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/dst.h> 70 #include <net/tcp.h> 71 #include <net/inet_common.h> 72 #include <linux/ipsec.h> 73 #include <asm/unaligned.h> 74 #include <net/netdma.h> 75 76 int sysctl_tcp_timestamps __read_mostly = 1; 77 int sysctl_tcp_window_scaling __read_mostly = 1; 78 int sysctl_tcp_sack __read_mostly = 1; 79 int sysctl_tcp_fack __read_mostly = 1; 80 int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH; 81 int sysctl_tcp_ecn __read_mostly; 82 int sysctl_tcp_dsack __read_mostly = 1; 83 int sysctl_tcp_app_win __read_mostly = 31; 84 int sysctl_tcp_adv_win_scale __read_mostly = 2; 85 86 int sysctl_tcp_stdurg __read_mostly; 87 int sysctl_tcp_rfc1337 __read_mostly; 88 int sysctl_tcp_max_orphans __read_mostly = NR_FILE; 89 int sysctl_tcp_frto __read_mostly = 2; 90 int sysctl_tcp_frto_response __read_mostly; 91 int sysctl_tcp_nometrics_save __read_mostly; 92 93 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1; 94 int sysctl_tcp_abc __read_mostly; 95 96 #define FLAG_DATA 0x01 /* Incoming frame contained data. */ 97 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */ 98 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */ 99 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */ 100 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */ 101 #define FLAG_DATA_SACKED 0x20 /* New SACK. */ 102 #define FLAG_ECE 0x40 /* ECE in this ACK */ 103 #define FLAG_DATA_LOST 0x80 /* SACK detected data lossage. */ 104 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/ 105 #define FLAG_ONLY_ORIG_SACKED 0x200 /* SACKs only non-rexmit sent before RTO */ 106 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */ 107 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */ 108 #define FLAG_NONHEAD_RETRANS_ACKED 0x1000 /* Non-head rexmitted data was ACKed */ 109 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */ 110 111 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED) 112 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED) 113 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE) 114 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED) 115 #define FLAG_ANY_PROGRESS (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED) 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, const struct sk_buff *skb) 124 { 125 struct inet_connection_sock *icsk = inet_csk(sk); 126 const unsigned int lss = icsk->icsk_ack.last_seg_size; 127 unsigned int len; 128 129 icsk->icsk_ack.last_seg_size = 0; 130 131 /* skb->len may jitter because of SACKs, even if peer 132 * sends good full-sized frames. 133 */ 134 len = skb_shinfo(skb)->gso_size ? : skb->len; 135 if (len >= icsk->icsk_ack.rcv_mss) { 136 icsk->icsk_ack.rcv_mss = len; 137 } else { 138 /* Otherwise, we make more careful check taking into account, 139 * that SACKs block is variable. 140 * 141 * "len" is invariant segment length, including TCP header. 142 */ 143 len += skb->data - skb_transport_header(skb); 144 if (len >= TCP_MIN_RCVMSS + sizeof(struct tcphdr) || 145 /* If PSH is not set, packet should be 146 * full sized, provided peer TCP is not badly broken. 147 * This observation (if it is correct 8)) allows 148 * to handle super-low mtu links fairly. 149 */ 150 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) && 151 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) { 152 /* Subtract also invariant (if peer is RFC compliant), 153 * tcp header plus fixed timestamp option length. 154 * Resulting "len" is MSS free of SACK jitter. 155 */ 156 len -= tcp_sk(sk)->tcp_header_len; 157 icsk->icsk_ack.last_seg_size = len; 158 if (len == lss) { 159 icsk->icsk_ack.rcv_mss = len; 160 return; 161 } 162 } 163 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED) 164 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2; 165 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED; 166 } 167 } 168 169 static void tcp_incr_quickack(struct sock *sk) 170 { 171 struct inet_connection_sock *icsk = inet_csk(sk); 172 unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss); 173 174 if (quickacks == 0) 175 quickacks = 2; 176 if (quickacks > icsk->icsk_ack.quick) 177 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS); 178 } 179 180 void tcp_enter_quickack_mode(struct sock *sk) 181 { 182 struct inet_connection_sock *icsk = inet_csk(sk); 183 tcp_incr_quickack(sk); 184 icsk->icsk_ack.pingpong = 0; 185 icsk->icsk_ack.ato = TCP_ATO_MIN; 186 } 187 188 /* Send ACKs quickly, if "quick" count is not exhausted 189 * and the session is not interactive. 190 */ 191 192 static inline int tcp_in_quickack_mode(const struct sock *sk) 193 { 194 const struct inet_connection_sock *icsk = inet_csk(sk); 195 return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong; 196 } 197 198 static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp) 199 { 200 if (tp->ecn_flags & TCP_ECN_OK) 201 tp->ecn_flags |= TCP_ECN_QUEUE_CWR; 202 } 203 204 static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, struct sk_buff *skb) 205 { 206 if (tcp_hdr(skb)->cwr) 207 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR; 208 } 209 210 static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp) 211 { 212 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR; 213 } 214 215 static inline void TCP_ECN_check_ce(struct tcp_sock *tp, struct sk_buff *skb) 216 { 217 if (tp->ecn_flags & TCP_ECN_OK) { 218 if (INET_ECN_is_ce(TCP_SKB_CB(skb)->flags)) 219 tp->ecn_flags |= TCP_ECN_DEMAND_CWR; 220 /* Funny extension: if ECT is not set on a segment, 221 * it is surely retransmit. It is not in ECN RFC, 222 * but Linux follows this rule. */ 223 else if (INET_ECN_is_not_ect((TCP_SKB_CB(skb)->flags))) 224 tcp_enter_quickack_mode((struct sock *)tp); 225 } 226 } 227 228 static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, struct tcphdr *th) 229 { 230 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr)) 231 tp->ecn_flags &= ~TCP_ECN_OK; 232 } 233 234 static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, struct tcphdr *th) 235 { 236 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr)) 237 tp->ecn_flags &= ~TCP_ECN_OK; 238 } 239 240 static inline int TCP_ECN_rcv_ecn_echo(struct tcp_sock *tp, struct tcphdr *th) 241 { 242 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK)) 243 return 1; 244 return 0; 245 } 246 247 /* Buffer size and advertised window tuning. 248 * 249 * 1. Tuning sk->sk_sndbuf, when connection enters established state. 250 */ 251 252 static void tcp_fixup_sndbuf(struct sock *sk) 253 { 254 int sndmem = tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER + 16 + 255 sizeof(struct sk_buff); 256 257 if (sk->sk_sndbuf < 3 * sndmem) 258 sk->sk_sndbuf = min(3 * sndmem, sysctl_tcp_wmem[2]); 259 } 260 261 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh) 262 * 263 * All tcp_full_space() is split to two parts: "network" buffer, allocated 264 * forward and advertised in receiver window (tp->rcv_wnd) and 265 * "application buffer", required to isolate scheduling/application 266 * latencies from network. 267 * window_clamp is maximal advertised window. It can be less than 268 * tcp_full_space(), in this case tcp_full_space() - window_clamp 269 * is reserved for "application" buffer. The less window_clamp is 270 * the smoother our behaviour from viewpoint of network, but the lower 271 * throughput and the higher sensitivity of the connection to losses. 8) 272 * 273 * rcv_ssthresh is more strict window_clamp used at "slow start" 274 * phase to predict further behaviour of this connection. 275 * It is used for two goals: 276 * - to enforce header prediction at sender, even when application 277 * requires some significant "application buffer". It is check #1. 278 * - to prevent pruning of receive queue because of misprediction 279 * of receiver window. Check #2. 280 * 281 * The scheme does not work when sender sends good segments opening 282 * window and then starts to feed us spaghetti. But it should work 283 * in common situations. Otherwise, we have to rely on queue collapsing. 284 */ 285 286 /* Slow part of check#2. */ 287 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb) 288 { 289 struct tcp_sock *tp = tcp_sk(sk); 290 /* Optimize this! */ 291 int truesize = tcp_win_from_space(skb->truesize) >> 1; 292 int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1; 293 294 while (tp->rcv_ssthresh <= window) { 295 if (truesize <= skb->len) 296 return 2 * inet_csk(sk)->icsk_ack.rcv_mss; 297 298 truesize >>= 1; 299 window >>= 1; 300 } 301 return 0; 302 } 303 304 static void tcp_grow_window(struct sock *sk, struct sk_buff *skb) 305 { 306 struct tcp_sock *tp = tcp_sk(sk); 307 308 /* Check #1 */ 309 if (tp->rcv_ssthresh < tp->window_clamp && 310 (int)tp->rcv_ssthresh < tcp_space(sk) && 311 !tcp_memory_pressure) { 312 int incr; 313 314 /* Check #2. Increase window, if skb with such overhead 315 * will fit to rcvbuf in future. 316 */ 317 if (tcp_win_from_space(skb->truesize) <= skb->len) 318 incr = 2 * tp->advmss; 319 else 320 incr = __tcp_grow_window(sk, skb); 321 322 if (incr) { 323 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr, 324 tp->window_clamp); 325 inet_csk(sk)->icsk_ack.quick |= 1; 326 } 327 } 328 } 329 330 /* 3. Tuning rcvbuf, when connection enters established state. */ 331 332 static void tcp_fixup_rcvbuf(struct sock *sk) 333 { 334 struct tcp_sock *tp = tcp_sk(sk); 335 int rcvmem = tp->advmss + MAX_TCP_HEADER + 16 + sizeof(struct sk_buff); 336 337 /* Try to select rcvbuf so that 4 mss-sized segments 338 * will fit to window and corresponding skbs will fit to our rcvbuf. 339 * (was 3; 4 is minimum to allow fast retransmit to work.) 340 */ 341 while (tcp_win_from_space(rcvmem) < tp->advmss) 342 rcvmem += 128; 343 if (sk->sk_rcvbuf < 4 * rcvmem) 344 sk->sk_rcvbuf = min(4 * rcvmem, sysctl_tcp_rmem[2]); 345 } 346 347 /* 4. Try to fixup all. It is made immediately after connection enters 348 * established state. 349 */ 350 static void tcp_init_buffer_space(struct sock *sk) 351 { 352 struct tcp_sock *tp = tcp_sk(sk); 353 int maxwin; 354 355 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) 356 tcp_fixup_rcvbuf(sk); 357 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK)) 358 tcp_fixup_sndbuf(sk); 359 360 tp->rcvq_space.space = tp->rcv_wnd; 361 362 maxwin = tcp_full_space(sk); 363 364 if (tp->window_clamp >= maxwin) { 365 tp->window_clamp = maxwin; 366 367 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss) 368 tp->window_clamp = max(maxwin - 369 (maxwin >> sysctl_tcp_app_win), 370 4 * tp->advmss); 371 } 372 373 /* Force reservation of one segment. */ 374 if (sysctl_tcp_app_win && 375 tp->window_clamp > 2 * tp->advmss && 376 tp->window_clamp + tp->advmss > maxwin) 377 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss); 378 379 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp); 380 tp->snd_cwnd_stamp = tcp_time_stamp; 381 } 382 383 /* 5. Recalculate window clamp after socket hit its memory bounds. */ 384 static void tcp_clamp_window(struct sock *sk) 385 { 386 struct tcp_sock *tp = tcp_sk(sk); 387 struct inet_connection_sock *icsk = inet_csk(sk); 388 389 icsk->icsk_ack.quick = 0; 390 391 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] && 392 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) && 393 !tcp_memory_pressure && 394 atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) { 395 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc), 396 sysctl_tcp_rmem[2]); 397 } 398 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf) 399 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss); 400 } 401 402 /* Initialize RCV_MSS value. 403 * RCV_MSS is an our guess about MSS used by the peer. 404 * We haven't any direct information about the MSS. 405 * It's better to underestimate the RCV_MSS rather than overestimate. 406 * Overestimations make us ACKing less frequently than needed. 407 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss(). 408 */ 409 void tcp_initialize_rcv_mss(struct sock *sk) 410 { 411 struct tcp_sock *tp = tcp_sk(sk); 412 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache); 413 414 hint = min(hint, tp->rcv_wnd / 2); 415 hint = min(hint, TCP_MIN_RCVMSS); 416 hint = max(hint, TCP_MIN_MSS); 417 418 inet_csk(sk)->icsk_ack.rcv_mss = hint; 419 } 420 421 /* Receiver "autotuning" code. 422 * 423 * The algorithm for RTT estimation w/o timestamps is based on 424 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL. 425 * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps> 426 * 427 * More detail on this code can be found at 428 * <http://www.psc.edu/~jheffner/senior_thesis.ps>, 429 * though this reference is out of date. A new paper 430 * is pending. 431 */ 432 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep) 433 { 434 u32 new_sample = tp->rcv_rtt_est.rtt; 435 long m = sample; 436 437 if (m == 0) 438 m = 1; 439 440 if (new_sample != 0) { 441 /* If we sample in larger samples in the non-timestamp 442 * case, we could grossly overestimate the RTT especially 443 * with chatty applications or bulk transfer apps which 444 * are stalled on filesystem I/O. 445 * 446 * Also, since we are only going for a minimum in the 447 * non-timestamp case, we do not smooth things out 448 * else with timestamps disabled convergence takes too 449 * long. 450 */ 451 if (!win_dep) { 452 m -= (new_sample >> 3); 453 new_sample += m; 454 } else if (m < new_sample) 455 new_sample = m << 3; 456 } else { 457 /* No previous measure. */ 458 new_sample = m << 3; 459 } 460 461 if (tp->rcv_rtt_est.rtt != new_sample) 462 tp->rcv_rtt_est.rtt = new_sample; 463 } 464 465 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp) 466 { 467 if (tp->rcv_rtt_est.time == 0) 468 goto new_measure; 469 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq)) 470 return; 471 tcp_rcv_rtt_update(tp, jiffies - tp->rcv_rtt_est.time, 1); 472 473 new_measure: 474 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd; 475 tp->rcv_rtt_est.time = tcp_time_stamp; 476 } 477 478 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, 479 const struct sk_buff *skb) 480 { 481 struct tcp_sock *tp = tcp_sk(sk); 482 if (tp->rx_opt.rcv_tsecr && 483 (TCP_SKB_CB(skb)->end_seq - 484 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss)) 485 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0); 486 } 487 488 /* 489 * This function should be called every time data is copied to user space. 490 * It calculates the appropriate TCP receive buffer space. 491 */ 492 void tcp_rcv_space_adjust(struct sock *sk) 493 { 494 struct tcp_sock *tp = tcp_sk(sk); 495 int time; 496 int space; 497 498 if (tp->rcvq_space.time == 0) 499 goto new_measure; 500 501 time = tcp_time_stamp - tp->rcvq_space.time; 502 if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0) 503 return; 504 505 space = 2 * (tp->copied_seq - tp->rcvq_space.seq); 506 507 space = max(tp->rcvq_space.space, space); 508 509 if (tp->rcvq_space.space != space) { 510 int rcvmem; 511 512 tp->rcvq_space.space = space; 513 514 if (sysctl_tcp_moderate_rcvbuf && 515 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) { 516 int new_clamp = space; 517 518 /* Receive space grows, normalize in order to 519 * take into account packet headers and sk_buff 520 * structure overhead. 521 */ 522 space /= tp->advmss; 523 if (!space) 524 space = 1; 525 rcvmem = (tp->advmss + MAX_TCP_HEADER + 526 16 + sizeof(struct sk_buff)); 527 while (tcp_win_from_space(rcvmem) < tp->advmss) 528 rcvmem += 128; 529 space *= rcvmem; 530 space = min(space, sysctl_tcp_rmem[2]); 531 if (space > sk->sk_rcvbuf) { 532 sk->sk_rcvbuf = space; 533 534 /* Make the window clamp follow along. */ 535 tp->window_clamp = new_clamp; 536 } 537 } 538 } 539 540 new_measure: 541 tp->rcvq_space.seq = tp->copied_seq; 542 tp->rcvq_space.time = tcp_time_stamp; 543 } 544 545 /* There is something which you must keep in mind when you analyze the 546 * behavior of the tp->ato delayed ack timeout interval. When a 547 * connection starts up, we want to ack as quickly as possible. The 548 * problem is that "good" TCP's do slow start at the beginning of data 549 * transmission. The means that until we send the first few ACK's the 550 * sender will sit on his end and only queue most of his data, because 551 * he can only send snd_cwnd unacked packets at any given time. For 552 * each ACK we send, he increments snd_cwnd and transmits more of his 553 * queue. -DaveM 554 */ 555 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb) 556 { 557 struct tcp_sock *tp = tcp_sk(sk); 558 struct inet_connection_sock *icsk = inet_csk(sk); 559 u32 now; 560 561 inet_csk_schedule_ack(sk); 562 563 tcp_measure_rcv_mss(sk, skb); 564 565 tcp_rcv_rtt_measure(tp); 566 567 now = tcp_time_stamp; 568 569 if (!icsk->icsk_ack.ato) { 570 /* The _first_ data packet received, initialize 571 * delayed ACK engine. 572 */ 573 tcp_incr_quickack(sk); 574 icsk->icsk_ack.ato = TCP_ATO_MIN; 575 } else { 576 int m = now - icsk->icsk_ack.lrcvtime; 577 578 if (m <= TCP_ATO_MIN / 2) { 579 /* The fastest case is the first. */ 580 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2; 581 } else if (m < icsk->icsk_ack.ato) { 582 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m; 583 if (icsk->icsk_ack.ato > icsk->icsk_rto) 584 icsk->icsk_ack.ato = icsk->icsk_rto; 585 } else if (m > icsk->icsk_rto) { 586 /* Too long gap. Apparently sender failed to 587 * restart window, so that we send ACKs quickly. 588 */ 589 tcp_incr_quickack(sk); 590 sk_mem_reclaim(sk); 591 } 592 } 593 icsk->icsk_ack.lrcvtime = now; 594 595 TCP_ECN_check_ce(tp, skb); 596 597 if (skb->len >= 128) 598 tcp_grow_window(sk, skb); 599 } 600 601 static u32 tcp_rto_min(struct sock *sk) 602 { 603 struct dst_entry *dst = __sk_dst_get(sk); 604 u32 rto_min = TCP_RTO_MIN; 605 606 if (dst && dst_metric_locked(dst, RTAX_RTO_MIN)) 607 rto_min = dst_metric(dst, RTAX_RTO_MIN); 608 return rto_min; 609 } 610 611 /* Called to compute a smoothed rtt estimate. The data fed to this 612 * routine either comes from timestamps, or from segments that were 613 * known _not_ to have been retransmitted [see Karn/Partridge 614 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88 615 * piece by Van Jacobson. 616 * NOTE: the next three routines used to be one big routine. 617 * To save cycles in the RFC 1323 implementation it was better to break 618 * it up into three procedures. -- erics 619 */ 620 static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt) 621 { 622 struct tcp_sock *tp = tcp_sk(sk); 623 long m = mrtt; /* RTT */ 624 625 /* The following amusing code comes from Jacobson's 626 * article in SIGCOMM '88. Note that rtt and mdev 627 * are scaled versions of rtt and mean deviation. 628 * This is designed to be as fast as possible 629 * m stands for "measurement". 630 * 631 * On a 1990 paper the rto value is changed to: 632 * RTO = rtt + 4 * mdev 633 * 634 * Funny. This algorithm seems to be very broken. 635 * These formulae increase RTO, when it should be decreased, increase 636 * too slowly, when it should be increased quickly, decrease too quickly 637 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely 638 * does not matter how to _calculate_ it. Seems, it was trap 639 * that VJ failed to avoid. 8) 640 */ 641 if (m == 0) 642 m = 1; 643 if (tp->srtt != 0) { 644 m -= (tp->srtt >> 3); /* m is now error in rtt est */ 645 tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */ 646 if (m < 0) { 647 m = -m; /* m is now abs(error) */ 648 m -= (tp->mdev >> 2); /* similar update on mdev */ 649 /* This is similar to one of Eifel findings. 650 * Eifel blocks mdev updates when rtt decreases. 651 * This solution is a bit different: we use finer gain 652 * for mdev in this case (alpha*beta). 653 * Like Eifel it also prevents growth of rto, 654 * but also it limits too fast rto decreases, 655 * happening in pure Eifel. 656 */ 657 if (m > 0) 658 m >>= 3; 659 } else { 660 m -= (tp->mdev >> 2); /* similar update on mdev */ 661 } 662 tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */ 663 if (tp->mdev > tp->mdev_max) { 664 tp->mdev_max = tp->mdev; 665 if (tp->mdev_max > tp->rttvar) 666 tp->rttvar = tp->mdev_max; 667 } 668 if (after(tp->snd_una, tp->rtt_seq)) { 669 if (tp->mdev_max < tp->rttvar) 670 tp->rttvar -= (tp->rttvar - tp->mdev_max) >> 2; 671 tp->rtt_seq = tp->snd_nxt; 672 tp->mdev_max = tcp_rto_min(sk); 673 } 674 } else { 675 /* no previous measure. */ 676 tp->srtt = m << 3; /* take the measured time to be rtt */ 677 tp->mdev = m << 1; /* make sure rto = 3*rtt */ 678 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk)); 679 tp->rtt_seq = tp->snd_nxt; 680 } 681 } 682 683 /* Calculate rto without backoff. This is the second half of Van Jacobson's 684 * routine referred to above. 685 */ 686 static inline void tcp_set_rto(struct sock *sk) 687 { 688 const struct tcp_sock *tp = tcp_sk(sk); 689 /* Old crap is replaced with new one. 8) 690 * 691 * More seriously: 692 * 1. If rtt variance happened to be less 50msec, it is hallucination. 693 * It cannot be less due to utterly erratic ACK generation made 694 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_ 695 * to do with delayed acks, because at cwnd>2 true delack timeout 696 * is invisible. Actually, Linux-2.4 also generates erratic 697 * ACKs in some circumstances. 698 */ 699 inet_csk(sk)->icsk_rto = (tp->srtt >> 3) + tp->rttvar; 700 701 /* 2. Fixups made earlier cannot be right. 702 * If we do not estimate RTO correctly without them, 703 * all the algo is pure shit and should be replaced 704 * with correct one. It is exactly, which we pretend to do. 705 */ 706 } 707 708 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo 709 * guarantees that rto is higher. 710 */ 711 static inline void tcp_bound_rto(struct sock *sk) 712 { 713 if (inet_csk(sk)->icsk_rto > TCP_RTO_MAX) 714 inet_csk(sk)->icsk_rto = TCP_RTO_MAX; 715 } 716 717 /* Save metrics learned by this TCP session. 718 This function is called only, when TCP finishes successfully 719 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE. 720 */ 721 void tcp_update_metrics(struct sock *sk) 722 { 723 struct tcp_sock *tp = tcp_sk(sk); 724 struct dst_entry *dst = __sk_dst_get(sk); 725 726 if (sysctl_tcp_nometrics_save) 727 return; 728 729 dst_confirm(dst); 730 731 if (dst && (dst->flags & DST_HOST)) { 732 const struct inet_connection_sock *icsk = inet_csk(sk); 733 int m; 734 735 if (icsk->icsk_backoff || !tp->srtt) { 736 /* This session failed to estimate rtt. Why? 737 * Probably, no packets returned in time. 738 * Reset our results. 739 */ 740 if (!(dst_metric_locked(dst, RTAX_RTT))) 741 dst->metrics[RTAX_RTT - 1] = 0; 742 return; 743 } 744 745 m = dst_metric(dst, RTAX_RTT) - tp->srtt; 746 747 /* If newly calculated rtt larger than stored one, 748 * store new one. Otherwise, use EWMA. Remember, 749 * rtt overestimation is always better than underestimation. 750 */ 751 if (!(dst_metric_locked(dst, RTAX_RTT))) { 752 if (m <= 0) 753 dst->metrics[RTAX_RTT - 1] = tp->srtt; 754 else 755 dst->metrics[RTAX_RTT - 1] -= (m >> 3); 756 } 757 758 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) { 759 if (m < 0) 760 m = -m; 761 762 /* Scale deviation to rttvar fixed point */ 763 m >>= 1; 764 if (m < tp->mdev) 765 m = tp->mdev; 766 767 if (m >= dst_metric(dst, RTAX_RTTVAR)) 768 dst->metrics[RTAX_RTTVAR - 1] = m; 769 else 770 dst->metrics[RTAX_RTTVAR-1] -= 771 (dst_metric(dst, RTAX_RTTVAR) - m)>>2; 772 } 773 774 if (tp->snd_ssthresh >= 0xFFFF) { 775 /* Slow start still did not finish. */ 776 if (dst_metric(dst, RTAX_SSTHRESH) && 777 !dst_metric_locked(dst, RTAX_SSTHRESH) && 778 (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH)) 779 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1; 780 if (!dst_metric_locked(dst, RTAX_CWND) && 781 tp->snd_cwnd > dst_metric(dst, RTAX_CWND)) 782 dst->metrics[RTAX_CWND - 1] = tp->snd_cwnd; 783 } else if (tp->snd_cwnd > tp->snd_ssthresh && 784 icsk->icsk_ca_state == TCP_CA_Open) { 785 /* Cong. avoidance phase, cwnd is reliable. */ 786 if (!dst_metric_locked(dst, RTAX_SSTHRESH)) 787 dst->metrics[RTAX_SSTHRESH-1] = 788 max(tp->snd_cwnd >> 1, tp->snd_ssthresh); 789 if (!dst_metric_locked(dst, RTAX_CWND)) 790 dst->metrics[RTAX_CWND-1] = (dst_metric(dst, RTAX_CWND) + tp->snd_cwnd) >> 1; 791 } else { 792 /* Else slow start did not finish, cwnd is non-sense, 793 ssthresh may be also invalid. 794 */ 795 if (!dst_metric_locked(dst, RTAX_CWND)) 796 dst->metrics[RTAX_CWND-1] = (dst_metric(dst, RTAX_CWND) + tp->snd_ssthresh) >> 1; 797 if (dst_metric(dst, RTAX_SSTHRESH) && 798 !dst_metric_locked(dst, RTAX_SSTHRESH) && 799 tp->snd_ssthresh > dst_metric(dst, RTAX_SSTHRESH)) 800 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh; 801 } 802 803 if (!dst_metric_locked(dst, RTAX_REORDERING)) { 804 if (dst_metric(dst, RTAX_REORDERING) < tp->reordering && 805 tp->reordering != sysctl_tcp_reordering) 806 dst->metrics[RTAX_REORDERING-1] = tp->reordering; 807 } 808 } 809 } 810 811 /* Numbers are taken from RFC3390. 812 * 813 * John Heffner states: 814 * 815 * The RFC specifies a window of no more than 4380 bytes 816 * unless 2*MSS > 4380. Reading the pseudocode in the RFC 817 * is a bit misleading because they use a clamp at 4380 bytes 818 * rather than use a multiplier in the relevant range. 819 */ 820 __u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst) 821 { 822 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0); 823 824 if (!cwnd) { 825 if (tp->mss_cache > 1460) 826 cwnd = 2; 827 else 828 cwnd = (tp->mss_cache > 1095) ? 3 : 4; 829 } 830 return min_t(__u32, cwnd, tp->snd_cwnd_clamp); 831 } 832 833 /* Set slow start threshold and cwnd not falling to slow start */ 834 void tcp_enter_cwr(struct sock *sk, const int set_ssthresh) 835 { 836 struct tcp_sock *tp = tcp_sk(sk); 837 const struct inet_connection_sock *icsk = inet_csk(sk); 838 839 tp->prior_ssthresh = 0; 840 tp->bytes_acked = 0; 841 if (icsk->icsk_ca_state < TCP_CA_CWR) { 842 tp->undo_marker = 0; 843 if (set_ssthresh) 844 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 845 tp->snd_cwnd = min(tp->snd_cwnd, 846 tcp_packets_in_flight(tp) + 1U); 847 tp->snd_cwnd_cnt = 0; 848 tp->high_seq = tp->snd_nxt; 849 tp->snd_cwnd_stamp = tcp_time_stamp; 850 TCP_ECN_queue_cwr(tp); 851 852 tcp_set_ca_state(sk, TCP_CA_CWR); 853 } 854 } 855 856 /* 857 * Packet counting of FACK is based on in-order assumptions, therefore TCP 858 * disables it when reordering is detected 859 */ 860 static void tcp_disable_fack(struct tcp_sock *tp) 861 { 862 /* RFC3517 uses different metric in lost marker => reset on change */ 863 if (tcp_is_fack(tp)) 864 tp->lost_skb_hint = NULL; 865 tp->rx_opt.sack_ok &= ~2; 866 } 867 868 /* Take a notice that peer is sending D-SACKs */ 869 static void tcp_dsack_seen(struct tcp_sock *tp) 870 { 871 tp->rx_opt.sack_ok |= 4; 872 } 873 874 /* Initialize metrics on socket. */ 875 876 static void tcp_init_metrics(struct sock *sk) 877 { 878 struct tcp_sock *tp = tcp_sk(sk); 879 struct dst_entry *dst = __sk_dst_get(sk); 880 881 if (dst == NULL) 882 goto reset; 883 884 dst_confirm(dst); 885 886 if (dst_metric_locked(dst, RTAX_CWND)) 887 tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND); 888 if (dst_metric(dst, RTAX_SSTHRESH)) { 889 tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH); 890 if (tp->snd_ssthresh > tp->snd_cwnd_clamp) 891 tp->snd_ssthresh = tp->snd_cwnd_clamp; 892 } 893 if (dst_metric(dst, RTAX_REORDERING) && 894 tp->reordering != dst_metric(dst, RTAX_REORDERING)) { 895 tcp_disable_fack(tp); 896 tp->reordering = dst_metric(dst, RTAX_REORDERING); 897 } 898 899 if (dst_metric(dst, RTAX_RTT) == 0) 900 goto reset; 901 902 if (!tp->srtt && dst_metric(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3)) 903 goto reset; 904 905 /* Initial rtt is determined from SYN,SYN-ACK. 906 * The segment is small and rtt may appear much 907 * less than real one. Use per-dst memory 908 * to make it more realistic. 909 * 910 * A bit of theory. RTT is time passed after "normal" sized packet 911 * is sent until it is ACKed. In normal circumstances sending small 912 * packets force peer to delay ACKs and calculation is correct too. 913 * The algorithm is adaptive and, provided we follow specs, it 914 * NEVER underestimate RTT. BUT! If peer tries to make some clever 915 * tricks sort of "quick acks" for time long enough to decrease RTT 916 * to low value, and then abruptly stops to do it and starts to delay 917 * ACKs, wait for troubles. 918 */ 919 if (dst_metric(dst, RTAX_RTT) > tp->srtt) { 920 tp->srtt = dst_metric(dst, RTAX_RTT); 921 tp->rtt_seq = tp->snd_nxt; 922 } 923 if (dst_metric(dst, RTAX_RTTVAR) > tp->mdev) { 924 tp->mdev = dst_metric(dst, RTAX_RTTVAR); 925 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk)); 926 } 927 tcp_set_rto(sk); 928 tcp_bound_rto(sk); 929 if (inet_csk(sk)->icsk_rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp) 930 goto reset; 931 tp->snd_cwnd = tcp_init_cwnd(tp, dst); 932 tp->snd_cwnd_stamp = tcp_time_stamp; 933 return; 934 935 reset: 936 /* Play conservative. If timestamps are not 937 * supported, TCP will fail to recalculate correct 938 * rtt, if initial rto is too small. FORGET ALL AND RESET! 939 */ 940 if (!tp->rx_opt.saw_tstamp && tp->srtt) { 941 tp->srtt = 0; 942 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT; 943 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; 944 } 945 } 946 947 static void tcp_update_reordering(struct sock *sk, const int metric, 948 const int ts) 949 { 950 struct tcp_sock *tp = tcp_sk(sk); 951 if (metric > tp->reordering) { 952 tp->reordering = min(TCP_MAX_REORDERING, metric); 953 954 /* This exciting event is worth to be remembered. 8) */ 955 if (ts) 956 NET_INC_STATS_BH(LINUX_MIB_TCPTSREORDER); 957 else if (tcp_is_reno(tp)) 958 NET_INC_STATS_BH(LINUX_MIB_TCPRENOREORDER); 959 else if (tcp_is_fack(tp)) 960 NET_INC_STATS_BH(LINUX_MIB_TCPFACKREORDER); 961 else 962 NET_INC_STATS_BH(LINUX_MIB_TCPSACKREORDER); 963 #if FASTRETRANS_DEBUG > 1 964 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n", 965 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state, 966 tp->reordering, 967 tp->fackets_out, 968 tp->sacked_out, 969 tp->undo_marker ? tp->undo_retrans : 0); 970 #endif 971 tcp_disable_fack(tp); 972 } 973 } 974 975 /* This procedure tags the retransmission queue when SACKs arrive. 976 * 977 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L). 978 * Packets in queue with these bits set are counted in variables 979 * sacked_out, retrans_out and lost_out, correspondingly. 980 * 981 * Valid combinations are: 982 * Tag InFlight Description 983 * 0 1 - orig segment is in flight. 984 * S 0 - nothing flies, orig reached receiver. 985 * L 0 - nothing flies, orig lost by net. 986 * R 2 - both orig and retransmit are in flight. 987 * L|R 1 - orig is lost, retransmit is in flight. 988 * S|R 1 - orig reached receiver, retrans is still in flight. 989 * (L|S|R is logically valid, it could occur when L|R is sacked, 990 * but it is equivalent to plain S and code short-curcuits it to S. 991 * L|S is logically invalid, it would mean -1 packet in flight 8)) 992 * 993 * These 6 states form finite state machine, controlled by the following events: 994 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue()) 995 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue()) 996 * 3. Loss detection event of one of three flavors: 997 * A. Scoreboard estimator decided the packet is lost. 998 * A'. Reno "three dupacks" marks head of queue lost. 999 * A''. Its FACK modfication, head until snd.fack is lost. 1000 * B. SACK arrives sacking data transmitted after never retransmitted 1001 * hole was sent out. 1002 * C. SACK arrives sacking SND.NXT at the moment, when the 1003 * segment was retransmitted. 1004 * 4. D-SACK added new rule: D-SACK changes any tag to S. 1005 * 1006 * It is pleasant to note, that state diagram turns out to be commutative, 1007 * so that we are allowed not to be bothered by order of our actions, 1008 * when multiple events arrive simultaneously. (see the function below). 1009 * 1010 * Reordering detection. 1011 * -------------------- 1012 * Reordering metric is maximal distance, which a packet can be displaced 1013 * in packet stream. With SACKs we can estimate it: 1014 * 1015 * 1. SACK fills old hole and the corresponding segment was not 1016 * ever retransmitted -> reordering. Alas, we cannot use it 1017 * when segment was retransmitted. 1018 * 2. The last flaw is solved with D-SACK. D-SACK arrives 1019 * for retransmitted and already SACKed segment -> reordering.. 1020 * Both of these heuristics are not used in Loss state, when we cannot 1021 * account for retransmits accurately. 1022 * 1023 * SACK block validation. 1024 * ---------------------- 1025 * 1026 * SACK block range validation checks that the received SACK block fits to 1027 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT. 1028 * Note that SND.UNA is not included to the range though being valid because 1029 * it means that the receiver is rather inconsistent with itself reporting 1030 * SACK reneging when it should advance SND.UNA. Such SACK block this is 1031 * perfectly valid, however, in light of RFC2018 which explicitly states 1032 * that "SACK block MUST reflect the newest segment. Even if the newest 1033 * segment is going to be discarded ...", not that it looks very clever 1034 * in case of head skb. Due to potentional receiver driven attacks, we 1035 * choose to avoid immediate execution of a walk in write queue due to 1036 * reneging and defer head skb's loss recovery to standard loss recovery 1037 * procedure that will eventually trigger (nothing forbids us doing this). 1038 * 1039 * Implements also blockage to start_seq wrap-around. Problem lies in the 1040 * fact that though start_seq (s) is before end_seq (i.e., not reversed), 1041 * there's no guarantee that it will be before snd_nxt (n). The problem 1042 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt 1043 * wrap (s_w): 1044 * 1045 * <- outs wnd -> <- wrapzone -> 1046 * u e n u_w e_w s n_w 1047 * | | | | | | | 1048 * |<------------+------+----- TCP seqno space --------------+---------->| 1049 * ...-- <2^31 ->| |<--------... 1050 * ...---- >2^31 ------>| |<--------... 1051 * 1052 * Current code wouldn't be vulnerable but it's better still to discard such 1053 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat 1054 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in 1055 * snd_nxt wrap -> snd_una region will then become "well defined", i.e., 1056 * equal to the ideal case (infinite seqno space without wrap caused issues). 1057 * 1058 * With D-SACK the lower bound is extended to cover sequence space below 1059 * SND.UNA down to undo_marker, which is the last point of interest. Yet 1060 * again, D-SACK block must not to go across snd_una (for the same reason as 1061 * for the normal SACK blocks, explained above). But there all simplicity 1062 * ends, TCP might receive valid D-SACKs below that. As long as they reside 1063 * fully below undo_marker they do not affect behavior in anyway and can 1064 * therefore be safely ignored. In rare cases (which are more or less 1065 * theoretical ones), the D-SACK will nicely cross that boundary due to skb 1066 * fragmentation and packet reordering past skb's retransmission. To consider 1067 * them correctly, the acceptable range must be extended even more though 1068 * the exact amount is rather hard to quantify. However, tp->max_window can 1069 * be used as an exaggerated estimate. 1070 */ 1071 static int tcp_is_sackblock_valid(struct tcp_sock *tp, int is_dsack, 1072 u32 start_seq, u32 end_seq) 1073 { 1074 /* Too far in future, or reversed (interpretation is ambiguous) */ 1075 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq)) 1076 return 0; 1077 1078 /* Nasty start_seq wrap-around check (see comments above) */ 1079 if (!before(start_seq, tp->snd_nxt)) 1080 return 0; 1081 1082 /* In outstanding window? ...This is valid exit for D-SACKs too. 1083 * start_seq == snd_una is non-sensical (see comments above) 1084 */ 1085 if (after(start_seq, tp->snd_una)) 1086 return 1; 1087 1088 if (!is_dsack || !tp->undo_marker) 1089 return 0; 1090 1091 /* ...Then it's D-SACK, and must reside below snd_una completely */ 1092 if (!after(end_seq, tp->snd_una)) 1093 return 0; 1094 1095 if (!before(start_seq, tp->undo_marker)) 1096 return 1; 1097 1098 /* Too old */ 1099 if (!after(end_seq, tp->undo_marker)) 1100 return 0; 1101 1102 /* Undo_marker boundary crossing (overestimates a lot). Known already: 1103 * start_seq < undo_marker and end_seq >= undo_marker. 1104 */ 1105 return !before(start_seq, end_seq - tp->max_window); 1106 } 1107 1108 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving". 1109 * Event "C". Later note: FACK people cheated me again 8), we have to account 1110 * for reordering! Ugly, but should help. 1111 * 1112 * Search retransmitted skbs from write_queue that were sent when snd_nxt was 1113 * less than what is now known to be received by the other end (derived from 1114 * highest SACK block). Also calculate the lowest snd_nxt among the remaining 1115 * retransmitted skbs to avoid some costly processing per ACKs. 1116 */ 1117 static void tcp_mark_lost_retrans(struct sock *sk) 1118 { 1119 const struct inet_connection_sock *icsk = inet_csk(sk); 1120 struct tcp_sock *tp = tcp_sk(sk); 1121 struct sk_buff *skb; 1122 int cnt = 0; 1123 u32 new_low_seq = tp->snd_nxt; 1124 u32 received_upto = tcp_highest_sack_seq(tp); 1125 1126 if (!tcp_is_fack(tp) || !tp->retrans_out || 1127 !after(received_upto, tp->lost_retrans_low) || 1128 icsk->icsk_ca_state != TCP_CA_Recovery) 1129 return; 1130 1131 tcp_for_write_queue(skb, sk) { 1132 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq; 1133 1134 if (skb == tcp_send_head(sk)) 1135 break; 1136 if (cnt == tp->retrans_out) 1137 break; 1138 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 1139 continue; 1140 1141 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)) 1142 continue; 1143 1144 if (after(received_upto, ack_seq) && 1145 (tcp_is_fack(tp) || 1146 !before(received_upto, 1147 ack_seq + tp->reordering * tp->mss_cache))) { 1148 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 1149 tp->retrans_out -= tcp_skb_pcount(skb); 1150 1151 /* clear lost hint */ 1152 tp->retransmit_skb_hint = NULL; 1153 1154 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) { 1155 tp->lost_out += tcp_skb_pcount(skb); 1156 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 1157 } 1158 NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT); 1159 } else { 1160 if (before(ack_seq, new_low_seq)) 1161 new_low_seq = ack_seq; 1162 cnt += tcp_skb_pcount(skb); 1163 } 1164 } 1165 1166 if (tp->retrans_out) 1167 tp->lost_retrans_low = new_low_seq; 1168 } 1169 1170 static int tcp_check_dsack(struct tcp_sock *tp, struct sk_buff *ack_skb, 1171 struct tcp_sack_block_wire *sp, int num_sacks, 1172 u32 prior_snd_una) 1173 { 1174 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq); 1175 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq); 1176 int dup_sack = 0; 1177 1178 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) { 1179 dup_sack = 1; 1180 tcp_dsack_seen(tp); 1181 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV); 1182 } else if (num_sacks > 1) { 1183 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq); 1184 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq); 1185 1186 if (!after(end_seq_0, end_seq_1) && 1187 !before(start_seq_0, start_seq_1)) { 1188 dup_sack = 1; 1189 tcp_dsack_seen(tp); 1190 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV); 1191 } 1192 } 1193 1194 /* D-SACK for already forgotten data... Do dumb counting. */ 1195 if (dup_sack && 1196 !after(end_seq_0, prior_snd_una) && 1197 after(end_seq_0, tp->undo_marker)) 1198 tp->undo_retrans--; 1199 1200 return dup_sack; 1201 } 1202 1203 /* Check if skb is fully within the SACK block. In presence of GSO skbs, 1204 * the incoming SACK may not exactly match but we can find smaller MSS 1205 * aligned portion of it that matches. Therefore we might need to fragment 1206 * which may fail and creates some hassle (caller must handle error case 1207 * returns). 1208 */ 1209 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb, 1210 u32 start_seq, u32 end_seq) 1211 { 1212 int in_sack, err; 1213 unsigned int pkt_len; 1214 1215 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && 1216 !before(end_seq, TCP_SKB_CB(skb)->end_seq); 1217 1218 if (tcp_skb_pcount(skb) > 1 && !in_sack && 1219 after(TCP_SKB_CB(skb)->end_seq, start_seq)) { 1220 1221 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); 1222 1223 if (!in_sack) 1224 pkt_len = start_seq - TCP_SKB_CB(skb)->seq; 1225 else 1226 pkt_len = end_seq - TCP_SKB_CB(skb)->seq; 1227 err = tcp_fragment(sk, skb, pkt_len, skb_shinfo(skb)->gso_size); 1228 if (err < 0) 1229 return err; 1230 } 1231 1232 return in_sack; 1233 } 1234 1235 static int tcp_sacktag_one(struct sk_buff *skb, struct sock *sk, 1236 int *reord, int dup_sack, int fack_count) 1237 { 1238 struct tcp_sock *tp = tcp_sk(sk); 1239 u8 sacked = TCP_SKB_CB(skb)->sacked; 1240 int flag = 0; 1241 1242 /* Account D-SACK for retransmitted packet. */ 1243 if (dup_sack && (sacked & TCPCB_RETRANS)) { 1244 if (after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker)) 1245 tp->undo_retrans--; 1246 if (sacked & TCPCB_SACKED_ACKED) 1247 *reord = min(fack_count, *reord); 1248 } 1249 1250 /* Nothing to do; acked frame is about to be dropped (was ACKed). */ 1251 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 1252 return flag; 1253 1254 if (!(sacked & TCPCB_SACKED_ACKED)) { 1255 if (sacked & TCPCB_SACKED_RETRANS) { 1256 /* If the segment is not tagged as lost, 1257 * we do not clear RETRANS, believing 1258 * that retransmission is still in flight. 1259 */ 1260 if (sacked & TCPCB_LOST) { 1261 TCP_SKB_CB(skb)->sacked &= 1262 ~(TCPCB_LOST|TCPCB_SACKED_RETRANS); 1263 tp->lost_out -= tcp_skb_pcount(skb); 1264 tp->retrans_out -= tcp_skb_pcount(skb); 1265 1266 /* clear lost hint */ 1267 tp->retransmit_skb_hint = NULL; 1268 } 1269 } else { 1270 if (!(sacked & TCPCB_RETRANS)) { 1271 /* New sack for not retransmitted frame, 1272 * which was in hole. It is reordering. 1273 */ 1274 if (before(TCP_SKB_CB(skb)->seq, 1275 tcp_highest_sack_seq(tp))) 1276 *reord = min(fack_count, *reord); 1277 1278 /* SACK enhanced F-RTO (RFC4138; Appendix B) */ 1279 if (!after(TCP_SKB_CB(skb)->end_seq, tp->frto_highmark)) 1280 flag |= FLAG_ONLY_ORIG_SACKED; 1281 } 1282 1283 if (sacked & TCPCB_LOST) { 1284 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; 1285 tp->lost_out -= tcp_skb_pcount(skb); 1286 1287 /* clear lost hint */ 1288 tp->retransmit_skb_hint = NULL; 1289 } 1290 } 1291 1292 TCP_SKB_CB(skb)->sacked |= TCPCB_SACKED_ACKED; 1293 flag |= FLAG_DATA_SACKED; 1294 tp->sacked_out += tcp_skb_pcount(skb); 1295 1296 fack_count += tcp_skb_pcount(skb); 1297 1298 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */ 1299 if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) && 1300 before(TCP_SKB_CB(skb)->seq, 1301 TCP_SKB_CB(tp->lost_skb_hint)->seq)) 1302 tp->lost_cnt_hint += tcp_skb_pcount(skb); 1303 1304 if (fack_count > tp->fackets_out) 1305 tp->fackets_out = fack_count; 1306 1307 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp))) 1308 tcp_advance_highest_sack(sk, skb); 1309 } 1310 1311 /* D-SACK. We can detect redundant retransmission in S|R and plain R 1312 * frames and clear it. undo_retrans is decreased above, L|R frames 1313 * are accounted above as well. 1314 */ 1315 if (dup_sack && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)) { 1316 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 1317 tp->retrans_out -= tcp_skb_pcount(skb); 1318 tp->retransmit_skb_hint = NULL; 1319 } 1320 1321 return flag; 1322 } 1323 1324 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk, 1325 struct tcp_sack_block *next_dup, 1326 u32 start_seq, u32 end_seq, 1327 int dup_sack_in, int *fack_count, 1328 int *reord, int *flag) 1329 { 1330 tcp_for_write_queue_from(skb, sk) { 1331 int in_sack = 0; 1332 int dup_sack = dup_sack_in; 1333 1334 if (skb == tcp_send_head(sk)) 1335 break; 1336 1337 /* queue is in-order => we can short-circuit the walk early */ 1338 if (!before(TCP_SKB_CB(skb)->seq, end_seq)) 1339 break; 1340 1341 if ((next_dup != NULL) && 1342 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) { 1343 in_sack = tcp_match_skb_to_sack(sk, skb, 1344 next_dup->start_seq, 1345 next_dup->end_seq); 1346 if (in_sack > 0) 1347 dup_sack = 1; 1348 } 1349 1350 if (in_sack <= 0) 1351 in_sack = tcp_match_skb_to_sack(sk, skb, start_seq, 1352 end_seq); 1353 if (unlikely(in_sack < 0)) 1354 break; 1355 1356 if (in_sack) 1357 *flag |= tcp_sacktag_one(skb, sk, reord, dup_sack, 1358 *fack_count); 1359 1360 *fack_count += tcp_skb_pcount(skb); 1361 } 1362 return skb; 1363 } 1364 1365 /* Avoid all extra work that is being done by sacktag while walking in 1366 * a normal way 1367 */ 1368 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk, 1369 u32 skip_to_seq, int *fack_count) 1370 { 1371 tcp_for_write_queue_from(skb, sk) { 1372 if (skb == tcp_send_head(sk)) 1373 break; 1374 1375 if (!before(TCP_SKB_CB(skb)->end_seq, skip_to_seq)) 1376 break; 1377 1378 *fack_count += tcp_skb_pcount(skb); 1379 } 1380 return skb; 1381 } 1382 1383 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb, 1384 struct sock *sk, 1385 struct tcp_sack_block *next_dup, 1386 u32 skip_to_seq, 1387 int *fack_count, int *reord, 1388 int *flag) 1389 { 1390 if (next_dup == NULL) 1391 return skb; 1392 1393 if (before(next_dup->start_seq, skip_to_seq)) { 1394 skb = tcp_sacktag_skip(skb, sk, next_dup->start_seq, fack_count); 1395 tcp_sacktag_walk(skb, sk, NULL, 1396 next_dup->start_seq, next_dup->end_seq, 1397 1, fack_count, reord, flag); 1398 } 1399 1400 return skb; 1401 } 1402 1403 static int tcp_sack_cache_ok(struct tcp_sock *tp, struct tcp_sack_block *cache) 1404 { 1405 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache); 1406 } 1407 1408 static int 1409 tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, 1410 u32 prior_snd_una) 1411 { 1412 const struct inet_connection_sock *icsk = inet_csk(sk); 1413 struct tcp_sock *tp = tcp_sk(sk); 1414 unsigned char *ptr = (skb_transport_header(ack_skb) + 1415 TCP_SKB_CB(ack_skb)->sacked); 1416 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2); 1417 struct tcp_sack_block sp[4]; 1418 struct tcp_sack_block *cache; 1419 struct sk_buff *skb; 1420 int num_sacks = (ptr[1] - TCPOLEN_SACK_BASE) >> 3; 1421 int used_sacks; 1422 int reord = tp->packets_out; 1423 int flag = 0; 1424 int found_dup_sack = 0; 1425 int fack_count; 1426 int i, j; 1427 int first_sack_index; 1428 1429 if (!tp->sacked_out) { 1430 if (WARN_ON(tp->fackets_out)) 1431 tp->fackets_out = 0; 1432 tcp_highest_sack_reset(sk); 1433 } 1434 1435 found_dup_sack = tcp_check_dsack(tp, ack_skb, sp_wire, 1436 num_sacks, prior_snd_una); 1437 if (found_dup_sack) 1438 flag |= FLAG_DSACKING_ACK; 1439 1440 /* Eliminate too old ACKs, but take into 1441 * account more or less fresh ones, they can 1442 * contain valid SACK info. 1443 */ 1444 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window)) 1445 return 0; 1446 1447 if (!tp->packets_out) 1448 goto out; 1449 1450 used_sacks = 0; 1451 first_sack_index = 0; 1452 for (i = 0; i < num_sacks; i++) { 1453 int dup_sack = !i && found_dup_sack; 1454 1455 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq); 1456 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq); 1457 1458 if (!tcp_is_sackblock_valid(tp, dup_sack, 1459 sp[used_sacks].start_seq, 1460 sp[used_sacks].end_seq)) { 1461 if (dup_sack) { 1462 if (!tp->undo_marker) 1463 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKIGNOREDNOUNDO); 1464 else 1465 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKIGNOREDOLD); 1466 } else { 1467 /* Don't count olds caused by ACK reordering */ 1468 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) && 1469 !after(sp[used_sacks].end_seq, tp->snd_una)) 1470 continue; 1471 NET_INC_STATS_BH(LINUX_MIB_TCPSACKDISCARD); 1472 } 1473 if (i == 0) 1474 first_sack_index = -1; 1475 continue; 1476 } 1477 1478 /* Ignore very old stuff early */ 1479 if (!after(sp[used_sacks].end_seq, prior_snd_una)) 1480 continue; 1481 1482 used_sacks++; 1483 } 1484 1485 /* order SACK blocks to allow in order walk of the retrans queue */ 1486 for (i = used_sacks - 1; i > 0; i--) { 1487 for (j = 0; j < i; j++) { 1488 if (after(sp[j].start_seq, sp[j + 1].start_seq)) { 1489 struct tcp_sack_block tmp; 1490 1491 tmp = sp[j]; 1492 sp[j] = sp[j + 1]; 1493 sp[j + 1] = tmp; 1494 1495 /* Track where the first SACK block goes to */ 1496 if (j == first_sack_index) 1497 first_sack_index = j + 1; 1498 } 1499 } 1500 } 1501 1502 skb = tcp_write_queue_head(sk); 1503 fack_count = 0; 1504 i = 0; 1505 1506 if (!tp->sacked_out) { 1507 /* It's already past, so skip checking against it */ 1508 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache); 1509 } else { 1510 cache = tp->recv_sack_cache; 1511 /* Skip empty blocks in at head of the cache */ 1512 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq && 1513 !cache->end_seq) 1514 cache++; 1515 } 1516 1517 while (i < used_sacks) { 1518 u32 start_seq = sp[i].start_seq; 1519 u32 end_seq = sp[i].end_seq; 1520 int dup_sack = (found_dup_sack && (i == first_sack_index)); 1521 struct tcp_sack_block *next_dup = NULL; 1522 1523 if (found_dup_sack && ((i + 1) == first_sack_index)) 1524 next_dup = &sp[i + 1]; 1525 1526 /* Event "B" in the comment above. */ 1527 if (after(end_seq, tp->high_seq)) 1528 flag |= FLAG_DATA_LOST; 1529 1530 /* Skip too early cached blocks */ 1531 while (tcp_sack_cache_ok(tp, cache) && 1532 !before(start_seq, cache->end_seq)) 1533 cache++; 1534 1535 /* Can skip some work by looking recv_sack_cache? */ 1536 if (tcp_sack_cache_ok(tp, cache) && !dup_sack && 1537 after(end_seq, cache->start_seq)) { 1538 1539 /* Head todo? */ 1540 if (before(start_seq, cache->start_seq)) { 1541 skb = tcp_sacktag_skip(skb, sk, start_seq, 1542 &fack_count); 1543 skb = tcp_sacktag_walk(skb, sk, next_dup, 1544 start_seq, 1545 cache->start_seq, 1546 dup_sack, &fack_count, 1547 &reord, &flag); 1548 } 1549 1550 /* Rest of the block already fully processed? */ 1551 if (!after(end_seq, cache->end_seq)) 1552 goto advance_sp; 1553 1554 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup, 1555 cache->end_seq, 1556 &fack_count, &reord, 1557 &flag); 1558 1559 /* ...tail remains todo... */ 1560 if (tcp_highest_sack_seq(tp) == cache->end_seq) { 1561 /* ...but better entrypoint exists! */ 1562 skb = tcp_highest_sack(sk); 1563 if (skb == NULL) 1564 break; 1565 fack_count = tp->fackets_out; 1566 cache++; 1567 goto walk; 1568 } 1569 1570 skb = tcp_sacktag_skip(skb, sk, cache->end_seq, 1571 &fack_count); 1572 /* Check overlap against next cached too (past this one already) */ 1573 cache++; 1574 continue; 1575 } 1576 1577 if (!before(start_seq, tcp_highest_sack_seq(tp))) { 1578 skb = tcp_highest_sack(sk); 1579 if (skb == NULL) 1580 break; 1581 fack_count = tp->fackets_out; 1582 } 1583 skb = tcp_sacktag_skip(skb, sk, start_seq, &fack_count); 1584 1585 walk: 1586 skb = tcp_sacktag_walk(skb, sk, next_dup, start_seq, end_seq, 1587 dup_sack, &fack_count, &reord, &flag); 1588 1589 advance_sp: 1590 /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct 1591 * due to in-order walk 1592 */ 1593 if (after(end_seq, tp->frto_highmark)) 1594 flag &= ~FLAG_ONLY_ORIG_SACKED; 1595 1596 i++; 1597 } 1598 1599 /* Clear the head of the cache sack blocks so we can skip it next time */ 1600 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) { 1601 tp->recv_sack_cache[i].start_seq = 0; 1602 tp->recv_sack_cache[i].end_seq = 0; 1603 } 1604 for (j = 0; j < used_sacks; j++) 1605 tp->recv_sack_cache[i++] = sp[j]; 1606 1607 tcp_mark_lost_retrans(sk); 1608 1609 tcp_verify_left_out(tp); 1610 1611 if ((reord < tp->fackets_out) && 1612 ((icsk->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker) && 1613 (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark))) 1614 tcp_update_reordering(sk, tp->fackets_out - reord, 0); 1615 1616 out: 1617 1618 #if FASTRETRANS_DEBUG > 0 1619 BUG_TRAP((int)tp->sacked_out >= 0); 1620 BUG_TRAP((int)tp->lost_out >= 0); 1621 BUG_TRAP((int)tp->retrans_out >= 0); 1622 BUG_TRAP((int)tcp_packets_in_flight(tp) >= 0); 1623 #endif 1624 return flag; 1625 } 1626 1627 /* Limits sacked_out so that sum with lost_out isn't ever larger than 1628 * packets_out. Returns zero if sacked_out adjustement wasn't necessary. 1629 */ 1630 int tcp_limit_reno_sacked(struct tcp_sock *tp) 1631 { 1632 u32 holes; 1633 1634 holes = max(tp->lost_out, 1U); 1635 holes = min(holes, tp->packets_out); 1636 1637 if ((tp->sacked_out + holes) > tp->packets_out) { 1638 tp->sacked_out = tp->packets_out - holes; 1639 return 1; 1640 } 1641 return 0; 1642 } 1643 1644 /* If we receive more dupacks than we expected counting segments 1645 * in assumption of absent reordering, interpret this as reordering. 1646 * The only another reason could be bug in receiver TCP. 1647 */ 1648 static void tcp_check_reno_reordering(struct sock *sk, const int addend) 1649 { 1650 struct tcp_sock *tp = tcp_sk(sk); 1651 if (tcp_limit_reno_sacked(tp)) 1652 tcp_update_reordering(sk, tp->packets_out + addend, 0); 1653 } 1654 1655 /* Emulate SACKs for SACKless connection: account for a new dupack. */ 1656 1657 static void tcp_add_reno_sack(struct sock *sk) 1658 { 1659 struct tcp_sock *tp = tcp_sk(sk); 1660 tp->sacked_out++; 1661 tcp_check_reno_reordering(sk, 0); 1662 tcp_verify_left_out(tp); 1663 } 1664 1665 /* Account for ACK, ACKing some data in Reno Recovery phase. */ 1666 1667 static void tcp_remove_reno_sacks(struct sock *sk, int acked) 1668 { 1669 struct tcp_sock *tp = tcp_sk(sk); 1670 1671 if (acked > 0) { 1672 /* One ACK acked hole. The rest eat duplicate ACKs. */ 1673 if (acked - 1 >= tp->sacked_out) 1674 tp->sacked_out = 0; 1675 else 1676 tp->sacked_out -= acked - 1; 1677 } 1678 tcp_check_reno_reordering(sk, acked); 1679 tcp_verify_left_out(tp); 1680 } 1681 1682 static inline void tcp_reset_reno_sack(struct tcp_sock *tp) 1683 { 1684 tp->sacked_out = 0; 1685 } 1686 1687 static int tcp_is_sackfrto(const struct tcp_sock *tp) 1688 { 1689 return (sysctl_tcp_frto == 0x2) && !tcp_is_reno(tp); 1690 } 1691 1692 /* F-RTO can only be used if TCP has never retransmitted anything other than 1693 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here) 1694 */ 1695 int tcp_use_frto(struct sock *sk) 1696 { 1697 const struct tcp_sock *tp = tcp_sk(sk); 1698 const struct inet_connection_sock *icsk = inet_csk(sk); 1699 struct sk_buff *skb; 1700 1701 if (!sysctl_tcp_frto) 1702 return 0; 1703 1704 /* MTU probe and F-RTO won't really play nicely along currently */ 1705 if (icsk->icsk_mtup.probe_size) 1706 return 0; 1707 1708 if (tcp_is_sackfrto(tp)) 1709 return 1; 1710 1711 /* Avoid expensive walking of rexmit queue if possible */ 1712 if (tp->retrans_out > 1) 1713 return 0; 1714 1715 skb = tcp_write_queue_head(sk); 1716 skb = tcp_write_queue_next(sk, skb); /* Skips head */ 1717 tcp_for_write_queue_from(skb, sk) { 1718 if (skb == tcp_send_head(sk)) 1719 break; 1720 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) 1721 return 0; 1722 /* Short-circuit when first non-SACKed skb has been checked */ 1723 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) 1724 break; 1725 } 1726 return 1; 1727 } 1728 1729 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO 1730 * recovery a bit and use heuristics in tcp_process_frto() to detect if 1731 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to 1732 * keep retrans_out counting accurate (with SACK F-RTO, other than head 1733 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS 1734 * bits are handled if the Loss state is really to be entered (in 1735 * tcp_enter_frto_loss). 1736 * 1737 * Do like tcp_enter_loss() would; when RTO expires the second time it 1738 * does: 1739 * "Reduce ssthresh if it has not yet been made inside this window." 1740 */ 1741 void tcp_enter_frto(struct sock *sk) 1742 { 1743 const struct inet_connection_sock *icsk = inet_csk(sk); 1744 struct tcp_sock *tp = tcp_sk(sk); 1745 struct sk_buff *skb; 1746 1747 if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) || 1748 tp->snd_una == tp->high_seq || 1749 ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) && 1750 !icsk->icsk_retransmits)) { 1751 tp->prior_ssthresh = tcp_current_ssthresh(sk); 1752 /* Our state is too optimistic in ssthresh() call because cwnd 1753 * is not reduced until tcp_enter_frto_loss() when previous F-RTO 1754 * recovery has not yet completed. Pattern would be this: RTO, 1755 * Cumulative ACK, RTO (2xRTO for the same segment does not end 1756 * up here twice). 1757 * RFC4138 should be more specific on what to do, even though 1758 * RTO is quite unlikely to occur after the first Cumulative ACK 1759 * due to back-off and complexity of triggering events ... 1760 */ 1761 if (tp->frto_counter) { 1762 u32 stored_cwnd; 1763 stored_cwnd = tp->snd_cwnd; 1764 tp->snd_cwnd = 2; 1765 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 1766 tp->snd_cwnd = stored_cwnd; 1767 } else { 1768 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 1769 } 1770 /* ... in theory, cong.control module could do "any tricks" in 1771 * ssthresh(), which means that ca_state, lost bits and lost_out 1772 * counter would have to be faked before the call occurs. We 1773 * consider that too expensive, unlikely and hacky, so modules 1774 * using these in ssthresh() must deal these incompatibility 1775 * issues if they receives CA_EVENT_FRTO and frto_counter != 0 1776 */ 1777 tcp_ca_event(sk, CA_EVENT_FRTO); 1778 } 1779 1780 tp->undo_marker = tp->snd_una; 1781 tp->undo_retrans = 0; 1782 1783 skb = tcp_write_queue_head(sk); 1784 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) 1785 tp->undo_marker = 0; 1786 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 1787 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 1788 tp->retrans_out -= tcp_skb_pcount(skb); 1789 } 1790 tcp_verify_left_out(tp); 1791 1792 /* Too bad if TCP was application limited */ 1793 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1); 1794 1795 /* Earlier loss recovery underway (see RFC4138; Appendix B). 1796 * The last condition is necessary at least in tp->frto_counter case. 1797 */ 1798 if (tcp_is_sackfrto(tp) && (tp->frto_counter || 1799 ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) && 1800 after(tp->high_seq, tp->snd_una)) { 1801 tp->frto_highmark = tp->high_seq; 1802 } else { 1803 tp->frto_highmark = tp->snd_nxt; 1804 } 1805 tcp_set_ca_state(sk, TCP_CA_Disorder); 1806 tp->high_seq = tp->snd_nxt; 1807 tp->frto_counter = 1; 1808 } 1809 1810 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO, 1811 * which indicates that we should follow the traditional RTO recovery, 1812 * i.e. mark everything lost and do go-back-N retransmission. 1813 */ 1814 static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag) 1815 { 1816 struct tcp_sock *tp = tcp_sk(sk); 1817 struct sk_buff *skb; 1818 1819 tp->lost_out = 0; 1820 tp->retrans_out = 0; 1821 if (tcp_is_reno(tp)) 1822 tcp_reset_reno_sack(tp); 1823 1824 tcp_for_write_queue(skb, sk) { 1825 if (skb == tcp_send_head(sk)) 1826 break; 1827 1828 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; 1829 /* 1830 * Count the retransmission made on RTO correctly (only when 1831 * waiting for the first ACK and did not get it)... 1832 */ 1833 if ((tp->frto_counter == 1) && !(flag & FLAG_DATA_ACKED)) { 1834 /* For some reason this R-bit might get cleared? */ 1835 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 1836 tp->retrans_out += tcp_skb_pcount(skb); 1837 /* ...enter this if branch just for the first segment */ 1838 flag |= FLAG_DATA_ACKED; 1839 } else { 1840 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) 1841 tp->undo_marker = 0; 1842 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 1843 } 1844 1845 /* Marking forward transmissions that were made after RTO lost 1846 * can cause unnecessary retransmissions in some scenarios, 1847 * SACK blocks will mitigate that in some but not in all cases. 1848 * We used to not mark them but it was causing break-ups with 1849 * receivers that do only in-order receival. 1850 * 1851 * TODO: we could detect presence of such receiver and select 1852 * different behavior per flow. 1853 */ 1854 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) { 1855 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 1856 tp->lost_out += tcp_skb_pcount(skb); 1857 } 1858 } 1859 tcp_verify_left_out(tp); 1860 1861 tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments; 1862 tp->snd_cwnd_cnt = 0; 1863 tp->snd_cwnd_stamp = tcp_time_stamp; 1864 tp->frto_counter = 0; 1865 tp->bytes_acked = 0; 1866 1867 tp->reordering = min_t(unsigned int, tp->reordering, 1868 sysctl_tcp_reordering); 1869 tcp_set_ca_state(sk, TCP_CA_Loss); 1870 tp->high_seq = tp->snd_nxt; 1871 TCP_ECN_queue_cwr(tp); 1872 1873 tcp_clear_retrans_hints_partial(tp); 1874 } 1875 1876 static void tcp_clear_retrans_partial(struct tcp_sock *tp) 1877 { 1878 tp->retrans_out = 0; 1879 tp->lost_out = 0; 1880 1881 tp->undo_marker = 0; 1882 tp->undo_retrans = 0; 1883 } 1884 1885 void tcp_clear_retrans(struct tcp_sock *tp) 1886 { 1887 tcp_clear_retrans_partial(tp); 1888 1889 tp->fackets_out = 0; 1890 tp->sacked_out = 0; 1891 } 1892 1893 /* Enter Loss state. If "how" is not zero, forget all SACK information 1894 * and reset tags completely, otherwise preserve SACKs. If receiver 1895 * dropped its ofo queue, we will know this due to reneging detection. 1896 */ 1897 void tcp_enter_loss(struct sock *sk, int how) 1898 { 1899 const struct inet_connection_sock *icsk = inet_csk(sk); 1900 struct tcp_sock *tp = tcp_sk(sk); 1901 struct sk_buff *skb; 1902 1903 /* Reduce ssthresh if it has not yet been made inside this window. */ 1904 if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq || 1905 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) { 1906 tp->prior_ssthresh = tcp_current_ssthresh(sk); 1907 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 1908 tcp_ca_event(sk, CA_EVENT_LOSS); 1909 } 1910 tp->snd_cwnd = 1; 1911 tp->snd_cwnd_cnt = 0; 1912 tp->snd_cwnd_stamp = tcp_time_stamp; 1913 1914 tp->bytes_acked = 0; 1915 tcp_clear_retrans_partial(tp); 1916 1917 if (tcp_is_reno(tp)) 1918 tcp_reset_reno_sack(tp); 1919 1920 if (!how) { 1921 /* Push undo marker, if it was plain RTO and nothing 1922 * was retransmitted. */ 1923 tp->undo_marker = tp->snd_una; 1924 tcp_clear_retrans_hints_partial(tp); 1925 } else { 1926 tp->sacked_out = 0; 1927 tp->fackets_out = 0; 1928 tcp_clear_all_retrans_hints(tp); 1929 } 1930 1931 tcp_for_write_queue(skb, sk) { 1932 if (skb == tcp_send_head(sk)) 1933 break; 1934 1935 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) 1936 tp->undo_marker = 0; 1937 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED; 1938 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) { 1939 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED; 1940 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 1941 tp->lost_out += tcp_skb_pcount(skb); 1942 } 1943 } 1944 tcp_verify_left_out(tp); 1945 1946 tp->reordering = min_t(unsigned int, tp->reordering, 1947 sysctl_tcp_reordering); 1948 tcp_set_ca_state(sk, TCP_CA_Loss); 1949 tp->high_seq = tp->snd_nxt; 1950 TCP_ECN_queue_cwr(tp); 1951 /* Abort F-RTO algorithm if one is in progress */ 1952 tp->frto_counter = 0; 1953 } 1954 1955 /* If ACK arrived pointing to a remembered SACK, it means that our 1956 * remembered SACKs do not reflect real state of receiver i.e. 1957 * receiver _host_ is heavily congested (or buggy). 1958 * 1959 * Do processing similar to RTO timeout. 1960 */ 1961 static int tcp_check_sack_reneging(struct sock *sk, int flag) 1962 { 1963 if (flag & FLAG_SACK_RENEGING) { 1964 struct inet_connection_sock *icsk = inet_csk(sk); 1965 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING); 1966 1967 tcp_enter_loss(sk, 1); 1968 icsk->icsk_retransmits++; 1969 tcp_retransmit_skb(sk, tcp_write_queue_head(sk)); 1970 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 1971 icsk->icsk_rto, TCP_RTO_MAX); 1972 return 1; 1973 } 1974 return 0; 1975 } 1976 1977 static inline int tcp_fackets_out(struct tcp_sock *tp) 1978 { 1979 return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out; 1980 } 1981 1982 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs 1983 * counter when SACK is enabled (without SACK, sacked_out is used for 1984 * that purpose). 1985 * 1986 * Instead, with FACK TCP uses fackets_out that includes both SACKed 1987 * segments up to the highest received SACK block so far and holes in 1988 * between them. 1989 * 1990 * With reordering, holes may still be in flight, so RFC3517 recovery 1991 * uses pure sacked_out (total number of SACKed segments) even though 1992 * it violates the RFC that uses duplicate ACKs, often these are equal 1993 * but when e.g. out-of-window ACKs or packet duplication occurs, 1994 * they differ. Since neither occurs due to loss, TCP should really 1995 * ignore them. 1996 */ 1997 static inline int tcp_dupack_heurestics(struct tcp_sock *tp) 1998 { 1999 return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1; 2000 } 2001 2002 static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb) 2003 { 2004 return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto); 2005 } 2006 2007 static inline int tcp_head_timedout(struct sock *sk) 2008 { 2009 struct tcp_sock *tp = tcp_sk(sk); 2010 2011 return tp->packets_out && 2012 tcp_skb_timedout(sk, tcp_write_queue_head(sk)); 2013 } 2014 2015 /* Linux NewReno/SACK/FACK/ECN state machine. 2016 * -------------------------------------- 2017 * 2018 * "Open" Normal state, no dubious events, fast path. 2019 * "Disorder" In all the respects it is "Open", 2020 * but requires a bit more attention. It is entered when 2021 * we see some SACKs or dupacks. It is split of "Open" 2022 * mainly to move some processing from fast path to slow one. 2023 * "CWR" CWND was reduced due to some Congestion Notification event. 2024 * It can be ECN, ICMP source quench, local device congestion. 2025 * "Recovery" CWND was reduced, we are fast-retransmitting. 2026 * "Loss" CWND was reduced due to RTO timeout or SACK reneging. 2027 * 2028 * tcp_fastretrans_alert() is entered: 2029 * - each incoming ACK, if state is not "Open" 2030 * - when arrived ACK is unusual, namely: 2031 * * SACK 2032 * * Duplicate ACK. 2033 * * ECN ECE. 2034 * 2035 * Counting packets in flight is pretty simple. 2036 * 2037 * in_flight = packets_out - left_out + retrans_out 2038 * 2039 * packets_out is SND.NXT-SND.UNA counted in packets. 2040 * 2041 * retrans_out is number of retransmitted segments. 2042 * 2043 * left_out is number of segments left network, but not ACKed yet. 2044 * 2045 * left_out = sacked_out + lost_out 2046 * 2047 * sacked_out: Packets, which arrived to receiver out of order 2048 * and hence not ACKed. With SACKs this number is simply 2049 * amount of SACKed data. Even without SACKs 2050 * it is easy to give pretty reliable estimate of this number, 2051 * counting duplicate ACKs. 2052 * 2053 * lost_out: Packets lost by network. TCP has no explicit 2054 * "loss notification" feedback from network (for now). 2055 * It means that this number can be only _guessed_. 2056 * Actually, it is the heuristics to predict lossage that 2057 * distinguishes different algorithms. 2058 * 2059 * F.e. after RTO, when all the queue is considered as lost, 2060 * lost_out = packets_out and in_flight = retrans_out. 2061 * 2062 * Essentially, we have now two algorithms counting 2063 * lost packets. 2064 * 2065 * FACK: It is the simplest heuristics. As soon as we decided 2066 * that something is lost, we decide that _all_ not SACKed 2067 * packets until the most forward SACK are lost. I.e. 2068 * lost_out = fackets_out - sacked_out and left_out = fackets_out. 2069 * It is absolutely correct estimate, if network does not reorder 2070 * packets. And it loses any connection to reality when reordering 2071 * takes place. We use FACK by default until reordering 2072 * is suspected on the path to this destination. 2073 * 2074 * NewReno: when Recovery is entered, we assume that one segment 2075 * is lost (classic Reno). While we are in Recovery and 2076 * a partial ACK arrives, we assume that one more packet 2077 * is lost (NewReno). This heuristics are the same in NewReno 2078 * and SACK. 2079 * 2080 * Imagine, that's all! Forget about all this shamanism about CWND inflation 2081 * deflation etc. CWND is real congestion window, never inflated, changes 2082 * only according to classic VJ rules. 2083 * 2084 * Really tricky (and requiring careful tuning) part of algorithm 2085 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue(). 2086 * The first determines the moment _when_ we should reduce CWND and, 2087 * hence, slow down forward transmission. In fact, it determines the moment 2088 * when we decide that hole is caused by loss, rather than by a reorder. 2089 * 2090 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill 2091 * holes, caused by lost packets. 2092 * 2093 * And the most logically complicated part of algorithm is undo 2094 * heuristics. We detect false retransmits due to both too early 2095 * fast retransmit (reordering) and underestimated RTO, analyzing 2096 * timestamps and D-SACKs. When we detect that some segments were 2097 * retransmitted by mistake and CWND reduction was wrong, we undo 2098 * window reduction and abort recovery phase. This logic is hidden 2099 * inside several functions named tcp_try_undo_<something>. 2100 */ 2101 2102 /* This function decides, when we should leave Disordered state 2103 * and enter Recovery phase, reducing congestion window. 2104 * 2105 * Main question: may we further continue forward transmission 2106 * with the same cwnd? 2107 */ 2108 static int tcp_time_to_recover(struct sock *sk) 2109 { 2110 struct tcp_sock *tp = tcp_sk(sk); 2111 __u32 packets_out; 2112 2113 /* Do not perform any recovery during F-RTO algorithm */ 2114 if (tp->frto_counter) 2115 return 0; 2116 2117 /* Trick#1: The loss is proven. */ 2118 if (tp->lost_out) 2119 return 1; 2120 2121 /* Not-A-Trick#2 : Classic rule... */ 2122 if (tcp_dupack_heurestics(tp) > tp->reordering) 2123 return 1; 2124 2125 /* Trick#3 : when we use RFC2988 timer restart, fast 2126 * retransmit can be triggered by timeout of queue head. 2127 */ 2128 if (tcp_is_fack(tp) && tcp_head_timedout(sk)) 2129 return 1; 2130 2131 /* Trick#4: It is still not OK... But will it be useful to delay 2132 * recovery more? 2133 */ 2134 packets_out = tp->packets_out; 2135 if (packets_out <= tp->reordering && 2136 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) && 2137 !tcp_may_send_now(sk)) { 2138 /* We have nothing to send. This connection is limited 2139 * either by receiver window or by application. 2140 */ 2141 return 1; 2142 } 2143 2144 return 0; 2145 } 2146 2147 /* RFC: This is from the original, I doubt that this is necessary at all: 2148 * clear xmit_retrans hint if seq of this skb is beyond hint. How could we 2149 * retransmitted past LOST markings in the first place? I'm not fully sure 2150 * about undo and end of connection cases, which can cause R without L? 2151 */ 2152 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb) 2153 { 2154 if ((tp->retransmit_skb_hint != NULL) && 2155 before(TCP_SKB_CB(skb)->seq, 2156 TCP_SKB_CB(tp->retransmit_skb_hint)->seq)) 2157 tp->retransmit_skb_hint = NULL; 2158 } 2159 2160 /* Mark head of queue up as lost. With RFC3517 SACK, the packets is 2161 * is against sacked "cnt", otherwise it's against facked "cnt" 2162 */ 2163 static void tcp_mark_head_lost(struct sock *sk, int packets) 2164 { 2165 struct tcp_sock *tp = tcp_sk(sk); 2166 struct sk_buff *skb; 2167 int cnt, oldcnt; 2168 int err; 2169 unsigned int mss; 2170 2171 BUG_TRAP(packets <= tp->packets_out); 2172 if (tp->lost_skb_hint) { 2173 skb = tp->lost_skb_hint; 2174 cnt = tp->lost_cnt_hint; 2175 } else { 2176 skb = tcp_write_queue_head(sk); 2177 cnt = 0; 2178 } 2179 2180 tcp_for_write_queue_from(skb, sk) { 2181 if (skb == tcp_send_head(sk)) 2182 break; 2183 /* TODO: do this better */ 2184 /* this is not the most efficient way to do this... */ 2185 tp->lost_skb_hint = skb; 2186 tp->lost_cnt_hint = cnt; 2187 2188 if (after(TCP_SKB_CB(skb)->end_seq, tp->high_seq)) 2189 break; 2190 2191 oldcnt = cnt; 2192 if (tcp_is_fack(tp) || tcp_is_reno(tp) || 2193 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) 2194 cnt += tcp_skb_pcount(skb); 2195 2196 if (cnt > packets) { 2197 if (tcp_is_sack(tp) || (oldcnt >= packets)) 2198 break; 2199 2200 mss = skb_shinfo(skb)->gso_size; 2201 err = tcp_fragment(sk, skb, (packets - oldcnt) * mss, mss); 2202 if (err < 0) 2203 break; 2204 cnt = packets; 2205 } 2206 2207 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_SACKED_ACKED|TCPCB_LOST))) { 2208 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 2209 tp->lost_out += tcp_skb_pcount(skb); 2210 tcp_verify_retransmit_hint(tp, skb); 2211 } 2212 } 2213 tcp_verify_left_out(tp); 2214 } 2215 2216 /* Account newly detected lost packet(s) */ 2217 2218 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit) 2219 { 2220 struct tcp_sock *tp = tcp_sk(sk); 2221 2222 if (tcp_is_reno(tp)) { 2223 tcp_mark_head_lost(sk, 1); 2224 } else if (tcp_is_fack(tp)) { 2225 int lost = tp->fackets_out - tp->reordering; 2226 if (lost <= 0) 2227 lost = 1; 2228 tcp_mark_head_lost(sk, lost); 2229 } else { 2230 int sacked_upto = tp->sacked_out - tp->reordering; 2231 if (sacked_upto < fast_rexmit) 2232 sacked_upto = fast_rexmit; 2233 tcp_mark_head_lost(sk, sacked_upto); 2234 } 2235 2236 /* New heuristics: it is possible only after we switched 2237 * to restart timer each time when something is ACKed. 2238 * Hence, we can detect timed out packets during fast 2239 * retransmit without falling to slow start. 2240 */ 2241 if (tcp_is_fack(tp) && tcp_head_timedout(sk)) { 2242 struct sk_buff *skb; 2243 2244 skb = tp->scoreboard_skb_hint ? tp->scoreboard_skb_hint 2245 : tcp_write_queue_head(sk); 2246 2247 tcp_for_write_queue_from(skb, sk) { 2248 if (skb == tcp_send_head(sk)) 2249 break; 2250 if (!tcp_skb_timedout(sk, skb)) 2251 break; 2252 2253 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_SACKED_ACKED|TCPCB_LOST))) { 2254 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 2255 tp->lost_out += tcp_skb_pcount(skb); 2256 tcp_verify_retransmit_hint(tp, skb); 2257 } 2258 } 2259 2260 tp->scoreboard_skb_hint = skb; 2261 2262 tcp_verify_left_out(tp); 2263 } 2264 } 2265 2266 /* CWND moderation, preventing bursts due to too big ACKs 2267 * in dubious situations. 2268 */ 2269 static inline void tcp_moderate_cwnd(struct tcp_sock *tp) 2270 { 2271 tp->snd_cwnd = min(tp->snd_cwnd, 2272 tcp_packets_in_flight(tp) + tcp_max_burst(tp)); 2273 tp->snd_cwnd_stamp = tcp_time_stamp; 2274 } 2275 2276 /* Lower bound on congestion window is slow start threshold 2277 * unless congestion avoidance choice decides to overide it. 2278 */ 2279 static inline u32 tcp_cwnd_min(const struct sock *sk) 2280 { 2281 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 2282 2283 return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh; 2284 } 2285 2286 /* Decrease cwnd each second ack. */ 2287 static void tcp_cwnd_down(struct sock *sk, int flag) 2288 { 2289 struct tcp_sock *tp = tcp_sk(sk); 2290 int decr = tp->snd_cwnd_cnt + 1; 2291 2292 if ((flag & (FLAG_ANY_PROGRESS | FLAG_DSACKING_ACK)) || 2293 (tcp_is_reno(tp) && !(flag & FLAG_NOT_DUP))) { 2294 tp->snd_cwnd_cnt = decr & 1; 2295 decr >>= 1; 2296 2297 if (decr && tp->snd_cwnd > tcp_cwnd_min(sk)) 2298 tp->snd_cwnd -= decr; 2299 2300 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1); 2301 tp->snd_cwnd_stamp = tcp_time_stamp; 2302 } 2303 } 2304 2305 /* Nothing was retransmitted or returned timestamp is less 2306 * than timestamp of the first retransmission. 2307 */ 2308 static inline int tcp_packet_delayed(struct tcp_sock *tp) 2309 { 2310 return !tp->retrans_stamp || 2311 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 2312 before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp)); 2313 } 2314 2315 /* Undo procedures. */ 2316 2317 #if FASTRETRANS_DEBUG > 1 2318 static void DBGUNDO(struct sock *sk, const char *msg) 2319 { 2320 struct tcp_sock *tp = tcp_sk(sk); 2321 struct inet_sock *inet = inet_sk(sk); 2322 2323 if (sk->sk_family == AF_INET) { 2324 printk(KERN_DEBUG "Undo %s " NIPQUAD_FMT "/%u c%u l%u ss%u/%u p%u\n", 2325 msg, 2326 NIPQUAD(inet->daddr), ntohs(inet->dport), 2327 tp->snd_cwnd, tcp_left_out(tp), 2328 tp->snd_ssthresh, tp->prior_ssthresh, 2329 tp->packets_out); 2330 } 2331 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE) 2332 else if (sk->sk_family == AF_INET6) { 2333 struct ipv6_pinfo *np = inet6_sk(sk); 2334 printk(KERN_DEBUG "Undo %s " NIP6_FMT "/%u c%u l%u ss%u/%u p%u\n", 2335 msg, 2336 NIP6(np->daddr), ntohs(inet->dport), 2337 tp->snd_cwnd, tcp_left_out(tp), 2338 tp->snd_ssthresh, tp->prior_ssthresh, 2339 tp->packets_out); 2340 } 2341 #endif 2342 } 2343 #else 2344 #define DBGUNDO(x...) do { } while (0) 2345 #endif 2346 2347 static void tcp_undo_cwr(struct sock *sk, const int undo) 2348 { 2349 struct tcp_sock *tp = tcp_sk(sk); 2350 2351 if (tp->prior_ssthresh) { 2352 const struct inet_connection_sock *icsk = inet_csk(sk); 2353 2354 if (icsk->icsk_ca_ops->undo_cwnd) 2355 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk); 2356 else 2357 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1); 2358 2359 if (undo && tp->prior_ssthresh > tp->snd_ssthresh) { 2360 tp->snd_ssthresh = tp->prior_ssthresh; 2361 TCP_ECN_withdraw_cwr(tp); 2362 } 2363 } else { 2364 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh); 2365 } 2366 tcp_moderate_cwnd(tp); 2367 tp->snd_cwnd_stamp = tcp_time_stamp; 2368 2369 /* There is something screwy going on with the retrans hints after 2370 an undo */ 2371 tcp_clear_all_retrans_hints(tp); 2372 } 2373 2374 static inline int tcp_may_undo(struct tcp_sock *tp) 2375 { 2376 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp)); 2377 } 2378 2379 /* People celebrate: "We love our President!" */ 2380 static int tcp_try_undo_recovery(struct sock *sk) 2381 { 2382 struct tcp_sock *tp = tcp_sk(sk); 2383 2384 if (tcp_may_undo(tp)) { 2385 /* Happy end! We did not retransmit anything 2386 * or our original transmission succeeded. 2387 */ 2388 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans"); 2389 tcp_undo_cwr(sk, 1); 2390 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss) 2391 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO); 2392 else 2393 NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO); 2394 tp->undo_marker = 0; 2395 } 2396 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) { 2397 /* Hold old state until something *above* high_seq 2398 * is ACKed. For Reno it is MUST to prevent false 2399 * fast retransmits (RFC2582). SACK TCP is safe. */ 2400 tcp_moderate_cwnd(tp); 2401 return 1; 2402 } 2403 tcp_set_ca_state(sk, TCP_CA_Open); 2404 return 0; 2405 } 2406 2407 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */ 2408 static void tcp_try_undo_dsack(struct sock *sk) 2409 { 2410 struct tcp_sock *tp = tcp_sk(sk); 2411 2412 if (tp->undo_marker && !tp->undo_retrans) { 2413 DBGUNDO(sk, "D-SACK"); 2414 tcp_undo_cwr(sk, 1); 2415 tp->undo_marker = 0; 2416 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO); 2417 } 2418 } 2419 2420 /* Undo during fast recovery after partial ACK. */ 2421 2422 static int tcp_try_undo_partial(struct sock *sk, int acked) 2423 { 2424 struct tcp_sock *tp = tcp_sk(sk); 2425 /* Partial ACK arrived. Force Hoe's retransmit. */ 2426 int failed = tcp_is_reno(tp) || (tcp_fackets_out(tp) > tp->reordering); 2427 2428 if (tcp_may_undo(tp)) { 2429 /* Plain luck! Hole if filled with delayed 2430 * packet, rather than with a retransmit. 2431 */ 2432 if (tp->retrans_out == 0) 2433 tp->retrans_stamp = 0; 2434 2435 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1); 2436 2437 DBGUNDO(sk, "Hoe"); 2438 tcp_undo_cwr(sk, 0); 2439 NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO); 2440 2441 /* So... Do not make Hoe's retransmit yet. 2442 * If the first packet was delayed, the rest 2443 * ones are most probably delayed as well. 2444 */ 2445 failed = 0; 2446 } 2447 return failed; 2448 } 2449 2450 /* Undo during loss recovery after partial ACK. */ 2451 static int tcp_try_undo_loss(struct sock *sk) 2452 { 2453 struct tcp_sock *tp = tcp_sk(sk); 2454 2455 if (tcp_may_undo(tp)) { 2456 struct sk_buff *skb; 2457 tcp_for_write_queue(skb, sk) { 2458 if (skb == tcp_send_head(sk)) 2459 break; 2460 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; 2461 } 2462 2463 tcp_clear_all_retrans_hints(tp); 2464 2465 DBGUNDO(sk, "partial loss"); 2466 tp->lost_out = 0; 2467 tcp_undo_cwr(sk, 1); 2468 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO); 2469 inet_csk(sk)->icsk_retransmits = 0; 2470 tp->undo_marker = 0; 2471 if (tcp_is_sack(tp)) 2472 tcp_set_ca_state(sk, TCP_CA_Open); 2473 return 1; 2474 } 2475 return 0; 2476 } 2477 2478 static inline void tcp_complete_cwr(struct sock *sk) 2479 { 2480 struct tcp_sock *tp = tcp_sk(sk); 2481 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh); 2482 tp->snd_cwnd_stamp = tcp_time_stamp; 2483 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR); 2484 } 2485 2486 static void tcp_try_to_open(struct sock *sk, int flag) 2487 { 2488 struct tcp_sock *tp = tcp_sk(sk); 2489 2490 tcp_verify_left_out(tp); 2491 2492 if (!tp->frto_counter && tp->retrans_out == 0) 2493 tp->retrans_stamp = 0; 2494 2495 if (flag & FLAG_ECE) 2496 tcp_enter_cwr(sk, 1); 2497 2498 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) { 2499 int state = TCP_CA_Open; 2500 2501 if (tcp_left_out(tp) || tp->retrans_out || tp->undo_marker) 2502 state = TCP_CA_Disorder; 2503 2504 if (inet_csk(sk)->icsk_ca_state != state) { 2505 tcp_set_ca_state(sk, state); 2506 tp->high_seq = tp->snd_nxt; 2507 } 2508 tcp_moderate_cwnd(tp); 2509 } else { 2510 tcp_cwnd_down(sk, flag); 2511 } 2512 } 2513 2514 static void tcp_mtup_probe_failed(struct sock *sk) 2515 { 2516 struct inet_connection_sock *icsk = inet_csk(sk); 2517 2518 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1; 2519 icsk->icsk_mtup.probe_size = 0; 2520 } 2521 2522 static void tcp_mtup_probe_success(struct sock *sk, struct sk_buff *skb) 2523 { 2524 struct tcp_sock *tp = tcp_sk(sk); 2525 struct inet_connection_sock *icsk = inet_csk(sk); 2526 2527 /* FIXME: breaks with very large cwnd */ 2528 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2529 tp->snd_cwnd = tp->snd_cwnd * 2530 tcp_mss_to_mtu(sk, tp->mss_cache) / 2531 icsk->icsk_mtup.probe_size; 2532 tp->snd_cwnd_cnt = 0; 2533 tp->snd_cwnd_stamp = tcp_time_stamp; 2534 tp->rcv_ssthresh = tcp_current_ssthresh(sk); 2535 2536 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size; 2537 icsk->icsk_mtup.probe_size = 0; 2538 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 2539 } 2540 2541 /* Process an event, which can update packets-in-flight not trivially. 2542 * Main goal of this function is to calculate new estimate for left_out, 2543 * taking into account both packets sitting in receiver's buffer and 2544 * packets lost by network. 2545 * 2546 * Besides that it does CWND reduction, when packet loss is detected 2547 * and changes state of machine. 2548 * 2549 * It does _not_ decide what to send, it is made in function 2550 * tcp_xmit_retransmit_queue(). 2551 */ 2552 static void tcp_fastretrans_alert(struct sock *sk, int pkts_acked, int flag) 2553 { 2554 struct inet_connection_sock *icsk = inet_csk(sk); 2555 struct tcp_sock *tp = tcp_sk(sk); 2556 int is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP)); 2557 int do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) && 2558 (tcp_fackets_out(tp) > tp->reordering)); 2559 int fast_rexmit = 0; 2560 2561 if (WARN_ON(!tp->packets_out && tp->sacked_out)) 2562 tp->sacked_out = 0; 2563 if (WARN_ON(!tp->sacked_out && tp->fackets_out)) 2564 tp->fackets_out = 0; 2565 2566 /* Now state machine starts. 2567 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */ 2568 if (flag & FLAG_ECE) 2569 tp->prior_ssthresh = 0; 2570 2571 /* B. In all the states check for reneging SACKs. */ 2572 if (tcp_check_sack_reneging(sk, flag)) 2573 return; 2574 2575 /* C. Process data loss notification, provided it is valid. */ 2576 if (tcp_is_fack(tp) && (flag & FLAG_DATA_LOST) && 2577 before(tp->snd_una, tp->high_seq) && 2578 icsk->icsk_ca_state != TCP_CA_Open && 2579 tp->fackets_out > tp->reordering) { 2580 tcp_mark_head_lost(sk, tp->fackets_out - tp->reordering); 2581 NET_INC_STATS_BH(LINUX_MIB_TCPLOSS); 2582 } 2583 2584 /* D. Check consistency of the current state. */ 2585 tcp_verify_left_out(tp); 2586 2587 /* E. Check state exit conditions. State can be terminated 2588 * when high_seq is ACKed. */ 2589 if (icsk->icsk_ca_state == TCP_CA_Open) { 2590 BUG_TRAP(tp->retrans_out == 0); 2591 tp->retrans_stamp = 0; 2592 } else if (!before(tp->snd_una, tp->high_seq)) { 2593 switch (icsk->icsk_ca_state) { 2594 case TCP_CA_Loss: 2595 icsk->icsk_retransmits = 0; 2596 if (tcp_try_undo_recovery(sk)) 2597 return; 2598 break; 2599 2600 case TCP_CA_CWR: 2601 /* CWR is to be held something *above* high_seq 2602 * is ACKed for CWR bit to reach receiver. */ 2603 if (tp->snd_una != tp->high_seq) { 2604 tcp_complete_cwr(sk); 2605 tcp_set_ca_state(sk, TCP_CA_Open); 2606 } 2607 break; 2608 2609 case TCP_CA_Disorder: 2610 tcp_try_undo_dsack(sk); 2611 if (!tp->undo_marker || 2612 /* For SACK case do not Open to allow to undo 2613 * catching for all duplicate ACKs. */ 2614 tcp_is_reno(tp) || tp->snd_una != tp->high_seq) { 2615 tp->undo_marker = 0; 2616 tcp_set_ca_state(sk, TCP_CA_Open); 2617 } 2618 break; 2619 2620 case TCP_CA_Recovery: 2621 if (tcp_is_reno(tp)) 2622 tcp_reset_reno_sack(tp); 2623 if (tcp_try_undo_recovery(sk)) 2624 return; 2625 tcp_complete_cwr(sk); 2626 break; 2627 } 2628 } 2629 2630 /* F. Process state. */ 2631 switch (icsk->icsk_ca_state) { 2632 case TCP_CA_Recovery: 2633 if (!(flag & FLAG_SND_UNA_ADVANCED)) { 2634 if (tcp_is_reno(tp) && is_dupack) 2635 tcp_add_reno_sack(sk); 2636 } else 2637 do_lost = tcp_try_undo_partial(sk, pkts_acked); 2638 break; 2639 case TCP_CA_Loss: 2640 if (flag & FLAG_DATA_ACKED) 2641 icsk->icsk_retransmits = 0; 2642 if (tcp_is_reno(tp) && flag & FLAG_SND_UNA_ADVANCED) 2643 tcp_reset_reno_sack(tp); 2644 if (!tcp_try_undo_loss(sk)) { 2645 tcp_moderate_cwnd(tp); 2646 tcp_xmit_retransmit_queue(sk); 2647 return; 2648 } 2649 if (icsk->icsk_ca_state != TCP_CA_Open) 2650 return; 2651 /* Loss is undone; fall through to processing in Open state. */ 2652 default: 2653 if (tcp_is_reno(tp)) { 2654 if (flag & FLAG_SND_UNA_ADVANCED) 2655 tcp_reset_reno_sack(tp); 2656 if (is_dupack) 2657 tcp_add_reno_sack(sk); 2658 } 2659 2660 if (icsk->icsk_ca_state == TCP_CA_Disorder) 2661 tcp_try_undo_dsack(sk); 2662 2663 if (!tcp_time_to_recover(sk)) { 2664 tcp_try_to_open(sk, flag); 2665 return; 2666 } 2667 2668 /* MTU probe failure: don't reduce cwnd */ 2669 if (icsk->icsk_ca_state < TCP_CA_CWR && 2670 icsk->icsk_mtup.probe_size && 2671 tp->snd_una == tp->mtu_probe.probe_seq_start) { 2672 tcp_mtup_probe_failed(sk); 2673 /* Restores the reduction we did in tcp_mtup_probe() */ 2674 tp->snd_cwnd++; 2675 tcp_simple_retransmit(sk); 2676 return; 2677 } 2678 2679 /* Otherwise enter Recovery state */ 2680 2681 if (tcp_is_reno(tp)) 2682 NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY); 2683 else 2684 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY); 2685 2686 tp->high_seq = tp->snd_nxt; 2687 tp->prior_ssthresh = 0; 2688 tp->undo_marker = tp->snd_una; 2689 tp->undo_retrans = tp->retrans_out; 2690 2691 if (icsk->icsk_ca_state < TCP_CA_CWR) { 2692 if (!(flag & FLAG_ECE)) 2693 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2694 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 2695 TCP_ECN_queue_cwr(tp); 2696 } 2697 2698 tp->bytes_acked = 0; 2699 tp->snd_cwnd_cnt = 0; 2700 tcp_set_ca_state(sk, TCP_CA_Recovery); 2701 fast_rexmit = 1; 2702 } 2703 2704 if (do_lost || (tcp_is_fack(tp) && tcp_head_timedout(sk))) 2705 tcp_update_scoreboard(sk, fast_rexmit); 2706 tcp_cwnd_down(sk, flag); 2707 tcp_xmit_retransmit_queue(sk); 2708 } 2709 2710 /* Read draft-ietf-tcplw-high-performance before mucking 2711 * with this code. (Supersedes RFC1323) 2712 */ 2713 static void tcp_ack_saw_tstamp(struct sock *sk, int flag) 2714 { 2715 /* RTTM Rule: A TSecr value received in a segment is used to 2716 * update the averaged RTT measurement only if the segment 2717 * acknowledges some new data, i.e., only if it advances the 2718 * left edge of the send window. 2719 * 2720 * See draft-ietf-tcplw-high-performance-00, section 3.3. 2721 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru> 2722 * 2723 * Changed: reset backoff as soon as we see the first valid sample. 2724 * If we do not, we get strongly overestimated rto. With timestamps 2725 * samples are accepted even from very old segments: f.e., when rtt=1 2726 * increases to 8, we retransmit 5 times and after 8 seconds delayed 2727 * answer arrives rto becomes 120 seconds! If at least one of segments 2728 * in window is lost... Voila. --ANK (010210) 2729 */ 2730 struct tcp_sock *tp = tcp_sk(sk); 2731 const __u32 seq_rtt = tcp_time_stamp - tp->rx_opt.rcv_tsecr; 2732 tcp_rtt_estimator(sk, seq_rtt); 2733 tcp_set_rto(sk); 2734 inet_csk(sk)->icsk_backoff = 0; 2735 tcp_bound_rto(sk); 2736 } 2737 2738 static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag) 2739 { 2740 /* We don't have a timestamp. Can only use 2741 * packets that are not retransmitted to determine 2742 * rtt estimates. Also, we must not reset the 2743 * backoff for rto until we get a non-retransmitted 2744 * packet. This allows us to deal with a situation 2745 * where the network delay has increased suddenly. 2746 * I.e. Karn's algorithm. (SIGCOMM '87, p5.) 2747 */ 2748 2749 if (flag & FLAG_RETRANS_DATA_ACKED) 2750 return; 2751 2752 tcp_rtt_estimator(sk, seq_rtt); 2753 tcp_set_rto(sk); 2754 inet_csk(sk)->icsk_backoff = 0; 2755 tcp_bound_rto(sk); 2756 } 2757 2758 static inline void tcp_ack_update_rtt(struct sock *sk, const int flag, 2759 const s32 seq_rtt) 2760 { 2761 const struct tcp_sock *tp = tcp_sk(sk); 2762 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */ 2763 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr) 2764 tcp_ack_saw_tstamp(sk, flag); 2765 else if (seq_rtt >= 0) 2766 tcp_ack_no_tstamp(sk, seq_rtt, flag); 2767 } 2768 2769 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight) 2770 { 2771 const struct inet_connection_sock *icsk = inet_csk(sk); 2772 icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight); 2773 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp; 2774 } 2775 2776 /* Restart timer after forward progress on connection. 2777 * RFC2988 recommends to restart timer to now+rto. 2778 */ 2779 static void tcp_rearm_rto(struct sock *sk) 2780 { 2781 struct tcp_sock *tp = tcp_sk(sk); 2782 2783 if (!tp->packets_out) { 2784 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS); 2785 } else { 2786 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 2787 inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 2788 } 2789 } 2790 2791 /* If we get here, the whole TSO packet has not been acked. */ 2792 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb) 2793 { 2794 struct tcp_sock *tp = tcp_sk(sk); 2795 u32 packets_acked; 2796 2797 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)); 2798 2799 packets_acked = tcp_skb_pcount(skb); 2800 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 2801 return 0; 2802 packets_acked -= tcp_skb_pcount(skb); 2803 2804 if (packets_acked) { 2805 BUG_ON(tcp_skb_pcount(skb) == 0); 2806 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)); 2807 } 2808 2809 return packets_acked; 2810 } 2811 2812 /* Remove acknowledged frames from the retransmission queue. If our packet 2813 * is before the ack sequence we can discard it as it's confirmed to have 2814 * arrived at the other end. 2815 */ 2816 static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets) 2817 { 2818 struct tcp_sock *tp = tcp_sk(sk); 2819 const struct inet_connection_sock *icsk = inet_csk(sk); 2820 struct sk_buff *skb; 2821 u32 now = tcp_time_stamp; 2822 int fully_acked = 1; 2823 int flag = 0; 2824 u32 pkts_acked = 0; 2825 u32 reord = tp->packets_out; 2826 s32 seq_rtt = -1; 2827 s32 ca_seq_rtt = -1; 2828 ktime_t last_ackt = net_invalid_timestamp(); 2829 2830 while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) { 2831 struct tcp_skb_cb *scb = TCP_SKB_CB(skb); 2832 u32 end_seq; 2833 u32 acked_pcount; 2834 u8 sacked = scb->sacked; 2835 2836 /* Determine how many packets and what bytes were acked, tso and else */ 2837 if (after(scb->end_seq, tp->snd_una)) { 2838 if (tcp_skb_pcount(skb) == 1 || 2839 !after(tp->snd_una, scb->seq)) 2840 break; 2841 2842 acked_pcount = tcp_tso_acked(sk, skb); 2843 if (!acked_pcount) 2844 break; 2845 2846 fully_acked = 0; 2847 end_seq = tp->snd_una; 2848 } else { 2849 acked_pcount = tcp_skb_pcount(skb); 2850 end_seq = scb->end_seq; 2851 } 2852 2853 /* MTU probing checks */ 2854 if (fully_acked && icsk->icsk_mtup.probe_size && 2855 !after(tp->mtu_probe.probe_seq_end, scb->end_seq)) { 2856 tcp_mtup_probe_success(sk, skb); 2857 } 2858 2859 if (sacked & TCPCB_RETRANS) { 2860 if (sacked & TCPCB_SACKED_RETRANS) 2861 tp->retrans_out -= acked_pcount; 2862 flag |= FLAG_RETRANS_DATA_ACKED; 2863 ca_seq_rtt = -1; 2864 seq_rtt = -1; 2865 if ((flag & FLAG_DATA_ACKED) || (acked_pcount > 1)) 2866 flag |= FLAG_NONHEAD_RETRANS_ACKED; 2867 } else { 2868 ca_seq_rtt = now - scb->when; 2869 last_ackt = skb->tstamp; 2870 if (seq_rtt < 0) { 2871 seq_rtt = ca_seq_rtt; 2872 } 2873 if (!(sacked & TCPCB_SACKED_ACKED)) 2874 reord = min(pkts_acked, reord); 2875 } 2876 2877 if (sacked & TCPCB_SACKED_ACKED) 2878 tp->sacked_out -= acked_pcount; 2879 if (sacked & TCPCB_LOST) 2880 tp->lost_out -= acked_pcount; 2881 2882 if (unlikely(tp->urg_mode && !before(end_seq, tp->snd_up))) 2883 tp->urg_mode = 0; 2884 2885 tp->packets_out -= acked_pcount; 2886 pkts_acked += acked_pcount; 2887 2888 /* Initial outgoing SYN's get put onto the write_queue 2889 * just like anything else we transmit. It is not 2890 * true data, and if we misinform our callers that 2891 * this ACK acks real data, we will erroneously exit 2892 * connection startup slow start one packet too 2893 * quickly. This is severely frowned upon behavior. 2894 */ 2895 if (!(scb->flags & TCPCB_FLAG_SYN)) { 2896 flag |= FLAG_DATA_ACKED; 2897 } else { 2898 flag |= FLAG_SYN_ACKED; 2899 tp->retrans_stamp = 0; 2900 } 2901 2902 if (!fully_acked) 2903 break; 2904 2905 tcp_unlink_write_queue(skb, sk); 2906 sk_wmem_free_skb(sk, skb); 2907 tcp_clear_all_retrans_hints(tp); 2908 } 2909 2910 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) 2911 flag |= FLAG_SACK_RENEGING; 2912 2913 if (flag & FLAG_ACKED) { 2914 const struct tcp_congestion_ops *ca_ops 2915 = inet_csk(sk)->icsk_ca_ops; 2916 2917 tcp_ack_update_rtt(sk, flag, seq_rtt); 2918 tcp_rearm_rto(sk); 2919 2920 if (tcp_is_reno(tp)) { 2921 tcp_remove_reno_sacks(sk, pkts_acked); 2922 } else { 2923 /* Non-retransmitted hole got filled? That's reordering */ 2924 if (reord < prior_fackets) 2925 tcp_update_reordering(sk, tp->fackets_out - reord, 0); 2926 } 2927 2928 tp->fackets_out -= min(pkts_acked, tp->fackets_out); 2929 2930 if (ca_ops->pkts_acked) { 2931 s32 rtt_us = -1; 2932 2933 /* Is the ACK triggering packet unambiguous? */ 2934 if (!(flag & FLAG_RETRANS_DATA_ACKED)) { 2935 /* High resolution needed and available? */ 2936 if (ca_ops->flags & TCP_CONG_RTT_STAMP && 2937 !ktime_equal(last_ackt, 2938 net_invalid_timestamp())) 2939 rtt_us = ktime_us_delta(ktime_get_real(), 2940 last_ackt); 2941 else if (ca_seq_rtt > 0) 2942 rtt_us = jiffies_to_usecs(ca_seq_rtt); 2943 } 2944 2945 ca_ops->pkts_acked(sk, pkts_acked, rtt_us); 2946 } 2947 } 2948 2949 #if FASTRETRANS_DEBUG > 0 2950 BUG_TRAP((int)tp->sacked_out >= 0); 2951 BUG_TRAP((int)tp->lost_out >= 0); 2952 BUG_TRAP((int)tp->retrans_out >= 0); 2953 if (!tp->packets_out && tcp_is_sack(tp)) { 2954 icsk = inet_csk(sk); 2955 if (tp->lost_out) { 2956 printk(KERN_DEBUG "Leak l=%u %d\n", 2957 tp->lost_out, icsk->icsk_ca_state); 2958 tp->lost_out = 0; 2959 } 2960 if (tp->sacked_out) { 2961 printk(KERN_DEBUG "Leak s=%u %d\n", 2962 tp->sacked_out, icsk->icsk_ca_state); 2963 tp->sacked_out = 0; 2964 } 2965 if (tp->retrans_out) { 2966 printk(KERN_DEBUG "Leak r=%u %d\n", 2967 tp->retrans_out, icsk->icsk_ca_state); 2968 tp->retrans_out = 0; 2969 } 2970 } 2971 #endif 2972 return flag; 2973 } 2974 2975 static void tcp_ack_probe(struct sock *sk) 2976 { 2977 const struct tcp_sock *tp = tcp_sk(sk); 2978 struct inet_connection_sock *icsk = inet_csk(sk); 2979 2980 /* Was it a usable window open? */ 2981 2982 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) { 2983 icsk->icsk_backoff = 0; 2984 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0); 2985 /* Socket must be waked up by subsequent tcp_data_snd_check(). 2986 * This function is not for random using! 2987 */ 2988 } else { 2989 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 2990 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), 2991 TCP_RTO_MAX); 2992 } 2993 } 2994 2995 static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag) 2996 { 2997 return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) || 2998 inet_csk(sk)->icsk_ca_state != TCP_CA_Open); 2999 } 3000 3001 static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag) 3002 { 3003 const struct tcp_sock *tp = tcp_sk(sk); 3004 return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) && 3005 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR)); 3006 } 3007 3008 /* Check that window update is acceptable. 3009 * The function assumes that snd_una<=ack<=snd_next. 3010 */ 3011 static inline int tcp_may_update_window(const struct tcp_sock *tp, 3012 const u32 ack, const u32 ack_seq, 3013 const u32 nwin) 3014 { 3015 return (after(ack, tp->snd_una) || 3016 after(ack_seq, tp->snd_wl1) || 3017 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd)); 3018 } 3019 3020 /* Update our send window. 3021 * 3022 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2 3023 * and in FreeBSD. NetBSD's one is even worse.) is wrong. 3024 */ 3025 static int tcp_ack_update_window(struct sock *sk, struct sk_buff *skb, u32 ack, 3026 u32 ack_seq) 3027 { 3028 struct tcp_sock *tp = tcp_sk(sk); 3029 int flag = 0; 3030 u32 nwin = ntohs(tcp_hdr(skb)->window); 3031 3032 if (likely(!tcp_hdr(skb)->syn)) 3033 nwin <<= tp->rx_opt.snd_wscale; 3034 3035 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) { 3036 flag |= FLAG_WIN_UPDATE; 3037 tcp_update_wl(tp, ack, ack_seq); 3038 3039 if (tp->snd_wnd != nwin) { 3040 tp->snd_wnd = nwin; 3041 3042 /* Note, it is the only place, where 3043 * fast path is recovered for sending TCP. 3044 */ 3045 tp->pred_flags = 0; 3046 tcp_fast_path_check(sk); 3047 3048 if (nwin > tp->max_window) { 3049 tp->max_window = nwin; 3050 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie); 3051 } 3052 } 3053 } 3054 3055 tp->snd_una = ack; 3056 3057 return flag; 3058 } 3059 3060 /* A very conservative spurious RTO response algorithm: reduce cwnd and 3061 * continue in congestion avoidance. 3062 */ 3063 static void tcp_conservative_spur_to_response(struct tcp_sock *tp) 3064 { 3065 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh); 3066 tp->snd_cwnd_cnt = 0; 3067 tp->bytes_acked = 0; 3068 TCP_ECN_queue_cwr(tp); 3069 tcp_moderate_cwnd(tp); 3070 } 3071 3072 /* A conservative spurious RTO response algorithm: reduce cwnd using 3073 * rate halving and continue in congestion avoidance. 3074 */ 3075 static void tcp_ratehalving_spur_to_response(struct sock *sk) 3076 { 3077 tcp_enter_cwr(sk, 0); 3078 } 3079 3080 static void tcp_undo_spur_to_response(struct sock *sk, int flag) 3081 { 3082 if (flag & FLAG_ECE) 3083 tcp_ratehalving_spur_to_response(sk); 3084 else 3085 tcp_undo_cwr(sk, 1); 3086 } 3087 3088 /* F-RTO spurious RTO detection algorithm (RFC4138) 3089 * 3090 * F-RTO affects during two new ACKs following RTO (well, almost, see inline 3091 * comments). State (ACK number) is kept in frto_counter. When ACK advances 3092 * window (but not to or beyond highest sequence sent before RTO): 3093 * On First ACK, send two new segments out. 3094 * On Second ACK, RTO was likely spurious. Do spurious response (response 3095 * algorithm is not part of the F-RTO detection algorithm 3096 * given in RFC4138 but can be selected separately). 3097 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss 3098 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding 3099 * of Nagle, this is done using frto_counter states 2 and 3, when a new data 3100 * segment of any size sent during F-RTO, state 2 is upgraded to 3. 3101 * 3102 * Rationale: if the RTO was spurious, new ACKs should arrive from the 3103 * original window even after we transmit two new data segments. 3104 * 3105 * SACK version: 3106 * on first step, wait until first cumulative ACK arrives, then move to 3107 * the second step. In second step, the next ACK decides. 3108 * 3109 * F-RTO is implemented (mainly) in four functions: 3110 * - tcp_use_frto() is used to determine if TCP is can use F-RTO 3111 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is 3112 * called when tcp_use_frto() showed green light 3113 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm 3114 * - tcp_enter_frto_loss() is called if there is not enough evidence 3115 * to prove that the RTO is indeed spurious. It transfers the control 3116 * from F-RTO to the conventional RTO recovery 3117 */ 3118 static int tcp_process_frto(struct sock *sk, int flag) 3119 { 3120 struct tcp_sock *tp = tcp_sk(sk); 3121 3122 tcp_verify_left_out(tp); 3123 3124 /* Duplicate the behavior from Loss state (fastretrans_alert) */ 3125 if (flag & FLAG_DATA_ACKED) 3126 inet_csk(sk)->icsk_retransmits = 0; 3127 3128 if ((flag & FLAG_NONHEAD_RETRANS_ACKED) || 3129 ((tp->frto_counter >= 2) && (flag & FLAG_RETRANS_DATA_ACKED))) 3130 tp->undo_marker = 0; 3131 3132 if (!before(tp->snd_una, tp->frto_highmark)) { 3133 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag); 3134 return 1; 3135 } 3136 3137 if (!tcp_is_sackfrto(tp)) { 3138 /* RFC4138 shortcoming in step 2; should also have case c): 3139 * ACK isn't duplicate nor advances window, e.g., opposite dir 3140 * data, winupdate 3141 */ 3142 if (!(flag & FLAG_ANY_PROGRESS) && (flag & FLAG_NOT_DUP)) 3143 return 1; 3144 3145 if (!(flag & FLAG_DATA_ACKED)) { 3146 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3), 3147 flag); 3148 return 1; 3149 } 3150 } else { 3151 if (!(flag & FLAG_DATA_ACKED) && (tp->frto_counter == 1)) { 3152 /* Prevent sending of new data. */ 3153 tp->snd_cwnd = min(tp->snd_cwnd, 3154 tcp_packets_in_flight(tp)); 3155 return 1; 3156 } 3157 3158 if ((tp->frto_counter >= 2) && 3159 (!(flag & FLAG_FORWARD_PROGRESS) || 3160 ((flag & FLAG_DATA_SACKED) && 3161 !(flag & FLAG_ONLY_ORIG_SACKED)))) { 3162 /* RFC4138 shortcoming (see comment above) */ 3163 if (!(flag & FLAG_FORWARD_PROGRESS) && 3164 (flag & FLAG_NOT_DUP)) 3165 return 1; 3166 3167 tcp_enter_frto_loss(sk, 3, flag); 3168 return 1; 3169 } 3170 } 3171 3172 if (tp->frto_counter == 1) { 3173 /* tcp_may_send_now needs to see updated state */ 3174 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2; 3175 tp->frto_counter = 2; 3176 3177 if (!tcp_may_send_now(sk)) 3178 tcp_enter_frto_loss(sk, 2, flag); 3179 3180 return 1; 3181 } else { 3182 switch (sysctl_tcp_frto_response) { 3183 case 2: 3184 tcp_undo_spur_to_response(sk, flag); 3185 break; 3186 case 1: 3187 tcp_conservative_spur_to_response(tp); 3188 break; 3189 default: 3190 tcp_ratehalving_spur_to_response(sk); 3191 break; 3192 } 3193 tp->frto_counter = 0; 3194 tp->undo_marker = 0; 3195 NET_INC_STATS_BH(LINUX_MIB_TCPSPURIOUSRTOS); 3196 } 3197 return 0; 3198 } 3199 3200 /* This routine deals with incoming acks, but not outgoing ones. */ 3201 static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag) 3202 { 3203 struct inet_connection_sock *icsk = inet_csk(sk); 3204 struct tcp_sock *tp = tcp_sk(sk); 3205 u32 prior_snd_una = tp->snd_una; 3206 u32 ack_seq = TCP_SKB_CB(skb)->seq; 3207 u32 ack = TCP_SKB_CB(skb)->ack_seq; 3208 u32 prior_in_flight; 3209 u32 prior_fackets; 3210 int prior_packets; 3211 int frto_cwnd = 0; 3212 3213 /* If the ack is newer than sent or older than previous acks 3214 * then we can probably ignore it. 3215 */ 3216 if (after(ack, tp->snd_nxt)) 3217 goto uninteresting_ack; 3218 3219 if (before(ack, prior_snd_una)) 3220 goto old_ack; 3221 3222 if (after(ack, prior_snd_una)) 3223 flag |= FLAG_SND_UNA_ADVANCED; 3224 3225 if (sysctl_tcp_abc) { 3226 if (icsk->icsk_ca_state < TCP_CA_CWR) 3227 tp->bytes_acked += ack - prior_snd_una; 3228 else if (icsk->icsk_ca_state == TCP_CA_Loss) 3229 /* we assume just one segment left network */ 3230 tp->bytes_acked += min(ack - prior_snd_una, 3231 tp->mss_cache); 3232 } 3233 3234 prior_fackets = tp->fackets_out; 3235 prior_in_flight = tcp_packets_in_flight(tp); 3236 3237 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) { 3238 /* Window is constant, pure forward advance. 3239 * No more checks are required. 3240 * Note, we use the fact that SND.UNA>=SND.WL2. 3241 */ 3242 tcp_update_wl(tp, ack, ack_seq); 3243 tp->snd_una = ack; 3244 flag |= FLAG_WIN_UPDATE; 3245 3246 tcp_ca_event(sk, CA_EVENT_FAST_ACK); 3247 3248 NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS); 3249 } else { 3250 if (ack_seq != TCP_SKB_CB(skb)->end_seq) 3251 flag |= FLAG_DATA; 3252 else 3253 NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS); 3254 3255 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq); 3256 3257 if (TCP_SKB_CB(skb)->sacked) 3258 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una); 3259 3260 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb))) 3261 flag |= FLAG_ECE; 3262 3263 tcp_ca_event(sk, CA_EVENT_SLOW_ACK); 3264 } 3265 3266 /* We passed data and got it acked, remove any soft error 3267 * log. Something worked... 3268 */ 3269 sk->sk_err_soft = 0; 3270 tp->rcv_tstamp = tcp_time_stamp; 3271 prior_packets = tp->packets_out; 3272 if (!prior_packets) 3273 goto no_queue; 3274 3275 /* See if we can take anything off of the retransmit queue. */ 3276 flag |= tcp_clean_rtx_queue(sk, prior_fackets); 3277 3278 if (tp->frto_counter) 3279 frto_cwnd = tcp_process_frto(sk, flag); 3280 /* Guarantee sacktag reordering detection against wrap-arounds */ 3281 if (before(tp->frto_highmark, tp->snd_una)) 3282 tp->frto_highmark = 0; 3283 3284 if (tcp_ack_is_dubious(sk, flag)) { 3285 /* Advance CWND, if state allows this. */ 3286 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd && 3287 tcp_may_raise_cwnd(sk, flag)) 3288 tcp_cong_avoid(sk, ack, prior_in_flight); 3289 tcp_fastretrans_alert(sk, prior_packets - tp->packets_out, 3290 flag); 3291 } else { 3292 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd) 3293 tcp_cong_avoid(sk, ack, prior_in_flight); 3294 } 3295 3296 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) 3297 dst_confirm(sk->sk_dst_cache); 3298 3299 return 1; 3300 3301 no_queue: 3302 icsk->icsk_probes_out = 0; 3303 3304 /* If this ack opens up a zero window, clear backoff. It was 3305 * being used to time the probes, and is probably far higher than 3306 * it needs to be for normal retransmission. 3307 */ 3308 if (tcp_send_head(sk)) 3309 tcp_ack_probe(sk); 3310 return 1; 3311 3312 old_ack: 3313 if (TCP_SKB_CB(skb)->sacked) 3314 tcp_sacktag_write_queue(sk, skb, prior_snd_una); 3315 3316 uninteresting_ack: 3317 SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt); 3318 return 0; 3319 } 3320 3321 /* Look for tcp options. Normally only called on SYN and SYNACK packets. 3322 * But, this can also be called on packets in the established flow when 3323 * the fast version below fails. 3324 */ 3325 void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx, 3326 int estab) 3327 { 3328 unsigned char *ptr; 3329 struct tcphdr *th = tcp_hdr(skb); 3330 int length = (th->doff * 4) - sizeof(struct tcphdr); 3331 3332 ptr = (unsigned char *)(th + 1); 3333 opt_rx->saw_tstamp = 0; 3334 3335 while (length > 0) { 3336 int opcode = *ptr++; 3337 int opsize; 3338 3339 switch (opcode) { 3340 case TCPOPT_EOL: 3341 return; 3342 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */ 3343 length--; 3344 continue; 3345 default: 3346 opsize = *ptr++; 3347 if (opsize < 2) /* "silly options" */ 3348 return; 3349 if (opsize > length) 3350 return; /* don't parse partial options */ 3351 switch (opcode) { 3352 case TCPOPT_MSS: 3353 if (opsize == TCPOLEN_MSS && th->syn && !estab) { 3354 u16 in_mss = get_unaligned_be16(ptr); 3355 if (in_mss) { 3356 if (opt_rx->user_mss && 3357 opt_rx->user_mss < in_mss) 3358 in_mss = opt_rx->user_mss; 3359 opt_rx->mss_clamp = in_mss; 3360 } 3361 } 3362 break; 3363 case TCPOPT_WINDOW: 3364 if (opsize == TCPOLEN_WINDOW && th->syn && 3365 !estab && sysctl_tcp_window_scaling) { 3366 __u8 snd_wscale = *(__u8 *)ptr; 3367 opt_rx->wscale_ok = 1; 3368 if (snd_wscale > 14) { 3369 if (net_ratelimit()) 3370 printk(KERN_INFO "tcp_parse_options: Illegal window " 3371 "scaling value %d >14 received.\n", 3372 snd_wscale); 3373 snd_wscale = 14; 3374 } 3375 opt_rx->snd_wscale = snd_wscale; 3376 } 3377 break; 3378 case TCPOPT_TIMESTAMP: 3379 if ((opsize == TCPOLEN_TIMESTAMP) && 3380 ((estab && opt_rx->tstamp_ok) || 3381 (!estab && sysctl_tcp_timestamps))) { 3382 opt_rx->saw_tstamp = 1; 3383 opt_rx->rcv_tsval = get_unaligned_be32(ptr); 3384 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4); 3385 } 3386 break; 3387 case TCPOPT_SACK_PERM: 3388 if (opsize == TCPOLEN_SACK_PERM && th->syn && 3389 !estab && sysctl_tcp_sack) { 3390 opt_rx->sack_ok = 1; 3391 tcp_sack_reset(opt_rx); 3392 } 3393 break; 3394 3395 case TCPOPT_SACK: 3396 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) && 3397 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) && 3398 opt_rx->sack_ok) { 3399 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th; 3400 } 3401 break; 3402 #ifdef CONFIG_TCP_MD5SIG 3403 case TCPOPT_MD5SIG: 3404 /* 3405 * The MD5 Hash has already been 3406 * checked (see tcp_v{4,6}_do_rcv()). 3407 */ 3408 break; 3409 #endif 3410 } 3411 3412 ptr += opsize-2; 3413 length -= opsize; 3414 } 3415 } 3416 } 3417 3418 /* Fast parse options. This hopes to only see timestamps. 3419 * If it is wrong it falls back on tcp_parse_options(). 3420 */ 3421 static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th, 3422 struct tcp_sock *tp) 3423 { 3424 if (th->doff == sizeof(struct tcphdr) >> 2) { 3425 tp->rx_opt.saw_tstamp = 0; 3426 return 0; 3427 } else if (tp->rx_opt.tstamp_ok && 3428 th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) { 3429 __be32 *ptr = (__be32 *)(th + 1); 3430 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) 3431 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) { 3432 tp->rx_opt.saw_tstamp = 1; 3433 ++ptr; 3434 tp->rx_opt.rcv_tsval = ntohl(*ptr); 3435 ++ptr; 3436 tp->rx_opt.rcv_tsecr = ntohl(*ptr); 3437 return 1; 3438 } 3439 } 3440 tcp_parse_options(skb, &tp->rx_opt, 1); 3441 return 1; 3442 } 3443 3444 static inline void tcp_store_ts_recent(struct tcp_sock *tp) 3445 { 3446 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval; 3447 tp->rx_opt.ts_recent_stamp = get_seconds(); 3448 } 3449 3450 static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq) 3451 { 3452 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) { 3453 /* PAWS bug workaround wrt. ACK frames, the PAWS discard 3454 * extra check below makes sure this can only happen 3455 * for pure ACK frames. -DaveM 3456 * 3457 * Not only, also it occurs for expired timestamps. 3458 */ 3459 3460 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) >= 0 || 3461 get_seconds() >= tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS) 3462 tcp_store_ts_recent(tp); 3463 } 3464 } 3465 3466 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM 3467 * 3468 * It is not fatal. If this ACK does _not_ change critical state (seqs, window) 3469 * it can pass through stack. So, the following predicate verifies that 3470 * this segment is not used for anything but congestion avoidance or 3471 * fast retransmit. Moreover, we even are able to eliminate most of such 3472 * second order effects, if we apply some small "replay" window (~RTO) 3473 * to timestamp space. 3474 * 3475 * All these measures still do not guarantee that we reject wrapped ACKs 3476 * on networks with high bandwidth, when sequence space is recycled fastly, 3477 * but it guarantees that such events will be very rare and do not affect 3478 * connection seriously. This doesn't look nice, but alas, PAWS is really 3479 * buggy extension. 3480 * 3481 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC 3482 * states that events when retransmit arrives after original data are rare. 3483 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is 3484 * the biggest problem on large power networks even with minor reordering. 3485 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe 3486 * up to bandwidth of 18Gigabit/sec. 8) ] 3487 */ 3488 3489 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb) 3490 { 3491 struct tcp_sock *tp = tcp_sk(sk); 3492 struct tcphdr *th = tcp_hdr(skb); 3493 u32 seq = TCP_SKB_CB(skb)->seq; 3494 u32 ack = TCP_SKB_CB(skb)->ack_seq; 3495 3496 return (/* 1. Pure ACK with correct sequence number. */ 3497 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) && 3498 3499 /* 2. ... and duplicate ACK. */ 3500 ack == tp->snd_una && 3501 3502 /* 3. ... and does not update window. */ 3503 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) && 3504 3505 /* 4. ... and sits in replay window. */ 3506 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ); 3507 } 3508 3509 static inline int tcp_paws_discard(const struct sock *sk, 3510 const struct sk_buff *skb) 3511 { 3512 const struct tcp_sock *tp = tcp_sk(sk); 3513 return ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > TCP_PAWS_WINDOW && 3514 get_seconds() < tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS && 3515 !tcp_disordered_ack(sk, skb)); 3516 } 3517 3518 /* Check segment sequence number for validity. 3519 * 3520 * Segment controls are considered valid, if the segment 3521 * fits to the window after truncation to the window. Acceptability 3522 * of data (and SYN, FIN, of course) is checked separately. 3523 * See tcp_data_queue(), for example. 3524 * 3525 * Also, controls (RST is main one) are accepted using RCV.WUP instead 3526 * of RCV.NXT. Peer still did not advance his SND.UNA when we 3527 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP. 3528 * (borrowed from freebsd) 3529 */ 3530 3531 static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq) 3532 { 3533 return !before(end_seq, tp->rcv_wup) && 3534 !after(seq, tp->rcv_nxt + tcp_receive_window(tp)); 3535 } 3536 3537 /* When we get a reset we do this. */ 3538 static void tcp_reset(struct sock *sk) 3539 { 3540 /* We want the right error as BSD sees it (and indeed as we do). */ 3541 switch (sk->sk_state) { 3542 case TCP_SYN_SENT: 3543 sk->sk_err = ECONNREFUSED; 3544 break; 3545 case TCP_CLOSE_WAIT: 3546 sk->sk_err = EPIPE; 3547 break; 3548 case TCP_CLOSE: 3549 return; 3550 default: 3551 sk->sk_err = ECONNRESET; 3552 } 3553 3554 if (!sock_flag(sk, SOCK_DEAD)) 3555 sk->sk_error_report(sk); 3556 3557 tcp_done(sk); 3558 } 3559 3560 /* 3561 * Process the FIN bit. This now behaves as it is supposed to work 3562 * and the FIN takes effect when it is validly part of sequence 3563 * space. Not before when we get holes. 3564 * 3565 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT 3566 * (and thence onto LAST-ACK and finally, CLOSE, we never enter 3567 * TIME-WAIT) 3568 * 3569 * If we are in FINWAIT-1, a received FIN indicates simultaneous 3570 * close and we go into CLOSING (and later onto TIME-WAIT) 3571 * 3572 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT. 3573 */ 3574 static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th) 3575 { 3576 struct tcp_sock *tp = tcp_sk(sk); 3577 3578 inet_csk_schedule_ack(sk); 3579 3580 sk->sk_shutdown |= RCV_SHUTDOWN; 3581 sock_set_flag(sk, SOCK_DONE); 3582 3583 switch (sk->sk_state) { 3584 case TCP_SYN_RECV: 3585 case TCP_ESTABLISHED: 3586 /* Move to CLOSE_WAIT */ 3587 tcp_set_state(sk, TCP_CLOSE_WAIT); 3588 inet_csk(sk)->icsk_ack.pingpong = 1; 3589 break; 3590 3591 case TCP_CLOSE_WAIT: 3592 case TCP_CLOSING: 3593 /* Received a retransmission of the FIN, do 3594 * nothing. 3595 */ 3596 break; 3597 case TCP_LAST_ACK: 3598 /* RFC793: Remain in the LAST-ACK state. */ 3599 break; 3600 3601 case TCP_FIN_WAIT1: 3602 /* This case occurs when a simultaneous close 3603 * happens, we must ack the received FIN and 3604 * enter the CLOSING state. 3605 */ 3606 tcp_send_ack(sk); 3607 tcp_set_state(sk, TCP_CLOSING); 3608 break; 3609 case TCP_FIN_WAIT2: 3610 /* Received a FIN -- send ACK and enter TIME_WAIT. */ 3611 tcp_send_ack(sk); 3612 tcp_time_wait(sk, TCP_TIME_WAIT, 0); 3613 break; 3614 default: 3615 /* Only TCP_LISTEN and TCP_CLOSE are left, in these 3616 * cases we should never reach this piece of code. 3617 */ 3618 printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n", 3619 __func__, sk->sk_state); 3620 break; 3621 } 3622 3623 /* It _is_ possible, that we have something out-of-order _after_ FIN. 3624 * Probably, we should reset in this case. For now drop them. 3625 */ 3626 __skb_queue_purge(&tp->out_of_order_queue); 3627 if (tcp_is_sack(tp)) 3628 tcp_sack_reset(&tp->rx_opt); 3629 sk_mem_reclaim(sk); 3630 3631 if (!sock_flag(sk, SOCK_DEAD)) { 3632 sk->sk_state_change(sk); 3633 3634 /* Do not send POLL_HUP for half duplex close. */ 3635 if (sk->sk_shutdown == SHUTDOWN_MASK || 3636 sk->sk_state == TCP_CLOSE) 3637 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP); 3638 else 3639 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN); 3640 } 3641 } 3642 3643 static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, 3644 u32 end_seq) 3645 { 3646 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) { 3647 if (before(seq, sp->start_seq)) 3648 sp->start_seq = seq; 3649 if (after(end_seq, sp->end_seq)) 3650 sp->end_seq = end_seq; 3651 return 1; 3652 } 3653 return 0; 3654 } 3655 3656 static void tcp_dsack_set(struct tcp_sock *tp, u32 seq, u32 end_seq) 3657 { 3658 if (tcp_is_sack(tp) && sysctl_tcp_dsack) { 3659 if (before(seq, tp->rcv_nxt)) 3660 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT); 3661 else 3662 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT); 3663 3664 tp->rx_opt.dsack = 1; 3665 tp->duplicate_sack[0].start_seq = seq; 3666 tp->duplicate_sack[0].end_seq = end_seq; 3667 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + 1, 3668 4 - tp->rx_opt.tstamp_ok); 3669 } 3670 } 3671 3672 static void tcp_dsack_extend(struct tcp_sock *tp, u32 seq, u32 end_seq) 3673 { 3674 if (!tp->rx_opt.dsack) 3675 tcp_dsack_set(tp, seq, end_seq); 3676 else 3677 tcp_sack_extend(tp->duplicate_sack, seq, end_seq); 3678 } 3679 3680 static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb) 3681 { 3682 struct tcp_sock *tp = tcp_sk(sk); 3683 3684 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 3685 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 3686 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST); 3687 tcp_enter_quickack_mode(sk); 3688 3689 if (tcp_is_sack(tp) && sysctl_tcp_dsack) { 3690 u32 end_seq = TCP_SKB_CB(skb)->end_seq; 3691 3692 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) 3693 end_seq = tp->rcv_nxt; 3694 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, end_seq); 3695 } 3696 } 3697 3698 tcp_send_ack(sk); 3699 } 3700 3701 /* These routines update the SACK block as out-of-order packets arrive or 3702 * in-order packets close up the sequence space. 3703 */ 3704 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp) 3705 { 3706 int this_sack; 3707 struct tcp_sack_block *sp = &tp->selective_acks[0]; 3708 struct tcp_sack_block *swalk = sp + 1; 3709 3710 /* See if the recent change to the first SACK eats into 3711 * or hits the sequence space of other SACK blocks, if so coalesce. 3712 */ 3713 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) { 3714 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) { 3715 int i; 3716 3717 /* Zap SWALK, by moving every further SACK up by one slot. 3718 * Decrease num_sacks. 3719 */ 3720 tp->rx_opt.num_sacks--; 3721 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + 3722 tp->rx_opt.dsack, 3723 4 - tp->rx_opt.tstamp_ok); 3724 for (i = this_sack; i < tp->rx_opt.num_sacks; i++) 3725 sp[i] = sp[i + 1]; 3726 continue; 3727 } 3728 this_sack++, swalk++; 3729 } 3730 } 3731 3732 static inline void tcp_sack_swap(struct tcp_sack_block *sack1, 3733 struct tcp_sack_block *sack2) 3734 { 3735 __u32 tmp; 3736 3737 tmp = sack1->start_seq; 3738 sack1->start_seq = sack2->start_seq; 3739 sack2->start_seq = tmp; 3740 3741 tmp = sack1->end_seq; 3742 sack1->end_seq = sack2->end_seq; 3743 sack2->end_seq = tmp; 3744 } 3745 3746 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq) 3747 { 3748 struct tcp_sock *tp = tcp_sk(sk); 3749 struct tcp_sack_block *sp = &tp->selective_acks[0]; 3750 int cur_sacks = tp->rx_opt.num_sacks; 3751 int this_sack; 3752 3753 if (!cur_sacks) 3754 goto new_sack; 3755 3756 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) { 3757 if (tcp_sack_extend(sp, seq, end_seq)) { 3758 /* Rotate this_sack to the first one. */ 3759 for (; this_sack > 0; this_sack--, sp--) 3760 tcp_sack_swap(sp, sp - 1); 3761 if (cur_sacks > 1) 3762 tcp_sack_maybe_coalesce(tp); 3763 return; 3764 } 3765 } 3766 3767 /* Could not find an adjacent existing SACK, build a new one, 3768 * put it at the front, and shift everyone else down. We 3769 * always know there is at least one SACK present already here. 3770 * 3771 * If the sack array is full, forget about the last one. 3772 */ 3773 if (this_sack >= 4) { 3774 this_sack--; 3775 tp->rx_opt.num_sacks--; 3776 sp--; 3777 } 3778 for (; this_sack > 0; this_sack--, sp--) 3779 *sp = *(sp - 1); 3780 3781 new_sack: 3782 /* Build the new head SACK, and we're done. */ 3783 sp->start_seq = seq; 3784 sp->end_seq = end_seq; 3785 tp->rx_opt.num_sacks++; 3786 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 3787 4 - tp->rx_opt.tstamp_ok); 3788 } 3789 3790 /* RCV.NXT advances, some SACKs should be eaten. */ 3791 3792 static void tcp_sack_remove(struct tcp_sock *tp) 3793 { 3794 struct tcp_sack_block *sp = &tp->selective_acks[0]; 3795 int num_sacks = tp->rx_opt.num_sacks; 3796 int this_sack; 3797 3798 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */ 3799 if (skb_queue_empty(&tp->out_of_order_queue)) { 3800 tp->rx_opt.num_sacks = 0; 3801 tp->rx_opt.eff_sacks = tp->rx_opt.dsack; 3802 return; 3803 } 3804 3805 for (this_sack = 0; this_sack < num_sacks;) { 3806 /* Check if the start of the sack is covered by RCV.NXT. */ 3807 if (!before(tp->rcv_nxt, sp->start_seq)) { 3808 int i; 3809 3810 /* RCV.NXT must cover all the block! */ 3811 BUG_TRAP(!before(tp->rcv_nxt, sp->end_seq)); 3812 3813 /* Zap this SACK, by moving forward any other SACKS. */ 3814 for (i=this_sack+1; i < num_sacks; i++) 3815 tp->selective_acks[i-1] = tp->selective_acks[i]; 3816 num_sacks--; 3817 continue; 3818 } 3819 this_sack++; 3820 sp++; 3821 } 3822 if (num_sacks != tp->rx_opt.num_sacks) { 3823 tp->rx_opt.num_sacks = num_sacks; 3824 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + 3825 tp->rx_opt.dsack, 3826 4 - tp->rx_opt.tstamp_ok); 3827 } 3828 } 3829 3830 /* This one checks to see if we can put data from the 3831 * out_of_order queue into the receive_queue. 3832 */ 3833 static void tcp_ofo_queue(struct sock *sk) 3834 { 3835 struct tcp_sock *tp = tcp_sk(sk); 3836 __u32 dsack_high = tp->rcv_nxt; 3837 struct sk_buff *skb; 3838 3839 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) { 3840 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) 3841 break; 3842 3843 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) { 3844 __u32 dsack = dsack_high; 3845 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high)) 3846 dsack_high = TCP_SKB_CB(skb)->end_seq; 3847 tcp_dsack_extend(tp, TCP_SKB_CB(skb)->seq, dsack); 3848 } 3849 3850 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) { 3851 SOCK_DEBUG(sk, "ofo packet was already received \n"); 3852 __skb_unlink(skb, &tp->out_of_order_queue); 3853 __kfree_skb(skb); 3854 continue; 3855 } 3856 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n", 3857 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, 3858 TCP_SKB_CB(skb)->end_seq); 3859 3860 __skb_unlink(skb, &tp->out_of_order_queue); 3861 __skb_queue_tail(&sk->sk_receive_queue, skb); 3862 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; 3863 if (tcp_hdr(skb)->fin) 3864 tcp_fin(skb, sk, tcp_hdr(skb)); 3865 } 3866 } 3867 3868 static int tcp_prune_ofo_queue(struct sock *sk); 3869 static int tcp_prune_queue(struct sock *sk); 3870 3871 static inline int tcp_try_rmem_schedule(struct sock *sk, unsigned int size) 3872 { 3873 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf || 3874 !sk_rmem_schedule(sk, size)) { 3875 3876 if (tcp_prune_queue(sk) < 0) 3877 return -1; 3878 3879 if (!sk_rmem_schedule(sk, size)) { 3880 if (!tcp_prune_ofo_queue(sk)) 3881 return -1; 3882 3883 if (!sk_rmem_schedule(sk, size)) 3884 return -1; 3885 } 3886 } 3887 return 0; 3888 } 3889 3890 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb) 3891 { 3892 struct tcphdr *th = tcp_hdr(skb); 3893 struct tcp_sock *tp = tcp_sk(sk); 3894 int eaten = -1; 3895 3896 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) 3897 goto drop; 3898 3899 __skb_pull(skb, th->doff * 4); 3900 3901 TCP_ECN_accept_cwr(tp, skb); 3902 3903 if (tp->rx_opt.dsack) { 3904 tp->rx_opt.dsack = 0; 3905 tp->rx_opt.eff_sacks = min_t(unsigned int, tp->rx_opt.num_sacks, 3906 4 - tp->rx_opt.tstamp_ok); 3907 } 3908 3909 /* Queue data for delivery to the user. 3910 * Packets in sequence go to the receive queue. 3911 * Out of sequence packets to the out_of_order_queue. 3912 */ 3913 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) { 3914 if (tcp_receive_window(tp) == 0) 3915 goto out_of_window; 3916 3917 /* Ok. In sequence. In window. */ 3918 if (tp->ucopy.task == current && 3919 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len && 3920 sock_owned_by_user(sk) && !tp->urg_data) { 3921 int chunk = min_t(unsigned int, skb->len, 3922 tp->ucopy.len); 3923 3924 __set_current_state(TASK_RUNNING); 3925 3926 local_bh_enable(); 3927 if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) { 3928 tp->ucopy.len -= chunk; 3929 tp->copied_seq += chunk; 3930 eaten = (chunk == skb->len && !th->fin); 3931 tcp_rcv_space_adjust(sk); 3932 } 3933 local_bh_disable(); 3934 } 3935 3936 if (eaten <= 0) { 3937 queue_and_out: 3938 if (eaten < 0 && 3939 tcp_try_rmem_schedule(sk, skb->truesize)) 3940 goto drop; 3941 3942 skb_set_owner_r(skb, sk); 3943 __skb_queue_tail(&sk->sk_receive_queue, skb); 3944 } 3945 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; 3946 if (skb->len) 3947 tcp_event_data_recv(sk, skb); 3948 if (th->fin) 3949 tcp_fin(skb, sk, th); 3950 3951 if (!skb_queue_empty(&tp->out_of_order_queue)) { 3952 tcp_ofo_queue(sk); 3953 3954 /* RFC2581. 4.2. SHOULD send immediate ACK, when 3955 * gap in queue is filled. 3956 */ 3957 if (skb_queue_empty(&tp->out_of_order_queue)) 3958 inet_csk(sk)->icsk_ack.pingpong = 0; 3959 } 3960 3961 if (tp->rx_opt.num_sacks) 3962 tcp_sack_remove(tp); 3963 3964 tcp_fast_path_check(sk); 3965 3966 if (eaten > 0) 3967 __kfree_skb(skb); 3968 else if (!sock_flag(sk, SOCK_DEAD)) 3969 sk->sk_data_ready(sk, 0); 3970 return; 3971 } 3972 3973 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) { 3974 /* A retransmit, 2nd most common case. Force an immediate ack. */ 3975 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST); 3976 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); 3977 3978 out_of_window: 3979 tcp_enter_quickack_mode(sk); 3980 inet_csk_schedule_ack(sk); 3981 drop: 3982 __kfree_skb(skb); 3983 return; 3984 } 3985 3986 /* Out of window. F.e. zero window probe. */ 3987 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp))) 3988 goto out_of_window; 3989 3990 tcp_enter_quickack_mode(sk); 3991 3992 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 3993 /* Partial packet, seq < rcv_next < end_seq */ 3994 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n", 3995 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, 3996 TCP_SKB_CB(skb)->end_seq); 3997 3998 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, tp->rcv_nxt); 3999 4000 /* If window is closed, drop tail of packet. But after 4001 * remembering D-SACK for its head made in previous line. 4002 */ 4003 if (!tcp_receive_window(tp)) 4004 goto out_of_window; 4005 goto queue_and_out; 4006 } 4007 4008 TCP_ECN_check_ce(tp, skb); 4009 4010 if (tcp_try_rmem_schedule(sk, skb->truesize)) 4011 goto drop; 4012 4013 /* Disable header prediction. */ 4014 tp->pred_flags = 0; 4015 inet_csk_schedule_ack(sk); 4016 4017 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n", 4018 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); 4019 4020 skb_set_owner_r(skb, sk); 4021 4022 if (!skb_peek(&tp->out_of_order_queue)) { 4023 /* Initial out of order segment, build 1 SACK. */ 4024 if (tcp_is_sack(tp)) { 4025 tp->rx_opt.num_sacks = 1; 4026 tp->rx_opt.dsack = 0; 4027 tp->rx_opt.eff_sacks = 1; 4028 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq; 4029 tp->selective_acks[0].end_seq = 4030 TCP_SKB_CB(skb)->end_seq; 4031 } 4032 __skb_queue_head(&tp->out_of_order_queue, skb); 4033 } else { 4034 struct sk_buff *skb1 = tp->out_of_order_queue.prev; 4035 u32 seq = TCP_SKB_CB(skb)->seq; 4036 u32 end_seq = TCP_SKB_CB(skb)->end_seq; 4037 4038 if (seq == TCP_SKB_CB(skb1)->end_seq) { 4039 __skb_queue_after(&tp->out_of_order_queue, skb1, skb); 4040 4041 if (!tp->rx_opt.num_sacks || 4042 tp->selective_acks[0].end_seq != seq) 4043 goto add_sack; 4044 4045 /* Common case: data arrive in order after hole. */ 4046 tp->selective_acks[0].end_seq = end_seq; 4047 return; 4048 } 4049 4050 /* Find place to insert this segment. */ 4051 do { 4052 if (!after(TCP_SKB_CB(skb1)->seq, seq)) 4053 break; 4054 } while ((skb1 = skb1->prev) != 4055 (struct sk_buff *)&tp->out_of_order_queue); 4056 4057 /* Do skb overlap to previous one? */ 4058 if (skb1 != (struct sk_buff *)&tp->out_of_order_queue && 4059 before(seq, TCP_SKB_CB(skb1)->end_seq)) { 4060 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) { 4061 /* All the bits are present. Drop. */ 4062 __kfree_skb(skb); 4063 tcp_dsack_set(tp, seq, end_seq); 4064 goto add_sack; 4065 } 4066 if (after(seq, TCP_SKB_CB(skb1)->seq)) { 4067 /* Partial overlap. */ 4068 tcp_dsack_set(tp, seq, 4069 TCP_SKB_CB(skb1)->end_seq); 4070 } else { 4071 skb1 = skb1->prev; 4072 } 4073 } 4074 __skb_insert(skb, skb1, skb1->next, &tp->out_of_order_queue); 4075 4076 /* And clean segments covered by new one as whole. */ 4077 while ((skb1 = skb->next) != 4078 (struct sk_buff *)&tp->out_of_order_queue && 4079 after(end_seq, TCP_SKB_CB(skb1)->seq)) { 4080 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) { 4081 tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, 4082 end_seq); 4083 break; 4084 } 4085 __skb_unlink(skb1, &tp->out_of_order_queue); 4086 tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, 4087 TCP_SKB_CB(skb1)->end_seq); 4088 __kfree_skb(skb1); 4089 } 4090 4091 add_sack: 4092 if (tcp_is_sack(tp)) 4093 tcp_sack_new_ofo_skb(sk, seq, end_seq); 4094 } 4095 } 4096 4097 /* Collapse contiguous sequence of skbs head..tail with 4098 * sequence numbers start..end. 4099 * Segments with FIN/SYN are not collapsed (only because this 4100 * simplifies code) 4101 */ 4102 static void 4103 tcp_collapse(struct sock *sk, struct sk_buff_head *list, 4104 struct sk_buff *head, struct sk_buff *tail, 4105 u32 start, u32 end) 4106 { 4107 struct sk_buff *skb; 4108 4109 /* First, check that queue is collapsible and find 4110 * the point where collapsing can be useful. */ 4111 for (skb = head; skb != tail;) { 4112 /* No new bits? It is possible on ofo queue. */ 4113 if (!before(start, TCP_SKB_CB(skb)->end_seq)) { 4114 struct sk_buff *next = skb->next; 4115 __skb_unlink(skb, list); 4116 __kfree_skb(skb); 4117 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED); 4118 skb = next; 4119 continue; 4120 } 4121 4122 /* The first skb to collapse is: 4123 * - not SYN/FIN and 4124 * - bloated or contains data before "start" or 4125 * overlaps to the next one. 4126 */ 4127 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin && 4128 (tcp_win_from_space(skb->truesize) > skb->len || 4129 before(TCP_SKB_CB(skb)->seq, start) || 4130 (skb->next != tail && 4131 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq))) 4132 break; 4133 4134 /* Decided to skip this, advance start seq. */ 4135 start = TCP_SKB_CB(skb)->end_seq; 4136 skb = skb->next; 4137 } 4138 if (skb == tail || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin) 4139 return; 4140 4141 while (before(start, end)) { 4142 struct sk_buff *nskb; 4143 unsigned int header = skb_headroom(skb); 4144 int copy = SKB_MAX_ORDER(header, 0); 4145 4146 /* Too big header? This can happen with IPv6. */ 4147 if (copy < 0) 4148 return; 4149 if (end - start < copy) 4150 copy = end - start; 4151 nskb = alloc_skb(copy + header, GFP_ATOMIC); 4152 if (!nskb) 4153 return; 4154 4155 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head); 4156 skb_set_network_header(nskb, (skb_network_header(skb) - 4157 skb->head)); 4158 skb_set_transport_header(nskb, (skb_transport_header(skb) - 4159 skb->head)); 4160 skb_reserve(nskb, header); 4161 memcpy(nskb->head, skb->head, header); 4162 memcpy(nskb->cb, skb->cb, sizeof(skb->cb)); 4163 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start; 4164 __skb_insert(nskb, skb->prev, skb, list); 4165 skb_set_owner_r(nskb, sk); 4166 4167 /* Copy data, releasing collapsed skbs. */ 4168 while (copy > 0) { 4169 int offset = start - TCP_SKB_CB(skb)->seq; 4170 int size = TCP_SKB_CB(skb)->end_seq - start; 4171 4172 BUG_ON(offset < 0); 4173 if (size > 0) { 4174 size = min(copy, size); 4175 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size)) 4176 BUG(); 4177 TCP_SKB_CB(nskb)->end_seq += size; 4178 copy -= size; 4179 start += size; 4180 } 4181 if (!before(start, TCP_SKB_CB(skb)->end_seq)) { 4182 struct sk_buff *next = skb->next; 4183 __skb_unlink(skb, list); 4184 __kfree_skb(skb); 4185 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED); 4186 skb = next; 4187 if (skb == tail || 4188 tcp_hdr(skb)->syn || 4189 tcp_hdr(skb)->fin) 4190 return; 4191 } 4192 } 4193 } 4194 } 4195 4196 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs 4197 * and tcp_collapse() them until all the queue is collapsed. 4198 */ 4199 static void tcp_collapse_ofo_queue(struct sock *sk) 4200 { 4201 struct tcp_sock *tp = tcp_sk(sk); 4202 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue); 4203 struct sk_buff *head; 4204 u32 start, end; 4205 4206 if (skb == NULL) 4207 return; 4208 4209 start = TCP_SKB_CB(skb)->seq; 4210 end = TCP_SKB_CB(skb)->end_seq; 4211 head = skb; 4212 4213 for (;;) { 4214 skb = skb->next; 4215 4216 /* Segment is terminated when we see gap or when 4217 * we are at the end of all the queue. */ 4218 if (skb == (struct sk_buff *)&tp->out_of_order_queue || 4219 after(TCP_SKB_CB(skb)->seq, end) || 4220 before(TCP_SKB_CB(skb)->end_seq, start)) { 4221 tcp_collapse(sk, &tp->out_of_order_queue, 4222 head, skb, start, end); 4223 head = skb; 4224 if (skb == (struct sk_buff *)&tp->out_of_order_queue) 4225 break; 4226 /* Start new segment */ 4227 start = TCP_SKB_CB(skb)->seq; 4228 end = TCP_SKB_CB(skb)->end_seq; 4229 } else { 4230 if (before(TCP_SKB_CB(skb)->seq, start)) 4231 start = TCP_SKB_CB(skb)->seq; 4232 if (after(TCP_SKB_CB(skb)->end_seq, end)) 4233 end = TCP_SKB_CB(skb)->end_seq; 4234 } 4235 } 4236 } 4237 4238 /* 4239 * Purge the out-of-order queue. 4240 * Return true if queue was pruned. 4241 */ 4242 static int tcp_prune_ofo_queue(struct sock *sk) 4243 { 4244 struct tcp_sock *tp = tcp_sk(sk); 4245 int res = 0; 4246 4247 if (!skb_queue_empty(&tp->out_of_order_queue)) { 4248 NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED); 4249 __skb_queue_purge(&tp->out_of_order_queue); 4250 4251 /* Reset SACK state. A conforming SACK implementation will 4252 * do the same at a timeout based retransmit. When a connection 4253 * is in a sad state like this, we care only about integrity 4254 * of the connection not performance. 4255 */ 4256 if (tp->rx_opt.sack_ok) 4257 tcp_sack_reset(&tp->rx_opt); 4258 sk_mem_reclaim(sk); 4259 res = 1; 4260 } 4261 return res; 4262 } 4263 4264 /* Reduce allocated memory if we can, trying to get 4265 * the socket within its memory limits again. 4266 * 4267 * Return less than zero if we should start dropping frames 4268 * until the socket owning process reads some of the data 4269 * to stabilize the situation. 4270 */ 4271 static int tcp_prune_queue(struct sock *sk) 4272 { 4273 struct tcp_sock *tp = tcp_sk(sk); 4274 4275 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq); 4276 4277 NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED); 4278 4279 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf) 4280 tcp_clamp_window(sk); 4281 else if (tcp_memory_pressure) 4282 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss); 4283 4284 tcp_collapse_ofo_queue(sk); 4285 tcp_collapse(sk, &sk->sk_receive_queue, 4286 sk->sk_receive_queue.next, 4287 (struct sk_buff *)&sk->sk_receive_queue, 4288 tp->copied_seq, tp->rcv_nxt); 4289 sk_mem_reclaim(sk); 4290 4291 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) 4292 return 0; 4293 4294 /* Collapsing did not help, destructive actions follow. 4295 * This must not ever occur. */ 4296 4297 tcp_prune_ofo_queue(sk); 4298 4299 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) 4300 return 0; 4301 4302 /* If we are really being abused, tell the caller to silently 4303 * drop receive data on the floor. It will get retransmitted 4304 * and hopefully then we'll have sufficient space. 4305 */ 4306 NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED); 4307 4308 /* Massive buffer overcommit. */ 4309 tp->pred_flags = 0; 4310 return -1; 4311 } 4312 4313 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. 4314 * As additional protections, we do not touch cwnd in retransmission phases, 4315 * and if application hit its sndbuf limit recently. 4316 */ 4317 void tcp_cwnd_application_limited(struct sock *sk) 4318 { 4319 struct tcp_sock *tp = tcp_sk(sk); 4320 4321 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && 4322 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { 4323 /* Limited by application or receiver window. */ 4324 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); 4325 u32 win_used = max(tp->snd_cwnd_used, init_win); 4326 if (win_used < tp->snd_cwnd) { 4327 tp->snd_ssthresh = tcp_current_ssthresh(sk); 4328 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1; 4329 } 4330 tp->snd_cwnd_used = 0; 4331 } 4332 tp->snd_cwnd_stamp = tcp_time_stamp; 4333 } 4334 4335 static int tcp_should_expand_sndbuf(struct sock *sk) 4336 { 4337 struct tcp_sock *tp = tcp_sk(sk); 4338 4339 /* If the user specified a specific send buffer setting, do 4340 * not modify it. 4341 */ 4342 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK) 4343 return 0; 4344 4345 /* If we are under global TCP memory pressure, do not expand. */ 4346 if (tcp_memory_pressure) 4347 return 0; 4348 4349 /* If we are under soft global TCP memory pressure, do not expand. */ 4350 if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0]) 4351 return 0; 4352 4353 /* If we filled the congestion window, do not expand. */ 4354 if (tp->packets_out >= tp->snd_cwnd) 4355 return 0; 4356 4357 return 1; 4358 } 4359 4360 /* When incoming ACK allowed to free some skb from write_queue, 4361 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket 4362 * on the exit from tcp input handler. 4363 * 4364 * PROBLEM: sndbuf expansion does not work well with largesend. 4365 */ 4366 static void tcp_new_space(struct sock *sk) 4367 { 4368 struct tcp_sock *tp = tcp_sk(sk); 4369 4370 if (tcp_should_expand_sndbuf(sk)) { 4371 int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) + 4372 MAX_TCP_HEADER + 16 + sizeof(struct sk_buff), 4373 demanded = max_t(unsigned int, tp->snd_cwnd, 4374 tp->reordering + 1); 4375 sndmem *= 2 * demanded; 4376 if (sndmem > sk->sk_sndbuf) 4377 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]); 4378 tp->snd_cwnd_stamp = tcp_time_stamp; 4379 } 4380 4381 sk->sk_write_space(sk); 4382 } 4383 4384 static void tcp_check_space(struct sock *sk) 4385 { 4386 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) { 4387 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK); 4388 if (sk->sk_socket && 4389 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) 4390 tcp_new_space(sk); 4391 } 4392 } 4393 4394 static inline void tcp_data_snd_check(struct sock *sk) 4395 { 4396 tcp_push_pending_frames(sk); 4397 tcp_check_space(sk); 4398 } 4399 4400 /* 4401 * Check if sending an ack is needed. 4402 */ 4403 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible) 4404 { 4405 struct tcp_sock *tp = tcp_sk(sk); 4406 4407 /* More than one full frame received... */ 4408 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss 4409 /* ... and right edge of window advances far enough. 4410 * (tcp_recvmsg() will send ACK otherwise). Or... 4411 */ 4412 && __tcp_select_window(sk) >= tp->rcv_wnd) || 4413 /* We ACK each frame or... */ 4414 tcp_in_quickack_mode(sk) || 4415 /* We have out of order data. */ 4416 (ofo_possible && skb_peek(&tp->out_of_order_queue))) { 4417 /* Then ack it now */ 4418 tcp_send_ack(sk); 4419 } else { 4420 /* Else, send delayed ack. */ 4421 tcp_send_delayed_ack(sk); 4422 } 4423 } 4424 4425 static inline void tcp_ack_snd_check(struct sock *sk) 4426 { 4427 if (!inet_csk_ack_scheduled(sk)) { 4428 /* We sent a data segment already. */ 4429 return; 4430 } 4431 __tcp_ack_snd_check(sk, 1); 4432 } 4433 4434 /* 4435 * This routine is only called when we have urgent data 4436 * signaled. Its the 'slow' part of tcp_urg. It could be 4437 * moved inline now as tcp_urg is only called from one 4438 * place. We handle URGent data wrong. We have to - as 4439 * BSD still doesn't use the correction from RFC961. 4440 * For 1003.1g we should support a new option TCP_STDURG to permit 4441 * either form (or just set the sysctl tcp_stdurg). 4442 */ 4443 4444 static void tcp_check_urg(struct sock *sk, struct tcphdr *th) 4445 { 4446 struct tcp_sock *tp = tcp_sk(sk); 4447 u32 ptr = ntohs(th->urg_ptr); 4448 4449 if (ptr && !sysctl_tcp_stdurg) 4450 ptr--; 4451 ptr += ntohl(th->seq); 4452 4453 /* Ignore urgent data that we've already seen and read. */ 4454 if (after(tp->copied_seq, ptr)) 4455 return; 4456 4457 /* Do not replay urg ptr. 4458 * 4459 * NOTE: interesting situation not covered by specs. 4460 * Misbehaving sender may send urg ptr, pointing to segment, 4461 * which we already have in ofo queue. We are not able to fetch 4462 * such data and will stay in TCP_URG_NOTYET until will be eaten 4463 * by recvmsg(). Seems, we are not obliged to handle such wicked 4464 * situations. But it is worth to think about possibility of some 4465 * DoSes using some hypothetical application level deadlock. 4466 */ 4467 if (before(ptr, tp->rcv_nxt)) 4468 return; 4469 4470 /* Do we already have a newer (or duplicate) urgent pointer? */ 4471 if (tp->urg_data && !after(ptr, tp->urg_seq)) 4472 return; 4473 4474 /* Tell the world about our new urgent pointer. */ 4475 sk_send_sigurg(sk); 4476 4477 /* We may be adding urgent data when the last byte read was 4478 * urgent. To do this requires some care. We cannot just ignore 4479 * tp->copied_seq since we would read the last urgent byte again 4480 * as data, nor can we alter copied_seq until this data arrives 4481 * or we break the semantics of SIOCATMARK (and thus sockatmark()) 4482 * 4483 * NOTE. Double Dutch. Rendering to plain English: author of comment 4484 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB); 4485 * and expect that both A and B disappear from stream. This is _wrong_. 4486 * Though this happens in BSD with high probability, this is occasional. 4487 * Any application relying on this is buggy. Note also, that fix "works" 4488 * only in this artificial test. Insert some normal data between A and B and we will 4489 * decline of BSD again. Verdict: it is better to remove to trap 4490 * buggy users. 4491 */ 4492 if (tp->urg_seq == tp->copied_seq && tp->urg_data && 4493 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) { 4494 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue); 4495 tp->copied_seq++; 4496 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) { 4497 __skb_unlink(skb, &sk->sk_receive_queue); 4498 __kfree_skb(skb); 4499 } 4500 } 4501 4502 tp->urg_data = TCP_URG_NOTYET; 4503 tp->urg_seq = ptr; 4504 4505 /* Disable header prediction. */ 4506 tp->pred_flags = 0; 4507 } 4508 4509 /* This is the 'fast' part of urgent handling. */ 4510 static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th) 4511 { 4512 struct tcp_sock *tp = tcp_sk(sk); 4513 4514 /* Check if we get a new urgent pointer - normally not. */ 4515 if (th->urg) 4516 tcp_check_urg(sk, th); 4517 4518 /* Do we wait for any urgent data? - normally not... */ 4519 if (tp->urg_data == TCP_URG_NOTYET) { 4520 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) - 4521 th->syn; 4522 4523 /* Is the urgent pointer pointing into this packet? */ 4524 if (ptr < skb->len) { 4525 u8 tmp; 4526 if (skb_copy_bits(skb, ptr, &tmp, 1)) 4527 BUG(); 4528 tp->urg_data = TCP_URG_VALID | tmp; 4529 if (!sock_flag(sk, SOCK_DEAD)) 4530 sk->sk_data_ready(sk, 0); 4531 } 4532 } 4533 } 4534 4535 static int tcp_defer_accept_check(struct sock *sk) 4536 { 4537 struct tcp_sock *tp = tcp_sk(sk); 4538 4539 if (tp->defer_tcp_accept.request) { 4540 int queued_data = tp->rcv_nxt - tp->copied_seq; 4541 int hasfin = !skb_queue_empty(&sk->sk_receive_queue) ? 4542 tcp_hdr((struct sk_buff *) 4543 sk->sk_receive_queue.prev)->fin : 0; 4544 4545 if (queued_data && hasfin) 4546 queued_data--; 4547 4548 if (queued_data && 4549 tp->defer_tcp_accept.listen_sk->sk_state == TCP_LISTEN) { 4550 if (sock_flag(sk, SOCK_KEEPOPEN)) { 4551 inet_csk_reset_keepalive_timer(sk, 4552 keepalive_time_when(tp)); 4553 } else { 4554 inet_csk_delete_keepalive_timer(sk); 4555 } 4556 4557 inet_csk_reqsk_queue_add( 4558 tp->defer_tcp_accept.listen_sk, 4559 tp->defer_tcp_accept.request, 4560 sk); 4561 4562 tp->defer_tcp_accept.listen_sk->sk_data_ready( 4563 tp->defer_tcp_accept.listen_sk, 0); 4564 4565 sock_put(tp->defer_tcp_accept.listen_sk); 4566 sock_put(sk); 4567 tp->defer_tcp_accept.listen_sk = NULL; 4568 tp->defer_tcp_accept.request = NULL; 4569 } else if (hasfin || 4570 tp->defer_tcp_accept.listen_sk->sk_state != TCP_LISTEN) { 4571 tcp_reset(sk); 4572 return -1; 4573 } 4574 } 4575 return 0; 4576 } 4577 4578 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen) 4579 { 4580 struct tcp_sock *tp = tcp_sk(sk); 4581 int chunk = skb->len - hlen; 4582 int err; 4583 4584 local_bh_enable(); 4585 if (skb_csum_unnecessary(skb)) 4586 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk); 4587 else 4588 err = skb_copy_and_csum_datagram_iovec(skb, hlen, 4589 tp->ucopy.iov); 4590 4591 if (!err) { 4592 tp->ucopy.len -= chunk; 4593 tp->copied_seq += chunk; 4594 tcp_rcv_space_adjust(sk); 4595 } 4596 4597 local_bh_disable(); 4598 return err; 4599 } 4600 4601 static __sum16 __tcp_checksum_complete_user(struct sock *sk, 4602 struct sk_buff *skb) 4603 { 4604 __sum16 result; 4605 4606 if (sock_owned_by_user(sk)) { 4607 local_bh_enable(); 4608 result = __tcp_checksum_complete(skb); 4609 local_bh_disable(); 4610 } else { 4611 result = __tcp_checksum_complete(skb); 4612 } 4613 return result; 4614 } 4615 4616 static inline int tcp_checksum_complete_user(struct sock *sk, 4617 struct sk_buff *skb) 4618 { 4619 return !skb_csum_unnecessary(skb) && 4620 __tcp_checksum_complete_user(sk, skb); 4621 } 4622 4623 #ifdef CONFIG_NET_DMA 4624 static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb, 4625 int hlen) 4626 { 4627 struct tcp_sock *tp = tcp_sk(sk); 4628 int chunk = skb->len - hlen; 4629 int dma_cookie; 4630 int copied_early = 0; 4631 4632 if (tp->ucopy.wakeup) 4633 return 0; 4634 4635 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list) 4636 tp->ucopy.dma_chan = get_softnet_dma(); 4637 4638 if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) { 4639 4640 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan, 4641 skb, hlen, 4642 tp->ucopy.iov, chunk, 4643 tp->ucopy.pinned_list); 4644 4645 if (dma_cookie < 0) 4646 goto out; 4647 4648 tp->ucopy.dma_cookie = dma_cookie; 4649 copied_early = 1; 4650 4651 tp->ucopy.len -= chunk; 4652 tp->copied_seq += chunk; 4653 tcp_rcv_space_adjust(sk); 4654 4655 if ((tp->ucopy.len == 0) || 4656 (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) || 4657 (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) { 4658 tp->ucopy.wakeup = 1; 4659 sk->sk_data_ready(sk, 0); 4660 } 4661 } else if (chunk > 0) { 4662 tp->ucopy.wakeup = 1; 4663 sk->sk_data_ready(sk, 0); 4664 } 4665 out: 4666 return copied_early; 4667 } 4668 #endif /* CONFIG_NET_DMA */ 4669 4670 /* 4671 * TCP receive function for the ESTABLISHED state. 4672 * 4673 * It is split into a fast path and a slow path. The fast path is 4674 * disabled when: 4675 * - A zero window was announced from us - zero window probing 4676 * is only handled properly in the slow path. 4677 * - Out of order segments arrived. 4678 * - Urgent data is expected. 4679 * - There is no buffer space left 4680 * - Unexpected TCP flags/window values/header lengths are received 4681 * (detected by checking the TCP header against pred_flags) 4682 * - Data is sent in both directions. Fast path only supports pure senders 4683 * or pure receivers (this means either the sequence number or the ack 4684 * value must stay constant) 4685 * - Unexpected TCP option. 4686 * 4687 * When these conditions are not satisfied it drops into a standard 4688 * receive procedure patterned after RFC793 to handle all cases. 4689 * The first three cases are guaranteed by proper pred_flags setting, 4690 * the rest is checked inline. Fast processing is turned on in 4691 * tcp_data_queue when everything is OK. 4692 */ 4693 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb, 4694 struct tcphdr *th, unsigned len) 4695 { 4696 struct tcp_sock *tp = tcp_sk(sk); 4697 4698 /* 4699 * Header prediction. 4700 * The code loosely follows the one in the famous 4701 * "30 instruction TCP receive" Van Jacobson mail. 4702 * 4703 * Van's trick is to deposit buffers into socket queue 4704 * on a device interrupt, to call tcp_recv function 4705 * on the receive process context and checksum and copy 4706 * the buffer to user space. smart... 4707 * 4708 * Our current scheme is not silly either but we take the 4709 * extra cost of the net_bh soft interrupt processing... 4710 * We do checksum and copy also but from device to kernel. 4711 */ 4712 4713 tp->rx_opt.saw_tstamp = 0; 4714 4715 /* pred_flags is 0xS?10 << 16 + snd_wnd 4716 * if header_prediction is to be made 4717 * 'S' will always be tp->tcp_header_len >> 2 4718 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to 4719 * turn it off (when there are holes in the receive 4720 * space for instance) 4721 * PSH flag is ignored. 4722 */ 4723 4724 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags && 4725 TCP_SKB_CB(skb)->seq == tp->rcv_nxt) { 4726 int tcp_header_len = tp->tcp_header_len; 4727 4728 /* Timestamp header prediction: tcp_header_len 4729 * is automatically equal to th->doff*4 due to pred_flags 4730 * match. 4731 */ 4732 4733 /* Check timestamp */ 4734 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) { 4735 __be32 *ptr = (__be32 *)(th + 1); 4736 4737 /* No? Slow path! */ 4738 if (*ptr != htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) 4739 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) 4740 goto slow_path; 4741 4742 tp->rx_opt.saw_tstamp = 1; 4743 ++ptr; 4744 tp->rx_opt.rcv_tsval = ntohl(*ptr); 4745 ++ptr; 4746 tp->rx_opt.rcv_tsecr = ntohl(*ptr); 4747 4748 /* If PAWS failed, check it more carefully in slow path */ 4749 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0) 4750 goto slow_path; 4751 4752 /* DO NOT update ts_recent here, if checksum fails 4753 * and timestamp was corrupted part, it will result 4754 * in a hung connection since we will drop all 4755 * future packets due to the PAWS test. 4756 */ 4757 } 4758 4759 if (len <= tcp_header_len) { 4760 /* Bulk data transfer: sender */ 4761 if (len == tcp_header_len) { 4762 /* Predicted packet is in window by definition. 4763 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 4764 * Hence, check seq<=rcv_wup reduces to: 4765 */ 4766 if (tcp_header_len == 4767 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && 4768 tp->rcv_nxt == tp->rcv_wup) 4769 tcp_store_ts_recent(tp); 4770 4771 /* We know that such packets are checksummed 4772 * on entry. 4773 */ 4774 tcp_ack(sk, skb, 0); 4775 __kfree_skb(skb); 4776 tcp_data_snd_check(sk); 4777 return 0; 4778 } else { /* Header too small */ 4779 TCP_INC_STATS_BH(TCP_MIB_INERRS); 4780 goto discard; 4781 } 4782 } else { 4783 int eaten = 0; 4784 int copied_early = 0; 4785 4786 if (tp->copied_seq == tp->rcv_nxt && 4787 len - tcp_header_len <= tp->ucopy.len) { 4788 #ifdef CONFIG_NET_DMA 4789 if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) { 4790 copied_early = 1; 4791 eaten = 1; 4792 } 4793 #endif 4794 if (tp->ucopy.task == current && 4795 sock_owned_by_user(sk) && !copied_early) { 4796 __set_current_state(TASK_RUNNING); 4797 4798 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len)) 4799 eaten = 1; 4800 } 4801 if (eaten) { 4802 /* Predicted packet is in window by definition. 4803 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 4804 * Hence, check seq<=rcv_wup reduces to: 4805 */ 4806 if (tcp_header_len == 4807 (sizeof(struct tcphdr) + 4808 TCPOLEN_TSTAMP_ALIGNED) && 4809 tp->rcv_nxt == tp->rcv_wup) 4810 tcp_store_ts_recent(tp); 4811 4812 tcp_rcv_rtt_measure_ts(sk, skb); 4813 4814 __skb_pull(skb, tcp_header_len); 4815 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; 4816 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER); 4817 } 4818 if (copied_early) 4819 tcp_cleanup_rbuf(sk, skb->len); 4820 } 4821 if (!eaten) { 4822 if (tcp_checksum_complete_user(sk, skb)) 4823 goto csum_error; 4824 4825 /* Predicted packet is in window by definition. 4826 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 4827 * Hence, check seq<=rcv_wup reduces to: 4828 */ 4829 if (tcp_header_len == 4830 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && 4831 tp->rcv_nxt == tp->rcv_wup) 4832 tcp_store_ts_recent(tp); 4833 4834 tcp_rcv_rtt_measure_ts(sk, skb); 4835 4836 if ((int)skb->truesize > sk->sk_forward_alloc) 4837 goto step5; 4838 4839 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS); 4840 4841 /* Bulk data transfer: receiver */ 4842 __skb_pull(skb, tcp_header_len); 4843 __skb_queue_tail(&sk->sk_receive_queue, skb); 4844 skb_set_owner_r(skb, sk); 4845 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; 4846 } 4847 4848 tcp_event_data_recv(sk, skb); 4849 4850 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) { 4851 /* Well, only one small jumplet in fast path... */ 4852 tcp_ack(sk, skb, FLAG_DATA); 4853 tcp_data_snd_check(sk); 4854 if (!inet_csk_ack_scheduled(sk)) 4855 goto no_ack; 4856 } 4857 4858 __tcp_ack_snd_check(sk, 0); 4859 no_ack: 4860 #ifdef CONFIG_NET_DMA 4861 if (copied_early) 4862 __skb_queue_tail(&sk->sk_async_wait_queue, skb); 4863 else 4864 #endif 4865 if (eaten) 4866 __kfree_skb(skb); 4867 else 4868 sk->sk_data_ready(sk, 0); 4869 return 0; 4870 } 4871 } 4872 4873 slow_path: 4874 if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb)) 4875 goto csum_error; 4876 4877 /* 4878 * RFC1323: H1. Apply PAWS check first. 4879 */ 4880 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp && 4881 tcp_paws_discard(sk, skb)) { 4882 if (!th->rst) { 4883 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED); 4884 tcp_send_dupack(sk, skb); 4885 goto discard; 4886 } 4887 /* Resets are accepted even if PAWS failed. 4888 4889 ts_recent update must be made after we are sure 4890 that the packet is in window. 4891 */ 4892 } 4893 4894 /* 4895 * Standard slow path. 4896 */ 4897 4898 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) { 4899 /* RFC793, page 37: "In all states except SYN-SENT, all reset 4900 * (RST) segments are validated by checking their SEQ-fields." 4901 * And page 69: "If an incoming segment is not acceptable, 4902 * an acknowledgment should be sent in reply (unless the RST bit 4903 * is set, if so drop the segment and return)". 4904 */ 4905 if (!th->rst) 4906 tcp_send_dupack(sk, skb); 4907 goto discard; 4908 } 4909 4910 if (th->rst) { 4911 tcp_reset(sk); 4912 goto discard; 4913 } 4914 4915 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq); 4916 4917 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 4918 TCP_INC_STATS_BH(TCP_MIB_INERRS); 4919 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN); 4920 tcp_reset(sk); 4921 return 1; 4922 } 4923 4924 step5: 4925 if (th->ack) 4926 tcp_ack(sk, skb, FLAG_SLOWPATH); 4927 4928 tcp_rcv_rtt_measure_ts(sk, skb); 4929 4930 /* Process urgent data. */ 4931 tcp_urg(sk, skb, th); 4932 4933 /* step 7: process the segment text */ 4934 tcp_data_queue(sk, skb); 4935 4936 tcp_data_snd_check(sk); 4937 tcp_ack_snd_check(sk); 4938 4939 tcp_defer_accept_check(sk); 4940 return 0; 4941 4942 csum_error: 4943 TCP_INC_STATS_BH(TCP_MIB_INERRS); 4944 4945 discard: 4946 __kfree_skb(skb); 4947 return 0; 4948 } 4949 4950 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb, 4951 struct tcphdr *th, unsigned len) 4952 { 4953 struct tcp_sock *tp = tcp_sk(sk); 4954 struct inet_connection_sock *icsk = inet_csk(sk); 4955 int saved_clamp = tp->rx_opt.mss_clamp; 4956 4957 tcp_parse_options(skb, &tp->rx_opt, 0); 4958 4959 if (th->ack) { 4960 /* rfc793: 4961 * "If the state is SYN-SENT then 4962 * first check the ACK bit 4963 * If the ACK bit is set 4964 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send 4965 * a reset (unless the RST bit is set, if so drop 4966 * the segment and return)" 4967 * 4968 * We do not send data with SYN, so that RFC-correct 4969 * test reduces to: 4970 */ 4971 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt) 4972 goto reset_and_undo; 4973 4974 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 4975 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp, 4976 tcp_time_stamp)) { 4977 NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED); 4978 goto reset_and_undo; 4979 } 4980 4981 /* Now ACK is acceptable. 4982 * 4983 * "If the RST bit is set 4984 * If the ACK was acceptable then signal the user "error: 4985 * connection reset", drop the segment, enter CLOSED state, 4986 * delete TCB, and return." 4987 */ 4988 4989 if (th->rst) { 4990 tcp_reset(sk); 4991 goto discard; 4992 } 4993 4994 /* rfc793: 4995 * "fifth, if neither of the SYN or RST bits is set then 4996 * drop the segment and return." 4997 * 4998 * See note below! 4999 * --ANK(990513) 5000 */ 5001 if (!th->syn) 5002 goto discard_and_undo; 5003 5004 /* rfc793: 5005 * "If the SYN bit is on ... 5006 * are acceptable then ... 5007 * (our SYN has been ACKed), change the connection 5008 * state to ESTABLISHED..." 5009 */ 5010 5011 TCP_ECN_rcv_synack(tp, th); 5012 5013 tp->snd_wl1 = TCP_SKB_CB(skb)->seq; 5014 tcp_ack(sk, skb, FLAG_SLOWPATH); 5015 5016 /* Ok.. it's good. Set up sequence numbers and 5017 * move to established. 5018 */ 5019 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; 5020 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; 5021 5022 /* RFC1323: The window in SYN & SYN/ACK segments is 5023 * never scaled. 5024 */ 5025 tp->snd_wnd = ntohs(th->window); 5026 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(skb)->seq); 5027 5028 if (!tp->rx_opt.wscale_ok) { 5029 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0; 5030 tp->window_clamp = min(tp->window_clamp, 65535U); 5031 } 5032 5033 if (tp->rx_opt.saw_tstamp) { 5034 tp->rx_opt.tstamp_ok = 1; 5035 tp->tcp_header_len = 5036 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 5037 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; 5038 tcp_store_ts_recent(tp); 5039 } else { 5040 tp->tcp_header_len = sizeof(struct tcphdr); 5041 } 5042 5043 if (tcp_is_sack(tp) && sysctl_tcp_fack) 5044 tcp_enable_fack(tp); 5045 5046 tcp_mtup_init(sk); 5047 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 5048 tcp_initialize_rcv_mss(sk); 5049 5050 /* Remember, tcp_poll() does not lock socket! 5051 * Change state from SYN-SENT only after copied_seq 5052 * is initialized. */ 5053 tp->copied_seq = tp->rcv_nxt; 5054 smp_mb(); 5055 tcp_set_state(sk, TCP_ESTABLISHED); 5056 5057 security_inet_conn_established(sk, skb); 5058 5059 /* Make sure socket is routed, for correct metrics. */ 5060 icsk->icsk_af_ops->rebuild_header(sk); 5061 5062 tcp_init_metrics(sk); 5063 5064 tcp_init_congestion_control(sk); 5065 5066 /* Prevent spurious tcp_cwnd_restart() on first data 5067 * packet. 5068 */ 5069 tp->lsndtime = tcp_time_stamp; 5070 5071 tcp_init_buffer_space(sk); 5072 5073 if (sock_flag(sk, SOCK_KEEPOPEN)) 5074 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp)); 5075 5076 if (!tp->rx_opt.snd_wscale) 5077 __tcp_fast_path_on(tp, tp->snd_wnd); 5078 else 5079 tp->pred_flags = 0; 5080 5081 if (!sock_flag(sk, SOCK_DEAD)) { 5082 sk->sk_state_change(sk); 5083 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT); 5084 } 5085 5086 if (sk->sk_write_pending || 5087 icsk->icsk_accept_queue.rskq_defer_accept || 5088 icsk->icsk_ack.pingpong) { 5089 /* Save one ACK. Data will be ready after 5090 * several ticks, if write_pending is set. 5091 * 5092 * It may be deleted, but with this feature tcpdumps 5093 * look so _wonderfully_ clever, that I was not able 5094 * to stand against the temptation 8) --ANK 5095 */ 5096 inet_csk_schedule_ack(sk); 5097 icsk->icsk_ack.lrcvtime = tcp_time_stamp; 5098 icsk->icsk_ack.ato = TCP_ATO_MIN; 5099 tcp_incr_quickack(sk); 5100 tcp_enter_quickack_mode(sk); 5101 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 5102 TCP_DELACK_MAX, TCP_RTO_MAX); 5103 5104 discard: 5105 __kfree_skb(skb); 5106 return 0; 5107 } else { 5108 tcp_send_ack(sk); 5109 } 5110 return -1; 5111 } 5112 5113 /* No ACK in the segment */ 5114 5115 if (th->rst) { 5116 /* rfc793: 5117 * "If the RST bit is set 5118 * 5119 * Otherwise (no ACK) drop the segment and return." 5120 */ 5121 5122 goto discard_and_undo; 5123 } 5124 5125 /* PAWS check. */ 5126 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && 5127 tcp_paws_check(&tp->rx_opt, 0)) 5128 goto discard_and_undo; 5129 5130 if (th->syn) { 5131 /* We see SYN without ACK. It is attempt of 5132 * simultaneous connect with crossed SYNs. 5133 * Particularly, it can be connect to self. 5134 */ 5135 tcp_set_state(sk, TCP_SYN_RECV); 5136 5137 if (tp->rx_opt.saw_tstamp) { 5138 tp->rx_opt.tstamp_ok = 1; 5139 tcp_store_ts_recent(tp); 5140 tp->tcp_header_len = 5141 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 5142 } else { 5143 tp->tcp_header_len = sizeof(struct tcphdr); 5144 } 5145 5146 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; 5147 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; 5148 5149 /* RFC1323: The window in SYN & SYN/ACK segments is 5150 * never scaled. 5151 */ 5152 tp->snd_wnd = ntohs(th->window); 5153 tp->snd_wl1 = TCP_SKB_CB(skb)->seq; 5154 tp->max_window = tp->snd_wnd; 5155 5156 TCP_ECN_rcv_syn(tp, th); 5157 5158 tcp_mtup_init(sk); 5159 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 5160 tcp_initialize_rcv_mss(sk); 5161 5162 tcp_send_synack(sk); 5163 #if 0 5164 /* Note, we could accept data and URG from this segment. 5165 * There are no obstacles to make this. 5166 * 5167 * However, if we ignore data in ACKless segments sometimes, 5168 * we have no reasons to accept it sometimes. 5169 * Also, seems the code doing it in step6 of tcp_rcv_state_process 5170 * is not flawless. So, discard packet for sanity. 5171 * Uncomment this return to process the data. 5172 */ 5173 return -1; 5174 #else 5175 goto discard; 5176 #endif 5177 } 5178 /* "fifth, if neither of the SYN or RST bits is set then 5179 * drop the segment and return." 5180 */ 5181 5182 discard_and_undo: 5183 tcp_clear_options(&tp->rx_opt); 5184 tp->rx_opt.mss_clamp = saved_clamp; 5185 goto discard; 5186 5187 reset_and_undo: 5188 tcp_clear_options(&tp->rx_opt); 5189 tp->rx_opt.mss_clamp = saved_clamp; 5190 return 1; 5191 } 5192 5193 /* 5194 * This function implements the receiving procedure of RFC 793 for 5195 * all states except ESTABLISHED and TIME_WAIT. 5196 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be 5197 * address independent. 5198 */ 5199 5200 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb, 5201 struct tcphdr *th, unsigned len) 5202 { 5203 struct tcp_sock *tp = tcp_sk(sk); 5204 struct inet_connection_sock *icsk = inet_csk(sk); 5205 int queued = 0; 5206 5207 tp->rx_opt.saw_tstamp = 0; 5208 5209 switch (sk->sk_state) { 5210 case TCP_CLOSE: 5211 goto discard; 5212 5213 case TCP_LISTEN: 5214 if (th->ack) 5215 return 1; 5216 5217 if (th->rst) 5218 goto discard; 5219 5220 if (th->syn) { 5221 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0) 5222 return 1; 5223 5224 /* Now we have several options: In theory there is 5225 * nothing else in the frame. KA9Q has an option to 5226 * send data with the syn, BSD accepts data with the 5227 * syn up to the [to be] advertised window and 5228 * Solaris 2.1 gives you a protocol error. For now 5229 * we just ignore it, that fits the spec precisely 5230 * and avoids incompatibilities. It would be nice in 5231 * future to drop through and process the data. 5232 * 5233 * Now that TTCP is starting to be used we ought to 5234 * queue this data. 5235 * But, this leaves one open to an easy denial of 5236 * service attack, and SYN cookies can't defend 5237 * against this problem. So, we drop the data 5238 * in the interest of security over speed unless 5239 * it's still in use. 5240 */ 5241 kfree_skb(skb); 5242 return 0; 5243 } 5244 goto discard; 5245 5246 case TCP_SYN_SENT: 5247 queued = tcp_rcv_synsent_state_process(sk, skb, th, len); 5248 if (queued >= 0) 5249 return queued; 5250 5251 /* Do step6 onward by hand. */ 5252 tcp_urg(sk, skb, th); 5253 __kfree_skb(skb); 5254 tcp_data_snd_check(sk); 5255 return 0; 5256 } 5257 5258 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp && 5259 tcp_paws_discard(sk, skb)) { 5260 if (!th->rst) { 5261 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED); 5262 tcp_send_dupack(sk, skb); 5263 goto discard; 5264 } 5265 /* Reset is accepted even if it did not pass PAWS. */ 5266 } 5267 5268 /* step 1: check sequence number */ 5269 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) { 5270 if (!th->rst) 5271 tcp_send_dupack(sk, skb); 5272 goto discard; 5273 } 5274 5275 /* step 2: check RST bit */ 5276 if (th->rst) { 5277 tcp_reset(sk); 5278 goto discard; 5279 } 5280 5281 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq); 5282 5283 /* step 3: check security and precedence [ignored] */ 5284 5285 /* step 4: 5286 * 5287 * Check for a SYN in window. 5288 */ 5289 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 5290 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN); 5291 tcp_reset(sk); 5292 return 1; 5293 } 5294 5295 /* step 5: check the ACK field */ 5296 if (th->ack) { 5297 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH); 5298 5299 switch (sk->sk_state) { 5300 case TCP_SYN_RECV: 5301 if (acceptable) { 5302 tp->copied_seq = tp->rcv_nxt; 5303 smp_mb(); 5304 tcp_set_state(sk, TCP_ESTABLISHED); 5305 sk->sk_state_change(sk); 5306 5307 /* Note, that this wakeup is only for marginal 5308 * crossed SYN case. Passively open sockets 5309 * are not waked up, because sk->sk_sleep == 5310 * NULL and sk->sk_socket == NULL. 5311 */ 5312 if (sk->sk_socket) 5313 sk_wake_async(sk, 5314 SOCK_WAKE_IO, POLL_OUT); 5315 5316 tp->snd_una = TCP_SKB_CB(skb)->ack_seq; 5317 tp->snd_wnd = ntohs(th->window) << 5318 tp->rx_opt.snd_wscale; 5319 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, 5320 TCP_SKB_CB(skb)->seq); 5321 5322 /* tcp_ack considers this ACK as duplicate 5323 * and does not calculate rtt. 5324 * Fix it at least with timestamps. 5325 */ 5326 if (tp->rx_opt.saw_tstamp && 5327 tp->rx_opt.rcv_tsecr && !tp->srtt) 5328 tcp_ack_saw_tstamp(sk, 0); 5329 5330 if (tp->rx_opt.tstamp_ok) 5331 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; 5332 5333 /* Make sure socket is routed, for 5334 * correct metrics. 5335 */ 5336 icsk->icsk_af_ops->rebuild_header(sk); 5337 5338 tcp_init_metrics(sk); 5339 5340 tcp_init_congestion_control(sk); 5341 5342 /* Prevent spurious tcp_cwnd_restart() on 5343 * first data packet. 5344 */ 5345 tp->lsndtime = tcp_time_stamp; 5346 5347 tcp_mtup_init(sk); 5348 tcp_initialize_rcv_mss(sk); 5349 tcp_init_buffer_space(sk); 5350 tcp_fast_path_on(tp); 5351 } else { 5352 return 1; 5353 } 5354 break; 5355 5356 case TCP_FIN_WAIT1: 5357 if (tp->snd_una == tp->write_seq) { 5358 tcp_set_state(sk, TCP_FIN_WAIT2); 5359 sk->sk_shutdown |= SEND_SHUTDOWN; 5360 dst_confirm(sk->sk_dst_cache); 5361 5362 if (!sock_flag(sk, SOCK_DEAD)) 5363 /* Wake up lingering close() */ 5364 sk->sk_state_change(sk); 5365 else { 5366 int tmo; 5367 5368 if (tp->linger2 < 0 || 5369 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 5370 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) { 5371 tcp_done(sk); 5372 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA); 5373 return 1; 5374 } 5375 5376 tmo = tcp_fin_time(sk); 5377 if (tmo > TCP_TIMEWAIT_LEN) { 5378 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN); 5379 } else if (th->fin || sock_owned_by_user(sk)) { 5380 /* Bad case. We could lose such FIN otherwise. 5381 * It is not a big problem, but it looks confusing 5382 * and not so rare event. We still can lose it now, 5383 * if it spins in bh_lock_sock(), but it is really 5384 * marginal case. 5385 */ 5386 inet_csk_reset_keepalive_timer(sk, tmo); 5387 } else { 5388 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo); 5389 goto discard; 5390 } 5391 } 5392 } 5393 break; 5394 5395 case TCP_CLOSING: 5396 if (tp->snd_una == tp->write_seq) { 5397 tcp_time_wait(sk, TCP_TIME_WAIT, 0); 5398 goto discard; 5399 } 5400 break; 5401 5402 case TCP_LAST_ACK: 5403 if (tp->snd_una == tp->write_seq) { 5404 tcp_update_metrics(sk); 5405 tcp_done(sk); 5406 goto discard; 5407 } 5408 break; 5409 } 5410 } else 5411 goto discard; 5412 5413 /* step 6: check the URG bit */ 5414 tcp_urg(sk, skb, th); 5415 5416 /* step 7: process the segment text */ 5417 switch (sk->sk_state) { 5418 case TCP_CLOSE_WAIT: 5419 case TCP_CLOSING: 5420 case TCP_LAST_ACK: 5421 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) 5422 break; 5423 case TCP_FIN_WAIT1: 5424 case TCP_FIN_WAIT2: 5425 /* RFC 793 says to queue data in these states, 5426 * RFC 1122 says we MUST send a reset. 5427 * BSD 4.4 also does reset. 5428 */ 5429 if (sk->sk_shutdown & RCV_SHUTDOWN) { 5430 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 5431 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) { 5432 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA); 5433 tcp_reset(sk); 5434 return 1; 5435 } 5436 } 5437 /* Fall through */ 5438 case TCP_ESTABLISHED: 5439 tcp_data_queue(sk, skb); 5440 queued = 1; 5441 break; 5442 } 5443 5444 /* tcp_data could move socket to TIME-WAIT */ 5445 if (sk->sk_state != TCP_CLOSE) { 5446 tcp_data_snd_check(sk); 5447 tcp_ack_snd_check(sk); 5448 } 5449 5450 if (!queued) { 5451 discard: 5452 __kfree_skb(skb); 5453 } 5454 return 0; 5455 } 5456 5457 EXPORT_SYMBOL(sysctl_tcp_ecn); 5458 EXPORT_SYMBOL(sysctl_tcp_reordering); 5459 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale); 5460 EXPORT_SYMBOL(tcp_parse_options); 5461 EXPORT_SYMBOL(tcp_rcv_established); 5462 EXPORT_SYMBOL(tcp_rcv_state_process); 5463 EXPORT_SYMBOL(tcp_initialize_rcv_mss); 5464