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