tcp_input.c revision cda42ebd67ee5fdf09d7057b5a4584d36fe8a335
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 * Authors: Ross Biro 9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> 10 * Mark Evans, <evansmp@uhura.aston.ac.uk> 11 * Corey Minyard <wf-rch!minyard@relay.EU.net> 12 * Florian La Roche, <flla@stud.uni-sb.de> 13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu> 14 * Linus Torvalds, <torvalds@cs.helsinki.fi> 15 * Alan Cox, <gw4pts@gw4pts.ampr.org> 16 * Matthew Dillon, <dillon@apollo.west.oic.com> 17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no> 18 * Jorge Cwik, <jorge@laser.satlink.net> 19 */ 20 21/* 22 * Changes: 23 * Pedro Roque : Fast Retransmit/Recovery. 24 * Two receive queues. 25 * Retransmit queue handled by TCP. 26 * Better retransmit timer handling. 27 * New congestion avoidance. 28 * Header prediction. 29 * Variable renaming. 30 * 31 * Eric : Fast Retransmit. 32 * Randy Scott : MSS option defines. 33 * Eric Schenk : Fixes to slow start algorithm. 34 * Eric Schenk : Yet another double ACK bug. 35 * Eric Schenk : Delayed ACK bug fixes. 36 * Eric Schenk : Floyd style fast retrans war avoidance. 37 * David S. Miller : Don't allow zero congestion window. 38 * Eric Schenk : Fix retransmitter so that it sends 39 * next packet on ack of previous packet. 40 * Andi Kleen : Moved open_request checking here 41 * and process RSTs for open_requests. 42 * Andi Kleen : Better prune_queue, and other fixes. 43 * Andrey Savochkin: Fix RTT measurements in the presence of 44 * timestamps. 45 * Andrey Savochkin: Check sequence numbers correctly when 46 * removing SACKs due to in sequence incoming 47 * data segments. 48 * Andi Kleen: Make sure we never ack data there is not 49 * enough room for. Also make this condition 50 * a fatal error if it might still happen. 51 * Andi Kleen: Add tcp_measure_rcv_mss to make 52 * connections with MSS<min(MTU,ann. MSS) 53 * work without delayed acks. 54 * Andi Kleen: Process packets with PSH set in the 55 * fast path. 56 * J Hadi Salim: ECN support 57 * Andrei Gurtov, 58 * Pasi Sarolahti, 59 * Panu Kuhlberg: Experimental audit of TCP (re)transmission 60 * engine. Lots of bugs are found. 61 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs 62 */ 63 64#include <linux/mm.h> 65#include <linux/module.h> 66#include <linux/sysctl.h> 67#include <linux/kernel.h> 68#include <net/dst.h> 69#include <net/tcp.h> 70#include <net/inet_common.h> 71#include <linux/ipsec.h> 72#include <asm/unaligned.h> 73#include <net/netdma.h> 74 75int sysctl_tcp_timestamps __read_mostly = 1; 76int sysctl_tcp_window_scaling __read_mostly = 1; 77int sysctl_tcp_sack __read_mostly = 1; 78int sysctl_tcp_fack __read_mostly = 1; 79int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH; 80int sysctl_tcp_ecn __read_mostly = 2; 81int sysctl_tcp_dsack __read_mostly = 1; 82int sysctl_tcp_app_win __read_mostly = 31; 83int sysctl_tcp_adv_win_scale __read_mostly = 2; 84 85int sysctl_tcp_stdurg __read_mostly; 86int sysctl_tcp_rfc1337 __read_mostly; 87int sysctl_tcp_max_orphans __read_mostly = NR_FILE; 88int sysctl_tcp_frto __read_mostly = 2; 89int sysctl_tcp_frto_response __read_mostly; 90int sysctl_tcp_nometrics_save __read_mostly; 91 92int sysctl_tcp_moderate_rcvbuf __read_mostly = 1; 93int sysctl_tcp_abc __read_mostly; 94 95#define FLAG_DATA 0x01 /* Incoming frame contained data. */ 96#define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */ 97#define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */ 98#define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */ 99#define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */ 100#define FLAG_DATA_SACKED 0x20 /* New SACK. */ 101#define FLAG_ECE 0x40 /* ECE in this ACK */ 102#define FLAG_DATA_LOST 0x80 /* SACK detected data lossage. */ 103#define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/ 104#define FLAG_ONLY_ORIG_SACKED 0x200 /* SACKs only non-rexmit sent before RTO */ 105#define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */ 106#define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */ 107#define FLAG_NONHEAD_RETRANS_ACKED 0x1000 /* Non-head rexmitted data was ACKed */ 108#define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */ 109 110#define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED) 111#define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED) 112#define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE) 113#define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED) 114#define FLAG_ANY_PROGRESS (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED) 115 116#define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH) 117#define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH)) 118 119/* Adapt the MSS value used to make delayed ack decision to the 120 * real world. 121 */ 122static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb) 123{ 124 struct inet_connection_sock *icsk = inet_csk(sk); 125 const unsigned int lss = icsk->icsk_ack.last_seg_size; 126 unsigned int len; 127 128 icsk->icsk_ack.last_seg_size = 0; 129 130 /* skb->len may jitter because of SACKs, even if peer 131 * sends good full-sized frames. 132 */ 133 len = skb_shinfo(skb)->gso_size ? : skb->len; 134 if (len >= icsk->icsk_ack.rcv_mss) { 135 icsk->icsk_ack.rcv_mss = len; 136 } else { 137 /* Otherwise, we make more careful check taking into account, 138 * that SACKs block is variable. 139 * 140 * "len" is invariant segment length, including TCP header. 141 */ 142 len += skb->data - skb_transport_header(skb); 143 if (len >= TCP_MIN_RCVMSS + sizeof(struct tcphdr) || 144 /* If PSH is not set, packet should be 145 * full sized, provided peer TCP is not badly broken. 146 * This observation (if it is correct 8)) allows 147 * to handle super-low mtu links fairly. 148 */ 149 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) && 150 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) { 151 /* Subtract also invariant (if peer is RFC compliant), 152 * tcp header plus fixed timestamp option length. 153 * Resulting "len" is MSS free of SACK jitter. 154 */ 155 len -= tcp_sk(sk)->tcp_header_len; 156 icsk->icsk_ack.last_seg_size = len; 157 if (len == lss) { 158 icsk->icsk_ack.rcv_mss = len; 159 return; 160 } 161 } 162 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED) 163 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2; 164 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED; 165 } 166} 167 168static void tcp_incr_quickack(struct sock *sk) 169{ 170 struct inet_connection_sock *icsk = inet_csk(sk); 171 unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss); 172 173 if (quickacks == 0) 174 quickacks = 2; 175 if (quickacks > icsk->icsk_ack.quick) 176 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS); 177} 178 179void tcp_enter_quickack_mode(struct sock *sk) 180{ 181 struct inet_connection_sock *icsk = inet_csk(sk); 182 tcp_incr_quickack(sk); 183 icsk->icsk_ack.pingpong = 0; 184 icsk->icsk_ack.ato = TCP_ATO_MIN; 185} 186 187/* Send ACKs quickly, if "quick" count is not exhausted 188 * and the session is not interactive. 189 */ 190 191static inline int tcp_in_quickack_mode(const struct sock *sk) 192{ 193 const struct inet_connection_sock *icsk = inet_csk(sk); 194 return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong; 195} 196 197static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp) 198{ 199 if (tp->ecn_flags & TCP_ECN_OK) 200 tp->ecn_flags |= TCP_ECN_QUEUE_CWR; 201} 202 203static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, struct sk_buff *skb) 204{ 205 if (tcp_hdr(skb)->cwr) 206 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR; 207} 208 209static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp) 210{ 211 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR; 212} 213 214static inline void TCP_ECN_check_ce(struct tcp_sock *tp, struct sk_buff *skb) 215{ 216 if (tp->ecn_flags & TCP_ECN_OK) { 217 if (INET_ECN_is_ce(TCP_SKB_CB(skb)->flags)) 218 tp->ecn_flags |= TCP_ECN_DEMAND_CWR; 219 /* Funny extension: if ECT is not set on a segment, 220 * it is surely retransmit. It is not in ECN RFC, 221 * but Linux follows this rule. */ 222 else if (INET_ECN_is_not_ect((TCP_SKB_CB(skb)->flags))) 223 tcp_enter_quickack_mode((struct sock *)tp); 224 } 225} 226 227static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, struct tcphdr *th) 228{ 229 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr)) 230 tp->ecn_flags &= ~TCP_ECN_OK; 231} 232 233static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, struct tcphdr *th) 234{ 235 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr)) 236 tp->ecn_flags &= ~TCP_ECN_OK; 237} 238 239static inline int TCP_ECN_rcv_ecn_echo(struct tcp_sock *tp, struct tcphdr *th) 240{ 241 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK)) 242 return 1; 243 return 0; 244} 245 246/* Buffer size and advertised window tuning. 247 * 248 * 1. Tuning sk->sk_sndbuf, when connection enters established state. 249 */ 250 251static void tcp_fixup_sndbuf(struct sock *sk) 252{ 253 int sndmem = tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER + 16 + 254 sizeof(struct sk_buff); 255 256 if (sk->sk_sndbuf < 3 * sndmem) 257 sk->sk_sndbuf = min(3 * sndmem, sysctl_tcp_wmem[2]); 258} 259 260/* 2. Tuning advertised window (window_clamp, rcv_ssthresh) 261 * 262 * All tcp_full_space() is split to two parts: "network" buffer, allocated 263 * forward and advertised in receiver window (tp->rcv_wnd) and 264 * "application buffer", required to isolate scheduling/application 265 * latencies from network. 266 * window_clamp is maximal advertised window. It can be less than 267 * tcp_full_space(), in this case tcp_full_space() - window_clamp 268 * is reserved for "application" buffer. The less window_clamp is 269 * the smoother our behaviour from viewpoint of network, but the lower 270 * throughput and the higher sensitivity of the connection to losses. 8) 271 * 272 * rcv_ssthresh is more strict window_clamp used at "slow start" 273 * phase to predict further behaviour of this connection. 274 * It is used for two goals: 275 * - to enforce header prediction at sender, even when application 276 * requires some significant "application buffer". It is check #1. 277 * - to prevent pruning of receive queue because of misprediction 278 * of receiver window. Check #2. 279 * 280 * The scheme does not work when sender sends good segments opening 281 * window and then starts to feed us spaghetti. But it should work 282 * in common situations. Otherwise, we have to rely on queue collapsing. 283 */ 284 285/* Slow part of check#2. */ 286static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb) 287{ 288 struct tcp_sock *tp = tcp_sk(sk); 289 /* Optimize this! */ 290 int truesize = tcp_win_from_space(skb->truesize) >> 1; 291 int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1; 292 293 while (tp->rcv_ssthresh <= window) { 294 if (truesize <= skb->len) 295 return 2 * inet_csk(sk)->icsk_ack.rcv_mss; 296 297 truesize >>= 1; 298 window >>= 1; 299 } 300 return 0; 301} 302 303static void tcp_grow_window(struct sock *sk, struct sk_buff *skb) 304{ 305 struct tcp_sock *tp = tcp_sk(sk); 306 307 /* Check #1 */ 308 if (tp->rcv_ssthresh < tp->window_clamp && 309 (int)tp->rcv_ssthresh < tcp_space(sk) && 310 !tcp_memory_pressure) { 311 int incr; 312 313 /* Check #2. Increase window, if skb with such overhead 314 * will fit to rcvbuf in future. 315 */ 316 if (tcp_win_from_space(skb->truesize) <= skb->len) 317 incr = 2 * tp->advmss; 318 else 319 incr = __tcp_grow_window(sk, skb); 320 321 if (incr) { 322 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr, 323 tp->window_clamp); 324 inet_csk(sk)->icsk_ack.quick |= 1; 325 } 326 } 327} 328 329/* 3. Tuning rcvbuf, when connection enters established state. */ 330 331static void tcp_fixup_rcvbuf(struct sock *sk) 332{ 333 struct tcp_sock *tp = tcp_sk(sk); 334 int rcvmem = tp->advmss + MAX_TCP_HEADER + 16 + sizeof(struct sk_buff); 335 336 /* Try to select rcvbuf so that 4 mss-sized segments 337 * will fit to window and corresponding skbs will fit to our rcvbuf. 338 * (was 3; 4 is minimum to allow fast retransmit to work.) 339 */ 340 while (tcp_win_from_space(rcvmem) < tp->advmss) 341 rcvmem += 128; 342 if (sk->sk_rcvbuf < 4 * rcvmem) 343 sk->sk_rcvbuf = min(4 * rcvmem, sysctl_tcp_rmem[2]); 344} 345 346/* 4. Try to fixup all. It is made immediately after connection enters 347 * established state. 348 */ 349static void tcp_init_buffer_space(struct sock *sk) 350{ 351 struct tcp_sock *tp = tcp_sk(sk); 352 int maxwin; 353 354 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) 355 tcp_fixup_rcvbuf(sk); 356 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK)) 357 tcp_fixup_sndbuf(sk); 358 359 tp->rcvq_space.space = tp->rcv_wnd; 360 361 maxwin = tcp_full_space(sk); 362 363 if (tp->window_clamp >= maxwin) { 364 tp->window_clamp = maxwin; 365 366 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss) 367 tp->window_clamp = max(maxwin - 368 (maxwin >> sysctl_tcp_app_win), 369 4 * tp->advmss); 370 } 371 372 /* Force reservation of one segment. */ 373 if (sysctl_tcp_app_win && 374 tp->window_clamp > 2 * tp->advmss && 375 tp->window_clamp + tp->advmss > maxwin) 376 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss); 377 378 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp); 379 tp->snd_cwnd_stamp = tcp_time_stamp; 380} 381 382/* 5. Recalculate window clamp after socket hit its memory bounds. */ 383static void tcp_clamp_window(struct sock *sk) 384{ 385 struct tcp_sock *tp = tcp_sk(sk); 386 struct inet_connection_sock *icsk = inet_csk(sk); 387 388 icsk->icsk_ack.quick = 0; 389 390 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] && 391 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) && 392 !tcp_memory_pressure && 393 atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) { 394 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc), 395 sysctl_tcp_rmem[2]); 396 } 397 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf) 398 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss); 399} 400 401/* Initialize RCV_MSS value. 402 * RCV_MSS is an our guess about MSS used by the peer. 403 * We haven't any direct information about the MSS. 404 * It's better to underestimate the RCV_MSS rather than overestimate. 405 * Overestimations make us ACKing less frequently than needed. 406 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss(). 407 */ 408void tcp_initialize_rcv_mss(struct sock *sk) 409{ 410 struct tcp_sock *tp = tcp_sk(sk); 411 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache); 412 413 hint = min(hint, tp->rcv_wnd / 2); 414 hint = min(hint, TCP_MIN_RCVMSS); 415 hint = max(hint, TCP_MIN_MSS); 416 417 inet_csk(sk)->icsk_ack.rcv_mss = hint; 418} 419 420/* Receiver "autotuning" code. 421 * 422 * The algorithm for RTT estimation w/o timestamps is based on 423 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL. 424 * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps> 425 * 426 * More detail on this code can be found at 427 * <http://www.psc.edu/~jheffner/senior_thesis.ps>, 428 * though this reference is out of date. A new paper 429 * is pending. 430 */ 431static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep) 432{ 433 u32 new_sample = tp->rcv_rtt_est.rtt; 434 long m = sample; 435 436 if (m == 0) 437 m = 1; 438 439 if (new_sample != 0) { 440 /* If we sample in larger samples in the non-timestamp 441 * case, we could grossly overestimate the RTT especially 442 * with chatty applications or bulk transfer apps which 443 * are stalled on filesystem I/O. 444 * 445 * Also, since we are only going for a minimum in the 446 * non-timestamp case, we do not smooth things out 447 * else with timestamps disabled convergence takes too 448 * long. 449 */ 450 if (!win_dep) { 451 m -= (new_sample >> 3); 452 new_sample += m; 453 } else if (m < new_sample) 454 new_sample = m << 3; 455 } else { 456 /* No previous measure. */ 457 new_sample = m << 3; 458 } 459 460 if (tp->rcv_rtt_est.rtt != new_sample) 461 tp->rcv_rtt_est.rtt = new_sample; 462} 463 464static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp) 465{ 466 if (tp->rcv_rtt_est.time == 0) 467 goto new_measure; 468 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq)) 469 return; 470 tcp_rcv_rtt_update(tp, jiffies - tp->rcv_rtt_est.time, 1); 471 472new_measure: 473 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd; 474 tp->rcv_rtt_est.time = tcp_time_stamp; 475} 476 477static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, 478 const struct sk_buff *skb) 479{ 480 struct tcp_sock *tp = tcp_sk(sk); 481 if (tp->rx_opt.rcv_tsecr && 482 (TCP_SKB_CB(skb)->end_seq - 483 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss)) 484 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0); 485} 486 487/* 488 * This function should be called every time data is copied to user space. 489 * It calculates the appropriate TCP receive buffer space. 490 */ 491void tcp_rcv_space_adjust(struct sock *sk) 492{ 493 struct tcp_sock *tp = tcp_sk(sk); 494 int time; 495 int space; 496 497 if (tp->rcvq_space.time == 0) 498 goto new_measure; 499 500 time = tcp_time_stamp - tp->rcvq_space.time; 501 if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0) 502 return; 503 504 space = 2 * (tp->copied_seq - tp->rcvq_space.seq); 505 506 space = max(tp->rcvq_space.space, space); 507 508 if (tp->rcvq_space.space != space) { 509 int rcvmem; 510 511 tp->rcvq_space.space = space; 512 513 if (sysctl_tcp_moderate_rcvbuf && 514 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) { 515 int new_clamp = space; 516 517 /* Receive space grows, normalize in order to 518 * take into account packet headers and sk_buff 519 * structure overhead. 520 */ 521 space /= tp->advmss; 522 if (!space) 523 space = 1; 524 rcvmem = (tp->advmss + MAX_TCP_HEADER + 525 16 + sizeof(struct sk_buff)); 526 while (tcp_win_from_space(rcvmem) < tp->advmss) 527 rcvmem += 128; 528 space *= rcvmem; 529 space = min(space, sysctl_tcp_rmem[2]); 530 if (space > sk->sk_rcvbuf) { 531 sk->sk_rcvbuf = space; 532 533 /* Make the window clamp follow along. */ 534 tp->window_clamp = new_clamp; 535 } 536 } 537 } 538 539new_measure: 540 tp->rcvq_space.seq = tp->copied_seq; 541 tp->rcvq_space.time = tcp_time_stamp; 542} 543 544/* There is something which you must keep in mind when you analyze the 545 * behavior of the tp->ato delayed ack timeout interval. When a 546 * connection starts up, we want to ack as quickly as possible. The 547 * problem is that "good" TCP's do slow start at the beginning of data 548 * transmission. The means that until we send the first few ACK's the 549 * sender will sit on his end and only queue most of his data, because 550 * he can only send snd_cwnd unacked packets at any given time. For 551 * each ACK we send, he increments snd_cwnd and transmits more of his 552 * queue. -DaveM 553 */ 554static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb) 555{ 556 struct tcp_sock *tp = tcp_sk(sk); 557 struct inet_connection_sock *icsk = inet_csk(sk); 558 u32 now; 559 560 inet_csk_schedule_ack(sk); 561 562 tcp_measure_rcv_mss(sk, skb); 563 564 tcp_rcv_rtt_measure(tp); 565 566 now = tcp_time_stamp; 567 568 if (!icsk->icsk_ack.ato) { 569 /* The _first_ data packet received, initialize 570 * delayed ACK engine. 571 */ 572 tcp_incr_quickack(sk); 573 icsk->icsk_ack.ato = TCP_ATO_MIN; 574 } else { 575 int m = now - icsk->icsk_ack.lrcvtime; 576 577 if (m <= TCP_ATO_MIN / 2) { 578 /* The fastest case is the first. */ 579 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2; 580 } else if (m < icsk->icsk_ack.ato) { 581 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m; 582 if (icsk->icsk_ack.ato > icsk->icsk_rto) 583 icsk->icsk_ack.ato = icsk->icsk_rto; 584 } else if (m > icsk->icsk_rto) { 585 /* Too long gap. Apparently sender failed to 586 * restart window, so that we send ACKs quickly. 587 */ 588 tcp_incr_quickack(sk); 589 sk_mem_reclaim(sk); 590 } 591 } 592 icsk->icsk_ack.lrcvtime = now; 593 594 TCP_ECN_check_ce(tp, skb); 595 596 if (skb->len >= 128) 597 tcp_grow_window(sk, skb); 598} 599 600/* Called to compute a smoothed rtt estimate. The data fed to this 601 * routine either comes from timestamps, or from segments that were 602 * known _not_ to have been retransmitted [see Karn/Partridge 603 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88 604 * piece by Van Jacobson. 605 * NOTE: the next three routines used to be one big routine. 606 * To save cycles in the RFC 1323 implementation it was better to break 607 * it up into three procedures. -- erics 608 */ 609static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt) 610{ 611 struct tcp_sock *tp = tcp_sk(sk); 612 long m = mrtt; /* RTT */ 613 614 /* The following amusing code comes from Jacobson's 615 * article in SIGCOMM '88. Note that rtt and mdev 616 * are scaled versions of rtt and mean deviation. 617 * This is designed to be as fast as possible 618 * m stands for "measurement". 619 * 620 * On a 1990 paper the rto value is changed to: 621 * RTO = rtt + 4 * mdev 622 * 623 * Funny. This algorithm seems to be very broken. 624 * These formulae increase RTO, when it should be decreased, increase 625 * too slowly, when it should be increased quickly, decrease too quickly 626 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely 627 * does not matter how to _calculate_ it. Seems, it was trap 628 * that VJ failed to avoid. 8) 629 */ 630 if (m == 0) 631 m = 1; 632 if (tp->srtt != 0) { 633 m -= (tp->srtt >> 3); /* m is now error in rtt est */ 634 tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */ 635 if (m < 0) { 636 m = -m; /* m is now abs(error) */ 637 m -= (tp->mdev >> 2); /* similar update on mdev */ 638 /* This is similar to one of Eifel findings. 639 * Eifel blocks mdev updates when rtt decreases. 640 * This solution is a bit different: we use finer gain 641 * for mdev in this case (alpha*beta). 642 * Like Eifel it also prevents growth of rto, 643 * but also it limits too fast rto decreases, 644 * happening in pure Eifel. 645 */ 646 if (m > 0) 647 m >>= 3; 648 } else { 649 m -= (tp->mdev >> 2); /* similar update on mdev */ 650 } 651 tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */ 652 if (tp->mdev > tp->mdev_max) { 653 tp->mdev_max = tp->mdev; 654 if (tp->mdev_max > tp->rttvar) 655 tp->rttvar = tp->mdev_max; 656 } 657 if (after(tp->snd_una, tp->rtt_seq)) { 658 if (tp->mdev_max < tp->rttvar) 659 tp->rttvar -= (tp->rttvar - tp->mdev_max) >> 2; 660 tp->rtt_seq = tp->snd_nxt; 661 tp->mdev_max = tcp_rto_min(sk); 662 } 663 } else { 664 /* no previous measure. */ 665 tp->srtt = m << 3; /* take the measured time to be rtt */ 666 tp->mdev = m << 1; /* make sure rto = 3*rtt */ 667 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk)); 668 tp->rtt_seq = tp->snd_nxt; 669 } 670} 671 672/* Calculate rto without backoff. This is the second half of Van Jacobson's 673 * routine referred to above. 674 */ 675static inline void tcp_set_rto(struct sock *sk) 676{ 677 const struct tcp_sock *tp = tcp_sk(sk); 678 /* Old crap is replaced with new one. 8) 679 * 680 * More seriously: 681 * 1. If rtt variance happened to be less 50msec, it is hallucination. 682 * It cannot be less due to utterly erratic ACK generation made 683 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_ 684 * to do with delayed acks, because at cwnd>2 true delack timeout 685 * is invisible. Actually, Linux-2.4 also generates erratic 686 * ACKs in some circumstances. 687 */ 688 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp); 689 690 /* 2. Fixups made earlier cannot be right. 691 * If we do not estimate RTO correctly without them, 692 * all the algo is pure shit and should be replaced 693 * with correct one. It is exactly, which we pretend to do. 694 */ 695 696 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo 697 * guarantees that rto is higher. 698 */ 699 tcp_bound_rto(sk); 700} 701 702/* Save metrics learned by this TCP session. 703 This function is called only, when TCP finishes successfully 704 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE. 705 */ 706void tcp_update_metrics(struct sock *sk) 707{ 708 struct tcp_sock *tp = tcp_sk(sk); 709 struct dst_entry *dst = __sk_dst_get(sk); 710 711 if (sysctl_tcp_nometrics_save) 712 return; 713 714 dst_confirm(dst); 715 716 if (dst && (dst->flags & DST_HOST)) { 717 const struct inet_connection_sock *icsk = inet_csk(sk); 718 int m; 719 unsigned long rtt; 720 721 if (icsk->icsk_backoff || !tp->srtt) { 722 /* This session failed to estimate rtt. Why? 723 * Probably, no packets returned in time. 724 * Reset our results. 725 */ 726 if (!(dst_metric_locked(dst, RTAX_RTT))) 727 dst->metrics[RTAX_RTT - 1] = 0; 728 return; 729 } 730 731 rtt = dst_metric_rtt(dst, RTAX_RTT); 732 m = rtt - tp->srtt; 733 734 /* If newly calculated rtt larger than stored one, 735 * store new one. Otherwise, use EWMA. Remember, 736 * rtt overestimation is always better than underestimation. 737 */ 738 if (!(dst_metric_locked(dst, RTAX_RTT))) { 739 if (m <= 0) 740 set_dst_metric_rtt(dst, RTAX_RTT, tp->srtt); 741 else 742 set_dst_metric_rtt(dst, RTAX_RTT, rtt - (m >> 3)); 743 } 744 745 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) { 746 unsigned long var; 747 if (m < 0) 748 m = -m; 749 750 /* Scale deviation to rttvar fixed point */ 751 m >>= 1; 752 if (m < tp->mdev) 753 m = tp->mdev; 754 755 var = dst_metric_rtt(dst, RTAX_RTTVAR); 756 if (m >= var) 757 var = m; 758 else 759 var -= (var - m) >> 2; 760 761 set_dst_metric_rtt(dst, RTAX_RTTVAR, var); 762 } 763 764 if (tcp_in_initial_slowstart(tp)) { 765 /* Slow start still did not finish. */ 766 if (dst_metric(dst, RTAX_SSTHRESH) && 767 !dst_metric_locked(dst, RTAX_SSTHRESH) && 768 (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH)) 769 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1; 770 if (!dst_metric_locked(dst, RTAX_CWND) && 771 tp->snd_cwnd > dst_metric(dst, RTAX_CWND)) 772 dst->metrics[RTAX_CWND - 1] = tp->snd_cwnd; 773 } else if (tp->snd_cwnd > tp->snd_ssthresh && 774 icsk->icsk_ca_state == TCP_CA_Open) { 775 /* Cong. avoidance phase, cwnd is reliable. */ 776 if (!dst_metric_locked(dst, RTAX_SSTHRESH)) 777 dst->metrics[RTAX_SSTHRESH-1] = 778 max(tp->snd_cwnd >> 1, tp->snd_ssthresh); 779 if (!dst_metric_locked(dst, RTAX_CWND)) 780 dst->metrics[RTAX_CWND-1] = (dst_metric(dst, RTAX_CWND) + tp->snd_cwnd) >> 1; 781 } else { 782 /* Else slow start did not finish, cwnd is non-sense, 783 ssthresh may be also invalid. 784 */ 785 if (!dst_metric_locked(dst, RTAX_CWND)) 786 dst->metrics[RTAX_CWND-1] = (dst_metric(dst, RTAX_CWND) + tp->snd_ssthresh) >> 1; 787 if (dst_metric(dst, RTAX_SSTHRESH) && 788 !dst_metric_locked(dst, RTAX_SSTHRESH) && 789 tp->snd_ssthresh > dst_metric(dst, RTAX_SSTHRESH)) 790 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh; 791 } 792 793 if (!dst_metric_locked(dst, RTAX_REORDERING)) { 794 if (dst_metric(dst, RTAX_REORDERING) < tp->reordering && 795 tp->reordering != sysctl_tcp_reordering) 796 dst->metrics[RTAX_REORDERING-1] = tp->reordering; 797 } 798 } 799} 800 801/* Numbers are taken from RFC3390. 802 * 803 * John Heffner states: 804 * 805 * The RFC specifies a window of no more than 4380 bytes 806 * unless 2*MSS > 4380. Reading the pseudocode in the RFC 807 * is a bit misleading because they use a clamp at 4380 bytes 808 * rather than use a multiplier in the relevant range. 809 */ 810__u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst) 811{ 812 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0); 813 814 if (!cwnd) { 815 if (tp->mss_cache > 1460) 816 cwnd = 2; 817 else 818 cwnd = (tp->mss_cache > 1095) ? 3 : 4; 819 } 820 return min_t(__u32, cwnd, tp->snd_cwnd_clamp); 821} 822 823/* Set slow start threshold and cwnd not falling to slow start */ 824void tcp_enter_cwr(struct sock *sk, const int set_ssthresh) 825{ 826 struct tcp_sock *tp = tcp_sk(sk); 827 const struct inet_connection_sock *icsk = inet_csk(sk); 828 829 tp->prior_ssthresh = 0; 830 tp->bytes_acked = 0; 831 if (icsk->icsk_ca_state < TCP_CA_CWR) { 832 tp->undo_marker = 0; 833 if (set_ssthresh) 834 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 835 tp->snd_cwnd = min(tp->snd_cwnd, 836 tcp_packets_in_flight(tp) + 1U); 837 tp->snd_cwnd_cnt = 0; 838 tp->high_seq = tp->snd_nxt; 839 tp->snd_cwnd_stamp = tcp_time_stamp; 840 TCP_ECN_queue_cwr(tp); 841 842 tcp_set_ca_state(sk, TCP_CA_CWR); 843 } 844} 845 846/* 847 * Packet counting of FACK is based on in-order assumptions, therefore TCP 848 * disables it when reordering is detected 849 */ 850static void tcp_disable_fack(struct tcp_sock *tp) 851{ 852 /* RFC3517 uses different metric in lost marker => reset on change */ 853 if (tcp_is_fack(tp)) 854 tp->lost_skb_hint = NULL; 855 tp->rx_opt.sack_ok &= ~2; 856} 857 858/* Take a notice that peer is sending D-SACKs */ 859static void tcp_dsack_seen(struct tcp_sock *tp) 860{ 861 tp->rx_opt.sack_ok |= 4; 862} 863 864/* Initialize metrics on socket. */ 865 866static void tcp_init_metrics(struct sock *sk) 867{ 868 struct tcp_sock *tp = tcp_sk(sk); 869 struct dst_entry *dst = __sk_dst_get(sk); 870 871 if (dst == NULL) 872 goto reset; 873 874 dst_confirm(dst); 875 876 if (dst_metric_locked(dst, RTAX_CWND)) 877 tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND); 878 if (dst_metric(dst, RTAX_SSTHRESH)) { 879 tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH); 880 if (tp->snd_ssthresh > tp->snd_cwnd_clamp) 881 tp->snd_ssthresh = tp->snd_cwnd_clamp; 882 } 883 if (dst_metric(dst, RTAX_REORDERING) && 884 tp->reordering != dst_metric(dst, RTAX_REORDERING)) { 885 tcp_disable_fack(tp); 886 tp->reordering = dst_metric(dst, RTAX_REORDERING); 887 } 888 889 if (dst_metric(dst, RTAX_RTT) == 0) 890 goto reset; 891 892 if (!tp->srtt && dst_metric_rtt(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3)) 893 goto reset; 894 895 /* Initial rtt is determined from SYN,SYN-ACK. 896 * The segment is small and rtt may appear much 897 * less than real one. Use per-dst memory 898 * to make it more realistic. 899 * 900 * A bit of theory. RTT is time passed after "normal" sized packet 901 * is sent until it is ACKed. In normal circumstances sending small 902 * packets force peer to delay ACKs and calculation is correct too. 903 * The algorithm is adaptive and, provided we follow specs, it 904 * NEVER underestimate RTT. BUT! If peer tries to make some clever 905 * tricks sort of "quick acks" for time long enough to decrease RTT 906 * to low value, and then abruptly stops to do it and starts to delay 907 * ACKs, wait for troubles. 908 */ 909 if (dst_metric_rtt(dst, RTAX_RTT) > tp->srtt) { 910 tp->srtt = dst_metric_rtt(dst, RTAX_RTT); 911 tp->rtt_seq = tp->snd_nxt; 912 } 913 if (dst_metric_rtt(dst, RTAX_RTTVAR) > tp->mdev) { 914 tp->mdev = dst_metric_rtt(dst, RTAX_RTTVAR); 915 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk)); 916 } 917 tcp_set_rto(sk); 918 if (inet_csk(sk)->icsk_rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp) 919 goto reset; 920 921cwnd: 922 tp->snd_cwnd = tcp_init_cwnd(tp, dst); 923 tp->snd_cwnd_stamp = tcp_time_stamp; 924 return; 925 926reset: 927 /* Play conservative. If timestamps are not 928 * supported, TCP will fail to recalculate correct 929 * rtt, if initial rto is too small. FORGET ALL AND RESET! 930 */ 931 if (!tp->rx_opt.saw_tstamp && tp->srtt) { 932 tp->srtt = 0; 933 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT; 934 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; 935 } 936 goto cwnd; 937} 938 939static void tcp_update_reordering(struct sock *sk, const int metric, 940 const int ts) 941{ 942 struct tcp_sock *tp = tcp_sk(sk); 943 if (metric > tp->reordering) { 944 int mib_idx; 945 946 tp->reordering = min(TCP_MAX_REORDERING, metric); 947 948 /* This exciting event is worth to be remembered. 8) */ 949 if (ts) 950 mib_idx = LINUX_MIB_TCPTSREORDER; 951 else if (tcp_is_reno(tp)) 952 mib_idx = LINUX_MIB_TCPRENOREORDER; 953 else if (tcp_is_fack(tp)) 954 mib_idx = LINUX_MIB_TCPFACKREORDER; 955 else 956 mib_idx = LINUX_MIB_TCPSACKREORDER; 957 958 NET_INC_STATS_BH(sock_net(sk), mib_idx); 959#if FASTRETRANS_DEBUG > 1 960 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n", 961 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state, 962 tp->reordering, 963 tp->fackets_out, 964 tp->sacked_out, 965 tp->undo_marker ? tp->undo_retrans : 0); 966#endif 967 tcp_disable_fack(tp); 968 } 969} 970 971/* This must be called before lost_out is incremented */ 972static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb) 973{ 974 if ((tp->retransmit_skb_hint == NULL) || 975 before(TCP_SKB_CB(skb)->seq, 976 TCP_SKB_CB(tp->retransmit_skb_hint)->seq)) 977 tp->retransmit_skb_hint = skb; 978 979 if (!tp->lost_out || 980 after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high)) 981 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq; 982} 983 984static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb) 985{ 986 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) { 987 tcp_verify_retransmit_hint(tp, skb); 988 989 tp->lost_out += tcp_skb_pcount(skb); 990 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 991 } 992} 993 994static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp, 995 struct sk_buff *skb) 996{ 997 tcp_verify_retransmit_hint(tp, skb); 998 999 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) { 1000 tp->lost_out += tcp_skb_pcount(skb); 1001 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 1002 } 1003} 1004 1005/* This procedure tags the retransmission queue when SACKs arrive. 1006 * 1007 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L). 1008 * Packets in queue with these bits set are counted in variables 1009 * sacked_out, retrans_out and lost_out, correspondingly. 1010 * 1011 * Valid combinations are: 1012 * Tag InFlight Description 1013 * 0 1 - orig segment is in flight. 1014 * S 0 - nothing flies, orig reached receiver. 1015 * L 0 - nothing flies, orig lost by net. 1016 * R 2 - both orig and retransmit are in flight. 1017 * L|R 1 - orig is lost, retransmit is in flight. 1018 * S|R 1 - orig reached receiver, retrans is still in flight. 1019 * (L|S|R is logically valid, it could occur when L|R is sacked, 1020 * but it is equivalent to plain S and code short-curcuits it to S. 1021 * L|S is logically invalid, it would mean -1 packet in flight 8)) 1022 * 1023 * These 6 states form finite state machine, controlled by the following events: 1024 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue()) 1025 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue()) 1026 * 3. Loss detection event of one of three flavors: 1027 * A. Scoreboard estimator decided the packet is lost. 1028 * A'. Reno "three dupacks" marks head of queue lost. 1029 * A''. Its FACK modfication, head until snd.fack is lost. 1030 * B. SACK arrives sacking data transmitted after never retransmitted 1031 * hole was sent out. 1032 * C. SACK arrives sacking SND.NXT at the moment, when the 1033 * segment was retransmitted. 1034 * 4. D-SACK added new rule: D-SACK changes any tag to S. 1035 * 1036 * It is pleasant to note, that state diagram turns out to be commutative, 1037 * so that we are allowed not to be bothered by order of our actions, 1038 * when multiple events arrive simultaneously. (see the function below). 1039 * 1040 * Reordering detection. 1041 * -------------------- 1042 * Reordering metric is maximal distance, which a packet can be displaced 1043 * in packet stream. With SACKs we can estimate it: 1044 * 1045 * 1. SACK fills old hole and the corresponding segment was not 1046 * ever retransmitted -> reordering. Alas, we cannot use it 1047 * when segment was retransmitted. 1048 * 2. The last flaw is solved with D-SACK. D-SACK arrives 1049 * for retransmitted and already SACKed segment -> reordering.. 1050 * Both of these heuristics are not used in Loss state, when we cannot 1051 * account for retransmits accurately. 1052 * 1053 * SACK block validation. 1054 * ---------------------- 1055 * 1056 * SACK block range validation checks that the received SACK block fits to 1057 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT. 1058 * Note that SND.UNA is not included to the range though being valid because 1059 * it means that the receiver is rather inconsistent with itself reporting 1060 * SACK reneging when it should advance SND.UNA. Such SACK block this is 1061 * perfectly valid, however, in light of RFC2018 which explicitly states 1062 * that "SACK block MUST reflect the newest segment. Even if the newest 1063 * segment is going to be discarded ...", not that it looks very clever 1064 * in case of head skb. Due to potentional receiver driven attacks, we 1065 * choose to avoid immediate execution of a walk in write queue due to 1066 * reneging and defer head skb's loss recovery to standard loss recovery 1067 * procedure that will eventually trigger (nothing forbids us doing this). 1068 * 1069 * Implements also blockage to start_seq wrap-around. Problem lies in the 1070 * fact that though start_seq (s) is before end_seq (i.e., not reversed), 1071 * there's no guarantee that it will be before snd_nxt (n). The problem 1072 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt 1073 * wrap (s_w): 1074 * 1075 * <- outs wnd -> <- wrapzone -> 1076 * u e n u_w e_w s n_w 1077 * | | | | | | | 1078 * |<------------+------+----- TCP seqno space --------------+---------->| 1079 * ...-- <2^31 ->| |<--------... 1080 * ...---- >2^31 ------>| |<--------... 1081 * 1082 * Current code wouldn't be vulnerable but it's better still to discard such 1083 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat 1084 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in 1085 * snd_nxt wrap -> snd_una region will then become "well defined", i.e., 1086 * equal to the ideal case (infinite seqno space without wrap caused issues). 1087 * 1088 * With D-SACK the lower bound is extended to cover sequence space below 1089 * SND.UNA down to undo_marker, which is the last point of interest. Yet 1090 * again, D-SACK block must not to go across snd_una (for the same reason as 1091 * for the normal SACK blocks, explained above). But there all simplicity 1092 * ends, TCP might receive valid D-SACKs below that. As long as they reside 1093 * fully below undo_marker they do not affect behavior in anyway and can 1094 * therefore be safely ignored. In rare cases (which are more or less 1095 * theoretical ones), the D-SACK will nicely cross that boundary due to skb 1096 * fragmentation and packet reordering past skb's retransmission. To consider 1097 * them correctly, the acceptable range must be extended even more though 1098 * the exact amount is rather hard to quantify. However, tp->max_window can 1099 * be used as an exaggerated estimate. 1100 */ 1101static int tcp_is_sackblock_valid(struct tcp_sock *tp, int is_dsack, 1102 u32 start_seq, u32 end_seq) 1103{ 1104 /* Too far in future, or reversed (interpretation is ambiguous) */ 1105 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq)) 1106 return 0; 1107 1108 /* Nasty start_seq wrap-around check (see comments above) */ 1109 if (!before(start_seq, tp->snd_nxt)) 1110 return 0; 1111 1112 /* In outstanding window? ...This is valid exit for D-SACKs too. 1113 * start_seq == snd_una is non-sensical (see comments above) 1114 */ 1115 if (after(start_seq, tp->snd_una)) 1116 return 1; 1117 1118 if (!is_dsack || !tp->undo_marker) 1119 return 0; 1120 1121 /* ...Then it's D-SACK, and must reside below snd_una completely */ 1122 if (!after(end_seq, tp->snd_una)) 1123 return 0; 1124 1125 if (!before(start_seq, tp->undo_marker)) 1126 return 1; 1127 1128 /* Too old */ 1129 if (!after(end_seq, tp->undo_marker)) 1130 return 0; 1131 1132 /* Undo_marker boundary crossing (overestimates a lot). Known already: 1133 * start_seq < undo_marker and end_seq >= undo_marker. 1134 */ 1135 return !before(start_seq, end_seq - tp->max_window); 1136} 1137 1138/* Check for lost retransmit. This superb idea is borrowed from "ratehalving". 1139 * Event "C". Later note: FACK people cheated me again 8), we have to account 1140 * for reordering! Ugly, but should help. 1141 * 1142 * Search retransmitted skbs from write_queue that were sent when snd_nxt was 1143 * less than what is now known to be received by the other end (derived from 1144 * highest SACK block). Also calculate the lowest snd_nxt among the remaining 1145 * retransmitted skbs to avoid some costly processing per ACKs. 1146 */ 1147static void tcp_mark_lost_retrans(struct sock *sk) 1148{ 1149 const struct inet_connection_sock *icsk = inet_csk(sk); 1150 struct tcp_sock *tp = tcp_sk(sk); 1151 struct sk_buff *skb; 1152 int cnt = 0; 1153 u32 new_low_seq = tp->snd_nxt; 1154 u32 received_upto = tcp_highest_sack_seq(tp); 1155 1156 if (!tcp_is_fack(tp) || !tp->retrans_out || 1157 !after(received_upto, tp->lost_retrans_low) || 1158 icsk->icsk_ca_state != TCP_CA_Recovery) 1159 return; 1160 1161 tcp_for_write_queue(skb, sk) { 1162 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq; 1163 1164 if (skb == tcp_send_head(sk)) 1165 break; 1166 if (cnt == tp->retrans_out) 1167 break; 1168 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 1169 continue; 1170 1171 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)) 1172 continue; 1173 1174 /* TODO: We would like to get rid of tcp_is_fack(tp) only 1175 * constraint here (see above) but figuring out that at 1176 * least tp->reordering SACK blocks reside between ack_seq 1177 * and received_upto is not easy task to do cheaply with 1178 * the available datastructures. 1179 * 1180 * Whether FACK should check here for tp->reordering segs 1181 * in-between one could argue for either way (it would be 1182 * rather simple to implement as we could count fack_count 1183 * during the walk and do tp->fackets_out - fack_count). 1184 */ 1185 if (after(received_upto, ack_seq)) { 1186 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 1187 tp->retrans_out -= tcp_skb_pcount(skb); 1188 1189 tcp_skb_mark_lost_uncond_verify(tp, skb); 1190 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT); 1191 } else { 1192 if (before(ack_seq, new_low_seq)) 1193 new_low_seq = ack_seq; 1194 cnt += tcp_skb_pcount(skb); 1195 } 1196 } 1197 1198 if (tp->retrans_out) 1199 tp->lost_retrans_low = new_low_seq; 1200} 1201 1202static int tcp_check_dsack(struct sock *sk, struct sk_buff *ack_skb, 1203 struct tcp_sack_block_wire *sp, int num_sacks, 1204 u32 prior_snd_una) 1205{ 1206 struct tcp_sock *tp = tcp_sk(sk); 1207 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq); 1208 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq); 1209 int dup_sack = 0; 1210 1211 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) { 1212 dup_sack = 1; 1213 tcp_dsack_seen(tp); 1214 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV); 1215 } else if (num_sacks > 1) { 1216 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq); 1217 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq); 1218 1219 if (!after(end_seq_0, end_seq_1) && 1220 !before(start_seq_0, start_seq_1)) { 1221 dup_sack = 1; 1222 tcp_dsack_seen(tp); 1223 NET_INC_STATS_BH(sock_net(sk), 1224 LINUX_MIB_TCPDSACKOFORECV); 1225 } 1226 } 1227 1228 /* D-SACK for already forgotten data... Do dumb counting. */ 1229 if (dup_sack && 1230 !after(end_seq_0, prior_snd_una) && 1231 after(end_seq_0, tp->undo_marker)) 1232 tp->undo_retrans--; 1233 1234 return dup_sack; 1235} 1236 1237struct tcp_sacktag_state { 1238 int reord; 1239 int fack_count; 1240 int flag; 1241}; 1242 1243/* Check if skb is fully within the SACK block. In presence of GSO skbs, 1244 * the incoming SACK may not exactly match but we can find smaller MSS 1245 * aligned portion of it that matches. Therefore we might need to fragment 1246 * which may fail and creates some hassle (caller must handle error case 1247 * returns). 1248 * 1249 * FIXME: this could be merged to shift decision code 1250 */ 1251static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb, 1252 u32 start_seq, u32 end_seq) 1253{ 1254 int in_sack, err; 1255 unsigned int pkt_len; 1256 unsigned int mss; 1257 1258 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && 1259 !before(end_seq, TCP_SKB_CB(skb)->end_seq); 1260 1261 if (tcp_skb_pcount(skb) > 1 && !in_sack && 1262 after(TCP_SKB_CB(skb)->end_seq, start_seq)) { 1263 mss = tcp_skb_mss(skb); 1264 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); 1265 1266 if (!in_sack) { 1267 pkt_len = start_seq - TCP_SKB_CB(skb)->seq; 1268 if (pkt_len < mss) 1269 pkt_len = mss; 1270 } else { 1271 pkt_len = end_seq - TCP_SKB_CB(skb)->seq; 1272 if (pkt_len < mss) 1273 return -EINVAL; 1274 } 1275 1276 /* Round if necessary so that SACKs cover only full MSSes 1277 * and/or the remaining small portion (if present) 1278 */ 1279 if (pkt_len > mss) { 1280 unsigned int new_len = (pkt_len / mss) * mss; 1281 if (!in_sack && new_len < pkt_len) { 1282 new_len += mss; 1283 if (new_len > skb->len) 1284 return 0; 1285 } 1286 pkt_len = new_len; 1287 } 1288 err = tcp_fragment(sk, skb, pkt_len, mss); 1289 if (err < 0) 1290 return err; 1291 } 1292 1293 return in_sack; 1294} 1295 1296static u8 tcp_sacktag_one(struct sk_buff *skb, struct sock *sk, 1297 struct tcp_sacktag_state *state, 1298 int dup_sack, int pcount) 1299{ 1300 struct tcp_sock *tp = tcp_sk(sk); 1301 u8 sacked = TCP_SKB_CB(skb)->sacked; 1302 int fack_count = state->fack_count; 1303 1304 /* Account D-SACK for retransmitted packet. */ 1305 if (dup_sack && (sacked & TCPCB_RETRANS)) { 1306 if (after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker)) 1307 tp->undo_retrans--; 1308 if (sacked & TCPCB_SACKED_ACKED) 1309 state->reord = min(fack_count, state->reord); 1310 } 1311 1312 /* Nothing to do; acked frame is about to be dropped (was ACKed). */ 1313 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 1314 return sacked; 1315 1316 if (!(sacked & TCPCB_SACKED_ACKED)) { 1317 if (sacked & TCPCB_SACKED_RETRANS) { 1318 /* If the segment is not tagged as lost, 1319 * we do not clear RETRANS, believing 1320 * that retransmission is still in flight. 1321 */ 1322 if (sacked & TCPCB_LOST) { 1323 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS); 1324 tp->lost_out -= pcount; 1325 tp->retrans_out -= pcount; 1326 } 1327 } else { 1328 if (!(sacked & TCPCB_RETRANS)) { 1329 /* New sack for not retransmitted frame, 1330 * which was in hole. It is reordering. 1331 */ 1332 if (before(TCP_SKB_CB(skb)->seq, 1333 tcp_highest_sack_seq(tp))) 1334 state->reord = min(fack_count, 1335 state->reord); 1336 1337 /* SACK enhanced F-RTO (RFC4138; Appendix B) */ 1338 if (!after(TCP_SKB_CB(skb)->end_seq, tp->frto_highmark)) 1339 state->flag |= FLAG_ONLY_ORIG_SACKED; 1340 } 1341 1342 if (sacked & TCPCB_LOST) { 1343 sacked &= ~TCPCB_LOST; 1344 tp->lost_out -= pcount; 1345 } 1346 } 1347 1348 sacked |= TCPCB_SACKED_ACKED; 1349 state->flag |= FLAG_DATA_SACKED; 1350 tp->sacked_out += pcount; 1351 1352 fack_count += pcount; 1353 1354 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */ 1355 if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) && 1356 before(TCP_SKB_CB(skb)->seq, 1357 TCP_SKB_CB(tp->lost_skb_hint)->seq)) 1358 tp->lost_cnt_hint += pcount; 1359 1360 if (fack_count > tp->fackets_out) 1361 tp->fackets_out = fack_count; 1362 } 1363 1364 /* D-SACK. We can detect redundant retransmission in S|R and plain R 1365 * frames and clear it. undo_retrans is decreased above, L|R frames 1366 * are accounted above as well. 1367 */ 1368 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) { 1369 sacked &= ~TCPCB_SACKED_RETRANS; 1370 tp->retrans_out -= pcount; 1371 } 1372 1373 return sacked; 1374} 1375 1376static int tcp_shifted_skb(struct sock *sk, struct sk_buff *skb, 1377 struct tcp_sacktag_state *state, 1378 unsigned int pcount, int shifted, int mss, 1379 int dup_sack) 1380{ 1381 struct tcp_sock *tp = tcp_sk(sk); 1382 struct sk_buff *prev = tcp_write_queue_prev(sk, skb); 1383 1384 BUG_ON(!pcount); 1385 1386 /* Tweak before seqno plays */ 1387 if (!tcp_is_fack(tp) && tcp_is_sack(tp) && tp->lost_skb_hint && 1388 !before(TCP_SKB_CB(tp->lost_skb_hint)->seq, TCP_SKB_CB(skb)->seq)) 1389 tp->lost_cnt_hint += pcount; 1390 1391 TCP_SKB_CB(prev)->end_seq += shifted; 1392 TCP_SKB_CB(skb)->seq += shifted; 1393 1394 skb_shinfo(prev)->gso_segs += pcount; 1395 BUG_ON(skb_shinfo(skb)->gso_segs < pcount); 1396 skb_shinfo(skb)->gso_segs -= pcount; 1397 1398 /* When we're adding to gso_segs == 1, gso_size will be zero, 1399 * in theory this shouldn't be necessary but as long as DSACK 1400 * code can come after this skb later on it's better to keep 1401 * setting gso_size to something. 1402 */ 1403 if (!skb_shinfo(prev)->gso_size) { 1404 skb_shinfo(prev)->gso_size = mss; 1405 skb_shinfo(prev)->gso_type = sk->sk_gso_type; 1406 } 1407 1408 /* CHECKME: To clear or not to clear? Mimics normal skb currently */ 1409 if (skb_shinfo(skb)->gso_segs <= 1) { 1410 skb_shinfo(skb)->gso_size = 0; 1411 skb_shinfo(skb)->gso_type = 0; 1412 } 1413 1414 /* We discard results */ 1415 tcp_sacktag_one(skb, sk, state, dup_sack, pcount); 1416 1417 /* Difference in this won't matter, both ACKed by the same cumul. ACK */ 1418 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS); 1419 1420 if (skb->len > 0) { 1421 BUG_ON(!tcp_skb_pcount(skb)); 1422 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED); 1423 return 0; 1424 } 1425 1426 /* Whole SKB was eaten :-) */ 1427 1428 if (skb == tp->retransmit_skb_hint) 1429 tp->retransmit_skb_hint = prev; 1430 if (skb == tp->scoreboard_skb_hint) 1431 tp->scoreboard_skb_hint = prev; 1432 if (skb == tp->lost_skb_hint) { 1433 tp->lost_skb_hint = prev; 1434 tp->lost_cnt_hint -= tcp_skb_pcount(prev); 1435 } 1436 1437 TCP_SKB_CB(skb)->flags |= TCP_SKB_CB(prev)->flags; 1438 if (skb == tcp_highest_sack(sk)) 1439 tcp_advance_highest_sack(sk, skb); 1440 1441 tcp_unlink_write_queue(skb, sk); 1442 sk_wmem_free_skb(sk, skb); 1443 1444 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED); 1445 1446 return 1; 1447} 1448 1449/* I wish gso_size would have a bit more sane initialization than 1450 * something-or-zero which complicates things 1451 */ 1452static int tcp_skb_seglen(struct sk_buff *skb) 1453{ 1454 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb); 1455} 1456 1457/* Shifting pages past head area doesn't work */ 1458static int skb_can_shift(struct sk_buff *skb) 1459{ 1460 return !skb_headlen(skb) && skb_is_nonlinear(skb); 1461} 1462 1463/* Try collapsing SACK blocks spanning across multiple skbs to a single 1464 * skb. 1465 */ 1466static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb, 1467 struct tcp_sacktag_state *state, 1468 u32 start_seq, u32 end_seq, 1469 int dup_sack) 1470{ 1471 struct tcp_sock *tp = tcp_sk(sk); 1472 struct sk_buff *prev; 1473 int mss; 1474 int pcount = 0; 1475 int len; 1476 int in_sack; 1477 1478 if (!sk_can_gso(sk)) 1479 goto fallback; 1480 1481 /* Normally R but no L won't result in plain S */ 1482 if (!dup_sack && 1483 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS) 1484 goto fallback; 1485 if (!skb_can_shift(skb)) 1486 goto fallback; 1487 /* This frame is about to be dropped (was ACKed). */ 1488 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 1489 goto fallback; 1490 1491 /* Can only happen with delayed DSACK + discard craziness */ 1492 if (unlikely(skb == tcp_write_queue_head(sk))) 1493 goto fallback; 1494 prev = tcp_write_queue_prev(sk, skb); 1495 1496 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) 1497 goto fallback; 1498 1499 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && 1500 !before(end_seq, TCP_SKB_CB(skb)->end_seq); 1501 1502 if (in_sack) { 1503 len = skb->len; 1504 pcount = tcp_skb_pcount(skb); 1505 mss = tcp_skb_seglen(skb); 1506 1507 /* TODO: Fix DSACKs to not fragment already SACKed and we can 1508 * drop this restriction as unnecessary 1509 */ 1510 if (mss != tcp_skb_seglen(prev)) 1511 goto fallback; 1512 } else { 1513 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq)) 1514 goto noop; 1515 /* CHECKME: This is non-MSS split case only?, this will 1516 * cause skipped skbs due to advancing loop btw, original 1517 * has that feature too 1518 */ 1519 if (tcp_skb_pcount(skb) <= 1) 1520 goto noop; 1521 1522 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); 1523 if (!in_sack) { 1524 /* TODO: head merge to next could be attempted here 1525 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)), 1526 * though it might not be worth of the additional hassle 1527 * 1528 * ...we can probably just fallback to what was done 1529 * previously. We could try merging non-SACKed ones 1530 * as well but it probably isn't going to buy off 1531 * because later SACKs might again split them, and 1532 * it would make skb timestamp tracking considerably 1533 * harder problem. 1534 */ 1535 goto fallback; 1536 } 1537 1538 len = end_seq - TCP_SKB_CB(skb)->seq; 1539 BUG_ON(len < 0); 1540 BUG_ON(len > skb->len); 1541 1542 /* MSS boundaries should be honoured or else pcount will 1543 * severely break even though it makes things bit trickier. 1544 * Optimize common case to avoid most of the divides 1545 */ 1546 mss = tcp_skb_mss(skb); 1547 1548 /* TODO: Fix DSACKs to not fragment already SACKed and we can 1549 * drop this restriction as unnecessary 1550 */ 1551 if (mss != tcp_skb_seglen(prev)) 1552 goto fallback; 1553 1554 if (len == mss) { 1555 pcount = 1; 1556 } else if (len < mss) { 1557 goto noop; 1558 } else { 1559 pcount = len / mss; 1560 len = pcount * mss; 1561 } 1562 } 1563 1564 if (!skb_shift(prev, skb, len)) 1565 goto fallback; 1566 if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack)) 1567 goto out; 1568 1569 /* Hole filled allows collapsing with the next as well, this is very 1570 * useful when hole on every nth skb pattern happens 1571 */ 1572 if (prev == tcp_write_queue_tail(sk)) 1573 goto out; 1574 skb = tcp_write_queue_next(sk, prev); 1575 1576 if (!skb_can_shift(skb) || 1577 (skb == tcp_send_head(sk)) || 1578 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) || 1579 (mss != tcp_skb_seglen(skb))) 1580 goto out; 1581 1582 len = skb->len; 1583 if (skb_shift(prev, skb, len)) { 1584 pcount += tcp_skb_pcount(skb); 1585 tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0); 1586 } 1587 1588out: 1589 state->fack_count += pcount; 1590 return prev; 1591 1592noop: 1593 return skb; 1594 1595fallback: 1596 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK); 1597 return NULL; 1598} 1599 1600static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk, 1601 struct tcp_sack_block *next_dup, 1602 struct tcp_sacktag_state *state, 1603 u32 start_seq, u32 end_seq, 1604 int dup_sack_in) 1605{ 1606 struct tcp_sock *tp = tcp_sk(sk); 1607 struct sk_buff *tmp; 1608 1609 tcp_for_write_queue_from(skb, sk) { 1610 int in_sack = 0; 1611 int dup_sack = dup_sack_in; 1612 1613 if (skb == tcp_send_head(sk)) 1614 break; 1615 1616 /* queue is in-order => we can short-circuit the walk early */ 1617 if (!before(TCP_SKB_CB(skb)->seq, end_seq)) 1618 break; 1619 1620 if ((next_dup != NULL) && 1621 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) { 1622 in_sack = tcp_match_skb_to_sack(sk, skb, 1623 next_dup->start_seq, 1624 next_dup->end_seq); 1625 if (in_sack > 0) 1626 dup_sack = 1; 1627 } 1628 1629 /* skb reference here is a bit tricky to get right, since 1630 * shifting can eat and free both this skb and the next, 1631 * so not even _safe variant of the loop is enough. 1632 */ 1633 if (in_sack <= 0) { 1634 tmp = tcp_shift_skb_data(sk, skb, state, 1635 start_seq, end_seq, dup_sack); 1636 if (tmp != NULL) { 1637 if (tmp != skb) { 1638 skb = tmp; 1639 continue; 1640 } 1641 1642 in_sack = 0; 1643 } else { 1644 in_sack = tcp_match_skb_to_sack(sk, skb, 1645 start_seq, 1646 end_seq); 1647 } 1648 } 1649 1650 if (unlikely(in_sack < 0)) 1651 break; 1652 1653 if (in_sack) { 1654 TCP_SKB_CB(skb)->sacked = tcp_sacktag_one(skb, sk, 1655 state, 1656 dup_sack, 1657 tcp_skb_pcount(skb)); 1658 1659 if (!before(TCP_SKB_CB(skb)->seq, 1660 tcp_highest_sack_seq(tp))) 1661 tcp_advance_highest_sack(sk, skb); 1662 } 1663 1664 state->fack_count += tcp_skb_pcount(skb); 1665 } 1666 return skb; 1667} 1668 1669/* Avoid all extra work that is being done by sacktag while walking in 1670 * a normal way 1671 */ 1672static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk, 1673 struct tcp_sacktag_state *state, 1674 u32 skip_to_seq) 1675{ 1676 tcp_for_write_queue_from(skb, sk) { 1677 if (skb == tcp_send_head(sk)) 1678 break; 1679 1680 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq)) 1681 break; 1682 1683 state->fack_count += tcp_skb_pcount(skb); 1684 } 1685 return skb; 1686} 1687 1688static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb, 1689 struct sock *sk, 1690 struct tcp_sack_block *next_dup, 1691 struct tcp_sacktag_state *state, 1692 u32 skip_to_seq) 1693{ 1694 if (next_dup == NULL) 1695 return skb; 1696 1697 if (before(next_dup->start_seq, skip_to_seq)) { 1698 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq); 1699 skb = tcp_sacktag_walk(skb, sk, NULL, state, 1700 next_dup->start_seq, next_dup->end_seq, 1701 1); 1702 } 1703 1704 return skb; 1705} 1706 1707static int tcp_sack_cache_ok(struct tcp_sock *tp, struct tcp_sack_block *cache) 1708{ 1709 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache); 1710} 1711 1712static int 1713tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, 1714 u32 prior_snd_una) 1715{ 1716 const struct inet_connection_sock *icsk = inet_csk(sk); 1717 struct tcp_sock *tp = tcp_sk(sk); 1718 unsigned char *ptr = (skb_transport_header(ack_skb) + 1719 TCP_SKB_CB(ack_skb)->sacked); 1720 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2); 1721 struct tcp_sack_block sp[TCP_NUM_SACKS]; 1722 struct tcp_sack_block *cache; 1723 struct tcp_sacktag_state state; 1724 struct sk_buff *skb; 1725 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3); 1726 int used_sacks; 1727 int found_dup_sack = 0; 1728 int i, j; 1729 int first_sack_index; 1730 1731 state.flag = 0; 1732 state.reord = tp->packets_out; 1733 1734 if (!tp->sacked_out) { 1735 if (WARN_ON(tp->fackets_out)) 1736 tp->fackets_out = 0; 1737 tcp_highest_sack_reset(sk); 1738 } 1739 1740 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire, 1741 num_sacks, prior_snd_una); 1742 if (found_dup_sack) 1743 state.flag |= FLAG_DSACKING_ACK; 1744 1745 /* Eliminate too old ACKs, but take into 1746 * account more or less fresh ones, they can 1747 * contain valid SACK info. 1748 */ 1749 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window)) 1750 return 0; 1751 1752 if (!tp->packets_out) 1753 goto out; 1754 1755 used_sacks = 0; 1756 first_sack_index = 0; 1757 for (i = 0; i < num_sacks; i++) { 1758 int dup_sack = !i && found_dup_sack; 1759 1760 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq); 1761 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq); 1762 1763 if (!tcp_is_sackblock_valid(tp, dup_sack, 1764 sp[used_sacks].start_seq, 1765 sp[used_sacks].end_seq)) { 1766 int mib_idx; 1767 1768 if (dup_sack) { 1769 if (!tp->undo_marker) 1770 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO; 1771 else 1772 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD; 1773 } else { 1774 /* Don't count olds caused by ACK reordering */ 1775 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) && 1776 !after(sp[used_sacks].end_seq, tp->snd_una)) 1777 continue; 1778 mib_idx = LINUX_MIB_TCPSACKDISCARD; 1779 } 1780 1781 NET_INC_STATS_BH(sock_net(sk), mib_idx); 1782 if (i == 0) 1783 first_sack_index = -1; 1784 continue; 1785 } 1786 1787 /* Ignore very old stuff early */ 1788 if (!after(sp[used_sacks].end_seq, prior_snd_una)) 1789 continue; 1790 1791 used_sacks++; 1792 } 1793 1794 /* order SACK blocks to allow in order walk of the retrans queue */ 1795 for (i = used_sacks - 1; i > 0; i--) { 1796 for (j = 0; j < i; j++) { 1797 if (after(sp[j].start_seq, sp[j + 1].start_seq)) { 1798 swap(sp[j], sp[j + 1]); 1799 1800 /* Track where the first SACK block goes to */ 1801 if (j == first_sack_index) 1802 first_sack_index = j + 1; 1803 } 1804 } 1805 } 1806 1807 skb = tcp_write_queue_head(sk); 1808 state.fack_count = 0; 1809 i = 0; 1810 1811 if (!tp->sacked_out) { 1812 /* It's already past, so skip checking against it */ 1813 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache); 1814 } else { 1815 cache = tp->recv_sack_cache; 1816 /* Skip empty blocks in at head of the cache */ 1817 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq && 1818 !cache->end_seq) 1819 cache++; 1820 } 1821 1822 while (i < used_sacks) { 1823 u32 start_seq = sp[i].start_seq; 1824 u32 end_seq = sp[i].end_seq; 1825 int dup_sack = (found_dup_sack && (i == first_sack_index)); 1826 struct tcp_sack_block *next_dup = NULL; 1827 1828 if (found_dup_sack && ((i + 1) == first_sack_index)) 1829 next_dup = &sp[i + 1]; 1830 1831 /* Event "B" in the comment above. */ 1832 if (after(end_seq, tp->high_seq)) 1833 state.flag |= FLAG_DATA_LOST; 1834 1835 /* Skip too early cached blocks */ 1836 while (tcp_sack_cache_ok(tp, cache) && 1837 !before(start_seq, cache->end_seq)) 1838 cache++; 1839 1840 /* Can skip some work by looking recv_sack_cache? */ 1841 if (tcp_sack_cache_ok(tp, cache) && !dup_sack && 1842 after(end_seq, cache->start_seq)) { 1843 1844 /* Head todo? */ 1845 if (before(start_seq, cache->start_seq)) { 1846 skb = tcp_sacktag_skip(skb, sk, &state, 1847 start_seq); 1848 skb = tcp_sacktag_walk(skb, sk, next_dup, 1849 &state, 1850 start_seq, 1851 cache->start_seq, 1852 dup_sack); 1853 } 1854 1855 /* Rest of the block already fully processed? */ 1856 if (!after(end_seq, cache->end_seq)) 1857 goto advance_sp; 1858 1859 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup, 1860 &state, 1861 cache->end_seq); 1862 1863 /* ...tail remains todo... */ 1864 if (tcp_highest_sack_seq(tp) == cache->end_seq) { 1865 /* ...but better entrypoint exists! */ 1866 skb = tcp_highest_sack(sk); 1867 if (skb == NULL) 1868 break; 1869 state.fack_count = tp->fackets_out; 1870 cache++; 1871 goto walk; 1872 } 1873 1874 skb = tcp_sacktag_skip(skb, sk, &state, cache->end_seq); 1875 /* Check overlap against next cached too (past this one already) */ 1876 cache++; 1877 continue; 1878 } 1879 1880 if (!before(start_seq, tcp_highest_sack_seq(tp))) { 1881 skb = tcp_highest_sack(sk); 1882 if (skb == NULL) 1883 break; 1884 state.fack_count = tp->fackets_out; 1885 } 1886 skb = tcp_sacktag_skip(skb, sk, &state, start_seq); 1887 1888walk: 1889 skb = tcp_sacktag_walk(skb, sk, next_dup, &state, 1890 start_seq, end_seq, dup_sack); 1891 1892advance_sp: 1893 /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct 1894 * due to in-order walk 1895 */ 1896 if (after(end_seq, tp->frto_highmark)) 1897 state.flag &= ~FLAG_ONLY_ORIG_SACKED; 1898 1899 i++; 1900 } 1901 1902 /* Clear the head of the cache sack blocks so we can skip it next time */ 1903 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) { 1904 tp->recv_sack_cache[i].start_seq = 0; 1905 tp->recv_sack_cache[i].end_seq = 0; 1906 } 1907 for (j = 0; j < used_sacks; j++) 1908 tp->recv_sack_cache[i++] = sp[j]; 1909 1910 tcp_mark_lost_retrans(sk); 1911 1912 tcp_verify_left_out(tp); 1913 1914 if ((state.reord < tp->fackets_out) && 1915 ((icsk->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker) && 1916 (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark))) 1917 tcp_update_reordering(sk, tp->fackets_out - state.reord, 0); 1918 1919out: 1920 1921#if FASTRETRANS_DEBUG > 0 1922 WARN_ON((int)tp->sacked_out < 0); 1923 WARN_ON((int)tp->lost_out < 0); 1924 WARN_ON((int)tp->retrans_out < 0); 1925 WARN_ON((int)tcp_packets_in_flight(tp) < 0); 1926#endif 1927 return state.flag; 1928} 1929 1930/* Limits sacked_out so that sum with lost_out isn't ever larger than 1931 * packets_out. Returns zero if sacked_out adjustement wasn't necessary. 1932 */ 1933static int tcp_limit_reno_sacked(struct tcp_sock *tp) 1934{ 1935 u32 holes; 1936 1937 holes = max(tp->lost_out, 1U); 1938 holes = min(holes, tp->packets_out); 1939 1940 if ((tp->sacked_out + holes) > tp->packets_out) { 1941 tp->sacked_out = tp->packets_out - holes; 1942 return 1; 1943 } 1944 return 0; 1945} 1946 1947/* If we receive more dupacks than we expected counting segments 1948 * in assumption of absent reordering, interpret this as reordering. 1949 * The only another reason could be bug in receiver TCP. 1950 */ 1951static void tcp_check_reno_reordering(struct sock *sk, const int addend) 1952{ 1953 struct tcp_sock *tp = tcp_sk(sk); 1954 if (tcp_limit_reno_sacked(tp)) 1955 tcp_update_reordering(sk, tp->packets_out + addend, 0); 1956} 1957 1958/* Emulate SACKs for SACKless connection: account for a new dupack. */ 1959 1960static void tcp_add_reno_sack(struct sock *sk) 1961{ 1962 struct tcp_sock *tp = tcp_sk(sk); 1963 tp->sacked_out++; 1964 tcp_check_reno_reordering(sk, 0); 1965 tcp_verify_left_out(tp); 1966} 1967 1968/* Account for ACK, ACKing some data in Reno Recovery phase. */ 1969 1970static void tcp_remove_reno_sacks(struct sock *sk, int acked) 1971{ 1972 struct tcp_sock *tp = tcp_sk(sk); 1973 1974 if (acked > 0) { 1975 /* One ACK acked hole. The rest eat duplicate ACKs. */ 1976 if (acked - 1 >= tp->sacked_out) 1977 tp->sacked_out = 0; 1978 else 1979 tp->sacked_out -= acked - 1; 1980 } 1981 tcp_check_reno_reordering(sk, acked); 1982 tcp_verify_left_out(tp); 1983} 1984 1985static inline void tcp_reset_reno_sack(struct tcp_sock *tp) 1986{ 1987 tp->sacked_out = 0; 1988} 1989 1990static int tcp_is_sackfrto(const struct tcp_sock *tp) 1991{ 1992 return (sysctl_tcp_frto == 0x2) && !tcp_is_reno(tp); 1993} 1994 1995/* F-RTO can only be used if TCP has never retransmitted anything other than 1996 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here) 1997 */ 1998int tcp_use_frto(struct sock *sk) 1999{ 2000 const struct tcp_sock *tp = tcp_sk(sk); 2001 const struct inet_connection_sock *icsk = inet_csk(sk); 2002 struct sk_buff *skb; 2003 2004 if (!sysctl_tcp_frto) 2005 return 0; 2006 2007 /* MTU probe and F-RTO won't really play nicely along currently */ 2008 if (icsk->icsk_mtup.probe_size) 2009 return 0; 2010 2011 if (tcp_is_sackfrto(tp)) 2012 return 1; 2013 2014 /* Avoid expensive walking of rexmit queue if possible */ 2015 if (tp->retrans_out > 1) 2016 return 0; 2017 2018 skb = tcp_write_queue_head(sk); 2019 if (tcp_skb_is_last(sk, skb)) 2020 return 1; 2021 skb = tcp_write_queue_next(sk, skb); /* Skips head */ 2022 tcp_for_write_queue_from(skb, sk) { 2023 if (skb == tcp_send_head(sk)) 2024 break; 2025 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) 2026 return 0; 2027 /* Short-circuit when first non-SACKed skb has been checked */ 2028 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) 2029 break; 2030 } 2031 return 1; 2032} 2033 2034/* RTO occurred, but do not yet enter Loss state. Instead, defer RTO 2035 * recovery a bit and use heuristics in tcp_process_frto() to detect if 2036 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to 2037 * keep retrans_out counting accurate (with SACK F-RTO, other than head 2038 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS 2039 * bits are handled if the Loss state is really to be entered (in 2040 * tcp_enter_frto_loss). 2041 * 2042 * Do like tcp_enter_loss() would; when RTO expires the second time it 2043 * does: 2044 * "Reduce ssthresh if it has not yet been made inside this window." 2045 */ 2046void tcp_enter_frto(struct sock *sk) 2047{ 2048 const struct inet_connection_sock *icsk = inet_csk(sk); 2049 struct tcp_sock *tp = tcp_sk(sk); 2050 struct sk_buff *skb; 2051 2052 if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) || 2053 tp->snd_una == tp->high_seq || 2054 ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) && 2055 !icsk->icsk_retransmits)) { 2056 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2057 /* Our state is too optimistic in ssthresh() call because cwnd 2058 * is not reduced until tcp_enter_frto_loss() when previous F-RTO 2059 * recovery has not yet completed. Pattern would be this: RTO, 2060 * Cumulative ACK, RTO (2xRTO for the same segment does not end 2061 * up here twice). 2062 * RFC4138 should be more specific on what to do, even though 2063 * RTO is quite unlikely to occur after the first Cumulative ACK 2064 * due to back-off and complexity of triggering events ... 2065 */ 2066 if (tp->frto_counter) { 2067 u32 stored_cwnd; 2068 stored_cwnd = tp->snd_cwnd; 2069 tp->snd_cwnd = 2; 2070 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 2071 tp->snd_cwnd = stored_cwnd; 2072 } else { 2073 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 2074 } 2075 /* ... in theory, cong.control module could do "any tricks" in 2076 * ssthresh(), which means that ca_state, lost bits and lost_out 2077 * counter would have to be faked before the call occurs. We 2078 * consider that too expensive, unlikely and hacky, so modules 2079 * using these in ssthresh() must deal these incompatibility 2080 * issues if they receives CA_EVENT_FRTO and frto_counter != 0 2081 */ 2082 tcp_ca_event(sk, CA_EVENT_FRTO); 2083 } 2084 2085 tp->undo_marker = tp->snd_una; 2086 tp->undo_retrans = 0; 2087 2088 skb = tcp_write_queue_head(sk); 2089 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) 2090 tp->undo_marker = 0; 2091 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 2092 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 2093 tp->retrans_out -= tcp_skb_pcount(skb); 2094 } 2095 tcp_verify_left_out(tp); 2096 2097 /* Too bad if TCP was application limited */ 2098 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1); 2099 2100 /* Earlier loss recovery underway (see RFC4138; Appendix B). 2101 * The last condition is necessary at least in tp->frto_counter case. 2102 */ 2103 if (tcp_is_sackfrto(tp) && (tp->frto_counter || 2104 ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) && 2105 after(tp->high_seq, tp->snd_una)) { 2106 tp->frto_highmark = tp->high_seq; 2107 } else { 2108 tp->frto_highmark = tp->snd_nxt; 2109 } 2110 tcp_set_ca_state(sk, TCP_CA_Disorder); 2111 tp->high_seq = tp->snd_nxt; 2112 tp->frto_counter = 1; 2113} 2114 2115/* Enter Loss state after F-RTO was applied. Dupack arrived after RTO, 2116 * which indicates that we should follow the traditional RTO recovery, 2117 * i.e. mark everything lost and do go-back-N retransmission. 2118 */ 2119static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag) 2120{ 2121 struct tcp_sock *tp = tcp_sk(sk); 2122 struct sk_buff *skb; 2123 2124 tp->lost_out = 0; 2125 tp->retrans_out = 0; 2126 if (tcp_is_reno(tp)) 2127 tcp_reset_reno_sack(tp); 2128 2129 tcp_for_write_queue(skb, sk) { 2130 if (skb == tcp_send_head(sk)) 2131 break; 2132 2133 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; 2134 /* 2135 * Count the retransmission made on RTO correctly (only when 2136 * waiting for the first ACK and did not get it)... 2137 */ 2138 if ((tp->frto_counter == 1) && !(flag & FLAG_DATA_ACKED)) { 2139 /* For some reason this R-bit might get cleared? */ 2140 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 2141 tp->retrans_out += tcp_skb_pcount(skb); 2142 /* ...enter this if branch just for the first segment */ 2143 flag |= FLAG_DATA_ACKED; 2144 } else { 2145 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) 2146 tp->undo_marker = 0; 2147 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 2148 } 2149 2150 /* Marking forward transmissions that were made after RTO lost 2151 * can cause unnecessary retransmissions in some scenarios, 2152 * SACK blocks will mitigate that in some but not in all cases. 2153 * We used to not mark them but it was causing break-ups with 2154 * receivers that do only in-order receival. 2155 * 2156 * TODO: we could detect presence of such receiver and select 2157 * different behavior per flow. 2158 */ 2159 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) { 2160 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 2161 tp->lost_out += tcp_skb_pcount(skb); 2162 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq; 2163 } 2164 } 2165 tcp_verify_left_out(tp); 2166 2167 tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments; 2168 tp->snd_cwnd_cnt = 0; 2169 tp->snd_cwnd_stamp = tcp_time_stamp; 2170 tp->frto_counter = 0; 2171 tp->bytes_acked = 0; 2172 2173 tp->reordering = min_t(unsigned int, tp->reordering, 2174 sysctl_tcp_reordering); 2175 tcp_set_ca_state(sk, TCP_CA_Loss); 2176 tp->high_seq = tp->snd_nxt; 2177 TCP_ECN_queue_cwr(tp); 2178 2179 tcp_clear_all_retrans_hints(tp); 2180} 2181 2182static void tcp_clear_retrans_partial(struct tcp_sock *tp) 2183{ 2184 tp->retrans_out = 0; 2185 tp->lost_out = 0; 2186 2187 tp->undo_marker = 0; 2188 tp->undo_retrans = 0; 2189} 2190 2191void tcp_clear_retrans(struct tcp_sock *tp) 2192{ 2193 tcp_clear_retrans_partial(tp); 2194 2195 tp->fackets_out = 0; 2196 tp->sacked_out = 0; 2197} 2198 2199/* Enter Loss state. If "how" is not zero, forget all SACK information 2200 * and reset tags completely, otherwise preserve SACKs. If receiver 2201 * dropped its ofo queue, we will know this due to reneging detection. 2202 */ 2203void tcp_enter_loss(struct sock *sk, int how) 2204{ 2205 const struct inet_connection_sock *icsk = inet_csk(sk); 2206 struct tcp_sock *tp = tcp_sk(sk); 2207 struct sk_buff *skb; 2208 2209 /* Reduce ssthresh if it has not yet been made inside this window. */ 2210 if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq || 2211 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) { 2212 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2213 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 2214 tcp_ca_event(sk, CA_EVENT_LOSS); 2215 } 2216 tp->snd_cwnd = 1; 2217 tp->snd_cwnd_cnt = 0; 2218 tp->snd_cwnd_stamp = tcp_time_stamp; 2219 2220 tp->bytes_acked = 0; 2221 tcp_clear_retrans_partial(tp); 2222 2223 if (tcp_is_reno(tp)) 2224 tcp_reset_reno_sack(tp); 2225 2226 if (!how) { 2227 /* Push undo marker, if it was plain RTO and nothing 2228 * was retransmitted. */ 2229 tp->undo_marker = tp->snd_una; 2230 } else { 2231 tp->sacked_out = 0; 2232 tp->fackets_out = 0; 2233 } 2234 tcp_clear_all_retrans_hints(tp); 2235 2236 tcp_for_write_queue(skb, sk) { 2237 if (skb == tcp_send_head(sk)) 2238 break; 2239 2240 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) 2241 tp->undo_marker = 0; 2242 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED; 2243 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) { 2244 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED; 2245 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 2246 tp->lost_out += tcp_skb_pcount(skb); 2247 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq; 2248 } 2249 } 2250 tcp_verify_left_out(tp); 2251 2252 tp->reordering = min_t(unsigned int, tp->reordering, 2253 sysctl_tcp_reordering); 2254 tcp_set_ca_state(sk, TCP_CA_Loss); 2255 tp->high_seq = tp->snd_nxt; 2256 TCP_ECN_queue_cwr(tp); 2257 /* Abort F-RTO algorithm if one is in progress */ 2258 tp->frto_counter = 0; 2259} 2260 2261/* If ACK arrived pointing to a remembered SACK, it means that our 2262 * remembered SACKs do not reflect real state of receiver i.e. 2263 * receiver _host_ is heavily congested (or buggy). 2264 * 2265 * Do processing similar to RTO timeout. 2266 */ 2267static int tcp_check_sack_reneging(struct sock *sk, int flag) 2268{ 2269 if (flag & FLAG_SACK_RENEGING) { 2270 struct inet_connection_sock *icsk = inet_csk(sk); 2271 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING); 2272 2273 tcp_enter_loss(sk, 1); 2274 icsk->icsk_retransmits++; 2275 tcp_retransmit_skb(sk, tcp_write_queue_head(sk)); 2276 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 2277 icsk->icsk_rto, TCP_RTO_MAX); 2278 return 1; 2279 } 2280 return 0; 2281} 2282 2283static inline int tcp_fackets_out(struct tcp_sock *tp) 2284{ 2285 return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out; 2286} 2287 2288/* Heurestics to calculate number of duplicate ACKs. There's no dupACKs 2289 * counter when SACK is enabled (without SACK, sacked_out is used for 2290 * that purpose). 2291 * 2292 * Instead, with FACK TCP uses fackets_out that includes both SACKed 2293 * segments up to the highest received SACK block so far and holes in 2294 * between them. 2295 * 2296 * With reordering, holes may still be in flight, so RFC3517 recovery 2297 * uses pure sacked_out (total number of SACKed segments) even though 2298 * it violates the RFC that uses duplicate ACKs, often these are equal 2299 * but when e.g. out-of-window ACKs or packet duplication occurs, 2300 * they differ. Since neither occurs due to loss, TCP should really 2301 * ignore them. 2302 */ 2303static inline int tcp_dupack_heuristics(struct tcp_sock *tp) 2304{ 2305 return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1; 2306} 2307 2308static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb) 2309{ 2310 return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto); 2311} 2312 2313static inline int tcp_head_timedout(struct sock *sk) 2314{ 2315 struct tcp_sock *tp = tcp_sk(sk); 2316 2317 return tp->packets_out && 2318 tcp_skb_timedout(sk, tcp_write_queue_head(sk)); 2319} 2320 2321/* Linux NewReno/SACK/FACK/ECN state machine. 2322 * -------------------------------------- 2323 * 2324 * "Open" Normal state, no dubious events, fast path. 2325 * "Disorder" In all the respects it is "Open", 2326 * but requires a bit more attention. It is entered when 2327 * we see some SACKs or dupacks. It is split of "Open" 2328 * mainly to move some processing from fast path to slow one. 2329 * "CWR" CWND was reduced due to some Congestion Notification event. 2330 * It can be ECN, ICMP source quench, local device congestion. 2331 * "Recovery" CWND was reduced, we are fast-retransmitting. 2332 * "Loss" CWND was reduced due to RTO timeout or SACK reneging. 2333 * 2334 * tcp_fastretrans_alert() is entered: 2335 * - each incoming ACK, if state is not "Open" 2336 * - when arrived ACK is unusual, namely: 2337 * * SACK 2338 * * Duplicate ACK. 2339 * * ECN ECE. 2340 * 2341 * Counting packets in flight is pretty simple. 2342 * 2343 * in_flight = packets_out - left_out + retrans_out 2344 * 2345 * packets_out is SND.NXT-SND.UNA counted in packets. 2346 * 2347 * retrans_out is number of retransmitted segments. 2348 * 2349 * left_out is number of segments left network, but not ACKed yet. 2350 * 2351 * left_out = sacked_out + lost_out 2352 * 2353 * sacked_out: Packets, which arrived to receiver out of order 2354 * and hence not ACKed. With SACKs this number is simply 2355 * amount of SACKed data. Even without SACKs 2356 * it is easy to give pretty reliable estimate of this number, 2357 * counting duplicate ACKs. 2358 * 2359 * lost_out: Packets lost by network. TCP has no explicit 2360 * "loss notification" feedback from network (for now). 2361 * It means that this number can be only _guessed_. 2362 * Actually, it is the heuristics to predict lossage that 2363 * distinguishes different algorithms. 2364 * 2365 * F.e. after RTO, when all the queue is considered as lost, 2366 * lost_out = packets_out and in_flight = retrans_out. 2367 * 2368 * Essentially, we have now two algorithms counting 2369 * lost packets. 2370 * 2371 * FACK: It is the simplest heuristics. As soon as we decided 2372 * that something is lost, we decide that _all_ not SACKed 2373 * packets until the most forward SACK are lost. I.e. 2374 * lost_out = fackets_out - sacked_out and left_out = fackets_out. 2375 * It is absolutely correct estimate, if network does not reorder 2376 * packets. And it loses any connection to reality when reordering 2377 * takes place. We use FACK by default until reordering 2378 * is suspected on the path to this destination. 2379 * 2380 * NewReno: when Recovery is entered, we assume that one segment 2381 * is lost (classic Reno). While we are in Recovery and 2382 * a partial ACK arrives, we assume that one more packet 2383 * is lost (NewReno). This heuristics are the same in NewReno 2384 * and SACK. 2385 * 2386 * Imagine, that's all! Forget about all this shamanism about CWND inflation 2387 * deflation etc. CWND is real congestion window, never inflated, changes 2388 * only according to classic VJ rules. 2389 * 2390 * Really tricky (and requiring careful tuning) part of algorithm 2391 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue(). 2392 * The first determines the moment _when_ we should reduce CWND and, 2393 * hence, slow down forward transmission. In fact, it determines the moment 2394 * when we decide that hole is caused by loss, rather than by a reorder. 2395 * 2396 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill 2397 * holes, caused by lost packets. 2398 * 2399 * And the most logically complicated part of algorithm is undo 2400 * heuristics. We detect false retransmits due to both too early 2401 * fast retransmit (reordering) and underestimated RTO, analyzing 2402 * timestamps and D-SACKs. When we detect that some segments were 2403 * retransmitted by mistake and CWND reduction was wrong, we undo 2404 * window reduction and abort recovery phase. This logic is hidden 2405 * inside several functions named tcp_try_undo_<something>. 2406 */ 2407 2408/* This function decides, when we should leave Disordered state 2409 * and enter Recovery phase, reducing congestion window. 2410 * 2411 * Main question: may we further continue forward transmission 2412 * with the same cwnd? 2413 */ 2414static int tcp_time_to_recover(struct sock *sk) 2415{ 2416 struct tcp_sock *tp = tcp_sk(sk); 2417 __u32 packets_out; 2418 2419 /* Do not perform any recovery during F-RTO algorithm */ 2420 if (tp->frto_counter) 2421 return 0; 2422 2423 /* Trick#1: The loss is proven. */ 2424 if (tp->lost_out) 2425 return 1; 2426 2427 /* Not-A-Trick#2 : Classic rule... */ 2428 if (tcp_dupack_heuristics(tp) > tp->reordering) 2429 return 1; 2430 2431 /* Trick#3 : when we use RFC2988 timer restart, fast 2432 * retransmit can be triggered by timeout of queue head. 2433 */ 2434 if (tcp_is_fack(tp) && tcp_head_timedout(sk)) 2435 return 1; 2436 2437 /* Trick#4: It is still not OK... But will it be useful to delay 2438 * recovery more? 2439 */ 2440 packets_out = tp->packets_out; 2441 if (packets_out <= tp->reordering && 2442 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) && 2443 !tcp_may_send_now(sk)) { 2444 /* We have nothing to send. This connection is limited 2445 * either by receiver window or by application. 2446 */ 2447 return 1; 2448 } 2449 2450 return 0; 2451} 2452 2453/* New heuristics: it is possible only after we switched to restart timer 2454 * each time when something is ACKed. Hence, we can detect timed out packets 2455 * during fast retransmit without falling to slow start. 2456 * 2457 * Usefulness of this as is very questionable, since we should know which of 2458 * the segments is the next to timeout which is relatively expensive to find 2459 * in general case unless we add some data structure just for that. The 2460 * current approach certainly won't find the right one too often and when it 2461 * finally does find _something_ it usually marks large part of the window 2462 * right away (because a retransmission with a larger timestamp blocks the 2463 * loop from advancing). -ij 2464 */ 2465static void tcp_timeout_skbs(struct sock *sk) 2466{ 2467 struct tcp_sock *tp = tcp_sk(sk); 2468 struct sk_buff *skb; 2469 2470 if (!tcp_is_fack(tp) || !tcp_head_timedout(sk)) 2471 return; 2472 2473 skb = tp->scoreboard_skb_hint; 2474 if (tp->scoreboard_skb_hint == NULL) 2475 skb = tcp_write_queue_head(sk); 2476 2477 tcp_for_write_queue_from(skb, sk) { 2478 if (skb == tcp_send_head(sk)) 2479 break; 2480 if (!tcp_skb_timedout(sk, skb)) 2481 break; 2482 2483 tcp_skb_mark_lost(tp, skb); 2484 } 2485 2486 tp->scoreboard_skb_hint = skb; 2487 2488 tcp_verify_left_out(tp); 2489} 2490 2491/* Mark head of queue up as lost. With RFC3517 SACK, the packets is 2492 * is against sacked "cnt", otherwise it's against facked "cnt" 2493 */ 2494static void tcp_mark_head_lost(struct sock *sk, int packets) 2495{ 2496 struct tcp_sock *tp = tcp_sk(sk); 2497 struct sk_buff *skb; 2498 int cnt, oldcnt; 2499 int err; 2500 unsigned int mss; 2501 2502 WARN_ON(packets > tp->packets_out); 2503 if (tp->lost_skb_hint) { 2504 skb = tp->lost_skb_hint; 2505 cnt = tp->lost_cnt_hint; 2506 } else { 2507 skb = tcp_write_queue_head(sk); 2508 cnt = 0; 2509 } 2510 2511 tcp_for_write_queue_from(skb, sk) { 2512 if (skb == tcp_send_head(sk)) 2513 break; 2514 /* TODO: do this better */ 2515 /* this is not the most efficient way to do this... */ 2516 tp->lost_skb_hint = skb; 2517 tp->lost_cnt_hint = cnt; 2518 2519 if (after(TCP_SKB_CB(skb)->end_seq, tp->high_seq)) 2520 break; 2521 2522 oldcnt = cnt; 2523 if (tcp_is_fack(tp) || tcp_is_reno(tp) || 2524 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) 2525 cnt += tcp_skb_pcount(skb); 2526 2527 if (cnt > packets) { 2528 if (tcp_is_sack(tp) || (oldcnt >= packets)) 2529 break; 2530 2531 mss = skb_shinfo(skb)->gso_size; 2532 err = tcp_fragment(sk, skb, (packets - oldcnt) * mss, mss); 2533 if (err < 0) 2534 break; 2535 cnt = packets; 2536 } 2537 2538 tcp_skb_mark_lost(tp, skb); 2539 } 2540 tcp_verify_left_out(tp); 2541} 2542 2543/* Account newly detected lost packet(s) */ 2544 2545static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit) 2546{ 2547 struct tcp_sock *tp = tcp_sk(sk); 2548 2549 if (tcp_is_reno(tp)) { 2550 tcp_mark_head_lost(sk, 1); 2551 } else if (tcp_is_fack(tp)) { 2552 int lost = tp->fackets_out - tp->reordering; 2553 if (lost <= 0) 2554 lost = 1; 2555 tcp_mark_head_lost(sk, lost); 2556 } else { 2557 int sacked_upto = tp->sacked_out - tp->reordering; 2558 if (sacked_upto < fast_rexmit) 2559 sacked_upto = fast_rexmit; 2560 tcp_mark_head_lost(sk, sacked_upto); 2561 } 2562 2563 tcp_timeout_skbs(sk); 2564} 2565 2566/* CWND moderation, preventing bursts due to too big ACKs 2567 * in dubious situations. 2568 */ 2569static inline void tcp_moderate_cwnd(struct tcp_sock *tp) 2570{ 2571 tp->snd_cwnd = min(tp->snd_cwnd, 2572 tcp_packets_in_flight(tp) + tcp_max_burst(tp)); 2573 tp->snd_cwnd_stamp = tcp_time_stamp; 2574} 2575 2576/* Lower bound on congestion window is slow start threshold 2577 * unless congestion avoidance choice decides to overide it. 2578 */ 2579static inline u32 tcp_cwnd_min(const struct sock *sk) 2580{ 2581 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 2582 2583 return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh; 2584} 2585 2586/* Decrease cwnd each second ack. */ 2587static void tcp_cwnd_down(struct sock *sk, int flag) 2588{ 2589 struct tcp_sock *tp = tcp_sk(sk); 2590 int decr = tp->snd_cwnd_cnt + 1; 2591 2592 if ((flag & (FLAG_ANY_PROGRESS | FLAG_DSACKING_ACK)) || 2593 (tcp_is_reno(tp) && !(flag & FLAG_NOT_DUP))) { 2594 tp->snd_cwnd_cnt = decr & 1; 2595 decr >>= 1; 2596 2597 if (decr && tp->snd_cwnd > tcp_cwnd_min(sk)) 2598 tp->snd_cwnd -= decr; 2599 2600 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1); 2601 tp->snd_cwnd_stamp = tcp_time_stamp; 2602 } 2603} 2604 2605/* Nothing was retransmitted or returned timestamp is less 2606 * than timestamp of the first retransmission. 2607 */ 2608static inline int tcp_packet_delayed(struct tcp_sock *tp) 2609{ 2610 return !tp->retrans_stamp || 2611 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 2612 before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp)); 2613} 2614 2615/* Undo procedures. */ 2616 2617#if FASTRETRANS_DEBUG > 1 2618static void DBGUNDO(struct sock *sk, const char *msg) 2619{ 2620 struct tcp_sock *tp = tcp_sk(sk); 2621 struct inet_sock *inet = inet_sk(sk); 2622 2623 if (sk->sk_family == AF_INET) { 2624 printk(KERN_DEBUG "Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n", 2625 msg, 2626 &inet->daddr, ntohs(inet->dport), 2627 tp->snd_cwnd, tcp_left_out(tp), 2628 tp->snd_ssthresh, tp->prior_ssthresh, 2629 tp->packets_out); 2630 } 2631#if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE) 2632 else if (sk->sk_family == AF_INET6) { 2633 struct ipv6_pinfo *np = inet6_sk(sk); 2634 printk(KERN_DEBUG "Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n", 2635 msg, 2636 &np->daddr, ntohs(inet->dport), 2637 tp->snd_cwnd, tcp_left_out(tp), 2638 tp->snd_ssthresh, tp->prior_ssthresh, 2639 tp->packets_out); 2640 } 2641#endif 2642} 2643#else 2644#define DBGUNDO(x...) do { } while (0) 2645#endif 2646 2647static void tcp_undo_cwr(struct sock *sk, const int undo) 2648{ 2649 struct tcp_sock *tp = tcp_sk(sk); 2650 2651 if (tp->prior_ssthresh) { 2652 const struct inet_connection_sock *icsk = inet_csk(sk); 2653 2654 if (icsk->icsk_ca_ops->undo_cwnd) 2655 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk); 2656 else 2657 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1); 2658 2659 if (undo && tp->prior_ssthresh > tp->snd_ssthresh) { 2660 tp->snd_ssthresh = tp->prior_ssthresh; 2661 TCP_ECN_withdraw_cwr(tp); 2662 } 2663 } else { 2664 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh); 2665 } 2666 tcp_moderate_cwnd(tp); 2667 tp->snd_cwnd_stamp = tcp_time_stamp; 2668} 2669 2670static inline int tcp_may_undo(struct tcp_sock *tp) 2671{ 2672 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp)); 2673} 2674 2675/* People celebrate: "We love our President!" */ 2676static int tcp_try_undo_recovery(struct sock *sk) 2677{ 2678 struct tcp_sock *tp = tcp_sk(sk); 2679 2680 if (tcp_may_undo(tp)) { 2681 int mib_idx; 2682 2683 /* Happy end! We did not retransmit anything 2684 * or our original transmission succeeded. 2685 */ 2686 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans"); 2687 tcp_undo_cwr(sk, 1); 2688 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss) 2689 mib_idx = LINUX_MIB_TCPLOSSUNDO; 2690 else 2691 mib_idx = LINUX_MIB_TCPFULLUNDO; 2692 2693 NET_INC_STATS_BH(sock_net(sk), mib_idx); 2694 tp->undo_marker = 0; 2695 } 2696 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) { 2697 /* Hold old state until something *above* high_seq 2698 * is ACKed. For Reno it is MUST to prevent false 2699 * fast retransmits (RFC2582). SACK TCP is safe. */ 2700 tcp_moderate_cwnd(tp); 2701 return 1; 2702 } 2703 tcp_set_ca_state(sk, TCP_CA_Open); 2704 return 0; 2705} 2706 2707/* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */ 2708static void tcp_try_undo_dsack(struct sock *sk) 2709{ 2710 struct tcp_sock *tp = tcp_sk(sk); 2711 2712 if (tp->undo_marker && !tp->undo_retrans) { 2713 DBGUNDO(sk, "D-SACK"); 2714 tcp_undo_cwr(sk, 1); 2715 tp->undo_marker = 0; 2716 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO); 2717 } 2718} 2719 2720/* Undo during fast recovery after partial ACK. */ 2721 2722static int tcp_try_undo_partial(struct sock *sk, int acked) 2723{ 2724 struct tcp_sock *tp = tcp_sk(sk); 2725 /* Partial ACK arrived. Force Hoe's retransmit. */ 2726 int failed = tcp_is_reno(tp) || (tcp_fackets_out(tp) > tp->reordering); 2727 2728 if (tcp_may_undo(tp)) { 2729 /* Plain luck! Hole if filled with delayed 2730 * packet, rather than with a retransmit. 2731 */ 2732 if (tp->retrans_out == 0) 2733 tp->retrans_stamp = 0; 2734 2735 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1); 2736 2737 DBGUNDO(sk, "Hoe"); 2738 tcp_undo_cwr(sk, 0); 2739 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO); 2740 2741 /* So... Do not make Hoe's retransmit yet. 2742 * If the first packet was delayed, the rest 2743 * ones are most probably delayed as well. 2744 */ 2745 failed = 0; 2746 } 2747 return failed; 2748} 2749 2750/* Undo during loss recovery after partial ACK. */ 2751static int tcp_try_undo_loss(struct sock *sk) 2752{ 2753 struct tcp_sock *tp = tcp_sk(sk); 2754 2755 if (tcp_may_undo(tp)) { 2756 struct sk_buff *skb; 2757 tcp_for_write_queue(skb, sk) { 2758 if (skb == tcp_send_head(sk)) 2759 break; 2760 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; 2761 } 2762 2763 tcp_clear_all_retrans_hints(tp); 2764 2765 DBGUNDO(sk, "partial loss"); 2766 tp->lost_out = 0; 2767 tcp_undo_cwr(sk, 1); 2768 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO); 2769 inet_csk(sk)->icsk_retransmits = 0; 2770 tp->undo_marker = 0; 2771 if (tcp_is_sack(tp)) 2772 tcp_set_ca_state(sk, TCP_CA_Open); 2773 return 1; 2774 } 2775 return 0; 2776} 2777 2778static inline void tcp_complete_cwr(struct sock *sk) 2779{ 2780 struct tcp_sock *tp = tcp_sk(sk); 2781 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh); 2782 tp->snd_cwnd_stamp = tcp_time_stamp; 2783 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR); 2784} 2785 2786static void tcp_try_keep_open(struct sock *sk) 2787{ 2788 struct tcp_sock *tp = tcp_sk(sk); 2789 int state = TCP_CA_Open; 2790 2791 if (tcp_left_out(tp) || tp->retrans_out || tp->undo_marker) 2792 state = TCP_CA_Disorder; 2793 2794 if (inet_csk(sk)->icsk_ca_state != state) { 2795 tcp_set_ca_state(sk, state); 2796 tp->high_seq = tp->snd_nxt; 2797 } 2798} 2799 2800static void tcp_try_to_open(struct sock *sk, int flag) 2801{ 2802 struct tcp_sock *tp = tcp_sk(sk); 2803 2804 tcp_verify_left_out(tp); 2805 2806 if (!tp->frto_counter && tp->retrans_out == 0) 2807 tp->retrans_stamp = 0; 2808 2809 if (flag & FLAG_ECE) 2810 tcp_enter_cwr(sk, 1); 2811 2812 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) { 2813 tcp_try_keep_open(sk); 2814 tcp_moderate_cwnd(tp); 2815 } else { 2816 tcp_cwnd_down(sk, flag); 2817 } 2818} 2819 2820static void tcp_mtup_probe_failed(struct sock *sk) 2821{ 2822 struct inet_connection_sock *icsk = inet_csk(sk); 2823 2824 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1; 2825 icsk->icsk_mtup.probe_size = 0; 2826} 2827 2828static void tcp_mtup_probe_success(struct sock *sk) 2829{ 2830 struct tcp_sock *tp = tcp_sk(sk); 2831 struct inet_connection_sock *icsk = inet_csk(sk); 2832 2833 /* FIXME: breaks with very large cwnd */ 2834 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2835 tp->snd_cwnd = tp->snd_cwnd * 2836 tcp_mss_to_mtu(sk, tp->mss_cache) / 2837 icsk->icsk_mtup.probe_size; 2838 tp->snd_cwnd_cnt = 0; 2839 tp->snd_cwnd_stamp = tcp_time_stamp; 2840 tp->rcv_ssthresh = tcp_current_ssthresh(sk); 2841 2842 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size; 2843 icsk->icsk_mtup.probe_size = 0; 2844 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 2845} 2846 2847/* Do a simple retransmit without using the backoff mechanisms in 2848 * tcp_timer. This is used for path mtu discovery. 2849 * The socket is already locked here. 2850 */ 2851void tcp_simple_retransmit(struct sock *sk) 2852{ 2853 const struct inet_connection_sock *icsk = inet_csk(sk); 2854 struct tcp_sock *tp = tcp_sk(sk); 2855 struct sk_buff *skb; 2856 unsigned int mss = tcp_current_mss(sk); 2857 u32 prior_lost = tp->lost_out; 2858 2859 tcp_for_write_queue(skb, sk) { 2860 if (skb == tcp_send_head(sk)) 2861 break; 2862 if (tcp_skb_seglen(skb) > mss && 2863 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) { 2864 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 2865 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 2866 tp->retrans_out -= tcp_skb_pcount(skb); 2867 } 2868 tcp_skb_mark_lost_uncond_verify(tp, skb); 2869 } 2870 } 2871 2872 tcp_clear_retrans_hints_partial(tp); 2873 2874 if (prior_lost == tp->lost_out) 2875 return; 2876 2877 if (tcp_is_reno(tp)) 2878 tcp_limit_reno_sacked(tp); 2879 2880 tcp_verify_left_out(tp); 2881 2882 /* Don't muck with the congestion window here. 2883 * Reason is that we do not increase amount of _data_ 2884 * in network, but units changed and effective 2885 * cwnd/ssthresh really reduced now. 2886 */ 2887 if (icsk->icsk_ca_state != TCP_CA_Loss) { 2888 tp->high_seq = tp->snd_nxt; 2889 tp->snd_ssthresh = tcp_current_ssthresh(sk); 2890 tp->prior_ssthresh = 0; 2891 tp->undo_marker = 0; 2892 tcp_set_ca_state(sk, TCP_CA_Loss); 2893 } 2894 tcp_xmit_retransmit_queue(sk); 2895} 2896 2897/* Process an event, which can update packets-in-flight not trivially. 2898 * Main goal of this function is to calculate new estimate for left_out, 2899 * taking into account both packets sitting in receiver's buffer and 2900 * packets lost by network. 2901 * 2902 * Besides that it does CWND reduction, when packet loss is detected 2903 * and changes state of machine. 2904 * 2905 * It does _not_ decide what to send, it is made in function 2906 * tcp_xmit_retransmit_queue(). 2907 */ 2908static void tcp_fastretrans_alert(struct sock *sk, int pkts_acked, int flag) 2909{ 2910 struct inet_connection_sock *icsk = inet_csk(sk); 2911 struct tcp_sock *tp = tcp_sk(sk); 2912 int is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP)); 2913 int do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) && 2914 (tcp_fackets_out(tp) > tp->reordering)); 2915 int fast_rexmit = 0, mib_idx; 2916 2917 if (WARN_ON(!tp->packets_out && tp->sacked_out)) 2918 tp->sacked_out = 0; 2919 if (WARN_ON(!tp->sacked_out && tp->fackets_out)) 2920 tp->fackets_out = 0; 2921 2922 /* Now state machine starts. 2923 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */ 2924 if (flag & FLAG_ECE) 2925 tp->prior_ssthresh = 0; 2926 2927 /* B. In all the states check for reneging SACKs. */ 2928 if (tcp_check_sack_reneging(sk, flag)) 2929 return; 2930 2931 /* C. Process data loss notification, provided it is valid. */ 2932 if (tcp_is_fack(tp) && (flag & FLAG_DATA_LOST) && 2933 before(tp->snd_una, tp->high_seq) && 2934 icsk->icsk_ca_state != TCP_CA_Open && 2935 tp->fackets_out > tp->reordering) { 2936 tcp_mark_head_lost(sk, tp->fackets_out - tp->reordering); 2937 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSS); 2938 } 2939 2940 /* D. Check consistency of the current state. */ 2941 tcp_verify_left_out(tp); 2942 2943 /* E. Check state exit conditions. State can be terminated 2944 * when high_seq is ACKed. */ 2945 if (icsk->icsk_ca_state == TCP_CA_Open) { 2946 WARN_ON(tp->retrans_out != 0); 2947 tp->retrans_stamp = 0; 2948 } else if (!before(tp->snd_una, tp->high_seq)) { 2949 switch (icsk->icsk_ca_state) { 2950 case TCP_CA_Loss: 2951 icsk->icsk_retransmits = 0; 2952 if (tcp_try_undo_recovery(sk)) 2953 return; 2954 break; 2955 2956 case TCP_CA_CWR: 2957 /* CWR is to be held something *above* high_seq 2958 * is ACKed for CWR bit to reach receiver. */ 2959 if (tp->snd_una != tp->high_seq) { 2960 tcp_complete_cwr(sk); 2961 tcp_set_ca_state(sk, TCP_CA_Open); 2962 } 2963 break; 2964 2965 case TCP_CA_Disorder: 2966 tcp_try_undo_dsack(sk); 2967 if (!tp->undo_marker || 2968 /* For SACK case do not Open to allow to undo 2969 * catching for all duplicate ACKs. */ 2970 tcp_is_reno(tp) || tp->snd_una != tp->high_seq) { 2971 tp->undo_marker = 0; 2972 tcp_set_ca_state(sk, TCP_CA_Open); 2973 } 2974 break; 2975 2976 case TCP_CA_Recovery: 2977 if (tcp_is_reno(tp)) 2978 tcp_reset_reno_sack(tp); 2979 if (tcp_try_undo_recovery(sk)) 2980 return; 2981 tcp_complete_cwr(sk); 2982 break; 2983 } 2984 } 2985 2986 /* F. Process state. */ 2987 switch (icsk->icsk_ca_state) { 2988 case TCP_CA_Recovery: 2989 if (!(flag & FLAG_SND_UNA_ADVANCED)) { 2990 if (tcp_is_reno(tp) && is_dupack) 2991 tcp_add_reno_sack(sk); 2992 } else 2993 do_lost = tcp_try_undo_partial(sk, pkts_acked); 2994 break; 2995 case TCP_CA_Loss: 2996 if (flag & FLAG_DATA_ACKED) 2997 icsk->icsk_retransmits = 0; 2998 if (tcp_is_reno(tp) && flag & FLAG_SND_UNA_ADVANCED) 2999 tcp_reset_reno_sack(tp); 3000 if (!tcp_try_undo_loss(sk)) { 3001 tcp_moderate_cwnd(tp); 3002 tcp_xmit_retransmit_queue(sk); 3003 return; 3004 } 3005 if (icsk->icsk_ca_state != TCP_CA_Open) 3006 return; 3007 /* Loss is undone; fall through to processing in Open state. */ 3008 default: 3009 if (tcp_is_reno(tp)) { 3010 if (flag & FLAG_SND_UNA_ADVANCED) 3011 tcp_reset_reno_sack(tp); 3012 if (is_dupack) 3013 tcp_add_reno_sack(sk); 3014 } 3015 3016 if (icsk->icsk_ca_state == TCP_CA_Disorder) 3017 tcp_try_undo_dsack(sk); 3018 3019 if (!tcp_time_to_recover(sk)) { 3020 tcp_try_to_open(sk, flag); 3021 return; 3022 } 3023 3024 /* MTU probe failure: don't reduce cwnd */ 3025 if (icsk->icsk_ca_state < TCP_CA_CWR && 3026 icsk->icsk_mtup.probe_size && 3027 tp->snd_una == tp->mtu_probe.probe_seq_start) { 3028 tcp_mtup_probe_failed(sk); 3029 /* Restores the reduction we did in tcp_mtup_probe() */ 3030 tp->snd_cwnd++; 3031 tcp_simple_retransmit(sk); 3032 return; 3033 } 3034 3035 /* Otherwise enter Recovery state */ 3036 3037 if (tcp_is_reno(tp)) 3038 mib_idx = LINUX_MIB_TCPRENORECOVERY; 3039 else 3040 mib_idx = LINUX_MIB_TCPSACKRECOVERY; 3041 3042 NET_INC_STATS_BH(sock_net(sk), mib_idx); 3043 3044 tp->high_seq = tp->snd_nxt; 3045 tp->prior_ssthresh = 0; 3046 tp->undo_marker = tp->snd_una; 3047 tp->undo_retrans = tp->retrans_out; 3048 3049 if (icsk->icsk_ca_state < TCP_CA_CWR) { 3050 if (!(flag & FLAG_ECE)) 3051 tp->prior_ssthresh = tcp_current_ssthresh(sk); 3052 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 3053 TCP_ECN_queue_cwr(tp); 3054 } 3055 3056 tp->bytes_acked = 0; 3057 tp->snd_cwnd_cnt = 0; 3058 tcp_set_ca_state(sk, TCP_CA_Recovery); 3059 fast_rexmit = 1; 3060 } 3061 3062 if (do_lost || (tcp_is_fack(tp) && tcp_head_timedout(sk))) 3063 tcp_update_scoreboard(sk, fast_rexmit); 3064 tcp_cwnd_down(sk, flag); 3065 tcp_xmit_retransmit_queue(sk); 3066} 3067 3068static void tcp_valid_rtt_meas(struct sock *sk, u32 seq_rtt) 3069{ 3070 tcp_rtt_estimator(sk, seq_rtt); 3071 tcp_set_rto(sk); 3072 inet_csk(sk)->icsk_backoff = 0; 3073} 3074 3075/* Read draft-ietf-tcplw-high-performance before mucking 3076 * with this code. (Supersedes RFC1323) 3077 */ 3078static void tcp_ack_saw_tstamp(struct sock *sk, int flag) 3079{ 3080 /* RTTM Rule: A TSecr value received in a segment is used to 3081 * update the averaged RTT measurement only if the segment 3082 * acknowledges some new data, i.e., only if it advances the 3083 * left edge of the send window. 3084 * 3085 * See draft-ietf-tcplw-high-performance-00, section 3.3. 3086 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru> 3087 * 3088 * Changed: reset backoff as soon as we see the first valid sample. 3089 * If we do not, we get strongly overestimated rto. With timestamps 3090 * samples are accepted even from very old segments: f.e., when rtt=1 3091 * increases to 8, we retransmit 5 times and after 8 seconds delayed 3092 * answer arrives rto becomes 120 seconds! If at least one of segments 3093 * in window is lost... Voila. --ANK (010210) 3094 */ 3095 struct tcp_sock *tp = tcp_sk(sk); 3096 3097 tcp_valid_rtt_meas(sk, tcp_time_stamp - tp->rx_opt.rcv_tsecr); 3098} 3099 3100static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag) 3101{ 3102 /* We don't have a timestamp. Can only use 3103 * packets that are not retransmitted to determine 3104 * rtt estimates. Also, we must not reset the 3105 * backoff for rto until we get a non-retransmitted 3106 * packet. This allows us to deal with a situation 3107 * where the network delay has increased suddenly. 3108 * I.e. Karn's algorithm. (SIGCOMM '87, p5.) 3109 */ 3110 3111 if (flag & FLAG_RETRANS_DATA_ACKED) 3112 return; 3113 3114 tcp_valid_rtt_meas(sk, seq_rtt); 3115} 3116 3117static inline void tcp_ack_update_rtt(struct sock *sk, const int flag, 3118 const s32 seq_rtt) 3119{ 3120 const struct tcp_sock *tp = tcp_sk(sk); 3121 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */ 3122 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr) 3123 tcp_ack_saw_tstamp(sk, flag); 3124 else if (seq_rtt >= 0) 3125 tcp_ack_no_tstamp(sk, seq_rtt, flag); 3126} 3127 3128static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight) 3129{ 3130 const struct inet_connection_sock *icsk = inet_csk(sk); 3131 icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight); 3132 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp; 3133} 3134 3135/* Restart timer after forward progress on connection. 3136 * RFC2988 recommends to restart timer to now+rto. 3137 */ 3138static void tcp_rearm_rto(struct sock *sk) 3139{ 3140 struct tcp_sock *tp = tcp_sk(sk); 3141 3142 if (!tp->packets_out) { 3143 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS); 3144 } else { 3145 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 3146 inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 3147 } 3148} 3149 3150/* If we get here, the whole TSO packet has not been acked. */ 3151static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb) 3152{ 3153 struct tcp_sock *tp = tcp_sk(sk); 3154 u32 packets_acked; 3155 3156 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)); 3157 3158 packets_acked = tcp_skb_pcount(skb); 3159 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 3160 return 0; 3161 packets_acked -= tcp_skb_pcount(skb); 3162 3163 if (packets_acked) { 3164 BUG_ON(tcp_skb_pcount(skb) == 0); 3165 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)); 3166 } 3167 3168 return packets_acked; 3169} 3170 3171/* Remove acknowledged frames from the retransmission queue. If our packet 3172 * is before the ack sequence we can discard it as it's confirmed to have 3173 * arrived at the other end. 3174 */ 3175static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets, 3176 u32 prior_snd_una) 3177{ 3178 struct tcp_sock *tp = tcp_sk(sk); 3179 const struct inet_connection_sock *icsk = inet_csk(sk); 3180 struct sk_buff *skb; 3181 u32 now = tcp_time_stamp; 3182 int fully_acked = 1; 3183 int flag = 0; 3184 u32 pkts_acked = 0; 3185 u32 reord = tp->packets_out; 3186 u32 prior_sacked = tp->sacked_out; 3187 s32 seq_rtt = -1; 3188 s32 ca_seq_rtt = -1; 3189 ktime_t last_ackt = net_invalid_timestamp(); 3190 3191 while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) { 3192 struct tcp_skb_cb *scb = TCP_SKB_CB(skb); 3193 u32 acked_pcount; 3194 u8 sacked = scb->sacked; 3195 3196 /* Determine how many packets and what bytes were acked, tso and else */ 3197 if (after(scb->end_seq, tp->snd_una)) { 3198 if (tcp_skb_pcount(skb) == 1 || 3199 !after(tp->snd_una, scb->seq)) 3200 break; 3201 3202 acked_pcount = tcp_tso_acked(sk, skb); 3203 if (!acked_pcount) 3204 break; 3205 3206 fully_acked = 0; 3207 } else { 3208 acked_pcount = tcp_skb_pcount(skb); 3209 } 3210 3211 if (sacked & TCPCB_RETRANS) { 3212 if (sacked & TCPCB_SACKED_RETRANS) 3213 tp->retrans_out -= acked_pcount; 3214 flag |= FLAG_RETRANS_DATA_ACKED; 3215 ca_seq_rtt = -1; 3216 seq_rtt = -1; 3217 if ((flag & FLAG_DATA_ACKED) || (acked_pcount > 1)) 3218 flag |= FLAG_NONHEAD_RETRANS_ACKED; 3219 } else { 3220 ca_seq_rtt = now - scb->when; 3221 last_ackt = skb->tstamp; 3222 if (seq_rtt < 0) { 3223 seq_rtt = ca_seq_rtt; 3224 } 3225 if (!(sacked & TCPCB_SACKED_ACKED)) 3226 reord = min(pkts_acked, reord); 3227 } 3228 3229 if (sacked & TCPCB_SACKED_ACKED) 3230 tp->sacked_out -= acked_pcount; 3231 if (sacked & TCPCB_LOST) 3232 tp->lost_out -= acked_pcount; 3233 3234 tp->packets_out -= acked_pcount; 3235 pkts_acked += acked_pcount; 3236 3237 /* Initial outgoing SYN's get put onto the write_queue 3238 * just like anything else we transmit. It is not 3239 * true data, and if we misinform our callers that 3240 * this ACK acks real data, we will erroneously exit 3241 * connection startup slow start one packet too 3242 * quickly. This is severely frowned upon behavior. 3243 */ 3244 if (!(scb->flags & TCPCB_FLAG_SYN)) { 3245 flag |= FLAG_DATA_ACKED; 3246 } else { 3247 flag |= FLAG_SYN_ACKED; 3248 tp->retrans_stamp = 0; 3249 } 3250 3251 if (!fully_acked) 3252 break; 3253 3254 tcp_unlink_write_queue(skb, sk); 3255 sk_wmem_free_skb(sk, skb); 3256 tp->scoreboard_skb_hint = NULL; 3257 if (skb == tp->retransmit_skb_hint) 3258 tp->retransmit_skb_hint = NULL; 3259 if (skb == tp->lost_skb_hint) 3260 tp->lost_skb_hint = NULL; 3261 } 3262 3263 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una))) 3264 tp->snd_up = tp->snd_una; 3265 3266 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) 3267 flag |= FLAG_SACK_RENEGING; 3268 3269 if (flag & FLAG_ACKED) { 3270 const struct tcp_congestion_ops *ca_ops 3271 = inet_csk(sk)->icsk_ca_ops; 3272 3273 if (unlikely(icsk->icsk_mtup.probe_size && 3274 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) { 3275 tcp_mtup_probe_success(sk); 3276 } 3277 3278 tcp_ack_update_rtt(sk, flag, seq_rtt); 3279 tcp_rearm_rto(sk); 3280 3281 if (tcp_is_reno(tp)) { 3282 tcp_remove_reno_sacks(sk, pkts_acked); 3283 } else { 3284 int delta; 3285 3286 /* Non-retransmitted hole got filled? That's reordering */ 3287 if (reord < prior_fackets) 3288 tcp_update_reordering(sk, tp->fackets_out - reord, 0); 3289 3290 delta = tcp_is_fack(tp) ? pkts_acked : 3291 prior_sacked - tp->sacked_out; 3292 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta); 3293 } 3294 3295 tp->fackets_out -= min(pkts_acked, tp->fackets_out); 3296 3297 if (ca_ops->pkts_acked) { 3298 s32 rtt_us = -1; 3299 3300 /* Is the ACK triggering packet unambiguous? */ 3301 if (!(flag & FLAG_RETRANS_DATA_ACKED)) { 3302 /* High resolution needed and available? */ 3303 if (ca_ops->flags & TCP_CONG_RTT_STAMP && 3304 !ktime_equal(last_ackt, 3305 net_invalid_timestamp())) 3306 rtt_us = ktime_us_delta(ktime_get_real(), 3307 last_ackt); 3308 else if (ca_seq_rtt > 0) 3309 rtt_us = jiffies_to_usecs(ca_seq_rtt); 3310 } 3311 3312 ca_ops->pkts_acked(sk, pkts_acked, rtt_us); 3313 } 3314 } 3315 3316#if FASTRETRANS_DEBUG > 0 3317 WARN_ON((int)tp->sacked_out < 0); 3318 WARN_ON((int)tp->lost_out < 0); 3319 WARN_ON((int)tp->retrans_out < 0); 3320 if (!tp->packets_out && tcp_is_sack(tp)) { 3321 icsk = inet_csk(sk); 3322 if (tp->lost_out) { 3323 printk(KERN_DEBUG "Leak l=%u %d\n", 3324 tp->lost_out, icsk->icsk_ca_state); 3325 tp->lost_out = 0; 3326 } 3327 if (tp->sacked_out) { 3328 printk(KERN_DEBUG "Leak s=%u %d\n", 3329 tp->sacked_out, icsk->icsk_ca_state); 3330 tp->sacked_out = 0; 3331 } 3332 if (tp->retrans_out) { 3333 printk(KERN_DEBUG "Leak r=%u %d\n", 3334 tp->retrans_out, icsk->icsk_ca_state); 3335 tp->retrans_out = 0; 3336 } 3337 } 3338#endif 3339 return flag; 3340} 3341 3342static void tcp_ack_probe(struct sock *sk) 3343{ 3344 const struct tcp_sock *tp = tcp_sk(sk); 3345 struct inet_connection_sock *icsk = inet_csk(sk); 3346 3347 /* Was it a usable window open? */ 3348 3349 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) { 3350 icsk->icsk_backoff = 0; 3351 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0); 3352 /* Socket must be waked up by subsequent tcp_data_snd_check(). 3353 * This function is not for random using! 3354 */ 3355 } else { 3356 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 3357 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), 3358 TCP_RTO_MAX); 3359 } 3360} 3361 3362static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag) 3363{ 3364 return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) || 3365 inet_csk(sk)->icsk_ca_state != TCP_CA_Open); 3366} 3367 3368static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag) 3369{ 3370 const struct tcp_sock *tp = tcp_sk(sk); 3371 return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) && 3372 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR)); 3373} 3374 3375/* Check that window update is acceptable. 3376 * The function assumes that snd_una<=ack<=snd_next. 3377 */ 3378static inline int tcp_may_update_window(const struct tcp_sock *tp, 3379 const u32 ack, const u32 ack_seq, 3380 const u32 nwin) 3381{ 3382 return (after(ack, tp->snd_una) || 3383 after(ack_seq, tp->snd_wl1) || 3384 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd)); 3385} 3386 3387/* Update our send window. 3388 * 3389 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2 3390 * and in FreeBSD. NetBSD's one is even worse.) is wrong. 3391 */ 3392static int tcp_ack_update_window(struct sock *sk, struct sk_buff *skb, u32 ack, 3393 u32 ack_seq) 3394{ 3395 struct tcp_sock *tp = tcp_sk(sk); 3396 int flag = 0; 3397 u32 nwin = ntohs(tcp_hdr(skb)->window); 3398 3399 if (likely(!tcp_hdr(skb)->syn)) 3400 nwin <<= tp->rx_opt.snd_wscale; 3401 3402 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) { 3403 flag |= FLAG_WIN_UPDATE; 3404 tcp_update_wl(tp, ack_seq); 3405 3406 if (tp->snd_wnd != nwin) { 3407 tp->snd_wnd = nwin; 3408 3409 /* Note, it is the only place, where 3410 * fast path is recovered for sending TCP. 3411 */ 3412 tp->pred_flags = 0; 3413 tcp_fast_path_check(sk); 3414 3415 if (nwin > tp->max_window) { 3416 tp->max_window = nwin; 3417 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie); 3418 } 3419 } 3420 } 3421 3422 tp->snd_una = ack; 3423 3424 return flag; 3425} 3426 3427/* A very conservative spurious RTO response algorithm: reduce cwnd and 3428 * continue in congestion avoidance. 3429 */ 3430static void tcp_conservative_spur_to_response(struct tcp_sock *tp) 3431{ 3432 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh); 3433 tp->snd_cwnd_cnt = 0; 3434 tp->bytes_acked = 0; 3435 TCP_ECN_queue_cwr(tp); 3436 tcp_moderate_cwnd(tp); 3437} 3438 3439/* A conservative spurious RTO response algorithm: reduce cwnd using 3440 * rate halving and continue in congestion avoidance. 3441 */ 3442static void tcp_ratehalving_spur_to_response(struct sock *sk) 3443{ 3444 tcp_enter_cwr(sk, 0); 3445} 3446 3447static void tcp_undo_spur_to_response(struct sock *sk, int flag) 3448{ 3449 if (flag & FLAG_ECE) 3450 tcp_ratehalving_spur_to_response(sk); 3451 else 3452 tcp_undo_cwr(sk, 1); 3453} 3454 3455/* F-RTO spurious RTO detection algorithm (RFC4138) 3456 * 3457 * F-RTO affects during two new ACKs following RTO (well, almost, see inline 3458 * comments). State (ACK number) is kept in frto_counter. When ACK advances 3459 * window (but not to or beyond highest sequence sent before RTO): 3460 * On First ACK, send two new segments out. 3461 * On Second ACK, RTO was likely spurious. Do spurious response (response 3462 * algorithm is not part of the F-RTO detection algorithm 3463 * given in RFC4138 but can be selected separately). 3464 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss 3465 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding 3466 * of Nagle, this is done using frto_counter states 2 and 3, when a new data 3467 * segment of any size sent during F-RTO, state 2 is upgraded to 3. 3468 * 3469 * Rationale: if the RTO was spurious, new ACKs should arrive from the 3470 * original window even after we transmit two new data segments. 3471 * 3472 * SACK version: 3473 * on first step, wait until first cumulative ACK arrives, then move to 3474 * the second step. In second step, the next ACK decides. 3475 * 3476 * F-RTO is implemented (mainly) in four functions: 3477 * - tcp_use_frto() is used to determine if TCP is can use F-RTO 3478 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is 3479 * called when tcp_use_frto() showed green light 3480 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm 3481 * - tcp_enter_frto_loss() is called if there is not enough evidence 3482 * to prove that the RTO is indeed spurious. It transfers the control 3483 * from F-RTO to the conventional RTO recovery 3484 */ 3485static int tcp_process_frto(struct sock *sk, int flag) 3486{ 3487 struct tcp_sock *tp = tcp_sk(sk); 3488 3489 tcp_verify_left_out(tp); 3490 3491 /* Duplicate the behavior from Loss state (fastretrans_alert) */ 3492 if (flag & FLAG_DATA_ACKED) 3493 inet_csk(sk)->icsk_retransmits = 0; 3494 3495 if ((flag & FLAG_NONHEAD_RETRANS_ACKED) || 3496 ((tp->frto_counter >= 2) && (flag & FLAG_RETRANS_DATA_ACKED))) 3497 tp->undo_marker = 0; 3498 3499 if (!before(tp->snd_una, tp->frto_highmark)) { 3500 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag); 3501 return 1; 3502 } 3503 3504 if (!tcp_is_sackfrto(tp)) { 3505 /* RFC4138 shortcoming in step 2; should also have case c): 3506 * ACK isn't duplicate nor advances window, e.g., opposite dir 3507 * data, winupdate 3508 */ 3509 if (!(flag & FLAG_ANY_PROGRESS) && (flag & FLAG_NOT_DUP)) 3510 return 1; 3511 3512 if (!(flag & FLAG_DATA_ACKED)) { 3513 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3), 3514 flag); 3515 return 1; 3516 } 3517 } else { 3518 if (!(flag & FLAG_DATA_ACKED) && (tp->frto_counter == 1)) { 3519 /* Prevent sending of new data. */ 3520 tp->snd_cwnd = min(tp->snd_cwnd, 3521 tcp_packets_in_flight(tp)); 3522 return 1; 3523 } 3524 3525 if ((tp->frto_counter >= 2) && 3526 (!(flag & FLAG_FORWARD_PROGRESS) || 3527 ((flag & FLAG_DATA_SACKED) && 3528 !(flag & FLAG_ONLY_ORIG_SACKED)))) { 3529 /* RFC4138 shortcoming (see comment above) */ 3530 if (!(flag & FLAG_FORWARD_PROGRESS) && 3531 (flag & FLAG_NOT_DUP)) 3532 return 1; 3533 3534 tcp_enter_frto_loss(sk, 3, flag); 3535 return 1; 3536 } 3537 } 3538 3539 if (tp->frto_counter == 1) { 3540 /* tcp_may_send_now needs to see updated state */ 3541 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2; 3542 tp->frto_counter = 2; 3543 3544 if (!tcp_may_send_now(sk)) 3545 tcp_enter_frto_loss(sk, 2, flag); 3546 3547 return 1; 3548 } else { 3549 switch (sysctl_tcp_frto_response) { 3550 case 2: 3551 tcp_undo_spur_to_response(sk, flag); 3552 break; 3553 case 1: 3554 tcp_conservative_spur_to_response(tp); 3555 break; 3556 default: 3557 tcp_ratehalving_spur_to_response(sk); 3558 break; 3559 } 3560 tp->frto_counter = 0; 3561 tp->undo_marker = 0; 3562 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSPURIOUSRTOS); 3563 } 3564 return 0; 3565} 3566 3567/* This routine deals with incoming acks, but not outgoing ones. */ 3568static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag) 3569{ 3570 struct inet_connection_sock *icsk = inet_csk(sk); 3571 struct tcp_sock *tp = tcp_sk(sk); 3572 u32 prior_snd_una = tp->snd_una; 3573 u32 ack_seq = TCP_SKB_CB(skb)->seq; 3574 u32 ack = TCP_SKB_CB(skb)->ack_seq; 3575 u32 prior_in_flight; 3576 u32 prior_fackets; 3577 int prior_packets; 3578 int frto_cwnd = 0; 3579 3580 /* If the ack is older than previous acks 3581 * then we can probably ignore it. 3582 */ 3583 if (before(ack, prior_snd_una)) 3584 goto old_ack; 3585 3586 /* If the ack includes data we haven't sent yet, discard 3587 * this segment (RFC793 Section 3.9). 3588 */ 3589 if (after(ack, tp->snd_nxt)) 3590 goto invalid_ack; 3591 3592 if (after(ack, prior_snd_una)) 3593 flag |= FLAG_SND_UNA_ADVANCED; 3594 3595 if (sysctl_tcp_abc) { 3596 if (icsk->icsk_ca_state < TCP_CA_CWR) 3597 tp->bytes_acked += ack - prior_snd_una; 3598 else if (icsk->icsk_ca_state == TCP_CA_Loss) 3599 /* we assume just one segment left network */ 3600 tp->bytes_acked += min(ack - prior_snd_una, 3601 tp->mss_cache); 3602 } 3603 3604 prior_fackets = tp->fackets_out; 3605 prior_in_flight = tcp_packets_in_flight(tp); 3606 3607 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) { 3608 /* Window is constant, pure forward advance. 3609 * No more checks are required. 3610 * Note, we use the fact that SND.UNA>=SND.WL2. 3611 */ 3612 tcp_update_wl(tp, ack_seq); 3613 tp->snd_una = ack; 3614 flag |= FLAG_WIN_UPDATE; 3615 3616 tcp_ca_event(sk, CA_EVENT_FAST_ACK); 3617 3618 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS); 3619 } else { 3620 if (ack_seq != TCP_SKB_CB(skb)->end_seq) 3621 flag |= FLAG_DATA; 3622 else 3623 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS); 3624 3625 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq); 3626 3627 if (TCP_SKB_CB(skb)->sacked) 3628 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una); 3629 3630 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb))) 3631 flag |= FLAG_ECE; 3632 3633 tcp_ca_event(sk, CA_EVENT_SLOW_ACK); 3634 } 3635 3636 /* We passed data and got it acked, remove any soft error 3637 * log. Something worked... 3638 */ 3639 sk->sk_err_soft = 0; 3640 icsk->icsk_probes_out = 0; 3641 tp->rcv_tstamp = tcp_time_stamp; 3642 prior_packets = tp->packets_out; 3643 if (!prior_packets) 3644 goto no_queue; 3645 3646 /* See if we can take anything off of the retransmit queue. */ 3647 flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una); 3648 3649 if (tp->frto_counter) 3650 frto_cwnd = tcp_process_frto(sk, flag); 3651 /* Guarantee sacktag reordering detection against wrap-arounds */ 3652 if (before(tp->frto_highmark, tp->snd_una)) 3653 tp->frto_highmark = 0; 3654 3655 if (tcp_ack_is_dubious(sk, flag)) { 3656 /* Advance CWND, if state allows this. */ 3657 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd && 3658 tcp_may_raise_cwnd(sk, flag)) 3659 tcp_cong_avoid(sk, ack, prior_in_flight); 3660 tcp_fastretrans_alert(sk, prior_packets - tp->packets_out, 3661 flag); 3662 } else { 3663 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd) 3664 tcp_cong_avoid(sk, ack, prior_in_flight); 3665 } 3666 3667 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) 3668 dst_confirm(sk->sk_dst_cache); 3669 3670 return 1; 3671 3672no_queue: 3673 /* If this ack opens up a zero window, clear backoff. It was 3674 * being used to time the probes, and is probably far higher than 3675 * it needs to be for normal retransmission. 3676 */ 3677 if (tcp_send_head(sk)) 3678 tcp_ack_probe(sk); 3679 return 1; 3680 3681invalid_ack: 3682 SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt); 3683 return -1; 3684 3685old_ack: 3686 if (TCP_SKB_CB(skb)->sacked) { 3687 tcp_sacktag_write_queue(sk, skb, prior_snd_una); 3688 if (icsk->icsk_ca_state == TCP_CA_Open) 3689 tcp_try_keep_open(sk); 3690 } 3691 3692 SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt); 3693 return 0; 3694} 3695 3696/* Look for tcp options. Normally only called on SYN and SYNACK packets. 3697 * But, this can also be called on packets in the established flow when 3698 * the fast version below fails. 3699 */ 3700void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx, 3701 int estab, struct dst_entry *dst) 3702{ 3703 unsigned char *ptr; 3704 struct tcphdr *th = tcp_hdr(skb); 3705 int length = (th->doff * 4) - sizeof(struct tcphdr); 3706 3707 BUG_ON(!estab && !dst); 3708 3709 ptr = (unsigned char *)(th + 1); 3710 opt_rx->saw_tstamp = 0; 3711 3712 while (length > 0) { 3713 int opcode = *ptr++; 3714 int opsize; 3715 3716 switch (opcode) { 3717 case TCPOPT_EOL: 3718 return; 3719 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */ 3720 length--; 3721 continue; 3722 default: 3723 opsize = *ptr++; 3724 if (opsize < 2) /* "silly options" */ 3725 return; 3726 if (opsize > length) 3727 return; /* don't parse partial options */ 3728 switch (opcode) { 3729 case TCPOPT_MSS: 3730 if (opsize == TCPOLEN_MSS && th->syn && !estab) { 3731 u16 in_mss = get_unaligned_be16(ptr); 3732 if (in_mss) { 3733 if (opt_rx->user_mss && 3734 opt_rx->user_mss < in_mss) 3735 in_mss = opt_rx->user_mss; 3736 opt_rx->mss_clamp = in_mss; 3737 } 3738 } 3739 break; 3740 case TCPOPT_WINDOW: 3741 if (opsize == TCPOLEN_WINDOW && th->syn && 3742 !estab && sysctl_tcp_window_scaling) { 3743 __u8 snd_wscale = *(__u8 *)ptr; 3744 opt_rx->wscale_ok = 1; 3745 if (snd_wscale > 14) { 3746 if (net_ratelimit()) 3747 printk(KERN_INFO "tcp_parse_options: Illegal window " 3748 "scaling value %d >14 received.\n", 3749 snd_wscale); 3750 snd_wscale = 14; 3751 } 3752 opt_rx->snd_wscale = snd_wscale; 3753 } 3754 break; 3755 case TCPOPT_TIMESTAMP: 3756 if ((opsize == TCPOLEN_TIMESTAMP) && 3757 ((estab && opt_rx->tstamp_ok) || 3758 (!estab && sysctl_tcp_timestamps && 3759 !dst_feature(dst, RTAX_FEATURE_NO_TSTAMP)))) { 3760 opt_rx->saw_tstamp = 1; 3761 opt_rx->rcv_tsval = get_unaligned_be32(ptr); 3762 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4); 3763 } 3764 break; 3765 case TCPOPT_SACK_PERM: 3766 if (opsize == TCPOLEN_SACK_PERM && th->syn && 3767 !estab && sysctl_tcp_sack && 3768 !dst_feature(dst, RTAX_FEATURE_NO_SACK)) { 3769 opt_rx->sack_ok = 1; 3770 tcp_sack_reset(opt_rx); 3771 } 3772 break; 3773 3774 case TCPOPT_SACK: 3775 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) && 3776 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) && 3777 opt_rx->sack_ok) { 3778 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th; 3779 } 3780 break; 3781#ifdef CONFIG_TCP_MD5SIG 3782 case TCPOPT_MD5SIG: 3783 /* 3784 * The MD5 Hash has already been 3785 * checked (see tcp_v{4,6}_do_rcv()). 3786 */ 3787 break; 3788#endif 3789 } 3790 3791 ptr += opsize-2; 3792 length -= opsize; 3793 } 3794 } 3795} 3796 3797static int tcp_parse_aligned_timestamp(struct tcp_sock *tp, struct tcphdr *th) 3798{ 3799 __be32 *ptr = (__be32 *)(th + 1); 3800 3801 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) 3802 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) { 3803 tp->rx_opt.saw_tstamp = 1; 3804 ++ptr; 3805 tp->rx_opt.rcv_tsval = ntohl(*ptr); 3806 ++ptr; 3807 tp->rx_opt.rcv_tsecr = ntohl(*ptr); 3808 return 1; 3809 } 3810 return 0; 3811} 3812 3813/* Fast parse options. This hopes to only see timestamps. 3814 * If it is wrong it falls back on tcp_parse_options(). 3815 */ 3816static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th, 3817 struct tcp_sock *tp) 3818{ 3819 if (th->doff == sizeof(struct tcphdr) >> 2) { 3820 tp->rx_opt.saw_tstamp = 0; 3821 return 0; 3822 } else if (tp->rx_opt.tstamp_ok && 3823 th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) { 3824 if (tcp_parse_aligned_timestamp(tp, th)) 3825 return 1; 3826 } 3827 tcp_parse_options(skb, &tp->rx_opt, 1, NULL); 3828 return 1; 3829} 3830 3831#ifdef CONFIG_TCP_MD5SIG 3832/* 3833 * Parse MD5 Signature option 3834 */ 3835u8 *tcp_parse_md5sig_option(struct tcphdr *th) 3836{ 3837 int length = (th->doff << 2) - sizeof (*th); 3838 u8 *ptr = (u8*)(th + 1); 3839 3840 /* If the TCP option is too short, we can short cut */ 3841 if (length < TCPOLEN_MD5SIG) 3842 return NULL; 3843 3844 while (length > 0) { 3845 int opcode = *ptr++; 3846 int opsize; 3847 3848 switch(opcode) { 3849 case TCPOPT_EOL: 3850 return NULL; 3851 case TCPOPT_NOP: 3852 length--; 3853 continue; 3854 default: 3855 opsize = *ptr++; 3856 if (opsize < 2 || opsize > length) 3857 return NULL; 3858 if (opcode == TCPOPT_MD5SIG) 3859 return ptr; 3860 } 3861 ptr += opsize - 2; 3862 length -= opsize; 3863 } 3864 return NULL; 3865} 3866#endif 3867 3868static inline void tcp_store_ts_recent(struct tcp_sock *tp) 3869{ 3870 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval; 3871 tp->rx_opt.ts_recent_stamp = get_seconds(); 3872} 3873 3874static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq) 3875{ 3876 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) { 3877 /* PAWS bug workaround wrt. ACK frames, the PAWS discard 3878 * extra check below makes sure this can only happen 3879 * for pure ACK frames. -DaveM 3880 * 3881 * Not only, also it occurs for expired timestamps. 3882 */ 3883 3884 if (tcp_paws_check(&tp->rx_opt, 0)) 3885 tcp_store_ts_recent(tp); 3886 } 3887} 3888 3889/* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM 3890 * 3891 * It is not fatal. If this ACK does _not_ change critical state (seqs, window) 3892 * it can pass through stack. So, the following predicate verifies that 3893 * this segment is not used for anything but congestion avoidance or 3894 * fast retransmit. Moreover, we even are able to eliminate most of such 3895 * second order effects, if we apply some small "replay" window (~RTO) 3896 * to timestamp space. 3897 * 3898 * All these measures still do not guarantee that we reject wrapped ACKs 3899 * on networks with high bandwidth, when sequence space is recycled fastly, 3900 * but it guarantees that such events will be very rare and do not affect 3901 * connection seriously. This doesn't look nice, but alas, PAWS is really 3902 * buggy extension. 3903 * 3904 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC 3905 * states that events when retransmit arrives after original data are rare. 3906 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is 3907 * the biggest problem on large power networks even with minor reordering. 3908 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe 3909 * up to bandwidth of 18Gigabit/sec. 8) ] 3910 */ 3911 3912static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb) 3913{ 3914 struct tcp_sock *tp = tcp_sk(sk); 3915 struct tcphdr *th = tcp_hdr(skb); 3916 u32 seq = TCP_SKB_CB(skb)->seq; 3917 u32 ack = TCP_SKB_CB(skb)->ack_seq; 3918 3919 return (/* 1. Pure ACK with correct sequence number. */ 3920 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) && 3921 3922 /* 2. ... and duplicate ACK. */ 3923 ack == tp->snd_una && 3924 3925 /* 3. ... and does not update window. */ 3926 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) && 3927 3928 /* 4. ... and sits in replay window. */ 3929 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ); 3930} 3931 3932static inline int tcp_paws_discard(const struct sock *sk, 3933 const struct sk_buff *skb) 3934{ 3935 const struct tcp_sock *tp = tcp_sk(sk); 3936 3937 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) && 3938 !tcp_disordered_ack(sk, skb); 3939} 3940 3941/* Check segment sequence number for validity. 3942 * 3943 * Segment controls are considered valid, if the segment 3944 * fits to the window after truncation to the window. Acceptability 3945 * of data (and SYN, FIN, of course) is checked separately. 3946 * See tcp_data_queue(), for example. 3947 * 3948 * Also, controls (RST is main one) are accepted using RCV.WUP instead 3949 * of RCV.NXT. Peer still did not advance his SND.UNA when we 3950 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP. 3951 * (borrowed from freebsd) 3952 */ 3953 3954static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq) 3955{ 3956 return !before(end_seq, tp->rcv_wup) && 3957 !after(seq, tp->rcv_nxt + tcp_receive_window(tp)); 3958} 3959 3960/* When we get a reset we do this. */ 3961static void tcp_reset(struct sock *sk) 3962{ 3963 /* We want the right error as BSD sees it (and indeed as we do). */ 3964 switch (sk->sk_state) { 3965 case TCP_SYN_SENT: 3966 sk->sk_err = ECONNREFUSED; 3967 break; 3968 case TCP_CLOSE_WAIT: 3969 sk->sk_err = EPIPE; 3970 break; 3971 case TCP_CLOSE: 3972 return; 3973 default: 3974 sk->sk_err = ECONNRESET; 3975 } 3976 3977 if (!sock_flag(sk, SOCK_DEAD)) 3978 sk->sk_error_report(sk); 3979 3980 tcp_done(sk); 3981} 3982 3983/* 3984 * Process the FIN bit. This now behaves as it is supposed to work 3985 * and the FIN takes effect when it is validly part of sequence 3986 * space. Not before when we get holes. 3987 * 3988 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT 3989 * (and thence onto LAST-ACK and finally, CLOSE, we never enter 3990 * TIME-WAIT) 3991 * 3992 * If we are in FINWAIT-1, a received FIN indicates simultaneous 3993 * close and we go into CLOSING (and later onto TIME-WAIT) 3994 * 3995 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT. 3996 */ 3997static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th) 3998{ 3999 struct tcp_sock *tp = tcp_sk(sk); 4000 4001 inet_csk_schedule_ack(sk); 4002 4003 sk->sk_shutdown |= RCV_SHUTDOWN; 4004 sock_set_flag(sk, SOCK_DONE); 4005 4006 switch (sk->sk_state) { 4007 case TCP_SYN_RECV: 4008 case TCP_ESTABLISHED: 4009 /* Move to CLOSE_WAIT */ 4010 tcp_set_state(sk, TCP_CLOSE_WAIT); 4011 inet_csk(sk)->icsk_ack.pingpong = 1; 4012 break; 4013 4014 case TCP_CLOSE_WAIT: 4015 case TCP_CLOSING: 4016 /* Received a retransmission of the FIN, do 4017 * nothing. 4018 */ 4019 break; 4020 case TCP_LAST_ACK: 4021 /* RFC793: Remain in the LAST-ACK state. */ 4022 break; 4023 4024 case TCP_FIN_WAIT1: 4025 /* This case occurs when a simultaneous close 4026 * happens, we must ack the received FIN and 4027 * enter the CLOSING state. 4028 */ 4029 tcp_send_ack(sk); 4030 tcp_set_state(sk, TCP_CLOSING); 4031 break; 4032 case TCP_FIN_WAIT2: 4033 /* Received a FIN -- send ACK and enter TIME_WAIT. */ 4034 tcp_send_ack(sk); 4035 tcp_time_wait(sk, TCP_TIME_WAIT, 0); 4036 break; 4037 default: 4038 /* Only TCP_LISTEN and TCP_CLOSE are left, in these 4039 * cases we should never reach this piece of code. 4040 */ 4041 printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n", 4042 __func__, sk->sk_state); 4043 break; 4044 } 4045 4046 /* It _is_ possible, that we have something out-of-order _after_ FIN. 4047 * Probably, we should reset in this case. For now drop them. 4048 */ 4049 __skb_queue_purge(&tp->out_of_order_queue); 4050 if (tcp_is_sack(tp)) 4051 tcp_sack_reset(&tp->rx_opt); 4052 sk_mem_reclaim(sk); 4053 4054 if (!sock_flag(sk, SOCK_DEAD)) { 4055 sk->sk_state_change(sk); 4056 4057 /* Do not send POLL_HUP for half duplex close. */ 4058 if (sk->sk_shutdown == SHUTDOWN_MASK || 4059 sk->sk_state == TCP_CLOSE) 4060 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP); 4061 else 4062 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN); 4063 } 4064} 4065 4066static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, 4067 u32 end_seq) 4068{ 4069 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) { 4070 if (before(seq, sp->start_seq)) 4071 sp->start_seq = seq; 4072 if (after(end_seq, sp->end_seq)) 4073 sp->end_seq = end_seq; 4074 return 1; 4075 } 4076 return 0; 4077} 4078 4079static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq) 4080{ 4081 struct tcp_sock *tp = tcp_sk(sk); 4082 4083 if (tcp_is_sack(tp) && sysctl_tcp_dsack) { 4084 int mib_idx; 4085 4086 if (before(seq, tp->rcv_nxt)) 4087 mib_idx = LINUX_MIB_TCPDSACKOLDSENT; 4088 else 4089 mib_idx = LINUX_MIB_TCPDSACKOFOSENT; 4090 4091 NET_INC_STATS_BH(sock_net(sk), mib_idx); 4092 4093 tp->rx_opt.dsack = 1; 4094 tp->duplicate_sack[0].start_seq = seq; 4095 tp->duplicate_sack[0].end_seq = end_seq; 4096 } 4097} 4098 4099static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq) 4100{ 4101 struct tcp_sock *tp = tcp_sk(sk); 4102 4103 if (!tp->rx_opt.dsack) 4104 tcp_dsack_set(sk, seq, end_seq); 4105 else 4106 tcp_sack_extend(tp->duplicate_sack, seq, end_seq); 4107} 4108 4109static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb) 4110{ 4111 struct tcp_sock *tp = tcp_sk(sk); 4112 4113 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 4114 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 4115 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST); 4116 tcp_enter_quickack_mode(sk); 4117 4118 if (tcp_is_sack(tp) && sysctl_tcp_dsack) { 4119 u32 end_seq = TCP_SKB_CB(skb)->end_seq; 4120 4121 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) 4122 end_seq = tp->rcv_nxt; 4123 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq); 4124 } 4125 } 4126 4127 tcp_send_ack(sk); 4128} 4129 4130/* These routines update the SACK block as out-of-order packets arrive or 4131 * in-order packets close up the sequence space. 4132 */ 4133static void tcp_sack_maybe_coalesce(struct tcp_sock *tp) 4134{ 4135 int this_sack; 4136 struct tcp_sack_block *sp = &tp->selective_acks[0]; 4137 struct tcp_sack_block *swalk = sp + 1; 4138 4139 /* See if the recent change to the first SACK eats into 4140 * or hits the sequence space of other SACK blocks, if so coalesce. 4141 */ 4142 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) { 4143 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) { 4144 int i; 4145 4146 /* Zap SWALK, by moving every further SACK up by one slot. 4147 * Decrease num_sacks. 4148 */ 4149 tp->rx_opt.num_sacks--; 4150 for (i = this_sack; i < tp->rx_opt.num_sacks; i++) 4151 sp[i] = sp[i + 1]; 4152 continue; 4153 } 4154 this_sack++, swalk++; 4155 } 4156} 4157 4158static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq) 4159{ 4160 struct tcp_sock *tp = tcp_sk(sk); 4161 struct tcp_sack_block *sp = &tp->selective_acks[0]; 4162 int cur_sacks = tp->rx_opt.num_sacks; 4163 int this_sack; 4164 4165 if (!cur_sacks) 4166 goto new_sack; 4167 4168 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) { 4169 if (tcp_sack_extend(sp, seq, end_seq)) { 4170 /* Rotate this_sack to the first one. */ 4171 for (; this_sack > 0; this_sack--, sp--) 4172 swap(*sp, *(sp - 1)); 4173 if (cur_sacks > 1) 4174 tcp_sack_maybe_coalesce(tp); 4175 return; 4176 } 4177 } 4178 4179 /* Could not find an adjacent existing SACK, build a new one, 4180 * put it at the front, and shift everyone else down. We 4181 * always know there is at least one SACK present already here. 4182 * 4183 * If the sack array is full, forget about the last one. 4184 */ 4185 if (this_sack >= TCP_NUM_SACKS) { 4186 this_sack--; 4187 tp->rx_opt.num_sacks--; 4188 sp--; 4189 } 4190 for (; this_sack > 0; this_sack--, sp--) 4191 *sp = *(sp - 1); 4192 4193new_sack: 4194 /* Build the new head SACK, and we're done. */ 4195 sp->start_seq = seq; 4196 sp->end_seq = end_seq; 4197 tp->rx_opt.num_sacks++; 4198} 4199 4200/* RCV.NXT advances, some SACKs should be eaten. */ 4201 4202static void tcp_sack_remove(struct tcp_sock *tp) 4203{ 4204 struct tcp_sack_block *sp = &tp->selective_acks[0]; 4205 int num_sacks = tp->rx_opt.num_sacks; 4206 int this_sack; 4207 4208 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */ 4209 if (skb_queue_empty(&tp->out_of_order_queue)) { 4210 tp->rx_opt.num_sacks = 0; 4211 return; 4212 } 4213 4214 for (this_sack = 0; this_sack < num_sacks;) { 4215 /* Check if the start of the sack is covered by RCV.NXT. */ 4216 if (!before(tp->rcv_nxt, sp->start_seq)) { 4217 int i; 4218 4219 /* RCV.NXT must cover all the block! */ 4220 WARN_ON(before(tp->rcv_nxt, sp->end_seq)); 4221 4222 /* Zap this SACK, by moving forward any other SACKS. */ 4223 for (i=this_sack+1; i < num_sacks; i++) 4224 tp->selective_acks[i-1] = tp->selective_acks[i]; 4225 num_sacks--; 4226 continue; 4227 } 4228 this_sack++; 4229 sp++; 4230 } 4231 tp->rx_opt.num_sacks = num_sacks; 4232} 4233 4234/* This one checks to see if we can put data from the 4235 * out_of_order queue into the receive_queue. 4236 */ 4237static void tcp_ofo_queue(struct sock *sk) 4238{ 4239 struct tcp_sock *tp = tcp_sk(sk); 4240 __u32 dsack_high = tp->rcv_nxt; 4241 struct sk_buff *skb; 4242 4243 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) { 4244 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) 4245 break; 4246 4247 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) { 4248 __u32 dsack = dsack_high; 4249 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high)) 4250 dsack_high = TCP_SKB_CB(skb)->end_seq; 4251 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack); 4252 } 4253 4254 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) { 4255 SOCK_DEBUG(sk, "ofo packet was already received \n"); 4256 __skb_unlink(skb, &tp->out_of_order_queue); 4257 __kfree_skb(skb); 4258 continue; 4259 } 4260 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n", 4261 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, 4262 TCP_SKB_CB(skb)->end_seq); 4263 4264 __skb_unlink(skb, &tp->out_of_order_queue); 4265 __skb_queue_tail(&sk->sk_receive_queue, skb); 4266 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; 4267 if (tcp_hdr(skb)->fin) 4268 tcp_fin(skb, sk, tcp_hdr(skb)); 4269 } 4270} 4271 4272static int tcp_prune_ofo_queue(struct sock *sk); 4273static int tcp_prune_queue(struct sock *sk); 4274 4275static inline int tcp_try_rmem_schedule(struct sock *sk, unsigned int size) 4276{ 4277 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf || 4278 !sk_rmem_schedule(sk, size)) { 4279 4280 if (tcp_prune_queue(sk) < 0) 4281 return -1; 4282 4283 if (!sk_rmem_schedule(sk, size)) { 4284 if (!tcp_prune_ofo_queue(sk)) 4285 return -1; 4286 4287 if (!sk_rmem_schedule(sk, size)) 4288 return -1; 4289 } 4290 } 4291 return 0; 4292} 4293 4294static void tcp_data_queue(struct sock *sk, struct sk_buff *skb) 4295{ 4296 struct tcphdr *th = tcp_hdr(skb); 4297 struct tcp_sock *tp = tcp_sk(sk); 4298 int eaten = -1; 4299 4300 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) 4301 goto drop; 4302 4303 __skb_pull(skb, th->doff * 4); 4304 4305 TCP_ECN_accept_cwr(tp, skb); 4306 4307 tp->rx_opt.dsack = 0; 4308 4309 /* Queue data for delivery to the user. 4310 * Packets in sequence go to the receive queue. 4311 * Out of sequence packets to the out_of_order_queue. 4312 */ 4313 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) { 4314 if (tcp_receive_window(tp) == 0) 4315 goto out_of_window; 4316 4317 /* Ok. In sequence. In window. */ 4318 if (tp->ucopy.task == current && 4319 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len && 4320 sock_owned_by_user(sk) && !tp->urg_data) { 4321 int chunk = min_t(unsigned int, skb->len, 4322 tp->ucopy.len); 4323 4324 __set_current_state(TASK_RUNNING); 4325 4326 local_bh_enable(); 4327 if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) { 4328 tp->ucopy.len -= chunk; 4329 tp->copied_seq += chunk; 4330 eaten = (chunk == skb->len && !th->fin); 4331 tcp_rcv_space_adjust(sk); 4332 } 4333 local_bh_disable(); 4334 } 4335 4336 if (eaten <= 0) { 4337queue_and_out: 4338 if (eaten < 0 && 4339 tcp_try_rmem_schedule(sk, skb->truesize)) 4340 goto drop; 4341 4342 skb_set_owner_r(skb, sk); 4343 __skb_queue_tail(&sk->sk_receive_queue, skb); 4344 } 4345 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; 4346 if (skb->len) 4347 tcp_event_data_recv(sk, skb); 4348 if (th->fin) 4349 tcp_fin(skb, sk, th); 4350 4351 if (!skb_queue_empty(&tp->out_of_order_queue)) { 4352 tcp_ofo_queue(sk); 4353 4354 /* RFC2581. 4.2. SHOULD send immediate ACK, when 4355 * gap in queue is filled. 4356 */ 4357 if (skb_queue_empty(&tp->out_of_order_queue)) 4358 inet_csk(sk)->icsk_ack.pingpong = 0; 4359 } 4360 4361 if (tp->rx_opt.num_sacks) 4362 tcp_sack_remove(tp); 4363 4364 tcp_fast_path_check(sk); 4365 4366 if (eaten > 0) 4367 __kfree_skb(skb); 4368 else if (!sock_flag(sk, SOCK_DEAD)) 4369 sk->sk_data_ready(sk, 0); 4370 return; 4371 } 4372 4373 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) { 4374 /* A retransmit, 2nd most common case. Force an immediate ack. */ 4375 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST); 4376 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); 4377 4378out_of_window: 4379 tcp_enter_quickack_mode(sk); 4380 inet_csk_schedule_ack(sk); 4381drop: 4382 __kfree_skb(skb); 4383 return; 4384 } 4385 4386 /* Out of window. F.e. zero window probe. */ 4387 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp))) 4388 goto out_of_window; 4389 4390 tcp_enter_quickack_mode(sk); 4391 4392 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 4393 /* Partial packet, seq < rcv_next < end_seq */ 4394 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n", 4395 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, 4396 TCP_SKB_CB(skb)->end_seq); 4397 4398 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt); 4399 4400 /* If window is closed, drop tail of packet. But after 4401 * remembering D-SACK for its head made in previous line. 4402 */ 4403 if (!tcp_receive_window(tp)) 4404 goto out_of_window; 4405 goto queue_and_out; 4406 } 4407 4408 TCP_ECN_check_ce(tp, skb); 4409 4410 if (tcp_try_rmem_schedule(sk, skb->truesize)) 4411 goto drop; 4412 4413 /* Disable header prediction. */ 4414 tp->pred_flags = 0; 4415 inet_csk_schedule_ack(sk); 4416 4417 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n", 4418 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); 4419 4420 skb_set_owner_r(skb, sk); 4421 4422 if (!skb_peek(&tp->out_of_order_queue)) { 4423 /* Initial out of order segment, build 1 SACK. */ 4424 if (tcp_is_sack(tp)) { 4425 tp->rx_opt.num_sacks = 1; 4426 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq; 4427 tp->selective_acks[0].end_seq = 4428 TCP_SKB_CB(skb)->end_seq; 4429 } 4430 __skb_queue_head(&tp->out_of_order_queue, skb); 4431 } else { 4432 struct sk_buff *skb1 = skb_peek_tail(&tp->out_of_order_queue); 4433 u32 seq = TCP_SKB_CB(skb)->seq; 4434 u32 end_seq = TCP_SKB_CB(skb)->end_seq; 4435 4436 if (seq == TCP_SKB_CB(skb1)->end_seq) { 4437 __skb_queue_after(&tp->out_of_order_queue, skb1, skb); 4438 4439 if (!tp->rx_opt.num_sacks || 4440 tp->selective_acks[0].end_seq != seq) 4441 goto add_sack; 4442 4443 /* Common case: data arrive in order after hole. */ 4444 tp->selective_acks[0].end_seq = end_seq; 4445 return; 4446 } 4447 4448 /* Find place to insert this segment. */ 4449 while (1) { 4450 if (!after(TCP_SKB_CB(skb1)->seq, seq)) 4451 break; 4452 if (skb_queue_is_first(&tp->out_of_order_queue, skb1)) { 4453 skb1 = NULL; 4454 break; 4455 } 4456 skb1 = skb_queue_prev(&tp->out_of_order_queue, skb1); 4457 } 4458 4459 /* Do skb overlap to previous one? */ 4460 if (skb1 && before(seq, TCP_SKB_CB(skb1)->end_seq)) { 4461 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) { 4462 /* All the bits are present. Drop. */ 4463 __kfree_skb(skb); 4464 tcp_dsack_set(sk, seq, end_seq); 4465 goto add_sack; 4466 } 4467 if (after(seq, TCP_SKB_CB(skb1)->seq)) { 4468 /* Partial overlap. */ 4469 tcp_dsack_set(sk, seq, 4470 TCP_SKB_CB(skb1)->end_seq); 4471 } else { 4472 if (skb_queue_is_first(&tp->out_of_order_queue, 4473 skb1)) 4474 skb1 = NULL; 4475 else 4476 skb1 = skb_queue_prev( 4477 &tp->out_of_order_queue, 4478 skb1); 4479 } 4480 } 4481 if (!skb1) 4482 __skb_queue_head(&tp->out_of_order_queue, skb); 4483 else 4484 __skb_queue_after(&tp->out_of_order_queue, skb1, skb); 4485 4486 /* And clean segments covered by new one as whole. */ 4487 while (!skb_queue_is_last(&tp->out_of_order_queue, skb)) { 4488 skb1 = skb_queue_next(&tp->out_of_order_queue, skb); 4489 4490 if (!after(end_seq, TCP_SKB_CB(skb1)->seq)) 4491 break; 4492 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) { 4493 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq, 4494 end_seq); 4495 break; 4496 } 4497 __skb_unlink(skb1, &tp->out_of_order_queue); 4498 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq, 4499 TCP_SKB_CB(skb1)->end_seq); 4500 __kfree_skb(skb1); 4501 } 4502 4503add_sack: 4504 if (tcp_is_sack(tp)) 4505 tcp_sack_new_ofo_skb(sk, seq, end_seq); 4506 } 4507} 4508 4509static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb, 4510 struct sk_buff_head *list) 4511{ 4512 struct sk_buff *next = NULL; 4513 4514 if (!skb_queue_is_last(list, skb)) 4515 next = skb_queue_next(list, skb); 4516 4517 __skb_unlink(skb, list); 4518 __kfree_skb(skb); 4519 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED); 4520 4521 return next; 4522} 4523 4524/* Collapse contiguous sequence of skbs head..tail with 4525 * sequence numbers start..end. 4526 * 4527 * If tail is NULL, this means until the end of the list. 4528 * 4529 * Segments with FIN/SYN are not collapsed (only because this 4530 * simplifies code) 4531 */ 4532static void 4533tcp_collapse(struct sock *sk, struct sk_buff_head *list, 4534 struct sk_buff *head, struct sk_buff *tail, 4535 u32 start, u32 end) 4536{ 4537 struct sk_buff *skb, *n; 4538 bool end_of_skbs; 4539 4540 /* First, check that queue is collapsible and find 4541 * the point where collapsing can be useful. */ 4542 skb = head; 4543restart: 4544 end_of_skbs = true; 4545 skb_queue_walk_from_safe(list, skb, n) { 4546 if (skb == tail) 4547 break; 4548 /* No new bits? It is possible on ofo queue. */ 4549 if (!before(start, TCP_SKB_CB(skb)->end_seq)) { 4550 skb = tcp_collapse_one(sk, skb, list); 4551 if (!skb) 4552 break; 4553 goto restart; 4554 } 4555 4556 /* The first skb to collapse is: 4557 * - not SYN/FIN and 4558 * - bloated or contains data before "start" or 4559 * overlaps to the next one. 4560 */ 4561 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin && 4562 (tcp_win_from_space(skb->truesize) > skb->len || 4563 before(TCP_SKB_CB(skb)->seq, start))) { 4564 end_of_skbs = false; 4565 break; 4566 } 4567 4568 if (!skb_queue_is_last(list, skb)) { 4569 struct sk_buff *next = skb_queue_next(list, skb); 4570 if (next != tail && 4571 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(next)->seq) { 4572 end_of_skbs = false; 4573 break; 4574 } 4575 } 4576 4577 /* Decided to skip this, advance start seq. */ 4578 start = TCP_SKB_CB(skb)->end_seq; 4579 } 4580 if (end_of_skbs || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin) 4581 return; 4582 4583 while (before(start, end)) { 4584 struct sk_buff *nskb; 4585 unsigned int header = skb_headroom(skb); 4586 int copy = SKB_MAX_ORDER(header, 0); 4587 4588 /* Too big header? This can happen with IPv6. */ 4589 if (copy < 0) 4590 return; 4591 if (end - start < copy) 4592 copy = end - start; 4593 nskb = alloc_skb(copy + header, GFP_ATOMIC); 4594 if (!nskb) 4595 return; 4596 4597 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head); 4598 skb_set_network_header(nskb, (skb_network_header(skb) - 4599 skb->head)); 4600 skb_set_transport_header(nskb, (skb_transport_header(skb) - 4601 skb->head)); 4602 skb_reserve(nskb, header); 4603 memcpy(nskb->head, skb->head, header); 4604 memcpy(nskb->cb, skb->cb, sizeof(skb->cb)); 4605 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start; 4606 __skb_queue_before(list, skb, nskb); 4607 skb_set_owner_r(nskb, sk); 4608 4609 /* Copy data, releasing collapsed skbs. */ 4610 while (copy > 0) { 4611 int offset = start - TCP_SKB_CB(skb)->seq; 4612 int size = TCP_SKB_CB(skb)->end_seq - start; 4613 4614 BUG_ON(offset < 0); 4615 if (size > 0) { 4616 size = min(copy, size); 4617 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size)) 4618 BUG(); 4619 TCP_SKB_CB(nskb)->end_seq += size; 4620 copy -= size; 4621 start += size; 4622 } 4623 if (!before(start, TCP_SKB_CB(skb)->end_seq)) { 4624 skb = tcp_collapse_one(sk, skb, list); 4625 if (!skb || 4626 skb == tail || 4627 tcp_hdr(skb)->syn || 4628 tcp_hdr(skb)->fin) 4629 return; 4630 } 4631 } 4632 } 4633} 4634 4635/* Collapse ofo queue. Algorithm: select contiguous sequence of skbs 4636 * and tcp_collapse() them until all the queue is collapsed. 4637 */ 4638static void tcp_collapse_ofo_queue(struct sock *sk) 4639{ 4640 struct tcp_sock *tp = tcp_sk(sk); 4641 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue); 4642 struct sk_buff *head; 4643 u32 start, end; 4644 4645 if (skb == NULL) 4646 return; 4647 4648 start = TCP_SKB_CB(skb)->seq; 4649 end = TCP_SKB_CB(skb)->end_seq; 4650 head = skb; 4651 4652 for (;;) { 4653 struct sk_buff *next = NULL; 4654 4655 if (!skb_queue_is_last(&tp->out_of_order_queue, skb)) 4656 next = skb_queue_next(&tp->out_of_order_queue, skb); 4657 skb = next; 4658 4659 /* Segment is terminated when we see gap or when 4660 * we are at the end of all the queue. */ 4661 if (!skb || 4662 after(TCP_SKB_CB(skb)->seq, end) || 4663 before(TCP_SKB_CB(skb)->end_seq, start)) { 4664 tcp_collapse(sk, &tp->out_of_order_queue, 4665 head, skb, start, end); 4666 head = skb; 4667 if (!skb) 4668 break; 4669 /* Start new segment */ 4670 start = TCP_SKB_CB(skb)->seq; 4671 end = TCP_SKB_CB(skb)->end_seq; 4672 } else { 4673 if (before(TCP_SKB_CB(skb)->seq, start)) 4674 start = TCP_SKB_CB(skb)->seq; 4675 if (after(TCP_SKB_CB(skb)->end_seq, end)) 4676 end = TCP_SKB_CB(skb)->end_seq; 4677 } 4678 } 4679} 4680 4681/* 4682 * Purge the out-of-order queue. 4683 * Return true if queue was pruned. 4684 */ 4685static int tcp_prune_ofo_queue(struct sock *sk) 4686{ 4687 struct tcp_sock *tp = tcp_sk(sk); 4688 int res = 0; 4689 4690 if (!skb_queue_empty(&tp->out_of_order_queue)) { 4691 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED); 4692 __skb_queue_purge(&tp->out_of_order_queue); 4693 4694 /* Reset SACK state. A conforming SACK implementation will 4695 * do the same at a timeout based retransmit. When a connection 4696 * is in a sad state like this, we care only about integrity 4697 * of the connection not performance. 4698 */ 4699 if (tp->rx_opt.sack_ok) 4700 tcp_sack_reset(&tp->rx_opt); 4701 sk_mem_reclaim(sk); 4702 res = 1; 4703 } 4704 return res; 4705} 4706 4707/* Reduce allocated memory if we can, trying to get 4708 * the socket within its memory limits again. 4709 * 4710 * Return less than zero if we should start dropping frames 4711 * until the socket owning process reads some of the data 4712 * to stabilize the situation. 4713 */ 4714static int tcp_prune_queue(struct sock *sk) 4715{ 4716 struct tcp_sock *tp = tcp_sk(sk); 4717 4718 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq); 4719 4720 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED); 4721 4722 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf) 4723 tcp_clamp_window(sk); 4724 else if (tcp_memory_pressure) 4725 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss); 4726 4727 tcp_collapse_ofo_queue(sk); 4728 if (!skb_queue_empty(&sk->sk_receive_queue)) 4729 tcp_collapse(sk, &sk->sk_receive_queue, 4730 skb_peek(&sk->sk_receive_queue), 4731 NULL, 4732 tp->copied_seq, tp->rcv_nxt); 4733 sk_mem_reclaim(sk); 4734 4735 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) 4736 return 0; 4737 4738 /* Collapsing did not help, destructive actions follow. 4739 * This must not ever occur. */ 4740 4741 tcp_prune_ofo_queue(sk); 4742 4743 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) 4744 return 0; 4745 4746 /* If we are really being abused, tell the caller to silently 4747 * drop receive data on the floor. It will get retransmitted 4748 * and hopefully then we'll have sufficient space. 4749 */ 4750 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED); 4751 4752 /* Massive buffer overcommit. */ 4753 tp->pred_flags = 0; 4754 return -1; 4755} 4756 4757/* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. 4758 * As additional protections, we do not touch cwnd in retransmission phases, 4759 * and if application hit its sndbuf limit recently. 4760 */ 4761void tcp_cwnd_application_limited(struct sock *sk) 4762{ 4763 struct tcp_sock *tp = tcp_sk(sk); 4764 4765 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && 4766 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { 4767 /* Limited by application or receiver window. */ 4768 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); 4769 u32 win_used = max(tp->snd_cwnd_used, init_win); 4770 if (win_used < tp->snd_cwnd) { 4771 tp->snd_ssthresh = tcp_current_ssthresh(sk); 4772 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1; 4773 } 4774 tp->snd_cwnd_used = 0; 4775 } 4776 tp->snd_cwnd_stamp = tcp_time_stamp; 4777} 4778 4779static int tcp_should_expand_sndbuf(struct sock *sk) 4780{ 4781 struct tcp_sock *tp = tcp_sk(sk); 4782 4783 /* If the user specified a specific send buffer setting, do 4784 * not modify it. 4785 */ 4786 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK) 4787 return 0; 4788 4789 /* If we are under global TCP memory pressure, do not expand. */ 4790 if (tcp_memory_pressure) 4791 return 0; 4792 4793 /* If we are under soft global TCP memory pressure, do not expand. */ 4794 if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0]) 4795 return 0; 4796 4797 /* If we filled the congestion window, do not expand. */ 4798 if (tp->packets_out >= tp->snd_cwnd) 4799 return 0; 4800 4801 return 1; 4802} 4803 4804/* When incoming ACK allowed to free some skb from write_queue, 4805 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket 4806 * on the exit from tcp input handler. 4807 * 4808 * PROBLEM: sndbuf expansion does not work well with largesend. 4809 */ 4810static void tcp_new_space(struct sock *sk) 4811{ 4812 struct tcp_sock *tp = tcp_sk(sk); 4813 4814 if (tcp_should_expand_sndbuf(sk)) { 4815 int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) + 4816 MAX_TCP_HEADER + 16 + sizeof(struct sk_buff); 4817 int demanded = max_t(unsigned int, tp->snd_cwnd, 4818 tp->reordering + 1); 4819 sndmem *= 2 * demanded; 4820 if (sndmem > sk->sk_sndbuf) 4821 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]); 4822 tp->snd_cwnd_stamp = tcp_time_stamp; 4823 } 4824 4825 sk->sk_write_space(sk); 4826} 4827 4828static void tcp_check_space(struct sock *sk) 4829{ 4830 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) { 4831 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK); 4832 if (sk->sk_socket && 4833 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) 4834 tcp_new_space(sk); 4835 } 4836} 4837 4838static inline void tcp_data_snd_check(struct sock *sk) 4839{ 4840 tcp_push_pending_frames(sk); 4841 tcp_check_space(sk); 4842} 4843 4844/* 4845 * Check if sending an ack is needed. 4846 */ 4847static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible) 4848{ 4849 struct tcp_sock *tp = tcp_sk(sk); 4850 4851 /* More than one full frame received... */ 4852 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss 4853 /* ... and right edge of window advances far enough. 4854 * (tcp_recvmsg() will send ACK otherwise). Or... 4855 */ 4856 && __tcp_select_window(sk) >= tp->rcv_wnd) || 4857 /* We ACK each frame or... */ 4858 tcp_in_quickack_mode(sk) || 4859 /* We have out of order data. */ 4860 (ofo_possible && skb_peek(&tp->out_of_order_queue))) { 4861 /* Then ack it now */ 4862 tcp_send_ack(sk); 4863 } else { 4864 /* Else, send delayed ack. */ 4865 tcp_send_delayed_ack(sk); 4866 } 4867} 4868 4869static inline void tcp_ack_snd_check(struct sock *sk) 4870{ 4871 if (!inet_csk_ack_scheduled(sk)) { 4872 /* We sent a data segment already. */ 4873 return; 4874 } 4875 __tcp_ack_snd_check(sk, 1); 4876} 4877 4878/* 4879 * This routine is only called when we have urgent data 4880 * signaled. Its the 'slow' part of tcp_urg. It could be 4881 * moved inline now as tcp_urg is only called from one 4882 * place. We handle URGent data wrong. We have to - as 4883 * BSD still doesn't use the correction from RFC961. 4884 * For 1003.1g we should support a new option TCP_STDURG to permit 4885 * either form (or just set the sysctl tcp_stdurg). 4886 */ 4887 4888static void tcp_check_urg(struct sock *sk, struct tcphdr *th) 4889{ 4890 struct tcp_sock *tp = tcp_sk(sk); 4891 u32 ptr = ntohs(th->urg_ptr); 4892 4893 if (ptr && !sysctl_tcp_stdurg) 4894 ptr--; 4895 ptr += ntohl(th->seq); 4896 4897 /* Ignore urgent data that we've already seen and read. */ 4898 if (after(tp->copied_seq, ptr)) 4899 return; 4900 4901 /* Do not replay urg ptr. 4902 * 4903 * NOTE: interesting situation not covered by specs. 4904 * Misbehaving sender may send urg ptr, pointing to segment, 4905 * which we already have in ofo queue. We are not able to fetch 4906 * such data and will stay in TCP_URG_NOTYET until will be eaten 4907 * by recvmsg(). Seems, we are not obliged to handle such wicked 4908 * situations. But it is worth to think about possibility of some 4909 * DoSes using some hypothetical application level deadlock. 4910 */ 4911 if (before(ptr, tp->rcv_nxt)) 4912 return; 4913 4914 /* Do we already have a newer (or duplicate) urgent pointer? */ 4915 if (tp->urg_data && !after(ptr, tp->urg_seq)) 4916 return; 4917 4918 /* Tell the world about our new urgent pointer. */ 4919 sk_send_sigurg(sk); 4920 4921 /* We may be adding urgent data when the last byte read was 4922 * urgent. To do this requires some care. We cannot just ignore 4923 * tp->copied_seq since we would read the last urgent byte again 4924 * as data, nor can we alter copied_seq until this data arrives 4925 * or we break the semantics of SIOCATMARK (and thus sockatmark()) 4926 * 4927 * NOTE. Double Dutch. Rendering to plain English: author of comment 4928 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB); 4929 * and expect that both A and B disappear from stream. This is _wrong_. 4930 * Though this happens in BSD with high probability, this is occasional. 4931 * Any application relying on this is buggy. Note also, that fix "works" 4932 * only in this artificial test. Insert some normal data between A and B and we will 4933 * decline of BSD again. Verdict: it is better to remove to trap 4934 * buggy users. 4935 */ 4936 if (tp->urg_seq == tp->copied_seq && tp->urg_data && 4937 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) { 4938 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue); 4939 tp->copied_seq++; 4940 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) { 4941 __skb_unlink(skb, &sk->sk_receive_queue); 4942 __kfree_skb(skb); 4943 } 4944 } 4945 4946 tp->urg_data = TCP_URG_NOTYET; 4947 tp->urg_seq = ptr; 4948 4949 /* Disable header prediction. */ 4950 tp->pred_flags = 0; 4951} 4952 4953/* This is the 'fast' part of urgent handling. */ 4954static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th) 4955{ 4956 struct tcp_sock *tp = tcp_sk(sk); 4957 4958 /* Check if we get a new urgent pointer - normally not. */ 4959 if (th->urg) 4960 tcp_check_urg(sk, th); 4961 4962 /* Do we wait for any urgent data? - normally not... */ 4963 if (tp->urg_data == TCP_URG_NOTYET) { 4964 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) - 4965 th->syn; 4966 4967 /* Is the urgent pointer pointing into this packet? */ 4968 if (ptr < skb->len) { 4969 u8 tmp; 4970 if (skb_copy_bits(skb, ptr, &tmp, 1)) 4971 BUG(); 4972 tp->urg_data = TCP_URG_VALID | tmp; 4973 if (!sock_flag(sk, SOCK_DEAD)) 4974 sk->sk_data_ready(sk, 0); 4975 } 4976 } 4977} 4978 4979static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen) 4980{ 4981 struct tcp_sock *tp = tcp_sk(sk); 4982 int chunk = skb->len - hlen; 4983 int err; 4984 4985 local_bh_enable(); 4986 if (skb_csum_unnecessary(skb)) 4987 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk); 4988 else 4989 err = skb_copy_and_csum_datagram_iovec(skb, hlen, 4990 tp->ucopy.iov); 4991 4992 if (!err) { 4993 tp->ucopy.len -= chunk; 4994 tp->copied_seq += chunk; 4995 tcp_rcv_space_adjust(sk); 4996 } 4997 4998 local_bh_disable(); 4999 return err; 5000} 5001 5002static __sum16 __tcp_checksum_complete_user(struct sock *sk, 5003 struct sk_buff *skb) 5004{ 5005 __sum16 result; 5006 5007 if (sock_owned_by_user(sk)) { 5008 local_bh_enable(); 5009 result = __tcp_checksum_complete(skb); 5010 local_bh_disable(); 5011 } else { 5012 result = __tcp_checksum_complete(skb); 5013 } 5014 return result; 5015} 5016 5017static inline int tcp_checksum_complete_user(struct sock *sk, 5018 struct sk_buff *skb) 5019{ 5020 return !skb_csum_unnecessary(skb) && 5021 __tcp_checksum_complete_user(sk, skb); 5022} 5023 5024#ifdef CONFIG_NET_DMA 5025static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb, 5026 int hlen) 5027{ 5028 struct tcp_sock *tp = tcp_sk(sk); 5029 int chunk = skb->len - hlen; 5030 int dma_cookie; 5031 int copied_early = 0; 5032 5033 if (tp->ucopy.wakeup) 5034 return 0; 5035 5036 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list) 5037 tp->ucopy.dma_chan = dma_find_channel(DMA_MEMCPY); 5038 5039 if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) { 5040 5041 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan, 5042 skb, hlen, 5043 tp->ucopy.iov, chunk, 5044 tp->ucopy.pinned_list); 5045 5046 if (dma_cookie < 0) 5047 goto out; 5048 5049 tp->ucopy.dma_cookie = dma_cookie; 5050 copied_early = 1; 5051 5052 tp->ucopy.len -= chunk; 5053 tp->copied_seq += chunk; 5054 tcp_rcv_space_adjust(sk); 5055 5056 if ((tp->ucopy.len == 0) || 5057 (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) || 5058 (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) { 5059 tp->ucopy.wakeup = 1; 5060 sk->sk_data_ready(sk, 0); 5061 } 5062 } else if (chunk > 0) { 5063 tp->ucopy.wakeup = 1; 5064 sk->sk_data_ready(sk, 0); 5065 } 5066out: 5067 return copied_early; 5068} 5069#endif /* CONFIG_NET_DMA */ 5070 5071/* Does PAWS and seqno based validation of an incoming segment, flags will 5072 * play significant role here. 5073 */ 5074static int tcp_validate_incoming(struct sock *sk, struct sk_buff *skb, 5075 struct tcphdr *th, int syn_inerr) 5076{ 5077 struct tcp_sock *tp = tcp_sk(sk); 5078 5079 /* RFC1323: H1. Apply PAWS check first. */ 5080 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp && 5081 tcp_paws_discard(sk, skb)) { 5082 if (!th->rst) { 5083 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED); 5084 tcp_send_dupack(sk, skb); 5085 goto discard; 5086 } 5087 /* Reset is accepted even if it did not pass PAWS. */ 5088 } 5089 5090 /* Step 1: check sequence number */ 5091 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) { 5092 /* RFC793, page 37: "In all states except SYN-SENT, all reset 5093 * (RST) segments are validated by checking their SEQ-fields." 5094 * And page 69: "If an incoming segment is not acceptable, 5095 * an acknowledgment should be sent in reply (unless the RST 5096 * bit is set, if so drop the segment and return)". 5097 */ 5098 if (!th->rst) 5099 tcp_send_dupack(sk, skb); 5100 goto discard; 5101 } 5102 5103 /* Step 2: check RST bit */ 5104 if (th->rst) { 5105 tcp_reset(sk); 5106 goto discard; 5107 } 5108 5109 /* ts_recent update must be made after we are sure that the packet 5110 * is in window. 5111 */ 5112 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq); 5113 5114 /* step 3: check security and precedence [ignored] */ 5115 5116 /* step 4: Check for a SYN in window. */ 5117 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 5118 if (syn_inerr) 5119 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS); 5120 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONSYN); 5121 tcp_reset(sk); 5122 return -1; 5123 } 5124 5125 return 1; 5126 5127discard: 5128 __kfree_skb(skb); 5129 return 0; 5130} 5131 5132/* 5133 * TCP receive function for the ESTABLISHED state. 5134 * 5135 * It is split into a fast path and a slow path. The fast path is 5136 * disabled when: 5137 * - A zero window was announced from us - zero window probing 5138 * is only handled properly in the slow path. 5139 * - Out of order segments arrived. 5140 * - Urgent data is expected. 5141 * - There is no buffer space left 5142 * - Unexpected TCP flags/window values/header lengths are received 5143 * (detected by checking the TCP header against pred_flags) 5144 * - Data is sent in both directions. Fast path only supports pure senders 5145 * or pure receivers (this means either the sequence number or the ack 5146 * value must stay constant) 5147 * - Unexpected TCP option. 5148 * 5149 * When these conditions are not satisfied it drops into a standard 5150 * receive procedure patterned after RFC793 to handle all cases. 5151 * The first three cases are guaranteed by proper pred_flags setting, 5152 * the rest is checked inline. Fast processing is turned on in 5153 * tcp_data_queue when everything is OK. 5154 */ 5155int tcp_rcv_established(struct sock *sk, struct sk_buff *skb, 5156 struct tcphdr *th, unsigned len) 5157{ 5158 struct tcp_sock *tp = tcp_sk(sk); 5159 int res; 5160 5161 /* 5162 * Header prediction. 5163 * The code loosely follows the one in the famous 5164 * "30 instruction TCP receive" Van Jacobson mail. 5165 * 5166 * Van's trick is to deposit buffers into socket queue 5167 * on a device interrupt, to call tcp_recv function 5168 * on the receive process context and checksum and copy 5169 * the buffer to user space. smart... 5170 * 5171 * Our current scheme is not silly either but we take the 5172 * extra cost of the net_bh soft interrupt processing... 5173 * We do checksum and copy also but from device to kernel. 5174 */ 5175 5176 tp->rx_opt.saw_tstamp = 0; 5177 5178 /* pred_flags is 0xS?10 << 16 + snd_wnd 5179 * if header_prediction is to be made 5180 * 'S' will always be tp->tcp_header_len >> 2 5181 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to 5182 * turn it off (when there are holes in the receive 5183 * space for instance) 5184 * PSH flag is ignored. 5185 */ 5186 5187 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags && 5188 TCP_SKB_CB(skb)->seq == tp->rcv_nxt && 5189 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) { 5190 int tcp_header_len = tp->tcp_header_len; 5191 5192 /* Timestamp header prediction: tcp_header_len 5193 * is automatically equal to th->doff*4 due to pred_flags 5194 * match. 5195 */ 5196 5197 /* Check timestamp */ 5198 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) { 5199 /* No? Slow path! */ 5200 if (!tcp_parse_aligned_timestamp(tp, th)) 5201 goto slow_path; 5202 5203 /* If PAWS failed, check it more carefully in slow path */ 5204 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0) 5205 goto slow_path; 5206 5207 /* DO NOT update ts_recent here, if checksum fails 5208 * and timestamp was corrupted part, it will result 5209 * in a hung connection since we will drop all 5210 * future packets due to the PAWS test. 5211 */ 5212 } 5213 5214 if (len <= tcp_header_len) { 5215 /* Bulk data transfer: sender */ 5216 if (len == tcp_header_len) { 5217 /* Predicted packet is in window by definition. 5218 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 5219 * Hence, check seq<=rcv_wup reduces to: 5220 */ 5221 if (tcp_header_len == 5222 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && 5223 tp->rcv_nxt == tp->rcv_wup) 5224 tcp_store_ts_recent(tp); 5225 5226 /* We know that such packets are checksummed 5227 * on entry. 5228 */ 5229 tcp_ack(sk, skb, 0); 5230 __kfree_skb(skb); 5231 tcp_data_snd_check(sk); 5232 return 0; 5233 } else { /* Header too small */ 5234 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS); 5235 goto discard; 5236 } 5237 } else { 5238 int eaten = 0; 5239 int copied_early = 0; 5240 5241 if (tp->copied_seq == tp->rcv_nxt && 5242 len - tcp_header_len <= tp->ucopy.len) { 5243#ifdef CONFIG_NET_DMA 5244 if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) { 5245 copied_early = 1; 5246 eaten = 1; 5247 } 5248#endif 5249 if (tp->ucopy.task == current && 5250 sock_owned_by_user(sk) && !copied_early) { 5251 __set_current_state(TASK_RUNNING); 5252 5253 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len)) 5254 eaten = 1; 5255 } 5256 if (eaten) { 5257 /* Predicted packet is in window by definition. 5258 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 5259 * Hence, check seq<=rcv_wup reduces to: 5260 */ 5261 if (tcp_header_len == 5262 (sizeof(struct tcphdr) + 5263 TCPOLEN_TSTAMP_ALIGNED) && 5264 tp->rcv_nxt == tp->rcv_wup) 5265 tcp_store_ts_recent(tp); 5266 5267 tcp_rcv_rtt_measure_ts(sk, skb); 5268 5269 __skb_pull(skb, tcp_header_len); 5270 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; 5271 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER); 5272 } 5273 if (copied_early) 5274 tcp_cleanup_rbuf(sk, skb->len); 5275 } 5276 if (!eaten) { 5277 if (tcp_checksum_complete_user(sk, skb)) 5278 goto csum_error; 5279 5280 /* Predicted packet is in window by definition. 5281 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 5282 * Hence, check seq<=rcv_wup reduces to: 5283 */ 5284 if (tcp_header_len == 5285 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && 5286 tp->rcv_nxt == tp->rcv_wup) 5287 tcp_store_ts_recent(tp); 5288 5289 tcp_rcv_rtt_measure_ts(sk, skb); 5290 5291 if ((int)skb->truesize > sk->sk_forward_alloc) 5292 goto step5; 5293 5294 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS); 5295 5296 /* Bulk data transfer: receiver */ 5297 __skb_pull(skb, tcp_header_len); 5298 __skb_queue_tail(&sk->sk_receive_queue, skb); 5299 skb_set_owner_r(skb, sk); 5300 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; 5301 } 5302 5303 tcp_event_data_recv(sk, skb); 5304 5305 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) { 5306 /* Well, only one small jumplet in fast path... */ 5307 tcp_ack(sk, skb, FLAG_DATA); 5308 tcp_data_snd_check(sk); 5309 if (!inet_csk_ack_scheduled(sk)) 5310 goto no_ack; 5311 } 5312 5313 if (!copied_early || tp->rcv_nxt != tp->rcv_wup) 5314 __tcp_ack_snd_check(sk, 0); 5315no_ack: 5316#ifdef CONFIG_NET_DMA 5317 if (copied_early) 5318 __skb_queue_tail(&sk->sk_async_wait_queue, skb); 5319 else 5320#endif 5321 if (eaten) 5322 __kfree_skb(skb); 5323 else 5324 sk->sk_data_ready(sk, 0); 5325 return 0; 5326 } 5327 } 5328 5329slow_path: 5330 if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb)) 5331 goto csum_error; 5332 5333 /* 5334 * Standard slow path. 5335 */ 5336 5337 res = tcp_validate_incoming(sk, skb, th, 1); 5338 if (res <= 0) 5339 return -res; 5340 5341step5: 5342 if (th->ack && tcp_ack(sk, skb, FLAG_SLOWPATH) < 0) 5343 goto discard; 5344 5345 tcp_rcv_rtt_measure_ts(sk, skb); 5346 5347 /* Process urgent data. */ 5348 tcp_urg(sk, skb, th); 5349 5350 /* step 7: process the segment text */ 5351 tcp_data_queue(sk, skb); 5352 5353 tcp_data_snd_check(sk); 5354 tcp_ack_snd_check(sk); 5355 return 0; 5356 5357csum_error: 5358 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS); 5359 5360discard: 5361 __kfree_skb(skb); 5362 return 0; 5363} 5364 5365static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb, 5366 struct tcphdr *th, unsigned len) 5367{ 5368 struct tcp_sock *tp = tcp_sk(sk); 5369 struct inet_connection_sock *icsk = inet_csk(sk); 5370 int saved_clamp = tp->rx_opt.mss_clamp; 5371 struct dst_entry *dst = __sk_dst_get(sk); 5372 5373 tcp_parse_options(skb, &tp->rx_opt, 0, dst); 5374 5375 if (th->ack) { 5376 /* rfc793: 5377 * "If the state is SYN-SENT then 5378 * first check the ACK bit 5379 * If the ACK bit is set 5380 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send 5381 * a reset (unless the RST bit is set, if so drop 5382 * the segment and return)" 5383 * 5384 * We do not send data with SYN, so that RFC-correct 5385 * test reduces to: 5386 */ 5387 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt) 5388 goto reset_and_undo; 5389 5390 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 5391 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp, 5392 tcp_time_stamp)) { 5393 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED); 5394 goto reset_and_undo; 5395 } 5396 5397 /* Now ACK is acceptable. 5398 * 5399 * "If the RST bit is set 5400 * If the ACK was acceptable then signal the user "error: 5401 * connection reset", drop the segment, enter CLOSED state, 5402 * delete TCB, and return." 5403 */ 5404 5405 if (th->rst) { 5406 tcp_reset(sk); 5407 goto discard; 5408 } 5409 5410 /* rfc793: 5411 * "fifth, if neither of the SYN or RST bits is set then 5412 * drop the segment and return." 5413 * 5414 * See note below! 5415 * --ANK(990513) 5416 */ 5417 if (!th->syn) 5418 goto discard_and_undo; 5419 5420 /* rfc793: 5421 * "If the SYN bit is on ... 5422 * are acceptable then ... 5423 * (our SYN has been ACKed), change the connection 5424 * state to ESTABLISHED..." 5425 */ 5426 5427 TCP_ECN_rcv_synack(tp, th); 5428 5429 tp->snd_wl1 = TCP_SKB_CB(skb)->seq; 5430 tcp_ack(sk, skb, FLAG_SLOWPATH); 5431 5432 /* Ok.. it's good. Set up sequence numbers and 5433 * move to established. 5434 */ 5435 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; 5436 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; 5437 5438 /* RFC1323: The window in SYN & SYN/ACK segments is 5439 * never scaled. 5440 */ 5441 tp->snd_wnd = ntohs(th->window); 5442 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq); 5443 5444 if (!tp->rx_opt.wscale_ok) { 5445 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0; 5446 tp->window_clamp = min(tp->window_clamp, 65535U); 5447 } 5448 5449 if (tp->rx_opt.saw_tstamp) { 5450 tp->rx_opt.tstamp_ok = 1; 5451 tp->tcp_header_len = 5452 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 5453 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; 5454 tcp_store_ts_recent(tp); 5455 } else { 5456 tp->tcp_header_len = sizeof(struct tcphdr); 5457 } 5458 5459 if (tcp_is_sack(tp) && sysctl_tcp_fack) 5460 tcp_enable_fack(tp); 5461 5462 tcp_mtup_init(sk); 5463 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 5464 tcp_initialize_rcv_mss(sk); 5465 5466 /* Remember, tcp_poll() does not lock socket! 5467 * Change state from SYN-SENT only after copied_seq 5468 * is initialized. */ 5469 tp->copied_seq = tp->rcv_nxt; 5470 smp_mb(); 5471 tcp_set_state(sk, TCP_ESTABLISHED); 5472 5473 security_inet_conn_established(sk, skb); 5474 5475 /* Make sure socket is routed, for correct metrics. */ 5476 icsk->icsk_af_ops->rebuild_header(sk); 5477 5478 tcp_init_metrics(sk); 5479 5480 tcp_init_congestion_control(sk); 5481 5482 /* Prevent spurious tcp_cwnd_restart() on first data 5483 * packet. 5484 */ 5485 tp->lsndtime = tcp_time_stamp; 5486 5487 tcp_init_buffer_space(sk); 5488 5489 if (sock_flag(sk, SOCK_KEEPOPEN)) 5490 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp)); 5491 5492 if (!tp->rx_opt.snd_wscale) 5493 __tcp_fast_path_on(tp, tp->snd_wnd); 5494 else 5495 tp->pred_flags = 0; 5496 5497 if (!sock_flag(sk, SOCK_DEAD)) { 5498 sk->sk_state_change(sk); 5499 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT); 5500 } 5501 5502 if (sk->sk_write_pending || 5503 icsk->icsk_accept_queue.rskq_defer_accept || 5504 icsk->icsk_ack.pingpong) { 5505 /* Save one ACK. Data will be ready after 5506 * several ticks, if write_pending is set. 5507 * 5508 * It may be deleted, but with this feature tcpdumps 5509 * look so _wonderfully_ clever, that I was not able 5510 * to stand against the temptation 8) --ANK 5511 */ 5512 inet_csk_schedule_ack(sk); 5513 icsk->icsk_ack.lrcvtime = tcp_time_stamp; 5514 icsk->icsk_ack.ato = TCP_ATO_MIN; 5515 tcp_incr_quickack(sk); 5516 tcp_enter_quickack_mode(sk); 5517 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 5518 TCP_DELACK_MAX, TCP_RTO_MAX); 5519 5520discard: 5521 __kfree_skb(skb); 5522 return 0; 5523 } else { 5524 tcp_send_ack(sk); 5525 } 5526 return -1; 5527 } 5528 5529 /* No ACK in the segment */ 5530 5531 if (th->rst) { 5532 /* rfc793: 5533 * "If the RST bit is set 5534 * 5535 * Otherwise (no ACK) drop the segment and return." 5536 */ 5537 5538 goto discard_and_undo; 5539 } 5540 5541 /* PAWS check. */ 5542 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && 5543 tcp_paws_reject(&tp->rx_opt, 0)) 5544 goto discard_and_undo; 5545 5546 if (th->syn) { 5547 /* We see SYN without ACK. It is attempt of 5548 * simultaneous connect with crossed SYNs. 5549 * Particularly, it can be connect to self. 5550 */ 5551 tcp_set_state(sk, TCP_SYN_RECV); 5552 5553 if (tp->rx_opt.saw_tstamp) { 5554 tp->rx_opt.tstamp_ok = 1; 5555 tcp_store_ts_recent(tp); 5556 tp->tcp_header_len = 5557 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 5558 } else { 5559 tp->tcp_header_len = sizeof(struct tcphdr); 5560 } 5561 5562 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; 5563 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; 5564 5565 /* RFC1323: The window in SYN & SYN/ACK segments is 5566 * never scaled. 5567 */ 5568 tp->snd_wnd = ntohs(th->window); 5569 tp->snd_wl1 = TCP_SKB_CB(skb)->seq; 5570 tp->max_window = tp->snd_wnd; 5571 5572 TCP_ECN_rcv_syn(tp, th); 5573 5574 tcp_mtup_init(sk); 5575 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 5576 tcp_initialize_rcv_mss(sk); 5577 5578 tcp_send_synack(sk); 5579#if 0 5580 /* Note, we could accept data and URG from this segment. 5581 * There are no obstacles to make this. 5582 * 5583 * However, if we ignore data in ACKless segments sometimes, 5584 * we have no reasons to accept it sometimes. 5585 * Also, seems the code doing it in step6 of tcp_rcv_state_process 5586 * is not flawless. So, discard packet for sanity. 5587 * Uncomment this return to process the data. 5588 */ 5589 return -1; 5590#else 5591 goto discard; 5592#endif 5593 } 5594 /* "fifth, if neither of the SYN or RST bits is set then 5595 * drop the segment and return." 5596 */ 5597 5598discard_and_undo: 5599 tcp_clear_options(&tp->rx_opt); 5600 tp->rx_opt.mss_clamp = saved_clamp; 5601 goto discard; 5602 5603reset_and_undo: 5604 tcp_clear_options(&tp->rx_opt); 5605 tp->rx_opt.mss_clamp = saved_clamp; 5606 return 1; 5607} 5608 5609/* 5610 * This function implements the receiving procedure of RFC 793 for 5611 * all states except ESTABLISHED and TIME_WAIT. 5612 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be 5613 * address independent. 5614 */ 5615 5616int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb, 5617 struct tcphdr *th, unsigned len) 5618{ 5619 struct tcp_sock *tp = tcp_sk(sk); 5620 struct inet_connection_sock *icsk = inet_csk(sk); 5621 int queued = 0; 5622 int res; 5623 5624 tp->rx_opt.saw_tstamp = 0; 5625 5626 switch (sk->sk_state) { 5627 case TCP_CLOSE: 5628 goto discard; 5629 5630 case TCP_LISTEN: 5631 if (th->ack) 5632 return 1; 5633 5634 if (th->rst) 5635 goto discard; 5636 5637 if (th->syn) { 5638 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0) 5639 return 1; 5640 5641 /* Now we have several options: In theory there is 5642 * nothing else in the frame. KA9Q has an option to 5643 * send data with the syn, BSD accepts data with the 5644 * syn up to the [to be] advertised window and 5645 * Solaris 2.1 gives you a protocol error. For now 5646 * we just ignore it, that fits the spec precisely 5647 * and avoids incompatibilities. It would be nice in 5648 * future to drop through and process the data. 5649 * 5650 * Now that TTCP is starting to be used we ought to 5651 * queue this data. 5652 * But, this leaves one open to an easy denial of 5653 * service attack, and SYN cookies can't defend 5654 * against this problem. So, we drop the data 5655 * in the interest of security over speed unless 5656 * it's still in use. 5657 */ 5658 kfree_skb(skb); 5659 return 0; 5660 } 5661 goto discard; 5662 5663 case TCP_SYN_SENT: 5664 queued = tcp_rcv_synsent_state_process(sk, skb, th, len); 5665 if (queued >= 0) 5666 return queued; 5667 5668 /* Do step6 onward by hand. */ 5669 tcp_urg(sk, skb, th); 5670 __kfree_skb(skb); 5671 tcp_data_snd_check(sk); 5672 return 0; 5673 } 5674 5675 res = tcp_validate_incoming(sk, skb, th, 0); 5676 if (res <= 0) 5677 return -res; 5678 5679 /* step 5: check the ACK field */ 5680 if (th->ack) { 5681 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH) > 0; 5682 5683 switch (sk->sk_state) { 5684 case TCP_SYN_RECV: 5685 if (acceptable) { 5686 tp->copied_seq = tp->rcv_nxt; 5687 smp_mb(); 5688 tcp_set_state(sk, TCP_ESTABLISHED); 5689 sk->sk_state_change(sk); 5690 5691 /* Note, that this wakeup is only for marginal 5692 * crossed SYN case. Passively open sockets 5693 * are not waked up, because sk->sk_sleep == 5694 * NULL and sk->sk_socket == NULL. 5695 */ 5696 if (sk->sk_socket) 5697 sk_wake_async(sk, 5698 SOCK_WAKE_IO, POLL_OUT); 5699 5700 tp->snd_una = TCP_SKB_CB(skb)->ack_seq; 5701 tp->snd_wnd = ntohs(th->window) << 5702 tp->rx_opt.snd_wscale; 5703 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq); 5704 5705 /* tcp_ack considers this ACK as duplicate 5706 * and does not calculate rtt. 5707 * Fix it at least with timestamps. 5708 */ 5709 if (tp->rx_opt.saw_tstamp && 5710 tp->rx_opt.rcv_tsecr && !tp->srtt) 5711 tcp_ack_saw_tstamp(sk, 0); 5712 5713 if (tp->rx_opt.tstamp_ok) 5714 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; 5715 5716 /* Make sure socket is routed, for 5717 * correct metrics. 5718 */ 5719 icsk->icsk_af_ops->rebuild_header(sk); 5720 5721 tcp_init_metrics(sk); 5722 5723 tcp_init_congestion_control(sk); 5724 5725 /* Prevent spurious tcp_cwnd_restart() on 5726 * first data packet. 5727 */ 5728 tp->lsndtime = tcp_time_stamp; 5729 5730 tcp_mtup_init(sk); 5731 tcp_initialize_rcv_mss(sk); 5732 tcp_init_buffer_space(sk); 5733 tcp_fast_path_on(tp); 5734 } else { 5735 return 1; 5736 } 5737 break; 5738 5739 case TCP_FIN_WAIT1: 5740 if (tp->snd_una == tp->write_seq) { 5741 tcp_set_state(sk, TCP_FIN_WAIT2); 5742 sk->sk_shutdown |= SEND_SHUTDOWN; 5743 dst_confirm(sk->sk_dst_cache); 5744 5745 if (!sock_flag(sk, SOCK_DEAD)) 5746 /* Wake up lingering close() */ 5747 sk->sk_state_change(sk); 5748 else { 5749 int tmo; 5750 5751 if (tp->linger2 < 0 || 5752 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 5753 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) { 5754 tcp_done(sk); 5755 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA); 5756 return 1; 5757 } 5758 5759 tmo = tcp_fin_time(sk); 5760 if (tmo > TCP_TIMEWAIT_LEN) { 5761 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN); 5762 } else if (th->fin || sock_owned_by_user(sk)) { 5763 /* Bad case. We could lose such FIN otherwise. 5764 * It is not a big problem, but it looks confusing 5765 * and not so rare event. We still can lose it now, 5766 * if it spins in bh_lock_sock(), but it is really 5767 * marginal case. 5768 */ 5769 inet_csk_reset_keepalive_timer(sk, tmo); 5770 } else { 5771 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo); 5772 goto discard; 5773 } 5774 } 5775 } 5776 break; 5777 5778 case TCP_CLOSING: 5779 if (tp->snd_una == tp->write_seq) { 5780 tcp_time_wait(sk, TCP_TIME_WAIT, 0); 5781 goto discard; 5782 } 5783 break; 5784 5785 case TCP_LAST_ACK: 5786 if (tp->snd_una == tp->write_seq) { 5787 tcp_update_metrics(sk); 5788 tcp_done(sk); 5789 goto discard; 5790 } 5791 break; 5792 } 5793 } else 5794 goto discard; 5795 5796 /* step 6: check the URG bit */ 5797 tcp_urg(sk, skb, th); 5798 5799 /* step 7: process the segment text */ 5800 switch (sk->sk_state) { 5801 case TCP_CLOSE_WAIT: 5802 case TCP_CLOSING: 5803 case TCP_LAST_ACK: 5804 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) 5805 break; 5806 case TCP_FIN_WAIT1: 5807 case TCP_FIN_WAIT2: 5808 /* RFC 793 says to queue data in these states, 5809 * RFC 1122 says we MUST send a reset. 5810 * BSD 4.4 also does reset. 5811 */ 5812 if (sk->sk_shutdown & RCV_SHUTDOWN) { 5813 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 5814 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) { 5815 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA); 5816 tcp_reset(sk); 5817 return 1; 5818 } 5819 } 5820 /* Fall through */ 5821 case TCP_ESTABLISHED: 5822 tcp_data_queue(sk, skb); 5823 queued = 1; 5824 break; 5825 } 5826 5827 /* tcp_data could move socket to TIME-WAIT */ 5828 if (sk->sk_state != TCP_CLOSE) { 5829 tcp_data_snd_check(sk); 5830 tcp_ack_snd_check(sk); 5831 } 5832 5833 if (!queued) { 5834discard: 5835 __kfree_skb(skb); 5836 } 5837 return 0; 5838} 5839 5840EXPORT_SYMBOL(sysctl_tcp_ecn); 5841EXPORT_SYMBOL(sysctl_tcp_reordering); 5842EXPORT_SYMBOL(sysctl_tcp_adv_win_scale); 5843EXPORT_SYMBOL(tcp_parse_options); 5844#ifdef CONFIG_TCP_MD5SIG 5845EXPORT_SYMBOL(tcp_parse_md5sig_option); 5846#endif 5847EXPORT_SYMBOL(tcp_rcv_established); 5848EXPORT_SYMBOL(tcp_rcv_state_process); 5849EXPORT_SYMBOL(tcp_initialize_rcv_mss); 5850