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 = 2; 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 #define FLAG_NONHEAD_RETRANS_ACKED 0x1000 /* Non-head rexmitted data was ACKed */ 108 109 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED) 110 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED) 111 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE) 112 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED) 113 #define FLAG_ANY_PROGRESS (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED) 114 115 #define IsSackFrto() (sysctl_tcp_frto == 0x2) 116 117 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH) 118 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH)) 119 120 /* Adapt the MSS value used to make delayed ack decision to the 121 * real world. 122 */ 123 static void tcp_measure_rcv_mss(struct sock *sk, 124 const struct sk_buff *skb) 125 { 126 struct inet_connection_sock *icsk = inet_csk(sk); 127 const unsigned int lss = icsk->icsk_ack.last_seg_size; 128 unsigned int len; 129 130 icsk->icsk_ack.last_seg_size = 0; 131 132 /* skb->len may jitter because of SACKs, even if peer 133 * sends good full-sized frames. 134 */ 135 len = skb_shinfo(skb)->gso_size ?: skb->len; 136 if (len >= icsk->icsk_ack.rcv_mss) { 137 icsk->icsk_ack.rcv_mss = len; 138 } else { 139 /* Otherwise, we make more careful check taking into account, 140 * that SACKs block is variable. 141 * 142 * "len" is invariant segment length, including TCP header. 143 */ 144 len += skb->data - skb_transport_header(skb); 145 if (len >= TCP_MIN_RCVMSS + sizeof(struct tcphdr) || 146 /* If PSH is not set, packet should be 147 * full sized, provided peer TCP is not badly broken. 148 * This observation (if it is correct 8)) allows 149 * to handle super-low mtu links fairly. 150 */ 151 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) && 152 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) { 153 /* Subtract also invariant (if peer is RFC compliant), 154 * tcp header plus fixed timestamp option length. 155 * Resulting "len" is MSS free of SACK jitter. 156 */ 157 len -= tcp_sk(sk)->tcp_header_len; 158 icsk->icsk_ack.last_seg_size = len; 159 if (len == lss) { 160 icsk->icsk_ack.rcv_mss = len; 161 return; 162 } 163 } 164 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED) 165 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2; 166 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED; 167 } 168 } 169 170 static void tcp_incr_quickack(struct sock *sk) 171 { 172 struct inet_connection_sock *icsk = inet_csk(sk); 173 unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss); 174 175 if (quickacks==0) 176 quickacks=2; 177 if (quickacks > icsk->icsk_ack.quick) 178 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS); 179 } 180 181 void tcp_enter_quickack_mode(struct sock *sk) 182 { 183 struct inet_connection_sock *icsk = inet_csk(sk); 184 tcp_incr_quickack(sk); 185 icsk->icsk_ack.pingpong = 0; 186 icsk->icsk_ack.ato = TCP_ATO_MIN; 187 } 188 189 /* Send ACKs quickly, if "quick" count is not exhausted 190 * and the session is not interactive. 191 */ 192 193 static inline int tcp_in_quickack_mode(const struct sock *sk) 194 { 195 const struct inet_connection_sock *icsk = inet_csk(sk); 196 return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong; 197 } 198 199 static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp) 200 { 201 if (tp->ecn_flags&TCP_ECN_OK) 202 tp->ecn_flags |= TCP_ECN_QUEUE_CWR; 203 } 204 205 static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, struct sk_buff *skb) 206 { 207 if (tcp_hdr(skb)->cwr) 208 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR; 209 } 210 211 static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp) 212 { 213 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR; 214 } 215 216 static inline void TCP_ECN_check_ce(struct tcp_sock *tp, struct sk_buff *skb) 217 { 218 if (tp->ecn_flags&TCP_ECN_OK) { 219 if (INET_ECN_is_ce(TCP_SKB_CB(skb)->flags)) 220 tp->ecn_flags |= TCP_ECN_DEMAND_CWR; 221 /* Funny extension: if ECT is not set on a segment, 222 * it is surely retransmit. It is not in ECN RFC, 223 * but Linux follows this rule. */ 224 else if (INET_ECN_is_not_ect((TCP_SKB_CB(skb)->flags))) 225 tcp_enter_quickack_mode((struct sock *)tp); 226 } 227 } 228 229 static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, struct tcphdr *th) 230 { 231 if ((tp->ecn_flags&TCP_ECN_OK) && (!th->ece || th->cwr)) 232 tp->ecn_flags &= ~TCP_ECN_OK; 233 } 234 235 static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, struct tcphdr *th) 236 { 237 if ((tp->ecn_flags&TCP_ECN_OK) && (!th->ece || !th->cwr)) 238 tp->ecn_flags &= ~TCP_ECN_OK; 239 } 240 241 static inline int TCP_ECN_rcv_ecn_echo(struct tcp_sock *tp, struct tcphdr *th) 242 { 243 if (th->ece && !th->syn && (tp->ecn_flags&TCP_ECN_OK)) 244 return 1; 245 return 0; 246 } 247 248 /* Buffer size and advertised window tuning. 249 * 250 * 1. Tuning sk->sk_sndbuf, when connection enters established state. 251 */ 252 253 static void tcp_fixup_sndbuf(struct sock *sk) 254 { 255 int sndmem = tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER + 16 + 256 sizeof(struct sk_buff); 257 258 if (sk->sk_sndbuf < 3 * sndmem) 259 sk->sk_sndbuf = min(3 * sndmem, sysctl_tcp_wmem[2]); 260 } 261 262 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh) 263 * 264 * All tcp_full_space() is split to two parts: "network" buffer, allocated 265 * forward and advertised in receiver window (tp->rcv_wnd) and 266 * "application buffer", required to isolate scheduling/application 267 * latencies from network. 268 * window_clamp is maximal advertised window. It can be less than 269 * tcp_full_space(), in this case tcp_full_space() - window_clamp 270 * is reserved for "application" buffer. The less window_clamp is 271 * the smoother our behaviour from viewpoint of network, but the lower 272 * throughput and the higher sensitivity of the connection to losses. 8) 273 * 274 * rcv_ssthresh is more strict window_clamp used at "slow start" 275 * phase to predict further behaviour of this connection. 276 * It is used for two goals: 277 * - to enforce header prediction at sender, even when application 278 * requires some significant "application buffer". It is check #1. 279 * - to prevent pruning of receive queue because of misprediction 280 * of receiver window. Check #2. 281 * 282 * The scheme does not work when sender sends good segments opening 283 * window and then starts to feed us spaghetti. But it should work 284 * in common situations. Otherwise, we have to rely on queue collapsing. 285 */ 286 287 /* Slow part of check#2. */ 288 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb) 289 { 290 struct tcp_sock *tp = tcp_sk(sk); 291 /* Optimize this! */ 292 int truesize = tcp_win_from_space(skb->truesize)/2; 293 int window = tcp_win_from_space(sysctl_tcp_rmem[2])/2; 294 295 while (tp->rcv_ssthresh <= window) { 296 if (truesize <= skb->len) 297 return 2 * inet_csk(sk)->icsk_ack.rcv_mss; 298 299 truesize >>= 1; 300 window >>= 1; 301 } 302 return 0; 303 } 304 305 static void tcp_grow_window(struct sock *sk, 306 struct sk_buff *skb) 307 { 308 struct tcp_sock *tp = tcp_sk(sk); 309 310 /* Check #1 */ 311 if (tp->rcv_ssthresh < tp->window_clamp && 312 (int)tp->rcv_ssthresh < tcp_space(sk) && 313 !tcp_memory_pressure) { 314 int incr; 315 316 /* Check #2. Increase window, if skb with such overhead 317 * will fit to rcvbuf in future. 318 */ 319 if (tcp_win_from_space(skb->truesize) <= skb->len) 320 incr = 2*tp->advmss; 321 else 322 incr = __tcp_grow_window(sk, skb); 323 324 if (incr) { 325 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr, tp->window_clamp); 326 inet_csk(sk)->icsk_ack.quick |= 1; 327 } 328 } 329 } 330 331 /* 3. Tuning rcvbuf, when connection enters established state. */ 332 333 static void tcp_fixup_rcvbuf(struct sock *sk) 334 { 335 struct tcp_sock *tp = tcp_sk(sk); 336 int rcvmem = tp->advmss + MAX_TCP_HEADER + 16 + sizeof(struct sk_buff); 337 338 /* Try to select rcvbuf so that 4 mss-sized segments 339 * will fit to window and corresponding skbs will fit to our rcvbuf. 340 * (was 3; 4 is minimum to allow fast retransmit to work.) 341 */ 342 while (tcp_win_from_space(rcvmem) < tp->advmss) 343 rcvmem += 128; 344 if (sk->sk_rcvbuf < 4 * rcvmem) 345 sk->sk_rcvbuf = min(4 * rcvmem, sysctl_tcp_rmem[2]); 346 } 347 348 /* 4. Try to fixup all. It is made immediately after connection enters 349 * established state. 350 */ 351 static void tcp_init_buffer_space(struct sock *sk) 352 { 353 struct tcp_sock *tp = tcp_sk(sk); 354 int maxwin; 355 356 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) 357 tcp_fixup_rcvbuf(sk); 358 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK)) 359 tcp_fixup_sndbuf(sk); 360 361 tp->rcvq_space.space = tp->rcv_wnd; 362 363 maxwin = tcp_full_space(sk); 364 365 if (tp->window_clamp >= maxwin) { 366 tp->window_clamp = maxwin; 367 368 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss) 369 tp->window_clamp = max(maxwin - 370 (maxwin >> sysctl_tcp_app_win), 371 4 * tp->advmss); 372 } 373 374 /* Force reservation of one segment. */ 375 if (sysctl_tcp_app_win && 376 tp->window_clamp > 2 * tp->advmss && 377 tp->window_clamp + tp->advmss > maxwin) 378 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss); 379 380 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp); 381 tp->snd_cwnd_stamp = tcp_time_stamp; 382 } 383 384 /* 5. Recalculate window clamp after socket hit its memory bounds. */ 385 static void tcp_clamp_window(struct sock *sk) 386 { 387 struct tcp_sock *tp = tcp_sk(sk); 388 struct inet_connection_sock *icsk = inet_csk(sk); 389 390 icsk->icsk_ack.quick = 0; 391 392 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] && 393 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) && 394 !tcp_memory_pressure && 395 atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) { 396 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc), 397 sysctl_tcp_rmem[2]); 398 } 399 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf) 400 tp->rcv_ssthresh = min(tp->window_clamp, 2U*tp->advmss); 401 } 402 403 404 /* Initialize RCV_MSS value. 405 * RCV_MSS is an our guess about MSS used by the peer. 406 * We haven't any direct information about the MSS. 407 * It's better to underestimate the RCV_MSS rather than overestimate. 408 * Overestimations make us ACKing less frequently than needed. 409 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss(). 410 */ 411 void tcp_initialize_rcv_mss(struct sock *sk) 412 { 413 struct tcp_sock *tp = tcp_sk(sk); 414 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache); 415 416 hint = min(hint, tp->rcv_wnd/2); 417 hint = min(hint, TCP_MIN_RCVMSS); 418 hint = max(hint, TCP_MIN_MSS); 419 420 inet_csk(sk)->icsk_ack.rcv_mss = hint; 421 } 422 423 /* Receiver "autotuning" code. 424 * 425 * The algorithm for RTT estimation w/o timestamps is based on 426 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL. 427 * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps> 428 * 429 * More detail on this code can be found at 430 * <http://www.psc.edu/~jheffner/senior_thesis.ps>, 431 * though this reference is out of date. A new paper 432 * is pending. 433 */ 434 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep) 435 { 436 u32 new_sample = tp->rcv_rtt_est.rtt; 437 long m = sample; 438 439 if (m == 0) 440 m = 1; 441 442 if (new_sample != 0) { 443 /* If we sample in larger samples in the non-timestamp 444 * case, we could grossly overestimate the RTT especially 445 * with chatty applications or bulk transfer apps which 446 * are stalled on filesystem I/O. 447 * 448 * Also, since we are only going for a minimum in the 449 * non-timestamp case, we do not smooth things out 450 * else with timestamps disabled convergence takes too 451 * long. 452 */ 453 if (!win_dep) { 454 m -= (new_sample >> 3); 455 new_sample += m; 456 } else if (m < new_sample) 457 new_sample = m << 3; 458 } else { 459 /* No previous measure. */ 460 new_sample = m << 3; 461 } 462 463 if (tp->rcv_rtt_est.rtt != new_sample) 464 tp->rcv_rtt_est.rtt = new_sample; 465 } 466 467 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp) 468 { 469 if (tp->rcv_rtt_est.time == 0) 470 goto new_measure; 471 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq)) 472 return; 473 tcp_rcv_rtt_update(tp, 474 jiffies - tp->rcv_rtt_est.time, 475 1); 476 477 new_measure: 478 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd; 479 tp->rcv_rtt_est.time = tcp_time_stamp; 480 } 481 482 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, const struct sk_buff *skb) 483 { 484 struct tcp_sock *tp = tcp_sk(sk); 485 if (tp->rx_opt.rcv_tsecr && 486 (TCP_SKB_CB(skb)->end_seq - 487 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss)) 488 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0); 489 } 490 491 /* 492 * This function should be called every time data is copied to user space. 493 * It calculates the appropriate TCP receive buffer space. 494 */ 495 void tcp_rcv_space_adjust(struct sock *sk) 496 { 497 struct tcp_sock *tp = tcp_sk(sk); 498 int time; 499 int space; 500 501 if (tp->rcvq_space.time == 0) 502 goto new_measure; 503 504 time = tcp_time_stamp - tp->rcvq_space.time; 505 if (time < (tp->rcv_rtt_est.rtt >> 3) || 506 tp->rcv_rtt_est.rtt == 0) 507 return; 508 509 space = 2 * (tp->copied_seq - tp->rcvq_space.seq); 510 511 space = max(tp->rcvq_space.space, space); 512 513 if (tp->rcvq_space.space != space) { 514 int rcvmem; 515 516 tp->rcvq_space.space = space; 517 518 if (sysctl_tcp_moderate_rcvbuf && 519 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) { 520 int new_clamp = space; 521 522 /* Receive space grows, normalize in order to 523 * take into account packet headers and sk_buff 524 * structure overhead. 525 */ 526 space /= tp->advmss; 527 if (!space) 528 space = 1; 529 rcvmem = (tp->advmss + MAX_TCP_HEADER + 530 16 + sizeof(struct sk_buff)); 531 while (tcp_win_from_space(rcvmem) < tp->advmss) 532 rcvmem += 128; 533 space *= rcvmem; 534 space = min(space, sysctl_tcp_rmem[2]); 535 if (space > sk->sk_rcvbuf) { 536 sk->sk_rcvbuf = space; 537 538 /* Make the window clamp follow along. */ 539 tp->window_clamp = new_clamp; 540 } 541 } 542 } 543 544 new_measure: 545 tp->rcvq_space.seq = tp->copied_seq; 546 tp->rcvq_space.time = tcp_time_stamp; 547 } 548 549 /* There is something which you must keep in mind when you analyze the 550 * behavior of the tp->ato delayed ack timeout interval. When a 551 * connection starts up, we want to ack as quickly as possible. The 552 * problem is that "good" TCP's do slow start at the beginning of data 553 * transmission. The means that until we send the first few ACK's the 554 * sender will sit on his end and only queue most of his data, because 555 * he can only send snd_cwnd unacked packets at any given time. For 556 * each ACK we send, he increments snd_cwnd and transmits more of his 557 * queue. -DaveM 558 */ 559 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb) 560 { 561 struct tcp_sock *tp = tcp_sk(sk); 562 struct inet_connection_sock *icsk = inet_csk(sk); 563 u32 now; 564 565 inet_csk_schedule_ack(sk); 566 567 tcp_measure_rcv_mss(sk, skb); 568 569 tcp_rcv_rtt_measure(tp); 570 571 now = tcp_time_stamp; 572 573 if (!icsk->icsk_ack.ato) { 574 /* The _first_ data packet received, initialize 575 * delayed ACK engine. 576 */ 577 tcp_incr_quickack(sk); 578 icsk->icsk_ack.ato = TCP_ATO_MIN; 579 } else { 580 int m = now - icsk->icsk_ack.lrcvtime; 581 582 if (m <= TCP_ATO_MIN/2) { 583 /* The fastest case is the first. */ 584 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2; 585 } else if (m < icsk->icsk_ack.ato) { 586 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m; 587 if (icsk->icsk_ack.ato > icsk->icsk_rto) 588 icsk->icsk_ack.ato = icsk->icsk_rto; 589 } else if (m > icsk->icsk_rto) { 590 /* Too long gap. Apparently sender failed to 591 * restart window, so that we send ACKs quickly. 592 */ 593 tcp_incr_quickack(sk); 594 sk_stream_mem_reclaim(sk); 595 } 596 } 597 icsk->icsk_ack.lrcvtime = now; 598 599 TCP_ECN_check_ce(tp, skb); 600 601 if (skb->len >= 128) 602 tcp_grow_window(sk, skb); 603 } 604 605 static u32 tcp_rto_min(struct sock *sk) 606 { 607 struct dst_entry *dst = __sk_dst_get(sk); 608 u32 rto_min = TCP_RTO_MIN; 609 610 if (dst && dst_metric_locked(dst, RTAX_RTO_MIN)) 611 rto_min = dst->metrics[RTAX_RTO_MIN-1]; 612 return rto_min; 613 } 614 615 /* Called to compute a smoothed rtt estimate. The data fed to this 616 * routine either comes from timestamps, or from segments that were 617 * known _not_ to have been retransmitted [see Karn/Partridge 618 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88 619 * piece by Van Jacobson. 620 * NOTE: the next three routines used to be one big routine. 621 * To save cycles in the RFC 1323 implementation it was better to break 622 * it up into three procedures. -- erics 623 */ 624 static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt) 625 { 626 struct tcp_sock *tp = tcp_sk(sk); 627 long m = mrtt; /* RTT */ 628 629 /* The following amusing code comes from Jacobson's 630 * article in SIGCOMM '88. Note that rtt and mdev 631 * are scaled versions of rtt and mean deviation. 632 * This is designed to be as fast as possible 633 * m stands for "measurement". 634 * 635 * On a 1990 paper the rto value is changed to: 636 * RTO = rtt + 4 * mdev 637 * 638 * Funny. This algorithm seems to be very broken. 639 * These formulae increase RTO, when it should be decreased, increase 640 * too slowly, when it should be increased quickly, decrease too quickly 641 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely 642 * does not matter how to _calculate_ it. Seems, it was trap 643 * that VJ failed to avoid. 8) 644 */ 645 if (m == 0) 646 m = 1; 647 if (tp->srtt != 0) { 648 m -= (tp->srtt >> 3); /* m is now error in rtt est */ 649 tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */ 650 if (m < 0) { 651 m = -m; /* m is now abs(error) */ 652 m -= (tp->mdev >> 2); /* similar update on mdev */ 653 /* This is similar to one of Eifel findings. 654 * Eifel blocks mdev updates when rtt decreases. 655 * This solution is a bit different: we use finer gain 656 * for mdev in this case (alpha*beta). 657 * Like Eifel it also prevents growth of rto, 658 * but also it limits too fast rto decreases, 659 * happening in pure Eifel. 660 */ 661 if (m > 0) 662 m >>= 3; 663 } else { 664 m -= (tp->mdev >> 2); /* similar update on mdev */ 665 } 666 tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */ 667 if (tp->mdev > tp->mdev_max) { 668 tp->mdev_max = tp->mdev; 669 if (tp->mdev_max > tp->rttvar) 670 tp->rttvar = tp->mdev_max; 671 } 672 if (after(tp->snd_una, tp->rtt_seq)) { 673 if (tp->mdev_max < tp->rttvar) 674 tp->rttvar -= (tp->rttvar-tp->mdev_max)>>2; 675 tp->rtt_seq = tp->snd_nxt; 676 tp->mdev_max = tcp_rto_min(sk); 677 } 678 } else { 679 /* no previous measure. */ 680 tp->srtt = m<<3; /* take the measured time to be rtt */ 681 tp->mdev = m<<1; /* make sure rto = 3*rtt */ 682 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk)); 683 tp->rtt_seq = tp->snd_nxt; 684 } 685 } 686 687 /* Calculate rto without backoff. This is the second half of Van Jacobson's 688 * routine referred to above. 689 */ 690 static inline void tcp_set_rto(struct sock *sk) 691 { 692 const struct tcp_sock *tp = tcp_sk(sk); 693 /* Old crap is replaced with new one. 8) 694 * 695 * More seriously: 696 * 1. If rtt variance happened to be less 50msec, it is hallucination. 697 * It cannot be less due to utterly erratic ACK generation made 698 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_ 699 * to do with delayed acks, because at cwnd>2 true delack timeout 700 * is invisible. Actually, Linux-2.4 also generates erratic 701 * ACKs in some circumstances. 702 */ 703 inet_csk(sk)->icsk_rto = (tp->srtt >> 3) + tp->rttvar; 704 705 /* 2. Fixups made earlier cannot be right. 706 * If we do not estimate RTO correctly without them, 707 * all the algo is pure shit and should be replaced 708 * with correct one. It is exactly, which we pretend to do. 709 */ 710 } 711 712 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo 713 * guarantees that rto is higher. 714 */ 715 static inline void tcp_bound_rto(struct sock *sk) 716 { 717 if (inet_csk(sk)->icsk_rto > TCP_RTO_MAX) 718 inet_csk(sk)->icsk_rto = TCP_RTO_MAX; 719 } 720 721 /* Save metrics learned by this TCP session. 722 This function is called only, when TCP finishes successfully 723 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE. 724 */ 725 void tcp_update_metrics(struct sock *sk) 726 { 727 struct tcp_sock *tp = tcp_sk(sk); 728 struct dst_entry *dst = __sk_dst_get(sk); 729 730 if (sysctl_tcp_nometrics_save) 731 return; 732 733 dst_confirm(dst); 734 735 if (dst && (dst->flags&DST_HOST)) { 736 const struct inet_connection_sock *icsk = inet_csk(sk); 737 int m; 738 739 if (icsk->icsk_backoff || !tp->srtt) { 740 /* This session failed to estimate rtt. Why? 741 * Probably, no packets returned in time. 742 * Reset our results. 743 */ 744 if (!(dst_metric_locked(dst, RTAX_RTT))) 745 dst->metrics[RTAX_RTT-1] = 0; 746 return; 747 } 748 749 m = dst_metric(dst, RTAX_RTT) - tp->srtt; 750 751 /* If newly calculated rtt larger than stored one, 752 * store new one. Otherwise, use EWMA. Remember, 753 * rtt overestimation is always better than underestimation. 754 */ 755 if (!(dst_metric_locked(dst, RTAX_RTT))) { 756 if (m <= 0) 757 dst->metrics[RTAX_RTT-1] = tp->srtt; 758 else 759 dst->metrics[RTAX_RTT-1] -= (m>>3); 760 } 761 762 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) { 763 if (m < 0) 764 m = -m; 765 766 /* Scale deviation to rttvar fixed point */ 767 m >>= 1; 768 if (m < tp->mdev) 769 m = tp->mdev; 770 771 if (m >= dst_metric(dst, RTAX_RTTVAR)) 772 dst->metrics[RTAX_RTTVAR-1] = m; 773 else 774 dst->metrics[RTAX_RTTVAR-1] -= 775 (dst->metrics[RTAX_RTTVAR-1] - m)>>2; 776 } 777 778 if (tp->snd_ssthresh >= 0xFFFF) { 779 /* Slow start still did not finish. */ 780 if (dst_metric(dst, RTAX_SSTHRESH) && 781 !dst_metric_locked(dst, RTAX_SSTHRESH) && 782 (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH)) 783 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1; 784 if (!dst_metric_locked(dst, RTAX_CWND) && 785 tp->snd_cwnd > dst_metric(dst, RTAX_CWND)) 786 dst->metrics[RTAX_CWND-1] = tp->snd_cwnd; 787 } else if (tp->snd_cwnd > tp->snd_ssthresh && 788 icsk->icsk_ca_state == TCP_CA_Open) { 789 /* Cong. avoidance phase, cwnd is reliable. */ 790 if (!dst_metric_locked(dst, RTAX_SSTHRESH)) 791 dst->metrics[RTAX_SSTHRESH-1] = 792 max(tp->snd_cwnd >> 1, tp->snd_ssthresh); 793 if (!dst_metric_locked(dst, RTAX_CWND)) 794 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_cwnd) >> 1; 795 } else { 796 /* Else slow start did not finish, cwnd is non-sense, 797 ssthresh may be also invalid. 798 */ 799 if (!dst_metric_locked(dst, RTAX_CWND)) 800 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_ssthresh) >> 1; 801 if (dst->metrics[RTAX_SSTHRESH-1] && 802 !dst_metric_locked(dst, RTAX_SSTHRESH) && 803 tp->snd_ssthresh > dst->metrics[RTAX_SSTHRESH-1]) 804 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh; 805 } 806 807 if (!dst_metric_locked(dst, RTAX_REORDERING)) { 808 if (dst->metrics[RTAX_REORDERING-1] < tp->reordering && 809 tp->reordering != sysctl_tcp_reordering) 810 dst->metrics[RTAX_REORDERING-1] = tp->reordering; 811 } 812 } 813 } 814 815 /* Numbers are taken from RFC3390. 816 * 817 * John Heffner states: 818 * 819 * The RFC specifies a window of no more than 4380 bytes 820 * unless 2*MSS > 4380. Reading the pseudocode in the RFC 821 * is a bit misleading because they use a clamp at 4380 bytes 822 * rather than use a multiplier in the relevant range. 823 */ 824 __u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst) 825 { 826 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0); 827 828 if (!cwnd) { 829 if (tp->mss_cache > 1460) 830 cwnd = 2; 831 else 832 cwnd = (tp->mss_cache > 1095) ? 3 : 4; 833 } 834 return min_t(__u32, cwnd, tp->snd_cwnd_clamp); 835 } 836 837 /* Set slow start threshold and cwnd not falling to slow start */ 838 void tcp_enter_cwr(struct sock *sk, const int set_ssthresh) 839 { 840 struct tcp_sock *tp = tcp_sk(sk); 841 const struct inet_connection_sock *icsk = inet_csk(sk); 842 843 tp->prior_ssthresh = 0; 844 tp->bytes_acked = 0; 845 if (icsk->icsk_ca_state < TCP_CA_CWR) { 846 tp->undo_marker = 0; 847 if (set_ssthresh) 848 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 849 tp->snd_cwnd = min(tp->snd_cwnd, 850 tcp_packets_in_flight(tp) + 1U); 851 tp->snd_cwnd_cnt = 0; 852 tp->high_seq = tp->snd_nxt; 853 tp->snd_cwnd_stamp = tcp_time_stamp; 854 TCP_ECN_queue_cwr(tp); 855 856 tcp_set_ca_state(sk, TCP_CA_CWR); 857 } 858 } 859 860 /* 861 * Packet counting of FACK is based on in-order assumptions, therefore TCP 862 * disables it when reordering is detected 863 */ 864 static void tcp_disable_fack(struct tcp_sock *tp) 865 { 866 tp->rx_opt.sack_ok &= ~2; 867 } 868 869 /* Take a notice that peer is sending DSACKs */ 870 static void tcp_dsack_seen(struct tcp_sock *tp) 871 { 872 tp->rx_opt.sack_ok |= 4; 873 } 874 875 /* Initialize metrics on socket. */ 876 877 static void tcp_init_metrics(struct sock *sk) 878 { 879 struct tcp_sock *tp = tcp_sk(sk); 880 struct dst_entry *dst = __sk_dst_get(sk); 881 882 if (dst == NULL) 883 goto reset; 884 885 dst_confirm(dst); 886 887 if (dst_metric_locked(dst, RTAX_CWND)) 888 tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND); 889 if (dst_metric(dst, RTAX_SSTHRESH)) { 890 tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH); 891 if (tp->snd_ssthresh > tp->snd_cwnd_clamp) 892 tp->snd_ssthresh = tp->snd_cwnd_clamp; 893 } 894 if (dst_metric(dst, RTAX_REORDERING) && 895 tp->reordering != dst_metric(dst, RTAX_REORDERING)) { 896 tcp_disable_fack(tp); 897 tp->reordering = dst_metric(dst, RTAX_REORDERING); 898 } 899 900 if (dst_metric(dst, RTAX_RTT) == 0) 901 goto reset; 902 903 if (!tp->srtt && dst_metric(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3)) 904 goto reset; 905 906 /* Initial rtt is determined from SYN,SYN-ACK. 907 * The segment is small and rtt may appear much 908 * less than real one. Use per-dst memory 909 * to make it more realistic. 910 * 911 * A bit of theory. RTT is time passed after "normal" sized packet 912 * is sent until it is ACKed. In normal circumstances sending small 913 * packets force peer to delay ACKs and calculation is correct too. 914 * The algorithm is adaptive and, provided we follow specs, it 915 * NEVER underestimate RTT. BUT! If peer tries to make some clever 916 * tricks sort of "quick acks" for time long enough to decrease RTT 917 * to low value, and then abruptly stops to do it and starts to delay 918 * ACKs, wait for troubles. 919 */ 920 if (dst_metric(dst, RTAX_RTT) > tp->srtt) { 921 tp->srtt = dst_metric(dst, RTAX_RTT); 922 tp->rtt_seq = tp->snd_nxt; 923 } 924 if (dst_metric(dst, RTAX_RTTVAR) > tp->mdev) { 925 tp->mdev = dst_metric(dst, RTAX_RTTVAR); 926 tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN); 927 } 928 tcp_set_rto(sk); 929 tcp_bound_rto(sk); 930 if (inet_csk(sk)->icsk_rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp) 931 goto reset; 932 tp->snd_cwnd = tcp_init_cwnd(tp, dst); 933 tp->snd_cwnd_stamp = tcp_time_stamp; 934 return; 935 936 reset: 937 /* Play conservative. If timestamps are not 938 * supported, TCP will fail to recalculate correct 939 * rtt, if initial rto is too small. FORGET ALL AND RESET! 940 */ 941 if (!tp->rx_opt.saw_tstamp && tp->srtt) { 942 tp->srtt = 0; 943 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT; 944 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; 945 } 946 } 947 948 static void tcp_update_reordering(struct sock *sk, const int metric, 949 const int ts) 950 { 951 struct tcp_sock *tp = tcp_sk(sk); 952 if (metric > tp->reordering) { 953 tp->reordering = min(TCP_MAX_REORDERING, metric); 954 955 /* This exciting event is worth to be remembered. 8) */ 956 if (ts) 957 NET_INC_STATS_BH(LINUX_MIB_TCPTSREORDER); 958 else if (tcp_is_reno(tp)) 959 NET_INC_STATS_BH(LINUX_MIB_TCPRENOREORDER); 960 else if (tcp_is_fack(tp)) 961 NET_INC_STATS_BH(LINUX_MIB_TCPFACKREORDER); 962 else 963 NET_INC_STATS_BH(LINUX_MIB_TCPSACKREORDER); 964 #if FASTRETRANS_DEBUG > 1 965 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n", 966 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state, 967 tp->reordering, 968 tp->fackets_out, 969 tp->sacked_out, 970 tp->undo_marker ? tp->undo_retrans : 0); 971 #endif 972 tcp_disable_fack(tp); 973 } 974 } 975 976 /* This procedure tags the retransmission queue when SACKs arrive. 977 * 978 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L). 979 * Packets in queue with these bits set are counted in variables 980 * sacked_out, retrans_out and lost_out, correspondingly. 981 * 982 * Valid combinations are: 983 * Tag InFlight Description 984 * 0 1 - orig segment is in flight. 985 * S 0 - nothing flies, orig reached receiver. 986 * L 0 - nothing flies, orig lost by net. 987 * R 2 - both orig and retransmit are in flight. 988 * L|R 1 - orig is lost, retransmit is in flight. 989 * S|R 1 - orig reached receiver, retrans is still in flight. 990 * (L|S|R is logically valid, it could occur when L|R is sacked, 991 * but it is equivalent to plain S and code short-curcuits it to S. 992 * L|S is logically invalid, it would mean -1 packet in flight 8)) 993 * 994 * These 6 states form finite state machine, controlled by the following events: 995 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue()) 996 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue()) 997 * 3. Loss detection event of one of three flavors: 998 * A. Scoreboard estimator decided the packet is lost. 999 * A'. Reno "three dupacks" marks head of queue lost. 1000 * A''. Its FACK modfication, head until snd.fack is lost. 1001 * B. SACK arrives sacking data transmitted after never retransmitted 1002 * hole was sent out. 1003 * C. SACK arrives sacking SND.NXT at the moment, when the 1004 * segment was retransmitted. 1005 * 4. D-SACK added new rule: D-SACK changes any tag to S. 1006 * 1007 * It is pleasant to note, that state diagram turns out to be commutative, 1008 * so that we are allowed not to be bothered by order of our actions, 1009 * when multiple events arrive simultaneously. (see the function below). 1010 * 1011 * Reordering detection. 1012 * -------------------- 1013 * Reordering metric is maximal distance, which a packet can be displaced 1014 * in packet stream. With SACKs we can estimate it: 1015 * 1016 * 1. SACK fills old hole and the corresponding segment was not 1017 * ever retransmitted -> reordering. Alas, we cannot use it 1018 * when segment was retransmitted. 1019 * 2. The last flaw is solved with D-SACK. D-SACK arrives 1020 * for retransmitted and already SACKed segment -> reordering.. 1021 * Both of these heuristics are not used in Loss state, when we cannot 1022 * account for retransmits accurately. 1023 * 1024 * SACK block validation. 1025 * ---------------------- 1026 * 1027 * SACK block range validation checks that the received SACK block fits to 1028 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT. 1029 * Note that SND.UNA is not included to the range though being valid because 1030 * it means that the receiver is rather inconsistent with itself reporting 1031 * SACK reneging when it should advance SND.UNA. Such SACK block this is 1032 * perfectly valid, however, in light of RFC2018 which explicitly states 1033 * that "SACK block MUST reflect the newest segment. Even if the newest 1034 * segment is going to be discarded ...", not that it looks very clever 1035 * in case of head skb. Due to potentional receiver driven attacks, we 1036 * choose to avoid immediate execution of a walk in write queue due to 1037 * reneging and defer head skb's loss recovery to standard loss recovery 1038 * procedure that will eventually trigger (nothing forbids us doing this). 1039 * 1040 * Implements also blockage to start_seq wrap-around. Problem lies in the 1041 * fact that though start_seq (s) is before end_seq (i.e., not reversed), 1042 * there's no guarantee that it will be before snd_nxt (n). The problem 1043 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt 1044 * wrap (s_w): 1045 * 1046 * <- outs wnd -> <- wrapzone -> 1047 * u e n u_w e_w s n_w 1048 * | | | | | | | 1049 * |<------------+------+----- TCP seqno space --------------+---------->| 1050 * ...-- <2^31 ->| |<--------... 1051 * ...---- >2^31 ------>| |<--------... 1052 * 1053 * Current code wouldn't be vulnerable but it's better still to discard such 1054 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat 1055 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in 1056 * snd_nxt wrap -> snd_una region will then become "well defined", i.e., 1057 * equal to the ideal case (infinite seqno space without wrap caused issues). 1058 * 1059 * With D-SACK the lower bound is extended to cover sequence space below 1060 * SND.UNA down to undo_marker, which is the last point of interest. Yet 1061 * again, DSACK block must not to go across snd_una (for the same reason as 1062 * for the normal SACK blocks, explained above). But there all simplicity 1063 * ends, TCP might receive valid D-SACKs below that. As long as they reside 1064 * fully below undo_marker they do not affect behavior in anyway and can 1065 * therefore be safely ignored. In rare cases (which are more or less 1066 * theoretical ones), the D-SACK will nicely cross that boundary due to skb 1067 * fragmentation and packet reordering past skb's retransmission. To consider 1068 * them correctly, the acceptable range must be extended even more though 1069 * the exact amount is rather hard to quantify. However, tp->max_window can 1070 * be used as an exaggerated estimate. 1071 */ 1072 static int tcp_is_sackblock_valid(struct tcp_sock *tp, int is_dsack, 1073 u32 start_seq, u32 end_seq) 1074 { 1075 /* Too far in future, or reversed (interpretation is ambiguous) */ 1076 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq)) 1077 return 0; 1078 1079 /* Nasty start_seq wrap-around check (see comments above) */ 1080 if (!before(start_seq, tp->snd_nxt)) 1081 return 0; 1082 1083 /* In outstanding window? ...This is valid exit for DSACKs too. 1084 * start_seq == snd_una is non-sensical (see comments above) 1085 */ 1086 if (after(start_seq, tp->snd_una)) 1087 return 1; 1088 1089 if (!is_dsack || !tp->undo_marker) 1090 return 0; 1091 1092 /* ...Then it's D-SACK, and must reside below snd_una completely */ 1093 if (!after(end_seq, tp->snd_una)) 1094 return 0; 1095 1096 if (!before(start_seq, tp->undo_marker)) 1097 return 1; 1098 1099 /* Too old */ 1100 if (!after(end_seq, tp->undo_marker)) 1101 return 0; 1102 1103 /* Undo_marker boundary crossing (overestimates a lot). Known already: 1104 * start_seq < undo_marker and end_seq >= undo_marker. 1105 */ 1106 return !before(start_seq, end_seq - tp->max_window); 1107 } 1108 1109 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving". 1110 * Event "C". Later note: FACK people cheated me again 8), we have to account 1111 * for reordering! Ugly, but should help. 1112 * 1113 * Search retransmitted skbs from write_queue that were sent when snd_nxt was 1114 * less than what is now known to be received by the other end (derived from 1115 * SACK blocks by the caller). Also calculate the lowest snd_nxt among the 1116 * remaining retransmitted skbs to avoid some costly processing per ACKs. 1117 */ 1118 static int tcp_mark_lost_retrans(struct sock *sk, u32 received_upto) 1119 { 1120 struct tcp_sock *tp = tcp_sk(sk); 1121 struct sk_buff *skb; 1122 int flag = 0; 1123 int cnt = 0; 1124 u32 new_low_seq = 0; 1125 1126 tcp_for_write_queue(skb, sk) { 1127 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq; 1128 1129 if (skb == tcp_send_head(sk)) 1130 break; 1131 if (cnt == tp->retrans_out) 1132 break; 1133 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 1134 continue; 1135 1136 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)) 1137 continue; 1138 1139 if (after(received_upto, ack_seq) && 1140 (tcp_is_fack(tp) || 1141 !before(received_upto, 1142 ack_seq + tp->reordering * tp->mss_cache))) { 1143 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 1144 tp->retrans_out -= tcp_skb_pcount(skb); 1145 1146 /* clear lost hint */ 1147 tp->retransmit_skb_hint = NULL; 1148 1149 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) { 1150 tp->lost_out += tcp_skb_pcount(skb); 1151 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 1152 flag |= FLAG_DATA_SACKED; 1153 NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT); 1154 } 1155 } else { 1156 if (!new_low_seq || before(ack_seq, new_low_seq)) 1157 new_low_seq = ack_seq; 1158 cnt += tcp_skb_pcount(skb); 1159 } 1160 } 1161 1162 if (tp->retrans_out) 1163 tp->lost_retrans_low = new_low_seq; 1164 1165 return flag; 1166 } 1167 1168 static int tcp_check_dsack(struct tcp_sock *tp, struct sk_buff *ack_skb, 1169 struct tcp_sack_block_wire *sp, int num_sacks, 1170 u32 prior_snd_una) 1171 { 1172 u32 start_seq_0 = ntohl(get_unaligned(&sp[0].start_seq)); 1173 u32 end_seq_0 = ntohl(get_unaligned(&sp[0].end_seq)); 1174 int dup_sack = 0; 1175 1176 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) { 1177 dup_sack = 1; 1178 tcp_dsack_seen(tp); 1179 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV); 1180 } else if (num_sacks > 1) { 1181 u32 end_seq_1 = ntohl(get_unaligned(&sp[1].end_seq)); 1182 u32 start_seq_1 = ntohl(get_unaligned(&sp[1].start_seq)); 1183 1184 if (!after(end_seq_0, end_seq_1) && 1185 !before(start_seq_0, start_seq_1)) { 1186 dup_sack = 1; 1187 tcp_dsack_seen(tp); 1188 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV); 1189 } 1190 } 1191 1192 /* D-SACK for already forgotten data... Do dumb counting. */ 1193 if (dup_sack && 1194 !after(end_seq_0, prior_snd_una) && 1195 after(end_seq_0, tp->undo_marker)) 1196 tp->undo_retrans--; 1197 1198 return dup_sack; 1199 } 1200 1201 /* Check if skb is fully within the SACK block. In presence of GSO skbs, 1202 * the incoming SACK may not exactly match but we can find smaller MSS 1203 * aligned portion of it that matches. Therefore we might need to fragment 1204 * which may fail and creates some hassle (caller must handle error case 1205 * returns). 1206 */ 1207 int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb, 1208 u32 start_seq, u32 end_seq) 1209 { 1210 int in_sack, err; 1211 unsigned int pkt_len; 1212 1213 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && 1214 !before(end_seq, TCP_SKB_CB(skb)->end_seq); 1215 1216 if (tcp_skb_pcount(skb) > 1 && !in_sack && 1217 after(TCP_SKB_CB(skb)->end_seq, start_seq)) { 1218 1219 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); 1220 1221 if (!in_sack) 1222 pkt_len = start_seq - TCP_SKB_CB(skb)->seq; 1223 else 1224 pkt_len = end_seq - TCP_SKB_CB(skb)->seq; 1225 err = tcp_fragment(sk, skb, pkt_len, skb_shinfo(skb)->gso_size); 1226 if (err < 0) 1227 return err; 1228 } 1229 1230 return in_sack; 1231 } 1232 1233 static int 1234 tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, u32 prior_snd_una) 1235 { 1236 const struct inet_connection_sock *icsk = inet_csk(sk); 1237 struct tcp_sock *tp = tcp_sk(sk); 1238 unsigned char *ptr = (skb_transport_header(ack_skb) + 1239 TCP_SKB_CB(ack_skb)->sacked); 1240 struct tcp_sack_block_wire *sp = (struct tcp_sack_block_wire *)(ptr+2); 1241 struct sk_buff *cached_skb; 1242 int num_sacks = (ptr[1] - TCPOLEN_SACK_BASE)>>3; 1243 int reord = tp->packets_out; 1244 int prior_fackets; 1245 u32 highest_sack_end_seq = 0; 1246 int flag = 0; 1247 int found_dup_sack = 0; 1248 int cached_fack_count; 1249 int i; 1250 int first_sack_index; 1251 1252 if (!tp->sacked_out) { 1253 if (WARN_ON(tp->fackets_out)) 1254 tp->fackets_out = 0; 1255 tp->highest_sack = tp->snd_una; 1256 } 1257 prior_fackets = tp->fackets_out; 1258 1259 found_dup_sack = tcp_check_dsack(tp, ack_skb, sp, 1260 num_sacks, prior_snd_una); 1261 if (found_dup_sack) 1262 flag |= FLAG_DSACKING_ACK; 1263 1264 /* Eliminate too old ACKs, but take into 1265 * account more or less fresh ones, they can 1266 * contain valid SACK info. 1267 */ 1268 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window)) 1269 return 0; 1270 1271 /* SACK fastpath: 1272 * if the only SACK change is the increase of the end_seq of 1273 * the first block then only apply that SACK block 1274 * and use retrans queue hinting otherwise slowpath */ 1275 flag = 1; 1276 for (i = 0; i < num_sacks; i++) { 1277 __be32 start_seq = sp[i].start_seq; 1278 __be32 end_seq = sp[i].end_seq; 1279 1280 if (i == 0) { 1281 if (tp->recv_sack_cache[i].start_seq != start_seq) 1282 flag = 0; 1283 } else { 1284 if ((tp->recv_sack_cache[i].start_seq != start_seq) || 1285 (tp->recv_sack_cache[i].end_seq != end_seq)) 1286 flag = 0; 1287 } 1288 tp->recv_sack_cache[i].start_seq = start_seq; 1289 tp->recv_sack_cache[i].end_seq = end_seq; 1290 } 1291 /* Clear the rest of the cache sack blocks so they won't match mistakenly. */ 1292 for (; i < ARRAY_SIZE(tp->recv_sack_cache); i++) { 1293 tp->recv_sack_cache[i].start_seq = 0; 1294 tp->recv_sack_cache[i].end_seq = 0; 1295 } 1296 1297 first_sack_index = 0; 1298 if (flag) 1299 num_sacks = 1; 1300 else { 1301 int j; 1302 tp->fastpath_skb_hint = NULL; 1303 1304 /* order SACK blocks to allow in order walk of the retrans queue */ 1305 for (i = num_sacks-1; i > 0; i--) { 1306 for (j = 0; j < i; j++){ 1307 if (after(ntohl(sp[j].start_seq), 1308 ntohl(sp[j+1].start_seq))){ 1309 struct tcp_sack_block_wire tmp; 1310 1311 tmp = sp[j]; 1312 sp[j] = sp[j+1]; 1313 sp[j+1] = tmp; 1314 1315 /* Track where the first SACK block goes to */ 1316 if (j == first_sack_index) 1317 first_sack_index = j+1; 1318 } 1319 1320 } 1321 } 1322 } 1323 1324 /* clear flag as used for different purpose in following code */ 1325 flag = 0; 1326 1327 /* Use SACK fastpath hint if valid */ 1328 cached_skb = tp->fastpath_skb_hint; 1329 cached_fack_count = tp->fastpath_cnt_hint; 1330 if (!cached_skb) { 1331 cached_skb = tcp_write_queue_head(sk); 1332 cached_fack_count = 0; 1333 } 1334 1335 for (i=0; i<num_sacks; i++, sp++) { 1336 struct sk_buff *skb; 1337 __u32 start_seq = ntohl(sp->start_seq); 1338 __u32 end_seq = ntohl(sp->end_seq); 1339 int fack_count; 1340 int dup_sack = (found_dup_sack && (i == first_sack_index)); 1341 1342 if (!tcp_is_sackblock_valid(tp, dup_sack, start_seq, end_seq)) { 1343 if (dup_sack) { 1344 if (!tp->undo_marker) 1345 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKIGNOREDNOUNDO); 1346 else 1347 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKIGNOREDOLD); 1348 } else { 1349 /* Don't count olds caused by ACK reordering */ 1350 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) && 1351 !after(end_seq, tp->snd_una)) 1352 continue; 1353 NET_INC_STATS_BH(LINUX_MIB_TCPSACKDISCARD); 1354 } 1355 continue; 1356 } 1357 1358 skb = cached_skb; 1359 fack_count = cached_fack_count; 1360 1361 /* Event "B" in the comment above. */ 1362 if (after(end_seq, tp->high_seq)) 1363 flag |= FLAG_DATA_LOST; 1364 1365 tcp_for_write_queue_from(skb, sk) { 1366 int in_sack; 1367 u8 sacked; 1368 1369 if (skb == tcp_send_head(sk)) 1370 break; 1371 1372 cached_skb = skb; 1373 cached_fack_count = fack_count; 1374 if (i == first_sack_index) { 1375 tp->fastpath_skb_hint = skb; 1376 tp->fastpath_cnt_hint = fack_count; 1377 } 1378 1379 /* The retransmission queue is always in order, so 1380 * we can short-circuit the walk early. 1381 */ 1382 if (!before(TCP_SKB_CB(skb)->seq, end_seq)) 1383 break; 1384 1385 in_sack = tcp_match_skb_to_sack(sk, skb, start_seq, end_seq); 1386 if (in_sack < 0) 1387 break; 1388 1389 fack_count += tcp_skb_pcount(skb); 1390 1391 sacked = TCP_SKB_CB(skb)->sacked; 1392 1393 /* Account D-SACK for retransmitted packet. */ 1394 if ((dup_sack && in_sack) && 1395 (sacked & TCPCB_RETRANS) && 1396 after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker)) 1397 tp->undo_retrans--; 1398 1399 /* The frame is ACKed. */ 1400 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) { 1401 if (sacked&TCPCB_RETRANS) { 1402 if ((dup_sack && in_sack) && 1403 (sacked&TCPCB_SACKED_ACKED)) 1404 reord = min(fack_count, reord); 1405 } else { 1406 /* If it was in a hole, we detected reordering. */ 1407 if (fack_count < prior_fackets && 1408 !(sacked&TCPCB_SACKED_ACKED)) 1409 reord = min(fack_count, reord); 1410 } 1411 1412 /* Nothing to do; acked frame is about to be dropped. */ 1413 continue; 1414 } 1415 1416 if (!in_sack) 1417 continue; 1418 1419 if (!(sacked&TCPCB_SACKED_ACKED)) { 1420 if (sacked & TCPCB_SACKED_RETRANS) { 1421 /* If the segment is not tagged as lost, 1422 * we do not clear RETRANS, believing 1423 * that retransmission is still in flight. 1424 */ 1425 if (sacked & TCPCB_LOST) { 1426 TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS); 1427 tp->lost_out -= tcp_skb_pcount(skb); 1428 tp->retrans_out -= tcp_skb_pcount(skb); 1429 1430 /* clear lost hint */ 1431 tp->retransmit_skb_hint = NULL; 1432 } 1433 } else { 1434 /* New sack for not retransmitted frame, 1435 * which was in hole. It is reordering. 1436 */ 1437 if (!(sacked & TCPCB_RETRANS) && 1438 fack_count < prior_fackets) 1439 reord = min(fack_count, reord); 1440 1441 if (sacked & TCPCB_LOST) { 1442 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; 1443 tp->lost_out -= tcp_skb_pcount(skb); 1444 1445 /* clear lost hint */ 1446 tp->retransmit_skb_hint = NULL; 1447 } 1448 /* SACK enhanced F-RTO detection. 1449 * Set flag if and only if non-rexmitted 1450 * segments below frto_highmark are 1451 * SACKed (RFC4138; Appendix B). 1452 * Clearing correct due to in-order walk 1453 */ 1454 if (after(end_seq, tp->frto_highmark)) { 1455 flag &= ~FLAG_ONLY_ORIG_SACKED; 1456 } else { 1457 if (!(sacked & TCPCB_RETRANS)) 1458 flag |= FLAG_ONLY_ORIG_SACKED; 1459 } 1460 } 1461 1462 TCP_SKB_CB(skb)->sacked |= TCPCB_SACKED_ACKED; 1463 flag |= FLAG_DATA_SACKED; 1464 tp->sacked_out += tcp_skb_pcount(skb); 1465 1466 if (fack_count > tp->fackets_out) 1467 tp->fackets_out = fack_count; 1468 1469 if (after(TCP_SKB_CB(skb)->seq, tp->highest_sack)) { 1470 tp->highest_sack = TCP_SKB_CB(skb)->seq; 1471 highest_sack_end_seq = TCP_SKB_CB(skb)->end_seq; 1472 } 1473 } else { 1474 if (dup_sack && (sacked&TCPCB_RETRANS)) 1475 reord = min(fack_count, reord); 1476 } 1477 1478 /* D-SACK. We can detect redundant retransmission 1479 * in S|R and plain R frames and clear it. 1480 * undo_retrans is decreased above, L|R frames 1481 * are accounted above as well. 1482 */ 1483 if (dup_sack && 1484 (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS)) { 1485 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 1486 tp->retrans_out -= tcp_skb_pcount(skb); 1487 tp->retransmit_skb_hint = NULL; 1488 } 1489 } 1490 } 1491 1492 if (tp->retrans_out && 1493 after(highest_sack_end_seq, tp->lost_retrans_low) && 1494 icsk->icsk_ca_state == TCP_CA_Recovery) 1495 flag |= tcp_mark_lost_retrans(sk, highest_sack_end_seq); 1496 1497 tcp_verify_left_out(tp); 1498 1499 if ((reord < tp->fackets_out) && icsk->icsk_ca_state != TCP_CA_Loss && 1500 (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark))) 1501 tcp_update_reordering(sk, ((tp->fackets_out + 1) - reord), 0); 1502 1503 #if FASTRETRANS_DEBUG > 0 1504 BUG_TRAP((int)tp->sacked_out >= 0); 1505 BUG_TRAP((int)tp->lost_out >= 0); 1506 BUG_TRAP((int)tp->retrans_out >= 0); 1507 BUG_TRAP((int)tcp_packets_in_flight(tp) >= 0); 1508 #endif 1509 return flag; 1510 } 1511 1512 /* If we receive more dupacks than we expected counting segments 1513 * in assumption of absent reordering, interpret this as reordering. 1514 * The only another reason could be bug in receiver TCP. 1515 */ 1516 static void tcp_check_reno_reordering(struct sock *sk, const int addend) 1517 { 1518 struct tcp_sock *tp = tcp_sk(sk); 1519 u32 holes; 1520 1521 holes = max(tp->lost_out, 1U); 1522 holes = min(holes, tp->packets_out); 1523 1524 if ((tp->sacked_out + holes) > tp->packets_out) { 1525 tp->sacked_out = tp->packets_out - holes; 1526 tcp_update_reordering(sk, tp->packets_out + addend, 0); 1527 } 1528 } 1529 1530 /* Emulate SACKs for SACKless connection: account for a new dupack. */ 1531 1532 static void tcp_add_reno_sack(struct sock *sk) 1533 { 1534 struct tcp_sock *tp = tcp_sk(sk); 1535 tp->sacked_out++; 1536 tcp_check_reno_reordering(sk, 0); 1537 tcp_verify_left_out(tp); 1538 } 1539 1540 /* Account for ACK, ACKing some data in Reno Recovery phase. */ 1541 1542 static void tcp_remove_reno_sacks(struct sock *sk, int acked) 1543 { 1544 struct tcp_sock *tp = tcp_sk(sk); 1545 1546 if (acked > 0) { 1547 /* One ACK acked hole. The rest eat duplicate ACKs. */ 1548 if (acked-1 >= tp->sacked_out) 1549 tp->sacked_out = 0; 1550 else 1551 tp->sacked_out -= acked-1; 1552 } 1553 tcp_check_reno_reordering(sk, acked); 1554 tcp_verify_left_out(tp); 1555 } 1556 1557 static inline void tcp_reset_reno_sack(struct tcp_sock *tp) 1558 { 1559 tp->sacked_out = 0; 1560 } 1561 1562 /* F-RTO can only be used if TCP has never retransmitted anything other than 1563 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here) 1564 */ 1565 int tcp_use_frto(struct sock *sk) 1566 { 1567 const struct tcp_sock *tp = tcp_sk(sk); 1568 struct sk_buff *skb; 1569 1570 if (!sysctl_tcp_frto) 1571 return 0; 1572 1573 if (IsSackFrto()) 1574 return 1; 1575 1576 /* Avoid expensive walking of rexmit queue if possible */ 1577 if (tp->retrans_out > 1) 1578 return 0; 1579 1580 skb = tcp_write_queue_head(sk); 1581 skb = tcp_write_queue_next(sk, skb); /* Skips head */ 1582 tcp_for_write_queue_from(skb, sk) { 1583 if (skb == tcp_send_head(sk)) 1584 break; 1585 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS) 1586 return 0; 1587 /* Short-circuit when first non-SACKed skb has been checked */ 1588 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) 1589 break; 1590 } 1591 return 1; 1592 } 1593 1594 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO 1595 * recovery a bit and use heuristics in tcp_process_frto() to detect if 1596 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to 1597 * keep retrans_out counting accurate (with SACK F-RTO, other than head 1598 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS 1599 * bits are handled if the Loss state is really to be entered (in 1600 * tcp_enter_frto_loss). 1601 * 1602 * Do like tcp_enter_loss() would; when RTO expires the second time it 1603 * does: 1604 * "Reduce ssthresh if it has not yet been made inside this window." 1605 */ 1606 void tcp_enter_frto(struct sock *sk) 1607 { 1608 const struct inet_connection_sock *icsk = inet_csk(sk); 1609 struct tcp_sock *tp = tcp_sk(sk); 1610 struct sk_buff *skb; 1611 1612 if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) || 1613 tp->snd_una == tp->high_seq || 1614 ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) && 1615 !icsk->icsk_retransmits)) { 1616 tp->prior_ssthresh = tcp_current_ssthresh(sk); 1617 /* Our state is too optimistic in ssthresh() call because cwnd 1618 * is not reduced until tcp_enter_frto_loss() when previous FRTO 1619 * recovery has not yet completed. Pattern would be this: RTO, 1620 * Cumulative ACK, RTO (2xRTO for the same segment does not end 1621 * up here twice). 1622 * RFC4138 should be more specific on what to do, even though 1623 * RTO is quite unlikely to occur after the first Cumulative ACK 1624 * due to back-off and complexity of triggering events ... 1625 */ 1626 if (tp->frto_counter) { 1627 u32 stored_cwnd; 1628 stored_cwnd = tp->snd_cwnd; 1629 tp->snd_cwnd = 2; 1630 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 1631 tp->snd_cwnd = stored_cwnd; 1632 } else { 1633 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 1634 } 1635 /* ... in theory, cong.control module could do "any tricks" in 1636 * ssthresh(), which means that ca_state, lost bits and lost_out 1637 * counter would have to be faked before the call occurs. We 1638 * consider that too expensive, unlikely and hacky, so modules 1639 * using these in ssthresh() must deal these incompatibility 1640 * issues if they receives CA_EVENT_FRTO and frto_counter != 0 1641 */ 1642 tcp_ca_event(sk, CA_EVENT_FRTO); 1643 } 1644 1645 tp->undo_marker = tp->snd_una; 1646 tp->undo_retrans = 0; 1647 1648 skb = tcp_write_queue_head(sk); 1649 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) 1650 tp->undo_marker = 0; 1651 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 1652 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 1653 tp->retrans_out -= tcp_skb_pcount(skb); 1654 } 1655 tcp_verify_left_out(tp); 1656 1657 /* Earlier loss recovery underway (see RFC4138; Appendix B). 1658 * The last condition is necessary at least in tp->frto_counter case. 1659 */ 1660 if (IsSackFrto() && (tp->frto_counter || 1661 ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) && 1662 after(tp->high_seq, tp->snd_una)) { 1663 tp->frto_highmark = tp->high_seq; 1664 } else { 1665 tp->frto_highmark = tp->snd_nxt; 1666 } 1667 tcp_set_ca_state(sk, TCP_CA_Disorder); 1668 tp->high_seq = tp->snd_nxt; 1669 tp->frto_counter = 1; 1670 } 1671 1672 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO, 1673 * which indicates that we should follow the traditional RTO recovery, 1674 * i.e. mark everything lost and do go-back-N retransmission. 1675 */ 1676 static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag) 1677 { 1678 struct tcp_sock *tp = tcp_sk(sk); 1679 struct sk_buff *skb; 1680 1681 tp->lost_out = 0; 1682 tp->retrans_out = 0; 1683 if (tcp_is_reno(tp)) 1684 tcp_reset_reno_sack(tp); 1685 1686 tcp_for_write_queue(skb, sk) { 1687 if (skb == tcp_send_head(sk)) 1688 break; 1689 /* 1690 * Count the retransmission made on RTO correctly (only when 1691 * waiting for the first ACK and did not get it)... 1692 */ 1693 if ((tp->frto_counter == 1) && !(flag&FLAG_DATA_ACKED)) { 1694 /* For some reason this R-bit might get cleared? */ 1695 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 1696 tp->retrans_out += tcp_skb_pcount(skb); 1697 /* ...enter this if branch just for the first segment */ 1698 flag |= FLAG_DATA_ACKED; 1699 } else { 1700 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) 1701 tp->undo_marker = 0; 1702 TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS); 1703 } 1704 1705 /* Don't lost mark skbs that were fwd transmitted after RTO */ 1706 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) && 1707 !after(TCP_SKB_CB(skb)->end_seq, tp->frto_highmark)) { 1708 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 1709 tp->lost_out += tcp_skb_pcount(skb); 1710 } 1711 } 1712 tcp_verify_left_out(tp); 1713 1714 tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments; 1715 tp->snd_cwnd_cnt = 0; 1716 tp->snd_cwnd_stamp = tcp_time_stamp; 1717 tp->frto_counter = 0; 1718 tp->bytes_acked = 0; 1719 1720 tp->reordering = min_t(unsigned int, tp->reordering, 1721 sysctl_tcp_reordering); 1722 tcp_set_ca_state(sk, TCP_CA_Loss); 1723 tp->high_seq = tp->frto_highmark; 1724 TCP_ECN_queue_cwr(tp); 1725 1726 tcp_clear_retrans_hints_partial(tp); 1727 } 1728 1729 static void tcp_clear_retrans_partial(struct tcp_sock *tp) 1730 { 1731 tp->retrans_out = 0; 1732 tp->lost_out = 0; 1733 1734 tp->undo_marker = 0; 1735 tp->undo_retrans = 0; 1736 } 1737 1738 void tcp_clear_retrans(struct tcp_sock *tp) 1739 { 1740 tcp_clear_retrans_partial(tp); 1741 1742 tp->fackets_out = 0; 1743 tp->sacked_out = 0; 1744 } 1745 1746 /* Enter Loss state. If "how" is not zero, forget all SACK information 1747 * and reset tags completely, otherwise preserve SACKs. If receiver 1748 * dropped its ofo queue, we will know this due to reneging detection. 1749 */ 1750 void tcp_enter_loss(struct sock *sk, int how) 1751 { 1752 const struct inet_connection_sock *icsk = inet_csk(sk); 1753 struct tcp_sock *tp = tcp_sk(sk); 1754 struct sk_buff *skb; 1755 1756 /* Reduce ssthresh if it has not yet been made inside this window. */ 1757 if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq || 1758 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) { 1759 tp->prior_ssthresh = tcp_current_ssthresh(sk); 1760 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 1761 tcp_ca_event(sk, CA_EVENT_LOSS); 1762 } 1763 tp->snd_cwnd = 1; 1764 tp->snd_cwnd_cnt = 0; 1765 tp->snd_cwnd_stamp = tcp_time_stamp; 1766 1767 tp->bytes_acked = 0; 1768 tcp_clear_retrans_partial(tp); 1769 1770 if (tcp_is_reno(tp)) 1771 tcp_reset_reno_sack(tp); 1772 1773 if (!how) { 1774 /* Push undo marker, if it was plain RTO and nothing 1775 * was retransmitted. */ 1776 tp->undo_marker = tp->snd_una; 1777 tcp_clear_retrans_hints_partial(tp); 1778 } else { 1779 tp->sacked_out = 0; 1780 tp->fackets_out = 0; 1781 tcp_clear_all_retrans_hints(tp); 1782 } 1783 1784 tcp_for_write_queue(skb, sk) { 1785 if (skb == tcp_send_head(sk)) 1786 break; 1787 1788 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS) 1789 tp->undo_marker = 0; 1790 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED; 1791 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) { 1792 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED; 1793 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 1794 tp->lost_out += tcp_skb_pcount(skb); 1795 } 1796 } 1797 tcp_verify_left_out(tp); 1798 1799 tp->reordering = min_t(unsigned int, tp->reordering, 1800 sysctl_tcp_reordering); 1801 tcp_set_ca_state(sk, TCP_CA_Loss); 1802 tp->high_seq = tp->snd_nxt; 1803 TCP_ECN_queue_cwr(tp); 1804 /* Abort FRTO algorithm if one is in progress */ 1805 tp->frto_counter = 0; 1806 } 1807 1808 static int tcp_check_sack_reneging(struct sock *sk) 1809 { 1810 struct sk_buff *skb; 1811 1812 /* If ACK arrived pointing to a remembered SACK, 1813 * it means that our remembered SACKs do not reflect 1814 * real state of receiver i.e. 1815 * receiver _host_ is heavily congested (or buggy). 1816 * Do processing similar to RTO timeout. 1817 */ 1818 if ((skb = tcp_write_queue_head(sk)) != NULL && 1819 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) { 1820 struct inet_connection_sock *icsk = inet_csk(sk); 1821 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING); 1822 1823 tcp_enter_loss(sk, 1); 1824 icsk->icsk_retransmits++; 1825 tcp_retransmit_skb(sk, tcp_write_queue_head(sk)); 1826 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 1827 icsk->icsk_rto, TCP_RTO_MAX); 1828 return 1; 1829 } 1830 return 0; 1831 } 1832 1833 static inline int tcp_fackets_out(struct tcp_sock *tp) 1834 { 1835 return tcp_is_reno(tp) ? tp->sacked_out+1 : tp->fackets_out; 1836 } 1837 1838 static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb) 1839 { 1840 return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto); 1841 } 1842 1843 static inline int tcp_head_timedout(struct sock *sk) 1844 { 1845 struct tcp_sock *tp = tcp_sk(sk); 1846 1847 return tp->packets_out && 1848 tcp_skb_timedout(sk, tcp_write_queue_head(sk)); 1849 } 1850 1851 /* Linux NewReno/SACK/FACK/ECN state machine. 1852 * -------------------------------------- 1853 * 1854 * "Open" Normal state, no dubious events, fast path. 1855 * "Disorder" In all the respects it is "Open", 1856 * but requires a bit more attention. It is entered when 1857 * we see some SACKs or dupacks. It is split of "Open" 1858 * mainly to move some processing from fast path to slow one. 1859 * "CWR" CWND was reduced due to some Congestion Notification event. 1860 * It can be ECN, ICMP source quench, local device congestion. 1861 * "Recovery" CWND was reduced, we are fast-retransmitting. 1862 * "Loss" CWND was reduced due to RTO timeout or SACK reneging. 1863 * 1864 * tcp_fastretrans_alert() is entered: 1865 * - each incoming ACK, if state is not "Open" 1866 * - when arrived ACK is unusual, namely: 1867 * * SACK 1868 * * Duplicate ACK. 1869 * * ECN ECE. 1870 * 1871 * Counting packets in flight is pretty simple. 1872 * 1873 * in_flight = packets_out - left_out + retrans_out 1874 * 1875 * packets_out is SND.NXT-SND.UNA counted in packets. 1876 * 1877 * retrans_out is number of retransmitted segments. 1878 * 1879 * left_out is number of segments left network, but not ACKed yet. 1880 * 1881 * left_out = sacked_out + lost_out 1882 * 1883 * sacked_out: Packets, which arrived to receiver out of order 1884 * and hence not ACKed. With SACKs this number is simply 1885 * amount of SACKed data. Even without SACKs 1886 * it is easy to give pretty reliable estimate of this number, 1887 * counting duplicate ACKs. 1888 * 1889 * lost_out: Packets lost by network. TCP has no explicit 1890 * "loss notification" feedback from network (for now). 1891 * It means that this number can be only _guessed_. 1892 * Actually, it is the heuristics to predict lossage that 1893 * distinguishes different algorithms. 1894 * 1895 * F.e. after RTO, when all the queue is considered as lost, 1896 * lost_out = packets_out and in_flight = retrans_out. 1897 * 1898 * Essentially, we have now two algorithms counting 1899 * lost packets. 1900 * 1901 * FACK: It is the simplest heuristics. As soon as we decided 1902 * that something is lost, we decide that _all_ not SACKed 1903 * packets until the most forward SACK are lost. I.e. 1904 * lost_out = fackets_out - sacked_out and left_out = fackets_out. 1905 * It is absolutely correct estimate, if network does not reorder 1906 * packets. And it loses any connection to reality when reordering 1907 * takes place. We use FACK by default until reordering 1908 * is suspected on the path to this destination. 1909 * 1910 * NewReno: when Recovery is entered, we assume that one segment 1911 * is lost (classic Reno). While we are in Recovery and 1912 * a partial ACK arrives, we assume that one more packet 1913 * is lost (NewReno). This heuristics are the same in NewReno 1914 * and SACK. 1915 * 1916 * Imagine, that's all! Forget about all this shamanism about CWND inflation 1917 * deflation etc. CWND is real congestion window, never inflated, changes 1918 * only according to classic VJ rules. 1919 * 1920 * Really tricky (and requiring careful tuning) part of algorithm 1921 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue(). 1922 * The first determines the moment _when_ we should reduce CWND and, 1923 * hence, slow down forward transmission. In fact, it determines the moment 1924 * when we decide that hole is caused by loss, rather than by a reorder. 1925 * 1926 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill 1927 * holes, caused by lost packets. 1928 * 1929 * And the most logically complicated part of algorithm is undo 1930 * heuristics. We detect false retransmits due to both too early 1931 * fast retransmit (reordering) and underestimated RTO, analyzing 1932 * timestamps and D-SACKs. When we detect that some segments were 1933 * retransmitted by mistake and CWND reduction was wrong, we undo 1934 * window reduction and abort recovery phase. This logic is hidden 1935 * inside several functions named tcp_try_undo_<something>. 1936 */ 1937 1938 /* This function decides, when we should leave Disordered state 1939 * and enter Recovery phase, reducing congestion window. 1940 * 1941 * Main question: may we further continue forward transmission 1942 * with the same cwnd? 1943 */ 1944 static int tcp_time_to_recover(struct sock *sk) 1945 { 1946 struct tcp_sock *tp = tcp_sk(sk); 1947 __u32 packets_out; 1948 1949 /* Do not perform any recovery during FRTO algorithm */ 1950 if (tp->frto_counter) 1951 return 0; 1952 1953 /* Trick#1: The loss is proven. */ 1954 if (tp->lost_out) 1955 return 1; 1956 1957 /* Not-A-Trick#2 : Classic rule... */ 1958 if (tcp_fackets_out(tp) > tp->reordering) 1959 return 1; 1960 1961 /* Trick#3 : when we use RFC2988 timer restart, fast 1962 * retransmit can be triggered by timeout of queue head. 1963 */ 1964 if (tcp_head_timedout(sk)) 1965 return 1; 1966 1967 /* Trick#4: It is still not OK... But will it be useful to delay 1968 * recovery more? 1969 */ 1970 packets_out = tp->packets_out; 1971 if (packets_out <= tp->reordering && 1972 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) && 1973 !tcp_may_send_now(sk)) { 1974 /* We have nothing to send. This connection is limited 1975 * either by receiver window or by application. 1976 */ 1977 return 1; 1978 } 1979 1980 return 0; 1981 } 1982 1983 /* RFC: This is from the original, I doubt that this is necessary at all: 1984 * clear xmit_retrans hint if seq of this skb is beyond hint. How could we 1985 * retransmitted past LOST markings in the first place? I'm not fully sure 1986 * about undo and end of connection cases, which can cause R without L? 1987 */ 1988 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, 1989 struct sk_buff *skb) 1990 { 1991 if ((tp->retransmit_skb_hint != NULL) && 1992 before(TCP_SKB_CB(skb)->seq, 1993 TCP_SKB_CB(tp->retransmit_skb_hint)->seq)) 1994 tp->retransmit_skb_hint = NULL; 1995 } 1996 1997 /* Mark head of queue up as lost. */ 1998 static void tcp_mark_head_lost(struct sock *sk, int packets) 1999 { 2000 struct tcp_sock *tp = tcp_sk(sk); 2001 struct sk_buff *skb; 2002 int cnt; 2003 2004 BUG_TRAP(packets <= tp->packets_out); 2005 if (tp->lost_skb_hint) { 2006 skb = tp->lost_skb_hint; 2007 cnt = tp->lost_cnt_hint; 2008 } else { 2009 skb = tcp_write_queue_head(sk); 2010 cnt = 0; 2011 } 2012 2013 tcp_for_write_queue_from(skb, sk) { 2014 if (skb == tcp_send_head(sk)) 2015 break; 2016 /* TODO: do this better */ 2017 /* this is not the most efficient way to do this... */ 2018 tp->lost_skb_hint = skb; 2019 tp->lost_cnt_hint = cnt; 2020 cnt += tcp_skb_pcount(skb); 2021 if (cnt > packets || after(TCP_SKB_CB(skb)->end_seq, tp->high_seq)) 2022 break; 2023 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_SACKED_ACKED|TCPCB_LOST))) { 2024 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 2025 tp->lost_out += tcp_skb_pcount(skb); 2026 tcp_verify_retransmit_hint(tp, skb); 2027 } 2028 } 2029 tcp_verify_left_out(tp); 2030 } 2031 2032 /* Account newly detected lost packet(s) */ 2033 2034 static void tcp_update_scoreboard(struct sock *sk) 2035 { 2036 struct tcp_sock *tp = tcp_sk(sk); 2037 2038 if (tcp_is_fack(tp)) { 2039 int lost = tp->fackets_out - tp->reordering; 2040 if (lost <= 0) 2041 lost = 1; 2042 tcp_mark_head_lost(sk, lost); 2043 } else { 2044 tcp_mark_head_lost(sk, 1); 2045 } 2046 2047 /* New heuristics: it is possible only after we switched 2048 * to restart timer each time when something is ACKed. 2049 * Hence, we can detect timed out packets during fast 2050 * retransmit without falling to slow start. 2051 */ 2052 if (!tcp_is_reno(tp) && tcp_head_timedout(sk)) { 2053 struct sk_buff *skb; 2054 2055 skb = tp->scoreboard_skb_hint ? tp->scoreboard_skb_hint 2056 : tcp_write_queue_head(sk); 2057 2058 tcp_for_write_queue_from(skb, sk) { 2059 if (skb == tcp_send_head(sk)) 2060 break; 2061 if (!tcp_skb_timedout(sk, skb)) 2062 break; 2063 2064 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) { 2065 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 2066 tp->lost_out += tcp_skb_pcount(skb); 2067 tcp_verify_retransmit_hint(tp, skb); 2068 } 2069 } 2070 2071 tp->scoreboard_skb_hint = skb; 2072 2073 tcp_verify_left_out(tp); 2074 } 2075 } 2076 2077 /* CWND moderation, preventing bursts due to too big ACKs 2078 * in dubious situations. 2079 */ 2080 static inline void tcp_moderate_cwnd(struct tcp_sock *tp) 2081 { 2082 tp->snd_cwnd = min(tp->snd_cwnd, 2083 tcp_packets_in_flight(tp)+tcp_max_burst(tp)); 2084 tp->snd_cwnd_stamp = tcp_time_stamp; 2085 } 2086 2087 /* Lower bound on congestion window is slow start threshold 2088 * unless congestion avoidance choice decides to overide it. 2089 */ 2090 static inline u32 tcp_cwnd_min(const struct sock *sk) 2091 { 2092 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 2093 2094 return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh; 2095 } 2096 2097 /* Decrease cwnd each second ack. */ 2098 static void tcp_cwnd_down(struct sock *sk, int flag) 2099 { 2100 struct tcp_sock *tp = tcp_sk(sk); 2101 int decr = tp->snd_cwnd_cnt + 1; 2102 2103 if ((flag&(FLAG_ANY_PROGRESS|FLAG_DSACKING_ACK)) || 2104 (tcp_is_reno(tp) && !(flag&FLAG_NOT_DUP))) { 2105 tp->snd_cwnd_cnt = decr&1; 2106 decr >>= 1; 2107 2108 if (decr && tp->snd_cwnd > tcp_cwnd_min(sk)) 2109 tp->snd_cwnd -= decr; 2110 2111 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp)+1); 2112 tp->snd_cwnd_stamp = tcp_time_stamp; 2113 } 2114 } 2115 2116 /* Nothing was retransmitted or returned timestamp is less 2117 * than timestamp of the first retransmission. 2118 */ 2119 static inline int tcp_packet_delayed(struct tcp_sock *tp) 2120 { 2121 return !tp->retrans_stamp || 2122 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 2123 (__s32)(tp->rx_opt.rcv_tsecr - tp->retrans_stamp) < 0); 2124 } 2125 2126 /* Undo procedures. */ 2127 2128 #if FASTRETRANS_DEBUG > 1 2129 static void DBGUNDO(struct sock *sk, const char *msg) 2130 { 2131 struct tcp_sock *tp = tcp_sk(sk); 2132 struct inet_sock *inet = inet_sk(sk); 2133 2134 printk(KERN_DEBUG "Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n", 2135 msg, 2136 NIPQUAD(inet->daddr), ntohs(inet->dport), 2137 tp->snd_cwnd, tcp_left_out(tp), 2138 tp->snd_ssthresh, tp->prior_ssthresh, 2139 tp->packets_out); 2140 } 2141 #else 2142 #define DBGUNDO(x...) do { } while (0) 2143 #endif 2144 2145 static void tcp_undo_cwr(struct sock *sk, const int undo) 2146 { 2147 struct tcp_sock *tp = tcp_sk(sk); 2148 2149 if (tp->prior_ssthresh) { 2150 const struct inet_connection_sock *icsk = inet_csk(sk); 2151 2152 if (icsk->icsk_ca_ops->undo_cwnd) 2153 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk); 2154 else 2155 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh<<1); 2156 2157 if (undo && tp->prior_ssthresh > tp->snd_ssthresh) { 2158 tp->snd_ssthresh = tp->prior_ssthresh; 2159 TCP_ECN_withdraw_cwr(tp); 2160 } 2161 } else { 2162 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh); 2163 } 2164 tcp_moderate_cwnd(tp); 2165 tp->snd_cwnd_stamp = tcp_time_stamp; 2166 2167 /* There is something screwy going on with the retrans hints after 2168 an undo */ 2169 tcp_clear_all_retrans_hints(tp); 2170 } 2171 2172 static inline int tcp_may_undo(struct tcp_sock *tp) 2173 { 2174 return tp->undo_marker && 2175 (!tp->undo_retrans || tcp_packet_delayed(tp)); 2176 } 2177 2178 /* People celebrate: "We love our President!" */ 2179 static int tcp_try_undo_recovery(struct sock *sk) 2180 { 2181 struct tcp_sock *tp = tcp_sk(sk); 2182 2183 if (tcp_may_undo(tp)) { 2184 /* Happy end! We did not retransmit anything 2185 * or our original transmission succeeded. 2186 */ 2187 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans"); 2188 tcp_undo_cwr(sk, 1); 2189 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss) 2190 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO); 2191 else 2192 NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO); 2193 tp->undo_marker = 0; 2194 } 2195 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) { 2196 /* Hold old state until something *above* high_seq 2197 * is ACKed. For Reno it is MUST to prevent false 2198 * fast retransmits (RFC2582). SACK TCP is safe. */ 2199 tcp_moderate_cwnd(tp); 2200 return 1; 2201 } 2202 tcp_set_ca_state(sk, TCP_CA_Open); 2203 return 0; 2204 } 2205 2206 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */ 2207 static void tcp_try_undo_dsack(struct sock *sk) 2208 { 2209 struct tcp_sock *tp = tcp_sk(sk); 2210 2211 if (tp->undo_marker && !tp->undo_retrans) { 2212 DBGUNDO(sk, "D-SACK"); 2213 tcp_undo_cwr(sk, 1); 2214 tp->undo_marker = 0; 2215 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO); 2216 } 2217 } 2218 2219 /* Undo during fast recovery after partial ACK. */ 2220 2221 static int tcp_try_undo_partial(struct sock *sk, int acked) 2222 { 2223 struct tcp_sock *tp = tcp_sk(sk); 2224 /* Partial ACK arrived. Force Hoe's retransmit. */ 2225 int failed = tcp_is_reno(tp) || tp->fackets_out>tp->reordering; 2226 2227 if (tcp_may_undo(tp)) { 2228 /* Plain luck! Hole if filled with delayed 2229 * packet, rather than with a retransmit. 2230 */ 2231 if (tp->retrans_out == 0) 2232 tp->retrans_stamp = 0; 2233 2234 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1); 2235 2236 DBGUNDO(sk, "Hoe"); 2237 tcp_undo_cwr(sk, 0); 2238 NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO); 2239 2240 /* So... Do not make Hoe's retransmit yet. 2241 * If the first packet was delayed, the rest 2242 * ones are most probably delayed as well. 2243 */ 2244 failed = 0; 2245 } 2246 return failed; 2247 } 2248 2249 /* Undo during loss recovery after partial ACK. */ 2250 static int tcp_try_undo_loss(struct sock *sk) 2251 { 2252 struct tcp_sock *tp = tcp_sk(sk); 2253 2254 if (tcp_may_undo(tp)) { 2255 struct sk_buff *skb; 2256 tcp_for_write_queue(skb, sk) { 2257 if (skb == tcp_send_head(sk)) 2258 break; 2259 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; 2260 } 2261 2262 tcp_clear_all_retrans_hints(tp); 2263 2264 DBGUNDO(sk, "partial loss"); 2265 tp->lost_out = 0; 2266 tcp_undo_cwr(sk, 1); 2267 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO); 2268 inet_csk(sk)->icsk_retransmits = 0; 2269 tp->undo_marker = 0; 2270 if (tcp_is_sack(tp)) 2271 tcp_set_ca_state(sk, TCP_CA_Open); 2272 return 1; 2273 } 2274 return 0; 2275 } 2276 2277 static inline void tcp_complete_cwr(struct sock *sk) 2278 { 2279 struct tcp_sock *tp = tcp_sk(sk); 2280 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh); 2281 tp->snd_cwnd_stamp = tcp_time_stamp; 2282 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR); 2283 } 2284 2285 static void tcp_try_to_open(struct sock *sk, int flag) 2286 { 2287 struct tcp_sock *tp = tcp_sk(sk); 2288 2289 tcp_verify_left_out(tp); 2290 2291 if (tp->retrans_out == 0) 2292 tp->retrans_stamp = 0; 2293 2294 if (flag&FLAG_ECE) 2295 tcp_enter_cwr(sk, 1); 2296 2297 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) { 2298 int state = TCP_CA_Open; 2299 2300 if (tcp_left_out(tp) || tp->retrans_out || tp->undo_marker) 2301 state = TCP_CA_Disorder; 2302 2303 if (inet_csk(sk)->icsk_ca_state != state) { 2304 tcp_set_ca_state(sk, state); 2305 tp->high_seq = tp->snd_nxt; 2306 } 2307 tcp_moderate_cwnd(tp); 2308 } else { 2309 tcp_cwnd_down(sk, flag); 2310 } 2311 } 2312 2313 static void tcp_mtup_probe_failed(struct sock *sk) 2314 { 2315 struct inet_connection_sock *icsk = inet_csk(sk); 2316 2317 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1; 2318 icsk->icsk_mtup.probe_size = 0; 2319 } 2320 2321 static void tcp_mtup_probe_success(struct sock *sk, struct sk_buff *skb) 2322 { 2323 struct tcp_sock *tp = tcp_sk(sk); 2324 struct inet_connection_sock *icsk = inet_csk(sk); 2325 2326 /* FIXME: breaks with very large cwnd */ 2327 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2328 tp->snd_cwnd = tp->snd_cwnd * 2329 tcp_mss_to_mtu(sk, tp->mss_cache) / 2330 icsk->icsk_mtup.probe_size; 2331 tp->snd_cwnd_cnt = 0; 2332 tp->snd_cwnd_stamp = tcp_time_stamp; 2333 tp->rcv_ssthresh = tcp_current_ssthresh(sk); 2334 2335 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size; 2336 icsk->icsk_mtup.probe_size = 0; 2337 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 2338 } 2339 2340 2341 /* Process an event, which can update packets-in-flight not trivially. 2342 * Main goal of this function is to calculate new estimate for left_out, 2343 * taking into account both packets sitting in receiver's buffer and 2344 * packets lost by network. 2345 * 2346 * Besides that it does CWND reduction, when packet loss is detected 2347 * and changes state of machine. 2348 * 2349 * It does _not_ decide what to send, it is made in function 2350 * tcp_xmit_retransmit_queue(). 2351 */ 2352 static void 2353 tcp_fastretrans_alert(struct sock *sk, int pkts_acked, int flag) 2354 { 2355 struct inet_connection_sock *icsk = inet_csk(sk); 2356 struct tcp_sock *tp = tcp_sk(sk); 2357 int is_dupack = !(flag&(FLAG_SND_UNA_ADVANCED|FLAG_NOT_DUP)); 2358 int do_lost = is_dupack || ((flag&FLAG_DATA_SACKED) && 2359 (tp->fackets_out > tp->reordering)); 2360 2361 /* Some technical things: 2362 * 1. Reno does not count dupacks (sacked_out) automatically. */ 2363 if (!tp->packets_out) 2364 tp->sacked_out = 0; 2365 2366 if (WARN_ON(!tp->sacked_out && tp->fackets_out)) 2367 tp->fackets_out = 0; 2368 2369 /* Now state machine starts. 2370 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */ 2371 if (flag&FLAG_ECE) 2372 tp->prior_ssthresh = 0; 2373 2374 /* B. In all the states check for reneging SACKs. */ 2375 if (tp->sacked_out && tcp_check_sack_reneging(sk)) 2376 return; 2377 2378 /* C. Process data loss notification, provided it is valid. */ 2379 if ((flag&FLAG_DATA_LOST) && 2380 before(tp->snd_una, tp->high_seq) && 2381 icsk->icsk_ca_state != TCP_CA_Open && 2382 tp->fackets_out > tp->reordering) { 2383 tcp_mark_head_lost(sk, tp->fackets_out - tp->reordering); 2384 NET_INC_STATS_BH(LINUX_MIB_TCPLOSS); 2385 } 2386 2387 /* D. Check consistency of the current state. */ 2388 tcp_verify_left_out(tp); 2389 2390 /* E. Check state exit conditions. State can be terminated 2391 * when high_seq is ACKed. */ 2392 if (icsk->icsk_ca_state == TCP_CA_Open) { 2393 BUG_TRAP(tp->retrans_out == 0); 2394 tp->retrans_stamp = 0; 2395 } else if (!before(tp->snd_una, tp->high_seq)) { 2396 switch (icsk->icsk_ca_state) { 2397 case TCP_CA_Loss: 2398 icsk->icsk_retransmits = 0; 2399 if (tcp_try_undo_recovery(sk)) 2400 return; 2401 break; 2402 2403 case TCP_CA_CWR: 2404 /* CWR is to be held something *above* high_seq 2405 * is ACKed for CWR bit to reach receiver. */ 2406 if (tp->snd_una != tp->high_seq) { 2407 tcp_complete_cwr(sk); 2408 tcp_set_ca_state(sk, TCP_CA_Open); 2409 } 2410 break; 2411 2412 case TCP_CA_Disorder: 2413 tcp_try_undo_dsack(sk); 2414 if (!tp->undo_marker || 2415 /* For SACK case do not Open to allow to undo 2416 * catching for all duplicate ACKs. */ 2417 tcp_is_reno(tp) || tp->snd_una != tp->high_seq) { 2418 tp->undo_marker = 0; 2419 tcp_set_ca_state(sk, TCP_CA_Open); 2420 } 2421 break; 2422 2423 case TCP_CA_Recovery: 2424 if (tcp_is_reno(tp)) 2425 tcp_reset_reno_sack(tp); 2426 if (tcp_try_undo_recovery(sk)) 2427 return; 2428 tcp_complete_cwr(sk); 2429 break; 2430 } 2431 } 2432 2433 /* F. Process state. */ 2434 switch (icsk->icsk_ca_state) { 2435 case TCP_CA_Recovery: 2436 if (!(flag & FLAG_SND_UNA_ADVANCED)) { 2437 if (tcp_is_reno(tp) && is_dupack) 2438 tcp_add_reno_sack(sk); 2439 } else 2440 do_lost = tcp_try_undo_partial(sk, pkts_acked); 2441 break; 2442 case TCP_CA_Loss: 2443 if (flag&FLAG_DATA_ACKED) 2444 icsk->icsk_retransmits = 0; 2445 if (!tcp_try_undo_loss(sk)) { 2446 tcp_moderate_cwnd(tp); 2447 tcp_xmit_retransmit_queue(sk); 2448 return; 2449 } 2450 if (icsk->icsk_ca_state != TCP_CA_Open) 2451 return; 2452 /* Loss is undone; fall through to processing in Open state. */ 2453 default: 2454 if (tcp_is_reno(tp)) { 2455 if (flag & FLAG_SND_UNA_ADVANCED) 2456 tcp_reset_reno_sack(tp); 2457 if (is_dupack) 2458 tcp_add_reno_sack(sk); 2459 } 2460 2461 if (icsk->icsk_ca_state == TCP_CA_Disorder) 2462 tcp_try_undo_dsack(sk); 2463 2464 if (!tcp_time_to_recover(sk)) { 2465 tcp_try_to_open(sk, flag); 2466 return; 2467 } 2468 2469 /* MTU probe failure: don't reduce cwnd */ 2470 if (icsk->icsk_ca_state < TCP_CA_CWR && 2471 icsk->icsk_mtup.probe_size && 2472 tp->snd_una == tp->mtu_probe.probe_seq_start) { 2473 tcp_mtup_probe_failed(sk); 2474 /* Restores the reduction we did in tcp_mtup_probe() */ 2475 tp->snd_cwnd++; 2476 tcp_simple_retransmit(sk); 2477 return; 2478 } 2479 2480 /* Otherwise enter Recovery state */ 2481 2482 if (tcp_is_reno(tp)) 2483 NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY); 2484 else 2485 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY); 2486 2487 tp->high_seq = tp->snd_nxt; 2488 tp->prior_ssthresh = 0; 2489 tp->undo_marker = tp->snd_una; 2490 tp->undo_retrans = tp->retrans_out; 2491 2492 if (icsk->icsk_ca_state < TCP_CA_CWR) { 2493 if (!(flag&FLAG_ECE)) 2494 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2495 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 2496 TCP_ECN_queue_cwr(tp); 2497 } 2498 2499 tp->bytes_acked = 0; 2500 tp->snd_cwnd_cnt = 0; 2501 tcp_set_ca_state(sk, TCP_CA_Recovery); 2502 } 2503 2504 if (do_lost || tcp_head_timedout(sk)) 2505 tcp_update_scoreboard(sk); 2506 tcp_cwnd_down(sk, flag); 2507 tcp_xmit_retransmit_queue(sk); 2508 } 2509 2510 /* Read draft-ietf-tcplw-high-performance before mucking 2511 * with this code. (Supersedes RFC1323) 2512 */ 2513 static void tcp_ack_saw_tstamp(struct sock *sk, int flag) 2514 { 2515 /* RTTM Rule: A TSecr value received in a segment is used to 2516 * update the averaged RTT measurement only if the segment 2517 * acknowledges some new data, i.e., only if it advances the 2518 * left edge of the send window. 2519 * 2520 * See draft-ietf-tcplw-high-performance-00, section 3.3. 2521 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru> 2522 * 2523 * Changed: reset backoff as soon as we see the first valid sample. 2524 * If we do not, we get strongly overestimated rto. With timestamps 2525 * samples are accepted even from very old segments: f.e., when rtt=1 2526 * increases to 8, we retransmit 5 times and after 8 seconds delayed 2527 * answer arrives rto becomes 120 seconds! If at least one of segments 2528 * in window is lost... Voila. --ANK (010210) 2529 */ 2530 struct tcp_sock *tp = tcp_sk(sk); 2531 const __u32 seq_rtt = tcp_time_stamp - tp->rx_opt.rcv_tsecr; 2532 tcp_rtt_estimator(sk, seq_rtt); 2533 tcp_set_rto(sk); 2534 inet_csk(sk)->icsk_backoff = 0; 2535 tcp_bound_rto(sk); 2536 } 2537 2538 static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag) 2539 { 2540 /* We don't have a timestamp. Can only use 2541 * packets that are not retransmitted to determine 2542 * rtt estimates. Also, we must not reset the 2543 * backoff for rto until we get a non-retransmitted 2544 * packet. This allows us to deal with a situation 2545 * where the network delay has increased suddenly. 2546 * I.e. Karn's algorithm. (SIGCOMM '87, p5.) 2547 */ 2548 2549 if (flag & FLAG_RETRANS_DATA_ACKED) 2550 return; 2551 2552 tcp_rtt_estimator(sk, seq_rtt); 2553 tcp_set_rto(sk); 2554 inet_csk(sk)->icsk_backoff = 0; 2555 tcp_bound_rto(sk); 2556 } 2557 2558 static inline void tcp_ack_update_rtt(struct sock *sk, const int flag, 2559 const s32 seq_rtt) 2560 { 2561 const struct tcp_sock *tp = tcp_sk(sk); 2562 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */ 2563 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr) 2564 tcp_ack_saw_tstamp(sk, flag); 2565 else if (seq_rtt >= 0) 2566 tcp_ack_no_tstamp(sk, seq_rtt, flag); 2567 } 2568 2569 static void tcp_cong_avoid(struct sock *sk, u32 ack, 2570 u32 in_flight, int good) 2571 { 2572 const struct inet_connection_sock *icsk = inet_csk(sk); 2573 icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight, good); 2574 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp; 2575 } 2576 2577 /* Restart timer after forward progress on connection. 2578 * RFC2988 recommends to restart timer to now+rto. 2579 */ 2580 static void tcp_rearm_rto(struct sock *sk) 2581 { 2582 struct tcp_sock *tp = tcp_sk(sk); 2583 2584 if (!tp->packets_out) { 2585 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS); 2586 } else { 2587 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 2588 } 2589 } 2590 2591 /* If we get here, the whole TSO packet has not been acked. */ 2592 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb) 2593 { 2594 struct tcp_sock *tp = tcp_sk(sk); 2595 u32 packets_acked; 2596 2597 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)); 2598 2599 packets_acked = tcp_skb_pcount(skb); 2600 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 2601 return 0; 2602 packets_acked -= tcp_skb_pcount(skb); 2603 2604 if (packets_acked) { 2605 BUG_ON(tcp_skb_pcount(skb) == 0); 2606 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)); 2607 } 2608 2609 return packets_acked; 2610 } 2611 2612 /* Remove acknowledged frames from the retransmission queue. If our packet 2613 * is before the ack sequence we can discard it as it's confirmed to have 2614 * arrived at the other end. 2615 */ 2616 static int tcp_clean_rtx_queue(struct sock *sk, s32 *seq_rtt_p) 2617 { 2618 struct tcp_sock *tp = tcp_sk(sk); 2619 const struct inet_connection_sock *icsk = inet_csk(sk); 2620 struct sk_buff *skb; 2621 u32 now = tcp_time_stamp; 2622 int fully_acked = 1; 2623 int flag = 0; 2624 int prior_packets = tp->packets_out; 2625 s32 seq_rtt = -1; 2626 ktime_t last_ackt = net_invalid_timestamp(); 2627 2628 while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) { 2629 struct tcp_skb_cb *scb = TCP_SKB_CB(skb); 2630 u32 end_seq; 2631 u32 packets_acked; 2632 u8 sacked = scb->sacked; 2633 2634 if (after(scb->end_seq, tp->snd_una)) { 2635 if (tcp_skb_pcount(skb) == 1 || 2636 !after(tp->snd_una, scb->seq)) 2637 break; 2638 2639 packets_acked = tcp_tso_acked(sk, skb); 2640 if (!packets_acked) 2641 break; 2642 2643 fully_acked = 0; 2644 end_seq = tp->snd_una; 2645 } else { 2646 packets_acked = tcp_skb_pcount(skb); 2647 end_seq = scb->end_seq; 2648 } 2649 2650 /* MTU probing checks */ 2651 if (fully_acked && icsk->icsk_mtup.probe_size && 2652 !after(tp->mtu_probe.probe_seq_end, scb->end_seq)) { 2653 tcp_mtup_probe_success(sk, skb); 2654 } 2655 2656 if (sacked) { 2657 if (sacked & TCPCB_RETRANS) { 2658 if (sacked & TCPCB_SACKED_RETRANS) 2659 tp->retrans_out -= packets_acked; 2660 flag |= FLAG_RETRANS_DATA_ACKED; 2661 seq_rtt = -1; 2662 if ((flag & FLAG_DATA_ACKED) || 2663 (packets_acked > 1)) 2664 flag |= FLAG_NONHEAD_RETRANS_ACKED; 2665 } else if (seq_rtt < 0) { 2666 seq_rtt = now - scb->when; 2667 if (fully_acked) 2668 last_ackt = skb->tstamp; 2669 } 2670 2671 if (sacked & TCPCB_SACKED_ACKED) 2672 tp->sacked_out -= packets_acked; 2673 if (sacked & TCPCB_LOST) 2674 tp->lost_out -= packets_acked; 2675 2676 if ((sacked & TCPCB_URG) && tp->urg_mode && 2677 !before(end_seq, tp->snd_up)) 2678 tp->urg_mode = 0; 2679 } else if (seq_rtt < 0) { 2680 seq_rtt = now - scb->when; 2681 if (fully_acked) 2682 last_ackt = skb->tstamp; 2683 } 2684 tp->packets_out -= packets_acked; 2685 2686 /* Initial outgoing SYN's get put onto the write_queue 2687 * just like anything else we transmit. It is not 2688 * true data, and if we misinform our callers that 2689 * this ACK acks real data, we will erroneously exit 2690 * connection startup slow start one packet too 2691 * quickly. This is severely frowned upon behavior. 2692 */ 2693 if (!(scb->flags & TCPCB_FLAG_SYN)) { 2694 flag |= FLAG_DATA_ACKED; 2695 } else { 2696 flag |= FLAG_SYN_ACKED; 2697 tp->retrans_stamp = 0; 2698 } 2699 2700 if (!fully_acked) 2701 break; 2702 2703 tcp_unlink_write_queue(skb, sk); 2704 sk_stream_free_skb(sk, skb); 2705 tcp_clear_all_retrans_hints(tp); 2706 } 2707 2708 if (flag & FLAG_ACKED) { 2709 u32 pkts_acked = prior_packets - tp->packets_out; 2710 const struct tcp_congestion_ops *ca_ops 2711 = inet_csk(sk)->icsk_ca_ops; 2712 2713 tcp_ack_update_rtt(sk, flag, seq_rtt); 2714 tcp_rearm_rto(sk); 2715 2716 tp->fackets_out -= min(pkts_acked, tp->fackets_out); 2717 /* hint's skb might be NULL but we don't need to care */ 2718 tp->fastpath_cnt_hint -= min_t(u32, pkts_acked, 2719 tp->fastpath_cnt_hint); 2720 if (tcp_is_reno(tp)) 2721 tcp_remove_reno_sacks(sk, pkts_acked); 2722 2723 if (ca_ops->pkts_acked) { 2724 s32 rtt_us = -1; 2725 2726 /* Is the ACK triggering packet unambiguous? */ 2727 if (!(flag & FLAG_RETRANS_DATA_ACKED)) { 2728 /* High resolution needed and available? */ 2729 if (ca_ops->flags & TCP_CONG_RTT_STAMP && 2730 !ktime_equal(last_ackt, 2731 net_invalid_timestamp())) 2732 rtt_us = ktime_us_delta(ktime_get_real(), 2733 last_ackt); 2734 else if (seq_rtt > 0) 2735 rtt_us = jiffies_to_usecs(seq_rtt); 2736 } 2737 2738 ca_ops->pkts_acked(sk, pkts_acked, rtt_us); 2739 } 2740 } 2741 2742 #if FASTRETRANS_DEBUG > 0 2743 BUG_TRAP((int)tp->sacked_out >= 0); 2744 BUG_TRAP((int)tp->lost_out >= 0); 2745 BUG_TRAP((int)tp->retrans_out >= 0); 2746 if (!tp->packets_out && tcp_is_sack(tp)) { 2747 icsk = inet_csk(sk); 2748 if (tp->lost_out) { 2749 printk(KERN_DEBUG "Leak l=%u %d\n", 2750 tp->lost_out, icsk->icsk_ca_state); 2751 tp->lost_out = 0; 2752 } 2753 if (tp->sacked_out) { 2754 printk(KERN_DEBUG "Leak s=%u %d\n", 2755 tp->sacked_out, icsk->icsk_ca_state); 2756 tp->sacked_out = 0; 2757 } 2758 if (tp->retrans_out) { 2759 printk(KERN_DEBUG "Leak r=%u %d\n", 2760 tp->retrans_out, icsk->icsk_ca_state); 2761 tp->retrans_out = 0; 2762 } 2763 } 2764 #endif 2765 *seq_rtt_p = seq_rtt; 2766 return flag; 2767 } 2768 2769 static void tcp_ack_probe(struct sock *sk) 2770 { 2771 const struct tcp_sock *tp = tcp_sk(sk); 2772 struct inet_connection_sock *icsk = inet_csk(sk); 2773 2774 /* Was it a usable window open? */ 2775 2776 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, 2777 tp->snd_una + tp->snd_wnd)) { 2778 icsk->icsk_backoff = 0; 2779 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0); 2780 /* Socket must be waked up by subsequent tcp_data_snd_check(). 2781 * This function is not for random using! 2782 */ 2783 } else { 2784 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 2785 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), 2786 TCP_RTO_MAX); 2787 } 2788 } 2789 2790 static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag) 2791 { 2792 return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) || 2793 inet_csk(sk)->icsk_ca_state != TCP_CA_Open); 2794 } 2795 2796 static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag) 2797 { 2798 const struct tcp_sock *tp = tcp_sk(sk); 2799 return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) && 2800 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR)); 2801 } 2802 2803 /* Check that window update is acceptable. 2804 * The function assumes that snd_una<=ack<=snd_next. 2805 */ 2806 static inline int tcp_may_update_window(const struct tcp_sock *tp, const u32 ack, 2807 const u32 ack_seq, const u32 nwin) 2808 { 2809 return (after(ack, tp->snd_una) || 2810 after(ack_seq, tp->snd_wl1) || 2811 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd)); 2812 } 2813 2814 /* Update our send window. 2815 * 2816 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2 2817 * and in FreeBSD. NetBSD's one is even worse.) is wrong. 2818 */ 2819 static int tcp_ack_update_window(struct sock *sk, struct sk_buff *skb, u32 ack, 2820 u32 ack_seq) 2821 { 2822 struct tcp_sock *tp = tcp_sk(sk); 2823 int flag = 0; 2824 u32 nwin = ntohs(tcp_hdr(skb)->window); 2825 2826 if (likely(!tcp_hdr(skb)->syn)) 2827 nwin <<= tp->rx_opt.snd_wscale; 2828 2829 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) { 2830 flag |= FLAG_WIN_UPDATE; 2831 tcp_update_wl(tp, ack, ack_seq); 2832 2833 if (tp->snd_wnd != nwin) { 2834 tp->snd_wnd = nwin; 2835 2836 /* Note, it is the only place, where 2837 * fast path is recovered for sending TCP. 2838 */ 2839 tp->pred_flags = 0; 2840 tcp_fast_path_check(sk); 2841 2842 if (nwin > tp->max_window) { 2843 tp->max_window = nwin; 2844 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie); 2845 } 2846 } 2847 } 2848 2849 tp->snd_una = ack; 2850 2851 return flag; 2852 } 2853 2854 /* A very conservative spurious RTO response algorithm: reduce cwnd and 2855 * continue in congestion avoidance. 2856 */ 2857 static void tcp_conservative_spur_to_response(struct tcp_sock *tp) 2858 { 2859 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh); 2860 tp->snd_cwnd_cnt = 0; 2861 tp->bytes_acked = 0; 2862 TCP_ECN_queue_cwr(tp); 2863 tcp_moderate_cwnd(tp); 2864 } 2865 2866 /* A conservative spurious RTO response algorithm: reduce cwnd using 2867 * rate halving and continue in congestion avoidance. 2868 */ 2869 static void tcp_ratehalving_spur_to_response(struct sock *sk) 2870 { 2871 tcp_enter_cwr(sk, 0); 2872 } 2873 2874 static void tcp_undo_spur_to_response(struct sock *sk, int flag) 2875 { 2876 if (flag&FLAG_ECE) 2877 tcp_ratehalving_spur_to_response(sk); 2878 else 2879 tcp_undo_cwr(sk, 1); 2880 } 2881 2882 /* F-RTO spurious RTO detection algorithm (RFC4138) 2883 * 2884 * F-RTO affects during two new ACKs following RTO (well, almost, see inline 2885 * comments). State (ACK number) is kept in frto_counter. When ACK advances 2886 * window (but not to or beyond highest sequence sent before RTO): 2887 * On First ACK, send two new segments out. 2888 * On Second ACK, RTO was likely spurious. Do spurious response (response 2889 * algorithm is not part of the F-RTO detection algorithm 2890 * given in RFC4138 but can be selected separately). 2891 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss 2892 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding 2893 * of Nagle, this is done using frto_counter states 2 and 3, when a new data 2894 * segment of any size sent during F-RTO, state 2 is upgraded to 3. 2895 * 2896 * Rationale: if the RTO was spurious, new ACKs should arrive from the 2897 * original window even after we transmit two new data segments. 2898 * 2899 * SACK version: 2900 * on first step, wait until first cumulative ACK arrives, then move to 2901 * the second step. In second step, the next ACK decides. 2902 * 2903 * F-RTO is implemented (mainly) in four functions: 2904 * - tcp_use_frto() is used to determine if TCP is can use F-RTO 2905 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is 2906 * called when tcp_use_frto() showed green light 2907 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm 2908 * - tcp_enter_frto_loss() is called if there is not enough evidence 2909 * to prove that the RTO is indeed spurious. It transfers the control 2910 * from F-RTO to the conventional RTO recovery 2911 */ 2912 static int tcp_process_frto(struct sock *sk, int flag) 2913 { 2914 struct tcp_sock *tp = tcp_sk(sk); 2915 2916 tcp_verify_left_out(tp); 2917 2918 /* Duplicate the behavior from Loss state (fastretrans_alert) */ 2919 if (flag&FLAG_DATA_ACKED) 2920 inet_csk(sk)->icsk_retransmits = 0; 2921 2922 if ((flag & FLAG_NONHEAD_RETRANS_ACKED) || 2923 ((tp->frto_counter >= 2) && (flag & FLAG_RETRANS_DATA_ACKED))) 2924 tp->undo_marker = 0; 2925 2926 if (!before(tp->snd_una, tp->frto_highmark)) { 2927 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag); 2928 return 1; 2929 } 2930 2931 if (!IsSackFrto() || tcp_is_reno(tp)) { 2932 /* RFC4138 shortcoming in step 2; should also have case c): 2933 * ACK isn't duplicate nor advances window, e.g., opposite dir 2934 * data, winupdate 2935 */ 2936 if (!(flag&FLAG_ANY_PROGRESS) && (flag&FLAG_NOT_DUP)) 2937 return 1; 2938 2939 if (!(flag&FLAG_DATA_ACKED)) { 2940 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3), 2941 flag); 2942 return 1; 2943 } 2944 } else { 2945 if (!(flag&FLAG_DATA_ACKED) && (tp->frto_counter == 1)) { 2946 /* Prevent sending of new data. */ 2947 tp->snd_cwnd = min(tp->snd_cwnd, 2948 tcp_packets_in_flight(tp)); 2949 return 1; 2950 } 2951 2952 if ((tp->frto_counter >= 2) && 2953 (!(flag&FLAG_FORWARD_PROGRESS) || 2954 ((flag&FLAG_DATA_SACKED) && !(flag&FLAG_ONLY_ORIG_SACKED)))) { 2955 /* RFC4138 shortcoming (see comment above) */ 2956 if (!(flag&FLAG_FORWARD_PROGRESS) && (flag&FLAG_NOT_DUP)) 2957 return 1; 2958 2959 tcp_enter_frto_loss(sk, 3, flag); 2960 return 1; 2961 } 2962 } 2963 2964 if (tp->frto_counter == 1) { 2965 /* Sending of the next skb must be allowed or no FRTO */ 2966 if (!tcp_send_head(sk) || 2967 after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, 2968 tp->snd_una + tp->snd_wnd)) { 2969 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), 2970 flag); 2971 return 1; 2972 } 2973 2974 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2; 2975 tp->frto_counter = 2; 2976 return 1; 2977 } else { 2978 switch (sysctl_tcp_frto_response) { 2979 case 2: 2980 tcp_undo_spur_to_response(sk, flag); 2981 break; 2982 case 1: 2983 tcp_conservative_spur_to_response(tp); 2984 break; 2985 default: 2986 tcp_ratehalving_spur_to_response(sk); 2987 break; 2988 } 2989 tp->frto_counter = 0; 2990 tp->undo_marker = 0; 2991 NET_INC_STATS_BH(LINUX_MIB_TCPSPURIOUSRTOS); 2992 } 2993 return 0; 2994 } 2995 2996 /* This routine deals with incoming acks, but not outgoing ones. */ 2997 static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag) 2998 { 2999 struct inet_connection_sock *icsk = inet_csk(sk); 3000 struct tcp_sock *tp = tcp_sk(sk); 3001 u32 prior_snd_una = tp->snd_una; 3002 u32 ack_seq = TCP_SKB_CB(skb)->seq; 3003 u32 ack = TCP_SKB_CB(skb)->ack_seq; 3004 u32 prior_in_flight; 3005 s32 seq_rtt; 3006 int prior_packets; 3007 int frto_cwnd = 0; 3008 3009 /* If the ack is newer than sent or older than previous acks 3010 * then we can probably ignore it. 3011 */ 3012 if (after(ack, tp->snd_nxt)) 3013 goto uninteresting_ack; 3014 3015 if (before(ack, prior_snd_una)) 3016 goto old_ack; 3017 3018 if (after(ack, prior_snd_una)) 3019 flag |= FLAG_SND_UNA_ADVANCED; 3020 3021 if (sysctl_tcp_abc) { 3022 if (icsk->icsk_ca_state < TCP_CA_CWR) 3023 tp->bytes_acked += ack - prior_snd_una; 3024 else if (icsk->icsk_ca_state == TCP_CA_Loss) 3025 /* we assume just one segment left network */ 3026 tp->bytes_acked += min(ack - prior_snd_una, tp->mss_cache); 3027 } 3028 3029 if (!(flag&FLAG_SLOWPATH) && after(ack, prior_snd_una)) { 3030 /* Window is constant, pure forward advance. 3031 * No more checks are required. 3032 * Note, we use the fact that SND.UNA>=SND.WL2. 3033 */ 3034 tcp_update_wl(tp, ack, ack_seq); 3035 tp->snd_una = ack; 3036 flag |= FLAG_WIN_UPDATE; 3037 3038 tcp_ca_event(sk, CA_EVENT_FAST_ACK); 3039 3040 NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS); 3041 } else { 3042 if (ack_seq != TCP_SKB_CB(skb)->end_seq) 3043 flag |= FLAG_DATA; 3044 else 3045 NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS); 3046 3047 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq); 3048 3049 if (TCP_SKB_CB(skb)->sacked) 3050 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una); 3051 3052 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb))) 3053 flag |= FLAG_ECE; 3054 3055 tcp_ca_event(sk, CA_EVENT_SLOW_ACK); 3056 } 3057 3058 /* We passed data and got it acked, remove any soft error 3059 * log. Something worked... 3060 */ 3061 sk->sk_err_soft = 0; 3062 tp->rcv_tstamp = tcp_time_stamp; 3063 prior_packets = tp->packets_out; 3064 if (!prior_packets) 3065 goto no_queue; 3066 3067 prior_in_flight = tcp_packets_in_flight(tp); 3068 3069 /* See if we can take anything off of the retransmit queue. */ 3070 flag |= tcp_clean_rtx_queue(sk, &seq_rtt); 3071 3072 /* Guarantee sacktag reordering detection against wrap-arounds */ 3073 if (before(tp->frto_highmark, tp->snd_una)) 3074 tp->frto_highmark = 0; 3075 if (tp->frto_counter) 3076 frto_cwnd = tcp_process_frto(sk, flag); 3077 3078 if (tcp_ack_is_dubious(sk, flag)) { 3079 /* Advance CWND, if state allows this. */ 3080 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd && 3081 tcp_may_raise_cwnd(sk, flag)) 3082 tcp_cong_avoid(sk, ack, prior_in_flight, 0); 3083 tcp_fastretrans_alert(sk, prior_packets - tp->packets_out, flag); 3084 } else { 3085 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd) 3086 tcp_cong_avoid(sk, ack, prior_in_flight, 1); 3087 } 3088 3089 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag&FLAG_NOT_DUP)) 3090 dst_confirm(sk->sk_dst_cache); 3091 3092 return 1; 3093 3094 no_queue: 3095 icsk->icsk_probes_out = 0; 3096 3097 /* If this ack opens up a zero window, clear backoff. It was 3098 * being used to time the probes, and is probably far higher than 3099 * it needs to be for normal retransmission. 3100 */ 3101 if (tcp_send_head(sk)) 3102 tcp_ack_probe(sk); 3103 return 1; 3104 3105 old_ack: 3106 if (TCP_SKB_CB(skb)->sacked) 3107 tcp_sacktag_write_queue(sk, skb, prior_snd_una); 3108 3109 uninteresting_ack: 3110 SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt); 3111 return 0; 3112 } 3113 3114 3115 /* Look for tcp options. Normally only called on SYN and SYNACK packets. 3116 * But, this can also be called on packets in the established flow when 3117 * the fast version below fails. 3118 */ 3119 void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx, int estab) 3120 { 3121 unsigned char *ptr; 3122 struct tcphdr *th = tcp_hdr(skb); 3123 int length=(th->doff*4)-sizeof(struct tcphdr); 3124 3125 ptr = (unsigned char *)(th + 1); 3126 opt_rx->saw_tstamp = 0; 3127 3128 while (length > 0) { 3129 int opcode=*ptr++; 3130 int opsize; 3131 3132 switch (opcode) { 3133 case TCPOPT_EOL: 3134 return; 3135 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */ 3136 length--; 3137 continue; 3138 default: 3139 opsize=*ptr++; 3140 if (opsize < 2) /* "silly options" */ 3141 return; 3142 if (opsize > length) 3143 return; /* don't parse partial options */ 3144 switch (opcode) { 3145 case TCPOPT_MSS: 3146 if (opsize==TCPOLEN_MSS && th->syn && !estab) { 3147 u16 in_mss = ntohs(get_unaligned((__be16 *)ptr)); 3148 if (in_mss) { 3149 if (opt_rx->user_mss && opt_rx->user_mss < in_mss) 3150 in_mss = opt_rx->user_mss; 3151 opt_rx->mss_clamp = in_mss; 3152 } 3153 } 3154 break; 3155 case TCPOPT_WINDOW: 3156 if (opsize==TCPOLEN_WINDOW && th->syn && !estab) 3157 if (sysctl_tcp_window_scaling) { 3158 __u8 snd_wscale = *(__u8 *) ptr; 3159 opt_rx->wscale_ok = 1; 3160 if (snd_wscale > 14) { 3161 if (net_ratelimit()) 3162 printk(KERN_INFO "tcp_parse_options: Illegal window " 3163 "scaling value %d >14 received.\n", 3164 snd_wscale); 3165 snd_wscale = 14; 3166 } 3167 opt_rx->snd_wscale = snd_wscale; 3168 } 3169 break; 3170 case TCPOPT_TIMESTAMP: 3171 if (opsize==TCPOLEN_TIMESTAMP) { 3172 if ((estab && opt_rx->tstamp_ok) || 3173 (!estab && sysctl_tcp_timestamps)) { 3174 opt_rx->saw_tstamp = 1; 3175 opt_rx->rcv_tsval = ntohl(get_unaligned((__be32 *)ptr)); 3176 opt_rx->rcv_tsecr = ntohl(get_unaligned((__be32 *)(ptr+4))); 3177 } 3178 } 3179 break; 3180 case TCPOPT_SACK_PERM: 3181 if (opsize==TCPOLEN_SACK_PERM && th->syn && !estab) { 3182 if (sysctl_tcp_sack) { 3183 opt_rx->sack_ok = 1; 3184 tcp_sack_reset(opt_rx); 3185 } 3186 } 3187 break; 3188 3189 case TCPOPT_SACK: 3190 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) && 3191 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) && 3192 opt_rx->sack_ok) { 3193 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th; 3194 } 3195 break; 3196 #ifdef CONFIG_TCP_MD5SIG 3197 case TCPOPT_MD5SIG: 3198 /* 3199 * The MD5 Hash has already been 3200 * checked (see tcp_v{4,6}_do_rcv()). 3201 */ 3202 break; 3203 #endif 3204 } 3205 3206 ptr+=opsize-2; 3207 length-=opsize; 3208 } 3209 } 3210 } 3211 3212 /* Fast parse options. This hopes to only see timestamps. 3213 * If it is wrong it falls back on tcp_parse_options(). 3214 */ 3215 static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th, 3216 struct tcp_sock *tp) 3217 { 3218 if (th->doff == sizeof(struct tcphdr)>>2) { 3219 tp->rx_opt.saw_tstamp = 0; 3220 return 0; 3221 } else if (tp->rx_opt.tstamp_ok && 3222 th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) { 3223 __be32 *ptr = (__be32 *)(th + 1); 3224 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) 3225 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) { 3226 tp->rx_opt.saw_tstamp = 1; 3227 ++ptr; 3228 tp->rx_opt.rcv_tsval = ntohl(*ptr); 3229 ++ptr; 3230 tp->rx_opt.rcv_tsecr = ntohl(*ptr); 3231 return 1; 3232 } 3233 } 3234 tcp_parse_options(skb, &tp->rx_opt, 1); 3235 return 1; 3236 } 3237 3238 static inline void tcp_store_ts_recent(struct tcp_sock *tp) 3239 { 3240 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval; 3241 tp->rx_opt.ts_recent_stamp = get_seconds(); 3242 } 3243 3244 static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq) 3245 { 3246 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) { 3247 /* PAWS bug workaround wrt. ACK frames, the PAWS discard 3248 * extra check below makes sure this can only happen 3249 * for pure ACK frames. -DaveM 3250 * 3251 * Not only, also it occurs for expired timestamps. 3252 */ 3253 3254 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) >= 0 || 3255 get_seconds() >= tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS) 3256 tcp_store_ts_recent(tp); 3257 } 3258 } 3259 3260 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM 3261 * 3262 * It is not fatal. If this ACK does _not_ change critical state (seqs, window) 3263 * it can pass through stack. So, the following predicate verifies that 3264 * this segment is not used for anything but congestion avoidance or 3265 * fast retransmit. Moreover, we even are able to eliminate most of such 3266 * second order effects, if we apply some small "replay" window (~RTO) 3267 * to timestamp space. 3268 * 3269 * All these measures still do not guarantee that we reject wrapped ACKs 3270 * on networks with high bandwidth, when sequence space is recycled fastly, 3271 * but it guarantees that such events will be very rare and do not affect 3272 * connection seriously. This doesn't look nice, but alas, PAWS is really 3273 * buggy extension. 3274 * 3275 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC 3276 * states that events when retransmit arrives after original data are rare. 3277 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is 3278 * the biggest problem on large power networks even with minor reordering. 3279 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe 3280 * up to bandwidth of 18Gigabit/sec. 8) ] 3281 */ 3282 3283 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb) 3284 { 3285 struct tcp_sock *tp = tcp_sk(sk); 3286 struct tcphdr *th = tcp_hdr(skb); 3287 u32 seq = TCP_SKB_CB(skb)->seq; 3288 u32 ack = TCP_SKB_CB(skb)->ack_seq; 3289 3290 return (/* 1. Pure ACK with correct sequence number. */ 3291 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) && 3292 3293 /* 2. ... and duplicate ACK. */ 3294 ack == tp->snd_una && 3295 3296 /* 3. ... and does not update window. */ 3297 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) && 3298 3299 /* 4. ... and sits in replay window. */ 3300 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ); 3301 } 3302 3303 static inline int tcp_paws_discard(const struct sock *sk, const struct sk_buff *skb) 3304 { 3305 const struct tcp_sock *tp = tcp_sk(sk); 3306 return ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > TCP_PAWS_WINDOW && 3307 get_seconds() < tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS && 3308 !tcp_disordered_ack(sk, skb)); 3309 } 3310 3311 /* Check segment sequence number for validity. 3312 * 3313 * Segment controls are considered valid, if the segment 3314 * fits to the window after truncation to the window. Acceptability 3315 * of data (and SYN, FIN, of course) is checked separately. 3316 * See tcp_data_queue(), for example. 3317 * 3318 * Also, controls (RST is main one) are accepted using RCV.WUP instead 3319 * of RCV.NXT. Peer still did not advance his SND.UNA when we 3320 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP. 3321 * (borrowed from freebsd) 3322 */ 3323 3324 static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq) 3325 { 3326 return !before(end_seq, tp->rcv_wup) && 3327 !after(seq, tp->rcv_nxt + tcp_receive_window(tp)); 3328 } 3329 3330 /* When we get a reset we do this. */ 3331 static void tcp_reset(struct sock *sk) 3332 { 3333 /* We want the right error as BSD sees it (and indeed as we do). */ 3334 switch (sk->sk_state) { 3335 case TCP_SYN_SENT: 3336 sk->sk_err = ECONNREFUSED; 3337 break; 3338 case TCP_CLOSE_WAIT: 3339 sk->sk_err = EPIPE; 3340 break; 3341 case TCP_CLOSE: 3342 return; 3343 default: 3344 sk->sk_err = ECONNRESET; 3345 } 3346 3347 if (!sock_flag(sk, SOCK_DEAD)) 3348 sk->sk_error_report(sk); 3349 3350 tcp_done(sk); 3351 } 3352 3353 /* 3354 * Process the FIN bit. This now behaves as it is supposed to work 3355 * and the FIN takes effect when it is validly part of sequence 3356 * space. Not before when we get holes. 3357 * 3358 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT 3359 * (and thence onto LAST-ACK and finally, CLOSE, we never enter 3360 * TIME-WAIT) 3361 * 3362 * If we are in FINWAIT-1, a received FIN indicates simultaneous 3363 * close and we go into CLOSING (and later onto TIME-WAIT) 3364 * 3365 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT. 3366 */ 3367 static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th) 3368 { 3369 struct tcp_sock *tp = tcp_sk(sk); 3370 3371 inet_csk_schedule_ack(sk); 3372 3373 sk->sk_shutdown |= RCV_SHUTDOWN; 3374 sock_set_flag(sk, SOCK_DONE); 3375 3376 switch (sk->sk_state) { 3377 case TCP_SYN_RECV: 3378 case TCP_ESTABLISHED: 3379 /* Move to CLOSE_WAIT */ 3380 tcp_set_state(sk, TCP_CLOSE_WAIT); 3381 inet_csk(sk)->icsk_ack.pingpong = 1; 3382 break; 3383 3384 case TCP_CLOSE_WAIT: 3385 case TCP_CLOSING: 3386 /* Received a retransmission of the FIN, do 3387 * nothing. 3388 */ 3389 break; 3390 case TCP_LAST_ACK: 3391 /* RFC793: Remain in the LAST-ACK state. */ 3392 break; 3393 3394 case TCP_FIN_WAIT1: 3395 /* This case occurs when a simultaneous close 3396 * happens, we must ack the received FIN and 3397 * enter the CLOSING state. 3398 */ 3399 tcp_send_ack(sk); 3400 tcp_set_state(sk, TCP_CLOSING); 3401 break; 3402 case TCP_FIN_WAIT2: 3403 /* Received a FIN -- send ACK and enter TIME_WAIT. */ 3404 tcp_send_ack(sk); 3405 tcp_time_wait(sk, TCP_TIME_WAIT, 0); 3406 break; 3407 default: 3408 /* Only TCP_LISTEN and TCP_CLOSE are left, in these 3409 * cases we should never reach this piece of code. 3410 */ 3411 printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n", 3412 __FUNCTION__, sk->sk_state); 3413 break; 3414 } 3415 3416 /* It _is_ possible, that we have something out-of-order _after_ FIN. 3417 * Probably, we should reset in this case. For now drop them. 3418 */ 3419 __skb_queue_purge(&tp->out_of_order_queue); 3420 if (tcp_is_sack(tp)) 3421 tcp_sack_reset(&tp->rx_opt); 3422 sk_stream_mem_reclaim(sk); 3423 3424 if (!sock_flag(sk, SOCK_DEAD)) { 3425 sk->sk_state_change(sk); 3426 3427 /* Do not send POLL_HUP for half duplex close. */ 3428 if (sk->sk_shutdown == SHUTDOWN_MASK || 3429 sk->sk_state == TCP_CLOSE) 3430 sk_wake_async(sk, 1, POLL_HUP); 3431 else 3432 sk_wake_async(sk, 1, POLL_IN); 3433 } 3434 } 3435 3436 static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, u32 end_seq) 3437 { 3438 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) { 3439 if (before(seq, sp->start_seq)) 3440 sp->start_seq = seq; 3441 if (after(end_seq, sp->end_seq)) 3442 sp->end_seq = end_seq; 3443 return 1; 3444 } 3445 return 0; 3446 } 3447 3448 static void tcp_dsack_set(struct tcp_sock *tp, u32 seq, u32 end_seq) 3449 { 3450 if (tcp_is_sack(tp) && sysctl_tcp_dsack) { 3451 if (before(seq, tp->rcv_nxt)) 3452 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT); 3453 else 3454 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT); 3455 3456 tp->rx_opt.dsack = 1; 3457 tp->duplicate_sack[0].start_seq = seq; 3458 tp->duplicate_sack[0].end_seq = end_seq; 3459 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + 1, 4 - tp->rx_opt.tstamp_ok); 3460 } 3461 } 3462 3463 static void tcp_dsack_extend(struct tcp_sock *tp, u32 seq, u32 end_seq) 3464 { 3465 if (!tp->rx_opt.dsack) 3466 tcp_dsack_set(tp, seq, end_seq); 3467 else 3468 tcp_sack_extend(tp->duplicate_sack, seq, end_seq); 3469 } 3470 3471 static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb) 3472 { 3473 struct tcp_sock *tp = tcp_sk(sk); 3474 3475 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 3476 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 3477 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST); 3478 tcp_enter_quickack_mode(sk); 3479 3480 if (tcp_is_sack(tp) && sysctl_tcp_dsack) { 3481 u32 end_seq = TCP_SKB_CB(skb)->end_seq; 3482 3483 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) 3484 end_seq = tp->rcv_nxt; 3485 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, end_seq); 3486 } 3487 } 3488 3489 tcp_send_ack(sk); 3490 } 3491 3492 /* These routines update the SACK block as out-of-order packets arrive or 3493 * in-order packets close up the sequence space. 3494 */ 3495 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp) 3496 { 3497 int this_sack; 3498 struct tcp_sack_block *sp = &tp->selective_acks[0]; 3499 struct tcp_sack_block *swalk = sp+1; 3500 3501 /* See if the recent change to the first SACK eats into 3502 * or hits the sequence space of other SACK blocks, if so coalesce. 3503 */ 3504 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks; ) { 3505 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) { 3506 int i; 3507 3508 /* Zap SWALK, by moving every further SACK up by one slot. 3509 * Decrease num_sacks. 3510 */ 3511 tp->rx_opt.num_sacks--; 3512 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok); 3513 for (i=this_sack; i < tp->rx_opt.num_sacks; i++) 3514 sp[i] = sp[i+1]; 3515 continue; 3516 } 3517 this_sack++, swalk++; 3518 } 3519 } 3520 3521 static inline void tcp_sack_swap(struct tcp_sack_block *sack1, struct tcp_sack_block *sack2) 3522 { 3523 __u32 tmp; 3524 3525 tmp = sack1->start_seq; 3526 sack1->start_seq = sack2->start_seq; 3527 sack2->start_seq = tmp; 3528 3529 tmp = sack1->end_seq; 3530 sack1->end_seq = sack2->end_seq; 3531 sack2->end_seq = tmp; 3532 } 3533 3534 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq) 3535 { 3536 struct tcp_sock *tp = tcp_sk(sk); 3537 struct tcp_sack_block *sp = &tp->selective_acks[0]; 3538 int cur_sacks = tp->rx_opt.num_sacks; 3539 int this_sack; 3540 3541 if (!cur_sacks) 3542 goto new_sack; 3543 3544 for (this_sack=0; this_sack<cur_sacks; this_sack++, sp++) { 3545 if (tcp_sack_extend(sp, seq, end_seq)) { 3546 /* Rotate this_sack to the first one. */ 3547 for (; this_sack>0; this_sack--, sp--) 3548 tcp_sack_swap(sp, sp-1); 3549 if (cur_sacks > 1) 3550 tcp_sack_maybe_coalesce(tp); 3551 return; 3552 } 3553 } 3554 3555 /* Could not find an adjacent existing SACK, build a new one, 3556 * put it at the front, and shift everyone else down. We 3557 * always know there is at least one SACK present already here. 3558 * 3559 * If the sack array is full, forget about the last one. 3560 */ 3561 if (this_sack >= 4) { 3562 this_sack--; 3563 tp->rx_opt.num_sacks--; 3564 sp--; 3565 } 3566 for (; this_sack > 0; this_sack--, sp--) 3567 *sp = *(sp-1); 3568 3569 new_sack: 3570 /* Build the new head SACK, and we're done. */ 3571 sp->start_seq = seq; 3572 sp->end_seq = end_seq; 3573 tp->rx_opt.num_sacks++; 3574 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok); 3575 } 3576 3577 /* RCV.NXT advances, some SACKs should be eaten. */ 3578 3579 static void tcp_sack_remove(struct tcp_sock *tp) 3580 { 3581 struct tcp_sack_block *sp = &tp->selective_acks[0]; 3582 int num_sacks = tp->rx_opt.num_sacks; 3583 int this_sack; 3584 3585 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */ 3586 if (skb_queue_empty(&tp->out_of_order_queue)) { 3587 tp->rx_opt.num_sacks = 0; 3588 tp->rx_opt.eff_sacks = tp->rx_opt.dsack; 3589 return; 3590 } 3591 3592 for (this_sack = 0; this_sack < num_sacks; ) { 3593 /* Check if the start of the sack is covered by RCV.NXT. */ 3594 if (!before(tp->rcv_nxt, sp->start_seq)) { 3595 int i; 3596 3597 /* RCV.NXT must cover all the block! */ 3598 BUG_TRAP(!before(tp->rcv_nxt, sp->end_seq)); 3599 3600 /* Zap this SACK, by moving forward any other SACKS. */ 3601 for (i=this_sack+1; i < num_sacks; i++) 3602 tp->selective_acks[i-1] = tp->selective_acks[i]; 3603 num_sacks--; 3604 continue; 3605 } 3606 this_sack++; 3607 sp++; 3608 } 3609 if (num_sacks != tp->rx_opt.num_sacks) { 3610 tp->rx_opt.num_sacks = num_sacks; 3611 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok); 3612 } 3613 } 3614 3615 /* This one checks to see if we can put data from the 3616 * out_of_order queue into the receive_queue. 3617 */ 3618 static void tcp_ofo_queue(struct sock *sk) 3619 { 3620 struct tcp_sock *tp = tcp_sk(sk); 3621 __u32 dsack_high = tp->rcv_nxt; 3622 struct sk_buff *skb; 3623 3624 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) { 3625 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) 3626 break; 3627 3628 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) { 3629 __u32 dsack = dsack_high; 3630 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high)) 3631 dsack_high = TCP_SKB_CB(skb)->end_seq; 3632 tcp_dsack_extend(tp, TCP_SKB_CB(skb)->seq, dsack); 3633 } 3634 3635 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) { 3636 SOCK_DEBUG(sk, "ofo packet was already received \n"); 3637 __skb_unlink(skb, &tp->out_of_order_queue); 3638 __kfree_skb(skb); 3639 continue; 3640 } 3641 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n", 3642 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, 3643 TCP_SKB_CB(skb)->end_seq); 3644 3645 __skb_unlink(skb, &tp->out_of_order_queue); 3646 __skb_queue_tail(&sk->sk_receive_queue, skb); 3647 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; 3648 if (tcp_hdr(skb)->fin) 3649 tcp_fin(skb, sk, tcp_hdr(skb)); 3650 } 3651 } 3652 3653 static int tcp_prune_queue(struct sock *sk); 3654 3655 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb) 3656 { 3657 struct tcphdr *th = tcp_hdr(skb); 3658 struct tcp_sock *tp = tcp_sk(sk); 3659 int eaten = -1; 3660 3661 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) 3662 goto drop; 3663 3664 __skb_pull(skb, th->doff*4); 3665 3666 TCP_ECN_accept_cwr(tp, skb); 3667 3668 if (tp->rx_opt.dsack) { 3669 tp->rx_opt.dsack = 0; 3670 tp->rx_opt.eff_sacks = min_t(unsigned int, tp->rx_opt.num_sacks, 3671 4 - tp->rx_opt.tstamp_ok); 3672 } 3673 3674 /* Queue data for delivery to the user. 3675 * Packets in sequence go to the receive queue. 3676 * Out of sequence packets to the out_of_order_queue. 3677 */ 3678 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) { 3679 if (tcp_receive_window(tp) == 0) 3680 goto out_of_window; 3681 3682 /* Ok. In sequence. In window. */ 3683 if (tp->ucopy.task == current && 3684 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len && 3685 sock_owned_by_user(sk) && !tp->urg_data) { 3686 int chunk = min_t(unsigned int, skb->len, 3687 tp->ucopy.len); 3688 3689 __set_current_state(TASK_RUNNING); 3690 3691 local_bh_enable(); 3692 if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) { 3693 tp->ucopy.len -= chunk; 3694 tp->copied_seq += chunk; 3695 eaten = (chunk == skb->len && !th->fin); 3696 tcp_rcv_space_adjust(sk); 3697 } 3698 local_bh_disable(); 3699 } 3700 3701 if (eaten <= 0) { 3702 queue_and_out: 3703 if (eaten < 0 && 3704 (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf || 3705 !sk_stream_rmem_schedule(sk, skb))) { 3706 if (tcp_prune_queue(sk) < 0 || 3707 !sk_stream_rmem_schedule(sk, skb)) 3708 goto drop; 3709 } 3710 sk_stream_set_owner_r(skb, sk); 3711 __skb_queue_tail(&sk->sk_receive_queue, skb); 3712 } 3713 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; 3714 if (skb->len) 3715 tcp_event_data_recv(sk, skb); 3716 if (th->fin) 3717 tcp_fin(skb, sk, th); 3718 3719 if (!skb_queue_empty(&tp->out_of_order_queue)) { 3720 tcp_ofo_queue(sk); 3721 3722 /* RFC2581. 4.2. SHOULD send immediate ACK, when 3723 * gap in queue is filled. 3724 */ 3725 if (skb_queue_empty(&tp->out_of_order_queue)) 3726 inet_csk(sk)->icsk_ack.pingpong = 0; 3727 } 3728 3729 if (tp->rx_opt.num_sacks) 3730 tcp_sack_remove(tp); 3731 3732 tcp_fast_path_check(sk); 3733 3734 if (eaten > 0) 3735 __kfree_skb(skb); 3736 else if (!sock_flag(sk, SOCK_DEAD)) 3737 sk->sk_data_ready(sk, 0); 3738 return; 3739 } 3740 3741 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) { 3742 /* A retransmit, 2nd most common case. Force an immediate ack. */ 3743 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST); 3744 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); 3745 3746 out_of_window: 3747 tcp_enter_quickack_mode(sk); 3748 inet_csk_schedule_ack(sk); 3749 drop: 3750 __kfree_skb(skb); 3751 return; 3752 } 3753 3754 /* Out of window. F.e. zero window probe. */ 3755 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp))) 3756 goto out_of_window; 3757 3758 tcp_enter_quickack_mode(sk); 3759 3760 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 3761 /* Partial packet, seq < rcv_next < end_seq */ 3762 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n", 3763 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, 3764 TCP_SKB_CB(skb)->end_seq); 3765 3766 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, tp->rcv_nxt); 3767 3768 /* If window is closed, drop tail of packet. But after 3769 * remembering D-SACK for its head made in previous line. 3770 */ 3771 if (!tcp_receive_window(tp)) 3772 goto out_of_window; 3773 goto queue_and_out; 3774 } 3775 3776 TCP_ECN_check_ce(tp, skb); 3777 3778 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf || 3779 !sk_stream_rmem_schedule(sk, skb)) { 3780 if (tcp_prune_queue(sk) < 0 || 3781 !sk_stream_rmem_schedule(sk, skb)) 3782 goto drop; 3783 } 3784 3785 /* Disable header prediction. */ 3786 tp->pred_flags = 0; 3787 inet_csk_schedule_ack(sk); 3788 3789 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n", 3790 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); 3791 3792 sk_stream_set_owner_r(skb, sk); 3793 3794 if (!skb_peek(&tp->out_of_order_queue)) { 3795 /* Initial out of order segment, build 1 SACK. */ 3796 if (tcp_is_sack(tp)) { 3797 tp->rx_opt.num_sacks = 1; 3798 tp->rx_opt.dsack = 0; 3799 tp->rx_opt.eff_sacks = 1; 3800 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq; 3801 tp->selective_acks[0].end_seq = 3802 TCP_SKB_CB(skb)->end_seq; 3803 } 3804 __skb_queue_head(&tp->out_of_order_queue,skb); 3805 } else { 3806 struct sk_buff *skb1 = tp->out_of_order_queue.prev; 3807 u32 seq = TCP_SKB_CB(skb)->seq; 3808 u32 end_seq = TCP_SKB_CB(skb)->end_seq; 3809 3810 if (seq == TCP_SKB_CB(skb1)->end_seq) { 3811 __skb_append(skb1, skb, &tp->out_of_order_queue); 3812 3813 if (!tp->rx_opt.num_sacks || 3814 tp->selective_acks[0].end_seq != seq) 3815 goto add_sack; 3816 3817 /* Common case: data arrive in order after hole. */ 3818 tp->selective_acks[0].end_seq = end_seq; 3819 return; 3820 } 3821 3822 /* Find place to insert this segment. */ 3823 do { 3824 if (!after(TCP_SKB_CB(skb1)->seq, seq)) 3825 break; 3826 } while ((skb1 = skb1->prev) != 3827 (struct sk_buff*)&tp->out_of_order_queue); 3828 3829 /* Do skb overlap to previous one? */ 3830 if (skb1 != (struct sk_buff*)&tp->out_of_order_queue && 3831 before(seq, TCP_SKB_CB(skb1)->end_seq)) { 3832 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) { 3833 /* All the bits are present. Drop. */ 3834 __kfree_skb(skb); 3835 tcp_dsack_set(tp, seq, end_seq); 3836 goto add_sack; 3837 } 3838 if (after(seq, TCP_SKB_CB(skb1)->seq)) { 3839 /* Partial overlap. */ 3840 tcp_dsack_set(tp, seq, TCP_SKB_CB(skb1)->end_seq); 3841 } else { 3842 skb1 = skb1->prev; 3843 } 3844 } 3845 __skb_insert(skb, skb1, skb1->next, &tp->out_of_order_queue); 3846 3847 /* And clean segments covered by new one as whole. */ 3848 while ((skb1 = skb->next) != 3849 (struct sk_buff*)&tp->out_of_order_queue && 3850 after(end_seq, TCP_SKB_CB(skb1)->seq)) { 3851 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) { 3852 tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, end_seq); 3853 break; 3854 } 3855 __skb_unlink(skb1, &tp->out_of_order_queue); 3856 tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, TCP_SKB_CB(skb1)->end_seq); 3857 __kfree_skb(skb1); 3858 } 3859 3860 add_sack: 3861 if (tcp_is_sack(tp)) 3862 tcp_sack_new_ofo_skb(sk, seq, end_seq); 3863 } 3864 } 3865 3866 /* Collapse contiguous sequence of skbs head..tail with 3867 * sequence numbers start..end. 3868 * Segments with FIN/SYN are not collapsed (only because this 3869 * simplifies code) 3870 */ 3871 static void 3872 tcp_collapse(struct sock *sk, struct sk_buff_head *list, 3873 struct sk_buff *head, struct sk_buff *tail, 3874 u32 start, u32 end) 3875 { 3876 struct sk_buff *skb; 3877 3878 /* First, check that queue is collapsible and find 3879 * the point where collapsing can be useful. */ 3880 for (skb = head; skb != tail; ) { 3881 /* No new bits? It is possible on ofo queue. */ 3882 if (!before(start, TCP_SKB_CB(skb)->end_seq)) { 3883 struct sk_buff *next = skb->next; 3884 __skb_unlink(skb, list); 3885 __kfree_skb(skb); 3886 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED); 3887 skb = next; 3888 continue; 3889 } 3890 3891 /* The first skb to collapse is: 3892 * - not SYN/FIN and 3893 * - bloated or contains data before "start" or 3894 * overlaps to the next one. 3895 */ 3896 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin && 3897 (tcp_win_from_space(skb->truesize) > skb->len || 3898 before(TCP_SKB_CB(skb)->seq, start) || 3899 (skb->next != tail && 3900 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq))) 3901 break; 3902 3903 /* Decided to skip this, advance start seq. */ 3904 start = TCP_SKB_CB(skb)->end_seq; 3905 skb = skb->next; 3906 } 3907 if (skb == tail || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin) 3908 return; 3909 3910 while (before(start, end)) { 3911 struct sk_buff *nskb; 3912 int header = skb_headroom(skb); 3913 int copy = SKB_MAX_ORDER(header, 0); 3914 3915 /* Too big header? This can happen with IPv6. */ 3916 if (copy < 0) 3917 return; 3918 if (end-start < copy) 3919 copy = end-start; 3920 nskb = alloc_skb(copy+header, GFP_ATOMIC); 3921 if (!nskb) 3922 return; 3923 3924 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head); 3925 skb_set_network_header(nskb, (skb_network_header(skb) - 3926 skb->head)); 3927 skb_set_transport_header(nskb, (skb_transport_header(skb) - 3928 skb->head)); 3929 skb_reserve(nskb, header); 3930 memcpy(nskb->head, skb->head, header); 3931 memcpy(nskb->cb, skb->cb, sizeof(skb->cb)); 3932 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start; 3933 __skb_insert(nskb, skb->prev, skb, list); 3934 sk_stream_set_owner_r(nskb, sk); 3935 3936 /* Copy data, releasing collapsed skbs. */ 3937 while (copy > 0) { 3938 int offset = start - TCP_SKB_CB(skb)->seq; 3939 int size = TCP_SKB_CB(skb)->end_seq - start; 3940 3941 BUG_ON(offset < 0); 3942 if (size > 0) { 3943 size = min(copy, size); 3944 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size)) 3945 BUG(); 3946 TCP_SKB_CB(nskb)->end_seq += size; 3947 copy -= size; 3948 start += size; 3949 } 3950 if (!before(start, TCP_SKB_CB(skb)->end_seq)) { 3951 struct sk_buff *next = skb->next; 3952 __skb_unlink(skb, list); 3953 __kfree_skb(skb); 3954 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED); 3955 skb = next; 3956 if (skb == tail || 3957 tcp_hdr(skb)->syn || 3958 tcp_hdr(skb)->fin) 3959 return; 3960 } 3961 } 3962 } 3963 } 3964 3965 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs 3966 * and tcp_collapse() them until all the queue is collapsed. 3967 */ 3968 static void tcp_collapse_ofo_queue(struct sock *sk) 3969 { 3970 struct tcp_sock *tp = tcp_sk(sk); 3971 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue); 3972 struct sk_buff *head; 3973 u32 start, end; 3974 3975 if (skb == NULL) 3976 return; 3977 3978 start = TCP_SKB_CB(skb)->seq; 3979 end = TCP_SKB_CB(skb)->end_seq; 3980 head = skb; 3981 3982 for (;;) { 3983 skb = skb->next; 3984 3985 /* Segment is terminated when we see gap or when 3986 * we are at the end of all the queue. */ 3987 if (skb == (struct sk_buff *)&tp->out_of_order_queue || 3988 after(TCP_SKB_CB(skb)->seq, end) || 3989 before(TCP_SKB_CB(skb)->end_seq, start)) { 3990 tcp_collapse(sk, &tp->out_of_order_queue, 3991 head, skb, start, end); 3992 head = skb; 3993 if (skb == (struct sk_buff *)&tp->out_of_order_queue) 3994 break; 3995 /* Start new segment */ 3996 start = TCP_SKB_CB(skb)->seq; 3997 end = TCP_SKB_CB(skb)->end_seq; 3998 } else { 3999 if (before(TCP_SKB_CB(skb)->seq, start)) 4000 start = TCP_SKB_CB(skb)->seq; 4001 if (after(TCP_SKB_CB(skb)->end_seq, end)) 4002 end = TCP_SKB_CB(skb)->end_seq; 4003 } 4004 } 4005 } 4006 4007 /* Reduce allocated memory if we can, trying to get 4008 * the socket within its memory limits again. 4009 * 4010 * Return less than zero if we should start dropping frames 4011 * until the socket owning process reads some of the data 4012 * to stabilize the situation. 4013 */ 4014 static int tcp_prune_queue(struct sock *sk) 4015 { 4016 struct tcp_sock *tp = tcp_sk(sk); 4017 4018 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq); 4019 4020 NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED); 4021 4022 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf) 4023 tcp_clamp_window(sk); 4024 else if (tcp_memory_pressure) 4025 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss); 4026 4027 tcp_collapse_ofo_queue(sk); 4028 tcp_collapse(sk, &sk->sk_receive_queue, 4029 sk->sk_receive_queue.next, 4030 (struct sk_buff*)&sk->sk_receive_queue, 4031 tp->copied_seq, tp->rcv_nxt); 4032 sk_stream_mem_reclaim(sk); 4033 4034 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) 4035 return 0; 4036 4037 /* Collapsing did not help, destructive actions follow. 4038 * This must not ever occur. */ 4039 4040 /* First, purge the out_of_order queue. */ 4041 if (!skb_queue_empty(&tp->out_of_order_queue)) { 4042 NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED); 4043 __skb_queue_purge(&tp->out_of_order_queue); 4044 4045 /* Reset SACK state. A conforming SACK implementation will 4046 * do the same at a timeout based retransmit. When a connection 4047 * is in a sad state like this, we care only about integrity 4048 * of the connection not performance. 4049 */ 4050 if (tcp_is_sack(tp)) 4051 tcp_sack_reset(&tp->rx_opt); 4052 sk_stream_mem_reclaim(sk); 4053 } 4054 4055 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) 4056 return 0; 4057 4058 /* If we are really being abused, tell the caller to silently 4059 * drop receive data on the floor. It will get retransmitted 4060 * and hopefully then we'll have sufficient space. 4061 */ 4062 NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED); 4063 4064 /* Massive buffer overcommit. */ 4065 tp->pred_flags = 0; 4066 return -1; 4067 } 4068 4069 4070 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. 4071 * As additional protections, we do not touch cwnd in retransmission phases, 4072 * and if application hit its sndbuf limit recently. 4073 */ 4074 void tcp_cwnd_application_limited(struct sock *sk) 4075 { 4076 struct tcp_sock *tp = tcp_sk(sk); 4077 4078 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && 4079 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { 4080 /* Limited by application or receiver window. */ 4081 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); 4082 u32 win_used = max(tp->snd_cwnd_used, init_win); 4083 if (win_used < tp->snd_cwnd) { 4084 tp->snd_ssthresh = tcp_current_ssthresh(sk); 4085 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1; 4086 } 4087 tp->snd_cwnd_used = 0; 4088 } 4089 tp->snd_cwnd_stamp = tcp_time_stamp; 4090 } 4091 4092 static int tcp_should_expand_sndbuf(struct sock *sk) 4093 { 4094 struct tcp_sock *tp = tcp_sk(sk); 4095 4096 /* If the user specified a specific send buffer setting, do 4097 * not modify it. 4098 */ 4099 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK) 4100 return 0; 4101 4102 /* If we are under global TCP memory pressure, do not expand. */ 4103 if (tcp_memory_pressure) 4104 return 0; 4105 4106 /* If we are under soft global TCP memory pressure, do not expand. */ 4107 if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0]) 4108 return 0; 4109 4110 /* If we filled the congestion window, do not expand. */ 4111 if (tp->packets_out >= tp->snd_cwnd) 4112 return 0; 4113 4114 return 1; 4115 } 4116 4117 /* When incoming ACK allowed to free some skb from write_queue, 4118 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket 4119 * on the exit from tcp input handler. 4120 * 4121 * PROBLEM: sndbuf expansion does not work well with largesend. 4122 */ 4123 static void tcp_new_space(struct sock *sk) 4124 { 4125 struct tcp_sock *tp = tcp_sk(sk); 4126 4127 if (tcp_should_expand_sndbuf(sk)) { 4128 int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) + 4129 MAX_TCP_HEADER + 16 + sizeof(struct sk_buff), 4130 demanded = max_t(unsigned int, tp->snd_cwnd, 4131 tp->reordering + 1); 4132 sndmem *= 2*demanded; 4133 if (sndmem > sk->sk_sndbuf) 4134 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]); 4135 tp->snd_cwnd_stamp = tcp_time_stamp; 4136 } 4137 4138 sk->sk_write_space(sk); 4139 } 4140 4141 static void tcp_check_space(struct sock *sk) 4142 { 4143 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) { 4144 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK); 4145 if (sk->sk_socket && 4146 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) 4147 tcp_new_space(sk); 4148 } 4149 } 4150 4151 static inline void tcp_data_snd_check(struct sock *sk) 4152 { 4153 tcp_push_pending_frames(sk); 4154 tcp_check_space(sk); 4155 } 4156 4157 /* 4158 * Check if sending an ack is needed. 4159 */ 4160 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible) 4161 { 4162 struct tcp_sock *tp = tcp_sk(sk); 4163 4164 /* More than one full frame received... */ 4165 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss 4166 /* ... and right edge of window advances far enough. 4167 * (tcp_recvmsg() will send ACK otherwise). Or... 4168 */ 4169 && __tcp_select_window(sk) >= tp->rcv_wnd) || 4170 /* We ACK each frame or... */ 4171 tcp_in_quickack_mode(sk) || 4172 /* We have out of order data. */ 4173 (ofo_possible && 4174 skb_peek(&tp->out_of_order_queue))) { 4175 /* Then ack it now */ 4176 tcp_send_ack(sk); 4177 } else { 4178 /* Else, send delayed ack. */ 4179 tcp_send_delayed_ack(sk); 4180 } 4181 } 4182 4183 static inline void tcp_ack_snd_check(struct sock *sk) 4184 { 4185 if (!inet_csk_ack_scheduled(sk)) { 4186 /* We sent a data segment already. */ 4187 return; 4188 } 4189 __tcp_ack_snd_check(sk, 1); 4190 } 4191 4192 /* 4193 * This routine is only called when we have urgent data 4194 * signaled. Its the 'slow' part of tcp_urg. It could be 4195 * moved inline now as tcp_urg is only called from one 4196 * place. We handle URGent data wrong. We have to - as 4197 * BSD still doesn't use the correction from RFC961. 4198 * For 1003.1g we should support a new option TCP_STDURG to permit 4199 * either form (or just set the sysctl tcp_stdurg). 4200 */ 4201 4202 static void tcp_check_urg(struct sock * sk, struct tcphdr * th) 4203 { 4204 struct tcp_sock *tp = tcp_sk(sk); 4205 u32 ptr = ntohs(th->urg_ptr); 4206 4207 if (ptr && !sysctl_tcp_stdurg) 4208 ptr--; 4209 ptr += ntohl(th->seq); 4210 4211 /* Ignore urgent data that we've already seen and read. */ 4212 if (after(tp->copied_seq, ptr)) 4213 return; 4214 4215 /* Do not replay urg ptr. 4216 * 4217 * NOTE: interesting situation not covered by specs. 4218 * Misbehaving sender may send urg ptr, pointing to segment, 4219 * which we already have in ofo queue. We are not able to fetch 4220 * such data and will stay in TCP_URG_NOTYET until will be eaten 4221 * by recvmsg(). Seems, we are not obliged to handle such wicked 4222 * situations. But it is worth to think about possibility of some 4223 * DoSes using some hypothetical application level deadlock. 4224 */ 4225 if (before(ptr, tp->rcv_nxt)) 4226 return; 4227 4228 /* Do we already have a newer (or duplicate) urgent pointer? */ 4229 if (tp->urg_data && !after(ptr, tp->urg_seq)) 4230 return; 4231 4232 /* Tell the world about our new urgent pointer. */ 4233 sk_send_sigurg(sk); 4234 4235 /* We may be adding urgent data when the last byte read was 4236 * urgent. To do this requires some care. We cannot just ignore 4237 * tp->copied_seq since we would read the last urgent byte again 4238 * as data, nor can we alter copied_seq until this data arrives 4239 * or we break the semantics of SIOCATMARK (and thus sockatmark()) 4240 * 4241 * NOTE. Double Dutch. Rendering to plain English: author of comment 4242 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB); 4243 * and expect that both A and B disappear from stream. This is _wrong_. 4244 * Though this happens in BSD with high probability, this is occasional. 4245 * Any application relying on this is buggy. Note also, that fix "works" 4246 * only in this artificial test. Insert some normal data between A and B and we will 4247 * decline of BSD again. Verdict: it is better to remove to trap 4248 * buggy users. 4249 */ 4250 if (tp->urg_seq == tp->copied_seq && tp->urg_data && 4251 !sock_flag(sk, SOCK_URGINLINE) && 4252 tp->copied_seq != tp->rcv_nxt) { 4253 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue); 4254 tp->copied_seq++; 4255 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) { 4256 __skb_unlink(skb, &sk->sk_receive_queue); 4257 __kfree_skb(skb); 4258 } 4259 } 4260 4261 tp->urg_data = TCP_URG_NOTYET; 4262 tp->urg_seq = ptr; 4263 4264 /* Disable header prediction. */ 4265 tp->pred_flags = 0; 4266 } 4267 4268 /* This is the 'fast' part of urgent handling. */ 4269 static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th) 4270 { 4271 struct tcp_sock *tp = tcp_sk(sk); 4272 4273 /* Check if we get a new urgent pointer - normally not. */ 4274 if (th->urg) 4275 tcp_check_urg(sk,th); 4276 4277 /* Do we wait for any urgent data? - normally not... */ 4278 if (tp->urg_data == TCP_URG_NOTYET) { 4279 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) - 4280 th->syn; 4281 4282 /* Is the urgent pointer pointing into this packet? */ 4283 if (ptr < skb->len) { 4284 u8 tmp; 4285 if (skb_copy_bits(skb, ptr, &tmp, 1)) 4286 BUG(); 4287 tp->urg_data = TCP_URG_VALID | tmp; 4288 if (!sock_flag(sk, SOCK_DEAD)) 4289 sk->sk_data_ready(sk, 0); 4290 } 4291 } 4292 } 4293 4294 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen) 4295 { 4296 struct tcp_sock *tp = tcp_sk(sk); 4297 int chunk = skb->len - hlen; 4298 int err; 4299 4300 local_bh_enable(); 4301 if (skb_csum_unnecessary(skb)) 4302 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk); 4303 else 4304 err = skb_copy_and_csum_datagram_iovec(skb, hlen, 4305 tp->ucopy.iov); 4306 4307 if (!err) { 4308 tp->ucopy.len -= chunk; 4309 tp->copied_seq += chunk; 4310 tcp_rcv_space_adjust(sk); 4311 } 4312 4313 local_bh_disable(); 4314 return err; 4315 } 4316 4317 static __sum16 __tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb) 4318 { 4319 __sum16 result; 4320 4321 if (sock_owned_by_user(sk)) { 4322 local_bh_enable(); 4323 result = __tcp_checksum_complete(skb); 4324 local_bh_disable(); 4325 } else { 4326 result = __tcp_checksum_complete(skb); 4327 } 4328 return result; 4329 } 4330 4331 static inline int tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb) 4332 { 4333 return !skb_csum_unnecessary(skb) && 4334 __tcp_checksum_complete_user(sk, skb); 4335 } 4336 4337 #ifdef CONFIG_NET_DMA 4338 static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb, int hlen) 4339 { 4340 struct tcp_sock *tp = tcp_sk(sk); 4341 int chunk = skb->len - hlen; 4342 int dma_cookie; 4343 int copied_early = 0; 4344 4345 if (tp->ucopy.wakeup) 4346 return 0; 4347 4348 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list) 4349 tp->ucopy.dma_chan = get_softnet_dma(); 4350 4351 if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) { 4352 4353 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan, 4354 skb, hlen, tp->ucopy.iov, chunk, tp->ucopy.pinned_list); 4355 4356 if (dma_cookie < 0) 4357 goto out; 4358 4359 tp->ucopy.dma_cookie = dma_cookie; 4360 copied_early = 1; 4361 4362 tp->ucopy.len -= chunk; 4363 tp->copied_seq += chunk; 4364 tcp_rcv_space_adjust(sk); 4365 4366 if ((tp->ucopy.len == 0) || 4367 (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) || 4368 (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) { 4369 tp->ucopy.wakeup = 1; 4370 sk->sk_data_ready(sk, 0); 4371 } 4372 } else if (chunk > 0) { 4373 tp->ucopy.wakeup = 1; 4374 sk->sk_data_ready(sk, 0); 4375 } 4376 out: 4377 return copied_early; 4378 } 4379 #endif /* CONFIG_NET_DMA */ 4380 4381 /* 4382 * TCP receive function for the ESTABLISHED state. 4383 * 4384 * It is split into a fast path and a slow path. The fast path is 4385 * disabled when: 4386 * - A zero window was announced from us - zero window probing 4387 * is only handled properly in the slow path. 4388 * - Out of order segments arrived. 4389 * - Urgent data is expected. 4390 * - There is no buffer space left 4391 * - Unexpected TCP flags/window values/header lengths are received 4392 * (detected by checking the TCP header against pred_flags) 4393 * - Data is sent in both directions. Fast path only supports pure senders 4394 * or pure receivers (this means either the sequence number or the ack 4395 * value must stay constant) 4396 * - Unexpected TCP option. 4397 * 4398 * When these conditions are not satisfied it drops into a standard 4399 * receive procedure patterned after RFC793 to handle all cases. 4400 * The first three cases are guaranteed by proper pred_flags setting, 4401 * the rest is checked inline. Fast processing is turned on in 4402 * tcp_data_queue when everything is OK. 4403 */ 4404 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb, 4405 struct tcphdr *th, unsigned len) 4406 { 4407 struct tcp_sock *tp = tcp_sk(sk); 4408 4409 /* 4410 * Header prediction. 4411 * The code loosely follows the one in the famous 4412 * "30 instruction TCP receive" Van Jacobson mail. 4413 * 4414 * Van's trick is to deposit buffers into socket queue 4415 * on a device interrupt, to call tcp_recv function 4416 * on the receive process context and checksum and copy 4417 * the buffer to user space. smart... 4418 * 4419 * Our current scheme is not silly either but we take the 4420 * extra cost of the net_bh soft interrupt processing... 4421 * We do checksum and copy also but from device to kernel. 4422 */ 4423 4424 tp->rx_opt.saw_tstamp = 0; 4425 4426 /* pred_flags is 0xS?10 << 16 + snd_wnd 4427 * if header_prediction is to be made 4428 * 'S' will always be tp->tcp_header_len >> 2 4429 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to 4430 * turn it off (when there are holes in the receive 4431 * space for instance) 4432 * PSH flag is ignored. 4433 */ 4434 4435 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags && 4436 TCP_SKB_CB(skb)->seq == tp->rcv_nxt) { 4437 int tcp_header_len = tp->tcp_header_len; 4438 4439 /* Timestamp header prediction: tcp_header_len 4440 * is automatically equal to th->doff*4 due to pred_flags 4441 * match. 4442 */ 4443 4444 /* Check timestamp */ 4445 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) { 4446 __be32 *ptr = (__be32 *)(th + 1); 4447 4448 /* No? Slow path! */ 4449 if (*ptr != htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) 4450 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) 4451 goto slow_path; 4452 4453 tp->rx_opt.saw_tstamp = 1; 4454 ++ptr; 4455 tp->rx_opt.rcv_tsval = ntohl(*ptr); 4456 ++ptr; 4457 tp->rx_opt.rcv_tsecr = ntohl(*ptr); 4458 4459 /* If PAWS failed, check it more carefully in slow path */ 4460 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0) 4461 goto slow_path; 4462 4463 /* DO NOT update ts_recent here, if checksum fails 4464 * and timestamp was corrupted part, it will result 4465 * in a hung connection since we will drop all 4466 * future packets due to the PAWS test. 4467 */ 4468 } 4469 4470 if (len <= tcp_header_len) { 4471 /* Bulk data transfer: sender */ 4472 if (len == tcp_header_len) { 4473 /* Predicted packet is in window by definition. 4474 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 4475 * Hence, check seq<=rcv_wup reduces to: 4476 */ 4477 if (tcp_header_len == 4478 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && 4479 tp->rcv_nxt == tp->rcv_wup) 4480 tcp_store_ts_recent(tp); 4481 4482 /* We know that such packets are checksummed 4483 * on entry. 4484 */ 4485 tcp_ack(sk, skb, 0); 4486 __kfree_skb(skb); 4487 tcp_data_snd_check(sk); 4488 return 0; 4489 } else { /* Header too small */ 4490 TCP_INC_STATS_BH(TCP_MIB_INERRS); 4491 goto discard; 4492 } 4493 } else { 4494 int eaten = 0; 4495 int copied_early = 0; 4496 4497 if (tp->copied_seq == tp->rcv_nxt && 4498 len - tcp_header_len <= tp->ucopy.len) { 4499 #ifdef CONFIG_NET_DMA 4500 if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) { 4501 copied_early = 1; 4502 eaten = 1; 4503 } 4504 #endif 4505 if (tp->ucopy.task == current && sock_owned_by_user(sk) && !copied_early) { 4506 __set_current_state(TASK_RUNNING); 4507 4508 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len)) 4509 eaten = 1; 4510 } 4511 if (eaten) { 4512 /* Predicted packet is in window by definition. 4513 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 4514 * Hence, check seq<=rcv_wup reduces to: 4515 */ 4516 if (tcp_header_len == 4517 (sizeof(struct tcphdr) + 4518 TCPOLEN_TSTAMP_ALIGNED) && 4519 tp->rcv_nxt == tp->rcv_wup) 4520 tcp_store_ts_recent(tp); 4521 4522 tcp_rcv_rtt_measure_ts(sk, skb); 4523 4524 __skb_pull(skb, tcp_header_len); 4525 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; 4526 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER); 4527 } 4528 if (copied_early) 4529 tcp_cleanup_rbuf(sk, skb->len); 4530 } 4531 if (!eaten) { 4532 if (tcp_checksum_complete_user(sk, skb)) 4533 goto csum_error; 4534 4535 /* Predicted packet is in window by definition. 4536 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 4537 * Hence, check seq<=rcv_wup reduces to: 4538 */ 4539 if (tcp_header_len == 4540 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && 4541 tp->rcv_nxt == tp->rcv_wup) 4542 tcp_store_ts_recent(tp); 4543 4544 tcp_rcv_rtt_measure_ts(sk, skb); 4545 4546 if ((int)skb->truesize > sk->sk_forward_alloc) 4547 goto step5; 4548 4549 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS); 4550 4551 /* Bulk data transfer: receiver */ 4552 __skb_pull(skb,tcp_header_len); 4553 __skb_queue_tail(&sk->sk_receive_queue, skb); 4554 sk_stream_set_owner_r(skb, sk); 4555 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; 4556 } 4557 4558 tcp_event_data_recv(sk, skb); 4559 4560 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) { 4561 /* Well, only one small jumplet in fast path... */ 4562 tcp_ack(sk, skb, FLAG_DATA); 4563 tcp_data_snd_check(sk); 4564 if (!inet_csk_ack_scheduled(sk)) 4565 goto no_ack; 4566 } 4567 4568 __tcp_ack_snd_check(sk, 0); 4569 no_ack: 4570 #ifdef CONFIG_NET_DMA 4571 if (copied_early) 4572 __skb_queue_tail(&sk->sk_async_wait_queue, skb); 4573 else 4574 #endif 4575 if (eaten) 4576 __kfree_skb(skb); 4577 else 4578 sk->sk_data_ready(sk, 0); 4579 return 0; 4580 } 4581 } 4582 4583 slow_path: 4584 if (len < (th->doff<<2) || tcp_checksum_complete_user(sk, skb)) 4585 goto csum_error; 4586 4587 /* 4588 * RFC1323: H1. Apply PAWS check first. 4589 */ 4590 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp && 4591 tcp_paws_discard(sk, skb)) { 4592 if (!th->rst) { 4593 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED); 4594 tcp_send_dupack(sk, skb); 4595 goto discard; 4596 } 4597 /* Resets are accepted even if PAWS failed. 4598 4599 ts_recent update must be made after we are sure 4600 that the packet is in window. 4601 */ 4602 } 4603 4604 /* 4605 * Standard slow path. 4606 */ 4607 4608 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) { 4609 /* RFC793, page 37: "In all states except SYN-SENT, all reset 4610 * (RST) segments are validated by checking their SEQ-fields." 4611 * And page 69: "If an incoming segment is not acceptable, 4612 * an acknowledgment should be sent in reply (unless the RST bit 4613 * is set, if so drop the segment and return)". 4614 */ 4615 if (!th->rst) 4616 tcp_send_dupack(sk, skb); 4617 goto discard; 4618 } 4619 4620 if (th->rst) { 4621 tcp_reset(sk); 4622 goto discard; 4623 } 4624 4625 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq); 4626 4627 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 4628 TCP_INC_STATS_BH(TCP_MIB_INERRS); 4629 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN); 4630 tcp_reset(sk); 4631 return 1; 4632 } 4633 4634 step5: 4635 if (th->ack) 4636 tcp_ack(sk, skb, FLAG_SLOWPATH); 4637 4638 tcp_rcv_rtt_measure_ts(sk, skb); 4639 4640 /* Process urgent data. */ 4641 tcp_urg(sk, skb, th); 4642 4643 /* step 7: process the segment text */ 4644 tcp_data_queue(sk, skb); 4645 4646 tcp_data_snd_check(sk); 4647 tcp_ack_snd_check(sk); 4648 return 0; 4649 4650 csum_error: 4651 TCP_INC_STATS_BH(TCP_MIB_INERRS); 4652 4653 discard: 4654 __kfree_skb(skb); 4655 return 0; 4656 } 4657 4658 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb, 4659 struct tcphdr *th, unsigned len) 4660 { 4661 struct tcp_sock *tp = tcp_sk(sk); 4662 struct inet_connection_sock *icsk = inet_csk(sk); 4663 int saved_clamp = tp->rx_opt.mss_clamp; 4664 4665 tcp_parse_options(skb, &tp->rx_opt, 0); 4666 4667 if (th->ack) { 4668 /* rfc793: 4669 * "If the state is SYN-SENT then 4670 * first check the ACK bit 4671 * If the ACK bit is set 4672 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send 4673 * a reset (unless the RST bit is set, if so drop 4674 * the segment and return)" 4675 * 4676 * We do not send data with SYN, so that RFC-correct 4677 * test reduces to: 4678 */ 4679 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt) 4680 goto reset_and_undo; 4681 4682 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 4683 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp, 4684 tcp_time_stamp)) { 4685 NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED); 4686 goto reset_and_undo; 4687 } 4688 4689 /* Now ACK is acceptable. 4690 * 4691 * "If the RST bit is set 4692 * If the ACK was acceptable then signal the user "error: 4693 * connection reset", drop the segment, enter CLOSED state, 4694 * delete TCB, and return." 4695 */ 4696 4697 if (th->rst) { 4698 tcp_reset(sk); 4699 goto discard; 4700 } 4701 4702 /* rfc793: 4703 * "fifth, if neither of the SYN or RST bits is set then 4704 * drop the segment and return." 4705 * 4706 * See note below! 4707 * --ANK(990513) 4708 */ 4709 if (!th->syn) 4710 goto discard_and_undo; 4711 4712 /* rfc793: 4713 * "If the SYN bit is on ... 4714 * are acceptable then ... 4715 * (our SYN has been ACKed), change the connection 4716 * state to ESTABLISHED..." 4717 */ 4718 4719 TCP_ECN_rcv_synack(tp, th); 4720 4721 tp->snd_wl1 = TCP_SKB_CB(skb)->seq; 4722 tcp_ack(sk, skb, FLAG_SLOWPATH); 4723 4724 /* Ok.. it's good. Set up sequence numbers and 4725 * move to established. 4726 */ 4727 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; 4728 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; 4729 4730 /* RFC1323: The window in SYN & SYN/ACK segments is 4731 * never scaled. 4732 */ 4733 tp->snd_wnd = ntohs(th->window); 4734 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(skb)->seq); 4735 4736 if (!tp->rx_opt.wscale_ok) { 4737 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0; 4738 tp->window_clamp = min(tp->window_clamp, 65535U); 4739 } 4740 4741 if (tp->rx_opt.saw_tstamp) { 4742 tp->rx_opt.tstamp_ok = 1; 4743 tp->tcp_header_len = 4744 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 4745 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; 4746 tcp_store_ts_recent(tp); 4747 } else { 4748 tp->tcp_header_len = sizeof(struct tcphdr); 4749 } 4750 4751 if (tcp_is_sack(tp) && sysctl_tcp_fack) 4752 tcp_enable_fack(tp); 4753 4754 tcp_mtup_init(sk); 4755 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 4756 tcp_initialize_rcv_mss(sk); 4757 4758 /* Remember, tcp_poll() does not lock socket! 4759 * Change state from SYN-SENT only after copied_seq 4760 * is initialized. */ 4761 tp->copied_seq = tp->rcv_nxt; 4762 smp_mb(); 4763 tcp_set_state(sk, TCP_ESTABLISHED); 4764 4765 security_inet_conn_established(sk, skb); 4766 4767 /* Make sure socket is routed, for correct metrics. */ 4768 icsk->icsk_af_ops->rebuild_header(sk); 4769 4770 tcp_init_metrics(sk); 4771 4772 tcp_init_congestion_control(sk); 4773 4774 /* Prevent spurious tcp_cwnd_restart() on first data 4775 * packet. 4776 */ 4777 tp->lsndtime = tcp_time_stamp; 4778 4779 tcp_init_buffer_space(sk); 4780 4781 if (sock_flag(sk, SOCK_KEEPOPEN)) 4782 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp)); 4783 4784 if (!tp->rx_opt.snd_wscale) 4785 __tcp_fast_path_on(tp, tp->snd_wnd); 4786 else 4787 tp->pred_flags = 0; 4788 4789 if (!sock_flag(sk, SOCK_DEAD)) { 4790 sk->sk_state_change(sk); 4791 sk_wake_async(sk, 0, POLL_OUT); 4792 } 4793 4794 if (sk->sk_write_pending || 4795 icsk->icsk_accept_queue.rskq_defer_accept || 4796 icsk->icsk_ack.pingpong) { 4797 /* Save one ACK. Data will be ready after 4798 * several ticks, if write_pending is set. 4799 * 4800 * It may be deleted, but with this feature tcpdumps 4801 * look so _wonderfully_ clever, that I was not able 4802 * to stand against the temptation 8) --ANK 4803 */ 4804 inet_csk_schedule_ack(sk); 4805 icsk->icsk_ack.lrcvtime = tcp_time_stamp; 4806 icsk->icsk_ack.ato = TCP_ATO_MIN; 4807 tcp_incr_quickack(sk); 4808 tcp_enter_quickack_mode(sk); 4809 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 4810 TCP_DELACK_MAX, TCP_RTO_MAX); 4811 4812 discard: 4813 __kfree_skb(skb); 4814 return 0; 4815 } else { 4816 tcp_send_ack(sk); 4817 } 4818 return -1; 4819 } 4820 4821 /* No ACK in the segment */ 4822 4823 if (th->rst) { 4824 /* rfc793: 4825 * "If the RST bit is set 4826 * 4827 * Otherwise (no ACK) drop the segment and return." 4828 */ 4829 4830 goto discard_and_undo; 4831 } 4832 4833 /* PAWS check. */ 4834 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && tcp_paws_check(&tp->rx_opt, 0)) 4835 goto discard_and_undo; 4836 4837 if (th->syn) { 4838 /* We see SYN without ACK. It is attempt of 4839 * simultaneous connect with crossed SYNs. 4840 * Particularly, it can be connect to self. 4841 */ 4842 tcp_set_state(sk, TCP_SYN_RECV); 4843 4844 if (tp->rx_opt.saw_tstamp) { 4845 tp->rx_opt.tstamp_ok = 1; 4846 tcp_store_ts_recent(tp); 4847 tp->tcp_header_len = 4848 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 4849 } else { 4850 tp->tcp_header_len = sizeof(struct tcphdr); 4851 } 4852 4853 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; 4854 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; 4855 4856 /* RFC1323: The window in SYN & SYN/ACK segments is 4857 * never scaled. 4858 */ 4859 tp->snd_wnd = ntohs(th->window); 4860 tp->snd_wl1 = TCP_SKB_CB(skb)->seq; 4861 tp->max_window = tp->snd_wnd; 4862 4863 TCP_ECN_rcv_syn(tp, th); 4864 4865 tcp_mtup_init(sk); 4866 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 4867 tcp_initialize_rcv_mss(sk); 4868 4869 4870 tcp_send_synack(sk); 4871 #if 0 4872 /* Note, we could accept data and URG from this segment. 4873 * There are no obstacles to make this. 4874 * 4875 * However, if we ignore data in ACKless segments sometimes, 4876 * we have no reasons to accept it sometimes. 4877 * Also, seems the code doing it in step6 of tcp_rcv_state_process 4878 * is not flawless. So, discard packet for sanity. 4879 * Uncomment this return to process the data. 4880 */ 4881 return -1; 4882 #else 4883 goto discard; 4884 #endif 4885 } 4886 /* "fifth, if neither of the SYN or RST bits is set then 4887 * drop the segment and return." 4888 */ 4889 4890 discard_and_undo: 4891 tcp_clear_options(&tp->rx_opt); 4892 tp->rx_opt.mss_clamp = saved_clamp; 4893 goto discard; 4894 4895 reset_and_undo: 4896 tcp_clear_options(&tp->rx_opt); 4897 tp->rx_opt.mss_clamp = saved_clamp; 4898 return 1; 4899 } 4900 4901 4902 /* 4903 * This function implements the receiving procedure of RFC 793 for 4904 * all states except ESTABLISHED and TIME_WAIT. 4905 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be 4906 * address independent. 4907 */ 4908 4909 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb, 4910 struct tcphdr *th, unsigned len) 4911 { 4912 struct tcp_sock *tp = tcp_sk(sk); 4913 struct inet_connection_sock *icsk = inet_csk(sk); 4914 int queued = 0; 4915 4916 tp->rx_opt.saw_tstamp = 0; 4917 4918 switch (sk->sk_state) { 4919 case TCP_CLOSE: 4920 goto discard; 4921 4922 case TCP_LISTEN: 4923 if (th->ack) 4924 return 1; 4925 4926 if (th->rst) 4927 goto discard; 4928 4929 if (th->syn) { 4930 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0) 4931 return 1; 4932 4933 /* Now we have several options: In theory there is 4934 * nothing else in the frame. KA9Q has an option to 4935 * send data with the syn, BSD accepts data with the 4936 * syn up to the [to be] advertised window and 4937 * Solaris 2.1 gives you a protocol error. For now 4938 * we just ignore it, that fits the spec precisely 4939 * and avoids incompatibilities. It would be nice in 4940 * future to drop through and process the data. 4941 * 4942 * Now that TTCP is starting to be used we ought to 4943 * queue this data. 4944 * But, this leaves one open to an easy denial of 4945 * service attack, and SYN cookies can't defend 4946 * against this problem. So, we drop the data 4947 * in the interest of security over speed unless 4948 * it's still in use. 4949 */ 4950 kfree_skb(skb); 4951 return 0; 4952 } 4953 goto discard; 4954 4955 case TCP_SYN_SENT: 4956 queued = tcp_rcv_synsent_state_process(sk, skb, th, len); 4957 if (queued >= 0) 4958 return queued; 4959 4960 /* Do step6 onward by hand. */ 4961 tcp_urg(sk, skb, th); 4962 __kfree_skb(skb); 4963 tcp_data_snd_check(sk); 4964 return 0; 4965 } 4966 4967 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp && 4968 tcp_paws_discard(sk, skb)) { 4969 if (!th->rst) { 4970 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED); 4971 tcp_send_dupack(sk, skb); 4972 goto discard; 4973 } 4974 /* Reset is accepted even if it did not pass PAWS. */ 4975 } 4976 4977 /* step 1: check sequence number */ 4978 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) { 4979 if (!th->rst) 4980 tcp_send_dupack(sk, skb); 4981 goto discard; 4982 } 4983 4984 /* step 2: check RST bit */ 4985 if (th->rst) { 4986 tcp_reset(sk); 4987 goto discard; 4988 } 4989 4990 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq); 4991 4992 /* step 3: check security and precedence [ignored] */ 4993 4994 /* step 4: 4995 * 4996 * Check for a SYN in window. 4997 */ 4998 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 4999 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN); 5000 tcp_reset(sk); 5001 return 1; 5002 } 5003 5004 /* step 5: check the ACK field */ 5005 if (th->ack) { 5006 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH); 5007 5008 switch (sk->sk_state) { 5009 case TCP_SYN_RECV: 5010 if (acceptable) { 5011 tp->copied_seq = tp->rcv_nxt; 5012 smp_mb(); 5013 tcp_set_state(sk, TCP_ESTABLISHED); 5014 sk->sk_state_change(sk); 5015 5016 /* Note, that this wakeup is only for marginal 5017 * crossed SYN case. Passively open sockets 5018 * are not waked up, because sk->sk_sleep == 5019 * NULL and sk->sk_socket == NULL. 5020 */ 5021 if (sk->sk_socket) { 5022 sk_wake_async(sk,0,POLL_OUT); 5023 } 5024 5025 tp->snd_una = TCP_SKB_CB(skb)->ack_seq; 5026 tp->snd_wnd = ntohs(th->window) << 5027 tp->rx_opt.snd_wscale; 5028 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, 5029 TCP_SKB_CB(skb)->seq); 5030 5031 /* tcp_ack considers this ACK as duplicate 5032 * and does not calculate rtt. 5033 * Fix it at least with timestamps. 5034 */ 5035 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 5036 !tp->srtt) 5037 tcp_ack_saw_tstamp(sk, 0); 5038 5039 if (tp->rx_opt.tstamp_ok) 5040 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; 5041 5042 /* Make sure socket is routed, for 5043 * correct metrics. 5044 */ 5045 icsk->icsk_af_ops->rebuild_header(sk); 5046 5047 tcp_init_metrics(sk); 5048 5049 tcp_init_congestion_control(sk); 5050 5051 /* Prevent spurious tcp_cwnd_restart() on 5052 * first data packet. 5053 */ 5054 tp->lsndtime = tcp_time_stamp; 5055 5056 tcp_mtup_init(sk); 5057 tcp_initialize_rcv_mss(sk); 5058 tcp_init_buffer_space(sk); 5059 tcp_fast_path_on(tp); 5060 } else { 5061 return 1; 5062 } 5063 break; 5064 5065 case TCP_FIN_WAIT1: 5066 if (tp->snd_una == tp->write_seq) { 5067 tcp_set_state(sk, TCP_FIN_WAIT2); 5068 sk->sk_shutdown |= SEND_SHUTDOWN; 5069 dst_confirm(sk->sk_dst_cache); 5070 5071 if (!sock_flag(sk, SOCK_DEAD)) 5072 /* Wake up lingering close() */ 5073 sk->sk_state_change(sk); 5074 else { 5075 int tmo; 5076 5077 if (tp->linger2 < 0 || 5078 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 5079 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) { 5080 tcp_done(sk); 5081 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA); 5082 return 1; 5083 } 5084 5085 tmo = tcp_fin_time(sk); 5086 if (tmo > TCP_TIMEWAIT_LEN) { 5087 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN); 5088 } else if (th->fin || sock_owned_by_user(sk)) { 5089 /* Bad case. We could lose such FIN otherwise. 5090 * It is not a big problem, but it looks confusing 5091 * and not so rare event. We still can lose it now, 5092 * if it spins in bh_lock_sock(), but it is really 5093 * marginal case. 5094 */ 5095 inet_csk_reset_keepalive_timer(sk, tmo); 5096 } else { 5097 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo); 5098 goto discard; 5099 } 5100 } 5101 } 5102 break; 5103 5104 case TCP_CLOSING: 5105 if (tp->snd_una == tp->write_seq) { 5106 tcp_time_wait(sk, TCP_TIME_WAIT, 0); 5107 goto discard; 5108 } 5109 break; 5110 5111 case TCP_LAST_ACK: 5112 if (tp->snd_una == tp->write_seq) { 5113 tcp_update_metrics(sk); 5114 tcp_done(sk); 5115 goto discard; 5116 } 5117 break; 5118 } 5119 } else 5120 goto discard; 5121 5122 /* step 6: check the URG bit */ 5123 tcp_urg(sk, skb, th); 5124 5125 /* step 7: process the segment text */ 5126 switch (sk->sk_state) { 5127 case TCP_CLOSE_WAIT: 5128 case TCP_CLOSING: 5129 case TCP_LAST_ACK: 5130 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) 5131 break; 5132 case TCP_FIN_WAIT1: 5133 case TCP_FIN_WAIT2: 5134 /* RFC 793 says to queue data in these states, 5135 * RFC 1122 says we MUST send a reset. 5136 * BSD 4.4 also does reset. 5137 */ 5138 if (sk->sk_shutdown & RCV_SHUTDOWN) { 5139 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 5140 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) { 5141 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA); 5142 tcp_reset(sk); 5143 return 1; 5144 } 5145 } 5146 /* Fall through */ 5147 case TCP_ESTABLISHED: 5148 tcp_data_queue(sk, skb); 5149 queued = 1; 5150 break; 5151 } 5152 5153 /* tcp_data could move socket to TIME-WAIT */ 5154 if (sk->sk_state != TCP_CLOSE) { 5155 tcp_data_snd_check(sk); 5156 tcp_ack_snd_check(sk); 5157 } 5158 5159 if (!queued) { 5160 discard: 5161 __kfree_skb(skb); 5162 } 5163 return 0; 5164 } 5165 5166 EXPORT_SYMBOL(sysctl_tcp_ecn); 5167 EXPORT_SYMBOL(sysctl_tcp_reordering); 5168 EXPORT_SYMBOL(tcp_parse_options); 5169 EXPORT_SYMBOL(tcp_rcv_established); 5170 EXPORT_SYMBOL(tcp_rcv_state_process); 5171 EXPORT_SYMBOL(tcp_initialize_rcv_mss); 5172