1 // SPDX-License-Identifier: GPL-2.0
3 * INET An implementation of the TCP/IP protocol suite for the LINUX
4 * operating system. INET is implemented using the BSD Socket
5 * interface as the means of communication with the user level.
7 * Implementation of the Transmission Control Protocol(TCP).
10 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
11 * Mark Evans, <evansmp@uhura.aston.ac.uk>
12 * Corey Minyard <wf-rch!minyard@relay.EU.net>
13 * Florian La Roche, <flla@stud.uni-sb.de>
14 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
15 * Linus Torvalds, <torvalds@cs.helsinki.fi>
16 * Alan Cox, <gw4pts@gw4pts.ampr.org>
17 * Matthew Dillon, <dillon@apollo.west.oic.com>
18 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
19 * Jorge Cwik, <jorge@laser.satlink.net>
24 * Pedro Roque : Fast Retransmit/Recovery.
26 * Retransmit queue handled by TCP.
27 * Better retransmit timer handling.
28 * New congestion avoidance.
32 * Eric : Fast Retransmit.
33 * Randy Scott : MSS option defines.
34 * Eric Schenk : Fixes to slow start algorithm.
35 * Eric Schenk : Yet another double ACK bug.
36 * Eric Schenk : Delayed ACK bug fixes.
37 * Eric Schenk : Floyd style fast retrans war avoidance.
38 * David S. Miller : Don't allow zero congestion window.
39 * Eric Schenk : Fix retransmitter so that it sends
40 * next packet on ack of previous packet.
41 * Andi Kleen : Moved open_request checking here
42 * and process RSTs for open_requests.
43 * Andi Kleen : Better prune_queue, and other fixes.
44 * Andrey Savochkin: Fix RTT measurements in the presence of
46 * Andrey Savochkin: Check sequence numbers correctly when
47 * removing SACKs due to in sequence incoming
49 * Andi Kleen: Make sure we never ack data there is not
50 * enough room for. Also make this condition
51 * a fatal error if it might still happen.
52 * Andi Kleen: Add tcp_measure_rcv_mss to make
53 * connections with MSS<min(MTU,ann. MSS)
54 * work without delayed acks.
55 * Andi Kleen: Process packets with PSH set in the
57 * J Hadi Salim: ECN support
60 * Panu Kuhlberg: Experimental audit of TCP (re)transmission
61 * engine. Lots of bugs are found.
62 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs
65 #define pr_fmt(fmt) "TCP: " fmt
68 #include <linux/slab.h>
69 #include <linux/module.h>
70 #include <linux/sysctl.h>
71 #include <linux/kernel.h>
72 #include <linux/prefetch.h>
75 #include <net/inet_common.h>
76 #include <linux/ipsec.h>
77 #include <asm/unaligned.h>
78 #include <linux/errqueue.h>
79 #include <trace/events/tcp.h>
80 #include <linux/static_key.h>
81 #include <net/busy_poll.h>
83 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
85 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
86 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
87 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
88 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
89 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
90 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
91 #define FLAG_ECE 0x40 /* ECE in this ACK */
92 #define FLAG_LOST_RETRANS 0x80 /* This ACK marks some retransmission lost */
93 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
94 #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
95 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
96 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
97 #define FLAG_SET_XMIT_TIMER 0x1000 /* Set TLP or RTO timer */
98 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
99 #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
100 #define FLAG_NO_CHALLENGE_ACK 0x8000 /* do not call tcp_send_challenge_ack() */
101 #define FLAG_ACK_MAYBE_DELAYED 0x10000 /* Likely a delayed ACK */
103 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
104 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
105 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE|FLAG_DSACKING_ACK)
106 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
108 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
109 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
111 #define REXMIT_NONE 0 /* no loss recovery to do */
112 #define REXMIT_LOST 1 /* retransmit packets marked lost */
113 #define REXMIT_NEW 2 /* FRTO-style transmit of unsent/new packets */
115 #if IS_ENABLED(CONFIG_TLS_DEVICE)
116 static DEFINE_STATIC_KEY_FALSE(clean_acked_data_enabled);
118 void clean_acked_data_enable(struct inet_connection_sock *icsk,
119 void (*cad)(struct sock *sk, u32 ack_seq))
121 icsk->icsk_clean_acked = cad;
122 static_branch_inc(&clean_acked_data_enabled);
124 EXPORT_SYMBOL_GPL(clean_acked_data_enable);
126 void clean_acked_data_disable(struct inet_connection_sock *icsk)
128 static_branch_dec(&clean_acked_data_enabled);
129 icsk->icsk_clean_acked = NULL;
131 EXPORT_SYMBOL_GPL(clean_acked_data_disable);
134 static void tcp_gro_dev_warn(struct sock *sk, const struct sk_buff *skb,
137 static bool __once __read_mostly;
140 struct net_device *dev;
145 dev = dev_get_by_index_rcu(sock_net(sk), skb->skb_iif);
146 if (!dev || len >= dev->mtu)
147 pr_warn("%s: Driver has suspect GRO implementation, TCP performance may be compromised.\n",
148 dev ? dev->name : "Unknown driver");
153 /* Adapt the MSS value used to make delayed ack decision to the
156 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
158 struct inet_connection_sock *icsk = inet_csk(sk);
159 const unsigned int lss = icsk->icsk_ack.last_seg_size;
162 icsk->icsk_ack.last_seg_size = 0;
164 /* skb->len may jitter because of SACKs, even if peer
165 * sends good full-sized frames.
167 len = skb_shinfo(skb)->gso_size ? : skb->len;
168 if (len >= icsk->icsk_ack.rcv_mss) {
169 icsk->icsk_ack.rcv_mss = min_t(unsigned int, len,
171 /* Account for possibly-removed options */
172 if (unlikely(len > icsk->icsk_ack.rcv_mss +
173 MAX_TCP_OPTION_SPACE))
174 tcp_gro_dev_warn(sk, skb, len);
176 /* Otherwise, we make more careful check taking into account,
177 * that SACKs block is variable.
179 * "len" is invariant segment length, including TCP header.
181 len += skb->data - skb_transport_header(skb);
182 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
183 /* If PSH is not set, packet should be
184 * full sized, provided peer TCP is not badly broken.
185 * This observation (if it is correct 8)) allows
186 * to handle super-low mtu links fairly.
188 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
189 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
190 /* Subtract also invariant (if peer is RFC compliant),
191 * tcp header plus fixed timestamp option length.
192 * Resulting "len" is MSS free of SACK jitter.
194 len -= tcp_sk(sk)->tcp_header_len;
195 icsk->icsk_ack.last_seg_size = len;
197 icsk->icsk_ack.rcv_mss = len;
201 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
202 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
203 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
207 static void tcp_incr_quickack(struct sock *sk, unsigned int max_quickacks)
209 struct inet_connection_sock *icsk = inet_csk(sk);
210 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
214 quickacks = min(quickacks, max_quickacks);
215 if (quickacks > icsk->icsk_ack.quick)
216 icsk->icsk_ack.quick = quickacks;
219 void tcp_enter_quickack_mode(struct sock *sk, unsigned int max_quickacks)
221 struct inet_connection_sock *icsk = inet_csk(sk);
223 tcp_incr_quickack(sk, max_quickacks);
224 icsk->icsk_ack.pingpong = 0;
225 icsk->icsk_ack.ato = TCP_ATO_MIN;
227 EXPORT_SYMBOL(tcp_enter_quickack_mode);
229 /* Send ACKs quickly, if "quick" count is not exhausted
230 * and the session is not interactive.
233 static bool tcp_in_quickack_mode(struct sock *sk)
235 const struct inet_connection_sock *icsk = inet_csk(sk);
236 const struct dst_entry *dst = __sk_dst_get(sk);
238 return (dst && dst_metric(dst, RTAX_QUICKACK)) ||
239 (icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong);
242 static void tcp_ecn_queue_cwr(struct tcp_sock *tp)
244 if (tp->ecn_flags & TCP_ECN_OK)
245 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
248 static void tcp_ecn_accept_cwr(struct sock *sk, const struct sk_buff *skb)
250 if (tcp_hdr(skb)->cwr) {
251 tcp_sk(sk)->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
253 /* If the sender is telling us it has entered CWR, then its
254 * cwnd may be very low (even just 1 packet), so we should ACK
257 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW;
261 static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp)
263 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
266 static void __tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb)
268 struct tcp_sock *tp = tcp_sk(sk);
270 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
271 case INET_ECN_NOT_ECT:
272 /* Funny extension: if ECT is not set on a segment,
273 * and we already seen ECT on a previous segment,
274 * it is probably a retransmit.
276 if (tp->ecn_flags & TCP_ECN_SEEN)
277 tcp_enter_quickack_mode(sk, 2);
280 if (tcp_ca_needs_ecn(sk))
281 tcp_ca_event(sk, CA_EVENT_ECN_IS_CE);
283 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
284 /* Better not delay acks, sender can have a very low cwnd */
285 tcp_enter_quickack_mode(sk, 2);
286 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
288 tp->ecn_flags |= TCP_ECN_SEEN;
291 if (tcp_ca_needs_ecn(sk))
292 tcp_ca_event(sk, CA_EVENT_ECN_NO_CE);
293 tp->ecn_flags |= TCP_ECN_SEEN;
298 static void tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb)
300 if (tcp_sk(sk)->ecn_flags & TCP_ECN_OK)
301 __tcp_ecn_check_ce(sk, skb);
304 static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
306 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
307 tp->ecn_flags &= ~TCP_ECN_OK;
310 static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
312 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
313 tp->ecn_flags &= ~TCP_ECN_OK;
316 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
318 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
323 /* Buffer size and advertised window tuning.
325 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
328 static void tcp_sndbuf_expand(struct sock *sk)
330 const struct tcp_sock *tp = tcp_sk(sk);
331 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
335 /* Worst case is non GSO/TSO : each frame consumes one skb
336 * and skb->head is kmalloced using power of two area of memory
338 per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
340 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
342 per_mss = roundup_pow_of_two(per_mss) +
343 SKB_DATA_ALIGN(sizeof(struct sk_buff));
345 nr_segs = max_t(u32, TCP_INIT_CWND, tp->snd_cwnd);
346 nr_segs = max_t(u32, nr_segs, tp->reordering + 1);
348 /* Fast Recovery (RFC 5681 3.2) :
349 * Cubic needs 1.7 factor, rounded to 2 to include
350 * extra cushion (application might react slowly to EPOLLOUT)
352 sndmem = ca_ops->sndbuf_expand ? ca_ops->sndbuf_expand(sk) : 2;
353 sndmem *= nr_segs * per_mss;
355 if (sk->sk_sndbuf < sndmem)
356 sk->sk_sndbuf = min(sndmem, sock_net(sk)->ipv4.sysctl_tcp_wmem[2]);
359 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
361 * All tcp_full_space() is split to two parts: "network" buffer, allocated
362 * forward and advertised in receiver window (tp->rcv_wnd) and
363 * "application buffer", required to isolate scheduling/application
364 * latencies from network.
365 * window_clamp is maximal advertised window. It can be less than
366 * tcp_full_space(), in this case tcp_full_space() - window_clamp
367 * is reserved for "application" buffer. The less window_clamp is
368 * the smoother our behaviour from viewpoint of network, but the lower
369 * throughput and the higher sensitivity of the connection to losses. 8)
371 * rcv_ssthresh is more strict window_clamp used at "slow start"
372 * phase to predict further behaviour of this connection.
373 * It is used for two goals:
374 * - to enforce header prediction at sender, even when application
375 * requires some significant "application buffer". It is check #1.
376 * - to prevent pruning of receive queue because of misprediction
377 * of receiver window. Check #2.
379 * The scheme does not work when sender sends good segments opening
380 * window and then starts to feed us spaghetti. But it should work
381 * in common situations. Otherwise, we have to rely on queue collapsing.
384 /* Slow part of check#2. */
385 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
387 struct tcp_sock *tp = tcp_sk(sk);
389 int truesize = tcp_win_from_space(sk, skb->truesize) >> 1;
390 int window = tcp_win_from_space(sk, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]) >> 1;
392 while (tp->rcv_ssthresh <= window) {
393 if (truesize <= skb->len)
394 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
402 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
404 struct tcp_sock *tp = tcp_sk(sk);
407 if (tp->rcv_ssthresh < tp->window_clamp &&
408 (int)tp->rcv_ssthresh < tcp_space(sk) &&
409 !tcp_under_memory_pressure(sk)) {
412 /* Check #2. Increase window, if skb with such overhead
413 * will fit to rcvbuf in future.
415 if (tcp_win_from_space(sk, skb->truesize) <= skb->len)
416 incr = 2 * tp->advmss;
418 incr = __tcp_grow_window(sk, skb);
421 incr = max_t(int, incr, 2 * skb->len);
422 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
424 inet_csk(sk)->icsk_ack.quick |= 1;
429 /* 3. Tuning rcvbuf, when connection enters established state. */
430 static void tcp_fixup_rcvbuf(struct sock *sk)
432 u32 mss = tcp_sk(sk)->advmss;
435 rcvmem = 2 * SKB_TRUESIZE(mss + MAX_TCP_HEADER) *
436 tcp_default_init_rwnd(mss);
438 /* Dynamic Right Sizing (DRS) has 2 to 3 RTT latency
439 * Allow enough cushion so that sender is not limited by our window
441 if (sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf)
444 if (sk->sk_rcvbuf < rcvmem)
445 sk->sk_rcvbuf = min(rcvmem, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
448 /* 4. Try to fixup all. It is made immediately after connection enters
451 void tcp_init_buffer_space(struct sock *sk)
453 int tcp_app_win = sock_net(sk)->ipv4.sysctl_tcp_app_win;
454 struct tcp_sock *tp = tcp_sk(sk);
457 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
458 tcp_fixup_rcvbuf(sk);
459 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
460 tcp_sndbuf_expand(sk);
462 tp->rcvq_space.space = tp->rcv_wnd;
463 tcp_mstamp_refresh(tp);
464 tp->rcvq_space.time = tp->tcp_mstamp;
465 tp->rcvq_space.seq = tp->copied_seq;
467 maxwin = tcp_full_space(sk);
469 if (tp->window_clamp >= maxwin) {
470 tp->window_clamp = maxwin;
472 if (tcp_app_win && maxwin > 4 * tp->advmss)
473 tp->window_clamp = max(maxwin -
474 (maxwin >> tcp_app_win),
478 /* Force reservation of one segment. */
480 tp->window_clamp > 2 * tp->advmss &&
481 tp->window_clamp + tp->advmss > maxwin)
482 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
484 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
485 tp->snd_cwnd_stamp = tcp_jiffies32;
488 /* 5. Recalculate window clamp after socket hit its memory bounds. */
489 static void tcp_clamp_window(struct sock *sk)
491 struct tcp_sock *tp = tcp_sk(sk);
492 struct inet_connection_sock *icsk = inet_csk(sk);
493 struct net *net = sock_net(sk);
495 icsk->icsk_ack.quick = 0;
497 if (sk->sk_rcvbuf < net->ipv4.sysctl_tcp_rmem[2] &&
498 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
499 !tcp_under_memory_pressure(sk) &&
500 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
501 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
502 net->ipv4.sysctl_tcp_rmem[2]);
504 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
505 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
508 /* Initialize RCV_MSS value.
509 * RCV_MSS is an our guess about MSS used by the peer.
510 * We haven't any direct information about the MSS.
511 * It's better to underestimate the RCV_MSS rather than overestimate.
512 * Overestimations make us ACKing less frequently than needed.
513 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
515 void tcp_initialize_rcv_mss(struct sock *sk)
517 const struct tcp_sock *tp = tcp_sk(sk);
518 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
520 hint = min(hint, tp->rcv_wnd / 2);
521 hint = min(hint, TCP_MSS_DEFAULT);
522 hint = max(hint, TCP_MIN_MSS);
524 inet_csk(sk)->icsk_ack.rcv_mss = hint;
526 EXPORT_SYMBOL(tcp_initialize_rcv_mss);
528 /* Receiver "autotuning" code.
530 * The algorithm for RTT estimation w/o timestamps is based on
531 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
532 * <http://public.lanl.gov/radiant/pubs.html#DRS>
534 * More detail on this code can be found at
535 * <http://staff.psc.edu/jheffner/>,
536 * though this reference is out of date. A new paper
539 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
541 u32 new_sample = tp->rcv_rtt_est.rtt_us;
544 if (new_sample != 0) {
545 /* If we sample in larger samples in the non-timestamp
546 * case, we could grossly overestimate the RTT especially
547 * with chatty applications or bulk transfer apps which
548 * are stalled on filesystem I/O.
550 * Also, since we are only going for a minimum in the
551 * non-timestamp case, we do not smooth things out
552 * else with timestamps disabled convergence takes too
556 m -= (new_sample >> 3);
564 /* No previous measure. */
568 tp->rcv_rtt_est.rtt_us = new_sample;
571 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
575 if (tp->rcv_rtt_est.time == 0)
577 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
579 delta_us = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcv_rtt_est.time);
582 tcp_rcv_rtt_update(tp, delta_us, 1);
585 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
586 tp->rcv_rtt_est.time = tp->tcp_mstamp;
589 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
590 const struct sk_buff *skb)
592 struct tcp_sock *tp = tcp_sk(sk);
594 if (tp->rx_opt.rcv_tsecr == tp->rcv_rtt_last_tsecr)
596 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr;
598 if (TCP_SKB_CB(skb)->end_seq -
599 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss) {
600 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr;
605 delta_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
606 tcp_rcv_rtt_update(tp, delta_us, 0);
611 * This function should be called every time data is copied to user space.
612 * It calculates the appropriate TCP receive buffer space.
614 void tcp_rcv_space_adjust(struct sock *sk)
616 struct tcp_sock *tp = tcp_sk(sk);
620 trace_tcp_rcv_space_adjust(sk);
622 tcp_mstamp_refresh(tp);
623 time = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcvq_space.time);
624 if (time < (tp->rcv_rtt_est.rtt_us >> 3) || tp->rcv_rtt_est.rtt_us == 0)
627 /* Number of bytes copied to user in last RTT */
628 copied = tp->copied_seq - tp->rcvq_space.seq;
629 if (copied <= tp->rcvq_space.space)
633 * copied = bytes received in previous RTT, our base window
634 * To cope with packet losses, we need a 2x factor
635 * To cope with slow start, and sender growing its cwin by 100 %
636 * every RTT, we need a 4x factor, because the ACK we are sending
637 * now is for the next RTT, not the current one :
638 * <prev RTT . ><current RTT .. ><next RTT .... >
641 if (sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf &&
642 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
646 /* minimal window to cope with packet losses, assuming
647 * steady state. Add some cushion because of small variations.
649 rcvwin = ((u64)copied << 1) + 16 * tp->advmss;
651 /* Accommodate for sender rate increase (eg. slow start) */
652 grow = rcvwin * (copied - tp->rcvq_space.space);
653 do_div(grow, tp->rcvq_space.space);
654 rcvwin += (grow << 1);
656 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
657 while (tcp_win_from_space(sk, rcvmem) < tp->advmss)
660 do_div(rcvwin, tp->advmss);
661 rcvbuf = min_t(u64, rcvwin * rcvmem,
662 sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
663 if (rcvbuf > sk->sk_rcvbuf) {
664 sk->sk_rcvbuf = rcvbuf;
666 /* Make the window clamp follow along. */
667 tp->window_clamp = tcp_win_from_space(sk, rcvbuf);
670 tp->rcvq_space.space = copied;
673 tp->rcvq_space.seq = tp->copied_seq;
674 tp->rcvq_space.time = tp->tcp_mstamp;
677 /* There is something which you must keep in mind when you analyze the
678 * behavior of the tp->ato delayed ack timeout interval. When a
679 * connection starts up, we want to ack as quickly as possible. The
680 * problem is that "good" TCP's do slow start at the beginning of data
681 * transmission. The means that until we send the first few ACK's the
682 * sender will sit on his end and only queue most of his data, because
683 * he can only send snd_cwnd unacked packets at any given time. For
684 * each ACK we send, he increments snd_cwnd and transmits more of his
687 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
689 struct tcp_sock *tp = tcp_sk(sk);
690 struct inet_connection_sock *icsk = inet_csk(sk);
693 inet_csk_schedule_ack(sk);
695 tcp_measure_rcv_mss(sk, skb);
697 tcp_rcv_rtt_measure(tp);
701 if (!icsk->icsk_ack.ato) {
702 /* The _first_ data packet received, initialize
703 * delayed ACK engine.
705 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
706 icsk->icsk_ack.ato = TCP_ATO_MIN;
708 int m = now - icsk->icsk_ack.lrcvtime;
710 if (m <= TCP_ATO_MIN / 2) {
711 /* The fastest case is the first. */
712 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
713 } else if (m < icsk->icsk_ack.ato) {
714 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
715 if (icsk->icsk_ack.ato > icsk->icsk_rto)
716 icsk->icsk_ack.ato = icsk->icsk_rto;
717 } else if (m > icsk->icsk_rto) {
718 /* Too long gap. Apparently sender failed to
719 * restart window, so that we send ACKs quickly.
721 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
725 icsk->icsk_ack.lrcvtime = now;
727 tcp_ecn_check_ce(sk, skb);
730 tcp_grow_window(sk, skb);
733 /* Called to compute a smoothed rtt estimate. The data fed to this
734 * routine either comes from timestamps, or from segments that were
735 * known _not_ to have been retransmitted [see Karn/Partridge
736 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
737 * piece by Van Jacobson.
738 * NOTE: the next three routines used to be one big routine.
739 * To save cycles in the RFC 1323 implementation it was better to break
740 * it up into three procedures. -- erics
742 static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
744 struct tcp_sock *tp = tcp_sk(sk);
745 long m = mrtt_us; /* RTT */
746 u32 srtt = tp->srtt_us;
748 /* The following amusing code comes from Jacobson's
749 * article in SIGCOMM '88. Note that rtt and mdev
750 * are scaled versions of rtt and mean deviation.
751 * This is designed to be as fast as possible
752 * m stands for "measurement".
754 * On a 1990 paper the rto value is changed to:
755 * RTO = rtt + 4 * mdev
757 * Funny. This algorithm seems to be very broken.
758 * These formulae increase RTO, when it should be decreased, increase
759 * too slowly, when it should be increased quickly, decrease too quickly
760 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
761 * does not matter how to _calculate_ it. Seems, it was trap
762 * that VJ failed to avoid. 8)
765 m -= (srtt >> 3); /* m is now error in rtt est */
766 srtt += m; /* rtt = 7/8 rtt + 1/8 new */
768 m = -m; /* m is now abs(error) */
769 m -= (tp->mdev_us >> 2); /* similar update on mdev */
770 /* This is similar to one of Eifel findings.
771 * Eifel blocks mdev updates when rtt decreases.
772 * This solution is a bit different: we use finer gain
773 * for mdev in this case (alpha*beta).
774 * Like Eifel it also prevents growth of rto,
775 * but also it limits too fast rto decreases,
776 * happening in pure Eifel.
781 m -= (tp->mdev_us >> 2); /* similar update on mdev */
783 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */
784 if (tp->mdev_us > tp->mdev_max_us) {
785 tp->mdev_max_us = tp->mdev_us;
786 if (tp->mdev_max_us > tp->rttvar_us)
787 tp->rttvar_us = tp->mdev_max_us;
789 if (after(tp->snd_una, tp->rtt_seq)) {
790 if (tp->mdev_max_us < tp->rttvar_us)
791 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
792 tp->rtt_seq = tp->snd_nxt;
793 tp->mdev_max_us = tcp_rto_min_us(sk);
796 /* no previous measure. */
797 srtt = m << 3; /* take the measured time to be rtt */
798 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */
799 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
800 tp->mdev_max_us = tp->rttvar_us;
801 tp->rtt_seq = tp->snd_nxt;
803 tp->srtt_us = max(1U, srtt);
806 static void tcp_update_pacing_rate(struct sock *sk)
808 const struct tcp_sock *tp = tcp_sk(sk);
811 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
812 rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3);
814 /* current rate is (cwnd * mss) / srtt
815 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
816 * In Congestion Avoidance phase, set it to 120 % the current rate.
818 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
819 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
820 * end of slow start and should slow down.
822 if (tp->snd_cwnd < tp->snd_ssthresh / 2)
823 rate *= sock_net(sk)->ipv4.sysctl_tcp_pacing_ss_ratio;
825 rate *= sock_net(sk)->ipv4.sysctl_tcp_pacing_ca_ratio;
827 rate *= max(tp->snd_cwnd, tp->packets_out);
829 if (likely(tp->srtt_us))
830 do_div(rate, tp->srtt_us);
832 /* WRITE_ONCE() is needed because sch_fq fetches sk_pacing_rate
833 * without any lock. We want to make sure compiler wont store
834 * intermediate values in this location.
836 WRITE_ONCE(sk->sk_pacing_rate, min_t(u64, rate,
837 sk->sk_max_pacing_rate));
840 /* Calculate rto without backoff. This is the second half of Van Jacobson's
841 * routine referred to above.
843 static void tcp_set_rto(struct sock *sk)
845 const struct tcp_sock *tp = tcp_sk(sk);
846 /* Old crap is replaced with new one. 8)
849 * 1. If rtt variance happened to be less 50msec, it is hallucination.
850 * It cannot be less due to utterly erratic ACK generation made
851 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
852 * to do with delayed acks, because at cwnd>2 true delack timeout
853 * is invisible. Actually, Linux-2.4 also generates erratic
854 * ACKs in some circumstances.
856 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
858 /* 2. Fixups made earlier cannot be right.
859 * If we do not estimate RTO correctly without them,
860 * all the algo is pure shit and should be replaced
861 * with correct one. It is exactly, which we pretend to do.
864 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
865 * guarantees that rto is higher.
870 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
872 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
875 cwnd = TCP_INIT_CWND;
876 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
879 /* Take a notice that peer is sending D-SACKs */
880 static void tcp_dsack_seen(struct tcp_sock *tp)
882 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
883 tp->rack.dsack_seen = 1;
887 /* It's reordering when higher sequence was delivered (i.e. sacked) before
888 * some lower never-retransmitted sequence ("low_seq"). The maximum reordering
889 * distance is approximated in full-mss packet distance ("reordering").
891 static void tcp_check_sack_reordering(struct sock *sk, const u32 low_seq,
894 struct tcp_sock *tp = tcp_sk(sk);
895 const u32 mss = tp->mss_cache;
898 fack = tcp_highest_sack_seq(tp);
899 if (!before(low_seq, fack))
902 metric = fack - low_seq;
903 if ((metric > tp->reordering * mss) && mss) {
904 #if FASTRETRANS_DEBUG > 1
905 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
906 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
910 tp->undo_marker ? tp->undo_retrans : 0);
912 tp->reordering = min_t(u32, (metric + mss - 1) / mss,
913 sock_net(sk)->ipv4.sysctl_tcp_max_reordering);
916 /* This exciting event is worth to be remembered. 8) */
918 NET_INC_STATS(sock_net(sk),
919 ts ? LINUX_MIB_TCPTSREORDER : LINUX_MIB_TCPSACKREORDER);
922 /* This must be called before lost_out is incremented */
923 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
925 if (!tp->retransmit_skb_hint ||
926 before(TCP_SKB_CB(skb)->seq,
927 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
928 tp->retransmit_skb_hint = skb;
931 /* Sum the number of packets on the wire we have marked as lost.
932 * There are two cases we care about here:
933 * a) Packet hasn't been marked lost (nor retransmitted),
934 * and this is the first loss.
935 * b) Packet has been marked both lost and retransmitted,
936 * and this means we think it was lost again.
938 static void tcp_sum_lost(struct tcp_sock *tp, struct sk_buff *skb)
940 __u8 sacked = TCP_SKB_CB(skb)->sacked;
942 if (!(sacked & TCPCB_LOST) ||
943 ((sacked & TCPCB_LOST) && (sacked & TCPCB_SACKED_RETRANS)))
944 tp->lost += tcp_skb_pcount(skb);
947 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
949 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
950 tcp_verify_retransmit_hint(tp, skb);
952 tp->lost_out += tcp_skb_pcount(skb);
953 tcp_sum_lost(tp, skb);
954 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
958 void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp, struct sk_buff *skb)
960 tcp_verify_retransmit_hint(tp, skb);
962 tcp_sum_lost(tp, skb);
963 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
964 tp->lost_out += tcp_skb_pcount(skb);
965 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
969 /* This procedure tags the retransmission queue when SACKs arrive.
971 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
972 * Packets in queue with these bits set are counted in variables
973 * sacked_out, retrans_out and lost_out, correspondingly.
975 * Valid combinations are:
976 * Tag InFlight Description
977 * 0 1 - orig segment is in flight.
978 * S 0 - nothing flies, orig reached receiver.
979 * L 0 - nothing flies, orig lost by net.
980 * R 2 - both orig and retransmit are in flight.
981 * L|R 1 - orig is lost, retransmit is in flight.
982 * S|R 1 - orig reached receiver, retrans is still in flight.
983 * (L|S|R is logically valid, it could occur when L|R is sacked,
984 * but it is equivalent to plain S and code short-curcuits it to S.
985 * L|S is logically invalid, it would mean -1 packet in flight 8))
987 * These 6 states form finite state machine, controlled by the following events:
988 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
989 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
990 * 3. Loss detection event of two flavors:
991 * A. Scoreboard estimator decided the packet is lost.
992 * A'. Reno "three dupacks" marks head of queue lost.
993 * B. SACK arrives sacking SND.NXT at the moment, when the
994 * segment was retransmitted.
995 * 4. D-SACK added new rule: D-SACK changes any tag to S.
997 * It is pleasant to note, that state diagram turns out to be commutative,
998 * so that we are allowed not to be bothered by order of our actions,
999 * when multiple events arrive simultaneously. (see the function below).
1001 * Reordering detection.
1002 * --------------------
1003 * Reordering metric is maximal distance, which a packet can be displaced
1004 * in packet stream. With SACKs we can estimate it:
1006 * 1. SACK fills old hole and the corresponding segment was not
1007 * ever retransmitted -> reordering. Alas, we cannot use it
1008 * when segment was retransmitted.
1009 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1010 * for retransmitted and already SACKed segment -> reordering..
1011 * Both of these heuristics are not used in Loss state, when we cannot
1012 * account for retransmits accurately.
1014 * SACK block validation.
1015 * ----------------------
1017 * SACK block range validation checks that the received SACK block fits to
1018 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1019 * Note that SND.UNA is not included to the range though being valid because
1020 * it means that the receiver is rather inconsistent with itself reporting
1021 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1022 * perfectly valid, however, in light of RFC2018 which explicitly states
1023 * that "SACK block MUST reflect the newest segment. Even if the newest
1024 * segment is going to be discarded ...", not that it looks very clever
1025 * in case of head skb. Due to potentional receiver driven attacks, we
1026 * choose to avoid immediate execution of a walk in write queue due to
1027 * reneging and defer head skb's loss recovery to standard loss recovery
1028 * procedure that will eventually trigger (nothing forbids us doing this).
1030 * Implements also blockage to start_seq wrap-around. Problem lies in the
1031 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1032 * there's no guarantee that it will be before snd_nxt (n). The problem
1033 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1036 * <- outs wnd -> <- wrapzone ->
1037 * u e n u_w e_w s n_w
1039 * |<------------+------+----- TCP seqno space --------------+---------->|
1040 * ...-- <2^31 ->| |<--------...
1041 * ...---- >2^31 ------>| |<--------...
1043 * Current code wouldn't be vulnerable but it's better still to discard such
1044 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1045 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1046 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1047 * equal to the ideal case (infinite seqno space without wrap caused issues).
1049 * With D-SACK the lower bound is extended to cover sequence space below
1050 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1051 * again, D-SACK block must not to go across snd_una (for the same reason as
1052 * for the normal SACK blocks, explained above). But there all simplicity
1053 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1054 * fully below undo_marker they do not affect behavior in anyway and can
1055 * therefore be safely ignored. In rare cases (which are more or less
1056 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1057 * fragmentation and packet reordering past skb's retransmission. To consider
1058 * them correctly, the acceptable range must be extended even more though
1059 * the exact amount is rather hard to quantify. However, tp->max_window can
1060 * be used as an exaggerated estimate.
1062 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
1063 u32 start_seq, u32 end_seq)
1065 /* Too far in future, or reversed (interpretation is ambiguous) */
1066 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1069 /* Nasty start_seq wrap-around check (see comments above) */
1070 if (!before(start_seq, tp->snd_nxt))
1073 /* In outstanding window? ...This is valid exit for D-SACKs too.
1074 * start_seq == snd_una is non-sensical (see comments above)
1076 if (after(start_seq, tp->snd_una))
1079 if (!is_dsack || !tp->undo_marker)
1082 /* ...Then it's D-SACK, and must reside below snd_una completely */
1083 if (after(end_seq, tp->snd_una))
1086 if (!before(start_seq, tp->undo_marker))
1090 if (!after(end_seq, tp->undo_marker))
1093 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1094 * start_seq < undo_marker and end_seq >= undo_marker.
1096 return !before(start_seq, end_seq - tp->max_window);
1099 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1100 struct tcp_sack_block_wire *sp, int num_sacks,
1103 struct tcp_sock *tp = tcp_sk(sk);
1104 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1105 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1106 bool dup_sack = false;
1108 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1111 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1112 } else if (num_sacks > 1) {
1113 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1114 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1116 if (!after(end_seq_0, end_seq_1) &&
1117 !before(start_seq_0, start_seq_1)) {
1120 NET_INC_STATS(sock_net(sk),
1121 LINUX_MIB_TCPDSACKOFORECV);
1125 /* D-SACK for already forgotten data... Do dumb counting. */
1126 if (dup_sack && tp->undo_marker && tp->undo_retrans > 0 &&
1127 !after(end_seq_0, prior_snd_una) &&
1128 after(end_seq_0, tp->undo_marker))
1134 struct tcp_sacktag_state {
1136 /* Timestamps for earliest and latest never-retransmitted segment
1137 * that was SACKed. RTO needs the earliest RTT to stay conservative,
1138 * but congestion control should still get an accurate delay signal.
1142 struct rate_sample *rate;
1144 unsigned int mss_now;
1147 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1148 * the incoming SACK may not exactly match but we can find smaller MSS
1149 * aligned portion of it that matches. Therefore we might need to fragment
1150 * which may fail and creates some hassle (caller must handle error case
1153 * FIXME: this could be merged to shift decision code
1155 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1156 u32 start_seq, u32 end_seq)
1160 unsigned int pkt_len;
1163 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1164 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1166 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1167 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1168 mss = tcp_skb_mss(skb);
1169 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1172 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1176 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1181 /* Round if necessary so that SACKs cover only full MSSes
1182 * and/or the remaining small portion (if present)
1184 if (pkt_len > mss) {
1185 unsigned int new_len = (pkt_len / mss) * mss;
1186 if (!in_sack && new_len < pkt_len)
1191 if (pkt_len >= skb->len && !in_sack)
1194 err = tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
1195 pkt_len, mss, GFP_ATOMIC);
1203 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1204 static u8 tcp_sacktag_one(struct sock *sk,
1205 struct tcp_sacktag_state *state, u8 sacked,
1206 u32 start_seq, u32 end_seq,
1207 int dup_sack, int pcount,
1210 struct tcp_sock *tp = tcp_sk(sk);
1212 /* Account D-SACK for retransmitted packet. */
1213 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1214 if (tp->undo_marker && tp->undo_retrans > 0 &&
1215 after(end_seq, tp->undo_marker))
1217 if ((sacked & TCPCB_SACKED_ACKED) &&
1218 before(start_seq, state->reord))
1219 state->reord = start_seq;
1222 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1223 if (!after(end_seq, tp->snd_una))
1226 if (!(sacked & TCPCB_SACKED_ACKED)) {
1227 tcp_rack_advance(tp, sacked, end_seq, xmit_time);
1229 if (sacked & TCPCB_SACKED_RETRANS) {
1230 /* If the segment is not tagged as lost,
1231 * we do not clear RETRANS, believing
1232 * that retransmission is still in flight.
1234 if (sacked & TCPCB_LOST) {
1235 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1236 tp->lost_out -= pcount;
1237 tp->retrans_out -= pcount;
1240 if (!(sacked & TCPCB_RETRANS)) {
1241 /* New sack for not retransmitted frame,
1242 * which was in hole. It is reordering.
1244 if (before(start_seq,
1245 tcp_highest_sack_seq(tp)) &&
1246 before(start_seq, state->reord))
1247 state->reord = start_seq;
1249 if (!after(end_seq, tp->high_seq))
1250 state->flag |= FLAG_ORIG_SACK_ACKED;
1251 if (state->first_sackt == 0)
1252 state->first_sackt = xmit_time;
1253 state->last_sackt = xmit_time;
1256 if (sacked & TCPCB_LOST) {
1257 sacked &= ~TCPCB_LOST;
1258 tp->lost_out -= pcount;
1262 sacked |= TCPCB_SACKED_ACKED;
1263 state->flag |= FLAG_DATA_SACKED;
1264 tp->sacked_out += pcount;
1265 tp->delivered += pcount; /* Out-of-order packets delivered */
1267 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1268 if (tp->lost_skb_hint &&
1269 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1270 tp->lost_cnt_hint += pcount;
1273 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1274 * frames and clear it. undo_retrans is decreased above, L|R frames
1275 * are accounted above as well.
1277 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1278 sacked &= ~TCPCB_SACKED_RETRANS;
1279 tp->retrans_out -= pcount;
1285 /* Shift newly-SACKed bytes from this skb to the immediately previous
1286 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1288 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *prev,
1289 struct sk_buff *skb,
1290 struct tcp_sacktag_state *state,
1291 unsigned int pcount, int shifted, int mss,
1294 struct tcp_sock *tp = tcp_sk(sk);
1295 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1296 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
1300 /* Adjust counters and hints for the newly sacked sequence
1301 * range but discard the return value since prev is already
1302 * marked. We must tag the range first because the seq
1303 * advancement below implicitly advances
1304 * tcp_highest_sack_seq() when skb is highest_sack.
1306 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1307 start_seq, end_seq, dup_sack, pcount,
1309 tcp_rate_skb_delivered(sk, skb, state->rate);
1311 if (skb == tp->lost_skb_hint)
1312 tp->lost_cnt_hint += pcount;
1314 TCP_SKB_CB(prev)->end_seq += shifted;
1315 TCP_SKB_CB(skb)->seq += shifted;
1317 tcp_skb_pcount_add(prev, pcount);
1318 BUG_ON(tcp_skb_pcount(skb) < pcount);
1319 tcp_skb_pcount_add(skb, -pcount);
1321 /* When we're adding to gso_segs == 1, gso_size will be zero,
1322 * in theory this shouldn't be necessary but as long as DSACK
1323 * code can come after this skb later on it's better to keep
1324 * setting gso_size to something.
1326 if (!TCP_SKB_CB(prev)->tcp_gso_size)
1327 TCP_SKB_CB(prev)->tcp_gso_size = mss;
1329 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1330 if (tcp_skb_pcount(skb) <= 1)
1331 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1333 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1334 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1337 BUG_ON(!tcp_skb_pcount(skb));
1338 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1342 /* Whole SKB was eaten :-) */
1344 if (skb == tp->retransmit_skb_hint)
1345 tp->retransmit_skb_hint = prev;
1346 if (skb == tp->lost_skb_hint) {
1347 tp->lost_skb_hint = prev;
1348 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1351 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1352 TCP_SKB_CB(prev)->eor = TCP_SKB_CB(skb)->eor;
1353 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1354 TCP_SKB_CB(prev)->end_seq++;
1356 if (skb == tcp_highest_sack(sk))
1357 tcp_advance_highest_sack(sk, skb);
1359 tcp_skb_collapse_tstamp(prev, skb);
1360 if (unlikely(TCP_SKB_CB(prev)->tx.delivered_mstamp))
1361 TCP_SKB_CB(prev)->tx.delivered_mstamp = 0;
1363 tcp_rtx_queue_unlink_and_free(skb, sk);
1365 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKMERGED);
1370 /* I wish gso_size would have a bit more sane initialization than
1371 * something-or-zero which complicates things
1373 static int tcp_skb_seglen(const struct sk_buff *skb)
1375 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1378 /* Shifting pages past head area doesn't work */
1379 static int skb_can_shift(const struct sk_buff *skb)
1381 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1384 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1387 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1388 struct tcp_sacktag_state *state,
1389 u32 start_seq, u32 end_seq,
1392 struct tcp_sock *tp = tcp_sk(sk);
1393 struct sk_buff *prev;
1399 /* Normally R but no L won't result in plain S */
1401 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1403 if (!skb_can_shift(skb))
1405 /* This frame is about to be dropped (was ACKed). */
1406 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1409 /* Can only happen with delayed DSACK + discard craziness */
1410 prev = skb_rb_prev(skb);
1414 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1417 if (!tcp_skb_can_collapse_to(prev))
1420 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1421 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1425 pcount = tcp_skb_pcount(skb);
1426 mss = tcp_skb_seglen(skb);
1428 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1429 * drop this restriction as unnecessary
1431 if (mss != tcp_skb_seglen(prev))
1434 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1436 /* CHECKME: This is non-MSS split case only?, this will
1437 * cause skipped skbs due to advancing loop btw, original
1438 * has that feature too
1440 if (tcp_skb_pcount(skb) <= 1)
1443 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1445 /* TODO: head merge to next could be attempted here
1446 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1447 * though it might not be worth of the additional hassle
1449 * ...we can probably just fallback to what was done
1450 * previously. We could try merging non-SACKed ones
1451 * as well but it probably isn't going to buy off
1452 * because later SACKs might again split them, and
1453 * it would make skb timestamp tracking considerably
1459 len = end_seq - TCP_SKB_CB(skb)->seq;
1461 BUG_ON(len > skb->len);
1463 /* MSS boundaries should be honoured or else pcount will
1464 * severely break even though it makes things bit trickier.
1465 * Optimize common case to avoid most of the divides
1467 mss = tcp_skb_mss(skb);
1469 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1470 * drop this restriction as unnecessary
1472 if (mss != tcp_skb_seglen(prev))
1477 } else if (len < mss) {
1485 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1486 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1489 if (!skb_shift(prev, skb, len))
1491 if (!tcp_shifted_skb(sk, prev, skb, state, pcount, len, mss, dup_sack))
1494 /* Hole filled allows collapsing with the next as well, this is very
1495 * useful when hole on every nth skb pattern happens
1497 skb = skb_rb_next(prev);
1501 if (!skb_can_shift(skb) ||
1502 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1503 (mss != tcp_skb_seglen(skb)))
1507 if (skb_shift(prev, skb, len)) {
1508 pcount += tcp_skb_pcount(skb);
1509 tcp_shifted_skb(sk, prev, skb, state, tcp_skb_pcount(skb),
1520 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1524 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1525 struct tcp_sack_block *next_dup,
1526 struct tcp_sacktag_state *state,
1527 u32 start_seq, u32 end_seq,
1530 struct tcp_sock *tp = tcp_sk(sk);
1531 struct sk_buff *tmp;
1533 skb_rbtree_walk_from(skb) {
1535 bool dup_sack = dup_sack_in;
1537 /* queue is in-order => we can short-circuit the walk early */
1538 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1542 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1543 in_sack = tcp_match_skb_to_sack(sk, skb,
1544 next_dup->start_seq,
1550 /* skb reference here is a bit tricky to get right, since
1551 * shifting can eat and free both this skb and the next,
1552 * so not even _safe variant of the loop is enough.
1555 tmp = tcp_shift_skb_data(sk, skb, state,
1556 start_seq, end_seq, dup_sack);
1565 in_sack = tcp_match_skb_to_sack(sk, skb,
1571 if (unlikely(in_sack < 0))
1575 TCP_SKB_CB(skb)->sacked =
1578 TCP_SKB_CB(skb)->sacked,
1579 TCP_SKB_CB(skb)->seq,
1580 TCP_SKB_CB(skb)->end_seq,
1582 tcp_skb_pcount(skb),
1584 tcp_rate_skb_delivered(sk, skb, state->rate);
1585 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1586 list_del_init(&skb->tcp_tsorted_anchor);
1588 if (!before(TCP_SKB_CB(skb)->seq,
1589 tcp_highest_sack_seq(tp)))
1590 tcp_advance_highest_sack(sk, skb);
1596 static struct sk_buff *tcp_sacktag_bsearch(struct sock *sk,
1597 struct tcp_sacktag_state *state,
1600 struct rb_node *parent, **p = &sk->tcp_rtx_queue.rb_node;
1601 struct sk_buff *skb;
1605 skb = rb_to_skb(parent);
1606 if (before(seq, TCP_SKB_CB(skb)->seq)) {
1607 p = &parent->rb_left;
1610 if (!before(seq, TCP_SKB_CB(skb)->end_seq)) {
1611 p = &parent->rb_right;
1619 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1620 struct tcp_sacktag_state *state,
1623 if (skb && after(TCP_SKB_CB(skb)->seq, skip_to_seq))
1626 return tcp_sacktag_bsearch(sk, state, skip_to_seq);
1629 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1631 struct tcp_sack_block *next_dup,
1632 struct tcp_sacktag_state *state,
1638 if (before(next_dup->start_seq, skip_to_seq)) {
1639 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1640 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1641 next_dup->start_seq, next_dup->end_seq,
1648 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1650 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1654 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1655 u32 prior_snd_una, struct tcp_sacktag_state *state)
1657 struct tcp_sock *tp = tcp_sk(sk);
1658 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1659 TCP_SKB_CB(ack_skb)->sacked);
1660 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1661 struct tcp_sack_block sp[TCP_NUM_SACKS];
1662 struct tcp_sack_block *cache;
1663 struct sk_buff *skb;
1664 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1666 bool found_dup_sack = false;
1668 int first_sack_index;
1671 state->reord = tp->snd_nxt;
1673 if (!tp->sacked_out)
1674 tcp_highest_sack_reset(sk);
1676 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1677 num_sacks, prior_snd_una);
1678 if (found_dup_sack) {
1679 state->flag |= FLAG_DSACKING_ACK;
1680 tp->delivered++; /* A spurious retransmission is delivered */
1683 /* Eliminate too old ACKs, but take into
1684 * account more or less fresh ones, they can
1685 * contain valid SACK info.
1687 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1690 if (!tp->packets_out)
1694 first_sack_index = 0;
1695 for (i = 0; i < num_sacks; i++) {
1696 bool dup_sack = !i && found_dup_sack;
1698 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1699 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1701 if (!tcp_is_sackblock_valid(tp, dup_sack,
1702 sp[used_sacks].start_seq,
1703 sp[used_sacks].end_seq)) {
1707 if (!tp->undo_marker)
1708 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1710 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1712 /* Don't count olds caused by ACK reordering */
1713 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1714 !after(sp[used_sacks].end_seq, tp->snd_una))
1716 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1719 NET_INC_STATS(sock_net(sk), mib_idx);
1721 first_sack_index = -1;
1725 /* Ignore very old stuff early */
1726 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1732 /* order SACK blocks to allow in order walk of the retrans queue */
1733 for (i = used_sacks - 1; i > 0; i--) {
1734 for (j = 0; j < i; j++) {
1735 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1736 swap(sp[j], sp[j + 1]);
1738 /* Track where the first SACK block goes to */
1739 if (j == first_sack_index)
1740 first_sack_index = j + 1;
1745 state->mss_now = tcp_current_mss(sk);
1749 if (!tp->sacked_out) {
1750 /* It's already past, so skip checking against it */
1751 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1753 cache = tp->recv_sack_cache;
1754 /* Skip empty blocks in at head of the cache */
1755 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1760 while (i < used_sacks) {
1761 u32 start_seq = sp[i].start_seq;
1762 u32 end_seq = sp[i].end_seq;
1763 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1764 struct tcp_sack_block *next_dup = NULL;
1766 if (found_dup_sack && ((i + 1) == first_sack_index))
1767 next_dup = &sp[i + 1];
1769 /* Skip too early cached blocks */
1770 while (tcp_sack_cache_ok(tp, cache) &&
1771 !before(start_seq, cache->end_seq))
1774 /* Can skip some work by looking recv_sack_cache? */
1775 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1776 after(end_seq, cache->start_seq)) {
1779 if (before(start_seq, cache->start_seq)) {
1780 skb = tcp_sacktag_skip(skb, sk, state,
1782 skb = tcp_sacktag_walk(skb, sk, next_dup,
1789 /* Rest of the block already fully processed? */
1790 if (!after(end_seq, cache->end_seq))
1793 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1797 /* ...tail remains todo... */
1798 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1799 /* ...but better entrypoint exists! */
1800 skb = tcp_highest_sack(sk);
1807 skb = tcp_sacktag_skip(skb, sk, state, cache->end_seq);
1808 /* Check overlap against next cached too (past this one already) */
1813 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1814 skb = tcp_highest_sack(sk);
1818 skb = tcp_sacktag_skip(skb, sk, state, start_seq);
1821 skb = tcp_sacktag_walk(skb, sk, next_dup, state,
1822 start_seq, end_seq, dup_sack);
1828 /* Clear the head of the cache sack blocks so we can skip it next time */
1829 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1830 tp->recv_sack_cache[i].start_seq = 0;
1831 tp->recv_sack_cache[i].end_seq = 0;
1833 for (j = 0; j < used_sacks; j++)
1834 tp->recv_sack_cache[i++] = sp[j];
1836 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Loss || tp->undo_marker)
1837 tcp_check_sack_reordering(sk, state->reord, 0);
1839 tcp_verify_left_out(tp);
1842 #if FASTRETRANS_DEBUG > 0
1843 WARN_ON((int)tp->sacked_out < 0);
1844 WARN_ON((int)tp->lost_out < 0);
1845 WARN_ON((int)tp->retrans_out < 0);
1846 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1851 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1852 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1854 static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
1858 holes = max(tp->lost_out, 1U);
1859 holes = min(holes, tp->packets_out);
1861 if ((tp->sacked_out + holes) > tp->packets_out) {
1862 tp->sacked_out = tp->packets_out - holes;
1868 /* If we receive more dupacks than we expected counting segments
1869 * in assumption of absent reordering, interpret this as reordering.
1870 * The only another reason could be bug in receiver TCP.
1872 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1874 struct tcp_sock *tp = tcp_sk(sk);
1876 if (!tcp_limit_reno_sacked(tp))
1879 tp->reordering = min_t(u32, tp->packets_out + addend,
1880 sock_net(sk)->ipv4.sysctl_tcp_max_reordering);
1882 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRENOREORDER);
1885 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1887 static void tcp_add_reno_sack(struct sock *sk)
1889 struct tcp_sock *tp = tcp_sk(sk);
1890 u32 prior_sacked = tp->sacked_out;
1893 tcp_check_reno_reordering(sk, 0);
1894 if (tp->sacked_out > prior_sacked)
1895 tp->delivered++; /* Some out-of-order packet is delivered */
1896 tcp_verify_left_out(tp);
1899 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1901 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1903 struct tcp_sock *tp = tcp_sk(sk);
1906 /* One ACK acked hole. The rest eat duplicate ACKs. */
1907 tp->delivered += max_t(int, acked - tp->sacked_out, 1);
1908 if (acked - 1 >= tp->sacked_out)
1911 tp->sacked_out -= acked - 1;
1913 tcp_check_reno_reordering(sk, acked);
1914 tcp_verify_left_out(tp);
1917 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1922 void tcp_clear_retrans(struct tcp_sock *tp)
1924 tp->retrans_out = 0;
1926 tp->undo_marker = 0;
1927 tp->undo_retrans = -1;
1931 static inline void tcp_init_undo(struct tcp_sock *tp)
1933 tp->undo_marker = tp->snd_una;
1934 /* Retransmission still in flight may cause DSACKs later. */
1935 tp->undo_retrans = tp->retrans_out ? : -1;
1938 static bool tcp_is_rack(const struct sock *sk)
1940 return sock_net(sk)->ipv4.sysctl_tcp_recovery & TCP_RACK_LOSS_DETECTION;
1943 /* If we detect SACK reneging, forget all SACK information
1944 * and reset tags completely, otherwise preserve SACKs. If receiver
1945 * dropped its ofo queue, we will know this due to reneging detection.
1947 static void tcp_timeout_mark_lost(struct sock *sk)
1949 struct tcp_sock *tp = tcp_sk(sk);
1950 struct sk_buff *skb, *head;
1951 bool is_reneg; /* is receiver reneging on SACKs? */
1953 head = tcp_rtx_queue_head(sk);
1954 is_reneg = head && (TCP_SKB_CB(head)->sacked & TCPCB_SACKED_ACKED);
1956 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
1958 /* Mark SACK reneging until we recover from this loss event. */
1959 tp->is_sack_reneg = 1;
1960 } else if (tcp_is_reno(tp)) {
1961 tcp_reset_reno_sack(tp);
1965 skb_rbtree_walk_from(skb) {
1967 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1968 else if (tcp_is_rack(sk) && skb != head &&
1969 tcp_rack_skb_timeout(tp, skb, 0) > 0)
1970 continue; /* Don't mark recently sent ones lost yet */
1971 tcp_mark_skb_lost(sk, skb);
1973 tcp_verify_left_out(tp);
1974 tcp_clear_all_retrans_hints(tp);
1977 /* Enter Loss state. */
1978 void tcp_enter_loss(struct sock *sk)
1980 const struct inet_connection_sock *icsk = inet_csk(sk);
1981 struct tcp_sock *tp = tcp_sk(sk);
1982 struct net *net = sock_net(sk);
1983 bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery;
1985 tcp_timeout_mark_lost(sk);
1987 /* Reduce ssthresh if it has not yet been made inside this window. */
1988 if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
1989 !after(tp->high_seq, tp->snd_una) ||
1990 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1991 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1992 tp->prior_cwnd = tp->snd_cwnd;
1993 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1994 tcp_ca_event(sk, CA_EVENT_LOSS);
1997 tp->snd_cwnd = tcp_packets_in_flight(tp) + 1;
1998 tp->snd_cwnd_cnt = 0;
1999 tp->snd_cwnd_stamp = tcp_jiffies32;
2001 /* Timeout in disordered state after receiving substantial DUPACKs
2002 * suggests that the degree of reordering is over-estimated.
2004 if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
2005 tp->sacked_out >= net->ipv4.sysctl_tcp_reordering)
2006 tp->reordering = min_t(unsigned int, tp->reordering,
2007 net->ipv4.sysctl_tcp_reordering);
2008 tcp_set_ca_state(sk, TCP_CA_Loss);
2009 tp->high_seq = tp->snd_nxt;
2010 tcp_ecn_queue_cwr(tp);
2012 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
2013 * loss recovery is underway except recurring timeout(s) on
2014 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
2016 tp->frto = net->ipv4.sysctl_tcp_frto &&
2017 (new_recovery || icsk->icsk_retransmits) &&
2018 !inet_csk(sk)->icsk_mtup.probe_size;
2021 /* If ACK arrived pointing to a remembered SACK, it means that our
2022 * remembered SACKs do not reflect real state of receiver i.e.
2023 * receiver _host_ is heavily congested (or buggy).
2025 * To avoid big spurious retransmission bursts due to transient SACK
2026 * scoreboard oddities that look like reneging, we give the receiver a
2027 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
2028 * restore sanity to the SACK scoreboard. If the apparent reneging
2029 * persists until this RTO then we'll clear the SACK scoreboard.
2031 static bool tcp_check_sack_reneging(struct sock *sk, int flag)
2033 if (flag & FLAG_SACK_RENEGING) {
2034 struct tcp_sock *tp = tcp_sk(sk);
2035 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4),
2036 msecs_to_jiffies(10));
2038 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2039 delay, TCP_RTO_MAX);
2045 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2046 * counter when SACK is enabled (without SACK, sacked_out is used for
2049 * With reordering, holes may still be in flight, so RFC3517 recovery
2050 * uses pure sacked_out (total number of SACKed segments) even though
2051 * it violates the RFC that uses duplicate ACKs, often these are equal
2052 * but when e.g. out-of-window ACKs or packet duplication occurs,
2053 * they differ. Since neither occurs due to loss, TCP should really
2056 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2058 return tp->sacked_out + 1;
2061 /* Linux NewReno/SACK/ECN state machine.
2062 * --------------------------------------
2064 * "Open" Normal state, no dubious events, fast path.
2065 * "Disorder" In all the respects it is "Open",
2066 * but requires a bit more attention. It is entered when
2067 * we see some SACKs or dupacks. It is split of "Open"
2068 * mainly to move some processing from fast path to slow one.
2069 * "CWR" CWND was reduced due to some Congestion Notification event.
2070 * It can be ECN, ICMP source quench, local device congestion.
2071 * "Recovery" CWND was reduced, we are fast-retransmitting.
2072 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2074 * tcp_fastretrans_alert() is entered:
2075 * - each incoming ACK, if state is not "Open"
2076 * - when arrived ACK is unusual, namely:
2081 * Counting packets in flight is pretty simple.
2083 * in_flight = packets_out - left_out + retrans_out
2085 * packets_out is SND.NXT-SND.UNA counted in packets.
2087 * retrans_out is number of retransmitted segments.
2089 * left_out is number of segments left network, but not ACKed yet.
2091 * left_out = sacked_out + lost_out
2093 * sacked_out: Packets, which arrived to receiver out of order
2094 * and hence not ACKed. With SACKs this number is simply
2095 * amount of SACKed data. Even without SACKs
2096 * it is easy to give pretty reliable estimate of this number,
2097 * counting duplicate ACKs.
2099 * lost_out: Packets lost by network. TCP has no explicit
2100 * "loss notification" feedback from network (for now).
2101 * It means that this number can be only _guessed_.
2102 * Actually, it is the heuristics to predict lossage that
2103 * distinguishes different algorithms.
2105 * F.e. after RTO, when all the queue is considered as lost,
2106 * lost_out = packets_out and in_flight = retrans_out.
2108 * Essentially, we have now a few algorithms detecting
2111 * If the receiver supports SACK:
2113 * RFC6675/3517: It is the conventional algorithm. A packet is
2114 * considered lost if the number of higher sequence packets
2115 * SACKed is greater than or equal the DUPACK thoreshold
2116 * (reordering). This is implemented in tcp_mark_head_lost and
2117 * tcp_update_scoreboard.
2119 * RACK (draft-ietf-tcpm-rack-01): it is a newer algorithm
2120 * (2017-) that checks timing instead of counting DUPACKs.
2121 * Essentially a packet is considered lost if it's not S/ACKed
2122 * after RTT + reordering_window, where both metrics are
2123 * dynamically measured and adjusted. This is implemented in
2124 * tcp_rack_mark_lost.
2126 * If the receiver does not support SACK:
2128 * NewReno (RFC6582): in Recovery we assume that one segment
2129 * is lost (classic Reno). While we are in Recovery and
2130 * a partial ACK arrives, we assume that one more packet
2131 * is lost (NewReno). This heuristics are the same in NewReno
2134 * Really tricky (and requiring careful tuning) part of algorithm
2135 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2136 * The first determines the moment _when_ we should reduce CWND and,
2137 * hence, slow down forward transmission. In fact, it determines the moment
2138 * when we decide that hole is caused by loss, rather than by a reorder.
2140 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2141 * holes, caused by lost packets.
2143 * And the most logically complicated part of algorithm is undo
2144 * heuristics. We detect false retransmits due to both too early
2145 * fast retransmit (reordering) and underestimated RTO, analyzing
2146 * timestamps and D-SACKs. When we detect that some segments were
2147 * retransmitted by mistake and CWND reduction was wrong, we undo
2148 * window reduction and abort recovery phase. This logic is hidden
2149 * inside several functions named tcp_try_undo_<something>.
2152 /* This function decides, when we should leave Disordered state
2153 * and enter Recovery phase, reducing congestion window.
2155 * Main question: may we further continue forward transmission
2156 * with the same cwnd?
2158 static bool tcp_time_to_recover(struct sock *sk, int flag)
2160 struct tcp_sock *tp = tcp_sk(sk);
2162 /* Trick#1: The loss is proven. */
2166 /* Not-A-Trick#2 : Classic rule... */
2167 if (!tcp_is_rack(sk) && tcp_dupack_heuristics(tp) > tp->reordering)
2173 /* Detect loss in event "A" above by marking head of queue up as lost.
2174 * For non-SACK(Reno) senders, the first "packets" number of segments
2175 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2176 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2177 * the maximum SACKed segments to pass before reaching this limit.
2179 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2181 struct tcp_sock *tp = tcp_sk(sk);
2182 struct sk_buff *skb;
2183 int cnt, oldcnt, lost;
2185 /* Use SACK to deduce losses of new sequences sent during recovery */
2186 const u32 loss_high = tcp_is_sack(tp) ? tp->snd_nxt : tp->high_seq;
2188 WARN_ON(packets > tp->packets_out);
2189 skb = tp->lost_skb_hint;
2191 /* Head already handled? */
2192 if (mark_head && after(TCP_SKB_CB(skb)->seq, tp->snd_una))
2194 cnt = tp->lost_cnt_hint;
2196 skb = tcp_rtx_queue_head(sk);
2200 skb_rbtree_walk_from(skb) {
2201 /* TODO: do this better */
2202 /* this is not the most efficient way to do this... */
2203 tp->lost_skb_hint = skb;
2204 tp->lost_cnt_hint = cnt;
2206 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2210 if (tcp_is_reno(tp) ||
2211 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2212 cnt += tcp_skb_pcount(skb);
2214 if (cnt > packets) {
2215 if (tcp_is_sack(tp) ||
2216 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
2217 (oldcnt >= packets))
2220 mss = tcp_skb_mss(skb);
2221 /* If needed, chop off the prefix to mark as lost. */
2222 lost = (packets - oldcnt) * mss;
2223 if (lost < skb->len &&
2224 tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
2225 lost, mss, GFP_ATOMIC) < 0)
2230 tcp_skb_mark_lost(tp, skb);
2235 tcp_verify_left_out(tp);
2238 /* Account newly detected lost packet(s) */
2240 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2242 struct tcp_sock *tp = tcp_sk(sk);
2244 if (tcp_is_sack(tp)) {
2245 int sacked_upto = tp->sacked_out - tp->reordering;
2246 if (sacked_upto >= 0)
2247 tcp_mark_head_lost(sk, sacked_upto, 0);
2248 else if (fast_rexmit)
2249 tcp_mark_head_lost(sk, 1, 1);
2253 static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when)
2255 return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2256 before(tp->rx_opt.rcv_tsecr, when);
2259 /* skb is spurious retransmitted if the returned timestamp echo
2260 * reply is prior to the skb transmission time
2262 static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp,
2263 const struct sk_buff *skb)
2265 return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) &&
2266 tcp_tsopt_ecr_before(tp, tcp_skb_timestamp(skb));
2269 /* Nothing was retransmitted or returned timestamp is less
2270 * than timestamp of the first retransmission.
2272 static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2274 return !tp->retrans_stamp ||
2275 tcp_tsopt_ecr_before(tp, tp->retrans_stamp);
2278 /* Undo procedures. */
2280 /* We can clear retrans_stamp when there are no retransmissions in the
2281 * window. It would seem that it is trivially available for us in
2282 * tp->retrans_out, however, that kind of assumptions doesn't consider
2283 * what will happen if errors occur when sending retransmission for the
2284 * second time. ...It could the that such segment has only
2285 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2286 * the head skb is enough except for some reneging corner cases that
2287 * are not worth the effort.
2289 * Main reason for all this complexity is the fact that connection dying
2290 * time now depends on the validity of the retrans_stamp, in particular,
2291 * that successive retransmissions of a segment must not advance
2292 * retrans_stamp under any conditions.
2294 static bool tcp_any_retrans_done(const struct sock *sk)
2296 const struct tcp_sock *tp = tcp_sk(sk);
2297 struct sk_buff *skb;
2299 if (tp->retrans_out)
2302 skb = tcp_rtx_queue_head(sk);
2303 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2309 static void DBGUNDO(struct sock *sk, const char *msg)
2311 #if FASTRETRANS_DEBUG > 1
2312 struct tcp_sock *tp = tcp_sk(sk);
2313 struct inet_sock *inet = inet_sk(sk);
2315 if (sk->sk_family == AF_INET) {
2316 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2318 &inet->inet_daddr, ntohs(inet->inet_dport),
2319 tp->snd_cwnd, tcp_left_out(tp),
2320 tp->snd_ssthresh, tp->prior_ssthresh,
2323 #if IS_ENABLED(CONFIG_IPV6)
2324 else if (sk->sk_family == AF_INET6) {
2325 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2327 &sk->sk_v6_daddr, ntohs(inet->inet_dport),
2328 tp->snd_cwnd, tcp_left_out(tp),
2329 tp->snd_ssthresh, tp->prior_ssthresh,
2336 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
2338 struct tcp_sock *tp = tcp_sk(sk);
2341 struct sk_buff *skb;
2343 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2344 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2347 tcp_clear_all_retrans_hints(tp);
2350 if (tp->prior_ssthresh) {
2351 const struct inet_connection_sock *icsk = inet_csk(sk);
2353 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2355 if (tp->prior_ssthresh > tp->snd_ssthresh) {
2356 tp->snd_ssthresh = tp->prior_ssthresh;
2357 tcp_ecn_withdraw_cwr(tp);
2360 tp->snd_cwnd_stamp = tcp_jiffies32;
2361 tp->undo_marker = 0;
2362 tp->rack.advanced = 1; /* Force RACK to re-exam losses */
2365 static inline bool tcp_may_undo(const struct tcp_sock *tp)
2367 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2370 /* People celebrate: "We love our President!" */
2371 static bool tcp_try_undo_recovery(struct sock *sk)
2373 struct tcp_sock *tp = tcp_sk(sk);
2375 if (tcp_may_undo(tp)) {
2378 /* Happy end! We did not retransmit anything
2379 * or our original transmission succeeded.
2381 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2382 tcp_undo_cwnd_reduction(sk, false);
2383 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2384 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2386 mib_idx = LINUX_MIB_TCPFULLUNDO;
2388 NET_INC_STATS(sock_net(sk), mib_idx);
2389 } else if (tp->rack.reo_wnd_persist) {
2390 tp->rack.reo_wnd_persist--;
2392 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2393 /* Hold old state until something *above* high_seq
2394 * is ACKed. For Reno it is MUST to prevent false
2395 * fast retransmits (RFC2582). SACK TCP is safe. */
2396 if (!tcp_any_retrans_done(sk))
2397 tp->retrans_stamp = 0;
2400 tcp_set_ca_state(sk, TCP_CA_Open);
2401 tp->is_sack_reneg = 0;
2405 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2406 static bool tcp_try_undo_dsack(struct sock *sk)
2408 struct tcp_sock *tp = tcp_sk(sk);
2410 if (tp->undo_marker && !tp->undo_retrans) {
2411 tp->rack.reo_wnd_persist = min(TCP_RACK_RECOVERY_THRESH,
2412 tp->rack.reo_wnd_persist + 1);
2413 DBGUNDO(sk, "D-SACK");
2414 tcp_undo_cwnd_reduction(sk, false);
2415 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2421 /* Undo during loss recovery after partial ACK or using F-RTO. */
2422 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2424 struct tcp_sock *tp = tcp_sk(sk);
2426 if (frto_undo || tcp_may_undo(tp)) {
2427 tcp_undo_cwnd_reduction(sk, true);
2429 DBGUNDO(sk, "partial loss");
2430 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2432 NET_INC_STATS(sock_net(sk),
2433 LINUX_MIB_TCPSPURIOUSRTOS);
2434 inet_csk(sk)->icsk_retransmits = 0;
2435 if (frto_undo || tcp_is_sack(tp)) {
2436 tcp_set_ca_state(sk, TCP_CA_Open);
2437 tp->is_sack_reneg = 0;
2444 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
2445 * It computes the number of packets to send (sndcnt) based on packets newly
2447 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2448 * cwnd reductions across a full RTT.
2449 * 2) Otherwise PRR uses packet conservation to send as much as delivered.
2450 * But when the retransmits are acked without further losses, PRR
2451 * slow starts cwnd up to ssthresh to speed up the recovery.
2453 static void tcp_init_cwnd_reduction(struct sock *sk)
2455 struct tcp_sock *tp = tcp_sk(sk);
2457 tp->high_seq = tp->snd_nxt;
2458 tp->tlp_high_seq = 0;
2459 tp->snd_cwnd_cnt = 0;
2460 tp->prior_cwnd = tp->snd_cwnd;
2461 tp->prr_delivered = 0;
2463 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2464 tcp_ecn_queue_cwr(tp);
2467 void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked, int flag)
2469 struct tcp_sock *tp = tcp_sk(sk);
2471 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2473 if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd))
2476 tp->prr_delivered += newly_acked_sacked;
2478 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2480 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2481 } else if ((flag & FLAG_RETRANS_DATA_ACKED) &&
2482 !(flag & FLAG_LOST_RETRANS)) {
2483 sndcnt = min_t(int, delta,
2484 max_t(int, tp->prr_delivered - tp->prr_out,
2485 newly_acked_sacked) + 1);
2487 sndcnt = min(delta, newly_acked_sacked);
2489 /* Force a fast retransmit upon entering fast recovery */
2490 sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1));
2491 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2494 static inline void tcp_end_cwnd_reduction(struct sock *sk)
2496 struct tcp_sock *tp = tcp_sk(sk);
2498 if (inet_csk(sk)->icsk_ca_ops->cong_control)
2501 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2502 if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH &&
2503 (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR || tp->undo_marker)) {
2504 tp->snd_cwnd = tp->snd_ssthresh;
2505 tp->snd_cwnd_stamp = tcp_jiffies32;
2507 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2510 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2511 void tcp_enter_cwr(struct sock *sk)
2513 struct tcp_sock *tp = tcp_sk(sk);
2515 tp->prior_ssthresh = 0;
2516 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2517 tp->undo_marker = 0;
2518 tcp_init_cwnd_reduction(sk);
2519 tcp_set_ca_state(sk, TCP_CA_CWR);
2522 EXPORT_SYMBOL(tcp_enter_cwr);
2524 static void tcp_try_keep_open(struct sock *sk)
2526 struct tcp_sock *tp = tcp_sk(sk);
2527 int state = TCP_CA_Open;
2529 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2530 state = TCP_CA_Disorder;
2532 if (inet_csk(sk)->icsk_ca_state != state) {
2533 tcp_set_ca_state(sk, state);
2534 tp->high_seq = tp->snd_nxt;
2538 static void tcp_try_to_open(struct sock *sk, int flag)
2540 struct tcp_sock *tp = tcp_sk(sk);
2542 tcp_verify_left_out(tp);
2544 if (!tcp_any_retrans_done(sk))
2545 tp->retrans_stamp = 0;
2547 if (flag & FLAG_ECE)
2550 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2551 tcp_try_keep_open(sk);
2555 static void tcp_mtup_probe_failed(struct sock *sk)
2557 struct inet_connection_sock *icsk = inet_csk(sk);
2559 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2560 icsk->icsk_mtup.probe_size = 0;
2561 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPFAIL);
2564 static void tcp_mtup_probe_success(struct sock *sk)
2566 struct tcp_sock *tp = tcp_sk(sk);
2567 struct inet_connection_sock *icsk = inet_csk(sk);
2569 /* FIXME: breaks with very large cwnd */
2570 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2571 tp->snd_cwnd = tp->snd_cwnd *
2572 tcp_mss_to_mtu(sk, tp->mss_cache) /
2573 icsk->icsk_mtup.probe_size;
2574 tp->snd_cwnd_cnt = 0;
2575 tp->snd_cwnd_stamp = tcp_jiffies32;
2576 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2578 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2579 icsk->icsk_mtup.probe_size = 0;
2580 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2581 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS);
2584 /* Do a simple retransmit without using the backoff mechanisms in
2585 * tcp_timer. This is used for path mtu discovery.
2586 * The socket is already locked here.
2588 void tcp_simple_retransmit(struct sock *sk)
2590 const struct inet_connection_sock *icsk = inet_csk(sk);
2591 struct tcp_sock *tp = tcp_sk(sk);
2592 struct sk_buff *skb;
2593 unsigned int mss = tcp_current_mss(sk);
2595 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2596 if (tcp_skb_seglen(skb) > mss &&
2597 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2598 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2599 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2600 tp->retrans_out -= tcp_skb_pcount(skb);
2602 tcp_skb_mark_lost_uncond_verify(tp, skb);
2606 tcp_clear_retrans_hints_partial(tp);
2611 if (tcp_is_reno(tp))
2612 tcp_limit_reno_sacked(tp);
2614 tcp_verify_left_out(tp);
2616 /* Don't muck with the congestion window here.
2617 * Reason is that we do not increase amount of _data_
2618 * in network, but units changed and effective
2619 * cwnd/ssthresh really reduced now.
2621 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2622 tp->high_seq = tp->snd_nxt;
2623 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2624 tp->prior_ssthresh = 0;
2625 tp->undo_marker = 0;
2626 tcp_set_ca_state(sk, TCP_CA_Loss);
2628 tcp_xmit_retransmit_queue(sk);
2630 EXPORT_SYMBOL(tcp_simple_retransmit);
2632 void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2634 struct tcp_sock *tp = tcp_sk(sk);
2637 if (tcp_is_reno(tp))
2638 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2640 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2642 NET_INC_STATS(sock_net(sk), mib_idx);
2644 tp->prior_ssthresh = 0;
2647 if (!tcp_in_cwnd_reduction(sk)) {
2649 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2650 tcp_init_cwnd_reduction(sk);
2652 tcp_set_ca_state(sk, TCP_CA_Recovery);
2655 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2656 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2658 static void tcp_process_loss(struct sock *sk, int flag, bool is_dupack,
2661 struct tcp_sock *tp = tcp_sk(sk);
2662 bool recovered = !before(tp->snd_una, tp->high_seq);
2664 if ((flag & FLAG_SND_UNA_ADVANCED) &&
2665 tcp_try_undo_loss(sk, false))
2668 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2669 /* Step 3.b. A timeout is spurious if not all data are
2670 * lost, i.e., never-retransmitted data are (s)acked.
2672 if ((flag & FLAG_ORIG_SACK_ACKED) &&
2673 tcp_try_undo_loss(sk, true))
2676 if (after(tp->snd_nxt, tp->high_seq)) {
2677 if (flag & FLAG_DATA_SACKED || is_dupack)
2678 tp->frto = 0; /* Step 3.a. loss was real */
2679 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2680 tp->high_seq = tp->snd_nxt;
2681 /* Step 2.b. Try send new data (but deferred until cwnd
2682 * is updated in tcp_ack()). Otherwise fall back to
2683 * the conventional recovery.
2685 if (!tcp_write_queue_empty(sk) &&
2686 after(tcp_wnd_end(tp), tp->snd_nxt)) {
2687 *rexmit = REXMIT_NEW;
2695 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2696 tcp_try_undo_recovery(sk);
2699 if (tcp_is_reno(tp)) {
2700 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2701 * delivered. Lower inflight to clock out (re)tranmissions.
2703 if (after(tp->snd_nxt, tp->high_seq) && is_dupack)
2704 tcp_add_reno_sack(sk);
2705 else if (flag & FLAG_SND_UNA_ADVANCED)
2706 tcp_reset_reno_sack(tp);
2708 *rexmit = REXMIT_LOST;
2711 /* Undo during fast recovery after partial ACK. */
2712 static bool tcp_try_undo_partial(struct sock *sk, u32 prior_snd_una)
2714 struct tcp_sock *tp = tcp_sk(sk);
2716 if (tp->undo_marker && tcp_packet_delayed(tp)) {
2717 /* Plain luck! Hole if filled with delayed
2718 * packet, rather than with a retransmit. Check reordering.
2720 tcp_check_sack_reordering(sk, prior_snd_una, 1);
2722 /* We are getting evidence that the reordering degree is higher
2723 * than we realized. If there are no retransmits out then we
2724 * can undo. Otherwise we clock out new packets but do not
2725 * mark more packets lost or retransmit more.
2727 if (tp->retrans_out)
2730 if (!tcp_any_retrans_done(sk))
2731 tp->retrans_stamp = 0;
2733 DBGUNDO(sk, "partial recovery");
2734 tcp_undo_cwnd_reduction(sk, true);
2735 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2736 tcp_try_keep_open(sk);
2742 static void tcp_identify_packet_loss(struct sock *sk, int *ack_flag)
2744 struct tcp_sock *tp = tcp_sk(sk);
2746 if (tcp_rtx_queue_empty(sk))
2749 if (unlikely(tcp_is_reno(tp))) {
2750 tcp_newreno_mark_lost(sk, *ack_flag & FLAG_SND_UNA_ADVANCED);
2751 } else if (tcp_is_rack(sk)) {
2752 u32 prior_retrans = tp->retrans_out;
2754 tcp_rack_mark_lost(sk);
2755 if (prior_retrans > tp->retrans_out)
2756 *ack_flag |= FLAG_LOST_RETRANS;
2760 static bool tcp_force_fast_retransmit(struct sock *sk)
2762 struct tcp_sock *tp = tcp_sk(sk);
2764 return after(tcp_highest_sack_seq(tp),
2765 tp->snd_una + tp->reordering * tp->mss_cache);
2768 /* Process an event, which can update packets-in-flight not trivially.
2769 * Main goal of this function is to calculate new estimate for left_out,
2770 * taking into account both packets sitting in receiver's buffer and
2771 * packets lost by network.
2773 * Besides that it updates the congestion state when packet loss or ECN
2774 * is detected. But it does not reduce the cwnd, it is done by the
2775 * congestion control later.
2777 * It does _not_ decide what to send, it is made in function
2778 * tcp_xmit_retransmit_queue().
2780 static void tcp_fastretrans_alert(struct sock *sk, const u32 prior_snd_una,
2781 bool is_dupack, int *ack_flag, int *rexmit)
2783 struct inet_connection_sock *icsk = inet_csk(sk);
2784 struct tcp_sock *tp = tcp_sk(sk);
2785 int fast_rexmit = 0, flag = *ack_flag;
2786 bool do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2787 tcp_force_fast_retransmit(sk));
2789 if (!tp->packets_out && tp->sacked_out)
2792 /* Now state machine starts.
2793 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2794 if (flag & FLAG_ECE)
2795 tp->prior_ssthresh = 0;
2797 /* B. In all the states check for reneging SACKs. */
2798 if (tcp_check_sack_reneging(sk, flag))
2801 /* C. Check consistency of the current state. */
2802 tcp_verify_left_out(tp);
2804 /* D. Check state exit conditions. State can be terminated
2805 * when high_seq is ACKed. */
2806 if (icsk->icsk_ca_state == TCP_CA_Open) {
2807 WARN_ON(tp->retrans_out != 0);
2808 tp->retrans_stamp = 0;
2809 } else if (!before(tp->snd_una, tp->high_seq)) {
2810 switch (icsk->icsk_ca_state) {
2812 /* CWR is to be held something *above* high_seq
2813 * is ACKed for CWR bit to reach receiver. */
2814 if (tp->snd_una != tp->high_seq) {
2815 tcp_end_cwnd_reduction(sk);
2816 tcp_set_ca_state(sk, TCP_CA_Open);
2820 case TCP_CA_Recovery:
2821 if (tcp_is_reno(tp))
2822 tcp_reset_reno_sack(tp);
2823 if (tcp_try_undo_recovery(sk))
2825 tcp_end_cwnd_reduction(sk);
2830 /* E. Process state. */
2831 switch (icsk->icsk_ca_state) {
2832 case TCP_CA_Recovery:
2833 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2834 if (tcp_is_reno(tp) && is_dupack)
2835 tcp_add_reno_sack(sk);
2837 if (tcp_try_undo_partial(sk, prior_snd_una))
2839 /* Partial ACK arrived. Force fast retransmit. */
2840 do_lost = tcp_is_reno(tp) ||
2841 tcp_force_fast_retransmit(sk);
2843 if (tcp_try_undo_dsack(sk)) {
2844 tcp_try_keep_open(sk);
2847 tcp_identify_packet_loss(sk, ack_flag);
2850 tcp_process_loss(sk, flag, is_dupack, rexmit);
2851 tcp_identify_packet_loss(sk, ack_flag);
2852 if (!(icsk->icsk_ca_state == TCP_CA_Open ||
2853 (*ack_flag & FLAG_LOST_RETRANS)))
2855 /* Change state if cwnd is undone or retransmits are lost */
2858 if (tcp_is_reno(tp)) {
2859 if (flag & FLAG_SND_UNA_ADVANCED)
2860 tcp_reset_reno_sack(tp);
2862 tcp_add_reno_sack(sk);
2865 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
2866 tcp_try_undo_dsack(sk);
2868 tcp_identify_packet_loss(sk, ack_flag);
2869 if (!tcp_time_to_recover(sk, flag)) {
2870 tcp_try_to_open(sk, flag);
2874 /* MTU probe failure: don't reduce cwnd */
2875 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2876 icsk->icsk_mtup.probe_size &&
2877 tp->snd_una == tp->mtu_probe.probe_seq_start) {
2878 tcp_mtup_probe_failed(sk);
2879 /* Restores the reduction we did in tcp_mtup_probe() */
2881 tcp_simple_retransmit(sk);
2885 /* Otherwise enter Recovery state */
2886 tcp_enter_recovery(sk, (flag & FLAG_ECE));
2890 if (!tcp_is_rack(sk) && do_lost)
2891 tcp_update_scoreboard(sk, fast_rexmit);
2892 *rexmit = REXMIT_LOST;
2895 static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us, const int flag)
2897 u32 wlen = sock_net(sk)->ipv4.sysctl_tcp_min_rtt_wlen * HZ;
2898 struct tcp_sock *tp = tcp_sk(sk);
2900 if ((flag & FLAG_ACK_MAYBE_DELAYED) && rtt_us > tcp_min_rtt(tp)) {
2901 /* If the remote keeps returning delayed ACKs, eventually
2902 * the min filter would pick it up and overestimate the
2903 * prop. delay when it expires. Skip suspected delayed ACKs.
2907 minmax_running_min(&tp->rtt_min, wlen, tcp_jiffies32,
2908 rtt_us ? : jiffies_to_usecs(1));
2911 static bool tcp_ack_update_rtt(struct sock *sk, const int flag,
2912 long seq_rtt_us, long sack_rtt_us,
2913 long ca_rtt_us, struct rate_sample *rs)
2915 const struct tcp_sock *tp = tcp_sk(sk);
2917 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2918 * broken middle-boxes or peers may corrupt TS-ECR fields. But
2919 * Karn's algorithm forbids taking RTT if some retransmitted data
2920 * is acked (RFC6298).
2923 seq_rtt_us = sack_rtt_us;
2925 /* RTTM Rule: A TSecr value received in a segment is used to
2926 * update the averaged RTT measurement only if the segment
2927 * acknowledges some new data, i.e., only if it advances the
2928 * left edge of the send window.
2929 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2931 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2932 flag & FLAG_ACKED) {
2933 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr;
2934 u32 delta_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
2936 seq_rtt_us = ca_rtt_us = delta_us;
2938 rs->rtt_us = ca_rtt_us; /* RTT of last (S)ACKed packet (or -1) */
2942 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
2943 * always taken together with ACK, SACK, or TS-opts. Any negative
2944 * values will be skipped with the seq_rtt_us < 0 check above.
2946 tcp_update_rtt_min(sk, ca_rtt_us, flag);
2947 tcp_rtt_estimator(sk, seq_rtt_us);
2950 /* RFC6298: only reset backoff on valid RTT measurement. */
2951 inet_csk(sk)->icsk_backoff = 0;
2955 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
2956 void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req)
2958 struct rate_sample rs;
2961 if (req && !req->num_retrans && tcp_rsk(req)->snt_synack)
2962 rtt_us = tcp_stamp_us_delta(tcp_clock_us(), tcp_rsk(req)->snt_synack);
2964 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us, &rs);
2968 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
2970 const struct inet_connection_sock *icsk = inet_csk(sk);
2972 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
2973 tcp_sk(sk)->snd_cwnd_stamp = tcp_jiffies32;
2976 /* Restart timer after forward progress on connection.
2977 * RFC2988 recommends to restart timer to now+rto.
2979 void tcp_rearm_rto(struct sock *sk)
2981 const struct inet_connection_sock *icsk = inet_csk(sk);
2982 struct tcp_sock *tp = tcp_sk(sk);
2984 /* If the retrans timer is currently being used by Fast Open
2985 * for SYN-ACK retrans purpose, stay put.
2987 if (tp->fastopen_rsk)
2990 if (!tp->packets_out) {
2991 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
2993 u32 rto = inet_csk(sk)->icsk_rto;
2994 /* Offset the time elapsed after installing regular RTO */
2995 if (icsk->icsk_pending == ICSK_TIME_REO_TIMEOUT ||
2996 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
2997 s64 delta_us = tcp_rto_delta_us(sk);
2998 /* delta_us may not be positive if the socket is locked
2999 * when the retrans timer fires and is rescheduled.
3001 rto = usecs_to_jiffies(max_t(int, delta_us, 1));
3003 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
3008 /* Try to schedule a loss probe; if that doesn't work, then schedule an RTO. */
3009 static void tcp_set_xmit_timer(struct sock *sk)
3011 if (!tcp_schedule_loss_probe(sk, true))
3015 /* If we get here, the whole TSO packet has not been acked. */
3016 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3018 struct tcp_sock *tp = tcp_sk(sk);
3021 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3023 packets_acked = tcp_skb_pcount(skb);
3024 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3026 packets_acked -= tcp_skb_pcount(skb);
3028 if (packets_acked) {
3029 BUG_ON(tcp_skb_pcount(skb) == 0);
3030 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3033 return packets_acked;
3036 static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
3039 const struct skb_shared_info *shinfo;
3041 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3042 if (likely(!TCP_SKB_CB(skb)->txstamp_ack))
3045 shinfo = skb_shinfo(skb);
3046 if (!before(shinfo->tskey, prior_snd_una) &&
3047 before(shinfo->tskey, tcp_sk(sk)->snd_una)) {
3048 tcp_skb_tsorted_save(skb) {
3049 __skb_tstamp_tx(skb, NULL, sk, SCM_TSTAMP_ACK);
3050 } tcp_skb_tsorted_restore(skb);
3054 /* Remove acknowledged frames from the retransmission queue. If our packet
3055 * is before the ack sequence we can discard it as it's confirmed to have
3056 * arrived at the other end.
3058 static int tcp_clean_rtx_queue(struct sock *sk, u32 prior_fack,
3060 struct tcp_sacktag_state *sack)
3062 const struct inet_connection_sock *icsk = inet_csk(sk);
3063 u64 first_ackt, last_ackt;
3064 struct tcp_sock *tp = tcp_sk(sk);
3065 u32 prior_sacked = tp->sacked_out;
3066 u32 reord = tp->snd_nxt; /* lowest acked un-retx un-sacked seq */
3067 struct sk_buff *skb, *next;
3068 bool fully_acked = true;
3069 long sack_rtt_us = -1L;
3070 long seq_rtt_us = -1L;
3071 long ca_rtt_us = -1L;
3073 u32 last_in_flight = 0;
3079 for (skb = skb_rb_first(&sk->tcp_rtx_queue); skb; skb = next) {
3080 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3081 const u32 start_seq = scb->seq;
3082 u8 sacked = scb->sacked;
3085 tcp_ack_tstamp(sk, skb, prior_snd_una);
3087 /* Determine how many packets and what bytes were acked, tso and else */
3088 if (after(scb->end_seq, tp->snd_una)) {
3089 if (tcp_skb_pcount(skb) == 1 ||
3090 !after(tp->snd_una, scb->seq))
3093 acked_pcount = tcp_tso_acked(sk, skb);
3096 fully_acked = false;
3098 acked_pcount = tcp_skb_pcount(skb);
3101 if (unlikely(sacked & TCPCB_RETRANS)) {
3102 if (sacked & TCPCB_SACKED_RETRANS)
3103 tp->retrans_out -= acked_pcount;
3104 flag |= FLAG_RETRANS_DATA_ACKED;
3105 } else if (!(sacked & TCPCB_SACKED_ACKED)) {
3106 last_ackt = skb->skb_mstamp;
3107 WARN_ON_ONCE(last_ackt == 0);
3109 first_ackt = last_ackt;
3111 last_in_flight = TCP_SKB_CB(skb)->tx.in_flight;
3112 if (before(start_seq, reord))
3114 if (!after(scb->end_seq, tp->high_seq))
3115 flag |= FLAG_ORIG_SACK_ACKED;
3118 if (sacked & TCPCB_SACKED_ACKED) {
3119 tp->sacked_out -= acked_pcount;
3120 } else if (tcp_is_sack(tp)) {
3121 tp->delivered += acked_pcount;
3122 if (!tcp_skb_spurious_retrans(tp, skb))
3123 tcp_rack_advance(tp, sacked, scb->end_seq,
3126 if (sacked & TCPCB_LOST)
3127 tp->lost_out -= acked_pcount;
3129 tp->packets_out -= acked_pcount;
3130 pkts_acked += acked_pcount;
3131 tcp_rate_skb_delivered(sk, skb, sack->rate);
3133 /* Initial outgoing SYN's get put onto the write_queue
3134 * just like anything else we transmit. It is not
3135 * true data, and if we misinform our callers that
3136 * this ACK acks real data, we will erroneously exit
3137 * connection startup slow start one packet too
3138 * quickly. This is severely frowned upon behavior.
3140 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
3141 flag |= FLAG_DATA_ACKED;
3143 flag |= FLAG_SYN_ACKED;
3144 tp->retrans_stamp = 0;
3150 next = skb_rb_next(skb);
3151 if (unlikely(skb == tp->retransmit_skb_hint))
3152 tp->retransmit_skb_hint = NULL;
3153 if (unlikely(skb == tp->lost_skb_hint))
3154 tp->lost_skb_hint = NULL;
3155 tcp_rtx_queue_unlink_and_free(skb, sk);
3159 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
3161 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3162 tp->snd_up = tp->snd_una;
3164 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3165 flag |= FLAG_SACK_RENEGING;
3167 if (likely(first_ackt) && !(flag & FLAG_RETRANS_DATA_ACKED)) {
3168 seq_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, first_ackt);
3169 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, last_ackt);
3171 if (pkts_acked == 1 && last_in_flight < tp->mss_cache &&
3172 last_in_flight && !prior_sacked && fully_acked &&
3173 sack->rate->prior_delivered + 1 == tp->delivered &&
3174 !(flag & (FLAG_CA_ALERT | FLAG_SYN_ACKED))) {
3175 /* Conservatively mark a delayed ACK. It's typically
3176 * from a lone runt packet over the round trip to
3177 * a receiver w/o out-of-order or CE events.
3179 flag |= FLAG_ACK_MAYBE_DELAYED;
3182 if (sack->first_sackt) {
3183 sack_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->first_sackt);
3184 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->last_sackt);
3186 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us,
3187 ca_rtt_us, sack->rate);
3189 if (flag & FLAG_ACKED) {
3190 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3191 if (unlikely(icsk->icsk_mtup.probe_size &&
3192 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3193 tcp_mtup_probe_success(sk);
3196 if (tcp_is_reno(tp)) {
3197 tcp_remove_reno_sacks(sk, pkts_acked);
3199 /* If any of the cumulatively ACKed segments was
3200 * retransmitted, non-SACK case cannot confirm that
3201 * progress was due to original transmission due to
3202 * lack of TCPCB_SACKED_ACKED bits even if some of
3203 * the packets may have been never retransmitted.
3205 if (flag & FLAG_RETRANS_DATA_ACKED)
3206 flag &= ~FLAG_ORIG_SACK_ACKED;
3210 /* Non-retransmitted hole got filled? That's reordering */
3211 if (before(reord, prior_fack))
3212 tcp_check_sack_reordering(sk, reord, 0);
3214 delta = prior_sacked - tp->sacked_out;
3215 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3217 } else if (skb && rtt_update && sack_rtt_us >= 0 &&
3218 sack_rtt_us > tcp_stamp_us_delta(tp->tcp_mstamp, skb->skb_mstamp)) {
3219 /* Do not re-arm RTO if the sack RTT is measured from data sent
3220 * after when the head was last (re)transmitted. Otherwise the
3221 * timeout may continue to extend in loss recovery.
3223 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3226 if (icsk->icsk_ca_ops->pkts_acked) {
3227 struct ack_sample sample = { .pkts_acked = pkts_acked,
3228 .rtt_us = sack->rate->rtt_us,
3229 .in_flight = last_in_flight };
3231 icsk->icsk_ca_ops->pkts_acked(sk, &sample);
3234 #if FASTRETRANS_DEBUG > 0
3235 WARN_ON((int)tp->sacked_out < 0);
3236 WARN_ON((int)tp->lost_out < 0);
3237 WARN_ON((int)tp->retrans_out < 0);
3238 if (!tp->packets_out && tcp_is_sack(tp)) {
3239 icsk = inet_csk(sk);
3241 pr_debug("Leak l=%u %d\n",
3242 tp->lost_out, icsk->icsk_ca_state);
3245 if (tp->sacked_out) {
3246 pr_debug("Leak s=%u %d\n",
3247 tp->sacked_out, icsk->icsk_ca_state);
3250 if (tp->retrans_out) {
3251 pr_debug("Leak r=%u %d\n",
3252 tp->retrans_out, icsk->icsk_ca_state);
3253 tp->retrans_out = 0;
3260 static void tcp_ack_probe(struct sock *sk)
3262 struct inet_connection_sock *icsk = inet_csk(sk);
3263 struct sk_buff *head = tcp_send_head(sk);
3264 const struct tcp_sock *tp = tcp_sk(sk);
3266 /* Was it a usable window open? */
3269 if (!after(TCP_SKB_CB(head)->end_seq, tcp_wnd_end(tp))) {
3270 icsk->icsk_backoff = 0;
3271 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3272 /* Socket must be waked up by subsequent tcp_data_snd_check().
3273 * This function is not for random using!
3276 unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX);
3278 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3283 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3285 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3286 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3289 /* Decide wheather to run the increase function of congestion control. */
3290 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3292 /* If reordering is high then always grow cwnd whenever data is
3293 * delivered regardless of its ordering. Otherwise stay conservative
3294 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3295 * new SACK or ECE mark may first advance cwnd here and later reduce
3296 * cwnd in tcp_fastretrans_alert() based on more states.
3298 if (tcp_sk(sk)->reordering > sock_net(sk)->ipv4.sysctl_tcp_reordering)
3299 return flag & FLAG_FORWARD_PROGRESS;
3301 return flag & FLAG_DATA_ACKED;
3304 /* The "ultimate" congestion control function that aims to replace the rigid
3305 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction).
3306 * It's called toward the end of processing an ACK with precise rate
3307 * information. All transmission or retransmission are delayed afterwards.
3309 static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked,
3310 int flag, const struct rate_sample *rs)
3312 const struct inet_connection_sock *icsk = inet_csk(sk);
3314 if (icsk->icsk_ca_ops->cong_control) {
3315 icsk->icsk_ca_ops->cong_control(sk, rs);
3319 if (tcp_in_cwnd_reduction(sk)) {
3320 /* Reduce cwnd if state mandates */
3321 tcp_cwnd_reduction(sk, acked_sacked, flag);
3322 } else if (tcp_may_raise_cwnd(sk, flag)) {
3323 /* Advance cwnd if state allows */
3324 tcp_cong_avoid(sk, ack, acked_sacked);
3326 tcp_update_pacing_rate(sk);
3329 /* Check that window update is acceptable.
3330 * The function assumes that snd_una<=ack<=snd_next.
3332 static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3333 const u32 ack, const u32 ack_seq,
3336 return after(ack, tp->snd_una) ||
3337 after(ack_seq, tp->snd_wl1) ||
3338 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3341 /* If we update tp->snd_una, also update tp->bytes_acked */
3342 static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack)
3344 u32 delta = ack - tp->snd_una;
3346 sock_owned_by_me((struct sock *)tp);
3347 tp->bytes_acked += delta;
3351 /* If we update tp->rcv_nxt, also update tp->bytes_received */
3352 static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq)
3354 u32 delta = seq - tp->rcv_nxt;
3356 sock_owned_by_me((struct sock *)tp);
3357 tp->bytes_received += delta;
3361 /* Update our send window.
3363 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3364 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3366 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3369 struct tcp_sock *tp = tcp_sk(sk);
3371 u32 nwin = ntohs(tcp_hdr(skb)->window);
3373 if (likely(!tcp_hdr(skb)->syn))
3374 nwin <<= tp->rx_opt.snd_wscale;
3376 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3377 flag |= FLAG_WIN_UPDATE;
3378 tcp_update_wl(tp, ack_seq);
3380 if (tp->snd_wnd != nwin) {
3383 /* Note, it is the only place, where
3384 * fast path is recovered for sending TCP.
3387 tcp_fast_path_check(sk);
3389 if (!tcp_write_queue_empty(sk))
3390 tcp_slow_start_after_idle_check(sk);
3392 if (nwin > tp->max_window) {
3393 tp->max_window = nwin;
3394 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3399 tcp_snd_una_update(tp, ack);
3404 static bool __tcp_oow_rate_limited(struct net *net, int mib_idx,
3405 u32 *last_oow_ack_time)
3407 if (*last_oow_ack_time) {
3408 s32 elapsed = (s32)(tcp_jiffies32 - *last_oow_ack_time);
3410 if (0 <= elapsed && elapsed < net->ipv4.sysctl_tcp_invalid_ratelimit) {
3411 NET_INC_STATS(net, mib_idx);
3412 return true; /* rate-limited: don't send yet! */
3416 *last_oow_ack_time = tcp_jiffies32;
3418 return false; /* not rate-limited: go ahead, send dupack now! */
3421 /* Return true if we're currently rate-limiting out-of-window ACKs and
3422 * thus shouldn't send a dupack right now. We rate-limit dupacks in
3423 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3424 * attacks that send repeated SYNs or ACKs for the same connection. To
3425 * do this, we do not send a duplicate SYNACK or ACK if the remote
3426 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3428 bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb,
3429 int mib_idx, u32 *last_oow_ack_time)
3431 /* Data packets without SYNs are not likely part of an ACK loop. */
3432 if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) &&
3436 return __tcp_oow_rate_limited(net, mib_idx, last_oow_ack_time);
3439 /* RFC 5961 7 [ACK Throttling] */
3440 static void tcp_send_challenge_ack(struct sock *sk, const struct sk_buff *skb)
3442 /* unprotected vars, we dont care of overwrites */
3443 static u32 challenge_timestamp;
3444 static unsigned int challenge_count;
3445 struct tcp_sock *tp = tcp_sk(sk);
3446 struct net *net = sock_net(sk);
3449 /* First check our per-socket dupack rate limit. */
3450 if (__tcp_oow_rate_limited(net,
3451 LINUX_MIB_TCPACKSKIPPEDCHALLENGE,
3452 &tp->last_oow_ack_time))
3455 /* Then check host-wide RFC 5961 rate limit. */
3457 if (now != challenge_timestamp) {
3458 u32 ack_limit = net->ipv4.sysctl_tcp_challenge_ack_limit;
3459 u32 half = (ack_limit + 1) >> 1;
3461 challenge_timestamp = now;
3462 WRITE_ONCE(challenge_count, half + prandom_u32_max(ack_limit));
3464 count = READ_ONCE(challenge_count);
3466 WRITE_ONCE(challenge_count, count - 1);
3467 NET_INC_STATS(net, LINUX_MIB_TCPCHALLENGEACK);
3472 static void tcp_store_ts_recent(struct tcp_sock *tp)
3474 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3475 tp->rx_opt.ts_recent_stamp = ktime_get_seconds();
3478 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3480 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3481 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3482 * extra check below makes sure this can only happen
3483 * for pure ACK frames. -DaveM
3485 * Not only, also it occurs for expired timestamps.
3488 if (tcp_paws_check(&tp->rx_opt, 0))
3489 tcp_store_ts_recent(tp);
3493 /* This routine deals with acks during a TLP episode.
3494 * We mark the end of a TLP episode on receiving TLP dupack or when
3495 * ack is after tlp_high_seq.
3496 * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3498 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3500 struct tcp_sock *tp = tcp_sk(sk);
3502 if (before(ack, tp->tlp_high_seq))
3505 if (flag & FLAG_DSACKING_ACK) {
3506 /* This DSACK means original and TLP probe arrived; no loss */
3507 tp->tlp_high_seq = 0;
3508 } else if (after(ack, tp->tlp_high_seq)) {
3509 /* ACK advances: there was a loss, so reduce cwnd. Reset
3510 * tlp_high_seq in tcp_init_cwnd_reduction()
3512 tcp_init_cwnd_reduction(sk);
3513 tcp_set_ca_state(sk, TCP_CA_CWR);
3514 tcp_end_cwnd_reduction(sk);
3515 tcp_try_keep_open(sk);
3516 NET_INC_STATS(sock_net(sk),
3517 LINUX_MIB_TCPLOSSPROBERECOVERY);
3518 } else if (!(flag & (FLAG_SND_UNA_ADVANCED |
3519 FLAG_NOT_DUP | FLAG_DATA_SACKED))) {
3520 /* Pure dupack: original and TLP probe arrived; no loss */
3521 tp->tlp_high_seq = 0;
3525 static inline void tcp_in_ack_event(struct sock *sk, u32 flags)
3527 const struct inet_connection_sock *icsk = inet_csk(sk);
3529 if (icsk->icsk_ca_ops->in_ack_event)
3530 icsk->icsk_ca_ops->in_ack_event(sk, flags);
3533 /* Congestion control has updated the cwnd already. So if we're in
3534 * loss recovery then now we do any new sends (for FRTO) or
3535 * retransmits (for CA_Loss or CA_recovery) that make sense.
3537 static void tcp_xmit_recovery(struct sock *sk, int rexmit)
3539 struct tcp_sock *tp = tcp_sk(sk);
3541 if (rexmit == REXMIT_NONE)
3544 if (unlikely(rexmit == 2)) {
3545 __tcp_push_pending_frames(sk, tcp_current_mss(sk),
3547 if (after(tp->snd_nxt, tp->high_seq))
3551 tcp_xmit_retransmit_queue(sk);
3554 /* Returns the number of packets newly acked or sacked by the current ACK */
3555 static u32 tcp_newly_delivered(struct sock *sk, u32 prior_delivered, int flag)
3557 const struct net *net = sock_net(sk);
3558 struct tcp_sock *tp = tcp_sk(sk);
3561 delivered = tp->delivered - prior_delivered;
3562 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVERED, delivered);
3563 if (flag & FLAG_ECE) {
3564 tp->delivered_ce += delivered;
3565 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVEREDCE, delivered);
3570 /* This routine deals with incoming acks, but not outgoing ones. */
3571 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3573 struct inet_connection_sock *icsk = inet_csk(sk);
3574 struct tcp_sock *tp = tcp_sk(sk);
3575 struct tcp_sacktag_state sack_state;
3576 struct rate_sample rs = { .prior_delivered = 0 };
3577 u32 prior_snd_una = tp->snd_una;
3578 bool is_sack_reneg = tp->is_sack_reneg;
3579 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3580 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3581 bool is_dupack = false;
3582 int prior_packets = tp->packets_out;
3583 u32 delivered = tp->delivered;
3584 u32 lost = tp->lost;
3585 int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */
3588 sack_state.first_sackt = 0;
3589 sack_state.rate = &rs;
3591 /* We very likely will need to access rtx queue. */
3592 prefetch(sk->tcp_rtx_queue.rb_node);
3594 /* If the ack is older than previous acks
3595 * then we can probably ignore it.
3597 if (before(ack, prior_snd_una)) {
3598 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3599 if (before(ack, prior_snd_una - tp->max_window)) {
3600 if (!(flag & FLAG_NO_CHALLENGE_ACK))
3601 tcp_send_challenge_ack(sk, skb);
3607 /* If the ack includes data we haven't sent yet, discard
3608 * this segment (RFC793 Section 3.9).
3610 if (after(ack, tp->snd_nxt))
3613 if (after(ack, prior_snd_una)) {
3614 flag |= FLAG_SND_UNA_ADVANCED;
3615 icsk->icsk_retransmits = 0;
3617 #if IS_ENABLED(CONFIG_TLS_DEVICE)
3618 if (static_branch_unlikely(&clean_acked_data_enabled))
3619 if (icsk->icsk_clean_acked)
3620 icsk->icsk_clean_acked(sk, ack);
3624 prior_fack = tcp_is_sack(tp) ? tcp_highest_sack_seq(tp) : tp->snd_una;
3625 rs.prior_in_flight = tcp_packets_in_flight(tp);
3627 /* ts_recent update must be made after we are sure that the packet
3630 if (flag & FLAG_UPDATE_TS_RECENT)
3631 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3633 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3634 /* Window is constant, pure forward advance.
3635 * No more checks are required.
3636 * Note, we use the fact that SND.UNA>=SND.WL2.
3638 tcp_update_wl(tp, ack_seq);
3639 tcp_snd_una_update(tp, ack);
3640 flag |= FLAG_WIN_UPDATE;
3642 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
3644 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPACKS);
3646 u32 ack_ev_flags = CA_ACK_SLOWPATH;
3648 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3651 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3653 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3655 if (TCP_SKB_CB(skb)->sacked)
3656 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3659 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) {
3661 ack_ev_flags |= CA_ACK_ECE;
3664 if (flag & FLAG_WIN_UPDATE)
3665 ack_ev_flags |= CA_ACK_WIN_UPDATE;
3667 tcp_in_ack_event(sk, ack_ev_flags);
3670 /* We passed data and got it acked, remove any soft error
3671 * log. Something worked...
3673 sk->sk_err_soft = 0;
3674 icsk->icsk_probes_out = 0;
3675 tp->rcv_tstamp = tcp_jiffies32;
3679 /* See if we can take anything off of the retransmit queue. */
3680 flag |= tcp_clean_rtx_queue(sk, prior_fack, prior_snd_una, &sack_state);
3682 tcp_rack_update_reo_wnd(sk, &rs);
3684 if (tp->tlp_high_seq)
3685 tcp_process_tlp_ack(sk, ack, flag);
3686 /* If needed, reset TLP/RTO timer; RACK may later override this. */
3687 if (flag & FLAG_SET_XMIT_TIMER)
3688 tcp_set_xmit_timer(sk);
3690 if (tcp_ack_is_dubious(sk, flag)) {
3691 is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
3692 tcp_fastretrans_alert(sk, prior_snd_una, is_dupack, &flag,
3696 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
3699 delivered = tcp_newly_delivered(sk, delivered, flag);
3700 lost = tp->lost - lost; /* freshly marked lost */
3701 rs.is_ack_delayed = !!(flag & FLAG_ACK_MAYBE_DELAYED);
3702 tcp_rate_gen(sk, delivered, lost, is_sack_reneg, sack_state.rate);
3703 tcp_cong_control(sk, ack, delivered, flag, sack_state.rate);
3704 tcp_xmit_recovery(sk, rexmit);
3708 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3709 if (flag & FLAG_DSACKING_ACK) {
3710 tcp_fastretrans_alert(sk, prior_snd_una, is_dupack, &flag,
3712 tcp_newly_delivered(sk, delivered, flag);
3714 /* If this ack opens up a zero window, clear backoff. It was
3715 * being used to time the probes, and is probably far higher than
3716 * it needs to be for normal retransmission.
3720 if (tp->tlp_high_seq)
3721 tcp_process_tlp_ack(sk, ack, flag);
3725 SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3729 /* If data was SACKed, tag it and see if we should send more data.
3730 * If data was DSACKed, see if we can undo a cwnd reduction.
3732 if (TCP_SKB_CB(skb)->sacked) {
3733 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3735 tcp_fastretrans_alert(sk, prior_snd_una, is_dupack, &flag,
3737 tcp_newly_delivered(sk, delivered, flag);
3738 tcp_xmit_recovery(sk, rexmit);
3741 SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3745 static void tcp_parse_fastopen_option(int len, const unsigned char *cookie,
3746 bool syn, struct tcp_fastopen_cookie *foc,
3749 /* Valid only in SYN or SYN-ACK with an even length. */
3750 if (!foc || !syn || len < 0 || (len & 1))
3753 if (len >= TCP_FASTOPEN_COOKIE_MIN &&
3754 len <= TCP_FASTOPEN_COOKIE_MAX)
3755 memcpy(foc->val, cookie, len);
3762 static void smc_parse_options(const struct tcphdr *th,
3763 struct tcp_options_received *opt_rx,
3764 const unsigned char *ptr,
3767 #if IS_ENABLED(CONFIG_SMC)
3768 if (static_branch_unlikely(&tcp_have_smc)) {
3769 if (th->syn && !(opsize & 1) &&
3770 opsize >= TCPOLEN_EXP_SMC_BASE &&
3771 get_unaligned_be32(ptr) == TCPOPT_SMC_MAGIC)
3777 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3778 * But, this can also be called on packets in the established flow when
3779 * the fast version below fails.
3781 void tcp_parse_options(const struct net *net,
3782 const struct sk_buff *skb,
3783 struct tcp_options_received *opt_rx, int estab,
3784 struct tcp_fastopen_cookie *foc)
3786 const unsigned char *ptr;
3787 const struct tcphdr *th = tcp_hdr(skb);
3788 int length = (th->doff * 4) - sizeof(struct tcphdr);
3790 ptr = (const unsigned char *)(th + 1);
3791 opt_rx->saw_tstamp = 0;
3793 while (length > 0) {
3794 int opcode = *ptr++;
3800 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3805 if (opsize < 2) /* "silly options" */
3807 if (opsize > length)
3808 return; /* don't parse partial options */
3811 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3812 u16 in_mss = get_unaligned_be16(ptr);
3814 if (opt_rx->user_mss &&
3815 opt_rx->user_mss < in_mss)
3816 in_mss = opt_rx->user_mss;
3817 opt_rx->mss_clamp = in_mss;
3822 if (opsize == TCPOLEN_WINDOW && th->syn &&
3823 !estab && net->ipv4.sysctl_tcp_window_scaling) {
3824 __u8 snd_wscale = *(__u8 *)ptr;
3825 opt_rx->wscale_ok = 1;
3826 if (snd_wscale > TCP_MAX_WSCALE) {
3827 net_info_ratelimited("%s: Illegal window scaling value %d > %u received\n",
3831 snd_wscale = TCP_MAX_WSCALE;
3833 opt_rx->snd_wscale = snd_wscale;
3836 case TCPOPT_TIMESTAMP:
3837 if ((opsize == TCPOLEN_TIMESTAMP) &&
3838 ((estab && opt_rx->tstamp_ok) ||
3839 (!estab && net->ipv4.sysctl_tcp_timestamps))) {
3840 opt_rx->saw_tstamp = 1;
3841 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3842 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3845 case TCPOPT_SACK_PERM:
3846 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3847 !estab && net->ipv4.sysctl_tcp_sack) {
3848 opt_rx->sack_ok = TCP_SACK_SEEN;
3849 tcp_sack_reset(opt_rx);
3854 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3855 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3857 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3860 #ifdef CONFIG_TCP_MD5SIG
3863 * The MD5 Hash has already been
3864 * checked (see tcp_v{4,6}_do_rcv()).
3868 case TCPOPT_FASTOPEN:
3869 tcp_parse_fastopen_option(
3870 opsize - TCPOLEN_FASTOPEN_BASE,
3871 ptr, th->syn, foc, false);
3875 /* Fast Open option shares code 254 using a
3876 * 16 bits magic number.
3878 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE &&
3879 get_unaligned_be16(ptr) ==
3880 TCPOPT_FASTOPEN_MAGIC)
3881 tcp_parse_fastopen_option(opsize -
3882 TCPOLEN_EXP_FASTOPEN_BASE,
3883 ptr + 2, th->syn, foc, true);
3885 smc_parse_options(th, opt_rx, ptr,
3895 EXPORT_SYMBOL(tcp_parse_options);
3897 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
3899 const __be32 *ptr = (const __be32 *)(th + 1);
3901 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3902 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3903 tp->rx_opt.saw_tstamp = 1;
3905 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3908 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
3910 tp->rx_opt.rcv_tsecr = 0;
3916 /* Fast parse options. This hopes to only see timestamps.
3917 * If it is wrong it falls back on tcp_parse_options().
3919 static bool tcp_fast_parse_options(const struct net *net,
3920 const struct sk_buff *skb,
3921 const struct tcphdr *th, struct tcp_sock *tp)
3923 /* In the spirit of fast parsing, compare doff directly to constant
3924 * values. Because equality is used, short doff can be ignored here.
3926 if (th->doff == (sizeof(*th) / 4)) {
3927 tp->rx_opt.saw_tstamp = 0;
3929 } else if (tp->rx_opt.tstamp_ok &&
3930 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3931 if (tcp_parse_aligned_timestamp(tp, th))
3935 tcp_parse_options(net, skb, &tp->rx_opt, 1, NULL);
3936 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3937 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
3942 #ifdef CONFIG_TCP_MD5SIG
3944 * Parse MD5 Signature option
3946 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
3948 int length = (th->doff << 2) - sizeof(*th);
3949 const u8 *ptr = (const u8 *)(th + 1);
3951 /* If not enough data remaining, we can short cut */
3952 while (length >= TCPOLEN_MD5SIG) {
3953 int opcode = *ptr++;
3964 if (opsize < 2 || opsize > length)
3966 if (opcode == TCPOPT_MD5SIG)
3967 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
3974 EXPORT_SYMBOL(tcp_parse_md5sig_option);
3977 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3979 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3980 * it can pass through stack. So, the following predicate verifies that
3981 * this segment is not used for anything but congestion avoidance or
3982 * fast retransmit. Moreover, we even are able to eliminate most of such
3983 * second order effects, if we apply some small "replay" window (~RTO)
3984 * to timestamp space.
3986 * All these measures still do not guarantee that we reject wrapped ACKs
3987 * on networks with high bandwidth, when sequence space is recycled fastly,
3988 * but it guarantees that such events will be very rare and do not affect
3989 * connection seriously. This doesn't look nice, but alas, PAWS is really
3992 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3993 * states that events when retransmit arrives after original data are rare.
3994 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3995 * the biggest problem on large power networks even with minor reordering.
3996 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3997 * up to bandwidth of 18Gigabit/sec. 8) ]
4000 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
4002 const struct tcp_sock *tp = tcp_sk(sk);
4003 const struct tcphdr *th = tcp_hdr(skb);
4004 u32 seq = TCP_SKB_CB(skb)->seq;
4005 u32 ack = TCP_SKB_CB(skb)->ack_seq;
4007 return (/* 1. Pure ACK with correct sequence number. */
4008 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
4010 /* 2. ... and duplicate ACK. */
4011 ack == tp->snd_una &&
4013 /* 3. ... and does not update window. */
4014 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
4016 /* 4. ... and sits in replay window. */
4017 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
4020 static inline bool tcp_paws_discard(const struct sock *sk,
4021 const struct sk_buff *skb)
4023 const struct tcp_sock *tp = tcp_sk(sk);
4025 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
4026 !tcp_disordered_ack(sk, skb);
4029 /* Check segment sequence number for validity.
4031 * Segment controls are considered valid, if the segment
4032 * fits to the window after truncation to the window. Acceptability
4033 * of data (and SYN, FIN, of course) is checked separately.
4034 * See tcp_data_queue(), for example.
4036 * Also, controls (RST is main one) are accepted using RCV.WUP instead
4037 * of RCV.NXT. Peer still did not advance his SND.UNA when we
4038 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4039 * (borrowed from freebsd)
4042 static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
4044 return !before(end_seq, tp->rcv_wup) &&
4045 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
4048 /* When we get a reset we do this. */
4049 void tcp_reset(struct sock *sk)
4051 trace_tcp_receive_reset(sk);
4053 /* We want the right error as BSD sees it (and indeed as we do). */
4054 switch (sk->sk_state) {
4056 sk->sk_err = ECONNREFUSED;
4058 case TCP_CLOSE_WAIT:
4064 sk->sk_err = ECONNRESET;
4066 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4069 tcp_write_queue_purge(sk);
4072 if (!sock_flag(sk, SOCK_DEAD))
4073 sk->sk_error_report(sk);
4077 * Process the FIN bit. This now behaves as it is supposed to work
4078 * and the FIN takes effect when it is validly part of sequence
4079 * space. Not before when we get holes.
4081 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4082 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4085 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4086 * close and we go into CLOSING (and later onto TIME-WAIT)
4088 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4090 void tcp_fin(struct sock *sk)
4092 struct tcp_sock *tp = tcp_sk(sk);
4094 inet_csk_schedule_ack(sk);
4096 sk->sk_shutdown |= RCV_SHUTDOWN;
4097 sock_set_flag(sk, SOCK_DONE);
4099 switch (sk->sk_state) {
4101 case TCP_ESTABLISHED:
4102 /* Move to CLOSE_WAIT */
4103 tcp_set_state(sk, TCP_CLOSE_WAIT);
4104 inet_csk(sk)->icsk_ack.pingpong = 1;
4107 case TCP_CLOSE_WAIT:
4109 /* Received a retransmission of the FIN, do
4114 /* RFC793: Remain in the LAST-ACK state. */
4118 /* This case occurs when a simultaneous close
4119 * happens, we must ack the received FIN and
4120 * enter the CLOSING state.
4123 tcp_set_state(sk, TCP_CLOSING);
4126 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4128 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4131 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4132 * cases we should never reach this piece of code.
4134 pr_err("%s: Impossible, sk->sk_state=%d\n",
4135 __func__, sk->sk_state);
4139 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4140 * Probably, we should reset in this case. For now drop them.
4142 skb_rbtree_purge(&tp->out_of_order_queue);
4143 if (tcp_is_sack(tp))
4144 tcp_sack_reset(&tp->rx_opt);
4147 if (!sock_flag(sk, SOCK_DEAD)) {
4148 sk->sk_state_change(sk);
4150 /* Do not send POLL_HUP for half duplex close. */
4151 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4152 sk->sk_state == TCP_CLOSE)
4153 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4155 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4159 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4162 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4163 if (before(seq, sp->start_seq))
4164 sp->start_seq = seq;
4165 if (after(end_seq, sp->end_seq))
4166 sp->end_seq = end_seq;
4172 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4174 struct tcp_sock *tp = tcp_sk(sk);
4176 if (tcp_is_sack(tp) && sock_net(sk)->ipv4.sysctl_tcp_dsack) {
4179 if (before(seq, tp->rcv_nxt))
4180 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4182 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4184 NET_INC_STATS(sock_net(sk), mib_idx);
4186 tp->rx_opt.dsack = 1;
4187 tp->duplicate_sack[0].start_seq = seq;
4188 tp->duplicate_sack[0].end_seq = end_seq;
4192 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4194 struct tcp_sock *tp = tcp_sk(sk);
4196 if (!tp->rx_opt.dsack)
4197 tcp_dsack_set(sk, seq, end_seq);
4199 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4202 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4204 struct tcp_sock *tp = tcp_sk(sk);
4206 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4207 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4208 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4209 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
4211 if (tcp_is_sack(tp) && sock_net(sk)->ipv4.sysctl_tcp_dsack) {
4212 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4214 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4215 end_seq = tp->rcv_nxt;
4216 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4223 /* These routines update the SACK block as out-of-order packets arrive or
4224 * in-order packets close up the sequence space.
4226 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4229 struct tcp_sack_block *sp = &tp->selective_acks[0];
4230 struct tcp_sack_block *swalk = sp + 1;
4232 /* See if the recent change to the first SACK eats into
4233 * or hits the sequence space of other SACK blocks, if so coalesce.
4235 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4236 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4239 /* Zap SWALK, by moving every further SACK up by one slot.
4240 * Decrease num_sacks.
4242 tp->rx_opt.num_sacks--;
4243 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4247 this_sack++, swalk++;
4251 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4253 struct tcp_sock *tp = tcp_sk(sk);
4254 struct tcp_sack_block *sp = &tp->selective_acks[0];
4255 int cur_sacks = tp->rx_opt.num_sacks;
4261 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4262 if (tcp_sack_extend(sp, seq, end_seq)) {
4263 /* Rotate this_sack to the first one. */
4264 for (; this_sack > 0; this_sack--, sp--)
4265 swap(*sp, *(sp - 1));
4267 tcp_sack_maybe_coalesce(tp);
4272 /* Could not find an adjacent existing SACK, build a new one,
4273 * put it at the front, and shift everyone else down. We
4274 * always know there is at least one SACK present already here.
4276 * If the sack array is full, forget about the last one.
4278 if (this_sack >= TCP_NUM_SACKS) {
4279 if (tp->compressed_ack)
4282 tp->rx_opt.num_sacks--;
4285 for (; this_sack > 0; this_sack--, sp--)
4289 /* Build the new head SACK, and we're done. */
4290 sp->start_seq = seq;
4291 sp->end_seq = end_seq;
4292 tp->rx_opt.num_sacks++;
4295 /* RCV.NXT advances, some SACKs should be eaten. */
4297 static void tcp_sack_remove(struct tcp_sock *tp)
4299 struct tcp_sack_block *sp = &tp->selective_acks[0];
4300 int num_sacks = tp->rx_opt.num_sacks;
4303 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4304 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4305 tp->rx_opt.num_sacks = 0;
4309 for (this_sack = 0; this_sack < num_sacks;) {
4310 /* Check if the start of the sack is covered by RCV.NXT. */
4311 if (!before(tp->rcv_nxt, sp->start_seq)) {
4314 /* RCV.NXT must cover all the block! */
4315 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4317 /* Zap this SACK, by moving forward any other SACKS. */
4318 for (i = this_sack+1; i < num_sacks; i++)
4319 tp->selective_acks[i-1] = tp->selective_acks[i];
4326 tp->rx_opt.num_sacks = num_sacks;
4330 * tcp_try_coalesce - try to merge skb to prior one
4332 * @dest: destination queue
4334 * @from: buffer to add in queue
4335 * @fragstolen: pointer to boolean
4337 * Before queueing skb @from after @to, try to merge them
4338 * to reduce overall memory use and queue lengths, if cost is small.
4339 * Packets in ofo or receive queues can stay a long time.
4340 * Better try to coalesce them right now to avoid future collapses.
4341 * Returns true if caller should free @from instead of queueing it
4343 static bool tcp_try_coalesce(struct sock *sk,
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 #ifdef CONFIG_TLS_DEVICE
4357 if (from->decrypted != to->decrypted)
4361 if (!skb_try_coalesce(to, from, fragstolen, &delta))
4364 atomic_add(delta, &sk->sk_rmem_alloc);
4365 sk_mem_charge(sk, delta);
4366 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4367 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4368 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4369 TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
4371 if (TCP_SKB_CB(from)->has_rxtstamp) {
4372 TCP_SKB_CB(to)->has_rxtstamp = true;
4373 to->tstamp = from->tstamp;
4379 static bool tcp_ooo_try_coalesce(struct sock *sk,
4381 struct sk_buff *from,
4384 bool res = tcp_try_coalesce(sk, to, from, fragstolen);
4386 /* In case tcp_drop() is called later, update to->gso_segs */
4388 u32 gso_segs = max_t(u16, 1, skb_shinfo(to)->gso_segs) +
4389 max_t(u16, 1, skb_shinfo(from)->gso_segs);
4391 skb_shinfo(to)->gso_segs = min_t(u32, gso_segs, 0xFFFF);
4396 static void tcp_drop(struct sock *sk, struct sk_buff *skb)
4398 sk_drops_add(sk, skb);
4402 /* This one checks to see if we can put data from the
4403 * out_of_order queue into the receive_queue.
4405 static void tcp_ofo_queue(struct sock *sk)
4407 struct tcp_sock *tp = tcp_sk(sk);
4408 __u32 dsack_high = tp->rcv_nxt;
4409 bool fin, fragstolen, eaten;
4410 struct sk_buff *skb, *tail;
4413 p = rb_first(&tp->out_of_order_queue);
4416 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4419 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4420 __u32 dsack = dsack_high;
4421 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4422 dsack_high = TCP_SKB_CB(skb)->end_seq;
4423 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4426 rb_erase(&skb->rbnode, &tp->out_of_order_queue);
4428 if (unlikely(!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))) {
4429 SOCK_DEBUG(sk, "ofo packet was already received\n");
4433 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4434 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4435 TCP_SKB_CB(skb)->end_seq);
4437 tail = skb_peek_tail(&sk->sk_receive_queue);
4438 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen);
4439 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4440 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
4442 __skb_queue_tail(&sk->sk_receive_queue, skb);
4444 kfree_skb_partial(skb, fragstolen);
4446 if (unlikely(fin)) {
4448 /* tcp_fin() purges tp->out_of_order_queue,
4449 * so we must end this loop right now.
4456 static bool tcp_prune_ofo_queue(struct sock *sk);
4457 static int tcp_prune_queue(struct sock *sk);
4459 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4462 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4463 !sk_rmem_schedule(sk, skb, size)) {
4465 if (tcp_prune_queue(sk) < 0)
4468 while (!sk_rmem_schedule(sk, skb, size)) {
4469 if (!tcp_prune_ofo_queue(sk))
4476 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4478 struct tcp_sock *tp = tcp_sk(sk);
4479 struct rb_node **p, *parent;
4480 struct sk_buff *skb1;
4484 tcp_ecn_check_ce(sk, skb);
4486 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4487 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFODROP);
4492 /* Disable header prediction. */
4494 inet_csk_schedule_ack(sk);
4496 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4497 seq = TCP_SKB_CB(skb)->seq;
4498 end_seq = TCP_SKB_CB(skb)->end_seq;
4499 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4500 tp->rcv_nxt, seq, end_seq);
4502 p = &tp->out_of_order_queue.rb_node;
4503 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4504 /* Initial out of order segment, build 1 SACK. */
4505 if (tcp_is_sack(tp)) {
4506 tp->rx_opt.num_sacks = 1;
4507 tp->selective_acks[0].start_seq = seq;
4508 tp->selective_acks[0].end_seq = end_seq;
4510 rb_link_node(&skb->rbnode, NULL, p);
4511 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4512 tp->ooo_last_skb = skb;
4516 /* In the typical case, we are adding an skb to the end of the list.
4517 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
4519 if (tcp_ooo_try_coalesce(sk, tp->ooo_last_skb,
4520 skb, &fragstolen)) {
4522 tcp_grow_window(sk, skb);
4523 kfree_skb_partial(skb, fragstolen);
4527 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
4528 if (!before(seq, TCP_SKB_CB(tp->ooo_last_skb)->end_seq)) {
4529 parent = &tp->ooo_last_skb->rbnode;
4530 p = &parent->rb_right;
4534 /* Find place to insert this segment. Handle overlaps on the way. */
4538 skb1 = rb_to_skb(parent);
4539 if (before(seq, TCP_SKB_CB(skb1)->seq)) {
4540 p = &parent->rb_left;
4543 if (before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4544 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4545 /* All the bits are present. Drop. */
4546 NET_INC_STATS(sock_net(sk),
4547 LINUX_MIB_TCPOFOMERGE);
4550 tcp_dsack_set(sk, seq, end_seq);
4553 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4554 /* Partial overlap. */
4555 tcp_dsack_set(sk, seq, TCP_SKB_CB(skb1)->end_seq);
4557 /* skb's seq == skb1's seq and skb covers skb1.
4558 * Replace skb1 with skb.
4560 rb_replace_node(&skb1->rbnode, &skb->rbnode,
4561 &tp->out_of_order_queue);
4562 tcp_dsack_extend(sk,
4563 TCP_SKB_CB(skb1)->seq,
4564 TCP_SKB_CB(skb1)->end_seq);
4565 NET_INC_STATS(sock_net(sk),
4566 LINUX_MIB_TCPOFOMERGE);
4570 } else if (tcp_ooo_try_coalesce(sk, skb1,
4571 skb, &fragstolen)) {
4574 p = &parent->rb_right;
4577 /* Insert segment into RB tree. */
4578 rb_link_node(&skb->rbnode, parent, p);
4579 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4582 /* Remove other segments covered by skb. */
4583 while ((skb1 = skb_rb_next(skb)) != NULL) {
4584 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4586 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4587 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4591 rb_erase(&skb1->rbnode, &tp->out_of_order_queue);
4592 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4593 TCP_SKB_CB(skb1)->end_seq);
4594 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4597 /* If there is no skb after us, we are the last_skb ! */
4599 tp->ooo_last_skb = skb;
4602 if (tcp_is_sack(tp))
4603 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4606 tcp_grow_window(sk, skb);
4608 skb_set_owner_r(skb, sk);
4612 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, int hdrlen,
4616 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4618 __skb_pull(skb, hdrlen);
4620 tcp_try_coalesce(sk, tail,
4621 skb, fragstolen)) ? 1 : 0;
4622 tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq);
4624 __skb_queue_tail(&sk->sk_receive_queue, skb);
4625 skb_set_owner_r(skb, sk);
4630 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4632 struct sk_buff *skb;
4640 if (size > PAGE_SIZE) {
4641 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS);
4643 data_len = npages << PAGE_SHIFT;
4644 size = data_len + (size & ~PAGE_MASK);
4646 skb = alloc_skb_with_frags(size - data_len, data_len,
4647 PAGE_ALLOC_COSTLY_ORDER,
4648 &err, sk->sk_allocation);
4652 skb_put(skb, size - data_len);
4653 skb->data_len = data_len;
4656 if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) {
4657 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP);
4661 err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size);
4665 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4666 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4667 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4669 if (tcp_queue_rcv(sk, skb, 0, &fragstolen)) {
4670 WARN_ON_ONCE(fragstolen); /* should not happen */
4682 void tcp_data_ready(struct sock *sk)
4684 const struct tcp_sock *tp = tcp_sk(sk);
4685 int avail = tp->rcv_nxt - tp->copied_seq;
4687 if (avail < sk->sk_rcvlowat && !sock_flag(sk, SOCK_DONE))
4690 sk->sk_data_ready(sk);
4693 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4695 struct tcp_sock *tp = tcp_sk(sk);
4699 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
4704 __skb_pull(skb, tcp_hdr(skb)->doff * 4);
4706 tcp_ecn_accept_cwr(sk, skb);
4708 tp->rx_opt.dsack = 0;
4710 /* Queue data for delivery to the user.
4711 * Packets in sequence go to the receive queue.
4712 * Out of sequence packets to the out_of_order_queue.
4714 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4715 if (tcp_receive_window(tp) == 0) {
4716 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP);
4720 /* Ok. In sequence. In window. */
4722 if (skb_queue_len(&sk->sk_receive_queue) == 0)
4723 sk_forced_mem_schedule(sk, skb->truesize);
4724 else if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) {
4725 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP);
4729 eaten = tcp_queue_rcv(sk, skb, 0, &fragstolen);
4731 tcp_event_data_recv(sk, skb);
4732 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
4735 if (!RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4738 /* RFC5681. 4.2. SHOULD send immediate ACK, when
4739 * gap in queue is filled.
4741 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
4742 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW;
4745 if (tp->rx_opt.num_sacks)
4746 tcp_sack_remove(tp);
4748 tcp_fast_path_check(sk);
4751 kfree_skb_partial(skb, fragstolen);
4752 if (!sock_flag(sk, SOCK_DEAD))
4757 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4758 /* A retransmit, 2nd most common case. Force an immediate ack. */
4759 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4760 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4763 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
4764 inet_csk_schedule_ack(sk);
4770 /* Out of window. F.e. zero window probe. */
4771 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4774 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4775 /* Partial packet, seq < rcv_next < end_seq */
4776 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4777 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4778 TCP_SKB_CB(skb)->end_seq);
4780 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4782 /* If window is closed, drop tail of packet. But after
4783 * remembering D-SACK for its head made in previous line.
4785 if (!tcp_receive_window(tp)) {
4786 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP);
4792 tcp_data_queue_ofo(sk, skb);
4795 static struct sk_buff *tcp_skb_next(struct sk_buff *skb, struct sk_buff_head *list)
4798 return !skb_queue_is_last(list, skb) ? skb->next : NULL;
4800 return skb_rb_next(skb);
4803 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4804 struct sk_buff_head *list,
4805 struct rb_root *root)
4807 struct sk_buff *next = tcp_skb_next(skb, list);
4810 __skb_unlink(skb, list);
4812 rb_erase(&skb->rbnode, root);
4815 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4820 /* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */
4821 void tcp_rbtree_insert(struct rb_root *root, struct sk_buff *skb)
4823 struct rb_node **p = &root->rb_node;
4824 struct rb_node *parent = NULL;
4825 struct sk_buff *skb1;
4829 skb1 = rb_to_skb(parent);
4830 if (before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb1)->seq))
4831 p = &parent->rb_left;
4833 p = &parent->rb_right;
4835 rb_link_node(&skb->rbnode, parent, p);
4836 rb_insert_color(&skb->rbnode, root);
4839 /* Collapse contiguous sequence of skbs head..tail with
4840 * sequence numbers start..end.
4842 * If tail is NULL, this means until the end of the queue.
4844 * Segments with FIN/SYN are not collapsed (only because this
4848 tcp_collapse(struct sock *sk, struct sk_buff_head *list, struct rb_root *root,
4849 struct sk_buff *head, struct sk_buff *tail, u32 start, u32 end)
4851 struct sk_buff *skb = head, *n;
4852 struct sk_buff_head tmp;
4855 /* First, check that queue is collapsible and find
4856 * the point where collapsing can be useful.
4859 for (end_of_skbs = true; skb != NULL && skb != tail; skb = n) {
4860 n = tcp_skb_next(skb, list);
4862 /* No new bits? It is possible on ofo queue. */
4863 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4864 skb = tcp_collapse_one(sk, skb, list, root);
4870 /* The first skb to collapse is:
4872 * - bloated or contains data before "start" or
4873 * overlaps to the next one.
4875 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
4876 (tcp_win_from_space(sk, skb->truesize) > skb->len ||
4877 before(TCP_SKB_CB(skb)->seq, start))) {
4878 end_of_skbs = false;
4882 if (n && n != tail &&
4883 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(n)->seq) {
4884 end_of_skbs = false;
4888 /* Decided to skip this, advance start seq. */
4889 start = TCP_SKB_CB(skb)->end_seq;
4892 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4895 __skb_queue_head_init(&tmp);
4897 while (before(start, end)) {
4898 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
4899 struct sk_buff *nskb;
4901 nskb = alloc_skb(copy, GFP_ATOMIC);
4905 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4906 #ifdef CONFIG_TLS_DEVICE
4907 nskb->decrypted = skb->decrypted;
4909 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4911 __skb_queue_before(list, skb, nskb);
4913 __skb_queue_tail(&tmp, nskb); /* defer rbtree insertion */
4914 skb_set_owner_r(nskb, sk);
4916 /* Copy data, releasing collapsed skbs. */
4918 int offset = start - TCP_SKB_CB(skb)->seq;
4919 int size = TCP_SKB_CB(skb)->end_seq - start;
4923 size = min(copy, size);
4924 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4926 TCP_SKB_CB(nskb)->end_seq += size;
4930 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4931 skb = tcp_collapse_one(sk, skb, list, root);
4934 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4936 #ifdef CONFIG_TLS_DEVICE
4937 if (skb->decrypted != nskb->decrypted)
4944 skb_queue_walk_safe(&tmp, skb, n)
4945 tcp_rbtree_insert(root, skb);
4948 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4949 * and tcp_collapse() them until all the queue is collapsed.
4951 static void tcp_collapse_ofo_queue(struct sock *sk)
4953 struct tcp_sock *tp = tcp_sk(sk);
4954 u32 range_truesize, sum_tiny = 0;
4955 struct sk_buff *skb, *head;
4958 skb = skb_rb_first(&tp->out_of_order_queue);
4961 tp->ooo_last_skb = skb_rb_last(&tp->out_of_order_queue);
4964 start = TCP_SKB_CB(skb)->seq;
4965 end = TCP_SKB_CB(skb)->end_seq;
4966 range_truesize = skb->truesize;
4968 for (head = skb;;) {
4969 skb = skb_rb_next(skb);
4971 /* Range is terminated when we see a gap or when
4972 * we are at the queue end.
4975 after(TCP_SKB_CB(skb)->seq, end) ||
4976 before(TCP_SKB_CB(skb)->end_seq, start)) {
4977 /* Do not attempt collapsing tiny skbs */
4978 if (range_truesize != head->truesize ||
4979 end - start >= SKB_WITH_OVERHEAD(SK_MEM_QUANTUM)) {
4980 tcp_collapse(sk, NULL, &tp->out_of_order_queue,
4981 head, skb, start, end);
4983 sum_tiny += range_truesize;
4984 if (sum_tiny > sk->sk_rcvbuf >> 3)
4990 range_truesize += skb->truesize;
4991 if (unlikely(before(TCP_SKB_CB(skb)->seq, start)))
4992 start = TCP_SKB_CB(skb)->seq;
4993 if (after(TCP_SKB_CB(skb)->end_seq, end))
4994 end = TCP_SKB_CB(skb)->end_seq;
4999 * Clean the out-of-order queue to make room.
5000 * We drop high sequences packets to :
5001 * 1) Let a chance for holes to be filled.
5002 * 2) not add too big latencies if thousands of packets sit there.
5003 * (But if application shrinks SO_RCVBUF, we could still end up
5004 * freeing whole queue here)
5005 * 3) Drop at least 12.5 % of sk_rcvbuf to avoid malicious attacks.
5007 * Return true if queue has shrunk.
5009 static bool tcp_prune_ofo_queue(struct sock *sk)
5011 struct tcp_sock *tp = tcp_sk(sk);
5012 struct rb_node *node, *prev;
5015 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
5018 NET_INC_STATS(sock_net(sk), LINUX_MIB_OFOPRUNED);
5019 goal = sk->sk_rcvbuf >> 3;
5020 node = &tp->ooo_last_skb->rbnode;
5022 prev = rb_prev(node);
5023 rb_erase(node, &tp->out_of_order_queue);
5024 goal -= rb_to_skb(node)->truesize;
5025 tcp_drop(sk, rb_to_skb(node));
5026 if (!prev || goal <= 0) {
5028 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf &&
5029 !tcp_under_memory_pressure(sk))
5031 goal = sk->sk_rcvbuf >> 3;
5035 tp->ooo_last_skb = rb_to_skb(prev);
5037 /* Reset SACK state. A conforming SACK implementation will
5038 * do the same at a timeout based retransmit. When a connection
5039 * is in a sad state like this, we care only about integrity
5040 * of the connection not performance.
5042 if (tp->rx_opt.sack_ok)
5043 tcp_sack_reset(&tp->rx_opt);
5047 /* Reduce allocated memory if we can, trying to get
5048 * the socket within its memory limits again.
5050 * Return less than zero if we should start dropping frames
5051 * until the socket owning process reads some of the data
5052 * to stabilize the situation.
5054 static int tcp_prune_queue(struct sock *sk)
5056 struct tcp_sock *tp = tcp_sk(sk);
5058 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
5060 NET_INC_STATS(sock_net(sk), LINUX_MIB_PRUNECALLED);
5062 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
5063 tcp_clamp_window(sk);
5064 else if (tcp_under_memory_pressure(sk))
5065 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
5067 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5070 tcp_collapse_ofo_queue(sk);
5071 if (!skb_queue_empty(&sk->sk_receive_queue))
5072 tcp_collapse(sk, &sk->sk_receive_queue, NULL,
5073 skb_peek(&sk->sk_receive_queue),
5075 tp->copied_seq, tp->rcv_nxt);
5078 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5081 /* Collapsing did not help, destructive actions follow.
5082 * This must not ever occur. */
5084 tcp_prune_ofo_queue(sk);
5086 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5089 /* If we are really being abused, tell the caller to silently
5090 * drop receive data on the floor. It will get retransmitted
5091 * and hopefully then we'll have sufficient space.
5093 NET_INC_STATS(sock_net(sk), LINUX_MIB_RCVPRUNED);
5095 /* Massive buffer overcommit. */
5100 static bool tcp_should_expand_sndbuf(const struct sock *sk)
5102 const struct tcp_sock *tp = tcp_sk(sk);
5104 /* If the user specified a specific send buffer setting, do
5107 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
5110 /* If we are under global TCP memory pressure, do not expand. */
5111 if (tcp_under_memory_pressure(sk))
5114 /* If we are under soft global TCP memory pressure, do not expand. */
5115 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
5118 /* If we filled the congestion window, do not expand. */
5119 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
5125 /* When incoming ACK allowed to free some skb from write_queue,
5126 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
5127 * on the exit from tcp input handler.
5129 * PROBLEM: sndbuf expansion does not work well with largesend.
5131 static void tcp_new_space(struct sock *sk)
5133 struct tcp_sock *tp = tcp_sk(sk);
5135 if (tcp_should_expand_sndbuf(sk)) {
5136 tcp_sndbuf_expand(sk);
5137 tp->snd_cwnd_stamp = tcp_jiffies32;
5140 sk->sk_write_space(sk);
5143 static void tcp_check_space(struct sock *sk)
5145 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
5146 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
5147 /* pairs with tcp_poll() */
5149 if (sk->sk_socket &&
5150 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
5152 if (!test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
5153 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
5158 static inline void tcp_data_snd_check(struct sock *sk)
5160 tcp_push_pending_frames(sk);
5161 tcp_check_space(sk);
5165 * Check if sending an ack is needed.
5167 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
5169 struct tcp_sock *tp = tcp_sk(sk);
5170 unsigned long rtt, delay;
5172 /* More than one full frame received... */
5173 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
5174 /* ... and right edge of window advances far enough.
5175 * (tcp_recvmsg() will send ACK otherwise).
5176 * If application uses SO_RCVLOWAT, we want send ack now if
5177 * we have not received enough bytes to satisfy the condition.
5179 (tp->rcv_nxt - tp->copied_seq < sk->sk_rcvlowat ||
5180 __tcp_select_window(sk) >= tp->rcv_wnd)) ||
5181 /* We ACK each frame or... */
5182 tcp_in_quickack_mode(sk) ||
5183 /* Protocol state mandates a one-time immediate ACK */
5184 inet_csk(sk)->icsk_ack.pending & ICSK_ACK_NOW) {
5190 if (!ofo_possible || RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
5191 tcp_send_delayed_ack(sk);
5195 if (!tcp_is_sack(tp) ||
5196 tp->compressed_ack >= sock_net(sk)->ipv4.sysctl_tcp_comp_sack_nr)
5198 tp->compressed_ack++;
5200 if (hrtimer_is_queued(&tp->compressed_ack_timer))
5203 /* compress ack timer : 5 % of rtt, but no more than tcp_comp_sack_delay_ns */
5205 rtt = tp->rcv_rtt_est.rtt_us;
5206 if (tp->srtt_us && tp->srtt_us < rtt)
5209 delay = min_t(unsigned long, sock_net(sk)->ipv4.sysctl_tcp_comp_sack_delay_ns,
5210 rtt * (NSEC_PER_USEC >> 3)/20);
5212 hrtimer_start(&tp->compressed_ack_timer, ns_to_ktime(delay),
5213 HRTIMER_MODE_REL_PINNED_SOFT);
5216 static inline void tcp_ack_snd_check(struct sock *sk)
5218 if (!inet_csk_ack_scheduled(sk)) {
5219 /* We sent a data segment already. */
5222 __tcp_ack_snd_check(sk, 1);
5226 * This routine is only called when we have urgent data
5227 * signaled. Its the 'slow' part of tcp_urg. It could be
5228 * moved inline now as tcp_urg is only called from one
5229 * place. We handle URGent data wrong. We have to - as
5230 * BSD still doesn't use the correction from RFC961.
5231 * For 1003.1g we should support a new option TCP_STDURG to permit
5232 * either form (or just set the sysctl tcp_stdurg).
5235 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
5237 struct tcp_sock *tp = tcp_sk(sk);
5238 u32 ptr = ntohs(th->urg_ptr);
5240 if (ptr && !sock_net(sk)->ipv4.sysctl_tcp_stdurg)
5242 ptr += ntohl(th->seq);
5244 /* Ignore urgent data that we've already seen and read. */
5245 if (after(tp->copied_seq, ptr))
5248 /* Do not replay urg ptr.
5250 * NOTE: interesting situation not covered by specs.
5251 * Misbehaving sender may send urg ptr, pointing to segment,
5252 * which we already have in ofo queue. We are not able to fetch
5253 * such data and will stay in TCP_URG_NOTYET until will be eaten
5254 * by recvmsg(). Seems, we are not obliged to handle such wicked
5255 * situations. But it is worth to think about possibility of some
5256 * DoSes using some hypothetical application level deadlock.
5258 if (before(ptr, tp->rcv_nxt))
5261 /* Do we already have a newer (or duplicate) urgent pointer? */
5262 if (tp->urg_data && !after(ptr, tp->urg_seq))
5265 /* Tell the world about our new urgent pointer. */
5268 /* We may be adding urgent data when the last byte read was
5269 * urgent. To do this requires some care. We cannot just ignore
5270 * tp->copied_seq since we would read the last urgent byte again
5271 * as data, nor can we alter copied_seq until this data arrives
5272 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5274 * NOTE. Double Dutch. Rendering to plain English: author of comment
5275 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5276 * and expect that both A and B disappear from stream. This is _wrong_.
5277 * Though this happens in BSD with high probability, this is occasional.
5278 * Any application relying on this is buggy. Note also, that fix "works"
5279 * only in this artificial test. Insert some normal data between A and B and we will
5280 * decline of BSD again. Verdict: it is better to remove to trap
5283 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5284 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5285 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5287 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5288 __skb_unlink(skb, &sk->sk_receive_queue);
5293 tp->urg_data = TCP_URG_NOTYET;
5296 /* Disable header prediction. */
5300 /* This is the 'fast' part of urgent handling. */
5301 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
5303 struct tcp_sock *tp = tcp_sk(sk);
5305 /* Check if we get a new urgent pointer - normally not. */
5307 tcp_check_urg(sk, th);
5309 /* Do we wait for any urgent data? - normally not... */
5310 if (tp->urg_data == TCP_URG_NOTYET) {
5311 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5314 /* Is the urgent pointer pointing into this packet? */
5315 if (ptr < skb->len) {
5317 if (skb_copy_bits(skb, ptr, &tmp, 1))
5319 tp->urg_data = TCP_URG_VALID | tmp;
5320 if (!sock_flag(sk, SOCK_DEAD))
5321 sk->sk_data_ready(sk);
5326 /* Accept RST for rcv_nxt - 1 after a FIN.
5327 * When tcp connections are abruptly terminated from Mac OSX (via ^C), a
5328 * FIN is sent followed by a RST packet. The RST is sent with the same
5329 * sequence number as the FIN, and thus according to RFC 5961 a challenge
5330 * ACK should be sent. However, Mac OSX rate limits replies to challenge
5331 * ACKs on the closed socket. In addition middleboxes can drop either the
5332 * challenge ACK or a subsequent RST.
5334 static bool tcp_reset_check(const struct sock *sk, const struct sk_buff *skb)
5336 struct tcp_sock *tp = tcp_sk(sk);
5338 return unlikely(TCP_SKB_CB(skb)->seq == (tp->rcv_nxt - 1) &&
5339 (1 << sk->sk_state) & (TCPF_CLOSE_WAIT | TCPF_LAST_ACK |
5343 /* Does PAWS and seqno based validation of an incoming segment, flags will
5344 * play significant role here.
5346 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5347 const struct tcphdr *th, int syn_inerr)
5349 struct tcp_sock *tp = tcp_sk(sk);
5350 bool rst_seq_match = false;
5352 /* RFC1323: H1. Apply PAWS check first. */
5353 if (tcp_fast_parse_options(sock_net(sk), skb, th, tp) &&
5354 tp->rx_opt.saw_tstamp &&
5355 tcp_paws_discard(sk, skb)) {
5357 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5358 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5359 LINUX_MIB_TCPACKSKIPPEDPAWS,
5360 &tp->last_oow_ack_time))
5361 tcp_send_dupack(sk, skb);
5364 /* Reset is accepted even if it did not pass PAWS. */
5367 /* Step 1: check sequence number */
5368 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5369 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5370 * (RST) segments are validated by checking their SEQ-fields."
5371 * And page 69: "If an incoming segment is not acceptable,
5372 * an acknowledgment should be sent in reply (unless the RST
5373 * bit is set, if so drop the segment and return)".
5378 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5379 LINUX_MIB_TCPACKSKIPPEDSEQ,
5380 &tp->last_oow_ack_time))
5381 tcp_send_dupack(sk, skb);
5382 } else if (tcp_reset_check(sk, skb)) {
5388 /* Step 2: check RST bit */
5390 /* RFC 5961 3.2 (extend to match against (RCV.NXT - 1) after a
5391 * FIN and SACK too if available):
5392 * If seq num matches RCV.NXT or (RCV.NXT - 1) after a FIN, or
5393 * the right-most SACK block,
5395 * RESET the connection
5397 * Send a challenge ACK
5399 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt ||
5400 tcp_reset_check(sk, skb)) {
5401 rst_seq_match = true;
5402 } else if (tcp_is_sack(tp) && tp->rx_opt.num_sacks > 0) {
5403 struct tcp_sack_block *sp = &tp->selective_acks[0];
5404 int max_sack = sp[0].end_seq;
5407 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;
5409 max_sack = after(sp[this_sack].end_seq,
5411 sp[this_sack].end_seq : max_sack;
5414 if (TCP_SKB_CB(skb)->seq == max_sack)
5415 rst_seq_match = true;
5421 /* Disable TFO if RST is out-of-order
5422 * and no data has been received
5423 * for current active TFO socket
5425 if (tp->syn_fastopen && !tp->data_segs_in &&
5426 sk->sk_state == TCP_ESTABLISHED)
5427 tcp_fastopen_active_disable(sk);
5428 tcp_send_challenge_ack(sk, skb);
5433 /* step 3: check security and precedence [ignored] */
5435 /* step 4: Check for a SYN
5436 * RFC 5961 4.2 : Send a challenge ack
5441 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5442 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5443 tcp_send_challenge_ack(sk, skb);
5455 * TCP receive function for the ESTABLISHED state.
5457 * It is split into a fast path and a slow path. The fast path is
5459 * - A zero window was announced from us - zero window probing
5460 * is only handled properly in the slow path.
5461 * - Out of order segments arrived.
5462 * - Urgent data is expected.
5463 * - There is no buffer space left
5464 * - Unexpected TCP flags/window values/header lengths are received
5465 * (detected by checking the TCP header against pred_flags)
5466 * - Data is sent in both directions. Fast path only supports pure senders
5467 * or pure receivers (this means either the sequence number or the ack
5468 * value must stay constant)
5469 * - Unexpected TCP option.
5471 * When these conditions are not satisfied it drops into a standard
5472 * receive procedure patterned after RFC793 to handle all cases.
5473 * The first three cases are guaranteed by proper pred_flags setting,
5474 * the rest is checked inline. Fast processing is turned on in
5475 * tcp_data_queue when everything is OK.
5477 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb)
5479 const struct tcphdr *th = (const struct tcphdr *)skb->data;
5480 struct tcp_sock *tp = tcp_sk(sk);
5481 unsigned int len = skb->len;
5483 /* TCP congestion window tracking */
5484 trace_tcp_probe(sk, skb);
5486 tcp_mstamp_refresh(tp);
5487 if (unlikely(!sk->sk_rx_dst))
5488 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
5490 * Header prediction.
5491 * The code loosely follows the one in the famous
5492 * "30 instruction TCP receive" Van Jacobson mail.
5494 * Van's trick is to deposit buffers into socket queue
5495 * on a device interrupt, to call tcp_recv function
5496 * on the receive process context and checksum and copy
5497 * the buffer to user space. smart...
5499 * Our current scheme is not silly either but we take the
5500 * extra cost of the net_bh soft interrupt processing...
5501 * We do checksum and copy also but from device to kernel.
5504 tp->rx_opt.saw_tstamp = 0;
5506 /* pred_flags is 0xS?10 << 16 + snd_wnd
5507 * if header_prediction is to be made
5508 * 'S' will always be tp->tcp_header_len >> 2
5509 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5510 * turn it off (when there are holes in the receive
5511 * space for instance)
5512 * PSH flag is ignored.
5515 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5516 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5517 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5518 int tcp_header_len = tp->tcp_header_len;
5520 /* Timestamp header prediction: tcp_header_len
5521 * is automatically equal to th->doff*4 due to pred_flags
5525 /* Check timestamp */
5526 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5527 /* No? Slow path! */
5528 if (!tcp_parse_aligned_timestamp(tp, th))
5531 /* If PAWS failed, check it more carefully in slow path */
5532 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5535 /* DO NOT update ts_recent here, if checksum fails
5536 * and timestamp was corrupted part, it will result
5537 * in a hung connection since we will drop all
5538 * future packets due to the PAWS test.
5542 if (len <= tcp_header_len) {
5543 /* Bulk data transfer: sender */
5544 if (len == tcp_header_len) {
5545 /* Predicted packet is in window by definition.
5546 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5547 * Hence, check seq<=rcv_wup reduces to:
5549 if (tcp_header_len ==
5550 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5551 tp->rcv_nxt == tp->rcv_wup)
5552 tcp_store_ts_recent(tp);
5554 /* We know that such packets are checksummed
5557 tcp_ack(sk, skb, 0);
5559 tcp_data_snd_check(sk);
5560 /* When receiving pure ack in fast path, update
5561 * last ts ecr directly instead of calling
5562 * tcp_rcv_rtt_measure_ts()
5564 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr;
5566 } else { /* Header too small */
5567 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5572 bool fragstolen = false;
5574 if (tcp_checksum_complete(skb))
5577 if ((int)skb->truesize > sk->sk_forward_alloc)
5580 /* Predicted packet is in window by definition.
5581 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5582 * Hence, check seq<=rcv_wup reduces to:
5584 if (tcp_header_len ==
5585 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5586 tp->rcv_nxt == tp->rcv_wup)
5587 tcp_store_ts_recent(tp);
5589 tcp_rcv_rtt_measure_ts(sk, skb);
5591 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPHITS);
5593 /* Bulk data transfer: receiver */
5594 eaten = tcp_queue_rcv(sk, skb, tcp_header_len,
5597 tcp_event_data_recv(sk, skb);
5599 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5600 /* Well, only one small jumplet in fast path... */
5601 tcp_ack(sk, skb, FLAG_DATA);
5602 tcp_data_snd_check(sk);
5603 if (!inet_csk_ack_scheduled(sk))
5607 __tcp_ack_snd_check(sk, 0);
5610 kfree_skb_partial(skb, fragstolen);
5617 if (len < (th->doff << 2) || tcp_checksum_complete(skb))
5620 if (!th->ack && !th->rst && !th->syn)
5624 * Standard slow path.
5627 if (!tcp_validate_incoming(sk, skb, th, 1))
5631 if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0)
5634 tcp_rcv_rtt_measure_ts(sk, skb);
5636 /* Process urgent data. */
5637 tcp_urg(sk, skb, th);
5639 /* step 7: process the segment text */
5640 tcp_data_queue(sk, skb);
5642 tcp_data_snd_check(sk);
5643 tcp_ack_snd_check(sk);
5647 TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS);
5648 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5653 EXPORT_SYMBOL(tcp_rcv_established);
5655 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5657 struct tcp_sock *tp = tcp_sk(sk);
5658 struct inet_connection_sock *icsk = inet_csk(sk);
5660 tcp_set_state(sk, TCP_ESTABLISHED);
5661 icsk->icsk_ack.lrcvtime = tcp_jiffies32;
5664 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
5665 security_inet_conn_established(sk, skb);
5666 sk_mark_napi_id(sk, skb);
5669 tcp_init_transfer(sk, BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB);
5671 /* Prevent spurious tcp_cwnd_restart() on first data
5674 tp->lsndtime = tcp_jiffies32;
5676 if (sock_flag(sk, SOCK_KEEPOPEN))
5677 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5679 if (!tp->rx_opt.snd_wscale)
5680 __tcp_fast_path_on(tp, tp->snd_wnd);
5685 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
5686 struct tcp_fastopen_cookie *cookie)
5688 struct tcp_sock *tp = tcp_sk(sk);
5689 struct sk_buff *data = tp->syn_data ? tcp_rtx_queue_head(sk) : NULL;
5690 u16 mss = tp->rx_opt.mss_clamp, try_exp = 0;
5691 bool syn_drop = false;
5693 if (mss == tp->rx_opt.user_mss) {
5694 struct tcp_options_received opt;
5696 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5697 tcp_clear_options(&opt);
5698 opt.user_mss = opt.mss_clamp = 0;
5699 tcp_parse_options(sock_net(sk), synack, &opt, 0, NULL);
5700 mss = opt.mss_clamp;
5703 if (!tp->syn_fastopen) {
5704 /* Ignore an unsolicited cookie */
5706 } else if (tp->total_retrans) {
5707 /* SYN timed out and the SYN-ACK neither has a cookie nor
5708 * acknowledges data. Presumably the remote received only
5709 * the retransmitted (regular) SYNs: either the original
5710 * SYN-data or the corresponding SYN-ACK was dropped.
5712 syn_drop = (cookie->len < 0 && data);
5713 } else if (cookie->len < 0 && !tp->syn_data) {
5714 /* We requested a cookie but didn't get it. If we did not use
5715 * the (old) exp opt format then try so next time (try_exp=1).
5716 * Otherwise we go back to use the RFC7413 opt (try_exp=2).
5718 try_exp = tp->syn_fastopen_exp ? 2 : 1;
5721 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp);
5723 if (data) { /* Retransmit unacked data in SYN */
5724 skb_rbtree_walk_from(data) {
5725 if (__tcp_retransmit_skb(sk, data, 1))
5729 NET_INC_STATS(sock_net(sk),
5730 LINUX_MIB_TCPFASTOPENACTIVEFAIL);
5733 tp->syn_data_acked = tp->syn_data;
5734 if (tp->syn_data_acked) {
5735 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
5736 /* SYN-data is counted as two separate packets in tcp_ack() */
5737 if (tp->delivered > 1)
5741 tcp_fastopen_add_skb(sk, synack);
5746 static void smc_check_reset_syn(struct tcp_sock *tp)
5748 #if IS_ENABLED(CONFIG_SMC)
5749 if (static_branch_unlikely(&tcp_have_smc)) {
5750 if (tp->syn_smc && !tp->rx_opt.smc_ok)
5756 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5757 const struct tcphdr *th)
5759 struct inet_connection_sock *icsk = inet_csk(sk);
5760 struct tcp_sock *tp = tcp_sk(sk);
5761 struct tcp_fastopen_cookie foc = { .len = -1 };
5762 int saved_clamp = tp->rx_opt.mss_clamp;
5765 tcp_parse_options(sock_net(sk), skb, &tp->rx_opt, 0, &foc);
5766 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
5767 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
5771 * "If the state is SYN-SENT then
5772 * first check the ACK bit
5773 * If the ACK bit is set
5774 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5775 * a reset (unless the RST bit is set, if so drop
5776 * the segment and return)"
5778 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
5779 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt))
5780 goto reset_and_undo;
5782 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5783 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5784 tcp_time_stamp(tp))) {
5785 NET_INC_STATS(sock_net(sk),
5786 LINUX_MIB_PAWSACTIVEREJECTED);
5787 goto reset_and_undo;
5790 /* Now ACK is acceptable.
5792 * "If the RST bit is set
5793 * If the ACK was acceptable then signal the user "error:
5794 * connection reset", drop the segment, enter CLOSED state,
5795 * delete TCB, and return."
5804 * "fifth, if neither of the SYN or RST bits is set then
5805 * drop the segment and return."
5811 goto discard_and_undo;
5814 * "If the SYN bit is on ...
5815 * are acceptable then ...
5816 * (our SYN has been ACKed), change the connection
5817 * state to ESTABLISHED..."
5820 tcp_ecn_rcv_synack(tp, th);
5822 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5823 tcp_ack(sk, skb, FLAG_SLOWPATH);
5825 /* Ok.. it's good. Set up sequence numbers and
5826 * move to established.
5828 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5829 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5831 /* RFC1323: The window in SYN & SYN/ACK segments is
5834 tp->snd_wnd = ntohs(th->window);
5836 if (!tp->rx_opt.wscale_ok) {
5837 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5838 tp->window_clamp = min(tp->window_clamp, 65535U);
5841 if (tp->rx_opt.saw_tstamp) {
5842 tp->rx_opt.tstamp_ok = 1;
5843 tp->tcp_header_len =
5844 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5845 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5846 tcp_store_ts_recent(tp);
5848 tp->tcp_header_len = sizeof(struct tcphdr);
5851 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5852 tcp_initialize_rcv_mss(sk);
5854 /* Remember, tcp_poll() does not lock socket!
5855 * Change state from SYN-SENT only after copied_seq
5856 * is initialized. */
5857 tp->copied_seq = tp->rcv_nxt;
5859 smc_check_reset_syn(tp);
5863 tcp_finish_connect(sk, skb);
5865 fastopen_fail = (tp->syn_fastopen || tp->syn_data) &&
5866 tcp_rcv_fastopen_synack(sk, skb, &foc);
5868 if (!sock_flag(sk, SOCK_DEAD)) {
5869 sk->sk_state_change(sk);
5870 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5874 if (sk->sk_write_pending ||
5875 icsk->icsk_accept_queue.rskq_defer_accept ||
5876 icsk->icsk_ack.pingpong) {
5877 /* Save one ACK. Data will be ready after
5878 * several ticks, if write_pending is set.
5880 * It may be deleted, but with this feature tcpdumps
5881 * look so _wonderfully_ clever, that I was not able
5882 * to stand against the temptation 8) --ANK
5884 inet_csk_schedule_ack(sk);
5885 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
5886 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5887 TCP_DELACK_MAX, TCP_RTO_MAX);
5898 /* No ACK in the segment */
5902 * "If the RST bit is set
5904 * Otherwise (no ACK) drop the segment and return."
5907 goto discard_and_undo;
5911 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5912 tcp_paws_reject(&tp->rx_opt, 0))
5913 goto discard_and_undo;
5916 /* We see SYN without ACK. It is attempt of
5917 * simultaneous connect with crossed SYNs.
5918 * Particularly, it can be connect to self.
5920 tcp_set_state(sk, TCP_SYN_RECV);
5922 if (tp->rx_opt.saw_tstamp) {
5923 tp->rx_opt.tstamp_ok = 1;
5924 tcp_store_ts_recent(tp);
5925 tp->tcp_header_len =
5926 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5928 tp->tcp_header_len = sizeof(struct tcphdr);
5931 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5932 tp->copied_seq = tp->rcv_nxt;
5933 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5935 /* RFC1323: The window in SYN & SYN/ACK segments is
5938 tp->snd_wnd = ntohs(th->window);
5939 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5940 tp->max_window = tp->snd_wnd;
5942 tcp_ecn_rcv_syn(tp, th);
5945 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5946 tcp_initialize_rcv_mss(sk);
5948 tcp_send_synack(sk);
5950 /* Note, we could accept data and URG from this segment.
5951 * There are no obstacles to make this (except that we must
5952 * either change tcp_recvmsg() to prevent it from returning data
5953 * before 3WHS completes per RFC793, or employ TCP Fast Open).
5955 * However, if we ignore data in ACKless segments sometimes,
5956 * we have no reasons to accept it sometimes.
5957 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5958 * is not flawless. So, discard packet for sanity.
5959 * Uncomment this return to process the data.
5966 /* "fifth, if neither of the SYN or RST bits is set then
5967 * drop the segment and return."
5971 tcp_clear_options(&tp->rx_opt);
5972 tp->rx_opt.mss_clamp = saved_clamp;
5976 tcp_clear_options(&tp->rx_opt);
5977 tp->rx_opt.mss_clamp = saved_clamp;
5982 * This function implements the receiving procedure of RFC 793 for
5983 * all states except ESTABLISHED and TIME_WAIT.
5984 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5985 * address independent.
5988 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb)
5990 struct tcp_sock *tp = tcp_sk(sk);
5991 struct inet_connection_sock *icsk = inet_csk(sk);
5992 const struct tcphdr *th = tcp_hdr(skb);
5993 struct request_sock *req;
5997 switch (sk->sk_state) {
6011 /* It is possible that we process SYN packets from backlog,
6012 * so we need to make sure to disable BH right there.
6015 acceptable = icsk->icsk_af_ops->conn_request(sk, skb) >= 0;
6026 tp->rx_opt.saw_tstamp = 0;
6027 tcp_mstamp_refresh(tp);
6028 queued = tcp_rcv_synsent_state_process(sk, skb, th);
6032 /* Do step6 onward by hand. */
6033 tcp_urg(sk, skb, th);
6035 tcp_data_snd_check(sk);
6039 tcp_mstamp_refresh(tp);
6040 tp->rx_opt.saw_tstamp = 0;
6041 req = tp->fastopen_rsk;
6045 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
6046 sk->sk_state != TCP_FIN_WAIT1);
6048 if (!tcp_check_req(sk, skb, req, true, &req_stolen))
6052 if (!th->ack && !th->rst && !th->syn)
6055 if (!tcp_validate_incoming(sk, skb, th, 0))
6058 /* step 5: check the ACK field */
6059 acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
6060 FLAG_UPDATE_TS_RECENT |
6061 FLAG_NO_CHALLENGE_ACK) > 0;
6064 if (sk->sk_state == TCP_SYN_RECV)
6065 return 1; /* send one RST */
6066 tcp_send_challenge_ack(sk, skb);
6069 switch (sk->sk_state) {
6071 tp->delivered++; /* SYN-ACK delivery isn't tracked in tcp_ack */
6073 tcp_synack_rtt_meas(sk, req);
6075 /* Once we leave TCP_SYN_RECV, we no longer need req
6079 inet_csk(sk)->icsk_retransmits = 0;
6080 reqsk_fastopen_remove(sk, req, false);
6081 /* Re-arm the timer because data may have been sent out.
6082 * This is similar to the regular data transmission case
6083 * when new data has just been ack'ed.
6085 * (TFO) - we could try to be more aggressive and
6086 * retransmitting any data sooner based on when they
6091 tcp_init_transfer(sk, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB);
6092 tp->copied_seq = tp->rcv_nxt;
6095 tcp_set_state(sk, TCP_ESTABLISHED);
6096 sk->sk_state_change(sk);
6098 /* Note, that this wakeup is only for marginal crossed SYN case.
6099 * Passively open sockets are not waked up, because
6100 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
6103 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
6105 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
6106 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
6107 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
6109 if (tp->rx_opt.tstamp_ok)
6110 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
6112 if (!inet_csk(sk)->icsk_ca_ops->cong_control)
6113 tcp_update_pacing_rate(sk);
6115 /* Prevent spurious tcp_cwnd_restart() on first data packet */
6116 tp->lsndtime = tcp_jiffies32;
6118 tcp_initialize_rcv_mss(sk);
6119 tcp_fast_path_on(tp);
6122 case TCP_FIN_WAIT1: {
6125 /* If we enter the TCP_FIN_WAIT1 state and we are a
6126 * Fast Open socket and this is the first acceptable
6127 * ACK we have received, this would have acknowledged
6128 * our SYNACK so stop the SYNACK timer.
6131 /* We no longer need the request sock. */
6132 reqsk_fastopen_remove(sk, req, false);
6135 if (tp->snd_una != tp->write_seq)
6138 tcp_set_state(sk, TCP_FIN_WAIT2);
6139 sk->sk_shutdown |= SEND_SHUTDOWN;
6143 if (!sock_flag(sk, SOCK_DEAD)) {
6144 /* Wake up lingering close() */
6145 sk->sk_state_change(sk);
6149 if (tp->linger2 < 0) {
6151 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6154 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6155 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6156 /* Receive out of order FIN after close() */
6157 if (tp->syn_fastopen && th->fin)
6158 tcp_fastopen_active_disable(sk);
6160 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6164 tmo = tcp_fin_time(sk);
6165 if (tmo > TCP_TIMEWAIT_LEN) {
6166 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
6167 } else if (th->fin || sock_owned_by_user(sk)) {
6168 /* Bad case. We could lose such FIN otherwise.
6169 * It is not a big problem, but it looks confusing
6170 * and not so rare event. We still can lose it now,
6171 * if it spins in bh_lock_sock(), but it is really
6174 inet_csk_reset_keepalive_timer(sk, tmo);
6176 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
6183 if (tp->snd_una == tp->write_seq) {
6184 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
6190 if (tp->snd_una == tp->write_seq) {
6191 tcp_update_metrics(sk);
6198 /* step 6: check the URG bit */
6199 tcp_urg(sk, skb, th);
6201 /* step 7: process the segment text */
6202 switch (sk->sk_state) {
6203 case TCP_CLOSE_WAIT:
6206 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
6211 /* RFC 793 says to queue data in these states,
6212 * RFC 1122 says we MUST send a reset.
6213 * BSD 4.4 also does reset.
6215 if (sk->sk_shutdown & RCV_SHUTDOWN) {
6216 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6217 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6218 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6224 case TCP_ESTABLISHED:
6225 tcp_data_queue(sk, skb);
6230 /* tcp_data could move socket to TIME-WAIT */
6231 if (sk->sk_state != TCP_CLOSE) {
6232 tcp_data_snd_check(sk);
6233 tcp_ack_snd_check(sk);
6242 EXPORT_SYMBOL(tcp_rcv_state_process);
6244 static inline void pr_drop_req(struct request_sock *req, __u16 port, int family)
6246 struct inet_request_sock *ireq = inet_rsk(req);
6248 if (family == AF_INET)
6249 net_dbg_ratelimited("drop open request from %pI4/%u\n",
6250 &ireq->ir_rmt_addr, port);
6251 #if IS_ENABLED(CONFIG_IPV6)
6252 else if (family == AF_INET6)
6253 net_dbg_ratelimited("drop open request from %pI6/%u\n",
6254 &ireq->ir_v6_rmt_addr, port);
6258 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
6260 * If we receive a SYN packet with these bits set, it means a
6261 * network is playing bad games with TOS bits. In order to
6262 * avoid possible false congestion notifications, we disable
6263 * TCP ECN negotiation.
6265 * Exception: tcp_ca wants ECN. This is required for DCTCP
6266 * congestion control: Linux DCTCP asserts ECT on all packets,
6267 * including SYN, which is most optimal solution; however,
6268 * others, such as FreeBSD do not.
6270 static void tcp_ecn_create_request(struct request_sock *req,
6271 const struct sk_buff *skb,
6272 const struct sock *listen_sk,
6273 const struct dst_entry *dst)
6275 const struct tcphdr *th = tcp_hdr(skb);
6276 const struct net *net = sock_net(listen_sk);
6277 bool th_ecn = th->ece && th->cwr;
6284 ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield);
6285 ecn_ok_dst = dst_feature(dst, DST_FEATURE_ECN_MASK);
6286 ecn_ok = net->ipv4.sysctl_tcp_ecn || ecn_ok_dst;
6288 if ((!ect && ecn_ok) || tcp_ca_needs_ecn(listen_sk) ||
6289 (ecn_ok_dst & DST_FEATURE_ECN_CA) ||
6290 tcp_bpf_ca_needs_ecn((struct sock *)req))
6291 inet_rsk(req)->ecn_ok = 1;
6294 static void tcp_openreq_init(struct request_sock *req,
6295 const struct tcp_options_received *rx_opt,
6296 struct sk_buff *skb, const struct sock *sk)
6298 struct inet_request_sock *ireq = inet_rsk(req);
6300 req->rsk_rcv_wnd = 0; /* So that tcp_send_synack() knows! */
6302 tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq;
6303 tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
6304 tcp_rsk(req)->snt_synack = tcp_clock_us();
6305 tcp_rsk(req)->last_oow_ack_time = 0;
6306 req->mss = rx_opt->mss_clamp;
6307 req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0;
6308 ireq->tstamp_ok = rx_opt->tstamp_ok;
6309 ireq->sack_ok = rx_opt->sack_ok;
6310 ireq->snd_wscale = rx_opt->snd_wscale;
6311 ireq->wscale_ok = rx_opt->wscale_ok;
6314 ireq->ir_rmt_port = tcp_hdr(skb)->source;
6315 ireq->ir_num = ntohs(tcp_hdr(skb)->dest);
6316 ireq->ir_mark = inet_request_mark(sk, skb);
6317 #if IS_ENABLED(CONFIG_SMC)
6318 ireq->smc_ok = rx_opt->smc_ok;
6322 struct request_sock *inet_reqsk_alloc(const struct request_sock_ops *ops,
6323 struct sock *sk_listener,
6324 bool attach_listener)
6326 struct request_sock *req = reqsk_alloc(ops, sk_listener,
6330 struct inet_request_sock *ireq = inet_rsk(req);
6332 ireq->ireq_opt = NULL;
6333 #if IS_ENABLED(CONFIG_IPV6)
6334 ireq->pktopts = NULL;
6336 atomic64_set(&ireq->ir_cookie, 0);
6337 ireq->ireq_state = TCP_NEW_SYN_RECV;
6338 write_pnet(&ireq->ireq_net, sock_net(sk_listener));
6339 ireq->ireq_family = sk_listener->sk_family;
6344 EXPORT_SYMBOL(inet_reqsk_alloc);
6347 * Return true if a syncookie should be sent
6349 static bool tcp_syn_flood_action(const struct sock *sk,
6350 const struct sk_buff *skb,
6353 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
6354 const char *msg = "Dropping request";
6355 bool want_cookie = false;
6356 struct net *net = sock_net(sk);
6358 #ifdef CONFIG_SYN_COOKIES
6359 if (net->ipv4.sysctl_tcp_syncookies) {
6360 msg = "Sending cookies";
6362 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES);
6365 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP);
6367 if (!queue->synflood_warned &&
6368 net->ipv4.sysctl_tcp_syncookies != 2 &&
6369 xchg(&queue->synflood_warned, 1) == 0)
6370 net_info_ratelimited("%s: Possible SYN flooding on port %d. %s. Check SNMP counters.\n",
6371 proto, ntohs(tcp_hdr(skb)->dest), msg);
6376 static void tcp_reqsk_record_syn(const struct sock *sk,
6377 struct request_sock *req,
6378 const struct sk_buff *skb)
6380 if (tcp_sk(sk)->save_syn) {
6381 u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb);
6384 copy = kmalloc(len + sizeof(u32), GFP_ATOMIC);
6387 memcpy(©[1], skb_network_header(skb), len);
6388 req->saved_syn = copy;
6393 int tcp_conn_request(struct request_sock_ops *rsk_ops,
6394 const struct tcp_request_sock_ops *af_ops,
6395 struct sock *sk, struct sk_buff *skb)
6397 struct tcp_fastopen_cookie foc = { .len = -1 };
6398 __u32 isn = TCP_SKB_CB(skb)->tcp_tw_isn;
6399 struct tcp_options_received tmp_opt;
6400 struct tcp_sock *tp = tcp_sk(sk);
6401 struct net *net = sock_net(sk);
6402 struct sock *fastopen_sk = NULL;
6403 struct request_sock *req;
6404 bool want_cookie = false;
6405 struct dst_entry *dst;
6408 /* TW buckets are converted to open requests without
6409 * limitations, they conserve resources and peer is
6410 * evidently real one.
6412 if ((net->ipv4.sysctl_tcp_syncookies == 2 ||
6413 inet_csk_reqsk_queue_is_full(sk)) && !isn) {
6414 want_cookie = tcp_syn_flood_action(sk, skb, rsk_ops->slab_name);
6419 if (sk_acceptq_is_full(sk)) {
6420 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
6424 req = inet_reqsk_alloc(rsk_ops, sk, !want_cookie);
6428 tcp_rsk(req)->af_specific = af_ops;
6429 tcp_rsk(req)->ts_off = 0;
6431 tcp_clear_options(&tmp_opt);
6432 tmp_opt.mss_clamp = af_ops->mss_clamp;
6433 tmp_opt.user_mss = tp->rx_opt.user_mss;
6434 tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0,
6435 want_cookie ? NULL : &foc);
6437 if (want_cookie && !tmp_opt.saw_tstamp)
6438 tcp_clear_options(&tmp_opt);
6440 if (IS_ENABLED(CONFIG_SMC) && want_cookie)
6443 tmp_opt.tstamp_ok = tmp_opt.saw_tstamp;
6444 tcp_openreq_init(req, &tmp_opt, skb, sk);
6445 inet_rsk(req)->no_srccheck = inet_sk(sk)->transparent;
6447 /* Note: tcp_v6_init_req() might override ir_iif for link locals */
6448 inet_rsk(req)->ir_iif = inet_request_bound_dev_if(sk, skb);
6450 af_ops->init_req(req, sk, skb);
6452 if (security_inet_conn_request(sk, skb, req))
6455 if (tmp_opt.tstamp_ok)
6456 tcp_rsk(req)->ts_off = af_ops->init_ts_off(net, skb);
6458 dst = af_ops->route_req(sk, &fl, req);
6462 if (!want_cookie && !isn) {
6463 /* Kill the following clause, if you dislike this way. */
6464 if (!net->ipv4.sysctl_tcp_syncookies &&
6465 (net->ipv4.sysctl_max_syn_backlog - inet_csk_reqsk_queue_len(sk) <
6466 (net->ipv4.sysctl_max_syn_backlog >> 2)) &&
6467 !tcp_peer_is_proven(req, dst)) {
6468 /* Without syncookies last quarter of
6469 * backlog is filled with destinations,
6470 * proven to be alive.
6471 * It means that we continue to communicate
6472 * to destinations, already remembered
6473 * to the moment of synflood.
6475 pr_drop_req(req, ntohs(tcp_hdr(skb)->source),
6477 goto drop_and_release;
6480 isn = af_ops->init_seq(skb);
6483 tcp_ecn_create_request(req, skb, sk, dst);
6486 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss);
6487 req->cookie_ts = tmp_opt.tstamp_ok;
6488 if (!tmp_opt.tstamp_ok)
6489 inet_rsk(req)->ecn_ok = 0;
6492 tcp_rsk(req)->snt_isn = isn;
6493 tcp_rsk(req)->txhash = net_tx_rndhash();
6494 tcp_openreq_init_rwin(req, sk, dst);
6495 sk_rx_queue_set(req_to_sk(req), skb);
6497 tcp_reqsk_record_syn(sk, req, skb);
6498 fastopen_sk = tcp_try_fastopen(sk, skb, req, &foc, dst);
6501 af_ops->send_synack(fastopen_sk, dst, &fl, req,
6502 &foc, TCP_SYNACK_FASTOPEN);
6503 /* Add the child socket directly into the accept queue */
6504 inet_csk_reqsk_queue_add(sk, req, fastopen_sk);
6505 sk->sk_data_ready(sk);
6506 bh_unlock_sock(fastopen_sk);
6507 sock_put(fastopen_sk);
6509 tcp_rsk(req)->tfo_listener = false;
6511 inet_csk_reqsk_queue_hash_add(sk, req,
6512 tcp_timeout_init((struct sock *)req));
6513 af_ops->send_synack(sk, dst, &fl, req, &foc,
6514 !want_cookie ? TCP_SYNACK_NORMAL :
6532 EXPORT_SYMBOL(tcp_conn_request);