1 // SPDX-License-Identifier: GPL-2.0
3 * INET An implementation of the TCP/IP protocol suite for the LINUX
4 * operating system. INET is implemented using the BSD Socket
5 * interface as the means of communication with the user level.
7 * Implementation of the Transmission Control Protocol(TCP).
10 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
11 * Mark Evans, <evansmp@uhura.aston.ac.uk>
12 * Corey Minyard <wf-rch!minyard@relay.EU.net>
13 * Florian La Roche, <flla@stud.uni-sb.de>
14 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
15 * Linus Torvalds, <torvalds@cs.helsinki.fi>
16 * Alan Cox, <gw4pts@gw4pts.ampr.org>
17 * Matthew Dillon, <dillon@apollo.west.oic.com>
18 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
19 * Jorge Cwik, <jorge@laser.satlink.net>
24 * Pedro Roque : Fast Retransmit/Recovery.
26 * Retransmit queue handled by TCP.
27 * Better retransmit timer handling.
28 * New congestion avoidance.
32 * Eric : Fast Retransmit.
33 * Randy Scott : MSS option defines.
34 * Eric Schenk : Fixes to slow start algorithm.
35 * Eric Schenk : Yet another double ACK bug.
36 * Eric Schenk : Delayed ACK bug fixes.
37 * Eric Schenk : Floyd style fast retrans war avoidance.
38 * David S. Miller : Don't allow zero congestion window.
39 * Eric Schenk : Fix retransmitter so that it sends
40 * next packet on ack of previous packet.
41 * Andi Kleen : Moved open_request checking here
42 * and process RSTs for open_requests.
43 * Andi Kleen : Better prune_queue, and other fixes.
44 * Andrey Savochkin: Fix RTT measurements in the presence of
46 * Andrey Savochkin: Check sequence numbers correctly when
47 * removing SACKs due to in sequence incoming
49 * Andi Kleen: Make sure we never ack data there is not
50 * enough room for. Also make this condition
51 * a fatal error if it might still happen.
52 * Andi Kleen: Add tcp_measure_rcv_mss to make
53 * connections with MSS<min(MTU,ann. MSS)
54 * work without delayed acks.
55 * Andi Kleen: Process packets with PSH set in the
57 * J Hadi Salim: ECN support
60 * Panu Kuhlberg: Experimental audit of TCP (re)transmission
61 * engine. Lots of bugs are found.
62 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs
65 #define pr_fmt(fmt) "TCP: " fmt
68 #include <linux/slab.h>
69 #include <linux/module.h>
70 #include <linux/sysctl.h>
71 #include <linux/kernel.h>
72 #include <linux/prefetch.h>
75 #include <net/inet_common.h>
76 #include <linux/ipsec.h>
77 #include <asm/unaligned.h>
78 #include <linux/errqueue.h>
79 #include <trace/events/tcp.h>
80 #include <linux/jump_label_ratelimit.h>
81 #include <net/busy_poll.h>
82 #include <net/mptcp.h>
84 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
86 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
87 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
88 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
89 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
90 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
91 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
92 #define FLAG_ECE 0x40 /* ECE in this ACK */
93 #define FLAG_LOST_RETRANS 0x80 /* This ACK marks some retransmission lost */
94 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
95 #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
96 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
97 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
98 #define FLAG_SET_XMIT_TIMER 0x1000 /* Set TLP or RTO timer */
99 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
100 #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
101 #define FLAG_NO_CHALLENGE_ACK 0x8000 /* do not call tcp_send_challenge_ack() */
102 #define FLAG_ACK_MAYBE_DELAYED 0x10000 /* Likely a delayed ACK */
104 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
105 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
106 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE|FLAG_DSACKING_ACK)
107 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
109 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
110 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
112 #define REXMIT_NONE 0 /* no loss recovery to do */
113 #define REXMIT_LOST 1 /* retransmit packets marked lost */
114 #define REXMIT_NEW 2 /* FRTO-style transmit of unsent/new packets */
116 #if IS_ENABLED(CONFIG_TLS_DEVICE)
117 static DEFINE_STATIC_KEY_DEFERRED_FALSE(clean_acked_data_enabled, HZ);
119 void clean_acked_data_enable(struct inet_connection_sock *icsk,
120 void (*cad)(struct sock *sk, u32 ack_seq))
122 icsk->icsk_clean_acked = cad;
123 static_branch_deferred_inc(&clean_acked_data_enabled);
125 EXPORT_SYMBOL_GPL(clean_acked_data_enable);
127 void clean_acked_data_disable(struct inet_connection_sock *icsk)
129 static_branch_slow_dec_deferred(&clean_acked_data_enabled);
130 icsk->icsk_clean_acked = NULL;
132 EXPORT_SYMBOL_GPL(clean_acked_data_disable);
134 void clean_acked_data_flush(void)
136 static_key_deferred_flush(&clean_acked_data_enabled);
138 EXPORT_SYMBOL_GPL(clean_acked_data_flush);
141 static void tcp_gro_dev_warn(struct sock *sk, const struct sk_buff *skb,
144 static bool __once __read_mostly;
147 struct net_device *dev;
152 dev = dev_get_by_index_rcu(sock_net(sk), skb->skb_iif);
153 if (!dev || len >= dev->mtu)
154 pr_warn("%s: Driver has suspect GRO implementation, TCP performance may be compromised.\n",
155 dev ? dev->name : "Unknown driver");
160 /* Adapt the MSS value used to make delayed ack decision to the
163 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
165 struct inet_connection_sock *icsk = inet_csk(sk);
166 const unsigned int lss = icsk->icsk_ack.last_seg_size;
169 icsk->icsk_ack.last_seg_size = 0;
171 /* skb->len may jitter because of SACKs, even if peer
172 * sends good full-sized frames.
174 len = skb_shinfo(skb)->gso_size ? : skb->len;
175 if (len >= icsk->icsk_ack.rcv_mss) {
176 icsk->icsk_ack.rcv_mss = min_t(unsigned int, len,
178 /* Account for possibly-removed options */
179 if (unlikely(len > icsk->icsk_ack.rcv_mss +
180 MAX_TCP_OPTION_SPACE))
181 tcp_gro_dev_warn(sk, skb, len);
183 /* Otherwise, we make more careful check taking into account,
184 * that SACKs block is variable.
186 * "len" is invariant segment length, including TCP header.
188 len += skb->data - skb_transport_header(skb);
189 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
190 /* If PSH is not set, packet should be
191 * full sized, provided peer TCP is not badly broken.
192 * This observation (if it is correct 8)) allows
193 * to handle super-low mtu links fairly.
195 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
196 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
197 /* Subtract also invariant (if peer is RFC compliant),
198 * tcp header plus fixed timestamp option length.
199 * Resulting "len" is MSS free of SACK jitter.
201 len -= tcp_sk(sk)->tcp_header_len;
202 icsk->icsk_ack.last_seg_size = len;
204 icsk->icsk_ack.rcv_mss = len;
208 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
209 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
210 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
214 static void tcp_incr_quickack(struct sock *sk, unsigned int max_quickacks)
216 struct inet_connection_sock *icsk = inet_csk(sk);
217 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
221 quickacks = min(quickacks, max_quickacks);
222 if (quickacks > icsk->icsk_ack.quick)
223 icsk->icsk_ack.quick = quickacks;
226 void tcp_enter_quickack_mode(struct sock *sk, unsigned int max_quickacks)
228 struct inet_connection_sock *icsk = inet_csk(sk);
230 tcp_incr_quickack(sk, max_quickacks);
231 inet_csk_exit_pingpong_mode(sk);
232 icsk->icsk_ack.ato = TCP_ATO_MIN;
234 EXPORT_SYMBOL(tcp_enter_quickack_mode);
236 /* Send ACKs quickly, if "quick" count is not exhausted
237 * and the session is not interactive.
240 static bool tcp_in_quickack_mode(struct sock *sk)
242 const struct inet_connection_sock *icsk = inet_csk(sk);
243 const struct dst_entry *dst = __sk_dst_get(sk);
245 return (dst && dst_metric(dst, RTAX_QUICKACK)) ||
246 (icsk->icsk_ack.quick && !inet_csk_in_pingpong_mode(sk));
249 static void tcp_ecn_queue_cwr(struct tcp_sock *tp)
251 if (tp->ecn_flags & TCP_ECN_OK)
252 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
255 static void tcp_ecn_accept_cwr(struct sock *sk, const struct sk_buff *skb)
257 if (tcp_hdr(skb)->cwr) {
258 tcp_sk(sk)->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
260 /* If the sender is telling us it has entered CWR, then its
261 * cwnd may be very low (even just 1 packet), so we should ACK
264 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW;
268 static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp)
270 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
273 static void __tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb)
275 struct tcp_sock *tp = tcp_sk(sk);
277 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
278 case INET_ECN_NOT_ECT:
279 /* Funny extension: if ECT is not set on a segment,
280 * and we already seen ECT on a previous segment,
281 * it is probably a retransmit.
283 if (tp->ecn_flags & TCP_ECN_SEEN)
284 tcp_enter_quickack_mode(sk, 2);
287 if (tcp_ca_needs_ecn(sk))
288 tcp_ca_event(sk, CA_EVENT_ECN_IS_CE);
290 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
291 /* Better not delay acks, sender can have a very low cwnd */
292 tcp_enter_quickack_mode(sk, 2);
293 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
295 tp->ecn_flags |= TCP_ECN_SEEN;
298 if (tcp_ca_needs_ecn(sk))
299 tcp_ca_event(sk, CA_EVENT_ECN_NO_CE);
300 tp->ecn_flags |= TCP_ECN_SEEN;
305 static void tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb)
307 if (tcp_sk(sk)->ecn_flags & TCP_ECN_OK)
308 __tcp_ecn_check_ce(sk, skb);
311 static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
313 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
314 tp->ecn_flags &= ~TCP_ECN_OK;
317 static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
319 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
320 tp->ecn_flags &= ~TCP_ECN_OK;
323 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
325 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
330 /* Buffer size and advertised window tuning.
332 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
335 static void tcp_sndbuf_expand(struct sock *sk)
337 const struct tcp_sock *tp = tcp_sk(sk);
338 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
342 /* Worst case is non GSO/TSO : each frame consumes one skb
343 * and skb->head is kmalloced using power of two area of memory
345 per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
347 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
349 per_mss = roundup_pow_of_two(per_mss) +
350 SKB_DATA_ALIGN(sizeof(struct sk_buff));
352 nr_segs = max_t(u32, TCP_INIT_CWND, tp->snd_cwnd);
353 nr_segs = max_t(u32, nr_segs, tp->reordering + 1);
355 /* Fast Recovery (RFC 5681 3.2) :
356 * Cubic needs 1.7 factor, rounded to 2 to include
357 * extra cushion (application might react slowly to EPOLLOUT)
359 sndmem = ca_ops->sndbuf_expand ? ca_ops->sndbuf_expand(sk) : 2;
360 sndmem *= nr_segs * per_mss;
362 if (sk->sk_sndbuf < sndmem)
363 WRITE_ONCE(sk->sk_sndbuf,
364 min(sndmem, sock_net(sk)->ipv4.sysctl_tcp_wmem[2]));
367 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
369 * All tcp_full_space() is split to two parts: "network" buffer, allocated
370 * forward and advertised in receiver window (tp->rcv_wnd) and
371 * "application buffer", required to isolate scheduling/application
372 * latencies from network.
373 * window_clamp is maximal advertised window. It can be less than
374 * tcp_full_space(), in this case tcp_full_space() - window_clamp
375 * is reserved for "application" buffer. The less window_clamp is
376 * the smoother our behaviour from viewpoint of network, but the lower
377 * throughput and the higher sensitivity of the connection to losses. 8)
379 * rcv_ssthresh is more strict window_clamp used at "slow start"
380 * phase to predict further behaviour of this connection.
381 * It is used for two goals:
382 * - to enforce header prediction at sender, even when application
383 * requires some significant "application buffer". It is check #1.
384 * - to prevent pruning of receive queue because of misprediction
385 * of receiver window. Check #2.
387 * The scheme does not work when sender sends good segments opening
388 * window and then starts to feed us spaghetti. But it should work
389 * in common situations. Otherwise, we have to rely on queue collapsing.
392 /* Slow part of check#2. */
393 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
395 struct tcp_sock *tp = tcp_sk(sk);
397 int truesize = tcp_win_from_space(sk, skb->truesize) >> 1;
398 int window = tcp_win_from_space(sk, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]) >> 1;
400 while (tp->rcv_ssthresh <= window) {
401 if (truesize <= skb->len)
402 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
410 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
412 struct tcp_sock *tp = tcp_sk(sk);
415 room = min_t(int, tp->window_clamp, tcp_space(sk)) - tp->rcv_ssthresh;
418 if (room > 0 && !tcp_under_memory_pressure(sk)) {
421 /* Check #2. Increase window, if skb with such overhead
422 * will fit to rcvbuf in future.
424 if (tcp_win_from_space(sk, skb->truesize) <= skb->len)
425 incr = 2 * tp->advmss;
427 incr = __tcp_grow_window(sk, skb);
430 incr = max_t(int, incr, 2 * skb->len);
431 tp->rcv_ssthresh += min(room, incr);
432 inet_csk(sk)->icsk_ack.quick |= 1;
437 /* 3. Try to fixup all. It is made immediately after connection enters
440 void tcp_init_buffer_space(struct sock *sk)
442 int tcp_app_win = sock_net(sk)->ipv4.sysctl_tcp_app_win;
443 struct tcp_sock *tp = tcp_sk(sk);
446 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
447 tcp_sndbuf_expand(sk);
449 tp->rcvq_space.space = min_t(u32, tp->rcv_wnd, TCP_INIT_CWND * tp->advmss);
450 tcp_mstamp_refresh(tp);
451 tp->rcvq_space.time = tp->tcp_mstamp;
452 tp->rcvq_space.seq = tp->copied_seq;
454 maxwin = tcp_full_space(sk);
456 if (tp->window_clamp >= maxwin) {
457 tp->window_clamp = maxwin;
459 if (tcp_app_win && maxwin > 4 * tp->advmss)
460 tp->window_clamp = max(maxwin -
461 (maxwin >> tcp_app_win),
465 /* Force reservation of one segment. */
467 tp->window_clamp > 2 * tp->advmss &&
468 tp->window_clamp + tp->advmss > maxwin)
469 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
471 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
472 tp->snd_cwnd_stamp = tcp_jiffies32;
475 /* 4. Recalculate window clamp after socket hit its memory bounds. */
476 static void tcp_clamp_window(struct sock *sk)
478 struct tcp_sock *tp = tcp_sk(sk);
479 struct inet_connection_sock *icsk = inet_csk(sk);
480 struct net *net = sock_net(sk);
482 icsk->icsk_ack.quick = 0;
484 if (sk->sk_rcvbuf < net->ipv4.sysctl_tcp_rmem[2] &&
485 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
486 !tcp_under_memory_pressure(sk) &&
487 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
488 WRITE_ONCE(sk->sk_rcvbuf,
489 min(atomic_read(&sk->sk_rmem_alloc),
490 net->ipv4.sysctl_tcp_rmem[2]));
492 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
493 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
496 /* Initialize RCV_MSS value.
497 * RCV_MSS is an our guess about MSS used by the peer.
498 * We haven't any direct information about the MSS.
499 * It's better to underestimate the RCV_MSS rather than overestimate.
500 * Overestimations make us ACKing less frequently than needed.
501 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
503 void tcp_initialize_rcv_mss(struct sock *sk)
505 const struct tcp_sock *tp = tcp_sk(sk);
506 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
508 hint = min(hint, tp->rcv_wnd / 2);
509 hint = min(hint, TCP_MSS_DEFAULT);
510 hint = max(hint, TCP_MIN_MSS);
512 inet_csk(sk)->icsk_ack.rcv_mss = hint;
514 EXPORT_SYMBOL(tcp_initialize_rcv_mss);
516 /* Receiver "autotuning" code.
518 * The algorithm for RTT estimation w/o timestamps is based on
519 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
520 * <http://public.lanl.gov/radiant/pubs.html#DRS>
522 * More detail on this code can be found at
523 * <http://staff.psc.edu/jheffner/>,
524 * though this reference is out of date. A new paper
527 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
529 u32 new_sample = tp->rcv_rtt_est.rtt_us;
532 if (new_sample != 0) {
533 /* If we sample in larger samples in the non-timestamp
534 * case, we could grossly overestimate the RTT especially
535 * with chatty applications or bulk transfer apps which
536 * are stalled on filesystem I/O.
538 * Also, since we are only going for a minimum in the
539 * non-timestamp case, we do not smooth things out
540 * else with timestamps disabled convergence takes too
544 m -= (new_sample >> 3);
552 /* No previous measure. */
556 tp->rcv_rtt_est.rtt_us = new_sample;
559 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
563 if (tp->rcv_rtt_est.time == 0)
565 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
567 delta_us = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcv_rtt_est.time);
570 tcp_rcv_rtt_update(tp, delta_us, 1);
573 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
574 tp->rcv_rtt_est.time = tp->tcp_mstamp;
577 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
578 const struct sk_buff *skb)
580 struct tcp_sock *tp = tcp_sk(sk);
582 if (tp->rx_opt.rcv_tsecr == tp->rcv_rtt_last_tsecr)
584 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr;
586 if (TCP_SKB_CB(skb)->end_seq -
587 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss) {
588 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr;
591 if (likely(delta < INT_MAX / (USEC_PER_SEC / TCP_TS_HZ))) {
594 delta_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
595 tcp_rcv_rtt_update(tp, delta_us, 0);
601 * This function should be called every time data is copied to user space.
602 * It calculates the appropriate TCP receive buffer space.
604 void tcp_rcv_space_adjust(struct sock *sk)
606 struct tcp_sock *tp = tcp_sk(sk);
610 trace_tcp_rcv_space_adjust(sk);
612 tcp_mstamp_refresh(tp);
613 time = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcvq_space.time);
614 if (time < (tp->rcv_rtt_est.rtt_us >> 3) || tp->rcv_rtt_est.rtt_us == 0)
617 /* Number of bytes copied to user in last RTT */
618 copied = tp->copied_seq - tp->rcvq_space.seq;
619 if (copied <= tp->rcvq_space.space)
623 * copied = bytes received in previous RTT, our base window
624 * To cope with packet losses, we need a 2x factor
625 * To cope with slow start, and sender growing its cwin by 100 %
626 * every RTT, we need a 4x factor, because the ACK we are sending
627 * now is for the next RTT, not the current one :
628 * <prev RTT . ><current RTT .. ><next RTT .... >
631 if (sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf &&
632 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
636 /* minimal window to cope with packet losses, assuming
637 * steady state. Add some cushion because of small variations.
639 rcvwin = ((u64)copied << 1) + 16 * tp->advmss;
641 /* Accommodate for sender rate increase (eg. slow start) */
642 grow = rcvwin * (copied - tp->rcvq_space.space);
643 do_div(grow, tp->rcvq_space.space);
644 rcvwin += (grow << 1);
646 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
647 while (tcp_win_from_space(sk, rcvmem) < tp->advmss)
650 do_div(rcvwin, tp->advmss);
651 rcvbuf = min_t(u64, rcvwin * rcvmem,
652 sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
653 if (rcvbuf > sk->sk_rcvbuf) {
654 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
656 /* Make the window clamp follow along. */
657 tp->window_clamp = tcp_win_from_space(sk, rcvbuf);
660 tp->rcvq_space.space = copied;
663 tp->rcvq_space.seq = tp->copied_seq;
664 tp->rcvq_space.time = tp->tcp_mstamp;
667 /* There is something which you must keep in mind when you analyze the
668 * behavior of the tp->ato delayed ack timeout interval. When a
669 * connection starts up, we want to ack as quickly as possible. The
670 * problem is that "good" TCP's do slow start at the beginning of data
671 * transmission. The means that until we send the first few ACK's the
672 * sender will sit on his end and only queue most of his data, because
673 * he can only send snd_cwnd unacked packets at any given time. For
674 * each ACK we send, he increments snd_cwnd and transmits more of his
677 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
679 struct tcp_sock *tp = tcp_sk(sk);
680 struct inet_connection_sock *icsk = inet_csk(sk);
683 inet_csk_schedule_ack(sk);
685 tcp_measure_rcv_mss(sk, skb);
687 tcp_rcv_rtt_measure(tp);
691 if (!icsk->icsk_ack.ato) {
692 /* The _first_ data packet received, initialize
693 * delayed ACK engine.
695 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
696 icsk->icsk_ack.ato = TCP_ATO_MIN;
698 int m = now - icsk->icsk_ack.lrcvtime;
700 if (m <= TCP_ATO_MIN / 2) {
701 /* The fastest case is the first. */
702 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
703 } else if (m < icsk->icsk_ack.ato) {
704 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
705 if (icsk->icsk_ack.ato > icsk->icsk_rto)
706 icsk->icsk_ack.ato = icsk->icsk_rto;
707 } else if (m > icsk->icsk_rto) {
708 /* Too long gap. Apparently sender failed to
709 * restart window, so that we send ACKs quickly.
711 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
715 icsk->icsk_ack.lrcvtime = now;
717 tcp_ecn_check_ce(sk, skb);
720 tcp_grow_window(sk, skb);
723 /* Called to compute a smoothed rtt estimate. The data fed to this
724 * routine either comes from timestamps, or from segments that were
725 * known _not_ to have been retransmitted [see Karn/Partridge
726 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
727 * piece by Van Jacobson.
728 * NOTE: the next three routines used to be one big routine.
729 * To save cycles in the RFC 1323 implementation it was better to break
730 * it up into three procedures. -- erics
732 static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
734 struct tcp_sock *tp = tcp_sk(sk);
735 long m = mrtt_us; /* RTT */
736 u32 srtt = tp->srtt_us;
738 /* The following amusing code comes from Jacobson's
739 * article in SIGCOMM '88. Note that rtt and mdev
740 * are scaled versions of rtt and mean deviation.
741 * This is designed to be as fast as possible
742 * m stands for "measurement".
744 * On a 1990 paper the rto value is changed to:
745 * RTO = rtt + 4 * mdev
747 * Funny. This algorithm seems to be very broken.
748 * These formulae increase RTO, when it should be decreased, increase
749 * too slowly, when it should be increased quickly, decrease too quickly
750 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
751 * does not matter how to _calculate_ it. Seems, it was trap
752 * that VJ failed to avoid. 8)
755 m -= (srtt >> 3); /* m is now error in rtt est */
756 srtt += m; /* rtt = 7/8 rtt + 1/8 new */
758 m = -m; /* m is now abs(error) */
759 m -= (tp->mdev_us >> 2); /* similar update on mdev */
760 /* This is similar to one of Eifel findings.
761 * Eifel blocks mdev updates when rtt decreases.
762 * This solution is a bit different: we use finer gain
763 * for mdev in this case (alpha*beta).
764 * Like Eifel it also prevents growth of rto,
765 * but also it limits too fast rto decreases,
766 * happening in pure Eifel.
771 m -= (tp->mdev_us >> 2); /* similar update on mdev */
773 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */
774 if (tp->mdev_us > tp->mdev_max_us) {
775 tp->mdev_max_us = tp->mdev_us;
776 if (tp->mdev_max_us > tp->rttvar_us)
777 tp->rttvar_us = tp->mdev_max_us;
779 if (after(tp->snd_una, tp->rtt_seq)) {
780 if (tp->mdev_max_us < tp->rttvar_us)
781 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
782 tp->rtt_seq = tp->snd_nxt;
783 tp->mdev_max_us = tcp_rto_min_us(sk);
788 /* no previous measure. */
789 srtt = m << 3; /* take the measured time to be rtt */
790 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */
791 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
792 tp->mdev_max_us = tp->rttvar_us;
793 tp->rtt_seq = tp->snd_nxt;
797 tp->srtt_us = max(1U, srtt);
800 static void tcp_update_pacing_rate(struct sock *sk)
802 const struct tcp_sock *tp = tcp_sk(sk);
805 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
806 rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3);
808 /* current rate is (cwnd * mss) / srtt
809 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
810 * In Congestion Avoidance phase, set it to 120 % the current rate.
812 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
813 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
814 * end of slow start and should slow down.
816 if (tp->snd_cwnd < tp->snd_ssthresh / 2)
817 rate *= sock_net(sk)->ipv4.sysctl_tcp_pacing_ss_ratio;
819 rate *= sock_net(sk)->ipv4.sysctl_tcp_pacing_ca_ratio;
821 rate *= max(tp->snd_cwnd, tp->packets_out);
823 if (likely(tp->srtt_us))
824 do_div(rate, tp->srtt_us);
826 /* WRITE_ONCE() is needed because sch_fq fetches sk_pacing_rate
827 * without any lock. We want to make sure compiler wont store
828 * intermediate values in this location.
830 WRITE_ONCE(sk->sk_pacing_rate, min_t(u64, rate,
831 sk->sk_max_pacing_rate));
834 /* Calculate rto without backoff. This is the second half of Van Jacobson's
835 * routine referred to above.
837 static void tcp_set_rto(struct sock *sk)
839 const struct tcp_sock *tp = tcp_sk(sk);
840 /* Old crap is replaced with new one. 8)
843 * 1. If rtt variance happened to be less 50msec, it is hallucination.
844 * It cannot be less due to utterly erratic ACK generation made
845 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
846 * to do with delayed acks, because at cwnd>2 true delack timeout
847 * is invisible. Actually, Linux-2.4 also generates erratic
848 * ACKs in some circumstances.
850 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
852 /* 2. Fixups made earlier cannot be right.
853 * If we do not estimate RTO correctly without them,
854 * all the algo is pure shit and should be replaced
855 * with correct one. It is exactly, which we pretend to do.
858 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
859 * guarantees that rto is higher.
864 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
866 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
869 cwnd = TCP_INIT_CWND;
870 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
873 /* Take a notice that peer is sending D-SACKs */
874 static void tcp_dsack_seen(struct tcp_sock *tp)
876 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
877 tp->rack.dsack_seen = 1;
881 /* It's reordering when higher sequence was delivered (i.e. sacked) before
882 * some lower never-retransmitted sequence ("low_seq"). The maximum reordering
883 * distance is approximated in full-mss packet distance ("reordering").
885 static void tcp_check_sack_reordering(struct sock *sk, const u32 low_seq,
888 struct tcp_sock *tp = tcp_sk(sk);
889 const u32 mss = tp->mss_cache;
892 fack = tcp_highest_sack_seq(tp);
893 if (!before(low_seq, fack))
896 metric = fack - low_seq;
897 if ((metric > tp->reordering * mss) && mss) {
898 #if FASTRETRANS_DEBUG > 1
899 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
900 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
904 tp->undo_marker ? tp->undo_retrans : 0);
906 tp->reordering = min_t(u32, (metric + mss - 1) / mss,
907 sock_net(sk)->ipv4.sysctl_tcp_max_reordering);
910 /* This exciting event is worth to be remembered. 8) */
912 NET_INC_STATS(sock_net(sk),
913 ts ? LINUX_MIB_TCPTSREORDER : LINUX_MIB_TCPSACKREORDER);
916 /* This must be called before lost_out is incremented */
917 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
919 if ((!tp->retransmit_skb_hint && tp->retrans_out >= tp->lost_out) ||
920 (tp->retransmit_skb_hint &&
921 before(TCP_SKB_CB(skb)->seq,
922 TCP_SKB_CB(tp->retransmit_skb_hint)->seq)))
923 tp->retransmit_skb_hint = skb;
926 /* Sum the number of packets on the wire we have marked as lost.
927 * There are two cases we care about here:
928 * a) Packet hasn't been marked lost (nor retransmitted),
929 * and this is the first loss.
930 * b) Packet has been marked both lost and retransmitted,
931 * and this means we think it was lost again.
933 static void tcp_sum_lost(struct tcp_sock *tp, struct sk_buff *skb)
935 __u8 sacked = TCP_SKB_CB(skb)->sacked;
937 if (!(sacked & TCPCB_LOST) ||
938 ((sacked & TCPCB_LOST) && (sacked & TCPCB_SACKED_RETRANS)))
939 tp->lost += tcp_skb_pcount(skb);
942 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
944 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
945 tcp_verify_retransmit_hint(tp, skb);
947 tp->lost_out += tcp_skb_pcount(skb);
948 tcp_sum_lost(tp, skb);
949 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
953 void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp, struct sk_buff *skb)
955 tcp_verify_retransmit_hint(tp, skb);
957 tcp_sum_lost(tp, skb);
958 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
959 tp->lost_out += tcp_skb_pcount(skb);
960 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
964 /* This procedure tags the retransmission queue when SACKs arrive.
966 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
967 * Packets in queue with these bits set are counted in variables
968 * sacked_out, retrans_out and lost_out, correspondingly.
970 * Valid combinations are:
971 * Tag InFlight Description
972 * 0 1 - orig segment is in flight.
973 * S 0 - nothing flies, orig reached receiver.
974 * L 0 - nothing flies, orig lost by net.
975 * R 2 - both orig and retransmit are in flight.
976 * L|R 1 - orig is lost, retransmit is in flight.
977 * S|R 1 - orig reached receiver, retrans is still in flight.
978 * (L|S|R is logically valid, it could occur when L|R is sacked,
979 * but it is equivalent to plain S and code short-curcuits it to S.
980 * L|S is logically invalid, it would mean -1 packet in flight 8))
982 * These 6 states form finite state machine, controlled by the following events:
983 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
984 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
985 * 3. Loss detection event of two flavors:
986 * A. Scoreboard estimator decided the packet is lost.
987 * A'. Reno "three dupacks" marks head of queue lost.
988 * B. SACK arrives sacking SND.NXT at the moment, when the
989 * segment was retransmitted.
990 * 4. D-SACK added new rule: D-SACK changes any tag to S.
992 * It is pleasant to note, that state diagram turns out to be commutative,
993 * so that we are allowed not to be bothered by order of our actions,
994 * when multiple events arrive simultaneously. (see the function below).
996 * Reordering detection.
997 * --------------------
998 * Reordering metric is maximal distance, which a packet can be displaced
999 * in packet stream. With SACKs we can estimate it:
1001 * 1. SACK fills old hole and the corresponding segment was not
1002 * ever retransmitted -> reordering. Alas, we cannot use it
1003 * when segment was retransmitted.
1004 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1005 * for retransmitted and already SACKed segment -> reordering..
1006 * Both of these heuristics are not used in Loss state, when we cannot
1007 * account for retransmits accurately.
1009 * SACK block validation.
1010 * ----------------------
1012 * SACK block range validation checks that the received SACK block fits to
1013 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1014 * Note that SND.UNA is not included to the range though being valid because
1015 * it means that the receiver is rather inconsistent with itself reporting
1016 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1017 * perfectly valid, however, in light of RFC2018 which explicitly states
1018 * that "SACK block MUST reflect the newest segment. Even if the newest
1019 * segment is going to be discarded ...", not that it looks very clever
1020 * in case of head skb. Due to potentional receiver driven attacks, we
1021 * choose to avoid immediate execution of a walk in write queue due to
1022 * reneging and defer head skb's loss recovery to standard loss recovery
1023 * procedure that will eventually trigger (nothing forbids us doing this).
1025 * Implements also blockage to start_seq wrap-around. Problem lies in the
1026 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1027 * there's no guarantee that it will be before snd_nxt (n). The problem
1028 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1031 * <- outs wnd -> <- wrapzone ->
1032 * u e n u_w e_w s n_w
1034 * |<------------+------+----- TCP seqno space --------------+---------->|
1035 * ...-- <2^31 ->| |<--------...
1036 * ...---- >2^31 ------>| |<--------...
1038 * Current code wouldn't be vulnerable but it's better still to discard such
1039 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1040 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1041 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1042 * equal to the ideal case (infinite seqno space without wrap caused issues).
1044 * With D-SACK the lower bound is extended to cover sequence space below
1045 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1046 * again, D-SACK block must not to go across snd_una (for the same reason as
1047 * for the normal SACK blocks, explained above). But there all simplicity
1048 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1049 * fully below undo_marker they do not affect behavior in anyway and can
1050 * therefore be safely ignored. In rare cases (which are more or less
1051 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1052 * fragmentation and packet reordering past skb's retransmission. To consider
1053 * them correctly, the acceptable range must be extended even more though
1054 * the exact amount is rather hard to quantify. However, tp->max_window can
1055 * be used as an exaggerated estimate.
1057 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
1058 u32 start_seq, u32 end_seq)
1060 /* Too far in future, or reversed (interpretation is ambiguous) */
1061 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1064 /* Nasty start_seq wrap-around check (see comments above) */
1065 if (!before(start_seq, tp->snd_nxt))
1068 /* In outstanding window? ...This is valid exit for D-SACKs too.
1069 * start_seq == snd_una is non-sensical (see comments above)
1071 if (after(start_seq, tp->snd_una))
1074 if (!is_dsack || !tp->undo_marker)
1077 /* ...Then it's D-SACK, and must reside below snd_una completely */
1078 if (after(end_seq, tp->snd_una))
1081 if (!before(start_seq, tp->undo_marker))
1085 if (!after(end_seq, tp->undo_marker))
1088 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1089 * start_seq < undo_marker and end_seq >= undo_marker.
1091 return !before(start_seq, end_seq - tp->max_window);
1094 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1095 struct tcp_sack_block_wire *sp, int num_sacks,
1098 struct tcp_sock *tp = tcp_sk(sk);
1099 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1100 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1101 bool dup_sack = false;
1103 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1106 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1107 } else if (num_sacks > 1) {
1108 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1109 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1111 if (!after(end_seq_0, end_seq_1) &&
1112 !before(start_seq_0, start_seq_1)) {
1115 NET_INC_STATS(sock_net(sk),
1116 LINUX_MIB_TCPDSACKOFORECV);
1120 /* D-SACK for already forgotten data... Do dumb counting. */
1121 if (dup_sack && tp->undo_marker && tp->undo_retrans > 0 &&
1122 !after(end_seq_0, prior_snd_una) &&
1123 after(end_seq_0, tp->undo_marker))
1129 struct tcp_sacktag_state {
1131 /* Timestamps for earliest and latest never-retransmitted segment
1132 * that was SACKed. RTO needs the earliest RTT to stay conservative,
1133 * but congestion control should still get an accurate delay signal.
1137 struct rate_sample *rate;
1139 unsigned int mss_now;
1142 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1143 * the incoming SACK may not exactly match but we can find smaller MSS
1144 * aligned portion of it that matches. Therefore we might need to fragment
1145 * which may fail and creates some hassle (caller must handle error case
1148 * FIXME: this could be merged to shift decision code
1150 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1151 u32 start_seq, u32 end_seq)
1155 unsigned int pkt_len;
1158 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1159 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1161 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1162 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1163 mss = tcp_skb_mss(skb);
1164 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1167 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1171 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1176 /* Round if necessary so that SACKs cover only full MSSes
1177 * and/or the remaining small portion (if present)
1179 if (pkt_len > mss) {
1180 unsigned int new_len = (pkt_len / mss) * mss;
1181 if (!in_sack && new_len < pkt_len)
1186 if (pkt_len >= skb->len && !in_sack)
1189 err = tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
1190 pkt_len, mss, GFP_ATOMIC);
1198 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1199 static u8 tcp_sacktag_one(struct sock *sk,
1200 struct tcp_sacktag_state *state, u8 sacked,
1201 u32 start_seq, u32 end_seq,
1202 int dup_sack, int pcount,
1205 struct tcp_sock *tp = tcp_sk(sk);
1207 /* Account D-SACK for retransmitted packet. */
1208 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1209 if (tp->undo_marker && tp->undo_retrans > 0 &&
1210 after(end_seq, tp->undo_marker))
1212 if ((sacked & TCPCB_SACKED_ACKED) &&
1213 before(start_seq, state->reord))
1214 state->reord = start_seq;
1217 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1218 if (!after(end_seq, tp->snd_una))
1221 if (!(sacked & TCPCB_SACKED_ACKED)) {
1222 tcp_rack_advance(tp, sacked, end_seq, xmit_time);
1224 if (sacked & TCPCB_SACKED_RETRANS) {
1225 /* If the segment is not tagged as lost,
1226 * we do not clear RETRANS, believing
1227 * that retransmission is still in flight.
1229 if (sacked & TCPCB_LOST) {
1230 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1231 tp->lost_out -= pcount;
1232 tp->retrans_out -= pcount;
1235 if (!(sacked & TCPCB_RETRANS)) {
1236 /* New sack for not retransmitted frame,
1237 * which was in hole. It is reordering.
1239 if (before(start_seq,
1240 tcp_highest_sack_seq(tp)) &&
1241 before(start_seq, state->reord))
1242 state->reord = start_seq;
1244 if (!after(end_seq, tp->high_seq))
1245 state->flag |= FLAG_ORIG_SACK_ACKED;
1246 if (state->first_sackt == 0)
1247 state->first_sackt = xmit_time;
1248 state->last_sackt = xmit_time;
1251 if (sacked & TCPCB_LOST) {
1252 sacked &= ~TCPCB_LOST;
1253 tp->lost_out -= pcount;
1257 sacked |= TCPCB_SACKED_ACKED;
1258 state->flag |= FLAG_DATA_SACKED;
1259 tp->sacked_out += pcount;
1260 tp->delivered += pcount; /* Out-of-order packets delivered */
1262 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1263 if (tp->lost_skb_hint &&
1264 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1265 tp->lost_cnt_hint += pcount;
1268 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1269 * frames and clear it. undo_retrans is decreased above, L|R frames
1270 * are accounted above as well.
1272 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1273 sacked &= ~TCPCB_SACKED_RETRANS;
1274 tp->retrans_out -= pcount;
1280 /* Shift newly-SACKed bytes from this skb to the immediately previous
1281 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1283 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *prev,
1284 struct sk_buff *skb,
1285 struct tcp_sacktag_state *state,
1286 unsigned int pcount, int shifted, int mss,
1289 struct tcp_sock *tp = tcp_sk(sk);
1290 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1291 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
1295 /* Adjust counters and hints for the newly sacked sequence
1296 * range but discard the return value since prev is already
1297 * marked. We must tag the range first because the seq
1298 * advancement below implicitly advances
1299 * tcp_highest_sack_seq() when skb is highest_sack.
1301 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1302 start_seq, end_seq, dup_sack, pcount,
1303 tcp_skb_timestamp_us(skb));
1304 tcp_rate_skb_delivered(sk, skb, state->rate);
1306 if (skb == tp->lost_skb_hint)
1307 tp->lost_cnt_hint += pcount;
1309 TCP_SKB_CB(prev)->end_seq += shifted;
1310 TCP_SKB_CB(skb)->seq += shifted;
1312 tcp_skb_pcount_add(prev, pcount);
1313 WARN_ON_ONCE(tcp_skb_pcount(skb) < pcount);
1314 tcp_skb_pcount_add(skb, -pcount);
1316 /* When we're adding to gso_segs == 1, gso_size will be zero,
1317 * in theory this shouldn't be necessary but as long as DSACK
1318 * code can come after this skb later on it's better to keep
1319 * setting gso_size to something.
1321 if (!TCP_SKB_CB(prev)->tcp_gso_size)
1322 TCP_SKB_CB(prev)->tcp_gso_size = mss;
1324 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1325 if (tcp_skb_pcount(skb) <= 1)
1326 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1328 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1329 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1332 BUG_ON(!tcp_skb_pcount(skb));
1333 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1337 /* Whole SKB was eaten :-) */
1339 if (skb == tp->retransmit_skb_hint)
1340 tp->retransmit_skb_hint = prev;
1341 if (skb == tp->lost_skb_hint) {
1342 tp->lost_skb_hint = prev;
1343 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1346 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1347 TCP_SKB_CB(prev)->eor = TCP_SKB_CB(skb)->eor;
1348 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1349 TCP_SKB_CB(prev)->end_seq++;
1351 if (skb == tcp_highest_sack(sk))
1352 tcp_advance_highest_sack(sk, skb);
1354 tcp_skb_collapse_tstamp(prev, skb);
1355 if (unlikely(TCP_SKB_CB(prev)->tx.delivered_mstamp))
1356 TCP_SKB_CB(prev)->tx.delivered_mstamp = 0;
1358 tcp_rtx_queue_unlink_and_free(skb, sk);
1360 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKMERGED);
1365 /* I wish gso_size would have a bit more sane initialization than
1366 * something-or-zero which complicates things
1368 static int tcp_skb_seglen(const struct sk_buff *skb)
1370 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1373 /* Shifting pages past head area doesn't work */
1374 static int skb_can_shift(const struct sk_buff *skb)
1376 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1379 int tcp_skb_shift(struct sk_buff *to, struct sk_buff *from,
1380 int pcount, int shiftlen)
1382 /* TCP min gso_size is 8 bytes (TCP_MIN_GSO_SIZE)
1383 * Since TCP_SKB_CB(skb)->tcp_gso_segs is 16 bits, we need
1384 * to make sure not storing more than 65535 * 8 bytes per skb,
1385 * even if current MSS is bigger.
1387 if (unlikely(to->len + shiftlen >= 65535 * TCP_MIN_GSO_SIZE))
1389 if (unlikely(tcp_skb_pcount(to) + pcount > 65535))
1391 return skb_shift(to, from, shiftlen);
1394 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1397 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1398 struct tcp_sacktag_state *state,
1399 u32 start_seq, u32 end_seq,
1402 struct tcp_sock *tp = tcp_sk(sk);
1403 struct sk_buff *prev;
1409 /* Normally R but no L won't result in plain S */
1411 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1413 if (!skb_can_shift(skb))
1415 /* This frame is about to be dropped (was ACKed). */
1416 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1419 /* Can only happen with delayed DSACK + discard craziness */
1420 prev = skb_rb_prev(skb);
1424 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1427 if (!tcp_skb_can_collapse(prev, skb))
1430 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1431 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1435 pcount = tcp_skb_pcount(skb);
1436 mss = tcp_skb_seglen(skb);
1438 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1439 * drop this restriction as unnecessary
1441 if (mss != tcp_skb_seglen(prev))
1444 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1446 /* CHECKME: This is non-MSS split case only?, this will
1447 * cause skipped skbs due to advancing loop btw, original
1448 * has that feature too
1450 if (tcp_skb_pcount(skb) <= 1)
1453 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1455 /* TODO: head merge to next could be attempted here
1456 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1457 * though it might not be worth of the additional hassle
1459 * ...we can probably just fallback to what was done
1460 * previously. We could try merging non-SACKed ones
1461 * as well but it probably isn't going to buy off
1462 * because later SACKs might again split them, and
1463 * it would make skb timestamp tracking considerably
1469 len = end_seq - TCP_SKB_CB(skb)->seq;
1471 BUG_ON(len > skb->len);
1473 /* MSS boundaries should be honoured or else pcount will
1474 * severely break even though it makes things bit trickier.
1475 * Optimize common case to avoid most of the divides
1477 mss = tcp_skb_mss(skb);
1479 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1480 * drop this restriction as unnecessary
1482 if (mss != tcp_skb_seglen(prev))
1487 } else if (len < mss) {
1495 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1496 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1499 if (!tcp_skb_shift(prev, skb, pcount, len))
1501 if (!tcp_shifted_skb(sk, prev, skb, state, pcount, len, mss, dup_sack))
1504 /* Hole filled allows collapsing with the next as well, this is very
1505 * useful when hole on every nth skb pattern happens
1507 skb = skb_rb_next(prev);
1511 if (!skb_can_shift(skb) ||
1512 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1513 (mss != tcp_skb_seglen(skb)))
1517 pcount = tcp_skb_pcount(skb);
1518 if (tcp_skb_shift(prev, skb, pcount, len))
1519 tcp_shifted_skb(sk, prev, skb, state, pcount,
1529 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1533 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1534 struct tcp_sack_block *next_dup,
1535 struct tcp_sacktag_state *state,
1536 u32 start_seq, u32 end_seq,
1539 struct tcp_sock *tp = tcp_sk(sk);
1540 struct sk_buff *tmp;
1542 skb_rbtree_walk_from(skb) {
1544 bool dup_sack = dup_sack_in;
1546 /* queue is in-order => we can short-circuit the walk early */
1547 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1551 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1552 in_sack = tcp_match_skb_to_sack(sk, skb,
1553 next_dup->start_seq,
1559 /* skb reference here is a bit tricky to get right, since
1560 * shifting can eat and free both this skb and the next,
1561 * so not even _safe variant of the loop is enough.
1564 tmp = tcp_shift_skb_data(sk, skb, state,
1565 start_seq, end_seq, dup_sack);
1574 in_sack = tcp_match_skb_to_sack(sk, skb,
1580 if (unlikely(in_sack < 0))
1584 TCP_SKB_CB(skb)->sacked =
1587 TCP_SKB_CB(skb)->sacked,
1588 TCP_SKB_CB(skb)->seq,
1589 TCP_SKB_CB(skb)->end_seq,
1591 tcp_skb_pcount(skb),
1592 tcp_skb_timestamp_us(skb));
1593 tcp_rate_skb_delivered(sk, skb, state->rate);
1594 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1595 list_del_init(&skb->tcp_tsorted_anchor);
1597 if (!before(TCP_SKB_CB(skb)->seq,
1598 tcp_highest_sack_seq(tp)))
1599 tcp_advance_highest_sack(sk, skb);
1605 static struct sk_buff *tcp_sacktag_bsearch(struct sock *sk, u32 seq)
1607 struct rb_node *parent, **p = &sk->tcp_rtx_queue.rb_node;
1608 struct sk_buff *skb;
1612 skb = rb_to_skb(parent);
1613 if (before(seq, TCP_SKB_CB(skb)->seq)) {
1614 p = &parent->rb_left;
1617 if (!before(seq, TCP_SKB_CB(skb)->end_seq)) {
1618 p = &parent->rb_right;
1626 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1629 if (skb && after(TCP_SKB_CB(skb)->seq, skip_to_seq))
1632 return tcp_sacktag_bsearch(sk, skip_to_seq);
1635 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1637 struct tcp_sack_block *next_dup,
1638 struct tcp_sacktag_state *state,
1644 if (before(next_dup->start_seq, skip_to_seq)) {
1645 skb = tcp_sacktag_skip(skb, sk, next_dup->start_seq);
1646 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1647 next_dup->start_seq, next_dup->end_seq,
1654 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1656 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1660 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1661 u32 prior_snd_una, struct tcp_sacktag_state *state)
1663 struct tcp_sock *tp = tcp_sk(sk);
1664 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1665 TCP_SKB_CB(ack_skb)->sacked);
1666 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1667 struct tcp_sack_block sp[TCP_NUM_SACKS];
1668 struct tcp_sack_block *cache;
1669 struct sk_buff *skb;
1670 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1672 bool found_dup_sack = false;
1674 int first_sack_index;
1677 state->reord = tp->snd_nxt;
1679 if (!tp->sacked_out)
1680 tcp_highest_sack_reset(sk);
1682 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1683 num_sacks, prior_snd_una);
1684 if (found_dup_sack) {
1685 state->flag |= FLAG_DSACKING_ACK;
1686 tp->delivered++; /* A spurious retransmission is delivered */
1689 /* Eliminate too old ACKs, but take into
1690 * account more or less fresh ones, they can
1691 * contain valid SACK info.
1693 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1696 if (!tp->packets_out)
1700 first_sack_index = 0;
1701 for (i = 0; i < num_sacks; i++) {
1702 bool dup_sack = !i && found_dup_sack;
1704 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1705 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1707 if (!tcp_is_sackblock_valid(tp, dup_sack,
1708 sp[used_sacks].start_seq,
1709 sp[used_sacks].end_seq)) {
1713 if (!tp->undo_marker)
1714 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1716 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1718 /* Don't count olds caused by ACK reordering */
1719 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1720 !after(sp[used_sacks].end_seq, tp->snd_una))
1722 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1725 NET_INC_STATS(sock_net(sk), mib_idx);
1727 first_sack_index = -1;
1731 /* Ignore very old stuff early */
1732 if (!after(sp[used_sacks].end_seq, prior_snd_una)) {
1734 first_sack_index = -1;
1741 /* order SACK blocks to allow in order walk of the retrans queue */
1742 for (i = used_sacks - 1; i > 0; i--) {
1743 for (j = 0; j < i; j++) {
1744 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1745 swap(sp[j], sp[j + 1]);
1747 /* Track where the first SACK block goes to */
1748 if (j == first_sack_index)
1749 first_sack_index = j + 1;
1754 state->mss_now = tcp_current_mss(sk);
1758 if (!tp->sacked_out) {
1759 /* It's already past, so skip checking against it */
1760 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1762 cache = tp->recv_sack_cache;
1763 /* Skip empty blocks in at head of the cache */
1764 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1769 while (i < used_sacks) {
1770 u32 start_seq = sp[i].start_seq;
1771 u32 end_seq = sp[i].end_seq;
1772 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1773 struct tcp_sack_block *next_dup = NULL;
1775 if (found_dup_sack && ((i + 1) == first_sack_index))
1776 next_dup = &sp[i + 1];
1778 /* Skip too early cached blocks */
1779 while (tcp_sack_cache_ok(tp, cache) &&
1780 !before(start_seq, cache->end_seq))
1783 /* Can skip some work by looking recv_sack_cache? */
1784 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1785 after(end_seq, cache->start_seq)) {
1788 if (before(start_seq, cache->start_seq)) {
1789 skb = tcp_sacktag_skip(skb, sk, start_seq);
1790 skb = tcp_sacktag_walk(skb, sk, next_dup,
1797 /* Rest of the block already fully processed? */
1798 if (!after(end_seq, cache->end_seq))
1801 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1805 /* ...tail remains todo... */
1806 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1807 /* ...but better entrypoint exists! */
1808 skb = tcp_highest_sack(sk);
1815 skb = tcp_sacktag_skip(skb, sk, cache->end_seq);
1816 /* Check overlap against next cached too (past this one already) */
1821 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1822 skb = tcp_highest_sack(sk);
1826 skb = tcp_sacktag_skip(skb, sk, start_seq);
1829 skb = tcp_sacktag_walk(skb, sk, next_dup, state,
1830 start_seq, end_seq, dup_sack);
1836 /* Clear the head of the cache sack blocks so we can skip it next time */
1837 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1838 tp->recv_sack_cache[i].start_seq = 0;
1839 tp->recv_sack_cache[i].end_seq = 0;
1841 for (j = 0; j < used_sacks; j++)
1842 tp->recv_sack_cache[i++] = sp[j];
1844 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Loss || tp->undo_marker)
1845 tcp_check_sack_reordering(sk, state->reord, 0);
1847 tcp_verify_left_out(tp);
1850 #if FASTRETRANS_DEBUG > 0
1851 WARN_ON((int)tp->sacked_out < 0);
1852 WARN_ON((int)tp->lost_out < 0);
1853 WARN_ON((int)tp->retrans_out < 0);
1854 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1859 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1860 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1862 static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
1866 holes = max(tp->lost_out, 1U);
1867 holes = min(holes, tp->packets_out);
1869 if ((tp->sacked_out + holes) > tp->packets_out) {
1870 tp->sacked_out = tp->packets_out - holes;
1876 /* If we receive more dupacks than we expected counting segments
1877 * in assumption of absent reordering, interpret this as reordering.
1878 * The only another reason could be bug in receiver TCP.
1880 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1882 struct tcp_sock *tp = tcp_sk(sk);
1884 if (!tcp_limit_reno_sacked(tp))
1887 tp->reordering = min_t(u32, tp->packets_out + addend,
1888 sock_net(sk)->ipv4.sysctl_tcp_max_reordering);
1890 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRENOREORDER);
1893 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1895 static void tcp_add_reno_sack(struct sock *sk, int num_dupack)
1898 struct tcp_sock *tp = tcp_sk(sk);
1899 u32 prior_sacked = tp->sacked_out;
1902 tp->sacked_out += num_dupack;
1903 tcp_check_reno_reordering(sk, 0);
1904 delivered = tp->sacked_out - prior_sacked;
1906 tp->delivered += delivered;
1907 tcp_verify_left_out(tp);
1911 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1913 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1915 struct tcp_sock *tp = tcp_sk(sk);
1918 /* One ACK acked hole. The rest eat duplicate ACKs. */
1919 tp->delivered += max_t(int, acked - tp->sacked_out, 1);
1920 if (acked - 1 >= tp->sacked_out)
1923 tp->sacked_out -= acked - 1;
1925 tcp_check_reno_reordering(sk, acked);
1926 tcp_verify_left_out(tp);
1929 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1934 void tcp_clear_retrans(struct tcp_sock *tp)
1936 tp->retrans_out = 0;
1938 tp->undo_marker = 0;
1939 tp->undo_retrans = -1;
1943 static inline void tcp_init_undo(struct tcp_sock *tp)
1945 tp->undo_marker = tp->snd_una;
1946 /* Retransmission still in flight may cause DSACKs later. */
1947 tp->undo_retrans = tp->retrans_out ? : -1;
1950 static bool tcp_is_rack(const struct sock *sk)
1952 return sock_net(sk)->ipv4.sysctl_tcp_recovery & TCP_RACK_LOSS_DETECTION;
1955 /* If we detect SACK reneging, forget all SACK information
1956 * and reset tags completely, otherwise preserve SACKs. If receiver
1957 * dropped its ofo queue, we will know this due to reneging detection.
1959 static void tcp_timeout_mark_lost(struct sock *sk)
1961 struct tcp_sock *tp = tcp_sk(sk);
1962 struct sk_buff *skb, *head;
1963 bool is_reneg; /* is receiver reneging on SACKs? */
1965 head = tcp_rtx_queue_head(sk);
1966 is_reneg = head && (TCP_SKB_CB(head)->sacked & TCPCB_SACKED_ACKED);
1968 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
1970 /* Mark SACK reneging until we recover from this loss event. */
1971 tp->is_sack_reneg = 1;
1972 } else if (tcp_is_reno(tp)) {
1973 tcp_reset_reno_sack(tp);
1977 skb_rbtree_walk_from(skb) {
1979 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1980 else if (tcp_is_rack(sk) && skb != head &&
1981 tcp_rack_skb_timeout(tp, skb, 0) > 0)
1982 continue; /* Don't mark recently sent ones lost yet */
1983 tcp_mark_skb_lost(sk, skb);
1985 tcp_verify_left_out(tp);
1986 tcp_clear_all_retrans_hints(tp);
1989 /* Enter Loss state. */
1990 void tcp_enter_loss(struct sock *sk)
1992 const struct inet_connection_sock *icsk = inet_csk(sk);
1993 struct tcp_sock *tp = tcp_sk(sk);
1994 struct net *net = sock_net(sk);
1995 bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery;
1997 tcp_timeout_mark_lost(sk);
1999 /* Reduce ssthresh if it has not yet been made inside this window. */
2000 if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
2001 !after(tp->high_seq, tp->snd_una) ||
2002 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
2003 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2004 tp->prior_cwnd = tp->snd_cwnd;
2005 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2006 tcp_ca_event(sk, CA_EVENT_LOSS);
2009 tp->snd_cwnd = tcp_packets_in_flight(tp) + 1;
2010 tp->snd_cwnd_cnt = 0;
2011 tp->snd_cwnd_stamp = tcp_jiffies32;
2013 /* Timeout in disordered state after receiving substantial DUPACKs
2014 * suggests that the degree of reordering is over-estimated.
2016 if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
2017 tp->sacked_out >= net->ipv4.sysctl_tcp_reordering)
2018 tp->reordering = min_t(unsigned int, tp->reordering,
2019 net->ipv4.sysctl_tcp_reordering);
2020 tcp_set_ca_state(sk, TCP_CA_Loss);
2021 tp->high_seq = tp->snd_nxt;
2022 tcp_ecn_queue_cwr(tp);
2024 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
2025 * loss recovery is underway except recurring timeout(s) on
2026 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
2028 tp->frto = net->ipv4.sysctl_tcp_frto &&
2029 (new_recovery || icsk->icsk_retransmits) &&
2030 !inet_csk(sk)->icsk_mtup.probe_size;
2033 /* If ACK arrived pointing to a remembered SACK, it means that our
2034 * remembered SACKs do not reflect real state of receiver i.e.
2035 * receiver _host_ is heavily congested (or buggy).
2037 * To avoid big spurious retransmission bursts due to transient SACK
2038 * scoreboard oddities that look like reneging, we give the receiver a
2039 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
2040 * restore sanity to the SACK scoreboard. If the apparent reneging
2041 * persists until this RTO then we'll clear the SACK scoreboard.
2043 static bool tcp_check_sack_reneging(struct sock *sk, int flag)
2045 if (flag & FLAG_SACK_RENEGING) {
2046 struct tcp_sock *tp = tcp_sk(sk);
2047 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4),
2048 msecs_to_jiffies(10));
2050 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2051 delay, TCP_RTO_MAX);
2057 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2058 * counter when SACK is enabled (without SACK, sacked_out is used for
2061 * With reordering, holes may still be in flight, so RFC3517 recovery
2062 * uses pure sacked_out (total number of SACKed segments) even though
2063 * it violates the RFC that uses duplicate ACKs, often these are equal
2064 * but when e.g. out-of-window ACKs or packet duplication occurs,
2065 * they differ. Since neither occurs due to loss, TCP should really
2068 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2070 return tp->sacked_out + 1;
2073 /* Linux NewReno/SACK/ECN state machine.
2074 * --------------------------------------
2076 * "Open" Normal state, no dubious events, fast path.
2077 * "Disorder" In all the respects it is "Open",
2078 * but requires a bit more attention. It is entered when
2079 * we see some SACKs or dupacks. It is split of "Open"
2080 * mainly to move some processing from fast path to slow one.
2081 * "CWR" CWND was reduced due to some Congestion Notification event.
2082 * It can be ECN, ICMP source quench, local device congestion.
2083 * "Recovery" CWND was reduced, we are fast-retransmitting.
2084 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2086 * tcp_fastretrans_alert() is entered:
2087 * - each incoming ACK, if state is not "Open"
2088 * - when arrived ACK is unusual, namely:
2093 * Counting packets in flight is pretty simple.
2095 * in_flight = packets_out - left_out + retrans_out
2097 * packets_out is SND.NXT-SND.UNA counted in packets.
2099 * retrans_out is number of retransmitted segments.
2101 * left_out is number of segments left network, but not ACKed yet.
2103 * left_out = sacked_out + lost_out
2105 * sacked_out: Packets, which arrived to receiver out of order
2106 * and hence not ACKed. With SACKs this number is simply
2107 * amount of SACKed data. Even without SACKs
2108 * it is easy to give pretty reliable estimate of this number,
2109 * counting duplicate ACKs.
2111 * lost_out: Packets lost by network. TCP has no explicit
2112 * "loss notification" feedback from network (for now).
2113 * It means that this number can be only _guessed_.
2114 * Actually, it is the heuristics to predict lossage that
2115 * distinguishes different algorithms.
2117 * F.e. after RTO, when all the queue is considered as lost,
2118 * lost_out = packets_out and in_flight = retrans_out.
2120 * Essentially, we have now a few algorithms detecting
2123 * If the receiver supports SACK:
2125 * RFC6675/3517: It is the conventional algorithm. A packet is
2126 * considered lost if the number of higher sequence packets
2127 * SACKed is greater than or equal the DUPACK thoreshold
2128 * (reordering). This is implemented in tcp_mark_head_lost and
2129 * tcp_update_scoreboard.
2131 * RACK (draft-ietf-tcpm-rack-01): it is a newer algorithm
2132 * (2017-) that checks timing instead of counting DUPACKs.
2133 * Essentially a packet is considered lost if it's not S/ACKed
2134 * after RTT + reordering_window, where both metrics are
2135 * dynamically measured and adjusted. This is implemented in
2136 * tcp_rack_mark_lost.
2138 * If the receiver does not support SACK:
2140 * NewReno (RFC6582): in Recovery we assume that one segment
2141 * is lost (classic Reno). While we are in Recovery and
2142 * a partial ACK arrives, we assume that one more packet
2143 * is lost (NewReno). This heuristics are the same in NewReno
2146 * Really tricky (and requiring careful tuning) part of algorithm
2147 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2148 * The first determines the moment _when_ we should reduce CWND and,
2149 * hence, slow down forward transmission. In fact, it determines the moment
2150 * when we decide that hole is caused by loss, rather than by a reorder.
2152 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2153 * holes, caused by lost packets.
2155 * And the most logically complicated part of algorithm is undo
2156 * heuristics. We detect false retransmits due to both too early
2157 * fast retransmit (reordering) and underestimated RTO, analyzing
2158 * timestamps and D-SACKs. When we detect that some segments were
2159 * retransmitted by mistake and CWND reduction was wrong, we undo
2160 * window reduction and abort recovery phase. This logic is hidden
2161 * inside several functions named tcp_try_undo_<something>.
2164 /* This function decides, when we should leave Disordered state
2165 * and enter Recovery phase, reducing congestion window.
2167 * Main question: may we further continue forward transmission
2168 * with the same cwnd?
2170 static bool tcp_time_to_recover(struct sock *sk, int flag)
2172 struct tcp_sock *tp = tcp_sk(sk);
2174 /* Trick#1: The loss is proven. */
2178 /* Not-A-Trick#2 : Classic rule... */
2179 if (!tcp_is_rack(sk) && tcp_dupack_heuristics(tp) > tp->reordering)
2185 /* Detect loss in event "A" above by marking head of queue up as lost.
2186 * For RFC3517 SACK, a segment is considered lost if it
2187 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2188 * the maximum SACKed segments to pass before reaching this limit.
2190 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2192 struct tcp_sock *tp = tcp_sk(sk);
2193 struct sk_buff *skb;
2195 /* Use SACK to deduce losses of new sequences sent during recovery */
2196 const u32 loss_high = tp->snd_nxt;
2198 WARN_ON(packets > tp->packets_out);
2199 skb = tp->lost_skb_hint;
2201 /* Head already handled? */
2202 if (mark_head && after(TCP_SKB_CB(skb)->seq, tp->snd_una))
2204 cnt = tp->lost_cnt_hint;
2206 skb = tcp_rtx_queue_head(sk);
2210 skb_rbtree_walk_from(skb) {
2211 /* TODO: do this better */
2212 /* this is not the most efficient way to do this... */
2213 tp->lost_skb_hint = skb;
2214 tp->lost_cnt_hint = cnt;
2216 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2219 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2220 cnt += tcp_skb_pcount(skb);
2225 tcp_skb_mark_lost(tp, skb);
2230 tcp_verify_left_out(tp);
2233 /* Account newly detected lost packet(s) */
2235 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2237 struct tcp_sock *tp = tcp_sk(sk);
2239 if (tcp_is_sack(tp)) {
2240 int sacked_upto = tp->sacked_out - tp->reordering;
2241 if (sacked_upto >= 0)
2242 tcp_mark_head_lost(sk, sacked_upto, 0);
2243 else if (fast_rexmit)
2244 tcp_mark_head_lost(sk, 1, 1);
2248 static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when)
2250 return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2251 before(tp->rx_opt.rcv_tsecr, when);
2254 /* skb is spurious retransmitted if the returned timestamp echo
2255 * reply is prior to the skb transmission time
2257 static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp,
2258 const struct sk_buff *skb)
2260 return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) &&
2261 tcp_tsopt_ecr_before(tp, tcp_skb_timestamp(skb));
2264 /* Nothing was retransmitted or returned timestamp is less
2265 * than timestamp of the first retransmission.
2267 static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2269 return tp->retrans_stamp &&
2270 tcp_tsopt_ecr_before(tp, tp->retrans_stamp);
2273 /* Undo procedures. */
2275 /* We can clear retrans_stamp when there are no retransmissions in the
2276 * window. It would seem that it is trivially available for us in
2277 * tp->retrans_out, however, that kind of assumptions doesn't consider
2278 * what will happen if errors occur when sending retransmission for the
2279 * second time. ...It could the that such segment has only
2280 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2281 * the head skb is enough except for some reneging corner cases that
2282 * are not worth the effort.
2284 * Main reason for all this complexity is the fact that connection dying
2285 * time now depends on the validity of the retrans_stamp, in particular,
2286 * that successive retransmissions of a segment must not advance
2287 * retrans_stamp under any conditions.
2289 static bool tcp_any_retrans_done(const struct sock *sk)
2291 const struct tcp_sock *tp = tcp_sk(sk);
2292 struct sk_buff *skb;
2294 if (tp->retrans_out)
2297 skb = tcp_rtx_queue_head(sk);
2298 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2304 static void DBGUNDO(struct sock *sk, const char *msg)
2306 #if FASTRETRANS_DEBUG > 1
2307 struct tcp_sock *tp = tcp_sk(sk);
2308 struct inet_sock *inet = inet_sk(sk);
2310 if (sk->sk_family == AF_INET) {
2311 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2313 &inet->inet_daddr, ntohs(inet->inet_dport),
2314 tp->snd_cwnd, tcp_left_out(tp),
2315 tp->snd_ssthresh, tp->prior_ssthresh,
2318 #if IS_ENABLED(CONFIG_IPV6)
2319 else if (sk->sk_family == AF_INET6) {
2320 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2322 &sk->sk_v6_daddr, ntohs(inet->inet_dport),
2323 tp->snd_cwnd, tcp_left_out(tp),
2324 tp->snd_ssthresh, tp->prior_ssthresh,
2331 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
2333 struct tcp_sock *tp = tcp_sk(sk);
2336 struct sk_buff *skb;
2338 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2339 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2342 tcp_clear_all_retrans_hints(tp);
2345 if (tp->prior_ssthresh) {
2346 const struct inet_connection_sock *icsk = inet_csk(sk);
2348 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2350 if (tp->prior_ssthresh > tp->snd_ssthresh) {
2351 tp->snd_ssthresh = tp->prior_ssthresh;
2352 tcp_ecn_withdraw_cwr(tp);
2355 tp->snd_cwnd_stamp = tcp_jiffies32;
2356 tp->undo_marker = 0;
2357 tp->rack.advanced = 1; /* Force RACK to re-exam losses */
2360 static inline bool tcp_may_undo(const struct tcp_sock *tp)
2362 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2365 /* People celebrate: "We love our President!" */
2366 static bool tcp_try_undo_recovery(struct sock *sk)
2368 struct tcp_sock *tp = tcp_sk(sk);
2370 if (tcp_may_undo(tp)) {
2373 /* Happy end! We did not retransmit anything
2374 * or our original transmission succeeded.
2376 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2377 tcp_undo_cwnd_reduction(sk, false);
2378 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2379 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2381 mib_idx = LINUX_MIB_TCPFULLUNDO;
2383 NET_INC_STATS(sock_net(sk), mib_idx);
2384 } else if (tp->rack.reo_wnd_persist) {
2385 tp->rack.reo_wnd_persist--;
2387 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2388 /* Hold old state until something *above* high_seq
2389 * is ACKed. For Reno it is MUST to prevent false
2390 * fast retransmits (RFC2582). SACK TCP is safe. */
2391 if (!tcp_any_retrans_done(sk))
2392 tp->retrans_stamp = 0;
2395 tcp_set_ca_state(sk, TCP_CA_Open);
2396 tp->is_sack_reneg = 0;
2400 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2401 static bool tcp_try_undo_dsack(struct sock *sk)
2403 struct tcp_sock *tp = tcp_sk(sk);
2405 if (tp->undo_marker && !tp->undo_retrans) {
2406 tp->rack.reo_wnd_persist = min(TCP_RACK_RECOVERY_THRESH,
2407 tp->rack.reo_wnd_persist + 1);
2408 DBGUNDO(sk, "D-SACK");
2409 tcp_undo_cwnd_reduction(sk, false);
2410 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2416 /* Undo during loss recovery after partial ACK or using F-RTO. */
2417 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2419 struct tcp_sock *tp = tcp_sk(sk);
2421 if (frto_undo || tcp_may_undo(tp)) {
2422 tcp_undo_cwnd_reduction(sk, true);
2424 DBGUNDO(sk, "partial loss");
2425 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2427 NET_INC_STATS(sock_net(sk),
2428 LINUX_MIB_TCPSPURIOUSRTOS);
2429 inet_csk(sk)->icsk_retransmits = 0;
2430 if (frto_undo || tcp_is_sack(tp)) {
2431 tcp_set_ca_state(sk, TCP_CA_Open);
2432 tp->is_sack_reneg = 0;
2439 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
2440 * It computes the number of packets to send (sndcnt) based on packets newly
2442 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2443 * cwnd reductions across a full RTT.
2444 * 2) Otherwise PRR uses packet conservation to send as much as delivered.
2445 * But when the retransmits are acked without further losses, PRR
2446 * slow starts cwnd up to ssthresh to speed up the recovery.
2448 static void tcp_init_cwnd_reduction(struct sock *sk)
2450 struct tcp_sock *tp = tcp_sk(sk);
2452 tp->high_seq = tp->snd_nxt;
2453 tp->tlp_high_seq = 0;
2454 tp->snd_cwnd_cnt = 0;
2455 tp->prior_cwnd = tp->snd_cwnd;
2456 tp->prr_delivered = 0;
2458 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2459 tcp_ecn_queue_cwr(tp);
2462 void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked, int flag)
2464 struct tcp_sock *tp = tcp_sk(sk);
2466 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2468 if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd))
2471 tp->prr_delivered += newly_acked_sacked;
2473 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2475 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2476 } else if ((flag & (FLAG_RETRANS_DATA_ACKED | FLAG_LOST_RETRANS)) ==
2477 FLAG_RETRANS_DATA_ACKED) {
2478 sndcnt = min_t(int, delta,
2479 max_t(int, tp->prr_delivered - tp->prr_out,
2480 newly_acked_sacked) + 1);
2482 sndcnt = min(delta, newly_acked_sacked);
2484 /* Force a fast retransmit upon entering fast recovery */
2485 sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1));
2486 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2489 static inline void tcp_end_cwnd_reduction(struct sock *sk)
2491 struct tcp_sock *tp = tcp_sk(sk);
2493 if (inet_csk(sk)->icsk_ca_ops->cong_control)
2496 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2497 if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH &&
2498 (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR || tp->undo_marker)) {
2499 tp->snd_cwnd = tp->snd_ssthresh;
2500 tp->snd_cwnd_stamp = tcp_jiffies32;
2502 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2505 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2506 void tcp_enter_cwr(struct sock *sk)
2508 struct tcp_sock *tp = tcp_sk(sk);
2510 tp->prior_ssthresh = 0;
2511 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2512 tp->undo_marker = 0;
2513 tcp_init_cwnd_reduction(sk);
2514 tcp_set_ca_state(sk, TCP_CA_CWR);
2517 EXPORT_SYMBOL(tcp_enter_cwr);
2519 static void tcp_try_keep_open(struct sock *sk)
2521 struct tcp_sock *tp = tcp_sk(sk);
2522 int state = TCP_CA_Open;
2524 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2525 state = TCP_CA_Disorder;
2527 if (inet_csk(sk)->icsk_ca_state != state) {
2528 tcp_set_ca_state(sk, state);
2529 tp->high_seq = tp->snd_nxt;
2533 static void tcp_try_to_open(struct sock *sk, int flag)
2535 struct tcp_sock *tp = tcp_sk(sk);
2537 tcp_verify_left_out(tp);
2539 if (!tcp_any_retrans_done(sk))
2540 tp->retrans_stamp = 0;
2542 if (flag & FLAG_ECE)
2545 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2546 tcp_try_keep_open(sk);
2550 static void tcp_mtup_probe_failed(struct sock *sk)
2552 struct inet_connection_sock *icsk = inet_csk(sk);
2554 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2555 icsk->icsk_mtup.probe_size = 0;
2556 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPFAIL);
2559 static void tcp_mtup_probe_success(struct sock *sk)
2561 struct tcp_sock *tp = tcp_sk(sk);
2562 struct inet_connection_sock *icsk = inet_csk(sk);
2564 /* FIXME: breaks with very large cwnd */
2565 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2566 tp->snd_cwnd = tp->snd_cwnd *
2567 tcp_mss_to_mtu(sk, tp->mss_cache) /
2568 icsk->icsk_mtup.probe_size;
2569 tp->snd_cwnd_cnt = 0;
2570 tp->snd_cwnd_stamp = tcp_jiffies32;
2571 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2573 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2574 icsk->icsk_mtup.probe_size = 0;
2575 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2576 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS);
2579 /* Do a simple retransmit without using the backoff mechanisms in
2580 * tcp_timer. This is used for path mtu discovery.
2581 * The socket is already locked here.
2583 void tcp_simple_retransmit(struct sock *sk)
2585 const struct inet_connection_sock *icsk = inet_csk(sk);
2586 struct tcp_sock *tp = tcp_sk(sk);
2587 struct sk_buff *skb;
2588 unsigned int mss = tcp_current_mss(sk);
2590 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2591 if (tcp_skb_seglen(skb) > mss &&
2592 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2593 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2594 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2595 tp->retrans_out -= tcp_skb_pcount(skb);
2597 tcp_skb_mark_lost_uncond_verify(tp, skb);
2601 tcp_clear_retrans_hints_partial(tp);
2606 if (tcp_is_reno(tp))
2607 tcp_limit_reno_sacked(tp);
2609 tcp_verify_left_out(tp);
2611 /* Don't muck with the congestion window here.
2612 * Reason is that we do not increase amount of _data_
2613 * in network, but units changed and effective
2614 * cwnd/ssthresh really reduced now.
2616 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2617 tp->high_seq = tp->snd_nxt;
2618 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2619 tp->prior_ssthresh = 0;
2620 tp->undo_marker = 0;
2621 tcp_set_ca_state(sk, TCP_CA_Loss);
2623 tcp_xmit_retransmit_queue(sk);
2625 EXPORT_SYMBOL(tcp_simple_retransmit);
2627 void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2629 struct tcp_sock *tp = tcp_sk(sk);
2632 if (tcp_is_reno(tp))
2633 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2635 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2637 NET_INC_STATS(sock_net(sk), mib_idx);
2639 tp->prior_ssthresh = 0;
2642 if (!tcp_in_cwnd_reduction(sk)) {
2644 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2645 tcp_init_cwnd_reduction(sk);
2647 tcp_set_ca_state(sk, TCP_CA_Recovery);
2650 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2651 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2653 static void tcp_process_loss(struct sock *sk, int flag, int num_dupack,
2656 struct tcp_sock *tp = tcp_sk(sk);
2657 bool recovered = !before(tp->snd_una, tp->high_seq);
2659 if ((flag & FLAG_SND_UNA_ADVANCED || rcu_access_pointer(tp->fastopen_rsk)) &&
2660 tcp_try_undo_loss(sk, false))
2663 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2664 /* Step 3.b. A timeout is spurious if not all data are
2665 * lost, i.e., never-retransmitted data are (s)acked.
2667 if ((flag & FLAG_ORIG_SACK_ACKED) &&
2668 tcp_try_undo_loss(sk, true))
2671 if (after(tp->snd_nxt, tp->high_seq)) {
2672 if (flag & FLAG_DATA_SACKED || num_dupack)
2673 tp->frto = 0; /* Step 3.a. loss was real */
2674 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2675 tp->high_seq = tp->snd_nxt;
2676 /* Step 2.b. Try send new data (but deferred until cwnd
2677 * is updated in tcp_ack()). Otherwise fall back to
2678 * the conventional recovery.
2680 if (!tcp_write_queue_empty(sk) &&
2681 after(tcp_wnd_end(tp), tp->snd_nxt)) {
2682 *rexmit = REXMIT_NEW;
2690 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2691 tcp_try_undo_recovery(sk);
2694 if (tcp_is_reno(tp)) {
2695 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2696 * delivered. Lower inflight to clock out (re)tranmissions.
2698 if (after(tp->snd_nxt, tp->high_seq) && num_dupack)
2699 tcp_add_reno_sack(sk, num_dupack);
2700 else if (flag & FLAG_SND_UNA_ADVANCED)
2701 tcp_reset_reno_sack(tp);
2703 *rexmit = REXMIT_LOST;
2706 /* Undo during fast recovery after partial ACK. */
2707 static bool tcp_try_undo_partial(struct sock *sk, u32 prior_snd_una)
2709 struct tcp_sock *tp = tcp_sk(sk);
2711 if (tp->undo_marker && tcp_packet_delayed(tp)) {
2712 /* Plain luck! Hole if filled with delayed
2713 * packet, rather than with a retransmit. Check reordering.
2715 tcp_check_sack_reordering(sk, prior_snd_una, 1);
2717 /* We are getting evidence that the reordering degree is higher
2718 * than we realized. If there are no retransmits out then we
2719 * can undo. Otherwise we clock out new packets but do not
2720 * mark more packets lost or retransmit more.
2722 if (tp->retrans_out)
2725 if (!tcp_any_retrans_done(sk))
2726 tp->retrans_stamp = 0;
2728 DBGUNDO(sk, "partial recovery");
2729 tcp_undo_cwnd_reduction(sk, true);
2730 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2731 tcp_try_keep_open(sk);
2737 static void tcp_identify_packet_loss(struct sock *sk, int *ack_flag)
2739 struct tcp_sock *tp = tcp_sk(sk);
2741 if (tcp_rtx_queue_empty(sk))
2744 if (unlikely(tcp_is_reno(tp))) {
2745 tcp_newreno_mark_lost(sk, *ack_flag & FLAG_SND_UNA_ADVANCED);
2746 } else if (tcp_is_rack(sk)) {
2747 u32 prior_retrans = tp->retrans_out;
2749 tcp_rack_mark_lost(sk);
2750 if (prior_retrans > tp->retrans_out)
2751 *ack_flag |= FLAG_LOST_RETRANS;
2755 static bool tcp_force_fast_retransmit(struct sock *sk)
2757 struct tcp_sock *tp = tcp_sk(sk);
2759 return after(tcp_highest_sack_seq(tp),
2760 tp->snd_una + tp->reordering * tp->mss_cache);
2763 /* Process an event, which can update packets-in-flight not trivially.
2764 * Main goal of this function is to calculate new estimate for left_out,
2765 * taking into account both packets sitting in receiver's buffer and
2766 * packets lost by network.
2768 * Besides that it updates the congestion state when packet loss or ECN
2769 * is detected. But it does not reduce the cwnd, it is done by the
2770 * congestion control later.
2772 * It does _not_ decide what to send, it is made in function
2773 * tcp_xmit_retransmit_queue().
2775 static void tcp_fastretrans_alert(struct sock *sk, const u32 prior_snd_una,
2776 int num_dupack, int *ack_flag, int *rexmit)
2778 struct inet_connection_sock *icsk = inet_csk(sk);
2779 struct tcp_sock *tp = tcp_sk(sk);
2780 int fast_rexmit = 0, flag = *ack_flag;
2781 bool do_lost = num_dupack || ((flag & FLAG_DATA_SACKED) &&
2782 tcp_force_fast_retransmit(sk));
2784 if (!tp->packets_out && tp->sacked_out)
2787 /* Now state machine starts.
2788 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2789 if (flag & FLAG_ECE)
2790 tp->prior_ssthresh = 0;
2792 /* B. In all the states check for reneging SACKs. */
2793 if (tcp_check_sack_reneging(sk, flag))
2796 /* C. Check consistency of the current state. */
2797 tcp_verify_left_out(tp);
2799 /* D. Check state exit conditions. State can be terminated
2800 * when high_seq is ACKed. */
2801 if (icsk->icsk_ca_state == TCP_CA_Open) {
2802 WARN_ON(tp->retrans_out != 0);
2803 tp->retrans_stamp = 0;
2804 } else if (!before(tp->snd_una, tp->high_seq)) {
2805 switch (icsk->icsk_ca_state) {
2807 /* CWR is to be held something *above* high_seq
2808 * is ACKed for CWR bit to reach receiver. */
2809 if (tp->snd_una != tp->high_seq) {
2810 tcp_end_cwnd_reduction(sk);
2811 tcp_set_ca_state(sk, TCP_CA_Open);
2815 case TCP_CA_Recovery:
2816 if (tcp_is_reno(tp))
2817 tcp_reset_reno_sack(tp);
2818 if (tcp_try_undo_recovery(sk))
2820 tcp_end_cwnd_reduction(sk);
2825 /* E. Process state. */
2826 switch (icsk->icsk_ca_state) {
2827 case TCP_CA_Recovery:
2828 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2829 if (tcp_is_reno(tp))
2830 tcp_add_reno_sack(sk, num_dupack);
2832 if (tcp_try_undo_partial(sk, prior_snd_una))
2834 /* Partial ACK arrived. Force fast retransmit. */
2835 do_lost = tcp_force_fast_retransmit(sk);
2837 if (tcp_try_undo_dsack(sk)) {
2838 tcp_try_keep_open(sk);
2841 tcp_identify_packet_loss(sk, ack_flag);
2844 tcp_process_loss(sk, flag, num_dupack, rexmit);
2845 tcp_identify_packet_loss(sk, ack_flag);
2846 if (!(icsk->icsk_ca_state == TCP_CA_Open ||
2847 (*ack_flag & FLAG_LOST_RETRANS)))
2849 /* Change state if cwnd is undone or retransmits are lost */
2852 if (tcp_is_reno(tp)) {
2853 if (flag & FLAG_SND_UNA_ADVANCED)
2854 tcp_reset_reno_sack(tp);
2855 tcp_add_reno_sack(sk, num_dupack);
2858 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
2859 tcp_try_undo_dsack(sk);
2861 tcp_identify_packet_loss(sk, ack_flag);
2862 if (!tcp_time_to_recover(sk, flag)) {
2863 tcp_try_to_open(sk, flag);
2867 /* MTU probe failure: don't reduce cwnd */
2868 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2869 icsk->icsk_mtup.probe_size &&
2870 tp->snd_una == tp->mtu_probe.probe_seq_start) {
2871 tcp_mtup_probe_failed(sk);
2872 /* Restores the reduction we did in tcp_mtup_probe() */
2874 tcp_simple_retransmit(sk);
2878 /* Otherwise enter Recovery state */
2879 tcp_enter_recovery(sk, (flag & FLAG_ECE));
2883 if (!tcp_is_rack(sk) && do_lost)
2884 tcp_update_scoreboard(sk, fast_rexmit);
2885 *rexmit = REXMIT_LOST;
2888 static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us, const int flag)
2890 u32 wlen = sock_net(sk)->ipv4.sysctl_tcp_min_rtt_wlen * HZ;
2891 struct tcp_sock *tp = tcp_sk(sk);
2893 if ((flag & FLAG_ACK_MAYBE_DELAYED) && rtt_us > tcp_min_rtt(tp)) {
2894 /* If the remote keeps returning delayed ACKs, eventually
2895 * the min filter would pick it up and overestimate the
2896 * prop. delay when it expires. Skip suspected delayed ACKs.
2900 minmax_running_min(&tp->rtt_min, wlen, tcp_jiffies32,
2901 rtt_us ? : jiffies_to_usecs(1));
2904 static bool tcp_ack_update_rtt(struct sock *sk, const int flag,
2905 long seq_rtt_us, long sack_rtt_us,
2906 long ca_rtt_us, struct rate_sample *rs)
2908 const struct tcp_sock *tp = tcp_sk(sk);
2910 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2911 * broken middle-boxes or peers may corrupt TS-ECR fields. But
2912 * Karn's algorithm forbids taking RTT if some retransmitted data
2913 * is acked (RFC6298).
2916 seq_rtt_us = sack_rtt_us;
2918 /* RTTM Rule: A TSecr value received in a segment is used to
2919 * update the averaged RTT measurement only if the segment
2920 * acknowledges some new data, i.e., only if it advances the
2921 * left edge of the send window.
2922 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2924 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2925 flag & FLAG_ACKED) {
2926 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr;
2928 if (likely(delta < INT_MAX / (USEC_PER_SEC / TCP_TS_HZ))) {
2929 seq_rtt_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
2930 ca_rtt_us = seq_rtt_us;
2933 rs->rtt_us = ca_rtt_us; /* RTT of last (S)ACKed packet (or -1) */
2937 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
2938 * always taken together with ACK, SACK, or TS-opts. Any negative
2939 * values will be skipped with the seq_rtt_us < 0 check above.
2941 tcp_update_rtt_min(sk, ca_rtt_us, flag);
2942 tcp_rtt_estimator(sk, seq_rtt_us);
2945 /* RFC6298: only reset backoff on valid RTT measurement. */
2946 inet_csk(sk)->icsk_backoff = 0;
2950 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
2951 void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req)
2953 struct rate_sample rs;
2956 if (req && !req->num_retrans && tcp_rsk(req)->snt_synack)
2957 rtt_us = tcp_stamp_us_delta(tcp_clock_us(), tcp_rsk(req)->snt_synack);
2959 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us, &rs);
2963 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
2965 const struct inet_connection_sock *icsk = inet_csk(sk);
2967 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
2968 tcp_sk(sk)->snd_cwnd_stamp = tcp_jiffies32;
2971 /* Restart timer after forward progress on connection.
2972 * RFC2988 recommends to restart timer to now+rto.
2974 void tcp_rearm_rto(struct sock *sk)
2976 const struct inet_connection_sock *icsk = inet_csk(sk);
2977 struct tcp_sock *tp = tcp_sk(sk);
2979 /* If the retrans timer is currently being used by Fast Open
2980 * for SYN-ACK retrans purpose, stay put.
2982 if (rcu_access_pointer(tp->fastopen_rsk))
2985 if (!tp->packets_out) {
2986 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
2988 u32 rto = inet_csk(sk)->icsk_rto;
2989 /* Offset the time elapsed after installing regular RTO */
2990 if (icsk->icsk_pending == ICSK_TIME_REO_TIMEOUT ||
2991 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
2992 s64 delta_us = tcp_rto_delta_us(sk);
2993 /* delta_us may not be positive if the socket is locked
2994 * when the retrans timer fires and is rescheduled.
2996 rto = usecs_to_jiffies(max_t(int, delta_us, 1));
2998 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
3003 /* Try to schedule a loss probe; if that doesn't work, then schedule an RTO. */
3004 static void tcp_set_xmit_timer(struct sock *sk)
3006 if (!tcp_schedule_loss_probe(sk, true))
3010 /* If we get here, the whole TSO packet has not been acked. */
3011 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3013 struct tcp_sock *tp = tcp_sk(sk);
3016 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3018 packets_acked = tcp_skb_pcount(skb);
3019 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3021 packets_acked -= tcp_skb_pcount(skb);
3023 if (packets_acked) {
3024 BUG_ON(tcp_skb_pcount(skb) == 0);
3025 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3028 return packets_acked;
3031 static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
3034 const struct skb_shared_info *shinfo;
3036 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3037 if (likely(!TCP_SKB_CB(skb)->txstamp_ack))
3040 shinfo = skb_shinfo(skb);
3041 if (!before(shinfo->tskey, prior_snd_una) &&
3042 before(shinfo->tskey, tcp_sk(sk)->snd_una)) {
3043 tcp_skb_tsorted_save(skb) {
3044 __skb_tstamp_tx(skb, NULL, sk, SCM_TSTAMP_ACK);
3045 } tcp_skb_tsorted_restore(skb);
3049 /* Remove acknowledged frames from the retransmission queue. If our packet
3050 * is before the ack sequence we can discard it as it's confirmed to have
3051 * arrived at the other end.
3053 static int tcp_clean_rtx_queue(struct sock *sk, u32 prior_fack,
3055 struct tcp_sacktag_state *sack)
3057 const struct inet_connection_sock *icsk = inet_csk(sk);
3058 u64 first_ackt, last_ackt;
3059 struct tcp_sock *tp = tcp_sk(sk);
3060 u32 prior_sacked = tp->sacked_out;
3061 u32 reord = tp->snd_nxt; /* lowest acked un-retx un-sacked seq */
3062 struct sk_buff *skb, *next;
3063 bool fully_acked = true;
3064 long sack_rtt_us = -1L;
3065 long seq_rtt_us = -1L;
3066 long ca_rtt_us = -1L;
3068 u32 last_in_flight = 0;
3074 for (skb = skb_rb_first(&sk->tcp_rtx_queue); skb; skb = next) {
3075 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3076 const u32 start_seq = scb->seq;
3077 u8 sacked = scb->sacked;
3080 tcp_ack_tstamp(sk, skb, prior_snd_una);
3082 /* Determine how many packets and what bytes were acked, tso and else */
3083 if (after(scb->end_seq, tp->snd_una)) {
3084 if (tcp_skb_pcount(skb) == 1 ||
3085 !after(tp->snd_una, scb->seq))
3088 acked_pcount = tcp_tso_acked(sk, skb);
3091 fully_acked = false;
3093 acked_pcount = tcp_skb_pcount(skb);
3096 if (unlikely(sacked & TCPCB_RETRANS)) {
3097 if (sacked & TCPCB_SACKED_RETRANS)
3098 tp->retrans_out -= acked_pcount;
3099 flag |= FLAG_RETRANS_DATA_ACKED;
3100 } else if (!(sacked & TCPCB_SACKED_ACKED)) {
3101 last_ackt = tcp_skb_timestamp_us(skb);
3102 WARN_ON_ONCE(last_ackt == 0);
3104 first_ackt = last_ackt;
3106 last_in_flight = TCP_SKB_CB(skb)->tx.in_flight;
3107 if (before(start_seq, reord))
3109 if (!after(scb->end_seq, tp->high_seq))
3110 flag |= FLAG_ORIG_SACK_ACKED;
3113 if (sacked & TCPCB_SACKED_ACKED) {
3114 tp->sacked_out -= acked_pcount;
3115 } else if (tcp_is_sack(tp)) {
3116 tp->delivered += acked_pcount;
3117 if (!tcp_skb_spurious_retrans(tp, skb))
3118 tcp_rack_advance(tp, sacked, scb->end_seq,
3119 tcp_skb_timestamp_us(skb));
3121 if (sacked & TCPCB_LOST)
3122 tp->lost_out -= acked_pcount;
3124 tp->packets_out -= acked_pcount;
3125 pkts_acked += acked_pcount;
3126 tcp_rate_skb_delivered(sk, skb, sack->rate);
3128 /* Initial outgoing SYN's get put onto the write_queue
3129 * just like anything else we transmit. It is not
3130 * true data, and if we misinform our callers that
3131 * this ACK acks real data, we will erroneously exit
3132 * connection startup slow start one packet too
3133 * quickly. This is severely frowned upon behavior.
3135 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
3136 flag |= FLAG_DATA_ACKED;
3138 flag |= FLAG_SYN_ACKED;
3139 tp->retrans_stamp = 0;
3145 next = skb_rb_next(skb);
3146 if (unlikely(skb == tp->retransmit_skb_hint))
3147 tp->retransmit_skb_hint = NULL;
3148 if (unlikely(skb == tp->lost_skb_hint))
3149 tp->lost_skb_hint = NULL;
3150 tcp_highest_sack_replace(sk, skb, next);
3151 tcp_rtx_queue_unlink_and_free(skb, sk);
3155 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
3157 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3158 tp->snd_up = tp->snd_una;
3160 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3161 flag |= FLAG_SACK_RENEGING;
3163 if (likely(first_ackt) && !(flag & FLAG_RETRANS_DATA_ACKED)) {
3164 seq_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, first_ackt);
3165 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, last_ackt);
3167 if (pkts_acked == 1 && last_in_flight < tp->mss_cache &&
3168 last_in_flight && !prior_sacked && fully_acked &&
3169 sack->rate->prior_delivered + 1 == tp->delivered &&
3170 !(flag & (FLAG_CA_ALERT | FLAG_SYN_ACKED))) {
3171 /* Conservatively mark a delayed ACK. It's typically
3172 * from a lone runt packet over the round trip to
3173 * a receiver w/o out-of-order or CE events.
3175 flag |= FLAG_ACK_MAYBE_DELAYED;
3178 if (sack->first_sackt) {
3179 sack_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->first_sackt);
3180 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->last_sackt);
3182 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us,
3183 ca_rtt_us, sack->rate);
3185 if (flag & FLAG_ACKED) {
3186 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3187 if (unlikely(icsk->icsk_mtup.probe_size &&
3188 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3189 tcp_mtup_probe_success(sk);
3192 if (tcp_is_reno(tp)) {
3193 tcp_remove_reno_sacks(sk, pkts_acked);
3195 /* If any of the cumulatively ACKed segments was
3196 * retransmitted, non-SACK case cannot confirm that
3197 * progress was due to original transmission due to
3198 * lack of TCPCB_SACKED_ACKED bits even if some of
3199 * the packets may have been never retransmitted.
3201 if (flag & FLAG_RETRANS_DATA_ACKED)
3202 flag &= ~FLAG_ORIG_SACK_ACKED;
3206 /* Non-retransmitted hole got filled? That's reordering */
3207 if (before(reord, prior_fack))
3208 tcp_check_sack_reordering(sk, reord, 0);
3210 delta = prior_sacked - tp->sacked_out;
3211 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3213 } else if (skb && rtt_update && sack_rtt_us >= 0 &&
3214 sack_rtt_us > tcp_stamp_us_delta(tp->tcp_mstamp,
3215 tcp_skb_timestamp_us(skb))) {
3216 /* Do not re-arm RTO if the sack RTT is measured from data sent
3217 * after when the head was last (re)transmitted. Otherwise the
3218 * timeout may continue to extend in loss recovery.
3220 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3223 if (icsk->icsk_ca_ops->pkts_acked) {
3224 struct ack_sample sample = { .pkts_acked = pkts_acked,
3225 .rtt_us = sack->rate->rtt_us,
3226 .in_flight = last_in_flight };
3228 icsk->icsk_ca_ops->pkts_acked(sk, &sample);
3231 #if FASTRETRANS_DEBUG > 0
3232 WARN_ON((int)tp->sacked_out < 0);
3233 WARN_ON((int)tp->lost_out < 0);
3234 WARN_ON((int)tp->retrans_out < 0);
3235 if (!tp->packets_out && tcp_is_sack(tp)) {
3236 icsk = inet_csk(sk);
3238 pr_debug("Leak l=%u %d\n",
3239 tp->lost_out, icsk->icsk_ca_state);
3242 if (tp->sacked_out) {
3243 pr_debug("Leak s=%u %d\n",
3244 tp->sacked_out, icsk->icsk_ca_state);
3247 if (tp->retrans_out) {
3248 pr_debug("Leak r=%u %d\n",
3249 tp->retrans_out, icsk->icsk_ca_state);
3250 tp->retrans_out = 0;
3257 static void tcp_ack_probe(struct sock *sk)
3259 struct inet_connection_sock *icsk = inet_csk(sk);
3260 struct sk_buff *head = tcp_send_head(sk);
3261 const struct tcp_sock *tp = tcp_sk(sk);
3263 /* Was it a usable window open? */
3266 if (!after(TCP_SKB_CB(head)->end_seq, tcp_wnd_end(tp))) {
3267 icsk->icsk_backoff = 0;
3268 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3269 /* Socket must be waked up by subsequent tcp_data_snd_check().
3270 * This function is not for random using!
3273 unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX);
3275 tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3280 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3282 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3283 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3286 /* Decide wheather to run the increase function of congestion control. */
3287 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3289 /* If reordering is high then always grow cwnd whenever data is
3290 * delivered regardless of its ordering. Otherwise stay conservative
3291 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3292 * new SACK or ECE mark may first advance cwnd here and later reduce
3293 * cwnd in tcp_fastretrans_alert() based on more states.
3295 if (tcp_sk(sk)->reordering > sock_net(sk)->ipv4.sysctl_tcp_reordering)
3296 return flag & FLAG_FORWARD_PROGRESS;
3298 return flag & FLAG_DATA_ACKED;
3301 /* The "ultimate" congestion control function that aims to replace the rigid
3302 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction).
3303 * It's called toward the end of processing an ACK with precise rate
3304 * information. All transmission or retransmission are delayed afterwards.
3306 static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked,
3307 int flag, const struct rate_sample *rs)
3309 const struct inet_connection_sock *icsk = inet_csk(sk);
3311 if (icsk->icsk_ca_ops->cong_control) {
3312 icsk->icsk_ca_ops->cong_control(sk, rs);
3316 if (tcp_in_cwnd_reduction(sk)) {
3317 /* Reduce cwnd if state mandates */
3318 tcp_cwnd_reduction(sk, acked_sacked, flag);
3319 } else if (tcp_may_raise_cwnd(sk, flag)) {
3320 /* Advance cwnd if state allows */
3321 tcp_cong_avoid(sk, ack, acked_sacked);
3323 tcp_update_pacing_rate(sk);
3326 /* Check that window update is acceptable.
3327 * The function assumes that snd_una<=ack<=snd_next.
3329 static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3330 const u32 ack, const u32 ack_seq,
3333 return after(ack, tp->snd_una) ||
3334 after(ack_seq, tp->snd_wl1) ||
3335 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3338 /* If we update tp->snd_una, also update tp->bytes_acked */
3339 static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack)
3341 u32 delta = ack - tp->snd_una;
3343 sock_owned_by_me((struct sock *)tp);
3344 tp->bytes_acked += delta;
3348 /* If we update tp->rcv_nxt, also update tp->bytes_received */
3349 static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq)
3351 u32 delta = seq - tp->rcv_nxt;
3353 sock_owned_by_me((struct sock *)tp);
3354 tp->bytes_received += delta;
3355 WRITE_ONCE(tp->rcv_nxt, seq);
3358 /* Update our send window.
3360 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3361 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3363 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3366 struct tcp_sock *tp = tcp_sk(sk);
3368 u32 nwin = ntohs(tcp_hdr(skb)->window);
3370 if (likely(!tcp_hdr(skb)->syn))
3371 nwin <<= tp->rx_opt.snd_wscale;
3373 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3374 flag |= FLAG_WIN_UPDATE;
3375 tcp_update_wl(tp, ack_seq);
3377 if (tp->snd_wnd != nwin) {
3380 /* Note, it is the only place, where
3381 * fast path is recovered for sending TCP.
3384 tcp_fast_path_check(sk);
3386 if (!tcp_write_queue_empty(sk))
3387 tcp_slow_start_after_idle_check(sk);
3389 if (nwin > tp->max_window) {
3390 tp->max_window = nwin;
3391 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3396 tcp_snd_una_update(tp, ack);
3401 static bool __tcp_oow_rate_limited(struct net *net, int mib_idx,
3402 u32 *last_oow_ack_time)
3404 if (*last_oow_ack_time) {
3405 s32 elapsed = (s32)(tcp_jiffies32 - *last_oow_ack_time);
3407 if (0 <= elapsed && elapsed < net->ipv4.sysctl_tcp_invalid_ratelimit) {
3408 NET_INC_STATS(net, mib_idx);
3409 return true; /* rate-limited: don't send yet! */
3413 *last_oow_ack_time = tcp_jiffies32;
3415 return false; /* not rate-limited: go ahead, send dupack now! */
3418 /* Return true if we're currently rate-limiting out-of-window ACKs and
3419 * thus shouldn't send a dupack right now. We rate-limit dupacks in
3420 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3421 * attacks that send repeated SYNs or ACKs for the same connection. To
3422 * do this, we do not send a duplicate SYNACK or ACK if the remote
3423 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3425 bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb,
3426 int mib_idx, u32 *last_oow_ack_time)
3428 /* Data packets without SYNs are not likely part of an ACK loop. */
3429 if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) &&
3433 return __tcp_oow_rate_limited(net, mib_idx, last_oow_ack_time);
3436 /* RFC 5961 7 [ACK Throttling] */
3437 static void tcp_send_challenge_ack(struct sock *sk, const struct sk_buff *skb)
3439 /* unprotected vars, we dont care of overwrites */
3440 static u32 challenge_timestamp;
3441 static unsigned int challenge_count;
3442 struct tcp_sock *tp = tcp_sk(sk);
3443 struct net *net = sock_net(sk);
3446 /* First check our per-socket dupack rate limit. */
3447 if (__tcp_oow_rate_limited(net,
3448 LINUX_MIB_TCPACKSKIPPEDCHALLENGE,
3449 &tp->last_oow_ack_time))
3452 /* Then check host-wide RFC 5961 rate limit. */
3454 if (now != challenge_timestamp) {
3455 u32 ack_limit = net->ipv4.sysctl_tcp_challenge_ack_limit;
3456 u32 half = (ack_limit + 1) >> 1;
3458 challenge_timestamp = now;
3459 WRITE_ONCE(challenge_count, half + prandom_u32_max(ack_limit));
3461 count = READ_ONCE(challenge_count);
3463 WRITE_ONCE(challenge_count, count - 1);
3464 NET_INC_STATS(net, LINUX_MIB_TCPCHALLENGEACK);
3469 static void tcp_store_ts_recent(struct tcp_sock *tp)
3471 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3472 tp->rx_opt.ts_recent_stamp = ktime_get_seconds();
3475 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3477 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3478 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3479 * extra check below makes sure this can only happen
3480 * for pure ACK frames. -DaveM
3482 * Not only, also it occurs for expired timestamps.
3485 if (tcp_paws_check(&tp->rx_opt, 0))
3486 tcp_store_ts_recent(tp);
3490 /* This routine deals with acks during a TLP episode.
3491 * We mark the end of a TLP episode on receiving TLP dupack or when
3492 * ack is after tlp_high_seq.
3493 * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3495 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3497 struct tcp_sock *tp = tcp_sk(sk);
3499 if (before(ack, tp->tlp_high_seq))
3502 if (flag & FLAG_DSACKING_ACK) {
3503 /* This DSACK means original and TLP probe arrived; no loss */
3504 tp->tlp_high_seq = 0;
3505 } else if (after(ack, tp->tlp_high_seq)) {
3506 /* ACK advances: there was a loss, so reduce cwnd. Reset
3507 * tlp_high_seq in tcp_init_cwnd_reduction()
3509 tcp_init_cwnd_reduction(sk);
3510 tcp_set_ca_state(sk, TCP_CA_CWR);
3511 tcp_end_cwnd_reduction(sk);
3512 tcp_try_keep_open(sk);
3513 NET_INC_STATS(sock_net(sk),
3514 LINUX_MIB_TCPLOSSPROBERECOVERY);
3515 } else if (!(flag & (FLAG_SND_UNA_ADVANCED |
3516 FLAG_NOT_DUP | FLAG_DATA_SACKED))) {
3517 /* Pure dupack: original and TLP probe arrived; no loss */
3518 tp->tlp_high_seq = 0;
3522 static inline void tcp_in_ack_event(struct sock *sk, u32 flags)
3524 const struct inet_connection_sock *icsk = inet_csk(sk);
3526 if (icsk->icsk_ca_ops->in_ack_event)
3527 icsk->icsk_ca_ops->in_ack_event(sk, flags);
3530 /* Congestion control has updated the cwnd already. So if we're in
3531 * loss recovery then now we do any new sends (for FRTO) or
3532 * retransmits (for CA_Loss or CA_recovery) that make sense.
3534 static void tcp_xmit_recovery(struct sock *sk, int rexmit)
3536 struct tcp_sock *tp = tcp_sk(sk);
3538 if (rexmit == REXMIT_NONE || sk->sk_state == TCP_SYN_SENT)
3541 if (unlikely(rexmit == REXMIT_NEW)) {
3542 __tcp_push_pending_frames(sk, tcp_current_mss(sk),
3544 if (after(tp->snd_nxt, tp->high_seq))
3548 tcp_xmit_retransmit_queue(sk);
3551 /* Returns the number of packets newly acked or sacked by the current ACK */
3552 static u32 tcp_newly_delivered(struct sock *sk, u32 prior_delivered, int flag)
3554 const struct net *net = sock_net(sk);
3555 struct tcp_sock *tp = tcp_sk(sk);
3558 delivered = tp->delivered - prior_delivered;
3559 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVERED, delivered);
3560 if (flag & FLAG_ECE) {
3561 tp->delivered_ce += delivered;
3562 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVEREDCE, delivered);
3567 /* This routine deals with incoming acks, but not outgoing ones. */
3568 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3570 struct inet_connection_sock *icsk = inet_csk(sk);
3571 struct tcp_sock *tp = tcp_sk(sk);
3572 struct tcp_sacktag_state sack_state;
3573 struct rate_sample rs = { .prior_delivered = 0 };
3574 u32 prior_snd_una = tp->snd_una;
3575 bool is_sack_reneg = tp->is_sack_reneg;
3576 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3577 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3579 int prior_packets = tp->packets_out;
3580 u32 delivered = tp->delivered;
3581 u32 lost = tp->lost;
3582 int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */
3585 sack_state.first_sackt = 0;
3586 sack_state.rate = &rs;
3588 /* We very likely will need to access rtx queue. */
3589 prefetch(sk->tcp_rtx_queue.rb_node);
3591 /* If the ack is older than previous acks
3592 * then we can probably ignore it.
3594 if (before(ack, prior_snd_una)) {
3595 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3596 if (before(ack, prior_snd_una - tp->max_window)) {
3597 if (!(flag & FLAG_NO_CHALLENGE_ACK))
3598 tcp_send_challenge_ack(sk, skb);
3604 /* If the ack includes data we haven't sent yet, discard
3605 * this segment (RFC793 Section 3.9).
3607 if (after(ack, tp->snd_nxt))
3610 if (after(ack, prior_snd_una)) {
3611 flag |= FLAG_SND_UNA_ADVANCED;
3612 icsk->icsk_retransmits = 0;
3614 #if IS_ENABLED(CONFIG_TLS_DEVICE)
3615 if (static_branch_unlikely(&clean_acked_data_enabled.key))
3616 if (icsk->icsk_clean_acked)
3617 icsk->icsk_clean_acked(sk, ack);
3621 prior_fack = tcp_is_sack(tp) ? tcp_highest_sack_seq(tp) : tp->snd_una;
3622 rs.prior_in_flight = tcp_packets_in_flight(tp);
3624 /* ts_recent update must be made after we are sure that the packet
3627 if (flag & FLAG_UPDATE_TS_RECENT)
3628 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3630 if ((flag & (FLAG_SLOWPATH | FLAG_SND_UNA_ADVANCED)) ==
3631 FLAG_SND_UNA_ADVANCED) {
3632 /* Window is constant, pure forward advance.
3633 * No more checks are required.
3634 * Note, we use the fact that SND.UNA>=SND.WL2.
3636 tcp_update_wl(tp, ack_seq);
3637 tcp_snd_una_update(tp, ack);
3638 flag |= FLAG_WIN_UPDATE;
3640 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
3642 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPACKS);
3644 u32 ack_ev_flags = CA_ACK_SLOWPATH;
3646 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3649 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3651 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3653 if (TCP_SKB_CB(skb)->sacked)
3654 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3657 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) {
3659 ack_ev_flags |= CA_ACK_ECE;
3662 if (flag & FLAG_WIN_UPDATE)
3663 ack_ev_flags |= CA_ACK_WIN_UPDATE;
3665 tcp_in_ack_event(sk, ack_ev_flags);
3668 /* We passed data and got it acked, remove any soft error
3669 * log. Something worked...
3671 sk->sk_err_soft = 0;
3672 icsk->icsk_probes_out = 0;
3673 tp->rcv_tstamp = tcp_jiffies32;
3677 /* See if we can take anything off of the retransmit queue. */
3678 flag |= tcp_clean_rtx_queue(sk, prior_fack, prior_snd_una, &sack_state);
3680 tcp_rack_update_reo_wnd(sk, &rs);
3682 if (tp->tlp_high_seq)
3683 tcp_process_tlp_ack(sk, ack, flag);
3684 /* If needed, reset TLP/RTO timer; RACK may later override this. */
3685 if (flag & FLAG_SET_XMIT_TIMER)
3686 tcp_set_xmit_timer(sk);
3688 if (tcp_ack_is_dubious(sk, flag)) {
3689 if (!(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP))) {
3691 /* Consider if pure acks were aggregated in tcp_add_backlog() */
3692 if (!(flag & FLAG_DATA))
3693 num_dupack = max_t(u16, 1, skb_shinfo(skb)->gso_segs);
3695 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
3699 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
3702 delivered = tcp_newly_delivered(sk, delivered, flag);
3703 lost = tp->lost - lost; /* freshly marked lost */
3704 rs.is_ack_delayed = !!(flag & FLAG_ACK_MAYBE_DELAYED);
3705 tcp_rate_gen(sk, delivered, lost, is_sack_reneg, sack_state.rate);
3706 tcp_cong_control(sk, ack, delivered, flag, sack_state.rate);
3707 tcp_xmit_recovery(sk, rexmit);
3711 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3712 if (flag & FLAG_DSACKING_ACK) {
3713 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
3715 tcp_newly_delivered(sk, delivered, flag);
3717 /* If this ack opens up a zero window, clear backoff. It was
3718 * being used to time the probes, and is probably far higher than
3719 * it needs to be for normal retransmission.
3723 if (tp->tlp_high_seq)
3724 tcp_process_tlp_ack(sk, ack, flag);
3728 /* If data was SACKed, tag it and see if we should send more data.
3729 * If data was DSACKed, see if we can undo a cwnd reduction.
3731 if (TCP_SKB_CB(skb)->sacked) {
3732 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3734 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
3736 tcp_newly_delivered(sk, delivered, flag);
3737 tcp_xmit_recovery(sk, rexmit);
3743 static void tcp_parse_fastopen_option(int len, const unsigned char *cookie,
3744 bool syn, struct tcp_fastopen_cookie *foc,
3747 /* Valid only in SYN or SYN-ACK with an even length. */
3748 if (!foc || !syn || len < 0 || (len & 1))
3751 if (len >= TCP_FASTOPEN_COOKIE_MIN &&
3752 len <= TCP_FASTOPEN_COOKIE_MAX)
3753 memcpy(foc->val, cookie, len);
3760 static void smc_parse_options(const struct tcphdr *th,
3761 struct tcp_options_received *opt_rx,
3762 const unsigned char *ptr,
3765 #if IS_ENABLED(CONFIG_SMC)
3766 if (static_branch_unlikely(&tcp_have_smc)) {
3767 if (th->syn && !(opsize & 1) &&
3768 opsize >= TCPOLEN_EXP_SMC_BASE &&
3769 get_unaligned_be32(ptr) == TCPOPT_SMC_MAGIC)
3775 /* Try to parse the MSS option from the TCP header. Return 0 on failure, clamped
3778 static u16 tcp_parse_mss_option(const struct tcphdr *th, u16 user_mss)
3780 const unsigned char *ptr = (const unsigned char *)(th + 1);
3781 int length = (th->doff * 4) - sizeof(struct tcphdr);
3784 while (length > 0) {
3785 int opcode = *ptr++;
3791 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3798 if (opsize < 2) /* "silly options" */
3800 if (opsize > length)
3801 return mss; /* fail on partial options */
3802 if (opcode == TCPOPT_MSS && opsize == TCPOLEN_MSS) {
3803 u16 in_mss = get_unaligned_be16(ptr);
3806 if (user_mss && user_mss < in_mss)
3818 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3819 * But, this can also be called on packets in the established flow when
3820 * the fast version below fails.
3822 void tcp_parse_options(const struct net *net,
3823 const struct sk_buff *skb,
3824 struct tcp_options_received *opt_rx, int estab,
3825 struct tcp_fastopen_cookie *foc)
3827 const unsigned char *ptr;
3828 const struct tcphdr *th = tcp_hdr(skb);
3829 int length = (th->doff * 4) - sizeof(struct tcphdr);
3831 ptr = (const unsigned char *)(th + 1);
3832 opt_rx->saw_tstamp = 0;
3834 while (length > 0) {
3835 int opcode = *ptr++;
3841 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3848 if (opsize < 2) /* "silly options" */
3850 if (opsize > length)
3851 return; /* don't parse partial options */
3854 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3855 u16 in_mss = get_unaligned_be16(ptr);
3857 if (opt_rx->user_mss &&
3858 opt_rx->user_mss < in_mss)
3859 in_mss = opt_rx->user_mss;
3860 opt_rx->mss_clamp = in_mss;
3865 if (opsize == TCPOLEN_WINDOW && th->syn &&
3866 !estab && net->ipv4.sysctl_tcp_window_scaling) {
3867 __u8 snd_wscale = *(__u8 *)ptr;
3868 opt_rx->wscale_ok = 1;
3869 if (snd_wscale > TCP_MAX_WSCALE) {
3870 net_info_ratelimited("%s: Illegal window scaling value %d > %u received\n",
3874 snd_wscale = TCP_MAX_WSCALE;
3876 opt_rx->snd_wscale = snd_wscale;
3879 case TCPOPT_TIMESTAMP:
3880 if ((opsize == TCPOLEN_TIMESTAMP) &&
3881 ((estab && opt_rx->tstamp_ok) ||
3882 (!estab && net->ipv4.sysctl_tcp_timestamps))) {
3883 opt_rx->saw_tstamp = 1;
3884 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3885 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3888 case TCPOPT_SACK_PERM:
3889 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3890 !estab && net->ipv4.sysctl_tcp_sack) {
3891 opt_rx->sack_ok = TCP_SACK_SEEN;
3892 tcp_sack_reset(opt_rx);
3897 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3898 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3900 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3903 #ifdef CONFIG_TCP_MD5SIG
3906 * The MD5 Hash has already been
3907 * checked (see tcp_v{4,6}_do_rcv()).
3911 case TCPOPT_FASTOPEN:
3912 tcp_parse_fastopen_option(
3913 opsize - TCPOLEN_FASTOPEN_BASE,
3914 ptr, th->syn, foc, false);
3918 /* Fast Open option shares code 254 using a
3919 * 16 bits magic number.
3921 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE &&
3922 get_unaligned_be16(ptr) ==
3923 TCPOPT_FASTOPEN_MAGIC)
3924 tcp_parse_fastopen_option(opsize -
3925 TCPOLEN_EXP_FASTOPEN_BASE,
3926 ptr + 2, th->syn, foc, true);
3928 smc_parse_options(th, opt_rx, ptr,
3938 EXPORT_SYMBOL(tcp_parse_options);
3940 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
3942 const __be32 *ptr = (const __be32 *)(th + 1);
3944 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3945 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3946 tp->rx_opt.saw_tstamp = 1;
3948 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3951 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
3953 tp->rx_opt.rcv_tsecr = 0;
3959 /* Fast parse options. This hopes to only see timestamps.
3960 * If it is wrong it falls back on tcp_parse_options().
3962 static bool tcp_fast_parse_options(const struct net *net,
3963 const struct sk_buff *skb,
3964 const struct tcphdr *th, struct tcp_sock *tp)
3966 /* In the spirit of fast parsing, compare doff directly to constant
3967 * values. Because equality is used, short doff can be ignored here.
3969 if (th->doff == (sizeof(*th) / 4)) {
3970 tp->rx_opt.saw_tstamp = 0;
3972 } else if (tp->rx_opt.tstamp_ok &&
3973 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3974 if (tcp_parse_aligned_timestamp(tp, th))
3978 tcp_parse_options(net, skb, &tp->rx_opt, 1, NULL);
3979 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3980 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
3985 #ifdef CONFIG_TCP_MD5SIG
3987 * Parse MD5 Signature option
3989 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
3991 int length = (th->doff << 2) - sizeof(*th);
3992 const u8 *ptr = (const u8 *)(th + 1);
3994 /* If not enough data remaining, we can short cut */
3995 while (length >= TCPOLEN_MD5SIG) {
3996 int opcode = *ptr++;
4007 if (opsize < 2 || opsize > length)
4009 if (opcode == TCPOPT_MD5SIG)
4010 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
4017 EXPORT_SYMBOL(tcp_parse_md5sig_option);
4020 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
4022 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
4023 * it can pass through stack. So, the following predicate verifies that
4024 * this segment is not used for anything but congestion avoidance or
4025 * fast retransmit. Moreover, we even are able to eliminate most of such
4026 * second order effects, if we apply some small "replay" window (~RTO)
4027 * to timestamp space.
4029 * All these measures still do not guarantee that we reject wrapped ACKs
4030 * on networks with high bandwidth, when sequence space is recycled fastly,
4031 * but it guarantees that such events will be very rare and do not affect
4032 * connection seriously. This doesn't look nice, but alas, PAWS is really
4035 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
4036 * states that events when retransmit arrives after original data are rare.
4037 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
4038 * the biggest problem on large power networks even with minor reordering.
4039 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
4040 * up to bandwidth of 18Gigabit/sec. 8) ]
4043 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
4045 const struct tcp_sock *tp = tcp_sk(sk);
4046 const struct tcphdr *th = tcp_hdr(skb);
4047 u32 seq = TCP_SKB_CB(skb)->seq;
4048 u32 ack = TCP_SKB_CB(skb)->ack_seq;
4050 return (/* 1. Pure ACK with correct sequence number. */
4051 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
4053 /* 2. ... and duplicate ACK. */
4054 ack == tp->snd_una &&
4056 /* 3. ... and does not update window. */
4057 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
4059 /* 4. ... and sits in replay window. */
4060 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
4063 static inline bool tcp_paws_discard(const struct sock *sk,
4064 const struct sk_buff *skb)
4066 const struct tcp_sock *tp = tcp_sk(sk);
4068 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
4069 !tcp_disordered_ack(sk, skb);
4072 /* Check segment sequence number for validity.
4074 * Segment controls are considered valid, if the segment
4075 * fits to the window after truncation to the window. Acceptability
4076 * of data (and SYN, FIN, of course) is checked separately.
4077 * See tcp_data_queue(), for example.
4079 * Also, controls (RST is main one) are accepted using RCV.WUP instead
4080 * of RCV.NXT. Peer still did not advance his SND.UNA when we
4081 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4082 * (borrowed from freebsd)
4085 static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
4087 return !before(end_seq, tp->rcv_wup) &&
4088 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
4091 /* When we get a reset we do this. */
4092 void tcp_reset(struct sock *sk)
4094 trace_tcp_receive_reset(sk);
4096 /* We want the right error as BSD sees it (and indeed as we do). */
4097 switch (sk->sk_state) {
4099 sk->sk_err = ECONNREFUSED;
4101 case TCP_CLOSE_WAIT:
4107 sk->sk_err = ECONNRESET;
4109 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4112 tcp_write_queue_purge(sk);
4115 if (!sock_flag(sk, SOCK_DEAD))
4116 sk->sk_error_report(sk);
4120 * Process the FIN bit. This now behaves as it is supposed to work
4121 * and the FIN takes effect when it is validly part of sequence
4122 * space. Not before when we get holes.
4124 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4125 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4128 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4129 * close and we go into CLOSING (and later onto TIME-WAIT)
4131 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4133 void tcp_fin(struct sock *sk)
4135 struct tcp_sock *tp = tcp_sk(sk);
4137 inet_csk_schedule_ack(sk);
4139 sk->sk_shutdown |= RCV_SHUTDOWN;
4140 sock_set_flag(sk, SOCK_DONE);
4142 switch (sk->sk_state) {
4144 case TCP_ESTABLISHED:
4145 /* Move to CLOSE_WAIT */
4146 tcp_set_state(sk, TCP_CLOSE_WAIT);
4147 inet_csk_enter_pingpong_mode(sk);
4150 case TCP_CLOSE_WAIT:
4152 /* Received a retransmission of the FIN, do
4157 /* RFC793: Remain in the LAST-ACK state. */
4161 /* This case occurs when a simultaneous close
4162 * happens, we must ack the received FIN and
4163 * enter the CLOSING state.
4166 tcp_set_state(sk, TCP_CLOSING);
4169 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4171 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4174 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4175 * cases we should never reach this piece of code.
4177 pr_err("%s: Impossible, sk->sk_state=%d\n",
4178 __func__, sk->sk_state);
4182 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4183 * Probably, we should reset in this case. For now drop them.
4185 skb_rbtree_purge(&tp->out_of_order_queue);
4186 if (tcp_is_sack(tp))
4187 tcp_sack_reset(&tp->rx_opt);
4190 if (!sock_flag(sk, SOCK_DEAD)) {
4191 sk->sk_state_change(sk);
4193 /* Do not send POLL_HUP for half duplex close. */
4194 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4195 sk->sk_state == TCP_CLOSE)
4196 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4198 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4202 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4205 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4206 if (before(seq, sp->start_seq))
4207 sp->start_seq = seq;
4208 if (after(end_seq, sp->end_seq))
4209 sp->end_seq = end_seq;
4215 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4217 struct tcp_sock *tp = tcp_sk(sk);
4219 if (tcp_is_sack(tp) && sock_net(sk)->ipv4.sysctl_tcp_dsack) {
4222 if (before(seq, tp->rcv_nxt))
4223 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4225 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4227 NET_INC_STATS(sock_net(sk), mib_idx);
4229 tp->rx_opt.dsack = 1;
4230 tp->duplicate_sack[0].start_seq = seq;
4231 tp->duplicate_sack[0].end_seq = end_seq;
4235 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4237 struct tcp_sock *tp = tcp_sk(sk);
4239 if (!tp->rx_opt.dsack)
4240 tcp_dsack_set(sk, seq, end_seq);
4242 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4245 static void tcp_rcv_spurious_retrans(struct sock *sk, const struct sk_buff *skb)
4247 /* When the ACK path fails or drops most ACKs, the sender would
4248 * timeout and spuriously retransmit the same segment repeatedly.
4249 * The receiver remembers and reflects via DSACKs. Leverage the
4250 * DSACK state and change the txhash to re-route speculatively.
4252 if (TCP_SKB_CB(skb)->seq == tcp_sk(sk)->duplicate_sack[0].start_seq) {
4253 sk_rethink_txhash(sk);
4254 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDUPLICATEDATAREHASH);
4258 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4260 struct tcp_sock *tp = tcp_sk(sk);
4262 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4263 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4264 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4265 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
4267 if (tcp_is_sack(tp) && sock_net(sk)->ipv4.sysctl_tcp_dsack) {
4268 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4270 tcp_rcv_spurious_retrans(sk, skb);
4271 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4272 end_seq = tp->rcv_nxt;
4273 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4280 /* These routines update the SACK block as out-of-order packets arrive or
4281 * in-order packets close up the sequence space.
4283 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4286 struct tcp_sack_block *sp = &tp->selective_acks[0];
4287 struct tcp_sack_block *swalk = sp + 1;
4289 /* See if the recent change to the first SACK eats into
4290 * or hits the sequence space of other SACK blocks, if so coalesce.
4292 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4293 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4296 /* Zap SWALK, by moving every further SACK up by one slot.
4297 * Decrease num_sacks.
4299 tp->rx_opt.num_sacks--;
4300 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4304 this_sack++, swalk++;
4308 static void tcp_sack_compress_send_ack(struct sock *sk)
4310 struct tcp_sock *tp = tcp_sk(sk);
4312 if (!tp->compressed_ack)
4315 if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1)
4318 /* Since we have to send one ack finally,
4319 * substract one from tp->compressed_ack to keep
4320 * LINUX_MIB_TCPACKCOMPRESSED accurate.
4322 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED,
4323 tp->compressed_ack - 1);
4325 tp->compressed_ack = 0;
4329 /* Reasonable amount of sack blocks included in TCP SACK option
4330 * The max is 4, but this becomes 3 if TCP timestamps are there.
4331 * Given that SACK packets might be lost, be conservative and use 2.
4333 #define TCP_SACK_BLOCKS_EXPECTED 2
4335 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4337 struct tcp_sock *tp = tcp_sk(sk);
4338 struct tcp_sack_block *sp = &tp->selective_acks[0];
4339 int cur_sacks = tp->rx_opt.num_sacks;
4345 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4346 if (tcp_sack_extend(sp, seq, end_seq)) {
4347 if (this_sack >= TCP_SACK_BLOCKS_EXPECTED)
4348 tcp_sack_compress_send_ack(sk);
4349 /* Rotate this_sack to the first one. */
4350 for (; this_sack > 0; this_sack--, sp--)
4351 swap(*sp, *(sp - 1));
4353 tcp_sack_maybe_coalesce(tp);
4358 if (this_sack >= TCP_SACK_BLOCKS_EXPECTED)
4359 tcp_sack_compress_send_ack(sk);
4361 /* Could not find an adjacent existing SACK, build a new one,
4362 * put it at the front, and shift everyone else down. We
4363 * always know there is at least one SACK present already here.
4365 * If the sack array is full, forget about the last one.
4367 if (this_sack >= TCP_NUM_SACKS) {
4369 tp->rx_opt.num_sacks--;
4372 for (; this_sack > 0; this_sack--, sp--)
4376 /* Build the new head SACK, and we're done. */
4377 sp->start_seq = seq;
4378 sp->end_seq = end_seq;
4379 tp->rx_opt.num_sacks++;
4382 /* RCV.NXT advances, some SACKs should be eaten. */
4384 static void tcp_sack_remove(struct tcp_sock *tp)
4386 struct tcp_sack_block *sp = &tp->selective_acks[0];
4387 int num_sacks = tp->rx_opt.num_sacks;
4390 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4391 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4392 tp->rx_opt.num_sacks = 0;
4396 for (this_sack = 0; this_sack < num_sacks;) {
4397 /* Check if the start of the sack is covered by RCV.NXT. */
4398 if (!before(tp->rcv_nxt, sp->start_seq)) {
4401 /* RCV.NXT must cover all the block! */
4402 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4404 /* Zap this SACK, by moving forward any other SACKS. */
4405 for (i = this_sack+1; i < num_sacks; i++)
4406 tp->selective_acks[i-1] = tp->selective_acks[i];
4413 tp->rx_opt.num_sacks = num_sacks;
4417 * tcp_try_coalesce - try to merge skb to prior one
4419 * @dest: destination queue
4421 * @from: buffer to add in queue
4422 * @fragstolen: pointer to boolean
4424 * Before queueing skb @from after @to, try to merge them
4425 * to reduce overall memory use and queue lengths, if cost is small.
4426 * Packets in ofo or receive queues can stay a long time.
4427 * Better try to coalesce them right now to avoid future collapses.
4428 * Returns true if caller should free @from instead of queueing it
4430 static bool tcp_try_coalesce(struct sock *sk,
4432 struct sk_buff *from,
4437 *fragstolen = false;
4439 /* Its possible this segment overlaps with prior segment in queue */
4440 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4443 if (!mptcp_skb_can_collapse(to, from))
4446 #ifdef CONFIG_TLS_DEVICE
4447 if (from->decrypted != to->decrypted)
4451 if (!skb_try_coalesce(to, from, fragstolen, &delta))
4454 atomic_add(delta, &sk->sk_rmem_alloc);
4455 sk_mem_charge(sk, delta);
4456 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4457 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4458 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4459 TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
4461 if (TCP_SKB_CB(from)->has_rxtstamp) {
4462 TCP_SKB_CB(to)->has_rxtstamp = true;
4463 to->tstamp = from->tstamp;
4464 skb_hwtstamps(to)->hwtstamp = skb_hwtstamps(from)->hwtstamp;
4470 static bool tcp_ooo_try_coalesce(struct sock *sk,
4472 struct sk_buff *from,
4475 bool res = tcp_try_coalesce(sk, to, from, fragstolen);
4477 /* In case tcp_drop() is called later, update to->gso_segs */
4479 u32 gso_segs = max_t(u16, 1, skb_shinfo(to)->gso_segs) +
4480 max_t(u16, 1, skb_shinfo(from)->gso_segs);
4482 skb_shinfo(to)->gso_segs = min_t(u32, gso_segs, 0xFFFF);
4487 static void tcp_drop(struct sock *sk, struct sk_buff *skb)
4489 sk_drops_add(sk, skb);
4493 /* This one checks to see if we can put data from the
4494 * out_of_order queue into the receive_queue.
4496 static void tcp_ofo_queue(struct sock *sk)
4498 struct tcp_sock *tp = tcp_sk(sk);
4499 __u32 dsack_high = tp->rcv_nxt;
4500 bool fin, fragstolen, eaten;
4501 struct sk_buff *skb, *tail;
4504 p = rb_first(&tp->out_of_order_queue);
4507 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4510 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4511 __u32 dsack = dsack_high;
4512 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4513 dsack_high = TCP_SKB_CB(skb)->end_seq;
4514 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4517 rb_erase(&skb->rbnode, &tp->out_of_order_queue);
4519 if (unlikely(!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))) {
4524 tail = skb_peek_tail(&sk->sk_receive_queue);
4525 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen);
4526 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4527 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
4529 __skb_queue_tail(&sk->sk_receive_queue, skb);
4531 kfree_skb_partial(skb, fragstolen);
4533 if (unlikely(fin)) {
4535 /* tcp_fin() purges tp->out_of_order_queue,
4536 * so we must end this loop right now.
4543 static bool tcp_prune_ofo_queue(struct sock *sk);
4544 static int tcp_prune_queue(struct sock *sk);
4546 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4549 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4550 !sk_rmem_schedule(sk, skb, size)) {
4552 if (tcp_prune_queue(sk) < 0)
4555 while (!sk_rmem_schedule(sk, skb, size)) {
4556 if (!tcp_prune_ofo_queue(sk))
4563 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4565 struct tcp_sock *tp = tcp_sk(sk);
4566 struct rb_node **p, *parent;
4567 struct sk_buff *skb1;
4571 tcp_ecn_check_ce(sk, skb);
4573 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4574 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFODROP);
4579 /* Disable header prediction. */
4581 inet_csk_schedule_ack(sk);
4583 tp->rcv_ooopack += max_t(u16, 1, skb_shinfo(skb)->gso_segs);
4584 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4585 seq = TCP_SKB_CB(skb)->seq;
4586 end_seq = TCP_SKB_CB(skb)->end_seq;
4588 p = &tp->out_of_order_queue.rb_node;
4589 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4590 /* Initial out of order segment, build 1 SACK. */
4591 if (tcp_is_sack(tp)) {
4592 tp->rx_opt.num_sacks = 1;
4593 tp->selective_acks[0].start_seq = seq;
4594 tp->selective_acks[0].end_seq = end_seq;
4596 rb_link_node(&skb->rbnode, NULL, p);
4597 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4598 tp->ooo_last_skb = skb;
4602 /* In the typical case, we are adding an skb to the end of the list.
4603 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
4605 if (tcp_ooo_try_coalesce(sk, tp->ooo_last_skb,
4606 skb, &fragstolen)) {
4608 tcp_grow_window(sk, skb);
4609 kfree_skb_partial(skb, fragstolen);
4613 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
4614 if (!before(seq, TCP_SKB_CB(tp->ooo_last_skb)->end_seq)) {
4615 parent = &tp->ooo_last_skb->rbnode;
4616 p = &parent->rb_right;
4620 /* Find place to insert this segment. Handle overlaps on the way. */
4624 skb1 = rb_to_skb(parent);
4625 if (before(seq, TCP_SKB_CB(skb1)->seq)) {
4626 p = &parent->rb_left;
4629 if (before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4630 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4631 /* All the bits are present. Drop. */
4632 NET_INC_STATS(sock_net(sk),
4633 LINUX_MIB_TCPOFOMERGE);
4636 tcp_dsack_set(sk, seq, end_seq);
4639 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4640 /* Partial overlap. */
4641 tcp_dsack_set(sk, seq, TCP_SKB_CB(skb1)->end_seq);
4643 /* skb's seq == skb1's seq and skb covers skb1.
4644 * Replace skb1 with skb.
4646 rb_replace_node(&skb1->rbnode, &skb->rbnode,
4647 &tp->out_of_order_queue);
4648 tcp_dsack_extend(sk,
4649 TCP_SKB_CB(skb1)->seq,
4650 TCP_SKB_CB(skb1)->end_seq);
4651 NET_INC_STATS(sock_net(sk),
4652 LINUX_MIB_TCPOFOMERGE);
4656 } else if (tcp_ooo_try_coalesce(sk, skb1,
4657 skb, &fragstolen)) {
4660 p = &parent->rb_right;
4663 /* Insert segment into RB tree. */
4664 rb_link_node(&skb->rbnode, parent, p);
4665 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4668 /* Remove other segments covered by skb. */
4669 while ((skb1 = skb_rb_next(skb)) != NULL) {
4670 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4672 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4673 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4677 rb_erase(&skb1->rbnode, &tp->out_of_order_queue);
4678 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4679 TCP_SKB_CB(skb1)->end_seq);
4680 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4683 /* If there is no skb after us, we are the last_skb ! */
4685 tp->ooo_last_skb = skb;
4688 if (tcp_is_sack(tp))
4689 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4692 tcp_grow_window(sk, skb);
4694 skb_set_owner_r(skb, sk);
4698 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb,
4702 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4705 tcp_try_coalesce(sk, tail,
4706 skb, fragstolen)) ? 1 : 0;
4707 tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq);
4709 __skb_queue_tail(&sk->sk_receive_queue, skb);
4710 skb_set_owner_r(skb, sk);
4715 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4717 struct sk_buff *skb;
4725 if (size > PAGE_SIZE) {
4726 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS);
4728 data_len = npages << PAGE_SHIFT;
4729 size = data_len + (size & ~PAGE_MASK);
4731 skb = alloc_skb_with_frags(size - data_len, data_len,
4732 PAGE_ALLOC_COSTLY_ORDER,
4733 &err, sk->sk_allocation);
4737 skb_put(skb, size - data_len);
4738 skb->data_len = data_len;
4741 if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) {
4742 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP);
4746 err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size);
4750 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4751 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4752 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4754 if (tcp_queue_rcv(sk, skb, &fragstolen)) {
4755 WARN_ON_ONCE(fragstolen); /* should not happen */
4767 void tcp_data_ready(struct sock *sk)
4769 const struct tcp_sock *tp = tcp_sk(sk);
4770 int avail = tp->rcv_nxt - tp->copied_seq;
4772 if (avail < sk->sk_rcvlowat && !tcp_rmem_pressure(sk) &&
4773 !sock_flag(sk, SOCK_DONE))
4776 sk->sk_data_ready(sk);
4779 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4781 struct tcp_sock *tp = tcp_sk(sk);
4785 if (sk_is_mptcp(sk))
4786 mptcp_incoming_options(sk, skb, &tp->rx_opt);
4788 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
4793 __skb_pull(skb, tcp_hdr(skb)->doff * 4);
4795 tcp_ecn_accept_cwr(sk, skb);
4797 tp->rx_opt.dsack = 0;
4799 /* Queue data for delivery to the user.
4800 * Packets in sequence go to the receive queue.
4801 * Out of sequence packets to the out_of_order_queue.
4803 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4804 if (tcp_receive_window(tp) == 0) {
4805 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP);
4809 /* Ok. In sequence. In window. */
4811 if (skb_queue_len(&sk->sk_receive_queue) == 0)
4812 sk_forced_mem_schedule(sk, skb->truesize);
4813 else if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) {
4814 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP);
4818 eaten = tcp_queue_rcv(sk, skb, &fragstolen);
4820 tcp_event_data_recv(sk, skb);
4821 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
4824 if (!RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4827 /* RFC5681. 4.2. SHOULD send immediate ACK, when
4828 * gap in queue is filled.
4830 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
4831 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW;
4834 if (tp->rx_opt.num_sacks)
4835 tcp_sack_remove(tp);
4837 tcp_fast_path_check(sk);
4840 kfree_skb_partial(skb, fragstolen);
4841 if (!sock_flag(sk, SOCK_DEAD))
4846 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4847 tcp_rcv_spurious_retrans(sk, skb);
4848 /* A retransmit, 2nd most common case. Force an immediate ack. */
4849 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4850 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4853 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
4854 inet_csk_schedule_ack(sk);
4860 /* Out of window. F.e. zero window probe. */
4861 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4864 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4865 /* Partial packet, seq < rcv_next < end_seq */
4866 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4868 /* If window is closed, drop tail of packet. But after
4869 * remembering D-SACK for its head made in previous line.
4871 if (!tcp_receive_window(tp)) {
4872 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP);
4878 tcp_data_queue_ofo(sk, skb);
4881 static struct sk_buff *tcp_skb_next(struct sk_buff *skb, struct sk_buff_head *list)
4884 return !skb_queue_is_last(list, skb) ? skb->next : NULL;
4886 return skb_rb_next(skb);
4889 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4890 struct sk_buff_head *list,
4891 struct rb_root *root)
4893 struct sk_buff *next = tcp_skb_next(skb, list);
4896 __skb_unlink(skb, list);
4898 rb_erase(&skb->rbnode, root);
4901 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4906 /* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */
4907 void tcp_rbtree_insert(struct rb_root *root, struct sk_buff *skb)
4909 struct rb_node **p = &root->rb_node;
4910 struct rb_node *parent = NULL;
4911 struct sk_buff *skb1;
4915 skb1 = rb_to_skb(parent);
4916 if (before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb1)->seq))
4917 p = &parent->rb_left;
4919 p = &parent->rb_right;
4921 rb_link_node(&skb->rbnode, parent, p);
4922 rb_insert_color(&skb->rbnode, root);
4925 /* Collapse contiguous sequence of skbs head..tail with
4926 * sequence numbers start..end.
4928 * If tail is NULL, this means until the end of the queue.
4930 * Segments with FIN/SYN are not collapsed (only because this
4934 tcp_collapse(struct sock *sk, struct sk_buff_head *list, struct rb_root *root,
4935 struct sk_buff *head, struct sk_buff *tail, u32 start, u32 end)
4937 struct sk_buff *skb = head, *n;
4938 struct sk_buff_head tmp;
4941 /* First, check that queue is collapsible and find
4942 * the point where collapsing can be useful.
4945 for (end_of_skbs = true; skb != NULL && skb != tail; skb = n) {
4946 n = tcp_skb_next(skb, list);
4948 /* No new bits? It is possible on ofo queue. */
4949 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4950 skb = tcp_collapse_one(sk, skb, list, root);
4956 /* The first skb to collapse is:
4958 * - bloated or contains data before "start" or
4959 * overlaps to the next one and mptcp allow collapsing.
4961 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
4962 (tcp_win_from_space(sk, skb->truesize) > skb->len ||
4963 before(TCP_SKB_CB(skb)->seq, start))) {
4964 end_of_skbs = false;
4968 if (n && n != tail && mptcp_skb_can_collapse(skb, n) &&
4969 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(n)->seq) {
4970 end_of_skbs = false;
4974 /* Decided to skip this, advance start seq. */
4975 start = TCP_SKB_CB(skb)->end_seq;
4978 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4981 __skb_queue_head_init(&tmp);
4983 while (before(start, end)) {
4984 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
4985 struct sk_buff *nskb;
4987 nskb = alloc_skb(copy, GFP_ATOMIC);
4991 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4992 #ifdef CONFIG_TLS_DEVICE
4993 nskb->decrypted = skb->decrypted;
4995 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4997 __skb_queue_before(list, skb, nskb);
4999 __skb_queue_tail(&tmp, nskb); /* defer rbtree insertion */
5000 skb_set_owner_r(nskb, sk);
5001 mptcp_skb_ext_move(nskb, skb);
5003 /* Copy data, releasing collapsed skbs. */
5005 int offset = start - TCP_SKB_CB(skb)->seq;
5006 int size = TCP_SKB_CB(skb)->end_seq - start;
5010 size = min(copy, size);
5011 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
5013 TCP_SKB_CB(nskb)->end_seq += size;
5017 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
5018 skb = tcp_collapse_one(sk, skb, list, root);
5021 !mptcp_skb_can_collapse(nskb, skb) ||
5022 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
5024 #ifdef CONFIG_TLS_DEVICE
5025 if (skb->decrypted != nskb->decrypted)
5032 skb_queue_walk_safe(&tmp, skb, n)
5033 tcp_rbtree_insert(root, skb);
5036 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
5037 * and tcp_collapse() them until all the queue is collapsed.
5039 static void tcp_collapse_ofo_queue(struct sock *sk)
5041 struct tcp_sock *tp = tcp_sk(sk);
5042 u32 range_truesize, sum_tiny = 0;
5043 struct sk_buff *skb, *head;
5046 skb = skb_rb_first(&tp->out_of_order_queue);
5049 tp->ooo_last_skb = skb_rb_last(&tp->out_of_order_queue);
5052 start = TCP_SKB_CB(skb)->seq;
5053 end = TCP_SKB_CB(skb)->end_seq;
5054 range_truesize = skb->truesize;
5056 for (head = skb;;) {
5057 skb = skb_rb_next(skb);
5059 /* Range is terminated when we see a gap or when
5060 * we are at the queue end.
5063 after(TCP_SKB_CB(skb)->seq, end) ||
5064 before(TCP_SKB_CB(skb)->end_seq, start)) {
5065 /* Do not attempt collapsing tiny skbs */
5066 if (range_truesize != head->truesize ||
5067 end - start >= SKB_WITH_OVERHEAD(SK_MEM_QUANTUM)) {
5068 tcp_collapse(sk, NULL, &tp->out_of_order_queue,
5069 head, skb, start, end);
5071 sum_tiny += range_truesize;
5072 if (sum_tiny > sk->sk_rcvbuf >> 3)
5078 range_truesize += skb->truesize;
5079 if (unlikely(before(TCP_SKB_CB(skb)->seq, start)))
5080 start = TCP_SKB_CB(skb)->seq;
5081 if (after(TCP_SKB_CB(skb)->end_seq, end))
5082 end = TCP_SKB_CB(skb)->end_seq;
5087 * Clean the out-of-order queue to make room.
5088 * We drop high sequences packets to :
5089 * 1) Let a chance for holes to be filled.
5090 * 2) not add too big latencies if thousands of packets sit there.
5091 * (But if application shrinks SO_RCVBUF, we could still end up
5092 * freeing whole queue here)
5093 * 3) Drop at least 12.5 % of sk_rcvbuf to avoid malicious attacks.
5095 * Return true if queue has shrunk.
5097 static bool tcp_prune_ofo_queue(struct sock *sk)
5099 struct tcp_sock *tp = tcp_sk(sk);
5100 struct rb_node *node, *prev;
5103 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
5106 NET_INC_STATS(sock_net(sk), LINUX_MIB_OFOPRUNED);
5107 goal = sk->sk_rcvbuf >> 3;
5108 node = &tp->ooo_last_skb->rbnode;
5110 prev = rb_prev(node);
5111 rb_erase(node, &tp->out_of_order_queue);
5112 goal -= rb_to_skb(node)->truesize;
5113 tcp_drop(sk, rb_to_skb(node));
5114 if (!prev || goal <= 0) {
5116 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf &&
5117 !tcp_under_memory_pressure(sk))
5119 goal = sk->sk_rcvbuf >> 3;
5123 tp->ooo_last_skb = rb_to_skb(prev);
5125 /* Reset SACK state. A conforming SACK implementation will
5126 * do the same at a timeout based retransmit. When a connection
5127 * is in a sad state like this, we care only about integrity
5128 * of the connection not performance.
5130 if (tp->rx_opt.sack_ok)
5131 tcp_sack_reset(&tp->rx_opt);
5135 /* Reduce allocated memory if we can, trying to get
5136 * the socket within its memory limits again.
5138 * Return less than zero if we should start dropping frames
5139 * until the socket owning process reads some of the data
5140 * to stabilize the situation.
5142 static int tcp_prune_queue(struct sock *sk)
5144 struct tcp_sock *tp = tcp_sk(sk);
5146 NET_INC_STATS(sock_net(sk), LINUX_MIB_PRUNECALLED);
5148 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
5149 tcp_clamp_window(sk);
5150 else if (tcp_under_memory_pressure(sk))
5151 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
5153 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5156 tcp_collapse_ofo_queue(sk);
5157 if (!skb_queue_empty(&sk->sk_receive_queue))
5158 tcp_collapse(sk, &sk->sk_receive_queue, NULL,
5159 skb_peek(&sk->sk_receive_queue),
5161 tp->copied_seq, tp->rcv_nxt);
5164 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5167 /* Collapsing did not help, destructive actions follow.
5168 * This must not ever occur. */
5170 tcp_prune_ofo_queue(sk);
5172 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5175 /* If we are really being abused, tell the caller to silently
5176 * drop receive data on the floor. It will get retransmitted
5177 * and hopefully then we'll have sufficient space.
5179 NET_INC_STATS(sock_net(sk), LINUX_MIB_RCVPRUNED);
5181 /* Massive buffer overcommit. */
5186 static bool tcp_should_expand_sndbuf(const struct sock *sk)
5188 const struct tcp_sock *tp = tcp_sk(sk);
5190 /* If the user specified a specific send buffer setting, do
5193 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
5196 /* If we are under global TCP memory pressure, do not expand. */
5197 if (tcp_under_memory_pressure(sk))
5200 /* If we are under soft global TCP memory pressure, do not expand. */
5201 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
5204 /* If we filled the congestion window, do not expand. */
5205 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
5211 /* When incoming ACK allowed to free some skb from write_queue,
5212 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
5213 * on the exit from tcp input handler.
5215 * PROBLEM: sndbuf expansion does not work well with largesend.
5217 static void tcp_new_space(struct sock *sk)
5219 struct tcp_sock *tp = tcp_sk(sk);
5221 if (tcp_should_expand_sndbuf(sk)) {
5222 tcp_sndbuf_expand(sk);
5223 tp->snd_cwnd_stamp = tcp_jiffies32;
5226 sk->sk_write_space(sk);
5229 static void tcp_check_space(struct sock *sk)
5231 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
5232 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
5233 /* pairs with tcp_poll() */
5235 if (sk->sk_socket &&
5236 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
5238 if (!test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
5239 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
5244 static inline void tcp_data_snd_check(struct sock *sk)
5246 tcp_push_pending_frames(sk);
5247 tcp_check_space(sk);
5251 * Check if sending an ack is needed.
5253 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
5255 struct tcp_sock *tp = tcp_sk(sk);
5256 unsigned long rtt, delay;
5258 /* More than one full frame received... */
5259 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
5260 /* ... and right edge of window advances far enough.
5261 * (tcp_recvmsg() will send ACK otherwise).
5262 * If application uses SO_RCVLOWAT, we want send ack now if
5263 * we have not received enough bytes to satisfy the condition.
5265 (tp->rcv_nxt - tp->copied_seq < sk->sk_rcvlowat ||
5266 __tcp_select_window(sk) >= tp->rcv_wnd)) ||
5267 /* We ACK each frame or... */
5268 tcp_in_quickack_mode(sk) ||
5269 /* Protocol state mandates a one-time immediate ACK */
5270 inet_csk(sk)->icsk_ack.pending & ICSK_ACK_NOW) {
5276 if (!ofo_possible || RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
5277 tcp_send_delayed_ack(sk);
5281 if (!tcp_is_sack(tp) ||
5282 tp->compressed_ack >= sock_net(sk)->ipv4.sysctl_tcp_comp_sack_nr)
5285 if (tp->compressed_ack_rcv_nxt != tp->rcv_nxt) {
5286 tp->compressed_ack_rcv_nxt = tp->rcv_nxt;
5287 tp->dup_ack_counter = 0;
5289 if (tp->dup_ack_counter < TCP_FASTRETRANS_THRESH) {
5290 tp->dup_ack_counter++;
5293 tp->compressed_ack++;
5294 if (hrtimer_is_queued(&tp->compressed_ack_timer))
5297 /* compress ack timer : 5 % of rtt, but no more than tcp_comp_sack_delay_ns */
5299 rtt = tp->rcv_rtt_est.rtt_us;
5300 if (tp->srtt_us && tp->srtt_us < rtt)
5303 delay = min_t(unsigned long, sock_net(sk)->ipv4.sysctl_tcp_comp_sack_delay_ns,
5304 rtt * (NSEC_PER_USEC >> 3)/20);
5306 hrtimer_start_range_ns(&tp->compressed_ack_timer, ns_to_ktime(delay),
5307 sock_net(sk)->ipv4.sysctl_tcp_comp_sack_slack_ns,
5308 HRTIMER_MODE_REL_PINNED_SOFT);
5311 static inline void tcp_ack_snd_check(struct sock *sk)
5313 if (!inet_csk_ack_scheduled(sk)) {
5314 /* We sent a data segment already. */
5317 __tcp_ack_snd_check(sk, 1);
5321 * This routine is only called when we have urgent data
5322 * signaled. Its the 'slow' part of tcp_urg. It could be
5323 * moved inline now as tcp_urg is only called from one
5324 * place. We handle URGent data wrong. We have to - as
5325 * BSD still doesn't use the correction from RFC961.
5326 * For 1003.1g we should support a new option TCP_STDURG to permit
5327 * either form (or just set the sysctl tcp_stdurg).
5330 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
5332 struct tcp_sock *tp = tcp_sk(sk);
5333 u32 ptr = ntohs(th->urg_ptr);
5335 if (ptr && !sock_net(sk)->ipv4.sysctl_tcp_stdurg)
5337 ptr += ntohl(th->seq);
5339 /* Ignore urgent data that we've already seen and read. */
5340 if (after(tp->copied_seq, ptr))
5343 /* Do not replay urg ptr.
5345 * NOTE: interesting situation not covered by specs.
5346 * Misbehaving sender may send urg ptr, pointing to segment,
5347 * which we already have in ofo queue. We are not able to fetch
5348 * such data and will stay in TCP_URG_NOTYET until will be eaten
5349 * by recvmsg(). Seems, we are not obliged to handle such wicked
5350 * situations. But it is worth to think about possibility of some
5351 * DoSes using some hypothetical application level deadlock.
5353 if (before(ptr, tp->rcv_nxt))
5356 /* Do we already have a newer (or duplicate) urgent pointer? */
5357 if (tp->urg_data && !after(ptr, tp->urg_seq))
5360 /* Tell the world about our new urgent pointer. */
5363 /* We may be adding urgent data when the last byte read was
5364 * urgent. To do this requires some care. We cannot just ignore
5365 * tp->copied_seq since we would read the last urgent byte again
5366 * as data, nor can we alter copied_seq until this data arrives
5367 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5369 * NOTE. Double Dutch. Rendering to plain English: author of comment
5370 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5371 * and expect that both A and B disappear from stream. This is _wrong_.
5372 * Though this happens in BSD with high probability, this is occasional.
5373 * Any application relying on this is buggy. Note also, that fix "works"
5374 * only in this artificial test. Insert some normal data between A and B and we will
5375 * decline of BSD again. Verdict: it is better to remove to trap
5378 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5379 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5380 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5382 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5383 __skb_unlink(skb, &sk->sk_receive_queue);
5388 tp->urg_data = TCP_URG_NOTYET;
5389 WRITE_ONCE(tp->urg_seq, ptr);
5391 /* Disable header prediction. */
5395 /* This is the 'fast' part of urgent handling. */
5396 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
5398 struct tcp_sock *tp = tcp_sk(sk);
5400 /* Check if we get a new urgent pointer - normally not. */
5402 tcp_check_urg(sk, th);
5404 /* Do we wait for any urgent data? - normally not... */
5405 if (tp->urg_data == TCP_URG_NOTYET) {
5406 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5409 /* Is the urgent pointer pointing into this packet? */
5410 if (ptr < skb->len) {
5412 if (skb_copy_bits(skb, ptr, &tmp, 1))
5414 tp->urg_data = TCP_URG_VALID | tmp;
5415 if (!sock_flag(sk, SOCK_DEAD))
5416 sk->sk_data_ready(sk);
5421 /* Accept RST for rcv_nxt - 1 after a FIN.
5422 * When tcp connections are abruptly terminated from Mac OSX (via ^C), a
5423 * FIN is sent followed by a RST packet. The RST is sent with the same
5424 * sequence number as the FIN, and thus according to RFC 5961 a challenge
5425 * ACK should be sent. However, Mac OSX rate limits replies to challenge
5426 * ACKs on the closed socket. In addition middleboxes can drop either the
5427 * challenge ACK or a subsequent RST.
5429 static bool tcp_reset_check(const struct sock *sk, const struct sk_buff *skb)
5431 struct tcp_sock *tp = tcp_sk(sk);
5433 return unlikely(TCP_SKB_CB(skb)->seq == (tp->rcv_nxt - 1) &&
5434 (1 << sk->sk_state) & (TCPF_CLOSE_WAIT | TCPF_LAST_ACK |
5438 /* Does PAWS and seqno based validation of an incoming segment, flags will
5439 * play significant role here.
5441 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5442 const struct tcphdr *th, int syn_inerr)
5444 struct tcp_sock *tp = tcp_sk(sk);
5445 bool rst_seq_match = false;
5447 /* RFC1323: H1. Apply PAWS check first. */
5448 if (tcp_fast_parse_options(sock_net(sk), skb, th, tp) &&
5449 tp->rx_opt.saw_tstamp &&
5450 tcp_paws_discard(sk, skb)) {
5452 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5453 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5454 LINUX_MIB_TCPACKSKIPPEDPAWS,
5455 &tp->last_oow_ack_time))
5456 tcp_send_dupack(sk, skb);
5459 /* Reset is accepted even if it did not pass PAWS. */
5462 /* Step 1: check sequence number */
5463 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5464 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5465 * (RST) segments are validated by checking their SEQ-fields."
5466 * And page 69: "If an incoming segment is not acceptable,
5467 * an acknowledgment should be sent in reply (unless the RST
5468 * bit is set, if so drop the segment and return)".
5473 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5474 LINUX_MIB_TCPACKSKIPPEDSEQ,
5475 &tp->last_oow_ack_time))
5476 tcp_send_dupack(sk, skb);
5477 } else if (tcp_reset_check(sk, skb)) {
5483 /* Step 2: check RST bit */
5485 /* RFC 5961 3.2 (extend to match against (RCV.NXT - 1) after a
5486 * FIN and SACK too if available):
5487 * If seq num matches RCV.NXT or (RCV.NXT - 1) after a FIN, or
5488 * the right-most SACK block,
5490 * RESET the connection
5492 * Send a challenge ACK
5494 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt ||
5495 tcp_reset_check(sk, skb)) {
5496 rst_seq_match = true;
5497 } else if (tcp_is_sack(tp) && tp->rx_opt.num_sacks > 0) {
5498 struct tcp_sack_block *sp = &tp->selective_acks[0];
5499 int max_sack = sp[0].end_seq;
5502 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;
5504 max_sack = after(sp[this_sack].end_seq,
5506 sp[this_sack].end_seq : max_sack;
5509 if (TCP_SKB_CB(skb)->seq == max_sack)
5510 rst_seq_match = true;
5516 /* Disable TFO if RST is out-of-order
5517 * and no data has been received
5518 * for current active TFO socket
5520 if (tp->syn_fastopen && !tp->data_segs_in &&
5521 sk->sk_state == TCP_ESTABLISHED)
5522 tcp_fastopen_active_disable(sk);
5523 tcp_send_challenge_ack(sk, skb);
5528 /* step 3: check security and precedence [ignored] */
5530 /* step 4: Check for a SYN
5531 * RFC 5961 4.2 : Send a challenge ack
5536 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5537 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5538 tcp_send_challenge_ack(sk, skb);
5550 * TCP receive function for the ESTABLISHED state.
5552 * It is split into a fast path and a slow path. The fast path is
5554 * - A zero window was announced from us - zero window probing
5555 * is only handled properly in the slow path.
5556 * - Out of order segments arrived.
5557 * - Urgent data is expected.
5558 * - There is no buffer space left
5559 * - Unexpected TCP flags/window values/header lengths are received
5560 * (detected by checking the TCP header against pred_flags)
5561 * - Data is sent in both directions. Fast path only supports pure senders
5562 * or pure receivers (this means either the sequence number or the ack
5563 * value must stay constant)
5564 * - Unexpected TCP option.
5566 * When these conditions are not satisfied it drops into a standard
5567 * receive procedure patterned after RFC793 to handle all cases.
5568 * The first three cases are guaranteed by proper pred_flags setting,
5569 * the rest is checked inline. Fast processing is turned on in
5570 * tcp_data_queue when everything is OK.
5572 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb)
5574 const struct tcphdr *th = (const struct tcphdr *)skb->data;
5575 struct tcp_sock *tp = tcp_sk(sk);
5576 unsigned int len = skb->len;
5578 /* TCP congestion window tracking */
5579 trace_tcp_probe(sk, skb);
5581 tcp_mstamp_refresh(tp);
5582 if (unlikely(!sk->sk_rx_dst))
5583 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
5585 * Header prediction.
5586 * The code loosely follows the one in the famous
5587 * "30 instruction TCP receive" Van Jacobson mail.
5589 * Van's trick is to deposit buffers into socket queue
5590 * on a device interrupt, to call tcp_recv function
5591 * on the receive process context and checksum and copy
5592 * the buffer to user space. smart...
5594 * Our current scheme is not silly either but we take the
5595 * extra cost of the net_bh soft interrupt processing...
5596 * We do checksum and copy also but from device to kernel.
5599 tp->rx_opt.saw_tstamp = 0;
5601 /* pred_flags is 0xS?10 << 16 + snd_wnd
5602 * if header_prediction is to be made
5603 * 'S' will always be tp->tcp_header_len >> 2
5604 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5605 * turn it off (when there are holes in the receive
5606 * space for instance)
5607 * PSH flag is ignored.
5610 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5611 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5612 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5613 int tcp_header_len = tp->tcp_header_len;
5615 /* Timestamp header prediction: tcp_header_len
5616 * is automatically equal to th->doff*4 due to pred_flags
5620 /* Check timestamp */
5621 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5622 /* No? Slow path! */
5623 if (!tcp_parse_aligned_timestamp(tp, th))
5626 /* If PAWS failed, check it more carefully in slow path */
5627 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5630 /* DO NOT update ts_recent here, if checksum fails
5631 * and timestamp was corrupted part, it will result
5632 * in a hung connection since we will drop all
5633 * future packets due to the PAWS test.
5637 if (len <= tcp_header_len) {
5638 /* Bulk data transfer: sender */
5639 if (len == tcp_header_len) {
5640 /* Predicted packet is in window by definition.
5641 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5642 * Hence, check seq<=rcv_wup reduces to:
5644 if (tcp_header_len ==
5645 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5646 tp->rcv_nxt == tp->rcv_wup)
5647 tcp_store_ts_recent(tp);
5649 /* We know that such packets are checksummed
5652 tcp_ack(sk, skb, 0);
5654 tcp_data_snd_check(sk);
5655 /* When receiving pure ack in fast path, update
5656 * last ts ecr directly instead of calling
5657 * tcp_rcv_rtt_measure_ts()
5659 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr;
5661 } else { /* Header too small */
5662 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5667 bool fragstolen = false;
5669 if (tcp_checksum_complete(skb))
5672 if ((int)skb->truesize > sk->sk_forward_alloc)
5675 /* Predicted packet is in window by definition.
5676 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5677 * Hence, check seq<=rcv_wup reduces to:
5679 if (tcp_header_len ==
5680 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5681 tp->rcv_nxt == tp->rcv_wup)
5682 tcp_store_ts_recent(tp);
5684 tcp_rcv_rtt_measure_ts(sk, skb);
5686 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPHITS);
5688 /* Bulk data transfer: receiver */
5689 __skb_pull(skb, tcp_header_len);
5690 eaten = tcp_queue_rcv(sk, skb, &fragstolen);
5692 tcp_event_data_recv(sk, skb);
5694 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5695 /* Well, only one small jumplet in fast path... */
5696 tcp_ack(sk, skb, FLAG_DATA);
5697 tcp_data_snd_check(sk);
5698 if (!inet_csk_ack_scheduled(sk))
5702 __tcp_ack_snd_check(sk, 0);
5705 kfree_skb_partial(skb, fragstolen);
5712 if (len < (th->doff << 2) || tcp_checksum_complete(skb))
5715 if (!th->ack && !th->rst && !th->syn)
5719 * Standard slow path.
5722 if (!tcp_validate_incoming(sk, skb, th, 1))
5726 if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0)
5729 tcp_rcv_rtt_measure_ts(sk, skb);
5731 /* Process urgent data. */
5732 tcp_urg(sk, skb, th);
5734 /* step 7: process the segment text */
5735 tcp_data_queue(sk, skb);
5737 tcp_data_snd_check(sk);
5738 tcp_ack_snd_check(sk);
5742 TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS);
5743 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5748 EXPORT_SYMBOL(tcp_rcv_established);
5750 void tcp_init_transfer(struct sock *sk, int bpf_op)
5752 struct inet_connection_sock *icsk = inet_csk(sk);
5753 struct tcp_sock *tp = tcp_sk(sk);
5756 icsk->icsk_af_ops->rebuild_header(sk);
5757 tcp_init_metrics(sk);
5759 /* Initialize the congestion window to start the transfer.
5760 * Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
5761 * retransmitted. In light of RFC6298 more aggressive 1sec
5762 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
5763 * retransmission has occurred.
5765 if (tp->total_retrans > 1 && tp->undo_marker)
5768 tp->snd_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk));
5769 tp->snd_cwnd_stamp = tcp_jiffies32;
5771 tcp_call_bpf(sk, bpf_op, 0, NULL);
5772 tcp_init_congestion_control(sk);
5773 tcp_init_buffer_space(sk);
5776 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5778 struct tcp_sock *tp = tcp_sk(sk);
5779 struct inet_connection_sock *icsk = inet_csk(sk);
5781 tcp_set_state(sk, TCP_ESTABLISHED);
5782 icsk->icsk_ack.lrcvtime = tcp_jiffies32;
5785 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
5786 security_inet_conn_established(sk, skb);
5787 sk_mark_napi_id(sk, skb);
5790 tcp_init_transfer(sk, BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB);
5792 /* Prevent spurious tcp_cwnd_restart() on first data
5795 tp->lsndtime = tcp_jiffies32;
5797 if (sock_flag(sk, SOCK_KEEPOPEN))
5798 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5800 if (!tp->rx_opt.snd_wscale)
5801 __tcp_fast_path_on(tp, tp->snd_wnd);
5806 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
5807 struct tcp_fastopen_cookie *cookie)
5809 struct tcp_sock *tp = tcp_sk(sk);
5810 struct sk_buff *data = tp->syn_data ? tcp_rtx_queue_head(sk) : NULL;
5811 u16 mss = tp->rx_opt.mss_clamp, try_exp = 0;
5812 bool syn_drop = false;
5814 if (mss == tp->rx_opt.user_mss) {
5815 struct tcp_options_received opt;
5817 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5818 tcp_clear_options(&opt);
5819 opt.user_mss = opt.mss_clamp = 0;
5820 tcp_parse_options(sock_net(sk), synack, &opt, 0, NULL);
5821 mss = opt.mss_clamp;
5824 if (!tp->syn_fastopen) {
5825 /* Ignore an unsolicited cookie */
5827 } else if (tp->total_retrans) {
5828 /* SYN timed out and the SYN-ACK neither has a cookie nor
5829 * acknowledges data. Presumably the remote received only
5830 * the retransmitted (regular) SYNs: either the original
5831 * SYN-data or the corresponding SYN-ACK was dropped.
5833 syn_drop = (cookie->len < 0 && data);
5834 } else if (cookie->len < 0 && !tp->syn_data) {
5835 /* We requested a cookie but didn't get it. If we did not use
5836 * the (old) exp opt format then try so next time (try_exp=1).
5837 * Otherwise we go back to use the RFC7413 opt (try_exp=2).
5839 try_exp = tp->syn_fastopen_exp ? 2 : 1;
5842 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp);
5844 if (data) { /* Retransmit unacked data in SYN */
5845 if (tp->total_retrans)
5846 tp->fastopen_client_fail = TFO_SYN_RETRANSMITTED;
5848 tp->fastopen_client_fail = TFO_DATA_NOT_ACKED;
5849 skb_rbtree_walk_from(data) {
5850 if (__tcp_retransmit_skb(sk, data, 1))
5854 NET_INC_STATS(sock_net(sk),
5855 LINUX_MIB_TCPFASTOPENACTIVEFAIL);
5858 tp->syn_data_acked = tp->syn_data;
5859 if (tp->syn_data_acked) {
5860 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
5861 /* SYN-data is counted as two separate packets in tcp_ack() */
5862 if (tp->delivered > 1)
5866 tcp_fastopen_add_skb(sk, synack);
5871 static void smc_check_reset_syn(struct tcp_sock *tp)
5873 #if IS_ENABLED(CONFIG_SMC)
5874 if (static_branch_unlikely(&tcp_have_smc)) {
5875 if (tp->syn_smc && !tp->rx_opt.smc_ok)
5881 static void tcp_try_undo_spurious_syn(struct sock *sk)
5883 struct tcp_sock *tp = tcp_sk(sk);
5886 /* undo_marker is set when SYN or SYNACK times out. The timeout is
5887 * spurious if the ACK's timestamp option echo value matches the
5888 * original SYN timestamp.
5890 syn_stamp = tp->retrans_stamp;
5891 if (tp->undo_marker && syn_stamp && tp->rx_opt.saw_tstamp &&
5892 syn_stamp == tp->rx_opt.rcv_tsecr)
5893 tp->undo_marker = 0;
5896 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5897 const struct tcphdr *th)
5899 struct inet_connection_sock *icsk = inet_csk(sk);
5900 struct tcp_sock *tp = tcp_sk(sk);
5901 struct tcp_fastopen_cookie foc = { .len = -1 };
5902 int saved_clamp = tp->rx_opt.mss_clamp;
5905 tcp_parse_options(sock_net(sk), skb, &tp->rx_opt, 0, &foc);
5906 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
5907 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
5911 * "If the state is SYN-SENT then
5912 * first check the ACK bit
5913 * If the ACK bit is set
5914 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5915 * a reset (unless the RST bit is set, if so drop
5916 * the segment and return)"
5918 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
5919 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5920 /* Previous FIN/ACK or RST/ACK might be ignored. */
5921 if (icsk->icsk_retransmits == 0)
5922 inet_csk_reset_xmit_timer(sk,
5924 TCP_TIMEOUT_MIN, TCP_RTO_MAX);
5925 goto reset_and_undo;
5928 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5929 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5930 tcp_time_stamp(tp))) {
5931 NET_INC_STATS(sock_net(sk),
5932 LINUX_MIB_PAWSACTIVEREJECTED);
5933 goto reset_and_undo;
5936 /* Now ACK is acceptable.
5938 * "If the RST bit is set
5939 * If the ACK was acceptable then signal the user "error:
5940 * connection reset", drop the segment, enter CLOSED state,
5941 * delete TCB, and return."
5950 * "fifth, if neither of the SYN or RST bits is set then
5951 * drop the segment and return."
5957 goto discard_and_undo;
5960 * "If the SYN bit is on ...
5961 * are acceptable then ...
5962 * (our SYN has been ACKed), change the connection
5963 * state to ESTABLISHED..."
5966 tcp_ecn_rcv_synack(tp, th);
5968 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5969 tcp_try_undo_spurious_syn(sk);
5970 tcp_ack(sk, skb, FLAG_SLOWPATH);
5972 /* Ok.. it's good. Set up sequence numbers and
5973 * move to established.
5975 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1);
5976 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5978 /* RFC1323: The window in SYN & SYN/ACK segments is
5981 tp->snd_wnd = ntohs(th->window);
5983 if (!tp->rx_opt.wscale_ok) {
5984 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5985 tp->window_clamp = min(tp->window_clamp, 65535U);
5988 if (tp->rx_opt.saw_tstamp) {
5989 tp->rx_opt.tstamp_ok = 1;
5990 tp->tcp_header_len =
5991 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5992 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5993 tcp_store_ts_recent(tp);
5995 tp->tcp_header_len = sizeof(struct tcphdr);
5998 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5999 tcp_initialize_rcv_mss(sk);
6001 /* Remember, tcp_poll() does not lock socket!
6002 * Change state from SYN-SENT only after copied_seq
6003 * is initialized. */
6004 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6006 smc_check_reset_syn(tp);
6010 tcp_finish_connect(sk, skb);
6012 fastopen_fail = (tp->syn_fastopen || tp->syn_data) &&
6013 tcp_rcv_fastopen_synack(sk, skb, &foc);
6015 if (!sock_flag(sk, SOCK_DEAD)) {
6016 sk->sk_state_change(sk);
6017 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
6021 if (sk->sk_write_pending ||
6022 icsk->icsk_accept_queue.rskq_defer_accept ||
6023 inet_csk_in_pingpong_mode(sk)) {
6024 /* Save one ACK. Data will be ready after
6025 * several ticks, if write_pending is set.
6027 * It may be deleted, but with this feature tcpdumps
6028 * look so _wonderfully_ clever, that I was not able
6029 * to stand against the temptation 8) --ANK
6031 inet_csk_schedule_ack(sk);
6032 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
6033 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
6034 TCP_DELACK_MAX, TCP_RTO_MAX);
6045 /* No ACK in the segment */
6049 * "If the RST bit is set
6051 * Otherwise (no ACK) drop the segment and return."
6054 goto discard_and_undo;
6058 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
6059 tcp_paws_reject(&tp->rx_opt, 0))
6060 goto discard_and_undo;
6063 /* We see SYN without ACK. It is attempt of
6064 * simultaneous connect with crossed SYNs.
6065 * Particularly, it can be connect to self.
6067 tcp_set_state(sk, TCP_SYN_RECV);
6069 if (tp->rx_opt.saw_tstamp) {
6070 tp->rx_opt.tstamp_ok = 1;
6071 tcp_store_ts_recent(tp);
6072 tp->tcp_header_len =
6073 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
6075 tp->tcp_header_len = sizeof(struct tcphdr);
6078 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1);
6079 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6080 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
6082 /* RFC1323: The window in SYN & SYN/ACK segments is
6085 tp->snd_wnd = ntohs(th->window);
6086 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
6087 tp->max_window = tp->snd_wnd;
6089 tcp_ecn_rcv_syn(tp, th);
6092 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
6093 tcp_initialize_rcv_mss(sk);
6095 tcp_send_synack(sk);
6097 /* Note, we could accept data and URG from this segment.
6098 * There are no obstacles to make this (except that we must
6099 * either change tcp_recvmsg() to prevent it from returning data
6100 * before 3WHS completes per RFC793, or employ TCP Fast Open).
6102 * However, if we ignore data in ACKless segments sometimes,
6103 * we have no reasons to accept it sometimes.
6104 * Also, seems the code doing it in step6 of tcp_rcv_state_process
6105 * is not flawless. So, discard packet for sanity.
6106 * Uncomment this return to process the data.
6113 /* "fifth, if neither of the SYN or RST bits is set then
6114 * drop the segment and return."
6118 tcp_clear_options(&tp->rx_opt);
6119 tp->rx_opt.mss_clamp = saved_clamp;
6123 tcp_clear_options(&tp->rx_opt);
6124 tp->rx_opt.mss_clamp = saved_clamp;
6128 static void tcp_rcv_synrecv_state_fastopen(struct sock *sk)
6130 struct request_sock *req;
6132 /* If we are still handling the SYNACK RTO, see if timestamp ECR allows
6133 * undo. If peer SACKs triggered fast recovery, we can't undo here.
6135 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
6136 tcp_try_undo_loss(sk, false);
6138 /* Reset rtx states to prevent spurious retransmits_timed_out() */
6139 tcp_sk(sk)->retrans_stamp = 0;
6140 inet_csk(sk)->icsk_retransmits = 0;
6142 /* Once we leave TCP_SYN_RECV or TCP_FIN_WAIT_1,
6143 * we no longer need req so release it.
6145 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk,
6146 lockdep_sock_is_held(sk));
6147 reqsk_fastopen_remove(sk, req, false);
6149 /* Re-arm the timer because data may have been sent out.
6150 * This is similar to the regular data transmission case
6151 * when new data has just been ack'ed.
6153 * (TFO) - we could try to be more aggressive and
6154 * retransmitting any data sooner based on when they
6161 * This function implements the receiving procedure of RFC 793 for
6162 * all states except ESTABLISHED and TIME_WAIT.
6163 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
6164 * address independent.
6167 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb)
6169 struct tcp_sock *tp = tcp_sk(sk);
6170 struct inet_connection_sock *icsk = inet_csk(sk);
6171 const struct tcphdr *th = tcp_hdr(skb);
6172 struct request_sock *req;
6176 switch (sk->sk_state) {
6190 /* It is possible that we process SYN packets from backlog,
6191 * so we need to make sure to disable BH and RCU right there.
6195 acceptable = icsk->icsk_af_ops->conn_request(sk, skb) >= 0;
6207 tp->rx_opt.saw_tstamp = 0;
6208 tcp_mstamp_refresh(tp);
6209 queued = tcp_rcv_synsent_state_process(sk, skb, th);
6213 /* Do step6 onward by hand. */
6214 tcp_urg(sk, skb, th);
6216 tcp_data_snd_check(sk);
6220 tcp_mstamp_refresh(tp);
6221 tp->rx_opt.saw_tstamp = 0;
6222 req = rcu_dereference_protected(tp->fastopen_rsk,
6223 lockdep_sock_is_held(sk));
6227 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
6228 sk->sk_state != TCP_FIN_WAIT1);
6230 if (!tcp_check_req(sk, skb, req, true, &req_stolen))
6234 if (!th->ack && !th->rst && !th->syn)
6237 if (!tcp_validate_incoming(sk, skb, th, 0))
6240 /* step 5: check the ACK field */
6241 acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
6242 FLAG_UPDATE_TS_RECENT |
6243 FLAG_NO_CHALLENGE_ACK) > 0;
6246 if (sk->sk_state == TCP_SYN_RECV)
6247 return 1; /* send one RST */
6248 tcp_send_challenge_ack(sk, skb);
6251 switch (sk->sk_state) {
6253 tp->delivered++; /* SYN-ACK delivery isn't tracked in tcp_ack */
6255 tcp_synack_rtt_meas(sk, req);
6258 tcp_rcv_synrecv_state_fastopen(sk);
6260 tcp_try_undo_spurious_syn(sk);
6261 tp->retrans_stamp = 0;
6262 tcp_init_transfer(sk, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB);
6263 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6266 tcp_set_state(sk, TCP_ESTABLISHED);
6267 sk->sk_state_change(sk);
6269 /* Note, that this wakeup is only for marginal crossed SYN case.
6270 * Passively open sockets are not waked up, because
6271 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
6274 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
6276 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
6277 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
6278 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
6280 if (tp->rx_opt.tstamp_ok)
6281 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
6283 if (!inet_csk(sk)->icsk_ca_ops->cong_control)
6284 tcp_update_pacing_rate(sk);
6286 /* Prevent spurious tcp_cwnd_restart() on first data packet */
6287 tp->lsndtime = tcp_jiffies32;
6289 tcp_initialize_rcv_mss(sk);
6290 tcp_fast_path_on(tp);
6293 case TCP_FIN_WAIT1: {
6297 tcp_rcv_synrecv_state_fastopen(sk);
6299 if (tp->snd_una != tp->write_seq)
6302 tcp_set_state(sk, TCP_FIN_WAIT2);
6303 sk->sk_shutdown |= SEND_SHUTDOWN;
6307 if (!sock_flag(sk, SOCK_DEAD)) {
6308 /* Wake up lingering close() */
6309 sk->sk_state_change(sk);
6313 if (tp->linger2 < 0) {
6315 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6318 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6319 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6320 /* Receive out of order FIN after close() */
6321 if (tp->syn_fastopen && th->fin)
6322 tcp_fastopen_active_disable(sk);
6324 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6328 tmo = tcp_fin_time(sk);
6329 if (tmo > TCP_TIMEWAIT_LEN) {
6330 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
6331 } else if (th->fin || sock_owned_by_user(sk)) {
6332 /* Bad case. We could lose such FIN otherwise.
6333 * It is not a big problem, but it looks confusing
6334 * and not so rare event. We still can lose it now,
6335 * if it spins in bh_lock_sock(), but it is really
6338 inet_csk_reset_keepalive_timer(sk, tmo);
6340 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
6347 if (tp->snd_una == tp->write_seq) {
6348 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
6354 if (tp->snd_una == tp->write_seq) {
6355 tcp_update_metrics(sk);
6362 /* step 6: check the URG bit */
6363 tcp_urg(sk, skb, th);
6365 /* step 7: process the segment text */
6366 switch (sk->sk_state) {
6367 case TCP_CLOSE_WAIT:
6370 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
6371 if (sk_is_mptcp(sk))
6372 mptcp_incoming_options(sk, skb, &tp->rx_opt);
6378 /* RFC 793 says to queue data in these states,
6379 * RFC 1122 says we MUST send a reset.
6380 * BSD 4.4 also does reset.
6382 if (sk->sk_shutdown & RCV_SHUTDOWN) {
6383 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6384 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6385 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6391 case TCP_ESTABLISHED:
6392 tcp_data_queue(sk, skb);
6397 /* tcp_data could move socket to TIME-WAIT */
6398 if (sk->sk_state != TCP_CLOSE) {
6399 tcp_data_snd_check(sk);
6400 tcp_ack_snd_check(sk);
6409 EXPORT_SYMBOL(tcp_rcv_state_process);
6411 static inline void pr_drop_req(struct request_sock *req, __u16 port, int family)
6413 struct inet_request_sock *ireq = inet_rsk(req);
6415 if (family == AF_INET)
6416 net_dbg_ratelimited("drop open request from %pI4/%u\n",
6417 &ireq->ir_rmt_addr, port);
6418 #if IS_ENABLED(CONFIG_IPV6)
6419 else if (family == AF_INET6)
6420 net_dbg_ratelimited("drop open request from %pI6/%u\n",
6421 &ireq->ir_v6_rmt_addr, port);
6425 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
6427 * If we receive a SYN packet with these bits set, it means a
6428 * network is playing bad games with TOS bits. In order to
6429 * avoid possible false congestion notifications, we disable
6430 * TCP ECN negotiation.
6432 * Exception: tcp_ca wants ECN. This is required for DCTCP
6433 * congestion control: Linux DCTCP asserts ECT on all packets,
6434 * including SYN, which is most optimal solution; however,
6435 * others, such as FreeBSD do not.
6437 * Exception: At least one of the reserved bits of the TCP header (th->res1) is
6438 * set, indicating the use of a future TCP extension (such as AccECN). See
6439 * RFC8311 §4.3 which updates RFC3168 to allow the development of such
6442 static void tcp_ecn_create_request(struct request_sock *req,
6443 const struct sk_buff *skb,
6444 const struct sock *listen_sk,
6445 const struct dst_entry *dst)
6447 const struct tcphdr *th = tcp_hdr(skb);
6448 const struct net *net = sock_net(listen_sk);
6449 bool th_ecn = th->ece && th->cwr;
6456 ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield);
6457 ecn_ok_dst = dst_feature(dst, DST_FEATURE_ECN_MASK);
6458 ecn_ok = net->ipv4.sysctl_tcp_ecn || ecn_ok_dst;
6460 if (((!ect || th->res1) && ecn_ok) || tcp_ca_needs_ecn(listen_sk) ||
6461 (ecn_ok_dst & DST_FEATURE_ECN_CA) ||
6462 tcp_bpf_ca_needs_ecn((struct sock *)req))
6463 inet_rsk(req)->ecn_ok = 1;
6466 static void tcp_openreq_init(struct request_sock *req,
6467 const struct tcp_options_received *rx_opt,
6468 struct sk_buff *skb, const struct sock *sk)
6470 struct inet_request_sock *ireq = inet_rsk(req);
6472 req->rsk_rcv_wnd = 0; /* So that tcp_send_synack() knows! */
6474 tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq;
6475 tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
6476 tcp_rsk(req)->snt_synack = 0;
6477 tcp_rsk(req)->last_oow_ack_time = 0;
6478 req->mss = rx_opt->mss_clamp;
6479 req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0;
6480 ireq->tstamp_ok = rx_opt->tstamp_ok;
6481 ireq->sack_ok = rx_opt->sack_ok;
6482 ireq->snd_wscale = rx_opt->snd_wscale;
6483 ireq->wscale_ok = rx_opt->wscale_ok;
6486 ireq->ir_rmt_port = tcp_hdr(skb)->source;
6487 ireq->ir_num = ntohs(tcp_hdr(skb)->dest);
6488 ireq->ir_mark = inet_request_mark(sk, skb);
6489 #if IS_ENABLED(CONFIG_SMC)
6490 ireq->smc_ok = rx_opt->smc_ok;
6494 struct request_sock *inet_reqsk_alloc(const struct request_sock_ops *ops,
6495 struct sock *sk_listener,
6496 bool attach_listener)
6498 struct request_sock *req = reqsk_alloc(ops, sk_listener,
6502 struct inet_request_sock *ireq = inet_rsk(req);
6504 ireq->ireq_opt = NULL;
6505 #if IS_ENABLED(CONFIG_IPV6)
6506 ireq->pktopts = NULL;
6508 atomic64_set(&ireq->ir_cookie, 0);
6509 ireq->ireq_state = TCP_NEW_SYN_RECV;
6510 write_pnet(&ireq->ireq_net, sock_net(sk_listener));
6511 ireq->ireq_family = sk_listener->sk_family;
6516 EXPORT_SYMBOL(inet_reqsk_alloc);
6519 * Return true if a syncookie should be sent
6521 static bool tcp_syn_flood_action(const struct sock *sk, const char *proto)
6523 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
6524 const char *msg = "Dropping request";
6525 bool want_cookie = false;
6526 struct net *net = sock_net(sk);
6528 #ifdef CONFIG_SYN_COOKIES
6529 if (net->ipv4.sysctl_tcp_syncookies) {
6530 msg = "Sending cookies";
6532 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES);
6535 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP);
6537 if (!queue->synflood_warned &&
6538 net->ipv4.sysctl_tcp_syncookies != 2 &&
6539 xchg(&queue->synflood_warned, 1) == 0)
6540 net_info_ratelimited("%s: Possible SYN flooding on port %d. %s. Check SNMP counters.\n",
6541 proto, sk->sk_num, msg);
6546 static void tcp_reqsk_record_syn(const struct sock *sk,
6547 struct request_sock *req,
6548 const struct sk_buff *skb)
6550 if (tcp_sk(sk)->save_syn) {
6551 u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb);
6554 copy = kmalloc(len + sizeof(u32), GFP_ATOMIC);
6557 memcpy(©[1], skb_network_header(skb), len);
6558 req->saved_syn = copy;
6563 /* If a SYN cookie is required and supported, returns a clamped MSS value to be
6564 * used for SYN cookie generation.
6566 u16 tcp_get_syncookie_mss(struct request_sock_ops *rsk_ops,
6567 const struct tcp_request_sock_ops *af_ops,
6568 struct sock *sk, struct tcphdr *th)
6570 struct tcp_sock *tp = tcp_sk(sk);
6573 if (sock_net(sk)->ipv4.sysctl_tcp_syncookies != 2 &&
6574 !inet_csk_reqsk_queue_is_full(sk))
6577 if (!tcp_syn_flood_action(sk, rsk_ops->slab_name))
6580 if (sk_acceptq_is_full(sk)) {
6581 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
6585 mss = tcp_parse_mss_option(th, tp->rx_opt.user_mss);
6587 mss = af_ops->mss_clamp;
6591 EXPORT_SYMBOL_GPL(tcp_get_syncookie_mss);
6593 int tcp_conn_request(struct request_sock_ops *rsk_ops,
6594 const struct tcp_request_sock_ops *af_ops,
6595 struct sock *sk, struct sk_buff *skb)
6597 struct tcp_fastopen_cookie foc = { .len = -1 };
6598 __u32 isn = TCP_SKB_CB(skb)->tcp_tw_isn;
6599 struct tcp_options_received tmp_opt;
6600 struct tcp_sock *tp = tcp_sk(sk);
6601 struct net *net = sock_net(sk);
6602 struct sock *fastopen_sk = NULL;
6603 struct request_sock *req;
6604 bool want_cookie = false;
6605 struct dst_entry *dst;
6608 /* TW buckets are converted to open requests without
6609 * limitations, they conserve resources and peer is
6610 * evidently real one.
6612 if ((net->ipv4.sysctl_tcp_syncookies == 2 ||
6613 inet_csk_reqsk_queue_is_full(sk)) && !isn) {
6614 want_cookie = tcp_syn_flood_action(sk, rsk_ops->slab_name);
6619 if (sk_acceptq_is_full(sk)) {
6620 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
6624 req = inet_reqsk_alloc(rsk_ops, sk, !want_cookie);
6628 tcp_rsk(req)->af_specific = af_ops;
6629 tcp_rsk(req)->ts_off = 0;
6630 #if IS_ENABLED(CONFIG_MPTCP)
6631 tcp_rsk(req)->is_mptcp = 0;
6634 tcp_clear_options(&tmp_opt);
6635 tmp_opt.mss_clamp = af_ops->mss_clamp;
6636 tmp_opt.user_mss = tp->rx_opt.user_mss;
6637 tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0,
6638 want_cookie ? NULL : &foc);
6640 if (want_cookie && !tmp_opt.saw_tstamp)
6641 tcp_clear_options(&tmp_opt);
6643 if (IS_ENABLED(CONFIG_SMC) && want_cookie)
6646 tmp_opt.tstamp_ok = tmp_opt.saw_tstamp;
6647 tcp_openreq_init(req, &tmp_opt, skb, sk);
6648 inet_rsk(req)->no_srccheck = inet_sk(sk)->transparent;
6650 /* Note: tcp_v6_init_req() might override ir_iif for link locals */
6651 inet_rsk(req)->ir_iif = inet_request_bound_dev_if(sk, skb);
6653 af_ops->init_req(req, sk, skb);
6655 if (IS_ENABLED(CONFIG_MPTCP) && want_cookie)
6656 tcp_rsk(req)->is_mptcp = 0;
6658 if (security_inet_conn_request(sk, skb, req))
6661 if (tmp_opt.tstamp_ok)
6662 tcp_rsk(req)->ts_off = af_ops->init_ts_off(net, skb);
6664 dst = af_ops->route_req(sk, &fl, req);
6668 if (!want_cookie && !isn) {
6669 /* Kill the following clause, if you dislike this way. */
6670 if (!net->ipv4.sysctl_tcp_syncookies &&
6671 (net->ipv4.sysctl_max_syn_backlog - inet_csk_reqsk_queue_len(sk) <
6672 (net->ipv4.sysctl_max_syn_backlog >> 2)) &&
6673 !tcp_peer_is_proven(req, dst)) {
6674 /* Without syncookies last quarter of
6675 * backlog is filled with destinations,
6676 * proven to be alive.
6677 * It means that we continue to communicate
6678 * to destinations, already remembered
6679 * to the moment of synflood.
6681 pr_drop_req(req, ntohs(tcp_hdr(skb)->source),
6683 goto drop_and_release;
6686 isn = af_ops->init_seq(skb);
6689 tcp_ecn_create_request(req, skb, sk, dst);
6692 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss);
6693 req->cookie_ts = tmp_opt.tstamp_ok;
6694 if (!tmp_opt.tstamp_ok)
6695 inet_rsk(req)->ecn_ok = 0;
6698 tcp_rsk(req)->snt_isn = isn;
6699 tcp_rsk(req)->txhash = net_tx_rndhash();
6700 tcp_openreq_init_rwin(req, sk, dst);
6701 sk_rx_queue_set(req_to_sk(req), skb);
6703 tcp_reqsk_record_syn(sk, req, skb);
6704 fastopen_sk = tcp_try_fastopen(sk, skb, req, &foc, dst);
6707 af_ops->send_synack(fastopen_sk, dst, &fl, req,
6708 &foc, TCP_SYNACK_FASTOPEN);
6709 /* Add the child socket directly into the accept queue */
6710 if (!inet_csk_reqsk_queue_add(sk, req, fastopen_sk)) {
6711 reqsk_fastopen_remove(fastopen_sk, req, false);
6712 bh_unlock_sock(fastopen_sk);
6713 sock_put(fastopen_sk);
6716 sk->sk_data_ready(sk);
6717 bh_unlock_sock(fastopen_sk);
6718 sock_put(fastopen_sk);
6720 tcp_rsk(req)->tfo_listener = false;
6722 inet_csk_reqsk_queue_hash_add(sk, req,
6723 tcp_timeout_init((struct sock *)req));
6724 af_ops->send_synack(sk, dst, &fl, req, &foc,
6725 !want_cookie ? TCP_SYNACK_NORMAL :
6743 EXPORT_SYMBOL(tcp_conn_request);