1 /* 2 * INET An implementation of the TCP/IP protocol suite for the LINUX 3 * operating system. INET is implemented using the BSD Socket 4 * interface as the means of communication with the user level. 5 * 6 * Implementation of the Transmission Control Protocol(TCP). 7 * 8 * Version: $Id: tcp_input.c,v 1.243 2002/02/01 22:01:04 davem Exp $ 9 * 10 * Authors: Ross Biro 11 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> 12 * Mark Evans, <evansmp@uhura.aston.ac.uk> 13 * Corey Minyard <wf-rch!minyard@relay.EU.net> 14 * Florian La Roche, <flla@stud.uni-sb.de> 15 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu> 16 * Linus Torvalds, <torvalds@cs.helsinki.fi> 17 * Alan Cox, <gw4pts@gw4pts.ampr.org> 18 * Matthew Dillon, <dillon@apollo.west.oic.com> 19 * Arnt Gulbrandsen, <agulbra@nvg.unit.no> 20 * Jorge Cwik, <jorge@laser.satlink.net> 21 */ 22 23 /* 24 * Changes: 25 * Pedro Roque : Fast Retransmit/Recovery. 26 * Two receive queues. 27 * Retransmit queue handled by TCP. 28 * Better retransmit timer handling. 29 * New congestion avoidance. 30 * Header prediction. 31 * Variable renaming. 32 * 33 * Eric : Fast Retransmit. 34 * Randy Scott : MSS option defines. 35 * Eric Schenk : Fixes to slow start algorithm. 36 * Eric Schenk : Yet another double ACK bug. 37 * Eric Schenk : Delayed ACK bug fixes. 38 * Eric Schenk : Floyd style fast retrans war avoidance. 39 * David S. Miller : Don't allow zero congestion window. 40 * Eric Schenk : Fix retransmitter so that it sends 41 * next packet on ack of previous packet. 42 * Andi Kleen : Moved open_request checking here 43 * and process RSTs for open_requests. 44 * Andi Kleen : Better prune_queue, and other fixes. 45 * Andrey Savochkin: Fix RTT measurements in the presence of 46 * timestamps. 47 * Andrey Savochkin: Check sequence numbers correctly when 48 * removing SACKs due to in sequence incoming 49 * data segments. 50 * Andi Kleen: Make sure we never ack data there is not 51 * enough room for. Also make this condition 52 * a fatal error if it might still happen. 53 * Andi Kleen: Add tcp_measure_rcv_mss to make 54 * connections with MSS<min(MTU,ann. MSS) 55 * work without delayed acks. 56 * Andi Kleen: Process packets with PSH set in the 57 * fast path. 58 * J Hadi Salim: ECN support 59 * Andrei Gurtov, 60 * Pasi Sarolahti, 61 * Panu Kuhlberg: Experimental audit of TCP (re)transmission 62 * engine. Lots of bugs are found. 63 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs 64 */ 65 66 #include <linux/mm.h> 67 #include <linux/module.h> 68 #include <linux/sysctl.h> 69 #include <net/tcp.h> 70 #include <net/inet_common.h> 71 #include <linux/ipsec.h> 72 #include <asm/unaligned.h> 73 #include <net/netdma.h> 74 75 int sysctl_tcp_timestamps __read_mostly = 1; 76 int sysctl_tcp_window_scaling __read_mostly = 1; 77 int sysctl_tcp_sack __read_mostly = 1; 78 int sysctl_tcp_fack __read_mostly = 1; 79 int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH; 80 int sysctl_tcp_ecn __read_mostly; 81 int sysctl_tcp_dsack __read_mostly = 1; 82 int sysctl_tcp_app_win __read_mostly = 31; 83 int sysctl_tcp_adv_win_scale __read_mostly = 2; 84 85 int sysctl_tcp_stdurg __read_mostly; 86 int sysctl_tcp_rfc1337 __read_mostly; 87 int sysctl_tcp_max_orphans __read_mostly = NR_FILE; 88 int sysctl_tcp_frto __read_mostly = 2; 89 int sysctl_tcp_frto_response __read_mostly; 90 int sysctl_tcp_nometrics_save __read_mostly; 91 92 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1; 93 int sysctl_tcp_abc __read_mostly; 94 95 #define FLAG_DATA 0x01 /* Incoming frame contained data. */ 96 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */ 97 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */ 98 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */ 99 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */ 100 #define FLAG_DATA_SACKED 0x20 /* New SACK. */ 101 #define FLAG_ECE 0x40 /* ECE in this ACK */ 102 #define FLAG_DATA_LOST 0x80 /* SACK detected data lossage. */ 103 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/ 104 #define FLAG_ONLY_ORIG_SACKED 0x200 /* SACKs only non-rexmit sent before RTO */ 105 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */ 106 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */ 107 #define FLAG_NONHEAD_RETRANS_ACKED 0x1000 /* Non-head rexmitted data was ACKed */ 108 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */ 109 110 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED) 111 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED) 112 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE) 113 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED) 114 #define FLAG_ANY_PROGRESS (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED) 115 116 #define IsSackFrto() (sysctl_tcp_frto == 0x2) 117 118 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH) 119 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH)) 120 121 /* Adapt the MSS value used to make delayed ack decision to the 122 * real world. 123 */ 124 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb) 125 { 126 struct inet_connection_sock *icsk = inet_csk(sk); 127 const unsigned int lss = icsk->icsk_ack.last_seg_size; 128 unsigned int len; 129 130 icsk->icsk_ack.last_seg_size = 0; 131 132 /* skb->len may jitter because of SACKs, even if peer 133 * sends good full-sized frames. 134 */ 135 len = skb_shinfo(skb)->gso_size ? : skb->len; 136 if (len >= icsk->icsk_ack.rcv_mss) { 137 icsk->icsk_ack.rcv_mss = len; 138 } else { 139 /* Otherwise, we make more careful check taking into account, 140 * that SACKs block is variable. 141 * 142 * "len" is invariant segment length, including TCP header. 143 */ 144 len += skb->data - skb_transport_header(skb); 145 if (len >= TCP_MIN_RCVMSS + sizeof(struct tcphdr) || 146 /* If PSH is not set, packet should be 147 * full sized, provided peer TCP is not badly broken. 148 * This observation (if it is correct 8)) allows 149 * to handle super-low mtu links fairly. 150 */ 151 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) && 152 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) { 153 /* Subtract also invariant (if peer is RFC compliant), 154 * tcp header plus fixed timestamp option length. 155 * Resulting "len" is MSS free of SACK jitter. 156 */ 157 len -= tcp_sk(sk)->tcp_header_len; 158 icsk->icsk_ack.last_seg_size = len; 159 if (len == lss) { 160 icsk->icsk_ack.rcv_mss = len; 161 return; 162 } 163 } 164 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED) 165 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2; 166 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED; 167 } 168 } 169 170 static void tcp_incr_quickack(struct sock *sk) 171 { 172 struct inet_connection_sock *icsk = inet_csk(sk); 173 unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss); 174 175 if (quickacks == 0) 176 quickacks = 2; 177 if (quickacks > icsk->icsk_ack.quick) 178 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS); 179 } 180 181 void tcp_enter_quickack_mode(struct sock *sk) 182 { 183 struct inet_connection_sock *icsk = inet_csk(sk); 184 tcp_incr_quickack(sk); 185 icsk->icsk_ack.pingpong = 0; 186 icsk->icsk_ack.ato = TCP_ATO_MIN; 187 } 188 189 /* Send ACKs quickly, if "quick" count is not exhausted 190 * and the session is not interactive. 191 */ 192 193 static inline int tcp_in_quickack_mode(const struct sock *sk) 194 { 195 const struct inet_connection_sock *icsk = inet_csk(sk); 196 return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong; 197 } 198 199 static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp) 200 { 201 if (tp->ecn_flags & TCP_ECN_OK) 202 tp->ecn_flags |= TCP_ECN_QUEUE_CWR; 203 } 204 205 static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, struct sk_buff *skb) 206 { 207 if (tcp_hdr(skb)->cwr) 208 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR; 209 } 210 211 static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp) 212 { 213 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR; 214 } 215 216 static inline void TCP_ECN_check_ce(struct tcp_sock *tp, struct sk_buff *skb) 217 { 218 if (tp->ecn_flags & TCP_ECN_OK) { 219 if (INET_ECN_is_ce(TCP_SKB_CB(skb)->flags)) 220 tp->ecn_flags |= TCP_ECN_DEMAND_CWR; 221 /* Funny extension: if ECT is not set on a segment, 222 * it is surely retransmit. It is not in ECN RFC, 223 * but Linux follows this rule. */ 224 else if (INET_ECN_is_not_ect((TCP_SKB_CB(skb)->flags))) 225 tcp_enter_quickack_mode((struct sock *)tp); 226 } 227 } 228 229 static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, struct tcphdr *th) 230 { 231 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr)) 232 tp->ecn_flags &= ~TCP_ECN_OK; 233 } 234 235 static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, struct tcphdr *th) 236 { 237 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr)) 238 tp->ecn_flags &= ~TCP_ECN_OK; 239 } 240 241 static inline int TCP_ECN_rcv_ecn_echo(struct tcp_sock *tp, struct tcphdr *th) 242 { 243 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK)) 244 return 1; 245 return 0; 246 } 247 248 /* Buffer size and advertised window tuning. 249 * 250 * 1. Tuning sk->sk_sndbuf, when connection enters established state. 251 */ 252 253 static void tcp_fixup_sndbuf(struct sock *sk) 254 { 255 int sndmem = tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER + 16 + 256 sizeof(struct sk_buff); 257 258 if (sk->sk_sndbuf < 3 * sndmem) 259 sk->sk_sndbuf = min(3 * sndmem, sysctl_tcp_wmem[2]); 260 } 261 262 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh) 263 * 264 * All tcp_full_space() is split to two parts: "network" buffer, allocated 265 * forward and advertised in receiver window (tp->rcv_wnd) and 266 * "application buffer", required to isolate scheduling/application 267 * latencies from network. 268 * window_clamp is maximal advertised window. It can be less than 269 * tcp_full_space(), in this case tcp_full_space() - window_clamp 270 * is reserved for "application" buffer. The less window_clamp is 271 * the smoother our behaviour from viewpoint of network, but the lower 272 * throughput and the higher sensitivity of the connection to losses. 8) 273 * 274 * rcv_ssthresh is more strict window_clamp used at "slow start" 275 * phase to predict further behaviour of this connection. 276 * It is used for two goals: 277 * - to enforce header prediction at sender, even when application 278 * requires some significant "application buffer". It is check #1. 279 * - to prevent pruning of receive queue because of misprediction 280 * of receiver window. Check #2. 281 * 282 * The scheme does not work when sender sends good segments opening 283 * window and then starts to feed us spaghetti. But it should work 284 * in common situations. Otherwise, we have to rely on queue collapsing. 285 */ 286 287 /* Slow part of check#2. */ 288 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb) 289 { 290 struct tcp_sock *tp = tcp_sk(sk); 291 /* Optimize this! */ 292 int truesize = tcp_win_from_space(skb->truesize) >> 1; 293 int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1; 294 295 while (tp->rcv_ssthresh <= window) { 296 if (truesize <= skb->len) 297 return 2 * inet_csk(sk)->icsk_ack.rcv_mss; 298 299 truesize >>= 1; 300 window >>= 1; 301 } 302 return 0; 303 } 304 305 static void tcp_grow_window(struct sock *sk, struct sk_buff *skb) 306 { 307 struct tcp_sock *tp = tcp_sk(sk); 308 309 /* Check #1 */ 310 if (tp->rcv_ssthresh < tp->window_clamp && 311 (int)tp->rcv_ssthresh < tcp_space(sk) && 312 !tcp_memory_pressure) { 313 int incr; 314 315 /* Check #2. Increase window, if skb with such overhead 316 * will fit to rcvbuf in future. 317 */ 318 if (tcp_win_from_space(skb->truesize) <= skb->len) 319 incr = 2 * tp->advmss; 320 else 321 incr = __tcp_grow_window(sk, skb); 322 323 if (incr) { 324 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr, 325 tp->window_clamp); 326 inet_csk(sk)->icsk_ack.quick |= 1; 327 } 328 } 329 } 330 331 /* 3. Tuning rcvbuf, when connection enters established state. */ 332 333 static void tcp_fixup_rcvbuf(struct sock *sk) 334 { 335 struct tcp_sock *tp = tcp_sk(sk); 336 int rcvmem = tp->advmss + MAX_TCP_HEADER + 16 + sizeof(struct sk_buff); 337 338 /* Try to select rcvbuf so that 4 mss-sized segments 339 * will fit to window and corresponding skbs will fit to our rcvbuf. 340 * (was 3; 4 is minimum to allow fast retransmit to work.) 341 */ 342 while (tcp_win_from_space(rcvmem) < tp->advmss) 343 rcvmem += 128; 344 if (sk->sk_rcvbuf < 4 * rcvmem) 345 sk->sk_rcvbuf = min(4 * rcvmem, sysctl_tcp_rmem[2]); 346 } 347 348 /* 4. Try to fixup all. It is made immediately after connection enters 349 * established state. 350 */ 351 static void tcp_init_buffer_space(struct sock *sk) 352 { 353 struct tcp_sock *tp = tcp_sk(sk); 354 int maxwin; 355 356 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) 357 tcp_fixup_rcvbuf(sk); 358 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK)) 359 tcp_fixup_sndbuf(sk); 360 361 tp->rcvq_space.space = tp->rcv_wnd; 362 363 maxwin = tcp_full_space(sk); 364 365 if (tp->window_clamp >= maxwin) { 366 tp->window_clamp = maxwin; 367 368 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss) 369 tp->window_clamp = max(maxwin - 370 (maxwin >> sysctl_tcp_app_win), 371 4 * tp->advmss); 372 } 373 374 /* Force reservation of one segment. */ 375 if (sysctl_tcp_app_win && 376 tp->window_clamp > 2 * tp->advmss && 377 tp->window_clamp + tp->advmss > maxwin) 378 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss); 379 380 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp); 381 tp->snd_cwnd_stamp = tcp_time_stamp; 382 } 383 384 /* 5. Recalculate window clamp after socket hit its memory bounds. */ 385 static void tcp_clamp_window(struct sock *sk) 386 { 387 struct tcp_sock *tp = tcp_sk(sk); 388 struct inet_connection_sock *icsk = inet_csk(sk); 389 390 icsk->icsk_ack.quick = 0; 391 392 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] && 393 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) && 394 !tcp_memory_pressure && 395 atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) { 396 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc), 397 sysctl_tcp_rmem[2]); 398 } 399 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf) 400 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss); 401 } 402 403 /* Initialize RCV_MSS value. 404 * RCV_MSS is an our guess about MSS used by the peer. 405 * We haven't any direct information about the MSS. 406 * It's better to underestimate the RCV_MSS rather than overestimate. 407 * Overestimations make us ACKing less frequently than needed. 408 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss(). 409 */ 410 void tcp_initialize_rcv_mss(struct sock *sk) 411 { 412 struct tcp_sock *tp = tcp_sk(sk); 413 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache); 414 415 hint = min(hint, tp->rcv_wnd / 2); 416 hint = min(hint, TCP_MIN_RCVMSS); 417 hint = max(hint, TCP_MIN_MSS); 418 419 inet_csk(sk)->icsk_ack.rcv_mss = hint; 420 } 421 422 /* Receiver "autotuning" code. 423 * 424 * The algorithm for RTT estimation w/o timestamps is based on 425 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL. 426 * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps> 427 * 428 * More detail on this code can be found at 429 * <http://www.psc.edu/~jheffner/senior_thesis.ps>, 430 * though this reference is out of date. A new paper 431 * is pending. 432 */ 433 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep) 434 { 435 u32 new_sample = tp->rcv_rtt_est.rtt; 436 long m = sample; 437 438 if (m == 0) 439 m = 1; 440 441 if (new_sample != 0) { 442 /* If we sample in larger samples in the non-timestamp 443 * case, we could grossly overestimate the RTT especially 444 * with chatty applications or bulk transfer apps which 445 * are stalled on filesystem I/O. 446 * 447 * Also, since we are only going for a minimum in the 448 * non-timestamp case, we do not smooth things out 449 * else with timestamps disabled convergence takes too 450 * long. 451 */ 452 if (!win_dep) { 453 m -= (new_sample >> 3); 454 new_sample += m; 455 } else if (m < new_sample) 456 new_sample = m << 3; 457 } else { 458 /* No previous measure. */ 459 new_sample = m << 3; 460 } 461 462 if (tp->rcv_rtt_est.rtt != new_sample) 463 tp->rcv_rtt_est.rtt = new_sample; 464 } 465 466 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp) 467 { 468 if (tp->rcv_rtt_est.time == 0) 469 goto new_measure; 470 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq)) 471 return; 472 tcp_rcv_rtt_update(tp, jiffies - tp->rcv_rtt_est.time, 1); 473 474 new_measure: 475 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd; 476 tp->rcv_rtt_est.time = tcp_time_stamp; 477 } 478 479 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, 480 const struct sk_buff *skb) 481 { 482 struct tcp_sock *tp = tcp_sk(sk); 483 if (tp->rx_opt.rcv_tsecr && 484 (TCP_SKB_CB(skb)->end_seq - 485 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss)) 486 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0); 487 } 488 489 /* 490 * This function should be called every time data is copied to user space. 491 * It calculates the appropriate TCP receive buffer space. 492 */ 493 void tcp_rcv_space_adjust(struct sock *sk) 494 { 495 struct tcp_sock *tp = tcp_sk(sk); 496 int time; 497 int space; 498 499 if (tp->rcvq_space.time == 0) 500 goto new_measure; 501 502 time = tcp_time_stamp - tp->rcvq_space.time; 503 if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0) 504 return; 505 506 space = 2 * (tp->copied_seq - tp->rcvq_space.seq); 507 508 space = max(tp->rcvq_space.space, space); 509 510 if (tp->rcvq_space.space != space) { 511 int rcvmem; 512 513 tp->rcvq_space.space = space; 514 515 if (sysctl_tcp_moderate_rcvbuf && 516 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) { 517 int new_clamp = space; 518 519 /* Receive space grows, normalize in order to 520 * take into account packet headers and sk_buff 521 * structure overhead. 522 */ 523 space /= tp->advmss; 524 if (!space) 525 space = 1; 526 rcvmem = (tp->advmss + MAX_TCP_HEADER + 527 16 + sizeof(struct sk_buff)); 528 while (tcp_win_from_space(rcvmem) < tp->advmss) 529 rcvmem += 128; 530 space *= rcvmem; 531 space = min(space, sysctl_tcp_rmem[2]); 532 if (space > sk->sk_rcvbuf) { 533 sk->sk_rcvbuf = space; 534 535 /* Make the window clamp follow along. */ 536 tp->window_clamp = new_clamp; 537 } 538 } 539 } 540 541 new_measure: 542 tp->rcvq_space.seq = tp->copied_seq; 543 tp->rcvq_space.time = tcp_time_stamp; 544 } 545 546 /* There is something which you must keep in mind when you analyze the 547 * behavior of the tp->ato delayed ack timeout interval. When a 548 * connection starts up, we want to ack as quickly as possible. The 549 * problem is that "good" TCP's do slow start at the beginning of data 550 * transmission. The means that until we send the first few ACK's the 551 * sender will sit on his end and only queue most of his data, because 552 * he can only send snd_cwnd unacked packets at any given time. For 553 * each ACK we send, he increments snd_cwnd and transmits more of his 554 * queue. -DaveM 555 */ 556 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb) 557 { 558 struct tcp_sock *tp = tcp_sk(sk); 559 struct inet_connection_sock *icsk = inet_csk(sk); 560 u32 now; 561 562 inet_csk_schedule_ack(sk); 563 564 tcp_measure_rcv_mss(sk, skb); 565 566 tcp_rcv_rtt_measure(tp); 567 568 now = tcp_time_stamp; 569 570 if (!icsk->icsk_ack.ato) { 571 /* The _first_ data packet received, initialize 572 * delayed ACK engine. 573 */ 574 tcp_incr_quickack(sk); 575 icsk->icsk_ack.ato = TCP_ATO_MIN; 576 } else { 577 int m = now - icsk->icsk_ack.lrcvtime; 578 579 if (m <= TCP_ATO_MIN / 2) { 580 /* The fastest case is the first. */ 581 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2; 582 } else if (m < icsk->icsk_ack.ato) { 583 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m; 584 if (icsk->icsk_ack.ato > icsk->icsk_rto) 585 icsk->icsk_ack.ato = icsk->icsk_rto; 586 } else if (m > icsk->icsk_rto) { 587 /* Too long gap. Apparently sender failed to 588 * restart window, so that we send ACKs quickly. 589 */ 590 tcp_incr_quickack(sk); 591 sk_mem_reclaim(sk); 592 } 593 } 594 icsk->icsk_ack.lrcvtime = now; 595 596 TCP_ECN_check_ce(tp, skb); 597 598 if (skb->len >= 128) 599 tcp_grow_window(sk, skb); 600 } 601 602 static u32 tcp_rto_min(struct sock *sk) 603 { 604 struct dst_entry *dst = __sk_dst_get(sk); 605 u32 rto_min = TCP_RTO_MIN; 606 607 if (dst && dst_metric_locked(dst, RTAX_RTO_MIN)) 608 rto_min = dst->metrics[RTAX_RTO_MIN - 1]; 609 return rto_min; 610 } 611 612 /* Called to compute a smoothed rtt estimate. The data fed to this 613 * routine either comes from timestamps, or from segments that were 614 * known _not_ to have been retransmitted [see Karn/Partridge 615 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88 616 * piece by Van Jacobson. 617 * NOTE: the next three routines used to be one big routine. 618 * To save cycles in the RFC 1323 implementation it was better to break 619 * it up into three procedures. -- erics 620 */ 621 static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt) 622 { 623 struct tcp_sock *tp = tcp_sk(sk); 624 long m = mrtt; /* RTT */ 625 626 /* The following amusing code comes from Jacobson's 627 * article in SIGCOMM '88. Note that rtt and mdev 628 * are scaled versions of rtt and mean deviation. 629 * This is designed to be as fast as possible 630 * m stands for "measurement". 631 * 632 * On a 1990 paper the rto value is changed to: 633 * RTO = rtt + 4 * mdev 634 * 635 * Funny. This algorithm seems to be very broken. 636 * These formulae increase RTO, when it should be decreased, increase 637 * too slowly, when it should be increased quickly, decrease too quickly 638 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely 639 * does not matter how to _calculate_ it. Seems, it was trap 640 * that VJ failed to avoid. 8) 641 */ 642 if (m == 0) 643 m = 1; 644 if (tp->srtt != 0) { 645 m -= (tp->srtt >> 3); /* m is now error in rtt est */ 646 tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */ 647 if (m < 0) { 648 m = -m; /* m is now abs(error) */ 649 m -= (tp->mdev >> 2); /* similar update on mdev */ 650 /* This is similar to one of Eifel findings. 651 * Eifel blocks mdev updates when rtt decreases. 652 * This solution is a bit different: we use finer gain 653 * for mdev in this case (alpha*beta). 654 * Like Eifel it also prevents growth of rto, 655 * but also it limits too fast rto decreases, 656 * happening in pure Eifel. 657 */ 658 if (m > 0) 659 m >>= 3; 660 } else { 661 m -= (tp->mdev >> 2); /* similar update on mdev */ 662 } 663 tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */ 664 if (tp->mdev > tp->mdev_max) { 665 tp->mdev_max = tp->mdev; 666 if (tp->mdev_max > tp->rttvar) 667 tp->rttvar = tp->mdev_max; 668 } 669 if (after(tp->snd_una, tp->rtt_seq)) { 670 if (tp->mdev_max < tp->rttvar) 671 tp->rttvar -= (tp->rttvar - tp->mdev_max) >> 2; 672 tp->rtt_seq = tp->snd_nxt; 673 tp->mdev_max = tcp_rto_min(sk); 674 } 675 } else { 676 /* no previous measure. */ 677 tp->srtt = m << 3; /* take the measured time to be rtt */ 678 tp->mdev = m << 1; /* make sure rto = 3*rtt */ 679 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk)); 680 tp->rtt_seq = tp->snd_nxt; 681 } 682 } 683 684 /* Calculate rto without backoff. This is the second half of Van Jacobson's 685 * routine referred to above. 686 */ 687 static inline void tcp_set_rto(struct sock *sk) 688 { 689 const struct tcp_sock *tp = tcp_sk(sk); 690 /* Old crap is replaced with new one. 8) 691 * 692 * More seriously: 693 * 1. If rtt variance happened to be less 50msec, it is hallucination. 694 * It cannot be less due to utterly erratic ACK generation made 695 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_ 696 * to do with delayed acks, because at cwnd>2 true delack timeout 697 * is invisible. Actually, Linux-2.4 also generates erratic 698 * ACKs in some circumstances. 699 */ 700 inet_csk(sk)->icsk_rto = (tp->srtt >> 3) + tp->rttvar; 701 702 /* 2. Fixups made earlier cannot be right. 703 * If we do not estimate RTO correctly without them, 704 * all the algo is pure shit and should be replaced 705 * with correct one. It is exactly, which we pretend to do. 706 */ 707 } 708 709 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo 710 * guarantees that rto is higher. 711 */ 712 static inline void tcp_bound_rto(struct sock *sk) 713 { 714 if (inet_csk(sk)->icsk_rto > TCP_RTO_MAX) 715 inet_csk(sk)->icsk_rto = TCP_RTO_MAX; 716 } 717 718 /* Save metrics learned by this TCP session. 719 This function is called only, when TCP finishes successfully 720 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE. 721 */ 722 void tcp_update_metrics(struct sock *sk) 723 { 724 struct tcp_sock *tp = tcp_sk(sk); 725 struct dst_entry *dst = __sk_dst_get(sk); 726 727 if (sysctl_tcp_nometrics_save) 728 return; 729 730 dst_confirm(dst); 731 732 if (dst && (dst->flags & DST_HOST)) { 733 const struct inet_connection_sock *icsk = inet_csk(sk); 734 int m; 735 736 if (icsk->icsk_backoff || !tp->srtt) { 737 /* This session failed to estimate rtt. Why? 738 * Probably, no packets returned in time. 739 * Reset our results. 740 */ 741 if (!(dst_metric_locked(dst, RTAX_RTT))) 742 dst->metrics[RTAX_RTT - 1] = 0; 743 return; 744 } 745 746 m = dst_metric(dst, RTAX_RTT) - tp->srtt; 747 748 /* If newly calculated rtt larger than stored one, 749 * store new one. Otherwise, use EWMA. Remember, 750 * rtt overestimation is always better than underestimation. 751 */ 752 if (!(dst_metric_locked(dst, RTAX_RTT))) { 753 if (m <= 0) 754 dst->metrics[RTAX_RTT - 1] = tp->srtt; 755 else 756 dst->metrics[RTAX_RTT - 1] -= (m >> 3); 757 } 758 759 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) { 760 if (m < 0) 761 m = -m; 762 763 /* Scale deviation to rttvar fixed point */ 764 m >>= 1; 765 if (m < tp->mdev) 766 m = tp->mdev; 767 768 if (m >= dst_metric(dst, RTAX_RTTVAR)) 769 dst->metrics[RTAX_RTTVAR - 1] = m; 770 else 771 dst->metrics[RTAX_RTTVAR-1] -= 772 (dst->metrics[RTAX_RTTVAR-1] - m)>>2; 773 } 774 775 if (tp->snd_ssthresh >= 0xFFFF) { 776 /* Slow start still did not finish. */ 777 if (dst_metric(dst, RTAX_SSTHRESH) && 778 !dst_metric_locked(dst, RTAX_SSTHRESH) && 779 (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH)) 780 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1; 781 if (!dst_metric_locked(dst, RTAX_CWND) && 782 tp->snd_cwnd > dst_metric(dst, RTAX_CWND)) 783 dst->metrics[RTAX_CWND - 1] = tp->snd_cwnd; 784 } else if (tp->snd_cwnd > tp->snd_ssthresh && 785 icsk->icsk_ca_state == TCP_CA_Open) { 786 /* Cong. avoidance phase, cwnd is reliable. */ 787 if (!dst_metric_locked(dst, RTAX_SSTHRESH)) 788 dst->metrics[RTAX_SSTHRESH-1] = 789 max(tp->snd_cwnd >> 1, tp->snd_ssthresh); 790 if (!dst_metric_locked(dst, RTAX_CWND)) 791 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_cwnd) >> 1; 792 } else { 793 /* Else slow start did not finish, cwnd is non-sense, 794 ssthresh may be also invalid. 795 */ 796 if (!dst_metric_locked(dst, RTAX_CWND)) 797 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_ssthresh) >> 1; 798 if (dst->metrics[RTAX_SSTHRESH-1] && 799 !dst_metric_locked(dst, RTAX_SSTHRESH) && 800 tp->snd_ssthresh > dst->metrics[RTAX_SSTHRESH-1]) 801 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh; 802 } 803 804 if (!dst_metric_locked(dst, RTAX_REORDERING)) { 805 if (dst->metrics[RTAX_REORDERING-1] < tp->reordering && 806 tp->reordering != sysctl_tcp_reordering) 807 dst->metrics[RTAX_REORDERING-1] = tp->reordering; 808 } 809 } 810 } 811 812 /* Numbers are taken from RFC3390. 813 * 814 * John Heffner states: 815 * 816 * The RFC specifies a window of no more than 4380 bytes 817 * unless 2*MSS > 4380. Reading the pseudocode in the RFC 818 * is a bit misleading because they use a clamp at 4380 bytes 819 * rather than use a multiplier in the relevant range. 820 */ 821 __u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst) 822 { 823 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0); 824 825 if (!cwnd) { 826 if (tp->mss_cache > 1460) 827 cwnd = 2; 828 else 829 cwnd = (tp->mss_cache > 1095) ? 3 : 4; 830 } 831 return min_t(__u32, cwnd, tp->snd_cwnd_clamp); 832 } 833 834 /* Set slow start threshold and cwnd not falling to slow start */ 835 void tcp_enter_cwr(struct sock *sk, const int set_ssthresh) 836 { 837 struct tcp_sock *tp = tcp_sk(sk); 838 const struct inet_connection_sock *icsk = inet_csk(sk); 839 840 tp->prior_ssthresh = 0; 841 tp->bytes_acked = 0; 842 if (icsk->icsk_ca_state < TCP_CA_CWR) { 843 tp->undo_marker = 0; 844 if (set_ssthresh) 845 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 846 tp->snd_cwnd = min(tp->snd_cwnd, 847 tcp_packets_in_flight(tp) + 1U); 848 tp->snd_cwnd_cnt = 0; 849 tp->high_seq = tp->snd_nxt; 850 tp->snd_cwnd_stamp = tcp_time_stamp; 851 TCP_ECN_queue_cwr(tp); 852 853 tcp_set_ca_state(sk, TCP_CA_CWR); 854 } 855 } 856 857 /* 858 * Packet counting of FACK is based on in-order assumptions, therefore TCP 859 * disables it when reordering is detected 860 */ 861 static void tcp_disable_fack(struct tcp_sock *tp) 862 { 863 /* RFC3517 uses different metric in lost marker => reset on change */ 864 if (tcp_is_fack(tp)) 865 tp->lost_skb_hint = NULL; 866 tp->rx_opt.sack_ok &= ~2; 867 } 868 869 /* Take a notice that peer is sending D-SACKs */ 870 static void tcp_dsack_seen(struct tcp_sock *tp) 871 { 872 tp->rx_opt.sack_ok |= 4; 873 } 874 875 /* Initialize metrics on socket. */ 876 877 static void tcp_init_metrics(struct sock *sk) 878 { 879 struct tcp_sock *tp = tcp_sk(sk); 880 struct dst_entry *dst = __sk_dst_get(sk); 881 882 if (dst == NULL) 883 goto reset; 884 885 dst_confirm(dst); 886 887 if (dst_metric_locked(dst, RTAX_CWND)) 888 tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND); 889 if (dst_metric(dst, RTAX_SSTHRESH)) { 890 tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH); 891 if (tp->snd_ssthresh > tp->snd_cwnd_clamp) 892 tp->snd_ssthresh = tp->snd_cwnd_clamp; 893 } 894 if (dst_metric(dst, RTAX_REORDERING) && 895 tp->reordering != dst_metric(dst, RTAX_REORDERING)) { 896 tcp_disable_fack(tp); 897 tp->reordering = dst_metric(dst, RTAX_REORDERING); 898 } 899 900 if (dst_metric(dst, RTAX_RTT) == 0) 901 goto reset; 902 903 if (!tp->srtt && dst_metric(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3)) 904 goto reset; 905 906 /* Initial rtt is determined from SYN,SYN-ACK. 907 * The segment is small and rtt may appear much 908 * less than real one. Use per-dst memory 909 * to make it more realistic. 910 * 911 * A bit of theory. RTT is time passed after "normal" sized packet 912 * is sent until it is ACKed. In normal circumstances sending small 913 * packets force peer to delay ACKs and calculation is correct too. 914 * The algorithm is adaptive and, provided we follow specs, it 915 * NEVER underestimate RTT. BUT! If peer tries to make some clever 916 * tricks sort of "quick acks" for time long enough to decrease RTT 917 * to low value, and then abruptly stops to do it and starts to delay 918 * ACKs, wait for troubles. 919 */ 920 if (dst_metric(dst, RTAX_RTT) > tp->srtt) { 921 tp->srtt = dst_metric(dst, RTAX_RTT); 922 tp->rtt_seq = tp->snd_nxt; 923 } 924 if (dst_metric(dst, RTAX_RTTVAR) > tp->mdev) { 925 tp->mdev = dst_metric(dst, RTAX_RTTVAR); 926 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk)); 927 } 928 tcp_set_rto(sk); 929 tcp_bound_rto(sk); 930 if (inet_csk(sk)->icsk_rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp) 931 goto reset; 932 tp->snd_cwnd = tcp_init_cwnd(tp, dst); 933 tp->snd_cwnd_stamp = tcp_time_stamp; 934 return; 935 936 reset: 937 /* Play conservative. If timestamps are not 938 * supported, TCP will fail to recalculate correct 939 * rtt, if initial rto is too small. FORGET ALL AND RESET! 940 */ 941 if (!tp->rx_opt.saw_tstamp && tp->srtt) { 942 tp->srtt = 0; 943 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT; 944 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; 945 } 946 } 947 948 static void tcp_update_reordering(struct sock *sk, const int metric, 949 const int ts) 950 { 951 struct tcp_sock *tp = tcp_sk(sk); 952 if (metric > tp->reordering) { 953 tp->reordering = min(TCP_MAX_REORDERING, metric); 954 955 /* This exciting event is worth to be remembered. 8) */ 956 if (ts) 957 NET_INC_STATS_BH(LINUX_MIB_TCPTSREORDER); 958 else if (tcp_is_reno(tp)) 959 NET_INC_STATS_BH(LINUX_MIB_TCPRENOREORDER); 960 else if (tcp_is_fack(tp)) 961 NET_INC_STATS_BH(LINUX_MIB_TCPFACKREORDER); 962 else 963 NET_INC_STATS_BH(LINUX_MIB_TCPSACKREORDER); 964 #if FASTRETRANS_DEBUG > 1 965 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n", 966 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state, 967 tp->reordering, 968 tp->fackets_out, 969 tp->sacked_out, 970 tp->undo_marker ? tp->undo_retrans : 0); 971 #endif 972 tcp_disable_fack(tp); 973 } 974 } 975 976 /* This procedure tags the retransmission queue when SACKs arrive. 977 * 978 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L). 979 * Packets in queue with these bits set are counted in variables 980 * sacked_out, retrans_out and lost_out, correspondingly. 981 * 982 * Valid combinations are: 983 * Tag InFlight Description 984 * 0 1 - orig segment is in flight. 985 * S 0 - nothing flies, orig reached receiver. 986 * L 0 - nothing flies, orig lost by net. 987 * R 2 - both orig and retransmit are in flight. 988 * L|R 1 - orig is lost, retransmit is in flight. 989 * S|R 1 - orig reached receiver, retrans is still in flight. 990 * (L|S|R is logically valid, it could occur when L|R is sacked, 991 * but it is equivalent to plain S and code short-curcuits it to S. 992 * L|S is logically invalid, it would mean -1 packet in flight 8)) 993 * 994 * These 6 states form finite state machine, controlled by the following events: 995 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue()) 996 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue()) 997 * 3. Loss detection event of one of three flavors: 998 * A. Scoreboard estimator decided the packet is lost. 999 * A'. Reno "three dupacks" marks head of queue lost. 1000 * A''. Its FACK modfication, head until snd.fack is lost. 1001 * B. SACK arrives sacking data transmitted after never retransmitted 1002 * hole was sent out. 1003 * C. SACK arrives sacking SND.NXT at the moment, when the 1004 * segment was retransmitted. 1005 * 4. D-SACK added new rule: D-SACK changes any tag to S. 1006 * 1007 * It is pleasant to note, that state diagram turns out to be commutative, 1008 * so that we are allowed not to be bothered by order of our actions, 1009 * when multiple events arrive simultaneously. (see the function below). 1010 * 1011 * Reordering detection. 1012 * -------------------- 1013 * Reordering metric is maximal distance, which a packet can be displaced 1014 * in packet stream. With SACKs we can estimate it: 1015 * 1016 * 1. SACK fills old hole and the corresponding segment was not 1017 * ever retransmitted -> reordering. Alas, we cannot use it 1018 * when segment was retransmitted. 1019 * 2. The last flaw is solved with D-SACK. D-SACK arrives 1020 * for retransmitted and already SACKed segment -> reordering.. 1021 * Both of these heuristics are not used in Loss state, when we cannot 1022 * account for retransmits accurately. 1023 * 1024 * SACK block validation. 1025 * ---------------------- 1026 * 1027 * SACK block range validation checks that the received SACK block fits to 1028 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT. 1029 * Note that SND.UNA is not included to the range though being valid because 1030 * it means that the receiver is rather inconsistent with itself reporting 1031 * SACK reneging when it should advance SND.UNA. Such SACK block this is 1032 * perfectly valid, however, in light of RFC2018 which explicitly states 1033 * that "SACK block MUST reflect the newest segment. Even if the newest 1034 * segment is going to be discarded ...", not that it looks very clever 1035 * in case of head skb. Due to potentional receiver driven attacks, we 1036 * choose to avoid immediate execution of a walk in write queue due to 1037 * reneging and defer head skb's loss recovery to standard loss recovery 1038 * procedure that will eventually trigger (nothing forbids us doing this). 1039 * 1040 * Implements also blockage to start_seq wrap-around. Problem lies in the 1041 * fact that though start_seq (s) is before end_seq (i.e., not reversed), 1042 * there's no guarantee that it will be before snd_nxt (n). The problem 1043 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt 1044 * wrap (s_w): 1045 * 1046 * <- outs wnd -> <- wrapzone -> 1047 * u e n u_w e_w s n_w 1048 * | | | | | | | 1049 * |<------------+------+----- TCP seqno space --------------+---------->| 1050 * ...-- <2^31 ->| |<--------... 1051 * ...---- >2^31 ------>| |<--------... 1052 * 1053 * Current code wouldn't be vulnerable but it's better still to discard such 1054 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat 1055 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in 1056 * snd_nxt wrap -> snd_una region will then become "well defined", i.e., 1057 * equal to the ideal case (infinite seqno space without wrap caused issues). 1058 * 1059 * With D-SACK the lower bound is extended to cover sequence space below 1060 * SND.UNA down to undo_marker, which is the last point of interest. Yet 1061 * again, D-SACK block must not to go across snd_una (for the same reason as 1062 * for the normal SACK blocks, explained above). But there all simplicity 1063 * ends, TCP might receive valid D-SACKs below that. As long as they reside 1064 * fully below undo_marker they do not affect behavior in anyway and can 1065 * therefore be safely ignored. In rare cases (which are more or less 1066 * theoretical ones), the D-SACK will nicely cross that boundary due to skb 1067 * fragmentation and packet reordering past skb's retransmission. To consider 1068 * them correctly, the acceptable range must be extended even more though 1069 * the exact amount is rather hard to quantify. However, tp->max_window can 1070 * be used as an exaggerated estimate. 1071 */ 1072 static int tcp_is_sackblock_valid(struct tcp_sock *tp, int is_dsack, 1073 u32 start_seq, u32 end_seq) 1074 { 1075 /* Too far in future, or reversed (interpretation is ambiguous) */ 1076 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq)) 1077 return 0; 1078 1079 /* Nasty start_seq wrap-around check (see comments above) */ 1080 if (!before(start_seq, tp->snd_nxt)) 1081 return 0; 1082 1083 /* In outstanding window? ...This is valid exit for D-SACKs too. 1084 * start_seq == snd_una is non-sensical (see comments above) 1085 */ 1086 if (after(start_seq, tp->snd_una)) 1087 return 1; 1088 1089 if (!is_dsack || !tp->undo_marker) 1090 return 0; 1091 1092 /* ...Then it's D-SACK, and must reside below snd_una completely */ 1093 if (!after(end_seq, tp->snd_una)) 1094 return 0; 1095 1096 if (!before(start_seq, tp->undo_marker)) 1097 return 1; 1098 1099 /* Too old */ 1100 if (!after(end_seq, tp->undo_marker)) 1101 return 0; 1102 1103 /* Undo_marker boundary crossing (overestimates a lot). Known already: 1104 * start_seq < undo_marker and end_seq >= undo_marker. 1105 */ 1106 return !before(start_seq, end_seq - tp->max_window); 1107 } 1108 1109 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving". 1110 * Event "C". Later note: FACK people cheated me again 8), we have to account 1111 * for reordering! Ugly, but should help. 1112 * 1113 * Search retransmitted skbs from write_queue that were sent when snd_nxt was 1114 * less than what is now known to be received by the other end (derived from 1115 * highest SACK block). Also calculate the lowest snd_nxt among the remaining 1116 * retransmitted skbs to avoid some costly processing per ACKs. 1117 */ 1118 static void tcp_mark_lost_retrans(struct sock *sk) 1119 { 1120 const struct inet_connection_sock *icsk = inet_csk(sk); 1121 struct tcp_sock *tp = tcp_sk(sk); 1122 struct sk_buff *skb; 1123 int cnt = 0; 1124 u32 new_low_seq = tp->snd_nxt; 1125 u32 received_upto = tcp_highest_sack_seq(tp); 1126 1127 if (!tcp_is_fack(tp) || !tp->retrans_out || 1128 !after(received_upto, tp->lost_retrans_low) || 1129 icsk->icsk_ca_state != TCP_CA_Recovery) 1130 return; 1131 1132 tcp_for_write_queue(skb, sk) { 1133 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq; 1134 1135 if (skb == tcp_send_head(sk)) 1136 break; 1137 if (cnt == tp->retrans_out) 1138 break; 1139 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 1140 continue; 1141 1142 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)) 1143 continue; 1144 1145 if (after(received_upto, ack_seq) && 1146 (tcp_is_fack(tp) || 1147 !before(received_upto, 1148 ack_seq + tp->reordering * tp->mss_cache))) { 1149 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 1150 tp->retrans_out -= tcp_skb_pcount(skb); 1151 1152 /* clear lost hint */ 1153 tp->retransmit_skb_hint = NULL; 1154 1155 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) { 1156 tp->lost_out += tcp_skb_pcount(skb); 1157 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 1158 } 1159 NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT); 1160 } else { 1161 if (before(ack_seq, new_low_seq)) 1162 new_low_seq = ack_seq; 1163 cnt += tcp_skb_pcount(skb); 1164 } 1165 } 1166 1167 if (tp->retrans_out) 1168 tp->lost_retrans_low = new_low_seq; 1169 } 1170 1171 static int tcp_check_dsack(struct tcp_sock *tp, struct sk_buff *ack_skb, 1172 struct tcp_sack_block_wire *sp, int num_sacks, 1173 u32 prior_snd_una) 1174 { 1175 u32 start_seq_0 = ntohl(get_unaligned(&sp[0].start_seq)); 1176 u32 end_seq_0 = ntohl(get_unaligned(&sp[0].end_seq)); 1177 int dup_sack = 0; 1178 1179 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) { 1180 dup_sack = 1; 1181 tcp_dsack_seen(tp); 1182 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV); 1183 } else if (num_sacks > 1) { 1184 u32 end_seq_1 = ntohl(get_unaligned(&sp[1].end_seq)); 1185 u32 start_seq_1 = ntohl(get_unaligned(&sp[1].start_seq)); 1186 1187 if (!after(end_seq_0, end_seq_1) && 1188 !before(start_seq_0, start_seq_1)) { 1189 dup_sack = 1; 1190 tcp_dsack_seen(tp); 1191 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV); 1192 } 1193 } 1194 1195 /* D-SACK for already forgotten data... Do dumb counting. */ 1196 if (dup_sack && 1197 !after(end_seq_0, prior_snd_una) && 1198 after(end_seq_0, tp->undo_marker)) 1199 tp->undo_retrans--; 1200 1201 return dup_sack; 1202 } 1203 1204 /* Check if skb is fully within the SACK block. In presence of GSO skbs, 1205 * the incoming SACK may not exactly match but we can find smaller MSS 1206 * aligned portion of it that matches. Therefore we might need to fragment 1207 * which may fail and creates some hassle (caller must handle error case 1208 * returns). 1209 */ 1210 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb, 1211 u32 start_seq, u32 end_seq) 1212 { 1213 int in_sack, err; 1214 unsigned int pkt_len; 1215 1216 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && 1217 !before(end_seq, TCP_SKB_CB(skb)->end_seq); 1218 1219 if (tcp_skb_pcount(skb) > 1 && !in_sack && 1220 after(TCP_SKB_CB(skb)->end_seq, start_seq)) { 1221 1222 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); 1223 1224 if (!in_sack) 1225 pkt_len = start_seq - TCP_SKB_CB(skb)->seq; 1226 else 1227 pkt_len = end_seq - TCP_SKB_CB(skb)->seq; 1228 err = tcp_fragment(sk, skb, pkt_len, skb_shinfo(skb)->gso_size); 1229 if (err < 0) 1230 return err; 1231 } 1232 1233 return in_sack; 1234 } 1235 1236 static int tcp_sacktag_one(struct sk_buff *skb, struct sock *sk, 1237 int *reord, int dup_sack, int fack_count) 1238 { 1239 struct tcp_sock *tp = tcp_sk(sk); 1240 u8 sacked = TCP_SKB_CB(skb)->sacked; 1241 int flag = 0; 1242 1243 /* Account D-SACK for retransmitted packet. */ 1244 if (dup_sack && (sacked & TCPCB_RETRANS)) { 1245 if (after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker)) 1246 tp->undo_retrans--; 1247 if (sacked & TCPCB_SACKED_ACKED) 1248 *reord = min(fack_count, *reord); 1249 } 1250 1251 /* Nothing to do; acked frame is about to be dropped (was ACKed). */ 1252 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 1253 return flag; 1254 1255 if (!(sacked & TCPCB_SACKED_ACKED)) { 1256 if (sacked & TCPCB_SACKED_RETRANS) { 1257 /* If the segment is not tagged as lost, 1258 * we do not clear RETRANS, believing 1259 * that retransmission is still in flight. 1260 */ 1261 if (sacked & TCPCB_LOST) { 1262 TCP_SKB_CB(skb)->sacked &= 1263 ~(TCPCB_LOST|TCPCB_SACKED_RETRANS); 1264 tp->lost_out -= tcp_skb_pcount(skb); 1265 tp->retrans_out -= tcp_skb_pcount(skb); 1266 1267 /* clear lost hint */ 1268 tp->retransmit_skb_hint = NULL; 1269 } 1270 } else { 1271 if (!(sacked & TCPCB_RETRANS)) { 1272 /* New sack for not retransmitted frame, 1273 * which was in hole. It is reordering. 1274 */ 1275 if (before(TCP_SKB_CB(skb)->seq, 1276 tcp_highest_sack_seq(tp))) 1277 *reord = min(fack_count, *reord); 1278 1279 /* SACK enhanced F-RTO (RFC4138; Appendix B) */ 1280 if (!after(TCP_SKB_CB(skb)->end_seq, tp->frto_highmark)) 1281 flag |= FLAG_ONLY_ORIG_SACKED; 1282 } 1283 1284 if (sacked & TCPCB_LOST) { 1285 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; 1286 tp->lost_out -= tcp_skb_pcount(skb); 1287 1288 /* clear lost hint */ 1289 tp->retransmit_skb_hint = NULL; 1290 } 1291 } 1292 1293 TCP_SKB_CB(skb)->sacked |= TCPCB_SACKED_ACKED; 1294 flag |= FLAG_DATA_SACKED; 1295 tp->sacked_out += tcp_skb_pcount(skb); 1296 1297 fack_count += tcp_skb_pcount(skb); 1298 1299 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */ 1300 if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) && 1301 before(TCP_SKB_CB(skb)->seq, 1302 TCP_SKB_CB(tp->lost_skb_hint)->seq)) 1303 tp->lost_cnt_hint += tcp_skb_pcount(skb); 1304 1305 if (fack_count > tp->fackets_out) 1306 tp->fackets_out = fack_count; 1307 1308 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp))) 1309 tcp_advance_highest_sack(sk, skb); 1310 } 1311 1312 /* D-SACK. We can detect redundant retransmission in S|R and plain R 1313 * frames and clear it. undo_retrans is decreased above, L|R frames 1314 * are accounted above as well. 1315 */ 1316 if (dup_sack && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)) { 1317 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 1318 tp->retrans_out -= tcp_skb_pcount(skb); 1319 tp->retransmit_skb_hint = NULL; 1320 } 1321 1322 return flag; 1323 } 1324 1325 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk, 1326 struct tcp_sack_block *next_dup, 1327 u32 start_seq, u32 end_seq, 1328 int dup_sack_in, int *fack_count, 1329 int *reord, int *flag) 1330 { 1331 tcp_for_write_queue_from(skb, sk) { 1332 int in_sack = 0; 1333 int dup_sack = dup_sack_in; 1334 1335 if (skb == tcp_send_head(sk)) 1336 break; 1337 1338 /* queue is in-order => we can short-circuit the walk early */ 1339 if (!before(TCP_SKB_CB(skb)->seq, end_seq)) 1340 break; 1341 1342 if ((next_dup != NULL) && 1343 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) { 1344 in_sack = tcp_match_skb_to_sack(sk, skb, 1345 next_dup->start_seq, 1346 next_dup->end_seq); 1347 if (in_sack > 0) 1348 dup_sack = 1; 1349 } 1350 1351 if (in_sack <= 0) 1352 in_sack = tcp_match_skb_to_sack(sk, skb, start_seq, 1353 end_seq); 1354 if (unlikely(in_sack < 0)) 1355 break; 1356 1357 if (in_sack) 1358 *flag |= tcp_sacktag_one(skb, sk, reord, dup_sack, 1359 *fack_count); 1360 1361 *fack_count += tcp_skb_pcount(skb); 1362 } 1363 return skb; 1364 } 1365 1366 /* Avoid all extra work that is being done by sacktag while walking in 1367 * a normal way 1368 */ 1369 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk, 1370 u32 skip_to_seq) 1371 { 1372 tcp_for_write_queue_from(skb, sk) { 1373 if (skb == tcp_send_head(sk)) 1374 break; 1375 1376 if (!before(TCP_SKB_CB(skb)->end_seq, skip_to_seq)) 1377 break; 1378 } 1379 return skb; 1380 } 1381 1382 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb, 1383 struct sock *sk, 1384 struct tcp_sack_block *next_dup, 1385 u32 skip_to_seq, 1386 int *fack_count, int *reord, 1387 int *flag) 1388 { 1389 if (next_dup == NULL) 1390 return skb; 1391 1392 if (before(next_dup->start_seq, skip_to_seq)) { 1393 skb = tcp_sacktag_skip(skb, sk, next_dup->start_seq); 1394 tcp_sacktag_walk(skb, sk, NULL, 1395 next_dup->start_seq, next_dup->end_seq, 1396 1, fack_count, reord, flag); 1397 } 1398 1399 return skb; 1400 } 1401 1402 static int tcp_sack_cache_ok(struct tcp_sock *tp, struct tcp_sack_block *cache) 1403 { 1404 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache); 1405 } 1406 1407 static int 1408 tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, 1409 u32 prior_snd_una) 1410 { 1411 const struct inet_connection_sock *icsk = inet_csk(sk); 1412 struct tcp_sock *tp = tcp_sk(sk); 1413 unsigned char *ptr = (skb_transport_header(ack_skb) + 1414 TCP_SKB_CB(ack_skb)->sacked); 1415 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2); 1416 struct tcp_sack_block sp[4]; 1417 struct tcp_sack_block *cache; 1418 struct sk_buff *skb; 1419 int num_sacks = (ptr[1] - TCPOLEN_SACK_BASE) >> 3; 1420 int used_sacks; 1421 int reord = tp->packets_out; 1422 int flag = 0; 1423 int found_dup_sack = 0; 1424 int fack_count; 1425 int i, j; 1426 int first_sack_index; 1427 1428 if (!tp->sacked_out) { 1429 if (WARN_ON(tp->fackets_out)) 1430 tp->fackets_out = 0; 1431 tcp_highest_sack_reset(sk); 1432 } 1433 1434 found_dup_sack = tcp_check_dsack(tp, ack_skb, sp_wire, 1435 num_sacks, prior_snd_una); 1436 if (found_dup_sack) 1437 flag |= FLAG_DSACKING_ACK; 1438 1439 /* Eliminate too old ACKs, but take into 1440 * account more or less fresh ones, they can 1441 * contain valid SACK info. 1442 */ 1443 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window)) 1444 return 0; 1445 1446 if (!tp->packets_out) 1447 goto out; 1448 1449 used_sacks = 0; 1450 first_sack_index = 0; 1451 for (i = 0; i < num_sacks; i++) { 1452 int dup_sack = !i && found_dup_sack; 1453 1454 sp[used_sacks].start_seq = ntohl(get_unaligned(&sp_wire[i].start_seq)); 1455 sp[used_sacks].end_seq = ntohl(get_unaligned(&sp_wire[i].end_seq)); 1456 1457 if (!tcp_is_sackblock_valid(tp, dup_sack, 1458 sp[used_sacks].start_seq, 1459 sp[used_sacks].end_seq)) { 1460 if (dup_sack) { 1461 if (!tp->undo_marker) 1462 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKIGNOREDNOUNDO); 1463 else 1464 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKIGNOREDOLD); 1465 } else { 1466 /* Don't count olds caused by ACK reordering */ 1467 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) && 1468 !after(sp[used_sacks].end_seq, tp->snd_una)) 1469 continue; 1470 NET_INC_STATS_BH(LINUX_MIB_TCPSACKDISCARD); 1471 } 1472 if (i == 0) 1473 first_sack_index = -1; 1474 continue; 1475 } 1476 1477 /* Ignore very old stuff early */ 1478 if (!after(sp[used_sacks].end_seq, prior_snd_una)) 1479 continue; 1480 1481 used_sacks++; 1482 } 1483 1484 /* order SACK blocks to allow in order walk of the retrans queue */ 1485 for (i = used_sacks - 1; i > 0; i--) { 1486 for (j = 0; j < i; j++) { 1487 if (after(sp[j].start_seq, sp[j + 1].start_seq)) { 1488 struct tcp_sack_block tmp; 1489 1490 tmp = sp[j]; 1491 sp[j] = sp[j + 1]; 1492 sp[j + 1] = tmp; 1493 1494 /* Track where the first SACK block goes to */ 1495 if (j == first_sack_index) 1496 first_sack_index = j + 1; 1497 } 1498 } 1499 } 1500 1501 skb = tcp_write_queue_head(sk); 1502 fack_count = 0; 1503 i = 0; 1504 1505 if (!tp->sacked_out) { 1506 /* It's already past, so skip checking against it */ 1507 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache); 1508 } else { 1509 cache = tp->recv_sack_cache; 1510 /* Skip empty blocks in at head of the cache */ 1511 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq && 1512 !cache->end_seq) 1513 cache++; 1514 } 1515 1516 while (i < used_sacks) { 1517 u32 start_seq = sp[i].start_seq; 1518 u32 end_seq = sp[i].end_seq; 1519 int dup_sack = (found_dup_sack && (i == first_sack_index)); 1520 struct tcp_sack_block *next_dup = NULL; 1521 1522 if (found_dup_sack && ((i + 1) == first_sack_index)) 1523 next_dup = &sp[i + 1]; 1524 1525 /* Event "B" in the comment above. */ 1526 if (after(end_seq, tp->high_seq)) 1527 flag |= FLAG_DATA_LOST; 1528 1529 /* Skip too early cached blocks */ 1530 while (tcp_sack_cache_ok(tp, cache) && 1531 !before(start_seq, cache->end_seq)) 1532 cache++; 1533 1534 /* Can skip some work by looking recv_sack_cache? */ 1535 if (tcp_sack_cache_ok(tp, cache) && !dup_sack && 1536 after(end_seq, cache->start_seq)) { 1537 1538 /* Head todo? */ 1539 if (before(start_seq, cache->start_seq)) { 1540 skb = tcp_sacktag_skip(skb, sk, start_seq); 1541 skb = tcp_sacktag_walk(skb, sk, next_dup, 1542 start_seq, 1543 cache->start_seq, 1544 dup_sack, &fack_count, 1545 &reord, &flag); 1546 } 1547 1548 /* Rest of the block already fully processed? */ 1549 if (!after(end_seq, cache->end_seq)) 1550 goto advance_sp; 1551 1552 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup, 1553 cache->end_seq, 1554 &fack_count, &reord, 1555 &flag); 1556 1557 /* ...tail remains todo... */ 1558 if (tcp_highest_sack_seq(tp) == cache->end_seq) { 1559 /* ...but better entrypoint exists! */ 1560 skb = tcp_highest_sack(sk); 1561 if (skb == NULL) 1562 break; 1563 fack_count = tp->fackets_out; 1564 cache++; 1565 goto walk; 1566 } 1567 1568 skb = tcp_sacktag_skip(skb, sk, cache->end_seq); 1569 /* Check overlap against next cached too (past this one already) */ 1570 cache++; 1571 continue; 1572 } 1573 1574 if (!before(start_seq, tcp_highest_sack_seq(tp))) { 1575 skb = tcp_highest_sack(sk); 1576 if (skb == NULL) 1577 break; 1578 fack_count = tp->fackets_out; 1579 } 1580 skb = tcp_sacktag_skip(skb, sk, start_seq); 1581 1582 walk: 1583 skb = tcp_sacktag_walk(skb, sk, next_dup, start_seq, end_seq, 1584 dup_sack, &fack_count, &reord, &flag); 1585 1586 advance_sp: 1587 /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct 1588 * due to in-order walk 1589 */ 1590 if (after(end_seq, tp->frto_highmark)) 1591 flag &= ~FLAG_ONLY_ORIG_SACKED; 1592 1593 i++; 1594 } 1595 1596 /* Clear the head of the cache sack blocks so we can skip it next time */ 1597 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) { 1598 tp->recv_sack_cache[i].start_seq = 0; 1599 tp->recv_sack_cache[i].end_seq = 0; 1600 } 1601 for (j = 0; j < used_sacks; j++) 1602 tp->recv_sack_cache[i++] = sp[j]; 1603 1604 tcp_mark_lost_retrans(sk); 1605 1606 tcp_verify_left_out(tp); 1607 1608 if ((reord < tp->fackets_out) && 1609 ((icsk->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker) && 1610 (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark))) 1611 tcp_update_reordering(sk, tp->fackets_out - reord, 0); 1612 1613 out: 1614 1615 #if FASTRETRANS_DEBUG > 0 1616 BUG_TRAP((int)tp->sacked_out >= 0); 1617 BUG_TRAP((int)tp->lost_out >= 0); 1618 BUG_TRAP((int)tp->retrans_out >= 0); 1619 BUG_TRAP((int)tcp_packets_in_flight(tp) >= 0); 1620 #endif 1621 return flag; 1622 } 1623 1624 /* If we receive more dupacks than we expected counting segments 1625 * in assumption of absent reordering, interpret this as reordering. 1626 * The only another reason could be bug in receiver TCP. 1627 */ 1628 static void tcp_check_reno_reordering(struct sock *sk, const int addend) 1629 { 1630 struct tcp_sock *tp = tcp_sk(sk); 1631 u32 holes; 1632 1633 holes = max(tp->lost_out, 1U); 1634 holes = min(holes, tp->packets_out); 1635 1636 if ((tp->sacked_out + holes) > tp->packets_out) { 1637 tp->sacked_out = tp->packets_out - holes; 1638 tcp_update_reordering(sk, tp->packets_out + addend, 0); 1639 } 1640 } 1641 1642 /* Emulate SACKs for SACKless connection: account for a new dupack. */ 1643 1644 static void tcp_add_reno_sack(struct sock *sk) 1645 { 1646 struct tcp_sock *tp = tcp_sk(sk); 1647 tp->sacked_out++; 1648 tcp_check_reno_reordering(sk, 0); 1649 tcp_verify_left_out(tp); 1650 } 1651 1652 /* Account for ACK, ACKing some data in Reno Recovery phase. */ 1653 1654 static void tcp_remove_reno_sacks(struct sock *sk, int acked) 1655 { 1656 struct tcp_sock *tp = tcp_sk(sk); 1657 1658 if (acked > 0) { 1659 /* One ACK acked hole. The rest eat duplicate ACKs. */ 1660 if (acked - 1 >= tp->sacked_out) 1661 tp->sacked_out = 0; 1662 else 1663 tp->sacked_out -= acked - 1; 1664 } 1665 tcp_check_reno_reordering(sk, acked); 1666 tcp_verify_left_out(tp); 1667 } 1668 1669 static inline void tcp_reset_reno_sack(struct tcp_sock *tp) 1670 { 1671 tp->sacked_out = 0; 1672 } 1673 1674 /* F-RTO can only be used if TCP has never retransmitted anything other than 1675 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here) 1676 */ 1677 int tcp_use_frto(struct sock *sk) 1678 { 1679 const struct tcp_sock *tp = tcp_sk(sk); 1680 struct sk_buff *skb; 1681 1682 if (!sysctl_tcp_frto) 1683 return 0; 1684 1685 if (IsSackFrto()) 1686 return 1; 1687 1688 /* Avoid expensive walking of rexmit queue if possible */ 1689 if (tp->retrans_out > 1) 1690 return 0; 1691 1692 skb = tcp_write_queue_head(sk); 1693 skb = tcp_write_queue_next(sk, skb); /* Skips head */ 1694 tcp_for_write_queue_from(skb, sk) { 1695 if (skb == tcp_send_head(sk)) 1696 break; 1697 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) 1698 return 0; 1699 /* Short-circuit when first non-SACKed skb has been checked */ 1700 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) 1701 break; 1702 } 1703 return 1; 1704 } 1705 1706 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO 1707 * recovery a bit and use heuristics in tcp_process_frto() to detect if 1708 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to 1709 * keep retrans_out counting accurate (with SACK F-RTO, other than head 1710 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS 1711 * bits are handled if the Loss state is really to be entered (in 1712 * tcp_enter_frto_loss). 1713 * 1714 * Do like tcp_enter_loss() would; when RTO expires the second time it 1715 * does: 1716 * "Reduce ssthresh if it has not yet been made inside this window." 1717 */ 1718 void tcp_enter_frto(struct sock *sk) 1719 { 1720 const struct inet_connection_sock *icsk = inet_csk(sk); 1721 struct tcp_sock *tp = tcp_sk(sk); 1722 struct sk_buff *skb; 1723 1724 if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) || 1725 tp->snd_una == tp->high_seq || 1726 ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) && 1727 !icsk->icsk_retransmits)) { 1728 tp->prior_ssthresh = tcp_current_ssthresh(sk); 1729 /* Our state is too optimistic in ssthresh() call because cwnd 1730 * is not reduced until tcp_enter_frto_loss() when previous F-RTO 1731 * recovery has not yet completed. Pattern would be this: RTO, 1732 * Cumulative ACK, RTO (2xRTO for the same segment does not end 1733 * up here twice). 1734 * RFC4138 should be more specific on what to do, even though 1735 * RTO is quite unlikely to occur after the first Cumulative ACK 1736 * due to back-off and complexity of triggering events ... 1737 */ 1738 if (tp->frto_counter) { 1739 u32 stored_cwnd; 1740 stored_cwnd = tp->snd_cwnd; 1741 tp->snd_cwnd = 2; 1742 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 1743 tp->snd_cwnd = stored_cwnd; 1744 } else { 1745 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 1746 } 1747 /* ... in theory, cong.control module could do "any tricks" in 1748 * ssthresh(), which means that ca_state, lost bits and lost_out 1749 * counter would have to be faked before the call occurs. We 1750 * consider that too expensive, unlikely and hacky, so modules 1751 * using these in ssthresh() must deal these incompatibility 1752 * issues if they receives CA_EVENT_FRTO and frto_counter != 0 1753 */ 1754 tcp_ca_event(sk, CA_EVENT_FRTO); 1755 } 1756 1757 tp->undo_marker = tp->snd_una; 1758 tp->undo_retrans = 0; 1759 1760 skb = tcp_write_queue_head(sk); 1761 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) 1762 tp->undo_marker = 0; 1763 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 1764 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 1765 tp->retrans_out -= tcp_skb_pcount(skb); 1766 } 1767 tcp_verify_left_out(tp); 1768 1769 /* Too bad if TCP was application limited */ 1770 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1); 1771 1772 /* Earlier loss recovery underway (see RFC4138; Appendix B). 1773 * The last condition is necessary at least in tp->frto_counter case. 1774 */ 1775 if (IsSackFrto() && (tp->frto_counter || 1776 ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) && 1777 after(tp->high_seq, tp->snd_una)) { 1778 tp->frto_highmark = tp->high_seq; 1779 } else { 1780 tp->frto_highmark = tp->snd_nxt; 1781 } 1782 tcp_set_ca_state(sk, TCP_CA_Disorder); 1783 tp->high_seq = tp->snd_nxt; 1784 tp->frto_counter = 1; 1785 } 1786 1787 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO, 1788 * which indicates that we should follow the traditional RTO recovery, 1789 * i.e. mark everything lost and do go-back-N retransmission. 1790 */ 1791 static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag) 1792 { 1793 struct tcp_sock *tp = tcp_sk(sk); 1794 struct sk_buff *skb; 1795 1796 tp->lost_out = 0; 1797 tp->retrans_out = 0; 1798 if (tcp_is_reno(tp)) 1799 tcp_reset_reno_sack(tp); 1800 1801 tcp_for_write_queue(skb, sk) { 1802 if (skb == tcp_send_head(sk)) 1803 break; 1804 1805 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; 1806 /* 1807 * Count the retransmission made on RTO correctly (only when 1808 * waiting for the first ACK and did not get it)... 1809 */ 1810 if ((tp->frto_counter == 1) && !(flag & FLAG_DATA_ACKED)) { 1811 /* For some reason this R-bit might get cleared? */ 1812 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 1813 tp->retrans_out += tcp_skb_pcount(skb); 1814 /* ...enter this if branch just for the first segment */ 1815 flag |= FLAG_DATA_ACKED; 1816 } else { 1817 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) 1818 tp->undo_marker = 0; 1819 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 1820 } 1821 1822 /* Don't lost mark skbs that were fwd transmitted after RTO */ 1823 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) && 1824 !after(TCP_SKB_CB(skb)->end_seq, tp->frto_highmark)) { 1825 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 1826 tp->lost_out += tcp_skb_pcount(skb); 1827 } 1828 } 1829 tcp_verify_left_out(tp); 1830 1831 tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments; 1832 tp->snd_cwnd_cnt = 0; 1833 tp->snd_cwnd_stamp = tcp_time_stamp; 1834 tp->frto_counter = 0; 1835 tp->bytes_acked = 0; 1836 1837 tp->reordering = min_t(unsigned int, tp->reordering, 1838 sysctl_tcp_reordering); 1839 tcp_set_ca_state(sk, TCP_CA_Loss); 1840 tp->high_seq = tp->frto_highmark; 1841 TCP_ECN_queue_cwr(tp); 1842 1843 tcp_clear_retrans_hints_partial(tp); 1844 } 1845 1846 static void tcp_clear_retrans_partial(struct tcp_sock *tp) 1847 { 1848 tp->retrans_out = 0; 1849 tp->lost_out = 0; 1850 1851 tp->undo_marker = 0; 1852 tp->undo_retrans = 0; 1853 } 1854 1855 void tcp_clear_retrans(struct tcp_sock *tp) 1856 { 1857 tcp_clear_retrans_partial(tp); 1858 1859 tp->fackets_out = 0; 1860 tp->sacked_out = 0; 1861 } 1862 1863 /* Enter Loss state. If "how" is not zero, forget all SACK information 1864 * and reset tags completely, otherwise preserve SACKs. If receiver 1865 * dropped its ofo queue, we will know this due to reneging detection. 1866 */ 1867 void tcp_enter_loss(struct sock *sk, int how) 1868 { 1869 const struct inet_connection_sock *icsk = inet_csk(sk); 1870 struct tcp_sock *tp = tcp_sk(sk); 1871 struct sk_buff *skb; 1872 1873 /* Reduce ssthresh if it has not yet been made inside this window. */ 1874 if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq || 1875 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) { 1876 tp->prior_ssthresh = tcp_current_ssthresh(sk); 1877 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 1878 tcp_ca_event(sk, CA_EVENT_LOSS); 1879 } 1880 tp->snd_cwnd = 1; 1881 tp->snd_cwnd_cnt = 0; 1882 tp->snd_cwnd_stamp = tcp_time_stamp; 1883 1884 tp->bytes_acked = 0; 1885 tcp_clear_retrans_partial(tp); 1886 1887 if (tcp_is_reno(tp)) 1888 tcp_reset_reno_sack(tp); 1889 1890 if (!how) { 1891 /* Push undo marker, if it was plain RTO and nothing 1892 * was retransmitted. */ 1893 tp->undo_marker = tp->snd_una; 1894 tcp_clear_retrans_hints_partial(tp); 1895 } else { 1896 tp->sacked_out = 0; 1897 tp->fackets_out = 0; 1898 tcp_clear_all_retrans_hints(tp); 1899 } 1900 1901 tcp_for_write_queue(skb, sk) { 1902 if (skb == tcp_send_head(sk)) 1903 break; 1904 1905 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) 1906 tp->undo_marker = 0; 1907 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED; 1908 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) { 1909 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED; 1910 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 1911 tp->lost_out += tcp_skb_pcount(skb); 1912 } 1913 } 1914 tcp_verify_left_out(tp); 1915 1916 tp->reordering = min_t(unsigned int, tp->reordering, 1917 sysctl_tcp_reordering); 1918 tcp_set_ca_state(sk, TCP_CA_Loss); 1919 tp->high_seq = tp->snd_nxt; 1920 TCP_ECN_queue_cwr(tp); 1921 /* Abort F-RTO algorithm if one is in progress */ 1922 tp->frto_counter = 0; 1923 } 1924 1925 /* If ACK arrived pointing to a remembered SACK, it means that our 1926 * remembered SACKs do not reflect real state of receiver i.e. 1927 * receiver _host_ is heavily congested (or buggy). 1928 * 1929 * Do processing similar to RTO timeout. 1930 */ 1931 static int tcp_check_sack_reneging(struct sock *sk, int flag) 1932 { 1933 if (flag & FLAG_SACK_RENEGING) { 1934 struct inet_connection_sock *icsk = inet_csk(sk); 1935 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING); 1936 1937 tcp_enter_loss(sk, 1); 1938 icsk->icsk_retransmits++; 1939 tcp_retransmit_skb(sk, tcp_write_queue_head(sk)); 1940 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 1941 icsk->icsk_rto, TCP_RTO_MAX); 1942 return 1; 1943 } 1944 return 0; 1945 } 1946 1947 static inline int tcp_fackets_out(struct tcp_sock *tp) 1948 { 1949 return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out; 1950 } 1951 1952 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs 1953 * counter when SACK is enabled (without SACK, sacked_out is used for 1954 * that purpose). 1955 * 1956 * Instead, with FACK TCP uses fackets_out that includes both SACKed 1957 * segments up to the highest received SACK block so far and holes in 1958 * between them. 1959 * 1960 * With reordering, holes may still be in flight, so RFC3517 recovery 1961 * uses pure sacked_out (total number of SACKed segments) even though 1962 * it violates the RFC that uses duplicate ACKs, often these are equal 1963 * but when e.g. out-of-window ACKs or packet duplication occurs, 1964 * they differ. Since neither occurs due to loss, TCP should really 1965 * ignore them. 1966 */ 1967 static inline int tcp_dupack_heurestics(struct tcp_sock *tp) 1968 { 1969 return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1; 1970 } 1971 1972 static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb) 1973 { 1974 return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto); 1975 } 1976 1977 static inline int tcp_head_timedout(struct sock *sk) 1978 { 1979 struct tcp_sock *tp = tcp_sk(sk); 1980 1981 return tp->packets_out && 1982 tcp_skb_timedout(sk, tcp_write_queue_head(sk)); 1983 } 1984 1985 /* Linux NewReno/SACK/FACK/ECN state machine. 1986 * -------------------------------------- 1987 * 1988 * "Open" Normal state, no dubious events, fast path. 1989 * "Disorder" In all the respects it is "Open", 1990 * but requires a bit more attention. It is entered when 1991 * we see some SACKs or dupacks. It is split of "Open" 1992 * mainly to move some processing from fast path to slow one. 1993 * "CWR" CWND was reduced due to some Congestion Notification event. 1994 * It can be ECN, ICMP source quench, local device congestion. 1995 * "Recovery" CWND was reduced, we are fast-retransmitting. 1996 * "Loss" CWND was reduced due to RTO timeout or SACK reneging. 1997 * 1998 * tcp_fastretrans_alert() is entered: 1999 * - each incoming ACK, if state is not "Open" 2000 * - when arrived ACK is unusual, namely: 2001 * * SACK 2002 * * Duplicate ACK. 2003 * * ECN ECE. 2004 * 2005 * Counting packets in flight is pretty simple. 2006 * 2007 * in_flight = packets_out - left_out + retrans_out 2008 * 2009 * packets_out is SND.NXT-SND.UNA counted in packets. 2010 * 2011 * retrans_out is number of retransmitted segments. 2012 * 2013 * left_out is number of segments left network, but not ACKed yet. 2014 * 2015 * left_out = sacked_out + lost_out 2016 * 2017 * sacked_out: Packets, which arrived to receiver out of order 2018 * and hence not ACKed. With SACKs this number is simply 2019 * amount of SACKed data. Even without SACKs 2020 * it is easy to give pretty reliable estimate of this number, 2021 * counting duplicate ACKs. 2022 * 2023 * lost_out: Packets lost by network. TCP has no explicit 2024 * "loss notification" feedback from network (for now). 2025 * It means that this number can be only _guessed_. 2026 * Actually, it is the heuristics to predict lossage that 2027 * distinguishes different algorithms. 2028 * 2029 * F.e. after RTO, when all the queue is considered as lost, 2030 * lost_out = packets_out and in_flight = retrans_out. 2031 * 2032 * Essentially, we have now two algorithms counting 2033 * lost packets. 2034 * 2035 * FACK: It is the simplest heuristics. As soon as we decided 2036 * that something is lost, we decide that _all_ not SACKed 2037 * packets until the most forward SACK are lost. I.e. 2038 * lost_out = fackets_out - sacked_out and left_out = fackets_out. 2039 * It is absolutely correct estimate, if network does not reorder 2040 * packets. And it loses any connection to reality when reordering 2041 * takes place. We use FACK by default until reordering 2042 * is suspected on the path to this destination. 2043 * 2044 * NewReno: when Recovery is entered, we assume that one segment 2045 * is lost (classic Reno). While we are in Recovery and 2046 * a partial ACK arrives, we assume that one more packet 2047 * is lost (NewReno). This heuristics are the same in NewReno 2048 * and SACK. 2049 * 2050 * Imagine, that's all! Forget about all this shamanism about CWND inflation 2051 * deflation etc. CWND is real congestion window, never inflated, changes 2052 * only according to classic VJ rules. 2053 * 2054 * Really tricky (and requiring careful tuning) part of algorithm 2055 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue(). 2056 * The first determines the moment _when_ we should reduce CWND and, 2057 * hence, slow down forward transmission. In fact, it determines the moment 2058 * when we decide that hole is caused by loss, rather than by a reorder. 2059 * 2060 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill 2061 * holes, caused by lost packets. 2062 * 2063 * And the most logically complicated part of algorithm is undo 2064 * heuristics. We detect false retransmits due to both too early 2065 * fast retransmit (reordering) and underestimated RTO, analyzing 2066 * timestamps and D-SACKs. When we detect that some segments were 2067 * retransmitted by mistake and CWND reduction was wrong, we undo 2068 * window reduction and abort recovery phase. This logic is hidden 2069 * inside several functions named tcp_try_undo_<something>. 2070 */ 2071 2072 /* This function decides, when we should leave Disordered state 2073 * and enter Recovery phase, reducing congestion window. 2074 * 2075 * Main question: may we further continue forward transmission 2076 * with the same cwnd? 2077 */ 2078 static int tcp_time_to_recover(struct sock *sk) 2079 { 2080 struct tcp_sock *tp = tcp_sk(sk); 2081 __u32 packets_out; 2082 2083 /* Do not perform any recovery during F-RTO algorithm */ 2084 if (tp->frto_counter) 2085 return 0; 2086 2087 /* Trick#1: The loss is proven. */ 2088 if (tp->lost_out) 2089 return 1; 2090 2091 /* Not-A-Trick#2 : Classic rule... */ 2092 if (tcp_dupack_heurestics(tp) > tp->reordering) 2093 return 1; 2094 2095 /* Trick#3 : when we use RFC2988 timer restart, fast 2096 * retransmit can be triggered by timeout of queue head. 2097 */ 2098 if (tcp_is_fack(tp) && tcp_head_timedout(sk)) 2099 return 1; 2100 2101 /* Trick#4: It is still not OK... But will it be useful to delay 2102 * recovery more? 2103 */ 2104 packets_out = tp->packets_out; 2105 if (packets_out <= tp->reordering && 2106 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) && 2107 !tcp_may_send_now(sk)) { 2108 /* We have nothing to send. This connection is limited 2109 * either by receiver window or by application. 2110 */ 2111 return 1; 2112 } 2113 2114 return 0; 2115 } 2116 2117 /* RFC: This is from the original, I doubt that this is necessary at all: 2118 * clear xmit_retrans hint if seq of this skb is beyond hint. How could we 2119 * retransmitted past LOST markings in the first place? I'm not fully sure 2120 * about undo and end of connection cases, which can cause R without L? 2121 */ 2122 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb) 2123 { 2124 if ((tp->retransmit_skb_hint != NULL) && 2125 before(TCP_SKB_CB(skb)->seq, 2126 TCP_SKB_CB(tp->retransmit_skb_hint)->seq)) 2127 tp->retransmit_skb_hint = NULL; 2128 } 2129 2130 /* Mark head of queue up as lost. With RFC3517 SACK, the packets is 2131 * is against sacked "cnt", otherwise it's against facked "cnt" 2132 */ 2133 static void tcp_mark_head_lost(struct sock *sk, int packets, int fast_rexmit) 2134 { 2135 struct tcp_sock *tp = tcp_sk(sk); 2136 struct sk_buff *skb; 2137 int cnt; 2138 2139 BUG_TRAP(packets <= tp->packets_out); 2140 if (tp->lost_skb_hint) { 2141 skb = tp->lost_skb_hint; 2142 cnt = tp->lost_cnt_hint; 2143 } else { 2144 skb = tcp_write_queue_head(sk); 2145 cnt = 0; 2146 } 2147 2148 tcp_for_write_queue_from(skb, sk) { 2149 if (skb == tcp_send_head(sk)) 2150 break; 2151 /* TODO: do this better */ 2152 /* this is not the most efficient way to do this... */ 2153 tp->lost_skb_hint = skb; 2154 tp->lost_cnt_hint = cnt; 2155 2156 if (tcp_is_fack(tp) || tcp_is_reno(tp) || 2157 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) 2158 cnt += tcp_skb_pcount(skb); 2159 2160 if (((!fast_rexmit || (tp->lost_out > 0)) && (cnt > packets)) || 2161 after(TCP_SKB_CB(skb)->end_seq, tp->high_seq)) 2162 break; 2163 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_SACKED_ACKED|TCPCB_LOST))) { 2164 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 2165 tp->lost_out += tcp_skb_pcount(skb); 2166 tcp_verify_retransmit_hint(tp, skb); 2167 } 2168 } 2169 tcp_verify_left_out(tp); 2170 } 2171 2172 /* Account newly detected lost packet(s) */ 2173 2174 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit) 2175 { 2176 struct tcp_sock *tp = tcp_sk(sk); 2177 2178 if (tcp_is_reno(tp)) { 2179 tcp_mark_head_lost(sk, 1, fast_rexmit); 2180 } else if (tcp_is_fack(tp)) { 2181 int lost = tp->fackets_out - tp->reordering; 2182 if (lost <= 0) 2183 lost = 1; 2184 tcp_mark_head_lost(sk, lost, fast_rexmit); 2185 } else { 2186 int sacked_upto = tp->sacked_out - tp->reordering; 2187 if (sacked_upto < 0) 2188 sacked_upto = 0; 2189 tcp_mark_head_lost(sk, sacked_upto, fast_rexmit); 2190 } 2191 2192 /* New heuristics: it is possible only after we switched 2193 * to restart timer each time when something is ACKed. 2194 * Hence, we can detect timed out packets during fast 2195 * retransmit without falling to slow start. 2196 */ 2197 if (tcp_is_fack(tp) && tcp_head_timedout(sk)) { 2198 struct sk_buff *skb; 2199 2200 skb = tp->scoreboard_skb_hint ? tp->scoreboard_skb_hint 2201 : tcp_write_queue_head(sk); 2202 2203 tcp_for_write_queue_from(skb, sk) { 2204 if (skb == tcp_send_head(sk)) 2205 break; 2206 if (!tcp_skb_timedout(sk, skb)) 2207 break; 2208 2209 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_SACKED_ACKED|TCPCB_LOST))) { 2210 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 2211 tp->lost_out += tcp_skb_pcount(skb); 2212 tcp_verify_retransmit_hint(tp, skb); 2213 } 2214 } 2215 2216 tp->scoreboard_skb_hint = skb; 2217 2218 tcp_verify_left_out(tp); 2219 } 2220 } 2221 2222 /* CWND moderation, preventing bursts due to too big ACKs 2223 * in dubious situations. 2224 */ 2225 static inline void tcp_moderate_cwnd(struct tcp_sock *tp) 2226 { 2227 tp->snd_cwnd = min(tp->snd_cwnd, 2228 tcp_packets_in_flight(tp) + tcp_max_burst(tp)); 2229 tp->snd_cwnd_stamp = tcp_time_stamp; 2230 } 2231 2232 /* Lower bound on congestion window is slow start threshold 2233 * unless congestion avoidance choice decides to overide it. 2234 */ 2235 static inline u32 tcp_cwnd_min(const struct sock *sk) 2236 { 2237 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 2238 2239 return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh; 2240 } 2241 2242 /* Decrease cwnd each second ack. */ 2243 static void tcp_cwnd_down(struct sock *sk, int flag) 2244 { 2245 struct tcp_sock *tp = tcp_sk(sk); 2246 int decr = tp->snd_cwnd_cnt + 1; 2247 2248 if ((flag & (FLAG_ANY_PROGRESS | FLAG_DSACKING_ACK)) || 2249 (tcp_is_reno(tp) && !(flag & FLAG_NOT_DUP))) { 2250 tp->snd_cwnd_cnt = decr & 1; 2251 decr >>= 1; 2252 2253 if (decr && tp->snd_cwnd > tcp_cwnd_min(sk)) 2254 tp->snd_cwnd -= decr; 2255 2256 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1); 2257 tp->snd_cwnd_stamp = tcp_time_stamp; 2258 } 2259 } 2260 2261 /* Nothing was retransmitted or returned timestamp is less 2262 * than timestamp of the first retransmission. 2263 */ 2264 static inline int tcp_packet_delayed(struct tcp_sock *tp) 2265 { 2266 return !tp->retrans_stamp || 2267 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 2268 (__s32)(tp->rx_opt.rcv_tsecr - tp->retrans_stamp) < 0); 2269 } 2270 2271 /* Undo procedures. */ 2272 2273 #if FASTRETRANS_DEBUG > 1 2274 static void DBGUNDO(struct sock *sk, const char *msg) 2275 { 2276 struct tcp_sock *tp = tcp_sk(sk); 2277 struct inet_sock *inet = inet_sk(sk); 2278 2279 printk(KERN_DEBUG "Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n", 2280 msg, 2281 NIPQUAD(inet->daddr), ntohs(inet->dport), 2282 tp->snd_cwnd, tcp_left_out(tp), 2283 tp->snd_ssthresh, tp->prior_ssthresh, 2284 tp->packets_out); 2285 } 2286 #else 2287 #define DBGUNDO(x...) do { } while (0) 2288 #endif 2289 2290 static void tcp_undo_cwr(struct sock *sk, const int undo) 2291 { 2292 struct tcp_sock *tp = tcp_sk(sk); 2293 2294 if (tp->prior_ssthresh) { 2295 const struct inet_connection_sock *icsk = inet_csk(sk); 2296 2297 if (icsk->icsk_ca_ops->undo_cwnd) 2298 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk); 2299 else 2300 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1); 2301 2302 if (undo && tp->prior_ssthresh > tp->snd_ssthresh) { 2303 tp->snd_ssthresh = tp->prior_ssthresh; 2304 TCP_ECN_withdraw_cwr(tp); 2305 } 2306 } else { 2307 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh); 2308 } 2309 tcp_moderate_cwnd(tp); 2310 tp->snd_cwnd_stamp = tcp_time_stamp; 2311 2312 /* There is something screwy going on with the retrans hints after 2313 an undo */ 2314 tcp_clear_all_retrans_hints(tp); 2315 } 2316 2317 static inline int tcp_may_undo(struct tcp_sock *tp) 2318 { 2319 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp)); 2320 } 2321 2322 /* People celebrate: "We love our President!" */ 2323 static int tcp_try_undo_recovery(struct sock *sk) 2324 { 2325 struct tcp_sock *tp = tcp_sk(sk); 2326 2327 if (tcp_may_undo(tp)) { 2328 /* Happy end! We did not retransmit anything 2329 * or our original transmission succeeded. 2330 */ 2331 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans"); 2332 tcp_undo_cwr(sk, 1); 2333 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss) 2334 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO); 2335 else 2336 NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO); 2337 tp->undo_marker = 0; 2338 } 2339 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) { 2340 /* Hold old state until something *above* high_seq 2341 * is ACKed. For Reno it is MUST to prevent false 2342 * fast retransmits (RFC2582). SACK TCP is safe. */ 2343 tcp_moderate_cwnd(tp); 2344 return 1; 2345 } 2346 tcp_set_ca_state(sk, TCP_CA_Open); 2347 return 0; 2348 } 2349 2350 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */ 2351 static void tcp_try_undo_dsack(struct sock *sk) 2352 { 2353 struct tcp_sock *tp = tcp_sk(sk); 2354 2355 if (tp->undo_marker && !tp->undo_retrans) { 2356 DBGUNDO(sk, "D-SACK"); 2357 tcp_undo_cwr(sk, 1); 2358 tp->undo_marker = 0; 2359 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO); 2360 } 2361 } 2362 2363 /* Undo during fast recovery after partial ACK. */ 2364 2365 static int tcp_try_undo_partial(struct sock *sk, int acked) 2366 { 2367 struct tcp_sock *tp = tcp_sk(sk); 2368 /* Partial ACK arrived. Force Hoe's retransmit. */ 2369 int failed = tcp_is_reno(tp) || (tcp_fackets_out(tp) > tp->reordering); 2370 2371 if (tcp_may_undo(tp)) { 2372 /* Plain luck! Hole if filled with delayed 2373 * packet, rather than with a retransmit. 2374 */ 2375 if (tp->retrans_out == 0) 2376 tp->retrans_stamp = 0; 2377 2378 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1); 2379 2380 DBGUNDO(sk, "Hoe"); 2381 tcp_undo_cwr(sk, 0); 2382 NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO); 2383 2384 /* So... Do not make Hoe's retransmit yet. 2385 * If the first packet was delayed, the rest 2386 * ones are most probably delayed as well. 2387 */ 2388 failed = 0; 2389 } 2390 return failed; 2391 } 2392 2393 /* Undo during loss recovery after partial ACK. */ 2394 static int tcp_try_undo_loss(struct sock *sk) 2395 { 2396 struct tcp_sock *tp = tcp_sk(sk); 2397 2398 if (tcp_may_undo(tp)) { 2399 struct sk_buff *skb; 2400 tcp_for_write_queue(skb, sk) { 2401 if (skb == tcp_send_head(sk)) 2402 break; 2403 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; 2404 } 2405 2406 tcp_clear_all_retrans_hints(tp); 2407 2408 DBGUNDO(sk, "partial loss"); 2409 tp->lost_out = 0; 2410 tcp_undo_cwr(sk, 1); 2411 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO); 2412 inet_csk(sk)->icsk_retransmits = 0; 2413 tp->undo_marker = 0; 2414 if (tcp_is_sack(tp)) 2415 tcp_set_ca_state(sk, TCP_CA_Open); 2416 return 1; 2417 } 2418 return 0; 2419 } 2420 2421 static inline void tcp_complete_cwr(struct sock *sk) 2422 { 2423 struct tcp_sock *tp = tcp_sk(sk); 2424 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh); 2425 tp->snd_cwnd_stamp = tcp_time_stamp; 2426 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR); 2427 } 2428 2429 static void tcp_try_to_open(struct sock *sk, int flag) 2430 { 2431 struct tcp_sock *tp = tcp_sk(sk); 2432 2433 tcp_verify_left_out(tp); 2434 2435 if (tp->retrans_out == 0) 2436 tp->retrans_stamp = 0; 2437 2438 if (flag & FLAG_ECE) 2439 tcp_enter_cwr(sk, 1); 2440 2441 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) { 2442 int state = TCP_CA_Open; 2443 2444 if (tcp_left_out(tp) || tp->retrans_out || tp->undo_marker) 2445 state = TCP_CA_Disorder; 2446 2447 if (inet_csk(sk)->icsk_ca_state != state) { 2448 tcp_set_ca_state(sk, state); 2449 tp->high_seq = tp->snd_nxt; 2450 } 2451 tcp_moderate_cwnd(tp); 2452 } else { 2453 tcp_cwnd_down(sk, flag); 2454 } 2455 } 2456 2457 static void tcp_mtup_probe_failed(struct sock *sk) 2458 { 2459 struct inet_connection_sock *icsk = inet_csk(sk); 2460 2461 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1; 2462 icsk->icsk_mtup.probe_size = 0; 2463 } 2464 2465 static void tcp_mtup_probe_success(struct sock *sk, struct sk_buff *skb) 2466 { 2467 struct tcp_sock *tp = tcp_sk(sk); 2468 struct inet_connection_sock *icsk = inet_csk(sk); 2469 2470 /* FIXME: breaks with very large cwnd */ 2471 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2472 tp->snd_cwnd = tp->snd_cwnd * 2473 tcp_mss_to_mtu(sk, tp->mss_cache) / 2474 icsk->icsk_mtup.probe_size; 2475 tp->snd_cwnd_cnt = 0; 2476 tp->snd_cwnd_stamp = tcp_time_stamp; 2477 tp->rcv_ssthresh = tcp_current_ssthresh(sk); 2478 2479 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size; 2480 icsk->icsk_mtup.probe_size = 0; 2481 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 2482 } 2483 2484 /* Process an event, which can update packets-in-flight not trivially. 2485 * Main goal of this function is to calculate new estimate for left_out, 2486 * taking into account both packets sitting in receiver's buffer and 2487 * packets lost by network. 2488 * 2489 * Besides that it does CWND reduction, when packet loss is detected 2490 * and changes state of machine. 2491 * 2492 * It does _not_ decide what to send, it is made in function 2493 * tcp_xmit_retransmit_queue(). 2494 */ 2495 static void tcp_fastretrans_alert(struct sock *sk, int pkts_acked, int flag) 2496 { 2497 struct inet_connection_sock *icsk = inet_csk(sk); 2498 struct tcp_sock *tp = tcp_sk(sk); 2499 int is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP)); 2500 int do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) && 2501 (tcp_fackets_out(tp) > tp->reordering)); 2502 int fast_rexmit = 0; 2503 2504 if (WARN_ON(!tp->packets_out && tp->sacked_out)) 2505 tp->sacked_out = 0; 2506 if (WARN_ON(!tp->sacked_out && tp->fackets_out)) 2507 tp->fackets_out = 0; 2508 2509 /* Now state machine starts. 2510 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */ 2511 if (flag & FLAG_ECE) 2512 tp->prior_ssthresh = 0; 2513 2514 /* B. In all the states check for reneging SACKs. */ 2515 if (tcp_check_sack_reneging(sk, flag)) 2516 return; 2517 2518 /* C. Process data loss notification, provided it is valid. */ 2519 if (tcp_is_fack(tp) && (flag & FLAG_DATA_LOST) && 2520 before(tp->snd_una, tp->high_seq) && 2521 icsk->icsk_ca_state != TCP_CA_Open && 2522 tp->fackets_out > tp->reordering) { 2523 tcp_mark_head_lost(sk, tp->fackets_out - tp->reordering, 0); 2524 NET_INC_STATS_BH(LINUX_MIB_TCPLOSS); 2525 } 2526 2527 /* D. Check consistency of the current state. */ 2528 tcp_verify_left_out(tp); 2529 2530 /* E. Check state exit conditions. State can be terminated 2531 * when high_seq is ACKed. */ 2532 if (icsk->icsk_ca_state == TCP_CA_Open) { 2533 BUG_TRAP(tp->retrans_out == 0); 2534 tp->retrans_stamp = 0; 2535 } else if (!before(tp->snd_una, tp->high_seq)) { 2536 switch (icsk->icsk_ca_state) { 2537 case TCP_CA_Loss: 2538 icsk->icsk_retransmits = 0; 2539 if (tcp_try_undo_recovery(sk)) 2540 return; 2541 break; 2542 2543 case TCP_CA_CWR: 2544 /* CWR is to be held something *above* high_seq 2545 * is ACKed for CWR bit to reach receiver. */ 2546 if (tp->snd_una != tp->high_seq) { 2547 tcp_complete_cwr(sk); 2548 tcp_set_ca_state(sk, TCP_CA_Open); 2549 } 2550 break; 2551 2552 case TCP_CA_Disorder: 2553 tcp_try_undo_dsack(sk); 2554 if (!tp->undo_marker || 2555 /* For SACK case do not Open to allow to undo 2556 * catching for all duplicate ACKs. */ 2557 tcp_is_reno(tp) || tp->snd_una != tp->high_seq) { 2558 tp->undo_marker = 0; 2559 tcp_set_ca_state(sk, TCP_CA_Open); 2560 } 2561 break; 2562 2563 case TCP_CA_Recovery: 2564 if (tcp_is_reno(tp)) 2565 tcp_reset_reno_sack(tp); 2566 if (tcp_try_undo_recovery(sk)) 2567 return; 2568 tcp_complete_cwr(sk); 2569 break; 2570 } 2571 } 2572 2573 /* F. Process state. */ 2574 switch (icsk->icsk_ca_state) { 2575 case TCP_CA_Recovery: 2576 if (!(flag & FLAG_SND_UNA_ADVANCED)) { 2577 if (tcp_is_reno(tp) && is_dupack) 2578 tcp_add_reno_sack(sk); 2579 } else 2580 do_lost = tcp_try_undo_partial(sk, pkts_acked); 2581 break; 2582 case TCP_CA_Loss: 2583 if (flag & FLAG_DATA_ACKED) 2584 icsk->icsk_retransmits = 0; 2585 if (!tcp_try_undo_loss(sk)) { 2586 tcp_moderate_cwnd(tp); 2587 tcp_xmit_retransmit_queue(sk); 2588 return; 2589 } 2590 if (icsk->icsk_ca_state != TCP_CA_Open) 2591 return; 2592 /* Loss is undone; fall through to processing in Open state. */ 2593 default: 2594 if (tcp_is_reno(tp)) { 2595 if (flag & FLAG_SND_UNA_ADVANCED) 2596 tcp_reset_reno_sack(tp); 2597 if (is_dupack) 2598 tcp_add_reno_sack(sk); 2599 } 2600 2601 if (icsk->icsk_ca_state == TCP_CA_Disorder) 2602 tcp_try_undo_dsack(sk); 2603 2604 if (!tcp_time_to_recover(sk)) { 2605 tcp_try_to_open(sk, flag); 2606 return; 2607 } 2608 2609 /* MTU probe failure: don't reduce cwnd */ 2610 if (icsk->icsk_ca_state < TCP_CA_CWR && 2611 icsk->icsk_mtup.probe_size && 2612 tp->snd_una == tp->mtu_probe.probe_seq_start) { 2613 tcp_mtup_probe_failed(sk); 2614 /* Restores the reduction we did in tcp_mtup_probe() */ 2615 tp->snd_cwnd++; 2616 tcp_simple_retransmit(sk); 2617 return; 2618 } 2619 2620 /* Otherwise enter Recovery state */ 2621 2622 if (tcp_is_reno(tp)) 2623 NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY); 2624 else 2625 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY); 2626 2627 tp->high_seq = tp->snd_nxt; 2628 tp->prior_ssthresh = 0; 2629 tp->undo_marker = tp->snd_una; 2630 tp->undo_retrans = tp->retrans_out; 2631 2632 if (icsk->icsk_ca_state < TCP_CA_CWR) { 2633 if (!(flag & FLAG_ECE)) 2634 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2635 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 2636 TCP_ECN_queue_cwr(tp); 2637 } 2638 2639 tp->bytes_acked = 0; 2640 tp->snd_cwnd_cnt = 0; 2641 tcp_set_ca_state(sk, TCP_CA_Recovery); 2642 fast_rexmit = 1; 2643 } 2644 2645 if (do_lost || (tcp_is_fack(tp) && tcp_head_timedout(sk))) 2646 tcp_update_scoreboard(sk, fast_rexmit); 2647 tcp_cwnd_down(sk, flag); 2648 tcp_xmit_retransmit_queue(sk); 2649 } 2650 2651 /* Read draft-ietf-tcplw-high-performance before mucking 2652 * with this code. (Supersedes RFC1323) 2653 */ 2654 static void tcp_ack_saw_tstamp(struct sock *sk, int flag) 2655 { 2656 /* RTTM Rule: A TSecr value received in a segment is used to 2657 * update the averaged RTT measurement only if the segment 2658 * acknowledges some new data, i.e., only if it advances the 2659 * left edge of the send window. 2660 * 2661 * See draft-ietf-tcplw-high-performance-00, section 3.3. 2662 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru> 2663 * 2664 * Changed: reset backoff as soon as we see the first valid sample. 2665 * If we do not, we get strongly overestimated rto. With timestamps 2666 * samples are accepted even from very old segments: f.e., when rtt=1 2667 * increases to 8, we retransmit 5 times and after 8 seconds delayed 2668 * answer arrives rto becomes 120 seconds! If at least one of segments 2669 * in window is lost... Voila. --ANK (010210) 2670 */ 2671 struct tcp_sock *tp = tcp_sk(sk); 2672 const __u32 seq_rtt = tcp_time_stamp - tp->rx_opt.rcv_tsecr; 2673 tcp_rtt_estimator(sk, seq_rtt); 2674 tcp_set_rto(sk); 2675 inet_csk(sk)->icsk_backoff = 0; 2676 tcp_bound_rto(sk); 2677 } 2678 2679 static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag) 2680 { 2681 /* We don't have a timestamp. Can only use 2682 * packets that are not retransmitted to determine 2683 * rtt estimates. Also, we must not reset the 2684 * backoff for rto until we get a non-retransmitted 2685 * packet. This allows us to deal with a situation 2686 * where the network delay has increased suddenly. 2687 * I.e. Karn's algorithm. (SIGCOMM '87, p5.) 2688 */ 2689 2690 if (flag & FLAG_RETRANS_DATA_ACKED) 2691 return; 2692 2693 tcp_rtt_estimator(sk, seq_rtt); 2694 tcp_set_rto(sk); 2695 inet_csk(sk)->icsk_backoff = 0; 2696 tcp_bound_rto(sk); 2697 } 2698 2699 static inline void tcp_ack_update_rtt(struct sock *sk, const int flag, 2700 const s32 seq_rtt) 2701 { 2702 const struct tcp_sock *tp = tcp_sk(sk); 2703 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */ 2704 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr) 2705 tcp_ack_saw_tstamp(sk, flag); 2706 else if (seq_rtt >= 0) 2707 tcp_ack_no_tstamp(sk, seq_rtt, flag); 2708 } 2709 2710 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight) 2711 { 2712 const struct inet_connection_sock *icsk = inet_csk(sk); 2713 icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight); 2714 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp; 2715 } 2716 2717 /* Restart timer after forward progress on connection. 2718 * RFC2988 recommends to restart timer to now+rto. 2719 */ 2720 static void tcp_rearm_rto(struct sock *sk) 2721 { 2722 struct tcp_sock *tp = tcp_sk(sk); 2723 2724 if (!tp->packets_out) { 2725 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS); 2726 } else { 2727 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 2728 inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 2729 } 2730 } 2731 2732 /* If we get here, the whole TSO packet has not been acked. */ 2733 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb) 2734 { 2735 struct tcp_sock *tp = tcp_sk(sk); 2736 u32 packets_acked; 2737 2738 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)); 2739 2740 packets_acked = tcp_skb_pcount(skb); 2741 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 2742 return 0; 2743 packets_acked -= tcp_skb_pcount(skb); 2744 2745 if (packets_acked) { 2746 BUG_ON(tcp_skb_pcount(skb) == 0); 2747 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)); 2748 } 2749 2750 return packets_acked; 2751 } 2752 2753 /* Remove acknowledged frames from the retransmission queue. If our packet 2754 * is before the ack sequence we can discard it as it's confirmed to have 2755 * arrived at the other end. 2756 */ 2757 static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets) 2758 { 2759 struct tcp_sock *tp = tcp_sk(sk); 2760 const struct inet_connection_sock *icsk = inet_csk(sk); 2761 struct sk_buff *skb; 2762 u32 now = tcp_time_stamp; 2763 int fully_acked = 1; 2764 int flag = 0; 2765 u32 pkts_acked = 0; 2766 u32 reord = tp->packets_out; 2767 s32 seq_rtt = -1; 2768 s32 ca_seq_rtt = -1; 2769 ktime_t last_ackt = net_invalid_timestamp(); 2770 2771 while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) { 2772 struct tcp_skb_cb *scb = TCP_SKB_CB(skb); 2773 u32 end_seq; 2774 u32 acked_pcount; 2775 u8 sacked = scb->sacked; 2776 2777 /* Determine how many packets and what bytes were acked, tso and else */ 2778 if (after(scb->end_seq, tp->snd_una)) { 2779 if (tcp_skb_pcount(skb) == 1 || 2780 !after(tp->snd_una, scb->seq)) 2781 break; 2782 2783 acked_pcount = tcp_tso_acked(sk, skb); 2784 if (!acked_pcount) 2785 break; 2786 2787 fully_acked = 0; 2788 end_seq = tp->snd_una; 2789 } else { 2790 acked_pcount = tcp_skb_pcount(skb); 2791 end_seq = scb->end_seq; 2792 } 2793 2794 /* MTU probing checks */ 2795 if (fully_acked && icsk->icsk_mtup.probe_size && 2796 !after(tp->mtu_probe.probe_seq_end, scb->end_seq)) { 2797 tcp_mtup_probe_success(sk, skb); 2798 } 2799 2800 if (sacked & TCPCB_RETRANS) { 2801 if (sacked & TCPCB_SACKED_RETRANS) 2802 tp->retrans_out -= acked_pcount; 2803 flag |= FLAG_RETRANS_DATA_ACKED; 2804 ca_seq_rtt = -1; 2805 seq_rtt = -1; 2806 if ((flag & FLAG_DATA_ACKED) || (acked_pcount > 1)) 2807 flag |= FLAG_NONHEAD_RETRANS_ACKED; 2808 } else { 2809 ca_seq_rtt = now - scb->when; 2810 last_ackt = skb->tstamp; 2811 if (seq_rtt < 0) { 2812 seq_rtt = ca_seq_rtt; 2813 } 2814 if (!(sacked & TCPCB_SACKED_ACKED)) 2815 reord = min(pkts_acked, reord); 2816 } 2817 2818 if (sacked & TCPCB_SACKED_ACKED) 2819 tp->sacked_out -= acked_pcount; 2820 if (sacked & TCPCB_LOST) 2821 tp->lost_out -= acked_pcount; 2822 2823 if (unlikely(tp->urg_mode && !before(end_seq, tp->snd_up))) 2824 tp->urg_mode = 0; 2825 2826 tp->packets_out -= acked_pcount; 2827 pkts_acked += acked_pcount; 2828 2829 /* Initial outgoing SYN's get put onto the write_queue 2830 * just like anything else we transmit. It is not 2831 * true data, and if we misinform our callers that 2832 * this ACK acks real data, we will erroneously exit 2833 * connection startup slow start one packet too 2834 * quickly. This is severely frowned upon behavior. 2835 */ 2836 if (!(scb->flags & TCPCB_FLAG_SYN)) { 2837 flag |= FLAG_DATA_ACKED; 2838 } else { 2839 flag |= FLAG_SYN_ACKED; 2840 tp->retrans_stamp = 0; 2841 } 2842 2843 if (!fully_acked) 2844 break; 2845 2846 tcp_unlink_write_queue(skb, sk); 2847 sk_wmem_free_skb(sk, skb); 2848 tcp_clear_all_retrans_hints(tp); 2849 } 2850 2851 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) 2852 flag |= FLAG_SACK_RENEGING; 2853 2854 if (flag & FLAG_ACKED) { 2855 const struct tcp_congestion_ops *ca_ops 2856 = inet_csk(sk)->icsk_ca_ops; 2857 2858 tcp_ack_update_rtt(sk, flag, seq_rtt); 2859 tcp_rearm_rto(sk); 2860 2861 if (tcp_is_reno(tp)) { 2862 tcp_remove_reno_sacks(sk, pkts_acked); 2863 } else { 2864 /* Non-retransmitted hole got filled? That's reordering */ 2865 if (reord < prior_fackets) 2866 tcp_update_reordering(sk, tp->fackets_out - reord, 0); 2867 } 2868 2869 tp->fackets_out -= min(pkts_acked, tp->fackets_out); 2870 2871 if (ca_ops->pkts_acked) { 2872 s32 rtt_us = -1; 2873 2874 /* Is the ACK triggering packet unambiguous? */ 2875 if (!(flag & FLAG_RETRANS_DATA_ACKED)) { 2876 /* High resolution needed and available? */ 2877 if (ca_ops->flags & TCP_CONG_RTT_STAMP && 2878 !ktime_equal(last_ackt, 2879 net_invalid_timestamp())) 2880 rtt_us = ktime_us_delta(ktime_get_real(), 2881 last_ackt); 2882 else if (ca_seq_rtt > 0) 2883 rtt_us = jiffies_to_usecs(ca_seq_rtt); 2884 } 2885 2886 ca_ops->pkts_acked(sk, pkts_acked, rtt_us); 2887 } 2888 } 2889 2890 #if FASTRETRANS_DEBUG > 0 2891 BUG_TRAP((int)tp->sacked_out >= 0); 2892 BUG_TRAP((int)tp->lost_out >= 0); 2893 BUG_TRAP((int)tp->retrans_out >= 0); 2894 if (!tp->packets_out && tcp_is_sack(tp)) { 2895 icsk = inet_csk(sk); 2896 if (tp->lost_out) { 2897 printk(KERN_DEBUG "Leak l=%u %d\n", 2898 tp->lost_out, icsk->icsk_ca_state); 2899 tp->lost_out = 0; 2900 } 2901 if (tp->sacked_out) { 2902 printk(KERN_DEBUG "Leak s=%u %d\n", 2903 tp->sacked_out, icsk->icsk_ca_state); 2904 tp->sacked_out = 0; 2905 } 2906 if (tp->retrans_out) { 2907 printk(KERN_DEBUG "Leak r=%u %d\n", 2908 tp->retrans_out, icsk->icsk_ca_state); 2909 tp->retrans_out = 0; 2910 } 2911 } 2912 #endif 2913 return flag; 2914 } 2915 2916 static void tcp_ack_probe(struct sock *sk) 2917 { 2918 const struct tcp_sock *tp = tcp_sk(sk); 2919 struct inet_connection_sock *icsk = inet_csk(sk); 2920 2921 /* Was it a usable window open? */ 2922 2923 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) { 2924 icsk->icsk_backoff = 0; 2925 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0); 2926 /* Socket must be waked up by subsequent tcp_data_snd_check(). 2927 * This function is not for random using! 2928 */ 2929 } else { 2930 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 2931 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), 2932 TCP_RTO_MAX); 2933 } 2934 } 2935 2936 static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag) 2937 { 2938 return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) || 2939 inet_csk(sk)->icsk_ca_state != TCP_CA_Open); 2940 } 2941 2942 static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag) 2943 { 2944 const struct tcp_sock *tp = tcp_sk(sk); 2945 return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) && 2946 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR)); 2947 } 2948 2949 /* Check that window update is acceptable. 2950 * The function assumes that snd_una<=ack<=snd_next. 2951 */ 2952 static inline int tcp_may_update_window(const struct tcp_sock *tp, 2953 const u32 ack, const u32 ack_seq, 2954 const u32 nwin) 2955 { 2956 return (after(ack, tp->snd_una) || 2957 after(ack_seq, tp->snd_wl1) || 2958 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd)); 2959 } 2960 2961 /* Update our send window. 2962 * 2963 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2 2964 * and in FreeBSD. NetBSD's one is even worse.) is wrong. 2965 */ 2966 static int tcp_ack_update_window(struct sock *sk, struct sk_buff *skb, u32 ack, 2967 u32 ack_seq) 2968 { 2969 struct tcp_sock *tp = tcp_sk(sk); 2970 int flag = 0; 2971 u32 nwin = ntohs(tcp_hdr(skb)->window); 2972 2973 if (likely(!tcp_hdr(skb)->syn)) 2974 nwin <<= tp->rx_opt.snd_wscale; 2975 2976 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) { 2977 flag |= FLAG_WIN_UPDATE; 2978 tcp_update_wl(tp, ack, ack_seq); 2979 2980 if (tp->snd_wnd != nwin) { 2981 tp->snd_wnd = nwin; 2982 2983 /* Note, it is the only place, where 2984 * fast path is recovered for sending TCP. 2985 */ 2986 tp->pred_flags = 0; 2987 tcp_fast_path_check(sk); 2988 2989 if (nwin > tp->max_window) { 2990 tp->max_window = nwin; 2991 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie); 2992 } 2993 } 2994 } 2995 2996 tp->snd_una = ack; 2997 2998 return flag; 2999 } 3000 3001 /* A very conservative spurious RTO response algorithm: reduce cwnd and 3002 * continue in congestion avoidance. 3003 */ 3004 static void tcp_conservative_spur_to_response(struct tcp_sock *tp) 3005 { 3006 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh); 3007 tp->snd_cwnd_cnt = 0; 3008 tp->bytes_acked = 0; 3009 TCP_ECN_queue_cwr(tp); 3010 tcp_moderate_cwnd(tp); 3011 } 3012 3013 /* A conservative spurious RTO response algorithm: reduce cwnd using 3014 * rate halving and continue in congestion avoidance. 3015 */ 3016 static void tcp_ratehalving_spur_to_response(struct sock *sk) 3017 { 3018 tcp_enter_cwr(sk, 0); 3019 } 3020 3021 static void tcp_undo_spur_to_response(struct sock *sk, int flag) 3022 { 3023 if (flag & FLAG_ECE) 3024 tcp_ratehalving_spur_to_response(sk); 3025 else 3026 tcp_undo_cwr(sk, 1); 3027 } 3028 3029 /* F-RTO spurious RTO detection algorithm (RFC4138) 3030 * 3031 * F-RTO affects during two new ACKs following RTO (well, almost, see inline 3032 * comments). State (ACK number) is kept in frto_counter. When ACK advances 3033 * window (but not to or beyond highest sequence sent before RTO): 3034 * On First ACK, send two new segments out. 3035 * On Second ACK, RTO was likely spurious. Do spurious response (response 3036 * algorithm is not part of the F-RTO detection algorithm 3037 * given in RFC4138 but can be selected separately). 3038 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss 3039 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding 3040 * of Nagle, this is done using frto_counter states 2 and 3, when a new data 3041 * segment of any size sent during F-RTO, state 2 is upgraded to 3. 3042 * 3043 * Rationale: if the RTO was spurious, new ACKs should arrive from the 3044 * original window even after we transmit two new data segments. 3045 * 3046 * SACK version: 3047 * on first step, wait until first cumulative ACK arrives, then move to 3048 * the second step. In second step, the next ACK decides. 3049 * 3050 * F-RTO is implemented (mainly) in four functions: 3051 * - tcp_use_frto() is used to determine if TCP is can use F-RTO 3052 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is 3053 * called when tcp_use_frto() showed green light 3054 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm 3055 * - tcp_enter_frto_loss() is called if there is not enough evidence 3056 * to prove that the RTO is indeed spurious. It transfers the control 3057 * from F-RTO to the conventional RTO recovery 3058 */ 3059 static int tcp_process_frto(struct sock *sk, int flag) 3060 { 3061 struct tcp_sock *tp = tcp_sk(sk); 3062 3063 tcp_verify_left_out(tp); 3064 3065 /* Duplicate the behavior from Loss state (fastretrans_alert) */ 3066 if (flag & FLAG_DATA_ACKED) 3067 inet_csk(sk)->icsk_retransmits = 0; 3068 3069 if ((flag & FLAG_NONHEAD_RETRANS_ACKED) || 3070 ((tp->frto_counter >= 2) && (flag & FLAG_RETRANS_DATA_ACKED))) 3071 tp->undo_marker = 0; 3072 3073 if (!before(tp->snd_una, tp->frto_highmark)) { 3074 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag); 3075 return 1; 3076 } 3077 3078 if (!IsSackFrto() || tcp_is_reno(tp)) { 3079 /* RFC4138 shortcoming in step 2; should also have case c): 3080 * ACK isn't duplicate nor advances window, e.g., opposite dir 3081 * data, winupdate 3082 */ 3083 if (!(flag & FLAG_ANY_PROGRESS) && (flag & FLAG_NOT_DUP)) 3084 return 1; 3085 3086 if (!(flag & FLAG_DATA_ACKED)) { 3087 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3), 3088 flag); 3089 return 1; 3090 } 3091 } else { 3092 if (!(flag & FLAG_DATA_ACKED) && (tp->frto_counter == 1)) { 3093 /* Prevent sending of new data. */ 3094 tp->snd_cwnd = min(tp->snd_cwnd, 3095 tcp_packets_in_flight(tp)); 3096 return 1; 3097 } 3098 3099 if ((tp->frto_counter >= 2) && 3100 (!(flag & FLAG_FORWARD_PROGRESS) || 3101 ((flag & FLAG_DATA_SACKED) && 3102 !(flag & FLAG_ONLY_ORIG_SACKED)))) { 3103 /* RFC4138 shortcoming (see comment above) */ 3104 if (!(flag & FLAG_FORWARD_PROGRESS) && 3105 (flag & FLAG_NOT_DUP)) 3106 return 1; 3107 3108 tcp_enter_frto_loss(sk, 3, flag); 3109 return 1; 3110 } 3111 } 3112 3113 if (tp->frto_counter == 1) { 3114 /* tcp_may_send_now needs to see updated state */ 3115 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2; 3116 tp->frto_counter = 2; 3117 3118 if (!tcp_may_send_now(sk)) 3119 tcp_enter_frto_loss(sk, 2, flag); 3120 3121 return 1; 3122 } else { 3123 switch (sysctl_tcp_frto_response) { 3124 case 2: 3125 tcp_undo_spur_to_response(sk, flag); 3126 break; 3127 case 1: 3128 tcp_conservative_spur_to_response(tp); 3129 break; 3130 default: 3131 tcp_ratehalving_spur_to_response(sk); 3132 break; 3133 } 3134 tp->frto_counter = 0; 3135 tp->undo_marker = 0; 3136 NET_INC_STATS_BH(LINUX_MIB_TCPSPURIOUSRTOS); 3137 } 3138 return 0; 3139 } 3140 3141 /* This routine deals with incoming acks, but not outgoing ones. */ 3142 static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag) 3143 { 3144 struct inet_connection_sock *icsk = inet_csk(sk); 3145 struct tcp_sock *tp = tcp_sk(sk); 3146 u32 prior_snd_una = tp->snd_una; 3147 u32 ack_seq = TCP_SKB_CB(skb)->seq; 3148 u32 ack = TCP_SKB_CB(skb)->ack_seq; 3149 u32 prior_in_flight; 3150 u32 prior_fackets; 3151 int prior_packets; 3152 int frto_cwnd = 0; 3153 3154 /* If the ack is newer than sent or older than previous acks 3155 * then we can probably ignore it. 3156 */ 3157 if (after(ack, tp->snd_nxt)) 3158 goto uninteresting_ack; 3159 3160 if (before(ack, prior_snd_una)) 3161 goto old_ack; 3162 3163 if (after(ack, prior_snd_una)) 3164 flag |= FLAG_SND_UNA_ADVANCED; 3165 3166 if (sysctl_tcp_abc) { 3167 if (icsk->icsk_ca_state < TCP_CA_CWR) 3168 tp->bytes_acked += ack - prior_snd_una; 3169 else if (icsk->icsk_ca_state == TCP_CA_Loss) 3170 /* we assume just one segment left network */ 3171 tp->bytes_acked += min(ack - prior_snd_una, 3172 tp->mss_cache); 3173 } 3174 3175 prior_fackets = tp->fackets_out; 3176 prior_in_flight = tcp_packets_in_flight(tp); 3177 3178 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) { 3179 /* Window is constant, pure forward advance. 3180 * No more checks are required. 3181 * Note, we use the fact that SND.UNA>=SND.WL2. 3182 */ 3183 tcp_update_wl(tp, ack, ack_seq); 3184 tp->snd_una = ack; 3185 flag |= FLAG_WIN_UPDATE; 3186 3187 tcp_ca_event(sk, CA_EVENT_FAST_ACK); 3188 3189 NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS); 3190 } else { 3191 if (ack_seq != TCP_SKB_CB(skb)->end_seq) 3192 flag |= FLAG_DATA; 3193 else 3194 NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS); 3195 3196 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq); 3197 3198 if (TCP_SKB_CB(skb)->sacked) 3199 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una); 3200 3201 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb))) 3202 flag |= FLAG_ECE; 3203 3204 tcp_ca_event(sk, CA_EVENT_SLOW_ACK); 3205 } 3206 3207 /* We passed data and got it acked, remove any soft error 3208 * log. Something worked... 3209 */ 3210 sk->sk_err_soft = 0; 3211 tp->rcv_tstamp = tcp_time_stamp; 3212 prior_packets = tp->packets_out; 3213 if (!prior_packets) 3214 goto no_queue; 3215 3216 /* See if we can take anything off of the retransmit queue. */ 3217 flag |= tcp_clean_rtx_queue(sk, prior_fackets); 3218 3219 if (tp->frto_counter) 3220 frto_cwnd = tcp_process_frto(sk, flag); 3221 /* Guarantee sacktag reordering detection against wrap-arounds */ 3222 if (before(tp->frto_highmark, tp->snd_una)) 3223 tp->frto_highmark = 0; 3224 3225 if (tcp_ack_is_dubious(sk, flag)) { 3226 /* Advance CWND, if state allows this. */ 3227 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd && 3228 tcp_may_raise_cwnd(sk, flag)) 3229 tcp_cong_avoid(sk, ack, prior_in_flight); 3230 tcp_fastretrans_alert(sk, prior_packets - tp->packets_out, 3231 flag); 3232 } else { 3233 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd) 3234 tcp_cong_avoid(sk, ack, prior_in_flight); 3235 } 3236 3237 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) 3238 dst_confirm(sk->sk_dst_cache); 3239 3240 return 1; 3241 3242 no_queue: 3243 icsk->icsk_probes_out = 0; 3244 3245 /* If this ack opens up a zero window, clear backoff. It was 3246 * being used to time the probes, and is probably far higher than 3247 * it needs to be for normal retransmission. 3248 */ 3249 if (tcp_send_head(sk)) 3250 tcp_ack_probe(sk); 3251 return 1; 3252 3253 old_ack: 3254 if (TCP_SKB_CB(skb)->sacked) 3255 tcp_sacktag_write_queue(sk, skb, prior_snd_una); 3256 3257 uninteresting_ack: 3258 SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt); 3259 return 0; 3260 } 3261 3262 /* Look for tcp options. Normally only called on SYN and SYNACK packets. 3263 * But, this can also be called on packets in the established flow when 3264 * the fast version below fails. 3265 */ 3266 void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx, 3267 int estab) 3268 { 3269 unsigned char *ptr; 3270 struct tcphdr *th = tcp_hdr(skb); 3271 int length = (th->doff * 4) - sizeof(struct tcphdr); 3272 3273 ptr = (unsigned char *)(th + 1); 3274 opt_rx->saw_tstamp = 0; 3275 3276 while (length > 0) { 3277 int opcode = *ptr++; 3278 int opsize; 3279 3280 switch (opcode) { 3281 case TCPOPT_EOL: 3282 return; 3283 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */ 3284 length--; 3285 continue; 3286 default: 3287 opsize = *ptr++; 3288 if (opsize < 2) /* "silly options" */ 3289 return; 3290 if (opsize > length) 3291 return; /* don't parse partial options */ 3292 switch (opcode) { 3293 case TCPOPT_MSS: 3294 if (opsize == TCPOLEN_MSS && th->syn && !estab) { 3295 u16 in_mss = ntohs(get_unaligned((__be16 *)ptr)); 3296 if (in_mss) { 3297 if (opt_rx->user_mss && 3298 opt_rx->user_mss < in_mss) 3299 in_mss = opt_rx->user_mss; 3300 opt_rx->mss_clamp = in_mss; 3301 } 3302 } 3303 break; 3304 case TCPOPT_WINDOW: 3305 if (opsize == TCPOLEN_WINDOW && th->syn && 3306 !estab && sysctl_tcp_window_scaling) { 3307 __u8 snd_wscale = *(__u8 *)ptr; 3308 opt_rx->wscale_ok = 1; 3309 if (snd_wscale > 14) { 3310 if (net_ratelimit()) 3311 printk(KERN_INFO "tcp_parse_options: Illegal window " 3312 "scaling value %d >14 received.\n", 3313 snd_wscale); 3314 snd_wscale = 14; 3315 } 3316 opt_rx->snd_wscale = snd_wscale; 3317 } 3318 break; 3319 case TCPOPT_TIMESTAMP: 3320 if ((opsize == TCPOLEN_TIMESTAMP) && 3321 ((estab && opt_rx->tstamp_ok) || 3322 (!estab && sysctl_tcp_timestamps))) { 3323 opt_rx->saw_tstamp = 1; 3324 opt_rx->rcv_tsval = ntohl(get_unaligned((__be32 *)ptr)); 3325 opt_rx->rcv_tsecr = ntohl(get_unaligned((__be32 *)(ptr+4))); 3326 } 3327 break; 3328 case TCPOPT_SACK_PERM: 3329 if (opsize == TCPOLEN_SACK_PERM && th->syn && 3330 !estab && sysctl_tcp_sack) { 3331 opt_rx->sack_ok = 1; 3332 tcp_sack_reset(opt_rx); 3333 } 3334 break; 3335 3336 case TCPOPT_SACK: 3337 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) && 3338 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) && 3339 opt_rx->sack_ok) { 3340 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th; 3341 } 3342 break; 3343 #ifdef CONFIG_TCP_MD5SIG 3344 case TCPOPT_MD5SIG: 3345 /* 3346 * The MD5 Hash has already been 3347 * checked (see tcp_v{4,6}_do_rcv()). 3348 */ 3349 break; 3350 #endif 3351 } 3352 3353 ptr += opsize-2; 3354 length -= opsize; 3355 } 3356 } 3357 } 3358 3359 /* Fast parse options. This hopes to only see timestamps. 3360 * If it is wrong it falls back on tcp_parse_options(). 3361 */ 3362 static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th, 3363 struct tcp_sock *tp) 3364 { 3365 if (th->doff == sizeof(struct tcphdr) >> 2) { 3366 tp->rx_opt.saw_tstamp = 0; 3367 return 0; 3368 } else if (tp->rx_opt.tstamp_ok && 3369 th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) { 3370 __be32 *ptr = (__be32 *)(th + 1); 3371 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) 3372 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) { 3373 tp->rx_opt.saw_tstamp = 1; 3374 ++ptr; 3375 tp->rx_opt.rcv_tsval = ntohl(*ptr); 3376 ++ptr; 3377 tp->rx_opt.rcv_tsecr = ntohl(*ptr); 3378 return 1; 3379 } 3380 } 3381 tcp_parse_options(skb, &tp->rx_opt, 1); 3382 return 1; 3383 } 3384 3385 static inline void tcp_store_ts_recent(struct tcp_sock *tp) 3386 { 3387 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval; 3388 tp->rx_opt.ts_recent_stamp = get_seconds(); 3389 } 3390 3391 static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq) 3392 { 3393 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) { 3394 /* PAWS bug workaround wrt. ACK frames, the PAWS discard 3395 * extra check below makes sure this can only happen 3396 * for pure ACK frames. -DaveM 3397 * 3398 * Not only, also it occurs for expired timestamps. 3399 */ 3400 3401 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) >= 0 || 3402 get_seconds() >= tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS) 3403 tcp_store_ts_recent(tp); 3404 } 3405 } 3406 3407 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM 3408 * 3409 * It is not fatal. If this ACK does _not_ change critical state (seqs, window) 3410 * it can pass through stack. So, the following predicate verifies that 3411 * this segment is not used for anything but congestion avoidance or 3412 * fast retransmit. Moreover, we even are able to eliminate most of such 3413 * second order effects, if we apply some small "replay" window (~RTO) 3414 * to timestamp space. 3415 * 3416 * All these measures still do not guarantee that we reject wrapped ACKs 3417 * on networks with high bandwidth, when sequence space is recycled fastly, 3418 * but it guarantees that such events will be very rare and do not affect 3419 * connection seriously. This doesn't look nice, but alas, PAWS is really 3420 * buggy extension. 3421 * 3422 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC 3423 * states that events when retransmit arrives after original data are rare. 3424 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is 3425 * the biggest problem on large power networks even with minor reordering. 3426 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe 3427 * up to bandwidth of 18Gigabit/sec. 8) ] 3428 */ 3429 3430 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb) 3431 { 3432 struct tcp_sock *tp = tcp_sk(sk); 3433 struct tcphdr *th = tcp_hdr(skb); 3434 u32 seq = TCP_SKB_CB(skb)->seq; 3435 u32 ack = TCP_SKB_CB(skb)->ack_seq; 3436 3437 return (/* 1. Pure ACK with correct sequence number. */ 3438 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) && 3439 3440 /* 2. ... and duplicate ACK. */ 3441 ack == tp->snd_una && 3442 3443 /* 3. ... and does not update window. */ 3444 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) && 3445 3446 /* 4. ... and sits in replay window. */ 3447 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ); 3448 } 3449 3450 static inline int tcp_paws_discard(const struct sock *sk, 3451 const struct sk_buff *skb) 3452 { 3453 const struct tcp_sock *tp = tcp_sk(sk); 3454 return ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > TCP_PAWS_WINDOW && 3455 get_seconds() < tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS && 3456 !tcp_disordered_ack(sk, skb)); 3457 } 3458 3459 /* Check segment sequence number for validity. 3460 * 3461 * Segment controls are considered valid, if the segment 3462 * fits to the window after truncation to the window. Acceptability 3463 * of data (and SYN, FIN, of course) is checked separately. 3464 * See tcp_data_queue(), for example. 3465 * 3466 * Also, controls (RST is main one) are accepted using RCV.WUP instead 3467 * of RCV.NXT. Peer still did not advance his SND.UNA when we 3468 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP. 3469 * (borrowed from freebsd) 3470 */ 3471 3472 static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq) 3473 { 3474 return !before(end_seq, tp->rcv_wup) && 3475 !after(seq, tp->rcv_nxt + tcp_receive_window(tp)); 3476 } 3477 3478 /* When we get a reset we do this. */ 3479 static void tcp_reset(struct sock *sk) 3480 { 3481 /* We want the right error as BSD sees it (and indeed as we do). */ 3482 switch (sk->sk_state) { 3483 case TCP_SYN_SENT: 3484 sk->sk_err = ECONNREFUSED; 3485 break; 3486 case TCP_CLOSE_WAIT: 3487 sk->sk_err = EPIPE; 3488 break; 3489 case TCP_CLOSE: 3490 return; 3491 default: 3492 sk->sk_err = ECONNRESET; 3493 } 3494 3495 if (!sock_flag(sk, SOCK_DEAD)) 3496 sk->sk_error_report(sk); 3497 3498 tcp_done(sk); 3499 } 3500 3501 /* 3502 * Process the FIN bit. This now behaves as it is supposed to work 3503 * and the FIN takes effect when it is validly part of sequence 3504 * space. Not before when we get holes. 3505 * 3506 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT 3507 * (and thence onto LAST-ACK and finally, CLOSE, we never enter 3508 * TIME-WAIT) 3509 * 3510 * If we are in FINWAIT-1, a received FIN indicates simultaneous 3511 * close and we go into CLOSING (and later onto TIME-WAIT) 3512 * 3513 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT. 3514 */ 3515 static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th) 3516 { 3517 struct tcp_sock *tp = tcp_sk(sk); 3518 3519 inet_csk_schedule_ack(sk); 3520 3521 sk->sk_shutdown |= RCV_SHUTDOWN; 3522 sock_set_flag(sk, SOCK_DONE); 3523 3524 switch (sk->sk_state) { 3525 case TCP_SYN_RECV: 3526 case TCP_ESTABLISHED: 3527 /* Move to CLOSE_WAIT */ 3528 tcp_set_state(sk, TCP_CLOSE_WAIT); 3529 inet_csk(sk)->icsk_ack.pingpong = 1; 3530 break; 3531 3532 case TCP_CLOSE_WAIT: 3533 case TCP_CLOSING: 3534 /* Received a retransmission of the FIN, do 3535 * nothing. 3536 */ 3537 break; 3538 case TCP_LAST_ACK: 3539 /* RFC793: Remain in the LAST-ACK state. */ 3540 break; 3541 3542 case TCP_FIN_WAIT1: 3543 /* This case occurs when a simultaneous close 3544 * happens, we must ack the received FIN and 3545 * enter the CLOSING state. 3546 */ 3547 tcp_send_ack(sk); 3548 tcp_set_state(sk, TCP_CLOSING); 3549 break; 3550 case TCP_FIN_WAIT2: 3551 /* Received a FIN -- send ACK and enter TIME_WAIT. */ 3552 tcp_send_ack(sk); 3553 tcp_time_wait(sk, TCP_TIME_WAIT, 0); 3554 break; 3555 default: 3556 /* Only TCP_LISTEN and TCP_CLOSE are left, in these 3557 * cases we should never reach this piece of code. 3558 */ 3559 printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n", 3560 __FUNCTION__, sk->sk_state); 3561 break; 3562 } 3563 3564 /* It _is_ possible, that we have something out-of-order _after_ FIN. 3565 * Probably, we should reset in this case. For now drop them. 3566 */ 3567 __skb_queue_purge(&tp->out_of_order_queue); 3568 if (tcp_is_sack(tp)) 3569 tcp_sack_reset(&tp->rx_opt); 3570 sk_mem_reclaim(sk); 3571 3572 if (!sock_flag(sk, SOCK_DEAD)) { 3573 sk->sk_state_change(sk); 3574 3575 /* Do not send POLL_HUP for half duplex close. */ 3576 if (sk->sk_shutdown == SHUTDOWN_MASK || 3577 sk->sk_state == TCP_CLOSE) 3578 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP); 3579 else 3580 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN); 3581 } 3582 } 3583 3584 static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, 3585 u32 end_seq) 3586 { 3587 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) { 3588 if (before(seq, sp->start_seq)) 3589 sp->start_seq = seq; 3590 if (after(end_seq, sp->end_seq)) 3591 sp->end_seq = end_seq; 3592 return 1; 3593 } 3594 return 0; 3595 } 3596 3597 static void tcp_dsack_set(struct tcp_sock *tp, u32 seq, u32 end_seq) 3598 { 3599 if (tcp_is_sack(tp) && sysctl_tcp_dsack) { 3600 if (before(seq, tp->rcv_nxt)) 3601 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT); 3602 else 3603 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT); 3604 3605 tp->rx_opt.dsack = 1; 3606 tp->duplicate_sack[0].start_seq = seq; 3607 tp->duplicate_sack[0].end_seq = end_seq; 3608 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + 1, 3609 4 - tp->rx_opt.tstamp_ok); 3610 } 3611 } 3612 3613 static void tcp_dsack_extend(struct tcp_sock *tp, u32 seq, u32 end_seq) 3614 { 3615 if (!tp->rx_opt.dsack) 3616 tcp_dsack_set(tp, seq, end_seq); 3617 else 3618 tcp_sack_extend(tp->duplicate_sack, seq, end_seq); 3619 } 3620 3621 static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb) 3622 { 3623 struct tcp_sock *tp = tcp_sk(sk); 3624 3625 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 3626 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 3627 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST); 3628 tcp_enter_quickack_mode(sk); 3629 3630 if (tcp_is_sack(tp) && sysctl_tcp_dsack) { 3631 u32 end_seq = TCP_SKB_CB(skb)->end_seq; 3632 3633 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) 3634 end_seq = tp->rcv_nxt; 3635 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, end_seq); 3636 } 3637 } 3638 3639 tcp_send_ack(sk); 3640 } 3641 3642 /* These routines update the SACK block as out-of-order packets arrive or 3643 * in-order packets close up the sequence space. 3644 */ 3645 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp) 3646 { 3647 int this_sack; 3648 struct tcp_sack_block *sp = &tp->selective_acks[0]; 3649 struct tcp_sack_block *swalk = sp + 1; 3650 3651 /* See if the recent change to the first SACK eats into 3652 * or hits the sequence space of other SACK blocks, if so coalesce. 3653 */ 3654 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) { 3655 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) { 3656 int i; 3657 3658 /* Zap SWALK, by moving every further SACK up by one slot. 3659 * Decrease num_sacks. 3660 */ 3661 tp->rx_opt.num_sacks--; 3662 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + 3663 tp->rx_opt.dsack, 3664 4 - tp->rx_opt.tstamp_ok); 3665 for (i = this_sack; i < tp->rx_opt.num_sacks; i++) 3666 sp[i] = sp[i + 1]; 3667 continue; 3668 } 3669 this_sack++, swalk++; 3670 } 3671 } 3672 3673 static inline void tcp_sack_swap(struct tcp_sack_block *sack1, 3674 struct tcp_sack_block *sack2) 3675 { 3676 __u32 tmp; 3677 3678 tmp = sack1->start_seq; 3679 sack1->start_seq = sack2->start_seq; 3680 sack2->start_seq = tmp; 3681 3682 tmp = sack1->end_seq; 3683 sack1->end_seq = sack2->end_seq; 3684 sack2->end_seq = tmp; 3685 } 3686 3687 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq) 3688 { 3689 struct tcp_sock *tp = tcp_sk(sk); 3690 struct tcp_sack_block *sp = &tp->selective_acks[0]; 3691 int cur_sacks = tp->rx_opt.num_sacks; 3692 int this_sack; 3693 3694 if (!cur_sacks) 3695 goto new_sack; 3696 3697 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) { 3698 if (tcp_sack_extend(sp, seq, end_seq)) { 3699 /* Rotate this_sack to the first one. */ 3700 for (; this_sack > 0; this_sack--, sp--) 3701 tcp_sack_swap(sp, sp - 1); 3702 if (cur_sacks > 1) 3703 tcp_sack_maybe_coalesce(tp); 3704 return; 3705 } 3706 } 3707 3708 /* Could not find an adjacent existing SACK, build a new one, 3709 * put it at the front, and shift everyone else down. We 3710 * always know there is at least one SACK present already here. 3711 * 3712 * If the sack array is full, forget about the last one. 3713 */ 3714 if (this_sack >= 4) { 3715 this_sack--; 3716 tp->rx_opt.num_sacks--; 3717 sp--; 3718 } 3719 for (; this_sack > 0; this_sack--, sp--) 3720 *sp = *(sp - 1); 3721 3722 new_sack: 3723 /* Build the new head SACK, and we're done. */ 3724 sp->start_seq = seq; 3725 sp->end_seq = end_seq; 3726 tp->rx_opt.num_sacks++; 3727 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 3728 4 - tp->rx_opt.tstamp_ok); 3729 } 3730 3731 /* RCV.NXT advances, some SACKs should be eaten. */ 3732 3733 static void tcp_sack_remove(struct tcp_sock *tp) 3734 { 3735 struct tcp_sack_block *sp = &tp->selective_acks[0]; 3736 int num_sacks = tp->rx_opt.num_sacks; 3737 int this_sack; 3738 3739 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */ 3740 if (skb_queue_empty(&tp->out_of_order_queue)) { 3741 tp->rx_opt.num_sacks = 0; 3742 tp->rx_opt.eff_sacks = tp->rx_opt.dsack; 3743 return; 3744 } 3745 3746 for (this_sack = 0; this_sack < num_sacks;) { 3747 /* Check if the start of the sack is covered by RCV.NXT. */ 3748 if (!before(tp->rcv_nxt, sp->start_seq)) { 3749 int i; 3750 3751 /* RCV.NXT must cover all the block! */ 3752 BUG_TRAP(!before(tp->rcv_nxt, sp->end_seq)); 3753 3754 /* Zap this SACK, by moving forward any other SACKS. */ 3755 for (i=this_sack+1; i < num_sacks; i++) 3756 tp->selective_acks[i-1] = tp->selective_acks[i]; 3757 num_sacks--; 3758 continue; 3759 } 3760 this_sack++; 3761 sp++; 3762 } 3763 if (num_sacks != tp->rx_opt.num_sacks) { 3764 tp->rx_opt.num_sacks = num_sacks; 3765 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + 3766 tp->rx_opt.dsack, 3767 4 - tp->rx_opt.tstamp_ok); 3768 } 3769 } 3770 3771 /* This one checks to see if we can put data from the 3772 * out_of_order queue into the receive_queue. 3773 */ 3774 static void tcp_ofo_queue(struct sock *sk) 3775 { 3776 struct tcp_sock *tp = tcp_sk(sk); 3777 __u32 dsack_high = tp->rcv_nxt; 3778 struct sk_buff *skb; 3779 3780 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) { 3781 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) 3782 break; 3783 3784 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) { 3785 __u32 dsack = dsack_high; 3786 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high)) 3787 dsack_high = TCP_SKB_CB(skb)->end_seq; 3788 tcp_dsack_extend(tp, TCP_SKB_CB(skb)->seq, dsack); 3789 } 3790 3791 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) { 3792 SOCK_DEBUG(sk, "ofo packet was already received \n"); 3793 __skb_unlink(skb, &tp->out_of_order_queue); 3794 __kfree_skb(skb); 3795 continue; 3796 } 3797 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n", 3798 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, 3799 TCP_SKB_CB(skb)->end_seq); 3800 3801 __skb_unlink(skb, &tp->out_of_order_queue); 3802 __skb_queue_tail(&sk->sk_receive_queue, skb); 3803 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; 3804 if (tcp_hdr(skb)->fin) 3805 tcp_fin(skb, sk, tcp_hdr(skb)); 3806 } 3807 } 3808 3809 static int tcp_prune_queue(struct sock *sk); 3810 3811 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb) 3812 { 3813 struct tcphdr *th = tcp_hdr(skb); 3814 struct tcp_sock *tp = tcp_sk(sk); 3815 int eaten = -1; 3816 3817 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) 3818 goto drop; 3819 3820 __skb_pull(skb, th->doff * 4); 3821 3822 TCP_ECN_accept_cwr(tp, skb); 3823 3824 if (tp->rx_opt.dsack) { 3825 tp->rx_opt.dsack = 0; 3826 tp->rx_opt.eff_sacks = min_t(unsigned int, tp->rx_opt.num_sacks, 3827 4 - tp->rx_opt.tstamp_ok); 3828 } 3829 3830 /* Queue data for delivery to the user. 3831 * Packets in sequence go to the receive queue. 3832 * Out of sequence packets to the out_of_order_queue. 3833 */ 3834 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) { 3835 if (tcp_receive_window(tp) == 0) 3836 goto out_of_window; 3837 3838 /* Ok. In sequence. In window. */ 3839 if (tp->ucopy.task == current && 3840 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len && 3841 sock_owned_by_user(sk) && !tp->urg_data) { 3842 int chunk = min_t(unsigned int, skb->len, 3843 tp->ucopy.len); 3844 3845 __set_current_state(TASK_RUNNING); 3846 3847 local_bh_enable(); 3848 if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) { 3849 tp->ucopy.len -= chunk; 3850 tp->copied_seq += chunk; 3851 eaten = (chunk == skb->len && !th->fin); 3852 tcp_rcv_space_adjust(sk); 3853 } 3854 local_bh_disable(); 3855 } 3856 3857 if (eaten <= 0) { 3858 queue_and_out: 3859 if (eaten < 0 && 3860 (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf || 3861 !sk_rmem_schedule(sk, skb->truesize))) { 3862 if (tcp_prune_queue(sk) < 0 || 3863 !sk_rmem_schedule(sk, skb->truesize)) 3864 goto drop; 3865 } 3866 skb_set_owner_r(skb, sk); 3867 __skb_queue_tail(&sk->sk_receive_queue, skb); 3868 } 3869 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; 3870 if (skb->len) 3871 tcp_event_data_recv(sk, skb); 3872 if (th->fin) 3873 tcp_fin(skb, sk, th); 3874 3875 if (!skb_queue_empty(&tp->out_of_order_queue)) { 3876 tcp_ofo_queue(sk); 3877 3878 /* RFC2581. 4.2. SHOULD send immediate ACK, when 3879 * gap in queue is filled. 3880 */ 3881 if (skb_queue_empty(&tp->out_of_order_queue)) 3882 inet_csk(sk)->icsk_ack.pingpong = 0; 3883 } 3884 3885 if (tp->rx_opt.num_sacks) 3886 tcp_sack_remove(tp); 3887 3888 tcp_fast_path_check(sk); 3889 3890 if (eaten > 0) 3891 __kfree_skb(skb); 3892 else if (!sock_flag(sk, SOCK_DEAD)) 3893 sk->sk_data_ready(sk, 0); 3894 return; 3895 } 3896 3897 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) { 3898 /* A retransmit, 2nd most common case. Force an immediate ack. */ 3899 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST); 3900 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); 3901 3902 out_of_window: 3903 tcp_enter_quickack_mode(sk); 3904 inet_csk_schedule_ack(sk); 3905 drop: 3906 __kfree_skb(skb); 3907 return; 3908 } 3909 3910 /* Out of window. F.e. zero window probe. */ 3911 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp))) 3912 goto out_of_window; 3913 3914 tcp_enter_quickack_mode(sk); 3915 3916 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 3917 /* Partial packet, seq < rcv_next < end_seq */ 3918 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n", 3919 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, 3920 TCP_SKB_CB(skb)->end_seq); 3921 3922 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, tp->rcv_nxt); 3923 3924 /* If window is closed, drop tail of packet. But after 3925 * remembering D-SACK for its head made in previous line. 3926 */ 3927 if (!tcp_receive_window(tp)) 3928 goto out_of_window; 3929 goto queue_and_out; 3930 } 3931 3932 TCP_ECN_check_ce(tp, skb); 3933 3934 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf || 3935 !sk_rmem_schedule(sk, skb->truesize)) { 3936 if (tcp_prune_queue(sk) < 0 || 3937 !sk_rmem_schedule(sk, skb->truesize)) 3938 goto drop; 3939 } 3940 3941 /* Disable header prediction. */ 3942 tp->pred_flags = 0; 3943 inet_csk_schedule_ack(sk); 3944 3945 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n", 3946 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); 3947 3948 skb_set_owner_r(skb, sk); 3949 3950 if (!skb_peek(&tp->out_of_order_queue)) { 3951 /* Initial out of order segment, build 1 SACK. */ 3952 if (tcp_is_sack(tp)) { 3953 tp->rx_opt.num_sacks = 1; 3954 tp->rx_opt.dsack = 0; 3955 tp->rx_opt.eff_sacks = 1; 3956 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq; 3957 tp->selective_acks[0].end_seq = 3958 TCP_SKB_CB(skb)->end_seq; 3959 } 3960 __skb_queue_head(&tp->out_of_order_queue, skb); 3961 } else { 3962 struct sk_buff *skb1 = tp->out_of_order_queue.prev; 3963 u32 seq = TCP_SKB_CB(skb)->seq; 3964 u32 end_seq = TCP_SKB_CB(skb)->end_seq; 3965 3966 if (seq == TCP_SKB_CB(skb1)->end_seq) { 3967 __skb_append(skb1, skb, &tp->out_of_order_queue); 3968 3969 if (!tp->rx_opt.num_sacks || 3970 tp->selective_acks[0].end_seq != seq) 3971 goto add_sack; 3972 3973 /* Common case: data arrive in order after hole. */ 3974 tp->selective_acks[0].end_seq = end_seq; 3975 return; 3976 } 3977 3978 /* Find place to insert this segment. */ 3979 do { 3980 if (!after(TCP_SKB_CB(skb1)->seq, seq)) 3981 break; 3982 } while ((skb1 = skb1->prev) != 3983 (struct sk_buff *)&tp->out_of_order_queue); 3984 3985 /* Do skb overlap to previous one? */ 3986 if (skb1 != (struct sk_buff *)&tp->out_of_order_queue && 3987 before(seq, TCP_SKB_CB(skb1)->end_seq)) { 3988 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) { 3989 /* All the bits are present. Drop. */ 3990 __kfree_skb(skb); 3991 tcp_dsack_set(tp, seq, end_seq); 3992 goto add_sack; 3993 } 3994 if (after(seq, TCP_SKB_CB(skb1)->seq)) { 3995 /* Partial overlap. */ 3996 tcp_dsack_set(tp, seq, 3997 TCP_SKB_CB(skb1)->end_seq); 3998 } else { 3999 skb1 = skb1->prev; 4000 } 4001 } 4002 __skb_insert(skb, skb1, skb1->next, &tp->out_of_order_queue); 4003 4004 /* And clean segments covered by new one as whole. */ 4005 while ((skb1 = skb->next) != 4006 (struct sk_buff *)&tp->out_of_order_queue && 4007 after(end_seq, TCP_SKB_CB(skb1)->seq)) { 4008 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) { 4009 tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, 4010 end_seq); 4011 break; 4012 } 4013 __skb_unlink(skb1, &tp->out_of_order_queue); 4014 tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, 4015 TCP_SKB_CB(skb1)->end_seq); 4016 __kfree_skb(skb1); 4017 } 4018 4019 add_sack: 4020 if (tcp_is_sack(tp)) 4021 tcp_sack_new_ofo_skb(sk, seq, end_seq); 4022 } 4023 } 4024 4025 /* Collapse contiguous sequence of skbs head..tail with 4026 * sequence numbers start..end. 4027 * Segments with FIN/SYN are not collapsed (only because this 4028 * simplifies code) 4029 */ 4030 static void 4031 tcp_collapse(struct sock *sk, struct sk_buff_head *list, 4032 struct sk_buff *head, struct sk_buff *tail, 4033 u32 start, u32 end) 4034 { 4035 struct sk_buff *skb; 4036 4037 /* First, check that queue is collapsible and find 4038 * the point where collapsing can be useful. */ 4039 for (skb = head; skb != tail;) { 4040 /* No new bits? It is possible on ofo queue. */ 4041 if (!before(start, TCP_SKB_CB(skb)->end_seq)) { 4042 struct sk_buff *next = skb->next; 4043 __skb_unlink(skb, list); 4044 __kfree_skb(skb); 4045 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED); 4046 skb = next; 4047 continue; 4048 } 4049 4050 /* The first skb to collapse is: 4051 * - not SYN/FIN and 4052 * - bloated or contains data before "start" or 4053 * overlaps to the next one. 4054 */ 4055 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin && 4056 (tcp_win_from_space(skb->truesize) > skb->len || 4057 before(TCP_SKB_CB(skb)->seq, start) || 4058 (skb->next != tail && 4059 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq))) 4060 break; 4061 4062 /* Decided to skip this, advance start seq. */ 4063 start = TCP_SKB_CB(skb)->end_seq; 4064 skb = skb->next; 4065 } 4066 if (skb == tail || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin) 4067 return; 4068 4069 while (before(start, end)) { 4070 struct sk_buff *nskb; 4071 unsigned int header = skb_headroom(skb); 4072 int copy = SKB_MAX_ORDER(header, 0); 4073 4074 /* Too big header? This can happen with IPv6. */ 4075 if (copy < 0) 4076 return; 4077 if (end - start < copy) 4078 copy = end - start; 4079 nskb = alloc_skb(copy + header, GFP_ATOMIC); 4080 if (!nskb) 4081 return; 4082 4083 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head); 4084 skb_set_network_header(nskb, (skb_network_header(skb) - 4085 skb->head)); 4086 skb_set_transport_header(nskb, (skb_transport_header(skb) - 4087 skb->head)); 4088 skb_reserve(nskb, header); 4089 memcpy(nskb->head, skb->head, header); 4090 memcpy(nskb->cb, skb->cb, sizeof(skb->cb)); 4091 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start; 4092 __skb_insert(nskb, skb->prev, skb, list); 4093 skb_set_owner_r(nskb, sk); 4094 4095 /* Copy data, releasing collapsed skbs. */ 4096 while (copy > 0) { 4097 int offset = start - TCP_SKB_CB(skb)->seq; 4098 int size = TCP_SKB_CB(skb)->end_seq - start; 4099 4100 BUG_ON(offset < 0); 4101 if (size > 0) { 4102 size = min(copy, size); 4103 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size)) 4104 BUG(); 4105 TCP_SKB_CB(nskb)->end_seq += size; 4106 copy -= size; 4107 start += size; 4108 } 4109 if (!before(start, TCP_SKB_CB(skb)->end_seq)) { 4110 struct sk_buff *next = skb->next; 4111 __skb_unlink(skb, list); 4112 __kfree_skb(skb); 4113 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED); 4114 skb = next; 4115 if (skb == tail || 4116 tcp_hdr(skb)->syn || 4117 tcp_hdr(skb)->fin) 4118 return; 4119 } 4120 } 4121 } 4122 } 4123 4124 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs 4125 * and tcp_collapse() them until all the queue is collapsed. 4126 */ 4127 static void tcp_collapse_ofo_queue(struct sock *sk) 4128 { 4129 struct tcp_sock *tp = tcp_sk(sk); 4130 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue); 4131 struct sk_buff *head; 4132 u32 start, end; 4133 4134 if (skb == NULL) 4135 return; 4136 4137 start = TCP_SKB_CB(skb)->seq; 4138 end = TCP_SKB_CB(skb)->end_seq; 4139 head = skb; 4140 4141 for (;;) { 4142 skb = skb->next; 4143 4144 /* Segment is terminated when we see gap or when 4145 * we are at the end of all the queue. */ 4146 if (skb == (struct sk_buff *)&tp->out_of_order_queue || 4147 after(TCP_SKB_CB(skb)->seq, end) || 4148 before(TCP_SKB_CB(skb)->end_seq, start)) { 4149 tcp_collapse(sk, &tp->out_of_order_queue, 4150 head, skb, start, end); 4151 head = skb; 4152 if (skb == (struct sk_buff *)&tp->out_of_order_queue) 4153 break; 4154 /* Start new segment */ 4155 start = TCP_SKB_CB(skb)->seq; 4156 end = TCP_SKB_CB(skb)->end_seq; 4157 } else { 4158 if (before(TCP_SKB_CB(skb)->seq, start)) 4159 start = TCP_SKB_CB(skb)->seq; 4160 if (after(TCP_SKB_CB(skb)->end_seq, end)) 4161 end = TCP_SKB_CB(skb)->end_seq; 4162 } 4163 } 4164 } 4165 4166 /* Reduce allocated memory if we can, trying to get 4167 * the socket within its memory limits again. 4168 * 4169 * Return less than zero if we should start dropping frames 4170 * until the socket owning process reads some of the data 4171 * to stabilize the situation. 4172 */ 4173 static int tcp_prune_queue(struct sock *sk) 4174 { 4175 struct tcp_sock *tp = tcp_sk(sk); 4176 4177 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq); 4178 4179 NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED); 4180 4181 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf) 4182 tcp_clamp_window(sk); 4183 else if (tcp_memory_pressure) 4184 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss); 4185 4186 tcp_collapse_ofo_queue(sk); 4187 tcp_collapse(sk, &sk->sk_receive_queue, 4188 sk->sk_receive_queue.next, 4189 (struct sk_buff *)&sk->sk_receive_queue, 4190 tp->copied_seq, tp->rcv_nxt); 4191 sk_mem_reclaim(sk); 4192 4193 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) 4194 return 0; 4195 4196 /* Collapsing did not help, destructive actions follow. 4197 * This must not ever occur. */ 4198 4199 /* First, purge the out_of_order queue. */ 4200 if (!skb_queue_empty(&tp->out_of_order_queue)) { 4201 NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED); 4202 __skb_queue_purge(&tp->out_of_order_queue); 4203 4204 /* Reset SACK state. A conforming SACK implementation will 4205 * do the same at a timeout based retransmit. When a connection 4206 * is in a sad state like this, we care only about integrity 4207 * of the connection not performance. 4208 */ 4209 if (tcp_is_sack(tp)) 4210 tcp_sack_reset(&tp->rx_opt); 4211 sk_mem_reclaim(sk); 4212 } 4213 4214 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) 4215 return 0; 4216 4217 /* If we are really being abused, tell the caller to silently 4218 * drop receive data on the floor. It will get retransmitted 4219 * and hopefully then we'll have sufficient space. 4220 */ 4221 NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED); 4222 4223 /* Massive buffer overcommit. */ 4224 tp->pred_flags = 0; 4225 return -1; 4226 } 4227 4228 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. 4229 * As additional protections, we do not touch cwnd in retransmission phases, 4230 * and if application hit its sndbuf limit recently. 4231 */ 4232 void tcp_cwnd_application_limited(struct sock *sk) 4233 { 4234 struct tcp_sock *tp = tcp_sk(sk); 4235 4236 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && 4237 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { 4238 /* Limited by application or receiver window. */ 4239 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); 4240 u32 win_used = max(tp->snd_cwnd_used, init_win); 4241 if (win_used < tp->snd_cwnd) { 4242 tp->snd_ssthresh = tcp_current_ssthresh(sk); 4243 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1; 4244 } 4245 tp->snd_cwnd_used = 0; 4246 } 4247 tp->snd_cwnd_stamp = tcp_time_stamp; 4248 } 4249 4250 static int tcp_should_expand_sndbuf(struct sock *sk) 4251 { 4252 struct tcp_sock *tp = tcp_sk(sk); 4253 4254 /* If the user specified a specific send buffer setting, do 4255 * not modify it. 4256 */ 4257 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK) 4258 return 0; 4259 4260 /* If we are under global TCP memory pressure, do not expand. */ 4261 if (tcp_memory_pressure) 4262 return 0; 4263 4264 /* If we are under soft global TCP memory pressure, do not expand. */ 4265 if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0]) 4266 return 0; 4267 4268 /* If we filled the congestion window, do not expand. */ 4269 if (tp->packets_out >= tp->snd_cwnd) 4270 return 0; 4271 4272 return 1; 4273 } 4274 4275 /* When incoming ACK allowed to free some skb from write_queue, 4276 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket 4277 * on the exit from tcp input handler. 4278 * 4279 * PROBLEM: sndbuf expansion does not work well with largesend. 4280 */ 4281 static void tcp_new_space(struct sock *sk) 4282 { 4283 struct tcp_sock *tp = tcp_sk(sk); 4284 4285 if (tcp_should_expand_sndbuf(sk)) { 4286 int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) + 4287 MAX_TCP_HEADER + 16 + sizeof(struct sk_buff), 4288 demanded = max_t(unsigned int, tp->snd_cwnd, 4289 tp->reordering + 1); 4290 sndmem *= 2 * demanded; 4291 if (sndmem > sk->sk_sndbuf) 4292 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]); 4293 tp->snd_cwnd_stamp = tcp_time_stamp; 4294 } 4295 4296 sk->sk_write_space(sk); 4297 } 4298 4299 static void tcp_check_space(struct sock *sk) 4300 { 4301 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) { 4302 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK); 4303 if (sk->sk_socket && 4304 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) 4305 tcp_new_space(sk); 4306 } 4307 } 4308 4309 static inline void tcp_data_snd_check(struct sock *sk) 4310 { 4311 tcp_push_pending_frames(sk); 4312 tcp_check_space(sk); 4313 } 4314 4315 /* 4316 * Check if sending an ack is needed. 4317 */ 4318 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible) 4319 { 4320 struct tcp_sock *tp = tcp_sk(sk); 4321 4322 /* More than one full frame received... */ 4323 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss 4324 /* ... and right edge of window advances far enough. 4325 * (tcp_recvmsg() will send ACK otherwise). Or... 4326 */ 4327 && __tcp_select_window(sk) >= tp->rcv_wnd) || 4328 /* We ACK each frame or... */ 4329 tcp_in_quickack_mode(sk) || 4330 /* We have out of order data. */ 4331 (ofo_possible && skb_peek(&tp->out_of_order_queue))) { 4332 /* Then ack it now */ 4333 tcp_send_ack(sk); 4334 } else { 4335 /* Else, send delayed ack. */ 4336 tcp_send_delayed_ack(sk); 4337 } 4338 } 4339 4340 static inline void tcp_ack_snd_check(struct sock *sk) 4341 { 4342 if (!inet_csk_ack_scheduled(sk)) { 4343 /* We sent a data segment already. */ 4344 return; 4345 } 4346 __tcp_ack_snd_check(sk, 1); 4347 } 4348 4349 /* 4350 * This routine is only called when we have urgent data 4351 * signaled. Its the 'slow' part of tcp_urg. It could be 4352 * moved inline now as tcp_urg is only called from one 4353 * place. We handle URGent data wrong. We have to - as 4354 * BSD still doesn't use the correction from RFC961. 4355 * For 1003.1g we should support a new option TCP_STDURG to permit 4356 * either form (or just set the sysctl tcp_stdurg). 4357 */ 4358 4359 static void tcp_check_urg(struct sock *sk, struct tcphdr *th) 4360 { 4361 struct tcp_sock *tp = tcp_sk(sk); 4362 u32 ptr = ntohs(th->urg_ptr); 4363 4364 if (ptr && !sysctl_tcp_stdurg) 4365 ptr--; 4366 ptr += ntohl(th->seq); 4367 4368 /* Ignore urgent data that we've already seen and read. */ 4369 if (after(tp->copied_seq, ptr)) 4370 return; 4371 4372 /* Do not replay urg ptr. 4373 * 4374 * NOTE: interesting situation not covered by specs. 4375 * Misbehaving sender may send urg ptr, pointing to segment, 4376 * which we already have in ofo queue. We are not able to fetch 4377 * such data and will stay in TCP_URG_NOTYET until will be eaten 4378 * by recvmsg(). Seems, we are not obliged to handle such wicked 4379 * situations. But it is worth to think about possibility of some 4380 * DoSes using some hypothetical application level deadlock. 4381 */ 4382 if (before(ptr, tp->rcv_nxt)) 4383 return; 4384 4385 /* Do we already have a newer (or duplicate) urgent pointer? */ 4386 if (tp->urg_data && !after(ptr, tp->urg_seq)) 4387 return; 4388 4389 /* Tell the world about our new urgent pointer. */ 4390 sk_send_sigurg(sk); 4391 4392 /* We may be adding urgent data when the last byte read was 4393 * urgent. To do this requires some care. We cannot just ignore 4394 * tp->copied_seq since we would read the last urgent byte again 4395 * as data, nor can we alter copied_seq until this data arrives 4396 * or we break the semantics of SIOCATMARK (and thus sockatmark()) 4397 * 4398 * NOTE. Double Dutch. Rendering to plain English: author of comment 4399 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB); 4400 * and expect that both A and B disappear from stream. This is _wrong_. 4401 * Though this happens in BSD with high probability, this is occasional. 4402 * Any application relying on this is buggy. Note also, that fix "works" 4403 * only in this artificial test. Insert some normal data between A and B and we will 4404 * decline of BSD again. Verdict: it is better to remove to trap 4405 * buggy users. 4406 */ 4407 if (tp->urg_seq == tp->copied_seq && tp->urg_data && 4408 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) { 4409 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue); 4410 tp->copied_seq++; 4411 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) { 4412 __skb_unlink(skb, &sk->sk_receive_queue); 4413 __kfree_skb(skb); 4414 } 4415 } 4416 4417 tp->urg_data = TCP_URG_NOTYET; 4418 tp->urg_seq = ptr; 4419 4420 /* Disable header prediction. */ 4421 tp->pred_flags = 0; 4422 } 4423 4424 /* This is the 'fast' part of urgent handling. */ 4425 static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th) 4426 { 4427 struct tcp_sock *tp = tcp_sk(sk); 4428 4429 /* Check if we get a new urgent pointer - normally not. */ 4430 if (th->urg) 4431 tcp_check_urg(sk, th); 4432 4433 /* Do we wait for any urgent data? - normally not... */ 4434 if (tp->urg_data == TCP_URG_NOTYET) { 4435 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) - 4436 th->syn; 4437 4438 /* Is the urgent pointer pointing into this packet? */ 4439 if (ptr < skb->len) { 4440 u8 tmp; 4441 if (skb_copy_bits(skb, ptr, &tmp, 1)) 4442 BUG(); 4443 tp->urg_data = TCP_URG_VALID | tmp; 4444 if (!sock_flag(sk, SOCK_DEAD)) 4445 sk->sk_data_ready(sk, 0); 4446 } 4447 } 4448 } 4449 4450 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen) 4451 { 4452 struct tcp_sock *tp = tcp_sk(sk); 4453 int chunk = skb->len - hlen; 4454 int err; 4455 4456 local_bh_enable(); 4457 if (skb_csum_unnecessary(skb)) 4458 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk); 4459 else 4460 err = skb_copy_and_csum_datagram_iovec(skb, hlen, 4461 tp->ucopy.iov); 4462 4463 if (!err) { 4464 tp->ucopy.len -= chunk; 4465 tp->copied_seq += chunk; 4466 tcp_rcv_space_adjust(sk); 4467 } 4468 4469 local_bh_disable(); 4470 return err; 4471 } 4472 4473 static __sum16 __tcp_checksum_complete_user(struct sock *sk, 4474 struct sk_buff *skb) 4475 { 4476 __sum16 result; 4477 4478 if (sock_owned_by_user(sk)) { 4479 local_bh_enable(); 4480 result = __tcp_checksum_complete(skb); 4481 local_bh_disable(); 4482 } else { 4483 result = __tcp_checksum_complete(skb); 4484 } 4485 return result; 4486 } 4487 4488 static inline int tcp_checksum_complete_user(struct sock *sk, 4489 struct sk_buff *skb) 4490 { 4491 return !skb_csum_unnecessary(skb) && 4492 __tcp_checksum_complete_user(sk, skb); 4493 } 4494 4495 #ifdef CONFIG_NET_DMA 4496 static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb, 4497 int hlen) 4498 { 4499 struct tcp_sock *tp = tcp_sk(sk); 4500 int chunk = skb->len - hlen; 4501 int dma_cookie; 4502 int copied_early = 0; 4503 4504 if (tp->ucopy.wakeup) 4505 return 0; 4506 4507 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list) 4508 tp->ucopy.dma_chan = get_softnet_dma(); 4509 4510 if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) { 4511 4512 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan, 4513 skb, hlen, 4514 tp->ucopy.iov, chunk, 4515 tp->ucopy.pinned_list); 4516 4517 if (dma_cookie < 0) 4518 goto out; 4519 4520 tp->ucopy.dma_cookie = dma_cookie; 4521 copied_early = 1; 4522 4523 tp->ucopy.len -= chunk; 4524 tp->copied_seq += chunk; 4525 tcp_rcv_space_adjust(sk); 4526 4527 if ((tp->ucopy.len == 0) || 4528 (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) || 4529 (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) { 4530 tp->ucopy.wakeup = 1; 4531 sk->sk_data_ready(sk, 0); 4532 } 4533 } else if (chunk > 0) { 4534 tp->ucopy.wakeup = 1; 4535 sk->sk_data_ready(sk, 0); 4536 } 4537 out: 4538 return copied_early; 4539 } 4540 #endif /* CONFIG_NET_DMA */ 4541 4542 /* 4543 * TCP receive function for the ESTABLISHED state. 4544 * 4545 * It is split into a fast path and a slow path. The fast path is 4546 * disabled when: 4547 * - A zero window was announced from us - zero window probing 4548 * is only handled properly in the slow path. 4549 * - Out of order segments arrived. 4550 * - Urgent data is expected. 4551 * - There is no buffer space left 4552 * - Unexpected TCP flags/window values/header lengths are received 4553 * (detected by checking the TCP header against pred_flags) 4554 * - Data is sent in both directions. Fast path only supports pure senders 4555 * or pure receivers (this means either the sequence number or the ack 4556 * value must stay constant) 4557 * - Unexpected TCP option. 4558 * 4559 * When these conditions are not satisfied it drops into a standard 4560 * receive procedure patterned after RFC793 to handle all cases. 4561 * The first three cases are guaranteed by proper pred_flags setting, 4562 * the rest is checked inline. Fast processing is turned on in 4563 * tcp_data_queue when everything is OK. 4564 */ 4565 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb, 4566 struct tcphdr *th, unsigned len) 4567 { 4568 struct tcp_sock *tp = tcp_sk(sk); 4569 4570 /* 4571 * Header prediction. 4572 * The code loosely follows the one in the famous 4573 * "30 instruction TCP receive" Van Jacobson mail. 4574 * 4575 * Van's trick is to deposit buffers into socket queue 4576 * on a device interrupt, to call tcp_recv function 4577 * on the receive process context and checksum and copy 4578 * the buffer to user space. smart... 4579 * 4580 * Our current scheme is not silly either but we take the 4581 * extra cost of the net_bh soft interrupt processing... 4582 * We do checksum and copy also but from device to kernel. 4583 */ 4584 4585 tp->rx_opt.saw_tstamp = 0; 4586 4587 /* pred_flags is 0xS?10 << 16 + snd_wnd 4588 * if header_prediction is to be made 4589 * 'S' will always be tp->tcp_header_len >> 2 4590 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to 4591 * turn it off (when there are holes in the receive 4592 * space for instance) 4593 * PSH flag is ignored. 4594 */ 4595 4596 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags && 4597 TCP_SKB_CB(skb)->seq == tp->rcv_nxt) { 4598 int tcp_header_len = tp->tcp_header_len; 4599 4600 /* Timestamp header prediction: tcp_header_len 4601 * is automatically equal to th->doff*4 due to pred_flags 4602 * match. 4603 */ 4604 4605 /* Check timestamp */ 4606 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) { 4607 __be32 *ptr = (__be32 *)(th + 1); 4608 4609 /* No? Slow path! */ 4610 if (*ptr != htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) 4611 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) 4612 goto slow_path; 4613 4614 tp->rx_opt.saw_tstamp = 1; 4615 ++ptr; 4616 tp->rx_opt.rcv_tsval = ntohl(*ptr); 4617 ++ptr; 4618 tp->rx_opt.rcv_tsecr = ntohl(*ptr); 4619 4620 /* If PAWS failed, check it more carefully in slow path */ 4621 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0) 4622 goto slow_path; 4623 4624 /* DO NOT update ts_recent here, if checksum fails 4625 * and timestamp was corrupted part, it will result 4626 * in a hung connection since we will drop all 4627 * future packets due to the PAWS test. 4628 */ 4629 } 4630 4631 if (len <= tcp_header_len) { 4632 /* Bulk data transfer: sender */ 4633 if (len == tcp_header_len) { 4634 /* Predicted packet is in window by definition. 4635 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 4636 * Hence, check seq<=rcv_wup reduces to: 4637 */ 4638 if (tcp_header_len == 4639 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && 4640 tp->rcv_nxt == tp->rcv_wup) 4641 tcp_store_ts_recent(tp); 4642 4643 /* We know that such packets are checksummed 4644 * on entry. 4645 */ 4646 tcp_ack(sk, skb, 0); 4647 __kfree_skb(skb); 4648 tcp_data_snd_check(sk); 4649 return 0; 4650 } else { /* Header too small */ 4651 TCP_INC_STATS_BH(TCP_MIB_INERRS); 4652 goto discard; 4653 } 4654 } else { 4655 int eaten = 0; 4656 int copied_early = 0; 4657 4658 if (tp->copied_seq == tp->rcv_nxt && 4659 len - tcp_header_len <= tp->ucopy.len) { 4660 #ifdef CONFIG_NET_DMA 4661 if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) { 4662 copied_early = 1; 4663 eaten = 1; 4664 } 4665 #endif 4666 if (tp->ucopy.task == current && 4667 sock_owned_by_user(sk) && !copied_early) { 4668 __set_current_state(TASK_RUNNING); 4669 4670 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len)) 4671 eaten = 1; 4672 } 4673 if (eaten) { 4674 /* Predicted packet is in window by definition. 4675 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 4676 * Hence, check seq<=rcv_wup reduces to: 4677 */ 4678 if (tcp_header_len == 4679 (sizeof(struct tcphdr) + 4680 TCPOLEN_TSTAMP_ALIGNED) && 4681 tp->rcv_nxt == tp->rcv_wup) 4682 tcp_store_ts_recent(tp); 4683 4684 tcp_rcv_rtt_measure_ts(sk, skb); 4685 4686 __skb_pull(skb, tcp_header_len); 4687 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; 4688 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER); 4689 } 4690 if (copied_early) 4691 tcp_cleanup_rbuf(sk, skb->len); 4692 } 4693 if (!eaten) { 4694 if (tcp_checksum_complete_user(sk, skb)) 4695 goto csum_error; 4696 4697 /* Predicted packet is in window by definition. 4698 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 4699 * Hence, check seq<=rcv_wup reduces to: 4700 */ 4701 if (tcp_header_len == 4702 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && 4703 tp->rcv_nxt == tp->rcv_wup) 4704 tcp_store_ts_recent(tp); 4705 4706 tcp_rcv_rtt_measure_ts(sk, skb); 4707 4708 if ((int)skb->truesize > sk->sk_forward_alloc) 4709 goto step5; 4710 4711 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS); 4712 4713 /* Bulk data transfer: receiver */ 4714 __skb_pull(skb, tcp_header_len); 4715 __skb_queue_tail(&sk->sk_receive_queue, skb); 4716 skb_set_owner_r(skb, sk); 4717 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; 4718 } 4719 4720 tcp_event_data_recv(sk, skb); 4721 4722 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) { 4723 /* Well, only one small jumplet in fast path... */ 4724 tcp_ack(sk, skb, FLAG_DATA); 4725 tcp_data_snd_check(sk); 4726 if (!inet_csk_ack_scheduled(sk)) 4727 goto no_ack; 4728 } 4729 4730 __tcp_ack_snd_check(sk, 0); 4731 no_ack: 4732 #ifdef CONFIG_NET_DMA 4733 if (copied_early) 4734 __skb_queue_tail(&sk->sk_async_wait_queue, skb); 4735 else 4736 #endif 4737 if (eaten) 4738 __kfree_skb(skb); 4739 else 4740 sk->sk_data_ready(sk, 0); 4741 return 0; 4742 } 4743 } 4744 4745 slow_path: 4746 if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb)) 4747 goto csum_error; 4748 4749 /* 4750 * RFC1323: H1. Apply PAWS check first. 4751 */ 4752 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp && 4753 tcp_paws_discard(sk, skb)) { 4754 if (!th->rst) { 4755 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED); 4756 tcp_send_dupack(sk, skb); 4757 goto discard; 4758 } 4759 /* Resets are accepted even if PAWS failed. 4760 4761 ts_recent update must be made after we are sure 4762 that the packet is in window. 4763 */ 4764 } 4765 4766 /* 4767 * Standard slow path. 4768 */ 4769 4770 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) { 4771 /* RFC793, page 37: "In all states except SYN-SENT, all reset 4772 * (RST) segments are validated by checking their SEQ-fields." 4773 * And page 69: "If an incoming segment is not acceptable, 4774 * an acknowledgment should be sent in reply (unless the RST bit 4775 * is set, if so drop the segment and return)". 4776 */ 4777 if (!th->rst) 4778 tcp_send_dupack(sk, skb); 4779 goto discard; 4780 } 4781 4782 if (th->rst) { 4783 tcp_reset(sk); 4784 goto discard; 4785 } 4786 4787 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq); 4788 4789 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 4790 TCP_INC_STATS_BH(TCP_MIB_INERRS); 4791 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN); 4792 tcp_reset(sk); 4793 return 1; 4794 } 4795 4796 step5: 4797 if (th->ack) 4798 tcp_ack(sk, skb, FLAG_SLOWPATH); 4799 4800 tcp_rcv_rtt_measure_ts(sk, skb); 4801 4802 /* Process urgent data. */ 4803 tcp_urg(sk, skb, th); 4804 4805 /* step 7: process the segment text */ 4806 tcp_data_queue(sk, skb); 4807 4808 tcp_data_snd_check(sk); 4809 tcp_ack_snd_check(sk); 4810 return 0; 4811 4812 csum_error: 4813 TCP_INC_STATS_BH(TCP_MIB_INERRS); 4814 4815 discard: 4816 __kfree_skb(skb); 4817 return 0; 4818 } 4819 4820 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb, 4821 struct tcphdr *th, unsigned len) 4822 { 4823 struct tcp_sock *tp = tcp_sk(sk); 4824 struct inet_connection_sock *icsk = inet_csk(sk); 4825 int saved_clamp = tp->rx_opt.mss_clamp; 4826 4827 tcp_parse_options(skb, &tp->rx_opt, 0); 4828 4829 if (th->ack) { 4830 /* rfc793: 4831 * "If the state is SYN-SENT then 4832 * first check the ACK bit 4833 * If the ACK bit is set 4834 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send 4835 * a reset (unless the RST bit is set, if so drop 4836 * the segment and return)" 4837 * 4838 * We do not send data with SYN, so that RFC-correct 4839 * test reduces to: 4840 */ 4841 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt) 4842 goto reset_and_undo; 4843 4844 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 4845 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp, 4846 tcp_time_stamp)) { 4847 NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED); 4848 goto reset_and_undo; 4849 } 4850 4851 /* Now ACK is acceptable. 4852 * 4853 * "If the RST bit is set 4854 * If the ACK was acceptable then signal the user "error: 4855 * connection reset", drop the segment, enter CLOSED state, 4856 * delete TCB, and return." 4857 */ 4858 4859 if (th->rst) { 4860 tcp_reset(sk); 4861 goto discard; 4862 } 4863 4864 /* rfc793: 4865 * "fifth, if neither of the SYN or RST bits is set then 4866 * drop the segment and return." 4867 * 4868 * See note below! 4869 * --ANK(990513) 4870 */ 4871 if (!th->syn) 4872 goto discard_and_undo; 4873 4874 /* rfc793: 4875 * "If the SYN bit is on ... 4876 * are acceptable then ... 4877 * (our SYN has been ACKed), change the connection 4878 * state to ESTABLISHED..." 4879 */ 4880 4881 TCP_ECN_rcv_synack(tp, th); 4882 4883 tp->snd_wl1 = TCP_SKB_CB(skb)->seq; 4884 tcp_ack(sk, skb, FLAG_SLOWPATH); 4885 4886 /* Ok.. it's good. Set up sequence numbers and 4887 * move to established. 4888 */ 4889 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; 4890 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; 4891 4892 /* RFC1323: The window in SYN & SYN/ACK segments is 4893 * never scaled. 4894 */ 4895 tp->snd_wnd = ntohs(th->window); 4896 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(skb)->seq); 4897 4898 if (!tp->rx_opt.wscale_ok) { 4899 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0; 4900 tp->window_clamp = min(tp->window_clamp, 65535U); 4901 } 4902 4903 if (tp->rx_opt.saw_tstamp) { 4904 tp->rx_opt.tstamp_ok = 1; 4905 tp->tcp_header_len = 4906 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 4907 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; 4908 tcp_store_ts_recent(tp); 4909 } else { 4910 tp->tcp_header_len = sizeof(struct tcphdr); 4911 } 4912 4913 if (tcp_is_sack(tp) && sysctl_tcp_fack) 4914 tcp_enable_fack(tp); 4915 4916 tcp_mtup_init(sk); 4917 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 4918 tcp_initialize_rcv_mss(sk); 4919 4920 /* Remember, tcp_poll() does not lock socket! 4921 * Change state from SYN-SENT only after copied_seq 4922 * is initialized. */ 4923 tp->copied_seq = tp->rcv_nxt; 4924 smp_mb(); 4925 tcp_set_state(sk, TCP_ESTABLISHED); 4926 4927 security_inet_conn_established(sk, skb); 4928 4929 /* Make sure socket is routed, for correct metrics. */ 4930 icsk->icsk_af_ops->rebuild_header(sk); 4931 4932 tcp_init_metrics(sk); 4933 4934 tcp_init_congestion_control(sk); 4935 4936 /* Prevent spurious tcp_cwnd_restart() on first data 4937 * packet. 4938 */ 4939 tp->lsndtime = tcp_time_stamp; 4940 4941 tcp_init_buffer_space(sk); 4942 4943 if (sock_flag(sk, SOCK_KEEPOPEN)) 4944 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp)); 4945 4946 if (!tp->rx_opt.snd_wscale) 4947 __tcp_fast_path_on(tp, tp->snd_wnd); 4948 else 4949 tp->pred_flags = 0; 4950 4951 if (!sock_flag(sk, SOCK_DEAD)) { 4952 sk->sk_state_change(sk); 4953 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT); 4954 } 4955 4956 if (sk->sk_write_pending || 4957 icsk->icsk_accept_queue.rskq_defer_accept || 4958 icsk->icsk_ack.pingpong) { 4959 /* Save one ACK. Data will be ready after 4960 * several ticks, if write_pending is set. 4961 * 4962 * It may be deleted, but with this feature tcpdumps 4963 * look so _wonderfully_ clever, that I was not able 4964 * to stand against the temptation 8) --ANK 4965 */ 4966 inet_csk_schedule_ack(sk); 4967 icsk->icsk_ack.lrcvtime = tcp_time_stamp; 4968 icsk->icsk_ack.ato = TCP_ATO_MIN; 4969 tcp_incr_quickack(sk); 4970 tcp_enter_quickack_mode(sk); 4971 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 4972 TCP_DELACK_MAX, TCP_RTO_MAX); 4973 4974 discard: 4975 __kfree_skb(skb); 4976 return 0; 4977 } else { 4978 tcp_send_ack(sk); 4979 } 4980 return -1; 4981 } 4982 4983 /* No ACK in the segment */ 4984 4985 if (th->rst) { 4986 /* rfc793: 4987 * "If the RST bit is set 4988 * 4989 * Otherwise (no ACK) drop the segment and return." 4990 */ 4991 4992 goto discard_and_undo; 4993 } 4994 4995 /* PAWS check. */ 4996 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && 4997 tcp_paws_check(&tp->rx_opt, 0)) 4998 goto discard_and_undo; 4999 5000 if (th->syn) { 5001 /* We see SYN without ACK. It is attempt of 5002 * simultaneous connect with crossed SYNs. 5003 * Particularly, it can be connect to self. 5004 */ 5005 tcp_set_state(sk, TCP_SYN_RECV); 5006 5007 if (tp->rx_opt.saw_tstamp) { 5008 tp->rx_opt.tstamp_ok = 1; 5009 tcp_store_ts_recent(tp); 5010 tp->tcp_header_len = 5011 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 5012 } else { 5013 tp->tcp_header_len = sizeof(struct tcphdr); 5014 } 5015 5016 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; 5017 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; 5018 5019 /* RFC1323: The window in SYN & SYN/ACK segments is 5020 * never scaled. 5021 */ 5022 tp->snd_wnd = ntohs(th->window); 5023 tp->snd_wl1 = TCP_SKB_CB(skb)->seq; 5024 tp->max_window = tp->snd_wnd; 5025 5026 TCP_ECN_rcv_syn(tp, th); 5027 5028 tcp_mtup_init(sk); 5029 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 5030 tcp_initialize_rcv_mss(sk); 5031 5032 tcp_send_synack(sk); 5033 #if 0 5034 /* Note, we could accept data and URG from this segment. 5035 * There are no obstacles to make this. 5036 * 5037 * However, if we ignore data in ACKless segments sometimes, 5038 * we have no reasons to accept it sometimes. 5039 * Also, seems the code doing it in step6 of tcp_rcv_state_process 5040 * is not flawless. So, discard packet for sanity. 5041 * Uncomment this return to process the data. 5042 */ 5043 return -1; 5044 #else 5045 goto discard; 5046 #endif 5047 } 5048 /* "fifth, if neither of the SYN or RST bits is set then 5049 * drop the segment and return." 5050 */ 5051 5052 discard_and_undo: 5053 tcp_clear_options(&tp->rx_opt); 5054 tp->rx_opt.mss_clamp = saved_clamp; 5055 goto discard; 5056 5057 reset_and_undo: 5058 tcp_clear_options(&tp->rx_opt); 5059 tp->rx_opt.mss_clamp = saved_clamp; 5060 return 1; 5061 } 5062 5063 /* 5064 * This function implements the receiving procedure of RFC 793 for 5065 * all states except ESTABLISHED and TIME_WAIT. 5066 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be 5067 * address independent. 5068 */ 5069 5070 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb, 5071 struct tcphdr *th, unsigned len) 5072 { 5073 struct tcp_sock *tp = tcp_sk(sk); 5074 struct inet_connection_sock *icsk = inet_csk(sk); 5075 int queued = 0; 5076 5077 tp->rx_opt.saw_tstamp = 0; 5078 5079 switch (sk->sk_state) { 5080 case TCP_CLOSE: 5081 goto discard; 5082 5083 case TCP_LISTEN: 5084 if (th->ack) 5085 return 1; 5086 5087 if (th->rst) 5088 goto discard; 5089 5090 if (th->syn) { 5091 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0) 5092 return 1; 5093 5094 /* Now we have several options: In theory there is 5095 * nothing else in the frame. KA9Q has an option to 5096 * send data with the syn, BSD accepts data with the 5097 * syn up to the [to be] advertised window and 5098 * Solaris 2.1 gives you a protocol error. For now 5099 * we just ignore it, that fits the spec precisely 5100 * and avoids incompatibilities. It would be nice in 5101 * future to drop through and process the data. 5102 * 5103 * Now that TTCP is starting to be used we ought to 5104 * queue this data. 5105 * But, this leaves one open to an easy denial of 5106 * service attack, and SYN cookies can't defend 5107 * against this problem. So, we drop the data 5108 * in the interest of security over speed unless 5109 * it's still in use. 5110 */ 5111 kfree_skb(skb); 5112 return 0; 5113 } 5114 goto discard; 5115 5116 case TCP_SYN_SENT: 5117 queued = tcp_rcv_synsent_state_process(sk, skb, th, len); 5118 if (queued >= 0) 5119 return queued; 5120 5121 /* Do step6 onward by hand. */ 5122 tcp_urg(sk, skb, th); 5123 __kfree_skb(skb); 5124 tcp_data_snd_check(sk); 5125 return 0; 5126 } 5127 5128 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp && 5129 tcp_paws_discard(sk, skb)) { 5130 if (!th->rst) { 5131 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED); 5132 tcp_send_dupack(sk, skb); 5133 goto discard; 5134 } 5135 /* Reset is accepted even if it did not pass PAWS. */ 5136 } 5137 5138 /* step 1: check sequence number */ 5139 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) { 5140 if (!th->rst) 5141 tcp_send_dupack(sk, skb); 5142 goto discard; 5143 } 5144 5145 /* step 2: check RST bit */ 5146 if (th->rst) { 5147 tcp_reset(sk); 5148 goto discard; 5149 } 5150 5151 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq); 5152 5153 /* step 3: check security and precedence [ignored] */ 5154 5155 /* step 4: 5156 * 5157 * Check for a SYN in window. 5158 */ 5159 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 5160 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN); 5161 tcp_reset(sk); 5162 return 1; 5163 } 5164 5165 /* step 5: check the ACK field */ 5166 if (th->ack) { 5167 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH); 5168 5169 switch (sk->sk_state) { 5170 case TCP_SYN_RECV: 5171 if (acceptable) { 5172 tp->copied_seq = tp->rcv_nxt; 5173 smp_mb(); 5174 tcp_set_state(sk, TCP_ESTABLISHED); 5175 sk->sk_state_change(sk); 5176 5177 /* Note, that this wakeup is only for marginal 5178 * crossed SYN case. Passively open sockets 5179 * are not waked up, because sk->sk_sleep == 5180 * NULL and sk->sk_socket == NULL. 5181 */ 5182 if (sk->sk_socket) 5183 sk_wake_async(sk, 5184 SOCK_WAKE_IO, POLL_OUT); 5185 5186 tp->snd_una = TCP_SKB_CB(skb)->ack_seq; 5187 tp->snd_wnd = ntohs(th->window) << 5188 tp->rx_opt.snd_wscale; 5189 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, 5190 TCP_SKB_CB(skb)->seq); 5191 5192 /* tcp_ack considers this ACK as duplicate 5193 * and does not calculate rtt. 5194 * Fix it at least with timestamps. 5195 */ 5196 if (tp->rx_opt.saw_tstamp && 5197 tp->rx_opt.rcv_tsecr && !tp->srtt) 5198 tcp_ack_saw_tstamp(sk, 0); 5199 5200 if (tp->rx_opt.tstamp_ok) 5201 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; 5202 5203 /* Make sure socket is routed, for 5204 * correct metrics. 5205 */ 5206 icsk->icsk_af_ops->rebuild_header(sk); 5207 5208 tcp_init_metrics(sk); 5209 5210 tcp_init_congestion_control(sk); 5211 5212 /* Prevent spurious tcp_cwnd_restart() on 5213 * first data packet. 5214 */ 5215 tp->lsndtime = tcp_time_stamp; 5216 5217 tcp_mtup_init(sk); 5218 tcp_initialize_rcv_mss(sk); 5219 tcp_init_buffer_space(sk); 5220 tcp_fast_path_on(tp); 5221 } else { 5222 return 1; 5223 } 5224 break; 5225 5226 case TCP_FIN_WAIT1: 5227 if (tp->snd_una == tp->write_seq) { 5228 tcp_set_state(sk, TCP_FIN_WAIT2); 5229 sk->sk_shutdown |= SEND_SHUTDOWN; 5230 dst_confirm(sk->sk_dst_cache); 5231 5232 if (!sock_flag(sk, SOCK_DEAD)) 5233 /* Wake up lingering close() */ 5234 sk->sk_state_change(sk); 5235 else { 5236 int tmo; 5237 5238 if (tp->linger2 < 0 || 5239 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 5240 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) { 5241 tcp_done(sk); 5242 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA); 5243 return 1; 5244 } 5245 5246 tmo = tcp_fin_time(sk); 5247 if (tmo > TCP_TIMEWAIT_LEN) { 5248 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN); 5249 } else if (th->fin || sock_owned_by_user(sk)) { 5250 /* Bad case. We could lose such FIN otherwise. 5251 * It is not a big problem, but it looks confusing 5252 * and not so rare event. We still can lose it now, 5253 * if it spins in bh_lock_sock(), but it is really 5254 * marginal case. 5255 */ 5256 inet_csk_reset_keepalive_timer(sk, tmo); 5257 } else { 5258 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo); 5259 goto discard; 5260 } 5261 } 5262 } 5263 break; 5264 5265 case TCP_CLOSING: 5266 if (tp->snd_una == tp->write_seq) { 5267 tcp_time_wait(sk, TCP_TIME_WAIT, 0); 5268 goto discard; 5269 } 5270 break; 5271 5272 case TCP_LAST_ACK: 5273 if (tp->snd_una == tp->write_seq) { 5274 tcp_update_metrics(sk); 5275 tcp_done(sk); 5276 goto discard; 5277 } 5278 break; 5279 } 5280 } else 5281 goto discard; 5282 5283 /* step 6: check the URG bit */ 5284 tcp_urg(sk, skb, th); 5285 5286 /* step 7: process the segment text */ 5287 switch (sk->sk_state) { 5288 case TCP_CLOSE_WAIT: 5289 case TCP_CLOSING: 5290 case TCP_LAST_ACK: 5291 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) 5292 break; 5293 case TCP_FIN_WAIT1: 5294 case TCP_FIN_WAIT2: 5295 /* RFC 793 says to queue data in these states, 5296 * RFC 1122 says we MUST send a reset. 5297 * BSD 4.4 also does reset. 5298 */ 5299 if (sk->sk_shutdown & RCV_SHUTDOWN) { 5300 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 5301 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) { 5302 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA); 5303 tcp_reset(sk); 5304 return 1; 5305 } 5306 } 5307 /* Fall through */ 5308 case TCP_ESTABLISHED: 5309 tcp_data_queue(sk, skb); 5310 queued = 1; 5311 break; 5312 } 5313 5314 /* tcp_data could move socket to TIME-WAIT */ 5315 if (sk->sk_state != TCP_CLOSE) { 5316 tcp_data_snd_check(sk); 5317 tcp_ack_snd_check(sk); 5318 } 5319 5320 if (!queued) { 5321 discard: 5322 __kfree_skb(skb); 5323 } 5324 return 0; 5325 } 5326 5327 EXPORT_SYMBOL(sysctl_tcp_ecn); 5328 EXPORT_SYMBOL(sysctl_tcp_reordering); 5329 EXPORT_SYMBOL(tcp_parse_options); 5330 EXPORT_SYMBOL(tcp_rcv_established); 5331 EXPORT_SYMBOL(tcp_rcv_state_process); 5332 EXPORT_SYMBOL(tcp_initialize_rcv_mss); 5333