1 // SPDX-License-Identifier: GPL-2.0
3 * INET An implementation of the TCP/IP protocol suite for the LINUX
4 * operating system. INET is implemented using the BSD Socket
5 * interface as the means of communication with the user level.
7 * Implementation of the Transmission Control Protocol(TCP).
10 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
11 * Mark Evans, <evansmp@uhura.aston.ac.uk>
12 * Corey Minyard <wf-rch!minyard@relay.EU.net>
13 * Florian La Roche, <flla@stud.uni-sb.de>
14 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
15 * Linus Torvalds, <torvalds@cs.helsinki.fi>
16 * Alan Cox, <gw4pts@gw4pts.ampr.org>
17 * Matthew Dillon, <dillon@apollo.west.oic.com>
18 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
19 * Jorge Cwik, <jorge@laser.satlink.net>
24 * Pedro Roque : Fast Retransmit/Recovery.
26 * Retransmit queue handled by TCP.
27 * Better retransmit timer handling.
28 * New congestion avoidance.
32 * Eric : Fast Retransmit.
33 * Randy Scott : MSS option defines.
34 * Eric Schenk : Fixes to slow start algorithm.
35 * Eric Schenk : Yet another double ACK bug.
36 * Eric Schenk : Delayed ACK bug fixes.
37 * Eric Schenk : Floyd style fast retrans war avoidance.
38 * David S. Miller : Don't allow zero congestion window.
39 * Eric Schenk : Fix retransmitter so that it sends
40 * next packet on ack of previous packet.
41 * Andi Kleen : Moved open_request checking here
42 * and process RSTs for open_requests.
43 * Andi Kleen : Better prune_queue, and other fixes.
44 * Andrey Savochkin: Fix RTT measurements in the presence of
46 * Andrey Savochkin: Check sequence numbers correctly when
47 * removing SACKs due to in sequence incoming
49 * Andi Kleen: Make sure we never ack data there is not
50 * enough room for. Also make this condition
51 * a fatal error if it might still happen.
52 * Andi Kleen: Add tcp_measure_rcv_mss to make
53 * connections with MSS<min(MTU,ann. MSS)
54 * work without delayed acks.
55 * Andi Kleen: Process packets with PSH set in the
57 * J Hadi Salim: ECN support
60 * Panu Kuhlberg: Experimental audit of TCP (re)transmission
61 * engine. Lots of bugs are found.
62 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs
65 #define pr_fmt(fmt) "TCP: " fmt
68 #include <linux/slab.h>
69 #include <linux/module.h>
70 #include <linux/sysctl.h>
71 #include <linux/kernel.h>
72 #include <linux/prefetch.h>
75 #include <net/inet_common.h>
76 #include <linux/ipsec.h>
77 #include <asm/unaligned.h>
78 #include <linux/errqueue.h>
79 #include <trace/events/tcp.h>
80 #include <linux/jump_label_ratelimit.h>
81 #include <net/busy_poll.h>
82 #include <net/mptcp.h>
84 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
86 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
87 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
88 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
89 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
90 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
91 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
92 #define FLAG_ECE 0x40 /* ECE in this ACK */
93 #define FLAG_LOST_RETRANS 0x80 /* This ACK marks some retransmission lost */
94 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
95 #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
96 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
97 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
98 #define FLAG_SET_XMIT_TIMER 0x1000 /* Set TLP or RTO timer */
99 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
100 #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
101 #define FLAG_NO_CHALLENGE_ACK 0x8000 /* do not call tcp_send_challenge_ack() */
102 #define FLAG_ACK_MAYBE_DELAYED 0x10000 /* Likely a delayed ACK */
104 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
105 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
106 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE|FLAG_DSACKING_ACK)
107 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
109 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
110 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
112 #define REXMIT_NONE 0 /* no loss recovery to do */
113 #define REXMIT_LOST 1 /* retransmit packets marked lost */
114 #define REXMIT_NEW 2 /* FRTO-style transmit of unsent/new packets */
116 #if IS_ENABLED(CONFIG_TLS_DEVICE)
117 static DEFINE_STATIC_KEY_DEFERRED_FALSE(clean_acked_data_enabled, HZ);
119 void clean_acked_data_enable(struct inet_connection_sock *icsk,
120 void (*cad)(struct sock *sk, u32 ack_seq))
122 icsk->icsk_clean_acked = cad;
123 static_branch_deferred_inc(&clean_acked_data_enabled);
125 EXPORT_SYMBOL_GPL(clean_acked_data_enable);
127 void clean_acked_data_disable(struct inet_connection_sock *icsk)
129 static_branch_slow_dec_deferred(&clean_acked_data_enabled);
130 icsk->icsk_clean_acked = NULL;
132 EXPORT_SYMBOL_GPL(clean_acked_data_disable);
134 void clean_acked_data_flush(void)
136 static_key_deferred_flush(&clean_acked_data_enabled);
138 EXPORT_SYMBOL_GPL(clean_acked_data_flush);
141 #ifdef CONFIG_CGROUP_BPF
142 static void bpf_skops_parse_hdr(struct sock *sk, struct sk_buff *skb)
144 bool unknown_opt = tcp_sk(sk)->rx_opt.saw_unknown &&
145 BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk),
146 BPF_SOCK_OPS_PARSE_UNKNOWN_HDR_OPT_CB_FLAG);
147 bool parse_all_opt = BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk),
148 BPF_SOCK_OPS_PARSE_ALL_HDR_OPT_CB_FLAG);
149 struct bpf_sock_ops_kern sock_ops;
151 if (likely(!unknown_opt && !parse_all_opt))
154 /* The skb will be handled in the
155 * bpf_skops_established() or
156 * bpf_skops_write_hdr_opt().
158 switch (sk->sk_state) {
165 sock_owned_by_me(sk);
167 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp));
168 sock_ops.op = BPF_SOCK_OPS_PARSE_HDR_OPT_CB;
169 sock_ops.is_fullsock = 1;
171 bpf_skops_init_skb(&sock_ops, skb, tcp_hdrlen(skb));
173 BPF_CGROUP_RUN_PROG_SOCK_OPS(&sock_ops);
176 static void bpf_skops_established(struct sock *sk, int bpf_op,
179 struct bpf_sock_ops_kern sock_ops;
181 sock_owned_by_me(sk);
183 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp));
184 sock_ops.op = bpf_op;
185 sock_ops.is_fullsock = 1;
187 /* sk with TCP_REPAIR_ON does not have skb in tcp_finish_connect */
189 bpf_skops_init_skb(&sock_ops, skb, tcp_hdrlen(skb));
191 BPF_CGROUP_RUN_PROG_SOCK_OPS(&sock_ops);
194 static void bpf_skops_parse_hdr(struct sock *sk, struct sk_buff *skb)
198 static void bpf_skops_established(struct sock *sk, int bpf_op,
204 static void tcp_gro_dev_warn(struct sock *sk, const struct sk_buff *skb,
207 static bool __once __read_mostly;
210 struct net_device *dev;
215 dev = dev_get_by_index_rcu(sock_net(sk), skb->skb_iif);
216 if (!dev || len >= dev->mtu)
217 pr_warn("%s: Driver has suspect GRO implementation, TCP performance may be compromised.\n",
218 dev ? dev->name : "Unknown driver");
223 /* Adapt the MSS value used to make delayed ack decision to the
226 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
228 struct inet_connection_sock *icsk = inet_csk(sk);
229 const unsigned int lss = icsk->icsk_ack.last_seg_size;
232 icsk->icsk_ack.last_seg_size = 0;
234 /* skb->len may jitter because of SACKs, even if peer
235 * sends good full-sized frames.
237 len = skb_shinfo(skb)->gso_size ? : skb->len;
238 if (len >= icsk->icsk_ack.rcv_mss) {
239 icsk->icsk_ack.rcv_mss = min_t(unsigned int, len,
241 /* Account for possibly-removed options */
242 if (unlikely(len > icsk->icsk_ack.rcv_mss +
243 MAX_TCP_OPTION_SPACE))
244 tcp_gro_dev_warn(sk, skb, len);
246 /* Otherwise, we make more careful check taking into account,
247 * that SACKs block is variable.
249 * "len" is invariant segment length, including TCP header.
251 len += skb->data - skb_transport_header(skb);
252 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
253 /* If PSH is not set, packet should be
254 * full sized, provided peer TCP is not badly broken.
255 * This observation (if it is correct 8)) allows
256 * to handle super-low mtu links fairly.
258 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
259 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
260 /* Subtract also invariant (if peer is RFC compliant),
261 * tcp header plus fixed timestamp option length.
262 * Resulting "len" is MSS free of SACK jitter.
264 len -= tcp_sk(sk)->tcp_header_len;
265 icsk->icsk_ack.last_seg_size = len;
267 icsk->icsk_ack.rcv_mss = len;
271 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
272 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
273 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
277 static void tcp_incr_quickack(struct sock *sk, unsigned int max_quickacks)
279 struct inet_connection_sock *icsk = inet_csk(sk);
280 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
284 quickacks = min(quickacks, max_quickacks);
285 if (quickacks > icsk->icsk_ack.quick)
286 icsk->icsk_ack.quick = quickacks;
289 void tcp_enter_quickack_mode(struct sock *sk, unsigned int max_quickacks)
291 struct inet_connection_sock *icsk = inet_csk(sk);
293 tcp_incr_quickack(sk, max_quickacks);
294 inet_csk_exit_pingpong_mode(sk);
295 icsk->icsk_ack.ato = TCP_ATO_MIN;
297 EXPORT_SYMBOL(tcp_enter_quickack_mode);
299 /* Send ACKs quickly, if "quick" count is not exhausted
300 * and the session is not interactive.
303 static bool tcp_in_quickack_mode(struct sock *sk)
305 const struct inet_connection_sock *icsk = inet_csk(sk);
306 const struct dst_entry *dst = __sk_dst_get(sk);
308 return (dst && dst_metric(dst, RTAX_QUICKACK)) ||
309 (icsk->icsk_ack.quick && !inet_csk_in_pingpong_mode(sk));
312 static void tcp_ecn_queue_cwr(struct tcp_sock *tp)
314 if (tp->ecn_flags & TCP_ECN_OK)
315 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
318 static void tcp_ecn_accept_cwr(struct sock *sk, const struct sk_buff *skb)
320 if (tcp_hdr(skb)->cwr) {
321 tcp_sk(sk)->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
323 /* If the sender is telling us it has entered CWR, then its
324 * cwnd may be very low (even just 1 packet), so we should ACK
327 if (TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq)
328 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW;
332 static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp)
334 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
337 static void __tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb)
339 struct tcp_sock *tp = tcp_sk(sk);
341 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
342 case INET_ECN_NOT_ECT:
343 /* Funny extension: if ECT is not set on a segment,
344 * and we already seen ECT on a previous segment,
345 * it is probably a retransmit.
347 if (tp->ecn_flags & TCP_ECN_SEEN)
348 tcp_enter_quickack_mode(sk, 2);
351 if (tcp_ca_needs_ecn(sk))
352 tcp_ca_event(sk, CA_EVENT_ECN_IS_CE);
354 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
355 /* Better not delay acks, sender can have a very low cwnd */
356 tcp_enter_quickack_mode(sk, 2);
357 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
359 tp->ecn_flags |= TCP_ECN_SEEN;
362 if (tcp_ca_needs_ecn(sk))
363 tcp_ca_event(sk, CA_EVENT_ECN_NO_CE);
364 tp->ecn_flags |= TCP_ECN_SEEN;
369 static void tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb)
371 if (tcp_sk(sk)->ecn_flags & TCP_ECN_OK)
372 __tcp_ecn_check_ce(sk, skb);
375 static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
377 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
378 tp->ecn_flags &= ~TCP_ECN_OK;
381 static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
383 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
384 tp->ecn_flags &= ~TCP_ECN_OK;
387 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
389 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
394 /* Buffer size and advertised window tuning.
396 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
399 static void tcp_sndbuf_expand(struct sock *sk)
401 const struct tcp_sock *tp = tcp_sk(sk);
402 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
406 /* Worst case is non GSO/TSO : each frame consumes one skb
407 * and skb->head is kmalloced using power of two area of memory
409 per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
411 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
413 per_mss = roundup_pow_of_two(per_mss) +
414 SKB_DATA_ALIGN(sizeof(struct sk_buff));
416 nr_segs = max_t(u32, TCP_INIT_CWND, tp->snd_cwnd);
417 nr_segs = max_t(u32, nr_segs, tp->reordering + 1);
419 /* Fast Recovery (RFC 5681 3.2) :
420 * Cubic needs 1.7 factor, rounded to 2 to include
421 * extra cushion (application might react slowly to EPOLLOUT)
423 sndmem = ca_ops->sndbuf_expand ? ca_ops->sndbuf_expand(sk) : 2;
424 sndmem *= nr_segs * per_mss;
426 if (sk->sk_sndbuf < sndmem)
427 WRITE_ONCE(sk->sk_sndbuf,
428 min(sndmem, sock_net(sk)->ipv4.sysctl_tcp_wmem[2]));
431 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
433 * All tcp_full_space() is split to two parts: "network" buffer, allocated
434 * forward and advertised in receiver window (tp->rcv_wnd) and
435 * "application buffer", required to isolate scheduling/application
436 * latencies from network.
437 * window_clamp is maximal advertised window. It can be less than
438 * tcp_full_space(), in this case tcp_full_space() - window_clamp
439 * is reserved for "application" buffer. The less window_clamp is
440 * the smoother our behaviour from viewpoint of network, but the lower
441 * throughput and the higher sensitivity of the connection to losses. 8)
443 * rcv_ssthresh is more strict window_clamp used at "slow start"
444 * phase to predict further behaviour of this connection.
445 * It is used for two goals:
446 * - to enforce header prediction at sender, even when application
447 * requires some significant "application buffer". It is check #1.
448 * - to prevent pruning of receive queue because of misprediction
449 * of receiver window. Check #2.
451 * The scheme does not work when sender sends good segments opening
452 * window and then starts to feed us spaghetti. But it should work
453 * in common situations. Otherwise, we have to rely on queue collapsing.
456 /* Slow part of check#2. */
457 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
459 struct tcp_sock *tp = tcp_sk(sk);
461 int truesize = tcp_win_from_space(sk, skb->truesize) >> 1;
462 int window = tcp_win_from_space(sk, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]) >> 1;
464 while (tp->rcv_ssthresh <= window) {
465 if (truesize <= skb->len)
466 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
474 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
476 struct tcp_sock *tp = tcp_sk(sk);
479 room = min_t(int, tp->window_clamp, tcp_space(sk)) - tp->rcv_ssthresh;
482 if (room > 0 && !tcp_under_memory_pressure(sk)) {
485 /* Check #2. Increase window, if skb with such overhead
486 * will fit to rcvbuf in future.
488 if (tcp_win_from_space(sk, skb->truesize) <= skb->len)
489 incr = 2 * tp->advmss;
491 incr = __tcp_grow_window(sk, skb);
494 incr = max_t(int, incr, 2 * skb->len);
495 tp->rcv_ssthresh += min(room, incr);
496 inet_csk(sk)->icsk_ack.quick |= 1;
501 /* 3. Try to fixup all. It is made immediately after connection enters
504 static void tcp_init_buffer_space(struct sock *sk)
506 int tcp_app_win = sock_net(sk)->ipv4.sysctl_tcp_app_win;
507 struct tcp_sock *tp = tcp_sk(sk);
510 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
511 tcp_sndbuf_expand(sk);
513 tcp_mstamp_refresh(tp);
514 tp->rcvq_space.time = tp->tcp_mstamp;
515 tp->rcvq_space.seq = tp->copied_seq;
517 maxwin = tcp_full_space(sk);
519 if (tp->window_clamp >= maxwin) {
520 tp->window_clamp = maxwin;
522 if (tcp_app_win && maxwin > 4 * tp->advmss)
523 tp->window_clamp = max(maxwin -
524 (maxwin >> tcp_app_win),
528 /* Force reservation of one segment. */
530 tp->window_clamp > 2 * tp->advmss &&
531 tp->window_clamp + tp->advmss > maxwin)
532 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
534 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
535 tp->snd_cwnd_stamp = tcp_jiffies32;
536 tp->rcvq_space.space = min3(tp->rcv_ssthresh, tp->rcv_wnd,
537 (u32)TCP_INIT_CWND * tp->advmss);
540 /* 4. Recalculate window clamp after socket hit its memory bounds. */
541 static void tcp_clamp_window(struct sock *sk)
543 struct tcp_sock *tp = tcp_sk(sk);
544 struct inet_connection_sock *icsk = inet_csk(sk);
545 struct net *net = sock_net(sk);
547 icsk->icsk_ack.quick = 0;
549 if (sk->sk_rcvbuf < net->ipv4.sysctl_tcp_rmem[2] &&
550 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
551 !tcp_under_memory_pressure(sk) &&
552 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
553 WRITE_ONCE(sk->sk_rcvbuf,
554 min(atomic_read(&sk->sk_rmem_alloc),
555 net->ipv4.sysctl_tcp_rmem[2]));
557 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
558 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
561 /* Initialize RCV_MSS value.
562 * RCV_MSS is an our guess about MSS used by the peer.
563 * We haven't any direct information about the MSS.
564 * It's better to underestimate the RCV_MSS rather than overestimate.
565 * Overestimations make us ACKing less frequently than needed.
566 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
568 void tcp_initialize_rcv_mss(struct sock *sk)
570 const struct tcp_sock *tp = tcp_sk(sk);
571 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
573 hint = min(hint, tp->rcv_wnd / 2);
574 hint = min(hint, TCP_MSS_DEFAULT);
575 hint = max(hint, TCP_MIN_MSS);
577 inet_csk(sk)->icsk_ack.rcv_mss = hint;
579 EXPORT_SYMBOL(tcp_initialize_rcv_mss);
581 /* Receiver "autotuning" code.
583 * The algorithm for RTT estimation w/o timestamps is based on
584 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
585 * <https://public.lanl.gov/radiant/pubs.html#DRS>
587 * More detail on this code can be found at
588 * <http://staff.psc.edu/jheffner/>,
589 * though this reference is out of date. A new paper
592 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
594 u32 new_sample = tp->rcv_rtt_est.rtt_us;
597 if (new_sample != 0) {
598 /* If we sample in larger samples in the non-timestamp
599 * case, we could grossly overestimate the RTT especially
600 * with chatty applications or bulk transfer apps which
601 * are stalled on filesystem I/O.
603 * Also, since we are only going for a minimum in the
604 * non-timestamp case, we do not smooth things out
605 * else with timestamps disabled convergence takes too
609 m -= (new_sample >> 3);
617 /* No previous measure. */
621 tp->rcv_rtt_est.rtt_us = new_sample;
624 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
628 if (tp->rcv_rtt_est.time == 0)
630 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
632 delta_us = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcv_rtt_est.time);
635 tcp_rcv_rtt_update(tp, delta_us, 1);
638 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
639 tp->rcv_rtt_est.time = tp->tcp_mstamp;
642 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
643 const struct sk_buff *skb)
645 struct tcp_sock *tp = tcp_sk(sk);
647 if (tp->rx_opt.rcv_tsecr == tp->rcv_rtt_last_tsecr)
649 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr;
651 if (TCP_SKB_CB(skb)->end_seq -
652 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss) {
653 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr;
656 if (likely(delta < INT_MAX / (USEC_PER_SEC / TCP_TS_HZ))) {
659 delta_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
660 tcp_rcv_rtt_update(tp, delta_us, 0);
666 * This function should be called every time data is copied to user space.
667 * It calculates the appropriate TCP receive buffer space.
669 void tcp_rcv_space_adjust(struct sock *sk)
671 struct tcp_sock *tp = tcp_sk(sk);
675 trace_tcp_rcv_space_adjust(sk);
677 tcp_mstamp_refresh(tp);
678 time = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcvq_space.time);
679 if (time < (tp->rcv_rtt_est.rtt_us >> 3) || tp->rcv_rtt_est.rtt_us == 0)
682 /* Number of bytes copied to user in last RTT */
683 copied = tp->copied_seq - tp->rcvq_space.seq;
684 if (copied <= tp->rcvq_space.space)
688 * copied = bytes received in previous RTT, our base window
689 * To cope with packet losses, we need a 2x factor
690 * To cope with slow start, and sender growing its cwin by 100 %
691 * every RTT, we need a 4x factor, because the ACK we are sending
692 * now is for the next RTT, not the current one :
693 * <prev RTT . ><current RTT .. ><next RTT .... >
696 if (sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf &&
697 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
701 /* minimal window to cope with packet losses, assuming
702 * steady state. Add some cushion because of small variations.
704 rcvwin = ((u64)copied << 1) + 16 * tp->advmss;
706 /* Accommodate for sender rate increase (eg. slow start) */
707 grow = rcvwin * (copied - tp->rcvq_space.space);
708 do_div(grow, tp->rcvq_space.space);
709 rcvwin += (grow << 1);
711 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
712 while (tcp_win_from_space(sk, rcvmem) < tp->advmss)
715 do_div(rcvwin, tp->advmss);
716 rcvbuf = min_t(u64, rcvwin * rcvmem,
717 sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
718 if (rcvbuf > sk->sk_rcvbuf) {
719 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
721 /* Make the window clamp follow along. */
722 tp->window_clamp = tcp_win_from_space(sk, rcvbuf);
725 tp->rcvq_space.space = copied;
728 tp->rcvq_space.seq = tp->copied_seq;
729 tp->rcvq_space.time = tp->tcp_mstamp;
732 /* There is something which you must keep in mind when you analyze the
733 * behavior of the tp->ato delayed ack timeout interval. When a
734 * connection starts up, we want to ack as quickly as possible. The
735 * problem is that "good" TCP's do slow start at the beginning of data
736 * transmission. The means that until we send the first few ACK's the
737 * sender will sit on his end and only queue most of his data, because
738 * he can only send snd_cwnd unacked packets at any given time. For
739 * each ACK we send, he increments snd_cwnd and transmits more of his
742 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
744 struct tcp_sock *tp = tcp_sk(sk);
745 struct inet_connection_sock *icsk = inet_csk(sk);
748 inet_csk_schedule_ack(sk);
750 tcp_measure_rcv_mss(sk, skb);
752 tcp_rcv_rtt_measure(tp);
756 if (!icsk->icsk_ack.ato) {
757 /* The _first_ data packet received, initialize
758 * delayed ACK engine.
760 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
761 icsk->icsk_ack.ato = TCP_ATO_MIN;
763 int m = now - icsk->icsk_ack.lrcvtime;
765 if (m <= TCP_ATO_MIN / 2) {
766 /* The fastest case is the first. */
767 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
768 } else if (m < icsk->icsk_ack.ato) {
769 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
770 if (icsk->icsk_ack.ato > icsk->icsk_rto)
771 icsk->icsk_ack.ato = icsk->icsk_rto;
772 } else if (m > icsk->icsk_rto) {
773 /* Too long gap. Apparently sender failed to
774 * restart window, so that we send ACKs quickly.
776 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
780 icsk->icsk_ack.lrcvtime = now;
782 tcp_ecn_check_ce(sk, skb);
785 tcp_grow_window(sk, skb);
788 /* Called to compute a smoothed rtt estimate. The data fed to this
789 * routine either comes from timestamps, or from segments that were
790 * known _not_ to have been retransmitted [see Karn/Partridge
791 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
792 * piece by Van Jacobson.
793 * NOTE: the next three routines used to be one big routine.
794 * To save cycles in the RFC 1323 implementation it was better to break
795 * it up into three procedures. -- erics
797 static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
799 struct tcp_sock *tp = tcp_sk(sk);
800 long m = mrtt_us; /* RTT */
801 u32 srtt = tp->srtt_us;
803 /* The following amusing code comes from Jacobson's
804 * article in SIGCOMM '88. Note that rtt and mdev
805 * are scaled versions of rtt and mean deviation.
806 * This is designed to be as fast as possible
807 * m stands for "measurement".
809 * On a 1990 paper the rto value is changed to:
810 * RTO = rtt + 4 * mdev
812 * Funny. This algorithm seems to be very broken.
813 * These formulae increase RTO, when it should be decreased, increase
814 * too slowly, when it should be increased quickly, decrease too quickly
815 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
816 * does not matter how to _calculate_ it. Seems, it was trap
817 * that VJ failed to avoid. 8)
820 m -= (srtt >> 3); /* m is now error in rtt est */
821 srtt += m; /* rtt = 7/8 rtt + 1/8 new */
823 m = -m; /* m is now abs(error) */
824 m -= (tp->mdev_us >> 2); /* similar update on mdev */
825 /* This is similar to one of Eifel findings.
826 * Eifel blocks mdev updates when rtt decreases.
827 * This solution is a bit different: we use finer gain
828 * for mdev in this case (alpha*beta).
829 * Like Eifel it also prevents growth of rto,
830 * but also it limits too fast rto decreases,
831 * happening in pure Eifel.
836 m -= (tp->mdev_us >> 2); /* similar update on mdev */
838 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */
839 if (tp->mdev_us > tp->mdev_max_us) {
840 tp->mdev_max_us = tp->mdev_us;
841 if (tp->mdev_max_us > tp->rttvar_us)
842 tp->rttvar_us = tp->mdev_max_us;
844 if (after(tp->snd_una, tp->rtt_seq)) {
845 if (tp->mdev_max_us < tp->rttvar_us)
846 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
847 tp->rtt_seq = tp->snd_nxt;
848 tp->mdev_max_us = tcp_rto_min_us(sk);
853 /* no previous measure. */
854 srtt = m << 3; /* take the measured time to be rtt */
855 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */
856 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
857 tp->mdev_max_us = tp->rttvar_us;
858 tp->rtt_seq = tp->snd_nxt;
862 tp->srtt_us = max(1U, srtt);
865 static void tcp_update_pacing_rate(struct sock *sk)
867 const struct tcp_sock *tp = tcp_sk(sk);
870 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
871 rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3);
873 /* current rate is (cwnd * mss) / srtt
874 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
875 * In Congestion Avoidance phase, set it to 120 % the current rate.
877 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
878 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
879 * end of slow start and should slow down.
881 if (tp->snd_cwnd < tp->snd_ssthresh / 2)
882 rate *= sock_net(sk)->ipv4.sysctl_tcp_pacing_ss_ratio;
884 rate *= sock_net(sk)->ipv4.sysctl_tcp_pacing_ca_ratio;
886 rate *= max(tp->snd_cwnd, tp->packets_out);
888 if (likely(tp->srtt_us))
889 do_div(rate, tp->srtt_us);
891 /* WRITE_ONCE() is needed because sch_fq fetches sk_pacing_rate
892 * without any lock. We want to make sure compiler wont store
893 * intermediate values in this location.
895 WRITE_ONCE(sk->sk_pacing_rate, min_t(u64, rate,
896 sk->sk_max_pacing_rate));
899 /* Calculate rto without backoff. This is the second half of Van Jacobson's
900 * routine referred to above.
902 static void tcp_set_rto(struct sock *sk)
904 const struct tcp_sock *tp = tcp_sk(sk);
905 /* Old crap is replaced with new one. 8)
908 * 1. If rtt variance happened to be less 50msec, it is hallucination.
909 * It cannot be less due to utterly erratic ACK generation made
910 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
911 * to do with delayed acks, because at cwnd>2 true delack timeout
912 * is invisible. Actually, Linux-2.4 also generates erratic
913 * ACKs in some circumstances.
915 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
917 /* 2. Fixups made earlier cannot be right.
918 * If we do not estimate RTO correctly without them,
919 * all the algo is pure shit and should be replaced
920 * with correct one. It is exactly, which we pretend to do.
923 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
924 * guarantees that rto is higher.
929 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
931 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
934 cwnd = TCP_INIT_CWND;
935 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
938 struct tcp_sacktag_state {
939 /* Timestamps for earliest and latest never-retransmitted segment
940 * that was SACKed. RTO needs the earliest RTT to stay conservative,
941 * but congestion control should still get an accurate delay signal.
948 unsigned int mss_now;
949 struct rate_sample *rate;
952 /* Take a notice that peer is sending D-SACKs. Skip update of data delivery
953 * and spurious retransmission information if this DSACK is unlikely caused by
955 * - DSACKed sequence range is larger than maximum receiver's window.
956 * - Total no. of DSACKed segments exceed the total no. of retransmitted segs.
958 static u32 tcp_dsack_seen(struct tcp_sock *tp, u32 start_seq,
959 u32 end_seq, struct tcp_sacktag_state *state)
961 u32 seq_len, dup_segs = 1;
963 if (!before(start_seq, end_seq))
966 seq_len = end_seq - start_seq;
967 /* Dubious DSACK: DSACKed range greater than maximum advertised rwnd */
968 if (seq_len > tp->max_window)
970 if (seq_len > tp->mss_cache)
971 dup_segs = DIV_ROUND_UP(seq_len, tp->mss_cache);
973 tp->dsack_dups += dup_segs;
974 /* Skip the DSACK if dup segs weren't retransmitted by sender */
975 if (tp->dsack_dups > tp->total_retrans)
978 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
979 tp->rack.dsack_seen = 1;
981 state->flag |= FLAG_DSACKING_ACK;
982 /* A spurious retransmission is delivered */
983 state->sack_delivered += dup_segs;
988 /* It's reordering when higher sequence was delivered (i.e. sacked) before
989 * some lower never-retransmitted sequence ("low_seq"). The maximum reordering
990 * distance is approximated in full-mss packet distance ("reordering").
992 static void tcp_check_sack_reordering(struct sock *sk, const u32 low_seq,
995 struct tcp_sock *tp = tcp_sk(sk);
996 const u32 mss = tp->mss_cache;
999 fack = tcp_highest_sack_seq(tp);
1000 if (!before(low_seq, fack))
1003 metric = fack - low_seq;
1004 if ((metric > tp->reordering * mss) && mss) {
1005 #if FASTRETRANS_DEBUG > 1
1006 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
1007 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
1011 tp->undo_marker ? tp->undo_retrans : 0);
1013 tp->reordering = min_t(u32, (metric + mss - 1) / mss,
1014 sock_net(sk)->ipv4.sysctl_tcp_max_reordering);
1017 /* This exciting event is worth to be remembered. 8) */
1019 NET_INC_STATS(sock_net(sk),
1020 ts ? LINUX_MIB_TCPTSREORDER : LINUX_MIB_TCPSACKREORDER);
1023 /* This must be called before lost_out or retrans_out are updated
1024 * on a new loss, because we want to know if all skbs previously
1025 * known to be lost have already been retransmitted, indicating
1026 * that this newly lost skb is our next skb to retransmit.
1028 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
1030 if ((!tp->retransmit_skb_hint && tp->retrans_out >= tp->lost_out) ||
1031 (tp->retransmit_skb_hint &&
1032 before(TCP_SKB_CB(skb)->seq,
1033 TCP_SKB_CB(tp->retransmit_skb_hint)->seq)))
1034 tp->retransmit_skb_hint = skb;
1037 /* Sum the number of packets on the wire we have marked as lost, and
1038 * notify the congestion control module that the given skb was marked lost.
1040 static void tcp_notify_skb_loss_event(struct tcp_sock *tp, const struct sk_buff *skb)
1042 tp->lost += tcp_skb_pcount(skb);
1045 void tcp_mark_skb_lost(struct sock *sk, struct sk_buff *skb)
1047 __u8 sacked = TCP_SKB_CB(skb)->sacked;
1048 struct tcp_sock *tp = tcp_sk(sk);
1050 if (sacked & TCPCB_SACKED_ACKED)
1053 tcp_verify_retransmit_hint(tp, skb);
1054 if (sacked & TCPCB_LOST) {
1055 if (sacked & TCPCB_SACKED_RETRANS) {
1056 /* Account for retransmits that are lost again */
1057 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1058 tp->retrans_out -= tcp_skb_pcount(skb);
1059 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT,
1060 tcp_skb_pcount(skb));
1061 tcp_notify_skb_loss_event(tp, skb);
1064 tp->lost_out += tcp_skb_pcount(skb);
1065 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1066 tcp_notify_skb_loss_event(tp, skb);
1070 /* Updates the delivered and delivered_ce counts */
1071 static void tcp_count_delivered(struct tcp_sock *tp, u32 delivered,
1074 tp->delivered += delivered;
1076 tp->delivered_ce += delivered;
1079 /* This procedure tags the retransmission queue when SACKs arrive.
1081 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
1082 * Packets in queue with these bits set are counted in variables
1083 * sacked_out, retrans_out and lost_out, correspondingly.
1085 * Valid combinations are:
1086 * Tag InFlight Description
1087 * 0 1 - orig segment is in flight.
1088 * S 0 - nothing flies, orig reached receiver.
1089 * L 0 - nothing flies, orig lost by net.
1090 * R 2 - both orig and retransmit are in flight.
1091 * L|R 1 - orig is lost, retransmit is in flight.
1092 * S|R 1 - orig reached receiver, retrans is still in flight.
1093 * (L|S|R is logically valid, it could occur when L|R is sacked,
1094 * but it is equivalent to plain S and code short-curcuits it to S.
1095 * L|S is logically invalid, it would mean -1 packet in flight 8))
1097 * These 6 states form finite state machine, controlled by the following events:
1098 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
1099 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
1100 * 3. Loss detection event of two flavors:
1101 * A. Scoreboard estimator decided the packet is lost.
1102 * A'. Reno "three dupacks" marks head of queue lost.
1103 * B. SACK arrives sacking SND.NXT at the moment, when the
1104 * segment was retransmitted.
1105 * 4. D-SACK added new rule: D-SACK changes any tag to S.
1107 * It is pleasant to note, that state diagram turns out to be commutative,
1108 * so that we are allowed not to be bothered by order of our actions,
1109 * when multiple events arrive simultaneously. (see the function below).
1111 * Reordering detection.
1112 * --------------------
1113 * Reordering metric is maximal distance, which a packet can be displaced
1114 * in packet stream. With SACKs we can estimate it:
1116 * 1. SACK fills old hole and the corresponding segment was not
1117 * ever retransmitted -> reordering. Alas, we cannot use it
1118 * when segment was retransmitted.
1119 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1120 * for retransmitted and already SACKed segment -> reordering..
1121 * Both of these heuristics are not used in Loss state, when we cannot
1122 * account for retransmits accurately.
1124 * SACK block validation.
1125 * ----------------------
1127 * SACK block range validation checks that the received SACK block fits to
1128 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1129 * Note that SND.UNA is not included to the range though being valid because
1130 * it means that the receiver is rather inconsistent with itself reporting
1131 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1132 * perfectly valid, however, in light of RFC2018 which explicitly states
1133 * that "SACK block MUST reflect the newest segment. Even if the newest
1134 * segment is going to be discarded ...", not that it looks very clever
1135 * in case of head skb. Due to potentional receiver driven attacks, we
1136 * choose to avoid immediate execution of a walk in write queue due to
1137 * reneging and defer head skb's loss recovery to standard loss recovery
1138 * procedure that will eventually trigger (nothing forbids us doing this).
1140 * Implements also blockage to start_seq wrap-around. Problem lies in the
1141 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1142 * there's no guarantee that it will be before snd_nxt (n). The problem
1143 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1146 * <- outs wnd -> <- wrapzone ->
1147 * u e n u_w e_w s n_w
1149 * |<------------+------+----- TCP seqno space --------------+---------->|
1150 * ...-- <2^31 ->| |<--------...
1151 * ...---- >2^31 ------>| |<--------...
1153 * Current code wouldn't be vulnerable but it's better still to discard such
1154 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1155 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1156 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1157 * equal to the ideal case (infinite seqno space without wrap caused issues).
1159 * With D-SACK the lower bound is extended to cover sequence space below
1160 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1161 * again, D-SACK block must not to go across snd_una (for the same reason as
1162 * for the normal SACK blocks, explained above). But there all simplicity
1163 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1164 * fully below undo_marker they do not affect behavior in anyway and can
1165 * therefore be safely ignored. In rare cases (which are more or less
1166 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1167 * fragmentation and packet reordering past skb's retransmission. To consider
1168 * them correctly, the acceptable range must be extended even more though
1169 * the exact amount is rather hard to quantify. However, tp->max_window can
1170 * be used as an exaggerated estimate.
1172 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
1173 u32 start_seq, u32 end_seq)
1175 /* Too far in future, or reversed (interpretation is ambiguous) */
1176 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1179 /* Nasty start_seq wrap-around check (see comments above) */
1180 if (!before(start_seq, tp->snd_nxt))
1183 /* In outstanding window? ...This is valid exit for D-SACKs too.
1184 * start_seq == snd_una is non-sensical (see comments above)
1186 if (after(start_seq, tp->snd_una))
1189 if (!is_dsack || !tp->undo_marker)
1192 /* ...Then it's D-SACK, and must reside below snd_una completely */
1193 if (after(end_seq, tp->snd_una))
1196 if (!before(start_seq, tp->undo_marker))
1200 if (!after(end_seq, tp->undo_marker))
1203 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1204 * start_seq < undo_marker and end_seq >= undo_marker.
1206 return !before(start_seq, end_seq - tp->max_window);
1209 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1210 struct tcp_sack_block_wire *sp, int num_sacks,
1211 u32 prior_snd_una, struct tcp_sacktag_state *state)
1213 struct tcp_sock *tp = tcp_sk(sk);
1214 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1215 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1218 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1219 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1220 } else if (num_sacks > 1) {
1221 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1222 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1224 if (after(end_seq_0, end_seq_1) || before(start_seq_0, start_seq_1))
1226 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKOFORECV);
1231 dup_segs = tcp_dsack_seen(tp, start_seq_0, end_seq_0, state);
1232 if (!dup_segs) { /* Skip dubious DSACK */
1233 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKIGNOREDDUBIOUS);
1237 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECVSEGS, dup_segs);
1239 /* D-SACK for already forgotten data... Do dumb counting. */
1240 if (tp->undo_marker && tp->undo_retrans > 0 &&
1241 !after(end_seq_0, prior_snd_una) &&
1242 after(end_seq_0, tp->undo_marker))
1243 tp->undo_retrans = max_t(int, 0, tp->undo_retrans - dup_segs);
1248 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1249 * the incoming SACK may not exactly match but we can find smaller MSS
1250 * aligned portion of it that matches. Therefore we might need to fragment
1251 * which may fail and creates some hassle (caller must handle error case
1254 * FIXME: this could be merged to shift decision code
1256 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1257 u32 start_seq, u32 end_seq)
1261 unsigned int pkt_len;
1264 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1265 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1267 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1268 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1269 mss = tcp_skb_mss(skb);
1270 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1273 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1277 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1282 /* Round if necessary so that SACKs cover only full MSSes
1283 * and/or the remaining small portion (if present)
1285 if (pkt_len > mss) {
1286 unsigned int new_len = (pkt_len / mss) * mss;
1287 if (!in_sack && new_len < pkt_len)
1292 if (pkt_len >= skb->len && !in_sack)
1295 err = tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
1296 pkt_len, mss, GFP_ATOMIC);
1304 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1305 static u8 tcp_sacktag_one(struct sock *sk,
1306 struct tcp_sacktag_state *state, u8 sacked,
1307 u32 start_seq, u32 end_seq,
1308 int dup_sack, int pcount,
1311 struct tcp_sock *tp = tcp_sk(sk);
1313 /* Account D-SACK for retransmitted packet. */
1314 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1315 if (tp->undo_marker && tp->undo_retrans > 0 &&
1316 after(end_seq, tp->undo_marker))
1318 if ((sacked & TCPCB_SACKED_ACKED) &&
1319 before(start_seq, state->reord))
1320 state->reord = start_seq;
1323 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1324 if (!after(end_seq, tp->snd_una))
1327 if (!(sacked & TCPCB_SACKED_ACKED)) {
1328 tcp_rack_advance(tp, sacked, end_seq, xmit_time);
1330 if (sacked & TCPCB_SACKED_RETRANS) {
1331 /* If the segment is not tagged as lost,
1332 * we do not clear RETRANS, believing
1333 * that retransmission is still in flight.
1335 if (sacked & TCPCB_LOST) {
1336 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1337 tp->lost_out -= pcount;
1338 tp->retrans_out -= pcount;
1341 if (!(sacked & TCPCB_RETRANS)) {
1342 /* New sack for not retransmitted frame,
1343 * which was in hole. It is reordering.
1345 if (before(start_seq,
1346 tcp_highest_sack_seq(tp)) &&
1347 before(start_seq, state->reord))
1348 state->reord = start_seq;
1350 if (!after(end_seq, tp->high_seq))
1351 state->flag |= FLAG_ORIG_SACK_ACKED;
1352 if (state->first_sackt == 0)
1353 state->first_sackt = xmit_time;
1354 state->last_sackt = xmit_time;
1357 if (sacked & TCPCB_LOST) {
1358 sacked &= ~TCPCB_LOST;
1359 tp->lost_out -= pcount;
1363 sacked |= TCPCB_SACKED_ACKED;
1364 state->flag |= FLAG_DATA_SACKED;
1365 tp->sacked_out += pcount;
1366 /* Out-of-order packets delivered */
1367 state->sack_delivered += pcount;
1369 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1370 if (tp->lost_skb_hint &&
1371 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1372 tp->lost_cnt_hint += pcount;
1375 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1376 * frames and clear it. undo_retrans is decreased above, L|R frames
1377 * are accounted above as well.
1379 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1380 sacked &= ~TCPCB_SACKED_RETRANS;
1381 tp->retrans_out -= pcount;
1387 /* Shift newly-SACKed bytes from this skb to the immediately previous
1388 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1390 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *prev,
1391 struct sk_buff *skb,
1392 struct tcp_sacktag_state *state,
1393 unsigned int pcount, int shifted, int mss,
1396 struct tcp_sock *tp = tcp_sk(sk);
1397 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1398 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
1402 /* Adjust counters and hints for the newly sacked sequence
1403 * range but discard the return value since prev is already
1404 * marked. We must tag the range first because the seq
1405 * advancement below implicitly advances
1406 * tcp_highest_sack_seq() when skb is highest_sack.
1408 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1409 start_seq, end_seq, dup_sack, pcount,
1410 tcp_skb_timestamp_us(skb));
1411 tcp_rate_skb_delivered(sk, skb, state->rate);
1413 if (skb == tp->lost_skb_hint)
1414 tp->lost_cnt_hint += pcount;
1416 TCP_SKB_CB(prev)->end_seq += shifted;
1417 TCP_SKB_CB(skb)->seq += shifted;
1419 tcp_skb_pcount_add(prev, pcount);
1420 WARN_ON_ONCE(tcp_skb_pcount(skb) < pcount);
1421 tcp_skb_pcount_add(skb, -pcount);
1423 /* When we're adding to gso_segs == 1, gso_size will be zero,
1424 * in theory this shouldn't be necessary but as long as DSACK
1425 * code can come after this skb later on it's better to keep
1426 * setting gso_size to something.
1428 if (!TCP_SKB_CB(prev)->tcp_gso_size)
1429 TCP_SKB_CB(prev)->tcp_gso_size = mss;
1431 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1432 if (tcp_skb_pcount(skb) <= 1)
1433 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1435 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1436 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1439 BUG_ON(!tcp_skb_pcount(skb));
1440 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1444 /* Whole SKB was eaten :-) */
1446 if (skb == tp->retransmit_skb_hint)
1447 tp->retransmit_skb_hint = prev;
1448 if (skb == tp->lost_skb_hint) {
1449 tp->lost_skb_hint = prev;
1450 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1453 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1454 TCP_SKB_CB(prev)->eor = TCP_SKB_CB(skb)->eor;
1455 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1456 TCP_SKB_CB(prev)->end_seq++;
1458 if (skb == tcp_highest_sack(sk))
1459 tcp_advance_highest_sack(sk, skb);
1461 tcp_skb_collapse_tstamp(prev, skb);
1462 if (unlikely(TCP_SKB_CB(prev)->tx.delivered_mstamp))
1463 TCP_SKB_CB(prev)->tx.delivered_mstamp = 0;
1465 tcp_rtx_queue_unlink_and_free(skb, sk);
1467 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKMERGED);
1472 /* I wish gso_size would have a bit more sane initialization than
1473 * something-or-zero which complicates things
1475 static int tcp_skb_seglen(const struct sk_buff *skb)
1477 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1480 /* Shifting pages past head area doesn't work */
1481 static int skb_can_shift(const struct sk_buff *skb)
1483 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1486 int tcp_skb_shift(struct sk_buff *to, struct sk_buff *from,
1487 int pcount, int shiftlen)
1489 /* TCP min gso_size is 8 bytes (TCP_MIN_GSO_SIZE)
1490 * Since TCP_SKB_CB(skb)->tcp_gso_segs is 16 bits, we need
1491 * to make sure not storing more than 65535 * 8 bytes per skb,
1492 * even if current MSS is bigger.
1494 if (unlikely(to->len + shiftlen >= 65535 * TCP_MIN_GSO_SIZE))
1496 if (unlikely(tcp_skb_pcount(to) + pcount > 65535))
1498 return skb_shift(to, from, shiftlen);
1501 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1504 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1505 struct tcp_sacktag_state *state,
1506 u32 start_seq, u32 end_seq,
1509 struct tcp_sock *tp = tcp_sk(sk);
1510 struct sk_buff *prev;
1516 /* Normally R but no L won't result in plain S */
1518 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1520 if (!skb_can_shift(skb))
1522 /* This frame is about to be dropped (was ACKed). */
1523 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1526 /* Can only happen with delayed DSACK + discard craziness */
1527 prev = skb_rb_prev(skb);
1531 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1534 if (!tcp_skb_can_collapse(prev, skb))
1537 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1538 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1542 pcount = tcp_skb_pcount(skb);
1543 mss = tcp_skb_seglen(skb);
1545 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1546 * drop this restriction as unnecessary
1548 if (mss != tcp_skb_seglen(prev))
1551 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1553 /* CHECKME: This is non-MSS split case only?, this will
1554 * cause skipped skbs due to advancing loop btw, original
1555 * has that feature too
1557 if (tcp_skb_pcount(skb) <= 1)
1560 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1562 /* TODO: head merge to next could be attempted here
1563 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1564 * though it might not be worth of the additional hassle
1566 * ...we can probably just fallback to what was done
1567 * previously. We could try merging non-SACKed ones
1568 * as well but it probably isn't going to buy off
1569 * because later SACKs might again split them, and
1570 * it would make skb timestamp tracking considerably
1576 len = end_seq - TCP_SKB_CB(skb)->seq;
1578 BUG_ON(len > skb->len);
1580 /* MSS boundaries should be honoured or else pcount will
1581 * severely break even though it makes things bit trickier.
1582 * Optimize common case to avoid most of the divides
1584 mss = tcp_skb_mss(skb);
1586 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1587 * drop this restriction as unnecessary
1589 if (mss != tcp_skb_seglen(prev))
1594 } else if (len < mss) {
1602 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1603 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1606 if (!tcp_skb_shift(prev, skb, pcount, len))
1608 if (!tcp_shifted_skb(sk, prev, skb, state, pcount, len, mss, dup_sack))
1611 /* Hole filled allows collapsing with the next as well, this is very
1612 * useful when hole on every nth skb pattern happens
1614 skb = skb_rb_next(prev);
1618 if (!skb_can_shift(skb) ||
1619 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1620 (mss != tcp_skb_seglen(skb)))
1624 pcount = tcp_skb_pcount(skb);
1625 if (tcp_skb_shift(prev, skb, pcount, len))
1626 tcp_shifted_skb(sk, prev, skb, state, pcount,
1636 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1640 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1641 struct tcp_sack_block *next_dup,
1642 struct tcp_sacktag_state *state,
1643 u32 start_seq, u32 end_seq,
1646 struct tcp_sock *tp = tcp_sk(sk);
1647 struct sk_buff *tmp;
1649 skb_rbtree_walk_from(skb) {
1651 bool dup_sack = dup_sack_in;
1653 /* queue is in-order => we can short-circuit the walk early */
1654 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1658 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1659 in_sack = tcp_match_skb_to_sack(sk, skb,
1660 next_dup->start_seq,
1666 /* skb reference here is a bit tricky to get right, since
1667 * shifting can eat and free both this skb and the next,
1668 * so not even _safe variant of the loop is enough.
1671 tmp = tcp_shift_skb_data(sk, skb, state,
1672 start_seq, end_seq, dup_sack);
1681 in_sack = tcp_match_skb_to_sack(sk, skb,
1687 if (unlikely(in_sack < 0))
1691 TCP_SKB_CB(skb)->sacked =
1694 TCP_SKB_CB(skb)->sacked,
1695 TCP_SKB_CB(skb)->seq,
1696 TCP_SKB_CB(skb)->end_seq,
1698 tcp_skb_pcount(skb),
1699 tcp_skb_timestamp_us(skb));
1700 tcp_rate_skb_delivered(sk, skb, state->rate);
1701 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1702 list_del_init(&skb->tcp_tsorted_anchor);
1704 if (!before(TCP_SKB_CB(skb)->seq,
1705 tcp_highest_sack_seq(tp)))
1706 tcp_advance_highest_sack(sk, skb);
1712 static struct sk_buff *tcp_sacktag_bsearch(struct sock *sk, u32 seq)
1714 struct rb_node *parent, **p = &sk->tcp_rtx_queue.rb_node;
1715 struct sk_buff *skb;
1719 skb = rb_to_skb(parent);
1720 if (before(seq, TCP_SKB_CB(skb)->seq)) {
1721 p = &parent->rb_left;
1724 if (!before(seq, TCP_SKB_CB(skb)->end_seq)) {
1725 p = &parent->rb_right;
1733 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1736 if (skb && after(TCP_SKB_CB(skb)->seq, skip_to_seq))
1739 return tcp_sacktag_bsearch(sk, skip_to_seq);
1742 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1744 struct tcp_sack_block *next_dup,
1745 struct tcp_sacktag_state *state,
1751 if (before(next_dup->start_seq, skip_to_seq)) {
1752 skb = tcp_sacktag_skip(skb, sk, next_dup->start_seq);
1753 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1754 next_dup->start_seq, next_dup->end_seq,
1761 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1763 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1767 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1768 u32 prior_snd_una, struct tcp_sacktag_state *state)
1770 struct tcp_sock *tp = tcp_sk(sk);
1771 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1772 TCP_SKB_CB(ack_skb)->sacked);
1773 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1774 struct tcp_sack_block sp[TCP_NUM_SACKS];
1775 struct tcp_sack_block *cache;
1776 struct sk_buff *skb;
1777 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1779 bool found_dup_sack = false;
1781 int first_sack_index;
1784 state->reord = tp->snd_nxt;
1786 if (!tp->sacked_out)
1787 tcp_highest_sack_reset(sk);
1789 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1790 num_sacks, prior_snd_una, state);
1792 /* Eliminate too old ACKs, but take into
1793 * account more or less fresh ones, they can
1794 * contain valid SACK info.
1796 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1799 if (!tp->packets_out)
1803 first_sack_index = 0;
1804 for (i = 0; i < num_sacks; i++) {
1805 bool dup_sack = !i && found_dup_sack;
1807 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1808 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1810 if (!tcp_is_sackblock_valid(tp, dup_sack,
1811 sp[used_sacks].start_seq,
1812 sp[used_sacks].end_seq)) {
1816 if (!tp->undo_marker)
1817 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1819 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1821 /* Don't count olds caused by ACK reordering */
1822 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1823 !after(sp[used_sacks].end_seq, tp->snd_una))
1825 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1828 NET_INC_STATS(sock_net(sk), mib_idx);
1830 first_sack_index = -1;
1834 /* Ignore very old stuff early */
1835 if (!after(sp[used_sacks].end_seq, prior_snd_una)) {
1837 first_sack_index = -1;
1844 /* order SACK blocks to allow in order walk of the retrans queue */
1845 for (i = used_sacks - 1; i > 0; i--) {
1846 for (j = 0; j < i; j++) {
1847 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1848 swap(sp[j], sp[j + 1]);
1850 /* Track where the first SACK block goes to */
1851 if (j == first_sack_index)
1852 first_sack_index = j + 1;
1857 state->mss_now = tcp_current_mss(sk);
1861 if (!tp->sacked_out) {
1862 /* It's already past, so skip checking against it */
1863 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1865 cache = tp->recv_sack_cache;
1866 /* Skip empty blocks in at head of the cache */
1867 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1872 while (i < used_sacks) {
1873 u32 start_seq = sp[i].start_seq;
1874 u32 end_seq = sp[i].end_seq;
1875 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1876 struct tcp_sack_block *next_dup = NULL;
1878 if (found_dup_sack && ((i + 1) == first_sack_index))
1879 next_dup = &sp[i + 1];
1881 /* Skip too early cached blocks */
1882 while (tcp_sack_cache_ok(tp, cache) &&
1883 !before(start_seq, cache->end_seq))
1886 /* Can skip some work by looking recv_sack_cache? */
1887 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1888 after(end_seq, cache->start_seq)) {
1891 if (before(start_seq, cache->start_seq)) {
1892 skb = tcp_sacktag_skip(skb, sk, start_seq);
1893 skb = tcp_sacktag_walk(skb, sk, next_dup,
1900 /* Rest of the block already fully processed? */
1901 if (!after(end_seq, cache->end_seq))
1904 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1908 /* ...tail remains todo... */
1909 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1910 /* ...but better entrypoint exists! */
1911 skb = tcp_highest_sack(sk);
1918 skb = tcp_sacktag_skip(skb, sk, cache->end_seq);
1919 /* Check overlap against next cached too (past this one already) */
1924 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1925 skb = tcp_highest_sack(sk);
1929 skb = tcp_sacktag_skip(skb, sk, start_seq);
1932 skb = tcp_sacktag_walk(skb, sk, next_dup, state,
1933 start_seq, end_seq, dup_sack);
1939 /* Clear the head of the cache sack blocks so we can skip it next time */
1940 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1941 tp->recv_sack_cache[i].start_seq = 0;
1942 tp->recv_sack_cache[i].end_seq = 0;
1944 for (j = 0; j < used_sacks; j++)
1945 tp->recv_sack_cache[i++] = sp[j];
1947 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Loss || tp->undo_marker)
1948 tcp_check_sack_reordering(sk, state->reord, 0);
1950 tcp_verify_left_out(tp);
1953 #if FASTRETRANS_DEBUG > 0
1954 WARN_ON((int)tp->sacked_out < 0);
1955 WARN_ON((int)tp->lost_out < 0);
1956 WARN_ON((int)tp->retrans_out < 0);
1957 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1962 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1963 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1965 static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
1969 holes = max(tp->lost_out, 1U);
1970 holes = min(holes, tp->packets_out);
1972 if ((tp->sacked_out + holes) > tp->packets_out) {
1973 tp->sacked_out = tp->packets_out - holes;
1979 /* If we receive more dupacks than we expected counting segments
1980 * in assumption of absent reordering, interpret this as reordering.
1981 * The only another reason could be bug in receiver TCP.
1983 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1985 struct tcp_sock *tp = tcp_sk(sk);
1987 if (!tcp_limit_reno_sacked(tp))
1990 tp->reordering = min_t(u32, tp->packets_out + addend,
1991 sock_net(sk)->ipv4.sysctl_tcp_max_reordering);
1993 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRENOREORDER);
1996 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1998 static void tcp_add_reno_sack(struct sock *sk, int num_dupack, bool ece_ack)
2001 struct tcp_sock *tp = tcp_sk(sk);
2002 u32 prior_sacked = tp->sacked_out;
2005 tp->sacked_out += num_dupack;
2006 tcp_check_reno_reordering(sk, 0);
2007 delivered = tp->sacked_out - prior_sacked;
2009 tcp_count_delivered(tp, delivered, ece_ack);
2010 tcp_verify_left_out(tp);
2014 /* Account for ACK, ACKing some data in Reno Recovery phase. */
2016 static void tcp_remove_reno_sacks(struct sock *sk, int acked, bool ece_ack)
2018 struct tcp_sock *tp = tcp_sk(sk);
2021 /* One ACK acked hole. The rest eat duplicate ACKs. */
2022 tcp_count_delivered(tp, max_t(int, acked - tp->sacked_out, 1),
2024 if (acked - 1 >= tp->sacked_out)
2027 tp->sacked_out -= acked - 1;
2029 tcp_check_reno_reordering(sk, acked);
2030 tcp_verify_left_out(tp);
2033 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
2038 void tcp_clear_retrans(struct tcp_sock *tp)
2040 tp->retrans_out = 0;
2042 tp->undo_marker = 0;
2043 tp->undo_retrans = -1;
2047 static inline void tcp_init_undo(struct tcp_sock *tp)
2049 tp->undo_marker = tp->snd_una;
2050 /* Retransmission still in flight may cause DSACKs later. */
2051 tp->undo_retrans = tp->retrans_out ? : -1;
2054 static bool tcp_is_rack(const struct sock *sk)
2056 return sock_net(sk)->ipv4.sysctl_tcp_recovery & TCP_RACK_LOSS_DETECTION;
2059 /* If we detect SACK reneging, forget all SACK information
2060 * and reset tags completely, otherwise preserve SACKs. If receiver
2061 * dropped its ofo queue, we will know this due to reneging detection.
2063 static void tcp_timeout_mark_lost(struct sock *sk)
2065 struct tcp_sock *tp = tcp_sk(sk);
2066 struct sk_buff *skb, *head;
2067 bool is_reneg; /* is receiver reneging on SACKs? */
2069 head = tcp_rtx_queue_head(sk);
2070 is_reneg = head && (TCP_SKB_CB(head)->sacked & TCPCB_SACKED_ACKED);
2072 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
2074 /* Mark SACK reneging until we recover from this loss event. */
2075 tp->is_sack_reneg = 1;
2076 } else if (tcp_is_reno(tp)) {
2077 tcp_reset_reno_sack(tp);
2081 skb_rbtree_walk_from(skb) {
2083 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
2084 else if (tcp_is_rack(sk) && skb != head &&
2085 tcp_rack_skb_timeout(tp, skb, 0) > 0)
2086 continue; /* Don't mark recently sent ones lost yet */
2087 tcp_mark_skb_lost(sk, skb);
2089 tcp_verify_left_out(tp);
2090 tcp_clear_all_retrans_hints(tp);
2093 /* Enter Loss state. */
2094 void tcp_enter_loss(struct sock *sk)
2096 const struct inet_connection_sock *icsk = inet_csk(sk);
2097 struct tcp_sock *tp = tcp_sk(sk);
2098 struct net *net = sock_net(sk);
2099 bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery;
2101 tcp_timeout_mark_lost(sk);
2103 /* Reduce ssthresh if it has not yet been made inside this window. */
2104 if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
2105 !after(tp->high_seq, tp->snd_una) ||
2106 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
2107 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2108 tp->prior_cwnd = tp->snd_cwnd;
2109 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2110 tcp_ca_event(sk, CA_EVENT_LOSS);
2113 tp->snd_cwnd = tcp_packets_in_flight(tp) + 1;
2114 tp->snd_cwnd_cnt = 0;
2115 tp->snd_cwnd_stamp = tcp_jiffies32;
2117 /* Timeout in disordered state after receiving substantial DUPACKs
2118 * suggests that the degree of reordering is over-estimated.
2120 if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
2121 tp->sacked_out >= net->ipv4.sysctl_tcp_reordering)
2122 tp->reordering = min_t(unsigned int, tp->reordering,
2123 net->ipv4.sysctl_tcp_reordering);
2124 tcp_set_ca_state(sk, TCP_CA_Loss);
2125 tp->high_seq = tp->snd_nxt;
2126 tcp_ecn_queue_cwr(tp);
2128 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
2129 * loss recovery is underway except recurring timeout(s) on
2130 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
2132 tp->frto = net->ipv4.sysctl_tcp_frto &&
2133 (new_recovery || icsk->icsk_retransmits) &&
2134 !inet_csk(sk)->icsk_mtup.probe_size;
2137 /* If ACK arrived pointing to a remembered SACK, it means that our
2138 * remembered SACKs do not reflect real state of receiver i.e.
2139 * receiver _host_ is heavily congested (or buggy).
2141 * To avoid big spurious retransmission bursts due to transient SACK
2142 * scoreboard oddities that look like reneging, we give the receiver a
2143 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
2144 * restore sanity to the SACK scoreboard. If the apparent reneging
2145 * persists until this RTO then we'll clear the SACK scoreboard.
2147 static bool tcp_check_sack_reneging(struct sock *sk, int flag)
2149 if (flag & FLAG_SACK_RENEGING) {
2150 struct tcp_sock *tp = tcp_sk(sk);
2151 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4),
2152 msecs_to_jiffies(10));
2154 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2155 delay, TCP_RTO_MAX);
2161 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2162 * counter when SACK is enabled (without SACK, sacked_out is used for
2165 * With reordering, holes may still be in flight, so RFC3517 recovery
2166 * uses pure sacked_out (total number of SACKed segments) even though
2167 * it violates the RFC that uses duplicate ACKs, often these are equal
2168 * but when e.g. out-of-window ACKs or packet duplication occurs,
2169 * they differ. Since neither occurs due to loss, TCP should really
2172 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2174 return tp->sacked_out + 1;
2177 /* Linux NewReno/SACK/ECN state machine.
2178 * --------------------------------------
2180 * "Open" Normal state, no dubious events, fast path.
2181 * "Disorder" In all the respects it is "Open",
2182 * but requires a bit more attention. It is entered when
2183 * we see some SACKs or dupacks. It is split of "Open"
2184 * mainly to move some processing from fast path to slow one.
2185 * "CWR" CWND was reduced due to some Congestion Notification event.
2186 * It can be ECN, ICMP source quench, local device congestion.
2187 * "Recovery" CWND was reduced, we are fast-retransmitting.
2188 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2190 * tcp_fastretrans_alert() is entered:
2191 * - each incoming ACK, if state is not "Open"
2192 * - when arrived ACK is unusual, namely:
2197 * Counting packets in flight is pretty simple.
2199 * in_flight = packets_out - left_out + retrans_out
2201 * packets_out is SND.NXT-SND.UNA counted in packets.
2203 * retrans_out is number of retransmitted segments.
2205 * left_out is number of segments left network, but not ACKed yet.
2207 * left_out = sacked_out + lost_out
2209 * sacked_out: Packets, which arrived to receiver out of order
2210 * and hence not ACKed. With SACKs this number is simply
2211 * amount of SACKed data. Even without SACKs
2212 * it is easy to give pretty reliable estimate of this number,
2213 * counting duplicate ACKs.
2215 * lost_out: Packets lost by network. TCP has no explicit
2216 * "loss notification" feedback from network (for now).
2217 * It means that this number can be only _guessed_.
2218 * Actually, it is the heuristics to predict lossage that
2219 * distinguishes different algorithms.
2221 * F.e. after RTO, when all the queue is considered as lost,
2222 * lost_out = packets_out and in_flight = retrans_out.
2224 * Essentially, we have now a few algorithms detecting
2227 * If the receiver supports SACK:
2229 * RFC6675/3517: It is the conventional algorithm. A packet is
2230 * considered lost if the number of higher sequence packets
2231 * SACKed is greater than or equal the DUPACK thoreshold
2232 * (reordering). This is implemented in tcp_mark_head_lost and
2233 * tcp_update_scoreboard.
2235 * RACK (draft-ietf-tcpm-rack-01): it is a newer algorithm
2236 * (2017-) that checks timing instead of counting DUPACKs.
2237 * Essentially a packet is considered lost if it's not S/ACKed
2238 * after RTT + reordering_window, where both metrics are
2239 * dynamically measured and adjusted. This is implemented in
2240 * tcp_rack_mark_lost.
2242 * If the receiver does not support SACK:
2244 * NewReno (RFC6582): in Recovery we assume that one segment
2245 * is lost (classic Reno). While we are in Recovery and
2246 * a partial ACK arrives, we assume that one more packet
2247 * is lost (NewReno). This heuristics are the same in NewReno
2250 * Really tricky (and requiring careful tuning) part of algorithm
2251 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2252 * The first determines the moment _when_ we should reduce CWND and,
2253 * hence, slow down forward transmission. In fact, it determines the moment
2254 * when we decide that hole is caused by loss, rather than by a reorder.
2256 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2257 * holes, caused by lost packets.
2259 * And the most logically complicated part of algorithm is undo
2260 * heuristics. We detect false retransmits due to both too early
2261 * fast retransmit (reordering) and underestimated RTO, analyzing
2262 * timestamps and D-SACKs. When we detect that some segments were
2263 * retransmitted by mistake and CWND reduction was wrong, we undo
2264 * window reduction and abort recovery phase. This logic is hidden
2265 * inside several functions named tcp_try_undo_<something>.
2268 /* This function decides, when we should leave Disordered state
2269 * and enter Recovery phase, reducing congestion window.
2271 * Main question: may we further continue forward transmission
2272 * with the same cwnd?
2274 static bool tcp_time_to_recover(struct sock *sk, int flag)
2276 struct tcp_sock *tp = tcp_sk(sk);
2278 /* Trick#1: The loss is proven. */
2282 /* Not-A-Trick#2 : Classic rule... */
2283 if (!tcp_is_rack(sk) && tcp_dupack_heuristics(tp) > tp->reordering)
2289 /* Detect loss in event "A" above by marking head of queue up as lost.
2290 * For RFC3517 SACK, a segment is considered lost if it
2291 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2292 * the maximum SACKed segments to pass before reaching this limit.
2294 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2296 struct tcp_sock *tp = tcp_sk(sk);
2297 struct sk_buff *skb;
2299 /* Use SACK to deduce losses of new sequences sent during recovery */
2300 const u32 loss_high = tp->snd_nxt;
2302 WARN_ON(packets > tp->packets_out);
2303 skb = tp->lost_skb_hint;
2305 /* Head already handled? */
2306 if (mark_head && after(TCP_SKB_CB(skb)->seq, tp->snd_una))
2308 cnt = tp->lost_cnt_hint;
2310 skb = tcp_rtx_queue_head(sk);
2314 skb_rbtree_walk_from(skb) {
2315 /* TODO: do this better */
2316 /* this is not the most efficient way to do this... */
2317 tp->lost_skb_hint = skb;
2318 tp->lost_cnt_hint = cnt;
2320 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2323 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2324 cnt += tcp_skb_pcount(skb);
2329 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_LOST))
2330 tcp_mark_skb_lost(sk, skb);
2335 tcp_verify_left_out(tp);
2338 /* Account newly detected lost packet(s) */
2340 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2342 struct tcp_sock *tp = tcp_sk(sk);
2344 if (tcp_is_sack(tp)) {
2345 int sacked_upto = tp->sacked_out - tp->reordering;
2346 if (sacked_upto >= 0)
2347 tcp_mark_head_lost(sk, sacked_upto, 0);
2348 else if (fast_rexmit)
2349 tcp_mark_head_lost(sk, 1, 1);
2353 static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when)
2355 return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2356 before(tp->rx_opt.rcv_tsecr, when);
2359 /* skb is spurious retransmitted if the returned timestamp echo
2360 * reply is prior to the skb transmission time
2362 static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp,
2363 const struct sk_buff *skb)
2365 return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) &&
2366 tcp_tsopt_ecr_before(tp, tcp_skb_timestamp(skb));
2369 /* Nothing was retransmitted or returned timestamp is less
2370 * than timestamp of the first retransmission.
2372 static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2374 return tp->retrans_stamp &&
2375 tcp_tsopt_ecr_before(tp, tp->retrans_stamp);
2378 /* Undo procedures. */
2380 /* We can clear retrans_stamp when there are no retransmissions in the
2381 * window. It would seem that it is trivially available for us in
2382 * tp->retrans_out, however, that kind of assumptions doesn't consider
2383 * what will happen if errors occur when sending retransmission for the
2384 * second time. ...It could the that such segment has only
2385 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2386 * the head skb is enough except for some reneging corner cases that
2387 * are not worth the effort.
2389 * Main reason for all this complexity is the fact that connection dying
2390 * time now depends on the validity of the retrans_stamp, in particular,
2391 * that successive retransmissions of a segment must not advance
2392 * retrans_stamp under any conditions.
2394 static bool tcp_any_retrans_done(const struct sock *sk)
2396 const struct tcp_sock *tp = tcp_sk(sk);
2397 struct sk_buff *skb;
2399 if (tp->retrans_out)
2402 skb = tcp_rtx_queue_head(sk);
2403 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2409 static void DBGUNDO(struct sock *sk, const char *msg)
2411 #if FASTRETRANS_DEBUG > 1
2412 struct tcp_sock *tp = tcp_sk(sk);
2413 struct inet_sock *inet = inet_sk(sk);
2415 if (sk->sk_family == AF_INET) {
2416 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2418 &inet->inet_daddr, ntohs(inet->inet_dport),
2419 tp->snd_cwnd, tcp_left_out(tp),
2420 tp->snd_ssthresh, tp->prior_ssthresh,
2423 #if IS_ENABLED(CONFIG_IPV6)
2424 else if (sk->sk_family == AF_INET6) {
2425 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2427 &sk->sk_v6_daddr, ntohs(inet->inet_dport),
2428 tp->snd_cwnd, tcp_left_out(tp),
2429 tp->snd_ssthresh, tp->prior_ssthresh,
2436 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
2438 struct tcp_sock *tp = tcp_sk(sk);
2441 struct sk_buff *skb;
2443 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2444 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2447 tcp_clear_all_retrans_hints(tp);
2450 if (tp->prior_ssthresh) {
2451 const struct inet_connection_sock *icsk = inet_csk(sk);
2453 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2455 if (tp->prior_ssthresh > tp->snd_ssthresh) {
2456 tp->snd_ssthresh = tp->prior_ssthresh;
2457 tcp_ecn_withdraw_cwr(tp);
2460 tp->snd_cwnd_stamp = tcp_jiffies32;
2461 tp->undo_marker = 0;
2462 tp->rack.advanced = 1; /* Force RACK to re-exam losses */
2465 static inline bool tcp_may_undo(const struct tcp_sock *tp)
2467 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2470 /* People celebrate: "We love our President!" */
2471 static bool tcp_try_undo_recovery(struct sock *sk)
2473 struct tcp_sock *tp = tcp_sk(sk);
2475 if (tcp_may_undo(tp)) {
2478 /* Happy end! We did not retransmit anything
2479 * or our original transmission succeeded.
2481 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2482 tcp_undo_cwnd_reduction(sk, false);
2483 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2484 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2486 mib_idx = LINUX_MIB_TCPFULLUNDO;
2488 NET_INC_STATS(sock_net(sk), mib_idx);
2489 } else if (tp->rack.reo_wnd_persist) {
2490 tp->rack.reo_wnd_persist--;
2492 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2493 /* Hold old state until something *above* high_seq
2494 * is ACKed. For Reno it is MUST to prevent false
2495 * fast retransmits (RFC2582). SACK TCP is safe. */
2496 if (!tcp_any_retrans_done(sk))
2497 tp->retrans_stamp = 0;
2500 tcp_set_ca_state(sk, TCP_CA_Open);
2501 tp->is_sack_reneg = 0;
2505 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2506 static bool tcp_try_undo_dsack(struct sock *sk)
2508 struct tcp_sock *tp = tcp_sk(sk);
2510 if (tp->undo_marker && !tp->undo_retrans) {
2511 tp->rack.reo_wnd_persist = min(TCP_RACK_RECOVERY_THRESH,
2512 tp->rack.reo_wnd_persist + 1);
2513 DBGUNDO(sk, "D-SACK");
2514 tcp_undo_cwnd_reduction(sk, false);
2515 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2521 /* Undo during loss recovery after partial ACK or using F-RTO. */
2522 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2524 struct tcp_sock *tp = tcp_sk(sk);
2526 if (frto_undo || tcp_may_undo(tp)) {
2527 tcp_undo_cwnd_reduction(sk, true);
2529 DBGUNDO(sk, "partial loss");
2530 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2532 NET_INC_STATS(sock_net(sk),
2533 LINUX_MIB_TCPSPURIOUSRTOS);
2534 inet_csk(sk)->icsk_retransmits = 0;
2535 if (frto_undo || tcp_is_sack(tp)) {
2536 tcp_set_ca_state(sk, TCP_CA_Open);
2537 tp->is_sack_reneg = 0;
2544 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
2545 * It computes the number of packets to send (sndcnt) based on packets newly
2547 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2548 * cwnd reductions across a full RTT.
2549 * 2) Otherwise PRR uses packet conservation to send as much as delivered.
2550 * But when SND_UNA is acked without further losses,
2551 * slow starts cwnd up to ssthresh to speed up the recovery.
2553 static void tcp_init_cwnd_reduction(struct sock *sk)
2555 struct tcp_sock *tp = tcp_sk(sk);
2557 tp->high_seq = tp->snd_nxt;
2558 tp->tlp_high_seq = 0;
2559 tp->snd_cwnd_cnt = 0;
2560 tp->prior_cwnd = tp->snd_cwnd;
2561 tp->prr_delivered = 0;
2563 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2564 tcp_ecn_queue_cwr(tp);
2567 void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked, int newly_lost, int flag)
2569 struct tcp_sock *tp = tcp_sk(sk);
2571 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2573 if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd))
2576 tp->prr_delivered += newly_acked_sacked;
2578 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2580 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2581 } else if (flag & FLAG_SND_UNA_ADVANCED && !newly_lost) {
2582 sndcnt = min_t(int, delta,
2583 max_t(int, tp->prr_delivered - tp->prr_out,
2584 newly_acked_sacked) + 1);
2586 sndcnt = min(delta, newly_acked_sacked);
2588 /* Force a fast retransmit upon entering fast recovery */
2589 sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1));
2590 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2593 static inline void tcp_end_cwnd_reduction(struct sock *sk)
2595 struct tcp_sock *tp = tcp_sk(sk);
2597 if (inet_csk(sk)->icsk_ca_ops->cong_control)
2600 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2601 if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH &&
2602 (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR || tp->undo_marker)) {
2603 tp->snd_cwnd = tp->snd_ssthresh;
2604 tp->snd_cwnd_stamp = tcp_jiffies32;
2606 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2609 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2610 void tcp_enter_cwr(struct sock *sk)
2612 struct tcp_sock *tp = tcp_sk(sk);
2614 tp->prior_ssthresh = 0;
2615 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2616 tp->undo_marker = 0;
2617 tcp_init_cwnd_reduction(sk);
2618 tcp_set_ca_state(sk, TCP_CA_CWR);
2621 EXPORT_SYMBOL(tcp_enter_cwr);
2623 static void tcp_try_keep_open(struct sock *sk)
2625 struct tcp_sock *tp = tcp_sk(sk);
2626 int state = TCP_CA_Open;
2628 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2629 state = TCP_CA_Disorder;
2631 if (inet_csk(sk)->icsk_ca_state != state) {
2632 tcp_set_ca_state(sk, state);
2633 tp->high_seq = tp->snd_nxt;
2637 static void tcp_try_to_open(struct sock *sk, int flag)
2639 struct tcp_sock *tp = tcp_sk(sk);
2641 tcp_verify_left_out(tp);
2643 if (!tcp_any_retrans_done(sk))
2644 tp->retrans_stamp = 0;
2646 if (flag & FLAG_ECE)
2649 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2650 tcp_try_keep_open(sk);
2654 static void tcp_mtup_probe_failed(struct sock *sk)
2656 struct inet_connection_sock *icsk = inet_csk(sk);
2658 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2659 icsk->icsk_mtup.probe_size = 0;
2660 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPFAIL);
2663 static void tcp_mtup_probe_success(struct sock *sk)
2665 struct tcp_sock *tp = tcp_sk(sk);
2666 struct inet_connection_sock *icsk = inet_csk(sk);
2668 /* FIXME: breaks with very large cwnd */
2669 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2670 tp->snd_cwnd = tp->snd_cwnd *
2671 tcp_mss_to_mtu(sk, tp->mss_cache) /
2672 icsk->icsk_mtup.probe_size;
2673 tp->snd_cwnd_cnt = 0;
2674 tp->snd_cwnd_stamp = tcp_jiffies32;
2675 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2677 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2678 icsk->icsk_mtup.probe_size = 0;
2679 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2680 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS);
2683 /* Do a simple retransmit without using the backoff mechanisms in
2684 * tcp_timer. This is used for path mtu discovery.
2685 * The socket is already locked here.
2687 void tcp_simple_retransmit(struct sock *sk)
2689 const struct inet_connection_sock *icsk = inet_csk(sk);
2690 struct tcp_sock *tp = tcp_sk(sk);
2691 struct sk_buff *skb;
2694 /* A fastopen SYN request is stored as two separate packets within
2695 * the retransmit queue, this is done by tcp_send_syn_data().
2696 * As a result simply checking the MSS of the frames in the queue
2697 * will not work for the SYN packet.
2699 * Us being here is an indication of a path MTU issue so we can
2700 * assume that the fastopen SYN was lost and just mark all the
2701 * frames in the retransmit queue as lost. We will use an MSS of
2702 * -1 to mark all frames as lost, otherwise compute the current MSS.
2704 if (tp->syn_data && sk->sk_state == TCP_SYN_SENT)
2707 mss = tcp_current_mss(sk);
2709 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2710 if (tcp_skb_seglen(skb) > mss)
2711 tcp_mark_skb_lost(sk, skb);
2714 tcp_clear_retrans_hints_partial(tp);
2719 if (tcp_is_reno(tp))
2720 tcp_limit_reno_sacked(tp);
2722 tcp_verify_left_out(tp);
2724 /* Don't muck with the congestion window here.
2725 * Reason is that we do not increase amount of _data_
2726 * in network, but units changed and effective
2727 * cwnd/ssthresh really reduced now.
2729 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2730 tp->high_seq = tp->snd_nxt;
2731 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2732 tp->prior_ssthresh = 0;
2733 tp->undo_marker = 0;
2734 tcp_set_ca_state(sk, TCP_CA_Loss);
2736 tcp_xmit_retransmit_queue(sk);
2738 EXPORT_SYMBOL(tcp_simple_retransmit);
2740 void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2742 struct tcp_sock *tp = tcp_sk(sk);
2745 if (tcp_is_reno(tp))
2746 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2748 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2750 NET_INC_STATS(sock_net(sk), mib_idx);
2752 tp->prior_ssthresh = 0;
2755 if (!tcp_in_cwnd_reduction(sk)) {
2757 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2758 tcp_init_cwnd_reduction(sk);
2760 tcp_set_ca_state(sk, TCP_CA_Recovery);
2763 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2764 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2766 static void tcp_process_loss(struct sock *sk, int flag, int num_dupack,
2769 struct tcp_sock *tp = tcp_sk(sk);
2770 bool recovered = !before(tp->snd_una, tp->high_seq);
2772 if ((flag & FLAG_SND_UNA_ADVANCED || rcu_access_pointer(tp->fastopen_rsk)) &&
2773 tcp_try_undo_loss(sk, false))
2776 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2777 /* Step 3.b. A timeout is spurious if not all data are
2778 * lost, i.e., never-retransmitted data are (s)acked.
2780 if ((flag & FLAG_ORIG_SACK_ACKED) &&
2781 tcp_try_undo_loss(sk, true))
2784 if (after(tp->snd_nxt, tp->high_seq)) {
2785 if (flag & FLAG_DATA_SACKED || num_dupack)
2786 tp->frto = 0; /* Step 3.a. loss was real */
2787 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2788 tp->high_seq = tp->snd_nxt;
2789 /* Step 2.b. Try send new data (but deferred until cwnd
2790 * is updated in tcp_ack()). Otherwise fall back to
2791 * the conventional recovery.
2793 if (!tcp_write_queue_empty(sk) &&
2794 after(tcp_wnd_end(tp), tp->snd_nxt)) {
2795 *rexmit = REXMIT_NEW;
2803 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2804 tcp_try_undo_recovery(sk);
2807 if (tcp_is_reno(tp)) {
2808 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2809 * delivered. Lower inflight to clock out (re)tranmissions.
2811 if (after(tp->snd_nxt, tp->high_seq) && num_dupack)
2812 tcp_add_reno_sack(sk, num_dupack, flag & FLAG_ECE);
2813 else if (flag & FLAG_SND_UNA_ADVANCED)
2814 tcp_reset_reno_sack(tp);
2816 *rexmit = REXMIT_LOST;
2819 /* Undo during fast recovery after partial ACK. */
2820 static bool tcp_try_undo_partial(struct sock *sk, u32 prior_snd_una)
2822 struct tcp_sock *tp = tcp_sk(sk);
2824 if (tp->undo_marker && tcp_packet_delayed(tp)) {
2825 /* Plain luck! Hole if filled with delayed
2826 * packet, rather than with a retransmit. Check reordering.
2828 tcp_check_sack_reordering(sk, prior_snd_una, 1);
2830 /* We are getting evidence that the reordering degree is higher
2831 * than we realized. If there are no retransmits out then we
2832 * can undo. Otherwise we clock out new packets but do not
2833 * mark more packets lost or retransmit more.
2835 if (tp->retrans_out)
2838 if (!tcp_any_retrans_done(sk))
2839 tp->retrans_stamp = 0;
2841 DBGUNDO(sk, "partial recovery");
2842 tcp_undo_cwnd_reduction(sk, true);
2843 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2844 tcp_try_keep_open(sk);
2850 static void tcp_identify_packet_loss(struct sock *sk, int *ack_flag)
2852 struct tcp_sock *tp = tcp_sk(sk);
2854 if (tcp_rtx_queue_empty(sk))
2857 if (unlikely(tcp_is_reno(tp))) {
2858 tcp_newreno_mark_lost(sk, *ack_flag & FLAG_SND_UNA_ADVANCED);
2859 } else if (tcp_is_rack(sk)) {
2860 u32 prior_retrans = tp->retrans_out;
2862 if (tcp_rack_mark_lost(sk))
2863 *ack_flag &= ~FLAG_SET_XMIT_TIMER;
2864 if (prior_retrans > tp->retrans_out)
2865 *ack_flag |= FLAG_LOST_RETRANS;
2869 static bool tcp_force_fast_retransmit(struct sock *sk)
2871 struct tcp_sock *tp = tcp_sk(sk);
2873 return after(tcp_highest_sack_seq(tp),
2874 tp->snd_una + tp->reordering * tp->mss_cache);
2877 /* Process an event, which can update packets-in-flight not trivially.
2878 * Main goal of this function is to calculate new estimate for left_out,
2879 * taking into account both packets sitting in receiver's buffer and
2880 * packets lost by network.
2882 * Besides that it updates the congestion state when packet loss or ECN
2883 * is detected. But it does not reduce the cwnd, it is done by the
2884 * congestion control later.
2886 * It does _not_ decide what to send, it is made in function
2887 * tcp_xmit_retransmit_queue().
2889 static void tcp_fastretrans_alert(struct sock *sk, const u32 prior_snd_una,
2890 int num_dupack, int *ack_flag, int *rexmit)
2892 struct inet_connection_sock *icsk = inet_csk(sk);
2893 struct tcp_sock *tp = tcp_sk(sk);
2894 int fast_rexmit = 0, flag = *ack_flag;
2895 bool ece_ack = flag & FLAG_ECE;
2896 bool do_lost = num_dupack || ((flag & FLAG_DATA_SACKED) &&
2897 tcp_force_fast_retransmit(sk));
2899 if (!tp->packets_out && tp->sacked_out)
2902 /* Now state machine starts.
2903 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2905 tp->prior_ssthresh = 0;
2907 /* B. In all the states check for reneging SACKs. */
2908 if (tcp_check_sack_reneging(sk, flag))
2911 /* C. Check consistency of the current state. */
2912 tcp_verify_left_out(tp);
2914 /* D. Check state exit conditions. State can be terminated
2915 * when high_seq is ACKed. */
2916 if (icsk->icsk_ca_state == TCP_CA_Open) {
2917 WARN_ON(tp->retrans_out != 0);
2918 tp->retrans_stamp = 0;
2919 } else if (!before(tp->snd_una, tp->high_seq)) {
2920 switch (icsk->icsk_ca_state) {
2922 /* CWR is to be held something *above* high_seq
2923 * is ACKed for CWR bit to reach receiver. */
2924 if (tp->snd_una != tp->high_seq) {
2925 tcp_end_cwnd_reduction(sk);
2926 tcp_set_ca_state(sk, TCP_CA_Open);
2930 case TCP_CA_Recovery:
2931 if (tcp_is_reno(tp))
2932 tcp_reset_reno_sack(tp);
2933 if (tcp_try_undo_recovery(sk))
2935 tcp_end_cwnd_reduction(sk);
2940 /* E. Process state. */
2941 switch (icsk->icsk_ca_state) {
2942 case TCP_CA_Recovery:
2943 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2944 if (tcp_is_reno(tp))
2945 tcp_add_reno_sack(sk, num_dupack, ece_ack);
2947 if (tcp_try_undo_partial(sk, prior_snd_una))
2949 /* Partial ACK arrived. Force fast retransmit. */
2950 do_lost = tcp_force_fast_retransmit(sk);
2952 if (tcp_try_undo_dsack(sk)) {
2953 tcp_try_keep_open(sk);
2956 tcp_identify_packet_loss(sk, ack_flag);
2959 tcp_process_loss(sk, flag, num_dupack, rexmit);
2960 tcp_identify_packet_loss(sk, ack_flag);
2961 if (!(icsk->icsk_ca_state == TCP_CA_Open ||
2962 (*ack_flag & FLAG_LOST_RETRANS)))
2964 /* Change state if cwnd is undone or retransmits are lost */
2967 if (tcp_is_reno(tp)) {
2968 if (flag & FLAG_SND_UNA_ADVANCED)
2969 tcp_reset_reno_sack(tp);
2970 tcp_add_reno_sack(sk, num_dupack, ece_ack);
2973 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
2974 tcp_try_undo_dsack(sk);
2976 tcp_identify_packet_loss(sk, ack_flag);
2977 if (!tcp_time_to_recover(sk, flag)) {
2978 tcp_try_to_open(sk, flag);
2982 /* MTU probe failure: don't reduce cwnd */
2983 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2984 icsk->icsk_mtup.probe_size &&
2985 tp->snd_una == tp->mtu_probe.probe_seq_start) {
2986 tcp_mtup_probe_failed(sk);
2987 /* Restores the reduction we did in tcp_mtup_probe() */
2989 tcp_simple_retransmit(sk);
2993 /* Otherwise enter Recovery state */
2994 tcp_enter_recovery(sk, ece_ack);
2998 if (!tcp_is_rack(sk) && do_lost)
2999 tcp_update_scoreboard(sk, fast_rexmit);
3000 *rexmit = REXMIT_LOST;
3003 static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us, const int flag)
3005 u32 wlen = sock_net(sk)->ipv4.sysctl_tcp_min_rtt_wlen * HZ;
3006 struct tcp_sock *tp = tcp_sk(sk);
3008 if ((flag & FLAG_ACK_MAYBE_DELAYED) && rtt_us > tcp_min_rtt(tp)) {
3009 /* If the remote keeps returning delayed ACKs, eventually
3010 * the min filter would pick it up and overestimate the
3011 * prop. delay when it expires. Skip suspected delayed ACKs.
3015 minmax_running_min(&tp->rtt_min, wlen, tcp_jiffies32,
3016 rtt_us ? : jiffies_to_usecs(1));
3019 static bool tcp_ack_update_rtt(struct sock *sk, const int flag,
3020 long seq_rtt_us, long sack_rtt_us,
3021 long ca_rtt_us, struct rate_sample *rs)
3023 const struct tcp_sock *tp = tcp_sk(sk);
3025 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
3026 * broken middle-boxes or peers may corrupt TS-ECR fields. But
3027 * Karn's algorithm forbids taking RTT if some retransmitted data
3028 * is acked (RFC6298).
3031 seq_rtt_us = sack_rtt_us;
3033 /* RTTM Rule: A TSecr value received in a segment is used to
3034 * update the averaged RTT measurement only if the segment
3035 * acknowledges some new data, i.e., only if it advances the
3036 * left edge of the send window.
3037 * See draft-ietf-tcplw-high-performance-00, section 3.3.
3039 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
3040 flag & FLAG_ACKED) {
3041 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr;
3043 if (likely(delta < INT_MAX / (USEC_PER_SEC / TCP_TS_HZ))) {
3046 seq_rtt_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
3047 ca_rtt_us = seq_rtt_us;
3050 rs->rtt_us = ca_rtt_us; /* RTT of last (S)ACKed packet (or -1) */
3054 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
3055 * always taken together with ACK, SACK, or TS-opts. Any negative
3056 * values will be skipped with the seq_rtt_us < 0 check above.
3058 tcp_update_rtt_min(sk, ca_rtt_us, flag);
3059 tcp_rtt_estimator(sk, seq_rtt_us);
3062 /* RFC6298: only reset backoff on valid RTT measurement. */
3063 inet_csk(sk)->icsk_backoff = 0;
3067 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
3068 void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req)
3070 struct rate_sample rs;
3073 if (req && !req->num_retrans && tcp_rsk(req)->snt_synack)
3074 rtt_us = tcp_stamp_us_delta(tcp_clock_us(), tcp_rsk(req)->snt_synack);
3076 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us, &rs);
3080 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
3082 const struct inet_connection_sock *icsk = inet_csk(sk);
3084 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
3085 tcp_sk(sk)->snd_cwnd_stamp = tcp_jiffies32;
3088 /* Restart timer after forward progress on connection.
3089 * RFC2988 recommends to restart timer to now+rto.
3091 void tcp_rearm_rto(struct sock *sk)
3093 const struct inet_connection_sock *icsk = inet_csk(sk);
3094 struct tcp_sock *tp = tcp_sk(sk);
3096 /* If the retrans timer is currently being used by Fast Open
3097 * for SYN-ACK retrans purpose, stay put.
3099 if (rcu_access_pointer(tp->fastopen_rsk))
3102 if (!tp->packets_out) {
3103 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
3105 u32 rto = inet_csk(sk)->icsk_rto;
3106 /* Offset the time elapsed after installing regular RTO */
3107 if (icsk->icsk_pending == ICSK_TIME_REO_TIMEOUT ||
3108 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
3109 s64 delta_us = tcp_rto_delta_us(sk);
3110 /* delta_us may not be positive if the socket is locked
3111 * when the retrans timer fires and is rescheduled.
3113 rto = usecs_to_jiffies(max_t(int, delta_us, 1));
3115 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
3120 /* Try to schedule a loss probe; if that doesn't work, then schedule an RTO. */
3121 static void tcp_set_xmit_timer(struct sock *sk)
3123 if (!tcp_schedule_loss_probe(sk, true))
3127 /* If we get here, the whole TSO packet has not been acked. */
3128 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3130 struct tcp_sock *tp = tcp_sk(sk);
3133 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3135 packets_acked = tcp_skb_pcount(skb);
3136 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3138 packets_acked -= tcp_skb_pcount(skb);
3140 if (packets_acked) {
3141 BUG_ON(tcp_skb_pcount(skb) == 0);
3142 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3145 return packets_acked;
3148 static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
3149 const struct sk_buff *ack_skb, u32 prior_snd_una)
3151 const struct skb_shared_info *shinfo;
3153 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3154 if (likely(!TCP_SKB_CB(skb)->txstamp_ack))
3157 shinfo = skb_shinfo(skb);
3158 if (!before(shinfo->tskey, prior_snd_una) &&
3159 before(shinfo->tskey, tcp_sk(sk)->snd_una)) {
3160 tcp_skb_tsorted_save(skb) {
3161 __skb_tstamp_tx(skb, ack_skb, NULL, sk, SCM_TSTAMP_ACK);
3162 } tcp_skb_tsorted_restore(skb);
3166 /* Remove acknowledged frames from the retransmission queue. If our packet
3167 * is before the ack sequence we can discard it as it's confirmed to have
3168 * arrived at the other end.
3170 static int tcp_clean_rtx_queue(struct sock *sk, const struct sk_buff *ack_skb,
3171 u32 prior_fack, u32 prior_snd_una,
3172 struct tcp_sacktag_state *sack, bool ece_ack)
3174 const struct inet_connection_sock *icsk = inet_csk(sk);
3175 u64 first_ackt, last_ackt;
3176 struct tcp_sock *tp = tcp_sk(sk);
3177 u32 prior_sacked = tp->sacked_out;
3178 u32 reord = tp->snd_nxt; /* lowest acked un-retx un-sacked seq */
3179 struct sk_buff *skb, *next;
3180 bool fully_acked = true;
3181 long sack_rtt_us = -1L;
3182 long seq_rtt_us = -1L;
3183 long ca_rtt_us = -1L;
3185 u32 last_in_flight = 0;
3191 for (skb = skb_rb_first(&sk->tcp_rtx_queue); skb; skb = next) {
3192 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3193 const u32 start_seq = scb->seq;
3194 u8 sacked = scb->sacked;
3197 /* Determine how many packets and what bytes were acked, tso and else */
3198 if (after(scb->end_seq, tp->snd_una)) {
3199 if (tcp_skb_pcount(skb) == 1 ||
3200 !after(tp->snd_una, scb->seq))
3203 acked_pcount = tcp_tso_acked(sk, skb);
3206 fully_acked = false;
3208 acked_pcount = tcp_skb_pcount(skb);
3211 if (unlikely(sacked & TCPCB_RETRANS)) {
3212 if (sacked & TCPCB_SACKED_RETRANS)
3213 tp->retrans_out -= acked_pcount;
3214 flag |= FLAG_RETRANS_DATA_ACKED;
3215 } else if (!(sacked & TCPCB_SACKED_ACKED)) {
3216 last_ackt = tcp_skb_timestamp_us(skb);
3217 WARN_ON_ONCE(last_ackt == 0);
3219 first_ackt = last_ackt;
3221 last_in_flight = TCP_SKB_CB(skb)->tx.in_flight;
3222 if (before(start_seq, reord))
3224 if (!after(scb->end_seq, tp->high_seq))
3225 flag |= FLAG_ORIG_SACK_ACKED;
3228 if (sacked & TCPCB_SACKED_ACKED) {
3229 tp->sacked_out -= acked_pcount;
3230 } else if (tcp_is_sack(tp)) {
3231 tcp_count_delivered(tp, acked_pcount, ece_ack);
3232 if (!tcp_skb_spurious_retrans(tp, skb))
3233 tcp_rack_advance(tp, sacked, scb->end_seq,
3234 tcp_skb_timestamp_us(skb));
3236 if (sacked & TCPCB_LOST)
3237 tp->lost_out -= acked_pcount;
3239 tp->packets_out -= acked_pcount;
3240 pkts_acked += acked_pcount;
3241 tcp_rate_skb_delivered(sk, skb, sack->rate);
3243 /* Initial outgoing SYN's get put onto the write_queue
3244 * just like anything else we transmit. It is not
3245 * true data, and if we misinform our callers that
3246 * this ACK acks real data, we will erroneously exit
3247 * connection startup slow start one packet too
3248 * quickly. This is severely frowned upon behavior.
3250 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
3251 flag |= FLAG_DATA_ACKED;
3253 flag |= FLAG_SYN_ACKED;
3254 tp->retrans_stamp = 0;
3260 tcp_ack_tstamp(sk, skb, ack_skb, prior_snd_una);
3262 next = skb_rb_next(skb);
3263 if (unlikely(skb == tp->retransmit_skb_hint))
3264 tp->retransmit_skb_hint = NULL;
3265 if (unlikely(skb == tp->lost_skb_hint))
3266 tp->lost_skb_hint = NULL;
3267 tcp_highest_sack_replace(sk, skb, next);
3268 tcp_rtx_queue_unlink_and_free(skb, sk);
3272 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
3274 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3275 tp->snd_up = tp->snd_una;
3278 tcp_ack_tstamp(sk, skb, ack_skb, prior_snd_una);
3279 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
3280 flag |= FLAG_SACK_RENEGING;
3283 if (likely(first_ackt) && !(flag & FLAG_RETRANS_DATA_ACKED)) {
3284 seq_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, first_ackt);
3285 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, last_ackt);
3287 if (pkts_acked == 1 && last_in_flight < tp->mss_cache &&
3288 last_in_flight && !prior_sacked && fully_acked &&
3289 sack->rate->prior_delivered + 1 == tp->delivered &&
3290 !(flag & (FLAG_CA_ALERT | FLAG_SYN_ACKED))) {
3291 /* Conservatively mark a delayed ACK. It's typically
3292 * from a lone runt packet over the round trip to
3293 * a receiver w/o out-of-order or CE events.
3295 flag |= FLAG_ACK_MAYBE_DELAYED;
3298 if (sack->first_sackt) {
3299 sack_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->first_sackt);
3300 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->last_sackt);
3302 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us,
3303 ca_rtt_us, sack->rate);
3305 if (flag & FLAG_ACKED) {
3306 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3307 if (unlikely(icsk->icsk_mtup.probe_size &&
3308 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3309 tcp_mtup_probe_success(sk);
3312 if (tcp_is_reno(tp)) {
3313 tcp_remove_reno_sacks(sk, pkts_acked, ece_ack);
3315 /* If any of the cumulatively ACKed segments was
3316 * retransmitted, non-SACK case cannot confirm that
3317 * progress was due to original transmission due to
3318 * lack of TCPCB_SACKED_ACKED bits even if some of
3319 * the packets may have been never retransmitted.
3321 if (flag & FLAG_RETRANS_DATA_ACKED)
3322 flag &= ~FLAG_ORIG_SACK_ACKED;
3326 /* Non-retransmitted hole got filled? That's reordering */
3327 if (before(reord, prior_fack))
3328 tcp_check_sack_reordering(sk, reord, 0);
3330 delta = prior_sacked - tp->sacked_out;
3331 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3333 } else if (skb && rtt_update && sack_rtt_us >= 0 &&
3334 sack_rtt_us > tcp_stamp_us_delta(tp->tcp_mstamp,
3335 tcp_skb_timestamp_us(skb))) {
3336 /* Do not re-arm RTO if the sack RTT is measured from data sent
3337 * after when the head was last (re)transmitted. Otherwise the
3338 * timeout may continue to extend in loss recovery.
3340 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3343 if (icsk->icsk_ca_ops->pkts_acked) {
3344 struct ack_sample sample = { .pkts_acked = pkts_acked,
3345 .rtt_us = sack->rate->rtt_us,
3346 .in_flight = last_in_flight };
3348 icsk->icsk_ca_ops->pkts_acked(sk, &sample);
3351 #if FASTRETRANS_DEBUG > 0
3352 WARN_ON((int)tp->sacked_out < 0);
3353 WARN_ON((int)tp->lost_out < 0);
3354 WARN_ON((int)tp->retrans_out < 0);
3355 if (!tp->packets_out && tcp_is_sack(tp)) {
3356 icsk = inet_csk(sk);
3358 pr_debug("Leak l=%u %d\n",
3359 tp->lost_out, icsk->icsk_ca_state);
3362 if (tp->sacked_out) {
3363 pr_debug("Leak s=%u %d\n",
3364 tp->sacked_out, icsk->icsk_ca_state);
3367 if (tp->retrans_out) {
3368 pr_debug("Leak r=%u %d\n",
3369 tp->retrans_out, icsk->icsk_ca_state);
3370 tp->retrans_out = 0;
3377 static void tcp_ack_probe(struct sock *sk)
3379 struct inet_connection_sock *icsk = inet_csk(sk);
3380 struct sk_buff *head = tcp_send_head(sk);
3381 const struct tcp_sock *tp = tcp_sk(sk);
3383 /* Was it a usable window open? */
3386 if (!after(TCP_SKB_CB(head)->end_seq, tcp_wnd_end(tp))) {
3387 icsk->icsk_backoff = 0;
3388 icsk->icsk_probes_tstamp = 0;
3389 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3390 /* Socket must be waked up by subsequent tcp_data_snd_check().
3391 * This function is not for random using!
3394 unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX);
3396 when = tcp_clamp_probe0_to_user_timeout(sk, when);
3397 tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, when, TCP_RTO_MAX);
3401 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3403 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3404 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3407 /* Decide wheather to run the increase function of congestion control. */
3408 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3410 /* If reordering is high then always grow cwnd whenever data is
3411 * delivered regardless of its ordering. Otherwise stay conservative
3412 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3413 * new SACK or ECE mark may first advance cwnd here and later reduce
3414 * cwnd in tcp_fastretrans_alert() based on more states.
3416 if (tcp_sk(sk)->reordering > sock_net(sk)->ipv4.sysctl_tcp_reordering)
3417 return flag & FLAG_FORWARD_PROGRESS;
3419 return flag & FLAG_DATA_ACKED;
3422 /* The "ultimate" congestion control function that aims to replace the rigid
3423 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction).
3424 * It's called toward the end of processing an ACK with precise rate
3425 * information. All transmission or retransmission are delayed afterwards.
3427 static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked,
3428 int flag, const struct rate_sample *rs)
3430 const struct inet_connection_sock *icsk = inet_csk(sk);
3432 if (icsk->icsk_ca_ops->cong_control) {
3433 icsk->icsk_ca_ops->cong_control(sk, rs);
3437 if (tcp_in_cwnd_reduction(sk)) {
3438 /* Reduce cwnd if state mandates */
3439 tcp_cwnd_reduction(sk, acked_sacked, rs->losses, flag);
3440 } else if (tcp_may_raise_cwnd(sk, flag)) {
3441 /* Advance cwnd if state allows */
3442 tcp_cong_avoid(sk, ack, acked_sacked);
3444 tcp_update_pacing_rate(sk);
3447 /* Check that window update is acceptable.
3448 * The function assumes that snd_una<=ack<=snd_next.
3450 static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3451 const u32 ack, const u32 ack_seq,
3454 return after(ack, tp->snd_una) ||
3455 after(ack_seq, tp->snd_wl1) ||
3456 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3459 /* If we update tp->snd_una, also update tp->bytes_acked */
3460 static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack)
3462 u32 delta = ack - tp->snd_una;
3464 sock_owned_by_me((struct sock *)tp);
3465 tp->bytes_acked += delta;
3469 /* If we update tp->rcv_nxt, also update tp->bytes_received */
3470 static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq)
3472 u32 delta = seq - tp->rcv_nxt;
3474 sock_owned_by_me((struct sock *)tp);
3475 tp->bytes_received += delta;
3476 WRITE_ONCE(tp->rcv_nxt, seq);
3479 /* Update our send window.
3481 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3482 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3484 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3487 struct tcp_sock *tp = tcp_sk(sk);
3489 u32 nwin = ntohs(tcp_hdr(skb)->window);
3491 if (likely(!tcp_hdr(skb)->syn))
3492 nwin <<= tp->rx_opt.snd_wscale;
3494 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3495 flag |= FLAG_WIN_UPDATE;
3496 tcp_update_wl(tp, ack_seq);
3498 if (tp->snd_wnd != nwin) {
3501 /* Note, it is the only place, where
3502 * fast path is recovered for sending TCP.
3505 tcp_fast_path_check(sk);
3507 if (!tcp_write_queue_empty(sk))
3508 tcp_slow_start_after_idle_check(sk);
3510 if (nwin > tp->max_window) {
3511 tp->max_window = nwin;
3512 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3517 tcp_snd_una_update(tp, ack);
3522 static bool __tcp_oow_rate_limited(struct net *net, int mib_idx,
3523 u32 *last_oow_ack_time)
3525 if (*last_oow_ack_time) {
3526 s32 elapsed = (s32)(tcp_jiffies32 - *last_oow_ack_time);
3528 if (0 <= elapsed && elapsed < net->ipv4.sysctl_tcp_invalid_ratelimit) {
3529 NET_INC_STATS(net, mib_idx);
3530 return true; /* rate-limited: don't send yet! */
3534 *last_oow_ack_time = tcp_jiffies32;
3536 return false; /* not rate-limited: go ahead, send dupack now! */
3539 /* Return true if we're currently rate-limiting out-of-window ACKs and
3540 * thus shouldn't send a dupack right now. We rate-limit dupacks in
3541 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3542 * attacks that send repeated SYNs or ACKs for the same connection. To
3543 * do this, we do not send a duplicate SYNACK or ACK if the remote
3544 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3546 bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb,
3547 int mib_idx, u32 *last_oow_ack_time)
3549 /* Data packets without SYNs are not likely part of an ACK loop. */
3550 if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) &&
3554 return __tcp_oow_rate_limited(net, mib_idx, last_oow_ack_time);
3557 /* RFC 5961 7 [ACK Throttling] */
3558 static void tcp_send_challenge_ack(struct sock *sk, const struct sk_buff *skb)
3560 /* unprotected vars, we dont care of overwrites */
3561 static u32 challenge_timestamp;
3562 static unsigned int challenge_count;
3563 struct tcp_sock *tp = tcp_sk(sk);
3564 struct net *net = sock_net(sk);
3567 /* First check our per-socket dupack rate limit. */
3568 if (__tcp_oow_rate_limited(net,
3569 LINUX_MIB_TCPACKSKIPPEDCHALLENGE,
3570 &tp->last_oow_ack_time))
3573 /* Then check host-wide RFC 5961 rate limit. */
3575 if (now != challenge_timestamp) {
3576 u32 ack_limit = net->ipv4.sysctl_tcp_challenge_ack_limit;
3577 u32 half = (ack_limit + 1) >> 1;
3579 challenge_timestamp = now;
3580 WRITE_ONCE(challenge_count, half + prandom_u32_max(ack_limit));
3582 count = READ_ONCE(challenge_count);
3584 WRITE_ONCE(challenge_count, count - 1);
3585 NET_INC_STATS(net, LINUX_MIB_TCPCHALLENGEACK);
3590 static void tcp_store_ts_recent(struct tcp_sock *tp)
3592 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3593 tp->rx_opt.ts_recent_stamp = ktime_get_seconds();
3596 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3598 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3599 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3600 * extra check below makes sure this can only happen
3601 * for pure ACK frames. -DaveM
3603 * Not only, also it occurs for expired timestamps.
3606 if (tcp_paws_check(&tp->rx_opt, 0))
3607 tcp_store_ts_recent(tp);
3611 /* This routine deals with acks during a TLP episode and ends an episode by
3612 * resetting tlp_high_seq. Ref: TLP algorithm in draft-ietf-tcpm-rack
3614 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3616 struct tcp_sock *tp = tcp_sk(sk);
3618 if (before(ack, tp->tlp_high_seq))
3621 if (!tp->tlp_retrans) {
3622 /* TLP of new data has been acknowledged */
3623 tp->tlp_high_seq = 0;
3624 } else if (flag & FLAG_DSACKING_ACK) {
3625 /* This DSACK means original and TLP probe arrived; no loss */
3626 tp->tlp_high_seq = 0;
3627 } else if (after(ack, tp->tlp_high_seq)) {
3628 /* ACK advances: there was a loss, so reduce cwnd. Reset
3629 * tlp_high_seq in tcp_init_cwnd_reduction()
3631 tcp_init_cwnd_reduction(sk);
3632 tcp_set_ca_state(sk, TCP_CA_CWR);
3633 tcp_end_cwnd_reduction(sk);
3634 tcp_try_keep_open(sk);
3635 NET_INC_STATS(sock_net(sk),
3636 LINUX_MIB_TCPLOSSPROBERECOVERY);
3637 } else if (!(flag & (FLAG_SND_UNA_ADVANCED |
3638 FLAG_NOT_DUP | FLAG_DATA_SACKED))) {
3639 /* Pure dupack: original and TLP probe arrived; no loss */
3640 tp->tlp_high_seq = 0;
3644 static inline void tcp_in_ack_event(struct sock *sk, u32 flags)
3646 const struct inet_connection_sock *icsk = inet_csk(sk);
3648 if (icsk->icsk_ca_ops->in_ack_event)
3649 icsk->icsk_ca_ops->in_ack_event(sk, flags);
3652 /* Congestion control has updated the cwnd already. So if we're in
3653 * loss recovery then now we do any new sends (for FRTO) or
3654 * retransmits (for CA_Loss or CA_recovery) that make sense.
3656 static void tcp_xmit_recovery(struct sock *sk, int rexmit)
3658 struct tcp_sock *tp = tcp_sk(sk);
3660 if (rexmit == REXMIT_NONE || sk->sk_state == TCP_SYN_SENT)
3663 if (unlikely(rexmit == REXMIT_NEW)) {
3664 __tcp_push_pending_frames(sk, tcp_current_mss(sk),
3666 if (after(tp->snd_nxt, tp->high_seq))
3670 tcp_xmit_retransmit_queue(sk);
3673 /* Returns the number of packets newly acked or sacked by the current ACK */
3674 static u32 tcp_newly_delivered(struct sock *sk, u32 prior_delivered, int flag)
3676 const struct net *net = sock_net(sk);
3677 struct tcp_sock *tp = tcp_sk(sk);
3680 delivered = tp->delivered - prior_delivered;
3681 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVERED, delivered);
3682 if (flag & FLAG_ECE)
3683 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVEREDCE, delivered);
3688 /* This routine deals with incoming acks, but not outgoing ones. */
3689 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3691 struct inet_connection_sock *icsk = inet_csk(sk);
3692 struct tcp_sock *tp = tcp_sk(sk);
3693 struct tcp_sacktag_state sack_state;
3694 struct rate_sample rs = { .prior_delivered = 0 };
3695 u32 prior_snd_una = tp->snd_una;
3696 bool is_sack_reneg = tp->is_sack_reneg;
3697 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3698 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3700 int prior_packets = tp->packets_out;
3701 u32 delivered = tp->delivered;
3702 u32 lost = tp->lost;
3703 int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */
3706 sack_state.first_sackt = 0;
3707 sack_state.rate = &rs;
3708 sack_state.sack_delivered = 0;
3710 /* We very likely will need to access rtx queue. */
3711 prefetch(sk->tcp_rtx_queue.rb_node);
3713 /* If the ack is older than previous acks
3714 * then we can probably ignore it.
3716 if (before(ack, prior_snd_una)) {
3717 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3718 if (before(ack, prior_snd_una - tp->max_window)) {
3719 if (!(flag & FLAG_NO_CHALLENGE_ACK))
3720 tcp_send_challenge_ack(sk, skb);
3726 /* If the ack includes data we haven't sent yet, discard
3727 * this segment (RFC793 Section 3.9).
3729 if (after(ack, tp->snd_nxt))
3732 if (after(ack, prior_snd_una)) {
3733 flag |= FLAG_SND_UNA_ADVANCED;
3734 icsk->icsk_retransmits = 0;
3736 #if IS_ENABLED(CONFIG_TLS_DEVICE)
3737 if (static_branch_unlikely(&clean_acked_data_enabled.key))
3738 if (icsk->icsk_clean_acked)
3739 icsk->icsk_clean_acked(sk, ack);
3743 prior_fack = tcp_is_sack(tp) ? tcp_highest_sack_seq(tp) : tp->snd_una;
3744 rs.prior_in_flight = tcp_packets_in_flight(tp);
3746 /* ts_recent update must be made after we are sure that the packet
3749 if (flag & FLAG_UPDATE_TS_RECENT)
3750 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3752 if ((flag & (FLAG_SLOWPATH | FLAG_SND_UNA_ADVANCED)) ==
3753 FLAG_SND_UNA_ADVANCED) {
3754 /* Window is constant, pure forward advance.
3755 * No more checks are required.
3756 * Note, we use the fact that SND.UNA>=SND.WL2.
3758 tcp_update_wl(tp, ack_seq);
3759 tcp_snd_una_update(tp, ack);
3760 flag |= FLAG_WIN_UPDATE;
3762 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
3764 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPACKS);
3766 u32 ack_ev_flags = CA_ACK_SLOWPATH;
3768 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3771 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3773 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3775 if (TCP_SKB_CB(skb)->sacked)
3776 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3779 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) {
3781 ack_ev_flags |= CA_ACK_ECE;
3784 if (sack_state.sack_delivered)
3785 tcp_count_delivered(tp, sack_state.sack_delivered,
3788 if (flag & FLAG_WIN_UPDATE)
3789 ack_ev_flags |= CA_ACK_WIN_UPDATE;
3791 tcp_in_ack_event(sk, ack_ev_flags);
3794 /* This is a deviation from RFC3168 since it states that:
3795 * "When the TCP data sender is ready to set the CWR bit after reducing
3796 * the congestion window, it SHOULD set the CWR bit only on the first
3797 * new data packet that it transmits."
3798 * We accept CWR on pure ACKs to be more robust
3799 * with widely-deployed TCP implementations that do this.
3801 tcp_ecn_accept_cwr(sk, skb);
3803 /* We passed data and got it acked, remove any soft error
3804 * log. Something worked...
3806 sk->sk_err_soft = 0;
3807 icsk->icsk_probes_out = 0;
3808 tp->rcv_tstamp = tcp_jiffies32;
3812 /* See if we can take anything off of the retransmit queue. */
3813 flag |= tcp_clean_rtx_queue(sk, skb, prior_fack, prior_snd_una,
3814 &sack_state, flag & FLAG_ECE);
3816 tcp_rack_update_reo_wnd(sk, &rs);
3818 if (tp->tlp_high_seq)
3819 tcp_process_tlp_ack(sk, ack, flag);
3821 if (tcp_ack_is_dubious(sk, flag)) {
3822 if (!(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP))) {
3824 /* Consider if pure acks were aggregated in tcp_add_backlog() */
3825 if (!(flag & FLAG_DATA))
3826 num_dupack = max_t(u16, 1, skb_shinfo(skb)->gso_segs);
3828 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
3832 /* If needed, reset TLP/RTO timer when RACK doesn't set. */
3833 if (flag & FLAG_SET_XMIT_TIMER)
3834 tcp_set_xmit_timer(sk);
3836 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
3839 delivered = tcp_newly_delivered(sk, delivered, flag);
3840 lost = tp->lost - lost; /* freshly marked lost */
3841 rs.is_ack_delayed = !!(flag & FLAG_ACK_MAYBE_DELAYED);
3842 tcp_rate_gen(sk, delivered, lost, is_sack_reneg, sack_state.rate);
3843 tcp_cong_control(sk, ack, delivered, flag, sack_state.rate);
3844 tcp_xmit_recovery(sk, rexmit);
3848 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3849 if (flag & FLAG_DSACKING_ACK) {
3850 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
3852 tcp_newly_delivered(sk, delivered, flag);
3854 /* If this ack opens up a zero window, clear backoff. It was
3855 * being used to time the probes, and is probably far higher than
3856 * it needs to be for normal retransmission.
3860 if (tp->tlp_high_seq)
3861 tcp_process_tlp_ack(sk, ack, flag);
3865 /* If data was SACKed, tag it and see if we should send more data.
3866 * If data was DSACKed, see if we can undo a cwnd reduction.
3868 if (TCP_SKB_CB(skb)->sacked) {
3869 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3871 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
3873 tcp_newly_delivered(sk, delivered, flag);
3874 tcp_xmit_recovery(sk, rexmit);
3880 static void tcp_parse_fastopen_option(int len, const unsigned char *cookie,
3881 bool syn, struct tcp_fastopen_cookie *foc,
3884 /* Valid only in SYN or SYN-ACK with an even length. */
3885 if (!foc || !syn || len < 0 || (len & 1))
3888 if (len >= TCP_FASTOPEN_COOKIE_MIN &&
3889 len <= TCP_FASTOPEN_COOKIE_MAX)
3890 memcpy(foc->val, cookie, len);
3897 static bool smc_parse_options(const struct tcphdr *th,
3898 struct tcp_options_received *opt_rx,
3899 const unsigned char *ptr,
3902 #if IS_ENABLED(CONFIG_SMC)
3903 if (static_branch_unlikely(&tcp_have_smc)) {
3904 if (th->syn && !(opsize & 1) &&
3905 opsize >= TCPOLEN_EXP_SMC_BASE &&
3906 get_unaligned_be32(ptr) == TCPOPT_SMC_MAGIC) {
3915 /* Try to parse the MSS option from the TCP header. Return 0 on failure, clamped
3918 static u16 tcp_parse_mss_option(const struct tcphdr *th, u16 user_mss)
3920 const unsigned char *ptr = (const unsigned char *)(th + 1);
3921 int length = (th->doff * 4) - sizeof(struct tcphdr);
3924 while (length > 0) {
3925 int opcode = *ptr++;
3931 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3938 if (opsize < 2) /* "silly options" */
3940 if (opsize > length)
3941 return mss; /* fail on partial options */
3942 if (opcode == TCPOPT_MSS && opsize == TCPOLEN_MSS) {
3943 u16 in_mss = get_unaligned_be16(ptr);
3946 if (user_mss && user_mss < in_mss)
3958 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3959 * But, this can also be called on packets in the established flow when
3960 * the fast version below fails.
3962 void tcp_parse_options(const struct net *net,
3963 const struct sk_buff *skb,
3964 struct tcp_options_received *opt_rx, int estab,
3965 struct tcp_fastopen_cookie *foc)
3967 const unsigned char *ptr;
3968 const struct tcphdr *th = tcp_hdr(skb);
3969 int length = (th->doff * 4) - sizeof(struct tcphdr);
3971 ptr = (const unsigned char *)(th + 1);
3972 opt_rx->saw_tstamp = 0;
3973 opt_rx->saw_unknown = 0;
3975 while (length > 0) {
3976 int opcode = *ptr++;
3982 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3989 if (opsize < 2) /* "silly options" */
3991 if (opsize > length)
3992 return; /* don't parse partial options */
3995 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3996 u16 in_mss = get_unaligned_be16(ptr);
3998 if (opt_rx->user_mss &&
3999 opt_rx->user_mss < in_mss)
4000 in_mss = opt_rx->user_mss;
4001 opt_rx->mss_clamp = in_mss;
4006 if (opsize == TCPOLEN_WINDOW && th->syn &&
4007 !estab && net->ipv4.sysctl_tcp_window_scaling) {
4008 __u8 snd_wscale = *(__u8 *)ptr;
4009 opt_rx->wscale_ok = 1;
4010 if (snd_wscale > TCP_MAX_WSCALE) {
4011 net_info_ratelimited("%s: Illegal window scaling value %d > %u received\n",
4015 snd_wscale = TCP_MAX_WSCALE;
4017 opt_rx->snd_wscale = snd_wscale;
4020 case TCPOPT_TIMESTAMP:
4021 if ((opsize == TCPOLEN_TIMESTAMP) &&
4022 ((estab && opt_rx->tstamp_ok) ||
4023 (!estab && net->ipv4.sysctl_tcp_timestamps))) {
4024 opt_rx->saw_tstamp = 1;
4025 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
4026 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
4029 case TCPOPT_SACK_PERM:
4030 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
4031 !estab && net->ipv4.sysctl_tcp_sack) {
4032 opt_rx->sack_ok = TCP_SACK_SEEN;
4033 tcp_sack_reset(opt_rx);
4038 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
4039 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
4041 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
4044 #ifdef CONFIG_TCP_MD5SIG
4047 * The MD5 Hash has already been
4048 * checked (see tcp_v{4,6}_do_rcv()).
4052 case TCPOPT_FASTOPEN:
4053 tcp_parse_fastopen_option(
4054 opsize - TCPOLEN_FASTOPEN_BASE,
4055 ptr, th->syn, foc, false);
4059 /* Fast Open option shares code 254 using a
4060 * 16 bits magic number.
4062 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE &&
4063 get_unaligned_be16(ptr) ==
4064 TCPOPT_FASTOPEN_MAGIC) {
4065 tcp_parse_fastopen_option(opsize -
4066 TCPOLEN_EXP_FASTOPEN_BASE,
4067 ptr + 2, th->syn, foc, true);
4071 if (smc_parse_options(th, opt_rx, ptr, opsize))
4074 opt_rx->saw_unknown = 1;
4078 opt_rx->saw_unknown = 1;
4085 EXPORT_SYMBOL(tcp_parse_options);
4087 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
4089 const __be32 *ptr = (const __be32 *)(th + 1);
4091 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
4092 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
4093 tp->rx_opt.saw_tstamp = 1;
4095 tp->rx_opt.rcv_tsval = ntohl(*ptr);
4098 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
4100 tp->rx_opt.rcv_tsecr = 0;
4106 /* Fast parse options. This hopes to only see timestamps.
4107 * If it is wrong it falls back on tcp_parse_options().
4109 static bool tcp_fast_parse_options(const struct net *net,
4110 const struct sk_buff *skb,
4111 const struct tcphdr *th, struct tcp_sock *tp)
4113 /* In the spirit of fast parsing, compare doff directly to constant
4114 * values. Because equality is used, short doff can be ignored here.
4116 if (th->doff == (sizeof(*th) / 4)) {
4117 tp->rx_opt.saw_tstamp = 0;
4119 } else if (tp->rx_opt.tstamp_ok &&
4120 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
4121 if (tcp_parse_aligned_timestamp(tp, th))
4125 tcp_parse_options(net, skb, &tp->rx_opt, 1, NULL);
4126 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
4127 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
4132 #ifdef CONFIG_TCP_MD5SIG
4134 * Parse MD5 Signature option
4136 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
4138 int length = (th->doff << 2) - sizeof(*th);
4139 const u8 *ptr = (const u8 *)(th + 1);
4141 /* If not enough data remaining, we can short cut */
4142 while (length >= TCPOLEN_MD5SIG) {
4143 int opcode = *ptr++;
4154 if (opsize < 2 || opsize > length)
4156 if (opcode == TCPOPT_MD5SIG)
4157 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
4164 EXPORT_SYMBOL(tcp_parse_md5sig_option);
4167 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
4169 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
4170 * it can pass through stack. So, the following predicate verifies that
4171 * this segment is not used for anything but congestion avoidance or
4172 * fast retransmit. Moreover, we even are able to eliminate most of such
4173 * second order effects, if we apply some small "replay" window (~RTO)
4174 * to timestamp space.
4176 * All these measures still do not guarantee that we reject wrapped ACKs
4177 * on networks with high bandwidth, when sequence space is recycled fastly,
4178 * but it guarantees that such events will be very rare and do not affect
4179 * connection seriously. This doesn't look nice, but alas, PAWS is really
4182 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
4183 * states that events when retransmit arrives after original data are rare.
4184 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
4185 * the biggest problem on large power networks even with minor reordering.
4186 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
4187 * up to bandwidth of 18Gigabit/sec. 8) ]
4190 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
4192 const struct tcp_sock *tp = tcp_sk(sk);
4193 const struct tcphdr *th = tcp_hdr(skb);
4194 u32 seq = TCP_SKB_CB(skb)->seq;
4195 u32 ack = TCP_SKB_CB(skb)->ack_seq;
4197 return (/* 1. Pure ACK with correct sequence number. */
4198 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
4200 /* 2. ... and duplicate ACK. */
4201 ack == tp->snd_una &&
4203 /* 3. ... and does not update window. */
4204 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
4206 /* 4. ... and sits in replay window. */
4207 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
4210 static inline bool tcp_paws_discard(const struct sock *sk,
4211 const struct sk_buff *skb)
4213 const struct tcp_sock *tp = tcp_sk(sk);
4215 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
4216 !tcp_disordered_ack(sk, skb);
4219 /* Check segment sequence number for validity.
4221 * Segment controls are considered valid, if the segment
4222 * fits to the window after truncation to the window. Acceptability
4223 * of data (and SYN, FIN, of course) is checked separately.
4224 * See tcp_data_queue(), for example.
4226 * Also, controls (RST is main one) are accepted using RCV.WUP instead
4227 * of RCV.NXT. Peer still did not advance his SND.UNA when we
4228 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4229 * (borrowed from freebsd)
4232 static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
4234 return !before(end_seq, tp->rcv_wup) &&
4235 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
4238 /* When we get a reset we do this. */
4239 void tcp_reset(struct sock *sk, struct sk_buff *skb)
4241 trace_tcp_receive_reset(sk);
4243 if (sk_is_mptcp(sk))
4244 mptcp_incoming_options(sk, skb);
4246 /* We want the right error as BSD sees it (and indeed as we do). */
4247 switch (sk->sk_state) {
4249 sk->sk_err = ECONNREFUSED;
4251 case TCP_CLOSE_WAIT:
4257 sk->sk_err = ECONNRESET;
4259 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4262 tcp_write_queue_purge(sk);
4265 if (!sock_flag(sk, SOCK_DEAD))
4266 sk->sk_error_report(sk);
4270 * Process the FIN bit. This now behaves as it is supposed to work
4271 * and the FIN takes effect when it is validly part of sequence
4272 * space. Not before when we get holes.
4274 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4275 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4278 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4279 * close and we go into CLOSING (and later onto TIME-WAIT)
4281 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4283 void tcp_fin(struct sock *sk)
4285 struct tcp_sock *tp = tcp_sk(sk);
4287 inet_csk_schedule_ack(sk);
4289 sk->sk_shutdown |= RCV_SHUTDOWN;
4290 sock_set_flag(sk, SOCK_DONE);
4292 switch (sk->sk_state) {
4294 case TCP_ESTABLISHED:
4295 /* Move to CLOSE_WAIT */
4296 tcp_set_state(sk, TCP_CLOSE_WAIT);
4297 inet_csk_enter_pingpong_mode(sk);
4300 case TCP_CLOSE_WAIT:
4302 /* Received a retransmission of the FIN, do
4307 /* RFC793: Remain in the LAST-ACK state. */
4311 /* This case occurs when a simultaneous close
4312 * happens, we must ack the received FIN and
4313 * enter the CLOSING state.
4316 tcp_set_state(sk, TCP_CLOSING);
4319 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4321 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4324 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4325 * cases we should never reach this piece of code.
4327 pr_err("%s: Impossible, sk->sk_state=%d\n",
4328 __func__, sk->sk_state);
4332 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4333 * Probably, we should reset in this case. For now drop them.
4335 skb_rbtree_purge(&tp->out_of_order_queue);
4336 if (tcp_is_sack(tp))
4337 tcp_sack_reset(&tp->rx_opt);
4340 if (!sock_flag(sk, SOCK_DEAD)) {
4341 sk->sk_state_change(sk);
4343 /* Do not send POLL_HUP for half duplex close. */
4344 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4345 sk->sk_state == TCP_CLOSE)
4346 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4348 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4352 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4355 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4356 if (before(seq, sp->start_seq))
4357 sp->start_seq = seq;
4358 if (after(end_seq, sp->end_seq))
4359 sp->end_seq = end_seq;
4365 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4367 struct tcp_sock *tp = tcp_sk(sk);
4369 if (tcp_is_sack(tp) && sock_net(sk)->ipv4.sysctl_tcp_dsack) {
4372 if (before(seq, tp->rcv_nxt))
4373 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4375 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4377 NET_INC_STATS(sock_net(sk), mib_idx);
4379 tp->rx_opt.dsack = 1;
4380 tp->duplicate_sack[0].start_seq = seq;
4381 tp->duplicate_sack[0].end_seq = end_seq;
4385 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4387 struct tcp_sock *tp = tcp_sk(sk);
4389 if (!tp->rx_opt.dsack)
4390 tcp_dsack_set(sk, seq, end_seq);
4392 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4395 static void tcp_rcv_spurious_retrans(struct sock *sk, const struct sk_buff *skb)
4397 /* When the ACK path fails or drops most ACKs, the sender would
4398 * timeout and spuriously retransmit the same segment repeatedly.
4399 * The receiver remembers and reflects via DSACKs. Leverage the
4400 * DSACK state and change the txhash to re-route speculatively.
4402 if (TCP_SKB_CB(skb)->seq == tcp_sk(sk)->duplicate_sack[0].start_seq &&
4403 sk_rethink_txhash(sk))
4404 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDUPLICATEDATAREHASH);
4407 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4409 struct tcp_sock *tp = tcp_sk(sk);
4411 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4412 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4413 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4414 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
4416 if (tcp_is_sack(tp) && sock_net(sk)->ipv4.sysctl_tcp_dsack) {
4417 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4419 tcp_rcv_spurious_retrans(sk, skb);
4420 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4421 end_seq = tp->rcv_nxt;
4422 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4429 /* These routines update the SACK block as out-of-order packets arrive or
4430 * in-order packets close up the sequence space.
4432 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4435 struct tcp_sack_block *sp = &tp->selective_acks[0];
4436 struct tcp_sack_block *swalk = sp + 1;
4438 /* See if the recent change to the first SACK eats into
4439 * or hits the sequence space of other SACK blocks, if so coalesce.
4441 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4442 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4445 /* Zap SWALK, by moving every further SACK up by one slot.
4446 * Decrease num_sacks.
4448 tp->rx_opt.num_sacks--;
4449 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4458 static void tcp_sack_compress_send_ack(struct sock *sk)
4460 struct tcp_sock *tp = tcp_sk(sk);
4462 if (!tp->compressed_ack)
4465 if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1)
4468 /* Since we have to send one ack finally,
4469 * substract one from tp->compressed_ack to keep
4470 * LINUX_MIB_TCPACKCOMPRESSED accurate.
4472 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED,
4473 tp->compressed_ack - 1);
4475 tp->compressed_ack = 0;
4479 /* Reasonable amount of sack blocks included in TCP SACK option
4480 * The max is 4, but this becomes 3 if TCP timestamps are there.
4481 * Given that SACK packets might be lost, be conservative and use 2.
4483 #define TCP_SACK_BLOCKS_EXPECTED 2
4485 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4487 struct tcp_sock *tp = tcp_sk(sk);
4488 struct tcp_sack_block *sp = &tp->selective_acks[0];
4489 int cur_sacks = tp->rx_opt.num_sacks;
4495 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4496 if (tcp_sack_extend(sp, seq, end_seq)) {
4497 if (this_sack >= TCP_SACK_BLOCKS_EXPECTED)
4498 tcp_sack_compress_send_ack(sk);
4499 /* Rotate this_sack to the first one. */
4500 for (; this_sack > 0; this_sack--, sp--)
4501 swap(*sp, *(sp - 1));
4503 tcp_sack_maybe_coalesce(tp);
4508 if (this_sack >= TCP_SACK_BLOCKS_EXPECTED)
4509 tcp_sack_compress_send_ack(sk);
4511 /* Could not find an adjacent existing SACK, build a new one,
4512 * put it at the front, and shift everyone else down. We
4513 * always know there is at least one SACK present already here.
4515 * If the sack array is full, forget about the last one.
4517 if (this_sack >= TCP_NUM_SACKS) {
4519 tp->rx_opt.num_sacks--;
4522 for (; this_sack > 0; this_sack--, sp--)
4526 /* Build the new head SACK, and we're done. */
4527 sp->start_seq = seq;
4528 sp->end_seq = end_seq;
4529 tp->rx_opt.num_sacks++;
4532 /* RCV.NXT advances, some SACKs should be eaten. */
4534 static void tcp_sack_remove(struct tcp_sock *tp)
4536 struct tcp_sack_block *sp = &tp->selective_acks[0];
4537 int num_sacks = tp->rx_opt.num_sacks;
4540 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4541 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4542 tp->rx_opt.num_sacks = 0;
4546 for (this_sack = 0; this_sack < num_sacks;) {
4547 /* Check if the start of the sack is covered by RCV.NXT. */
4548 if (!before(tp->rcv_nxt, sp->start_seq)) {
4551 /* RCV.NXT must cover all the block! */
4552 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4554 /* Zap this SACK, by moving forward any other SACKS. */
4555 for (i = this_sack+1; i < num_sacks; i++)
4556 tp->selective_acks[i-1] = tp->selective_acks[i];
4563 tp->rx_opt.num_sacks = num_sacks;
4567 * tcp_try_coalesce - try to merge skb to prior one
4570 * @from: buffer to add in queue
4571 * @fragstolen: pointer to boolean
4573 * Before queueing skb @from after @to, try to merge them
4574 * to reduce overall memory use and queue lengths, if cost is small.
4575 * Packets in ofo or receive queues can stay a long time.
4576 * Better try to coalesce them right now to avoid future collapses.
4577 * Returns true if caller should free @from instead of queueing it
4579 static bool tcp_try_coalesce(struct sock *sk,
4581 struct sk_buff *from,
4586 *fragstolen = false;
4588 /* Its possible this segment overlaps with prior segment in queue */
4589 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4592 if (!mptcp_skb_can_collapse(to, from))
4595 #ifdef CONFIG_TLS_DEVICE
4596 if (from->decrypted != to->decrypted)
4600 if (!skb_try_coalesce(to, from, fragstolen, &delta))
4603 atomic_add(delta, &sk->sk_rmem_alloc);
4604 sk_mem_charge(sk, delta);
4605 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4606 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4607 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4608 TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
4610 if (TCP_SKB_CB(from)->has_rxtstamp) {
4611 TCP_SKB_CB(to)->has_rxtstamp = true;
4612 to->tstamp = from->tstamp;
4613 skb_hwtstamps(to)->hwtstamp = skb_hwtstamps(from)->hwtstamp;
4619 static bool tcp_ooo_try_coalesce(struct sock *sk,
4621 struct sk_buff *from,
4624 bool res = tcp_try_coalesce(sk, to, from, fragstolen);
4626 /* In case tcp_drop() is called later, update to->gso_segs */
4628 u32 gso_segs = max_t(u16, 1, skb_shinfo(to)->gso_segs) +
4629 max_t(u16, 1, skb_shinfo(from)->gso_segs);
4631 skb_shinfo(to)->gso_segs = min_t(u32, gso_segs, 0xFFFF);
4636 static void tcp_drop(struct sock *sk, struct sk_buff *skb)
4638 sk_drops_add(sk, skb);
4642 /* This one checks to see if we can put data from the
4643 * out_of_order queue into the receive_queue.
4645 static void tcp_ofo_queue(struct sock *sk)
4647 struct tcp_sock *tp = tcp_sk(sk);
4648 __u32 dsack_high = tp->rcv_nxt;
4649 bool fin, fragstolen, eaten;
4650 struct sk_buff *skb, *tail;
4653 p = rb_first(&tp->out_of_order_queue);
4656 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4659 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4660 __u32 dsack = dsack_high;
4661 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4662 dsack_high = TCP_SKB_CB(skb)->end_seq;
4663 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4666 rb_erase(&skb->rbnode, &tp->out_of_order_queue);
4668 if (unlikely(!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))) {
4673 tail = skb_peek_tail(&sk->sk_receive_queue);
4674 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen);
4675 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4676 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
4678 __skb_queue_tail(&sk->sk_receive_queue, skb);
4680 kfree_skb_partial(skb, fragstolen);
4682 if (unlikely(fin)) {
4684 /* tcp_fin() purges tp->out_of_order_queue,
4685 * so we must end this loop right now.
4692 static bool tcp_prune_ofo_queue(struct sock *sk);
4693 static int tcp_prune_queue(struct sock *sk);
4695 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4698 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4699 !sk_rmem_schedule(sk, skb, size)) {
4701 if (tcp_prune_queue(sk) < 0)
4704 while (!sk_rmem_schedule(sk, skb, size)) {
4705 if (!tcp_prune_ofo_queue(sk))
4712 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4714 struct tcp_sock *tp = tcp_sk(sk);
4715 struct rb_node **p, *parent;
4716 struct sk_buff *skb1;
4720 tcp_ecn_check_ce(sk, skb);
4722 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4723 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFODROP);
4724 sk->sk_data_ready(sk);
4729 /* Disable header prediction. */
4731 inet_csk_schedule_ack(sk);
4733 tp->rcv_ooopack += max_t(u16, 1, skb_shinfo(skb)->gso_segs);
4734 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4735 seq = TCP_SKB_CB(skb)->seq;
4736 end_seq = TCP_SKB_CB(skb)->end_seq;
4738 p = &tp->out_of_order_queue.rb_node;
4739 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4740 /* Initial out of order segment, build 1 SACK. */
4741 if (tcp_is_sack(tp)) {
4742 tp->rx_opt.num_sacks = 1;
4743 tp->selective_acks[0].start_seq = seq;
4744 tp->selective_acks[0].end_seq = end_seq;
4746 rb_link_node(&skb->rbnode, NULL, p);
4747 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4748 tp->ooo_last_skb = skb;
4752 /* In the typical case, we are adding an skb to the end of the list.
4753 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
4755 if (tcp_ooo_try_coalesce(sk, tp->ooo_last_skb,
4756 skb, &fragstolen)) {
4758 /* For non sack flows, do not grow window to force DUPACK
4759 * and trigger fast retransmit.
4761 if (tcp_is_sack(tp))
4762 tcp_grow_window(sk, skb);
4763 kfree_skb_partial(skb, fragstolen);
4767 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
4768 if (!before(seq, TCP_SKB_CB(tp->ooo_last_skb)->end_seq)) {
4769 parent = &tp->ooo_last_skb->rbnode;
4770 p = &parent->rb_right;
4774 /* Find place to insert this segment. Handle overlaps on the way. */
4778 skb1 = rb_to_skb(parent);
4779 if (before(seq, TCP_SKB_CB(skb1)->seq)) {
4780 p = &parent->rb_left;
4783 if (before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4784 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4785 /* All the bits are present. Drop. */
4786 NET_INC_STATS(sock_net(sk),
4787 LINUX_MIB_TCPOFOMERGE);
4790 tcp_dsack_set(sk, seq, end_seq);
4793 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4794 /* Partial overlap. */
4795 tcp_dsack_set(sk, seq, TCP_SKB_CB(skb1)->end_seq);
4797 /* skb's seq == skb1's seq and skb covers skb1.
4798 * Replace skb1 with skb.
4800 rb_replace_node(&skb1->rbnode, &skb->rbnode,
4801 &tp->out_of_order_queue);
4802 tcp_dsack_extend(sk,
4803 TCP_SKB_CB(skb1)->seq,
4804 TCP_SKB_CB(skb1)->end_seq);
4805 NET_INC_STATS(sock_net(sk),
4806 LINUX_MIB_TCPOFOMERGE);
4810 } else if (tcp_ooo_try_coalesce(sk, skb1,
4811 skb, &fragstolen)) {
4814 p = &parent->rb_right;
4817 /* Insert segment into RB tree. */
4818 rb_link_node(&skb->rbnode, parent, p);
4819 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4822 /* Remove other segments covered by skb. */
4823 while ((skb1 = skb_rb_next(skb)) != NULL) {
4824 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4826 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4827 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4831 rb_erase(&skb1->rbnode, &tp->out_of_order_queue);
4832 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4833 TCP_SKB_CB(skb1)->end_seq);
4834 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4837 /* If there is no skb after us, we are the last_skb ! */
4839 tp->ooo_last_skb = skb;
4842 if (tcp_is_sack(tp))
4843 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4846 /* For non sack flows, do not grow window to force DUPACK
4847 * and trigger fast retransmit.
4849 if (tcp_is_sack(tp))
4850 tcp_grow_window(sk, skb);
4852 skb_set_owner_r(skb, sk);
4856 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb,
4860 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4863 tcp_try_coalesce(sk, tail,
4864 skb, fragstolen)) ? 1 : 0;
4865 tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq);
4867 __skb_queue_tail(&sk->sk_receive_queue, skb);
4868 skb_set_owner_r(skb, sk);
4873 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4875 struct sk_buff *skb;
4883 if (size > PAGE_SIZE) {
4884 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS);
4886 data_len = npages << PAGE_SHIFT;
4887 size = data_len + (size & ~PAGE_MASK);
4889 skb = alloc_skb_with_frags(size - data_len, data_len,
4890 PAGE_ALLOC_COSTLY_ORDER,
4891 &err, sk->sk_allocation);
4895 skb_put(skb, size - data_len);
4896 skb->data_len = data_len;
4899 if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) {
4900 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP);
4904 err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size);
4908 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4909 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4910 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4912 if (tcp_queue_rcv(sk, skb, &fragstolen)) {
4913 WARN_ON_ONCE(fragstolen); /* should not happen */
4925 void tcp_data_ready(struct sock *sk)
4927 if (tcp_epollin_ready(sk, sk->sk_rcvlowat) || sock_flag(sk, SOCK_DONE))
4928 sk->sk_data_ready(sk);
4931 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4933 struct tcp_sock *tp = tcp_sk(sk);
4937 if (sk_is_mptcp(sk))
4938 mptcp_incoming_options(sk, skb);
4940 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
4945 __skb_pull(skb, tcp_hdr(skb)->doff * 4);
4947 tp->rx_opt.dsack = 0;
4949 /* Queue data for delivery to the user.
4950 * Packets in sequence go to the receive queue.
4951 * Out of sequence packets to the out_of_order_queue.
4953 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4954 if (tcp_receive_window(tp) == 0) {
4955 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP);
4959 /* Ok. In sequence. In window. */
4961 if (skb_queue_len(&sk->sk_receive_queue) == 0)
4962 sk_forced_mem_schedule(sk, skb->truesize);
4963 else if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) {
4964 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP);
4965 sk->sk_data_ready(sk);
4969 eaten = tcp_queue_rcv(sk, skb, &fragstolen);
4971 tcp_event_data_recv(sk, skb);
4972 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
4975 if (!RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4978 /* RFC5681. 4.2. SHOULD send immediate ACK, when
4979 * gap in queue is filled.
4981 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
4982 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW;
4985 if (tp->rx_opt.num_sacks)
4986 tcp_sack_remove(tp);
4988 tcp_fast_path_check(sk);
4991 kfree_skb_partial(skb, fragstolen);
4992 if (!sock_flag(sk, SOCK_DEAD))
4997 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4998 tcp_rcv_spurious_retrans(sk, skb);
4999 /* A retransmit, 2nd most common case. Force an immediate ack. */
5000 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
5001 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
5004 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
5005 inet_csk_schedule_ack(sk);
5011 /* Out of window. F.e. zero window probe. */
5012 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
5015 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
5016 /* Partial packet, seq < rcv_next < end_seq */
5017 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
5019 /* If window is closed, drop tail of packet. But after
5020 * remembering D-SACK for its head made in previous line.
5022 if (!tcp_receive_window(tp)) {
5023 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP);
5029 tcp_data_queue_ofo(sk, skb);
5032 static struct sk_buff *tcp_skb_next(struct sk_buff *skb, struct sk_buff_head *list)
5035 return !skb_queue_is_last(list, skb) ? skb->next : NULL;
5037 return skb_rb_next(skb);
5040 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
5041 struct sk_buff_head *list,
5042 struct rb_root *root)
5044 struct sk_buff *next = tcp_skb_next(skb, list);
5047 __skb_unlink(skb, list);
5049 rb_erase(&skb->rbnode, root);
5052 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
5057 /* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */
5058 void tcp_rbtree_insert(struct rb_root *root, struct sk_buff *skb)
5060 struct rb_node **p = &root->rb_node;
5061 struct rb_node *parent = NULL;
5062 struct sk_buff *skb1;
5066 skb1 = rb_to_skb(parent);
5067 if (before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb1)->seq))
5068 p = &parent->rb_left;
5070 p = &parent->rb_right;
5072 rb_link_node(&skb->rbnode, parent, p);
5073 rb_insert_color(&skb->rbnode, root);
5076 /* Collapse contiguous sequence of skbs head..tail with
5077 * sequence numbers start..end.
5079 * If tail is NULL, this means until the end of the queue.
5081 * Segments with FIN/SYN are not collapsed (only because this
5085 tcp_collapse(struct sock *sk, struct sk_buff_head *list, struct rb_root *root,
5086 struct sk_buff *head, struct sk_buff *tail, u32 start, u32 end)
5088 struct sk_buff *skb = head, *n;
5089 struct sk_buff_head tmp;
5092 /* First, check that queue is collapsible and find
5093 * the point where collapsing can be useful.
5096 for (end_of_skbs = true; skb != NULL && skb != tail; skb = n) {
5097 n = tcp_skb_next(skb, list);
5099 /* No new bits? It is possible on ofo queue. */
5100 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
5101 skb = tcp_collapse_one(sk, skb, list, root);
5107 /* The first skb to collapse is:
5109 * - bloated or contains data before "start" or
5110 * overlaps to the next one and mptcp allow collapsing.
5112 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
5113 (tcp_win_from_space(sk, skb->truesize) > skb->len ||
5114 before(TCP_SKB_CB(skb)->seq, start))) {
5115 end_of_skbs = false;
5119 if (n && n != tail && mptcp_skb_can_collapse(skb, n) &&
5120 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(n)->seq) {
5121 end_of_skbs = false;
5125 /* Decided to skip this, advance start seq. */
5126 start = TCP_SKB_CB(skb)->end_seq;
5129 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
5132 __skb_queue_head_init(&tmp);
5134 while (before(start, end)) {
5135 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
5136 struct sk_buff *nskb;
5138 nskb = alloc_skb(copy, GFP_ATOMIC);
5142 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
5143 #ifdef CONFIG_TLS_DEVICE
5144 nskb->decrypted = skb->decrypted;
5146 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
5148 __skb_queue_before(list, skb, nskb);
5150 __skb_queue_tail(&tmp, nskb); /* defer rbtree insertion */
5151 skb_set_owner_r(nskb, sk);
5152 mptcp_skb_ext_move(nskb, skb);
5154 /* Copy data, releasing collapsed skbs. */
5156 int offset = start - TCP_SKB_CB(skb)->seq;
5157 int size = TCP_SKB_CB(skb)->end_seq - start;
5161 size = min(copy, size);
5162 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
5164 TCP_SKB_CB(nskb)->end_seq += size;
5168 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
5169 skb = tcp_collapse_one(sk, skb, list, root);
5172 !mptcp_skb_can_collapse(nskb, skb) ||
5173 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
5175 #ifdef CONFIG_TLS_DEVICE
5176 if (skb->decrypted != nskb->decrypted)
5183 skb_queue_walk_safe(&tmp, skb, n)
5184 tcp_rbtree_insert(root, skb);
5187 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
5188 * and tcp_collapse() them until all the queue is collapsed.
5190 static void tcp_collapse_ofo_queue(struct sock *sk)
5192 struct tcp_sock *tp = tcp_sk(sk);
5193 u32 range_truesize, sum_tiny = 0;
5194 struct sk_buff *skb, *head;
5197 skb = skb_rb_first(&tp->out_of_order_queue);
5200 tp->ooo_last_skb = skb_rb_last(&tp->out_of_order_queue);
5203 start = TCP_SKB_CB(skb)->seq;
5204 end = TCP_SKB_CB(skb)->end_seq;
5205 range_truesize = skb->truesize;
5207 for (head = skb;;) {
5208 skb = skb_rb_next(skb);
5210 /* Range is terminated when we see a gap or when
5211 * we are at the queue end.
5214 after(TCP_SKB_CB(skb)->seq, end) ||
5215 before(TCP_SKB_CB(skb)->end_seq, start)) {
5216 /* Do not attempt collapsing tiny skbs */
5217 if (range_truesize != head->truesize ||
5218 end - start >= SKB_WITH_OVERHEAD(SK_MEM_QUANTUM)) {
5219 tcp_collapse(sk, NULL, &tp->out_of_order_queue,
5220 head, skb, start, end);
5222 sum_tiny += range_truesize;
5223 if (sum_tiny > sk->sk_rcvbuf >> 3)
5229 range_truesize += skb->truesize;
5230 if (unlikely(before(TCP_SKB_CB(skb)->seq, start)))
5231 start = TCP_SKB_CB(skb)->seq;
5232 if (after(TCP_SKB_CB(skb)->end_seq, end))
5233 end = TCP_SKB_CB(skb)->end_seq;
5238 * Clean the out-of-order queue to make room.
5239 * We drop high sequences packets to :
5240 * 1) Let a chance for holes to be filled.
5241 * 2) not add too big latencies if thousands of packets sit there.
5242 * (But if application shrinks SO_RCVBUF, we could still end up
5243 * freeing whole queue here)
5244 * 3) Drop at least 12.5 % of sk_rcvbuf to avoid malicious attacks.
5246 * Return true if queue has shrunk.
5248 static bool tcp_prune_ofo_queue(struct sock *sk)
5250 struct tcp_sock *tp = tcp_sk(sk);
5251 struct rb_node *node, *prev;
5254 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
5257 NET_INC_STATS(sock_net(sk), LINUX_MIB_OFOPRUNED);
5258 goal = sk->sk_rcvbuf >> 3;
5259 node = &tp->ooo_last_skb->rbnode;
5261 prev = rb_prev(node);
5262 rb_erase(node, &tp->out_of_order_queue);
5263 goal -= rb_to_skb(node)->truesize;
5264 tcp_drop(sk, rb_to_skb(node));
5265 if (!prev || goal <= 0) {
5267 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf &&
5268 !tcp_under_memory_pressure(sk))
5270 goal = sk->sk_rcvbuf >> 3;
5274 tp->ooo_last_skb = rb_to_skb(prev);
5276 /* Reset SACK state. A conforming SACK implementation will
5277 * do the same at a timeout based retransmit. When a connection
5278 * is in a sad state like this, we care only about integrity
5279 * of the connection not performance.
5281 if (tp->rx_opt.sack_ok)
5282 tcp_sack_reset(&tp->rx_opt);
5286 /* Reduce allocated memory if we can, trying to get
5287 * the socket within its memory limits again.
5289 * Return less than zero if we should start dropping frames
5290 * until the socket owning process reads some of the data
5291 * to stabilize the situation.
5293 static int tcp_prune_queue(struct sock *sk)
5295 struct tcp_sock *tp = tcp_sk(sk);
5297 NET_INC_STATS(sock_net(sk), LINUX_MIB_PRUNECALLED);
5299 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
5300 tcp_clamp_window(sk);
5301 else if (tcp_under_memory_pressure(sk))
5302 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
5304 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5307 tcp_collapse_ofo_queue(sk);
5308 if (!skb_queue_empty(&sk->sk_receive_queue))
5309 tcp_collapse(sk, &sk->sk_receive_queue, NULL,
5310 skb_peek(&sk->sk_receive_queue),
5312 tp->copied_seq, tp->rcv_nxt);
5315 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5318 /* Collapsing did not help, destructive actions follow.
5319 * This must not ever occur. */
5321 tcp_prune_ofo_queue(sk);
5323 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5326 /* If we are really being abused, tell the caller to silently
5327 * drop receive data on the floor. It will get retransmitted
5328 * and hopefully then we'll have sufficient space.
5330 NET_INC_STATS(sock_net(sk), LINUX_MIB_RCVPRUNED);
5332 /* Massive buffer overcommit. */
5337 static bool tcp_should_expand_sndbuf(const struct sock *sk)
5339 const struct tcp_sock *tp = tcp_sk(sk);
5341 /* If the user specified a specific send buffer setting, do
5344 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
5347 /* If we are under global TCP memory pressure, do not expand. */
5348 if (tcp_under_memory_pressure(sk))
5351 /* If we are under soft global TCP memory pressure, do not expand. */
5352 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
5355 /* If we filled the congestion window, do not expand. */
5356 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
5362 static void tcp_new_space(struct sock *sk)
5364 struct tcp_sock *tp = tcp_sk(sk);
5366 if (tcp_should_expand_sndbuf(sk)) {
5367 tcp_sndbuf_expand(sk);
5368 tp->snd_cwnd_stamp = tcp_jiffies32;
5371 sk->sk_write_space(sk);
5374 static void tcp_check_space(struct sock *sk)
5376 /* pairs with tcp_poll() */
5378 if (sk->sk_socket &&
5379 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
5381 if (!test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
5382 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
5386 static inline void tcp_data_snd_check(struct sock *sk)
5388 tcp_push_pending_frames(sk);
5389 tcp_check_space(sk);
5393 * Check if sending an ack is needed.
5395 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
5397 struct tcp_sock *tp = tcp_sk(sk);
5398 unsigned long rtt, delay;
5400 /* More than one full frame received... */
5401 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
5402 /* ... and right edge of window advances far enough.
5403 * (tcp_recvmsg() will send ACK otherwise).
5404 * If application uses SO_RCVLOWAT, we want send ack now if
5405 * we have not received enough bytes to satisfy the condition.
5407 (tp->rcv_nxt - tp->copied_seq < sk->sk_rcvlowat ||
5408 __tcp_select_window(sk) >= tp->rcv_wnd)) ||
5409 /* We ACK each frame or... */
5410 tcp_in_quickack_mode(sk) ||
5411 /* Protocol state mandates a one-time immediate ACK */
5412 inet_csk(sk)->icsk_ack.pending & ICSK_ACK_NOW) {
5418 if (!ofo_possible || RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
5419 tcp_send_delayed_ack(sk);
5423 if (!tcp_is_sack(tp) ||
5424 tp->compressed_ack >= sock_net(sk)->ipv4.sysctl_tcp_comp_sack_nr)
5427 if (tp->compressed_ack_rcv_nxt != tp->rcv_nxt) {
5428 tp->compressed_ack_rcv_nxt = tp->rcv_nxt;
5429 tp->dup_ack_counter = 0;
5431 if (tp->dup_ack_counter < TCP_FASTRETRANS_THRESH) {
5432 tp->dup_ack_counter++;
5435 tp->compressed_ack++;
5436 if (hrtimer_is_queued(&tp->compressed_ack_timer))
5439 /* compress ack timer : 5 % of rtt, but no more than tcp_comp_sack_delay_ns */
5441 rtt = tp->rcv_rtt_est.rtt_us;
5442 if (tp->srtt_us && tp->srtt_us < rtt)
5445 delay = min_t(unsigned long, sock_net(sk)->ipv4.sysctl_tcp_comp_sack_delay_ns,
5446 rtt * (NSEC_PER_USEC >> 3)/20);
5448 hrtimer_start_range_ns(&tp->compressed_ack_timer, ns_to_ktime(delay),
5449 sock_net(sk)->ipv4.sysctl_tcp_comp_sack_slack_ns,
5450 HRTIMER_MODE_REL_PINNED_SOFT);
5453 static inline void tcp_ack_snd_check(struct sock *sk)
5455 if (!inet_csk_ack_scheduled(sk)) {
5456 /* We sent a data segment already. */
5459 __tcp_ack_snd_check(sk, 1);
5463 * This routine is only called when we have urgent data
5464 * signaled. Its the 'slow' part of tcp_urg. It could be
5465 * moved inline now as tcp_urg is only called from one
5466 * place. We handle URGent data wrong. We have to - as
5467 * BSD still doesn't use the correction from RFC961.
5468 * For 1003.1g we should support a new option TCP_STDURG to permit
5469 * either form (or just set the sysctl tcp_stdurg).
5472 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
5474 struct tcp_sock *tp = tcp_sk(sk);
5475 u32 ptr = ntohs(th->urg_ptr);
5477 if (ptr && !sock_net(sk)->ipv4.sysctl_tcp_stdurg)
5479 ptr += ntohl(th->seq);
5481 /* Ignore urgent data that we've already seen and read. */
5482 if (after(tp->copied_seq, ptr))
5485 /* Do not replay urg ptr.
5487 * NOTE: interesting situation not covered by specs.
5488 * Misbehaving sender may send urg ptr, pointing to segment,
5489 * which we already have in ofo queue. We are not able to fetch
5490 * such data and will stay in TCP_URG_NOTYET until will be eaten
5491 * by recvmsg(). Seems, we are not obliged to handle such wicked
5492 * situations. But it is worth to think about possibility of some
5493 * DoSes using some hypothetical application level deadlock.
5495 if (before(ptr, tp->rcv_nxt))
5498 /* Do we already have a newer (or duplicate) urgent pointer? */
5499 if (tp->urg_data && !after(ptr, tp->urg_seq))
5502 /* Tell the world about our new urgent pointer. */
5505 /* We may be adding urgent data when the last byte read was
5506 * urgent. To do this requires some care. We cannot just ignore
5507 * tp->copied_seq since we would read the last urgent byte again
5508 * as data, nor can we alter copied_seq until this data arrives
5509 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5511 * NOTE. Double Dutch. Rendering to plain English: author of comment
5512 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5513 * and expect that both A and B disappear from stream. This is _wrong_.
5514 * Though this happens in BSD with high probability, this is occasional.
5515 * Any application relying on this is buggy. Note also, that fix "works"
5516 * only in this artificial test. Insert some normal data between A and B and we will
5517 * decline of BSD again. Verdict: it is better to remove to trap
5520 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5521 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5522 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5524 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5525 __skb_unlink(skb, &sk->sk_receive_queue);
5530 tp->urg_data = TCP_URG_NOTYET;
5531 WRITE_ONCE(tp->urg_seq, ptr);
5533 /* Disable header prediction. */
5537 /* This is the 'fast' part of urgent handling. */
5538 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
5540 struct tcp_sock *tp = tcp_sk(sk);
5542 /* Check if we get a new urgent pointer - normally not. */
5544 tcp_check_urg(sk, th);
5546 /* Do we wait for any urgent data? - normally not... */
5547 if (tp->urg_data == TCP_URG_NOTYET) {
5548 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5551 /* Is the urgent pointer pointing into this packet? */
5552 if (ptr < skb->len) {
5554 if (skb_copy_bits(skb, ptr, &tmp, 1))
5556 tp->urg_data = TCP_URG_VALID | tmp;
5557 if (!sock_flag(sk, SOCK_DEAD))
5558 sk->sk_data_ready(sk);
5563 /* Accept RST for rcv_nxt - 1 after a FIN.
5564 * When tcp connections are abruptly terminated from Mac OSX (via ^C), a
5565 * FIN is sent followed by a RST packet. The RST is sent with the same
5566 * sequence number as the FIN, and thus according to RFC 5961 a challenge
5567 * ACK should be sent. However, Mac OSX rate limits replies to challenge
5568 * ACKs on the closed socket. In addition middleboxes can drop either the
5569 * challenge ACK or a subsequent RST.
5571 static bool tcp_reset_check(const struct sock *sk, const struct sk_buff *skb)
5573 struct tcp_sock *tp = tcp_sk(sk);
5575 return unlikely(TCP_SKB_CB(skb)->seq == (tp->rcv_nxt - 1) &&
5576 (1 << sk->sk_state) & (TCPF_CLOSE_WAIT | TCPF_LAST_ACK |
5580 /* Does PAWS and seqno based validation of an incoming segment, flags will
5581 * play significant role here.
5583 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5584 const struct tcphdr *th, int syn_inerr)
5586 struct tcp_sock *tp = tcp_sk(sk);
5587 bool rst_seq_match = false;
5589 /* RFC1323: H1. Apply PAWS check first. */
5590 if (tcp_fast_parse_options(sock_net(sk), skb, th, tp) &&
5591 tp->rx_opt.saw_tstamp &&
5592 tcp_paws_discard(sk, skb)) {
5594 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5595 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5596 LINUX_MIB_TCPACKSKIPPEDPAWS,
5597 &tp->last_oow_ack_time))
5598 tcp_send_dupack(sk, skb);
5601 /* Reset is accepted even if it did not pass PAWS. */
5604 /* Step 1: check sequence number */
5605 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5606 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5607 * (RST) segments are validated by checking their SEQ-fields."
5608 * And page 69: "If an incoming segment is not acceptable,
5609 * an acknowledgment should be sent in reply (unless the RST
5610 * bit is set, if so drop the segment and return)".
5615 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5616 LINUX_MIB_TCPACKSKIPPEDSEQ,
5617 &tp->last_oow_ack_time))
5618 tcp_send_dupack(sk, skb);
5619 } else if (tcp_reset_check(sk, skb)) {
5625 /* Step 2: check RST bit */
5627 /* RFC 5961 3.2 (extend to match against (RCV.NXT - 1) after a
5628 * FIN and SACK too if available):
5629 * If seq num matches RCV.NXT or (RCV.NXT - 1) after a FIN, or
5630 * the right-most SACK block,
5632 * RESET the connection
5634 * Send a challenge ACK
5636 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt ||
5637 tcp_reset_check(sk, skb)) {
5638 rst_seq_match = true;
5639 } else if (tcp_is_sack(tp) && tp->rx_opt.num_sacks > 0) {
5640 struct tcp_sack_block *sp = &tp->selective_acks[0];
5641 int max_sack = sp[0].end_seq;
5644 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;
5646 max_sack = after(sp[this_sack].end_seq,
5648 sp[this_sack].end_seq : max_sack;
5651 if (TCP_SKB_CB(skb)->seq == max_sack)
5652 rst_seq_match = true;
5658 /* Disable TFO if RST is out-of-order
5659 * and no data has been received
5660 * for current active TFO socket
5662 if (tp->syn_fastopen && !tp->data_segs_in &&
5663 sk->sk_state == TCP_ESTABLISHED)
5664 tcp_fastopen_active_disable(sk);
5665 tcp_send_challenge_ack(sk, skb);
5670 /* step 3: check security and precedence [ignored] */
5672 /* step 4: Check for a SYN
5673 * RFC 5961 4.2 : Send a challenge ack
5678 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5679 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5680 tcp_send_challenge_ack(sk, skb);
5684 bpf_skops_parse_hdr(sk, skb);
5694 * TCP receive function for the ESTABLISHED state.
5696 * It is split into a fast path and a slow path. The fast path is
5698 * - A zero window was announced from us - zero window probing
5699 * is only handled properly in the slow path.
5700 * - Out of order segments arrived.
5701 * - Urgent data is expected.
5702 * - There is no buffer space left
5703 * - Unexpected TCP flags/window values/header lengths are received
5704 * (detected by checking the TCP header against pred_flags)
5705 * - Data is sent in both directions. Fast path only supports pure senders
5706 * or pure receivers (this means either the sequence number or the ack
5707 * value must stay constant)
5708 * - Unexpected TCP option.
5710 * When these conditions are not satisfied it drops into a standard
5711 * receive procedure patterned after RFC793 to handle all cases.
5712 * The first three cases are guaranteed by proper pred_flags setting,
5713 * the rest is checked inline. Fast processing is turned on in
5714 * tcp_data_queue when everything is OK.
5716 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb)
5718 const struct tcphdr *th = (const struct tcphdr *)skb->data;
5719 struct tcp_sock *tp = tcp_sk(sk);
5720 unsigned int len = skb->len;
5722 /* TCP congestion window tracking */
5723 trace_tcp_probe(sk, skb);
5725 tcp_mstamp_refresh(tp);
5726 if (unlikely(!sk->sk_rx_dst))
5727 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
5729 * Header prediction.
5730 * The code loosely follows the one in the famous
5731 * "30 instruction TCP receive" Van Jacobson mail.
5733 * Van's trick is to deposit buffers into socket queue
5734 * on a device interrupt, to call tcp_recv function
5735 * on the receive process context and checksum and copy
5736 * the buffer to user space. smart...
5738 * Our current scheme is not silly either but we take the
5739 * extra cost of the net_bh soft interrupt processing...
5740 * We do checksum and copy also but from device to kernel.
5743 tp->rx_opt.saw_tstamp = 0;
5745 /* pred_flags is 0xS?10 << 16 + snd_wnd
5746 * if header_prediction is to be made
5747 * 'S' will always be tp->tcp_header_len >> 2
5748 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5749 * turn it off (when there are holes in the receive
5750 * space for instance)
5751 * PSH flag is ignored.
5754 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5755 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5756 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5757 int tcp_header_len = tp->tcp_header_len;
5759 /* Timestamp header prediction: tcp_header_len
5760 * is automatically equal to th->doff*4 due to pred_flags
5764 /* Check timestamp */
5765 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5766 /* No? Slow path! */
5767 if (!tcp_parse_aligned_timestamp(tp, th))
5770 /* If PAWS failed, check it more carefully in slow path */
5771 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5774 /* DO NOT update ts_recent here, if checksum fails
5775 * and timestamp was corrupted part, it will result
5776 * in a hung connection since we will drop all
5777 * future packets due to the PAWS test.
5781 if (len <= tcp_header_len) {
5782 /* Bulk data transfer: sender */
5783 if (len == tcp_header_len) {
5784 /* Predicted packet is in window by definition.
5785 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5786 * Hence, check seq<=rcv_wup reduces to:
5788 if (tcp_header_len ==
5789 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5790 tp->rcv_nxt == tp->rcv_wup)
5791 tcp_store_ts_recent(tp);
5793 /* We know that such packets are checksummed
5796 tcp_ack(sk, skb, 0);
5798 tcp_data_snd_check(sk);
5799 /* When receiving pure ack in fast path, update
5800 * last ts ecr directly instead of calling
5801 * tcp_rcv_rtt_measure_ts()
5803 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr;
5805 } else { /* Header too small */
5806 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5811 bool fragstolen = false;
5813 if (tcp_checksum_complete(skb))
5816 if ((int)skb->truesize > sk->sk_forward_alloc)
5819 /* Predicted packet is in window by definition.
5820 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5821 * Hence, check seq<=rcv_wup reduces to:
5823 if (tcp_header_len ==
5824 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5825 tp->rcv_nxt == tp->rcv_wup)
5826 tcp_store_ts_recent(tp);
5828 tcp_rcv_rtt_measure_ts(sk, skb);
5830 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPHITS);
5832 /* Bulk data transfer: receiver */
5833 __skb_pull(skb, tcp_header_len);
5834 eaten = tcp_queue_rcv(sk, skb, &fragstolen);
5836 tcp_event_data_recv(sk, skb);
5838 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5839 /* Well, only one small jumplet in fast path... */
5840 tcp_ack(sk, skb, FLAG_DATA);
5841 tcp_data_snd_check(sk);
5842 if (!inet_csk_ack_scheduled(sk))
5845 tcp_update_wl(tp, TCP_SKB_CB(skb)->seq);
5848 __tcp_ack_snd_check(sk, 0);
5851 kfree_skb_partial(skb, fragstolen);
5858 if (len < (th->doff << 2) || tcp_checksum_complete(skb))
5861 if (!th->ack && !th->rst && !th->syn)
5865 * Standard slow path.
5868 if (!tcp_validate_incoming(sk, skb, th, 1))
5872 if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0)
5875 tcp_rcv_rtt_measure_ts(sk, skb);
5877 /* Process urgent data. */
5878 tcp_urg(sk, skb, th);
5880 /* step 7: process the segment text */
5881 tcp_data_queue(sk, skb);
5883 tcp_data_snd_check(sk);
5884 tcp_ack_snd_check(sk);
5888 TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS);
5889 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5894 EXPORT_SYMBOL(tcp_rcv_established);
5896 void tcp_init_transfer(struct sock *sk, int bpf_op, struct sk_buff *skb)
5898 struct inet_connection_sock *icsk = inet_csk(sk);
5899 struct tcp_sock *tp = tcp_sk(sk);
5902 icsk->icsk_af_ops->rebuild_header(sk);
5903 tcp_init_metrics(sk);
5905 /* Initialize the congestion window to start the transfer.
5906 * Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
5907 * retransmitted. In light of RFC6298 more aggressive 1sec
5908 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
5909 * retransmission has occurred.
5911 if (tp->total_retrans > 1 && tp->undo_marker)
5914 tp->snd_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk));
5915 tp->snd_cwnd_stamp = tcp_jiffies32;
5917 icsk->icsk_ca_initialized = 0;
5918 bpf_skops_established(sk, bpf_op, skb);
5919 if (!icsk->icsk_ca_initialized)
5920 tcp_init_congestion_control(sk);
5921 tcp_init_buffer_space(sk);
5924 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5926 struct tcp_sock *tp = tcp_sk(sk);
5927 struct inet_connection_sock *icsk = inet_csk(sk);
5929 tcp_set_state(sk, TCP_ESTABLISHED);
5930 icsk->icsk_ack.lrcvtime = tcp_jiffies32;
5933 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
5934 security_inet_conn_established(sk, skb);
5935 sk_mark_napi_id(sk, skb);
5938 tcp_init_transfer(sk, BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB, skb);
5940 /* Prevent spurious tcp_cwnd_restart() on first data
5943 tp->lsndtime = tcp_jiffies32;
5945 if (sock_flag(sk, SOCK_KEEPOPEN))
5946 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5948 if (!tp->rx_opt.snd_wscale)
5949 __tcp_fast_path_on(tp, tp->snd_wnd);
5954 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
5955 struct tcp_fastopen_cookie *cookie)
5957 struct tcp_sock *tp = tcp_sk(sk);
5958 struct sk_buff *data = tp->syn_data ? tcp_rtx_queue_head(sk) : NULL;
5959 u16 mss = tp->rx_opt.mss_clamp, try_exp = 0;
5960 bool syn_drop = false;
5962 if (mss == tp->rx_opt.user_mss) {
5963 struct tcp_options_received opt;
5965 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5966 tcp_clear_options(&opt);
5967 opt.user_mss = opt.mss_clamp = 0;
5968 tcp_parse_options(sock_net(sk), synack, &opt, 0, NULL);
5969 mss = opt.mss_clamp;
5972 if (!tp->syn_fastopen) {
5973 /* Ignore an unsolicited cookie */
5975 } else if (tp->total_retrans) {
5976 /* SYN timed out and the SYN-ACK neither has a cookie nor
5977 * acknowledges data. Presumably the remote received only
5978 * the retransmitted (regular) SYNs: either the original
5979 * SYN-data or the corresponding SYN-ACK was dropped.
5981 syn_drop = (cookie->len < 0 && data);
5982 } else if (cookie->len < 0 && !tp->syn_data) {
5983 /* We requested a cookie but didn't get it. If we did not use
5984 * the (old) exp opt format then try so next time (try_exp=1).
5985 * Otherwise we go back to use the RFC7413 opt (try_exp=2).
5987 try_exp = tp->syn_fastopen_exp ? 2 : 1;
5990 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp);
5992 if (data) { /* Retransmit unacked data in SYN */
5993 if (tp->total_retrans)
5994 tp->fastopen_client_fail = TFO_SYN_RETRANSMITTED;
5996 tp->fastopen_client_fail = TFO_DATA_NOT_ACKED;
5997 skb_rbtree_walk_from(data) {
5998 if (__tcp_retransmit_skb(sk, data, 1))
6002 NET_INC_STATS(sock_net(sk),
6003 LINUX_MIB_TCPFASTOPENACTIVEFAIL);
6006 tp->syn_data_acked = tp->syn_data;
6007 if (tp->syn_data_acked) {
6008 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
6009 /* SYN-data is counted as two separate packets in tcp_ack() */
6010 if (tp->delivered > 1)
6014 tcp_fastopen_add_skb(sk, synack);
6019 static void smc_check_reset_syn(struct tcp_sock *tp)
6021 #if IS_ENABLED(CONFIG_SMC)
6022 if (static_branch_unlikely(&tcp_have_smc)) {
6023 if (tp->syn_smc && !tp->rx_opt.smc_ok)
6029 static void tcp_try_undo_spurious_syn(struct sock *sk)
6031 struct tcp_sock *tp = tcp_sk(sk);
6034 /* undo_marker is set when SYN or SYNACK times out. The timeout is
6035 * spurious if the ACK's timestamp option echo value matches the
6036 * original SYN timestamp.
6038 syn_stamp = tp->retrans_stamp;
6039 if (tp->undo_marker && syn_stamp && tp->rx_opt.saw_tstamp &&
6040 syn_stamp == tp->rx_opt.rcv_tsecr)
6041 tp->undo_marker = 0;
6044 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
6045 const struct tcphdr *th)
6047 struct inet_connection_sock *icsk = inet_csk(sk);
6048 struct tcp_sock *tp = tcp_sk(sk);
6049 struct tcp_fastopen_cookie foc = { .len = -1 };
6050 int saved_clamp = tp->rx_opt.mss_clamp;
6053 tcp_parse_options(sock_net(sk), skb, &tp->rx_opt, 0, &foc);
6054 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
6055 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
6059 * "If the state is SYN-SENT then
6060 * first check the ACK bit
6061 * If the ACK bit is set
6062 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
6063 * a reset (unless the RST bit is set, if so drop
6064 * the segment and return)"
6066 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
6067 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
6068 /* Previous FIN/ACK or RST/ACK might be ignored. */
6069 if (icsk->icsk_retransmits == 0)
6070 inet_csk_reset_xmit_timer(sk,
6072 TCP_TIMEOUT_MIN, TCP_RTO_MAX);
6073 goto reset_and_undo;
6076 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
6077 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
6078 tcp_time_stamp(tp))) {
6079 NET_INC_STATS(sock_net(sk),
6080 LINUX_MIB_PAWSACTIVEREJECTED);
6081 goto reset_and_undo;
6084 /* Now ACK is acceptable.
6086 * "If the RST bit is set
6087 * If the ACK was acceptable then signal the user "error:
6088 * connection reset", drop the segment, enter CLOSED state,
6089 * delete TCB, and return."
6098 * "fifth, if neither of the SYN or RST bits is set then
6099 * drop the segment and return."
6105 goto discard_and_undo;
6108 * "If the SYN bit is on ...
6109 * are acceptable then ...
6110 * (our SYN has been ACKed), change the connection
6111 * state to ESTABLISHED..."
6114 tcp_ecn_rcv_synack(tp, th);
6116 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
6117 tcp_try_undo_spurious_syn(sk);
6118 tcp_ack(sk, skb, FLAG_SLOWPATH);
6120 /* Ok.. it's good. Set up sequence numbers and
6121 * move to established.
6123 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1);
6124 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
6126 /* RFC1323: The window in SYN & SYN/ACK segments is
6129 tp->snd_wnd = ntohs(th->window);
6131 if (!tp->rx_opt.wscale_ok) {
6132 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
6133 tp->window_clamp = min(tp->window_clamp, 65535U);
6136 if (tp->rx_opt.saw_tstamp) {
6137 tp->rx_opt.tstamp_ok = 1;
6138 tp->tcp_header_len =
6139 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
6140 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
6141 tcp_store_ts_recent(tp);
6143 tp->tcp_header_len = sizeof(struct tcphdr);
6146 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
6147 tcp_initialize_rcv_mss(sk);
6149 /* Remember, tcp_poll() does not lock socket!
6150 * Change state from SYN-SENT only after copied_seq
6151 * is initialized. */
6152 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6154 smc_check_reset_syn(tp);
6158 tcp_finish_connect(sk, skb);
6160 fastopen_fail = (tp->syn_fastopen || tp->syn_data) &&
6161 tcp_rcv_fastopen_synack(sk, skb, &foc);
6163 if (!sock_flag(sk, SOCK_DEAD)) {
6164 sk->sk_state_change(sk);
6165 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
6169 if (sk->sk_write_pending ||
6170 icsk->icsk_accept_queue.rskq_defer_accept ||
6171 inet_csk_in_pingpong_mode(sk)) {
6172 /* Save one ACK. Data will be ready after
6173 * several ticks, if write_pending is set.
6175 * It may be deleted, but with this feature tcpdumps
6176 * look so _wonderfully_ clever, that I was not able
6177 * to stand against the temptation 8) --ANK
6179 inet_csk_schedule_ack(sk);
6180 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
6181 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
6182 TCP_DELACK_MAX, TCP_RTO_MAX);
6193 /* No ACK in the segment */
6197 * "If the RST bit is set
6199 * Otherwise (no ACK) drop the segment and return."
6202 goto discard_and_undo;
6206 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
6207 tcp_paws_reject(&tp->rx_opt, 0))
6208 goto discard_and_undo;
6211 /* We see SYN without ACK. It is attempt of
6212 * simultaneous connect with crossed SYNs.
6213 * Particularly, it can be connect to self.
6215 tcp_set_state(sk, TCP_SYN_RECV);
6217 if (tp->rx_opt.saw_tstamp) {
6218 tp->rx_opt.tstamp_ok = 1;
6219 tcp_store_ts_recent(tp);
6220 tp->tcp_header_len =
6221 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
6223 tp->tcp_header_len = sizeof(struct tcphdr);
6226 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1);
6227 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6228 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
6230 /* RFC1323: The window in SYN & SYN/ACK segments is
6233 tp->snd_wnd = ntohs(th->window);
6234 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
6235 tp->max_window = tp->snd_wnd;
6237 tcp_ecn_rcv_syn(tp, th);
6240 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
6241 tcp_initialize_rcv_mss(sk);
6243 tcp_send_synack(sk);
6245 /* Note, we could accept data and URG from this segment.
6246 * There are no obstacles to make this (except that we must
6247 * either change tcp_recvmsg() to prevent it from returning data
6248 * before 3WHS completes per RFC793, or employ TCP Fast Open).
6250 * However, if we ignore data in ACKless segments sometimes,
6251 * we have no reasons to accept it sometimes.
6252 * Also, seems the code doing it in step6 of tcp_rcv_state_process
6253 * is not flawless. So, discard packet for sanity.
6254 * Uncomment this return to process the data.
6261 /* "fifth, if neither of the SYN or RST bits is set then
6262 * drop the segment and return."
6266 tcp_clear_options(&tp->rx_opt);
6267 tp->rx_opt.mss_clamp = saved_clamp;
6271 tcp_clear_options(&tp->rx_opt);
6272 tp->rx_opt.mss_clamp = saved_clamp;
6276 static void tcp_rcv_synrecv_state_fastopen(struct sock *sk)
6278 struct request_sock *req;
6280 /* If we are still handling the SYNACK RTO, see if timestamp ECR allows
6281 * undo. If peer SACKs triggered fast recovery, we can't undo here.
6283 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
6284 tcp_try_undo_loss(sk, false);
6286 /* Reset rtx states to prevent spurious retransmits_timed_out() */
6287 tcp_sk(sk)->retrans_stamp = 0;
6288 inet_csk(sk)->icsk_retransmits = 0;
6290 /* Once we leave TCP_SYN_RECV or TCP_FIN_WAIT_1,
6291 * we no longer need req so release it.
6293 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk,
6294 lockdep_sock_is_held(sk));
6295 reqsk_fastopen_remove(sk, req, false);
6297 /* Re-arm the timer because data may have been sent out.
6298 * This is similar to the regular data transmission case
6299 * when new data has just been ack'ed.
6301 * (TFO) - we could try to be more aggressive and
6302 * retransmitting any data sooner based on when they
6309 * This function implements the receiving procedure of RFC 793 for
6310 * all states except ESTABLISHED and TIME_WAIT.
6311 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
6312 * address independent.
6315 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb)
6317 struct tcp_sock *tp = tcp_sk(sk);
6318 struct inet_connection_sock *icsk = inet_csk(sk);
6319 const struct tcphdr *th = tcp_hdr(skb);
6320 struct request_sock *req;
6324 switch (sk->sk_state) {
6338 /* It is possible that we process SYN packets from backlog,
6339 * so we need to make sure to disable BH and RCU right there.
6343 acceptable = icsk->icsk_af_ops->conn_request(sk, skb) >= 0;
6355 tp->rx_opt.saw_tstamp = 0;
6356 tcp_mstamp_refresh(tp);
6357 queued = tcp_rcv_synsent_state_process(sk, skb, th);
6361 /* Do step6 onward by hand. */
6362 tcp_urg(sk, skb, th);
6364 tcp_data_snd_check(sk);
6368 tcp_mstamp_refresh(tp);
6369 tp->rx_opt.saw_tstamp = 0;
6370 req = rcu_dereference_protected(tp->fastopen_rsk,
6371 lockdep_sock_is_held(sk));
6375 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
6376 sk->sk_state != TCP_FIN_WAIT1);
6378 if (!tcp_check_req(sk, skb, req, true, &req_stolen))
6382 if (!th->ack && !th->rst && !th->syn)
6385 if (!tcp_validate_incoming(sk, skb, th, 0))
6388 /* step 5: check the ACK field */
6389 acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
6390 FLAG_UPDATE_TS_RECENT |
6391 FLAG_NO_CHALLENGE_ACK) > 0;
6394 if (sk->sk_state == TCP_SYN_RECV)
6395 return 1; /* send one RST */
6396 tcp_send_challenge_ack(sk, skb);
6399 switch (sk->sk_state) {
6401 tp->delivered++; /* SYN-ACK delivery isn't tracked in tcp_ack */
6403 tcp_synack_rtt_meas(sk, req);
6406 tcp_rcv_synrecv_state_fastopen(sk);
6408 tcp_try_undo_spurious_syn(sk);
6409 tp->retrans_stamp = 0;
6410 tcp_init_transfer(sk, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB,
6412 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6415 tcp_set_state(sk, TCP_ESTABLISHED);
6416 sk->sk_state_change(sk);
6418 /* Note, that this wakeup is only for marginal crossed SYN case.
6419 * Passively open sockets are not waked up, because
6420 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
6423 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
6425 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
6426 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
6427 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
6429 if (tp->rx_opt.tstamp_ok)
6430 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
6432 if (!inet_csk(sk)->icsk_ca_ops->cong_control)
6433 tcp_update_pacing_rate(sk);
6435 /* Prevent spurious tcp_cwnd_restart() on first data packet */
6436 tp->lsndtime = tcp_jiffies32;
6438 tcp_initialize_rcv_mss(sk);
6439 tcp_fast_path_on(tp);
6442 case TCP_FIN_WAIT1: {
6446 tcp_rcv_synrecv_state_fastopen(sk);
6448 if (tp->snd_una != tp->write_seq)
6451 tcp_set_state(sk, TCP_FIN_WAIT2);
6452 sk->sk_shutdown |= SEND_SHUTDOWN;
6456 if (!sock_flag(sk, SOCK_DEAD)) {
6457 /* Wake up lingering close() */
6458 sk->sk_state_change(sk);
6462 if (tp->linger2 < 0) {
6464 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6467 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6468 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6469 /* Receive out of order FIN after close() */
6470 if (tp->syn_fastopen && th->fin)
6471 tcp_fastopen_active_disable(sk);
6473 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6477 tmo = tcp_fin_time(sk);
6478 if (tmo > TCP_TIMEWAIT_LEN) {
6479 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
6480 } else if (th->fin || sock_owned_by_user(sk)) {
6481 /* Bad case. We could lose such FIN otherwise.
6482 * It is not a big problem, but it looks confusing
6483 * and not so rare event. We still can lose it now,
6484 * if it spins in bh_lock_sock(), but it is really
6487 inet_csk_reset_keepalive_timer(sk, tmo);
6489 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
6496 if (tp->snd_una == tp->write_seq) {
6497 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
6503 if (tp->snd_una == tp->write_seq) {
6504 tcp_update_metrics(sk);
6511 /* step 6: check the URG bit */
6512 tcp_urg(sk, skb, th);
6514 /* step 7: process the segment text */
6515 switch (sk->sk_state) {
6516 case TCP_CLOSE_WAIT:
6519 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
6520 if (sk_is_mptcp(sk))
6521 mptcp_incoming_options(sk, skb);
6527 /* RFC 793 says to queue data in these states,
6528 * RFC 1122 says we MUST send a reset.
6529 * BSD 4.4 also does reset.
6531 if (sk->sk_shutdown & RCV_SHUTDOWN) {
6532 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6533 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6534 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6540 case TCP_ESTABLISHED:
6541 tcp_data_queue(sk, skb);
6546 /* tcp_data could move socket to TIME-WAIT */
6547 if (sk->sk_state != TCP_CLOSE) {
6548 tcp_data_snd_check(sk);
6549 tcp_ack_snd_check(sk);
6558 EXPORT_SYMBOL(tcp_rcv_state_process);
6560 static inline void pr_drop_req(struct request_sock *req, __u16 port, int family)
6562 struct inet_request_sock *ireq = inet_rsk(req);
6564 if (family == AF_INET)
6565 net_dbg_ratelimited("drop open request from %pI4/%u\n",
6566 &ireq->ir_rmt_addr, port);
6567 #if IS_ENABLED(CONFIG_IPV6)
6568 else if (family == AF_INET6)
6569 net_dbg_ratelimited("drop open request from %pI6/%u\n",
6570 &ireq->ir_v6_rmt_addr, port);
6574 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
6576 * If we receive a SYN packet with these bits set, it means a
6577 * network is playing bad games with TOS bits. In order to
6578 * avoid possible false congestion notifications, we disable
6579 * TCP ECN negotiation.
6581 * Exception: tcp_ca wants ECN. This is required for DCTCP
6582 * congestion control: Linux DCTCP asserts ECT on all packets,
6583 * including SYN, which is most optimal solution; however,
6584 * others, such as FreeBSD do not.
6586 * Exception: At least one of the reserved bits of the TCP header (th->res1) is
6587 * set, indicating the use of a future TCP extension (such as AccECN). See
6588 * RFC8311 §4.3 which updates RFC3168 to allow the development of such
6591 static void tcp_ecn_create_request(struct request_sock *req,
6592 const struct sk_buff *skb,
6593 const struct sock *listen_sk,
6594 const struct dst_entry *dst)
6596 const struct tcphdr *th = tcp_hdr(skb);
6597 const struct net *net = sock_net(listen_sk);
6598 bool th_ecn = th->ece && th->cwr;
6605 ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield);
6606 ecn_ok_dst = dst_feature(dst, DST_FEATURE_ECN_MASK);
6607 ecn_ok = net->ipv4.sysctl_tcp_ecn || ecn_ok_dst;
6609 if (((!ect || th->res1) && ecn_ok) || tcp_ca_needs_ecn(listen_sk) ||
6610 (ecn_ok_dst & DST_FEATURE_ECN_CA) ||
6611 tcp_bpf_ca_needs_ecn((struct sock *)req))
6612 inet_rsk(req)->ecn_ok = 1;
6615 static void tcp_openreq_init(struct request_sock *req,
6616 const struct tcp_options_received *rx_opt,
6617 struct sk_buff *skb, const struct sock *sk)
6619 struct inet_request_sock *ireq = inet_rsk(req);
6621 req->rsk_rcv_wnd = 0; /* So that tcp_send_synack() knows! */
6622 tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq;
6623 tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
6624 tcp_rsk(req)->snt_synack = 0;
6625 tcp_rsk(req)->last_oow_ack_time = 0;
6626 req->mss = rx_opt->mss_clamp;
6627 req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0;
6628 ireq->tstamp_ok = rx_opt->tstamp_ok;
6629 ireq->sack_ok = rx_opt->sack_ok;
6630 ireq->snd_wscale = rx_opt->snd_wscale;
6631 ireq->wscale_ok = rx_opt->wscale_ok;
6634 ireq->ir_rmt_port = tcp_hdr(skb)->source;
6635 ireq->ir_num = ntohs(tcp_hdr(skb)->dest);
6636 ireq->ir_mark = inet_request_mark(sk, skb);
6637 #if IS_ENABLED(CONFIG_SMC)
6638 ireq->smc_ok = rx_opt->smc_ok;
6642 struct request_sock *inet_reqsk_alloc(const struct request_sock_ops *ops,
6643 struct sock *sk_listener,
6644 bool attach_listener)
6646 struct request_sock *req = reqsk_alloc(ops, sk_listener,
6650 struct inet_request_sock *ireq = inet_rsk(req);
6652 ireq->ireq_opt = NULL;
6653 #if IS_ENABLED(CONFIG_IPV6)
6654 ireq->pktopts = NULL;
6656 atomic64_set(&ireq->ir_cookie, 0);
6657 ireq->ireq_state = TCP_NEW_SYN_RECV;
6658 write_pnet(&ireq->ireq_net, sock_net(sk_listener));
6659 ireq->ireq_family = sk_listener->sk_family;
6664 EXPORT_SYMBOL(inet_reqsk_alloc);
6667 * Return true if a syncookie should be sent
6669 static bool tcp_syn_flood_action(const struct sock *sk, const char *proto)
6671 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
6672 const char *msg = "Dropping request";
6673 bool want_cookie = false;
6674 struct net *net = sock_net(sk);
6676 #ifdef CONFIG_SYN_COOKIES
6677 if (net->ipv4.sysctl_tcp_syncookies) {
6678 msg = "Sending cookies";
6680 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES);
6683 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP);
6685 if (!queue->synflood_warned &&
6686 net->ipv4.sysctl_tcp_syncookies != 2 &&
6687 xchg(&queue->synflood_warned, 1) == 0)
6688 net_info_ratelimited("%s: Possible SYN flooding on port %d. %s. Check SNMP counters.\n",
6689 proto, sk->sk_num, msg);
6694 static void tcp_reqsk_record_syn(const struct sock *sk,
6695 struct request_sock *req,
6696 const struct sk_buff *skb)
6698 if (tcp_sk(sk)->save_syn) {
6699 u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb);
6700 struct saved_syn *saved_syn;
6704 if (tcp_sk(sk)->save_syn == 2) { /* Save full header. */
6705 base = skb_mac_header(skb);
6706 mac_hdrlen = skb_mac_header_len(skb);
6709 base = skb_network_header(skb);
6713 saved_syn = kmalloc(struct_size(saved_syn, data, len),
6716 saved_syn->mac_hdrlen = mac_hdrlen;
6717 saved_syn->network_hdrlen = skb_network_header_len(skb);
6718 saved_syn->tcp_hdrlen = tcp_hdrlen(skb);
6719 memcpy(saved_syn->data, base, len);
6720 req->saved_syn = saved_syn;
6725 /* If a SYN cookie is required and supported, returns a clamped MSS value to be
6726 * used for SYN cookie generation.
6728 u16 tcp_get_syncookie_mss(struct request_sock_ops *rsk_ops,
6729 const struct tcp_request_sock_ops *af_ops,
6730 struct sock *sk, struct tcphdr *th)
6732 struct tcp_sock *tp = tcp_sk(sk);
6735 if (sock_net(sk)->ipv4.sysctl_tcp_syncookies != 2 &&
6736 !inet_csk_reqsk_queue_is_full(sk))
6739 if (!tcp_syn_flood_action(sk, rsk_ops->slab_name))
6742 if (sk_acceptq_is_full(sk)) {
6743 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
6747 mss = tcp_parse_mss_option(th, tp->rx_opt.user_mss);
6749 mss = af_ops->mss_clamp;
6753 EXPORT_SYMBOL_GPL(tcp_get_syncookie_mss);
6755 int tcp_conn_request(struct request_sock_ops *rsk_ops,
6756 const struct tcp_request_sock_ops *af_ops,
6757 struct sock *sk, struct sk_buff *skb)
6759 struct tcp_fastopen_cookie foc = { .len = -1 };
6760 __u32 isn = TCP_SKB_CB(skb)->tcp_tw_isn;
6761 struct tcp_options_received tmp_opt;
6762 struct tcp_sock *tp = tcp_sk(sk);
6763 struct net *net = sock_net(sk);
6764 struct sock *fastopen_sk = NULL;
6765 struct request_sock *req;
6766 bool want_cookie = false;
6767 struct dst_entry *dst;
6770 /* TW buckets are converted to open requests without
6771 * limitations, they conserve resources and peer is
6772 * evidently real one.
6774 if ((net->ipv4.sysctl_tcp_syncookies == 2 ||
6775 inet_csk_reqsk_queue_is_full(sk)) && !isn) {
6776 want_cookie = tcp_syn_flood_action(sk, rsk_ops->slab_name);
6781 if (sk_acceptq_is_full(sk)) {
6782 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
6786 req = inet_reqsk_alloc(rsk_ops, sk, !want_cookie);
6790 req->syncookie = want_cookie;
6791 tcp_rsk(req)->af_specific = af_ops;
6792 tcp_rsk(req)->ts_off = 0;
6793 #if IS_ENABLED(CONFIG_MPTCP)
6794 tcp_rsk(req)->is_mptcp = 0;
6797 tcp_clear_options(&tmp_opt);
6798 tmp_opt.mss_clamp = af_ops->mss_clamp;
6799 tmp_opt.user_mss = tp->rx_opt.user_mss;
6800 tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0,
6801 want_cookie ? NULL : &foc);
6803 if (want_cookie && !tmp_opt.saw_tstamp)
6804 tcp_clear_options(&tmp_opt);
6806 if (IS_ENABLED(CONFIG_SMC) && want_cookie)
6809 tmp_opt.tstamp_ok = tmp_opt.saw_tstamp;
6810 tcp_openreq_init(req, &tmp_opt, skb, sk);
6811 inet_rsk(req)->no_srccheck = inet_sk(sk)->transparent;
6813 /* Note: tcp_v6_init_req() might override ir_iif for link locals */
6814 inet_rsk(req)->ir_iif = inet_request_bound_dev_if(sk, skb);
6816 dst = af_ops->route_req(sk, skb, &fl, req);
6820 if (tmp_opt.tstamp_ok)
6821 tcp_rsk(req)->ts_off = af_ops->init_ts_off(net, skb);
6823 if (!want_cookie && !isn) {
6824 /* Kill the following clause, if you dislike this way. */
6825 if (!net->ipv4.sysctl_tcp_syncookies &&
6826 (net->ipv4.sysctl_max_syn_backlog - inet_csk_reqsk_queue_len(sk) <
6827 (net->ipv4.sysctl_max_syn_backlog >> 2)) &&
6828 !tcp_peer_is_proven(req, dst)) {
6829 /* Without syncookies last quarter of
6830 * backlog is filled with destinations,
6831 * proven to be alive.
6832 * It means that we continue to communicate
6833 * to destinations, already remembered
6834 * to the moment of synflood.
6836 pr_drop_req(req, ntohs(tcp_hdr(skb)->source),
6838 goto drop_and_release;
6841 isn = af_ops->init_seq(skb);
6844 tcp_ecn_create_request(req, skb, sk, dst);
6847 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss);
6848 if (!tmp_opt.tstamp_ok)
6849 inet_rsk(req)->ecn_ok = 0;
6852 tcp_rsk(req)->snt_isn = isn;
6853 tcp_rsk(req)->txhash = net_tx_rndhash();
6854 tcp_rsk(req)->syn_tos = TCP_SKB_CB(skb)->ip_dsfield;
6855 tcp_openreq_init_rwin(req, sk, dst);
6856 sk_rx_queue_set(req_to_sk(req), skb);
6858 tcp_reqsk_record_syn(sk, req, skb);
6859 fastopen_sk = tcp_try_fastopen(sk, skb, req, &foc, dst);
6862 af_ops->send_synack(fastopen_sk, dst, &fl, req,
6863 &foc, TCP_SYNACK_FASTOPEN, skb);
6864 /* Add the child socket directly into the accept queue */
6865 if (!inet_csk_reqsk_queue_add(sk, req, fastopen_sk)) {
6866 reqsk_fastopen_remove(fastopen_sk, req, false);
6867 bh_unlock_sock(fastopen_sk);
6868 sock_put(fastopen_sk);
6871 sk->sk_data_ready(sk);
6872 bh_unlock_sock(fastopen_sk);
6873 sock_put(fastopen_sk);
6875 tcp_rsk(req)->tfo_listener = false;
6877 inet_csk_reqsk_queue_hash_add(sk, req,
6878 tcp_timeout_init((struct sock *)req));
6879 af_ops->send_synack(sk, dst, &fl, req, &foc,
6880 !want_cookie ? TCP_SYNACK_NORMAL :
6899 EXPORT_SYMBOL(tcp_conn_request);