1 /* 2 * INET An implementation of the TCP/IP protocol suite for the LINUX 3 * operating system. INET is implemented using the BSD Socket 4 * interface as the means of communication with the user level. 5 * 6 * Implementation of the Transmission Control Protocol(TCP). 7 * 8 * Version: $Id: tcp_input.c,v 1.243 2002/02/01 22:01:04 davem Exp $ 9 * 10 * Authors: Ross Biro 11 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> 12 * Mark Evans, <evansmp@uhura.aston.ac.uk> 13 * Corey Minyard <wf-rch!minyard@relay.EU.net> 14 * Florian La Roche, <flla@stud.uni-sb.de> 15 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu> 16 * Linus Torvalds, <torvalds@cs.helsinki.fi> 17 * Alan Cox, <gw4pts@gw4pts.ampr.org> 18 * Matthew Dillon, <dillon@apollo.west.oic.com> 19 * Arnt Gulbrandsen, <agulbra@nvg.unit.no> 20 * Jorge Cwik, <jorge@laser.satlink.net> 21 */ 22 23 /* 24 * Changes: 25 * Pedro Roque : Fast Retransmit/Recovery. 26 * Two receive queues. 27 * Retransmit queue handled by TCP. 28 * Better retransmit timer handling. 29 * New congestion avoidance. 30 * Header prediction. 31 * Variable renaming. 32 * 33 * Eric : Fast Retransmit. 34 * Randy Scott : MSS option defines. 35 * Eric Schenk : Fixes to slow start algorithm. 36 * Eric Schenk : Yet another double ACK bug. 37 * Eric Schenk : Delayed ACK bug fixes. 38 * Eric Schenk : Floyd style fast retrans war avoidance. 39 * David S. Miller : Don't allow zero congestion window. 40 * Eric Schenk : Fix retransmitter so that it sends 41 * next packet on ack of previous packet. 42 * Andi Kleen : Moved open_request checking here 43 * and process RSTs for open_requests. 44 * Andi Kleen : Better prune_queue, and other fixes. 45 * Andrey Savochkin: Fix RTT measurements in the presence of 46 * timestamps. 47 * Andrey Savochkin: Check sequence numbers correctly when 48 * removing SACKs due to in sequence incoming 49 * data segments. 50 * Andi Kleen: Make sure we never ack data there is not 51 * enough room for. Also make this condition 52 * a fatal error if it might still happen. 53 * Andi Kleen: Add tcp_measure_rcv_mss to make 54 * connections with MSS<min(MTU,ann. MSS) 55 * work without delayed acks. 56 * Andi Kleen: Process packets with PSH set in the 57 * fast path. 58 * J Hadi Salim: ECN support 59 * Andrei Gurtov, 60 * Pasi Sarolahti, 61 * Panu Kuhlberg: Experimental audit of TCP (re)transmission 62 * engine. Lots of bugs are found. 63 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs 64 */ 65 66 #include <linux/mm.h> 67 #include <linux/module.h> 68 #include <linux/sysctl.h> 69 #include <net/tcp.h> 70 #include <net/inet_common.h> 71 #include <linux/ipsec.h> 72 #include <asm/unaligned.h> 73 #include <net/netdma.h> 74 75 int sysctl_tcp_timestamps __read_mostly = 1; 76 int sysctl_tcp_window_scaling __read_mostly = 1; 77 int sysctl_tcp_sack __read_mostly = 1; 78 int sysctl_tcp_fack __read_mostly = 1; 79 int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH; 80 int sysctl_tcp_ecn __read_mostly; 81 int sysctl_tcp_dsack __read_mostly = 1; 82 int sysctl_tcp_app_win __read_mostly = 31; 83 int sysctl_tcp_adv_win_scale __read_mostly = 2; 84 85 int sysctl_tcp_stdurg __read_mostly; 86 int sysctl_tcp_rfc1337 __read_mostly; 87 int sysctl_tcp_max_orphans __read_mostly = NR_FILE; 88 int sysctl_tcp_frto __read_mostly; 89 int sysctl_tcp_frto_response __read_mostly; 90 int sysctl_tcp_nometrics_save __read_mostly; 91 92 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1; 93 int sysctl_tcp_abc __read_mostly; 94 95 #define FLAG_DATA 0x01 /* Incoming frame contained data. */ 96 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */ 97 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */ 98 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */ 99 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */ 100 #define FLAG_DATA_SACKED 0x20 /* New SACK. */ 101 #define FLAG_ECE 0x40 /* ECE in this ACK */ 102 #define FLAG_DATA_LOST 0x80 /* SACK detected data lossage. */ 103 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/ 104 #define FLAG_ONLY_ORIG_SACKED 0x200 /* SACKs only non-rexmit sent before RTO */ 105 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */ 106 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained DSACK info */ 107 108 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED) 109 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED) 110 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE) 111 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED) 112 #define FLAG_ANY_PROGRESS (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED) 113 114 #define IsReno(tp) ((tp)->rx_opt.sack_ok == 0) 115 #define IsFack(tp) ((tp)->rx_opt.sack_ok & 2) 116 #define IsDSack(tp) ((tp)->rx_opt.sack_ok & 4) 117 118 #define IsSackFrto() (sysctl_tcp_frto == 0x2) 119 120 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH) 121 122 /* Adapt the MSS value used to make delayed ack decision to the 123 * real world. 124 */ 125 static void tcp_measure_rcv_mss(struct sock *sk, 126 const struct sk_buff *skb) 127 { 128 struct inet_connection_sock *icsk = inet_csk(sk); 129 const unsigned int lss = icsk->icsk_ack.last_seg_size; 130 unsigned int len; 131 132 icsk->icsk_ack.last_seg_size = 0; 133 134 /* skb->len may jitter because of SACKs, even if peer 135 * sends good full-sized frames. 136 */ 137 len = skb_shinfo(skb)->gso_size ?: skb->len; 138 if (len >= icsk->icsk_ack.rcv_mss) { 139 icsk->icsk_ack.rcv_mss = len; 140 } else { 141 /* Otherwise, we make more careful check taking into account, 142 * that SACKs block is variable. 143 * 144 * "len" is invariant segment length, including TCP header. 145 */ 146 len += skb->data - skb_transport_header(skb); 147 if (len >= TCP_MIN_RCVMSS + sizeof(struct tcphdr) || 148 /* If PSH is not set, packet should be 149 * full sized, provided peer TCP is not badly broken. 150 * This observation (if it is correct 8)) allows 151 * to handle super-low mtu links fairly. 152 */ 153 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) && 154 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) { 155 /* Subtract also invariant (if peer is RFC compliant), 156 * tcp header plus fixed timestamp option length. 157 * Resulting "len" is MSS free of SACK jitter. 158 */ 159 len -= tcp_sk(sk)->tcp_header_len; 160 icsk->icsk_ack.last_seg_size = len; 161 if (len == lss) { 162 icsk->icsk_ack.rcv_mss = len; 163 return; 164 } 165 } 166 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED) 167 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2; 168 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED; 169 } 170 } 171 172 static void tcp_incr_quickack(struct sock *sk) 173 { 174 struct inet_connection_sock *icsk = inet_csk(sk); 175 unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss); 176 177 if (quickacks==0) 178 quickacks=2; 179 if (quickacks > icsk->icsk_ack.quick) 180 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS); 181 } 182 183 void tcp_enter_quickack_mode(struct sock *sk) 184 { 185 struct inet_connection_sock *icsk = inet_csk(sk); 186 tcp_incr_quickack(sk); 187 icsk->icsk_ack.pingpong = 0; 188 icsk->icsk_ack.ato = TCP_ATO_MIN; 189 } 190 191 /* Send ACKs quickly, if "quick" count is not exhausted 192 * and the session is not interactive. 193 */ 194 195 static inline int tcp_in_quickack_mode(const struct sock *sk) 196 { 197 const struct inet_connection_sock *icsk = inet_csk(sk); 198 return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong; 199 } 200 201 /* Buffer size and advertised window tuning. 202 * 203 * 1. Tuning sk->sk_sndbuf, when connection enters established state. 204 */ 205 206 static void tcp_fixup_sndbuf(struct sock *sk) 207 { 208 int sndmem = tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER + 16 + 209 sizeof(struct sk_buff); 210 211 if (sk->sk_sndbuf < 3 * sndmem) 212 sk->sk_sndbuf = min(3 * sndmem, sysctl_tcp_wmem[2]); 213 } 214 215 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh) 216 * 217 * All tcp_full_space() is split to two parts: "network" buffer, allocated 218 * forward and advertised in receiver window (tp->rcv_wnd) and 219 * "application buffer", required to isolate scheduling/application 220 * latencies from network. 221 * window_clamp is maximal advertised window. It can be less than 222 * tcp_full_space(), in this case tcp_full_space() - window_clamp 223 * is reserved for "application" buffer. The less window_clamp is 224 * the smoother our behaviour from viewpoint of network, but the lower 225 * throughput and the higher sensitivity of the connection to losses. 8) 226 * 227 * rcv_ssthresh is more strict window_clamp used at "slow start" 228 * phase to predict further behaviour of this connection. 229 * It is used for two goals: 230 * - to enforce header prediction at sender, even when application 231 * requires some significant "application buffer". It is check #1. 232 * - to prevent pruning of receive queue because of misprediction 233 * of receiver window. Check #2. 234 * 235 * The scheme does not work when sender sends good segments opening 236 * window and then starts to feed us spaghetti. But it should work 237 * in common situations. Otherwise, we have to rely on queue collapsing. 238 */ 239 240 /* Slow part of check#2. */ 241 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb) 242 { 243 struct tcp_sock *tp = tcp_sk(sk); 244 /* Optimize this! */ 245 int truesize = tcp_win_from_space(skb->truesize)/2; 246 int window = tcp_win_from_space(sysctl_tcp_rmem[2])/2; 247 248 while (tp->rcv_ssthresh <= window) { 249 if (truesize <= skb->len) 250 return 2 * inet_csk(sk)->icsk_ack.rcv_mss; 251 252 truesize >>= 1; 253 window >>= 1; 254 } 255 return 0; 256 } 257 258 static void tcp_grow_window(struct sock *sk, 259 struct sk_buff *skb) 260 { 261 struct tcp_sock *tp = tcp_sk(sk); 262 263 /* Check #1 */ 264 if (tp->rcv_ssthresh < tp->window_clamp && 265 (int)tp->rcv_ssthresh < tcp_space(sk) && 266 !tcp_memory_pressure) { 267 int incr; 268 269 /* Check #2. Increase window, if skb with such overhead 270 * will fit to rcvbuf in future. 271 */ 272 if (tcp_win_from_space(skb->truesize) <= skb->len) 273 incr = 2*tp->advmss; 274 else 275 incr = __tcp_grow_window(sk, skb); 276 277 if (incr) { 278 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr, tp->window_clamp); 279 inet_csk(sk)->icsk_ack.quick |= 1; 280 } 281 } 282 } 283 284 /* 3. Tuning rcvbuf, when connection enters established state. */ 285 286 static void tcp_fixup_rcvbuf(struct sock *sk) 287 { 288 struct tcp_sock *tp = tcp_sk(sk); 289 int rcvmem = tp->advmss + MAX_TCP_HEADER + 16 + sizeof(struct sk_buff); 290 291 /* Try to select rcvbuf so that 4 mss-sized segments 292 * will fit to window and corresponding skbs will fit to our rcvbuf. 293 * (was 3; 4 is minimum to allow fast retransmit to work.) 294 */ 295 while (tcp_win_from_space(rcvmem) < tp->advmss) 296 rcvmem += 128; 297 if (sk->sk_rcvbuf < 4 * rcvmem) 298 sk->sk_rcvbuf = min(4 * rcvmem, sysctl_tcp_rmem[2]); 299 } 300 301 /* 4. Try to fixup all. It is made immediately after connection enters 302 * established state. 303 */ 304 static void tcp_init_buffer_space(struct sock *sk) 305 { 306 struct tcp_sock *tp = tcp_sk(sk); 307 int maxwin; 308 309 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) 310 tcp_fixup_rcvbuf(sk); 311 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK)) 312 tcp_fixup_sndbuf(sk); 313 314 tp->rcvq_space.space = tp->rcv_wnd; 315 316 maxwin = tcp_full_space(sk); 317 318 if (tp->window_clamp >= maxwin) { 319 tp->window_clamp = maxwin; 320 321 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss) 322 tp->window_clamp = max(maxwin - 323 (maxwin >> sysctl_tcp_app_win), 324 4 * tp->advmss); 325 } 326 327 /* Force reservation of one segment. */ 328 if (sysctl_tcp_app_win && 329 tp->window_clamp > 2 * tp->advmss && 330 tp->window_clamp + tp->advmss > maxwin) 331 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss); 332 333 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp); 334 tp->snd_cwnd_stamp = tcp_time_stamp; 335 } 336 337 /* 5. Recalculate window clamp after socket hit its memory bounds. */ 338 static void tcp_clamp_window(struct sock *sk) 339 { 340 struct tcp_sock *tp = tcp_sk(sk); 341 struct inet_connection_sock *icsk = inet_csk(sk); 342 343 icsk->icsk_ack.quick = 0; 344 345 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] && 346 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) && 347 !tcp_memory_pressure && 348 atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) { 349 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc), 350 sysctl_tcp_rmem[2]); 351 } 352 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf) 353 tp->rcv_ssthresh = min(tp->window_clamp, 2U*tp->advmss); 354 } 355 356 357 /* Initialize RCV_MSS value. 358 * RCV_MSS is an our guess about MSS used by the peer. 359 * We haven't any direct information about the MSS. 360 * It's better to underestimate the RCV_MSS rather than overestimate. 361 * Overestimations make us ACKing less frequently than needed. 362 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss(). 363 */ 364 void tcp_initialize_rcv_mss(struct sock *sk) 365 { 366 struct tcp_sock *tp = tcp_sk(sk); 367 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache); 368 369 hint = min(hint, tp->rcv_wnd/2); 370 hint = min(hint, TCP_MIN_RCVMSS); 371 hint = max(hint, TCP_MIN_MSS); 372 373 inet_csk(sk)->icsk_ack.rcv_mss = hint; 374 } 375 376 /* Receiver "autotuning" code. 377 * 378 * The algorithm for RTT estimation w/o timestamps is based on 379 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL. 380 * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps> 381 * 382 * More detail on this code can be found at 383 * <http://www.psc.edu/~jheffner/senior_thesis.ps>, 384 * though this reference is out of date. A new paper 385 * is pending. 386 */ 387 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep) 388 { 389 u32 new_sample = tp->rcv_rtt_est.rtt; 390 long m = sample; 391 392 if (m == 0) 393 m = 1; 394 395 if (new_sample != 0) { 396 /* If we sample in larger samples in the non-timestamp 397 * case, we could grossly overestimate the RTT especially 398 * with chatty applications or bulk transfer apps which 399 * are stalled on filesystem I/O. 400 * 401 * Also, since we are only going for a minimum in the 402 * non-timestamp case, we do not smooth things out 403 * else with timestamps disabled convergence takes too 404 * long. 405 */ 406 if (!win_dep) { 407 m -= (new_sample >> 3); 408 new_sample += m; 409 } else if (m < new_sample) 410 new_sample = m << 3; 411 } else { 412 /* No previous measure. */ 413 new_sample = m << 3; 414 } 415 416 if (tp->rcv_rtt_est.rtt != new_sample) 417 tp->rcv_rtt_est.rtt = new_sample; 418 } 419 420 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp) 421 { 422 if (tp->rcv_rtt_est.time == 0) 423 goto new_measure; 424 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq)) 425 return; 426 tcp_rcv_rtt_update(tp, 427 jiffies - tp->rcv_rtt_est.time, 428 1); 429 430 new_measure: 431 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd; 432 tp->rcv_rtt_est.time = tcp_time_stamp; 433 } 434 435 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, const struct sk_buff *skb) 436 { 437 struct tcp_sock *tp = tcp_sk(sk); 438 if (tp->rx_opt.rcv_tsecr && 439 (TCP_SKB_CB(skb)->end_seq - 440 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss)) 441 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0); 442 } 443 444 /* 445 * This function should be called every time data is copied to user space. 446 * It calculates the appropriate TCP receive buffer space. 447 */ 448 void tcp_rcv_space_adjust(struct sock *sk) 449 { 450 struct tcp_sock *tp = tcp_sk(sk); 451 int time; 452 int space; 453 454 if (tp->rcvq_space.time == 0) 455 goto new_measure; 456 457 time = tcp_time_stamp - tp->rcvq_space.time; 458 if (time < (tp->rcv_rtt_est.rtt >> 3) || 459 tp->rcv_rtt_est.rtt == 0) 460 return; 461 462 space = 2 * (tp->copied_seq - tp->rcvq_space.seq); 463 464 space = max(tp->rcvq_space.space, space); 465 466 if (tp->rcvq_space.space != space) { 467 int rcvmem; 468 469 tp->rcvq_space.space = space; 470 471 if (sysctl_tcp_moderate_rcvbuf && 472 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) { 473 int new_clamp = space; 474 475 /* Receive space grows, normalize in order to 476 * take into account packet headers and sk_buff 477 * structure overhead. 478 */ 479 space /= tp->advmss; 480 if (!space) 481 space = 1; 482 rcvmem = (tp->advmss + MAX_TCP_HEADER + 483 16 + sizeof(struct sk_buff)); 484 while (tcp_win_from_space(rcvmem) < tp->advmss) 485 rcvmem += 128; 486 space *= rcvmem; 487 space = min(space, sysctl_tcp_rmem[2]); 488 if (space > sk->sk_rcvbuf) { 489 sk->sk_rcvbuf = space; 490 491 /* Make the window clamp follow along. */ 492 tp->window_clamp = new_clamp; 493 } 494 } 495 } 496 497 new_measure: 498 tp->rcvq_space.seq = tp->copied_seq; 499 tp->rcvq_space.time = tcp_time_stamp; 500 } 501 502 /* There is something which you must keep in mind when you analyze the 503 * behavior of the tp->ato delayed ack timeout interval. When a 504 * connection starts up, we want to ack as quickly as possible. The 505 * problem is that "good" TCP's do slow start at the beginning of data 506 * transmission. The means that until we send the first few ACK's the 507 * sender will sit on his end and only queue most of his data, because 508 * he can only send snd_cwnd unacked packets at any given time. For 509 * each ACK we send, he increments snd_cwnd and transmits more of his 510 * queue. -DaveM 511 */ 512 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb) 513 { 514 struct tcp_sock *tp = tcp_sk(sk); 515 struct inet_connection_sock *icsk = inet_csk(sk); 516 u32 now; 517 518 inet_csk_schedule_ack(sk); 519 520 tcp_measure_rcv_mss(sk, skb); 521 522 tcp_rcv_rtt_measure(tp); 523 524 now = tcp_time_stamp; 525 526 if (!icsk->icsk_ack.ato) { 527 /* The _first_ data packet received, initialize 528 * delayed ACK engine. 529 */ 530 tcp_incr_quickack(sk); 531 icsk->icsk_ack.ato = TCP_ATO_MIN; 532 } else { 533 int m = now - icsk->icsk_ack.lrcvtime; 534 535 if (m <= TCP_ATO_MIN/2) { 536 /* The fastest case is the first. */ 537 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2; 538 } else if (m < icsk->icsk_ack.ato) { 539 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m; 540 if (icsk->icsk_ack.ato > icsk->icsk_rto) 541 icsk->icsk_ack.ato = icsk->icsk_rto; 542 } else if (m > icsk->icsk_rto) { 543 /* Too long gap. Apparently sender failed to 544 * restart window, so that we send ACKs quickly. 545 */ 546 tcp_incr_quickack(sk); 547 sk_stream_mem_reclaim(sk); 548 } 549 } 550 icsk->icsk_ack.lrcvtime = now; 551 552 TCP_ECN_check_ce(tp, skb); 553 554 if (skb->len >= 128) 555 tcp_grow_window(sk, skb); 556 } 557 558 static u32 tcp_rto_min(struct sock *sk) 559 { 560 struct dst_entry *dst = __sk_dst_get(sk); 561 u32 rto_min = TCP_RTO_MIN; 562 563 if (dst && dst_metric_locked(dst, RTAX_RTO_MIN)) 564 rto_min = dst->metrics[RTAX_RTO_MIN-1]; 565 return rto_min; 566 } 567 568 /* Called to compute a smoothed rtt estimate. The data fed to this 569 * routine either comes from timestamps, or from segments that were 570 * known _not_ to have been retransmitted [see Karn/Partridge 571 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88 572 * piece by Van Jacobson. 573 * NOTE: the next three routines used to be one big routine. 574 * To save cycles in the RFC 1323 implementation it was better to break 575 * it up into three procedures. -- erics 576 */ 577 static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt) 578 { 579 struct tcp_sock *tp = tcp_sk(sk); 580 long m = mrtt; /* RTT */ 581 582 /* The following amusing code comes from Jacobson's 583 * article in SIGCOMM '88. Note that rtt and mdev 584 * are scaled versions of rtt and mean deviation. 585 * This is designed to be as fast as possible 586 * m stands for "measurement". 587 * 588 * On a 1990 paper the rto value is changed to: 589 * RTO = rtt + 4 * mdev 590 * 591 * Funny. This algorithm seems to be very broken. 592 * These formulae increase RTO, when it should be decreased, increase 593 * too slowly, when it should be increased quickly, decrease too quickly 594 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely 595 * does not matter how to _calculate_ it. Seems, it was trap 596 * that VJ failed to avoid. 8) 597 */ 598 if (m == 0) 599 m = 1; 600 if (tp->srtt != 0) { 601 m -= (tp->srtt >> 3); /* m is now error in rtt est */ 602 tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */ 603 if (m < 0) { 604 m = -m; /* m is now abs(error) */ 605 m -= (tp->mdev >> 2); /* similar update on mdev */ 606 /* This is similar to one of Eifel findings. 607 * Eifel blocks mdev updates when rtt decreases. 608 * This solution is a bit different: we use finer gain 609 * for mdev in this case (alpha*beta). 610 * Like Eifel it also prevents growth of rto, 611 * but also it limits too fast rto decreases, 612 * happening in pure Eifel. 613 */ 614 if (m > 0) 615 m >>= 3; 616 } else { 617 m -= (tp->mdev >> 2); /* similar update on mdev */ 618 } 619 tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */ 620 if (tp->mdev > tp->mdev_max) { 621 tp->mdev_max = tp->mdev; 622 if (tp->mdev_max > tp->rttvar) 623 tp->rttvar = tp->mdev_max; 624 } 625 if (after(tp->snd_una, tp->rtt_seq)) { 626 if (tp->mdev_max < tp->rttvar) 627 tp->rttvar -= (tp->rttvar-tp->mdev_max)>>2; 628 tp->rtt_seq = tp->snd_nxt; 629 tp->mdev_max = tcp_rto_min(sk); 630 } 631 } else { 632 /* no previous measure. */ 633 tp->srtt = m<<3; /* take the measured time to be rtt */ 634 tp->mdev = m<<1; /* make sure rto = 3*rtt */ 635 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk)); 636 tp->rtt_seq = tp->snd_nxt; 637 } 638 } 639 640 /* Calculate rto without backoff. This is the second half of Van Jacobson's 641 * routine referred to above. 642 */ 643 static inline void tcp_set_rto(struct sock *sk) 644 { 645 const struct tcp_sock *tp = tcp_sk(sk); 646 /* Old crap is replaced with new one. 8) 647 * 648 * More seriously: 649 * 1. If rtt variance happened to be less 50msec, it is hallucination. 650 * It cannot be less due to utterly erratic ACK generation made 651 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_ 652 * to do with delayed acks, because at cwnd>2 true delack timeout 653 * is invisible. Actually, Linux-2.4 also generates erratic 654 * ACKs in some circumstances. 655 */ 656 inet_csk(sk)->icsk_rto = (tp->srtt >> 3) + tp->rttvar; 657 658 /* 2. Fixups made earlier cannot be right. 659 * If we do not estimate RTO correctly without them, 660 * all the algo is pure shit and should be replaced 661 * with correct one. It is exactly, which we pretend to do. 662 */ 663 } 664 665 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo 666 * guarantees that rto is higher. 667 */ 668 static inline void tcp_bound_rto(struct sock *sk) 669 { 670 if (inet_csk(sk)->icsk_rto > TCP_RTO_MAX) 671 inet_csk(sk)->icsk_rto = TCP_RTO_MAX; 672 } 673 674 /* Save metrics learned by this TCP session. 675 This function is called only, when TCP finishes successfully 676 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE. 677 */ 678 void tcp_update_metrics(struct sock *sk) 679 { 680 struct tcp_sock *tp = tcp_sk(sk); 681 struct dst_entry *dst = __sk_dst_get(sk); 682 683 if (sysctl_tcp_nometrics_save) 684 return; 685 686 dst_confirm(dst); 687 688 if (dst && (dst->flags&DST_HOST)) { 689 const struct inet_connection_sock *icsk = inet_csk(sk); 690 int m; 691 692 if (icsk->icsk_backoff || !tp->srtt) { 693 /* This session failed to estimate rtt. Why? 694 * Probably, no packets returned in time. 695 * Reset our results. 696 */ 697 if (!(dst_metric_locked(dst, RTAX_RTT))) 698 dst->metrics[RTAX_RTT-1] = 0; 699 return; 700 } 701 702 m = dst_metric(dst, RTAX_RTT) - tp->srtt; 703 704 /* If newly calculated rtt larger than stored one, 705 * store new one. Otherwise, use EWMA. Remember, 706 * rtt overestimation is always better than underestimation. 707 */ 708 if (!(dst_metric_locked(dst, RTAX_RTT))) { 709 if (m <= 0) 710 dst->metrics[RTAX_RTT-1] = tp->srtt; 711 else 712 dst->metrics[RTAX_RTT-1] -= (m>>3); 713 } 714 715 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) { 716 if (m < 0) 717 m = -m; 718 719 /* Scale deviation to rttvar fixed point */ 720 m >>= 1; 721 if (m < tp->mdev) 722 m = tp->mdev; 723 724 if (m >= dst_metric(dst, RTAX_RTTVAR)) 725 dst->metrics[RTAX_RTTVAR-1] = m; 726 else 727 dst->metrics[RTAX_RTTVAR-1] -= 728 (dst->metrics[RTAX_RTTVAR-1] - m)>>2; 729 } 730 731 if (tp->snd_ssthresh >= 0xFFFF) { 732 /* Slow start still did not finish. */ 733 if (dst_metric(dst, RTAX_SSTHRESH) && 734 !dst_metric_locked(dst, RTAX_SSTHRESH) && 735 (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH)) 736 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1; 737 if (!dst_metric_locked(dst, RTAX_CWND) && 738 tp->snd_cwnd > dst_metric(dst, RTAX_CWND)) 739 dst->metrics[RTAX_CWND-1] = tp->snd_cwnd; 740 } else if (tp->snd_cwnd > tp->snd_ssthresh && 741 icsk->icsk_ca_state == TCP_CA_Open) { 742 /* Cong. avoidance phase, cwnd is reliable. */ 743 if (!dst_metric_locked(dst, RTAX_SSTHRESH)) 744 dst->metrics[RTAX_SSTHRESH-1] = 745 max(tp->snd_cwnd >> 1, tp->snd_ssthresh); 746 if (!dst_metric_locked(dst, RTAX_CWND)) 747 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_cwnd) >> 1; 748 } else { 749 /* Else slow start did not finish, cwnd is non-sense, 750 ssthresh may be also invalid. 751 */ 752 if (!dst_metric_locked(dst, RTAX_CWND)) 753 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_ssthresh) >> 1; 754 if (dst->metrics[RTAX_SSTHRESH-1] && 755 !dst_metric_locked(dst, RTAX_SSTHRESH) && 756 tp->snd_ssthresh > dst->metrics[RTAX_SSTHRESH-1]) 757 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh; 758 } 759 760 if (!dst_metric_locked(dst, RTAX_REORDERING)) { 761 if (dst->metrics[RTAX_REORDERING-1] < tp->reordering && 762 tp->reordering != sysctl_tcp_reordering) 763 dst->metrics[RTAX_REORDERING-1] = tp->reordering; 764 } 765 } 766 } 767 768 /* Numbers are taken from RFC3390. 769 * 770 * John Heffner states: 771 * 772 * The RFC specifies a window of no more than 4380 bytes 773 * unless 2*MSS > 4380. Reading the pseudocode in the RFC 774 * is a bit misleading because they use a clamp at 4380 bytes 775 * rather than use a multiplier in the relevant range. 776 */ 777 __u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst) 778 { 779 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0); 780 781 if (!cwnd) { 782 if (tp->mss_cache > 1460) 783 cwnd = 2; 784 else 785 cwnd = (tp->mss_cache > 1095) ? 3 : 4; 786 } 787 return min_t(__u32, cwnd, tp->snd_cwnd_clamp); 788 } 789 790 /* Set slow start threshold and cwnd not falling to slow start */ 791 void tcp_enter_cwr(struct sock *sk, const int set_ssthresh) 792 { 793 struct tcp_sock *tp = tcp_sk(sk); 794 const struct inet_connection_sock *icsk = inet_csk(sk); 795 796 tp->prior_ssthresh = 0; 797 tp->bytes_acked = 0; 798 if (icsk->icsk_ca_state < TCP_CA_CWR) { 799 tp->undo_marker = 0; 800 if (set_ssthresh) 801 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 802 tp->snd_cwnd = min(tp->snd_cwnd, 803 tcp_packets_in_flight(tp) + 1U); 804 tp->snd_cwnd_cnt = 0; 805 tp->high_seq = tp->snd_nxt; 806 tp->snd_cwnd_stamp = tcp_time_stamp; 807 TCP_ECN_queue_cwr(tp); 808 809 tcp_set_ca_state(sk, TCP_CA_CWR); 810 } 811 } 812 813 /* Initialize metrics on socket. */ 814 815 static void tcp_init_metrics(struct sock *sk) 816 { 817 struct tcp_sock *tp = tcp_sk(sk); 818 struct dst_entry *dst = __sk_dst_get(sk); 819 820 if (dst == NULL) 821 goto reset; 822 823 dst_confirm(dst); 824 825 if (dst_metric_locked(dst, RTAX_CWND)) 826 tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND); 827 if (dst_metric(dst, RTAX_SSTHRESH)) { 828 tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH); 829 if (tp->snd_ssthresh > tp->snd_cwnd_clamp) 830 tp->snd_ssthresh = tp->snd_cwnd_clamp; 831 } 832 if (dst_metric(dst, RTAX_REORDERING) && 833 tp->reordering != dst_metric(dst, RTAX_REORDERING)) { 834 tp->rx_opt.sack_ok &= ~2; 835 tp->reordering = dst_metric(dst, RTAX_REORDERING); 836 } 837 838 if (dst_metric(dst, RTAX_RTT) == 0) 839 goto reset; 840 841 if (!tp->srtt && dst_metric(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3)) 842 goto reset; 843 844 /* Initial rtt is determined from SYN,SYN-ACK. 845 * The segment is small and rtt may appear much 846 * less than real one. Use per-dst memory 847 * to make it more realistic. 848 * 849 * A bit of theory. RTT is time passed after "normal" sized packet 850 * is sent until it is ACKed. In normal circumstances sending small 851 * packets force peer to delay ACKs and calculation is correct too. 852 * The algorithm is adaptive and, provided we follow specs, it 853 * NEVER underestimate RTT. BUT! If peer tries to make some clever 854 * tricks sort of "quick acks" for time long enough to decrease RTT 855 * to low value, and then abruptly stops to do it and starts to delay 856 * ACKs, wait for troubles. 857 */ 858 if (dst_metric(dst, RTAX_RTT) > tp->srtt) { 859 tp->srtt = dst_metric(dst, RTAX_RTT); 860 tp->rtt_seq = tp->snd_nxt; 861 } 862 if (dst_metric(dst, RTAX_RTTVAR) > tp->mdev) { 863 tp->mdev = dst_metric(dst, RTAX_RTTVAR); 864 tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN); 865 } 866 tcp_set_rto(sk); 867 tcp_bound_rto(sk); 868 if (inet_csk(sk)->icsk_rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp) 869 goto reset; 870 tp->snd_cwnd = tcp_init_cwnd(tp, dst); 871 tp->snd_cwnd_stamp = tcp_time_stamp; 872 return; 873 874 reset: 875 /* Play conservative. If timestamps are not 876 * supported, TCP will fail to recalculate correct 877 * rtt, if initial rto is too small. FORGET ALL AND RESET! 878 */ 879 if (!tp->rx_opt.saw_tstamp && tp->srtt) { 880 tp->srtt = 0; 881 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT; 882 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; 883 } 884 } 885 886 static void tcp_update_reordering(struct sock *sk, const int metric, 887 const int ts) 888 { 889 struct tcp_sock *tp = tcp_sk(sk); 890 if (metric > tp->reordering) { 891 tp->reordering = min(TCP_MAX_REORDERING, metric); 892 893 /* This exciting event is worth to be remembered. 8) */ 894 if (ts) 895 NET_INC_STATS_BH(LINUX_MIB_TCPTSREORDER); 896 else if (IsReno(tp)) 897 NET_INC_STATS_BH(LINUX_MIB_TCPRENOREORDER); 898 else if (IsFack(tp)) 899 NET_INC_STATS_BH(LINUX_MIB_TCPFACKREORDER); 900 else 901 NET_INC_STATS_BH(LINUX_MIB_TCPSACKREORDER); 902 #if FASTRETRANS_DEBUG > 1 903 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n", 904 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state, 905 tp->reordering, 906 tp->fackets_out, 907 tp->sacked_out, 908 tp->undo_marker ? tp->undo_retrans : 0); 909 #endif 910 /* Disable FACK yet. */ 911 tp->rx_opt.sack_ok &= ~2; 912 } 913 } 914 915 /* This procedure tags the retransmission queue when SACKs arrive. 916 * 917 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L). 918 * Packets in queue with these bits set are counted in variables 919 * sacked_out, retrans_out and lost_out, correspondingly. 920 * 921 * Valid combinations are: 922 * Tag InFlight Description 923 * 0 1 - orig segment is in flight. 924 * S 0 - nothing flies, orig reached receiver. 925 * L 0 - nothing flies, orig lost by net. 926 * R 2 - both orig and retransmit are in flight. 927 * L|R 1 - orig is lost, retransmit is in flight. 928 * S|R 1 - orig reached receiver, retrans is still in flight. 929 * (L|S|R is logically valid, it could occur when L|R is sacked, 930 * but it is equivalent to plain S and code short-curcuits it to S. 931 * L|S is logically invalid, it would mean -1 packet in flight 8)) 932 * 933 * These 6 states form finite state machine, controlled by the following events: 934 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue()) 935 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue()) 936 * 3. Loss detection event of one of three flavors: 937 * A. Scoreboard estimator decided the packet is lost. 938 * A'. Reno "three dupacks" marks head of queue lost. 939 * A''. Its FACK modfication, head until snd.fack is lost. 940 * B. SACK arrives sacking data transmitted after never retransmitted 941 * hole was sent out. 942 * C. SACK arrives sacking SND.NXT at the moment, when the 943 * segment was retransmitted. 944 * 4. D-SACK added new rule: D-SACK changes any tag to S. 945 * 946 * It is pleasant to note, that state diagram turns out to be commutative, 947 * so that we are allowed not to be bothered by order of our actions, 948 * when multiple events arrive simultaneously. (see the function below). 949 * 950 * Reordering detection. 951 * -------------------- 952 * Reordering metric is maximal distance, which a packet can be displaced 953 * in packet stream. With SACKs we can estimate it: 954 * 955 * 1. SACK fills old hole and the corresponding segment was not 956 * ever retransmitted -> reordering. Alas, we cannot use it 957 * when segment was retransmitted. 958 * 2. The last flaw is solved with D-SACK. D-SACK arrives 959 * for retransmitted and already SACKed segment -> reordering.. 960 * Both of these heuristics are not used in Loss state, when we cannot 961 * account for retransmits accurately. 962 */ 963 static int 964 tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, u32 prior_snd_una) 965 { 966 const struct inet_connection_sock *icsk = inet_csk(sk); 967 struct tcp_sock *tp = tcp_sk(sk); 968 unsigned char *ptr = (skb_transport_header(ack_skb) + 969 TCP_SKB_CB(ack_skb)->sacked); 970 struct tcp_sack_block_wire *sp = (struct tcp_sack_block_wire *)(ptr+2); 971 struct sk_buff *cached_skb; 972 int num_sacks = (ptr[1] - TCPOLEN_SACK_BASE)>>3; 973 int reord = tp->packets_out; 974 int prior_fackets; 975 u32 lost_retrans = 0; 976 int flag = 0; 977 int found_dup_sack = 0; 978 int cached_fack_count; 979 int i; 980 int first_sack_index; 981 982 if (!tp->sacked_out) 983 tp->fackets_out = 0; 984 prior_fackets = tp->fackets_out; 985 986 /* Check for D-SACK. */ 987 if (before(ntohl(sp[0].start_seq), TCP_SKB_CB(ack_skb)->ack_seq)) { 988 flag |= FLAG_DSACKING_ACK; 989 found_dup_sack = 1; 990 tp->rx_opt.sack_ok |= 4; 991 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV); 992 } else if (num_sacks > 1 && 993 !after(ntohl(sp[0].end_seq), ntohl(sp[1].end_seq)) && 994 !before(ntohl(sp[0].start_seq), ntohl(sp[1].start_seq))) { 995 flag |= FLAG_DSACKING_ACK; 996 found_dup_sack = 1; 997 tp->rx_opt.sack_ok |= 4; 998 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV); 999 } 1000 1001 /* D-SACK for already forgotten data... 1002 * Do dumb counting. */ 1003 if (found_dup_sack && 1004 !after(ntohl(sp[0].end_seq), prior_snd_una) && 1005 after(ntohl(sp[0].end_seq), tp->undo_marker)) 1006 tp->undo_retrans--; 1007 1008 /* Eliminate too old ACKs, but take into 1009 * account more or less fresh ones, they can 1010 * contain valid SACK info. 1011 */ 1012 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window)) 1013 return 0; 1014 1015 /* SACK fastpath: 1016 * if the only SACK change is the increase of the end_seq of 1017 * the first block then only apply that SACK block 1018 * and use retrans queue hinting otherwise slowpath */ 1019 flag = 1; 1020 for (i = 0; i < num_sacks; i++) { 1021 __be32 start_seq = sp[i].start_seq; 1022 __be32 end_seq = sp[i].end_seq; 1023 1024 if (i == 0) { 1025 if (tp->recv_sack_cache[i].start_seq != start_seq) 1026 flag = 0; 1027 } else { 1028 if ((tp->recv_sack_cache[i].start_seq != start_seq) || 1029 (tp->recv_sack_cache[i].end_seq != end_seq)) 1030 flag = 0; 1031 } 1032 tp->recv_sack_cache[i].start_seq = start_seq; 1033 tp->recv_sack_cache[i].end_seq = end_seq; 1034 } 1035 /* Clear the rest of the cache sack blocks so they won't match mistakenly. */ 1036 for (; i < ARRAY_SIZE(tp->recv_sack_cache); i++) { 1037 tp->recv_sack_cache[i].start_seq = 0; 1038 tp->recv_sack_cache[i].end_seq = 0; 1039 } 1040 1041 first_sack_index = 0; 1042 if (flag) 1043 num_sacks = 1; 1044 else { 1045 int j; 1046 tp->fastpath_skb_hint = NULL; 1047 1048 /* order SACK blocks to allow in order walk of the retrans queue */ 1049 for (i = num_sacks-1; i > 0; i--) { 1050 for (j = 0; j < i; j++){ 1051 if (after(ntohl(sp[j].start_seq), 1052 ntohl(sp[j+1].start_seq))){ 1053 struct tcp_sack_block_wire tmp; 1054 1055 tmp = sp[j]; 1056 sp[j] = sp[j+1]; 1057 sp[j+1] = tmp; 1058 1059 /* Track where the first SACK block goes to */ 1060 if (j == first_sack_index) 1061 first_sack_index = j+1; 1062 } 1063 1064 } 1065 } 1066 } 1067 1068 /* clear flag as used for different purpose in following code */ 1069 flag = 0; 1070 1071 /* Use SACK fastpath hint if valid */ 1072 cached_skb = tp->fastpath_skb_hint; 1073 cached_fack_count = tp->fastpath_cnt_hint; 1074 if (!cached_skb) { 1075 cached_skb = tcp_write_queue_head(sk); 1076 cached_fack_count = 0; 1077 } 1078 1079 for (i=0; i<num_sacks; i++, sp++) { 1080 struct sk_buff *skb; 1081 __u32 start_seq = ntohl(sp->start_seq); 1082 __u32 end_seq = ntohl(sp->end_seq); 1083 int fack_count; 1084 int dup_sack = (found_dup_sack && (i == first_sack_index)); 1085 1086 skb = cached_skb; 1087 fack_count = cached_fack_count; 1088 1089 /* Event "B" in the comment above. */ 1090 if (after(end_seq, tp->high_seq)) 1091 flag |= FLAG_DATA_LOST; 1092 1093 tcp_for_write_queue_from(skb, sk) { 1094 int in_sack, pcount; 1095 u8 sacked; 1096 1097 if (skb == tcp_send_head(sk)) 1098 break; 1099 1100 cached_skb = skb; 1101 cached_fack_count = fack_count; 1102 if (i == first_sack_index) { 1103 tp->fastpath_skb_hint = skb; 1104 tp->fastpath_cnt_hint = fack_count; 1105 } 1106 1107 /* The retransmission queue is always in order, so 1108 * we can short-circuit the walk early. 1109 */ 1110 if (!before(TCP_SKB_CB(skb)->seq, end_seq)) 1111 break; 1112 1113 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && 1114 !before(end_seq, TCP_SKB_CB(skb)->end_seq); 1115 1116 pcount = tcp_skb_pcount(skb); 1117 1118 if (pcount > 1 && !in_sack && 1119 after(TCP_SKB_CB(skb)->end_seq, start_seq)) { 1120 unsigned int pkt_len; 1121 1122 in_sack = !after(start_seq, 1123 TCP_SKB_CB(skb)->seq); 1124 1125 if (!in_sack) 1126 pkt_len = (start_seq - 1127 TCP_SKB_CB(skb)->seq); 1128 else 1129 pkt_len = (end_seq - 1130 TCP_SKB_CB(skb)->seq); 1131 if (tcp_fragment(sk, skb, pkt_len, skb_shinfo(skb)->gso_size)) 1132 break; 1133 pcount = tcp_skb_pcount(skb); 1134 } 1135 1136 fack_count += pcount; 1137 1138 sacked = TCP_SKB_CB(skb)->sacked; 1139 1140 /* Account D-SACK for retransmitted packet. */ 1141 if ((dup_sack && in_sack) && 1142 (sacked & TCPCB_RETRANS) && 1143 after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker)) 1144 tp->undo_retrans--; 1145 1146 /* The frame is ACKed. */ 1147 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) { 1148 if (sacked&TCPCB_RETRANS) { 1149 if ((dup_sack && in_sack) && 1150 (sacked&TCPCB_SACKED_ACKED)) 1151 reord = min(fack_count, reord); 1152 } else { 1153 /* If it was in a hole, we detected reordering. */ 1154 if (fack_count < prior_fackets && 1155 !(sacked&TCPCB_SACKED_ACKED)) 1156 reord = min(fack_count, reord); 1157 } 1158 1159 /* Nothing to do; acked frame is about to be dropped. */ 1160 continue; 1161 } 1162 1163 if ((sacked&TCPCB_SACKED_RETRANS) && 1164 after(end_seq, TCP_SKB_CB(skb)->ack_seq) && 1165 (!lost_retrans || after(end_seq, lost_retrans))) 1166 lost_retrans = end_seq; 1167 1168 if (!in_sack) 1169 continue; 1170 1171 if (!(sacked&TCPCB_SACKED_ACKED)) { 1172 if (sacked & TCPCB_SACKED_RETRANS) { 1173 /* If the segment is not tagged as lost, 1174 * we do not clear RETRANS, believing 1175 * that retransmission is still in flight. 1176 */ 1177 if (sacked & TCPCB_LOST) { 1178 TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS); 1179 tp->lost_out -= tcp_skb_pcount(skb); 1180 tp->retrans_out -= tcp_skb_pcount(skb); 1181 1182 /* clear lost hint */ 1183 tp->retransmit_skb_hint = NULL; 1184 } 1185 } else { 1186 /* New sack for not retransmitted frame, 1187 * which was in hole. It is reordering. 1188 */ 1189 if (!(sacked & TCPCB_RETRANS) && 1190 fack_count < prior_fackets) 1191 reord = min(fack_count, reord); 1192 1193 if (sacked & TCPCB_LOST) { 1194 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; 1195 tp->lost_out -= tcp_skb_pcount(skb); 1196 1197 /* clear lost hint */ 1198 tp->retransmit_skb_hint = NULL; 1199 } 1200 /* SACK enhanced F-RTO detection. 1201 * Set flag if and only if non-rexmitted 1202 * segments below frto_highmark are 1203 * SACKed (RFC4138; Appendix B). 1204 * Clearing correct due to in-order walk 1205 */ 1206 if (after(end_seq, tp->frto_highmark)) { 1207 flag &= ~FLAG_ONLY_ORIG_SACKED; 1208 } else { 1209 if (!(sacked & TCPCB_RETRANS)) 1210 flag |= FLAG_ONLY_ORIG_SACKED; 1211 } 1212 } 1213 1214 TCP_SKB_CB(skb)->sacked |= TCPCB_SACKED_ACKED; 1215 flag |= FLAG_DATA_SACKED; 1216 tp->sacked_out += tcp_skb_pcount(skb); 1217 1218 if (fack_count > tp->fackets_out) 1219 tp->fackets_out = fack_count; 1220 } else { 1221 if (dup_sack && (sacked&TCPCB_RETRANS)) 1222 reord = min(fack_count, reord); 1223 } 1224 1225 /* D-SACK. We can detect redundant retransmission 1226 * in S|R and plain R frames and clear it. 1227 * undo_retrans is decreased above, L|R frames 1228 * are accounted above as well. 1229 */ 1230 if (dup_sack && 1231 (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS)) { 1232 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 1233 tp->retrans_out -= tcp_skb_pcount(skb); 1234 tp->retransmit_skb_hint = NULL; 1235 } 1236 } 1237 } 1238 1239 /* Check for lost retransmit. This superb idea is 1240 * borrowed from "ratehalving". Event "C". 1241 * Later note: FACK people cheated me again 8), 1242 * we have to account for reordering! Ugly, 1243 * but should help. 1244 */ 1245 if (lost_retrans && icsk->icsk_ca_state == TCP_CA_Recovery) { 1246 struct sk_buff *skb; 1247 1248 tcp_for_write_queue(skb, sk) { 1249 if (skb == tcp_send_head(sk)) 1250 break; 1251 if (after(TCP_SKB_CB(skb)->seq, lost_retrans)) 1252 break; 1253 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 1254 continue; 1255 if ((TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) && 1256 after(lost_retrans, TCP_SKB_CB(skb)->ack_seq) && 1257 (IsFack(tp) || 1258 !before(lost_retrans, 1259 TCP_SKB_CB(skb)->ack_seq + tp->reordering * 1260 tp->mss_cache))) { 1261 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 1262 tp->retrans_out -= tcp_skb_pcount(skb); 1263 1264 /* clear lost hint */ 1265 tp->retransmit_skb_hint = NULL; 1266 1267 if (!(TCP_SKB_CB(skb)->sacked&(TCPCB_LOST|TCPCB_SACKED_ACKED))) { 1268 tp->lost_out += tcp_skb_pcount(skb); 1269 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 1270 flag |= FLAG_DATA_SACKED; 1271 NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT); 1272 } 1273 } 1274 } 1275 } 1276 1277 tp->left_out = tp->sacked_out + tp->lost_out; 1278 1279 if ((reord < tp->fackets_out) && icsk->icsk_ca_state != TCP_CA_Loss && 1280 (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark))) 1281 tcp_update_reordering(sk, ((tp->fackets_out + 1) - reord), 0); 1282 1283 #if FASTRETRANS_DEBUG > 0 1284 BUG_TRAP((int)tp->sacked_out >= 0); 1285 BUG_TRAP((int)tp->lost_out >= 0); 1286 BUG_TRAP((int)tp->retrans_out >= 0); 1287 BUG_TRAP((int)tcp_packets_in_flight(tp) >= 0); 1288 #endif 1289 return flag; 1290 } 1291 1292 /* F-RTO can only be used if TCP has never retransmitted anything other than 1293 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here) 1294 */ 1295 int tcp_use_frto(struct sock *sk) 1296 { 1297 const struct tcp_sock *tp = tcp_sk(sk); 1298 struct sk_buff *skb; 1299 1300 if (!sysctl_tcp_frto) 1301 return 0; 1302 1303 if (IsSackFrto()) 1304 return 1; 1305 1306 /* Avoid expensive walking of rexmit queue if possible */ 1307 if (tp->retrans_out > 1) 1308 return 0; 1309 1310 skb = tcp_write_queue_head(sk); 1311 skb = tcp_write_queue_next(sk, skb); /* Skips head */ 1312 tcp_for_write_queue_from(skb, sk) { 1313 if (skb == tcp_send_head(sk)) 1314 break; 1315 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS) 1316 return 0; 1317 /* Short-circuit when first non-SACKed skb has been checked */ 1318 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) 1319 break; 1320 } 1321 return 1; 1322 } 1323 1324 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO 1325 * recovery a bit and use heuristics in tcp_process_frto() to detect if 1326 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to 1327 * keep retrans_out counting accurate (with SACK F-RTO, other than head 1328 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS 1329 * bits are handled if the Loss state is really to be entered (in 1330 * tcp_enter_frto_loss). 1331 * 1332 * Do like tcp_enter_loss() would; when RTO expires the second time it 1333 * does: 1334 * "Reduce ssthresh if it has not yet been made inside this window." 1335 */ 1336 void tcp_enter_frto(struct sock *sk) 1337 { 1338 const struct inet_connection_sock *icsk = inet_csk(sk); 1339 struct tcp_sock *tp = tcp_sk(sk); 1340 struct sk_buff *skb; 1341 1342 if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) || 1343 tp->snd_una == tp->high_seq || 1344 ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) && 1345 !icsk->icsk_retransmits)) { 1346 tp->prior_ssthresh = tcp_current_ssthresh(sk); 1347 /* Our state is too optimistic in ssthresh() call because cwnd 1348 * is not reduced until tcp_enter_frto_loss() when previous FRTO 1349 * recovery has not yet completed. Pattern would be this: RTO, 1350 * Cumulative ACK, RTO (2xRTO for the same segment does not end 1351 * up here twice). 1352 * RFC4138 should be more specific on what to do, even though 1353 * RTO is quite unlikely to occur after the first Cumulative ACK 1354 * due to back-off and complexity of triggering events ... 1355 */ 1356 if (tp->frto_counter) { 1357 u32 stored_cwnd; 1358 stored_cwnd = tp->snd_cwnd; 1359 tp->snd_cwnd = 2; 1360 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 1361 tp->snd_cwnd = stored_cwnd; 1362 } else { 1363 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 1364 } 1365 /* ... in theory, cong.control module could do "any tricks" in 1366 * ssthresh(), which means that ca_state, lost bits and lost_out 1367 * counter would have to be faked before the call occurs. We 1368 * consider that too expensive, unlikely and hacky, so modules 1369 * using these in ssthresh() must deal these incompatibility 1370 * issues if they receives CA_EVENT_FRTO and frto_counter != 0 1371 */ 1372 tcp_ca_event(sk, CA_EVENT_FRTO); 1373 } 1374 1375 tp->undo_marker = tp->snd_una; 1376 tp->undo_retrans = 0; 1377 1378 skb = tcp_write_queue_head(sk); 1379 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 1380 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 1381 tp->retrans_out -= tcp_skb_pcount(skb); 1382 } 1383 tcp_sync_left_out(tp); 1384 1385 /* Earlier loss recovery underway (see RFC4138; Appendix B). 1386 * The last condition is necessary at least in tp->frto_counter case. 1387 */ 1388 if (IsSackFrto() && (tp->frto_counter || 1389 ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) && 1390 after(tp->high_seq, tp->snd_una)) { 1391 tp->frto_highmark = tp->high_seq; 1392 } else { 1393 tp->frto_highmark = tp->snd_nxt; 1394 } 1395 tcp_set_ca_state(sk, TCP_CA_Disorder); 1396 tp->high_seq = tp->snd_nxt; 1397 tp->frto_counter = 1; 1398 } 1399 1400 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO, 1401 * which indicates that we should follow the traditional RTO recovery, 1402 * i.e. mark everything lost and do go-back-N retransmission. 1403 */ 1404 static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag) 1405 { 1406 struct tcp_sock *tp = tcp_sk(sk); 1407 struct sk_buff *skb; 1408 int cnt = 0; 1409 1410 tp->sacked_out = 0; 1411 tp->lost_out = 0; 1412 tp->fackets_out = 0; 1413 tp->retrans_out = 0; 1414 1415 tcp_for_write_queue(skb, sk) { 1416 if (skb == tcp_send_head(sk)) 1417 break; 1418 cnt += tcp_skb_pcount(skb); 1419 /* 1420 * Count the retransmission made on RTO correctly (only when 1421 * waiting for the first ACK and did not get it)... 1422 */ 1423 if ((tp->frto_counter == 1) && !(flag&FLAG_DATA_ACKED)) { 1424 /* For some reason this R-bit might get cleared? */ 1425 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 1426 tp->retrans_out += tcp_skb_pcount(skb); 1427 /* ...enter this if branch just for the first segment */ 1428 flag |= FLAG_DATA_ACKED; 1429 } else { 1430 TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS); 1431 } 1432 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) { 1433 1434 /* Do not mark those segments lost that were 1435 * forward transmitted after RTO 1436 */ 1437 if (!after(TCP_SKB_CB(skb)->end_seq, 1438 tp->frto_highmark)) { 1439 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 1440 tp->lost_out += tcp_skb_pcount(skb); 1441 } 1442 } else { 1443 tp->sacked_out += tcp_skb_pcount(skb); 1444 tp->fackets_out = cnt; 1445 } 1446 } 1447 tcp_sync_left_out(tp); 1448 1449 tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments; 1450 tp->snd_cwnd_cnt = 0; 1451 tp->snd_cwnd_stamp = tcp_time_stamp; 1452 tp->undo_marker = 0; 1453 tp->frto_counter = 0; 1454 1455 tp->reordering = min_t(unsigned int, tp->reordering, 1456 sysctl_tcp_reordering); 1457 tcp_set_ca_state(sk, TCP_CA_Loss); 1458 tp->high_seq = tp->frto_highmark; 1459 TCP_ECN_queue_cwr(tp); 1460 1461 clear_all_retrans_hints(tp); 1462 } 1463 1464 void tcp_clear_retrans(struct tcp_sock *tp) 1465 { 1466 tp->left_out = 0; 1467 tp->retrans_out = 0; 1468 1469 tp->fackets_out = 0; 1470 tp->sacked_out = 0; 1471 tp->lost_out = 0; 1472 1473 tp->undo_marker = 0; 1474 tp->undo_retrans = 0; 1475 } 1476 1477 /* Enter Loss state. If "how" is not zero, forget all SACK information 1478 * and reset tags completely, otherwise preserve SACKs. If receiver 1479 * dropped its ofo queue, we will know this due to reneging detection. 1480 */ 1481 void tcp_enter_loss(struct sock *sk, int how) 1482 { 1483 const struct inet_connection_sock *icsk = inet_csk(sk); 1484 struct tcp_sock *tp = tcp_sk(sk); 1485 struct sk_buff *skb; 1486 int cnt = 0; 1487 1488 /* Reduce ssthresh if it has not yet been made inside this window. */ 1489 if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq || 1490 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) { 1491 tp->prior_ssthresh = tcp_current_ssthresh(sk); 1492 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 1493 tcp_ca_event(sk, CA_EVENT_LOSS); 1494 } 1495 tp->snd_cwnd = 1; 1496 tp->snd_cwnd_cnt = 0; 1497 tp->snd_cwnd_stamp = tcp_time_stamp; 1498 1499 tp->bytes_acked = 0; 1500 tcp_clear_retrans(tp); 1501 1502 /* Push undo marker, if it was plain RTO and nothing 1503 * was retransmitted. */ 1504 if (!how) 1505 tp->undo_marker = tp->snd_una; 1506 1507 tcp_for_write_queue(skb, sk) { 1508 if (skb == tcp_send_head(sk)) 1509 break; 1510 cnt += tcp_skb_pcount(skb); 1511 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS) 1512 tp->undo_marker = 0; 1513 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED; 1514 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) { 1515 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED; 1516 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 1517 tp->lost_out += tcp_skb_pcount(skb); 1518 } else { 1519 tp->sacked_out += tcp_skb_pcount(skb); 1520 tp->fackets_out = cnt; 1521 } 1522 } 1523 tcp_sync_left_out(tp); 1524 1525 tp->reordering = min_t(unsigned int, tp->reordering, 1526 sysctl_tcp_reordering); 1527 tcp_set_ca_state(sk, TCP_CA_Loss); 1528 tp->high_seq = tp->snd_nxt; 1529 TCP_ECN_queue_cwr(tp); 1530 /* Abort FRTO algorithm if one is in progress */ 1531 tp->frto_counter = 0; 1532 1533 clear_all_retrans_hints(tp); 1534 } 1535 1536 static int tcp_check_sack_reneging(struct sock *sk) 1537 { 1538 struct sk_buff *skb; 1539 1540 /* If ACK arrived pointing to a remembered SACK, 1541 * it means that our remembered SACKs do not reflect 1542 * real state of receiver i.e. 1543 * receiver _host_ is heavily congested (or buggy). 1544 * Do processing similar to RTO timeout. 1545 */ 1546 if ((skb = tcp_write_queue_head(sk)) != NULL && 1547 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) { 1548 struct inet_connection_sock *icsk = inet_csk(sk); 1549 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING); 1550 1551 tcp_enter_loss(sk, 1); 1552 icsk->icsk_retransmits++; 1553 tcp_retransmit_skb(sk, tcp_write_queue_head(sk)); 1554 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 1555 icsk->icsk_rto, TCP_RTO_MAX); 1556 return 1; 1557 } 1558 return 0; 1559 } 1560 1561 static inline int tcp_fackets_out(struct tcp_sock *tp) 1562 { 1563 return IsReno(tp) ? tp->sacked_out+1 : tp->fackets_out; 1564 } 1565 1566 static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb) 1567 { 1568 return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto); 1569 } 1570 1571 static inline int tcp_head_timedout(struct sock *sk) 1572 { 1573 struct tcp_sock *tp = tcp_sk(sk); 1574 1575 return tp->packets_out && 1576 tcp_skb_timedout(sk, tcp_write_queue_head(sk)); 1577 } 1578 1579 /* Linux NewReno/SACK/FACK/ECN state machine. 1580 * -------------------------------------- 1581 * 1582 * "Open" Normal state, no dubious events, fast path. 1583 * "Disorder" In all the respects it is "Open", 1584 * but requires a bit more attention. It is entered when 1585 * we see some SACKs or dupacks. It is split of "Open" 1586 * mainly to move some processing from fast path to slow one. 1587 * "CWR" CWND was reduced due to some Congestion Notification event. 1588 * It can be ECN, ICMP source quench, local device congestion. 1589 * "Recovery" CWND was reduced, we are fast-retransmitting. 1590 * "Loss" CWND was reduced due to RTO timeout or SACK reneging. 1591 * 1592 * tcp_fastretrans_alert() is entered: 1593 * - each incoming ACK, if state is not "Open" 1594 * - when arrived ACK is unusual, namely: 1595 * * SACK 1596 * * Duplicate ACK. 1597 * * ECN ECE. 1598 * 1599 * Counting packets in flight is pretty simple. 1600 * 1601 * in_flight = packets_out - left_out + retrans_out 1602 * 1603 * packets_out is SND.NXT-SND.UNA counted in packets. 1604 * 1605 * retrans_out is number of retransmitted segments. 1606 * 1607 * left_out is number of segments left network, but not ACKed yet. 1608 * 1609 * left_out = sacked_out + lost_out 1610 * 1611 * sacked_out: Packets, which arrived to receiver out of order 1612 * and hence not ACKed. With SACKs this number is simply 1613 * amount of SACKed data. Even without SACKs 1614 * it is easy to give pretty reliable estimate of this number, 1615 * counting duplicate ACKs. 1616 * 1617 * lost_out: Packets lost by network. TCP has no explicit 1618 * "loss notification" feedback from network (for now). 1619 * It means that this number can be only _guessed_. 1620 * Actually, it is the heuristics to predict lossage that 1621 * distinguishes different algorithms. 1622 * 1623 * F.e. after RTO, when all the queue is considered as lost, 1624 * lost_out = packets_out and in_flight = retrans_out. 1625 * 1626 * Essentially, we have now two algorithms counting 1627 * lost packets. 1628 * 1629 * FACK: It is the simplest heuristics. As soon as we decided 1630 * that something is lost, we decide that _all_ not SACKed 1631 * packets until the most forward SACK are lost. I.e. 1632 * lost_out = fackets_out - sacked_out and left_out = fackets_out. 1633 * It is absolutely correct estimate, if network does not reorder 1634 * packets. And it loses any connection to reality when reordering 1635 * takes place. We use FACK by default until reordering 1636 * is suspected on the path to this destination. 1637 * 1638 * NewReno: when Recovery is entered, we assume that one segment 1639 * is lost (classic Reno). While we are in Recovery and 1640 * a partial ACK arrives, we assume that one more packet 1641 * is lost (NewReno). This heuristics are the same in NewReno 1642 * and SACK. 1643 * 1644 * Imagine, that's all! Forget about all this shamanism about CWND inflation 1645 * deflation etc. CWND is real congestion window, never inflated, changes 1646 * only according to classic VJ rules. 1647 * 1648 * Really tricky (and requiring careful tuning) part of algorithm 1649 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue(). 1650 * The first determines the moment _when_ we should reduce CWND and, 1651 * hence, slow down forward transmission. In fact, it determines the moment 1652 * when we decide that hole is caused by loss, rather than by a reorder. 1653 * 1654 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill 1655 * holes, caused by lost packets. 1656 * 1657 * And the most logically complicated part of algorithm is undo 1658 * heuristics. We detect false retransmits due to both too early 1659 * fast retransmit (reordering) and underestimated RTO, analyzing 1660 * timestamps and D-SACKs. When we detect that some segments were 1661 * retransmitted by mistake and CWND reduction was wrong, we undo 1662 * window reduction and abort recovery phase. This logic is hidden 1663 * inside several functions named tcp_try_undo_<something>. 1664 */ 1665 1666 /* This function decides, when we should leave Disordered state 1667 * and enter Recovery phase, reducing congestion window. 1668 * 1669 * Main question: may we further continue forward transmission 1670 * with the same cwnd? 1671 */ 1672 static int tcp_time_to_recover(struct sock *sk) 1673 { 1674 struct tcp_sock *tp = tcp_sk(sk); 1675 __u32 packets_out; 1676 1677 /* Do not perform any recovery during FRTO algorithm */ 1678 if (tp->frto_counter) 1679 return 0; 1680 1681 /* Trick#1: The loss is proven. */ 1682 if (tp->lost_out) 1683 return 1; 1684 1685 /* Not-A-Trick#2 : Classic rule... */ 1686 if (tcp_fackets_out(tp) > tp->reordering) 1687 return 1; 1688 1689 /* Trick#3 : when we use RFC2988 timer restart, fast 1690 * retransmit can be triggered by timeout of queue head. 1691 */ 1692 if (tcp_head_timedout(sk)) 1693 return 1; 1694 1695 /* Trick#4: It is still not OK... But will it be useful to delay 1696 * recovery more? 1697 */ 1698 packets_out = tp->packets_out; 1699 if (packets_out <= tp->reordering && 1700 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) && 1701 !tcp_may_send_now(sk)) { 1702 /* We have nothing to send. This connection is limited 1703 * either by receiver window or by application. 1704 */ 1705 return 1; 1706 } 1707 1708 return 0; 1709 } 1710 1711 /* If we receive more dupacks than we expected counting segments 1712 * in assumption of absent reordering, interpret this as reordering. 1713 * The only another reason could be bug in receiver TCP. 1714 */ 1715 static void tcp_check_reno_reordering(struct sock *sk, const int addend) 1716 { 1717 struct tcp_sock *tp = tcp_sk(sk); 1718 u32 holes; 1719 1720 holes = max(tp->lost_out, 1U); 1721 holes = min(holes, tp->packets_out); 1722 1723 if ((tp->sacked_out + holes) > tp->packets_out) { 1724 tp->sacked_out = tp->packets_out - holes; 1725 tcp_update_reordering(sk, tp->packets_out + addend, 0); 1726 } 1727 } 1728 1729 /* Emulate SACKs for SACKless connection: account for a new dupack. */ 1730 1731 static void tcp_add_reno_sack(struct sock *sk) 1732 { 1733 struct tcp_sock *tp = tcp_sk(sk); 1734 tp->sacked_out++; 1735 tcp_check_reno_reordering(sk, 0); 1736 tcp_sync_left_out(tp); 1737 } 1738 1739 /* Account for ACK, ACKing some data in Reno Recovery phase. */ 1740 1741 static void tcp_remove_reno_sacks(struct sock *sk, int acked) 1742 { 1743 struct tcp_sock *tp = tcp_sk(sk); 1744 1745 if (acked > 0) { 1746 /* One ACK acked hole. The rest eat duplicate ACKs. */ 1747 if (acked-1 >= tp->sacked_out) 1748 tp->sacked_out = 0; 1749 else 1750 tp->sacked_out -= acked-1; 1751 } 1752 tcp_check_reno_reordering(sk, acked); 1753 tcp_sync_left_out(tp); 1754 } 1755 1756 static inline void tcp_reset_reno_sack(struct tcp_sock *tp) 1757 { 1758 tp->sacked_out = 0; 1759 tp->left_out = tp->lost_out; 1760 } 1761 1762 /* Mark head of queue up as lost. */ 1763 static void tcp_mark_head_lost(struct sock *sk, 1764 int packets, u32 high_seq) 1765 { 1766 struct tcp_sock *tp = tcp_sk(sk); 1767 struct sk_buff *skb; 1768 int cnt; 1769 1770 BUG_TRAP(packets <= tp->packets_out); 1771 if (tp->lost_skb_hint) { 1772 skb = tp->lost_skb_hint; 1773 cnt = tp->lost_cnt_hint; 1774 } else { 1775 skb = tcp_write_queue_head(sk); 1776 cnt = 0; 1777 } 1778 1779 tcp_for_write_queue_from(skb, sk) { 1780 if (skb == tcp_send_head(sk)) 1781 break; 1782 /* TODO: do this better */ 1783 /* this is not the most efficient way to do this... */ 1784 tp->lost_skb_hint = skb; 1785 tp->lost_cnt_hint = cnt; 1786 cnt += tcp_skb_pcount(skb); 1787 if (cnt > packets || after(TCP_SKB_CB(skb)->end_seq, high_seq)) 1788 break; 1789 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) { 1790 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 1791 tp->lost_out += tcp_skb_pcount(skb); 1792 1793 /* clear xmit_retransmit_queue hints 1794 * if this is beyond hint */ 1795 if (tp->retransmit_skb_hint != NULL && 1796 before(TCP_SKB_CB(skb)->seq, 1797 TCP_SKB_CB(tp->retransmit_skb_hint)->seq)) 1798 tp->retransmit_skb_hint = NULL; 1799 1800 } 1801 } 1802 tcp_sync_left_out(tp); 1803 } 1804 1805 /* Account newly detected lost packet(s) */ 1806 1807 static void tcp_update_scoreboard(struct sock *sk) 1808 { 1809 struct tcp_sock *tp = tcp_sk(sk); 1810 1811 if (IsFack(tp)) { 1812 int lost = tp->fackets_out - tp->reordering; 1813 if (lost <= 0) 1814 lost = 1; 1815 tcp_mark_head_lost(sk, lost, tp->high_seq); 1816 } else { 1817 tcp_mark_head_lost(sk, 1, tp->high_seq); 1818 } 1819 1820 /* New heuristics: it is possible only after we switched 1821 * to restart timer each time when something is ACKed. 1822 * Hence, we can detect timed out packets during fast 1823 * retransmit without falling to slow start. 1824 */ 1825 if (!IsReno(tp) && tcp_head_timedout(sk)) { 1826 struct sk_buff *skb; 1827 1828 skb = tp->scoreboard_skb_hint ? tp->scoreboard_skb_hint 1829 : tcp_write_queue_head(sk); 1830 1831 tcp_for_write_queue_from(skb, sk) { 1832 if (skb == tcp_send_head(sk)) 1833 break; 1834 if (!tcp_skb_timedout(sk, skb)) 1835 break; 1836 1837 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) { 1838 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 1839 tp->lost_out += tcp_skb_pcount(skb); 1840 1841 /* clear xmit_retrans hint */ 1842 if (tp->retransmit_skb_hint && 1843 before(TCP_SKB_CB(skb)->seq, 1844 TCP_SKB_CB(tp->retransmit_skb_hint)->seq)) 1845 1846 tp->retransmit_skb_hint = NULL; 1847 } 1848 } 1849 1850 tp->scoreboard_skb_hint = skb; 1851 1852 tcp_sync_left_out(tp); 1853 } 1854 } 1855 1856 /* CWND moderation, preventing bursts due to too big ACKs 1857 * in dubious situations. 1858 */ 1859 static inline void tcp_moderate_cwnd(struct tcp_sock *tp) 1860 { 1861 tp->snd_cwnd = min(tp->snd_cwnd, 1862 tcp_packets_in_flight(tp)+tcp_max_burst(tp)); 1863 tp->snd_cwnd_stamp = tcp_time_stamp; 1864 } 1865 1866 /* Lower bound on congestion window is slow start threshold 1867 * unless congestion avoidance choice decides to overide it. 1868 */ 1869 static inline u32 tcp_cwnd_min(const struct sock *sk) 1870 { 1871 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 1872 1873 return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh; 1874 } 1875 1876 /* Decrease cwnd each second ack. */ 1877 static void tcp_cwnd_down(struct sock *sk, int flag) 1878 { 1879 struct tcp_sock *tp = tcp_sk(sk); 1880 int decr = tp->snd_cwnd_cnt + 1; 1881 1882 if ((flag&(FLAG_ANY_PROGRESS|FLAG_DSACKING_ACK)) || 1883 (IsReno(tp) && !(flag&FLAG_NOT_DUP))) { 1884 tp->snd_cwnd_cnt = decr&1; 1885 decr >>= 1; 1886 1887 if (decr && tp->snd_cwnd > tcp_cwnd_min(sk)) 1888 tp->snd_cwnd -= decr; 1889 1890 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp)+1); 1891 tp->snd_cwnd_stamp = tcp_time_stamp; 1892 } 1893 } 1894 1895 /* Nothing was retransmitted or returned timestamp is less 1896 * than timestamp of the first retransmission. 1897 */ 1898 static inline int tcp_packet_delayed(struct tcp_sock *tp) 1899 { 1900 return !tp->retrans_stamp || 1901 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 1902 (__s32)(tp->rx_opt.rcv_tsecr - tp->retrans_stamp) < 0); 1903 } 1904 1905 /* Undo procedures. */ 1906 1907 #if FASTRETRANS_DEBUG > 1 1908 static void DBGUNDO(struct sock *sk, const char *msg) 1909 { 1910 struct tcp_sock *tp = tcp_sk(sk); 1911 struct inet_sock *inet = inet_sk(sk); 1912 1913 printk(KERN_DEBUG "Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n", 1914 msg, 1915 NIPQUAD(inet->daddr), ntohs(inet->dport), 1916 tp->snd_cwnd, tp->left_out, 1917 tp->snd_ssthresh, tp->prior_ssthresh, 1918 tp->packets_out); 1919 } 1920 #else 1921 #define DBGUNDO(x...) do { } while (0) 1922 #endif 1923 1924 static void tcp_undo_cwr(struct sock *sk, const int undo) 1925 { 1926 struct tcp_sock *tp = tcp_sk(sk); 1927 1928 if (tp->prior_ssthresh) { 1929 const struct inet_connection_sock *icsk = inet_csk(sk); 1930 1931 if (icsk->icsk_ca_ops->undo_cwnd) 1932 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk); 1933 else 1934 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh<<1); 1935 1936 if (undo && tp->prior_ssthresh > tp->snd_ssthresh) { 1937 tp->snd_ssthresh = tp->prior_ssthresh; 1938 TCP_ECN_withdraw_cwr(tp); 1939 } 1940 } else { 1941 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh); 1942 } 1943 tcp_moderate_cwnd(tp); 1944 tp->snd_cwnd_stamp = tcp_time_stamp; 1945 1946 /* There is something screwy going on with the retrans hints after 1947 an undo */ 1948 clear_all_retrans_hints(tp); 1949 } 1950 1951 static inline int tcp_may_undo(struct tcp_sock *tp) 1952 { 1953 return tp->undo_marker && 1954 (!tp->undo_retrans || tcp_packet_delayed(tp)); 1955 } 1956 1957 /* People celebrate: "We love our President!" */ 1958 static int tcp_try_undo_recovery(struct sock *sk) 1959 { 1960 struct tcp_sock *tp = tcp_sk(sk); 1961 1962 if (tcp_may_undo(tp)) { 1963 /* Happy end! We did not retransmit anything 1964 * or our original transmission succeeded. 1965 */ 1966 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans"); 1967 tcp_undo_cwr(sk, 1); 1968 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss) 1969 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO); 1970 else 1971 NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO); 1972 tp->undo_marker = 0; 1973 } 1974 if (tp->snd_una == tp->high_seq && IsReno(tp)) { 1975 /* Hold old state until something *above* high_seq 1976 * is ACKed. For Reno it is MUST to prevent false 1977 * fast retransmits (RFC2582). SACK TCP is safe. */ 1978 tcp_moderate_cwnd(tp); 1979 return 1; 1980 } 1981 tcp_set_ca_state(sk, TCP_CA_Open); 1982 return 0; 1983 } 1984 1985 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */ 1986 static void tcp_try_undo_dsack(struct sock *sk) 1987 { 1988 struct tcp_sock *tp = tcp_sk(sk); 1989 1990 if (tp->undo_marker && !tp->undo_retrans) { 1991 DBGUNDO(sk, "D-SACK"); 1992 tcp_undo_cwr(sk, 1); 1993 tp->undo_marker = 0; 1994 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO); 1995 } 1996 } 1997 1998 /* Undo during fast recovery after partial ACK. */ 1999 2000 static int tcp_try_undo_partial(struct sock *sk, int acked) 2001 { 2002 struct tcp_sock *tp = tcp_sk(sk); 2003 /* Partial ACK arrived. Force Hoe's retransmit. */ 2004 int failed = IsReno(tp) || tp->fackets_out>tp->reordering; 2005 2006 if (tcp_may_undo(tp)) { 2007 /* Plain luck! Hole if filled with delayed 2008 * packet, rather than with a retransmit. 2009 */ 2010 if (tp->retrans_out == 0) 2011 tp->retrans_stamp = 0; 2012 2013 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1); 2014 2015 DBGUNDO(sk, "Hoe"); 2016 tcp_undo_cwr(sk, 0); 2017 NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO); 2018 2019 /* So... Do not make Hoe's retransmit yet. 2020 * If the first packet was delayed, the rest 2021 * ones are most probably delayed as well. 2022 */ 2023 failed = 0; 2024 } 2025 return failed; 2026 } 2027 2028 /* Undo during loss recovery after partial ACK. */ 2029 static int tcp_try_undo_loss(struct sock *sk) 2030 { 2031 struct tcp_sock *tp = tcp_sk(sk); 2032 2033 if (tcp_may_undo(tp)) { 2034 struct sk_buff *skb; 2035 tcp_for_write_queue(skb, sk) { 2036 if (skb == tcp_send_head(sk)) 2037 break; 2038 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; 2039 } 2040 2041 clear_all_retrans_hints(tp); 2042 2043 DBGUNDO(sk, "partial loss"); 2044 tp->lost_out = 0; 2045 tp->left_out = tp->sacked_out; 2046 tcp_undo_cwr(sk, 1); 2047 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO); 2048 inet_csk(sk)->icsk_retransmits = 0; 2049 tp->undo_marker = 0; 2050 if (!IsReno(tp)) 2051 tcp_set_ca_state(sk, TCP_CA_Open); 2052 return 1; 2053 } 2054 return 0; 2055 } 2056 2057 static inline void tcp_complete_cwr(struct sock *sk) 2058 { 2059 struct tcp_sock *tp = tcp_sk(sk); 2060 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh); 2061 tp->snd_cwnd_stamp = tcp_time_stamp; 2062 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR); 2063 } 2064 2065 static void tcp_try_to_open(struct sock *sk, int flag) 2066 { 2067 struct tcp_sock *tp = tcp_sk(sk); 2068 2069 tcp_sync_left_out(tp); 2070 2071 if (tp->retrans_out == 0) 2072 tp->retrans_stamp = 0; 2073 2074 if (flag&FLAG_ECE) 2075 tcp_enter_cwr(sk, 1); 2076 2077 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) { 2078 int state = TCP_CA_Open; 2079 2080 if (tp->left_out || tp->retrans_out || tp->undo_marker) 2081 state = TCP_CA_Disorder; 2082 2083 if (inet_csk(sk)->icsk_ca_state != state) { 2084 tcp_set_ca_state(sk, state); 2085 tp->high_seq = tp->snd_nxt; 2086 } 2087 tcp_moderate_cwnd(tp); 2088 } else { 2089 tcp_cwnd_down(sk, flag); 2090 } 2091 } 2092 2093 static void tcp_mtup_probe_failed(struct sock *sk) 2094 { 2095 struct inet_connection_sock *icsk = inet_csk(sk); 2096 2097 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1; 2098 icsk->icsk_mtup.probe_size = 0; 2099 } 2100 2101 static void tcp_mtup_probe_success(struct sock *sk, struct sk_buff *skb) 2102 { 2103 struct tcp_sock *tp = tcp_sk(sk); 2104 struct inet_connection_sock *icsk = inet_csk(sk); 2105 2106 /* FIXME: breaks with very large cwnd */ 2107 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2108 tp->snd_cwnd = tp->snd_cwnd * 2109 tcp_mss_to_mtu(sk, tp->mss_cache) / 2110 icsk->icsk_mtup.probe_size; 2111 tp->snd_cwnd_cnt = 0; 2112 tp->snd_cwnd_stamp = tcp_time_stamp; 2113 tp->rcv_ssthresh = tcp_current_ssthresh(sk); 2114 2115 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size; 2116 icsk->icsk_mtup.probe_size = 0; 2117 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 2118 } 2119 2120 2121 /* Process an event, which can update packets-in-flight not trivially. 2122 * Main goal of this function is to calculate new estimate for left_out, 2123 * taking into account both packets sitting in receiver's buffer and 2124 * packets lost by network. 2125 * 2126 * Besides that it does CWND reduction, when packet loss is detected 2127 * and changes state of machine. 2128 * 2129 * It does _not_ decide what to send, it is made in function 2130 * tcp_xmit_retransmit_queue(). 2131 */ 2132 static void 2133 tcp_fastretrans_alert(struct sock *sk, int prior_packets, int flag) 2134 { 2135 struct inet_connection_sock *icsk = inet_csk(sk); 2136 struct tcp_sock *tp = tcp_sk(sk); 2137 int is_dupack = !(flag&(FLAG_SND_UNA_ADVANCED|FLAG_NOT_DUP)); 2138 int do_lost = is_dupack || ((flag&FLAG_DATA_SACKED) && 2139 (tp->fackets_out > tp->reordering)); 2140 2141 /* Some technical things: 2142 * 1. Reno does not count dupacks (sacked_out) automatically. */ 2143 if (!tp->packets_out) 2144 tp->sacked_out = 0; 2145 /* 2. SACK counts snd_fack in packets inaccurately. */ 2146 if (tp->sacked_out == 0) 2147 tp->fackets_out = 0; 2148 2149 /* Now state machine starts. 2150 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */ 2151 if (flag&FLAG_ECE) 2152 tp->prior_ssthresh = 0; 2153 2154 /* B. In all the states check for reneging SACKs. */ 2155 if (tp->sacked_out && tcp_check_sack_reneging(sk)) 2156 return; 2157 2158 /* C. Process data loss notification, provided it is valid. */ 2159 if ((flag&FLAG_DATA_LOST) && 2160 before(tp->snd_una, tp->high_seq) && 2161 icsk->icsk_ca_state != TCP_CA_Open && 2162 tp->fackets_out > tp->reordering) { 2163 tcp_mark_head_lost(sk, tp->fackets_out-tp->reordering, tp->high_seq); 2164 NET_INC_STATS_BH(LINUX_MIB_TCPLOSS); 2165 } 2166 2167 /* D. Synchronize left_out to current state. */ 2168 tcp_sync_left_out(tp); 2169 2170 /* E. Check state exit conditions. State can be terminated 2171 * when high_seq is ACKed. */ 2172 if (icsk->icsk_ca_state == TCP_CA_Open) { 2173 BUG_TRAP(tp->retrans_out == 0); 2174 tp->retrans_stamp = 0; 2175 } else if (!before(tp->snd_una, tp->high_seq)) { 2176 switch (icsk->icsk_ca_state) { 2177 case TCP_CA_Loss: 2178 icsk->icsk_retransmits = 0; 2179 if (tcp_try_undo_recovery(sk)) 2180 return; 2181 break; 2182 2183 case TCP_CA_CWR: 2184 /* CWR is to be held something *above* high_seq 2185 * is ACKed for CWR bit to reach receiver. */ 2186 if (tp->snd_una != tp->high_seq) { 2187 tcp_complete_cwr(sk); 2188 tcp_set_ca_state(sk, TCP_CA_Open); 2189 } 2190 break; 2191 2192 case TCP_CA_Disorder: 2193 tcp_try_undo_dsack(sk); 2194 if (!tp->undo_marker || 2195 /* For SACK case do not Open to allow to undo 2196 * catching for all duplicate ACKs. */ 2197 IsReno(tp) || tp->snd_una != tp->high_seq) { 2198 tp->undo_marker = 0; 2199 tcp_set_ca_state(sk, TCP_CA_Open); 2200 } 2201 break; 2202 2203 case TCP_CA_Recovery: 2204 if (IsReno(tp)) 2205 tcp_reset_reno_sack(tp); 2206 if (tcp_try_undo_recovery(sk)) 2207 return; 2208 tcp_complete_cwr(sk); 2209 break; 2210 } 2211 } 2212 2213 /* F. Process state. */ 2214 switch (icsk->icsk_ca_state) { 2215 case TCP_CA_Recovery: 2216 if (!(flag & FLAG_SND_UNA_ADVANCED)) { 2217 if (IsReno(tp) && is_dupack) 2218 tcp_add_reno_sack(sk); 2219 } else { 2220 int acked = prior_packets - tp->packets_out; 2221 if (IsReno(tp)) 2222 tcp_remove_reno_sacks(sk, acked); 2223 do_lost = tcp_try_undo_partial(sk, acked); 2224 } 2225 break; 2226 case TCP_CA_Loss: 2227 if (flag&FLAG_DATA_ACKED) 2228 icsk->icsk_retransmits = 0; 2229 if (!tcp_try_undo_loss(sk)) { 2230 tcp_moderate_cwnd(tp); 2231 tcp_xmit_retransmit_queue(sk); 2232 return; 2233 } 2234 if (icsk->icsk_ca_state != TCP_CA_Open) 2235 return; 2236 /* Loss is undone; fall through to processing in Open state. */ 2237 default: 2238 if (IsReno(tp)) { 2239 if (flag & FLAG_SND_UNA_ADVANCED) 2240 tcp_reset_reno_sack(tp); 2241 if (is_dupack) 2242 tcp_add_reno_sack(sk); 2243 } 2244 2245 if (icsk->icsk_ca_state == TCP_CA_Disorder) 2246 tcp_try_undo_dsack(sk); 2247 2248 if (!tcp_time_to_recover(sk)) { 2249 tcp_try_to_open(sk, flag); 2250 return; 2251 } 2252 2253 /* MTU probe failure: don't reduce cwnd */ 2254 if (icsk->icsk_ca_state < TCP_CA_CWR && 2255 icsk->icsk_mtup.probe_size && 2256 tp->snd_una == tp->mtu_probe.probe_seq_start) { 2257 tcp_mtup_probe_failed(sk); 2258 /* Restores the reduction we did in tcp_mtup_probe() */ 2259 tp->snd_cwnd++; 2260 tcp_simple_retransmit(sk); 2261 return; 2262 } 2263 2264 /* Otherwise enter Recovery state */ 2265 2266 if (IsReno(tp)) 2267 NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY); 2268 else 2269 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY); 2270 2271 tp->high_seq = tp->snd_nxt; 2272 tp->prior_ssthresh = 0; 2273 tp->undo_marker = tp->snd_una; 2274 tp->undo_retrans = tp->retrans_out; 2275 2276 if (icsk->icsk_ca_state < TCP_CA_CWR) { 2277 if (!(flag&FLAG_ECE)) 2278 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2279 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 2280 TCP_ECN_queue_cwr(tp); 2281 } 2282 2283 tp->bytes_acked = 0; 2284 tp->snd_cwnd_cnt = 0; 2285 tcp_set_ca_state(sk, TCP_CA_Recovery); 2286 } 2287 2288 if (do_lost || tcp_head_timedout(sk)) 2289 tcp_update_scoreboard(sk); 2290 tcp_cwnd_down(sk, flag); 2291 tcp_xmit_retransmit_queue(sk); 2292 } 2293 2294 /* Read draft-ietf-tcplw-high-performance before mucking 2295 * with this code. (Supersedes RFC1323) 2296 */ 2297 static void tcp_ack_saw_tstamp(struct sock *sk, int flag) 2298 { 2299 /* RTTM Rule: A TSecr value received in a segment is used to 2300 * update the averaged RTT measurement only if the segment 2301 * acknowledges some new data, i.e., only if it advances the 2302 * left edge of the send window. 2303 * 2304 * See draft-ietf-tcplw-high-performance-00, section 3.3. 2305 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru> 2306 * 2307 * Changed: reset backoff as soon as we see the first valid sample. 2308 * If we do not, we get strongly overestimated rto. With timestamps 2309 * samples are accepted even from very old segments: f.e., when rtt=1 2310 * increases to 8, we retransmit 5 times and after 8 seconds delayed 2311 * answer arrives rto becomes 120 seconds! If at least one of segments 2312 * in window is lost... Voila. --ANK (010210) 2313 */ 2314 struct tcp_sock *tp = tcp_sk(sk); 2315 const __u32 seq_rtt = tcp_time_stamp - tp->rx_opt.rcv_tsecr; 2316 tcp_rtt_estimator(sk, seq_rtt); 2317 tcp_set_rto(sk); 2318 inet_csk(sk)->icsk_backoff = 0; 2319 tcp_bound_rto(sk); 2320 } 2321 2322 static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag) 2323 { 2324 /* We don't have a timestamp. Can only use 2325 * packets that are not retransmitted to determine 2326 * rtt estimates. Also, we must not reset the 2327 * backoff for rto until we get a non-retransmitted 2328 * packet. This allows us to deal with a situation 2329 * where the network delay has increased suddenly. 2330 * I.e. Karn's algorithm. (SIGCOMM '87, p5.) 2331 */ 2332 2333 if (flag & FLAG_RETRANS_DATA_ACKED) 2334 return; 2335 2336 tcp_rtt_estimator(sk, seq_rtt); 2337 tcp_set_rto(sk); 2338 inet_csk(sk)->icsk_backoff = 0; 2339 tcp_bound_rto(sk); 2340 } 2341 2342 static inline void tcp_ack_update_rtt(struct sock *sk, const int flag, 2343 const s32 seq_rtt) 2344 { 2345 const struct tcp_sock *tp = tcp_sk(sk); 2346 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */ 2347 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr) 2348 tcp_ack_saw_tstamp(sk, flag); 2349 else if (seq_rtt >= 0) 2350 tcp_ack_no_tstamp(sk, seq_rtt, flag); 2351 } 2352 2353 static void tcp_cong_avoid(struct sock *sk, u32 ack, 2354 u32 in_flight, int good) 2355 { 2356 const struct inet_connection_sock *icsk = inet_csk(sk); 2357 icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight, good); 2358 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp; 2359 } 2360 2361 /* Restart timer after forward progress on connection. 2362 * RFC2988 recommends to restart timer to now+rto. 2363 */ 2364 2365 static void tcp_ack_packets_out(struct sock *sk) 2366 { 2367 struct tcp_sock *tp = tcp_sk(sk); 2368 2369 if (!tp->packets_out) { 2370 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS); 2371 } else { 2372 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 2373 } 2374 } 2375 2376 static int tcp_tso_acked(struct sock *sk, struct sk_buff *skb, 2377 __u32 now, __s32 *seq_rtt) 2378 { 2379 struct tcp_sock *tp = tcp_sk(sk); 2380 struct tcp_skb_cb *scb = TCP_SKB_CB(skb); 2381 __u32 seq = tp->snd_una; 2382 __u32 packets_acked; 2383 int acked = 0; 2384 2385 /* If we get here, the whole TSO packet has not been 2386 * acked. 2387 */ 2388 BUG_ON(!after(scb->end_seq, seq)); 2389 2390 packets_acked = tcp_skb_pcount(skb); 2391 if (tcp_trim_head(sk, skb, seq - scb->seq)) 2392 return 0; 2393 packets_acked -= tcp_skb_pcount(skb); 2394 2395 if (packets_acked) { 2396 __u8 sacked = scb->sacked; 2397 2398 acked |= FLAG_DATA_ACKED; 2399 if (sacked) { 2400 if (sacked & TCPCB_RETRANS) { 2401 if (sacked & TCPCB_SACKED_RETRANS) 2402 tp->retrans_out -= packets_acked; 2403 acked |= FLAG_RETRANS_DATA_ACKED; 2404 *seq_rtt = -1; 2405 } else if (*seq_rtt < 0) 2406 *seq_rtt = now - scb->when; 2407 if (sacked & TCPCB_SACKED_ACKED) 2408 tp->sacked_out -= packets_acked; 2409 if (sacked & TCPCB_LOST) 2410 tp->lost_out -= packets_acked; 2411 if (sacked & TCPCB_URG) { 2412 if (tp->urg_mode && 2413 !before(seq, tp->snd_up)) 2414 tp->urg_mode = 0; 2415 } 2416 } else if (*seq_rtt < 0) 2417 *seq_rtt = now - scb->when; 2418 2419 if (tp->fackets_out) { 2420 __u32 dval = min(tp->fackets_out, packets_acked); 2421 tp->fackets_out -= dval; 2422 } 2423 /* hint's skb might be NULL but we don't need to care */ 2424 tp->fastpath_cnt_hint -= min_t(u32, packets_acked, 2425 tp->fastpath_cnt_hint); 2426 tp->packets_out -= packets_acked; 2427 2428 BUG_ON(tcp_skb_pcount(skb) == 0); 2429 BUG_ON(!before(scb->seq, scb->end_seq)); 2430 } 2431 2432 return acked; 2433 } 2434 2435 /* Remove acknowledged frames from the retransmission queue. */ 2436 static int tcp_clean_rtx_queue(struct sock *sk, __s32 *seq_rtt_p) 2437 { 2438 struct tcp_sock *tp = tcp_sk(sk); 2439 const struct inet_connection_sock *icsk = inet_csk(sk); 2440 struct sk_buff *skb; 2441 __u32 now = tcp_time_stamp; 2442 int acked = 0; 2443 int prior_packets = tp->packets_out; 2444 __s32 seq_rtt = -1; 2445 ktime_t last_ackt = net_invalid_timestamp(); 2446 2447 while ((skb = tcp_write_queue_head(sk)) && 2448 skb != tcp_send_head(sk)) { 2449 struct tcp_skb_cb *scb = TCP_SKB_CB(skb); 2450 __u8 sacked = scb->sacked; 2451 2452 /* If our packet is before the ack sequence we can 2453 * discard it as it's confirmed to have arrived at 2454 * the other end. 2455 */ 2456 if (after(scb->end_seq, tp->snd_una)) { 2457 if (tcp_skb_pcount(skb) > 1 && 2458 after(tp->snd_una, scb->seq)) 2459 acked |= tcp_tso_acked(sk, skb, 2460 now, &seq_rtt); 2461 break; 2462 } 2463 2464 /* Initial outgoing SYN's get put onto the write_queue 2465 * just like anything else we transmit. It is not 2466 * true data, and if we misinform our callers that 2467 * this ACK acks real data, we will erroneously exit 2468 * connection startup slow start one packet too 2469 * quickly. This is severely frowned upon behavior. 2470 */ 2471 if (!(scb->flags & TCPCB_FLAG_SYN)) { 2472 acked |= FLAG_DATA_ACKED; 2473 } else { 2474 acked |= FLAG_SYN_ACKED; 2475 tp->retrans_stamp = 0; 2476 } 2477 2478 /* MTU probing checks */ 2479 if (icsk->icsk_mtup.probe_size) { 2480 if (!after(tp->mtu_probe.probe_seq_end, TCP_SKB_CB(skb)->end_seq)) { 2481 tcp_mtup_probe_success(sk, skb); 2482 } 2483 } 2484 2485 if (sacked) { 2486 if (sacked & TCPCB_RETRANS) { 2487 if (sacked & TCPCB_SACKED_RETRANS) 2488 tp->retrans_out -= tcp_skb_pcount(skb); 2489 acked |= FLAG_RETRANS_DATA_ACKED; 2490 seq_rtt = -1; 2491 } else if (seq_rtt < 0) { 2492 seq_rtt = now - scb->when; 2493 last_ackt = skb->tstamp; 2494 } 2495 if (sacked & TCPCB_SACKED_ACKED) 2496 tp->sacked_out -= tcp_skb_pcount(skb); 2497 if (sacked & TCPCB_LOST) 2498 tp->lost_out -= tcp_skb_pcount(skb); 2499 if (sacked & TCPCB_URG) { 2500 if (tp->urg_mode && 2501 !before(scb->end_seq, tp->snd_up)) 2502 tp->urg_mode = 0; 2503 } 2504 } else if (seq_rtt < 0) { 2505 seq_rtt = now - scb->when; 2506 last_ackt = skb->tstamp; 2507 } 2508 tcp_dec_pcount_approx(&tp->fackets_out, skb); 2509 tcp_packets_out_dec(tp, skb); 2510 tcp_unlink_write_queue(skb, sk); 2511 sk_stream_free_skb(sk, skb); 2512 clear_all_retrans_hints(tp); 2513 } 2514 2515 if (acked&FLAG_ACKED) { 2516 u32 pkts_acked = prior_packets - tp->packets_out; 2517 const struct tcp_congestion_ops *ca_ops 2518 = inet_csk(sk)->icsk_ca_ops; 2519 2520 tcp_ack_update_rtt(sk, acked, seq_rtt); 2521 tcp_ack_packets_out(sk); 2522 2523 if (ca_ops->pkts_acked) { 2524 s32 rtt_us = -1; 2525 2526 /* Is the ACK triggering packet unambiguous? */ 2527 if (!(acked & FLAG_RETRANS_DATA_ACKED)) { 2528 /* High resolution needed and available? */ 2529 if (ca_ops->flags & TCP_CONG_RTT_STAMP && 2530 !ktime_equal(last_ackt, 2531 net_invalid_timestamp())) 2532 rtt_us = ktime_us_delta(ktime_get_real(), 2533 last_ackt); 2534 else if (seq_rtt > 0) 2535 rtt_us = jiffies_to_usecs(seq_rtt); 2536 } 2537 2538 ca_ops->pkts_acked(sk, pkts_acked, rtt_us); 2539 } 2540 } 2541 2542 #if FASTRETRANS_DEBUG > 0 2543 BUG_TRAP((int)tp->sacked_out >= 0); 2544 BUG_TRAP((int)tp->lost_out >= 0); 2545 BUG_TRAP((int)tp->retrans_out >= 0); 2546 if (!tp->packets_out && tp->rx_opt.sack_ok) { 2547 const struct inet_connection_sock *icsk = inet_csk(sk); 2548 if (tp->lost_out) { 2549 printk(KERN_DEBUG "Leak l=%u %d\n", 2550 tp->lost_out, icsk->icsk_ca_state); 2551 tp->lost_out = 0; 2552 } 2553 if (tp->sacked_out) { 2554 printk(KERN_DEBUG "Leak s=%u %d\n", 2555 tp->sacked_out, icsk->icsk_ca_state); 2556 tp->sacked_out = 0; 2557 } 2558 if (tp->retrans_out) { 2559 printk(KERN_DEBUG "Leak r=%u %d\n", 2560 tp->retrans_out, icsk->icsk_ca_state); 2561 tp->retrans_out = 0; 2562 } 2563 } 2564 #endif 2565 *seq_rtt_p = seq_rtt; 2566 return acked; 2567 } 2568 2569 static void tcp_ack_probe(struct sock *sk) 2570 { 2571 const struct tcp_sock *tp = tcp_sk(sk); 2572 struct inet_connection_sock *icsk = inet_csk(sk); 2573 2574 /* Was it a usable window open? */ 2575 2576 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, 2577 tp->snd_una + tp->snd_wnd)) { 2578 icsk->icsk_backoff = 0; 2579 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0); 2580 /* Socket must be waked up by subsequent tcp_data_snd_check(). 2581 * This function is not for random using! 2582 */ 2583 } else { 2584 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 2585 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), 2586 TCP_RTO_MAX); 2587 } 2588 } 2589 2590 static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag) 2591 { 2592 return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) || 2593 inet_csk(sk)->icsk_ca_state != TCP_CA_Open); 2594 } 2595 2596 static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag) 2597 { 2598 const struct tcp_sock *tp = tcp_sk(sk); 2599 return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) && 2600 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR)); 2601 } 2602 2603 /* Check that window update is acceptable. 2604 * The function assumes that snd_una<=ack<=snd_next. 2605 */ 2606 static inline int tcp_may_update_window(const struct tcp_sock *tp, const u32 ack, 2607 const u32 ack_seq, const u32 nwin) 2608 { 2609 return (after(ack, tp->snd_una) || 2610 after(ack_seq, tp->snd_wl1) || 2611 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd)); 2612 } 2613 2614 /* Update our send window. 2615 * 2616 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2 2617 * and in FreeBSD. NetBSD's one is even worse.) is wrong. 2618 */ 2619 static int tcp_ack_update_window(struct sock *sk, struct sk_buff *skb, u32 ack, 2620 u32 ack_seq) 2621 { 2622 struct tcp_sock *tp = tcp_sk(sk); 2623 int flag = 0; 2624 u32 nwin = ntohs(tcp_hdr(skb)->window); 2625 2626 if (likely(!tcp_hdr(skb)->syn)) 2627 nwin <<= tp->rx_opt.snd_wscale; 2628 2629 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) { 2630 flag |= FLAG_WIN_UPDATE; 2631 tcp_update_wl(tp, ack, ack_seq); 2632 2633 if (tp->snd_wnd != nwin) { 2634 tp->snd_wnd = nwin; 2635 2636 /* Note, it is the only place, where 2637 * fast path is recovered for sending TCP. 2638 */ 2639 tp->pred_flags = 0; 2640 tcp_fast_path_check(sk); 2641 2642 if (nwin > tp->max_window) { 2643 tp->max_window = nwin; 2644 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie); 2645 } 2646 } 2647 } 2648 2649 tp->snd_una = ack; 2650 2651 return flag; 2652 } 2653 2654 /* A very conservative spurious RTO response algorithm: reduce cwnd and 2655 * continue in congestion avoidance. 2656 */ 2657 static void tcp_conservative_spur_to_response(struct tcp_sock *tp) 2658 { 2659 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh); 2660 tp->snd_cwnd_cnt = 0; 2661 TCP_ECN_queue_cwr(tp); 2662 tcp_moderate_cwnd(tp); 2663 } 2664 2665 /* A conservative spurious RTO response algorithm: reduce cwnd using 2666 * rate halving and continue in congestion avoidance. 2667 */ 2668 static void tcp_ratehalving_spur_to_response(struct sock *sk) 2669 { 2670 tcp_enter_cwr(sk, 0); 2671 } 2672 2673 static void tcp_undo_spur_to_response(struct sock *sk, int flag) 2674 { 2675 if (flag&FLAG_ECE) 2676 tcp_ratehalving_spur_to_response(sk); 2677 else 2678 tcp_undo_cwr(sk, 1); 2679 } 2680 2681 /* F-RTO spurious RTO detection algorithm (RFC4138) 2682 * 2683 * F-RTO affects during two new ACKs following RTO (well, almost, see inline 2684 * comments). State (ACK number) is kept in frto_counter. When ACK advances 2685 * window (but not to or beyond highest sequence sent before RTO): 2686 * On First ACK, send two new segments out. 2687 * On Second ACK, RTO was likely spurious. Do spurious response (response 2688 * algorithm is not part of the F-RTO detection algorithm 2689 * given in RFC4138 but can be selected separately). 2690 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss 2691 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding 2692 * of Nagle, this is done using frto_counter states 2 and 3, when a new data 2693 * segment of any size sent during F-RTO, state 2 is upgraded to 3. 2694 * 2695 * Rationale: if the RTO was spurious, new ACKs should arrive from the 2696 * original window even after we transmit two new data segments. 2697 * 2698 * SACK version: 2699 * on first step, wait until first cumulative ACK arrives, then move to 2700 * the second step. In second step, the next ACK decides. 2701 * 2702 * F-RTO is implemented (mainly) in four functions: 2703 * - tcp_use_frto() is used to determine if TCP is can use F-RTO 2704 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is 2705 * called when tcp_use_frto() showed green light 2706 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm 2707 * - tcp_enter_frto_loss() is called if there is not enough evidence 2708 * to prove that the RTO is indeed spurious. It transfers the control 2709 * from F-RTO to the conventional RTO recovery 2710 */ 2711 static int tcp_process_frto(struct sock *sk, int flag) 2712 { 2713 struct tcp_sock *tp = tcp_sk(sk); 2714 2715 tcp_sync_left_out(tp); 2716 2717 /* Duplicate the behavior from Loss state (fastretrans_alert) */ 2718 if (flag&FLAG_DATA_ACKED) 2719 inet_csk(sk)->icsk_retransmits = 0; 2720 2721 if (!before(tp->snd_una, tp->frto_highmark)) { 2722 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag); 2723 return 1; 2724 } 2725 2726 if (!IsSackFrto() || IsReno(tp)) { 2727 /* RFC4138 shortcoming in step 2; should also have case c): 2728 * ACK isn't duplicate nor advances window, e.g., opposite dir 2729 * data, winupdate 2730 */ 2731 if (!(flag&FLAG_ANY_PROGRESS) && (flag&FLAG_NOT_DUP)) 2732 return 1; 2733 2734 if (!(flag&FLAG_DATA_ACKED)) { 2735 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3), 2736 flag); 2737 return 1; 2738 } 2739 } else { 2740 if (!(flag&FLAG_DATA_ACKED) && (tp->frto_counter == 1)) { 2741 /* Prevent sending of new data. */ 2742 tp->snd_cwnd = min(tp->snd_cwnd, 2743 tcp_packets_in_flight(tp)); 2744 return 1; 2745 } 2746 2747 if ((tp->frto_counter >= 2) && 2748 (!(flag&FLAG_FORWARD_PROGRESS) || 2749 ((flag&FLAG_DATA_SACKED) && !(flag&FLAG_ONLY_ORIG_SACKED)))) { 2750 /* RFC4138 shortcoming (see comment above) */ 2751 if (!(flag&FLAG_FORWARD_PROGRESS) && (flag&FLAG_NOT_DUP)) 2752 return 1; 2753 2754 tcp_enter_frto_loss(sk, 3, flag); 2755 return 1; 2756 } 2757 } 2758 2759 if (tp->frto_counter == 1) { 2760 /* Sending of the next skb must be allowed or no FRTO */ 2761 if (!tcp_send_head(sk) || 2762 after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, 2763 tp->snd_una + tp->snd_wnd)) { 2764 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), 2765 flag); 2766 return 1; 2767 } 2768 2769 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2; 2770 tp->frto_counter = 2; 2771 return 1; 2772 } else { 2773 switch (sysctl_tcp_frto_response) { 2774 case 2: 2775 tcp_undo_spur_to_response(sk, flag); 2776 break; 2777 case 1: 2778 tcp_conservative_spur_to_response(tp); 2779 break; 2780 default: 2781 tcp_ratehalving_spur_to_response(sk); 2782 break; 2783 } 2784 tp->frto_counter = 0; 2785 } 2786 return 0; 2787 } 2788 2789 /* This routine deals with incoming acks, but not outgoing ones. */ 2790 static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag) 2791 { 2792 struct inet_connection_sock *icsk = inet_csk(sk); 2793 struct tcp_sock *tp = tcp_sk(sk); 2794 u32 prior_snd_una = tp->snd_una; 2795 u32 ack_seq = TCP_SKB_CB(skb)->seq; 2796 u32 ack = TCP_SKB_CB(skb)->ack_seq; 2797 u32 prior_in_flight; 2798 s32 seq_rtt; 2799 int prior_packets; 2800 int frto_cwnd = 0; 2801 2802 /* If the ack is newer than sent or older than previous acks 2803 * then we can probably ignore it. 2804 */ 2805 if (after(ack, tp->snd_nxt)) 2806 goto uninteresting_ack; 2807 2808 if (before(ack, prior_snd_una)) 2809 goto old_ack; 2810 2811 if (after(ack, prior_snd_una)) 2812 flag |= FLAG_SND_UNA_ADVANCED; 2813 2814 if (sysctl_tcp_abc) { 2815 if (icsk->icsk_ca_state < TCP_CA_CWR) 2816 tp->bytes_acked += ack - prior_snd_una; 2817 else if (icsk->icsk_ca_state == TCP_CA_Loss) 2818 /* we assume just one segment left network */ 2819 tp->bytes_acked += min(ack - prior_snd_una, tp->mss_cache); 2820 } 2821 2822 if (!(flag&FLAG_SLOWPATH) && after(ack, prior_snd_una)) { 2823 /* Window is constant, pure forward advance. 2824 * No more checks are required. 2825 * Note, we use the fact that SND.UNA>=SND.WL2. 2826 */ 2827 tcp_update_wl(tp, ack, ack_seq); 2828 tp->snd_una = ack; 2829 flag |= FLAG_WIN_UPDATE; 2830 2831 tcp_ca_event(sk, CA_EVENT_FAST_ACK); 2832 2833 NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS); 2834 } else { 2835 if (ack_seq != TCP_SKB_CB(skb)->end_seq) 2836 flag |= FLAG_DATA; 2837 else 2838 NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS); 2839 2840 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq); 2841 2842 if (TCP_SKB_CB(skb)->sacked) 2843 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una); 2844 2845 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb))) 2846 flag |= FLAG_ECE; 2847 2848 tcp_ca_event(sk, CA_EVENT_SLOW_ACK); 2849 } 2850 2851 /* We passed data and got it acked, remove any soft error 2852 * log. Something worked... 2853 */ 2854 sk->sk_err_soft = 0; 2855 tp->rcv_tstamp = tcp_time_stamp; 2856 prior_packets = tp->packets_out; 2857 if (!prior_packets) 2858 goto no_queue; 2859 2860 prior_in_flight = tcp_packets_in_flight(tp); 2861 2862 /* See if we can take anything off of the retransmit queue. */ 2863 flag |= tcp_clean_rtx_queue(sk, &seq_rtt); 2864 2865 if (tp->frto_counter) 2866 frto_cwnd = tcp_process_frto(sk, flag); 2867 2868 if (tcp_ack_is_dubious(sk, flag)) { 2869 /* Advance CWND, if state allows this. */ 2870 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd && 2871 tcp_may_raise_cwnd(sk, flag)) 2872 tcp_cong_avoid(sk, ack, prior_in_flight, 0); 2873 tcp_fastretrans_alert(sk, prior_packets, flag); 2874 } else { 2875 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd) 2876 tcp_cong_avoid(sk, ack, prior_in_flight, 1); 2877 } 2878 2879 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag&FLAG_NOT_DUP)) 2880 dst_confirm(sk->sk_dst_cache); 2881 2882 return 1; 2883 2884 no_queue: 2885 icsk->icsk_probes_out = 0; 2886 2887 /* If this ack opens up a zero window, clear backoff. It was 2888 * being used to time the probes, and is probably far higher than 2889 * it needs to be for normal retransmission. 2890 */ 2891 if (tcp_send_head(sk)) 2892 tcp_ack_probe(sk); 2893 return 1; 2894 2895 old_ack: 2896 if (TCP_SKB_CB(skb)->sacked) 2897 tcp_sacktag_write_queue(sk, skb, prior_snd_una); 2898 2899 uninteresting_ack: 2900 SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt); 2901 return 0; 2902 } 2903 2904 2905 /* Look for tcp options. Normally only called on SYN and SYNACK packets. 2906 * But, this can also be called on packets in the established flow when 2907 * the fast version below fails. 2908 */ 2909 void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx, int estab) 2910 { 2911 unsigned char *ptr; 2912 struct tcphdr *th = tcp_hdr(skb); 2913 int length=(th->doff*4)-sizeof(struct tcphdr); 2914 2915 ptr = (unsigned char *)(th + 1); 2916 opt_rx->saw_tstamp = 0; 2917 2918 while (length > 0) { 2919 int opcode=*ptr++; 2920 int opsize; 2921 2922 switch (opcode) { 2923 case TCPOPT_EOL: 2924 return; 2925 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */ 2926 length--; 2927 continue; 2928 default: 2929 opsize=*ptr++; 2930 if (opsize < 2) /* "silly options" */ 2931 return; 2932 if (opsize > length) 2933 return; /* don't parse partial options */ 2934 switch (opcode) { 2935 case TCPOPT_MSS: 2936 if (opsize==TCPOLEN_MSS && th->syn && !estab) { 2937 u16 in_mss = ntohs(get_unaligned((__be16 *)ptr)); 2938 if (in_mss) { 2939 if (opt_rx->user_mss && opt_rx->user_mss < in_mss) 2940 in_mss = opt_rx->user_mss; 2941 opt_rx->mss_clamp = in_mss; 2942 } 2943 } 2944 break; 2945 case TCPOPT_WINDOW: 2946 if (opsize==TCPOLEN_WINDOW && th->syn && !estab) 2947 if (sysctl_tcp_window_scaling) { 2948 __u8 snd_wscale = *(__u8 *) ptr; 2949 opt_rx->wscale_ok = 1; 2950 if (snd_wscale > 14) { 2951 if (net_ratelimit()) 2952 printk(KERN_INFO "tcp_parse_options: Illegal window " 2953 "scaling value %d >14 received.\n", 2954 snd_wscale); 2955 snd_wscale = 14; 2956 } 2957 opt_rx->snd_wscale = snd_wscale; 2958 } 2959 break; 2960 case TCPOPT_TIMESTAMP: 2961 if (opsize==TCPOLEN_TIMESTAMP) { 2962 if ((estab && opt_rx->tstamp_ok) || 2963 (!estab && sysctl_tcp_timestamps)) { 2964 opt_rx->saw_tstamp = 1; 2965 opt_rx->rcv_tsval = ntohl(get_unaligned((__be32 *)ptr)); 2966 opt_rx->rcv_tsecr = ntohl(get_unaligned((__be32 *)(ptr+4))); 2967 } 2968 } 2969 break; 2970 case TCPOPT_SACK_PERM: 2971 if (opsize==TCPOLEN_SACK_PERM && th->syn && !estab) { 2972 if (sysctl_tcp_sack) { 2973 opt_rx->sack_ok = 1; 2974 tcp_sack_reset(opt_rx); 2975 } 2976 } 2977 break; 2978 2979 case TCPOPT_SACK: 2980 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) && 2981 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) && 2982 opt_rx->sack_ok) { 2983 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th; 2984 } 2985 break; 2986 #ifdef CONFIG_TCP_MD5SIG 2987 case TCPOPT_MD5SIG: 2988 /* 2989 * The MD5 Hash has already been 2990 * checked (see tcp_v{4,6}_do_rcv()). 2991 */ 2992 break; 2993 #endif 2994 } 2995 2996 ptr+=opsize-2; 2997 length-=opsize; 2998 } 2999 } 3000 } 3001 3002 /* Fast parse options. This hopes to only see timestamps. 3003 * If it is wrong it falls back on tcp_parse_options(). 3004 */ 3005 static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th, 3006 struct tcp_sock *tp) 3007 { 3008 if (th->doff == sizeof(struct tcphdr)>>2) { 3009 tp->rx_opt.saw_tstamp = 0; 3010 return 0; 3011 } else if (tp->rx_opt.tstamp_ok && 3012 th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) { 3013 __be32 *ptr = (__be32 *)(th + 1); 3014 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) 3015 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) { 3016 tp->rx_opt.saw_tstamp = 1; 3017 ++ptr; 3018 tp->rx_opt.rcv_tsval = ntohl(*ptr); 3019 ++ptr; 3020 tp->rx_opt.rcv_tsecr = ntohl(*ptr); 3021 return 1; 3022 } 3023 } 3024 tcp_parse_options(skb, &tp->rx_opt, 1); 3025 return 1; 3026 } 3027 3028 static inline void tcp_store_ts_recent(struct tcp_sock *tp) 3029 { 3030 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval; 3031 tp->rx_opt.ts_recent_stamp = get_seconds(); 3032 } 3033 3034 static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq) 3035 { 3036 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) { 3037 /* PAWS bug workaround wrt. ACK frames, the PAWS discard 3038 * extra check below makes sure this can only happen 3039 * for pure ACK frames. -DaveM 3040 * 3041 * Not only, also it occurs for expired timestamps. 3042 */ 3043 3044 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) >= 0 || 3045 get_seconds() >= tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS) 3046 tcp_store_ts_recent(tp); 3047 } 3048 } 3049 3050 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM 3051 * 3052 * It is not fatal. If this ACK does _not_ change critical state (seqs, window) 3053 * it can pass through stack. So, the following predicate verifies that 3054 * this segment is not used for anything but congestion avoidance or 3055 * fast retransmit. Moreover, we even are able to eliminate most of such 3056 * second order effects, if we apply some small "replay" window (~RTO) 3057 * to timestamp space. 3058 * 3059 * All these measures still do not guarantee that we reject wrapped ACKs 3060 * on networks with high bandwidth, when sequence space is recycled fastly, 3061 * but it guarantees that such events will be very rare and do not affect 3062 * connection seriously. This doesn't look nice, but alas, PAWS is really 3063 * buggy extension. 3064 * 3065 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC 3066 * states that events when retransmit arrives after original data are rare. 3067 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is 3068 * the biggest problem on large power networks even with minor reordering. 3069 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe 3070 * up to bandwidth of 18Gigabit/sec. 8) ] 3071 */ 3072 3073 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb) 3074 { 3075 struct tcp_sock *tp = tcp_sk(sk); 3076 struct tcphdr *th = tcp_hdr(skb); 3077 u32 seq = TCP_SKB_CB(skb)->seq; 3078 u32 ack = TCP_SKB_CB(skb)->ack_seq; 3079 3080 return (/* 1. Pure ACK with correct sequence number. */ 3081 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) && 3082 3083 /* 2. ... and duplicate ACK. */ 3084 ack == tp->snd_una && 3085 3086 /* 3. ... and does not update window. */ 3087 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) && 3088 3089 /* 4. ... and sits in replay window. */ 3090 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ); 3091 } 3092 3093 static inline int tcp_paws_discard(const struct sock *sk, const struct sk_buff *skb) 3094 { 3095 const struct tcp_sock *tp = tcp_sk(sk); 3096 return ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > TCP_PAWS_WINDOW && 3097 get_seconds() < tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS && 3098 !tcp_disordered_ack(sk, skb)); 3099 } 3100 3101 /* Check segment sequence number for validity. 3102 * 3103 * Segment controls are considered valid, if the segment 3104 * fits to the window after truncation to the window. Acceptability 3105 * of data (and SYN, FIN, of course) is checked separately. 3106 * See tcp_data_queue(), for example. 3107 * 3108 * Also, controls (RST is main one) are accepted using RCV.WUP instead 3109 * of RCV.NXT. Peer still did not advance his SND.UNA when we 3110 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP. 3111 * (borrowed from freebsd) 3112 */ 3113 3114 static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq) 3115 { 3116 return !before(end_seq, tp->rcv_wup) && 3117 !after(seq, tp->rcv_nxt + tcp_receive_window(tp)); 3118 } 3119 3120 /* When we get a reset we do this. */ 3121 static void tcp_reset(struct sock *sk) 3122 { 3123 /* We want the right error as BSD sees it (and indeed as we do). */ 3124 switch (sk->sk_state) { 3125 case TCP_SYN_SENT: 3126 sk->sk_err = ECONNREFUSED; 3127 break; 3128 case TCP_CLOSE_WAIT: 3129 sk->sk_err = EPIPE; 3130 break; 3131 case TCP_CLOSE: 3132 return; 3133 default: 3134 sk->sk_err = ECONNRESET; 3135 } 3136 3137 if (!sock_flag(sk, SOCK_DEAD)) 3138 sk->sk_error_report(sk); 3139 3140 tcp_done(sk); 3141 } 3142 3143 /* 3144 * Process the FIN bit. This now behaves as it is supposed to work 3145 * and the FIN takes effect when it is validly part of sequence 3146 * space. Not before when we get holes. 3147 * 3148 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT 3149 * (and thence onto LAST-ACK and finally, CLOSE, we never enter 3150 * TIME-WAIT) 3151 * 3152 * If we are in FINWAIT-1, a received FIN indicates simultaneous 3153 * close and we go into CLOSING (and later onto TIME-WAIT) 3154 * 3155 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT. 3156 */ 3157 static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th) 3158 { 3159 struct tcp_sock *tp = tcp_sk(sk); 3160 3161 inet_csk_schedule_ack(sk); 3162 3163 sk->sk_shutdown |= RCV_SHUTDOWN; 3164 sock_set_flag(sk, SOCK_DONE); 3165 3166 switch (sk->sk_state) { 3167 case TCP_SYN_RECV: 3168 case TCP_ESTABLISHED: 3169 /* Move to CLOSE_WAIT */ 3170 tcp_set_state(sk, TCP_CLOSE_WAIT); 3171 inet_csk(sk)->icsk_ack.pingpong = 1; 3172 break; 3173 3174 case TCP_CLOSE_WAIT: 3175 case TCP_CLOSING: 3176 /* Received a retransmission of the FIN, do 3177 * nothing. 3178 */ 3179 break; 3180 case TCP_LAST_ACK: 3181 /* RFC793: Remain in the LAST-ACK state. */ 3182 break; 3183 3184 case TCP_FIN_WAIT1: 3185 /* This case occurs when a simultaneous close 3186 * happens, we must ack the received FIN and 3187 * enter the CLOSING state. 3188 */ 3189 tcp_send_ack(sk); 3190 tcp_set_state(sk, TCP_CLOSING); 3191 break; 3192 case TCP_FIN_WAIT2: 3193 /* Received a FIN -- send ACK and enter TIME_WAIT. */ 3194 tcp_send_ack(sk); 3195 tcp_time_wait(sk, TCP_TIME_WAIT, 0); 3196 break; 3197 default: 3198 /* Only TCP_LISTEN and TCP_CLOSE are left, in these 3199 * cases we should never reach this piece of code. 3200 */ 3201 printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n", 3202 __FUNCTION__, sk->sk_state); 3203 break; 3204 } 3205 3206 /* It _is_ possible, that we have something out-of-order _after_ FIN. 3207 * Probably, we should reset in this case. For now drop them. 3208 */ 3209 __skb_queue_purge(&tp->out_of_order_queue); 3210 if (tp->rx_opt.sack_ok) 3211 tcp_sack_reset(&tp->rx_opt); 3212 sk_stream_mem_reclaim(sk); 3213 3214 if (!sock_flag(sk, SOCK_DEAD)) { 3215 sk->sk_state_change(sk); 3216 3217 /* Do not send POLL_HUP for half duplex close. */ 3218 if (sk->sk_shutdown == SHUTDOWN_MASK || 3219 sk->sk_state == TCP_CLOSE) 3220 sk_wake_async(sk, 1, POLL_HUP); 3221 else 3222 sk_wake_async(sk, 1, POLL_IN); 3223 } 3224 } 3225 3226 static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, u32 end_seq) 3227 { 3228 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) { 3229 if (before(seq, sp->start_seq)) 3230 sp->start_seq = seq; 3231 if (after(end_seq, sp->end_seq)) 3232 sp->end_seq = end_seq; 3233 return 1; 3234 } 3235 return 0; 3236 } 3237 3238 static void tcp_dsack_set(struct tcp_sock *tp, u32 seq, u32 end_seq) 3239 { 3240 if (tp->rx_opt.sack_ok && sysctl_tcp_dsack) { 3241 if (before(seq, tp->rcv_nxt)) 3242 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT); 3243 else 3244 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT); 3245 3246 tp->rx_opt.dsack = 1; 3247 tp->duplicate_sack[0].start_seq = seq; 3248 tp->duplicate_sack[0].end_seq = end_seq; 3249 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + 1, 4 - tp->rx_opt.tstamp_ok); 3250 } 3251 } 3252 3253 static void tcp_dsack_extend(struct tcp_sock *tp, u32 seq, u32 end_seq) 3254 { 3255 if (!tp->rx_opt.dsack) 3256 tcp_dsack_set(tp, seq, end_seq); 3257 else 3258 tcp_sack_extend(tp->duplicate_sack, seq, end_seq); 3259 } 3260 3261 static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb) 3262 { 3263 struct tcp_sock *tp = tcp_sk(sk); 3264 3265 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 3266 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 3267 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST); 3268 tcp_enter_quickack_mode(sk); 3269 3270 if (tp->rx_opt.sack_ok && sysctl_tcp_dsack) { 3271 u32 end_seq = TCP_SKB_CB(skb)->end_seq; 3272 3273 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) 3274 end_seq = tp->rcv_nxt; 3275 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, end_seq); 3276 } 3277 } 3278 3279 tcp_send_ack(sk); 3280 } 3281 3282 /* These routines update the SACK block as out-of-order packets arrive or 3283 * in-order packets close up the sequence space. 3284 */ 3285 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp) 3286 { 3287 int this_sack; 3288 struct tcp_sack_block *sp = &tp->selective_acks[0]; 3289 struct tcp_sack_block *swalk = sp+1; 3290 3291 /* See if the recent change to the first SACK eats into 3292 * or hits the sequence space of other SACK blocks, if so coalesce. 3293 */ 3294 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks; ) { 3295 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) { 3296 int i; 3297 3298 /* Zap SWALK, by moving every further SACK up by one slot. 3299 * Decrease num_sacks. 3300 */ 3301 tp->rx_opt.num_sacks--; 3302 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok); 3303 for (i=this_sack; i < tp->rx_opt.num_sacks; i++) 3304 sp[i] = sp[i+1]; 3305 continue; 3306 } 3307 this_sack++, swalk++; 3308 } 3309 } 3310 3311 static inline void tcp_sack_swap(struct tcp_sack_block *sack1, struct tcp_sack_block *sack2) 3312 { 3313 __u32 tmp; 3314 3315 tmp = sack1->start_seq; 3316 sack1->start_seq = sack2->start_seq; 3317 sack2->start_seq = tmp; 3318 3319 tmp = sack1->end_seq; 3320 sack1->end_seq = sack2->end_seq; 3321 sack2->end_seq = tmp; 3322 } 3323 3324 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq) 3325 { 3326 struct tcp_sock *tp = tcp_sk(sk); 3327 struct tcp_sack_block *sp = &tp->selective_acks[0]; 3328 int cur_sacks = tp->rx_opt.num_sacks; 3329 int this_sack; 3330 3331 if (!cur_sacks) 3332 goto new_sack; 3333 3334 for (this_sack=0; this_sack<cur_sacks; this_sack++, sp++) { 3335 if (tcp_sack_extend(sp, seq, end_seq)) { 3336 /* Rotate this_sack to the first one. */ 3337 for (; this_sack>0; this_sack--, sp--) 3338 tcp_sack_swap(sp, sp-1); 3339 if (cur_sacks > 1) 3340 tcp_sack_maybe_coalesce(tp); 3341 return; 3342 } 3343 } 3344 3345 /* Could not find an adjacent existing SACK, build a new one, 3346 * put it at the front, and shift everyone else down. We 3347 * always know there is at least one SACK present already here. 3348 * 3349 * If the sack array is full, forget about the last one. 3350 */ 3351 if (this_sack >= 4) { 3352 this_sack--; 3353 tp->rx_opt.num_sacks--; 3354 sp--; 3355 } 3356 for (; this_sack > 0; this_sack--, sp--) 3357 *sp = *(sp-1); 3358 3359 new_sack: 3360 /* Build the new head SACK, and we're done. */ 3361 sp->start_seq = seq; 3362 sp->end_seq = end_seq; 3363 tp->rx_opt.num_sacks++; 3364 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok); 3365 } 3366 3367 /* RCV.NXT advances, some SACKs should be eaten. */ 3368 3369 static void tcp_sack_remove(struct tcp_sock *tp) 3370 { 3371 struct tcp_sack_block *sp = &tp->selective_acks[0]; 3372 int num_sacks = tp->rx_opt.num_sacks; 3373 int this_sack; 3374 3375 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */ 3376 if (skb_queue_empty(&tp->out_of_order_queue)) { 3377 tp->rx_opt.num_sacks = 0; 3378 tp->rx_opt.eff_sacks = tp->rx_opt.dsack; 3379 return; 3380 } 3381 3382 for (this_sack = 0; this_sack < num_sacks; ) { 3383 /* Check if the start of the sack is covered by RCV.NXT. */ 3384 if (!before(tp->rcv_nxt, sp->start_seq)) { 3385 int i; 3386 3387 /* RCV.NXT must cover all the block! */ 3388 BUG_TRAP(!before(tp->rcv_nxt, sp->end_seq)); 3389 3390 /* Zap this SACK, by moving forward any other SACKS. */ 3391 for (i=this_sack+1; i < num_sacks; i++) 3392 tp->selective_acks[i-1] = tp->selective_acks[i]; 3393 num_sacks--; 3394 continue; 3395 } 3396 this_sack++; 3397 sp++; 3398 } 3399 if (num_sacks != tp->rx_opt.num_sacks) { 3400 tp->rx_opt.num_sacks = num_sacks; 3401 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok); 3402 } 3403 } 3404 3405 /* This one checks to see if we can put data from the 3406 * out_of_order queue into the receive_queue. 3407 */ 3408 static void tcp_ofo_queue(struct sock *sk) 3409 { 3410 struct tcp_sock *tp = tcp_sk(sk); 3411 __u32 dsack_high = tp->rcv_nxt; 3412 struct sk_buff *skb; 3413 3414 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) { 3415 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) 3416 break; 3417 3418 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) { 3419 __u32 dsack = dsack_high; 3420 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high)) 3421 dsack_high = TCP_SKB_CB(skb)->end_seq; 3422 tcp_dsack_extend(tp, TCP_SKB_CB(skb)->seq, dsack); 3423 } 3424 3425 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) { 3426 SOCK_DEBUG(sk, "ofo packet was already received \n"); 3427 __skb_unlink(skb, &tp->out_of_order_queue); 3428 __kfree_skb(skb); 3429 continue; 3430 } 3431 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n", 3432 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, 3433 TCP_SKB_CB(skb)->end_seq); 3434 3435 __skb_unlink(skb, &tp->out_of_order_queue); 3436 __skb_queue_tail(&sk->sk_receive_queue, skb); 3437 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; 3438 if (tcp_hdr(skb)->fin) 3439 tcp_fin(skb, sk, tcp_hdr(skb)); 3440 } 3441 } 3442 3443 static int tcp_prune_queue(struct sock *sk); 3444 3445 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb) 3446 { 3447 struct tcphdr *th = tcp_hdr(skb); 3448 struct tcp_sock *tp = tcp_sk(sk); 3449 int eaten = -1; 3450 3451 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) 3452 goto drop; 3453 3454 __skb_pull(skb, th->doff*4); 3455 3456 TCP_ECN_accept_cwr(tp, skb); 3457 3458 if (tp->rx_opt.dsack) { 3459 tp->rx_opt.dsack = 0; 3460 tp->rx_opt.eff_sacks = min_t(unsigned int, tp->rx_opt.num_sacks, 3461 4 - tp->rx_opt.tstamp_ok); 3462 } 3463 3464 /* Queue data for delivery to the user. 3465 * Packets in sequence go to the receive queue. 3466 * Out of sequence packets to the out_of_order_queue. 3467 */ 3468 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) { 3469 if (tcp_receive_window(tp) == 0) 3470 goto out_of_window; 3471 3472 /* Ok. In sequence. In window. */ 3473 if (tp->ucopy.task == current && 3474 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len && 3475 sock_owned_by_user(sk) && !tp->urg_data) { 3476 int chunk = min_t(unsigned int, skb->len, 3477 tp->ucopy.len); 3478 3479 __set_current_state(TASK_RUNNING); 3480 3481 local_bh_enable(); 3482 if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) { 3483 tp->ucopy.len -= chunk; 3484 tp->copied_seq += chunk; 3485 eaten = (chunk == skb->len && !th->fin); 3486 tcp_rcv_space_adjust(sk); 3487 } 3488 local_bh_disable(); 3489 } 3490 3491 if (eaten <= 0) { 3492 queue_and_out: 3493 if (eaten < 0 && 3494 (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf || 3495 !sk_stream_rmem_schedule(sk, skb))) { 3496 if (tcp_prune_queue(sk) < 0 || 3497 !sk_stream_rmem_schedule(sk, skb)) 3498 goto drop; 3499 } 3500 sk_stream_set_owner_r(skb, sk); 3501 __skb_queue_tail(&sk->sk_receive_queue, skb); 3502 } 3503 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; 3504 if (skb->len) 3505 tcp_event_data_recv(sk, skb); 3506 if (th->fin) 3507 tcp_fin(skb, sk, th); 3508 3509 if (!skb_queue_empty(&tp->out_of_order_queue)) { 3510 tcp_ofo_queue(sk); 3511 3512 /* RFC2581. 4.2. SHOULD send immediate ACK, when 3513 * gap in queue is filled. 3514 */ 3515 if (skb_queue_empty(&tp->out_of_order_queue)) 3516 inet_csk(sk)->icsk_ack.pingpong = 0; 3517 } 3518 3519 if (tp->rx_opt.num_sacks) 3520 tcp_sack_remove(tp); 3521 3522 tcp_fast_path_check(sk); 3523 3524 if (eaten > 0) 3525 __kfree_skb(skb); 3526 else if (!sock_flag(sk, SOCK_DEAD)) 3527 sk->sk_data_ready(sk, 0); 3528 return; 3529 } 3530 3531 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) { 3532 /* A retransmit, 2nd most common case. Force an immediate ack. */ 3533 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST); 3534 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); 3535 3536 out_of_window: 3537 tcp_enter_quickack_mode(sk); 3538 inet_csk_schedule_ack(sk); 3539 drop: 3540 __kfree_skb(skb); 3541 return; 3542 } 3543 3544 /* Out of window. F.e. zero window probe. */ 3545 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp))) 3546 goto out_of_window; 3547 3548 tcp_enter_quickack_mode(sk); 3549 3550 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 3551 /* Partial packet, seq < rcv_next < end_seq */ 3552 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n", 3553 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, 3554 TCP_SKB_CB(skb)->end_seq); 3555 3556 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, tp->rcv_nxt); 3557 3558 /* If window is closed, drop tail of packet. But after 3559 * remembering D-SACK for its head made in previous line. 3560 */ 3561 if (!tcp_receive_window(tp)) 3562 goto out_of_window; 3563 goto queue_and_out; 3564 } 3565 3566 TCP_ECN_check_ce(tp, skb); 3567 3568 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf || 3569 !sk_stream_rmem_schedule(sk, skb)) { 3570 if (tcp_prune_queue(sk) < 0 || 3571 !sk_stream_rmem_schedule(sk, skb)) 3572 goto drop; 3573 } 3574 3575 /* Disable header prediction. */ 3576 tp->pred_flags = 0; 3577 inet_csk_schedule_ack(sk); 3578 3579 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n", 3580 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); 3581 3582 sk_stream_set_owner_r(skb, sk); 3583 3584 if (!skb_peek(&tp->out_of_order_queue)) { 3585 /* Initial out of order segment, build 1 SACK. */ 3586 if (tp->rx_opt.sack_ok) { 3587 tp->rx_opt.num_sacks = 1; 3588 tp->rx_opt.dsack = 0; 3589 tp->rx_opt.eff_sacks = 1; 3590 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq; 3591 tp->selective_acks[0].end_seq = 3592 TCP_SKB_CB(skb)->end_seq; 3593 } 3594 __skb_queue_head(&tp->out_of_order_queue,skb); 3595 } else { 3596 struct sk_buff *skb1 = tp->out_of_order_queue.prev; 3597 u32 seq = TCP_SKB_CB(skb)->seq; 3598 u32 end_seq = TCP_SKB_CB(skb)->end_seq; 3599 3600 if (seq == TCP_SKB_CB(skb1)->end_seq) { 3601 __skb_append(skb1, skb, &tp->out_of_order_queue); 3602 3603 if (!tp->rx_opt.num_sacks || 3604 tp->selective_acks[0].end_seq != seq) 3605 goto add_sack; 3606 3607 /* Common case: data arrive in order after hole. */ 3608 tp->selective_acks[0].end_seq = end_seq; 3609 return; 3610 } 3611 3612 /* Find place to insert this segment. */ 3613 do { 3614 if (!after(TCP_SKB_CB(skb1)->seq, seq)) 3615 break; 3616 } while ((skb1 = skb1->prev) != 3617 (struct sk_buff*)&tp->out_of_order_queue); 3618 3619 /* Do skb overlap to previous one? */ 3620 if (skb1 != (struct sk_buff*)&tp->out_of_order_queue && 3621 before(seq, TCP_SKB_CB(skb1)->end_seq)) { 3622 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) { 3623 /* All the bits are present. Drop. */ 3624 __kfree_skb(skb); 3625 tcp_dsack_set(tp, seq, end_seq); 3626 goto add_sack; 3627 } 3628 if (after(seq, TCP_SKB_CB(skb1)->seq)) { 3629 /* Partial overlap. */ 3630 tcp_dsack_set(tp, seq, TCP_SKB_CB(skb1)->end_seq); 3631 } else { 3632 skb1 = skb1->prev; 3633 } 3634 } 3635 __skb_insert(skb, skb1, skb1->next, &tp->out_of_order_queue); 3636 3637 /* And clean segments covered by new one as whole. */ 3638 while ((skb1 = skb->next) != 3639 (struct sk_buff*)&tp->out_of_order_queue && 3640 after(end_seq, TCP_SKB_CB(skb1)->seq)) { 3641 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) { 3642 tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, end_seq); 3643 break; 3644 } 3645 __skb_unlink(skb1, &tp->out_of_order_queue); 3646 tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, TCP_SKB_CB(skb1)->end_seq); 3647 __kfree_skb(skb1); 3648 } 3649 3650 add_sack: 3651 if (tp->rx_opt.sack_ok) 3652 tcp_sack_new_ofo_skb(sk, seq, end_seq); 3653 } 3654 } 3655 3656 /* Collapse contiguous sequence of skbs head..tail with 3657 * sequence numbers start..end. 3658 * Segments with FIN/SYN are not collapsed (only because this 3659 * simplifies code) 3660 */ 3661 static void 3662 tcp_collapse(struct sock *sk, struct sk_buff_head *list, 3663 struct sk_buff *head, struct sk_buff *tail, 3664 u32 start, u32 end) 3665 { 3666 struct sk_buff *skb; 3667 3668 /* First, check that queue is collapsible and find 3669 * the point where collapsing can be useful. */ 3670 for (skb = head; skb != tail; ) { 3671 /* No new bits? It is possible on ofo queue. */ 3672 if (!before(start, TCP_SKB_CB(skb)->end_seq)) { 3673 struct sk_buff *next = skb->next; 3674 __skb_unlink(skb, list); 3675 __kfree_skb(skb); 3676 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED); 3677 skb = next; 3678 continue; 3679 } 3680 3681 /* The first skb to collapse is: 3682 * - not SYN/FIN and 3683 * - bloated or contains data before "start" or 3684 * overlaps to the next one. 3685 */ 3686 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin && 3687 (tcp_win_from_space(skb->truesize) > skb->len || 3688 before(TCP_SKB_CB(skb)->seq, start) || 3689 (skb->next != tail && 3690 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq))) 3691 break; 3692 3693 /* Decided to skip this, advance start seq. */ 3694 start = TCP_SKB_CB(skb)->end_seq; 3695 skb = skb->next; 3696 } 3697 if (skb == tail || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin) 3698 return; 3699 3700 while (before(start, end)) { 3701 struct sk_buff *nskb; 3702 int header = skb_headroom(skb); 3703 int copy = SKB_MAX_ORDER(header, 0); 3704 3705 /* Too big header? This can happen with IPv6. */ 3706 if (copy < 0) 3707 return; 3708 if (end-start < copy) 3709 copy = end-start; 3710 nskb = alloc_skb(copy+header, GFP_ATOMIC); 3711 if (!nskb) 3712 return; 3713 3714 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head); 3715 skb_set_network_header(nskb, (skb_network_header(skb) - 3716 skb->head)); 3717 skb_set_transport_header(nskb, (skb_transport_header(skb) - 3718 skb->head)); 3719 skb_reserve(nskb, header); 3720 memcpy(nskb->head, skb->head, header); 3721 memcpy(nskb->cb, skb->cb, sizeof(skb->cb)); 3722 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start; 3723 __skb_insert(nskb, skb->prev, skb, list); 3724 sk_stream_set_owner_r(nskb, sk); 3725 3726 /* Copy data, releasing collapsed skbs. */ 3727 while (copy > 0) { 3728 int offset = start - TCP_SKB_CB(skb)->seq; 3729 int size = TCP_SKB_CB(skb)->end_seq - start; 3730 3731 BUG_ON(offset < 0); 3732 if (size > 0) { 3733 size = min(copy, size); 3734 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size)) 3735 BUG(); 3736 TCP_SKB_CB(nskb)->end_seq += size; 3737 copy -= size; 3738 start += size; 3739 } 3740 if (!before(start, TCP_SKB_CB(skb)->end_seq)) { 3741 struct sk_buff *next = skb->next; 3742 __skb_unlink(skb, list); 3743 __kfree_skb(skb); 3744 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED); 3745 skb = next; 3746 if (skb == tail || 3747 tcp_hdr(skb)->syn || 3748 tcp_hdr(skb)->fin) 3749 return; 3750 } 3751 } 3752 } 3753 } 3754 3755 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs 3756 * and tcp_collapse() them until all the queue is collapsed. 3757 */ 3758 static void tcp_collapse_ofo_queue(struct sock *sk) 3759 { 3760 struct tcp_sock *tp = tcp_sk(sk); 3761 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue); 3762 struct sk_buff *head; 3763 u32 start, end; 3764 3765 if (skb == NULL) 3766 return; 3767 3768 start = TCP_SKB_CB(skb)->seq; 3769 end = TCP_SKB_CB(skb)->end_seq; 3770 head = skb; 3771 3772 for (;;) { 3773 skb = skb->next; 3774 3775 /* Segment is terminated when we see gap or when 3776 * we are at the end of all the queue. */ 3777 if (skb == (struct sk_buff *)&tp->out_of_order_queue || 3778 after(TCP_SKB_CB(skb)->seq, end) || 3779 before(TCP_SKB_CB(skb)->end_seq, start)) { 3780 tcp_collapse(sk, &tp->out_of_order_queue, 3781 head, skb, start, end); 3782 head = skb; 3783 if (skb == (struct sk_buff *)&tp->out_of_order_queue) 3784 break; 3785 /* Start new segment */ 3786 start = TCP_SKB_CB(skb)->seq; 3787 end = TCP_SKB_CB(skb)->end_seq; 3788 } else { 3789 if (before(TCP_SKB_CB(skb)->seq, start)) 3790 start = TCP_SKB_CB(skb)->seq; 3791 if (after(TCP_SKB_CB(skb)->end_seq, end)) 3792 end = TCP_SKB_CB(skb)->end_seq; 3793 } 3794 } 3795 } 3796 3797 /* Reduce allocated memory if we can, trying to get 3798 * the socket within its memory limits again. 3799 * 3800 * Return less than zero if we should start dropping frames 3801 * until the socket owning process reads some of the data 3802 * to stabilize the situation. 3803 */ 3804 static int tcp_prune_queue(struct sock *sk) 3805 { 3806 struct tcp_sock *tp = tcp_sk(sk); 3807 3808 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq); 3809 3810 NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED); 3811 3812 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf) 3813 tcp_clamp_window(sk); 3814 else if (tcp_memory_pressure) 3815 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss); 3816 3817 tcp_collapse_ofo_queue(sk); 3818 tcp_collapse(sk, &sk->sk_receive_queue, 3819 sk->sk_receive_queue.next, 3820 (struct sk_buff*)&sk->sk_receive_queue, 3821 tp->copied_seq, tp->rcv_nxt); 3822 sk_stream_mem_reclaim(sk); 3823 3824 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) 3825 return 0; 3826 3827 /* Collapsing did not help, destructive actions follow. 3828 * This must not ever occur. */ 3829 3830 /* First, purge the out_of_order queue. */ 3831 if (!skb_queue_empty(&tp->out_of_order_queue)) { 3832 NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED); 3833 __skb_queue_purge(&tp->out_of_order_queue); 3834 3835 /* Reset SACK state. A conforming SACK implementation will 3836 * do the same at a timeout based retransmit. When a connection 3837 * is in a sad state like this, we care only about integrity 3838 * of the connection not performance. 3839 */ 3840 if (tp->rx_opt.sack_ok) 3841 tcp_sack_reset(&tp->rx_opt); 3842 sk_stream_mem_reclaim(sk); 3843 } 3844 3845 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) 3846 return 0; 3847 3848 /* If we are really being abused, tell the caller to silently 3849 * drop receive data on the floor. It will get retransmitted 3850 * and hopefully then we'll have sufficient space. 3851 */ 3852 NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED); 3853 3854 /* Massive buffer overcommit. */ 3855 tp->pred_flags = 0; 3856 return -1; 3857 } 3858 3859 3860 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. 3861 * As additional protections, we do not touch cwnd in retransmission phases, 3862 * and if application hit its sndbuf limit recently. 3863 */ 3864 void tcp_cwnd_application_limited(struct sock *sk) 3865 { 3866 struct tcp_sock *tp = tcp_sk(sk); 3867 3868 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && 3869 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { 3870 /* Limited by application or receiver window. */ 3871 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); 3872 u32 win_used = max(tp->snd_cwnd_used, init_win); 3873 if (win_used < tp->snd_cwnd) { 3874 tp->snd_ssthresh = tcp_current_ssthresh(sk); 3875 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1; 3876 } 3877 tp->snd_cwnd_used = 0; 3878 } 3879 tp->snd_cwnd_stamp = tcp_time_stamp; 3880 } 3881 3882 static int tcp_should_expand_sndbuf(struct sock *sk) 3883 { 3884 struct tcp_sock *tp = tcp_sk(sk); 3885 3886 /* If the user specified a specific send buffer setting, do 3887 * not modify it. 3888 */ 3889 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK) 3890 return 0; 3891 3892 /* If we are under global TCP memory pressure, do not expand. */ 3893 if (tcp_memory_pressure) 3894 return 0; 3895 3896 /* If we are under soft global TCP memory pressure, do not expand. */ 3897 if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0]) 3898 return 0; 3899 3900 /* If we filled the congestion window, do not expand. */ 3901 if (tp->packets_out >= tp->snd_cwnd) 3902 return 0; 3903 3904 return 1; 3905 } 3906 3907 /* When incoming ACK allowed to free some skb from write_queue, 3908 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket 3909 * on the exit from tcp input handler. 3910 * 3911 * PROBLEM: sndbuf expansion does not work well with largesend. 3912 */ 3913 static void tcp_new_space(struct sock *sk) 3914 { 3915 struct tcp_sock *tp = tcp_sk(sk); 3916 3917 if (tcp_should_expand_sndbuf(sk)) { 3918 int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) + 3919 MAX_TCP_HEADER + 16 + sizeof(struct sk_buff), 3920 demanded = max_t(unsigned int, tp->snd_cwnd, 3921 tp->reordering + 1); 3922 sndmem *= 2*demanded; 3923 if (sndmem > sk->sk_sndbuf) 3924 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]); 3925 tp->snd_cwnd_stamp = tcp_time_stamp; 3926 } 3927 3928 sk->sk_write_space(sk); 3929 } 3930 3931 static void tcp_check_space(struct sock *sk) 3932 { 3933 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) { 3934 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK); 3935 if (sk->sk_socket && 3936 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) 3937 tcp_new_space(sk); 3938 } 3939 } 3940 3941 static inline void tcp_data_snd_check(struct sock *sk) 3942 { 3943 tcp_push_pending_frames(sk); 3944 tcp_check_space(sk); 3945 } 3946 3947 /* 3948 * Check if sending an ack is needed. 3949 */ 3950 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible) 3951 { 3952 struct tcp_sock *tp = tcp_sk(sk); 3953 3954 /* More than one full frame received... */ 3955 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss 3956 /* ... and right edge of window advances far enough. 3957 * (tcp_recvmsg() will send ACK otherwise). Or... 3958 */ 3959 && __tcp_select_window(sk) >= tp->rcv_wnd) || 3960 /* We ACK each frame or... */ 3961 tcp_in_quickack_mode(sk) || 3962 /* We have out of order data. */ 3963 (ofo_possible && 3964 skb_peek(&tp->out_of_order_queue))) { 3965 /* Then ack it now */ 3966 tcp_send_ack(sk); 3967 } else { 3968 /* Else, send delayed ack. */ 3969 tcp_send_delayed_ack(sk); 3970 } 3971 } 3972 3973 static inline void tcp_ack_snd_check(struct sock *sk) 3974 { 3975 if (!inet_csk_ack_scheduled(sk)) { 3976 /* We sent a data segment already. */ 3977 return; 3978 } 3979 __tcp_ack_snd_check(sk, 1); 3980 } 3981 3982 /* 3983 * This routine is only called when we have urgent data 3984 * signaled. Its the 'slow' part of tcp_urg. It could be 3985 * moved inline now as tcp_urg is only called from one 3986 * place. We handle URGent data wrong. We have to - as 3987 * BSD still doesn't use the correction from RFC961. 3988 * For 1003.1g we should support a new option TCP_STDURG to permit 3989 * either form (or just set the sysctl tcp_stdurg). 3990 */ 3991 3992 static void tcp_check_urg(struct sock * sk, struct tcphdr * th) 3993 { 3994 struct tcp_sock *tp = tcp_sk(sk); 3995 u32 ptr = ntohs(th->urg_ptr); 3996 3997 if (ptr && !sysctl_tcp_stdurg) 3998 ptr--; 3999 ptr += ntohl(th->seq); 4000 4001 /* Ignore urgent data that we've already seen and read. */ 4002 if (after(tp->copied_seq, ptr)) 4003 return; 4004 4005 /* Do not replay urg ptr. 4006 * 4007 * NOTE: interesting situation not covered by specs. 4008 * Misbehaving sender may send urg ptr, pointing to segment, 4009 * which we already have in ofo queue. We are not able to fetch 4010 * such data and will stay in TCP_URG_NOTYET until will be eaten 4011 * by recvmsg(). Seems, we are not obliged to handle such wicked 4012 * situations. But it is worth to think about possibility of some 4013 * DoSes using some hypothetical application level deadlock. 4014 */ 4015 if (before(ptr, tp->rcv_nxt)) 4016 return; 4017 4018 /* Do we already have a newer (or duplicate) urgent pointer? */ 4019 if (tp->urg_data && !after(ptr, tp->urg_seq)) 4020 return; 4021 4022 /* Tell the world about our new urgent pointer. */ 4023 sk_send_sigurg(sk); 4024 4025 /* We may be adding urgent data when the last byte read was 4026 * urgent. To do this requires some care. We cannot just ignore 4027 * tp->copied_seq since we would read the last urgent byte again 4028 * as data, nor can we alter copied_seq until this data arrives 4029 * or we break the semantics of SIOCATMARK (and thus sockatmark()) 4030 * 4031 * NOTE. Double Dutch. Rendering to plain English: author of comment 4032 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB); 4033 * and expect that both A and B disappear from stream. This is _wrong_. 4034 * Though this happens in BSD with high probability, this is occasional. 4035 * Any application relying on this is buggy. Note also, that fix "works" 4036 * only in this artificial test. Insert some normal data between A and B and we will 4037 * decline of BSD again. Verdict: it is better to remove to trap 4038 * buggy users. 4039 */ 4040 if (tp->urg_seq == tp->copied_seq && tp->urg_data && 4041 !sock_flag(sk, SOCK_URGINLINE) && 4042 tp->copied_seq != tp->rcv_nxt) { 4043 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue); 4044 tp->copied_seq++; 4045 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) { 4046 __skb_unlink(skb, &sk->sk_receive_queue); 4047 __kfree_skb(skb); 4048 } 4049 } 4050 4051 tp->urg_data = TCP_URG_NOTYET; 4052 tp->urg_seq = ptr; 4053 4054 /* Disable header prediction. */ 4055 tp->pred_flags = 0; 4056 } 4057 4058 /* This is the 'fast' part of urgent handling. */ 4059 static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th) 4060 { 4061 struct tcp_sock *tp = tcp_sk(sk); 4062 4063 /* Check if we get a new urgent pointer - normally not. */ 4064 if (th->urg) 4065 tcp_check_urg(sk,th); 4066 4067 /* Do we wait for any urgent data? - normally not... */ 4068 if (tp->urg_data == TCP_URG_NOTYET) { 4069 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) - 4070 th->syn; 4071 4072 /* Is the urgent pointer pointing into this packet? */ 4073 if (ptr < skb->len) { 4074 u8 tmp; 4075 if (skb_copy_bits(skb, ptr, &tmp, 1)) 4076 BUG(); 4077 tp->urg_data = TCP_URG_VALID | tmp; 4078 if (!sock_flag(sk, SOCK_DEAD)) 4079 sk->sk_data_ready(sk, 0); 4080 } 4081 } 4082 } 4083 4084 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen) 4085 { 4086 struct tcp_sock *tp = tcp_sk(sk); 4087 int chunk = skb->len - hlen; 4088 int err; 4089 4090 local_bh_enable(); 4091 if (skb_csum_unnecessary(skb)) 4092 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk); 4093 else 4094 err = skb_copy_and_csum_datagram_iovec(skb, hlen, 4095 tp->ucopy.iov); 4096 4097 if (!err) { 4098 tp->ucopy.len -= chunk; 4099 tp->copied_seq += chunk; 4100 tcp_rcv_space_adjust(sk); 4101 } 4102 4103 local_bh_disable(); 4104 return err; 4105 } 4106 4107 static __sum16 __tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb) 4108 { 4109 __sum16 result; 4110 4111 if (sock_owned_by_user(sk)) { 4112 local_bh_enable(); 4113 result = __tcp_checksum_complete(skb); 4114 local_bh_disable(); 4115 } else { 4116 result = __tcp_checksum_complete(skb); 4117 } 4118 return result; 4119 } 4120 4121 static inline int tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb) 4122 { 4123 return !skb_csum_unnecessary(skb) && 4124 __tcp_checksum_complete_user(sk, skb); 4125 } 4126 4127 #ifdef CONFIG_NET_DMA 4128 static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb, int hlen) 4129 { 4130 struct tcp_sock *tp = tcp_sk(sk); 4131 int chunk = skb->len - hlen; 4132 int dma_cookie; 4133 int copied_early = 0; 4134 4135 if (tp->ucopy.wakeup) 4136 return 0; 4137 4138 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list) 4139 tp->ucopy.dma_chan = get_softnet_dma(); 4140 4141 if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) { 4142 4143 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan, 4144 skb, hlen, tp->ucopy.iov, chunk, tp->ucopy.pinned_list); 4145 4146 if (dma_cookie < 0) 4147 goto out; 4148 4149 tp->ucopy.dma_cookie = dma_cookie; 4150 copied_early = 1; 4151 4152 tp->ucopy.len -= chunk; 4153 tp->copied_seq += chunk; 4154 tcp_rcv_space_adjust(sk); 4155 4156 if ((tp->ucopy.len == 0) || 4157 (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) || 4158 (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) { 4159 tp->ucopy.wakeup = 1; 4160 sk->sk_data_ready(sk, 0); 4161 } 4162 } else if (chunk > 0) { 4163 tp->ucopy.wakeup = 1; 4164 sk->sk_data_ready(sk, 0); 4165 } 4166 out: 4167 return copied_early; 4168 } 4169 #endif /* CONFIG_NET_DMA */ 4170 4171 /* 4172 * TCP receive function for the ESTABLISHED state. 4173 * 4174 * It is split into a fast path and a slow path. The fast path is 4175 * disabled when: 4176 * - A zero window was announced from us - zero window probing 4177 * is only handled properly in the slow path. 4178 * - Out of order segments arrived. 4179 * - Urgent data is expected. 4180 * - There is no buffer space left 4181 * - Unexpected TCP flags/window values/header lengths are received 4182 * (detected by checking the TCP header against pred_flags) 4183 * - Data is sent in both directions. Fast path only supports pure senders 4184 * or pure receivers (this means either the sequence number or the ack 4185 * value must stay constant) 4186 * - Unexpected TCP option. 4187 * 4188 * When these conditions are not satisfied it drops into a standard 4189 * receive procedure patterned after RFC793 to handle all cases. 4190 * The first three cases are guaranteed by proper pred_flags setting, 4191 * the rest is checked inline. Fast processing is turned on in 4192 * tcp_data_queue when everything is OK. 4193 */ 4194 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb, 4195 struct tcphdr *th, unsigned len) 4196 { 4197 struct tcp_sock *tp = tcp_sk(sk); 4198 4199 /* 4200 * Header prediction. 4201 * The code loosely follows the one in the famous 4202 * "30 instruction TCP receive" Van Jacobson mail. 4203 * 4204 * Van's trick is to deposit buffers into socket queue 4205 * on a device interrupt, to call tcp_recv function 4206 * on the receive process context and checksum and copy 4207 * the buffer to user space. smart... 4208 * 4209 * Our current scheme is not silly either but we take the 4210 * extra cost of the net_bh soft interrupt processing... 4211 * We do checksum and copy also but from device to kernel. 4212 */ 4213 4214 tp->rx_opt.saw_tstamp = 0; 4215 4216 /* pred_flags is 0xS?10 << 16 + snd_wnd 4217 * if header_prediction is to be made 4218 * 'S' will always be tp->tcp_header_len >> 2 4219 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to 4220 * turn it off (when there are holes in the receive 4221 * space for instance) 4222 * PSH flag is ignored. 4223 */ 4224 4225 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags && 4226 TCP_SKB_CB(skb)->seq == tp->rcv_nxt) { 4227 int tcp_header_len = tp->tcp_header_len; 4228 4229 /* Timestamp header prediction: tcp_header_len 4230 * is automatically equal to th->doff*4 due to pred_flags 4231 * match. 4232 */ 4233 4234 /* Check timestamp */ 4235 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) { 4236 __be32 *ptr = (__be32 *)(th + 1); 4237 4238 /* No? Slow path! */ 4239 if (*ptr != htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) 4240 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) 4241 goto slow_path; 4242 4243 tp->rx_opt.saw_tstamp = 1; 4244 ++ptr; 4245 tp->rx_opt.rcv_tsval = ntohl(*ptr); 4246 ++ptr; 4247 tp->rx_opt.rcv_tsecr = ntohl(*ptr); 4248 4249 /* If PAWS failed, check it more carefully in slow path */ 4250 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0) 4251 goto slow_path; 4252 4253 /* DO NOT update ts_recent here, if checksum fails 4254 * and timestamp was corrupted part, it will result 4255 * in a hung connection since we will drop all 4256 * future packets due to the PAWS test. 4257 */ 4258 } 4259 4260 if (len <= tcp_header_len) { 4261 /* Bulk data transfer: sender */ 4262 if (len == tcp_header_len) { 4263 /* Predicted packet is in window by definition. 4264 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 4265 * Hence, check seq<=rcv_wup reduces to: 4266 */ 4267 if (tcp_header_len == 4268 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && 4269 tp->rcv_nxt == tp->rcv_wup) 4270 tcp_store_ts_recent(tp); 4271 4272 /* We know that such packets are checksummed 4273 * on entry. 4274 */ 4275 tcp_ack(sk, skb, 0); 4276 __kfree_skb(skb); 4277 tcp_data_snd_check(sk); 4278 return 0; 4279 } else { /* Header too small */ 4280 TCP_INC_STATS_BH(TCP_MIB_INERRS); 4281 goto discard; 4282 } 4283 } else { 4284 int eaten = 0; 4285 int copied_early = 0; 4286 4287 if (tp->copied_seq == tp->rcv_nxt && 4288 len - tcp_header_len <= tp->ucopy.len) { 4289 #ifdef CONFIG_NET_DMA 4290 if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) { 4291 copied_early = 1; 4292 eaten = 1; 4293 } 4294 #endif 4295 if (tp->ucopy.task == current && sock_owned_by_user(sk) && !copied_early) { 4296 __set_current_state(TASK_RUNNING); 4297 4298 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len)) 4299 eaten = 1; 4300 } 4301 if (eaten) { 4302 /* Predicted packet is in window by definition. 4303 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 4304 * Hence, check seq<=rcv_wup reduces to: 4305 */ 4306 if (tcp_header_len == 4307 (sizeof(struct tcphdr) + 4308 TCPOLEN_TSTAMP_ALIGNED) && 4309 tp->rcv_nxt == tp->rcv_wup) 4310 tcp_store_ts_recent(tp); 4311 4312 tcp_rcv_rtt_measure_ts(sk, skb); 4313 4314 __skb_pull(skb, tcp_header_len); 4315 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; 4316 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER); 4317 } 4318 if (copied_early) 4319 tcp_cleanup_rbuf(sk, skb->len); 4320 } 4321 if (!eaten) { 4322 if (tcp_checksum_complete_user(sk, skb)) 4323 goto csum_error; 4324 4325 /* Predicted packet is in window by definition. 4326 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 4327 * Hence, check seq<=rcv_wup reduces to: 4328 */ 4329 if (tcp_header_len == 4330 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && 4331 tp->rcv_nxt == tp->rcv_wup) 4332 tcp_store_ts_recent(tp); 4333 4334 tcp_rcv_rtt_measure_ts(sk, skb); 4335 4336 if ((int)skb->truesize > sk->sk_forward_alloc) 4337 goto step5; 4338 4339 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS); 4340 4341 /* Bulk data transfer: receiver */ 4342 __skb_pull(skb,tcp_header_len); 4343 __skb_queue_tail(&sk->sk_receive_queue, skb); 4344 sk_stream_set_owner_r(skb, sk); 4345 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; 4346 } 4347 4348 tcp_event_data_recv(sk, skb); 4349 4350 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) { 4351 /* Well, only one small jumplet in fast path... */ 4352 tcp_ack(sk, skb, FLAG_DATA); 4353 tcp_data_snd_check(sk); 4354 if (!inet_csk_ack_scheduled(sk)) 4355 goto no_ack; 4356 } 4357 4358 __tcp_ack_snd_check(sk, 0); 4359 no_ack: 4360 #ifdef CONFIG_NET_DMA 4361 if (copied_early) 4362 __skb_queue_tail(&sk->sk_async_wait_queue, skb); 4363 else 4364 #endif 4365 if (eaten) 4366 __kfree_skb(skb); 4367 else 4368 sk->sk_data_ready(sk, 0); 4369 return 0; 4370 } 4371 } 4372 4373 slow_path: 4374 if (len < (th->doff<<2) || tcp_checksum_complete_user(sk, skb)) 4375 goto csum_error; 4376 4377 /* 4378 * RFC1323: H1. Apply PAWS check first. 4379 */ 4380 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp && 4381 tcp_paws_discard(sk, skb)) { 4382 if (!th->rst) { 4383 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED); 4384 tcp_send_dupack(sk, skb); 4385 goto discard; 4386 } 4387 /* Resets are accepted even if PAWS failed. 4388 4389 ts_recent update must be made after we are sure 4390 that the packet is in window. 4391 */ 4392 } 4393 4394 /* 4395 * Standard slow path. 4396 */ 4397 4398 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) { 4399 /* RFC793, page 37: "In all states except SYN-SENT, all reset 4400 * (RST) segments are validated by checking their SEQ-fields." 4401 * And page 69: "If an incoming segment is not acceptable, 4402 * an acknowledgment should be sent in reply (unless the RST bit 4403 * is set, if so drop the segment and return)". 4404 */ 4405 if (!th->rst) 4406 tcp_send_dupack(sk, skb); 4407 goto discard; 4408 } 4409 4410 if (th->rst) { 4411 tcp_reset(sk); 4412 goto discard; 4413 } 4414 4415 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq); 4416 4417 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 4418 TCP_INC_STATS_BH(TCP_MIB_INERRS); 4419 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN); 4420 tcp_reset(sk); 4421 return 1; 4422 } 4423 4424 step5: 4425 if (th->ack) 4426 tcp_ack(sk, skb, FLAG_SLOWPATH); 4427 4428 tcp_rcv_rtt_measure_ts(sk, skb); 4429 4430 /* Process urgent data. */ 4431 tcp_urg(sk, skb, th); 4432 4433 /* step 7: process the segment text */ 4434 tcp_data_queue(sk, skb); 4435 4436 tcp_data_snd_check(sk); 4437 tcp_ack_snd_check(sk); 4438 return 0; 4439 4440 csum_error: 4441 TCP_INC_STATS_BH(TCP_MIB_INERRS); 4442 4443 discard: 4444 __kfree_skb(skb); 4445 return 0; 4446 } 4447 4448 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb, 4449 struct tcphdr *th, unsigned len) 4450 { 4451 struct tcp_sock *tp = tcp_sk(sk); 4452 struct inet_connection_sock *icsk = inet_csk(sk); 4453 int saved_clamp = tp->rx_opt.mss_clamp; 4454 4455 tcp_parse_options(skb, &tp->rx_opt, 0); 4456 4457 if (th->ack) { 4458 /* rfc793: 4459 * "If the state is SYN-SENT then 4460 * first check the ACK bit 4461 * If the ACK bit is set 4462 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send 4463 * a reset (unless the RST bit is set, if so drop 4464 * the segment and return)" 4465 * 4466 * We do not send data with SYN, so that RFC-correct 4467 * test reduces to: 4468 */ 4469 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt) 4470 goto reset_and_undo; 4471 4472 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 4473 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp, 4474 tcp_time_stamp)) { 4475 NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED); 4476 goto reset_and_undo; 4477 } 4478 4479 /* Now ACK is acceptable. 4480 * 4481 * "If the RST bit is set 4482 * If the ACK was acceptable then signal the user "error: 4483 * connection reset", drop the segment, enter CLOSED state, 4484 * delete TCB, and return." 4485 */ 4486 4487 if (th->rst) { 4488 tcp_reset(sk); 4489 goto discard; 4490 } 4491 4492 /* rfc793: 4493 * "fifth, if neither of the SYN or RST bits is set then 4494 * drop the segment and return." 4495 * 4496 * See note below! 4497 * --ANK(990513) 4498 */ 4499 if (!th->syn) 4500 goto discard_and_undo; 4501 4502 /* rfc793: 4503 * "If the SYN bit is on ... 4504 * are acceptable then ... 4505 * (our SYN has been ACKed), change the connection 4506 * state to ESTABLISHED..." 4507 */ 4508 4509 TCP_ECN_rcv_synack(tp, th); 4510 4511 tp->snd_wl1 = TCP_SKB_CB(skb)->seq; 4512 tcp_ack(sk, skb, FLAG_SLOWPATH); 4513 4514 /* Ok.. it's good. Set up sequence numbers and 4515 * move to established. 4516 */ 4517 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; 4518 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; 4519 4520 /* RFC1323: The window in SYN & SYN/ACK segments is 4521 * never scaled. 4522 */ 4523 tp->snd_wnd = ntohs(th->window); 4524 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(skb)->seq); 4525 4526 if (!tp->rx_opt.wscale_ok) { 4527 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0; 4528 tp->window_clamp = min(tp->window_clamp, 65535U); 4529 } 4530 4531 if (tp->rx_opt.saw_tstamp) { 4532 tp->rx_opt.tstamp_ok = 1; 4533 tp->tcp_header_len = 4534 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 4535 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; 4536 tcp_store_ts_recent(tp); 4537 } else { 4538 tp->tcp_header_len = sizeof(struct tcphdr); 4539 } 4540 4541 if (tp->rx_opt.sack_ok && sysctl_tcp_fack) 4542 tp->rx_opt.sack_ok |= 2; 4543 4544 tcp_mtup_init(sk); 4545 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 4546 tcp_initialize_rcv_mss(sk); 4547 4548 /* Remember, tcp_poll() does not lock socket! 4549 * Change state from SYN-SENT only after copied_seq 4550 * is initialized. */ 4551 tp->copied_seq = tp->rcv_nxt; 4552 smp_mb(); 4553 tcp_set_state(sk, TCP_ESTABLISHED); 4554 4555 security_inet_conn_established(sk, skb); 4556 4557 /* Make sure socket is routed, for correct metrics. */ 4558 icsk->icsk_af_ops->rebuild_header(sk); 4559 4560 tcp_init_metrics(sk); 4561 4562 tcp_init_congestion_control(sk); 4563 4564 /* Prevent spurious tcp_cwnd_restart() on first data 4565 * packet. 4566 */ 4567 tp->lsndtime = tcp_time_stamp; 4568 4569 tcp_init_buffer_space(sk); 4570 4571 if (sock_flag(sk, SOCK_KEEPOPEN)) 4572 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp)); 4573 4574 if (!tp->rx_opt.snd_wscale) 4575 __tcp_fast_path_on(tp, tp->snd_wnd); 4576 else 4577 tp->pred_flags = 0; 4578 4579 if (!sock_flag(sk, SOCK_DEAD)) { 4580 sk->sk_state_change(sk); 4581 sk_wake_async(sk, 0, POLL_OUT); 4582 } 4583 4584 if (sk->sk_write_pending || 4585 icsk->icsk_accept_queue.rskq_defer_accept || 4586 icsk->icsk_ack.pingpong) { 4587 /* Save one ACK. Data will be ready after 4588 * several ticks, if write_pending is set. 4589 * 4590 * It may be deleted, but with this feature tcpdumps 4591 * look so _wonderfully_ clever, that I was not able 4592 * to stand against the temptation 8) --ANK 4593 */ 4594 inet_csk_schedule_ack(sk); 4595 icsk->icsk_ack.lrcvtime = tcp_time_stamp; 4596 icsk->icsk_ack.ato = TCP_ATO_MIN; 4597 tcp_incr_quickack(sk); 4598 tcp_enter_quickack_mode(sk); 4599 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 4600 TCP_DELACK_MAX, TCP_RTO_MAX); 4601 4602 discard: 4603 __kfree_skb(skb); 4604 return 0; 4605 } else { 4606 tcp_send_ack(sk); 4607 } 4608 return -1; 4609 } 4610 4611 /* No ACK in the segment */ 4612 4613 if (th->rst) { 4614 /* rfc793: 4615 * "If the RST bit is set 4616 * 4617 * Otherwise (no ACK) drop the segment and return." 4618 */ 4619 4620 goto discard_and_undo; 4621 } 4622 4623 /* PAWS check. */ 4624 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && tcp_paws_check(&tp->rx_opt, 0)) 4625 goto discard_and_undo; 4626 4627 if (th->syn) { 4628 /* We see SYN without ACK. It is attempt of 4629 * simultaneous connect with crossed SYNs. 4630 * Particularly, it can be connect to self. 4631 */ 4632 tcp_set_state(sk, TCP_SYN_RECV); 4633 4634 if (tp->rx_opt.saw_tstamp) { 4635 tp->rx_opt.tstamp_ok = 1; 4636 tcp_store_ts_recent(tp); 4637 tp->tcp_header_len = 4638 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 4639 } else { 4640 tp->tcp_header_len = sizeof(struct tcphdr); 4641 } 4642 4643 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; 4644 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; 4645 4646 /* RFC1323: The window in SYN & SYN/ACK segments is 4647 * never scaled. 4648 */ 4649 tp->snd_wnd = ntohs(th->window); 4650 tp->snd_wl1 = TCP_SKB_CB(skb)->seq; 4651 tp->max_window = tp->snd_wnd; 4652 4653 TCP_ECN_rcv_syn(tp, th); 4654 4655 tcp_mtup_init(sk); 4656 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 4657 tcp_initialize_rcv_mss(sk); 4658 4659 4660 tcp_send_synack(sk); 4661 #if 0 4662 /* Note, we could accept data and URG from this segment. 4663 * There are no obstacles to make this. 4664 * 4665 * However, if we ignore data in ACKless segments sometimes, 4666 * we have no reasons to accept it sometimes. 4667 * Also, seems the code doing it in step6 of tcp_rcv_state_process 4668 * is not flawless. So, discard packet for sanity. 4669 * Uncomment this return to process the data. 4670 */ 4671 return -1; 4672 #else 4673 goto discard; 4674 #endif 4675 } 4676 /* "fifth, if neither of the SYN or RST bits is set then 4677 * drop the segment and return." 4678 */ 4679 4680 discard_and_undo: 4681 tcp_clear_options(&tp->rx_opt); 4682 tp->rx_opt.mss_clamp = saved_clamp; 4683 goto discard; 4684 4685 reset_and_undo: 4686 tcp_clear_options(&tp->rx_opt); 4687 tp->rx_opt.mss_clamp = saved_clamp; 4688 return 1; 4689 } 4690 4691 4692 /* 4693 * This function implements the receiving procedure of RFC 793 for 4694 * all states except ESTABLISHED and TIME_WAIT. 4695 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be 4696 * address independent. 4697 */ 4698 4699 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb, 4700 struct tcphdr *th, unsigned len) 4701 { 4702 struct tcp_sock *tp = tcp_sk(sk); 4703 struct inet_connection_sock *icsk = inet_csk(sk); 4704 int queued = 0; 4705 4706 tp->rx_opt.saw_tstamp = 0; 4707 4708 switch (sk->sk_state) { 4709 case TCP_CLOSE: 4710 goto discard; 4711 4712 case TCP_LISTEN: 4713 if (th->ack) 4714 return 1; 4715 4716 if (th->rst) 4717 goto discard; 4718 4719 if (th->syn) { 4720 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0) 4721 return 1; 4722 4723 /* Now we have several options: In theory there is 4724 * nothing else in the frame. KA9Q has an option to 4725 * send data with the syn, BSD accepts data with the 4726 * syn up to the [to be] advertised window and 4727 * Solaris 2.1 gives you a protocol error. For now 4728 * we just ignore it, that fits the spec precisely 4729 * and avoids incompatibilities. It would be nice in 4730 * future to drop through and process the data. 4731 * 4732 * Now that TTCP is starting to be used we ought to 4733 * queue this data. 4734 * But, this leaves one open to an easy denial of 4735 * service attack, and SYN cookies can't defend 4736 * against this problem. So, we drop the data 4737 * in the interest of security over speed unless 4738 * it's still in use. 4739 */ 4740 kfree_skb(skb); 4741 return 0; 4742 } 4743 goto discard; 4744 4745 case TCP_SYN_SENT: 4746 queued = tcp_rcv_synsent_state_process(sk, skb, th, len); 4747 if (queued >= 0) 4748 return queued; 4749 4750 /* Do step6 onward by hand. */ 4751 tcp_urg(sk, skb, th); 4752 __kfree_skb(skb); 4753 tcp_data_snd_check(sk); 4754 return 0; 4755 } 4756 4757 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp && 4758 tcp_paws_discard(sk, skb)) { 4759 if (!th->rst) { 4760 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED); 4761 tcp_send_dupack(sk, skb); 4762 goto discard; 4763 } 4764 /* Reset is accepted even if it did not pass PAWS. */ 4765 } 4766 4767 /* step 1: check sequence number */ 4768 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) { 4769 if (!th->rst) 4770 tcp_send_dupack(sk, skb); 4771 goto discard; 4772 } 4773 4774 /* step 2: check RST bit */ 4775 if (th->rst) { 4776 tcp_reset(sk); 4777 goto discard; 4778 } 4779 4780 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq); 4781 4782 /* step 3: check security and precedence [ignored] */ 4783 4784 /* step 4: 4785 * 4786 * Check for a SYN in window. 4787 */ 4788 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 4789 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN); 4790 tcp_reset(sk); 4791 return 1; 4792 } 4793 4794 /* step 5: check the ACK field */ 4795 if (th->ack) { 4796 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH); 4797 4798 switch (sk->sk_state) { 4799 case TCP_SYN_RECV: 4800 if (acceptable) { 4801 tp->copied_seq = tp->rcv_nxt; 4802 smp_mb(); 4803 tcp_set_state(sk, TCP_ESTABLISHED); 4804 sk->sk_state_change(sk); 4805 4806 /* Note, that this wakeup is only for marginal 4807 * crossed SYN case. Passively open sockets 4808 * are not waked up, because sk->sk_sleep == 4809 * NULL and sk->sk_socket == NULL. 4810 */ 4811 if (sk->sk_socket) { 4812 sk_wake_async(sk,0,POLL_OUT); 4813 } 4814 4815 tp->snd_una = TCP_SKB_CB(skb)->ack_seq; 4816 tp->snd_wnd = ntohs(th->window) << 4817 tp->rx_opt.snd_wscale; 4818 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, 4819 TCP_SKB_CB(skb)->seq); 4820 4821 /* tcp_ack considers this ACK as duplicate 4822 * and does not calculate rtt. 4823 * Fix it at least with timestamps. 4824 */ 4825 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 4826 !tp->srtt) 4827 tcp_ack_saw_tstamp(sk, 0); 4828 4829 if (tp->rx_opt.tstamp_ok) 4830 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; 4831 4832 /* Make sure socket is routed, for 4833 * correct metrics. 4834 */ 4835 icsk->icsk_af_ops->rebuild_header(sk); 4836 4837 tcp_init_metrics(sk); 4838 4839 tcp_init_congestion_control(sk); 4840 4841 /* Prevent spurious tcp_cwnd_restart() on 4842 * first data packet. 4843 */ 4844 tp->lsndtime = tcp_time_stamp; 4845 4846 tcp_mtup_init(sk); 4847 tcp_initialize_rcv_mss(sk); 4848 tcp_init_buffer_space(sk); 4849 tcp_fast_path_on(tp); 4850 } else { 4851 return 1; 4852 } 4853 break; 4854 4855 case TCP_FIN_WAIT1: 4856 if (tp->snd_una == tp->write_seq) { 4857 tcp_set_state(sk, TCP_FIN_WAIT2); 4858 sk->sk_shutdown |= SEND_SHUTDOWN; 4859 dst_confirm(sk->sk_dst_cache); 4860 4861 if (!sock_flag(sk, SOCK_DEAD)) 4862 /* Wake up lingering close() */ 4863 sk->sk_state_change(sk); 4864 else { 4865 int tmo; 4866 4867 if (tp->linger2 < 0 || 4868 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 4869 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) { 4870 tcp_done(sk); 4871 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA); 4872 return 1; 4873 } 4874 4875 tmo = tcp_fin_time(sk); 4876 if (tmo > TCP_TIMEWAIT_LEN) { 4877 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN); 4878 } else if (th->fin || sock_owned_by_user(sk)) { 4879 /* Bad case. We could lose such FIN otherwise. 4880 * It is not a big problem, but it looks confusing 4881 * and not so rare event. We still can lose it now, 4882 * if it spins in bh_lock_sock(), but it is really 4883 * marginal case. 4884 */ 4885 inet_csk_reset_keepalive_timer(sk, tmo); 4886 } else { 4887 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo); 4888 goto discard; 4889 } 4890 } 4891 } 4892 break; 4893 4894 case TCP_CLOSING: 4895 if (tp->snd_una == tp->write_seq) { 4896 tcp_time_wait(sk, TCP_TIME_WAIT, 0); 4897 goto discard; 4898 } 4899 break; 4900 4901 case TCP_LAST_ACK: 4902 if (tp->snd_una == tp->write_seq) { 4903 tcp_update_metrics(sk); 4904 tcp_done(sk); 4905 goto discard; 4906 } 4907 break; 4908 } 4909 } else 4910 goto discard; 4911 4912 /* step 6: check the URG bit */ 4913 tcp_urg(sk, skb, th); 4914 4915 /* step 7: process the segment text */ 4916 switch (sk->sk_state) { 4917 case TCP_CLOSE_WAIT: 4918 case TCP_CLOSING: 4919 case TCP_LAST_ACK: 4920 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) 4921 break; 4922 case TCP_FIN_WAIT1: 4923 case TCP_FIN_WAIT2: 4924 /* RFC 793 says to queue data in these states, 4925 * RFC 1122 says we MUST send a reset. 4926 * BSD 4.4 also does reset. 4927 */ 4928 if (sk->sk_shutdown & RCV_SHUTDOWN) { 4929 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 4930 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) { 4931 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA); 4932 tcp_reset(sk); 4933 return 1; 4934 } 4935 } 4936 /* Fall through */ 4937 case TCP_ESTABLISHED: 4938 tcp_data_queue(sk, skb); 4939 queued = 1; 4940 break; 4941 } 4942 4943 /* tcp_data could move socket to TIME-WAIT */ 4944 if (sk->sk_state != TCP_CLOSE) { 4945 tcp_data_snd_check(sk); 4946 tcp_ack_snd_check(sk); 4947 } 4948 4949 if (!queued) { 4950 discard: 4951 __kfree_skb(skb); 4952 } 4953 return 0; 4954 } 4955 4956 EXPORT_SYMBOL(sysctl_tcp_ecn); 4957 EXPORT_SYMBOL(sysctl_tcp_reordering); 4958 EXPORT_SYMBOL(tcp_parse_options); 4959 EXPORT_SYMBOL(tcp_rcv_established); 4960 EXPORT_SYMBOL(tcp_rcv_state_process); 4961 EXPORT_SYMBOL(tcp_initialize_rcv_mss); 4962