->prior_ssthresh = tcp_current_ssthresh(sk); 1453 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 1454 tcp_ca_event(sk, CA_EVENT_LOSS); 1455 } 1456 tp->snd_cwnd = 1; 1457 tp->snd_cwnd_cnt = 0; 1458 tp->snd_cwnd_stamp = tcp_time_stamp; 1459 1460 tp->bytes_acked = 0; 1461 tcp_clear_retrans(tp); 1462 1463 /* Push undo marker, if it was plain RTO and nothing 1464 * was retransmitted. */ 1465 if (!how) 1466 tp->undo_marker = tp->snd_una; 1467 1468 sk_stream_for_retrans_queue(skb, sk) { 1469 cnt += tcp_skb_pcount(skb); 1470 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS) 1471 tp->undo_marker = 0; 1472 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED; 1473 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) { 1474 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED; 1475 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 1476 tp->lost_out += tcp_skb_pcount(skb); 1477 } else { 1478 tp->sacked_out += tcp_skb_pcount(skb); 1479 tp->fackets_out = cnt; 1480 } 1481 } 1482 tcp_sync_left_out(tp); 1483 1484 tp->reordering = min_t(unsigned int, tp->reordering, 1485 sysctl_tcp_reordering); 1486 tcp_set_ca_state(sk, TCP_CA_Loss); 1487 tp->high_seq = tp->snd_nxt; 1488 TCP_ECN_queue_cwr(tp); 1489 1490 clear_all_retrans_hints(tp); 1491} 1492 1493static int tcp_check_sack_reneging(struct sock *sk) 1494{ 1495 struct sk_buff *skb; 1496 1497 /* If ACK arrived pointing to a remembered SACK, 1498 * it means that our remembered SACKs do not reflect 1499 * real state of receiver i.e. 1500 * receiver _host_ is heavily congested (or buggy). 1501 * Do processing similar to RTO timeout. 1502 */ 1503 if ((skb = skb_peek(&sk->sk_write_queue)) != NULL && 1504 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) { 1505 struct inet_connection_sock *icsk = inet_csk(sk); 1506 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING); 1507 1508 tcp_enter_loss(sk, 1); 1509 icsk->icsk_retransmits++; 1510 tcp_retransmit_skb(sk, skb_peek(&sk->sk_write_queue)); 1511 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 1512 icsk->icsk_rto, TCP_RTO_MAX); 1513 return 1; 1514 } 1515 return 0; 1516} 1517 1518static inline int tcp_fackets_out(struct tcp_sock *tp) 1519{ 1520 return IsReno(tp) ? tp->sacked_out+1 : tp->fackets_out; 1521} 1522 1523static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb) 1524{ 1525 return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto); 1526} 1527 1528static inline int tcp_head_timedout(struct sock *sk, struct tcp_sock *tp) 1529{ 1530 return tp->packets_out && 1531 tcp_skb_timedout(sk, skb_peek(&sk->sk_write_queue)); 1532} 1533 1534/* Linux NewReno/SACK/FACK/ECN state machine. 1535 * -------------------------------------- 1536 * 1537 * "Open" Normal state, no dubious events, fast path. 1538 * "Disorder" In all the respects it is "Open", 1539 * but requires a bit more attention. It is entered when 1540 * we see some SACKs or dupacks. It is split of "Open" 1541 * mainly to move some processing from fast path to slow one. 1542 * "CWR" CWND was reduced due to some Congestion Notification event. 1543 * It can be ECN, ICMP source quench, local device congestion. 1544 * "Recovery" CWND was reduced, we are fast-retransmitting. 1545 * "Loss" CWND was reduced due to RTO timeout or SACK reneging. 1546 * 1547 * tcp_fastretrans_alert() is entered: 1548 * - each incoming ACK, if state is not "Open" 1549 * - when arrived ACK is unusual, namely: 1550 * * SACK 1551 * * Duplicate ACK. 1552 * * ECN ECE. 1553 * 1554 * Counting packets in flight is pretty simple. 1555 * 1556 * in_flight = packets_out - left_out + retrans_out 1557 * 1558 * packets_out is SND.NXT-SND.UNA counted in packets. 1559 * 1560 * retrans_out is number of retransmitted segments. 1561 * 1562 * left_out is number of segments left network, but not ACKed yet. 1563 * 1564 * left_out = sacked_out + lost_out 1565 * 1566 * sacked_out: Packets, which arrived to receiver out of order 1567 * and hence not ACKed. With SACKs this number is simply 1568 * amount of SACKed data. Even without SACKs 1569 * it is easy to give pretty reliable estimate of this number, 1570 * counting duplicate ACKs. 1571 * 1572 * lost_out: Packets lost by network. TCP has no explicit 1573 * "loss notification" feedback from network (for now). 1574 * It means that this number can be only _guessed_. 1575 * Actually, it is the heuristics to predict lossage that 1576 * distinguishes different algorithms. 1577 * 1578 * F.e. after RTO, when all the queue is considered as lost, 1579 * lost_out = packets_out and in_flight = retrans_out. 1580 * 1581 * Essentially, we have now two algorithms counting 1582 * lost packets. 1583 * 1584 * FACK: It is the simplest heuristics. As soon as we decided 1585 * that something is lost, we decide that _all_ not SACKed 1586 * packets until the most forward SACK are lost. I.e. 1587 * lost_out = fackets_out - sacked_out and left_out = fackets_out. 1588 * It is absolutely correct estimate, if network does not reorder 1589 * packets. And it loses any connection to reality when reordering 1590 * takes place. We use FACK by default until reordering 1591 * is suspected on the path to this destination. 1592 * 1593 * NewReno: when Recovery is entered, we assume that one segment 1594 * is lost (classic Reno). While we are in Recovery and 1595 * a partial ACK arrives, we assume that one more packet 1596 * is lost (NewReno). This heuristics are the same in NewReno 1597 * and SACK. 1598 * 1599 * Imagine, that's all! Forget about all this shamanism about CWND inflation 1600 * deflation etc. CWND is real congestion window, never inflated, changes 1601 * only according to classic VJ rules. 1602 * 1603 * Really tricky (and requiring careful tuning) part of algorithm 1604 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue(). 1605 * The first determines the moment _when_ we should reduce CWND and, 1606 * hence, slow down forward transmission. In fact, it determines the moment 1607 * when we decide that hole is caused by loss, rather than by a reorder. 1608 * 1609 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill 1610 * holes, caused by lost packets. 1611 * 1612 * And the most logically complicated part of algorithm is undo 1613 * heuristics. We detect false retransmits due to both too early 1614 * fast retransmit (reordering) and underestimated RTO, analyzing 1615 * timestamps and D-SACKs. When we detect that some segments were 1616 * retransmitted by mistake and CWND reduction was wrong, we undo 1617 * window reduction and abort recovery phase. This logic is hidden 1618 * inside several functions named tcp_try_undo_<something>. 1619 */ 1620 1621/* This function decides, when we should leave Disordered state 1622 * and enter Recovery phase, reducing congestion window. 1623 * 1624 * Main question: may we further continue forward transmission 1625 * with the same cwnd? 1626 */ 1627static int tcp_time_to_recover(struct sock *sk, struct tcp_sock *tp) 1628{ 1629 __u32 packets_out; 1630 1631 /* Do not perform any recovery during FRTO algorithm */ 1632 if (tp->frto_counter) 1633 return 0; 1634 1635 /* Trick#1: The loss is proven. */ 1636 if (tp->lost_out) 1637 return 1; 1638 1639 /* Not-A-Trick#2 : Classic rule... */ 1640 if (tcp_fackets_out(tp) > tp->reordering) 1641 return 1; 1642 1643 /* Trick#3 : when we use RFC2988 timer restart, fast 1644 * retransmit can be triggered by timeout of queue head. 1645 */ 1646 if (tcp_head_timedout(sk, tp)) 1647 return 1; 1648 1649 /* Trick#4: It is still not OK... But will it be useful to delay 1650 * recovery more? 1651 */ 1652 packets_out = tp->packets_out; 1653 if (packets_out <= tp->reordering && 1654 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) && 1655 !tcp_may_send_now(sk, tp)) { 1656 /* We have nothing to send. This connection is limited 1657 * either by receiver window or by application. 1658 */ 1659 return 1; 1660 } 1661 1662 return 0; 1663} 1664 1665/* If we receive more dupacks than we expected counting segments 1666 * in assumption of absent reordering, interpret this as reordering. 1667 * The only another reason could be bug in receiver TCP. 1668 */ 1669static void tcp_check_reno_reordering(struct sock *sk, const int addend) 1670{ 1671 struct tcp_sock *tp = tcp_sk(sk); 1672 u32 holes; 1673 1674 holes = max(tp->lost_out, 1U); 1675 holes = min(holes, tp->packets_out); 1676 1677 if ((tp->sacked_out + holes) > tp->packets_out) { 1678 tp->sacked_out = tp->packets_out - holes; 1679 tcp_update_reordering(sk, tp->packets_out + addend, 0); 1680 } 1681} 1682 1683/* Emulate SACKs for SACKless connection: account for a new dupack. */ 1684 1685static void tcp_add_reno_sack(struct sock *sk) 1686{ 1687 struct tcp_sock *tp = tcp_sk(sk); 1688 tp->sacked_out++; 1689 tcp_check_reno_reordering(sk, 0); 1690 tcp_sync_left_out(tp); 1691} 1692 1693/* Account for ACK, ACKing some data in Reno Recovery phase. */ 1694 1695static void tcp_remove_reno_sacks(struct sock *sk, struct tcp_sock *tp, int acked) 1696{ 1697 if (acked > 0) { 1698 /* One ACK acked hole. The rest eat duplicate ACKs. */ 1699 if (acked-1 >= tp->sacked_out) 1700 tp->sacked_out = 0; 1701 else 1702 tp->sacked_out -= acked-1; 1703 } 1704 tcp_check_reno_reordering(sk, acked); 1705 tcp_sync_left_out(tp); 1706} 1707 1708static inline void tcp_reset_reno_sack(struct tcp_sock *tp) 1709{ 1710 tp->sacked_out = 0; 1711 tp->left_out = tp->lost_out; 1712} 1713 1714/* Mark head of queue up as lost. */ 1715static void tcp_mark_head_lost(struct sock *sk, struct tcp_sock *tp, 1716 int packets, u32 high_seq) 1717{ 1718 struct sk_buff *skb; 1719 int cnt; 1720 1721 BUG_TRAP(packets <= tp->packets_out); 1722 if (tp->lost_skb_hint) { 1723 skb = tp->lost_skb_hint; 1724 cnt = tp->lost_cnt_hint; 1725 } else { 1726 skb = sk->sk_write_queue.next; 1727 cnt = 0; 1728 } 1729 1730 sk_stream_for_retrans_queue_from(skb, sk) { 1731 /* TODO: do this better */ 1732 /* this is not the most efficient way to do this... */ 1733 tp->lost_skb_hint = skb; 1734 tp->lost_cnt_hint = cnt; 1735 cnt += tcp_skb_pcount(skb); 1736 if (cnt > packets || after(TCP_SKB_CB(skb)->end_seq, high_seq)) 1737 break; 1738 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) { 1739 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 1740 tp->lost_out += tcp_skb_pcount(skb); 1741 1742 /* clear xmit_retransmit_queue hints 1743 * if this is beyond hint */ 1744 if(tp->retransmit_skb_hint != NULL && 1745 before(TCP_SKB_CB(skb)->seq, 1746 TCP_SKB_CB(tp->retransmit_skb_hint)->seq)) { 1747 1748 tp->retransmit_skb_hint = NULL; 1749 } 1750 } 1751 } 1752 tcp_sync_left_out(tp); 1753} 1754 1755/* Account newly detected lost packet(s) */ 1756 1757static void tcp_update_scoreboard(struct sock *sk, struct tcp_sock *tp) 1758{ 1759 if (IsFack(tp)) { 1760 int lost = tp->fackets_out - tp->reordering; 1761 if (lost <= 0) 1762 lost = 1; 1763 tcp_mark_head_lost(sk, tp, lost, tp->high_seq); 1764 } else { 1765 tcp_mark_head_lost(sk, tp, 1, tp->high_seq); 1766 } 1767 1768 /* New heuristics: it is possible only after we switched 1769 * to restart timer each time when something is ACKed. 1770 * Hence, we can detect timed out packets during fast 1771 * retransmit without falling to slow start. 1772 */ 1773 if (!IsReno(tp) && tcp_head_timedout(sk, tp)) { 1774 struct sk_buff *skb; 1775 1776 skb = tp->scoreboard_skb_hint ? tp->scoreboard_skb_hint 1777 : sk->sk_write_queue.next; 1778 1779 sk_stream_for_retrans_queue_from(skb, sk) { 1780 if (!tcp_skb_timedout(sk, skb)) 1781 break; 1782 1783 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) { 1784 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 1785 tp->lost_out += tcp_skb_pcount(skb); 1786 1787 /* clear xmit_retrans hint */ 1788 if (tp->retransmit_skb_hint && 1789 before(TCP_SKB_CB(skb)->seq, 1790 TCP_SKB_CB(tp->retransmit_skb_hint)->seq)) 1791 1792 tp->retransmit_skb_hint = NULL; 1793 } 1794 } 1795 1796 tp->scoreboard_skb_hint = skb; 1797 1798 tcp_sync_left_out(tp); 1799 } 1800} 1801 1802/* CWND moderation, preventing bursts due to too big ACKs 1803 * in dubious situations. 1804 */ 1805static inline void tcp_moderate_cwnd(struct tcp_sock *tp) 1806{ 1807 tp->snd_cwnd = min(tp->snd_cwnd, 1808 tcp_packets_in_flight(tp)+tcp_max_burst(tp)); 1809 tp->snd_cwnd_stamp = tcp_time_stamp; 1810} 1811 1812/* Lower bound on congestion window is slow start threshold 1813 * unless congestion avoidance choice decides to overide it. 1814 */ 1815static inline u32 tcp_cwnd_min(const struct sock *sk) 1816{ 1817 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 1818 1819 return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh; 1820} 1821 1822/* Decrease cwnd each second ack. */ 1823static void tcp_cwnd_down(struct sock *sk) 1824{ 1825 struct tcp_sock *tp = tcp_sk(sk); 1826 int decr = tp->snd_cwnd_cnt + 1; 1827 1828 tp->snd_cwnd_cnt = decr&1; 1829 decr >>= 1; 1830 1831 if (decr && tp->snd_cwnd > tcp_cwnd_min(sk)) 1832 tp->snd_cwnd -= decr; 1833 1834 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp)+1); 1835 tp->snd_cwnd_stamp = tcp_time_stamp; 1836} 1837 1838/* Nothing was retransmitted or returned timestamp is less 1839 * than timestamp of the first retransmission. 1840 */ 1841static inline int tcp_packet_delayed(struct tcp_sock *tp) 1842{ 1843 return !tp->retrans_stamp || 1844 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 1845 (__s32)(tp->rx_opt.rcv_tsecr - tp->retrans_stamp) < 0); 1846} 1847 1848/* Undo procedures. */ 1849 1850#if FASTRETRANS_DEBUG > 1 1851static void DBGUNDO(struct sock *sk, struct tcp_sock *tp, const char *msg) 1852{ 1853 struct inet_sock *inet = inet_sk(sk); 1854 printk(KERN_DEBUG "Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n", 1855 msg, 1856 NIPQUAD(inet->daddr), ntohs(inet->dport), 1857 tp->snd_cwnd, tp->left_out, 1858 tp->snd_ssthresh, tp->prior_ssthresh, 1859 tp->packets_out); 1860} 1861#else 1862#define DBGUNDO(x...) do { } while (0) 1863#endif 1864 1865static void tcp_undo_cwr(struct sock *sk, const int undo) 1866{ 1867 struct tcp_sock *tp = tcp_sk(sk); 1868 1869 if (tp->prior_ssthresh) { 1870 const struct inet_connection_sock *icsk = inet_csk(sk); 1871 1872 if (icsk->icsk_ca_ops->undo_cwnd) 1873 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk); 1874 else 1875 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh<<1); 1876 1877 if (undo && tp->prior_ssthresh > tp->snd_ssthresh) { 1878 tp->snd_ssthresh = tp->prior_ssthresh; 1879 TCP_ECN_withdraw_cwr(tp); 1880 } 1881 } else { 1882 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh); 1883 } 1884 tcp_moderate_cwnd(tp); 1885 tp->snd_cwnd_stamp = tcp_time_stamp; 1886 1887 /* There is something screwy going on with the retrans hints after 1888 an undo */ 1889 clear_all_retrans_hints(tp); 1890} 1891 1892static inline int tcp_may_undo(struct tcp_sock *tp) 1893{ 1894 return tp->undo_marker && 1895 (!tp->undo_retrans || tcp_packet_delayed(tp)); 1896} 1897 1898/* People celebrate: "We love our President!" */ 1899static int tcp_try_undo_recovery(struct sock *sk, struct tcp_sock *tp) 1900{ 1901 if (tcp_may_undo(tp)) { 1902 /* Happy end! We did not retransmit anything 1903 * or our original transmission succeeded. 1904 */ 1905 DBGUNDO(sk, tp, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans"); 1906 tcp_undo_cwr(sk, 1); 1907 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss) 1908 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO); 1909 else 1910 NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO); 1911 tp->undo_marker = 0; 1912 } 1913 if (tp->snd_una == tp->high_seq && IsReno(tp)) { 1914 /* Hold old state until something *above* high_seq 1915 * is ACKed. For Reno it is MUST to prevent false 1916 * fast retransmits (RFC2582). SACK TCP is safe. */ 1917 tcp_moderate_cwnd(tp); 1918 return 1; 1919 } 1920 tcp_set_ca_state(sk, TCP_CA_Open); 1921 return 0; 1922} 1923 1924/* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */ 1925static void tcp_try_undo_dsack(struct sock *sk, struct tcp_sock *tp) 1926{ 1927 if (tp->undo_marker && !tp->undo_retrans) { 1928 DBGUNDO(sk, tp, "D-SACK"); 1929 tcp_undo_cwr(sk, 1); 1930 tp->undo_marker = 0; 1931 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO); 1932 } 1933} 1934 1935/* Undo during fast recovery after partial ACK. */ 1936 1937static int tcp_try_undo_partial(struct sock *sk, struct tcp_sock *tp, 1938 int acked) 1939{ 1940 /* Partial ACK arrived. Force Hoe's retransmit. */ 1941 int failed = IsReno(tp) || tp->fackets_out>tp->reordering; 1942 1943 if (tcp_may_undo(tp)) { 1944 /* Plain luck! Hole if filled with delayed 1945 * packet, rather than with a retransmit. 1946 */ 1947 if (tp->retrans_out == 0) 1948 tp->retrans_stamp = 0; 1949 1950 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1); 1951 1952 DBGUNDO(sk, tp, "Hoe"); 1953 tcp_undo_cwr(sk, 0); 1954 NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO); 1955 1956 /* So... Do not make Hoe's retransmit yet. 1957 * If the first packet was delayed, the rest 1958 * ones are most probably delayed as well. 1959 */ 1960 failed = 0; 1961 } 1962 return failed; 1963} 1964 1965/* Undo during loss recovery after partial ACK. */ 1966static int tcp_try_undo_loss(struct sock *sk, struct tcp_sock *tp) 1967{ 1968 if (tcp_may_undo(tp)) { 1969 struct sk_buff *skb; 1970 sk_stream_for_retrans_queue(skb, sk) { 1971 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; 1972 } 1973 1974 clear_all_retrans_hints(tp); 1975 1976 DBGUNDO(sk, tp, "partial loss"); 1977 tp->lost_out = 0; 1978 tp->left_out = tp->sacked_out; 1979 tcp_undo_cwr(sk, 1); 1980 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO); 1981 inet_csk(sk)->icsk_retransmits = 0; 1982 tp->undo_marker = 0; 1983 if (!IsReno(tp)) 1984 tcp_set_ca_state(sk, TCP_CA_Open); 1985 return 1; 1986 } 1987 return 0; 1988} 1989 1990static inline void tcp_complete_cwr(struct sock *sk) 1991{ 1992 struct tcp_sock *tp = tcp_sk(sk); 1993 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh); 1994 tp->snd_cwnd_stamp = tcp_time_stamp; 1995 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR); 1996} 1997 1998static void tcp_try_to_open(struct sock *sk, struct tcp_sock *tp, int flag) 1999{ 2000 tp->left_out = tp->sacked_out; 2001 2002 if (tp->retrans_out == 0) 2003 tp->retrans_stamp = 0; 2004 2005 if (flag&FLAG_ECE) 2006 tcp_enter_cwr(sk); 2007 2008 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) { 2009 int state = TCP_CA_Open; 2010 2011 if (tp->left_out || tp->retrans_out || tp->undo_marker) 2012 state = TCP_CA_Disorder; 2013 2014 if (inet_csk(sk)->icsk_ca_state != state) { 2015 tcp_set_ca_state(sk, state); 2016 tp->high_seq = tp->snd_nxt; 2017 } 2018 tcp_moderate_cwnd(tp); 2019 } else { 2020 tcp_cwnd_down(sk); 2021 } 2022} 2023 2024static void tcp_mtup_probe_failed(struct sock *sk) 2025{ 2026 struct inet_connection_sock *icsk = inet_csk(sk); 2027 2028 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1; 2029 icsk->icsk_mtup.probe_size = 0; 2030} 2031 2032static void tcp_mtup_probe_success(struct sock *sk, struct sk_buff *skb) 2033{ 2034 struct tcp_sock *tp = tcp_sk(sk); 2035 struct inet_connection_sock *icsk = inet_csk(sk); 2036 2037 /* FIXME: breaks with very large cwnd */ 2038 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2039 tp->snd_cwnd = tp->snd_cwnd * 2040 tcp_mss_to_mtu(sk, tp->mss_cache) / 2041 icsk->icsk_mtup.probe_size; 2042 tp->snd_cwnd_cnt = 0; 2043 tp->snd_cwnd_stamp = tcp_time_stamp; 2044 tp->rcv_ssthresh = tcp_current_ssthresh(sk); 2045 2046 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size; 2047 icsk->icsk_mtup.probe_size = 0; 2048 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 2049} 2050 2051 2052/* Process an event, which can update packets-in-flight not trivially. 2053 * Main goal of this function is to calculate new estimate for left_out, 2054 * taking into account both packets sitting in receiver's buffer and 2055 * packets lost by network. 2056 * 2057 * Besides that it does CWND reduction, when packet loss is detected 2058 * and changes state of machine. 2059 * 2060 * It does _not_ decide what to send, it is made in function 2061 * tcp_xmit_retransmit_queue(). 2062 */ 2063static void 2064tcp_fastretrans_alert(struct sock *sk, u32 prior_snd_una, 2065 int prior_packets, int flag) 2066{ 2067 struct inet_connection_sock *icsk = inet_csk(sk); 2068 struct tcp_sock *tp = tcp_sk(sk); 2069 int is_dupack = (tp->snd_una == prior_snd_una && !(flag&FLAG_NOT_DUP)); 2070 2071 /* Some technical things: 2072 * 1. Reno does not count dupacks (sacked_out) automatically. */ 2073 if (!tp->packets_out) 2074 tp->sacked_out = 0; 2075 /* 2. SACK counts snd_fack in packets inaccurately. */ 2076 if (tp->sacked_out == 0) 2077 tp->fackets_out = 0; 2078 2079 /* Now state machine starts. 2080 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */ 2081 if (flag&FLAG_ECE) 2082 tp->prior_ssthresh = 0; 2083 2084 /* B. In all the states check for reneging SACKs. */ 2085 if (tp->sacked_out && tcp_check_sack_reneging(sk)) 2086 return; 2087 2088 /* C. Process data loss notification, provided it is valid. */ 2089 if ((flag&FLAG_DATA_LOST) && 2090 before(tp->snd_una, tp->high_seq) && 2091 icsk->icsk_ca_state != TCP_CA_Open && 2092 tp->fackets_out > tp->reordering) { 2093 tcp_mark_head_lost(sk, tp, tp->fackets_out-tp->reordering, tp->high_seq); 2094 NET_INC_STATS_BH(LINUX_MIB_TCPLOSS); 2095 } 2096 2097 /* D. Synchronize left_out to current state. */ 2098 tcp_sync_left_out(tp); 2099 2100 /* E. Check state exit conditions. State can be terminated 2101 * when high_seq is ACKed. */ 2102 if (icsk->icsk_ca_state == TCP_CA_Open) { 2103 BUG_TRAP(tp->retrans_out == 0); 2104 tp->retrans_stamp = 0; 2105 } else if (!before(tp->snd_una, tp->high_seq)) { 2106 switch (icsk->icsk_ca_state) { 2107 case TCP_CA_Loss: 2108 icsk->icsk_retransmits = 0; 2109 if (tcp_try_undo_recovery(sk, tp)) 2110 return; 2111 break; 2112 2113 case TCP_CA_CWR: 2114 /* CWR is to be held something *above* high_seq 2115 * is ACKed for CWR bit to reach receiver. */ 2116 if (tp->snd_una != tp->high_seq) { 2117 tcp_complete_cwr(sk); 2118 tcp_set_ca_state(sk, TCP_CA_Open); 2119 } 2120 break; 2121 2122 case TCP_CA_Disorder: 2123 tcp_try_undo_dsack(sk, tp); 2124 if (!tp->undo_marker || 2125 /* For SACK case do not Open to allow to undo 2126 * catching for all duplicate ACKs. */ 2127 IsReno(tp) || tp->snd_una != tp->high_seq) { 2128 tp->undo_marker = 0; 2129 tcp_set_ca_state(sk, TCP_CA_Open); 2130 } 2131 break; 2132 2133 case TCP_CA_Recovery: 2134 if (IsReno(tp)) 2135 tcp_reset_reno_sack(tp); 2136 if (tcp_try_undo_recovery(sk, tp)) 2137 return; 2138 tcp_complete_cwr(sk); 2139 break; 2140 } 2141 } 2142 2143 /* F. Process state. */ 2144 switch (icsk->icsk_ca_state) { 2145 case TCP_CA_Recovery: 2146 if (prior_snd_una == tp->snd_una) { 2147 if (IsReno(tp) && is_dupack) 2148 tcp_add_reno_sack(sk); 2149 } else { 2150 int acked = prior_packets - tp->packets_out; 2151 if (IsReno(tp)) 2152 tcp_remove_reno_sacks(sk, tp, acked); 2153 is_dupack = tcp_try_undo_partial(sk, tp, acked); 2154 } 2155 break; 2156 case TCP_CA_Loss: 2157 if (flag&FLAG_DATA_ACKED) 2158 icsk->icsk_retransmits = 0; 2159 if (!tcp_try_undo_loss(sk, tp)) { 2160 tcp_moderate_cwnd(tp); 2161 tcp_xmit_retransmit_queue(sk); 2162 return; 2163 } 2164 if (icsk->icsk_ca_state != TCP_CA_Open) 2165 return; 2166 /* Loss is undone; fall through to processing in Open state. */ 2167 default: 2168 if (IsReno(tp)) { 2169 if (tp->snd_una != prior_snd_una) 2170 tcp_reset_reno_sack(tp); 2171 if (is_dupack) 2172 tcp_add_reno_sack(sk); 2173 } 2174 2175 if (icsk->icsk_ca_state == TCP_CA_Disorder) 2176 tcp_try_undo_dsack(sk, tp); 2177 2178 if (!tcp_time_to_recover(sk, tp)) { 2179 tcp_try_to_open(sk, tp, flag); 2180 return; 2181 } 2182 2183 /* MTU probe failure: don't reduce cwnd */ 2184 if (icsk->icsk_ca_state < TCP_CA_CWR && 2185 icsk->icsk_mtup.probe_size && 2186 tp->snd_una == tp->mtu_probe.probe_seq_start) { 2187 tcp_mtup_probe_failed(sk); 2188 /* Restores the reduction we did in tcp_mtup_probe() */ 2189 tp->snd_cwnd++; 2190 tcp_simple_retransmit(sk); 2191 return; 2192 } 2193 2194 /* Otherwise enter Recovery state */ 2195 2196 if (IsReno(tp)) 2197 NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY); 2198 else 2199 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY); 2200 2201 tp->high_seq = tp->snd_nxt; 2202 tp->prior_ssthresh = 0; 2203 tp->undo_marker = tp->snd_una; 2204 tp->undo_retrans = tp->retrans_out; 2205 2206 if (icsk->icsk_ca_state < TCP_CA_CWR) { 2207 if (!(flag&FLAG_ECE)) 2208 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2209 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 2210 TCP_ECN_queue_cwr(tp); 2211 } 2212 2213 tp->bytes_acked = 0; 2214 tp->snd_cwnd_cnt = 0; 2215 tcp_set_ca_state(sk, TCP_CA_Recovery); 2216 } 2217 2218 if (is_dupack || tcp_head_timedout(sk, tp)) 2219 tcp_update_scoreboard(sk, tp); 2220 tcp_cwnd_down(sk); 2221 tcp_xmit_retransmit_queue(sk); 2222} 2223 2224/* Read draft-ietf-tcplw-high-performance before mucking 2225 * with this code. (Supersedes RFC1323) 2226 */ 2227static void tcp_ack_saw_tstamp(struct sock *sk, int flag) 2228{ 2229 /* RTTM Rule: A TSecr value received in a segment is used to 2230 * update the averaged RTT measurement only if the segment 2231 * acknowledges some new data, i.e., only if it advances the 2232 * left edge of the send window. 2233 * 2234 * See draft-ietf-tcplw-high-performance-00, section 3.3. 2235 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru> 2236 * 2237 * Changed: reset backoff as soon as we see the first valid sample. 2238 * If we do not, we get strongly overestimated rto. With timestamps 2239 * samples are accepted even from very old segments: f.e., when rtt=1 2240 * increases to 8, we retransmit 5 times and after 8 seconds delayed 2241 * answer arrives rto becomes 120 seconds! If at least one of segments 2242 * in window is lost... Voila. --ANK (010210) 2243 */ 2244 struct tcp_sock *tp = tcp_sk(sk); 2245 const __u32 seq_rtt = tcp_time_stamp - tp->rx_opt.rcv_tsecr; 2246 tcp_rtt_estimator(sk, seq_rtt); 2247 tcp_set_rto(sk); 2248 inet_csk(sk)->icsk_backoff = 0; 2249 tcp_bound_rto(sk); 2250} 2251 2252static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag) 2253{ 2254 /* We don't have a timestamp. Can only use 2255 * packets that are not retransmitted to determine 2256 * rtt estimates. Also, we must not reset the 2257 * backoff for rto until we get a non-retransmitted 2258 * packet. This allows us to deal with a situation 2259 * where the network delay has increased suddenly. 2260 * I.e. Karn's algorithm. (SIGCOMM '87, p5.) 2261 */ 2262 2263 if (flag & FLAG_RETRANS_DATA_ACKED) 2264 return; 2265 2266 tcp_rtt_estimator(sk, seq_rtt); 2267 tcp_set_rto(sk); 2268 inet_csk(sk)->icsk_backoff = 0; 2269 tcp_bound_rto(sk); 2270} 2271 2272static inline void tcp_ack_update_rtt(struct sock *sk, const int flag, 2273 const s32 seq_rtt) 2274{ 2275 const struct tcp_sock *tp = tcp_sk(sk); 2276 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */ 2277 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr) 2278 tcp_ack_saw_tstamp(sk, flag); 2279 else if (seq_rtt >= 0) 2280 tcp_ack_no_tstamp(sk, seq_rtt, flag); 2281} 2282 2283static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 rtt, 2284 u32 in_flight, int good) 2285{ 2286 const struct inet_connection_sock *icsk = inet_csk(sk); 2287 icsk->icsk_ca_ops->cong_avoid(sk, ack, rtt, in_flight, good); 2288 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp; 2289} 2290 2291/* Restart timer after forward progress on connection. 2292 * RFC2988 recommends to restart timer to now+rto. 2293 */ 2294 2295static void tcp_ack_packets_out(struct sock *sk, struct tcp_sock *tp) 2296{ 2297 if (!tp->packets_out) { 2298 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS); 2299 } else { 2300 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 2301 } 2302} 2303 2304static int tcp_tso_acked(struct sock *sk, struct sk_buff *skb, 2305 __u32 now, __s32 *seq_rtt) 2306{ 2307 struct tcp_sock *tp = tcp_sk(sk); 2308 struct tcp_skb_cb *scb = TCP_SKB_CB(skb); 2309 __u32 seq = tp->snd_una; 2310 __u32 packets_acked; 2311 int acked = 0; 2312 2313 /* If we get here, the whole TSO packet has not been 2314 * acked. 2315 */ 2316 BUG_ON(!after(scb->end_seq, seq)); 2317 2318 packets_acked = tcp_skb_pcount(skb); 2319 if (tcp_trim_head(sk, skb, seq - scb->seq)) 2320 return 0; 2321 packets_acked -= tcp_skb_pcount(skb); 2322 2323 if (packets_acked) { 2324 __u8 sacked = scb->sacked; 2325 2326 acked |= FLAG_DATA_ACKED; 2327 if (sacked) { 2328 if (sacked & TCPCB_RETRANS) { 2329 if (sacked & TCPCB_SACKED_RETRANS) 2330 tp->retrans_out -= packets_acked; 2331 acked |= FLAG_RETRANS_DATA_ACKED; 2332 *seq_rtt = -1; 2333 } else if (*seq_rtt < 0) 2334 *seq_rtt = now - scb->when; 2335 if (sacked & TCPCB_SACKED_ACKED) 2336 tp->sacked_out -= packets_acked; 2337 if (sacked & TCPCB_LOST) 2338 tp->lost_out -= packets_acked; 2339 if (sacked & TCPCB_URG) { 2340 if (tp->urg_mode && 2341 !before(seq, tp->snd_up)) 2342 tp->urg_mode = 0; 2343 } 2344 } else if (*seq_rtt < 0) 2345 *seq_rtt = now - scb->when; 2346 2347 if (tp->fackets_out) { 2348 __u32 dval = min(tp->fackets_out, packets_acked); 2349 tp->fackets_out -= dval; 2350 } 2351 tp->packets_out -= packets_acked; 2352 2353 BUG_ON(tcp_skb_pcount(skb) == 0); 2354 BUG_ON(!before(scb->seq, scb->end_seq)); 2355 } 2356 2357 return acked; 2358} 2359 2360static u32 tcp_usrtt(struct timeval *tv) 2361{ 2362 struct timeval now; 2363 2364 do_gettimeofday(&now); 2365 return (now.tv_sec - tv->tv_sec) * 1000000 + (now.tv_usec - tv->tv_usec); 2366} 2367 2368/* Remove acknowledged frames from the retransmission queue. */ 2369static int tcp_clean_rtx_queue(struct sock *sk, __s32 *seq_rtt_p) 2370{ 2371 struct tcp_sock *tp = tcp_sk(sk); 2372 const struct inet_connection_sock *icsk = inet_csk(sk); 2373 struct sk_buff *skb; 2374 __u32 now = tcp_time_stamp; 2375 int acked = 0; 2376 __s32 seq_rtt = -1; 2377 u32 pkts_acked = 0; 2378 void (*rtt_sample)(struct sock *sk, u32 usrtt) 2379 = icsk->icsk_ca_ops->rtt_sample; 2380 struct timeval tv = { .tv_sec = 0, .tv_usec = 0 }; 2381 2382 while ((skb = skb_peek(&sk->sk_write_queue)) && 2383 skb != sk->sk_send_head) { 2384 struct tcp_skb_cb *scb = TCP_SKB_CB(skb); 2385 __u8 sacked = scb->sacked; 2386 2387 /* If our packet is before the ack sequence we can 2388 * discard it as it's confirmed to have arrived at 2389 * the other end. 2390 */ 2391 if (after(scb->end_seq, tp->snd_una)) { 2392 if (tcp_skb_pcount(skb) > 1 && 2393 after(tp->snd_una, scb->seq)) 2394 acked |= tcp_tso_acked(sk, skb, 2395 now, &seq_rtt); 2396 break; 2397 } 2398 2399 /* Initial outgoing SYN's get put onto the write_queue 2400 * just like anything else we transmit. It is not 2401 * true data, and if we misinform our callers that 2402 * this ACK acks real data, we will erroneously exit 2403 * connection startup slow start one packet too 2404 * quickly. This is severely frowned upon behavior. 2405 */ 2406 if (!(scb->flags & TCPCB_FLAG_SYN)) { 2407 acked |= FLAG_DATA_ACKED; 2408 ++pkts_acked; 2409 } else { 2410 acked |= FLAG_SYN_ACKED; 2411 tp->retrans_stamp = 0; 2412 } 2413 2414 /* MTU probing checks */ 2415 if (icsk->icsk_mtup.probe_size) { 2416 if (!after(tp->mtu_probe.probe_seq_end, TCP_SKB_CB(skb)->end_seq)) { 2417 tcp_mtup_probe_success(sk, skb); 2418 } 2419 } 2420 2421 if (sacked) { 2422 if (sacked & TCPCB_RETRANS) { 2423 if(sacked & TCPCB_SACKED_RETRANS) 2424 tp->retrans_out -= tcp_skb_pcount(skb); 2425 acked |= FLAG_RETRANS_DATA_ACKED; 2426 seq_rtt = -1; 2427 } else if (seq_rtt < 0) { 2428 seq_rtt = now - scb->when; 2429 skb_get_timestamp(skb, &tv); 2430 } 2431 if (sacked & TCPCB_SACKED_ACKED) 2432 tp->sacked_out -= tcp_skb_pcount(skb); 2433 if (sacked & TCPCB_LOST) 2434 tp->lost_out -= tcp_skb_pcount(skb); 2435 if (sacked & TCPCB_URG) { 2436 if (tp->urg_mode && 2437 !before(scb->end_seq, tp->snd_up)) 2438 tp->urg_mode = 0; 2439 } 2440 } else if (seq_rtt < 0) { 2441 seq_rtt = now - scb->when; 2442 skb_get_timestamp(skb, &tv); 2443 } 2444 tcp_dec_pcount_approx(&tp->fackets_out, skb); 2445 tcp_packets_out_dec(tp, skb); 2446 __skb_unlink(skb, &sk->sk_write_queue); 2447 sk_stream_free_skb(sk, skb); 2448 clear_all_retrans_hints(tp); 2449 } 2450 2451 if (acked&FLAG_ACKED) { 2452 tcp_ack_update_rtt(sk, acked, seq_rtt); 2453 tcp_ack_packets_out(sk, tp); 2454 if (rtt_sample && !(acked & FLAG_RETRANS_DATA_ACKED)) 2455 (*rtt_sample)(sk, tcp_usrtt(&tv)); 2456 2457 if (icsk->icsk_ca_ops->pkts_acked) 2458 icsk->icsk_ca_ops->pkts_acked(sk, pkts_acked); 2459 } 2460 2461#if FASTRETRANS_DEBUG > 0 2462 BUG_TRAP((int)tp->sacked_out >= 0); 2463 BUG_TRAP((int)tp->lost_out >= 0); 2464 BUG_TRAP((int)tp->retrans_out >= 0); 2465 if (!tp->packets_out && tp->rx_opt.sack_ok) { 2466 const struct inet_connection_sock *icsk = inet_csk(sk); 2467 if (tp->lost_out) { 2468 printk(KERN_DEBUG "Leak l=%u %d\n", 2469 tp->lost_out, icsk->icsk_ca_state); 2470 tp->lost_out = 0; 2471 } 2472 if (tp->sacked_out) { 2473 printk(KERN_DEBUG "Leak s=%u %d\n", 2474 tp->sacked_out, icsk->icsk_ca_state); 2475 tp->sacked_out = 0; 2476 } 2477 if (tp->retrans_out) { 2478 printk(KERN_DEBUG "Leak r=%u %d\n", 2479 tp->retrans_out, icsk->icsk_ca_state); 2480 tp->retrans_out = 0; 2481 } 2482 } 2483#endif 2484 *seq_rtt_p = seq_rtt; 2485 return acked; 2486} 2487 2488static void tcp_ack_probe(struct sock *sk) 2489{ 2490 const struct tcp_sock *tp = tcp_sk(sk); 2491 struct inet_connection_sock *icsk = inet_csk(sk); 2492 2493 /* Was it a usable window open? */ 2494 2495 if (!after(TCP_SKB_CB(sk->sk_send_head)->end_seq, 2496 tp->snd_una + tp->snd_wnd)) { 2497 icsk->icsk_backoff = 0; 2498 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0); 2499 /* Socket must be waked up by subsequent tcp_data_snd_check(). 2500 * This function is not for random using! 2501 */ 2502 } else { 2503 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 2504 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), 2505 TCP_RTO_MAX); 2506 } 2507} 2508 2509static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag) 2510{ 2511 return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) || 2512 inet_csk(sk)->icsk_ca_state != TCP_CA_Open); 2513} 2514 2515static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag) 2516{ 2517 const struct tcp_sock *tp = tcp_sk(sk); 2518 return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) && 2519 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR)); 2520} 2521 2522/* Check that window update is acceptable. 2523 * The function assumes that snd_una<=ack<=snd_next. 2524 */ 2525static inline int tcp_may_update_window(const struct tcp_sock *tp, const u32 ack, 2526 const u32 ack_seq, const u32 nwin) 2527{ 2528 return (after(ack, tp->snd_una) || 2529 after(ack_seq, tp->snd_wl1) || 2530 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd)); 2531} 2532 2533/* Update our send window. 2534 * 2535 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2 2536 * and in FreeBSD. NetBSD's one is even worse.) is wrong. 2537 */ 2538static int tcp_ack_update_window(struct sock *sk, struct tcp_sock *tp, 2539 struct sk_buff *skb, u32 ack, u32 ack_seq) 2540{ 2541 int flag = 0; 2542 u32 nwin = ntohs(skb->h.th->window); 2543 2544 if (likely(!skb->h.th->syn)) 2545 nwin <<= tp->rx_opt.snd_wscale; 2546 2547 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) { 2548 flag |= FLAG_WIN_UPDATE; 2549 tcp_update_wl(tp, ack, ack_seq); 2550 2551 if (tp->snd_wnd != nwin) { 2552 tp->snd_wnd = nwin; 2553 2554 /* Note, it is the only place, where 2555 * fast path is recovered for sending TCP. 2556 */ 2557 tp->pred_flags = 0; 2558 tcp_fast_path_check(sk, tp); 2559 2560 if (nwin > tp->max_window) { 2561 tp->max_window = nwin; 2562 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie); 2563 } 2564 } 2565 } 2566 2567 tp->snd_una = ack; 2568 2569 return flag; 2570} 2571 2572/* A very conservative spurious RTO response algorithm: reduce cwnd and 2573 * continue in congestion avoidance. 2574 */ 2575static void tcp_conservative_spur_to_response(struct tcp_sock *tp) 2576{ 2577 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh); 2578 tp->snd_cwnd_cnt = 0; 2579 tcp_moderate_cwnd(tp); 2580} 2581 2582/* F-RTO spurious RTO detection algorithm (RFC4138) 2583 * 2584 * F-RTO affects during two new ACKs following RTO (well, almost, see inline 2585 * comments). State (ACK number) is kept in frto_counter. When ACK advances 2586 * window (but not to or beyond highest sequence sent before RTO): 2587 * On First ACK, send two new segments out. 2588 * On Second ACK, RTO was likely spurious. Do spurious response (response 2589 * algorithm is not part of the F-RTO detection algorithm 2590 * given in RFC4138 but can be selected separately). 2591 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss 2592 * and TCP falls back to conventional RTO recovery. 2593 * 2594 * Rationale: if the RTO was spurious, new ACKs should arrive from the 2595 * original window even after we transmit two new data segments. 2596 * 2597 * SACK version: 2598 * on first step, wait until first cumulative ACK arrives, then move to 2599 * the second step. In second step, the next ACK decides. 2600 * 2601 * F-RTO is implemented (mainly) in four functions: 2602 * - tcp_use_frto() is used to determine if TCP is can use F-RTO 2603 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is 2604 * called when tcp_use_frto() showed green light 2605 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm 2606 * - tcp_enter_frto_loss() is called if there is not enough evidence 2607 * to prove that the RTO is indeed spurious. It transfers the control 2608 * from F-RTO to the conventional RTO recovery 2609 */ 2610static int tcp_process_frto(struct sock *sk, u32 prior_snd_una, int flag) 2611{ 2612 struct tcp_sock *tp = tcp_sk(sk); 2613 2614 tcp_sync_left_out(tp); 2615 2616 /* Duplicate the behavior from Loss state (fastretrans_alert) */ 2617 if (flag&FLAG_DATA_ACKED) 2618 inet_csk(sk)->icsk_retransmits = 0; 2619 2620 if (!before(tp->snd_una, tp->frto_highmark)) { 2621 tcp_enter_frto_loss(sk, tp->frto_counter + 1, flag); 2622 return 1; 2623 } 2624 2625 if (!IsSackFrto() || IsReno(tp)) { 2626 /* RFC4138 shortcoming in step 2; should also have case c): 2627 * ACK isn't duplicate nor advances window, e.g., opposite dir 2628 * data, winupdate 2629 */ 2630 if ((tp->snd_una == prior_snd_una) && (flag&FLAG_NOT_DUP) && 2631 !(flag&FLAG_FORWARD_PROGRESS)) 2632 return 1; 2633 2634 if (!(flag&FLAG_DATA_ACKED)) { 2635 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3), 2636 flag); 2637 return 1; 2638 } 2639 } else { 2640 if (!(flag&FLAG_DATA_ACKED) && (tp->frto_counter == 1)) { 2641 /* Prevent sending of new data. */ 2642 tp->snd_cwnd = min(tp->snd_cwnd, 2643 tcp_packets_in_flight(tp)); 2644 return 1; 2645 } 2646 2647 if ((tp->frto_counter == 2) && 2648 (!(flag&FLAG_FORWARD_PROGRESS) || 2649 ((flag&FLAG_DATA_SACKED) && !(flag&FLAG_ONLY_ORIG_SACKED)))) { 2650 /* RFC4138 shortcoming (see comment above) */ 2651 if (!(flag&FLAG_FORWARD_PROGRESS) && (flag&FLAG_NOT_DUP)) 2652 return 1; 2653 2654 tcp_enter_frto_loss(sk, 3, flag); 2655 return 1; 2656 } 2657 } 2658 2659 if (tp->frto_counter == 1) { 2660 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2; 2661 tp->frto_counter = 2; 2662 return 1; 2663 } else /* frto_counter == 2 */ { 2664 tcp_conservative_spur_to_response(tp); 2665 tp->frto_counter = 0; 2666 } 2667 return 0; 2668} 2669 2670/* This routine deals with incoming acks, but not outgoing ones. */ 2671static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag) 2672{ 2673 struct inet_connection_sock *icsk = inet_csk(sk); 2674 struct tcp_sock *tp = tcp_sk(sk); 2675 u32 prior_snd_una = tp->snd_una; 2676 u32 ack_seq = TCP_SKB_CB(skb)->seq; 2677 u32 ack = TCP_SKB_CB(skb)->ack_seq; 2678 u32 prior_in_flight; 2679 s32 seq_rtt; 2680 int prior_packets; 2681 int frto_cwnd = 0; 2682 2683 /* If the ack is newer than sent or older than previous acks 2684 * then we can probably ignore it. 2685 */ 2686 if (after(ack, tp->snd_nxt)) 2687 goto uninteresting_ack; 2688 2689 if (before(ack, prior_snd_una)) 2690 goto old_ack; 2691 2692 if (sysctl_tcp_abc) { 2693 if (icsk->icsk_ca_state < TCP_CA_CWR) 2694 tp->bytes_acked += ack - prior_snd_una; 2695 else if (icsk->icsk_ca_state == TCP_CA_Loss) 2696 /* we assume just one segment left network */ 2697 tp->bytes_acked += min(ack - prior_snd_una, tp->mss_cache); 2698 } 2699 2700 if (!(flag&FLAG_SLOWPATH) && after(ack, prior_snd_una)) { 2701 /* Window is constant, pure forward advance. 2702 * No more checks are required. 2703 * Note, we use the fact that SND.UNA>=SND.WL2. 2704 */ 2705 tcp_update_wl(tp, ack, ack_seq); 2706 tp->snd_una = ack; 2707 flag |= FLAG_WIN_UPDATE; 2708 2709 tcp_ca_event(sk, CA_EVENT_FAST_ACK); 2710 2711 NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS); 2712 } else { 2713 if (ack_seq != TCP_SKB_CB(skb)->end_seq) 2714 flag |= FLAG_DATA; 2715 else 2716 NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS); 2717 2718 flag |= tcp_ack_update_window(sk, tp, skb, ack, ack_seq); 2719 2720 if (TCP_SKB_CB(skb)->sacked) 2721 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una); 2722 2723 if (TCP_ECN_rcv_ecn_echo(tp, skb->h.th)) 2724 flag |= FLAG_ECE; 2725 2726 tcp_ca_event(sk, CA_EVENT_SLOW_ACK); 2727 } 2728 2729 /* We passed data and got it acked, remove any soft error 2730 * log. Something worked... 2731 */ 2732 sk->sk_err_soft = 0; 2733 tp->rcv_tstamp = tcp_time_stamp; 2734 prior_packets = tp->packets_out; 2735 if (!prior_packets) 2736 goto no_queue; 2737 2738 prior_in_flight = tcp_packets_in_flight(tp); 2739 2740 /* See if we can take anything off of the retransmit queue. */ 2741 flag |= tcp_clean_rtx_queue(sk, &seq_rtt); 2742 2743 if (tp->frto_counter) 2744 frto_cwnd = tcp_process_frto(sk, prior_snd_una, flag); 2745 2746 if (tcp_ack_is_dubious(sk, flag)) { 2747 /* Advance CWND, if state allows this. */ 2748 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd && 2749 tcp_may_raise_cwnd(sk, flag)) 2750 tcp_cong_avoid(sk, ack, seq_rtt, prior_in_flight, 0); 2751 tcp_fastretrans_alert(sk, prior_snd_una, prior_packets, flag); 2752 } else { 2753 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd) 2754 tcp_cong_avoid(sk, ack, seq_rtt, prior_in_flight, 1); 2755 } 2756 2757 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag&FLAG_NOT_DUP)) 2758 dst_confirm(sk->sk_dst_cache); 2759 2760 return 1; 2761 2762no_queue: 2763 icsk->icsk_probes_out = 0; 2764 2765 /* If this ack opens up a zero window, clear backoff. It was 2766 * being used to time the probes, and is probably far higher than 2767 * it needs to be for normal retransmission. 2768 */ 2769 if (sk->sk_send_head) 2770 tcp_ack_probe(sk); 2771 return 1; 2772 2773old_ack: 2774 if (TCP_SKB_CB(skb)->sacked) 2775 tcp_sacktag_write_queue(sk, skb, prior_snd_una); 2776 2777uninteresting_ack: 2778 SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt); 2779 return 0; 2780} 2781 2782 2783/* Look for tcp options. Normally only called on SYN and SYNACK packets. 2784 * But, this can also be called on packets in the established flow when 2785 * the fast version below fails. 2786 */ 2787void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx, int estab) 2788{ 2789 unsigned char *ptr; 2790 struct tcphdr *th = skb->h.th; 2791 int length=(th->doff*4)-sizeof(struct tcphdr); 2792 2793 ptr = (unsigned char *)(th + 1); 2794 opt_rx->saw_tstamp = 0; 2795 2796 while(length>0) { 2797 int opcode=*ptr++; 2798 int opsize; 2799 2800 switch (opcode) { 2801 case TCPOPT_EOL: 2802 return; 2803 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */ 2804 length--; 2805 continue; 2806 default: 2807 opsize=*ptr++; 2808 if (opsize < 2) /* "silly options" */ 2809 return; 2810 if (opsize > length) 2811 return; /* don't parse partial options */ 2812 switch(opcode) { 2813 case TCPOPT_MSS: 2814 if(opsize==TCPOLEN_MSS && th->syn && !estab) { 2815 u16 in_mss = ntohs(get_unaligned((__be16 *)ptr)); 2816 if (in_mss) { 2817 if (opt_rx->user_mss && opt_rx->user_mss < in_mss) 2818 in_mss = opt_rx->user_mss; 2819 opt_rx->mss_clamp = in_mss; 2820 } 2821 } 2822 break; 2823 case TCPOPT_WINDOW: 2824 if(opsize==TCPOLEN_WINDOW && th->syn && !estab) 2825 if (sysctl_tcp_window_scaling) { 2826 __u8 snd_wscale = *(__u8 *) ptr; 2827 opt_rx->wscale_ok = 1; 2828 if (snd_wscale > 14) { 2829 if(net_ratelimit()) 2830 printk(KERN_INFO "tcp_parse_options: Illegal window " 2831 "scaling value %d >14 received.\n", 2832 snd_wscale); 2833 snd_wscale = 14; 2834 } 2835 opt_rx->snd_wscale = snd_wscale; 2836 } 2837 break; 2838 case TCPOPT_TIMESTAMP: 2839 if(opsize==TCPOLEN_TIMESTAMP) { 2840 if ((estab && opt_rx->tstamp_ok) || 2841 (!estab && sysctl_tcp_timestamps)) { 2842 opt_rx->saw_tstamp = 1; 2843 opt_rx->rcv_tsval = ntohl(get_unaligned((__be32 *)ptr)); 2844 opt_rx->rcv_tsecr = ntohl(get_unaligned((__be32 *)(ptr+4))); 2845 } 2846 } 2847 break; 2848 case TCPOPT_SACK_PERM: 2849 if(opsize==TCPOLEN_SACK_PERM && th->syn && !estab) { 2850 if (sysctl_tcp_sack) { 2851 opt_rx->sack_ok = 1; 2852 tcp_sack_reset(opt_rx); 2853 } 2854 } 2855 break; 2856 2857 case TCPOPT_SACK: 2858 if((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) && 2859 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) && 2860 opt_rx->sack_ok) { 2861 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th; 2862 } 2863#ifdef CONFIG_TCP_MD5SIG 2864 case TCPOPT_MD5SIG: 2865 /* 2866 * The MD5 Hash has already been 2867 * checked (see tcp_v{4,6}_do_rcv()). 2868 */ 2869 break; 2870#endif 2871 }; 2872 ptr+=opsize-2; 2873 length-=opsize; 2874 }; 2875 } 2876} 2877 2878/* Fast parse options. This hopes to only see timestamps. 2879 * If it is wrong it falls back on tcp_parse_options(). 2880 */ 2881static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th, 2882 struct tcp_sock *tp) 2883{ 2884 if (th->doff == sizeof(struct tcphdr)>>2) { 2885 tp->rx_opt.saw_tstamp = 0; 2886 return 0; 2887 } else if (tp->rx_opt.tstamp_ok && 2888 th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) { 2889 __be32 *ptr = (__be32 *)(th + 1); 2890 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) 2891 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) { 2892 tp->rx_opt.saw_tstamp = 1; 2893 ++ptr; 2894 tp->rx_opt.rcv_tsval = ntohl(*ptr); 2895 ++ptr; 2896 tp->rx_opt.rcv_tsecr = ntohl(*ptr); 2897 return 1; 2898 } 2899 } 2900 tcp_parse_options(skb, &tp->rx_opt, 1); 2901 return 1; 2902} 2903 2904static inline void tcp_store_ts_recent(struct tcp_sock *tp) 2905{ 2906 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval; 2907 tp->rx_opt.ts_recent_stamp = xtime.tv_sec; 2908} 2909 2910static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq) 2911{ 2912 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) { 2913 /* PAWS bug workaround wrt. ACK frames, the PAWS discard 2914 * extra check below makes sure this can only happen 2915 * for pure ACK frames. -DaveM 2916 * 2917 * Not only, also it occurs for expired timestamps. 2918 */ 2919 2920 if((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) >= 0 || 2921 xtime.tv_sec >= tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS) 2922 tcp_store_ts_recent(tp); 2923 } 2924} 2925 2926/* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM 2927 * 2928 * It is not fatal. If this ACK does _not_ change critical state (seqs, window) 2929 * it can pass through stack. So, the following predicate verifies that 2930 * this segment is not used for anything but congestion avoidance or 2931 * fast retransmit. Moreover, we even are able to eliminate most of such 2932 * second order effects, if we apply some small "replay" window (~RTO) 2933 * to timestamp space. 2934 * 2935 * All these measures still do not guarantee that we reject wrapped ACKs 2936 * on networks with high bandwidth, when sequence space is recycled fastly, 2937 * but it guarantees that such events will be very rare and do not affect 2938 * connection seriously. This doesn't look nice, but alas, PAWS is really 2939 * buggy extension. 2940 * 2941 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC 2942 * states that events when retransmit arrives after original data are rare. 2943 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is 2944 * the biggest problem on large power networks even with minor reordering. 2945 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe 2946 * up to bandwidth of 18Gigabit/sec. 8) ] 2947 */ 2948 2949static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb) 2950{ 2951 struct tcp_sock *tp = tcp_sk(sk); 2952 struct tcphdr *th = skb->h.th; 2953 u32 seq = TCP_SKB_CB(skb)->seq; 2954 u32 ack = TCP_SKB_CB(skb)->ack_seq; 2955 2956 return (/* 1. Pure ACK with correct sequence number. */ 2957 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) && 2958 2959 /* 2. ... and duplicate ACK. */ 2960 ack == tp->snd_una && 2961 2962 /* 3. ... and does not update window. */ 2963 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) && 2964 2965 /* 4. ... and sits in replay window. */ 2966 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ); 2967} 2968 2969static inline int tcp_paws_discard(const struct sock *sk, const struct sk_buff *skb) 2970{ 2971 const struct tcp_sock *tp = tcp_sk(sk); 2972 return ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > TCP_PAWS_WINDOW && 2973 xtime.tv_sec < tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS && 2974 !tcp_disordered_ack(sk, skb)); 2975} 2976 2977/* Check segment sequence number for validity. 2978 * 2979 * Segment controls are considered valid, if the segment 2980 * fits to the window after truncation to the window. Acceptability 2981 * of data (and SYN, FIN, of course) is checked separately. 2982 * See tcp_data_queue(), for example. 2983 * 2984 * Also, controls (RST is main one) are accepted using RCV.WUP instead 2985 * of RCV.NXT. Peer still did not advance his SND.UNA when we 2986 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP. 2987 * (borrowed from freebsd) 2988 */ 2989 2990static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq) 2991{ 2992 return !before(end_seq, tp->rcv_wup) && 2993 !after(seq, tp->rcv_nxt + tcp_receive_window(tp)); 2994} 2995 2996/* When we get a reset we do this. */ 2997static void tcp_reset(struct sock *sk) 2998{ 2999 /* We want the right error as BSD sees it (and indeed as we do). */ 3000 switch (sk->sk_state) { 3001 case TCP_SYN_SENT: 3002 sk->sk_err = ECONNREFUSED; 3003 break; 3004 case TCP_CLOSE_WAIT: 3005 sk->sk_err = EPIPE; 3006 break; 3007 case TCP_CLOSE: 3008 return; 3009 default: 3010 sk->sk_err = ECONNRESET; 3011 } 3012 3013 if (!sock_flag(sk, SOCK_DEAD)) 3014 sk->sk_error_report(sk); 3015 3016 tcp_done(sk); 3017} 3018 3019/* 3020 * Process the FIN bit. This now behaves as it is supposed to work 3021 * and the FIN takes effect when it is validly part of sequence 3022 * space. Not before when we get holes. 3023 * 3024 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT 3025 * (and thence onto LAST-ACK and finally, CLOSE, we never enter 3026 * TIME-WAIT) 3027 * 3028 * If we are in FINWAIT-1, a received FIN indicates simultaneous 3029 * close and we go into CLOSING (and later onto TIME-WAIT) 3030 * 3031 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT. 3032 */ 3033static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th) 3034{ 3035 struct tcp_sock *tp = tcp_sk(sk); 3036 3037 inet_csk_schedule_ack(sk); 3038 3039 sk->sk_shutdown |= RCV_SHUTDOWN; 3040 sock_set_flag(sk, SOCK_DONE); 3041 3042 switch (sk->sk_state) { 3043 case TCP_SYN_RECV: 3044 case TCP_ESTABLISHED: 3045 /* Move to CLOSE_WAIT */ 3046 tcp_set_state(sk, TCP_CLOSE_WAIT); 3047 inet_csk(sk)->icsk_ack.pingpong = 1; 3048 break; 3049 3050 case TCP_CLOSE_WAIT: 3051 case TCP_CLOSING: 3052 /* Received a retransmission of the FIN, do 3053 * nothing. 3054 */ 3055 break; 3056 case TCP_LAST_ACK: 3057 /* RFC793: Remain in the LAST-ACK state. */ 3058 break; 3059 3060 case TCP_FIN_WAIT1: 3061 /* This case occurs when a simultaneous close 3062 * happens, we must ack the received FIN and 3063 * enter the CLOSING state. 3064 */ 3065 tcp_send_ack(sk); 3066 tcp_set_state(sk, TCP_CLOSING); 3067 break; 3068 case TCP_FIN_WAIT2: 3069 /* Received a FIN -- send ACK and enter TIME_WAIT. */ 3070 tcp_send_ack(sk); 3071 tcp_time_wait(sk, TCP_TIME_WAIT, 0); 3072 break; 3073 default: 3074 /* Only TCP_LISTEN and TCP_CLOSE are left, in these 3075 * cases we should never reach this piece of code. 3076 */ 3077 printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n", 3078 __FUNCTION__, sk->sk_state); 3079 break; 3080 }; 3081 3082 /* It _is_ possible, that we have something out-of-order _after_ FIN. 3083 * Probably, we should reset in this case. For now drop them. 3084 */ 3085 __skb_queue_purge(&tp->out_of_order_queue); 3086 if (tp->rx_opt.sack_ok) 3087 tcp_sack_reset(&tp->rx_opt); 3088 sk_stream_mem_reclaim(sk); 3089 3090 if (!sock_flag(sk, SOCK_DEAD)) { 3091 sk->sk_state_change(sk); 3092 3093 /* Do not send POLL_HUP for half duplex close. */ 3094 if (sk->sk_shutdown == SHUTDOWN_MASK || 3095 sk->sk_state == TCP_CLOSE) 3096 sk_wake_async(sk, 1, POLL_HUP); 3097 else 3098 sk_wake_async(sk, 1, POLL_IN); 3099 } 3100} 3101 3102static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, u32 end_seq) 3103{ 3104 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) { 3105 if (before(seq, sp->start_seq)) 3106 sp->start_seq = seq; 3107 if (after(end_seq, sp->end_seq)) 3108 sp->end_seq = end_seq; 3109 return 1; 3110 } 3111 return 0; 3112} 3113 3114static void tcp_dsack_set(struct tcp_sock *tp, u32 seq, u32 end_seq) 3115{ 3116 if (tp->rx_opt.sack_ok && sysctl_tcp_dsack) { 3117 if (before(seq, tp->rcv_nxt)) 3118 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT); 3119 else 3120 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT); 3121 3122 tp->rx_opt.dsack = 1; 3123 tp->duplicate_sack[0].start_seq = seq; 3124 tp->duplicate_sack[0].end_seq = end_seq; 3125 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + 1, 4 - tp->rx_opt.tstamp_ok); 3126 } 3127} 3128 3129static void tcp_dsack_extend(struct tcp_sock *tp, u32 seq, u32 end_seq) 3130{ 3131 if (!tp->rx_opt.dsack) 3132 tcp_dsack_set(tp, seq, end_seq); 3133 else 3134 tcp_sack_extend(tp->duplicate_sack, seq, end_seq); 3135} 3136 3137static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb) 3138{ 3139 struct tcp_sock *tp = tcp_sk(sk); 3140 3141 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 3142 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 3143 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST); 3144 tcp_enter_quickack_mode(sk); 3145 3146 if (tp->rx_opt.sack_ok && sysctl_tcp_dsack) { 3147 u32 end_seq = TCP_SKB_CB(skb)->end_seq; 3148 3149 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) 3150 end_seq = tp->rcv_nxt; 3151 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, end_seq); 3152 } 3153 } 3154 3155 tcp_send_ack(sk); 3156} 3157 3158/* These routines update the SACK block as out-of-order packets arrive or 3159 * in-order packets close up the sequence space. 3160 */ 3161static void tcp_sack_maybe_coalesce(struct tcp_sock *tp) 3162{ 3163 int this_sack; 3164 struct tcp_sack_block *sp = &tp->selective_acks[0]; 3165 struct tcp_sack_block *swalk = sp+1; 3166 3167 /* See if the recent change to the first SACK eats into 3168 * or hits the sequence space of other SACK blocks, if so coalesce. 3169 */ 3170 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks; ) { 3171 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) { 3172 int i; 3173 3174 /* Zap SWALK, by moving every further SACK up by one slot. 3175 * Decrease num_sacks. 3176 */ 3177 tp->rx_opt.num_sacks--; 3178 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok); 3179 for(i=this_sack; i < tp->rx_opt.num_sacks; i++) 3180 sp[i] = sp[i+1]; 3181 continue; 3182 } 3183 this_sack++, swalk++; 3184 } 3185} 3186 3187static inline void tcp_sack_swap(struct tcp_sack_block *sack1, struct tcp_sack_block *sack2) 3188{ 3189 __u32 tmp; 3190 3191 tmp = sack1->start_seq; 3192 sack1->start_seq = sack2->start_seq; 3193 sack2->start_seq = tmp; 3194 3195 tmp = sack1->end_seq; 3196 sack1->end_seq = sack2->end_seq; 3197 sack2->end_seq = tmp; 3198} 3199 3200static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq) 3201{ 3202 struct tcp_sock *tp = tcp_sk(sk); 3203 struct tcp_sack_block *sp = &tp->selective_acks[0]; 3204 int cur_sacks = tp->rx_opt.num_sacks; 3205 int this_sack; 3206 3207 if (!cur_sacks) 3208 goto new_sack; 3209 3210 for (this_sack=0; this_sack<cur_sacks; this_sack++, sp++) { 3211 if (tcp_sack_extend(sp, seq, end_seq)) { 3212 /* Rotate this_sack to the first one. */ 3213 for (; this_sack>0; this_sack--, sp--) 3214 tcp_sack_swap(sp, sp-1); 3215 if (cur_sacks > 1) 3216 tcp_sack_maybe_coalesce(tp); 3217 return; 3218 } 3219 } 3220 3221 /* Could not find an adjacent existing SACK, build a new one, 3222 * put it at the front, and shift everyone else down. We 3223 * always know there is at least one SACK present already here. 3224 * 3225 * If the sack array is full, forget about the last one. 3226 */ 3227 if (this_sack >= 4) { 3228 this_sack--; 3229 tp->rx_opt.num_sacks--; 3230 sp--; 3231 } 3232 for(; this_sack > 0; this_sack--, sp--) 3233 *sp = *(sp-1); 3234 3235new_sack: 3236 /* Build the new head SACK, and we're done. */ 3237 sp->start_seq = seq; 3238 sp->end_seq = end_seq; 3239 tp->rx_opt.num_sacks++; 3240 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok); 3241} 3242 3243/* RCV.NXT advances, some SACKs should be eaten. */ 3244 3245static void tcp_sack_remove(struct tcp_sock *tp) 3246{ 3247 struct tcp_sack_block *sp = &tp->selective_acks[0]; 3248 int num_sacks = tp->rx_opt.num_sacks; 3249 int this_sack; 3250 3251 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */ 3252 if (skb_queue_empty(&tp->out_of_order_queue)) { 3253 tp->rx_opt.num_sacks = 0; 3254 tp->rx_opt.eff_sacks = tp->rx_opt.dsack; 3255 return; 3256 } 3257 3258 for(this_sack = 0; this_sack < num_sacks; ) { 3259 /* Check if the start of the sack is covered by RCV.NXT. */ 3260 if (!before(tp->rcv_nxt, sp->start_seq)) { 3261 int i; 3262 3263 /* RCV.NXT must cover all the block! */ 3264 BUG_TRAP(!before(tp->rcv_nxt, sp->end_seq)); 3265 3266 /* Zap this SACK, by moving forward any other SACKS. */ 3267 for (i=this_sack+1; i < num_sacks; i++) 3268 tp->selective_acks[i-1] = tp->selective_acks[i]; 3269 num_sacks--; 3270 continue; 3271 } 3272 this_sack++; 3273 sp++; 3274 } 3275 if (num_sacks != tp->rx_opt.num_sacks) { 3276 tp->rx_opt.num_sacks = num_sacks; 3277 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok); 3278 } 3279} 3280 3281/* This one checks to see if we can put data from the 3282 * out_of_order queue into the receive_queue. 3283 */ 3284static void tcp_ofo_queue(struct sock *sk) 3285{ 3286 struct tcp_sock *tp = tcp_sk(sk); 3287 __u32 dsack_high = tp->rcv_nxt; 3288 struct sk_buff *skb; 3289 3290 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) { 3291 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) 3292 break; 3293 3294 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) { 3295 __u32 dsack = dsack_high; 3296 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high)) 3297 dsack_high = TCP_SKB_CB(skb)->end_seq; 3298 tcp_dsack_extend(tp, TCP_SKB_CB(skb)->seq, dsack); 3299 } 3300 3301 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) { 3302 SOCK_DEBUG(sk, "ofo packet was already received \n"); 3303 __skb_unlink(skb, &tp->out_of_order_queue); 3304 __kfree_skb(skb); 3305 continue; 3306 } 3307 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n", 3308 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, 3309 TCP_SKB_CB(skb)->end_seq); 3310 3311 __skb_unlink(skb, &tp->out_of_order_queue); 3312 __skb_queue_tail(&sk->sk_receive_queue, skb); 3313 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; 3314 if(skb->h.th->fin) 3315 tcp_fin(skb, sk, skb->h.th); 3316 } 3317} 3318 3319static int tcp_prune_queue(struct sock *sk); 3320 3321static void tcp_data_queue(struct sock *sk, struct sk_buff *skb) 3322{ 3323 struct tcphdr *th = skb->h.th; 3324 struct tcp_sock *tp = tcp_sk(sk); 3325 int eaten = -1; 3326 3327 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) 3328 goto drop; 3329 3330 __skb_pull(skb, th->doff*4); 3331 3332 TCP_ECN_accept_cwr(tp, skb); 3333 3334 if (tp->rx_opt.dsack) { 3335 tp->rx_opt.dsack = 0; 3336 tp->rx_opt.eff_sacks = min_t(unsigned int, tp->rx_opt.num_sacks, 3337 4 - tp->rx_opt.tstamp_ok); 3338 } 3339 3340 /* Queue data for delivery to the user. 3341 * Packets in sequence go to the receive queue. 3342 * Out of sequence packets to the out_of_order_queue. 3343 */ 3344 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) { 3345 if (tcp_receive_window(tp) == 0) 3346 goto out_of_window; 3347 3348 /* Ok. In sequence. In window. */ 3349 if (tp->ucopy.task == current && 3350 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len && 3351 sock_owned_by_user(sk) && !tp->urg_data) { 3352 int chunk = min_t(unsigned int, skb->len, 3353 tp->ucopy.len); 3354 3355 __set_current_state(TASK_RUNNING); 3356 3357 local_bh_enable(); 3358 if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) { 3359 tp->ucopy.len -= chunk; 3360 tp->copied_seq += chunk; 3361 eaten = (chunk == skb->len && !th->fin); 3362 tcp_rcv_space_adjust(sk); 3363 } 3364 local_bh_disable(); 3365 } 3366 3367 if (eaten <= 0) { 3368queue_and_out: 3369 if (eaten < 0 && 3370 (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf || 3371 !sk_stream_rmem_schedule(sk, skb))) { 3372 if (tcp_prune_queue(sk) < 0 || 3373 !sk_stream_rmem_schedule(sk, skb)) 3374 goto drop; 3375 } 3376 sk_stream_set_owner_r(skb, sk); 3377 __skb_queue_tail(&sk->sk_receive_queue, skb); 3378 } 3379 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; 3380 if(skb->len) 3381 tcp_event_data_recv(sk, tp, skb); 3382 if(th->fin) 3383 tcp_fin(skb, sk, th); 3384 3385 if (!skb_queue_empty(&tp->out_of_order_queue)) { 3386 tcp_ofo_queue(sk); 3387 3388 /* RFC2581. 4.2. SHOULD send immediate ACK, when 3389 * gap in queue is filled. 3390 */ 3391 if (skb_queue_empty(&tp->out_of_order_queue)) 3392 inet_csk(sk)->icsk_ack.pingpong = 0; 3393 } 3394 3395 if (tp->rx_opt.num_sacks) 3396 tcp_sack_remove(tp); 3397 3398 tcp_fast_path_check(sk, tp); 3399 3400 if (eaten > 0) 3401 __kfree_skb(skb); 3402 else if (!sock_flag(sk, SOCK_DEAD)) 3403 sk->sk_data_ready(sk, 0); 3404 return; 3405 } 3406 3407 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) { 3408 /* A retransmit, 2nd most common case. Force an immediate ack. */ 3409 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST); 3410 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); 3411 3412out_of_window: 3413 tcp_enter_quickack_mode(sk); 3414 inet_csk_schedule_ack(sk); 3415drop: 3416 __kfree_skb(skb); 3417 return; 3418 } 3419 3420 /* Out of window. F.e. zero window probe. */ 3421 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp))) 3422 goto out_of_window; 3423 3424 tcp_enter_quickack_mode(sk); 3425 3426 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 3427 /* Partial packet, seq < rcv_next < end_seq */ 3428 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n", 3429 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, 3430 TCP_SKB_CB(skb)->end_seq); 3431 3432 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, tp->rcv_nxt); 3433 3434 /* If window is closed, drop tail of packet. But after 3435 * remembering D-SACK for its head made in previous line. 3436 */ 3437 if (!tcp_receive_window(tp)) 3438 goto out_of_window; 3439 goto queue_and_out; 3440 } 3441 3442 TCP_ECN_check_ce(tp, skb); 3443 3444 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf || 3445 !sk_stream_rmem_schedule(sk, skb)) { 3446 if (tcp_prune_queue(sk) < 0 || 3447 !sk_stream_rmem_schedule(sk, skb)) 3448 goto drop; 3449 } 3450 3451 /* Disable header prediction. */ 3452 tp->pred_flags = 0; 3453 inet_csk_schedule_ack(sk); 3454 3455 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n", 3456 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); 3457 3458 sk_stream_set_owner_r(skb, sk); 3459 3460 if (!skb_peek(&tp->out_of_order_queue)) { 3461 /* Initial out of order segment, build 1 SACK. */ 3462 if (tp->rx_opt.sack_ok) { 3463 tp->rx_opt.num_sacks = 1; 3464 tp->rx_opt.dsack = 0; 3465 tp->rx_opt.eff_sacks = 1; 3466 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq; 3467 tp->selective_acks[0].end_seq = 3468 TCP_SKB_CB(skb)->end_seq; 3469 } 3470 __skb_queue_head(&tp->out_of_order_queue,skb); 3471 } else { 3472 struct sk_buff *skb1 = tp->out_of_order_queue.prev; 3473 u32 seq = TCP_SKB_CB(skb)->seq; 3474 u32 end_seq = TCP_SKB_CB(skb)->end_seq; 3475 3476 if (seq == TCP_SKB_CB(skb1)->end_seq) { 3477 __skb_append(skb1, skb, &tp->out_of_order_queue); 3478 3479 if (!tp->rx_opt.num_sacks || 3480 tp->selective_acks[0].end_seq != seq) 3481 goto add_sack; 3482 3483 /* Common case: data arrive in order after hole. */ 3484 tp->selective_acks[0].end_seq = end_seq; 3485 return; 3486 } 3487 3488 /* Find place to insert this segment. */ 3489 do { 3490 if (!after(TCP_SKB_CB(skb1)->seq, seq)) 3491 break; 3492 } while ((skb1 = skb1->prev) != 3493 (struct sk_buff*)&tp->out_of_order_queue); 3494 3495 /* Do skb overlap to previous one? */ 3496 if (skb1 != (struct sk_buff*)&tp->out_of_order_queue && 3497 before(seq, TCP_SKB_CB(skb1)->end_seq)) { 3498 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) { 3499 /* All the bits are present. Drop. */ 3500 __kfree_skb(skb); 3501 tcp_dsack_set(tp, seq, end_seq); 3502 goto add_sack; 3503 } 3504 if (after(seq, TCP_SKB_CB(skb1)->seq)) { 3505 /* Partial overlap. */ 3506 tcp_dsack_set(tp, seq, TCP_SKB_CB(skb1)->end_seq); 3507 } else { 3508 skb1 = skb1->prev; 3509 } 3510 } 3511 __skb_insert(skb, skb1, skb1->next, &tp->out_of_order_queue); 3512 3513 /* And clean segments covered by new one as whole. */ 3514 while ((skb1 = skb->next) != 3515 (struct sk_buff*)&tp->out_of_order_queue && 3516 after(end_seq, TCP_SKB_CB(skb1)->seq)) { 3517 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) { 3518 tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, end_seq); 3519 break; 3520 } 3521 __skb_unlink(skb1, &tp->out_of_order_queue); 3522 tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, TCP_SKB_CB(skb1)->end_seq); 3523 __kfree_skb(skb1); 3524 } 3525 3526add_sack: 3527 if (tp->rx_opt.sack_ok) 3528 tcp_sack_new_ofo_skb(sk, seq, end_seq); 3529 } 3530} 3531 3532/* Collapse contiguous sequence of skbs head..tail with 3533 * sequence numbers start..end. 3534 * Segments with FIN/SYN are not collapsed (only because this 3535 * simplifies code) 3536 */ 3537static void 3538tcp_collapse(struct sock *sk, struct sk_buff_head *list, 3539 struct sk_buff *head, struct sk_buff *tail, 3540 u32 start, u32 end) 3541{ 3542 struct sk_buff *skb; 3543 3544 /* First, check that queue is collapsible and find 3545 * the point where collapsing can be useful. */ 3546 for (skb = head; skb != tail; ) { 3547 /* No new bits? It is possible on ofo queue. */ 3548 if (!before(start, TCP_SKB_CB(skb)->end_seq)) { 3549 struct sk_buff *next = skb->next; 3550 __skb_unlink(skb, list); 3551 __kfree_skb(skb); 3552 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED); 3553 skb = next; 3554 continue; 3555 } 3556 3557 /* The first skb to collapse is: 3558 * - not SYN/FIN and 3559 * - bloated or contains data before "start" or 3560 * overlaps to the next one. 3561 */ 3562 if (!skb->h.th->syn && !skb->h.th->fin && 3563 (tcp_win_from_space(skb->truesize) > skb->len || 3564 before(TCP_SKB_CB(skb)->seq, start) || 3565 (skb->next != tail && 3566 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq))) 3567 break; 3568 3569 /* Decided to skip this, advance start seq. */ 3570 start = TCP_SKB_CB(skb)->end_seq; 3571 skb = skb->next; 3572 } 3573 if (skb == tail || skb->h.th->syn || skb->h.th->fin) 3574 return; 3575 3576 while (before(start, end)) { 3577 struct sk_buff *nskb; 3578 int header = skb_headroom(skb); 3579 int copy = SKB_MAX_ORDER(header, 0); 3580 3581 /* Too big header? This can happen with IPv6. */ 3582 if (copy < 0) 3583 return; 3584 if (end-start < copy) 3585 copy = end-start; 3586 nskb = alloc_skb(copy+header, GFP_ATOMIC); 3587 if (!nskb) 3588 return; 3589 skb_reserve(nskb, header); 3590 memcpy(nskb->head, skb->head, header); 3591 nskb->nh.raw = nskb->head + (skb->nh.raw-skb->head); 3592 nskb->h.raw = nskb->head + (skb->h.raw-skb->head); 3593 nskb->mac.raw = nskb->head + (skb->mac.raw-skb->head); 3594 memcpy(nskb->cb, skb->cb, sizeof(skb->cb)); 3595 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start; 3596 __skb_insert(nskb, skb->prev, skb, list); 3597 sk_stream_set_owner_r(nskb, sk); 3598 3599 /* Copy data, releasing collapsed skbs. */ 3600 while (copy > 0) { 3601 int offset = start - TCP_SKB_CB(skb)->seq; 3602 int size = TCP_SKB_CB(skb)->end_seq - start; 3603 3604 BUG_ON(offset < 0); 3605 if (size > 0) { 3606 size = min(copy, size); 3607 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size)) 3608 BUG(); 3609 TCP_SKB_CB(nskb)->end_seq += size; 3610 copy -= size; 3611 start += size; 3612 } 3613 if (!before(start, TCP_SKB_CB(skb)->end_seq)) { 3614 struct sk_buff *next = skb->next; 3615 __skb_unlink(skb, list); 3616 __kfree_skb(skb); 3617 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED); 3618 skb = next; 3619 if (skb == tail || skb->h.th->syn || skb->h.th->fin) 3620 return; 3621 } 3622 } 3623 } 3624} 3625 3626/* Collapse ofo queue. Algorithm: select contiguous sequence of skbs 3627 * and tcp_collapse() them until all the queue is collapsed. 3628 */ 3629static void tcp_collapse_ofo_queue(struct sock *sk) 3630{ 3631 struct tcp_sock *tp = tcp_sk(sk); 3632 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue); 3633 struct sk_buff *head; 3634 u32 start, end; 3635 3636 if (skb == NULL) 3637 return; 3638 3639 start = TCP_SKB_CB(skb)->seq; 3640 end = TCP_SKB_CB(skb)->end_seq; 3641 head = skb; 3642 3643 for (;;) { 3644 skb = skb->next; 3645 3646 /* Segment is terminated when we see gap or when 3647 * we are at the end of all the queue. */ 3648 if (skb == (struct sk_buff *)&tp->out_of_order_queue || 3649 after(TCP_SKB_CB(skb)->seq, end) || 3650 before(TCP_SKB_CB(skb)->end_seq, start)) { 3651 tcp_collapse(sk, &tp->out_of_order_queue, 3652 head, skb, start, end); 3653 head = skb; 3654 if (skb == (struct sk_buff *)&tp->out_of_order_queue) 3655 break; 3656 /* Start new segment */ 3657 start = TCP_SKB_CB(skb)->seq; 3658 end = TCP_SKB_CB(skb)->end_seq; 3659 } else { 3660 if (before(TCP_SKB_CB(skb)->seq, start)) 3661 start = TCP_SKB_CB(skb)->seq; 3662 if (after(TCP_SKB_CB(skb)->end_seq, end)) 3663 end = TCP_SKB_CB(skb)->end_seq; 3664 } 3665 } 3666} 3667 3668/* Reduce allocated memory if we can, trying to get 3669 * the socket within its memory limits again. 3670 * 3671 * Return less than zero if we should start dropping frames 3672 * until the socket owning process reads some of the data 3673 * to stabilize the situation. 3674 */ 3675static int tcp_prune_queue(struct sock *sk) 3676{ 3677 struct tcp_sock *tp = tcp_sk(sk); 3678 3679 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq); 3680 3681 NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED); 3682 3683 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf) 3684 tcp_clamp_window(sk, tp); 3685 else if (tcp_memory_pressure) 3686 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss); 3687 3688 tcp_collapse_ofo_queue(sk); 3689 tcp_collapse(sk, &sk->sk_receive_queue, 3690 sk->sk_receive_queue.next, 3691 (struct sk_buff*)&sk->sk_receive_queue, 3692 tp->copied_seq, tp->rcv_nxt); 3693 sk_stream_mem_reclaim(sk); 3694 3695 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) 3696 return 0; 3697 3698 /* Collapsing did not help, destructive actions follow. 3699 * This must not ever occur. */ 3700 3701 /* First, purge the out_of_order queue. */ 3702 if (!skb_queue_empty(&tp->out_of_order_queue)) { 3703 NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED); 3704 __skb_queue_purge(&tp->out_of_order_queue); 3705 3706 /* Reset SACK state. A conforming SACK implementation will 3707 * do the same at a timeout based retransmit. When a connection 3708 * is in a sad state like this, we care only about integrity 3709 * of the connection not performance. 3710 */ 3711 if (tp->rx_opt.sack_ok) 3712 tcp_sack_reset(&tp->rx_opt); 3713 sk_stream_mem_reclaim(sk); 3714 } 3715 3716 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) 3717 return 0; 3718 3719 /* If we are really being abused, tell the caller to silently 3720 * drop receive data on the floor. It will get retransmitted 3721 * and hopefully then we'll have sufficient space. 3722 */ 3723 NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED); 3724 3725 /* Massive buffer overcommit. */ 3726 tp->pred_flags = 0; 3727 return -1; 3728} 3729 3730 3731/* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. 3732 * As additional protections, we do not touch cwnd in retransmission phases, 3733 * and if application hit its sndbuf limit recently. 3734 */ 3735void tcp_cwnd_application_limited(struct sock *sk) 3736{ 3737 struct tcp_sock *tp = tcp_sk(sk); 3738 3739 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && 3740 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { 3741 /* Limited by application or receiver window. */ 3742 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); 3743 u32 win_used = max(tp->snd_cwnd_used, init_win); 3744 if (win_used < tp->snd_cwnd) { 3745 tp->snd_ssthresh = tcp_current_ssthresh(sk); 3746 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1; 3747 } 3748 tp->snd_cwnd_used = 0; 3749 } 3750 tp->snd_cwnd_stamp = tcp_time_stamp; 3751} 3752 3753static int tcp_should_expand_sndbuf(struct sock *sk, struct tcp_sock *tp) 3754{ 3755 /* If the user specified a specific send buffer setting, do 3756 * not modify it. 3757 */ 3758 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK) 3759 return 0; 3760 3761 /* If we are under global TCP memory pressure, do not expand. */ 3762 if (tcp_memory_pressure) 3763 return 0; 3764 3765 /* If we are under soft global TCP memory pressure, do not expand. */ 3766 if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0]) 3767 return 0; 3768 3769 /* If we filled the congestion window, do not expand. */ 3770 if (tp->packets_out >= tp->snd_cwnd) 3771 return 0; 3772 3773 return 1; 3774} 3775 3776/* When incoming ACK allowed to free some skb from write_queue, 3777 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket 3778 * on the exit from tcp input handler. 3779 * 3780 * PROBLEM: sndbuf expansion does not work well with largesend. 3781 */ 3782static void tcp_new_space(struct sock *sk) 3783{ 3784 struct tcp_sock *tp = tcp_sk(sk); 3785 3786 if (tcp_should_expand_sndbuf(sk, tp)) { 3787 int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) + 3788 MAX_TCP_HEADER + 16 + sizeof(struct sk_buff), 3789 demanded = max_t(unsigned int, tp->snd_cwnd, 3790 tp->reordering + 1); 3791 sndmem *= 2*demanded; 3792 if (sndmem > sk->sk_sndbuf) 3793 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]); 3794 tp->snd_cwnd_stamp = tcp_time_stamp; 3795 } 3796 3797 sk->sk_write_space(sk); 3798} 3799 3800static void tcp_check_space(struct sock *sk) 3801{ 3802 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) { 3803 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK); 3804 if (sk->sk_socket && 3805 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) 3806 tcp_new_space(sk); 3807 } 3808} 3809 3810static inline void tcp_data_snd_check(struct sock *sk, struct tcp_sock *tp) 3811{ 3812 tcp_push_pending_frames(sk, tp); 3813 tcp_check_space(sk); 3814} 3815 3816/* 3817 * Check if sending an ack is needed. 3818 */ 3819static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible) 3820{ 3821 struct tcp_sock *tp = tcp_sk(sk); 3822 3823 /* More than one full frame received... */ 3824 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss 3825 /* ... and right edge of window advances far enough. 3826 * (tcp_recvmsg() will send ACK otherwise). Or... 3827 */ 3828 && __tcp_select_window(sk) >= tp->rcv_wnd) || 3829 /* We ACK each frame or... */ 3830 tcp_in_quickack_mode(sk) || 3831 /* We have out of order data. */ 3832 (ofo_possible && 3833 skb_peek(&tp->out_of_order_queue))) { 3834 /* Then ack it now */ 3835 tcp_send_ack(sk); 3836 } else { 3837 /* Else, send delayed ack. */ 3838 tcp_send_delayed_ack(sk); 3839 } 3840} 3841 3842static inline void tcp_ack_snd_check(struct sock *sk) 3843{ 3844 if (!inet_csk_ack_scheduled(sk)) { 3845 /* We sent a data segment already. */ 3846 return; 3847 } 3848 __tcp_ack_snd_check(sk, 1); 3849} 3850 3851/* 3852 * This routine is only called when we have urgent data 3853 * signaled. Its the 'slow' part of tcp_urg. It could be 3854 * moved inline now as tcp_urg is only called from one 3855 * place. We handle URGent data wrong. We have to - as 3856 * BSD still doesn't use the correction from RFC961. 3857 * For 1003.1g we should support a new option TCP_STDURG to permit 3858 * either form (or just set the sysctl tcp_stdurg). 3859 */ 3860 3861static void tcp_check_urg(struct sock * sk, struct tcphdr * th) 3862{ 3863 struct tcp_sock *tp = tcp_sk(sk); 3864 u32 ptr = ntohs(th->urg_ptr); 3865 3866 if (ptr && !sysctl_tcp_stdurg) 3867 ptr--; 3868 ptr += ntohl(th->seq); 3869 3870 /* Ignore urgent data that we've already seen and read. */ 3871 if (after(tp->copied_seq, ptr)) 3872 return; 3873 3874 /* Do not replay urg ptr. 3875 * 3876 * NOTE: interesting situation not covered by specs. 3877 * Misbehaving sender may send urg ptr, pointing to segment, 3878 * which we already have in ofo queue. We are not able to fetch 3879 * such data and will stay in TCP_URG_NOTYET until will be eaten 3880 * by recvmsg(). Seems, we are not obliged to handle such wicked 3881 * situations. But it is worth to think about possibility of some 3882 * DoSes using some hypothetical application level deadlock. 3883 */ 3884 if (before(ptr, tp->rcv_nxt)) 3885 return; 3886 3887 /* Do we already have a newer (or duplicate) urgent pointer? */ 3888 if (tp->urg_data && !after(ptr, tp->urg_seq)) 3889 return; 3890 3891 /* Tell the world about our new urgent pointer. */ 3892 sk_send_sigurg(sk); 3893 3894 /* We may be adding urgent data when the last byte read was 3895 * urgent. To do this requires some care. We cannot just ignore 3896 * tp->copied_seq since we would read the last urgent byte again 3897 * as data, nor can we alter copied_seq until this data arrives 3898 * or we break the semantics of SIOCATMARK (and thus sockatmark()) 3899 * 3900 * NOTE. Double Dutch. Rendering to plain English: author of comment 3901 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB); 3902 * and expect that both A and B disappear from stream. This is _wrong_. 3903 * Though this happens in BSD with high probability, this is occasional. 3904 * Any application relying on this is buggy. Note also, that fix "works" 3905 * only in this artificial test. Insert some normal data between A and B and we will 3906 * decline of BSD again. Verdict: it is better to remove to trap 3907 * buggy users. 3908 */ 3909 if (tp->urg_seq == tp->copied_seq && tp->urg_data && 3910 !sock_flag(sk, SOCK_URGINLINE) && 3911 tp->copied_seq != tp->rcv_nxt) { 3912 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue); 3913 tp->copied_seq++; 3914 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) { 3915 __skb_unlink(skb, &sk->sk_receive_queue); 3916 __kfree_skb(skb); 3917 } 3918 } 3919 3920 tp->urg_data = TCP_URG_NOTYET; 3921 tp->urg_seq = ptr; 3922 3923 /* Disable header prediction. */ 3924 tp->pred_flags = 0; 3925} 3926 3927/* This is the 'fast' part of urgent handling. */ 3928static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th) 3929{ 3930 struct tcp_sock *tp = tcp_sk(sk); 3931 3932 /* Check if we get a new urgent pointer - normally not. */ 3933 if (th->urg) 3934 tcp_check_urg(sk,th); 3935 3936 /* Do we wait for any urgent data? - normally not... */ 3937 if (tp->urg_data == TCP_URG_NOTYET) { 3938 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) - 3939 th->syn; 3940 3941 /* Is the urgent pointer pointing into this packet? */ 3942 if (ptr < skb->len) { 3943 u8 tmp; 3944 if (skb_copy_bits(skb, ptr, &tmp, 1)) 3945 BUG(); 3946 tp->urg_data = TCP_URG_VALID | tmp; 3947 if (!sock_flag(sk, SOCK_DEAD)) 3948 sk->sk_data_ready(sk, 0); 3949 } 3950 } 3951} 3952 3953static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen) 3954{ 3955 struct tcp_sock *tp = tcp_sk(sk); 3956 int chunk = skb->len - hlen; 3957 int err; 3958 3959 local_bh_enable(); 3960 if (skb->ip_summed==CHECKSUM_UNNECESSARY) 3961 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk); 3962 else 3963 err = skb_copy_and_csum_datagram_iovec(skb, hlen, 3964 tp->ucopy.iov); 3965 3966 if (!err) { 3967 tp->ucopy.len -= chunk; 3968 tp->copied_seq += chunk; 3969 tcp_rcv_space_adjust(sk); 3970 } 3971 3972 local_bh_disable(); 3973 return err; 3974} 3975 3976static __sum16 __tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb) 3977{ 3978 __sum16 result; 3979 3980 if (sock_owned_by_user(sk)) { 3981 local_bh_enable(); 3982 result = __tcp_checksum_complete(skb); 3983 local_bh_disable(); 3984 } else { 3985 result = __tcp_checksum_complete(skb); 3986 } 3987 return result; 3988} 3989 3990static inline int tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb) 3991{ 3992 return skb->ip_summed != CHECKSUM_UNNECESSARY && 3993 __tcp_checksum_complete_user(sk, skb); 3994} 3995 3996#ifdef CONFIG_NET_DMA 3997static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb, int hlen) 3998{ 3999 struct tcp_sock *tp = tcp_sk(sk); 4000 int chunk = skb->len - hlen; 4001 int dma_cookie; 4002 int copied_early = 0; 4003 4004 if (tp->ucopy.wakeup) 4005 return 0; 4006 4007 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list) 4008 tp->ucopy.dma_chan = get_softnet_dma(); 4009 4010 if (tp->ucopy.dma_chan && skb->ip_summed == CHECKSUM_UNNECESSARY) { 4011 4012 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan, 4013 skb, hlen, tp->ucopy.iov, chunk, tp->ucopy.pinned_list); 4014 4015 if (dma_cookie < 0) 4016 goto out; 4017 4018 tp->ucopy.dma_cookie = dma_cookie; 4019 copied_early = 1; 4020 4021 tp->ucopy.len -= chunk; 4022 tp->copied_seq += chunk; 4023 tcp_rcv_space_adjust(sk); 4024 4025 if ((tp->ucopy.len == 0) || 4026 (tcp_flag_word(skb->h.th) & TCP_FLAG_PSH) || 4027 (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) { 4028 tp->ucopy.wakeup = 1; 4029 sk->sk_data_ready(sk, 0); 4030 } 4031 } else if (chunk > 0) { 4032 tp->ucopy.wakeup = 1; 4033 sk->sk_data_ready(sk, 0); 4034 } 4035out: 4036 return copied_early; 4037} 4038#endif /* CONFIG_NET_DMA */ 4039 4040/* 4041 * TCP receive function for the ESTABLISHED state. 4042 * 4043 * It is split into a fast path and a slow path. The fast path is 4044 * disabled when: 4045 * - A zero window was announced from us - zero window probing 4046 * is only handled properly in the slow path. 4047 * - Out of order segments arrived. 4048 * - Urgent data is expected. 4049 * - There is no buffer space left 4050 * - Unexpected TCP flags/window values/header lengths are received 4051 * (detected by checking the TCP header against pred_flags) 4052 * - Data is sent in both directions. Fast path only supports pure senders 4053 * or pure receivers (this means either the sequence number or the ack 4054 * value must stay constant) 4055 * - Unexpected TCP option. 4056 * 4057 * When these conditions are not satisfied it drops into a standard 4058 * receive procedure patterned after RFC793 to handle all cases. 4059 * The first three cases are guaranteed by proper pred_flags setting, 4060 * the rest is checked inline. Fast processing is turned on in 4061 * tcp_data_queue when everything is OK. 4062 */ 4063int tcp_rcv_established(struct sock *sk, struct sk_buff *skb, 4064 struct tcphdr *th, unsigned len) 4065{ 4066 struct tcp_sock *tp = tcp_sk(sk); 4067 4068 /* 4069 * Header prediction. 4070 * The code loosely follows the one in the famous 4071 * "30 instruction TCP receive" Van Jacobson mail. 4072 * 4073 * Van's trick is to deposit buffers into socket queue 4074 * on a device interrupt, to call tcp_recv function 4075 * on the receive process context and checksum and copy 4076 * the buffer to user space. smart... 4077 * 4078 * Our current scheme is not silly either but we take the 4079 * extra cost of the net_bh soft interrupt processing... 4080 * We do checksum and copy also but from device to kernel. 4081 */ 4082 4083 tp->rx_opt.saw_tstamp = 0; 4084 4085 /* pred_flags is 0xS?10 << 16 + snd_wnd 4086 * if header_prediction is to be made 4087 * 'S' will always be tp->tcp_header_len >> 2 4088 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to 4089 * turn it off (when there are holes in the receive 4090 * space for instance) 4091 * PSH flag is ignored. 4092 */ 4093 4094 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags && 4095 TCP_SKB_CB(skb)->seq == tp->rcv_nxt) { 4096 int tcp_header_len = tp->tcp_header_len; 4097 4098 /* Timestamp header prediction: tcp_header_len 4099 * is automatically equal to th->doff*4 due to pred_flags 4100 * match. 4101 */ 4102 4103 /* Check timestamp */ 4104 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) { 4105 __be32 *ptr = (__be32 *)(th + 1); 4106 4107 /* No? Slow path! */ 4108 if (*ptr != htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) 4109 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) 4110 goto slow_path; 4111 4112 tp->rx_opt.saw_tstamp = 1; 4113 ++ptr; 4114 tp->rx_opt.rcv_tsval = ntohl(*ptr); 4115 ++ptr; 4116 tp->rx_opt.rcv_tsecr = ntohl(*ptr); 4117 4118 /* If PAWS failed, check it more carefully in slow path */ 4119 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0) 4120 goto slow_path; 4121 4122 /* DO NOT update ts_recent here, if checksum fails 4123 * and timestamp was corrupted part, it will result 4124 * in a hung connection since we will drop all 4125 * future packets due to the PAWS test. 4126 */ 4127 } 4128 4129 if (len <= tcp_header_len) { 4130 /* Bulk data transfer: sender */ 4131 if (len == tcp_header_len) { 4132 /* Predicted packet is in window by definition. 4133 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 4134 * Hence, check seq<=rcv_wup reduces to: 4135 */ 4136 if (tcp_header_len == 4137 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && 4138 tp->rcv_nxt == tp->rcv_wup) 4139 tcp_store_ts_recent(tp); 4140 4141 /* We know that such packets are checksummed 4142 * on entry. 4143 */ 4144 tcp_ack(sk, skb, 0); 4145 __kfree_skb(skb); 4146 tcp_data_snd_check(sk, tp); 4147 return 0; 4148 } else { /* Header too small */ 4149 TCP_INC_STATS_BH(TCP_MIB_INERRS); 4150 goto discard; 4151 } 4152 } else { 4153 int eaten = 0; 4154 int copied_early = 0; 4155 4156 if (tp->copied_seq == tp->rcv_nxt && 4157 len - tcp_header_len <= tp->ucopy.len) { 4158#ifdef CONFIG_NET_DMA 4159 if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) { 4160 copied_early = 1; 4161 eaten = 1; 4162 } 4163#endif 4164 if (tp->ucopy.task == current && sock_owned_by_user(sk) && !copied_early) { 4165 __set_current_state(TASK_RUNNING); 4166 4167 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len)) 4168 eaten = 1; 4169 } 4170 if (eaten) { 4171 /* Predicted packet is in window by definition. 4172 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 4173 * Hence, check seq<=rcv_wup reduces to: 4174 */ 4175 if (tcp_header_len == 4176 (sizeof(struct tcphdr) + 4177 TCPOLEN_TSTAMP_ALIGNED) && 4178 tp->rcv_nxt == tp->rcv_wup) 4179 tcp_store_ts_recent(tp); 4180 4181 tcp_rcv_rtt_measure_ts(sk, skb); 4182 4183 __skb_pull(skb, tcp_header_len); 4184 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; 4185 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER); 4186 } 4187 if (copied_early) 4188 tcp_cleanup_rbuf(sk, skb->len); 4189 } 4190 if (!eaten) { 4191 if (tcp_checksum_complete_user(sk, skb)) 4192 goto csum_error; 4193 4194 /* Predicted packet is in window by definition. 4195 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 4196 * Hence, check seq<=rcv_wup reduces to: 4197 */ 4198 if (tcp_header_len == 4199 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && 4200 tp->rcv_nxt == tp->rcv_wup) 4201 tcp_store_ts_recent(tp); 4202 4203 tcp_rcv_rtt_measure_ts(sk, skb); 4204 4205 if ((int)skb->truesize > sk->sk_forward_alloc) 4206 goto step5; 4207 4208 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS); 4209 4210 /* Bulk data transfer: receiver */ 4211 __skb_pull(skb,tcp_header_len); 4212 __skb_queue_tail(&sk->sk_receive_queue, skb); 4213 sk_stream_set_owner_r(skb, sk); 4214 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; 4215 } 4216 4217 tcp_event_data_recv(sk, tp, skb); 4218 4219 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) { 4220 /* Well, only one small jumplet in fast path... */ 4221 tcp_ack(sk, skb, FLAG_DATA); 4222 tcp_data_snd_check(sk, tp); 4223 if (!inet_csk_ack_scheduled(sk)) 4224 goto no_ack; 4225 } 4226 4227 __tcp_ack_snd_check(sk, 0); 4228no_ack: 4229#ifdef CONFIG_NET_DMA 4230 if (copied_early) 4231 __skb_queue_tail(&sk->sk_async_wait_queue, skb); 4232 else 4233#endif 4234 if (eaten) 4235 __kfree_skb(skb); 4236 else 4237 sk->sk_data_ready(sk, 0); 4238 return 0; 4239 } 4240 } 4241 4242slow_path: 4243 if (len < (th->doff<<2) || tcp_checksum_complete_user(sk, skb)) 4244 goto csum_error; 4245 4246 /* 4247 * RFC1323: H1. Apply PAWS check first. 4248 */ 4249 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp && 4250 tcp_paws_discard(sk, skb)) { 4251 if (!th->rst) { 4252 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED); 4253 tcp_send_dupack(sk, skb); 4254 goto discard; 4255 } 4256 /* Resets are accepted even if PAWS failed. 4257 4258 ts_recent update must be made after we are sure 4259 that the packet is in window. 4260 */ 4261 } 4262 4263 /* 4264 * Standard slow path. 4265 */ 4266 4267 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) { 4268 /* RFC793, page 37: "In all states except SYN-SENT, all reset 4269 * (RST) segments are validated by checking their SEQ-fields." 4270 * And page 69: "If an incoming segment is not acceptable, 4271 * an acknowledgment should be sent in reply (unless the RST bit 4272 * is set, if so drop the segment and return)". 4273 */ 4274 if (!th->rst) 4275 tcp_send_dupack(sk, skb); 4276 goto discard; 4277 } 4278 4279 if(th->rst) { 4280 tcp_reset(sk); 4281 goto discard; 4282 } 4283 4284 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq); 4285 4286 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 4287 TCP_INC_STATS_BH(TCP_MIB_INERRS); 4288 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN); 4289 tcp_reset(sk); 4290 return 1; 4291 } 4292 4293step5: 4294 if(th->ack) 4295 tcp_ack(sk, skb, FLAG_SLOWPATH); 4296 4297 tcp_rcv_rtt_measure_ts(sk, skb); 4298 4299 /* Process urgent data. */ 4300 tcp_urg(sk, skb, th); 4301 4302 /* step 7: process the segment text */ 4303 tcp_data_queue(sk, skb); 4304 4305 tcp_data_snd_check(sk, tp); 4306 tcp_ack_snd_check(sk); 4307 return 0; 4308 4309csum_error: 4310 TCP_INC_STATS_BH(TCP_MIB_INERRS); 4311 4312discard: 4313 __kfree_skb(skb); 4314 return 0; 4315} 4316 4317static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb, 4318 struct tcphdr *th, unsigned len) 4319{ 4320 struct tcp_sock *tp = tcp_sk(sk); 4321 struct inet_connection_sock *icsk = inet_csk(sk); 4322 int saved_clamp = tp->rx_opt.mss_clamp; 4323 4324 tcp_parse_options(skb, &tp->rx_opt, 0); 4325 4326 if (th->ack) { 4327 /* rfc793: 4328 * "If the state is SYN-SENT then 4329 * first check the ACK bit 4330 * If the ACK bit is set 4331 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send 4332 * a reset (unless the RST bit is set, if so drop 4333 * the segment and return)" 4334 * 4335 * We do not send data with SYN, so that RFC-correct 4336 * test reduces to: 4337 */ 4338 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt) 4339 goto reset_and_undo; 4340 4341 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 4342 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp, 4343 tcp_time_stamp)) { 4344 NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED); 4345 goto reset_and_undo; 4346 } 4347 4348 /* Now ACK is acceptable. 4349 * 4350 * "If the RST bit is set 4351 * If the ACK was acceptable then signal the user "error: 4352 * connection reset", drop the segment, enter CLOSED state, 4353 * delete TCB, and return." 4354 */ 4355 4356 if (th->rst) { 4357 tcp_reset(sk); 4358 goto discard; 4359 } 4360 4361 /* rfc793: 4362 * "fifth, if neither of the SYN or RST bits is set then 4363 * drop the segment and return." 4364 * 4365 * See note below! 4366 * --ANK(990513) 4367 */ 4368 if (!th->syn) 4369 goto discard_and_undo; 4370 4371 /* rfc793: 4372 * "If the SYN bit is on ... 4373 * are acceptable then ... 4374 * (our SYN has been ACKed), change the connection 4375 * state to ESTABLISHED..." 4376 */ 4377 4378 TCP_ECN_rcv_synack(tp, th); 4379 4380 tp->snd_wl1 = TCP_SKB_CB(skb)->seq; 4381 tcp_ack(sk, skb, FLAG_SLOWPATH); 4382 4383 /* Ok.. it's good. Set up sequence numbers and 4384 * move to established. 4385 */ 4386 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; 4387 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; 4388 4389 /* RFC1323: The window in SYN & SYN/ACK segments is 4390 * never scaled. 4391 */ 4392 tp->snd_wnd = ntohs(th->window); 4393 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(skb)->seq); 4394 4395 if (!tp->rx_opt.wscale_ok) { 4396 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0; 4397 tp->window_clamp = min(tp->window_clamp, 65535U); 4398 } 4399 4400 if (tp->rx_opt.saw_tstamp) { 4401 tp->rx_opt.tstamp_ok = 1; 4402 tp->tcp_header_len = 4403 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 4404 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; 4405 tcp_store_ts_recent(tp); 4406 } else { 4407 tp->tcp_header_len = sizeof(struct tcphdr); 4408 } 4409 4410 if (tp->rx_opt.sack_ok && sysctl_tcp_fack) 4411 tp->rx_opt.sack_ok |= 2; 4412 4413 tcp_mtup_init(sk); 4414 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 4415 tcp_initialize_rcv_mss(sk); 4416 4417 /* Remember, tcp_poll() does not lock socket! 4418 * Change state from SYN-SENT only after copied_seq 4419 * is initialized. */ 4420 tp->copied_seq = tp->rcv_nxt; 4421 smp_mb(); 4422 tcp_set_state(sk, TCP_ESTABLISHED); 4423 4424 security_inet_conn_established(sk, skb); 4425 4426 /* Make sure socket is routed, for correct metrics. */ 4427 icsk->icsk_af_ops->rebuild_header(sk); 4428 4429 tcp_init_metrics(sk); 4430 4431 tcp_init_congestion_control(sk); 4432 4433 /* Prevent spurious tcp_cwnd_restart() on first data 4434 * packet. 4435 */ 4436 tp->lsndtime = tcp_time_stamp; 4437 4438 tcp_init_buffer_space(sk); 4439 4440 if (sock_flag(sk, SOCK_KEEPOPEN)) 4441 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp)); 4442 4443 if (!tp->rx_opt.snd_wscale) 4444 __tcp_fast_path_on(tp, tp->snd_wnd); 4445 else 4446 tp->pred_flags = 0; 4447 4448 if (!sock_flag(sk, SOCK_DEAD)) { 4449 sk->sk_state_change(sk); 4450 sk_wake_async(sk, 0, POLL_OUT); 4451 } 4452 4453 if (sk->sk_write_pending || 4454 icsk->icsk_accept_queue.rskq_defer_accept || 4455 icsk->icsk_ack.pingpong) { 4456 /* Save one ACK. Data will be ready after 4457 * several ticks, if write_pending is set. 4458 * 4459 * It may be deleted, but with this feature tcpdumps 4460 * look so _wonderfully_ clever, that I was not able 4461 * to stand against the temptation 8) --ANK 4462 */ 4463 inet_csk_schedule_ack(sk); 4464 icsk->icsk_ack.lrcvtime = tcp_time_stamp; 4465 icsk->icsk_ack.ato = TCP_ATO_MIN; 4466 tcp_incr_quickack(sk); 4467 tcp_enter_quickack_mode(sk); 4468 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 4469 TCP_DELACK_MAX, TCP_RTO_MAX); 4470 4471discard: 4472 __kfree_skb(skb); 4473 return 0; 4474 } else { 4475 tcp_send_ack(sk); 4476 } 4477 return -1; 4478 } 4479 4480 /* No ACK in the segment */ 4481 4482 if (th->rst) { 4483 /* rfc793: 4484 * "If the RST bit is set 4485 * 4486 * Otherwise (no ACK) drop the segment and return." 4487 */ 4488 4489 goto discard_and_undo; 4490 } 4491 4492 /* PAWS check. */ 4493 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && tcp_paws_check(&tp->rx_opt, 0)) 4494 goto discard_and_undo; 4495 4496 if (th->syn) { 4497 /* We see SYN without ACK. It is attempt of 4498 * simultaneous connect with crossed SYNs. 4499 * Particularly, it can be connect to self. 4500 */ 4501 tcp_set_state(sk, TCP_SYN_RECV); 4502 4503 if (tp->rx_opt.saw_tstamp) { 4504 tp->rx_opt.tstamp_ok = 1; 4505 tcp_store_ts_recent(tp); 4506 tp->tcp_header_len = 4507 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 4508 } else { 4509 tp->tcp_header_len = sizeof(struct tcphdr); 4510 } 4511 4512 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; 4513 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; 4514 4515 /* RFC1323: The window in SYN & SYN/ACK segments is 4516 * never scaled. 4517 */ 4518 tp->snd_wnd = ntohs(th->window); 4519 tp->snd_wl1 = TCP_SKB_CB(skb)->seq; 4520 tp->max_window = tp->snd_wnd; 4521 4522 TCP_ECN_rcv_syn(tp, th); 4523 4524 tcp_mtup_init(sk); 4525 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 4526 tcp_initialize_rcv_mss(sk); 4527 4528 4529 tcp_send_synack(sk); 4530#if 0 4531 /* Note, we could accept data and URG from this segment. 4532 * There are no obstacles to make this. 4533 * 4534 * However, if we ignore data in ACKless segments sometimes, 4535 * we have no reasons to accept it sometimes. 4536 * Also, seems the code doing it in step6 of tcp_rcv_state_process 4537 * is not flawless. So, discard packet for sanity. 4538 * Uncomment this return to process the data. 4539 */ 4540 return -1; 4541#else 4542 goto discard; 4543#endif 4544 } 4545 /* "fifth, if neither of the SYN or RST bits is set then 4546 * drop the segment and return." 4547 */ 4548 4549discard_and_undo: 4550 tcp_clear_options(&tp->rx_opt); 4551 tp->rx_opt.mss_clamp = saved_clamp; 4552 goto discard; 4553 4554reset_and_undo: 4555 tcp_clear_options(&tp->rx_opt); 4556 tp->rx_opt.mss_clamp = saved_clamp; 4557 return 1; 4558} 4559 4560 4561/* 4562 * This function implements the receiving procedure of RFC 793 for 4563 * all states except ESTABLISHED and TIME_WAIT. 4564 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be 4565 * address independent. 4566 */ 4567 4568int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb, 4569 struct tcphdr *th, unsigned len) 4570{ 4571 struct tcp_sock *tp = tcp_sk(sk); 4572 struct inet_connection_sock *icsk = inet_csk(sk); 4573 int queued = 0; 4574 4575 tp->rx_opt.saw_tstamp = 0; 4576 4577 switch (sk->sk_state) { 4578 case TCP_CLOSE: 4579 goto discard; 4580 4581 case TCP_LISTEN: 4582 if(th->ack) 4583 return 1; 4584 4585 if(th->rst) 4586 goto discard; 4587 4588 if(th->syn) { 4589 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0) 4590 return 1; 4591 4592 /* Now we have several options: In theory there is 4593 * nothing else in the frame. KA9Q has an option to 4594 * send data with the syn, BSD accepts data with the 4595 * syn up to the [to be] advertised window and 4596 * Solaris 2.1 gives you a protocol error. For now 4597 * we just ignore it, that fits the spec precisely 4598 * and avoids incompatibilities. It would be nice in 4599 * future to drop through and process the data. 4600 * 4601 * Now that TTCP is starting to be used we ought to 4602 * queue this data. 4603 * But, this leaves one open to an easy denial of 4604 * service attack, and SYN cookies can't defend 4605 * against this problem. So, we drop the data 4606 * in the interest of security over speed unless 4607 * it's still in use. 4608 */ 4609 kfree_skb(skb); 4610 return 0; 4611 } 4612 goto discard; 4613 4614 case TCP_SYN_SENT: 4615 queued = tcp_rcv_synsent_state_process(sk, skb, th, len); 4616 if (queued >= 0) 4617 return queued; 4618 4619 /* Do step6 onward by hand. */ 4620 tcp_urg(sk, skb, th); 4621 __kfree_skb(skb); 4622 tcp_data_snd_check(sk, tp); 4623 return 0; 4624 } 4625 4626 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp && 4627 tcp_paws_discard(sk, skb)) { 4628 if (!th->rst) { 4629 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED); 4630 tcp_send_dupack(sk, skb); 4631 goto discard; 4632 } 4633 /* Reset is accepted even if it did not pass PAWS. */ 4634 } 4635 4636 /* step 1: check sequence number */ 4637 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) { 4638 if (!th->rst) 4639 tcp_send_dupack(sk, skb); 4640 goto discard; 4641 } 4642 4643 /* step 2: check RST bit */ 4644 if(th->rst) { 4645 tcp_reset(sk); 4646 goto discard; 4647 } 4648 4649 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq); 4650 4651 /* step 3: check security and precedence [ignored] */ 4652 4653 /* step 4: 4654 * 4655 * Check for a SYN in window. 4656 */ 4657 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 4658 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN); 4659 tcp_reset(sk); 4660 return 1; 4661 } 4662 4663 /* step 5: check the ACK field */ 4664 if (th->ack) { 4665 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH); 4666 4667 switch(sk->sk_state) { 4668 case TCP_SYN_RECV: 4669 if (acceptable) { 4670 tp->copied_seq = tp->rcv_nxt; 4671 smp_mb(); 4672 tcp_set_state(sk, TCP_ESTABLISHED); 4673 sk->sk_state_change(sk); 4674 4675 /* Note, that this wakeup is only for marginal 4676 * crossed SYN case. Passively open sockets 4677 * are not waked up, because sk->sk_sleep == 4678 * NULL and sk->sk_socket == NULL. 4679 */ 4680 if (sk->sk_socket) { 4681 sk_wake_async(sk,0,POLL_OUT); 4682 } 4683 4684 tp->snd_una = TCP_SKB_CB(skb)->ack_seq; 4685 tp->snd_wnd = ntohs(th->window) << 4686 tp->rx_opt.snd_wscale; 4687 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, 4688 TCP_SKB_CB(skb)->seq); 4689 4690 /* tcp_ack considers this ACK as duplicate 4691 * and does not calculate rtt. 4692 * Fix it at least with timestamps. 4693 */ 4694 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 4695 !tp->srtt) 4696 tcp_ack_saw_tstamp(sk, 0); 4697 4698 if (tp->rx_opt.tstamp_ok) 4699 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; 4700 4701 /* Make sure socket is routed, for 4702 * correct metrics. 4703 */ 4704 icsk->icsk_af_ops->rebuild_header(sk); 4705 4706 tcp_init_metrics(sk); 4707 4708 tcp_init_congestion_control(sk); 4709 4710 /* Prevent spurious tcp_cwnd_restart() on 4711 * first data packet. 4712 */ 4713 tp->lsndtime = tcp_time_stamp; 4714 4715 tcp_mtup_init(sk); 4716 tcp_initialize_rcv_mss(sk); 4717 tcp_init_buffer_space(sk); 4718 tcp_fast_path_on(tp); 4719 } else { 4720 return 1; 4721 } 4722 break; 4723 4724 case TCP_FIN_WAIT1: 4725 if (tp->snd_una == tp->write_seq) { 4726 tcp_set_state(sk, TCP_FIN_WAIT2); 4727 sk->sk_shutdown |= SEND_SHUTDOWN; 4728 dst_confirm(sk->sk_dst_cache); 4729 4730 if (!sock_flag(sk, SOCK_DEAD)) 4731 /* Wake up lingering close() */ 4732 sk->sk_state_change(sk); 4733 else { 4734 int tmo; 4735 4736 if (tp->linger2 < 0 || 4737 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 4738 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) { 4739 tcp_done(sk); 4740 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA); 4741 return 1; 4742 } 4743 4744 tmo = tcp_fin_time(sk); 4745 if (tmo > TCP_TIMEWAIT_LEN) { 4746 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN); 4747 } else if (th->fin || sock_owned_by_user(sk)) { 4748 /* Bad case. We could lose such FIN otherwise. 4749 * It is not a big problem, but it looks confusing 4750 * and not so rare event. We still can lose it now, 4751 * if it spins in bh_lock_sock(), but it is really 4752 * marginal case. 4753 */ 4754 inet_csk_reset_keepalive_timer(sk, tmo); 4755 } else { 4756 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo); 4757 goto discard; 4758 } 4759 } 4760 } 4761 break; 4762 4763 case TCP_CLOSING: 4764 if (tp->snd_una == tp->write_seq) { 4765 tcp_time_wait(sk, TCP_TIME_WAIT, 0); 4766 goto discard; 4767 } 4768 break; 4769 4770 case TCP_LAST_ACK: 4771 if (tp->snd_una == tp->write_seq) { 4772 tcp_update_metrics(sk); 4773 tcp_done(sk); 4774 goto discard; 4775 } 4776 break; 4777 } 4778 } else 4779 goto discard; 4780 4781 /* step 6: check the URG bit */ 4782 tcp_urg(sk, skb, th); 4783 4784 /* step 7: process the segment text */ 4785 switch (sk->sk_state) { 4786 case TCP_CLOSE_WAIT: 4787 case TCP_CLOSING: 4788 case TCP_LAST_ACK: 4789 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) 4790 break; 4791 case TCP_FIN_WAIT1: 4792 case TCP_FIN_WAIT2: 4793 /* RFC 793 says to queue data in these states, 4794 * RFC 1122 says we MUST send a reset. 4795 * BSD 4.4 also does reset. 4796 */ 4797 if (sk->sk_shutdown & RCV_SHUTDOWN) { 4798 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 4799 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) { 4800 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA); 4801 tcp_reset(sk); 4802 return 1; 4803 } 4804 } 4805 /* Fall through */ 4806 case TCP_ESTABLISHED: 4807 tcp_data_queue(sk, skb); 4808 queued = 1; 4809 break; 4810 } 4811 4812 /* tcp_data could move socket to TIME-WAIT */ 4813 if (sk->sk_state != TCP_CLOSE) { 4814 tcp_data_snd_check(sk, tp); 4815 tcp_ack_snd_check(sk); 4816 } 4817 4818 if (!queued) { 4819discard: 4820 __kfree_skb(skb); 4821 } 4822 return 0; 4823} 4824 4825EXPORT_SYMBOL(sysctl_tcp_ecn); 4826EXPORT_SYMBOL(sysctl_tcp_reordering); 4827EXPORT_SYMBOL(tcp_parse_options); 4828EXPORT_SYMBOL(tcp_rcv_established); 4829EXPORT_SYMBOL(tcp_rcv_state_process); 4830EXPORT_SYMBOL(tcp_initialize_rcv_mss); 4831