tcp_input.c revision 923dd347b8904c24bcac89bf038ed4da87f8aa90
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; 102int sysctl_tcp_early_retrans __read_mostly = 2; 103 104#define FLAG_DATA 0x01 /* Incoming frame contained data. */ 105#define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */ 106#define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */ 107#define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */ 108#define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */ 109#define FLAG_DATA_SACKED 0x20 /* New SACK. */ 110#define FLAG_ECE 0x40 /* ECE in this ACK */ 111#define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/ 112#define FLAG_ONLY_ORIG_SACKED 0x200 /* SACKs only non-rexmit sent before RTO */ 113#define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */ 114#define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */ 115#define FLAG_NONHEAD_RETRANS_ACKED 0x1000 /* Non-head rexmitted data was ACKed */ 116#define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */ 117 118#define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED) 119#define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED) 120#define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE) 121#define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED) 122#define FLAG_ANY_PROGRESS (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED) 123 124#define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH) 125#define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH)) 126 127/* Adapt the MSS value used to make delayed ack decision to the 128 * real world. 129 */ 130static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb) 131{ 132 struct inet_connection_sock *icsk = inet_csk(sk); 133 const unsigned int lss = icsk->icsk_ack.last_seg_size; 134 unsigned int len; 135 136 icsk->icsk_ack.last_seg_size = 0; 137 138 /* skb->len may jitter because of SACKs, even if peer 139 * sends good full-sized frames. 140 */ 141 len = skb_shinfo(skb)->gso_size ? : skb->len; 142 if (len >= icsk->icsk_ack.rcv_mss) { 143 icsk->icsk_ack.rcv_mss = len; 144 } else { 145 /* Otherwise, we make more careful check taking into account, 146 * that SACKs block is variable. 147 * 148 * "len" is invariant segment length, including TCP header. 149 */ 150 len += skb->data - skb_transport_header(skb); 151 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) || 152 /* If PSH is not set, packet should be 153 * full sized, provided peer TCP is not badly broken. 154 * This observation (if it is correct 8)) allows 155 * to handle super-low mtu links fairly. 156 */ 157 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) && 158 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) { 159 /* Subtract also invariant (if peer is RFC compliant), 160 * tcp header plus fixed timestamp option length. 161 * Resulting "len" is MSS free of SACK jitter. 162 */ 163 len -= tcp_sk(sk)->tcp_header_len; 164 icsk->icsk_ack.last_seg_size = len; 165 if (len == lss) { 166 icsk->icsk_ack.rcv_mss = len; 167 return; 168 } 169 } 170 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED) 171 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2; 172 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED; 173 } 174} 175 176static void tcp_incr_quickack(struct sock *sk) 177{ 178 struct inet_connection_sock *icsk = inet_csk(sk); 179 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss); 180 181 if (quickacks == 0) 182 quickacks = 2; 183 if (quickacks > icsk->icsk_ack.quick) 184 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS); 185} 186 187static void tcp_enter_quickack_mode(struct sock *sk) 188{ 189 struct inet_connection_sock *icsk = inet_csk(sk); 190 tcp_incr_quickack(sk); 191 icsk->icsk_ack.pingpong = 0; 192 icsk->icsk_ack.ato = TCP_ATO_MIN; 193} 194 195/* Send ACKs quickly, if "quick" count is not exhausted 196 * and the session is not interactive. 197 */ 198 199static inline int tcp_in_quickack_mode(const struct sock *sk) 200{ 201 const struct inet_connection_sock *icsk = inet_csk(sk); 202 return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong; 203} 204 205static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp) 206{ 207 if (tp->ecn_flags & TCP_ECN_OK) 208 tp->ecn_flags |= TCP_ECN_QUEUE_CWR; 209} 210 211static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb) 212{ 213 if (tcp_hdr(skb)->cwr) 214 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR; 215} 216 217static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp) 218{ 219 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR; 220} 221 222static inline void TCP_ECN_check_ce(struct tcp_sock *tp, const struct sk_buff *skb) 223{ 224 if (!(tp->ecn_flags & TCP_ECN_OK)) 225 return; 226 227 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) { 228 case INET_ECN_NOT_ECT: 229 /* Funny extension: if ECT is not set on a segment, 230 * and we already seen ECT on a previous segment, 231 * it is probably a retransmit. 232 */ 233 if (tp->ecn_flags & TCP_ECN_SEEN) 234 tcp_enter_quickack_mode((struct sock *)tp); 235 break; 236 case INET_ECN_CE: 237 tp->ecn_flags |= TCP_ECN_DEMAND_CWR; 238 /* fallinto */ 239 default: 240 tp->ecn_flags |= TCP_ECN_SEEN; 241 } 242} 243 244static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th) 245{ 246 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr)) 247 tp->ecn_flags &= ~TCP_ECN_OK; 248} 249 250static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th) 251{ 252 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr)) 253 tp->ecn_flags &= ~TCP_ECN_OK; 254} 255 256static inline int TCP_ECN_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th) 257{ 258 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK)) 259 return 1; 260 return 0; 261} 262 263/* Buffer size and advertised window tuning. 264 * 265 * 1. Tuning sk->sk_sndbuf, when connection enters established state. 266 */ 267 268static void tcp_fixup_sndbuf(struct sock *sk) 269{ 270 int sndmem = SKB_TRUESIZE(tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER); 271 272 sndmem *= TCP_INIT_CWND; 273 if (sk->sk_sndbuf < sndmem) 274 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]); 275} 276 277/* 2. Tuning advertised window (window_clamp, rcv_ssthresh) 278 * 279 * All tcp_full_space() is split to two parts: "network" buffer, allocated 280 * forward and advertised in receiver window (tp->rcv_wnd) and 281 * "application buffer", required to isolate scheduling/application 282 * latencies from network. 283 * window_clamp is maximal advertised window. It can be less than 284 * tcp_full_space(), in this case tcp_full_space() - window_clamp 285 * is reserved for "application" buffer. The less window_clamp is 286 * the smoother our behaviour from viewpoint of network, but the lower 287 * throughput and the higher sensitivity of the connection to losses. 8) 288 * 289 * rcv_ssthresh is more strict window_clamp used at "slow start" 290 * phase to predict further behaviour of this connection. 291 * It is used for two goals: 292 * - to enforce header prediction at sender, even when application 293 * requires some significant "application buffer". It is check #1. 294 * - to prevent pruning of receive queue because of misprediction 295 * of receiver window. Check #2. 296 * 297 * The scheme does not work when sender sends good segments opening 298 * window and then starts to feed us spaghetti. But it should work 299 * in common situations. Otherwise, we have to rely on queue collapsing. 300 */ 301 302/* Slow part of check#2. */ 303static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb) 304{ 305 struct tcp_sock *tp = tcp_sk(sk); 306 /* Optimize this! */ 307 int truesize = tcp_win_from_space(skb->truesize) >> 1; 308 int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1; 309 310 while (tp->rcv_ssthresh <= window) { 311 if (truesize <= skb->len) 312 return 2 * inet_csk(sk)->icsk_ack.rcv_mss; 313 314 truesize >>= 1; 315 window >>= 1; 316 } 317 return 0; 318} 319 320static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb) 321{ 322 struct tcp_sock *tp = tcp_sk(sk); 323 324 /* Check #1 */ 325 if (tp->rcv_ssthresh < tp->window_clamp && 326 (int)tp->rcv_ssthresh < tcp_space(sk) && 327 !sk_under_memory_pressure(sk)) { 328 int incr; 329 330 /* Check #2. Increase window, if skb with such overhead 331 * will fit to rcvbuf in future. 332 */ 333 if (tcp_win_from_space(skb->truesize) <= skb->len) 334 incr = 2 * tp->advmss; 335 else 336 incr = __tcp_grow_window(sk, skb); 337 338 if (incr) { 339 incr = max_t(int, incr, 2 * skb->len); 340 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr, 341 tp->window_clamp); 342 inet_csk(sk)->icsk_ack.quick |= 1; 343 } 344 } 345} 346 347/* 3. Tuning rcvbuf, when connection enters established state. */ 348 349static void tcp_fixup_rcvbuf(struct sock *sk) 350{ 351 u32 mss = tcp_sk(sk)->advmss; 352 u32 icwnd = TCP_DEFAULT_INIT_RCVWND; 353 int rcvmem; 354 355 /* Limit to 10 segments if mss <= 1460, 356 * or 14600/mss segments, with a minimum of two segments. 357 */ 358 if (mss > 1460) 359 icwnd = max_t(u32, (1460 * TCP_DEFAULT_INIT_RCVWND) / mss, 2); 360 361 rcvmem = SKB_TRUESIZE(mss + MAX_TCP_HEADER); 362 while (tcp_win_from_space(rcvmem) < mss) 363 rcvmem += 128; 364 365 rcvmem *= icwnd; 366 367 if (sk->sk_rcvbuf < rcvmem) 368 sk->sk_rcvbuf = min(rcvmem, sysctl_tcp_rmem[2]); 369} 370 371/* 4. Try to fixup all. It is made immediately after connection enters 372 * established state. 373 */ 374static void tcp_init_buffer_space(struct sock *sk) 375{ 376 struct tcp_sock *tp = tcp_sk(sk); 377 int maxwin; 378 379 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) 380 tcp_fixup_rcvbuf(sk); 381 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK)) 382 tcp_fixup_sndbuf(sk); 383 384 tp->rcvq_space.space = tp->rcv_wnd; 385 386 maxwin = tcp_full_space(sk); 387 388 if (tp->window_clamp >= maxwin) { 389 tp->window_clamp = maxwin; 390 391 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss) 392 tp->window_clamp = max(maxwin - 393 (maxwin >> sysctl_tcp_app_win), 394 4 * tp->advmss); 395 } 396 397 /* Force reservation of one segment. */ 398 if (sysctl_tcp_app_win && 399 tp->window_clamp > 2 * tp->advmss && 400 tp->window_clamp + tp->advmss > maxwin) 401 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss); 402 403 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp); 404 tp->snd_cwnd_stamp = tcp_time_stamp; 405} 406 407/* 5. Recalculate window clamp after socket hit its memory bounds. */ 408static void tcp_clamp_window(struct sock *sk) 409{ 410 struct tcp_sock *tp = tcp_sk(sk); 411 struct inet_connection_sock *icsk = inet_csk(sk); 412 413 icsk->icsk_ack.quick = 0; 414 415 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] && 416 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) && 417 !sk_under_memory_pressure(sk) && 418 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) { 419 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc), 420 sysctl_tcp_rmem[2]); 421 } 422 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf) 423 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss); 424} 425 426/* Initialize RCV_MSS value. 427 * RCV_MSS is an our guess about MSS used by the peer. 428 * We haven't any direct information about the MSS. 429 * It's better to underestimate the RCV_MSS rather than overestimate. 430 * Overestimations make us ACKing less frequently than needed. 431 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss(). 432 */ 433void tcp_initialize_rcv_mss(struct sock *sk) 434{ 435 const struct tcp_sock *tp = tcp_sk(sk); 436 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache); 437 438 hint = min(hint, tp->rcv_wnd / 2); 439 hint = min(hint, TCP_MSS_DEFAULT); 440 hint = max(hint, TCP_MIN_MSS); 441 442 inet_csk(sk)->icsk_ack.rcv_mss = hint; 443} 444EXPORT_SYMBOL(tcp_initialize_rcv_mss); 445 446/* Receiver "autotuning" code. 447 * 448 * The algorithm for RTT estimation w/o timestamps is based on 449 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL. 450 * <http://public.lanl.gov/radiant/pubs.html#DRS> 451 * 452 * More detail on this code can be found at 453 * <http://staff.psc.edu/jheffner/>, 454 * though this reference is out of date. A new paper 455 * is pending. 456 */ 457static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep) 458{ 459 u32 new_sample = tp->rcv_rtt_est.rtt; 460 long m = sample; 461 462 if (m == 0) 463 m = 1; 464 465 if (new_sample != 0) { 466 /* If we sample in larger samples in the non-timestamp 467 * case, we could grossly overestimate the RTT especially 468 * with chatty applications or bulk transfer apps which 469 * are stalled on filesystem I/O. 470 * 471 * Also, since we are only going for a minimum in the 472 * non-timestamp case, we do not smooth things out 473 * else with timestamps disabled convergence takes too 474 * long. 475 */ 476 if (!win_dep) { 477 m -= (new_sample >> 3); 478 new_sample += m; 479 } else { 480 m <<= 3; 481 if (m < new_sample) 482 new_sample = m; 483 } 484 } else { 485 /* No previous measure. */ 486 new_sample = m << 3; 487 } 488 489 if (tp->rcv_rtt_est.rtt != new_sample) 490 tp->rcv_rtt_est.rtt = new_sample; 491} 492 493static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp) 494{ 495 if (tp->rcv_rtt_est.time == 0) 496 goto new_measure; 497 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq)) 498 return; 499 tcp_rcv_rtt_update(tp, jiffies - tp->rcv_rtt_est.time, 1); 500 501new_measure: 502 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd; 503 tp->rcv_rtt_est.time = tcp_time_stamp; 504} 505 506static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, 507 const struct sk_buff *skb) 508{ 509 struct tcp_sock *tp = tcp_sk(sk); 510 if (tp->rx_opt.rcv_tsecr && 511 (TCP_SKB_CB(skb)->end_seq - 512 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss)) 513 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0); 514} 515 516/* 517 * This function should be called every time data is copied to user space. 518 * It calculates the appropriate TCP receive buffer space. 519 */ 520void tcp_rcv_space_adjust(struct sock *sk) 521{ 522 struct tcp_sock *tp = tcp_sk(sk); 523 int time; 524 int space; 525 526 if (tp->rcvq_space.time == 0) 527 goto new_measure; 528 529 time = tcp_time_stamp - tp->rcvq_space.time; 530 if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0) 531 return; 532 533 space = 2 * (tp->copied_seq - tp->rcvq_space.seq); 534 535 space = max(tp->rcvq_space.space, space); 536 537 if (tp->rcvq_space.space != space) { 538 int rcvmem; 539 540 tp->rcvq_space.space = space; 541 542 if (sysctl_tcp_moderate_rcvbuf && 543 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) { 544 int new_clamp = space; 545 546 /* Receive space grows, normalize in order to 547 * take into account packet headers and sk_buff 548 * structure overhead. 549 */ 550 space /= tp->advmss; 551 if (!space) 552 space = 1; 553 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER); 554 while (tcp_win_from_space(rcvmem) < tp->advmss) 555 rcvmem += 128; 556 space *= rcvmem; 557 space = min(space, sysctl_tcp_rmem[2]); 558 if (space > sk->sk_rcvbuf) { 559 sk->sk_rcvbuf = space; 560 561 /* Make the window clamp follow along. */ 562 tp->window_clamp = new_clamp; 563 } 564 } 565 } 566 567new_measure: 568 tp->rcvq_space.seq = tp->copied_seq; 569 tp->rcvq_space.time = tcp_time_stamp; 570} 571 572/* There is something which you must keep in mind when you analyze the 573 * behavior of the tp->ato delayed ack timeout interval. When a 574 * connection starts up, we want to ack as quickly as possible. The 575 * problem is that "good" TCP's do slow start at the beginning of data 576 * transmission. The means that until we send the first few ACK's the 577 * sender will sit on his end and only queue most of his data, because 578 * he can only send snd_cwnd unacked packets at any given time. For 579 * each ACK we send, he increments snd_cwnd and transmits more of his 580 * queue. -DaveM 581 */ 582static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb) 583{ 584 struct tcp_sock *tp = tcp_sk(sk); 585 struct inet_connection_sock *icsk = inet_csk(sk); 586 u32 now; 587 588 inet_csk_schedule_ack(sk); 589 590 tcp_measure_rcv_mss(sk, skb); 591 592 tcp_rcv_rtt_measure(tp); 593 594 now = tcp_time_stamp; 595 596 if (!icsk->icsk_ack.ato) { 597 /* The _first_ data packet received, initialize 598 * delayed ACK engine. 599 */ 600 tcp_incr_quickack(sk); 601 icsk->icsk_ack.ato = TCP_ATO_MIN; 602 } else { 603 int m = now - icsk->icsk_ack.lrcvtime; 604 605 if (m <= TCP_ATO_MIN / 2) { 606 /* The fastest case is the first. */ 607 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2; 608 } else if (m < icsk->icsk_ack.ato) { 609 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m; 610 if (icsk->icsk_ack.ato > icsk->icsk_rto) 611 icsk->icsk_ack.ato = icsk->icsk_rto; 612 } else if (m > icsk->icsk_rto) { 613 /* Too long gap. Apparently sender failed to 614 * restart window, so that we send ACKs quickly. 615 */ 616 tcp_incr_quickack(sk); 617 sk_mem_reclaim(sk); 618 } 619 } 620 icsk->icsk_ack.lrcvtime = now; 621 622 TCP_ECN_check_ce(tp, skb); 623 624 if (skb->len >= 128) 625 tcp_grow_window(sk, skb); 626} 627 628/* Called to compute a smoothed rtt estimate. The data fed to this 629 * routine either comes from timestamps, or from segments that were 630 * known _not_ to have been retransmitted [see Karn/Partridge 631 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88 632 * piece by Van Jacobson. 633 * NOTE: the next three routines used to be one big routine. 634 * To save cycles in the RFC 1323 implementation it was better to break 635 * it up into three procedures. -- erics 636 */ 637static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt) 638{ 639 struct tcp_sock *tp = tcp_sk(sk); 640 long m = mrtt; /* RTT */ 641 642 /* The following amusing code comes from Jacobson's 643 * article in SIGCOMM '88. Note that rtt and mdev 644 * are scaled versions of rtt and mean deviation. 645 * This is designed to be as fast as possible 646 * m stands for "measurement". 647 * 648 * On a 1990 paper the rto value is changed to: 649 * RTO = rtt + 4 * mdev 650 * 651 * Funny. This algorithm seems to be very broken. 652 * These formulae increase RTO, when it should be decreased, increase 653 * too slowly, when it should be increased quickly, decrease too quickly 654 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely 655 * does not matter how to _calculate_ it. Seems, it was trap 656 * that VJ failed to avoid. 8) 657 */ 658 if (m == 0) 659 m = 1; 660 if (tp->srtt != 0) { 661 m -= (tp->srtt >> 3); /* m is now error in rtt est */ 662 tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */ 663 if (m < 0) { 664 m = -m; /* m is now abs(error) */ 665 m -= (tp->mdev >> 2); /* similar update on mdev */ 666 /* This is similar to one of Eifel findings. 667 * Eifel blocks mdev updates when rtt decreases. 668 * This solution is a bit different: we use finer gain 669 * for mdev in this case (alpha*beta). 670 * Like Eifel it also prevents growth of rto, 671 * but also it limits too fast rto decreases, 672 * happening in pure Eifel. 673 */ 674 if (m > 0) 675 m >>= 3; 676 } else { 677 m -= (tp->mdev >> 2); /* similar update on mdev */ 678 } 679 tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */ 680 if (tp->mdev > tp->mdev_max) { 681 tp->mdev_max = tp->mdev; 682 if (tp->mdev_max > tp->rttvar) 683 tp->rttvar = tp->mdev_max; 684 } 685 if (after(tp->snd_una, tp->rtt_seq)) { 686 if (tp->mdev_max < tp->rttvar) 687 tp->rttvar -= (tp->rttvar - tp->mdev_max) >> 2; 688 tp->rtt_seq = tp->snd_nxt; 689 tp->mdev_max = tcp_rto_min(sk); 690 } 691 } else { 692 /* no previous measure. */ 693 tp->srtt = m << 3; /* take the measured time to be rtt */ 694 tp->mdev = m << 1; /* make sure rto = 3*rtt */ 695 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk)); 696 tp->rtt_seq = tp->snd_nxt; 697 } 698} 699 700/* Calculate rto without backoff. This is the second half of Van Jacobson's 701 * routine referred to above. 702 */ 703static inline void tcp_set_rto(struct sock *sk) 704{ 705 const struct tcp_sock *tp = tcp_sk(sk); 706 /* Old crap is replaced with new one. 8) 707 * 708 * More seriously: 709 * 1. If rtt variance happened to be less 50msec, it is hallucination. 710 * It cannot be less due to utterly erratic ACK generation made 711 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_ 712 * to do with delayed acks, because at cwnd>2 true delack timeout 713 * is invisible. Actually, Linux-2.4 also generates erratic 714 * ACKs in some circumstances. 715 */ 716 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp); 717 718 /* 2. Fixups made earlier cannot be right. 719 * If we do not estimate RTO correctly without them, 720 * all the algo is pure shit and should be replaced 721 * with correct one. It is exactly, which we pretend to do. 722 */ 723 724 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo 725 * guarantees that rto is higher. 726 */ 727 tcp_bound_rto(sk); 728} 729 730/* Save metrics learned by this TCP session. 731 This function is called only, when TCP finishes successfully 732 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE. 733 */ 734void tcp_update_metrics(struct sock *sk) 735{ 736 struct tcp_sock *tp = tcp_sk(sk); 737 struct dst_entry *dst = __sk_dst_get(sk); 738 739 if (sysctl_tcp_nometrics_save) 740 return; 741 742 dst_confirm(dst); 743 744 if (dst && (dst->flags & DST_HOST)) { 745 const struct inet_connection_sock *icsk = inet_csk(sk); 746 int m; 747 unsigned long rtt; 748 749 if (icsk->icsk_backoff || !tp->srtt) { 750 /* This session failed to estimate rtt. Why? 751 * Probably, no packets returned in time. 752 * Reset our results. 753 */ 754 if (!(dst_metric_locked(dst, RTAX_RTT))) 755 dst_metric_set(dst, RTAX_RTT, 0); 756 return; 757 } 758 759 rtt = dst_metric_rtt(dst, RTAX_RTT); 760 m = rtt - tp->srtt; 761 762 /* If newly calculated rtt larger than stored one, 763 * store new one. Otherwise, use EWMA. Remember, 764 * rtt overestimation is always better than underestimation. 765 */ 766 if (!(dst_metric_locked(dst, RTAX_RTT))) { 767 if (m <= 0) 768 set_dst_metric_rtt(dst, RTAX_RTT, tp->srtt); 769 else 770 set_dst_metric_rtt(dst, RTAX_RTT, rtt - (m >> 3)); 771 } 772 773 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) { 774 unsigned long var; 775 if (m < 0) 776 m = -m; 777 778 /* Scale deviation to rttvar fixed point */ 779 m >>= 1; 780 if (m < tp->mdev) 781 m = tp->mdev; 782 783 var = dst_metric_rtt(dst, RTAX_RTTVAR); 784 if (m >= var) 785 var = m; 786 else 787 var -= (var - m) >> 2; 788 789 set_dst_metric_rtt(dst, RTAX_RTTVAR, var); 790 } 791 792 if (tcp_in_initial_slowstart(tp)) { 793 /* Slow start still did not finish. */ 794 if (dst_metric(dst, RTAX_SSTHRESH) && 795 !dst_metric_locked(dst, RTAX_SSTHRESH) && 796 (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH)) 797 dst_metric_set(dst, RTAX_SSTHRESH, tp->snd_cwnd >> 1); 798 if (!dst_metric_locked(dst, RTAX_CWND) && 799 tp->snd_cwnd > dst_metric(dst, RTAX_CWND)) 800 dst_metric_set(dst, RTAX_CWND, tp->snd_cwnd); 801 } else if (tp->snd_cwnd > tp->snd_ssthresh && 802 icsk->icsk_ca_state == TCP_CA_Open) { 803 /* Cong. avoidance phase, cwnd is reliable. */ 804 if (!dst_metric_locked(dst, RTAX_SSTHRESH)) 805 dst_metric_set(dst, RTAX_SSTHRESH, 806 max(tp->snd_cwnd >> 1, tp->snd_ssthresh)); 807 if (!dst_metric_locked(dst, RTAX_CWND)) 808 dst_metric_set(dst, RTAX_CWND, 809 (dst_metric(dst, RTAX_CWND) + 810 tp->snd_cwnd) >> 1); 811 } else { 812 /* Else slow start did not finish, cwnd is non-sense, 813 ssthresh may be also invalid. 814 */ 815 if (!dst_metric_locked(dst, RTAX_CWND)) 816 dst_metric_set(dst, RTAX_CWND, 817 (dst_metric(dst, RTAX_CWND) + 818 tp->snd_ssthresh) >> 1); 819 if (dst_metric(dst, RTAX_SSTHRESH) && 820 !dst_metric_locked(dst, RTAX_SSTHRESH) && 821 tp->snd_ssthresh > dst_metric(dst, RTAX_SSTHRESH)) 822 dst_metric_set(dst, RTAX_SSTHRESH, tp->snd_ssthresh); 823 } 824 825 if (!dst_metric_locked(dst, RTAX_REORDERING)) { 826 if (dst_metric(dst, RTAX_REORDERING) < tp->reordering && 827 tp->reordering != sysctl_tcp_reordering) 828 dst_metric_set(dst, RTAX_REORDERING, tp->reordering); 829 } 830 } 831} 832 833__u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst) 834{ 835 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0); 836 837 if (!cwnd) 838 cwnd = TCP_INIT_CWND; 839 return min_t(__u32, cwnd, tp->snd_cwnd_clamp); 840} 841 842/* Set slow start threshold and cwnd not falling to slow start */ 843void tcp_enter_cwr(struct sock *sk, const int set_ssthresh) 844{ 845 struct tcp_sock *tp = tcp_sk(sk); 846 const struct inet_connection_sock *icsk = inet_csk(sk); 847 848 tp->prior_ssthresh = 0; 849 tp->bytes_acked = 0; 850 if (icsk->icsk_ca_state < TCP_CA_CWR) { 851 tp->undo_marker = 0; 852 if (set_ssthresh) 853 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 854 tp->snd_cwnd = min(tp->snd_cwnd, 855 tcp_packets_in_flight(tp) + 1U); 856 tp->snd_cwnd_cnt = 0; 857 tp->high_seq = tp->snd_nxt; 858 tp->snd_cwnd_stamp = tcp_time_stamp; 859 TCP_ECN_queue_cwr(tp); 860 861 tcp_set_ca_state(sk, TCP_CA_CWR); 862 } 863} 864 865/* 866 * Packet counting of FACK is based on in-order assumptions, therefore TCP 867 * disables it when reordering is detected 868 */ 869static void tcp_disable_fack(struct tcp_sock *tp) 870{ 871 /* RFC3517 uses different metric in lost marker => reset on change */ 872 if (tcp_is_fack(tp)) 873 tp->lost_skb_hint = NULL; 874 tp->rx_opt.sack_ok &= ~TCP_FACK_ENABLED; 875} 876 877/* Take a notice that peer is sending D-SACKs */ 878static void tcp_dsack_seen(struct tcp_sock *tp) 879{ 880 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN; 881} 882 883/* Initialize metrics on socket. */ 884 885static void tcp_init_metrics(struct sock *sk) 886{ 887 struct tcp_sock *tp = tcp_sk(sk); 888 struct dst_entry *dst = __sk_dst_get(sk); 889 890 if (dst == NULL) 891 goto reset; 892 893 dst_confirm(dst); 894 895 if (dst_metric_locked(dst, RTAX_CWND)) 896 tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND); 897 if (dst_metric(dst, RTAX_SSTHRESH)) { 898 tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH); 899 if (tp->snd_ssthresh > tp->snd_cwnd_clamp) 900 tp->snd_ssthresh = tp->snd_cwnd_clamp; 901 } else { 902 /* ssthresh may have been reduced unnecessarily during. 903 * 3WHS. Restore it back to its initial default. 904 */ 905 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH; 906 } 907 if (dst_metric(dst, RTAX_REORDERING) && 908 tp->reordering != dst_metric(dst, RTAX_REORDERING)) { 909 tcp_disable_fack(tp); 910 tcp_disable_early_retrans(tp); 911 tp->reordering = dst_metric(dst, RTAX_REORDERING); 912 } 913 914 if (dst_metric(dst, RTAX_RTT) == 0 || tp->srtt == 0) 915 goto reset; 916 917 /* Initial rtt is determined from SYN,SYN-ACK. 918 * The segment is small and rtt may appear much 919 * less than real one. Use per-dst memory 920 * to make it more realistic. 921 * 922 * A bit of theory. RTT is time passed after "normal" sized packet 923 * is sent until it is ACKed. In normal circumstances sending small 924 * packets force peer to delay ACKs and calculation is correct too. 925 * The algorithm is adaptive and, provided we follow specs, it 926 * NEVER underestimate RTT. BUT! If peer tries to make some clever 927 * tricks sort of "quick acks" for time long enough to decrease RTT 928 * to low value, and then abruptly stops to do it and starts to delay 929 * ACKs, wait for troubles. 930 */ 931 if (dst_metric_rtt(dst, RTAX_RTT) > tp->srtt) { 932 tp->srtt = dst_metric_rtt(dst, RTAX_RTT); 933 tp->rtt_seq = tp->snd_nxt; 934 } 935 if (dst_metric_rtt(dst, RTAX_RTTVAR) > tp->mdev) { 936 tp->mdev = dst_metric_rtt(dst, RTAX_RTTVAR); 937 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk)); 938 } 939 tcp_set_rto(sk); 940reset: 941 if (tp->srtt == 0) { 942 /* RFC6298: 5.7 We've failed to get a valid RTT sample from 943 * 3WHS. This is most likely due to retransmission, 944 * including spurious one. Reset the RTO back to 3secs 945 * from the more aggressive 1sec to avoid more spurious 946 * retransmission. 947 */ 948 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_FALLBACK; 949 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_FALLBACK; 950 } 951 /* Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been 952 * retransmitted. In light of RFC6298 more aggressive 1sec 953 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK 954 * retransmission has occurred. 955 */ 956 if (tp->total_retrans > 1) 957 tp->snd_cwnd = 1; 958 else 959 tp->snd_cwnd = tcp_init_cwnd(tp, dst); 960 tp->snd_cwnd_stamp = tcp_time_stamp; 961} 962 963static void tcp_update_reordering(struct sock *sk, const int metric, 964 const int ts) 965{ 966 struct tcp_sock *tp = tcp_sk(sk); 967 if (metric > tp->reordering) { 968 int mib_idx; 969 970 tp->reordering = min(TCP_MAX_REORDERING, metric); 971 972 /* This exciting event is worth to be remembered. 8) */ 973 if (ts) 974 mib_idx = LINUX_MIB_TCPTSREORDER; 975 else if (tcp_is_reno(tp)) 976 mib_idx = LINUX_MIB_TCPRENOREORDER; 977 else if (tcp_is_fack(tp)) 978 mib_idx = LINUX_MIB_TCPFACKREORDER; 979 else 980 mib_idx = LINUX_MIB_TCPSACKREORDER; 981 982 NET_INC_STATS_BH(sock_net(sk), mib_idx); 983#if FASTRETRANS_DEBUG > 1 984 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n", 985 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state, 986 tp->reordering, 987 tp->fackets_out, 988 tp->sacked_out, 989 tp->undo_marker ? tp->undo_retrans : 0); 990#endif 991 tcp_disable_fack(tp); 992 } 993 994 if (metric > 0) 995 tcp_disable_early_retrans(tp); 996} 997 998/* This must be called before lost_out is incremented */ 999static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb) 1000{ 1001 if ((tp->retransmit_skb_hint == NULL) || 1002 before(TCP_SKB_CB(skb)->seq, 1003 TCP_SKB_CB(tp->retransmit_skb_hint)->seq)) 1004 tp->retransmit_skb_hint = skb; 1005 1006 if (!tp->lost_out || 1007 after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high)) 1008 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq; 1009} 1010 1011static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb) 1012{ 1013 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) { 1014 tcp_verify_retransmit_hint(tp, skb); 1015 1016 tp->lost_out += tcp_skb_pcount(skb); 1017 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 1018 } 1019} 1020 1021static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp, 1022 struct sk_buff *skb) 1023{ 1024 tcp_verify_retransmit_hint(tp, skb); 1025 1026 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) { 1027 tp->lost_out += tcp_skb_pcount(skb); 1028 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 1029 } 1030} 1031 1032/* This procedure tags the retransmission queue when SACKs arrive. 1033 * 1034 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L). 1035 * Packets in queue with these bits set are counted in variables 1036 * sacked_out, retrans_out and lost_out, correspondingly. 1037 * 1038 * Valid combinations are: 1039 * Tag InFlight Description 1040 * 0 1 - orig segment is in flight. 1041 * S 0 - nothing flies, orig reached receiver. 1042 * L 0 - nothing flies, orig lost by net. 1043 * R 2 - both orig and retransmit are in flight. 1044 * L|R 1 - orig is lost, retransmit is in flight. 1045 * S|R 1 - orig reached receiver, retrans is still in flight. 1046 * (L|S|R is logically valid, it could occur when L|R is sacked, 1047 * but it is equivalent to plain S and code short-curcuits it to S. 1048 * L|S is logically invalid, it would mean -1 packet in flight 8)) 1049 * 1050 * These 6 states form finite state machine, controlled by the following events: 1051 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue()) 1052 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue()) 1053 * 3. Loss detection event of two flavors: 1054 * A. Scoreboard estimator decided the packet is lost. 1055 * A'. Reno "three dupacks" marks head of queue lost. 1056 * A''. Its FACK modification, head until snd.fack is lost. 1057 * B. SACK arrives sacking SND.NXT at the moment, when the 1058 * segment was retransmitted. 1059 * 4. D-SACK added new rule: D-SACK changes any tag to S. 1060 * 1061 * It is pleasant to note, that state diagram turns out to be commutative, 1062 * so that we are allowed not to be bothered by order of our actions, 1063 * when multiple events arrive simultaneously. (see the function below). 1064 * 1065 * Reordering detection. 1066 * -------------------- 1067 * Reordering metric is maximal distance, which a packet can be displaced 1068 * in packet stream. With SACKs we can estimate it: 1069 * 1070 * 1. SACK fills old hole and the corresponding segment was not 1071 * ever retransmitted -> reordering. Alas, we cannot use it 1072 * when segment was retransmitted. 1073 * 2. The last flaw is solved with D-SACK. D-SACK arrives 1074 * for retransmitted and already SACKed segment -> reordering.. 1075 * Both of these heuristics are not used in Loss state, when we cannot 1076 * account for retransmits accurately. 1077 * 1078 * SACK block validation. 1079 * ---------------------- 1080 * 1081 * SACK block range validation checks that the received SACK block fits to 1082 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT. 1083 * Note that SND.UNA is not included to the range though being valid because 1084 * it means that the receiver is rather inconsistent with itself reporting 1085 * SACK reneging when it should advance SND.UNA. Such SACK block this is 1086 * perfectly valid, however, in light of RFC2018 which explicitly states 1087 * that "SACK block MUST reflect the newest segment. Even if the newest 1088 * segment is going to be discarded ...", not that it looks very clever 1089 * in case of head skb. Due to potentional receiver driven attacks, we 1090 * choose to avoid immediate execution of a walk in write queue due to 1091 * reneging and defer head skb's loss recovery to standard loss recovery 1092 * procedure that will eventually trigger (nothing forbids us doing this). 1093 * 1094 * Implements also blockage to start_seq wrap-around. Problem lies in the 1095 * fact that though start_seq (s) is before end_seq (i.e., not reversed), 1096 * there's no guarantee that it will be before snd_nxt (n). The problem 1097 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt 1098 * wrap (s_w): 1099 * 1100 * <- outs wnd -> <- wrapzone -> 1101 * u e n u_w e_w s n_w 1102 * | | | | | | | 1103 * |<------------+------+----- TCP seqno space --------------+---------->| 1104 * ...-- <2^31 ->| |<--------... 1105 * ...---- >2^31 ------>| |<--------... 1106 * 1107 * Current code wouldn't be vulnerable but it's better still to discard such 1108 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat 1109 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in 1110 * snd_nxt wrap -> snd_una region will then become "well defined", i.e., 1111 * equal to the ideal case (infinite seqno space without wrap caused issues). 1112 * 1113 * With D-SACK the lower bound is extended to cover sequence space below 1114 * SND.UNA down to undo_marker, which is the last point of interest. Yet 1115 * again, D-SACK block must not to go across snd_una (for the same reason as 1116 * for the normal SACK blocks, explained above). But there all simplicity 1117 * ends, TCP might receive valid D-SACKs below that. As long as they reside 1118 * fully below undo_marker they do not affect behavior in anyway and can 1119 * therefore be safely ignored. In rare cases (which are more or less 1120 * theoretical ones), the D-SACK will nicely cross that boundary due to skb 1121 * fragmentation and packet reordering past skb's retransmission. To consider 1122 * them correctly, the acceptable range must be extended even more though 1123 * the exact amount is rather hard to quantify. However, tp->max_window can 1124 * be used as an exaggerated estimate. 1125 */ 1126static int tcp_is_sackblock_valid(struct tcp_sock *tp, int is_dsack, 1127 u32 start_seq, u32 end_seq) 1128{ 1129 /* Too far in future, or reversed (interpretation is ambiguous) */ 1130 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq)) 1131 return 0; 1132 1133 /* Nasty start_seq wrap-around check (see comments above) */ 1134 if (!before(start_seq, tp->snd_nxt)) 1135 return 0; 1136 1137 /* In outstanding window? ...This is valid exit for D-SACKs too. 1138 * start_seq == snd_una is non-sensical (see comments above) 1139 */ 1140 if (after(start_seq, tp->snd_una)) 1141 return 1; 1142 1143 if (!is_dsack || !tp->undo_marker) 1144 return 0; 1145 1146 /* ...Then it's D-SACK, and must reside below snd_una completely */ 1147 if (after(end_seq, tp->snd_una)) 1148 return 0; 1149 1150 if (!before(start_seq, tp->undo_marker)) 1151 return 1; 1152 1153 /* Too old */ 1154 if (!after(end_seq, tp->undo_marker)) 1155 return 0; 1156 1157 /* Undo_marker boundary crossing (overestimates a lot). Known already: 1158 * start_seq < undo_marker and end_seq >= undo_marker. 1159 */ 1160 return !before(start_seq, end_seq - tp->max_window); 1161} 1162 1163/* Check for lost retransmit. This superb idea is borrowed from "ratehalving". 1164 * Event "B". Later note: FACK people cheated me again 8), we have to account 1165 * for reordering! Ugly, but should help. 1166 * 1167 * Search retransmitted skbs from write_queue that were sent when snd_nxt was 1168 * less than what is now known to be received by the other end (derived from 1169 * highest SACK block). Also calculate the lowest snd_nxt among the remaining 1170 * retransmitted skbs to avoid some costly processing per ACKs. 1171 */ 1172static void tcp_mark_lost_retrans(struct sock *sk) 1173{ 1174 const struct inet_connection_sock *icsk = inet_csk(sk); 1175 struct tcp_sock *tp = tcp_sk(sk); 1176 struct sk_buff *skb; 1177 int cnt = 0; 1178 u32 new_low_seq = tp->snd_nxt; 1179 u32 received_upto = tcp_highest_sack_seq(tp); 1180 1181 if (!tcp_is_fack(tp) || !tp->retrans_out || 1182 !after(received_upto, tp->lost_retrans_low) || 1183 icsk->icsk_ca_state != TCP_CA_Recovery) 1184 return; 1185 1186 tcp_for_write_queue(skb, sk) { 1187 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq; 1188 1189 if (skb == tcp_send_head(sk)) 1190 break; 1191 if (cnt == tp->retrans_out) 1192 break; 1193 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 1194 continue; 1195 1196 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)) 1197 continue; 1198 1199 /* TODO: We would like to get rid of tcp_is_fack(tp) only 1200 * constraint here (see above) but figuring out that at 1201 * least tp->reordering SACK blocks reside between ack_seq 1202 * and received_upto is not easy task to do cheaply with 1203 * the available datastructures. 1204 * 1205 * Whether FACK should check here for tp->reordering segs 1206 * in-between one could argue for either way (it would be 1207 * rather simple to implement as we could count fack_count 1208 * during the walk and do tp->fackets_out - fack_count). 1209 */ 1210 if (after(received_upto, ack_seq)) { 1211 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 1212 tp->retrans_out -= tcp_skb_pcount(skb); 1213 1214 tcp_skb_mark_lost_uncond_verify(tp, skb); 1215 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT); 1216 } else { 1217 if (before(ack_seq, new_low_seq)) 1218 new_low_seq = ack_seq; 1219 cnt += tcp_skb_pcount(skb); 1220 } 1221 } 1222 1223 if (tp->retrans_out) 1224 tp->lost_retrans_low = new_low_seq; 1225} 1226 1227static int tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb, 1228 struct tcp_sack_block_wire *sp, int num_sacks, 1229 u32 prior_snd_una) 1230{ 1231 struct tcp_sock *tp = tcp_sk(sk); 1232 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq); 1233 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq); 1234 int dup_sack = 0; 1235 1236 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) { 1237 dup_sack = 1; 1238 tcp_dsack_seen(tp); 1239 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV); 1240 } else if (num_sacks > 1) { 1241 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq); 1242 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq); 1243 1244 if (!after(end_seq_0, end_seq_1) && 1245 !before(start_seq_0, start_seq_1)) { 1246 dup_sack = 1; 1247 tcp_dsack_seen(tp); 1248 NET_INC_STATS_BH(sock_net(sk), 1249 LINUX_MIB_TCPDSACKOFORECV); 1250 } 1251 } 1252 1253 /* D-SACK for already forgotten data... Do dumb counting. */ 1254 if (dup_sack && tp->undo_marker && tp->undo_retrans && 1255 !after(end_seq_0, prior_snd_una) && 1256 after(end_seq_0, tp->undo_marker)) 1257 tp->undo_retrans--; 1258 1259 return dup_sack; 1260} 1261 1262struct tcp_sacktag_state { 1263 int reord; 1264 int fack_count; 1265 int flag; 1266}; 1267 1268/* Check if skb is fully within the SACK block. In presence of GSO skbs, 1269 * the incoming SACK may not exactly match but we can find smaller MSS 1270 * aligned portion of it that matches. Therefore we might need to fragment 1271 * which may fail and creates some hassle (caller must handle error case 1272 * returns). 1273 * 1274 * FIXME: this could be merged to shift decision code 1275 */ 1276static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb, 1277 u32 start_seq, u32 end_seq) 1278{ 1279 int in_sack, err; 1280 unsigned int pkt_len; 1281 unsigned int mss; 1282 1283 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && 1284 !before(end_seq, TCP_SKB_CB(skb)->end_seq); 1285 1286 if (tcp_skb_pcount(skb) > 1 && !in_sack && 1287 after(TCP_SKB_CB(skb)->end_seq, start_seq)) { 1288 mss = tcp_skb_mss(skb); 1289 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); 1290 1291 if (!in_sack) { 1292 pkt_len = start_seq - TCP_SKB_CB(skb)->seq; 1293 if (pkt_len < mss) 1294 pkt_len = mss; 1295 } else { 1296 pkt_len = end_seq - TCP_SKB_CB(skb)->seq; 1297 if (pkt_len < mss) 1298 return -EINVAL; 1299 } 1300 1301 /* Round if necessary so that SACKs cover only full MSSes 1302 * and/or the remaining small portion (if present) 1303 */ 1304 if (pkt_len > mss) { 1305 unsigned int new_len = (pkt_len / mss) * mss; 1306 if (!in_sack && new_len < pkt_len) { 1307 new_len += mss; 1308 if (new_len > skb->len) 1309 return 0; 1310 } 1311 pkt_len = new_len; 1312 } 1313 err = tcp_fragment(sk, skb, pkt_len, mss); 1314 if (err < 0) 1315 return err; 1316 } 1317 1318 return in_sack; 1319} 1320 1321/* Mark the given newly-SACKed range as such, adjusting counters and hints. */ 1322static u8 tcp_sacktag_one(struct sock *sk, 1323 struct tcp_sacktag_state *state, u8 sacked, 1324 u32 start_seq, u32 end_seq, 1325 int dup_sack, int pcount) 1326{ 1327 struct tcp_sock *tp = tcp_sk(sk); 1328 int fack_count = state->fack_count; 1329 1330 /* Account D-SACK for retransmitted packet. */ 1331 if (dup_sack && (sacked & TCPCB_RETRANS)) { 1332 if (tp->undo_marker && tp->undo_retrans && 1333 after(end_seq, tp->undo_marker)) 1334 tp->undo_retrans--; 1335 if (sacked & TCPCB_SACKED_ACKED) 1336 state->reord = min(fack_count, state->reord); 1337 } 1338 1339 /* Nothing to do; acked frame is about to be dropped (was ACKed). */ 1340 if (!after(end_seq, tp->snd_una)) 1341 return sacked; 1342 1343 if (!(sacked & TCPCB_SACKED_ACKED)) { 1344 if (sacked & TCPCB_SACKED_RETRANS) { 1345 /* If the segment is not tagged as lost, 1346 * we do not clear RETRANS, believing 1347 * that retransmission is still in flight. 1348 */ 1349 if (sacked & TCPCB_LOST) { 1350 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS); 1351 tp->lost_out -= pcount; 1352 tp->retrans_out -= pcount; 1353 } 1354 } else { 1355 if (!(sacked & TCPCB_RETRANS)) { 1356 /* New sack for not retransmitted frame, 1357 * which was in hole. It is reordering. 1358 */ 1359 if (before(start_seq, 1360 tcp_highest_sack_seq(tp))) 1361 state->reord = min(fack_count, 1362 state->reord); 1363 1364 /* SACK enhanced F-RTO (RFC4138; Appendix B) */ 1365 if (!after(end_seq, tp->frto_highmark)) 1366 state->flag |= FLAG_ONLY_ORIG_SACKED; 1367 } 1368 1369 if (sacked & TCPCB_LOST) { 1370 sacked &= ~TCPCB_LOST; 1371 tp->lost_out -= pcount; 1372 } 1373 } 1374 1375 sacked |= TCPCB_SACKED_ACKED; 1376 state->flag |= FLAG_DATA_SACKED; 1377 tp->sacked_out += pcount; 1378 1379 fack_count += pcount; 1380 1381 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */ 1382 if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) && 1383 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq)) 1384 tp->lost_cnt_hint += pcount; 1385 1386 if (fack_count > tp->fackets_out) 1387 tp->fackets_out = fack_count; 1388 } 1389 1390 /* D-SACK. We can detect redundant retransmission in S|R and plain R 1391 * frames and clear it. undo_retrans is decreased above, L|R frames 1392 * are accounted above as well. 1393 */ 1394 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) { 1395 sacked &= ~TCPCB_SACKED_RETRANS; 1396 tp->retrans_out -= pcount; 1397 } 1398 1399 return sacked; 1400} 1401 1402/* Shift newly-SACKed bytes from this skb to the immediately previous 1403 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such. 1404 */ 1405static int tcp_shifted_skb(struct sock *sk, struct sk_buff *skb, 1406 struct tcp_sacktag_state *state, 1407 unsigned int pcount, int shifted, int mss, 1408 int dup_sack) 1409{ 1410 struct tcp_sock *tp = tcp_sk(sk); 1411 struct sk_buff *prev = tcp_write_queue_prev(sk, skb); 1412 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */ 1413 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */ 1414 1415 BUG_ON(!pcount); 1416 1417 /* Adjust counters and hints for the newly sacked sequence 1418 * range but discard the return value since prev is already 1419 * marked. We must tag the range first because the seq 1420 * advancement below implicitly advances 1421 * tcp_highest_sack_seq() when skb is highest_sack. 1422 */ 1423 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked, 1424 start_seq, end_seq, dup_sack, pcount); 1425 1426 if (skb == tp->lost_skb_hint) 1427 tp->lost_cnt_hint += pcount; 1428 1429 TCP_SKB_CB(prev)->end_seq += shifted; 1430 TCP_SKB_CB(skb)->seq += shifted; 1431 1432 skb_shinfo(prev)->gso_segs += pcount; 1433 BUG_ON(skb_shinfo(skb)->gso_segs < pcount); 1434 skb_shinfo(skb)->gso_segs -= pcount; 1435 1436 /* When we're adding to gso_segs == 1, gso_size will be zero, 1437 * in theory this shouldn't be necessary but as long as DSACK 1438 * code can come after this skb later on it's better to keep 1439 * setting gso_size to something. 1440 */ 1441 if (!skb_shinfo(prev)->gso_size) { 1442 skb_shinfo(prev)->gso_size = mss; 1443 skb_shinfo(prev)->gso_type = sk->sk_gso_type; 1444 } 1445 1446 /* CHECKME: To clear or not to clear? Mimics normal skb currently */ 1447 if (skb_shinfo(skb)->gso_segs <= 1) { 1448 skb_shinfo(skb)->gso_size = 0; 1449 skb_shinfo(skb)->gso_type = 0; 1450 } 1451 1452 /* Difference in this won't matter, both ACKed by the same cumul. ACK */ 1453 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS); 1454 1455 if (skb->len > 0) { 1456 BUG_ON(!tcp_skb_pcount(skb)); 1457 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED); 1458 return 0; 1459 } 1460 1461 /* Whole SKB was eaten :-) */ 1462 1463 if (skb == tp->retransmit_skb_hint) 1464 tp->retransmit_skb_hint = prev; 1465 if (skb == tp->scoreboard_skb_hint) 1466 tp->scoreboard_skb_hint = prev; 1467 if (skb == tp->lost_skb_hint) { 1468 tp->lost_skb_hint = prev; 1469 tp->lost_cnt_hint -= tcp_skb_pcount(prev); 1470 } 1471 1472 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(prev)->tcp_flags; 1473 if (skb == tcp_highest_sack(sk)) 1474 tcp_advance_highest_sack(sk, skb); 1475 1476 tcp_unlink_write_queue(skb, sk); 1477 sk_wmem_free_skb(sk, skb); 1478 1479 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED); 1480 1481 return 1; 1482} 1483 1484/* I wish gso_size would have a bit more sane initialization than 1485 * something-or-zero which complicates things 1486 */ 1487static int tcp_skb_seglen(const struct sk_buff *skb) 1488{ 1489 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb); 1490} 1491 1492/* Shifting pages past head area doesn't work */ 1493static int skb_can_shift(const struct sk_buff *skb) 1494{ 1495 return !skb_headlen(skb) && skb_is_nonlinear(skb); 1496} 1497 1498/* Try collapsing SACK blocks spanning across multiple skbs to a single 1499 * skb. 1500 */ 1501static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb, 1502 struct tcp_sacktag_state *state, 1503 u32 start_seq, u32 end_seq, 1504 int dup_sack) 1505{ 1506 struct tcp_sock *tp = tcp_sk(sk); 1507 struct sk_buff *prev; 1508 int mss; 1509 int pcount = 0; 1510 int len; 1511 int in_sack; 1512 1513 if (!sk_can_gso(sk)) 1514 goto fallback; 1515 1516 /* Normally R but no L won't result in plain S */ 1517 if (!dup_sack && 1518 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS) 1519 goto fallback; 1520 if (!skb_can_shift(skb)) 1521 goto fallback; 1522 /* This frame is about to be dropped (was ACKed). */ 1523 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 1524 goto fallback; 1525 1526 /* Can only happen with delayed DSACK + discard craziness */ 1527 if (unlikely(skb == tcp_write_queue_head(sk))) 1528 goto fallback; 1529 prev = tcp_write_queue_prev(sk, skb); 1530 1531 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) 1532 goto fallback; 1533 1534 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && 1535 !before(end_seq, TCP_SKB_CB(skb)->end_seq); 1536 1537 if (in_sack) { 1538 len = skb->len; 1539 pcount = tcp_skb_pcount(skb); 1540 mss = tcp_skb_seglen(skb); 1541 1542 /* TODO: Fix DSACKs to not fragment already SACKed and we can 1543 * drop this restriction as unnecessary 1544 */ 1545 if (mss != tcp_skb_seglen(prev)) 1546 goto fallback; 1547 } else { 1548 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq)) 1549 goto noop; 1550 /* CHECKME: This is non-MSS split case only?, this will 1551 * cause skipped skbs due to advancing loop btw, original 1552 * has that feature too 1553 */ 1554 if (tcp_skb_pcount(skb) <= 1) 1555 goto noop; 1556 1557 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); 1558 if (!in_sack) { 1559 /* TODO: head merge to next could be attempted here 1560 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)), 1561 * though it might not be worth of the additional hassle 1562 * 1563 * ...we can probably just fallback to what was done 1564 * previously. We could try merging non-SACKed ones 1565 * as well but it probably isn't going to buy off 1566 * because later SACKs might again split them, and 1567 * it would make skb timestamp tracking considerably 1568 * harder problem. 1569 */ 1570 goto fallback; 1571 } 1572 1573 len = end_seq - TCP_SKB_CB(skb)->seq; 1574 BUG_ON(len < 0); 1575 BUG_ON(len > skb->len); 1576 1577 /* MSS boundaries should be honoured or else pcount will 1578 * severely break even though it makes things bit trickier. 1579 * Optimize common case to avoid most of the divides 1580 */ 1581 mss = tcp_skb_mss(skb); 1582 1583 /* TODO: Fix DSACKs to not fragment already SACKed and we can 1584 * drop this restriction as unnecessary 1585 */ 1586 if (mss != tcp_skb_seglen(prev)) 1587 goto fallback; 1588 1589 if (len == mss) { 1590 pcount = 1; 1591 } else if (len < mss) { 1592 goto noop; 1593 } else { 1594 pcount = len / mss; 1595 len = pcount * mss; 1596 } 1597 } 1598 1599 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */ 1600 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una)) 1601 goto fallback; 1602 1603 if (!skb_shift(prev, skb, len)) 1604 goto fallback; 1605 if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack)) 1606 goto out; 1607 1608 /* Hole filled allows collapsing with the next as well, this is very 1609 * useful when hole on every nth skb pattern happens 1610 */ 1611 if (prev == tcp_write_queue_tail(sk)) 1612 goto out; 1613 skb = tcp_write_queue_next(sk, prev); 1614 1615 if (!skb_can_shift(skb) || 1616 (skb == tcp_send_head(sk)) || 1617 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) || 1618 (mss != tcp_skb_seglen(skb))) 1619 goto out; 1620 1621 len = skb->len; 1622 if (skb_shift(prev, skb, len)) { 1623 pcount += tcp_skb_pcount(skb); 1624 tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0); 1625 } 1626 1627out: 1628 state->fack_count += pcount; 1629 return prev; 1630 1631noop: 1632 return skb; 1633 1634fallback: 1635 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK); 1636 return NULL; 1637} 1638 1639static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk, 1640 struct tcp_sack_block *next_dup, 1641 struct tcp_sacktag_state *state, 1642 u32 start_seq, u32 end_seq, 1643 int dup_sack_in) 1644{ 1645 struct tcp_sock *tp = tcp_sk(sk); 1646 struct sk_buff *tmp; 1647 1648 tcp_for_write_queue_from(skb, sk) { 1649 int in_sack = 0; 1650 int dup_sack = dup_sack_in; 1651 1652 if (skb == tcp_send_head(sk)) 1653 break; 1654 1655 /* queue is in-order => we can short-circuit the walk early */ 1656 if (!before(TCP_SKB_CB(skb)->seq, end_seq)) 1657 break; 1658 1659 if ((next_dup != NULL) && 1660 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) { 1661 in_sack = tcp_match_skb_to_sack(sk, skb, 1662 next_dup->start_seq, 1663 next_dup->end_seq); 1664 if (in_sack > 0) 1665 dup_sack = 1; 1666 } 1667 1668 /* skb reference here is a bit tricky to get right, since 1669 * shifting can eat and free both this skb and the next, 1670 * so not even _safe variant of the loop is enough. 1671 */ 1672 if (in_sack <= 0) { 1673 tmp = tcp_shift_skb_data(sk, skb, state, 1674 start_seq, end_seq, dup_sack); 1675 if (tmp != NULL) { 1676 if (tmp != skb) { 1677 skb = tmp; 1678 continue; 1679 } 1680 1681 in_sack = 0; 1682 } else { 1683 in_sack = tcp_match_skb_to_sack(sk, skb, 1684 start_seq, 1685 end_seq); 1686 } 1687 } 1688 1689 if (unlikely(in_sack < 0)) 1690 break; 1691 1692 if (in_sack) { 1693 TCP_SKB_CB(skb)->sacked = 1694 tcp_sacktag_one(sk, 1695 state, 1696 TCP_SKB_CB(skb)->sacked, 1697 TCP_SKB_CB(skb)->seq, 1698 TCP_SKB_CB(skb)->end_seq, 1699 dup_sack, 1700 tcp_skb_pcount(skb)); 1701 1702 if (!before(TCP_SKB_CB(skb)->seq, 1703 tcp_highest_sack_seq(tp))) 1704 tcp_advance_highest_sack(sk, skb); 1705 } 1706 1707 state->fack_count += tcp_skb_pcount(skb); 1708 } 1709 return skb; 1710} 1711 1712/* Avoid all extra work that is being done by sacktag while walking in 1713 * a normal way 1714 */ 1715static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk, 1716 struct tcp_sacktag_state *state, 1717 u32 skip_to_seq) 1718{ 1719 tcp_for_write_queue_from(skb, sk) { 1720 if (skb == tcp_send_head(sk)) 1721 break; 1722 1723 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq)) 1724 break; 1725 1726 state->fack_count += tcp_skb_pcount(skb); 1727 } 1728 return skb; 1729} 1730 1731static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb, 1732 struct sock *sk, 1733 struct tcp_sack_block *next_dup, 1734 struct tcp_sacktag_state *state, 1735 u32 skip_to_seq) 1736{ 1737 if (next_dup == NULL) 1738 return skb; 1739 1740 if (before(next_dup->start_seq, skip_to_seq)) { 1741 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq); 1742 skb = tcp_sacktag_walk(skb, sk, NULL, state, 1743 next_dup->start_seq, next_dup->end_seq, 1744 1); 1745 } 1746 1747 return skb; 1748} 1749 1750static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache) 1751{ 1752 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache); 1753} 1754 1755static int 1756tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb, 1757 u32 prior_snd_una) 1758{ 1759 const struct inet_connection_sock *icsk = inet_csk(sk); 1760 struct tcp_sock *tp = tcp_sk(sk); 1761 const unsigned char *ptr = (skb_transport_header(ack_skb) + 1762 TCP_SKB_CB(ack_skb)->sacked); 1763 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2); 1764 struct tcp_sack_block sp[TCP_NUM_SACKS]; 1765 struct tcp_sack_block *cache; 1766 struct tcp_sacktag_state state; 1767 struct sk_buff *skb; 1768 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3); 1769 int used_sacks; 1770 int found_dup_sack = 0; 1771 int i, j; 1772 int first_sack_index; 1773 1774 state.flag = 0; 1775 state.reord = tp->packets_out; 1776 1777 if (!tp->sacked_out) { 1778 if (WARN_ON(tp->fackets_out)) 1779 tp->fackets_out = 0; 1780 tcp_highest_sack_reset(sk); 1781 } 1782 1783 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire, 1784 num_sacks, prior_snd_una); 1785 if (found_dup_sack) 1786 state.flag |= FLAG_DSACKING_ACK; 1787 1788 /* Eliminate too old ACKs, but take into 1789 * account more or less fresh ones, they can 1790 * contain valid SACK info. 1791 */ 1792 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window)) 1793 return 0; 1794 1795 if (!tp->packets_out) 1796 goto out; 1797 1798 used_sacks = 0; 1799 first_sack_index = 0; 1800 for (i = 0; i < num_sacks; i++) { 1801 int dup_sack = !i && found_dup_sack; 1802 1803 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq); 1804 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq); 1805 1806 if (!tcp_is_sackblock_valid(tp, dup_sack, 1807 sp[used_sacks].start_seq, 1808 sp[used_sacks].end_seq)) { 1809 int mib_idx; 1810 1811 if (dup_sack) { 1812 if (!tp->undo_marker) 1813 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO; 1814 else 1815 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD; 1816 } else { 1817 /* Don't count olds caused by ACK reordering */ 1818 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) && 1819 !after(sp[used_sacks].end_seq, tp->snd_una)) 1820 continue; 1821 mib_idx = LINUX_MIB_TCPSACKDISCARD; 1822 } 1823 1824 NET_INC_STATS_BH(sock_net(sk), mib_idx); 1825 if (i == 0) 1826 first_sack_index = -1; 1827 continue; 1828 } 1829 1830 /* Ignore very old stuff early */ 1831 if (!after(sp[used_sacks].end_seq, prior_snd_una)) 1832 continue; 1833 1834 used_sacks++; 1835 } 1836 1837 /* order SACK blocks to allow in order walk of the retrans queue */ 1838 for (i = used_sacks - 1; i > 0; i--) { 1839 for (j = 0; j < i; j++) { 1840 if (after(sp[j].start_seq, sp[j + 1].start_seq)) { 1841 swap(sp[j], sp[j + 1]); 1842 1843 /* Track where the first SACK block goes to */ 1844 if (j == first_sack_index) 1845 first_sack_index = j + 1; 1846 } 1847 } 1848 } 1849 1850 skb = tcp_write_queue_head(sk); 1851 state.fack_count = 0; 1852 i = 0; 1853 1854 if (!tp->sacked_out) { 1855 /* It's already past, so skip checking against it */ 1856 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache); 1857 } else { 1858 cache = tp->recv_sack_cache; 1859 /* Skip empty blocks in at head of the cache */ 1860 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq && 1861 !cache->end_seq) 1862 cache++; 1863 } 1864 1865 while (i < used_sacks) { 1866 u32 start_seq = sp[i].start_seq; 1867 u32 end_seq = sp[i].end_seq; 1868 int dup_sack = (found_dup_sack && (i == first_sack_index)); 1869 struct tcp_sack_block *next_dup = NULL; 1870 1871 if (found_dup_sack && ((i + 1) == first_sack_index)) 1872 next_dup = &sp[i + 1]; 1873 1874 /* Skip too early cached blocks */ 1875 while (tcp_sack_cache_ok(tp, cache) && 1876 !before(start_seq, cache->end_seq)) 1877 cache++; 1878 1879 /* Can skip some work by looking recv_sack_cache? */ 1880 if (tcp_sack_cache_ok(tp, cache) && !dup_sack && 1881 after(end_seq, cache->start_seq)) { 1882 1883 /* Head todo? */ 1884 if (before(start_seq, cache->start_seq)) { 1885 skb = tcp_sacktag_skip(skb, sk, &state, 1886 start_seq); 1887 skb = tcp_sacktag_walk(skb, sk, next_dup, 1888 &state, 1889 start_seq, 1890 cache->start_seq, 1891 dup_sack); 1892 } 1893 1894 /* Rest of the block already fully processed? */ 1895 if (!after(end_seq, cache->end_seq)) 1896 goto advance_sp; 1897 1898 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup, 1899 &state, 1900 cache->end_seq); 1901 1902 /* ...tail remains todo... */ 1903 if (tcp_highest_sack_seq(tp) == cache->end_seq) { 1904 /* ...but better entrypoint exists! */ 1905 skb = tcp_highest_sack(sk); 1906 if (skb == NULL) 1907 break; 1908 state.fack_count = tp->fackets_out; 1909 cache++; 1910 goto walk; 1911 } 1912 1913 skb = tcp_sacktag_skip(skb, sk, &state, cache->end_seq); 1914 /* Check overlap against next cached too (past this one already) */ 1915 cache++; 1916 continue; 1917 } 1918 1919 if (!before(start_seq, tcp_highest_sack_seq(tp))) { 1920 skb = tcp_highest_sack(sk); 1921 if (skb == NULL) 1922 break; 1923 state.fack_count = tp->fackets_out; 1924 } 1925 skb = tcp_sacktag_skip(skb, sk, &state, start_seq); 1926 1927walk: 1928 skb = tcp_sacktag_walk(skb, sk, next_dup, &state, 1929 start_seq, end_seq, dup_sack); 1930 1931advance_sp: 1932 /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct 1933 * due to in-order walk 1934 */ 1935 if (after(end_seq, tp->frto_highmark)) 1936 state.flag &= ~FLAG_ONLY_ORIG_SACKED; 1937 1938 i++; 1939 } 1940 1941 /* Clear the head of the cache sack blocks so we can skip it next time */ 1942 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) { 1943 tp->recv_sack_cache[i].start_seq = 0; 1944 tp->recv_sack_cache[i].end_seq = 0; 1945 } 1946 for (j = 0; j < used_sacks; j++) 1947 tp->recv_sack_cache[i++] = sp[j]; 1948 1949 tcp_mark_lost_retrans(sk); 1950 1951 tcp_verify_left_out(tp); 1952 1953 if ((state.reord < tp->fackets_out) && 1954 ((icsk->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker) && 1955 (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark))) 1956 tcp_update_reordering(sk, tp->fackets_out - state.reord, 0); 1957 1958out: 1959 1960#if FASTRETRANS_DEBUG > 0 1961 WARN_ON((int)tp->sacked_out < 0); 1962 WARN_ON((int)tp->lost_out < 0); 1963 WARN_ON((int)tp->retrans_out < 0); 1964 WARN_ON((int)tcp_packets_in_flight(tp) < 0); 1965#endif 1966 return state.flag; 1967} 1968 1969/* Limits sacked_out so that sum with lost_out isn't ever larger than 1970 * packets_out. Returns zero if sacked_out adjustement wasn't necessary. 1971 */ 1972static int tcp_limit_reno_sacked(struct tcp_sock *tp) 1973{ 1974 u32 holes; 1975 1976 holes = max(tp->lost_out, 1U); 1977 holes = min(holes, tp->packets_out); 1978 1979 if ((tp->sacked_out + holes) > tp->packets_out) { 1980 tp->sacked_out = tp->packets_out - holes; 1981 return 1; 1982 } 1983 return 0; 1984} 1985 1986/* If we receive more dupacks than we expected counting segments 1987 * in assumption of absent reordering, interpret this as reordering. 1988 * The only another reason could be bug in receiver TCP. 1989 */ 1990static void tcp_check_reno_reordering(struct sock *sk, const int addend) 1991{ 1992 struct tcp_sock *tp = tcp_sk(sk); 1993 if (tcp_limit_reno_sacked(tp)) 1994 tcp_update_reordering(sk, tp->packets_out + addend, 0); 1995} 1996 1997/* Emulate SACKs for SACKless connection: account for a new dupack. */ 1998 1999static void tcp_add_reno_sack(struct sock *sk) 2000{ 2001 struct tcp_sock *tp = tcp_sk(sk); 2002 tp->sacked_out++; 2003 tcp_check_reno_reordering(sk, 0); 2004 tcp_verify_left_out(tp); 2005} 2006 2007/* Account for ACK, ACKing some data in Reno Recovery phase. */ 2008 2009static void tcp_remove_reno_sacks(struct sock *sk, int acked) 2010{ 2011 struct tcp_sock *tp = tcp_sk(sk); 2012 2013 if (acked > 0) { 2014 /* One ACK acked hole. The rest eat duplicate ACKs. */ 2015 if (acked - 1 >= tp->sacked_out) 2016 tp->sacked_out = 0; 2017 else 2018 tp->sacked_out -= acked - 1; 2019 } 2020 tcp_check_reno_reordering(sk, acked); 2021 tcp_verify_left_out(tp); 2022} 2023 2024static inline void tcp_reset_reno_sack(struct tcp_sock *tp) 2025{ 2026 tp->sacked_out = 0; 2027} 2028 2029static int tcp_is_sackfrto(const struct tcp_sock *tp) 2030{ 2031 return (sysctl_tcp_frto == 0x2) && !tcp_is_reno(tp); 2032} 2033 2034/* F-RTO can only be used if TCP has never retransmitted anything other than 2035 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here) 2036 */ 2037int tcp_use_frto(struct sock *sk) 2038{ 2039 const struct tcp_sock *tp = tcp_sk(sk); 2040 const struct inet_connection_sock *icsk = inet_csk(sk); 2041 struct sk_buff *skb; 2042 2043 if (!sysctl_tcp_frto) 2044 return 0; 2045 2046 /* MTU probe and F-RTO won't really play nicely along currently */ 2047 if (icsk->icsk_mtup.probe_size) 2048 return 0; 2049 2050 if (tcp_is_sackfrto(tp)) 2051 return 1; 2052 2053 /* Avoid expensive walking of rexmit queue if possible */ 2054 if (tp->retrans_out > 1) 2055 return 0; 2056 2057 skb = tcp_write_queue_head(sk); 2058 if (tcp_skb_is_last(sk, skb)) 2059 return 1; 2060 skb = tcp_write_queue_next(sk, skb); /* Skips head */ 2061 tcp_for_write_queue_from(skb, sk) { 2062 if (skb == tcp_send_head(sk)) 2063 break; 2064 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) 2065 return 0; 2066 /* Short-circuit when first non-SACKed skb has been checked */ 2067 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) 2068 break; 2069 } 2070 return 1; 2071} 2072 2073/* RTO occurred, but do not yet enter Loss state. Instead, defer RTO 2074 * recovery a bit and use heuristics in tcp_process_frto() to detect if 2075 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to 2076 * keep retrans_out counting accurate (with SACK F-RTO, other than head 2077 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS 2078 * bits are handled if the Loss state is really to be entered (in 2079 * tcp_enter_frto_loss). 2080 * 2081 * Do like tcp_enter_loss() would; when RTO expires the second time it 2082 * does: 2083 * "Reduce ssthresh if it has not yet been made inside this window." 2084 */ 2085void tcp_enter_frto(struct sock *sk) 2086{ 2087 const struct inet_connection_sock *icsk = inet_csk(sk); 2088 struct tcp_sock *tp = tcp_sk(sk); 2089 struct sk_buff *skb; 2090 2091 if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) || 2092 tp->snd_una == tp->high_seq || 2093 ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) && 2094 !icsk->icsk_retransmits)) { 2095 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2096 /* Our state is too optimistic in ssthresh() call because cwnd 2097 * is not reduced until tcp_enter_frto_loss() when previous F-RTO 2098 * recovery has not yet completed. Pattern would be this: RTO, 2099 * Cumulative ACK, RTO (2xRTO for the same segment does not end 2100 * up here twice). 2101 * RFC4138 should be more specific on what to do, even though 2102 * RTO is quite unlikely to occur after the first Cumulative ACK 2103 * due to back-off and complexity of triggering events ... 2104 */ 2105 if (tp->frto_counter) { 2106 u32 stored_cwnd; 2107 stored_cwnd = tp->snd_cwnd; 2108 tp->snd_cwnd = 2; 2109 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 2110 tp->snd_cwnd = stored_cwnd; 2111 } else { 2112 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 2113 } 2114 /* ... in theory, cong.control module could do "any tricks" in 2115 * ssthresh(), which means that ca_state, lost bits and lost_out 2116 * counter would have to be faked before the call occurs. We 2117 * consider that too expensive, unlikely and hacky, so modules 2118 * using these in ssthresh() must deal these incompatibility 2119 * issues if they receives CA_EVENT_FRTO and frto_counter != 0 2120 */ 2121 tcp_ca_event(sk, CA_EVENT_FRTO); 2122 } 2123 2124 tp->undo_marker = tp->snd_una; 2125 tp->undo_retrans = 0; 2126 2127 skb = tcp_write_queue_head(sk); 2128 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) 2129 tp->undo_marker = 0; 2130 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 2131 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 2132 tp->retrans_out -= tcp_skb_pcount(skb); 2133 } 2134 tcp_verify_left_out(tp); 2135 2136 /* Too bad if TCP was application limited */ 2137 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1); 2138 2139 /* Earlier loss recovery underway (see RFC4138; Appendix B). 2140 * The last condition is necessary at least in tp->frto_counter case. 2141 */ 2142 if (tcp_is_sackfrto(tp) && (tp->frto_counter || 2143 ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) && 2144 after(tp->high_seq, tp->snd_una)) { 2145 tp->frto_highmark = tp->high_seq; 2146 } else { 2147 tp->frto_highmark = tp->snd_nxt; 2148 } 2149 tcp_set_ca_state(sk, TCP_CA_Disorder); 2150 tp->high_seq = tp->snd_nxt; 2151 tp->frto_counter = 1; 2152} 2153 2154/* Enter Loss state after F-RTO was applied. Dupack arrived after RTO, 2155 * which indicates that we should follow the traditional RTO recovery, 2156 * i.e. mark everything lost and do go-back-N retransmission. 2157 */ 2158static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag) 2159{ 2160 struct tcp_sock *tp = tcp_sk(sk); 2161 struct sk_buff *skb; 2162 2163 tp->lost_out = 0; 2164 tp->retrans_out = 0; 2165 if (tcp_is_reno(tp)) 2166 tcp_reset_reno_sack(tp); 2167 2168 tcp_for_write_queue(skb, sk) { 2169 if (skb == tcp_send_head(sk)) 2170 break; 2171 2172 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; 2173 /* 2174 * Count the retransmission made on RTO correctly (only when 2175 * waiting for the first ACK and did not get it)... 2176 */ 2177 if ((tp->frto_counter == 1) && !(flag & FLAG_DATA_ACKED)) { 2178 /* For some reason this R-bit might get cleared? */ 2179 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 2180 tp->retrans_out += tcp_skb_pcount(skb); 2181 /* ...enter this if branch just for the first segment */ 2182 flag |= FLAG_DATA_ACKED; 2183 } else { 2184 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) 2185 tp->undo_marker = 0; 2186 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 2187 } 2188 2189 /* Marking forward transmissions that were made after RTO lost 2190 * can cause unnecessary retransmissions in some scenarios, 2191 * SACK blocks will mitigate that in some but not in all cases. 2192 * We used to not mark them but it was causing break-ups with 2193 * receivers that do only in-order receival. 2194 * 2195 * TODO: we could detect presence of such receiver and select 2196 * different behavior per flow. 2197 */ 2198 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) { 2199 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 2200 tp->lost_out += tcp_skb_pcount(skb); 2201 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq; 2202 } 2203 } 2204 tcp_verify_left_out(tp); 2205 2206 tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments; 2207 tp->snd_cwnd_cnt = 0; 2208 tp->snd_cwnd_stamp = tcp_time_stamp; 2209 tp->frto_counter = 0; 2210 tp->bytes_acked = 0; 2211 2212 tp->reordering = min_t(unsigned int, tp->reordering, 2213 sysctl_tcp_reordering); 2214 tcp_set_ca_state(sk, TCP_CA_Loss); 2215 tp->high_seq = tp->snd_nxt; 2216 TCP_ECN_queue_cwr(tp); 2217 2218 tcp_clear_all_retrans_hints(tp); 2219} 2220 2221static void tcp_clear_retrans_partial(struct tcp_sock *tp) 2222{ 2223 tp->retrans_out = 0; 2224 tp->lost_out = 0; 2225 2226 tp->undo_marker = 0; 2227 tp->undo_retrans = 0; 2228} 2229 2230void tcp_clear_retrans(struct tcp_sock *tp) 2231{ 2232 tcp_clear_retrans_partial(tp); 2233 2234 tp->fackets_out = 0; 2235 tp->sacked_out = 0; 2236} 2237 2238/* Enter Loss state. If "how" is not zero, forget all SACK information 2239 * and reset tags completely, otherwise preserve SACKs. If receiver 2240 * dropped its ofo queue, we will know this due to reneging detection. 2241 */ 2242void tcp_enter_loss(struct sock *sk, int how) 2243{ 2244 const struct inet_connection_sock *icsk = inet_csk(sk); 2245 struct tcp_sock *tp = tcp_sk(sk); 2246 struct sk_buff *skb; 2247 2248 /* Reduce ssthresh if it has not yet been made inside this window. */ 2249 if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq || 2250 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) { 2251 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2252 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 2253 tcp_ca_event(sk, CA_EVENT_LOSS); 2254 } 2255 tp->snd_cwnd = 1; 2256 tp->snd_cwnd_cnt = 0; 2257 tp->snd_cwnd_stamp = tcp_time_stamp; 2258 2259 tp->bytes_acked = 0; 2260 tcp_clear_retrans_partial(tp); 2261 2262 if (tcp_is_reno(tp)) 2263 tcp_reset_reno_sack(tp); 2264 2265 if (!how) { 2266 /* Push undo marker, if it was plain RTO and nothing 2267 * was retransmitted. */ 2268 tp->undo_marker = tp->snd_una; 2269 } else { 2270 tp->sacked_out = 0; 2271 tp->fackets_out = 0; 2272 } 2273 tcp_clear_all_retrans_hints(tp); 2274 2275 tcp_for_write_queue(skb, sk) { 2276 if (skb == tcp_send_head(sk)) 2277 break; 2278 2279 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) 2280 tp->undo_marker = 0; 2281 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED; 2282 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) { 2283 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED; 2284 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 2285 tp->lost_out += tcp_skb_pcount(skb); 2286 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq; 2287 } 2288 } 2289 tcp_verify_left_out(tp); 2290 2291 tp->reordering = min_t(unsigned int, tp->reordering, 2292 sysctl_tcp_reordering); 2293 tcp_set_ca_state(sk, TCP_CA_Loss); 2294 tp->high_seq = tp->snd_nxt; 2295 TCP_ECN_queue_cwr(tp); 2296 /* Abort F-RTO algorithm if one is in progress */ 2297 tp->frto_counter = 0; 2298} 2299 2300/* If ACK arrived pointing to a remembered SACK, it means that our 2301 * remembered SACKs do not reflect real state of receiver i.e. 2302 * receiver _host_ is heavily congested (or buggy). 2303 * 2304 * Do processing similar to RTO timeout. 2305 */ 2306static int tcp_check_sack_reneging(struct sock *sk, int flag) 2307{ 2308 if (flag & FLAG_SACK_RENEGING) { 2309 struct inet_connection_sock *icsk = inet_csk(sk); 2310 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING); 2311 2312 tcp_enter_loss(sk, 1); 2313 icsk->icsk_retransmits++; 2314 tcp_retransmit_skb(sk, tcp_write_queue_head(sk)); 2315 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 2316 icsk->icsk_rto, TCP_RTO_MAX); 2317 return 1; 2318 } 2319 return 0; 2320} 2321 2322static inline int tcp_fackets_out(const struct tcp_sock *tp) 2323{ 2324 return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out; 2325} 2326 2327/* Heurestics to calculate number of duplicate ACKs. There's no dupACKs 2328 * counter when SACK is enabled (without SACK, sacked_out is used for 2329 * that purpose). 2330 * 2331 * Instead, with FACK TCP uses fackets_out that includes both SACKed 2332 * segments up to the highest received SACK block so far and holes in 2333 * between them. 2334 * 2335 * With reordering, holes may still be in flight, so RFC3517 recovery 2336 * uses pure sacked_out (total number of SACKed segments) even though 2337 * it violates the RFC that uses duplicate ACKs, often these are equal 2338 * but when e.g. out-of-window ACKs or packet duplication occurs, 2339 * they differ. Since neither occurs due to loss, TCP should really 2340 * ignore them. 2341 */ 2342static inline int tcp_dupack_heuristics(const struct tcp_sock *tp) 2343{ 2344 return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1; 2345} 2346 2347static bool tcp_pause_early_retransmit(struct sock *sk, int flag) 2348{ 2349 struct tcp_sock *tp = tcp_sk(sk); 2350 unsigned long delay; 2351 2352 /* Delay early retransmit and entering fast recovery for 2353 * max(RTT/4, 2msec) unless ack has ECE mark, no RTT samples 2354 * available, or RTO is scheduled to fire first. 2355 */ 2356 if (sysctl_tcp_early_retrans < 2 || (flag & FLAG_ECE) || !tp->srtt) 2357 return false; 2358 2359 delay = max_t(unsigned long, (tp->srtt >> 5), msecs_to_jiffies(2)); 2360 if (!time_after(inet_csk(sk)->icsk_timeout, (jiffies + delay))) 2361 return false; 2362 2363 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, delay, TCP_RTO_MAX); 2364 tp->early_retrans_delayed = 1; 2365 return true; 2366} 2367 2368static inline int tcp_skb_timedout(const struct sock *sk, 2369 const struct sk_buff *skb) 2370{ 2371 return tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto; 2372} 2373 2374static inline int tcp_head_timedout(const struct sock *sk) 2375{ 2376 const struct tcp_sock *tp = tcp_sk(sk); 2377 2378 return tp->packets_out && 2379 tcp_skb_timedout(sk, tcp_write_queue_head(sk)); 2380} 2381 2382/* Linux NewReno/SACK/FACK/ECN state machine. 2383 * -------------------------------------- 2384 * 2385 * "Open" Normal state, no dubious events, fast path. 2386 * "Disorder" In all the respects it is "Open", 2387 * but requires a bit more attention. It is entered when 2388 * we see some SACKs or dupacks. It is split of "Open" 2389 * mainly to move some processing from fast path to slow one. 2390 * "CWR" CWND was reduced due to some Congestion Notification event. 2391 * It can be ECN, ICMP source quench, local device congestion. 2392 * "Recovery" CWND was reduced, we are fast-retransmitting. 2393 * "Loss" CWND was reduced due to RTO timeout or SACK reneging. 2394 * 2395 * tcp_fastretrans_alert() is entered: 2396 * - each incoming ACK, if state is not "Open" 2397 * - when arrived ACK is unusual, namely: 2398 * * SACK 2399 * * Duplicate ACK. 2400 * * ECN ECE. 2401 * 2402 * Counting packets in flight is pretty simple. 2403 * 2404 * in_flight = packets_out - left_out + retrans_out 2405 * 2406 * packets_out is SND.NXT-SND.UNA counted in packets. 2407 * 2408 * retrans_out is number of retransmitted segments. 2409 * 2410 * left_out is number of segments left network, but not ACKed yet. 2411 * 2412 * left_out = sacked_out + lost_out 2413 * 2414 * sacked_out: Packets, which arrived to receiver out of order 2415 * and hence not ACKed. With SACKs this number is simply 2416 * amount of SACKed data. Even without SACKs 2417 * it is easy to give pretty reliable estimate of this number, 2418 * counting duplicate ACKs. 2419 * 2420 * lost_out: Packets lost by network. TCP has no explicit 2421 * "loss notification" feedback from network (for now). 2422 * It means that this number can be only _guessed_. 2423 * Actually, it is the heuristics to predict lossage that 2424 * distinguishes different algorithms. 2425 * 2426 * F.e. after RTO, when all the queue is considered as lost, 2427 * lost_out = packets_out and in_flight = retrans_out. 2428 * 2429 * Essentially, we have now two algorithms counting 2430 * lost packets. 2431 * 2432 * FACK: It is the simplest heuristics. As soon as we decided 2433 * that something is lost, we decide that _all_ not SACKed 2434 * packets until the most forward SACK are lost. I.e. 2435 * lost_out = fackets_out - sacked_out and left_out = fackets_out. 2436 * It is absolutely correct estimate, if network does not reorder 2437 * packets. And it loses any connection to reality when reordering 2438 * takes place. We use FACK by default until reordering 2439 * is suspected on the path to this destination. 2440 * 2441 * NewReno: when Recovery is entered, we assume that one segment 2442 * is lost (classic Reno). While we are in Recovery and 2443 * a partial ACK arrives, we assume that one more packet 2444 * is lost (NewReno). This heuristics are the same in NewReno 2445 * and SACK. 2446 * 2447 * Imagine, that's all! Forget about all this shamanism about CWND inflation 2448 * deflation etc. CWND is real congestion window, never inflated, changes 2449 * only according to classic VJ rules. 2450 * 2451 * Really tricky (and requiring careful tuning) part of algorithm 2452 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue(). 2453 * The first determines the moment _when_ we should reduce CWND and, 2454 * hence, slow down forward transmission. In fact, it determines the moment 2455 * when we decide that hole is caused by loss, rather than by a reorder. 2456 * 2457 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill 2458 * holes, caused by lost packets. 2459 * 2460 * And the most logically complicated part of algorithm is undo 2461 * heuristics. We detect false retransmits due to both too early 2462 * fast retransmit (reordering) and underestimated RTO, analyzing 2463 * timestamps and D-SACKs. When we detect that some segments were 2464 * retransmitted by mistake and CWND reduction was wrong, we undo 2465 * window reduction and abort recovery phase. This logic is hidden 2466 * inside several functions named tcp_try_undo_<something>. 2467 */ 2468 2469/* This function decides, when we should leave Disordered state 2470 * and enter Recovery phase, reducing congestion window. 2471 * 2472 * Main question: may we further continue forward transmission 2473 * with the same cwnd? 2474 */ 2475static int tcp_time_to_recover(struct sock *sk, int flag) 2476{ 2477 struct tcp_sock *tp = tcp_sk(sk); 2478 __u32 packets_out; 2479 2480 /* Do not perform any recovery during F-RTO algorithm */ 2481 if (tp->frto_counter) 2482 return 0; 2483 2484 /* Trick#1: The loss is proven. */ 2485 if (tp->lost_out) 2486 return 1; 2487 2488 /* Not-A-Trick#2 : Classic rule... */ 2489 if (tcp_dupack_heuristics(tp) > tp->reordering) 2490 return 1; 2491 2492 /* Trick#3 : when we use RFC2988 timer restart, fast 2493 * retransmit can be triggered by timeout of queue head. 2494 */ 2495 if (tcp_is_fack(tp) && tcp_head_timedout(sk)) 2496 return 1; 2497 2498 /* Trick#4: It is still not OK... But will it be useful to delay 2499 * recovery more? 2500 */ 2501 packets_out = tp->packets_out; 2502 if (packets_out <= tp->reordering && 2503 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) && 2504 !tcp_may_send_now(sk)) { 2505 /* We have nothing to send. This connection is limited 2506 * either by receiver window or by application. 2507 */ 2508 return 1; 2509 } 2510 2511 /* If a thin stream is detected, retransmit after first 2512 * received dupack. Employ only if SACK is supported in order 2513 * to avoid possible corner-case series of spurious retransmissions 2514 * Use only if there are no unsent data. 2515 */ 2516 if ((tp->thin_dupack || sysctl_tcp_thin_dupack) && 2517 tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 && 2518 tcp_is_sack(tp) && !tcp_send_head(sk)) 2519 return 1; 2520 2521 /* Trick#6: TCP early retransmit, per RFC5827. To avoid spurious 2522 * retransmissions due to small network reorderings, we implement 2523 * Mitigation A.3 in the RFC and delay the retransmission for a short 2524 * interval if appropriate. 2525 */ 2526 if (tp->do_early_retrans && !tp->retrans_out && tp->sacked_out && 2527 (tp->packets_out == (tp->sacked_out + 1) && tp->packets_out < 4) && 2528 !tcp_may_send_now(sk)) 2529 return !tcp_pause_early_retransmit(sk, flag); 2530 2531 return 0; 2532} 2533 2534/* New heuristics: it is possible only after we switched to restart timer 2535 * each time when something is ACKed. Hence, we can detect timed out packets 2536 * during fast retransmit without falling to slow start. 2537 * 2538 * Usefulness of this as is very questionable, since we should know which of 2539 * the segments is the next to timeout which is relatively expensive to find 2540 * in general case unless we add some data structure just for that. The 2541 * current approach certainly won't find the right one too often and when it 2542 * finally does find _something_ it usually marks large part of the window 2543 * right away (because a retransmission with a larger timestamp blocks the 2544 * loop from advancing). -ij 2545 */ 2546static void tcp_timeout_skbs(struct sock *sk) 2547{ 2548 struct tcp_sock *tp = tcp_sk(sk); 2549 struct sk_buff *skb; 2550 2551 if (!tcp_is_fack(tp) || !tcp_head_timedout(sk)) 2552 return; 2553 2554 skb = tp->scoreboard_skb_hint; 2555 if (tp->scoreboard_skb_hint == NULL) 2556 skb = tcp_write_queue_head(sk); 2557 2558 tcp_for_write_queue_from(skb, sk) { 2559 if (skb == tcp_send_head(sk)) 2560 break; 2561 if (!tcp_skb_timedout(sk, skb)) 2562 break; 2563 2564 tcp_skb_mark_lost(tp, skb); 2565 } 2566 2567 tp->scoreboard_skb_hint = skb; 2568 2569 tcp_verify_left_out(tp); 2570} 2571 2572/* Detect loss in event "A" above by marking head of queue up as lost. 2573 * For FACK or non-SACK(Reno) senders, the first "packets" number of segments 2574 * are considered lost. For RFC3517 SACK, a segment is considered lost if it 2575 * has at least tp->reordering SACKed seqments above it; "packets" refers to 2576 * the maximum SACKed segments to pass before reaching this limit. 2577 */ 2578static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head) 2579{ 2580 struct tcp_sock *tp = tcp_sk(sk); 2581 struct sk_buff *skb; 2582 int cnt, oldcnt; 2583 int err; 2584 unsigned int mss; 2585 /* Use SACK to deduce losses of new sequences sent during recovery */ 2586 const u32 loss_high = tcp_is_sack(tp) ? tp->snd_nxt : tp->high_seq; 2587 2588 WARN_ON(packets > tp->packets_out); 2589 if (tp->lost_skb_hint) { 2590 skb = tp->lost_skb_hint; 2591 cnt = tp->lost_cnt_hint; 2592 /* Head already handled? */ 2593 if (mark_head && skb != tcp_write_queue_head(sk)) 2594 return; 2595 } else { 2596 skb = tcp_write_queue_head(sk); 2597 cnt = 0; 2598 } 2599 2600 tcp_for_write_queue_from(skb, sk) { 2601 if (skb == tcp_send_head(sk)) 2602 break; 2603 /* TODO: do this better */ 2604 /* this is not the most efficient way to do this... */ 2605 tp->lost_skb_hint = skb; 2606 tp->lost_cnt_hint = cnt; 2607 2608 if (after(TCP_SKB_CB(skb)->end_seq, loss_high)) 2609 break; 2610 2611 oldcnt = cnt; 2612 if (tcp_is_fack(tp) || tcp_is_reno(tp) || 2613 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) 2614 cnt += tcp_skb_pcount(skb); 2615 2616 if (cnt > packets) { 2617 if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) || 2618 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) || 2619 (oldcnt >= packets)) 2620 break; 2621 2622 mss = skb_shinfo(skb)->gso_size; 2623 err = tcp_fragment(sk, skb, (packets - oldcnt) * mss, mss); 2624 if (err < 0) 2625 break; 2626 cnt = packets; 2627 } 2628 2629 tcp_skb_mark_lost(tp, skb); 2630 2631 if (mark_head) 2632 break; 2633 } 2634 tcp_verify_left_out(tp); 2635} 2636 2637/* Account newly detected lost packet(s) */ 2638 2639static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit) 2640{ 2641 struct tcp_sock *tp = tcp_sk(sk); 2642 2643 if (tcp_is_reno(tp)) { 2644 tcp_mark_head_lost(sk, 1, 1); 2645 } else if (tcp_is_fack(tp)) { 2646 int lost = tp->fackets_out - tp->reordering; 2647 if (lost <= 0) 2648 lost = 1; 2649 tcp_mark_head_lost(sk, lost, 0); 2650 } else { 2651 int sacked_upto = tp->sacked_out - tp->reordering; 2652 if (sacked_upto >= 0) 2653 tcp_mark_head_lost(sk, sacked_upto, 0); 2654 else if (fast_rexmit) 2655 tcp_mark_head_lost(sk, 1, 1); 2656 } 2657 2658 tcp_timeout_skbs(sk); 2659} 2660 2661/* CWND moderation, preventing bursts due to too big ACKs 2662 * in dubious situations. 2663 */ 2664static inline void tcp_moderate_cwnd(struct tcp_sock *tp) 2665{ 2666 tp->snd_cwnd = min(tp->snd_cwnd, 2667 tcp_packets_in_flight(tp) + tcp_max_burst(tp)); 2668 tp->snd_cwnd_stamp = tcp_time_stamp; 2669} 2670 2671/* Lower bound on congestion window is slow start threshold 2672 * unless congestion avoidance choice decides to overide it. 2673 */ 2674static inline u32 tcp_cwnd_min(const struct sock *sk) 2675{ 2676 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 2677 2678 return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh; 2679} 2680 2681/* Decrease cwnd each second ack. */ 2682static void tcp_cwnd_down(struct sock *sk, int flag) 2683{ 2684 struct tcp_sock *tp = tcp_sk(sk); 2685 int decr = tp->snd_cwnd_cnt + 1; 2686 2687 if ((flag & (FLAG_ANY_PROGRESS | FLAG_DSACKING_ACK)) || 2688 (tcp_is_reno(tp) && !(flag & FLAG_NOT_DUP))) { 2689 tp->snd_cwnd_cnt = decr & 1; 2690 decr >>= 1; 2691 2692 if (decr && tp->snd_cwnd > tcp_cwnd_min(sk)) 2693 tp->snd_cwnd -= decr; 2694 2695 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1); 2696 tp->snd_cwnd_stamp = tcp_time_stamp; 2697 } 2698} 2699 2700/* Nothing was retransmitted or returned timestamp is less 2701 * than timestamp of the first retransmission. 2702 */ 2703static inline int tcp_packet_delayed(const struct tcp_sock *tp) 2704{ 2705 return !tp->retrans_stamp || 2706 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 2707 before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp)); 2708} 2709 2710/* Undo procedures. */ 2711 2712#if FASTRETRANS_DEBUG > 1 2713static void DBGUNDO(struct sock *sk, const char *msg) 2714{ 2715 struct tcp_sock *tp = tcp_sk(sk); 2716 struct inet_sock *inet = inet_sk(sk); 2717 2718 if (sk->sk_family == AF_INET) { 2719 printk(KERN_DEBUG "Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n", 2720 msg, 2721 &inet->inet_daddr, ntohs(inet->inet_dport), 2722 tp->snd_cwnd, tcp_left_out(tp), 2723 tp->snd_ssthresh, tp->prior_ssthresh, 2724 tp->packets_out); 2725 } 2726#if IS_ENABLED(CONFIG_IPV6) 2727 else if (sk->sk_family == AF_INET6) { 2728 struct ipv6_pinfo *np = inet6_sk(sk); 2729 printk(KERN_DEBUG "Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n", 2730 msg, 2731 &np->daddr, ntohs(inet->inet_dport), 2732 tp->snd_cwnd, tcp_left_out(tp), 2733 tp->snd_ssthresh, tp->prior_ssthresh, 2734 tp->packets_out); 2735 } 2736#endif 2737} 2738#else 2739#define DBGUNDO(x...) do { } while (0) 2740#endif 2741 2742static void tcp_undo_cwr(struct sock *sk, const bool undo_ssthresh) 2743{ 2744 struct tcp_sock *tp = tcp_sk(sk); 2745 2746 if (tp->prior_ssthresh) { 2747 const struct inet_connection_sock *icsk = inet_csk(sk); 2748 2749 if (icsk->icsk_ca_ops->undo_cwnd) 2750 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk); 2751 else 2752 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1); 2753 2754 if (undo_ssthresh && tp->prior_ssthresh > tp->snd_ssthresh) { 2755 tp->snd_ssthresh = tp->prior_ssthresh; 2756 TCP_ECN_withdraw_cwr(tp); 2757 } 2758 } else { 2759 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh); 2760 } 2761 tp->snd_cwnd_stamp = tcp_time_stamp; 2762} 2763 2764static inline int tcp_may_undo(const struct tcp_sock *tp) 2765{ 2766 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp)); 2767} 2768 2769/* People celebrate: "We love our President!" */ 2770static int tcp_try_undo_recovery(struct sock *sk) 2771{ 2772 struct tcp_sock *tp = tcp_sk(sk); 2773 2774 if (tcp_may_undo(tp)) { 2775 int mib_idx; 2776 2777 /* Happy end! We did not retransmit anything 2778 * or our original transmission succeeded. 2779 */ 2780 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans"); 2781 tcp_undo_cwr(sk, true); 2782 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss) 2783 mib_idx = LINUX_MIB_TCPLOSSUNDO; 2784 else 2785 mib_idx = LINUX_MIB_TCPFULLUNDO; 2786 2787 NET_INC_STATS_BH(sock_net(sk), mib_idx); 2788 tp->undo_marker = 0; 2789 } 2790 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) { 2791 /* Hold old state until something *above* high_seq 2792 * is ACKed. For Reno it is MUST to prevent false 2793 * fast retransmits (RFC2582). SACK TCP is safe. */ 2794 tcp_moderate_cwnd(tp); 2795 return 1; 2796 } 2797 tcp_set_ca_state(sk, TCP_CA_Open); 2798 return 0; 2799} 2800 2801/* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */ 2802static void tcp_try_undo_dsack(struct sock *sk) 2803{ 2804 struct tcp_sock *tp = tcp_sk(sk); 2805 2806 if (tp->undo_marker && !tp->undo_retrans) { 2807 DBGUNDO(sk, "D-SACK"); 2808 tcp_undo_cwr(sk, true); 2809 tp->undo_marker = 0; 2810 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO); 2811 } 2812} 2813 2814/* We can clear retrans_stamp when there are no retransmissions in the 2815 * window. It would seem that it is trivially available for us in 2816 * tp->retrans_out, however, that kind of assumptions doesn't consider 2817 * what will happen if errors occur when sending retransmission for the 2818 * second time. ...It could the that such segment has only 2819 * TCPCB_EVER_RETRANS set at the present time. It seems that checking 2820 * the head skb is enough except for some reneging corner cases that 2821 * are not worth the effort. 2822 * 2823 * Main reason for all this complexity is the fact that connection dying 2824 * time now depends on the validity of the retrans_stamp, in particular, 2825 * that successive retransmissions of a segment must not advance 2826 * retrans_stamp under any conditions. 2827 */ 2828static int tcp_any_retrans_done(const struct sock *sk) 2829{ 2830 const struct tcp_sock *tp = tcp_sk(sk); 2831 struct sk_buff *skb; 2832 2833 if (tp->retrans_out) 2834 return 1; 2835 2836 skb = tcp_write_queue_head(sk); 2837 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS)) 2838 return 1; 2839 2840 return 0; 2841} 2842 2843/* Undo during fast recovery after partial ACK. */ 2844 2845static int tcp_try_undo_partial(struct sock *sk, int acked) 2846{ 2847 struct tcp_sock *tp = tcp_sk(sk); 2848 /* Partial ACK arrived. Force Hoe's retransmit. */ 2849 int failed = tcp_is_reno(tp) || (tcp_fackets_out(tp) > tp->reordering); 2850 2851 if (tcp_may_undo(tp)) { 2852 /* Plain luck! Hole if filled with delayed 2853 * packet, rather than with a retransmit. 2854 */ 2855 if (!tcp_any_retrans_done(sk)) 2856 tp->retrans_stamp = 0; 2857 2858 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1); 2859 2860 DBGUNDO(sk, "Hoe"); 2861 tcp_undo_cwr(sk, false); 2862 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO); 2863 2864 /* So... Do not make Hoe's retransmit yet. 2865 * If the first packet was delayed, the rest 2866 * ones are most probably delayed as well. 2867 */ 2868 failed = 0; 2869 } 2870 return failed; 2871} 2872 2873/* Undo during loss recovery after partial ACK. */ 2874static int tcp_try_undo_loss(struct sock *sk) 2875{ 2876 struct tcp_sock *tp = tcp_sk(sk); 2877 2878 if (tcp_may_undo(tp)) { 2879 struct sk_buff *skb; 2880 tcp_for_write_queue(skb, sk) { 2881 if (skb == tcp_send_head(sk)) 2882 break; 2883 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; 2884 } 2885 2886 tcp_clear_all_retrans_hints(tp); 2887 2888 DBGUNDO(sk, "partial loss"); 2889 tp->lost_out = 0; 2890 tcp_undo_cwr(sk, true); 2891 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO); 2892 inet_csk(sk)->icsk_retransmits = 0; 2893 tp->undo_marker = 0; 2894 if (tcp_is_sack(tp)) 2895 tcp_set_ca_state(sk, TCP_CA_Open); 2896 return 1; 2897 } 2898 return 0; 2899} 2900 2901static inline void tcp_complete_cwr(struct sock *sk) 2902{ 2903 struct tcp_sock *tp = tcp_sk(sk); 2904 2905 /* Do not moderate cwnd if it's already undone in cwr or recovery. */ 2906 if (tp->undo_marker) { 2907 if (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR) 2908 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh); 2909 else /* PRR */ 2910 tp->snd_cwnd = tp->snd_ssthresh; 2911 tp->snd_cwnd_stamp = tcp_time_stamp; 2912 } 2913 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR); 2914} 2915 2916static void tcp_try_keep_open(struct sock *sk) 2917{ 2918 struct tcp_sock *tp = tcp_sk(sk); 2919 int state = TCP_CA_Open; 2920 2921 if (tcp_left_out(tp) || tcp_any_retrans_done(sk)) 2922 state = TCP_CA_Disorder; 2923 2924 if (inet_csk(sk)->icsk_ca_state != state) { 2925 tcp_set_ca_state(sk, state); 2926 tp->high_seq = tp->snd_nxt; 2927 } 2928} 2929 2930static void tcp_try_to_open(struct sock *sk, int flag) 2931{ 2932 struct tcp_sock *tp = tcp_sk(sk); 2933 2934 tcp_verify_left_out(tp); 2935 2936 if (!tp->frto_counter && !tcp_any_retrans_done(sk)) 2937 tp->retrans_stamp = 0; 2938 2939 if (flag & FLAG_ECE) 2940 tcp_enter_cwr(sk, 1); 2941 2942 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) { 2943 tcp_try_keep_open(sk); 2944 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Open) 2945 tcp_moderate_cwnd(tp); 2946 } else { 2947 tcp_cwnd_down(sk, flag); 2948 } 2949} 2950 2951static void tcp_mtup_probe_failed(struct sock *sk) 2952{ 2953 struct inet_connection_sock *icsk = inet_csk(sk); 2954 2955 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1; 2956 icsk->icsk_mtup.probe_size = 0; 2957} 2958 2959static void tcp_mtup_probe_success(struct sock *sk) 2960{ 2961 struct tcp_sock *tp = tcp_sk(sk); 2962 struct inet_connection_sock *icsk = inet_csk(sk); 2963 2964 /* FIXME: breaks with very large cwnd */ 2965 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2966 tp->snd_cwnd = tp->snd_cwnd * 2967 tcp_mss_to_mtu(sk, tp->mss_cache) / 2968 icsk->icsk_mtup.probe_size; 2969 tp->snd_cwnd_cnt = 0; 2970 tp->snd_cwnd_stamp = tcp_time_stamp; 2971 tp->snd_ssthresh = tcp_current_ssthresh(sk); 2972 2973 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size; 2974 icsk->icsk_mtup.probe_size = 0; 2975 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 2976} 2977 2978/* Do a simple retransmit without using the backoff mechanisms in 2979 * tcp_timer. This is used for path mtu discovery. 2980 * The socket is already locked here. 2981 */ 2982void tcp_simple_retransmit(struct sock *sk) 2983{ 2984 const struct inet_connection_sock *icsk = inet_csk(sk); 2985 struct tcp_sock *tp = tcp_sk(sk); 2986 struct sk_buff *skb; 2987 unsigned int mss = tcp_current_mss(sk); 2988 u32 prior_lost = tp->lost_out; 2989 2990 tcp_for_write_queue(skb, sk) { 2991 if (skb == tcp_send_head(sk)) 2992 break; 2993 if (tcp_skb_seglen(skb) > mss && 2994 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) { 2995 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 2996 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 2997 tp->retrans_out -= tcp_skb_pcount(skb); 2998 } 2999 tcp_skb_mark_lost_uncond_verify(tp, skb); 3000 } 3001 } 3002 3003 tcp_clear_retrans_hints_partial(tp); 3004 3005 if (prior_lost == tp->lost_out) 3006 return; 3007 3008 if (tcp_is_reno(tp)) 3009 tcp_limit_reno_sacked(tp); 3010 3011 tcp_verify_left_out(tp); 3012 3013 /* Don't muck with the congestion window here. 3014 * Reason is that we do not increase amount of _data_ 3015 * in network, but units changed and effective 3016 * cwnd/ssthresh really reduced now. 3017 */ 3018 if (icsk->icsk_ca_state != TCP_CA_Loss) { 3019 tp->high_seq = tp->snd_nxt; 3020 tp->snd_ssthresh = tcp_current_ssthresh(sk); 3021 tp->prior_ssthresh = 0; 3022 tp->undo_marker = 0; 3023 tcp_set_ca_state(sk, TCP_CA_Loss); 3024 } 3025 tcp_xmit_retransmit_queue(sk); 3026} 3027EXPORT_SYMBOL(tcp_simple_retransmit); 3028 3029/* This function implements the PRR algorithm, specifcally the PRR-SSRB 3030 * (proportional rate reduction with slow start reduction bound) as described in 3031 * http://www.ietf.org/id/draft-mathis-tcpm-proportional-rate-reduction-01.txt. 3032 * It computes the number of packets to send (sndcnt) based on packets newly 3033 * delivered: 3034 * 1) If the packets in flight is larger than ssthresh, PRR spreads the 3035 * cwnd reductions across a full RTT. 3036 * 2) If packets in flight is lower than ssthresh (such as due to excess 3037 * losses and/or application stalls), do not perform any further cwnd 3038 * reductions, but instead slow start up to ssthresh. 3039 */ 3040static void tcp_update_cwnd_in_recovery(struct sock *sk, int newly_acked_sacked, 3041 int fast_rexmit, int flag) 3042{ 3043 struct tcp_sock *tp = tcp_sk(sk); 3044 int sndcnt = 0; 3045 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp); 3046 3047 if (tcp_packets_in_flight(tp) > tp->snd_ssthresh) { 3048 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered + 3049 tp->prior_cwnd - 1; 3050 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out; 3051 } else { 3052 sndcnt = min_t(int, delta, 3053 max_t(int, tp->prr_delivered - tp->prr_out, 3054 newly_acked_sacked) + 1); 3055 } 3056 3057 sndcnt = max(sndcnt, (fast_rexmit ? 1 : 0)); 3058 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt; 3059} 3060 3061static void tcp_enter_recovery(struct sock *sk, bool ece_ack) 3062{ 3063 struct tcp_sock *tp = tcp_sk(sk); 3064 int mib_idx; 3065 3066 if (tcp_is_reno(tp)) 3067 mib_idx = LINUX_MIB_TCPRENORECOVERY; 3068 else 3069 mib_idx = LINUX_MIB_TCPSACKRECOVERY; 3070 3071 NET_INC_STATS_BH(sock_net(sk), mib_idx); 3072 3073 tp->high_seq = tp->snd_nxt; 3074 tp->prior_ssthresh = 0; 3075 tp->undo_marker = tp->snd_una; 3076 tp->undo_retrans = tp->retrans_out; 3077 3078 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) { 3079 if (!ece_ack) 3080 tp->prior_ssthresh = tcp_current_ssthresh(sk); 3081 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk); 3082 TCP_ECN_queue_cwr(tp); 3083 } 3084 3085 tp->bytes_acked = 0; 3086 tp->snd_cwnd_cnt = 0; 3087 tp->prior_cwnd = tp->snd_cwnd; 3088 tp->prr_delivered = 0; 3089 tp->prr_out = 0; 3090 tcp_set_ca_state(sk, TCP_CA_Recovery); 3091} 3092 3093/* Process an event, which can update packets-in-flight not trivially. 3094 * Main goal of this function is to calculate new estimate for left_out, 3095 * taking into account both packets sitting in receiver's buffer and 3096 * packets lost by network. 3097 * 3098 * Besides that it does CWND reduction, when packet loss is detected 3099 * and changes state of machine. 3100 * 3101 * It does _not_ decide what to send, it is made in function 3102 * tcp_xmit_retransmit_queue(). 3103 */ 3104static void tcp_fastretrans_alert(struct sock *sk, int pkts_acked, 3105 int newly_acked_sacked, bool is_dupack, 3106 int flag) 3107{ 3108 struct inet_connection_sock *icsk = inet_csk(sk); 3109 struct tcp_sock *tp = tcp_sk(sk); 3110 int do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) && 3111 (tcp_fackets_out(tp) > tp->reordering)); 3112 int fast_rexmit = 0; 3113 3114 if (WARN_ON(!tp->packets_out && tp->sacked_out)) 3115 tp->sacked_out = 0; 3116 if (WARN_ON(!tp->sacked_out && tp->fackets_out)) 3117 tp->fackets_out = 0; 3118 3119 /* Now state machine starts. 3120 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */ 3121 if (flag & FLAG_ECE) 3122 tp->prior_ssthresh = 0; 3123 3124 /* B. In all the states check for reneging SACKs. */ 3125 if (tcp_check_sack_reneging(sk, flag)) 3126 return; 3127 3128 /* C. Check consistency of the current state. */ 3129 tcp_verify_left_out(tp); 3130 3131 /* D. Check state exit conditions. State can be terminated 3132 * when high_seq is ACKed. */ 3133 if (icsk->icsk_ca_state == TCP_CA_Open) { 3134 WARN_ON(tp->retrans_out != 0); 3135 tp->retrans_stamp = 0; 3136 } else if (!before(tp->snd_una, tp->high_seq)) { 3137 switch (icsk->icsk_ca_state) { 3138 case TCP_CA_Loss: 3139 icsk->icsk_retransmits = 0; 3140 if (tcp_try_undo_recovery(sk)) 3141 return; 3142 break; 3143 3144 case TCP_CA_CWR: 3145 /* CWR is to be held something *above* high_seq 3146 * is ACKed for CWR bit to reach receiver. */ 3147 if (tp->snd_una != tp->high_seq) { 3148 tcp_complete_cwr(sk); 3149 tcp_set_ca_state(sk, TCP_CA_Open); 3150 } 3151 break; 3152 3153 case TCP_CA_Recovery: 3154 if (tcp_is_reno(tp)) 3155 tcp_reset_reno_sack(tp); 3156 if (tcp_try_undo_recovery(sk)) 3157 return; 3158 tcp_complete_cwr(sk); 3159 break; 3160 } 3161 } 3162 3163 /* E. Process state. */ 3164 switch (icsk->icsk_ca_state) { 3165 case TCP_CA_Recovery: 3166 if (!(flag & FLAG_SND_UNA_ADVANCED)) { 3167 if (tcp_is_reno(tp) && is_dupack) 3168 tcp_add_reno_sack(sk); 3169 } else 3170 do_lost = tcp_try_undo_partial(sk, pkts_acked); 3171 break; 3172 case TCP_CA_Loss: 3173 if (flag & FLAG_DATA_ACKED) 3174 icsk->icsk_retransmits = 0; 3175 if (tcp_is_reno(tp) && flag & FLAG_SND_UNA_ADVANCED) 3176 tcp_reset_reno_sack(tp); 3177 if (!tcp_try_undo_loss(sk)) { 3178 tcp_moderate_cwnd(tp); 3179 tcp_xmit_retransmit_queue(sk); 3180 return; 3181 } 3182 if (icsk->icsk_ca_state != TCP_CA_Open) 3183 return; 3184 /* Loss is undone; fall through to processing in Open state. */ 3185 default: 3186 if (tcp_is_reno(tp)) { 3187 if (flag & FLAG_SND_UNA_ADVANCED) 3188 tcp_reset_reno_sack(tp); 3189 if (is_dupack) 3190 tcp_add_reno_sack(sk); 3191 } 3192 3193 if (icsk->icsk_ca_state <= TCP_CA_Disorder) 3194 tcp_try_undo_dsack(sk); 3195 3196 if (!tcp_time_to_recover(sk, flag)) { 3197 tcp_try_to_open(sk, flag); 3198 return; 3199 } 3200 3201 /* MTU probe failure: don't reduce cwnd */ 3202 if (icsk->icsk_ca_state < TCP_CA_CWR && 3203 icsk->icsk_mtup.probe_size && 3204 tp->snd_una == tp->mtu_probe.probe_seq_start) { 3205 tcp_mtup_probe_failed(sk); 3206 /* Restores the reduction we did in tcp_mtup_probe() */ 3207 tp->snd_cwnd++; 3208 tcp_simple_retransmit(sk); 3209 return; 3210 } 3211 3212 /* Otherwise enter Recovery state */ 3213 tcp_enter_recovery(sk, (flag & FLAG_ECE)); 3214 fast_rexmit = 1; 3215 } 3216 3217 if (do_lost || (tcp_is_fack(tp) && tcp_head_timedout(sk))) 3218 tcp_update_scoreboard(sk, fast_rexmit); 3219 tp->prr_delivered += newly_acked_sacked; 3220 tcp_update_cwnd_in_recovery(sk, newly_acked_sacked, fast_rexmit, flag); 3221 tcp_xmit_retransmit_queue(sk); 3222} 3223 3224void tcp_valid_rtt_meas(struct sock *sk, u32 seq_rtt) 3225{ 3226 tcp_rtt_estimator(sk, seq_rtt); 3227 tcp_set_rto(sk); 3228 inet_csk(sk)->icsk_backoff = 0; 3229} 3230EXPORT_SYMBOL(tcp_valid_rtt_meas); 3231 3232/* Read draft-ietf-tcplw-high-performance before mucking 3233 * with this code. (Supersedes RFC1323) 3234 */ 3235static void tcp_ack_saw_tstamp(struct sock *sk, int flag) 3236{ 3237 /* RTTM Rule: A TSecr value received in a segment is used to 3238 * update the averaged RTT measurement only if the segment 3239 * acknowledges some new data, i.e., only if it advances the 3240 * left edge of the send window. 3241 * 3242 * See draft-ietf-tcplw-high-performance-00, section 3.3. 3243 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru> 3244 * 3245 * Changed: reset backoff as soon as we see the first valid sample. 3246 * If we do not, we get strongly overestimated rto. With timestamps 3247 * samples are accepted even from very old segments: f.e., when rtt=1 3248 * increases to 8, we retransmit 5 times and after 8 seconds delayed 3249 * answer arrives rto becomes 120 seconds! If at least one of segments 3250 * in window is lost... Voila. --ANK (010210) 3251 */ 3252 struct tcp_sock *tp = tcp_sk(sk); 3253 3254 tcp_valid_rtt_meas(sk, tcp_time_stamp - tp->rx_opt.rcv_tsecr); 3255} 3256 3257static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag) 3258{ 3259 /* We don't have a timestamp. Can only use 3260 * packets that are not retransmitted to determine 3261 * rtt estimates. Also, we must not reset the 3262 * backoff for rto until we get a non-retransmitted 3263 * packet. This allows us to deal with a situation 3264 * where the network delay has increased suddenly. 3265 * I.e. Karn's algorithm. (SIGCOMM '87, p5.) 3266 */ 3267 3268 if (flag & FLAG_RETRANS_DATA_ACKED) 3269 return; 3270 3271 tcp_valid_rtt_meas(sk, seq_rtt); 3272} 3273 3274static inline void tcp_ack_update_rtt(struct sock *sk, const int flag, 3275 const s32 seq_rtt) 3276{ 3277 const struct tcp_sock *tp = tcp_sk(sk); 3278 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */ 3279 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr) 3280 tcp_ack_saw_tstamp(sk, flag); 3281 else if (seq_rtt >= 0) 3282 tcp_ack_no_tstamp(sk, seq_rtt, flag); 3283} 3284 3285static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight) 3286{ 3287 const struct inet_connection_sock *icsk = inet_csk(sk); 3288 icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight); 3289 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp; 3290} 3291 3292/* Restart timer after forward progress on connection. 3293 * RFC2988 recommends to restart timer to now+rto. 3294 */ 3295void tcp_rearm_rto(struct sock *sk) 3296{ 3297 struct tcp_sock *tp = tcp_sk(sk); 3298 3299 if (!tp->packets_out) { 3300 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS); 3301 } else { 3302 u32 rto = inet_csk(sk)->icsk_rto; 3303 /* Offset the time elapsed after installing regular RTO */ 3304 if (tp->early_retrans_delayed) { 3305 struct sk_buff *skb = tcp_write_queue_head(sk); 3306 const u32 rto_time_stamp = TCP_SKB_CB(skb)->when + rto; 3307 s32 delta = (s32)(rto_time_stamp - tcp_time_stamp); 3308 /* delta may not be positive if the socket is locked 3309 * when the delayed ER timer fires and is rescheduled. 3310 */ 3311 if (delta > 0) 3312 rto = delta; 3313 } 3314 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto, 3315 TCP_RTO_MAX); 3316 } 3317 tp->early_retrans_delayed = 0; 3318} 3319 3320/* This function is called when the delayed ER timer fires. TCP enters 3321 * fast recovery and performs fast-retransmit. 3322 */ 3323void tcp_resume_early_retransmit(struct sock *sk) 3324{ 3325 struct tcp_sock *tp = tcp_sk(sk); 3326 3327 tcp_rearm_rto(sk); 3328 3329 /* Stop if ER is disabled after the delayed ER timer is scheduled */ 3330 if (!tp->do_early_retrans) 3331 return; 3332 3333 tcp_enter_recovery(sk, false); 3334 tcp_update_scoreboard(sk, 1); 3335 tcp_xmit_retransmit_queue(sk); 3336} 3337 3338/* If we get here, the whole TSO packet has not been acked. */ 3339static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb) 3340{ 3341 struct tcp_sock *tp = tcp_sk(sk); 3342 u32 packets_acked; 3343 3344 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)); 3345 3346 packets_acked = tcp_skb_pcount(skb); 3347 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 3348 return 0; 3349 packets_acked -= tcp_skb_pcount(skb); 3350 3351 if (packets_acked) { 3352 BUG_ON(tcp_skb_pcount(skb) == 0); 3353 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)); 3354 } 3355 3356 return packets_acked; 3357} 3358 3359/* Remove acknowledged frames from the retransmission queue. If our packet 3360 * is before the ack sequence we can discard it as it's confirmed to have 3361 * arrived at the other end. 3362 */ 3363static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets, 3364 u32 prior_snd_una) 3365{ 3366 struct tcp_sock *tp = tcp_sk(sk); 3367 const struct inet_connection_sock *icsk = inet_csk(sk); 3368 struct sk_buff *skb; 3369 u32 now = tcp_time_stamp; 3370 int fully_acked = 1; 3371 int flag = 0; 3372 u32 pkts_acked = 0; 3373 u32 reord = tp->packets_out; 3374 u32 prior_sacked = tp->sacked_out; 3375 s32 seq_rtt = -1; 3376 s32 ca_seq_rtt = -1; 3377 ktime_t last_ackt = net_invalid_timestamp(); 3378 3379 while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) { 3380 struct tcp_skb_cb *scb = TCP_SKB_CB(skb); 3381 u32 acked_pcount; 3382 u8 sacked = scb->sacked; 3383 3384 /* Determine how many packets and what bytes were acked, tso and else */ 3385 if (after(scb->end_seq, tp->snd_una)) { 3386 if (tcp_skb_pcount(skb) == 1 || 3387 !after(tp->snd_una, scb->seq)) 3388 break; 3389 3390 acked_pcount = tcp_tso_acked(sk, skb); 3391 if (!acked_pcount) 3392 break; 3393 3394 fully_acked = 0; 3395 } else { 3396 acked_pcount = tcp_skb_pcount(skb); 3397 } 3398 3399 if (sacked & TCPCB_RETRANS) { 3400 if (sacked & TCPCB_SACKED_RETRANS) 3401 tp->retrans_out -= acked_pcount; 3402 flag |= FLAG_RETRANS_DATA_ACKED; 3403 ca_seq_rtt = -1; 3404 seq_rtt = -1; 3405 if ((flag & FLAG_DATA_ACKED) || (acked_pcount > 1)) 3406 flag |= FLAG_NONHEAD_RETRANS_ACKED; 3407 } else { 3408 ca_seq_rtt = now - scb->when; 3409 last_ackt = skb->tstamp; 3410 if (seq_rtt < 0) { 3411 seq_rtt = ca_seq_rtt; 3412 } 3413 if (!(sacked & TCPCB_SACKED_ACKED)) 3414 reord = min(pkts_acked, reord); 3415 } 3416 3417 if (sacked & TCPCB_SACKED_ACKED) 3418 tp->sacked_out -= acked_pcount; 3419 if (sacked & TCPCB_LOST) 3420 tp->lost_out -= acked_pcount; 3421 3422 tp->packets_out -= acked_pcount; 3423 pkts_acked += acked_pcount; 3424 3425 /* Initial outgoing SYN's get put onto the write_queue 3426 * just like anything else we transmit. It is not 3427 * true data, and if we misinform our callers that 3428 * this ACK acks real data, we will erroneously exit 3429 * connection startup slow start one packet too 3430 * quickly. This is severely frowned upon behavior. 3431 */ 3432 if (!(scb->tcp_flags & TCPHDR_SYN)) { 3433 flag |= FLAG_DATA_ACKED; 3434 } else { 3435 flag |= FLAG_SYN_ACKED; 3436 tp->retrans_stamp = 0; 3437 } 3438 3439 if (!fully_acked) 3440 break; 3441 3442 tcp_unlink_write_queue(skb, sk); 3443 sk_wmem_free_skb(sk, skb); 3444 tp->scoreboard_skb_hint = NULL; 3445 if (skb == tp->retransmit_skb_hint) 3446 tp->retransmit_skb_hint = NULL; 3447 if (skb == tp->lost_skb_hint) 3448 tp->lost_skb_hint = NULL; 3449 } 3450 3451 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una))) 3452 tp->snd_up = tp->snd_una; 3453 3454 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) 3455 flag |= FLAG_SACK_RENEGING; 3456 3457 if (flag & FLAG_ACKED) { 3458 const struct tcp_congestion_ops *ca_ops 3459 = inet_csk(sk)->icsk_ca_ops; 3460 3461 if (unlikely(icsk->icsk_mtup.probe_size && 3462 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) { 3463 tcp_mtup_probe_success(sk); 3464 } 3465 3466 tcp_ack_update_rtt(sk, flag, seq_rtt); 3467 tcp_rearm_rto(sk); 3468 3469 if (tcp_is_reno(tp)) { 3470 tcp_remove_reno_sacks(sk, pkts_acked); 3471 } else { 3472 int delta; 3473 3474 /* Non-retransmitted hole got filled? That's reordering */ 3475 if (reord < prior_fackets) 3476 tcp_update_reordering(sk, tp->fackets_out - reord, 0); 3477 3478 delta = tcp_is_fack(tp) ? pkts_acked : 3479 prior_sacked - tp->sacked_out; 3480 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta); 3481 } 3482 3483 tp->fackets_out -= min(pkts_acked, tp->fackets_out); 3484 3485 if (ca_ops->pkts_acked) { 3486 s32 rtt_us = -1; 3487 3488 /* Is the ACK triggering packet unambiguous? */ 3489 if (!(flag & FLAG_RETRANS_DATA_ACKED)) { 3490 /* High resolution needed and available? */ 3491 if (ca_ops->flags & TCP_CONG_RTT_STAMP && 3492 !ktime_equal(last_ackt, 3493 net_invalid_timestamp())) 3494 rtt_us = ktime_us_delta(ktime_get_real(), 3495 last_ackt); 3496 else if (ca_seq_rtt >= 0) 3497 rtt_us = jiffies_to_usecs(ca_seq_rtt); 3498 } 3499 3500 ca_ops->pkts_acked(sk, pkts_acked, rtt_us); 3501 } 3502 } 3503 3504#if FASTRETRANS_DEBUG > 0 3505 WARN_ON((int)tp->sacked_out < 0); 3506 WARN_ON((int)tp->lost_out < 0); 3507 WARN_ON((int)tp->retrans_out < 0); 3508 if (!tp->packets_out && tcp_is_sack(tp)) { 3509 icsk = inet_csk(sk); 3510 if (tp->lost_out) { 3511 printk(KERN_DEBUG "Leak l=%u %d\n", 3512 tp->lost_out, icsk->icsk_ca_state); 3513 tp->lost_out = 0; 3514 } 3515 if (tp->sacked_out) { 3516 printk(KERN_DEBUG "Leak s=%u %d\n", 3517 tp->sacked_out, icsk->icsk_ca_state); 3518 tp->sacked_out = 0; 3519 } 3520 if (tp->retrans_out) { 3521 printk(KERN_DEBUG "Leak r=%u %d\n", 3522 tp->retrans_out, icsk->icsk_ca_state); 3523 tp->retrans_out = 0; 3524 } 3525 } 3526#endif 3527 return flag; 3528} 3529 3530static void tcp_ack_probe(struct sock *sk) 3531{ 3532 const struct tcp_sock *tp = tcp_sk(sk); 3533 struct inet_connection_sock *icsk = inet_csk(sk); 3534 3535 /* Was it a usable window open? */ 3536 3537 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) { 3538 icsk->icsk_backoff = 0; 3539 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0); 3540 /* Socket must be waked up by subsequent tcp_data_snd_check(). 3541 * This function is not for random using! 3542 */ 3543 } else { 3544 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 3545 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), 3546 TCP_RTO_MAX); 3547 } 3548} 3549 3550static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag) 3551{ 3552 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) || 3553 inet_csk(sk)->icsk_ca_state != TCP_CA_Open; 3554} 3555 3556static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag) 3557{ 3558 const struct tcp_sock *tp = tcp_sk(sk); 3559 return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) && 3560 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR)); 3561} 3562 3563/* Check that window update is acceptable. 3564 * The function assumes that snd_una<=ack<=snd_next. 3565 */ 3566static inline int tcp_may_update_window(const struct tcp_sock *tp, 3567 const u32 ack, const u32 ack_seq, 3568 const u32 nwin) 3569{ 3570 return after(ack, tp->snd_una) || 3571 after(ack_seq, tp->snd_wl1) || 3572 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd); 3573} 3574 3575/* Update our send window. 3576 * 3577 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2 3578 * and in FreeBSD. NetBSD's one is even worse.) is wrong. 3579 */ 3580static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack, 3581 u32 ack_seq) 3582{ 3583 struct tcp_sock *tp = tcp_sk(sk); 3584 int flag = 0; 3585 u32 nwin = ntohs(tcp_hdr(skb)->window); 3586 3587 if (likely(!tcp_hdr(skb)->syn)) 3588 nwin <<= tp->rx_opt.snd_wscale; 3589 3590 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) { 3591 flag |= FLAG_WIN_UPDATE; 3592 tcp_update_wl(tp, ack_seq); 3593 3594 if (tp->snd_wnd != nwin) { 3595 tp->snd_wnd = nwin; 3596 3597 /* Note, it is the only place, where 3598 * fast path is recovered for sending TCP. 3599 */ 3600 tp->pred_flags = 0; 3601 tcp_fast_path_check(sk); 3602 3603 if (nwin > tp->max_window) { 3604 tp->max_window = nwin; 3605 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie); 3606 } 3607 } 3608 } 3609 3610 tp->snd_una = ack; 3611 3612 return flag; 3613} 3614 3615/* A very conservative spurious RTO response algorithm: reduce cwnd and 3616 * continue in congestion avoidance. 3617 */ 3618static void tcp_conservative_spur_to_response(struct tcp_sock *tp) 3619{ 3620 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh); 3621 tp->snd_cwnd_cnt = 0; 3622 tp->bytes_acked = 0; 3623 TCP_ECN_queue_cwr(tp); 3624 tcp_moderate_cwnd(tp); 3625} 3626 3627/* A conservative spurious RTO response algorithm: reduce cwnd using 3628 * rate halving and continue in congestion avoidance. 3629 */ 3630static void tcp_ratehalving_spur_to_response(struct sock *sk) 3631{ 3632 tcp_enter_cwr(sk, 0); 3633} 3634 3635static void tcp_undo_spur_to_response(struct sock *sk, int flag) 3636{ 3637 if (flag & FLAG_ECE) 3638 tcp_ratehalving_spur_to_response(sk); 3639 else 3640 tcp_undo_cwr(sk, true); 3641} 3642 3643/* F-RTO spurious RTO detection algorithm (RFC4138) 3644 * 3645 * F-RTO affects during two new ACKs following RTO (well, almost, see inline 3646 * comments). State (ACK number) is kept in frto_counter. When ACK advances 3647 * window (but not to or beyond highest sequence sent before RTO): 3648 * On First ACK, send two new segments out. 3649 * On Second ACK, RTO was likely spurious. Do spurious response (response 3650 * algorithm is not part of the F-RTO detection algorithm 3651 * given in RFC4138 but can be selected separately). 3652 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss 3653 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding 3654 * of Nagle, this is done using frto_counter states 2 and 3, when a new data 3655 * segment of any size sent during F-RTO, state 2 is upgraded to 3. 3656 * 3657 * Rationale: if the RTO was spurious, new ACKs should arrive from the 3658 * original window even after we transmit two new data segments. 3659 * 3660 * SACK version: 3661 * on first step, wait until first cumulative ACK arrives, then move to 3662 * the second step. In second step, the next ACK decides. 3663 * 3664 * F-RTO is implemented (mainly) in four functions: 3665 * - tcp_use_frto() is used to determine if TCP is can use F-RTO 3666 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is 3667 * called when tcp_use_frto() showed green light 3668 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm 3669 * - tcp_enter_frto_loss() is called if there is not enough evidence 3670 * to prove that the RTO is indeed spurious. It transfers the control 3671 * from F-RTO to the conventional RTO recovery 3672 */ 3673static int tcp_process_frto(struct sock *sk, int flag) 3674{ 3675 struct tcp_sock *tp = tcp_sk(sk); 3676 3677 tcp_verify_left_out(tp); 3678 3679 /* Duplicate the behavior from Loss state (fastretrans_alert) */ 3680 if (flag & FLAG_DATA_ACKED) 3681 inet_csk(sk)->icsk_retransmits = 0; 3682 3683 if ((flag & FLAG_NONHEAD_RETRANS_ACKED) || 3684 ((tp->frto_counter >= 2) && (flag & FLAG_RETRANS_DATA_ACKED))) 3685 tp->undo_marker = 0; 3686 3687 if (!before(tp->snd_una, tp->frto_highmark)) { 3688 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag); 3689 return 1; 3690 } 3691 3692 if (!tcp_is_sackfrto(tp)) { 3693 /* RFC4138 shortcoming in step 2; should also have case c): 3694 * ACK isn't duplicate nor advances window, e.g., opposite dir 3695 * data, winupdate 3696 */ 3697 if (!(flag & FLAG_ANY_PROGRESS) && (flag & FLAG_NOT_DUP)) 3698 return 1; 3699 3700 if (!(flag & FLAG_DATA_ACKED)) { 3701 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3), 3702 flag); 3703 return 1; 3704 } 3705 } else { 3706 if (!(flag & FLAG_DATA_ACKED) && (tp->frto_counter == 1)) { 3707 /* Prevent sending of new data. */ 3708 tp->snd_cwnd = min(tp->snd_cwnd, 3709 tcp_packets_in_flight(tp)); 3710 return 1; 3711 } 3712 3713 if ((tp->frto_counter >= 2) && 3714 (!(flag & FLAG_FORWARD_PROGRESS) || 3715 ((flag & FLAG_DATA_SACKED) && 3716 !(flag & FLAG_ONLY_ORIG_SACKED)))) { 3717 /* RFC4138 shortcoming (see comment above) */ 3718 if (!(flag & FLAG_FORWARD_PROGRESS) && 3719 (flag & FLAG_NOT_DUP)) 3720 return 1; 3721 3722 tcp_enter_frto_loss(sk, 3, flag); 3723 return 1; 3724 } 3725 } 3726 3727 if (tp->frto_counter == 1) { 3728 /* tcp_may_send_now needs to see updated state */ 3729 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2; 3730 tp->frto_counter = 2; 3731 3732 if (!tcp_may_send_now(sk)) 3733 tcp_enter_frto_loss(sk, 2, flag); 3734 3735 return 1; 3736 } else { 3737 switch (sysctl_tcp_frto_response) { 3738 case 2: 3739 tcp_undo_spur_to_response(sk, flag); 3740 break; 3741 case 1: 3742 tcp_conservative_spur_to_response(tp); 3743 break; 3744 default: 3745 tcp_ratehalving_spur_to_response(sk); 3746 break; 3747 } 3748 tp->frto_counter = 0; 3749 tp->undo_marker = 0; 3750 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSPURIOUSRTOS); 3751 } 3752 return 0; 3753} 3754 3755/* This routine deals with incoming acks, but not outgoing ones. */ 3756static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag) 3757{ 3758 struct inet_connection_sock *icsk = inet_csk(sk); 3759 struct tcp_sock *tp = tcp_sk(sk); 3760 u32 prior_snd_una = tp->snd_una; 3761 u32 ack_seq = TCP_SKB_CB(skb)->seq; 3762 u32 ack = TCP_SKB_CB(skb)->ack_seq; 3763 bool is_dupack = false; 3764 u32 prior_in_flight; 3765 u32 prior_fackets; 3766 int prior_packets; 3767 int prior_sacked = tp->sacked_out; 3768 int pkts_acked = 0; 3769 int newly_acked_sacked = 0; 3770 int frto_cwnd = 0; 3771 3772 /* If the ack is older than previous acks 3773 * then we can probably ignore it. 3774 */ 3775 if (before(ack, prior_snd_una)) 3776 goto old_ack; 3777 3778 /* If the ack includes data we haven't sent yet, discard 3779 * this segment (RFC793 Section 3.9). 3780 */ 3781 if (after(ack, tp->snd_nxt)) 3782 goto invalid_ack; 3783 3784 if (tp->early_retrans_delayed) 3785 tcp_rearm_rto(sk); 3786 3787 if (after(ack, prior_snd_una)) 3788 flag |= FLAG_SND_UNA_ADVANCED; 3789 3790 if (sysctl_tcp_abc) { 3791 if (icsk->icsk_ca_state < TCP_CA_CWR) 3792 tp->bytes_acked += ack - prior_snd_una; 3793 else if (icsk->icsk_ca_state == TCP_CA_Loss) 3794 /* we assume just one segment left network */ 3795 tp->bytes_acked += min(ack - prior_snd_una, 3796 tp->mss_cache); 3797 } 3798 3799 prior_fackets = tp->fackets_out; 3800 prior_in_flight = tcp_packets_in_flight(tp); 3801 3802 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) { 3803 /* Window is constant, pure forward advance. 3804 * No more checks are required. 3805 * Note, we use the fact that SND.UNA>=SND.WL2. 3806 */ 3807 tcp_update_wl(tp, ack_seq); 3808 tp->snd_una = ack; 3809 flag |= FLAG_WIN_UPDATE; 3810 3811 tcp_ca_event(sk, CA_EVENT_FAST_ACK); 3812 3813 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS); 3814 } else { 3815 if (ack_seq != TCP_SKB_CB(skb)->end_seq) 3816 flag |= FLAG_DATA; 3817 else 3818 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS); 3819 3820 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq); 3821 3822 if (TCP_SKB_CB(skb)->sacked) 3823 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una); 3824 3825 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb))) 3826 flag |= FLAG_ECE; 3827 3828 tcp_ca_event(sk, CA_EVENT_SLOW_ACK); 3829 } 3830 3831 /* We passed data and got it acked, remove any soft error 3832 * log. Something worked... 3833 */ 3834 sk->sk_err_soft = 0; 3835 icsk->icsk_probes_out = 0; 3836 tp->rcv_tstamp = tcp_time_stamp; 3837 prior_packets = tp->packets_out; 3838 if (!prior_packets) 3839 goto no_queue; 3840 3841 /* See if we can take anything off of the retransmit queue. */ 3842 flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una); 3843 3844 pkts_acked = prior_packets - tp->packets_out; 3845 newly_acked_sacked = (prior_packets - prior_sacked) - 3846 (tp->packets_out - tp->sacked_out); 3847 3848 if (tp->frto_counter) 3849 frto_cwnd = tcp_process_frto(sk, flag); 3850 /* Guarantee sacktag reordering detection against wrap-arounds */ 3851 if (before(tp->frto_highmark, tp->snd_una)) 3852 tp->frto_highmark = 0; 3853 3854 if (tcp_ack_is_dubious(sk, flag)) { 3855 /* Advance CWND, if state allows this. */ 3856 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd && 3857 tcp_may_raise_cwnd(sk, flag)) 3858 tcp_cong_avoid(sk, ack, prior_in_flight); 3859 is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP)); 3860 tcp_fastretrans_alert(sk, pkts_acked, newly_acked_sacked, 3861 is_dupack, flag); 3862 } else { 3863 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd) 3864 tcp_cong_avoid(sk, ack, prior_in_flight); 3865 } 3866 3867 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) 3868 dst_confirm(__sk_dst_get(sk)); 3869 3870 return 1; 3871 3872no_queue: 3873 /* If data was DSACKed, see if we can undo a cwnd reduction. */ 3874 if (flag & FLAG_DSACKING_ACK) 3875 tcp_fastretrans_alert(sk, pkts_acked, newly_acked_sacked, 3876 is_dupack, flag); 3877 /* If this ack opens up a zero window, clear backoff. It was 3878 * being used to time the probes, and is probably far higher than 3879 * it needs to be for normal retransmission. 3880 */ 3881 if (tcp_send_head(sk)) 3882 tcp_ack_probe(sk); 3883 return 1; 3884 3885invalid_ack: 3886 SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt); 3887 return -1; 3888 3889old_ack: 3890 /* If data was SACKed, tag it and see if we should send more data. 3891 * If data was DSACKed, see if we can undo a cwnd reduction. 3892 */ 3893 if (TCP_SKB_CB(skb)->sacked) { 3894 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una); 3895 newly_acked_sacked = tp->sacked_out - prior_sacked; 3896 tcp_fastretrans_alert(sk, pkts_acked, newly_acked_sacked, 3897 is_dupack, flag); 3898 } 3899 3900 SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt); 3901 return 0; 3902} 3903 3904/* Look for tcp options. Normally only called on SYN and SYNACK packets. 3905 * But, this can also be called on packets in the established flow when 3906 * the fast version below fails. 3907 */ 3908void tcp_parse_options(const struct sk_buff *skb, struct tcp_options_received *opt_rx, 3909 const u8 **hvpp, int estab) 3910{ 3911 const unsigned char *ptr; 3912 const struct tcphdr *th = tcp_hdr(skb); 3913 int length = (th->doff * 4) - sizeof(struct tcphdr); 3914 3915 ptr = (const unsigned char *)(th + 1); 3916 opt_rx->saw_tstamp = 0; 3917 3918 while (length > 0) { 3919 int opcode = *ptr++; 3920 int opsize; 3921 3922 switch (opcode) { 3923 case TCPOPT_EOL: 3924 return; 3925 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */ 3926 length--; 3927 continue; 3928 default: 3929 opsize = *ptr++; 3930 if (opsize < 2) /* "silly options" */ 3931 return; 3932 if (opsize > length) 3933 return; /* don't parse partial options */ 3934 switch (opcode) { 3935 case TCPOPT_MSS: 3936 if (opsize == TCPOLEN_MSS && th->syn && !estab) { 3937 u16 in_mss = get_unaligned_be16(ptr); 3938 if (in_mss) { 3939 if (opt_rx->user_mss && 3940 opt_rx->user_mss < in_mss) 3941 in_mss = opt_rx->user_mss; 3942 opt_rx->mss_clamp = in_mss; 3943 } 3944 } 3945 break; 3946 case TCPOPT_WINDOW: 3947 if (opsize == TCPOLEN_WINDOW && th->syn && 3948 !estab && sysctl_tcp_window_scaling) { 3949 __u8 snd_wscale = *(__u8 *)ptr; 3950 opt_rx->wscale_ok = 1; 3951 if (snd_wscale > 14) { 3952 if (net_ratelimit()) 3953 pr_info("%s: Illegal window scaling value %d >14 received\n", 3954 __func__, 3955 snd_wscale); 3956 snd_wscale = 14; 3957 } 3958 opt_rx->snd_wscale = snd_wscale; 3959 } 3960 break; 3961 case TCPOPT_TIMESTAMP: 3962 if ((opsize == TCPOLEN_TIMESTAMP) && 3963 ((estab && opt_rx->tstamp_ok) || 3964 (!estab && sysctl_tcp_timestamps))) { 3965 opt_rx->saw_tstamp = 1; 3966 opt_rx->rcv_tsval = get_unaligned_be32(ptr); 3967 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4); 3968 } 3969 break; 3970 case TCPOPT_SACK_PERM: 3971 if (opsize == TCPOLEN_SACK_PERM && th->syn && 3972 !estab && sysctl_tcp_sack) { 3973 opt_rx->sack_ok = TCP_SACK_SEEN; 3974 tcp_sack_reset(opt_rx); 3975 } 3976 break; 3977 3978 case TCPOPT_SACK: 3979 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) && 3980 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) && 3981 opt_rx->sack_ok) { 3982 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th; 3983 } 3984 break; 3985#ifdef CONFIG_TCP_MD5SIG 3986 case TCPOPT_MD5SIG: 3987 /* 3988 * The MD5 Hash has already been 3989 * checked (see tcp_v{4,6}_do_rcv()). 3990 */ 3991 break; 3992#endif 3993 case TCPOPT_COOKIE: 3994 /* This option is variable length. 3995 */ 3996 switch (opsize) { 3997 case TCPOLEN_COOKIE_BASE: 3998 /* not yet implemented */ 3999 break; 4000 case TCPOLEN_COOKIE_PAIR: 4001 /* not yet implemented */ 4002 break; 4003 case TCPOLEN_COOKIE_MIN+0: 4004 case TCPOLEN_COOKIE_MIN+2: 4005 case TCPOLEN_COOKIE_MIN+4: 4006 case TCPOLEN_COOKIE_MIN+6: 4007 case TCPOLEN_COOKIE_MAX: 4008 /* 16-bit multiple */ 4009 opt_rx->cookie_plus = opsize; 4010 *hvpp = ptr; 4011 break; 4012 default: 4013 /* ignore option */ 4014 break; 4015 } 4016 break; 4017 } 4018 4019 ptr += opsize-2; 4020 length -= opsize; 4021 } 4022 } 4023} 4024EXPORT_SYMBOL(tcp_parse_options); 4025 4026static int tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th) 4027{ 4028 const __be32 *ptr = (const __be32 *)(th + 1); 4029 4030 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) 4031 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) { 4032 tp->rx_opt.saw_tstamp = 1; 4033 ++ptr; 4034 tp->rx_opt.rcv_tsval = ntohl(*ptr); 4035 ++ptr; 4036 tp->rx_opt.rcv_tsecr = ntohl(*ptr); 4037 return 1; 4038 } 4039 return 0; 4040} 4041 4042/* Fast parse options. This hopes to only see timestamps. 4043 * If it is wrong it falls back on tcp_parse_options(). 4044 */ 4045static int tcp_fast_parse_options(const struct sk_buff *skb, 4046 const struct tcphdr *th, 4047 struct tcp_sock *tp, const u8 **hvpp) 4048{ 4049 /* In the spirit of fast parsing, compare doff directly to constant 4050 * values. Because equality is used, short doff can be ignored here. 4051 */ 4052 if (th->doff == (sizeof(*th) / 4)) { 4053 tp->rx_opt.saw_tstamp = 0; 4054 return 0; 4055 } else if (tp->rx_opt.tstamp_ok && 4056 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) { 4057 if (tcp_parse_aligned_timestamp(tp, th)) 4058 return 1; 4059 } 4060 tcp_parse_options(skb, &tp->rx_opt, hvpp, 1); 4061 return 1; 4062} 4063 4064#ifdef CONFIG_TCP_MD5SIG 4065/* 4066 * Parse MD5 Signature option 4067 */ 4068const u8 *tcp_parse_md5sig_option(const struct tcphdr *th) 4069{ 4070 int length = (th->doff << 2) - sizeof(*th); 4071 const u8 *ptr = (const u8 *)(th + 1); 4072 4073 /* If the TCP option is too short, we can short cut */ 4074 if (length < TCPOLEN_MD5SIG) 4075 return NULL; 4076 4077 while (length > 0) { 4078 int opcode = *ptr++; 4079 int opsize; 4080 4081 switch(opcode) { 4082 case TCPOPT_EOL: 4083 return NULL; 4084 case TCPOPT_NOP: 4085 length--; 4086 continue; 4087 default: 4088 opsize = *ptr++; 4089 if (opsize < 2 || opsize > length) 4090 return NULL; 4091 if (opcode == TCPOPT_MD5SIG) 4092 return opsize == TCPOLEN_MD5SIG ? ptr : NULL; 4093 } 4094 ptr += opsize - 2; 4095 length -= opsize; 4096 } 4097 return NULL; 4098} 4099EXPORT_SYMBOL(tcp_parse_md5sig_option); 4100#endif 4101 4102static inline void tcp_store_ts_recent(struct tcp_sock *tp) 4103{ 4104 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval; 4105 tp->rx_opt.ts_recent_stamp = get_seconds(); 4106} 4107 4108static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq) 4109{ 4110 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) { 4111 /* PAWS bug workaround wrt. ACK frames, the PAWS discard 4112 * extra check below makes sure this can only happen 4113 * for pure ACK frames. -DaveM 4114 * 4115 * Not only, also it occurs for expired timestamps. 4116 */ 4117 4118 if (tcp_paws_check(&tp->rx_opt, 0)) 4119 tcp_store_ts_recent(tp); 4120 } 4121} 4122 4123/* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM 4124 * 4125 * It is not fatal. If this ACK does _not_ change critical state (seqs, window) 4126 * it can pass through stack. So, the following predicate verifies that 4127 * this segment is not used for anything but congestion avoidance or 4128 * fast retransmit. Moreover, we even are able to eliminate most of such 4129 * second order effects, if we apply some small "replay" window (~RTO) 4130 * to timestamp space. 4131 * 4132 * All these measures still do not guarantee that we reject wrapped ACKs 4133 * on networks with high bandwidth, when sequence space is recycled fastly, 4134 * but it guarantees that such events will be very rare and do not affect 4135 * connection seriously. This doesn't look nice, but alas, PAWS is really 4136 * buggy extension. 4137 * 4138 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC 4139 * states that events when retransmit arrives after original data are rare. 4140 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is 4141 * the biggest problem on large power networks even with minor reordering. 4142 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe 4143 * up to bandwidth of 18Gigabit/sec. 8) ] 4144 */ 4145 4146static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb) 4147{ 4148 const struct tcp_sock *tp = tcp_sk(sk); 4149 const struct tcphdr *th = tcp_hdr(skb); 4150 u32 seq = TCP_SKB_CB(skb)->seq; 4151 u32 ack = TCP_SKB_CB(skb)->ack_seq; 4152 4153 return (/* 1. Pure ACK with correct sequence number. */ 4154 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) && 4155 4156 /* 2. ... and duplicate ACK. */ 4157 ack == tp->snd_una && 4158 4159 /* 3. ... and does not update window. */ 4160 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) && 4161 4162 /* 4. ... and sits in replay window. */ 4163 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ); 4164} 4165 4166static inline int tcp_paws_discard(const struct sock *sk, 4167 const struct sk_buff *skb) 4168{ 4169 const struct tcp_sock *tp = tcp_sk(sk); 4170 4171 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) && 4172 !tcp_disordered_ack(sk, skb); 4173} 4174 4175/* Check segment sequence number for validity. 4176 * 4177 * Segment controls are considered valid, if the segment 4178 * fits to the window after truncation to the window. Acceptability 4179 * of data (and SYN, FIN, of course) is checked separately. 4180 * See tcp_data_queue(), for example. 4181 * 4182 * Also, controls (RST is main one) are accepted using RCV.WUP instead 4183 * of RCV.NXT. Peer still did not advance his SND.UNA when we 4184 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP. 4185 * (borrowed from freebsd) 4186 */ 4187 4188static inline int tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq) 4189{ 4190 return !before(end_seq, tp->rcv_wup) && 4191 !after(seq, tp->rcv_nxt + tcp_receive_window(tp)); 4192} 4193 4194/* When we get a reset we do this. */ 4195static void tcp_reset(struct sock *sk) 4196{ 4197 /* We want the right error as BSD sees it (and indeed as we do). */ 4198 switch (sk->sk_state) { 4199 case TCP_SYN_SENT: 4200 sk->sk_err = ECONNREFUSED; 4201 break; 4202 case TCP_CLOSE_WAIT: 4203 sk->sk_err = EPIPE; 4204 break; 4205 case TCP_CLOSE: 4206 return; 4207 default: 4208 sk->sk_err = ECONNRESET; 4209 } 4210 /* This barrier is coupled with smp_rmb() in tcp_poll() */ 4211 smp_wmb(); 4212 4213 if (!sock_flag(sk, SOCK_DEAD)) 4214 sk->sk_error_report(sk); 4215 4216 tcp_done(sk); 4217} 4218 4219/* 4220 * Process the FIN bit. This now behaves as it is supposed to work 4221 * and the FIN takes effect when it is validly part of sequence 4222 * space. Not before when we get holes. 4223 * 4224 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT 4225 * (and thence onto LAST-ACK and finally, CLOSE, we never enter 4226 * TIME-WAIT) 4227 * 4228 * If we are in FINWAIT-1, a received FIN indicates simultaneous 4229 * close and we go into CLOSING (and later onto TIME-WAIT) 4230 * 4231 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT. 4232 */ 4233static void tcp_fin(struct sock *sk) 4234{ 4235 struct tcp_sock *tp = tcp_sk(sk); 4236 4237 inet_csk_schedule_ack(sk); 4238 4239 sk->sk_shutdown |= RCV_SHUTDOWN; 4240 sock_set_flag(sk, SOCK_DONE); 4241 4242 switch (sk->sk_state) { 4243 case TCP_SYN_RECV: 4244 case TCP_ESTABLISHED: 4245 /* Move to CLOSE_WAIT */ 4246 tcp_set_state(sk, TCP_CLOSE_WAIT); 4247 inet_csk(sk)->icsk_ack.pingpong = 1; 4248 break; 4249 4250 case TCP_CLOSE_WAIT: 4251 case TCP_CLOSING: 4252 /* Received a retransmission of the FIN, do 4253 * nothing. 4254 */ 4255 break; 4256 case TCP_LAST_ACK: 4257 /* RFC793: Remain in the LAST-ACK state. */ 4258 break; 4259 4260 case TCP_FIN_WAIT1: 4261 /* This case occurs when a simultaneous close 4262 * happens, we must ack the received FIN and 4263 * enter the CLOSING state. 4264 */ 4265 tcp_send_ack(sk); 4266 tcp_set_state(sk, TCP_CLOSING); 4267 break; 4268 case TCP_FIN_WAIT2: 4269 /* Received a FIN -- send ACK and enter TIME_WAIT. */ 4270 tcp_send_ack(sk); 4271 tcp_time_wait(sk, TCP_TIME_WAIT, 0); 4272 break; 4273 default: 4274 /* Only TCP_LISTEN and TCP_CLOSE are left, in these 4275 * cases we should never reach this piece of code. 4276 */ 4277 pr_err("%s: Impossible, sk->sk_state=%d\n", 4278 __func__, sk->sk_state); 4279 break; 4280 } 4281 4282 /* It _is_ possible, that we have something out-of-order _after_ FIN. 4283 * Probably, we should reset in this case. For now drop them. 4284 */ 4285 __skb_queue_purge(&tp->out_of_order_queue); 4286 if (tcp_is_sack(tp)) 4287 tcp_sack_reset(&tp->rx_opt); 4288 sk_mem_reclaim(sk); 4289 4290 if (!sock_flag(sk, SOCK_DEAD)) { 4291 sk->sk_state_change(sk); 4292 4293 /* Do not send POLL_HUP for half duplex close. */ 4294 if (sk->sk_shutdown == SHUTDOWN_MASK || 4295 sk->sk_state == TCP_CLOSE) 4296 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP); 4297 else 4298 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN); 4299 } 4300} 4301 4302static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, 4303 u32 end_seq) 4304{ 4305 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) { 4306 if (before(seq, sp->start_seq)) 4307 sp->start_seq = seq; 4308 if (after(end_seq, sp->end_seq)) 4309 sp->end_seq = end_seq; 4310 return 1; 4311 } 4312 return 0; 4313} 4314 4315static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq) 4316{ 4317 struct tcp_sock *tp = tcp_sk(sk); 4318 4319 if (tcp_is_sack(tp) && sysctl_tcp_dsack) { 4320 int mib_idx; 4321 4322 if (before(seq, tp->rcv_nxt)) 4323 mib_idx = LINUX_MIB_TCPDSACKOLDSENT; 4324 else 4325 mib_idx = LINUX_MIB_TCPDSACKOFOSENT; 4326 4327 NET_INC_STATS_BH(sock_net(sk), mib_idx); 4328 4329 tp->rx_opt.dsack = 1; 4330 tp->duplicate_sack[0].start_seq = seq; 4331 tp->duplicate_sack[0].end_seq = end_seq; 4332 } 4333} 4334 4335static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq) 4336{ 4337 struct tcp_sock *tp = tcp_sk(sk); 4338 4339 if (!tp->rx_opt.dsack) 4340 tcp_dsack_set(sk, seq, end_seq); 4341 else 4342 tcp_sack_extend(tp->duplicate_sack, seq, end_seq); 4343} 4344 4345static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb) 4346{ 4347 struct tcp_sock *tp = tcp_sk(sk); 4348 4349 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 4350 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 4351 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST); 4352 tcp_enter_quickack_mode(sk); 4353 4354 if (tcp_is_sack(tp) && sysctl_tcp_dsack) { 4355 u32 end_seq = TCP_SKB_CB(skb)->end_seq; 4356 4357 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) 4358 end_seq = tp->rcv_nxt; 4359 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq); 4360 } 4361 } 4362 4363 tcp_send_ack(sk); 4364} 4365 4366/* These routines update the SACK block as out-of-order packets arrive or 4367 * in-order packets close up the sequence space. 4368 */ 4369static void tcp_sack_maybe_coalesce(struct tcp_sock *tp) 4370{ 4371 int this_sack; 4372 struct tcp_sack_block *sp = &tp->selective_acks[0]; 4373 struct tcp_sack_block *swalk = sp + 1; 4374 4375 /* See if the recent change to the first SACK eats into 4376 * or hits the sequence space of other SACK blocks, if so coalesce. 4377 */ 4378 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) { 4379 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) { 4380 int i; 4381 4382 /* Zap SWALK, by moving every further SACK up by one slot. 4383 * Decrease num_sacks. 4384 */ 4385 tp->rx_opt.num_sacks--; 4386 for (i = this_sack; i < tp->rx_opt.num_sacks; i++) 4387 sp[i] = sp[i + 1]; 4388 continue; 4389 } 4390 this_sack++, swalk++; 4391 } 4392} 4393 4394static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq) 4395{ 4396 struct tcp_sock *tp = tcp_sk(sk); 4397 struct tcp_sack_block *sp = &tp->selective_acks[0]; 4398 int cur_sacks = tp->rx_opt.num_sacks; 4399 int this_sack; 4400 4401 if (!cur_sacks) 4402 goto new_sack; 4403 4404 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) { 4405 if (tcp_sack_extend(sp, seq, end_seq)) { 4406 /* Rotate this_sack to the first one. */ 4407 for (; this_sack > 0; this_sack--, sp--) 4408 swap(*sp, *(sp - 1)); 4409 if (cur_sacks > 1) 4410 tcp_sack_maybe_coalesce(tp); 4411 return; 4412 } 4413 } 4414 4415 /* Could not find an adjacent existing SACK, build a new one, 4416 * put it at the front, and shift everyone else down. We 4417 * always know there is at least one SACK present already here. 4418 * 4419 * If the sack array is full, forget about the last one. 4420 */ 4421 if (this_sack >= TCP_NUM_SACKS) { 4422 this_sack--; 4423 tp->rx_opt.num_sacks--; 4424 sp--; 4425 } 4426 for (; this_sack > 0; this_sack--, sp--) 4427 *sp = *(sp - 1); 4428 4429new_sack: 4430 /* Build the new head SACK, and we're done. */ 4431 sp->start_seq = seq; 4432 sp->end_seq = end_seq; 4433 tp->rx_opt.num_sacks++; 4434} 4435 4436/* RCV.NXT advances, some SACKs should be eaten. */ 4437 4438static void tcp_sack_remove(struct tcp_sock *tp) 4439{ 4440 struct tcp_sack_block *sp = &tp->selective_acks[0]; 4441 int num_sacks = tp->rx_opt.num_sacks; 4442 int this_sack; 4443 4444 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */ 4445 if (skb_queue_empty(&tp->out_of_order_queue)) { 4446 tp->rx_opt.num_sacks = 0; 4447 return; 4448 } 4449 4450 for (this_sack = 0; this_sack < num_sacks;) { 4451 /* Check if the start of the sack is covered by RCV.NXT. */ 4452 if (!before(tp->rcv_nxt, sp->start_seq)) { 4453 int i; 4454 4455 /* RCV.NXT must cover all the block! */ 4456 WARN_ON(before(tp->rcv_nxt, sp->end_seq)); 4457 4458 /* Zap this SACK, by moving forward any other SACKS. */ 4459 for (i=this_sack+1; i < num_sacks; i++) 4460 tp->selective_acks[i-1] = tp->selective_acks[i]; 4461 num_sacks--; 4462 continue; 4463 } 4464 this_sack++; 4465 sp++; 4466 } 4467 tp->rx_opt.num_sacks = num_sacks; 4468} 4469 4470/* This one checks to see if we can put data from the 4471 * out_of_order queue into the receive_queue. 4472 */ 4473static void tcp_ofo_queue(struct sock *sk) 4474{ 4475 struct tcp_sock *tp = tcp_sk(sk); 4476 __u32 dsack_high = tp->rcv_nxt; 4477 struct sk_buff *skb; 4478 4479 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) { 4480 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) 4481 break; 4482 4483 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) { 4484 __u32 dsack = dsack_high; 4485 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high)) 4486 dsack_high = TCP_SKB_CB(skb)->end_seq; 4487 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack); 4488 } 4489 4490 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) { 4491 SOCK_DEBUG(sk, "ofo packet was already received\n"); 4492 __skb_unlink(skb, &tp->out_of_order_queue); 4493 __kfree_skb(skb); 4494 continue; 4495 } 4496 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n", 4497 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, 4498 TCP_SKB_CB(skb)->end_seq); 4499 4500 __skb_unlink(skb, &tp->out_of_order_queue); 4501 __skb_queue_tail(&sk->sk_receive_queue, skb); 4502 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; 4503 if (tcp_hdr(skb)->fin) 4504 tcp_fin(sk); 4505 } 4506} 4507 4508static int tcp_prune_ofo_queue(struct sock *sk); 4509static int tcp_prune_queue(struct sock *sk); 4510 4511static inline int tcp_try_rmem_schedule(struct sock *sk, unsigned int size) 4512{ 4513 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf || 4514 !sk_rmem_schedule(sk, size)) { 4515 4516 if (tcp_prune_queue(sk) < 0) 4517 return -1; 4518 4519 if (!sk_rmem_schedule(sk, size)) { 4520 if (!tcp_prune_ofo_queue(sk)) 4521 return -1; 4522 4523 if (!sk_rmem_schedule(sk, size)) 4524 return -1; 4525 } 4526 } 4527 return 0; 4528} 4529 4530/** 4531 * tcp_try_coalesce - try to merge skb to prior one 4532 * @sk: socket 4533 * @to: prior buffer 4534 * @from: buffer to add in queue 4535 * @fragstolen: pointer to boolean 4536 * 4537 * Before queueing skb @from after @to, try to merge them 4538 * to reduce overall memory use and queue lengths, if cost is small. 4539 * Packets in ofo or receive queues can stay a long time. 4540 * Better try to coalesce them right now to avoid future collapses. 4541 * Returns true if caller should free @from instead of queueing it 4542 */ 4543static bool tcp_try_coalesce(struct sock *sk, 4544 struct sk_buff *to, 4545 struct sk_buff *from, 4546 bool *fragstolen) 4547{ 4548 int i, delta, len = from->len; 4549 4550 *fragstolen = false; 4551 if (tcp_hdr(from)->fin || skb_cloned(to)) 4552 return false; 4553 if (len <= skb_tailroom(to)) { 4554 BUG_ON(skb_copy_bits(from, 0, skb_put(to, len), len)); 4555merge: 4556 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE); 4557 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq; 4558 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq; 4559 return true; 4560 } 4561 4562 if (skb_has_frag_list(to) || skb_has_frag_list(from)) 4563 return false; 4564 4565 if (skb_headlen(from) == 0 && 4566 (skb_shinfo(to)->nr_frags + 4567 skb_shinfo(from)->nr_frags <= MAX_SKB_FRAGS)) { 4568 WARN_ON_ONCE(from->head_frag); 4569 delta = from->truesize - ksize(from->head) - 4570 SKB_DATA_ALIGN(sizeof(struct sk_buff)); 4571 4572 WARN_ON_ONCE(delta < len); 4573copyfrags: 4574 memcpy(skb_shinfo(to)->frags + skb_shinfo(to)->nr_frags, 4575 skb_shinfo(from)->frags, 4576 skb_shinfo(from)->nr_frags * sizeof(skb_frag_t)); 4577 skb_shinfo(to)->nr_frags += skb_shinfo(from)->nr_frags; 4578 4579 if (skb_cloned(from)) 4580 for (i = 0; i < skb_shinfo(from)->nr_frags; i++) 4581 skb_frag_ref(from, i); 4582 else 4583 skb_shinfo(from)->nr_frags = 0; 4584 4585 to->truesize += delta; 4586 atomic_add(delta, &sk->sk_rmem_alloc); 4587 sk_mem_charge(sk, delta); 4588 to->len += len; 4589 to->data_len += len; 4590 goto merge; 4591 } 4592 if (from->head_frag) { 4593 struct page *page; 4594 unsigned int offset; 4595 4596 if (skb_shinfo(to)->nr_frags + skb_shinfo(from)->nr_frags >= MAX_SKB_FRAGS) 4597 return false; 4598 page = virt_to_head_page(from->head); 4599 offset = from->data - (unsigned char *)page_address(page); 4600 skb_fill_page_desc(to, skb_shinfo(to)->nr_frags, 4601 page, offset, skb_headlen(from)); 4602 4603 if (skb_cloned(from)) 4604 get_page(page); 4605 else 4606 *fragstolen = true; 4607 4608 delta = len; /* we dont know real truesize... */ 4609 goto copyfrags; 4610 } 4611 return false; 4612} 4613 4614static void kfree_skb_partial(struct sk_buff *skb, bool head_stolen) 4615{ 4616 if (head_stolen) 4617 kmem_cache_free(skbuff_head_cache, skb); 4618 else 4619 __kfree_skb(skb); 4620} 4621 4622static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb) 4623{ 4624 struct tcp_sock *tp = tcp_sk(sk); 4625 struct sk_buff *skb1; 4626 u32 seq, end_seq; 4627 4628 TCP_ECN_check_ce(tp, skb); 4629 4630 if (tcp_try_rmem_schedule(sk, skb->truesize)) { 4631 /* TODO: should increment a counter */ 4632 __kfree_skb(skb); 4633 return; 4634 } 4635 4636 /* Disable header prediction. */ 4637 tp->pred_flags = 0; 4638 inet_csk_schedule_ack(sk); 4639 4640 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n", 4641 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); 4642 4643 skb1 = skb_peek_tail(&tp->out_of_order_queue); 4644 if (!skb1) { 4645 /* Initial out of order segment, build 1 SACK. */ 4646 if (tcp_is_sack(tp)) { 4647 tp->rx_opt.num_sacks = 1; 4648 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq; 4649 tp->selective_acks[0].end_seq = 4650 TCP_SKB_CB(skb)->end_seq; 4651 } 4652 __skb_queue_head(&tp->out_of_order_queue, skb); 4653 goto end; 4654 } 4655 4656 seq = TCP_SKB_CB(skb)->seq; 4657 end_seq = TCP_SKB_CB(skb)->end_seq; 4658 4659 if (seq == TCP_SKB_CB(skb1)->end_seq) { 4660 bool fragstolen; 4661 4662 if (!tcp_try_coalesce(sk, skb1, skb, &fragstolen)) { 4663 __skb_queue_after(&tp->out_of_order_queue, skb1, skb); 4664 } else { 4665 kfree_skb_partial(skb, fragstolen); 4666 skb = NULL; 4667 } 4668 4669 if (!tp->rx_opt.num_sacks || 4670 tp->selective_acks[0].end_seq != seq) 4671 goto add_sack; 4672 4673 /* Common case: data arrive in order after hole. */ 4674 tp->selective_acks[0].end_seq = end_seq; 4675 goto end; 4676 } 4677 4678 /* Find place to insert this segment. */ 4679 while (1) { 4680 if (!after(TCP_SKB_CB(skb1)->seq, seq)) 4681 break; 4682 if (skb_queue_is_first(&tp->out_of_order_queue, skb1)) { 4683 skb1 = NULL; 4684 break; 4685 } 4686 skb1 = skb_queue_prev(&tp->out_of_order_queue, skb1); 4687 } 4688 4689 /* Do skb overlap to previous one? */ 4690 if (skb1 && before(seq, TCP_SKB_CB(skb1)->end_seq)) { 4691 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) { 4692 /* All the bits are present. Drop. */ 4693 __kfree_skb(skb); 4694 skb = NULL; 4695 tcp_dsack_set(sk, seq, end_seq); 4696 goto add_sack; 4697 } 4698 if (after(seq, TCP_SKB_CB(skb1)->seq)) { 4699 /* Partial overlap. */ 4700 tcp_dsack_set(sk, seq, 4701 TCP_SKB_CB(skb1)->end_seq); 4702 } else { 4703 if (skb_queue_is_first(&tp->out_of_order_queue, 4704 skb1)) 4705 skb1 = NULL; 4706 else 4707 skb1 = skb_queue_prev( 4708 &tp->out_of_order_queue, 4709 skb1); 4710 } 4711 } 4712 if (!skb1) 4713 __skb_queue_head(&tp->out_of_order_queue, skb); 4714 else 4715 __skb_queue_after(&tp->out_of_order_queue, skb1, skb); 4716 4717 /* And clean segments covered by new one as whole. */ 4718 while (!skb_queue_is_last(&tp->out_of_order_queue, skb)) { 4719 skb1 = skb_queue_next(&tp->out_of_order_queue, skb); 4720 4721 if (!after(end_seq, TCP_SKB_CB(skb1)->seq)) 4722 break; 4723 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) { 4724 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq, 4725 end_seq); 4726 break; 4727 } 4728 __skb_unlink(skb1, &tp->out_of_order_queue); 4729 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq, 4730 TCP_SKB_CB(skb1)->end_seq); 4731 __kfree_skb(skb1); 4732 } 4733 4734add_sack: 4735 if (tcp_is_sack(tp)) 4736 tcp_sack_new_ofo_skb(sk, seq, end_seq); 4737end: 4738 if (skb) 4739 skb_set_owner_r(skb, sk); 4740} 4741 4742 4743static void tcp_data_queue(struct sock *sk, struct sk_buff *skb) 4744{ 4745 const struct tcphdr *th = tcp_hdr(skb); 4746 struct tcp_sock *tp = tcp_sk(sk); 4747 int eaten = -1; 4748 bool fragstolen = false; 4749 4750 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) 4751 goto drop; 4752 4753 skb_dst_drop(skb); 4754 __skb_pull(skb, th->doff * 4); 4755 4756 TCP_ECN_accept_cwr(tp, skb); 4757 4758 tp->rx_opt.dsack = 0; 4759 4760 /* Queue data for delivery to the user. 4761 * Packets in sequence go to the receive queue. 4762 * Out of sequence packets to the out_of_order_queue. 4763 */ 4764 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) { 4765 if (tcp_receive_window(tp) == 0) 4766 goto out_of_window; 4767 4768 /* Ok. In sequence. In window. */ 4769 if (tp->ucopy.task == current && 4770 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len && 4771 sock_owned_by_user(sk) && !tp->urg_data) { 4772 int chunk = min_t(unsigned int, skb->len, 4773 tp->ucopy.len); 4774 4775 __set_current_state(TASK_RUNNING); 4776 4777 local_bh_enable(); 4778 if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) { 4779 tp->ucopy.len -= chunk; 4780 tp->copied_seq += chunk; 4781 eaten = (chunk == skb->len); 4782 tcp_rcv_space_adjust(sk); 4783 } 4784 local_bh_disable(); 4785 } 4786 4787 if (eaten <= 0) { 4788 struct sk_buff *tail; 4789queue_and_out: 4790 if (eaten < 0 && 4791 tcp_try_rmem_schedule(sk, skb->truesize)) 4792 goto drop; 4793 4794 tail = skb_peek_tail(&sk->sk_receive_queue); 4795 eaten = (tail && 4796 tcp_try_coalesce(sk, tail, skb, 4797 &fragstolen)) ? 1 : 0; 4798 if (eaten <= 0) { 4799 skb_set_owner_r(skb, sk); 4800 __skb_queue_tail(&sk->sk_receive_queue, skb); 4801 } 4802 } 4803 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; 4804 if (skb->len) 4805 tcp_event_data_recv(sk, skb); 4806 if (th->fin) 4807 tcp_fin(sk); 4808 4809 if (!skb_queue_empty(&tp->out_of_order_queue)) { 4810 tcp_ofo_queue(sk); 4811 4812 /* RFC2581. 4.2. SHOULD send immediate ACK, when 4813 * gap in queue is filled. 4814 */ 4815 if (skb_queue_empty(&tp->out_of_order_queue)) 4816 inet_csk(sk)->icsk_ack.pingpong = 0; 4817 } 4818 4819 if (tp->rx_opt.num_sacks) 4820 tcp_sack_remove(tp); 4821 4822 tcp_fast_path_check(sk); 4823 4824 if (eaten > 0) 4825 kfree_skb_partial(skb, fragstolen); 4826 else if (!sock_flag(sk, SOCK_DEAD)) 4827 sk->sk_data_ready(sk, 0); 4828 return; 4829 } 4830 4831 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) { 4832 /* A retransmit, 2nd most common case. Force an immediate ack. */ 4833 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST); 4834 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); 4835 4836out_of_window: 4837 tcp_enter_quickack_mode(sk); 4838 inet_csk_schedule_ack(sk); 4839drop: 4840 __kfree_skb(skb); 4841 return; 4842 } 4843 4844 /* Out of window. F.e. zero window probe. */ 4845 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp))) 4846 goto out_of_window; 4847 4848 tcp_enter_quickack_mode(sk); 4849 4850 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 4851 /* Partial packet, seq < rcv_next < end_seq */ 4852 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n", 4853 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, 4854 TCP_SKB_CB(skb)->end_seq); 4855 4856 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt); 4857 4858 /* If window is closed, drop tail of packet. But after 4859 * remembering D-SACK for its head made in previous line. 4860 */ 4861 if (!tcp_receive_window(tp)) 4862 goto out_of_window; 4863 goto queue_and_out; 4864 } 4865 4866 tcp_data_queue_ofo(sk, skb); 4867} 4868 4869static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb, 4870 struct sk_buff_head *list) 4871{ 4872 struct sk_buff *next = NULL; 4873 4874 if (!skb_queue_is_last(list, skb)) 4875 next = skb_queue_next(list, skb); 4876 4877 __skb_unlink(skb, list); 4878 __kfree_skb(skb); 4879 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED); 4880 4881 return next; 4882} 4883 4884/* Collapse contiguous sequence of skbs head..tail with 4885 * sequence numbers start..end. 4886 * 4887 * If tail is NULL, this means until the end of the list. 4888 * 4889 * Segments with FIN/SYN are not collapsed (only because this 4890 * simplifies code) 4891 */ 4892static void 4893tcp_collapse(struct sock *sk, struct sk_buff_head *list, 4894 struct sk_buff *head, struct sk_buff *tail, 4895 u32 start, u32 end) 4896{ 4897 struct sk_buff *skb, *n; 4898 bool end_of_skbs; 4899 4900 /* First, check that queue is collapsible and find 4901 * the point where collapsing can be useful. */ 4902 skb = head; 4903restart: 4904 end_of_skbs = true; 4905 skb_queue_walk_from_safe(list, skb, n) { 4906 if (skb == tail) 4907 break; 4908 /* No new bits? It is possible on ofo queue. */ 4909 if (!before(start, TCP_SKB_CB(skb)->end_seq)) { 4910 skb = tcp_collapse_one(sk, skb, list); 4911 if (!skb) 4912 break; 4913 goto restart; 4914 } 4915 4916 /* The first skb to collapse is: 4917 * - not SYN/FIN and 4918 * - bloated or contains data before "start" or 4919 * overlaps to the next one. 4920 */ 4921 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin && 4922 (tcp_win_from_space(skb->truesize) > skb->len || 4923 before(TCP_SKB_CB(skb)->seq, start))) { 4924 end_of_skbs = false; 4925 break; 4926 } 4927 4928 if (!skb_queue_is_last(list, skb)) { 4929 struct sk_buff *next = skb_queue_next(list, skb); 4930 if (next != tail && 4931 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(next)->seq) { 4932 end_of_skbs = false; 4933 break; 4934 } 4935 } 4936 4937 /* Decided to skip this, advance start seq. */ 4938 start = TCP_SKB_CB(skb)->end_seq; 4939 } 4940 if (end_of_skbs || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin) 4941 return; 4942 4943 while (before(start, end)) { 4944 struct sk_buff *nskb; 4945 unsigned int header = skb_headroom(skb); 4946 int copy = SKB_MAX_ORDER(header, 0); 4947 4948 /* Too big header? This can happen with IPv6. */ 4949 if (copy < 0) 4950 return; 4951 if (end - start < copy) 4952 copy = end - start; 4953 nskb = alloc_skb(copy + header, GFP_ATOMIC); 4954 if (!nskb) 4955 return; 4956 4957 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head); 4958 skb_set_network_header(nskb, (skb_network_header(skb) - 4959 skb->head)); 4960 skb_set_transport_header(nskb, (skb_transport_header(skb) - 4961 skb->head)); 4962 skb_reserve(nskb, header); 4963 memcpy(nskb->head, skb->head, header); 4964 memcpy(nskb->cb, skb->cb, sizeof(skb->cb)); 4965 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start; 4966 __skb_queue_before(list, skb, nskb); 4967 skb_set_owner_r(nskb, sk); 4968 4969 /* Copy data, releasing collapsed skbs. */ 4970 while (copy > 0) { 4971 int offset = start - TCP_SKB_CB(skb)->seq; 4972 int size = TCP_SKB_CB(skb)->end_seq - start; 4973 4974 BUG_ON(offset < 0); 4975 if (size > 0) { 4976 size = min(copy, size); 4977 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size)) 4978 BUG(); 4979 TCP_SKB_CB(nskb)->end_seq += size; 4980 copy -= size; 4981 start += size; 4982 } 4983 if (!before(start, TCP_SKB_CB(skb)->end_seq)) { 4984 skb = tcp_collapse_one(sk, skb, list); 4985 if (!skb || 4986 skb == tail || 4987 tcp_hdr(skb)->syn || 4988 tcp_hdr(skb)->fin) 4989 return; 4990 } 4991 } 4992 } 4993} 4994 4995/* Collapse ofo queue. Algorithm: select contiguous sequence of skbs 4996 * and tcp_collapse() them until all the queue is collapsed. 4997 */ 4998static void tcp_collapse_ofo_queue(struct sock *sk) 4999{ 5000 struct tcp_sock *tp = tcp_sk(sk); 5001 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue); 5002 struct sk_buff *head; 5003 u32 start, end; 5004 5005 if (skb == NULL) 5006 return; 5007 5008 start = TCP_SKB_CB(skb)->seq; 5009 end = TCP_SKB_CB(skb)->end_seq; 5010 head = skb; 5011 5012 for (;;) { 5013 struct sk_buff *next = NULL; 5014 5015 if (!skb_queue_is_last(&tp->out_of_order_queue, skb)) 5016 next = skb_queue_next(&tp->out_of_order_queue, skb); 5017 skb = next; 5018 5019 /* Segment is terminated when we see gap or when 5020 * we are at the end of all the queue. */ 5021 if (!skb || 5022 after(TCP_SKB_CB(skb)->seq, end) || 5023 before(TCP_SKB_CB(skb)->end_seq, start)) { 5024 tcp_collapse(sk, &tp->out_of_order_queue, 5025 head, skb, start, end); 5026 head = skb; 5027 if (!skb) 5028 break; 5029 /* Start new segment */ 5030 start = TCP_SKB_CB(skb)->seq; 5031 end = TCP_SKB_CB(skb)->end_seq; 5032 } else { 5033 if (before(TCP_SKB_CB(skb)->seq, start)) 5034 start = TCP_SKB_CB(skb)->seq; 5035 if (after(TCP_SKB_CB(skb)->end_seq, end)) 5036 end = TCP_SKB_CB(skb)->end_seq; 5037 } 5038 } 5039} 5040 5041/* 5042 * Purge the out-of-order queue. 5043 * Return true if queue was pruned. 5044 */ 5045static int tcp_prune_ofo_queue(struct sock *sk) 5046{ 5047 struct tcp_sock *tp = tcp_sk(sk); 5048 int res = 0; 5049 5050 if (!skb_queue_empty(&tp->out_of_order_queue)) { 5051 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED); 5052 __skb_queue_purge(&tp->out_of_order_queue); 5053 5054 /* Reset SACK state. A conforming SACK implementation will 5055 * do the same at a timeout based retransmit. When a connection 5056 * is in a sad state like this, we care only about integrity 5057 * of the connection not performance. 5058 */ 5059 if (tp->rx_opt.sack_ok) 5060 tcp_sack_reset(&tp->rx_opt); 5061 sk_mem_reclaim(sk); 5062 res = 1; 5063 } 5064 return res; 5065} 5066 5067/* Reduce allocated memory if we can, trying to get 5068 * the socket within its memory limits again. 5069 * 5070 * Return less than zero if we should start dropping frames 5071 * until the socket owning process reads some of the data 5072 * to stabilize the situation. 5073 */ 5074static int tcp_prune_queue(struct sock *sk) 5075{ 5076 struct tcp_sock *tp = tcp_sk(sk); 5077 5078 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq); 5079 5080 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED); 5081 5082 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf) 5083 tcp_clamp_window(sk); 5084 else if (sk_under_memory_pressure(sk)) 5085 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss); 5086 5087 tcp_collapse_ofo_queue(sk); 5088 if (!skb_queue_empty(&sk->sk_receive_queue)) 5089 tcp_collapse(sk, &sk->sk_receive_queue, 5090 skb_peek(&sk->sk_receive_queue), 5091 NULL, 5092 tp->copied_seq, tp->rcv_nxt); 5093 sk_mem_reclaim(sk); 5094 5095 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) 5096 return 0; 5097 5098 /* Collapsing did not help, destructive actions follow. 5099 * This must not ever occur. */ 5100 5101 tcp_prune_ofo_queue(sk); 5102 5103 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) 5104 return 0; 5105 5106 /* If we are really being abused, tell the caller to silently 5107 * drop receive data on the floor. It will get retransmitted 5108 * and hopefully then we'll have sufficient space. 5109 */ 5110 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED); 5111 5112 /* Massive buffer overcommit. */ 5113 tp->pred_flags = 0; 5114 return -1; 5115} 5116 5117/* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. 5118 * As additional protections, we do not touch cwnd in retransmission phases, 5119 * and if application hit its sndbuf limit recently. 5120 */ 5121void tcp_cwnd_application_limited(struct sock *sk) 5122{ 5123 struct tcp_sock *tp = tcp_sk(sk); 5124 5125 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && 5126 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { 5127 /* Limited by application or receiver window. */ 5128 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); 5129 u32 win_used = max(tp->snd_cwnd_used, init_win); 5130 if (win_used < tp->snd_cwnd) { 5131 tp->snd_ssthresh = tcp_current_ssthresh(sk); 5132 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1; 5133 } 5134 tp->snd_cwnd_used = 0; 5135 } 5136 tp->snd_cwnd_stamp = tcp_time_stamp; 5137} 5138 5139static int tcp_should_expand_sndbuf(const struct sock *sk) 5140{ 5141 const struct tcp_sock *tp = tcp_sk(sk); 5142 5143 /* If the user specified a specific send buffer setting, do 5144 * not modify it. 5145 */ 5146 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK) 5147 return 0; 5148 5149 /* If we are under global TCP memory pressure, do not expand. */ 5150 if (sk_under_memory_pressure(sk)) 5151 return 0; 5152 5153 /* If we are under soft global TCP memory pressure, do not expand. */ 5154 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0)) 5155 return 0; 5156 5157 /* If we filled the congestion window, do not expand. */ 5158 if (tp->packets_out >= tp->snd_cwnd) 5159 return 0; 5160 5161 return 1; 5162} 5163 5164/* When incoming ACK allowed to free some skb from write_queue, 5165 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket 5166 * on the exit from tcp input handler. 5167 * 5168 * PROBLEM: sndbuf expansion does not work well with largesend. 5169 */ 5170static void tcp_new_space(struct sock *sk) 5171{ 5172 struct tcp_sock *tp = tcp_sk(sk); 5173 5174 if (tcp_should_expand_sndbuf(sk)) { 5175 int sndmem = SKB_TRUESIZE(max_t(u32, 5176 tp->rx_opt.mss_clamp, 5177 tp->mss_cache) + 5178 MAX_TCP_HEADER); 5179 int demanded = max_t(unsigned int, tp->snd_cwnd, 5180 tp->reordering + 1); 5181 sndmem *= 2 * demanded; 5182 if (sndmem > sk->sk_sndbuf) 5183 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]); 5184 tp->snd_cwnd_stamp = tcp_time_stamp; 5185 } 5186 5187 sk->sk_write_space(sk); 5188} 5189 5190static void tcp_check_space(struct sock *sk) 5191{ 5192 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) { 5193 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK); 5194 if (sk->sk_socket && 5195 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) 5196 tcp_new_space(sk); 5197 } 5198} 5199 5200static inline void tcp_data_snd_check(struct sock *sk) 5201{ 5202 tcp_push_pending_frames(sk); 5203 tcp_check_space(sk); 5204} 5205 5206/* 5207 * Check if sending an ack is needed. 5208 */ 5209static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible) 5210{ 5211 struct tcp_sock *tp = tcp_sk(sk); 5212 5213 /* More than one full frame received... */ 5214 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss && 5215 /* ... and right edge of window advances far enough. 5216 * (tcp_recvmsg() will send ACK otherwise). Or... 5217 */ 5218 __tcp_select_window(sk) >= tp->rcv_wnd) || 5219 /* We ACK each frame or... */ 5220 tcp_in_quickack_mode(sk) || 5221 /* We have out of order data. */ 5222 (ofo_possible && skb_peek(&tp->out_of_order_queue))) { 5223 /* Then ack it now */ 5224 tcp_send_ack(sk); 5225 } else { 5226 /* Else, send delayed ack. */ 5227 tcp_send_delayed_ack(sk); 5228 } 5229} 5230 5231static inline void tcp_ack_snd_check(struct sock *sk) 5232{ 5233 if (!inet_csk_ack_scheduled(sk)) { 5234 /* We sent a data segment already. */ 5235 return; 5236 } 5237 __tcp_ack_snd_check(sk, 1); 5238} 5239 5240/* 5241 * This routine is only called when we have urgent data 5242 * signaled. Its the 'slow' part of tcp_urg. It could be 5243 * moved inline now as tcp_urg is only called from one 5244 * place. We handle URGent data wrong. We have to - as 5245 * BSD still doesn't use the correction from RFC961. 5246 * For 1003.1g we should support a new option TCP_STDURG to permit 5247 * either form (or just set the sysctl tcp_stdurg). 5248 */ 5249 5250static void tcp_check_urg(struct sock *sk, const struct tcphdr *th) 5251{ 5252 struct tcp_sock *tp = tcp_sk(sk); 5253 u32 ptr = ntohs(th->urg_ptr); 5254 5255 if (ptr && !sysctl_tcp_stdurg) 5256 ptr--; 5257 ptr += ntohl(th->seq); 5258 5259 /* Ignore urgent data that we've already seen and read. */ 5260 if (after(tp->copied_seq, ptr)) 5261 return; 5262 5263 /* Do not replay urg ptr. 5264 * 5265 * NOTE: interesting situation not covered by specs. 5266 * Misbehaving sender may send urg ptr, pointing to segment, 5267 * which we already have in ofo queue. We are not able to fetch 5268 * such data and will stay in TCP_URG_NOTYET until will be eaten 5269 * by recvmsg(). Seems, we are not obliged to handle such wicked 5270 * situations. But it is worth to think about possibility of some 5271 * DoSes using some hypothetical application level deadlock. 5272 */ 5273 if (before(ptr, tp->rcv_nxt)) 5274 return; 5275 5276 /* Do we already have a newer (or duplicate) urgent pointer? */ 5277 if (tp->urg_data && !after(ptr, tp->urg_seq)) 5278 return; 5279 5280 /* Tell the world about our new urgent pointer. */ 5281 sk_send_sigurg(sk); 5282 5283 /* We may be adding urgent data when the last byte read was 5284 * urgent. To do this requires some care. We cannot just ignore 5285 * tp->copied_seq since we would read the last urgent byte again 5286 * as data, nor can we alter copied_seq until this data arrives 5287 * or we break the semantics of SIOCATMARK (and thus sockatmark()) 5288 * 5289 * NOTE. Double Dutch. Rendering to plain English: author of comment 5290 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB); 5291 * and expect that both A and B disappear from stream. This is _wrong_. 5292 * Though this happens in BSD with high probability, this is occasional. 5293 * Any application relying on this is buggy. Note also, that fix "works" 5294 * only in this artificial test. Insert some normal data between A and B and we will 5295 * decline of BSD again. Verdict: it is better to remove to trap 5296 * buggy users. 5297 */ 5298 if (tp->urg_seq == tp->copied_seq && tp->urg_data && 5299 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) { 5300 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue); 5301 tp->copied_seq++; 5302 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) { 5303 __skb_unlink(skb, &sk->sk_receive_queue); 5304 __kfree_skb(skb); 5305 } 5306 } 5307 5308 tp->urg_data = TCP_URG_NOTYET; 5309 tp->urg_seq = ptr; 5310 5311 /* Disable header prediction. */ 5312 tp->pred_flags = 0; 5313} 5314 5315/* This is the 'fast' part of urgent handling. */ 5316static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th) 5317{ 5318 struct tcp_sock *tp = tcp_sk(sk); 5319 5320 /* Check if we get a new urgent pointer - normally not. */ 5321 if (th->urg) 5322 tcp_check_urg(sk, th); 5323 5324 /* Do we wait for any urgent data? - normally not... */ 5325 if (tp->urg_data == TCP_URG_NOTYET) { 5326 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) - 5327 th->syn; 5328 5329 /* Is the urgent pointer pointing into this packet? */ 5330 if (ptr < skb->len) { 5331 u8 tmp; 5332 if (skb_copy_bits(skb, ptr, &tmp, 1)) 5333 BUG(); 5334 tp->urg_data = TCP_URG_VALID | tmp; 5335 if (!sock_flag(sk, SOCK_DEAD)) 5336 sk->sk_data_ready(sk, 0); 5337 } 5338 } 5339} 5340 5341static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen) 5342{ 5343 struct tcp_sock *tp = tcp_sk(sk); 5344 int chunk = skb->len - hlen; 5345 int err; 5346 5347 local_bh_enable(); 5348 if (skb_csum_unnecessary(skb)) 5349 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk); 5350 else 5351 err = skb_copy_and_csum_datagram_iovec(skb, hlen, 5352 tp->ucopy.iov); 5353 5354 if (!err) { 5355 tp->ucopy.len -= chunk; 5356 tp->copied_seq += chunk; 5357 tcp_rcv_space_adjust(sk); 5358 } 5359 5360 local_bh_disable(); 5361 return err; 5362} 5363 5364static __sum16 __tcp_checksum_complete_user(struct sock *sk, 5365 struct sk_buff *skb) 5366{ 5367 __sum16 result; 5368 5369 if (sock_owned_by_user(sk)) { 5370 local_bh_enable(); 5371 result = __tcp_checksum_complete(skb); 5372 local_bh_disable(); 5373 } else { 5374 result = __tcp_checksum_complete(skb); 5375 } 5376 return result; 5377} 5378 5379static inline int tcp_checksum_complete_user(struct sock *sk, 5380 struct sk_buff *skb) 5381{ 5382 return !skb_csum_unnecessary(skb) && 5383 __tcp_checksum_complete_user(sk, skb); 5384} 5385 5386#ifdef CONFIG_NET_DMA 5387static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb, 5388 int hlen) 5389{ 5390 struct tcp_sock *tp = tcp_sk(sk); 5391 int chunk = skb->len - hlen; 5392 int dma_cookie; 5393 int copied_early = 0; 5394 5395 if (tp->ucopy.wakeup) 5396 return 0; 5397 5398 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list) 5399 tp->ucopy.dma_chan = net_dma_find_channel(); 5400 5401 if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) { 5402 5403 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan, 5404 skb, hlen, 5405 tp->ucopy.iov, chunk, 5406 tp->ucopy.pinned_list); 5407 5408 if (dma_cookie < 0) 5409 goto out; 5410 5411 tp->ucopy.dma_cookie = dma_cookie; 5412 copied_early = 1; 5413 5414 tp->ucopy.len -= chunk; 5415 tp->copied_seq += chunk; 5416 tcp_rcv_space_adjust(sk); 5417 5418 if ((tp->ucopy.len == 0) || 5419 (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) || 5420 (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) { 5421 tp->ucopy.wakeup = 1; 5422 sk->sk_data_ready(sk, 0); 5423 } 5424 } else if (chunk > 0) { 5425 tp->ucopy.wakeup = 1; 5426 sk->sk_data_ready(sk, 0); 5427 } 5428out: 5429 return copied_early; 5430} 5431#endif /* CONFIG_NET_DMA */ 5432 5433/* Does PAWS and seqno based validation of an incoming segment, flags will 5434 * play significant role here. 5435 */ 5436static int tcp_validate_incoming(struct sock *sk, struct sk_buff *skb, 5437 const struct tcphdr *th, int syn_inerr) 5438{ 5439 const u8 *hash_location; 5440 struct tcp_sock *tp = tcp_sk(sk); 5441 5442 /* RFC1323: H1. Apply PAWS check first. */ 5443 if (tcp_fast_parse_options(skb, th, tp, &hash_location) && 5444 tp->rx_opt.saw_tstamp && 5445 tcp_paws_discard(sk, skb)) { 5446 if (!th->rst) { 5447 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED); 5448 tcp_send_dupack(sk, skb); 5449 goto discard; 5450 } 5451 /* Reset is accepted even if it did not pass PAWS. */ 5452 } 5453 5454 /* Step 1: check sequence number */ 5455 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) { 5456 /* RFC793, page 37: "In all states except SYN-SENT, all reset 5457 * (RST) segments are validated by checking their SEQ-fields." 5458 * And page 69: "If an incoming segment is not acceptable, 5459 * an acknowledgment should be sent in reply (unless the RST 5460 * bit is set, if so drop the segment and return)". 5461 */ 5462 if (!th->rst) 5463 tcp_send_dupack(sk, skb); 5464 goto discard; 5465 } 5466 5467 /* Step 2: check RST bit */ 5468 if (th->rst) { 5469 tcp_reset(sk); 5470 goto discard; 5471 } 5472 5473 /* ts_recent update must be made after we are sure that the packet 5474 * is in window. 5475 */ 5476 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq); 5477 5478 /* step 3: check security and precedence [ignored] */ 5479 5480 /* step 4: Check for a SYN in window. */ 5481 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 5482 if (syn_inerr) 5483 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS); 5484 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONSYN); 5485 tcp_reset(sk); 5486 return -1; 5487 } 5488 5489 return 1; 5490 5491discard: 5492 __kfree_skb(skb); 5493 return 0; 5494} 5495 5496void tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, int hdrlen) 5497{ 5498 __skb_pull(skb, hdrlen); 5499 __skb_queue_tail(&sk->sk_receive_queue, skb); 5500 skb_set_owner_r(skb, sk); 5501 tcp_sk(sk)->rcv_nxt = TCP_SKB_CB(skb)->end_seq; 5502} 5503 5504/* 5505 * TCP receive function for the ESTABLISHED state. 5506 * 5507 * It is split into a fast path and a slow path. The fast path is 5508 * disabled when: 5509 * - A zero window was announced from us - zero window probing 5510 * is only handled properly in the slow path. 5511 * - Out of order segments arrived. 5512 * - Urgent data is expected. 5513 * - There is no buffer space left 5514 * - Unexpected TCP flags/window values/header lengths are received 5515 * (detected by checking the TCP header against pred_flags) 5516 * - Data is sent in both directions. Fast path only supports pure senders 5517 * or pure receivers (this means either the sequence number or the ack 5518 * value must stay constant) 5519 * - Unexpected TCP option. 5520 * 5521 * When these conditions are not satisfied it drops into a standard 5522 * receive procedure patterned after RFC793 to handle all cases. 5523 * The first three cases are guaranteed by proper pred_flags setting, 5524 * the rest is checked inline. Fast processing is turned on in 5525 * tcp_data_queue when everything is OK. 5526 */ 5527int tcp_rcv_established(struct sock *sk, struct sk_buff *skb, 5528 const struct tcphdr *th, unsigned int len) 5529{ 5530 struct tcp_sock *tp = tcp_sk(sk); 5531 int res; 5532 5533 /* 5534 * Header prediction. 5535 * The code loosely follows the one in the famous 5536 * "30 instruction TCP receive" Van Jacobson mail. 5537 * 5538 * Van's trick is to deposit buffers into socket queue 5539 * on a device interrupt, to call tcp_recv function 5540 * on the receive process context and checksum and copy 5541 * the buffer to user space. smart... 5542 * 5543 * Our current scheme is not silly either but we take the 5544 * extra cost of the net_bh soft interrupt processing... 5545 * We do checksum and copy also but from device to kernel. 5546 */ 5547 5548 tp->rx_opt.saw_tstamp = 0; 5549 5550 /* pred_flags is 0xS?10 << 16 + snd_wnd 5551 * if header_prediction is to be made 5552 * 'S' will always be tp->tcp_header_len >> 2 5553 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to 5554 * turn it off (when there are holes in the receive 5555 * space for instance) 5556 * PSH flag is ignored. 5557 */ 5558 5559 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags && 5560 TCP_SKB_CB(skb)->seq == tp->rcv_nxt && 5561 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) { 5562 int tcp_header_len = tp->tcp_header_len; 5563 5564 /* Timestamp header prediction: tcp_header_len 5565 * is automatically equal to th->doff*4 due to pred_flags 5566 * match. 5567 */ 5568 5569 /* Check timestamp */ 5570 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) { 5571 /* No? Slow path! */ 5572 if (!tcp_parse_aligned_timestamp(tp, th)) 5573 goto slow_path; 5574 5575 /* If PAWS failed, check it more carefully in slow path */ 5576 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0) 5577 goto slow_path; 5578 5579 /* DO NOT update ts_recent here, if checksum fails 5580 * and timestamp was corrupted part, it will result 5581 * in a hung connection since we will drop all 5582 * future packets due to the PAWS test. 5583 */ 5584 } 5585 5586 if (len <= tcp_header_len) { 5587 /* Bulk data transfer: sender */ 5588 if (len == tcp_header_len) { 5589 /* Predicted packet is in window by definition. 5590 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 5591 * Hence, check seq<=rcv_wup reduces to: 5592 */ 5593 if (tcp_header_len == 5594 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && 5595 tp->rcv_nxt == tp->rcv_wup) 5596 tcp_store_ts_recent(tp); 5597 5598 /* We know that such packets are checksummed 5599 * on entry. 5600 */ 5601 tcp_ack(sk, skb, 0); 5602 __kfree_skb(skb); 5603 tcp_data_snd_check(sk); 5604 return 0; 5605 } else { /* Header too small */ 5606 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS); 5607 goto discard; 5608 } 5609 } else { 5610 int eaten = 0; 5611 int copied_early = 0; 5612 5613 if (tp->copied_seq == tp->rcv_nxt && 5614 len - tcp_header_len <= tp->ucopy.len) { 5615#ifdef CONFIG_NET_DMA 5616 if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) { 5617 copied_early = 1; 5618 eaten = 1; 5619 } 5620#endif 5621 if (tp->ucopy.task == current && 5622 sock_owned_by_user(sk) && !copied_early) { 5623 __set_current_state(TASK_RUNNING); 5624 5625 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len)) 5626 eaten = 1; 5627 } 5628 if (eaten) { 5629 /* Predicted packet is in window by definition. 5630 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 5631 * Hence, check seq<=rcv_wup reduces to: 5632 */ 5633 if (tcp_header_len == 5634 (sizeof(struct tcphdr) + 5635 TCPOLEN_TSTAMP_ALIGNED) && 5636 tp->rcv_nxt == tp->rcv_wup) 5637 tcp_store_ts_recent(tp); 5638 5639 tcp_rcv_rtt_measure_ts(sk, skb); 5640 5641 __skb_pull(skb, tcp_header_len); 5642 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; 5643 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER); 5644 } 5645 if (copied_early) 5646 tcp_cleanup_rbuf(sk, skb->len); 5647 } 5648 if (!eaten) { 5649 if (tcp_checksum_complete_user(sk, skb)) 5650 goto csum_error; 5651 5652 /* Predicted packet is in window by definition. 5653 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 5654 * Hence, check seq<=rcv_wup reduces to: 5655 */ 5656 if (tcp_header_len == 5657 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && 5658 tp->rcv_nxt == tp->rcv_wup) 5659 tcp_store_ts_recent(tp); 5660 5661 tcp_rcv_rtt_measure_ts(sk, skb); 5662 5663 if ((int)skb->truesize > sk->sk_forward_alloc) 5664 goto step5; 5665 5666 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS); 5667 5668 /* Bulk data transfer: receiver */ 5669 tcp_queue_rcv(sk, skb, tcp_header_len); 5670 } 5671 5672 tcp_event_data_recv(sk, skb); 5673 5674 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) { 5675 /* Well, only one small jumplet in fast path... */ 5676 tcp_ack(sk, skb, FLAG_DATA); 5677 tcp_data_snd_check(sk); 5678 if (!inet_csk_ack_scheduled(sk)) 5679 goto no_ack; 5680 } 5681 5682 if (!copied_early || tp->rcv_nxt != tp->rcv_wup) 5683 __tcp_ack_snd_check(sk, 0); 5684no_ack: 5685#ifdef CONFIG_NET_DMA 5686 if (copied_early) 5687 __skb_queue_tail(&sk->sk_async_wait_queue, skb); 5688 else 5689#endif 5690 if (eaten) 5691 __kfree_skb(skb); 5692 else 5693 sk->sk_data_ready(sk, 0); 5694 return 0; 5695 } 5696 } 5697 5698slow_path: 5699 if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb)) 5700 goto csum_error; 5701 5702 /* 5703 * Standard slow path. 5704 */ 5705 5706 res = tcp_validate_incoming(sk, skb, th, 1); 5707 if (res <= 0) 5708 return -res; 5709 5710step5: 5711 if (th->ack && tcp_ack(sk, skb, FLAG_SLOWPATH) < 0) 5712 goto discard; 5713 5714 tcp_rcv_rtt_measure_ts(sk, skb); 5715 5716 /* Process urgent data. */ 5717 tcp_urg(sk, skb, th); 5718 5719 /* step 7: process the segment text */ 5720 tcp_data_queue(sk, skb); 5721 5722 tcp_data_snd_check(sk); 5723 tcp_ack_snd_check(sk); 5724 return 0; 5725 5726csum_error: 5727 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS); 5728 5729discard: 5730 __kfree_skb(skb); 5731 return 0; 5732} 5733EXPORT_SYMBOL(tcp_rcv_established); 5734 5735void tcp_finish_connect(struct sock *sk, struct sk_buff *skb) 5736{ 5737 struct tcp_sock *tp = tcp_sk(sk); 5738 struct inet_connection_sock *icsk = inet_csk(sk); 5739 5740 tcp_set_state(sk, TCP_ESTABLISHED); 5741 5742 if (skb != NULL) 5743 security_inet_conn_established(sk, skb); 5744 5745 /* Make sure socket is routed, for correct metrics. */ 5746 icsk->icsk_af_ops->rebuild_header(sk); 5747 5748 tcp_init_metrics(sk); 5749 5750 tcp_init_congestion_control(sk); 5751 5752 /* Prevent spurious tcp_cwnd_restart() on first data 5753 * packet. 5754 */ 5755 tp->lsndtime = tcp_time_stamp; 5756 5757 tcp_init_buffer_space(sk); 5758 5759 if (sock_flag(sk, SOCK_KEEPOPEN)) 5760 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp)); 5761 5762 if (!tp->rx_opt.snd_wscale) 5763 __tcp_fast_path_on(tp, tp->snd_wnd); 5764 else 5765 tp->pred_flags = 0; 5766 5767 if (!sock_flag(sk, SOCK_DEAD)) { 5768 sk->sk_state_change(sk); 5769 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT); 5770 } 5771} 5772 5773static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb, 5774 const struct tcphdr *th, unsigned int len) 5775{ 5776 const u8 *hash_location; 5777 struct inet_connection_sock *icsk = inet_csk(sk); 5778 struct tcp_sock *tp = tcp_sk(sk); 5779 struct tcp_cookie_values *cvp = tp->cookie_values; 5780 int saved_clamp = tp->rx_opt.mss_clamp; 5781 5782 tcp_parse_options(skb, &tp->rx_opt, &hash_location, 0); 5783 5784 if (th->ack) { 5785 /* rfc793: 5786 * "If the state is SYN-SENT then 5787 * first check the ACK bit 5788 * If the ACK bit is set 5789 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send 5790 * a reset (unless the RST bit is set, if so drop 5791 * the segment and return)" 5792 * 5793 * We do not send data with SYN, so that RFC-correct 5794 * test reduces to: 5795 */ 5796 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt) 5797 goto reset_and_undo; 5798 5799 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 5800 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp, 5801 tcp_time_stamp)) { 5802 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED); 5803 goto reset_and_undo; 5804 } 5805 5806 /* Now ACK is acceptable. 5807 * 5808 * "If the RST bit is set 5809 * If the ACK was acceptable then signal the user "error: 5810 * connection reset", drop the segment, enter CLOSED state, 5811 * delete TCB, and return." 5812 */ 5813 5814 if (th->rst) { 5815 tcp_reset(sk); 5816 goto discard; 5817 } 5818 5819 /* rfc793: 5820 * "fifth, if neither of the SYN or RST bits is set then 5821 * drop the segment and return." 5822 * 5823 * See note below! 5824 * --ANK(990513) 5825 */ 5826 if (!th->syn) 5827 goto discard_and_undo; 5828 5829 /* rfc793: 5830 * "If the SYN bit is on ... 5831 * are acceptable then ... 5832 * (our SYN has been ACKed), change the connection 5833 * state to ESTABLISHED..." 5834 */ 5835 5836 TCP_ECN_rcv_synack(tp, th); 5837 5838 tp->snd_wl1 = TCP_SKB_CB(skb)->seq; 5839 tcp_ack(sk, skb, FLAG_SLOWPATH); 5840 5841 /* Ok.. it's good. Set up sequence numbers and 5842 * move to established. 5843 */ 5844 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; 5845 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; 5846 5847 /* RFC1323: The window in SYN & SYN/ACK segments is 5848 * never scaled. 5849 */ 5850 tp->snd_wnd = ntohs(th->window); 5851 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq); 5852 5853 if (!tp->rx_opt.wscale_ok) { 5854 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0; 5855 tp->window_clamp = min(tp->window_clamp, 65535U); 5856 } 5857 5858 if (tp->rx_opt.saw_tstamp) { 5859 tp->rx_opt.tstamp_ok = 1; 5860 tp->tcp_header_len = 5861 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 5862 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; 5863 tcp_store_ts_recent(tp); 5864 } else { 5865 tp->tcp_header_len = sizeof(struct tcphdr); 5866 } 5867 5868 if (tcp_is_sack(tp) && sysctl_tcp_fack) 5869 tcp_enable_fack(tp); 5870 5871 tcp_mtup_init(sk); 5872 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 5873 tcp_initialize_rcv_mss(sk); 5874 5875 /* Remember, tcp_poll() does not lock socket! 5876 * Change state from SYN-SENT only after copied_seq 5877 * is initialized. */ 5878 tp->copied_seq = tp->rcv_nxt; 5879 5880 if (cvp != NULL && 5881 cvp->cookie_pair_size > 0 && 5882 tp->rx_opt.cookie_plus > 0) { 5883 int cookie_size = tp->rx_opt.cookie_plus 5884 - TCPOLEN_COOKIE_BASE; 5885 int cookie_pair_size = cookie_size 5886 + cvp->cookie_desired; 5887 5888 /* A cookie extension option was sent and returned. 5889 * Note that each incoming SYNACK replaces the 5890 * Responder cookie. The initial exchange is most 5891 * fragile, as protection against spoofing relies 5892 * entirely upon the sequence and timestamp (above). 5893 * This replacement strategy allows the correct pair to 5894 * pass through, while any others will be filtered via 5895 * Responder verification later. 5896 */ 5897 if (sizeof(cvp->cookie_pair) >= cookie_pair_size) { 5898 memcpy(&cvp->cookie_pair[cvp->cookie_desired], 5899 hash_location, cookie_size); 5900 cvp->cookie_pair_size = cookie_pair_size; 5901 } 5902 } 5903 5904 smp_mb(); 5905 5906 tcp_finish_connect(sk, skb); 5907 5908 if (sk->sk_write_pending || 5909 icsk->icsk_accept_queue.rskq_defer_accept || 5910 icsk->icsk_ack.pingpong) { 5911 /* Save one ACK. Data will be ready after 5912 * several ticks, if write_pending is set. 5913 * 5914 * It may be deleted, but with this feature tcpdumps 5915 * look so _wonderfully_ clever, that I was not able 5916 * to stand against the temptation 8) --ANK 5917 */ 5918 inet_csk_schedule_ack(sk); 5919 icsk->icsk_ack.lrcvtime = tcp_time_stamp; 5920 tcp_enter_quickack_mode(sk); 5921 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 5922 TCP_DELACK_MAX, TCP_RTO_MAX); 5923 5924discard: 5925 __kfree_skb(skb); 5926 return 0; 5927 } else { 5928 tcp_send_ack(sk); 5929 } 5930 return -1; 5931 } 5932 5933 /* No ACK in the segment */ 5934 5935 if (th->rst) { 5936 /* rfc793: 5937 * "If the RST bit is set 5938 * 5939 * Otherwise (no ACK) drop the segment and return." 5940 */ 5941 5942 goto discard_and_undo; 5943 } 5944 5945 /* PAWS check. */ 5946 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && 5947 tcp_paws_reject(&tp->rx_opt, 0)) 5948 goto discard_and_undo; 5949 5950 if (th->syn) { 5951 /* We see SYN without ACK. It is attempt of 5952 * simultaneous connect with crossed SYNs. 5953 * Particularly, it can be connect to self. 5954 */ 5955 tcp_set_state(sk, TCP_SYN_RECV); 5956 5957 if (tp->rx_opt.saw_tstamp) { 5958 tp->rx_opt.tstamp_ok = 1; 5959 tcp_store_ts_recent(tp); 5960 tp->tcp_header_len = 5961 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 5962 } else { 5963 tp->tcp_header_len = sizeof(struct tcphdr); 5964 } 5965 5966 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; 5967 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; 5968 5969 /* RFC1323: The window in SYN & SYN/ACK segments is 5970 * never scaled. 5971 */ 5972 tp->snd_wnd = ntohs(th->window); 5973 tp->snd_wl1 = TCP_SKB_CB(skb)->seq; 5974 tp->max_window = tp->snd_wnd; 5975 5976 TCP_ECN_rcv_syn(tp, th); 5977 5978 tcp_mtup_init(sk); 5979 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 5980 tcp_initialize_rcv_mss(sk); 5981 5982 tcp_send_synack(sk); 5983#if 0 5984 /* Note, we could accept data and URG from this segment. 5985 * There are no obstacles to make this. 5986 * 5987 * However, if we ignore data in ACKless segments sometimes, 5988 * we have no reasons to accept it sometimes. 5989 * Also, seems the code doing it in step6 of tcp_rcv_state_process 5990 * is not flawless. So, discard packet for sanity. 5991 * Uncomment this return to process the data. 5992 */ 5993 return -1; 5994#else 5995 goto discard; 5996#endif 5997 } 5998 /* "fifth, if neither of the SYN or RST bits is set then 5999 * drop the segment and return." 6000 */ 6001 6002discard_and_undo: 6003 tcp_clear_options(&tp->rx_opt); 6004 tp->rx_opt.mss_clamp = saved_clamp; 6005 goto discard; 6006 6007reset_and_undo: 6008 tcp_clear_options(&tp->rx_opt); 6009 tp->rx_opt.mss_clamp = saved_clamp; 6010 return 1; 6011} 6012 6013/* 6014 * This function implements the receiving procedure of RFC 793 for 6015 * all states except ESTABLISHED and TIME_WAIT. 6016 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be 6017 * address independent. 6018 */ 6019 6020int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb, 6021 const struct tcphdr *th, unsigned int len) 6022{ 6023 struct tcp_sock *tp = tcp_sk(sk); 6024 struct inet_connection_sock *icsk = inet_csk(sk); 6025 int queued = 0; 6026 int res; 6027 6028 tp->rx_opt.saw_tstamp = 0; 6029 6030 switch (sk->sk_state) { 6031 case TCP_CLOSE: 6032 goto discard; 6033 6034 case TCP_LISTEN: 6035 if (th->ack) 6036 return 1; 6037 6038 if (th->rst) 6039 goto discard; 6040 6041 if (th->syn) { 6042 if (th->fin) 6043 goto discard; 6044 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0) 6045 return 1; 6046 6047 /* Now we have several options: In theory there is 6048 * nothing else in the frame. KA9Q has an option to 6049 * send data with the syn, BSD accepts data with the 6050 * syn up to the [to be] advertised window and 6051 * Solaris 2.1 gives you a protocol error. For now 6052 * we just ignore it, that fits the spec precisely 6053 * and avoids incompatibilities. It would be nice in 6054 * future to drop through and process the data. 6055 * 6056 * Now that TTCP is starting to be used we ought to 6057 * queue this data. 6058 * But, this leaves one open to an easy denial of 6059 * service attack, and SYN cookies can't defend 6060 * against this problem. So, we drop the data 6061 * in the interest of security over speed unless 6062 * it's still in use. 6063 */ 6064 kfree_skb(skb); 6065 return 0; 6066 } 6067 goto discard; 6068 6069 case TCP_SYN_SENT: 6070 queued = tcp_rcv_synsent_state_process(sk, skb, th, len); 6071 if (queued >= 0) 6072 return queued; 6073 6074 /* Do step6 onward by hand. */ 6075 tcp_urg(sk, skb, th); 6076 __kfree_skb(skb); 6077 tcp_data_snd_check(sk); 6078 return 0; 6079 } 6080 6081 res = tcp_validate_incoming(sk, skb, th, 0); 6082 if (res <= 0) 6083 return -res; 6084 6085 /* step 5: check the ACK field */ 6086 if (th->ack) { 6087 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH) > 0; 6088 6089 switch (sk->sk_state) { 6090 case TCP_SYN_RECV: 6091 if (acceptable) { 6092 tp->copied_seq = tp->rcv_nxt; 6093 smp_mb(); 6094 tcp_set_state(sk, TCP_ESTABLISHED); 6095 sk->sk_state_change(sk); 6096 6097 /* Note, that this wakeup is only for marginal 6098 * crossed SYN case. Passively open sockets 6099 * are not waked up, because sk->sk_sleep == 6100 * NULL and sk->sk_socket == NULL. 6101 */ 6102 if (sk->sk_socket) 6103 sk_wake_async(sk, 6104 SOCK_WAKE_IO, POLL_OUT); 6105 6106 tp->snd_una = TCP_SKB_CB(skb)->ack_seq; 6107 tp->snd_wnd = ntohs(th->window) << 6108 tp->rx_opt.snd_wscale; 6109 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq); 6110 6111 if (tp->rx_opt.tstamp_ok) 6112 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; 6113 6114 /* Make sure socket is routed, for 6115 * correct metrics. 6116 */ 6117 icsk->icsk_af_ops->rebuild_header(sk); 6118 6119 tcp_init_metrics(sk); 6120 6121 tcp_init_congestion_control(sk); 6122 6123 /* Prevent spurious tcp_cwnd_restart() on 6124 * first data packet. 6125 */ 6126 tp->lsndtime = tcp_time_stamp; 6127 6128 tcp_mtup_init(sk); 6129 tcp_initialize_rcv_mss(sk); 6130 tcp_init_buffer_space(sk); 6131 tcp_fast_path_on(tp); 6132 } else { 6133 return 1; 6134 } 6135 break; 6136 6137 case TCP_FIN_WAIT1: 6138 if (tp->snd_una == tp->write_seq) { 6139 tcp_set_state(sk, TCP_FIN_WAIT2); 6140 sk->sk_shutdown |= SEND_SHUTDOWN; 6141 dst_confirm(__sk_dst_get(sk)); 6142 6143 if (!sock_flag(sk, SOCK_DEAD)) 6144 /* Wake up lingering close() */ 6145 sk->sk_state_change(sk); 6146 else { 6147 int tmo; 6148 6149 if (tp->linger2 < 0 || 6150 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 6151 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) { 6152 tcp_done(sk); 6153 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA); 6154 return 1; 6155 } 6156 6157 tmo = tcp_fin_time(sk); 6158 if (tmo > TCP_TIMEWAIT_LEN) { 6159 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN); 6160 } else if (th->fin || sock_owned_by_user(sk)) { 6161 /* Bad case. We could lose such FIN otherwise. 6162 * It is not a big problem, but it looks confusing 6163 * and not so rare event. We still can lose it now, 6164 * if it spins in bh_lock_sock(), but it is really 6165 * marginal case. 6166 */ 6167 inet_csk_reset_keepalive_timer(sk, tmo); 6168 } else { 6169 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo); 6170 goto discard; 6171 } 6172 } 6173 } 6174 break; 6175 6176 case TCP_CLOSING: 6177 if (tp->snd_una == tp->write_seq) { 6178 tcp_time_wait(sk, TCP_TIME_WAIT, 0); 6179 goto discard; 6180 } 6181 break; 6182 6183 case TCP_LAST_ACK: 6184 if (tp->snd_una == tp->write_seq) { 6185 tcp_update_metrics(sk); 6186 tcp_done(sk); 6187 goto discard; 6188 } 6189 break; 6190 } 6191 } else 6192 goto discard; 6193 6194 /* step 6: check the URG bit */ 6195 tcp_urg(sk, skb, th); 6196 6197 /* step 7: process the segment text */ 6198 switch (sk->sk_state) { 6199 case TCP_CLOSE_WAIT: 6200 case TCP_CLOSING: 6201 case TCP_LAST_ACK: 6202 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) 6203 break; 6204 case TCP_FIN_WAIT1: 6205 case TCP_FIN_WAIT2: 6206 /* RFC 793 says to queue data in these states, 6207 * RFC 1122 says we MUST send a reset. 6208 * BSD 4.4 also does reset. 6209 */ 6210 if (sk->sk_shutdown & RCV_SHUTDOWN) { 6211 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 6212 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) { 6213 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA); 6214 tcp_reset(sk); 6215 return 1; 6216 } 6217 } 6218 /* Fall through */ 6219 case TCP_ESTABLISHED: 6220 tcp_data_queue(sk, skb); 6221 queued = 1; 6222 break; 6223 } 6224 6225 /* tcp_data could move socket to TIME-WAIT */ 6226 if (sk->sk_state != TCP_CLOSE) { 6227 tcp_data_snd_check(sk); 6228 tcp_ack_snd_check(sk); 6229 } 6230 6231 if (!queued) { 6232discard: 6233 __kfree_skb(skb); 6234 } 6235 return 0; 6236} 6237EXPORT_SYMBOL(tcp_rcv_state_process); 6238