1 /* 2 * INET An implementation of the TCP/IP protocol suite for the LINUX 3 * operating system. INET is implemented using the BSD Socket 4 * interface as the means of communication with the user level. 5 * 6 * Implementation of the Transmission Control Protocol(TCP). 7 * 8 * Authors: Ross Biro 9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> 10 * Mark Evans, <evansmp@uhura.aston.ac.uk> 11 * Corey Minyard <wf-rch!minyard@relay.EU.net> 12 * Florian La Roche, <flla@stud.uni-sb.de> 13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu> 14 * Linus Torvalds, <torvalds@cs.helsinki.fi> 15 * Alan Cox, <gw4pts@gw4pts.ampr.org> 16 * Matthew Dillon, <dillon@apollo.west.oic.com> 17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no> 18 * Jorge Cwik, <jorge@laser.satlink.net> 19 */ 20 21 /* 22 * Changes: 23 * Pedro Roque : Fast Retransmit/Recovery. 24 * Two receive queues. 25 * Retransmit queue handled by TCP. 26 * Better retransmit timer handling. 27 * New congestion avoidance. 28 * Header prediction. 29 * Variable renaming. 30 * 31 * Eric : Fast Retransmit. 32 * Randy Scott : MSS option defines. 33 * Eric Schenk : Fixes to slow start algorithm. 34 * Eric Schenk : Yet another double ACK bug. 35 * Eric Schenk : Delayed ACK bug fixes. 36 * Eric Schenk : Floyd style fast retrans war avoidance. 37 * David S. Miller : Don't allow zero congestion window. 38 * Eric Schenk : Fix retransmitter so that it sends 39 * next packet on ack of previous packet. 40 * Andi Kleen : Moved open_request checking here 41 * and process RSTs for open_requests. 42 * Andi Kleen : Better prune_queue, and other fixes. 43 * Andrey Savochkin: Fix RTT measurements in the presence of 44 * timestamps. 45 * Andrey Savochkin: Check sequence numbers correctly when 46 * removing SACKs due to in sequence incoming 47 * data segments. 48 * Andi Kleen: Make sure we never ack data there is not 49 * enough room for. Also make this condition 50 * a fatal error if it might still happen. 51 * Andi Kleen: Add tcp_measure_rcv_mss to make 52 * connections with MSS<min(MTU,ann. MSS) 53 * work without delayed acks. 54 * Andi Kleen: Process packets with PSH set in the 55 * fast path. 56 * J Hadi Salim: ECN support 57 * Andrei Gurtov, 58 * Pasi Sarolahti, 59 * Panu Kuhlberg: Experimental audit of TCP (re)transmission 60 * engine. Lots of bugs are found. 61 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs 62 */ 63 64 #include <linux/mm.h> 65 #include <linux/slab.h> 66 #include <linux/module.h> 67 #include <linux/sysctl.h> 68 #include <linux/kernel.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 EXPORT_SYMBOL(sysctl_tcp_reordering); 82 int sysctl_tcp_ecn __read_mostly = 2; 83 EXPORT_SYMBOL(sysctl_tcp_ecn); 84 int sysctl_tcp_dsack __read_mostly = 1; 85 int sysctl_tcp_app_win __read_mostly = 31; 86 int sysctl_tcp_adv_win_scale __read_mostly = 2; 87 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale); 88 89 int sysctl_tcp_stdurg __read_mostly; 90 int sysctl_tcp_rfc1337 __read_mostly; 91 int sysctl_tcp_max_orphans __read_mostly = NR_FILE; 92 int sysctl_tcp_frto __read_mostly = 2; 93 int sysctl_tcp_frto_response __read_mostly; 94 int sysctl_tcp_nometrics_save __read_mostly; 95 96 int sysctl_tcp_thin_dupack __read_mostly; 97 98 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1; 99 int sysctl_tcp_abc __read_mostly; 100 101 #define FLAG_DATA 0x01 /* Incoming frame contained data. */ 102 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */ 103 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */ 104 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */ 105 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */ 106 #define FLAG_DATA_SACKED 0x20 /* New SACK. */ 107 #define FLAG_ECE 0x40 /* ECE in this ACK */ 108 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/ 109 #define FLAG_ONLY_ORIG_SACKED 0x200 /* SACKs only non-rexmit sent before RTO */ 110 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */ 111 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */ 112 #define FLAG_NONHEAD_RETRANS_ACKED 0x1000 /* Non-head rexmitted data was ACKed */ 113 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */ 114 115 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED) 116 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED) 117 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE) 118 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED) 119 #define FLAG_ANY_PROGRESS (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED) 120 121 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH) 122 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH)) 123 124 /* Adapt the MSS value used to make delayed ack decision to the 125 * real world. 126 */ 127 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb) 128 { 129 struct inet_connection_sock *icsk = inet_csk(sk); 130 const unsigned int lss = icsk->icsk_ack.last_seg_size; 131 unsigned int len; 132 133 icsk->icsk_ack.last_seg_size = 0; 134 135 /* skb->len may jitter because of SACKs, even if peer 136 * sends good full-sized frames. 137 */ 138 len = skb_shinfo(skb)->gso_size ? : skb->len; 139 if (len >= icsk->icsk_ack.rcv_mss) { 140 icsk->icsk_ack.rcv_mss = len; 141 } else { 142 /* Otherwise, we make more careful check taking into account, 143 * that SACKs block is variable. 144 * 145 * "len" is invariant segment length, including TCP header. 146 */ 147 len += skb->data - skb_transport_header(skb); 148 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) || 149 /* If PSH is not set, packet should be 150 * full sized, provided peer TCP is not badly broken. 151 * This observation (if it is correct 8)) allows 152 * to handle super-low mtu links fairly. 153 */ 154 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) && 155 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) { 156 /* Subtract also invariant (if peer is RFC compliant), 157 * tcp header plus fixed timestamp option length. 158 * Resulting "len" is MSS free of SACK jitter. 159 */ 160 len -= tcp_sk(sk)->tcp_header_len; 161 icsk->icsk_ack.last_seg_size = len; 162 if (len == lss) { 163 icsk->icsk_ack.rcv_mss = len; 164 return; 165 } 166 } 167 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED) 168 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2; 169 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED; 170 } 171 } 172 173 static void tcp_incr_quickack(struct sock *sk) 174 { 175 struct inet_connection_sock *icsk = inet_csk(sk); 176 unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss); 177 178 if (quickacks == 0) 179 quickacks = 2; 180 if (quickacks > icsk->icsk_ack.quick) 181 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS); 182 } 183 184 static void tcp_enter_quickack_mode(struct sock *sk) 185 { 186 struct inet_connection_sock *icsk = inet_csk(sk); 187 tcp_incr_quickack(sk); 188 icsk->icsk_ack.pingpong = 0; 189 icsk->icsk_ack.ato = TCP_ATO_MIN; 190 } 191 192 /* Send ACKs quickly, if "quick" count is not exhausted 193 * and the session is not interactive. 194 */ 195 196 static inline int tcp_in_quickack_mode(const struct sock *sk) 197 { 198 const struct inet_connection_sock *icsk = inet_csk(sk); 199 return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong; 200 } 201 202 static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp) 203 { 204 if (tp->ecn_flags & TCP_ECN_OK) 205 tp->ecn_flags |= TCP_ECN_QUEUE_CWR; 206 } 207 208 static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb) 209 { 210 if (tcp_hdr(skb)->cwr) 211 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR; 212 } 213 214 static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp) 215 { 216 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR; 217 } 218 219 static inline void TCP_ECN_check_ce(struct tcp_sock *tp, const struct sk_buff *skb) 220 { 221 if (!(tp->ecn_flags & TCP_ECN_OK)) 222 return; 223 224 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) { 225 case INET_ECN_NOT_ECT: 226 /* Funny extension: if ECT is not set on a segment, 227 * and we already seen ECT on a previous segment, 228 * it is probably a retransmit. 229 */ 230 if (tp->ecn_flags & TCP_ECN_SEEN) 231 tcp_enter_quickack_mode((struct sock *)tp); 232 break; 233 case INET_ECN_CE: 234 tp->ecn_flags |= TCP_ECN_DEMAND_CWR; 235 /* fallinto */ 236 default: 237 tp->ecn_flags |= TCP_ECN_SEEN; 238 } 239 } 240 241 static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th) 242 { 243 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr)) 244 tp->ecn_flags &= ~TCP_ECN_OK; 245 } 246 247 static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th) 248 { 249 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr)) 250 tp->ecn_flags &= ~TCP_ECN_OK; 251 } 252 253 static inline int TCP_ECN_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th) 254 { 255 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK)) 256 return 1; 257 return 0; 258 } 259 260 /* Buffer size and advertised window tuning. 261 * 262 * 1. Tuning sk->sk_sndbuf, when connection enters established state. 263 */ 264 265 static void tcp_fixup_sndbuf(struct sock *sk) 266 { 267 int sndmem = SKB_TRUESIZE(tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER); 268 269 sndmem *= TCP_INIT_CWND; 270 if (sk->sk_sndbuf < sndmem) 271 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]); 272 } 273 274 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh) 275 * 276 * All tcp_full_space() is split to two parts: "network" buffer, allocated 277 * forward and advertised in receiver window (tp->rcv_wnd) and 278 * "application buffer", required to isolate scheduling/application 279 * latencies from network. 280 * window_clamp is maximal advertised window. It can be less than 281 * tcp_full_space(), in this case tcp_full_space() - window_clamp 282 * is reserved for "application" buffer. The less window_clamp is 283 * the smoother our behaviour from viewpoint of network, but the lower 284 * throughput and the higher sensitivity of the connection to losses. 8) 285 * 286 * rcv_ssthresh is more strict window_clamp used at "slow start" 287 * phase to predict further behaviour of this connection. 288 * It is used for two goals: 289 * - to enforce header prediction at sender, even when application 290 * requires some significant "application buffer". It is check #1. 291 * - to prevent pruning of receive queue because of misprediction 292 * of receiver window. Check #2. 293 * 294 * The scheme does not work when sender sends good segments opening 295 * window and then starts to feed us spaghetti. But it should work 296 * in common situations. Otherwise, we have to rely on queue collapsing. 297 */ 298 299 /* Slow part of check#2. */ 300 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb) 301 { 302 struct tcp_sock *tp = tcp_sk(sk); 303 /* Optimize this! */ 304 int truesize = tcp_win_from_space(skb->truesize) >> 1; 305 int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1; 306 307 while (tp->rcv_ssthresh <= window) { 308 if (truesize <= skb->len) 309 return 2 * inet_csk(sk)->icsk_ack.rcv_mss; 310 311 truesize >>= 1; 312 window >>= 1; 313 } 314 return 0; 315 } 316 317 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb) 318 { 319 struct tcp_sock *tp = tcp_sk(sk); 320 321 /* Check #1 */ 322 if (tp->rcv_ssthresh < tp->window_clamp && 323 (int)tp->rcv_ssthresh < tcp_space(sk) && 324 !sk_under_memory_pressure(sk)) { 325 int incr; 326 327 /* Check #2. Increase window, if skb with such overhead 328 * will fit to rcvbuf in future. 329 */ 330 if (tcp_win_from_space(skb->truesize) <= skb->len) 331 incr = 2 * tp->advmss; 332 else 333 incr = __tcp_grow_window(sk, skb); 334 335 if (incr) { 336 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr, 337 tp->window_clamp); 338 inet_csk(sk)->icsk_ack.quick |= 1; 339 } 340 } 341 } 342 343 /* 3. Tuning rcvbuf, when connection enters established state. */ 344 345 static void tcp_fixup_rcvbuf(struct sock *sk) 346 { 347 u32 mss = tcp_sk(sk)->advmss; 348 u32 icwnd = TCP_DEFAULT_INIT_RCVWND; 349 int rcvmem; 350 351 /* Limit to 10 segments if mss <= 1460, 352 * or 14600/mss segments, with a minimum of two segments. 353 */ 354 if (mss > 1460) 355 icwnd = max_t(u32, (1460 * TCP_DEFAULT_INIT_RCVWND) / mss, 2); 356 357 rcvmem = SKB_TRUESIZE(mss + MAX_TCP_HEADER); 358 while (tcp_win_from_space(rcvmem) < mss) 359 rcvmem += 128; 360 361 rcvmem *= icwnd; 362 363 if (sk->sk_rcvbuf < rcvmem) 364 sk->sk_rcvbuf = min(rcvmem, sysctl_tcp_rmem[2]); 365 } 366 367 /* 4. Try to fixup all. It is made immediately after connection enters 368 * established state. 369 */ 370 static void tcp_init_buffer_space(struct sock *sk) 371 { 372 struct tcp_sock *tp = tcp_sk(sk); 373 int maxwin; 374 375 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) 376 tcp_fixup_rcvbuf(sk); 377 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK)) 378 tcp_fixup_sndbuf(sk); 379 380 tp->rcvq_space.space = tp->rcv_wnd; 381 382 maxwin = tcp_full_space(sk); 383 384 if (tp->window_clamp >= maxwin) { 385 tp->window_clamp = maxwin; 386 387 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss) 388 tp->window_clamp = max(maxwin - 389 (maxwin >> sysctl_tcp_app_win), 390 4 * tp->advmss); 391 } 392 393 /* Force reservation of one segment. */ 394 if (sysctl_tcp_app_win && 395 tp->window_clamp > 2 * tp->advmss && 396 tp->window_clamp + tp->advmss > maxwin) 397 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss); 398 399 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp); 400 tp->snd_cwnd_stamp = tcp_time_stamp; 401 } 402 403 /* 5. Recalculate window clamp after socket hit its memory bounds. */ 404 static void tcp_clamp_window(struct sock *sk) 405 { 406 struct tcp_sock *tp = tcp_sk(sk); 407 struct inet_connection_sock *icsk = inet_csk(sk); 408 409 icsk->icsk_ack.quick = 0; 410 411 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] && 412 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) && 413 !sk_under_memory_pressure(sk) && 414 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) { 415 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc), 416 sysctl_tcp_rmem[2]); 417 } 418 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf) 419 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss); 420 } 421 422 /* Initialize RCV_MSS value. 423 * RCV_MSS is an our guess about MSS used by the peer. 424 * We haven't any direct information about the MSS. 425 * It's better to underestimate the RCV_MSS rather than overestimate. 426 * Overestimations make us ACKing less frequently than needed. 427 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss(). 428 */ 429 void tcp_initialize_rcv_mss(struct sock *sk) 430 { 431 const struct tcp_sock *tp = tcp_sk(sk); 432 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache); 433 434 hint = min(hint, tp->rcv_wnd / 2); 435 hint = min(hint, TCP_MSS_DEFAULT); 436 hint = max(hint, TCP_MIN_MSS); 437 438 inet_csk(sk)->icsk_ack.rcv_mss = hint; 439 } 440 EXPORT_SYMBOL(tcp_initialize_rcv_mss); 441 442 /* Receiver "autotuning" code. 443 * 444 * The algorithm for RTT estimation w/o timestamps is based on 445 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL. 446 * <http://public.lanl.gov/radiant/pubs.html#DRS> 447 * 448 * More detail on this code can be found at 449 * <http://staff.psc.edu/jheffner/>, 450 * though this reference is out of date. A new paper 451 * is pending. 452 */ 453 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep) 454 { 455 u32 new_sample = tp->rcv_rtt_est.rtt; 456 long m = sample; 457 458 if (m == 0) 459 m = 1; 460 461 if (new_sample != 0) { 462 /* If we sample in larger samples in the non-timestamp 463 * case, we could grossly overestimate the RTT especially 464 * with chatty applications or bulk transfer apps which 465 * are stalled on filesystem I/O. 466 * 467 * Also, since we are only going for a minimum in the 468 * non-timestamp case, we do not smooth things out 469 * else with timestamps disabled convergence takes too 470 * long. 471 */ 472 if (!win_dep) { 473 m -= (new_sample >> 3); 474 new_sample += m; 475 } else if (m < new_sample) 476 new_sample = m << 3; 477 } else { 478 /* No previous measure. */ 479 new_sample = m << 3; 480 } 481 482 if (tp->rcv_rtt_est.rtt != new_sample) 483 tp->rcv_rtt_est.rtt = new_sample; 484 } 485 486 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp) 487 { 488 if (tp->rcv_rtt_est.time == 0) 489 goto new_measure; 490 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq)) 491 return; 492 tcp_rcv_rtt_update(tp, jiffies - tp->rcv_rtt_est.time, 1); 493 494 new_measure: 495 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd; 496 tp->rcv_rtt_est.time = tcp_time_stamp; 497 } 498 499 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, 500 const struct sk_buff *skb) 501 { 502 struct tcp_sock *tp = tcp_sk(sk); 503 if (tp->rx_opt.rcv_tsecr && 504 (TCP_SKB_CB(skb)->end_seq - 505 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss)) 506 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0); 507 } 508 509 /* 510 * This function should be called every time data is copied to user space. 511 * It calculates the appropriate TCP receive buffer space. 512 */ 513 void tcp_rcv_space_adjust(struct sock *sk) 514 { 515 struct tcp_sock *tp = tcp_sk(sk); 516 int time; 517 int space; 518 519 if (tp->rcvq_space.time == 0) 520 goto new_measure; 521 522 time = tcp_time_stamp - tp->rcvq_space.time; 523 if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0) 524 return; 525 526 space = 2 * (tp->copied_seq - tp->rcvq_space.seq); 527 528 space = max(tp->rcvq_space.space, space); 529 530 if (tp->rcvq_space.space != space) { 531 int rcvmem; 532 533 tp->rcvq_space.space = space; 534 535 if (sysctl_tcp_moderate_rcvbuf && 536 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) { 537 int new_clamp = space; 538 539 /* Receive space grows, normalize in order to 540 * take into account packet headers and sk_buff 541 * structure overhead. 542 */ 543 space /= tp->advmss; 544 if (!space) 545 space = 1; 546 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER); 547 while (tcp_win_from_space(rcvmem) < tp->advmss) 548 rcvmem += 128; 549 space *= rcvmem; 550 space = min(space, sysctl_tcp_rmem[2]); 551 if (space > sk->sk_rcvbuf) { 552 sk->sk_rcvbuf = space; 553 554 /* Make the window clamp follow along. */ 555 tp->window_clamp = new_clamp; 556 } 557 } 558 } 559 560 new_measure: 561 tp->rcvq_space.seq = tp->copied_seq; 562 tp->rcvq_space.time = tcp_time_stamp; 563 } 564 565 /* There is something which you must keep in mind when you analyze the 566 * behavior of the tp->ato delayed ack timeout interval. When a 567 * connection starts up, we want to ack as quickly as possible. The 568 * problem is that "good" TCP's do slow start at the beginning of data 569 * transmission. The means that until we send the first few ACK's the 570 * sender will sit on his end and only queue most of his data, because 571 * he can only send snd_cwnd unacked packets at any given time. For 572 * each ACK we send, he increments snd_cwnd and transmits more of his 573 * queue. -DaveM 574 */ 575 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb) 576 { 577 struct tcp_sock *tp = tcp_sk(sk); 578 struct inet_connection_sock *icsk = inet_csk(sk); 579 u32 now; 580 581 inet_csk_schedule_ack(sk); 582 583 tcp_measure_rcv_mss(sk, skb); 584 585 tcp_rcv_rtt_measure(tp); 586 587 now = tcp_time_stamp; 588 589 if (!icsk->icsk_ack.ato) { 590 /* The _first_ data packet received, initialize 591 * delayed ACK engine. 592 */ 593 tcp_incr_quickack(sk); 594 icsk->icsk_ack.ato = TCP_ATO_MIN; 595 } else { 596 int m = now - icsk->icsk_ack.lrcvtime; 597 598 if (m <= TCP_ATO_MIN / 2) { 599 /* The fastest case is the first. */ 600 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2; 601 } else if (m < icsk->icsk_ack.ato) { 602 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m; 603 if (icsk->icsk_ack.ato > icsk->icsk_rto) 604 icsk->icsk_ack.ato = icsk->icsk_rto; 605 } else if (m > icsk->icsk_rto) { 606 /* Too long gap. Apparently sender failed to 607 * restart window, so that we send ACKs quickly. 608 */ 609 tcp_incr_quickack(sk); 610 sk_mem_reclaim(sk); 611 } 612 } 613 icsk->icsk_ack.lrcvtime = now; 614 615 TCP_ECN_check_ce(tp, skb); 616 617 if (skb->len >= 128) 618 tcp_grow_window(sk, skb); 619 } 620 621 /* Called to compute a smoothed rtt estimate. The data fed to this 622 * routine either comes from timestamps, or from segments that were 623 * known _not_ to have been retransmitted [see Karn/Partridge 624 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88 625 * piece by Van Jacobson. 626 * NOTE: the next three routines used to be one big routine. 627 * To save cycles in the RFC 1323 implementation it was better to break 628 * it up into three procedures. -- erics 629 */ 630 static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt) 631 { 632 struct tcp_sock *tp = tcp_sk(sk); 633 long m = mrtt; /* RTT */ 634 635 /* The following amusing code comes from Jacobson's 636 * article in SIGCOMM '88. Note that rtt and mdev 637 * are scaled versions of rtt and mean deviation. 638 * This is designed to be as fast as possible 639 * m stands for "measurement". 640 * 641 * On a 1990 paper the rto value is changed to: 642 * RTO = rtt + 4 * mdev 643 * 644 * Funny. This algorithm seems to be very broken. 645 * These formulae increase RTO, when it should be decreased, increase 646 * too slowly, when it should be increased quickly, decrease too quickly 647 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely 648 * does not matter how to _calculate_ it. Seems, it was trap 649 * that VJ failed to avoid. 8) 650 */ 651 if (m == 0) 652 m = 1; 653 if (tp->srtt != 0) { 654 m -= (tp->srtt >> 3); /* m is now error in rtt est */ 655 tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */ 656 if (m < 0) { 657 m = -m; /* m is now abs(error) */ 658 m -= (tp->mdev >> 2); /* similar update on mdev */ 659 /* This is similar to one of Eifel findings. 660 * Eifel blocks mdev updates when rtt decreases. 661 * This solution is a bit different: we use finer gain 662 * for mdev in this case (alpha*beta). 663 * Like Eifel it also prevents growth of rto, 664 * but also it limits too fast rto decreases, 665 * happening in pure Eifel. 666 */ 667 if (m > 0) 668 m >>= 3; 669 } else { 670 m -= (tp->mdev >> 2); /* similar update on mdev */ 671 } 672 tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */ 673 if (tp->mdev > tp->mdev_max) { 674 tp->mdev_max = tp->mdev; 675 if (tp->mdev_max > tp->rttvar) 676 tp->rttvar = tp->mdev_max; 677 } 678 if (after(tp->snd_una, tp->rtt_seq)) { 679 if (tp->mdev_max < tp->rttvar) 680 tp->rttvar -= (tp->rttvar - tp->mdev_max) >> 2; 681 tp->rtt_seq = tp->snd_nxt; 682 tp->mdev_max = tcp_rto_min(sk); 683 } 684 } else { 685 /* no previous measure. */ 686 tp->srtt = m << 3; /* take the measured time to be rtt */ 687 tp->mdev = m << 1; /* make sure rto = 3*rtt */ 688 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk)); 689 tp->rtt_seq = tp->snd_nxt; 690 } 691 } 692 693 /* Calculate rto without backoff. This is the second half of Van Jacobson's 694 * routine referred to above. 695 */ 696 static inline void tcp_set_rto(struct sock *sk) 697 { 698 const struct tcp_sock *tp = tcp_sk(sk); 699 /* Old crap is replaced with new one. 8) 700 * 701 * More seriously: 702 * 1. If rtt variance happened to be less 50msec, it is hallucination. 703 * It cannot be less due to utterly erratic ACK generation made 704 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_ 705 * to do with delayed acks, because at cwnd>2 true delack timeout 706 * is invisible. Actually, Linux-2.4 also generates erratic 707 * ACKs in some circumstances. 708 */ 709 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp); 710 711 /* 2. Fixups made earlier cannot be right. 712 * If we do not estimate RTO correctly without them, 713 * all the algo is pure shit and should be replaced 714 * with correct one. It is exactly, which we pretend to do. 715 */ 716 717 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo 718 * guarantees that rto is higher. 719 */ 720 tcp_bound_rto(sk); 721 } 722 723 /* Save metrics learned by this TCP session. 724 This function is called only, when TCP finishes successfully 725 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE. 726 */ 727 void tcp_update_metrics(struct sock *sk) 728 { 729 struct tcp_sock *tp = tcp_sk(sk); 730 struct dst_entry *dst = __sk_dst_get(sk); 731 732 if (sysctl_tcp_nometrics_save) 733 return; 734 735 dst_confirm(dst); 736 737 if (dst && (dst->flags & DST_HOST)) { 738 const struct inet_connection_sock *icsk = inet_csk(sk); 739 int m; 740 unsigned long rtt; 741 742 if (icsk->icsk_backoff || !tp->srtt) { 743 /* This session failed to estimate rtt. Why? 744 * Probably, no packets returned in time. 745 * Reset our results. 746 */ 747 if (!(dst_metric_locked(dst, RTAX_RTT))) 748 dst_metric_set(dst, RTAX_RTT, 0); 749 return; 750 } 751 752 rtt = dst_metric_rtt(dst, RTAX_RTT); 753 m = rtt - tp->srtt; 754 755 /* If newly calculated rtt larger than stored one, 756 * store new one. Otherwise, use EWMA. Remember, 757 * rtt overestimation is always better than underestimation. 758 */ 759 if (!(dst_metric_locked(dst, RTAX_RTT))) { 760 if (m <= 0) 761 set_dst_metric_rtt(dst, RTAX_RTT, tp->srtt); 762 else 763 set_dst_metric_rtt(dst, RTAX_RTT, rtt - (m >> 3)); 764 } 765 766 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) { 767 unsigned long var; 768 if (m < 0) 769 m = -m; 770 771 /* Scale deviation to rttvar fixed point */ 772 m >>= 1; 773 if (m < tp->mdev) 774 m = tp->mdev; 775 776 var = dst_metric_rtt(dst, RTAX_RTTVAR); 777 if (m >= var) 778 var = m; 779 else 780 var -= (var - m) >> 2; 781 782 set_dst_metric_rtt(dst, RTAX_RTTVAR, var); 783 } 784 785 if (tcp_in_initial_slowstart(tp)) { 786 /* Slow start still did not finish. */ 787 if (dst_metric(dst, RTAX_SSTHRESH) && 788 !dst_metric_locked(dst, RTAX_SSTHRESH) && 789 (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH)) 790 dst_metric_set(dst, RTAX_SSTHRESH, tp->snd_cwnd >> 1); 791 if (!dst_metric_locked(dst, RTAX_CWND) && 792 tp->snd_cwnd > dst_metric(dst, RTAX_CWND)) 793 dst_metric_set(dst, RTAX_CWND, tp->snd_cwnd); 794 } else if (tp->snd_cwnd > tp->snd_ssthresh && 795 icsk->icsk_ca_state == TCP_CA_Open) { 796 /* Cong. avoidance phase, cwnd is reliable. */ 797 if (!dst_metric_locked(dst, RTAX_SSTHRESH)) 798 dst_metric_set(dst, RTAX_SSTHRESH, 799 max(tp->snd_cwnd >> 1, tp->snd_ssthresh)); 800 if (!dst_metric_locked(dst, RTAX_CWND)) 801 dst_metric_set(dst, RTAX_CWND, 802 (dst_metric(dst, RTAX_CWND) + 803 tp->snd_cwnd) >> 1); 804 } else { 805 /* Else slow start did not finish, cwnd is non-sense, 806 ssthresh may be also invalid. 807 */ 808 if (!dst_metric_locked(dst, RTAX_CWND)) 809 dst_metric_set(dst, RTAX_CWND, 810 (dst_metric(dst, RTAX_CWND) + 811 tp->snd_ssthresh) >> 1); 812 if (dst_metric(dst, RTAX_SSTHRESH) && 813 !dst_metric_locked(dst, RTAX_SSTHRESH) && 814 tp->snd_ssthresh > dst_metric(dst, RTAX_SSTHRESH)) 815 dst_metric_set(dst, RTAX_SSTHRESH, tp->snd_ssthresh); 816 } 817 818 if (!dst_metric_locked(dst, RTAX_REORDERING)) { 819 if (dst_metric(dst, RTAX_REORDERING) < tp->reordering && 820 tp->reordering != sysctl_tcp_reordering) 821 dst_metric_set(dst, RTAX_REORDERING, tp->reordering); 822 } 823 } 824 } 825 826 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst) 827 { 828 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0); 829 830 if (!cwnd) 831 cwnd = TCP_INIT_CWND; 832 return min_t(__u32, cwnd, tp->snd_cwnd_clamp); 833 } 834 835 /* Set slow start threshold and cwnd not falling to slow start */ 836 void tcp_enter_cwr(struct sock *sk, const int set_ssthresh) 837 { 838 struct tcp_sock *tp = tcp_sk(sk); 839 const struct inet_connection_sock *icsk = inet_csk(sk); 840 841 tp->prior_ssthresh = 0; 842 tp->bytes_acked = 0; 843 if (icsk->icsk_ca_state < TCP_CA_CWR) { 844 tp->undo_marker = 0; 845 if (set_ssthresh) 846 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 847 tp->snd_cwnd = min(tp->snd_cwnd, 848 tcp_packets_in_flight(tp) + 1U); 849 tp->snd_cwnd_cnt = 0; 850 tp->high_seq = tp->snd_nxt; 851 tp->snd_cwnd_stamp = tcp_time_stamp; 852 TCP_ECN_queue_cwr(tp); 853 854 tcp_set_ca_state(sk, TCP_CA_CWR); 855 } 856 } 857 858 /* 859 * Packet counting of FACK is based on in-order assumptions, therefore TCP 860 * disables it when reordering is detected 861 */ 862 static void tcp_disable_fack(struct tcp_sock *tp) 863 { 864 /* RFC3517 uses different metric in lost marker => reset on change */ 865 if (tcp_is_fack(tp)) 866 tp->lost_skb_hint = NULL; 867 tp->rx_opt.sack_ok &= ~TCP_FACK_ENABLED; 868 } 869 870 /* Take a notice that peer is sending D-SACKs */ 871 static void tcp_dsack_seen(struct tcp_sock *tp) 872 { 873 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN; 874 } 875 876 /* Initialize metrics on socket. */ 877 878 static void tcp_init_metrics(struct sock *sk) 879 { 880 struct tcp_sock *tp = tcp_sk(sk); 881 struct dst_entry *dst = __sk_dst_get(sk); 882 883 if (dst == NULL) 884 goto reset; 885 886 dst_confirm(dst); 887 888 if (dst_metric_locked(dst, RTAX_CWND)) 889 tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND); 890 if (dst_metric(dst, RTAX_SSTHRESH)) { 891 tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH); 892 if (tp->snd_ssthresh > tp->snd_cwnd_clamp) 893 tp->snd_ssthresh = tp->snd_cwnd_clamp; 894 } else { 895 /* ssthresh may have been reduced unnecessarily during. 896 * 3WHS. Restore it back to its initial default. 897 */ 898 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH; 899 } 900 if (dst_metric(dst, RTAX_REORDERING) && 901 tp->reordering != dst_metric(dst, RTAX_REORDERING)) { 902 tcp_disable_fack(tp); 903 tp->reordering = dst_metric(dst, RTAX_REORDERING); 904 } 905 906 if (dst_metric(dst, RTAX_RTT) == 0 || tp->srtt == 0) 907 goto reset; 908 909 /* Initial rtt is determined from SYN,SYN-ACK. 910 * The segment is small and rtt may appear much 911 * less than real one. Use per-dst memory 912 * to make it more realistic. 913 * 914 * A bit of theory. RTT is time passed after "normal" sized packet 915 * is sent until it is ACKed. In normal circumstances sending small 916 * packets force peer to delay ACKs and calculation is correct too. 917 * The algorithm is adaptive and, provided we follow specs, it 918 * NEVER underestimate RTT. BUT! If peer tries to make some clever 919 * tricks sort of "quick acks" for time long enough to decrease RTT 920 * to low value, and then abruptly stops to do it and starts to delay 921 * ACKs, wait for troubles. 922 */ 923 if (dst_metric_rtt(dst, RTAX_RTT) > tp->srtt) { 924 tp->srtt = dst_metric_rtt(dst, RTAX_RTT); 925 tp->rtt_seq = tp->snd_nxt; 926 } 927 if (dst_metric_rtt(dst, RTAX_RTTVAR) > tp->mdev) { 928 tp->mdev = dst_metric_rtt(dst, RTAX_RTTVAR); 929 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk)); 930 } 931 tcp_set_rto(sk); 932 reset: 933 if (tp->srtt == 0) { 934 /* RFC2988bis: We've failed to get a valid RTT sample from 935 * 3WHS. This is most likely due to retransmission, 936 * including spurious one. Reset the RTO back to 3secs 937 * from the more aggressive 1sec to avoid more spurious 938 * retransmission. 939 */ 940 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_FALLBACK; 941 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_FALLBACK; 942 } 943 /* Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been 944 * retransmitted. In light of RFC2988bis' more aggressive 1sec 945 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK 946 * retransmission has occurred. 947 */ 948 if (tp->total_retrans > 1) 949 tp->snd_cwnd = 1; 950 else 951 tp->snd_cwnd = tcp_init_cwnd(tp, dst); 952 tp->snd_cwnd_stamp = tcp_time_stamp; 953 } 954 955 static void tcp_update_reordering(struct sock *sk, const int metric, 956 const int ts) 957 { 958 struct tcp_sock *tp = tcp_sk(sk); 959 if (metric > tp->reordering) { 960 int mib_idx; 961 962 tp->reordering = min(TCP_MAX_REORDERING, metric); 963 964 /* This exciting event is worth to be remembered. 8) */ 965 if (ts) 966 mib_idx = LINUX_MIB_TCPTSREORDER; 967 else if (tcp_is_reno(tp)) 968 mib_idx = LINUX_MIB_TCPRENOREORDER; 969 else if (tcp_is_fack(tp)) 970 mib_idx = LINUX_MIB_TCPFACKREORDER; 971 else 972 mib_idx = LINUX_MIB_TCPSACKREORDER; 973 974 NET_INC_STATS_BH(sock_net(sk), mib_idx); 975 #if FASTRETRANS_DEBUG > 1 976 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n", 977 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state, 978 tp->reordering, 979 tp->fackets_out, 980 tp->sacked_out, 981 tp->undo_marker ? tp->undo_retrans : 0); 982 #endif 983 tcp_disable_fack(tp); 984 } 985 } 986 987 /* This must be called before lost_out is incremented */ 988 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb) 989 { 990 if ((tp->retransmit_skb_hint == NULL) || 991 before(TCP_SKB_CB(skb)->seq, 992 TCP_SKB_CB(tp->retransmit_skb_hint)->seq)) 993 tp->retransmit_skb_hint = skb; 994 995 if (!tp->lost_out || 996 after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high)) 997 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq; 998 } 999 1000 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb) 1001 { 1002 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) { 1003 tcp_verify_retransmit_hint(tp, skb); 1004 1005 tp->lost_out += tcp_skb_pcount(skb); 1006 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 1007 } 1008 } 1009 1010 static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp, 1011 struct sk_buff *skb) 1012 { 1013 tcp_verify_retransmit_hint(tp, skb); 1014 1015 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) { 1016 tp->lost_out += tcp_skb_pcount(skb); 1017 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 1018 } 1019 } 1020 1021 /* This procedure tags the retransmission queue when SACKs arrive. 1022 * 1023 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L). 1024 * Packets in queue with these bits set are counted in variables 1025 * sacked_out, retrans_out and lost_out, correspondingly. 1026 * 1027 * Valid combinations are: 1028 * Tag InFlight Description 1029 * 0 1 - orig segment is in flight. 1030 * S 0 - nothing flies, orig reached receiver. 1031 * L 0 - nothing flies, orig lost by net. 1032 * R 2 - both orig and retransmit are in flight. 1033 * L|R 1 - orig is lost, retransmit is in flight. 1034 * S|R 1 - orig reached receiver, retrans is still in flight. 1035 * (L|S|R is logically valid, it could occur when L|R is sacked, 1036 * but it is equivalent to plain S and code short-curcuits it to S. 1037 * L|S is logically invalid, it would mean -1 packet in flight 8)) 1038 * 1039 * These 6 states form finite state machine, controlled by the following events: 1040 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue()) 1041 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue()) 1042 * 3. Loss detection event of two flavors: 1043 * A. Scoreboard estimator decided the packet is lost. 1044 * A'. Reno "three dupacks" marks head of queue lost. 1045 * A''. Its FACK modification, head until snd.fack is lost. 1046 * B. SACK arrives sacking SND.NXT at the moment, when the 1047 * segment was retransmitted. 1048 * 4. D-SACK added new rule: D-SACK changes any tag to S. 1049 * 1050 * It is pleasant to note, that state diagram turns out to be commutative, 1051 * so that we are allowed not to be bothered by order of our actions, 1052 * when multiple events arrive simultaneously. (see the function below). 1053 * 1054 * Reordering detection. 1055 * -------------------- 1056 * Reordering metric is maximal distance, which a packet can be displaced 1057 * in packet stream. With SACKs we can estimate it: 1058 * 1059 * 1. SACK fills old hole and the corresponding segment was not 1060 * ever retransmitted -> reordering. Alas, we cannot use it 1061 * when segment was retransmitted. 1062 * 2. The last flaw is solved with D-SACK. D-SACK arrives 1063 * for retransmitted and already SACKed segment -> reordering.. 1064 * Both of these heuristics are not used in Loss state, when we cannot 1065 * account for retransmits accurately. 1066 * 1067 * SACK block validation. 1068 * ---------------------- 1069 * 1070 * SACK block range validation checks that the received SACK block fits to 1071 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT. 1072 * Note that SND.UNA is not included to the range though being valid because 1073 * it means that the receiver is rather inconsistent with itself reporting 1074 * SACK reneging when it should advance SND.UNA. Such SACK block this is 1075 * perfectly valid, however, in light of RFC2018 which explicitly states 1076 * that "SACK block MUST reflect the newest segment. Even if the newest 1077 * segment is going to be discarded ...", not that it looks very clever 1078 * in case of head skb. Due to potentional receiver driven attacks, we 1079 * choose to avoid immediate execution of a walk in write queue due to 1080 * reneging and defer head skb's loss recovery to standard loss recovery 1081 * procedure that will eventually trigger (nothing forbids us doing this). 1082 * 1083 * Implements also blockage to start_seq wrap-around. Problem lies in the 1084 * fact that though start_seq (s) is before end_seq (i.e., not reversed), 1085 * there's no guarantee that it will be before snd_nxt (n). The problem 1086 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt 1087 * wrap (s_w): 1088 * 1089 * <- outs wnd -> <- wrapzone -> 1090 * u e n u_w e_w s n_w 1091 * | | | | | | | 1092 * |<------------+------+----- TCP seqno space --------------+---------->| 1093 * ...-- <2^31 ->| |<--------... 1094 * ...---- >2^31 ------>| |<--------... 1095 * 1096 * Current code wouldn't be vulnerable but it's better still to discard such 1097 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat 1098 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in 1099 * snd_nxt wrap -> snd_una region will then become "well defined", i.e., 1100 * equal to the ideal case (infinite seqno space without wrap caused issues). 1101 * 1102 * With D-SACK the lower bound is extended to cover sequence space below 1103 * SND.UNA down to undo_marker, which is the last point of interest. Yet 1104 * again, D-SACK block must not to go across snd_una (for the same reason as 1105 * for the normal SACK blocks, explained above). But there all simplicity 1106 * ends, TCP might receive valid D-SACKs below that. As long as they reside 1107 * fully below undo_marker they do not affect behavior in anyway and can 1108 * therefore be safely ignored. In rare cases (which are more or less 1109 * theoretical ones), the D-SACK will nicely cross that boundary due to skb 1110 * fragmentation and packet reordering past skb's retransmission. To consider 1111 * them correctly, the acceptable range must be extended even more though 1112 * the exact amount is rather hard to quantify. However, tp->max_window can 1113 * be used as an exaggerated estimate. 1114 */ 1115 static int tcp_is_sackblock_valid(struct tcp_sock *tp, int is_dsack, 1116 u32 start_seq, u32 end_seq) 1117 { 1118 /* Too far in future, or reversed (interpretation is ambiguous) */ 1119 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq)) 1120 return 0; 1121 1122 /* Nasty start_seq wrap-around check (see comments above) */ 1123 if (!before(start_seq, tp->snd_nxt)) 1124 return 0; 1125 1126 /* In outstanding window? ...This is valid exit for D-SACKs too. 1127 * start_seq == snd_una is non-sensical (see comments above) 1128 */ 1129 if (after(start_seq, tp->snd_una)) 1130 return 1; 1131 1132 if (!is_dsack || !tp->undo_marker) 1133 return 0; 1134 1135 /* ...Then it's D-SACK, and must reside below snd_una completely */ 1136 if (after(end_seq, tp->snd_una)) 1137 return 0; 1138 1139 if (!before(start_seq, tp->undo_marker)) 1140 return 1; 1141 1142 /* Too old */ 1143 if (!after(end_seq, tp->undo_marker)) 1144 return 0; 1145 1146 /* Undo_marker boundary crossing (overestimates a lot). Known already: 1147 * start_seq < undo_marker and end_seq >= undo_marker. 1148 */ 1149 return !before(start_seq, end_seq - tp->max_window); 1150 } 1151 1152 /* Check for lost retransmit. This superb idea is borrowed from "ratehalving". 1153 * Event "B". Later note: FACK people cheated me again 8), we have to account 1154 * for reordering! Ugly, but should help. 1155 * 1156 * Search retransmitted skbs from write_queue that were sent when snd_nxt was 1157 * less than what is now known to be received by the other end (derived from 1158 * highest SACK block). Also calculate the lowest snd_nxt among the remaining 1159 * retransmitted skbs to avoid some costly processing per ACKs. 1160 */ 1161 static void tcp_mark_lost_retrans(struct sock *sk) 1162 { 1163 const struct inet_connection_sock *icsk = inet_csk(sk); 1164 struct tcp_sock *tp = tcp_sk(sk); 1165 struct sk_buff *skb; 1166 int cnt = 0; 1167 u32 new_low_seq = tp->snd_nxt; 1168 u32 received_upto = tcp_highest_sack_seq(tp); 1169 1170 if (!tcp_is_fack(tp) || !tp->retrans_out || 1171 !after(received_upto, tp->lost_retrans_low) || 1172 icsk->icsk_ca_state != TCP_CA_Recovery) 1173 return; 1174 1175 tcp_for_write_queue(skb, sk) { 1176 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq; 1177 1178 if (skb == tcp_send_head(sk)) 1179 break; 1180 if (cnt == tp->retrans_out) 1181 break; 1182 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 1183 continue; 1184 1185 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)) 1186 continue; 1187 1188 /* TODO: We would like to get rid of tcp_is_fack(tp) only 1189 * constraint here (see above) but figuring out that at 1190 * least tp->reordering SACK blocks reside between ack_seq 1191 * and received_upto is not easy task to do cheaply with 1192 * the available datastructures. 1193 * 1194 * Whether FACK should check here for tp->reordering segs 1195 * in-between one could argue for either way (it would be 1196 * rather simple to implement as we could count fack_count 1197 * during the walk and do tp->fackets_out - fack_count). 1198 */ 1199 if (after(received_upto, ack_seq)) { 1200 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 1201 tp->retrans_out -= tcp_skb_pcount(skb); 1202 1203 tcp_skb_mark_lost_uncond_verify(tp, skb); 1204 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT); 1205 } else { 1206 if (before(ack_seq, new_low_seq)) 1207 new_low_seq = ack_seq; 1208 cnt += tcp_skb_pcount(skb); 1209 } 1210 } 1211 1212 if (tp->retrans_out) 1213 tp->lost_retrans_low = new_low_seq; 1214 } 1215 1216 static int tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb, 1217 struct tcp_sack_block_wire *sp, int num_sacks, 1218 u32 prior_snd_una) 1219 { 1220 struct tcp_sock *tp = tcp_sk(sk); 1221 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq); 1222 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq); 1223 int dup_sack = 0; 1224 1225 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) { 1226 dup_sack = 1; 1227 tcp_dsack_seen(tp); 1228 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV); 1229 } else if (num_sacks > 1) { 1230 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq); 1231 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq); 1232 1233 if (!after(end_seq_0, end_seq_1) && 1234 !before(start_seq_0, start_seq_1)) { 1235 dup_sack = 1; 1236 tcp_dsack_seen(tp); 1237 NET_INC_STATS_BH(sock_net(sk), 1238 LINUX_MIB_TCPDSACKOFORECV); 1239 } 1240 } 1241 1242 /* D-SACK for already forgotten data... Do dumb counting. */ 1243 if (dup_sack && tp->undo_marker && tp->undo_retrans && 1244 !after(end_seq_0, prior_snd_una) && 1245 after(end_seq_0, tp->undo_marker)) 1246 tp->undo_retrans--; 1247 1248 return dup_sack; 1249 } 1250 1251 struct tcp_sacktag_state { 1252 int reord; 1253 int fack_count; 1254 int flag; 1255 }; 1256 1257 /* Check if skb is fully within the SACK block. In presence of GSO skbs, 1258 * the incoming SACK may not exactly match but we can find smaller MSS 1259 * aligned portion of it that matches. Therefore we might need to fragment 1260 * which may fail and creates some hassle (caller must handle error case 1261 * returns). 1262 * 1263 * FIXME: this could be merged to shift decision code 1264 */ 1265 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb, 1266 u32 start_seq, u32 end_seq) 1267 { 1268 int in_sack, err; 1269 unsigned int pkt_len; 1270 unsigned int mss; 1271 1272 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && 1273 !before(end_seq, TCP_SKB_CB(skb)->end_seq); 1274 1275 if (tcp_skb_pcount(skb) > 1 && !in_sack && 1276 after(TCP_SKB_CB(skb)->end_seq, start_seq)) { 1277 mss = tcp_skb_mss(skb); 1278 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); 1279 1280 if (!in_sack) { 1281 pkt_len = start_seq - TCP_SKB_CB(skb)->seq; 1282 if (pkt_len < mss) 1283 pkt_len = mss; 1284 } else { 1285 pkt_len = end_seq - TCP_SKB_CB(skb)->seq; 1286 if (pkt_len < mss) 1287 return -EINVAL; 1288 } 1289 1290 /* Round if necessary so that SACKs cover only full MSSes 1291 * and/or the remaining small portion (if present) 1292 */ 1293 if (pkt_len > mss) { 1294 unsigned int new_len = (pkt_len / mss) * mss; 1295 if (!in_sack && new_len < pkt_len) { 1296 new_len += mss; 1297 if (new_len > skb->len) 1298 return 0; 1299 } 1300 pkt_len = new_len; 1301 } 1302 err = tcp_fragment(sk, skb, pkt_len, mss); 1303 if (err < 0) 1304 return err; 1305 } 1306 1307 return in_sack; 1308 } 1309 1310 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */ 1311 static u8 tcp_sacktag_one(struct sock *sk, 1312 struct tcp_sacktag_state *state, u8 sacked, 1313 u32 start_seq, u32 end_seq, 1314 int dup_sack, int pcount) 1315 { 1316 struct tcp_sock *tp = tcp_sk(sk); 1317 int fack_count = state->fack_count; 1318 1319 /* Account D-SACK for retransmitted packet. */ 1320 if (dup_sack && (sacked & TCPCB_RETRANS)) { 1321 if (tp->undo_marker && tp->undo_retrans && 1322 after(end_seq, tp->undo_marker)) 1323 tp->undo_retrans--; 1324 if (sacked & TCPCB_SACKED_ACKED) 1325 state->reord = min(fack_count, state->reord); 1326 } 1327 1328 /* Nothing to do; acked frame is about to be dropped (was ACKed). */ 1329 if (!after(end_seq, tp->snd_una)) 1330 return sacked; 1331 1332 if (!(sacked & TCPCB_SACKED_ACKED)) { 1333 if (sacked & TCPCB_SACKED_RETRANS) { 1334 /* If the segment is not tagged as lost, 1335 * we do not clear RETRANS, believing 1336 * that retransmission is still in flight. 1337 */ 1338 if (sacked & TCPCB_LOST) { 1339 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS); 1340 tp->lost_out -= pcount; 1341 tp->retrans_out -= pcount; 1342 } 1343 } else { 1344 if (!(sacked & TCPCB_RETRANS)) { 1345 /* New sack for not retransmitted frame, 1346 * which was in hole. It is reordering. 1347 */ 1348 if (before(start_seq, 1349 tcp_highest_sack_seq(tp))) 1350 state->reord = min(fack_count, 1351 state->reord); 1352 1353 /* SACK enhanced F-RTO (RFC4138; Appendix B) */ 1354 if (!after(end_seq, tp->frto_highmark)) 1355 state->flag |= FLAG_ONLY_ORIG_SACKED; 1356 } 1357 1358 if (sacked & TCPCB_LOST) { 1359 sacked &= ~TCPCB_LOST; 1360 tp->lost_out -= pcount; 1361 } 1362 } 1363 1364 sacked |= TCPCB_SACKED_ACKED; 1365 state->flag |= FLAG_DATA_SACKED; 1366 tp->sacked_out += pcount; 1367 1368 fack_count += pcount; 1369 1370 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */ 1371 if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) && 1372 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq)) 1373 tp->lost_cnt_hint += pcount; 1374 1375 if (fack_count > tp->fackets_out) 1376 tp->fackets_out = fack_count; 1377 } 1378 1379 /* D-SACK. We can detect redundant retransmission in S|R and plain R 1380 * frames and clear it. undo_retrans is decreased above, L|R frames 1381 * are accounted above as well. 1382 */ 1383 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) { 1384 sacked &= ~TCPCB_SACKED_RETRANS; 1385 tp->retrans_out -= pcount; 1386 } 1387 1388 return sacked; 1389 } 1390 1391 /* Shift newly-SACKed bytes from this skb to the immediately previous 1392 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such. 1393 */ 1394 static int tcp_shifted_skb(struct sock *sk, struct sk_buff *skb, 1395 struct tcp_sacktag_state *state, 1396 unsigned int pcount, int shifted, int mss, 1397 int dup_sack) 1398 { 1399 struct tcp_sock *tp = tcp_sk(sk); 1400 struct sk_buff *prev = tcp_write_queue_prev(sk, skb); 1401 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */ 1402 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */ 1403 1404 BUG_ON(!pcount); 1405 1406 /* Adjust counters and hints for the newly sacked sequence 1407 * range but discard the return value since prev is already 1408 * marked. We must tag the range first because the seq 1409 * advancement below implicitly advances 1410 * tcp_highest_sack_seq() when skb is highest_sack. 1411 */ 1412 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked, 1413 start_seq, end_seq, dup_sack, pcount); 1414 1415 if (skb == tp->lost_skb_hint) 1416 tp->lost_cnt_hint += pcount; 1417 1418 TCP_SKB_CB(prev)->end_seq += shifted; 1419 TCP_SKB_CB(skb)->seq += shifted; 1420 1421 skb_shinfo(prev)->gso_segs += pcount; 1422 BUG_ON(skb_shinfo(skb)->gso_segs < pcount); 1423 skb_shinfo(skb)->gso_segs -= pcount; 1424 1425 /* When we're adding to gso_segs == 1, gso_size will be zero, 1426 * in theory this shouldn't be necessary but as long as DSACK 1427 * code can come after this skb later on it's better to keep 1428 * setting gso_size to something. 1429 */ 1430 if (!skb_shinfo(prev)->gso_size) { 1431 skb_shinfo(prev)->gso_size = mss; 1432 skb_shinfo(prev)->gso_type = sk->sk_gso_type; 1433 } 1434 1435 /* CHECKME: To clear or not to clear? Mimics normal skb currently */ 1436 if (skb_shinfo(skb)->gso_segs <= 1) { 1437 skb_shinfo(skb)->gso_size = 0; 1438 skb_shinfo(skb)->gso_type = 0; 1439 } 1440 1441 /* Difference in this won't matter, both ACKed by the same cumul. ACK */ 1442 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS); 1443 1444 if (skb->len > 0) { 1445 BUG_ON(!tcp_skb_pcount(skb)); 1446 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED); 1447 return 0; 1448 } 1449 1450 /* Whole SKB was eaten :-) */ 1451 1452 if (skb == tp->retransmit_skb_hint) 1453 tp->retransmit_skb_hint = prev; 1454 if (skb == tp->scoreboard_skb_hint) 1455 tp->scoreboard_skb_hint = prev; 1456 if (skb == tp->lost_skb_hint) { 1457 tp->lost_skb_hint = prev; 1458 tp->lost_cnt_hint -= tcp_skb_pcount(prev); 1459 } 1460 1461 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(prev)->tcp_flags; 1462 if (skb == tcp_highest_sack(sk)) 1463 tcp_advance_highest_sack(sk, skb); 1464 1465 tcp_unlink_write_queue(skb, sk); 1466 sk_wmem_free_skb(sk, skb); 1467 1468 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED); 1469 1470 return 1; 1471 } 1472 1473 /* I wish gso_size would have a bit more sane initialization than 1474 * something-or-zero which complicates things 1475 */ 1476 static int tcp_skb_seglen(const struct sk_buff *skb) 1477 { 1478 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb); 1479 } 1480 1481 /* Shifting pages past head area doesn't work */ 1482 static int skb_can_shift(const struct sk_buff *skb) 1483 { 1484 return !skb_headlen(skb) && skb_is_nonlinear(skb); 1485 } 1486 1487 /* Try collapsing SACK blocks spanning across multiple skbs to a single 1488 * skb. 1489 */ 1490 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb, 1491 struct tcp_sacktag_state *state, 1492 u32 start_seq, u32 end_seq, 1493 int dup_sack) 1494 { 1495 struct tcp_sock *tp = tcp_sk(sk); 1496 struct sk_buff *prev; 1497 int mss; 1498 int pcount = 0; 1499 int len; 1500 int in_sack; 1501 1502 if (!sk_can_gso(sk)) 1503 goto fallback; 1504 1505 /* Normally R but no L won't result in plain S */ 1506 if (!dup_sack && 1507 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS) 1508 goto fallback; 1509 if (!skb_can_shift(skb)) 1510 goto fallback; 1511 /* This frame is about to be dropped (was ACKed). */ 1512 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 1513 goto fallback; 1514 1515 /* Can only happen with delayed DSACK + discard craziness */ 1516 if (unlikely(skb == tcp_write_queue_head(sk))) 1517 goto fallback; 1518 prev = tcp_write_queue_prev(sk, skb); 1519 1520 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) 1521 goto fallback; 1522 1523 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && 1524 !before(end_seq, TCP_SKB_CB(skb)->end_seq); 1525 1526 if (in_sack) { 1527 len = skb->len; 1528 pcount = tcp_skb_pcount(skb); 1529 mss = tcp_skb_seglen(skb); 1530 1531 /* TODO: Fix DSACKs to not fragment already SACKed and we can 1532 * drop this restriction as unnecessary 1533 */ 1534 if (mss != tcp_skb_seglen(prev)) 1535 goto fallback; 1536 } else { 1537 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq)) 1538 goto noop; 1539 /* CHECKME: This is non-MSS split case only?, this will 1540 * cause skipped skbs due to advancing loop btw, original 1541 * has that feature too 1542 */ 1543 if (tcp_skb_pcount(skb) <= 1) 1544 goto noop; 1545 1546 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); 1547 if (!in_sack) { 1548 /* TODO: head merge to next could be attempted here 1549 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)), 1550 * though it might not be worth of the additional hassle 1551 * 1552 * ...we can probably just fallback to what was done 1553 * previously. We could try merging non-SACKed ones 1554 * as well but it probably isn't going to buy off 1555 * because later SACKs might again split them, and 1556 * it would make skb timestamp tracking considerably 1557 * harder problem. 1558 */ 1559 goto fallback; 1560 } 1561 1562 len = end_seq - TCP_SKB_CB(skb)->seq; 1563 BUG_ON(len < 0); 1564 BUG_ON(len > skb->len); 1565 1566 /* MSS boundaries should be honoured or else pcount will 1567 * severely break even though it makes things bit trickier. 1568 * Optimize common case to avoid most of the divides 1569 */ 1570 mss = tcp_skb_mss(skb); 1571 1572 /* TODO: Fix DSACKs to not fragment already SACKed and we can 1573 * drop this restriction as unnecessary 1574 */ 1575 if (mss != tcp_skb_seglen(prev)) 1576 goto fallback; 1577 1578 if (len == mss) { 1579 pcount = 1; 1580 } else if (len < mss) { 1581 goto noop; 1582 } else { 1583 pcount = len / mss; 1584 len = pcount * mss; 1585 } 1586 } 1587 1588 if (!skb_shift(prev, skb, len)) 1589 goto fallback; 1590 if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack)) 1591 goto out; 1592 1593 /* Hole filled allows collapsing with the next as well, this is very 1594 * useful when hole on every nth skb pattern happens 1595 */ 1596 if (prev == tcp_write_queue_tail(sk)) 1597 goto out; 1598 skb = tcp_write_queue_next(sk, prev); 1599 1600 if (!skb_can_shift(skb) || 1601 (skb == tcp_send_head(sk)) || 1602 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) || 1603 (mss != tcp_skb_seglen(skb))) 1604 goto out; 1605 1606 len = skb->len; 1607 if (skb_shift(prev, skb, len)) { 1608 pcount += tcp_skb_pcount(skb); 1609 tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0); 1610 } 1611 1612 out: 1613 state->fack_count += pcount; 1614 return prev; 1615 1616 noop: 1617 return skb; 1618 1619 fallback: 1620 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK); 1621 return NULL; 1622 } 1623 1624 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk, 1625 struct tcp_sack_block *next_dup, 1626 struct tcp_sacktag_state *state, 1627 u32 start_seq, u32 end_seq, 1628 int dup_sack_in) 1629 { 1630 struct tcp_sock *tp = tcp_sk(sk); 1631 struct sk_buff *tmp; 1632 1633 tcp_for_write_queue_from(skb, sk) { 1634 int in_sack = 0; 1635 int dup_sack = dup_sack_in; 1636 1637 if (skb == tcp_send_head(sk)) 1638 break; 1639 1640 /* queue is in-order => we can short-circuit the walk early */ 1641 if (!before(TCP_SKB_CB(skb)->seq, end_seq)) 1642 break; 1643 1644 if ((next_dup != NULL) && 1645 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) { 1646 in_sack = tcp_match_skb_to_sack(sk, skb, 1647 next_dup->start_seq, 1648 next_dup->end_seq); 1649 if (in_sack > 0) 1650 dup_sack = 1; 1651 } 1652 1653 /* skb reference here is a bit tricky to get right, since 1654 * shifting can eat and free both this skb and the next, 1655 * so not even _safe variant of the loop is enough. 1656 */ 1657 if (in_sack <= 0) { 1658 tmp = tcp_shift_skb_data(sk, skb, state, 1659 start_seq, end_seq, dup_sack); 1660 if (tmp != NULL) { 1661 if (tmp != skb) { 1662 skb = tmp; 1663 continue; 1664 } 1665 1666 in_sack = 0; 1667 } else { 1668 in_sack = tcp_match_skb_to_sack(sk, skb, 1669 start_seq, 1670 end_seq); 1671 } 1672 } 1673 1674 if (unlikely(in_sack < 0)) 1675 break; 1676 1677 if (in_sack) { 1678 TCP_SKB_CB(skb)->sacked = 1679 tcp_sacktag_one(sk, 1680 state, 1681 TCP_SKB_CB(skb)->sacked, 1682 TCP_SKB_CB(skb)->seq, 1683 TCP_SKB_CB(skb)->end_seq, 1684 dup_sack, 1685 tcp_skb_pcount(skb)); 1686 1687 if (!before(TCP_SKB_CB(skb)->seq, 1688 tcp_highest_sack_seq(tp))) 1689 tcp_advance_highest_sack(sk, skb); 1690 } 1691 1692 state->fack_count += tcp_skb_pcount(skb); 1693 } 1694 return skb; 1695 } 1696 1697 /* Avoid all extra work that is being done by sacktag while walking in 1698 * a normal way 1699 */ 1700 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk, 1701 struct tcp_sacktag_state *state, 1702 u32 skip_to_seq) 1703 { 1704 tcp_for_write_queue_from(skb, sk) { 1705 if (skb == tcp_send_head(sk)) 1706 break; 1707 1708 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq)) 1709 break; 1710 1711 state->fack_count += tcp_skb_pcount(skb); 1712 } 1713 return skb; 1714 } 1715 1716 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb, 1717 struct sock *sk, 1718 struct tcp_sack_block *next_dup, 1719 struct tcp_sacktag_state *state, 1720 u32 skip_to_seq) 1721 { 1722 if (next_dup == NULL) 1723 return skb; 1724 1725 if (before(next_dup->start_seq, skip_to_seq)) { 1726 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq); 1727 skb = tcp_sacktag_walk(skb, sk, NULL, state, 1728 next_dup->start_seq, next_dup->end_seq, 1729 1); 1730 } 1731 1732 return skb; 1733 } 1734 1735 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache) 1736 { 1737 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache); 1738 } 1739 1740 static int 1741 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb, 1742 u32 prior_snd_una) 1743 { 1744 const struct inet_connection_sock *icsk = inet_csk(sk); 1745 struct tcp_sock *tp = tcp_sk(sk); 1746 const unsigned char *ptr = (skb_transport_header(ack_skb) + 1747 TCP_SKB_CB(ack_skb)->sacked); 1748 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2); 1749 struct tcp_sack_block sp[TCP_NUM_SACKS]; 1750 struct tcp_sack_block *cache; 1751 struct tcp_sacktag_state state; 1752 struct sk_buff *skb; 1753 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3); 1754 int used_sacks; 1755 int found_dup_sack = 0; 1756 int i, j; 1757 int first_sack_index; 1758 1759 state.flag = 0; 1760 state.reord = tp->packets_out; 1761 1762 if (!tp->sacked_out) { 1763 if (WARN_ON(tp->fackets_out)) 1764 tp->fackets_out = 0; 1765 tcp_highest_sack_reset(sk); 1766 } 1767 1768 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire, 1769 num_sacks, prior_snd_una); 1770 if (found_dup_sack) 1771 state.flag |= FLAG_DSACKING_ACK; 1772 1773 /* Eliminate too old ACKs, but take into 1774 * account more or less fresh ones, they can 1775 * contain valid SACK info. 1776 */ 1777 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window)) 1778 return 0; 1779 1780 if (!tp->packets_out) 1781 goto out; 1782 1783 used_sacks = 0; 1784 first_sack_index = 0; 1785 for (i = 0; i < num_sacks; i++) { 1786 int dup_sack = !i && found_dup_sack; 1787 1788 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq); 1789 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq); 1790 1791 if (!tcp_is_sackblock_valid(tp, dup_sack, 1792 sp[used_sacks].start_seq, 1793 sp[used_sacks].end_seq)) { 1794 int mib_idx; 1795 1796 if (dup_sack) { 1797 if (!tp->undo_marker) 1798 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO; 1799 else 1800 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD; 1801 } else { 1802 /* Don't count olds caused by ACK reordering */ 1803 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) && 1804 !after(sp[used_sacks].end_seq, tp->snd_una)) 1805 continue; 1806 mib_idx = LINUX_MIB_TCPSACKDISCARD; 1807 } 1808 1809 NET_INC_STATS_BH(sock_net(sk), mib_idx); 1810 if (i == 0) 1811 first_sack_index = -1; 1812 continue; 1813 } 1814 1815 /* Ignore very old stuff early */ 1816 if (!after(sp[used_sacks].end_seq, prior_snd_una)) 1817 continue; 1818 1819 used_sacks++; 1820 } 1821 1822 /* order SACK blocks to allow in order walk of the retrans queue */ 1823 for (i = used_sacks - 1; i > 0; i--) { 1824 for (j = 0; j < i; j++) { 1825 if (after(sp[j].start_seq, sp[j + 1].start_seq)) { 1826 swap(sp[j], sp[j + 1]); 1827 1828 /* Track where the first SACK block goes to */ 1829 if (j == first_sack_index) 1830 first_sack_index = j + 1; 1831 } 1832 } 1833 } 1834 1835 skb = tcp_write_queue_head(sk); 1836 state.fack_count = 0; 1837 i = 0; 1838 1839 if (!tp->sacked_out) { 1840 /* It's already past, so skip checking against it */ 1841 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache); 1842 } else { 1843 cache = tp->recv_sack_cache; 1844 /* Skip empty blocks in at head of the cache */ 1845 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq && 1846 !cache->end_seq) 1847 cache++; 1848 } 1849 1850 while (i < used_sacks) { 1851 u32 start_seq = sp[i].start_seq; 1852 u32 end_seq = sp[i].end_seq; 1853 int dup_sack = (found_dup_sack && (i == first_sack_index)); 1854 struct tcp_sack_block *next_dup = NULL; 1855 1856 if (found_dup_sack && ((i + 1) == first_sack_index)) 1857 next_dup = &sp[i + 1]; 1858 1859 /* Skip too early cached blocks */ 1860 while (tcp_sack_cache_ok(tp, cache) && 1861 !before(start_seq, cache->end_seq)) 1862 cache++; 1863 1864 /* Can skip some work by looking recv_sack_cache? */ 1865 if (tcp_sack_cache_ok(tp, cache) && !dup_sack && 1866 after(end_seq, cache->start_seq)) { 1867 1868 /* Head todo? */ 1869 if (before(start_seq, cache->start_seq)) { 1870 skb = tcp_sacktag_skip(skb, sk, &state, 1871 start_seq); 1872 skb = tcp_sacktag_walk(skb, sk, next_dup, 1873 &state, 1874 start_seq, 1875 cache->start_seq, 1876 dup_sack); 1877 } 1878 1879 /* Rest of the block already fully processed? */ 1880 if (!after(end_seq, cache->end_seq)) 1881 goto advance_sp; 1882 1883 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup, 1884 &state, 1885 cache->end_seq); 1886 1887 /* ...tail remains todo... */ 1888 if (tcp_highest_sack_seq(tp) == cache->end_seq) { 1889 /* ...but better entrypoint exists! */ 1890 skb = tcp_highest_sack(sk); 1891 if (skb == NULL) 1892 break; 1893 state.fack_count = tp->fackets_out; 1894 cache++; 1895 goto walk; 1896 } 1897 1898 skb = tcp_sacktag_skip(skb, sk, &state, cache->end_seq); 1899 /* Check overlap against next cached too (past this one already) */ 1900 cache++; 1901 continue; 1902 } 1903 1904 if (!before(start_seq, tcp_highest_sack_seq(tp))) { 1905 skb = tcp_highest_sack(sk); 1906 if (skb == NULL) 1907 break; 1908 state.fack_count = tp->fackets_out; 1909 } 1910 skb = tcp_sacktag_skip(skb, sk, &state, start_seq); 1911 1912 walk: 1913 skb = tcp_sacktag_walk(skb, sk, next_dup, &state, 1914 start_seq, end_seq, dup_sack); 1915 1916 advance_sp: 1917 /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct 1918 * due to in-order walk 1919 */ 1920 if (after(end_seq, tp->frto_highmark)) 1921 state.flag &= ~FLAG_ONLY_ORIG_SACKED; 1922 1923 i++; 1924 } 1925 1926 /* Clear the head of the cache sack blocks so we can skip it next time */ 1927 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) { 1928 tp->recv_sack_cache[i].start_seq = 0; 1929 tp->recv_sack_cache[i].end_seq = 0; 1930 } 1931 for (j = 0; j < used_sacks; j++) 1932 tp->recv_sack_cache[i++] = sp[j]; 1933 1934 tcp_mark_lost_retrans(sk); 1935 1936 tcp_verify_left_out(tp); 1937 1938 if ((state.reord < tp->fackets_out) && 1939 ((icsk->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker) && 1940 (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark))) 1941 tcp_update_reordering(sk, tp->fackets_out - state.reord, 0); 1942 1943 out: 1944 1945 #if FASTRETRANS_DEBUG > 0 1946 WARN_ON((int)tp->sacked_out < 0); 1947 WARN_ON((int)tp->lost_out < 0); 1948 WARN_ON((int)tp->retrans_out < 0); 1949 WARN_ON((int)tcp_packets_in_flight(tp) < 0); 1950 #endif 1951 return state.flag; 1952 } 1953 1954 /* Limits sacked_out so that sum with lost_out isn't ever larger than 1955 * packets_out. Returns zero if sacked_out adjustement wasn't necessary. 1956 */ 1957 static int tcp_limit_reno_sacked(struct tcp_sock *tp) 1958 { 1959 u32 holes; 1960 1961 holes = max(tp->lost_out, 1U); 1962 holes = min(holes, tp->packets_out); 1963 1964 if ((tp->sacked_out + holes) > tp->packets_out) { 1965 tp->sacked_out = tp->packets_out - holes; 1966 return 1; 1967 } 1968 return 0; 1969 } 1970 1971 /* If we receive more dupacks than we expected counting segments 1972 * in assumption of absent reordering, interpret this as reordering. 1973 * The only another reason could be bug in receiver TCP. 1974 */ 1975 static void tcp_check_reno_reordering(struct sock *sk, const int addend) 1976 { 1977 struct tcp_sock *tp = tcp_sk(sk); 1978 if (tcp_limit_reno_sacked(tp)) 1979 tcp_update_reordering(sk, tp->packets_out + addend, 0); 1980 } 1981 1982 /* Emulate SACKs for SACKless connection: account for a new dupack. */ 1983 1984 static void tcp_add_reno_sack(struct sock *sk) 1985 { 1986 struct tcp_sock *tp = tcp_sk(sk); 1987 tp->sacked_out++; 1988 tcp_check_reno_reordering(sk, 0); 1989 tcp_verify_left_out(tp); 1990 } 1991 1992 /* Account for ACK, ACKing some data in Reno Recovery phase. */ 1993 1994 static void tcp_remove_reno_sacks(struct sock *sk, int acked) 1995 { 1996 struct tcp_sock *tp = tcp_sk(sk); 1997 1998 if (acked > 0) { 1999 /* One ACK acked hole. The rest eat duplicate ACKs. */ 2000 if (acked - 1 >= tp->sacked_out) 2001 tp->sacked_out = 0; 2002 else 2003 tp->sacked_out -= acked - 1; 2004 } 2005 tcp_check_reno_reordering(sk, acked); 2006 tcp_verify_left_out(tp); 2007 } 2008 2009 static inline void tcp_reset_reno_sack(struct tcp_sock *tp) 2010 { 2011 tp->sacked_out = 0; 2012 } 2013 2014 static int tcp_is_sackfrto(const struct tcp_sock *tp) 2015 { 2016 return (sysctl_tcp_frto == 0x2) && !tcp_is_reno(tp); 2017 } 2018 2019 /* F-RTO can only be used if TCP has never retransmitted anything other than 2020 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here) 2021 */ 2022 int tcp_use_frto(struct sock *sk) 2023 { 2024 const struct tcp_sock *tp = tcp_sk(sk); 2025 const struct inet_connection_sock *icsk = inet_csk(sk); 2026 struct sk_buff *skb; 2027 2028 if (!sysctl_tcp_frto) 2029 return 0; 2030 2031 /* MTU probe and F-RTO won't really play nicely along currently */ 2032 if (icsk->icsk_mtup.probe_size) 2033 return 0; 2034 2035 if (tcp_is_sackfrto(tp)) 2036 return 1; 2037 2038 /* Avoid expensive walking of rexmit queue if possible */ 2039 if (tp->retrans_out > 1) 2040 return 0; 2041 2042 skb = tcp_write_queue_head(sk); 2043 if (tcp_skb_is_last(sk, skb)) 2044 return 1; 2045 skb = tcp_write_queue_next(sk, skb); /* Skips head */ 2046 tcp_for_write_queue_from(skb, sk) { 2047 if (skb == tcp_send_head(sk)) 2048 break; 2049 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) 2050 return 0; 2051 /* Short-circuit when first non-SACKed skb has been checked */ 2052 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) 2053 break; 2054 } 2055 return 1; 2056 } 2057 2058 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO 2059 * recovery a bit and use heuristics in tcp_process_frto() to detect if 2060 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to 2061 * keep retrans_out counting accurate (with SACK F-RTO, other than head 2062 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS 2063 * bits are handled if the Loss state is really to be entered (in 2064 * tcp_enter_frto_loss). 2065 * 2066 * Do like tcp_enter_loss() would; when RTO expires the second time it 2067 * does: 2068 * "Reduce ssthresh if it has not yet been made inside this window." 2069 */ 2070 void tcp_enter_frto(struct sock *sk) 2071 { 2072 const struct inet_connection_sock *icsk = inet_csk(sk); 2073 struct tcp_sock *tp = tcp_sk(sk); 2074 struct sk_buff *skb; 2075 2076 if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) || 2077 tp->snd_una == tp->high_seq || 2078 ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) && 2079 !icsk->icsk_retransmits)) { 2080 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2081 /* Our state is too optimistic in ssthresh() call because cwnd 2082 * is not reduced until tcp_enter_frto_loss() when previous F-RTO 2083 * recovery has not yet completed. Pattern would be this: RTO, 2084 * Cumulative ACK, RTO (2xRTO for the same segment does not end 2085 * up here twice). 2086 * RFC4138 should be more specific on what to do, even though 2087 * RTO is quite unlikely to occur after the first Cumulative ACK 2088 * due to back-off and complexity of triggering events ... 2089 */ 2090 if (tp->frto_counter) { 2091 u32 stored_cwnd; 2092 stored_cwnd = tp->snd_cwnd; 2093 tp->snd_cwnd = 2; 2094 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 2095 tp->snd_cwnd = stored_cwnd; 2096 } else { 2097 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 2098 } 2099 /* ... in theory, cong.control module could do "any tricks" in 2100 * ssthresh(), which means that ca_state, lost bits and lost_out 2101 * counter would have to be faked before the call occurs. We 2102 * consider that too expensive, unlikely and hacky, so modules 2103 * using these in ssthresh() must deal these incompatibility 2104 * issues if they receives CA_EVENT_FRTO and frto_counter != 0 2105 */ 2106 tcp_ca_event(sk, CA_EVENT_FRTO); 2107 } 2108 2109 tp->undo_marker = tp->snd_una; 2110 tp->undo_retrans = 0; 2111 2112 skb = tcp_write_queue_head(sk); 2113 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) 2114 tp->undo_marker = 0; 2115 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 2116 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 2117 tp->retrans_out -= tcp_skb_pcount(skb); 2118 } 2119 tcp_verify_left_out(tp); 2120 2121 /* Too bad if TCP was application limited */ 2122 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1); 2123 2124 /* Earlier loss recovery underway (see RFC4138; Appendix B). 2125 * The last condition is necessary at least in tp->frto_counter case. 2126 */ 2127 if (tcp_is_sackfrto(tp) && (tp->frto_counter || 2128 ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) && 2129 after(tp->high_seq, tp->snd_una)) { 2130 tp->frto_highmark = tp->high_seq; 2131 } else { 2132 tp->frto_highmark = tp->snd_nxt; 2133 } 2134 tcp_set_ca_state(sk, TCP_CA_Disorder); 2135 tp->high_seq = tp->snd_nxt; 2136 tp->frto_counter = 1; 2137 } 2138 2139 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO, 2140 * which indicates that we should follow the traditional RTO recovery, 2141 * i.e. mark everything lost and do go-back-N retransmission. 2142 */ 2143 static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag) 2144 { 2145 struct tcp_sock *tp = tcp_sk(sk); 2146 struct sk_buff *skb; 2147 2148 tp->lost_out = 0; 2149 tp->retrans_out = 0; 2150 if (tcp_is_reno(tp)) 2151 tcp_reset_reno_sack(tp); 2152 2153 tcp_for_write_queue(skb, sk) { 2154 if (skb == tcp_send_head(sk)) 2155 break; 2156 2157 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; 2158 /* 2159 * Count the retransmission made on RTO correctly (only when 2160 * waiting for the first ACK and did not get it)... 2161 */ 2162 if ((tp->frto_counter == 1) && !(flag & FLAG_DATA_ACKED)) { 2163 /* For some reason this R-bit might get cleared? */ 2164 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 2165 tp->retrans_out += tcp_skb_pcount(skb); 2166 /* ...enter this if branch just for the first segment */ 2167 flag |= FLAG_DATA_ACKED; 2168 } else { 2169 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) 2170 tp->undo_marker = 0; 2171 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 2172 } 2173 2174 /* Marking forward transmissions that were made after RTO lost 2175 * can cause unnecessary retransmissions in some scenarios, 2176 * SACK blocks will mitigate that in some but not in all cases. 2177 * We used to not mark them but it was causing break-ups with 2178 * receivers that do only in-order receival. 2179 * 2180 * TODO: we could detect presence of such receiver and select 2181 * different behavior per flow. 2182 */ 2183 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) { 2184 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 2185 tp->lost_out += tcp_skb_pcount(skb); 2186 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq; 2187 } 2188 } 2189 tcp_verify_left_out(tp); 2190 2191 tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments; 2192 tp->snd_cwnd_cnt = 0; 2193 tp->snd_cwnd_stamp = tcp_time_stamp; 2194 tp->frto_counter = 0; 2195 tp->bytes_acked = 0; 2196 2197 tp->reordering = min_t(unsigned int, tp->reordering, 2198 sysctl_tcp_reordering); 2199 tcp_set_ca_state(sk, TCP_CA_Loss); 2200 tp->high_seq = tp->snd_nxt; 2201 TCP_ECN_queue_cwr(tp); 2202 2203 tcp_clear_all_retrans_hints(tp); 2204 } 2205 2206 static void tcp_clear_retrans_partial(struct tcp_sock *tp) 2207 { 2208 tp->retrans_out = 0; 2209 tp->lost_out = 0; 2210 2211 tp->undo_marker = 0; 2212 tp->undo_retrans = 0; 2213 } 2214 2215 void tcp_clear_retrans(struct tcp_sock *tp) 2216 { 2217 tcp_clear_retrans_partial(tp); 2218 2219 tp->fackets_out = 0; 2220 tp->sacked_out = 0; 2221 } 2222 2223 /* Enter Loss state. If "how" is not zero, forget all SACK information 2224 * and reset tags completely, otherwise preserve SACKs. If receiver 2225 * dropped its ofo queue, we will know this due to reneging detection. 2226 */ 2227 void tcp_enter_loss(struct sock *sk, int how) 2228 { 2229 const struct inet_connection_sock *icsk = inet_csk(sk); 2230 struct tcp_sock *tp = tcp_sk(sk); 2231 struct sk_buff *skb; 2232 2233 /* Reduce ssthresh if it has not yet been made inside this window. */ 2234 if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq || 2235 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) { 2236 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2237 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 2238 tcp_ca_event(sk, CA_EVENT_LOSS); 2239 } 2240 tp->snd_cwnd = 1; 2241 tp->snd_cwnd_cnt = 0; 2242 tp->snd_cwnd_stamp = tcp_time_stamp; 2243 2244 tp->bytes_acked = 0; 2245 tcp_clear_retrans_partial(tp); 2246 2247 if (tcp_is_reno(tp)) 2248 tcp_reset_reno_sack(tp); 2249 2250 if (!how) { 2251 /* Push undo marker, if it was plain RTO and nothing 2252 * was retransmitted. */ 2253 tp->undo_marker = tp->snd_una; 2254 } else { 2255 tp->sacked_out = 0; 2256 tp->fackets_out = 0; 2257 } 2258 tcp_clear_all_retrans_hints(tp); 2259 2260 tcp_for_write_queue(skb, sk) { 2261 if (skb == tcp_send_head(sk)) 2262 break; 2263 2264 if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) 2265 tp->undo_marker = 0; 2266 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED; 2267 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) { 2268 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED; 2269 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 2270 tp->lost_out += tcp_skb_pcount(skb); 2271 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq; 2272 } 2273 } 2274 tcp_verify_left_out(tp); 2275 2276 tp->reordering = min_t(unsigned int, tp->reordering, 2277 sysctl_tcp_reordering); 2278 tcp_set_ca_state(sk, TCP_CA_Loss); 2279 tp->high_seq = tp->snd_nxt; 2280 TCP_ECN_queue_cwr(tp); 2281 /* Abort F-RTO algorithm if one is in progress */ 2282 tp->frto_counter = 0; 2283 } 2284 2285 /* If ACK arrived pointing to a remembered SACK, it means that our 2286 * remembered SACKs do not reflect real state of receiver i.e. 2287 * receiver _host_ is heavily congested (or buggy). 2288 * 2289 * Do processing similar to RTO timeout. 2290 */ 2291 static int tcp_check_sack_reneging(struct sock *sk, int flag) 2292 { 2293 if (flag & FLAG_SACK_RENEGING) { 2294 struct inet_connection_sock *icsk = inet_csk(sk); 2295 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING); 2296 2297 tcp_enter_loss(sk, 1); 2298 icsk->icsk_retransmits++; 2299 tcp_retransmit_skb(sk, tcp_write_queue_head(sk)); 2300 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 2301 icsk->icsk_rto, TCP_RTO_MAX); 2302 return 1; 2303 } 2304 return 0; 2305 } 2306 2307 static inline int tcp_fackets_out(const struct tcp_sock *tp) 2308 { 2309 return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out; 2310 } 2311 2312 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs 2313 * counter when SACK is enabled (without SACK, sacked_out is used for 2314 * that purpose). 2315 * 2316 * Instead, with FACK TCP uses fackets_out that includes both SACKed 2317 * segments up to the highest received SACK block so far and holes in 2318 * between them. 2319 * 2320 * With reordering, holes may still be in flight, so RFC3517 recovery 2321 * uses pure sacked_out (total number of SACKed segments) even though 2322 * it violates the RFC that uses duplicate ACKs, often these are equal 2323 * but when e.g. out-of-window ACKs or packet duplication occurs, 2324 * they differ. Since neither occurs due to loss, TCP should really 2325 * ignore them. 2326 */ 2327 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp) 2328 { 2329 return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1; 2330 } 2331 2332 static inline int tcp_skb_timedout(const struct sock *sk, 2333 const struct sk_buff *skb) 2334 { 2335 return tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto; 2336 } 2337 2338 static inline int tcp_head_timedout(const struct sock *sk) 2339 { 2340 const struct tcp_sock *tp = tcp_sk(sk); 2341 2342 return tp->packets_out && 2343 tcp_skb_timedout(sk, tcp_write_queue_head(sk)); 2344 } 2345 2346 /* Linux NewReno/SACK/FACK/ECN state machine. 2347 * -------------------------------------- 2348 * 2349 * "Open" Normal state, no dubious events, fast path. 2350 * "Disorder" In all the respects it is "Open", 2351 * but requires a bit more attention. It is entered when 2352 * we see some SACKs or dupacks. It is split of "Open" 2353 * mainly to move some processing from fast path to slow one. 2354 * "CWR" CWND was reduced due to some Congestion Notification event. 2355 * It can be ECN, ICMP source quench, local device congestion. 2356 * "Recovery" CWND was reduced, we are fast-retransmitting. 2357 * "Loss" CWND was reduced due to RTO timeout or SACK reneging. 2358 * 2359 * tcp_fastretrans_alert() is entered: 2360 * - each incoming ACK, if state is not "Open" 2361 * - when arrived ACK is unusual, namely: 2362 * * SACK 2363 * * Duplicate ACK. 2364 * * ECN ECE. 2365 * 2366 * Counting packets in flight is pretty simple. 2367 * 2368 * in_flight = packets_out - left_out + retrans_out 2369 * 2370 * packets_out is SND.NXT-SND.UNA counted in packets. 2371 * 2372 * retrans_out is number of retransmitted segments. 2373 * 2374 * left_out is number of segments left network, but not ACKed yet. 2375 * 2376 * left_out = sacked_out + lost_out 2377 * 2378 * sacked_out: Packets, which arrived to receiver out of order 2379 * and hence not ACKed. With SACKs this number is simply 2380 * amount of SACKed data. Even without SACKs 2381 * it is easy to give pretty reliable estimate of this number, 2382 * counting duplicate ACKs. 2383 * 2384 * lost_out: Packets lost by network. TCP has no explicit 2385 * "loss notification" feedback from network (for now). 2386 * It means that this number can be only _guessed_. 2387 * Actually, it is the heuristics to predict lossage that 2388 * distinguishes different algorithms. 2389 * 2390 * F.e. after RTO, when all the queue is considered as lost, 2391 * lost_out = packets_out and in_flight = retrans_out. 2392 * 2393 * Essentially, we have now two algorithms counting 2394 * lost packets. 2395 * 2396 * FACK: It is the simplest heuristics. As soon as we decided 2397 * that something is lost, we decide that _all_ not SACKed 2398 * packets until the most forward SACK are lost. I.e. 2399 * lost_out = fackets_out - sacked_out and left_out = fackets_out. 2400 * It is absolutely correct estimate, if network does not reorder 2401 * packets. And it loses any connection to reality when reordering 2402 * takes place. We use FACK by default until reordering 2403 * is suspected on the path to this destination. 2404 * 2405 * NewReno: when Recovery is entered, we assume that one segment 2406 * is lost (classic Reno). While we are in Recovery and 2407 * a partial ACK arrives, we assume that one more packet 2408 * is lost (NewReno). This heuristics are the same in NewReno 2409 * and SACK. 2410 * 2411 * Imagine, that's all! Forget about all this shamanism about CWND inflation 2412 * deflation etc. CWND is real congestion window, never inflated, changes 2413 * only according to classic VJ rules. 2414 * 2415 * Really tricky (and requiring careful tuning) part of algorithm 2416 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue(). 2417 * The first determines the moment _when_ we should reduce CWND and, 2418 * hence, slow down forward transmission. In fact, it determines the moment 2419 * when we decide that hole is caused by loss, rather than by a reorder. 2420 * 2421 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill 2422 * holes, caused by lost packets. 2423 * 2424 * And the most logically complicated part of algorithm is undo 2425 * heuristics. We detect false retransmits due to both too early 2426 * fast retransmit (reordering) and underestimated RTO, analyzing 2427 * timestamps and D-SACKs. When we detect that some segments were 2428 * retransmitted by mistake and CWND reduction was wrong, we undo 2429 * window reduction and abort recovery phase. This logic is hidden 2430 * inside several functions named tcp_try_undo_<something>. 2431 */ 2432 2433 /* This function decides, when we should leave Disordered state 2434 * and enter Recovery phase, reducing congestion window. 2435 * 2436 * Main question: may we further continue forward transmission 2437 * with the same cwnd? 2438 */ 2439 static int tcp_time_to_recover(struct sock *sk) 2440 { 2441 struct tcp_sock *tp = tcp_sk(sk); 2442 __u32 packets_out; 2443 2444 /* Do not perform any recovery during F-RTO algorithm */ 2445 if (tp->frto_counter) 2446 return 0; 2447 2448 /* Trick#1: The loss is proven. */ 2449 if (tp->lost_out) 2450 return 1; 2451 2452 /* Not-A-Trick#2 : Classic rule... */ 2453 if (tcp_dupack_heuristics(tp) > tp->reordering) 2454 return 1; 2455 2456 /* Trick#3 : when we use RFC2988 timer restart, fast 2457 * retransmit can be triggered by timeout of queue head. 2458 */ 2459 if (tcp_is_fack(tp) && tcp_head_timedout(sk)) 2460 return 1; 2461 2462 /* Trick#4: It is still not OK... But will it be useful to delay 2463 * recovery more? 2464 */ 2465 packets_out = tp->packets_out; 2466 if (packets_out <= tp->reordering && 2467 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) && 2468 !tcp_may_send_now(sk)) { 2469 /* We have nothing to send. This connection is limited 2470 * either by receiver window or by application. 2471 */ 2472 return 1; 2473 } 2474 2475 /* If a thin stream is detected, retransmit after first 2476 * received dupack. Employ only if SACK is supported in order 2477 * to avoid possible corner-case series of spurious retransmissions 2478 * Use only if there are no unsent data. 2479 */ 2480 if ((tp->thin_dupack || sysctl_tcp_thin_dupack) && 2481 tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 && 2482 tcp_is_sack(tp) && !tcp_send_head(sk)) 2483 return 1; 2484 2485 return 0; 2486 } 2487 2488 /* New heuristics: it is possible only after we switched to restart timer 2489 * each time when something is ACKed. Hence, we can detect timed out packets 2490 * during fast retransmit without falling to slow start. 2491 * 2492 * Usefulness of this as is very questionable, since we should know which of 2493 * the segments is the next to timeout which is relatively expensive to find 2494 * in general case unless we add some data structure just for that. The 2495 * current approach certainly won't find the right one too often and when it 2496 * finally does find _something_ it usually marks large part of the window 2497 * right away (because a retransmission with a larger timestamp blocks the 2498 * loop from advancing). -ij 2499 */ 2500 static void tcp_timeout_skbs(struct sock *sk) 2501 { 2502 struct tcp_sock *tp = tcp_sk(sk); 2503 struct sk_buff *skb; 2504 2505 if (!tcp_is_fack(tp) || !tcp_head_timedout(sk)) 2506 return; 2507 2508 skb = tp->scoreboard_skb_hint; 2509 if (tp->scoreboard_skb_hint == NULL) 2510 skb = tcp_write_queue_head(sk); 2511 2512 tcp_for_write_queue_from(skb, sk) { 2513 if (skb == tcp_send_head(sk)) 2514 break; 2515 if (!tcp_skb_timedout(sk, skb)) 2516 break; 2517 2518 tcp_skb_mark_lost(tp, skb); 2519 } 2520 2521 tp->scoreboard_skb_hint = skb; 2522 2523 tcp_verify_left_out(tp); 2524 } 2525 2526 /* Detect loss in event "A" above by marking head of queue up as lost. 2527 * For FACK or non-SACK(Reno) senders, the first "packets" number of segments 2528 * are considered lost. For RFC3517 SACK, a segment is considered lost if it 2529 * has at least tp->reordering SACKed seqments above it; "packets" refers to 2530 * the maximum SACKed segments to pass before reaching this limit. 2531 */ 2532 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head) 2533 { 2534 struct tcp_sock *tp = tcp_sk(sk); 2535 struct sk_buff *skb; 2536 int cnt, oldcnt; 2537 int err; 2538 unsigned int mss; 2539 /* Use SACK to deduce losses of new sequences sent during recovery */ 2540 const u32 loss_high = tcp_is_sack(tp) ? tp->snd_nxt : tp->high_seq; 2541 2542 WARN_ON(packets > tp->packets_out); 2543 if (tp->lost_skb_hint) { 2544 skb = tp->lost_skb_hint; 2545 cnt = tp->lost_cnt_hint; 2546 /* Head already handled? */ 2547 if (mark_head && skb != tcp_write_queue_head(sk)) 2548 return; 2549 } else { 2550 skb = tcp_write_queue_head(sk); 2551 cnt = 0; 2552 } 2553 2554 tcp_for_write_queue_from(skb, sk) { 2555 if (skb == tcp_send_head(sk)) 2556 break; 2557 /* TODO: do this better */ 2558 /* this is not the most efficient way to do this... */ 2559 tp->lost_skb_hint = skb; 2560 tp->lost_cnt_hint = cnt; 2561 2562 if (after(TCP_SKB_CB(skb)->end_seq, loss_high)) 2563 break; 2564 2565 oldcnt = cnt; 2566 if (tcp_is_fack(tp) || tcp_is_reno(tp) || 2567 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) 2568 cnt += tcp_skb_pcount(skb); 2569 2570 if (cnt > packets) { 2571 if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) || 2572 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) || 2573 (oldcnt >= packets)) 2574 break; 2575 2576 mss = skb_shinfo(skb)->gso_size; 2577 err = tcp_fragment(sk, skb, (packets - oldcnt) * mss, mss); 2578 if (err < 0) 2579 break; 2580 cnt = packets; 2581 } 2582 2583 tcp_skb_mark_lost(tp, skb); 2584 2585 if (mark_head) 2586 break; 2587 } 2588 tcp_verify_left_out(tp); 2589 } 2590 2591 /* Account newly detected lost packet(s) */ 2592 2593 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit) 2594 { 2595 struct tcp_sock *tp = tcp_sk(sk); 2596 2597 if (tcp_is_reno(tp)) { 2598 tcp_mark_head_lost(sk, 1, 1); 2599 } else if (tcp_is_fack(tp)) { 2600 int lost = tp->fackets_out - tp->reordering; 2601 if (lost <= 0) 2602 lost = 1; 2603 tcp_mark_head_lost(sk, lost, 0); 2604 } else { 2605 int sacked_upto = tp->sacked_out - tp->reordering; 2606 if (sacked_upto >= 0) 2607 tcp_mark_head_lost(sk, sacked_upto, 0); 2608 else if (fast_rexmit) 2609 tcp_mark_head_lost(sk, 1, 1); 2610 } 2611 2612 tcp_timeout_skbs(sk); 2613 } 2614 2615 /* CWND moderation, preventing bursts due to too big ACKs 2616 * in dubious situations. 2617 */ 2618 static inline void tcp_moderate_cwnd(struct tcp_sock *tp) 2619 { 2620 tp->snd_cwnd = min(tp->snd_cwnd, 2621 tcp_packets_in_flight(tp) + tcp_max_burst(tp)); 2622 tp->snd_cwnd_stamp = tcp_time_stamp; 2623 } 2624 2625 /* Lower bound on congestion window is slow start threshold 2626 * unless congestion avoidance choice decides to overide it. 2627 */ 2628 static inline u32 tcp_cwnd_min(const struct sock *sk) 2629 { 2630 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 2631 2632 return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh; 2633 } 2634 2635 /* Decrease cwnd each second ack. */ 2636 static void tcp_cwnd_down(struct sock *sk, int flag) 2637 { 2638 struct tcp_sock *tp = tcp_sk(sk); 2639 int decr = tp->snd_cwnd_cnt + 1; 2640 2641 if ((flag & (FLAG_ANY_PROGRESS | FLAG_DSACKING_ACK)) || 2642 (tcp_is_reno(tp) && !(flag & FLAG_NOT_DUP))) { 2643 tp->snd_cwnd_cnt = decr & 1; 2644 decr >>= 1; 2645 2646 if (decr && tp->snd_cwnd > tcp_cwnd_min(sk)) 2647 tp->snd_cwnd -= decr; 2648 2649 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1); 2650 tp->snd_cwnd_stamp = tcp_time_stamp; 2651 } 2652 } 2653 2654 /* Nothing was retransmitted or returned timestamp is less 2655 * than timestamp of the first retransmission. 2656 */ 2657 static inline int tcp_packet_delayed(const struct tcp_sock *tp) 2658 { 2659 return !tp->retrans_stamp || 2660 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 2661 before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp)); 2662 } 2663 2664 /* Undo procedures. */ 2665 2666 #if FASTRETRANS_DEBUG > 1 2667 static void DBGUNDO(struct sock *sk, const char *msg) 2668 { 2669 struct tcp_sock *tp = tcp_sk(sk); 2670 struct inet_sock *inet = inet_sk(sk); 2671 2672 if (sk->sk_family == AF_INET) { 2673 printk(KERN_DEBUG "Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n", 2674 msg, 2675 &inet->inet_daddr, ntohs(inet->inet_dport), 2676 tp->snd_cwnd, tcp_left_out(tp), 2677 tp->snd_ssthresh, tp->prior_ssthresh, 2678 tp->packets_out); 2679 } 2680 #if IS_ENABLED(CONFIG_IPV6) 2681 else if (sk->sk_family == AF_INET6) { 2682 struct ipv6_pinfo *np = inet6_sk(sk); 2683 printk(KERN_DEBUG "Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n", 2684 msg, 2685 &np->daddr, ntohs(inet->inet_dport), 2686 tp->snd_cwnd, tcp_left_out(tp), 2687 tp->snd_ssthresh, tp->prior_ssthresh, 2688 tp->packets_out); 2689 } 2690 #endif 2691 } 2692 #else 2693 #define DBGUNDO(x...) do { } while (0) 2694 #endif 2695 2696 static void tcp_undo_cwr(struct sock *sk, const bool undo_ssthresh) 2697 { 2698 struct tcp_sock *tp = tcp_sk(sk); 2699 2700 if (tp->prior_ssthresh) { 2701 const struct inet_connection_sock *icsk = inet_csk(sk); 2702 2703 if (icsk->icsk_ca_ops->undo_cwnd) 2704 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk); 2705 else 2706 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1); 2707 2708 if (undo_ssthresh && tp->prior_ssthresh > tp->snd_ssthresh) { 2709 tp->snd_ssthresh = tp->prior_ssthresh; 2710 TCP_ECN_withdraw_cwr(tp); 2711 } 2712 } else { 2713 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh); 2714 } 2715 tp->snd_cwnd_stamp = tcp_time_stamp; 2716 } 2717 2718 static inline int tcp_may_undo(const struct tcp_sock *tp) 2719 { 2720 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp)); 2721 } 2722 2723 /* People celebrate: "We love our President!" */ 2724 static int tcp_try_undo_recovery(struct sock *sk) 2725 { 2726 struct tcp_sock *tp = tcp_sk(sk); 2727 2728 if (tcp_may_undo(tp)) { 2729 int mib_idx; 2730 2731 /* Happy end! We did not retransmit anything 2732 * or our original transmission succeeded. 2733 */ 2734 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans"); 2735 tcp_undo_cwr(sk, true); 2736 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss) 2737 mib_idx = LINUX_MIB_TCPLOSSUNDO; 2738 else 2739 mib_idx = LINUX_MIB_TCPFULLUNDO; 2740 2741 NET_INC_STATS_BH(sock_net(sk), mib_idx); 2742 tp->undo_marker = 0; 2743 } 2744 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) { 2745 /* Hold old state until something *above* high_seq 2746 * is ACKed. For Reno it is MUST to prevent false 2747 * fast retransmits (RFC2582). SACK TCP is safe. */ 2748 tcp_moderate_cwnd(tp); 2749 return 1; 2750 } 2751 tcp_set_ca_state(sk, TCP_CA_Open); 2752 return 0; 2753 } 2754 2755 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */ 2756 static void tcp_try_undo_dsack(struct sock *sk) 2757 { 2758 struct tcp_sock *tp = tcp_sk(sk); 2759 2760 if (tp->undo_marker && !tp->undo_retrans) { 2761 DBGUNDO(sk, "D-SACK"); 2762 tcp_undo_cwr(sk, true); 2763 tp->undo_marker = 0; 2764 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO); 2765 } 2766 } 2767 2768 /* We can clear retrans_stamp when there are no retransmissions in the 2769 * window. It would seem that it is trivially available for us in 2770 * tp->retrans_out, however, that kind of assumptions doesn't consider 2771 * what will happen if errors occur when sending retransmission for the 2772 * second time. ...It could the that such segment has only 2773 * TCPCB_EVER_RETRANS set at the present time. It seems that checking 2774 * the head skb is enough except for some reneging corner cases that 2775 * are not worth the effort. 2776 * 2777 * Main reason for all this complexity is the fact that connection dying 2778 * time now depends on the validity of the retrans_stamp, in particular, 2779 * that successive retransmissions of a segment must not advance 2780 * retrans_stamp under any conditions. 2781 */ 2782 static int tcp_any_retrans_done(const struct sock *sk) 2783 { 2784 const struct tcp_sock *tp = tcp_sk(sk); 2785 struct sk_buff *skb; 2786 2787 if (tp->retrans_out) 2788 return 1; 2789 2790 skb = tcp_write_queue_head(sk); 2791 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS)) 2792 return 1; 2793 2794 return 0; 2795 } 2796 2797 /* Undo during fast recovery after partial ACK. */ 2798 2799 static int tcp_try_undo_partial(struct sock *sk, int acked) 2800 { 2801 struct tcp_sock *tp = tcp_sk(sk); 2802 /* Partial ACK arrived. Force Hoe's retransmit. */ 2803 int failed = tcp_is_reno(tp) || (tcp_fackets_out(tp) > tp->reordering); 2804 2805 if (tcp_may_undo(tp)) { 2806 /* Plain luck! Hole if filled with delayed 2807 * packet, rather than with a retransmit. 2808 */ 2809 if (!tcp_any_retrans_done(sk)) 2810 tp->retrans_stamp = 0; 2811 2812 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1); 2813 2814 DBGUNDO(sk, "Hoe"); 2815 tcp_undo_cwr(sk, false); 2816 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO); 2817 2818 /* So... Do not make Hoe's retransmit yet. 2819 * If the first packet was delayed, the rest 2820 * ones are most probably delayed as well. 2821 */ 2822 failed = 0; 2823 } 2824 return failed; 2825 } 2826 2827 /* Undo during loss recovery after partial ACK. */ 2828 static int tcp_try_undo_loss(struct sock *sk) 2829 { 2830 struct tcp_sock *tp = tcp_sk(sk); 2831 2832 if (tcp_may_undo(tp)) { 2833 struct sk_buff *skb; 2834 tcp_for_write_queue(skb, sk) { 2835 if (skb == tcp_send_head(sk)) 2836 break; 2837 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; 2838 } 2839 2840 tcp_clear_all_retrans_hints(tp); 2841 2842 DBGUNDO(sk, "partial loss"); 2843 tp->lost_out = 0; 2844 tcp_undo_cwr(sk, true); 2845 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO); 2846 inet_csk(sk)->icsk_retransmits = 0; 2847 tp->undo_marker = 0; 2848 if (tcp_is_sack(tp)) 2849 tcp_set_ca_state(sk, TCP_CA_Open); 2850 return 1; 2851 } 2852 return 0; 2853 } 2854 2855 static inline void tcp_complete_cwr(struct sock *sk) 2856 { 2857 struct tcp_sock *tp = tcp_sk(sk); 2858 2859 /* Do not moderate cwnd if it's already undone in cwr or recovery. */ 2860 if (tp->undo_marker) { 2861 if (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR) 2862 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh); 2863 else /* PRR */ 2864 tp->snd_cwnd = tp->snd_ssthresh; 2865 tp->snd_cwnd_stamp = tcp_time_stamp; 2866 } 2867 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR); 2868 } 2869 2870 static void tcp_try_keep_open(struct sock *sk) 2871 { 2872 struct tcp_sock *tp = tcp_sk(sk); 2873 int state = TCP_CA_Open; 2874 2875 if (tcp_left_out(tp) || tcp_any_retrans_done(sk)) 2876 state = TCP_CA_Disorder; 2877 2878 if (inet_csk(sk)->icsk_ca_state != state) { 2879 tcp_set_ca_state(sk, state); 2880 tp->high_seq = tp->snd_nxt; 2881 } 2882 } 2883 2884 static void tcp_try_to_open(struct sock *sk, int flag) 2885 { 2886 struct tcp_sock *tp = tcp_sk(sk); 2887 2888 tcp_verify_left_out(tp); 2889 2890 if (!tp->frto_counter && !tcp_any_retrans_done(sk)) 2891 tp->retrans_stamp = 0; 2892 2893 if (flag & FLAG_ECE) 2894 tcp_enter_cwr(sk, 1); 2895 2896 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) { 2897 tcp_try_keep_open(sk); 2898 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Open) 2899 tcp_moderate_cwnd(tp); 2900 } else { 2901 tcp_cwnd_down(sk, flag); 2902 } 2903 } 2904 2905 static void tcp_mtup_probe_failed(struct sock *sk) 2906 { 2907 struct inet_connection_sock *icsk = inet_csk(sk); 2908 2909 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1; 2910 icsk->icsk_mtup.probe_size = 0; 2911 } 2912 2913 static void tcp_mtup_probe_success(struct sock *sk) 2914 { 2915 struct tcp_sock *tp = tcp_sk(sk); 2916 struct inet_connection_sock *icsk = inet_csk(sk); 2917 2918 /* FIXME: breaks with very large cwnd */ 2919 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2920 tp->snd_cwnd = tp->snd_cwnd * 2921 tcp_mss_to_mtu(sk, tp->mss_cache) / 2922 icsk->icsk_mtup.probe_size; 2923 tp->snd_cwnd_cnt = 0; 2924 tp->snd_cwnd_stamp = tcp_time_stamp; 2925 tp->snd_ssthresh = tcp_current_ssthresh(sk); 2926 2927 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size; 2928 icsk->icsk_mtup.probe_size = 0; 2929 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 2930 } 2931 2932 /* Do a simple retransmit without using the backoff mechanisms in 2933 * tcp_timer. This is used for path mtu discovery. 2934 * The socket is already locked here. 2935 */ 2936 void tcp_simple_retransmit(struct sock *sk) 2937 { 2938 const struct inet_connection_sock *icsk = inet_csk(sk); 2939 struct tcp_sock *tp = tcp_sk(sk); 2940 struct sk_buff *skb; 2941 unsigned int mss = tcp_current_mss(sk); 2942 u32 prior_lost = tp->lost_out; 2943 2944 tcp_for_write_queue(skb, sk) { 2945 if (skb == tcp_send_head(sk)) 2946 break; 2947 if (tcp_skb_seglen(skb) > mss && 2948 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) { 2949 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 2950 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 2951 tp->retrans_out -= tcp_skb_pcount(skb); 2952 } 2953 tcp_skb_mark_lost_uncond_verify(tp, skb); 2954 } 2955 } 2956 2957 tcp_clear_retrans_hints_partial(tp); 2958 2959 if (prior_lost == tp->lost_out) 2960 return; 2961 2962 if (tcp_is_reno(tp)) 2963 tcp_limit_reno_sacked(tp); 2964 2965 tcp_verify_left_out(tp); 2966 2967 /* Don't muck with the congestion window here. 2968 * Reason is that we do not increase amount of _data_ 2969 * in network, but units changed and effective 2970 * cwnd/ssthresh really reduced now. 2971 */ 2972 if (icsk->icsk_ca_state != TCP_CA_Loss) { 2973 tp->high_seq = tp->snd_nxt; 2974 tp->snd_ssthresh = tcp_current_ssthresh(sk); 2975 tp->prior_ssthresh = 0; 2976 tp->undo_marker = 0; 2977 tcp_set_ca_state(sk, TCP_CA_Loss); 2978 } 2979 tcp_xmit_retransmit_queue(sk); 2980 } 2981 EXPORT_SYMBOL(tcp_simple_retransmit); 2982 2983 /* This function implements the PRR algorithm, specifcally the PRR-SSRB 2984 * (proportional rate reduction with slow start reduction bound) as described in 2985 * http://www.ietf.org/id/draft-mathis-tcpm-proportional-rate-reduction-01.txt. 2986 * It computes the number of packets to send (sndcnt) based on packets newly 2987 * delivered: 2988 * 1) If the packets in flight is larger than ssthresh, PRR spreads the 2989 * cwnd reductions across a full RTT. 2990 * 2) If packets in flight is lower than ssthresh (such as due to excess 2991 * losses and/or application stalls), do not perform any further cwnd 2992 * reductions, but instead slow start up to ssthresh. 2993 */ 2994 static void tcp_update_cwnd_in_recovery(struct sock *sk, int newly_acked_sacked, 2995 int fast_rexmit, int flag) 2996 { 2997 struct tcp_sock *tp = tcp_sk(sk); 2998 int sndcnt = 0; 2999 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp); 3000 3001 if (tcp_packets_in_flight(tp) > tp->snd_ssthresh) { 3002 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered + 3003 tp->prior_cwnd - 1; 3004 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out; 3005 } else { 3006 sndcnt = min_t(int, delta, 3007 max_t(int, tp->prr_delivered - tp->prr_out, 3008 newly_acked_sacked) + 1); 3009 } 3010 3011 sndcnt = max(sndcnt, (fast_rexmit ? 1 : 0)); 3012 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt; 3013 } 3014 3015 /* Process an event, which can update packets-in-flight not trivially. 3016 * Main goal of this function is to calculate new estimate for left_out, 3017 * taking into account both packets sitting in receiver's buffer and 3018 * packets lost by network. 3019 * 3020 * Besides that it does CWND reduction, when packet loss is detected 3021 * and changes state of machine. 3022 * 3023 * It does _not_ decide what to send, it is made in function 3024 * tcp_xmit_retransmit_queue(). 3025 */ 3026 static void tcp_fastretrans_alert(struct sock *sk, int pkts_acked, 3027 int newly_acked_sacked, bool is_dupack, 3028 int flag) 3029 { 3030 struct inet_connection_sock *icsk = inet_csk(sk); 3031 struct tcp_sock *tp = tcp_sk(sk); 3032 int do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) && 3033 (tcp_fackets_out(tp) > tp->reordering)); 3034 int fast_rexmit = 0, mib_idx; 3035 3036 if (WARN_ON(!tp->packets_out && tp->sacked_out)) 3037 tp->sacked_out = 0; 3038 if (WARN_ON(!tp->sacked_out && tp->fackets_out)) 3039 tp->fackets_out = 0; 3040 3041 /* Now state machine starts. 3042 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */ 3043 if (flag & FLAG_ECE) 3044 tp->prior_ssthresh = 0; 3045 3046 /* B. In all the states check for reneging SACKs. */ 3047 if (tcp_check_sack_reneging(sk, flag)) 3048 return; 3049 3050 /* C. Check consistency of the current state. */ 3051 tcp_verify_left_out(tp); 3052 3053 /* D. Check state exit conditions. State can be terminated 3054 * when high_seq is ACKed. */ 3055 if (icsk->icsk_ca_state == TCP_CA_Open) { 3056 WARN_ON(tp->retrans_out != 0); 3057 tp->retrans_stamp = 0; 3058 } else if (!before(tp->snd_una, tp->high_seq)) { 3059 switch (icsk->icsk_ca_state) { 3060 case TCP_CA_Loss: 3061 icsk->icsk_retransmits = 0; 3062 if (tcp_try_undo_recovery(sk)) 3063 return; 3064 break; 3065 3066 case TCP_CA_CWR: 3067 /* CWR is to be held something *above* high_seq 3068 * is ACKed for CWR bit to reach receiver. */ 3069 if (tp->snd_una != tp->high_seq) { 3070 tcp_complete_cwr(sk); 3071 tcp_set_ca_state(sk, TCP_CA_Open); 3072 } 3073 break; 3074 3075 case TCP_CA_Recovery: 3076 if (tcp_is_reno(tp)) 3077 tcp_reset_reno_sack(tp); 3078 if (tcp_try_undo_recovery(sk)) 3079 return; 3080 tcp_complete_cwr(sk); 3081 break; 3082 } 3083 } 3084 3085 /* E. Process state. */ 3086 switch (icsk->icsk_ca_state) { 3087 case TCP_CA_Recovery: 3088 if (!(flag & FLAG_SND_UNA_ADVANCED)) { 3089 if (tcp_is_reno(tp) && is_dupack) 3090 tcp_add_reno_sack(sk); 3091 } else 3092 do_lost = tcp_try_undo_partial(sk, pkts_acked); 3093 break; 3094 case TCP_CA_Loss: 3095 if (flag & FLAG_DATA_ACKED) 3096 icsk->icsk_retransmits = 0; 3097 if (tcp_is_reno(tp) && flag & FLAG_SND_UNA_ADVANCED) 3098 tcp_reset_reno_sack(tp); 3099 if (!tcp_try_undo_loss(sk)) { 3100 tcp_moderate_cwnd(tp); 3101 tcp_xmit_retransmit_queue(sk); 3102 return; 3103 } 3104 if (icsk->icsk_ca_state != TCP_CA_Open) 3105 return; 3106 /* Loss is undone; fall through to processing in Open state. */ 3107 default: 3108 if (tcp_is_reno(tp)) { 3109 if (flag & FLAG_SND_UNA_ADVANCED) 3110 tcp_reset_reno_sack(tp); 3111 if (is_dupack) 3112 tcp_add_reno_sack(sk); 3113 } 3114 3115 if (icsk->icsk_ca_state <= TCP_CA_Disorder) 3116 tcp_try_undo_dsack(sk); 3117 3118 if (!tcp_time_to_recover(sk)) { 3119 tcp_try_to_open(sk, flag); 3120 return; 3121 } 3122 3123 /* MTU probe failure: don't reduce cwnd */ 3124 if (icsk->icsk_ca_state < TCP_CA_CWR && 3125 icsk->icsk_mtup.probe_size && 3126 tp->snd_una == tp->mtu_probe.probe_seq_start) { 3127 tcp_mtup_probe_failed(sk); 3128 /* Restores the reduction we did in tcp_mtup_probe() */ 3129 tp->snd_cwnd++; 3130 tcp_simple_retransmit(sk); 3131 return; 3132 } 3133 3134 /* Otherwise enter Recovery state */ 3135 3136 if (tcp_is_reno(tp)) 3137 mib_idx = LINUX_MIB_TCPRENORECOVERY; 3138 else 3139 mib_idx = LINUX_MIB_TCPSACKRECOVERY; 3140 3141 NET_INC_STATS_BH(sock_net(sk), mib_idx); 3142 3143 tp->high_seq = tp->snd_nxt; 3144 tp->prior_ssthresh = 0; 3145 tp->undo_marker = tp->snd_una; 3146 tp->undo_retrans = tp->retrans_out; 3147 3148 if (icsk->icsk_ca_state < TCP_CA_CWR) { 3149 if (!(flag & FLAG_ECE)) 3150 tp->prior_ssthresh = tcp_current_ssthresh(sk); 3151 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 3152 TCP_ECN_queue_cwr(tp); 3153 } 3154 3155 tp->bytes_acked = 0; 3156 tp->snd_cwnd_cnt = 0; 3157 tp->prior_cwnd = tp->snd_cwnd; 3158 tp->prr_delivered = 0; 3159 tp->prr_out = 0; 3160 tcp_set_ca_state(sk, TCP_CA_Recovery); 3161 fast_rexmit = 1; 3162 } 3163 3164 if (do_lost || (tcp_is_fack(tp) && tcp_head_timedout(sk))) 3165 tcp_update_scoreboard(sk, fast_rexmit); 3166 tp->prr_delivered += newly_acked_sacked; 3167 tcp_update_cwnd_in_recovery(sk, newly_acked_sacked, fast_rexmit, flag); 3168 tcp_xmit_retransmit_queue(sk); 3169 } 3170 3171 void tcp_valid_rtt_meas(struct sock *sk, u32 seq_rtt) 3172 { 3173 tcp_rtt_estimator(sk, seq_rtt); 3174 tcp_set_rto(sk); 3175 inet_csk(sk)->icsk_backoff = 0; 3176 } 3177 EXPORT_SYMBOL(tcp_valid_rtt_meas); 3178 3179 /* Read draft-ietf-tcplw-high-performance before mucking 3180 * with this code. (Supersedes RFC1323) 3181 */ 3182 static void tcp_ack_saw_tstamp(struct sock *sk, int flag) 3183 { 3184 /* RTTM Rule: A TSecr value received in a segment is used to 3185 * update the averaged RTT measurement only if the segment 3186 * acknowledges some new data, i.e., only if it advances the 3187 * left edge of the send window. 3188 * 3189 * See draft-ietf-tcplw-high-performance-00, section 3.3. 3190 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru> 3191 * 3192 * Changed: reset backoff as soon as we see the first valid sample. 3193 * If we do not, we get strongly overestimated rto. With timestamps 3194 * samples are accepted even from very old segments: f.e., when rtt=1 3195 * increases to 8, we retransmit 5 times and after 8 seconds delayed 3196 * answer arrives rto becomes 120 seconds! If at least one of segments 3197 * in window is lost... Voila. --ANK (010210) 3198 */ 3199 struct tcp_sock *tp = tcp_sk(sk); 3200 3201 tcp_valid_rtt_meas(sk, tcp_time_stamp - tp->rx_opt.rcv_tsecr); 3202 } 3203 3204 static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag) 3205 { 3206 /* We don't have a timestamp. Can only use 3207 * packets that are not retransmitted to determine 3208 * rtt estimates. Also, we must not reset the 3209 * backoff for rto until we get a non-retransmitted 3210 * packet. This allows us to deal with a situation 3211 * where the network delay has increased suddenly. 3212 * I.e. Karn's algorithm. (SIGCOMM '87, p5.) 3213 */ 3214 3215 if (flag & FLAG_RETRANS_DATA_ACKED) 3216 return; 3217 3218 tcp_valid_rtt_meas(sk, seq_rtt); 3219 } 3220 3221 static inline void tcp_ack_update_rtt(struct sock *sk, const int flag, 3222 const s32 seq_rtt) 3223 { 3224 const struct tcp_sock *tp = tcp_sk(sk); 3225 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */ 3226 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr) 3227 tcp_ack_saw_tstamp(sk, flag); 3228 else if (seq_rtt >= 0) 3229 tcp_ack_no_tstamp(sk, seq_rtt, flag); 3230 } 3231 3232 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight) 3233 { 3234 const struct inet_connection_sock *icsk = inet_csk(sk); 3235 icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight); 3236 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp; 3237 } 3238 3239 /* Restart timer after forward progress on connection. 3240 * RFC2988 recommends to restart timer to now+rto. 3241 */ 3242 static void tcp_rearm_rto(struct sock *sk) 3243 { 3244 const struct tcp_sock *tp = tcp_sk(sk); 3245 3246 if (!tp->packets_out) { 3247 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS); 3248 } else { 3249 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 3250 inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 3251 } 3252 } 3253 3254 /* If we get here, the whole TSO packet has not been acked. */ 3255 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb) 3256 { 3257 struct tcp_sock *tp = tcp_sk(sk); 3258 u32 packets_acked; 3259 3260 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)); 3261 3262 packets_acked = tcp_skb_pcount(skb); 3263 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 3264 return 0; 3265 packets_acked -= tcp_skb_pcount(skb); 3266 3267 if (packets_acked) { 3268 BUG_ON(tcp_skb_pcount(skb) == 0); 3269 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)); 3270 } 3271 3272 return packets_acked; 3273 } 3274 3275 /* Remove acknowledged frames from the retransmission queue. If our packet 3276 * is before the ack sequence we can discard it as it's confirmed to have 3277 * arrived at the other end. 3278 */ 3279 static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets, 3280 u32 prior_snd_una) 3281 { 3282 struct tcp_sock *tp = tcp_sk(sk); 3283 const struct inet_connection_sock *icsk = inet_csk(sk); 3284 struct sk_buff *skb; 3285 u32 now = tcp_time_stamp; 3286 int fully_acked = 1; 3287 int flag = 0; 3288 u32 pkts_acked = 0; 3289 u32 reord = tp->packets_out; 3290 u32 prior_sacked = tp->sacked_out; 3291 s32 seq_rtt = -1; 3292 s32 ca_seq_rtt = -1; 3293 ktime_t last_ackt = net_invalid_timestamp(); 3294 3295 while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) { 3296 struct tcp_skb_cb *scb = TCP_SKB_CB(skb); 3297 u32 acked_pcount; 3298 u8 sacked = scb->sacked; 3299 3300 /* Determine how many packets and what bytes were acked, tso and else */ 3301 if (after(scb->end_seq, tp->snd_una)) { 3302 if (tcp_skb_pcount(skb) == 1 || 3303 !after(tp->snd_una, scb->seq)) 3304 break; 3305 3306 acked_pcount = tcp_tso_acked(sk, skb); 3307 if (!acked_pcount) 3308 break; 3309 3310 fully_acked = 0; 3311 } else { 3312 acked_pcount = tcp_skb_pcount(skb); 3313 } 3314 3315 if (sacked & TCPCB_RETRANS) { 3316 if (sacked & TCPCB_SACKED_RETRANS) 3317 tp->retrans_out -= acked_pcount; 3318 flag |= FLAG_RETRANS_DATA_ACKED; 3319 ca_seq_rtt = -1; 3320 seq_rtt = -1; 3321 if ((flag & FLAG_DATA_ACKED) || (acked_pcount > 1)) 3322 flag |= FLAG_NONHEAD_RETRANS_ACKED; 3323 } else { 3324 ca_seq_rtt = now - scb->when; 3325 last_ackt = skb->tstamp; 3326 if (seq_rtt < 0) { 3327 seq_rtt = ca_seq_rtt; 3328 } 3329 if (!(sacked & TCPCB_SACKED_ACKED)) 3330 reord = min(pkts_acked, reord); 3331 } 3332 3333 if (sacked & TCPCB_SACKED_ACKED) 3334 tp->sacked_out -= acked_pcount; 3335 if (sacked & TCPCB_LOST) 3336 tp->lost_out -= acked_pcount; 3337 3338 tp->packets_out -= acked_pcount; 3339 pkts_acked += acked_pcount; 3340 3341 /* Initial outgoing SYN's get put onto the write_queue 3342 * just like anything else we transmit. It is not 3343 * true data, and if we misinform our callers that 3344 * this ACK acks real data, we will erroneously exit 3345 * connection startup slow start one packet too 3346 * quickly. This is severely frowned upon behavior. 3347 */ 3348 if (!(scb->tcp_flags & TCPHDR_SYN)) { 3349 flag |= FLAG_DATA_ACKED; 3350 } else { 3351 flag |= FLAG_SYN_ACKED; 3352 tp->retrans_stamp = 0; 3353 } 3354 3355 if (!fully_acked) 3356 break; 3357 3358 tcp_unlink_write_queue(skb, sk); 3359 sk_wmem_free_skb(sk, skb); 3360 tp->scoreboard_skb_hint = NULL; 3361 if (skb == tp->retransmit_skb_hint) 3362 tp->retransmit_skb_hint = NULL; 3363 if (skb == tp->lost_skb_hint) 3364 tp->lost_skb_hint = NULL; 3365 } 3366 3367 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una))) 3368 tp->snd_up = tp->snd_una; 3369 3370 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) 3371 flag |= FLAG_SACK_RENEGING; 3372 3373 if (flag & FLAG_ACKED) { 3374 const struct tcp_congestion_ops *ca_ops 3375 = inet_csk(sk)->icsk_ca_ops; 3376 3377 if (unlikely(icsk->icsk_mtup.probe_size && 3378 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) { 3379 tcp_mtup_probe_success(sk); 3380 } 3381 3382 tcp_ack_update_rtt(sk, flag, seq_rtt); 3383 tcp_rearm_rto(sk); 3384 3385 if (tcp_is_reno(tp)) { 3386 tcp_remove_reno_sacks(sk, pkts_acked); 3387 } else { 3388 int delta; 3389 3390 /* Non-retransmitted hole got filled? That's reordering */ 3391 if (reord < prior_fackets) 3392 tcp_update_reordering(sk, tp->fackets_out - reord, 0); 3393 3394 delta = tcp_is_fack(tp) ? pkts_acked : 3395 prior_sacked - tp->sacked_out; 3396 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta); 3397 } 3398 3399 tp->fackets_out -= min(pkts_acked, tp->fackets_out); 3400 3401 if (ca_ops->pkts_acked) { 3402 s32 rtt_us = -1; 3403 3404 /* Is the ACK triggering packet unambiguous? */ 3405 if (!(flag & FLAG_RETRANS_DATA_ACKED)) { 3406 /* High resolution needed and available? */ 3407 if (ca_ops->flags & TCP_CONG_RTT_STAMP && 3408 !ktime_equal(last_ackt, 3409 net_invalid_timestamp())) 3410 rtt_us = ktime_us_delta(ktime_get_real(), 3411 last_ackt); 3412 else if (ca_seq_rtt >= 0) 3413 rtt_us = jiffies_to_usecs(ca_seq_rtt); 3414 } 3415 3416 ca_ops->pkts_acked(sk, pkts_acked, rtt_us); 3417 } 3418 } 3419 3420 #if FASTRETRANS_DEBUG > 0 3421 WARN_ON((int)tp->sacked_out < 0); 3422 WARN_ON((int)tp->lost_out < 0); 3423 WARN_ON((int)tp->retrans_out < 0); 3424 if (!tp->packets_out && tcp_is_sack(tp)) { 3425 icsk = inet_csk(sk); 3426 if (tp->lost_out) { 3427 printk(KERN_DEBUG "Leak l=%u %d\n", 3428 tp->lost_out, icsk->icsk_ca_state); 3429 tp->lost_out = 0; 3430 } 3431 if (tp->sacked_out) { 3432 printk(KERN_DEBUG "Leak s=%u %d\n", 3433 tp->sacked_out, icsk->icsk_ca_state); 3434 tp->sacked_out = 0; 3435 } 3436 if (tp->retrans_out) { 3437 printk(KERN_DEBUG "Leak r=%u %d\n", 3438 tp->retrans_out, icsk->icsk_ca_state); 3439 tp->retrans_out = 0; 3440 } 3441 } 3442 #endif 3443 return flag; 3444 } 3445 3446 static void tcp_ack_probe(struct sock *sk) 3447 { 3448 const struct tcp_sock *tp = tcp_sk(sk); 3449 struct inet_connection_sock *icsk = inet_csk(sk); 3450 3451 /* Was it a usable window open? */ 3452 3453 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) { 3454 icsk->icsk_backoff = 0; 3455 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0); 3456 /* Socket must be waked up by subsequent tcp_data_snd_check(). 3457 * This function is not for random using! 3458 */ 3459 } else { 3460 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 3461 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), 3462 TCP_RTO_MAX); 3463 } 3464 } 3465 3466 static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag) 3467 { 3468 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) || 3469 inet_csk(sk)->icsk_ca_state != TCP_CA_Open; 3470 } 3471 3472 static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag) 3473 { 3474 const struct tcp_sock *tp = tcp_sk(sk); 3475 return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) && 3476 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR)); 3477 } 3478 3479 /* Check that window update is acceptable. 3480 * The function assumes that snd_una<=ack<=snd_next. 3481 */ 3482 static inline int tcp_may_update_window(const struct tcp_sock *tp, 3483 const u32 ack, const u32 ack_seq, 3484 const u32 nwin) 3485 { 3486 return after(ack, tp->snd_una) || 3487 after(ack_seq, tp->snd_wl1) || 3488 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd); 3489 } 3490 3491 /* Update our send window. 3492 * 3493 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2 3494 * and in FreeBSD. NetBSD's one is even worse.) is wrong. 3495 */ 3496 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack, 3497 u32 ack_seq) 3498 { 3499 struct tcp_sock *tp = tcp_sk(sk); 3500 int flag = 0; 3501 u32 nwin = ntohs(tcp_hdr(skb)->window); 3502 3503 if (likely(!tcp_hdr(skb)->syn)) 3504 nwin <<= tp->rx_opt.snd_wscale; 3505 3506 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) { 3507 flag |= FLAG_WIN_UPDATE; 3508 tcp_update_wl(tp, ack_seq); 3509 3510 if (tp->snd_wnd != nwin) { 3511 tp->snd_wnd = nwin; 3512 3513 /* Note, it is the only place, where 3514 * fast path is recovered for sending TCP. 3515 */ 3516 tp->pred_flags = 0; 3517 tcp_fast_path_check(sk); 3518 3519 if (nwin > tp->max_window) { 3520 tp->max_window = nwin; 3521 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie); 3522 } 3523 } 3524 } 3525 3526 tp->snd_una = ack; 3527 3528 return flag; 3529 } 3530 3531 /* A very conservative spurious RTO response algorithm: reduce cwnd and 3532 * continue in congestion avoidance. 3533 */ 3534 static void tcp_conservative_spur_to_response(struct tcp_sock *tp) 3535 { 3536 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh); 3537 tp->snd_cwnd_cnt = 0; 3538 tp->bytes_acked = 0; 3539 TCP_ECN_queue_cwr(tp); 3540 tcp_moderate_cwnd(tp); 3541 } 3542 3543 /* A conservative spurious RTO response algorithm: reduce cwnd using 3544 * rate halving and continue in congestion avoidance. 3545 */ 3546 static void tcp_ratehalving_spur_to_response(struct sock *sk) 3547 { 3548 tcp_enter_cwr(sk, 0); 3549 } 3550 3551 static void tcp_undo_spur_to_response(struct sock *sk, int flag) 3552 { 3553 if (flag & FLAG_ECE) 3554 tcp_ratehalving_spur_to_response(sk); 3555 else 3556 tcp_undo_cwr(sk, true); 3557 } 3558 3559 /* F-RTO spurious RTO detection algorithm (RFC4138) 3560 * 3561 * F-RTO affects during two new ACKs following RTO (well, almost, see inline 3562 * comments). State (ACK number) is kept in frto_counter. When ACK advances 3563 * window (but not to or beyond highest sequence sent before RTO): 3564 * On First ACK, send two new segments out. 3565 * On Second ACK, RTO was likely spurious. Do spurious response (response 3566 * algorithm is not part of the F-RTO detection algorithm 3567 * given in RFC4138 but can be selected separately). 3568 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss 3569 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding 3570 * of Nagle, this is done using frto_counter states 2 and 3, when a new data 3571 * segment of any size sent during F-RTO, state 2 is upgraded to 3. 3572 * 3573 * Rationale: if the RTO was spurious, new ACKs should arrive from the 3574 * original window even after we transmit two new data segments. 3575 * 3576 * SACK version: 3577 * on first step, wait until first cumulative ACK arrives, then move to 3578 * the second step. In second step, the next ACK decides. 3579 * 3580 * F-RTO is implemented (mainly) in four functions: 3581 * - tcp_use_frto() is used to determine if TCP is can use F-RTO 3582 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is 3583 * called when tcp_use_frto() showed green light 3584 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm 3585 * - tcp_enter_frto_loss() is called if there is not enough evidence 3586 * to prove that the RTO is indeed spurious. It transfers the control 3587 * from F-RTO to the conventional RTO recovery 3588 */ 3589 static int tcp_process_frto(struct sock *sk, int flag) 3590 { 3591 struct tcp_sock *tp = tcp_sk(sk); 3592 3593 tcp_verify_left_out(tp); 3594 3595 /* Duplicate the behavior from Loss state (fastretrans_alert) */ 3596 if (flag & FLAG_DATA_ACKED) 3597 inet_csk(sk)->icsk_retransmits = 0; 3598 3599 if ((flag & FLAG_NONHEAD_RETRANS_ACKED) || 3600 ((tp->frto_counter >= 2) && (flag & FLAG_RETRANS_DATA_ACKED))) 3601 tp->undo_marker = 0; 3602 3603 if (!before(tp->snd_una, tp->frto_highmark)) { 3604 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag); 3605 return 1; 3606 } 3607 3608 if (!tcp_is_sackfrto(tp)) { 3609 /* RFC4138 shortcoming in step 2; should also have case c): 3610 * ACK isn't duplicate nor advances window, e.g., opposite dir 3611 * data, winupdate 3612 */ 3613 if (!(flag & FLAG_ANY_PROGRESS) && (flag & FLAG_NOT_DUP)) 3614 return 1; 3615 3616 if (!(flag & FLAG_DATA_ACKED)) { 3617 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3), 3618 flag); 3619 return 1; 3620 } 3621 } else { 3622 if (!(flag & FLAG_DATA_ACKED) && (tp->frto_counter == 1)) { 3623 /* Prevent sending of new data. */ 3624 tp->snd_cwnd = min(tp->snd_cwnd, 3625 tcp_packets_in_flight(tp)); 3626 return 1; 3627 } 3628 3629 if ((tp->frto_counter >= 2) && 3630 (!(flag & FLAG_FORWARD_PROGRESS) || 3631 ((flag & FLAG_DATA_SACKED) && 3632 !(flag & FLAG_ONLY_ORIG_SACKED)))) { 3633 /* RFC4138 shortcoming (see comment above) */ 3634 if (!(flag & FLAG_FORWARD_PROGRESS) && 3635 (flag & FLAG_NOT_DUP)) 3636 return 1; 3637 3638 tcp_enter_frto_loss(sk, 3, flag); 3639 return 1; 3640 } 3641 } 3642 3643 if (tp->frto_counter == 1) { 3644 /* tcp_may_send_now needs to see updated state */ 3645 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2; 3646 tp->frto_counter = 2; 3647 3648 if (!tcp_may_send_now(sk)) 3649 tcp_enter_frto_loss(sk, 2, flag); 3650 3651 return 1; 3652 } else { 3653 switch (sysctl_tcp_frto_response) { 3654 case 2: 3655 tcp_undo_spur_to_response(sk, flag); 3656 break; 3657 case 1: 3658 tcp_conservative_spur_to_response(tp); 3659 break; 3660 default: 3661 tcp_ratehalving_spur_to_response(sk); 3662 break; 3663 } 3664 tp->frto_counter = 0; 3665 tp->undo_marker = 0; 3666 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSPURIOUSRTOS); 3667 } 3668 return 0; 3669 } 3670 3671 /* This routine deals with incoming acks, but not outgoing ones. */ 3672 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag) 3673 { 3674 struct inet_connection_sock *icsk = inet_csk(sk); 3675 struct tcp_sock *tp = tcp_sk(sk); 3676 u32 prior_snd_una = tp->snd_una; 3677 u32 ack_seq = TCP_SKB_CB(skb)->seq; 3678 u32 ack = TCP_SKB_CB(skb)->ack_seq; 3679 bool is_dupack = false; 3680 u32 prior_in_flight; 3681 u32 prior_fackets; 3682 int prior_packets; 3683 int prior_sacked = tp->sacked_out; 3684 int pkts_acked = 0; 3685 int newly_acked_sacked = 0; 3686 int frto_cwnd = 0; 3687 3688 /* If the ack is older than previous acks 3689 * then we can probably ignore it. 3690 */ 3691 if (before(ack, prior_snd_una)) 3692 goto old_ack; 3693 3694 /* If the ack includes data we haven't sent yet, discard 3695 * this segment (RFC793 Section 3.9). 3696 */ 3697 if (after(ack, tp->snd_nxt)) 3698 goto invalid_ack; 3699 3700 if (after(ack, prior_snd_una)) 3701 flag |= FLAG_SND_UNA_ADVANCED; 3702 3703 if (sysctl_tcp_abc) { 3704 if (icsk->icsk_ca_state < TCP_CA_CWR) 3705 tp->bytes_acked += ack - prior_snd_una; 3706 else if (icsk->icsk_ca_state == TCP_CA_Loss) 3707 /* we assume just one segment left network */ 3708 tp->bytes_acked += min(ack - prior_snd_una, 3709 tp->mss_cache); 3710 } 3711 3712 prior_fackets = tp->fackets_out; 3713 prior_in_flight = tcp_packets_in_flight(tp); 3714 3715 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) { 3716 /* Window is constant, pure forward advance. 3717 * No more checks are required. 3718 * Note, we use the fact that SND.UNA>=SND.WL2. 3719 */ 3720 tcp_update_wl(tp, ack_seq); 3721 tp->snd_una = ack; 3722 flag |= FLAG_WIN_UPDATE; 3723 3724 tcp_ca_event(sk, CA_EVENT_FAST_ACK); 3725 3726 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS); 3727 } else { 3728 if (ack_seq != TCP_SKB_CB(skb)->end_seq) 3729 flag |= FLAG_DATA; 3730 else 3731 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS); 3732 3733 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq); 3734 3735 if (TCP_SKB_CB(skb)->sacked) 3736 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una); 3737 3738 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb))) 3739 flag |= FLAG_ECE; 3740 3741 tcp_ca_event(sk, CA_EVENT_SLOW_ACK); 3742 } 3743 3744 /* We passed data and got it acked, remove any soft error 3745 * log. Something worked... 3746 */ 3747 sk->sk_err_soft = 0; 3748 icsk->icsk_probes_out = 0; 3749 tp->rcv_tstamp = tcp_time_stamp; 3750 prior_packets = tp->packets_out; 3751 if (!prior_packets) 3752 goto no_queue; 3753 3754 /* See if we can take anything off of the retransmit queue. */ 3755 flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una); 3756 3757 pkts_acked = prior_packets - tp->packets_out; 3758 newly_acked_sacked = (prior_packets - prior_sacked) - 3759 (tp->packets_out - tp->sacked_out); 3760 3761 if (tp->frto_counter) 3762 frto_cwnd = tcp_process_frto(sk, flag); 3763 /* Guarantee sacktag reordering detection against wrap-arounds */ 3764 if (before(tp->frto_highmark, tp->snd_una)) 3765 tp->frto_highmark = 0; 3766 3767 if (tcp_ack_is_dubious(sk, flag)) { 3768 /* Advance CWND, if state allows this. */ 3769 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd && 3770 tcp_may_raise_cwnd(sk, flag)) 3771 tcp_cong_avoid(sk, ack, prior_in_flight); 3772 is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP)); 3773 tcp_fastretrans_alert(sk, pkts_acked, newly_acked_sacked, 3774 is_dupack, flag); 3775 } else { 3776 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd) 3777 tcp_cong_avoid(sk, ack, prior_in_flight); 3778 } 3779 3780 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) 3781 dst_confirm(__sk_dst_get(sk)); 3782 3783 return 1; 3784 3785 no_queue: 3786 /* If data was DSACKed, see if we can undo a cwnd reduction. */ 3787 if (flag & FLAG_DSACKING_ACK) 3788 tcp_fastretrans_alert(sk, pkts_acked, newly_acked_sacked, 3789 is_dupack, flag); 3790 /* If this ack opens up a zero window, clear backoff. It was 3791 * being used to time the probes, and is probably far higher than 3792 * it needs to be for normal retransmission. 3793 */ 3794 if (tcp_send_head(sk)) 3795 tcp_ack_probe(sk); 3796 return 1; 3797 3798 invalid_ack: 3799 SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt); 3800 return -1; 3801 3802 old_ack: 3803 /* If data was SACKed, tag it and see if we should send more data. 3804 * If data was DSACKed, see if we can undo a cwnd reduction. 3805 */ 3806 if (TCP_SKB_CB(skb)->sacked) { 3807 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una); 3808 newly_acked_sacked = tp->sacked_out - prior_sacked; 3809 tcp_fastretrans_alert(sk, pkts_acked, newly_acked_sacked, 3810 is_dupack, flag); 3811 } 3812 3813 SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt); 3814 return 0; 3815 } 3816 3817 /* Look for tcp options. Normally only called on SYN and SYNACK packets. 3818 * But, this can also be called on packets in the established flow when 3819 * the fast version below fails. 3820 */ 3821 void tcp_parse_options(const struct sk_buff *skb, struct tcp_options_received *opt_rx, 3822 const u8 **hvpp, int estab) 3823 { 3824 const unsigned char *ptr; 3825 const struct tcphdr *th = tcp_hdr(skb); 3826 int length = (th->doff * 4) - sizeof(struct tcphdr); 3827 3828 ptr = (const unsigned char *)(th + 1); 3829 opt_rx->saw_tstamp = 0; 3830 3831 while (length > 0) { 3832 int opcode = *ptr++; 3833 int opsize; 3834 3835 switch (opcode) { 3836 case TCPOPT_EOL: 3837 return; 3838 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */ 3839 length--; 3840 continue; 3841 default: 3842 opsize = *ptr++; 3843 if (opsize < 2) /* "silly options" */ 3844 return; 3845 if (opsize > length) 3846 return; /* don't parse partial options */ 3847 switch (opcode) { 3848 case TCPOPT_MSS: 3849 if (opsize == TCPOLEN_MSS && th->syn && !estab) { 3850 u16 in_mss = get_unaligned_be16(ptr); 3851 if (in_mss) { 3852 if (opt_rx->user_mss && 3853 opt_rx->user_mss < in_mss) 3854 in_mss = opt_rx->user_mss; 3855 opt_rx->mss_clamp = in_mss; 3856 } 3857 } 3858 break; 3859 case TCPOPT_WINDOW: 3860 if (opsize == TCPOLEN_WINDOW && th->syn && 3861 !estab && sysctl_tcp_window_scaling) { 3862 __u8 snd_wscale = *(__u8 *)ptr; 3863 opt_rx->wscale_ok = 1; 3864 if (snd_wscale > 14) { 3865 if (net_ratelimit()) 3866 printk(KERN_INFO "tcp_parse_options: Illegal window " 3867 "scaling value %d >14 received.\n", 3868 snd_wscale); 3869 snd_wscale = 14; 3870 } 3871 opt_rx->snd_wscale = snd_wscale; 3872 } 3873 break; 3874 case TCPOPT_TIMESTAMP: 3875 if ((opsize == TCPOLEN_TIMESTAMP) && 3876 ((estab && opt_rx->tstamp_ok) || 3877 (!estab && sysctl_tcp_timestamps))) { 3878 opt_rx->saw_tstamp = 1; 3879 opt_rx->rcv_tsval = get_unaligned_be32(ptr); 3880 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4); 3881 } 3882 break; 3883 case TCPOPT_SACK_PERM: 3884 if (opsize == TCPOLEN_SACK_PERM && th->syn && 3885 !estab && sysctl_tcp_sack) { 3886 opt_rx->sack_ok = TCP_SACK_SEEN; 3887 tcp_sack_reset(opt_rx); 3888 } 3889 break; 3890 3891 case TCPOPT_SACK: 3892 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) && 3893 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) && 3894 opt_rx->sack_ok) { 3895 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th; 3896 } 3897 break; 3898 #ifdef CONFIG_TCP_MD5SIG 3899 case TCPOPT_MD5SIG: 3900 /* 3901 * The MD5 Hash has already been 3902 * checked (see tcp_v{4,6}_do_rcv()). 3903 */ 3904 break; 3905 #endif 3906 case TCPOPT_COOKIE: 3907 /* This option is variable length. 3908 */ 3909 switch (opsize) { 3910 case TCPOLEN_COOKIE_BASE: 3911 /* not yet implemented */ 3912 break; 3913 case TCPOLEN_COOKIE_PAIR: 3914 /* not yet implemented */ 3915 break; 3916 case TCPOLEN_COOKIE_MIN+0: 3917 case TCPOLEN_COOKIE_MIN+2: 3918 case TCPOLEN_COOKIE_MIN+4: 3919 case TCPOLEN_COOKIE_MIN+6: 3920 case TCPOLEN_COOKIE_MAX: 3921 /* 16-bit multiple */ 3922 opt_rx->cookie_plus = opsize; 3923 *hvpp = ptr; 3924 break; 3925 default: 3926 /* ignore option */ 3927 break; 3928 } 3929 break; 3930 } 3931 3932 ptr += opsize-2; 3933 length -= opsize; 3934 } 3935 } 3936 } 3937 EXPORT_SYMBOL(tcp_parse_options); 3938 3939 static int tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th) 3940 { 3941 const __be32 *ptr = (const __be32 *)(th + 1); 3942 3943 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) 3944 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) { 3945 tp->rx_opt.saw_tstamp = 1; 3946 ++ptr; 3947 tp->rx_opt.rcv_tsval = ntohl(*ptr); 3948 ++ptr; 3949 tp->rx_opt.rcv_tsecr = ntohl(*ptr); 3950 return 1; 3951 } 3952 return 0; 3953 } 3954 3955 /* Fast parse options. This hopes to only see timestamps. 3956 * If it is wrong it falls back on tcp_parse_options(). 3957 */ 3958 static int tcp_fast_parse_options(const struct sk_buff *skb, 3959 const struct tcphdr *th, 3960 struct tcp_sock *tp, const u8 **hvpp) 3961 { 3962 /* In the spirit of fast parsing, compare doff directly to constant 3963 * values. Because equality is used, short doff can be ignored here. 3964 */ 3965 if (th->doff == (sizeof(*th) / 4)) { 3966 tp->rx_opt.saw_tstamp = 0; 3967 return 0; 3968 } else if (tp->rx_opt.tstamp_ok && 3969 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) { 3970 if (tcp_parse_aligned_timestamp(tp, th)) 3971 return 1; 3972 } 3973 tcp_parse_options(skb, &tp->rx_opt, hvpp, 1); 3974 return 1; 3975 } 3976 3977 #ifdef CONFIG_TCP_MD5SIG 3978 /* 3979 * Parse MD5 Signature option 3980 */ 3981 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th) 3982 { 3983 int length = (th->doff << 2) - sizeof(*th); 3984 const u8 *ptr = (const u8 *)(th + 1); 3985 3986 /* If the TCP option is too short, we can short cut */ 3987 if (length < TCPOLEN_MD5SIG) 3988 return NULL; 3989 3990 while (length > 0) { 3991 int opcode = *ptr++; 3992 int opsize; 3993 3994 switch(opcode) { 3995 case TCPOPT_EOL: 3996 return NULL; 3997 case TCPOPT_NOP: 3998 length--; 3999 continue; 4000 default: 4001 opsize = *ptr++; 4002 if (opsize < 2 || opsize > length) 4003 return NULL; 4004 if (opcode == TCPOPT_MD5SIG) 4005 return opsize == TCPOLEN_MD5SIG ? ptr : NULL; 4006 } 4007 ptr += opsize - 2; 4008 length -= opsize; 4009 } 4010 return NULL; 4011 } 4012 EXPORT_SYMBOL(tcp_parse_md5sig_option); 4013 #endif 4014 4015 static inline void tcp_store_ts_recent(struct tcp_sock *tp) 4016 { 4017 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval; 4018 tp->rx_opt.ts_recent_stamp = get_seconds(); 4019 } 4020 4021 static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq) 4022 { 4023 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) { 4024 /* PAWS bug workaround wrt. ACK frames, the PAWS discard 4025 * extra check below makes sure this can only happen 4026 * for pure ACK frames. -DaveM 4027 * 4028 * Not only, also it occurs for expired timestamps. 4029 */ 4030 4031 if (tcp_paws_check(&tp->rx_opt, 0)) 4032 tcp_store_ts_recent(tp); 4033 } 4034 } 4035 4036 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM 4037 * 4038 * It is not fatal. If this ACK does _not_ change critical state (seqs, window) 4039 * it can pass through stack. So, the following predicate verifies that 4040 * this segment is not used for anything but congestion avoidance or 4041 * fast retransmit. Moreover, we even are able to eliminate most of such 4042 * second order effects, if we apply some small "replay" window (~RTO) 4043 * to timestamp space. 4044 * 4045 * All these measures still do not guarantee that we reject wrapped ACKs 4046 * on networks with high bandwidth, when sequence space is recycled fastly, 4047 * but it guarantees that such events will be very rare and do not affect 4048 * connection seriously. This doesn't look nice, but alas, PAWS is really 4049 * buggy extension. 4050 * 4051 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC 4052 * states that events when retransmit arrives after original data are rare. 4053 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is 4054 * the biggest problem on large power networks even with minor reordering. 4055 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe 4056 * up to bandwidth of 18Gigabit/sec. 8) ] 4057 */ 4058 4059 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb) 4060 { 4061 const struct tcp_sock *tp = tcp_sk(sk); 4062 const struct tcphdr *th = tcp_hdr(skb); 4063 u32 seq = TCP_SKB_CB(skb)->seq; 4064 u32 ack = TCP_SKB_CB(skb)->ack_seq; 4065 4066 return (/* 1. Pure ACK with correct sequence number. */ 4067 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) && 4068 4069 /* 2. ... and duplicate ACK. */ 4070 ack == tp->snd_una && 4071 4072 /* 3. ... and does not update window. */ 4073 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) && 4074 4075 /* 4. ... and sits in replay window. */ 4076 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ); 4077 } 4078 4079 static inline int tcp_paws_discard(const struct sock *sk, 4080 const struct sk_buff *skb) 4081 { 4082 const struct tcp_sock *tp = tcp_sk(sk); 4083 4084 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) && 4085 !tcp_disordered_ack(sk, skb); 4086 } 4087 4088 /* Check segment sequence number for validity. 4089 * 4090 * Segment controls are considered valid, if the segment 4091 * fits to the window after truncation to the window. Acceptability 4092 * of data (and SYN, FIN, of course) is checked separately. 4093 * See tcp_data_queue(), for example. 4094 * 4095 * Also, controls (RST is main one) are accepted using RCV.WUP instead 4096 * of RCV.NXT. Peer still did not advance his SND.UNA when we 4097 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP. 4098 * (borrowed from freebsd) 4099 */ 4100 4101 static inline int tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq) 4102 { 4103 return !before(end_seq, tp->rcv_wup) && 4104 !after(seq, tp->rcv_nxt + tcp_receive_window(tp)); 4105 } 4106 4107 /* When we get a reset we do this. */ 4108 static void tcp_reset(struct sock *sk) 4109 { 4110 /* We want the right error as BSD sees it (and indeed as we do). */ 4111 switch (sk->sk_state) { 4112 case TCP_SYN_SENT: 4113 sk->sk_err = ECONNREFUSED; 4114 break; 4115 case TCP_CLOSE_WAIT: 4116 sk->sk_err = EPIPE; 4117 break; 4118 case TCP_CLOSE: 4119 return; 4120 default: 4121 sk->sk_err = ECONNRESET; 4122 } 4123 /* This barrier is coupled with smp_rmb() in tcp_poll() */ 4124 smp_wmb(); 4125 4126 if (!sock_flag(sk, SOCK_DEAD)) 4127 sk->sk_error_report(sk); 4128 4129 tcp_done(sk); 4130 } 4131 4132 /* 4133 * Process the FIN bit. This now behaves as it is supposed to work 4134 * and the FIN takes effect when it is validly part of sequence 4135 * space. Not before when we get holes. 4136 * 4137 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT 4138 * (and thence onto LAST-ACK and finally, CLOSE, we never enter 4139 * TIME-WAIT) 4140 * 4141 * If we are in FINWAIT-1, a received FIN indicates simultaneous 4142 * close and we go into CLOSING (and later onto TIME-WAIT) 4143 * 4144 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT. 4145 */ 4146 static void tcp_fin(struct sock *sk) 4147 { 4148 struct tcp_sock *tp = tcp_sk(sk); 4149 4150 inet_csk_schedule_ack(sk); 4151 4152 sk->sk_shutdown |= RCV_SHUTDOWN; 4153 sock_set_flag(sk, SOCK_DONE); 4154 4155 switch (sk->sk_state) { 4156 case TCP_SYN_RECV: 4157 case TCP_ESTABLISHED: 4158 /* Move to CLOSE_WAIT */ 4159 tcp_set_state(sk, TCP_CLOSE_WAIT); 4160 inet_csk(sk)->icsk_ack.pingpong = 1; 4161 break; 4162 4163 case TCP_CLOSE_WAIT: 4164 case TCP_CLOSING: 4165 /* Received a retransmission of the FIN, do 4166 * nothing. 4167 */ 4168 break; 4169 case TCP_LAST_ACK: 4170 /* RFC793: Remain in the LAST-ACK state. */ 4171 break; 4172 4173 case TCP_FIN_WAIT1: 4174 /* This case occurs when a simultaneous close 4175 * happens, we must ack the received FIN and 4176 * enter the CLOSING state. 4177 */ 4178 tcp_send_ack(sk); 4179 tcp_set_state(sk, TCP_CLOSING); 4180 break; 4181 case TCP_FIN_WAIT2: 4182 /* Received a FIN -- send ACK and enter TIME_WAIT. */ 4183 tcp_send_ack(sk); 4184 tcp_time_wait(sk, TCP_TIME_WAIT, 0); 4185 break; 4186 default: 4187 /* Only TCP_LISTEN and TCP_CLOSE are left, in these 4188 * cases we should never reach this piece of code. 4189 */ 4190 printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n", 4191 __func__, sk->sk_state); 4192 break; 4193 } 4194 4195 /* It _is_ possible, that we have something out-of-order _after_ FIN. 4196 * Probably, we should reset in this case. For now drop them. 4197 */ 4198 __skb_queue_purge(&tp->out_of_order_queue); 4199 if (tcp_is_sack(tp)) 4200 tcp_sack_reset(&tp->rx_opt); 4201 sk_mem_reclaim(sk); 4202 4203 if (!sock_flag(sk, SOCK_DEAD)) { 4204 sk->sk_state_change(sk); 4205 4206 /* Do not send POLL_HUP for half duplex close. */ 4207 if (sk->sk_shutdown == SHUTDOWN_MASK || 4208 sk->sk_state == TCP_CLOSE) 4209 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP); 4210 else 4211 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN); 4212 } 4213 } 4214 4215 static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, 4216 u32 end_seq) 4217 { 4218 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) { 4219 if (before(seq, sp->start_seq)) 4220 sp->start_seq = seq; 4221 if (after(end_seq, sp->end_seq)) 4222 sp->end_seq = end_seq; 4223 return 1; 4224 } 4225 return 0; 4226 } 4227 4228 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq) 4229 { 4230 struct tcp_sock *tp = tcp_sk(sk); 4231 4232 if (tcp_is_sack(tp) && sysctl_tcp_dsack) { 4233 int mib_idx; 4234 4235 if (before(seq, tp->rcv_nxt)) 4236 mib_idx = LINUX_MIB_TCPDSACKOLDSENT; 4237 else 4238 mib_idx = LINUX_MIB_TCPDSACKOFOSENT; 4239 4240 NET_INC_STATS_BH(sock_net(sk), mib_idx); 4241 4242 tp->rx_opt.dsack = 1; 4243 tp->duplicate_sack[0].start_seq = seq; 4244 tp->duplicate_sack[0].end_seq = end_seq; 4245 } 4246 } 4247 4248 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq) 4249 { 4250 struct tcp_sock *tp = tcp_sk(sk); 4251 4252 if (!tp->rx_opt.dsack) 4253 tcp_dsack_set(sk, seq, end_seq); 4254 else 4255 tcp_sack_extend(tp->duplicate_sack, seq, end_seq); 4256 } 4257 4258 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb) 4259 { 4260 struct tcp_sock *tp = tcp_sk(sk); 4261 4262 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 4263 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 4264 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST); 4265 tcp_enter_quickack_mode(sk); 4266 4267 if (tcp_is_sack(tp) && sysctl_tcp_dsack) { 4268 u32 end_seq = TCP_SKB_CB(skb)->end_seq; 4269 4270 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) 4271 end_seq = tp->rcv_nxt; 4272 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq); 4273 } 4274 } 4275 4276 tcp_send_ack(sk); 4277 } 4278 4279 /* These routines update the SACK block as out-of-order packets arrive or 4280 * in-order packets close up the sequence space. 4281 */ 4282 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp) 4283 { 4284 int this_sack; 4285 struct tcp_sack_block *sp = &tp->selective_acks[0]; 4286 struct tcp_sack_block *swalk = sp + 1; 4287 4288 /* See if the recent change to the first SACK eats into 4289 * or hits the sequence space of other SACK blocks, if so coalesce. 4290 */ 4291 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) { 4292 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) { 4293 int i; 4294 4295 /* Zap SWALK, by moving every further SACK up by one slot. 4296 * Decrease num_sacks. 4297 */ 4298 tp->rx_opt.num_sacks--; 4299 for (i = this_sack; i < tp->rx_opt.num_sacks; i++) 4300 sp[i] = sp[i + 1]; 4301 continue; 4302 } 4303 this_sack++, swalk++; 4304 } 4305 } 4306 4307 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq) 4308 { 4309 struct tcp_sock *tp = tcp_sk(sk); 4310 struct tcp_sack_block *sp = &tp->selective_acks[0]; 4311 int cur_sacks = tp->rx_opt.num_sacks; 4312 int this_sack; 4313 4314 if (!cur_sacks) 4315 goto new_sack; 4316 4317 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) { 4318 if (tcp_sack_extend(sp, seq, end_seq)) { 4319 /* Rotate this_sack to the first one. */ 4320 for (; this_sack > 0; this_sack--, sp--) 4321 swap(*sp, *(sp - 1)); 4322 if (cur_sacks > 1) 4323 tcp_sack_maybe_coalesce(tp); 4324 return; 4325 } 4326 } 4327 4328 /* Could not find an adjacent existing SACK, build a new one, 4329 * put it at the front, and shift everyone else down. We 4330 * always know there is at least one SACK present already here. 4331 * 4332 * If the sack array is full, forget about the last one. 4333 */ 4334 if (this_sack >= TCP_NUM_SACKS) { 4335 this_sack--; 4336 tp->rx_opt.num_sacks--; 4337 sp--; 4338 } 4339 for (; this_sack > 0; this_sack--, sp--) 4340 *sp = *(sp - 1); 4341 4342 new_sack: 4343 /* Build the new head SACK, and we're done. */ 4344 sp->start_seq = seq; 4345 sp->end_seq = end_seq; 4346 tp->rx_opt.num_sacks++; 4347 } 4348 4349 /* RCV.NXT advances, some SACKs should be eaten. */ 4350 4351 static void tcp_sack_remove(struct tcp_sock *tp) 4352 { 4353 struct tcp_sack_block *sp = &tp->selective_acks[0]; 4354 int num_sacks = tp->rx_opt.num_sacks; 4355 int this_sack; 4356 4357 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */ 4358 if (skb_queue_empty(&tp->out_of_order_queue)) { 4359 tp->rx_opt.num_sacks = 0; 4360 return; 4361 } 4362 4363 for (this_sack = 0; this_sack < num_sacks;) { 4364 /* Check if the start of the sack is covered by RCV.NXT. */ 4365 if (!before(tp->rcv_nxt, sp->start_seq)) { 4366 int i; 4367 4368 /* RCV.NXT must cover all the block! */ 4369 WARN_ON(before(tp->rcv_nxt, sp->end_seq)); 4370 4371 /* Zap this SACK, by moving forward any other SACKS. */ 4372 for (i=this_sack+1; i < num_sacks; i++) 4373 tp->selective_acks[i-1] = tp->selective_acks[i]; 4374 num_sacks--; 4375 continue; 4376 } 4377 this_sack++; 4378 sp++; 4379 } 4380 tp->rx_opt.num_sacks = num_sacks; 4381 } 4382 4383 /* This one checks to see if we can put data from the 4384 * out_of_order queue into the receive_queue. 4385 */ 4386 static void tcp_ofo_queue(struct sock *sk) 4387 { 4388 struct tcp_sock *tp = tcp_sk(sk); 4389 __u32 dsack_high = tp->rcv_nxt; 4390 struct sk_buff *skb; 4391 4392 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) { 4393 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) 4394 break; 4395 4396 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) { 4397 __u32 dsack = dsack_high; 4398 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high)) 4399 dsack_high = TCP_SKB_CB(skb)->end_seq; 4400 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack); 4401 } 4402 4403 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) { 4404 SOCK_DEBUG(sk, "ofo packet was already received\n"); 4405 __skb_unlink(skb, &tp->out_of_order_queue); 4406 __kfree_skb(skb); 4407 continue; 4408 } 4409 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n", 4410 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, 4411 TCP_SKB_CB(skb)->end_seq); 4412 4413 __skb_unlink(skb, &tp->out_of_order_queue); 4414 __skb_queue_tail(&sk->sk_receive_queue, skb); 4415 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; 4416 if (tcp_hdr(skb)->fin) 4417 tcp_fin(sk); 4418 } 4419 } 4420 4421 static int tcp_prune_ofo_queue(struct sock *sk); 4422 static int tcp_prune_queue(struct sock *sk); 4423 4424 static inline int tcp_try_rmem_schedule(struct sock *sk, unsigned int size) 4425 { 4426 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf || 4427 !sk_rmem_schedule(sk, size)) { 4428 4429 if (tcp_prune_queue(sk) < 0) 4430 return -1; 4431 4432 if (!sk_rmem_schedule(sk, size)) { 4433 if (!tcp_prune_ofo_queue(sk)) 4434 return -1; 4435 4436 if (!sk_rmem_schedule(sk, size)) 4437 return -1; 4438 } 4439 } 4440 return 0; 4441 } 4442 4443 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb) 4444 { 4445 const struct tcphdr *th = tcp_hdr(skb); 4446 struct tcp_sock *tp = tcp_sk(sk); 4447 int eaten = -1; 4448 4449 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) 4450 goto drop; 4451 4452 skb_dst_drop(skb); 4453 __skb_pull(skb, th->doff * 4); 4454 4455 TCP_ECN_accept_cwr(tp, skb); 4456 4457 tp->rx_opt.dsack = 0; 4458 4459 /* Queue data for delivery to the user. 4460 * Packets in sequence go to the receive queue. 4461 * Out of sequence packets to the out_of_order_queue. 4462 */ 4463 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) { 4464 if (tcp_receive_window(tp) == 0) 4465 goto out_of_window; 4466 4467 /* Ok. In sequence. In window. */ 4468 if (tp->ucopy.task == current && 4469 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len && 4470 sock_owned_by_user(sk) && !tp->urg_data) { 4471 int chunk = min_t(unsigned int, skb->len, 4472 tp->ucopy.len); 4473 4474 __set_current_state(TASK_RUNNING); 4475 4476 local_bh_enable(); 4477 if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) { 4478 tp->ucopy.len -= chunk; 4479 tp->copied_seq += chunk; 4480 eaten = (chunk == skb->len); 4481 tcp_rcv_space_adjust(sk); 4482 } 4483 local_bh_disable(); 4484 } 4485 4486 if (eaten <= 0) { 4487 queue_and_out: 4488 if (eaten < 0 && 4489 tcp_try_rmem_schedule(sk, skb->truesize)) 4490 goto drop; 4491 4492 skb_set_owner_r(skb, sk); 4493 __skb_queue_tail(&sk->sk_receive_queue, skb); 4494 } 4495 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; 4496 if (skb->len) 4497 tcp_event_data_recv(sk, skb); 4498 if (th->fin) 4499 tcp_fin(sk); 4500 4501 if (!skb_queue_empty(&tp->out_of_order_queue)) { 4502 tcp_ofo_queue(sk); 4503 4504 /* RFC2581. 4.2. SHOULD send immediate ACK, when 4505 * gap in queue is filled. 4506 */ 4507 if (skb_queue_empty(&tp->out_of_order_queue)) 4508 inet_csk(sk)->icsk_ack.pingpong = 0; 4509 } 4510 4511 if (tp->rx_opt.num_sacks) 4512 tcp_sack_remove(tp); 4513 4514 tcp_fast_path_check(sk); 4515 4516 if (eaten > 0) 4517 __kfree_skb(skb); 4518 else if (!sock_flag(sk, SOCK_DEAD)) 4519 sk->sk_data_ready(sk, 0); 4520 return; 4521 } 4522 4523 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) { 4524 /* A retransmit, 2nd most common case. Force an immediate ack. */ 4525 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST); 4526 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); 4527 4528 out_of_window: 4529 tcp_enter_quickack_mode(sk); 4530 inet_csk_schedule_ack(sk); 4531 drop: 4532 __kfree_skb(skb); 4533 return; 4534 } 4535 4536 /* Out of window. F.e. zero window probe. */ 4537 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp))) 4538 goto out_of_window; 4539 4540 tcp_enter_quickack_mode(sk); 4541 4542 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 4543 /* Partial packet, seq < rcv_next < end_seq */ 4544 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n", 4545 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, 4546 TCP_SKB_CB(skb)->end_seq); 4547 4548 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt); 4549 4550 /* If window is closed, drop tail of packet. But after 4551 * remembering D-SACK for its head made in previous line. 4552 */ 4553 if (!tcp_receive_window(tp)) 4554 goto out_of_window; 4555 goto queue_and_out; 4556 } 4557 4558 TCP_ECN_check_ce(tp, skb); 4559 4560 if (tcp_try_rmem_schedule(sk, skb->truesize)) 4561 goto drop; 4562 4563 /* Disable header prediction. */ 4564 tp->pred_flags = 0; 4565 inet_csk_schedule_ack(sk); 4566 4567 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n", 4568 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); 4569 4570 skb_set_owner_r(skb, sk); 4571 4572 if (!skb_peek(&tp->out_of_order_queue)) { 4573 /* Initial out of order segment, build 1 SACK. */ 4574 if (tcp_is_sack(tp)) { 4575 tp->rx_opt.num_sacks = 1; 4576 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq; 4577 tp->selective_acks[0].end_seq = 4578 TCP_SKB_CB(skb)->end_seq; 4579 } 4580 __skb_queue_head(&tp->out_of_order_queue, skb); 4581 } else { 4582 struct sk_buff *skb1 = skb_peek_tail(&tp->out_of_order_queue); 4583 u32 seq = TCP_SKB_CB(skb)->seq; 4584 u32 end_seq = TCP_SKB_CB(skb)->end_seq; 4585 4586 if (seq == TCP_SKB_CB(skb1)->end_seq) { 4587 __skb_queue_after(&tp->out_of_order_queue, skb1, skb); 4588 4589 if (!tp->rx_opt.num_sacks || 4590 tp->selective_acks[0].end_seq != seq) 4591 goto add_sack; 4592 4593 /* Common case: data arrive in order after hole. */ 4594 tp->selective_acks[0].end_seq = end_seq; 4595 return; 4596 } 4597 4598 /* Find place to insert this segment. */ 4599 while (1) { 4600 if (!after(TCP_SKB_CB(skb1)->seq, seq)) 4601 break; 4602 if (skb_queue_is_first(&tp->out_of_order_queue, skb1)) { 4603 skb1 = NULL; 4604 break; 4605 } 4606 skb1 = skb_queue_prev(&tp->out_of_order_queue, skb1); 4607 } 4608 4609 /* Do skb overlap to previous one? */ 4610 if (skb1 && before(seq, TCP_SKB_CB(skb1)->end_seq)) { 4611 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) { 4612 /* All the bits are present. Drop. */ 4613 __kfree_skb(skb); 4614 tcp_dsack_set(sk, seq, end_seq); 4615 goto add_sack; 4616 } 4617 if (after(seq, TCP_SKB_CB(skb1)->seq)) { 4618 /* Partial overlap. */ 4619 tcp_dsack_set(sk, seq, 4620 TCP_SKB_CB(skb1)->end_seq); 4621 } else { 4622 if (skb_queue_is_first(&tp->out_of_order_queue, 4623 skb1)) 4624 skb1 = NULL; 4625 else 4626 skb1 = skb_queue_prev( 4627 &tp->out_of_order_queue, 4628 skb1); 4629 } 4630 } 4631 if (!skb1) 4632 __skb_queue_head(&tp->out_of_order_queue, skb); 4633 else 4634 __skb_queue_after(&tp->out_of_order_queue, skb1, skb); 4635 4636 /* And clean segments covered by new one as whole. */ 4637 while (!skb_queue_is_last(&tp->out_of_order_queue, skb)) { 4638 skb1 = skb_queue_next(&tp->out_of_order_queue, skb); 4639 4640 if (!after(end_seq, TCP_SKB_CB(skb1)->seq)) 4641 break; 4642 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) { 4643 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq, 4644 end_seq); 4645 break; 4646 } 4647 __skb_unlink(skb1, &tp->out_of_order_queue); 4648 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq, 4649 TCP_SKB_CB(skb1)->end_seq); 4650 __kfree_skb(skb1); 4651 } 4652 4653 add_sack: 4654 if (tcp_is_sack(tp)) 4655 tcp_sack_new_ofo_skb(sk, seq, end_seq); 4656 } 4657 } 4658 4659 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb, 4660 struct sk_buff_head *list) 4661 { 4662 struct sk_buff *next = NULL; 4663 4664 if (!skb_queue_is_last(list, skb)) 4665 next = skb_queue_next(list, skb); 4666 4667 __skb_unlink(skb, list); 4668 __kfree_skb(skb); 4669 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED); 4670 4671 return next; 4672 } 4673 4674 /* Collapse contiguous sequence of skbs head..tail with 4675 * sequence numbers start..end. 4676 * 4677 * If tail is NULL, this means until the end of the list. 4678 * 4679 * Segments with FIN/SYN are not collapsed (only because this 4680 * simplifies code) 4681 */ 4682 static void 4683 tcp_collapse(struct sock *sk, struct sk_buff_head *list, 4684 struct sk_buff *head, struct sk_buff *tail, 4685 u32 start, u32 end) 4686 { 4687 struct sk_buff *skb, *n; 4688 bool end_of_skbs; 4689 4690 /* First, check that queue is collapsible and find 4691 * the point where collapsing can be useful. */ 4692 skb = head; 4693 restart: 4694 end_of_skbs = true; 4695 skb_queue_walk_from_safe(list, skb, n) { 4696 if (skb == tail) 4697 break; 4698 /* No new bits? It is possible on ofo queue. */ 4699 if (!before(start, TCP_SKB_CB(skb)->end_seq)) { 4700 skb = tcp_collapse_one(sk, skb, list); 4701 if (!skb) 4702 break; 4703 goto restart; 4704 } 4705 4706 /* The first skb to collapse is: 4707 * - not SYN/FIN and 4708 * - bloated or contains data before "start" or 4709 * overlaps to the next one. 4710 */ 4711 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin && 4712 (tcp_win_from_space(skb->truesize) > skb->len || 4713 before(TCP_SKB_CB(skb)->seq, start))) { 4714 end_of_skbs = false; 4715 break; 4716 } 4717 4718 if (!skb_queue_is_last(list, skb)) { 4719 struct sk_buff *next = skb_queue_next(list, skb); 4720 if (next != tail && 4721 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(next)->seq) { 4722 end_of_skbs = false; 4723 break; 4724 } 4725 } 4726 4727 /* Decided to skip this, advance start seq. */ 4728 start = TCP_SKB_CB(skb)->end_seq; 4729 } 4730 if (end_of_skbs || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin) 4731 return; 4732 4733 while (before(start, end)) { 4734 struct sk_buff *nskb; 4735 unsigned int header = skb_headroom(skb); 4736 int copy = SKB_MAX_ORDER(header, 0); 4737 4738 /* Too big header? This can happen with IPv6. */ 4739 if (copy < 0) 4740 return; 4741 if (end - start < copy) 4742 copy = end - start; 4743 nskb = alloc_skb(copy + header, GFP_ATOMIC); 4744 if (!nskb) 4745 return; 4746 4747 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head); 4748 skb_set_network_header(nskb, (skb_network_header(skb) - 4749 skb->head)); 4750 skb_set_transport_header(nskb, (skb_transport_header(skb) - 4751 skb->head)); 4752 skb_reserve(nskb, header); 4753 memcpy(nskb->head, skb->head, header); 4754 memcpy(nskb->cb, skb->cb, sizeof(skb->cb)); 4755 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start; 4756 __skb_queue_before(list, skb, nskb); 4757 skb_set_owner_r(nskb, sk); 4758 4759 /* Copy data, releasing collapsed skbs. */ 4760 while (copy > 0) { 4761 int offset = start - TCP_SKB_CB(skb)->seq; 4762 int size = TCP_SKB_CB(skb)->end_seq - start; 4763 4764 BUG_ON(offset < 0); 4765 if (size > 0) { 4766 size = min(copy, size); 4767 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size)) 4768 BUG(); 4769 TCP_SKB_CB(nskb)->end_seq += size; 4770 copy -= size; 4771 start += size; 4772 } 4773 if (!before(start, TCP_SKB_CB(skb)->end_seq)) { 4774 skb = tcp_collapse_one(sk, skb, list); 4775 if (!skb || 4776 skb == tail || 4777 tcp_hdr(skb)->syn || 4778 tcp_hdr(skb)->fin) 4779 return; 4780 } 4781 } 4782 } 4783 } 4784 4785 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs 4786 * and tcp_collapse() them until all the queue is collapsed. 4787 */ 4788 static void tcp_collapse_ofo_queue(struct sock *sk) 4789 { 4790 struct tcp_sock *tp = tcp_sk(sk); 4791 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue); 4792 struct sk_buff *head; 4793 u32 start, end; 4794 4795 if (skb == NULL) 4796 return; 4797 4798 start = TCP_SKB_CB(skb)->seq; 4799 end = TCP_SKB_CB(skb)->end_seq; 4800 head = skb; 4801 4802 for (;;) { 4803 struct sk_buff *next = NULL; 4804 4805 if (!skb_queue_is_last(&tp->out_of_order_queue, skb)) 4806 next = skb_queue_next(&tp->out_of_order_queue, skb); 4807 skb = next; 4808 4809 /* Segment is terminated when we see gap or when 4810 * we are at the end of all the queue. */ 4811 if (!skb || 4812 after(TCP_SKB_CB(skb)->seq, end) || 4813 before(TCP_SKB_CB(skb)->end_seq, start)) { 4814 tcp_collapse(sk, &tp->out_of_order_queue, 4815 head, skb, start, end); 4816 head = skb; 4817 if (!skb) 4818 break; 4819 /* Start new segment */ 4820 start = TCP_SKB_CB(skb)->seq; 4821 end = TCP_SKB_CB(skb)->end_seq; 4822 } else { 4823 if (before(TCP_SKB_CB(skb)->seq, start)) 4824 start = TCP_SKB_CB(skb)->seq; 4825 if (after(TCP_SKB_CB(skb)->end_seq, end)) 4826 end = TCP_SKB_CB(skb)->end_seq; 4827 } 4828 } 4829 } 4830 4831 /* 4832 * Purge the out-of-order queue. 4833 * Return true if queue was pruned. 4834 */ 4835 static int tcp_prune_ofo_queue(struct sock *sk) 4836 { 4837 struct tcp_sock *tp = tcp_sk(sk); 4838 int res = 0; 4839 4840 if (!skb_queue_empty(&tp->out_of_order_queue)) { 4841 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED); 4842 __skb_queue_purge(&tp->out_of_order_queue); 4843 4844 /* Reset SACK state. A conforming SACK implementation will 4845 * do the same at a timeout based retransmit. When a connection 4846 * is in a sad state like this, we care only about integrity 4847 * of the connection not performance. 4848 */ 4849 if (tp->rx_opt.sack_ok) 4850 tcp_sack_reset(&tp->rx_opt); 4851 sk_mem_reclaim(sk); 4852 res = 1; 4853 } 4854 return res; 4855 } 4856 4857 /* Reduce allocated memory if we can, trying to get 4858 * the socket within its memory limits again. 4859 * 4860 * Return less than zero if we should start dropping frames 4861 * until the socket owning process reads some of the data 4862 * to stabilize the situation. 4863 */ 4864 static int tcp_prune_queue(struct sock *sk) 4865 { 4866 struct tcp_sock *tp = tcp_sk(sk); 4867 4868 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq); 4869 4870 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED); 4871 4872 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf) 4873 tcp_clamp_window(sk); 4874 else if (sk_under_memory_pressure(sk)) 4875 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss); 4876 4877 tcp_collapse_ofo_queue(sk); 4878 if (!skb_queue_empty(&sk->sk_receive_queue)) 4879 tcp_collapse(sk, &sk->sk_receive_queue, 4880 skb_peek(&sk->sk_receive_queue), 4881 NULL, 4882 tp->copied_seq, tp->rcv_nxt); 4883 sk_mem_reclaim(sk); 4884 4885 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) 4886 return 0; 4887 4888 /* Collapsing did not help, destructive actions follow. 4889 * This must not ever occur. */ 4890 4891 tcp_prune_ofo_queue(sk); 4892 4893 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) 4894 return 0; 4895 4896 /* If we are really being abused, tell the caller to silently 4897 * drop receive data on the floor. It will get retransmitted 4898 * and hopefully then we'll have sufficient space. 4899 */ 4900 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED); 4901 4902 /* Massive buffer overcommit. */ 4903 tp->pred_flags = 0; 4904 return -1; 4905 } 4906 4907 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. 4908 * As additional protections, we do not touch cwnd in retransmission phases, 4909 * and if application hit its sndbuf limit recently. 4910 */ 4911 void tcp_cwnd_application_limited(struct sock *sk) 4912 { 4913 struct tcp_sock *tp = tcp_sk(sk); 4914 4915 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && 4916 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { 4917 /* Limited by application or receiver window. */ 4918 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); 4919 u32 win_used = max(tp->snd_cwnd_used, init_win); 4920 if (win_used < tp->snd_cwnd) { 4921 tp->snd_ssthresh = tcp_current_ssthresh(sk); 4922 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1; 4923 } 4924 tp->snd_cwnd_used = 0; 4925 } 4926 tp->snd_cwnd_stamp = tcp_time_stamp; 4927 } 4928 4929 static int tcp_should_expand_sndbuf(const struct sock *sk) 4930 { 4931 const struct tcp_sock *tp = tcp_sk(sk); 4932 4933 /* If the user specified a specific send buffer setting, do 4934 * not modify it. 4935 */ 4936 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK) 4937 return 0; 4938 4939 /* If we are under global TCP memory pressure, do not expand. */ 4940 if (sk_under_memory_pressure(sk)) 4941 return 0; 4942 4943 /* If we are under soft global TCP memory pressure, do not expand. */ 4944 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0)) 4945 return 0; 4946 4947 /* If we filled the congestion window, do not expand. */ 4948 if (tp->packets_out >= tp->snd_cwnd) 4949 return 0; 4950 4951 return 1; 4952 } 4953 4954 /* When incoming ACK allowed to free some skb from write_queue, 4955 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket 4956 * on the exit from tcp input handler. 4957 * 4958 * PROBLEM: sndbuf expansion does not work well with largesend. 4959 */ 4960 static void tcp_new_space(struct sock *sk) 4961 { 4962 struct tcp_sock *tp = tcp_sk(sk); 4963 4964 if (tcp_should_expand_sndbuf(sk)) { 4965 int sndmem = SKB_TRUESIZE(max_t(u32, 4966 tp->rx_opt.mss_clamp, 4967 tp->mss_cache) + 4968 MAX_TCP_HEADER); 4969 int demanded = max_t(unsigned int, tp->snd_cwnd, 4970 tp->reordering + 1); 4971 sndmem *= 2 * demanded; 4972 if (sndmem > sk->sk_sndbuf) 4973 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]); 4974 tp->snd_cwnd_stamp = tcp_time_stamp; 4975 } 4976 4977 sk->sk_write_space(sk); 4978 } 4979 4980 static void tcp_check_space(struct sock *sk) 4981 { 4982 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) { 4983 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK); 4984 if (sk->sk_socket && 4985 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) 4986 tcp_new_space(sk); 4987 } 4988 } 4989 4990 static inline void tcp_data_snd_check(struct sock *sk) 4991 { 4992 tcp_push_pending_frames(sk); 4993 tcp_check_space(sk); 4994 } 4995 4996 /* 4997 * Check if sending an ack is needed. 4998 */ 4999 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible) 5000 { 5001 struct tcp_sock *tp = tcp_sk(sk); 5002 5003 /* More than one full frame received... */ 5004 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss && 5005 /* ... and right edge of window advances far enough. 5006 * (tcp_recvmsg() will send ACK otherwise). Or... 5007 */ 5008 __tcp_select_window(sk) >= tp->rcv_wnd) || 5009 /* We ACK each frame or... */ 5010 tcp_in_quickack_mode(sk) || 5011 /* We have out of order data. */ 5012 (ofo_possible && skb_peek(&tp->out_of_order_queue))) { 5013 /* Then ack it now */ 5014 tcp_send_ack(sk); 5015 } else { 5016 /* Else, send delayed ack. */ 5017 tcp_send_delayed_ack(sk); 5018 } 5019 } 5020 5021 static inline void tcp_ack_snd_check(struct sock *sk) 5022 { 5023 if (!inet_csk_ack_scheduled(sk)) { 5024 /* We sent a data segment already. */ 5025 return; 5026 } 5027 __tcp_ack_snd_check(sk, 1); 5028 } 5029 5030 /* 5031 * This routine is only called when we have urgent data 5032 * signaled. Its the 'slow' part of tcp_urg. It could be 5033 * moved inline now as tcp_urg is only called from one 5034 * place. We handle URGent data wrong. We have to - as 5035 * BSD still doesn't use the correction from RFC961. 5036 * For 1003.1g we should support a new option TCP_STDURG to permit 5037 * either form (or just set the sysctl tcp_stdurg). 5038 */ 5039 5040 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th) 5041 { 5042 struct tcp_sock *tp = tcp_sk(sk); 5043 u32 ptr = ntohs(th->urg_ptr); 5044 5045 if (ptr && !sysctl_tcp_stdurg) 5046 ptr--; 5047 ptr += ntohl(th->seq); 5048 5049 /* Ignore urgent data that we've already seen and read. */ 5050 if (after(tp->copied_seq, ptr)) 5051 return; 5052 5053 /* Do not replay urg ptr. 5054 * 5055 * NOTE: interesting situation not covered by specs. 5056 * Misbehaving sender may send urg ptr, pointing to segment, 5057 * which we already have in ofo queue. We are not able to fetch 5058 * such data and will stay in TCP_URG_NOTYET until will be eaten 5059 * by recvmsg(). Seems, we are not obliged to handle such wicked 5060 * situations. But it is worth to think about possibility of some 5061 * DoSes using some hypothetical application level deadlock. 5062 */ 5063 if (before(ptr, tp->rcv_nxt)) 5064 return; 5065 5066 /* Do we already have a newer (or duplicate) urgent pointer? */ 5067 if (tp->urg_data && !after(ptr, tp->urg_seq)) 5068 return; 5069 5070 /* Tell the world about our new urgent pointer. */ 5071 sk_send_sigurg(sk); 5072 5073 /* We may be adding urgent data when the last byte read was 5074 * urgent. To do this requires some care. We cannot just ignore 5075 * tp->copied_seq since we would read the last urgent byte again 5076 * as data, nor can we alter copied_seq until this data arrives 5077 * or we break the semantics of SIOCATMARK (and thus sockatmark()) 5078 * 5079 * NOTE. Double Dutch. Rendering to plain English: author of comment 5080 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB); 5081 * and expect that both A and B disappear from stream. This is _wrong_. 5082 * Though this happens in BSD with high probability, this is occasional. 5083 * Any application relying on this is buggy. Note also, that fix "works" 5084 * only in this artificial test. Insert some normal data between A and B and we will 5085 * decline of BSD again. Verdict: it is better to remove to trap 5086 * buggy users. 5087 */ 5088 if (tp->urg_seq == tp->copied_seq && tp->urg_data && 5089 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) { 5090 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue); 5091 tp->copied_seq++; 5092 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) { 5093 __skb_unlink(skb, &sk->sk_receive_queue); 5094 __kfree_skb(skb); 5095 } 5096 } 5097 5098 tp->urg_data = TCP_URG_NOTYET; 5099 tp->urg_seq = ptr; 5100 5101 /* Disable header prediction. */ 5102 tp->pred_flags = 0; 5103 } 5104 5105 /* This is the 'fast' part of urgent handling. */ 5106 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th) 5107 { 5108 struct tcp_sock *tp = tcp_sk(sk); 5109 5110 /* Check if we get a new urgent pointer - normally not. */ 5111 if (th->urg) 5112 tcp_check_urg(sk, th); 5113 5114 /* Do we wait for any urgent data? - normally not... */ 5115 if (tp->urg_data == TCP_URG_NOTYET) { 5116 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) - 5117 th->syn; 5118 5119 /* Is the urgent pointer pointing into this packet? */ 5120 if (ptr < skb->len) { 5121 u8 tmp; 5122 if (skb_copy_bits(skb, ptr, &tmp, 1)) 5123 BUG(); 5124 tp->urg_data = TCP_URG_VALID | tmp; 5125 if (!sock_flag(sk, SOCK_DEAD)) 5126 sk->sk_data_ready(sk, 0); 5127 } 5128 } 5129 } 5130 5131 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen) 5132 { 5133 struct tcp_sock *tp = tcp_sk(sk); 5134 int chunk = skb->len - hlen; 5135 int err; 5136 5137 local_bh_enable(); 5138 if (skb_csum_unnecessary(skb)) 5139 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk); 5140 else 5141 err = skb_copy_and_csum_datagram_iovec(skb, hlen, 5142 tp->ucopy.iov); 5143 5144 if (!err) { 5145 tp->ucopy.len -= chunk; 5146 tp->copied_seq += chunk; 5147 tcp_rcv_space_adjust(sk); 5148 } 5149 5150 local_bh_disable(); 5151 return err; 5152 } 5153 5154 static __sum16 __tcp_checksum_complete_user(struct sock *sk, 5155 struct sk_buff *skb) 5156 { 5157 __sum16 result; 5158 5159 if (sock_owned_by_user(sk)) { 5160 local_bh_enable(); 5161 result = __tcp_checksum_complete(skb); 5162 local_bh_disable(); 5163 } else { 5164 result = __tcp_checksum_complete(skb); 5165 } 5166 return result; 5167 } 5168 5169 static inline int tcp_checksum_complete_user(struct sock *sk, 5170 struct sk_buff *skb) 5171 { 5172 return !skb_csum_unnecessary(skb) && 5173 __tcp_checksum_complete_user(sk, skb); 5174 } 5175 5176 #ifdef CONFIG_NET_DMA 5177 static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb, 5178 int hlen) 5179 { 5180 struct tcp_sock *tp = tcp_sk(sk); 5181 int chunk = skb->len - hlen; 5182 int dma_cookie; 5183 int copied_early = 0; 5184 5185 if (tp->ucopy.wakeup) 5186 return 0; 5187 5188 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list) 5189 tp->ucopy.dma_chan = dma_find_channel(DMA_MEMCPY); 5190 5191 if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) { 5192 5193 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan, 5194 skb, hlen, 5195 tp->ucopy.iov, chunk, 5196 tp->ucopy.pinned_list); 5197 5198 if (dma_cookie < 0) 5199 goto out; 5200 5201 tp->ucopy.dma_cookie = dma_cookie; 5202 copied_early = 1; 5203 5204 tp->ucopy.len -= chunk; 5205 tp->copied_seq += chunk; 5206 tcp_rcv_space_adjust(sk); 5207 5208 if ((tp->ucopy.len == 0) || 5209 (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) || 5210 (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) { 5211 tp->ucopy.wakeup = 1; 5212 sk->sk_data_ready(sk, 0); 5213 } 5214 } else if (chunk > 0) { 5215 tp->ucopy.wakeup = 1; 5216 sk->sk_data_ready(sk, 0); 5217 } 5218 out: 5219 return copied_early; 5220 } 5221 #endif /* CONFIG_NET_DMA */ 5222 5223 /* Does PAWS and seqno based validation of an incoming segment, flags will 5224 * play significant role here. 5225 */ 5226 static int tcp_validate_incoming(struct sock *sk, struct sk_buff *skb, 5227 const struct tcphdr *th, int syn_inerr) 5228 { 5229 const u8 *hash_location; 5230 struct tcp_sock *tp = tcp_sk(sk); 5231 5232 /* RFC1323: H1. Apply PAWS check first. */ 5233 if (tcp_fast_parse_options(skb, th, tp, &hash_location) && 5234 tp->rx_opt.saw_tstamp && 5235 tcp_paws_discard(sk, skb)) { 5236 if (!th->rst) { 5237 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED); 5238 tcp_send_dupack(sk, skb); 5239 goto discard; 5240 } 5241 /* Reset is accepted even if it did not pass PAWS. */ 5242 } 5243 5244 /* Step 1: check sequence number */ 5245 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) { 5246 /* RFC793, page 37: "In all states except SYN-SENT, all reset 5247 * (RST) segments are validated by checking their SEQ-fields." 5248 * And page 69: "If an incoming segment is not acceptable, 5249 * an acknowledgment should be sent in reply (unless the RST 5250 * bit is set, if so drop the segment and return)". 5251 */ 5252 if (!th->rst) 5253 tcp_send_dupack(sk, skb); 5254 goto discard; 5255 } 5256 5257 /* Step 2: check RST bit */ 5258 if (th->rst) { 5259 tcp_reset(sk); 5260 goto discard; 5261 } 5262 5263 /* ts_recent update must be made after we are sure that the packet 5264 * is in window. 5265 */ 5266 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq); 5267 5268 /* step 3: check security and precedence [ignored] */ 5269 5270 /* step 4: Check for a SYN in window. */ 5271 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 5272 if (syn_inerr) 5273 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS); 5274 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONSYN); 5275 tcp_reset(sk); 5276 return -1; 5277 } 5278 5279 return 1; 5280 5281 discard: 5282 __kfree_skb(skb); 5283 return 0; 5284 } 5285 5286 /* 5287 * TCP receive function for the ESTABLISHED state. 5288 * 5289 * It is split into a fast path and a slow path. The fast path is 5290 * disabled when: 5291 * - A zero window was announced from us - zero window probing 5292 * is only handled properly in the slow path. 5293 * - Out of order segments arrived. 5294 * - Urgent data is expected. 5295 * - There is no buffer space left 5296 * - Unexpected TCP flags/window values/header lengths are received 5297 * (detected by checking the TCP header against pred_flags) 5298 * - Data is sent in both directions. Fast path only supports pure senders 5299 * or pure receivers (this means either the sequence number or the ack 5300 * value must stay constant) 5301 * - Unexpected TCP option. 5302 * 5303 * When these conditions are not satisfied it drops into a standard 5304 * receive procedure patterned after RFC793 to handle all cases. 5305 * The first three cases are guaranteed by proper pred_flags setting, 5306 * the rest is checked inline. Fast processing is turned on in 5307 * tcp_data_queue when everything is OK. 5308 */ 5309 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb, 5310 const struct tcphdr *th, unsigned int len) 5311 { 5312 struct tcp_sock *tp = tcp_sk(sk); 5313 int res; 5314 5315 /* 5316 * Header prediction. 5317 * The code loosely follows the one in the famous 5318 * "30 instruction TCP receive" Van Jacobson mail. 5319 * 5320 * Van's trick is to deposit buffers into socket queue 5321 * on a device interrupt, to call tcp_recv function 5322 * on the receive process context and checksum and copy 5323 * the buffer to user space. smart... 5324 * 5325 * Our current scheme is not silly either but we take the 5326 * extra cost of the net_bh soft interrupt processing... 5327 * We do checksum and copy also but from device to kernel. 5328 */ 5329 5330 tp->rx_opt.saw_tstamp = 0; 5331 5332 /* pred_flags is 0xS?10 << 16 + snd_wnd 5333 * if header_prediction is to be made 5334 * 'S' will always be tp->tcp_header_len >> 2 5335 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to 5336 * turn it off (when there are holes in the receive 5337 * space for instance) 5338 * PSH flag is ignored. 5339 */ 5340 5341 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags && 5342 TCP_SKB_CB(skb)->seq == tp->rcv_nxt && 5343 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) { 5344 int tcp_header_len = tp->tcp_header_len; 5345 5346 /* Timestamp header prediction: tcp_header_len 5347 * is automatically equal to th->doff*4 due to pred_flags 5348 * match. 5349 */ 5350 5351 /* Check timestamp */ 5352 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) { 5353 /* No? Slow path! */ 5354 if (!tcp_parse_aligned_timestamp(tp, th)) 5355 goto slow_path; 5356 5357 /* If PAWS failed, check it more carefully in slow path */ 5358 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0) 5359 goto slow_path; 5360 5361 /* DO NOT update ts_recent here, if checksum fails 5362 * and timestamp was corrupted part, it will result 5363 * in a hung connection since we will drop all 5364 * future packets due to the PAWS test. 5365 */ 5366 } 5367 5368 if (len <= tcp_header_len) { 5369 /* Bulk data transfer: sender */ 5370 if (len == tcp_header_len) { 5371 /* Predicted packet is in window by definition. 5372 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 5373 * Hence, check seq<=rcv_wup reduces to: 5374 */ 5375 if (tcp_header_len == 5376 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && 5377 tp->rcv_nxt == tp->rcv_wup) 5378 tcp_store_ts_recent(tp); 5379 5380 /* We know that such packets are checksummed 5381 * on entry. 5382 */ 5383 tcp_ack(sk, skb, 0); 5384 __kfree_skb(skb); 5385 tcp_data_snd_check(sk); 5386 return 0; 5387 } else { /* Header too small */ 5388 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS); 5389 goto discard; 5390 } 5391 } else { 5392 int eaten = 0; 5393 int copied_early = 0; 5394 5395 if (tp->copied_seq == tp->rcv_nxt && 5396 len - tcp_header_len <= tp->ucopy.len) { 5397 #ifdef CONFIG_NET_DMA 5398 if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) { 5399 copied_early = 1; 5400 eaten = 1; 5401 } 5402 #endif 5403 if (tp->ucopy.task == current && 5404 sock_owned_by_user(sk) && !copied_early) { 5405 __set_current_state(TASK_RUNNING); 5406 5407 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len)) 5408 eaten = 1; 5409 } 5410 if (eaten) { 5411 /* Predicted packet is in window by definition. 5412 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 5413 * Hence, check seq<=rcv_wup reduces to: 5414 */ 5415 if (tcp_header_len == 5416 (sizeof(struct tcphdr) + 5417 TCPOLEN_TSTAMP_ALIGNED) && 5418 tp->rcv_nxt == tp->rcv_wup) 5419 tcp_store_ts_recent(tp); 5420 5421 tcp_rcv_rtt_measure_ts(sk, skb); 5422 5423 __skb_pull(skb, tcp_header_len); 5424 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; 5425 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER); 5426 } 5427 if (copied_early) 5428 tcp_cleanup_rbuf(sk, skb->len); 5429 } 5430 if (!eaten) { 5431 if (tcp_checksum_complete_user(sk, skb)) 5432 goto csum_error; 5433 5434 /* Predicted packet is in window by definition. 5435 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 5436 * Hence, check seq<=rcv_wup reduces to: 5437 */ 5438 if (tcp_header_len == 5439 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && 5440 tp->rcv_nxt == tp->rcv_wup) 5441 tcp_store_ts_recent(tp); 5442 5443 tcp_rcv_rtt_measure_ts(sk, skb); 5444 5445 if ((int)skb->truesize > sk->sk_forward_alloc) 5446 goto step5; 5447 5448 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS); 5449 5450 /* Bulk data transfer: receiver */ 5451 __skb_pull(skb, tcp_header_len); 5452 __skb_queue_tail(&sk->sk_receive_queue, skb); 5453 skb_set_owner_r(skb, sk); 5454 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; 5455 } 5456 5457 tcp_event_data_recv(sk, skb); 5458 5459 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) { 5460 /* Well, only one small jumplet in fast path... */ 5461 tcp_ack(sk, skb, FLAG_DATA); 5462 tcp_data_snd_check(sk); 5463 if (!inet_csk_ack_scheduled(sk)) 5464 goto no_ack; 5465 } 5466 5467 if (!copied_early || tp->rcv_nxt != tp->rcv_wup) 5468 __tcp_ack_snd_check(sk, 0); 5469 no_ack: 5470 #ifdef CONFIG_NET_DMA 5471 if (copied_early) 5472 __skb_queue_tail(&sk->sk_async_wait_queue, skb); 5473 else 5474 #endif 5475 if (eaten) 5476 __kfree_skb(skb); 5477 else 5478 sk->sk_data_ready(sk, 0); 5479 return 0; 5480 } 5481 } 5482 5483 slow_path: 5484 if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb)) 5485 goto csum_error; 5486 5487 /* 5488 * Standard slow path. 5489 */ 5490 5491 res = tcp_validate_incoming(sk, skb, th, 1); 5492 if (res <= 0) 5493 return -res; 5494 5495 step5: 5496 if (th->ack && tcp_ack(sk, skb, FLAG_SLOWPATH) < 0) 5497 goto discard; 5498 5499 tcp_rcv_rtt_measure_ts(sk, skb); 5500 5501 /* Process urgent data. */ 5502 tcp_urg(sk, skb, th); 5503 5504 /* step 7: process the segment text */ 5505 tcp_data_queue(sk, skb); 5506 5507 tcp_data_snd_check(sk); 5508 tcp_ack_snd_check(sk); 5509 return 0; 5510 5511 csum_error: 5512 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS); 5513 5514 discard: 5515 __kfree_skb(skb); 5516 return 0; 5517 } 5518 EXPORT_SYMBOL(tcp_rcv_established); 5519 5520 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb, 5521 const struct tcphdr *th, unsigned int len) 5522 { 5523 const u8 *hash_location; 5524 struct inet_connection_sock *icsk = inet_csk(sk); 5525 struct tcp_sock *tp = tcp_sk(sk); 5526 struct tcp_cookie_values *cvp = tp->cookie_values; 5527 int saved_clamp = tp->rx_opt.mss_clamp; 5528 5529 tcp_parse_options(skb, &tp->rx_opt, &hash_location, 0); 5530 5531 if (th->ack) { 5532 /* rfc793: 5533 * "If the state is SYN-SENT then 5534 * first check the ACK bit 5535 * If the ACK bit is set 5536 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send 5537 * a reset (unless the RST bit is set, if so drop 5538 * the segment and return)" 5539 * 5540 * We do not send data with SYN, so that RFC-correct 5541 * test reduces to: 5542 */ 5543 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt) 5544 goto reset_and_undo; 5545 5546 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 5547 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp, 5548 tcp_time_stamp)) { 5549 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED); 5550 goto reset_and_undo; 5551 } 5552 5553 /* Now ACK is acceptable. 5554 * 5555 * "If the RST bit is set 5556 * If the ACK was acceptable then signal the user "error: 5557 * connection reset", drop the segment, enter CLOSED state, 5558 * delete TCB, and return." 5559 */ 5560 5561 if (th->rst) { 5562 tcp_reset(sk); 5563 goto discard; 5564 } 5565 5566 /* rfc793: 5567 * "fifth, if neither of the SYN or RST bits is set then 5568 * drop the segment and return." 5569 * 5570 * See note below! 5571 * --ANK(990513) 5572 */ 5573 if (!th->syn) 5574 goto discard_and_undo; 5575 5576 /* rfc793: 5577 * "If the SYN bit is on ... 5578 * are acceptable then ... 5579 * (our SYN has been ACKed), change the connection 5580 * state to ESTABLISHED..." 5581 */ 5582 5583 TCP_ECN_rcv_synack(tp, th); 5584 5585 tp->snd_wl1 = TCP_SKB_CB(skb)->seq; 5586 tcp_ack(sk, skb, FLAG_SLOWPATH); 5587 5588 /* Ok.. it's good. Set up sequence numbers and 5589 * move to established. 5590 */ 5591 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; 5592 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; 5593 5594 /* RFC1323: The window in SYN & SYN/ACK segments is 5595 * never scaled. 5596 */ 5597 tp->snd_wnd = ntohs(th->window); 5598 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq); 5599 5600 if (!tp->rx_opt.wscale_ok) { 5601 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0; 5602 tp->window_clamp = min(tp->window_clamp, 65535U); 5603 } 5604 5605 if (tp->rx_opt.saw_tstamp) { 5606 tp->rx_opt.tstamp_ok = 1; 5607 tp->tcp_header_len = 5608 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 5609 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; 5610 tcp_store_ts_recent(tp); 5611 } else { 5612 tp->tcp_header_len = sizeof(struct tcphdr); 5613 } 5614 5615 if (tcp_is_sack(tp) && sysctl_tcp_fack) 5616 tcp_enable_fack(tp); 5617 5618 tcp_mtup_init(sk); 5619 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 5620 tcp_initialize_rcv_mss(sk); 5621 5622 /* Remember, tcp_poll() does not lock socket! 5623 * Change state from SYN-SENT only after copied_seq 5624 * is initialized. */ 5625 tp->copied_seq = tp->rcv_nxt; 5626 5627 if (cvp != NULL && 5628 cvp->cookie_pair_size > 0 && 5629 tp->rx_opt.cookie_plus > 0) { 5630 int cookie_size = tp->rx_opt.cookie_plus 5631 - TCPOLEN_COOKIE_BASE; 5632 int cookie_pair_size = cookie_size 5633 + cvp->cookie_desired; 5634 5635 /* A cookie extension option was sent and returned. 5636 * Note that each incoming SYNACK replaces the 5637 * Responder cookie. The initial exchange is most 5638 * fragile, as protection against spoofing relies 5639 * entirely upon the sequence and timestamp (above). 5640 * This replacement strategy allows the correct pair to 5641 * pass through, while any others will be filtered via 5642 * Responder verification later. 5643 */ 5644 if (sizeof(cvp->cookie_pair) >= cookie_pair_size) { 5645 memcpy(&cvp->cookie_pair[cvp->cookie_desired], 5646 hash_location, cookie_size); 5647 cvp->cookie_pair_size = cookie_pair_size; 5648 } 5649 } 5650 5651 smp_mb(); 5652 tcp_set_state(sk, TCP_ESTABLISHED); 5653 5654 security_inet_conn_established(sk, skb); 5655 5656 /* Make sure socket is routed, for correct metrics. */ 5657 icsk->icsk_af_ops->rebuild_header(sk); 5658 5659 tcp_init_metrics(sk); 5660 5661 tcp_init_congestion_control(sk); 5662 5663 /* Prevent spurious tcp_cwnd_restart() on first data 5664 * packet. 5665 */ 5666 tp->lsndtime = tcp_time_stamp; 5667 5668 tcp_init_buffer_space(sk); 5669 5670 if (sock_flag(sk, SOCK_KEEPOPEN)) 5671 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp)); 5672 5673 if (!tp->rx_opt.snd_wscale) 5674 __tcp_fast_path_on(tp, tp->snd_wnd); 5675 else 5676 tp->pred_flags = 0; 5677 5678 if (!sock_flag(sk, SOCK_DEAD)) { 5679 sk->sk_state_change(sk); 5680 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT); 5681 } 5682 5683 if (sk->sk_write_pending || 5684 icsk->icsk_accept_queue.rskq_defer_accept || 5685 icsk->icsk_ack.pingpong) { 5686 /* Save one ACK. Data will be ready after 5687 * several ticks, if write_pending is set. 5688 * 5689 * It may be deleted, but with this feature tcpdumps 5690 * look so _wonderfully_ clever, that I was not able 5691 * to stand against the temptation 8) --ANK 5692 */ 5693 inet_csk_schedule_ack(sk); 5694 icsk->icsk_ack.lrcvtime = tcp_time_stamp; 5695 icsk->icsk_ack.ato = TCP_ATO_MIN; 5696 tcp_incr_quickack(sk); 5697 tcp_enter_quickack_mode(sk); 5698 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 5699 TCP_DELACK_MAX, TCP_RTO_MAX); 5700 5701 discard: 5702 __kfree_skb(skb); 5703 return 0; 5704 } else { 5705 tcp_send_ack(sk); 5706 } 5707 return -1; 5708 } 5709 5710 /* No ACK in the segment */ 5711 5712 if (th->rst) { 5713 /* rfc793: 5714 * "If the RST bit is set 5715 * 5716 * Otherwise (no ACK) drop the segment and return." 5717 */ 5718 5719 goto discard_and_undo; 5720 } 5721 5722 /* PAWS check. */ 5723 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && 5724 tcp_paws_reject(&tp->rx_opt, 0)) 5725 goto discard_and_undo; 5726 5727 if (th->syn) { 5728 /* We see SYN without ACK. It is attempt of 5729 * simultaneous connect with crossed SYNs. 5730 * Particularly, it can be connect to self. 5731 */ 5732 tcp_set_state(sk, TCP_SYN_RECV); 5733 5734 if (tp->rx_opt.saw_tstamp) { 5735 tp->rx_opt.tstamp_ok = 1; 5736 tcp_store_ts_recent(tp); 5737 tp->tcp_header_len = 5738 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 5739 } else { 5740 tp->tcp_header_len = sizeof(struct tcphdr); 5741 } 5742 5743 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; 5744 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; 5745 5746 /* RFC1323: The window in SYN & SYN/ACK segments is 5747 * never scaled. 5748 */ 5749 tp->snd_wnd = ntohs(th->window); 5750 tp->snd_wl1 = TCP_SKB_CB(skb)->seq; 5751 tp->max_window = tp->snd_wnd; 5752 5753 TCP_ECN_rcv_syn(tp, th); 5754 5755 tcp_mtup_init(sk); 5756 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 5757 tcp_initialize_rcv_mss(sk); 5758 5759 tcp_send_synack(sk); 5760 #if 0 5761 /* Note, we could accept data and URG from this segment. 5762 * There are no obstacles to make this. 5763 * 5764 * However, if we ignore data in ACKless segments sometimes, 5765 * we have no reasons to accept it sometimes. 5766 * Also, seems the code doing it in step6 of tcp_rcv_state_process 5767 * is not flawless. So, discard packet for sanity. 5768 * Uncomment this return to process the data. 5769 */ 5770 return -1; 5771 #else 5772 goto discard; 5773 #endif 5774 } 5775 /* "fifth, if neither of the SYN or RST bits is set then 5776 * drop the segment and return." 5777 */ 5778 5779 discard_and_undo: 5780 tcp_clear_options(&tp->rx_opt); 5781 tp->rx_opt.mss_clamp = saved_clamp; 5782 goto discard; 5783 5784 reset_and_undo: 5785 tcp_clear_options(&tp->rx_opt); 5786 tp->rx_opt.mss_clamp = saved_clamp; 5787 return 1; 5788 } 5789 5790 /* 5791 * This function implements the receiving procedure of RFC 793 for 5792 * all states except ESTABLISHED and TIME_WAIT. 5793 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be 5794 * address independent. 5795 */ 5796 5797 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb, 5798 const struct tcphdr *th, unsigned int len) 5799 { 5800 struct tcp_sock *tp = tcp_sk(sk); 5801 struct inet_connection_sock *icsk = inet_csk(sk); 5802 int queued = 0; 5803 int res; 5804 5805 tp->rx_opt.saw_tstamp = 0; 5806 5807 switch (sk->sk_state) { 5808 case TCP_CLOSE: 5809 goto discard; 5810 5811 case TCP_LISTEN: 5812 if (th->ack) 5813 return 1; 5814 5815 if (th->rst) 5816 goto discard; 5817 5818 if (th->syn) { 5819 if (th->fin) 5820 goto discard; 5821 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0) 5822 return 1; 5823 5824 /* Now we have several options: In theory there is 5825 * nothing else in the frame. KA9Q has an option to 5826 * send data with the syn, BSD accepts data with the 5827 * syn up to the [to be] advertised window and 5828 * Solaris 2.1 gives you a protocol error. For now 5829 * we just ignore it, that fits the spec precisely 5830 * and avoids incompatibilities. It would be nice in 5831 * future to drop through and process the data. 5832 * 5833 * Now that TTCP is starting to be used we ought to 5834 * queue this data. 5835 * But, this leaves one open to an easy denial of 5836 * service attack, and SYN cookies can't defend 5837 * against this problem. So, we drop the data 5838 * in the interest of security over speed unless 5839 * it's still in use. 5840 */ 5841 kfree_skb(skb); 5842 return 0; 5843 } 5844 goto discard; 5845 5846 case TCP_SYN_SENT: 5847 queued = tcp_rcv_synsent_state_process(sk, skb, th, len); 5848 if (queued >= 0) 5849 return queued; 5850 5851 /* Do step6 onward by hand. */ 5852 tcp_urg(sk, skb, th); 5853 __kfree_skb(skb); 5854 tcp_data_snd_check(sk); 5855 return 0; 5856 } 5857 5858 res = tcp_validate_incoming(sk, skb, th, 0); 5859 if (res <= 0) 5860 return -res; 5861 5862 /* step 5: check the ACK field */ 5863 if (th->ack) { 5864 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH) > 0; 5865 5866 switch (sk->sk_state) { 5867 case TCP_SYN_RECV: 5868 if (acceptable) { 5869 tp->copied_seq = tp->rcv_nxt; 5870 smp_mb(); 5871 tcp_set_state(sk, TCP_ESTABLISHED); 5872 sk->sk_state_change(sk); 5873 5874 /* Note, that this wakeup is only for marginal 5875 * crossed SYN case. Passively open sockets 5876 * are not waked up, because sk->sk_sleep == 5877 * NULL and sk->sk_socket == NULL. 5878 */ 5879 if (sk->sk_socket) 5880 sk_wake_async(sk, 5881 SOCK_WAKE_IO, POLL_OUT); 5882 5883 tp->snd_una = TCP_SKB_CB(skb)->ack_seq; 5884 tp->snd_wnd = ntohs(th->window) << 5885 tp->rx_opt.snd_wscale; 5886 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq); 5887 5888 if (tp->rx_opt.tstamp_ok) 5889 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; 5890 5891 /* Make sure socket is routed, for 5892 * correct metrics. 5893 */ 5894 icsk->icsk_af_ops->rebuild_header(sk); 5895 5896 tcp_init_metrics(sk); 5897 5898 tcp_init_congestion_control(sk); 5899 5900 /* Prevent spurious tcp_cwnd_restart() on 5901 * first data packet. 5902 */ 5903 tp->lsndtime = tcp_time_stamp; 5904 5905 tcp_mtup_init(sk); 5906 tcp_initialize_rcv_mss(sk); 5907 tcp_init_buffer_space(sk); 5908 tcp_fast_path_on(tp); 5909 } else { 5910 return 1; 5911 } 5912 break; 5913 5914 case TCP_FIN_WAIT1: 5915 if (tp->snd_una == tp->write_seq) { 5916 tcp_set_state(sk, TCP_FIN_WAIT2); 5917 sk->sk_shutdown |= SEND_SHUTDOWN; 5918 dst_confirm(__sk_dst_get(sk)); 5919 5920 if (!sock_flag(sk, SOCK_DEAD)) 5921 /* Wake up lingering close() */ 5922 sk->sk_state_change(sk); 5923 else { 5924 int tmo; 5925 5926 if (tp->linger2 < 0 || 5927 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 5928 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) { 5929 tcp_done(sk); 5930 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA); 5931 return 1; 5932 } 5933 5934 tmo = tcp_fin_time(sk); 5935 if (tmo > TCP_TIMEWAIT_LEN) { 5936 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN); 5937 } else if (th->fin || sock_owned_by_user(sk)) { 5938 /* Bad case. We could lose such FIN otherwise. 5939 * It is not a big problem, but it looks confusing 5940 * and not so rare event. We still can lose it now, 5941 * if it spins in bh_lock_sock(), but it is really 5942 * marginal case. 5943 */ 5944 inet_csk_reset_keepalive_timer(sk, tmo); 5945 } else { 5946 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo); 5947 goto discard; 5948 } 5949 } 5950 } 5951 break; 5952 5953 case TCP_CLOSING: 5954 if (tp->snd_una == tp->write_seq) { 5955 tcp_time_wait(sk, TCP_TIME_WAIT, 0); 5956 goto discard; 5957 } 5958 break; 5959 5960 case TCP_LAST_ACK: 5961 if (tp->snd_una == tp->write_seq) { 5962 tcp_update_metrics(sk); 5963 tcp_done(sk); 5964 goto discard; 5965 } 5966 break; 5967 } 5968 } else 5969 goto discard; 5970 5971 /* step 6: check the URG bit */ 5972 tcp_urg(sk, skb, th); 5973 5974 /* step 7: process the segment text */ 5975 switch (sk->sk_state) { 5976 case TCP_CLOSE_WAIT: 5977 case TCP_CLOSING: 5978 case TCP_LAST_ACK: 5979 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) 5980 break; 5981 case TCP_FIN_WAIT1: 5982 case TCP_FIN_WAIT2: 5983 /* RFC 793 says to queue data in these states, 5984 * RFC 1122 says we MUST send a reset. 5985 * BSD 4.4 also does reset. 5986 */ 5987 if (sk->sk_shutdown & RCV_SHUTDOWN) { 5988 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 5989 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) { 5990 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA); 5991 tcp_reset(sk); 5992 return 1; 5993 } 5994 } 5995 /* Fall through */ 5996 case TCP_ESTABLISHED: 5997 tcp_data_queue(sk, skb); 5998 queued = 1; 5999 break; 6000 } 6001 6002 /* tcp_data could move socket to TIME-WAIT */ 6003 if (sk->sk_state != TCP_CLOSE) { 6004 tcp_data_snd_check(sk); 6005 tcp_ack_snd_check(sk); 6006 } 6007 6008 if (!queued) { 6009 discard: 6010 __kfree_skb(skb); 6011 } 6012 return 0; 6013 } 6014 EXPORT_SYMBOL(tcp_rcv_state_process); 6015