tcp_input.c revision 7363a5b233734dba339f2874ff6ed6c489d3d865
1/* 2 * INET An implementation of the TCP/IP protocol suite for the LINUX 3 * operating system. INET is implemented using the BSD Socket 4 * interface as the means of communication with the user level. 5 * 6 * Implementation of the Transmission Control Protocol(TCP). 7 * 8 * Authors: Ross Biro 9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> 10 * Mark Evans, <evansmp@uhura.aston.ac.uk> 11 * Corey Minyard <wf-rch!minyard@relay.EU.net> 12 * Florian La Roche, <flla@stud.uni-sb.de> 13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu> 14 * Linus Torvalds, <torvalds@cs.helsinki.fi> 15 * Alan Cox, <gw4pts@gw4pts.ampr.org> 16 * Matthew Dillon, <dillon@apollo.west.oic.com> 17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no> 18 * Jorge Cwik, <jorge@laser.satlink.net> 19 */ 20 21/* 22 * Changes: 23 * Pedro Roque : Fast Retransmit/Recovery. 24 * Two receive queues. 25 * Retransmit queue handled by TCP. 26 * Better retransmit timer handling. 27 * New congestion avoidance. 28 * Header prediction. 29 * Variable renaming. 30 * 31 * Eric : Fast Retransmit. 32 * Randy Scott : MSS option defines. 33 * Eric Schenk : Fixes to slow start algorithm. 34 * Eric Schenk : Yet another double ACK bug. 35 * Eric Schenk : Delayed ACK bug fixes. 36 * Eric Schenk : Floyd style fast retrans war avoidance. 37 * David S. Miller : Don't allow zero congestion window. 38 * Eric Schenk : Fix retransmitter so that it sends 39 * next packet on ack of previous packet. 40 * Andi Kleen : Moved open_request checking here 41 * and process RSTs for open_requests. 42 * Andi Kleen : Better prune_queue, and other fixes. 43 * Andrey Savochkin: Fix RTT measurements in the presence of 44 * timestamps. 45 * Andrey Savochkin: Check sequence numbers correctly when 46 * removing SACKs due to in sequence incoming 47 * data segments. 48 * Andi Kleen: Make sure we never ack data there is not 49 * enough room for. Also make this condition 50 * a fatal error if it might still happen. 51 * Andi Kleen: Add tcp_measure_rcv_mss to make 52 * connections with MSS<min(MTU,ann. MSS) 53 * work without delayed acks. 54 * Andi Kleen: Process packets with PSH set in the 55 * fast path. 56 * J Hadi Salim: ECN support 57 * Andrei Gurtov, 58 * Pasi Sarolahti, 59 * Panu Kuhlberg: Experimental audit of TCP (re)transmission 60 * engine. Lots of bugs are found. 61 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs 62 */ 63 64#include <linux/mm.h> 65#include <linux/module.h> 66#include <linux/sysctl.h> 67#include <net/dst.h> 68#include <net/tcp.h> 69#include <net/inet_common.h> 70#include <linux/ipsec.h> 71#include <asm/unaligned.h> 72#include <net/netdma.h> 73 74int sysctl_tcp_timestamps __read_mostly = 1; 75int sysctl_tcp_window_scaling __read_mostly = 1; 76int sysctl_tcp_sack __read_mostly = 1; 77int sysctl_tcp_fack __read_mostly = 1; 78int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH; 79int sysctl_tcp_ecn __read_mostly; 80int sysctl_tcp_dsack __read_mostly = 1; 81int sysctl_tcp_app_win __read_mostly = 31; 82int sysctl_tcp_adv_win_scale __read_mostly = 2; 83 84int sysctl_tcp_stdurg __read_mostly; 85int sysctl_tcp_rfc1337 __read_mostly; 86int sysctl_tcp_max_orphans __read_mostly = NR_FILE; 87int sysctl_tcp_frto __read_mostly = 2; 88int sysctl_tcp_frto_response __read_mostly; 89int sysctl_tcp_nometrics_save __read_mostly; 90 91int sysctl_tcp_moderate_rcvbuf __read_mostly = 1; 92int sysctl_tcp_abc __read_mostly; 93 94#define FLAG_DATA 0x01 /* Incoming frame contained data. */ 95#define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */ 96#define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */ 97#define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */ 98#define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */ 99#define FLAG_DATA_SACKED 0x20 /* New SACK. */ 100#define FLAG_ECE 0x40 /* ECE in this ACK */ 101#define FLAG_DATA_LOST 0x80 /* SACK detected data lossage. */ 102#define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/ 103#define FLAG_ONLY_ORIG_SACKED 0x200 /* SACKs only non-rexmit sent before RTO */ 104#define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */ 105#define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */ 106#define FLAG_NONHEAD_RETRANS_ACKED 0x1000 /* Non-head rexmitted data was ACKed */ 107#define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */ 108 109#define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED) 110#define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED) 111#define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE) 112#define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED) 113#define FLAG_ANY_PROGRESS (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED) 114 115#define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH) 116#define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH)) 117 118/* Adapt the MSS value used to make delayed ack decision to the 119 * real world. 120 */ 121static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb) 122{ 123 struct inet_connection_sock *icsk = inet_csk(sk); 124 const unsigned int lss = icsk->icsk_ack.last_seg_size; 125 unsigned int len; 126 127 icsk->icsk_ack.last_seg_size = 0; 128 129 /* skb->len may jitter because of SACKs, even if peer 130 * sends good full-sized frames. 131 */ 132 len = skb_shinfo(skb)->gso_size ? : skb->len; 133 if (len >= icsk->icsk_ack.rcv_mss) { 134 icsk->icsk_ack.rcv_mss = len; 135 } else { 136 /* Otherwise, we make more careful check taking into account, 137 * that SACKs block is variable. 138 * 139 * "len" is invariant segment length, including TCP header. 140 */ 141 len += skb->data - skb_transport_header(skb); 142 if (len >= TCP_MIN_RCVMSS + sizeof(struct tcphdr) || 143 /* If PSH is not set, packet should be 144 * full sized, provided peer TCP is not badly broken. 145 * This observation (if it is correct 8)) allows 146 * to handle super-low mtu links fairly. 147 */ 148 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) && 149 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) { 150 /* Subtract also invariant (if peer is RFC compliant), 151 * tcp header plus fixed timestamp option length. 152 * Resulting "len" is MSS free of SACK jitter. 153 */ 154 len -= tcp_sk(sk)->tcp_header_len; 155 icsk->icsk_ack.last_seg_size = len; 156 if (len == lss) { 157 icsk->icsk_ack.rcv_mss = len; 158 return; 159 } 160 } 161 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED) 162 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2; 163 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED; 164 } 165} 166 167static void tcp_incr_quickack(struct sock *sk) 168{ 169 struct inet_connection_sock *icsk = inet_csk(sk); 170 unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss); 171 172 if (quickacks == 0) 173 quickacks = 2; 174 if (quickacks > icsk->icsk_ack.quick) 175 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS); 176} 177 178void tcp_enter_quickack_mode(struct sock *sk) 179{ 180 struct inet_connection_sock *icsk = inet_csk(sk); 181 tcp_incr_quickack(sk); 182 icsk->icsk_ack.pingpong = 0; 183 icsk->icsk_ack.ato = TCP_ATO_MIN; 184} 185 186/* Send ACKs quickly, if "quick" count is not exhausted 187 * and the session is not interactive. 188 */ 189 190static inline int tcp_in_quickack_mode(const struct sock *sk) 191{ 192 const struct inet_connection_sock *icsk = inet_csk(sk); 193 return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong; 194} 195 196static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp) 197{ 198 if (tp->ecn_flags & TCP_ECN_OK) 199 tp->ecn_flags |= TCP_ECN_QUEUE_CWR; 200} 201 202static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, struct sk_buff *skb) 203{ 204 if (tcp_hdr(skb)->cwr) 205 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR; 206} 207 208static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp) 209{ 210 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR; 211} 212 213static inline void TCP_ECN_check_ce(struct tcp_sock *tp, struct sk_buff *skb) 214{ 215 if (tp->ecn_flags & TCP_ECN_OK) { 216 if (INET_ECN_is_ce(TCP_SKB_CB(skb)->flags)) 217 tp->ecn_flags |= TCP_ECN_DEMAND_CWR; 218 /* Funny extension: if ECT is not set on a segment, 219 * it is surely retransmit. It is not in ECN RFC, 220 * but Linux follows this rule. */ 221 else if (INET_ECN_is_not_ect((TCP_SKB_CB(skb)->flags))) 222 tcp_enter_quickack_mode((struct sock *)tp); 223 } 224} 225 226static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, struct tcphdr *th) 227{ 228 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr)) 229 tp->ecn_flags &= ~TCP_ECN_OK; 230} 231 232static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, struct tcphdr *th) 233{ 234 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr)) 235 tp->ecn_flags &= ~TCP_ECN_OK; 236} 237 238static inline int TCP_ECN_rcv_ecn_echo(struct tcp_sock *tp, struct tcphdr *th) 239{ 240 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK)) 241 return 1; 242 return 0; 243} 244 245/* Buffer size and advertised window tuning. 246 * 247 * 1. Tuning sk->sk_sndbuf, when connection enters established state. 248 */ 249 250static void tcp_fixup_sndbuf(struct sock *sk) 251{ 252 int sndmem = tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER + 16 + 253 sizeof(struct sk_buff); 254 255 if (sk->sk_sndbuf < 3 * sndmem) 256 sk->sk_sndbuf = min(3 * sndmem, sysctl_tcp_wmem[2]); 257} 258 259/* 2. Tuning advertised window (window_clamp, rcv_ssthresh) 260 * 261 * All tcp_full_space() is split to two parts: "network" buffer, allocated 262 * forward and advertised in receiver window (tp->rcv_wnd) and 263 * "application buffer", required to isolate scheduling/application 264 * latencies from network. 265 * window_clamp is maximal advertised window. It can be less than 266 * tcp_full_space(), in this case tcp_full_space() - window_clamp 267 * is reserved for "application" buffer. The less window_clamp is 268 * the smoother our behaviour from viewpoint of network, but the lower 269 * throughput and the higher sensitivity of the connection to losses. 8) 270 * 271 * rcv_ssthresh is more strict window_clamp used at "slow start" 272 * phase to predict further behaviour of this connection. 273 * It is used for two goals: 274 * - to enforce header prediction at sender, even when application 275 * requires some significant "application buffer". It is check #1. 276 * - to prevent pruning of receive queue because of misprediction 277 * of receiver window. Check #2. 278 * 279 * The scheme does not work when sender sends good segments opening 280 * window and then starts to feed us spaghetti. But it should work 281 * in common situations. Otherwise, we have to rely on queue collapsing. 282 */ 283 284/* Slow part of check#2. */ 285static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb) 286{ 287 struct tcp_sock *tp = tcp_sk(sk); 288 /* Optimize this! */ 289 int truesize = tcp_win_from_space(skb->truesize) >> 1; 290 int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1; 291 292 while (tp->rcv_ssthresh <= window) { 293 if (truesize <= skb->len) 294 return 2 * inet_csk(sk)->icsk_ack.rcv_mss; 295 296 truesize >>= 1; 297 window >>= 1; 298 } 299 return 0; 300} 301 302static void tcp_grow_window(struct sock *sk, struct sk_buff *skb) 303{ 304 struct tcp_sock *tp = tcp_sk(sk); 305 306 /* Check #1 */ 307 if (tp->rcv_ssthresh < tp->window_clamp && 308 (int)tp->rcv_ssthresh < tcp_space(sk) && 309 !tcp_memory_pressure) { 310 int incr; 311 312 /* Check #2. Increase window, if skb with such overhead 313 * will fit to rcvbuf in future. 314 */ 315 if (tcp_win_from_space(skb->truesize) <= skb->len) 316 incr = 2 * tp->advmss; 317 else 318 incr = __tcp_grow_window(sk, skb); 319 320 if (incr) { 321 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr, 322 tp->window_clamp); 323 inet_csk(sk)->icsk_ack.quick |= 1; 324 } 325 } 326} 327 328/* 3. Tuning rcvbuf, when connection enters established state. */ 329 330static void tcp_fixup_rcvbuf(struct sock *sk) 331{ 332 struct tcp_sock *tp = tcp_sk(sk); 333 int rcvmem = tp->advmss + MAX_TCP_HEADER + 16 + sizeof(struct sk_buff); 334 335 /* Try to select rcvbuf so that 4 mss-sized segments 336 * will fit to window and corresponding skbs will fit to our rcvbuf. 337 * (was 3; 4 is minimum to allow fast retransmit to work.) 338 */ 339 while (tcp_win_from_space(rcvmem) < tp->advmss) 340 rcvmem += 128; 341 if (sk->sk_rcvbuf < 4 * rcvmem) 342 sk->sk_rcvbuf = min(4 * rcvmem, sysctl_tcp_rmem[2]); 343} 344 345/* 4. Try to fixup all. It is made immediately after connection enters 346 * established state. 347 */ 348static void tcp_init_buffer_space(struct sock *sk) 349{ 350 struct tcp_sock *tp = tcp_sk(sk); 351 int maxwin; 352 353 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) 354 tcp_fixup_rcvbuf(sk); 355 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK)) 356 tcp_fixup_sndbuf(sk); 357 358 tp->rcvq_space.space = tp->rcv_wnd; 359 360 maxwin = tcp_full_space(sk); 361 362 if (tp->window_clamp >= maxwin) { 363 tp->window_clamp = maxwin; 364 365 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss) 366 tp->window_clamp = max(maxwin - 367 (maxwin >> sysctl_tcp_app_win), 368 4 * tp->advmss); 369 } 370 371 /* Force reservation of one segment. */ 372 if (sysctl_tcp_app_win && 373 tp->window_clamp > 2 * tp->advmss && 374 tp->window_clamp + tp->advmss > maxwin) 375 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss); 376 377 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp); 378 tp->snd_cwnd_stamp = tcp_time_stamp; 379} 380 381/* 5. Recalculate window clamp after socket hit its memory bounds. */ 382static void tcp_clamp_window(struct sock *sk) 383{ 384 struct tcp_sock *tp = tcp_sk(sk); 385 struct inet_connection_sock *icsk = inet_csk(sk); 386 387 icsk->icsk_ack.quick = 0; 388 389 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] && 390 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) && 391 !tcp_memory_pressure && 392 atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) { 393 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc), 394 sysctl_tcp_rmem[2]); 395 } 396 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf) 397 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss); 398} 399 400/* Initialize RCV_MSS value. 401 * RCV_MSS is an our guess about MSS used by the peer. 402 * We haven't any direct information about the MSS. 403 * It's better to underestimate the RCV_MSS rather than overestimate. 404 * Overestimations make us ACKing less frequently than needed. 405 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss(). 406 */ 407void tcp_initialize_rcv_mss(struct sock *sk) 408{ 409 struct tcp_sock *tp = tcp_sk(sk); 410 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache); 411 412 hint = min(hint, tp->rcv_wnd / 2); 413 hint = min(hint, TCP_MIN_RCVMSS); 414 hint = max(hint, TCP_MIN_MSS); 415 416 inet_csk(sk)->icsk_ack.rcv_mss = hint; 417} 418 419/* Receiver "autotuning" code. 420 * 421 * The algorithm for RTT estimation w/o timestamps is based on 422 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL. 423 * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps> 424 * 425 * More detail on this code can be found at 426 * <http://www.psc.edu/~jheffner/senior_thesis.ps>, 427 * though this reference is out of date. A new paper 428 * is pending. 429 */ 430static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep) 431{ 432 u32 new_sample = tp->rcv_rtt_est.rtt; 433 long m = sample; 434 435 if (m == 0) 436 m = 1; 437 438 if (new_sample != 0) { 439 /* If we sample in larger samples in the non-timestamp 440 * case, we could grossly overestimate the RTT especially 441 * with chatty applications or bulk transfer apps which 442 * are stalled on filesystem I/O. 443 * 444 * Also, since we are only going for a minimum in the 445 * non-timestamp case, we do not smooth things out 446 * else with timestamps disabled convergence takes too 447 * long. 448 */ 449 if (!win_dep) { 450 m -= (new_sample >> 3); 451 new_sample += m; 452 } else if (m < new_sample) 453 new_sample = m << 3; 454 } else { 455 /* No previous measure. */ 456 new_sample = m << 3; 457 } 458 459 if (tp->rcv_rtt_est.rtt != new_sample) 460 tp->rcv_rtt_est.rtt = new_sample; 461} 462 463static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp) 464{ 465 if (tp->rcv_rtt_est.time == 0) 466 goto new_measure; 467 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq)) 468 return; 469 tcp_rcv_rtt_update(tp, jiffies - tp->rcv_rtt_est.time, 1); 470 471new_measure: 472 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd; 473 tp->rcv_rtt_est.time = tcp_time_stamp; 474} 475 476static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, 477 const struct sk_buff *skb) 478{ 479 struct tcp_sock *tp = tcp_sk(sk); 480 if (tp->rx_opt.rcv_tsecr && 481 (TCP_SKB_CB(skb)->end_seq - 482 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss)) 483 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0); 484} 485 486/* 487 * This function should be called every time data is copied to user space. 488 * It calculates the appropriate TCP receive buffer space. 489 */ 490void tcp_rcv_space_adjust(struct sock *sk) 491{ 492 struct tcp_sock *tp = tcp_sk(sk); 493 int time; 494 int space; 495 496 if (tp->rcvq_space.time == 0) 497 goto new_measure; 498 499 time = tcp_time_stamp - tp->rcvq_space.time; 500 if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0) 501 return; 502 503 space = 2 * (tp->copied_seq - tp->rcvq_space.seq); 504 505 space = max(tp->rcvq_space.space, space); 506 507 if (tp->rcvq_space.space != space) { 508 int rcvmem; 509 510 tp->rcvq_space.space = space; 511 512 if (sysctl_tcp_moderate_rcvbuf && 513 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) { 514 int new_clamp = space; 515 516 /* Receive space grows, normalize in order to 517 * take into account packet headers and sk_buff 518 * structure overhead. 519 */ 520 space /= tp->advmss; 521 if (!space) 522 space = 1; 523 rcvmem = (tp->advmss + MAX_TCP_HEADER + 524 16 + sizeof(struct sk_buff)); 525 while (tcp_win_from_space(rcvmem) < tp->advmss) 526 rcvmem += 128; 527 space *= rcvmem; 528 space = min(space, sysctl_tcp_rmem[2]); 529 if (space > sk->sk_rcvbuf) { 530 sk->sk_rcvbuf = space; 531 532 /* Make the window clamp follow along. */ 533 tp->window_clamp = new_clamp; 534 } 535 } 536 } 537 538new_measure: 539 tp->rcvq_space.seq = tp->copied_seq; 540 tp->rcvq_space.time = tcp_time_stamp; 541} 542 543/* There is something which you must keep in mind when you analyze the 544 * behavior of the tp->ato delayed ack timeout interval. When a 545 * connection starts up, we want to ack as quickly as possible. The 546 * problem is that "good" TCP's do slow start at the beginning of data 547 * transmission. The means that until we send the first few ACK's the 548 * sender will sit on his end and only queue most of his data, because 549 * he can only send snd_cwnd unacked packets at any given time. For 550 * each ACK we send, he increments snd_cwnd and transmits more of his 551 * queue. -DaveM 552 */ 553static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb) 554{ 555 struct tcp_sock *tp = tcp_sk(sk); 556 struct inet_connection_sock *icsk = inet_csk(sk); 557 u32 now; 558 559 inet_csk_schedule_ack(sk); 560 561 tcp_measure_rcv_mss(sk, skb); 562 563 tcp_rcv_rtt_measure(tp); 564 565 now = tcp_time_stamp; 566 567 if (!icsk->icsk_ack.ato) { 568 /* The _first_ data packet received, initialize 569 * delayed ACK engine. 570 */ 571 tcp_incr_quickack(sk); 572 icsk->icsk_ack.ato = TCP_ATO_MIN; 573 } else { 574 int m = now - icsk->icsk_ack.lrcvtime; 575 576 if (m <= TCP_ATO_MIN / 2) { 577 /* The fastest case is the first. */ 578 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2; 579 } else if (m < icsk->icsk_ack.ato) { 580 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m; 581 if (icsk->icsk_ack.ato > icsk->icsk_rto) 582 icsk->icsk_ack.ato = icsk->icsk_rto; 583 } else if (m > icsk->icsk_rto) { 584 /* Too long gap. Apparently sender failed to 585 * restart window, so that we send ACKs quickly. 586 */ 587 tcp_incr_quickack(sk); 588 sk_mem_reclaim(sk); 589 } 590 } 591 icsk->icsk_ack.lrcvtime = now; 592 593 TCP_ECN_check_ce(tp, skb); 594 595 if (skb->len >= 128) 596 tcp_grow_window(sk, skb); 597} 598 599static u32 tcp_rto_min(struct sock *sk) 600{ 601 struct dst_entry *dst = __sk_dst_get(sk); 602 u32 rto_min = TCP_RTO_MIN; 603 604 if (dst && dst_metric_locked(dst, RTAX_RTO_MIN)) 605 rto_min = dst_metric_rtt(dst, RTAX_RTO_MIN); 606 return rto_min; 607} 608 609/* Called to compute a smoothed rtt estimate. The data fed to this 610 * routine either comes from timestamps, or from segments that were 611 * known _not_ to have been retransmitted [see Karn/Partridge 612 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88 613 * piece by Van Jacobson. 614 * NOTE: the next three routines used to be one big routine. 615 * To save cycles in the RFC 1323 implementation it was better to break 616 * it up into three procedures. -- erics 617 */ 618static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt) 619{ 620 struct tcp_sock *tp = tcp_sk(sk); 621 long m = mrtt; /* RTT */ 622 623 /* The following amusing code comes from Jacobson's 624 * article in SIGCOMM '88. Note that rtt and mdev 625 * are scaled versions of rtt and mean deviation. 626 * This is designed to be as fast as possible 627 * m stands for "measurement". 628 * 629 * On a 1990 paper the rto value is changed to: 630 * RTO = rtt + 4 * mdev 631 * 632 * Funny. This algorithm seems to be very broken. 633 * These formulae increase RTO, when it should be decreased, increase 634 * too slowly, when it should be increased quickly, decrease too quickly 635 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely 636 * does not matter how to _calculate_ it. Seems, it was trap 637 * that VJ failed to avoid. 8) 638 */ 639 if (m == 0) 640 m = 1; 641 if (tp->srtt != 0) { 642 m -= (tp->srtt >> 3); /* m is now error in rtt est */ 643 tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */ 644 if (m < 0) { 645 m = -m; /* m is now abs(error) */ 646 m -= (tp->mdev >> 2); /* similar update on mdev */ 647 /* This is similar to one of Eifel findings. 648 * Eifel blocks mdev updates when rtt decreases. 649 * This solution is a bit different: we use finer gain 650 * for mdev in this case (alpha*beta). 651 * Like Eifel it also prevents growth of rto, 652 * but also it limits too fast rto decreases, 653 * happening in pure Eifel. 654 */ 655 if (m > 0) 656 m >>= 3; 657 } else { 658 m -= (tp->mdev >> 2); /* similar update on mdev */ 659 } 660 tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */ 661 if (tp->mdev > tp->mdev_max) { 662 tp->mdev_max = tp->mdev; 663 if (tp->mdev_max > tp->rttvar) 664 tp->rttvar = tp->mdev_max; 665 } 666 if (after(tp->snd_una, tp->rtt_seq)) { 667 if (tp->mdev_max < tp->rttvar) 668 tp->rttvar -= (tp->rttvar - tp->mdev_max) >> 2; 669 tp->rtt_seq = tp->snd_nxt; 670 tp->mdev_max = tcp_rto_min(sk); 671 } 672 } else { 673 /* no previous measure. */ 674 tp->srtt = m << 3; /* take the measured time to be rtt */ 675 tp->mdev = m << 1; /* make sure rto = 3*rtt */ 676 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk)); 677 tp->rtt_seq = tp->snd_nxt; 678 } 679} 680 681/* Calculate rto without backoff. This is the second half of Van Jacobson's 682 * routine referred to above. 683 */ 684static inline void tcp_set_rto(struct sock *sk) 685{ 686 const struct tcp_sock *tp = tcp_sk(sk); 687 /* Old crap is replaced with new one. 8) 688 * 689 * More seriously: 690 * 1. If rtt variance happened to be less 50msec, it is hallucination. 691 * It cannot be less due to utterly erratic ACK generation made 692 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_ 693 * to do with delayed acks, because at cwnd>2 true delack timeout 694 * is invisible. Actually, Linux-2.4 also generates erratic 695 * ACKs in some circumstances. 696 */ 697 inet_csk(sk)->icsk_rto = (tp->srtt >> 3) + tp->rttvar; 698 699 /* 2. Fixups made earlier cannot be right. 700 * If we do not estimate RTO correctly without them, 701 * all the algo is pure shit and should be replaced 702 * with correct one. It is exactly, which we pretend to do. 703 */ 704 705 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo 706 * guarantees that rto is higher. 707 */ 708 if (inet_csk(sk)->icsk_rto > TCP_RTO_MAX) 709 inet_csk(sk)->icsk_rto = TCP_RTO_MAX; 710} 711 712/* Save metrics learned by this TCP session. 713 This function is called only, when TCP finishes successfully 714 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE. 715 */ 716void tcp_update_metrics(struct sock *sk) 717{ 718 struct tcp_sock *tp = tcp_sk(sk); 719 struct dst_entry *dst = __sk_dst_get(sk); 720 721 if (sysctl_tcp_nometrics_save) 722 return; 723 724 dst_confirm(dst); 725 726 if (dst && (dst->flags & DST_HOST)) { 727 const struct inet_connection_sock *icsk = inet_csk(sk); 728 int m; 729 unsigned long rtt; 730 731 if (icsk->icsk_backoff || !tp->srtt) { 732 /* This session failed to estimate rtt. Why? 733 * Probably, no packets returned in time. 734 * Reset our results. 735 */ 736 if (!(dst_metric_locked(dst, RTAX_RTT))) 737 dst->metrics[RTAX_RTT - 1] = 0; 738 return; 739 } 740 741 rtt = dst_metric_rtt(dst, RTAX_RTT); 742 m = rtt - tp->srtt; 743 744 /* If newly calculated rtt larger than stored one, 745 * store new one. Otherwise, use EWMA. Remember, 746 * rtt overestimation is always better than underestimation. 747 */ 748 if (!(dst_metric_locked(dst, RTAX_RTT))) { 749 if (m <= 0) 750 set_dst_metric_rtt(dst, RTAX_RTT, tp->srtt); 751 else 752 set_dst_metric_rtt(dst, RTAX_RTT, rtt - (m >> 3)); 753 } 754 755 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) { 756 unsigned long var; 757 if (m < 0) 758 m = -m; 759 760 /* Scale deviation to rttvar fixed point */ 761 m >>= 1; 762 if (m < tp->mdev) 763 m = tp->mdev; 764 765 var = dst_metric_rtt(dst, RTAX_RTTVAR); 766 if (m >= var) 767 var = m; 768 else 769 var -= (var - m) >> 2; 770 771 set_dst_metric_rtt(dst, RTAX_RTTVAR, var); 772 } 773 774 if (tp->snd_ssthresh >= 0xFFFF) { 775 /* Slow start still did not finish. */ 776 if (dst_metric(dst, RTAX_SSTHRESH) && 777 !dst_metric_locked(dst, RTAX_SSTHRESH) && 778 (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH)) 779 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1; 780 if (!dst_metric_locked(dst, RTAX_CWND) && 781 tp->snd_cwnd > dst_metric(dst, RTAX_CWND)) 782 dst->metrics[RTAX_CWND - 1] = tp->snd_cwnd; 783 } else if (tp->snd_cwnd > tp->snd_ssthresh && 784 icsk->icsk_ca_state == TCP_CA_Open) { 785 /* Cong. avoidance phase, cwnd is reliable. */ 786 if (!dst_metric_locked(dst, RTAX_SSTHRESH)) 787 dst->metrics[RTAX_SSTHRESH-1] = 788 max(tp->snd_cwnd >> 1, tp->snd_ssthresh); 789 if (!dst_metric_locked(dst, RTAX_CWND)) 790 dst->metrics[RTAX_CWND-1] = (dst_metric(dst, RTAX_CWND) + tp->snd_cwnd) >> 1; 791 } else { 792 /* Else slow start did not finish, cwnd is non-sense, 793 ssthresh may be also invalid. 794 */ 795 if (!dst_metric_locked(dst, RTAX_CWND)) 796 dst->metrics[RTAX_CWND-1] = (dst_metric(dst, RTAX_CWND) + tp->snd_ssthresh) >> 1; 797 if (dst_metric(dst, RTAX_SSTHRESH) && 798 !dst_metric_locked(dst, RTAX_SSTHRESH) && 799 tp->snd_ssthresh > dst_metric(dst, RTAX_SSTHRESH)) 800 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh; 801 } 802 803 if (!dst_metric_locked(dst, RTAX_REORDERING)) { 804 if (dst_metric(dst, RTAX_REORDERING) < tp->reordering && 805 tp->reordering != sysctl_tcp_reordering) 806 dst->metrics[RTAX_REORDERING-1] = tp->reordering; 807 } 808 } 809} 810 811/* Numbers are taken from RFC3390. 812 * 813 * John Heffner states: 814 * 815 * The RFC specifies a window of no more than 4380 bytes 816 * unless 2*MSS > 4380. Reading the pseudocode in the RFC 817 * is a bit misleading because they use a clamp at 4380 bytes 818 * rather than use a multiplier in the relevant range. 819 */ 820__u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst) 821{ 822 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0); 823 824 if (!cwnd) { 825 if (tp->mss_cache > 1460) 826 cwnd = 2; 827 else 828 cwnd = (tp->mss_cache > 1095) ? 3 : 4; 829 } 830 return min_t(__u32, cwnd, tp->snd_cwnd_clamp); 831} 832 833/* Set slow start threshold and cwnd not falling to slow start */ 834void tcp_enter_cwr(struct sock *sk, const int set_ssthresh) 835{ 836 struct tcp_sock *tp = tcp_sk(sk); 837 const struct inet_connection_sock *icsk = inet_csk(sk); 838 839 tp->prior_ssthresh = 0; 840 tp->bytes_acked = 0; 841 if (icsk->icsk_ca_state < TCP_CA_CWR) { 842 tp->undo_marker = 0; 843 if (set_ssthresh) 844 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 845 tp->snd_cwnd = min(tp->snd_cwnd, 846 tcp_packets_in_flight(tp) + 1U); 847 tp->snd_cwnd_cnt = 0; 848 tp->high_seq = tp->snd_nxt; 849 tp->snd_cwnd_stamp = tcp_time_stamp; 850 TCP_ECN_queue_cwr(tp); 851 852 tcp_set_ca_state(sk, TCP_CA_CWR); 853 } 854} 855 856/* 857 * Packet counting of FACK is based on in-order assumptions, therefore TCP 858 * disables it when reordering is detected 859 */ 860static void tcp_disable_fack(struct tcp_sock *tp) 861{ 862 /* RFC3517 uses different metric in lost marker => reset on change */ 863 if (tcp_is_fack(tp)) 864 tp->lost_skb_hint = NULL; 865 tp->rx_opt.sack_ok &= ~2; 866} 867 868/* Take a notice that peer is sending D-SACKs */ 869static void tcp_dsack_seen(struct tcp_sock *tp) 870{ 871 tp->rx_opt.sack_ok |= 4; 872} 873 874/* Initialize metrics on socket. */ 875 876static void tcp_init_metrics(struct sock *sk) 877{ 878 struct tcp_sock *tp = tcp_sk(sk); 879 struct dst_entry *dst = __sk_dst_get(sk); 880 881 if (dst == NULL) 882 goto reset; 883 884 dst_confirm(dst); 885 886 if (dst_metric_locked(dst, RTAX_CWND)) 887 tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND); 888 if (dst_metric(dst, RTAX_SSTHRESH)) { 889 tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH); 890 if (tp->snd_ssthresh > tp->snd_cwnd_clamp) 891 tp->snd_ssthresh = tp->snd_cwnd_clamp; 892 } 893 if (dst_metric(dst, RTAX_REORDERING) && 894 tp->reordering != dst_metric(dst, RTAX_REORDERING)) { 895 tcp_disable_fack(tp); 896 tp->reordering = dst_metric(dst, RTAX_REORDERING); 897 } 898 899 if (dst_metric(dst, RTAX_RTT) == 0) 900 goto reset; 901 902 if (!tp->srtt && dst_metric_rtt(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3)) 903 goto reset; 904 905 /* Initial rtt is determined from SYN,SYN-ACK. 906 * The segment is small and rtt may appear much 907 * less than real one. Use per-dst memory 908 * to make it more realistic. 909 * 910 * A bit of theory. RTT is time passed after "normal" sized packet 911 * is sent until it is ACKed. In normal circumstances sending small 912 * packets force peer to delay ACKs and calculation is correct too. 913 * The algorithm is adaptive and, provided we follow specs, it 914 * NEVER underestimate RTT. BUT! If peer tries to make some clever 915 * tricks sort of "quick acks" for time long enough to decrease RTT 916 * to low value, and then abruptly stops to do it and starts to delay 917 * ACKs, wait for troubles. 918 */ 919 if (dst_metric_rtt(dst, RTAX_RTT) > tp->srtt) { 920 tp->srtt = dst_metric_rtt(dst, RTAX_RTT); 921 tp->rtt_seq = tp->snd_nxt; 922 } 923 if (dst_metric_rtt(dst, RTAX_RTTVAR) > tp->mdev) { 924 tp->mdev = dst_metric_rtt(dst, RTAX_RTTVAR); 925 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk)); 926 } 927 tcp_set_rto(sk); 928 if (inet_csk(sk)->icsk_rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp) 929 goto reset; 930 tp->snd_cwnd = tcp_init_cwnd(tp, dst); 931 tp->snd_cwnd_stamp = tcp_time_stamp; 932 return; 933 934reset: 935 /* Play conservative. If timestamps are not 936 * supported, TCP will fail to recalculate correct 937 * rtt, if initial rto is too small. FORGET ALL AND RESET! 938 */ 939 if (!tp->rx_opt.saw_tstamp && tp->srtt) { 940 tp->srtt = 0; 941 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT; 942 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; 943 } 944} 945 946static void tcp_update_reordering(struct sock *sk, const int metric, 947 const int ts) 948{ 949 struct tcp_sock *tp = tcp_sk(sk); 950 if (metric > tp->reordering) { 951 int mib_idx; 952 953 tp->reordering = min(TCP_MAX_REORDERING, metric); 954 955 /* This exciting event is worth to be remembered. 8) */ 956 if (ts) 957 mib_idx = LINUX_MIB_TCPTSREORDER; 958 else if (tcp_is_reno(tp)) 959 mib_idx = LINUX_MIB_TCPRENOREORDER; 960 else if (tcp_is_fack(tp)) 961 mib_idx = LINUX_MIB_TCPFACKREORDER; 962 else 963 mib_idx = LINUX_MIB_TCPSACKREORDER; 964 965 NET_INC_STATS_BH(sock_net(sk), mib_idx); 966#if FASTRETRANS_DEBUG > 1 967 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n", 968 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state, 969 tp->reordering, 970 tp->fackets_out, 971 tp->sacked_out, 972 tp->undo_marker ? tp->undo_retrans : 0); 973#endif 974 tcp_disable_fack(tp); 975 } 976} 977 978/* This must be called before lost_out is incremented */ 979static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb) 980{ 981 if ((tp->retransmit_skb_hint == NULL) || 982 before(TCP_SKB_CB(skb)->seq, 983 TCP_SKB_CB(tp->retransmit_skb_hint)->seq)) 984 tp->retransmit_skb_hint = skb; 985 986 if (!tp->lost_out || 987 after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high)) 988 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq; 989} 990 991static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb) 992{ 993 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) { 994 tcp_verify_retransmit_hint(tp, skb); 995 996 tp->lost_out += tcp_skb_pcount(skb); 997 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 998 } 999} 1000 1001static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp, 1002 struct sk_buff *skb) 1003{ 1004 tcp_verify_retransmit_hint(tp, skb); 1005 1006 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) { 1007 tp->lost_out += tcp_skb_pcount(skb); 1008 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 1009 } 1010} 1011 1012/* This procedure tags the retransmission queue when SACKs arrive. 1013 * 1014 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L). 1015 * Packets in queue with these bits set are counted in variables 1016 * sacked_out, retrans_out and lost_out, correspondingly. 1017 * 1018 * Valid combinations are: 1019 * Tag InFlight Description 1020 * 0 1 - orig segment is in flight. 1021 * S 0 - nothing flies, orig reached receiver. 1022 * L 0 - nothing flies, orig lost by net. 1023 * R 2 - both orig and retransmit are in flight. 1024 * L|R 1 - orig is lost, retransmit is in flight. 1025 * S|R 1 - orig reached receiver, retrans is still in flight. 1026 * (L|S|R is logically valid, it could occur when L|R is sacked, 1027 * but it is equivalent to plain S and code short-curcuits it to S. 1028 * L|S is logically invalid, it would mean -1 packet in flight 8)) 1029 * 1030 * These 6 states form finite state machine, controlled by the following events: 1031 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue()) 1032 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue()) 1033 * 3. Loss detection event of one of three flavors: 1034 * A. Scoreboard estimator decided the packet is lost. 1035 * A'. Reno "three dupacks" marks head of queue lost. 1036 * A''. Its FACK modfication, head until snd.fack is lost. 1037 * B. SACK arrives sacking data transmitted after never retransmitted 1038 * hole was sent out. 1039 * C. SACK arrives sacking SND.NXT at the moment, when the 1040 * segment was retransmitted. 1041 * 4. D-SACK added new rule: D-SACK changes any tag to S. 1042 * 1043 * It is pleasant to note, that state diagram turns out to be commutative, 1044 * so that we are allowed not to be bothered by order of our actions, 1045 * when multiple events arrive simultaneously. (see the function below). 1046 * 1047 * Reordering detection. 1048 * -------------------- 1049 * Reordering metric is maximal distance, which a packet can be displaced 1050 * in packet stream. With SACKs we can estimate it: 1051 * 1052 * 1. SACK fills old hole and the corresponding segment was not 1053 * ever retransmitted -> reordering. Alas, we cannot use it 1054 * when segment was retransmitted. 1055 * 2. The last flaw is solved with D-SACK. D-SACK arrives 1056 * for retransmitted and already SACKed segment -> reordering.. 1057 * Both of these heuristics are not used in Loss state, when we cannot 1058 * account for retransmits accurately. 1059 * 1060 * SACK block validation. 1061 * ---------------------- 1062 * 1063 * SACK block range validation checks that the received SACK block fits to 1064 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT. 1065 * Note that SND.UNA is not included to the range though being valid because 1066 * it means that the receiver is rather inconsistent with itself reporting 1067 * SACK reneging when it should advance SND.UNA. Such SACK block this is 1068 * perfectly valid, however, in light of RFC2018 which explicitly states 1069 * that "SACK block MUST reflect the newest segment. Even if the newest 1070 * segment is going to be discarded ...", not that it looks very clever 1071 * in case of head skb. Due to potentional receiver driven attacks, we 1072 * choose to avoid immediate execution of a walk in write queue due to 1073 * reneging and defer head skb's loss recovery to standard loss recovery 1074 * procedure that will eventually trigger (nothing forbids us doing this). 1075 * 1076 * Implements also blockage to start_seq wrap-around. Problem lies in the 1077 * fact that though start_seq (s) is before end_seq (i.e., not reversed), 1078 * there's no guarantee that it will be before snd_nxt (n). The problem 1079 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt 1080 * wrap (s_w): 1081 * 1082 * <- outs wnd -> <- wrapzone -> 1083 * u e n u_w e_w s n_w 1084 * | | | | | | | 1085 * |<------------+------+----- TCP seqno space --------------+---------->| 1086 * ...-- <2^31 ->| |<--------... 1087 * ...---- >2^31 ------>| |<--------... 1088 * 1089 * Current code wouldn't be vulnerable but it's better still to discard such 1090 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat 1091 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in 1092 * snd_nxt wrap -> snd_una region will then become "well defined", i.e., 1093 * equal to the ideal case (infinite seqno space without wrap caused issues). 1094 * 1095 * With D-SACK the lower bound is extended to cover sequence space below 1096 * SND.UNA down to undo_marker, which is the last point of interest. Yet 1097 * again, D-SACK block must not to go across snd_una (for the same reason as 1098 * for the normal SACK blocks, explained above). But there all simplicity 1099 * ends, TCP might receive valid D-SACKs below that. As long as they reside 1100 * fully below undo_marker they do not affect behavior in anyway and can 1101 * therefore be safely ignored. In rare cases (which are more or less 1102 * theoretical ones), the D-SACK will nicely cross that boundary due to skb 1103 * fragmentation and packet reordering past skb's retransmission. To consider 1104 * them correctly, the acceptable range must be extended even more though 1105 * the exact amount is rather hard to quantify. However, tp->max_window can 1106 * be used as an exaggerated estimate. 1107 */ 1108static int tcp_is_sackblock_valid(struct tcp_sock *tp, int is_dsack, 1109 u32 start_seq, u32 end_seq) 1110{ 1111 /* Too far in future, or reversed (interpretation is ambiguous) */ 1112 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq)) 1113 return 0; 1114 1115 /* Nasty start_seq wrap-around check (see comments above) */ 1116 if (!before(start_seq, tp->snd_nxt)) 1117 return 0; 1118 1119 /* In outstanding window? ...This is valid exit for D-SACKs too. 1120 * start_seq == snd_una is non-sensical (see comments above) 1121 */ 1122 if (after(start_seq, tp->snd_una)) 1123 return 1; 1124 1125 if (!is_dsack || !tp->undo_marker) 1126 return 0; 1127 1128 /* ...Then it's D-SACK, and must reside below snd_una completely */ 1129 if (!after(end_seq, tp->snd_una)) 1130 return 0; 1131 1132 if (!before(start_seq, tp->undo_marker)) 1133 return 1; 1134 1135 /* Too old */ 1136 if (!after(end_seq, tp->undo_marker)) 1137 return 0; 1138 1139 /* Undo_marker boundary crossing (overestimates a lot). Known already: 1140 * start_seq < undo_marker and end_seq >= undo_marker. 1141 */ 1142 return !before(start_seq, end_seq - tp->max_window); 1143} 1144 1145/* Check for lost retransmit. This superb idea is borrowed from "ratehalving". 1146 * Event "C". Later note: FACK people cheated me again 8), we have to account 1147 * for reordering! Ugly, but should help. 1148 * 1149 * Search retransmitted skbs from write_queue that were sent when snd_nxt was 1150 * less than what is now known to be received by the other end (derived from 1151 * highest SACK block). Also calculate the lowest snd_nxt among the remaining 1152 * retransmitted skbs to avoid some costly processing per ACKs. 1153 */ 1154static void tcp_mark_lost_retrans(struct sock *sk) 1155{ 1156 const struct inet_connection_sock *icsk = inet_csk(sk); 1157 struct tcp_sock *tp = tcp_sk(sk); 1158 struct sk_buff *skb; 1159 int cnt = 0; 1160 u32 new_low_seq = tp->snd_nxt; 1161 u32 received_upto = tcp_highest_sack_seq(tp); 1162 1163 if (!tcp_is_fack(tp) || !tp->retrans_out || 1164 !after(received_upto, tp->lost_retrans_low) || 1165 icsk->icsk_ca_state != TCP_CA_Recovery) 1166 return; 1167 1168 tcp_for_write_queue(skb, sk) { 1169 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq; 1170 1171 if (skb == tcp_send_head(sk)) 1172 break; 1173 if (cnt == tp->retrans_out) 1174 break; 1175 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 1176 continue; 1177 1178 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)) 1179 continue; 1180 1181 /* TODO: We would like to get rid of tcp_is_fack(tp) only 1182 * constraint here (see above) but figuring out that at 1183 * least tp->reordering SACK blocks reside between ack_seq 1184 * and received_upto is not easy task to do cheaply with 1185 * the available datastructures. 1186 * 1187 * Whether FACK should check here for tp->reordering segs 1188 * in-between one could argue for either way (it would be 1189 * rather simple to implement as we could count fack_count 1190 * during the walk and do tp->fackets_out - fack_count). 1191 */ 1192 if (after(received_upto, ack_seq)) { 1193 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 1194 tp->retrans_out -= tcp_skb_pcount(skb); 1195 1196 tcp_skb_mark_lost_uncond_verify(tp, skb); 1197 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT); 1198 } else { 1199 if (before(ack_seq, new_low_seq)) 1200 new_low_seq = ack_seq; 1201 cnt += tcp_skb_pcount(skb); 1202 } 1203 } 1204 1205 if (tp->retrans_out) 1206 tp->lost_retrans_low = new_low_seq; 1207} 1208 1209static int tcp_check_dsack(struct sock *sk, struct sk_buff *ack_skb, 1210 struct tcp_sack_block_wire *sp, int num_sacks, 1211 u32 prior_snd_una) 1212{ 1213 struct tcp_sock *tp = tcp_sk(sk); 1214 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq); 1215 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq); 1216 int dup_sack = 0; 1217 1218 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) { 1219 dup_sack = 1; 1220 tcp_dsack_seen(tp); 1221 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV); 1222 } else if (num_sacks > 1) { 1223 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq); 1224 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq); 1225 1226 if (!after(end_seq_0, end_seq_1) && 1227 !before(start_seq_0, start_seq_1)) { 1228 dup_sack = 1; 1229 tcp_dsack_seen(tp); 1230 NET_INC_STATS_BH(sock_net(sk), 1231 LINUX_MIB_TCPDSACKOFORECV); 1232 } 1233 } 1234 1235 /* D-SACK for already forgotten data... Do dumb counting. */ 1236 if (dup_sack && 1237 !after(end_seq_0, prior_snd_una) && 1238 after(end_seq_0, tp->undo_marker)) 1239 tp->undo_retrans--; 1240 1241 return dup_sack; 1242} 1243 1244struct tcp_sacktag_state { 1245 int reord; 1246 int fack_count; 1247 int flag; 1248}; 1249 1250/* Check if skb is fully within the SACK block. In presence of GSO skbs, 1251 * the incoming SACK may not exactly match but we can find smaller MSS 1252 * aligned portion of it that matches. Therefore we might need to fragment 1253 * which may fail and creates some hassle (caller must handle error case 1254 * returns). 1255 * 1256 * FIXME: this could be merged to shift decision code 1257 */ 1258static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb, 1259 u32 start_seq, u32 end_seq) 1260{ 1261 int in_sack, err; 1262 unsigned int pkt_len; 1263 unsigned int mss; 1264 1265 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && 1266 !before(end_seq, TCP_SKB_CB(skb)->end_seq); 1267 1268 if (tcp_skb_pcount(skb) > 1 && !in_sack && 1269 after(TCP_SKB_CB(skb)->end_seq, start_seq)) { 1270 mss = tcp_skb_mss(skb); 1271 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); 1272 1273 if (!in_sack) { 1274 pkt_len = start_seq - TCP_SKB_CB(skb)->seq; 1275 if (pkt_len < mss) 1276 pkt_len = mss; 1277 } else { 1278 pkt_len = end_seq - TCP_SKB_CB(skb)->seq; 1279 if (pkt_len < mss) 1280 return -EINVAL; 1281 } 1282 1283 /* Round if necessary so that SACKs cover only full MSSes 1284 * and/or the remaining small portion (if present) 1285 */ 1286 if (pkt_len > mss) { 1287 unsigned int new_len = (pkt_len / mss) * mss; 1288 if (!in_sack && new_len < pkt_len) { 1289 new_len += mss; 1290 if (new_len > skb->len) 1291 return 0; 1292 } 1293 pkt_len = new_len; 1294 } 1295 err = tcp_fragment(sk, skb, pkt_len, mss); 1296 if (err < 0) 1297 return err; 1298 } 1299 1300 return in_sack; 1301} 1302 1303static u8 tcp_sacktag_one(struct sk_buff *skb, struct sock *sk, 1304 struct tcp_sacktag_state *state, 1305 int dup_sack, int pcount) 1306{ 1307 struct tcp_sock *tp = tcp_sk(sk); 1308 u8 sacked = TCP_SKB_CB(skb)->sacked; 1309 int fack_count = state->fack_count; 1310 1311 /* Account D-SACK for retransmitted packet. */ 1312 if (dup_sack && (sacked & TCPCB_RETRANS)) { 1313 if (after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker)) 1314 tp->undo_retrans--; 1315 if (sacked & TCPCB_SACKED_ACKED) 1316 state->reord = min(fack_count, state->reord); 1317 } 1318 1319 /* Nothing to do; acked frame is about to be dropped (was ACKed). */ 1320 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 1321 return sacked; 1322 1323 if (!(sacked & TCPCB_SACKED_ACKED)) { 1324 if (sacked & TCPCB_SACKED_RETRANS) { 1325 /* If the segment is not tagged as lost, 1326 * we do not clear RETRANS, believing 1327 * that retransmission is still in flight. 1328 */ 1329 if (sacked & TCPCB_LOST) { 1330 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS); 1331 tp->lost_out -= pcount; 1332 tp->retrans_out -= pcount; 1333 } 1334 } else { 1335 if (!(sacked & TCPCB_RETRANS)) { 1336 /* New sack for not retransmitted frame, 1337 * which was in hole. It is reordering. 1338 */ 1339 if (before(TCP_SKB_CB(skb)->seq, 1340 tcp_highest_sack_seq(tp))) 1341 state->reord = min(fack_count, 1342 state->reord); 1343 1344 /* SACK enhanced F-RTO (RFC4138; Appendix B) */ 1345 if (!after(TCP_SKB_CB(skb)->end_seq, tp->frto_highmark)) 1346 state->flag |= FLAG_ONLY_ORIG_SACKED; 1347 } 1348 1349 if (sacked & TCPCB_LOST) { 1350 sacked &= ~TCPCB_LOST; 1351 tp->lost_out -= pcount; 1352 } 1353 } 1354 1355 sacked |= TCPCB_SACKED_ACKED; 1356 state->flag |= FLAG_DATA_SACKED; 1357 tp->sacked_out += pcount; 1358 1359 fack_count += pcount; 1360 1361 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */ 1362 if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) && 1363 before(TCP_SKB_CB(skb)->seq, 1364 TCP_SKB_CB(tp->lost_skb_hint)->seq)) 1365 tp->lost_cnt_hint += pcount; 1366 1367 if (fack_count > tp->fackets_out) 1368 tp->fackets_out = fack_count; 1369 } 1370 1371 /* D-SACK. We can detect redundant retransmission in S|R and plain R 1372 * frames and clear it. undo_retrans is decreased above, L|R frames 1373 * are accounted above as well. 1374 */ 1375 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) { 1376 sacked &= ~TCPCB_SACKED_RETRANS; 1377 tp->retrans_out -= pcount; 1378 } 1379 1380 return sacked; 1381} 1382 1383static int tcp_shifted_skb(struct sock *sk, struct sk_buff *skb, 1384 struct tcp_sacktag_state *state, 1385 unsigned int pcount, int shifted, int mss, 1386 int dup_sack) 1387{ 1388 struct tcp_sock *tp = tcp_sk(sk); 1389 struct sk_buff *prev = tcp_write_queue_prev(sk, skb); 1390 1391 BUG_ON(!pcount); 1392 1393 /* Tweak before seqno plays */ 1394 if (!tcp_is_fack(tp) && tcp_is_sack(tp) && tp->lost_skb_hint && 1395 !before(TCP_SKB_CB(tp->lost_skb_hint)->seq, TCP_SKB_CB(skb)->seq)) 1396 tp->lost_cnt_hint += pcount; 1397 1398 TCP_SKB_CB(prev)->end_seq += shifted; 1399 TCP_SKB_CB(skb)->seq += shifted; 1400 1401 skb_shinfo(prev)->gso_segs += pcount; 1402 BUG_ON(skb_shinfo(skb)->gso_segs < pcount); 1403 skb_shinfo(skb)->gso_segs -= pcount; 1404 1405 /* When we're adding to gso_segs == 1, gso_size will be zero, 1406 * in theory this shouldn't be necessary but as long as DSACK 1407 * code can come after this skb later on it's better to keep 1408 * setting gso_size to something. 1409 */ 1410 if (!skb_shinfo(prev)->gso_size) { 1411 skb_shinfo(prev)->gso_size = mss; 1412 skb_shinfo(prev)->gso_type = sk->sk_gso_type; 1413 } 1414 1415 /* CHECKME: To clear or not to clear? Mimics normal skb currently */ 1416 if (skb_shinfo(skb)->gso_segs <= 1) { 1417 skb_shinfo(skb)->gso_size = 0; 1418 skb_shinfo(skb)->gso_type = 0; 1419 } 1420 1421 /* We discard results */ 1422 tcp_sacktag_one(skb, sk, state, dup_sack, pcount); 1423 1424 /* Difference in this won't matter, both ACKed by the same cumul. ACK */ 1425 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS); 1426 1427 if (skb->len > 0) { 1428 BUG_ON(!tcp_skb_pcount(skb)); 1429 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED); 1430 return 0; 1431 } 1432 1433 /* Whole SKB was eaten :-) */ 1434 1435 if (skb == tp->retransmit_skb_hint) 1436 tp->retransmit_skb_hint = prev; 1437 if (skb == tp->scoreboard_skb_hint) 1438 tp->scoreboard_skb_hint = prev; 1439 if (skb == tp->lost_skb_hint) { 1440 tp->lost_skb_hint = prev; 1441 tp->lost_cnt_hint -= tcp_skb_pcount(prev); 1442 } 1443 1444 TCP_SKB_CB(skb)->flags |= TCP_SKB_CB(prev)->flags; 1445 if (skb == tcp_highest_sack(sk)) 1446 tcp_advance_highest_sack(sk, skb); 1447 1448 tcp_unlink_write_queue(skb, sk); 1449 sk_wmem_free_skb(sk, skb); 1450 1451 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED); 1452 1453 return 1; 1454} 1455 1456/* I wish gso_size would have a bit more sane initialization than 1457 * something-or-zero which complicates things 1458 */ 1459static int tcp_skb_seglen(struct sk_buff *skb) 1460{ 1461 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb); 1462} 1463 1464/* Shifting pages past head area doesn't work */ 1465static int skb_can_shift(struct sk_buff *skb) 1466{ 1467 return !skb_headlen(skb) && skb_is_nonlinear(skb); 1468} 1469 1470/* Try collapsing SACK blocks spanning across multiple skbs to a single 1471 * skb. 1472 */ 1473static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb, 1474 struct tcp_sacktag_state *state, 1475 u32 start_seq, u32 end_seq, 1476 int dup_sack) 1477{ 1478 struct tcp_sock *tp = tcp_sk(sk); 1479 struct sk_buff *prev; 1480 int mss; 1481 int pcount = 0; 1482 int len; 1483 int in_sack; 1484 1485 if (!sk_can_gso(sk)) 1486 goto fallback; 1487 1488 /* Normally R but no L won't result in plain S */ 1489 if (!dup_sack && 1490 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS) 1491 goto fallback; 1492 if (!skb_can_shift(skb)) 1493 goto fallback; 1494 /* This frame is about to be dropped (was ACKed). */ 1495 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 1496 goto fallback; 1497 1498 /* Can only happen with delayed DSACK + discard craziness */ 1499 if (unlikely(skb == tcp_write_queue_head(sk))) 1500 goto fallback; 1501 prev = tcp_write_queue_prev(sk, skb); 1502 1503 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) 1504 goto fallback; 1505 1506 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && 1507 !before(end_seq, TCP_SKB_CB(skb)->end_seq); 1508 1509 if (in_sack) { 1510 len = skb->len; 1511 pcount = tcp_skb_pcount(skb); 1512 mss = tcp_skb_seglen(skb); 1513 1514 /* TODO: Fix DSACKs to not fragment already SACKed and we can 1515 * drop this restriction as unnecessary 1516 */ 1517 if (mss != tcp_skb_seglen(prev)) 1518 goto fallback; 1519 } else { 1520 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq)) 1521 goto noop; 1522 /* CHECKME: This is non-MSS split case only?, this will 1523 * cause skipped skbs due to advancing loop btw, original 1524 * has that feature too 1525 */ 1526 if (tcp_skb_pcount(skb) <= 1) 1527 goto noop; 1528 1529 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); 1530 if (!in_sack) { 1531 /* TODO: head merge to next could be attempted here 1532 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)), 1533 * though it might not be worth of the additional hassle 1534 * 1535 * ...we can probably just fallback to what was done 1536 * previously. We could try merging non-SACKed ones 1537 * as well but it probably isn't going to buy off 1538 * because later SACKs might again split them, and 1539 * it would make skb timestamp tracking considerably 1540 * harder problem. 1541 */ 1542 goto fallback; 1543 } 1544 1545 len = end_seq - TCP_SKB_CB(skb)->seq; 1546 BUG_ON(len < 0); 1547 BUG_ON(len > skb->len); 1548 1549 /* MSS boundaries should be honoured or else pcount will 1550 * severely break even though it makes things bit trickier. 1551 * Optimize common case to avoid most of the divides 1552 */ 1553 mss = tcp_skb_mss(skb); 1554 1555 /* TODO: Fix DSACKs to not fragment already SACKed and we can 1556 * drop this restriction as unnecessary 1557 */ 1558 if (mss != tcp_skb_seglen(prev)) 1559 goto fallback; 1560 1561 if (len == mss) { 1562 pcount = 1; 1563 } else if (len < mss) { 1564 goto noop; 1565 } else { 1566 pcount = len / mss; 1567 len = pcount * mss; 1568 } 1569 } 1570 1571 if (!skb_shift(prev, skb, len)) 1572 goto fallback; 1573 if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack)) 1574 goto out; 1575 1576 /* Hole filled allows collapsing with the next as well, this is very 1577 * useful when hole on every nth skb pattern happens 1578 */ 1579 if (prev == tcp_write_queue_tail(sk)) 1580 goto out; 1581 skb = tcp_write_queue_next(sk, prev); 1582 1583 if (!skb_can_shift(skb) || 1584 (skb == tcp_send_head(sk)) || 1585 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) || 1586 (mss != tcp_skb_seglen(skb))) 1587 goto out; 1588 1589 len = skb->len; 1590 if (skb_shift(prev, skb, len)) { 1591 pcount += tcp_skb_pcount(skb); 1592 tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0); 1593 } 1594 1595out: 1596 state->fack_count += pcount; 1597 return prev; 1598 1599noop: 1600 return skb; 1601 1602fallback: 1603 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK); 1604 return NULL; 1605} 1606 1607static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk, 1608 struct tcp_sack_block *next_dup, 1609 struct tcp_sacktag_state *state, 1610 u32 start_seq, u32 end_seq, 1611 int dup_sack_in) 1612{ 1613 struct tcp_sock *tp = tcp_sk(sk); 1614 struct sk_buff *tmp; 1615 1616 tcp_for_write_queue_from(skb, sk) { 1617 int in_sack = 0; 1618 int dup_sack = dup_sack_in; 1619 1620 if (skb == tcp_send_head(sk)) 1621 break; 1622 1623 /* queue is in-order => we can short-circuit the walk early */ 1624 if (!before(TCP_SKB_CB(skb)->seq, end_seq)) 1625 break; 1626 1627 if ((next_dup != NULL) && 1628 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) { 1629 in_sack = tcp_match_skb_to_sack(sk, skb, 1630 next_dup->start_seq, 1631 next_dup->end_seq); 1632 if (in_sack > 0) 1633 dup_sack = 1; 1634 } 1635 1636 /* skb reference here is a bit tricky to get right, since 1637 * shifting can eat and free both this skb and the next, 1638 * so not even _safe variant of the loop is enough. 1639 */ 1640 if (in_sack <= 0) { 1641 tmp = tcp_shift_skb_data(sk, skb, state, 1642 start_seq, end_seq, dup_sack); 1643 if (tmp != NULL) { 1644 if (tmp != skb) { 1645 skb = tmp; 1646 continue; 1647 } 1648 1649 in_sack = 0; 1650 } else { 1651 in_sack = tcp_match_skb_to_sack(sk, skb, 1652 start_seq, 1653 end_seq); 1654 } 1655 } 1656 1657 if (unlikely(in_sack < 0)) 1658 break; 1659 1660 if (in_sack) { 1661 TCP_SKB_CB(skb)->sacked = tcp_sacktag_one(skb, sk, 1662 state, 1663 dup_sack, 1664 tcp_skb_pcount(skb)); 1665 1666 if (!before(TCP_SKB_CB(skb)->seq, 1667 tcp_highest_sack_seq(tp))) 1668 tcp_advance_highest_sack(sk, skb); 1669 } 1670 1671 state->fack_count += tcp_skb_pcount(skb); 1672 } 1673 return skb; 1674} 1675 1676/* Avoid all extra work that is being done by sacktag while walking in 1677 * a normal way 1678 */ 1679static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk, 1680 struct tcp_sacktag_state *state, 1681 u32 skip_to_seq) 1682{ 1683 tcp_for_write_queue_from(skb, sk) { 1684 if (skb == tcp_send_head(sk)) 1685 break; 1686 1687 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq)) 1688 break; 1689 1690 state->fack_count += tcp_skb_pcount(skb); 1691 } 1692 return skb; 1693} 1694 1695static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb, 1696 struct sock *sk, 1697 struct tcp_sack_block *next_dup, 1698 struct tcp_sacktag_state *state, 1699 u32 skip_to_seq) 1700{ 1701 if (next_dup == NULL) 1702 return skb; 1703 1704 if (before(next_dup->start_seq, skip_to_seq)) { 1705 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq); 1706 skb = tcp_sacktag_walk(skb, sk, NULL, state, 1707 next_dup->start_seq, next_dup->end_seq, 1708 1); 1709 } 1710 1711 return skb; 1712} 1713 1714static int tcp_sack_cache_ok(struct tcp_sock *tp, struct tcp_sack_block *cache) 1715{ 1716 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache); 1717} 1718 1719static int 1720tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, 1721 u32 prior_snd_una) 1722{ 1723 const struct inet_connection_sock *icsk = inet_csk(sk); 1724 struct tcp_sock *tp = tcp_sk(sk); 1725 unsigned char *ptr = (skb_transport_header(ack_skb) + 1726 TCP_SKB_CB(ack_skb)->sacked); 1727 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2); 1728 struct tcp_sack_block sp[TCP_NUM_SACKS]; 1729 struct tcp_sack_block *cache; 1730 struct tcp_sacktag_state state; 1731 struct sk_buff *skb; 1732 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3); 1733 int used_sacks; 1734 int found_dup_sack = 0; 1735 int i, j; 1736 int first_sack_index; 1737 1738 state.flag = 0; 1739 state.reord = tp->packets_out; 1740 1741 if (!tp->sacked_out) { 1742 if (WARN_ON(tp->fackets_out)) 1743 tp->fackets_out = 0; 1744 tcp_highest_sack_reset(sk); 1745 } 1746 1747 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire, 1748 num_sacks, prior_snd_una); 1749 if (found_dup_sack) 1750 state.flag |= FLAG_DSACKING_ACK; 1751 1752 /* Eliminate too old ACKs, but take into 1753 * account more or less fresh ones, they can 1754 * contain valid SACK info. 1755 */ 1756 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window)) 1757 return 0; 1758 1759 if (!tp->packets_out) 1760 goto out; 1761 1762 used_sacks = 0; 1763 first_sack_index = 0; 1764 for (i = 0; i < num_sacks; i++) { 1765 int dup_sack = !i && found_dup_sack; 1766 1767 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq); 1768 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq); 1769 1770 if (!tcp_is_sackblock_valid(tp, dup_sack, 1771 sp[used_sacks].start_seq, 1772 sp[used_sacks].end_seq)) { 1773 int mib_idx; 1774 1775 if (dup_sack) { 1776 if (!tp->undo_marker) 1777 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO; 1778 else 1779 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD; 1780 } else { 1781 /* Don't count olds caused by ACK reordering */ 1782 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) && 1783 !after(sp[used_sacks].end_seq, tp->snd_una)) 1784 continue; 1785 mib_idx = LINUX_MIB_TCPSACKDISCARD; 1786 } 1787 1788 NET_INC_STATS_BH(sock_net(sk), mib_idx); 1789 if (i == 0) 1790 first_sack_index = -1; 1791 continue; 1792 } 1793 1794 /* Ignore very old stuff early */ 1795 if (!after(sp[used_sacks].end_seq, prior_snd_una)) 1796 continue; 1797 1798 used_sacks++; 1799 } 1800 1801 /* order SACK blocks to allow in order walk of the retrans queue */ 1802 for (i = used_sacks - 1; i > 0; i--) { 1803 for (j = 0; j < i; j++) { 1804 if (after(sp[j].start_seq, sp[j + 1].start_seq)) { 1805 struct tcp_sack_block tmp; 1806 1807 tmp = sp[j]; 1808 sp[j] = sp[j + 1]; 1809 sp[j + 1] = tmp; 1810 1811 /* Track where the first SACK block goes to */ 1812 if (j == first_sack_index) 1813 first_sack_index = j + 1; 1814 } 1815 } 1816 } 1817 1818 skb = tcp_write_queue_head(sk); 1819 state.fack_count = 0; 1820 i = 0; 1821 1822 if (!tp->sacked_out) { 1823 /* It's already past, so skip checking against it */ 1824 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache); 1825 } else { 1826 cache = tp->recv_sack_cache; 1827 /* Skip empty blocks in at head of the cache */ 1828 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq && 1829 !cache->end_seq) 1830 cache++; 1831 } 1832 1833 while (i < used_sacks) { 1834 u32 start_seq = sp[i].start_seq; 1835 u32 end_seq = sp[i].end_seq; 1836 int dup_sack = (found_dup_sack && (i == first_sack_index)); 1837 struct tcp_sack_block *next_dup = NULL; 1838 1839 if (found_dup_sack && ((i + 1) == first_sack_index)) 1840 next_dup = &sp[i + 1]; 1841 1842 /* Event "B" in the comment above. */ 1843 if (after(end_seq, tp->high_seq)) 1844 state.flag |= FLAG_DATA_LOST; 1845 1846 /* Skip too early cached blocks */ 1847 while (tcp_sack_cache_ok(tp, cache) && 1848 !before(start_seq, cache->end_seq)) 1849 cache++; 1850 1851 /* Can skip some work by looking recv_sack_cache? */ 1852 if (tcp_sack_cache_ok(tp, cache) && !dup_sack && 1853 after(end_seq, cache->start_seq)) { 1854 1855 /* Head todo? */ 1856 if (before(start_seq, cache->start_seq)) { 1857 skb = tcp_sacktag_skip(skb, sk, &state, 1858 start_seq); 1859 skb = tcp_sacktag_walk(skb, sk, next_dup, 1860 &state, 1861 start_seq, 1862 cache->start_seq, 1863 dup_sack); 1864 } 1865 1866 /* Rest of the block already fully processed? */ 1867 if (!after(end_seq, cache->end_seq)) 1868 goto advance_sp; 1869 1870 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup, 1871 &state, 1872 cache->end_seq); 1873 1874 /* ...tail remains todo... */ 1875 if (tcp_highest_sack_seq(tp) == cache->end_seq) { 1876 /* ...but better entrypoint exists! */ 1877 skb = tcp_highest_sack(sk); 1878 if (skb == NULL) 1879 break; 1880 state.fack_count = tp->fackets_out; 1881 cache++; 1882 goto walk; 1883 } 1884 1885 skb = tcp_sacktag_skip(skb, sk, &state, cache->end_seq); 1886 /* Check overlap against next cached too (past this one already) */ 1887 cache++; 1888 continue; 1889 } 1890 1891 if (!before(start_seq, tcp_highest_sack_seq(tp))) { 1892 skb = tcp_highest_sack(sk); 1893 if (skb == NULL) 1894 break; 1895 state.fack_count = tp->fackets_out; 1896 } 1897 skb = tcp_sacktag_skip(skb, sk, &state, start_seq); 1898 1899walk: 1900 skb = tcp_sacktag_walk(skb, sk, next_dup, &state, 1901 start_seq, end_seq, dup_sack); 1902 1903advance_sp: 1904 /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct 1905 * due to in-order walk 1906 */ 1907 if (after(end_seq, tp->frto_highmark)) 1908 state.flag &= ~FLAG_ONLY_ORIG_SACKED; 1909 1910 i++; 1911 } 1912 1913 /* Clear the head of the cache sack blocks so we can skip it next time */ 1914 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) { 1915 tp->recv_sack_cache[i].start_seq = 0; 1916 tp->recv_sack_cache[i].end_seq = 0; 1917 } 1918 for (j = 0; j < used_sacks; j++) 1919 tp->recv_sack_cache[i++] = sp[j]; 1920 1921 tcp_mark_lost_retrans(sk); 1922 1923 tcp_verify_left_out(tp); 1924 1925 if ((state.reord < tp->fackets_out) && 1926 ((icsk->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker) && 1927 (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark))) 1928 tcp_update_reordering(sk, tp->fackets_out - state.reord, 0); 1929 1930out: 1931 1932#if FASTRETRANS_DEBUG > 0 1933 WARN_ON((int)tp->sacked_out < 0); 1934 WARN_ON((int)tp->lost_out < 0); 1935 WARN_ON((int)tp->retrans_out < 0); 1936 WARN_ON((int)tcp_packets_in_flight(tp) < 0); 1937#endif 1938 return state.flag; 1939} 1940 1941/* Limits sacked_out so that sum with lost_out isn't ever larger than 1942 * packets_out. Returns zero if sacked_out adjustement wasn't necessary. 1943 */ 1944static int tcp_limit_reno_sacked(struct tcp_sock *tp) 1945{ 1946 u32 holes; 1947 1948 holes = max(tp->lost_out, 1U); 1949 holes = min(holes, tp->packets_out); 1950 1951 if ((tp->sacked_out + holes) > tp->packets_out) { 1952 tp->sacked_out = tp->packets_out - holes; 1953 return 1; 1954 } 1955 return 0; 1956} 1957 1958/* If we receive more dupacks than we expected counting segments 1959 * in assumption of absent reordering, interpret this as reordering. 1960 * The only another reason could be bug in receiver TCP. 1961 */ 1962static void tcp_check_reno_reordering(struct sock *sk, const int addend) 1963{ 1964 struct tcp_sock *tp = tcp_sk(sk); 1965 if (tcp_limit_reno_sacked(tp)) 1966 tcp_update_reordering(sk, tp->packets_out + addend, 0); 1967} 1968 1969/* Emulate SACKs for SACKless connection: account for a new dupack. */ 1970 1971static void tcp_add_reno_sack(struct sock *sk) 1972{ 1973 struct tcp_sock *tp = tcp_sk(sk); 1974 tp->sacked_out++; 1975 tcp_check_reno_reordering(sk, 0); 1976 tcp_verify_left_out(tp); 1977} 1978 1979/* Account for ACK, ACKing some data in Reno Recovery phase. */ 1980 1981static void tcp_remove_reno_sacks(struct sock *sk, int acked) 1982{ 1983 struct tcp_sock *tp = tcp_sk(sk); 1984 1985 if (acked > 0) { 1986 /* One ACK acked hole. The rest eat duplicate ACKs. */ 1987 if (acked - 1 >= tp->sacked_out) 1988 tp->sacked_out = 0; 1989 else 1990 tp->sacked_out -= acked - 1; 1991 } 1992 tcp_check_reno_reordering(sk, acked); 1993 tcp_verify_left_out(tp); 1994} 1995 1996static inline void tcp_reset_reno_sack(struct tcp_sock *tp) 1997{ 1998 tp->sacked_out = 0; 1999} 2000 2001static int tcp_is_sackfrto(const struct tcp_sock *tp) 2002{ 2003 return (sysctl_tcp_frto == 0x2) && !tcp_is_reno(tp); 2004} 2005 2006/* F-RTO can only be used if TCP has never retransmitted anything other than 2007 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here) 2008 */ 2009int tcp_use_frto(struct sock *sk) 2010{ 2011 const struct tcp_sock *tp = tcp_sk(sk); 2012 const struct inet_connection_sock *icsk = inet_csk(sk); 2013 struct sk_buff *skb; 2014 2015 if (!sysctl_tcp_frto) 2016 return 0; 2017 2018 /* MTU probe and F-RTO won't really play nicely along currently */ 2019 if (icsk->icsk_mtup.probe_size) 2020 return 0; 2021 2022 if (tcp_is_sackfrto(tp)) 2023 return 1; 2024 2025 /* Avoid expensive walking of rexmit queue if possible */ 2026 if (tp->retrans_out > 1) 2027 return 0; 2028 2029 skb = tcp_write_queue_head(sk); 2030 if (tcp_skb_is_last(sk, skb)) 2031 return 1; 2032 skb = tcp_write_queue_next(sk, skb); /* Skips head */ 2033 tcp_for_write_queue_from(skb, sk) { 2034 if (skb == tcp_send_head(sk)) 2035 break; 2036 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) 2037 return 0; 2038 /* Short-circuit when first non-SACKed skb has been checked */ 2039 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) 2040 break; 2041 } 2042 return 1; 2043} 2044 2045/* RTO occurred, but do not yet enter Loss state. Instead, defer RTO 2046 * recovery a bit and use heuristics in tcp_process_frto() to detect if 2047 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to 2048 * keep retrans_out counting accurate (with SACK F-RTO, other than head 2049 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS 2050 * bits are handled if the Loss state is really to be entered (in 2051 * tcp_enter_frto_loss). 2052 * 2053 * Do like tcp_enter_loss() would; when RTO expires the second time it 2054 * does: 2055 * "Reduce ssthresh if it has not yet been made inside this window." 2056 */ 2057void tcp_enter_frto(struct sock *sk) 2058{ 2059 const struct inet_connection_sock *icsk = inet_csk(sk); 2060 struct tcp_sock *tp = tcp_sk(sk); 2061 struct sk_buff *skb; 2062 2063 if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) || 2064 tp->snd_una == tp->high_seq || 2065 ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) && 2066 !icsk->icsk_retransmits)) { 2067 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2068 /* Our state is too optimistic in ssthresh() call because cwnd 2069 * is not reduced until tcp_enter_frto_loss() when previous F-RTO 2070 * recovery has not yet completed. Pattern would be this: RTO, 2071 * Cumulative ACK, RTO (2xRTO for the same segment does not end 2072 * up here twice). 2073 * RFC4138 should be more specific on what to do, even though 2074 * RTO is quite unlikely to occur after the first Cumulative ACK 2075 * due to back-off and complexity of triggering events ... 2076 */ 2077 if (tp->frto_counter) { 2078 u32 stored_cwnd; 2079 stored_cwnd = tp->snd_cwnd; 2080 tp->snd_cwnd = 2; 2081 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 2082 tp->snd_cwnd = stored_cwnd; 2083 } else { 2084 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 2085 } 2086 /* ... in theory, cong.control module could do "any tricks" in 2087 * ssthresh(), which means that ca_state, lost bits and lost_out 2088 * counter would have to be faked before the call occurs. We 2089 * consider that too expensive, unlikely and hacky, so modules 2090 * using these in ssthresh() must deal these incompatibility 2091 * issues if they receives CA_EVENT_FRTO and frto_counter != 0 2092 */ 2093 tcp_ca_event(sk, CA_EVENT_FRTO); 2094 } 2095 2096 tp->undo_marker = tp->snd_una; 2097 tp->undo_retrans = 0; 2098 2099 skb = tcp_write_queue_head(sk); 2100 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) 2101 tp->undo_marker = 0; 2102 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 2103 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 2104 tp->retrans_out -= tcp_skb_pcount(skb); 2105 } 2106 tcp_verify_left_out(tp); 2107 2108 /* Too bad if TCP was application limited */ 2109 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1); 2110 2111 /* Earlier loss recovery underway (see RFC4138; Appendix B). 2112 * The last condition is necessary at least in tp->frto_counter case. 2113 */ 2114 if (tcp_is_sackfrto(tp) && (tp->frto_counter || 2115 ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) && 2116 after(tp->high_seq, tp->snd_una)) { 2117 tp->frto_highmark = tp->high_seq; 2118 } else { 2119 tp->frto_highmark = tp->snd_nxt; 2120 } 2121 tcp_set_ca_state(sk, TCP_CA_Disorder); 2122 tp->high_seq = tp->snd_nxt; 2123 tp->frto_counter = 1; 2124} 2125 2126/* Enter Loss state after F-RTO was applied. Dupack arrived after RTO, 2127 * which indicates that we should follow the traditional RTO recovery, 2128 * i.e. mark everything lost and do go-back-N retransmission. 2129 */ 2130static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag) 2131{ 2132 struct tcp_sock *tp = tcp_sk(sk); 2133 struct sk_buff *skb; 2134 2135 tp->lost_out = 0; 2136 tp->retrans_out = 0; 2137 if (tcp_is_reno(tp)) 2138 tcp_reset_reno_sack(tp); 2139 2140 tcp_for_write_queue(skb, sk) { 2141 if (skb == tcp_send_head(sk)) 2142 break; 2143 2144 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; 2145 /* 2146 * Count the retransmission made on RTO correctly (only when 2147 * waiting for the first ACK and did not get it)... 2148 */ 2149 if ((tp->frto_counter == 1) && !(flag & FLAG_DATA_ACKED)) { 2150 /* For some reason this R-bit might get cleared? */ 2151 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 2152 tp->retrans_out += tcp_skb_pcount(skb); 2153 /* ...enter this if branch just for the first segment */ 2154 flag |= FLAG_DATA_ACKED; 2155 } else { 2156 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) 2157 tp->undo_marker = 0; 2158 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 2159 } 2160 2161 /* Marking forward transmissions that were made after RTO lost 2162 * can cause unnecessary retransmissions in some scenarios, 2163 * SACK blocks will mitigate that in some but not in all cases. 2164 * We used to not mark them but it was causing break-ups with 2165 * receivers that do only in-order receival. 2166 * 2167 * TODO: we could detect presence of such receiver and select 2168 * different behavior per flow. 2169 */ 2170 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) { 2171 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 2172 tp->lost_out += tcp_skb_pcount(skb); 2173 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq; 2174 } 2175 } 2176 tcp_verify_left_out(tp); 2177 2178 tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments; 2179 tp->snd_cwnd_cnt = 0; 2180 tp->snd_cwnd_stamp = tcp_time_stamp; 2181 tp->frto_counter = 0; 2182 tp->bytes_acked = 0; 2183 2184 tp->reordering = min_t(unsigned int, tp->reordering, 2185 sysctl_tcp_reordering); 2186 tcp_set_ca_state(sk, TCP_CA_Loss); 2187 tp->high_seq = tp->snd_nxt; 2188 TCP_ECN_queue_cwr(tp); 2189 2190 tcp_clear_all_retrans_hints(tp); 2191} 2192 2193static void tcp_clear_retrans_partial(struct tcp_sock *tp) 2194{ 2195 tp->retrans_out = 0; 2196 tp->lost_out = 0; 2197 2198 tp->undo_marker = 0; 2199 tp->undo_retrans = 0; 2200} 2201 2202void tcp_clear_retrans(struct tcp_sock *tp) 2203{ 2204 tcp_clear_retrans_partial(tp); 2205 2206 tp->fackets_out = 0; 2207 tp->sacked_out = 0; 2208} 2209 2210/* Enter Loss state. If "how" is not zero, forget all SACK information 2211 * and reset tags completely, otherwise preserve SACKs. If receiver 2212 * dropped its ofo queue, we will know this due to reneging detection. 2213 */ 2214void tcp_enter_loss(struct sock *sk, int how) 2215{ 2216 const struct inet_connection_sock *icsk = inet_csk(sk); 2217 struct tcp_sock *tp = tcp_sk(sk); 2218 struct sk_buff *skb; 2219 2220 /* Reduce ssthresh if it has not yet been made inside this window. */ 2221 if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq || 2222 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) { 2223 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2224 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 2225 tcp_ca_event(sk, CA_EVENT_LOSS); 2226 } 2227 tp->snd_cwnd = 1; 2228 tp->snd_cwnd_cnt = 0; 2229 tp->snd_cwnd_stamp = tcp_time_stamp; 2230 2231 tp->bytes_acked = 0; 2232 tcp_clear_retrans_partial(tp); 2233 2234 if (tcp_is_reno(tp)) 2235 tcp_reset_reno_sack(tp); 2236 2237 if (!how) { 2238 /* Push undo marker, if it was plain RTO and nothing 2239 * was retransmitted. */ 2240 tp->undo_marker = tp->snd_una; 2241 } else { 2242 tp->sacked_out = 0; 2243 tp->fackets_out = 0; 2244 } 2245 tcp_clear_all_retrans_hints(tp); 2246 2247 tcp_for_write_queue(skb, sk) { 2248 if (skb == tcp_send_head(sk)) 2249 break; 2250 2251 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) 2252 tp->undo_marker = 0; 2253 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED; 2254 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) { 2255 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED; 2256 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 2257 tp->lost_out += tcp_skb_pcount(skb); 2258 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq; 2259 } 2260 } 2261 tcp_verify_left_out(tp); 2262 2263 tp->reordering = min_t(unsigned int, tp->reordering, 2264 sysctl_tcp_reordering); 2265 tcp_set_ca_state(sk, TCP_CA_Loss); 2266 tp->high_seq = tp->snd_nxt; 2267 TCP_ECN_queue_cwr(tp); 2268 /* Abort F-RTO algorithm if one is in progress */ 2269 tp->frto_counter = 0; 2270} 2271 2272/* If ACK arrived pointing to a remembered SACK, it means that our 2273 * remembered SACKs do not reflect real state of receiver i.e. 2274 * receiver _host_ is heavily congested (or buggy). 2275 * 2276 * Do processing similar to RTO timeout. 2277 */ 2278static int tcp_check_sack_reneging(struct sock *sk, int flag) 2279{ 2280 if (flag & FLAG_SACK_RENEGING) { 2281 struct inet_connection_sock *icsk = inet_csk(sk); 2282 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING); 2283 2284 tcp_enter_loss(sk, 1); 2285 icsk->icsk_retransmits++; 2286 tcp_retransmit_skb(sk, tcp_write_queue_head(sk)); 2287 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 2288 icsk->icsk_rto, TCP_RTO_MAX); 2289 return 1; 2290 } 2291 return 0; 2292} 2293 2294static inline int tcp_fackets_out(struct tcp_sock *tp) 2295{ 2296 return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out; 2297} 2298 2299/* Heurestics to calculate number of duplicate ACKs. There's no dupACKs 2300 * counter when SACK is enabled (without SACK, sacked_out is used for 2301 * that purpose). 2302 * 2303 * Instead, with FACK TCP uses fackets_out that includes both SACKed 2304 * segments up to the highest received SACK block so far and holes in 2305 * between them. 2306 * 2307 * With reordering, holes may still be in flight, so RFC3517 recovery 2308 * uses pure sacked_out (total number of SACKed segments) even though 2309 * it violates the RFC that uses duplicate ACKs, often these are equal 2310 * but when e.g. out-of-window ACKs or packet duplication occurs, 2311 * they differ. Since neither occurs due to loss, TCP should really 2312 * ignore them. 2313 */ 2314static inline int tcp_dupack_heurestics(struct tcp_sock *tp) 2315{ 2316 return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1; 2317} 2318 2319static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb) 2320{ 2321 return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto); 2322} 2323 2324static inline int tcp_head_timedout(struct sock *sk) 2325{ 2326 struct tcp_sock *tp = tcp_sk(sk); 2327 2328 return tp->packets_out && 2329 tcp_skb_timedout(sk, tcp_write_queue_head(sk)); 2330} 2331 2332/* Linux NewReno/SACK/FACK/ECN state machine. 2333 * -------------------------------------- 2334 * 2335 * "Open" Normal state, no dubious events, fast path. 2336 * "Disorder" In all the respects it is "Open", 2337 * but requires a bit more attention. It is entered when 2338 * we see some SACKs or dupacks. It is split of "Open" 2339 * mainly to move some processing from fast path to slow one. 2340 * "CWR" CWND was reduced due to some Congestion Notification event. 2341 * It can be ECN, ICMP source quench, local device congestion. 2342 * "Recovery" CWND was reduced, we are fast-retransmitting. 2343 * "Loss" CWND was reduced due to RTO timeout or SACK reneging. 2344 * 2345 * tcp_fastretrans_alert() is entered: 2346 * - each incoming ACK, if state is not "Open" 2347 * - when arrived ACK is unusual, namely: 2348 * * SACK 2349 * * Duplicate ACK. 2350 * * ECN ECE. 2351 * 2352 * Counting packets in flight is pretty simple. 2353 * 2354 * in_flight = packets_out - left_out + retrans_out 2355 * 2356 * packets_out is SND.NXT-SND.UNA counted in packets. 2357 * 2358 * retrans_out is number of retransmitted segments. 2359 * 2360 * left_out is number of segments left network, but not ACKed yet. 2361 * 2362 * left_out = sacked_out + lost_out 2363 * 2364 * sacked_out: Packets, which arrived to receiver out of order 2365 * and hence not ACKed. With SACKs this number is simply 2366 * amount of SACKed data. Even without SACKs 2367 * it is easy to give pretty reliable estimate of this number, 2368 * counting duplicate ACKs. 2369 * 2370 * lost_out: Packets lost by network. TCP has no explicit 2371 * "loss notification" feedback from network (for now). 2372 * It means that this number can be only _guessed_. 2373 * Actually, it is the heuristics to predict lossage that 2374 * distinguishes different algorithms. 2375 * 2376 * F.e. after RTO, when all the queue is considered as lost, 2377 * lost_out = packets_out and in_flight = retrans_out. 2378 * 2379 * Essentially, we have now two algorithms counting 2380 * lost packets. 2381 * 2382 * FACK: It is the simplest heuristics. As soon as we decided 2383 * that something is lost, we decide that _all_ not SACKed 2384 * packets until the most forward SACK are lost. I.e. 2385 * lost_out = fackets_out - sacked_out and left_out = fackets_out. 2386 * It is absolutely correct estimate, if network does not reorder 2387 * packets. And it loses any connection to reality when reordering 2388 * takes place. We use FACK by default until reordering 2389 * is suspected on the path to this destination. 2390 * 2391 * NewReno: when Recovery is entered, we assume that one segment 2392 * is lost (classic Reno). While we are in Recovery and 2393 * a partial ACK arrives, we assume that one more packet 2394 * is lost (NewReno). This heuristics are the same in NewReno 2395 * and SACK. 2396 * 2397 * Imagine, that's all! Forget about all this shamanism about CWND inflation 2398 * deflation etc. CWND is real congestion window, never inflated, changes 2399 * only according to classic VJ rules. 2400 * 2401 * Really tricky (and requiring careful tuning) part of algorithm 2402 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue(). 2403 * The first determines the moment _when_ we should reduce CWND and, 2404 * hence, slow down forward transmission. In fact, it determines the moment 2405 * when we decide that hole is caused by loss, rather than by a reorder. 2406 * 2407 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill 2408 * holes, caused by lost packets. 2409 * 2410 * And the most logically complicated part of algorithm is undo 2411 * heuristics. We detect false retransmits due to both too early 2412 * fast retransmit (reordering) and underestimated RTO, analyzing 2413 * timestamps and D-SACKs. When we detect that some segments were 2414 * retransmitted by mistake and CWND reduction was wrong, we undo 2415 * window reduction and abort recovery phase. This logic is hidden 2416 * inside several functions named tcp_try_undo_<something>. 2417 */ 2418 2419/* This function decides, when we should leave Disordered state 2420 * and enter Recovery phase, reducing congestion window. 2421 * 2422 * Main question: may we further continue forward transmission 2423 * with the same cwnd? 2424 */ 2425static int tcp_time_to_recover(struct sock *sk) 2426{ 2427 struct tcp_sock *tp = tcp_sk(sk); 2428 __u32 packets_out; 2429 2430 /* Do not perform any recovery during F-RTO algorithm */ 2431 if (tp->frto_counter) 2432 return 0; 2433 2434 /* Trick#1: The loss is proven. */ 2435 if (tp->lost_out) 2436 return 1; 2437 2438 /* Not-A-Trick#2 : Classic rule... */ 2439 if (tcp_dupack_heurestics(tp) > tp->reordering) 2440 return 1; 2441 2442 /* Trick#3 : when we use RFC2988 timer restart, fast 2443 * retransmit can be triggered by timeout of queue head. 2444 */ 2445 if (tcp_is_fack(tp) && tcp_head_timedout(sk)) 2446 return 1; 2447 2448 /* Trick#4: It is still not OK... But will it be useful to delay 2449 * recovery more? 2450 */ 2451 packets_out = tp->packets_out; 2452 if (packets_out <= tp->reordering && 2453 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) && 2454 !tcp_may_send_now(sk)) { 2455 /* We have nothing to send. This connection is limited 2456 * either by receiver window or by application. 2457 */ 2458 return 1; 2459 } 2460 2461 return 0; 2462} 2463 2464/* New heuristics: it is possible only after we switched to restart timer 2465 * each time when something is ACKed. Hence, we can detect timed out packets 2466 * during fast retransmit without falling to slow start. 2467 * 2468 * Usefulness of this as is very questionable, since we should know which of 2469 * the segments is the next to timeout which is relatively expensive to find 2470 * in general case unless we add some data structure just for that. The 2471 * current approach certainly won't find the right one too often and when it 2472 * finally does find _something_ it usually marks large part of the window 2473 * right away (because a retransmission with a larger timestamp blocks the 2474 * loop from advancing). -ij 2475 */ 2476static void tcp_timeout_skbs(struct sock *sk) 2477{ 2478 struct tcp_sock *tp = tcp_sk(sk); 2479 struct sk_buff *skb; 2480 2481 if (!tcp_is_fack(tp) || !tcp_head_timedout(sk)) 2482 return; 2483 2484 skb = tp->scoreboard_skb_hint; 2485 if (tp->scoreboard_skb_hint == NULL) 2486 skb = tcp_write_queue_head(sk); 2487 2488 tcp_for_write_queue_from(skb, sk) { 2489 if (skb == tcp_send_head(sk)) 2490 break; 2491 if (!tcp_skb_timedout(sk, skb)) 2492 break; 2493 2494 tcp_skb_mark_lost(tp, skb); 2495 } 2496 2497 tp->scoreboard_skb_hint = skb; 2498 2499 tcp_verify_left_out(tp); 2500} 2501 2502/* Mark head of queue up as lost. With RFC3517 SACK, the packets is 2503 * is against sacked "cnt", otherwise it's against facked "cnt" 2504 */ 2505static void tcp_mark_head_lost(struct sock *sk, int packets) 2506{ 2507 struct tcp_sock *tp = tcp_sk(sk); 2508 struct sk_buff *skb; 2509 int cnt, oldcnt; 2510 int err; 2511 unsigned int mss; 2512 2513 WARN_ON(packets > tp->packets_out); 2514 if (tp->lost_skb_hint) { 2515 skb = tp->lost_skb_hint; 2516 cnt = tp->lost_cnt_hint; 2517 } else { 2518 skb = tcp_write_queue_head(sk); 2519 cnt = 0; 2520 } 2521 2522 tcp_for_write_queue_from(skb, sk) { 2523 if (skb == tcp_send_head(sk)) 2524 break; 2525 /* TODO: do this better */ 2526 /* this is not the most efficient way to do this... */ 2527 tp->lost_skb_hint = skb; 2528 tp->lost_cnt_hint = cnt; 2529 2530 if (after(TCP_SKB_CB(skb)->end_seq, tp->high_seq)) 2531 break; 2532 2533 oldcnt = cnt; 2534 if (tcp_is_fack(tp) || tcp_is_reno(tp) || 2535 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) 2536 cnt += tcp_skb_pcount(skb); 2537 2538 if (cnt > packets) { 2539 if (tcp_is_sack(tp) || (oldcnt >= packets)) 2540 break; 2541 2542 mss = skb_shinfo(skb)->gso_size; 2543 err = tcp_fragment(sk, skb, (packets - oldcnt) * mss, mss); 2544 if (err < 0) 2545 break; 2546 cnt = packets; 2547 } 2548 2549 tcp_skb_mark_lost(tp, skb); 2550 } 2551 tcp_verify_left_out(tp); 2552} 2553 2554/* Account newly detected lost packet(s) */ 2555 2556static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit) 2557{ 2558 struct tcp_sock *tp = tcp_sk(sk); 2559 2560 if (tcp_is_reno(tp)) { 2561 tcp_mark_head_lost(sk, 1); 2562 } else if (tcp_is_fack(tp)) { 2563 int lost = tp->fackets_out - tp->reordering; 2564 if (lost <= 0) 2565 lost = 1; 2566 tcp_mark_head_lost(sk, lost); 2567 } else { 2568 int sacked_upto = tp->sacked_out - tp->reordering; 2569 if (sacked_upto < fast_rexmit) 2570 sacked_upto = fast_rexmit; 2571 tcp_mark_head_lost(sk, sacked_upto); 2572 } 2573 2574 tcp_timeout_skbs(sk); 2575} 2576 2577/* CWND moderation, preventing bursts due to too big ACKs 2578 * in dubious situations. 2579 */ 2580static inline void tcp_moderate_cwnd(struct tcp_sock *tp) 2581{ 2582 tp->snd_cwnd = min(tp->snd_cwnd, 2583 tcp_packets_in_flight(tp) + tcp_max_burst(tp)); 2584 tp->snd_cwnd_stamp = tcp_time_stamp; 2585} 2586 2587/* Lower bound on congestion window is slow start threshold 2588 * unless congestion avoidance choice decides to overide it. 2589 */ 2590static inline u32 tcp_cwnd_min(const struct sock *sk) 2591{ 2592 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 2593 2594 return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh; 2595} 2596 2597/* Decrease cwnd each second ack. */ 2598static void tcp_cwnd_down(struct sock *sk, int flag) 2599{ 2600 struct tcp_sock *tp = tcp_sk(sk); 2601 int decr = tp->snd_cwnd_cnt + 1; 2602 2603 if ((flag & (FLAG_ANY_PROGRESS | FLAG_DSACKING_ACK)) || 2604 (tcp_is_reno(tp) && !(flag & FLAG_NOT_DUP))) { 2605 tp->snd_cwnd_cnt = decr & 1; 2606 decr >>= 1; 2607 2608 if (decr && tp->snd_cwnd > tcp_cwnd_min(sk)) 2609 tp->snd_cwnd -= decr; 2610 2611 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1); 2612 tp->snd_cwnd_stamp = tcp_time_stamp; 2613 } 2614} 2615 2616/* Nothing was retransmitted or returned timestamp is less 2617 * than timestamp of the first retransmission. 2618 */ 2619static inline int tcp_packet_delayed(struct tcp_sock *tp) 2620{ 2621 return !tp->retrans_stamp || 2622 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 2623 before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp)); 2624} 2625 2626/* Undo procedures. */ 2627 2628#if FASTRETRANS_DEBUG > 1 2629static void DBGUNDO(struct sock *sk, const char *msg) 2630{ 2631 struct tcp_sock *tp = tcp_sk(sk); 2632 struct inet_sock *inet = inet_sk(sk); 2633 2634 if (sk->sk_family == AF_INET) { 2635 printk(KERN_DEBUG "Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n", 2636 msg, 2637 &inet->daddr, ntohs(inet->dport), 2638 tp->snd_cwnd, tcp_left_out(tp), 2639 tp->snd_ssthresh, tp->prior_ssthresh, 2640 tp->packets_out); 2641 } 2642#if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE) 2643 else if (sk->sk_family == AF_INET6) { 2644 struct ipv6_pinfo *np = inet6_sk(sk); 2645 printk(KERN_DEBUG "Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n", 2646 msg, 2647 &np->daddr, ntohs(inet->dport), 2648 tp->snd_cwnd, tcp_left_out(tp), 2649 tp->snd_ssthresh, tp->prior_ssthresh, 2650 tp->packets_out); 2651 } 2652#endif 2653} 2654#else 2655#define DBGUNDO(x...) do { } while (0) 2656#endif 2657 2658static void tcp_undo_cwr(struct sock *sk, const int undo) 2659{ 2660 struct tcp_sock *tp = tcp_sk(sk); 2661 2662 if (tp->prior_ssthresh) { 2663 const struct inet_connection_sock *icsk = inet_csk(sk); 2664 2665 if (icsk->icsk_ca_ops->undo_cwnd) 2666 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk); 2667 else 2668 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1); 2669 2670 if (undo && tp->prior_ssthresh > tp->snd_ssthresh) { 2671 tp->snd_ssthresh = tp->prior_ssthresh; 2672 TCP_ECN_withdraw_cwr(tp); 2673 } 2674 } else { 2675 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh); 2676 } 2677 tcp_moderate_cwnd(tp); 2678 tp->snd_cwnd_stamp = tcp_time_stamp; 2679} 2680 2681static inline int tcp_may_undo(struct tcp_sock *tp) 2682{ 2683 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp)); 2684} 2685 2686/* People celebrate: "We love our President!" */ 2687static int tcp_try_undo_recovery(struct sock *sk) 2688{ 2689 struct tcp_sock *tp = tcp_sk(sk); 2690 2691 if (tcp_may_undo(tp)) { 2692 int mib_idx; 2693 2694 /* Happy end! We did not retransmit anything 2695 * or our original transmission succeeded. 2696 */ 2697 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans"); 2698 tcp_undo_cwr(sk, 1); 2699 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss) 2700 mib_idx = LINUX_MIB_TCPLOSSUNDO; 2701 else 2702 mib_idx = LINUX_MIB_TCPFULLUNDO; 2703 2704 NET_INC_STATS_BH(sock_net(sk), mib_idx); 2705 tp->undo_marker = 0; 2706 } 2707 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) { 2708 /* Hold old state until something *above* high_seq 2709 * is ACKed. For Reno it is MUST to prevent false 2710 * fast retransmits (RFC2582). SACK TCP is safe. */ 2711 tcp_moderate_cwnd(tp); 2712 return 1; 2713 } 2714 tcp_set_ca_state(sk, TCP_CA_Open); 2715 return 0; 2716} 2717 2718/* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */ 2719static void tcp_try_undo_dsack(struct sock *sk) 2720{ 2721 struct tcp_sock *tp = tcp_sk(sk); 2722 2723 if (tp->undo_marker && !tp->undo_retrans) { 2724 DBGUNDO(sk, "D-SACK"); 2725 tcp_undo_cwr(sk, 1); 2726 tp->undo_marker = 0; 2727 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO); 2728 } 2729} 2730 2731/* Undo during fast recovery after partial ACK. */ 2732 2733static int tcp_try_undo_partial(struct sock *sk, int acked) 2734{ 2735 struct tcp_sock *tp = tcp_sk(sk); 2736 /* Partial ACK arrived. Force Hoe's retransmit. */ 2737 int failed = tcp_is_reno(tp) || (tcp_fackets_out(tp) > tp->reordering); 2738 2739 if (tcp_may_undo(tp)) { 2740 /* Plain luck! Hole if filled with delayed 2741 * packet, rather than with a retransmit. 2742 */ 2743 if (tp->retrans_out == 0) 2744 tp->retrans_stamp = 0; 2745 2746 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1); 2747 2748 DBGUNDO(sk, "Hoe"); 2749 tcp_undo_cwr(sk, 0); 2750 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO); 2751 2752 /* So... Do not make Hoe's retransmit yet. 2753 * If the first packet was delayed, the rest 2754 * ones are most probably delayed as well. 2755 */ 2756 failed = 0; 2757 } 2758 return failed; 2759} 2760 2761/* Undo during loss recovery after partial ACK. */ 2762static int tcp_try_undo_loss(struct sock *sk) 2763{ 2764 struct tcp_sock *tp = tcp_sk(sk); 2765 2766 if (tcp_may_undo(tp)) { 2767 struct sk_buff *skb; 2768 tcp_for_write_queue(skb, sk) { 2769 if (skb == tcp_send_head(sk)) 2770 break; 2771 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; 2772 } 2773 2774 tcp_clear_all_retrans_hints(tp); 2775 2776 DBGUNDO(sk, "partial loss"); 2777 tp->lost_out = 0; 2778 tcp_undo_cwr(sk, 1); 2779 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO); 2780 inet_csk(sk)->icsk_retransmits = 0; 2781 tp->undo_marker = 0; 2782 if (tcp_is_sack(tp)) 2783 tcp_set_ca_state(sk, TCP_CA_Open); 2784 return 1; 2785 } 2786 return 0; 2787} 2788 2789static inline void tcp_complete_cwr(struct sock *sk) 2790{ 2791 struct tcp_sock *tp = tcp_sk(sk); 2792 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh); 2793 tp->snd_cwnd_stamp = tcp_time_stamp; 2794 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR); 2795} 2796 2797static void tcp_try_keep_open(struct sock *sk) 2798{ 2799 struct tcp_sock *tp = tcp_sk(sk); 2800 int state = TCP_CA_Open; 2801 2802 if (tcp_left_out(tp) || tp->retrans_out || tp->undo_marker) 2803 state = TCP_CA_Disorder; 2804 2805 if (inet_csk(sk)->icsk_ca_state != state) { 2806 tcp_set_ca_state(sk, state); 2807 tp->high_seq = tp->snd_nxt; 2808 } 2809} 2810 2811static void tcp_try_to_open(struct sock *sk, int flag) 2812{ 2813 struct tcp_sock *tp = tcp_sk(sk); 2814 2815 tcp_verify_left_out(tp); 2816 2817 if (!tp->frto_counter && tp->retrans_out == 0) 2818 tp->retrans_stamp = 0; 2819 2820 if (flag & FLAG_ECE) 2821 tcp_enter_cwr(sk, 1); 2822 2823 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) { 2824 tcp_try_keep_open(sk); 2825 tcp_moderate_cwnd(tp); 2826 } else { 2827 tcp_cwnd_down(sk, flag); 2828 } 2829} 2830 2831static void tcp_mtup_probe_failed(struct sock *sk) 2832{ 2833 struct inet_connection_sock *icsk = inet_csk(sk); 2834 2835 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1; 2836 icsk->icsk_mtup.probe_size = 0; 2837} 2838 2839static void tcp_mtup_probe_success(struct sock *sk, struct sk_buff *skb) 2840{ 2841 struct tcp_sock *tp = tcp_sk(sk); 2842 struct inet_connection_sock *icsk = inet_csk(sk); 2843 2844 /* FIXME: breaks with very large cwnd */ 2845 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2846 tp->snd_cwnd = tp->snd_cwnd * 2847 tcp_mss_to_mtu(sk, tp->mss_cache) / 2848 icsk->icsk_mtup.probe_size; 2849 tp->snd_cwnd_cnt = 0; 2850 tp->snd_cwnd_stamp = tcp_time_stamp; 2851 tp->rcv_ssthresh = tcp_current_ssthresh(sk); 2852 2853 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size; 2854 icsk->icsk_mtup.probe_size = 0; 2855 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 2856} 2857 2858/* Do a simple retransmit without using the backoff mechanisms in 2859 * tcp_timer. This is used for path mtu discovery. 2860 * The socket is already locked here. 2861 */ 2862void tcp_simple_retransmit(struct sock *sk) 2863{ 2864 const struct inet_connection_sock *icsk = inet_csk(sk); 2865 struct tcp_sock *tp = tcp_sk(sk); 2866 struct sk_buff *skb; 2867 unsigned int mss = tcp_current_mss(sk, 0); 2868 u32 prior_lost = tp->lost_out; 2869 2870 tcp_for_write_queue(skb, sk) { 2871 if (skb == tcp_send_head(sk)) 2872 break; 2873 if (tcp_skb_seglen(skb) > mss && 2874 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) { 2875 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 2876 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 2877 tp->retrans_out -= tcp_skb_pcount(skb); 2878 } 2879 tcp_skb_mark_lost_uncond_verify(tp, skb); 2880 } 2881 } 2882 2883 tcp_clear_retrans_hints_partial(tp); 2884 2885 if (prior_lost == tp->lost_out) 2886 return; 2887 2888 if (tcp_is_reno(tp)) 2889 tcp_limit_reno_sacked(tp); 2890 2891 tcp_verify_left_out(tp); 2892 2893 /* Don't muck with the congestion window here. 2894 * Reason is that we do not increase amount of _data_ 2895 * in network, but units changed and effective 2896 * cwnd/ssthresh really reduced now. 2897 */ 2898 if (icsk->icsk_ca_state != TCP_CA_Loss) { 2899 tp->high_seq = tp->snd_nxt; 2900 tp->snd_ssthresh = tcp_current_ssthresh(sk); 2901 tp->prior_ssthresh = 0; 2902 tp->undo_marker = 0; 2903 tcp_set_ca_state(sk, TCP_CA_Loss); 2904 } 2905 tcp_xmit_retransmit_queue(sk); 2906} 2907 2908/* Process an event, which can update packets-in-flight not trivially. 2909 * Main goal of this function is to calculate new estimate for left_out, 2910 * taking into account both packets sitting in receiver's buffer and 2911 * packets lost by network. 2912 * 2913 * Besides that it does CWND reduction, when packet loss is detected 2914 * and changes state of machine. 2915 * 2916 * It does _not_ decide what to send, it is made in function 2917 * tcp_xmit_retransmit_queue(). 2918 */ 2919static void tcp_fastretrans_alert(struct sock *sk, int pkts_acked, int flag) 2920{ 2921 struct inet_connection_sock *icsk = inet_csk(sk); 2922 struct tcp_sock *tp = tcp_sk(sk); 2923 int is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP)); 2924 int do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) && 2925 (tcp_fackets_out(tp) > tp->reordering)); 2926 int fast_rexmit = 0, mib_idx; 2927 2928 if (WARN_ON(!tp->packets_out && tp->sacked_out)) 2929 tp->sacked_out = 0; 2930 if (WARN_ON(!tp->sacked_out && tp->fackets_out)) 2931 tp->fackets_out = 0; 2932 2933 /* Now state machine starts. 2934 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */ 2935 if (flag & FLAG_ECE) 2936 tp->prior_ssthresh = 0; 2937 2938 /* B. In all the states check for reneging SACKs. */ 2939 if (tcp_check_sack_reneging(sk, flag)) 2940 return; 2941 2942 /* C. Process data loss notification, provided it is valid. */ 2943 if (tcp_is_fack(tp) && (flag & FLAG_DATA_LOST) && 2944 before(tp->snd_una, tp->high_seq) && 2945 icsk->icsk_ca_state != TCP_CA_Open && 2946 tp->fackets_out > tp->reordering) { 2947 tcp_mark_head_lost(sk, tp->fackets_out - tp->reordering); 2948 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSS); 2949 } 2950 2951 /* D. Check consistency of the current state. */ 2952 tcp_verify_left_out(tp); 2953 2954 /* E. Check state exit conditions. State can be terminated 2955 * when high_seq is ACKed. */ 2956 if (icsk->icsk_ca_state == TCP_CA_Open) { 2957 WARN_ON(tp->retrans_out != 0); 2958 tp->retrans_stamp = 0; 2959 } else if (!before(tp->snd_una, tp->high_seq)) { 2960 switch (icsk->icsk_ca_state) { 2961 case TCP_CA_Loss: 2962 icsk->icsk_retransmits = 0; 2963 if (tcp_try_undo_recovery(sk)) 2964 return; 2965 break; 2966 2967 case TCP_CA_CWR: 2968 /* CWR is to be held something *above* high_seq 2969 * is ACKed for CWR bit to reach receiver. */ 2970 if (tp->snd_una != tp->high_seq) { 2971 tcp_complete_cwr(sk); 2972 tcp_set_ca_state(sk, TCP_CA_Open); 2973 } 2974 break; 2975 2976 case TCP_CA_Disorder: 2977 tcp_try_undo_dsack(sk); 2978 if (!tp->undo_marker || 2979 /* For SACK case do not Open to allow to undo 2980 * catching for all duplicate ACKs. */ 2981 tcp_is_reno(tp) || tp->snd_una != tp->high_seq) { 2982 tp->undo_marker = 0; 2983 tcp_set_ca_state(sk, TCP_CA_Open); 2984 } 2985 break; 2986 2987 case TCP_CA_Recovery: 2988 if (tcp_is_reno(tp)) 2989 tcp_reset_reno_sack(tp); 2990 if (tcp_try_undo_recovery(sk)) 2991 return; 2992 tcp_complete_cwr(sk); 2993 break; 2994 } 2995 } 2996 2997 /* F. Process state. */ 2998 switch (icsk->icsk_ca_state) { 2999 case TCP_CA_Recovery: 3000 if (!(flag & FLAG_SND_UNA_ADVANCED)) { 3001 if (tcp_is_reno(tp) && is_dupack) 3002 tcp_add_reno_sack(sk); 3003 } else 3004 do_lost = tcp_try_undo_partial(sk, pkts_acked); 3005 break; 3006 case TCP_CA_Loss: 3007 if (flag & FLAG_DATA_ACKED) 3008 icsk->icsk_retransmits = 0; 3009 if (tcp_is_reno(tp) && flag & FLAG_SND_UNA_ADVANCED) 3010 tcp_reset_reno_sack(tp); 3011 if (!tcp_try_undo_loss(sk)) { 3012 tcp_moderate_cwnd(tp); 3013 tcp_xmit_retransmit_queue(sk); 3014 return; 3015 } 3016 if (icsk->icsk_ca_state != TCP_CA_Open) 3017 return; 3018 /* Loss is undone; fall through to processing in Open state. */ 3019 default: 3020 if (tcp_is_reno(tp)) { 3021 if (flag & FLAG_SND_UNA_ADVANCED) 3022 tcp_reset_reno_sack(tp); 3023 if (is_dupack) 3024 tcp_add_reno_sack(sk); 3025 } 3026 3027 if (icsk->icsk_ca_state == TCP_CA_Disorder) 3028 tcp_try_undo_dsack(sk); 3029 3030 if (!tcp_time_to_recover(sk)) { 3031 tcp_try_to_open(sk, flag); 3032 return; 3033 } 3034 3035 /* MTU probe failure: don't reduce cwnd */ 3036 if (icsk->icsk_ca_state < TCP_CA_CWR && 3037 icsk->icsk_mtup.probe_size && 3038 tp->snd_una == tp->mtu_probe.probe_seq_start) { 3039 tcp_mtup_probe_failed(sk); 3040 /* Restores the reduction we did in tcp_mtup_probe() */ 3041 tp->snd_cwnd++; 3042 tcp_simple_retransmit(sk); 3043 return; 3044 } 3045 3046 /* Otherwise enter Recovery state */ 3047 3048 if (tcp_is_reno(tp)) 3049 mib_idx = LINUX_MIB_TCPRENORECOVERY; 3050 else 3051 mib_idx = LINUX_MIB_TCPSACKRECOVERY; 3052 3053 NET_INC_STATS_BH(sock_net(sk), mib_idx); 3054 3055 tp->high_seq = tp->snd_nxt; 3056 tp->prior_ssthresh = 0; 3057 tp->undo_marker = tp->snd_una; 3058 tp->undo_retrans = tp->retrans_out; 3059 3060 if (icsk->icsk_ca_state < TCP_CA_CWR) { 3061 if (!(flag & FLAG_ECE)) 3062 tp->prior_ssthresh = tcp_current_ssthresh(sk); 3063 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 3064 TCP_ECN_queue_cwr(tp); 3065 } 3066 3067 tp->bytes_acked = 0; 3068 tp->snd_cwnd_cnt = 0; 3069 tcp_set_ca_state(sk, TCP_CA_Recovery); 3070 fast_rexmit = 1; 3071 } 3072 3073 if (do_lost || (tcp_is_fack(tp) && tcp_head_timedout(sk))) 3074 tcp_update_scoreboard(sk, fast_rexmit); 3075 tcp_cwnd_down(sk, flag); 3076 tcp_xmit_retransmit_queue(sk); 3077} 3078 3079static void tcp_valid_rtt_meas(struct sock *sk, u32 seq_rtt) 3080{ 3081 tcp_rtt_estimator(sk, seq_rtt); 3082 tcp_set_rto(sk); 3083 inet_csk(sk)->icsk_backoff = 0; 3084} 3085 3086/* Read draft-ietf-tcplw-high-performance before mucking 3087 * with this code. (Supersedes RFC1323) 3088 */ 3089static void tcp_ack_saw_tstamp(struct sock *sk, int flag) 3090{ 3091 /* RTTM Rule: A TSecr value received in a segment is used to 3092 * update the averaged RTT measurement only if the segment 3093 * acknowledges some new data, i.e., only if it advances the 3094 * left edge of the send window. 3095 * 3096 * See draft-ietf-tcplw-high-performance-00, section 3.3. 3097 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru> 3098 * 3099 * Changed: reset backoff as soon as we see the first valid sample. 3100 * If we do not, we get strongly overestimated rto. With timestamps 3101 * samples are accepted even from very old segments: f.e., when rtt=1 3102 * increases to 8, we retransmit 5 times and after 8 seconds delayed 3103 * answer arrives rto becomes 120 seconds! If at least one of segments 3104 * in window is lost... Voila. --ANK (010210) 3105 */ 3106 struct tcp_sock *tp = tcp_sk(sk); 3107 3108 tcp_valid_rtt_meas(sk, tcp_time_stamp - tp->rx_opt.rcv_tsecr); 3109} 3110 3111static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag) 3112{ 3113 /* We don't have a timestamp. Can only use 3114 * packets that are not retransmitted to determine 3115 * rtt estimates. Also, we must not reset the 3116 * backoff for rto until we get a non-retransmitted 3117 * packet. This allows us to deal with a situation 3118 * where the network delay has increased suddenly. 3119 * I.e. Karn's algorithm. (SIGCOMM '87, p5.) 3120 */ 3121 3122 if (flag & FLAG_RETRANS_DATA_ACKED) 3123 return; 3124 3125 tcp_valid_rtt_meas(sk, seq_rtt); 3126} 3127 3128static inline void tcp_ack_update_rtt(struct sock *sk, const int flag, 3129 const s32 seq_rtt) 3130{ 3131 const struct tcp_sock *tp = tcp_sk(sk); 3132 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */ 3133 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr) 3134 tcp_ack_saw_tstamp(sk, flag); 3135 else if (seq_rtt >= 0) 3136 tcp_ack_no_tstamp(sk, seq_rtt, flag); 3137} 3138 3139static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight) 3140{ 3141 const struct inet_connection_sock *icsk = inet_csk(sk); 3142 icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight); 3143 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp; 3144} 3145 3146/* Restart timer after forward progress on connection. 3147 * RFC2988 recommends to restart timer to now+rto. 3148 */ 3149static void tcp_rearm_rto(struct sock *sk) 3150{ 3151 struct tcp_sock *tp = tcp_sk(sk); 3152 3153 if (!tp->packets_out) { 3154 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS); 3155 } else { 3156 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 3157 inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 3158 } 3159} 3160 3161/* If we get here, the whole TSO packet has not been acked. */ 3162static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb) 3163{ 3164 struct tcp_sock *tp = tcp_sk(sk); 3165 u32 packets_acked; 3166 3167 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)); 3168 3169 packets_acked = tcp_skb_pcount(skb); 3170 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 3171 return 0; 3172 packets_acked -= tcp_skb_pcount(skb); 3173 3174 if (packets_acked) { 3175 BUG_ON(tcp_skb_pcount(skb) == 0); 3176 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)); 3177 } 3178 3179 return packets_acked; 3180} 3181 3182/* Remove acknowledged frames from the retransmission queue. If our packet 3183 * is before the ack sequence we can discard it as it's confirmed to have 3184 * arrived at the other end. 3185 */ 3186static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets, 3187 u32 prior_snd_una) 3188{ 3189 struct tcp_sock *tp = tcp_sk(sk); 3190 const struct inet_connection_sock *icsk = inet_csk(sk); 3191 struct sk_buff *skb; 3192 u32 now = tcp_time_stamp; 3193 int fully_acked = 1; 3194 int flag = 0; 3195 u32 pkts_acked = 0; 3196 u32 reord = tp->packets_out; 3197 u32 prior_sacked = tp->sacked_out; 3198 s32 seq_rtt = -1; 3199 s32 ca_seq_rtt = -1; 3200 ktime_t last_ackt = net_invalid_timestamp(); 3201 3202 while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) { 3203 struct tcp_skb_cb *scb = TCP_SKB_CB(skb); 3204 u32 end_seq; 3205 u32 acked_pcount; 3206 u8 sacked = scb->sacked; 3207 3208 /* Determine how many packets and what bytes were acked, tso and else */ 3209 if (after(scb->end_seq, tp->snd_una)) { 3210 if (tcp_skb_pcount(skb) == 1 || 3211 !after(tp->snd_una, scb->seq)) 3212 break; 3213 3214 acked_pcount = tcp_tso_acked(sk, skb); 3215 if (!acked_pcount) 3216 break; 3217 3218 fully_acked = 0; 3219 end_seq = tp->snd_una; 3220 } else { 3221 acked_pcount = tcp_skb_pcount(skb); 3222 end_seq = scb->end_seq; 3223 } 3224 3225 /* MTU probing checks */ 3226 if (fully_acked && icsk->icsk_mtup.probe_size && 3227 !after(tp->mtu_probe.probe_seq_end, scb->end_seq)) { 3228 tcp_mtup_probe_success(sk, skb); 3229 } 3230 3231 if (sacked & TCPCB_RETRANS) { 3232 if (sacked & TCPCB_SACKED_RETRANS) 3233 tp->retrans_out -= acked_pcount; 3234 flag |= FLAG_RETRANS_DATA_ACKED; 3235 ca_seq_rtt = -1; 3236 seq_rtt = -1; 3237 if ((flag & FLAG_DATA_ACKED) || (acked_pcount > 1)) 3238 flag |= FLAG_NONHEAD_RETRANS_ACKED; 3239 } else { 3240 ca_seq_rtt = now - scb->when; 3241 last_ackt = skb->tstamp; 3242 if (seq_rtt < 0) { 3243 seq_rtt = ca_seq_rtt; 3244 } 3245 if (!(sacked & TCPCB_SACKED_ACKED)) 3246 reord = min(pkts_acked, reord); 3247 } 3248 3249 if (sacked & TCPCB_SACKED_ACKED) 3250 tp->sacked_out -= acked_pcount; 3251 if (sacked & TCPCB_LOST) 3252 tp->lost_out -= acked_pcount; 3253 3254 tp->packets_out -= acked_pcount; 3255 pkts_acked += acked_pcount; 3256 3257 /* Initial outgoing SYN's get put onto the write_queue 3258 * just like anything else we transmit. It is not 3259 * true data, and if we misinform our callers that 3260 * this ACK acks real data, we will erroneously exit 3261 * connection startup slow start one packet too 3262 * quickly. This is severely frowned upon behavior. 3263 */ 3264 if (!(scb->flags & TCPCB_FLAG_SYN)) { 3265 flag |= FLAG_DATA_ACKED; 3266 } else { 3267 flag |= FLAG_SYN_ACKED; 3268 tp->retrans_stamp = 0; 3269 } 3270 3271 if (!fully_acked) 3272 break; 3273 3274 tcp_unlink_write_queue(skb, sk); 3275 sk_wmem_free_skb(sk, skb); 3276 tp->scoreboard_skb_hint = NULL; 3277 if (skb == tp->retransmit_skb_hint) 3278 tp->retransmit_skb_hint = NULL; 3279 if (skb == tp->lost_skb_hint) 3280 tp->lost_skb_hint = NULL; 3281 } 3282 3283 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una))) 3284 tp->snd_up = tp->snd_una; 3285 3286 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) 3287 flag |= FLAG_SACK_RENEGING; 3288 3289 if (flag & FLAG_ACKED) { 3290 const struct tcp_congestion_ops *ca_ops 3291 = inet_csk(sk)->icsk_ca_ops; 3292 3293 tcp_ack_update_rtt(sk, flag, seq_rtt); 3294 tcp_rearm_rto(sk); 3295 3296 if (tcp_is_reno(tp)) { 3297 tcp_remove_reno_sacks(sk, pkts_acked); 3298 } else { 3299 int delta; 3300 3301 /* Non-retransmitted hole got filled? That's reordering */ 3302 if (reord < prior_fackets) 3303 tcp_update_reordering(sk, tp->fackets_out - reord, 0); 3304 3305 delta = tcp_is_fack(tp) ? pkts_acked : 3306 prior_sacked - tp->sacked_out; 3307 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta); 3308 } 3309 3310 tp->fackets_out -= min(pkts_acked, tp->fackets_out); 3311 3312 if (ca_ops->pkts_acked) { 3313 s32 rtt_us = -1; 3314 3315 /* Is the ACK triggering packet unambiguous? */ 3316 if (!(flag & FLAG_RETRANS_DATA_ACKED)) { 3317 /* High resolution needed and available? */ 3318 if (ca_ops->flags & TCP_CONG_RTT_STAMP && 3319 !ktime_equal(last_ackt, 3320 net_invalid_timestamp())) 3321 rtt_us = ktime_us_delta(ktime_get_real(), 3322 last_ackt); 3323 else if (ca_seq_rtt > 0) 3324 rtt_us = jiffies_to_usecs(ca_seq_rtt); 3325 } 3326 3327 ca_ops->pkts_acked(sk, pkts_acked, rtt_us); 3328 } 3329 } 3330 3331#if FASTRETRANS_DEBUG > 0 3332 WARN_ON((int)tp->sacked_out < 0); 3333 WARN_ON((int)tp->lost_out < 0); 3334 WARN_ON((int)tp->retrans_out < 0); 3335 if (!tp->packets_out && tcp_is_sack(tp)) { 3336 icsk = inet_csk(sk); 3337 if (tp->lost_out) { 3338 printk(KERN_DEBUG "Leak l=%u %d\n", 3339 tp->lost_out, icsk->icsk_ca_state); 3340 tp->lost_out = 0; 3341 } 3342 if (tp->sacked_out) { 3343 printk(KERN_DEBUG "Leak s=%u %d\n", 3344 tp->sacked_out, icsk->icsk_ca_state); 3345 tp->sacked_out = 0; 3346 } 3347 if (tp->retrans_out) { 3348 printk(KERN_DEBUG "Leak r=%u %d\n", 3349 tp->retrans_out, icsk->icsk_ca_state); 3350 tp->retrans_out = 0; 3351 } 3352 } 3353#endif 3354 return flag; 3355} 3356 3357static void tcp_ack_probe(struct sock *sk) 3358{ 3359 const struct tcp_sock *tp = tcp_sk(sk); 3360 struct inet_connection_sock *icsk = inet_csk(sk); 3361 3362 /* Was it a usable window open? */ 3363 3364 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) { 3365 icsk->icsk_backoff = 0; 3366 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0); 3367 /* Socket must be waked up by subsequent tcp_data_snd_check(). 3368 * This function is not for random using! 3369 */ 3370 } else { 3371 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 3372 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), 3373 TCP_RTO_MAX); 3374 } 3375} 3376 3377static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag) 3378{ 3379 return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) || 3380 inet_csk(sk)->icsk_ca_state != TCP_CA_Open); 3381} 3382 3383static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag) 3384{ 3385 const struct tcp_sock *tp = tcp_sk(sk); 3386 return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) && 3387 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR)); 3388} 3389 3390/* Check that window update is acceptable. 3391 * The function assumes that snd_una<=ack<=snd_next. 3392 */ 3393static inline int tcp_may_update_window(const struct tcp_sock *tp, 3394 const u32 ack, const u32 ack_seq, 3395 const u32 nwin) 3396{ 3397 return (after(ack, tp->snd_una) || 3398 after(ack_seq, tp->snd_wl1) || 3399 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd)); 3400} 3401 3402/* Update our send window. 3403 * 3404 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2 3405 * and in FreeBSD. NetBSD's one is even worse.) is wrong. 3406 */ 3407static int tcp_ack_update_window(struct sock *sk, struct sk_buff *skb, u32 ack, 3408 u32 ack_seq) 3409{ 3410 struct tcp_sock *tp = tcp_sk(sk); 3411 int flag = 0; 3412 u32 nwin = ntohs(tcp_hdr(skb)->window); 3413 3414 if (likely(!tcp_hdr(skb)->syn)) 3415 nwin <<= tp->rx_opt.snd_wscale; 3416 3417 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) { 3418 flag |= FLAG_WIN_UPDATE; 3419 tcp_update_wl(tp, ack, ack_seq); 3420 3421 if (tp->snd_wnd != nwin) { 3422 tp->snd_wnd = nwin; 3423 3424 /* Note, it is the only place, where 3425 * fast path is recovered for sending TCP. 3426 */ 3427 tp->pred_flags = 0; 3428 tcp_fast_path_check(sk); 3429 3430 if (nwin > tp->max_window) { 3431 tp->max_window = nwin; 3432 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie); 3433 } 3434 } 3435 } 3436 3437 tp->snd_una = ack; 3438 3439 return flag; 3440} 3441 3442/* A very conservative spurious RTO response algorithm: reduce cwnd and 3443 * continue in congestion avoidance. 3444 */ 3445static void tcp_conservative_spur_to_response(struct tcp_sock *tp) 3446{ 3447 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh); 3448 tp->snd_cwnd_cnt = 0; 3449 tp->bytes_acked = 0; 3450 TCP_ECN_queue_cwr(tp); 3451 tcp_moderate_cwnd(tp); 3452} 3453 3454/* A conservative spurious RTO response algorithm: reduce cwnd using 3455 * rate halving and continue in congestion avoidance. 3456 */ 3457static void tcp_ratehalving_spur_to_response(struct sock *sk) 3458{ 3459 tcp_enter_cwr(sk, 0); 3460} 3461 3462static void tcp_undo_spur_to_response(struct sock *sk, int flag) 3463{ 3464 if (flag & FLAG_ECE) 3465 tcp_ratehalving_spur_to_response(sk); 3466 else 3467 tcp_undo_cwr(sk, 1); 3468} 3469 3470/* F-RTO spurious RTO detection algorithm (RFC4138) 3471 * 3472 * F-RTO affects during two new ACKs following RTO (well, almost, see inline 3473 * comments). State (ACK number) is kept in frto_counter. When ACK advances 3474 * window (but not to or beyond highest sequence sent before RTO): 3475 * On First ACK, send two new segments out. 3476 * On Second ACK, RTO was likely spurious. Do spurious response (response 3477 * algorithm is not part of the F-RTO detection algorithm 3478 * given in RFC4138 but can be selected separately). 3479 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss 3480 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding 3481 * of Nagle, this is done using frto_counter states 2 and 3, when a new data 3482 * segment of any size sent during F-RTO, state 2 is upgraded to 3. 3483 * 3484 * Rationale: if the RTO was spurious, new ACKs should arrive from the 3485 * original window even after we transmit two new data segments. 3486 * 3487 * SACK version: 3488 * on first step, wait until first cumulative ACK arrives, then move to 3489 * the second step. In second step, the next ACK decides. 3490 * 3491 * F-RTO is implemented (mainly) in four functions: 3492 * - tcp_use_frto() is used to determine if TCP is can use F-RTO 3493 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is 3494 * called when tcp_use_frto() showed green light 3495 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm 3496 * - tcp_enter_frto_loss() is called if there is not enough evidence 3497 * to prove that the RTO is indeed spurious. It transfers the control 3498 * from F-RTO to the conventional RTO recovery 3499 */ 3500static int tcp_process_frto(struct sock *sk, int flag) 3501{ 3502 struct tcp_sock *tp = tcp_sk(sk); 3503 3504 tcp_verify_left_out(tp); 3505 3506 /* Duplicate the behavior from Loss state (fastretrans_alert) */ 3507 if (flag & FLAG_DATA_ACKED) 3508 inet_csk(sk)->icsk_retransmits = 0; 3509 3510 if ((flag & FLAG_NONHEAD_RETRANS_ACKED) || 3511 ((tp->frto_counter >= 2) && (flag & FLAG_RETRANS_DATA_ACKED))) 3512 tp->undo_marker = 0; 3513 3514 if (!before(tp->snd_una, tp->frto_highmark)) { 3515 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag); 3516 return 1; 3517 } 3518 3519 if (!tcp_is_sackfrto(tp)) { 3520 /* RFC4138 shortcoming in step 2; should also have case c): 3521 * ACK isn't duplicate nor advances window, e.g., opposite dir 3522 * data, winupdate 3523 */ 3524 if (!(flag & FLAG_ANY_PROGRESS) && (flag & FLAG_NOT_DUP)) 3525 return 1; 3526 3527 if (!(flag & FLAG_DATA_ACKED)) { 3528 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3), 3529 flag); 3530 return 1; 3531 } 3532 } else { 3533 if (!(flag & FLAG_DATA_ACKED) && (tp->frto_counter == 1)) { 3534 /* Prevent sending of new data. */ 3535 tp->snd_cwnd = min(tp->snd_cwnd, 3536 tcp_packets_in_flight(tp)); 3537 return 1; 3538 } 3539 3540 if ((tp->frto_counter >= 2) && 3541 (!(flag & FLAG_FORWARD_PROGRESS) || 3542 ((flag & FLAG_DATA_SACKED) && 3543 !(flag & FLAG_ONLY_ORIG_SACKED)))) { 3544 /* RFC4138 shortcoming (see comment above) */ 3545 if (!(flag & FLAG_FORWARD_PROGRESS) && 3546 (flag & FLAG_NOT_DUP)) 3547 return 1; 3548 3549 tcp_enter_frto_loss(sk, 3, flag); 3550 return 1; 3551 } 3552 } 3553 3554 if (tp->frto_counter == 1) { 3555 /* tcp_may_send_now needs to see updated state */ 3556 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2; 3557 tp->frto_counter = 2; 3558 3559 if (!tcp_may_send_now(sk)) 3560 tcp_enter_frto_loss(sk, 2, flag); 3561 3562 return 1; 3563 } else { 3564 switch (sysctl_tcp_frto_response) { 3565 case 2: 3566 tcp_undo_spur_to_response(sk, flag); 3567 break; 3568 case 1: 3569 tcp_conservative_spur_to_response(tp); 3570 break; 3571 default: 3572 tcp_ratehalving_spur_to_response(sk); 3573 break; 3574 } 3575 tp->frto_counter = 0; 3576 tp->undo_marker = 0; 3577 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSPURIOUSRTOS); 3578 } 3579 return 0; 3580} 3581 3582/* This routine deals with incoming acks, but not outgoing ones. */ 3583static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag) 3584{ 3585 struct inet_connection_sock *icsk = inet_csk(sk); 3586 struct tcp_sock *tp = tcp_sk(sk); 3587 u32 prior_snd_una = tp->snd_una; 3588 u32 ack_seq = TCP_SKB_CB(skb)->seq; 3589 u32 ack = TCP_SKB_CB(skb)->ack_seq; 3590 u32 prior_in_flight; 3591 u32 prior_fackets; 3592 int prior_packets; 3593 int frto_cwnd = 0; 3594 3595 /* If the ack is newer than sent or older than previous acks 3596 * then we can probably ignore it. 3597 */ 3598 if (after(ack, tp->snd_nxt)) 3599 goto uninteresting_ack; 3600 3601 if (before(ack, prior_snd_una)) 3602 goto old_ack; 3603 3604 if (after(ack, prior_snd_una)) 3605 flag |= FLAG_SND_UNA_ADVANCED; 3606 3607 if (sysctl_tcp_abc) { 3608 if (icsk->icsk_ca_state < TCP_CA_CWR) 3609 tp->bytes_acked += ack - prior_snd_una; 3610 else if (icsk->icsk_ca_state == TCP_CA_Loss) 3611 /* we assume just one segment left network */ 3612 tp->bytes_acked += min(ack - prior_snd_una, 3613 tp->mss_cache); 3614 } 3615 3616 prior_fackets = tp->fackets_out; 3617 prior_in_flight = tcp_packets_in_flight(tp); 3618 3619 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) { 3620 /* Window is constant, pure forward advance. 3621 * No more checks are required. 3622 * Note, we use the fact that SND.UNA>=SND.WL2. 3623 */ 3624 tcp_update_wl(tp, ack, ack_seq); 3625 tp->snd_una = ack; 3626 flag |= FLAG_WIN_UPDATE; 3627 3628 tcp_ca_event(sk, CA_EVENT_FAST_ACK); 3629 3630 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS); 3631 } else { 3632 if (ack_seq != TCP_SKB_CB(skb)->end_seq) 3633 flag |= FLAG_DATA; 3634 else 3635 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS); 3636 3637 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq); 3638 3639 if (TCP_SKB_CB(skb)->sacked) 3640 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una); 3641 3642 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb))) 3643 flag |= FLAG_ECE; 3644 3645 tcp_ca_event(sk, CA_EVENT_SLOW_ACK); 3646 } 3647 3648 /* We passed data and got it acked, remove any soft error 3649 * log. Something worked... 3650 */ 3651 sk->sk_err_soft = 0; 3652 icsk->icsk_probes_out = 0; 3653 tp->rcv_tstamp = tcp_time_stamp; 3654 prior_packets = tp->packets_out; 3655 if (!prior_packets) 3656 goto no_queue; 3657 3658 /* See if we can take anything off of the retransmit queue. */ 3659 flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una); 3660 3661 if (tp->frto_counter) 3662 frto_cwnd = tcp_process_frto(sk, flag); 3663 /* Guarantee sacktag reordering detection against wrap-arounds */ 3664 if (before(tp->frto_highmark, tp->snd_una)) 3665 tp->frto_highmark = 0; 3666 3667 if (tcp_ack_is_dubious(sk, flag)) { 3668 /* Advance CWND, if state allows this. */ 3669 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd && 3670 tcp_may_raise_cwnd(sk, flag)) 3671 tcp_cong_avoid(sk, ack, prior_in_flight); 3672 tcp_fastretrans_alert(sk, prior_packets - tp->packets_out, 3673 flag); 3674 } else { 3675 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd) 3676 tcp_cong_avoid(sk, ack, prior_in_flight); 3677 } 3678 3679 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) 3680 dst_confirm(sk->sk_dst_cache); 3681 3682 return 1; 3683 3684no_queue: 3685 /* If this ack opens up a zero window, clear backoff. It was 3686 * being used to time the probes, and is probably far higher than 3687 * it needs to be for normal retransmission. 3688 */ 3689 if (tcp_send_head(sk)) 3690 tcp_ack_probe(sk); 3691 return 1; 3692 3693old_ack: 3694 if (TCP_SKB_CB(skb)->sacked) { 3695 tcp_sacktag_write_queue(sk, skb, prior_snd_una); 3696 if (icsk->icsk_ca_state == TCP_CA_Open) 3697 tcp_try_keep_open(sk); 3698 } 3699 3700uninteresting_ack: 3701 SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt); 3702 return 0; 3703} 3704 3705/* Look for tcp options. Normally only called on SYN and SYNACK packets. 3706 * But, this can also be called on packets in the established flow when 3707 * the fast version below fails. 3708 */ 3709void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx, 3710 int estab) 3711{ 3712 unsigned char *ptr; 3713 struct tcphdr *th = tcp_hdr(skb); 3714 int length = (th->doff * 4) - sizeof(struct tcphdr); 3715 3716 ptr = (unsigned char *)(th + 1); 3717 opt_rx->saw_tstamp = 0; 3718 3719 while (length > 0) { 3720 int opcode = *ptr++; 3721 int opsize; 3722 3723 switch (opcode) { 3724 case TCPOPT_EOL: 3725 return; 3726 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */ 3727 length--; 3728 continue; 3729 default: 3730 opsize = *ptr++; 3731 if (opsize < 2) /* "silly options" */ 3732 return; 3733 if (opsize > length) 3734 return; /* don't parse partial options */ 3735 switch (opcode) { 3736 case TCPOPT_MSS: 3737 if (opsize == TCPOLEN_MSS && th->syn && !estab) { 3738 u16 in_mss = get_unaligned_be16(ptr); 3739 if (in_mss) { 3740 if (opt_rx->user_mss && 3741 opt_rx->user_mss < in_mss) 3742 in_mss = opt_rx->user_mss; 3743 opt_rx->mss_clamp = in_mss; 3744 } 3745 } 3746 break; 3747 case TCPOPT_WINDOW: 3748 if (opsize == TCPOLEN_WINDOW && th->syn && 3749 !estab && sysctl_tcp_window_scaling) { 3750 __u8 snd_wscale = *(__u8 *)ptr; 3751 opt_rx->wscale_ok = 1; 3752 if (snd_wscale > 14) { 3753 if (net_ratelimit()) 3754 printk(KERN_INFO "tcp_parse_options: Illegal window " 3755 "scaling value %d >14 received.\n", 3756 snd_wscale); 3757 snd_wscale = 14; 3758 } 3759 opt_rx->snd_wscale = snd_wscale; 3760 } 3761 break; 3762 case TCPOPT_TIMESTAMP: 3763 if ((opsize == TCPOLEN_TIMESTAMP) && 3764 ((estab && opt_rx->tstamp_ok) || 3765 (!estab && sysctl_tcp_timestamps))) { 3766 opt_rx->saw_tstamp = 1; 3767 opt_rx->rcv_tsval = get_unaligned_be32(ptr); 3768 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4); 3769 } 3770 break; 3771 case TCPOPT_SACK_PERM: 3772 if (opsize == TCPOLEN_SACK_PERM && th->syn && 3773 !estab && sysctl_tcp_sack) { 3774 opt_rx->sack_ok = 1; 3775 tcp_sack_reset(opt_rx); 3776 } 3777 break; 3778 3779 case TCPOPT_SACK: 3780 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) && 3781 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) && 3782 opt_rx->sack_ok) { 3783 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th; 3784 } 3785 break; 3786#ifdef CONFIG_TCP_MD5SIG 3787 case TCPOPT_MD5SIG: 3788 /* 3789 * The MD5 Hash has already been 3790 * checked (see tcp_v{4,6}_do_rcv()). 3791 */ 3792 break; 3793#endif 3794 } 3795 3796 ptr += opsize-2; 3797 length -= opsize; 3798 } 3799 } 3800} 3801 3802static int tcp_parse_aligned_timestamp(struct tcp_sock *tp, struct tcphdr *th) 3803{ 3804 __be32 *ptr = (__be32 *)(th + 1); 3805 3806 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) 3807 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) { 3808 tp->rx_opt.saw_tstamp = 1; 3809 ++ptr; 3810 tp->rx_opt.rcv_tsval = ntohl(*ptr); 3811 ++ptr; 3812 tp->rx_opt.rcv_tsecr = ntohl(*ptr); 3813 return 1; 3814 } 3815 return 0; 3816} 3817 3818/* Fast parse options. This hopes to only see timestamps. 3819 * If it is wrong it falls back on tcp_parse_options(). 3820 */ 3821static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th, 3822 struct tcp_sock *tp) 3823{ 3824 if (th->doff == sizeof(struct tcphdr) >> 2) { 3825 tp->rx_opt.saw_tstamp = 0; 3826 return 0; 3827 } else if (tp->rx_opt.tstamp_ok && 3828 th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) { 3829 if (tcp_parse_aligned_timestamp(tp, th)) 3830 return 1; 3831 } 3832 tcp_parse_options(skb, &tp->rx_opt, 1); 3833 return 1; 3834} 3835 3836#ifdef CONFIG_TCP_MD5SIG 3837/* 3838 * Parse MD5 Signature option 3839 */ 3840u8 *tcp_parse_md5sig_option(struct tcphdr *th) 3841{ 3842 int length = (th->doff << 2) - sizeof (*th); 3843 u8 *ptr = (u8*)(th + 1); 3844 3845 /* If the TCP option is too short, we can short cut */ 3846 if (length < TCPOLEN_MD5SIG) 3847 return NULL; 3848 3849 while (length > 0) { 3850 int opcode = *ptr++; 3851 int opsize; 3852 3853 switch(opcode) { 3854 case TCPOPT_EOL: 3855 return NULL; 3856 case TCPOPT_NOP: 3857 length--; 3858 continue; 3859 default: 3860 opsize = *ptr++; 3861 if (opsize < 2 || opsize > length) 3862 return NULL; 3863 if (opcode == TCPOPT_MD5SIG) 3864 return ptr; 3865 } 3866 ptr += opsize - 2; 3867 length -= opsize; 3868 } 3869 return NULL; 3870} 3871#endif 3872 3873static inline void tcp_store_ts_recent(struct tcp_sock *tp) 3874{ 3875 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval; 3876 tp->rx_opt.ts_recent_stamp = get_seconds(); 3877} 3878 3879static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq) 3880{ 3881 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) { 3882 /* PAWS bug workaround wrt. ACK frames, the PAWS discard 3883 * extra check below makes sure this can only happen 3884 * for pure ACK frames. -DaveM 3885 * 3886 * Not only, also it occurs for expired timestamps. 3887 */ 3888 3889 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) >= 0 || 3890 get_seconds() >= tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS) 3891 tcp_store_ts_recent(tp); 3892 } 3893} 3894 3895/* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM 3896 * 3897 * It is not fatal. If this ACK does _not_ change critical state (seqs, window) 3898 * it can pass through stack. So, the following predicate verifies that 3899 * this segment is not used for anything but congestion avoidance or 3900 * fast retransmit. Moreover, we even are able to eliminate most of such 3901 * second order effects, if we apply some small "replay" window (~RTO) 3902 * to timestamp space. 3903 * 3904 * All these measures still do not guarantee that we reject wrapped ACKs 3905 * on networks with high bandwidth, when sequence space is recycled fastly, 3906 * but it guarantees that such events will be very rare and do not affect 3907 * connection seriously. This doesn't look nice, but alas, PAWS is really 3908 * buggy extension. 3909 * 3910 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC 3911 * states that events when retransmit arrives after original data are rare. 3912 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is 3913 * the biggest problem on large power networks even with minor reordering. 3914 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe 3915 * up to bandwidth of 18Gigabit/sec. 8) ] 3916 */ 3917 3918static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb) 3919{ 3920 struct tcp_sock *tp = tcp_sk(sk); 3921 struct tcphdr *th = tcp_hdr(skb); 3922 u32 seq = TCP_SKB_CB(skb)->seq; 3923 u32 ack = TCP_SKB_CB(skb)->ack_seq; 3924 3925 return (/* 1. Pure ACK with correct sequence number. */ 3926 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) && 3927 3928 /* 2. ... and duplicate ACK. */ 3929 ack == tp->snd_una && 3930 3931 /* 3. ... and does not update window. */ 3932 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) && 3933 3934 /* 4. ... and sits in replay window. */ 3935 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ); 3936} 3937 3938static inline int tcp_paws_discard(const struct sock *sk, 3939 const struct sk_buff *skb) 3940{ 3941 const struct tcp_sock *tp = tcp_sk(sk); 3942 return ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > TCP_PAWS_WINDOW && 3943 get_seconds() < tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS && 3944 !tcp_disordered_ack(sk, skb)); 3945} 3946 3947/* Check segment sequence number for validity. 3948 * 3949 * Segment controls are considered valid, if the segment 3950 * fits to the window after truncation to the window. Acceptability 3951 * of data (and SYN, FIN, of course) is checked separately. 3952 * See tcp_data_queue(), for example. 3953 * 3954 * Also, controls (RST is main one) are accepted using RCV.WUP instead 3955 * of RCV.NXT. Peer still did not advance his SND.UNA when we 3956 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP. 3957 * (borrowed from freebsd) 3958 */ 3959 3960static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq) 3961{ 3962 return !before(end_seq, tp->rcv_wup) && 3963 !after(seq, tp->rcv_nxt + tcp_receive_window(tp)); 3964} 3965 3966/* When we get a reset we do this. */ 3967static void tcp_reset(struct sock *sk) 3968{ 3969 /* We want the right error as BSD sees it (and indeed as we do). */ 3970 switch (sk->sk_state) { 3971 case TCP_SYN_SENT: 3972 sk->sk_err = ECONNREFUSED; 3973 break; 3974 case TCP_CLOSE_WAIT: 3975 sk->sk_err = EPIPE; 3976 break; 3977 case TCP_CLOSE: 3978 return; 3979 default: 3980 sk->sk_err = ECONNRESET; 3981 } 3982 3983 if (!sock_flag(sk, SOCK_DEAD)) 3984 sk->sk_error_report(sk); 3985 3986 tcp_done(sk); 3987} 3988 3989/* 3990 * Process the FIN bit. This now behaves as it is supposed to work 3991 * and the FIN takes effect when it is validly part of sequence 3992 * space. Not before when we get holes. 3993 * 3994 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT 3995 * (and thence onto LAST-ACK and finally, CLOSE, we never enter 3996 * TIME-WAIT) 3997 * 3998 * If we are in FINWAIT-1, a received FIN indicates simultaneous 3999 * close and we go into CLOSING (and later onto TIME-WAIT) 4000 * 4001 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT. 4002 */ 4003static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th) 4004{ 4005 struct tcp_sock *tp = tcp_sk(sk); 4006 4007 inet_csk_schedule_ack(sk); 4008 4009 sk->sk_shutdown |= RCV_SHUTDOWN; 4010 sock_set_flag(sk, SOCK_DONE); 4011 4012 switch (sk->sk_state) { 4013 case TCP_SYN_RECV: 4014 case TCP_ESTABLISHED: 4015 /* Move to CLOSE_WAIT */ 4016 tcp_set_state(sk, TCP_CLOSE_WAIT); 4017 inet_csk(sk)->icsk_ack.pingpong = 1; 4018 break; 4019 4020 case TCP_CLOSE_WAIT: 4021 case TCP_CLOSING: 4022 /* Received a retransmission of the FIN, do 4023 * nothing. 4024 */ 4025 break; 4026 case TCP_LAST_ACK: 4027 /* RFC793: Remain in the LAST-ACK state. */ 4028 break; 4029 4030 case TCP_FIN_WAIT1: 4031 /* This case occurs when a simultaneous close 4032 * happens, we must ack the received FIN and 4033 * enter the CLOSING state. 4034 */ 4035 tcp_send_ack(sk); 4036 tcp_set_state(sk, TCP_CLOSING); 4037 break; 4038 case TCP_FIN_WAIT2: 4039 /* Received a FIN -- send ACK and enter TIME_WAIT. */ 4040 tcp_send_ack(sk); 4041 tcp_time_wait(sk, TCP_TIME_WAIT, 0); 4042 break; 4043 default: 4044 /* Only TCP_LISTEN and TCP_CLOSE are left, in these 4045 * cases we should never reach this piece of code. 4046 */ 4047 printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n", 4048 __func__, sk->sk_state); 4049 break; 4050 } 4051 4052 /* It _is_ possible, that we have something out-of-order _after_ FIN. 4053 * Probably, we should reset in this case. For now drop them. 4054 */ 4055 __skb_queue_purge(&tp->out_of_order_queue); 4056 if (tcp_is_sack(tp)) 4057 tcp_sack_reset(&tp->rx_opt); 4058 sk_mem_reclaim(sk); 4059 4060 if (!sock_flag(sk, SOCK_DEAD)) { 4061 sk->sk_state_change(sk); 4062 4063 /* Do not send POLL_HUP for half duplex close. */ 4064 if (sk->sk_shutdown == SHUTDOWN_MASK || 4065 sk->sk_state == TCP_CLOSE) 4066 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP); 4067 else 4068 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN); 4069 } 4070} 4071 4072static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, 4073 u32 end_seq) 4074{ 4075 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) { 4076 if (before(seq, sp->start_seq)) 4077 sp->start_seq = seq; 4078 if (after(end_seq, sp->end_seq)) 4079 sp->end_seq = end_seq; 4080 return 1; 4081 } 4082 return 0; 4083} 4084 4085static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq) 4086{ 4087 struct tcp_sock *tp = tcp_sk(sk); 4088 4089 if (tcp_is_sack(tp) && sysctl_tcp_dsack) { 4090 int mib_idx; 4091 4092 if (before(seq, tp->rcv_nxt)) 4093 mib_idx = LINUX_MIB_TCPDSACKOLDSENT; 4094 else 4095 mib_idx = LINUX_MIB_TCPDSACKOFOSENT; 4096 4097 NET_INC_STATS_BH(sock_net(sk), mib_idx); 4098 4099 tp->rx_opt.dsack = 1; 4100 tp->duplicate_sack[0].start_seq = seq; 4101 tp->duplicate_sack[0].end_seq = end_seq; 4102 tp->rx_opt.eff_sacks = tp->rx_opt.num_sacks + 1; 4103 } 4104} 4105 4106static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq) 4107{ 4108 struct tcp_sock *tp = tcp_sk(sk); 4109 4110 if (!tp->rx_opt.dsack) 4111 tcp_dsack_set(sk, seq, end_seq); 4112 else 4113 tcp_sack_extend(tp->duplicate_sack, seq, end_seq); 4114} 4115 4116static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb) 4117{ 4118 struct tcp_sock *tp = tcp_sk(sk); 4119 4120 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 4121 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 4122 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST); 4123 tcp_enter_quickack_mode(sk); 4124 4125 if (tcp_is_sack(tp) && sysctl_tcp_dsack) { 4126 u32 end_seq = TCP_SKB_CB(skb)->end_seq; 4127 4128 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) 4129 end_seq = tp->rcv_nxt; 4130 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq); 4131 } 4132 } 4133 4134 tcp_send_ack(sk); 4135} 4136 4137/* These routines update the SACK block as out-of-order packets arrive or 4138 * in-order packets close up the sequence space. 4139 */ 4140static void tcp_sack_maybe_coalesce(struct tcp_sock *tp) 4141{ 4142 int this_sack; 4143 struct tcp_sack_block *sp = &tp->selective_acks[0]; 4144 struct tcp_sack_block *swalk = sp + 1; 4145 4146 /* See if the recent change to the first SACK eats into 4147 * or hits the sequence space of other SACK blocks, if so coalesce. 4148 */ 4149 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) { 4150 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) { 4151 int i; 4152 4153 /* Zap SWALK, by moving every further SACK up by one slot. 4154 * Decrease num_sacks. 4155 */ 4156 tp->rx_opt.num_sacks--; 4157 tp->rx_opt.eff_sacks = tp->rx_opt.num_sacks + 4158 tp->rx_opt.dsack; 4159 for (i = this_sack; i < tp->rx_opt.num_sacks; i++) 4160 sp[i] = sp[i + 1]; 4161 continue; 4162 } 4163 this_sack++, swalk++; 4164 } 4165} 4166 4167static inline void tcp_sack_swap(struct tcp_sack_block *sack1, 4168 struct tcp_sack_block *sack2) 4169{ 4170 __u32 tmp; 4171 4172 tmp = sack1->start_seq; 4173 sack1->start_seq = sack2->start_seq; 4174 sack2->start_seq = tmp; 4175 4176 tmp = sack1->end_seq; 4177 sack1->end_seq = sack2->end_seq; 4178 sack2->end_seq = tmp; 4179} 4180 4181static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq) 4182{ 4183 struct tcp_sock *tp = tcp_sk(sk); 4184 struct tcp_sack_block *sp = &tp->selective_acks[0]; 4185 int cur_sacks = tp->rx_opt.num_sacks; 4186 int this_sack; 4187 4188 if (!cur_sacks) 4189 goto new_sack; 4190 4191 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) { 4192 if (tcp_sack_extend(sp, seq, end_seq)) { 4193 /* Rotate this_sack to the first one. */ 4194 for (; this_sack > 0; this_sack--, sp--) 4195 tcp_sack_swap(sp, sp - 1); 4196 if (cur_sacks > 1) 4197 tcp_sack_maybe_coalesce(tp); 4198 return; 4199 } 4200 } 4201 4202 /* Could not find an adjacent existing SACK, build a new one, 4203 * put it at the front, and shift everyone else down. We 4204 * always know there is at least one SACK present already here. 4205 * 4206 * If the sack array is full, forget about the last one. 4207 */ 4208 if (this_sack >= TCP_NUM_SACKS) { 4209 this_sack--; 4210 tp->rx_opt.num_sacks--; 4211 sp--; 4212 } 4213 for (; this_sack > 0; this_sack--, sp--) 4214 *sp = *(sp - 1); 4215 4216new_sack: 4217 /* Build the new head SACK, and we're done. */ 4218 sp->start_seq = seq; 4219 sp->end_seq = end_seq; 4220 tp->rx_opt.num_sacks++; 4221 tp->rx_opt.eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; 4222} 4223 4224/* RCV.NXT advances, some SACKs should be eaten. */ 4225 4226static void tcp_sack_remove(struct tcp_sock *tp) 4227{ 4228 struct tcp_sack_block *sp = &tp->selective_acks[0]; 4229 int num_sacks = tp->rx_opt.num_sacks; 4230 int this_sack; 4231 4232 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */ 4233 if (skb_queue_empty(&tp->out_of_order_queue)) { 4234 tp->rx_opt.num_sacks = 0; 4235 tp->rx_opt.eff_sacks = tp->rx_opt.dsack; 4236 return; 4237 } 4238 4239 for (this_sack = 0; this_sack < num_sacks;) { 4240 /* Check if the start of the sack is covered by RCV.NXT. */ 4241 if (!before(tp->rcv_nxt, sp->start_seq)) { 4242 int i; 4243 4244 /* RCV.NXT must cover all the block! */ 4245 WARN_ON(before(tp->rcv_nxt, sp->end_seq)); 4246 4247 /* Zap this SACK, by moving forward any other SACKS. */ 4248 for (i=this_sack+1; i < num_sacks; i++) 4249 tp->selective_acks[i-1] = tp->selective_acks[i]; 4250 num_sacks--; 4251 continue; 4252 } 4253 this_sack++; 4254 sp++; 4255 } 4256 if (num_sacks != tp->rx_opt.num_sacks) { 4257 tp->rx_opt.num_sacks = num_sacks; 4258 tp->rx_opt.eff_sacks = tp->rx_opt.num_sacks + 4259 tp->rx_opt.dsack; 4260 } 4261} 4262 4263/* This one checks to see if we can put data from the 4264 * out_of_order queue into the receive_queue. 4265 */ 4266static void tcp_ofo_queue(struct sock *sk) 4267{ 4268 struct tcp_sock *tp = tcp_sk(sk); 4269 __u32 dsack_high = tp->rcv_nxt; 4270 struct sk_buff *skb; 4271 4272 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) { 4273 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) 4274 break; 4275 4276 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) { 4277 __u32 dsack = dsack_high; 4278 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high)) 4279 dsack_high = TCP_SKB_CB(skb)->end_seq; 4280 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack); 4281 } 4282 4283 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) { 4284 SOCK_DEBUG(sk, "ofo packet was already received \n"); 4285 __skb_unlink(skb, &tp->out_of_order_queue); 4286 __kfree_skb(skb); 4287 continue; 4288 } 4289 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n", 4290 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, 4291 TCP_SKB_CB(skb)->end_seq); 4292 4293 __skb_unlink(skb, &tp->out_of_order_queue); 4294 __skb_queue_tail(&sk->sk_receive_queue, skb); 4295 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; 4296 if (tcp_hdr(skb)->fin) 4297 tcp_fin(skb, sk, tcp_hdr(skb)); 4298 } 4299} 4300 4301static int tcp_prune_ofo_queue(struct sock *sk); 4302static int tcp_prune_queue(struct sock *sk); 4303 4304static inline int tcp_try_rmem_schedule(struct sock *sk, unsigned int size) 4305{ 4306 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf || 4307 !sk_rmem_schedule(sk, size)) { 4308 4309 if (tcp_prune_queue(sk) < 0) 4310 return -1; 4311 4312 if (!sk_rmem_schedule(sk, size)) { 4313 if (!tcp_prune_ofo_queue(sk)) 4314 return -1; 4315 4316 if (!sk_rmem_schedule(sk, size)) 4317 return -1; 4318 } 4319 } 4320 return 0; 4321} 4322 4323static void tcp_data_queue(struct sock *sk, struct sk_buff *skb) 4324{ 4325 struct tcphdr *th = tcp_hdr(skb); 4326 struct tcp_sock *tp = tcp_sk(sk); 4327 int eaten = -1; 4328 4329 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) 4330 goto drop; 4331 4332 __skb_pull(skb, th->doff * 4); 4333 4334 TCP_ECN_accept_cwr(tp, skb); 4335 4336 if (tp->rx_opt.dsack) { 4337 tp->rx_opt.dsack = 0; 4338 tp->rx_opt.eff_sacks = tp->rx_opt.num_sacks; 4339 } 4340 4341 /* Queue data for delivery to the user. 4342 * Packets in sequence go to the receive queue. 4343 * Out of sequence packets to the out_of_order_queue. 4344 */ 4345 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) { 4346 if (tcp_receive_window(tp) == 0) 4347 goto out_of_window; 4348 4349 /* Ok. In sequence. In window. */ 4350 if (tp->ucopy.task == current && 4351 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len && 4352 sock_owned_by_user(sk) && !tp->urg_data) { 4353 int chunk = min_t(unsigned int, skb->len, 4354 tp->ucopy.len); 4355 4356 __set_current_state(TASK_RUNNING); 4357 4358 local_bh_enable(); 4359 if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) { 4360 tp->ucopy.len -= chunk; 4361 tp->copied_seq += chunk; 4362 eaten = (chunk == skb->len && !th->fin); 4363 tcp_rcv_space_adjust(sk); 4364 } 4365 local_bh_disable(); 4366 } 4367 4368 if (eaten <= 0) { 4369queue_and_out: 4370 if (eaten < 0 && 4371 tcp_try_rmem_schedule(sk, skb->truesize)) 4372 goto drop; 4373 4374 skb_set_owner_r(skb, sk); 4375 __skb_queue_tail(&sk->sk_receive_queue, skb); 4376 } 4377 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; 4378 if (skb->len) 4379 tcp_event_data_recv(sk, skb); 4380 if (th->fin) 4381 tcp_fin(skb, sk, th); 4382 4383 if (!skb_queue_empty(&tp->out_of_order_queue)) { 4384 tcp_ofo_queue(sk); 4385 4386 /* RFC2581. 4.2. SHOULD send immediate ACK, when 4387 * gap in queue is filled. 4388 */ 4389 if (skb_queue_empty(&tp->out_of_order_queue)) 4390 inet_csk(sk)->icsk_ack.pingpong = 0; 4391 } 4392 4393 if (tp->rx_opt.num_sacks) 4394 tcp_sack_remove(tp); 4395 4396 tcp_fast_path_check(sk); 4397 4398 if (eaten > 0) 4399 __kfree_skb(skb); 4400 else if (!sock_flag(sk, SOCK_DEAD)) 4401 sk->sk_data_ready(sk, 0); 4402 return; 4403 } 4404 4405 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) { 4406 /* A retransmit, 2nd most common case. Force an immediate ack. */ 4407 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST); 4408 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); 4409 4410out_of_window: 4411 tcp_enter_quickack_mode(sk); 4412 inet_csk_schedule_ack(sk); 4413drop: 4414 __kfree_skb(skb); 4415 return; 4416 } 4417 4418 /* Out of window. F.e. zero window probe. */ 4419 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp))) 4420 goto out_of_window; 4421 4422 tcp_enter_quickack_mode(sk); 4423 4424 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 4425 /* Partial packet, seq < rcv_next < end_seq */ 4426 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n", 4427 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, 4428 TCP_SKB_CB(skb)->end_seq); 4429 4430 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt); 4431 4432 /* If window is closed, drop tail of packet. But after 4433 * remembering D-SACK for its head made in previous line. 4434 */ 4435 if (!tcp_receive_window(tp)) 4436 goto out_of_window; 4437 goto queue_and_out; 4438 } 4439 4440 TCP_ECN_check_ce(tp, skb); 4441 4442 if (tcp_try_rmem_schedule(sk, skb->truesize)) 4443 goto drop; 4444 4445 /* Disable header prediction. */ 4446 tp->pred_flags = 0; 4447 inet_csk_schedule_ack(sk); 4448 4449 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n", 4450 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); 4451 4452 skb_set_owner_r(skb, sk); 4453 4454 if (!skb_peek(&tp->out_of_order_queue)) { 4455 /* Initial out of order segment, build 1 SACK. */ 4456 if (tcp_is_sack(tp)) { 4457 tp->rx_opt.num_sacks = 1; 4458 tp->rx_opt.dsack = 0; 4459 tp->rx_opt.eff_sacks = 1; 4460 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq; 4461 tp->selective_acks[0].end_seq = 4462 TCP_SKB_CB(skb)->end_seq; 4463 } 4464 __skb_queue_head(&tp->out_of_order_queue, skb); 4465 } else { 4466 struct sk_buff *skb1 = tp->out_of_order_queue.prev; 4467 u32 seq = TCP_SKB_CB(skb)->seq; 4468 u32 end_seq = TCP_SKB_CB(skb)->end_seq; 4469 4470 if (seq == TCP_SKB_CB(skb1)->end_seq) { 4471 __skb_queue_after(&tp->out_of_order_queue, skb1, skb); 4472 4473 if (!tp->rx_opt.num_sacks || 4474 tp->selective_acks[0].end_seq != seq) 4475 goto add_sack; 4476 4477 /* Common case: data arrive in order after hole. */ 4478 tp->selective_acks[0].end_seq = end_seq; 4479 return; 4480 } 4481 4482 /* Find place to insert this segment. */ 4483 do { 4484 if (!after(TCP_SKB_CB(skb1)->seq, seq)) 4485 break; 4486 } while ((skb1 = skb1->prev) != 4487 (struct sk_buff *)&tp->out_of_order_queue); 4488 4489 /* Do skb overlap to previous one? */ 4490 if (skb1 != (struct sk_buff *)&tp->out_of_order_queue && 4491 before(seq, TCP_SKB_CB(skb1)->end_seq)) { 4492 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) { 4493 /* All the bits are present. Drop. */ 4494 __kfree_skb(skb); 4495 tcp_dsack_set(sk, seq, end_seq); 4496 goto add_sack; 4497 } 4498 if (after(seq, TCP_SKB_CB(skb1)->seq)) { 4499 /* Partial overlap. */ 4500 tcp_dsack_set(sk, seq, 4501 TCP_SKB_CB(skb1)->end_seq); 4502 } else { 4503 skb1 = skb1->prev; 4504 } 4505 } 4506 __skb_queue_after(&tp->out_of_order_queue, skb1, skb); 4507 4508 /* And clean segments covered by new one as whole. */ 4509 while ((skb1 = skb->next) != 4510 (struct sk_buff *)&tp->out_of_order_queue && 4511 after(end_seq, TCP_SKB_CB(skb1)->seq)) { 4512 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) { 4513 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq, 4514 end_seq); 4515 break; 4516 } 4517 __skb_unlink(skb1, &tp->out_of_order_queue); 4518 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq, 4519 TCP_SKB_CB(skb1)->end_seq); 4520 __kfree_skb(skb1); 4521 } 4522 4523add_sack: 4524 if (tcp_is_sack(tp)) 4525 tcp_sack_new_ofo_skb(sk, seq, end_seq); 4526 } 4527} 4528 4529static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb, 4530 struct sk_buff_head *list) 4531{ 4532 struct sk_buff *next = skb->next; 4533 4534 __skb_unlink(skb, list); 4535 __kfree_skb(skb); 4536 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED); 4537 4538 return next; 4539} 4540 4541/* Collapse contiguous sequence of skbs head..tail with 4542 * sequence numbers start..end. 4543 * Segments with FIN/SYN are not collapsed (only because this 4544 * simplifies code) 4545 */ 4546static void 4547tcp_collapse(struct sock *sk, struct sk_buff_head *list, 4548 struct sk_buff *head, struct sk_buff *tail, 4549 u32 start, u32 end) 4550{ 4551 struct sk_buff *skb; 4552 4553 /* First, check that queue is collapsible and find 4554 * the point where collapsing can be useful. */ 4555 for (skb = head; skb != tail;) { 4556 /* No new bits? It is possible on ofo queue. */ 4557 if (!before(start, TCP_SKB_CB(skb)->end_seq)) { 4558 skb = tcp_collapse_one(sk, skb, list); 4559 continue; 4560 } 4561 4562 /* The first skb to collapse is: 4563 * - not SYN/FIN and 4564 * - bloated or contains data before "start" or 4565 * overlaps to the next one. 4566 */ 4567 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin && 4568 (tcp_win_from_space(skb->truesize) > skb->len || 4569 before(TCP_SKB_CB(skb)->seq, start) || 4570 (skb->next != tail && 4571 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq))) 4572 break; 4573 4574 /* Decided to skip this, advance start seq. */ 4575 start = TCP_SKB_CB(skb)->end_seq; 4576 skb = skb->next; 4577 } 4578 if (skb == tail || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin) 4579 return; 4580 4581 while (before(start, end)) { 4582 struct sk_buff *nskb; 4583 unsigned int header = skb_headroom(skb); 4584 int copy = SKB_MAX_ORDER(header, 0); 4585 4586 /* Too big header? This can happen with IPv6. */ 4587 if (copy < 0) 4588 return; 4589 if (end - start < copy) 4590 copy = end - start; 4591 nskb = alloc_skb(copy + header, GFP_ATOMIC); 4592 if (!nskb) 4593 return; 4594 4595 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head); 4596 skb_set_network_header(nskb, (skb_network_header(skb) - 4597 skb->head)); 4598 skb_set_transport_header(nskb, (skb_transport_header(skb) - 4599 skb->head)); 4600 skb_reserve(nskb, header); 4601 memcpy(nskb->head, skb->head, header); 4602 memcpy(nskb->cb, skb->cb, sizeof(skb->cb)); 4603 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start; 4604 __skb_queue_before(list, skb, nskb); 4605 skb_set_owner_r(nskb, sk); 4606 4607 /* Copy data, releasing collapsed skbs. */ 4608 while (copy > 0) { 4609 int offset = start - TCP_SKB_CB(skb)->seq; 4610 int size = TCP_SKB_CB(skb)->end_seq - start; 4611 4612 BUG_ON(offset < 0); 4613 if (size > 0) { 4614 size = min(copy, size); 4615 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size)) 4616 BUG(); 4617 TCP_SKB_CB(nskb)->end_seq += size; 4618 copy -= size; 4619 start += size; 4620 } 4621 if (!before(start, TCP_SKB_CB(skb)->end_seq)) { 4622 skb = tcp_collapse_one(sk, skb, list); 4623 if (skb == tail || 4624 tcp_hdr(skb)->syn || 4625 tcp_hdr(skb)->fin) 4626 return; 4627 } 4628 } 4629 } 4630} 4631 4632/* Collapse ofo queue. Algorithm: select contiguous sequence of skbs 4633 * and tcp_collapse() them until all the queue is collapsed. 4634 */ 4635static void tcp_collapse_ofo_queue(struct sock *sk) 4636{ 4637 struct tcp_sock *tp = tcp_sk(sk); 4638 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue); 4639 struct sk_buff *head; 4640 u32 start, end; 4641 4642 if (skb == NULL) 4643 return; 4644 4645 start = TCP_SKB_CB(skb)->seq; 4646 end = TCP_SKB_CB(skb)->end_seq; 4647 head = skb; 4648 4649 for (;;) { 4650 skb = skb->next; 4651 4652 /* Segment is terminated when we see gap or when 4653 * we are at the end of all the queue. */ 4654 if (skb == (struct sk_buff *)&tp->out_of_order_queue || 4655 after(TCP_SKB_CB(skb)->seq, end) || 4656 before(TCP_SKB_CB(skb)->end_seq, start)) { 4657 tcp_collapse(sk, &tp->out_of_order_queue, 4658 head, skb, start, end); 4659 head = skb; 4660 if (skb == (struct sk_buff *)&tp->out_of_order_queue) 4661 break; 4662 /* Start new segment */ 4663 start = TCP_SKB_CB(skb)->seq; 4664 end = TCP_SKB_CB(skb)->end_seq; 4665 } else { 4666 if (before(TCP_SKB_CB(skb)->seq, start)) 4667 start = TCP_SKB_CB(skb)->seq; 4668 if (after(TCP_SKB_CB(skb)->end_seq, end)) 4669 end = TCP_SKB_CB(skb)->end_seq; 4670 } 4671 } 4672} 4673 4674/* 4675 * Purge the out-of-order queue. 4676 * Return true if queue was pruned. 4677 */ 4678static int tcp_prune_ofo_queue(struct sock *sk) 4679{ 4680 struct tcp_sock *tp = tcp_sk(sk); 4681 int res = 0; 4682 4683 if (!skb_queue_empty(&tp->out_of_order_queue)) { 4684 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED); 4685 __skb_queue_purge(&tp->out_of_order_queue); 4686 4687 /* Reset SACK state. A conforming SACK implementation will 4688 * do the same at a timeout based retransmit. When a connection 4689 * is in a sad state like this, we care only about integrity 4690 * of the connection not performance. 4691 */ 4692 if (tp->rx_opt.sack_ok) 4693 tcp_sack_reset(&tp->rx_opt); 4694 sk_mem_reclaim(sk); 4695 res = 1; 4696 } 4697 return res; 4698} 4699 4700/* Reduce allocated memory if we can, trying to get 4701 * the socket within its memory limits again. 4702 * 4703 * Return less than zero if we should start dropping frames 4704 * until the socket owning process reads some of the data 4705 * to stabilize the situation. 4706 */ 4707static int tcp_prune_queue(struct sock *sk) 4708{ 4709 struct tcp_sock *tp = tcp_sk(sk); 4710 4711 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq); 4712 4713 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED); 4714 4715 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf) 4716 tcp_clamp_window(sk); 4717 else if (tcp_memory_pressure) 4718 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss); 4719 4720 tcp_collapse_ofo_queue(sk); 4721 tcp_collapse(sk, &sk->sk_receive_queue, 4722 sk->sk_receive_queue.next, 4723 (struct sk_buff *)&sk->sk_receive_queue, 4724 tp->copied_seq, tp->rcv_nxt); 4725 sk_mem_reclaim(sk); 4726 4727 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) 4728 return 0; 4729 4730 /* Collapsing did not help, destructive actions follow. 4731 * This must not ever occur. */ 4732 4733 tcp_prune_ofo_queue(sk); 4734 4735 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) 4736 return 0; 4737 4738 /* If we are really being abused, tell the caller to silently 4739 * drop receive data on the floor. It will get retransmitted 4740 * and hopefully then we'll have sufficient space. 4741 */ 4742 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED); 4743 4744 /* Massive buffer overcommit. */ 4745 tp->pred_flags = 0; 4746 return -1; 4747} 4748 4749/* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. 4750 * As additional protections, we do not touch cwnd in retransmission phases, 4751 * and if application hit its sndbuf limit recently. 4752 */ 4753void tcp_cwnd_application_limited(struct sock *sk) 4754{ 4755 struct tcp_sock *tp = tcp_sk(sk); 4756 4757 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && 4758 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { 4759 /* Limited by application or receiver window. */ 4760 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); 4761 u32 win_used = max(tp->snd_cwnd_used, init_win); 4762 if (win_used < tp->snd_cwnd) { 4763 tp->snd_ssthresh = tcp_current_ssthresh(sk); 4764 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1; 4765 } 4766 tp->snd_cwnd_used = 0; 4767 } 4768 tp->snd_cwnd_stamp = tcp_time_stamp; 4769} 4770 4771static int tcp_should_expand_sndbuf(struct sock *sk) 4772{ 4773 struct tcp_sock *tp = tcp_sk(sk); 4774 4775 /* If the user specified a specific send buffer setting, do 4776 * not modify it. 4777 */ 4778 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK) 4779 return 0; 4780 4781 /* If we are under global TCP memory pressure, do not expand. */ 4782 if (tcp_memory_pressure) 4783 return 0; 4784 4785 /* If we are under soft global TCP memory pressure, do not expand. */ 4786 if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0]) 4787 return 0; 4788 4789 /* If we filled the congestion window, do not expand. */ 4790 if (tp->packets_out >= tp->snd_cwnd) 4791 return 0; 4792 4793 return 1; 4794} 4795 4796/* When incoming ACK allowed to free some skb from write_queue, 4797 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket 4798 * on the exit from tcp input handler. 4799 * 4800 * PROBLEM: sndbuf expansion does not work well with largesend. 4801 */ 4802static void tcp_new_space(struct sock *sk) 4803{ 4804 struct tcp_sock *tp = tcp_sk(sk); 4805 4806 if (tcp_should_expand_sndbuf(sk)) { 4807 int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) + 4808 MAX_TCP_HEADER + 16 + sizeof(struct sk_buff); 4809 int demanded = max_t(unsigned int, tp->snd_cwnd, 4810 tp->reordering + 1); 4811 sndmem *= 2 * demanded; 4812 if (sndmem > sk->sk_sndbuf) 4813 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]); 4814 tp->snd_cwnd_stamp = tcp_time_stamp; 4815 } 4816 4817 sk->sk_write_space(sk); 4818} 4819 4820static void tcp_check_space(struct sock *sk) 4821{ 4822 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) { 4823 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK); 4824 if (sk->sk_socket && 4825 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) 4826 tcp_new_space(sk); 4827 } 4828} 4829 4830static inline void tcp_data_snd_check(struct sock *sk) 4831{ 4832 tcp_push_pending_frames(sk); 4833 tcp_check_space(sk); 4834} 4835 4836/* 4837 * Check if sending an ack is needed. 4838 */ 4839static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible) 4840{ 4841 struct tcp_sock *tp = tcp_sk(sk); 4842 4843 /* More than one full frame received... */ 4844 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss 4845 /* ... and right edge of window advances far enough. 4846 * (tcp_recvmsg() will send ACK otherwise). Or... 4847 */ 4848 && __tcp_select_window(sk) >= tp->rcv_wnd) || 4849 /* We ACK each frame or... */ 4850 tcp_in_quickack_mode(sk) || 4851 /* We have out of order data. */ 4852 (ofo_possible && skb_peek(&tp->out_of_order_queue))) { 4853 /* Then ack it now */ 4854 tcp_send_ack(sk); 4855 } else { 4856 /* Else, send delayed ack. */ 4857 tcp_send_delayed_ack(sk); 4858 } 4859} 4860 4861static inline void tcp_ack_snd_check(struct sock *sk) 4862{ 4863 if (!inet_csk_ack_scheduled(sk)) { 4864 /* We sent a data segment already. */ 4865 return; 4866 } 4867 __tcp_ack_snd_check(sk, 1); 4868} 4869 4870/* 4871 * This routine is only called when we have urgent data 4872 * signaled. Its the 'slow' part of tcp_urg. It could be 4873 * moved inline now as tcp_urg is only called from one 4874 * place. We handle URGent data wrong. We have to - as 4875 * BSD still doesn't use the correction from RFC961. 4876 * For 1003.1g we should support a new option TCP_STDURG to permit 4877 * either form (or just set the sysctl tcp_stdurg). 4878 */ 4879 4880static void tcp_check_urg(struct sock *sk, struct tcphdr *th) 4881{ 4882 struct tcp_sock *tp = tcp_sk(sk); 4883 u32 ptr = ntohs(th->urg_ptr); 4884 4885 if (ptr && !sysctl_tcp_stdurg) 4886 ptr--; 4887 ptr += ntohl(th->seq); 4888 4889 /* Ignore urgent data that we've already seen and read. */ 4890 if (after(tp->copied_seq, ptr)) 4891 return; 4892 4893 /* Do not replay urg ptr. 4894 * 4895 * NOTE: interesting situation not covered by specs. 4896 * Misbehaving sender may send urg ptr, pointing to segment, 4897 * which we already have in ofo queue. We are not able to fetch 4898 * such data and will stay in TCP_URG_NOTYET until will be eaten 4899 * by recvmsg(). Seems, we are not obliged to handle such wicked 4900 * situations. But it is worth to think about possibility of some 4901 * DoSes using some hypothetical application level deadlock. 4902 */ 4903 if (before(ptr, tp->rcv_nxt)) 4904 return; 4905 4906 /* Do we already have a newer (or duplicate) urgent pointer? */ 4907 if (tp->urg_data && !after(ptr, tp->urg_seq)) 4908 return; 4909 4910 /* Tell the world about our new urgent pointer. */ 4911 sk_send_sigurg(sk); 4912 4913 /* We may be adding urgent data when the last byte read was 4914 * urgent. To do this requires some care. We cannot just ignore 4915 * tp->copied_seq since we would read the last urgent byte again 4916 * as data, nor can we alter copied_seq until this data arrives 4917 * or we break the semantics of SIOCATMARK (and thus sockatmark()) 4918 * 4919 * NOTE. Double Dutch. Rendering to plain English: author of comment 4920 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB); 4921 * and expect that both A and B disappear from stream. This is _wrong_. 4922 * Though this happens in BSD with high probability, this is occasional. 4923 * Any application relying on this is buggy. Note also, that fix "works" 4924 * only in this artificial test. Insert some normal data between A and B and we will 4925 * decline of BSD again. Verdict: it is better to remove to trap 4926 * buggy users. 4927 */ 4928 if (tp->urg_seq == tp->copied_seq && tp->urg_data && 4929 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) { 4930 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue); 4931 tp->copied_seq++; 4932 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) { 4933 __skb_unlink(skb, &sk->sk_receive_queue); 4934 __kfree_skb(skb); 4935 } 4936 } 4937 4938 tp->urg_data = TCP_URG_NOTYET; 4939 tp->urg_seq = ptr; 4940 4941 /* Disable header prediction. */ 4942 tp->pred_flags = 0; 4943} 4944 4945/* This is the 'fast' part of urgent handling. */ 4946static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th) 4947{ 4948 struct tcp_sock *tp = tcp_sk(sk); 4949 4950 /* Check if we get a new urgent pointer - normally not. */ 4951 if (th->urg) 4952 tcp_check_urg(sk, th); 4953 4954 /* Do we wait for any urgent data? - normally not... */ 4955 if (tp->urg_data == TCP_URG_NOTYET) { 4956 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) - 4957 th->syn; 4958 4959 /* Is the urgent pointer pointing into this packet? */ 4960 if (ptr < skb->len) { 4961 u8 tmp; 4962 if (skb_copy_bits(skb, ptr, &tmp, 1)) 4963 BUG(); 4964 tp->urg_data = TCP_URG_VALID | tmp; 4965 if (!sock_flag(sk, SOCK_DEAD)) 4966 sk->sk_data_ready(sk, 0); 4967 } 4968 } 4969} 4970 4971static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen) 4972{ 4973 struct tcp_sock *tp = tcp_sk(sk); 4974 int chunk = skb->len - hlen; 4975 int err; 4976 4977 local_bh_enable(); 4978 if (skb_csum_unnecessary(skb)) 4979 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk); 4980 else 4981 err = skb_copy_and_csum_datagram_iovec(skb, hlen, 4982 tp->ucopy.iov); 4983 4984 if (!err) { 4985 tp->ucopy.len -= chunk; 4986 tp->copied_seq += chunk; 4987 tcp_rcv_space_adjust(sk); 4988 } 4989 4990 local_bh_disable(); 4991 return err; 4992} 4993 4994static __sum16 __tcp_checksum_complete_user(struct sock *sk, 4995 struct sk_buff *skb) 4996{ 4997 __sum16 result; 4998 4999 if (sock_owned_by_user(sk)) { 5000 local_bh_enable(); 5001 result = __tcp_checksum_complete(skb); 5002 local_bh_disable(); 5003 } else { 5004 result = __tcp_checksum_complete(skb); 5005 } 5006 return result; 5007} 5008 5009static inline int tcp_checksum_complete_user(struct sock *sk, 5010 struct sk_buff *skb) 5011{ 5012 return !skb_csum_unnecessary(skb) && 5013 __tcp_checksum_complete_user(sk, skb); 5014} 5015 5016#ifdef CONFIG_NET_DMA 5017static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb, 5018 int hlen) 5019{ 5020 struct tcp_sock *tp = tcp_sk(sk); 5021 int chunk = skb->len - hlen; 5022 int dma_cookie; 5023 int copied_early = 0; 5024 5025 if (tp->ucopy.wakeup) 5026 return 0; 5027 5028 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list) 5029 tp->ucopy.dma_chan = dma_find_channel(DMA_MEMCPY); 5030 5031 if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) { 5032 5033 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan, 5034 skb, hlen, 5035 tp->ucopy.iov, chunk, 5036 tp->ucopy.pinned_list); 5037 5038 if (dma_cookie < 0) 5039 goto out; 5040 5041 tp->ucopy.dma_cookie = dma_cookie; 5042 copied_early = 1; 5043 5044 tp->ucopy.len -= chunk; 5045 tp->copied_seq += chunk; 5046 tcp_rcv_space_adjust(sk); 5047 5048 if ((tp->ucopy.len == 0) || 5049 (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) || 5050 (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) { 5051 tp->ucopy.wakeup = 1; 5052 sk->sk_data_ready(sk, 0); 5053 } 5054 } else if (chunk > 0) { 5055 tp->ucopy.wakeup = 1; 5056 sk->sk_data_ready(sk, 0); 5057 } 5058out: 5059 return copied_early; 5060} 5061#endif /* CONFIG_NET_DMA */ 5062 5063/* Does PAWS and seqno based validation of an incoming segment, flags will 5064 * play significant role here. 5065 */ 5066static int tcp_validate_incoming(struct sock *sk, struct sk_buff *skb, 5067 struct tcphdr *th, int syn_inerr) 5068{ 5069 struct tcp_sock *tp = tcp_sk(sk); 5070 5071 /* RFC1323: H1. Apply PAWS check first. */ 5072 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp && 5073 tcp_paws_discard(sk, skb)) { 5074 if (!th->rst) { 5075 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED); 5076 tcp_send_dupack(sk, skb); 5077 goto discard; 5078 } 5079 /* Reset is accepted even if it did not pass PAWS. */ 5080 } 5081 5082 /* Step 1: check sequence number */ 5083 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) { 5084 /* RFC793, page 37: "In all states except SYN-SENT, all reset 5085 * (RST) segments are validated by checking their SEQ-fields." 5086 * And page 69: "If an incoming segment is not acceptable, 5087 * an acknowledgment should be sent in reply (unless the RST 5088 * bit is set, if so drop the segment and return)". 5089 */ 5090 if (!th->rst) 5091 tcp_send_dupack(sk, skb); 5092 goto discard; 5093 } 5094 5095 /* Step 2: check RST bit */ 5096 if (th->rst) { 5097 tcp_reset(sk); 5098 goto discard; 5099 } 5100 5101 /* ts_recent update must be made after we are sure that the packet 5102 * is in window. 5103 */ 5104 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq); 5105 5106 /* step 3: check security and precedence [ignored] */ 5107 5108 /* step 4: Check for a SYN in window. */ 5109 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 5110 if (syn_inerr) 5111 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS); 5112 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONSYN); 5113 tcp_reset(sk); 5114 return -1; 5115 } 5116 5117 return 1; 5118 5119discard: 5120 __kfree_skb(skb); 5121 return 0; 5122} 5123 5124/* 5125 * TCP receive function for the ESTABLISHED state. 5126 * 5127 * It is split into a fast path and a slow path. The fast path is 5128 * disabled when: 5129 * - A zero window was announced from us - zero window probing 5130 * is only handled properly in the slow path. 5131 * - Out of order segments arrived. 5132 * - Urgent data is expected. 5133 * - There is no buffer space left 5134 * - Unexpected TCP flags/window values/header lengths are received 5135 * (detected by checking the TCP header against pred_flags) 5136 * - Data is sent in both directions. Fast path only supports pure senders 5137 * or pure receivers (this means either the sequence number or the ack 5138 * value must stay constant) 5139 * - Unexpected TCP option. 5140 * 5141 * When these conditions are not satisfied it drops into a standard 5142 * receive procedure patterned after RFC793 to handle all cases. 5143 * The first three cases are guaranteed by proper pred_flags setting, 5144 * the rest is checked inline. Fast processing is turned on in 5145 * tcp_data_queue when everything is OK. 5146 */ 5147int tcp_rcv_established(struct sock *sk, struct sk_buff *skb, 5148 struct tcphdr *th, unsigned len) 5149{ 5150 struct tcp_sock *tp = tcp_sk(sk); 5151 int res; 5152 5153 /* 5154 * Header prediction. 5155 * The code loosely follows the one in the famous 5156 * "30 instruction TCP receive" Van Jacobson mail. 5157 * 5158 * Van's trick is to deposit buffers into socket queue 5159 * on a device interrupt, to call tcp_recv function 5160 * on the receive process context and checksum and copy 5161 * the buffer to user space. smart... 5162 * 5163 * Our current scheme is not silly either but we take the 5164 * extra cost of the net_bh soft interrupt processing... 5165 * We do checksum and copy also but from device to kernel. 5166 */ 5167 5168 tp->rx_opt.saw_tstamp = 0; 5169 5170 /* pred_flags is 0xS?10 << 16 + snd_wnd 5171 * if header_prediction is to be made 5172 * 'S' will always be tp->tcp_header_len >> 2 5173 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to 5174 * turn it off (when there are holes in the receive 5175 * space for instance) 5176 * PSH flag is ignored. 5177 */ 5178 5179 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags && 5180 TCP_SKB_CB(skb)->seq == tp->rcv_nxt) { 5181 int tcp_header_len = tp->tcp_header_len; 5182 5183 /* Timestamp header prediction: tcp_header_len 5184 * is automatically equal to th->doff*4 due to pred_flags 5185 * match. 5186 */ 5187 5188 /* Check timestamp */ 5189 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) { 5190 /* No? Slow path! */ 5191 if (!tcp_parse_aligned_timestamp(tp, th)) 5192 goto slow_path; 5193 5194 /* If PAWS failed, check it more carefully in slow path */ 5195 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0) 5196 goto slow_path; 5197 5198 /* DO NOT update ts_recent here, if checksum fails 5199 * and timestamp was corrupted part, it will result 5200 * in a hung connection since we will drop all 5201 * future packets due to the PAWS test. 5202 */ 5203 } 5204 5205 if (len <= tcp_header_len) { 5206 /* Bulk data transfer: sender */ 5207 if (len == tcp_header_len) { 5208 /* Predicted packet is in window by definition. 5209 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 5210 * Hence, check seq<=rcv_wup reduces to: 5211 */ 5212 if (tcp_header_len == 5213 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && 5214 tp->rcv_nxt == tp->rcv_wup) 5215 tcp_store_ts_recent(tp); 5216 5217 /* We know that such packets are checksummed 5218 * on entry. 5219 */ 5220 tcp_ack(sk, skb, 0); 5221 __kfree_skb(skb); 5222 tcp_data_snd_check(sk); 5223 return 0; 5224 } else { /* Header too small */ 5225 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS); 5226 goto discard; 5227 } 5228 } else { 5229 int eaten = 0; 5230 int copied_early = 0; 5231 5232 if (tp->copied_seq == tp->rcv_nxt && 5233 len - tcp_header_len <= tp->ucopy.len) { 5234#ifdef CONFIG_NET_DMA 5235 if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) { 5236 copied_early = 1; 5237 eaten = 1; 5238 } 5239#endif 5240 if (tp->ucopy.task == current && 5241 sock_owned_by_user(sk) && !copied_early) { 5242 __set_current_state(TASK_RUNNING); 5243 5244 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len)) 5245 eaten = 1; 5246 } 5247 if (eaten) { 5248 /* Predicted packet is in window by definition. 5249 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 5250 * Hence, check seq<=rcv_wup reduces to: 5251 */ 5252 if (tcp_header_len == 5253 (sizeof(struct tcphdr) + 5254 TCPOLEN_TSTAMP_ALIGNED) && 5255 tp->rcv_nxt == tp->rcv_wup) 5256 tcp_store_ts_recent(tp); 5257 5258 tcp_rcv_rtt_measure_ts(sk, skb); 5259 5260 __skb_pull(skb, tcp_header_len); 5261 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; 5262 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER); 5263 } 5264 if (copied_early) 5265 tcp_cleanup_rbuf(sk, skb->len); 5266 } 5267 if (!eaten) { 5268 if (tcp_checksum_complete_user(sk, skb)) 5269 goto csum_error; 5270 5271 /* Predicted packet is in window by definition. 5272 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 5273 * Hence, check seq<=rcv_wup reduces to: 5274 */ 5275 if (tcp_header_len == 5276 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && 5277 tp->rcv_nxt == tp->rcv_wup) 5278 tcp_store_ts_recent(tp); 5279 5280 tcp_rcv_rtt_measure_ts(sk, skb); 5281 5282 if ((int)skb->truesize > sk->sk_forward_alloc) 5283 goto step5; 5284 5285 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS); 5286 5287 /* Bulk data transfer: receiver */ 5288 __skb_pull(skb, tcp_header_len); 5289 __skb_queue_tail(&sk->sk_receive_queue, skb); 5290 skb_set_owner_r(skb, sk); 5291 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; 5292 } 5293 5294 tcp_event_data_recv(sk, skb); 5295 5296 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) { 5297 /* Well, only one small jumplet in fast path... */ 5298 tcp_ack(sk, skb, FLAG_DATA); 5299 tcp_data_snd_check(sk); 5300 if (!inet_csk_ack_scheduled(sk)) 5301 goto no_ack; 5302 } 5303 5304 if (!copied_early || tp->rcv_nxt != tp->rcv_wup) 5305 __tcp_ack_snd_check(sk, 0); 5306no_ack: 5307#ifdef CONFIG_NET_DMA 5308 if (copied_early) 5309 __skb_queue_tail(&sk->sk_async_wait_queue, skb); 5310 else 5311#endif 5312 if (eaten) 5313 __kfree_skb(skb); 5314 else 5315 sk->sk_data_ready(sk, 0); 5316 return 0; 5317 } 5318 } 5319 5320slow_path: 5321 if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb)) 5322 goto csum_error; 5323 5324 /* 5325 * Standard slow path. 5326 */ 5327 5328 res = tcp_validate_incoming(sk, skb, th, 1); 5329 if (res <= 0) 5330 return -res; 5331 5332step5: 5333 if (th->ack) 5334 tcp_ack(sk, skb, FLAG_SLOWPATH); 5335 5336 tcp_rcv_rtt_measure_ts(sk, skb); 5337 5338 /* Process urgent data. */ 5339 tcp_urg(sk, skb, th); 5340 5341 /* step 7: process the segment text */ 5342 tcp_data_queue(sk, skb); 5343 5344 tcp_data_snd_check(sk); 5345 tcp_ack_snd_check(sk); 5346 return 0; 5347 5348csum_error: 5349 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS); 5350 5351discard: 5352 __kfree_skb(skb); 5353 return 0; 5354} 5355 5356static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb, 5357 struct tcphdr *th, unsigned len) 5358{ 5359 struct tcp_sock *tp = tcp_sk(sk); 5360 struct inet_connection_sock *icsk = inet_csk(sk); 5361 int saved_clamp = tp->rx_opt.mss_clamp; 5362 5363 tcp_parse_options(skb, &tp->rx_opt, 0); 5364 5365 if (th->ack) { 5366 /* rfc793: 5367 * "If the state is SYN-SENT then 5368 * first check the ACK bit 5369 * If the ACK bit is set 5370 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send 5371 * a reset (unless the RST bit is set, if so drop 5372 * the segment and return)" 5373 * 5374 * We do not send data with SYN, so that RFC-correct 5375 * test reduces to: 5376 */ 5377 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt) 5378 goto reset_and_undo; 5379 5380 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 5381 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp, 5382 tcp_time_stamp)) { 5383 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED); 5384 goto reset_and_undo; 5385 } 5386 5387 /* Now ACK is acceptable. 5388 * 5389 * "If the RST bit is set 5390 * If the ACK was acceptable then signal the user "error: 5391 * connection reset", drop the segment, enter CLOSED state, 5392 * delete TCB, and return." 5393 */ 5394 5395 if (th->rst) { 5396 tcp_reset(sk); 5397 goto discard; 5398 } 5399 5400 /* rfc793: 5401 * "fifth, if neither of the SYN or RST bits is set then 5402 * drop the segment and return." 5403 * 5404 * See note below! 5405 * --ANK(990513) 5406 */ 5407 if (!th->syn) 5408 goto discard_and_undo; 5409 5410 /* rfc793: 5411 * "If the SYN bit is on ... 5412 * are acceptable then ... 5413 * (our SYN has been ACKed), change the connection 5414 * state to ESTABLISHED..." 5415 */ 5416 5417 TCP_ECN_rcv_synack(tp, th); 5418 5419 tp->snd_wl1 = TCP_SKB_CB(skb)->seq; 5420 tcp_ack(sk, skb, FLAG_SLOWPATH); 5421 5422 /* Ok.. it's good. Set up sequence numbers and 5423 * move to established. 5424 */ 5425 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; 5426 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; 5427 5428 /* RFC1323: The window in SYN & SYN/ACK segments is 5429 * never scaled. 5430 */ 5431 tp->snd_wnd = ntohs(th->window); 5432 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(skb)->seq); 5433 5434 if (!tp->rx_opt.wscale_ok) { 5435 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0; 5436 tp->window_clamp = min(tp->window_clamp, 65535U); 5437 } 5438 5439 if (tp->rx_opt.saw_tstamp) { 5440 tp->rx_opt.tstamp_ok = 1; 5441 tp->tcp_header_len = 5442 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 5443 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; 5444 tcp_store_ts_recent(tp); 5445 } else { 5446 tp->tcp_header_len = sizeof(struct tcphdr); 5447 } 5448 5449 if (tcp_is_sack(tp) && sysctl_tcp_fack) 5450 tcp_enable_fack(tp); 5451 5452 tcp_mtup_init(sk); 5453 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 5454 tcp_initialize_rcv_mss(sk); 5455 5456 /* Remember, tcp_poll() does not lock socket! 5457 * Change state from SYN-SENT only after copied_seq 5458 * is initialized. */ 5459 tp->copied_seq = tp->rcv_nxt; 5460 smp_mb(); 5461 tcp_set_state(sk, TCP_ESTABLISHED); 5462 5463 security_inet_conn_established(sk, skb); 5464 5465 /* Make sure socket is routed, for correct metrics. */ 5466 icsk->icsk_af_ops->rebuild_header(sk); 5467 5468 tcp_init_metrics(sk); 5469 5470 tcp_init_congestion_control(sk); 5471 5472 /* Prevent spurious tcp_cwnd_restart() on first data 5473 * packet. 5474 */ 5475 tp->lsndtime = tcp_time_stamp; 5476 5477 tcp_init_buffer_space(sk); 5478 5479 if (sock_flag(sk, SOCK_KEEPOPEN)) 5480 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp)); 5481 5482 if (!tp->rx_opt.snd_wscale) 5483 __tcp_fast_path_on(tp, tp->snd_wnd); 5484 else 5485 tp->pred_flags = 0; 5486 5487 if (!sock_flag(sk, SOCK_DEAD)) { 5488 sk->sk_state_change(sk); 5489 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT); 5490 } 5491 5492 if (sk->sk_write_pending || 5493 icsk->icsk_accept_queue.rskq_defer_accept || 5494 icsk->icsk_ack.pingpong) { 5495 /* Save one ACK. Data will be ready after 5496 * several ticks, if write_pending is set. 5497 * 5498 * It may be deleted, but with this feature tcpdumps 5499 * look so _wonderfully_ clever, that I was not able 5500 * to stand against the temptation 8) --ANK 5501 */ 5502 inet_csk_schedule_ack(sk); 5503 icsk->icsk_ack.lrcvtime = tcp_time_stamp; 5504 icsk->icsk_ack.ato = TCP_ATO_MIN; 5505 tcp_incr_quickack(sk); 5506 tcp_enter_quickack_mode(sk); 5507 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 5508 TCP_DELACK_MAX, TCP_RTO_MAX); 5509 5510discard: 5511 __kfree_skb(skb); 5512 return 0; 5513 } else { 5514 tcp_send_ack(sk); 5515 } 5516 return -1; 5517 } 5518 5519 /* No ACK in the segment */ 5520 5521 if (th->rst) { 5522 /* rfc793: 5523 * "If the RST bit is set 5524 * 5525 * Otherwise (no ACK) drop the segment and return." 5526 */ 5527 5528 goto discard_and_undo; 5529 } 5530 5531 /* PAWS check. */ 5532 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && 5533 tcp_paws_check(&tp->rx_opt, 0)) 5534 goto discard_and_undo; 5535 5536 if (th->syn) { 5537 /* We see SYN without ACK. It is attempt of 5538 * simultaneous connect with crossed SYNs. 5539 * Particularly, it can be connect to self. 5540 */ 5541 tcp_set_state(sk, TCP_SYN_RECV); 5542 5543 if (tp->rx_opt.saw_tstamp) { 5544 tp->rx_opt.tstamp_ok = 1; 5545 tcp_store_ts_recent(tp); 5546 tp->tcp_header_len = 5547 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 5548 } else { 5549 tp->tcp_header_len = sizeof(struct tcphdr); 5550 } 5551 5552 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; 5553 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; 5554 5555 /* RFC1323: The window in SYN & SYN/ACK segments is 5556 * never scaled. 5557 */ 5558 tp->snd_wnd = ntohs(th->window); 5559 tp->snd_wl1 = TCP_SKB_CB(skb)->seq; 5560 tp->max_window = tp->snd_wnd; 5561 5562 TCP_ECN_rcv_syn(tp, th); 5563 5564 tcp_mtup_init(sk); 5565 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 5566 tcp_initialize_rcv_mss(sk); 5567 5568 tcp_send_synack(sk); 5569#if 0 5570 /* Note, we could accept data and URG from this segment. 5571 * There are no obstacles to make this. 5572 * 5573 * However, if we ignore data in ACKless segments sometimes, 5574 * we have no reasons to accept it sometimes. 5575 * Also, seems the code doing it in step6 of tcp_rcv_state_process 5576 * is not flawless. So, discard packet for sanity. 5577 * Uncomment this return to process the data. 5578 */ 5579 return -1; 5580#else 5581 goto discard; 5582#endif 5583 } 5584 /* "fifth, if neither of the SYN or RST bits is set then 5585 * drop the segment and return." 5586 */ 5587 5588discard_and_undo: 5589 tcp_clear_options(&tp->rx_opt); 5590 tp->rx_opt.mss_clamp = saved_clamp; 5591 goto discard; 5592 5593reset_and_undo: 5594 tcp_clear_options(&tp->rx_opt); 5595 tp->rx_opt.mss_clamp = saved_clamp; 5596 return 1; 5597} 5598 5599/* 5600 * This function implements the receiving procedure of RFC 793 for 5601 * all states except ESTABLISHED and TIME_WAIT. 5602 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be 5603 * address independent. 5604 */ 5605 5606int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb, 5607 struct tcphdr *th, unsigned len) 5608{ 5609 struct tcp_sock *tp = tcp_sk(sk); 5610 struct inet_connection_sock *icsk = inet_csk(sk); 5611 int queued = 0; 5612 int res; 5613 5614 tp->rx_opt.saw_tstamp = 0; 5615 5616 switch (sk->sk_state) { 5617 case TCP_CLOSE: 5618 goto discard; 5619 5620 case TCP_LISTEN: 5621 if (th->ack) 5622 return 1; 5623 5624 if (th->rst) 5625 goto discard; 5626 5627 if (th->syn) { 5628 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0) 5629 return 1; 5630 5631 /* Now we have several options: In theory there is 5632 * nothing else in the frame. KA9Q has an option to 5633 * send data with the syn, BSD accepts data with the 5634 * syn up to the [to be] advertised window and 5635 * Solaris 2.1 gives you a protocol error. For now 5636 * we just ignore it, that fits the spec precisely 5637 * and avoids incompatibilities. It would be nice in 5638 * future to drop through and process the data. 5639 * 5640 * Now that TTCP is starting to be used we ought to 5641 * queue this data. 5642 * But, this leaves one open to an easy denial of 5643 * service attack, and SYN cookies can't defend 5644 * against this problem. So, we drop the data 5645 * in the interest of security over speed unless 5646 * it's still in use. 5647 */ 5648 kfree_skb(skb); 5649 return 0; 5650 } 5651 goto discard; 5652 5653 case TCP_SYN_SENT: 5654 queued = tcp_rcv_synsent_state_process(sk, skb, th, len); 5655 if (queued >= 0) 5656 return queued; 5657 5658 /* Do step6 onward by hand. */ 5659 tcp_urg(sk, skb, th); 5660 __kfree_skb(skb); 5661 tcp_data_snd_check(sk); 5662 return 0; 5663 } 5664 5665 res = tcp_validate_incoming(sk, skb, th, 0); 5666 if (res <= 0) 5667 return -res; 5668 5669 /* step 5: check the ACK field */ 5670 if (th->ack) { 5671 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH); 5672 5673 switch (sk->sk_state) { 5674 case TCP_SYN_RECV: 5675 if (acceptable) { 5676 tp->copied_seq = tp->rcv_nxt; 5677 smp_mb(); 5678 tcp_set_state(sk, TCP_ESTABLISHED); 5679 sk->sk_state_change(sk); 5680 5681 /* Note, that this wakeup is only for marginal 5682 * crossed SYN case. Passively open sockets 5683 * are not waked up, because sk->sk_sleep == 5684 * NULL and sk->sk_socket == NULL. 5685 */ 5686 if (sk->sk_socket) 5687 sk_wake_async(sk, 5688 SOCK_WAKE_IO, POLL_OUT); 5689 5690 tp->snd_una = TCP_SKB_CB(skb)->ack_seq; 5691 tp->snd_wnd = ntohs(th->window) << 5692 tp->rx_opt.snd_wscale; 5693 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, 5694 TCP_SKB_CB(skb)->seq); 5695 5696 /* tcp_ack considers this ACK as duplicate 5697 * and does not calculate rtt. 5698 * Fix it at least with timestamps. 5699 */ 5700 if (tp->rx_opt.saw_tstamp && 5701 tp->rx_opt.rcv_tsecr && !tp->srtt) 5702 tcp_ack_saw_tstamp(sk, 0); 5703 5704 if (tp->rx_opt.tstamp_ok) 5705 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; 5706 5707 /* Make sure socket is routed, for 5708 * correct metrics. 5709 */ 5710 icsk->icsk_af_ops->rebuild_header(sk); 5711 5712 tcp_init_metrics(sk); 5713 5714 tcp_init_congestion_control(sk); 5715 5716 /* Prevent spurious tcp_cwnd_restart() on 5717 * first data packet. 5718 */ 5719 tp->lsndtime = tcp_time_stamp; 5720 5721 tcp_mtup_init(sk); 5722 tcp_initialize_rcv_mss(sk); 5723 tcp_init_buffer_space(sk); 5724 tcp_fast_path_on(tp); 5725 } else { 5726 return 1; 5727 } 5728 break; 5729 5730 case TCP_FIN_WAIT1: 5731 if (tp->snd_una == tp->write_seq) { 5732 tcp_set_state(sk, TCP_FIN_WAIT2); 5733 sk->sk_shutdown |= SEND_SHUTDOWN; 5734 dst_confirm(sk->sk_dst_cache); 5735 5736 if (!sock_flag(sk, SOCK_DEAD)) 5737 /* Wake up lingering close() */ 5738 sk->sk_state_change(sk); 5739 else { 5740 int tmo; 5741 5742 if (tp->linger2 < 0 || 5743 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 5744 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) { 5745 tcp_done(sk); 5746 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA); 5747 return 1; 5748 } 5749 5750 tmo = tcp_fin_time(sk); 5751 if (tmo > TCP_TIMEWAIT_LEN) { 5752 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN); 5753 } else if (th->fin || sock_owned_by_user(sk)) { 5754 /* Bad case. We could lose such FIN otherwise. 5755 * It is not a big problem, but it looks confusing 5756 * and not so rare event. We still can lose it now, 5757 * if it spins in bh_lock_sock(), but it is really 5758 * marginal case. 5759 */ 5760 inet_csk_reset_keepalive_timer(sk, tmo); 5761 } else { 5762 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo); 5763 goto discard; 5764 } 5765 } 5766 } 5767 break; 5768 5769 case TCP_CLOSING: 5770 if (tp->snd_una == tp->write_seq) { 5771 tcp_time_wait(sk, TCP_TIME_WAIT, 0); 5772 goto discard; 5773 } 5774 break; 5775 5776 case TCP_LAST_ACK: 5777 if (tp->snd_una == tp->write_seq) { 5778 tcp_update_metrics(sk); 5779 tcp_done(sk); 5780 goto discard; 5781 } 5782 break; 5783 } 5784 } else 5785 goto discard; 5786 5787 /* step 6: check the URG bit */ 5788 tcp_urg(sk, skb, th); 5789 5790 /* step 7: process the segment text */ 5791 switch (sk->sk_state) { 5792 case TCP_CLOSE_WAIT: 5793 case TCP_CLOSING: 5794 case TCP_LAST_ACK: 5795 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) 5796 break; 5797 case TCP_FIN_WAIT1: 5798 case TCP_FIN_WAIT2: 5799 /* RFC 793 says to queue data in these states, 5800 * RFC 1122 says we MUST send a reset. 5801 * BSD 4.4 also does reset. 5802 */ 5803 if (sk->sk_shutdown & RCV_SHUTDOWN) { 5804 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 5805 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) { 5806 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA); 5807 tcp_reset(sk); 5808 return 1; 5809 } 5810 } 5811 /* Fall through */ 5812 case TCP_ESTABLISHED: 5813 tcp_data_queue(sk, skb); 5814 queued = 1; 5815 break; 5816 } 5817 5818 /* tcp_data could move socket to TIME-WAIT */ 5819 if (sk->sk_state != TCP_CLOSE) { 5820 tcp_data_snd_check(sk); 5821 tcp_ack_snd_check(sk); 5822 } 5823 5824 if (!queued) { 5825discard: 5826 __kfree_skb(skb); 5827 } 5828 return 0; 5829} 5830 5831EXPORT_SYMBOL(sysctl_tcp_ecn); 5832EXPORT_SYMBOL(sysctl_tcp_reordering); 5833EXPORT_SYMBOL(sysctl_tcp_adv_win_scale); 5834EXPORT_SYMBOL(tcp_parse_options); 5835#ifdef CONFIG_TCP_MD5SIG 5836EXPORT_SYMBOL(tcp_parse_md5sig_option); 5837#endif 5838EXPORT_SYMBOL(tcp_rcv_established); 5839EXPORT_SYMBOL(tcp_rcv_state_process); 5840EXPORT_SYMBOL(tcp_initialize_rcv_mss); 5841