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 75 int sysctl_tcp_timestamps __read_mostly = 1; 76 int sysctl_tcp_window_scaling __read_mostly = 1; 77 int sysctl_tcp_sack __read_mostly = 1; 78 int sysctl_tcp_fack __read_mostly = 1; 79 int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH; 80 int sysctl_tcp_ecn __read_mostly; 81 int sysctl_tcp_dsack __read_mostly = 1; 82 int sysctl_tcp_app_win __read_mostly = 31; 83 int sysctl_tcp_adv_win_scale __read_mostly = 2; 84 85 int sysctl_tcp_stdurg __read_mostly; 86 int sysctl_tcp_rfc1337 __read_mostly; 87 int sysctl_tcp_max_orphans __read_mostly = NR_FILE; 88 int sysctl_tcp_frto __read_mostly; 89 int sysctl_tcp_frto_response __read_mostly; 90 int sysctl_tcp_nometrics_save __read_mostly; 91 92 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1; 93 int 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 */ 122 static 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 } 168 169 static 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 } 179 180 void 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 */ 191 192 static 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 */ 202 203 static 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. */ 238 static 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 } 254 255 static 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. */ 282 283 static 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 */ 301 static 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. */ 335 static 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 */ 361 void 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 */ 384 static 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 } 416 417 static 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); 426 427 new_measure: 428 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd; 429 tp->rcv_rtt_est.time = tcp_time_stamp; 430 } 431 432 static 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 */ 445 void 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 } 493 494 new_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 */ 509 static 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 */ 564 static 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 */ 630 static 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 */ 655 static 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 */ 665 void 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 */ 770 void 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. */ 793 794 static 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; 852 853 reset: 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 } 864 865 static 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 */ 942 static int 943 tcp_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 found_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 found_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 found_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 (found_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 int dup_sack = (found_dup_sack && (i == first_sack_index)); 1062 1063 skb = cached_skb; 1064 fack_count = cached_fack_count; 1065 1066 /* Event "B" in the comment above. */ 1067 if (after(end_seq, tp->high_seq)) 1068 flag |= FLAG_DATA_LOST; 1069 1070 tcp_for_write_queue_from(skb, sk) { 1071 int in_sack, pcount; 1072 u8 sacked; 1073 1074 if (skb == tcp_send_head(sk)) 1075 break; 1076 1077 cached_skb = skb; 1078 cached_fack_count = fack_count; 1079 if (i == first_sack_index) { 1080 tp->fastpath_skb_hint = skb; 1081 tp->fastpath_cnt_hint = fack_count; 1082 } 1083 1084 /* The retransmission queue is always in order, so 1085 * we can short-circuit the walk early. 1086 */ 1087 if (!before(TCP_SKB_CB(skb)->seq, end_seq)) 1088 break; 1089 1090 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && 1091 !before(end_seq, TCP_SKB_CB(skb)->end_seq); 1092 1093 pcount = tcp_skb_pcount(skb); 1094 1095 if (pcount > 1 && !in_sack && 1096 after(TCP_SKB_CB(skb)->end_seq, start_seq)) { 1097 unsigned int pkt_len; 1098 1099 in_sack = !after(start_seq, 1100 TCP_SKB_CB(skb)->seq); 1101 1102 if (!in_sack) 1103 pkt_len = (start_seq - 1104 TCP_SKB_CB(skb)->seq); 1105 else 1106 pkt_len = (end_seq - 1107 TCP_SKB_CB(skb)->seq); 1108 if (tcp_fragment(sk, skb, pkt_len, skb_shinfo(skb)->gso_size)) 1109 break; 1110 pcount = tcp_skb_pcount(skb); 1111 } 1112 1113 fack_count += pcount; 1114 1115 sacked = TCP_SKB_CB(skb)->sacked; 1116 1117 /* Account D-SACK for retransmitted packet. */ 1118 if ((dup_sack && in_sack) && 1119 (sacked & TCPCB_RETRANS) && 1120 after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker)) 1121 tp->undo_retrans--; 1122 1123 /* The frame is ACKed. */ 1124 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) { 1125 if (sacked&TCPCB_RETRANS) { 1126 if ((dup_sack && in_sack) && 1127 (sacked&TCPCB_SACKED_ACKED)) 1128 reord = min(fack_count, reord); 1129 } else { 1130 /* If it was in a hole, we detected reordering. */ 1131 if (fack_count < prior_fackets && 1132 !(sacked&TCPCB_SACKED_ACKED)) 1133 reord = min(fack_count, reord); 1134 } 1135 1136 /* Nothing to do; acked frame is about to be dropped. */ 1137 continue; 1138 } 1139 1140 if ((sacked&TCPCB_SACKED_RETRANS) && 1141 after(end_seq, TCP_SKB_CB(skb)->ack_seq) && 1142 (!lost_retrans || after(end_seq, lost_retrans))) 1143 lost_retrans = end_seq; 1144 1145 if (!in_sack) 1146 continue; 1147 1148 if (!(sacked&TCPCB_SACKED_ACKED)) { 1149 if (sacked & TCPCB_SACKED_RETRANS) { 1150 /* If the segment is not tagged as lost, 1151 * we do not clear RETRANS, believing 1152 * that retransmission is still in flight. 1153 */ 1154 if (sacked & TCPCB_LOST) { 1155 TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS); 1156 tp->lost_out -= tcp_skb_pcount(skb); 1157 tp->retrans_out -= tcp_skb_pcount(skb); 1158 1159 /* clear lost hint */ 1160 tp->retransmit_skb_hint = NULL; 1161 } 1162 } else { 1163 /* New sack for not retransmitted frame, 1164 * which was in hole. It is reordering. 1165 */ 1166 if (!(sacked & TCPCB_RETRANS) && 1167 fack_count < prior_fackets) 1168 reord = min(fack_count, reord); 1169 1170 if (sacked & TCPCB_LOST) { 1171 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; 1172 tp->lost_out -= tcp_skb_pcount(skb); 1173 1174 /* clear lost hint */ 1175 tp->retransmit_skb_hint = NULL; 1176 } 1177 /* SACK enhanced F-RTO detection. 1178 * Set flag if and only if non-rexmitted 1179 * segments below frto_highmark are 1180 * SACKed (RFC4138; Appendix B). 1181 * Clearing correct due to in-order walk 1182 */ 1183 if (after(end_seq, tp->frto_highmark)) { 1184 flag &= ~FLAG_ONLY_ORIG_SACKED; 1185 } else { 1186 if (!(sacked & TCPCB_RETRANS)) 1187 flag |= FLAG_ONLY_ORIG_SACKED; 1188 } 1189 } 1190 1191 TCP_SKB_CB(skb)->sacked |= TCPCB_SACKED_ACKED; 1192 flag |= FLAG_DATA_SACKED; 1193 tp->sacked_out += tcp_skb_pcount(skb); 1194 1195 if (fack_count > tp->fackets_out) 1196 tp->fackets_out = fack_count; 1197 } else { 1198 if (dup_sack && (sacked&TCPCB_RETRANS)) 1199 reord = min(fack_count, reord); 1200 } 1201 1202 /* D-SACK. We can detect redundant retransmission 1203 * in S|R and plain R frames and clear it. 1204 * undo_retrans is decreased above, L|R frames 1205 * are accounted above as well. 1206 */ 1207 if (dup_sack && 1208 (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS)) { 1209 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 1210 tp->retrans_out -= tcp_skb_pcount(skb); 1211 tp->retransmit_skb_hint = NULL; 1212 } 1213 } 1214 } 1215 1216 /* Check for lost retransmit. This superb idea is 1217 * borrowed from "ratehalving". Event "C". 1218 * Later note: FACK people cheated me again 8), 1219 * we have to account for reordering! Ugly, 1220 * but should help. 1221 */ 1222 if (lost_retrans && icsk->icsk_ca_state == TCP_CA_Recovery) { 1223 struct sk_buff *skb; 1224 1225 tcp_for_write_queue(skb, sk) { 1226 if (skb == tcp_send_head(sk)) 1227 break; 1228 if (after(TCP_SKB_CB(skb)->seq, lost_retrans)) 1229 break; 1230 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 1231 continue; 1232 if ((TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) && 1233 after(lost_retrans, TCP_SKB_CB(skb)->ack_seq) && 1234 (IsFack(tp) || 1235 !before(lost_retrans, 1236 TCP_SKB_CB(skb)->ack_seq + tp->reordering * 1237 tp->mss_cache))) { 1238 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 1239 tp->retrans_out -= tcp_skb_pcount(skb); 1240 1241 /* clear lost hint */ 1242 tp->retransmit_skb_hint = NULL; 1243 1244 if (!(TCP_SKB_CB(skb)->sacked&(TCPCB_LOST|TCPCB_SACKED_ACKED))) { 1245 tp->lost_out += tcp_skb_pcount(skb); 1246 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 1247 flag |= FLAG_DATA_SACKED; 1248 NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT); 1249 } 1250 } 1251 } 1252 } 1253 1254 tp->left_out = tp->sacked_out + tp->lost_out; 1255 1256 if ((reord < tp->fackets_out) && icsk->icsk_ca_state != TCP_CA_Loss && 1257 (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark))) 1258 tcp_update_reordering(sk, ((tp->fackets_out + 1) - reord), 0); 1259 1260 #if FASTRETRANS_DEBUG > 0 1261 BUG_TRAP((int)tp->sacked_out >= 0); 1262 BUG_TRAP((int)tp->lost_out >= 0); 1263 BUG_TRAP((int)tp->retrans_out >= 0); 1264 BUG_TRAP((int)tcp_packets_in_flight(tp) >= 0); 1265 #endif 1266 return flag; 1267 } 1268 1269 /* F-RTO can only be used if TCP has never retransmitted anything other than 1270 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here) 1271 */ 1272 int tcp_use_frto(struct sock *sk) 1273 { 1274 const struct tcp_sock *tp = tcp_sk(sk); 1275 struct sk_buff *skb; 1276 1277 if (!sysctl_tcp_frto) 1278 return 0; 1279 1280 if (IsSackFrto()) 1281 return 1; 1282 1283 /* Avoid expensive walking of rexmit queue if possible */ 1284 if (tp->retrans_out > 1) 1285 return 0; 1286 1287 skb = tcp_write_queue_head(sk); 1288 skb = tcp_write_queue_next(sk, skb); /* Skips head */ 1289 tcp_for_write_queue_from(skb, sk) { 1290 if (skb == tcp_send_head(sk)) 1291 break; 1292 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS) 1293 return 0; 1294 /* Short-circuit when first non-SACKed skb has been checked */ 1295 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) 1296 break; 1297 } 1298 return 1; 1299 } 1300 1301 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO 1302 * recovery a bit and use heuristics in tcp_process_frto() to detect if 1303 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to 1304 * keep retrans_out counting accurate (with SACK F-RTO, other than head 1305 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS 1306 * bits are handled if the Loss state is really to be entered (in 1307 * tcp_enter_frto_loss). 1308 * 1309 * Do like tcp_enter_loss() would; when RTO expires the second time it 1310 * does: 1311 * "Reduce ssthresh if it has not yet been made inside this window." 1312 */ 1313 void tcp_enter_frto(struct sock *sk) 1314 { 1315 const struct inet_connection_sock *icsk = inet_csk(sk); 1316 struct tcp_sock *tp = tcp_sk(sk); 1317 struct sk_buff *skb; 1318 1319 if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) || 1320 tp->snd_una == tp->high_seq || 1321 ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) && 1322 !icsk->icsk_retransmits)) { 1323 tp->prior_ssthresh = tcp_current_ssthresh(sk); 1324 /* Our state is too optimistic in ssthresh() call because cwnd 1325 * is not reduced until tcp_enter_frto_loss() when previous FRTO 1326 * recovery has not yet completed. Pattern would be this: RTO, 1327 * Cumulative ACK, RTO (2xRTO for the same segment does not end 1328 * up here twice). 1329 * RFC4138 should be more specific on what to do, even though 1330 * RTO is quite unlikely to occur after the first Cumulative ACK 1331 * due to back-off and complexity of triggering events ... 1332 */ 1333 if (tp->frto_counter) { 1334 u32 stored_cwnd; 1335 stored_cwnd = tp->snd_cwnd; 1336 tp->snd_cwnd = 2; 1337 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 1338 tp->snd_cwnd = stored_cwnd; 1339 } else { 1340 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 1341 } 1342 /* ... in theory, cong.control module could do "any tricks" in 1343 * ssthresh(), which means that ca_state, lost bits and lost_out 1344 * counter would have to be faked before the call occurs. We 1345 * consider that too expensive, unlikely and hacky, so modules 1346 * using these in ssthresh() must deal these incompatibility 1347 * issues if they receives CA_EVENT_FRTO and frto_counter != 0 1348 */ 1349 tcp_ca_event(sk, CA_EVENT_FRTO); 1350 } 1351 1352 tp->undo_marker = tp->snd_una; 1353 tp->undo_retrans = 0; 1354 1355 skb = tcp_write_queue_head(sk); 1356 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 1357 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 1358 tp->retrans_out -= tcp_skb_pcount(skb); 1359 } 1360 tcp_sync_left_out(tp); 1361 1362 /* Earlier loss recovery underway (see RFC4138; Appendix B). 1363 * The last condition is necessary at least in tp->frto_counter case. 1364 */ 1365 if (IsSackFrto() && (tp->frto_counter || 1366 ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) && 1367 after(tp->high_seq, tp->snd_una)) { 1368 tp->frto_highmark = tp->high_seq; 1369 } else { 1370 tp->frto_highmark = tp->snd_nxt; 1371 } 1372 tcp_set_ca_state(sk, TCP_CA_Disorder); 1373 tp->high_seq = tp->snd_nxt; 1374 tp->frto_counter = 1; 1375 } 1376 1377 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO, 1378 * which indicates that we should follow the traditional RTO recovery, 1379 * i.e. mark everything lost and do go-back-N retransmission. 1380 */ 1381 static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag) 1382 { 1383 struct tcp_sock *tp = tcp_sk(sk); 1384 struct sk_buff *skb; 1385 int cnt = 0; 1386 1387 tp->sacked_out = 0; 1388 tp->lost_out = 0; 1389 tp->fackets_out = 0; 1390 tp->retrans_out = 0; 1391 1392 tcp_for_write_queue(skb, sk) { 1393 if (skb == tcp_send_head(sk)) 1394 break; 1395 cnt += tcp_skb_pcount(skb); 1396 /* 1397 * Count the retransmission made on RTO correctly (only when 1398 * waiting for the first ACK and did not get it)... 1399 */ 1400 if ((tp->frto_counter == 1) && !(flag&FLAG_DATA_ACKED)) { 1401 tp->retrans_out += tcp_skb_pcount(skb); 1402 /* ...enter this if branch just for the first segment */ 1403 flag |= FLAG_DATA_ACKED; 1404 } else { 1405 TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS); 1406 } 1407 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) { 1408 1409 /* Do not mark those segments lost that were 1410 * forward transmitted after RTO 1411 */ 1412 if (!after(TCP_SKB_CB(skb)->end_seq, 1413 tp->frto_highmark)) { 1414 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 1415 tp->lost_out += tcp_skb_pcount(skb); 1416 } 1417 } else { 1418 tp->sacked_out += tcp_skb_pcount(skb); 1419 tp->fackets_out = cnt; 1420 } 1421 } 1422 tcp_sync_left_out(tp); 1423 1424 tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments; 1425 tp->snd_cwnd_cnt = 0; 1426 tp->snd_cwnd_stamp = tcp_time_stamp; 1427 tp->undo_marker = 0; 1428 tp->frto_counter = 0; 1429 1430 tp->reordering = min_t(unsigned int, tp->reordering, 1431 sysctl_tcp_reordering); 1432 tcp_set_ca_state(sk, TCP_CA_Loss); 1433 tp->high_seq = tp->frto_highmark; 1434 TCP_ECN_queue_cwr(tp); 1435 1436 clear_all_retrans_hints(tp); 1437 } 1438 1439 void tcp_clear_retrans(struct tcp_sock *tp) 1440 { 1441 tp->left_out = 0; 1442 tp->retrans_out = 0; 1443 1444 tp->fackets_out = 0; 1445 tp->sacked_out = 0; 1446 tp->lost_out = 0; 1447 1448 tp->undo_marker = 0; 1449 tp->undo_retrans = 0; 1450 } 1451 1452 /* Enter Loss state. If "how" is not zero, forget all SACK information 1453 * and reset tags completely, otherwise preserve SACKs. If receiver 1454 * dropped its ofo queue, we will know this due to reneging detection. 1455 */ 1456 void tcp_enter_loss(struct sock *sk, int how) 1457 { 1458 const struct inet_connection_sock *icsk = inet_csk(sk); 1459 struct tcp_sock *tp = tcp_sk(sk); 1460 struct sk_buff *skb; 1461 int cnt = 0; 1462 1463 /* Reduce ssthresh if it has not yet been made inside this window. */ 1464 if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq || 1465 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) { 1466 tp->prior_ssthresh = tcp_current_ssthresh(sk); 1467 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 1468 tcp_ca_event(sk, CA_EVENT_LOSS); 1469 } 1470 tp->snd_cwnd = 1; 1471 tp->snd_cwnd_cnt = 0; 1472 tp->snd_cwnd_stamp = tcp_time_stamp; 1473 1474 tp->bytes_acked = 0; 1475 tcp_clear_retrans(tp); 1476 1477 /* Push undo marker, if it was plain RTO and nothing 1478 * was retransmitted. */ 1479 if (!how) 1480 tp->undo_marker = tp->snd_una; 1481 1482 tcp_for_write_queue(skb, sk) { 1483 if (skb == tcp_send_head(sk)) 1484 break; 1485 cnt += tcp_skb_pcount(skb); 1486 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS) 1487 tp->undo_marker = 0; 1488 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED; 1489 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) { 1490 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED; 1491 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 1492 tp->lost_out += tcp_skb_pcount(skb); 1493 } else { 1494 tp->sacked_out += tcp_skb_pcount(skb); 1495 tp->fackets_out = cnt; 1496 } 1497 } 1498 tcp_sync_left_out(tp); 1499 1500 tp->reordering = min_t(unsigned int, tp->reordering, 1501 sysctl_tcp_reordering); 1502 tcp_set_ca_state(sk, TCP_CA_Loss); 1503 tp->high_seq = tp->snd_nxt; 1504 TCP_ECN_queue_cwr(tp); 1505 /* Abort FRTO algorithm if one is in progress */ 1506 tp->frto_counter = 0; 1507 1508 clear_all_retrans_hints(tp); 1509 } 1510 1511 static int tcp_check_sack_reneging(struct sock *sk) 1512 { 1513 struct sk_buff *skb; 1514 1515 /* If ACK arrived pointing to a remembered SACK, 1516 * it means that our remembered SACKs do not reflect 1517 * real state of receiver i.e. 1518 * receiver _host_ is heavily congested (or buggy). 1519 * Do processing similar to RTO timeout. 1520 */ 1521 if ((skb = tcp_write_queue_head(sk)) != NULL && 1522 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) { 1523 struct inet_connection_sock *icsk = inet_csk(sk); 1524 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING); 1525 1526 tcp_enter_loss(sk, 1); 1527 icsk->icsk_retransmits++; 1528 tcp_retransmit_skb(sk, tcp_write_queue_head(sk)); 1529 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 1530 icsk->icsk_rto, TCP_RTO_MAX); 1531 return 1; 1532 } 1533 return 0; 1534 } 1535 1536 static inline int tcp_fackets_out(struct tcp_sock *tp) 1537 { 1538 return IsReno(tp) ? tp->sacked_out+1 : tp->fackets_out; 1539 } 1540 1541 static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb) 1542 { 1543 return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto); 1544 } 1545 1546 static inline int tcp_head_timedout(struct sock *sk) 1547 { 1548 struct tcp_sock *tp = tcp_sk(sk); 1549 1550 return tp->packets_out && 1551 tcp_skb_timedout(sk, tcp_write_queue_head(sk)); 1552 } 1553 1554 /* Linux NewReno/SACK/FACK/ECN state machine. 1555 * -------------------------------------- 1556 * 1557 * "Open" Normal state, no dubious events, fast path. 1558 * "Disorder" In all the respects it is "Open", 1559 * but requires a bit more attention. It is entered when 1560 * we see some SACKs or dupacks. It is split of "Open" 1561 * mainly to move some processing from fast path to slow one. 1562 * "CWR" CWND was reduced due to some Congestion Notification event. 1563 * It can be ECN, ICMP source quench, local device congestion. 1564 * "Recovery" CWND was reduced, we are fast-retransmitting. 1565 * "Loss" CWND was reduced due to RTO timeout or SACK reneging. 1566 * 1567 * tcp_fastretrans_alert() is entered: 1568 * - each incoming ACK, if state is not "Open" 1569 * - when arrived ACK is unusual, namely: 1570 * * SACK 1571 * * Duplicate ACK. 1572 * * ECN ECE. 1573 * 1574 * Counting packets in flight is pretty simple. 1575 * 1576 * in_flight = packets_out - left_out + retrans_out 1577 * 1578 * packets_out is SND.NXT-SND.UNA counted in packets. 1579 * 1580 * retrans_out is number of retransmitted segments. 1581 * 1582 * left_out is number of segments left network, but not ACKed yet. 1583 * 1584 * left_out = sacked_out + lost_out 1585 * 1586 * sacked_out: Packets, which arrived to receiver out of order 1587 * and hence not ACKed. With SACKs this number is simply 1588 * amount of SACKed data. Even without SACKs 1589 * it is easy to give pretty reliable estimate of this number, 1590 * counting duplicate ACKs. 1591 * 1592 * lost_out: Packets lost by network. TCP has no explicit 1593 * "loss notification" feedback from network (for now). 1594 * It means that this number can be only _guessed_. 1595 * Actually, it is the heuristics to predict lossage that 1596 * distinguishes different algorithms. 1597 * 1598 * F.e. after RTO, when all the queue is considered as lost, 1599 * lost_out = packets_out and in_flight = retrans_out. 1600 * 1601 * Essentially, we have now two algorithms counting 1602 * lost packets. 1603 * 1604 * FACK: It is the simplest heuristics. As soon as we decided 1605 * that something is lost, we decide that _all_ not SACKed 1606 * packets until the most forward SACK are lost. I.e. 1607 * lost_out = fackets_out - sacked_out and left_out = fackets_out. 1608 * It is absolutely correct estimate, if network does not reorder 1609 * packets. And it loses any connection to reality when reordering 1610 * takes place. We use FACK by default until reordering 1611 * is suspected on the path to this destination. 1612 * 1613 * NewReno: when Recovery is entered, we assume that one segment 1614 * is lost (classic Reno). While we are in Recovery and 1615 * a partial ACK arrives, we assume that one more packet 1616 * is lost (NewReno). This heuristics are the same in NewReno 1617 * and SACK. 1618 * 1619 * Imagine, that's all! Forget about all this shamanism about CWND inflation 1620 * deflation etc. CWND is real congestion window, never inflated, changes 1621 * only according to classic VJ rules. 1622 * 1623 * Really tricky (and requiring careful tuning) part of algorithm 1624 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue(). 1625 * The first determines the moment _when_ we should reduce CWND and, 1626 * hence, slow down forward transmission. In fact, it determines the moment 1627 * when we decide that hole is caused by loss, rather than by a reorder. 1628 * 1629 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill 1630 * holes, caused by lost packets. 1631 * 1632 * And the most logically complicated part of algorithm is undo 1633 * heuristics. We detect false retransmits due to both too early 1634 * fast retransmit (reordering) and underestimated RTO, analyzing 1635 * timestamps and D-SACKs. When we detect that some segments were 1636 * retransmitted by mistake and CWND reduction was wrong, we undo 1637 * window reduction and abort recovery phase. This logic is hidden 1638 * inside several functions named tcp_try_undo_<something>. 1639 */ 1640 1641 /* This function decides, when we should leave Disordered state 1642 * and enter Recovery phase, reducing congestion window. 1643 * 1644 * Main question: may we further continue forward transmission 1645 * with the same cwnd? 1646 */ 1647 static int tcp_time_to_recover(struct sock *sk) 1648 { 1649 struct tcp_sock *tp = tcp_sk(sk); 1650 __u32 packets_out; 1651 1652 /* Do not perform any recovery during FRTO algorithm */ 1653 if (tp->frto_counter) 1654 return 0; 1655 1656 /* Trick#1: The loss is proven. */ 1657 if (tp->lost_out) 1658 return 1; 1659 1660 /* Not-A-Trick#2 : Classic rule... */ 1661 if (tcp_fackets_out(tp) > tp->reordering) 1662 return 1; 1663 1664 /* Trick#3 : when we use RFC2988 timer restart, fast 1665 * retransmit can be triggered by timeout of queue head. 1666 */ 1667 if (tcp_head_timedout(sk)) 1668 return 1; 1669 1670 /* Trick#4: It is still not OK... But will it be useful to delay 1671 * recovery more? 1672 */ 1673 packets_out = tp->packets_out; 1674 if (packets_out <= tp->reordering && 1675 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) && 1676 !tcp_may_send_now(sk)) { 1677 /* We have nothing to send. This connection is limited 1678 * either by receiver window or by application. 1679 */ 1680 return 1; 1681 } 1682 1683 return 0; 1684 } 1685 1686 /* If we receive more dupacks than we expected counting segments 1687 * in assumption of absent reordering, interpret this as reordering. 1688 * The only another reason could be bug in receiver TCP. 1689 */ 1690 static void tcp_check_reno_reordering(struct sock *sk, const int addend) 1691 { 1692 struct tcp_sock *tp = tcp_sk(sk); 1693 u32 holes; 1694 1695 holes = max(tp->lost_out, 1U); 1696 holes = min(holes, tp->packets_out); 1697 1698 if ((tp->sacked_out + holes) > tp->packets_out) { 1699 tp->sacked_out = tp->packets_out - holes; 1700 tcp_update_reordering(sk, tp->packets_out + addend, 0); 1701 } 1702 } 1703 1704 /* Emulate SACKs for SACKless connection: account for a new dupack. */ 1705 1706 static void tcp_add_reno_sack(struct sock *sk) 1707 { 1708 struct tcp_sock *tp = tcp_sk(sk); 1709 tp->sacked_out++; 1710 tcp_check_reno_reordering(sk, 0); 1711 tcp_sync_left_out(tp); 1712 } 1713 1714 /* Account for ACK, ACKing some data in Reno Recovery phase. */ 1715 1716 static void tcp_remove_reno_sacks(struct sock *sk, int acked) 1717 { 1718 struct tcp_sock *tp = tcp_sk(sk); 1719 1720 if (acked > 0) { 1721 /* One ACK acked hole. The rest eat duplicate ACKs. */ 1722 if (acked-1 >= tp->sacked_out) 1723 tp->sacked_out = 0; 1724 else 1725 tp->sacked_out -= acked-1; 1726 } 1727 tcp_check_reno_reordering(sk, acked); 1728 tcp_sync_left_out(tp); 1729 } 1730 1731 static inline void tcp_reset_reno_sack(struct tcp_sock *tp) 1732 { 1733 tp->sacked_out = 0; 1734 tp->left_out = tp->lost_out; 1735 } 1736 1737 /* Mark head of queue up as lost. */ 1738 static void tcp_mark_head_lost(struct sock *sk, 1739 int packets, u32 high_seq) 1740 { 1741 struct tcp_sock *tp = tcp_sk(sk); 1742 struct sk_buff *skb; 1743 int cnt; 1744 1745 BUG_TRAP(packets <= tp->packets_out); 1746 if (tp->lost_skb_hint) { 1747 skb = tp->lost_skb_hint; 1748 cnt = tp->lost_cnt_hint; 1749 } else { 1750 skb = tcp_write_queue_head(sk); 1751 cnt = 0; 1752 } 1753 1754 tcp_for_write_queue_from(skb, sk) { 1755 if (skb == tcp_send_head(sk)) 1756 break; 1757 /* TODO: do this better */ 1758 /* this is not the most efficient way to do this... */ 1759 tp->lost_skb_hint = skb; 1760 tp->lost_cnt_hint = cnt; 1761 cnt += tcp_skb_pcount(skb); 1762 if (cnt > packets || after(TCP_SKB_CB(skb)->end_seq, high_seq)) 1763 break; 1764 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) { 1765 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 1766 tp->lost_out += tcp_skb_pcount(skb); 1767 1768 /* clear xmit_retransmit_queue hints 1769 * if this is beyond hint */ 1770 if (tp->retransmit_skb_hint != NULL && 1771 before(TCP_SKB_CB(skb)->seq, 1772 TCP_SKB_CB(tp->retransmit_skb_hint)->seq)) 1773 tp->retransmit_skb_hint = NULL; 1774 1775 } 1776 } 1777 tcp_sync_left_out(tp); 1778 } 1779 1780 /* Account newly detected lost packet(s) */ 1781 1782 static void tcp_update_scoreboard(struct sock *sk) 1783 { 1784 struct tcp_sock *tp = tcp_sk(sk); 1785 1786 if (IsFack(tp)) { 1787 int lost = tp->fackets_out - tp->reordering; 1788 if (lost <= 0) 1789 lost = 1; 1790 tcp_mark_head_lost(sk, lost, tp->high_seq); 1791 } else { 1792 tcp_mark_head_lost(sk, 1, tp->high_seq); 1793 } 1794 1795 /* New heuristics: it is possible only after we switched 1796 * to restart timer each time when something is ACKed. 1797 * Hence, we can detect timed out packets during fast 1798 * retransmit without falling to slow start. 1799 */ 1800 if (!IsReno(tp) && tcp_head_timedout(sk)) { 1801 struct sk_buff *skb; 1802 1803 skb = tp->scoreboard_skb_hint ? tp->scoreboard_skb_hint 1804 : tcp_write_queue_head(sk); 1805 1806 tcp_for_write_queue_from(skb, sk) { 1807 if (skb == tcp_send_head(sk)) 1808 break; 1809 if (!tcp_skb_timedout(sk, skb)) 1810 break; 1811 1812 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) { 1813 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 1814 tp->lost_out += tcp_skb_pcount(skb); 1815 1816 /* clear xmit_retrans hint */ 1817 if (tp->retransmit_skb_hint && 1818 before(TCP_SKB_CB(skb)->seq, 1819 TCP_SKB_CB(tp->retransmit_skb_hint)->seq)) 1820 1821 tp->retransmit_skb_hint = NULL; 1822 } 1823 } 1824 1825 tp->scoreboard_skb_hint = skb; 1826 1827 tcp_sync_left_out(tp); 1828 } 1829 } 1830 1831 /* CWND moderation, preventing bursts due to too big ACKs 1832 * in dubious situations. 1833 */ 1834 static inline void tcp_moderate_cwnd(struct tcp_sock *tp) 1835 { 1836 tp->snd_cwnd = min(tp->snd_cwnd, 1837 tcp_packets_in_flight(tp)+tcp_max_burst(tp)); 1838 tp->snd_cwnd_stamp = tcp_time_stamp; 1839 } 1840 1841 /* Lower bound on congestion window is slow start threshold 1842 * unless congestion avoidance choice decides to overide it. 1843 */ 1844 static inline u32 tcp_cwnd_min(const struct sock *sk) 1845 { 1846 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 1847 1848 return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh; 1849 } 1850 1851 /* Decrease cwnd each second ack. */ 1852 static void tcp_cwnd_down(struct sock *sk) 1853 { 1854 struct tcp_sock *tp = tcp_sk(sk); 1855 int decr = tp->snd_cwnd_cnt + 1; 1856 1857 tp->snd_cwnd_cnt = decr&1; 1858 decr >>= 1; 1859 1860 if (decr && tp->snd_cwnd > tcp_cwnd_min(sk)) 1861 tp->snd_cwnd -= decr; 1862 1863 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp)+1); 1864 tp->snd_cwnd_stamp = tcp_time_stamp; 1865 } 1866 1867 /* Nothing was retransmitted or returned timestamp is less 1868 * than timestamp of the first retransmission. 1869 */ 1870 static inline int tcp_packet_delayed(struct tcp_sock *tp) 1871 { 1872 return !tp->retrans_stamp || 1873 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 1874 (__s32)(tp->rx_opt.rcv_tsecr - tp->retrans_stamp) < 0); 1875 } 1876 1877 /* Undo procedures. */ 1878 1879 #if FASTRETRANS_DEBUG > 1 1880 static void DBGUNDO(struct sock *sk, const char *msg) 1881 { 1882 struct tcp_sock *tp = tcp_sk(sk); 1883 struct inet_sock *inet = inet_sk(sk); 1884 1885 printk(KERN_DEBUG "Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n", 1886 msg, 1887 NIPQUAD(inet->daddr), ntohs(inet->dport), 1888 tp->snd_cwnd, tp->left_out, 1889 tp->snd_ssthresh, tp->prior_ssthresh, 1890 tp->packets_out); 1891 } 1892 #else 1893 #define DBGUNDO(x...) do { } while (0) 1894 #endif 1895 1896 static void tcp_undo_cwr(struct sock *sk, const int undo) 1897 { 1898 struct tcp_sock *tp = tcp_sk(sk); 1899 1900 if (tp->prior_ssthresh) { 1901 const struct inet_connection_sock *icsk = inet_csk(sk); 1902 1903 if (icsk->icsk_ca_ops->undo_cwnd) 1904 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk); 1905 else 1906 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh<<1); 1907 1908 if (undo && tp->prior_ssthresh > tp->snd_ssthresh) { 1909 tp->snd_ssthresh = tp->prior_ssthresh; 1910 TCP_ECN_withdraw_cwr(tp); 1911 } 1912 } else { 1913 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh); 1914 } 1915 tcp_moderate_cwnd(tp); 1916 tp->snd_cwnd_stamp = tcp_time_stamp; 1917 1918 /* There is something screwy going on with the retrans hints after 1919 an undo */ 1920 clear_all_retrans_hints(tp); 1921 } 1922 1923 static inline int tcp_may_undo(struct tcp_sock *tp) 1924 { 1925 return tp->undo_marker && 1926 (!tp->undo_retrans || tcp_packet_delayed(tp)); 1927 } 1928 1929 /* People celebrate: "We love our President!" */ 1930 static int tcp_try_undo_recovery(struct sock *sk) 1931 { 1932 struct tcp_sock *tp = tcp_sk(sk); 1933 1934 if (tcp_may_undo(tp)) { 1935 /* Happy end! We did not retransmit anything 1936 * or our original transmission succeeded. 1937 */ 1938 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans"); 1939 tcp_undo_cwr(sk, 1); 1940 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss) 1941 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO); 1942 else 1943 NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO); 1944 tp->undo_marker = 0; 1945 } 1946 if (tp->snd_una == tp->high_seq && IsReno(tp)) { 1947 /* Hold old state until something *above* high_seq 1948 * is ACKed. For Reno it is MUST to prevent false 1949 * fast retransmits (RFC2582). SACK TCP is safe. */ 1950 tcp_moderate_cwnd(tp); 1951 return 1; 1952 } 1953 tcp_set_ca_state(sk, TCP_CA_Open); 1954 return 0; 1955 } 1956 1957 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */ 1958 static void tcp_try_undo_dsack(struct sock *sk) 1959 { 1960 struct tcp_sock *tp = tcp_sk(sk); 1961 1962 if (tp->undo_marker && !tp->undo_retrans) { 1963 DBGUNDO(sk, "D-SACK"); 1964 tcp_undo_cwr(sk, 1); 1965 tp->undo_marker = 0; 1966 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO); 1967 } 1968 } 1969 1970 /* Undo during fast recovery after partial ACK. */ 1971 1972 static int tcp_try_undo_partial(struct sock *sk, int acked) 1973 { 1974 struct tcp_sock *tp = tcp_sk(sk); 1975 /* Partial ACK arrived. Force Hoe's retransmit. */ 1976 int failed = IsReno(tp) || tp->fackets_out>tp->reordering; 1977 1978 if (tcp_may_undo(tp)) { 1979 /* Plain luck! Hole if filled with delayed 1980 * packet, rather than with a retransmit. 1981 */ 1982 if (tp->retrans_out == 0) 1983 tp->retrans_stamp = 0; 1984 1985 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1); 1986 1987 DBGUNDO(sk, "Hoe"); 1988 tcp_undo_cwr(sk, 0); 1989 NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO); 1990 1991 /* So... Do not make Hoe's retransmit yet. 1992 * If the first packet was delayed, the rest 1993 * ones are most probably delayed as well. 1994 */ 1995 failed = 0; 1996 } 1997 return failed; 1998 } 1999 2000 /* Undo during loss recovery after partial ACK. */ 2001 static int tcp_try_undo_loss(struct sock *sk) 2002 { 2003 struct tcp_sock *tp = tcp_sk(sk); 2004 2005 if (tcp_may_undo(tp)) { 2006 struct sk_buff *skb; 2007 tcp_for_write_queue(skb, sk) { 2008 if (skb == tcp_send_head(sk)) 2009 break; 2010 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; 2011 } 2012 2013 clear_all_retrans_hints(tp); 2014 2015 DBGUNDO(sk, "partial loss"); 2016 tp->lost_out = 0; 2017 tp->left_out = tp->sacked_out; 2018 tcp_undo_cwr(sk, 1); 2019 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO); 2020 inet_csk(sk)->icsk_retransmits = 0; 2021 tp->undo_marker = 0; 2022 if (!IsReno(tp)) 2023 tcp_set_ca_state(sk, TCP_CA_Open); 2024 return 1; 2025 } 2026 return 0; 2027 } 2028 2029 static inline void tcp_complete_cwr(struct sock *sk) 2030 { 2031 struct tcp_sock *tp = tcp_sk(sk); 2032 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh); 2033 tp->snd_cwnd_stamp = tcp_time_stamp; 2034 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR); 2035 } 2036 2037 static void tcp_try_to_open(struct sock *sk, int flag) 2038 { 2039 struct tcp_sock *tp = tcp_sk(sk); 2040 2041 tcp_sync_left_out(tp); 2042 2043 if (tp->retrans_out == 0) 2044 tp->retrans_stamp = 0; 2045 2046 if (flag&FLAG_ECE) 2047 tcp_enter_cwr(sk, 1); 2048 2049 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) { 2050 int state = TCP_CA_Open; 2051 2052 if (tp->left_out || tp->retrans_out || tp->undo_marker) 2053 state = TCP_CA_Disorder; 2054 2055 if (inet_csk(sk)->icsk_ca_state != state) { 2056 tcp_set_ca_state(sk, state); 2057 tp->high_seq = tp->snd_nxt; 2058 } 2059 tcp_moderate_cwnd(tp); 2060 } else { 2061 tcp_cwnd_down(sk); 2062 } 2063 } 2064 2065 static void tcp_mtup_probe_failed(struct sock *sk) 2066 { 2067 struct inet_connection_sock *icsk = inet_csk(sk); 2068 2069 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1; 2070 icsk->icsk_mtup.probe_size = 0; 2071 } 2072 2073 static void tcp_mtup_probe_success(struct sock *sk, struct sk_buff *skb) 2074 { 2075 struct tcp_sock *tp = tcp_sk(sk); 2076 struct inet_connection_sock *icsk = inet_csk(sk); 2077 2078 /* FIXME: breaks with very large cwnd */ 2079 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2080 tp->snd_cwnd = tp->snd_cwnd * 2081 tcp_mss_to_mtu(sk, tp->mss_cache) / 2082 icsk->icsk_mtup.probe_size; 2083 tp->snd_cwnd_cnt = 0; 2084 tp->snd_cwnd_stamp = tcp_time_stamp; 2085 tp->rcv_ssthresh = tcp_current_ssthresh(sk); 2086 2087 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size; 2088 icsk->icsk_mtup.probe_size = 0; 2089 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 2090 } 2091 2092 2093 /* Process an event, which can update packets-in-flight not trivially. 2094 * Main goal of this function is to calculate new estimate for left_out, 2095 * taking into account both packets sitting in receiver's buffer and 2096 * packets lost by network. 2097 * 2098 * Besides that it does CWND reduction, when packet loss is detected 2099 * and changes state of machine. 2100 * 2101 * It does _not_ decide what to send, it is made in function 2102 * tcp_xmit_retransmit_queue(). 2103 */ 2104 static void 2105 tcp_fastretrans_alert(struct sock *sk, u32 prior_snd_una, 2106 int prior_packets, int flag) 2107 { 2108 struct inet_connection_sock *icsk = inet_csk(sk); 2109 struct tcp_sock *tp = tcp_sk(sk); 2110 int is_dupack = (tp->snd_una == prior_snd_una && !(flag&FLAG_NOT_DUP)); 2111 2112 /* Some technical things: 2113 * 1. Reno does not count dupacks (sacked_out) automatically. */ 2114 if (!tp->packets_out) 2115 tp->sacked_out = 0; 2116 /* 2. SACK counts snd_fack in packets inaccurately. */ 2117 if (tp->sacked_out == 0) 2118 tp->fackets_out = 0; 2119 2120 /* Now state machine starts. 2121 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */ 2122 if (flag&FLAG_ECE) 2123 tp->prior_ssthresh = 0; 2124 2125 /* B. In all the states check for reneging SACKs. */ 2126 if (tp->sacked_out && tcp_check_sack_reneging(sk)) 2127 return; 2128 2129 /* C. Process data loss notification, provided it is valid. */ 2130 if ((flag&FLAG_DATA_LOST) && 2131 before(tp->snd_una, tp->high_seq) && 2132 icsk->icsk_ca_state != TCP_CA_Open && 2133 tp->fackets_out > tp->reordering) { 2134 tcp_mark_head_lost(sk, tp->fackets_out-tp->reordering, tp->high_seq); 2135 NET_INC_STATS_BH(LINUX_MIB_TCPLOSS); 2136 } 2137 2138 /* D. Synchronize left_out to current state. */ 2139 tcp_sync_left_out(tp); 2140 2141 /* E. Check state exit conditions. State can be terminated 2142 * when high_seq is ACKed. */ 2143 if (icsk->icsk_ca_state == TCP_CA_Open) { 2144 BUG_TRAP(tp->retrans_out == 0); 2145 tp->retrans_stamp = 0; 2146 } else if (!before(tp->snd_una, tp->high_seq)) { 2147 switch (icsk->icsk_ca_state) { 2148 case TCP_CA_Loss: 2149 icsk->icsk_retransmits = 0; 2150 if (tcp_try_undo_recovery(sk)) 2151 return; 2152 break; 2153 2154 case TCP_CA_CWR: 2155 /* CWR is to be held something *above* high_seq 2156 * is ACKed for CWR bit to reach receiver. */ 2157 if (tp->snd_una != tp->high_seq) { 2158 tcp_complete_cwr(sk); 2159 tcp_set_ca_state(sk, TCP_CA_Open); 2160 } 2161 break; 2162 2163 case TCP_CA_Disorder: 2164 tcp_try_undo_dsack(sk); 2165 if (!tp->undo_marker || 2166 /* For SACK case do not Open to allow to undo 2167 * catching for all duplicate ACKs. */ 2168 IsReno(tp) || tp->snd_una != tp->high_seq) { 2169 tp->undo_marker = 0; 2170 tcp_set_ca_state(sk, TCP_CA_Open); 2171 } 2172 break; 2173 2174 case TCP_CA_Recovery: 2175 if (IsReno(tp)) 2176 tcp_reset_reno_sack(tp); 2177 if (tcp_try_undo_recovery(sk)) 2178 return; 2179 tcp_complete_cwr(sk); 2180 break; 2181 } 2182 } 2183 2184 /* F. Process state. */ 2185 switch (icsk->icsk_ca_state) { 2186 case TCP_CA_Recovery: 2187 if (prior_snd_una == tp->snd_una) { 2188 if (IsReno(tp) && is_dupack) 2189 tcp_add_reno_sack(sk); 2190 } else { 2191 int acked = prior_packets - tp->packets_out; 2192 if (IsReno(tp)) 2193 tcp_remove_reno_sacks(sk, acked); 2194 is_dupack = tcp_try_undo_partial(sk, acked); 2195 } 2196 break; 2197 case TCP_CA_Loss: 2198 if (flag&FLAG_DATA_ACKED) 2199 icsk->icsk_retransmits = 0; 2200 if (!tcp_try_undo_loss(sk)) { 2201 tcp_moderate_cwnd(tp); 2202 tcp_xmit_retransmit_queue(sk); 2203 return; 2204 } 2205 if (icsk->icsk_ca_state != TCP_CA_Open) 2206 return; 2207 /* Loss is undone; fall through to processing in Open state. */ 2208 default: 2209 if (IsReno(tp)) { 2210 if (tp->snd_una != prior_snd_una) 2211 tcp_reset_reno_sack(tp); 2212 if (is_dupack) 2213 tcp_add_reno_sack(sk); 2214 } 2215 2216 if (icsk->icsk_ca_state == TCP_CA_Disorder) 2217 tcp_try_undo_dsack(sk); 2218 2219 if (!tcp_time_to_recover(sk)) { 2220 tcp_try_to_open(sk, flag); 2221 return; 2222 } 2223 2224 /* MTU probe failure: don't reduce cwnd */ 2225 if (icsk->icsk_ca_state < TCP_CA_CWR && 2226 icsk->icsk_mtup.probe_size && 2227 tp->snd_una == tp->mtu_probe.probe_seq_start) { 2228 tcp_mtup_probe_failed(sk); 2229 /* Restores the reduction we did in tcp_mtup_probe() */ 2230 tp->snd_cwnd++; 2231 tcp_simple_retransmit(sk); 2232 return; 2233 } 2234 2235 /* Otherwise enter Recovery state */ 2236 2237 if (IsReno(tp)) 2238 NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY); 2239 else 2240 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY); 2241 2242 tp->high_seq = tp->snd_nxt; 2243 tp->prior_ssthresh = 0; 2244 tp->undo_marker = tp->snd_una; 2245 tp->undo_retrans = tp->retrans_out; 2246 2247 if (icsk->icsk_ca_state < TCP_CA_CWR) { 2248 if (!(flag&FLAG_ECE)) 2249 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2250 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 2251 TCP_ECN_queue_cwr(tp); 2252 } 2253 2254 tp->bytes_acked = 0; 2255 tp->snd_cwnd_cnt = 0; 2256 tcp_set_ca_state(sk, TCP_CA_Recovery); 2257 } 2258 2259 if (is_dupack || tcp_head_timedout(sk)) 2260 tcp_update_scoreboard(sk); 2261 tcp_cwnd_down(sk); 2262 tcp_xmit_retransmit_queue(sk); 2263 } 2264 2265 /* Read draft-ietf-tcplw-high-performance before mucking 2266 * with this code. (Supersedes RFC1323) 2267 */ 2268 static void tcp_ack_saw_tstamp(struct sock *sk, int flag) 2269 { 2270 /* RTTM Rule: A TSecr value received in a segment is used to 2271 * update the averaged RTT measurement only if the segment 2272 * acknowledges some new data, i.e., only if it advances the 2273 * left edge of the send window. 2274 * 2275 * See draft-ietf-tcplw-high-performance-00, section 3.3. 2276 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru> 2277 * 2278 * Changed: reset backoff as soon as we see the first valid sample. 2279 * If we do not, we get strongly overestimated rto. With timestamps 2280 * samples are accepted even from very old segments: f.e., when rtt=1 2281 * increases to 8, we retransmit 5 times and after 8 seconds delayed 2282 * answer arrives rto becomes 120 seconds! If at least one of segments 2283 * in window is lost... Voila. --ANK (010210) 2284 */ 2285 struct tcp_sock *tp = tcp_sk(sk); 2286 const __u32 seq_rtt = tcp_time_stamp - tp->rx_opt.rcv_tsecr; 2287 tcp_rtt_estimator(sk, seq_rtt); 2288 tcp_set_rto(sk); 2289 inet_csk(sk)->icsk_backoff = 0; 2290 tcp_bound_rto(sk); 2291 } 2292 2293 static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag) 2294 { 2295 /* We don't have a timestamp. Can only use 2296 * packets that are not retransmitted to determine 2297 * rtt estimates. Also, we must not reset the 2298 * backoff for rto until we get a non-retransmitted 2299 * packet. This allows us to deal with a situation 2300 * where the network delay has increased suddenly. 2301 * I.e. Karn's algorithm. (SIGCOMM '87, p5.) 2302 */ 2303 2304 if (flag & FLAG_RETRANS_DATA_ACKED) 2305 return; 2306 2307 tcp_rtt_estimator(sk, seq_rtt); 2308 tcp_set_rto(sk); 2309 inet_csk(sk)->icsk_backoff = 0; 2310 tcp_bound_rto(sk); 2311 } 2312 2313 static inline void tcp_ack_update_rtt(struct sock *sk, const int flag, 2314 const s32 seq_rtt) 2315 { 2316 const struct tcp_sock *tp = tcp_sk(sk); 2317 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */ 2318 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr) 2319 tcp_ack_saw_tstamp(sk, flag); 2320 else if (seq_rtt >= 0) 2321 tcp_ack_no_tstamp(sk, seq_rtt, flag); 2322 } 2323 2324 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 rtt, 2325 u32 in_flight, int good) 2326 { 2327 const struct inet_connection_sock *icsk = inet_csk(sk); 2328 icsk->icsk_ca_ops->cong_avoid(sk, ack, rtt, in_flight, good); 2329 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp; 2330 } 2331 2332 /* Restart timer after forward progress on connection. 2333 * RFC2988 recommends to restart timer to now+rto. 2334 */ 2335 2336 static void tcp_ack_packets_out(struct sock *sk) 2337 { 2338 struct tcp_sock *tp = tcp_sk(sk); 2339 2340 if (!tp->packets_out) { 2341 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS); 2342 } else { 2343 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 2344 } 2345 } 2346 2347 static int tcp_tso_acked(struct sock *sk, struct sk_buff *skb, 2348 __u32 now, __s32 *seq_rtt) 2349 { 2350 struct tcp_sock *tp = tcp_sk(sk); 2351 struct tcp_skb_cb *scb = TCP_SKB_CB(skb); 2352 __u32 seq = tp->snd_una; 2353 __u32 packets_acked; 2354 int acked = 0; 2355 2356 /* If we get here, the whole TSO packet has not been 2357 * acked. 2358 */ 2359 BUG_ON(!after(scb->end_seq, seq)); 2360 2361 packets_acked = tcp_skb_pcount(skb); 2362 if (tcp_trim_head(sk, skb, seq - scb->seq)) 2363 return 0; 2364 packets_acked -= tcp_skb_pcount(skb); 2365 2366 if (packets_acked) { 2367 __u8 sacked = scb->sacked; 2368 2369 acked |= FLAG_DATA_ACKED; 2370 if (sacked) { 2371 if (sacked & TCPCB_RETRANS) { 2372 if (sacked & TCPCB_SACKED_RETRANS) 2373 tp->retrans_out -= packets_acked; 2374 acked |= FLAG_RETRANS_DATA_ACKED; 2375 *seq_rtt = -1; 2376 } else if (*seq_rtt < 0) 2377 *seq_rtt = now - scb->when; 2378 if (sacked & TCPCB_SACKED_ACKED) 2379 tp->sacked_out -= packets_acked; 2380 if (sacked & TCPCB_LOST) 2381 tp->lost_out -= packets_acked; 2382 if (sacked & TCPCB_URG) { 2383 if (tp->urg_mode && 2384 !before(seq, tp->snd_up)) 2385 tp->urg_mode = 0; 2386 } 2387 } else if (*seq_rtt < 0) 2388 *seq_rtt = now - scb->when; 2389 2390 if (tp->fackets_out) { 2391 __u32 dval = min(tp->fackets_out, packets_acked); 2392 tp->fackets_out -= dval; 2393 } 2394 tp->packets_out -= packets_acked; 2395 2396 BUG_ON(tcp_skb_pcount(skb) == 0); 2397 BUG_ON(!before(scb->seq, scb->end_seq)); 2398 } 2399 2400 return acked; 2401 } 2402 2403 /* Remove acknowledged frames from the retransmission queue. */ 2404 static int tcp_clean_rtx_queue(struct sock *sk, __s32 *seq_rtt_p) 2405 { 2406 struct tcp_sock *tp = tcp_sk(sk); 2407 const struct inet_connection_sock *icsk = inet_csk(sk); 2408 struct sk_buff *skb; 2409 __u32 now = tcp_time_stamp; 2410 int acked = 0; 2411 int prior_packets = tp->packets_out; 2412 __s32 seq_rtt = -1; 2413 ktime_t last_ackt = net_invalid_timestamp(); 2414 2415 while ((skb = tcp_write_queue_head(sk)) && 2416 skb != tcp_send_head(sk)) { 2417 struct tcp_skb_cb *scb = TCP_SKB_CB(skb); 2418 __u8 sacked = scb->sacked; 2419 2420 /* If our packet is before the ack sequence we can 2421 * discard it as it's confirmed to have arrived at 2422 * the other end. 2423 */ 2424 if (after(scb->end_seq, tp->snd_una)) { 2425 if (tcp_skb_pcount(skb) > 1 && 2426 after(tp->snd_una, scb->seq)) 2427 acked |= tcp_tso_acked(sk, skb, 2428 now, &seq_rtt); 2429 break; 2430 } 2431 2432 /* Initial outgoing SYN's get put onto the write_queue 2433 * just like anything else we transmit. It is not 2434 * true data, and if we misinform our callers that 2435 * this ACK acks real data, we will erroneously exit 2436 * connection startup slow start one packet too 2437 * quickly. This is severely frowned upon behavior. 2438 */ 2439 if (!(scb->flags & TCPCB_FLAG_SYN)) { 2440 acked |= FLAG_DATA_ACKED; 2441 } else { 2442 acked |= FLAG_SYN_ACKED; 2443 tp->retrans_stamp = 0; 2444 } 2445 2446 /* MTU probing checks */ 2447 if (icsk->icsk_mtup.probe_size) { 2448 if (!after(tp->mtu_probe.probe_seq_end, TCP_SKB_CB(skb)->end_seq)) { 2449 tcp_mtup_probe_success(sk, skb); 2450 } 2451 } 2452 2453 if (sacked) { 2454 if (sacked & TCPCB_RETRANS) { 2455 if (sacked & TCPCB_SACKED_RETRANS) 2456 tp->retrans_out -= tcp_skb_pcount(skb); 2457 acked |= FLAG_RETRANS_DATA_ACKED; 2458 seq_rtt = -1; 2459 } else if (seq_rtt < 0) { 2460 seq_rtt = now - scb->when; 2461 last_ackt = skb->tstamp; 2462 } 2463 if (sacked & TCPCB_SACKED_ACKED) 2464 tp->sacked_out -= tcp_skb_pcount(skb); 2465 if (sacked & TCPCB_LOST) 2466 tp->lost_out -= tcp_skb_pcount(skb); 2467 if (sacked & TCPCB_URG) { 2468 if (tp->urg_mode && 2469 !before(scb->end_seq, tp->snd_up)) 2470 tp->urg_mode = 0; 2471 } 2472 } else if (seq_rtt < 0) { 2473 seq_rtt = now - scb->when; 2474 last_ackt = skb->tstamp; 2475 } 2476 tcp_dec_pcount_approx(&tp->fackets_out, skb); 2477 tcp_packets_out_dec(tp, skb); 2478 tcp_unlink_write_queue(skb, sk); 2479 sk_stream_free_skb(sk, skb); 2480 clear_all_retrans_hints(tp); 2481 } 2482 2483 if (acked&FLAG_ACKED) { 2484 u32 pkts_acked = prior_packets - tp->packets_out; 2485 const struct tcp_congestion_ops *ca_ops 2486 = inet_csk(sk)->icsk_ca_ops; 2487 2488 tcp_ack_update_rtt(sk, acked, seq_rtt); 2489 tcp_ack_packets_out(sk); 2490 2491 /* Is the ACK triggering packet unambiguous? */ 2492 if (acked & FLAG_RETRANS_DATA_ACKED) 2493 last_ackt = net_invalid_timestamp(); 2494 2495 if (ca_ops->pkts_acked) 2496 ca_ops->pkts_acked(sk, pkts_acked, last_ackt); 2497 } 2498 2499 #if FASTRETRANS_DEBUG > 0 2500 BUG_TRAP((int)tp->sacked_out >= 0); 2501 BUG_TRAP((int)tp->lost_out >= 0); 2502 BUG_TRAP((int)tp->retrans_out >= 0); 2503 if (!tp->packets_out && tp->rx_opt.sack_ok) { 2504 const struct inet_connection_sock *icsk = inet_csk(sk); 2505 if (tp->lost_out) { 2506 printk(KERN_DEBUG "Leak l=%u %d\n", 2507 tp->lost_out, icsk->icsk_ca_state); 2508 tp->lost_out = 0; 2509 } 2510 if (tp->sacked_out) { 2511 printk(KERN_DEBUG "Leak s=%u %d\n", 2512 tp->sacked_out, icsk->icsk_ca_state); 2513 tp->sacked_out = 0; 2514 } 2515 if (tp->retrans_out) { 2516 printk(KERN_DEBUG "Leak r=%u %d\n", 2517 tp->retrans_out, icsk->icsk_ca_state); 2518 tp->retrans_out = 0; 2519 } 2520 } 2521 #endif 2522 *seq_rtt_p = seq_rtt; 2523 return acked; 2524 } 2525 2526 static void tcp_ack_probe(struct sock *sk) 2527 { 2528 const struct tcp_sock *tp = tcp_sk(sk); 2529 struct inet_connection_sock *icsk = inet_csk(sk); 2530 2531 /* Was it a usable window open? */ 2532 2533 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, 2534 tp->snd_una + tp->snd_wnd)) { 2535 icsk->icsk_backoff = 0; 2536 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0); 2537 /* Socket must be waked up by subsequent tcp_data_snd_check(). 2538 * This function is not for random using! 2539 */ 2540 } else { 2541 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 2542 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), 2543 TCP_RTO_MAX); 2544 } 2545 } 2546 2547 static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag) 2548 { 2549 return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) || 2550 inet_csk(sk)->icsk_ca_state != TCP_CA_Open); 2551 } 2552 2553 static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag) 2554 { 2555 const struct tcp_sock *tp = tcp_sk(sk); 2556 return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) && 2557 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR)); 2558 } 2559 2560 /* Check that window update is acceptable. 2561 * The function assumes that snd_una<=ack<=snd_next. 2562 */ 2563 static inline int tcp_may_update_window(const struct tcp_sock *tp, const u32 ack, 2564 const u32 ack_seq, const u32 nwin) 2565 { 2566 return (after(ack, tp->snd_una) || 2567 after(ack_seq, tp->snd_wl1) || 2568 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd)); 2569 } 2570 2571 /* Update our send window. 2572 * 2573 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2 2574 * and in FreeBSD. NetBSD's one is even worse.) is wrong. 2575 */ 2576 static int tcp_ack_update_window(struct sock *sk, struct sk_buff *skb, u32 ack, 2577 u32 ack_seq) 2578 { 2579 struct tcp_sock *tp = tcp_sk(sk); 2580 int flag = 0; 2581 u32 nwin = ntohs(tcp_hdr(skb)->window); 2582 2583 if (likely(!tcp_hdr(skb)->syn)) 2584 nwin <<= tp->rx_opt.snd_wscale; 2585 2586 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) { 2587 flag |= FLAG_WIN_UPDATE; 2588 tcp_update_wl(tp, ack, ack_seq); 2589 2590 if (tp->snd_wnd != nwin) { 2591 tp->snd_wnd = nwin; 2592 2593 /* Note, it is the only place, where 2594 * fast path is recovered for sending TCP. 2595 */ 2596 tp->pred_flags = 0; 2597 tcp_fast_path_check(sk); 2598 2599 if (nwin > tp->max_window) { 2600 tp->max_window = nwin; 2601 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie); 2602 } 2603 } 2604 } 2605 2606 tp->snd_una = ack; 2607 2608 return flag; 2609 } 2610 2611 /* A very conservative spurious RTO response algorithm: reduce cwnd and 2612 * continue in congestion avoidance. 2613 */ 2614 static void tcp_conservative_spur_to_response(struct tcp_sock *tp) 2615 { 2616 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh); 2617 tp->snd_cwnd_cnt = 0; 2618 TCP_ECN_queue_cwr(tp); 2619 tcp_moderate_cwnd(tp); 2620 } 2621 2622 /* A conservative spurious RTO response algorithm: reduce cwnd using 2623 * rate halving and continue in congestion avoidance. 2624 */ 2625 static void tcp_ratehalving_spur_to_response(struct sock *sk) 2626 { 2627 tcp_enter_cwr(sk, 0); 2628 } 2629 2630 static void tcp_undo_spur_to_response(struct sock *sk, int flag) 2631 { 2632 if (flag&FLAG_ECE) 2633 tcp_ratehalving_spur_to_response(sk); 2634 else 2635 tcp_undo_cwr(sk, 1); 2636 } 2637 2638 /* F-RTO spurious RTO detection algorithm (RFC4138) 2639 * 2640 * F-RTO affects during two new ACKs following RTO (well, almost, see inline 2641 * comments). State (ACK number) is kept in frto_counter. When ACK advances 2642 * window (but not to or beyond highest sequence sent before RTO): 2643 * On First ACK, send two new segments out. 2644 * On Second ACK, RTO was likely spurious. Do spurious response (response 2645 * algorithm is not part of the F-RTO detection algorithm 2646 * given in RFC4138 but can be selected separately). 2647 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss 2648 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding 2649 * of Nagle, this is done using frto_counter states 2 and 3, when a new data 2650 * segment of any size sent during F-RTO, state 2 is upgraded to 3. 2651 * 2652 * Rationale: if the RTO was spurious, new ACKs should arrive from the 2653 * original window even after we transmit two new data segments. 2654 * 2655 * SACK version: 2656 * on first step, wait until first cumulative ACK arrives, then move to 2657 * the second step. In second step, the next ACK decides. 2658 * 2659 * F-RTO is implemented (mainly) in four functions: 2660 * - tcp_use_frto() is used to determine if TCP is can use F-RTO 2661 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is 2662 * called when tcp_use_frto() showed green light 2663 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm 2664 * - tcp_enter_frto_loss() is called if there is not enough evidence 2665 * to prove that the RTO is indeed spurious. It transfers the control 2666 * from F-RTO to the conventional RTO recovery 2667 */ 2668 static int tcp_process_frto(struct sock *sk, u32 prior_snd_una, int flag) 2669 { 2670 struct tcp_sock *tp = tcp_sk(sk); 2671 2672 tcp_sync_left_out(tp); 2673 2674 /* Duplicate the behavior from Loss state (fastretrans_alert) */ 2675 if (flag&FLAG_DATA_ACKED) 2676 inet_csk(sk)->icsk_retransmits = 0; 2677 2678 if (!before(tp->snd_una, tp->frto_highmark)) { 2679 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag); 2680 return 1; 2681 } 2682 2683 if (!IsSackFrto() || IsReno(tp)) { 2684 /* RFC4138 shortcoming in step 2; should also have case c): 2685 * ACK isn't duplicate nor advances window, e.g., opposite dir 2686 * data, winupdate 2687 */ 2688 if ((tp->snd_una == prior_snd_una) && (flag&FLAG_NOT_DUP) && 2689 !(flag&FLAG_FORWARD_PROGRESS)) 2690 return 1; 2691 2692 if (!(flag&FLAG_DATA_ACKED)) { 2693 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3), 2694 flag); 2695 return 1; 2696 } 2697 } else { 2698 if (!(flag&FLAG_DATA_ACKED) && (tp->frto_counter == 1)) { 2699 /* Prevent sending of new data. */ 2700 tp->snd_cwnd = min(tp->snd_cwnd, 2701 tcp_packets_in_flight(tp)); 2702 return 1; 2703 } 2704 2705 if ((tp->frto_counter >= 2) && 2706 (!(flag&FLAG_FORWARD_PROGRESS) || 2707 ((flag&FLAG_DATA_SACKED) && !(flag&FLAG_ONLY_ORIG_SACKED)))) { 2708 /* RFC4138 shortcoming (see comment above) */ 2709 if (!(flag&FLAG_FORWARD_PROGRESS) && (flag&FLAG_NOT_DUP)) 2710 return 1; 2711 2712 tcp_enter_frto_loss(sk, 3, flag); 2713 return 1; 2714 } 2715 } 2716 2717 if (tp->frto_counter == 1) { 2718 /* Sending of the next skb must be allowed or no FRTO */ 2719 if (!tcp_send_head(sk) || 2720 after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, 2721 tp->snd_una + tp->snd_wnd)) { 2722 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), 2723 flag); 2724 return 1; 2725 } 2726 2727 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2; 2728 tp->frto_counter = 2; 2729 return 1; 2730 } else { 2731 switch (sysctl_tcp_frto_response) { 2732 case 2: 2733 tcp_undo_spur_to_response(sk, flag); 2734 break; 2735 case 1: 2736 tcp_conservative_spur_to_response(tp); 2737 break; 2738 default: 2739 tcp_ratehalving_spur_to_response(sk); 2740 break; 2741 } 2742 tp->frto_counter = 0; 2743 } 2744 return 0; 2745 } 2746 2747 /* This routine deals with incoming acks, but not outgoing ones. */ 2748 static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag) 2749 { 2750 struct inet_connection_sock *icsk = inet_csk(sk); 2751 struct tcp_sock *tp = tcp_sk(sk); 2752 u32 prior_snd_una = tp->snd_una; 2753 u32 ack_seq = TCP_SKB_CB(skb)->seq; 2754 u32 ack = TCP_SKB_CB(skb)->ack_seq; 2755 u32 prior_in_flight; 2756 s32 seq_rtt; 2757 int prior_packets; 2758 int frto_cwnd = 0; 2759 2760 /* If the ack is newer than sent or older than previous acks 2761 * then we can probably ignore it. 2762 */ 2763 if (after(ack, tp->snd_nxt)) 2764 goto uninteresting_ack; 2765 2766 if (before(ack, prior_snd_una)) 2767 goto old_ack; 2768 2769 if (sysctl_tcp_abc) { 2770 if (icsk->icsk_ca_state < TCP_CA_CWR) 2771 tp->bytes_acked += ack - prior_snd_una; 2772 else if (icsk->icsk_ca_state == TCP_CA_Loss) 2773 /* we assume just one segment left network */ 2774 tp->bytes_acked += min(ack - prior_snd_una, tp->mss_cache); 2775 } 2776 2777 if (!(flag&FLAG_SLOWPATH) && after(ack, prior_snd_una)) { 2778 /* Window is constant, pure forward advance. 2779 * No more checks are required. 2780 * Note, we use the fact that SND.UNA>=SND.WL2. 2781 */ 2782 tcp_update_wl(tp, ack, ack_seq); 2783 tp->snd_una = ack; 2784 flag |= FLAG_WIN_UPDATE; 2785 2786 tcp_ca_event(sk, CA_EVENT_FAST_ACK); 2787 2788 NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS); 2789 } else { 2790 if (ack_seq != TCP_SKB_CB(skb)->end_seq) 2791 flag |= FLAG_DATA; 2792 else 2793 NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS); 2794 2795 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq); 2796 2797 if (TCP_SKB_CB(skb)->sacked) 2798 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una); 2799 2800 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb))) 2801 flag |= FLAG_ECE; 2802 2803 tcp_ca_event(sk, CA_EVENT_SLOW_ACK); 2804 } 2805 2806 /* We passed data and got it acked, remove any soft error 2807 * log. Something worked... 2808 */ 2809 sk->sk_err_soft = 0; 2810 tp->rcv_tstamp = tcp_time_stamp; 2811 prior_packets = tp->packets_out; 2812 if (!prior_packets) 2813 goto no_queue; 2814 2815 prior_in_flight = tcp_packets_in_flight(tp); 2816 2817 /* See if we can take anything off of the retransmit queue. */ 2818 flag |= tcp_clean_rtx_queue(sk, &seq_rtt); 2819 2820 if (tp->frto_counter) 2821 frto_cwnd = tcp_process_frto(sk, prior_snd_una, flag); 2822 2823 if (tcp_ack_is_dubious(sk, flag)) { 2824 /* Advance CWND, if state allows this. */ 2825 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd && 2826 tcp_may_raise_cwnd(sk, flag)) 2827 tcp_cong_avoid(sk, ack, seq_rtt, prior_in_flight, 0); 2828 tcp_fastretrans_alert(sk, prior_snd_una, prior_packets, flag); 2829 } else { 2830 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd) 2831 tcp_cong_avoid(sk, ack, seq_rtt, prior_in_flight, 1); 2832 } 2833 2834 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag&FLAG_NOT_DUP)) 2835 dst_confirm(sk->sk_dst_cache); 2836 2837 return 1; 2838 2839 no_queue: 2840 icsk->icsk_probes_out = 0; 2841 2842 /* If this ack opens up a zero window, clear backoff. It was 2843 * being used to time the probes, and is probably far higher than 2844 * it needs to be for normal retransmission. 2845 */ 2846 if (tcp_send_head(sk)) 2847 tcp_ack_probe(sk); 2848 return 1; 2849 2850 old_ack: 2851 if (TCP_SKB_CB(skb)->sacked) 2852 tcp_sacktag_write_queue(sk, skb, prior_snd_una); 2853 2854 uninteresting_ack: 2855 SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt); 2856 return 0; 2857 } 2858 2859 2860 /* Look for tcp options. Normally only called on SYN and SYNACK packets. 2861 * But, this can also be called on packets in the established flow when 2862 * the fast version below fails. 2863 */ 2864 void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx, int estab) 2865 { 2866 unsigned char *ptr; 2867 struct tcphdr *th = tcp_hdr(skb); 2868 int length=(th->doff*4)-sizeof(struct tcphdr); 2869 2870 ptr = (unsigned char *)(th + 1); 2871 opt_rx->saw_tstamp = 0; 2872 2873 while (length > 0) { 2874 int opcode=*ptr++; 2875 int opsize; 2876 2877 switch (opcode) { 2878 case TCPOPT_EOL: 2879 return; 2880 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */ 2881 length--; 2882 continue; 2883 default: 2884 opsize=*ptr++; 2885 if (opsize < 2) /* "silly options" */ 2886 return; 2887 if (opsize > length) 2888 return; /* don't parse partial options */ 2889 switch (opcode) { 2890 case TCPOPT_MSS: 2891 if (opsize==TCPOLEN_MSS && th->syn && !estab) { 2892 u16 in_mss = ntohs(get_unaligned((__be16 *)ptr)); 2893 if (in_mss) { 2894 if (opt_rx->user_mss && opt_rx->user_mss < in_mss) 2895 in_mss = opt_rx->user_mss; 2896 opt_rx->mss_clamp = in_mss; 2897 } 2898 } 2899 break; 2900 case TCPOPT_WINDOW: 2901 if (opsize==TCPOLEN_WINDOW && th->syn && !estab) 2902 if (sysctl_tcp_window_scaling) { 2903 __u8 snd_wscale = *(__u8 *) ptr; 2904 opt_rx->wscale_ok = 1; 2905 if (snd_wscale > 14) { 2906 if (net_ratelimit()) 2907 printk(KERN_INFO "tcp_parse_options: Illegal window " 2908 "scaling value %d >14 received.\n", 2909 snd_wscale); 2910 snd_wscale = 14; 2911 } 2912 opt_rx->snd_wscale = snd_wscale; 2913 } 2914 break; 2915 case TCPOPT_TIMESTAMP: 2916 if (opsize==TCPOLEN_TIMESTAMP) { 2917 if ((estab && opt_rx->tstamp_ok) || 2918 (!estab && sysctl_tcp_timestamps)) { 2919 opt_rx->saw_tstamp = 1; 2920 opt_rx->rcv_tsval = ntohl(get_unaligned((__be32 *)ptr)); 2921 opt_rx->rcv_tsecr = ntohl(get_unaligned((__be32 *)(ptr+4))); 2922 } 2923 } 2924 break; 2925 case TCPOPT_SACK_PERM: 2926 if (opsize==TCPOLEN_SACK_PERM && th->syn && !estab) { 2927 if (sysctl_tcp_sack) { 2928 opt_rx->sack_ok = 1; 2929 tcp_sack_reset(opt_rx); 2930 } 2931 } 2932 break; 2933 2934 case TCPOPT_SACK: 2935 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) && 2936 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) && 2937 opt_rx->sack_ok) { 2938 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th; 2939 } 2940 break; 2941 #ifdef CONFIG_TCP_MD5SIG 2942 case TCPOPT_MD5SIG: 2943 /* 2944 * The MD5 Hash has already been 2945 * checked (see tcp_v{4,6}_do_rcv()). 2946 */ 2947 break; 2948 #endif 2949 } 2950 2951 ptr+=opsize-2; 2952 length-=opsize; 2953 } 2954 } 2955 } 2956 2957 /* Fast parse options. This hopes to only see timestamps. 2958 * If it is wrong it falls back on tcp_parse_options(). 2959 */ 2960 static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th, 2961 struct tcp_sock *tp) 2962 { 2963 if (th->doff == sizeof(struct tcphdr)>>2) { 2964 tp->rx_opt.saw_tstamp = 0; 2965 return 0; 2966 } else if (tp->rx_opt.tstamp_ok && 2967 th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) { 2968 __be32 *ptr = (__be32 *)(th + 1); 2969 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) 2970 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) { 2971 tp->rx_opt.saw_tstamp = 1; 2972 ++ptr; 2973 tp->rx_opt.rcv_tsval = ntohl(*ptr); 2974 ++ptr; 2975 tp->rx_opt.rcv_tsecr = ntohl(*ptr); 2976 return 1; 2977 } 2978 } 2979 tcp_parse_options(skb, &tp->rx_opt, 1); 2980 return 1; 2981 } 2982 2983 static inline void tcp_store_ts_recent(struct tcp_sock *tp) 2984 { 2985 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval; 2986 tp->rx_opt.ts_recent_stamp = get_seconds(); 2987 } 2988 2989 static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq) 2990 { 2991 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) { 2992 /* PAWS bug workaround wrt. ACK frames, the PAWS discard 2993 * extra check below makes sure this can only happen 2994 * for pure ACK frames. -DaveM 2995 * 2996 * Not only, also it occurs for expired timestamps. 2997 */ 2998 2999 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) >= 0 || 3000 get_seconds() >= tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS) 3001 tcp_store_ts_recent(tp); 3002 } 3003 } 3004 3005 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM 3006 * 3007 * It is not fatal. If this ACK does _not_ change critical state (seqs, window) 3008 * it can pass through stack. So, the following predicate verifies that 3009 * this segment is not used for anything but congestion avoidance or 3010 * fast retransmit. Moreover, we even are able to eliminate most of such 3011 * second order effects, if we apply some small "replay" window (~RTO) 3012 * to timestamp space. 3013 * 3014 * All these measures still do not guarantee that we reject wrapped ACKs 3015 * on networks with high bandwidth, when sequence space is recycled fastly, 3016 * but it guarantees that such events will be very rare and do not affect 3017 * connection seriously. This doesn't look nice, but alas, PAWS is really 3018 * buggy extension. 3019 * 3020 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC 3021 * states that events when retransmit arrives after original data are rare. 3022 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is 3023 * the biggest problem on large power networks even with minor reordering. 3024 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe 3025 * up to bandwidth of 18Gigabit/sec. 8) ] 3026 */ 3027 3028 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb) 3029 { 3030 struct tcp_sock *tp = tcp_sk(sk); 3031 struct tcphdr *th = tcp_hdr(skb); 3032 u32 seq = TCP_SKB_CB(skb)->seq; 3033 u32 ack = TCP_SKB_CB(skb)->ack_seq; 3034 3035 return (/* 1. Pure ACK with correct sequence number. */ 3036 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) && 3037 3038 /* 2. ... and duplicate ACK. */ 3039 ack == tp->snd_una && 3040 3041 /* 3. ... and does not update window. */ 3042 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) && 3043 3044 /* 4. ... and sits in replay window. */ 3045 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ); 3046 } 3047 3048 static inline int tcp_paws_discard(const struct sock *sk, const struct sk_buff *skb) 3049 { 3050 const struct tcp_sock *tp = tcp_sk(sk); 3051 return ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > TCP_PAWS_WINDOW && 3052 get_seconds() < tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS && 3053 !tcp_disordered_ack(sk, skb)); 3054 } 3055 3056 /* Check segment sequence number for validity. 3057 * 3058 * Segment controls are considered valid, if the segment 3059 * fits to the window after truncation to the window. Acceptability 3060 * of data (and SYN, FIN, of course) is checked separately. 3061 * See tcp_data_queue(), for example. 3062 * 3063 * Also, controls (RST is main one) are accepted using RCV.WUP instead 3064 * of RCV.NXT. Peer still did not advance his SND.UNA when we 3065 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP. 3066 * (borrowed from freebsd) 3067 */ 3068 3069 static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq) 3070 { 3071 return !before(end_seq, tp->rcv_wup) && 3072 !after(seq, tp->rcv_nxt + tcp_receive_window(tp)); 3073 } 3074 3075 /* When we get a reset we do this. */ 3076 static void tcp_reset(struct sock *sk) 3077 { 3078 /* We want the right error as BSD sees it (and indeed as we do). */ 3079 switch (sk->sk_state) { 3080 case TCP_SYN_SENT: 3081 sk->sk_err = ECONNREFUSED; 3082 break; 3083 case TCP_CLOSE_WAIT: 3084 sk->sk_err = EPIPE; 3085 break; 3086 case TCP_CLOSE: 3087 return; 3088 default: 3089 sk->sk_err = ECONNRESET; 3090 } 3091 3092 if (!sock_flag(sk, SOCK_DEAD)) 3093 sk->sk_error_report(sk); 3094 3095 tcp_done(sk); 3096 } 3097 3098 /* 3099 * Process the FIN bit. This now behaves as it is supposed to work 3100 * and the FIN takes effect when it is validly part of sequence 3101 * space. Not before when we get holes. 3102 * 3103 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT 3104 * (and thence onto LAST-ACK and finally, CLOSE, we never enter 3105 * TIME-WAIT) 3106 * 3107 * If we are in FINWAIT-1, a received FIN indicates simultaneous 3108 * close and we go into CLOSING (and later onto TIME-WAIT) 3109 * 3110 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT. 3111 */ 3112 static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th) 3113 { 3114 struct tcp_sock *tp = tcp_sk(sk); 3115 3116 inet_csk_schedule_ack(sk); 3117 3118 sk->sk_shutdown |= RCV_SHUTDOWN; 3119 sock_set_flag(sk, SOCK_DONE); 3120 3121 switch (sk->sk_state) { 3122 case TCP_SYN_RECV: 3123 case TCP_ESTABLISHED: 3124 /* Move to CLOSE_WAIT */ 3125 tcp_set_state(sk, TCP_CLOSE_WAIT); 3126 inet_csk(sk)->icsk_ack.pingpong = 1; 3127 break; 3128 3129 case TCP_CLOSE_WAIT: 3130 case TCP_CLOSING: 3131 /* Received a retransmission of the FIN, do 3132 * nothing. 3133 */ 3134 break; 3135 case TCP_LAST_ACK: 3136 /* RFC793: Remain in the LAST-ACK state. */ 3137 break; 3138 3139 case TCP_FIN_WAIT1: 3140 /* This case occurs when a simultaneous close 3141 * happens, we must ack the received FIN and 3142 * enter the CLOSING state. 3143 */ 3144 tcp_send_ack(sk); 3145 tcp_set_state(sk, TCP_CLOSING); 3146 break; 3147 case TCP_FIN_WAIT2: 3148 /* Received a FIN -- send ACK and enter TIME_WAIT. */ 3149 tcp_send_ack(sk); 3150 tcp_time_wait(sk, TCP_TIME_WAIT, 0); 3151 break; 3152 default: 3153 /* Only TCP_LISTEN and TCP_CLOSE are left, in these 3154 * cases we should never reach this piece of code. 3155 */ 3156 printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n", 3157 __FUNCTION__, sk->sk_state); 3158 break; 3159 } 3160 3161 /* It _is_ possible, that we have something out-of-order _after_ FIN. 3162 * Probably, we should reset in this case. For now drop them. 3163 */ 3164 __skb_queue_purge(&tp->out_of_order_queue); 3165 if (tp->rx_opt.sack_ok) 3166 tcp_sack_reset(&tp->rx_opt); 3167 sk_stream_mem_reclaim(sk); 3168 3169 if (!sock_flag(sk, SOCK_DEAD)) { 3170 sk->sk_state_change(sk); 3171 3172 /* Do not send POLL_HUP for half duplex close. */ 3173 if (sk->sk_shutdown == SHUTDOWN_MASK || 3174 sk->sk_state == TCP_CLOSE) 3175 sk_wake_async(sk, 1, POLL_HUP); 3176 else 3177 sk_wake_async(sk, 1, POLL_IN); 3178 } 3179 } 3180 3181 static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, u32 end_seq) 3182 { 3183 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) { 3184 if (before(seq, sp->start_seq)) 3185 sp->start_seq = seq; 3186 if (after(end_seq, sp->end_seq)) 3187 sp->end_seq = end_seq; 3188 return 1; 3189 } 3190 return 0; 3191 } 3192 3193 static void tcp_dsack_set(struct tcp_sock *tp, u32 seq, u32 end_seq) 3194 { 3195 if (tp->rx_opt.sack_ok && sysctl_tcp_dsack) { 3196 if (before(seq, tp->rcv_nxt)) 3197 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT); 3198 else 3199 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT); 3200 3201 tp->rx_opt.dsack = 1; 3202 tp->duplicate_sack[0].start_seq = seq; 3203 tp->duplicate_sack[0].end_seq = end_seq; 3204 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + 1, 4 - tp->rx_opt.tstamp_ok); 3205 } 3206 } 3207 3208 static void tcp_dsack_extend(struct tcp_sock *tp, u32 seq, u32 end_seq) 3209 { 3210 if (!tp->rx_opt.dsack) 3211 tcp_dsack_set(tp, seq, end_seq); 3212 else 3213 tcp_sack_extend(tp->duplicate_sack, seq, end_seq); 3214 } 3215 3216 static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb) 3217 { 3218 struct tcp_sock *tp = tcp_sk(sk); 3219 3220 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 3221 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 3222 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST); 3223 tcp_enter_quickack_mode(sk); 3224 3225 if (tp->rx_opt.sack_ok && sysctl_tcp_dsack) { 3226 u32 end_seq = TCP_SKB_CB(skb)->end_seq; 3227 3228 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) 3229 end_seq = tp->rcv_nxt; 3230 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, end_seq); 3231 } 3232 } 3233 3234 tcp_send_ack(sk); 3235 } 3236 3237 /* These routines update the SACK block as out-of-order packets arrive or 3238 * in-order packets close up the sequence space. 3239 */ 3240 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp) 3241 { 3242 int this_sack; 3243 struct tcp_sack_block *sp = &tp->selective_acks[0]; 3244 struct tcp_sack_block *swalk = sp+1; 3245 3246 /* See if the recent change to the first SACK eats into 3247 * or hits the sequence space of other SACK blocks, if so coalesce. 3248 */ 3249 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks; ) { 3250 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) { 3251 int i; 3252 3253 /* Zap SWALK, by moving every further SACK up by one slot. 3254 * Decrease num_sacks. 3255 */ 3256 tp->rx_opt.num_sacks--; 3257 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok); 3258 for (i=this_sack; i < tp->rx_opt.num_sacks; i++) 3259 sp[i] = sp[i+1]; 3260 continue; 3261 } 3262 this_sack++, swalk++; 3263 } 3264 } 3265 3266 static inline void tcp_sack_swap(struct tcp_sack_block *sack1, struct tcp_sack_block *sack2) 3267 { 3268 __u32 tmp; 3269 3270 tmp = sack1->start_seq; 3271 sack1->start_seq = sack2->start_seq; 3272 sack2->start_seq = tmp; 3273 3274 tmp = sack1->end_seq; 3275 sack1->end_seq = sack2->end_seq; 3276 sack2->end_seq = tmp; 3277 } 3278 3279 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq) 3280 { 3281 struct tcp_sock *tp = tcp_sk(sk); 3282 struct tcp_sack_block *sp = &tp->selective_acks[0]; 3283 int cur_sacks = tp->rx_opt.num_sacks; 3284 int this_sack; 3285 3286 if (!cur_sacks) 3287 goto new_sack; 3288 3289 for (this_sack=0; this_sack<cur_sacks; this_sack++, sp++) { 3290 if (tcp_sack_extend(sp, seq, end_seq)) { 3291 /* Rotate this_sack to the first one. */ 3292 for (; this_sack>0; this_sack--, sp--) 3293 tcp_sack_swap(sp, sp-1); 3294 if (cur_sacks > 1) 3295 tcp_sack_maybe_coalesce(tp); 3296 return; 3297 } 3298 } 3299 3300 /* Could not find an adjacent existing SACK, build a new one, 3301 * put it at the front, and shift everyone else down. We 3302 * always know there is at least one SACK present already here. 3303 * 3304 * If the sack array is full, forget about the last one. 3305 */ 3306 if (this_sack >= 4) { 3307 this_sack--; 3308 tp->rx_opt.num_sacks--; 3309 sp--; 3310 } 3311 for (; this_sack > 0; this_sack--, sp--) 3312 *sp = *(sp-1); 3313 3314 new_sack: 3315 /* Build the new head SACK, and we're done. */ 3316 sp->start_seq = seq; 3317 sp->end_seq = end_seq; 3318 tp->rx_opt.num_sacks++; 3319 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok); 3320 } 3321 3322 /* RCV.NXT advances, some SACKs should be eaten. */ 3323 3324 static void tcp_sack_remove(struct tcp_sock *tp) 3325 { 3326 struct tcp_sack_block *sp = &tp->selective_acks[0]; 3327 int num_sacks = tp->rx_opt.num_sacks; 3328 int this_sack; 3329 3330 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */ 3331 if (skb_queue_empty(&tp->out_of_order_queue)) { 3332 tp->rx_opt.num_sacks = 0; 3333 tp->rx_opt.eff_sacks = tp->rx_opt.dsack; 3334 return; 3335 } 3336 3337 for (this_sack = 0; this_sack < num_sacks; ) { 3338 /* Check if the start of the sack is covered by RCV.NXT. */ 3339 if (!before(tp->rcv_nxt, sp->start_seq)) { 3340 int i; 3341 3342 /* RCV.NXT must cover all the block! */ 3343 BUG_TRAP(!before(tp->rcv_nxt, sp->end_seq)); 3344 3345 /* Zap this SACK, by moving forward any other SACKS. */ 3346 for (i=this_sack+1; i < num_sacks; i++) 3347 tp->selective_acks[i-1] = tp->selective_acks[i]; 3348 num_sacks--; 3349 continue; 3350 } 3351 this_sack++; 3352 sp++; 3353 } 3354 if (num_sacks != tp->rx_opt.num_sacks) { 3355 tp->rx_opt.num_sacks = num_sacks; 3356 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok); 3357 } 3358 } 3359 3360 /* This one checks to see if we can put data from the 3361 * out_of_order queue into the receive_queue. 3362 */ 3363 static void tcp_ofo_queue(struct sock *sk) 3364 { 3365 struct tcp_sock *tp = tcp_sk(sk); 3366 __u32 dsack_high = tp->rcv_nxt; 3367 struct sk_buff *skb; 3368 3369 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) { 3370 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) 3371 break; 3372 3373 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) { 3374 __u32 dsack = dsack_high; 3375 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high)) 3376 dsack_high = TCP_SKB_CB(skb)->end_seq; 3377 tcp_dsack_extend(tp, TCP_SKB_CB(skb)->seq, dsack); 3378 } 3379 3380 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) { 3381 SOCK_DEBUG(sk, "ofo packet was already received \n"); 3382 __skb_unlink(skb, &tp->out_of_order_queue); 3383 __kfree_skb(skb); 3384 continue; 3385 } 3386 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n", 3387 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, 3388 TCP_SKB_CB(skb)->end_seq); 3389 3390 __skb_unlink(skb, &tp->out_of_order_queue); 3391 __skb_queue_tail(&sk->sk_receive_queue, skb); 3392 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; 3393 if (tcp_hdr(skb)->fin) 3394 tcp_fin(skb, sk, tcp_hdr(skb)); 3395 } 3396 } 3397 3398 static int tcp_prune_queue(struct sock *sk); 3399 3400 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb) 3401 { 3402 struct tcphdr *th = tcp_hdr(skb); 3403 struct tcp_sock *tp = tcp_sk(sk); 3404 int eaten = -1; 3405 3406 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) 3407 goto drop; 3408 3409 __skb_pull(skb, th->doff*4); 3410 3411 TCP_ECN_accept_cwr(tp, skb); 3412 3413 if (tp->rx_opt.dsack) { 3414 tp->rx_opt.dsack = 0; 3415 tp->rx_opt.eff_sacks = min_t(unsigned int, tp->rx_opt.num_sacks, 3416 4 - tp->rx_opt.tstamp_ok); 3417 } 3418 3419 /* Queue data for delivery to the user. 3420 * Packets in sequence go to the receive queue. 3421 * Out of sequence packets to the out_of_order_queue. 3422 */ 3423 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) { 3424 if (tcp_receive_window(tp) == 0) 3425 goto out_of_window; 3426 3427 /* Ok. In sequence. In window. */ 3428 if (tp->ucopy.task == current && 3429 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len && 3430 sock_owned_by_user(sk) && !tp->urg_data) { 3431 int chunk = min_t(unsigned int, skb->len, 3432 tp->ucopy.len); 3433 3434 __set_current_state(TASK_RUNNING); 3435 3436 local_bh_enable(); 3437 if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) { 3438 tp->ucopy.len -= chunk; 3439 tp->copied_seq += chunk; 3440 eaten = (chunk == skb->len && !th->fin); 3441 tcp_rcv_space_adjust(sk); 3442 } 3443 local_bh_disable(); 3444 } 3445 3446 if (eaten <= 0) { 3447 queue_and_out: 3448 if (eaten < 0 && 3449 (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf || 3450 !sk_stream_rmem_schedule(sk, skb))) { 3451 if (tcp_prune_queue(sk) < 0 || 3452 !sk_stream_rmem_schedule(sk, skb)) 3453 goto drop; 3454 } 3455 sk_stream_set_owner_r(skb, sk); 3456 __skb_queue_tail(&sk->sk_receive_queue, skb); 3457 } 3458 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; 3459 if (skb->len) 3460 tcp_event_data_recv(sk, skb); 3461 if (th->fin) 3462 tcp_fin(skb, sk, th); 3463 3464 if (!skb_queue_empty(&tp->out_of_order_queue)) { 3465 tcp_ofo_queue(sk); 3466 3467 /* RFC2581. 4.2. SHOULD send immediate ACK, when 3468 * gap in queue is filled. 3469 */ 3470 if (skb_queue_empty(&tp->out_of_order_queue)) 3471 inet_csk(sk)->icsk_ack.pingpong = 0; 3472 } 3473 3474 if (tp->rx_opt.num_sacks) 3475 tcp_sack_remove(tp); 3476 3477 tcp_fast_path_check(sk); 3478 3479 if (eaten > 0) 3480 __kfree_skb(skb); 3481 else if (!sock_flag(sk, SOCK_DEAD)) 3482 sk->sk_data_ready(sk, 0); 3483 return; 3484 } 3485 3486 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) { 3487 /* A retransmit, 2nd most common case. Force an immediate ack. */ 3488 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST); 3489 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); 3490 3491 out_of_window: 3492 tcp_enter_quickack_mode(sk); 3493 inet_csk_schedule_ack(sk); 3494 drop: 3495 __kfree_skb(skb); 3496 return; 3497 } 3498 3499 /* Out of window. F.e. zero window probe. */ 3500 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp))) 3501 goto out_of_window; 3502 3503 tcp_enter_quickack_mode(sk); 3504 3505 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 3506 /* Partial packet, seq < rcv_next < end_seq */ 3507 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n", 3508 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, 3509 TCP_SKB_CB(skb)->end_seq); 3510 3511 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, tp->rcv_nxt); 3512 3513 /* If window is closed, drop tail of packet. But after 3514 * remembering D-SACK for its head made in previous line. 3515 */ 3516 if (!tcp_receive_window(tp)) 3517 goto out_of_window; 3518 goto queue_and_out; 3519 } 3520 3521 TCP_ECN_check_ce(tp, skb); 3522 3523 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf || 3524 !sk_stream_rmem_schedule(sk, skb)) { 3525 if (tcp_prune_queue(sk) < 0 || 3526 !sk_stream_rmem_schedule(sk, skb)) 3527 goto drop; 3528 } 3529 3530 /* Disable header prediction. */ 3531 tp->pred_flags = 0; 3532 inet_csk_schedule_ack(sk); 3533 3534 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n", 3535 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); 3536 3537 sk_stream_set_owner_r(skb, sk); 3538 3539 if (!skb_peek(&tp->out_of_order_queue)) { 3540 /* Initial out of order segment, build 1 SACK. */ 3541 if (tp->rx_opt.sack_ok) { 3542 tp->rx_opt.num_sacks = 1; 3543 tp->rx_opt.dsack = 0; 3544 tp->rx_opt.eff_sacks = 1; 3545 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq; 3546 tp->selective_acks[0].end_seq = 3547 TCP_SKB_CB(skb)->end_seq; 3548 } 3549 __skb_queue_head(&tp->out_of_order_queue,skb); 3550 } else { 3551 struct sk_buff *skb1 = tp->out_of_order_queue.prev; 3552 u32 seq = TCP_SKB_CB(skb)->seq; 3553 u32 end_seq = TCP_SKB_CB(skb)->end_seq; 3554 3555 if (seq == TCP_SKB_CB(skb1)->end_seq) { 3556 __skb_append(skb1, skb, &tp->out_of_order_queue); 3557 3558 if (!tp->rx_opt.num_sacks || 3559 tp->selective_acks[0].end_seq != seq) 3560 goto add_sack; 3561 3562 /* Common case: data arrive in order after hole. */ 3563 tp->selective_acks[0].end_seq = end_seq; 3564 return; 3565 } 3566 3567 /* Find place to insert this segment. */ 3568 do { 3569 if (!after(TCP_SKB_CB(skb1)->seq, seq)) 3570 break; 3571 } while ((skb1 = skb1->prev) != 3572 (struct sk_buff*)&tp->out_of_order_queue); 3573 3574 /* Do skb overlap to previous one? */ 3575 if (skb1 != (struct sk_buff*)&tp->out_of_order_queue && 3576 before(seq, TCP_SKB_CB(skb1)->end_seq)) { 3577 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) { 3578 /* All the bits are present. Drop. */ 3579 __kfree_skb(skb); 3580 tcp_dsack_set(tp, seq, end_seq); 3581 goto add_sack; 3582 } 3583 if (after(seq, TCP_SKB_CB(skb1)->seq)) { 3584 /* Partial overlap. */ 3585 tcp_dsack_set(tp, seq, TCP_SKB_CB(skb1)->end_seq); 3586 } else { 3587 skb1 = skb1->prev; 3588 } 3589 } 3590 __skb_insert(skb, skb1, skb1->next, &tp->out_of_order_queue); 3591 3592 /* And clean segments covered by new one as whole. */ 3593 while ((skb1 = skb->next) != 3594 (struct sk_buff*)&tp->out_of_order_queue && 3595 after(end_seq, TCP_SKB_CB(skb1)->seq)) { 3596 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) { 3597 tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, end_seq); 3598 break; 3599 } 3600 __skb_unlink(skb1, &tp->out_of_order_queue); 3601 tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, TCP_SKB_CB(skb1)->end_seq); 3602 __kfree_skb(skb1); 3603 } 3604 3605 add_sack: 3606 if (tp->rx_opt.sack_ok) 3607 tcp_sack_new_ofo_skb(sk, seq, end_seq); 3608 } 3609 } 3610 3611 /* Collapse contiguous sequence of skbs head..tail with 3612 * sequence numbers start..end. 3613 * Segments with FIN/SYN are not collapsed (only because this 3614 * simplifies code) 3615 */ 3616 static void 3617 tcp_collapse(struct sock *sk, struct sk_buff_head *list, 3618 struct sk_buff *head, struct sk_buff *tail, 3619 u32 start, u32 end) 3620 { 3621 struct sk_buff *skb; 3622 3623 /* First, check that queue is collapsible and find 3624 * the point where collapsing can be useful. */ 3625 for (skb = head; skb != tail; ) { 3626 /* No new bits? It is possible on ofo queue. */ 3627 if (!before(start, TCP_SKB_CB(skb)->end_seq)) { 3628 struct sk_buff *next = skb->next; 3629 __skb_unlink(skb, list); 3630 __kfree_skb(skb); 3631 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED); 3632 skb = next; 3633 continue; 3634 } 3635 3636 /* The first skb to collapse is: 3637 * - not SYN/FIN and 3638 * - bloated or contains data before "start" or 3639 * overlaps to the next one. 3640 */ 3641 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin && 3642 (tcp_win_from_space(skb->truesize) > skb->len || 3643 before(TCP_SKB_CB(skb)->seq, start) || 3644 (skb->next != tail && 3645 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq))) 3646 break; 3647 3648 /* Decided to skip this, advance start seq. */ 3649 start = TCP_SKB_CB(skb)->end_seq; 3650 skb = skb->next; 3651 } 3652 if (skb == tail || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin) 3653 return; 3654 3655 while (before(start, end)) { 3656 struct sk_buff *nskb; 3657 int header = skb_headroom(skb); 3658 int copy = SKB_MAX_ORDER(header, 0); 3659 3660 /* Too big header? This can happen with IPv6. */ 3661 if (copy < 0) 3662 return; 3663 if (end-start < copy) 3664 copy = end-start; 3665 nskb = alloc_skb(copy+header, GFP_ATOMIC); 3666 if (!nskb) 3667 return; 3668 3669 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head); 3670 skb_set_network_header(nskb, (skb_network_header(skb) - 3671 skb->head)); 3672 skb_set_transport_header(nskb, (skb_transport_header(skb) - 3673 skb->head)); 3674 skb_reserve(nskb, header); 3675 memcpy(nskb->head, skb->head, header); 3676 memcpy(nskb->cb, skb->cb, sizeof(skb->cb)); 3677 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start; 3678 __skb_insert(nskb, skb->prev, skb, list); 3679 sk_stream_set_owner_r(nskb, sk); 3680 3681 /* Copy data, releasing collapsed skbs. */ 3682 while (copy > 0) { 3683 int offset = start - TCP_SKB_CB(skb)->seq; 3684 int size = TCP_SKB_CB(skb)->end_seq - start; 3685 3686 BUG_ON(offset < 0); 3687 if (size > 0) { 3688 size = min(copy, size); 3689 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size)) 3690 BUG(); 3691 TCP_SKB_CB(nskb)->end_seq += size; 3692 copy -= size; 3693 start += size; 3694 } 3695 if (!before(start, TCP_SKB_CB(skb)->end_seq)) { 3696 struct sk_buff *next = skb->next; 3697 __skb_unlink(skb, list); 3698 __kfree_skb(skb); 3699 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED); 3700 skb = next; 3701 if (skb == tail || 3702 tcp_hdr(skb)->syn || 3703 tcp_hdr(skb)->fin) 3704 return; 3705 } 3706 } 3707 } 3708 } 3709 3710 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs 3711 * and tcp_collapse() them until all the queue is collapsed. 3712 */ 3713 static void tcp_collapse_ofo_queue(struct sock *sk) 3714 { 3715 struct tcp_sock *tp = tcp_sk(sk); 3716 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue); 3717 struct sk_buff *head; 3718 u32 start, end; 3719 3720 if (skb == NULL) 3721 return; 3722 3723 start = TCP_SKB_CB(skb)->seq; 3724 end = TCP_SKB_CB(skb)->end_seq; 3725 head = skb; 3726 3727 for (;;) { 3728 skb = skb->next; 3729 3730 /* Segment is terminated when we see gap or when 3731 * we are at the end of all the queue. */ 3732 if (skb == (struct sk_buff *)&tp->out_of_order_queue || 3733 after(TCP_SKB_CB(skb)->seq, end) || 3734 before(TCP_SKB_CB(skb)->end_seq, start)) { 3735 tcp_collapse(sk, &tp->out_of_order_queue, 3736 head, skb, start, end); 3737 head = skb; 3738 if (skb == (struct sk_buff *)&tp->out_of_order_queue) 3739 break; 3740 /* Start new segment */ 3741 start = TCP_SKB_CB(skb)->seq; 3742 end = TCP_SKB_CB(skb)->end_seq; 3743 } else { 3744 if (before(TCP_SKB_CB(skb)->seq, start)) 3745 start = TCP_SKB_CB(skb)->seq; 3746 if (after(TCP_SKB_CB(skb)->end_seq, end)) 3747 end = TCP_SKB_CB(skb)->end_seq; 3748 } 3749 } 3750 } 3751 3752 /* Reduce allocated memory if we can, trying to get 3753 * the socket within its memory limits again. 3754 * 3755 * Return less than zero if we should start dropping frames 3756 * until the socket owning process reads some of the data 3757 * to stabilize the situation. 3758 */ 3759 static int tcp_prune_queue(struct sock *sk) 3760 { 3761 struct tcp_sock *tp = tcp_sk(sk); 3762 3763 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq); 3764 3765 NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED); 3766 3767 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf) 3768 tcp_clamp_window(sk); 3769 else if (tcp_memory_pressure) 3770 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss); 3771 3772 tcp_collapse_ofo_queue(sk); 3773 tcp_collapse(sk, &sk->sk_receive_queue, 3774 sk->sk_receive_queue.next, 3775 (struct sk_buff*)&sk->sk_receive_queue, 3776 tp->copied_seq, tp->rcv_nxt); 3777 sk_stream_mem_reclaim(sk); 3778 3779 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) 3780 return 0; 3781 3782 /* Collapsing did not help, destructive actions follow. 3783 * This must not ever occur. */ 3784 3785 /* First, purge the out_of_order queue. */ 3786 if (!skb_queue_empty(&tp->out_of_order_queue)) { 3787 NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED); 3788 __skb_queue_purge(&tp->out_of_order_queue); 3789 3790 /* Reset SACK state. A conforming SACK implementation will 3791 * do the same at a timeout based retransmit. When a connection 3792 * is in a sad state like this, we care only about integrity 3793 * of the connection not performance. 3794 */ 3795 if (tp->rx_opt.sack_ok) 3796 tcp_sack_reset(&tp->rx_opt); 3797 sk_stream_mem_reclaim(sk); 3798 } 3799 3800 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) 3801 return 0; 3802 3803 /* If we are really being abused, tell the caller to silently 3804 * drop receive data on the floor. It will get retransmitted 3805 * and hopefully then we'll have sufficient space. 3806 */ 3807 NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED); 3808 3809 /* Massive buffer overcommit. */ 3810 tp->pred_flags = 0; 3811 return -1; 3812 } 3813 3814 3815 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. 3816 * As additional protections, we do not touch cwnd in retransmission phases, 3817 * and if application hit its sndbuf limit recently. 3818 */ 3819 void tcp_cwnd_application_limited(struct sock *sk) 3820 { 3821 struct tcp_sock *tp = tcp_sk(sk); 3822 3823 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && 3824 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { 3825 /* Limited by application or receiver window. */ 3826 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); 3827 u32 win_used = max(tp->snd_cwnd_used, init_win); 3828 if (win_used < tp->snd_cwnd) { 3829 tp->snd_ssthresh = tcp_current_ssthresh(sk); 3830 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1; 3831 } 3832 tp->snd_cwnd_used = 0; 3833 } 3834 tp->snd_cwnd_stamp = tcp_time_stamp; 3835 } 3836 3837 static int tcp_should_expand_sndbuf(struct sock *sk) 3838 { 3839 struct tcp_sock *tp = tcp_sk(sk); 3840 3841 /* If the user specified a specific send buffer setting, do 3842 * not modify it. 3843 */ 3844 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK) 3845 return 0; 3846 3847 /* If we are under global TCP memory pressure, do not expand. */ 3848 if (tcp_memory_pressure) 3849 return 0; 3850 3851 /* If we are under soft global TCP memory pressure, do not expand. */ 3852 if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0]) 3853 return 0; 3854 3855 /* If we filled the congestion window, do not expand. */ 3856 if (tp->packets_out >= tp->snd_cwnd) 3857 return 0; 3858 3859 return 1; 3860 } 3861 3862 /* When incoming ACK allowed to free some skb from write_queue, 3863 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket 3864 * on the exit from tcp input handler. 3865 * 3866 * PROBLEM: sndbuf expansion does not work well with largesend. 3867 */ 3868 static void tcp_new_space(struct sock *sk) 3869 { 3870 struct tcp_sock *tp = tcp_sk(sk); 3871 3872 if (tcp_should_expand_sndbuf(sk)) { 3873 int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) + 3874 MAX_TCP_HEADER + 16 + sizeof(struct sk_buff), 3875 demanded = max_t(unsigned int, tp->snd_cwnd, 3876 tp->reordering + 1); 3877 sndmem *= 2*demanded; 3878 if (sndmem > sk->sk_sndbuf) 3879 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]); 3880 tp->snd_cwnd_stamp = tcp_time_stamp; 3881 } 3882 3883 sk->sk_write_space(sk); 3884 } 3885 3886 static void tcp_check_space(struct sock *sk) 3887 { 3888 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) { 3889 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK); 3890 if (sk->sk_socket && 3891 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) 3892 tcp_new_space(sk); 3893 } 3894 } 3895 3896 static inline void tcp_data_snd_check(struct sock *sk) 3897 { 3898 tcp_push_pending_frames(sk); 3899 tcp_check_space(sk); 3900 } 3901 3902 /* 3903 * Check if sending an ack is needed. 3904 */ 3905 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible) 3906 { 3907 struct tcp_sock *tp = tcp_sk(sk); 3908 3909 /* More than one full frame received... */ 3910 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss 3911 /* ... and right edge of window advances far enough. 3912 * (tcp_recvmsg() will send ACK otherwise). Or... 3913 */ 3914 && __tcp_select_window(sk) >= tp->rcv_wnd) || 3915 /* We ACK each frame or... */ 3916 tcp_in_quickack_mode(sk) || 3917 /* We have out of order data. */ 3918 (ofo_possible && 3919 skb_peek(&tp->out_of_order_queue))) { 3920 /* Then ack it now */ 3921 tcp_send_ack(sk); 3922 } else { 3923 /* Else, send delayed ack. */ 3924 tcp_send_delayed_ack(sk); 3925 } 3926 } 3927 3928 static inline void tcp_ack_snd_check(struct sock *sk) 3929 { 3930 if (!inet_csk_ack_scheduled(sk)) { 3931 /* We sent a data segment already. */ 3932 return; 3933 } 3934 __tcp_ack_snd_check(sk, 1); 3935 } 3936 3937 /* 3938 * This routine is only called when we have urgent data 3939 * signaled. Its the 'slow' part of tcp_urg. It could be 3940 * moved inline now as tcp_urg is only called from one 3941 * place. We handle URGent data wrong. We have to - as 3942 * BSD still doesn't use the correction from RFC961. 3943 * For 1003.1g we should support a new option TCP_STDURG to permit 3944 * either form (or just set the sysctl tcp_stdurg). 3945 */ 3946 3947 static void tcp_check_urg(struct sock * sk, struct tcphdr * th) 3948 { 3949 struct tcp_sock *tp = tcp_sk(sk); 3950 u32 ptr = ntohs(th->urg_ptr); 3951 3952 if (ptr && !sysctl_tcp_stdurg) 3953 ptr--; 3954 ptr += ntohl(th->seq); 3955 3956 /* Ignore urgent data that we've already seen and read. */ 3957 if (after(tp->copied_seq, ptr)) 3958 return; 3959 3960 /* Do not replay urg ptr. 3961 * 3962 * NOTE: interesting situation not covered by specs. 3963 * Misbehaving sender may send urg ptr, pointing to segment, 3964 * which we already have in ofo queue. We are not able to fetch 3965 * such data and will stay in TCP_URG_NOTYET until will be eaten 3966 * by recvmsg(). Seems, we are not obliged to handle such wicked 3967 * situations. But it is worth to think about possibility of some 3968 * DoSes using some hypothetical application level deadlock. 3969 */ 3970 if (before(ptr, tp->rcv_nxt)) 3971 return; 3972 3973 /* Do we already have a newer (or duplicate) urgent pointer? */ 3974 if (tp->urg_data && !after(ptr, tp->urg_seq)) 3975 return; 3976 3977 /* Tell the world about our new urgent pointer. */ 3978 sk_send_sigurg(sk); 3979 3980 /* We may be adding urgent data when the last byte read was 3981 * urgent. To do this requires some care. We cannot just ignore 3982 * tp->copied_seq since we would read the last urgent byte again 3983 * as data, nor can we alter copied_seq until this data arrives 3984 * or we break the semantics of SIOCATMARK (and thus sockatmark()) 3985 * 3986 * NOTE. Double Dutch. Rendering to plain English: author of comment 3987 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB); 3988 * and expect that both A and B disappear from stream. This is _wrong_. 3989 * Though this happens in BSD with high probability, this is occasional. 3990 * Any application relying on this is buggy. Note also, that fix "works" 3991 * only in this artificial test. Insert some normal data between A and B and we will 3992 * decline of BSD again. Verdict: it is better to remove to trap 3993 * buggy users. 3994 */ 3995 if (tp->urg_seq == tp->copied_seq && tp->urg_data && 3996 !sock_flag(sk, SOCK_URGINLINE) && 3997 tp->copied_seq != tp->rcv_nxt) { 3998 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue); 3999 tp->copied_seq++; 4000 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) { 4001 __skb_unlink(skb, &sk->sk_receive_queue); 4002 __kfree_skb(skb); 4003 } 4004 } 4005 4006 tp->urg_data = TCP_URG_NOTYET; 4007 tp->urg_seq = ptr; 4008 4009 /* Disable header prediction. */ 4010 tp->pred_flags = 0; 4011 } 4012 4013 /* This is the 'fast' part of urgent handling. */ 4014 static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th) 4015 { 4016 struct tcp_sock *tp = tcp_sk(sk); 4017 4018 /* Check if we get a new urgent pointer - normally not. */ 4019 if (th->urg) 4020 tcp_check_urg(sk,th); 4021 4022 /* Do we wait for any urgent data? - normally not... */ 4023 if (tp->urg_data == TCP_URG_NOTYET) { 4024 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) - 4025 th->syn; 4026 4027 /* Is the urgent pointer pointing into this packet? */ 4028 if (ptr < skb->len) { 4029 u8 tmp; 4030 if (skb_copy_bits(skb, ptr, &tmp, 1)) 4031 BUG(); 4032 tp->urg_data = TCP_URG_VALID | tmp; 4033 if (!sock_flag(sk, SOCK_DEAD)) 4034 sk->sk_data_ready(sk, 0); 4035 } 4036 } 4037 } 4038 4039 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen) 4040 { 4041 struct tcp_sock *tp = tcp_sk(sk); 4042 int chunk = skb->len - hlen; 4043 int err; 4044 4045 local_bh_enable(); 4046 if (skb_csum_unnecessary(skb)) 4047 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk); 4048 else 4049 err = skb_copy_and_csum_datagram_iovec(skb, hlen, 4050 tp->ucopy.iov); 4051 4052 if (!err) { 4053 tp->ucopy.len -= chunk; 4054 tp->copied_seq += chunk; 4055 tcp_rcv_space_adjust(sk); 4056 } 4057 4058 local_bh_disable(); 4059 return err; 4060 } 4061 4062 static __sum16 __tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb) 4063 { 4064 __sum16 result; 4065 4066 if (sock_owned_by_user(sk)) { 4067 local_bh_enable(); 4068 result = __tcp_checksum_complete(skb); 4069 local_bh_disable(); 4070 } else { 4071 result = __tcp_checksum_complete(skb); 4072 } 4073 return result; 4074 } 4075 4076 static inline int tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb) 4077 { 4078 return !skb_csum_unnecessary(skb) && 4079 __tcp_checksum_complete_user(sk, skb); 4080 } 4081 4082 #ifdef CONFIG_NET_DMA 4083 static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb, int hlen) 4084 { 4085 struct tcp_sock *tp = tcp_sk(sk); 4086 int chunk = skb->len - hlen; 4087 int dma_cookie; 4088 int copied_early = 0; 4089 4090 if (tp->ucopy.wakeup) 4091 return 0; 4092 4093 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list) 4094 tp->ucopy.dma_chan = get_softnet_dma(); 4095 4096 if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) { 4097 4098 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan, 4099 skb, hlen, tp->ucopy.iov, chunk, tp->ucopy.pinned_list); 4100 4101 if (dma_cookie < 0) 4102 goto out; 4103 4104 tp->ucopy.dma_cookie = dma_cookie; 4105 copied_early = 1; 4106 4107 tp->ucopy.len -= chunk; 4108 tp->copied_seq += chunk; 4109 tcp_rcv_space_adjust(sk); 4110 4111 if ((tp->ucopy.len == 0) || 4112 (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) || 4113 (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) { 4114 tp->ucopy.wakeup = 1; 4115 sk->sk_data_ready(sk, 0); 4116 } 4117 } else if (chunk > 0) { 4118 tp->ucopy.wakeup = 1; 4119 sk->sk_data_ready(sk, 0); 4120 } 4121 out: 4122 return copied_early; 4123 } 4124 #endif /* CONFIG_NET_DMA */ 4125 4126 /* 4127 * TCP receive function for the ESTABLISHED state. 4128 * 4129 * It is split into a fast path and a slow path. The fast path is 4130 * disabled when: 4131 * - A zero window was announced from us - zero window probing 4132 * is only handled properly in the slow path. 4133 * - Out of order segments arrived. 4134 * - Urgent data is expected. 4135 * - There is no buffer space left 4136 * - Unexpected TCP flags/window values/header lengths are received 4137 * (detected by checking the TCP header against pred_flags) 4138 * - Data is sent in both directions. Fast path only supports pure senders 4139 * or pure receivers (this means either the sequence number or the ack 4140 * value must stay constant) 4141 * - Unexpected TCP option. 4142 * 4143 * When these conditions are not satisfied it drops into a standard 4144 * receive procedure patterned after RFC793 to handle all cases. 4145 * The first three cases are guaranteed by proper pred_flags setting, 4146 * the rest is checked inline. Fast processing is turned on in 4147 * tcp_data_queue when everything is OK. 4148 */ 4149 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb, 4150 struct tcphdr *th, unsigned len) 4151 { 4152 struct tcp_sock *tp = tcp_sk(sk); 4153 4154 /* 4155 * Header prediction. 4156 * The code loosely follows the one in the famous 4157 * "30 instruction TCP receive" Van Jacobson mail. 4158 * 4159 * Van's trick is to deposit buffers into socket queue 4160 * on a device interrupt, to call tcp_recv function 4161 * on the receive process context and checksum and copy 4162 * the buffer to user space. smart... 4163 * 4164 * Our current scheme is not silly either but we take the 4165 * extra cost of the net_bh soft interrupt processing... 4166 * We do checksum and copy also but from device to kernel. 4167 */ 4168 4169 tp->rx_opt.saw_tstamp = 0; 4170 4171 /* pred_flags is 0xS?10 << 16 + snd_wnd 4172 * if header_prediction is to be made 4173 * 'S' will always be tp->tcp_header_len >> 2 4174 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to 4175 * turn it off (when there are holes in the receive 4176 * space for instance) 4177 * PSH flag is ignored. 4178 */ 4179 4180 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags && 4181 TCP_SKB_CB(skb)->seq == tp->rcv_nxt) { 4182 int tcp_header_len = tp->tcp_header_len; 4183 4184 /* Timestamp header prediction: tcp_header_len 4185 * is automatically equal to th->doff*4 due to pred_flags 4186 * match. 4187 */ 4188 4189 /* Check timestamp */ 4190 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) { 4191 __be32 *ptr = (__be32 *)(th + 1); 4192 4193 /* No? Slow path! */ 4194 if (*ptr != htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) 4195 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) 4196 goto slow_path; 4197 4198 tp->rx_opt.saw_tstamp = 1; 4199 ++ptr; 4200 tp->rx_opt.rcv_tsval = ntohl(*ptr); 4201 ++ptr; 4202 tp->rx_opt.rcv_tsecr = ntohl(*ptr); 4203 4204 /* If PAWS failed, check it more carefully in slow path */ 4205 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0) 4206 goto slow_path; 4207 4208 /* DO NOT update ts_recent here, if checksum fails 4209 * and timestamp was corrupted part, it will result 4210 * in a hung connection since we will drop all 4211 * future packets due to the PAWS test. 4212 */ 4213 } 4214 4215 if (len <= tcp_header_len) { 4216 /* Bulk data transfer: sender */ 4217 if (len == tcp_header_len) { 4218 /* Predicted packet is in window by definition. 4219 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 4220 * Hence, check seq<=rcv_wup reduces to: 4221 */ 4222 if (tcp_header_len == 4223 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && 4224 tp->rcv_nxt == tp->rcv_wup) 4225 tcp_store_ts_recent(tp); 4226 4227 /* We know that such packets are checksummed 4228 * on entry. 4229 */ 4230 tcp_ack(sk, skb, 0); 4231 __kfree_skb(skb); 4232 tcp_data_snd_check(sk); 4233 return 0; 4234 } else { /* Header too small */ 4235 TCP_INC_STATS_BH(TCP_MIB_INERRS); 4236 goto discard; 4237 } 4238 } else { 4239 int eaten = 0; 4240 int copied_early = 0; 4241 4242 if (tp->copied_seq == tp->rcv_nxt && 4243 len - tcp_header_len <= tp->ucopy.len) { 4244 #ifdef CONFIG_NET_DMA 4245 if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) { 4246 copied_early = 1; 4247 eaten = 1; 4248 } 4249 #endif 4250 if (tp->ucopy.task == current && sock_owned_by_user(sk) && !copied_early) { 4251 __set_current_state(TASK_RUNNING); 4252 4253 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len)) 4254 eaten = 1; 4255 } 4256 if (eaten) { 4257 /* Predicted packet is in window by definition. 4258 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 4259 * Hence, check seq<=rcv_wup reduces to: 4260 */ 4261 if (tcp_header_len == 4262 (sizeof(struct tcphdr) + 4263 TCPOLEN_TSTAMP_ALIGNED) && 4264 tp->rcv_nxt == tp->rcv_wup) 4265 tcp_store_ts_recent(tp); 4266 4267 tcp_rcv_rtt_measure_ts(sk, skb); 4268 4269 __skb_pull(skb, tcp_header_len); 4270 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; 4271 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER); 4272 } 4273 if (copied_early) 4274 tcp_cleanup_rbuf(sk, skb->len); 4275 } 4276 if (!eaten) { 4277 if (tcp_checksum_complete_user(sk, skb)) 4278 goto csum_error; 4279 4280 /* Predicted packet is in window by definition. 4281 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 4282 * Hence, check seq<=rcv_wup reduces to: 4283 */ 4284 if (tcp_header_len == 4285 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && 4286 tp->rcv_nxt == tp->rcv_wup) 4287 tcp_store_ts_recent(tp); 4288 4289 tcp_rcv_rtt_measure_ts(sk, skb); 4290 4291 if ((int)skb->truesize > sk->sk_forward_alloc) 4292 goto step5; 4293 4294 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS); 4295 4296 /* Bulk data transfer: receiver */ 4297 __skb_pull(skb,tcp_header_len); 4298 __skb_queue_tail(&sk->sk_receive_queue, skb); 4299 sk_stream_set_owner_r(skb, sk); 4300 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; 4301 } 4302 4303 tcp_event_data_recv(sk, skb); 4304 4305 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) { 4306 /* Well, only one small jumplet in fast path... */ 4307 tcp_ack(sk, skb, FLAG_DATA); 4308 tcp_data_snd_check(sk); 4309 if (!inet_csk_ack_scheduled(sk)) 4310 goto no_ack; 4311 } 4312 4313 __tcp_ack_snd_check(sk, 0); 4314 no_ack: 4315 #ifdef CONFIG_NET_DMA 4316 if (copied_early) 4317 __skb_queue_tail(&sk->sk_async_wait_queue, skb); 4318 else 4319 #endif 4320 if (eaten) 4321 __kfree_skb(skb); 4322 else 4323 sk->sk_data_ready(sk, 0); 4324 return 0; 4325 } 4326 } 4327 4328 slow_path: 4329 if (len < (th->doff<<2) || tcp_checksum_complete_user(sk, skb)) 4330 goto csum_error; 4331 4332 /* 4333 * RFC1323: H1. Apply PAWS check first. 4334 */ 4335 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp && 4336 tcp_paws_discard(sk, skb)) { 4337 if (!th->rst) { 4338 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED); 4339 tcp_send_dupack(sk, skb); 4340 goto discard; 4341 } 4342 /* Resets are accepted even if PAWS failed. 4343 4344 ts_recent update must be made after we are sure 4345 that the packet is in window. 4346 */ 4347 } 4348 4349 /* 4350 * Standard slow path. 4351 */ 4352 4353 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) { 4354 /* RFC793, page 37: "In all states except SYN-SENT, all reset 4355 * (RST) segments are validated by checking their SEQ-fields." 4356 * And page 69: "If an incoming segment is not acceptable, 4357 * an acknowledgment should be sent in reply (unless the RST bit 4358 * is set, if so drop the segment and return)". 4359 */ 4360 if (!th->rst) 4361 tcp_send_dupack(sk, skb); 4362 goto discard; 4363 } 4364 4365 if (th->rst) { 4366 tcp_reset(sk); 4367 goto discard; 4368 } 4369 4370 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq); 4371 4372 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 4373 TCP_INC_STATS_BH(TCP_MIB_INERRS); 4374 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN); 4375 tcp_reset(sk); 4376 return 1; 4377 } 4378 4379 step5: 4380 if (th->ack) 4381 tcp_ack(sk, skb, FLAG_SLOWPATH); 4382 4383 tcp_rcv_rtt_measure_ts(sk, skb); 4384 4385 /* Process urgent data. */ 4386 tcp_urg(sk, skb, th); 4387 4388 /* step 7: process the segment text */ 4389 tcp_data_queue(sk, skb); 4390 4391 tcp_data_snd_check(sk); 4392 tcp_ack_snd_check(sk); 4393 return 0; 4394 4395 csum_error: 4396 TCP_INC_STATS_BH(TCP_MIB_INERRS); 4397 4398 discard: 4399 __kfree_skb(skb); 4400 return 0; 4401 } 4402 4403 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb, 4404 struct tcphdr *th, unsigned len) 4405 { 4406 struct tcp_sock *tp = tcp_sk(sk); 4407 struct inet_connection_sock *icsk = inet_csk(sk); 4408 int saved_clamp = tp->rx_opt.mss_clamp; 4409 4410 tcp_parse_options(skb, &tp->rx_opt, 0); 4411 4412 if (th->ack) { 4413 /* rfc793: 4414 * "If the state is SYN-SENT then 4415 * first check the ACK bit 4416 * If the ACK bit is set 4417 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send 4418 * a reset (unless the RST bit is set, if so drop 4419 * the segment and return)" 4420 * 4421 * We do not send data with SYN, so that RFC-correct 4422 * test reduces to: 4423 */ 4424 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt) 4425 goto reset_and_undo; 4426 4427 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 4428 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp, 4429 tcp_time_stamp)) { 4430 NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED); 4431 goto reset_and_undo; 4432 } 4433 4434 /* Now ACK is acceptable. 4435 * 4436 * "If the RST bit is set 4437 * If the ACK was acceptable then signal the user "error: 4438 * connection reset", drop the segment, enter CLOSED state, 4439 * delete TCB, and return." 4440 */ 4441 4442 if (th->rst) { 4443 tcp_reset(sk); 4444 goto discard; 4445 } 4446 4447 /* rfc793: 4448 * "fifth, if neither of the SYN or RST bits is set then 4449 * drop the segment and return." 4450 * 4451 * See note below! 4452 * --ANK(990513) 4453 */ 4454 if (!th->syn) 4455 goto discard_and_undo; 4456 4457 /* rfc793: 4458 * "If the SYN bit is on ... 4459 * are acceptable then ... 4460 * (our SYN has been ACKed), change the connection 4461 * state to ESTABLISHED..." 4462 */ 4463 4464 TCP_ECN_rcv_synack(tp, th); 4465 4466 tp->snd_wl1 = TCP_SKB_CB(skb)->seq; 4467 tcp_ack(sk, skb, FLAG_SLOWPATH); 4468 4469 /* Ok.. it's good. Set up sequence numbers and 4470 * move to established. 4471 */ 4472 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; 4473 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; 4474 4475 /* RFC1323: The window in SYN & SYN/ACK segments is 4476 * never scaled. 4477 */ 4478 tp->snd_wnd = ntohs(th->window); 4479 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(skb)->seq); 4480 4481 if (!tp->rx_opt.wscale_ok) { 4482 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0; 4483 tp->window_clamp = min(tp->window_clamp, 65535U); 4484 } 4485 4486 if (tp->rx_opt.saw_tstamp) { 4487 tp->rx_opt.tstamp_ok = 1; 4488 tp->tcp_header_len = 4489 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 4490 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; 4491 tcp_store_ts_recent(tp); 4492 } else { 4493 tp->tcp_header_len = sizeof(struct tcphdr); 4494 } 4495 4496 if (tp->rx_opt.sack_ok && sysctl_tcp_fack) 4497 tp->rx_opt.sack_ok |= 2; 4498 4499 tcp_mtup_init(sk); 4500 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 4501 tcp_initialize_rcv_mss(sk); 4502 4503 /* Remember, tcp_poll() does not lock socket! 4504 * Change state from SYN-SENT only after copied_seq 4505 * is initialized. */ 4506 tp->copied_seq = tp->rcv_nxt; 4507 smp_mb(); 4508 tcp_set_state(sk, TCP_ESTABLISHED); 4509 4510 security_inet_conn_established(sk, skb); 4511 4512 /* Make sure socket is routed, for correct metrics. */ 4513 icsk->icsk_af_ops->rebuild_header(sk); 4514 4515 tcp_init_metrics(sk); 4516 4517 tcp_init_congestion_control(sk); 4518 4519 /* Prevent spurious tcp_cwnd_restart() on first data 4520 * packet. 4521 */ 4522 tp->lsndtime = tcp_time_stamp; 4523 4524 tcp_init_buffer_space(sk); 4525 4526 if (sock_flag(sk, SOCK_KEEPOPEN)) 4527 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp)); 4528 4529 if (!tp->rx_opt.snd_wscale) 4530 __tcp_fast_path_on(tp, tp->snd_wnd); 4531 else 4532 tp->pred_flags = 0; 4533 4534 if (!sock_flag(sk, SOCK_DEAD)) { 4535 sk->sk_state_change(sk); 4536 sk_wake_async(sk, 0, POLL_OUT); 4537 } 4538 4539 if (sk->sk_write_pending || 4540 icsk->icsk_accept_queue.rskq_defer_accept || 4541 icsk->icsk_ack.pingpong) { 4542 /* Save one ACK. Data will be ready after 4543 * several ticks, if write_pending is set. 4544 * 4545 * It may be deleted, but with this feature tcpdumps 4546 * look so _wonderfully_ clever, that I was not able 4547 * to stand against the temptation 8) --ANK 4548 */ 4549 inet_csk_schedule_ack(sk); 4550 icsk->icsk_ack.lrcvtime = tcp_time_stamp; 4551 icsk->icsk_ack.ato = TCP_ATO_MIN; 4552 tcp_incr_quickack(sk); 4553 tcp_enter_quickack_mode(sk); 4554 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 4555 TCP_DELACK_MAX, TCP_RTO_MAX); 4556 4557 discard: 4558 __kfree_skb(skb); 4559 return 0; 4560 } else { 4561 tcp_send_ack(sk); 4562 } 4563 return -1; 4564 } 4565 4566 /* No ACK in the segment */ 4567 4568 if (th->rst) { 4569 /* rfc793: 4570 * "If the RST bit is set 4571 * 4572 * Otherwise (no ACK) drop the segment and return." 4573 */ 4574 4575 goto discard_and_undo; 4576 } 4577 4578 /* PAWS check. */ 4579 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && tcp_paws_check(&tp->rx_opt, 0)) 4580 goto discard_and_undo; 4581 4582 if (th->syn) { 4583 /* We see SYN without ACK. It is attempt of 4584 * simultaneous connect with crossed SYNs. 4585 * Particularly, it can be connect to self. 4586 */ 4587 tcp_set_state(sk, TCP_SYN_RECV); 4588 4589 if (tp->rx_opt.saw_tstamp) { 4590 tp->rx_opt.tstamp_ok = 1; 4591 tcp_store_ts_recent(tp); 4592 tp->tcp_header_len = 4593 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 4594 } else { 4595 tp->tcp_header_len = sizeof(struct tcphdr); 4596 } 4597 4598 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; 4599 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; 4600 4601 /* RFC1323: The window in SYN & SYN/ACK segments is 4602 * never scaled. 4603 */ 4604 tp->snd_wnd = ntohs(th->window); 4605 tp->snd_wl1 = TCP_SKB_CB(skb)->seq; 4606 tp->max_window = tp->snd_wnd; 4607 4608 TCP_ECN_rcv_syn(tp, th); 4609 4610 tcp_mtup_init(sk); 4611 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 4612 tcp_initialize_rcv_mss(sk); 4613 4614 4615 tcp_send_synack(sk); 4616 #if 0 4617 /* Note, we could accept data and URG from this segment. 4618 * There are no obstacles to make this. 4619 * 4620 * However, if we ignore data in ACKless segments sometimes, 4621 * we have no reasons to accept it sometimes. 4622 * Also, seems the code doing it in step6 of tcp_rcv_state_process 4623 * is not flawless. So, discard packet for sanity. 4624 * Uncomment this return to process the data. 4625 */ 4626 return -1; 4627 #else 4628 goto discard; 4629 #endif 4630 } 4631 /* "fifth, if neither of the SYN or RST bits is set then 4632 * drop the segment and return." 4633 */ 4634 4635 discard_and_undo: 4636 tcp_clear_options(&tp->rx_opt); 4637 tp->rx_opt.mss_clamp = saved_clamp; 4638 goto discard; 4639 4640 reset_and_undo: 4641 tcp_clear_options(&tp->rx_opt); 4642 tp->rx_opt.mss_clamp = saved_clamp; 4643 return 1; 4644 } 4645 4646 4647 /* 4648 * This function implements the receiving procedure of RFC 793 for 4649 * all states except ESTABLISHED and TIME_WAIT. 4650 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be 4651 * address independent. 4652 */ 4653 4654 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb, 4655 struct tcphdr *th, unsigned len) 4656 { 4657 struct tcp_sock *tp = tcp_sk(sk); 4658 struct inet_connection_sock *icsk = inet_csk(sk); 4659 int queued = 0; 4660 4661 tp->rx_opt.saw_tstamp = 0; 4662 4663 switch (sk->sk_state) { 4664 case TCP_CLOSE: 4665 goto discard; 4666 4667 case TCP_LISTEN: 4668 if (th->ack) 4669 return 1; 4670 4671 if (th->rst) 4672 goto discard; 4673 4674 if (th->syn) { 4675 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0) 4676 return 1; 4677 4678 /* Now we have several options: In theory there is 4679 * nothing else in the frame. KA9Q has an option to 4680 * send data with the syn, BSD accepts data with the 4681 * syn up to the [to be] advertised window and 4682 * Solaris 2.1 gives you a protocol error. For now 4683 * we just ignore it, that fits the spec precisely 4684 * and avoids incompatibilities. It would be nice in 4685 * future to drop through and process the data. 4686 * 4687 * Now that TTCP is starting to be used we ought to 4688 * queue this data. 4689 * But, this leaves one open to an easy denial of 4690 * service attack, and SYN cookies can't defend 4691 * against this problem. So, we drop the data 4692 * in the interest of security over speed unless 4693 * it's still in use. 4694 */ 4695 kfree_skb(skb); 4696 return 0; 4697 } 4698 goto discard; 4699 4700 case TCP_SYN_SENT: 4701 queued = tcp_rcv_synsent_state_process(sk, skb, th, len); 4702 if (queued >= 0) 4703 return queued; 4704 4705 /* Do step6 onward by hand. */ 4706 tcp_urg(sk, skb, th); 4707 __kfree_skb(skb); 4708 tcp_data_snd_check(sk); 4709 return 0; 4710 } 4711 4712 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp && 4713 tcp_paws_discard(sk, skb)) { 4714 if (!th->rst) { 4715 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED); 4716 tcp_send_dupack(sk, skb); 4717 goto discard; 4718 } 4719 /* Reset is accepted even if it did not pass PAWS. */ 4720 } 4721 4722 /* step 1: check sequence number */ 4723 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) { 4724 if (!th->rst) 4725 tcp_send_dupack(sk, skb); 4726 goto discard; 4727 } 4728 4729 /* step 2: check RST bit */ 4730 if (th->rst) { 4731 tcp_reset(sk); 4732 goto discard; 4733 } 4734 4735 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq); 4736 4737 /* step 3: check security and precedence [ignored] */ 4738 4739 /* step 4: 4740 * 4741 * Check for a SYN in window. 4742 */ 4743 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 4744 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN); 4745 tcp_reset(sk); 4746 return 1; 4747 } 4748 4749 /* step 5: check the ACK field */ 4750 if (th->ack) { 4751 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH); 4752 4753 switch (sk->sk_state) { 4754 case TCP_SYN_RECV: 4755 if (acceptable) { 4756 tp->copied_seq = tp->rcv_nxt; 4757 smp_mb(); 4758 tcp_set_state(sk, TCP_ESTABLISHED); 4759 sk->sk_state_change(sk); 4760 4761 /* Note, that this wakeup is only for marginal 4762 * crossed SYN case. Passively open sockets 4763 * are not waked up, because sk->sk_sleep == 4764 * NULL and sk->sk_socket == NULL. 4765 */ 4766 if (sk->sk_socket) { 4767 sk_wake_async(sk,0,POLL_OUT); 4768 } 4769 4770 tp->snd_una = TCP_SKB_CB(skb)->ack_seq; 4771 tp->snd_wnd = ntohs(th->window) << 4772 tp->rx_opt.snd_wscale; 4773 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, 4774 TCP_SKB_CB(skb)->seq); 4775 4776 /* tcp_ack considers this ACK as duplicate 4777 * and does not calculate rtt. 4778 * Fix it at least with timestamps. 4779 */ 4780 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 4781 !tp->srtt) 4782 tcp_ack_saw_tstamp(sk, 0); 4783 4784 if (tp->rx_opt.tstamp_ok) 4785 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; 4786 4787 /* Make sure socket is routed, for 4788 * correct metrics. 4789 */ 4790 icsk->icsk_af_ops->rebuild_header(sk); 4791 4792 tcp_init_metrics(sk); 4793 4794 tcp_init_congestion_control(sk); 4795 4796 /* Prevent spurious tcp_cwnd_restart() on 4797 * first data packet. 4798 */ 4799 tp->lsndtime = tcp_time_stamp; 4800 4801 tcp_mtup_init(sk); 4802 tcp_initialize_rcv_mss(sk); 4803 tcp_init_buffer_space(sk); 4804 tcp_fast_path_on(tp); 4805 } else { 4806 return 1; 4807 } 4808 break; 4809 4810 case TCP_FIN_WAIT1: 4811 if (tp->snd_una == tp->write_seq) { 4812 tcp_set_state(sk, TCP_FIN_WAIT2); 4813 sk->sk_shutdown |= SEND_SHUTDOWN; 4814 dst_confirm(sk->sk_dst_cache); 4815 4816 if (!sock_flag(sk, SOCK_DEAD)) 4817 /* Wake up lingering close() */ 4818 sk->sk_state_change(sk); 4819 else { 4820 int tmo; 4821 4822 if (tp->linger2 < 0 || 4823 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 4824 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) { 4825 tcp_done(sk); 4826 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA); 4827 return 1; 4828 } 4829 4830 tmo = tcp_fin_time(sk); 4831 if (tmo > TCP_TIMEWAIT_LEN) { 4832 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN); 4833 } else if (th->fin || sock_owned_by_user(sk)) { 4834 /* Bad case. We could lose such FIN otherwise. 4835 * It is not a big problem, but it looks confusing 4836 * and not so rare event. We still can lose it now, 4837 * if it spins in bh_lock_sock(), but it is really 4838 * marginal case. 4839 */ 4840 inet_csk_reset_keepalive_timer(sk, tmo); 4841 } else { 4842 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo); 4843 goto discard; 4844 } 4845 } 4846 } 4847 break; 4848 4849 case TCP_CLOSING: 4850 if (tp->snd_una == tp->write_seq) { 4851 tcp_time_wait(sk, TCP_TIME_WAIT, 0); 4852 goto discard; 4853 } 4854 break; 4855 4856 case TCP_LAST_ACK: 4857 if (tp->snd_una == tp->write_seq) { 4858 tcp_update_metrics(sk); 4859 tcp_done(sk); 4860 goto discard; 4861 } 4862 break; 4863 } 4864 } else 4865 goto discard; 4866 4867 /* step 6: check the URG bit */ 4868 tcp_urg(sk, skb, th); 4869 4870 /* step 7: process the segment text */ 4871 switch (sk->sk_state) { 4872 case TCP_CLOSE_WAIT: 4873 case TCP_CLOSING: 4874 case TCP_LAST_ACK: 4875 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) 4876 break; 4877 case TCP_FIN_WAIT1: 4878 case TCP_FIN_WAIT2: 4879 /* RFC 793 says to queue data in these states, 4880 * RFC 1122 says we MUST send a reset. 4881 * BSD 4.4 also does reset. 4882 */ 4883 if (sk->sk_shutdown & RCV_SHUTDOWN) { 4884 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 4885 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) { 4886 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA); 4887 tcp_reset(sk); 4888 return 1; 4889 } 4890 } 4891 /* Fall through */ 4892 case TCP_ESTABLISHED: 4893 tcp_data_queue(sk, skb); 4894 queued = 1; 4895 break; 4896 } 4897 4898 /* tcp_data could move socket to TIME-WAIT */ 4899 if (sk->sk_state != TCP_CLOSE) { 4900 tcp_data_snd_check(sk); 4901 tcp_ack_snd_check(sk); 4902 } 4903 4904 if (!queued) { 4905 discard: 4906 __kfree_skb(skb); 4907 } 4908 return 0; 4909 } 4910 4911 EXPORT_SYMBOL(sysctl_tcp_ecn); 4912 EXPORT_SYMBOL(sysctl_tcp_reordering); 4913 EXPORT_SYMBOL(tcp_parse_options); 4914 EXPORT_SYMBOL(tcp_rcv_established); 4915 EXPORT_SYMBOL(tcp_rcv_state_process); 4916 EXPORT_SYMBOL(tcp_initialize_rcv_mss); 4917