tcp_input.c revision 575ee7140dabe9b9c4f66f4f867039b97e548867
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_frto_response __read_mostly; 90int sysctl_tcp_nometrics_save __read_mostly; 91 92int sysctl_tcp_moderate_rcvbuf __read_mostly = 1; 93int sysctl_tcp_abc __read_mostly; 94 95#define FLAG_DATA 0x01 /* Incoming frame contained data. */ 96#define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */ 97#define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */ 98#define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */ 99#define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */ 100#define FLAG_DATA_SACKED 0x20 /* New SACK. */ 101#define FLAG_ECE 0x40 /* ECE in this ACK */ 102#define FLAG_DATA_LOST 0x80 /* SACK detected data lossage. */ 103#define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/ 104#define FLAG_ONLY_ORIG_SACKED 0x200 /* SACKs only non-rexmit sent before RTO */ 105 106#define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED) 107#define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED) 108#define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE) 109#define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED) 110 111#define IsReno(tp) ((tp)->rx_opt.sack_ok == 0) 112#define IsFack(tp) ((tp)->rx_opt.sack_ok & 2) 113#define IsDSack(tp) ((tp)->rx_opt.sack_ok & 4) 114 115#define IsSackFrto() (sysctl_tcp_frto == 0x2) 116 117#define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH) 118 119/* Adapt the MSS value used to make delayed ack decision to the 120 * real world. 121 */ 122static void tcp_measure_rcv_mss(struct sock *sk, 123 const struct sk_buff *skb) 124{ 125 struct inet_connection_sock *icsk = inet_csk(sk); 126 const unsigned int lss = icsk->icsk_ack.last_seg_size; 127 unsigned int len; 128 129 icsk->icsk_ack.last_seg_size = 0; 130 131 /* skb->len may jitter because of SACKs, even if peer 132 * sends good full-sized frames. 133 */ 134 len = skb_shinfo(skb)->gso_size ?: skb->len; 135 if (len >= icsk->icsk_ack.rcv_mss) { 136 icsk->icsk_ack.rcv_mss = len; 137 } else { 138 /* Otherwise, we make more careful check taking into account, 139 * that SACKs block is variable. 140 * 141 * "len" is invariant segment length, including TCP header. 142 */ 143 len += skb->data - skb_transport_header(skb); 144 if (len >= TCP_MIN_RCVMSS + sizeof(struct tcphdr) || 145 /* If PSH is not set, packet should be 146 * full sized, provided peer TCP is not badly broken. 147 * This observation (if it is correct 8)) allows 148 * to handle super-low mtu links fairly. 149 */ 150 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) && 151 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) { 152 /* Subtract also invariant (if peer is RFC compliant), 153 * tcp header plus fixed timestamp option length. 154 * Resulting "len" is MSS free of SACK jitter. 155 */ 156 len -= tcp_sk(sk)->tcp_header_len; 157 icsk->icsk_ack.last_seg_size = len; 158 if (len == lss) { 159 icsk->icsk_ack.rcv_mss = len; 160 return; 161 } 162 } 163 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED) 164 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2; 165 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED; 166 } 167} 168 169static void tcp_incr_quickack(struct sock *sk) 170{ 171 struct inet_connection_sock *icsk = inet_csk(sk); 172 unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss); 173 174 if (quickacks==0) 175 quickacks=2; 176 if (quickacks > icsk->icsk_ack.quick) 177 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS); 178} 179 180void tcp_enter_quickack_mode(struct sock *sk) 181{ 182 struct inet_connection_sock *icsk = inet_csk(sk); 183 tcp_incr_quickack(sk); 184 icsk->icsk_ack.pingpong = 0; 185 icsk->icsk_ack.ato = TCP_ATO_MIN; 186} 187 188/* Send ACKs quickly, if "quick" count is not exhausted 189 * and the session is not interactive. 190 */ 191 192static inline int tcp_in_quickack_mode(const struct sock *sk) 193{ 194 const struct inet_connection_sock *icsk = inet_csk(sk); 195 return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong; 196} 197 198/* Buffer size and advertised window tuning. 199 * 200 * 1. Tuning sk->sk_sndbuf, when connection enters established state. 201 */ 202 203static void tcp_fixup_sndbuf(struct sock *sk) 204{ 205 int sndmem = tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER + 16 + 206 sizeof(struct sk_buff); 207 208 if (sk->sk_sndbuf < 3 * sndmem) 209 sk->sk_sndbuf = min(3 * sndmem, sysctl_tcp_wmem[2]); 210} 211 212/* 2. Tuning advertised window (window_clamp, rcv_ssthresh) 213 * 214 * All tcp_full_space() is split to two parts: "network" buffer, allocated 215 * forward and advertised in receiver window (tp->rcv_wnd) and 216 * "application buffer", required to isolate scheduling/application 217 * latencies from network. 218 * window_clamp is maximal advertised window. It can be less than 219 * tcp_full_space(), in this case tcp_full_space() - window_clamp 220 * is reserved for "application" buffer. The less window_clamp is 221 * the smoother our behaviour from viewpoint of network, but the lower 222 * throughput and the higher sensitivity of the connection to losses. 8) 223 * 224 * rcv_ssthresh is more strict window_clamp used at "slow start" 225 * phase to predict further behaviour of this connection. 226 * It is used for two goals: 227 * - to enforce header prediction at sender, even when application 228 * requires some significant "application buffer". It is check #1. 229 * - to prevent pruning of receive queue because of misprediction 230 * of receiver window. Check #2. 231 * 232 * The scheme does not work when sender sends good segments opening 233 * window and then starts to feed us spaghetti. But it should work 234 * in common situations. Otherwise, we have to rely on queue collapsing. 235 */ 236 237/* Slow part of check#2. */ 238static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb) 239{ 240 struct tcp_sock *tp = tcp_sk(sk); 241 /* Optimize this! */ 242 int truesize = tcp_win_from_space(skb->truesize)/2; 243 int window = tcp_win_from_space(sysctl_tcp_rmem[2])/2; 244 245 while (tp->rcv_ssthresh <= window) { 246 if (truesize <= skb->len) 247 return 2 * inet_csk(sk)->icsk_ack.rcv_mss; 248 249 truesize >>= 1; 250 window >>= 1; 251 } 252 return 0; 253} 254 255static void tcp_grow_window(struct sock *sk, 256 struct sk_buff *skb) 257{ 258 struct tcp_sock *tp = tcp_sk(sk); 259 260 /* Check #1 */ 261 if (tp->rcv_ssthresh < tp->window_clamp && 262 (int)tp->rcv_ssthresh < tcp_space(sk) && 263 !tcp_memory_pressure) { 264 int incr; 265 266 /* Check #2. Increase window, if skb with such overhead 267 * will fit to rcvbuf in future. 268 */ 269 if (tcp_win_from_space(skb->truesize) <= skb->len) 270 incr = 2*tp->advmss; 271 else 272 incr = __tcp_grow_window(sk, skb); 273 274 if (incr) { 275 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr, tp->window_clamp); 276 inet_csk(sk)->icsk_ack.quick |= 1; 277 } 278 } 279} 280 281/* 3. Tuning rcvbuf, when connection enters established state. */ 282 283static void tcp_fixup_rcvbuf(struct sock *sk) 284{ 285 struct tcp_sock *tp = tcp_sk(sk); 286 int rcvmem = tp->advmss + MAX_TCP_HEADER + 16 + sizeof(struct sk_buff); 287 288 /* Try to select rcvbuf so that 4 mss-sized segments 289 * will fit to window and corresponding skbs will fit to our rcvbuf. 290 * (was 3; 4 is minimum to allow fast retransmit to work.) 291 */ 292 while (tcp_win_from_space(rcvmem) < tp->advmss) 293 rcvmem += 128; 294 if (sk->sk_rcvbuf < 4 * rcvmem) 295 sk->sk_rcvbuf = min(4 * rcvmem, sysctl_tcp_rmem[2]); 296} 297 298/* 4. Try to fixup all. It is made immediately after connection enters 299 * established state. 300 */ 301static void tcp_init_buffer_space(struct sock *sk) 302{ 303 struct tcp_sock *tp = tcp_sk(sk); 304 int maxwin; 305 306 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) 307 tcp_fixup_rcvbuf(sk); 308 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK)) 309 tcp_fixup_sndbuf(sk); 310 311 tp->rcvq_space.space = tp->rcv_wnd; 312 313 maxwin = tcp_full_space(sk); 314 315 if (tp->window_clamp >= maxwin) { 316 tp->window_clamp = maxwin; 317 318 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss) 319 tp->window_clamp = max(maxwin - 320 (maxwin >> sysctl_tcp_app_win), 321 4 * tp->advmss); 322 } 323 324 /* Force reservation of one segment. */ 325 if (sysctl_tcp_app_win && 326 tp->window_clamp > 2 * tp->advmss && 327 tp->window_clamp + tp->advmss > maxwin) 328 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss); 329 330 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp); 331 tp->snd_cwnd_stamp = tcp_time_stamp; 332} 333 334/* 5. Recalculate window clamp after socket hit its memory bounds. */ 335static void tcp_clamp_window(struct sock *sk) 336{ 337 struct tcp_sock *tp = tcp_sk(sk); 338 struct inet_connection_sock *icsk = inet_csk(sk); 339 340 icsk->icsk_ack.quick = 0; 341 342 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] && 343 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) && 344 !tcp_memory_pressure && 345 atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) { 346 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc), 347 sysctl_tcp_rmem[2]); 348 } 349 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf) 350 tp->rcv_ssthresh = min(tp->window_clamp, 2U*tp->advmss); 351} 352 353 354/* Initialize RCV_MSS value. 355 * RCV_MSS is an our guess about MSS used by the peer. 356 * We haven't any direct information about the MSS. 357 * It's better to underestimate the RCV_MSS rather than overestimate. 358 * Overestimations make us ACKing less frequently than needed. 359 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss(). 360 */ 361void tcp_initialize_rcv_mss(struct sock *sk) 362{ 363 struct tcp_sock *tp = tcp_sk(sk); 364 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache); 365 366 hint = min(hint, tp->rcv_wnd/2); 367 hint = min(hint, TCP_MIN_RCVMSS); 368 hint = max(hint, TCP_MIN_MSS); 369 370 inet_csk(sk)->icsk_ack.rcv_mss = hint; 371} 372 373/* Receiver "autotuning" code. 374 * 375 * The algorithm for RTT estimation w/o timestamps is based on 376 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL. 377 * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps> 378 * 379 * More detail on this code can be found at 380 * <http://www.psc.edu/~jheffner/senior_thesis.ps>, 381 * though this reference is out of date. A new paper 382 * is pending. 383 */ 384static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep) 385{ 386 u32 new_sample = tp->rcv_rtt_est.rtt; 387 long m = sample; 388 389 if (m == 0) 390 m = 1; 391 392 if (new_sample != 0) { 393 /* If we sample in larger samples in the non-timestamp 394 * case, we could grossly overestimate the RTT especially 395 * with chatty applications or bulk transfer apps which 396 * are stalled on filesystem I/O. 397 * 398 * Also, since we are only going for a minimum in the 399 * non-timestamp case, we do not smooth things out 400 * else with timestamps disabled convergence takes too 401 * long. 402 */ 403 if (!win_dep) { 404 m -= (new_sample >> 3); 405 new_sample += m; 406 } else if (m < new_sample) 407 new_sample = m << 3; 408 } else { 409 /* No previous measure. */ 410 new_sample = m << 3; 411 } 412 413 if (tp->rcv_rtt_est.rtt != new_sample) 414 tp->rcv_rtt_est.rtt = new_sample; 415} 416 417static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp) 418{ 419 if (tp->rcv_rtt_est.time == 0) 420 goto new_measure; 421 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq)) 422 return; 423 tcp_rcv_rtt_update(tp, 424 jiffies - tp->rcv_rtt_est.time, 425 1); 426 427new_measure: 428 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd; 429 tp->rcv_rtt_est.time = tcp_time_stamp; 430} 431 432static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, const struct sk_buff *skb) 433{ 434 struct tcp_sock *tp = tcp_sk(sk); 435 if (tp->rx_opt.rcv_tsecr && 436 (TCP_SKB_CB(skb)->end_seq - 437 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss)) 438 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0); 439} 440 441/* 442 * This function should be called every time data is copied to user space. 443 * It calculates the appropriate TCP receive buffer space. 444 */ 445void tcp_rcv_space_adjust(struct sock *sk) 446{ 447 struct tcp_sock *tp = tcp_sk(sk); 448 int time; 449 int space; 450 451 if (tp->rcvq_space.time == 0) 452 goto new_measure; 453 454 time = tcp_time_stamp - tp->rcvq_space.time; 455 if (time < (tp->rcv_rtt_est.rtt >> 3) || 456 tp->rcv_rtt_est.rtt == 0) 457 return; 458 459 space = 2 * (tp->copied_seq - tp->rcvq_space.seq); 460 461 space = max(tp->rcvq_space.space, space); 462 463 if (tp->rcvq_space.space != space) { 464 int rcvmem; 465 466 tp->rcvq_space.space = space; 467 468 if (sysctl_tcp_moderate_rcvbuf && 469 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) { 470 int new_clamp = space; 471 472 /* Receive space grows, normalize in order to 473 * take into account packet headers and sk_buff 474 * structure overhead. 475 */ 476 space /= tp->advmss; 477 if (!space) 478 space = 1; 479 rcvmem = (tp->advmss + MAX_TCP_HEADER + 480 16 + sizeof(struct sk_buff)); 481 while (tcp_win_from_space(rcvmem) < tp->advmss) 482 rcvmem += 128; 483 space *= rcvmem; 484 space = min(space, sysctl_tcp_rmem[2]); 485 if (space > sk->sk_rcvbuf) { 486 sk->sk_rcvbuf = space; 487 488 /* Make the window clamp follow along. */ 489 tp->window_clamp = new_clamp; 490 } 491 } 492 } 493 494new_measure: 495 tp->rcvq_space.seq = tp->copied_seq; 496 tp->rcvq_space.time = tcp_time_stamp; 497} 498 499/* There is something which you must keep in mind when you analyze the 500 * behavior of the tp->ato delayed ack timeout interval. When a 501 * connection starts up, we want to ack as quickly as possible. The 502 * problem is that "good" TCP's do slow start at the beginning of data 503 * transmission. The means that until we send the first few ACK's the 504 * sender will sit on his end and only queue most of his data, because 505 * he can only send snd_cwnd unacked packets at any given time. For 506 * each ACK we send, he increments snd_cwnd and transmits more of his 507 * queue. -DaveM 508 */ 509static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb) 510{ 511 struct tcp_sock *tp = tcp_sk(sk); 512 struct inet_connection_sock *icsk = inet_csk(sk); 513 u32 now; 514 515 inet_csk_schedule_ack(sk); 516 517 tcp_measure_rcv_mss(sk, skb); 518 519 tcp_rcv_rtt_measure(tp); 520 521 now = tcp_time_stamp; 522 523 if (!icsk->icsk_ack.ato) { 524 /* The _first_ data packet received, initialize 525 * delayed ACK engine. 526 */ 527 tcp_incr_quickack(sk); 528 icsk->icsk_ack.ato = TCP_ATO_MIN; 529 } else { 530 int m = now - icsk->icsk_ack.lrcvtime; 531 532 if (m <= TCP_ATO_MIN/2) { 533 /* The fastest case is the first. */ 534 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2; 535 } else if (m < icsk->icsk_ack.ato) { 536 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m; 537 if (icsk->icsk_ack.ato > icsk->icsk_rto) 538 icsk->icsk_ack.ato = icsk->icsk_rto; 539 } else if (m > icsk->icsk_rto) { 540 /* Too long gap. Apparently sender failed to 541 * restart window, so that we send ACKs quickly. 542 */ 543 tcp_incr_quickack(sk); 544 sk_stream_mem_reclaim(sk); 545 } 546 } 547 icsk->icsk_ack.lrcvtime = now; 548 549 TCP_ECN_check_ce(tp, skb); 550 551 if (skb->len >= 128) 552 tcp_grow_window(sk, skb); 553} 554 555/* Called to compute a smoothed rtt estimate. The data fed to this 556 * routine either comes from timestamps, or from segments that were 557 * known _not_ to have been retransmitted [see Karn/Partridge 558 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88 559 * piece by Van Jacobson. 560 * NOTE: the next three routines used to be one big routine. 561 * To save cycles in the RFC 1323 implementation it was better to break 562 * it up into three procedures. -- erics 563 */ 564static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt) 565{ 566 struct tcp_sock *tp = tcp_sk(sk); 567 long m = mrtt; /* RTT */ 568 569 /* The following amusing code comes from Jacobson's 570 * article in SIGCOMM '88. Note that rtt and mdev 571 * are scaled versions of rtt and mean deviation. 572 * This is designed to be as fast as possible 573 * m stands for "measurement". 574 * 575 * On a 1990 paper the rto value is changed to: 576 * RTO = rtt + 4 * mdev 577 * 578 * Funny. This algorithm seems to be very broken. 579 * These formulae increase RTO, when it should be decreased, increase 580 * too slowly, when it should be increased quickly, decrease too quickly 581 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely 582 * does not matter how to _calculate_ it. Seems, it was trap 583 * that VJ failed to avoid. 8) 584 */ 585 if (m == 0) 586 m = 1; 587 if (tp->srtt != 0) { 588 m -= (tp->srtt >> 3); /* m is now error in rtt est */ 589 tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */ 590 if (m < 0) { 591 m = -m; /* m is now abs(error) */ 592 m -= (tp->mdev >> 2); /* similar update on mdev */ 593 /* This is similar to one of Eifel findings. 594 * Eifel blocks mdev updates when rtt decreases. 595 * This solution is a bit different: we use finer gain 596 * for mdev in this case (alpha*beta). 597 * Like Eifel it also prevents growth of rto, 598 * but also it limits too fast rto decreases, 599 * happening in pure Eifel. 600 */ 601 if (m > 0) 602 m >>= 3; 603 } else { 604 m -= (tp->mdev >> 2); /* similar update on mdev */ 605 } 606 tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */ 607 if (tp->mdev > tp->mdev_max) { 608 tp->mdev_max = tp->mdev; 609 if (tp->mdev_max > tp->rttvar) 610 tp->rttvar = tp->mdev_max; 611 } 612 if (after(tp->snd_una, tp->rtt_seq)) { 613 if (tp->mdev_max < tp->rttvar) 614 tp->rttvar -= (tp->rttvar-tp->mdev_max)>>2; 615 tp->rtt_seq = tp->snd_nxt; 616 tp->mdev_max = TCP_RTO_MIN; 617 } 618 } else { 619 /* no previous measure. */ 620 tp->srtt = m<<3; /* take the measured time to be rtt */ 621 tp->mdev = m<<1; /* make sure rto = 3*rtt */ 622 tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN); 623 tp->rtt_seq = tp->snd_nxt; 624 } 625} 626 627/* Calculate rto without backoff. This is the second half of Van Jacobson's 628 * routine referred to above. 629 */ 630static inline void tcp_set_rto(struct sock *sk) 631{ 632 const struct tcp_sock *tp = tcp_sk(sk); 633 /* Old crap is replaced with new one. 8) 634 * 635 * More seriously: 636 * 1. If rtt variance happened to be less 50msec, it is hallucination. 637 * It cannot be less due to utterly erratic ACK generation made 638 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_ 639 * to do with delayed acks, because at cwnd>2 true delack timeout 640 * is invisible. Actually, Linux-2.4 also generates erratic 641 * ACKs in some circumstances. 642 */ 643 inet_csk(sk)->icsk_rto = (tp->srtt >> 3) + tp->rttvar; 644 645 /* 2. Fixups made earlier cannot be right. 646 * If we do not estimate RTO correctly without them, 647 * all the algo is pure shit and should be replaced 648 * with correct one. It is exactly, which we pretend to do. 649 */ 650} 651 652/* NOTE: clamping at TCP_RTO_MIN is not required, current algo 653 * guarantees that rto is higher. 654 */ 655static inline void tcp_bound_rto(struct sock *sk) 656{ 657 if (inet_csk(sk)->icsk_rto > TCP_RTO_MAX) 658 inet_csk(sk)->icsk_rto = TCP_RTO_MAX; 659} 660 661/* Save metrics learned by this TCP session. 662 This function is called only, when TCP finishes successfully 663 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE. 664 */ 665void tcp_update_metrics(struct sock *sk) 666{ 667 struct tcp_sock *tp = tcp_sk(sk); 668 struct dst_entry *dst = __sk_dst_get(sk); 669 670 if (sysctl_tcp_nometrics_save) 671 return; 672 673 dst_confirm(dst); 674 675 if (dst && (dst->flags&DST_HOST)) { 676 const struct inet_connection_sock *icsk = inet_csk(sk); 677 int m; 678 679 if (icsk->icsk_backoff || !tp->srtt) { 680 /* This session failed to estimate rtt. Why? 681 * Probably, no packets returned in time. 682 * Reset our results. 683 */ 684 if (!(dst_metric_locked(dst, RTAX_RTT))) 685 dst->metrics[RTAX_RTT-1] = 0; 686 return; 687 } 688 689 m = dst_metric(dst, RTAX_RTT) - tp->srtt; 690 691 /* If newly calculated rtt larger than stored one, 692 * store new one. Otherwise, use EWMA. Remember, 693 * rtt overestimation is always better than underestimation. 694 */ 695 if (!(dst_metric_locked(dst, RTAX_RTT))) { 696 if (m <= 0) 697 dst->metrics[RTAX_RTT-1] = tp->srtt; 698 else 699 dst->metrics[RTAX_RTT-1] -= (m>>3); 700 } 701 702 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) { 703 if (m < 0) 704 m = -m; 705 706 /* Scale deviation to rttvar fixed point */ 707 m >>= 1; 708 if (m < tp->mdev) 709 m = tp->mdev; 710 711 if (m >= dst_metric(dst, RTAX_RTTVAR)) 712 dst->metrics[RTAX_RTTVAR-1] = m; 713 else 714 dst->metrics[RTAX_RTTVAR-1] -= 715 (dst->metrics[RTAX_RTTVAR-1] - m)>>2; 716 } 717 718 if (tp->snd_ssthresh >= 0xFFFF) { 719 /* Slow start still did not finish. */ 720 if (dst_metric(dst, RTAX_SSTHRESH) && 721 !dst_metric_locked(dst, RTAX_SSTHRESH) && 722 (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH)) 723 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1; 724 if (!dst_metric_locked(dst, RTAX_CWND) && 725 tp->snd_cwnd > dst_metric(dst, RTAX_CWND)) 726 dst->metrics[RTAX_CWND-1] = tp->snd_cwnd; 727 } else if (tp->snd_cwnd > tp->snd_ssthresh && 728 icsk->icsk_ca_state == TCP_CA_Open) { 729 /* Cong. avoidance phase, cwnd is reliable. */ 730 if (!dst_metric_locked(dst, RTAX_SSTHRESH)) 731 dst->metrics[RTAX_SSTHRESH-1] = 732 max(tp->snd_cwnd >> 1, tp->snd_ssthresh); 733 if (!dst_metric_locked(dst, RTAX_CWND)) 734 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_cwnd) >> 1; 735 } else { 736 /* Else slow start did not finish, cwnd is non-sense, 737 ssthresh may be also invalid. 738 */ 739 if (!dst_metric_locked(dst, RTAX_CWND)) 740 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_ssthresh) >> 1; 741 if (dst->metrics[RTAX_SSTHRESH-1] && 742 !dst_metric_locked(dst, RTAX_SSTHRESH) && 743 tp->snd_ssthresh > dst->metrics[RTAX_SSTHRESH-1]) 744 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh; 745 } 746 747 if (!dst_metric_locked(dst, RTAX_REORDERING)) { 748 if (dst->metrics[RTAX_REORDERING-1] < tp->reordering && 749 tp->reordering != sysctl_tcp_reordering) 750 dst->metrics[RTAX_REORDERING-1] = tp->reordering; 751 } 752 } 753} 754 755/* Numbers are taken from RFC2414. */ 756__u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst) 757{ 758 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0); 759 760 if (!cwnd) { 761 if (tp->mss_cache > 1460) 762 cwnd = 2; 763 else 764 cwnd = (tp->mss_cache > 1095) ? 3 : 4; 765 } 766 return min_t(__u32, cwnd, tp->snd_cwnd_clamp); 767} 768 769/* Set slow start threshold and cwnd not falling to slow start */ 770void tcp_enter_cwr(struct sock *sk, const int set_ssthresh) 771{ 772 struct tcp_sock *tp = tcp_sk(sk); 773 const struct inet_connection_sock *icsk = inet_csk(sk); 774 775 tp->prior_ssthresh = 0; 776 tp->bytes_acked = 0; 777 if (icsk->icsk_ca_state < TCP_CA_CWR) { 778 tp->undo_marker = 0; 779 if (set_ssthresh) 780 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 781 tp->snd_cwnd = min(tp->snd_cwnd, 782 tcp_packets_in_flight(tp) + 1U); 783 tp->snd_cwnd_cnt = 0; 784 tp->high_seq = tp->snd_nxt; 785 tp->snd_cwnd_stamp = tcp_time_stamp; 786 TCP_ECN_queue_cwr(tp); 787 788 tcp_set_ca_state(sk, TCP_CA_CWR); 789 } 790} 791 792/* Initialize metrics on socket. */ 793 794static void tcp_init_metrics(struct sock *sk) 795{ 796 struct tcp_sock *tp = tcp_sk(sk); 797 struct dst_entry *dst = __sk_dst_get(sk); 798 799 if (dst == NULL) 800 goto reset; 801 802 dst_confirm(dst); 803 804 if (dst_metric_locked(dst, RTAX_CWND)) 805 tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND); 806 if (dst_metric(dst, RTAX_SSTHRESH)) { 807 tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH); 808 if (tp->snd_ssthresh > tp->snd_cwnd_clamp) 809 tp->snd_ssthresh = tp->snd_cwnd_clamp; 810 } 811 if (dst_metric(dst, RTAX_REORDERING) && 812 tp->reordering != dst_metric(dst, RTAX_REORDERING)) { 813 tp->rx_opt.sack_ok &= ~2; 814 tp->reordering = dst_metric(dst, RTAX_REORDERING); 815 } 816 817 if (dst_metric(dst, RTAX_RTT) == 0) 818 goto reset; 819 820 if (!tp->srtt && dst_metric(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3)) 821 goto reset; 822 823 /* Initial rtt is determined from SYN,SYN-ACK. 824 * The segment is small and rtt may appear much 825 * less than real one. Use per-dst memory 826 * to make it more realistic. 827 * 828 * A bit of theory. RTT is time passed after "normal" sized packet 829 * is sent until it is ACKed. In normal circumstances sending small 830 * packets force peer to delay ACKs and calculation is correct too. 831 * The algorithm is adaptive and, provided we follow specs, it 832 * NEVER underestimate RTT. BUT! If peer tries to make some clever 833 * tricks sort of "quick acks" for time long enough to decrease RTT 834 * to low value, and then abruptly stops to do it and starts to delay 835 * ACKs, wait for troubles. 836 */ 837 if (dst_metric(dst, RTAX_RTT) > tp->srtt) { 838 tp->srtt = dst_metric(dst, RTAX_RTT); 839 tp->rtt_seq = tp->snd_nxt; 840 } 841 if (dst_metric(dst, RTAX_RTTVAR) > tp->mdev) { 842 tp->mdev = dst_metric(dst, RTAX_RTTVAR); 843 tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN); 844 } 845 tcp_set_rto(sk); 846 tcp_bound_rto(sk); 847 if (inet_csk(sk)->icsk_rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp) 848 goto reset; 849 tp->snd_cwnd = tcp_init_cwnd(tp, dst); 850 tp->snd_cwnd_stamp = tcp_time_stamp; 851 return; 852 853reset: 854 /* Play conservative. If timestamps are not 855 * supported, TCP will fail to recalculate correct 856 * rtt, if initial rto is too small. FORGET ALL AND RESET! 857 */ 858 if (!tp->rx_opt.saw_tstamp && tp->srtt) { 859 tp->srtt = 0; 860 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT; 861 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; 862 } 863} 864 865static void tcp_update_reordering(struct sock *sk, const int metric, 866 const int ts) 867{ 868 struct tcp_sock *tp = tcp_sk(sk); 869 if (metric > tp->reordering) { 870 tp->reordering = min(TCP_MAX_REORDERING, metric); 871 872 /* This exciting event is worth to be remembered. 8) */ 873 if (ts) 874 NET_INC_STATS_BH(LINUX_MIB_TCPTSREORDER); 875 else if (IsReno(tp)) 876 NET_INC_STATS_BH(LINUX_MIB_TCPRENOREORDER); 877 else if (IsFack(tp)) 878 NET_INC_STATS_BH(LINUX_MIB_TCPFACKREORDER); 879 else 880 NET_INC_STATS_BH(LINUX_MIB_TCPSACKREORDER); 881#if FASTRETRANS_DEBUG > 1 882 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n", 883 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state, 884 tp->reordering, 885 tp->fackets_out, 886 tp->sacked_out, 887 tp->undo_marker ? tp->undo_retrans : 0); 888#endif 889 /* Disable FACK yet. */ 890 tp->rx_opt.sack_ok &= ~2; 891 } 892} 893 894/* This procedure tags the retransmission queue when SACKs arrive. 895 * 896 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L). 897 * Packets in queue with these bits set are counted in variables 898 * sacked_out, retrans_out and lost_out, correspondingly. 899 * 900 * Valid combinations are: 901 * Tag InFlight Description 902 * 0 1 - orig segment is in flight. 903 * S 0 - nothing flies, orig reached receiver. 904 * L 0 - nothing flies, orig lost by net. 905 * R 2 - both orig and retransmit are in flight. 906 * L|R 1 - orig is lost, retransmit is in flight. 907 * S|R 1 - orig reached receiver, retrans is still in flight. 908 * (L|S|R is logically valid, it could occur when L|R is sacked, 909 * but it is equivalent to plain S and code short-curcuits it to S. 910 * L|S is logically invalid, it would mean -1 packet in flight 8)) 911 * 912 * These 6 states form finite state machine, controlled by the following events: 913 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue()) 914 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue()) 915 * 3. Loss detection event of one of three flavors: 916 * A. Scoreboard estimator decided the packet is lost. 917 * A'. Reno "three dupacks" marks head of queue lost. 918 * A''. Its FACK modfication, head until snd.fack is lost. 919 * B. SACK arrives sacking data transmitted after never retransmitted 920 * hole was sent out. 921 * C. SACK arrives sacking SND.NXT at the moment, when the 922 * segment was retransmitted. 923 * 4. D-SACK added new rule: D-SACK changes any tag to S. 924 * 925 * It is pleasant to note, that state diagram turns out to be commutative, 926 * so that we are allowed not to be bothered by order of our actions, 927 * when multiple events arrive simultaneously. (see the function below). 928 * 929 * Reordering detection. 930 * -------------------- 931 * Reordering metric is maximal distance, which a packet can be displaced 932 * in packet stream. With SACKs we can estimate it: 933 * 934 * 1. SACK fills old hole and the corresponding segment was not 935 * ever retransmitted -> reordering. Alas, we cannot use it 936 * when segment was retransmitted. 937 * 2. The last flaw is solved with D-SACK. D-SACK arrives 938 * for retransmitted and already SACKed segment -> reordering.. 939 * Both of these heuristics are not used in Loss state, when we cannot 940 * account for retransmits accurately. 941 */ 942static int 943tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, u32 prior_snd_una) 944{ 945 const struct inet_connection_sock *icsk = inet_csk(sk); 946 struct tcp_sock *tp = tcp_sk(sk); 947 unsigned char *ptr = (skb_transport_header(ack_skb) + 948 TCP_SKB_CB(ack_skb)->sacked); 949 struct tcp_sack_block_wire *sp = (struct tcp_sack_block_wire *)(ptr+2); 950 struct sk_buff *cached_skb; 951 int num_sacks = (ptr[1] - TCPOLEN_SACK_BASE)>>3; 952 int reord = tp->packets_out; 953 int prior_fackets; 954 u32 lost_retrans = 0; 955 int flag = 0; 956 int dup_sack = 0; 957 int cached_fack_count; 958 int i; 959 int first_sack_index; 960 961 if (!tp->sacked_out) 962 tp->fackets_out = 0; 963 prior_fackets = tp->fackets_out; 964 965 /* Check for D-SACK. */ 966 if (before(ntohl(sp[0].start_seq), TCP_SKB_CB(ack_skb)->ack_seq)) { 967 dup_sack = 1; 968 tp->rx_opt.sack_ok |= 4; 969 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV); 970 } else if (num_sacks > 1 && 971 !after(ntohl(sp[0].end_seq), ntohl(sp[1].end_seq)) && 972 !before(ntohl(sp[0].start_seq), ntohl(sp[1].start_seq))) { 973 dup_sack = 1; 974 tp->rx_opt.sack_ok |= 4; 975 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV); 976 } 977 978 /* D-SACK for already forgotten data... 979 * Do dumb counting. */ 980 if (dup_sack && 981 !after(ntohl(sp[0].end_seq), prior_snd_una) && 982 after(ntohl(sp[0].end_seq), tp->undo_marker)) 983 tp->undo_retrans--; 984 985 /* Eliminate too old ACKs, but take into 986 * account more or less fresh ones, they can 987 * contain valid SACK info. 988 */ 989 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window)) 990 return 0; 991 992 /* SACK fastpath: 993 * if the only SACK change is the increase of the end_seq of 994 * the first block then only apply that SACK block 995 * and use retrans queue hinting otherwise slowpath */ 996 flag = 1; 997 for (i = 0; i < num_sacks; i++) { 998 __be32 start_seq = sp[i].start_seq; 999 __be32 end_seq = sp[i].end_seq; 1000 1001 if (i == 0) { 1002 if (tp->recv_sack_cache[i].start_seq != start_seq) 1003 flag = 0; 1004 } else { 1005 if ((tp->recv_sack_cache[i].start_seq != start_seq) || 1006 (tp->recv_sack_cache[i].end_seq != end_seq)) 1007 flag = 0; 1008 } 1009 tp->recv_sack_cache[i].start_seq = start_seq; 1010 tp->recv_sack_cache[i].end_seq = end_seq; 1011 } 1012 /* Clear the rest of the cache sack blocks so they won't match mistakenly. */ 1013 for (; i < ARRAY_SIZE(tp->recv_sack_cache); i++) { 1014 tp->recv_sack_cache[i].start_seq = 0; 1015 tp->recv_sack_cache[i].end_seq = 0; 1016 } 1017 1018 first_sack_index = 0; 1019 if (flag) 1020 num_sacks = 1; 1021 else { 1022 int j; 1023 tp->fastpath_skb_hint = NULL; 1024 1025 /* order SACK blocks to allow in order walk of the retrans queue */ 1026 for (i = num_sacks-1; i > 0; i--) { 1027 for (j = 0; j < i; j++){ 1028 if (after(ntohl(sp[j].start_seq), 1029 ntohl(sp[j+1].start_seq))){ 1030 struct tcp_sack_block_wire tmp; 1031 1032 tmp = sp[j]; 1033 sp[j] = sp[j+1]; 1034 sp[j+1] = tmp; 1035 1036 /* Track where the first SACK block goes to */ 1037 if (j == first_sack_index) 1038 first_sack_index = j+1; 1039 } 1040 1041 } 1042 } 1043 } 1044 1045 /* clear flag as used for different purpose in following code */ 1046 flag = 0; 1047 1048 /* Use SACK fastpath hint if valid */ 1049 cached_skb = tp->fastpath_skb_hint; 1050 cached_fack_count = tp->fastpath_cnt_hint; 1051 if (!cached_skb) { 1052 cached_skb = tcp_write_queue_head(sk); 1053 cached_fack_count = 0; 1054 } 1055 1056 for (i=0; i<num_sacks; i++, sp++) { 1057 struct sk_buff *skb; 1058 __u32 start_seq = ntohl(sp->start_seq); 1059 __u32 end_seq = ntohl(sp->end_seq); 1060 int fack_count; 1061 1062 skb = cached_skb; 1063 fack_count = cached_fack_count; 1064 1065 /* Event "B" in the comment above. */ 1066 if (after(end_seq, tp->high_seq)) 1067 flag |= FLAG_DATA_LOST; 1068 1069 tcp_for_write_queue_from(skb, sk) { 1070 int in_sack, pcount; 1071 u8 sacked; 1072 1073 if (skb == tcp_send_head(sk)) 1074 break; 1075 1076 cached_skb = skb; 1077 cached_fack_count = fack_count; 1078 if (i == first_sack_index) { 1079 tp->fastpath_skb_hint = skb; 1080 tp->fastpath_cnt_hint = fack_count; 1081 } 1082 1083 /* The retransmission queue is always in order, so 1084 * we can short-circuit the walk early. 1085 */ 1086 if (!before(TCP_SKB_CB(skb)->seq, end_seq)) 1087 break; 1088 1089 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && 1090 !before(end_seq, TCP_SKB_CB(skb)->end_seq); 1091 1092 pcount = tcp_skb_pcount(skb); 1093 1094 if (pcount > 1 && !in_sack && 1095 after(TCP_SKB_CB(skb)->end_seq, start_seq)) { 1096 unsigned int pkt_len; 1097 1098 in_sack = !after(start_seq, 1099 TCP_SKB_CB(skb)->seq); 1100 1101 if (!in_sack) 1102 pkt_len = (start_seq - 1103 TCP_SKB_CB(skb)->seq); 1104 else 1105 pkt_len = (end_seq - 1106 TCP_SKB_CB(skb)->seq); 1107 if (tcp_fragment(sk, skb, pkt_len, skb_shinfo(skb)->gso_size)) 1108 break; 1109 pcount = tcp_skb_pcount(skb); 1110 } 1111 1112 fack_count += pcount; 1113 1114 sacked = TCP_SKB_CB(skb)->sacked; 1115 1116 /* Account D-SACK for retransmitted packet. */ 1117 if ((dup_sack && in_sack) && 1118 (sacked & TCPCB_RETRANS) && 1119 after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker)) 1120 tp->undo_retrans--; 1121 1122 /* The frame is ACKed. */ 1123 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) { 1124 if (sacked&TCPCB_RETRANS) { 1125 if ((dup_sack && in_sack) && 1126 (sacked&TCPCB_SACKED_ACKED)) 1127 reord = min(fack_count, reord); 1128 } else { 1129 /* If it was in a hole, we detected reordering. */ 1130 if (fack_count < prior_fackets && 1131 !(sacked&TCPCB_SACKED_ACKED)) 1132 reord = min(fack_count, reord); 1133 } 1134 1135 /* Nothing to do; acked frame is about to be dropped. */ 1136 continue; 1137 } 1138 1139 if ((sacked&TCPCB_SACKED_RETRANS) && 1140 after(end_seq, TCP_SKB_CB(skb)->ack_seq) && 1141 (!lost_retrans || after(end_seq, lost_retrans))) 1142 lost_retrans = end_seq; 1143 1144 if (!in_sack) 1145 continue; 1146 1147 if (!(sacked&TCPCB_SACKED_ACKED)) { 1148 if (sacked & TCPCB_SACKED_RETRANS) { 1149 /* If the segment is not tagged as lost, 1150 * we do not clear RETRANS, believing 1151 * that retransmission is still in flight. 1152 */ 1153 if (sacked & TCPCB_LOST) { 1154 TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS); 1155 tp->lost_out -= tcp_skb_pcount(skb); 1156 tp->retrans_out -= tcp_skb_pcount(skb); 1157 1158 /* clear lost hint */ 1159 tp->retransmit_skb_hint = NULL; 1160 } 1161 } else { 1162 /* New sack for not retransmitted frame, 1163 * which was in hole. It is reordering. 1164 */ 1165 if (!(sacked & TCPCB_RETRANS) && 1166 fack_count < prior_fackets) 1167 reord = min(fack_count, reord); 1168 1169 if (sacked & TCPCB_LOST) { 1170 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; 1171 tp->lost_out -= tcp_skb_pcount(skb); 1172 1173 /* clear lost hint */ 1174 tp->retransmit_skb_hint = NULL; 1175 } 1176 /* SACK enhanced F-RTO detection. 1177 * Set flag if and only if non-rexmitted 1178 * segments below frto_highmark are 1179 * SACKed (RFC4138; Appendix B). 1180 * Clearing correct due to in-order walk 1181 */ 1182 if (after(end_seq, tp->frto_highmark)) { 1183 flag &= ~FLAG_ONLY_ORIG_SACKED; 1184 } else { 1185 if (!(sacked & TCPCB_RETRANS)) 1186 flag |= FLAG_ONLY_ORIG_SACKED; 1187 } 1188 } 1189 1190 TCP_SKB_CB(skb)->sacked |= TCPCB_SACKED_ACKED; 1191 flag |= FLAG_DATA_SACKED; 1192 tp->sacked_out += tcp_skb_pcount(skb); 1193 1194 if (fack_count > tp->fackets_out) 1195 tp->fackets_out = fack_count; 1196 } else { 1197 if (dup_sack && (sacked&TCPCB_RETRANS)) 1198 reord = min(fack_count, reord); 1199 } 1200 1201 /* D-SACK. We can detect redundant retransmission 1202 * in S|R and plain R frames and clear it. 1203 * undo_retrans is decreased above, L|R frames 1204 * are accounted above as well. 1205 */ 1206 if (dup_sack && 1207 (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS)) { 1208 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 1209 tp->retrans_out -= tcp_skb_pcount(skb); 1210 tp->retransmit_skb_hint = NULL; 1211 } 1212 } 1213 } 1214 1215 /* Check for lost retransmit. This superb idea is 1216 * borrowed from "ratehalving". Event "C". 1217 * Later note: FACK people cheated me again 8), 1218 * we have to account for reordering! Ugly, 1219 * but should help. 1220 */ 1221 if (lost_retrans && icsk->icsk_ca_state == TCP_CA_Recovery) { 1222 struct sk_buff *skb; 1223 1224 tcp_for_write_queue(skb, sk) { 1225 if (skb == tcp_send_head(sk)) 1226 break; 1227 if (after(TCP_SKB_CB(skb)->seq, lost_retrans)) 1228 break; 1229 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 1230 continue; 1231 if ((TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) && 1232 after(lost_retrans, TCP_SKB_CB(skb)->ack_seq) && 1233 (IsFack(tp) || 1234 !before(lost_retrans, 1235 TCP_SKB_CB(skb)->ack_seq + tp->reordering * 1236 tp->mss_cache))) { 1237 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 1238 tp->retrans_out -= tcp_skb_pcount(skb); 1239 1240 /* clear lost hint */ 1241 tp->retransmit_skb_hint = NULL; 1242 1243 if (!(TCP_SKB_CB(skb)->sacked&(TCPCB_LOST|TCPCB_SACKED_ACKED))) { 1244 tp->lost_out += tcp_skb_pcount(skb); 1245 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 1246 flag |= FLAG_DATA_SACKED; 1247 NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT); 1248 } 1249 } 1250 } 1251 } 1252 1253 tp->left_out = tp->sacked_out + tp->lost_out; 1254 1255 if ((reord < tp->fackets_out) && icsk->icsk_ca_state != TCP_CA_Loss && 1256 (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark))) 1257 tcp_update_reordering(sk, ((tp->fackets_out + 1) - reord), 0); 1258 1259#if FASTRETRANS_DEBUG > 0 1260 BUG_TRAP((int)tp->sacked_out >= 0); 1261 BUG_TRAP((int)tp->lost_out >= 0); 1262 BUG_TRAP((int)tp->retrans_out >= 0); 1263 BUG_TRAP((int)tcp_packets_in_flight(tp) >= 0); 1264#endif 1265 return flag; 1266} 1267 1268/* F-RTO can only be used if TCP has never retransmitted anything other than 1269 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here) 1270 */ 1271int tcp_use_frto(struct sock *sk) 1272{ 1273 const struct tcp_sock *tp = tcp_sk(sk); 1274 struct sk_buff *skb; 1275 1276 if (!sysctl_tcp_frto) 1277 return 0; 1278 1279 if (IsSackFrto()) 1280 return 1; 1281 1282 /* Avoid expensive walking of rexmit queue if possible */ 1283 if (tp->retrans_out > 1) 1284 return 0; 1285 1286 skb = tcp_write_queue_head(sk); 1287 skb = tcp_write_queue_next(sk, skb); /* Skips head */ 1288 tcp_for_write_queue_from(skb, sk) { 1289 if (skb == tcp_send_head(sk)) 1290 break; 1291 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS) 1292 return 0; 1293 /* Short-circuit when first non-SACKed skb has been checked */ 1294 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) 1295 break; 1296 } 1297 return 1; 1298} 1299 1300/* RTO occurred, but do not yet enter Loss state. Instead, defer RTO 1301 * recovery a bit and use heuristics in tcp_process_frto() to detect if 1302 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to 1303 * keep retrans_out counting accurate (with SACK F-RTO, other than head 1304 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS 1305 * bits are handled if the Loss state is really to be entered (in 1306 * tcp_enter_frto_loss). 1307 * 1308 * Do like tcp_enter_loss() would; when RTO expires the second time it 1309 * does: 1310 * "Reduce ssthresh if it has not yet been made inside this window." 1311 */ 1312void tcp_enter_frto(struct sock *sk) 1313{ 1314 const struct inet_connection_sock *icsk = inet_csk(sk); 1315 struct tcp_sock *tp = tcp_sk(sk); 1316 struct sk_buff *skb; 1317 1318 if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) || 1319 tp->snd_una == tp->high_seq || 1320 ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) && 1321 !icsk->icsk_retransmits)) { 1322 tp->prior_ssthresh = tcp_current_ssthresh(sk); 1323 /* Our state is too optimistic in ssthresh() call because cwnd 1324 * is not reduced until tcp_enter_frto_loss() when previous FRTO 1325 * recovery has not yet completed. Pattern would be this: RTO, 1326 * Cumulative ACK, RTO (2xRTO for the same segment does not end 1327 * up here twice). 1328 * RFC4138 should be more specific on what to do, even though 1329 * RTO is quite unlikely to occur after the first Cumulative ACK 1330 * due to back-off and complexity of triggering events ... 1331 */ 1332 if (tp->frto_counter) { 1333 u32 stored_cwnd; 1334 stored_cwnd = tp->snd_cwnd; 1335 tp->snd_cwnd = 2; 1336 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 1337 tp->snd_cwnd = stored_cwnd; 1338 } else { 1339 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 1340 } 1341 /* ... in theory, cong.control module could do "any tricks" in 1342 * ssthresh(), which means that ca_state, lost bits and lost_out 1343 * counter would have to be faked before the call occurs. We 1344 * consider that too expensive, unlikely and hacky, so modules 1345 * using these in ssthresh() must deal these incompatibility 1346 * issues if they receives CA_EVENT_FRTO and frto_counter != 0 1347 */ 1348 tcp_ca_event(sk, CA_EVENT_FRTO); 1349 } 1350 1351 tp->undo_marker = tp->snd_una; 1352 tp->undo_retrans = 0; 1353 1354 skb = tcp_write_queue_head(sk); 1355 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 1356 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 1357 tp->retrans_out -= tcp_skb_pcount(skb); 1358 } 1359 tcp_sync_left_out(tp); 1360 1361 /* Earlier loss recovery underway (see RFC4138; Appendix B). 1362 * The last condition is necessary at least in tp->frto_counter case. 1363 */ 1364 if (IsSackFrto() && (tp->frto_counter || 1365 ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) && 1366 after(tp->high_seq, tp->snd_una)) { 1367 tp->frto_highmark = tp->high_seq; 1368 } else { 1369 tp->frto_highmark = tp->snd_nxt; 1370 } 1371 tcp_set_ca_state(sk, TCP_CA_Disorder); 1372 tp->high_seq = tp->snd_nxt; 1373 tp->frto_counter = 1; 1374} 1375 1376/* Enter Loss state after F-RTO was applied. Dupack arrived after RTO, 1377 * which indicates that we should follow the traditional RTO recovery, 1378 * i.e. mark everything lost and do go-back-N retransmission. 1379 */ 1380static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag) 1381{ 1382 struct tcp_sock *tp = tcp_sk(sk); 1383 struct sk_buff *skb; 1384 int cnt = 0; 1385 1386 tp->sacked_out = 0; 1387 tp->lost_out = 0; 1388 tp->fackets_out = 0; 1389 tp->retrans_out = 0; 1390 1391 tcp_for_write_queue(skb, sk) { 1392 if (skb == tcp_send_head(sk)) 1393 break; 1394 cnt += tcp_skb_pcount(skb); 1395 /* 1396 * Count the retransmission made on RTO correctly (only when 1397 * waiting for the first ACK and did not get it)... 1398 */ 1399 if ((tp->frto_counter == 1) && !(flag&FLAG_DATA_ACKED)) { 1400 tp->retrans_out += tcp_skb_pcount(skb); 1401 /* ...enter this if branch just for the first segment */ 1402 flag |= FLAG_DATA_ACKED; 1403 } else { 1404 TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS); 1405 } 1406 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) { 1407 1408 /* Do not mark those segments lost that were 1409 * forward transmitted after RTO 1410 */ 1411 if (!after(TCP_SKB_CB(skb)->end_seq, 1412 tp->frto_highmark)) { 1413 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 1414 tp->lost_out += tcp_skb_pcount(skb); 1415 } 1416 } else { 1417 tp->sacked_out += tcp_skb_pcount(skb); 1418 tp->fackets_out = cnt; 1419 } 1420 } 1421 tcp_sync_left_out(tp); 1422 1423 tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments; 1424 tp->snd_cwnd_cnt = 0; 1425 tp->snd_cwnd_stamp = tcp_time_stamp; 1426 tp->undo_marker = 0; 1427 tp->frto_counter = 0; 1428 1429 tp->reordering = min_t(unsigned int, tp->reordering, 1430 sysctl_tcp_reordering); 1431 tcp_set_ca_state(sk, TCP_CA_Loss); 1432 tp->high_seq = tp->frto_highmark; 1433 TCP_ECN_queue_cwr(tp); 1434 1435 clear_all_retrans_hints(tp); 1436} 1437 1438void tcp_clear_retrans(struct tcp_sock *tp) 1439{ 1440 tp->left_out = 0; 1441 tp->retrans_out = 0; 1442 1443 tp->fackets_out = 0; 1444 tp->sacked_out = 0; 1445 tp->lost_out = 0; 1446 1447 tp->undo_marker = 0; 1448 tp->undo_retrans = 0; 1449} 1450 1451/* Enter Loss state. If "how" is not zero, forget all SACK information 1452 * and reset tags completely, otherwise preserve SACKs. If receiver 1453 * dropped its ofo queue, we will know this due to reneging detection. 1454 */ 1455void tcp_enter_loss(struct sock *sk, int how) 1456{ 1457 const struct inet_connection_sock *icsk = inet_csk(sk); 1458 struct tcp_sock *tp = tcp_sk(sk); 1459 struct sk_buff *skb; 1460 int cnt = 0; 1461 1462 /* Reduce ssthresh if it has not yet been made inside this window. */ 1463 if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq || 1464 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) { 1465 tp->prior_ssthresh = tcp_current_ssthresh(sk); 1466 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 1467 tcp_ca_event(sk, CA_EVENT_LOSS); 1468 } 1469 tp->snd_cwnd = 1; 1470 tp->snd_cwnd_cnt = 0; 1471 tp->snd_cwnd_stamp = tcp_time_stamp; 1472 1473 tp->bytes_acked = 0; 1474 tcp_clear_retrans(tp); 1475 1476 /* Push undo marker, if it was plain RTO and nothing 1477 * was retransmitted. */ 1478 if (!how) 1479 tp->undo_marker = tp->snd_una; 1480 1481 tcp_for_write_queue(skb, sk) { 1482 if (skb == tcp_send_head(sk)) 1483 break; 1484 cnt += tcp_skb_pcount(skb); 1485 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS) 1486 tp->undo_marker = 0; 1487 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED; 1488 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) { 1489 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED; 1490 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 1491 tp->lost_out += tcp_skb_pcount(skb); 1492 } else { 1493 tp->sacked_out += tcp_skb_pcount(skb); 1494 tp->fackets_out = cnt; 1495 } 1496 } 1497 tcp_sync_left_out(tp); 1498 1499 tp->reordering = min_t(unsigned int, tp->reordering, 1500 sysctl_tcp_reordering); 1501 tcp_set_ca_state(sk, TCP_CA_Loss); 1502 tp->high_seq = tp->snd_nxt; 1503 TCP_ECN_queue_cwr(tp); 1504 1505 clear_all_retrans_hints(tp); 1506} 1507 1508static int tcp_check_sack_reneging(struct sock *sk) 1509{ 1510 struct sk_buff *skb; 1511 1512 /* If ACK arrived pointing to a remembered SACK, 1513 * it means that our remembered SACKs do not reflect 1514 * real state of receiver i.e. 1515 * receiver _host_ is heavily congested (or buggy). 1516 * Do processing similar to RTO timeout. 1517 */ 1518 if ((skb = tcp_write_queue_head(sk)) != NULL && 1519 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) { 1520 struct inet_connection_sock *icsk = inet_csk(sk); 1521 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING); 1522 1523 tcp_enter_loss(sk, 1); 1524 icsk->icsk_retransmits++; 1525 tcp_retransmit_skb(sk, tcp_write_queue_head(sk)); 1526 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 1527 icsk->icsk_rto, TCP_RTO_MAX); 1528 return 1; 1529 } 1530 return 0; 1531} 1532 1533static inline int tcp_fackets_out(struct tcp_sock *tp) 1534{ 1535 return IsReno(tp) ? tp->sacked_out+1 : tp->fackets_out; 1536} 1537 1538static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb) 1539{ 1540 return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto); 1541} 1542 1543static inline int tcp_head_timedout(struct sock *sk) 1544{ 1545 struct tcp_sock *tp = tcp_sk(sk); 1546 1547 return tp->packets_out && 1548 tcp_skb_timedout(sk, tcp_write_queue_head(sk)); 1549} 1550 1551/* Linux NewReno/SACK/FACK/ECN state machine. 1552 * -------------------------------------- 1553 * 1554 * "Open" Normal state, no dubious events, fast path. 1555 * "Disorder" In all the respects it is "Open", 1556 * but requires a bit more attention. It is entered when 1557 * we see some SACKs or dupacks. It is split of "Open" 1558 * mainly to move some processing from fast path to slow one. 1559 * "CWR" CWND was reduced due to some Congestion Notification event. 1560 * It can be ECN, ICMP source quench, local device congestion. 1561 * "Recovery" CWND was reduced, we are fast-retransmitting. 1562 * "Loss" CWND was reduced due to RTO timeout or SACK reneging. 1563 * 1564 * tcp_fastretrans_alert() is entered: 1565 * - each incoming ACK, if state is not "Open" 1566 * - when arrived ACK is unusual, namely: 1567 * * SACK 1568 * * Duplicate ACK. 1569 * * ECN ECE. 1570 * 1571 * Counting packets in flight is pretty simple. 1572 * 1573 * in_flight = packets_out - left_out + retrans_out 1574 * 1575 * packets_out is SND.NXT-SND.UNA counted in packets. 1576 * 1577 * retrans_out is number of retransmitted segments. 1578 * 1579 * left_out is number of segments left network, but not ACKed yet. 1580 * 1581 * left_out = sacked_out + lost_out 1582 * 1583 * sacked_out: Packets, which arrived to receiver out of order 1584 * and hence not ACKed. With SACKs this number is simply 1585 * amount of SACKed data. Even without SACKs 1586 * it is easy to give pretty reliable estimate of this number, 1587 * counting duplicate ACKs. 1588 * 1589 * lost_out: Packets lost by network. TCP has no explicit 1590 * "loss notification" feedback from network (for now). 1591 * It means that this number can be only _guessed_. 1592 * Actually, it is the heuristics to predict lossage that 1593 * distinguishes different algorithms. 1594 * 1595 * F.e. after RTO, when all the queue is considered as lost, 1596 * lost_out = packets_out and in_flight = retrans_out. 1597 * 1598 * Essentially, we have now two algorithms counting 1599 * lost packets. 1600 * 1601 * FACK: It is the simplest heuristics. As soon as we decided 1602 * that something is lost, we decide that _all_ not SACKed 1603 * packets until the most forward SACK are lost. I.e. 1604 * lost_out = fackets_out - sacked_out and left_out = fackets_out. 1605 * It is absolutely correct estimate, if network does not reorder 1606 * packets. And it loses any connection to reality when reordering 1607 * takes place. We use FACK by default until reordering 1608 * is suspected on the path to this destination. 1609 * 1610 * NewReno: when Recovery is entered, we assume that one segment 1611 * is lost (classic Reno). While we are in Recovery and 1612 * a partial ACK arrives, we assume that one more packet 1613 * is lost (NewReno). This heuristics are the same in NewReno 1614 * and SACK. 1615 * 1616 * Imagine, that's all! Forget about all this shamanism about CWND inflation 1617 * deflation etc. CWND is real congestion window, never inflated, changes 1618 * only according to classic VJ rules. 1619 * 1620 * Really tricky (and requiring careful tuning) part of algorithm 1621 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue(). 1622 * The first determines the moment _when_ we should reduce CWND and, 1623 * hence, slow down forward transmission. In fact, it determines the moment 1624 * when we decide that hole is caused by loss, rather than by a reorder. 1625 * 1626 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill 1627 * holes, caused by lost packets. 1628 * 1629 * And the most logically complicated part of algorithm is undo 1630 * heuristics. We detect false retransmits due to both too early 1631 * fast retransmit (reordering) and underestimated RTO, analyzing 1632 * timestamps and D-SACKs. When we detect that some segments were 1633 * retransmitted by mistake and CWND reduction was wrong, we undo 1634 * window reduction and abort recovery phase. This logic is hidden 1635 * inside several functions named tcp_try_undo_<something>. 1636 */ 1637 1638/* This function decides, when we should leave Disordered state 1639 * and enter Recovery phase, reducing congestion window. 1640 * 1641 * Main question: may we further continue forward transmission 1642 * with the same cwnd? 1643 */ 1644static int tcp_time_to_recover(struct sock *sk) 1645{ 1646 struct tcp_sock *tp = tcp_sk(sk); 1647 __u32 packets_out; 1648 1649 /* Do not perform any recovery during FRTO algorithm */ 1650 if (tp->frto_counter) 1651 return 0; 1652 1653 /* Trick#1: The loss is proven. */ 1654 if (tp->lost_out) 1655 return 1; 1656 1657 /* Not-A-Trick#2 : Classic rule... */ 1658 if (tcp_fackets_out(tp) > tp->reordering) 1659 return 1; 1660 1661 /* Trick#3 : when we use RFC2988 timer restart, fast 1662 * retransmit can be triggered by timeout of queue head. 1663 */ 1664 if (tcp_head_timedout(sk)) 1665 return 1; 1666 1667 /* Trick#4: It is still not OK... But will it be useful to delay 1668 * recovery more? 1669 */ 1670 packets_out = tp->packets_out; 1671 if (packets_out <= tp->reordering && 1672 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) && 1673 !tcp_may_send_now(sk)) { 1674 /* We have nothing to send. This connection is limited 1675 * either by receiver window or by application. 1676 */ 1677 return 1; 1678 } 1679 1680 return 0; 1681} 1682 1683/* If we receive more dupacks than we expected counting segments 1684 * in assumption of absent reordering, interpret this as reordering. 1685 * The only another reason could be bug in receiver TCP. 1686 */ 1687static void tcp_check_reno_reordering(struct sock *sk, const int addend) 1688{ 1689 struct tcp_sock *tp = tcp_sk(sk); 1690 u32 holes; 1691 1692 holes = max(tp->lost_out, 1U); 1693 holes = min(holes, tp->packets_out); 1694 1695 if ((tp->sacked_out + holes) > tp->packets_out) { 1696 tp->sacked_out = tp->packets_out - holes; 1697 tcp_update_reordering(sk, tp->packets_out + addend, 0); 1698 } 1699} 1700 1701/* Emulate SACKs for SACKless connection: account for a new dupack. */ 1702 1703static void tcp_add_reno_sack(struct sock *sk) 1704{ 1705 struct tcp_sock *tp = tcp_sk(sk); 1706 tp->sacked_out++; 1707 tcp_check_reno_reordering(sk, 0); 1708 tcp_sync_left_out(tp); 1709} 1710 1711/* Account for ACK, ACKing some data in Reno Recovery phase. */ 1712 1713static void tcp_remove_reno_sacks(struct sock *sk, int acked) 1714{ 1715 struct tcp_sock *tp = tcp_sk(sk); 1716 1717 if (acked > 0) { 1718 /* One ACK acked hole. The rest eat duplicate ACKs. */ 1719 if (acked-1 >= tp->sacked_out) 1720 tp->sacked_out = 0; 1721 else 1722 tp->sacked_out -= acked-1; 1723 } 1724 tcp_check_reno_reordering(sk, acked); 1725 tcp_sync_left_out(tp); 1726} 1727 1728static inline void tcp_reset_reno_sack(struct tcp_sock *tp) 1729{ 1730 tp->sacked_out = 0; 1731 tp->left_out = tp->lost_out; 1732} 1733 1734/* Mark head of queue up as lost. */ 1735static void tcp_mark_head_lost(struct sock *sk, 1736 int packets, u32 high_seq) 1737{ 1738 struct tcp_sock *tp = tcp_sk(sk); 1739 struct sk_buff *skb; 1740 int cnt; 1741 1742 BUG_TRAP(packets <= tp->packets_out); 1743 if (tp->lost_skb_hint) { 1744 skb = tp->lost_skb_hint; 1745 cnt = tp->lost_cnt_hint; 1746 } else { 1747 skb = tcp_write_queue_head(sk); 1748 cnt = 0; 1749 } 1750 1751 tcp_for_write_queue_from(skb, sk) { 1752 if (skb == tcp_send_head(sk)) 1753 break; 1754 /* TODO: do this better */ 1755 /* this is not the most efficient way to do this... */ 1756 tp->lost_skb_hint = skb; 1757 tp->lost_cnt_hint = cnt; 1758 cnt += tcp_skb_pcount(skb); 1759 if (cnt > packets || after(TCP_SKB_CB(skb)->end_seq, high_seq)) 1760 break; 1761 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) { 1762 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 1763 tp->lost_out += tcp_skb_pcount(skb); 1764 1765 /* clear xmit_retransmit_queue hints 1766 * if this is beyond hint */ 1767 if (tp->retransmit_skb_hint != NULL && 1768 before(TCP_SKB_CB(skb)->seq, 1769 TCP_SKB_CB(tp->retransmit_skb_hint)->seq)) 1770 tp->retransmit_skb_hint = NULL; 1771 1772 } 1773 } 1774 tcp_sync_left_out(tp); 1775} 1776 1777/* Account newly detected lost packet(s) */ 1778 1779static void tcp_update_scoreboard(struct sock *sk) 1780{ 1781 struct tcp_sock *tp = tcp_sk(sk); 1782 1783 if (IsFack(tp)) { 1784 int lost = tp->fackets_out - tp->reordering; 1785 if (lost <= 0) 1786 lost = 1; 1787 tcp_mark_head_lost(sk, lost, tp->high_seq); 1788 } else { 1789 tcp_mark_head_lost(sk, 1, tp->high_seq); 1790 } 1791 1792 /* New heuristics: it is possible only after we switched 1793 * to restart timer each time when something is ACKed. 1794 * Hence, we can detect timed out packets during fast 1795 * retransmit without falling to slow start. 1796 */ 1797 if (!IsReno(tp) && tcp_head_timedout(sk)) { 1798 struct sk_buff *skb; 1799 1800 skb = tp->scoreboard_skb_hint ? tp->scoreboard_skb_hint 1801 : tcp_write_queue_head(sk); 1802 1803 tcp_for_write_queue_from(skb, sk) { 1804 if (skb == tcp_send_head(sk)) 1805 break; 1806 if (!tcp_skb_timedout(sk, skb)) 1807 break; 1808 1809 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) { 1810 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 1811 tp->lost_out += tcp_skb_pcount(skb); 1812 1813 /* clear xmit_retrans hint */ 1814 if (tp->retransmit_skb_hint && 1815 before(TCP_SKB_CB(skb)->seq, 1816 TCP_SKB_CB(tp->retransmit_skb_hint)->seq)) 1817 1818 tp->retransmit_skb_hint = NULL; 1819 } 1820 } 1821 1822 tp->scoreboard_skb_hint = skb; 1823 1824 tcp_sync_left_out(tp); 1825 } 1826} 1827 1828/* CWND moderation, preventing bursts due to too big ACKs 1829 * in dubious situations. 1830 */ 1831static inline void tcp_moderate_cwnd(struct tcp_sock *tp) 1832{ 1833 tp->snd_cwnd = min(tp->snd_cwnd, 1834 tcp_packets_in_flight(tp)+tcp_max_burst(tp)); 1835 tp->snd_cwnd_stamp = tcp_time_stamp; 1836} 1837 1838/* Lower bound on congestion window is slow start threshold 1839 * unless congestion avoidance choice decides to overide it. 1840 */ 1841static inline u32 tcp_cwnd_min(const struct sock *sk) 1842{ 1843 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 1844 1845 return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh; 1846} 1847 1848/* Decrease cwnd each second ack. */ 1849static void tcp_cwnd_down(struct sock *sk) 1850{ 1851 struct tcp_sock *tp = tcp_sk(sk); 1852 int decr = tp->snd_cwnd_cnt + 1; 1853 1854 tp->snd_cwnd_cnt = decr&1; 1855 decr >>= 1; 1856 1857 if (decr && tp->snd_cwnd > tcp_cwnd_min(sk)) 1858 tp->snd_cwnd -= decr; 1859 1860 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp)+1); 1861 tp->snd_cwnd_stamp = tcp_time_stamp; 1862} 1863 1864/* Nothing was retransmitted or returned timestamp is less 1865 * than timestamp of the first retransmission. 1866 */ 1867static inline int tcp_packet_delayed(struct tcp_sock *tp) 1868{ 1869 return !tp->retrans_stamp || 1870 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 1871 (__s32)(tp->rx_opt.rcv_tsecr - tp->retrans_stamp) < 0); 1872} 1873 1874/* Undo procedures. */ 1875 1876#if FASTRETRANS_DEBUG > 1 1877static void DBGUNDO(struct sock *sk, const char *msg) 1878{ 1879 struct tcp_sock *tp = tcp_sk(sk); 1880 struct inet_sock *inet = inet_sk(sk); 1881 1882 printk(KERN_DEBUG "Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n", 1883 msg, 1884 NIPQUAD(inet->daddr), ntohs(inet->dport), 1885 tp->snd_cwnd, tp->left_out, 1886 tp->snd_ssthresh, tp->prior_ssthresh, 1887 tp->packets_out); 1888} 1889#else 1890#define DBGUNDO(x...) do { } while (0) 1891#endif 1892 1893static void tcp_undo_cwr(struct sock *sk, const int undo) 1894{ 1895 struct tcp_sock *tp = tcp_sk(sk); 1896 1897 if (tp->prior_ssthresh) { 1898 const struct inet_connection_sock *icsk = inet_csk(sk); 1899 1900 if (icsk->icsk_ca_ops->undo_cwnd) 1901 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk); 1902 else 1903 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh<<1); 1904 1905 if (undo && tp->prior_ssthresh > tp->snd_ssthresh) { 1906 tp->snd_ssthresh = tp->prior_ssthresh; 1907 TCP_ECN_withdraw_cwr(tp); 1908 } 1909 } else { 1910 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh); 1911 } 1912 tcp_moderate_cwnd(tp); 1913 tp->snd_cwnd_stamp = tcp_time_stamp; 1914 1915 /* There is something screwy going on with the retrans hints after 1916 an undo */ 1917 clear_all_retrans_hints(tp); 1918} 1919 1920static inline int tcp_may_undo(struct tcp_sock *tp) 1921{ 1922 return tp->undo_marker && 1923 (!tp->undo_retrans || tcp_packet_delayed(tp)); 1924} 1925 1926/* People celebrate: "We love our President!" */ 1927static int tcp_try_undo_recovery(struct sock *sk) 1928{ 1929 struct tcp_sock *tp = tcp_sk(sk); 1930 1931 if (tcp_may_undo(tp)) { 1932 /* Happy end! We did not retransmit anything 1933 * or our original transmission succeeded. 1934 */ 1935 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans"); 1936 tcp_undo_cwr(sk, 1); 1937 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss) 1938 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO); 1939 else 1940 NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO); 1941 tp->undo_marker = 0; 1942 } 1943 if (tp->snd_una == tp->high_seq && IsReno(tp)) { 1944 /* Hold old state until something *above* high_seq 1945 * is ACKed. For Reno it is MUST to prevent false 1946 * fast retransmits (RFC2582). SACK TCP is safe. */ 1947 tcp_moderate_cwnd(tp); 1948 return 1; 1949 } 1950 tcp_set_ca_state(sk, TCP_CA_Open); 1951 return 0; 1952} 1953 1954/* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */ 1955static void tcp_try_undo_dsack(struct sock *sk) 1956{ 1957 struct tcp_sock *tp = tcp_sk(sk); 1958 1959 if (tp->undo_marker && !tp->undo_retrans) { 1960 DBGUNDO(sk, "D-SACK"); 1961 tcp_undo_cwr(sk, 1); 1962 tp->undo_marker = 0; 1963 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO); 1964 } 1965} 1966 1967/* Undo during fast recovery after partial ACK. */ 1968 1969static int tcp_try_undo_partial(struct sock *sk, int acked) 1970{ 1971 struct tcp_sock *tp = tcp_sk(sk); 1972 /* Partial ACK arrived. Force Hoe's retransmit. */ 1973 int failed = IsReno(tp) || tp->fackets_out>tp->reordering; 1974 1975 if (tcp_may_undo(tp)) { 1976 /* Plain luck! Hole if filled with delayed 1977 * packet, rather than with a retransmit. 1978 */ 1979 if (tp->retrans_out == 0) 1980 tp->retrans_stamp = 0; 1981 1982 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1); 1983 1984 DBGUNDO(sk, "Hoe"); 1985 tcp_undo_cwr(sk, 0); 1986 NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO); 1987 1988 /* So... Do not make Hoe's retransmit yet. 1989 * If the first packet was delayed, the rest 1990 * ones are most probably delayed as well. 1991 */ 1992 failed = 0; 1993 } 1994 return failed; 1995} 1996 1997/* Undo during loss recovery after partial ACK. */ 1998static int tcp_try_undo_loss(struct sock *sk) 1999{ 2000 struct tcp_sock *tp = tcp_sk(sk); 2001 2002 if (tcp_may_undo(tp)) { 2003 struct sk_buff *skb; 2004 tcp_for_write_queue(skb, sk) { 2005 if (skb == tcp_send_head(sk)) 2006 break; 2007 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; 2008 } 2009 2010 clear_all_retrans_hints(tp); 2011 2012 DBGUNDO(sk, "partial loss"); 2013 tp->lost_out = 0; 2014 tp->left_out = tp->sacked_out; 2015 tcp_undo_cwr(sk, 1); 2016 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO); 2017 inet_csk(sk)->icsk_retransmits = 0; 2018 tp->undo_marker = 0; 2019 if (!IsReno(tp)) 2020 tcp_set_ca_state(sk, TCP_CA_Open); 2021 return 1; 2022 } 2023 return 0; 2024} 2025 2026static inline void tcp_complete_cwr(struct sock *sk) 2027{ 2028 struct tcp_sock *tp = tcp_sk(sk); 2029 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh); 2030 tp->snd_cwnd_stamp = tcp_time_stamp; 2031 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR); 2032} 2033 2034static void tcp_try_to_open(struct sock *sk, int flag) 2035{ 2036 struct tcp_sock *tp = tcp_sk(sk); 2037 2038 tp->left_out = tp->sacked_out; 2039 2040 if (tp->retrans_out == 0) 2041 tp->retrans_stamp = 0; 2042 2043 if (flag&FLAG_ECE) 2044 tcp_enter_cwr(sk, 1); 2045 2046 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) { 2047 int state = TCP_CA_Open; 2048 2049 if (tp->left_out || tp->retrans_out || tp->undo_marker) 2050 state = TCP_CA_Disorder; 2051 2052 if (inet_csk(sk)->icsk_ca_state != state) { 2053 tcp_set_ca_state(sk, state); 2054 tp->high_seq = tp->snd_nxt; 2055 } 2056 tcp_moderate_cwnd(tp); 2057 } else { 2058 tcp_cwnd_down(sk); 2059 } 2060} 2061 2062static void tcp_mtup_probe_failed(struct sock *sk) 2063{ 2064 struct inet_connection_sock *icsk = inet_csk(sk); 2065 2066 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1; 2067 icsk->icsk_mtup.probe_size = 0; 2068} 2069 2070static void tcp_mtup_probe_success(struct sock *sk, struct sk_buff *skb) 2071{ 2072 struct tcp_sock *tp = tcp_sk(sk); 2073 struct inet_connection_sock *icsk = inet_csk(sk); 2074 2075 /* FIXME: breaks with very large cwnd */ 2076 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2077 tp->snd_cwnd = tp->snd_cwnd * 2078 tcp_mss_to_mtu(sk, tp->mss_cache) / 2079 icsk->icsk_mtup.probe_size; 2080 tp->snd_cwnd_cnt = 0; 2081 tp->snd_cwnd_stamp = tcp_time_stamp; 2082 tp->rcv_ssthresh = tcp_current_ssthresh(sk); 2083 2084 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size; 2085 icsk->icsk_mtup.probe_size = 0; 2086 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 2087} 2088 2089 2090/* Process an event, which can update packets-in-flight not trivially. 2091 * Main goal of this function is to calculate new estimate for left_out, 2092 * taking into account both packets sitting in receiver's buffer and 2093 * packets lost by network. 2094 * 2095 * Besides that it does CWND reduction, when packet loss is detected 2096 * and changes state of machine. 2097 * 2098 * It does _not_ decide what to send, it is made in function 2099 * tcp_xmit_retransmit_queue(). 2100 */ 2101static void 2102tcp_fastretrans_alert(struct sock *sk, u32 prior_snd_una, 2103 int prior_packets, int flag) 2104{ 2105 struct inet_connection_sock *icsk = inet_csk(sk); 2106 struct tcp_sock *tp = tcp_sk(sk); 2107 int is_dupack = (tp->snd_una == prior_snd_una && !(flag&FLAG_NOT_DUP)); 2108 2109 /* Some technical things: 2110 * 1. Reno does not count dupacks (sacked_out) automatically. */ 2111 if (!tp->packets_out) 2112 tp->sacked_out = 0; 2113 /* 2. SACK counts snd_fack in packets inaccurately. */ 2114 if (tp->sacked_out == 0) 2115 tp->fackets_out = 0; 2116 2117 /* Now state machine starts. 2118 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */ 2119 if (flag&FLAG_ECE) 2120 tp->prior_ssthresh = 0; 2121 2122 /* B. In all the states check for reneging SACKs. */ 2123 if (tp->sacked_out && tcp_check_sack_reneging(sk)) 2124 return; 2125 2126 /* C. Process data loss notification, provided it is valid. */ 2127 if ((flag&FLAG_DATA_LOST) && 2128 before(tp->snd_una, tp->high_seq) && 2129 icsk->icsk_ca_state != TCP_CA_Open && 2130 tp->fackets_out > tp->reordering) { 2131 tcp_mark_head_lost(sk, tp->fackets_out-tp->reordering, tp->high_seq); 2132 NET_INC_STATS_BH(LINUX_MIB_TCPLOSS); 2133 } 2134 2135 /* D. Synchronize left_out to current state. */ 2136 tcp_sync_left_out(tp); 2137 2138 /* E. Check state exit conditions. State can be terminated 2139 * when high_seq is ACKed. */ 2140 if (icsk->icsk_ca_state == TCP_CA_Open) { 2141 BUG_TRAP(tp->retrans_out == 0); 2142 tp->retrans_stamp = 0; 2143 } else if (!before(tp->snd_una, tp->high_seq)) { 2144 switch (icsk->icsk_ca_state) { 2145 case TCP_CA_Loss: 2146 icsk->icsk_retransmits = 0; 2147 if (tcp_try_undo_recovery(sk)) 2148 return; 2149 break; 2150 2151 case TCP_CA_CWR: 2152 /* CWR is to be held something *above* high_seq 2153 * is ACKed for CWR bit to reach receiver. */ 2154 if (tp->snd_una != tp->high_seq) { 2155 tcp_complete_cwr(sk); 2156 tcp_set_ca_state(sk, TCP_CA_Open); 2157 } 2158 break; 2159 2160 case TCP_CA_Disorder: 2161 tcp_try_undo_dsack(sk); 2162 if (!tp->undo_marker || 2163 /* For SACK case do not Open to allow to undo 2164 * catching for all duplicate ACKs. */ 2165 IsReno(tp) || tp->snd_una != tp->high_seq) { 2166 tp->undo_marker = 0; 2167 tcp_set_ca_state(sk, TCP_CA_Open); 2168 } 2169 break; 2170 2171 case TCP_CA_Recovery: 2172 if (IsReno(tp)) 2173 tcp_reset_reno_sack(tp); 2174 if (tcp_try_undo_recovery(sk)) 2175 return; 2176 tcp_complete_cwr(sk); 2177 break; 2178 } 2179 } 2180 2181 /* F. Process state. */ 2182 switch (icsk->icsk_ca_state) { 2183 case TCP_CA_Recovery: 2184 if (prior_snd_una == tp->snd_una) { 2185 if (IsReno(tp) && is_dupack) 2186 tcp_add_reno_sack(sk); 2187 } else { 2188 int acked = prior_packets - tp->packets_out; 2189 if (IsReno(tp)) 2190 tcp_remove_reno_sacks(sk, acked); 2191 is_dupack = tcp_try_undo_partial(sk, acked); 2192 } 2193 break; 2194 case TCP_CA_Loss: 2195 if (flag&FLAG_DATA_ACKED) 2196 icsk->icsk_retransmits = 0; 2197 if (!tcp_try_undo_loss(sk)) { 2198 tcp_moderate_cwnd(tp); 2199 tcp_xmit_retransmit_queue(sk); 2200 return; 2201 } 2202 if (icsk->icsk_ca_state != TCP_CA_Open) 2203 return; 2204 /* Loss is undone; fall through to processing in Open state. */ 2205 default: 2206 if (IsReno(tp)) { 2207 if (tp->snd_una != prior_snd_una) 2208 tcp_reset_reno_sack(tp); 2209 if (is_dupack) 2210 tcp_add_reno_sack(sk); 2211 } 2212 2213 if (icsk->icsk_ca_state == TCP_CA_Disorder) 2214 tcp_try_undo_dsack(sk); 2215 2216 if (!tcp_time_to_recover(sk)) { 2217 tcp_try_to_open(sk, flag); 2218 return; 2219 } 2220 2221 /* MTU probe failure: don't reduce cwnd */ 2222 if (icsk->icsk_ca_state < TCP_CA_CWR && 2223 icsk->icsk_mtup.probe_size && 2224 tp->snd_una == tp->mtu_probe.probe_seq_start) { 2225 tcp_mtup_probe_failed(sk); 2226 /* Restores the reduction we did in tcp_mtup_probe() */ 2227 tp->snd_cwnd++; 2228 tcp_simple_retransmit(sk); 2229 return; 2230 } 2231 2232 /* Otherwise enter Recovery state */ 2233 2234 if (IsReno(tp)) 2235 NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY); 2236 else 2237 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY); 2238 2239 tp->high_seq = tp->snd_nxt; 2240 tp->prior_ssthresh = 0; 2241 tp->undo_marker = tp->snd_una; 2242 tp->undo_retrans = tp->retrans_out; 2243 2244 if (icsk->icsk_ca_state < TCP_CA_CWR) { 2245 if (!(flag&FLAG_ECE)) 2246 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2247 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 2248 TCP_ECN_queue_cwr(tp); 2249 } 2250 2251 tp->bytes_acked = 0; 2252 tp->snd_cwnd_cnt = 0; 2253 tcp_set_ca_state(sk, TCP_CA_Recovery); 2254 } 2255 2256 if (is_dupack || tcp_head_timedout(sk)) 2257 tcp_update_scoreboard(sk); 2258 tcp_cwnd_down(sk); 2259 tcp_xmit_retransmit_queue(sk); 2260} 2261 2262/* Read draft-ietf-tcplw-high-performance before mucking 2263 * with this code. (Supersedes RFC1323) 2264 */ 2265static void tcp_ack_saw_tstamp(struct sock *sk, int flag) 2266{ 2267 /* RTTM Rule: A TSecr value received in a segment is used to 2268 * update the averaged RTT measurement only if the segment 2269 * acknowledges some new data, i.e., only if it advances the 2270 * left edge of the send window. 2271 * 2272 * See draft-ietf-tcplw-high-performance-00, section 3.3. 2273 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru> 2274 * 2275 * Changed: reset backoff as soon as we see the first valid sample. 2276 * If we do not, we get strongly overestimated rto. With timestamps 2277 * samples are accepted even from very old segments: f.e., when rtt=1 2278 * increases to 8, we retransmit 5 times and after 8 seconds delayed 2279 * answer arrives rto becomes 120 seconds! If at least one of segments 2280 * in window is lost... Voila. --ANK (010210) 2281 */ 2282 struct tcp_sock *tp = tcp_sk(sk); 2283 const __u32 seq_rtt = tcp_time_stamp - tp->rx_opt.rcv_tsecr; 2284 tcp_rtt_estimator(sk, seq_rtt); 2285 tcp_set_rto(sk); 2286 inet_csk(sk)->icsk_backoff = 0; 2287 tcp_bound_rto(sk); 2288} 2289 2290static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag) 2291{ 2292 /* We don't have a timestamp. Can only use 2293 * packets that are not retransmitted to determine 2294 * rtt estimates. Also, we must not reset the 2295 * backoff for rto until we get a non-retransmitted 2296 * packet. This allows us to deal with a situation 2297 * where the network delay has increased suddenly. 2298 * I.e. Karn's algorithm. (SIGCOMM '87, p5.) 2299 */ 2300 2301 if (flag & FLAG_RETRANS_DATA_ACKED) 2302 return; 2303 2304 tcp_rtt_estimator(sk, seq_rtt); 2305 tcp_set_rto(sk); 2306 inet_csk(sk)->icsk_backoff = 0; 2307 tcp_bound_rto(sk); 2308} 2309 2310static inline void tcp_ack_update_rtt(struct sock *sk, const int flag, 2311 const s32 seq_rtt) 2312{ 2313 const struct tcp_sock *tp = tcp_sk(sk); 2314 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */ 2315 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr) 2316 tcp_ack_saw_tstamp(sk, flag); 2317 else if (seq_rtt >= 0) 2318 tcp_ack_no_tstamp(sk, seq_rtt, flag); 2319} 2320 2321static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 rtt, 2322 u32 in_flight, int good) 2323{ 2324 const struct inet_connection_sock *icsk = inet_csk(sk); 2325 icsk->icsk_ca_ops->cong_avoid(sk, ack, rtt, in_flight, good); 2326 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp; 2327} 2328 2329/* Restart timer after forward progress on connection. 2330 * RFC2988 recommends to restart timer to now+rto. 2331 */ 2332 2333static void tcp_ack_packets_out(struct sock *sk) 2334{ 2335 struct tcp_sock *tp = tcp_sk(sk); 2336 2337 if (!tp->packets_out) { 2338 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS); 2339 } else { 2340 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 2341 } 2342} 2343 2344static int tcp_tso_acked(struct sock *sk, struct sk_buff *skb, 2345 __u32 now, __s32 *seq_rtt) 2346{ 2347 struct tcp_sock *tp = tcp_sk(sk); 2348 struct tcp_skb_cb *scb = TCP_SKB_CB(skb); 2349 __u32 seq = tp->snd_una; 2350 __u32 packets_acked; 2351 int acked = 0; 2352 2353 /* If we get here, the whole TSO packet has not been 2354 * acked. 2355 */ 2356 BUG_ON(!after(scb->end_seq, seq)); 2357 2358 packets_acked = tcp_skb_pcount(skb); 2359 if (tcp_trim_head(sk, skb, seq - scb->seq)) 2360 return 0; 2361 packets_acked -= tcp_skb_pcount(skb); 2362 2363 if (packets_acked) { 2364 __u8 sacked = scb->sacked; 2365 2366 acked |= FLAG_DATA_ACKED; 2367 if (sacked) { 2368 if (sacked & TCPCB_RETRANS) { 2369 if (sacked & TCPCB_SACKED_RETRANS) 2370 tp->retrans_out -= packets_acked; 2371 acked |= FLAG_RETRANS_DATA_ACKED; 2372 *seq_rtt = -1; 2373 } else if (*seq_rtt < 0) 2374 *seq_rtt = now - scb->when; 2375 if (sacked & TCPCB_SACKED_ACKED) 2376 tp->sacked_out -= packets_acked; 2377 if (sacked & TCPCB_LOST) 2378 tp->lost_out -= packets_acked; 2379 if (sacked & TCPCB_URG) { 2380 if (tp->urg_mode && 2381 !before(seq, tp->snd_up)) 2382 tp->urg_mode = 0; 2383 } 2384 } else if (*seq_rtt < 0) 2385 *seq_rtt = now - scb->when; 2386 2387 if (tp->fackets_out) { 2388 __u32 dval = min(tp->fackets_out, packets_acked); 2389 tp->fackets_out -= dval; 2390 } 2391 tp->packets_out -= packets_acked; 2392 2393 BUG_ON(tcp_skb_pcount(skb) == 0); 2394 BUG_ON(!before(scb->seq, scb->end_seq)); 2395 } 2396 2397 return acked; 2398} 2399 2400/* Remove acknowledged frames from the retransmission queue. */ 2401static int tcp_clean_rtx_queue(struct sock *sk, __s32 *seq_rtt_p) 2402{ 2403 struct tcp_sock *tp = tcp_sk(sk); 2404 const struct inet_connection_sock *icsk = inet_csk(sk); 2405 struct sk_buff *skb; 2406 __u32 now = tcp_time_stamp; 2407 int acked = 0; 2408 __s32 seq_rtt = -1; 2409 u32 pkts_acked = 0; 2410 ktime_t last_ackt = ktime_set(0,0); 2411 2412 while ((skb = tcp_write_queue_head(sk)) && 2413 skb != tcp_send_head(sk)) { 2414 struct tcp_skb_cb *scb = TCP_SKB_CB(skb); 2415 __u8 sacked = scb->sacked; 2416 2417 /* If our packet is before the ack sequence we can 2418 * discard it as it's confirmed to have arrived at 2419 * the other end. 2420 */ 2421 if (after(scb->end_seq, tp->snd_una)) { 2422 if (tcp_skb_pcount(skb) > 1 && 2423 after(tp->snd_una, scb->seq)) 2424 acked |= tcp_tso_acked(sk, skb, 2425 now, &seq_rtt); 2426 break; 2427 } 2428 2429 /* Initial outgoing SYN's get put onto the write_queue 2430 * just like anything else we transmit. It is not 2431 * true data, and if we misinform our callers that 2432 * this ACK acks real data, we will erroneously exit 2433 * connection startup slow start one packet too 2434 * quickly. This is severely frowned upon behavior. 2435 */ 2436 if (!(scb->flags & TCPCB_FLAG_SYN)) { 2437 acked |= FLAG_DATA_ACKED; 2438 ++pkts_acked; 2439 } else { 2440 acked |= FLAG_SYN_ACKED; 2441 tp->retrans_stamp = 0; 2442 } 2443 2444 /* MTU probing checks */ 2445 if (icsk->icsk_mtup.probe_size) { 2446 if (!after(tp->mtu_probe.probe_seq_end, TCP_SKB_CB(skb)->end_seq)) { 2447 tcp_mtup_probe_success(sk, skb); 2448 } 2449 } 2450 2451 if (sacked) { 2452 if (sacked & TCPCB_RETRANS) { 2453 if (sacked & TCPCB_SACKED_RETRANS) 2454 tp->retrans_out -= tcp_skb_pcount(skb); 2455 acked |= FLAG_RETRANS_DATA_ACKED; 2456 seq_rtt = -1; 2457 } else if (seq_rtt < 0) { 2458 seq_rtt = now - scb->when; 2459 last_ackt = skb->tstamp; 2460 } 2461 if (sacked & TCPCB_SACKED_ACKED) 2462 tp->sacked_out -= tcp_skb_pcount(skb); 2463 if (sacked & TCPCB_LOST) 2464 tp->lost_out -= tcp_skb_pcount(skb); 2465 if (sacked & TCPCB_URG) { 2466 if (tp->urg_mode && 2467 !before(scb->end_seq, tp->snd_up)) 2468 tp->urg_mode = 0; 2469 } 2470 } else if (seq_rtt < 0) { 2471 seq_rtt = now - scb->when; 2472 last_ackt = skb->tstamp; 2473 } 2474 tcp_dec_pcount_approx(&tp->fackets_out, skb); 2475 tcp_packets_out_dec(tp, skb); 2476 tcp_unlink_write_queue(skb, sk); 2477 sk_stream_free_skb(sk, skb); 2478 clear_all_retrans_hints(tp); 2479 } 2480 2481 if (acked&FLAG_ACKED) { 2482 const struct tcp_congestion_ops *ca_ops 2483 = inet_csk(sk)->icsk_ca_ops; 2484 2485 tcp_ack_update_rtt(sk, acked, seq_rtt); 2486 tcp_ack_packets_out(sk); 2487 2488 if (ca_ops->pkts_acked) 2489 ca_ops->pkts_acked(sk, pkts_acked, last_ackt); 2490 } 2491 2492#if FASTRETRANS_DEBUG > 0 2493 BUG_TRAP((int)tp->sacked_out >= 0); 2494 BUG_TRAP((int)tp->lost_out >= 0); 2495 BUG_TRAP((int)tp->retrans_out >= 0); 2496 if (!tp->packets_out && tp->rx_opt.sack_ok) { 2497 const struct inet_connection_sock *icsk = inet_csk(sk); 2498 if (tp->lost_out) { 2499 printk(KERN_DEBUG "Leak l=%u %d\n", 2500 tp->lost_out, icsk->icsk_ca_state); 2501 tp->lost_out = 0; 2502 } 2503 if (tp->sacked_out) { 2504 printk(KERN_DEBUG "Leak s=%u %d\n", 2505 tp->sacked_out, icsk->icsk_ca_state); 2506 tp->sacked_out = 0; 2507 } 2508 if (tp->retrans_out) { 2509 printk(KERN_DEBUG "Leak r=%u %d\n", 2510 tp->retrans_out, icsk->icsk_ca_state); 2511 tp->retrans_out = 0; 2512 } 2513 } 2514#endif 2515 *seq_rtt_p = seq_rtt; 2516 return acked; 2517} 2518 2519static void tcp_ack_probe(struct sock *sk) 2520{ 2521 const struct tcp_sock *tp = tcp_sk(sk); 2522 struct inet_connection_sock *icsk = inet_csk(sk); 2523 2524 /* Was it a usable window open? */ 2525 2526 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, 2527 tp->snd_una + tp->snd_wnd)) { 2528 icsk->icsk_backoff = 0; 2529 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0); 2530 /* Socket must be waked up by subsequent tcp_data_snd_check(). 2531 * This function is not for random using! 2532 */ 2533 } else { 2534 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 2535 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), 2536 TCP_RTO_MAX); 2537 } 2538} 2539 2540static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag) 2541{ 2542 return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) || 2543 inet_csk(sk)->icsk_ca_state != TCP_CA_Open); 2544} 2545 2546static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag) 2547{ 2548 const struct tcp_sock *tp = tcp_sk(sk); 2549 return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) && 2550 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR)); 2551} 2552 2553/* Check that window update is acceptable. 2554 * The function assumes that snd_una<=ack<=snd_next. 2555 */ 2556static inline int tcp_may_update_window(const struct tcp_sock *tp, const u32 ack, 2557 const u32 ack_seq, const u32 nwin) 2558{ 2559 return (after(ack, tp->snd_una) || 2560 after(ack_seq, tp->snd_wl1) || 2561 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd)); 2562} 2563 2564/* Update our send window. 2565 * 2566 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2 2567 * and in FreeBSD. NetBSD's one is even worse.) is wrong. 2568 */ 2569static int tcp_ack_update_window(struct sock *sk, struct sk_buff *skb, u32 ack, 2570 u32 ack_seq) 2571{ 2572 struct tcp_sock *tp = tcp_sk(sk); 2573 int flag = 0; 2574 u32 nwin = ntohs(tcp_hdr(skb)->window); 2575 2576 if (likely(!tcp_hdr(skb)->syn)) 2577 nwin <<= tp->rx_opt.snd_wscale; 2578 2579 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) { 2580 flag |= FLAG_WIN_UPDATE; 2581 tcp_update_wl(tp, ack, ack_seq); 2582 2583 if (tp->snd_wnd != nwin) { 2584 tp->snd_wnd = nwin; 2585 2586 /* Note, it is the only place, where 2587 * fast path is recovered for sending TCP. 2588 */ 2589 tp->pred_flags = 0; 2590 tcp_fast_path_check(sk); 2591 2592 if (nwin > tp->max_window) { 2593 tp->max_window = nwin; 2594 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie); 2595 } 2596 } 2597 } 2598 2599 tp->snd_una = ack; 2600 2601 return flag; 2602} 2603 2604/* A very conservative spurious RTO response algorithm: reduce cwnd and 2605 * continue in congestion avoidance. 2606 */ 2607static void tcp_conservative_spur_to_response(struct tcp_sock *tp) 2608{ 2609 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh); 2610 tp->snd_cwnd_cnt = 0; 2611 tcp_moderate_cwnd(tp); 2612} 2613 2614/* A conservative spurious RTO response algorithm: reduce cwnd using 2615 * rate halving and continue in congestion avoidance. 2616 */ 2617static void tcp_ratehalving_spur_to_response(struct sock *sk) 2618{ 2619 tcp_enter_cwr(sk, 0); 2620} 2621 2622static void tcp_undo_spur_to_response(struct sock *sk, int flag) 2623{ 2624 if (flag&FLAG_ECE) 2625 tcp_ratehalving_spur_to_response(sk); 2626 else 2627 tcp_undo_cwr(sk, 1); 2628} 2629 2630/* F-RTO spurious RTO detection algorithm (RFC4138) 2631 * 2632 * F-RTO affects during two new ACKs following RTO (well, almost, see inline 2633 * comments). State (ACK number) is kept in frto_counter. When ACK advances 2634 * window (but not to or beyond highest sequence sent before RTO): 2635 * On First ACK, send two new segments out. 2636 * On Second ACK, RTO was likely spurious. Do spurious response (response 2637 * algorithm is not part of the F-RTO detection algorithm 2638 * given in RFC4138 but can be selected separately). 2639 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss 2640 * and TCP falls back to conventional RTO recovery. 2641 * 2642 * Rationale: if the RTO was spurious, new ACKs should arrive from the 2643 * original window even after we transmit two new data segments. 2644 * 2645 * SACK version: 2646 * on first step, wait until first cumulative ACK arrives, then move to 2647 * the second step. In second step, the next ACK decides. 2648 * 2649 * F-RTO is implemented (mainly) in four functions: 2650 * - tcp_use_frto() is used to determine if TCP is can use F-RTO 2651 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is 2652 * called when tcp_use_frto() showed green light 2653 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm 2654 * - tcp_enter_frto_loss() is called if there is not enough evidence 2655 * to prove that the RTO is indeed spurious. It transfers the control 2656 * from F-RTO to the conventional RTO recovery 2657 */ 2658static int tcp_process_frto(struct sock *sk, u32 prior_snd_una, int flag) 2659{ 2660 struct tcp_sock *tp = tcp_sk(sk); 2661 2662 tcp_sync_left_out(tp); 2663 2664 /* Duplicate the behavior from Loss state (fastretrans_alert) */ 2665 if (flag&FLAG_DATA_ACKED) 2666 inet_csk(sk)->icsk_retransmits = 0; 2667 2668 if (!before(tp->snd_una, tp->frto_highmark)) { 2669 tcp_enter_frto_loss(sk, tp->frto_counter + 1, flag); 2670 return 1; 2671 } 2672 2673 if (!IsSackFrto() || IsReno(tp)) { 2674 /* RFC4138 shortcoming in step 2; should also have case c): 2675 * ACK isn't duplicate nor advances window, e.g., opposite dir 2676 * data, winupdate 2677 */ 2678 if ((tp->snd_una == prior_snd_una) && (flag&FLAG_NOT_DUP) && 2679 !(flag&FLAG_FORWARD_PROGRESS)) 2680 return 1; 2681 2682 if (!(flag&FLAG_DATA_ACKED)) { 2683 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3), 2684 flag); 2685 return 1; 2686 } 2687 } else { 2688 if (!(flag&FLAG_DATA_ACKED) && (tp->frto_counter == 1)) { 2689 /* Prevent sending of new data. */ 2690 tp->snd_cwnd = min(tp->snd_cwnd, 2691 tcp_packets_in_flight(tp)); 2692 return 1; 2693 } 2694 2695 if ((tp->frto_counter == 2) && 2696 (!(flag&FLAG_FORWARD_PROGRESS) || 2697 ((flag&FLAG_DATA_SACKED) && !(flag&FLAG_ONLY_ORIG_SACKED)))) { 2698 /* RFC4138 shortcoming (see comment above) */ 2699 if (!(flag&FLAG_FORWARD_PROGRESS) && (flag&FLAG_NOT_DUP)) 2700 return 1; 2701 2702 tcp_enter_frto_loss(sk, 3, flag); 2703 return 1; 2704 } 2705 } 2706 2707 if (tp->frto_counter == 1) { 2708 /* Sending of the next skb must be allowed or no FRTO */ 2709 if (!tcp_send_head(sk) || 2710 after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, 2711 tp->snd_una + tp->snd_wnd)) { 2712 tcp_enter_frto_loss(sk, tp->frto_counter + 1, flag); 2713 return 1; 2714 } 2715 2716 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2; 2717 tp->frto_counter = 2; 2718 return 1; 2719 } else /* frto_counter == 2 */ { 2720 switch (sysctl_tcp_frto_response) { 2721 case 2: 2722 tcp_undo_spur_to_response(sk, flag); 2723 break; 2724 case 1: 2725 tcp_conservative_spur_to_response(tp); 2726 break; 2727 default: 2728 tcp_ratehalving_spur_to_response(sk); 2729 break; 2730 } 2731 tp->frto_counter = 0; 2732 } 2733 return 0; 2734} 2735 2736/* This routine deals with incoming acks, but not outgoing ones. */ 2737static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag) 2738{ 2739 struct inet_connection_sock *icsk = inet_csk(sk); 2740 struct tcp_sock *tp = tcp_sk(sk); 2741 u32 prior_snd_una = tp->snd_una; 2742 u32 ack_seq = TCP_SKB_CB(skb)->seq; 2743 u32 ack = TCP_SKB_CB(skb)->ack_seq; 2744 u32 prior_in_flight; 2745 s32 seq_rtt; 2746 int prior_packets; 2747 int frto_cwnd = 0; 2748 2749 /* If the ack is newer than sent or older than previous acks 2750 * then we can probably ignore it. 2751 */ 2752 if (after(ack, tp->snd_nxt)) 2753 goto uninteresting_ack; 2754 2755 if (before(ack, prior_snd_una)) 2756 goto old_ack; 2757 2758 if (sysctl_tcp_abc) { 2759 if (icsk->icsk_ca_state < TCP_CA_CWR) 2760 tp->bytes_acked += ack - prior_snd_una; 2761 else if (icsk->icsk_ca_state == TCP_CA_Loss) 2762 /* we assume just one segment left network */ 2763 tp->bytes_acked += min(ack - prior_snd_una, tp->mss_cache); 2764 } 2765 2766 if (!(flag&FLAG_SLOWPATH) && after(ack, prior_snd_una)) { 2767 /* Window is constant, pure forward advance. 2768 * No more checks are required. 2769 * Note, we use the fact that SND.UNA>=SND.WL2. 2770 */ 2771 tcp_update_wl(tp, ack, ack_seq); 2772 tp->snd_una = ack; 2773 flag |= FLAG_WIN_UPDATE; 2774 2775 tcp_ca_event(sk, CA_EVENT_FAST_ACK); 2776 2777 NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS); 2778 } else { 2779 if (ack_seq != TCP_SKB_CB(skb)->end_seq) 2780 flag |= FLAG_DATA; 2781 else 2782 NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS); 2783 2784 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq); 2785 2786 if (TCP_SKB_CB(skb)->sacked) 2787 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una); 2788 2789 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb))) 2790 flag |= FLAG_ECE; 2791 2792 tcp_ca_event(sk, CA_EVENT_SLOW_ACK); 2793 } 2794 2795 /* We passed data and got it acked, remove any soft error 2796 * log. Something worked... 2797 */ 2798 sk->sk_err_soft = 0; 2799 tp->rcv_tstamp = tcp_time_stamp; 2800 prior_packets = tp->packets_out; 2801 if (!prior_packets) 2802 goto no_queue; 2803 2804 prior_in_flight = tcp_packets_in_flight(tp); 2805 2806 /* See if we can take anything off of the retransmit queue. */ 2807 flag |= tcp_clean_rtx_queue(sk, &seq_rtt); 2808 2809 if (tp->frto_counter) 2810 frto_cwnd = tcp_process_frto(sk, prior_snd_una, flag); 2811 2812 if (tcp_ack_is_dubious(sk, flag)) { 2813 /* Advance CWND, if state allows this. */ 2814 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd && 2815 tcp_may_raise_cwnd(sk, flag)) 2816 tcp_cong_avoid(sk, ack, seq_rtt, prior_in_flight, 0); 2817 tcp_fastretrans_alert(sk, prior_snd_una, prior_packets, flag); 2818 } else { 2819 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd) 2820 tcp_cong_avoid(sk, ack, seq_rtt, prior_in_flight, 1); 2821 } 2822 2823 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag&FLAG_NOT_DUP)) 2824 dst_confirm(sk->sk_dst_cache); 2825 2826 return 1; 2827 2828no_queue: 2829 icsk->icsk_probes_out = 0; 2830 2831 /* If this ack opens up a zero window, clear backoff. It was 2832 * being used to time the probes, and is probably far higher than 2833 * it needs to be for normal retransmission. 2834 */ 2835 if (tcp_send_head(sk)) 2836 tcp_ack_probe(sk); 2837 return 1; 2838 2839old_ack: 2840 if (TCP_SKB_CB(skb)->sacked) 2841 tcp_sacktag_write_queue(sk, skb, prior_snd_una); 2842 2843uninteresting_ack: 2844 SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt); 2845 return 0; 2846} 2847 2848 2849/* Look for tcp options. Normally only called on SYN and SYNACK packets. 2850 * But, this can also be called on packets in the established flow when 2851 * the fast version below fails. 2852 */ 2853void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx, int estab) 2854{ 2855 unsigned char *ptr; 2856 struct tcphdr *th = tcp_hdr(skb); 2857 int length=(th->doff*4)-sizeof(struct tcphdr); 2858 2859 ptr = (unsigned char *)(th + 1); 2860 opt_rx->saw_tstamp = 0; 2861 2862 while (length > 0) { 2863 int opcode=*ptr++; 2864 int opsize; 2865 2866 switch (opcode) { 2867 case TCPOPT_EOL: 2868 return; 2869 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */ 2870 length--; 2871 continue; 2872 default: 2873 opsize=*ptr++; 2874 if (opsize < 2) /* "silly options" */ 2875 return; 2876 if (opsize > length) 2877 return; /* don't parse partial options */ 2878 switch (opcode) { 2879 case TCPOPT_MSS: 2880 if (opsize==TCPOLEN_MSS && th->syn && !estab) { 2881 u16 in_mss = ntohs(get_unaligned((__be16 *)ptr)); 2882 if (in_mss) { 2883 if (opt_rx->user_mss && opt_rx->user_mss < in_mss) 2884 in_mss = opt_rx->user_mss; 2885 opt_rx->mss_clamp = in_mss; 2886 } 2887 } 2888 break; 2889 case TCPOPT_WINDOW: 2890 if (opsize==TCPOLEN_WINDOW && th->syn && !estab) 2891 if (sysctl_tcp_window_scaling) { 2892 __u8 snd_wscale = *(__u8 *) ptr; 2893 opt_rx->wscale_ok = 1; 2894 if (snd_wscale > 14) { 2895 if (net_ratelimit()) 2896 printk(KERN_INFO "tcp_parse_options: Illegal window " 2897 "scaling value %d >14 received.\n", 2898 snd_wscale); 2899 snd_wscale = 14; 2900 } 2901 opt_rx->snd_wscale = snd_wscale; 2902 } 2903 break; 2904 case TCPOPT_TIMESTAMP: 2905 if (opsize==TCPOLEN_TIMESTAMP) { 2906 if ((estab && opt_rx->tstamp_ok) || 2907 (!estab && sysctl_tcp_timestamps)) { 2908 opt_rx->saw_tstamp = 1; 2909 opt_rx->rcv_tsval = ntohl(get_unaligned((__be32 *)ptr)); 2910 opt_rx->rcv_tsecr = ntohl(get_unaligned((__be32 *)(ptr+4))); 2911 } 2912 } 2913 break; 2914 case TCPOPT_SACK_PERM: 2915 if (opsize==TCPOLEN_SACK_PERM && th->syn && !estab) { 2916 if (sysctl_tcp_sack) { 2917 opt_rx->sack_ok = 1; 2918 tcp_sack_reset(opt_rx); 2919 } 2920 } 2921 break; 2922 2923 case TCPOPT_SACK: 2924 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) && 2925 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) && 2926 opt_rx->sack_ok) { 2927 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th; 2928 } 2929#ifdef CONFIG_TCP_MD5SIG 2930 case TCPOPT_MD5SIG: 2931 /* 2932 * The MD5 Hash has already been 2933 * checked (see tcp_v{4,6}_do_rcv()). 2934 */ 2935 break; 2936#endif 2937 } 2938 2939 ptr+=opsize-2; 2940 length-=opsize; 2941 } 2942 } 2943} 2944 2945/* Fast parse options. This hopes to only see timestamps. 2946 * If it is wrong it falls back on tcp_parse_options(). 2947 */ 2948static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th, 2949 struct tcp_sock *tp) 2950{ 2951 if (th->doff == sizeof(struct tcphdr)>>2) { 2952 tp->rx_opt.saw_tstamp = 0; 2953 return 0; 2954 } else if (tp->rx_opt.tstamp_ok && 2955 th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) { 2956 __be32 *ptr = (__be32 *)(th + 1); 2957 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) 2958 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) { 2959 tp->rx_opt.saw_tstamp = 1; 2960 ++ptr; 2961 tp->rx_opt.rcv_tsval = ntohl(*ptr); 2962 ++ptr; 2963 tp->rx_opt.rcv_tsecr = ntohl(*ptr); 2964 return 1; 2965 } 2966 } 2967 tcp_parse_options(skb, &tp->rx_opt, 1); 2968 return 1; 2969} 2970 2971static inline void tcp_store_ts_recent(struct tcp_sock *tp) 2972{ 2973 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval; 2974 tp->rx_opt.ts_recent_stamp = get_seconds(); 2975} 2976 2977static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq) 2978{ 2979 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) { 2980 /* PAWS bug workaround wrt. ACK frames, the PAWS discard 2981 * extra check below makes sure this can only happen 2982 * for pure ACK frames. -DaveM 2983 * 2984 * Not only, also it occurs for expired timestamps. 2985 */ 2986 2987 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) >= 0 || 2988 get_seconds() >= tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS) 2989 tcp_store_ts_recent(tp); 2990 } 2991} 2992 2993/* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM 2994 * 2995 * It is not fatal. If this ACK does _not_ change critical state (seqs, window) 2996 * it can pass through stack. So, the following predicate verifies that 2997 * this segment is not used for anything but congestion avoidance or 2998 * fast retransmit. Moreover, we even are able to eliminate most of such 2999 * second order effects, if we apply some small "replay" window (~RTO) 3000 * to timestamp space. 3001 * 3002 * All these measures still do not guarantee that we reject wrapped ACKs 3003 * on networks with high bandwidth, when sequence space is recycled fastly, 3004 * but it guarantees that such events will be very rare and do not affect 3005 * connection seriously. This doesn't look nice, but alas, PAWS is really 3006 * buggy extension. 3007 * 3008 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC 3009 * states that events when retransmit arrives after original data are rare. 3010 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is 3011 * the biggest problem on large power networks even with minor reordering. 3012 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe 3013 * up to bandwidth of 18Gigabit/sec. 8) ] 3014 */ 3015 3016static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb) 3017{ 3018 struct tcp_sock *tp = tcp_sk(sk); 3019 struct tcphdr *th = tcp_hdr(skb); 3020 u32 seq = TCP_SKB_CB(skb)->seq; 3021 u32 ack = TCP_SKB_CB(skb)->ack_seq; 3022 3023 return (/* 1. Pure ACK with correct sequence number. */ 3024 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) && 3025 3026 /* 2. ... and duplicate ACK. */ 3027 ack == tp->snd_una && 3028 3029 /* 3. ... and does not update window. */ 3030 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) && 3031 3032 /* 4. ... and sits in replay window. */ 3033 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ); 3034} 3035 3036static inline int tcp_paws_discard(const struct sock *sk, const struct sk_buff *skb) 3037{ 3038 const struct tcp_sock *tp = tcp_sk(sk); 3039 return ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > TCP_PAWS_WINDOW && 3040 get_seconds() < tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS && 3041 !tcp_disordered_ack(sk, skb)); 3042} 3043 3044/* Check segment sequence number for validity. 3045 * 3046 * Segment controls are considered valid, if the segment 3047 * fits to the window after truncation to the window. Acceptability 3048 * of data (and SYN, FIN, of course) is checked separately. 3049 * See tcp_data_queue(), for example. 3050 * 3051 * Also, controls (RST is main one) are accepted using RCV.WUP instead 3052 * of RCV.NXT. Peer still did not advance his SND.UNA when we 3053 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP. 3054 * (borrowed from freebsd) 3055 */ 3056 3057static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq) 3058{ 3059 return !before(end_seq, tp->rcv_wup) && 3060 !after(seq, tp->rcv_nxt + tcp_receive_window(tp)); 3061} 3062 3063/* When we get a reset we do this. */ 3064static void tcp_reset(struct sock *sk) 3065{ 3066 /* We want the right error as BSD sees it (and indeed as we do). */ 3067 switch (sk->sk_state) { 3068 case TCP_SYN_SENT: 3069 sk->sk_err = ECONNREFUSED; 3070 break; 3071 case TCP_CLOSE_WAIT: 3072 sk->sk_err = EPIPE; 3073 break; 3074 case TCP_CLOSE: 3075 return; 3076 default: 3077 sk->sk_err = ECONNRESET; 3078 } 3079 3080 if (!sock_flag(sk, SOCK_DEAD)) 3081 sk->sk_error_report(sk); 3082 3083 tcp_done(sk); 3084} 3085 3086/* 3087 * Process the FIN bit. This now behaves as it is supposed to work 3088 * and the FIN takes effect when it is validly part of sequence 3089 * space. Not before when we get holes. 3090 * 3091 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT 3092 * (and thence onto LAST-ACK and finally, CLOSE, we never enter 3093 * TIME-WAIT) 3094 * 3095 * If we are in FINWAIT-1, a received FIN indicates simultaneous 3096 * close and we go into CLOSING (and later onto TIME-WAIT) 3097 * 3098 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT. 3099 */ 3100static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th) 3101{ 3102 struct tcp_sock *tp = tcp_sk(sk); 3103 3104 inet_csk_schedule_ack(sk); 3105 3106 sk->sk_shutdown |= RCV_SHUTDOWN; 3107 sock_set_flag(sk, SOCK_DONE); 3108 3109 switch (sk->sk_state) { 3110 case TCP_SYN_RECV: 3111 case TCP_ESTABLISHED: 3112 /* Move to CLOSE_WAIT */ 3113 tcp_set_state(sk, TCP_CLOSE_WAIT); 3114 inet_csk(sk)->icsk_ack.pingpong = 1; 3115 break; 3116 3117 case TCP_CLOSE_WAIT: 3118 case TCP_CLOSING: 3119 /* Received a retransmission of the FIN, do 3120 * nothing. 3121 */ 3122 break; 3123 case TCP_LAST_ACK: 3124 /* RFC793: Remain in the LAST-ACK state. */ 3125 break; 3126 3127 case TCP_FIN_WAIT1: 3128 /* This case occurs when a simultaneous close 3129 * happens, we must ack the received FIN and 3130 * enter the CLOSING state. 3131 */ 3132 tcp_send_ack(sk); 3133 tcp_set_state(sk, TCP_CLOSING); 3134 break; 3135 case TCP_FIN_WAIT2: 3136 /* Received a FIN -- send ACK and enter TIME_WAIT. */ 3137 tcp_send_ack(sk); 3138 tcp_time_wait(sk, TCP_TIME_WAIT, 0); 3139 break; 3140 default: 3141 /* Only TCP_LISTEN and TCP_CLOSE are left, in these 3142 * cases we should never reach this piece of code. 3143 */ 3144 printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n", 3145 __FUNCTION__, sk->sk_state); 3146 break; 3147 } 3148 3149 /* It _is_ possible, that we have something out-of-order _after_ FIN. 3150 * Probably, we should reset in this case. For now drop them. 3151 */ 3152 __skb_queue_purge(&tp->out_of_order_queue); 3153 if (tp->rx_opt.sack_ok) 3154 tcp_sack_reset(&tp->rx_opt); 3155 sk_stream_mem_reclaim(sk); 3156 3157 if (!sock_flag(sk, SOCK_DEAD)) { 3158 sk->sk_state_change(sk); 3159 3160 /* Do not send POLL_HUP for half duplex close. */ 3161 if (sk->sk_shutdown == SHUTDOWN_MASK || 3162 sk->sk_state == TCP_CLOSE) 3163 sk_wake_async(sk, 1, POLL_HUP); 3164 else 3165 sk_wake_async(sk, 1, POLL_IN); 3166 } 3167} 3168 3169static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, u32 end_seq) 3170{ 3171 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) { 3172 if (before(seq, sp->start_seq)) 3173 sp->start_seq = seq; 3174 if (after(end_seq, sp->end_seq)) 3175 sp->end_seq = end_seq; 3176 return 1; 3177 } 3178 return 0; 3179} 3180 3181static void tcp_dsack_set(struct tcp_sock *tp, u32 seq, u32 end_seq) 3182{ 3183 if (tp->rx_opt.sack_ok && sysctl_tcp_dsack) { 3184 if (before(seq, tp->rcv_nxt)) 3185 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT); 3186 else 3187 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT); 3188 3189 tp->rx_opt.dsack = 1; 3190 tp->duplicate_sack[0].start_seq = seq; 3191 tp->duplicate_sack[0].end_seq = end_seq; 3192 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + 1, 4 - tp->rx_opt.tstamp_ok); 3193 } 3194} 3195 3196static void tcp_dsack_extend(struct tcp_sock *tp, u32 seq, u32 end_seq) 3197{ 3198 if (!tp->rx_opt.dsack) 3199 tcp_dsack_set(tp, seq, end_seq); 3200 else 3201 tcp_sack_extend(tp->duplicate_sack, seq, end_seq); 3202} 3203 3204static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb) 3205{ 3206 struct tcp_sock *tp = tcp_sk(sk); 3207 3208 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 3209 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 3210 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST); 3211 tcp_enter_quickack_mode(sk); 3212 3213 if (tp->rx_opt.sack_ok && sysctl_tcp_dsack) { 3214 u32 end_seq = TCP_SKB_CB(skb)->end_seq; 3215 3216 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) 3217 end_seq = tp->rcv_nxt; 3218 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, end_seq); 3219 } 3220 } 3221 3222 tcp_send_ack(sk); 3223} 3224 3225/* These routines update the SACK block as out-of-order packets arrive or 3226 * in-order packets close up the sequence space. 3227 */ 3228static void tcp_sack_maybe_coalesce(struct tcp_sock *tp) 3229{ 3230 int this_sack; 3231 struct tcp_sack_block *sp = &tp->selective_acks[0]; 3232 struct tcp_sack_block *swalk = sp+1; 3233 3234 /* See if the recent change to the first SACK eats into 3235 * or hits the sequence space of other SACK blocks, if so coalesce. 3236 */ 3237 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks; ) { 3238 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) { 3239 int i; 3240 3241 /* Zap SWALK, by moving every further SACK up by one slot. 3242 * Decrease num_sacks. 3243 */ 3244 tp->rx_opt.num_sacks--; 3245 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok); 3246 for (i=this_sack; i < tp->rx_opt.num_sacks; i++) 3247 sp[i] = sp[i+1]; 3248 continue; 3249 } 3250 this_sack++, swalk++; 3251 } 3252} 3253 3254static inline void tcp_sack_swap(struct tcp_sack_block *sack1, struct tcp_sack_block *sack2) 3255{ 3256 __u32 tmp; 3257 3258 tmp = sack1->start_seq; 3259 sack1->start_seq = sack2->start_seq; 3260 sack2->start_seq = tmp; 3261 3262 tmp = sack1->end_seq; 3263 sack1->end_seq = sack2->end_seq; 3264 sack2->end_seq = tmp; 3265} 3266 3267static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq) 3268{ 3269 struct tcp_sock *tp = tcp_sk(sk); 3270 struct tcp_sack_block *sp = &tp->selective_acks[0]; 3271 int cur_sacks = tp->rx_opt.num_sacks; 3272 int this_sack; 3273 3274 if (!cur_sacks) 3275 goto new_sack; 3276 3277 for (this_sack=0; this_sack<cur_sacks; this_sack++, sp++) { 3278 if (tcp_sack_extend(sp, seq, end_seq)) { 3279 /* Rotate this_sack to the first one. */ 3280 for (; this_sack>0; this_sack--, sp--) 3281 tcp_sack_swap(sp, sp-1); 3282 if (cur_sacks > 1) 3283 tcp_sack_maybe_coalesce(tp); 3284 return; 3285 } 3286 } 3287 3288 /* Could not find an adjacent existing SACK, build a new one, 3289 * put it at the front, and shift everyone else down. We 3290 * always know there is at least one SACK present already here. 3291 * 3292 * If the sack array is full, forget about the last one. 3293 */ 3294 if (this_sack >= 4) { 3295 this_sack--; 3296 tp->rx_opt.num_sacks--; 3297 sp--; 3298 } 3299 for (; this_sack > 0; this_sack--, sp--) 3300 *sp = *(sp-1); 3301 3302new_sack: 3303 /* Build the new head SACK, and we're done. */ 3304 sp->start_seq = seq; 3305 sp->end_seq = end_seq; 3306 tp->rx_opt.num_sacks++; 3307 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok); 3308} 3309 3310/* RCV.NXT advances, some SACKs should be eaten. */ 3311 3312static void tcp_sack_remove(struct tcp_sock *tp) 3313{ 3314 struct tcp_sack_block *sp = &tp->selective_acks[0]; 3315 int num_sacks = tp->rx_opt.num_sacks; 3316 int this_sack; 3317 3318 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */ 3319 if (skb_queue_empty(&tp->out_of_order_queue)) { 3320 tp->rx_opt.num_sacks = 0; 3321 tp->rx_opt.eff_sacks = tp->rx_opt.dsack; 3322 return; 3323 } 3324 3325 for (this_sack = 0; this_sack < num_sacks; ) { 3326 /* Check if the start of the sack is covered by RCV.NXT. */ 3327 if (!before(tp->rcv_nxt, sp->start_seq)) { 3328 int i; 3329 3330 /* RCV.NXT must cover all the block! */ 3331 BUG_TRAP(!before(tp->rcv_nxt, sp->end_seq)); 3332 3333 /* Zap this SACK, by moving forward any other SACKS. */ 3334 for (i=this_sack+1; i < num_sacks; i++) 3335 tp->selective_acks[i-1] = tp->selective_acks[i]; 3336 num_sacks--; 3337 continue; 3338 } 3339 this_sack++; 3340 sp++; 3341 } 3342 if (num_sacks != tp->rx_opt.num_sacks) { 3343 tp->rx_opt.num_sacks = num_sacks; 3344 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok); 3345 } 3346} 3347 3348/* This one checks to see if we can put data from the 3349 * out_of_order queue into the receive_queue. 3350 */ 3351static void tcp_ofo_queue(struct sock *sk) 3352{ 3353 struct tcp_sock *tp = tcp_sk(sk); 3354 __u32 dsack_high = tp->rcv_nxt; 3355 struct sk_buff *skb; 3356 3357 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) { 3358 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) 3359 break; 3360 3361 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) { 3362 __u32 dsack = dsack_high; 3363 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high)) 3364 dsack_high = TCP_SKB_CB(skb)->end_seq; 3365 tcp_dsack_extend(tp, TCP_SKB_CB(skb)->seq, dsack); 3366 } 3367 3368 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) { 3369 SOCK_DEBUG(sk, "ofo packet was already received \n"); 3370 __skb_unlink(skb, &tp->out_of_order_queue); 3371 __kfree_skb(skb); 3372 continue; 3373 } 3374 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n", 3375 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, 3376 TCP_SKB_CB(skb)->end_seq); 3377 3378 __skb_unlink(skb, &tp->out_of_order_queue); 3379 __skb_queue_tail(&sk->sk_receive_queue, skb); 3380 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; 3381 if (tcp_hdr(skb)->fin) 3382 tcp_fin(skb, sk, tcp_hdr(skb)); 3383 } 3384} 3385 3386static int tcp_prune_queue(struct sock *sk); 3387 3388static void tcp_data_queue(struct sock *sk, struct sk_buff *skb) 3389{ 3390 struct tcphdr *th = tcp_hdr(skb); 3391 struct tcp_sock *tp = tcp_sk(sk); 3392 int eaten = -1; 3393 3394 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) 3395 goto drop; 3396 3397 __skb_pull(skb, th->doff*4); 3398 3399 TCP_ECN_accept_cwr(tp, skb); 3400 3401 if (tp->rx_opt.dsack) { 3402 tp->rx_opt.dsack = 0; 3403 tp->rx_opt.eff_sacks = min_t(unsigned int, tp->rx_opt.num_sacks, 3404 4 - tp->rx_opt.tstamp_ok); 3405 } 3406 3407 /* Queue data for delivery to the user. 3408 * Packets in sequence go to the receive queue. 3409 * Out of sequence packets to the out_of_order_queue. 3410 */ 3411 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) { 3412 if (tcp_receive_window(tp) == 0) 3413 goto out_of_window; 3414 3415 /* Ok. In sequence. In window. */ 3416 if (tp->ucopy.task == current && 3417 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len && 3418 sock_owned_by_user(sk) && !tp->urg_data) { 3419 int chunk = min_t(unsigned int, skb->len, 3420 tp->ucopy.len); 3421 3422 __set_current_state(TASK_RUNNING); 3423 3424 local_bh_enable(); 3425 if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) { 3426 tp->ucopy.len -= chunk; 3427 tp->copied_seq += chunk; 3428 eaten = (chunk == skb->len && !th->fin); 3429 tcp_rcv_space_adjust(sk); 3430 } 3431 local_bh_disable(); 3432 } 3433 3434 if (eaten <= 0) { 3435queue_and_out: 3436 if (eaten < 0 && 3437 (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf || 3438 !sk_stream_rmem_schedule(sk, skb))) { 3439 if (tcp_prune_queue(sk) < 0 || 3440 !sk_stream_rmem_schedule(sk, skb)) 3441 goto drop; 3442 } 3443 sk_stream_set_owner_r(skb, sk); 3444 __skb_queue_tail(&sk->sk_receive_queue, skb); 3445 } 3446 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; 3447 if (skb->len) 3448 tcp_event_data_recv(sk, skb); 3449 if (th->fin) 3450 tcp_fin(skb, sk, th); 3451 3452 if (!skb_queue_empty(&tp->out_of_order_queue)) { 3453 tcp_ofo_queue(sk); 3454 3455 /* RFC2581. 4.2. SHOULD send immediate ACK, when 3456 * gap in queue is filled. 3457 */ 3458 if (skb_queue_empty(&tp->out_of_order_queue)) 3459 inet_csk(sk)->icsk_ack.pingpong = 0; 3460 } 3461 3462 if (tp->rx_opt.num_sacks) 3463 tcp_sack_remove(tp); 3464 3465 tcp_fast_path_check(sk); 3466 3467 if (eaten > 0) 3468 __kfree_skb(skb); 3469 else if (!sock_flag(sk, SOCK_DEAD)) 3470 sk->sk_data_ready(sk, 0); 3471 return; 3472 } 3473 3474 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) { 3475 /* A retransmit, 2nd most common case. Force an immediate ack. */ 3476 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST); 3477 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); 3478 3479out_of_window: 3480 tcp_enter_quickack_mode(sk); 3481 inet_csk_schedule_ack(sk); 3482drop: 3483 __kfree_skb(skb); 3484 return; 3485 } 3486 3487 /* Out of window. F.e. zero window probe. */ 3488 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp))) 3489 goto out_of_window; 3490 3491 tcp_enter_quickack_mode(sk); 3492 3493 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 3494 /* Partial packet, seq < rcv_next < end_seq */ 3495 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n", 3496 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, 3497 TCP_SKB_CB(skb)->end_seq); 3498 3499 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, tp->rcv_nxt); 3500 3501 /* If window is closed, drop tail of packet. But after 3502 * remembering D-SACK for its head made in previous line. 3503 */ 3504 if (!tcp_receive_window(tp)) 3505 goto out_of_window; 3506 goto queue_and_out; 3507 } 3508 3509 TCP_ECN_check_ce(tp, skb); 3510 3511 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf || 3512 !sk_stream_rmem_schedule(sk, skb)) { 3513 if (tcp_prune_queue(sk) < 0 || 3514 !sk_stream_rmem_schedule(sk, skb)) 3515 goto drop; 3516 } 3517 3518 /* Disable header prediction. */ 3519 tp->pred_flags = 0; 3520 inet_csk_schedule_ack(sk); 3521 3522 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n", 3523 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); 3524 3525 sk_stream_set_owner_r(skb, sk); 3526 3527 if (!skb_peek(&tp->out_of_order_queue)) { 3528 /* Initial out of order segment, build 1 SACK. */ 3529 if (tp->rx_opt.sack_ok) { 3530 tp->rx_opt.num_sacks = 1; 3531 tp->rx_opt.dsack = 0; 3532 tp->rx_opt.eff_sacks = 1; 3533 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq; 3534 tp->selective_acks[0].end_seq = 3535 TCP_SKB_CB(skb)->end_seq; 3536 } 3537 __skb_queue_head(&tp->out_of_order_queue,skb); 3538 } else { 3539 struct sk_buff *skb1 = tp->out_of_order_queue.prev; 3540 u32 seq = TCP_SKB_CB(skb)->seq; 3541 u32 end_seq = TCP_SKB_CB(skb)->end_seq; 3542 3543 if (seq == TCP_SKB_CB(skb1)->end_seq) { 3544 __skb_append(skb1, skb, &tp->out_of_order_queue); 3545 3546 if (!tp->rx_opt.num_sacks || 3547 tp->selective_acks[0].end_seq != seq) 3548 goto add_sack; 3549 3550 /* Common case: data arrive in order after hole. */ 3551 tp->selective_acks[0].end_seq = end_seq; 3552 return; 3553 } 3554 3555 /* Find place to insert this segment. */ 3556 do { 3557 if (!after(TCP_SKB_CB(skb1)->seq, seq)) 3558 break; 3559 } while ((skb1 = skb1->prev) != 3560 (struct sk_buff*)&tp->out_of_order_queue); 3561 3562 /* Do skb overlap to previous one? */ 3563 if (skb1 != (struct sk_buff*)&tp->out_of_order_queue && 3564 before(seq, TCP_SKB_CB(skb1)->end_seq)) { 3565 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) { 3566 /* All the bits are present. Drop. */ 3567 __kfree_skb(skb); 3568 tcp_dsack_set(tp, seq, end_seq); 3569 goto add_sack; 3570 } 3571 if (after(seq, TCP_SKB_CB(skb1)->seq)) { 3572 /* Partial overlap. */ 3573 tcp_dsack_set(tp, seq, TCP_SKB_CB(skb1)->end_seq); 3574 } else { 3575 skb1 = skb1->prev; 3576 } 3577 } 3578 __skb_insert(skb, skb1, skb1->next, &tp->out_of_order_queue); 3579 3580 /* And clean segments covered by new one as whole. */ 3581 while ((skb1 = skb->next) != 3582 (struct sk_buff*)&tp->out_of_order_queue && 3583 after(end_seq, TCP_SKB_CB(skb1)->seq)) { 3584 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) { 3585 tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, end_seq); 3586 break; 3587 } 3588 __skb_unlink(skb1, &tp->out_of_order_queue); 3589 tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, TCP_SKB_CB(skb1)->end_seq); 3590 __kfree_skb(skb1); 3591 } 3592 3593add_sack: 3594 if (tp->rx_opt.sack_ok) 3595 tcp_sack_new_ofo_skb(sk, seq, end_seq); 3596 } 3597} 3598 3599/* Collapse contiguous sequence of skbs head..tail with 3600 * sequence numbers start..end. 3601 * Segments with FIN/SYN are not collapsed (only because this 3602 * simplifies code) 3603 */ 3604static void 3605tcp_collapse(struct sock *sk, struct sk_buff_head *list, 3606 struct sk_buff *head, struct sk_buff *tail, 3607 u32 start, u32 end) 3608{ 3609 struct sk_buff *skb; 3610 3611 /* First, check that queue is collapsible and find 3612 * the point where collapsing can be useful. */ 3613 for (skb = head; skb != tail; ) { 3614 /* No new bits? It is possible on ofo queue. */ 3615 if (!before(start, TCP_SKB_CB(skb)->end_seq)) { 3616 struct sk_buff *next = skb->next; 3617 __skb_unlink(skb, list); 3618 __kfree_skb(skb); 3619 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED); 3620 skb = next; 3621 continue; 3622 } 3623 3624 /* The first skb to collapse is: 3625 * - not SYN/FIN and 3626 * - bloated or contains data before "start" or 3627 * overlaps to the next one. 3628 */ 3629 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin && 3630 (tcp_win_from_space(skb->truesize) > skb->len || 3631 before(TCP_SKB_CB(skb)->seq, start) || 3632 (skb->next != tail && 3633 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq))) 3634 break; 3635 3636 /* Decided to skip this, advance start seq. */ 3637 start = TCP_SKB_CB(skb)->end_seq; 3638 skb = skb->next; 3639 } 3640 if (skb == tail || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin) 3641 return; 3642 3643 while (before(start, end)) { 3644 struct sk_buff *nskb; 3645 int header = skb_headroom(skb); 3646 int copy = SKB_MAX_ORDER(header, 0); 3647 3648 /* Too big header? This can happen with IPv6. */ 3649 if (copy < 0) 3650 return; 3651 if (end-start < copy) 3652 copy = end-start; 3653 nskb = alloc_skb(copy+header, GFP_ATOMIC); 3654 if (!nskb) 3655 return; 3656 3657 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head); 3658 skb_set_network_header(nskb, (skb_network_header(skb) - 3659 skb->head)); 3660 skb_set_transport_header(nskb, (skb_transport_header(skb) - 3661 skb->head)); 3662 skb_reserve(nskb, header); 3663 memcpy(nskb->head, skb->head, header); 3664 memcpy(nskb->cb, skb->cb, sizeof(skb->cb)); 3665 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start; 3666 __skb_insert(nskb, skb->prev, skb, list); 3667 sk_stream_set_owner_r(nskb, sk); 3668 3669 /* Copy data, releasing collapsed skbs. */ 3670 while (copy > 0) { 3671 int offset = start - TCP_SKB_CB(skb)->seq; 3672 int size = TCP_SKB_CB(skb)->end_seq - start; 3673 3674 BUG_ON(offset < 0); 3675 if (size > 0) { 3676 size = min(copy, size); 3677 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size)) 3678 BUG(); 3679 TCP_SKB_CB(nskb)->end_seq += size; 3680 copy -= size; 3681 start += size; 3682 } 3683 if (!before(start, TCP_SKB_CB(skb)->end_seq)) { 3684 struct sk_buff *next = skb->next; 3685 __skb_unlink(skb, list); 3686 __kfree_skb(skb); 3687 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED); 3688 skb = next; 3689 if (skb == tail || 3690 tcp_hdr(skb)->syn || 3691 tcp_hdr(skb)->fin) 3692 return; 3693 } 3694 } 3695 } 3696} 3697 3698/* Collapse ofo queue. Algorithm: select contiguous sequence of skbs 3699 * and tcp_collapse() them until all the queue is collapsed. 3700 */ 3701static void tcp_collapse_ofo_queue(struct sock *sk) 3702{ 3703 struct tcp_sock *tp = tcp_sk(sk); 3704 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue); 3705 struct sk_buff *head; 3706 u32 start, end; 3707 3708 if (skb == NULL) 3709 return; 3710 3711 start = TCP_SKB_CB(skb)->seq; 3712 end = TCP_SKB_CB(skb)->end_seq; 3713 head = skb; 3714 3715 for (;;) { 3716 skb = skb->next; 3717 3718 /* Segment is terminated when we see gap or when 3719 * we are at the end of all the queue. */ 3720 if (skb == (struct sk_buff *)&tp->out_of_order_queue || 3721 after(TCP_SKB_CB(skb)->seq, end) || 3722 before(TCP_SKB_CB(skb)->end_seq, start)) { 3723 tcp_collapse(sk, &tp->out_of_order_queue, 3724 head, skb, start, end); 3725 head = skb; 3726 if (skb == (struct sk_buff *)&tp->out_of_order_queue) 3727 break; 3728 /* Start new segment */ 3729 start = TCP_SKB_CB(skb)->seq; 3730 end = TCP_SKB_CB(skb)->end_seq; 3731 } else { 3732 if (before(TCP_SKB_CB(skb)->seq, start)) 3733 start = TCP_SKB_CB(skb)->seq; 3734 if (after(TCP_SKB_CB(skb)->end_seq, end)) 3735 end = TCP_SKB_CB(skb)->end_seq; 3736 } 3737 } 3738} 3739 3740/* Reduce allocated memory if we can, trying to get 3741 * the socket within its memory limits again. 3742 * 3743 * Return less than zero if we should start dropping frames 3744 * until the socket owning process reads some of the data 3745 * to stabilize the situation. 3746 */ 3747static int tcp_prune_queue(struct sock *sk) 3748{ 3749 struct tcp_sock *tp = tcp_sk(sk); 3750 3751 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq); 3752 3753 NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED); 3754 3755 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf) 3756 tcp_clamp_window(sk); 3757 else if (tcp_memory_pressure) 3758 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss); 3759 3760 tcp_collapse_ofo_queue(sk); 3761 tcp_collapse(sk, &sk->sk_receive_queue, 3762 sk->sk_receive_queue.next, 3763 (struct sk_buff*)&sk->sk_receive_queue, 3764 tp->copied_seq, tp->rcv_nxt); 3765 sk_stream_mem_reclaim(sk); 3766 3767 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) 3768 return 0; 3769 3770 /* Collapsing did not help, destructive actions follow. 3771 * This must not ever occur. */ 3772 3773 /* First, purge the out_of_order queue. */ 3774 if (!skb_queue_empty(&tp->out_of_order_queue)) { 3775 NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED); 3776 __skb_queue_purge(&tp->out_of_order_queue); 3777 3778 /* Reset SACK state. A conforming SACK implementation will 3779 * do the same at a timeout based retransmit. When a connection 3780 * is in a sad state like this, we care only about integrity 3781 * of the connection not performance. 3782 */ 3783 if (tp->rx_opt.sack_ok) 3784 tcp_sack_reset(&tp->rx_opt); 3785 sk_stream_mem_reclaim(sk); 3786 } 3787 3788 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) 3789 return 0; 3790 3791 /* If we are really being abused, tell the caller to silently 3792 * drop receive data on the floor. It will get retransmitted 3793 * and hopefully then we'll have sufficient space. 3794 */ 3795 NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED); 3796 3797 /* Massive buffer overcommit. */ 3798 tp->pred_flags = 0; 3799 return -1; 3800} 3801 3802 3803/* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. 3804 * As additional protections, we do not touch cwnd in retransmission phases, 3805 * and if application hit its sndbuf limit recently. 3806 */ 3807void tcp_cwnd_application_limited(struct sock *sk) 3808{ 3809 struct tcp_sock *tp = tcp_sk(sk); 3810 3811 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && 3812 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { 3813 /* Limited by application or receiver window. */ 3814 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); 3815 u32 win_used = max(tp->snd_cwnd_used, init_win); 3816 if (win_used < tp->snd_cwnd) { 3817 tp->snd_ssthresh = tcp_current_ssthresh(sk); 3818 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1; 3819 } 3820 tp->snd_cwnd_used = 0; 3821 } 3822 tp->snd_cwnd_stamp = tcp_time_stamp; 3823} 3824 3825static int tcp_should_expand_sndbuf(struct sock *sk) 3826{ 3827 struct tcp_sock *tp = tcp_sk(sk); 3828 3829 /* If the user specified a specific send buffer setting, do 3830 * not modify it. 3831 */ 3832 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK) 3833 return 0; 3834 3835 /* If we are under global TCP memory pressure, do not expand. */ 3836 if (tcp_memory_pressure) 3837 return 0; 3838 3839 /* If we are under soft global TCP memory pressure, do not expand. */ 3840 if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0]) 3841 return 0; 3842 3843 /* If we filled the congestion window, do not expand. */ 3844 if (tp->packets_out >= tp->snd_cwnd) 3845 return 0; 3846 3847 return 1; 3848} 3849 3850/* When incoming ACK allowed to free some skb from write_queue, 3851 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket 3852 * on the exit from tcp input handler. 3853 * 3854 * PROBLEM: sndbuf expansion does not work well with largesend. 3855 */ 3856static void tcp_new_space(struct sock *sk) 3857{ 3858 struct tcp_sock *tp = tcp_sk(sk); 3859 3860 if (tcp_should_expand_sndbuf(sk)) { 3861 int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) + 3862 MAX_TCP_HEADER + 16 + sizeof(struct sk_buff), 3863 demanded = max_t(unsigned int, tp->snd_cwnd, 3864 tp->reordering + 1); 3865 sndmem *= 2*demanded; 3866 if (sndmem > sk->sk_sndbuf) 3867 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]); 3868 tp->snd_cwnd_stamp = tcp_time_stamp; 3869 } 3870 3871 sk->sk_write_space(sk); 3872} 3873 3874static void tcp_check_space(struct sock *sk) 3875{ 3876 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) { 3877 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK); 3878 if (sk->sk_socket && 3879 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) 3880 tcp_new_space(sk); 3881 } 3882} 3883 3884static inline void tcp_data_snd_check(struct sock *sk) 3885{ 3886 tcp_push_pending_frames(sk); 3887 tcp_check_space(sk); 3888} 3889 3890/* 3891 * Check if sending an ack is needed. 3892 */ 3893static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible) 3894{ 3895 struct tcp_sock *tp = tcp_sk(sk); 3896 3897 /* More than one full frame received... */ 3898 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss 3899 /* ... and right edge of window advances far enough. 3900 * (tcp_recvmsg() will send ACK otherwise). Or... 3901 */ 3902 && __tcp_select_window(sk) >= tp->rcv_wnd) || 3903 /* We ACK each frame or... */ 3904 tcp_in_quickack_mode(sk) || 3905 /* We have out of order data. */ 3906 (ofo_possible && 3907 skb_peek(&tp->out_of_order_queue))) { 3908 /* Then ack it now */ 3909 tcp_send_ack(sk); 3910 } else { 3911 /* Else, send delayed ack. */ 3912 tcp_send_delayed_ack(sk); 3913 } 3914} 3915 3916static inline void tcp_ack_snd_check(struct sock *sk) 3917{ 3918 if (!inet_csk_ack_scheduled(sk)) { 3919 /* We sent a data segment already. */ 3920 return; 3921 } 3922 __tcp_ack_snd_check(sk, 1); 3923} 3924 3925/* 3926 * This routine is only called when we have urgent data 3927 * signaled. Its the 'slow' part of tcp_urg. It could be 3928 * moved inline now as tcp_urg is only called from one 3929 * place. We handle URGent data wrong. We have to - as 3930 * BSD still doesn't use the correction from RFC961. 3931 * For 1003.1g we should support a new option TCP_STDURG to permit 3932 * either form (or just set the sysctl tcp_stdurg). 3933 */ 3934 3935static void tcp_check_urg(struct sock * sk, struct tcphdr * th) 3936{ 3937 struct tcp_sock *tp = tcp_sk(sk); 3938 u32 ptr = ntohs(th->urg_ptr); 3939 3940 if (ptr && !sysctl_tcp_stdurg) 3941 ptr--; 3942 ptr += ntohl(th->seq); 3943 3944 /* Ignore urgent data that we've already seen and read. */ 3945 if (after(tp->copied_seq, ptr)) 3946 return; 3947 3948 /* Do not replay urg ptr. 3949 * 3950 * NOTE: interesting situation not covered by specs. 3951 * Misbehaving sender may send urg ptr, pointing to segment, 3952 * which we already have in ofo queue. We are not able to fetch 3953 * such data and will stay in TCP_URG_NOTYET until will be eaten 3954 * by recvmsg(). Seems, we are not obliged to handle such wicked 3955 * situations. But it is worth to think about possibility of some 3956 * DoSes using some hypothetical application level deadlock. 3957 */ 3958 if (before(ptr, tp->rcv_nxt)) 3959 return; 3960 3961 /* Do we already have a newer (or duplicate) urgent pointer? */ 3962 if (tp->urg_data && !after(ptr, tp->urg_seq)) 3963 return; 3964 3965 /* Tell the world about our new urgent pointer. */ 3966 sk_send_sigurg(sk); 3967 3968 /* We may be adding urgent data when the last byte read was 3969 * urgent. To do this requires some care. We cannot just ignore 3970 * tp->copied_seq since we would read the last urgent byte again 3971 * as data, nor can we alter copied_seq until this data arrives 3972 * or we break the semantics of SIOCATMARK (and thus sockatmark()) 3973 * 3974 * NOTE. Double Dutch. Rendering to plain English: author of comment 3975 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB); 3976 * and expect that both A and B disappear from stream. This is _wrong_. 3977 * Though this happens in BSD with high probability, this is occasional. 3978 * Any application relying on this is buggy. Note also, that fix "works" 3979 * only in this artificial test. Insert some normal data between A and B and we will 3980 * decline of BSD again. Verdict: it is better to remove to trap 3981 * buggy users. 3982 */ 3983 if (tp->urg_seq == tp->copied_seq && tp->urg_data && 3984 !sock_flag(sk, SOCK_URGINLINE) && 3985 tp->copied_seq != tp->rcv_nxt) { 3986 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue); 3987 tp->copied_seq++; 3988 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) { 3989 __skb_unlink(skb, &sk->sk_receive_queue); 3990 __kfree_skb(skb); 3991 } 3992 } 3993 3994 tp->urg_data = TCP_URG_NOTYET; 3995 tp->urg_seq = ptr; 3996 3997 /* Disable header prediction. */ 3998 tp->pred_flags = 0; 3999} 4000 4001/* This is the 'fast' part of urgent handling. */ 4002static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th) 4003{ 4004 struct tcp_sock *tp = tcp_sk(sk); 4005 4006 /* Check if we get a new urgent pointer - normally not. */ 4007 if (th->urg) 4008 tcp_check_urg(sk,th); 4009 4010 /* Do we wait for any urgent data? - normally not... */ 4011 if (tp->urg_data == TCP_URG_NOTYET) { 4012 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) - 4013 th->syn; 4014 4015 /* Is the urgent pointer pointing into this packet? */ 4016 if (ptr < skb->len) { 4017 u8 tmp; 4018 if (skb_copy_bits(skb, ptr, &tmp, 1)) 4019 BUG(); 4020 tp->urg_data = TCP_URG_VALID | tmp; 4021 if (!sock_flag(sk, SOCK_DEAD)) 4022 sk->sk_data_ready(sk, 0); 4023 } 4024 } 4025} 4026 4027static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen) 4028{ 4029 struct tcp_sock *tp = tcp_sk(sk); 4030 int chunk = skb->len - hlen; 4031 int err; 4032 4033 local_bh_enable(); 4034 if (skb_csum_unnecessary(skb)) 4035 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk); 4036 else 4037 err = skb_copy_and_csum_datagram_iovec(skb, hlen, 4038 tp->ucopy.iov); 4039 4040 if (!err) { 4041 tp->ucopy.len -= chunk; 4042 tp->copied_seq += chunk; 4043 tcp_rcv_space_adjust(sk); 4044 } 4045 4046 local_bh_disable(); 4047 return err; 4048} 4049 4050static __sum16 __tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb) 4051{ 4052 __sum16 result; 4053 4054 if (sock_owned_by_user(sk)) { 4055 local_bh_enable(); 4056 result = __tcp_checksum_complete(skb); 4057 local_bh_disable(); 4058 } else { 4059 result = __tcp_checksum_complete(skb); 4060 } 4061 return result; 4062} 4063 4064static inline int tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb) 4065{ 4066 return !skb_csum_unnecessary(skb) && 4067 __tcp_checksum_complete_user(sk, skb); 4068} 4069 4070#ifdef CONFIG_NET_DMA 4071static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb, int hlen) 4072{ 4073 struct tcp_sock *tp = tcp_sk(sk); 4074 int chunk = skb->len - hlen; 4075 int dma_cookie; 4076 int copied_early = 0; 4077 4078 if (tp->ucopy.wakeup) 4079 return 0; 4080 4081 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list) 4082 tp->ucopy.dma_chan = get_softnet_dma(); 4083 4084 if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) { 4085 4086 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan, 4087 skb, hlen, tp->ucopy.iov, chunk, tp->ucopy.pinned_list); 4088 4089 if (dma_cookie < 0) 4090 goto out; 4091 4092 tp->ucopy.dma_cookie = dma_cookie; 4093 copied_early = 1; 4094 4095 tp->ucopy.len -= chunk; 4096 tp->copied_seq += chunk; 4097 tcp_rcv_space_adjust(sk); 4098 4099 if ((tp->ucopy.len == 0) || 4100 (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) || 4101 (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) { 4102 tp->ucopy.wakeup = 1; 4103 sk->sk_data_ready(sk, 0); 4104 } 4105 } else if (chunk > 0) { 4106 tp->ucopy.wakeup = 1; 4107 sk->sk_data_ready(sk, 0); 4108 } 4109out: 4110 return copied_early; 4111} 4112#endif /* CONFIG_NET_DMA */ 4113 4114/* 4115 * TCP receive function for the ESTABLISHED state. 4116 * 4117 * It is split into a fast path and a slow path. The fast path is 4118 * disabled when: 4119 * - A zero window was announced from us - zero window probing 4120 * is only handled properly in the slow path. 4121 * - Out of order segments arrived. 4122 * - Urgent data is expected. 4123 * - There is no buffer space left 4124 * - Unexpected TCP flags/window values/header lengths are received 4125 * (detected by checking the TCP header against pred_flags) 4126 * - Data is sent in both directions. Fast path only supports pure senders 4127 * or pure receivers (this means either the sequence number or the ack 4128 * value must stay constant) 4129 * - Unexpected TCP option. 4130 * 4131 * When these conditions are not satisfied it drops into a standard 4132 * receive procedure patterned after RFC793 to handle all cases. 4133 * The first three cases are guaranteed by proper pred_flags setting, 4134 * the rest is checked inline. Fast processing is turned on in 4135 * tcp_data_queue when everything is OK. 4136 */ 4137int tcp_rcv_established(struct sock *sk, struct sk_buff *skb, 4138 struct tcphdr *th, unsigned len) 4139{ 4140 struct tcp_sock *tp = tcp_sk(sk); 4141 4142 /* 4143 * Header prediction. 4144 * The code loosely follows the one in the famous 4145 * "30 instruction TCP receive" Van Jacobson mail. 4146 * 4147 * Van's trick is to deposit buffers into socket queue 4148 * on a device interrupt, to call tcp_recv function 4149 * on the receive process context and checksum and copy 4150 * the buffer to user space. smart... 4151 * 4152 * Our current scheme is not silly either but we take the 4153 * extra cost of the net_bh soft interrupt processing... 4154 * We do checksum and copy also but from device to kernel. 4155 */ 4156 4157 tp->rx_opt.saw_tstamp = 0; 4158 4159 /* pred_flags is 0xS?10 << 16 + snd_wnd 4160 * if header_prediction is to be made 4161 * 'S' will always be tp->tcp_header_len >> 2 4162 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to 4163 * turn it off (when there are holes in the receive 4164 * space for instance) 4165 * PSH flag is ignored. 4166 */ 4167 4168 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags && 4169 TCP_SKB_CB(skb)->seq == tp->rcv_nxt) { 4170 int tcp_header_len = tp->tcp_header_len; 4171 4172 /* Timestamp header prediction: tcp_header_len 4173 * is automatically equal to th->doff*4 due to pred_flags 4174 * match. 4175 */ 4176 4177 /* Check timestamp */ 4178 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) { 4179 __be32 *ptr = (__be32 *)(th + 1); 4180 4181 /* No? Slow path! */ 4182 if (*ptr != htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) 4183 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) 4184 goto slow_path; 4185 4186 tp->rx_opt.saw_tstamp = 1; 4187 ++ptr; 4188 tp->rx_opt.rcv_tsval = ntohl(*ptr); 4189 ++ptr; 4190 tp->rx_opt.rcv_tsecr = ntohl(*ptr); 4191 4192 /* If PAWS failed, check it more carefully in slow path */ 4193 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0) 4194 goto slow_path; 4195 4196 /* DO NOT update ts_recent here, if checksum fails 4197 * and timestamp was corrupted part, it will result 4198 * in a hung connection since we will drop all 4199 * future packets due to the PAWS test. 4200 */ 4201 } 4202 4203 if (len <= tcp_header_len) { 4204 /* Bulk data transfer: sender */ 4205 if (len == tcp_header_len) { 4206 /* Predicted packet is in window by definition. 4207 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 4208 * Hence, check seq<=rcv_wup reduces to: 4209 */ 4210 if (tcp_header_len == 4211 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && 4212 tp->rcv_nxt == tp->rcv_wup) 4213 tcp_store_ts_recent(tp); 4214 4215 /* We know that such packets are checksummed 4216 * on entry. 4217 */ 4218 tcp_ack(sk, skb, 0); 4219 __kfree_skb(skb); 4220 tcp_data_snd_check(sk); 4221 return 0; 4222 } else { /* Header too small */ 4223 TCP_INC_STATS_BH(TCP_MIB_INERRS); 4224 goto discard; 4225 } 4226 } else { 4227 int eaten = 0; 4228 int copied_early = 0; 4229 4230 if (tp->copied_seq == tp->rcv_nxt && 4231 len - tcp_header_len <= tp->ucopy.len) { 4232#ifdef CONFIG_NET_DMA 4233 if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) { 4234 copied_early = 1; 4235 eaten = 1; 4236 } 4237#endif 4238 if (tp->ucopy.task == current && sock_owned_by_user(sk) && !copied_early) { 4239 __set_current_state(TASK_RUNNING); 4240 4241 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len)) 4242 eaten = 1; 4243 } 4244 if (eaten) { 4245 /* Predicted packet is in window by definition. 4246 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 4247 * Hence, check seq<=rcv_wup reduces to: 4248 */ 4249 if (tcp_header_len == 4250 (sizeof(struct tcphdr) + 4251 TCPOLEN_TSTAMP_ALIGNED) && 4252 tp->rcv_nxt == tp->rcv_wup) 4253 tcp_store_ts_recent(tp); 4254 4255 tcp_rcv_rtt_measure_ts(sk, skb); 4256 4257 __skb_pull(skb, tcp_header_len); 4258 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; 4259 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER); 4260 } 4261 if (copied_early) 4262 tcp_cleanup_rbuf(sk, skb->len); 4263 } 4264 if (!eaten) { 4265 if (tcp_checksum_complete_user(sk, skb)) 4266 goto csum_error; 4267 4268 /* Predicted packet is in window by definition. 4269 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 4270 * Hence, check seq<=rcv_wup reduces to: 4271 */ 4272 if (tcp_header_len == 4273 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && 4274 tp->rcv_nxt == tp->rcv_wup) 4275 tcp_store_ts_recent(tp); 4276 4277 tcp_rcv_rtt_measure_ts(sk, skb); 4278 4279 if ((int)skb->truesize > sk->sk_forward_alloc) 4280 goto step5; 4281 4282 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS); 4283 4284 /* Bulk data transfer: receiver */ 4285 __skb_pull(skb,tcp_header_len); 4286 __skb_queue_tail(&sk->sk_receive_queue, skb); 4287 sk_stream_set_owner_r(skb, sk); 4288 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; 4289 } 4290 4291 tcp_event_data_recv(sk, skb); 4292 4293 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) { 4294 /* Well, only one small jumplet in fast path... */ 4295 tcp_ack(sk, skb, FLAG_DATA); 4296 tcp_data_snd_check(sk); 4297 if (!inet_csk_ack_scheduled(sk)) 4298 goto no_ack; 4299 } 4300 4301 __tcp_ack_snd_check(sk, 0); 4302no_ack: 4303#ifdef CONFIG_NET_DMA 4304 if (copied_early) 4305 __skb_queue_tail(&sk->sk_async_wait_queue, skb); 4306 else 4307#endif 4308 if (eaten) 4309 __kfree_skb(skb); 4310 else 4311 sk->sk_data_ready(sk, 0); 4312 return 0; 4313 } 4314 } 4315 4316slow_path: 4317 if (len < (th->doff<<2) || tcp_checksum_complete_user(sk, skb)) 4318 goto csum_error; 4319 4320 /* 4321 * RFC1323: H1. Apply PAWS check first. 4322 */ 4323 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp && 4324 tcp_paws_discard(sk, skb)) { 4325 if (!th->rst) { 4326 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED); 4327 tcp_send_dupack(sk, skb); 4328 goto discard; 4329 } 4330 /* Resets are accepted even if PAWS failed. 4331 4332 ts_recent update must be made after we are sure 4333 that the packet is in window. 4334 */ 4335 } 4336 4337 /* 4338 * Standard slow path. 4339 */ 4340 4341 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) { 4342 /* RFC793, page 37: "In all states except SYN-SENT, all reset 4343 * (RST) segments are validated by checking their SEQ-fields." 4344 * And page 69: "If an incoming segment is not acceptable, 4345 * an acknowledgment should be sent in reply (unless the RST bit 4346 * is set, if so drop the segment and return)". 4347 */ 4348 if (!th->rst) 4349 tcp_send_dupack(sk, skb); 4350 goto discard; 4351 } 4352 4353 if (th->rst) { 4354 tcp_reset(sk); 4355 goto discard; 4356 } 4357 4358 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq); 4359 4360 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 4361 TCP_INC_STATS_BH(TCP_MIB_INERRS); 4362 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN); 4363 tcp_reset(sk); 4364 return 1; 4365 } 4366 4367step5: 4368 if (th->ack) 4369 tcp_ack(sk, skb, FLAG_SLOWPATH); 4370 4371 tcp_rcv_rtt_measure_ts(sk, skb); 4372 4373 /* Process urgent data. */ 4374 tcp_urg(sk, skb, th); 4375 4376 /* step 7: process the segment text */ 4377 tcp_data_queue(sk, skb); 4378 4379 tcp_data_snd_check(sk); 4380 tcp_ack_snd_check(sk); 4381 return 0; 4382 4383csum_error: 4384 TCP_INC_STATS_BH(TCP_MIB_INERRS); 4385 4386discard: 4387 __kfree_skb(skb); 4388 return 0; 4389} 4390 4391static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb, 4392 struct tcphdr *th, unsigned len) 4393{ 4394 struct tcp_sock *tp = tcp_sk(sk); 4395 struct inet_connection_sock *icsk = inet_csk(sk); 4396 int saved_clamp = tp->rx_opt.mss_clamp; 4397 4398 tcp_parse_options(skb, &tp->rx_opt, 0); 4399 4400 if (th->ack) { 4401 /* rfc793: 4402 * "If the state is SYN-SENT then 4403 * first check the ACK bit 4404 * If the ACK bit is set 4405 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send 4406 * a reset (unless the RST bit is set, if so drop 4407 * the segment and return)" 4408 * 4409 * We do not send data with SYN, so that RFC-correct 4410 * test reduces to: 4411 */ 4412 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt) 4413 goto reset_and_undo; 4414 4415 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 4416 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp, 4417 tcp_time_stamp)) { 4418 NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED); 4419 goto reset_and_undo; 4420 } 4421 4422 /* Now ACK is acceptable. 4423 * 4424 * "If the RST bit is set 4425 * If the ACK was acceptable then signal the user "error: 4426 * connection reset", drop the segment, enter CLOSED state, 4427 * delete TCB, and return." 4428 */ 4429 4430 if (th->rst) { 4431 tcp_reset(sk); 4432 goto discard; 4433 } 4434 4435 /* rfc793: 4436 * "fifth, if neither of the SYN or RST bits is set then 4437 * drop the segment and return." 4438 * 4439 * See note below! 4440 * --ANK(990513) 4441 */ 4442 if (!th->syn) 4443 goto discard_and_undo; 4444 4445 /* rfc793: 4446 * "If the SYN bit is on ... 4447 * are acceptable then ... 4448 * (our SYN has been ACKed), change the connection 4449 * state to ESTABLISHED..." 4450 */ 4451 4452 TCP_ECN_rcv_synack(tp, th); 4453 4454 tp->snd_wl1 = TCP_SKB_CB(skb)->seq; 4455 tcp_ack(sk, skb, FLAG_SLOWPATH); 4456 4457 /* Ok.. it's good. Set up sequence numbers and 4458 * move to established. 4459 */ 4460 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; 4461 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; 4462 4463 /* RFC1323: The window in SYN & SYN/ACK segments is 4464 * never scaled. 4465 */ 4466 tp->snd_wnd = ntohs(th->window); 4467 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(skb)->seq); 4468 4469 if (!tp->rx_opt.wscale_ok) { 4470 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0; 4471 tp->window_clamp = min(tp->window_clamp, 65535U); 4472 } 4473 4474 if (tp->rx_opt.saw_tstamp) { 4475 tp->rx_opt.tstamp_ok = 1; 4476 tp->tcp_header_len = 4477 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 4478 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; 4479 tcp_store_ts_recent(tp); 4480 } else { 4481 tp->tcp_header_len = sizeof(struct tcphdr); 4482 } 4483 4484 if (tp->rx_opt.sack_ok && sysctl_tcp_fack) 4485 tp->rx_opt.sack_ok |= 2; 4486 4487 tcp_mtup_init(sk); 4488 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 4489 tcp_initialize_rcv_mss(sk); 4490 4491 /* Remember, tcp_poll() does not lock socket! 4492 * Change state from SYN-SENT only after copied_seq 4493 * is initialized. */ 4494 tp->copied_seq = tp->rcv_nxt; 4495 smp_mb(); 4496 tcp_set_state(sk, TCP_ESTABLISHED); 4497 4498 security_inet_conn_established(sk, skb); 4499 4500 /* Make sure socket is routed, for correct metrics. */ 4501 icsk->icsk_af_ops->rebuild_header(sk); 4502 4503 tcp_init_metrics(sk); 4504 4505 tcp_init_congestion_control(sk); 4506 4507 /* Prevent spurious tcp_cwnd_restart() on first data 4508 * packet. 4509 */ 4510 tp->lsndtime = tcp_time_stamp; 4511 4512 tcp_init_buffer_space(sk); 4513 4514 if (sock_flag(sk, SOCK_KEEPOPEN)) 4515 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp)); 4516 4517 if (!tp->rx_opt.snd_wscale) 4518 __tcp_fast_path_on(tp, tp->snd_wnd); 4519 else 4520 tp->pred_flags = 0; 4521 4522 if (!sock_flag(sk, SOCK_DEAD)) { 4523 sk->sk_state_change(sk); 4524 sk_wake_async(sk, 0, POLL_OUT); 4525 } 4526 4527 if (sk->sk_write_pending || 4528 icsk->icsk_accept_queue.rskq_defer_accept || 4529 icsk->icsk_ack.pingpong) { 4530 /* Save one ACK. Data will be ready after 4531 * several ticks, if write_pending is set. 4532 * 4533 * It may be deleted, but with this feature tcpdumps 4534 * look so _wonderfully_ clever, that I was not able 4535 * to stand against the temptation 8) --ANK 4536 */ 4537 inet_csk_schedule_ack(sk); 4538 icsk->icsk_ack.lrcvtime = tcp_time_stamp; 4539 icsk->icsk_ack.ato = TCP_ATO_MIN; 4540 tcp_incr_quickack(sk); 4541 tcp_enter_quickack_mode(sk); 4542 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 4543 TCP_DELACK_MAX, TCP_RTO_MAX); 4544 4545discard: 4546 __kfree_skb(skb); 4547 return 0; 4548 } else { 4549 tcp_send_ack(sk); 4550 } 4551 return -1; 4552 } 4553 4554 /* No ACK in the segment */ 4555 4556 if (th->rst) { 4557 /* rfc793: 4558 * "If the RST bit is set 4559 * 4560 * Otherwise (no ACK) drop the segment and return." 4561 */ 4562 4563 goto discard_and_undo; 4564 } 4565 4566 /* PAWS check. */ 4567 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && tcp_paws_check(&tp->rx_opt, 0)) 4568 goto discard_and_undo; 4569 4570 if (th->syn) { 4571 /* We see SYN without ACK. It is attempt of 4572 * simultaneous connect with crossed SYNs. 4573 * Particularly, it can be connect to self. 4574 */ 4575 tcp_set_state(sk, TCP_SYN_RECV); 4576 4577 if (tp->rx_opt.saw_tstamp) { 4578 tp->rx_opt.tstamp_ok = 1; 4579 tcp_store_ts_recent(tp); 4580 tp->tcp_header_len = 4581 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 4582 } else { 4583 tp->tcp_header_len = sizeof(struct tcphdr); 4584 } 4585 4586 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; 4587 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; 4588 4589 /* RFC1323: The window in SYN & SYN/ACK segments is 4590 * never scaled. 4591 */ 4592 tp->snd_wnd = ntohs(th->window); 4593 tp->snd_wl1 = TCP_SKB_CB(skb)->seq; 4594 tp->max_window = tp->snd_wnd; 4595 4596 TCP_ECN_rcv_syn(tp, th); 4597 4598 tcp_mtup_init(sk); 4599 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 4600 tcp_initialize_rcv_mss(sk); 4601 4602 4603 tcp_send_synack(sk); 4604#if 0 4605 /* Note, we could accept data and URG from this segment. 4606 * There are no obstacles to make this. 4607 * 4608 * However, if we ignore data in ACKless segments sometimes, 4609 * we have no reasons to accept it sometimes. 4610 * Also, seems the code doing it in step6 of tcp_rcv_state_process 4611 * is not flawless. So, discard packet for sanity. 4612 * Uncomment this return to process the data. 4613 */ 4614 return -1; 4615#else 4616 goto discard; 4617#endif 4618 } 4619 /* "fifth, if neither of the SYN or RST bits is set then 4620 * drop the segment and return." 4621 */ 4622 4623discard_and_undo: 4624 tcp_clear_options(&tp->rx_opt); 4625 tp->rx_opt.mss_clamp = saved_clamp; 4626 goto discard; 4627 4628reset_and_undo: 4629 tcp_clear_options(&tp->rx_opt); 4630 tp->rx_opt.mss_clamp = saved_clamp; 4631 return 1; 4632} 4633 4634 4635/* 4636 * This function implements the receiving procedure of RFC 793 for 4637 * all states except ESTABLISHED and TIME_WAIT. 4638 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be 4639 * address independent. 4640 */ 4641 4642int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb, 4643 struct tcphdr *th, unsigned len) 4644{ 4645 struct tcp_sock *tp = tcp_sk(sk); 4646 struct inet_connection_sock *icsk = inet_csk(sk); 4647 int queued = 0; 4648 4649 tp->rx_opt.saw_tstamp = 0; 4650 4651 switch (sk->sk_state) { 4652 case TCP_CLOSE: 4653 goto discard; 4654 4655 case TCP_LISTEN: 4656 if (th->ack) 4657 return 1; 4658 4659 if (th->rst) 4660 goto discard; 4661 4662 if (th->syn) { 4663 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0) 4664 return 1; 4665 4666 /* Now we have several options: In theory there is 4667 * nothing else in the frame. KA9Q has an option to 4668 * send data with the syn, BSD accepts data with the 4669 * syn up to the [to be] advertised window and 4670 * Solaris 2.1 gives you a protocol error. For now 4671 * we just ignore it, that fits the spec precisely 4672 * and avoids incompatibilities. It would be nice in 4673 * future to drop through and process the data. 4674 * 4675 * Now that TTCP is starting to be used we ought to 4676 * queue this data. 4677 * But, this leaves one open to an easy denial of 4678 * service attack, and SYN cookies can't defend 4679 * against this problem. So, we drop the data 4680 * in the interest of security over speed unless 4681 * it's still in use. 4682 */ 4683 kfree_skb(skb); 4684 return 0; 4685 } 4686 goto discard; 4687 4688 case TCP_SYN_SENT: 4689 queued = tcp_rcv_synsent_state_process(sk, skb, th, len); 4690 if (queued >= 0) 4691 return queued; 4692 4693 /* Do step6 onward by hand. */ 4694 tcp_urg(sk, skb, th); 4695 __kfree_skb(skb); 4696 tcp_data_snd_check(sk); 4697 return 0; 4698 } 4699 4700 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp && 4701 tcp_paws_discard(sk, skb)) { 4702 if (!th->rst) { 4703 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED); 4704 tcp_send_dupack(sk, skb); 4705 goto discard; 4706 } 4707 /* Reset is accepted even if it did not pass PAWS. */ 4708 } 4709 4710 /* step 1: check sequence number */ 4711 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) { 4712 if (!th->rst) 4713 tcp_send_dupack(sk, skb); 4714 goto discard; 4715 } 4716 4717 /* step 2: check RST bit */ 4718 if (th->rst) { 4719 tcp_reset(sk); 4720 goto discard; 4721 } 4722 4723 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq); 4724 4725 /* step 3: check security and precedence [ignored] */ 4726 4727 /* step 4: 4728 * 4729 * Check for a SYN in window. 4730 */ 4731 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 4732 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN); 4733 tcp_reset(sk); 4734 return 1; 4735 } 4736 4737 /* step 5: check the ACK field */ 4738 if (th->ack) { 4739 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH); 4740 4741 switch (sk->sk_state) { 4742 case TCP_SYN_RECV: 4743 if (acceptable) { 4744 tp->copied_seq = tp->rcv_nxt; 4745 smp_mb(); 4746 tcp_set_state(sk, TCP_ESTABLISHED); 4747 sk->sk_state_change(sk); 4748 4749 /* Note, that this wakeup is only for marginal 4750 * crossed SYN case. Passively open sockets 4751 * are not waked up, because sk->sk_sleep == 4752 * NULL and sk->sk_socket == NULL. 4753 */ 4754 if (sk->sk_socket) { 4755 sk_wake_async(sk,0,POLL_OUT); 4756 } 4757 4758 tp->snd_una = TCP_SKB_CB(skb)->ack_seq; 4759 tp->snd_wnd = ntohs(th->window) << 4760 tp->rx_opt.snd_wscale; 4761 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, 4762 TCP_SKB_CB(skb)->seq); 4763 4764 /* tcp_ack considers this ACK as duplicate 4765 * and does not calculate rtt. 4766 * Fix it at least with timestamps. 4767 */ 4768 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 4769 !tp->srtt) 4770 tcp_ack_saw_tstamp(sk, 0); 4771 4772 if (tp->rx_opt.tstamp_ok) 4773 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; 4774 4775 /* Make sure socket is routed, for 4776 * correct metrics. 4777 */ 4778 icsk->icsk_af_ops->rebuild_header(sk); 4779 4780 tcp_init_metrics(sk); 4781 4782 tcp_init_congestion_control(sk); 4783 4784 /* Prevent spurious tcp_cwnd_restart() on 4785 * first data packet. 4786 */ 4787 tp->lsndtime = tcp_time_stamp; 4788 4789 tcp_mtup_init(sk); 4790 tcp_initialize_rcv_mss(sk); 4791 tcp_init_buffer_space(sk); 4792 tcp_fast_path_on(tp); 4793 } else { 4794 return 1; 4795 } 4796 break; 4797 4798 case TCP_FIN_WAIT1: 4799 if (tp->snd_una == tp->write_seq) { 4800 tcp_set_state(sk, TCP_FIN_WAIT2); 4801 sk->sk_shutdown |= SEND_SHUTDOWN; 4802 dst_confirm(sk->sk_dst_cache); 4803 4804 if (!sock_flag(sk, SOCK_DEAD)) 4805 /* Wake up lingering close() */ 4806 sk->sk_state_change(sk); 4807 else { 4808 int tmo; 4809 4810 if (tp->linger2 < 0 || 4811 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 4812 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) { 4813 tcp_done(sk); 4814 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA); 4815 return 1; 4816 } 4817 4818 tmo = tcp_fin_time(sk); 4819 if (tmo > TCP_TIMEWAIT_LEN) { 4820 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN); 4821 } else if (th->fin || sock_owned_by_user(sk)) { 4822 /* Bad case. We could lose such FIN otherwise. 4823 * It is not a big problem, but it looks confusing 4824 * and not so rare event. We still can lose it now, 4825 * if it spins in bh_lock_sock(), but it is really 4826 * marginal case. 4827 */ 4828 inet_csk_reset_keepalive_timer(sk, tmo); 4829 } else { 4830 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo); 4831 goto discard; 4832 } 4833 } 4834 } 4835 break; 4836 4837 case TCP_CLOSING: 4838 if (tp->snd_una == tp->write_seq) { 4839 tcp_time_wait(sk, TCP_TIME_WAIT, 0); 4840 goto discard; 4841 } 4842 break; 4843 4844 case TCP_LAST_ACK: 4845 if (tp->snd_una == tp->write_seq) { 4846 tcp_update_metrics(sk); 4847 tcp_done(sk); 4848 goto discard; 4849 } 4850 break; 4851 } 4852 } else 4853 goto discard; 4854 4855 /* step 6: check the URG bit */ 4856 tcp_urg(sk, skb, th); 4857 4858 /* step 7: process the segment text */ 4859 switch (sk->sk_state) { 4860 case TCP_CLOSE_WAIT: 4861 case TCP_CLOSING: 4862 case TCP_LAST_ACK: 4863 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) 4864 break; 4865 case TCP_FIN_WAIT1: 4866 case TCP_FIN_WAIT2: 4867 /* RFC 793 says to queue data in these states, 4868 * RFC 1122 says we MUST send a reset. 4869 * BSD 4.4 also does reset. 4870 */ 4871 if (sk->sk_shutdown & RCV_SHUTDOWN) { 4872 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 4873 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) { 4874 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA); 4875 tcp_reset(sk); 4876 return 1; 4877 } 4878 } 4879 /* Fall through */ 4880 case TCP_ESTABLISHED: 4881 tcp_data_queue(sk, skb); 4882 queued = 1; 4883 break; 4884 } 4885 4886 /* tcp_data could move socket to TIME-WAIT */ 4887 if (sk->sk_state != TCP_CLOSE) { 4888 tcp_data_snd_check(sk); 4889 tcp_ack_snd_check(sk); 4890 } 4891 4892 if (!queued) { 4893discard: 4894 __kfree_skb(skb); 4895 } 4896 return 0; 4897} 4898 4899EXPORT_SYMBOL(sysctl_tcp_ecn); 4900EXPORT_SYMBOL(sysctl_tcp_reordering); 4901EXPORT_SYMBOL(tcp_parse_options); 4902EXPORT_SYMBOL(tcp_rcv_established); 4903EXPORT_SYMBOL(tcp_rcv_state_process); 4904EXPORT_SYMBOL(tcp_initialize_rcv_mss); 4905