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