tcp_input.c revision 651913ce9de2bbcedef608c5d6cf39c244248509
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#define pr_fmt(fmt) "TCP: " fmt 65 66#include <linux/mm.h> 67#include <linux/slab.h> 68#include <linux/module.h> 69#include <linux/sysctl.h> 70#include <linux/kernel.h> 71#include <net/dst.h> 72#include <net/tcp.h> 73#include <net/inet_common.h> 74#include <linux/ipsec.h> 75#include <asm/unaligned.h> 76#include <net/netdma.h> 77 78int sysctl_tcp_timestamps __read_mostly = 1; 79int sysctl_tcp_window_scaling __read_mostly = 1; 80int sysctl_tcp_sack __read_mostly = 1; 81int sysctl_tcp_fack __read_mostly = 1; 82int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH; 83EXPORT_SYMBOL(sysctl_tcp_reordering); 84int sysctl_tcp_ecn __read_mostly = 2; 85EXPORT_SYMBOL(sysctl_tcp_ecn); 86int sysctl_tcp_dsack __read_mostly = 1; 87int sysctl_tcp_app_win __read_mostly = 31; 88int sysctl_tcp_adv_win_scale __read_mostly = 2; 89EXPORT_SYMBOL(sysctl_tcp_adv_win_scale); 90 91int sysctl_tcp_stdurg __read_mostly; 92int sysctl_tcp_rfc1337 __read_mostly; 93int sysctl_tcp_max_orphans __read_mostly = NR_FILE; 94int sysctl_tcp_frto __read_mostly = 2; 95int sysctl_tcp_frto_response __read_mostly; 96int sysctl_tcp_nometrics_save __read_mostly; 97 98int sysctl_tcp_thin_dupack __read_mostly; 99 100int sysctl_tcp_moderate_rcvbuf __read_mostly = 1; 101int sysctl_tcp_abc __read_mostly; 102 103#define FLAG_DATA 0x01 /* Incoming frame contained data. */ 104#define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */ 105#define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */ 106#define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */ 107#define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */ 108#define FLAG_DATA_SACKED 0x20 /* New SACK. */ 109#define FLAG_ECE 0x40 /* ECE in this ACK */ 110#define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/ 111#define FLAG_ONLY_ORIG_SACKED 0x200 /* SACKs only non-rexmit sent before RTO */ 112#define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */ 113#define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */ 114#define FLAG_NONHEAD_RETRANS_ACKED 0x1000 /* Non-head rexmitted data was ACKed */ 115#define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */ 116 117#define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED) 118#define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED) 119#define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE) 120#define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED) 121#define FLAG_ANY_PROGRESS (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED) 122 123#define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH) 124#define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH)) 125 126/* Adapt the MSS value used to make delayed ack decision to the 127 * real world. 128 */ 129static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb) 130{ 131 struct inet_connection_sock *icsk = inet_csk(sk); 132 const unsigned int lss = icsk->icsk_ack.last_seg_size; 133 unsigned int len; 134 135 icsk->icsk_ack.last_seg_size = 0; 136 137 /* skb->len may jitter because of SACKs, even if peer 138 * sends good full-sized frames. 139 */ 140 len = skb_shinfo(skb)->gso_size ? : skb->len; 141 if (len >= icsk->icsk_ack.rcv_mss) { 142 icsk->icsk_ack.rcv_mss = len; 143 } else { 144 /* Otherwise, we make more careful check taking into account, 145 * that SACKs block is variable. 146 * 147 * "len" is invariant segment length, including TCP header. 148 */ 149 len += skb->data - skb_transport_header(skb); 150 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) || 151 /* If PSH is not set, packet should be 152 * full sized, provided peer TCP is not badly broken. 153 * This observation (if it is correct 8)) allows 154 * to handle super-low mtu links fairly. 155 */ 156 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) && 157 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) { 158 /* Subtract also invariant (if peer is RFC compliant), 159 * tcp header plus fixed timestamp option length. 160 * Resulting "len" is MSS free of SACK jitter. 161 */ 162 len -= tcp_sk(sk)->tcp_header_len; 163 icsk->icsk_ack.last_seg_size = len; 164 if (len == lss) { 165 icsk->icsk_ack.rcv_mss = len; 166 return; 167 } 168 } 169 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED) 170 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2; 171 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED; 172 } 173} 174 175static void tcp_incr_quickack(struct sock *sk) 176{ 177 struct inet_connection_sock *icsk = inet_csk(sk); 178 unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss); 179 180 if (quickacks == 0) 181 quickacks = 2; 182 if (quickacks > icsk->icsk_ack.quick) 183 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS); 184} 185 186static void tcp_enter_quickack_mode(struct sock *sk) 187{ 188 struct inet_connection_sock *icsk = inet_csk(sk); 189 tcp_incr_quickack(sk); 190 icsk->icsk_ack.pingpong = 0; 191 icsk->icsk_ack.ato = TCP_ATO_MIN; 192} 193 194/* Send ACKs quickly, if "quick" count is not exhausted 195 * and the session is not interactive. 196 */ 197 198static inline int tcp_in_quickack_mode(const struct sock *sk) 199{ 200 const struct inet_connection_sock *icsk = inet_csk(sk); 201 return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong; 202} 203 204static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp) 205{ 206 if (tp->ecn_flags & TCP_ECN_OK) 207 tp->ecn_flags |= TCP_ECN_QUEUE_CWR; 208} 209 210static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb) 211{ 212 if (tcp_hdr(skb)->cwr) 213 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR; 214} 215 216static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp) 217{ 218 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR; 219} 220 221static inline void TCP_ECN_check_ce(struct tcp_sock *tp, const struct sk_buff *skb) 222{ 223 if (!(tp->ecn_flags & TCP_ECN_OK)) 224 return; 225 226 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) { 227 case INET_ECN_NOT_ECT: 228 /* Funny extension: if ECT is not set on a segment, 229 * and we already seen ECT on a previous segment, 230 * it is probably a retransmit. 231 */ 232 if (tp->ecn_flags & TCP_ECN_SEEN) 233 tcp_enter_quickack_mode((struct sock *)tp); 234 break; 235 case INET_ECN_CE: 236 tp->ecn_flags |= TCP_ECN_DEMAND_CWR; 237 /* fallinto */ 238 default: 239 tp->ecn_flags |= TCP_ECN_SEEN; 240 } 241} 242 243static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th) 244{ 245 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr)) 246 tp->ecn_flags &= ~TCP_ECN_OK; 247} 248 249static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th) 250{ 251 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr)) 252 tp->ecn_flags &= ~TCP_ECN_OK; 253} 254 255static inline int TCP_ECN_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th) 256{ 257 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK)) 258 return 1; 259 return 0; 260} 261 262/* Buffer size and advertised window tuning. 263 * 264 * 1. Tuning sk->sk_sndbuf, when connection enters established state. 265 */ 266 267static void tcp_fixup_sndbuf(struct sock *sk) 268{ 269 int sndmem = SKB_TRUESIZE(tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER); 270 271 sndmem *= TCP_INIT_CWND; 272 if (sk->sk_sndbuf < sndmem) 273 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]); 274} 275 276/* 2. Tuning advertised window (window_clamp, rcv_ssthresh) 277 * 278 * All tcp_full_space() is split to two parts: "network" buffer, allocated 279 * forward and advertised in receiver window (tp->rcv_wnd) and 280 * "application buffer", required to isolate scheduling/application 281 * latencies from network. 282 * window_clamp is maximal advertised window. It can be less than 283 * tcp_full_space(), in this case tcp_full_space() - window_clamp 284 * is reserved for "application" buffer. The less window_clamp is 285 * the smoother our behaviour from viewpoint of network, but the lower 286 * throughput and the higher sensitivity of the connection to losses. 8) 287 * 288 * rcv_ssthresh is more strict window_clamp used at "slow start" 289 * phase to predict further behaviour of this connection. 290 * It is used for two goals: 291 * - to enforce header prediction at sender, even when application 292 * requires some significant "application buffer". It is check #1. 293 * - to prevent pruning of receive queue because of misprediction 294 * of receiver window. Check #2. 295 * 296 * The scheme does not work when sender sends good segments opening 297 * window and then starts to feed us spaghetti. But it should work 298 * in common situations. Otherwise, we have to rely on queue collapsing. 299 */ 300 301/* Slow part of check#2. */ 302static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb) 303{ 304 struct tcp_sock *tp = tcp_sk(sk); 305 /* Optimize this! */ 306 int truesize = tcp_win_from_space(skb->truesize) >> 1; 307 int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1; 308 309 while (tp->rcv_ssthresh <= window) { 310 if (truesize <= skb->len) 311 return 2 * inet_csk(sk)->icsk_ack.rcv_mss; 312 313 truesize >>= 1; 314 window >>= 1; 315 } 316 return 0; 317} 318 319static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb) 320{ 321 struct tcp_sock *tp = tcp_sk(sk); 322 323 /* Check #1 */ 324 if (tp->rcv_ssthresh < tp->window_clamp && 325 (int)tp->rcv_ssthresh < tcp_space(sk) && 326 !sk_under_memory_pressure(sk)) { 327 int incr; 328 329 /* Check #2. Increase window, if skb with such overhead 330 * will fit to rcvbuf in future. 331 */ 332 if (tcp_win_from_space(skb->truesize) <= skb->len) 333 incr = 2 * tp->advmss; 334 else 335 incr = __tcp_grow_window(sk, skb); 336 337 if (incr) { 338 incr = max_t(int, incr, 2 * skb->len); 339 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr, 340 tp->window_clamp); 341 inet_csk(sk)->icsk_ack.quick |= 1; 342 } 343 } 344} 345 346/* 3. Tuning rcvbuf, when connection enters established state. */ 347 348static void tcp_fixup_rcvbuf(struct sock *sk) 349{ 350 u32 mss = tcp_sk(sk)->advmss; 351 u32 icwnd = TCP_DEFAULT_INIT_RCVWND; 352 int rcvmem; 353 354 /* Limit to 10 segments if mss <= 1460, 355 * or 14600/mss segments, with a minimum of two segments. 356 */ 357 if (mss > 1460) 358 icwnd = max_t(u32, (1460 * TCP_DEFAULT_INIT_RCVWND) / mss, 2); 359 360 rcvmem = SKB_TRUESIZE(mss + MAX_TCP_HEADER); 361 while (tcp_win_from_space(rcvmem) < mss) 362 rcvmem += 128; 363 364 rcvmem *= icwnd; 365 366 if (sk->sk_rcvbuf < rcvmem) 367 sk->sk_rcvbuf = min(rcvmem, sysctl_tcp_rmem[2]); 368} 369 370/* 4. Try to fixup all. It is made immediately after connection enters 371 * established state. 372 */ 373static void tcp_init_buffer_space(struct sock *sk) 374{ 375 struct tcp_sock *tp = tcp_sk(sk); 376 int maxwin; 377 378 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) 379 tcp_fixup_rcvbuf(sk); 380 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK)) 381 tcp_fixup_sndbuf(sk); 382 383 tp->rcvq_space.space = tp->rcv_wnd; 384 385 maxwin = tcp_full_space(sk); 386 387 if (tp->window_clamp >= maxwin) { 388 tp->window_clamp = maxwin; 389 390 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss) 391 tp->window_clamp = max(maxwin - 392 (maxwin >> sysctl_tcp_app_win), 393 4 * tp->advmss); 394 } 395 396 /* Force reservation of one segment. */ 397 if (sysctl_tcp_app_win && 398 tp->window_clamp > 2 * tp->advmss && 399 tp->window_clamp + tp->advmss > maxwin) 400 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss); 401 402 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp); 403 tp->snd_cwnd_stamp = tcp_time_stamp; 404} 405 406/* 5. Recalculate window clamp after socket hit its memory bounds. */ 407static void tcp_clamp_window(struct sock *sk) 408{ 409 struct tcp_sock *tp = tcp_sk(sk); 410 struct inet_connection_sock *icsk = inet_csk(sk); 411 412 icsk->icsk_ack.quick = 0; 413 414 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] && 415 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) && 416 !sk_under_memory_pressure(sk) && 417 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) { 418 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc), 419 sysctl_tcp_rmem[2]); 420 } 421 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf) 422 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss); 423} 424 425/* Initialize RCV_MSS value. 426 * RCV_MSS is an our guess about MSS used by the peer. 427 * We haven't any direct information about the MSS. 428 * It's better to underestimate the RCV_MSS rather than overestimate. 429 * Overestimations make us ACKing less frequently than needed. 430 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss(). 431 */ 432void tcp_initialize_rcv_mss(struct sock *sk) 433{ 434 const struct tcp_sock *tp = tcp_sk(sk); 435 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache); 436 437 hint = min(hint, tp->rcv_wnd / 2); 438 hint = min(hint, TCP_MSS_DEFAULT); 439 hint = max(hint, TCP_MIN_MSS); 440 441 inet_csk(sk)->icsk_ack.rcv_mss = hint; 442} 443EXPORT_SYMBOL(tcp_initialize_rcv_mss); 444 445/* Receiver "autotuning" code. 446 * 447 * The algorithm for RTT estimation w/o timestamps is based on 448 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL. 449 * <http://public.lanl.gov/radiant/pubs.html#DRS> 450 * 451 * More detail on this code can be found at 452 * <http://staff.psc.edu/jheffner/>, 453 * though this reference is out of date. A new paper 454 * is pending. 455 */ 456static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep) 457{ 458 u32 new_sample = tp->rcv_rtt_est.rtt; 459 long m = sample; 460 461 if (m == 0) 462 m = 1; 463 464 if (new_sample != 0) { 465 /* If we sample in larger samples in the non-timestamp 466 * case, we could grossly overestimate the RTT especially 467 * with chatty applications or bulk transfer apps which 468 * are stalled on filesystem I/O. 469 * 470 * Also, since we are only going for a minimum in the 471 * non-timestamp case, we do not smooth things out 472 * else with timestamps disabled convergence takes too 473 * long. 474 */ 475 if (!win_dep) { 476 m -= (new_sample >> 3); 477 new_sample += m; 478 } else { 479 m <<= 3; 480 if (m < new_sample) 481 new_sample = m; 482 } 483 } else { 484 /* No previous measure. */ 485 new_sample = m << 3; 486 } 487 488 if (tp->rcv_rtt_est.rtt != new_sample) 489 tp->rcv_rtt_est.rtt = new_sample; 490} 491 492static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp) 493{ 494 if (tp->rcv_rtt_est.time == 0) 495 goto new_measure; 496 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq)) 497 return; 498 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rcv_rtt_est.time, 1); 499 500new_measure: 501 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd; 502 tp->rcv_rtt_est.time = tcp_time_stamp; 503} 504 505static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, 506 const struct sk_buff *skb) 507{ 508 struct tcp_sock *tp = tcp_sk(sk); 509 if (tp->rx_opt.rcv_tsecr && 510 (TCP_SKB_CB(skb)->end_seq - 511 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss)) 512 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0); 513} 514 515/* 516 * This function should be called every time data is copied to user space. 517 * It calculates the appropriate TCP receive buffer space. 518 */ 519void tcp_rcv_space_adjust(struct sock *sk) 520{ 521 struct tcp_sock *tp = tcp_sk(sk); 522 int time; 523 int space; 524 525 if (tp->rcvq_space.time == 0) 526 goto new_measure; 527 528 time = tcp_time_stamp - tp->rcvq_space.time; 529 if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0) 530 return; 531 532 space = 2 * (tp->copied_seq - tp->rcvq_space.seq); 533 534 space = max(tp->rcvq_space.space, space); 535 536 if (tp->rcvq_space.space != space) { 537 int rcvmem; 538 539 tp->rcvq_space.space = space; 540 541 if (sysctl_tcp_moderate_rcvbuf && 542 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) { 543 int new_clamp = space; 544 545 /* Receive space grows, normalize in order to 546 * take into account packet headers and sk_buff 547 * structure overhead. 548 */ 549 space /= tp->advmss; 550 if (!space) 551 space = 1; 552 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER); 553 while (tcp_win_from_space(rcvmem) < tp->advmss) 554 rcvmem += 128; 555 space *= rcvmem; 556 space = min(space, sysctl_tcp_rmem[2]); 557 if (space > sk->sk_rcvbuf) { 558 sk->sk_rcvbuf = space; 559 560 /* Make the window clamp follow along. */ 561 tp->window_clamp = new_clamp; 562 } 563 } 564 } 565 566new_measure: 567 tp->rcvq_space.seq = tp->copied_seq; 568 tp->rcvq_space.time = tcp_time_stamp; 569} 570 571/* There is something which you must keep in mind when you analyze the 572 * behavior of the tp->ato delayed ack timeout interval. When a 573 * connection starts up, we want to ack as quickly as possible. The 574 * problem is that "good" TCP's do slow start at the beginning of data 575 * transmission. The means that until we send the first few ACK's the 576 * sender will sit on his end and only queue most of his data, because 577 * he can only send snd_cwnd unacked packets at any given time. For 578 * each ACK we send, he increments snd_cwnd and transmits more of his 579 * queue. -DaveM 580 */ 581static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb) 582{ 583 struct tcp_sock *tp = tcp_sk(sk); 584 struct inet_connection_sock *icsk = inet_csk(sk); 585 u32 now; 586 587 inet_csk_schedule_ack(sk); 588 589 tcp_measure_rcv_mss(sk, skb); 590 591 tcp_rcv_rtt_measure(tp); 592 593 now = tcp_time_stamp; 594 595 if (!icsk->icsk_ack.ato) { 596 /* The _first_ data packet received, initialize 597 * delayed ACK engine. 598 */ 599 tcp_incr_quickack(sk); 600 icsk->icsk_ack.ato = TCP_ATO_MIN; 601 } else { 602 int m = now - icsk->icsk_ack.lrcvtime; 603 604 if (m <= TCP_ATO_MIN / 2) { 605 /* The fastest case is the first. */ 606 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2; 607 } else if (m < icsk->icsk_ack.ato) { 608 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m; 609 if (icsk->icsk_ack.ato > icsk->icsk_rto) 610 icsk->icsk_ack.ato = icsk->icsk_rto; 611 } else if (m > icsk->icsk_rto) { 612 /* Too long gap. Apparently sender failed to 613 * restart window, so that we send ACKs quickly. 614 */ 615 tcp_incr_quickack(sk); 616 sk_mem_reclaim(sk); 617 } 618 } 619 icsk->icsk_ack.lrcvtime = now; 620 621 TCP_ECN_check_ce(tp, skb); 622 623 if (skb->len >= 128) 624 tcp_grow_window(sk, skb); 625} 626 627/* Called to compute a smoothed rtt estimate. The data fed to this 628 * routine either comes from timestamps, or from segments that were 629 * known _not_ to have been retransmitted [see Karn/Partridge 630 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88 631 * piece by Van Jacobson. 632 * NOTE: the next three routines used to be one big routine. 633 * To save cycles in the RFC 1323 implementation it was better to break 634 * it up into three procedures. -- erics 635 */ 636static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt) 637{ 638 struct tcp_sock *tp = tcp_sk(sk); 639 long m = mrtt; /* RTT */ 640 641 /* The following amusing code comes from Jacobson's 642 * article in SIGCOMM '88. Note that rtt and mdev 643 * are scaled versions of rtt and mean deviation. 644 * This is designed to be as fast as possible 645 * m stands for "measurement". 646 * 647 * On a 1990 paper the rto value is changed to: 648 * RTO = rtt + 4 * mdev 649 * 650 * Funny. This algorithm seems to be very broken. 651 * These formulae increase RTO, when it should be decreased, increase 652 * too slowly, when it should be increased quickly, decrease too quickly 653 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely 654 * does not matter how to _calculate_ it. Seems, it was trap 655 * that VJ failed to avoid. 8) 656 */ 657 if (m == 0) 658 m = 1; 659 if (tp->srtt != 0) { 660 m -= (tp->srtt >> 3); /* m is now error in rtt est */ 661 tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */ 662 if (m < 0) { 663 m = -m; /* m is now abs(error) */ 664 m -= (tp->mdev >> 2); /* similar update on mdev */ 665 /* This is similar to one of Eifel findings. 666 * Eifel blocks mdev updates when rtt decreases. 667 * This solution is a bit different: we use finer gain 668 * for mdev in this case (alpha*beta). 669 * Like Eifel it also prevents growth of rto, 670 * but also it limits too fast rto decreases, 671 * happening in pure Eifel. 672 */ 673 if (m > 0) 674 m >>= 3; 675 } else { 676 m -= (tp->mdev >> 2); /* similar update on mdev */ 677 } 678 tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */ 679 if (tp->mdev > tp->mdev_max) { 680 tp->mdev_max = tp->mdev; 681 if (tp->mdev_max > tp->rttvar) 682 tp->rttvar = tp->mdev_max; 683 } 684 if (after(tp->snd_una, tp->rtt_seq)) { 685 if (tp->mdev_max < tp->rttvar) 686 tp->rttvar -= (tp->rttvar - tp->mdev_max) >> 2; 687 tp->rtt_seq = tp->snd_nxt; 688 tp->mdev_max = tcp_rto_min(sk); 689 } 690 } else { 691 /* no previous measure. */ 692 tp->srtt = m << 3; /* take the measured time to be rtt */ 693 tp->mdev = m << 1; /* make sure rto = 3*rtt */ 694 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk)); 695 tp->rtt_seq = tp->snd_nxt; 696 } 697} 698 699/* Calculate rto without backoff. This is the second half of Van Jacobson's 700 * routine referred to above. 701 */ 702static inline void tcp_set_rto(struct sock *sk) 703{ 704 const struct tcp_sock *tp = tcp_sk(sk); 705 /* Old crap is replaced with new one. 8) 706 * 707 * More seriously: 708 * 1. If rtt variance happened to be less 50msec, it is hallucination. 709 * It cannot be less due to utterly erratic ACK generation made 710 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_ 711 * to do with delayed acks, because at cwnd>2 true delack timeout 712 * is invisible. Actually, Linux-2.4 also generates erratic 713 * ACKs in some circumstances. 714 */ 715 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp); 716 717 /* 2. Fixups made earlier cannot be right. 718 * If we do not estimate RTO correctly without them, 719 * all the algo is pure shit and should be replaced 720 * with correct one. It is exactly, which we pretend to do. 721 */ 722 723 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo 724 * guarantees that rto is higher. 725 */ 726 tcp_bound_rto(sk); 727} 728 729/* Save metrics learned by this TCP session. 730 This function is called only, when TCP finishes successfully 731 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE. 732 */ 733void tcp_update_metrics(struct sock *sk) 734{ 735 struct tcp_sock *tp = tcp_sk(sk); 736 struct dst_entry *dst = __sk_dst_get(sk); 737 738 if (sysctl_tcp_nometrics_save) 739 return; 740 741 dst_confirm(dst); 742 743 if (dst && (dst->flags & DST_HOST)) { 744 const struct inet_connection_sock *icsk = inet_csk(sk); 745 int m; 746 unsigned long rtt; 747 748 if (icsk->icsk_backoff || !tp->srtt) { 749 /* This session failed to estimate rtt. Why? 750 * Probably, no packets returned in time. 751 * Reset our results. 752 */ 753 if (!(dst_metric_locked(dst, RTAX_RTT))) 754 dst_metric_set(dst, RTAX_RTT, 0); 755 return; 756 } 757 758 rtt = dst_metric_rtt(dst, RTAX_RTT); 759 m = rtt - tp->srtt; 760 761 /* If newly calculated rtt larger than stored one, 762 * store new one. Otherwise, use EWMA. Remember, 763 * rtt overestimation is always better than underestimation. 764 */ 765 if (!(dst_metric_locked(dst, RTAX_RTT))) { 766 if (m <= 0) 767 set_dst_metric_rtt(dst, RTAX_RTT, tp->srtt); 768 else 769 set_dst_metric_rtt(dst, RTAX_RTT, rtt - (m >> 3)); 770 } 771 772 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) { 773 unsigned long var; 774 if (m < 0) 775 m = -m; 776 777 /* Scale deviation to rttvar fixed point */ 778 m >>= 1; 779 if (m < tp->mdev) 780 m = tp->mdev; 781 782 var = dst_metric_rtt(dst, RTAX_RTTVAR); 783 if (m >= var) 784 var = m; 785 else 786 var -= (var - m) >> 2; 787 788 set_dst_metric_rtt(dst, RTAX_RTTVAR, var); 789 } 790 791 if (tcp_in_initial_slowstart(tp)) { 792 /* Slow start still did not finish. */ 793 if (dst_metric(dst, RTAX_SSTHRESH) && 794 !dst_metric_locked(dst, RTAX_SSTHRESH) && 795 (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH)) 796 dst_metric_set(dst, RTAX_SSTHRESH, tp->snd_cwnd >> 1); 797 if (!dst_metric_locked(dst, RTAX_CWND) && 798 tp->snd_cwnd > dst_metric(dst, RTAX_CWND)) 799 dst_metric_set(dst, RTAX_CWND, tp->snd_cwnd); 800 } else if (tp->snd_cwnd > tp->snd_ssthresh && 801 icsk->icsk_ca_state == TCP_CA_Open) { 802 /* Cong. avoidance phase, cwnd is reliable. */ 803 if (!dst_metric_locked(dst, RTAX_SSTHRESH)) 804 dst_metric_set(dst, RTAX_SSTHRESH, 805 max(tp->snd_cwnd >> 1, tp->snd_ssthresh)); 806 if (!dst_metric_locked(dst, RTAX_CWND)) 807 dst_metric_set(dst, RTAX_CWND, 808 (dst_metric(dst, RTAX_CWND) + 809 tp->snd_cwnd) >> 1); 810 } else { 811 /* Else slow start did not finish, cwnd is non-sense, 812 ssthresh may be also invalid. 813 */ 814 if (!dst_metric_locked(dst, RTAX_CWND)) 815 dst_metric_set(dst, RTAX_CWND, 816 (dst_metric(dst, RTAX_CWND) + 817 tp->snd_ssthresh) >> 1); 818 if (dst_metric(dst, RTAX_SSTHRESH) && 819 !dst_metric_locked(dst, RTAX_SSTHRESH) && 820 tp->snd_ssthresh > dst_metric(dst, RTAX_SSTHRESH)) 821 dst_metric_set(dst, RTAX_SSTHRESH, tp->snd_ssthresh); 822 } 823 824 if (!dst_metric_locked(dst, RTAX_REORDERING)) { 825 if (dst_metric(dst, RTAX_REORDERING) < tp->reordering && 826 tp->reordering != sysctl_tcp_reordering) 827 dst_metric_set(dst, RTAX_REORDERING, tp->reordering); 828 } 829 } 830} 831 832__u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst) 833{ 834 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0); 835 836 if (!cwnd) 837 cwnd = TCP_INIT_CWND; 838 return min_t(__u32, cwnd, tp->snd_cwnd_clamp); 839} 840 841/* Set slow start threshold and cwnd not falling to slow start */ 842void tcp_enter_cwr(struct sock *sk, const int set_ssthresh) 843{ 844 struct tcp_sock *tp = tcp_sk(sk); 845 const struct inet_connection_sock *icsk = inet_csk(sk); 846 847 tp->prior_ssthresh = 0; 848 tp->bytes_acked = 0; 849 if (icsk->icsk_ca_state < TCP_CA_CWR) { 850 tp->undo_marker = 0; 851 if (set_ssthresh) 852 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 853 tp->snd_cwnd = min(tp->snd_cwnd, 854 tcp_packets_in_flight(tp) + 1U); 855 tp->snd_cwnd_cnt = 0; 856 tp->high_seq = tp->snd_nxt; 857 tp->snd_cwnd_stamp = tcp_time_stamp; 858 TCP_ECN_queue_cwr(tp); 859 860 tcp_set_ca_state(sk, TCP_CA_CWR); 861 } 862} 863 864/* 865 * Packet counting of FACK is based on in-order assumptions, therefore TCP 866 * disables it when reordering is detected 867 */ 868static void tcp_disable_fack(struct tcp_sock *tp) 869{ 870 /* RFC3517 uses different metric in lost marker => reset on change */ 871 if (tcp_is_fack(tp)) 872 tp->lost_skb_hint = NULL; 873 tp->rx_opt.sack_ok &= ~TCP_FACK_ENABLED; 874} 875 876/* Take a notice that peer is sending D-SACKs */ 877static void tcp_dsack_seen(struct tcp_sock *tp) 878{ 879 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN; 880} 881 882/* Initialize metrics on socket. */ 883 884static void tcp_init_metrics(struct sock *sk) 885{ 886 struct tcp_sock *tp = tcp_sk(sk); 887 struct dst_entry *dst = __sk_dst_get(sk); 888 889 if (dst == NULL) 890 goto reset; 891 892 dst_confirm(dst); 893 894 if (dst_metric_locked(dst, RTAX_CWND)) 895 tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND); 896 if (dst_metric(dst, RTAX_SSTHRESH)) { 897 tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH); 898 if (tp->snd_ssthresh > tp->snd_cwnd_clamp) 899 tp->snd_ssthresh = tp->snd_cwnd_clamp; 900 } else { 901 /* ssthresh may have been reduced unnecessarily during. 902 * 3WHS. Restore it back to its initial default. 903 */ 904 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH; 905 } 906 if (dst_metric(dst, RTAX_REORDERING) && 907 tp->reordering != dst_metric(dst, RTAX_REORDERING)) { 908 tcp_disable_fack(tp); 909 tp->reordering = dst_metric(dst, RTAX_REORDERING); 910 } 911 912 if (dst_metric(dst, RTAX_RTT) == 0 || tp->srtt == 0) 913 goto reset; 914 915 /* Initial rtt is determined from SYN,SYN-ACK. 916 * The segment is small and rtt may appear much 917 * less than real one. Use per-dst memory 918 * to make it more realistic. 919 * 920 * A bit of theory. RTT is time passed after "normal" sized packet 921 * is sent until it is ACKed. In normal circumstances sending small 922 * packets force peer to delay ACKs and calculation is correct too. 923 * The algorithm is adaptive and, provided we follow specs, it 924 * NEVER underestimate RTT. BUT! If peer tries to make some clever 925 * tricks sort of "quick acks" for time long enough to decrease RTT 926 * to low value, and then abruptly stops to do it and starts to delay 927 * ACKs, wait for troubles. 928 */ 929 if (dst_metric_rtt(dst, RTAX_RTT) > tp->srtt) { 930 tp->srtt = dst_metric_rtt(dst, RTAX_RTT); 931 tp->rtt_seq = tp->snd_nxt; 932 } 933 if (dst_metric_rtt(dst, RTAX_RTTVAR) > tp->mdev) { 934 tp->mdev = dst_metric_rtt(dst, RTAX_RTTVAR); 935 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk)); 936 } 937 tcp_set_rto(sk); 938reset: 939 if (tp->srtt == 0) { 940 /* RFC2988bis: We've failed to get a valid RTT sample from 941 * 3WHS. This is most likely due to retransmission, 942 * including spurious one. Reset the RTO back to 3secs 943 * from the more aggressive 1sec to avoid more spurious 944 * retransmission. 945 */ 946 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_FALLBACK; 947 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_FALLBACK; 948 } 949 /* Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been 950 * retransmitted. In light of RFC2988bis' more aggressive 1sec 951 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK 952 * retransmission has occurred. 953 */ 954 if (tp->total_retrans > 1) 955 tp->snd_cwnd = 1; 956 else 957 tp->snd_cwnd = tcp_init_cwnd(tp, dst); 958 tp->snd_cwnd_stamp = tcp_time_stamp; 959} 960 961static void tcp_update_reordering(struct sock *sk, const int metric, 962 const int ts) 963{ 964 struct tcp_sock *tp = tcp_sk(sk); 965 if (metric > tp->reordering) { 966 int mib_idx; 967 968 tp->reordering = min(TCP_MAX_REORDERING, metric); 969 970 /* This exciting event is worth to be remembered. 8) */ 971 if (ts) 972 mib_idx = LINUX_MIB_TCPTSREORDER; 973 else if (tcp_is_reno(tp)) 974 mib_idx = LINUX_MIB_TCPRENOREORDER; 975 else if (tcp_is_fack(tp)) 976 mib_idx = LINUX_MIB_TCPFACKREORDER; 977 else 978 mib_idx = LINUX_MIB_TCPSACKREORDER; 979 980 NET_INC_STATS_BH(sock_net(sk), mib_idx); 981#if FASTRETRANS_DEBUG > 1 982 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n", 983 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state, 984 tp->reordering, 985 tp->fackets_out, 986 tp->sacked_out, 987 tp->undo_marker ? tp->undo_retrans : 0); 988#endif 989 tcp_disable_fack(tp); 990 } 991} 992 993/* This must be called before lost_out is incremented */ 994static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb) 995{ 996 if ((tp->retransmit_skb_hint == NULL) || 997 before(TCP_SKB_CB(skb)->seq, 998 TCP_SKB_CB(tp->retransmit_skb_hint)->seq)) 999 tp->retransmit_skb_hint = skb; 1000 1001 if (!tp->lost_out || 1002 after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high)) 1003 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq; 1004} 1005 1006static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb) 1007{ 1008 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) { 1009 tcp_verify_retransmit_hint(tp, skb); 1010 1011 tp->lost_out += tcp_skb_pcount(skb); 1012 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 1013 } 1014} 1015 1016static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp, 1017 struct sk_buff *skb) 1018{ 1019 tcp_verify_retransmit_hint(tp, skb); 1020 1021 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) { 1022 tp->lost_out += tcp_skb_pcount(skb); 1023 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 1024 } 1025} 1026 1027/* This procedure tags the retransmission queue when SACKs arrive. 1028 * 1029 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L). 1030 * Packets in queue with these bits set are counted in variables 1031 * sacked_out, retrans_out and lost_out, correspondingly. 1032 * 1033 * Valid combinations are: 1034 * Tag InFlight Description 1035 * 0 1 - orig segment is in flight. 1036 * S 0 - nothing flies, orig reached receiver. 1037 * L 0 - nothing flies, orig lost by net. 1038 * R 2 - both orig and retransmit are in flight. 1039 * L|R 1 - orig is lost, retransmit is in flight. 1040 * S|R 1 - orig reached receiver, retrans is still in flight. 1041 * (L|S|R is logically valid, it could occur when L|R is sacked, 1042 * but it is equivalent to plain S and code short-curcuits it to S. 1043 * L|S is logically invalid, it would mean -1 packet in flight 8)) 1044 * 1045 * These 6 states form finite state machine, controlled by the following events: 1046 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue()) 1047 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue()) 1048 * 3. Loss detection event of two flavors: 1049 * A. Scoreboard estimator decided the packet is lost. 1050 * A'. Reno "three dupacks" marks head of queue lost. 1051 * A''. Its FACK modification, head until snd.fack is lost. 1052 * B. SACK arrives sacking SND.NXT at the moment, when the 1053 * segment was retransmitted. 1054 * 4. D-SACK added new rule: D-SACK changes any tag to S. 1055 * 1056 * It is pleasant to note, that state diagram turns out to be commutative, 1057 * so that we are allowed not to be bothered by order of our actions, 1058 * when multiple events arrive simultaneously. (see the function below). 1059 * 1060 * Reordering detection. 1061 * -------------------- 1062 * Reordering metric is maximal distance, which a packet can be displaced 1063 * in packet stream. With SACKs we can estimate it: 1064 * 1065 * 1. SACK fills old hole and the corresponding segment was not 1066 * ever retransmitted -> reordering. Alas, we cannot use it 1067 * when segment was retransmitted. 1068 * 2. The last flaw is solved with D-SACK. D-SACK arrives 1069 * for retransmitted and already SACKed segment -> reordering.. 1070 * Both of these heuristics are not used in Loss state, when we cannot 1071 * account for retransmits accurately. 1072 * 1073 * SACK block validation. 1074 * ---------------------- 1075 * 1076 * SACK block range validation checks that the received SACK block fits to 1077 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT. 1078 * Note that SND.UNA is not included to the range though being valid because 1079 * it means that the receiver is rather inconsistent with itself reporting 1080 * SACK reneging when it should advance SND.UNA. Such SACK block this is 1081 * perfectly valid, however, in light of RFC2018 which explicitly states 1082 * that "SACK block MUST reflect the newest segment. Even if the newest 1083 * segment is going to be discarded ...", not that it looks very clever 1084 * in case of head skb. Due to potentional receiver driven attacks, we 1085 * choose to avoid immediate execution of a walk in write queue due to 1086 * reneging and defer head skb's loss recovery to standard loss recovery 1087 * procedure that will eventually trigger (nothing forbids us doing this). 1088 * 1089 * Implements also blockage to start_seq wrap-around. Problem lies in the 1090 * fact that though start_seq (s) is before end_seq (i.e., not reversed), 1091 * there's no guarantee that it will be before snd_nxt (n). The problem 1092 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt 1093 * wrap (s_w): 1094 * 1095 * <- outs wnd -> <- wrapzone -> 1096 * u e n u_w e_w s n_w 1097 * | | | | | | | 1098 * |<------------+------+----- TCP seqno space --------------+---------->| 1099 * ...-- <2^31 ->| |<--------... 1100 * ...---- >2^31 ------>| |<--------... 1101 * 1102 * Current code wouldn't be vulnerable but it's better still to discard such 1103 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat 1104 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in 1105 * snd_nxt wrap -> snd_una region will then become "well defined", i.e., 1106 * equal to the ideal case (infinite seqno space without wrap caused issues). 1107 * 1108 * With D-SACK the lower bound is extended to cover sequence space below 1109 * SND.UNA down to undo_marker, which is the last point of interest. Yet 1110 * again, D-SACK block must not to go across snd_una (for the same reason as 1111 * for the normal SACK blocks, explained above). But there all simplicity 1112 * ends, TCP might receive valid D-SACKs below that. As long as they reside 1113 * fully below undo_marker they do not affect behavior in anyway and can 1114 * therefore be safely ignored. In rare cases (which are more or less 1115 * theoretical ones), the D-SACK will nicely cross that boundary due to skb 1116 * fragmentation and packet reordering past skb's retransmission. To consider 1117 * them correctly, the acceptable range must be extended even more though 1118 * the exact amount is rather hard to quantify. However, tp->max_window can 1119 * be used as an exaggerated estimate. 1120 */ 1121static int tcp_is_sackblock_valid(struct tcp_sock *tp, int is_dsack, 1122 u32 start_seq, u32 end_seq) 1123{ 1124 /* Too far in future, or reversed (interpretation is ambiguous) */ 1125 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq)) 1126 return 0; 1127 1128 /* Nasty start_seq wrap-around check (see comments above) */ 1129 if (!before(start_seq, tp->snd_nxt)) 1130 return 0; 1131 1132 /* In outstanding window? ...This is valid exit for D-SACKs too. 1133 * start_seq == snd_una is non-sensical (see comments above) 1134 */ 1135 if (after(start_seq, tp->snd_una)) 1136 return 1; 1137 1138 if (!is_dsack || !tp->undo_marker) 1139 return 0; 1140 1141 /* ...Then it's D-SACK, and must reside below snd_una completely */ 1142 if (after(end_seq, tp->snd_una)) 1143 return 0; 1144 1145 if (!before(start_seq, tp->undo_marker)) 1146 return 1; 1147 1148 /* Too old */ 1149 if (!after(end_seq, tp->undo_marker)) 1150 return 0; 1151 1152 /* Undo_marker boundary crossing (overestimates a lot). Known already: 1153 * start_seq < undo_marker and end_seq >= undo_marker. 1154 */ 1155 return !before(start_seq, end_seq - tp->max_window); 1156} 1157 1158/* Check for lost retransmit. This superb idea is borrowed from "ratehalving". 1159 * Event "B". Later note: FACK people cheated me again 8), we have to account 1160 * for reordering! Ugly, but should help. 1161 * 1162 * Search retransmitted skbs from write_queue that were sent when snd_nxt was 1163 * less than what is now known to be received by the other end (derived from 1164 * highest SACK block). Also calculate the lowest snd_nxt among the remaining 1165 * retransmitted skbs to avoid some costly processing per ACKs. 1166 */ 1167static void tcp_mark_lost_retrans(struct sock *sk) 1168{ 1169 const struct inet_connection_sock *icsk = inet_csk(sk); 1170 struct tcp_sock *tp = tcp_sk(sk); 1171 struct sk_buff *skb; 1172 int cnt = 0; 1173 u32 new_low_seq = tp->snd_nxt; 1174 u32 received_upto = tcp_highest_sack_seq(tp); 1175 1176 if (!tcp_is_fack(tp) || !tp->retrans_out || 1177 !after(received_upto, tp->lost_retrans_low) || 1178 icsk->icsk_ca_state != TCP_CA_Recovery) 1179 return; 1180 1181 tcp_for_write_queue(skb, sk) { 1182 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq; 1183 1184 if (skb == tcp_send_head(sk)) 1185 break; 1186 if (cnt == tp->retrans_out) 1187 break; 1188 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 1189 continue; 1190 1191 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)) 1192 continue; 1193 1194 /* TODO: We would like to get rid of tcp_is_fack(tp) only 1195 * constraint here (see above) but figuring out that at 1196 * least tp->reordering SACK blocks reside between ack_seq 1197 * and received_upto is not easy task to do cheaply with 1198 * the available datastructures. 1199 * 1200 * Whether FACK should check here for tp->reordering segs 1201 * in-between one could argue for either way (it would be 1202 * rather simple to implement as we could count fack_count 1203 * during the walk and do tp->fackets_out - fack_count). 1204 */ 1205 if (after(received_upto, ack_seq)) { 1206 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 1207 tp->retrans_out -= tcp_skb_pcount(skb); 1208 1209 tcp_skb_mark_lost_uncond_verify(tp, skb); 1210 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT); 1211 } else { 1212 if (before(ack_seq, new_low_seq)) 1213 new_low_seq = ack_seq; 1214 cnt += tcp_skb_pcount(skb); 1215 } 1216 } 1217 1218 if (tp->retrans_out) 1219 tp->lost_retrans_low = new_low_seq; 1220} 1221 1222static int tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb, 1223 struct tcp_sack_block_wire *sp, int num_sacks, 1224 u32 prior_snd_una) 1225{ 1226 struct tcp_sock *tp = tcp_sk(sk); 1227 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq); 1228 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq); 1229 int dup_sack = 0; 1230 1231 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) { 1232 dup_sack = 1; 1233 tcp_dsack_seen(tp); 1234 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV); 1235 } else if (num_sacks > 1) { 1236 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq); 1237 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq); 1238 1239 if (!after(end_seq_0, end_seq_1) && 1240 !before(start_seq_0, start_seq_1)) { 1241 dup_sack = 1; 1242 tcp_dsack_seen(tp); 1243 NET_INC_STATS_BH(sock_net(sk), 1244 LINUX_MIB_TCPDSACKOFORECV); 1245 } 1246 } 1247 1248 /* D-SACK for already forgotten data... Do dumb counting. */ 1249 if (dup_sack && tp->undo_marker && tp->undo_retrans && 1250 !after(end_seq_0, prior_snd_una) && 1251 after(end_seq_0, tp->undo_marker)) 1252 tp->undo_retrans--; 1253 1254 return dup_sack; 1255} 1256 1257struct tcp_sacktag_state { 1258 int reord; 1259 int fack_count; 1260 int flag; 1261}; 1262 1263/* Check if skb is fully within the SACK block. In presence of GSO skbs, 1264 * the incoming SACK may not exactly match but we can find smaller MSS 1265 * aligned portion of it that matches. Therefore we might need to fragment 1266 * which may fail and creates some hassle (caller must handle error case 1267 * returns). 1268 * 1269 * FIXME: this could be merged to shift decision code 1270 */ 1271static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb, 1272 u32 start_seq, u32 end_seq) 1273{ 1274 int in_sack, err; 1275 unsigned int pkt_len; 1276 unsigned int mss; 1277 1278 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && 1279 !before(end_seq, TCP_SKB_CB(skb)->end_seq); 1280 1281 if (tcp_skb_pcount(skb) > 1 && !in_sack && 1282 after(TCP_SKB_CB(skb)->end_seq, start_seq)) { 1283 mss = tcp_skb_mss(skb); 1284 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); 1285 1286 if (!in_sack) { 1287 pkt_len = start_seq - TCP_SKB_CB(skb)->seq; 1288 if (pkt_len < mss) 1289 pkt_len = mss; 1290 } else { 1291 pkt_len = end_seq - TCP_SKB_CB(skb)->seq; 1292 if (pkt_len < mss) 1293 return -EINVAL; 1294 } 1295 1296 /* Round if necessary so that SACKs cover only full MSSes 1297 * and/or the remaining small portion (if present) 1298 */ 1299 if (pkt_len > mss) { 1300 unsigned int new_len = (pkt_len / mss) * mss; 1301 if (!in_sack && new_len < pkt_len) { 1302 new_len += mss; 1303 if (new_len > skb->len) 1304 return 0; 1305 } 1306 pkt_len = new_len; 1307 } 1308 err = tcp_fragment(sk, skb, pkt_len, mss); 1309 if (err < 0) 1310 return err; 1311 } 1312 1313 return in_sack; 1314} 1315 1316/* Mark the given newly-SACKed range as such, adjusting counters and hints. */ 1317static u8 tcp_sacktag_one(struct sock *sk, 1318 struct tcp_sacktag_state *state, u8 sacked, 1319 u32 start_seq, u32 end_seq, 1320 int dup_sack, int pcount) 1321{ 1322 struct tcp_sock *tp = tcp_sk(sk); 1323 int fack_count = state->fack_count; 1324 1325 /* Account D-SACK for retransmitted packet. */ 1326 if (dup_sack && (sacked & TCPCB_RETRANS)) { 1327 if (tp->undo_marker && tp->undo_retrans && 1328 after(end_seq, tp->undo_marker)) 1329 tp->undo_retrans--; 1330 if (sacked & TCPCB_SACKED_ACKED) 1331 state->reord = min(fack_count, state->reord); 1332 } 1333 1334 /* Nothing to do; acked frame is about to be dropped (was ACKed). */ 1335 if (!after(end_seq, tp->snd_una)) 1336 return sacked; 1337 1338 if (!(sacked & TCPCB_SACKED_ACKED)) { 1339 if (sacked & TCPCB_SACKED_RETRANS) { 1340 /* If the segment is not tagged as lost, 1341 * we do not clear RETRANS, believing 1342 * that retransmission is still in flight. 1343 */ 1344 if (sacked & TCPCB_LOST) { 1345 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS); 1346 tp->lost_out -= pcount; 1347 tp->retrans_out -= pcount; 1348 } 1349 } else { 1350 if (!(sacked & TCPCB_RETRANS)) { 1351 /* New sack for not retransmitted frame, 1352 * which was in hole. It is reordering. 1353 */ 1354 if (before(start_seq, 1355 tcp_highest_sack_seq(tp))) 1356 state->reord = min(fack_count, 1357 state->reord); 1358 1359 /* SACK enhanced F-RTO (RFC4138; Appendix B) */ 1360 if (!after(end_seq, tp->frto_highmark)) 1361 state->flag |= FLAG_ONLY_ORIG_SACKED; 1362 } 1363 1364 if (sacked & TCPCB_LOST) { 1365 sacked &= ~TCPCB_LOST; 1366 tp->lost_out -= pcount; 1367 } 1368 } 1369 1370 sacked |= TCPCB_SACKED_ACKED; 1371 state->flag |= FLAG_DATA_SACKED; 1372 tp->sacked_out += pcount; 1373 1374 fack_count += pcount; 1375 1376 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */ 1377 if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) && 1378 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq)) 1379 tp->lost_cnt_hint += pcount; 1380 1381 if (fack_count > tp->fackets_out) 1382 tp->fackets_out = fack_count; 1383 } 1384 1385 /* D-SACK. We can detect redundant retransmission in S|R and plain R 1386 * frames and clear it. undo_retrans is decreased above, L|R frames 1387 * are accounted above as well. 1388 */ 1389 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) { 1390 sacked &= ~TCPCB_SACKED_RETRANS; 1391 tp->retrans_out -= pcount; 1392 } 1393 1394 return sacked; 1395} 1396 1397/* Shift newly-SACKed bytes from this skb to the immediately previous 1398 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such. 1399 */ 1400static int tcp_shifted_skb(struct sock *sk, struct sk_buff *skb, 1401 struct tcp_sacktag_state *state, 1402 unsigned int pcount, int shifted, int mss, 1403 int dup_sack) 1404{ 1405 struct tcp_sock *tp = tcp_sk(sk); 1406 struct sk_buff *prev = tcp_write_queue_prev(sk, skb); 1407 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */ 1408 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */ 1409 1410 BUG_ON(!pcount); 1411 1412 /* Adjust counters and hints for the newly sacked sequence 1413 * range but discard the return value since prev is already 1414 * marked. We must tag the range first because the seq 1415 * advancement below implicitly advances 1416 * tcp_highest_sack_seq() when skb is highest_sack. 1417 */ 1418 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked, 1419 start_seq, end_seq, dup_sack, pcount); 1420 1421 if (skb == tp->lost_skb_hint) 1422 tp->lost_cnt_hint += pcount; 1423 1424 TCP_SKB_CB(prev)->end_seq += shifted; 1425 TCP_SKB_CB(skb)->seq += shifted; 1426 1427 skb_shinfo(prev)->gso_segs += pcount; 1428 BUG_ON(skb_shinfo(skb)->gso_segs < pcount); 1429 skb_shinfo(skb)->gso_segs -= pcount; 1430 1431 /* When we're adding to gso_segs == 1, gso_size will be zero, 1432 * in theory this shouldn't be necessary but as long as DSACK 1433 * code can come after this skb later on it's better to keep 1434 * setting gso_size to something. 1435 */ 1436 if (!skb_shinfo(prev)->gso_size) { 1437 skb_shinfo(prev)->gso_size = mss; 1438 skb_shinfo(prev)->gso_type = sk->sk_gso_type; 1439 } 1440 1441 /* CHECKME: To clear or not to clear? Mimics normal skb currently */ 1442 if (skb_shinfo(skb)->gso_segs <= 1) { 1443 skb_shinfo(skb)->gso_size = 0; 1444 skb_shinfo(skb)->gso_type = 0; 1445 } 1446 1447 /* Difference in this won't matter, both ACKed by the same cumul. ACK */ 1448 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS); 1449 1450 if (skb->len > 0) { 1451 BUG_ON(!tcp_skb_pcount(skb)); 1452 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED); 1453 return 0; 1454 } 1455 1456 /* Whole SKB was eaten :-) */ 1457 1458 if (skb == tp->retransmit_skb_hint) 1459 tp->retransmit_skb_hint = prev; 1460 if (skb == tp->scoreboard_skb_hint) 1461 tp->scoreboard_skb_hint = prev; 1462 if (skb == tp->lost_skb_hint) { 1463 tp->lost_skb_hint = prev; 1464 tp->lost_cnt_hint -= tcp_skb_pcount(prev); 1465 } 1466 1467 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(prev)->tcp_flags; 1468 if (skb == tcp_highest_sack(sk)) 1469 tcp_advance_highest_sack(sk, skb); 1470 1471 tcp_unlink_write_queue(skb, sk); 1472 sk_wmem_free_skb(sk, skb); 1473 1474 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED); 1475 1476 return 1; 1477} 1478 1479/* I wish gso_size would have a bit more sane initialization than 1480 * something-or-zero which complicates things 1481 */ 1482static int tcp_skb_seglen(const struct sk_buff *skb) 1483{ 1484 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb); 1485} 1486 1487/* Shifting pages past head area doesn't work */ 1488static int skb_can_shift(const struct sk_buff *skb) 1489{ 1490 return !skb_headlen(skb) && skb_is_nonlinear(skb); 1491} 1492 1493/* Try collapsing SACK blocks spanning across multiple skbs to a single 1494 * skb. 1495 */ 1496static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb, 1497 struct tcp_sacktag_state *state, 1498 u32 start_seq, u32 end_seq, 1499 int dup_sack) 1500{ 1501 struct tcp_sock *tp = tcp_sk(sk); 1502 struct sk_buff *prev; 1503 int mss; 1504 int pcount = 0; 1505 int len; 1506 int in_sack; 1507 1508 if (!sk_can_gso(sk)) 1509 goto fallback; 1510 1511 /* Normally R but no L won't result in plain S */ 1512 if (!dup_sack && 1513 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS) 1514 goto fallback; 1515 if (!skb_can_shift(skb)) 1516 goto fallback; 1517 /* This frame is about to be dropped (was ACKed). */ 1518 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 1519 goto fallback; 1520 1521 /* Can only happen with delayed DSACK + discard craziness */ 1522 if (unlikely(skb == tcp_write_queue_head(sk))) 1523 goto fallback; 1524 prev = tcp_write_queue_prev(sk, skb); 1525 1526 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) 1527 goto fallback; 1528 1529 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && 1530 !before(end_seq, TCP_SKB_CB(skb)->end_seq); 1531 1532 if (in_sack) { 1533 len = skb->len; 1534 pcount = tcp_skb_pcount(skb); 1535 mss = tcp_skb_seglen(skb); 1536 1537 /* TODO: Fix DSACKs to not fragment already SACKed and we can 1538 * drop this restriction as unnecessary 1539 */ 1540 if (mss != tcp_skb_seglen(prev)) 1541 goto fallback; 1542 } else { 1543 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq)) 1544 goto noop; 1545 /* CHECKME: This is non-MSS split case only?, this will 1546 * cause skipped skbs due to advancing loop btw, original 1547 * has that feature too 1548 */ 1549 if (tcp_skb_pcount(skb) <= 1) 1550 goto noop; 1551 1552 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); 1553 if (!in_sack) { 1554 /* TODO: head merge to next could be attempted here 1555 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)), 1556 * though it might not be worth of the additional hassle 1557 * 1558 * ...we can probably just fallback to what was done 1559 * previously. We could try merging non-SACKed ones 1560 * as well but it probably isn't going to buy off 1561 * because later SACKs might again split them, and 1562 * it would make skb timestamp tracking considerably 1563 * harder problem. 1564 */ 1565 goto fallback; 1566 } 1567 1568 len = end_seq - TCP_SKB_CB(skb)->seq; 1569 BUG_ON(len < 0); 1570 BUG_ON(len > skb->len); 1571 1572 /* MSS boundaries should be honoured or else pcount will 1573 * severely break even though it makes things bit trickier. 1574 * Optimize common case to avoid most of the divides 1575 */ 1576 mss = tcp_skb_mss(skb); 1577 1578 /* TODO: Fix DSACKs to not fragment already SACKed and we can 1579 * drop this restriction as unnecessary 1580 */ 1581 if (mss != tcp_skb_seglen(prev)) 1582 goto fallback; 1583 1584 if (len == mss) { 1585 pcount = 1; 1586 } else if (len < mss) { 1587 goto noop; 1588 } else { 1589 pcount = len / mss; 1590 len = pcount * mss; 1591 } 1592 } 1593 1594 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */ 1595 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una)) 1596 goto fallback; 1597 1598 if (!skb_shift(prev, skb, len)) 1599 goto fallback; 1600 if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack)) 1601 goto out; 1602 1603 /* Hole filled allows collapsing with the next as well, this is very 1604 * useful when hole on every nth skb pattern happens 1605 */ 1606 if (prev == tcp_write_queue_tail(sk)) 1607 goto out; 1608 skb = tcp_write_queue_next(sk, prev); 1609 1610 if (!skb_can_shift(skb) || 1611 (skb == tcp_send_head(sk)) || 1612 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) || 1613 (mss != tcp_skb_seglen(skb))) 1614 goto out; 1615 1616 len = skb->len; 1617 if (skb_shift(prev, skb, len)) { 1618 pcount += tcp_skb_pcount(skb); 1619 tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0); 1620 } 1621 1622out: 1623 state->fack_count += pcount; 1624 return prev; 1625 1626noop: 1627 return skb; 1628 1629fallback: 1630 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK); 1631 return NULL; 1632} 1633 1634static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk, 1635 struct tcp_sack_block *next_dup, 1636 struct tcp_sacktag_state *state, 1637 u32 start_seq, u32 end_seq, 1638 int dup_sack_in) 1639{ 1640 struct tcp_sock *tp = tcp_sk(sk); 1641 struct sk_buff *tmp; 1642 1643 tcp_for_write_queue_from(skb, sk) { 1644 int in_sack = 0; 1645 int dup_sack = dup_sack_in; 1646 1647 if (skb == tcp_send_head(sk)) 1648 break; 1649 1650 /* queue is in-order => we can short-circuit the walk early */ 1651 if (!before(TCP_SKB_CB(skb)->seq, end_seq)) 1652 break; 1653 1654 if ((next_dup != NULL) && 1655 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) { 1656 in_sack = tcp_match_skb_to_sack(sk, skb, 1657 next_dup->start_seq, 1658 next_dup->end_seq); 1659 if (in_sack > 0) 1660 dup_sack = 1; 1661 } 1662 1663 /* skb reference here is a bit tricky to get right, since 1664 * shifting can eat and free both this skb and the next, 1665 * so not even _safe variant of the loop is enough. 1666 */ 1667 if (in_sack <= 0) { 1668 tmp = tcp_shift_skb_data(sk, skb, state, 1669 start_seq, end_seq, dup_sack); 1670 if (tmp != NULL) { 1671 if (tmp != skb) { 1672 skb = tmp; 1673 continue; 1674 } 1675 1676 in_sack = 0; 1677 } else { 1678 in_sack = tcp_match_skb_to_sack(sk, skb, 1679 start_seq, 1680 end_seq); 1681 } 1682 } 1683 1684 if (unlikely(in_sack < 0)) 1685 break; 1686 1687 if (in_sack) { 1688 TCP_SKB_CB(skb)->sacked = 1689 tcp_sacktag_one(sk, 1690 state, 1691 TCP_SKB_CB(skb)->sacked, 1692 TCP_SKB_CB(skb)->seq, 1693 TCP_SKB_CB(skb)->end_seq, 1694 dup_sack, 1695 tcp_skb_pcount(skb)); 1696 1697 if (!before(TCP_SKB_CB(skb)->seq, 1698 tcp_highest_sack_seq(tp))) 1699 tcp_advance_highest_sack(sk, skb); 1700 } 1701 1702 state->fack_count += tcp_skb_pcount(skb); 1703 } 1704 return skb; 1705} 1706 1707/* Avoid all extra work that is being done by sacktag while walking in 1708 * a normal way 1709 */ 1710static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk, 1711 struct tcp_sacktag_state *state, 1712 u32 skip_to_seq) 1713{ 1714 tcp_for_write_queue_from(skb, sk) { 1715 if (skb == tcp_send_head(sk)) 1716 break; 1717 1718 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq)) 1719 break; 1720 1721 state->fack_count += tcp_skb_pcount(skb); 1722 } 1723 return skb; 1724} 1725 1726static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb, 1727 struct sock *sk, 1728 struct tcp_sack_block *next_dup, 1729 struct tcp_sacktag_state *state, 1730 u32 skip_to_seq) 1731{ 1732 if (next_dup == NULL) 1733 return skb; 1734 1735 if (before(next_dup->start_seq, skip_to_seq)) { 1736 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq); 1737 skb = tcp_sacktag_walk(skb, sk, NULL, state, 1738 next_dup->start_seq, next_dup->end_seq, 1739 1); 1740 } 1741 1742 return skb; 1743} 1744 1745static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache) 1746{ 1747 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache); 1748} 1749 1750static int 1751tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb, 1752 u32 prior_snd_una) 1753{ 1754 const struct inet_connection_sock *icsk = inet_csk(sk); 1755 struct tcp_sock *tp = tcp_sk(sk); 1756 const unsigned char *ptr = (skb_transport_header(ack_skb) + 1757 TCP_SKB_CB(ack_skb)->sacked); 1758 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2); 1759 struct tcp_sack_block sp[TCP_NUM_SACKS]; 1760 struct tcp_sack_block *cache; 1761 struct tcp_sacktag_state state; 1762 struct sk_buff *skb; 1763 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3); 1764 int used_sacks; 1765 int found_dup_sack = 0; 1766 int i, j; 1767 int first_sack_index; 1768 1769 state.flag = 0; 1770 state.reord = tp->packets_out; 1771 1772 if (!tp->sacked_out) { 1773 if (WARN_ON(tp->fackets_out)) 1774 tp->fackets_out = 0; 1775 tcp_highest_sack_reset(sk); 1776 } 1777 1778 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire, 1779 num_sacks, prior_snd_una); 1780 if (found_dup_sack) 1781 state.flag |= FLAG_DSACKING_ACK; 1782 1783 /* Eliminate too old ACKs, but take into 1784 * account more or less fresh ones, they can 1785 * contain valid SACK info. 1786 */ 1787 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window)) 1788 return 0; 1789 1790 if (!tp->packets_out) 1791 goto out; 1792 1793 used_sacks = 0; 1794 first_sack_index = 0; 1795 for (i = 0; i < num_sacks; i++) { 1796 int dup_sack = !i && found_dup_sack; 1797 1798 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq); 1799 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq); 1800 1801 if (!tcp_is_sackblock_valid(tp, dup_sack, 1802 sp[used_sacks].start_seq, 1803 sp[used_sacks].end_seq)) { 1804 int mib_idx; 1805 1806 if (dup_sack) { 1807 if (!tp->undo_marker) 1808 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO; 1809 else 1810 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD; 1811 } else { 1812 /* Don't count olds caused by ACK reordering */ 1813 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) && 1814 !after(sp[used_sacks].end_seq, tp->snd_una)) 1815 continue; 1816 mib_idx = LINUX_MIB_TCPSACKDISCARD; 1817 } 1818 1819 NET_INC_STATS_BH(sock_net(sk), mib_idx); 1820 if (i == 0) 1821 first_sack_index = -1; 1822 continue; 1823 } 1824 1825 /* Ignore very old stuff early */ 1826 if (!after(sp[used_sacks].end_seq, prior_snd_una)) 1827 continue; 1828 1829 used_sacks++; 1830 } 1831 1832 /* order SACK blocks to allow in order walk of the retrans queue */ 1833 for (i = used_sacks - 1; i > 0; i--) { 1834 for (j = 0; j < i; j++) { 1835 if (after(sp[j].start_seq, sp[j + 1].start_seq)) { 1836 swap(sp[j], sp[j + 1]); 1837 1838 /* Track where the first SACK block goes to */ 1839 if (j == first_sack_index) 1840 first_sack_index = j + 1; 1841 } 1842 } 1843 } 1844 1845 skb = tcp_write_queue_head(sk); 1846 state.fack_count = 0; 1847 i = 0; 1848 1849 if (!tp->sacked_out) { 1850 /* It's already past, so skip checking against it */ 1851 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache); 1852 } else { 1853 cache = tp->recv_sack_cache; 1854 /* Skip empty blocks in at head of the cache */ 1855 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq && 1856 !cache->end_seq) 1857 cache++; 1858 } 1859 1860 while (i < used_sacks) { 1861 u32 start_seq = sp[i].start_seq; 1862 u32 end_seq = sp[i].end_seq; 1863 int dup_sack = (found_dup_sack && (i == first_sack_index)); 1864 struct tcp_sack_block *next_dup = NULL; 1865 1866 if (found_dup_sack && ((i + 1) == first_sack_index)) 1867 next_dup = &sp[i + 1]; 1868 1869 /* Skip too early cached blocks */ 1870 while (tcp_sack_cache_ok(tp, cache) && 1871 !before(start_seq, cache->end_seq)) 1872 cache++; 1873 1874 /* Can skip some work by looking recv_sack_cache? */ 1875 if (tcp_sack_cache_ok(tp, cache) && !dup_sack && 1876 after(end_seq, cache->start_seq)) { 1877 1878 /* Head todo? */ 1879 if (before(start_seq, cache->start_seq)) { 1880 skb = tcp_sacktag_skip(skb, sk, &state, 1881 start_seq); 1882 skb = tcp_sacktag_walk(skb, sk, next_dup, 1883 &state, 1884 start_seq, 1885 cache->start_seq, 1886 dup_sack); 1887 } 1888 1889 /* Rest of the block already fully processed? */ 1890 if (!after(end_seq, cache->end_seq)) 1891 goto advance_sp; 1892 1893 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup, 1894 &state, 1895 cache->end_seq); 1896 1897 /* ...tail remains todo... */ 1898 if (tcp_highest_sack_seq(tp) == cache->end_seq) { 1899 /* ...but better entrypoint exists! */ 1900 skb = tcp_highest_sack(sk); 1901 if (skb == NULL) 1902 break; 1903 state.fack_count = tp->fackets_out; 1904 cache++; 1905 goto walk; 1906 } 1907 1908 skb = tcp_sacktag_skip(skb, sk, &state, cache->end_seq); 1909 /* Check overlap against next cached too (past this one already) */ 1910 cache++; 1911 continue; 1912 } 1913 1914 if (!before(start_seq, tcp_highest_sack_seq(tp))) { 1915 skb = tcp_highest_sack(sk); 1916 if (skb == NULL) 1917 break; 1918 state.fack_count = tp->fackets_out; 1919 } 1920 skb = tcp_sacktag_skip(skb, sk, &state, start_seq); 1921 1922walk: 1923 skb = tcp_sacktag_walk(skb, sk, next_dup, &state, 1924 start_seq, end_seq, dup_sack); 1925 1926advance_sp: 1927 /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct 1928 * due to in-order walk 1929 */ 1930 if (after(end_seq, tp->frto_highmark)) 1931 state.flag &= ~FLAG_ONLY_ORIG_SACKED; 1932 1933 i++; 1934 } 1935 1936 /* Clear the head of the cache sack blocks so we can skip it next time */ 1937 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) { 1938 tp->recv_sack_cache[i].start_seq = 0; 1939 tp->recv_sack_cache[i].end_seq = 0; 1940 } 1941 for (j = 0; j < used_sacks; j++) 1942 tp->recv_sack_cache[i++] = sp[j]; 1943 1944 tcp_mark_lost_retrans(sk); 1945 1946 tcp_verify_left_out(tp); 1947 1948 if ((state.reord < tp->fackets_out) && 1949 ((icsk->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker) && 1950 (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark))) 1951 tcp_update_reordering(sk, tp->fackets_out - state.reord, 0); 1952 1953out: 1954 1955#if FASTRETRANS_DEBUG > 0 1956 WARN_ON((int)tp->sacked_out < 0); 1957 WARN_ON((int)tp->lost_out < 0); 1958 WARN_ON((int)tp->retrans_out < 0); 1959 WARN_ON((int)tcp_packets_in_flight(tp) < 0); 1960#endif 1961 return state.flag; 1962} 1963 1964/* Limits sacked_out so that sum with lost_out isn't ever larger than 1965 * packets_out. Returns zero if sacked_out adjustement wasn't necessary. 1966 */ 1967static int tcp_limit_reno_sacked(struct tcp_sock *tp) 1968{ 1969 u32 holes; 1970 1971 holes = max(tp->lost_out, 1U); 1972 holes = min(holes, tp->packets_out); 1973 1974 if ((tp->sacked_out + holes) > tp->packets_out) { 1975 tp->sacked_out = tp->packets_out - holes; 1976 return 1; 1977 } 1978 return 0; 1979} 1980 1981/* If we receive more dupacks than we expected counting segments 1982 * in assumption of absent reordering, interpret this as reordering. 1983 * The only another reason could be bug in receiver TCP. 1984 */ 1985static void tcp_check_reno_reordering(struct sock *sk, const int addend) 1986{ 1987 struct tcp_sock *tp = tcp_sk(sk); 1988 if (tcp_limit_reno_sacked(tp)) 1989 tcp_update_reordering(sk, tp->packets_out + addend, 0); 1990} 1991 1992/* Emulate SACKs for SACKless connection: account for a new dupack. */ 1993 1994static void tcp_add_reno_sack(struct sock *sk) 1995{ 1996 struct tcp_sock *tp = tcp_sk(sk); 1997 tp->sacked_out++; 1998 tcp_check_reno_reordering(sk, 0); 1999 tcp_verify_left_out(tp); 2000} 2001 2002/* Account for ACK, ACKing some data in Reno Recovery phase. */ 2003 2004static void tcp_remove_reno_sacks(struct sock *sk, int acked) 2005{ 2006 struct tcp_sock *tp = tcp_sk(sk); 2007 2008 if (acked > 0) { 2009 /* One ACK acked hole. The rest eat duplicate ACKs. */ 2010 if (acked - 1 >= tp->sacked_out) 2011 tp->sacked_out = 0; 2012 else 2013 tp->sacked_out -= acked - 1; 2014 } 2015 tcp_check_reno_reordering(sk, acked); 2016 tcp_verify_left_out(tp); 2017} 2018 2019static inline void tcp_reset_reno_sack(struct tcp_sock *tp) 2020{ 2021 tp->sacked_out = 0; 2022} 2023 2024static int tcp_is_sackfrto(const struct tcp_sock *tp) 2025{ 2026 return (sysctl_tcp_frto == 0x2) && !tcp_is_reno(tp); 2027} 2028 2029/* F-RTO can only be used if TCP has never retransmitted anything other than 2030 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here) 2031 */ 2032int tcp_use_frto(struct sock *sk) 2033{ 2034 const struct tcp_sock *tp = tcp_sk(sk); 2035 const struct inet_connection_sock *icsk = inet_csk(sk); 2036 struct sk_buff *skb; 2037 2038 if (!sysctl_tcp_frto) 2039 return 0; 2040 2041 /* MTU probe and F-RTO won't really play nicely along currently */ 2042 if (icsk->icsk_mtup.probe_size) 2043 return 0; 2044 2045 if (tcp_is_sackfrto(tp)) 2046 return 1; 2047 2048 /* Avoid expensive walking of rexmit queue if possible */ 2049 if (tp->retrans_out > 1) 2050 return 0; 2051 2052 skb = tcp_write_queue_head(sk); 2053 if (tcp_skb_is_last(sk, skb)) 2054 return 1; 2055 skb = tcp_write_queue_next(sk, skb); /* Skips head */ 2056 tcp_for_write_queue_from(skb, sk) { 2057 if (skb == tcp_send_head(sk)) 2058 break; 2059 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) 2060 return 0; 2061 /* Short-circuit when first non-SACKed skb has been checked */ 2062 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) 2063 break; 2064 } 2065 return 1; 2066} 2067 2068/* RTO occurred, but do not yet enter Loss state. Instead, defer RTO 2069 * recovery a bit and use heuristics in tcp_process_frto() to detect if 2070 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to 2071 * keep retrans_out counting accurate (with SACK F-RTO, other than head 2072 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS 2073 * bits are handled if the Loss state is really to be entered (in 2074 * tcp_enter_frto_loss). 2075 * 2076 * Do like tcp_enter_loss() would; when RTO expires the second time it 2077 * does: 2078 * "Reduce ssthresh if it has not yet been made inside this window." 2079 */ 2080void tcp_enter_frto(struct sock *sk) 2081{ 2082 const struct inet_connection_sock *icsk = inet_csk(sk); 2083 struct tcp_sock *tp = tcp_sk(sk); 2084 struct sk_buff *skb; 2085 2086 if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) || 2087 tp->snd_una == tp->high_seq || 2088 ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) && 2089 !icsk->icsk_retransmits)) { 2090 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2091 /* Our state is too optimistic in ssthresh() call because cwnd 2092 * is not reduced until tcp_enter_frto_loss() when previous F-RTO 2093 * recovery has not yet completed. Pattern would be this: RTO, 2094 * Cumulative ACK, RTO (2xRTO for the same segment does not end 2095 * up here twice). 2096 * RFC4138 should be more specific on what to do, even though 2097 * RTO is quite unlikely to occur after the first Cumulative ACK 2098 * due to back-off and complexity of triggering events ... 2099 */ 2100 if (tp->frto_counter) { 2101 u32 stored_cwnd; 2102 stored_cwnd = tp->snd_cwnd; 2103 tp->snd_cwnd = 2; 2104 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 2105 tp->snd_cwnd = stored_cwnd; 2106 } else { 2107 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 2108 } 2109 /* ... in theory, cong.control module could do "any tricks" in 2110 * ssthresh(), which means that ca_state, lost bits and lost_out 2111 * counter would have to be faked before the call occurs. We 2112 * consider that too expensive, unlikely and hacky, so modules 2113 * using these in ssthresh() must deal these incompatibility 2114 * issues if they receives CA_EVENT_FRTO and frto_counter != 0 2115 */ 2116 tcp_ca_event(sk, CA_EVENT_FRTO); 2117 } 2118 2119 tp->undo_marker = tp->snd_una; 2120 tp->undo_retrans = 0; 2121 2122 skb = tcp_write_queue_head(sk); 2123 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) 2124 tp->undo_marker = 0; 2125 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 2126 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 2127 tp->retrans_out -= tcp_skb_pcount(skb); 2128 } 2129 tcp_verify_left_out(tp); 2130 2131 /* Too bad if TCP was application limited */ 2132 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1); 2133 2134 /* Earlier loss recovery underway (see RFC4138; Appendix B). 2135 * The last condition is necessary at least in tp->frto_counter case. 2136 */ 2137 if (tcp_is_sackfrto(tp) && (tp->frto_counter || 2138 ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) && 2139 after(tp->high_seq, tp->snd_una)) { 2140 tp->frto_highmark = tp->high_seq; 2141 } else { 2142 tp->frto_highmark = tp->snd_nxt; 2143 } 2144 tcp_set_ca_state(sk, TCP_CA_Disorder); 2145 tp->high_seq = tp->snd_nxt; 2146 tp->frto_counter = 1; 2147} 2148 2149/* Enter Loss state after F-RTO was applied. Dupack arrived after RTO, 2150 * which indicates that we should follow the traditional RTO recovery, 2151 * i.e. mark everything lost and do go-back-N retransmission. 2152 */ 2153static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag) 2154{ 2155 struct tcp_sock *tp = tcp_sk(sk); 2156 struct sk_buff *skb; 2157 2158 tp->lost_out = 0; 2159 tp->retrans_out = 0; 2160 if (tcp_is_reno(tp)) 2161 tcp_reset_reno_sack(tp); 2162 2163 tcp_for_write_queue(skb, sk) { 2164 if (skb == tcp_send_head(sk)) 2165 break; 2166 2167 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; 2168 /* 2169 * Count the retransmission made on RTO correctly (only when 2170 * waiting for the first ACK and did not get it)... 2171 */ 2172 if ((tp->frto_counter == 1) && !(flag & FLAG_DATA_ACKED)) { 2173 /* For some reason this R-bit might get cleared? */ 2174 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 2175 tp->retrans_out += tcp_skb_pcount(skb); 2176 /* ...enter this if branch just for the first segment */ 2177 flag |= FLAG_DATA_ACKED; 2178 } else { 2179 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) 2180 tp->undo_marker = 0; 2181 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 2182 } 2183 2184 /* Marking forward transmissions that were made after RTO lost 2185 * can cause unnecessary retransmissions in some scenarios, 2186 * SACK blocks will mitigate that in some but not in all cases. 2187 * We used to not mark them but it was causing break-ups with 2188 * receivers that do only in-order receival. 2189 * 2190 * TODO: we could detect presence of such receiver and select 2191 * different behavior per flow. 2192 */ 2193 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) { 2194 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 2195 tp->lost_out += tcp_skb_pcount(skb); 2196 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq; 2197 } 2198 } 2199 tcp_verify_left_out(tp); 2200 2201 tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments; 2202 tp->snd_cwnd_cnt = 0; 2203 tp->snd_cwnd_stamp = tcp_time_stamp; 2204 tp->frto_counter = 0; 2205 tp->bytes_acked = 0; 2206 2207 tp->reordering = min_t(unsigned int, tp->reordering, 2208 sysctl_tcp_reordering); 2209 tcp_set_ca_state(sk, TCP_CA_Loss); 2210 tp->high_seq = tp->snd_nxt; 2211 TCP_ECN_queue_cwr(tp); 2212 2213 tcp_clear_all_retrans_hints(tp); 2214} 2215 2216static void tcp_clear_retrans_partial(struct tcp_sock *tp) 2217{ 2218 tp->retrans_out = 0; 2219 tp->lost_out = 0; 2220 2221 tp->undo_marker = 0; 2222 tp->undo_retrans = 0; 2223} 2224 2225void tcp_clear_retrans(struct tcp_sock *tp) 2226{ 2227 tcp_clear_retrans_partial(tp); 2228 2229 tp->fackets_out = 0; 2230 tp->sacked_out = 0; 2231} 2232 2233/* Enter Loss state. If "how" is not zero, forget all SACK information 2234 * and reset tags completely, otherwise preserve SACKs. If receiver 2235 * dropped its ofo queue, we will know this due to reneging detection. 2236 */ 2237void tcp_enter_loss(struct sock *sk, int how) 2238{ 2239 const struct inet_connection_sock *icsk = inet_csk(sk); 2240 struct tcp_sock *tp = tcp_sk(sk); 2241 struct sk_buff *skb; 2242 2243 /* Reduce ssthresh if it has not yet been made inside this window. */ 2244 if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq || 2245 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) { 2246 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2247 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 2248 tcp_ca_event(sk, CA_EVENT_LOSS); 2249 } 2250 tp->snd_cwnd = 1; 2251 tp->snd_cwnd_cnt = 0; 2252 tp->snd_cwnd_stamp = tcp_time_stamp; 2253 2254 tp->bytes_acked = 0; 2255 tcp_clear_retrans_partial(tp); 2256 2257 if (tcp_is_reno(tp)) 2258 tcp_reset_reno_sack(tp); 2259 2260 if (!how) { 2261 /* Push undo marker, if it was plain RTO and nothing 2262 * was retransmitted. */ 2263 tp->undo_marker = tp->snd_una; 2264 } else { 2265 tp->sacked_out = 0; 2266 tp->fackets_out = 0; 2267 } 2268 tcp_clear_all_retrans_hints(tp); 2269 2270 tcp_for_write_queue(skb, sk) { 2271 if (skb == tcp_send_head(sk)) 2272 break; 2273 2274 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) 2275 tp->undo_marker = 0; 2276 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED; 2277 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) { 2278 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED; 2279 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 2280 tp->lost_out += tcp_skb_pcount(skb); 2281 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq; 2282 } 2283 } 2284 tcp_verify_left_out(tp); 2285 2286 tp->reordering = min_t(unsigned int, tp->reordering, 2287 sysctl_tcp_reordering); 2288 tcp_set_ca_state(sk, TCP_CA_Loss); 2289 tp->high_seq = tp->snd_nxt; 2290 TCP_ECN_queue_cwr(tp); 2291 /* Abort F-RTO algorithm if one is in progress */ 2292 tp->frto_counter = 0; 2293} 2294 2295/* If ACK arrived pointing to a remembered SACK, it means that our 2296 * remembered SACKs do not reflect real state of receiver i.e. 2297 * receiver _host_ is heavily congested (or buggy). 2298 * 2299 * Do processing similar to RTO timeout. 2300 */ 2301static int tcp_check_sack_reneging(struct sock *sk, int flag) 2302{ 2303 if (flag & FLAG_SACK_RENEGING) { 2304 struct inet_connection_sock *icsk = inet_csk(sk); 2305 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING); 2306 2307 tcp_enter_loss(sk, 1); 2308 icsk->icsk_retransmits++; 2309 tcp_retransmit_skb(sk, tcp_write_queue_head(sk)); 2310 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 2311 icsk->icsk_rto, TCP_RTO_MAX); 2312 return 1; 2313 } 2314 return 0; 2315} 2316 2317static inline int tcp_fackets_out(const struct tcp_sock *tp) 2318{ 2319 return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out; 2320} 2321 2322/* Heurestics to calculate number of duplicate ACKs. There's no dupACKs 2323 * counter when SACK is enabled (without SACK, sacked_out is used for 2324 * that purpose). 2325 * 2326 * Instead, with FACK TCP uses fackets_out that includes both SACKed 2327 * segments up to the highest received SACK block so far and holes in 2328 * between them. 2329 * 2330 * With reordering, holes may still be in flight, so RFC3517 recovery 2331 * uses pure sacked_out (total number of SACKed segments) even though 2332 * it violates the RFC that uses duplicate ACKs, often these are equal 2333 * but when e.g. out-of-window ACKs or packet duplication occurs, 2334 * they differ. Since neither occurs due to loss, TCP should really 2335 * ignore them. 2336 */ 2337static inline int tcp_dupack_heuristics(const struct tcp_sock *tp) 2338{ 2339 return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1; 2340} 2341 2342static inline int tcp_skb_timedout(const struct sock *sk, 2343 const struct sk_buff *skb) 2344{ 2345 return tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto; 2346} 2347 2348static inline int tcp_head_timedout(const struct sock *sk) 2349{ 2350 const struct tcp_sock *tp = tcp_sk(sk); 2351 2352 return tp->packets_out && 2353 tcp_skb_timedout(sk, tcp_write_queue_head(sk)); 2354} 2355 2356/* Linux NewReno/SACK/FACK/ECN state machine. 2357 * -------------------------------------- 2358 * 2359 * "Open" Normal state, no dubious events, fast path. 2360 * "Disorder" In all the respects it is "Open", 2361 * but requires a bit more attention. It is entered when 2362 * we see some SACKs or dupacks. It is split of "Open" 2363 * mainly to move some processing from fast path to slow one. 2364 * "CWR" CWND was reduced due to some Congestion Notification event. 2365 * It can be ECN, ICMP source quench, local device congestion. 2366 * "Recovery" CWND was reduced, we are fast-retransmitting. 2367 * "Loss" CWND was reduced due to RTO timeout or SACK reneging. 2368 * 2369 * tcp_fastretrans_alert() is entered: 2370 * - each incoming ACK, if state is not "Open" 2371 * - when arrived ACK is unusual, namely: 2372 * * SACK 2373 * * Duplicate ACK. 2374 * * ECN ECE. 2375 * 2376 * Counting packets in flight is pretty simple. 2377 * 2378 * in_flight = packets_out - left_out + retrans_out 2379 * 2380 * packets_out is SND.NXT-SND.UNA counted in packets. 2381 * 2382 * retrans_out is number of retransmitted segments. 2383 * 2384 * left_out is number of segments left network, but not ACKed yet. 2385 * 2386 * left_out = sacked_out + lost_out 2387 * 2388 * sacked_out: Packets, which arrived to receiver out of order 2389 * and hence not ACKed. With SACKs this number is simply 2390 * amount of SACKed data. Even without SACKs 2391 * it is easy to give pretty reliable estimate of this number, 2392 * counting duplicate ACKs. 2393 * 2394 * lost_out: Packets lost by network. TCP has no explicit 2395 * "loss notification" feedback from network (for now). 2396 * It means that this number can be only _guessed_. 2397 * Actually, it is the heuristics to predict lossage that 2398 * distinguishes different algorithms. 2399 * 2400 * F.e. after RTO, when all the queue is considered as lost, 2401 * lost_out = packets_out and in_flight = retrans_out. 2402 * 2403 * Essentially, we have now two algorithms counting 2404 * lost packets. 2405 * 2406 * FACK: It is the simplest heuristics. As soon as we decided 2407 * that something is lost, we decide that _all_ not SACKed 2408 * packets until the most forward SACK are lost. I.e. 2409 * lost_out = fackets_out - sacked_out and left_out = fackets_out. 2410 * It is absolutely correct estimate, if network does not reorder 2411 * packets. And it loses any connection to reality when reordering 2412 * takes place. We use FACK by default until reordering 2413 * is suspected on the path to this destination. 2414 * 2415 * NewReno: when Recovery is entered, we assume that one segment 2416 * is lost (classic Reno). While we are in Recovery and 2417 * a partial ACK arrives, we assume that one more packet 2418 * is lost (NewReno). This heuristics are the same in NewReno 2419 * and SACK. 2420 * 2421 * Imagine, that's all! Forget about all this shamanism about CWND inflation 2422 * deflation etc. CWND is real congestion window, never inflated, changes 2423 * only according to classic VJ rules. 2424 * 2425 * Really tricky (and requiring careful tuning) part of algorithm 2426 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue(). 2427 * The first determines the moment _when_ we should reduce CWND and, 2428 * hence, slow down forward transmission. In fact, it determines the moment 2429 * when we decide that hole is caused by loss, rather than by a reorder. 2430 * 2431 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill 2432 * holes, caused by lost packets. 2433 * 2434 * And the most logically complicated part of algorithm is undo 2435 * heuristics. We detect false retransmits due to both too early 2436 * fast retransmit (reordering) and underestimated RTO, analyzing 2437 * timestamps and D-SACKs. When we detect that some segments were 2438 * retransmitted by mistake and CWND reduction was wrong, we undo 2439 * window reduction and abort recovery phase. This logic is hidden 2440 * inside several functions named tcp_try_undo_<something>. 2441 */ 2442 2443/* This function decides, when we should leave Disordered state 2444 * and enter Recovery phase, reducing congestion window. 2445 * 2446 * Main question: may we further continue forward transmission 2447 * with the same cwnd? 2448 */ 2449static int tcp_time_to_recover(struct sock *sk) 2450{ 2451 struct tcp_sock *tp = tcp_sk(sk); 2452 __u32 packets_out; 2453 2454 /* Do not perform any recovery during F-RTO algorithm */ 2455 if (tp->frto_counter) 2456 return 0; 2457 2458 /* Trick#1: The loss is proven. */ 2459 if (tp->lost_out) 2460 return 1; 2461 2462 /* Not-A-Trick#2 : Classic rule... */ 2463 if (tcp_dupack_heuristics(tp) > tp->reordering) 2464 return 1; 2465 2466 /* Trick#3 : when we use RFC2988 timer restart, fast 2467 * retransmit can be triggered by timeout of queue head. 2468 */ 2469 if (tcp_is_fack(tp) && tcp_head_timedout(sk)) 2470 return 1; 2471 2472 /* Trick#4: It is still not OK... But will it be useful to delay 2473 * recovery more? 2474 */ 2475 packets_out = tp->packets_out; 2476 if (packets_out <= tp->reordering && 2477 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) && 2478 !tcp_may_send_now(sk)) { 2479 /* We have nothing to send. This connection is limited 2480 * either by receiver window or by application. 2481 */ 2482 return 1; 2483 } 2484 2485 /* If a thin stream is detected, retransmit after first 2486 * received dupack. Employ only if SACK is supported in order 2487 * to avoid possible corner-case series of spurious retransmissions 2488 * Use only if there are no unsent data. 2489 */ 2490 if ((tp->thin_dupack || sysctl_tcp_thin_dupack) && 2491 tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 && 2492 tcp_is_sack(tp) && !tcp_send_head(sk)) 2493 return 1; 2494 2495 return 0; 2496} 2497 2498/* New heuristics: it is possible only after we switched to restart timer 2499 * each time when something is ACKed. Hence, we can detect timed out packets 2500 * during fast retransmit without falling to slow start. 2501 * 2502 * Usefulness of this as is very questionable, since we should know which of 2503 * the segments is the next to timeout which is relatively expensive to find 2504 * in general case unless we add some data structure just for that. The 2505 * current approach certainly won't find the right one too often and when it 2506 * finally does find _something_ it usually marks large part of the window 2507 * right away (because a retransmission with a larger timestamp blocks the 2508 * loop from advancing). -ij 2509 */ 2510static void tcp_timeout_skbs(struct sock *sk) 2511{ 2512 struct tcp_sock *tp = tcp_sk(sk); 2513 struct sk_buff *skb; 2514 2515 if (!tcp_is_fack(tp) || !tcp_head_timedout(sk)) 2516 return; 2517 2518 skb = tp->scoreboard_skb_hint; 2519 if (tp->scoreboard_skb_hint == NULL) 2520 skb = tcp_write_queue_head(sk); 2521 2522 tcp_for_write_queue_from(skb, sk) { 2523 if (skb == tcp_send_head(sk)) 2524 break; 2525 if (!tcp_skb_timedout(sk, skb)) 2526 break; 2527 2528 tcp_skb_mark_lost(tp, skb); 2529 } 2530 2531 tp->scoreboard_skb_hint = skb; 2532 2533 tcp_verify_left_out(tp); 2534} 2535 2536/* Detect loss in event "A" above by marking head of queue up as lost. 2537 * For FACK or non-SACK(Reno) senders, the first "packets" number of segments 2538 * are considered lost. For RFC3517 SACK, a segment is considered lost if it 2539 * has at least tp->reordering SACKed seqments above it; "packets" refers to 2540 * the maximum SACKed segments to pass before reaching this limit. 2541 */ 2542static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head) 2543{ 2544 struct tcp_sock *tp = tcp_sk(sk); 2545 struct sk_buff *skb; 2546 int cnt, oldcnt; 2547 int err; 2548 unsigned int mss; 2549 /* Use SACK to deduce losses of new sequences sent during recovery */ 2550 const u32 loss_high = tcp_is_sack(tp) ? tp->snd_nxt : tp->high_seq; 2551 2552 WARN_ON(packets > tp->packets_out); 2553 if (tp->lost_skb_hint) { 2554 skb = tp->lost_skb_hint; 2555 cnt = tp->lost_cnt_hint; 2556 /* Head already handled? */ 2557 if (mark_head && skb != tcp_write_queue_head(sk)) 2558 return; 2559 } else { 2560 skb = tcp_write_queue_head(sk); 2561 cnt = 0; 2562 } 2563 2564 tcp_for_write_queue_from(skb, sk) { 2565 if (skb == tcp_send_head(sk)) 2566 break; 2567 /* TODO: do this better */ 2568 /* this is not the most efficient way to do this... */ 2569 tp->lost_skb_hint = skb; 2570 tp->lost_cnt_hint = cnt; 2571 2572 if (after(TCP_SKB_CB(skb)->end_seq, loss_high)) 2573 break; 2574 2575 oldcnt = cnt; 2576 if (tcp_is_fack(tp) || tcp_is_reno(tp) || 2577 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) 2578 cnt += tcp_skb_pcount(skb); 2579 2580 if (cnt > packets) { 2581 if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) || 2582 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) || 2583 (oldcnt >= packets)) 2584 break; 2585 2586 mss = skb_shinfo(skb)->gso_size; 2587 err = tcp_fragment(sk, skb, (packets - oldcnt) * mss, mss); 2588 if (err < 0) 2589 break; 2590 cnt = packets; 2591 } 2592 2593 tcp_skb_mark_lost(tp, skb); 2594 2595 if (mark_head) 2596 break; 2597 } 2598 tcp_verify_left_out(tp); 2599} 2600 2601/* Account newly detected lost packet(s) */ 2602 2603static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit) 2604{ 2605 struct tcp_sock *tp = tcp_sk(sk); 2606 2607 if (tcp_is_reno(tp)) { 2608 tcp_mark_head_lost(sk, 1, 1); 2609 } else if (tcp_is_fack(tp)) { 2610 int lost = tp->fackets_out - tp->reordering; 2611 if (lost <= 0) 2612 lost = 1; 2613 tcp_mark_head_lost(sk, lost, 0); 2614 } else { 2615 int sacked_upto = tp->sacked_out - tp->reordering; 2616 if (sacked_upto >= 0) 2617 tcp_mark_head_lost(sk, sacked_upto, 0); 2618 else if (fast_rexmit) 2619 tcp_mark_head_lost(sk, 1, 1); 2620 } 2621 2622 tcp_timeout_skbs(sk); 2623} 2624 2625/* CWND moderation, preventing bursts due to too big ACKs 2626 * in dubious situations. 2627 */ 2628static inline void tcp_moderate_cwnd(struct tcp_sock *tp) 2629{ 2630 tp->snd_cwnd = min(tp->snd_cwnd, 2631 tcp_packets_in_flight(tp) + tcp_max_burst(tp)); 2632 tp->snd_cwnd_stamp = tcp_time_stamp; 2633} 2634 2635/* Lower bound on congestion window is slow start threshold 2636 * unless congestion avoidance choice decides to overide it. 2637 */ 2638static inline u32 tcp_cwnd_min(const struct sock *sk) 2639{ 2640 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 2641 2642 return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh; 2643} 2644 2645/* Decrease cwnd each second ack. */ 2646static void tcp_cwnd_down(struct sock *sk, int flag) 2647{ 2648 struct tcp_sock *tp = tcp_sk(sk); 2649 int decr = tp->snd_cwnd_cnt + 1; 2650 2651 if ((flag & (FLAG_ANY_PROGRESS | FLAG_DSACKING_ACK)) || 2652 (tcp_is_reno(tp) && !(flag & FLAG_NOT_DUP))) { 2653 tp->snd_cwnd_cnt = decr & 1; 2654 decr >>= 1; 2655 2656 if (decr && tp->snd_cwnd > tcp_cwnd_min(sk)) 2657 tp->snd_cwnd -= decr; 2658 2659 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1); 2660 tp->snd_cwnd_stamp = tcp_time_stamp; 2661 } 2662} 2663 2664/* Nothing was retransmitted or returned timestamp is less 2665 * than timestamp of the first retransmission. 2666 */ 2667static inline int tcp_packet_delayed(const struct tcp_sock *tp) 2668{ 2669 return !tp->retrans_stamp || 2670 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 2671 before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp)); 2672} 2673 2674/* Undo procedures. */ 2675 2676#if FASTRETRANS_DEBUG > 1 2677static void DBGUNDO(struct sock *sk, const char *msg) 2678{ 2679 struct tcp_sock *tp = tcp_sk(sk); 2680 struct inet_sock *inet = inet_sk(sk); 2681 2682 if (sk->sk_family == AF_INET) { 2683 printk(KERN_DEBUG "Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n", 2684 msg, 2685 &inet->inet_daddr, ntohs(inet->inet_dport), 2686 tp->snd_cwnd, tcp_left_out(tp), 2687 tp->snd_ssthresh, tp->prior_ssthresh, 2688 tp->packets_out); 2689 } 2690#if IS_ENABLED(CONFIG_IPV6) 2691 else if (sk->sk_family == AF_INET6) { 2692 struct ipv6_pinfo *np = inet6_sk(sk); 2693 printk(KERN_DEBUG "Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n", 2694 msg, 2695 &np->daddr, ntohs(inet->inet_dport), 2696 tp->snd_cwnd, tcp_left_out(tp), 2697 tp->snd_ssthresh, tp->prior_ssthresh, 2698 tp->packets_out); 2699 } 2700#endif 2701} 2702#else 2703#define DBGUNDO(x...) do { } while (0) 2704#endif 2705 2706static void tcp_undo_cwr(struct sock *sk, const bool undo_ssthresh) 2707{ 2708 struct tcp_sock *tp = tcp_sk(sk); 2709 2710 if (tp->prior_ssthresh) { 2711 const struct inet_connection_sock *icsk = inet_csk(sk); 2712 2713 if (icsk->icsk_ca_ops->undo_cwnd) 2714 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk); 2715 else 2716 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1); 2717 2718 if (undo_ssthresh && tp->prior_ssthresh > tp->snd_ssthresh) { 2719 tp->snd_ssthresh = tp->prior_ssthresh; 2720 TCP_ECN_withdraw_cwr(tp); 2721 } 2722 } else { 2723 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh); 2724 } 2725 tp->snd_cwnd_stamp = tcp_time_stamp; 2726} 2727 2728static inline int tcp_may_undo(const struct tcp_sock *tp) 2729{ 2730 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp)); 2731} 2732 2733/* People celebrate: "We love our President!" */ 2734static int tcp_try_undo_recovery(struct sock *sk) 2735{ 2736 struct tcp_sock *tp = tcp_sk(sk); 2737 2738 if (tcp_may_undo(tp)) { 2739 int mib_idx; 2740 2741 /* Happy end! We did not retransmit anything 2742 * or our original transmission succeeded. 2743 */ 2744 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans"); 2745 tcp_undo_cwr(sk, true); 2746 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss) 2747 mib_idx = LINUX_MIB_TCPLOSSUNDO; 2748 else 2749 mib_idx = LINUX_MIB_TCPFULLUNDO; 2750 2751 NET_INC_STATS_BH(sock_net(sk), mib_idx); 2752 tp->undo_marker = 0; 2753 } 2754 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) { 2755 /* Hold old state until something *above* high_seq 2756 * is ACKed. For Reno it is MUST to prevent false 2757 * fast retransmits (RFC2582). SACK TCP is safe. */ 2758 tcp_moderate_cwnd(tp); 2759 return 1; 2760 } 2761 tcp_set_ca_state(sk, TCP_CA_Open); 2762 return 0; 2763} 2764 2765/* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */ 2766static void tcp_try_undo_dsack(struct sock *sk) 2767{ 2768 struct tcp_sock *tp = tcp_sk(sk); 2769 2770 if (tp->undo_marker && !tp->undo_retrans) { 2771 DBGUNDO(sk, "D-SACK"); 2772 tcp_undo_cwr(sk, true); 2773 tp->undo_marker = 0; 2774 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO); 2775 } 2776} 2777 2778/* We can clear retrans_stamp when there are no retransmissions in the 2779 * window. It would seem that it is trivially available for us in 2780 * tp->retrans_out, however, that kind of assumptions doesn't consider 2781 * what will happen if errors occur when sending retransmission for the 2782 * second time. ...It could the that such segment has only 2783 * TCPCB_EVER_RETRANS set at the present time. It seems that checking 2784 * the head skb is enough except for some reneging corner cases that 2785 * are not worth the effort. 2786 * 2787 * Main reason for all this complexity is the fact that connection dying 2788 * time now depends on the validity of the retrans_stamp, in particular, 2789 * that successive retransmissions of a segment must not advance 2790 * retrans_stamp under any conditions. 2791 */ 2792static int tcp_any_retrans_done(const struct sock *sk) 2793{ 2794 const struct tcp_sock *tp = tcp_sk(sk); 2795 struct sk_buff *skb; 2796 2797 if (tp->retrans_out) 2798 return 1; 2799 2800 skb = tcp_write_queue_head(sk); 2801 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS)) 2802 return 1; 2803 2804 return 0; 2805} 2806 2807/* Undo during fast recovery after partial ACK. */ 2808 2809static int tcp_try_undo_partial(struct sock *sk, int acked) 2810{ 2811 struct tcp_sock *tp = tcp_sk(sk); 2812 /* Partial ACK arrived. Force Hoe's retransmit. */ 2813 int failed = tcp_is_reno(tp) || (tcp_fackets_out(tp) > tp->reordering); 2814 2815 if (tcp_may_undo(tp)) { 2816 /* Plain luck! Hole if filled with delayed 2817 * packet, rather than with a retransmit. 2818 */ 2819 if (!tcp_any_retrans_done(sk)) 2820 tp->retrans_stamp = 0; 2821 2822 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1); 2823 2824 DBGUNDO(sk, "Hoe"); 2825 tcp_undo_cwr(sk, false); 2826 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO); 2827 2828 /* So... Do not make Hoe's retransmit yet. 2829 * If the first packet was delayed, the rest 2830 * ones are most probably delayed as well. 2831 */ 2832 failed = 0; 2833 } 2834 return failed; 2835} 2836 2837/* Undo during loss recovery after partial ACK. */ 2838static int tcp_try_undo_loss(struct sock *sk) 2839{ 2840 struct tcp_sock *tp = tcp_sk(sk); 2841 2842 if (tcp_may_undo(tp)) { 2843 struct sk_buff *skb; 2844 tcp_for_write_queue(skb, sk) { 2845 if (skb == tcp_send_head(sk)) 2846 break; 2847 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; 2848 } 2849 2850 tcp_clear_all_retrans_hints(tp); 2851 2852 DBGUNDO(sk, "partial loss"); 2853 tp->lost_out = 0; 2854 tcp_undo_cwr(sk, true); 2855 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO); 2856 inet_csk(sk)->icsk_retransmits = 0; 2857 tp->undo_marker = 0; 2858 if (tcp_is_sack(tp)) 2859 tcp_set_ca_state(sk, TCP_CA_Open); 2860 return 1; 2861 } 2862 return 0; 2863} 2864 2865static inline void tcp_complete_cwr(struct sock *sk) 2866{ 2867 struct tcp_sock *tp = tcp_sk(sk); 2868 2869 /* Do not moderate cwnd if it's already undone in cwr or recovery. */ 2870 if (tp->undo_marker) { 2871 if (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR) 2872 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh); 2873 else /* PRR */ 2874 tp->snd_cwnd = tp->snd_ssthresh; 2875 tp->snd_cwnd_stamp = tcp_time_stamp; 2876 } 2877 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR); 2878} 2879 2880static void tcp_try_keep_open(struct sock *sk) 2881{ 2882 struct tcp_sock *tp = tcp_sk(sk); 2883 int state = TCP_CA_Open; 2884 2885 if (tcp_left_out(tp) || tcp_any_retrans_done(sk)) 2886 state = TCP_CA_Disorder; 2887 2888 if (inet_csk(sk)->icsk_ca_state != state) { 2889 tcp_set_ca_state(sk, state); 2890 tp->high_seq = tp->snd_nxt; 2891 } 2892} 2893 2894static void tcp_try_to_open(struct sock *sk, int flag) 2895{ 2896 struct tcp_sock *tp = tcp_sk(sk); 2897 2898 tcp_verify_left_out(tp); 2899 2900 if (!tp->frto_counter && !tcp_any_retrans_done(sk)) 2901 tp->retrans_stamp = 0; 2902 2903 if (flag & FLAG_ECE) 2904 tcp_enter_cwr(sk, 1); 2905 2906 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) { 2907 tcp_try_keep_open(sk); 2908 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Open) 2909 tcp_moderate_cwnd(tp); 2910 } else { 2911 tcp_cwnd_down(sk, flag); 2912 } 2913} 2914 2915static void tcp_mtup_probe_failed(struct sock *sk) 2916{ 2917 struct inet_connection_sock *icsk = inet_csk(sk); 2918 2919 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1; 2920 icsk->icsk_mtup.probe_size = 0; 2921} 2922 2923static void tcp_mtup_probe_success(struct sock *sk) 2924{ 2925 struct tcp_sock *tp = tcp_sk(sk); 2926 struct inet_connection_sock *icsk = inet_csk(sk); 2927 2928 /* FIXME: breaks with very large cwnd */ 2929 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2930 tp->snd_cwnd = tp->snd_cwnd * 2931 tcp_mss_to_mtu(sk, tp->mss_cache) / 2932 icsk->icsk_mtup.probe_size; 2933 tp->snd_cwnd_cnt = 0; 2934 tp->snd_cwnd_stamp = tcp_time_stamp; 2935 tp->snd_ssthresh = tcp_current_ssthresh(sk); 2936 2937 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size; 2938 icsk->icsk_mtup.probe_size = 0; 2939 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 2940} 2941 2942/* Do a simple retransmit without using the backoff mechanisms in 2943 * tcp_timer. This is used for path mtu discovery. 2944 * The socket is already locked here. 2945 */ 2946void tcp_simple_retransmit(struct sock *sk) 2947{ 2948 const struct inet_connection_sock *icsk = inet_csk(sk); 2949 struct tcp_sock *tp = tcp_sk(sk); 2950 struct sk_buff *skb; 2951 unsigned int mss = tcp_current_mss(sk); 2952 u32 prior_lost = tp->lost_out; 2953 2954 tcp_for_write_queue(skb, sk) { 2955 if (skb == tcp_send_head(sk)) 2956 break; 2957 if (tcp_skb_seglen(skb) > mss && 2958 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) { 2959 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 2960 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 2961 tp->retrans_out -= tcp_skb_pcount(skb); 2962 } 2963 tcp_skb_mark_lost_uncond_verify(tp, skb); 2964 } 2965 } 2966 2967 tcp_clear_retrans_hints_partial(tp); 2968 2969 if (prior_lost == tp->lost_out) 2970 return; 2971 2972 if (tcp_is_reno(tp)) 2973 tcp_limit_reno_sacked(tp); 2974 2975 tcp_verify_left_out(tp); 2976 2977 /* Don't muck with the congestion window here. 2978 * Reason is that we do not increase amount of _data_ 2979 * in network, but units changed and effective 2980 * cwnd/ssthresh really reduced now. 2981 */ 2982 if (icsk->icsk_ca_state != TCP_CA_Loss) { 2983 tp->high_seq = tp->snd_nxt; 2984 tp->snd_ssthresh = tcp_current_ssthresh(sk); 2985 tp->prior_ssthresh = 0; 2986 tp->undo_marker = 0; 2987 tcp_set_ca_state(sk, TCP_CA_Loss); 2988 } 2989 tcp_xmit_retransmit_queue(sk); 2990} 2991EXPORT_SYMBOL(tcp_simple_retransmit); 2992 2993/* This function implements the PRR algorithm, specifcally the PRR-SSRB 2994 * (proportional rate reduction with slow start reduction bound) as described in 2995 * http://www.ietf.org/id/draft-mathis-tcpm-proportional-rate-reduction-01.txt. 2996 * It computes the number of packets to send (sndcnt) based on packets newly 2997 * delivered: 2998 * 1) If the packets in flight is larger than ssthresh, PRR spreads the 2999 * cwnd reductions across a full RTT. 3000 * 2) If packets in flight is lower than ssthresh (such as due to excess 3001 * losses and/or application stalls), do not perform any further cwnd 3002 * reductions, but instead slow start up to ssthresh. 3003 */ 3004static void tcp_update_cwnd_in_recovery(struct sock *sk, int newly_acked_sacked, 3005 int fast_rexmit, int flag) 3006{ 3007 struct tcp_sock *tp = tcp_sk(sk); 3008 int sndcnt = 0; 3009 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp); 3010 3011 if (tcp_packets_in_flight(tp) > tp->snd_ssthresh) { 3012 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered + 3013 tp->prior_cwnd - 1; 3014 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out; 3015 } else { 3016 sndcnt = min_t(int, delta, 3017 max_t(int, tp->prr_delivered - tp->prr_out, 3018 newly_acked_sacked) + 1); 3019 } 3020 3021 sndcnt = max(sndcnt, (fast_rexmit ? 1 : 0)); 3022 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt; 3023} 3024 3025/* Process an event, which can update packets-in-flight not trivially. 3026 * Main goal of this function is to calculate new estimate for left_out, 3027 * taking into account both packets sitting in receiver's buffer and 3028 * packets lost by network. 3029 * 3030 * Besides that it does CWND reduction, when packet loss is detected 3031 * and changes state of machine. 3032 * 3033 * It does _not_ decide what to send, it is made in function 3034 * tcp_xmit_retransmit_queue(). 3035 */ 3036static void tcp_fastretrans_alert(struct sock *sk, int pkts_acked, 3037 int newly_acked_sacked, bool is_dupack, 3038 int flag) 3039{ 3040 struct inet_connection_sock *icsk = inet_csk(sk); 3041 struct tcp_sock *tp = tcp_sk(sk); 3042 int do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) && 3043 (tcp_fackets_out(tp) > tp->reordering)); 3044 int fast_rexmit = 0, mib_idx; 3045 3046 if (WARN_ON(!tp->packets_out && tp->sacked_out)) 3047 tp->sacked_out = 0; 3048 if (WARN_ON(!tp->sacked_out && tp->fackets_out)) 3049 tp->fackets_out = 0; 3050 3051 /* Now state machine starts. 3052 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */ 3053 if (flag & FLAG_ECE) 3054 tp->prior_ssthresh = 0; 3055 3056 /* B. In all the states check for reneging SACKs. */ 3057 if (tcp_check_sack_reneging(sk, flag)) 3058 return; 3059 3060 /* C. Check consistency of the current state. */ 3061 tcp_verify_left_out(tp); 3062 3063 /* D. Check state exit conditions. State can be terminated 3064 * when high_seq is ACKed. */ 3065 if (icsk->icsk_ca_state == TCP_CA_Open) { 3066 WARN_ON(tp->retrans_out != 0); 3067 tp->retrans_stamp = 0; 3068 } else if (!before(tp->snd_una, tp->high_seq)) { 3069 switch (icsk->icsk_ca_state) { 3070 case TCP_CA_Loss: 3071 icsk->icsk_retransmits = 0; 3072 if (tcp_try_undo_recovery(sk)) 3073 return; 3074 break; 3075 3076 case TCP_CA_CWR: 3077 /* CWR is to be held something *above* high_seq 3078 * is ACKed for CWR bit to reach receiver. */ 3079 if (tp->snd_una != tp->high_seq) { 3080 tcp_complete_cwr(sk); 3081 tcp_set_ca_state(sk, TCP_CA_Open); 3082 } 3083 break; 3084 3085 case TCP_CA_Recovery: 3086 if (tcp_is_reno(tp)) 3087 tcp_reset_reno_sack(tp); 3088 if (tcp_try_undo_recovery(sk)) 3089 return; 3090 tcp_complete_cwr(sk); 3091 break; 3092 } 3093 } 3094 3095 /* E. Process state. */ 3096 switch (icsk->icsk_ca_state) { 3097 case TCP_CA_Recovery: 3098 if (!(flag & FLAG_SND_UNA_ADVANCED)) { 3099 if (tcp_is_reno(tp) && is_dupack) 3100 tcp_add_reno_sack(sk); 3101 } else 3102 do_lost = tcp_try_undo_partial(sk, pkts_acked); 3103 break; 3104 case TCP_CA_Loss: 3105 if (flag & FLAG_DATA_ACKED) 3106 icsk->icsk_retransmits = 0; 3107 if (tcp_is_reno(tp) && flag & FLAG_SND_UNA_ADVANCED) 3108 tcp_reset_reno_sack(tp); 3109 if (!tcp_try_undo_loss(sk)) { 3110 tcp_moderate_cwnd(tp); 3111 tcp_xmit_retransmit_queue(sk); 3112 return; 3113 } 3114 if (icsk->icsk_ca_state != TCP_CA_Open) 3115 return; 3116 /* Loss is undone; fall through to processing in Open state. */ 3117 default: 3118 if (tcp_is_reno(tp)) { 3119 if (flag & FLAG_SND_UNA_ADVANCED) 3120 tcp_reset_reno_sack(tp); 3121 if (is_dupack) 3122 tcp_add_reno_sack(sk); 3123 } 3124 3125 if (icsk->icsk_ca_state <= TCP_CA_Disorder) 3126 tcp_try_undo_dsack(sk); 3127 3128 if (!tcp_time_to_recover(sk)) { 3129 tcp_try_to_open(sk, flag); 3130 return; 3131 } 3132 3133 /* MTU probe failure: don't reduce cwnd */ 3134 if (icsk->icsk_ca_state < TCP_CA_CWR && 3135 icsk->icsk_mtup.probe_size && 3136 tp->snd_una == tp->mtu_probe.probe_seq_start) { 3137 tcp_mtup_probe_failed(sk); 3138 /* Restores the reduction we did in tcp_mtup_probe() */ 3139 tp->snd_cwnd++; 3140 tcp_simple_retransmit(sk); 3141 return; 3142 } 3143 3144 /* Otherwise enter Recovery state */ 3145 3146 if (tcp_is_reno(tp)) 3147 mib_idx = LINUX_MIB_TCPRENORECOVERY; 3148 else 3149 mib_idx = LINUX_MIB_TCPSACKRECOVERY; 3150 3151 NET_INC_STATS_BH(sock_net(sk), mib_idx); 3152 3153 tp->high_seq = tp->snd_nxt; 3154 tp->prior_ssthresh = 0; 3155 tp->undo_marker = tp->snd_una; 3156 tp->undo_retrans = tp->retrans_out; 3157 3158 if (icsk->icsk_ca_state < TCP_CA_CWR) { 3159 if (!(flag & FLAG_ECE)) 3160 tp->prior_ssthresh = tcp_current_ssthresh(sk); 3161 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 3162 TCP_ECN_queue_cwr(tp); 3163 } 3164 3165 tp->bytes_acked = 0; 3166 tp->snd_cwnd_cnt = 0; 3167 tp->prior_cwnd = tp->snd_cwnd; 3168 tp->prr_delivered = 0; 3169 tp->prr_out = 0; 3170 tcp_set_ca_state(sk, TCP_CA_Recovery); 3171 fast_rexmit = 1; 3172 } 3173 3174 if (do_lost || (tcp_is_fack(tp) && tcp_head_timedout(sk))) 3175 tcp_update_scoreboard(sk, fast_rexmit); 3176 tp->prr_delivered += newly_acked_sacked; 3177 tcp_update_cwnd_in_recovery(sk, newly_acked_sacked, fast_rexmit, flag); 3178 tcp_xmit_retransmit_queue(sk); 3179} 3180 3181void tcp_valid_rtt_meas(struct sock *sk, u32 seq_rtt) 3182{ 3183 tcp_rtt_estimator(sk, seq_rtt); 3184 tcp_set_rto(sk); 3185 inet_csk(sk)->icsk_backoff = 0; 3186} 3187EXPORT_SYMBOL(tcp_valid_rtt_meas); 3188 3189/* Read draft-ietf-tcplw-high-performance before mucking 3190 * with this code. (Supersedes RFC1323) 3191 */ 3192static void tcp_ack_saw_tstamp(struct sock *sk, int flag) 3193{ 3194 /* RTTM Rule: A TSecr value received in a segment is used to 3195 * update the averaged RTT measurement only if the segment 3196 * acknowledges some new data, i.e., only if it advances the 3197 * left edge of the send window. 3198 * 3199 * See draft-ietf-tcplw-high-performance-00, section 3.3. 3200 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru> 3201 * 3202 * Changed: reset backoff as soon as we see the first valid sample. 3203 * If we do not, we get strongly overestimated rto. With timestamps 3204 * samples are accepted even from very old segments: f.e., when rtt=1 3205 * increases to 8, we retransmit 5 times and after 8 seconds delayed 3206 * answer arrives rto becomes 120 seconds! If at least one of segments 3207 * in window is lost... Voila. --ANK (010210) 3208 */ 3209 struct tcp_sock *tp = tcp_sk(sk); 3210 3211 tcp_valid_rtt_meas(sk, tcp_time_stamp - tp->rx_opt.rcv_tsecr); 3212} 3213 3214static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag) 3215{ 3216 /* We don't have a timestamp. Can only use 3217 * packets that are not retransmitted to determine 3218 * rtt estimates. Also, we must not reset the 3219 * backoff for rto until we get a non-retransmitted 3220 * packet. This allows us to deal with a situation 3221 * where the network delay has increased suddenly. 3222 * I.e. Karn's algorithm. (SIGCOMM '87, p5.) 3223 */ 3224 3225 if (flag & FLAG_RETRANS_DATA_ACKED) 3226 return; 3227 3228 tcp_valid_rtt_meas(sk, seq_rtt); 3229} 3230 3231static inline void tcp_ack_update_rtt(struct sock *sk, const int flag, 3232 const s32 seq_rtt) 3233{ 3234 const struct tcp_sock *tp = tcp_sk(sk); 3235 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */ 3236 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr) 3237 tcp_ack_saw_tstamp(sk, flag); 3238 else if (seq_rtt >= 0) 3239 tcp_ack_no_tstamp(sk, seq_rtt, flag); 3240} 3241 3242static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight) 3243{ 3244 const struct inet_connection_sock *icsk = inet_csk(sk); 3245 icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight); 3246 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp; 3247} 3248 3249/* Restart timer after forward progress on connection. 3250 * RFC2988 recommends to restart timer to now+rto. 3251 */ 3252static void tcp_rearm_rto(struct sock *sk) 3253{ 3254 const struct tcp_sock *tp = tcp_sk(sk); 3255 3256 if (!tp->packets_out) { 3257 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS); 3258 } else { 3259 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 3260 inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 3261 } 3262} 3263 3264/* If we get here, the whole TSO packet has not been acked. */ 3265static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb) 3266{ 3267 struct tcp_sock *tp = tcp_sk(sk); 3268 u32 packets_acked; 3269 3270 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)); 3271 3272 packets_acked = tcp_skb_pcount(skb); 3273 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 3274 return 0; 3275 packets_acked -= tcp_skb_pcount(skb); 3276 3277 if (packets_acked) { 3278 BUG_ON(tcp_skb_pcount(skb) == 0); 3279 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)); 3280 } 3281 3282 return packets_acked; 3283} 3284 3285/* Remove acknowledged frames from the retransmission queue. If our packet 3286 * is before the ack sequence we can discard it as it's confirmed to have 3287 * arrived at the other end. 3288 */ 3289static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets, 3290 u32 prior_snd_una) 3291{ 3292 struct tcp_sock *tp = tcp_sk(sk); 3293 const struct inet_connection_sock *icsk = inet_csk(sk); 3294 struct sk_buff *skb; 3295 u32 now = tcp_time_stamp; 3296 int fully_acked = 1; 3297 int flag = 0; 3298 u32 pkts_acked = 0; 3299 u32 reord = tp->packets_out; 3300 u32 prior_sacked = tp->sacked_out; 3301 s32 seq_rtt = -1; 3302 s32 ca_seq_rtt = -1; 3303 ktime_t last_ackt = net_invalid_timestamp(); 3304 3305 while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) { 3306 struct tcp_skb_cb *scb = TCP_SKB_CB(skb); 3307 u32 acked_pcount; 3308 u8 sacked = scb->sacked; 3309 3310 /* Determine how many packets and what bytes were acked, tso and else */ 3311 if (after(scb->end_seq, tp->snd_una)) { 3312 if (tcp_skb_pcount(skb) == 1 || 3313 !after(tp->snd_una, scb->seq)) 3314 break; 3315 3316 acked_pcount = tcp_tso_acked(sk, skb); 3317 if (!acked_pcount) 3318 break; 3319 3320 fully_acked = 0; 3321 } else { 3322 acked_pcount = tcp_skb_pcount(skb); 3323 } 3324 3325 if (sacked & TCPCB_RETRANS) { 3326 if (sacked & TCPCB_SACKED_RETRANS) 3327 tp->retrans_out -= acked_pcount; 3328 flag |= FLAG_RETRANS_DATA_ACKED; 3329 ca_seq_rtt = -1; 3330 seq_rtt = -1; 3331 if ((flag & FLAG_DATA_ACKED) || (acked_pcount > 1)) 3332 flag |= FLAG_NONHEAD_RETRANS_ACKED; 3333 } else { 3334 ca_seq_rtt = now - scb->when; 3335 last_ackt = skb->tstamp; 3336 if (seq_rtt < 0) { 3337 seq_rtt = ca_seq_rtt; 3338 } 3339 if (!(sacked & TCPCB_SACKED_ACKED)) 3340 reord = min(pkts_acked, reord); 3341 } 3342 3343 if (sacked & TCPCB_SACKED_ACKED) 3344 tp->sacked_out -= acked_pcount; 3345 if (sacked & TCPCB_LOST) 3346 tp->lost_out -= acked_pcount; 3347 3348 tp->packets_out -= acked_pcount; 3349 pkts_acked += acked_pcount; 3350 3351 /* Initial outgoing SYN's get put onto the write_queue 3352 * just like anything else we transmit. It is not 3353 * true data, and if we misinform our callers that 3354 * this ACK acks real data, we will erroneously exit 3355 * connection startup slow start one packet too 3356 * quickly. This is severely frowned upon behavior. 3357 */ 3358 if (!(scb->tcp_flags & TCPHDR_SYN)) { 3359 flag |= FLAG_DATA_ACKED; 3360 } else { 3361 flag |= FLAG_SYN_ACKED; 3362 tp->retrans_stamp = 0; 3363 } 3364 3365 if (!fully_acked) 3366 break; 3367 3368 tcp_unlink_write_queue(skb, sk); 3369 sk_wmem_free_skb(sk, skb); 3370 tp->scoreboard_skb_hint = NULL; 3371 if (skb == tp->retransmit_skb_hint) 3372 tp->retransmit_skb_hint = NULL; 3373 if (skb == tp->lost_skb_hint) 3374 tp->lost_skb_hint = NULL; 3375 } 3376 3377 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una))) 3378 tp->snd_up = tp->snd_una; 3379 3380 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) 3381 flag |= FLAG_SACK_RENEGING; 3382 3383 if (flag & FLAG_ACKED) { 3384 const struct tcp_congestion_ops *ca_ops 3385 = inet_csk(sk)->icsk_ca_ops; 3386 3387 if (unlikely(icsk->icsk_mtup.probe_size && 3388 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) { 3389 tcp_mtup_probe_success(sk); 3390 } 3391 3392 tcp_ack_update_rtt(sk, flag, seq_rtt); 3393 tcp_rearm_rto(sk); 3394 3395 if (tcp_is_reno(tp)) { 3396 tcp_remove_reno_sacks(sk, pkts_acked); 3397 } else { 3398 int delta; 3399 3400 /* Non-retransmitted hole got filled? That's reordering */ 3401 if (reord < prior_fackets) 3402 tcp_update_reordering(sk, tp->fackets_out - reord, 0); 3403 3404 delta = tcp_is_fack(tp) ? pkts_acked : 3405 prior_sacked - tp->sacked_out; 3406 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta); 3407 } 3408 3409 tp->fackets_out -= min(pkts_acked, tp->fackets_out); 3410 3411 if (ca_ops->pkts_acked) { 3412 s32 rtt_us = -1; 3413 3414 /* Is the ACK triggering packet unambiguous? */ 3415 if (!(flag & FLAG_RETRANS_DATA_ACKED)) { 3416 /* High resolution needed and available? */ 3417 if (ca_ops->flags & TCP_CONG_RTT_STAMP && 3418 !ktime_equal(last_ackt, 3419 net_invalid_timestamp())) 3420 rtt_us = ktime_us_delta(ktime_get_real(), 3421 last_ackt); 3422 else if (ca_seq_rtt >= 0) 3423 rtt_us = jiffies_to_usecs(ca_seq_rtt); 3424 } 3425 3426 ca_ops->pkts_acked(sk, pkts_acked, rtt_us); 3427 } 3428 } 3429 3430#if FASTRETRANS_DEBUG > 0 3431 WARN_ON((int)tp->sacked_out < 0); 3432 WARN_ON((int)tp->lost_out < 0); 3433 WARN_ON((int)tp->retrans_out < 0); 3434 if (!tp->packets_out && tcp_is_sack(tp)) { 3435 icsk = inet_csk(sk); 3436 if (tp->lost_out) { 3437 printk(KERN_DEBUG "Leak l=%u %d\n", 3438 tp->lost_out, icsk->icsk_ca_state); 3439 tp->lost_out = 0; 3440 } 3441 if (tp->sacked_out) { 3442 printk(KERN_DEBUG "Leak s=%u %d\n", 3443 tp->sacked_out, icsk->icsk_ca_state); 3444 tp->sacked_out = 0; 3445 } 3446 if (tp->retrans_out) { 3447 printk(KERN_DEBUG "Leak r=%u %d\n", 3448 tp->retrans_out, icsk->icsk_ca_state); 3449 tp->retrans_out = 0; 3450 } 3451 } 3452#endif 3453 return flag; 3454} 3455 3456static void tcp_ack_probe(struct sock *sk) 3457{ 3458 const struct tcp_sock *tp = tcp_sk(sk); 3459 struct inet_connection_sock *icsk = inet_csk(sk); 3460 3461 /* Was it a usable window open? */ 3462 3463 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) { 3464 icsk->icsk_backoff = 0; 3465 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0); 3466 /* Socket must be waked up by subsequent tcp_data_snd_check(). 3467 * This function is not for random using! 3468 */ 3469 } else { 3470 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 3471 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), 3472 TCP_RTO_MAX); 3473 } 3474} 3475 3476static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag) 3477{ 3478 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) || 3479 inet_csk(sk)->icsk_ca_state != TCP_CA_Open; 3480} 3481 3482static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag) 3483{ 3484 const struct tcp_sock *tp = tcp_sk(sk); 3485 return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) && 3486 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR)); 3487} 3488 3489/* Check that window update is acceptable. 3490 * The function assumes that snd_una<=ack<=snd_next. 3491 */ 3492static inline int tcp_may_update_window(const struct tcp_sock *tp, 3493 const u32 ack, const u32 ack_seq, 3494 const u32 nwin) 3495{ 3496 return after(ack, tp->snd_una) || 3497 after(ack_seq, tp->snd_wl1) || 3498 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd); 3499} 3500 3501/* Update our send window. 3502 * 3503 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2 3504 * and in FreeBSD. NetBSD's one is even worse.) is wrong. 3505 */ 3506static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack, 3507 u32 ack_seq) 3508{ 3509 struct tcp_sock *tp = tcp_sk(sk); 3510 int flag = 0; 3511 u32 nwin = ntohs(tcp_hdr(skb)->window); 3512 3513 if (likely(!tcp_hdr(skb)->syn)) 3514 nwin <<= tp->rx_opt.snd_wscale; 3515 3516 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) { 3517 flag |= FLAG_WIN_UPDATE; 3518 tcp_update_wl(tp, ack_seq); 3519 3520 if (tp->snd_wnd != nwin) { 3521 tp->snd_wnd = nwin; 3522 3523 /* Note, it is the only place, where 3524 * fast path is recovered for sending TCP. 3525 */ 3526 tp->pred_flags = 0; 3527 tcp_fast_path_check(sk); 3528 3529 if (nwin > tp->max_window) { 3530 tp->max_window = nwin; 3531 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie); 3532 } 3533 } 3534 } 3535 3536 tp->snd_una = ack; 3537 3538 return flag; 3539} 3540 3541/* A very conservative spurious RTO response algorithm: reduce cwnd and 3542 * continue in congestion avoidance. 3543 */ 3544static void tcp_conservative_spur_to_response(struct tcp_sock *tp) 3545{ 3546 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh); 3547 tp->snd_cwnd_cnt = 0; 3548 tp->bytes_acked = 0; 3549 TCP_ECN_queue_cwr(tp); 3550 tcp_moderate_cwnd(tp); 3551} 3552 3553/* A conservative spurious RTO response algorithm: reduce cwnd using 3554 * rate halving and continue in congestion avoidance. 3555 */ 3556static void tcp_ratehalving_spur_to_response(struct sock *sk) 3557{ 3558 tcp_enter_cwr(sk, 0); 3559} 3560 3561static void tcp_undo_spur_to_response(struct sock *sk, int flag) 3562{ 3563 if (flag & FLAG_ECE) 3564 tcp_ratehalving_spur_to_response(sk); 3565 else 3566 tcp_undo_cwr(sk, true); 3567} 3568 3569/* F-RTO spurious RTO detection algorithm (RFC4138) 3570 * 3571 * F-RTO affects during two new ACKs following RTO (well, almost, see inline 3572 * comments). State (ACK number) is kept in frto_counter. When ACK advances 3573 * window (but not to or beyond highest sequence sent before RTO): 3574 * On First ACK, send two new segments out. 3575 * On Second ACK, RTO was likely spurious. Do spurious response (response 3576 * algorithm is not part of the F-RTO detection algorithm 3577 * given in RFC4138 but can be selected separately). 3578 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss 3579 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding 3580 * of Nagle, this is done using frto_counter states 2 and 3, when a new data 3581 * segment of any size sent during F-RTO, state 2 is upgraded to 3. 3582 * 3583 * Rationale: if the RTO was spurious, new ACKs should arrive from the 3584 * original window even after we transmit two new data segments. 3585 * 3586 * SACK version: 3587 * on first step, wait until first cumulative ACK arrives, then move to 3588 * the second step. In second step, the next ACK decides. 3589 * 3590 * F-RTO is implemented (mainly) in four functions: 3591 * - tcp_use_frto() is used to determine if TCP is can use F-RTO 3592 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is 3593 * called when tcp_use_frto() showed green light 3594 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm 3595 * - tcp_enter_frto_loss() is called if there is not enough evidence 3596 * to prove that the RTO is indeed spurious. It transfers the control 3597 * from F-RTO to the conventional RTO recovery 3598 */ 3599static int tcp_process_frto(struct sock *sk, int flag) 3600{ 3601 struct tcp_sock *tp = tcp_sk(sk); 3602 3603 tcp_verify_left_out(tp); 3604 3605 /* Duplicate the behavior from Loss state (fastretrans_alert) */ 3606 if (flag & FLAG_DATA_ACKED) 3607 inet_csk(sk)->icsk_retransmits = 0; 3608 3609 if ((flag & FLAG_NONHEAD_RETRANS_ACKED) || 3610 ((tp->frto_counter >= 2) && (flag & FLAG_RETRANS_DATA_ACKED))) 3611 tp->undo_marker = 0; 3612 3613 if (!before(tp->snd_una, tp->frto_highmark)) { 3614 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag); 3615 return 1; 3616 } 3617 3618 if (!tcp_is_sackfrto(tp)) { 3619 /* RFC4138 shortcoming in step 2; should also have case c): 3620 * ACK isn't duplicate nor advances window, e.g., opposite dir 3621 * data, winupdate 3622 */ 3623 if (!(flag & FLAG_ANY_PROGRESS) && (flag & FLAG_NOT_DUP)) 3624 return 1; 3625 3626 if (!(flag & FLAG_DATA_ACKED)) { 3627 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3), 3628 flag); 3629 return 1; 3630 } 3631 } else { 3632 if (!(flag & FLAG_DATA_ACKED) && (tp->frto_counter == 1)) { 3633 /* Prevent sending of new data. */ 3634 tp->snd_cwnd = min(tp->snd_cwnd, 3635 tcp_packets_in_flight(tp)); 3636 return 1; 3637 } 3638 3639 if ((tp->frto_counter >= 2) && 3640 (!(flag & FLAG_FORWARD_PROGRESS) || 3641 ((flag & FLAG_DATA_SACKED) && 3642 !(flag & FLAG_ONLY_ORIG_SACKED)))) { 3643 /* RFC4138 shortcoming (see comment above) */ 3644 if (!(flag & FLAG_FORWARD_PROGRESS) && 3645 (flag & FLAG_NOT_DUP)) 3646 return 1; 3647 3648 tcp_enter_frto_loss(sk, 3, flag); 3649 return 1; 3650 } 3651 } 3652 3653 if (tp->frto_counter == 1) { 3654 /* tcp_may_send_now needs to see updated state */ 3655 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2; 3656 tp->frto_counter = 2; 3657 3658 if (!tcp_may_send_now(sk)) 3659 tcp_enter_frto_loss(sk, 2, flag); 3660 3661 return 1; 3662 } else { 3663 switch (sysctl_tcp_frto_response) { 3664 case 2: 3665 tcp_undo_spur_to_response(sk, flag); 3666 break; 3667 case 1: 3668 tcp_conservative_spur_to_response(tp); 3669 break; 3670 default: 3671 tcp_ratehalving_spur_to_response(sk); 3672 break; 3673 } 3674 tp->frto_counter = 0; 3675 tp->undo_marker = 0; 3676 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSPURIOUSRTOS); 3677 } 3678 return 0; 3679} 3680 3681/* This routine deals with incoming acks, but not outgoing ones. */ 3682static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag) 3683{ 3684 struct inet_connection_sock *icsk = inet_csk(sk); 3685 struct tcp_sock *tp = tcp_sk(sk); 3686 u32 prior_snd_una = tp->snd_una; 3687 u32 ack_seq = TCP_SKB_CB(skb)->seq; 3688 u32 ack = TCP_SKB_CB(skb)->ack_seq; 3689 bool is_dupack = false; 3690 u32 prior_in_flight; 3691 u32 prior_fackets; 3692 int prior_packets; 3693 int prior_sacked = tp->sacked_out; 3694 int pkts_acked = 0; 3695 int newly_acked_sacked = 0; 3696 int frto_cwnd = 0; 3697 3698 /* If the ack is older than previous acks 3699 * then we can probably ignore it. 3700 */ 3701 if (before(ack, prior_snd_una)) 3702 goto old_ack; 3703 3704 /* If the ack includes data we haven't sent yet, discard 3705 * this segment (RFC793 Section 3.9). 3706 */ 3707 if (after(ack, tp->snd_nxt)) 3708 goto invalid_ack; 3709 3710 if (after(ack, prior_snd_una)) 3711 flag |= FLAG_SND_UNA_ADVANCED; 3712 3713 if (sysctl_tcp_abc) { 3714 if (icsk->icsk_ca_state < TCP_CA_CWR) 3715 tp->bytes_acked += ack - prior_snd_una; 3716 else if (icsk->icsk_ca_state == TCP_CA_Loss) 3717 /* we assume just one segment left network */ 3718 tp->bytes_acked += min(ack - prior_snd_una, 3719 tp->mss_cache); 3720 } 3721 3722 prior_fackets = tp->fackets_out; 3723 prior_in_flight = tcp_packets_in_flight(tp); 3724 3725 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) { 3726 /* Window is constant, pure forward advance. 3727 * No more checks are required. 3728 * Note, we use the fact that SND.UNA>=SND.WL2. 3729 */ 3730 tcp_update_wl(tp, ack_seq); 3731 tp->snd_una = ack; 3732 flag |= FLAG_WIN_UPDATE; 3733 3734 tcp_ca_event(sk, CA_EVENT_FAST_ACK); 3735 3736 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS); 3737 } else { 3738 if (ack_seq != TCP_SKB_CB(skb)->end_seq) 3739 flag |= FLAG_DATA; 3740 else 3741 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS); 3742 3743 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq); 3744 3745 if (TCP_SKB_CB(skb)->sacked) 3746 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una); 3747 3748 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb))) 3749 flag |= FLAG_ECE; 3750 3751 tcp_ca_event(sk, CA_EVENT_SLOW_ACK); 3752 } 3753 3754 /* We passed data and got it acked, remove any soft error 3755 * log. Something worked... 3756 */ 3757 sk->sk_err_soft = 0; 3758 icsk->icsk_probes_out = 0; 3759 tp->rcv_tstamp = tcp_time_stamp; 3760 prior_packets = tp->packets_out; 3761 if (!prior_packets) 3762 goto no_queue; 3763 3764 /* See if we can take anything off of the retransmit queue. */ 3765 flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una); 3766 3767 pkts_acked = prior_packets - tp->packets_out; 3768 newly_acked_sacked = (prior_packets - prior_sacked) - 3769 (tp->packets_out - tp->sacked_out); 3770 3771 if (tp->frto_counter) 3772 frto_cwnd = tcp_process_frto(sk, flag); 3773 /* Guarantee sacktag reordering detection against wrap-arounds */ 3774 if (before(tp->frto_highmark, tp->snd_una)) 3775 tp->frto_highmark = 0; 3776 3777 if (tcp_ack_is_dubious(sk, flag)) { 3778 /* Advance CWND, if state allows this. */ 3779 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd && 3780 tcp_may_raise_cwnd(sk, flag)) 3781 tcp_cong_avoid(sk, ack, prior_in_flight); 3782 is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP)); 3783 tcp_fastretrans_alert(sk, pkts_acked, newly_acked_sacked, 3784 is_dupack, flag); 3785 } else { 3786 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd) 3787 tcp_cong_avoid(sk, ack, prior_in_flight); 3788 } 3789 3790 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) 3791 dst_confirm(__sk_dst_get(sk)); 3792 3793 return 1; 3794 3795no_queue: 3796 /* If data was DSACKed, see if we can undo a cwnd reduction. */ 3797 if (flag & FLAG_DSACKING_ACK) 3798 tcp_fastretrans_alert(sk, pkts_acked, newly_acked_sacked, 3799 is_dupack, flag); 3800 /* If this ack opens up a zero window, clear backoff. It was 3801 * being used to time the probes, and is probably far higher than 3802 * it needs to be for normal retransmission. 3803 */ 3804 if (tcp_send_head(sk)) 3805 tcp_ack_probe(sk); 3806 return 1; 3807 3808invalid_ack: 3809 SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt); 3810 return -1; 3811 3812old_ack: 3813 /* If data was SACKed, tag it and see if we should send more data. 3814 * If data was DSACKed, see if we can undo a cwnd reduction. 3815 */ 3816 if (TCP_SKB_CB(skb)->sacked) { 3817 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una); 3818 newly_acked_sacked = tp->sacked_out - prior_sacked; 3819 tcp_fastretrans_alert(sk, pkts_acked, newly_acked_sacked, 3820 is_dupack, flag); 3821 } 3822 3823 SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt); 3824 return 0; 3825} 3826 3827/* Look for tcp options. Normally only called on SYN and SYNACK packets. 3828 * But, this can also be called on packets in the established flow when 3829 * the fast version below fails. 3830 */ 3831void tcp_parse_options(const struct sk_buff *skb, struct tcp_options_received *opt_rx, 3832 const u8 **hvpp, int estab) 3833{ 3834 const unsigned char *ptr; 3835 const struct tcphdr *th = tcp_hdr(skb); 3836 int length = (th->doff * 4) - sizeof(struct tcphdr); 3837 3838 ptr = (const unsigned char *)(th + 1); 3839 opt_rx->saw_tstamp = 0; 3840 3841 while (length > 0) { 3842 int opcode = *ptr++; 3843 int opsize; 3844 3845 switch (opcode) { 3846 case TCPOPT_EOL: 3847 return; 3848 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */ 3849 length--; 3850 continue; 3851 default: 3852 opsize = *ptr++; 3853 if (opsize < 2) /* "silly options" */ 3854 return; 3855 if (opsize > length) 3856 return; /* don't parse partial options */ 3857 switch (opcode) { 3858 case TCPOPT_MSS: 3859 if (opsize == TCPOLEN_MSS && th->syn && !estab) { 3860 u16 in_mss = get_unaligned_be16(ptr); 3861 if (in_mss) { 3862 if (opt_rx->user_mss && 3863 opt_rx->user_mss < in_mss) 3864 in_mss = opt_rx->user_mss; 3865 opt_rx->mss_clamp = in_mss; 3866 } 3867 } 3868 break; 3869 case TCPOPT_WINDOW: 3870 if (opsize == TCPOLEN_WINDOW && th->syn && 3871 !estab && sysctl_tcp_window_scaling) { 3872 __u8 snd_wscale = *(__u8 *)ptr; 3873 opt_rx->wscale_ok = 1; 3874 if (snd_wscale > 14) { 3875 if (net_ratelimit()) 3876 pr_info("%s: Illegal window scaling value %d >14 received\n", 3877 __func__, 3878 snd_wscale); 3879 snd_wscale = 14; 3880 } 3881 opt_rx->snd_wscale = snd_wscale; 3882 } 3883 break; 3884 case TCPOPT_TIMESTAMP: 3885 if ((opsize == TCPOLEN_TIMESTAMP) && 3886 ((estab && opt_rx->tstamp_ok) || 3887 (!estab && sysctl_tcp_timestamps))) { 3888 opt_rx->saw_tstamp = 1; 3889 opt_rx->rcv_tsval = get_unaligned_be32(ptr); 3890 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4); 3891 } 3892 break; 3893 case TCPOPT_SACK_PERM: 3894 if (opsize == TCPOLEN_SACK_PERM && th->syn && 3895 !estab && sysctl_tcp_sack) { 3896 opt_rx->sack_ok = TCP_SACK_SEEN; 3897 tcp_sack_reset(opt_rx); 3898 } 3899 break; 3900 3901 case TCPOPT_SACK: 3902 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) && 3903 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) && 3904 opt_rx->sack_ok) { 3905 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th; 3906 } 3907 break; 3908#ifdef CONFIG_TCP_MD5SIG 3909 case TCPOPT_MD5SIG: 3910 /* 3911 * The MD5 Hash has already been 3912 * checked (see tcp_v{4,6}_do_rcv()). 3913 */ 3914 break; 3915#endif 3916 case TCPOPT_COOKIE: 3917 /* This option is variable length. 3918 */ 3919 switch (opsize) { 3920 case TCPOLEN_COOKIE_BASE: 3921 /* not yet implemented */ 3922 break; 3923 case TCPOLEN_COOKIE_PAIR: 3924 /* not yet implemented */ 3925 break; 3926 case TCPOLEN_COOKIE_MIN+0: 3927 case TCPOLEN_COOKIE_MIN+2: 3928 case TCPOLEN_COOKIE_MIN+4: 3929 case TCPOLEN_COOKIE_MIN+6: 3930 case TCPOLEN_COOKIE_MAX: 3931 /* 16-bit multiple */ 3932 opt_rx->cookie_plus = opsize; 3933 *hvpp = ptr; 3934 break; 3935 default: 3936 /* ignore option */ 3937 break; 3938 } 3939 break; 3940 } 3941 3942 ptr += opsize-2; 3943 length -= opsize; 3944 } 3945 } 3946} 3947EXPORT_SYMBOL(tcp_parse_options); 3948 3949static int tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th) 3950{ 3951 const __be32 *ptr = (const __be32 *)(th + 1); 3952 3953 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) 3954 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) { 3955 tp->rx_opt.saw_tstamp = 1; 3956 ++ptr; 3957 tp->rx_opt.rcv_tsval = ntohl(*ptr); 3958 ++ptr; 3959 tp->rx_opt.rcv_tsecr = ntohl(*ptr); 3960 return 1; 3961 } 3962 return 0; 3963} 3964 3965/* Fast parse options. This hopes to only see timestamps. 3966 * If it is wrong it falls back on tcp_parse_options(). 3967 */ 3968static int tcp_fast_parse_options(const struct sk_buff *skb, 3969 const struct tcphdr *th, 3970 struct tcp_sock *tp, const u8 **hvpp) 3971{ 3972 /* In the spirit of fast parsing, compare doff directly to constant 3973 * values. Because equality is used, short doff can be ignored here. 3974 */ 3975 if (th->doff == (sizeof(*th) / 4)) { 3976 tp->rx_opt.saw_tstamp = 0; 3977 return 0; 3978 } else if (tp->rx_opt.tstamp_ok && 3979 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) { 3980 if (tcp_parse_aligned_timestamp(tp, th)) 3981 return 1; 3982 } 3983 tcp_parse_options(skb, &tp->rx_opt, hvpp, 1); 3984 return 1; 3985} 3986 3987#ifdef CONFIG_TCP_MD5SIG 3988/* 3989 * Parse MD5 Signature option 3990 */ 3991const u8 *tcp_parse_md5sig_option(const struct tcphdr *th) 3992{ 3993 int length = (th->doff << 2) - sizeof(*th); 3994 const u8 *ptr = (const u8 *)(th + 1); 3995 3996 /* If the TCP option is too short, we can short cut */ 3997 if (length < TCPOLEN_MD5SIG) 3998 return NULL; 3999 4000 while (length > 0) { 4001 int opcode = *ptr++; 4002 int opsize; 4003 4004 switch(opcode) { 4005 case TCPOPT_EOL: 4006 return NULL; 4007 case TCPOPT_NOP: 4008 length--; 4009 continue; 4010 default: 4011 opsize = *ptr++; 4012 if (opsize < 2 || opsize > length) 4013 return NULL; 4014 if (opcode == TCPOPT_MD5SIG) 4015 return opsize == TCPOLEN_MD5SIG ? ptr : NULL; 4016 } 4017 ptr += opsize - 2; 4018 length -= opsize; 4019 } 4020 return NULL; 4021} 4022EXPORT_SYMBOL(tcp_parse_md5sig_option); 4023#endif 4024 4025static inline void tcp_store_ts_recent(struct tcp_sock *tp) 4026{ 4027 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval; 4028 tp->rx_opt.ts_recent_stamp = get_seconds(); 4029} 4030 4031static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq) 4032{ 4033 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) { 4034 /* PAWS bug workaround wrt. ACK frames, the PAWS discard 4035 * extra check below makes sure this can only happen 4036 * for pure ACK frames. -DaveM 4037 * 4038 * Not only, also it occurs for expired timestamps. 4039 */ 4040 4041 if (tcp_paws_check(&tp->rx_opt, 0)) 4042 tcp_store_ts_recent(tp); 4043 } 4044} 4045 4046/* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM 4047 * 4048 * It is not fatal. If this ACK does _not_ change critical state (seqs, window) 4049 * it can pass through stack. So, the following predicate verifies that 4050 * this segment is not used for anything but congestion avoidance or 4051 * fast retransmit. Moreover, we even are able to eliminate most of such 4052 * second order effects, if we apply some small "replay" window (~RTO) 4053 * to timestamp space. 4054 * 4055 * All these measures still do not guarantee that we reject wrapped ACKs 4056 * on networks with high bandwidth, when sequence space is recycled fastly, 4057 * but it guarantees that such events will be very rare and do not affect 4058 * connection seriously. This doesn't look nice, but alas, PAWS is really 4059 * buggy extension. 4060 * 4061 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC 4062 * states that events when retransmit arrives after original data are rare. 4063 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is 4064 * the biggest problem on large power networks even with minor reordering. 4065 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe 4066 * up to bandwidth of 18Gigabit/sec. 8) ] 4067 */ 4068 4069static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb) 4070{ 4071 const struct tcp_sock *tp = tcp_sk(sk); 4072 const struct tcphdr *th = tcp_hdr(skb); 4073 u32 seq = TCP_SKB_CB(skb)->seq; 4074 u32 ack = TCP_SKB_CB(skb)->ack_seq; 4075 4076 return (/* 1. Pure ACK with correct sequence number. */ 4077 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) && 4078 4079 /* 2. ... and duplicate ACK. */ 4080 ack == tp->snd_una && 4081 4082 /* 3. ... and does not update window. */ 4083 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) && 4084 4085 /* 4. ... and sits in replay window. */ 4086 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ); 4087} 4088 4089static inline int tcp_paws_discard(const struct sock *sk, 4090 const struct sk_buff *skb) 4091{ 4092 const struct tcp_sock *tp = tcp_sk(sk); 4093 4094 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) && 4095 !tcp_disordered_ack(sk, skb); 4096} 4097 4098/* Check segment sequence number for validity. 4099 * 4100 * Segment controls are considered valid, if the segment 4101 * fits to the window after truncation to the window. Acceptability 4102 * of data (and SYN, FIN, of course) is checked separately. 4103 * See tcp_data_queue(), for example. 4104 * 4105 * Also, controls (RST is main one) are accepted using RCV.WUP instead 4106 * of RCV.NXT. Peer still did not advance his SND.UNA when we 4107 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP. 4108 * (borrowed from freebsd) 4109 */ 4110 4111static inline int tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq) 4112{ 4113 return !before(end_seq, tp->rcv_wup) && 4114 !after(seq, tp->rcv_nxt + tcp_receive_window(tp)); 4115} 4116 4117/* When we get a reset we do this. */ 4118static void tcp_reset(struct sock *sk) 4119{ 4120 /* We want the right error as BSD sees it (and indeed as we do). */ 4121 switch (sk->sk_state) { 4122 case TCP_SYN_SENT: 4123 sk->sk_err = ECONNREFUSED; 4124 break; 4125 case TCP_CLOSE_WAIT: 4126 sk->sk_err = EPIPE; 4127 break; 4128 case TCP_CLOSE: 4129 return; 4130 default: 4131 sk->sk_err = ECONNRESET; 4132 } 4133 /* This barrier is coupled with smp_rmb() in tcp_poll() */ 4134 smp_wmb(); 4135 4136 if (!sock_flag(sk, SOCK_DEAD)) 4137 sk->sk_error_report(sk); 4138 4139 tcp_done(sk); 4140} 4141 4142/* 4143 * Process the FIN bit. This now behaves as it is supposed to work 4144 * and the FIN takes effect when it is validly part of sequence 4145 * space. Not before when we get holes. 4146 * 4147 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT 4148 * (and thence onto LAST-ACK and finally, CLOSE, we never enter 4149 * TIME-WAIT) 4150 * 4151 * If we are in FINWAIT-1, a received FIN indicates simultaneous 4152 * close and we go into CLOSING (and later onto TIME-WAIT) 4153 * 4154 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT. 4155 */ 4156static void tcp_fin(struct sock *sk) 4157{ 4158 struct tcp_sock *tp = tcp_sk(sk); 4159 4160 inet_csk_schedule_ack(sk); 4161 4162 sk->sk_shutdown |= RCV_SHUTDOWN; 4163 sock_set_flag(sk, SOCK_DONE); 4164 4165 switch (sk->sk_state) { 4166 case TCP_SYN_RECV: 4167 case TCP_ESTABLISHED: 4168 /* Move to CLOSE_WAIT */ 4169 tcp_set_state(sk, TCP_CLOSE_WAIT); 4170 inet_csk(sk)->icsk_ack.pingpong = 1; 4171 break; 4172 4173 case TCP_CLOSE_WAIT: 4174 case TCP_CLOSING: 4175 /* Received a retransmission of the FIN, do 4176 * nothing. 4177 */ 4178 break; 4179 case TCP_LAST_ACK: 4180 /* RFC793: Remain in the LAST-ACK state. */ 4181 break; 4182 4183 case TCP_FIN_WAIT1: 4184 /* This case occurs when a simultaneous close 4185 * happens, we must ack the received FIN and 4186 * enter the CLOSING state. 4187 */ 4188 tcp_send_ack(sk); 4189 tcp_set_state(sk, TCP_CLOSING); 4190 break; 4191 case TCP_FIN_WAIT2: 4192 /* Received a FIN -- send ACK and enter TIME_WAIT. */ 4193 tcp_send_ack(sk); 4194 tcp_time_wait(sk, TCP_TIME_WAIT, 0); 4195 break; 4196 default: 4197 /* Only TCP_LISTEN and TCP_CLOSE are left, in these 4198 * cases we should never reach this piece of code. 4199 */ 4200 pr_err("%s: Impossible, sk->sk_state=%d\n", 4201 __func__, sk->sk_state); 4202 break; 4203 } 4204 4205 /* It _is_ possible, that we have something out-of-order _after_ FIN. 4206 * Probably, we should reset in this case. For now drop them. 4207 */ 4208 __skb_queue_purge(&tp->out_of_order_queue); 4209 if (tcp_is_sack(tp)) 4210 tcp_sack_reset(&tp->rx_opt); 4211 sk_mem_reclaim(sk); 4212 4213 if (!sock_flag(sk, SOCK_DEAD)) { 4214 sk->sk_state_change(sk); 4215 4216 /* Do not send POLL_HUP for half duplex close. */ 4217 if (sk->sk_shutdown == SHUTDOWN_MASK || 4218 sk->sk_state == TCP_CLOSE) 4219 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP); 4220 else 4221 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN); 4222 } 4223} 4224 4225static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, 4226 u32 end_seq) 4227{ 4228 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) { 4229 if (before(seq, sp->start_seq)) 4230 sp->start_seq = seq; 4231 if (after(end_seq, sp->end_seq)) 4232 sp->end_seq = end_seq; 4233 return 1; 4234 } 4235 return 0; 4236} 4237 4238static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq) 4239{ 4240 struct tcp_sock *tp = tcp_sk(sk); 4241 4242 if (tcp_is_sack(tp) && sysctl_tcp_dsack) { 4243 int mib_idx; 4244 4245 if (before(seq, tp->rcv_nxt)) 4246 mib_idx = LINUX_MIB_TCPDSACKOLDSENT; 4247 else 4248 mib_idx = LINUX_MIB_TCPDSACKOFOSENT; 4249 4250 NET_INC_STATS_BH(sock_net(sk), mib_idx); 4251 4252 tp->rx_opt.dsack = 1; 4253 tp->duplicate_sack[0].start_seq = seq; 4254 tp->duplicate_sack[0].end_seq = end_seq; 4255 } 4256} 4257 4258static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq) 4259{ 4260 struct tcp_sock *tp = tcp_sk(sk); 4261 4262 if (!tp->rx_opt.dsack) 4263 tcp_dsack_set(sk, seq, end_seq); 4264 else 4265 tcp_sack_extend(tp->duplicate_sack, seq, end_seq); 4266} 4267 4268static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb) 4269{ 4270 struct tcp_sock *tp = tcp_sk(sk); 4271 4272 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 4273 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 4274 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST); 4275 tcp_enter_quickack_mode(sk); 4276 4277 if (tcp_is_sack(tp) && sysctl_tcp_dsack) { 4278 u32 end_seq = TCP_SKB_CB(skb)->end_seq; 4279 4280 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) 4281 end_seq = tp->rcv_nxt; 4282 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq); 4283 } 4284 } 4285 4286 tcp_send_ack(sk); 4287} 4288 4289/* These routines update the SACK block as out-of-order packets arrive or 4290 * in-order packets close up the sequence space. 4291 */ 4292static void tcp_sack_maybe_coalesce(struct tcp_sock *tp) 4293{ 4294 int this_sack; 4295 struct tcp_sack_block *sp = &tp->selective_acks[0]; 4296 struct tcp_sack_block *swalk = sp + 1; 4297 4298 /* See if the recent change to the first SACK eats into 4299 * or hits the sequence space of other SACK blocks, if so coalesce. 4300 */ 4301 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) { 4302 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) { 4303 int i; 4304 4305 /* Zap SWALK, by moving every further SACK up by one slot. 4306 * Decrease num_sacks. 4307 */ 4308 tp->rx_opt.num_sacks--; 4309 for (i = this_sack; i < tp->rx_opt.num_sacks; i++) 4310 sp[i] = sp[i + 1]; 4311 continue; 4312 } 4313 this_sack++, swalk++; 4314 } 4315} 4316 4317static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq) 4318{ 4319 struct tcp_sock *tp = tcp_sk(sk); 4320 struct tcp_sack_block *sp = &tp->selective_acks[0]; 4321 int cur_sacks = tp->rx_opt.num_sacks; 4322 int this_sack; 4323 4324 if (!cur_sacks) 4325 goto new_sack; 4326 4327 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) { 4328 if (tcp_sack_extend(sp, seq, end_seq)) { 4329 /* Rotate this_sack to the first one. */ 4330 for (; this_sack > 0; this_sack--, sp--) 4331 swap(*sp, *(sp - 1)); 4332 if (cur_sacks > 1) 4333 tcp_sack_maybe_coalesce(tp); 4334 return; 4335 } 4336 } 4337 4338 /* Could not find an adjacent existing SACK, build a new one, 4339 * put it at the front, and shift everyone else down. We 4340 * always know there is at least one SACK present already here. 4341 * 4342 * If the sack array is full, forget about the last one. 4343 */ 4344 if (this_sack >= TCP_NUM_SACKS) { 4345 this_sack--; 4346 tp->rx_opt.num_sacks--; 4347 sp--; 4348 } 4349 for (; this_sack > 0; this_sack--, sp--) 4350 *sp = *(sp - 1); 4351 4352new_sack: 4353 /* Build the new head SACK, and we're done. */ 4354 sp->start_seq = seq; 4355 sp->end_seq = end_seq; 4356 tp->rx_opt.num_sacks++; 4357} 4358 4359/* RCV.NXT advances, some SACKs should be eaten. */ 4360 4361static void tcp_sack_remove(struct tcp_sock *tp) 4362{ 4363 struct tcp_sack_block *sp = &tp->selective_acks[0]; 4364 int num_sacks = tp->rx_opt.num_sacks; 4365 int this_sack; 4366 4367 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */ 4368 if (skb_queue_empty(&tp->out_of_order_queue)) { 4369 tp->rx_opt.num_sacks = 0; 4370 return; 4371 } 4372 4373 for (this_sack = 0; this_sack < num_sacks;) { 4374 /* Check if the start of the sack is covered by RCV.NXT. */ 4375 if (!before(tp->rcv_nxt, sp->start_seq)) { 4376 int i; 4377 4378 /* RCV.NXT must cover all the block! */ 4379 WARN_ON(before(tp->rcv_nxt, sp->end_seq)); 4380 4381 /* Zap this SACK, by moving forward any other SACKS. */ 4382 for (i=this_sack+1; i < num_sacks; i++) 4383 tp->selective_acks[i-1] = tp->selective_acks[i]; 4384 num_sacks--; 4385 continue; 4386 } 4387 this_sack++; 4388 sp++; 4389 } 4390 tp->rx_opt.num_sacks = num_sacks; 4391} 4392 4393/* This one checks to see if we can put data from the 4394 * out_of_order queue into the receive_queue. 4395 */ 4396static void tcp_ofo_queue(struct sock *sk) 4397{ 4398 struct tcp_sock *tp = tcp_sk(sk); 4399 __u32 dsack_high = tp->rcv_nxt; 4400 struct sk_buff *skb; 4401 4402 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) { 4403 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) 4404 break; 4405 4406 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) { 4407 __u32 dsack = dsack_high; 4408 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high)) 4409 dsack_high = TCP_SKB_CB(skb)->end_seq; 4410 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack); 4411 } 4412 4413 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) { 4414 SOCK_DEBUG(sk, "ofo packet was already received\n"); 4415 __skb_unlink(skb, &tp->out_of_order_queue); 4416 __kfree_skb(skb); 4417 continue; 4418 } 4419 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n", 4420 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, 4421 TCP_SKB_CB(skb)->end_seq); 4422 4423 __skb_unlink(skb, &tp->out_of_order_queue); 4424 __skb_queue_tail(&sk->sk_receive_queue, skb); 4425 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; 4426 if (tcp_hdr(skb)->fin) 4427 tcp_fin(sk); 4428 } 4429} 4430 4431static int tcp_prune_ofo_queue(struct sock *sk); 4432static int tcp_prune_queue(struct sock *sk); 4433 4434static inline int tcp_try_rmem_schedule(struct sock *sk, unsigned int size) 4435{ 4436 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf || 4437 !sk_rmem_schedule(sk, size)) { 4438 4439 if (tcp_prune_queue(sk) < 0) 4440 return -1; 4441 4442 if (!sk_rmem_schedule(sk, size)) { 4443 if (!tcp_prune_ofo_queue(sk)) 4444 return -1; 4445 4446 if (!sk_rmem_schedule(sk, size)) 4447 return -1; 4448 } 4449 } 4450 return 0; 4451} 4452 4453static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb) 4454{ 4455 struct tcp_sock *tp = tcp_sk(sk); 4456 struct sk_buff *skb1; 4457 u32 seq, end_seq; 4458 4459 TCP_ECN_check_ce(tp, skb); 4460 4461 if (tcp_try_rmem_schedule(sk, skb->truesize)) { 4462 /* TODO: should increment a counter */ 4463 __kfree_skb(skb); 4464 return; 4465 } 4466 4467 /* Disable header prediction. */ 4468 tp->pred_flags = 0; 4469 inet_csk_schedule_ack(sk); 4470 4471 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n", 4472 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); 4473 4474 skb1 = skb_peek_tail(&tp->out_of_order_queue); 4475 if (!skb1) { 4476 /* Initial out of order segment, build 1 SACK. */ 4477 if (tcp_is_sack(tp)) { 4478 tp->rx_opt.num_sacks = 1; 4479 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq; 4480 tp->selective_acks[0].end_seq = 4481 TCP_SKB_CB(skb)->end_seq; 4482 } 4483 __skb_queue_head(&tp->out_of_order_queue, skb); 4484 goto end; 4485 } 4486 4487 seq = TCP_SKB_CB(skb)->seq; 4488 end_seq = TCP_SKB_CB(skb)->end_seq; 4489 4490 if (seq == TCP_SKB_CB(skb1)->end_seq) { 4491 /* Packets in ofo can stay in queue a long time. 4492 * Better try to coalesce them right now 4493 * to avoid future tcp_collapse_ofo_queue(), 4494 * probably the most expensive function in tcp stack. 4495 */ 4496 if (skb->len <= skb_tailroom(skb1) && !tcp_hdr(skb)->fin) { 4497 NET_INC_STATS_BH(sock_net(sk), 4498 LINUX_MIB_TCPRCVCOALESCE); 4499 BUG_ON(skb_copy_bits(skb, 0, 4500 skb_put(skb1, skb->len), 4501 skb->len)); 4502 TCP_SKB_CB(skb1)->end_seq = end_seq; 4503 TCP_SKB_CB(skb1)->ack_seq = TCP_SKB_CB(skb)->ack_seq; 4504 __kfree_skb(skb); 4505 skb = NULL; 4506 } else { 4507 __skb_queue_after(&tp->out_of_order_queue, skb1, skb); 4508 } 4509 4510 if (!tp->rx_opt.num_sacks || 4511 tp->selective_acks[0].end_seq != seq) 4512 goto add_sack; 4513 4514 /* Common case: data arrive in order after hole. */ 4515 tp->selective_acks[0].end_seq = end_seq; 4516 goto end; 4517 } 4518 4519 /* Find place to insert this segment. */ 4520 while (1) { 4521 if (!after(TCP_SKB_CB(skb1)->seq, seq)) 4522 break; 4523 if (skb_queue_is_first(&tp->out_of_order_queue, skb1)) { 4524 skb1 = NULL; 4525 break; 4526 } 4527 skb1 = skb_queue_prev(&tp->out_of_order_queue, skb1); 4528 } 4529 4530 /* Do skb overlap to previous one? */ 4531 if (skb1 && before(seq, TCP_SKB_CB(skb1)->end_seq)) { 4532 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) { 4533 /* All the bits are present. Drop. */ 4534 __kfree_skb(skb); 4535 skb = NULL; 4536 tcp_dsack_set(sk, seq, end_seq); 4537 goto add_sack; 4538 } 4539 if (after(seq, TCP_SKB_CB(skb1)->seq)) { 4540 /* Partial overlap. */ 4541 tcp_dsack_set(sk, seq, 4542 TCP_SKB_CB(skb1)->end_seq); 4543 } else { 4544 if (skb_queue_is_first(&tp->out_of_order_queue, 4545 skb1)) 4546 skb1 = NULL; 4547 else 4548 skb1 = skb_queue_prev( 4549 &tp->out_of_order_queue, 4550 skb1); 4551 } 4552 } 4553 if (!skb1) 4554 __skb_queue_head(&tp->out_of_order_queue, skb); 4555 else 4556 __skb_queue_after(&tp->out_of_order_queue, skb1, skb); 4557 4558 /* And clean segments covered by new one as whole. */ 4559 while (!skb_queue_is_last(&tp->out_of_order_queue, skb)) { 4560 skb1 = skb_queue_next(&tp->out_of_order_queue, skb); 4561 4562 if (!after(end_seq, TCP_SKB_CB(skb1)->seq)) 4563 break; 4564 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) { 4565 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq, 4566 end_seq); 4567 break; 4568 } 4569 __skb_unlink(skb1, &tp->out_of_order_queue); 4570 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq, 4571 TCP_SKB_CB(skb1)->end_seq); 4572 __kfree_skb(skb1); 4573 } 4574 4575add_sack: 4576 if (tcp_is_sack(tp)) 4577 tcp_sack_new_ofo_skb(sk, seq, end_seq); 4578end: 4579 if (skb) 4580 skb_set_owner_r(skb, sk); 4581} 4582 4583 4584static void tcp_data_queue(struct sock *sk, struct sk_buff *skb) 4585{ 4586 const struct tcphdr *th = tcp_hdr(skb); 4587 struct tcp_sock *tp = tcp_sk(sk); 4588 int eaten = -1; 4589 4590 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) 4591 goto drop; 4592 4593 skb_dst_drop(skb); 4594 __skb_pull(skb, th->doff * 4); 4595 4596 TCP_ECN_accept_cwr(tp, skb); 4597 4598 tp->rx_opt.dsack = 0; 4599 4600 /* Queue data for delivery to the user. 4601 * Packets in sequence go to the receive queue. 4602 * Out of sequence packets to the out_of_order_queue. 4603 */ 4604 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) { 4605 if (tcp_receive_window(tp) == 0) 4606 goto out_of_window; 4607 4608 /* Ok. In sequence. In window. */ 4609 if (tp->ucopy.task == current && 4610 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len && 4611 sock_owned_by_user(sk) && !tp->urg_data) { 4612 int chunk = min_t(unsigned int, skb->len, 4613 tp->ucopy.len); 4614 4615 __set_current_state(TASK_RUNNING); 4616 4617 local_bh_enable(); 4618 if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) { 4619 tp->ucopy.len -= chunk; 4620 tp->copied_seq += chunk; 4621 eaten = (chunk == skb->len); 4622 tcp_rcv_space_adjust(sk); 4623 } 4624 local_bh_disable(); 4625 } 4626 4627 if (eaten <= 0) { 4628queue_and_out: 4629 if (eaten < 0 && 4630 tcp_try_rmem_schedule(sk, skb->truesize)) 4631 goto drop; 4632 4633 skb_set_owner_r(skb, sk); 4634 __skb_queue_tail(&sk->sk_receive_queue, skb); 4635 } 4636 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; 4637 if (skb->len) 4638 tcp_event_data_recv(sk, skb); 4639 if (th->fin) 4640 tcp_fin(sk); 4641 4642 if (!skb_queue_empty(&tp->out_of_order_queue)) { 4643 tcp_ofo_queue(sk); 4644 4645 /* RFC2581. 4.2. SHOULD send immediate ACK, when 4646 * gap in queue is filled. 4647 */ 4648 if (skb_queue_empty(&tp->out_of_order_queue)) 4649 inet_csk(sk)->icsk_ack.pingpong = 0; 4650 } 4651 4652 if (tp->rx_opt.num_sacks) 4653 tcp_sack_remove(tp); 4654 4655 tcp_fast_path_check(sk); 4656 4657 if (eaten > 0) 4658 __kfree_skb(skb); 4659 else if (!sock_flag(sk, SOCK_DEAD)) 4660 sk->sk_data_ready(sk, 0); 4661 return; 4662 } 4663 4664 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) { 4665 /* A retransmit, 2nd most common case. Force an immediate ack. */ 4666 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST); 4667 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); 4668 4669out_of_window: 4670 tcp_enter_quickack_mode(sk); 4671 inet_csk_schedule_ack(sk); 4672drop: 4673 __kfree_skb(skb); 4674 return; 4675 } 4676 4677 /* Out of window. F.e. zero window probe. */ 4678 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp))) 4679 goto out_of_window; 4680 4681 tcp_enter_quickack_mode(sk); 4682 4683 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 4684 /* Partial packet, seq < rcv_next < end_seq */ 4685 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n", 4686 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, 4687 TCP_SKB_CB(skb)->end_seq); 4688 4689 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt); 4690 4691 /* If window is closed, drop tail of packet. But after 4692 * remembering D-SACK for its head made in previous line. 4693 */ 4694 if (!tcp_receive_window(tp)) 4695 goto out_of_window; 4696 goto queue_and_out; 4697 } 4698 4699 tcp_data_queue_ofo(sk, skb); 4700} 4701 4702static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb, 4703 struct sk_buff_head *list) 4704{ 4705 struct sk_buff *next = NULL; 4706 4707 if (!skb_queue_is_last(list, skb)) 4708 next = skb_queue_next(list, skb); 4709 4710 __skb_unlink(skb, list); 4711 __kfree_skb(skb); 4712 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED); 4713 4714 return next; 4715} 4716 4717/* Collapse contiguous sequence of skbs head..tail with 4718 * sequence numbers start..end. 4719 * 4720 * If tail is NULL, this means until the end of the list. 4721 * 4722 * Segments with FIN/SYN are not collapsed (only because this 4723 * simplifies code) 4724 */ 4725static void 4726tcp_collapse(struct sock *sk, struct sk_buff_head *list, 4727 struct sk_buff *head, struct sk_buff *tail, 4728 u32 start, u32 end) 4729{ 4730 struct sk_buff *skb, *n; 4731 bool end_of_skbs; 4732 4733 /* First, check that queue is collapsible and find 4734 * the point where collapsing can be useful. */ 4735 skb = head; 4736restart: 4737 end_of_skbs = true; 4738 skb_queue_walk_from_safe(list, skb, n) { 4739 if (skb == tail) 4740 break; 4741 /* No new bits? It is possible on ofo queue. */ 4742 if (!before(start, TCP_SKB_CB(skb)->end_seq)) { 4743 skb = tcp_collapse_one(sk, skb, list); 4744 if (!skb) 4745 break; 4746 goto restart; 4747 } 4748 4749 /* The first skb to collapse is: 4750 * - not SYN/FIN and 4751 * - bloated or contains data before "start" or 4752 * overlaps to the next one. 4753 */ 4754 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin && 4755 (tcp_win_from_space(skb->truesize) > skb->len || 4756 before(TCP_SKB_CB(skb)->seq, start))) { 4757 end_of_skbs = false; 4758 break; 4759 } 4760 4761 if (!skb_queue_is_last(list, skb)) { 4762 struct sk_buff *next = skb_queue_next(list, skb); 4763 if (next != tail && 4764 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(next)->seq) { 4765 end_of_skbs = false; 4766 break; 4767 } 4768 } 4769 4770 /* Decided to skip this, advance start seq. */ 4771 start = TCP_SKB_CB(skb)->end_seq; 4772 } 4773 if (end_of_skbs || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin) 4774 return; 4775 4776 while (before(start, end)) { 4777 struct sk_buff *nskb; 4778 unsigned int header = skb_headroom(skb); 4779 int copy = SKB_MAX_ORDER(header, 0); 4780 4781 /* Too big header? This can happen with IPv6. */ 4782 if (copy < 0) 4783 return; 4784 if (end - start < copy) 4785 copy = end - start; 4786 nskb = alloc_skb(copy + header, GFP_ATOMIC); 4787 if (!nskb) 4788 return; 4789 4790 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head); 4791 skb_set_network_header(nskb, (skb_network_header(skb) - 4792 skb->head)); 4793 skb_set_transport_header(nskb, (skb_transport_header(skb) - 4794 skb->head)); 4795 skb_reserve(nskb, header); 4796 memcpy(nskb->head, skb->head, header); 4797 memcpy(nskb->cb, skb->cb, sizeof(skb->cb)); 4798 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start; 4799 __skb_queue_before(list, skb, nskb); 4800 skb_set_owner_r(nskb, sk); 4801 4802 /* Copy data, releasing collapsed skbs. */ 4803 while (copy > 0) { 4804 int offset = start - TCP_SKB_CB(skb)->seq; 4805 int size = TCP_SKB_CB(skb)->end_seq - start; 4806 4807 BUG_ON(offset < 0); 4808 if (size > 0) { 4809 size = min(copy, size); 4810 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size)) 4811 BUG(); 4812 TCP_SKB_CB(nskb)->end_seq += size; 4813 copy -= size; 4814 start += size; 4815 } 4816 if (!before(start, TCP_SKB_CB(skb)->end_seq)) { 4817 skb = tcp_collapse_one(sk, skb, list); 4818 if (!skb || 4819 skb == tail || 4820 tcp_hdr(skb)->syn || 4821 tcp_hdr(skb)->fin) 4822 return; 4823 } 4824 } 4825 } 4826} 4827 4828/* Collapse ofo queue. Algorithm: select contiguous sequence of skbs 4829 * and tcp_collapse() them until all the queue is collapsed. 4830 */ 4831static void tcp_collapse_ofo_queue(struct sock *sk) 4832{ 4833 struct tcp_sock *tp = tcp_sk(sk); 4834 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue); 4835 struct sk_buff *head; 4836 u32 start, end; 4837 4838 if (skb == NULL) 4839 return; 4840 4841 start = TCP_SKB_CB(skb)->seq; 4842 end = TCP_SKB_CB(skb)->end_seq; 4843 head = skb; 4844 4845 for (;;) { 4846 struct sk_buff *next = NULL; 4847 4848 if (!skb_queue_is_last(&tp->out_of_order_queue, skb)) 4849 next = skb_queue_next(&tp->out_of_order_queue, skb); 4850 skb = next; 4851 4852 /* Segment is terminated when we see gap or when 4853 * we are at the end of all the queue. */ 4854 if (!skb || 4855 after(TCP_SKB_CB(skb)->seq, end) || 4856 before(TCP_SKB_CB(skb)->end_seq, start)) { 4857 tcp_collapse(sk, &tp->out_of_order_queue, 4858 head, skb, start, end); 4859 head = skb; 4860 if (!skb) 4861 break; 4862 /* Start new segment */ 4863 start = TCP_SKB_CB(skb)->seq; 4864 end = TCP_SKB_CB(skb)->end_seq; 4865 } else { 4866 if (before(TCP_SKB_CB(skb)->seq, start)) 4867 start = TCP_SKB_CB(skb)->seq; 4868 if (after(TCP_SKB_CB(skb)->end_seq, end)) 4869 end = TCP_SKB_CB(skb)->end_seq; 4870 } 4871 } 4872} 4873 4874/* 4875 * Purge the out-of-order queue. 4876 * Return true if queue was pruned. 4877 */ 4878static int tcp_prune_ofo_queue(struct sock *sk) 4879{ 4880 struct tcp_sock *tp = tcp_sk(sk); 4881 int res = 0; 4882 4883 if (!skb_queue_empty(&tp->out_of_order_queue)) { 4884 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED); 4885 __skb_queue_purge(&tp->out_of_order_queue); 4886 4887 /* Reset SACK state. A conforming SACK implementation will 4888 * do the same at a timeout based retransmit. When a connection 4889 * is in a sad state like this, we care only about integrity 4890 * of the connection not performance. 4891 */ 4892 if (tp->rx_opt.sack_ok) 4893 tcp_sack_reset(&tp->rx_opt); 4894 sk_mem_reclaim(sk); 4895 res = 1; 4896 } 4897 return res; 4898} 4899 4900/* Reduce allocated memory if we can, trying to get 4901 * the socket within its memory limits again. 4902 * 4903 * Return less than zero if we should start dropping frames 4904 * until the socket owning process reads some of the data 4905 * to stabilize the situation. 4906 */ 4907static int tcp_prune_queue(struct sock *sk) 4908{ 4909 struct tcp_sock *tp = tcp_sk(sk); 4910 4911 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq); 4912 4913 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED); 4914 4915 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf) 4916 tcp_clamp_window(sk); 4917 else if (sk_under_memory_pressure(sk)) 4918 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss); 4919 4920 tcp_collapse_ofo_queue(sk); 4921 if (!skb_queue_empty(&sk->sk_receive_queue)) 4922 tcp_collapse(sk, &sk->sk_receive_queue, 4923 skb_peek(&sk->sk_receive_queue), 4924 NULL, 4925 tp->copied_seq, tp->rcv_nxt); 4926 sk_mem_reclaim(sk); 4927 4928 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) 4929 return 0; 4930 4931 /* Collapsing did not help, destructive actions follow. 4932 * This must not ever occur. */ 4933 4934 tcp_prune_ofo_queue(sk); 4935 4936 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) 4937 return 0; 4938 4939 /* If we are really being abused, tell the caller to silently 4940 * drop receive data on the floor. It will get retransmitted 4941 * and hopefully then we'll have sufficient space. 4942 */ 4943 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED); 4944 4945 /* Massive buffer overcommit. */ 4946 tp->pred_flags = 0; 4947 return -1; 4948} 4949 4950/* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. 4951 * As additional protections, we do not touch cwnd in retransmission phases, 4952 * and if application hit its sndbuf limit recently. 4953 */ 4954void tcp_cwnd_application_limited(struct sock *sk) 4955{ 4956 struct tcp_sock *tp = tcp_sk(sk); 4957 4958 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && 4959 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { 4960 /* Limited by application or receiver window. */ 4961 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); 4962 u32 win_used = max(tp->snd_cwnd_used, init_win); 4963 if (win_used < tp->snd_cwnd) { 4964 tp->snd_ssthresh = tcp_current_ssthresh(sk); 4965 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1; 4966 } 4967 tp->snd_cwnd_used = 0; 4968 } 4969 tp->snd_cwnd_stamp = tcp_time_stamp; 4970} 4971 4972static int tcp_should_expand_sndbuf(const struct sock *sk) 4973{ 4974 const struct tcp_sock *tp = tcp_sk(sk); 4975 4976 /* If the user specified a specific send buffer setting, do 4977 * not modify it. 4978 */ 4979 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK) 4980 return 0; 4981 4982 /* If we are under global TCP memory pressure, do not expand. */ 4983 if (sk_under_memory_pressure(sk)) 4984 return 0; 4985 4986 /* If we are under soft global TCP memory pressure, do not expand. */ 4987 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0)) 4988 return 0; 4989 4990 /* If we filled the congestion window, do not expand. */ 4991 if (tp->packets_out >= tp->snd_cwnd) 4992 return 0; 4993 4994 return 1; 4995} 4996 4997/* When incoming ACK allowed to free some skb from write_queue, 4998 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket 4999 * on the exit from tcp input handler. 5000 * 5001 * PROBLEM: sndbuf expansion does not work well with largesend. 5002 */ 5003static void tcp_new_space(struct sock *sk) 5004{ 5005 struct tcp_sock *tp = tcp_sk(sk); 5006 5007 if (tcp_should_expand_sndbuf(sk)) { 5008 int sndmem = SKB_TRUESIZE(max_t(u32, 5009 tp->rx_opt.mss_clamp, 5010 tp->mss_cache) + 5011 MAX_TCP_HEADER); 5012 int demanded = max_t(unsigned int, tp->snd_cwnd, 5013 tp->reordering + 1); 5014 sndmem *= 2 * demanded; 5015 if (sndmem > sk->sk_sndbuf) 5016 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]); 5017 tp->snd_cwnd_stamp = tcp_time_stamp; 5018 } 5019 5020 sk->sk_write_space(sk); 5021} 5022 5023static void tcp_check_space(struct sock *sk) 5024{ 5025 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) { 5026 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK); 5027 if (sk->sk_socket && 5028 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) 5029 tcp_new_space(sk); 5030 } 5031} 5032 5033static inline void tcp_data_snd_check(struct sock *sk) 5034{ 5035 tcp_push_pending_frames(sk); 5036 tcp_check_space(sk); 5037} 5038 5039/* 5040 * Check if sending an ack is needed. 5041 */ 5042static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible) 5043{ 5044 struct tcp_sock *tp = tcp_sk(sk); 5045 5046 /* More than one full frame received... */ 5047 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss && 5048 /* ... and right edge of window advances far enough. 5049 * (tcp_recvmsg() will send ACK otherwise). Or... 5050 */ 5051 __tcp_select_window(sk) >= tp->rcv_wnd) || 5052 /* We ACK each frame or... */ 5053 tcp_in_quickack_mode(sk) || 5054 /* We have out of order data. */ 5055 (ofo_possible && skb_peek(&tp->out_of_order_queue))) { 5056 /* Then ack it now */ 5057 tcp_send_ack(sk); 5058 } else { 5059 /* Else, send delayed ack. */ 5060 tcp_send_delayed_ack(sk); 5061 } 5062} 5063 5064static inline void tcp_ack_snd_check(struct sock *sk) 5065{ 5066 if (!inet_csk_ack_scheduled(sk)) { 5067 /* We sent a data segment already. */ 5068 return; 5069 } 5070 __tcp_ack_snd_check(sk, 1); 5071} 5072 5073/* 5074 * This routine is only called when we have urgent data 5075 * signaled. Its the 'slow' part of tcp_urg. It could be 5076 * moved inline now as tcp_urg is only called from one 5077 * place. We handle URGent data wrong. We have to - as 5078 * BSD still doesn't use the correction from RFC961. 5079 * For 1003.1g we should support a new option TCP_STDURG to permit 5080 * either form (or just set the sysctl tcp_stdurg). 5081 */ 5082 5083static void tcp_check_urg(struct sock *sk, const struct tcphdr *th) 5084{ 5085 struct tcp_sock *tp = tcp_sk(sk); 5086 u32 ptr = ntohs(th->urg_ptr); 5087 5088 if (ptr && !sysctl_tcp_stdurg) 5089 ptr--; 5090 ptr += ntohl(th->seq); 5091 5092 /* Ignore urgent data that we've already seen and read. */ 5093 if (after(tp->copied_seq, ptr)) 5094 return; 5095 5096 /* Do not replay urg ptr. 5097 * 5098 * NOTE: interesting situation not covered by specs. 5099 * Misbehaving sender may send urg ptr, pointing to segment, 5100 * which we already have in ofo queue. We are not able to fetch 5101 * such data and will stay in TCP_URG_NOTYET until will be eaten 5102 * by recvmsg(). Seems, we are not obliged to handle such wicked 5103 * situations. But it is worth to think about possibility of some 5104 * DoSes using some hypothetical application level deadlock. 5105 */ 5106 if (before(ptr, tp->rcv_nxt)) 5107 return; 5108 5109 /* Do we already have a newer (or duplicate) urgent pointer? */ 5110 if (tp->urg_data && !after(ptr, tp->urg_seq)) 5111 return; 5112 5113 /* Tell the world about our new urgent pointer. */ 5114 sk_send_sigurg(sk); 5115 5116 /* We may be adding urgent data when the last byte read was 5117 * urgent. To do this requires some care. We cannot just ignore 5118 * tp->copied_seq since we would read the last urgent byte again 5119 * as data, nor can we alter copied_seq until this data arrives 5120 * or we break the semantics of SIOCATMARK (and thus sockatmark()) 5121 * 5122 * NOTE. Double Dutch. Rendering to plain English: author of comment 5123 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB); 5124 * and expect that both A and B disappear from stream. This is _wrong_. 5125 * Though this happens in BSD with high probability, this is occasional. 5126 * Any application relying on this is buggy. Note also, that fix "works" 5127 * only in this artificial test. Insert some normal data between A and B and we will 5128 * decline of BSD again. Verdict: it is better to remove to trap 5129 * buggy users. 5130 */ 5131 if (tp->urg_seq == tp->copied_seq && tp->urg_data && 5132 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) { 5133 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue); 5134 tp->copied_seq++; 5135 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) { 5136 __skb_unlink(skb, &sk->sk_receive_queue); 5137 __kfree_skb(skb); 5138 } 5139 } 5140 5141 tp->urg_data = TCP_URG_NOTYET; 5142 tp->urg_seq = ptr; 5143 5144 /* Disable header prediction. */ 5145 tp->pred_flags = 0; 5146} 5147 5148/* This is the 'fast' part of urgent handling. */ 5149static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th) 5150{ 5151 struct tcp_sock *tp = tcp_sk(sk); 5152 5153 /* Check if we get a new urgent pointer - normally not. */ 5154 if (th->urg) 5155 tcp_check_urg(sk, th); 5156 5157 /* Do we wait for any urgent data? - normally not... */ 5158 if (tp->urg_data == TCP_URG_NOTYET) { 5159 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) - 5160 th->syn; 5161 5162 /* Is the urgent pointer pointing into this packet? */ 5163 if (ptr < skb->len) { 5164 u8 tmp; 5165 if (skb_copy_bits(skb, ptr, &tmp, 1)) 5166 BUG(); 5167 tp->urg_data = TCP_URG_VALID | tmp; 5168 if (!sock_flag(sk, SOCK_DEAD)) 5169 sk->sk_data_ready(sk, 0); 5170 } 5171 } 5172} 5173 5174static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen) 5175{ 5176 struct tcp_sock *tp = tcp_sk(sk); 5177 int chunk = skb->len - hlen; 5178 int err; 5179 5180 local_bh_enable(); 5181 if (skb_csum_unnecessary(skb)) 5182 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk); 5183 else 5184 err = skb_copy_and_csum_datagram_iovec(skb, hlen, 5185 tp->ucopy.iov); 5186 5187 if (!err) { 5188 tp->ucopy.len -= chunk; 5189 tp->copied_seq += chunk; 5190 tcp_rcv_space_adjust(sk); 5191 } 5192 5193 local_bh_disable(); 5194 return err; 5195} 5196 5197static __sum16 __tcp_checksum_complete_user(struct sock *sk, 5198 struct sk_buff *skb) 5199{ 5200 __sum16 result; 5201 5202 if (sock_owned_by_user(sk)) { 5203 local_bh_enable(); 5204 result = __tcp_checksum_complete(skb); 5205 local_bh_disable(); 5206 } else { 5207 result = __tcp_checksum_complete(skb); 5208 } 5209 return result; 5210} 5211 5212static inline int tcp_checksum_complete_user(struct sock *sk, 5213 struct sk_buff *skb) 5214{ 5215 return !skb_csum_unnecessary(skb) && 5216 __tcp_checksum_complete_user(sk, skb); 5217} 5218 5219#ifdef CONFIG_NET_DMA 5220static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb, 5221 int hlen) 5222{ 5223 struct tcp_sock *tp = tcp_sk(sk); 5224 int chunk = skb->len - hlen; 5225 int dma_cookie; 5226 int copied_early = 0; 5227 5228 if (tp->ucopy.wakeup) 5229 return 0; 5230 5231 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list) 5232 tp->ucopy.dma_chan = net_dma_find_channel(); 5233 5234 if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) { 5235 5236 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan, 5237 skb, hlen, 5238 tp->ucopy.iov, chunk, 5239 tp->ucopy.pinned_list); 5240 5241 if (dma_cookie < 0) 5242 goto out; 5243 5244 tp->ucopy.dma_cookie = dma_cookie; 5245 copied_early = 1; 5246 5247 tp->ucopy.len -= chunk; 5248 tp->copied_seq += chunk; 5249 tcp_rcv_space_adjust(sk); 5250 5251 if ((tp->ucopy.len == 0) || 5252 (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) || 5253 (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) { 5254 tp->ucopy.wakeup = 1; 5255 sk->sk_data_ready(sk, 0); 5256 } 5257 } else if (chunk > 0) { 5258 tp->ucopy.wakeup = 1; 5259 sk->sk_data_ready(sk, 0); 5260 } 5261out: 5262 return copied_early; 5263} 5264#endif /* CONFIG_NET_DMA */ 5265 5266/* Does PAWS and seqno based validation of an incoming segment, flags will 5267 * play significant role here. 5268 */ 5269static int tcp_validate_incoming(struct sock *sk, struct sk_buff *skb, 5270 const struct tcphdr *th, int syn_inerr) 5271{ 5272 const u8 *hash_location; 5273 struct tcp_sock *tp = tcp_sk(sk); 5274 5275 /* RFC1323: H1. Apply PAWS check first. */ 5276 if (tcp_fast_parse_options(skb, th, tp, &hash_location) && 5277 tp->rx_opt.saw_tstamp && 5278 tcp_paws_discard(sk, skb)) { 5279 if (!th->rst) { 5280 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED); 5281 tcp_send_dupack(sk, skb); 5282 goto discard; 5283 } 5284 /* Reset is accepted even if it did not pass PAWS. */ 5285 } 5286 5287 /* Step 1: check sequence number */ 5288 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) { 5289 /* RFC793, page 37: "In all states except SYN-SENT, all reset 5290 * (RST) segments are validated by checking their SEQ-fields." 5291 * And page 69: "If an incoming segment is not acceptable, 5292 * an acknowledgment should be sent in reply (unless the RST 5293 * bit is set, if so drop the segment and return)". 5294 */ 5295 if (!th->rst) 5296 tcp_send_dupack(sk, skb); 5297 goto discard; 5298 } 5299 5300 /* Step 2: check RST bit */ 5301 if (th->rst) { 5302 tcp_reset(sk); 5303 goto discard; 5304 } 5305 5306 /* ts_recent update must be made after we are sure that the packet 5307 * is in window. 5308 */ 5309 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq); 5310 5311 /* step 3: check security and precedence [ignored] */ 5312 5313 /* step 4: Check for a SYN in window. */ 5314 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 5315 if (syn_inerr) 5316 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS); 5317 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONSYN); 5318 tcp_reset(sk); 5319 return -1; 5320 } 5321 5322 return 1; 5323 5324discard: 5325 __kfree_skb(skb); 5326 return 0; 5327} 5328 5329/* 5330 * TCP receive function for the ESTABLISHED state. 5331 * 5332 * It is split into a fast path and a slow path. The fast path is 5333 * disabled when: 5334 * - A zero window was announced from us - zero window probing 5335 * is only handled properly in the slow path. 5336 * - Out of order segments arrived. 5337 * - Urgent data is expected. 5338 * - There is no buffer space left 5339 * - Unexpected TCP flags/window values/header lengths are received 5340 * (detected by checking the TCP header against pred_flags) 5341 * - Data is sent in both directions. Fast path only supports pure senders 5342 * or pure receivers (this means either the sequence number or the ack 5343 * value must stay constant) 5344 * - Unexpected TCP option. 5345 * 5346 * When these conditions are not satisfied it drops into a standard 5347 * receive procedure patterned after RFC793 to handle all cases. 5348 * The first three cases are guaranteed by proper pred_flags setting, 5349 * the rest is checked inline. Fast processing is turned on in 5350 * tcp_data_queue when everything is OK. 5351 */ 5352int tcp_rcv_established(struct sock *sk, struct sk_buff *skb, 5353 const struct tcphdr *th, unsigned int len) 5354{ 5355 struct tcp_sock *tp = tcp_sk(sk); 5356 int res; 5357 5358 /* 5359 * Header prediction. 5360 * The code loosely follows the one in the famous 5361 * "30 instruction TCP receive" Van Jacobson mail. 5362 * 5363 * Van's trick is to deposit buffers into socket queue 5364 * on a device interrupt, to call tcp_recv function 5365 * on the receive process context and checksum and copy 5366 * the buffer to user space. smart... 5367 * 5368 * Our current scheme is not silly either but we take the 5369 * extra cost of the net_bh soft interrupt processing... 5370 * We do checksum and copy also but from device to kernel. 5371 */ 5372 5373 tp->rx_opt.saw_tstamp = 0; 5374 5375 /* pred_flags is 0xS?10 << 16 + snd_wnd 5376 * if header_prediction is to be made 5377 * 'S' will always be tp->tcp_header_len >> 2 5378 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to 5379 * turn it off (when there are holes in the receive 5380 * space for instance) 5381 * PSH flag is ignored. 5382 */ 5383 5384 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags && 5385 TCP_SKB_CB(skb)->seq == tp->rcv_nxt && 5386 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) { 5387 int tcp_header_len = tp->tcp_header_len; 5388 5389 /* Timestamp header prediction: tcp_header_len 5390 * is automatically equal to th->doff*4 due to pred_flags 5391 * match. 5392 */ 5393 5394 /* Check timestamp */ 5395 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) { 5396 /* No? Slow path! */ 5397 if (!tcp_parse_aligned_timestamp(tp, th)) 5398 goto slow_path; 5399 5400 /* If PAWS failed, check it more carefully in slow path */ 5401 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0) 5402 goto slow_path; 5403 5404 /* DO NOT update ts_recent here, if checksum fails 5405 * and timestamp was corrupted part, it will result 5406 * in a hung connection since we will drop all 5407 * future packets due to the PAWS test. 5408 */ 5409 } 5410 5411 if (len <= tcp_header_len) { 5412 /* Bulk data transfer: sender */ 5413 if (len == tcp_header_len) { 5414 /* Predicted packet is in window by definition. 5415 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 5416 * Hence, check seq<=rcv_wup reduces to: 5417 */ 5418 if (tcp_header_len == 5419 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && 5420 tp->rcv_nxt == tp->rcv_wup) 5421 tcp_store_ts_recent(tp); 5422 5423 /* We know that such packets are checksummed 5424 * on entry. 5425 */ 5426 tcp_ack(sk, skb, 0); 5427 __kfree_skb(skb); 5428 tcp_data_snd_check(sk); 5429 return 0; 5430 } else { /* Header too small */ 5431 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS); 5432 goto discard; 5433 } 5434 } else { 5435 int eaten = 0; 5436 int copied_early = 0; 5437 5438 if (tp->copied_seq == tp->rcv_nxt && 5439 len - tcp_header_len <= tp->ucopy.len) { 5440#ifdef CONFIG_NET_DMA 5441 if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) { 5442 copied_early = 1; 5443 eaten = 1; 5444 } 5445#endif 5446 if (tp->ucopy.task == current && 5447 sock_owned_by_user(sk) && !copied_early) { 5448 __set_current_state(TASK_RUNNING); 5449 5450 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len)) 5451 eaten = 1; 5452 } 5453 if (eaten) { 5454 /* Predicted packet is in window by definition. 5455 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 5456 * Hence, check seq<=rcv_wup reduces to: 5457 */ 5458 if (tcp_header_len == 5459 (sizeof(struct tcphdr) + 5460 TCPOLEN_TSTAMP_ALIGNED) && 5461 tp->rcv_nxt == tp->rcv_wup) 5462 tcp_store_ts_recent(tp); 5463 5464 tcp_rcv_rtt_measure_ts(sk, skb); 5465 5466 __skb_pull(skb, tcp_header_len); 5467 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; 5468 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER); 5469 } 5470 if (copied_early) 5471 tcp_cleanup_rbuf(sk, skb->len); 5472 } 5473 if (!eaten) { 5474 if (tcp_checksum_complete_user(sk, skb)) 5475 goto csum_error; 5476 5477 /* Predicted packet is in window by definition. 5478 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 5479 * Hence, check seq<=rcv_wup reduces to: 5480 */ 5481 if (tcp_header_len == 5482 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && 5483 tp->rcv_nxt == tp->rcv_wup) 5484 tcp_store_ts_recent(tp); 5485 5486 tcp_rcv_rtt_measure_ts(sk, skb); 5487 5488 if ((int)skb->truesize > sk->sk_forward_alloc) 5489 goto step5; 5490 5491 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS); 5492 5493 /* Bulk data transfer: receiver */ 5494 __skb_pull(skb, tcp_header_len); 5495 __skb_queue_tail(&sk->sk_receive_queue, skb); 5496 skb_set_owner_r(skb, sk); 5497 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; 5498 } 5499 5500 tcp_event_data_recv(sk, skb); 5501 5502 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) { 5503 /* Well, only one small jumplet in fast path... */ 5504 tcp_ack(sk, skb, FLAG_DATA); 5505 tcp_data_snd_check(sk); 5506 if (!inet_csk_ack_scheduled(sk)) 5507 goto no_ack; 5508 } 5509 5510 if (!copied_early || tp->rcv_nxt != tp->rcv_wup) 5511 __tcp_ack_snd_check(sk, 0); 5512no_ack: 5513#ifdef CONFIG_NET_DMA 5514 if (copied_early) 5515 __skb_queue_tail(&sk->sk_async_wait_queue, skb); 5516 else 5517#endif 5518 if (eaten) 5519 __kfree_skb(skb); 5520 else 5521 sk->sk_data_ready(sk, 0); 5522 return 0; 5523 } 5524 } 5525 5526slow_path: 5527 if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb)) 5528 goto csum_error; 5529 5530 /* 5531 * Standard slow path. 5532 */ 5533 5534 res = tcp_validate_incoming(sk, skb, th, 1); 5535 if (res <= 0) 5536 return -res; 5537 5538step5: 5539 if (th->ack && tcp_ack(sk, skb, FLAG_SLOWPATH) < 0) 5540 goto discard; 5541 5542 tcp_rcv_rtt_measure_ts(sk, skb); 5543 5544 /* Process urgent data. */ 5545 tcp_urg(sk, skb, th); 5546 5547 /* step 7: process the segment text */ 5548 tcp_data_queue(sk, skb); 5549 5550 tcp_data_snd_check(sk); 5551 tcp_ack_snd_check(sk); 5552 return 0; 5553 5554csum_error: 5555 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS); 5556 5557discard: 5558 __kfree_skb(skb); 5559 return 0; 5560} 5561EXPORT_SYMBOL(tcp_rcv_established); 5562 5563static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb, 5564 const struct tcphdr *th, unsigned int len) 5565{ 5566 const u8 *hash_location; 5567 struct inet_connection_sock *icsk = inet_csk(sk); 5568 struct tcp_sock *tp = tcp_sk(sk); 5569 struct tcp_cookie_values *cvp = tp->cookie_values; 5570 int saved_clamp = tp->rx_opt.mss_clamp; 5571 5572 tcp_parse_options(skb, &tp->rx_opt, &hash_location, 0); 5573 5574 if (th->ack) { 5575 /* rfc793: 5576 * "If the state is SYN-SENT then 5577 * first check the ACK bit 5578 * If the ACK bit is set 5579 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send 5580 * a reset (unless the RST bit is set, if so drop 5581 * the segment and return)" 5582 * 5583 * We do not send data with SYN, so that RFC-correct 5584 * test reduces to: 5585 */ 5586 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt) 5587 goto reset_and_undo; 5588 5589 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 5590 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp, 5591 tcp_time_stamp)) { 5592 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED); 5593 goto reset_and_undo; 5594 } 5595 5596 /* Now ACK is acceptable. 5597 * 5598 * "If the RST bit is set 5599 * If the ACK was acceptable then signal the user "error: 5600 * connection reset", drop the segment, enter CLOSED state, 5601 * delete TCB, and return." 5602 */ 5603 5604 if (th->rst) { 5605 tcp_reset(sk); 5606 goto discard; 5607 } 5608 5609 /* rfc793: 5610 * "fifth, if neither of the SYN or RST bits is set then 5611 * drop the segment and return." 5612 * 5613 * See note below! 5614 * --ANK(990513) 5615 */ 5616 if (!th->syn) 5617 goto discard_and_undo; 5618 5619 /* rfc793: 5620 * "If the SYN bit is on ... 5621 * are acceptable then ... 5622 * (our SYN has been ACKed), change the connection 5623 * state to ESTABLISHED..." 5624 */ 5625 5626 TCP_ECN_rcv_synack(tp, th); 5627 5628 tp->snd_wl1 = TCP_SKB_CB(skb)->seq; 5629 tcp_ack(sk, skb, FLAG_SLOWPATH); 5630 5631 /* Ok.. it's good. Set up sequence numbers and 5632 * move to established. 5633 */ 5634 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; 5635 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; 5636 5637 /* RFC1323: The window in SYN & SYN/ACK segments is 5638 * never scaled. 5639 */ 5640 tp->snd_wnd = ntohs(th->window); 5641 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq); 5642 5643 if (!tp->rx_opt.wscale_ok) { 5644 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0; 5645 tp->window_clamp = min(tp->window_clamp, 65535U); 5646 } 5647 5648 if (tp->rx_opt.saw_tstamp) { 5649 tp->rx_opt.tstamp_ok = 1; 5650 tp->tcp_header_len = 5651 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 5652 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; 5653 tcp_store_ts_recent(tp); 5654 } else { 5655 tp->tcp_header_len = sizeof(struct tcphdr); 5656 } 5657 5658 if (tcp_is_sack(tp) && sysctl_tcp_fack) 5659 tcp_enable_fack(tp); 5660 5661 tcp_mtup_init(sk); 5662 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 5663 tcp_initialize_rcv_mss(sk); 5664 5665 /* Remember, tcp_poll() does not lock socket! 5666 * Change state from SYN-SENT only after copied_seq 5667 * is initialized. */ 5668 tp->copied_seq = tp->rcv_nxt; 5669 5670 if (cvp != NULL && 5671 cvp->cookie_pair_size > 0 && 5672 tp->rx_opt.cookie_plus > 0) { 5673 int cookie_size = tp->rx_opt.cookie_plus 5674 - TCPOLEN_COOKIE_BASE; 5675 int cookie_pair_size = cookie_size 5676 + cvp->cookie_desired; 5677 5678 /* A cookie extension option was sent and returned. 5679 * Note that each incoming SYNACK replaces the 5680 * Responder cookie. The initial exchange is most 5681 * fragile, as protection against spoofing relies 5682 * entirely upon the sequence and timestamp (above). 5683 * This replacement strategy allows the correct pair to 5684 * pass through, while any others will be filtered via 5685 * Responder verification later. 5686 */ 5687 if (sizeof(cvp->cookie_pair) >= cookie_pair_size) { 5688 memcpy(&cvp->cookie_pair[cvp->cookie_desired], 5689 hash_location, cookie_size); 5690 cvp->cookie_pair_size = cookie_pair_size; 5691 } 5692 } 5693 5694 smp_mb(); 5695 tcp_set_state(sk, TCP_ESTABLISHED); 5696 5697 security_inet_conn_established(sk, skb); 5698 5699 /* Make sure socket is routed, for correct metrics. */ 5700 icsk->icsk_af_ops->rebuild_header(sk); 5701 5702 tcp_init_metrics(sk); 5703 5704 tcp_init_congestion_control(sk); 5705 5706 /* Prevent spurious tcp_cwnd_restart() on first data 5707 * packet. 5708 */ 5709 tp->lsndtime = tcp_time_stamp; 5710 5711 tcp_init_buffer_space(sk); 5712 5713 if (sock_flag(sk, SOCK_KEEPOPEN)) 5714 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp)); 5715 5716 if (!tp->rx_opt.snd_wscale) 5717 __tcp_fast_path_on(tp, tp->snd_wnd); 5718 else 5719 tp->pred_flags = 0; 5720 5721 if (!sock_flag(sk, SOCK_DEAD)) { 5722 sk->sk_state_change(sk); 5723 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT); 5724 } 5725 5726 if (sk->sk_write_pending || 5727 icsk->icsk_accept_queue.rskq_defer_accept || 5728 icsk->icsk_ack.pingpong) { 5729 /* Save one ACK. Data will be ready after 5730 * several ticks, if write_pending is set. 5731 * 5732 * It may be deleted, but with this feature tcpdumps 5733 * look so _wonderfully_ clever, that I was not able 5734 * to stand against the temptation 8) --ANK 5735 */ 5736 inet_csk_schedule_ack(sk); 5737 icsk->icsk_ack.lrcvtime = tcp_time_stamp; 5738 icsk->icsk_ack.ato = TCP_ATO_MIN; 5739 tcp_incr_quickack(sk); 5740 tcp_enter_quickack_mode(sk); 5741 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 5742 TCP_DELACK_MAX, TCP_RTO_MAX); 5743 5744discard: 5745 __kfree_skb(skb); 5746 return 0; 5747 } else { 5748 tcp_send_ack(sk); 5749 } 5750 return -1; 5751 } 5752 5753 /* No ACK in the segment */ 5754 5755 if (th->rst) { 5756 /* rfc793: 5757 * "If the RST bit is set 5758 * 5759 * Otherwise (no ACK) drop the segment and return." 5760 */ 5761 5762 goto discard_and_undo; 5763 } 5764 5765 /* PAWS check. */ 5766 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && 5767 tcp_paws_reject(&tp->rx_opt, 0)) 5768 goto discard_and_undo; 5769 5770 if (th->syn) { 5771 /* We see SYN without ACK. It is attempt of 5772 * simultaneous connect with crossed SYNs. 5773 * Particularly, it can be connect to self. 5774 */ 5775 tcp_set_state(sk, TCP_SYN_RECV); 5776 5777 if (tp->rx_opt.saw_tstamp) { 5778 tp->rx_opt.tstamp_ok = 1; 5779 tcp_store_ts_recent(tp); 5780 tp->tcp_header_len = 5781 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 5782 } else { 5783 tp->tcp_header_len = sizeof(struct tcphdr); 5784 } 5785 5786 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; 5787 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; 5788 5789 /* RFC1323: The window in SYN & SYN/ACK segments is 5790 * never scaled. 5791 */ 5792 tp->snd_wnd = ntohs(th->window); 5793 tp->snd_wl1 = TCP_SKB_CB(skb)->seq; 5794 tp->max_window = tp->snd_wnd; 5795 5796 TCP_ECN_rcv_syn(tp, th); 5797 5798 tcp_mtup_init(sk); 5799 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 5800 tcp_initialize_rcv_mss(sk); 5801 5802 tcp_send_synack(sk); 5803#if 0 5804 /* Note, we could accept data and URG from this segment. 5805 * There are no obstacles to make this. 5806 * 5807 * However, if we ignore data in ACKless segments sometimes, 5808 * we have no reasons to accept it sometimes. 5809 * Also, seems the code doing it in step6 of tcp_rcv_state_process 5810 * is not flawless. So, discard packet for sanity. 5811 * Uncomment this return to process the data. 5812 */ 5813 return -1; 5814#else 5815 goto discard; 5816#endif 5817 } 5818 /* "fifth, if neither of the SYN or RST bits is set then 5819 * drop the segment and return." 5820 */ 5821 5822discard_and_undo: 5823 tcp_clear_options(&tp->rx_opt); 5824 tp->rx_opt.mss_clamp = saved_clamp; 5825 goto discard; 5826 5827reset_and_undo: 5828 tcp_clear_options(&tp->rx_opt); 5829 tp->rx_opt.mss_clamp = saved_clamp; 5830 return 1; 5831} 5832 5833/* 5834 * This function implements the receiving procedure of RFC 793 for 5835 * all states except ESTABLISHED and TIME_WAIT. 5836 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be 5837 * address independent. 5838 */ 5839 5840int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb, 5841 const struct tcphdr *th, unsigned int len) 5842{ 5843 struct tcp_sock *tp = tcp_sk(sk); 5844 struct inet_connection_sock *icsk = inet_csk(sk); 5845 int queued = 0; 5846 int res; 5847 5848 tp->rx_opt.saw_tstamp = 0; 5849 5850 switch (sk->sk_state) { 5851 case TCP_CLOSE: 5852 goto discard; 5853 5854 case TCP_LISTEN: 5855 if (th->ack) 5856 return 1; 5857 5858 if (th->rst) 5859 goto discard; 5860 5861 if (th->syn) { 5862 if (th->fin) 5863 goto discard; 5864 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0) 5865 return 1; 5866 5867 /* Now we have several options: In theory there is 5868 * nothing else in the frame. KA9Q has an option to 5869 * send data with the syn, BSD accepts data with the 5870 * syn up to the [to be] advertised window and 5871 * Solaris 2.1 gives you a protocol error. For now 5872 * we just ignore it, that fits the spec precisely 5873 * and avoids incompatibilities. It would be nice in 5874 * future to drop through and process the data. 5875 * 5876 * Now that TTCP is starting to be used we ought to 5877 * queue this data. 5878 * But, this leaves one open to an easy denial of 5879 * service attack, and SYN cookies can't defend 5880 * against this problem. So, we drop the data 5881 * in the interest of security over speed unless 5882 * it's still in use. 5883 */ 5884 kfree_skb(skb); 5885 return 0; 5886 } 5887 goto discard; 5888 5889 case TCP_SYN_SENT: 5890 queued = tcp_rcv_synsent_state_process(sk, skb, th, len); 5891 if (queued >= 0) 5892 return queued; 5893 5894 /* Do step6 onward by hand. */ 5895 tcp_urg(sk, skb, th); 5896 __kfree_skb(skb); 5897 tcp_data_snd_check(sk); 5898 return 0; 5899 } 5900 5901 res = tcp_validate_incoming(sk, skb, th, 0); 5902 if (res <= 0) 5903 return -res; 5904 5905 /* step 5: check the ACK field */ 5906 if (th->ack) { 5907 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH) > 0; 5908 5909 switch (sk->sk_state) { 5910 case TCP_SYN_RECV: 5911 if (acceptable) { 5912 tp->copied_seq = tp->rcv_nxt; 5913 smp_mb(); 5914 tcp_set_state(sk, TCP_ESTABLISHED); 5915 sk->sk_state_change(sk); 5916 5917 /* Note, that this wakeup is only for marginal 5918 * crossed SYN case. Passively open sockets 5919 * are not waked up, because sk->sk_sleep == 5920 * NULL and sk->sk_socket == NULL. 5921 */ 5922 if (sk->sk_socket) 5923 sk_wake_async(sk, 5924 SOCK_WAKE_IO, POLL_OUT); 5925 5926 tp->snd_una = TCP_SKB_CB(skb)->ack_seq; 5927 tp->snd_wnd = ntohs(th->window) << 5928 tp->rx_opt.snd_wscale; 5929 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq); 5930 5931 if (tp->rx_opt.tstamp_ok) 5932 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; 5933 5934 /* Make sure socket is routed, for 5935 * correct metrics. 5936 */ 5937 icsk->icsk_af_ops->rebuild_header(sk); 5938 5939 tcp_init_metrics(sk); 5940 5941 tcp_init_congestion_control(sk); 5942 5943 /* Prevent spurious tcp_cwnd_restart() on 5944 * first data packet. 5945 */ 5946 tp->lsndtime = tcp_time_stamp; 5947 5948 tcp_mtup_init(sk); 5949 tcp_initialize_rcv_mss(sk); 5950 tcp_init_buffer_space(sk); 5951 tcp_fast_path_on(tp); 5952 } else { 5953 return 1; 5954 } 5955 break; 5956 5957 case TCP_FIN_WAIT1: 5958 if (tp->snd_una == tp->write_seq) { 5959 tcp_set_state(sk, TCP_FIN_WAIT2); 5960 sk->sk_shutdown |= SEND_SHUTDOWN; 5961 dst_confirm(__sk_dst_get(sk)); 5962 5963 if (!sock_flag(sk, SOCK_DEAD)) 5964 /* Wake up lingering close() */ 5965 sk->sk_state_change(sk); 5966 else { 5967 int tmo; 5968 5969 if (tp->linger2 < 0 || 5970 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 5971 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) { 5972 tcp_done(sk); 5973 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA); 5974 return 1; 5975 } 5976 5977 tmo = tcp_fin_time(sk); 5978 if (tmo > TCP_TIMEWAIT_LEN) { 5979 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN); 5980 } else if (th->fin || sock_owned_by_user(sk)) { 5981 /* Bad case. We could lose such FIN otherwise. 5982 * It is not a big problem, but it looks confusing 5983 * and not so rare event. We still can lose it now, 5984 * if it spins in bh_lock_sock(), but it is really 5985 * marginal case. 5986 */ 5987 inet_csk_reset_keepalive_timer(sk, tmo); 5988 } else { 5989 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo); 5990 goto discard; 5991 } 5992 } 5993 } 5994 break; 5995 5996 case TCP_CLOSING: 5997 if (tp->snd_una == tp->write_seq) { 5998 tcp_time_wait(sk, TCP_TIME_WAIT, 0); 5999 goto discard; 6000 } 6001 break; 6002 6003 case TCP_LAST_ACK: 6004 if (tp->snd_una == tp->write_seq) { 6005 tcp_update_metrics(sk); 6006 tcp_done(sk); 6007 goto discard; 6008 } 6009 break; 6010 } 6011 } else 6012 goto discard; 6013 6014 /* step 6: check the URG bit */ 6015 tcp_urg(sk, skb, th); 6016 6017 /* step 7: process the segment text */ 6018 switch (sk->sk_state) { 6019 case TCP_CLOSE_WAIT: 6020 case TCP_CLOSING: 6021 case TCP_LAST_ACK: 6022 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) 6023 break; 6024 case TCP_FIN_WAIT1: 6025 case TCP_FIN_WAIT2: 6026 /* RFC 793 says to queue data in these states, 6027 * RFC 1122 says we MUST send a reset. 6028 * BSD 4.4 also does reset. 6029 */ 6030 if (sk->sk_shutdown & RCV_SHUTDOWN) { 6031 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 6032 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) { 6033 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA); 6034 tcp_reset(sk); 6035 return 1; 6036 } 6037 } 6038 /* Fall through */ 6039 case TCP_ESTABLISHED: 6040 tcp_data_queue(sk, skb); 6041 queued = 1; 6042 break; 6043 } 6044 6045 /* tcp_data could move socket to TIME-WAIT */ 6046 if (sk->sk_state != TCP_CLOSE) { 6047 tcp_data_snd_check(sk); 6048 tcp_ack_snd_check(sk); 6049 } 6050 6051 if (!queued) { 6052discard: 6053 __kfree_skb(skb); 6054 } 6055 return 0; 6056} 6057EXPORT_SYMBOL(tcp_rcv_state_process); 6058