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