tcp_input.c revision 712a72213fad36cc9e6ec706b5e020d7eb6e03bc
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#define pr_fmt(fmt) "TCP: " fmt 65 66#include <linux/mm.h> 67#include <linux/slab.h> 68#include <linux/module.h> 69#include <linux/sysctl.h> 70#include <linux/kernel.h> 71#include <net/dst.h> 72#include <net/tcp.h> 73#include <net/inet_common.h> 74#include <linux/ipsec.h> 75#include <asm/unaligned.h> 76#include <net/netdma.h> 77#include <linux/errqueue.h> 78 79int sysctl_tcp_timestamps __read_mostly = 1; 80int sysctl_tcp_window_scaling __read_mostly = 1; 81int sysctl_tcp_sack __read_mostly = 1; 82int sysctl_tcp_fack __read_mostly = 1; 83int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH; 84EXPORT_SYMBOL(sysctl_tcp_reordering); 85int sysctl_tcp_dsack __read_mostly = 1; 86int sysctl_tcp_app_win __read_mostly = 31; 87int sysctl_tcp_adv_win_scale __read_mostly = 1; 88EXPORT_SYMBOL(sysctl_tcp_adv_win_scale); 89 90/* rfc5961 challenge ack rate limiting */ 91int sysctl_tcp_challenge_ack_limit = 100; 92 93int sysctl_tcp_stdurg __read_mostly; 94int sysctl_tcp_rfc1337 __read_mostly; 95int sysctl_tcp_max_orphans __read_mostly = NR_FILE; 96int sysctl_tcp_frto __read_mostly = 2; 97 98int sysctl_tcp_thin_dupack __read_mostly; 99 100int sysctl_tcp_moderate_rcvbuf __read_mostly = 1; 101int sysctl_tcp_early_retrans __read_mostly = 3; 102 103#define FLAG_DATA 0x01 /* Incoming frame contained data. */ 104#define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */ 105#define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */ 106#define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */ 107#define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */ 108#define FLAG_DATA_SACKED 0x20 /* New SACK. */ 109#define FLAG_ECE 0x40 /* ECE in this ACK */ 110#define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/ 111#define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */ 112#define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */ 113#define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */ 114#define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */ 115#define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */ 116 117#define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED) 118#define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED) 119#define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE) 120#define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED) 121 122#define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH) 123#define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH)) 124 125/* Adapt the MSS value used to make delayed ack decision to the 126 * real world. 127 */ 128static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb) 129{ 130 struct inet_connection_sock *icsk = inet_csk(sk); 131 const unsigned int lss = icsk->icsk_ack.last_seg_size; 132 unsigned int len; 133 134 icsk->icsk_ack.last_seg_size = 0; 135 136 /* skb->len may jitter because of SACKs, even if peer 137 * sends good full-sized frames. 138 */ 139 len = skb_shinfo(skb)->gso_size ? : skb->len; 140 if (len >= icsk->icsk_ack.rcv_mss) { 141 icsk->icsk_ack.rcv_mss = len; 142 } else { 143 /* Otherwise, we make more careful check taking into account, 144 * that SACKs block is variable. 145 * 146 * "len" is invariant segment length, including TCP header. 147 */ 148 len += skb->data - skb_transport_header(skb); 149 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) || 150 /* If PSH is not set, packet should be 151 * full sized, provided peer TCP is not badly broken. 152 * This observation (if it is correct 8)) allows 153 * to handle super-low mtu links fairly. 154 */ 155 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) && 156 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) { 157 /* Subtract also invariant (if peer is RFC compliant), 158 * tcp header plus fixed timestamp option length. 159 * Resulting "len" is MSS free of SACK jitter. 160 */ 161 len -= tcp_sk(sk)->tcp_header_len; 162 icsk->icsk_ack.last_seg_size = len; 163 if (len == lss) { 164 icsk->icsk_ack.rcv_mss = len; 165 return; 166 } 167 } 168 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED) 169 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2; 170 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED; 171 } 172} 173 174static void tcp_incr_quickack(struct sock *sk) 175{ 176 struct inet_connection_sock *icsk = inet_csk(sk); 177 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss); 178 179 if (quickacks == 0) 180 quickacks = 2; 181 if (quickacks > icsk->icsk_ack.quick) 182 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS); 183} 184 185static void tcp_enter_quickack_mode(struct sock *sk) 186{ 187 struct inet_connection_sock *icsk = inet_csk(sk); 188 tcp_incr_quickack(sk); 189 icsk->icsk_ack.pingpong = 0; 190 icsk->icsk_ack.ato = TCP_ATO_MIN; 191} 192 193/* Send ACKs quickly, if "quick" count is not exhausted 194 * and the session is not interactive. 195 */ 196 197static inline bool tcp_in_quickack_mode(const struct sock *sk) 198{ 199 const struct inet_connection_sock *icsk = inet_csk(sk); 200 201 return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong; 202} 203 204static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp) 205{ 206 if (tp->ecn_flags & TCP_ECN_OK) 207 tp->ecn_flags |= TCP_ECN_QUEUE_CWR; 208} 209 210static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb) 211{ 212 if (tcp_hdr(skb)->cwr) 213 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR; 214} 215 216static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp) 217{ 218 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR; 219} 220 221static inline void TCP_ECN_check_ce(struct tcp_sock *tp, const struct sk_buff *skb) 222{ 223 if (!(tp->ecn_flags & TCP_ECN_OK)) 224 return; 225 226 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) { 227 case INET_ECN_NOT_ECT: 228 /* Funny extension: if ECT is not set on a segment, 229 * and we already seen ECT on a previous segment, 230 * it is probably a retransmit. 231 */ 232 if (tp->ecn_flags & TCP_ECN_SEEN) 233 tcp_enter_quickack_mode((struct sock *)tp); 234 break; 235 case INET_ECN_CE: 236 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) { 237 /* Better not delay acks, sender can have a very low cwnd */ 238 tcp_enter_quickack_mode((struct sock *)tp); 239 tp->ecn_flags |= TCP_ECN_DEMAND_CWR; 240 } 241 /* fallinto */ 242 default: 243 tp->ecn_flags |= TCP_ECN_SEEN; 244 } 245} 246 247static inline void TCP_ECN_rcv_synack(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 253static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th) 254{ 255 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr)) 256 tp->ecn_flags &= ~TCP_ECN_OK; 257} 258 259static bool TCP_ECN_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th) 260{ 261 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK)) 262 return true; 263 return false; 264} 265 266/* Buffer size and advertised window tuning. 267 * 268 * 1. Tuning sk->sk_sndbuf, when connection enters established state. 269 */ 270 271static void tcp_sndbuf_expand(struct sock *sk) 272{ 273 const struct tcp_sock *tp = tcp_sk(sk); 274 int sndmem, per_mss; 275 u32 nr_segs; 276 277 /* Worst case is non GSO/TSO : each frame consumes one skb 278 * and skb->head is kmalloced using power of two area of memory 279 */ 280 per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) + 281 MAX_TCP_HEADER + 282 SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); 283 284 per_mss = roundup_pow_of_two(per_mss) + 285 SKB_DATA_ALIGN(sizeof(struct sk_buff)); 286 287 nr_segs = max_t(u32, TCP_INIT_CWND, tp->snd_cwnd); 288 nr_segs = max_t(u32, nr_segs, tp->reordering + 1); 289 290 /* Fast Recovery (RFC 5681 3.2) : 291 * Cubic needs 1.7 factor, rounded to 2 to include 292 * extra cushion (application might react slowly to POLLOUT) 293 */ 294 sndmem = 2 * nr_segs * per_mss; 295 296 if (sk->sk_sndbuf < sndmem) 297 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]); 298} 299 300/* 2. Tuning advertised window (window_clamp, rcv_ssthresh) 301 * 302 * All tcp_full_space() is split to two parts: "network" buffer, allocated 303 * forward and advertised in receiver window (tp->rcv_wnd) and 304 * "application buffer", required to isolate scheduling/application 305 * latencies from network. 306 * window_clamp is maximal advertised window. It can be less than 307 * tcp_full_space(), in this case tcp_full_space() - window_clamp 308 * is reserved for "application" buffer. The less window_clamp is 309 * the smoother our behaviour from viewpoint of network, but the lower 310 * throughput and the higher sensitivity of the connection to losses. 8) 311 * 312 * rcv_ssthresh is more strict window_clamp used at "slow start" 313 * phase to predict further behaviour of this connection. 314 * It is used for two goals: 315 * - to enforce header prediction at sender, even when application 316 * requires some significant "application buffer". It is check #1. 317 * - to prevent pruning of receive queue because of misprediction 318 * of receiver window. Check #2. 319 * 320 * The scheme does not work when sender sends good segments opening 321 * window and then starts to feed us spaghetti. But it should work 322 * in common situations. Otherwise, we have to rely on queue collapsing. 323 */ 324 325/* Slow part of check#2. */ 326static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb) 327{ 328 struct tcp_sock *tp = tcp_sk(sk); 329 /* Optimize this! */ 330 int truesize = tcp_win_from_space(skb->truesize) >> 1; 331 int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1; 332 333 while (tp->rcv_ssthresh <= window) { 334 if (truesize <= skb->len) 335 return 2 * inet_csk(sk)->icsk_ack.rcv_mss; 336 337 truesize >>= 1; 338 window >>= 1; 339 } 340 return 0; 341} 342 343static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb) 344{ 345 struct tcp_sock *tp = tcp_sk(sk); 346 347 /* Check #1 */ 348 if (tp->rcv_ssthresh < tp->window_clamp && 349 (int)tp->rcv_ssthresh < tcp_space(sk) && 350 !sk_under_memory_pressure(sk)) { 351 int incr; 352 353 /* Check #2. Increase window, if skb with such overhead 354 * will fit to rcvbuf in future. 355 */ 356 if (tcp_win_from_space(skb->truesize) <= skb->len) 357 incr = 2 * tp->advmss; 358 else 359 incr = __tcp_grow_window(sk, skb); 360 361 if (incr) { 362 incr = max_t(int, incr, 2 * skb->len); 363 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr, 364 tp->window_clamp); 365 inet_csk(sk)->icsk_ack.quick |= 1; 366 } 367 } 368} 369 370/* 3. Tuning rcvbuf, when connection enters established state. */ 371static void tcp_fixup_rcvbuf(struct sock *sk) 372{ 373 u32 mss = tcp_sk(sk)->advmss; 374 int rcvmem; 375 376 rcvmem = 2 * SKB_TRUESIZE(mss + MAX_TCP_HEADER) * 377 tcp_default_init_rwnd(mss); 378 379 /* Dynamic Right Sizing (DRS) has 2 to 3 RTT latency 380 * Allow enough cushion so that sender is not limited by our window 381 */ 382 if (sysctl_tcp_moderate_rcvbuf) 383 rcvmem <<= 2; 384 385 if (sk->sk_rcvbuf < rcvmem) 386 sk->sk_rcvbuf = min(rcvmem, sysctl_tcp_rmem[2]); 387} 388 389/* 4. Try to fixup all. It is made immediately after connection enters 390 * established state. 391 */ 392void tcp_init_buffer_space(struct sock *sk) 393{ 394 struct tcp_sock *tp = tcp_sk(sk); 395 int maxwin; 396 397 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) 398 tcp_fixup_rcvbuf(sk); 399 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK)) 400 tcp_sndbuf_expand(sk); 401 402 tp->rcvq_space.space = tp->rcv_wnd; 403 tp->rcvq_space.time = tcp_time_stamp; 404 tp->rcvq_space.seq = tp->copied_seq; 405 406 maxwin = tcp_full_space(sk); 407 408 if (tp->window_clamp >= maxwin) { 409 tp->window_clamp = maxwin; 410 411 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss) 412 tp->window_clamp = max(maxwin - 413 (maxwin >> sysctl_tcp_app_win), 414 4 * tp->advmss); 415 } 416 417 /* Force reservation of one segment. */ 418 if (sysctl_tcp_app_win && 419 tp->window_clamp > 2 * tp->advmss && 420 tp->window_clamp + tp->advmss > maxwin) 421 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss); 422 423 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp); 424 tp->snd_cwnd_stamp = tcp_time_stamp; 425} 426 427/* 5. Recalculate window clamp after socket hit its memory bounds. */ 428static void tcp_clamp_window(struct sock *sk) 429{ 430 struct tcp_sock *tp = tcp_sk(sk); 431 struct inet_connection_sock *icsk = inet_csk(sk); 432 433 icsk->icsk_ack.quick = 0; 434 435 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] && 436 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) && 437 !sk_under_memory_pressure(sk) && 438 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) { 439 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc), 440 sysctl_tcp_rmem[2]); 441 } 442 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf) 443 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss); 444} 445 446/* Initialize RCV_MSS value. 447 * RCV_MSS is an our guess about MSS used by the peer. 448 * We haven't any direct information about the MSS. 449 * It's better to underestimate the RCV_MSS rather than overestimate. 450 * Overestimations make us ACKing less frequently than needed. 451 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss(). 452 */ 453void tcp_initialize_rcv_mss(struct sock *sk) 454{ 455 const struct tcp_sock *tp = tcp_sk(sk); 456 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache); 457 458 hint = min(hint, tp->rcv_wnd / 2); 459 hint = min(hint, TCP_MSS_DEFAULT); 460 hint = max(hint, TCP_MIN_MSS); 461 462 inet_csk(sk)->icsk_ack.rcv_mss = hint; 463} 464EXPORT_SYMBOL(tcp_initialize_rcv_mss); 465 466/* Receiver "autotuning" code. 467 * 468 * The algorithm for RTT estimation w/o timestamps is based on 469 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL. 470 * <http://public.lanl.gov/radiant/pubs.html#DRS> 471 * 472 * More detail on this code can be found at 473 * <http://staff.psc.edu/jheffner/>, 474 * though this reference is out of date. A new paper 475 * is pending. 476 */ 477static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep) 478{ 479 u32 new_sample = tp->rcv_rtt_est.rtt; 480 long m = sample; 481 482 if (m == 0) 483 m = 1; 484 485 if (new_sample != 0) { 486 /* If we sample in larger samples in the non-timestamp 487 * case, we could grossly overestimate the RTT especially 488 * with chatty applications or bulk transfer apps which 489 * are stalled on filesystem I/O. 490 * 491 * Also, since we are only going for a minimum in the 492 * non-timestamp case, we do not smooth things out 493 * else with timestamps disabled convergence takes too 494 * long. 495 */ 496 if (!win_dep) { 497 m -= (new_sample >> 3); 498 new_sample += m; 499 } else { 500 m <<= 3; 501 if (m < new_sample) 502 new_sample = m; 503 } 504 } else { 505 /* No previous measure. */ 506 new_sample = m << 3; 507 } 508 509 if (tp->rcv_rtt_est.rtt != new_sample) 510 tp->rcv_rtt_est.rtt = new_sample; 511} 512 513static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp) 514{ 515 if (tp->rcv_rtt_est.time == 0) 516 goto new_measure; 517 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq)) 518 return; 519 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rcv_rtt_est.time, 1); 520 521new_measure: 522 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd; 523 tp->rcv_rtt_est.time = tcp_time_stamp; 524} 525 526static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, 527 const struct sk_buff *skb) 528{ 529 struct tcp_sock *tp = tcp_sk(sk); 530 if (tp->rx_opt.rcv_tsecr && 531 (TCP_SKB_CB(skb)->end_seq - 532 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss)) 533 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0); 534} 535 536/* 537 * This function should be called every time data is copied to user space. 538 * It calculates the appropriate TCP receive buffer space. 539 */ 540void tcp_rcv_space_adjust(struct sock *sk) 541{ 542 struct tcp_sock *tp = tcp_sk(sk); 543 int time; 544 int copied; 545 546 time = tcp_time_stamp - tp->rcvq_space.time; 547 if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0) 548 return; 549 550 /* Number of bytes copied to user in last RTT */ 551 copied = tp->copied_seq - tp->rcvq_space.seq; 552 if (copied <= tp->rcvq_space.space) 553 goto new_measure; 554 555 /* A bit of theory : 556 * copied = bytes received in previous RTT, our base window 557 * To cope with packet losses, we need a 2x factor 558 * To cope with slow start, and sender growing its cwin by 100 % 559 * every RTT, we need a 4x factor, because the ACK we are sending 560 * now is for the next RTT, not the current one : 561 * <prev RTT . ><current RTT .. ><next RTT .... > 562 */ 563 564 if (sysctl_tcp_moderate_rcvbuf && 565 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) { 566 int rcvwin, rcvmem, rcvbuf; 567 568 /* minimal window to cope with packet losses, assuming 569 * steady state. Add some cushion because of small variations. 570 */ 571 rcvwin = (copied << 1) + 16 * tp->advmss; 572 573 /* If rate increased by 25%, 574 * assume slow start, rcvwin = 3 * copied 575 * If rate increased by 50%, 576 * assume sender can use 2x growth, rcvwin = 4 * copied 577 */ 578 if (copied >= 579 tp->rcvq_space.space + (tp->rcvq_space.space >> 2)) { 580 if (copied >= 581 tp->rcvq_space.space + (tp->rcvq_space.space >> 1)) 582 rcvwin <<= 1; 583 else 584 rcvwin += (rcvwin >> 1); 585 } 586 587 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER); 588 while (tcp_win_from_space(rcvmem) < tp->advmss) 589 rcvmem += 128; 590 591 rcvbuf = min(rcvwin / tp->advmss * rcvmem, sysctl_tcp_rmem[2]); 592 if (rcvbuf > sk->sk_rcvbuf) { 593 sk->sk_rcvbuf = rcvbuf; 594 595 /* Make the window clamp follow along. */ 596 tp->window_clamp = rcvwin; 597 } 598 } 599 tp->rcvq_space.space = copied; 600 601new_measure: 602 tp->rcvq_space.seq = tp->copied_seq; 603 tp->rcvq_space.time = tcp_time_stamp; 604} 605 606/* There is something which you must keep in mind when you analyze the 607 * behavior of the tp->ato delayed ack timeout interval. When a 608 * connection starts up, we want to ack as quickly as possible. The 609 * problem is that "good" TCP's do slow start at the beginning of data 610 * transmission. The means that until we send the first few ACK's the 611 * sender will sit on his end and only queue most of his data, because 612 * he can only send snd_cwnd unacked packets at any given time. For 613 * each ACK we send, he increments snd_cwnd and transmits more of his 614 * queue. -DaveM 615 */ 616static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb) 617{ 618 struct tcp_sock *tp = tcp_sk(sk); 619 struct inet_connection_sock *icsk = inet_csk(sk); 620 u32 now; 621 622 inet_csk_schedule_ack(sk); 623 624 tcp_measure_rcv_mss(sk, skb); 625 626 tcp_rcv_rtt_measure(tp); 627 628 now = tcp_time_stamp; 629 630 if (!icsk->icsk_ack.ato) { 631 /* The _first_ data packet received, initialize 632 * delayed ACK engine. 633 */ 634 tcp_incr_quickack(sk); 635 icsk->icsk_ack.ato = TCP_ATO_MIN; 636 } else { 637 int m = now - icsk->icsk_ack.lrcvtime; 638 639 if (m <= TCP_ATO_MIN / 2) { 640 /* The fastest case is the first. */ 641 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2; 642 } else if (m < icsk->icsk_ack.ato) { 643 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m; 644 if (icsk->icsk_ack.ato > icsk->icsk_rto) 645 icsk->icsk_ack.ato = icsk->icsk_rto; 646 } else if (m > icsk->icsk_rto) { 647 /* Too long gap. Apparently sender failed to 648 * restart window, so that we send ACKs quickly. 649 */ 650 tcp_incr_quickack(sk); 651 sk_mem_reclaim(sk); 652 } 653 } 654 icsk->icsk_ack.lrcvtime = now; 655 656 TCP_ECN_check_ce(tp, skb); 657 658 if (skb->len >= 128) 659 tcp_grow_window(sk, skb); 660} 661 662/* Called to compute a smoothed rtt estimate. The data fed to this 663 * routine either comes from timestamps, or from segments that were 664 * known _not_ to have been retransmitted [see Karn/Partridge 665 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88 666 * piece by Van Jacobson. 667 * NOTE: the next three routines used to be one big routine. 668 * To save cycles in the RFC 1323 implementation it was better to break 669 * it up into three procedures. -- erics 670 */ 671static void tcp_rtt_estimator(struct sock *sk, long mrtt_us) 672{ 673 struct tcp_sock *tp = tcp_sk(sk); 674 long m = mrtt_us; /* RTT */ 675 u32 srtt = tp->srtt_us; 676 677 /* The following amusing code comes from Jacobson's 678 * article in SIGCOMM '88. Note that rtt and mdev 679 * are scaled versions of rtt and mean deviation. 680 * This is designed to be as fast as possible 681 * m stands for "measurement". 682 * 683 * On a 1990 paper the rto value is changed to: 684 * RTO = rtt + 4 * mdev 685 * 686 * Funny. This algorithm seems to be very broken. 687 * These formulae increase RTO, when it should be decreased, increase 688 * too slowly, when it should be increased quickly, decrease too quickly 689 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely 690 * does not matter how to _calculate_ it. Seems, it was trap 691 * that VJ failed to avoid. 8) 692 */ 693 if (srtt != 0) { 694 m -= (srtt >> 3); /* m is now error in rtt est */ 695 srtt += m; /* rtt = 7/8 rtt + 1/8 new */ 696 if (m < 0) { 697 m = -m; /* m is now abs(error) */ 698 m -= (tp->mdev_us >> 2); /* similar update on mdev */ 699 /* This is similar to one of Eifel findings. 700 * Eifel blocks mdev updates when rtt decreases. 701 * This solution is a bit different: we use finer gain 702 * for mdev in this case (alpha*beta). 703 * Like Eifel it also prevents growth of rto, 704 * but also it limits too fast rto decreases, 705 * happening in pure Eifel. 706 */ 707 if (m > 0) 708 m >>= 3; 709 } else { 710 m -= (tp->mdev_us >> 2); /* similar update on mdev */ 711 } 712 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */ 713 if (tp->mdev_us > tp->mdev_max_us) { 714 tp->mdev_max_us = tp->mdev_us; 715 if (tp->mdev_max_us > tp->rttvar_us) 716 tp->rttvar_us = tp->mdev_max_us; 717 } 718 if (after(tp->snd_una, tp->rtt_seq)) { 719 if (tp->mdev_max_us < tp->rttvar_us) 720 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2; 721 tp->rtt_seq = tp->snd_nxt; 722 tp->mdev_max_us = tcp_rto_min_us(sk); 723 } 724 } else { 725 /* no previous measure. */ 726 srtt = m << 3; /* take the measured time to be rtt */ 727 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */ 728 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk)); 729 tp->mdev_max_us = tp->rttvar_us; 730 tp->rtt_seq = tp->snd_nxt; 731 } 732 tp->srtt_us = max(1U, srtt); 733} 734 735/* Set the sk_pacing_rate to allow proper sizing of TSO packets. 736 * Note: TCP stack does not yet implement pacing. 737 * FQ packet scheduler can be used to implement cheap but effective 738 * TCP pacing, to smooth the burst on large writes when packets 739 * in flight is significantly lower than cwnd (or rwin) 740 */ 741static void tcp_update_pacing_rate(struct sock *sk) 742{ 743 const struct tcp_sock *tp = tcp_sk(sk); 744 u64 rate; 745 746 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */ 747 rate = (u64)tp->mss_cache * 2 * (USEC_PER_SEC << 3); 748 749 rate *= max(tp->snd_cwnd, tp->packets_out); 750 751 if (likely(tp->srtt_us)) 752 do_div(rate, tp->srtt_us); 753 754 /* ACCESS_ONCE() is needed because sch_fq fetches sk_pacing_rate 755 * without any lock. We want to make sure compiler wont store 756 * intermediate values in this location. 757 */ 758 ACCESS_ONCE(sk->sk_pacing_rate) = min_t(u64, rate, 759 sk->sk_max_pacing_rate); 760} 761 762/* Calculate rto without backoff. This is the second half of Van Jacobson's 763 * routine referred to above. 764 */ 765static void tcp_set_rto(struct sock *sk) 766{ 767 const struct tcp_sock *tp = tcp_sk(sk); 768 /* Old crap is replaced with new one. 8) 769 * 770 * More seriously: 771 * 1. If rtt variance happened to be less 50msec, it is hallucination. 772 * It cannot be less due to utterly erratic ACK generation made 773 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_ 774 * to do with delayed acks, because at cwnd>2 true delack timeout 775 * is invisible. Actually, Linux-2.4 also generates erratic 776 * ACKs in some circumstances. 777 */ 778 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp); 779 780 /* 2. Fixups made earlier cannot be right. 781 * If we do not estimate RTO correctly without them, 782 * all the algo is pure shit and should be replaced 783 * with correct one. It is exactly, which we pretend to do. 784 */ 785 786 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo 787 * guarantees that rto is higher. 788 */ 789 tcp_bound_rto(sk); 790} 791 792__u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst) 793{ 794 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0); 795 796 if (!cwnd) 797 cwnd = TCP_INIT_CWND; 798 return min_t(__u32, cwnd, tp->snd_cwnd_clamp); 799} 800 801/* 802 * Packet counting of FACK is based on in-order assumptions, therefore TCP 803 * disables it when reordering is detected 804 */ 805void tcp_disable_fack(struct tcp_sock *tp) 806{ 807 /* RFC3517 uses different metric in lost marker => reset on change */ 808 if (tcp_is_fack(tp)) 809 tp->lost_skb_hint = NULL; 810 tp->rx_opt.sack_ok &= ~TCP_FACK_ENABLED; 811} 812 813/* Take a notice that peer is sending D-SACKs */ 814static void tcp_dsack_seen(struct tcp_sock *tp) 815{ 816 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN; 817} 818 819static void tcp_update_reordering(struct sock *sk, const int metric, 820 const int ts) 821{ 822 struct tcp_sock *tp = tcp_sk(sk); 823 if (metric > tp->reordering) { 824 int mib_idx; 825 826 tp->reordering = min(TCP_MAX_REORDERING, metric); 827 828 /* This exciting event is worth to be remembered. 8) */ 829 if (ts) 830 mib_idx = LINUX_MIB_TCPTSREORDER; 831 else if (tcp_is_reno(tp)) 832 mib_idx = LINUX_MIB_TCPRENOREORDER; 833 else if (tcp_is_fack(tp)) 834 mib_idx = LINUX_MIB_TCPFACKREORDER; 835 else 836 mib_idx = LINUX_MIB_TCPSACKREORDER; 837 838 NET_INC_STATS_BH(sock_net(sk), mib_idx); 839#if FASTRETRANS_DEBUG > 1 840 pr_debug("Disorder%d %d %u f%u s%u rr%d\n", 841 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state, 842 tp->reordering, 843 tp->fackets_out, 844 tp->sacked_out, 845 tp->undo_marker ? tp->undo_retrans : 0); 846#endif 847 tcp_disable_fack(tp); 848 } 849 850 if (metric > 0) 851 tcp_disable_early_retrans(tp); 852} 853 854/* This must be called before lost_out is incremented */ 855static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb) 856{ 857 if ((tp->retransmit_skb_hint == NULL) || 858 before(TCP_SKB_CB(skb)->seq, 859 TCP_SKB_CB(tp->retransmit_skb_hint)->seq)) 860 tp->retransmit_skb_hint = skb; 861 862 if (!tp->lost_out || 863 after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high)) 864 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq; 865} 866 867static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb) 868{ 869 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) { 870 tcp_verify_retransmit_hint(tp, skb); 871 872 tp->lost_out += tcp_skb_pcount(skb); 873 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 874 } 875} 876 877static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp, 878 struct sk_buff *skb) 879{ 880 tcp_verify_retransmit_hint(tp, skb); 881 882 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) { 883 tp->lost_out += tcp_skb_pcount(skb); 884 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 885 } 886} 887 888/* This procedure tags the retransmission queue when SACKs arrive. 889 * 890 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L). 891 * Packets in queue with these bits set are counted in variables 892 * sacked_out, retrans_out and lost_out, correspondingly. 893 * 894 * Valid combinations are: 895 * Tag InFlight Description 896 * 0 1 - orig segment is in flight. 897 * S 0 - nothing flies, orig reached receiver. 898 * L 0 - nothing flies, orig lost by net. 899 * R 2 - both orig and retransmit are in flight. 900 * L|R 1 - orig is lost, retransmit is in flight. 901 * S|R 1 - orig reached receiver, retrans is still in flight. 902 * (L|S|R is logically valid, it could occur when L|R is sacked, 903 * but it is equivalent to plain S and code short-curcuits it to S. 904 * L|S is logically invalid, it would mean -1 packet in flight 8)) 905 * 906 * These 6 states form finite state machine, controlled by the following events: 907 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue()) 908 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue()) 909 * 3. Loss detection event of two flavors: 910 * A. Scoreboard estimator decided the packet is lost. 911 * A'. Reno "three dupacks" marks head of queue lost. 912 * A''. Its FACK modification, head until snd.fack is lost. 913 * B. SACK arrives sacking SND.NXT at the moment, when the 914 * segment was retransmitted. 915 * 4. D-SACK added new rule: D-SACK changes any tag to S. 916 * 917 * It is pleasant to note, that state diagram turns out to be commutative, 918 * so that we are allowed not to be bothered by order of our actions, 919 * when multiple events arrive simultaneously. (see the function below). 920 * 921 * Reordering detection. 922 * -------------------- 923 * Reordering metric is maximal distance, which a packet can be displaced 924 * in packet stream. With SACKs we can estimate it: 925 * 926 * 1. SACK fills old hole and the corresponding segment was not 927 * ever retransmitted -> reordering. Alas, we cannot use it 928 * when segment was retransmitted. 929 * 2. The last flaw is solved with D-SACK. D-SACK arrives 930 * for retransmitted and already SACKed segment -> reordering.. 931 * Both of these heuristics are not used in Loss state, when we cannot 932 * account for retransmits accurately. 933 * 934 * SACK block validation. 935 * ---------------------- 936 * 937 * SACK block range validation checks that the received SACK block fits to 938 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT. 939 * Note that SND.UNA is not included to the range though being valid because 940 * it means that the receiver is rather inconsistent with itself reporting 941 * SACK reneging when it should advance SND.UNA. Such SACK block this is 942 * perfectly valid, however, in light of RFC2018 which explicitly states 943 * that "SACK block MUST reflect the newest segment. Even if the newest 944 * segment is going to be discarded ...", not that it looks very clever 945 * in case of head skb. Due to potentional receiver driven attacks, we 946 * choose to avoid immediate execution of a walk in write queue due to 947 * reneging and defer head skb's loss recovery to standard loss recovery 948 * procedure that will eventually trigger (nothing forbids us doing this). 949 * 950 * Implements also blockage to start_seq wrap-around. Problem lies in the 951 * fact that though start_seq (s) is before end_seq (i.e., not reversed), 952 * there's no guarantee that it will be before snd_nxt (n). The problem 953 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt 954 * wrap (s_w): 955 * 956 * <- outs wnd -> <- wrapzone -> 957 * u e n u_w e_w s n_w 958 * | | | | | | | 959 * |<------------+------+----- TCP seqno space --------------+---------->| 960 * ...-- <2^31 ->| |<--------... 961 * ...---- >2^31 ------>| |<--------... 962 * 963 * Current code wouldn't be vulnerable but it's better still to discard such 964 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat 965 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in 966 * snd_nxt wrap -> snd_una region will then become "well defined", i.e., 967 * equal to the ideal case (infinite seqno space without wrap caused issues). 968 * 969 * With D-SACK the lower bound is extended to cover sequence space below 970 * SND.UNA down to undo_marker, which is the last point of interest. Yet 971 * again, D-SACK block must not to go across snd_una (for the same reason as 972 * for the normal SACK blocks, explained above). But there all simplicity 973 * ends, TCP might receive valid D-SACKs below that. As long as they reside 974 * fully below undo_marker they do not affect behavior in anyway and can 975 * therefore be safely ignored. In rare cases (which are more or less 976 * theoretical ones), the D-SACK will nicely cross that boundary due to skb 977 * fragmentation and packet reordering past skb's retransmission. To consider 978 * them correctly, the acceptable range must be extended even more though 979 * the exact amount is rather hard to quantify. However, tp->max_window can 980 * be used as an exaggerated estimate. 981 */ 982static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack, 983 u32 start_seq, u32 end_seq) 984{ 985 /* Too far in future, or reversed (interpretation is ambiguous) */ 986 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq)) 987 return false; 988 989 /* Nasty start_seq wrap-around check (see comments above) */ 990 if (!before(start_seq, tp->snd_nxt)) 991 return false; 992 993 /* In outstanding window? ...This is valid exit for D-SACKs too. 994 * start_seq == snd_una is non-sensical (see comments above) 995 */ 996 if (after(start_seq, tp->snd_una)) 997 return true; 998 999 if (!is_dsack || !tp->undo_marker) 1000 return false; 1001 1002 /* ...Then it's D-SACK, and must reside below snd_una completely */ 1003 if (after(end_seq, tp->snd_una)) 1004 return false; 1005 1006 if (!before(start_seq, tp->undo_marker)) 1007 return true; 1008 1009 /* Too old */ 1010 if (!after(end_seq, tp->undo_marker)) 1011 return false; 1012 1013 /* Undo_marker boundary crossing (overestimates a lot). Known already: 1014 * start_seq < undo_marker and end_seq >= undo_marker. 1015 */ 1016 return !before(start_seq, end_seq - tp->max_window); 1017} 1018 1019/* Check for lost retransmit. This superb idea is borrowed from "ratehalving". 1020 * Event "B". Later note: FACK people cheated me again 8), we have to account 1021 * for reordering! Ugly, but should help. 1022 * 1023 * Search retransmitted skbs from write_queue that were sent when snd_nxt was 1024 * less than what is now known to be received by the other end (derived from 1025 * highest SACK block). Also calculate the lowest snd_nxt among the remaining 1026 * retransmitted skbs to avoid some costly processing per ACKs. 1027 */ 1028static void tcp_mark_lost_retrans(struct sock *sk) 1029{ 1030 const struct inet_connection_sock *icsk = inet_csk(sk); 1031 struct tcp_sock *tp = tcp_sk(sk); 1032 struct sk_buff *skb; 1033 int cnt = 0; 1034 u32 new_low_seq = tp->snd_nxt; 1035 u32 received_upto = tcp_highest_sack_seq(tp); 1036 1037 if (!tcp_is_fack(tp) || !tp->retrans_out || 1038 !after(received_upto, tp->lost_retrans_low) || 1039 icsk->icsk_ca_state != TCP_CA_Recovery) 1040 return; 1041 1042 tcp_for_write_queue(skb, sk) { 1043 u32 ack_seq = TCP_SKB_CB(skb)->ack_seq; 1044 1045 if (skb == tcp_send_head(sk)) 1046 break; 1047 if (cnt == tp->retrans_out) 1048 break; 1049 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 1050 continue; 1051 1052 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)) 1053 continue; 1054 1055 /* TODO: We would like to get rid of tcp_is_fack(tp) only 1056 * constraint here (see above) but figuring out that at 1057 * least tp->reordering SACK blocks reside between ack_seq 1058 * and received_upto is not easy task to do cheaply with 1059 * the available datastructures. 1060 * 1061 * Whether FACK should check here for tp->reordering segs 1062 * in-between one could argue for either way (it would be 1063 * rather simple to implement as we could count fack_count 1064 * during the walk and do tp->fackets_out - fack_count). 1065 */ 1066 if (after(received_upto, ack_seq)) { 1067 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 1068 tp->retrans_out -= tcp_skb_pcount(skb); 1069 1070 tcp_skb_mark_lost_uncond_verify(tp, skb); 1071 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT); 1072 } else { 1073 if (before(ack_seq, new_low_seq)) 1074 new_low_seq = ack_seq; 1075 cnt += tcp_skb_pcount(skb); 1076 } 1077 } 1078 1079 if (tp->retrans_out) 1080 tp->lost_retrans_low = new_low_seq; 1081} 1082 1083static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb, 1084 struct tcp_sack_block_wire *sp, int num_sacks, 1085 u32 prior_snd_una) 1086{ 1087 struct tcp_sock *tp = tcp_sk(sk); 1088 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq); 1089 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq); 1090 bool dup_sack = false; 1091 1092 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) { 1093 dup_sack = true; 1094 tcp_dsack_seen(tp); 1095 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV); 1096 } else if (num_sacks > 1) { 1097 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq); 1098 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq); 1099 1100 if (!after(end_seq_0, end_seq_1) && 1101 !before(start_seq_0, start_seq_1)) { 1102 dup_sack = true; 1103 tcp_dsack_seen(tp); 1104 NET_INC_STATS_BH(sock_net(sk), 1105 LINUX_MIB_TCPDSACKOFORECV); 1106 } 1107 } 1108 1109 /* D-SACK for already forgotten data... Do dumb counting. */ 1110 if (dup_sack && tp->undo_marker && tp->undo_retrans > 0 && 1111 !after(end_seq_0, prior_snd_una) && 1112 after(end_seq_0, tp->undo_marker)) 1113 tp->undo_retrans--; 1114 1115 return dup_sack; 1116} 1117 1118struct tcp_sacktag_state { 1119 int reord; 1120 int fack_count; 1121 long rtt_us; /* RTT measured by SACKing never-retransmitted data */ 1122 int flag; 1123}; 1124 1125/* Check if skb is fully within the SACK block. In presence of GSO skbs, 1126 * the incoming SACK may not exactly match but we can find smaller MSS 1127 * aligned portion of it that matches. Therefore we might need to fragment 1128 * which may fail and creates some hassle (caller must handle error case 1129 * returns). 1130 * 1131 * FIXME: this could be merged to shift decision code 1132 */ 1133static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb, 1134 u32 start_seq, u32 end_seq) 1135{ 1136 int err; 1137 bool in_sack; 1138 unsigned int pkt_len; 1139 unsigned int mss; 1140 1141 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && 1142 !before(end_seq, TCP_SKB_CB(skb)->end_seq); 1143 1144 if (tcp_skb_pcount(skb) > 1 && !in_sack && 1145 after(TCP_SKB_CB(skb)->end_seq, start_seq)) { 1146 mss = tcp_skb_mss(skb); 1147 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); 1148 1149 if (!in_sack) { 1150 pkt_len = start_seq - TCP_SKB_CB(skb)->seq; 1151 if (pkt_len < mss) 1152 pkt_len = mss; 1153 } else { 1154 pkt_len = end_seq - TCP_SKB_CB(skb)->seq; 1155 if (pkt_len < mss) 1156 return -EINVAL; 1157 } 1158 1159 /* Round if necessary so that SACKs cover only full MSSes 1160 * and/or the remaining small portion (if present) 1161 */ 1162 if (pkt_len > mss) { 1163 unsigned int new_len = (pkt_len / mss) * mss; 1164 if (!in_sack && new_len < pkt_len) { 1165 new_len += mss; 1166 if (new_len >= skb->len) 1167 return 0; 1168 } 1169 pkt_len = new_len; 1170 } 1171 err = tcp_fragment(sk, skb, pkt_len, mss, GFP_ATOMIC); 1172 if (err < 0) 1173 return err; 1174 } 1175 1176 return in_sack; 1177} 1178 1179/* Mark the given newly-SACKed range as such, adjusting counters and hints. */ 1180static u8 tcp_sacktag_one(struct sock *sk, 1181 struct tcp_sacktag_state *state, u8 sacked, 1182 u32 start_seq, u32 end_seq, 1183 int dup_sack, int pcount, 1184 const struct skb_mstamp *xmit_time) 1185{ 1186 struct tcp_sock *tp = tcp_sk(sk); 1187 int fack_count = state->fack_count; 1188 1189 /* Account D-SACK for retransmitted packet. */ 1190 if (dup_sack && (sacked & TCPCB_RETRANS)) { 1191 if (tp->undo_marker && tp->undo_retrans > 0 && 1192 after(end_seq, tp->undo_marker)) 1193 tp->undo_retrans--; 1194 if (sacked & TCPCB_SACKED_ACKED) 1195 state->reord = min(fack_count, state->reord); 1196 } 1197 1198 /* Nothing to do; acked frame is about to be dropped (was ACKed). */ 1199 if (!after(end_seq, tp->snd_una)) 1200 return sacked; 1201 1202 if (!(sacked & TCPCB_SACKED_ACKED)) { 1203 if (sacked & TCPCB_SACKED_RETRANS) { 1204 /* If the segment is not tagged as lost, 1205 * we do not clear RETRANS, believing 1206 * that retransmission is still in flight. 1207 */ 1208 if (sacked & TCPCB_LOST) { 1209 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS); 1210 tp->lost_out -= pcount; 1211 tp->retrans_out -= pcount; 1212 } 1213 } else { 1214 if (!(sacked & TCPCB_RETRANS)) { 1215 /* New sack for not retransmitted frame, 1216 * which was in hole. It is reordering. 1217 */ 1218 if (before(start_seq, 1219 tcp_highest_sack_seq(tp))) 1220 state->reord = min(fack_count, 1221 state->reord); 1222 if (!after(end_seq, tp->high_seq)) 1223 state->flag |= FLAG_ORIG_SACK_ACKED; 1224 /* Pick the earliest sequence sacked for RTT */ 1225 if (state->rtt_us < 0) { 1226 struct skb_mstamp now; 1227 1228 skb_mstamp_get(&now); 1229 state->rtt_us = skb_mstamp_us_delta(&now, 1230 xmit_time); 1231 } 1232 } 1233 1234 if (sacked & TCPCB_LOST) { 1235 sacked &= ~TCPCB_LOST; 1236 tp->lost_out -= pcount; 1237 } 1238 } 1239 1240 sacked |= TCPCB_SACKED_ACKED; 1241 state->flag |= FLAG_DATA_SACKED; 1242 tp->sacked_out += pcount; 1243 1244 fack_count += pcount; 1245 1246 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */ 1247 if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) && 1248 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq)) 1249 tp->lost_cnt_hint += pcount; 1250 1251 if (fack_count > tp->fackets_out) 1252 tp->fackets_out = fack_count; 1253 } 1254 1255 /* D-SACK. We can detect redundant retransmission in S|R and plain R 1256 * frames and clear it. undo_retrans is decreased above, L|R frames 1257 * are accounted above as well. 1258 */ 1259 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) { 1260 sacked &= ~TCPCB_SACKED_RETRANS; 1261 tp->retrans_out -= pcount; 1262 } 1263 1264 return sacked; 1265} 1266 1267/* Shift newly-SACKed bytes from this skb to the immediately previous 1268 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such. 1269 */ 1270static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *skb, 1271 struct tcp_sacktag_state *state, 1272 unsigned int pcount, int shifted, int mss, 1273 bool dup_sack) 1274{ 1275 struct tcp_sock *tp = tcp_sk(sk); 1276 struct sk_buff *prev = tcp_write_queue_prev(sk, skb); 1277 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */ 1278 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */ 1279 1280 BUG_ON(!pcount); 1281 1282 /* Adjust counters and hints for the newly sacked sequence 1283 * range but discard the return value since prev is already 1284 * marked. We must tag the range first because the seq 1285 * advancement below implicitly advances 1286 * tcp_highest_sack_seq() when skb is highest_sack. 1287 */ 1288 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked, 1289 start_seq, end_seq, dup_sack, pcount, 1290 &skb->skb_mstamp); 1291 1292 if (skb == tp->lost_skb_hint) 1293 tp->lost_cnt_hint += pcount; 1294 1295 TCP_SKB_CB(prev)->end_seq += shifted; 1296 TCP_SKB_CB(skb)->seq += shifted; 1297 1298 skb_shinfo(prev)->gso_segs += pcount; 1299 BUG_ON(skb_shinfo(skb)->gso_segs < pcount); 1300 skb_shinfo(skb)->gso_segs -= pcount; 1301 1302 /* When we're adding to gso_segs == 1, gso_size will be zero, 1303 * in theory this shouldn't be necessary but as long as DSACK 1304 * code can come after this skb later on it's better to keep 1305 * setting gso_size to something. 1306 */ 1307 if (!skb_shinfo(prev)->gso_size) { 1308 skb_shinfo(prev)->gso_size = mss; 1309 skb_shinfo(prev)->gso_type = sk->sk_gso_type; 1310 } 1311 1312 /* CHECKME: To clear or not to clear? Mimics normal skb currently */ 1313 if (skb_shinfo(skb)->gso_segs <= 1) { 1314 skb_shinfo(skb)->gso_size = 0; 1315 skb_shinfo(skb)->gso_type = 0; 1316 } 1317 1318 /* Difference in this won't matter, both ACKed by the same cumul. ACK */ 1319 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS); 1320 1321 if (skb->len > 0) { 1322 BUG_ON(!tcp_skb_pcount(skb)); 1323 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED); 1324 return false; 1325 } 1326 1327 /* Whole SKB was eaten :-) */ 1328 1329 if (skb == tp->retransmit_skb_hint) 1330 tp->retransmit_skb_hint = prev; 1331 if (skb == tp->lost_skb_hint) { 1332 tp->lost_skb_hint = prev; 1333 tp->lost_cnt_hint -= tcp_skb_pcount(prev); 1334 } 1335 1336 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; 1337 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 1338 TCP_SKB_CB(prev)->end_seq++; 1339 1340 if (skb == tcp_highest_sack(sk)) 1341 tcp_advance_highest_sack(sk, skb); 1342 1343 tcp_unlink_write_queue(skb, sk); 1344 sk_wmem_free_skb(sk, skb); 1345 1346 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED); 1347 1348 return true; 1349} 1350 1351/* I wish gso_size would have a bit more sane initialization than 1352 * something-or-zero which complicates things 1353 */ 1354static int tcp_skb_seglen(const struct sk_buff *skb) 1355{ 1356 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb); 1357} 1358 1359/* Shifting pages past head area doesn't work */ 1360static int skb_can_shift(const struct sk_buff *skb) 1361{ 1362 return !skb_headlen(skb) && skb_is_nonlinear(skb); 1363} 1364 1365/* Try collapsing SACK blocks spanning across multiple skbs to a single 1366 * skb. 1367 */ 1368static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb, 1369 struct tcp_sacktag_state *state, 1370 u32 start_seq, u32 end_seq, 1371 bool dup_sack) 1372{ 1373 struct tcp_sock *tp = tcp_sk(sk); 1374 struct sk_buff *prev; 1375 int mss; 1376 int pcount = 0; 1377 int len; 1378 int in_sack; 1379 1380 if (!sk_can_gso(sk)) 1381 goto fallback; 1382 1383 /* Normally R but no L won't result in plain S */ 1384 if (!dup_sack && 1385 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS) 1386 goto fallback; 1387 if (!skb_can_shift(skb)) 1388 goto fallback; 1389 /* This frame is about to be dropped (was ACKed). */ 1390 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 1391 goto fallback; 1392 1393 /* Can only happen with delayed DSACK + discard craziness */ 1394 if (unlikely(skb == tcp_write_queue_head(sk))) 1395 goto fallback; 1396 prev = tcp_write_queue_prev(sk, skb); 1397 1398 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) 1399 goto fallback; 1400 1401 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && 1402 !before(end_seq, TCP_SKB_CB(skb)->end_seq); 1403 1404 if (in_sack) { 1405 len = skb->len; 1406 pcount = tcp_skb_pcount(skb); 1407 mss = tcp_skb_seglen(skb); 1408 1409 /* TODO: Fix DSACKs to not fragment already SACKed and we can 1410 * drop this restriction as unnecessary 1411 */ 1412 if (mss != tcp_skb_seglen(prev)) 1413 goto fallback; 1414 } else { 1415 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq)) 1416 goto noop; 1417 /* CHECKME: This is non-MSS split case only?, this will 1418 * cause skipped skbs due to advancing loop btw, original 1419 * has that feature too 1420 */ 1421 if (tcp_skb_pcount(skb) <= 1) 1422 goto noop; 1423 1424 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); 1425 if (!in_sack) { 1426 /* TODO: head merge to next could be attempted here 1427 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)), 1428 * though it might not be worth of the additional hassle 1429 * 1430 * ...we can probably just fallback to what was done 1431 * previously. We could try merging non-SACKed ones 1432 * as well but it probably isn't going to buy off 1433 * because later SACKs might again split them, and 1434 * it would make skb timestamp tracking considerably 1435 * harder problem. 1436 */ 1437 goto fallback; 1438 } 1439 1440 len = end_seq - TCP_SKB_CB(skb)->seq; 1441 BUG_ON(len < 0); 1442 BUG_ON(len > skb->len); 1443 1444 /* MSS boundaries should be honoured or else pcount will 1445 * severely break even though it makes things bit trickier. 1446 * Optimize common case to avoid most of the divides 1447 */ 1448 mss = tcp_skb_mss(skb); 1449 1450 /* TODO: Fix DSACKs to not fragment already SACKed and we can 1451 * drop this restriction as unnecessary 1452 */ 1453 if (mss != tcp_skb_seglen(prev)) 1454 goto fallback; 1455 1456 if (len == mss) { 1457 pcount = 1; 1458 } else if (len < mss) { 1459 goto noop; 1460 } else { 1461 pcount = len / mss; 1462 len = pcount * mss; 1463 } 1464 } 1465 1466 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */ 1467 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una)) 1468 goto fallback; 1469 1470 if (!skb_shift(prev, skb, len)) 1471 goto fallback; 1472 if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack)) 1473 goto out; 1474 1475 /* Hole filled allows collapsing with the next as well, this is very 1476 * useful when hole on every nth skb pattern happens 1477 */ 1478 if (prev == tcp_write_queue_tail(sk)) 1479 goto out; 1480 skb = tcp_write_queue_next(sk, prev); 1481 1482 if (!skb_can_shift(skb) || 1483 (skb == tcp_send_head(sk)) || 1484 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) || 1485 (mss != tcp_skb_seglen(skb))) 1486 goto out; 1487 1488 len = skb->len; 1489 if (skb_shift(prev, skb, len)) { 1490 pcount += tcp_skb_pcount(skb); 1491 tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0); 1492 } 1493 1494out: 1495 state->fack_count += pcount; 1496 return prev; 1497 1498noop: 1499 return skb; 1500 1501fallback: 1502 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK); 1503 return NULL; 1504} 1505 1506static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk, 1507 struct tcp_sack_block *next_dup, 1508 struct tcp_sacktag_state *state, 1509 u32 start_seq, u32 end_seq, 1510 bool dup_sack_in) 1511{ 1512 struct tcp_sock *tp = tcp_sk(sk); 1513 struct sk_buff *tmp; 1514 1515 tcp_for_write_queue_from(skb, sk) { 1516 int in_sack = 0; 1517 bool dup_sack = dup_sack_in; 1518 1519 if (skb == tcp_send_head(sk)) 1520 break; 1521 1522 /* queue is in-order => we can short-circuit the walk early */ 1523 if (!before(TCP_SKB_CB(skb)->seq, end_seq)) 1524 break; 1525 1526 if ((next_dup != NULL) && 1527 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) { 1528 in_sack = tcp_match_skb_to_sack(sk, skb, 1529 next_dup->start_seq, 1530 next_dup->end_seq); 1531 if (in_sack > 0) 1532 dup_sack = true; 1533 } 1534 1535 /* skb reference here is a bit tricky to get right, since 1536 * shifting can eat and free both this skb and the next, 1537 * so not even _safe variant of the loop is enough. 1538 */ 1539 if (in_sack <= 0) { 1540 tmp = tcp_shift_skb_data(sk, skb, state, 1541 start_seq, end_seq, dup_sack); 1542 if (tmp != NULL) { 1543 if (tmp != skb) { 1544 skb = tmp; 1545 continue; 1546 } 1547 1548 in_sack = 0; 1549 } else { 1550 in_sack = tcp_match_skb_to_sack(sk, skb, 1551 start_seq, 1552 end_seq); 1553 } 1554 } 1555 1556 if (unlikely(in_sack < 0)) 1557 break; 1558 1559 if (in_sack) { 1560 TCP_SKB_CB(skb)->sacked = 1561 tcp_sacktag_one(sk, 1562 state, 1563 TCP_SKB_CB(skb)->sacked, 1564 TCP_SKB_CB(skb)->seq, 1565 TCP_SKB_CB(skb)->end_seq, 1566 dup_sack, 1567 tcp_skb_pcount(skb), 1568 &skb->skb_mstamp); 1569 1570 if (!before(TCP_SKB_CB(skb)->seq, 1571 tcp_highest_sack_seq(tp))) 1572 tcp_advance_highest_sack(sk, skb); 1573 } 1574 1575 state->fack_count += tcp_skb_pcount(skb); 1576 } 1577 return skb; 1578} 1579 1580/* Avoid all extra work that is being done by sacktag while walking in 1581 * a normal way 1582 */ 1583static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk, 1584 struct tcp_sacktag_state *state, 1585 u32 skip_to_seq) 1586{ 1587 tcp_for_write_queue_from(skb, sk) { 1588 if (skb == tcp_send_head(sk)) 1589 break; 1590 1591 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq)) 1592 break; 1593 1594 state->fack_count += tcp_skb_pcount(skb); 1595 } 1596 return skb; 1597} 1598 1599static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb, 1600 struct sock *sk, 1601 struct tcp_sack_block *next_dup, 1602 struct tcp_sacktag_state *state, 1603 u32 skip_to_seq) 1604{ 1605 if (next_dup == NULL) 1606 return skb; 1607 1608 if (before(next_dup->start_seq, skip_to_seq)) { 1609 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq); 1610 skb = tcp_sacktag_walk(skb, sk, NULL, state, 1611 next_dup->start_seq, next_dup->end_seq, 1612 1); 1613 } 1614 1615 return skb; 1616} 1617 1618static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache) 1619{ 1620 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache); 1621} 1622 1623static int 1624tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb, 1625 u32 prior_snd_una, long *sack_rtt_us) 1626{ 1627 struct tcp_sock *tp = tcp_sk(sk); 1628 const unsigned char *ptr = (skb_transport_header(ack_skb) + 1629 TCP_SKB_CB(ack_skb)->sacked); 1630 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2); 1631 struct tcp_sack_block sp[TCP_NUM_SACKS]; 1632 struct tcp_sack_block *cache; 1633 struct tcp_sacktag_state state; 1634 struct sk_buff *skb; 1635 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3); 1636 int used_sacks; 1637 bool found_dup_sack = false; 1638 int i, j; 1639 int first_sack_index; 1640 1641 state.flag = 0; 1642 state.reord = tp->packets_out; 1643 state.rtt_us = -1L; 1644 1645 if (!tp->sacked_out) { 1646 if (WARN_ON(tp->fackets_out)) 1647 tp->fackets_out = 0; 1648 tcp_highest_sack_reset(sk); 1649 } 1650 1651 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire, 1652 num_sacks, prior_snd_una); 1653 if (found_dup_sack) 1654 state.flag |= FLAG_DSACKING_ACK; 1655 1656 /* Eliminate too old ACKs, but take into 1657 * account more or less fresh ones, they can 1658 * contain valid SACK info. 1659 */ 1660 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window)) 1661 return 0; 1662 1663 if (!tp->packets_out) 1664 goto out; 1665 1666 used_sacks = 0; 1667 first_sack_index = 0; 1668 for (i = 0; i < num_sacks; i++) { 1669 bool dup_sack = !i && found_dup_sack; 1670 1671 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq); 1672 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq); 1673 1674 if (!tcp_is_sackblock_valid(tp, dup_sack, 1675 sp[used_sacks].start_seq, 1676 sp[used_sacks].end_seq)) { 1677 int mib_idx; 1678 1679 if (dup_sack) { 1680 if (!tp->undo_marker) 1681 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO; 1682 else 1683 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD; 1684 } else { 1685 /* Don't count olds caused by ACK reordering */ 1686 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) && 1687 !after(sp[used_sacks].end_seq, tp->snd_una)) 1688 continue; 1689 mib_idx = LINUX_MIB_TCPSACKDISCARD; 1690 } 1691 1692 NET_INC_STATS_BH(sock_net(sk), mib_idx); 1693 if (i == 0) 1694 first_sack_index = -1; 1695 continue; 1696 } 1697 1698 /* Ignore very old stuff early */ 1699 if (!after(sp[used_sacks].end_seq, prior_snd_una)) 1700 continue; 1701 1702 used_sacks++; 1703 } 1704 1705 /* order SACK blocks to allow in order walk of the retrans queue */ 1706 for (i = used_sacks - 1; i > 0; i--) { 1707 for (j = 0; j < i; j++) { 1708 if (after(sp[j].start_seq, sp[j + 1].start_seq)) { 1709 swap(sp[j], sp[j + 1]); 1710 1711 /* Track where the first SACK block goes to */ 1712 if (j == first_sack_index) 1713 first_sack_index = j + 1; 1714 } 1715 } 1716 } 1717 1718 skb = tcp_write_queue_head(sk); 1719 state.fack_count = 0; 1720 i = 0; 1721 1722 if (!tp->sacked_out) { 1723 /* It's already past, so skip checking against it */ 1724 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache); 1725 } else { 1726 cache = tp->recv_sack_cache; 1727 /* Skip empty blocks in at head of the cache */ 1728 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq && 1729 !cache->end_seq) 1730 cache++; 1731 } 1732 1733 while (i < used_sacks) { 1734 u32 start_seq = sp[i].start_seq; 1735 u32 end_seq = sp[i].end_seq; 1736 bool dup_sack = (found_dup_sack && (i == first_sack_index)); 1737 struct tcp_sack_block *next_dup = NULL; 1738 1739 if (found_dup_sack && ((i + 1) == first_sack_index)) 1740 next_dup = &sp[i + 1]; 1741 1742 /* Skip too early cached blocks */ 1743 while (tcp_sack_cache_ok(tp, cache) && 1744 !before(start_seq, cache->end_seq)) 1745 cache++; 1746 1747 /* Can skip some work by looking recv_sack_cache? */ 1748 if (tcp_sack_cache_ok(tp, cache) && !dup_sack && 1749 after(end_seq, cache->start_seq)) { 1750 1751 /* Head todo? */ 1752 if (before(start_seq, cache->start_seq)) { 1753 skb = tcp_sacktag_skip(skb, sk, &state, 1754 start_seq); 1755 skb = tcp_sacktag_walk(skb, sk, next_dup, 1756 &state, 1757 start_seq, 1758 cache->start_seq, 1759 dup_sack); 1760 } 1761 1762 /* Rest of the block already fully processed? */ 1763 if (!after(end_seq, cache->end_seq)) 1764 goto advance_sp; 1765 1766 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup, 1767 &state, 1768 cache->end_seq); 1769 1770 /* ...tail remains todo... */ 1771 if (tcp_highest_sack_seq(tp) == cache->end_seq) { 1772 /* ...but better entrypoint exists! */ 1773 skb = tcp_highest_sack(sk); 1774 if (skb == NULL) 1775 break; 1776 state.fack_count = tp->fackets_out; 1777 cache++; 1778 goto walk; 1779 } 1780 1781 skb = tcp_sacktag_skip(skb, sk, &state, cache->end_seq); 1782 /* Check overlap against next cached too (past this one already) */ 1783 cache++; 1784 continue; 1785 } 1786 1787 if (!before(start_seq, tcp_highest_sack_seq(tp))) { 1788 skb = tcp_highest_sack(sk); 1789 if (skb == NULL) 1790 break; 1791 state.fack_count = tp->fackets_out; 1792 } 1793 skb = tcp_sacktag_skip(skb, sk, &state, start_seq); 1794 1795walk: 1796 skb = tcp_sacktag_walk(skb, sk, next_dup, &state, 1797 start_seq, end_seq, dup_sack); 1798 1799advance_sp: 1800 i++; 1801 } 1802 1803 /* Clear the head of the cache sack blocks so we can skip it next time */ 1804 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) { 1805 tp->recv_sack_cache[i].start_seq = 0; 1806 tp->recv_sack_cache[i].end_seq = 0; 1807 } 1808 for (j = 0; j < used_sacks; j++) 1809 tp->recv_sack_cache[i++] = sp[j]; 1810 1811 tcp_mark_lost_retrans(sk); 1812 1813 tcp_verify_left_out(tp); 1814 1815 if ((state.reord < tp->fackets_out) && 1816 ((inet_csk(sk)->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker)) 1817 tcp_update_reordering(sk, tp->fackets_out - state.reord, 0); 1818 1819out: 1820 1821#if FASTRETRANS_DEBUG > 0 1822 WARN_ON((int)tp->sacked_out < 0); 1823 WARN_ON((int)tp->lost_out < 0); 1824 WARN_ON((int)tp->retrans_out < 0); 1825 WARN_ON((int)tcp_packets_in_flight(tp) < 0); 1826#endif 1827 *sack_rtt_us = state.rtt_us; 1828 return state.flag; 1829} 1830 1831/* Limits sacked_out so that sum with lost_out isn't ever larger than 1832 * packets_out. Returns false if sacked_out adjustement wasn't necessary. 1833 */ 1834static bool tcp_limit_reno_sacked(struct tcp_sock *tp) 1835{ 1836 u32 holes; 1837 1838 holes = max(tp->lost_out, 1U); 1839 holes = min(holes, tp->packets_out); 1840 1841 if ((tp->sacked_out + holes) > tp->packets_out) { 1842 tp->sacked_out = tp->packets_out - holes; 1843 return true; 1844 } 1845 return false; 1846} 1847 1848/* If we receive more dupacks than we expected counting segments 1849 * in assumption of absent reordering, interpret this as reordering. 1850 * The only another reason could be bug in receiver TCP. 1851 */ 1852static void tcp_check_reno_reordering(struct sock *sk, const int addend) 1853{ 1854 struct tcp_sock *tp = tcp_sk(sk); 1855 if (tcp_limit_reno_sacked(tp)) 1856 tcp_update_reordering(sk, tp->packets_out + addend, 0); 1857} 1858 1859/* Emulate SACKs for SACKless connection: account for a new dupack. */ 1860 1861static void tcp_add_reno_sack(struct sock *sk) 1862{ 1863 struct tcp_sock *tp = tcp_sk(sk); 1864 tp->sacked_out++; 1865 tcp_check_reno_reordering(sk, 0); 1866 tcp_verify_left_out(tp); 1867} 1868 1869/* Account for ACK, ACKing some data in Reno Recovery phase. */ 1870 1871static void tcp_remove_reno_sacks(struct sock *sk, int acked) 1872{ 1873 struct tcp_sock *tp = tcp_sk(sk); 1874 1875 if (acked > 0) { 1876 /* One ACK acked hole. The rest eat duplicate ACKs. */ 1877 if (acked - 1 >= tp->sacked_out) 1878 tp->sacked_out = 0; 1879 else 1880 tp->sacked_out -= acked - 1; 1881 } 1882 tcp_check_reno_reordering(sk, acked); 1883 tcp_verify_left_out(tp); 1884} 1885 1886static inline void tcp_reset_reno_sack(struct tcp_sock *tp) 1887{ 1888 tp->sacked_out = 0; 1889} 1890 1891static void tcp_clear_retrans_partial(struct tcp_sock *tp) 1892{ 1893 tp->retrans_out = 0; 1894 tp->lost_out = 0; 1895 1896 tp->undo_marker = 0; 1897 tp->undo_retrans = -1; 1898} 1899 1900void tcp_clear_retrans(struct tcp_sock *tp) 1901{ 1902 tcp_clear_retrans_partial(tp); 1903 1904 tp->fackets_out = 0; 1905 tp->sacked_out = 0; 1906} 1907 1908/* Enter Loss state. If we detect SACK reneging, forget all SACK information 1909 * and reset tags completely, otherwise preserve SACKs. If receiver 1910 * dropped its ofo queue, we will know this due to reneging detection. 1911 */ 1912void tcp_enter_loss(struct sock *sk) 1913{ 1914 const struct inet_connection_sock *icsk = inet_csk(sk); 1915 struct tcp_sock *tp = tcp_sk(sk); 1916 struct sk_buff *skb; 1917 bool new_recovery = false; 1918 bool is_reneg; /* is receiver reneging on SACKs? */ 1919 1920 /* Reduce ssthresh if it has not yet been made inside this window. */ 1921 if (icsk->icsk_ca_state <= TCP_CA_Disorder || 1922 !after(tp->high_seq, tp->snd_una) || 1923 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) { 1924 new_recovery = true; 1925 tp->prior_ssthresh = tcp_current_ssthresh(sk); 1926 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 1927 tcp_ca_event(sk, CA_EVENT_LOSS); 1928 } 1929 tp->snd_cwnd = 1; 1930 tp->snd_cwnd_cnt = 0; 1931 tp->snd_cwnd_stamp = tcp_time_stamp; 1932 1933 tcp_clear_retrans_partial(tp); 1934 1935 if (tcp_is_reno(tp)) 1936 tcp_reset_reno_sack(tp); 1937 1938 tp->undo_marker = tp->snd_una; 1939 1940 skb = tcp_write_queue_head(sk); 1941 is_reneg = skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED); 1942 if (is_reneg) { 1943 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING); 1944 tp->sacked_out = 0; 1945 tp->fackets_out = 0; 1946 } 1947 tcp_clear_all_retrans_hints(tp); 1948 1949 tcp_for_write_queue(skb, sk) { 1950 if (skb == tcp_send_head(sk)) 1951 break; 1952 1953 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 1954 tp->undo_marker = 0; 1955 1956 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED; 1957 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || is_reneg) { 1958 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED; 1959 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 1960 tp->lost_out += tcp_skb_pcount(skb); 1961 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq; 1962 } 1963 } 1964 tcp_verify_left_out(tp); 1965 1966 /* Timeout in disordered state after receiving substantial DUPACKs 1967 * suggests that the degree of reordering is over-estimated. 1968 */ 1969 if (icsk->icsk_ca_state <= TCP_CA_Disorder && 1970 tp->sacked_out >= sysctl_tcp_reordering) 1971 tp->reordering = min_t(unsigned int, tp->reordering, 1972 sysctl_tcp_reordering); 1973 tcp_set_ca_state(sk, TCP_CA_Loss); 1974 tp->high_seq = tp->snd_nxt; 1975 TCP_ECN_queue_cwr(tp); 1976 1977 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous 1978 * loss recovery is underway except recurring timeout(s) on 1979 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing 1980 */ 1981 tp->frto = sysctl_tcp_frto && 1982 (new_recovery || icsk->icsk_retransmits) && 1983 !inet_csk(sk)->icsk_mtup.probe_size; 1984} 1985 1986/* If ACK arrived pointing to a remembered SACK, it means that our 1987 * remembered SACKs do not reflect real state of receiver i.e. 1988 * receiver _host_ is heavily congested (or buggy). 1989 * 1990 * To avoid big spurious retransmission bursts due to transient SACK 1991 * scoreboard oddities that look like reneging, we give the receiver a 1992 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will 1993 * restore sanity to the SACK scoreboard. If the apparent reneging 1994 * persists until this RTO then we'll clear the SACK scoreboard. 1995 */ 1996static bool tcp_check_sack_reneging(struct sock *sk, int flag) 1997{ 1998 if (flag & FLAG_SACK_RENEGING) { 1999 struct tcp_sock *tp = tcp_sk(sk); 2000 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4), 2001 msecs_to_jiffies(10)); 2002 2003 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 2004 delay, TCP_RTO_MAX); 2005 return true; 2006 } 2007 return false; 2008} 2009 2010static inline int tcp_fackets_out(const struct tcp_sock *tp) 2011{ 2012 return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out; 2013} 2014 2015/* Heurestics to calculate number of duplicate ACKs. There's no dupACKs 2016 * counter when SACK is enabled (without SACK, sacked_out is used for 2017 * that purpose). 2018 * 2019 * Instead, with FACK TCP uses fackets_out that includes both SACKed 2020 * segments up to the highest received SACK block so far and holes in 2021 * between them. 2022 * 2023 * With reordering, holes may still be in flight, so RFC3517 recovery 2024 * uses pure sacked_out (total number of SACKed segments) even though 2025 * it violates the RFC that uses duplicate ACKs, often these are equal 2026 * but when e.g. out-of-window ACKs or packet duplication occurs, 2027 * they differ. Since neither occurs due to loss, TCP should really 2028 * ignore them. 2029 */ 2030static inline int tcp_dupack_heuristics(const struct tcp_sock *tp) 2031{ 2032 return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1; 2033} 2034 2035static bool tcp_pause_early_retransmit(struct sock *sk, int flag) 2036{ 2037 struct tcp_sock *tp = tcp_sk(sk); 2038 unsigned long delay; 2039 2040 /* Delay early retransmit and entering fast recovery for 2041 * max(RTT/4, 2msec) unless ack has ECE mark, no RTT samples 2042 * available, or RTO is scheduled to fire first. 2043 */ 2044 if (sysctl_tcp_early_retrans < 2 || sysctl_tcp_early_retrans > 3 || 2045 (flag & FLAG_ECE) || !tp->srtt_us) 2046 return false; 2047 2048 delay = max(usecs_to_jiffies(tp->srtt_us >> 5), 2049 msecs_to_jiffies(2)); 2050 2051 if (!time_after(inet_csk(sk)->icsk_timeout, (jiffies + delay))) 2052 return false; 2053 2054 inet_csk_reset_xmit_timer(sk, ICSK_TIME_EARLY_RETRANS, delay, 2055 TCP_RTO_MAX); 2056 return true; 2057} 2058 2059/* Linux NewReno/SACK/FACK/ECN state machine. 2060 * -------------------------------------- 2061 * 2062 * "Open" Normal state, no dubious events, fast path. 2063 * "Disorder" In all the respects it is "Open", 2064 * but requires a bit more attention. It is entered when 2065 * we see some SACKs or dupacks. It is split of "Open" 2066 * mainly to move some processing from fast path to slow one. 2067 * "CWR" CWND was reduced due to some Congestion Notification event. 2068 * It can be ECN, ICMP source quench, local device congestion. 2069 * "Recovery" CWND was reduced, we are fast-retransmitting. 2070 * "Loss" CWND was reduced due to RTO timeout or SACK reneging. 2071 * 2072 * tcp_fastretrans_alert() is entered: 2073 * - each incoming ACK, if state is not "Open" 2074 * - when arrived ACK is unusual, namely: 2075 * * SACK 2076 * * Duplicate ACK. 2077 * * ECN ECE. 2078 * 2079 * Counting packets in flight is pretty simple. 2080 * 2081 * in_flight = packets_out - left_out + retrans_out 2082 * 2083 * packets_out is SND.NXT-SND.UNA counted in packets. 2084 * 2085 * retrans_out is number of retransmitted segments. 2086 * 2087 * left_out is number of segments left network, but not ACKed yet. 2088 * 2089 * left_out = sacked_out + lost_out 2090 * 2091 * sacked_out: Packets, which arrived to receiver out of order 2092 * and hence not ACKed. With SACKs this number is simply 2093 * amount of SACKed data. Even without SACKs 2094 * it is easy to give pretty reliable estimate of this number, 2095 * counting duplicate ACKs. 2096 * 2097 * lost_out: Packets lost by network. TCP has no explicit 2098 * "loss notification" feedback from network (for now). 2099 * It means that this number can be only _guessed_. 2100 * Actually, it is the heuristics to predict lossage that 2101 * distinguishes different algorithms. 2102 * 2103 * F.e. after RTO, when all the queue is considered as lost, 2104 * lost_out = packets_out and in_flight = retrans_out. 2105 * 2106 * Essentially, we have now two algorithms counting 2107 * lost packets. 2108 * 2109 * FACK: It is the simplest heuristics. As soon as we decided 2110 * that something is lost, we decide that _all_ not SACKed 2111 * packets until the most forward SACK are lost. I.e. 2112 * lost_out = fackets_out - sacked_out and left_out = fackets_out. 2113 * It is absolutely correct estimate, if network does not reorder 2114 * packets. And it loses any connection to reality when reordering 2115 * takes place. We use FACK by default until reordering 2116 * is suspected on the path to this destination. 2117 * 2118 * NewReno: when Recovery is entered, we assume that one segment 2119 * is lost (classic Reno). While we are in Recovery and 2120 * a partial ACK arrives, we assume that one more packet 2121 * is lost (NewReno). This heuristics are the same in NewReno 2122 * and SACK. 2123 * 2124 * Imagine, that's all! Forget about all this shamanism about CWND inflation 2125 * deflation etc. CWND is real congestion window, never inflated, changes 2126 * only according to classic VJ rules. 2127 * 2128 * Really tricky (and requiring careful tuning) part of algorithm 2129 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue(). 2130 * The first determines the moment _when_ we should reduce CWND and, 2131 * hence, slow down forward transmission. In fact, it determines the moment 2132 * when we decide that hole is caused by loss, rather than by a reorder. 2133 * 2134 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill 2135 * holes, caused by lost packets. 2136 * 2137 * And the most logically complicated part of algorithm is undo 2138 * heuristics. We detect false retransmits due to both too early 2139 * fast retransmit (reordering) and underestimated RTO, analyzing 2140 * timestamps and D-SACKs. When we detect that some segments were 2141 * retransmitted by mistake and CWND reduction was wrong, we undo 2142 * window reduction and abort recovery phase. This logic is hidden 2143 * inside several functions named tcp_try_undo_<something>. 2144 */ 2145 2146/* This function decides, when we should leave Disordered state 2147 * and enter Recovery phase, reducing congestion window. 2148 * 2149 * Main question: may we further continue forward transmission 2150 * with the same cwnd? 2151 */ 2152static bool tcp_time_to_recover(struct sock *sk, int flag) 2153{ 2154 struct tcp_sock *tp = tcp_sk(sk); 2155 __u32 packets_out; 2156 2157 /* Trick#1: The loss is proven. */ 2158 if (tp->lost_out) 2159 return true; 2160 2161 /* Not-A-Trick#2 : Classic rule... */ 2162 if (tcp_dupack_heuristics(tp) > tp->reordering) 2163 return true; 2164 2165 /* Trick#4: It is still not OK... But will it be useful to delay 2166 * recovery more? 2167 */ 2168 packets_out = tp->packets_out; 2169 if (packets_out <= tp->reordering && 2170 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) && 2171 !tcp_may_send_now(sk)) { 2172 /* We have nothing to send. This connection is limited 2173 * either by receiver window or by application. 2174 */ 2175 return true; 2176 } 2177 2178 /* If a thin stream is detected, retransmit after first 2179 * received dupack. Employ only if SACK is supported in order 2180 * to avoid possible corner-case series of spurious retransmissions 2181 * Use only if there are no unsent data. 2182 */ 2183 if ((tp->thin_dupack || sysctl_tcp_thin_dupack) && 2184 tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 && 2185 tcp_is_sack(tp) && !tcp_send_head(sk)) 2186 return true; 2187 2188 /* Trick#6: TCP early retransmit, per RFC5827. To avoid spurious 2189 * retransmissions due to small network reorderings, we implement 2190 * Mitigation A.3 in the RFC and delay the retransmission for a short 2191 * interval if appropriate. 2192 */ 2193 if (tp->do_early_retrans && !tp->retrans_out && tp->sacked_out && 2194 (tp->packets_out >= (tp->sacked_out + 1) && tp->packets_out < 4) && 2195 !tcp_may_send_now(sk)) 2196 return !tcp_pause_early_retransmit(sk, flag); 2197 2198 return false; 2199} 2200 2201/* Detect loss in event "A" above by marking head of queue up as lost. 2202 * For FACK or non-SACK(Reno) senders, the first "packets" number of segments 2203 * are considered lost. For RFC3517 SACK, a segment is considered lost if it 2204 * has at least tp->reordering SACKed seqments above it; "packets" refers to 2205 * the maximum SACKed segments to pass before reaching this limit. 2206 */ 2207static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head) 2208{ 2209 struct tcp_sock *tp = tcp_sk(sk); 2210 struct sk_buff *skb; 2211 int cnt, oldcnt; 2212 int err; 2213 unsigned int mss; 2214 /* Use SACK to deduce losses of new sequences sent during recovery */ 2215 const u32 loss_high = tcp_is_sack(tp) ? tp->snd_nxt : tp->high_seq; 2216 2217 WARN_ON(packets > tp->packets_out); 2218 if (tp->lost_skb_hint) { 2219 skb = tp->lost_skb_hint; 2220 cnt = tp->lost_cnt_hint; 2221 /* Head already handled? */ 2222 if (mark_head && skb != tcp_write_queue_head(sk)) 2223 return; 2224 } else { 2225 skb = tcp_write_queue_head(sk); 2226 cnt = 0; 2227 } 2228 2229 tcp_for_write_queue_from(skb, sk) { 2230 if (skb == tcp_send_head(sk)) 2231 break; 2232 /* TODO: do this better */ 2233 /* this is not the most efficient way to do this... */ 2234 tp->lost_skb_hint = skb; 2235 tp->lost_cnt_hint = cnt; 2236 2237 if (after(TCP_SKB_CB(skb)->end_seq, loss_high)) 2238 break; 2239 2240 oldcnt = cnt; 2241 if (tcp_is_fack(tp) || tcp_is_reno(tp) || 2242 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) 2243 cnt += tcp_skb_pcount(skb); 2244 2245 if (cnt > packets) { 2246 if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) || 2247 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) || 2248 (oldcnt >= packets)) 2249 break; 2250 2251 mss = skb_shinfo(skb)->gso_size; 2252 err = tcp_fragment(sk, skb, (packets - oldcnt) * mss, 2253 mss, GFP_ATOMIC); 2254 if (err < 0) 2255 break; 2256 cnt = packets; 2257 } 2258 2259 tcp_skb_mark_lost(tp, skb); 2260 2261 if (mark_head) 2262 break; 2263 } 2264 tcp_verify_left_out(tp); 2265} 2266 2267/* Account newly detected lost packet(s) */ 2268 2269static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit) 2270{ 2271 struct tcp_sock *tp = tcp_sk(sk); 2272 2273 if (tcp_is_reno(tp)) { 2274 tcp_mark_head_lost(sk, 1, 1); 2275 } else if (tcp_is_fack(tp)) { 2276 int lost = tp->fackets_out - tp->reordering; 2277 if (lost <= 0) 2278 lost = 1; 2279 tcp_mark_head_lost(sk, lost, 0); 2280 } else { 2281 int sacked_upto = tp->sacked_out - tp->reordering; 2282 if (sacked_upto >= 0) 2283 tcp_mark_head_lost(sk, sacked_upto, 0); 2284 else if (fast_rexmit) 2285 tcp_mark_head_lost(sk, 1, 1); 2286 } 2287} 2288 2289/* CWND moderation, preventing bursts due to too big ACKs 2290 * in dubious situations. 2291 */ 2292static inline void tcp_moderate_cwnd(struct tcp_sock *tp) 2293{ 2294 tp->snd_cwnd = min(tp->snd_cwnd, 2295 tcp_packets_in_flight(tp) + tcp_max_burst(tp)); 2296 tp->snd_cwnd_stamp = tcp_time_stamp; 2297} 2298 2299/* Nothing was retransmitted or returned timestamp is less 2300 * than timestamp of the first retransmission. 2301 */ 2302static inline bool tcp_packet_delayed(const struct tcp_sock *tp) 2303{ 2304 return !tp->retrans_stamp || 2305 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 2306 before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp)); 2307} 2308 2309/* Undo procedures. */ 2310 2311#if FASTRETRANS_DEBUG > 1 2312static void DBGUNDO(struct sock *sk, const char *msg) 2313{ 2314 struct tcp_sock *tp = tcp_sk(sk); 2315 struct inet_sock *inet = inet_sk(sk); 2316 2317 if (sk->sk_family == AF_INET) { 2318 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n", 2319 msg, 2320 &inet->inet_daddr, ntohs(inet->inet_dport), 2321 tp->snd_cwnd, tcp_left_out(tp), 2322 tp->snd_ssthresh, tp->prior_ssthresh, 2323 tp->packets_out); 2324 } 2325#if IS_ENABLED(CONFIG_IPV6) 2326 else if (sk->sk_family == AF_INET6) { 2327 struct ipv6_pinfo *np = inet6_sk(sk); 2328 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n", 2329 msg, 2330 &np->daddr, ntohs(inet->inet_dport), 2331 tp->snd_cwnd, tcp_left_out(tp), 2332 tp->snd_ssthresh, tp->prior_ssthresh, 2333 tp->packets_out); 2334 } 2335#endif 2336} 2337#else 2338#define DBGUNDO(x...) do { } while (0) 2339#endif 2340 2341static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss) 2342{ 2343 struct tcp_sock *tp = tcp_sk(sk); 2344 2345 if (unmark_loss) { 2346 struct sk_buff *skb; 2347 2348 tcp_for_write_queue(skb, sk) { 2349 if (skb == tcp_send_head(sk)) 2350 break; 2351 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; 2352 } 2353 tp->lost_out = 0; 2354 tcp_clear_all_retrans_hints(tp); 2355 } 2356 2357 if (tp->prior_ssthresh) { 2358 const struct inet_connection_sock *icsk = inet_csk(sk); 2359 2360 if (icsk->icsk_ca_ops->undo_cwnd) 2361 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk); 2362 else 2363 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1); 2364 2365 if (tp->prior_ssthresh > tp->snd_ssthresh) { 2366 tp->snd_ssthresh = tp->prior_ssthresh; 2367 TCP_ECN_withdraw_cwr(tp); 2368 } 2369 } else { 2370 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh); 2371 } 2372 tp->snd_cwnd_stamp = tcp_time_stamp; 2373 tp->undo_marker = 0; 2374} 2375 2376static inline bool tcp_may_undo(const struct tcp_sock *tp) 2377{ 2378 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp)); 2379} 2380 2381/* People celebrate: "We love our President!" */ 2382static bool tcp_try_undo_recovery(struct sock *sk) 2383{ 2384 struct tcp_sock *tp = tcp_sk(sk); 2385 2386 if (tcp_may_undo(tp)) { 2387 int mib_idx; 2388 2389 /* Happy end! We did not retransmit anything 2390 * or our original transmission succeeded. 2391 */ 2392 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans"); 2393 tcp_undo_cwnd_reduction(sk, false); 2394 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss) 2395 mib_idx = LINUX_MIB_TCPLOSSUNDO; 2396 else 2397 mib_idx = LINUX_MIB_TCPFULLUNDO; 2398 2399 NET_INC_STATS_BH(sock_net(sk), mib_idx); 2400 } 2401 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) { 2402 /* Hold old state until something *above* high_seq 2403 * is ACKed. For Reno it is MUST to prevent false 2404 * fast retransmits (RFC2582). SACK TCP is safe. */ 2405 tcp_moderate_cwnd(tp); 2406 return true; 2407 } 2408 tcp_set_ca_state(sk, TCP_CA_Open); 2409 return false; 2410} 2411 2412/* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */ 2413static bool tcp_try_undo_dsack(struct sock *sk) 2414{ 2415 struct tcp_sock *tp = tcp_sk(sk); 2416 2417 if (tp->undo_marker && !tp->undo_retrans) { 2418 DBGUNDO(sk, "D-SACK"); 2419 tcp_undo_cwnd_reduction(sk, false); 2420 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO); 2421 return true; 2422 } 2423 return false; 2424} 2425 2426/* We can clear retrans_stamp when there are no retransmissions in the 2427 * window. It would seem that it is trivially available for us in 2428 * tp->retrans_out, however, that kind of assumptions doesn't consider 2429 * what will happen if errors occur when sending retransmission for the 2430 * second time. ...It could the that such segment has only 2431 * TCPCB_EVER_RETRANS set at the present time. It seems that checking 2432 * the head skb is enough except for some reneging corner cases that 2433 * are not worth the effort. 2434 * 2435 * Main reason for all this complexity is the fact that connection dying 2436 * time now depends on the validity of the retrans_stamp, in particular, 2437 * that successive retransmissions of a segment must not advance 2438 * retrans_stamp under any conditions. 2439 */ 2440static bool tcp_any_retrans_done(const struct sock *sk) 2441{ 2442 const struct tcp_sock *tp = tcp_sk(sk); 2443 struct sk_buff *skb; 2444 2445 if (tp->retrans_out) 2446 return true; 2447 2448 skb = tcp_write_queue_head(sk); 2449 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS)) 2450 return true; 2451 2452 return false; 2453} 2454 2455/* Undo during loss recovery after partial ACK or using F-RTO. */ 2456static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo) 2457{ 2458 struct tcp_sock *tp = tcp_sk(sk); 2459 2460 if (frto_undo || tcp_may_undo(tp)) { 2461 tcp_undo_cwnd_reduction(sk, true); 2462 2463 DBGUNDO(sk, "partial loss"); 2464 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO); 2465 if (frto_undo) 2466 NET_INC_STATS_BH(sock_net(sk), 2467 LINUX_MIB_TCPSPURIOUSRTOS); 2468 inet_csk(sk)->icsk_retransmits = 0; 2469 if (frto_undo || tcp_is_sack(tp)) 2470 tcp_set_ca_state(sk, TCP_CA_Open); 2471 return true; 2472 } 2473 return false; 2474} 2475 2476/* The cwnd reduction in CWR and Recovery use the PRR algorithm 2477 * https://datatracker.ietf.org/doc/draft-ietf-tcpm-proportional-rate-reduction/ 2478 * It computes the number of packets to send (sndcnt) based on packets newly 2479 * delivered: 2480 * 1) If the packets in flight is larger than ssthresh, PRR spreads the 2481 * cwnd reductions across a full RTT. 2482 * 2) If packets in flight is lower than ssthresh (such as due to excess 2483 * losses and/or application stalls), do not perform any further cwnd 2484 * reductions, but instead slow start up to ssthresh. 2485 */ 2486static void tcp_init_cwnd_reduction(struct sock *sk) 2487{ 2488 struct tcp_sock *tp = tcp_sk(sk); 2489 2490 tp->high_seq = tp->snd_nxt; 2491 tp->tlp_high_seq = 0; 2492 tp->snd_cwnd_cnt = 0; 2493 tp->prior_cwnd = tp->snd_cwnd; 2494 tp->prr_delivered = 0; 2495 tp->prr_out = 0; 2496 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk); 2497 TCP_ECN_queue_cwr(tp); 2498} 2499 2500static void tcp_cwnd_reduction(struct sock *sk, const int prior_unsacked, 2501 int fast_rexmit) 2502{ 2503 struct tcp_sock *tp = tcp_sk(sk); 2504 int sndcnt = 0; 2505 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp); 2506 int newly_acked_sacked = prior_unsacked - 2507 (tp->packets_out - tp->sacked_out); 2508 2509 tp->prr_delivered += newly_acked_sacked; 2510 if (tcp_packets_in_flight(tp) > tp->snd_ssthresh) { 2511 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered + 2512 tp->prior_cwnd - 1; 2513 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out; 2514 } else { 2515 sndcnt = min_t(int, delta, 2516 max_t(int, tp->prr_delivered - tp->prr_out, 2517 newly_acked_sacked) + 1); 2518 } 2519 2520 sndcnt = max(sndcnt, (fast_rexmit ? 1 : 0)); 2521 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt; 2522} 2523 2524static inline void tcp_end_cwnd_reduction(struct sock *sk) 2525{ 2526 struct tcp_sock *tp = tcp_sk(sk); 2527 2528 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */ 2529 if (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR || 2530 (tp->undo_marker && tp->snd_ssthresh < TCP_INFINITE_SSTHRESH)) { 2531 tp->snd_cwnd = tp->snd_ssthresh; 2532 tp->snd_cwnd_stamp = tcp_time_stamp; 2533 } 2534 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR); 2535} 2536 2537/* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */ 2538void tcp_enter_cwr(struct sock *sk) 2539{ 2540 struct tcp_sock *tp = tcp_sk(sk); 2541 2542 tp->prior_ssthresh = 0; 2543 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) { 2544 tp->undo_marker = 0; 2545 tcp_init_cwnd_reduction(sk); 2546 tcp_set_ca_state(sk, TCP_CA_CWR); 2547 } 2548} 2549 2550static void tcp_try_keep_open(struct sock *sk) 2551{ 2552 struct tcp_sock *tp = tcp_sk(sk); 2553 int state = TCP_CA_Open; 2554 2555 if (tcp_left_out(tp) || tcp_any_retrans_done(sk)) 2556 state = TCP_CA_Disorder; 2557 2558 if (inet_csk(sk)->icsk_ca_state != state) { 2559 tcp_set_ca_state(sk, state); 2560 tp->high_seq = tp->snd_nxt; 2561 } 2562} 2563 2564static void tcp_try_to_open(struct sock *sk, int flag, const int prior_unsacked) 2565{ 2566 struct tcp_sock *tp = tcp_sk(sk); 2567 2568 tcp_verify_left_out(tp); 2569 2570 if (!tcp_any_retrans_done(sk)) 2571 tp->retrans_stamp = 0; 2572 2573 if (flag & FLAG_ECE) 2574 tcp_enter_cwr(sk); 2575 2576 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) { 2577 tcp_try_keep_open(sk); 2578 } else { 2579 tcp_cwnd_reduction(sk, prior_unsacked, 0); 2580 } 2581} 2582 2583static void tcp_mtup_probe_failed(struct sock *sk) 2584{ 2585 struct inet_connection_sock *icsk = inet_csk(sk); 2586 2587 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1; 2588 icsk->icsk_mtup.probe_size = 0; 2589} 2590 2591static void tcp_mtup_probe_success(struct sock *sk) 2592{ 2593 struct tcp_sock *tp = tcp_sk(sk); 2594 struct inet_connection_sock *icsk = inet_csk(sk); 2595 2596 /* FIXME: breaks with very large cwnd */ 2597 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2598 tp->snd_cwnd = tp->snd_cwnd * 2599 tcp_mss_to_mtu(sk, tp->mss_cache) / 2600 icsk->icsk_mtup.probe_size; 2601 tp->snd_cwnd_cnt = 0; 2602 tp->snd_cwnd_stamp = tcp_time_stamp; 2603 tp->snd_ssthresh = tcp_current_ssthresh(sk); 2604 2605 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size; 2606 icsk->icsk_mtup.probe_size = 0; 2607 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 2608} 2609 2610/* Do a simple retransmit without using the backoff mechanisms in 2611 * tcp_timer. This is used for path mtu discovery. 2612 * The socket is already locked here. 2613 */ 2614void tcp_simple_retransmit(struct sock *sk) 2615{ 2616 const struct inet_connection_sock *icsk = inet_csk(sk); 2617 struct tcp_sock *tp = tcp_sk(sk); 2618 struct sk_buff *skb; 2619 unsigned int mss = tcp_current_mss(sk); 2620 u32 prior_lost = tp->lost_out; 2621 2622 tcp_for_write_queue(skb, sk) { 2623 if (skb == tcp_send_head(sk)) 2624 break; 2625 if (tcp_skb_seglen(skb) > mss && 2626 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) { 2627 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 2628 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 2629 tp->retrans_out -= tcp_skb_pcount(skb); 2630 } 2631 tcp_skb_mark_lost_uncond_verify(tp, skb); 2632 } 2633 } 2634 2635 tcp_clear_retrans_hints_partial(tp); 2636 2637 if (prior_lost == tp->lost_out) 2638 return; 2639 2640 if (tcp_is_reno(tp)) 2641 tcp_limit_reno_sacked(tp); 2642 2643 tcp_verify_left_out(tp); 2644 2645 /* Don't muck with the congestion window here. 2646 * Reason is that we do not increase amount of _data_ 2647 * in network, but units changed and effective 2648 * cwnd/ssthresh really reduced now. 2649 */ 2650 if (icsk->icsk_ca_state != TCP_CA_Loss) { 2651 tp->high_seq = tp->snd_nxt; 2652 tp->snd_ssthresh = tcp_current_ssthresh(sk); 2653 tp->prior_ssthresh = 0; 2654 tp->undo_marker = 0; 2655 tcp_set_ca_state(sk, TCP_CA_Loss); 2656 } 2657 tcp_xmit_retransmit_queue(sk); 2658} 2659EXPORT_SYMBOL(tcp_simple_retransmit); 2660 2661static void tcp_enter_recovery(struct sock *sk, bool ece_ack) 2662{ 2663 struct tcp_sock *tp = tcp_sk(sk); 2664 int mib_idx; 2665 2666 if (tcp_is_reno(tp)) 2667 mib_idx = LINUX_MIB_TCPRENORECOVERY; 2668 else 2669 mib_idx = LINUX_MIB_TCPSACKRECOVERY; 2670 2671 NET_INC_STATS_BH(sock_net(sk), mib_idx); 2672 2673 tp->prior_ssthresh = 0; 2674 tp->undo_marker = tp->snd_una; 2675 tp->undo_retrans = tp->retrans_out ? : -1; 2676 2677 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) { 2678 if (!ece_ack) 2679 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2680 tcp_init_cwnd_reduction(sk); 2681 } 2682 tcp_set_ca_state(sk, TCP_CA_Recovery); 2683} 2684 2685/* Process an ACK in CA_Loss state. Move to CA_Open if lost data are 2686 * recovered or spurious. Otherwise retransmits more on partial ACKs. 2687 */ 2688static void tcp_process_loss(struct sock *sk, int flag, bool is_dupack) 2689{ 2690 struct inet_connection_sock *icsk = inet_csk(sk); 2691 struct tcp_sock *tp = tcp_sk(sk); 2692 bool recovered = !before(tp->snd_una, tp->high_seq); 2693 2694 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */ 2695 /* Step 3.b. A timeout is spurious if not all data are 2696 * lost, i.e., never-retransmitted data are (s)acked. 2697 */ 2698 if (tcp_try_undo_loss(sk, flag & FLAG_ORIG_SACK_ACKED)) 2699 return; 2700 2701 if (after(tp->snd_nxt, tp->high_seq) && 2702 (flag & FLAG_DATA_SACKED || is_dupack)) { 2703 tp->frto = 0; /* Loss was real: 2nd part of step 3.a */ 2704 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) { 2705 tp->high_seq = tp->snd_nxt; 2706 __tcp_push_pending_frames(sk, tcp_current_mss(sk), 2707 TCP_NAGLE_OFF); 2708 if (after(tp->snd_nxt, tp->high_seq)) 2709 return; /* Step 2.b */ 2710 tp->frto = 0; 2711 } 2712 } 2713 2714 if (recovered) { 2715 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */ 2716 icsk->icsk_retransmits = 0; 2717 tcp_try_undo_recovery(sk); 2718 return; 2719 } 2720 if (flag & FLAG_DATA_ACKED) 2721 icsk->icsk_retransmits = 0; 2722 if (tcp_is_reno(tp)) { 2723 /* A Reno DUPACK means new data in F-RTO step 2.b above are 2724 * delivered. Lower inflight to clock out (re)tranmissions. 2725 */ 2726 if (after(tp->snd_nxt, tp->high_seq) && is_dupack) 2727 tcp_add_reno_sack(sk); 2728 else if (flag & FLAG_SND_UNA_ADVANCED) 2729 tcp_reset_reno_sack(tp); 2730 } 2731 if (tcp_try_undo_loss(sk, false)) 2732 return; 2733 tcp_xmit_retransmit_queue(sk); 2734} 2735 2736/* Undo during fast recovery after partial ACK. */ 2737static bool tcp_try_undo_partial(struct sock *sk, const int acked, 2738 const int prior_unsacked) 2739{ 2740 struct tcp_sock *tp = tcp_sk(sk); 2741 2742 if (tp->undo_marker && tcp_packet_delayed(tp)) { 2743 /* Plain luck! Hole if filled with delayed 2744 * packet, rather than with a retransmit. 2745 */ 2746 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1); 2747 2748 /* We are getting evidence that the reordering degree is higher 2749 * than we realized. If there are no retransmits out then we 2750 * can undo. Otherwise we clock out new packets but do not 2751 * mark more packets lost or retransmit more. 2752 */ 2753 if (tp->retrans_out) { 2754 tcp_cwnd_reduction(sk, prior_unsacked, 0); 2755 return true; 2756 } 2757 2758 if (!tcp_any_retrans_done(sk)) 2759 tp->retrans_stamp = 0; 2760 2761 DBGUNDO(sk, "partial recovery"); 2762 tcp_undo_cwnd_reduction(sk, true); 2763 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO); 2764 tcp_try_keep_open(sk); 2765 return true; 2766 } 2767 return false; 2768} 2769 2770/* Process an event, which can update packets-in-flight not trivially. 2771 * Main goal of this function is to calculate new estimate for left_out, 2772 * taking into account both packets sitting in receiver's buffer and 2773 * packets lost by network. 2774 * 2775 * Besides that it does CWND reduction, when packet loss is detected 2776 * and changes state of machine. 2777 * 2778 * It does _not_ decide what to send, it is made in function 2779 * tcp_xmit_retransmit_queue(). 2780 */ 2781static void tcp_fastretrans_alert(struct sock *sk, const int acked, 2782 const int prior_unsacked, 2783 bool is_dupack, int flag) 2784{ 2785 struct inet_connection_sock *icsk = inet_csk(sk); 2786 struct tcp_sock *tp = tcp_sk(sk); 2787 bool do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) && 2788 (tcp_fackets_out(tp) > tp->reordering)); 2789 int fast_rexmit = 0; 2790 2791 if (WARN_ON(!tp->packets_out && tp->sacked_out)) 2792 tp->sacked_out = 0; 2793 if (WARN_ON(!tp->sacked_out && tp->fackets_out)) 2794 tp->fackets_out = 0; 2795 2796 /* Now state machine starts. 2797 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */ 2798 if (flag & FLAG_ECE) 2799 tp->prior_ssthresh = 0; 2800 2801 /* B. In all the states check for reneging SACKs. */ 2802 if (tcp_check_sack_reneging(sk, flag)) 2803 return; 2804 2805 /* C. Check consistency of the current state. */ 2806 tcp_verify_left_out(tp); 2807 2808 /* D. Check state exit conditions. State can be terminated 2809 * when high_seq is ACKed. */ 2810 if (icsk->icsk_ca_state == TCP_CA_Open) { 2811 WARN_ON(tp->retrans_out != 0); 2812 tp->retrans_stamp = 0; 2813 } else if (!before(tp->snd_una, tp->high_seq)) { 2814 switch (icsk->icsk_ca_state) { 2815 case TCP_CA_CWR: 2816 /* CWR is to be held something *above* high_seq 2817 * is ACKed for CWR bit to reach receiver. */ 2818 if (tp->snd_una != tp->high_seq) { 2819 tcp_end_cwnd_reduction(sk); 2820 tcp_set_ca_state(sk, TCP_CA_Open); 2821 } 2822 break; 2823 2824 case TCP_CA_Recovery: 2825 if (tcp_is_reno(tp)) 2826 tcp_reset_reno_sack(tp); 2827 if (tcp_try_undo_recovery(sk)) 2828 return; 2829 tcp_end_cwnd_reduction(sk); 2830 break; 2831 } 2832 } 2833 2834 /* E. Process state. */ 2835 switch (icsk->icsk_ca_state) { 2836 case TCP_CA_Recovery: 2837 if (!(flag & FLAG_SND_UNA_ADVANCED)) { 2838 if (tcp_is_reno(tp) && is_dupack) 2839 tcp_add_reno_sack(sk); 2840 } else { 2841 if (tcp_try_undo_partial(sk, acked, prior_unsacked)) 2842 return; 2843 /* Partial ACK arrived. Force fast retransmit. */ 2844 do_lost = tcp_is_reno(tp) || 2845 tcp_fackets_out(tp) > tp->reordering; 2846 } 2847 if (tcp_try_undo_dsack(sk)) { 2848 tcp_try_keep_open(sk); 2849 return; 2850 } 2851 break; 2852 case TCP_CA_Loss: 2853 tcp_process_loss(sk, flag, is_dupack); 2854 if (icsk->icsk_ca_state != TCP_CA_Open) 2855 return; 2856 /* Fall through to processing in Open state. */ 2857 default: 2858 if (tcp_is_reno(tp)) { 2859 if (flag & FLAG_SND_UNA_ADVANCED) 2860 tcp_reset_reno_sack(tp); 2861 if (is_dupack) 2862 tcp_add_reno_sack(sk); 2863 } 2864 2865 if (icsk->icsk_ca_state <= TCP_CA_Disorder) 2866 tcp_try_undo_dsack(sk); 2867 2868 if (!tcp_time_to_recover(sk, flag)) { 2869 tcp_try_to_open(sk, flag, prior_unsacked); 2870 return; 2871 } 2872 2873 /* MTU probe failure: don't reduce cwnd */ 2874 if (icsk->icsk_ca_state < TCP_CA_CWR && 2875 icsk->icsk_mtup.probe_size && 2876 tp->snd_una == tp->mtu_probe.probe_seq_start) { 2877 tcp_mtup_probe_failed(sk); 2878 /* Restores the reduction we did in tcp_mtup_probe() */ 2879 tp->snd_cwnd++; 2880 tcp_simple_retransmit(sk); 2881 return; 2882 } 2883 2884 /* Otherwise enter Recovery state */ 2885 tcp_enter_recovery(sk, (flag & FLAG_ECE)); 2886 fast_rexmit = 1; 2887 } 2888 2889 if (do_lost) 2890 tcp_update_scoreboard(sk, fast_rexmit); 2891 tcp_cwnd_reduction(sk, prior_unsacked, fast_rexmit); 2892 tcp_xmit_retransmit_queue(sk); 2893} 2894 2895static inline bool tcp_ack_update_rtt(struct sock *sk, const int flag, 2896 long seq_rtt_us, long sack_rtt_us) 2897{ 2898 const struct tcp_sock *tp = tcp_sk(sk); 2899 2900 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because 2901 * broken middle-boxes or peers may corrupt TS-ECR fields. But 2902 * Karn's algorithm forbids taking RTT if some retransmitted data 2903 * is acked (RFC6298). 2904 */ 2905 if (flag & FLAG_RETRANS_DATA_ACKED) 2906 seq_rtt_us = -1L; 2907 2908 if (seq_rtt_us < 0) 2909 seq_rtt_us = sack_rtt_us; 2910 2911 /* RTTM Rule: A TSecr value received in a segment is used to 2912 * update the averaged RTT measurement only if the segment 2913 * acknowledges some new data, i.e., only if it advances the 2914 * left edge of the send window. 2915 * See draft-ietf-tcplw-high-performance-00, section 3.3. 2916 */ 2917 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 2918 flag & FLAG_ACKED) 2919 seq_rtt_us = jiffies_to_usecs(tcp_time_stamp - tp->rx_opt.rcv_tsecr); 2920 2921 if (seq_rtt_us < 0) 2922 return false; 2923 2924 tcp_rtt_estimator(sk, seq_rtt_us); 2925 tcp_set_rto(sk); 2926 2927 /* RFC6298: only reset backoff on valid RTT measurement. */ 2928 inet_csk(sk)->icsk_backoff = 0; 2929 return true; 2930} 2931 2932/* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */ 2933static void tcp_synack_rtt_meas(struct sock *sk, const u32 synack_stamp) 2934{ 2935 struct tcp_sock *tp = tcp_sk(sk); 2936 long seq_rtt_us = -1L; 2937 2938 if (synack_stamp && !tp->total_retrans) 2939 seq_rtt_us = jiffies_to_usecs(tcp_time_stamp - synack_stamp); 2940 2941 /* If the ACK acks both the SYNACK and the (Fast Open'd) data packets 2942 * sent in SYN_RECV, SYNACK RTT is the smooth RTT computed in tcp_ack() 2943 */ 2944 if (!tp->srtt_us) 2945 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, seq_rtt_us, -1L); 2946} 2947 2948static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked) 2949{ 2950 const struct inet_connection_sock *icsk = inet_csk(sk); 2951 2952 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked); 2953 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp; 2954} 2955 2956/* Restart timer after forward progress on connection. 2957 * RFC2988 recommends to restart timer to now+rto. 2958 */ 2959void tcp_rearm_rto(struct sock *sk) 2960{ 2961 const struct inet_connection_sock *icsk = inet_csk(sk); 2962 struct tcp_sock *tp = tcp_sk(sk); 2963 2964 /* If the retrans timer is currently being used by Fast Open 2965 * for SYN-ACK retrans purpose, stay put. 2966 */ 2967 if (tp->fastopen_rsk) 2968 return; 2969 2970 if (!tp->packets_out) { 2971 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS); 2972 } else { 2973 u32 rto = inet_csk(sk)->icsk_rto; 2974 /* Offset the time elapsed after installing regular RTO */ 2975 if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS || 2976 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) { 2977 struct sk_buff *skb = tcp_write_queue_head(sk); 2978 const u32 rto_time_stamp = TCP_SKB_CB(skb)->when + rto; 2979 s32 delta = (s32)(rto_time_stamp - tcp_time_stamp); 2980 /* delta may not be positive if the socket is locked 2981 * when the retrans timer fires and is rescheduled. 2982 */ 2983 if (delta > 0) 2984 rto = delta; 2985 } 2986 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto, 2987 TCP_RTO_MAX); 2988 } 2989} 2990 2991/* This function is called when the delayed ER timer fires. TCP enters 2992 * fast recovery and performs fast-retransmit. 2993 */ 2994void tcp_resume_early_retransmit(struct sock *sk) 2995{ 2996 struct tcp_sock *tp = tcp_sk(sk); 2997 2998 tcp_rearm_rto(sk); 2999 3000 /* Stop if ER is disabled after the delayed ER timer is scheduled */ 3001 if (!tp->do_early_retrans) 3002 return; 3003 3004 tcp_enter_recovery(sk, false); 3005 tcp_update_scoreboard(sk, 1); 3006 tcp_xmit_retransmit_queue(sk); 3007} 3008 3009/* If we get here, the whole TSO packet has not been acked. */ 3010static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb) 3011{ 3012 struct tcp_sock *tp = tcp_sk(sk); 3013 u32 packets_acked; 3014 3015 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)); 3016 3017 packets_acked = tcp_skb_pcount(skb); 3018 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 3019 return 0; 3020 packets_acked -= tcp_skb_pcount(skb); 3021 3022 if (packets_acked) { 3023 BUG_ON(tcp_skb_pcount(skb) == 0); 3024 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)); 3025 } 3026 3027 return packets_acked; 3028} 3029 3030/* Remove acknowledged frames from the retransmission queue. If our packet 3031 * is before the ack sequence we can discard it as it's confirmed to have 3032 * arrived at the other end. 3033 */ 3034static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets, 3035 u32 prior_snd_una, long sack_rtt_us) 3036{ 3037 const struct inet_connection_sock *icsk = inet_csk(sk); 3038 struct skb_mstamp first_ackt, last_ackt, now; 3039 struct tcp_sock *tp = tcp_sk(sk); 3040 u32 prior_sacked = tp->sacked_out; 3041 u32 reord = tp->packets_out; 3042 bool fully_acked = true; 3043 long ca_seq_rtt_us = -1L; 3044 long seq_rtt_us = -1L; 3045 struct sk_buff *skb; 3046 u32 pkts_acked = 0; 3047 bool rtt_update; 3048 int flag = 0; 3049 3050 first_ackt.v64 = 0; 3051 3052 while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) { 3053 struct skb_shared_info *shinfo = skb_shinfo(skb); 3054 struct tcp_skb_cb *scb = TCP_SKB_CB(skb); 3055 u8 sacked = scb->sacked; 3056 u32 acked_pcount; 3057 3058 if (unlikely(shinfo->tx_flags & SKBTX_ACK_TSTAMP) && 3059 between(shinfo->tskey, prior_snd_una, tp->snd_una - 1)) 3060 __skb_tstamp_tx(skb, NULL, sk, SCM_TSTAMP_ACK); 3061 3062 /* Determine how many packets and what bytes were acked, tso and else */ 3063 if (after(scb->end_seq, tp->snd_una)) { 3064 if (tcp_skb_pcount(skb) == 1 || 3065 !after(tp->snd_una, scb->seq)) 3066 break; 3067 3068 acked_pcount = tcp_tso_acked(sk, skb); 3069 if (!acked_pcount) 3070 break; 3071 3072 fully_acked = false; 3073 } else { 3074 acked_pcount = tcp_skb_pcount(skb); 3075 } 3076 3077 if (sacked & TCPCB_RETRANS) { 3078 if (sacked & TCPCB_SACKED_RETRANS) 3079 tp->retrans_out -= acked_pcount; 3080 flag |= FLAG_RETRANS_DATA_ACKED; 3081 } else { 3082 last_ackt = skb->skb_mstamp; 3083 WARN_ON_ONCE(last_ackt.v64 == 0); 3084 if (!first_ackt.v64) 3085 first_ackt = last_ackt; 3086 3087 if (!(sacked & TCPCB_SACKED_ACKED)) 3088 reord = min(pkts_acked, reord); 3089 if (!after(scb->end_seq, tp->high_seq)) 3090 flag |= FLAG_ORIG_SACK_ACKED; 3091 } 3092 3093 if (sacked & TCPCB_SACKED_ACKED) 3094 tp->sacked_out -= acked_pcount; 3095 if (sacked & TCPCB_LOST) 3096 tp->lost_out -= acked_pcount; 3097 3098 tp->packets_out -= acked_pcount; 3099 pkts_acked += acked_pcount; 3100 3101 /* Initial outgoing SYN's get put onto the write_queue 3102 * just like anything else we transmit. It is not 3103 * true data, and if we misinform our callers that 3104 * this ACK acks real data, we will erroneously exit 3105 * connection startup slow start one packet too 3106 * quickly. This is severely frowned upon behavior. 3107 */ 3108 if (!(scb->tcp_flags & TCPHDR_SYN)) { 3109 flag |= FLAG_DATA_ACKED; 3110 } else { 3111 flag |= FLAG_SYN_ACKED; 3112 tp->retrans_stamp = 0; 3113 } 3114 3115 if (!fully_acked) 3116 break; 3117 3118 tcp_unlink_write_queue(skb, sk); 3119 sk_wmem_free_skb(sk, skb); 3120 if (skb == tp->retransmit_skb_hint) 3121 tp->retransmit_skb_hint = NULL; 3122 if (skb == tp->lost_skb_hint) 3123 tp->lost_skb_hint = NULL; 3124 } 3125 3126 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una))) 3127 tp->snd_up = tp->snd_una; 3128 3129 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) 3130 flag |= FLAG_SACK_RENEGING; 3131 3132 skb_mstamp_get(&now); 3133 if (first_ackt.v64) { 3134 seq_rtt_us = skb_mstamp_us_delta(&now, &first_ackt); 3135 ca_seq_rtt_us = skb_mstamp_us_delta(&now, &last_ackt); 3136 } 3137 3138 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us); 3139 3140 if (flag & FLAG_ACKED) { 3141 const struct tcp_congestion_ops *ca_ops 3142 = inet_csk(sk)->icsk_ca_ops; 3143 3144 tcp_rearm_rto(sk); 3145 if (unlikely(icsk->icsk_mtup.probe_size && 3146 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) { 3147 tcp_mtup_probe_success(sk); 3148 } 3149 3150 if (tcp_is_reno(tp)) { 3151 tcp_remove_reno_sacks(sk, pkts_acked); 3152 } else { 3153 int delta; 3154 3155 /* Non-retransmitted hole got filled? That's reordering */ 3156 if (reord < prior_fackets) 3157 tcp_update_reordering(sk, tp->fackets_out - reord, 0); 3158 3159 delta = tcp_is_fack(tp) ? pkts_acked : 3160 prior_sacked - tp->sacked_out; 3161 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta); 3162 } 3163 3164 tp->fackets_out -= min(pkts_acked, tp->fackets_out); 3165 3166 if (ca_ops->pkts_acked) 3167 ca_ops->pkts_acked(sk, pkts_acked, ca_seq_rtt_us); 3168 3169 } else if (skb && rtt_update && sack_rtt_us >= 0 && 3170 sack_rtt_us > skb_mstamp_us_delta(&now, &skb->skb_mstamp)) { 3171 /* Do not re-arm RTO if the sack RTT is measured from data sent 3172 * after when the head was last (re)transmitted. Otherwise the 3173 * timeout may continue to extend in loss recovery. 3174 */ 3175 tcp_rearm_rto(sk); 3176 } 3177 3178#if FASTRETRANS_DEBUG > 0 3179 WARN_ON((int)tp->sacked_out < 0); 3180 WARN_ON((int)tp->lost_out < 0); 3181 WARN_ON((int)tp->retrans_out < 0); 3182 if (!tp->packets_out && tcp_is_sack(tp)) { 3183 icsk = inet_csk(sk); 3184 if (tp->lost_out) { 3185 pr_debug("Leak l=%u %d\n", 3186 tp->lost_out, icsk->icsk_ca_state); 3187 tp->lost_out = 0; 3188 } 3189 if (tp->sacked_out) { 3190 pr_debug("Leak s=%u %d\n", 3191 tp->sacked_out, icsk->icsk_ca_state); 3192 tp->sacked_out = 0; 3193 } 3194 if (tp->retrans_out) { 3195 pr_debug("Leak r=%u %d\n", 3196 tp->retrans_out, icsk->icsk_ca_state); 3197 tp->retrans_out = 0; 3198 } 3199 } 3200#endif 3201 return flag; 3202} 3203 3204static void tcp_ack_probe(struct sock *sk) 3205{ 3206 const struct tcp_sock *tp = tcp_sk(sk); 3207 struct inet_connection_sock *icsk = inet_csk(sk); 3208 3209 /* Was it a usable window open? */ 3210 3211 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) { 3212 icsk->icsk_backoff = 0; 3213 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0); 3214 /* Socket must be waked up by subsequent tcp_data_snd_check(). 3215 * This function is not for random using! 3216 */ 3217 } else { 3218 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 3219 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), 3220 TCP_RTO_MAX); 3221 } 3222} 3223 3224static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag) 3225{ 3226 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) || 3227 inet_csk(sk)->icsk_ca_state != TCP_CA_Open; 3228} 3229 3230/* Decide wheather to run the increase function of congestion control. */ 3231static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag) 3232{ 3233 if (tcp_in_cwnd_reduction(sk)) 3234 return false; 3235 3236 /* If reordering is high then always grow cwnd whenever data is 3237 * delivered regardless of its ordering. Otherwise stay conservative 3238 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/ 3239 * new SACK or ECE mark may first advance cwnd here and later reduce 3240 * cwnd in tcp_fastretrans_alert() based on more states. 3241 */ 3242 if (tcp_sk(sk)->reordering > sysctl_tcp_reordering) 3243 return flag & FLAG_FORWARD_PROGRESS; 3244 3245 return flag & FLAG_DATA_ACKED; 3246} 3247 3248/* Check that window update is acceptable. 3249 * The function assumes that snd_una<=ack<=snd_next. 3250 */ 3251static inline bool tcp_may_update_window(const struct tcp_sock *tp, 3252 const u32 ack, const u32 ack_seq, 3253 const u32 nwin) 3254{ 3255 return after(ack, tp->snd_una) || 3256 after(ack_seq, tp->snd_wl1) || 3257 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd); 3258} 3259 3260/* Update our send window. 3261 * 3262 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2 3263 * and in FreeBSD. NetBSD's one is even worse.) is wrong. 3264 */ 3265static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack, 3266 u32 ack_seq) 3267{ 3268 struct tcp_sock *tp = tcp_sk(sk); 3269 int flag = 0; 3270 u32 nwin = ntohs(tcp_hdr(skb)->window); 3271 3272 if (likely(!tcp_hdr(skb)->syn)) 3273 nwin <<= tp->rx_opt.snd_wscale; 3274 3275 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) { 3276 flag |= FLAG_WIN_UPDATE; 3277 tcp_update_wl(tp, ack_seq); 3278 3279 if (tp->snd_wnd != nwin) { 3280 tp->snd_wnd = nwin; 3281 3282 /* Note, it is the only place, where 3283 * fast path is recovered for sending TCP. 3284 */ 3285 tp->pred_flags = 0; 3286 tcp_fast_path_check(sk); 3287 3288 if (nwin > tp->max_window) { 3289 tp->max_window = nwin; 3290 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie); 3291 } 3292 } 3293 } 3294 3295 tp->snd_una = ack; 3296 3297 return flag; 3298} 3299 3300/* RFC 5961 7 [ACK Throttling] */ 3301static void tcp_send_challenge_ack(struct sock *sk) 3302{ 3303 /* unprotected vars, we dont care of overwrites */ 3304 static u32 challenge_timestamp; 3305 static unsigned int challenge_count; 3306 u32 now = jiffies / HZ; 3307 3308 if (now != challenge_timestamp) { 3309 challenge_timestamp = now; 3310 challenge_count = 0; 3311 } 3312 if (++challenge_count <= sysctl_tcp_challenge_ack_limit) { 3313 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPCHALLENGEACK); 3314 tcp_send_ack(sk); 3315 } 3316} 3317 3318static void tcp_store_ts_recent(struct tcp_sock *tp) 3319{ 3320 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval; 3321 tp->rx_opt.ts_recent_stamp = get_seconds(); 3322} 3323 3324static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq) 3325{ 3326 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) { 3327 /* PAWS bug workaround wrt. ACK frames, the PAWS discard 3328 * extra check below makes sure this can only happen 3329 * for pure ACK frames. -DaveM 3330 * 3331 * Not only, also it occurs for expired timestamps. 3332 */ 3333 3334 if (tcp_paws_check(&tp->rx_opt, 0)) 3335 tcp_store_ts_recent(tp); 3336 } 3337} 3338 3339/* This routine deals with acks during a TLP episode. 3340 * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe. 3341 */ 3342static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag) 3343{ 3344 struct tcp_sock *tp = tcp_sk(sk); 3345 bool is_tlp_dupack = (ack == tp->tlp_high_seq) && 3346 !(flag & (FLAG_SND_UNA_ADVANCED | 3347 FLAG_NOT_DUP | FLAG_DATA_SACKED)); 3348 3349 /* Mark the end of TLP episode on receiving TLP dupack or when 3350 * ack is after tlp_high_seq. 3351 */ 3352 if (is_tlp_dupack) { 3353 tp->tlp_high_seq = 0; 3354 return; 3355 } 3356 3357 if (after(ack, tp->tlp_high_seq)) { 3358 tp->tlp_high_seq = 0; 3359 /* Don't reduce cwnd if DSACK arrives for TLP retrans. */ 3360 if (!(flag & FLAG_DSACKING_ACK)) { 3361 tcp_init_cwnd_reduction(sk); 3362 tcp_set_ca_state(sk, TCP_CA_CWR); 3363 tcp_end_cwnd_reduction(sk); 3364 tcp_try_keep_open(sk); 3365 NET_INC_STATS_BH(sock_net(sk), 3366 LINUX_MIB_TCPLOSSPROBERECOVERY); 3367 } 3368 } 3369} 3370 3371/* This routine deals with incoming acks, but not outgoing ones. */ 3372static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag) 3373{ 3374 struct inet_connection_sock *icsk = inet_csk(sk); 3375 struct tcp_sock *tp = tcp_sk(sk); 3376 u32 prior_snd_una = tp->snd_una; 3377 u32 ack_seq = TCP_SKB_CB(skb)->seq; 3378 u32 ack = TCP_SKB_CB(skb)->ack_seq; 3379 bool is_dupack = false; 3380 u32 prior_fackets; 3381 int prior_packets = tp->packets_out; 3382 const int prior_unsacked = tp->packets_out - tp->sacked_out; 3383 int acked = 0; /* Number of packets newly acked */ 3384 long sack_rtt_us = -1L; 3385 3386 /* If the ack is older than previous acks 3387 * then we can probably ignore it. 3388 */ 3389 if (before(ack, prior_snd_una)) { 3390 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */ 3391 if (before(ack, prior_snd_una - tp->max_window)) { 3392 tcp_send_challenge_ack(sk); 3393 return -1; 3394 } 3395 goto old_ack; 3396 } 3397 3398 /* If the ack includes data we haven't sent yet, discard 3399 * this segment (RFC793 Section 3.9). 3400 */ 3401 if (after(ack, tp->snd_nxt)) 3402 goto invalid_ack; 3403 3404 if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS || 3405 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) 3406 tcp_rearm_rto(sk); 3407 3408 if (after(ack, prior_snd_una)) 3409 flag |= FLAG_SND_UNA_ADVANCED; 3410 3411 prior_fackets = tp->fackets_out; 3412 3413 /* ts_recent update must be made after we are sure that the packet 3414 * is in window. 3415 */ 3416 if (flag & FLAG_UPDATE_TS_RECENT) 3417 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq); 3418 3419 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) { 3420 /* Window is constant, pure forward advance. 3421 * No more checks are required. 3422 * Note, we use the fact that SND.UNA>=SND.WL2. 3423 */ 3424 tcp_update_wl(tp, ack_seq); 3425 tp->snd_una = ack; 3426 flag |= FLAG_WIN_UPDATE; 3427 3428 tcp_ca_event(sk, CA_EVENT_FAST_ACK); 3429 3430 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS); 3431 } else { 3432 if (ack_seq != TCP_SKB_CB(skb)->end_seq) 3433 flag |= FLAG_DATA; 3434 else 3435 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS); 3436 3437 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq); 3438 3439 if (TCP_SKB_CB(skb)->sacked) 3440 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una, 3441 &sack_rtt_us); 3442 3443 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb))) 3444 flag |= FLAG_ECE; 3445 3446 tcp_ca_event(sk, CA_EVENT_SLOW_ACK); 3447 } 3448 3449 /* We passed data and got it acked, remove any soft error 3450 * log. Something worked... 3451 */ 3452 sk->sk_err_soft = 0; 3453 icsk->icsk_probes_out = 0; 3454 tp->rcv_tstamp = tcp_time_stamp; 3455 if (!prior_packets) 3456 goto no_queue; 3457 3458 /* See if we can take anything off of the retransmit queue. */ 3459 acked = tp->packets_out; 3460 flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una, 3461 sack_rtt_us); 3462 acked -= tp->packets_out; 3463 3464 /* Advance cwnd if state allows */ 3465 if (tcp_may_raise_cwnd(sk, flag)) 3466 tcp_cong_avoid(sk, ack, acked); 3467 3468 if (tcp_ack_is_dubious(sk, flag)) { 3469 is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP)); 3470 tcp_fastretrans_alert(sk, acked, prior_unsacked, 3471 is_dupack, flag); 3472 } 3473 if (tp->tlp_high_seq) 3474 tcp_process_tlp_ack(sk, ack, flag); 3475 3476 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) { 3477 struct dst_entry *dst = __sk_dst_get(sk); 3478 if (dst) 3479 dst_confirm(dst); 3480 } 3481 3482 if (icsk->icsk_pending == ICSK_TIME_RETRANS) 3483 tcp_schedule_loss_probe(sk); 3484 tcp_update_pacing_rate(sk); 3485 return 1; 3486 3487no_queue: 3488 /* If data was DSACKed, see if we can undo a cwnd reduction. */ 3489 if (flag & FLAG_DSACKING_ACK) 3490 tcp_fastretrans_alert(sk, acked, prior_unsacked, 3491 is_dupack, flag); 3492 /* If this ack opens up a zero window, clear backoff. It was 3493 * being used to time the probes, and is probably far higher than 3494 * it needs to be for normal retransmission. 3495 */ 3496 if (tcp_send_head(sk)) 3497 tcp_ack_probe(sk); 3498 3499 if (tp->tlp_high_seq) 3500 tcp_process_tlp_ack(sk, ack, flag); 3501 return 1; 3502 3503invalid_ack: 3504 SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt); 3505 return -1; 3506 3507old_ack: 3508 /* If data was SACKed, tag it and see if we should send more data. 3509 * If data was DSACKed, see if we can undo a cwnd reduction. 3510 */ 3511 if (TCP_SKB_CB(skb)->sacked) { 3512 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una, 3513 &sack_rtt_us); 3514 tcp_fastretrans_alert(sk, acked, prior_unsacked, 3515 is_dupack, flag); 3516 } 3517 3518 SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt); 3519 return 0; 3520} 3521 3522/* Look for tcp options. Normally only called on SYN and SYNACK packets. 3523 * But, this can also be called on packets in the established flow when 3524 * the fast version below fails. 3525 */ 3526void tcp_parse_options(const struct sk_buff *skb, 3527 struct tcp_options_received *opt_rx, int estab, 3528 struct tcp_fastopen_cookie *foc) 3529{ 3530 const unsigned char *ptr; 3531 const struct tcphdr *th = tcp_hdr(skb); 3532 int length = (th->doff * 4) - sizeof(struct tcphdr); 3533 3534 ptr = (const unsigned char *)(th + 1); 3535 opt_rx->saw_tstamp = 0; 3536 3537 while (length > 0) { 3538 int opcode = *ptr++; 3539 int opsize; 3540 3541 switch (opcode) { 3542 case TCPOPT_EOL: 3543 return; 3544 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */ 3545 length--; 3546 continue; 3547 default: 3548 opsize = *ptr++; 3549 if (opsize < 2) /* "silly options" */ 3550 return; 3551 if (opsize > length) 3552 return; /* don't parse partial options */ 3553 switch (opcode) { 3554 case TCPOPT_MSS: 3555 if (opsize == TCPOLEN_MSS && th->syn && !estab) { 3556 u16 in_mss = get_unaligned_be16(ptr); 3557 if (in_mss) { 3558 if (opt_rx->user_mss && 3559 opt_rx->user_mss < in_mss) 3560 in_mss = opt_rx->user_mss; 3561 opt_rx->mss_clamp = in_mss; 3562 } 3563 } 3564 break; 3565 case TCPOPT_WINDOW: 3566 if (opsize == TCPOLEN_WINDOW && th->syn && 3567 !estab && sysctl_tcp_window_scaling) { 3568 __u8 snd_wscale = *(__u8 *)ptr; 3569 opt_rx->wscale_ok = 1; 3570 if (snd_wscale > 14) { 3571 net_info_ratelimited("%s: Illegal window scaling value %d >14 received\n", 3572 __func__, 3573 snd_wscale); 3574 snd_wscale = 14; 3575 } 3576 opt_rx->snd_wscale = snd_wscale; 3577 } 3578 break; 3579 case TCPOPT_TIMESTAMP: 3580 if ((opsize == TCPOLEN_TIMESTAMP) && 3581 ((estab && opt_rx->tstamp_ok) || 3582 (!estab && sysctl_tcp_timestamps))) { 3583 opt_rx->saw_tstamp = 1; 3584 opt_rx->rcv_tsval = get_unaligned_be32(ptr); 3585 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4); 3586 } 3587 break; 3588 case TCPOPT_SACK_PERM: 3589 if (opsize == TCPOLEN_SACK_PERM && th->syn && 3590 !estab && sysctl_tcp_sack) { 3591 opt_rx->sack_ok = TCP_SACK_SEEN; 3592 tcp_sack_reset(opt_rx); 3593 } 3594 break; 3595 3596 case TCPOPT_SACK: 3597 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) && 3598 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) && 3599 opt_rx->sack_ok) { 3600 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th; 3601 } 3602 break; 3603#ifdef CONFIG_TCP_MD5SIG 3604 case TCPOPT_MD5SIG: 3605 /* 3606 * The MD5 Hash has already been 3607 * checked (see tcp_v{4,6}_do_rcv()). 3608 */ 3609 break; 3610#endif 3611 case TCPOPT_EXP: 3612 /* Fast Open option shares code 254 using a 3613 * 16 bits magic number. It's valid only in 3614 * SYN or SYN-ACK with an even size. 3615 */ 3616 if (opsize < TCPOLEN_EXP_FASTOPEN_BASE || 3617 get_unaligned_be16(ptr) != TCPOPT_FASTOPEN_MAGIC || 3618 foc == NULL || !th->syn || (opsize & 1)) 3619 break; 3620 foc->len = opsize - TCPOLEN_EXP_FASTOPEN_BASE; 3621 if (foc->len >= TCP_FASTOPEN_COOKIE_MIN && 3622 foc->len <= TCP_FASTOPEN_COOKIE_MAX) 3623 memcpy(foc->val, ptr + 2, foc->len); 3624 else if (foc->len != 0) 3625 foc->len = -1; 3626 break; 3627 3628 } 3629 ptr += opsize-2; 3630 length -= opsize; 3631 } 3632 } 3633} 3634EXPORT_SYMBOL(tcp_parse_options); 3635 3636static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th) 3637{ 3638 const __be32 *ptr = (const __be32 *)(th + 1); 3639 3640 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) 3641 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) { 3642 tp->rx_opt.saw_tstamp = 1; 3643 ++ptr; 3644 tp->rx_opt.rcv_tsval = ntohl(*ptr); 3645 ++ptr; 3646 if (*ptr) 3647 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset; 3648 else 3649 tp->rx_opt.rcv_tsecr = 0; 3650 return true; 3651 } 3652 return false; 3653} 3654 3655/* Fast parse options. This hopes to only see timestamps. 3656 * If it is wrong it falls back on tcp_parse_options(). 3657 */ 3658static bool tcp_fast_parse_options(const struct sk_buff *skb, 3659 const struct tcphdr *th, struct tcp_sock *tp) 3660{ 3661 /* In the spirit of fast parsing, compare doff directly to constant 3662 * values. Because equality is used, short doff can be ignored here. 3663 */ 3664 if (th->doff == (sizeof(*th) / 4)) { 3665 tp->rx_opt.saw_tstamp = 0; 3666 return false; 3667 } else if (tp->rx_opt.tstamp_ok && 3668 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) { 3669 if (tcp_parse_aligned_timestamp(tp, th)) 3670 return true; 3671 } 3672 3673 tcp_parse_options(skb, &tp->rx_opt, 1, NULL); 3674 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr) 3675 tp->rx_opt.rcv_tsecr -= tp->tsoffset; 3676 3677 return true; 3678} 3679 3680#ifdef CONFIG_TCP_MD5SIG 3681/* 3682 * Parse MD5 Signature option 3683 */ 3684const u8 *tcp_parse_md5sig_option(const struct tcphdr *th) 3685{ 3686 int length = (th->doff << 2) - sizeof(*th); 3687 const u8 *ptr = (const u8 *)(th + 1); 3688 3689 /* If the TCP option is too short, we can short cut */ 3690 if (length < TCPOLEN_MD5SIG) 3691 return NULL; 3692 3693 while (length > 0) { 3694 int opcode = *ptr++; 3695 int opsize; 3696 3697 switch (opcode) { 3698 case TCPOPT_EOL: 3699 return NULL; 3700 case TCPOPT_NOP: 3701 length--; 3702 continue; 3703 default: 3704 opsize = *ptr++; 3705 if (opsize < 2 || opsize > length) 3706 return NULL; 3707 if (opcode == TCPOPT_MD5SIG) 3708 return opsize == TCPOLEN_MD5SIG ? ptr : NULL; 3709 } 3710 ptr += opsize - 2; 3711 length -= opsize; 3712 } 3713 return NULL; 3714} 3715EXPORT_SYMBOL(tcp_parse_md5sig_option); 3716#endif 3717 3718/* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM 3719 * 3720 * It is not fatal. If this ACK does _not_ change critical state (seqs, window) 3721 * it can pass through stack. So, the following predicate verifies that 3722 * this segment is not used for anything but congestion avoidance or 3723 * fast retransmit. Moreover, we even are able to eliminate most of such 3724 * second order effects, if we apply some small "replay" window (~RTO) 3725 * to timestamp space. 3726 * 3727 * All these measures still do not guarantee that we reject wrapped ACKs 3728 * on networks with high bandwidth, when sequence space is recycled fastly, 3729 * but it guarantees that such events will be very rare and do not affect 3730 * connection seriously. This doesn't look nice, but alas, PAWS is really 3731 * buggy extension. 3732 * 3733 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC 3734 * states that events when retransmit arrives after original data are rare. 3735 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is 3736 * the biggest problem on large power networks even with minor reordering. 3737 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe 3738 * up to bandwidth of 18Gigabit/sec. 8) ] 3739 */ 3740 3741static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb) 3742{ 3743 const struct tcp_sock *tp = tcp_sk(sk); 3744 const struct tcphdr *th = tcp_hdr(skb); 3745 u32 seq = TCP_SKB_CB(skb)->seq; 3746 u32 ack = TCP_SKB_CB(skb)->ack_seq; 3747 3748 return (/* 1. Pure ACK with correct sequence number. */ 3749 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) && 3750 3751 /* 2. ... and duplicate ACK. */ 3752 ack == tp->snd_una && 3753 3754 /* 3. ... and does not update window. */ 3755 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) && 3756 3757 /* 4. ... and sits in replay window. */ 3758 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ); 3759} 3760 3761static inline bool tcp_paws_discard(const struct sock *sk, 3762 const struct sk_buff *skb) 3763{ 3764 const struct tcp_sock *tp = tcp_sk(sk); 3765 3766 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) && 3767 !tcp_disordered_ack(sk, skb); 3768} 3769 3770/* Check segment sequence number for validity. 3771 * 3772 * Segment controls are considered valid, if the segment 3773 * fits to the window after truncation to the window. Acceptability 3774 * of data (and SYN, FIN, of course) is checked separately. 3775 * See tcp_data_queue(), for example. 3776 * 3777 * Also, controls (RST is main one) are accepted using RCV.WUP instead 3778 * of RCV.NXT. Peer still did not advance his SND.UNA when we 3779 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP. 3780 * (borrowed from freebsd) 3781 */ 3782 3783static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq) 3784{ 3785 return !before(end_seq, tp->rcv_wup) && 3786 !after(seq, tp->rcv_nxt + tcp_receive_window(tp)); 3787} 3788 3789/* When we get a reset we do this. */ 3790void tcp_reset(struct sock *sk) 3791{ 3792 /* We want the right error as BSD sees it (and indeed as we do). */ 3793 switch (sk->sk_state) { 3794 case TCP_SYN_SENT: 3795 sk->sk_err = ECONNREFUSED; 3796 break; 3797 case TCP_CLOSE_WAIT: 3798 sk->sk_err = EPIPE; 3799 break; 3800 case TCP_CLOSE: 3801 return; 3802 default: 3803 sk->sk_err = ECONNRESET; 3804 } 3805 /* This barrier is coupled with smp_rmb() in tcp_poll() */ 3806 smp_wmb(); 3807 3808 if (!sock_flag(sk, SOCK_DEAD)) 3809 sk->sk_error_report(sk); 3810 3811 tcp_done(sk); 3812} 3813 3814/* 3815 * Process the FIN bit. This now behaves as it is supposed to work 3816 * and the FIN takes effect when it is validly part of sequence 3817 * space. Not before when we get holes. 3818 * 3819 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT 3820 * (and thence onto LAST-ACK and finally, CLOSE, we never enter 3821 * TIME-WAIT) 3822 * 3823 * If we are in FINWAIT-1, a received FIN indicates simultaneous 3824 * close and we go into CLOSING (and later onto TIME-WAIT) 3825 * 3826 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT. 3827 */ 3828static void tcp_fin(struct sock *sk) 3829{ 3830 struct tcp_sock *tp = tcp_sk(sk); 3831 const struct dst_entry *dst; 3832 3833 inet_csk_schedule_ack(sk); 3834 3835 sk->sk_shutdown |= RCV_SHUTDOWN; 3836 sock_set_flag(sk, SOCK_DONE); 3837 3838 switch (sk->sk_state) { 3839 case TCP_SYN_RECV: 3840 case TCP_ESTABLISHED: 3841 /* Move to CLOSE_WAIT */ 3842 tcp_set_state(sk, TCP_CLOSE_WAIT); 3843 dst = __sk_dst_get(sk); 3844 if (!dst || !dst_metric(dst, RTAX_QUICKACK)) 3845 inet_csk(sk)->icsk_ack.pingpong = 1; 3846 break; 3847 3848 case TCP_CLOSE_WAIT: 3849 case TCP_CLOSING: 3850 /* Received a retransmission of the FIN, do 3851 * nothing. 3852 */ 3853 break; 3854 case TCP_LAST_ACK: 3855 /* RFC793: Remain in the LAST-ACK state. */ 3856 break; 3857 3858 case TCP_FIN_WAIT1: 3859 /* This case occurs when a simultaneous close 3860 * happens, we must ack the received FIN and 3861 * enter the CLOSING state. 3862 */ 3863 tcp_send_ack(sk); 3864 tcp_set_state(sk, TCP_CLOSING); 3865 break; 3866 case TCP_FIN_WAIT2: 3867 /* Received a FIN -- send ACK and enter TIME_WAIT. */ 3868 tcp_send_ack(sk); 3869 tcp_time_wait(sk, TCP_TIME_WAIT, 0); 3870 break; 3871 default: 3872 /* Only TCP_LISTEN and TCP_CLOSE are left, in these 3873 * cases we should never reach this piece of code. 3874 */ 3875 pr_err("%s: Impossible, sk->sk_state=%d\n", 3876 __func__, sk->sk_state); 3877 break; 3878 } 3879 3880 /* It _is_ possible, that we have something out-of-order _after_ FIN. 3881 * Probably, we should reset in this case. For now drop them. 3882 */ 3883 __skb_queue_purge(&tp->out_of_order_queue); 3884 if (tcp_is_sack(tp)) 3885 tcp_sack_reset(&tp->rx_opt); 3886 sk_mem_reclaim(sk); 3887 3888 if (!sock_flag(sk, SOCK_DEAD)) { 3889 sk->sk_state_change(sk); 3890 3891 /* Do not send POLL_HUP for half duplex close. */ 3892 if (sk->sk_shutdown == SHUTDOWN_MASK || 3893 sk->sk_state == TCP_CLOSE) 3894 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP); 3895 else 3896 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN); 3897 } 3898} 3899 3900static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, 3901 u32 end_seq) 3902{ 3903 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) { 3904 if (before(seq, sp->start_seq)) 3905 sp->start_seq = seq; 3906 if (after(end_seq, sp->end_seq)) 3907 sp->end_seq = end_seq; 3908 return true; 3909 } 3910 return false; 3911} 3912 3913static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq) 3914{ 3915 struct tcp_sock *tp = tcp_sk(sk); 3916 3917 if (tcp_is_sack(tp) && sysctl_tcp_dsack) { 3918 int mib_idx; 3919 3920 if (before(seq, tp->rcv_nxt)) 3921 mib_idx = LINUX_MIB_TCPDSACKOLDSENT; 3922 else 3923 mib_idx = LINUX_MIB_TCPDSACKOFOSENT; 3924 3925 NET_INC_STATS_BH(sock_net(sk), mib_idx); 3926 3927 tp->rx_opt.dsack = 1; 3928 tp->duplicate_sack[0].start_seq = seq; 3929 tp->duplicate_sack[0].end_seq = end_seq; 3930 } 3931} 3932 3933static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq) 3934{ 3935 struct tcp_sock *tp = tcp_sk(sk); 3936 3937 if (!tp->rx_opt.dsack) 3938 tcp_dsack_set(sk, seq, end_seq); 3939 else 3940 tcp_sack_extend(tp->duplicate_sack, seq, end_seq); 3941} 3942 3943static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb) 3944{ 3945 struct tcp_sock *tp = tcp_sk(sk); 3946 3947 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 3948 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 3949 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST); 3950 tcp_enter_quickack_mode(sk); 3951 3952 if (tcp_is_sack(tp) && sysctl_tcp_dsack) { 3953 u32 end_seq = TCP_SKB_CB(skb)->end_seq; 3954 3955 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) 3956 end_seq = tp->rcv_nxt; 3957 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq); 3958 } 3959 } 3960 3961 tcp_send_ack(sk); 3962} 3963 3964/* These routines update the SACK block as out-of-order packets arrive or 3965 * in-order packets close up the sequence space. 3966 */ 3967static void tcp_sack_maybe_coalesce(struct tcp_sock *tp) 3968{ 3969 int this_sack; 3970 struct tcp_sack_block *sp = &tp->selective_acks[0]; 3971 struct tcp_sack_block *swalk = sp + 1; 3972 3973 /* See if the recent change to the first SACK eats into 3974 * or hits the sequence space of other SACK blocks, if so coalesce. 3975 */ 3976 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) { 3977 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) { 3978 int i; 3979 3980 /* Zap SWALK, by moving every further SACK up by one slot. 3981 * Decrease num_sacks. 3982 */ 3983 tp->rx_opt.num_sacks--; 3984 for (i = this_sack; i < tp->rx_opt.num_sacks; i++) 3985 sp[i] = sp[i + 1]; 3986 continue; 3987 } 3988 this_sack++, swalk++; 3989 } 3990} 3991 3992static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq) 3993{ 3994 struct tcp_sock *tp = tcp_sk(sk); 3995 struct tcp_sack_block *sp = &tp->selective_acks[0]; 3996 int cur_sacks = tp->rx_opt.num_sacks; 3997 int this_sack; 3998 3999 if (!cur_sacks) 4000 goto new_sack; 4001 4002 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) { 4003 if (tcp_sack_extend(sp, seq, end_seq)) { 4004 /* Rotate this_sack to the first one. */ 4005 for (; this_sack > 0; this_sack--, sp--) 4006 swap(*sp, *(sp - 1)); 4007 if (cur_sacks > 1) 4008 tcp_sack_maybe_coalesce(tp); 4009 return; 4010 } 4011 } 4012 4013 /* Could not find an adjacent existing SACK, build a new one, 4014 * put it at the front, and shift everyone else down. We 4015 * always know there is at least one SACK present already here. 4016 * 4017 * If the sack array is full, forget about the last one. 4018 */ 4019 if (this_sack >= TCP_NUM_SACKS) { 4020 this_sack--; 4021 tp->rx_opt.num_sacks--; 4022 sp--; 4023 } 4024 for (; this_sack > 0; this_sack--, sp--) 4025 *sp = *(sp - 1); 4026 4027new_sack: 4028 /* Build the new head SACK, and we're done. */ 4029 sp->start_seq = seq; 4030 sp->end_seq = end_seq; 4031 tp->rx_opt.num_sacks++; 4032} 4033 4034/* RCV.NXT advances, some SACKs should be eaten. */ 4035 4036static void tcp_sack_remove(struct tcp_sock *tp) 4037{ 4038 struct tcp_sack_block *sp = &tp->selective_acks[0]; 4039 int num_sacks = tp->rx_opt.num_sacks; 4040 int this_sack; 4041 4042 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */ 4043 if (skb_queue_empty(&tp->out_of_order_queue)) { 4044 tp->rx_opt.num_sacks = 0; 4045 return; 4046 } 4047 4048 for (this_sack = 0; this_sack < num_sacks;) { 4049 /* Check if the start of the sack is covered by RCV.NXT. */ 4050 if (!before(tp->rcv_nxt, sp->start_seq)) { 4051 int i; 4052 4053 /* RCV.NXT must cover all the block! */ 4054 WARN_ON(before(tp->rcv_nxt, sp->end_seq)); 4055 4056 /* Zap this SACK, by moving forward any other SACKS. */ 4057 for (i = this_sack+1; i < num_sacks; i++) 4058 tp->selective_acks[i-1] = tp->selective_acks[i]; 4059 num_sacks--; 4060 continue; 4061 } 4062 this_sack++; 4063 sp++; 4064 } 4065 tp->rx_opt.num_sacks = num_sacks; 4066} 4067 4068/* This one checks to see if we can put data from the 4069 * out_of_order queue into the receive_queue. 4070 */ 4071static void tcp_ofo_queue(struct sock *sk) 4072{ 4073 struct tcp_sock *tp = tcp_sk(sk); 4074 __u32 dsack_high = tp->rcv_nxt; 4075 struct sk_buff *skb; 4076 4077 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) { 4078 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) 4079 break; 4080 4081 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) { 4082 __u32 dsack = dsack_high; 4083 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high)) 4084 dsack_high = TCP_SKB_CB(skb)->end_seq; 4085 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack); 4086 } 4087 4088 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) { 4089 SOCK_DEBUG(sk, "ofo packet was already received\n"); 4090 __skb_unlink(skb, &tp->out_of_order_queue); 4091 __kfree_skb(skb); 4092 continue; 4093 } 4094 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n", 4095 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, 4096 TCP_SKB_CB(skb)->end_seq); 4097 4098 __skb_unlink(skb, &tp->out_of_order_queue); 4099 __skb_queue_tail(&sk->sk_receive_queue, skb); 4100 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; 4101 if (tcp_hdr(skb)->fin) 4102 tcp_fin(sk); 4103 } 4104} 4105 4106static bool tcp_prune_ofo_queue(struct sock *sk); 4107static int tcp_prune_queue(struct sock *sk); 4108 4109static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb, 4110 unsigned int size) 4111{ 4112 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf || 4113 !sk_rmem_schedule(sk, skb, size)) { 4114 4115 if (tcp_prune_queue(sk) < 0) 4116 return -1; 4117 4118 if (!sk_rmem_schedule(sk, skb, size)) { 4119 if (!tcp_prune_ofo_queue(sk)) 4120 return -1; 4121 4122 if (!sk_rmem_schedule(sk, skb, size)) 4123 return -1; 4124 } 4125 } 4126 return 0; 4127} 4128 4129/** 4130 * tcp_try_coalesce - try to merge skb to prior one 4131 * @sk: socket 4132 * @to: prior buffer 4133 * @from: buffer to add in queue 4134 * @fragstolen: pointer to boolean 4135 * 4136 * Before queueing skb @from after @to, try to merge them 4137 * to reduce overall memory use and queue lengths, if cost is small. 4138 * Packets in ofo or receive queues can stay a long time. 4139 * Better try to coalesce them right now to avoid future collapses. 4140 * Returns true if caller should free @from instead of queueing it 4141 */ 4142static bool tcp_try_coalesce(struct sock *sk, 4143 struct sk_buff *to, 4144 struct sk_buff *from, 4145 bool *fragstolen) 4146{ 4147 int delta; 4148 4149 *fragstolen = false; 4150 4151 if (tcp_hdr(from)->fin) 4152 return false; 4153 4154 /* Its possible this segment overlaps with prior segment in queue */ 4155 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq) 4156 return false; 4157 4158 if (!skb_try_coalesce(to, from, fragstolen, &delta)) 4159 return false; 4160 4161 atomic_add(delta, &sk->sk_rmem_alloc); 4162 sk_mem_charge(sk, delta); 4163 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE); 4164 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq; 4165 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq; 4166 return true; 4167} 4168 4169static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb) 4170{ 4171 struct tcp_sock *tp = tcp_sk(sk); 4172 struct sk_buff *skb1; 4173 u32 seq, end_seq; 4174 4175 TCP_ECN_check_ce(tp, skb); 4176 4177 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) { 4178 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFODROP); 4179 __kfree_skb(skb); 4180 return; 4181 } 4182 4183 /* Disable header prediction. */ 4184 tp->pred_flags = 0; 4185 inet_csk_schedule_ack(sk); 4186 4187 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOQUEUE); 4188 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n", 4189 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); 4190 4191 skb1 = skb_peek_tail(&tp->out_of_order_queue); 4192 if (!skb1) { 4193 /* Initial out of order segment, build 1 SACK. */ 4194 if (tcp_is_sack(tp)) { 4195 tp->rx_opt.num_sacks = 1; 4196 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq; 4197 tp->selective_acks[0].end_seq = 4198 TCP_SKB_CB(skb)->end_seq; 4199 } 4200 __skb_queue_head(&tp->out_of_order_queue, skb); 4201 goto end; 4202 } 4203 4204 seq = TCP_SKB_CB(skb)->seq; 4205 end_seq = TCP_SKB_CB(skb)->end_seq; 4206 4207 if (seq == TCP_SKB_CB(skb1)->end_seq) { 4208 bool fragstolen; 4209 4210 if (!tcp_try_coalesce(sk, skb1, skb, &fragstolen)) { 4211 __skb_queue_after(&tp->out_of_order_queue, skb1, skb); 4212 } else { 4213 tcp_grow_window(sk, skb); 4214 kfree_skb_partial(skb, fragstolen); 4215 skb = NULL; 4216 } 4217 4218 if (!tp->rx_opt.num_sacks || 4219 tp->selective_acks[0].end_seq != seq) 4220 goto add_sack; 4221 4222 /* Common case: data arrive in order after hole. */ 4223 tp->selective_acks[0].end_seq = end_seq; 4224 goto end; 4225 } 4226 4227 /* Find place to insert this segment. */ 4228 while (1) { 4229 if (!after(TCP_SKB_CB(skb1)->seq, seq)) 4230 break; 4231 if (skb_queue_is_first(&tp->out_of_order_queue, skb1)) { 4232 skb1 = NULL; 4233 break; 4234 } 4235 skb1 = skb_queue_prev(&tp->out_of_order_queue, skb1); 4236 } 4237 4238 /* Do skb overlap to previous one? */ 4239 if (skb1 && before(seq, TCP_SKB_CB(skb1)->end_seq)) { 4240 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) { 4241 /* All the bits are present. Drop. */ 4242 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE); 4243 __kfree_skb(skb); 4244 skb = NULL; 4245 tcp_dsack_set(sk, seq, end_seq); 4246 goto add_sack; 4247 } 4248 if (after(seq, TCP_SKB_CB(skb1)->seq)) { 4249 /* Partial overlap. */ 4250 tcp_dsack_set(sk, seq, 4251 TCP_SKB_CB(skb1)->end_seq); 4252 } else { 4253 if (skb_queue_is_first(&tp->out_of_order_queue, 4254 skb1)) 4255 skb1 = NULL; 4256 else 4257 skb1 = skb_queue_prev( 4258 &tp->out_of_order_queue, 4259 skb1); 4260 } 4261 } 4262 if (!skb1) 4263 __skb_queue_head(&tp->out_of_order_queue, skb); 4264 else 4265 __skb_queue_after(&tp->out_of_order_queue, skb1, skb); 4266 4267 /* And clean segments covered by new one as whole. */ 4268 while (!skb_queue_is_last(&tp->out_of_order_queue, skb)) { 4269 skb1 = skb_queue_next(&tp->out_of_order_queue, skb); 4270 4271 if (!after(end_seq, TCP_SKB_CB(skb1)->seq)) 4272 break; 4273 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) { 4274 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq, 4275 end_seq); 4276 break; 4277 } 4278 __skb_unlink(skb1, &tp->out_of_order_queue); 4279 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq, 4280 TCP_SKB_CB(skb1)->end_seq); 4281 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE); 4282 __kfree_skb(skb1); 4283 } 4284 4285add_sack: 4286 if (tcp_is_sack(tp)) 4287 tcp_sack_new_ofo_skb(sk, seq, end_seq); 4288end: 4289 if (skb) { 4290 tcp_grow_window(sk, skb); 4291 skb_set_owner_r(skb, sk); 4292 } 4293} 4294 4295static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, int hdrlen, 4296 bool *fragstolen) 4297{ 4298 int eaten; 4299 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue); 4300 4301 __skb_pull(skb, hdrlen); 4302 eaten = (tail && 4303 tcp_try_coalesce(sk, tail, skb, fragstolen)) ? 1 : 0; 4304 tcp_sk(sk)->rcv_nxt = TCP_SKB_CB(skb)->end_seq; 4305 if (!eaten) { 4306 __skb_queue_tail(&sk->sk_receive_queue, skb); 4307 skb_set_owner_r(skb, sk); 4308 } 4309 return eaten; 4310} 4311 4312int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size) 4313{ 4314 struct sk_buff *skb = NULL; 4315 struct tcphdr *th; 4316 bool fragstolen; 4317 4318 if (size == 0) 4319 return 0; 4320 4321 skb = alloc_skb(size + sizeof(*th), sk->sk_allocation); 4322 if (!skb) 4323 goto err; 4324 4325 if (tcp_try_rmem_schedule(sk, skb, size + sizeof(*th))) 4326 goto err_free; 4327 4328 th = (struct tcphdr *)skb_put(skb, sizeof(*th)); 4329 skb_reset_transport_header(skb); 4330 memset(th, 0, sizeof(*th)); 4331 4332 if (memcpy_fromiovec(skb_put(skb, size), msg->msg_iov, size)) 4333 goto err_free; 4334 4335 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt; 4336 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size; 4337 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1; 4338 4339 if (tcp_queue_rcv(sk, skb, sizeof(*th), &fragstolen)) { 4340 WARN_ON_ONCE(fragstolen); /* should not happen */ 4341 __kfree_skb(skb); 4342 } 4343 return size; 4344 4345err_free: 4346 kfree_skb(skb); 4347err: 4348 return -ENOMEM; 4349} 4350 4351static void tcp_data_queue(struct sock *sk, struct sk_buff *skb) 4352{ 4353 const struct tcphdr *th = tcp_hdr(skb); 4354 struct tcp_sock *tp = tcp_sk(sk); 4355 int eaten = -1; 4356 bool fragstolen = false; 4357 4358 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) 4359 goto drop; 4360 4361 skb_dst_drop(skb); 4362 __skb_pull(skb, th->doff * 4); 4363 4364 TCP_ECN_accept_cwr(tp, skb); 4365 4366 tp->rx_opt.dsack = 0; 4367 4368 /* Queue data for delivery to the user. 4369 * Packets in sequence go to the receive queue. 4370 * Out of sequence packets to the out_of_order_queue. 4371 */ 4372 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) { 4373 if (tcp_receive_window(tp) == 0) 4374 goto out_of_window; 4375 4376 /* Ok. In sequence. In window. */ 4377 if (tp->ucopy.task == current && 4378 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len && 4379 sock_owned_by_user(sk) && !tp->urg_data) { 4380 int chunk = min_t(unsigned int, skb->len, 4381 tp->ucopy.len); 4382 4383 __set_current_state(TASK_RUNNING); 4384 4385 local_bh_enable(); 4386 if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) { 4387 tp->ucopy.len -= chunk; 4388 tp->copied_seq += chunk; 4389 eaten = (chunk == skb->len); 4390 tcp_rcv_space_adjust(sk); 4391 } 4392 local_bh_disable(); 4393 } 4394 4395 if (eaten <= 0) { 4396queue_and_out: 4397 if (eaten < 0 && 4398 tcp_try_rmem_schedule(sk, skb, skb->truesize)) 4399 goto drop; 4400 4401 eaten = tcp_queue_rcv(sk, skb, 0, &fragstolen); 4402 } 4403 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; 4404 if (skb->len) 4405 tcp_event_data_recv(sk, skb); 4406 if (th->fin) 4407 tcp_fin(sk); 4408 4409 if (!skb_queue_empty(&tp->out_of_order_queue)) { 4410 tcp_ofo_queue(sk); 4411 4412 /* RFC2581. 4.2. SHOULD send immediate ACK, when 4413 * gap in queue is filled. 4414 */ 4415 if (skb_queue_empty(&tp->out_of_order_queue)) 4416 inet_csk(sk)->icsk_ack.pingpong = 0; 4417 } 4418 4419 if (tp->rx_opt.num_sacks) 4420 tcp_sack_remove(tp); 4421 4422 tcp_fast_path_check(sk); 4423 4424 if (eaten > 0) 4425 kfree_skb_partial(skb, fragstolen); 4426 if (!sock_flag(sk, SOCK_DEAD)) 4427 sk->sk_data_ready(sk); 4428 return; 4429 } 4430 4431 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) { 4432 /* A retransmit, 2nd most common case. Force an immediate ack. */ 4433 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST); 4434 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); 4435 4436out_of_window: 4437 tcp_enter_quickack_mode(sk); 4438 inet_csk_schedule_ack(sk); 4439drop: 4440 __kfree_skb(skb); 4441 return; 4442 } 4443 4444 /* Out of window. F.e. zero window probe. */ 4445 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp))) 4446 goto out_of_window; 4447 4448 tcp_enter_quickack_mode(sk); 4449 4450 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 4451 /* Partial packet, seq < rcv_next < end_seq */ 4452 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n", 4453 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, 4454 TCP_SKB_CB(skb)->end_seq); 4455 4456 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt); 4457 4458 /* If window is closed, drop tail of packet. But after 4459 * remembering D-SACK for its head made in previous line. 4460 */ 4461 if (!tcp_receive_window(tp)) 4462 goto out_of_window; 4463 goto queue_and_out; 4464 } 4465 4466 tcp_data_queue_ofo(sk, skb); 4467} 4468 4469static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb, 4470 struct sk_buff_head *list) 4471{ 4472 struct sk_buff *next = NULL; 4473 4474 if (!skb_queue_is_last(list, skb)) 4475 next = skb_queue_next(list, skb); 4476 4477 __skb_unlink(skb, list); 4478 __kfree_skb(skb); 4479 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED); 4480 4481 return next; 4482} 4483 4484/* Collapse contiguous sequence of skbs head..tail with 4485 * sequence numbers start..end. 4486 * 4487 * If tail is NULL, this means until the end of the list. 4488 * 4489 * Segments with FIN/SYN are not collapsed (only because this 4490 * simplifies code) 4491 */ 4492static void 4493tcp_collapse(struct sock *sk, struct sk_buff_head *list, 4494 struct sk_buff *head, struct sk_buff *tail, 4495 u32 start, u32 end) 4496{ 4497 struct sk_buff *skb, *n; 4498 bool end_of_skbs; 4499 4500 /* First, check that queue is collapsible and find 4501 * the point where collapsing can be useful. */ 4502 skb = head; 4503restart: 4504 end_of_skbs = true; 4505 skb_queue_walk_from_safe(list, skb, n) { 4506 if (skb == tail) 4507 break; 4508 /* No new bits? It is possible on ofo queue. */ 4509 if (!before(start, TCP_SKB_CB(skb)->end_seq)) { 4510 skb = tcp_collapse_one(sk, skb, list); 4511 if (!skb) 4512 break; 4513 goto restart; 4514 } 4515 4516 /* The first skb to collapse is: 4517 * - not SYN/FIN and 4518 * - bloated or contains data before "start" or 4519 * overlaps to the next one. 4520 */ 4521 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin && 4522 (tcp_win_from_space(skb->truesize) > skb->len || 4523 before(TCP_SKB_CB(skb)->seq, start))) { 4524 end_of_skbs = false; 4525 break; 4526 } 4527 4528 if (!skb_queue_is_last(list, skb)) { 4529 struct sk_buff *next = skb_queue_next(list, skb); 4530 if (next != tail && 4531 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(next)->seq) { 4532 end_of_skbs = false; 4533 break; 4534 } 4535 } 4536 4537 /* Decided to skip this, advance start seq. */ 4538 start = TCP_SKB_CB(skb)->end_seq; 4539 } 4540 if (end_of_skbs || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin) 4541 return; 4542 4543 while (before(start, end)) { 4544 struct sk_buff *nskb; 4545 unsigned int header = skb_headroom(skb); 4546 int copy = SKB_MAX_ORDER(header, 0); 4547 4548 /* Too big header? This can happen with IPv6. */ 4549 if (copy < 0) 4550 return; 4551 if (end - start < copy) 4552 copy = end - start; 4553 nskb = alloc_skb(copy + header, GFP_ATOMIC); 4554 if (!nskb) 4555 return; 4556 4557 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head); 4558 skb_set_network_header(nskb, (skb_network_header(skb) - 4559 skb->head)); 4560 skb_set_transport_header(nskb, (skb_transport_header(skb) - 4561 skb->head)); 4562 skb_reserve(nskb, header); 4563 memcpy(nskb->head, skb->head, header); 4564 memcpy(nskb->cb, skb->cb, sizeof(skb->cb)); 4565 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start; 4566 __skb_queue_before(list, skb, nskb); 4567 skb_set_owner_r(nskb, sk); 4568 4569 /* Copy data, releasing collapsed skbs. */ 4570 while (copy > 0) { 4571 int offset = start - TCP_SKB_CB(skb)->seq; 4572 int size = TCP_SKB_CB(skb)->end_seq - start; 4573 4574 BUG_ON(offset < 0); 4575 if (size > 0) { 4576 size = min(copy, size); 4577 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size)) 4578 BUG(); 4579 TCP_SKB_CB(nskb)->end_seq += size; 4580 copy -= size; 4581 start += size; 4582 } 4583 if (!before(start, TCP_SKB_CB(skb)->end_seq)) { 4584 skb = tcp_collapse_one(sk, skb, list); 4585 if (!skb || 4586 skb == tail || 4587 tcp_hdr(skb)->syn || 4588 tcp_hdr(skb)->fin) 4589 return; 4590 } 4591 } 4592 } 4593} 4594 4595/* Collapse ofo queue. Algorithm: select contiguous sequence of skbs 4596 * and tcp_collapse() them until all the queue is collapsed. 4597 */ 4598static void tcp_collapse_ofo_queue(struct sock *sk) 4599{ 4600 struct tcp_sock *tp = tcp_sk(sk); 4601 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue); 4602 struct sk_buff *head; 4603 u32 start, end; 4604 4605 if (skb == NULL) 4606 return; 4607 4608 start = TCP_SKB_CB(skb)->seq; 4609 end = TCP_SKB_CB(skb)->end_seq; 4610 head = skb; 4611 4612 for (;;) { 4613 struct sk_buff *next = NULL; 4614 4615 if (!skb_queue_is_last(&tp->out_of_order_queue, skb)) 4616 next = skb_queue_next(&tp->out_of_order_queue, skb); 4617 skb = next; 4618 4619 /* Segment is terminated when we see gap or when 4620 * we are at the end of all the queue. */ 4621 if (!skb || 4622 after(TCP_SKB_CB(skb)->seq, end) || 4623 before(TCP_SKB_CB(skb)->end_seq, start)) { 4624 tcp_collapse(sk, &tp->out_of_order_queue, 4625 head, skb, start, end); 4626 head = skb; 4627 if (!skb) 4628 break; 4629 /* Start new segment */ 4630 start = TCP_SKB_CB(skb)->seq; 4631 end = TCP_SKB_CB(skb)->end_seq; 4632 } else { 4633 if (before(TCP_SKB_CB(skb)->seq, start)) 4634 start = TCP_SKB_CB(skb)->seq; 4635 if (after(TCP_SKB_CB(skb)->end_seq, end)) 4636 end = TCP_SKB_CB(skb)->end_seq; 4637 } 4638 } 4639} 4640 4641/* 4642 * Purge the out-of-order queue. 4643 * Return true if queue was pruned. 4644 */ 4645static bool tcp_prune_ofo_queue(struct sock *sk) 4646{ 4647 struct tcp_sock *tp = tcp_sk(sk); 4648 bool res = false; 4649 4650 if (!skb_queue_empty(&tp->out_of_order_queue)) { 4651 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED); 4652 __skb_queue_purge(&tp->out_of_order_queue); 4653 4654 /* Reset SACK state. A conforming SACK implementation will 4655 * do the same at a timeout based retransmit. When a connection 4656 * is in a sad state like this, we care only about integrity 4657 * of the connection not performance. 4658 */ 4659 if (tp->rx_opt.sack_ok) 4660 tcp_sack_reset(&tp->rx_opt); 4661 sk_mem_reclaim(sk); 4662 res = true; 4663 } 4664 return res; 4665} 4666 4667/* Reduce allocated memory if we can, trying to get 4668 * the socket within its memory limits again. 4669 * 4670 * Return less than zero if we should start dropping frames 4671 * until the socket owning process reads some of the data 4672 * to stabilize the situation. 4673 */ 4674static int tcp_prune_queue(struct sock *sk) 4675{ 4676 struct tcp_sock *tp = tcp_sk(sk); 4677 4678 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq); 4679 4680 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED); 4681 4682 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf) 4683 tcp_clamp_window(sk); 4684 else if (sk_under_memory_pressure(sk)) 4685 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss); 4686 4687 tcp_collapse_ofo_queue(sk); 4688 if (!skb_queue_empty(&sk->sk_receive_queue)) 4689 tcp_collapse(sk, &sk->sk_receive_queue, 4690 skb_peek(&sk->sk_receive_queue), 4691 NULL, 4692 tp->copied_seq, tp->rcv_nxt); 4693 sk_mem_reclaim(sk); 4694 4695 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) 4696 return 0; 4697 4698 /* Collapsing did not help, destructive actions follow. 4699 * This must not ever occur. */ 4700 4701 tcp_prune_ofo_queue(sk); 4702 4703 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) 4704 return 0; 4705 4706 /* If we are really being abused, tell the caller to silently 4707 * drop receive data on the floor. It will get retransmitted 4708 * and hopefully then we'll have sufficient space. 4709 */ 4710 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED); 4711 4712 /* Massive buffer overcommit. */ 4713 tp->pred_flags = 0; 4714 return -1; 4715} 4716 4717static bool tcp_should_expand_sndbuf(const struct sock *sk) 4718{ 4719 const struct tcp_sock *tp = tcp_sk(sk); 4720 4721 /* If the user specified a specific send buffer setting, do 4722 * not modify it. 4723 */ 4724 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK) 4725 return false; 4726 4727 /* If we are under global TCP memory pressure, do not expand. */ 4728 if (sk_under_memory_pressure(sk)) 4729 return false; 4730 4731 /* If we are under soft global TCP memory pressure, do not expand. */ 4732 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0)) 4733 return false; 4734 4735 /* If we filled the congestion window, do not expand. */ 4736 if (tp->packets_out >= tp->snd_cwnd) 4737 return false; 4738 4739 return true; 4740} 4741 4742/* When incoming ACK allowed to free some skb from write_queue, 4743 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket 4744 * on the exit from tcp input handler. 4745 * 4746 * PROBLEM: sndbuf expansion does not work well with largesend. 4747 */ 4748static void tcp_new_space(struct sock *sk) 4749{ 4750 struct tcp_sock *tp = tcp_sk(sk); 4751 4752 if (tcp_should_expand_sndbuf(sk)) { 4753 tcp_sndbuf_expand(sk); 4754 tp->snd_cwnd_stamp = tcp_time_stamp; 4755 } 4756 4757 sk->sk_write_space(sk); 4758} 4759 4760static void tcp_check_space(struct sock *sk) 4761{ 4762 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) { 4763 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK); 4764 if (sk->sk_socket && 4765 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) 4766 tcp_new_space(sk); 4767 } 4768} 4769 4770static inline void tcp_data_snd_check(struct sock *sk) 4771{ 4772 tcp_push_pending_frames(sk); 4773 tcp_check_space(sk); 4774} 4775 4776/* 4777 * Check if sending an ack is needed. 4778 */ 4779static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible) 4780{ 4781 struct tcp_sock *tp = tcp_sk(sk); 4782 4783 /* More than one full frame received... */ 4784 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss && 4785 /* ... and right edge of window advances far enough. 4786 * (tcp_recvmsg() will send ACK otherwise). Or... 4787 */ 4788 __tcp_select_window(sk) >= tp->rcv_wnd) || 4789 /* We ACK each frame or... */ 4790 tcp_in_quickack_mode(sk) || 4791 /* We have out of order data. */ 4792 (ofo_possible && skb_peek(&tp->out_of_order_queue))) { 4793 /* Then ack it now */ 4794 tcp_send_ack(sk); 4795 } else { 4796 /* Else, send delayed ack. */ 4797 tcp_send_delayed_ack(sk); 4798 } 4799} 4800 4801static inline void tcp_ack_snd_check(struct sock *sk) 4802{ 4803 if (!inet_csk_ack_scheduled(sk)) { 4804 /* We sent a data segment already. */ 4805 return; 4806 } 4807 __tcp_ack_snd_check(sk, 1); 4808} 4809 4810/* 4811 * This routine is only called when we have urgent data 4812 * signaled. Its the 'slow' part of tcp_urg. It could be 4813 * moved inline now as tcp_urg is only called from one 4814 * place. We handle URGent data wrong. We have to - as 4815 * BSD still doesn't use the correction from RFC961. 4816 * For 1003.1g we should support a new option TCP_STDURG to permit 4817 * either form (or just set the sysctl tcp_stdurg). 4818 */ 4819 4820static void tcp_check_urg(struct sock *sk, const struct tcphdr *th) 4821{ 4822 struct tcp_sock *tp = tcp_sk(sk); 4823 u32 ptr = ntohs(th->urg_ptr); 4824 4825 if (ptr && !sysctl_tcp_stdurg) 4826 ptr--; 4827 ptr += ntohl(th->seq); 4828 4829 /* Ignore urgent data that we've already seen and read. */ 4830 if (after(tp->copied_seq, ptr)) 4831 return; 4832 4833 /* Do not replay urg ptr. 4834 * 4835 * NOTE: interesting situation not covered by specs. 4836 * Misbehaving sender may send urg ptr, pointing to segment, 4837 * which we already have in ofo queue. We are not able to fetch 4838 * such data and will stay in TCP_URG_NOTYET until will be eaten 4839 * by recvmsg(). Seems, we are not obliged to handle such wicked 4840 * situations. But it is worth to think about possibility of some 4841 * DoSes using some hypothetical application level deadlock. 4842 */ 4843 if (before(ptr, tp->rcv_nxt)) 4844 return; 4845 4846 /* Do we already have a newer (or duplicate) urgent pointer? */ 4847 if (tp->urg_data && !after(ptr, tp->urg_seq)) 4848 return; 4849 4850 /* Tell the world about our new urgent pointer. */ 4851 sk_send_sigurg(sk); 4852 4853 /* We may be adding urgent data when the last byte read was 4854 * urgent. To do this requires some care. We cannot just ignore 4855 * tp->copied_seq since we would read the last urgent byte again 4856 * as data, nor can we alter copied_seq until this data arrives 4857 * or we break the semantics of SIOCATMARK (and thus sockatmark()) 4858 * 4859 * NOTE. Double Dutch. Rendering to plain English: author of comment 4860 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB); 4861 * and expect that both A and B disappear from stream. This is _wrong_. 4862 * Though this happens in BSD with high probability, this is occasional. 4863 * Any application relying on this is buggy. Note also, that fix "works" 4864 * only in this artificial test. Insert some normal data between A and B and we will 4865 * decline of BSD again. Verdict: it is better to remove to trap 4866 * buggy users. 4867 */ 4868 if (tp->urg_seq == tp->copied_seq && tp->urg_data && 4869 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) { 4870 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue); 4871 tp->copied_seq++; 4872 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) { 4873 __skb_unlink(skb, &sk->sk_receive_queue); 4874 __kfree_skb(skb); 4875 } 4876 } 4877 4878 tp->urg_data = TCP_URG_NOTYET; 4879 tp->urg_seq = ptr; 4880 4881 /* Disable header prediction. */ 4882 tp->pred_flags = 0; 4883} 4884 4885/* This is the 'fast' part of urgent handling. */ 4886static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th) 4887{ 4888 struct tcp_sock *tp = tcp_sk(sk); 4889 4890 /* Check if we get a new urgent pointer - normally not. */ 4891 if (th->urg) 4892 tcp_check_urg(sk, th); 4893 4894 /* Do we wait for any urgent data? - normally not... */ 4895 if (tp->urg_data == TCP_URG_NOTYET) { 4896 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) - 4897 th->syn; 4898 4899 /* Is the urgent pointer pointing into this packet? */ 4900 if (ptr < skb->len) { 4901 u8 tmp; 4902 if (skb_copy_bits(skb, ptr, &tmp, 1)) 4903 BUG(); 4904 tp->urg_data = TCP_URG_VALID | tmp; 4905 if (!sock_flag(sk, SOCK_DEAD)) 4906 sk->sk_data_ready(sk); 4907 } 4908 } 4909} 4910 4911static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen) 4912{ 4913 struct tcp_sock *tp = tcp_sk(sk); 4914 int chunk = skb->len - hlen; 4915 int err; 4916 4917 local_bh_enable(); 4918 if (skb_csum_unnecessary(skb)) 4919 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk); 4920 else 4921 err = skb_copy_and_csum_datagram_iovec(skb, hlen, 4922 tp->ucopy.iov); 4923 4924 if (!err) { 4925 tp->ucopy.len -= chunk; 4926 tp->copied_seq += chunk; 4927 tcp_rcv_space_adjust(sk); 4928 } 4929 4930 local_bh_disable(); 4931 return err; 4932} 4933 4934static __sum16 __tcp_checksum_complete_user(struct sock *sk, 4935 struct sk_buff *skb) 4936{ 4937 __sum16 result; 4938 4939 if (sock_owned_by_user(sk)) { 4940 local_bh_enable(); 4941 result = __tcp_checksum_complete(skb); 4942 local_bh_disable(); 4943 } else { 4944 result = __tcp_checksum_complete(skb); 4945 } 4946 return result; 4947} 4948 4949static inline bool tcp_checksum_complete_user(struct sock *sk, 4950 struct sk_buff *skb) 4951{ 4952 return !skb_csum_unnecessary(skb) && 4953 __tcp_checksum_complete_user(sk, skb); 4954} 4955 4956#ifdef CONFIG_NET_DMA 4957static bool tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb, 4958 int hlen) 4959{ 4960 struct tcp_sock *tp = tcp_sk(sk); 4961 int chunk = skb->len - hlen; 4962 int dma_cookie; 4963 bool copied_early = false; 4964 4965 if (tp->ucopy.wakeup) 4966 return false; 4967 4968 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list) 4969 tp->ucopy.dma_chan = net_dma_find_channel(); 4970 4971 if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) { 4972 4973 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan, 4974 skb, hlen, 4975 tp->ucopy.iov, chunk, 4976 tp->ucopy.pinned_list); 4977 4978 if (dma_cookie < 0) 4979 goto out; 4980 4981 tp->ucopy.dma_cookie = dma_cookie; 4982 copied_early = true; 4983 4984 tp->ucopy.len -= chunk; 4985 tp->copied_seq += chunk; 4986 tcp_rcv_space_adjust(sk); 4987 4988 if ((tp->ucopy.len == 0) || 4989 (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) || 4990 (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) { 4991 tp->ucopy.wakeup = 1; 4992 sk->sk_data_ready(sk); 4993 } 4994 } else if (chunk > 0) { 4995 tp->ucopy.wakeup = 1; 4996 sk->sk_data_ready(sk); 4997 } 4998out: 4999 return copied_early; 5000} 5001#endif /* CONFIG_NET_DMA */ 5002 5003/* Does PAWS and seqno based validation of an incoming segment, flags will 5004 * play significant role here. 5005 */ 5006static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb, 5007 const struct tcphdr *th, int syn_inerr) 5008{ 5009 struct tcp_sock *tp = tcp_sk(sk); 5010 5011 /* RFC1323: H1. Apply PAWS check first. */ 5012 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp && 5013 tcp_paws_discard(sk, skb)) { 5014 if (!th->rst) { 5015 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED); 5016 tcp_send_dupack(sk, skb); 5017 goto discard; 5018 } 5019 /* Reset is accepted even if it did not pass PAWS. */ 5020 } 5021 5022 /* Step 1: check sequence number */ 5023 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) { 5024 /* RFC793, page 37: "In all states except SYN-SENT, all reset 5025 * (RST) segments are validated by checking their SEQ-fields." 5026 * And page 69: "If an incoming segment is not acceptable, 5027 * an acknowledgment should be sent in reply (unless the RST 5028 * bit is set, if so drop the segment and return)". 5029 */ 5030 if (!th->rst) { 5031 if (th->syn) 5032 goto syn_challenge; 5033 tcp_send_dupack(sk, skb); 5034 } 5035 goto discard; 5036 } 5037 5038 /* Step 2: check RST bit */ 5039 if (th->rst) { 5040 /* RFC 5961 3.2 : 5041 * If sequence number exactly matches RCV.NXT, then 5042 * RESET the connection 5043 * else 5044 * Send a challenge ACK 5045 */ 5046 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) 5047 tcp_reset(sk); 5048 else 5049 tcp_send_challenge_ack(sk); 5050 goto discard; 5051 } 5052 5053 /* step 3: check security and precedence [ignored] */ 5054 5055 /* step 4: Check for a SYN 5056 * RFC 5691 4.2 : Send a challenge ack 5057 */ 5058 if (th->syn) { 5059syn_challenge: 5060 if (syn_inerr) 5061 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS); 5062 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE); 5063 tcp_send_challenge_ack(sk); 5064 goto discard; 5065 } 5066 5067 return true; 5068 5069discard: 5070 __kfree_skb(skb); 5071 return false; 5072} 5073 5074/* 5075 * TCP receive function for the ESTABLISHED state. 5076 * 5077 * It is split into a fast path and a slow path. The fast path is 5078 * disabled when: 5079 * - A zero window was announced from us - zero window probing 5080 * is only handled properly in the slow path. 5081 * - Out of order segments arrived. 5082 * - Urgent data is expected. 5083 * - There is no buffer space left 5084 * - Unexpected TCP flags/window values/header lengths are received 5085 * (detected by checking the TCP header against pred_flags) 5086 * - Data is sent in both directions. Fast path only supports pure senders 5087 * or pure receivers (this means either the sequence number or the ack 5088 * value must stay constant) 5089 * - Unexpected TCP option. 5090 * 5091 * When these conditions are not satisfied it drops into a standard 5092 * receive procedure patterned after RFC793 to handle all cases. 5093 * The first three cases are guaranteed by proper pred_flags setting, 5094 * the rest is checked inline. Fast processing is turned on in 5095 * tcp_data_queue when everything is OK. 5096 */ 5097void tcp_rcv_established(struct sock *sk, struct sk_buff *skb, 5098 const struct tcphdr *th, unsigned int len) 5099{ 5100 struct tcp_sock *tp = tcp_sk(sk); 5101 5102 if (unlikely(sk->sk_rx_dst == NULL)) 5103 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb); 5104 /* 5105 * Header prediction. 5106 * The code loosely follows the one in the famous 5107 * "30 instruction TCP receive" Van Jacobson mail. 5108 * 5109 * Van's trick is to deposit buffers into socket queue 5110 * on a device interrupt, to call tcp_recv function 5111 * on the receive process context and checksum and copy 5112 * the buffer to user space. smart... 5113 * 5114 * Our current scheme is not silly either but we take the 5115 * extra cost of the net_bh soft interrupt processing... 5116 * We do checksum and copy also but from device to kernel. 5117 */ 5118 5119 tp->rx_opt.saw_tstamp = 0; 5120 5121 /* pred_flags is 0xS?10 << 16 + snd_wnd 5122 * if header_prediction is to be made 5123 * 'S' will always be tp->tcp_header_len >> 2 5124 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to 5125 * turn it off (when there are holes in the receive 5126 * space for instance) 5127 * PSH flag is ignored. 5128 */ 5129 5130 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags && 5131 TCP_SKB_CB(skb)->seq == tp->rcv_nxt && 5132 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) { 5133 int tcp_header_len = tp->tcp_header_len; 5134 5135 /* Timestamp header prediction: tcp_header_len 5136 * is automatically equal to th->doff*4 due to pred_flags 5137 * match. 5138 */ 5139 5140 /* Check timestamp */ 5141 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) { 5142 /* No? Slow path! */ 5143 if (!tcp_parse_aligned_timestamp(tp, th)) 5144 goto slow_path; 5145 5146 /* If PAWS failed, check it more carefully in slow path */ 5147 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0) 5148 goto slow_path; 5149 5150 /* DO NOT update ts_recent here, if checksum fails 5151 * and timestamp was corrupted part, it will result 5152 * in a hung connection since we will drop all 5153 * future packets due to the PAWS test. 5154 */ 5155 } 5156 5157 if (len <= tcp_header_len) { 5158 /* Bulk data transfer: sender */ 5159 if (len == tcp_header_len) { 5160 /* Predicted packet is in window by definition. 5161 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 5162 * Hence, check seq<=rcv_wup reduces to: 5163 */ 5164 if (tcp_header_len == 5165 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && 5166 tp->rcv_nxt == tp->rcv_wup) 5167 tcp_store_ts_recent(tp); 5168 5169 /* We know that such packets are checksummed 5170 * on entry. 5171 */ 5172 tcp_ack(sk, skb, 0); 5173 __kfree_skb(skb); 5174 tcp_data_snd_check(sk); 5175 return; 5176 } else { /* Header too small */ 5177 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS); 5178 goto discard; 5179 } 5180 } else { 5181 int eaten = 0; 5182 int copied_early = 0; 5183 bool fragstolen = false; 5184 5185 if (tp->copied_seq == tp->rcv_nxt && 5186 len - tcp_header_len <= tp->ucopy.len) { 5187#ifdef CONFIG_NET_DMA 5188 if (tp->ucopy.task == current && 5189 sock_owned_by_user(sk) && 5190 tcp_dma_try_early_copy(sk, skb, tcp_header_len)) { 5191 copied_early = 1; 5192 eaten = 1; 5193 } 5194#endif 5195 if (tp->ucopy.task == current && 5196 sock_owned_by_user(sk) && !copied_early) { 5197 __set_current_state(TASK_RUNNING); 5198 5199 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len)) 5200 eaten = 1; 5201 } 5202 if (eaten) { 5203 /* Predicted packet is in window by definition. 5204 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 5205 * Hence, check seq<=rcv_wup reduces to: 5206 */ 5207 if (tcp_header_len == 5208 (sizeof(struct tcphdr) + 5209 TCPOLEN_TSTAMP_ALIGNED) && 5210 tp->rcv_nxt == tp->rcv_wup) 5211 tcp_store_ts_recent(tp); 5212 5213 tcp_rcv_rtt_measure_ts(sk, skb); 5214 5215 __skb_pull(skb, tcp_header_len); 5216 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; 5217 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER); 5218 } 5219 if (copied_early) 5220 tcp_cleanup_rbuf(sk, skb->len); 5221 } 5222 if (!eaten) { 5223 if (tcp_checksum_complete_user(sk, skb)) 5224 goto csum_error; 5225 5226 if ((int)skb->truesize > sk->sk_forward_alloc) 5227 goto step5; 5228 5229 /* Predicted packet is in window by definition. 5230 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 5231 * Hence, check seq<=rcv_wup reduces to: 5232 */ 5233 if (tcp_header_len == 5234 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && 5235 tp->rcv_nxt == tp->rcv_wup) 5236 tcp_store_ts_recent(tp); 5237 5238 tcp_rcv_rtt_measure_ts(sk, skb); 5239 5240 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS); 5241 5242 /* Bulk data transfer: receiver */ 5243 eaten = tcp_queue_rcv(sk, skb, tcp_header_len, 5244 &fragstolen); 5245 } 5246 5247 tcp_event_data_recv(sk, skb); 5248 5249 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) { 5250 /* Well, only one small jumplet in fast path... */ 5251 tcp_ack(sk, skb, FLAG_DATA); 5252 tcp_data_snd_check(sk); 5253 if (!inet_csk_ack_scheduled(sk)) 5254 goto no_ack; 5255 } 5256 5257 if (!copied_early || tp->rcv_nxt != tp->rcv_wup) 5258 __tcp_ack_snd_check(sk, 0); 5259no_ack: 5260#ifdef CONFIG_NET_DMA 5261 if (copied_early) 5262 __skb_queue_tail(&sk->sk_async_wait_queue, skb); 5263 else 5264#endif 5265 if (eaten) 5266 kfree_skb_partial(skb, fragstolen); 5267 sk->sk_data_ready(sk); 5268 return; 5269 } 5270 } 5271 5272slow_path: 5273 if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb)) 5274 goto csum_error; 5275 5276 if (!th->ack && !th->rst) 5277 goto discard; 5278 5279 /* 5280 * Standard slow path. 5281 */ 5282 5283 if (!tcp_validate_incoming(sk, skb, th, 1)) 5284 return; 5285 5286step5: 5287 if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0) 5288 goto discard; 5289 5290 tcp_rcv_rtt_measure_ts(sk, skb); 5291 5292 /* Process urgent data. */ 5293 tcp_urg(sk, skb, th); 5294 5295 /* step 7: process the segment text */ 5296 tcp_data_queue(sk, skb); 5297 5298 tcp_data_snd_check(sk); 5299 tcp_ack_snd_check(sk); 5300 return; 5301 5302csum_error: 5303 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_CSUMERRORS); 5304 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS); 5305 5306discard: 5307 __kfree_skb(skb); 5308} 5309EXPORT_SYMBOL(tcp_rcv_established); 5310 5311void tcp_finish_connect(struct sock *sk, struct sk_buff *skb) 5312{ 5313 struct tcp_sock *tp = tcp_sk(sk); 5314 struct inet_connection_sock *icsk = inet_csk(sk); 5315 5316 tcp_set_state(sk, TCP_ESTABLISHED); 5317 5318 if (skb != NULL) { 5319 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb); 5320 security_inet_conn_established(sk, skb); 5321 } 5322 5323 /* Make sure socket is routed, for correct metrics. */ 5324 icsk->icsk_af_ops->rebuild_header(sk); 5325 5326 tcp_init_metrics(sk); 5327 5328 tcp_init_congestion_control(sk); 5329 5330 /* Prevent spurious tcp_cwnd_restart() on first data 5331 * packet. 5332 */ 5333 tp->lsndtime = tcp_time_stamp; 5334 5335 tcp_init_buffer_space(sk); 5336 5337 if (sock_flag(sk, SOCK_KEEPOPEN)) 5338 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp)); 5339 5340 if (!tp->rx_opt.snd_wscale) 5341 __tcp_fast_path_on(tp, tp->snd_wnd); 5342 else 5343 tp->pred_flags = 0; 5344 5345 if (!sock_flag(sk, SOCK_DEAD)) { 5346 sk->sk_state_change(sk); 5347 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT); 5348 } 5349} 5350 5351static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack, 5352 struct tcp_fastopen_cookie *cookie) 5353{ 5354 struct tcp_sock *tp = tcp_sk(sk); 5355 struct sk_buff *data = tp->syn_data ? tcp_write_queue_head(sk) : NULL; 5356 u16 mss = tp->rx_opt.mss_clamp; 5357 bool syn_drop; 5358 5359 if (mss == tp->rx_opt.user_mss) { 5360 struct tcp_options_received opt; 5361 5362 /* Get original SYNACK MSS value if user MSS sets mss_clamp */ 5363 tcp_clear_options(&opt); 5364 opt.user_mss = opt.mss_clamp = 0; 5365 tcp_parse_options(synack, &opt, 0, NULL); 5366 mss = opt.mss_clamp; 5367 } 5368 5369 if (!tp->syn_fastopen) /* Ignore an unsolicited cookie */ 5370 cookie->len = -1; 5371 5372 /* The SYN-ACK neither has cookie nor acknowledges the data. Presumably 5373 * the remote receives only the retransmitted (regular) SYNs: either 5374 * the original SYN-data or the corresponding SYN-ACK is lost. 5375 */ 5376 syn_drop = (cookie->len <= 0 && data && tp->total_retrans); 5377 5378 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop); 5379 5380 if (data) { /* Retransmit unacked data in SYN */ 5381 tcp_for_write_queue_from(data, sk) { 5382 if (data == tcp_send_head(sk) || 5383 __tcp_retransmit_skb(sk, data)) 5384 break; 5385 } 5386 tcp_rearm_rto(sk); 5387 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVEFAIL); 5388 return true; 5389 } 5390 tp->syn_data_acked = tp->syn_data; 5391 if (tp->syn_data_acked) 5392 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE); 5393 return false; 5394} 5395 5396static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb, 5397 const struct tcphdr *th, unsigned int len) 5398{ 5399 struct inet_connection_sock *icsk = inet_csk(sk); 5400 struct tcp_sock *tp = tcp_sk(sk); 5401 struct tcp_fastopen_cookie foc = { .len = -1 }; 5402 int saved_clamp = tp->rx_opt.mss_clamp; 5403 5404 tcp_parse_options(skb, &tp->rx_opt, 0, &foc); 5405 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr) 5406 tp->rx_opt.rcv_tsecr -= tp->tsoffset; 5407 5408 if (th->ack) { 5409 /* rfc793: 5410 * "If the state is SYN-SENT then 5411 * first check the ACK bit 5412 * If the ACK bit is set 5413 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send 5414 * a reset (unless the RST bit is set, if so drop 5415 * the segment and return)" 5416 */ 5417 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) || 5418 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) 5419 goto reset_and_undo; 5420 5421 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 5422 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp, 5423 tcp_time_stamp)) { 5424 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED); 5425 goto reset_and_undo; 5426 } 5427 5428 /* Now ACK is acceptable. 5429 * 5430 * "If the RST bit is set 5431 * If the ACK was acceptable then signal the user "error: 5432 * connection reset", drop the segment, enter CLOSED state, 5433 * delete TCB, and return." 5434 */ 5435 5436 if (th->rst) { 5437 tcp_reset(sk); 5438 goto discard; 5439 } 5440 5441 /* rfc793: 5442 * "fifth, if neither of the SYN or RST bits is set then 5443 * drop the segment and return." 5444 * 5445 * See note below! 5446 * --ANK(990513) 5447 */ 5448 if (!th->syn) 5449 goto discard_and_undo; 5450 5451 /* rfc793: 5452 * "If the SYN bit is on ... 5453 * are acceptable then ... 5454 * (our SYN has been ACKed), change the connection 5455 * state to ESTABLISHED..." 5456 */ 5457 5458 TCP_ECN_rcv_synack(tp, th); 5459 5460 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq); 5461 tcp_ack(sk, skb, FLAG_SLOWPATH); 5462 5463 /* Ok.. it's good. Set up sequence numbers and 5464 * move to established. 5465 */ 5466 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; 5467 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; 5468 5469 /* RFC1323: The window in SYN & SYN/ACK segments is 5470 * never scaled. 5471 */ 5472 tp->snd_wnd = ntohs(th->window); 5473 5474 if (!tp->rx_opt.wscale_ok) { 5475 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0; 5476 tp->window_clamp = min(tp->window_clamp, 65535U); 5477 } 5478 5479 if (tp->rx_opt.saw_tstamp) { 5480 tp->rx_opt.tstamp_ok = 1; 5481 tp->tcp_header_len = 5482 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 5483 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; 5484 tcp_store_ts_recent(tp); 5485 } else { 5486 tp->tcp_header_len = sizeof(struct tcphdr); 5487 } 5488 5489 if (tcp_is_sack(tp) && sysctl_tcp_fack) 5490 tcp_enable_fack(tp); 5491 5492 tcp_mtup_init(sk); 5493 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 5494 tcp_initialize_rcv_mss(sk); 5495 5496 /* Remember, tcp_poll() does not lock socket! 5497 * Change state from SYN-SENT only after copied_seq 5498 * is initialized. */ 5499 tp->copied_seq = tp->rcv_nxt; 5500 5501 smp_mb(); 5502 5503 tcp_finish_connect(sk, skb); 5504 5505 if ((tp->syn_fastopen || tp->syn_data) && 5506 tcp_rcv_fastopen_synack(sk, skb, &foc)) 5507 return -1; 5508 5509 if (sk->sk_write_pending || 5510 icsk->icsk_accept_queue.rskq_defer_accept || 5511 icsk->icsk_ack.pingpong) { 5512 /* Save one ACK. Data will be ready after 5513 * several ticks, if write_pending is set. 5514 * 5515 * It may be deleted, but with this feature tcpdumps 5516 * look so _wonderfully_ clever, that I was not able 5517 * to stand against the temptation 8) --ANK 5518 */ 5519 inet_csk_schedule_ack(sk); 5520 icsk->icsk_ack.lrcvtime = tcp_time_stamp; 5521 tcp_enter_quickack_mode(sk); 5522 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 5523 TCP_DELACK_MAX, TCP_RTO_MAX); 5524 5525discard: 5526 __kfree_skb(skb); 5527 return 0; 5528 } else { 5529 tcp_send_ack(sk); 5530 } 5531 return -1; 5532 } 5533 5534 /* No ACK in the segment */ 5535 5536 if (th->rst) { 5537 /* rfc793: 5538 * "If the RST bit is set 5539 * 5540 * Otherwise (no ACK) drop the segment and return." 5541 */ 5542 5543 goto discard_and_undo; 5544 } 5545 5546 /* PAWS check. */ 5547 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && 5548 tcp_paws_reject(&tp->rx_opt, 0)) 5549 goto discard_and_undo; 5550 5551 if (th->syn) { 5552 /* We see SYN without ACK. It is attempt of 5553 * simultaneous connect with crossed SYNs. 5554 * Particularly, it can be connect to self. 5555 */ 5556 tcp_set_state(sk, TCP_SYN_RECV); 5557 5558 if (tp->rx_opt.saw_tstamp) { 5559 tp->rx_opt.tstamp_ok = 1; 5560 tcp_store_ts_recent(tp); 5561 tp->tcp_header_len = 5562 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 5563 } else { 5564 tp->tcp_header_len = sizeof(struct tcphdr); 5565 } 5566 5567 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; 5568 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; 5569 5570 /* RFC1323: The window in SYN & SYN/ACK segments is 5571 * never scaled. 5572 */ 5573 tp->snd_wnd = ntohs(th->window); 5574 tp->snd_wl1 = TCP_SKB_CB(skb)->seq; 5575 tp->max_window = tp->snd_wnd; 5576 5577 TCP_ECN_rcv_syn(tp, th); 5578 5579 tcp_mtup_init(sk); 5580 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 5581 tcp_initialize_rcv_mss(sk); 5582 5583 tcp_send_synack(sk); 5584#if 0 5585 /* Note, we could accept data and URG from this segment. 5586 * There are no obstacles to make this (except that we must 5587 * either change tcp_recvmsg() to prevent it from returning data 5588 * before 3WHS completes per RFC793, or employ TCP Fast Open). 5589 * 5590 * However, if we ignore data in ACKless segments sometimes, 5591 * we have no reasons to accept it sometimes. 5592 * Also, seems the code doing it in step6 of tcp_rcv_state_process 5593 * is not flawless. So, discard packet for sanity. 5594 * Uncomment this return to process the data. 5595 */ 5596 return -1; 5597#else 5598 goto discard; 5599#endif 5600 } 5601 /* "fifth, if neither of the SYN or RST bits is set then 5602 * drop the segment and return." 5603 */ 5604 5605discard_and_undo: 5606 tcp_clear_options(&tp->rx_opt); 5607 tp->rx_opt.mss_clamp = saved_clamp; 5608 goto discard; 5609 5610reset_and_undo: 5611 tcp_clear_options(&tp->rx_opt); 5612 tp->rx_opt.mss_clamp = saved_clamp; 5613 return 1; 5614} 5615 5616/* 5617 * This function implements the receiving procedure of RFC 793 for 5618 * all states except ESTABLISHED and TIME_WAIT. 5619 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be 5620 * address independent. 5621 */ 5622 5623int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb, 5624 const struct tcphdr *th, unsigned int len) 5625{ 5626 struct tcp_sock *tp = tcp_sk(sk); 5627 struct inet_connection_sock *icsk = inet_csk(sk); 5628 struct request_sock *req; 5629 int queued = 0; 5630 bool acceptable; 5631 u32 synack_stamp; 5632 5633 tp->rx_opt.saw_tstamp = 0; 5634 5635 switch (sk->sk_state) { 5636 case TCP_CLOSE: 5637 goto discard; 5638 5639 case TCP_LISTEN: 5640 if (th->ack) 5641 return 1; 5642 5643 if (th->rst) 5644 goto discard; 5645 5646 if (th->syn) { 5647 if (th->fin) 5648 goto discard; 5649 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0) 5650 return 1; 5651 5652 /* Now we have several options: In theory there is 5653 * nothing else in the frame. KA9Q has an option to 5654 * send data with the syn, BSD accepts data with the 5655 * syn up to the [to be] advertised window and 5656 * Solaris 2.1 gives you a protocol error. For now 5657 * we just ignore it, that fits the spec precisely 5658 * and avoids incompatibilities. It would be nice in 5659 * future to drop through and process the data. 5660 * 5661 * Now that TTCP is starting to be used we ought to 5662 * queue this data. 5663 * But, this leaves one open to an easy denial of 5664 * service attack, and SYN cookies can't defend 5665 * against this problem. So, we drop the data 5666 * in the interest of security over speed unless 5667 * it's still in use. 5668 */ 5669 kfree_skb(skb); 5670 return 0; 5671 } 5672 goto discard; 5673 5674 case TCP_SYN_SENT: 5675 queued = tcp_rcv_synsent_state_process(sk, skb, th, len); 5676 if (queued >= 0) 5677 return queued; 5678 5679 /* Do step6 onward by hand. */ 5680 tcp_urg(sk, skb, th); 5681 __kfree_skb(skb); 5682 tcp_data_snd_check(sk); 5683 return 0; 5684 } 5685 5686 req = tp->fastopen_rsk; 5687 if (req != NULL) { 5688 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV && 5689 sk->sk_state != TCP_FIN_WAIT1); 5690 5691 if (tcp_check_req(sk, skb, req, NULL, true) == NULL) 5692 goto discard; 5693 } 5694 5695 if (!th->ack && !th->rst) 5696 goto discard; 5697 5698 if (!tcp_validate_incoming(sk, skb, th, 0)) 5699 return 0; 5700 5701 /* step 5: check the ACK field */ 5702 acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH | 5703 FLAG_UPDATE_TS_RECENT) > 0; 5704 5705 switch (sk->sk_state) { 5706 case TCP_SYN_RECV: 5707 if (!acceptable) 5708 return 1; 5709 5710 /* Once we leave TCP_SYN_RECV, we no longer need req 5711 * so release it. 5712 */ 5713 if (req) { 5714 synack_stamp = tcp_rsk(req)->snt_synack; 5715 tp->total_retrans = req->num_retrans; 5716 reqsk_fastopen_remove(sk, req, false); 5717 } else { 5718 synack_stamp = tp->lsndtime; 5719 /* Make sure socket is routed, for correct metrics. */ 5720 icsk->icsk_af_ops->rebuild_header(sk); 5721 tcp_init_congestion_control(sk); 5722 5723 tcp_mtup_init(sk); 5724 tp->copied_seq = tp->rcv_nxt; 5725 tcp_init_buffer_space(sk); 5726 } 5727 smp_mb(); 5728 tcp_set_state(sk, TCP_ESTABLISHED); 5729 sk->sk_state_change(sk); 5730 5731 /* Note, that this wakeup is only for marginal crossed SYN case. 5732 * Passively open sockets are not waked up, because 5733 * sk->sk_sleep == NULL and sk->sk_socket == NULL. 5734 */ 5735 if (sk->sk_socket) 5736 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT); 5737 5738 tp->snd_una = TCP_SKB_CB(skb)->ack_seq; 5739 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale; 5740 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq); 5741 tcp_synack_rtt_meas(sk, synack_stamp); 5742 5743 if (tp->rx_opt.tstamp_ok) 5744 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; 5745 5746 if (req) { 5747 /* Re-arm the timer because data may have been sent out. 5748 * This is similar to the regular data transmission case 5749 * when new data has just been ack'ed. 5750 * 5751 * (TFO) - we could try to be more aggressive and 5752 * retransmitting any data sooner based on when they 5753 * are sent out. 5754 */ 5755 tcp_rearm_rto(sk); 5756 } else 5757 tcp_init_metrics(sk); 5758 5759 tcp_update_pacing_rate(sk); 5760 5761 /* Prevent spurious tcp_cwnd_restart() on first data packet */ 5762 tp->lsndtime = tcp_time_stamp; 5763 5764 tcp_initialize_rcv_mss(sk); 5765 tcp_fast_path_on(tp); 5766 break; 5767 5768 case TCP_FIN_WAIT1: { 5769 struct dst_entry *dst; 5770 int tmo; 5771 5772 /* If we enter the TCP_FIN_WAIT1 state and we are a 5773 * Fast Open socket and this is the first acceptable 5774 * ACK we have received, this would have acknowledged 5775 * our SYNACK so stop the SYNACK timer. 5776 */ 5777 if (req != NULL) { 5778 /* Return RST if ack_seq is invalid. 5779 * Note that RFC793 only says to generate a 5780 * DUPACK for it but for TCP Fast Open it seems 5781 * better to treat this case like TCP_SYN_RECV 5782 * above. 5783 */ 5784 if (!acceptable) 5785 return 1; 5786 /* We no longer need the request sock. */ 5787 reqsk_fastopen_remove(sk, req, false); 5788 tcp_rearm_rto(sk); 5789 } 5790 if (tp->snd_una != tp->write_seq) 5791 break; 5792 5793 tcp_set_state(sk, TCP_FIN_WAIT2); 5794 sk->sk_shutdown |= SEND_SHUTDOWN; 5795 5796 dst = __sk_dst_get(sk); 5797 if (dst) 5798 dst_confirm(dst); 5799 5800 if (!sock_flag(sk, SOCK_DEAD)) { 5801 /* Wake up lingering close() */ 5802 sk->sk_state_change(sk); 5803 break; 5804 } 5805 5806 if (tp->linger2 < 0 || 5807 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 5808 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) { 5809 tcp_done(sk); 5810 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA); 5811 return 1; 5812 } 5813 5814 tmo = tcp_fin_time(sk); 5815 if (tmo > TCP_TIMEWAIT_LEN) { 5816 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN); 5817 } else if (th->fin || sock_owned_by_user(sk)) { 5818 /* Bad case. We could lose such FIN otherwise. 5819 * It is not a big problem, but it looks confusing 5820 * and not so rare event. We still can lose it now, 5821 * if it spins in bh_lock_sock(), but it is really 5822 * marginal case. 5823 */ 5824 inet_csk_reset_keepalive_timer(sk, tmo); 5825 } else { 5826 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo); 5827 goto discard; 5828 } 5829 break; 5830 } 5831 5832 case TCP_CLOSING: 5833 if (tp->snd_una == tp->write_seq) { 5834 tcp_time_wait(sk, TCP_TIME_WAIT, 0); 5835 goto discard; 5836 } 5837 break; 5838 5839 case TCP_LAST_ACK: 5840 if (tp->snd_una == tp->write_seq) { 5841 tcp_update_metrics(sk); 5842 tcp_done(sk); 5843 goto discard; 5844 } 5845 break; 5846 } 5847 5848 /* step 6: check the URG bit */ 5849 tcp_urg(sk, skb, th); 5850 5851 /* step 7: process the segment text */ 5852 switch (sk->sk_state) { 5853 case TCP_CLOSE_WAIT: 5854 case TCP_CLOSING: 5855 case TCP_LAST_ACK: 5856 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) 5857 break; 5858 case TCP_FIN_WAIT1: 5859 case TCP_FIN_WAIT2: 5860 /* RFC 793 says to queue data in these states, 5861 * RFC 1122 says we MUST send a reset. 5862 * BSD 4.4 also does reset. 5863 */ 5864 if (sk->sk_shutdown & RCV_SHUTDOWN) { 5865 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 5866 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) { 5867 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA); 5868 tcp_reset(sk); 5869 return 1; 5870 } 5871 } 5872 /* Fall through */ 5873 case TCP_ESTABLISHED: 5874 tcp_data_queue(sk, skb); 5875 queued = 1; 5876 break; 5877 } 5878 5879 /* tcp_data could move socket to TIME-WAIT */ 5880 if (sk->sk_state != TCP_CLOSE) { 5881 tcp_data_snd_check(sk); 5882 tcp_ack_snd_check(sk); 5883 } 5884 5885 if (!queued) { 5886discard: 5887 __kfree_skb(skb); 5888 } 5889 return 0; 5890} 5891EXPORT_SYMBOL(tcp_rcv_state_process); 5892 5893static inline void pr_drop_req(struct request_sock *req, __u16 port, int family) 5894{ 5895 struct inet_request_sock *ireq = inet_rsk(req); 5896 5897 if (family == AF_INET) 5898 LIMIT_NETDEBUG(KERN_DEBUG pr_fmt("drop open request from %pI4/%u\n"), 5899 &ireq->ir_rmt_addr, port); 5900#if IS_ENABLED(CONFIG_IPV6) 5901 else if (family == AF_INET6) 5902 LIMIT_NETDEBUG(KERN_DEBUG pr_fmt("drop open request from %pI6/%u\n"), 5903 &ireq->ir_v6_rmt_addr, port); 5904#endif 5905} 5906 5907int tcp_conn_request(struct request_sock_ops *rsk_ops, 5908 const struct tcp_request_sock_ops *af_ops, 5909 struct sock *sk, struct sk_buff *skb) 5910{ 5911 struct tcp_options_received tmp_opt; 5912 struct request_sock *req; 5913 struct tcp_sock *tp = tcp_sk(sk); 5914 struct dst_entry *dst = NULL; 5915 __u32 isn = TCP_SKB_CB(skb)->when; 5916 bool want_cookie = false, fastopen; 5917 struct flowi fl; 5918 struct tcp_fastopen_cookie foc = { .len = -1 }; 5919 int err; 5920 5921 5922 /* TW buckets are converted to open requests without 5923 * limitations, they conserve resources and peer is 5924 * evidently real one. 5925 */ 5926 if ((sysctl_tcp_syncookies == 2 || 5927 inet_csk_reqsk_queue_is_full(sk)) && !isn) { 5928 want_cookie = tcp_syn_flood_action(sk, skb, rsk_ops->slab_name); 5929 if (!want_cookie) 5930 goto drop; 5931 } 5932 5933 5934 /* Accept backlog is full. If we have already queued enough 5935 * of warm entries in syn queue, drop request. It is better than 5936 * clogging syn queue with openreqs with exponentially increasing 5937 * timeout. 5938 */ 5939 if (sk_acceptq_is_full(sk) && inet_csk_reqsk_queue_young(sk) > 1) { 5940 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS); 5941 goto drop; 5942 } 5943 5944 req = inet_reqsk_alloc(rsk_ops); 5945 if (!req) 5946 goto drop; 5947 5948 tcp_rsk(req)->af_specific = af_ops; 5949 5950 tcp_clear_options(&tmp_opt); 5951 tmp_opt.mss_clamp = af_ops->mss_clamp; 5952 tmp_opt.user_mss = tp->rx_opt.user_mss; 5953 tcp_parse_options(skb, &tmp_opt, 0, want_cookie ? NULL : &foc); 5954 5955 if (want_cookie && !tmp_opt.saw_tstamp) 5956 tcp_clear_options(&tmp_opt); 5957 5958 tmp_opt.tstamp_ok = tmp_opt.saw_tstamp; 5959 tcp_openreq_init(req, &tmp_opt, skb, sk); 5960 5961 af_ops->init_req(req, sk, skb); 5962 5963 if (security_inet_conn_request(sk, skb, req)) 5964 goto drop_and_free; 5965 5966 if (!want_cookie || tmp_opt.tstamp_ok) 5967 TCP_ECN_create_request(req, skb, sock_net(sk)); 5968 5969 if (want_cookie) { 5970 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss); 5971 req->cookie_ts = tmp_opt.tstamp_ok; 5972 } else if (!isn) { 5973 /* VJ's idea. We save last timestamp seen 5974 * from the destination in peer table, when entering 5975 * state TIME-WAIT, and check against it before 5976 * accepting new connection request. 5977 * 5978 * If "isn" is not zero, this request hit alive 5979 * timewait bucket, so that all the necessary checks 5980 * are made in the function processing timewait state. 5981 */ 5982 if (tmp_opt.saw_tstamp && tcp_death_row.sysctl_tw_recycle) { 5983 bool strict; 5984 5985 dst = af_ops->route_req(sk, &fl, req, &strict); 5986 if (dst && strict && 5987 !tcp_peer_is_proven(req, dst, true)) { 5988 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSPASSIVEREJECTED); 5989 goto drop_and_release; 5990 } 5991 } 5992 /* Kill the following clause, if you dislike this way. */ 5993 else if (!sysctl_tcp_syncookies && 5994 (sysctl_max_syn_backlog - inet_csk_reqsk_queue_len(sk) < 5995 (sysctl_max_syn_backlog >> 2)) && 5996 !tcp_peer_is_proven(req, dst, false)) { 5997 /* Without syncookies last quarter of 5998 * backlog is filled with destinations, 5999 * proven to be alive. 6000 * It means that we continue to communicate 6001 * to destinations, already remembered 6002 * to the moment of synflood. 6003 */ 6004 pr_drop_req(req, ntohs(tcp_hdr(skb)->source), 6005 rsk_ops->family); 6006 goto drop_and_release; 6007 } 6008 6009 isn = af_ops->init_seq(skb); 6010 } 6011 if (!dst) { 6012 dst = af_ops->route_req(sk, &fl, req, NULL); 6013 if (!dst) 6014 goto drop_and_free; 6015 } 6016 6017 tcp_rsk(req)->snt_isn = isn; 6018 tcp_openreq_init_rwin(req, sk, dst); 6019 fastopen = !want_cookie && 6020 tcp_try_fastopen(sk, skb, req, &foc, dst); 6021 err = af_ops->send_synack(sk, dst, &fl, req, 6022 skb_get_queue_mapping(skb), &foc); 6023 if (!fastopen) { 6024 if (err || want_cookie) 6025 goto drop_and_free; 6026 6027 tcp_rsk(req)->listener = NULL; 6028 af_ops->queue_hash_add(sk, req, TCP_TIMEOUT_INIT); 6029 } 6030 6031 return 0; 6032 6033drop_and_release: 6034 dst_release(dst); 6035drop_and_free: 6036 reqsk_free(req); 6037drop: 6038 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_LISTENDROPS); 6039 return 0; 6040} 6041EXPORT_SYMBOL(tcp_conn_request); 6042