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>
80 int sysctl_tcp_fack __read_mostly;
81 int sysctl_tcp_max_reordering __read_mostly = 300;
82 int sysctl_tcp_dsack __read_mostly = 1;
83 int sysctl_tcp_app_win __read_mostly = 31;
84 int sysctl_tcp_adv_win_scale __read_mostly = 1;
85 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale);
87 /* rfc5961 challenge ack rate limiting */
88 int sysctl_tcp_challenge_ack_limit = 1000;
90 int sysctl_tcp_stdurg __read_mostly;
91 int sysctl_tcp_rfc1337 __read_mostly;
92 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
93 int sysctl_tcp_frto __read_mostly = 2;
94 int sysctl_tcp_min_rtt_wlen __read_mostly = 300;
95 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
96 int sysctl_tcp_early_retrans __read_mostly = 3;
97 int sysctl_tcp_invalid_ratelimit __read_mostly = HZ/2;
99 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
100 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
101 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
102 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
103 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
104 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
105 #define FLAG_ECE 0x40 /* ECE in this ACK */
106 #define FLAG_LOST_RETRANS 0x80 /* This ACK marks some retransmission lost */
107 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
108 #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
109 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
110 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
111 #define FLAG_SET_XMIT_TIMER 0x1000 /* Set TLP or RTO timer */
112 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
113 #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
114 #define FLAG_NO_CHALLENGE_ACK 0x8000 /* do not call tcp_send_challenge_ack() */
116 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
117 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
118 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE|FLAG_DSACKING_ACK)
119 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
121 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
122 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
124 #define REXMIT_NONE 0 /* no loss recovery to do */
125 #define REXMIT_LOST 1 /* retransmit packets marked lost */
126 #define REXMIT_NEW 2 /* FRTO-style transmit of unsent/new packets */
128 static void tcp_gro_dev_warn(struct sock *sk, const struct sk_buff *skb,
131 static bool __once __read_mostly;
134 struct net_device *dev;
139 dev = dev_get_by_index_rcu(sock_net(sk), skb->skb_iif);
140 if (!dev || len >= dev->mtu)
141 pr_warn("%s: Driver has suspect GRO implementation, TCP performance may be compromised.\n",
142 dev ? dev->name : "Unknown driver");
147 /* Adapt the MSS value used to make delayed ack decision to the
150 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
152 struct inet_connection_sock *icsk = inet_csk(sk);
153 const unsigned int lss = icsk->icsk_ack.last_seg_size;
156 icsk->icsk_ack.last_seg_size = 0;
158 /* skb->len may jitter because of SACKs, even if peer
159 * sends good full-sized frames.
161 len = skb_shinfo(skb)->gso_size ? : skb->len;
162 if (len >= icsk->icsk_ack.rcv_mss) {
163 icsk->icsk_ack.rcv_mss = min_t(unsigned int, len,
165 /* Account for possibly-removed options */
166 if (unlikely(len > icsk->icsk_ack.rcv_mss +
167 MAX_TCP_OPTION_SPACE))
168 tcp_gro_dev_warn(sk, skb, len);
170 /* Otherwise, we make more careful check taking into account,
171 * that SACKs block is variable.
173 * "len" is invariant segment length, including TCP header.
175 len += skb->data - skb_transport_header(skb);
176 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
177 /* If PSH is not set, packet should be
178 * full sized, provided peer TCP is not badly broken.
179 * This observation (if it is correct 8)) allows
180 * to handle super-low mtu links fairly.
182 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
183 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
184 /* Subtract also invariant (if peer is RFC compliant),
185 * tcp header plus fixed timestamp option length.
186 * Resulting "len" is MSS free of SACK jitter.
188 len -= tcp_sk(sk)->tcp_header_len;
189 icsk->icsk_ack.last_seg_size = len;
191 icsk->icsk_ack.rcv_mss = len;
195 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
196 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
197 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
201 static void tcp_incr_quickack(struct sock *sk, unsigned int max_quickacks)
203 struct inet_connection_sock *icsk = inet_csk(sk);
204 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
208 quickacks = min(quickacks, max_quickacks);
209 if (quickacks > icsk->icsk_ack.quick)
210 icsk->icsk_ack.quick = quickacks;
213 void tcp_enter_quickack_mode(struct sock *sk, unsigned int max_quickacks)
215 struct inet_connection_sock *icsk = inet_csk(sk);
217 tcp_incr_quickack(sk, max_quickacks);
218 icsk->icsk_ack.pingpong = 0;
219 icsk->icsk_ack.ato = TCP_ATO_MIN;
221 EXPORT_SYMBOL(tcp_enter_quickack_mode);
223 /* Send ACKs quickly, if "quick" count is not exhausted
224 * and the session is not interactive.
227 static bool tcp_in_quickack_mode(struct sock *sk)
229 const struct inet_connection_sock *icsk = inet_csk(sk);
230 const struct dst_entry *dst = __sk_dst_get(sk);
232 return (dst && dst_metric(dst, RTAX_QUICKACK)) ||
233 (icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong);
236 static void tcp_ecn_queue_cwr(struct tcp_sock *tp)
238 if (tp->ecn_flags & TCP_ECN_OK)
239 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
242 static void tcp_ecn_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb)
244 if (tcp_hdr(skb)->cwr)
245 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
248 static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp)
250 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
253 static void __tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb)
255 struct tcp_sock *tp = tcp_sk(sk);
257 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
258 case INET_ECN_NOT_ECT:
259 /* Funny extension: if ECT is not set on a segment,
260 * and we already seen ECT on a previous segment,
261 * it is probably a retransmit.
263 if (tp->ecn_flags & TCP_ECN_SEEN)
264 tcp_enter_quickack_mode(sk, 2);
267 if (tcp_ca_needs_ecn(sk))
268 tcp_ca_event(sk, CA_EVENT_ECN_IS_CE);
270 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
271 /* Better not delay acks, sender can have a very low cwnd */
272 tcp_enter_quickack_mode(sk, 2);
273 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
275 tp->ecn_flags |= TCP_ECN_SEEN;
278 if (tcp_ca_needs_ecn(sk))
279 tcp_ca_event(sk, CA_EVENT_ECN_NO_CE);
280 tp->ecn_flags |= TCP_ECN_SEEN;
285 static void tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb)
287 if (tcp_sk(sk)->ecn_flags & TCP_ECN_OK)
288 __tcp_ecn_check_ce(sk, skb);
291 static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
293 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
294 tp->ecn_flags &= ~TCP_ECN_OK;
297 static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
299 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
300 tp->ecn_flags &= ~TCP_ECN_OK;
303 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
305 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
310 /* Buffer size and advertised window tuning.
312 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
315 static void tcp_sndbuf_expand(struct sock *sk)
317 const struct tcp_sock *tp = tcp_sk(sk);
318 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
322 /* Worst case is non GSO/TSO : each frame consumes one skb
323 * and skb->head is kmalloced using power of two area of memory
325 per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
327 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
329 per_mss = roundup_pow_of_two(per_mss) +
330 SKB_DATA_ALIGN(sizeof(struct sk_buff));
332 nr_segs = max_t(u32, TCP_INIT_CWND, tp->snd_cwnd);
333 nr_segs = max_t(u32, nr_segs, tp->reordering + 1);
335 /* Fast Recovery (RFC 5681 3.2) :
336 * Cubic needs 1.7 factor, rounded to 2 to include
337 * extra cushion (application might react slowly to POLLOUT)
339 sndmem = ca_ops->sndbuf_expand ? ca_ops->sndbuf_expand(sk) : 2;
340 sndmem *= nr_segs * per_mss;
342 if (sk->sk_sndbuf < sndmem)
343 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
346 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
348 * All tcp_full_space() is split to two parts: "network" buffer, allocated
349 * forward and advertised in receiver window (tp->rcv_wnd) and
350 * "application buffer", required to isolate scheduling/application
351 * latencies from network.
352 * window_clamp is maximal advertised window. It can be less than
353 * tcp_full_space(), in this case tcp_full_space() - window_clamp
354 * is reserved for "application" buffer. The less window_clamp is
355 * the smoother our behaviour from viewpoint of network, but the lower
356 * throughput and the higher sensitivity of the connection to losses. 8)
358 * rcv_ssthresh is more strict window_clamp used at "slow start"
359 * phase to predict further behaviour of this connection.
360 * It is used for two goals:
361 * - to enforce header prediction at sender, even when application
362 * requires some significant "application buffer". It is check #1.
363 * - to prevent pruning of receive queue because of misprediction
364 * of receiver window. Check #2.
366 * The scheme does not work when sender sends good segments opening
367 * window and then starts to feed us spaghetti. But it should work
368 * in common situations. Otherwise, we have to rely on queue collapsing.
371 /* Slow part of check#2. */
372 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
374 struct tcp_sock *tp = tcp_sk(sk);
376 int truesize = tcp_win_from_space(skb->truesize) >> 1;
377 int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1;
379 while (tp->rcv_ssthresh <= window) {
380 if (truesize <= skb->len)
381 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
389 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
391 struct tcp_sock *tp = tcp_sk(sk);
394 room = min_t(int, tp->window_clamp, tcp_space(sk)) - tp->rcv_ssthresh;
397 if (room > 0 && !tcp_under_memory_pressure(sk)) {
400 /* Check #2. Increase window, if skb with such overhead
401 * will fit to rcvbuf in future.
403 if (tcp_win_from_space(skb->truesize) <= skb->len)
404 incr = 2 * tp->advmss;
406 incr = __tcp_grow_window(sk, skb);
409 incr = max_t(int, incr, 2 * skb->len);
410 tp->rcv_ssthresh += min(room, incr);
411 inet_csk(sk)->icsk_ack.quick |= 1;
416 /* 3. Tuning rcvbuf, when connection enters established state. */
417 static void tcp_fixup_rcvbuf(struct sock *sk)
419 u32 mss = tcp_sk(sk)->advmss;
422 rcvmem = 2 * SKB_TRUESIZE(mss + MAX_TCP_HEADER) *
423 tcp_default_init_rwnd(mss);
425 /* Dynamic Right Sizing (DRS) has 2 to 3 RTT latency
426 * Allow enough cushion so that sender is not limited by our window
428 if (sysctl_tcp_moderate_rcvbuf)
431 if (sk->sk_rcvbuf < rcvmem)
432 sk->sk_rcvbuf = min(rcvmem, sysctl_tcp_rmem[2]);
435 /* 4. Try to fixup all. It is made immediately after connection enters
438 void tcp_init_buffer_space(struct sock *sk)
440 struct tcp_sock *tp = tcp_sk(sk);
443 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
444 tcp_fixup_rcvbuf(sk);
445 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
446 tcp_sndbuf_expand(sk);
448 tp->rcvq_space.space = tp->rcv_wnd;
449 tcp_mstamp_refresh(tp);
450 tp->rcvq_space.time = tp->tcp_mstamp;
451 tp->rcvq_space.seq = tp->copied_seq;
453 maxwin = tcp_full_space(sk);
455 if (tp->window_clamp >= maxwin) {
456 tp->window_clamp = maxwin;
458 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
459 tp->window_clamp = max(maxwin -
460 (maxwin >> sysctl_tcp_app_win),
464 /* Force reservation of one segment. */
465 if (sysctl_tcp_app_win &&
466 tp->window_clamp > 2 * tp->advmss &&
467 tp->window_clamp + tp->advmss > maxwin)
468 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
470 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
471 tp->snd_cwnd_stamp = tcp_jiffies32;
474 /* 5. Recalculate window clamp after socket hit its memory bounds. */
475 static void tcp_clamp_window(struct sock *sk)
477 struct tcp_sock *tp = tcp_sk(sk);
478 struct inet_connection_sock *icsk = inet_csk(sk);
480 icsk->icsk_ack.quick = 0;
482 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
483 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
484 !tcp_under_memory_pressure(sk) &&
485 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
486 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
489 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
490 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
493 /* Initialize RCV_MSS value.
494 * RCV_MSS is an our guess about MSS used by the peer.
495 * We haven't any direct information about the MSS.
496 * It's better to underestimate the RCV_MSS rather than overestimate.
497 * Overestimations make us ACKing less frequently than needed.
498 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
500 void tcp_initialize_rcv_mss(struct sock *sk)
502 const struct tcp_sock *tp = tcp_sk(sk);
503 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
505 hint = min(hint, tp->rcv_wnd / 2);
506 hint = min(hint, TCP_MSS_DEFAULT);
507 hint = max(hint, TCP_MIN_MSS);
509 inet_csk(sk)->icsk_ack.rcv_mss = hint;
511 EXPORT_SYMBOL(tcp_initialize_rcv_mss);
513 /* Receiver "autotuning" code.
515 * The algorithm for RTT estimation w/o timestamps is based on
516 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
517 * <http://public.lanl.gov/radiant/pubs.html#DRS>
519 * More detail on this code can be found at
520 * <http://staff.psc.edu/jheffner/>,
521 * though this reference is out of date. A new paper
524 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
526 u32 new_sample = tp->rcv_rtt_est.rtt_us;
529 if (new_sample != 0) {
530 /* If we sample in larger samples in the non-timestamp
531 * case, we could grossly overestimate the RTT especially
532 * with chatty applications or bulk transfer apps which
533 * are stalled on filesystem I/O.
535 * Also, since we are only going for a minimum in the
536 * non-timestamp case, we do not smooth things out
537 * else with timestamps disabled convergence takes too
541 m -= (new_sample >> 3);
549 /* No previous measure. */
553 tp->rcv_rtt_est.rtt_us = new_sample;
556 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
560 if (tp->rcv_rtt_est.time == 0)
562 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
564 delta_us = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcv_rtt_est.time);
567 tcp_rcv_rtt_update(tp, delta_us, 1);
570 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
571 tp->rcv_rtt_est.time = tp->tcp_mstamp;
574 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
575 const struct sk_buff *skb)
577 struct tcp_sock *tp = tcp_sk(sk);
579 if (tp->rx_opt.rcv_tsecr &&
580 (TCP_SKB_CB(skb)->end_seq -
581 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss)) {
582 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr;
587 delta_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
588 tcp_rcv_rtt_update(tp, delta_us, 0);
593 * This function should be called every time data is copied to user space.
594 * It calculates the appropriate TCP receive buffer space.
596 void tcp_rcv_space_adjust(struct sock *sk)
598 struct tcp_sock *tp = tcp_sk(sk);
602 tcp_mstamp_refresh(tp);
603 time = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcvq_space.time);
604 if (time < (tp->rcv_rtt_est.rtt_us >> 3) || tp->rcv_rtt_est.rtt_us == 0)
607 /* Number of bytes copied to user in last RTT */
608 copied = tp->copied_seq - tp->rcvq_space.seq;
609 if (copied <= tp->rcvq_space.space)
613 * copied = bytes received in previous RTT, our base window
614 * To cope with packet losses, we need a 2x factor
615 * To cope with slow start, and sender growing its cwin by 100 %
616 * every RTT, we need a 4x factor, because the ACK we are sending
617 * now is for the next RTT, not the current one :
618 * <prev RTT . ><current RTT .. ><next RTT .... >
621 if (sysctl_tcp_moderate_rcvbuf &&
622 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
626 /* minimal window to cope with packet losses, assuming
627 * steady state. Add some cushion because of small variations.
629 rcvwin = ((u64)copied << 1) + 16 * tp->advmss;
631 /* If rate increased by 25%,
632 * assume slow start, rcvwin = 3 * copied
633 * If rate increased by 50%,
634 * assume sender can use 2x growth, rcvwin = 4 * copied
637 tp->rcvq_space.space + (tp->rcvq_space.space >> 2)) {
639 tp->rcvq_space.space + (tp->rcvq_space.space >> 1))
642 rcvwin += (rcvwin >> 1);
645 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
646 while (tcp_win_from_space(rcvmem) < tp->advmss)
649 do_div(rcvwin, tp->advmss);
650 rcvbuf = min_t(u64, rcvwin * rcvmem, sysctl_tcp_rmem[2]);
651 if (rcvbuf > sk->sk_rcvbuf) {
652 sk->sk_rcvbuf = rcvbuf;
654 /* Make the window clamp follow along. */
655 tp->window_clamp = tcp_win_from_space(rcvbuf);
658 tp->rcvq_space.space = copied;
661 tp->rcvq_space.seq = tp->copied_seq;
662 tp->rcvq_space.time = tp->tcp_mstamp;
665 /* There is something which you must keep in mind when you analyze the
666 * behavior of the tp->ato delayed ack timeout interval. When a
667 * connection starts up, we want to ack as quickly as possible. The
668 * problem is that "good" TCP's do slow start at the beginning of data
669 * transmission. The means that until we send the first few ACK's the
670 * sender will sit on his end and only queue most of his data, because
671 * he can only send snd_cwnd unacked packets at any given time. For
672 * each ACK we send, he increments snd_cwnd and transmits more of his
675 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
677 struct tcp_sock *tp = tcp_sk(sk);
678 struct inet_connection_sock *icsk = inet_csk(sk);
681 inet_csk_schedule_ack(sk);
683 tcp_measure_rcv_mss(sk, skb);
685 tcp_rcv_rtt_measure(tp);
689 if (!icsk->icsk_ack.ato) {
690 /* The _first_ data packet received, initialize
691 * delayed ACK engine.
693 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
694 icsk->icsk_ack.ato = TCP_ATO_MIN;
696 int m = now - icsk->icsk_ack.lrcvtime;
698 if (m <= TCP_ATO_MIN / 2) {
699 /* The fastest case is the first. */
700 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
701 } else if (m < icsk->icsk_ack.ato) {
702 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
703 if (icsk->icsk_ack.ato > icsk->icsk_rto)
704 icsk->icsk_ack.ato = icsk->icsk_rto;
705 } else if (m > icsk->icsk_rto) {
706 /* Too long gap. Apparently sender failed to
707 * restart window, so that we send ACKs quickly.
709 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
713 icsk->icsk_ack.lrcvtime = now;
715 tcp_ecn_check_ce(sk, skb);
718 tcp_grow_window(sk, skb);
721 /* Called to compute a smoothed rtt estimate. The data fed to this
722 * routine either comes from timestamps, or from segments that were
723 * known _not_ to have been retransmitted [see Karn/Partridge
724 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
725 * piece by Van Jacobson.
726 * NOTE: the next three routines used to be one big routine.
727 * To save cycles in the RFC 1323 implementation it was better to break
728 * it up into three procedures. -- erics
730 static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
732 struct tcp_sock *tp = tcp_sk(sk);
733 long m = mrtt_us; /* RTT */
734 u32 srtt = tp->srtt_us;
736 /* The following amusing code comes from Jacobson's
737 * article in SIGCOMM '88. Note that rtt and mdev
738 * are scaled versions of rtt and mean deviation.
739 * This is designed to be as fast as possible
740 * m stands for "measurement".
742 * On a 1990 paper the rto value is changed to:
743 * RTO = rtt + 4 * mdev
745 * Funny. This algorithm seems to be very broken.
746 * These formulae increase RTO, when it should be decreased, increase
747 * too slowly, when it should be increased quickly, decrease too quickly
748 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
749 * does not matter how to _calculate_ it. Seems, it was trap
750 * that VJ failed to avoid. 8)
753 m -= (srtt >> 3); /* m is now error in rtt est */
754 srtt += m; /* rtt = 7/8 rtt + 1/8 new */
756 m = -m; /* m is now abs(error) */
757 m -= (tp->mdev_us >> 2); /* similar update on mdev */
758 /* This is similar to one of Eifel findings.
759 * Eifel blocks mdev updates when rtt decreases.
760 * This solution is a bit different: we use finer gain
761 * for mdev in this case (alpha*beta).
762 * Like Eifel it also prevents growth of rto,
763 * but also it limits too fast rto decreases,
764 * happening in pure Eifel.
769 m -= (tp->mdev_us >> 2); /* similar update on mdev */
771 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */
772 if (tp->mdev_us > tp->mdev_max_us) {
773 tp->mdev_max_us = tp->mdev_us;
774 if (tp->mdev_max_us > tp->rttvar_us)
775 tp->rttvar_us = tp->mdev_max_us;
777 if (after(tp->snd_una, tp->rtt_seq)) {
778 if (tp->mdev_max_us < tp->rttvar_us)
779 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
780 tp->rtt_seq = tp->snd_nxt;
781 tp->mdev_max_us = tcp_rto_min_us(sk);
784 /* no previous measure. */
785 srtt = m << 3; /* take the measured time to be rtt */
786 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */
787 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
788 tp->mdev_max_us = tp->rttvar_us;
789 tp->rtt_seq = tp->snd_nxt;
791 tp->srtt_us = max(1U, srtt);
794 /* Set the sk_pacing_rate to allow proper sizing of TSO packets.
795 * Note: TCP stack does not yet implement pacing.
796 * FQ packet scheduler can be used to implement cheap but effective
797 * TCP pacing, to smooth the burst on large writes when packets
798 * in flight is significantly lower than cwnd (or rwin)
800 int sysctl_tcp_pacing_ss_ratio __read_mostly = 200;
801 int sysctl_tcp_pacing_ca_ratio __read_mostly = 120;
803 static void tcp_update_pacing_rate(struct sock *sk)
805 const struct tcp_sock *tp = tcp_sk(sk);
808 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
809 rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3);
811 /* current rate is (cwnd * mss) / srtt
812 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
813 * In Congestion Avoidance phase, set it to 120 % the current rate.
815 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
816 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
817 * end of slow start and should slow down.
819 if (tp->snd_cwnd < tp->snd_ssthresh / 2)
820 rate *= sysctl_tcp_pacing_ss_ratio;
822 rate *= sysctl_tcp_pacing_ca_ratio;
824 rate *= max(tp->snd_cwnd, tp->packets_out);
826 if (likely(tp->srtt_us))
827 do_div(rate, tp->srtt_us);
829 /* ACCESS_ONCE() is needed because sch_fq fetches sk_pacing_rate
830 * without any lock. We want to make sure compiler wont store
831 * intermediate values in this location.
833 ACCESS_ONCE(sk->sk_pacing_rate) = min_t(u64, rate,
834 sk->sk_max_pacing_rate);
837 /* Calculate rto without backoff. This is the second half of Van Jacobson's
838 * routine referred to above.
840 static void tcp_set_rto(struct sock *sk)
842 const struct tcp_sock *tp = tcp_sk(sk);
843 /* Old crap is replaced with new one. 8)
846 * 1. If rtt variance happened to be less 50msec, it is hallucination.
847 * It cannot be less due to utterly erratic ACK generation made
848 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
849 * to do with delayed acks, because at cwnd>2 true delack timeout
850 * is invisible. Actually, Linux-2.4 also generates erratic
851 * ACKs in some circumstances.
853 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
855 /* 2. Fixups made earlier cannot be right.
856 * If we do not estimate RTO correctly without them,
857 * all the algo is pure shit and should be replaced
858 * with correct one. It is exactly, which we pretend to do.
861 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
862 * guarantees that rto is higher.
867 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
869 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
872 cwnd = TCP_INIT_CWND;
873 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
877 * Packet counting of FACK is based on in-order assumptions, therefore TCP
878 * disables it when reordering is detected
880 void tcp_disable_fack(struct tcp_sock *tp)
882 /* RFC3517 uses different metric in lost marker => reset on change */
884 tp->lost_skb_hint = NULL;
885 tp->rx_opt.sack_ok &= ~TCP_FACK_ENABLED;
888 /* Take a notice that peer is sending D-SACKs */
889 static void tcp_dsack_seen(struct tcp_sock *tp)
891 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
894 static void tcp_update_reordering(struct sock *sk, const int metric,
897 struct tcp_sock *tp = tcp_sk(sk);
900 if (WARN_ON_ONCE(metric < 0))
903 if (metric > tp->reordering) {
904 tp->reordering = min(sysctl_tcp_max_reordering, metric);
906 #if FASTRETRANS_DEBUG > 1
907 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
908 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
912 tp->undo_marker ? tp->undo_retrans : 0);
914 tcp_disable_fack(tp);
919 /* This exciting event is worth to be remembered. 8) */
921 mib_idx = LINUX_MIB_TCPTSREORDER;
922 else if (tcp_is_reno(tp))
923 mib_idx = LINUX_MIB_TCPRENOREORDER;
924 else if (tcp_is_fack(tp))
925 mib_idx = LINUX_MIB_TCPFACKREORDER;
927 mib_idx = LINUX_MIB_TCPSACKREORDER;
929 NET_INC_STATS(sock_net(sk), mib_idx);
932 /* This must be called before lost_out is incremented */
933 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
935 if ((!tp->retransmit_skb_hint && tp->retrans_out >= tp->lost_out) ||
936 (tp->retransmit_skb_hint &&
937 before(TCP_SKB_CB(skb)->seq,
938 TCP_SKB_CB(tp->retransmit_skb_hint)->seq)))
939 tp->retransmit_skb_hint = skb;
942 /* Sum the number of packets on the wire we have marked as lost.
943 * There are two cases we care about here:
944 * a) Packet hasn't been marked lost (nor retransmitted),
945 * and this is the first loss.
946 * b) Packet has been marked both lost and retransmitted,
947 * and this means we think it was lost again.
949 static void tcp_sum_lost(struct tcp_sock *tp, struct sk_buff *skb)
951 __u8 sacked = TCP_SKB_CB(skb)->sacked;
953 if (!(sacked & TCPCB_LOST) ||
954 ((sacked & TCPCB_LOST) && (sacked & TCPCB_SACKED_RETRANS)))
955 tp->lost += tcp_skb_pcount(skb);
958 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
960 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
961 tcp_verify_retransmit_hint(tp, skb);
963 tp->lost_out += tcp_skb_pcount(skb);
964 tcp_sum_lost(tp, skb);
965 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
969 void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp, struct sk_buff *skb)
971 tcp_verify_retransmit_hint(tp, skb);
973 tcp_sum_lost(tp, skb);
974 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
975 tp->lost_out += tcp_skb_pcount(skb);
976 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
980 /* This procedure tags the retransmission queue when SACKs arrive.
982 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
983 * Packets in queue with these bits set are counted in variables
984 * sacked_out, retrans_out and lost_out, correspondingly.
986 * Valid combinations are:
987 * Tag InFlight Description
988 * 0 1 - orig segment is in flight.
989 * S 0 - nothing flies, orig reached receiver.
990 * L 0 - nothing flies, orig lost by net.
991 * R 2 - both orig and retransmit are in flight.
992 * L|R 1 - orig is lost, retransmit is in flight.
993 * S|R 1 - orig reached receiver, retrans is still in flight.
994 * (L|S|R is logically valid, it could occur when L|R is sacked,
995 * but it is equivalent to plain S and code short-curcuits it to S.
996 * L|S is logically invalid, it would mean -1 packet in flight 8))
998 * These 6 states form finite state machine, controlled by the following events:
999 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
1000 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
1001 * 3. Loss detection event of two flavors:
1002 * A. Scoreboard estimator decided the packet is lost.
1003 * A'. Reno "three dupacks" marks head of queue lost.
1004 * A''. Its FACK modification, head until snd.fack is lost.
1005 * B. SACK arrives sacking SND.NXT at the moment, when the
1006 * segment was retransmitted.
1007 * 4. D-SACK added new rule: D-SACK changes any tag to S.
1009 * It is pleasant to note, that state diagram turns out to be commutative,
1010 * so that we are allowed not to be bothered by order of our actions,
1011 * when multiple events arrive simultaneously. (see the function below).
1013 * Reordering detection.
1014 * --------------------
1015 * Reordering metric is maximal distance, which a packet can be displaced
1016 * in packet stream. With SACKs we can estimate it:
1018 * 1. SACK fills old hole and the corresponding segment was not
1019 * ever retransmitted -> reordering. Alas, we cannot use it
1020 * when segment was retransmitted.
1021 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1022 * for retransmitted and already SACKed segment -> reordering..
1023 * Both of these heuristics are not used in Loss state, when we cannot
1024 * account for retransmits accurately.
1026 * SACK block validation.
1027 * ----------------------
1029 * SACK block range validation checks that the received SACK block fits to
1030 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1031 * Note that SND.UNA is not included to the range though being valid because
1032 * it means that the receiver is rather inconsistent with itself reporting
1033 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1034 * perfectly valid, however, in light of RFC2018 which explicitly states
1035 * that "SACK block MUST reflect the newest segment. Even if the newest
1036 * segment is going to be discarded ...", not that it looks very clever
1037 * in case of head skb. Due to potentional receiver driven attacks, we
1038 * choose to avoid immediate execution of a walk in write queue due to
1039 * reneging and defer head skb's loss recovery to standard loss recovery
1040 * procedure that will eventually trigger (nothing forbids us doing this).
1042 * Implements also blockage to start_seq wrap-around. Problem lies in the
1043 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1044 * there's no guarantee that it will be before snd_nxt (n). The problem
1045 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1048 * <- outs wnd -> <- wrapzone ->
1049 * u e n u_w e_w s n_w
1051 * |<------------+------+----- TCP seqno space --------------+---------->|
1052 * ...-- <2^31 ->| |<--------...
1053 * ...---- >2^31 ------>| |<--------...
1055 * Current code wouldn't be vulnerable but it's better still to discard such
1056 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1057 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1058 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1059 * equal to the ideal case (infinite seqno space without wrap caused issues).
1061 * With D-SACK the lower bound is extended to cover sequence space below
1062 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1063 * again, D-SACK block must not to go across snd_una (for the same reason as
1064 * for the normal SACK blocks, explained above). But there all simplicity
1065 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1066 * fully below undo_marker they do not affect behavior in anyway and can
1067 * therefore be safely ignored. In rare cases (which are more or less
1068 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1069 * fragmentation and packet reordering past skb's retransmission. To consider
1070 * them correctly, the acceptable range must be extended even more though
1071 * the exact amount is rather hard to quantify. However, tp->max_window can
1072 * be used as an exaggerated estimate.
1074 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
1075 u32 start_seq, u32 end_seq)
1077 /* Too far in future, or reversed (interpretation is ambiguous) */
1078 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1081 /* Nasty start_seq wrap-around check (see comments above) */
1082 if (!before(start_seq, tp->snd_nxt))
1085 /* In outstanding window? ...This is valid exit for D-SACKs too.
1086 * start_seq == snd_una is non-sensical (see comments above)
1088 if (after(start_seq, tp->snd_una))
1091 if (!is_dsack || !tp->undo_marker)
1094 /* ...Then it's D-SACK, and must reside below snd_una completely */
1095 if (after(end_seq, tp->snd_una))
1098 if (!before(start_seq, tp->undo_marker))
1102 if (!after(end_seq, tp->undo_marker))
1105 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1106 * start_seq < undo_marker and end_seq >= undo_marker.
1108 return !before(start_seq, end_seq - tp->max_window);
1111 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1112 struct tcp_sack_block_wire *sp, int num_sacks,
1115 struct tcp_sock *tp = tcp_sk(sk);
1116 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1117 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1118 bool dup_sack = false;
1120 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1123 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1124 } else if (num_sacks > 1) {
1125 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1126 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1128 if (!after(end_seq_0, end_seq_1) &&
1129 !before(start_seq_0, start_seq_1)) {
1132 NET_INC_STATS(sock_net(sk),
1133 LINUX_MIB_TCPDSACKOFORECV);
1137 /* D-SACK for already forgotten data... Do dumb counting. */
1138 if (dup_sack && tp->undo_marker && tp->undo_retrans > 0 &&
1139 !after(end_seq_0, prior_snd_una) &&
1140 after(end_seq_0, tp->undo_marker))
1146 struct tcp_sacktag_state {
1149 /* Timestamps for earliest and latest never-retransmitted segment
1150 * that was SACKed. RTO needs the earliest RTT to stay conservative,
1151 * but congestion control should still get an accurate delay signal.
1155 struct rate_sample *rate;
1159 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1160 * the incoming SACK may not exactly match but we can find smaller MSS
1161 * aligned portion of it that matches. Therefore we might need to fragment
1162 * which may fail and creates some hassle (caller must handle error case
1165 * FIXME: this could be merged to shift decision code
1167 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1168 u32 start_seq, u32 end_seq)
1172 unsigned int pkt_len;
1175 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1176 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1178 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1179 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1180 mss = tcp_skb_mss(skb);
1181 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1184 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1188 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1193 /* Round if necessary so that SACKs cover only full MSSes
1194 * and/or the remaining small portion (if present)
1196 if (pkt_len > mss) {
1197 unsigned int new_len = (pkt_len / mss) * mss;
1198 if (!in_sack && new_len < pkt_len)
1203 if (pkt_len >= skb->len && !in_sack)
1206 err = tcp_fragment(sk, skb, pkt_len, mss, GFP_ATOMIC);
1214 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1215 static u8 tcp_sacktag_one(struct sock *sk,
1216 struct tcp_sacktag_state *state, u8 sacked,
1217 u32 start_seq, u32 end_seq,
1218 int dup_sack, int pcount,
1221 struct tcp_sock *tp = tcp_sk(sk);
1222 int fack_count = state->fack_count;
1224 /* Account D-SACK for retransmitted packet. */
1225 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1226 if (tp->undo_marker && tp->undo_retrans > 0 &&
1227 after(end_seq, tp->undo_marker))
1228 tp->undo_retrans = max_t(int, 0, tp->undo_retrans - pcount);
1229 if (sacked & TCPCB_SACKED_ACKED)
1230 state->reord = min(fack_count, state->reord);
1233 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1234 if (!after(end_seq, tp->snd_una))
1237 if (!(sacked & TCPCB_SACKED_ACKED)) {
1238 tcp_rack_advance(tp, sacked, end_seq, xmit_time);
1240 if (sacked & TCPCB_SACKED_RETRANS) {
1241 /* If the segment is not tagged as lost,
1242 * we do not clear RETRANS, believing
1243 * that retransmission is still in flight.
1245 if (sacked & TCPCB_LOST) {
1246 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1247 tp->lost_out -= pcount;
1248 tp->retrans_out -= pcount;
1251 if (!(sacked & TCPCB_RETRANS)) {
1252 /* New sack for not retransmitted frame,
1253 * which was in hole. It is reordering.
1255 if (before(start_seq,
1256 tcp_highest_sack_seq(tp)))
1257 state->reord = min(fack_count,
1259 if (!after(end_seq, tp->high_seq))
1260 state->flag |= FLAG_ORIG_SACK_ACKED;
1261 if (state->first_sackt == 0)
1262 state->first_sackt = xmit_time;
1263 state->last_sackt = xmit_time;
1266 if (sacked & TCPCB_LOST) {
1267 sacked &= ~TCPCB_LOST;
1268 tp->lost_out -= pcount;
1272 sacked |= TCPCB_SACKED_ACKED;
1273 state->flag |= FLAG_DATA_SACKED;
1274 tp->sacked_out += pcount;
1275 tp->delivered += pcount; /* Out-of-order packets delivered */
1277 fack_count += pcount;
1279 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1280 if (!tcp_is_fack(tp) && tp->lost_skb_hint &&
1281 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1282 tp->lost_cnt_hint += pcount;
1284 if (fack_count > tp->fackets_out)
1285 tp->fackets_out = fack_count;
1288 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1289 * frames and clear it. undo_retrans is decreased above, L|R frames
1290 * are accounted above as well.
1292 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1293 sacked &= ~TCPCB_SACKED_RETRANS;
1294 tp->retrans_out -= pcount;
1300 /* Shift newly-SACKed bytes from this skb to the immediately previous
1301 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1303 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
1304 struct tcp_sacktag_state *state,
1305 unsigned int pcount, int shifted, int mss,
1308 struct tcp_sock *tp = tcp_sk(sk);
1309 struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
1310 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1311 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
1315 /* Adjust counters and hints for the newly sacked sequence
1316 * range but discard the return value since prev is already
1317 * marked. We must tag the range first because the seq
1318 * advancement below implicitly advances
1319 * tcp_highest_sack_seq() when skb is highest_sack.
1321 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1322 start_seq, end_seq, dup_sack, pcount,
1324 tcp_rate_skb_delivered(sk, skb, state->rate);
1326 if (skb == tp->lost_skb_hint)
1327 tp->lost_cnt_hint += pcount;
1329 TCP_SKB_CB(prev)->end_seq += shifted;
1330 TCP_SKB_CB(skb)->seq += shifted;
1332 tcp_skb_pcount_add(prev, pcount);
1333 WARN_ON_ONCE(tcp_skb_pcount(skb) < pcount);
1334 tcp_skb_pcount_add(skb, -pcount);
1336 /* When we're adding to gso_segs == 1, gso_size will be zero,
1337 * in theory this shouldn't be necessary but as long as DSACK
1338 * code can come after this skb later on it's better to keep
1339 * setting gso_size to something.
1341 if (!TCP_SKB_CB(prev)->tcp_gso_size)
1342 TCP_SKB_CB(prev)->tcp_gso_size = mss;
1344 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1345 if (tcp_skb_pcount(skb) <= 1)
1346 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1348 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1349 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1352 BUG_ON(!tcp_skb_pcount(skb));
1353 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1357 /* Whole SKB was eaten :-) */
1359 if (skb == tp->retransmit_skb_hint)
1360 tp->retransmit_skb_hint = prev;
1361 if (skb == tp->lost_skb_hint) {
1362 tp->lost_skb_hint = prev;
1363 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1366 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1367 TCP_SKB_CB(prev)->eor = TCP_SKB_CB(skb)->eor;
1368 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1369 TCP_SKB_CB(prev)->end_seq++;
1371 if (skb == tcp_highest_sack(sk))
1372 tcp_advance_highest_sack(sk, skb);
1374 tcp_skb_collapse_tstamp(prev, skb);
1375 if (unlikely(TCP_SKB_CB(prev)->tx.delivered_mstamp))
1376 TCP_SKB_CB(prev)->tx.delivered_mstamp = 0;
1378 tcp_unlink_write_queue(skb, sk);
1379 sk_wmem_free_skb(sk, skb);
1381 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKMERGED);
1386 /* I wish gso_size would have a bit more sane initialization than
1387 * something-or-zero which complicates things
1389 static int tcp_skb_seglen(const struct sk_buff *skb)
1391 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1394 /* Shifting pages past head area doesn't work */
1395 static int skb_can_shift(const struct sk_buff *skb)
1397 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1400 int tcp_skb_shift(struct sk_buff *to, struct sk_buff *from,
1401 int pcount, int shiftlen)
1403 /* TCP min gso_size is 8 bytes (TCP_MIN_GSO_SIZE)
1404 * Since TCP_SKB_CB(skb)->tcp_gso_segs is 16 bits, we need
1405 * to make sure not storing more than 65535 * 8 bytes per skb,
1406 * even if current MSS is bigger.
1408 if (unlikely(to->len + shiftlen >= 65535 * TCP_MIN_GSO_SIZE))
1410 if (unlikely(tcp_skb_pcount(to) + pcount > 65535))
1412 return skb_shift(to, from, shiftlen);
1415 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1418 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1419 struct tcp_sacktag_state *state,
1420 u32 start_seq, u32 end_seq,
1423 struct tcp_sock *tp = tcp_sk(sk);
1424 struct sk_buff *prev;
1431 if (!sk_can_gso(sk))
1434 /* Normally R but no L won't result in plain S */
1436 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1438 if (!skb_can_shift(skb))
1440 /* This frame is about to be dropped (was ACKed). */
1441 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1444 /* Can only happen with delayed DSACK + discard craziness */
1445 if (unlikely(skb == tcp_write_queue_head(sk)))
1447 prev = tcp_write_queue_prev(sk, skb);
1449 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1452 if (!tcp_skb_can_collapse_to(prev))
1455 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1456 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1460 pcount = tcp_skb_pcount(skb);
1461 mss = tcp_skb_seglen(skb);
1463 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1464 * drop this restriction as unnecessary
1466 if (mss != tcp_skb_seglen(prev))
1469 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1471 /* CHECKME: This is non-MSS split case only?, this will
1472 * cause skipped skbs due to advancing loop btw, original
1473 * has that feature too
1475 if (tcp_skb_pcount(skb) <= 1)
1478 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1480 /* TODO: head merge to next could be attempted here
1481 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1482 * though it might not be worth of the additional hassle
1484 * ...we can probably just fallback to what was done
1485 * previously. We could try merging non-SACKed ones
1486 * as well but it probably isn't going to buy off
1487 * because later SACKs might again split them, and
1488 * it would make skb timestamp tracking considerably
1494 len = end_seq - TCP_SKB_CB(skb)->seq;
1496 BUG_ON(len > skb->len);
1498 /* MSS boundaries should be honoured or else pcount will
1499 * severely break even though it makes things bit trickier.
1500 * Optimize common case to avoid most of the divides
1502 mss = tcp_skb_mss(skb);
1504 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1505 * drop this restriction as unnecessary
1507 if (mss != tcp_skb_seglen(prev))
1512 } else if (len < mss) {
1520 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1521 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1524 if (!tcp_skb_shift(prev, skb, pcount, len))
1526 if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack))
1529 /* Hole filled allows collapsing with the next as well, this is very
1530 * useful when hole on every nth skb pattern happens
1532 if (prev == tcp_write_queue_tail(sk))
1534 skb = tcp_write_queue_next(sk, prev);
1536 if (!skb_can_shift(skb) ||
1537 (skb == tcp_send_head(sk)) ||
1538 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1539 (mss != tcp_skb_seglen(skb)))
1543 next_pcount = tcp_skb_pcount(skb);
1544 if (tcp_skb_shift(prev, skb, next_pcount, len)) {
1545 pcount += next_pcount;
1546 tcp_shifted_skb(sk, skb, state, next_pcount, len, mss, 0);
1549 state->fack_count += pcount;
1556 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1560 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1561 struct tcp_sack_block *next_dup,
1562 struct tcp_sacktag_state *state,
1563 u32 start_seq, u32 end_seq,
1566 struct tcp_sock *tp = tcp_sk(sk);
1567 struct sk_buff *tmp;
1569 tcp_for_write_queue_from(skb, sk) {
1571 bool dup_sack = dup_sack_in;
1573 if (skb == tcp_send_head(sk))
1576 /* queue is in-order => we can short-circuit the walk early */
1577 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1581 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1582 in_sack = tcp_match_skb_to_sack(sk, skb,
1583 next_dup->start_seq,
1589 /* skb reference here is a bit tricky to get right, since
1590 * shifting can eat and free both this skb and the next,
1591 * so not even _safe variant of the loop is enough.
1594 tmp = tcp_shift_skb_data(sk, skb, state,
1595 start_seq, end_seq, dup_sack);
1604 in_sack = tcp_match_skb_to_sack(sk, skb,
1610 if (unlikely(in_sack < 0))
1614 TCP_SKB_CB(skb)->sacked =
1617 TCP_SKB_CB(skb)->sacked,
1618 TCP_SKB_CB(skb)->seq,
1619 TCP_SKB_CB(skb)->end_seq,
1621 tcp_skb_pcount(skb),
1623 tcp_rate_skb_delivered(sk, skb, state->rate);
1625 if (!before(TCP_SKB_CB(skb)->seq,
1626 tcp_highest_sack_seq(tp)))
1627 tcp_advance_highest_sack(sk, skb);
1630 state->fack_count += tcp_skb_pcount(skb);
1635 /* Avoid all extra work that is being done by sacktag while walking in
1638 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1639 struct tcp_sacktag_state *state,
1642 tcp_for_write_queue_from(skb, sk) {
1643 if (skb == tcp_send_head(sk))
1646 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1649 state->fack_count += tcp_skb_pcount(skb);
1654 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1656 struct tcp_sack_block *next_dup,
1657 struct tcp_sacktag_state *state,
1663 if (before(next_dup->start_seq, skip_to_seq)) {
1664 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1665 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1666 next_dup->start_seq, next_dup->end_seq,
1673 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1675 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1679 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1680 u32 prior_snd_una, struct tcp_sacktag_state *state)
1682 struct tcp_sock *tp = tcp_sk(sk);
1683 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1684 TCP_SKB_CB(ack_skb)->sacked);
1685 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1686 struct tcp_sack_block sp[TCP_NUM_SACKS];
1687 struct tcp_sack_block *cache;
1688 struct sk_buff *skb;
1689 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1691 bool found_dup_sack = false;
1693 int first_sack_index;
1696 state->reord = tp->packets_out;
1698 if (!tp->sacked_out) {
1699 if (WARN_ON(tp->fackets_out))
1700 tp->fackets_out = 0;
1701 tcp_highest_sack_reset(sk);
1704 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1705 num_sacks, prior_snd_una);
1706 if (found_dup_sack) {
1707 state->flag |= FLAG_DSACKING_ACK;
1708 tp->delivered++; /* A spurious retransmission is delivered */
1711 /* Eliminate too old ACKs, but take into
1712 * account more or less fresh ones, they can
1713 * contain valid SACK info.
1715 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1718 if (!tp->packets_out)
1722 first_sack_index = 0;
1723 for (i = 0; i < num_sacks; i++) {
1724 bool dup_sack = !i && found_dup_sack;
1726 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1727 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1729 if (!tcp_is_sackblock_valid(tp, dup_sack,
1730 sp[used_sacks].start_seq,
1731 sp[used_sacks].end_seq)) {
1735 if (!tp->undo_marker)
1736 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1738 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1740 /* Don't count olds caused by ACK reordering */
1741 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1742 !after(sp[used_sacks].end_seq, tp->snd_una))
1744 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1747 NET_INC_STATS(sock_net(sk), mib_idx);
1749 first_sack_index = -1;
1753 /* Ignore very old stuff early */
1754 if (!after(sp[used_sacks].end_seq, prior_snd_una)) {
1756 first_sack_index = -1;
1763 /* order SACK blocks to allow in order walk of the retrans queue */
1764 for (i = used_sacks - 1; i > 0; i--) {
1765 for (j = 0; j < i; j++) {
1766 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1767 swap(sp[j], sp[j + 1]);
1769 /* Track where the first SACK block goes to */
1770 if (j == first_sack_index)
1771 first_sack_index = j + 1;
1776 skb = tcp_write_queue_head(sk);
1777 state->fack_count = 0;
1780 if (!tp->sacked_out) {
1781 /* It's already past, so skip checking against it */
1782 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1784 cache = tp->recv_sack_cache;
1785 /* Skip empty blocks in at head of the cache */
1786 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1791 while (i < used_sacks) {
1792 u32 start_seq = sp[i].start_seq;
1793 u32 end_seq = sp[i].end_seq;
1794 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1795 struct tcp_sack_block *next_dup = NULL;
1797 if (found_dup_sack && ((i + 1) == first_sack_index))
1798 next_dup = &sp[i + 1];
1800 /* Skip too early cached blocks */
1801 while (tcp_sack_cache_ok(tp, cache) &&
1802 !before(start_seq, cache->end_seq))
1805 /* Can skip some work by looking recv_sack_cache? */
1806 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1807 after(end_seq, cache->start_seq)) {
1810 if (before(start_seq, cache->start_seq)) {
1811 skb = tcp_sacktag_skip(skb, sk, state,
1813 skb = tcp_sacktag_walk(skb, sk, next_dup,
1820 /* Rest of the block already fully processed? */
1821 if (!after(end_seq, cache->end_seq))
1824 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1828 /* ...tail remains todo... */
1829 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1830 /* ...but better entrypoint exists! */
1831 skb = tcp_highest_sack(sk);
1834 state->fack_count = tp->fackets_out;
1839 skb = tcp_sacktag_skip(skb, sk, state, cache->end_seq);
1840 /* Check overlap against next cached too (past this one already) */
1845 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1846 skb = tcp_highest_sack(sk);
1849 state->fack_count = tp->fackets_out;
1851 skb = tcp_sacktag_skip(skb, sk, state, start_seq);
1854 skb = tcp_sacktag_walk(skb, sk, next_dup, state,
1855 start_seq, end_seq, dup_sack);
1861 /* Clear the head of the cache sack blocks so we can skip it next time */
1862 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1863 tp->recv_sack_cache[i].start_seq = 0;
1864 tp->recv_sack_cache[i].end_seq = 0;
1866 for (j = 0; j < used_sacks; j++)
1867 tp->recv_sack_cache[i++] = sp[j];
1869 if ((state->reord < tp->fackets_out) &&
1870 ((inet_csk(sk)->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker))
1871 tcp_update_reordering(sk, tp->fackets_out - state->reord, 0);
1873 tcp_verify_left_out(tp);
1876 #if FASTRETRANS_DEBUG > 0
1877 WARN_ON((int)tp->sacked_out < 0);
1878 WARN_ON((int)tp->lost_out < 0);
1879 WARN_ON((int)tp->retrans_out < 0);
1880 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1885 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1886 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1888 static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
1892 holes = max(tp->lost_out, 1U);
1893 holes = min(holes, tp->packets_out);
1895 if ((tp->sacked_out + holes) > tp->packets_out) {
1896 tp->sacked_out = tp->packets_out - holes;
1902 /* If we receive more dupacks than we expected counting segments
1903 * in assumption of absent reordering, interpret this as reordering.
1904 * The only another reason could be bug in receiver TCP.
1906 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1908 struct tcp_sock *tp = tcp_sk(sk);
1909 if (tcp_limit_reno_sacked(tp))
1910 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1913 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1915 static void tcp_add_reno_sack(struct sock *sk)
1917 struct tcp_sock *tp = tcp_sk(sk);
1918 u32 prior_sacked = tp->sacked_out;
1921 tcp_check_reno_reordering(sk, 0);
1922 if (tp->sacked_out > prior_sacked)
1923 tp->delivered++; /* Some out-of-order packet is delivered */
1924 tcp_verify_left_out(tp);
1927 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1929 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1931 struct tcp_sock *tp = tcp_sk(sk);
1934 /* One ACK acked hole. The rest eat duplicate ACKs. */
1935 tp->delivered += max_t(int, acked - tp->sacked_out, 1);
1936 if (acked - 1 >= tp->sacked_out)
1939 tp->sacked_out -= acked - 1;
1941 tcp_check_reno_reordering(sk, acked);
1942 tcp_verify_left_out(tp);
1945 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1950 void tcp_clear_retrans(struct tcp_sock *tp)
1952 tp->retrans_out = 0;
1954 tp->undo_marker = 0;
1955 tp->undo_retrans = -1;
1956 tp->fackets_out = 0;
1960 static inline void tcp_init_undo(struct tcp_sock *tp)
1962 tp->undo_marker = tp->snd_una;
1963 /* Retransmission still in flight may cause DSACKs later. */
1964 tp->undo_retrans = tp->retrans_out ? : -1;
1967 /* Enter Loss state. If we detect SACK reneging, forget all SACK information
1968 * and reset tags completely, otherwise preserve SACKs. If receiver
1969 * dropped its ofo queue, we will know this due to reneging detection.
1971 void tcp_enter_loss(struct sock *sk)
1973 const struct inet_connection_sock *icsk = inet_csk(sk);
1974 struct tcp_sock *tp = tcp_sk(sk);
1975 struct net *net = sock_net(sk);
1976 struct sk_buff *skb;
1977 bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery;
1978 bool is_reneg; /* is receiver reneging on SACKs? */
1981 /* Reduce ssthresh if it has not yet been made inside this window. */
1982 if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
1983 !after(tp->high_seq, tp->snd_una) ||
1984 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1985 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1986 tp->prior_cwnd = tp->snd_cwnd;
1987 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1988 tcp_ca_event(sk, CA_EVENT_LOSS);
1992 tp->snd_cwnd_cnt = 0;
1993 tp->snd_cwnd_stamp = tcp_jiffies32;
1995 tp->retrans_out = 0;
1998 if (tcp_is_reno(tp))
1999 tcp_reset_reno_sack(tp);
2001 skb = tcp_write_queue_head(sk);
2002 is_reneg = skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED);
2004 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
2006 tp->fackets_out = 0;
2007 /* Mark SACK reneging until we recover from this loss event. */
2008 tp->is_sack_reneg = 1;
2010 tcp_clear_all_retrans_hints(tp);
2012 tcp_for_write_queue(skb, sk) {
2013 if (skb == tcp_send_head(sk))
2016 mark_lost = (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
2019 tcp_sum_lost(tp, skb);
2020 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
2022 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
2023 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
2024 tp->lost_out += tcp_skb_pcount(skb);
2027 tcp_verify_left_out(tp);
2029 /* Timeout in disordered state after receiving substantial DUPACKs
2030 * suggests that the degree of reordering is over-estimated.
2032 if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
2033 tp->sacked_out >= net->ipv4.sysctl_tcp_reordering)
2034 tp->reordering = min_t(unsigned int, tp->reordering,
2035 net->ipv4.sysctl_tcp_reordering);
2036 tcp_set_ca_state(sk, TCP_CA_Loss);
2037 tp->high_seq = tp->snd_nxt;
2038 tcp_ecn_queue_cwr(tp);
2040 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
2041 * loss recovery is underway except recurring timeout(s) on
2042 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
2044 tp->frto = sysctl_tcp_frto &&
2045 (new_recovery || icsk->icsk_retransmits) &&
2046 !inet_csk(sk)->icsk_mtup.probe_size;
2049 /* If ACK arrived pointing to a remembered SACK, it means that our
2050 * remembered SACKs do not reflect real state of receiver i.e.
2051 * receiver _host_ is heavily congested (or buggy).
2053 * To avoid big spurious retransmission bursts due to transient SACK
2054 * scoreboard oddities that look like reneging, we give the receiver a
2055 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
2056 * restore sanity to the SACK scoreboard. If the apparent reneging
2057 * persists until this RTO then we'll clear the SACK scoreboard.
2059 static bool tcp_check_sack_reneging(struct sock *sk, int flag)
2061 if (flag & FLAG_SACK_RENEGING &&
2062 flag & FLAG_SND_UNA_ADVANCED) {
2063 struct tcp_sock *tp = tcp_sk(sk);
2064 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4),
2065 msecs_to_jiffies(10));
2067 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2068 delay, TCP_RTO_MAX);
2074 static inline int tcp_fackets_out(const struct tcp_sock *tp)
2076 return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
2079 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2080 * counter when SACK is enabled (without SACK, sacked_out is used for
2083 * Instead, with FACK TCP uses fackets_out that includes both SACKed
2084 * segments up to the highest received SACK block so far and holes in
2087 * With reordering, holes may still be in flight, so RFC3517 recovery
2088 * uses pure sacked_out (total number of SACKed segments) even though
2089 * it violates the RFC that uses duplicate ACKs, often these are equal
2090 * but when e.g. out-of-window ACKs or packet duplication occurs,
2091 * they differ. Since neither occurs due to loss, TCP should really
2094 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2096 return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
2099 /* Linux NewReno/SACK/FACK/ECN state machine.
2100 * --------------------------------------
2102 * "Open" Normal state, no dubious events, fast path.
2103 * "Disorder" In all the respects it is "Open",
2104 * but requires a bit more attention. It is entered when
2105 * we see some SACKs or dupacks. It is split of "Open"
2106 * mainly to move some processing from fast path to slow one.
2107 * "CWR" CWND was reduced due to some Congestion Notification event.
2108 * It can be ECN, ICMP source quench, local device congestion.
2109 * "Recovery" CWND was reduced, we are fast-retransmitting.
2110 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2112 * tcp_fastretrans_alert() is entered:
2113 * - each incoming ACK, if state is not "Open"
2114 * - when arrived ACK is unusual, namely:
2119 * Counting packets in flight is pretty simple.
2121 * in_flight = packets_out - left_out + retrans_out
2123 * packets_out is SND.NXT-SND.UNA counted in packets.
2125 * retrans_out is number of retransmitted segments.
2127 * left_out is number of segments left network, but not ACKed yet.
2129 * left_out = sacked_out + lost_out
2131 * sacked_out: Packets, which arrived to receiver out of order
2132 * and hence not ACKed. With SACKs this number is simply
2133 * amount of SACKed data. Even without SACKs
2134 * it is easy to give pretty reliable estimate of this number,
2135 * counting duplicate ACKs.
2137 * lost_out: Packets lost by network. TCP has no explicit
2138 * "loss notification" feedback from network (for now).
2139 * It means that this number can be only _guessed_.
2140 * Actually, it is the heuristics to predict lossage that
2141 * distinguishes different algorithms.
2143 * F.e. after RTO, when all the queue is considered as lost,
2144 * lost_out = packets_out and in_flight = retrans_out.
2146 * Essentially, we have now a few algorithms detecting
2149 * If the receiver supports SACK:
2151 * RFC6675/3517: It is the conventional algorithm. A packet is
2152 * considered lost if the number of higher sequence packets
2153 * SACKed is greater than or equal the DUPACK thoreshold
2154 * (reordering). This is implemented in tcp_mark_head_lost and
2155 * tcp_update_scoreboard.
2157 * RACK (draft-ietf-tcpm-rack-01): it is a newer algorithm
2158 * (2017-) that checks timing instead of counting DUPACKs.
2159 * Essentially a packet is considered lost if it's not S/ACKed
2160 * after RTT + reordering_window, where both metrics are
2161 * dynamically measured and adjusted. This is implemented in
2162 * tcp_rack_mark_lost.
2164 * FACK (Disabled by default. Subsumbed by RACK):
2165 * It is the simplest heuristics. As soon as we decided
2166 * that something is lost, we decide that _all_ not SACKed
2167 * packets until the most forward SACK are lost. I.e.
2168 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2169 * It is absolutely correct estimate, if network does not reorder
2170 * packets. And it loses any connection to reality when reordering
2171 * takes place. We use FACK by default until reordering
2172 * is suspected on the path to this destination.
2174 * If the receiver does not support SACK:
2176 * NewReno (RFC6582): in Recovery we assume that one segment
2177 * is lost (classic Reno). While we are in Recovery and
2178 * a partial ACK arrives, we assume that one more packet
2179 * is lost (NewReno). This heuristics are the same in NewReno
2182 * Really tricky (and requiring careful tuning) part of algorithm
2183 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2184 * The first determines the moment _when_ we should reduce CWND and,
2185 * hence, slow down forward transmission. In fact, it determines the moment
2186 * when we decide that hole is caused by loss, rather than by a reorder.
2188 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2189 * holes, caused by lost packets.
2191 * And the most logically complicated part of algorithm is undo
2192 * heuristics. We detect false retransmits due to both too early
2193 * fast retransmit (reordering) and underestimated RTO, analyzing
2194 * timestamps and D-SACKs. When we detect that some segments were
2195 * retransmitted by mistake and CWND reduction was wrong, we undo
2196 * window reduction and abort recovery phase. This logic is hidden
2197 * inside several functions named tcp_try_undo_<something>.
2200 /* This function decides, when we should leave Disordered state
2201 * and enter Recovery phase, reducing congestion window.
2203 * Main question: may we further continue forward transmission
2204 * with the same cwnd?
2206 static bool tcp_time_to_recover(struct sock *sk, int flag)
2208 struct tcp_sock *tp = tcp_sk(sk);
2210 /* Trick#1: The loss is proven. */
2214 /* Not-A-Trick#2 : Classic rule... */
2215 if (tcp_dupack_heuristics(tp) > tp->reordering)
2221 /* Detect loss in event "A" above by marking head of queue up as lost.
2222 * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2223 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2224 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2225 * the maximum SACKed segments to pass before reaching this limit.
2227 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2229 struct tcp_sock *tp = tcp_sk(sk);
2230 struct sk_buff *skb;
2231 int cnt, oldcnt, lost;
2233 /* Use SACK to deduce losses of new sequences sent during recovery */
2234 const u32 loss_high = tcp_is_sack(tp) ? tp->snd_nxt : tp->high_seq;
2236 WARN_ON(packets > tp->packets_out);
2237 if (tp->lost_skb_hint) {
2238 skb = tp->lost_skb_hint;
2239 cnt = tp->lost_cnt_hint;
2240 /* Head already handled? */
2241 if (mark_head && skb != tcp_write_queue_head(sk))
2244 skb = tcp_write_queue_head(sk);
2248 tcp_for_write_queue_from(skb, sk) {
2249 if (skb == tcp_send_head(sk))
2251 /* TODO: do this better */
2252 /* this is not the most efficient way to do this... */
2253 tp->lost_skb_hint = skb;
2254 tp->lost_cnt_hint = cnt;
2256 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2260 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2261 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2262 cnt += tcp_skb_pcount(skb);
2264 if (cnt > packets) {
2265 if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) ||
2266 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
2267 (oldcnt >= packets))
2270 mss = tcp_skb_mss(skb);
2271 /* If needed, chop off the prefix to mark as lost. */
2272 lost = (packets - oldcnt) * mss;
2273 if (lost < skb->len &&
2274 tcp_fragment(sk, skb, lost, mss, GFP_ATOMIC) < 0)
2279 tcp_skb_mark_lost(tp, skb);
2284 tcp_verify_left_out(tp);
2287 /* Account newly detected lost packet(s) */
2289 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2291 struct tcp_sock *tp = tcp_sk(sk);
2293 if (tcp_is_reno(tp)) {
2294 tcp_mark_head_lost(sk, 1, 1);
2295 } else if (tcp_is_fack(tp)) {
2296 int lost = tp->fackets_out - tp->reordering;
2299 tcp_mark_head_lost(sk, lost, 0);
2301 int sacked_upto = tp->sacked_out - tp->reordering;
2302 if (sacked_upto >= 0)
2303 tcp_mark_head_lost(sk, sacked_upto, 0);
2304 else if (fast_rexmit)
2305 tcp_mark_head_lost(sk, 1, 1);
2309 static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when)
2311 return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2312 before(tp->rx_opt.rcv_tsecr, when);
2315 /* skb is spurious retransmitted if the returned timestamp echo
2316 * reply is prior to the skb transmission time
2318 static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp,
2319 const struct sk_buff *skb)
2321 return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) &&
2322 tcp_tsopt_ecr_before(tp, tcp_skb_timestamp(skb));
2325 /* Nothing was retransmitted or returned timestamp is less
2326 * than timestamp of the first retransmission.
2328 static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2330 return !tp->retrans_stamp ||
2331 tcp_tsopt_ecr_before(tp, tp->retrans_stamp);
2334 /* Undo procedures. */
2336 /* We can clear retrans_stamp when there are no retransmissions in the
2337 * window. It would seem that it is trivially available for us in
2338 * tp->retrans_out, however, that kind of assumptions doesn't consider
2339 * what will happen if errors occur when sending retransmission for the
2340 * second time. ...It could the that such segment has only
2341 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2342 * the head skb is enough except for some reneging corner cases that
2343 * are not worth the effort.
2345 * Main reason for all this complexity is the fact that connection dying
2346 * time now depends on the validity of the retrans_stamp, in particular,
2347 * that successive retransmissions of a segment must not advance
2348 * retrans_stamp under any conditions.
2350 static bool tcp_any_retrans_done(const struct sock *sk)
2352 const struct tcp_sock *tp = tcp_sk(sk);
2353 struct sk_buff *skb;
2355 if (tp->retrans_out)
2358 skb = tcp_write_queue_head(sk);
2359 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2365 #if FASTRETRANS_DEBUG > 1
2366 static void DBGUNDO(struct sock *sk, const char *msg)
2368 struct tcp_sock *tp = tcp_sk(sk);
2369 struct inet_sock *inet = inet_sk(sk);
2371 if (sk->sk_family == AF_INET) {
2372 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2374 &inet->inet_daddr, ntohs(inet->inet_dport),
2375 tp->snd_cwnd, tcp_left_out(tp),
2376 tp->snd_ssthresh, tp->prior_ssthresh,
2379 #if IS_ENABLED(CONFIG_IPV6)
2380 else if (sk->sk_family == AF_INET6) {
2381 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2383 &sk->sk_v6_daddr, ntohs(inet->inet_dport),
2384 tp->snd_cwnd, tcp_left_out(tp),
2385 tp->snd_ssthresh, tp->prior_ssthresh,
2391 #define DBGUNDO(x...) do { } while (0)
2394 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
2396 struct tcp_sock *tp = tcp_sk(sk);
2399 struct sk_buff *skb;
2401 tcp_for_write_queue(skb, sk) {
2402 if (skb == tcp_send_head(sk))
2404 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2407 tcp_clear_all_retrans_hints(tp);
2410 if (tp->prior_ssthresh) {
2411 const struct inet_connection_sock *icsk = inet_csk(sk);
2413 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2415 if (tp->prior_ssthresh > tp->snd_ssthresh) {
2416 tp->snd_ssthresh = tp->prior_ssthresh;
2417 tcp_ecn_withdraw_cwr(tp);
2420 tp->snd_cwnd_stamp = tcp_jiffies32;
2421 tp->undo_marker = 0;
2424 static inline bool tcp_may_undo(const struct tcp_sock *tp)
2426 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2429 static bool tcp_is_non_sack_preventing_reopen(struct sock *sk)
2431 struct tcp_sock *tp = tcp_sk(sk);
2433 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2434 /* Hold old state until something *above* high_seq
2435 * is ACKed. For Reno it is MUST to prevent false
2436 * fast retransmits (RFC2582). SACK TCP is safe. */
2437 if (!tcp_any_retrans_done(sk))
2438 tp->retrans_stamp = 0;
2444 /* People celebrate: "We love our President!" */
2445 static bool tcp_try_undo_recovery(struct sock *sk)
2447 struct tcp_sock *tp = tcp_sk(sk);
2449 if (tcp_may_undo(tp)) {
2452 /* Happy end! We did not retransmit anything
2453 * or our original transmission succeeded.
2455 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2456 tcp_undo_cwnd_reduction(sk, false);
2457 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2458 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2460 mib_idx = LINUX_MIB_TCPFULLUNDO;
2462 NET_INC_STATS(sock_net(sk), mib_idx);
2464 if (tcp_is_non_sack_preventing_reopen(sk))
2466 tcp_set_ca_state(sk, TCP_CA_Open);
2467 tp->is_sack_reneg = 0;
2471 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2472 static bool tcp_try_undo_dsack(struct sock *sk)
2474 struct tcp_sock *tp = tcp_sk(sk);
2476 if (tp->undo_marker && !tp->undo_retrans) {
2477 DBGUNDO(sk, "D-SACK");
2478 tcp_undo_cwnd_reduction(sk, false);
2479 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2485 /* Undo during loss recovery after partial ACK or using F-RTO. */
2486 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2488 struct tcp_sock *tp = tcp_sk(sk);
2490 if (frto_undo || tcp_may_undo(tp)) {
2491 tcp_undo_cwnd_reduction(sk, true);
2493 DBGUNDO(sk, "partial loss");
2494 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2496 NET_INC_STATS(sock_net(sk),
2497 LINUX_MIB_TCPSPURIOUSRTOS);
2498 inet_csk(sk)->icsk_retransmits = 0;
2499 if (tcp_is_non_sack_preventing_reopen(sk))
2501 if (frto_undo || tcp_is_sack(tp)) {
2502 tcp_set_ca_state(sk, TCP_CA_Open);
2503 tp->is_sack_reneg = 0;
2510 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
2511 * It computes the number of packets to send (sndcnt) based on packets newly
2513 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2514 * cwnd reductions across a full RTT.
2515 * 2) Otherwise PRR uses packet conservation to send as much as delivered.
2516 * But when the retransmits are acked without further losses, PRR
2517 * slow starts cwnd up to ssthresh to speed up the recovery.
2519 static void tcp_init_cwnd_reduction(struct sock *sk)
2521 struct tcp_sock *tp = tcp_sk(sk);
2523 tp->high_seq = tp->snd_nxt;
2524 tp->tlp_high_seq = 0;
2525 tp->snd_cwnd_cnt = 0;
2526 tp->prior_cwnd = tp->snd_cwnd;
2527 tp->prr_delivered = 0;
2529 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2530 tcp_ecn_queue_cwr(tp);
2533 void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked, int flag)
2535 struct tcp_sock *tp = tcp_sk(sk);
2537 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2539 if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd))
2542 tp->prr_delivered += newly_acked_sacked;
2544 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2546 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2547 } else if ((flag & FLAG_RETRANS_DATA_ACKED) &&
2548 !(flag & FLAG_LOST_RETRANS)) {
2549 sndcnt = min_t(int, delta,
2550 max_t(int, tp->prr_delivered - tp->prr_out,
2551 newly_acked_sacked) + 1);
2553 sndcnt = min(delta, newly_acked_sacked);
2555 /* Force a fast retransmit upon entering fast recovery */
2556 sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1));
2557 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2560 static inline void tcp_end_cwnd_reduction(struct sock *sk)
2562 struct tcp_sock *tp = tcp_sk(sk);
2564 if (inet_csk(sk)->icsk_ca_ops->cong_control)
2567 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2568 if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH &&
2569 (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR || tp->undo_marker)) {
2570 tp->snd_cwnd = tp->snd_ssthresh;
2571 tp->snd_cwnd_stamp = tcp_jiffies32;
2573 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2576 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2577 void tcp_enter_cwr(struct sock *sk)
2579 struct tcp_sock *tp = tcp_sk(sk);
2581 tp->prior_ssthresh = 0;
2582 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2583 tp->undo_marker = 0;
2584 tcp_init_cwnd_reduction(sk);
2585 tcp_set_ca_state(sk, TCP_CA_CWR);
2588 EXPORT_SYMBOL(tcp_enter_cwr);
2590 static void tcp_try_keep_open(struct sock *sk)
2592 struct tcp_sock *tp = tcp_sk(sk);
2593 int state = TCP_CA_Open;
2595 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2596 state = TCP_CA_Disorder;
2598 if (inet_csk(sk)->icsk_ca_state != state) {
2599 tcp_set_ca_state(sk, state);
2600 tp->high_seq = tp->snd_nxt;
2604 static void tcp_try_to_open(struct sock *sk, int flag)
2606 struct tcp_sock *tp = tcp_sk(sk);
2608 tcp_verify_left_out(tp);
2610 if (!tcp_any_retrans_done(sk))
2611 tp->retrans_stamp = 0;
2613 if (flag & FLAG_ECE)
2616 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2617 tcp_try_keep_open(sk);
2621 static void tcp_mtup_probe_failed(struct sock *sk)
2623 struct inet_connection_sock *icsk = inet_csk(sk);
2625 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2626 icsk->icsk_mtup.probe_size = 0;
2627 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPFAIL);
2630 static void tcp_mtup_probe_success(struct sock *sk)
2632 struct tcp_sock *tp = tcp_sk(sk);
2633 struct inet_connection_sock *icsk = inet_csk(sk);
2636 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2638 val = (u64)tp->snd_cwnd * tcp_mss_to_mtu(sk, tp->mss_cache);
2639 do_div(val, icsk->icsk_mtup.probe_size);
2640 WARN_ON_ONCE((u32)val != val);
2641 tp->snd_cwnd = max_t(u32, 1U, val);
2643 tp->snd_cwnd_cnt = 0;
2644 tp->snd_cwnd_stamp = tcp_jiffies32;
2645 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2647 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2648 icsk->icsk_mtup.probe_size = 0;
2649 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2650 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS);
2653 /* Do a simple retransmit without using the backoff mechanisms in
2654 * tcp_timer. This is used for path mtu discovery.
2655 * The socket is already locked here.
2657 void tcp_simple_retransmit(struct sock *sk)
2659 const struct inet_connection_sock *icsk = inet_csk(sk);
2660 struct tcp_sock *tp = tcp_sk(sk);
2661 struct sk_buff *skb;
2662 unsigned int mss = tcp_current_mss(sk);
2664 tcp_for_write_queue(skb, sk) {
2665 if (skb == tcp_send_head(sk))
2667 if (tcp_skb_seglen(skb) > mss &&
2668 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2669 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2670 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2671 tp->retrans_out -= tcp_skb_pcount(skb);
2673 tcp_skb_mark_lost_uncond_verify(tp, skb);
2677 tcp_clear_retrans_hints_partial(tp);
2682 if (tcp_is_reno(tp))
2683 tcp_limit_reno_sacked(tp);
2685 tcp_verify_left_out(tp);
2687 /* Don't muck with the congestion window here.
2688 * Reason is that we do not increase amount of _data_
2689 * in network, but units changed and effective
2690 * cwnd/ssthresh really reduced now.
2692 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2693 tp->high_seq = tp->snd_nxt;
2694 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2695 tp->prior_ssthresh = 0;
2696 tp->undo_marker = 0;
2697 tcp_set_ca_state(sk, TCP_CA_Loss);
2699 tcp_xmit_retransmit_queue(sk);
2701 EXPORT_SYMBOL(tcp_simple_retransmit);
2703 void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2705 struct tcp_sock *tp = tcp_sk(sk);
2708 if (tcp_is_reno(tp))
2709 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2711 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2713 NET_INC_STATS(sock_net(sk), mib_idx);
2715 tp->prior_ssthresh = 0;
2718 if (!tcp_in_cwnd_reduction(sk)) {
2720 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2721 tcp_init_cwnd_reduction(sk);
2723 tcp_set_ca_state(sk, TCP_CA_Recovery);
2726 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2727 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2729 static void tcp_process_loss(struct sock *sk, int flag, bool is_dupack,
2732 struct tcp_sock *tp = tcp_sk(sk);
2733 bool recovered = !before(tp->snd_una, tp->high_seq);
2735 if ((flag & FLAG_SND_UNA_ADVANCED) &&
2736 tcp_try_undo_loss(sk, false))
2739 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2740 /* Step 3.b. A timeout is spurious if not all data are
2741 * lost, i.e., never-retransmitted data are (s)acked.
2743 if ((flag & FLAG_ORIG_SACK_ACKED) &&
2744 tcp_try_undo_loss(sk, true))
2747 if (after(tp->snd_nxt, tp->high_seq)) {
2748 if (flag & FLAG_DATA_SACKED || is_dupack)
2749 tp->frto = 0; /* Step 3.a. loss was real */
2750 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2751 tp->high_seq = tp->snd_nxt;
2752 /* Step 2.b. Try send new data (but deferred until cwnd
2753 * is updated in tcp_ack()). Otherwise fall back to
2754 * the conventional recovery.
2756 if (tcp_send_head(sk) &&
2757 after(tcp_wnd_end(tp), tp->snd_nxt)) {
2758 *rexmit = REXMIT_NEW;
2766 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2767 tcp_try_undo_recovery(sk);
2770 if (tcp_is_reno(tp)) {
2771 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2772 * delivered. Lower inflight to clock out (re)tranmissions.
2774 if (after(tp->snd_nxt, tp->high_seq) && is_dupack)
2775 tcp_add_reno_sack(sk);
2776 else if (flag & FLAG_SND_UNA_ADVANCED)
2777 tcp_reset_reno_sack(tp);
2779 *rexmit = REXMIT_LOST;
2782 /* Undo during fast recovery after partial ACK. */
2783 static bool tcp_try_undo_partial(struct sock *sk, const int acked)
2785 struct tcp_sock *tp = tcp_sk(sk);
2787 if (tp->undo_marker && tcp_packet_delayed(tp)) {
2788 /* Plain luck! Hole if filled with delayed
2789 * packet, rather than with a retransmit.
2791 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2793 /* We are getting evidence that the reordering degree is higher
2794 * than we realized. If there are no retransmits out then we
2795 * can undo. Otherwise we clock out new packets but do not
2796 * mark more packets lost or retransmit more.
2798 if (tp->retrans_out)
2801 if (!tcp_any_retrans_done(sk))
2802 tp->retrans_stamp = 0;
2804 DBGUNDO(sk, "partial recovery");
2805 tcp_undo_cwnd_reduction(sk, true);
2806 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2807 tcp_try_keep_open(sk);
2813 static void tcp_rack_identify_loss(struct sock *sk, int *ack_flag)
2815 struct tcp_sock *tp = tcp_sk(sk);
2817 /* Use RACK to detect loss */
2818 if (sysctl_tcp_recovery & TCP_RACK_LOSS_DETECTION) {
2819 u32 prior_retrans = tp->retrans_out;
2821 if (tcp_rack_mark_lost(sk))
2822 *ack_flag &= ~FLAG_SET_XMIT_TIMER;
2823 if (prior_retrans > tp->retrans_out)
2824 *ack_flag |= FLAG_LOST_RETRANS;
2828 /* Process an event, which can update packets-in-flight not trivially.
2829 * Main goal of this function is to calculate new estimate for left_out,
2830 * taking into account both packets sitting in receiver's buffer and
2831 * packets lost by network.
2833 * Besides that it updates the congestion state when packet loss or ECN
2834 * is detected. But it does not reduce the cwnd, it is done by the
2835 * congestion control later.
2837 * It does _not_ decide what to send, it is made in function
2838 * tcp_xmit_retransmit_queue().
2840 static void tcp_fastretrans_alert(struct sock *sk, const int acked,
2841 bool is_dupack, int *ack_flag, int *rexmit)
2843 struct inet_connection_sock *icsk = inet_csk(sk);
2844 struct tcp_sock *tp = tcp_sk(sk);
2845 int fast_rexmit = 0, flag = *ack_flag;
2846 bool do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2847 (tcp_fackets_out(tp) > tp->reordering));
2849 if (!tp->packets_out && tp->sacked_out)
2851 if (!tp->sacked_out && tp->fackets_out)
2852 tp->fackets_out = 0;
2854 /* Now state machine starts.
2855 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2856 if (flag & FLAG_ECE)
2857 tp->prior_ssthresh = 0;
2859 /* B. In all the states check for reneging SACKs. */
2860 if (tcp_check_sack_reneging(sk, flag))
2863 /* C. Check consistency of the current state. */
2864 tcp_verify_left_out(tp);
2866 /* D. Check state exit conditions. State can be terminated
2867 * when high_seq is ACKed. */
2868 if (icsk->icsk_ca_state == TCP_CA_Open) {
2869 WARN_ON(tp->retrans_out != 0);
2870 tp->retrans_stamp = 0;
2871 } else if (!before(tp->snd_una, tp->high_seq)) {
2872 switch (icsk->icsk_ca_state) {
2874 /* CWR is to be held something *above* high_seq
2875 * is ACKed for CWR bit to reach receiver. */
2876 if (tp->snd_una != tp->high_seq) {
2877 tcp_end_cwnd_reduction(sk);
2878 tcp_set_ca_state(sk, TCP_CA_Open);
2882 case TCP_CA_Recovery:
2883 if (tcp_is_reno(tp))
2884 tcp_reset_reno_sack(tp);
2885 if (tcp_try_undo_recovery(sk))
2887 tcp_end_cwnd_reduction(sk);
2892 /* E. Process state. */
2893 switch (icsk->icsk_ca_state) {
2894 case TCP_CA_Recovery:
2895 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2896 if (tcp_is_reno(tp) && is_dupack)
2897 tcp_add_reno_sack(sk);
2899 if (tcp_try_undo_partial(sk, acked))
2901 /* Partial ACK arrived. Force fast retransmit. */
2902 do_lost = tcp_is_reno(tp) ||
2903 tcp_fackets_out(tp) > tp->reordering;
2905 if (tcp_try_undo_dsack(sk)) {
2906 tcp_try_keep_open(sk);
2909 tcp_rack_identify_loss(sk, ack_flag);
2912 tcp_process_loss(sk, flag, is_dupack, rexmit);
2913 tcp_rack_identify_loss(sk, ack_flag);
2914 if (!(icsk->icsk_ca_state == TCP_CA_Open ||
2915 (*ack_flag & FLAG_LOST_RETRANS)))
2917 /* Change state if cwnd is undone or retransmits are lost */
2919 if (tcp_is_reno(tp)) {
2920 if (flag & FLAG_SND_UNA_ADVANCED)
2921 tcp_reset_reno_sack(tp);
2923 tcp_add_reno_sack(sk);
2926 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
2927 tcp_try_undo_dsack(sk);
2929 tcp_rack_identify_loss(sk, ack_flag);
2930 if (!tcp_time_to_recover(sk, flag)) {
2931 tcp_try_to_open(sk, flag);
2935 /* MTU probe failure: don't reduce cwnd */
2936 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2937 icsk->icsk_mtup.probe_size &&
2938 tp->snd_una == tp->mtu_probe.probe_seq_start) {
2939 tcp_mtup_probe_failed(sk);
2940 /* Restores the reduction we did in tcp_mtup_probe() */
2942 tcp_simple_retransmit(sk);
2946 /* Otherwise enter Recovery state */
2947 tcp_enter_recovery(sk, (flag & FLAG_ECE));
2952 tcp_update_scoreboard(sk, fast_rexmit);
2953 *rexmit = REXMIT_LOST;
2956 static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us)
2958 struct tcp_sock *tp = tcp_sk(sk);
2959 u32 wlen = sysctl_tcp_min_rtt_wlen * HZ;
2961 minmax_running_min(&tp->rtt_min, wlen, tcp_jiffies32,
2962 rtt_us ? : jiffies_to_usecs(1));
2965 static bool tcp_ack_update_rtt(struct sock *sk, const int flag,
2966 long seq_rtt_us, long sack_rtt_us,
2967 long ca_rtt_us, struct rate_sample *rs)
2969 const struct tcp_sock *tp = tcp_sk(sk);
2971 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2972 * broken middle-boxes or peers may corrupt TS-ECR fields. But
2973 * Karn's algorithm forbids taking RTT if some retransmitted data
2974 * is acked (RFC6298).
2977 seq_rtt_us = sack_rtt_us;
2979 /* RTTM Rule: A TSecr value received in a segment is used to
2980 * update the averaged RTT measurement only if the segment
2981 * acknowledges some new data, i.e., only if it advances the
2982 * left edge of the send window.
2983 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2985 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2986 flag & FLAG_ACKED) {
2987 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr;
2988 u32 delta_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
2990 seq_rtt_us = ca_rtt_us = delta_us;
2992 rs->rtt_us = ca_rtt_us; /* RTT of last (S)ACKed packet (or -1) */
2996 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
2997 * always taken together with ACK, SACK, or TS-opts. Any negative
2998 * values will be skipped with the seq_rtt_us < 0 check above.
3000 tcp_update_rtt_min(sk, ca_rtt_us);
3001 tcp_rtt_estimator(sk, seq_rtt_us);
3004 /* RFC6298: only reset backoff on valid RTT measurement. */
3005 inet_csk(sk)->icsk_backoff = 0;
3009 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
3010 void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req)
3012 struct rate_sample rs;
3015 if (req && !req->num_retrans && tcp_rsk(req)->snt_synack)
3016 rtt_us = tcp_stamp_us_delta(tcp_clock_us(), tcp_rsk(req)->snt_synack);
3018 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us, &rs);
3022 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
3024 const struct inet_connection_sock *icsk = inet_csk(sk);
3026 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
3027 tcp_sk(sk)->snd_cwnd_stamp = tcp_jiffies32;
3030 /* Restart timer after forward progress on connection.
3031 * RFC2988 recommends to restart timer to now+rto.
3033 void tcp_rearm_rto(struct sock *sk)
3035 const struct inet_connection_sock *icsk = inet_csk(sk);
3036 struct tcp_sock *tp = tcp_sk(sk);
3038 /* If the retrans timer is currently being used by Fast Open
3039 * for SYN-ACK retrans purpose, stay put.
3041 if (tp->fastopen_rsk)
3044 if (!tp->packets_out) {
3045 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
3047 u32 rto = inet_csk(sk)->icsk_rto;
3048 /* Offset the time elapsed after installing regular RTO */
3049 if (icsk->icsk_pending == ICSK_TIME_REO_TIMEOUT ||
3050 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
3051 s64 delta_us = tcp_rto_delta_us(sk);
3052 /* delta_us may not be positive if the socket is locked
3053 * when the retrans timer fires and is rescheduled.
3055 rto = usecs_to_jiffies(max_t(int, delta_us, 1));
3057 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
3062 /* Try to schedule a loss probe; if that doesn't work, then schedule an RTO. */
3063 static void tcp_set_xmit_timer(struct sock *sk)
3065 if (!tcp_schedule_loss_probe(sk, true))
3069 /* If we get here, the whole TSO packet has not been acked. */
3070 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3072 struct tcp_sock *tp = tcp_sk(sk);
3075 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3077 packets_acked = tcp_skb_pcount(skb);
3078 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3080 packets_acked -= tcp_skb_pcount(skb);
3082 if (packets_acked) {
3083 BUG_ON(tcp_skb_pcount(skb) == 0);
3084 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3087 return packets_acked;
3090 static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
3093 const struct skb_shared_info *shinfo;
3095 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3096 if (likely(!TCP_SKB_CB(skb)->txstamp_ack))
3099 shinfo = skb_shinfo(skb);
3100 if (!before(shinfo->tskey, prior_snd_una) &&
3101 before(shinfo->tskey, tcp_sk(sk)->snd_una))
3102 __skb_tstamp_tx(skb, NULL, sk, SCM_TSTAMP_ACK);
3105 /* Remove acknowledged frames from the retransmission queue. If our packet
3106 * is before the ack sequence we can discard it as it's confirmed to have
3107 * arrived at the other end.
3109 static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
3110 u32 prior_snd_una, int *acked,
3111 struct tcp_sacktag_state *sack)
3113 const struct inet_connection_sock *icsk = inet_csk(sk);
3114 u64 first_ackt, last_ackt;
3115 struct tcp_sock *tp = tcp_sk(sk);
3116 u32 prior_sacked = tp->sacked_out;
3117 u32 reord = tp->packets_out;
3118 bool fully_acked = true;
3119 long sack_rtt_us = -1L;
3120 long seq_rtt_us = -1L;
3121 long ca_rtt_us = -1L;
3122 struct sk_buff *skb;
3124 u32 last_in_flight = 0;
3130 while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
3131 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3132 u8 sacked = scb->sacked;
3135 tcp_ack_tstamp(sk, skb, prior_snd_una);
3137 /* Determine how many packets and what bytes were acked, tso and else */
3138 if (after(scb->end_seq, tp->snd_una)) {
3139 if (tcp_skb_pcount(skb) == 1 ||
3140 !after(tp->snd_una, scb->seq))
3143 acked_pcount = tcp_tso_acked(sk, skb);
3146 fully_acked = false;
3148 /* Speedup tcp_unlink_write_queue() and next loop */
3149 prefetchw(skb->next);
3150 acked_pcount = tcp_skb_pcount(skb);
3153 if (unlikely(sacked & TCPCB_RETRANS)) {
3154 if (sacked & TCPCB_SACKED_RETRANS)
3155 tp->retrans_out -= acked_pcount;
3156 flag |= FLAG_RETRANS_DATA_ACKED;
3157 } else if (!(sacked & TCPCB_SACKED_ACKED)) {
3158 last_ackt = skb->skb_mstamp;
3159 WARN_ON_ONCE(last_ackt == 0);
3161 first_ackt = last_ackt;
3163 last_in_flight = TCP_SKB_CB(skb)->tx.in_flight;
3164 reord = min(pkts_acked, reord);
3165 if (!after(scb->end_seq, tp->high_seq))
3166 flag |= FLAG_ORIG_SACK_ACKED;
3169 if (sacked & TCPCB_SACKED_ACKED) {
3170 tp->sacked_out -= acked_pcount;
3171 } else if (tcp_is_sack(tp)) {
3172 tp->delivered += acked_pcount;
3173 if (!tcp_skb_spurious_retrans(tp, skb))
3174 tcp_rack_advance(tp, sacked, scb->end_seq,
3177 if (sacked & TCPCB_LOST)
3178 tp->lost_out -= acked_pcount;
3180 tp->packets_out -= acked_pcount;
3181 pkts_acked += acked_pcount;
3182 tcp_rate_skb_delivered(sk, skb, sack->rate);
3184 /* Initial outgoing SYN's get put onto the write_queue
3185 * just like anything else we transmit. It is not
3186 * true data, and if we misinform our callers that
3187 * this ACK acks real data, we will erroneously exit
3188 * connection startup slow start one packet too
3189 * quickly. This is severely frowned upon behavior.
3191 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
3192 flag |= FLAG_DATA_ACKED;
3194 flag |= FLAG_SYN_ACKED;
3195 tp->retrans_stamp = 0;
3201 tcp_unlink_write_queue(skb, sk);
3202 sk_wmem_free_skb(sk, skb);
3203 if (unlikely(skb == tp->retransmit_skb_hint))
3204 tp->retransmit_skb_hint = NULL;
3205 if (unlikely(skb == tp->lost_skb_hint))
3206 tp->lost_skb_hint = NULL;
3210 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
3212 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3213 tp->snd_up = tp->snd_una;
3215 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3216 flag |= FLAG_SACK_RENEGING;
3218 if (likely(first_ackt) && !(flag & FLAG_RETRANS_DATA_ACKED)) {
3219 seq_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, first_ackt);
3220 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, last_ackt);
3222 if (sack->first_sackt) {
3223 sack_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->first_sackt);
3224 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->last_sackt);
3226 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us,
3227 ca_rtt_us, sack->rate);
3229 if (flag & FLAG_ACKED) {
3230 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3231 if (unlikely(icsk->icsk_mtup.probe_size &&
3232 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3233 tcp_mtup_probe_success(sk);
3236 if (tcp_is_reno(tp)) {
3237 tcp_remove_reno_sacks(sk, pkts_acked);
3239 /* If any of the cumulatively ACKed segments was
3240 * retransmitted, non-SACK case cannot confirm that
3241 * progress was due to original transmission due to
3242 * lack of TCPCB_SACKED_ACKED bits even if some of
3243 * the packets may have been never retransmitted.
3245 if (flag & FLAG_RETRANS_DATA_ACKED)
3246 flag &= ~FLAG_ORIG_SACK_ACKED;
3250 /* Non-retransmitted hole got filled? That's reordering */
3251 if (reord < prior_fackets && reord <= tp->fackets_out)
3252 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
3254 delta = tcp_is_fack(tp) ? pkts_acked :
3255 prior_sacked - tp->sacked_out;
3256 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3259 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
3261 } else if (skb && rtt_update && sack_rtt_us >= 0 &&
3262 sack_rtt_us > tcp_stamp_us_delta(tp->tcp_mstamp, skb->skb_mstamp)) {
3263 /* Do not re-arm RTO if the sack RTT is measured from data sent
3264 * after when the head was last (re)transmitted. Otherwise the
3265 * timeout may continue to extend in loss recovery.
3267 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3270 if (icsk->icsk_ca_ops->pkts_acked) {
3271 struct ack_sample sample = { .pkts_acked = pkts_acked,
3272 .rtt_us = sack->rate->rtt_us,
3273 .in_flight = last_in_flight };
3275 icsk->icsk_ca_ops->pkts_acked(sk, &sample);
3278 #if FASTRETRANS_DEBUG > 0
3279 WARN_ON((int)tp->sacked_out < 0);
3280 WARN_ON((int)tp->lost_out < 0);
3281 WARN_ON((int)tp->retrans_out < 0);
3282 if (!tp->packets_out && tcp_is_sack(tp)) {
3283 icsk = inet_csk(sk);
3285 pr_debug("Leak l=%u %d\n",
3286 tp->lost_out, icsk->icsk_ca_state);
3289 if (tp->sacked_out) {
3290 pr_debug("Leak s=%u %d\n",
3291 tp->sacked_out, icsk->icsk_ca_state);
3294 if (tp->retrans_out) {
3295 pr_debug("Leak r=%u %d\n",
3296 tp->retrans_out, icsk->icsk_ca_state);
3297 tp->retrans_out = 0;
3301 *acked = pkts_acked;
3305 static void tcp_ack_probe(struct sock *sk)
3307 const struct tcp_sock *tp = tcp_sk(sk);
3308 struct inet_connection_sock *icsk = inet_csk(sk);
3310 /* Was it a usable window open? */
3312 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
3313 icsk->icsk_backoff = 0;
3314 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3315 /* Socket must be waked up by subsequent tcp_data_snd_check().
3316 * This function is not for random using!
3319 unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX);
3321 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3326 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3328 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3329 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3332 /* Decide wheather to run the increase function of congestion control. */
3333 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3335 /* If reordering is high then always grow cwnd whenever data is
3336 * delivered regardless of its ordering. Otherwise stay conservative
3337 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3338 * new SACK or ECE mark may first advance cwnd here and later reduce
3339 * cwnd in tcp_fastretrans_alert() based on more states.
3341 if (tcp_sk(sk)->reordering > sock_net(sk)->ipv4.sysctl_tcp_reordering)
3342 return flag & FLAG_FORWARD_PROGRESS;
3344 return flag & FLAG_DATA_ACKED;
3347 /* The "ultimate" congestion control function that aims to replace the rigid
3348 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction).
3349 * It's called toward the end of processing an ACK with precise rate
3350 * information. All transmission or retransmission are delayed afterwards.
3352 static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked,
3353 int flag, const struct rate_sample *rs)
3355 const struct inet_connection_sock *icsk = inet_csk(sk);
3357 if (icsk->icsk_ca_ops->cong_control) {
3358 icsk->icsk_ca_ops->cong_control(sk, rs);
3362 if (tcp_in_cwnd_reduction(sk)) {
3363 /* Reduce cwnd if state mandates */
3364 tcp_cwnd_reduction(sk, acked_sacked, flag);
3365 } else if (tcp_may_raise_cwnd(sk, flag)) {
3366 /* Advance cwnd if state allows */
3367 tcp_cong_avoid(sk, ack, acked_sacked);
3369 tcp_update_pacing_rate(sk);
3372 /* Check that window update is acceptable.
3373 * The function assumes that snd_una<=ack<=snd_next.
3375 static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3376 const u32 ack, const u32 ack_seq,
3379 return after(ack, tp->snd_una) ||
3380 after(ack_seq, tp->snd_wl1) ||
3381 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3384 /* If we update tp->snd_una, also update tp->bytes_acked */
3385 static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack)
3387 u32 delta = ack - tp->snd_una;
3389 sock_owned_by_me((struct sock *)tp);
3390 tp->bytes_acked += delta;
3394 /* If we update tp->rcv_nxt, also update tp->bytes_received */
3395 static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq)
3397 u32 delta = seq - tp->rcv_nxt;
3399 sock_owned_by_me((struct sock *)tp);
3400 tp->bytes_received += delta;
3404 /* Update our send window.
3406 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3407 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3409 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3412 struct tcp_sock *tp = tcp_sk(sk);
3414 u32 nwin = ntohs(tcp_hdr(skb)->window);
3416 if (likely(!tcp_hdr(skb)->syn))
3417 nwin <<= tp->rx_opt.snd_wscale;
3419 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3420 flag |= FLAG_WIN_UPDATE;
3421 tcp_update_wl(tp, ack_seq);
3423 if (tp->snd_wnd != nwin) {
3426 /* Note, it is the only place, where
3427 * fast path is recovered for sending TCP.
3430 tcp_fast_path_check(sk);
3432 if (tcp_send_head(sk))
3433 tcp_slow_start_after_idle_check(sk);
3435 if (nwin > tp->max_window) {
3436 tp->max_window = nwin;
3437 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3442 tcp_snd_una_update(tp, ack);
3447 static bool __tcp_oow_rate_limited(struct net *net, int mib_idx,
3448 u32 *last_oow_ack_time)
3450 if (*last_oow_ack_time) {
3451 s32 elapsed = (s32)(tcp_jiffies32 - *last_oow_ack_time);
3453 if (0 <= elapsed && elapsed < sysctl_tcp_invalid_ratelimit) {
3454 NET_INC_STATS(net, mib_idx);
3455 return true; /* rate-limited: don't send yet! */
3459 *last_oow_ack_time = tcp_jiffies32;
3461 return false; /* not rate-limited: go ahead, send dupack now! */
3464 /* Return true if we're currently rate-limiting out-of-window ACKs and
3465 * thus shouldn't send a dupack right now. We rate-limit dupacks in
3466 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3467 * attacks that send repeated SYNs or ACKs for the same connection. To
3468 * do this, we do not send a duplicate SYNACK or ACK if the remote
3469 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3471 bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb,
3472 int mib_idx, u32 *last_oow_ack_time)
3474 /* Data packets without SYNs are not likely part of an ACK loop. */
3475 if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) &&
3479 return __tcp_oow_rate_limited(net, mib_idx, last_oow_ack_time);
3482 /* RFC 5961 7 [ACK Throttling] */
3483 static void tcp_send_challenge_ack(struct sock *sk, const struct sk_buff *skb)
3485 /* unprotected vars, we dont care of overwrites */
3486 static u32 challenge_timestamp;
3487 static unsigned int challenge_count;
3488 struct tcp_sock *tp = tcp_sk(sk);
3491 /* First check our per-socket dupack rate limit. */
3492 if (__tcp_oow_rate_limited(sock_net(sk),
3493 LINUX_MIB_TCPACKSKIPPEDCHALLENGE,
3494 &tp->last_oow_ack_time))
3497 /* Then check host-wide RFC 5961 rate limit. */
3499 if (now != challenge_timestamp) {
3500 u32 half = (sysctl_tcp_challenge_ack_limit + 1) >> 1;
3502 challenge_timestamp = now;
3503 WRITE_ONCE(challenge_count, half +
3504 prandom_u32_max(sysctl_tcp_challenge_ack_limit));
3506 count = READ_ONCE(challenge_count);
3508 WRITE_ONCE(challenge_count, count - 1);
3509 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPCHALLENGEACK);
3514 static void tcp_store_ts_recent(struct tcp_sock *tp)
3516 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3517 tp->rx_opt.ts_recent_stamp = get_seconds();
3520 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3522 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3523 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3524 * extra check below makes sure this can only happen
3525 * for pure ACK frames. -DaveM
3527 * Not only, also it occurs for expired timestamps.
3530 if (tcp_paws_check(&tp->rx_opt, 0))
3531 tcp_store_ts_recent(tp);
3535 /* This routine deals with acks during a TLP episode and ends an episode by
3536 * resetting tlp_high_seq. Ref: TLP algorithm in draft-ietf-tcpm-rack
3538 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3540 struct tcp_sock *tp = tcp_sk(sk);
3542 if (before(ack, tp->tlp_high_seq))
3545 if (!tp->tlp_retrans) {
3546 /* TLP of new data has been acknowledged */
3547 tp->tlp_high_seq = 0;
3548 } else if (flag & FLAG_DSACKING_ACK) {
3549 /* This DSACK means original and TLP probe arrived; no loss */
3550 tp->tlp_high_seq = 0;
3551 } else if (after(ack, tp->tlp_high_seq)) {
3552 /* ACK advances: there was a loss, so reduce cwnd. Reset
3553 * tlp_high_seq in tcp_init_cwnd_reduction()
3555 tcp_init_cwnd_reduction(sk);
3556 tcp_set_ca_state(sk, TCP_CA_CWR);
3557 tcp_end_cwnd_reduction(sk);
3558 tcp_try_keep_open(sk);
3559 NET_INC_STATS(sock_net(sk),
3560 LINUX_MIB_TCPLOSSPROBERECOVERY);
3561 } else if (!(flag & (FLAG_SND_UNA_ADVANCED |
3562 FLAG_NOT_DUP | FLAG_DATA_SACKED))) {
3563 /* Pure dupack: original and TLP probe arrived; no loss */
3564 tp->tlp_high_seq = 0;
3568 static inline void tcp_in_ack_event(struct sock *sk, u32 flags)
3570 const struct inet_connection_sock *icsk = inet_csk(sk);
3572 if (icsk->icsk_ca_ops->in_ack_event)
3573 icsk->icsk_ca_ops->in_ack_event(sk, flags);
3576 /* Congestion control has updated the cwnd already. So if we're in
3577 * loss recovery then now we do any new sends (for FRTO) or
3578 * retransmits (for CA_Loss or CA_recovery) that make sense.
3580 static void tcp_xmit_recovery(struct sock *sk, int rexmit)
3582 struct tcp_sock *tp = tcp_sk(sk);
3584 if (rexmit == REXMIT_NONE)
3587 if (unlikely(rexmit == 2)) {
3588 __tcp_push_pending_frames(sk, tcp_current_mss(sk),
3590 if (after(tp->snd_nxt, tp->high_seq))
3594 tcp_xmit_retransmit_queue(sk);
3597 /* This routine deals with incoming acks, but not outgoing ones. */
3598 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3600 struct inet_connection_sock *icsk = inet_csk(sk);
3601 struct tcp_sock *tp = tcp_sk(sk);
3602 struct tcp_sacktag_state sack_state;
3603 struct rate_sample rs = { .prior_delivered = 0 };
3604 u32 prior_snd_una = tp->snd_una;
3605 bool is_sack_reneg = tp->is_sack_reneg;
3606 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3607 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3608 bool is_dupack = false;
3610 int prior_packets = tp->packets_out;
3611 u32 delivered = tp->delivered;
3612 u32 lost = tp->lost;
3613 int acked = 0; /* Number of packets newly acked */
3614 int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */
3616 sack_state.first_sackt = 0;
3617 sack_state.rate = &rs;
3619 /* We very likely will need to access write queue head. */
3620 prefetchw(sk->sk_write_queue.next);
3622 /* If the ack is older than previous acks
3623 * then we can probably ignore it.
3625 if (before(ack, prior_snd_una)) {
3626 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3627 if (before(ack, prior_snd_una - tp->max_window)) {
3628 if (!(flag & FLAG_NO_CHALLENGE_ACK))
3629 tcp_send_challenge_ack(sk, skb);
3635 /* If the ack includes data we haven't sent yet, discard
3636 * this segment (RFC793 Section 3.9).
3638 if (after(ack, tp->snd_nxt))
3641 if (after(ack, prior_snd_una)) {
3642 flag |= FLAG_SND_UNA_ADVANCED;
3643 icsk->icsk_retransmits = 0;
3646 prior_fackets = tp->fackets_out;
3647 rs.prior_in_flight = tcp_packets_in_flight(tp);
3649 /* ts_recent update must be made after we are sure that the packet
3652 if (flag & FLAG_UPDATE_TS_RECENT)
3653 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3655 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3656 /* Window is constant, pure forward advance.
3657 * No more checks are required.
3658 * Note, we use the fact that SND.UNA>=SND.WL2.
3660 tcp_update_wl(tp, ack_seq);
3661 tcp_snd_una_update(tp, ack);
3662 flag |= FLAG_WIN_UPDATE;
3664 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
3666 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPACKS);
3668 u32 ack_ev_flags = CA_ACK_SLOWPATH;
3670 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3673 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3675 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3677 if (TCP_SKB_CB(skb)->sacked)
3678 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3681 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) {
3683 ack_ev_flags |= CA_ACK_ECE;
3686 if (flag & FLAG_WIN_UPDATE)
3687 ack_ev_flags |= CA_ACK_WIN_UPDATE;
3689 tcp_in_ack_event(sk, ack_ev_flags);
3692 /* We passed data and got it acked, remove any soft error
3693 * log. Something worked...
3695 sk->sk_err_soft = 0;
3696 icsk->icsk_probes_out = 0;
3697 tp->rcv_tstamp = tcp_jiffies32;
3701 /* See if we can take anything off of the retransmit queue. */
3702 flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una, &acked,
3705 if (tp->tlp_high_seq)
3706 tcp_process_tlp_ack(sk, ack, flag);
3708 if (tcp_ack_is_dubious(sk, flag)) {
3709 is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
3710 tcp_fastretrans_alert(sk, acked, is_dupack, &flag, &rexmit);
3713 /* If needed, reset TLP/RTO timer when RACK doesn't set. */
3714 if (flag & FLAG_SET_XMIT_TIMER)
3715 tcp_set_xmit_timer(sk);
3717 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
3720 delivered = tp->delivered - delivered; /* freshly ACKed or SACKed */
3721 lost = tp->lost - lost; /* freshly marked lost */
3722 tcp_rate_gen(sk, delivered, lost, is_sack_reneg, sack_state.rate);
3723 tcp_cong_control(sk, ack, delivered, flag, sack_state.rate);
3724 tcp_xmit_recovery(sk, rexmit);
3728 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3729 if (flag & FLAG_DSACKING_ACK)
3730 tcp_fastretrans_alert(sk, acked, is_dupack, &flag, &rexmit);
3731 /* If this ack opens up a zero window, clear backoff. It was
3732 * being used to time the probes, and is probably far higher than
3733 * it needs to be for normal retransmission.
3735 if (tcp_send_head(sk))
3738 if (tp->tlp_high_seq)
3739 tcp_process_tlp_ack(sk, ack, flag);
3743 SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3747 /* If data was SACKed, tag it and see if we should send more data.
3748 * If data was DSACKed, see if we can undo a cwnd reduction.
3750 if (TCP_SKB_CB(skb)->sacked) {
3751 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3753 tcp_fastretrans_alert(sk, acked, is_dupack, &flag, &rexmit);
3754 tcp_xmit_recovery(sk, rexmit);
3757 SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3761 static void tcp_parse_fastopen_option(int len, const unsigned char *cookie,
3762 bool syn, struct tcp_fastopen_cookie *foc,
3765 /* Valid only in SYN or SYN-ACK with an even length. */
3766 if (!foc || !syn || len < 0 || (len & 1))
3769 if (len >= TCP_FASTOPEN_COOKIE_MIN &&
3770 len <= TCP_FASTOPEN_COOKIE_MAX)
3771 memcpy(foc->val, cookie, len);
3778 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3779 * But, this can also be called on packets in the established flow when
3780 * the fast version below fails.
3782 void tcp_parse_options(const struct net *net,
3783 const struct sk_buff *skb,
3784 struct tcp_options_received *opt_rx, int estab,
3785 struct tcp_fastopen_cookie *foc)
3787 const unsigned char *ptr;
3788 const struct tcphdr *th = tcp_hdr(skb);
3789 int length = (th->doff * 4) - sizeof(struct tcphdr);
3791 ptr = (const unsigned char *)(th + 1);
3792 opt_rx->saw_tstamp = 0;
3794 while (length > 0) {
3795 int opcode = *ptr++;
3801 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3806 if (opsize < 2) /* "silly options" */
3808 if (opsize > length)
3809 return; /* don't parse partial options */
3812 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3813 u16 in_mss = get_unaligned_be16(ptr);
3815 if (opt_rx->user_mss &&
3816 opt_rx->user_mss < in_mss)
3817 in_mss = opt_rx->user_mss;
3818 opt_rx->mss_clamp = in_mss;
3823 if (opsize == TCPOLEN_WINDOW && th->syn &&
3824 !estab && net->ipv4.sysctl_tcp_window_scaling) {
3825 __u8 snd_wscale = *(__u8 *)ptr;
3826 opt_rx->wscale_ok = 1;
3827 if (snd_wscale > TCP_MAX_WSCALE) {
3828 net_info_ratelimited("%s: Illegal window scaling value %d > %u received\n",
3832 snd_wscale = TCP_MAX_WSCALE;
3834 opt_rx->snd_wscale = snd_wscale;
3837 case TCPOPT_TIMESTAMP:
3838 if ((opsize == TCPOLEN_TIMESTAMP) &&
3839 ((estab && opt_rx->tstamp_ok) ||
3840 (!estab && net->ipv4.sysctl_tcp_timestamps))) {
3841 opt_rx->saw_tstamp = 1;
3842 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3843 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3846 case TCPOPT_SACK_PERM:
3847 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3848 !estab && net->ipv4.sysctl_tcp_sack) {
3849 opt_rx->sack_ok = TCP_SACK_SEEN;
3850 tcp_sack_reset(opt_rx);
3855 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3856 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3858 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3861 #ifdef CONFIG_TCP_MD5SIG
3864 * The MD5 Hash has already been
3865 * checked (see tcp_v{4,6}_do_rcv()).
3869 case TCPOPT_FASTOPEN:
3870 tcp_parse_fastopen_option(
3871 opsize - TCPOLEN_FASTOPEN_BASE,
3872 ptr, th->syn, foc, false);
3876 /* Fast Open option shares code 254 using a
3877 * 16 bits magic number.
3879 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE &&
3880 get_unaligned_be16(ptr) ==
3881 TCPOPT_FASTOPEN_MAGIC)
3882 tcp_parse_fastopen_option(opsize -
3883 TCPOLEN_EXP_FASTOPEN_BASE,
3884 ptr + 2, th->syn, foc, true);
3893 EXPORT_SYMBOL(tcp_parse_options);
3895 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
3897 const __be32 *ptr = (const __be32 *)(th + 1);
3899 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3900 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3901 tp->rx_opt.saw_tstamp = 1;
3903 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3906 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
3908 tp->rx_opt.rcv_tsecr = 0;
3914 /* Fast parse options. This hopes to only see timestamps.
3915 * If it is wrong it falls back on tcp_parse_options().
3917 static bool tcp_fast_parse_options(const struct net *net,
3918 const struct sk_buff *skb,
3919 const struct tcphdr *th, struct tcp_sock *tp)
3921 /* In the spirit of fast parsing, compare doff directly to constant
3922 * values. Because equality is used, short doff can be ignored here.
3924 if (th->doff == (sizeof(*th) / 4)) {
3925 tp->rx_opt.saw_tstamp = 0;
3927 } else if (tp->rx_opt.tstamp_ok &&
3928 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3929 if (tcp_parse_aligned_timestamp(tp, th))
3933 tcp_parse_options(net, skb, &tp->rx_opt, 1, NULL);
3934 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3935 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
3940 #ifdef CONFIG_TCP_MD5SIG
3942 * Parse MD5 Signature option
3944 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
3946 int length = (th->doff << 2) - sizeof(*th);
3947 const u8 *ptr = (const u8 *)(th + 1);
3949 /* If not enough data remaining, we can short cut */
3950 while (length >= TCPOLEN_MD5SIG) {
3951 int opcode = *ptr++;
3962 if (opsize < 2 || opsize > length)
3964 if (opcode == TCPOPT_MD5SIG)
3965 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
3972 EXPORT_SYMBOL(tcp_parse_md5sig_option);
3975 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3977 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3978 * it can pass through stack. So, the following predicate verifies that
3979 * this segment is not used for anything but congestion avoidance or
3980 * fast retransmit. Moreover, we even are able to eliminate most of such
3981 * second order effects, if we apply some small "replay" window (~RTO)
3982 * to timestamp space.
3984 * All these measures still do not guarantee that we reject wrapped ACKs
3985 * on networks with high bandwidth, when sequence space is recycled fastly,
3986 * but it guarantees that such events will be very rare and do not affect
3987 * connection seriously. This doesn't look nice, but alas, PAWS is really
3990 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3991 * states that events when retransmit arrives after original data are rare.
3992 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3993 * the biggest problem on large power networks even with minor reordering.
3994 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3995 * up to bandwidth of 18Gigabit/sec. 8) ]
3998 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
4000 const struct tcp_sock *tp = tcp_sk(sk);
4001 const struct tcphdr *th = tcp_hdr(skb);
4002 u32 seq = TCP_SKB_CB(skb)->seq;
4003 u32 ack = TCP_SKB_CB(skb)->ack_seq;
4005 return (/* 1. Pure ACK with correct sequence number. */
4006 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
4008 /* 2. ... and duplicate ACK. */
4009 ack == tp->snd_una &&
4011 /* 3. ... and does not update window. */
4012 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
4014 /* 4. ... and sits in replay window. */
4015 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
4018 static inline bool tcp_paws_discard(const struct sock *sk,
4019 const struct sk_buff *skb)
4021 const struct tcp_sock *tp = tcp_sk(sk);
4023 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
4024 !tcp_disordered_ack(sk, skb);
4027 /* Check segment sequence number for validity.
4029 * Segment controls are considered valid, if the segment
4030 * fits to the window after truncation to the window. Acceptability
4031 * of data (and SYN, FIN, of course) is checked separately.
4032 * See tcp_data_queue(), for example.
4034 * Also, controls (RST is main one) are accepted using RCV.WUP instead
4035 * of RCV.NXT. Peer still did not advance his SND.UNA when we
4036 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4037 * (borrowed from freebsd)
4040 static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
4042 return !before(end_seq, tp->rcv_wup) &&
4043 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
4046 /* When we get a reset we do this. */
4047 void tcp_reset(struct sock *sk)
4049 /* We want the right error as BSD sees it (and indeed as we do). */
4050 switch (sk->sk_state) {
4052 sk->sk_err = ECONNREFUSED;
4054 case TCP_CLOSE_WAIT:
4060 sk->sk_err = ECONNRESET;
4062 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4065 tcp_write_queue_purge(sk);
4068 if (!sock_flag(sk, SOCK_DEAD))
4069 sk->sk_error_report(sk);
4073 * Process the FIN bit. This now behaves as it is supposed to work
4074 * and the FIN takes effect when it is validly part of sequence
4075 * space. Not before when we get holes.
4077 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4078 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4081 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4082 * close and we go into CLOSING (and later onto TIME-WAIT)
4084 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4086 void tcp_fin(struct sock *sk)
4088 struct tcp_sock *tp = tcp_sk(sk);
4090 inet_csk_schedule_ack(sk);
4092 sk->sk_shutdown |= RCV_SHUTDOWN;
4093 sock_set_flag(sk, SOCK_DONE);
4095 switch (sk->sk_state) {
4097 case TCP_ESTABLISHED:
4098 /* Move to CLOSE_WAIT */
4099 tcp_set_state(sk, TCP_CLOSE_WAIT);
4100 inet_csk(sk)->icsk_ack.pingpong = 1;
4103 case TCP_CLOSE_WAIT:
4105 /* Received a retransmission of the FIN, do
4110 /* RFC793: Remain in the LAST-ACK state. */
4114 /* This case occurs when a simultaneous close
4115 * happens, we must ack the received FIN and
4116 * enter the CLOSING state.
4119 tcp_set_state(sk, TCP_CLOSING);
4122 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4124 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4127 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4128 * cases we should never reach this piece of code.
4130 pr_err("%s: Impossible, sk->sk_state=%d\n",
4131 __func__, sk->sk_state);
4135 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4136 * Probably, we should reset in this case. For now drop them.
4138 skb_rbtree_purge(&tp->out_of_order_queue);
4139 if (tcp_is_sack(tp))
4140 tcp_sack_reset(&tp->rx_opt);
4143 if (!sock_flag(sk, SOCK_DEAD)) {
4144 sk->sk_state_change(sk);
4146 /* Do not send POLL_HUP for half duplex close. */
4147 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4148 sk->sk_state == TCP_CLOSE)
4149 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4151 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4155 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4158 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4159 if (before(seq, sp->start_seq))
4160 sp->start_seq = seq;
4161 if (after(end_seq, sp->end_seq))
4162 sp->end_seq = end_seq;
4168 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4170 struct tcp_sock *tp = tcp_sk(sk);
4172 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4175 if (before(seq, tp->rcv_nxt))
4176 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4178 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4180 NET_INC_STATS(sock_net(sk), mib_idx);
4182 tp->rx_opt.dsack = 1;
4183 tp->duplicate_sack[0].start_seq = seq;
4184 tp->duplicate_sack[0].end_seq = end_seq;
4188 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4190 struct tcp_sock *tp = tcp_sk(sk);
4192 if (!tp->rx_opt.dsack)
4193 tcp_dsack_set(sk, seq, end_seq);
4195 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4198 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4200 struct tcp_sock *tp = tcp_sk(sk);
4202 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4203 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4204 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4205 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
4207 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4208 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4210 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4211 end_seq = tp->rcv_nxt;
4212 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4219 /* These routines update the SACK block as out-of-order packets arrive or
4220 * in-order packets close up the sequence space.
4222 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4225 struct tcp_sack_block *sp = &tp->selective_acks[0];
4226 struct tcp_sack_block *swalk = sp + 1;
4228 /* See if the recent change to the first SACK eats into
4229 * or hits the sequence space of other SACK blocks, if so coalesce.
4231 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4232 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4235 /* Zap SWALK, by moving every further SACK up by one slot.
4236 * Decrease num_sacks.
4238 tp->rx_opt.num_sacks--;
4239 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4243 this_sack++, swalk++;
4247 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4249 struct tcp_sock *tp = tcp_sk(sk);
4250 struct tcp_sack_block *sp = &tp->selective_acks[0];
4251 int cur_sacks = tp->rx_opt.num_sacks;
4257 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4258 if (tcp_sack_extend(sp, seq, end_seq)) {
4259 /* Rotate this_sack to the first one. */
4260 for (; this_sack > 0; this_sack--, sp--)
4261 swap(*sp, *(sp - 1));
4263 tcp_sack_maybe_coalesce(tp);
4268 /* Could not find an adjacent existing SACK, build a new one,
4269 * put it at the front, and shift everyone else down. We
4270 * always know there is at least one SACK present already here.
4272 * If the sack array is full, forget about the last one.
4274 if (this_sack >= TCP_NUM_SACKS) {
4276 tp->rx_opt.num_sacks--;
4279 for (; this_sack > 0; this_sack--, sp--)
4283 /* Build the new head SACK, and we're done. */
4284 sp->start_seq = seq;
4285 sp->end_seq = end_seq;
4286 tp->rx_opt.num_sacks++;
4289 /* RCV.NXT advances, some SACKs should be eaten. */
4291 static void tcp_sack_remove(struct tcp_sock *tp)
4293 struct tcp_sack_block *sp = &tp->selective_acks[0];
4294 int num_sacks = tp->rx_opt.num_sacks;
4297 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4298 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4299 tp->rx_opt.num_sacks = 0;
4303 for (this_sack = 0; this_sack < num_sacks;) {
4304 /* Check if the start of the sack is covered by RCV.NXT. */
4305 if (!before(tp->rcv_nxt, sp->start_seq)) {
4308 /* RCV.NXT must cover all the block! */
4309 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4311 /* Zap this SACK, by moving forward any other SACKS. */
4312 for (i = this_sack+1; i < num_sacks; i++)
4313 tp->selective_acks[i-1] = tp->selective_acks[i];
4320 tp->rx_opt.num_sacks = num_sacks;
4329 * tcp_try_coalesce - try to merge skb to prior one
4331 * @dest: destination queue
4333 * @from: buffer to add in queue
4334 * @fragstolen: pointer to boolean
4336 * Before queueing skb @from after @to, try to merge them
4337 * to reduce overall memory use and queue lengths, if cost is small.
4338 * Packets in ofo or receive queues can stay a long time.
4339 * Better try to coalesce them right now to avoid future collapses.
4340 * Returns true if caller should free @from instead of queueing it
4342 static bool tcp_try_coalesce(struct sock *sk,
4343 enum tcp_queue dest,
4345 struct sk_buff *from,
4350 *fragstolen = false;
4352 /* Its possible this segment overlaps with prior segment in queue */
4353 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4356 if (!skb_try_coalesce(to, from, fragstolen, &delta))
4359 atomic_add(delta, &sk->sk_rmem_alloc);
4360 sk_mem_charge(sk, delta);
4361 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4362 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4363 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4364 TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
4366 if (TCP_SKB_CB(from)->has_rxtstamp) {
4367 TCP_SKB_CB(to)->has_rxtstamp = true;
4368 if (dest == OOO_QUEUE)
4369 TCP_SKB_CB(to)->swtstamp = TCP_SKB_CB(from)->swtstamp;
4371 to->tstamp = from->tstamp;
4377 static bool tcp_ooo_try_coalesce(struct sock *sk,
4379 struct sk_buff *from,
4382 bool res = tcp_try_coalesce(sk, OOO_QUEUE, to, from, fragstolen);
4384 /* In case tcp_drop() is called later, update to->gso_segs */
4386 u32 gso_segs = max_t(u16, 1, skb_shinfo(to)->gso_segs) +
4387 max_t(u16, 1, skb_shinfo(from)->gso_segs);
4389 skb_shinfo(to)->gso_segs = min_t(u32, gso_segs, 0xFFFF);
4394 static void tcp_drop(struct sock *sk, struct sk_buff *skb)
4396 sk_drops_add(sk, skb);
4400 /* This one checks to see if we can put data from the
4401 * out_of_order queue into the receive_queue.
4403 static void tcp_ofo_queue(struct sock *sk)
4405 struct tcp_sock *tp = tcp_sk(sk);
4406 __u32 dsack_high = tp->rcv_nxt;
4407 bool fin, fragstolen, eaten;
4408 struct sk_buff *skb, *tail;
4411 p = rb_first(&tp->out_of_order_queue);
4414 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4417 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4418 __u32 dsack = dsack_high;
4419 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4420 dsack_high = TCP_SKB_CB(skb)->end_seq;
4421 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4424 rb_erase(&skb->rbnode, &tp->out_of_order_queue);
4425 /* Replace tstamp which was stomped by rbnode */
4426 if (TCP_SKB_CB(skb)->has_rxtstamp)
4427 skb->tstamp = TCP_SKB_CB(skb)->swtstamp;
4429 if (unlikely(!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))) {
4430 SOCK_DEBUG(sk, "ofo packet was already received\n");
4434 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4435 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4436 TCP_SKB_CB(skb)->end_seq);
4438 tail = skb_peek_tail(&sk->sk_receive_queue);
4439 eaten = tail && tcp_try_coalesce(sk, RCV_QUEUE,
4440 tail, skb, &fragstolen);
4441 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4442 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
4444 __skb_queue_tail(&sk->sk_receive_queue, skb);
4446 kfree_skb_partial(skb, fragstolen);
4448 if (unlikely(fin)) {
4450 /* tcp_fin() purges tp->out_of_order_queue,
4451 * so we must end this loop right now.
4458 static bool tcp_prune_ofo_queue(struct sock *sk);
4459 static int tcp_prune_queue(struct sock *sk);
4461 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4464 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4465 !sk_rmem_schedule(sk, skb, size)) {
4467 if (tcp_prune_queue(sk) < 0)
4470 while (!sk_rmem_schedule(sk, skb, size)) {
4471 if (!tcp_prune_ofo_queue(sk))
4478 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4480 struct tcp_sock *tp = tcp_sk(sk);
4481 struct rb_node **p, *parent;
4482 struct sk_buff *skb1;
4486 tcp_ecn_check_ce(sk, skb);
4488 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4489 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFODROP);
4494 /* Stash tstamp to avoid being stomped on by rbnode */
4495 if (TCP_SKB_CB(skb)->has_rxtstamp)
4496 TCP_SKB_CB(skb)->swtstamp = skb->tstamp;
4498 /* Disable header prediction. */
4500 inet_csk_schedule_ack(sk);
4502 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4503 seq = TCP_SKB_CB(skb)->seq;
4504 end_seq = TCP_SKB_CB(skb)->end_seq;
4505 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4506 tp->rcv_nxt, seq, end_seq);
4508 p = &tp->out_of_order_queue.rb_node;
4509 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4510 /* Initial out of order segment, build 1 SACK. */
4511 if (tcp_is_sack(tp)) {
4512 tp->rx_opt.num_sacks = 1;
4513 tp->selective_acks[0].start_seq = seq;
4514 tp->selective_acks[0].end_seq = end_seq;
4516 rb_link_node(&skb->rbnode, NULL, p);
4517 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4518 tp->ooo_last_skb = skb;
4522 /* In the typical case, we are adding an skb to the end of the list.
4523 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
4525 if (tcp_ooo_try_coalesce(sk, tp->ooo_last_skb,
4526 skb, &fragstolen)) {
4528 /* For non sack flows, do not grow window to force DUPACK
4529 * and trigger fast retransmit.
4531 if (tcp_is_sack(tp))
4532 tcp_grow_window(sk, skb);
4533 kfree_skb_partial(skb, fragstolen);
4537 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
4538 if (!before(seq, TCP_SKB_CB(tp->ooo_last_skb)->end_seq)) {
4539 parent = &tp->ooo_last_skb->rbnode;
4540 p = &parent->rb_right;
4544 /* Find place to insert this segment. Handle overlaps on the way. */
4548 skb1 = rb_to_skb(parent);
4549 if (before(seq, TCP_SKB_CB(skb1)->seq)) {
4550 p = &parent->rb_left;
4553 if (before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4554 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4555 /* All the bits are present. Drop. */
4556 NET_INC_STATS(sock_net(sk),
4557 LINUX_MIB_TCPOFOMERGE);
4560 tcp_dsack_set(sk, seq, end_seq);
4563 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4564 /* Partial overlap. */
4565 tcp_dsack_set(sk, seq, TCP_SKB_CB(skb1)->end_seq);
4567 /* skb's seq == skb1's seq and skb covers skb1.
4568 * Replace skb1 with skb.
4570 rb_replace_node(&skb1->rbnode, &skb->rbnode,
4571 &tp->out_of_order_queue);
4572 tcp_dsack_extend(sk,
4573 TCP_SKB_CB(skb1)->seq,
4574 TCP_SKB_CB(skb1)->end_seq);
4575 NET_INC_STATS(sock_net(sk),
4576 LINUX_MIB_TCPOFOMERGE);
4580 } else if (tcp_ooo_try_coalesce(sk, skb1,
4581 skb, &fragstolen)) {
4584 p = &parent->rb_right;
4587 /* Insert segment into RB tree. */
4588 rb_link_node(&skb->rbnode, parent, p);
4589 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4592 /* Remove other segments covered by skb. */
4593 while ((skb1 = skb_rb_next(skb)) != NULL) {
4594 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4596 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4597 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4601 rb_erase(&skb1->rbnode, &tp->out_of_order_queue);
4602 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4603 TCP_SKB_CB(skb1)->end_seq);
4604 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4607 /* If there is no skb after us, we are the last_skb ! */
4609 tp->ooo_last_skb = skb;
4612 if (tcp_is_sack(tp))
4613 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4616 /* For non sack flows, do not grow window to force DUPACK
4617 * and trigger fast retransmit.
4619 if (tcp_is_sack(tp))
4620 tcp_grow_window(sk, skb);
4622 skb_set_owner_r(skb, sk);
4626 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, int hdrlen,
4630 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4632 __skb_pull(skb, hdrlen);
4634 tcp_try_coalesce(sk, RCV_QUEUE, tail,
4635 skb, fragstolen)) ? 1 : 0;
4636 tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq);
4638 __skb_queue_tail(&sk->sk_receive_queue, skb);
4639 skb_set_owner_r(skb, sk);
4644 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4646 struct sk_buff *skb;
4654 if (size > PAGE_SIZE) {
4655 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS);
4657 data_len = npages << PAGE_SHIFT;
4658 size = data_len + (size & ~PAGE_MASK);
4660 skb = alloc_skb_with_frags(size - data_len, data_len,
4661 PAGE_ALLOC_COSTLY_ORDER,
4662 &err, sk->sk_allocation);
4666 skb_put(skb, size - data_len);
4667 skb->data_len = data_len;
4670 if (tcp_try_rmem_schedule(sk, skb, skb->truesize))
4673 err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size);
4677 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4678 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4679 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4681 if (tcp_queue_rcv(sk, skb, 0, &fragstolen)) {
4682 WARN_ON_ONCE(fragstolen); /* should not happen */
4694 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4696 struct tcp_sock *tp = tcp_sk(sk);
4700 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
4705 __skb_pull(skb, tcp_hdr(skb)->doff * 4);
4707 tcp_ecn_accept_cwr(tp, skb);
4709 tp->rx_opt.dsack = 0;
4711 /* Queue data for delivery to the user.
4712 * Packets in sequence go to the receive queue.
4713 * Out of sequence packets to the out_of_order_queue.
4715 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4716 if (tcp_receive_window(tp) == 0)
4719 /* Ok. In sequence. In window. */
4721 if (skb_queue_len(&sk->sk_receive_queue) == 0)
4722 sk_forced_mem_schedule(sk, skb->truesize);
4723 else if (tcp_try_rmem_schedule(sk, skb, skb->truesize))
4726 eaten = tcp_queue_rcv(sk, skb, 0, &fragstolen);
4727 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4729 tcp_event_data_recv(sk, skb);
4730 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
4733 if (!RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4736 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4737 * gap in queue is filled.
4739 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
4740 inet_csk(sk)->icsk_ack.pingpong = 0;
4743 if (tp->rx_opt.num_sacks)
4744 tcp_sack_remove(tp);
4746 tcp_fast_path_check(sk);
4749 kfree_skb_partial(skb, fragstolen);
4750 if (!sock_flag(sk, SOCK_DEAD))
4751 sk->sk_data_ready(sk);
4755 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4756 /* A retransmit, 2nd most common case. Force an immediate ack. */
4757 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4758 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4761 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
4762 inet_csk_schedule_ack(sk);
4768 /* Out of window. F.e. zero window probe. */
4769 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4772 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4773 /* Partial packet, seq < rcv_next < end_seq */
4774 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4775 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4776 TCP_SKB_CB(skb)->end_seq);
4778 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4780 /* If window is closed, drop tail of packet. But after
4781 * remembering D-SACK for its head made in previous line.
4783 if (!tcp_receive_window(tp))
4788 tcp_data_queue_ofo(sk, skb);
4791 static struct sk_buff *tcp_skb_next(struct sk_buff *skb, struct sk_buff_head *list)
4794 return !skb_queue_is_last(list, skb) ? skb->next : NULL;
4796 return skb_rb_next(skb);
4799 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4800 struct sk_buff_head *list,
4801 struct rb_root *root)
4803 struct sk_buff *next = tcp_skb_next(skb, list);
4806 __skb_unlink(skb, list);
4808 rb_erase(&skb->rbnode, root);
4811 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4816 /* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */
4817 static void tcp_rbtree_insert(struct rb_root *root, struct sk_buff *skb)
4819 struct rb_node **p = &root->rb_node;
4820 struct rb_node *parent = NULL;
4821 struct sk_buff *skb1;
4825 skb1 = rb_to_skb(parent);
4826 if (before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb1)->seq))
4827 p = &parent->rb_left;
4829 p = &parent->rb_right;
4831 rb_link_node(&skb->rbnode, parent, p);
4832 rb_insert_color(&skb->rbnode, root);
4835 /* Collapse contiguous sequence of skbs head..tail with
4836 * sequence numbers start..end.
4838 * If tail is NULL, this means until the end of the queue.
4840 * Segments with FIN/SYN are not collapsed (only because this
4844 tcp_collapse(struct sock *sk, struct sk_buff_head *list, struct rb_root *root,
4845 struct sk_buff *head, struct sk_buff *tail, u32 start, u32 end)
4847 struct sk_buff *skb = head, *n;
4848 struct sk_buff_head tmp;
4851 /* First, check that queue is collapsible and find
4852 * the point where collapsing can be useful.
4855 for (end_of_skbs = true; skb != NULL && skb != tail; skb = n) {
4856 n = tcp_skb_next(skb, list);
4858 /* No new bits? It is possible on ofo queue. */
4859 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4860 skb = tcp_collapse_one(sk, skb, list, root);
4866 /* The first skb to collapse is:
4868 * - bloated or contains data before "start" or
4869 * overlaps to the next one.
4871 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
4872 (tcp_win_from_space(skb->truesize) > skb->len ||
4873 before(TCP_SKB_CB(skb)->seq, start))) {
4874 end_of_skbs = false;
4878 if (n && n != tail &&
4879 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(n)->seq) {
4880 end_of_skbs = false;
4884 /* Decided to skip this, advance start seq. */
4885 start = TCP_SKB_CB(skb)->end_seq;
4888 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4891 __skb_queue_head_init(&tmp);
4893 while (before(start, end)) {
4894 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
4895 struct sk_buff *nskb;
4897 nskb = alloc_skb(copy, GFP_ATOMIC);
4901 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4902 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4904 __skb_queue_before(list, skb, nskb);
4906 __skb_queue_tail(&tmp, nskb); /* defer rbtree insertion */
4907 skb_set_owner_r(nskb, sk);
4909 /* Copy data, releasing collapsed skbs. */
4911 int offset = start - TCP_SKB_CB(skb)->seq;
4912 int size = TCP_SKB_CB(skb)->end_seq - start;
4916 size = min(copy, size);
4917 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4919 TCP_SKB_CB(nskb)->end_seq += size;
4923 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4924 skb = tcp_collapse_one(sk, skb, list, root);
4927 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4933 skb_queue_walk_safe(&tmp, skb, n)
4934 tcp_rbtree_insert(root, skb);
4937 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4938 * and tcp_collapse() them until all the queue is collapsed.
4940 static void tcp_collapse_ofo_queue(struct sock *sk)
4942 struct tcp_sock *tp = tcp_sk(sk);
4943 u32 range_truesize, sum_tiny = 0;
4944 struct sk_buff *skb, *head;
4947 skb = skb_rb_first(&tp->out_of_order_queue);
4950 tp->ooo_last_skb = skb_rb_last(&tp->out_of_order_queue);
4953 start = TCP_SKB_CB(skb)->seq;
4954 end = TCP_SKB_CB(skb)->end_seq;
4955 range_truesize = skb->truesize;
4957 for (head = skb;;) {
4958 skb = skb_rb_next(skb);
4960 /* Range is terminated when we see a gap or when
4961 * we are at the queue end.
4964 after(TCP_SKB_CB(skb)->seq, end) ||
4965 before(TCP_SKB_CB(skb)->end_seq, start)) {
4966 /* Do not attempt collapsing tiny skbs */
4967 if (range_truesize != head->truesize ||
4968 end - start >= SKB_WITH_OVERHEAD(SK_MEM_QUANTUM)) {
4969 tcp_collapse(sk, NULL, &tp->out_of_order_queue,
4970 head, skb, start, end);
4972 sum_tiny += range_truesize;
4973 if (sum_tiny > sk->sk_rcvbuf >> 3)
4979 range_truesize += skb->truesize;
4980 if (unlikely(before(TCP_SKB_CB(skb)->seq, start)))
4981 start = TCP_SKB_CB(skb)->seq;
4982 if (after(TCP_SKB_CB(skb)->end_seq, end))
4983 end = TCP_SKB_CB(skb)->end_seq;
4988 * Clean the out-of-order queue to make room.
4989 * We drop high sequences packets to :
4990 * 1) Let a chance for holes to be filled.
4991 * 2) not add too big latencies if thousands of packets sit there.
4992 * (But if application shrinks SO_RCVBUF, we could still end up
4993 * freeing whole queue here)
4994 * 3) Drop at least 12.5 % of sk_rcvbuf to avoid malicious attacks.
4996 * Return true if queue has shrunk.
4998 static bool tcp_prune_ofo_queue(struct sock *sk)
5000 struct tcp_sock *tp = tcp_sk(sk);
5001 struct rb_node *node, *prev;
5004 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
5007 NET_INC_STATS(sock_net(sk), LINUX_MIB_OFOPRUNED);
5008 goal = sk->sk_rcvbuf >> 3;
5009 node = &tp->ooo_last_skb->rbnode;
5011 prev = rb_prev(node);
5012 rb_erase(node, &tp->out_of_order_queue);
5013 goal -= rb_to_skb(node)->truesize;
5014 tcp_drop(sk, rb_to_skb(node));
5015 if (!prev || goal <= 0) {
5017 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf &&
5018 !tcp_under_memory_pressure(sk))
5020 goal = sk->sk_rcvbuf >> 3;
5024 tp->ooo_last_skb = rb_to_skb(prev);
5026 /* Reset SACK state. A conforming SACK implementation will
5027 * do the same at a timeout based retransmit. When a connection
5028 * is in a sad state like this, we care only about integrity
5029 * of the connection not performance.
5031 if (tp->rx_opt.sack_ok)
5032 tcp_sack_reset(&tp->rx_opt);
5036 /* Reduce allocated memory if we can, trying to get
5037 * the socket within its memory limits again.
5039 * Return less than zero if we should start dropping frames
5040 * until the socket owning process reads some of the data
5041 * to stabilize the situation.
5043 static int tcp_prune_queue(struct sock *sk)
5045 struct tcp_sock *tp = tcp_sk(sk);
5047 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
5049 NET_INC_STATS(sock_net(sk), LINUX_MIB_PRUNECALLED);
5051 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
5052 tcp_clamp_window(sk);
5053 else if (tcp_under_memory_pressure(sk))
5054 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
5056 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5059 tcp_collapse_ofo_queue(sk);
5060 if (!skb_queue_empty(&sk->sk_receive_queue))
5061 tcp_collapse(sk, &sk->sk_receive_queue, NULL,
5062 skb_peek(&sk->sk_receive_queue),
5064 tp->copied_seq, tp->rcv_nxt);
5067 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5070 /* Collapsing did not help, destructive actions follow.
5071 * This must not ever occur. */
5073 tcp_prune_ofo_queue(sk);
5075 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5078 /* If we are really being abused, tell the caller to silently
5079 * drop receive data on the floor. It will get retransmitted
5080 * and hopefully then we'll have sufficient space.
5082 NET_INC_STATS(sock_net(sk), LINUX_MIB_RCVPRUNED);
5084 /* Massive buffer overcommit. */
5089 static bool tcp_should_expand_sndbuf(const struct sock *sk)
5091 const struct tcp_sock *tp = tcp_sk(sk);
5093 /* If the user specified a specific send buffer setting, do
5096 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
5099 /* If we are under global TCP memory pressure, do not expand. */
5100 if (tcp_under_memory_pressure(sk))
5103 /* If we are under soft global TCP memory pressure, do not expand. */
5104 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
5107 /* If we filled the congestion window, do not expand. */
5108 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
5114 /* When incoming ACK allowed to free some skb from write_queue,
5115 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
5116 * on the exit from tcp input handler.
5118 * PROBLEM: sndbuf expansion does not work well with largesend.
5120 static void tcp_new_space(struct sock *sk)
5122 struct tcp_sock *tp = tcp_sk(sk);
5124 if (tcp_should_expand_sndbuf(sk)) {
5125 tcp_sndbuf_expand(sk);
5126 tp->snd_cwnd_stamp = tcp_jiffies32;
5129 sk->sk_write_space(sk);
5132 /* Caller made space either from:
5133 * 1) Freeing skbs in rtx queues (after tp->snd_una has advanced)
5134 * 2) Sent skbs from output queue (and thus advancing tp->snd_nxt)
5136 * We might be able to generate EPOLLOUT to the application if:
5137 * 1) Space consumed in output/rtx queues is below sk->sk_sndbuf/2
5138 * 2) notsent amount (tp->write_seq - tp->snd_nxt) became
5139 * small enough that tcp_stream_memory_free() decides it
5140 * is time to generate EPOLLOUT.
5142 void tcp_check_space(struct sock *sk)
5144 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
5145 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
5146 /* pairs with tcp_poll() */
5148 if (sk->sk_socket &&
5149 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
5151 if (!test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
5152 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
5157 static inline void tcp_data_snd_check(struct sock *sk)
5159 tcp_push_pending_frames(sk);
5160 tcp_check_space(sk);
5164 * Check if sending an ack is needed.
5166 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
5168 struct tcp_sock *tp = tcp_sk(sk);
5170 /* More than one full frame received... */
5171 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
5172 /* ... and right edge of window advances far enough.
5173 * (tcp_recvmsg() will send ACK otherwise). Or...
5175 __tcp_select_window(sk) >= tp->rcv_wnd) ||
5176 /* We ACK each frame or... */
5177 tcp_in_quickack_mode(sk) ||
5178 /* We have out of order data. */
5179 (ofo_possible && !RB_EMPTY_ROOT(&tp->out_of_order_queue))) {
5180 /* Then ack it now */
5183 /* Else, send delayed ack. */
5184 tcp_send_delayed_ack(sk);
5188 static inline void tcp_ack_snd_check(struct sock *sk)
5190 if (!inet_csk_ack_scheduled(sk)) {
5191 /* We sent a data segment already. */
5194 __tcp_ack_snd_check(sk, 1);
5198 * This routine is only called when we have urgent data
5199 * signaled. Its the 'slow' part of tcp_urg. It could be
5200 * moved inline now as tcp_urg is only called from one
5201 * place. We handle URGent data wrong. We have to - as
5202 * BSD still doesn't use the correction from RFC961.
5203 * For 1003.1g we should support a new option TCP_STDURG to permit
5204 * either form (or just set the sysctl tcp_stdurg).
5207 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
5209 struct tcp_sock *tp = tcp_sk(sk);
5210 u32 ptr = ntohs(th->urg_ptr);
5212 if (ptr && !sysctl_tcp_stdurg)
5214 ptr += ntohl(th->seq);
5216 /* Ignore urgent data that we've already seen and read. */
5217 if (after(tp->copied_seq, ptr))
5220 /* Do not replay urg ptr.
5222 * NOTE: interesting situation not covered by specs.
5223 * Misbehaving sender may send urg ptr, pointing to segment,
5224 * which we already have in ofo queue. We are not able to fetch
5225 * such data and will stay in TCP_URG_NOTYET until will be eaten
5226 * by recvmsg(). Seems, we are not obliged to handle such wicked
5227 * situations. But it is worth to think about possibility of some
5228 * DoSes using some hypothetical application level deadlock.
5230 if (before(ptr, tp->rcv_nxt))
5233 /* Do we already have a newer (or duplicate) urgent pointer? */
5234 if (tp->urg_data && !after(ptr, tp->urg_seq))
5237 /* Tell the world about our new urgent pointer. */
5240 /* We may be adding urgent data when the last byte read was
5241 * urgent. To do this requires some care. We cannot just ignore
5242 * tp->copied_seq since we would read the last urgent byte again
5243 * as data, nor can we alter copied_seq until this data arrives
5244 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5246 * NOTE. Double Dutch. Rendering to plain English: author of comment
5247 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5248 * and expect that both A and B disappear from stream. This is _wrong_.
5249 * Though this happens in BSD with high probability, this is occasional.
5250 * Any application relying on this is buggy. Note also, that fix "works"
5251 * only in this artificial test. Insert some normal data between A and B and we will
5252 * decline of BSD again. Verdict: it is better to remove to trap
5255 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5256 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5257 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5259 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5260 __skb_unlink(skb, &sk->sk_receive_queue);
5265 tp->urg_data = TCP_URG_NOTYET;
5268 /* Disable header prediction. */
5272 /* This is the 'fast' part of urgent handling. */
5273 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
5275 struct tcp_sock *tp = tcp_sk(sk);
5277 /* Check if we get a new urgent pointer - normally not. */
5279 tcp_check_urg(sk, th);
5281 /* Do we wait for any urgent data? - normally not... */
5282 if (tp->urg_data == TCP_URG_NOTYET) {
5283 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5286 /* Is the urgent pointer pointing into this packet? */
5287 if (ptr < skb->len) {
5289 if (skb_copy_bits(skb, ptr, &tmp, 1))
5291 tp->urg_data = TCP_URG_VALID | tmp;
5292 if (!sock_flag(sk, SOCK_DEAD))
5293 sk->sk_data_ready(sk);
5298 /* Accept RST for rcv_nxt - 1 after a FIN.
5299 * When tcp connections are abruptly terminated from Mac OSX (via ^C), a
5300 * FIN is sent followed by a RST packet. The RST is sent with the same
5301 * sequence number as the FIN, and thus according to RFC 5961 a challenge
5302 * ACK should be sent. However, Mac OSX rate limits replies to challenge
5303 * ACKs on the closed socket. In addition middleboxes can drop either the
5304 * challenge ACK or a subsequent RST.
5306 static bool tcp_reset_check(const struct sock *sk, const struct sk_buff *skb)
5308 struct tcp_sock *tp = tcp_sk(sk);
5310 return unlikely(TCP_SKB_CB(skb)->seq == (tp->rcv_nxt - 1) &&
5311 (1 << sk->sk_state) & (TCPF_CLOSE_WAIT | TCPF_LAST_ACK |
5315 /* Does PAWS and seqno based validation of an incoming segment, flags will
5316 * play significant role here.
5318 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5319 const struct tcphdr *th, int syn_inerr)
5321 struct tcp_sock *tp = tcp_sk(sk);
5322 bool rst_seq_match = false;
5324 /* RFC1323: H1. Apply PAWS check first. */
5325 if (tcp_fast_parse_options(sock_net(sk), skb, th, tp) &&
5326 tp->rx_opt.saw_tstamp &&
5327 tcp_paws_discard(sk, skb)) {
5329 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5330 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5331 LINUX_MIB_TCPACKSKIPPEDPAWS,
5332 &tp->last_oow_ack_time))
5333 tcp_send_dupack(sk, skb);
5336 /* Reset is accepted even if it did not pass PAWS. */
5339 /* Step 1: check sequence number */
5340 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5341 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5342 * (RST) segments are validated by checking their SEQ-fields."
5343 * And page 69: "If an incoming segment is not acceptable,
5344 * an acknowledgment should be sent in reply (unless the RST
5345 * bit is set, if so drop the segment and return)".
5350 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5351 LINUX_MIB_TCPACKSKIPPEDSEQ,
5352 &tp->last_oow_ack_time))
5353 tcp_send_dupack(sk, skb);
5354 } else if (tcp_reset_check(sk, skb)) {
5360 /* Step 2: check RST bit */
5362 /* RFC 5961 3.2 (extend to match against (RCV.NXT - 1) after a
5363 * FIN and SACK too if available):
5364 * If seq num matches RCV.NXT or (RCV.NXT - 1) after a FIN, or
5365 * the right-most SACK block,
5367 * RESET the connection
5369 * Send a challenge ACK
5371 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt ||
5372 tcp_reset_check(sk, skb)) {
5373 rst_seq_match = true;
5374 } else if (tcp_is_sack(tp) && tp->rx_opt.num_sacks > 0) {
5375 struct tcp_sack_block *sp = &tp->selective_acks[0];
5376 int max_sack = sp[0].end_seq;
5379 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;
5381 max_sack = after(sp[this_sack].end_seq,
5383 sp[this_sack].end_seq : max_sack;
5386 if (TCP_SKB_CB(skb)->seq == max_sack)
5387 rst_seq_match = true;
5393 /* Disable TFO if RST is out-of-order
5394 * and no data has been received
5395 * for current active TFO socket
5397 if (tp->syn_fastopen && !tp->data_segs_in &&
5398 sk->sk_state == TCP_ESTABLISHED)
5399 tcp_fastopen_active_disable(sk);
5400 tcp_send_challenge_ack(sk, skb);
5405 /* step 3: check security and precedence [ignored] */
5407 /* step 4: Check for a SYN
5408 * RFC 5961 4.2 : Send a challenge ack
5413 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5414 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5415 tcp_send_challenge_ack(sk, skb);
5427 * TCP receive function for the ESTABLISHED state.
5429 * It is split into a fast path and a slow path. The fast path is
5431 * - A zero window was announced from us - zero window probing
5432 * is only handled properly in the slow path.
5433 * - Out of order segments arrived.
5434 * - Urgent data is expected.
5435 * - There is no buffer space left
5436 * - Unexpected TCP flags/window values/header lengths are received
5437 * (detected by checking the TCP header against pred_flags)
5438 * - Data is sent in both directions. Fast path only supports pure senders
5439 * or pure receivers (this means either the sequence number or the ack
5440 * value must stay constant)
5441 * - Unexpected TCP option.
5443 * When these conditions are not satisfied it drops into a standard
5444 * receive procedure patterned after RFC793 to handle all cases.
5445 * The first three cases are guaranteed by proper pred_flags setting,
5446 * the rest is checked inline. Fast processing is turned on in
5447 * tcp_data_queue when everything is OK.
5449 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5450 const struct tcphdr *th)
5452 unsigned int len = skb->len;
5453 struct tcp_sock *tp = tcp_sk(sk);
5455 tcp_mstamp_refresh(tp);
5456 if (unlikely(!rcu_access_pointer(sk->sk_rx_dst)))
5457 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
5459 * Header prediction.
5460 * The code loosely follows the one in the famous
5461 * "30 instruction TCP receive" Van Jacobson mail.
5463 * Van's trick is to deposit buffers into socket queue
5464 * on a device interrupt, to call tcp_recv function
5465 * on the receive process context and checksum and copy
5466 * the buffer to user space. smart...
5468 * Our current scheme is not silly either but we take the
5469 * extra cost of the net_bh soft interrupt processing...
5470 * We do checksum and copy also but from device to kernel.
5473 tp->rx_opt.saw_tstamp = 0;
5475 /* pred_flags is 0xS?10 << 16 + snd_wnd
5476 * if header_prediction is to be made
5477 * 'S' will always be tp->tcp_header_len >> 2
5478 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5479 * turn it off (when there are holes in the receive
5480 * space for instance)
5481 * PSH flag is ignored.
5484 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5485 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5486 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5487 int tcp_header_len = tp->tcp_header_len;
5489 /* Timestamp header prediction: tcp_header_len
5490 * is automatically equal to th->doff*4 due to pred_flags
5494 /* Check timestamp */
5495 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5496 /* No? Slow path! */
5497 if (!tcp_parse_aligned_timestamp(tp, th))
5500 /* If PAWS failed, check it more carefully in slow path */
5501 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5504 /* DO NOT update ts_recent here, if checksum fails
5505 * and timestamp was corrupted part, it will result
5506 * in a hung connection since we will drop all
5507 * future packets due to the PAWS test.
5511 if (len <= tcp_header_len) {
5512 /* Bulk data transfer: sender */
5513 if (len == tcp_header_len) {
5514 /* Predicted packet is in window by definition.
5515 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5516 * Hence, check seq<=rcv_wup reduces to:
5518 if (tcp_header_len ==
5519 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5520 tp->rcv_nxt == tp->rcv_wup)
5521 tcp_store_ts_recent(tp);
5523 /* We know that such packets are checksummed
5526 tcp_ack(sk, skb, 0);
5528 tcp_data_snd_check(sk);
5530 } else { /* Header too small */
5531 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5536 bool fragstolen = false;
5538 if (tcp_checksum_complete(skb))
5541 if ((int)skb->truesize > sk->sk_forward_alloc)
5544 /* Predicted packet is in window by definition.
5545 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5546 * Hence, check seq<=rcv_wup reduces to:
5548 if (tcp_header_len ==
5549 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5550 tp->rcv_nxt == tp->rcv_wup)
5551 tcp_store_ts_recent(tp);
5553 tcp_rcv_rtt_measure_ts(sk, skb);
5555 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPHITS);
5557 /* Bulk data transfer: receiver */
5558 eaten = tcp_queue_rcv(sk, skb, tcp_header_len,
5561 tcp_event_data_recv(sk, skb);
5563 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5564 /* Well, only one small jumplet in fast path... */
5565 tcp_ack(sk, skb, FLAG_DATA);
5566 tcp_data_snd_check(sk);
5567 if (!inet_csk_ack_scheduled(sk))
5570 tcp_update_wl(tp, TCP_SKB_CB(skb)->seq);
5573 __tcp_ack_snd_check(sk, 0);
5576 kfree_skb_partial(skb, fragstolen);
5577 sk->sk_data_ready(sk);
5583 if (len < (th->doff << 2) || tcp_checksum_complete(skb))
5586 if (!th->ack && !th->rst && !th->syn)
5590 * Standard slow path.
5593 if (!tcp_validate_incoming(sk, skb, th, 1))
5597 if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0)
5600 tcp_rcv_rtt_measure_ts(sk, skb);
5602 /* Process urgent data. */
5603 tcp_urg(sk, skb, th);
5605 /* step 7: process the segment text */
5606 tcp_data_queue(sk, skb);
5608 tcp_data_snd_check(sk);
5609 tcp_ack_snd_check(sk);
5613 TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS);
5614 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5619 EXPORT_SYMBOL(tcp_rcv_established);
5621 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5623 struct tcp_sock *tp = tcp_sk(sk);
5624 struct inet_connection_sock *icsk = inet_csk(sk);
5626 tcp_set_state(sk, TCP_ESTABLISHED);
5627 icsk->icsk_ack.lrcvtime = tcp_jiffies32;
5630 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
5631 security_inet_conn_established(sk, skb);
5634 /* Make sure socket is routed, for correct metrics. */
5635 icsk->icsk_af_ops->rebuild_header(sk);
5637 tcp_init_metrics(sk);
5638 tcp_call_bpf(sk, BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB);
5639 tcp_init_congestion_control(sk);
5641 /* Prevent spurious tcp_cwnd_restart() on first data
5644 tp->lsndtime = tcp_jiffies32;
5646 tcp_init_buffer_space(sk);
5648 if (sock_flag(sk, SOCK_KEEPOPEN))
5649 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5651 if (!tp->rx_opt.snd_wscale)
5652 __tcp_fast_path_on(tp, tp->snd_wnd);
5657 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
5658 struct tcp_fastopen_cookie *cookie)
5660 struct tcp_sock *tp = tcp_sk(sk);
5661 struct sk_buff *data = tp->syn_data ? tcp_write_queue_head(sk) : NULL;
5662 u16 mss = tp->rx_opt.mss_clamp, try_exp = 0;
5663 bool syn_drop = false;
5665 if (mss == tp->rx_opt.user_mss) {
5666 struct tcp_options_received opt;
5668 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5669 tcp_clear_options(&opt);
5670 opt.user_mss = opt.mss_clamp = 0;
5671 tcp_parse_options(sock_net(sk), synack, &opt, 0, NULL);
5672 mss = opt.mss_clamp;
5675 if (!tp->syn_fastopen) {
5676 /* Ignore an unsolicited cookie */
5678 } else if (tp->total_retrans) {
5679 /* SYN timed out and the SYN-ACK neither has a cookie nor
5680 * acknowledges data. Presumably the remote received only
5681 * the retransmitted (regular) SYNs: either the original
5682 * SYN-data or the corresponding SYN-ACK was dropped.
5684 syn_drop = (cookie->len < 0 && data);
5685 } else if (cookie->len < 0 && !tp->syn_data) {
5686 /* We requested a cookie but didn't get it. If we did not use
5687 * the (old) exp opt format then try so next time (try_exp=1).
5688 * Otherwise we go back to use the RFC7413 opt (try_exp=2).
5690 try_exp = tp->syn_fastopen_exp ? 2 : 1;
5693 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp);
5695 if (data) { /* Retransmit unacked data in SYN */
5696 tcp_for_write_queue_from(data, sk) {
5697 if (data == tcp_send_head(sk) ||
5698 __tcp_retransmit_skb(sk, data, 1))
5702 NET_INC_STATS(sock_net(sk),
5703 LINUX_MIB_TCPFASTOPENACTIVEFAIL);
5706 tp->syn_data_acked = tp->syn_data;
5707 if (tp->syn_data_acked)
5708 NET_INC_STATS(sock_net(sk),
5709 LINUX_MIB_TCPFASTOPENACTIVE);
5711 tcp_fastopen_add_skb(sk, synack);
5716 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5717 const struct tcphdr *th)
5719 struct inet_connection_sock *icsk = inet_csk(sk);
5720 struct tcp_sock *tp = tcp_sk(sk);
5721 struct tcp_fastopen_cookie foc = { .len = -1 };
5722 int saved_clamp = tp->rx_opt.mss_clamp;
5725 tcp_parse_options(sock_net(sk), skb, &tp->rx_opt, 0, &foc);
5726 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
5727 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
5731 * "If the state is SYN-SENT then
5732 * first check the ACK bit
5733 * If the ACK bit is set
5734 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5735 * a reset (unless the RST bit is set, if so drop
5736 * the segment and return)"
5738 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
5739 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt))
5740 goto reset_and_undo;
5742 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5743 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5744 tcp_time_stamp(tp))) {
5745 NET_INC_STATS(sock_net(sk),
5746 LINUX_MIB_PAWSACTIVEREJECTED);
5747 goto reset_and_undo;
5750 /* Now ACK is acceptable.
5752 * "If the RST bit is set
5753 * If the ACK was acceptable then signal the user "error:
5754 * connection reset", drop the segment, enter CLOSED state,
5755 * delete TCB, and return."
5764 * "fifth, if neither of the SYN or RST bits is set then
5765 * drop the segment and return."
5771 goto discard_and_undo;
5774 * "If the SYN bit is on ...
5775 * are acceptable then ...
5776 * (our SYN has been ACKed), change the connection
5777 * state to ESTABLISHED..."
5780 tcp_ecn_rcv_synack(tp, th);
5782 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5783 tcp_ack(sk, skb, FLAG_SLOWPATH);
5785 /* Ok.. it's good. Set up sequence numbers and
5786 * move to established.
5788 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5789 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5791 /* RFC1323: The window in SYN & SYN/ACK segments is
5794 tp->snd_wnd = ntohs(th->window);
5796 if (!tp->rx_opt.wscale_ok) {
5797 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5798 tp->window_clamp = min(tp->window_clamp, 65535U);
5801 if (tp->rx_opt.saw_tstamp) {
5802 tp->rx_opt.tstamp_ok = 1;
5803 tp->tcp_header_len =
5804 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5805 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5806 tcp_store_ts_recent(tp);
5808 tp->tcp_header_len = sizeof(struct tcphdr);
5811 if (tcp_is_sack(tp) && sysctl_tcp_fack)
5812 tcp_enable_fack(tp);
5815 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5816 tcp_initialize_rcv_mss(sk);
5818 /* Remember, tcp_poll() does not lock socket!
5819 * Change state from SYN-SENT only after copied_seq
5820 * is initialized. */
5821 tp->copied_seq = tp->rcv_nxt;
5825 tcp_finish_connect(sk, skb);
5827 fastopen_fail = (tp->syn_fastopen || tp->syn_data) &&
5828 tcp_rcv_fastopen_synack(sk, skb, &foc);
5830 if (!sock_flag(sk, SOCK_DEAD)) {
5831 sk->sk_state_change(sk);
5832 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5836 if (sk->sk_write_pending ||
5837 icsk->icsk_accept_queue.rskq_defer_accept ||
5838 icsk->icsk_ack.pingpong) {
5839 /* Save one ACK. Data will be ready after
5840 * several ticks, if write_pending is set.
5842 * It may be deleted, but with this feature tcpdumps
5843 * look so _wonderfully_ clever, that I was not able
5844 * to stand against the temptation 8) --ANK
5846 inet_csk_schedule_ack(sk);
5847 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
5848 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5849 TCP_DELACK_MAX, TCP_RTO_MAX);
5860 /* No ACK in the segment */
5864 * "If the RST bit is set
5866 * Otherwise (no ACK) drop the segment and return."
5869 goto discard_and_undo;
5873 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5874 tcp_paws_reject(&tp->rx_opt, 0))
5875 goto discard_and_undo;
5878 /* We see SYN without ACK. It is attempt of
5879 * simultaneous connect with crossed SYNs.
5880 * Particularly, it can be connect to self.
5882 tcp_set_state(sk, TCP_SYN_RECV);
5884 if (tp->rx_opt.saw_tstamp) {
5885 tp->rx_opt.tstamp_ok = 1;
5886 tcp_store_ts_recent(tp);
5887 tp->tcp_header_len =
5888 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5890 tp->tcp_header_len = sizeof(struct tcphdr);
5893 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5894 tp->copied_seq = tp->rcv_nxt;
5895 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5897 /* RFC1323: The window in SYN & SYN/ACK segments is
5900 tp->snd_wnd = ntohs(th->window);
5901 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5902 tp->max_window = tp->snd_wnd;
5904 tcp_ecn_rcv_syn(tp, th);
5907 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5908 tcp_initialize_rcv_mss(sk);
5910 tcp_send_synack(sk);
5912 /* Note, we could accept data and URG from this segment.
5913 * There are no obstacles to make this (except that we must
5914 * either change tcp_recvmsg() to prevent it from returning data
5915 * before 3WHS completes per RFC793, or employ TCP Fast Open).
5917 * However, if we ignore data in ACKless segments sometimes,
5918 * we have no reasons to accept it sometimes.
5919 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5920 * is not flawless. So, discard packet for sanity.
5921 * Uncomment this return to process the data.
5928 /* "fifth, if neither of the SYN or RST bits is set then
5929 * drop the segment and return."
5933 tcp_clear_options(&tp->rx_opt);
5934 tp->rx_opt.mss_clamp = saved_clamp;
5938 tcp_clear_options(&tp->rx_opt);
5939 tp->rx_opt.mss_clamp = saved_clamp;
5944 * This function implements the receiving procedure of RFC 793 for
5945 * all states except ESTABLISHED and TIME_WAIT.
5946 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5947 * address independent.
5950 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb)
5952 struct tcp_sock *tp = tcp_sk(sk);
5953 struct inet_connection_sock *icsk = inet_csk(sk);
5954 const struct tcphdr *th = tcp_hdr(skb);
5955 struct request_sock *req;
5959 switch (sk->sk_state) {
5973 /* It is possible that we process SYN packets from backlog,
5974 * so we need to make sure to disable BH and RCU right there.
5978 acceptable = icsk->icsk_af_ops->conn_request(sk, skb) >= 0;
5990 tp->rx_opt.saw_tstamp = 0;
5991 tcp_mstamp_refresh(tp);
5992 queued = tcp_rcv_synsent_state_process(sk, skb, th);
5996 /* Do step6 onward by hand. */
5997 tcp_urg(sk, skb, th);
5999 tcp_data_snd_check(sk);
6003 tcp_mstamp_refresh(tp);
6004 tp->rx_opt.saw_tstamp = 0;
6005 req = tp->fastopen_rsk;
6007 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
6008 sk->sk_state != TCP_FIN_WAIT1);
6010 if (!tcp_check_req(sk, skb, req, true))
6014 if (!th->ack && !th->rst && !th->syn)
6017 if (!tcp_validate_incoming(sk, skb, th, 0))
6020 /* step 5: check the ACK field */
6021 acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
6022 FLAG_UPDATE_TS_RECENT |
6023 FLAG_NO_CHALLENGE_ACK) > 0;
6026 if (sk->sk_state == TCP_SYN_RECV)
6027 return 1; /* send one RST */
6028 tcp_send_challenge_ack(sk, skb);
6031 switch (sk->sk_state) {
6034 tcp_synack_rtt_meas(sk, req);
6036 /* Once we leave TCP_SYN_RECV, we no longer need req
6040 inet_csk(sk)->icsk_retransmits = 0;
6041 reqsk_fastopen_remove(sk, req, false);
6043 /* Make sure socket is routed, for correct metrics. */
6044 icsk->icsk_af_ops->rebuild_header(sk);
6045 tcp_call_bpf(sk, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB);
6046 tcp_init_congestion_control(sk);
6049 tp->copied_seq = tp->rcv_nxt;
6050 tcp_init_buffer_space(sk);
6053 tcp_set_state(sk, TCP_ESTABLISHED);
6054 sk->sk_state_change(sk);
6056 /* Note, that this wakeup is only for marginal crossed SYN case.
6057 * Passively open sockets are not waked up, because
6058 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
6061 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
6063 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
6064 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
6065 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
6067 if (tp->rx_opt.tstamp_ok)
6068 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
6071 /* Re-arm the timer because data may have been sent out.
6072 * This is similar to the regular data transmission case
6073 * when new data has just been ack'ed.
6075 * (TFO) - we could try to be more aggressive and
6076 * retransmitting any data sooner based on when they
6081 tcp_init_metrics(sk);
6083 if (!inet_csk(sk)->icsk_ca_ops->cong_control)
6084 tcp_update_pacing_rate(sk);
6086 /* Prevent spurious tcp_cwnd_restart() on first data packet */
6087 tp->lsndtime = tcp_jiffies32;
6089 tcp_initialize_rcv_mss(sk);
6090 tcp_fast_path_on(tp);
6093 case TCP_FIN_WAIT1: {
6096 /* If we enter the TCP_FIN_WAIT1 state and we are a
6097 * Fast Open socket and this is the first acceptable
6098 * ACK we have received, this would have acknowledged
6099 * our SYNACK so stop the SYNACK timer.
6102 /* We no longer need the request sock. */
6103 reqsk_fastopen_remove(sk, req, false);
6106 if (tp->snd_una != tp->write_seq)
6109 tcp_set_state(sk, TCP_FIN_WAIT2);
6110 sk->sk_shutdown |= SEND_SHUTDOWN;
6114 if (!sock_flag(sk, SOCK_DEAD)) {
6115 /* Wake up lingering close() */
6116 sk->sk_state_change(sk);
6120 if (tp->linger2 < 0) {
6122 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6125 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6126 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6127 /* Receive out of order FIN after close() */
6128 if (tp->syn_fastopen && th->fin)
6129 tcp_fastopen_active_disable(sk);
6131 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6135 tmo = tcp_fin_time(sk);
6136 if (tmo > TCP_TIMEWAIT_LEN) {
6137 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
6138 } else if (th->fin || sock_owned_by_user(sk)) {
6139 /* Bad case. We could lose such FIN otherwise.
6140 * It is not a big problem, but it looks confusing
6141 * and not so rare event. We still can lose it now,
6142 * if it spins in bh_lock_sock(), but it is really
6145 inet_csk_reset_keepalive_timer(sk, tmo);
6147 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
6154 if (tp->snd_una == tp->write_seq) {
6155 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
6161 if (tp->snd_una == tp->write_seq) {
6162 tcp_update_metrics(sk);
6169 /* step 6: check the URG bit */
6170 tcp_urg(sk, skb, th);
6172 /* step 7: process the segment text */
6173 switch (sk->sk_state) {
6174 case TCP_CLOSE_WAIT:
6177 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
6181 /* RFC 793 says to queue data in these states,
6182 * RFC 1122 says we MUST send a reset.
6183 * BSD 4.4 also does reset.
6185 if (sk->sk_shutdown & RCV_SHUTDOWN) {
6186 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6187 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6188 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6194 case TCP_ESTABLISHED:
6195 tcp_data_queue(sk, skb);
6200 /* tcp_data could move socket to TIME-WAIT */
6201 if (sk->sk_state != TCP_CLOSE) {
6202 tcp_data_snd_check(sk);
6203 tcp_ack_snd_check(sk);
6212 EXPORT_SYMBOL(tcp_rcv_state_process);
6214 static inline void pr_drop_req(struct request_sock *req, __u16 port, int family)
6216 struct inet_request_sock *ireq = inet_rsk(req);
6218 if (family == AF_INET)
6219 net_dbg_ratelimited("drop open request from %pI4/%u\n",
6220 &ireq->ir_rmt_addr, port);
6221 #if IS_ENABLED(CONFIG_IPV6)
6222 else if (family == AF_INET6)
6223 net_dbg_ratelimited("drop open request from %pI6/%u\n",
6224 &ireq->ir_v6_rmt_addr, port);
6228 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
6230 * If we receive a SYN packet with these bits set, it means a
6231 * network is playing bad games with TOS bits. In order to
6232 * avoid possible false congestion notifications, we disable
6233 * TCP ECN negotiation.
6235 * Exception: tcp_ca wants ECN. This is required for DCTCP
6236 * congestion control: Linux DCTCP asserts ECT on all packets,
6237 * including SYN, which is most optimal solution; however,
6238 * others, such as FreeBSD do not.
6240 static void tcp_ecn_create_request(struct request_sock *req,
6241 const struct sk_buff *skb,
6242 const struct sock *listen_sk,
6243 const struct dst_entry *dst)
6245 const struct tcphdr *th = tcp_hdr(skb);
6246 const struct net *net = sock_net(listen_sk);
6247 bool th_ecn = th->ece && th->cwr;
6254 ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield);
6255 ecn_ok_dst = dst_feature(dst, DST_FEATURE_ECN_MASK);
6256 ecn_ok = net->ipv4.sysctl_tcp_ecn || ecn_ok_dst;
6258 if ((!ect && ecn_ok) || tcp_ca_needs_ecn(listen_sk) ||
6259 (ecn_ok_dst & DST_FEATURE_ECN_CA) ||
6260 tcp_bpf_ca_needs_ecn((struct sock *)req))
6261 inet_rsk(req)->ecn_ok = 1;
6264 static void tcp_openreq_init(struct request_sock *req,
6265 const struct tcp_options_received *rx_opt,
6266 struct sk_buff *skb, const struct sock *sk)
6268 struct inet_request_sock *ireq = inet_rsk(req);
6270 req->rsk_rcv_wnd = 0; /* So that tcp_send_synack() knows! */
6272 tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq;
6273 tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
6274 tcp_rsk(req)->snt_synack = tcp_clock_us();
6275 tcp_rsk(req)->last_oow_ack_time = 0;
6276 req->mss = rx_opt->mss_clamp;
6277 req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0;
6278 ireq->tstamp_ok = rx_opt->tstamp_ok;
6279 ireq->sack_ok = rx_opt->sack_ok;
6280 ireq->snd_wscale = rx_opt->snd_wscale;
6281 ireq->wscale_ok = rx_opt->wscale_ok;
6284 ireq->ir_rmt_port = tcp_hdr(skb)->source;
6285 ireq->ir_num = ntohs(tcp_hdr(skb)->dest);
6286 ireq->ir_mark = inet_request_mark(sk, skb);
6289 struct request_sock *inet_reqsk_alloc(const struct request_sock_ops *ops,
6290 struct sock *sk_listener,
6291 bool attach_listener)
6293 struct request_sock *req = reqsk_alloc(ops, sk_listener,
6297 struct inet_request_sock *ireq = inet_rsk(req);
6299 ireq->ireq_opt = NULL;
6300 #if IS_ENABLED(CONFIG_IPV6)
6301 ireq->pktopts = NULL;
6303 atomic64_set(&ireq->ir_cookie, 0);
6304 ireq->ireq_state = TCP_NEW_SYN_RECV;
6305 write_pnet(&ireq->ireq_net, sock_net(sk_listener));
6306 ireq->ireq_family = sk_listener->sk_family;
6311 EXPORT_SYMBOL(inet_reqsk_alloc);
6314 * Return true if a syncookie should be sent
6316 static bool tcp_syn_flood_action(const struct sock *sk,
6317 const struct sk_buff *skb,
6320 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
6321 const char *msg = "Dropping request";
6322 bool want_cookie = false;
6323 struct net *net = sock_net(sk);
6325 #ifdef CONFIG_SYN_COOKIES
6326 if (net->ipv4.sysctl_tcp_syncookies) {
6327 msg = "Sending cookies";
6329 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES);
6332 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP);
6334 if (!queue->synflood_warned &&
6335 net->ipv4.sysctl_tcp_syncookies != 2 &&
6336 xchg(&queue->synflood_warned, 1) == 0)
6337 pr_info("%s: Possible SYN flooding on port %d. %s. Check SNMP counters.\n",
6338 proto, ntohs(tcp_hdr(skb)->dest), msg);
6343 static void tcp_reqsk_record_syn(const struct sock *sk,
6344 struct request_sock *req,
6345 const struct sk_buff *skb)
6347 if (tcp_sk(sk)->save_syn) {
6348 u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb);
6351 copy = kmalloc(len + sizeof(u32), GFP_ATOMIC);
6354 memcpy(©[1], skb_network_header(skb), len);
6355 req->saved_syn = copy;
6360 int tcp_conn_request(struct request_sock_ops *rsk_ops,
6361 const struct tcp_request_sock_ops *af_ops,
6362 struct sock *sk, struct sk_buff *skb)
6364 struct tcp_fastopen_cookie foc = { .len = -1 };
6365 __u32 isn = TCP_SKB_CB(skb)->tcp_tw_isn;
6366 struct tcp_options_received tmp_opt;
6367 struct tcp_sock *tp = tcp_sk(sk);
6368 struct net *net = sock_net(sk);
6369 struct sock *fastopen_sk = NULL;
6370 struct request_sock *req;
6371 bool want_cookie = false;
6372 struct dst_entry *dst;
6375 /* TW buckets are converted to open requests without
6376 * limitations, they conserve resources and peer is
6377 * evidently real one.
6379 if ((net->ipv4.sysctl_tcp_syncookies == 2 ||
6380 inet_csk_reqsk_queue_is_full(sk)) && !isn) {
6381 want_cookie = tcp_syn_flood_action(sk, skb, rsk_ops->slab_name);
6386 if (sk_acceptq_is_full(sk)) {
6387 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
6391 req = inet_reqsk_alloc(rsk_ops, sk, !want_cookie);
6395 tcp_rsk(req)->af_specific = af_ops;
6396 tcp_rsk(req)->ts_off = 0;
6398 tcp_clear_options(&tmp_opt);
6399 tmp_opt.mss_clamp = af_ops->mss_clamp;
6400 tmp_opt.user_mss = tp->rx_opt.user_mss;
6401 tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0,
6402 want_cookie ? NULL : &foc);
6404 if (want_cookie && !tmp_opt.saw_tstamp)
6405 tcp_clear_options(&tmp_opt);
6407 tmp_opt.tstamp_ok = tmp_opt.saw_tstamp;
6408 tcp_openreq_init(req, &tmp_opt, skb, sk);
6409 inet_rsk(req)->no_srccheck = inet_sk(sk)->transparent;
6411 /* Note: tcp_v6_init_req() might override ir_iif for link locals */
6412 inet_rsk(req)->ir_iif = inet_request_bound_dev_if(sk, skb);
6414 af_ops->init_req(req, sk, skb);
6416 if (security_inet_conn_request(sk, skb, req))
6419 if (tmp_opt.tstamp_ok)
6420 tcp_rsk(req)->ts_off = af_ops->init_ts_off(net, skb);
6422 dst = af_ops->route_req(sk, &fl, req);
6426 if (!want_cookie && !isn) {
6427 /* Kill the following clause, if you dislike this way. */
6428 if (!net->ipv4.sysctl_tcp_syncookies &&
6429 (net->ipv4.sysctl_max_syn_backlog - inet_csk_reqsk_queue_len(sk) <
6430 (net->ipv4.sysctl_max_syn_backlog >> 2)) &&
6431 !tcp_peer_is_proven(req, dst)) {
6432 /* Without syncookies last quarter of
6433 * backlog is filled with destinations,
6434 * proven to be alive.
6435 * It means that we continue to communicate
6436 * to destinations, already remembered
6437 * to the moment of synflood.
6439 pr_drop_req(req, ntohs(tcp_hdr(skb)->source),
6441 goto drop_and_release;
6444 isn = af_ops->init_seq(skb);
6447 tcp_ecn_create_request(req, skb, sk, dst);
6450 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss);
6451 req->cookie_ts = tmp_opt.tstamp_ok;
6452 if (!tmp_opt.tstamp_ok)
6453 inet_rsk(req)->ecn_ok = 0;
6456 tcp_rsk(req)->snt_isn = isn;
6457 tcp_rsk(req)->txhash = net_tx_rndhash();
6458 tcp_openreq_init_rwin(req, sk, dst);
6460 tcp_reqsk_record_syn(sk, req, skb);
6461 fastopen_sk = tcp_try_fastopen(sk, skb, req, &foc);
6464 af_ops->send_synack(fastopen_sk, dst, &fl, req,
6465 &foc, TCP_SYNACK_FASTOPEN);
6466 /* Add the child socket directly into the accept queue */
6467 if (!inet_csk_reqsk_queue_add(sk, req, fastopen_sk)) {
6468 reqsk_fastopen_remove(fastopen_sk, req, false);
6469 bh_unlock_sock(fastopen_sk);
6470 sock_put(fastopen_sk);
6474 sk->sk_data_ready(sk);
6475 bh_unlock_sock(fastopen_sk);
6476 sock_put(fastopen_sk);
6478 tcp_rsk(req)->tfo_listener = false;
6480 inet_csk_reqsk_queue_hash_add(sk, req,
6481 tcp_timeout_init((struct sock *)req));
6482 af_ops->send_synack(sk, dst, &fl, req, &foc,
6483 !want_cookie ? TCP_SYNACK_NORMAL :
6501 EXPORT_SYMBOL(tcp_conn_request);