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>
83 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
85 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
86 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
87 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
88 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
89 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
90 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
91 #define FLAG_ECE 0x40 /* ECE in this ACK */
92 #define FLAG_LOST_RETRANS 0x80 /* This ACK marks some retransmission lost */
93 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
94 #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
95 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
96 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
97 #define FLAG_SET_XMIT_TIMER 0x1000 /* Set TLP or RTO timer */
98 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
99 #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
100 #define FLAG_NO_CHALLENGE_ACK 0x8000 /* do not call tcp_send_challenge_ack() */
101 #define FLAG_ACK_MAYBE_DELAYED 0x10000 /* Likely a delayed ACK */
103 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
104 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
105 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE|FLAG_DSACKING_ACK)
106 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
108 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
109 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
111 #define REXMIT_NONE 0 /* no loss recovery to do */
112 #define REXMIT_LOST 1 /* retransmit packets marked lost */
113 #define REXMIT_NEW 2 /* FRTO-style transmit of unsent/new packets */
115 #if IS_ENABLED(CONFIG_TLS_DEVICE)
116 static DEFINE_STATIC_KEY_DEFERRED_FALSE(clean_acked_data_enabled, HZ);
118 void clean_acked_data_enable(struct inet_connection_sock *icsk,
119 void (*cad)(struct sock *sk, u32 ack_seq))
121 icsk->icsk_clean_acked = cad;
122 static_branch_deferred_inc(&clean_acked_data_enabled);
124 EXPORT_SYMBOL_GPL(clean_acked_data_enable);
126 void clean_acked_data_disable(struct inet_connection_sock *icsk)
128 static_branch_slow_dec_deferred(&clean_acked_data_enabled);
129 icsk->icsk_clean_acked = NULL;
131 EXPORT_SYMBOL_GPL(clean_acked_data_disable);
133 void clean_acked_data_flush(void)
135 static_key_deferred_flush(&clean_acked_data_enabled);
137 EXPORT_SYMBOL_GPL(clean_acked_data_flush);
140 static void tcp_gro_dev_warn(struct sock *sk, const struct sk_buff *skb,
143 static bool __once __read_mostly;
146 struct net_device *dev;
151 dev = dev_get_by_index_rcu(sock_net(sk), skb->skb_iif);
152 if (!dev || len >= dev->mtu)
153 pr_warn("%s: Driver has suspect GRO implementation, TCP performance may be compromised.\n",
154 dev ? dev->name : "Unknown driver");
159 /* Adapt the MSS value used to make delayed ack decision to the
162 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
164 struct inet_connection_sock *icsk = inet_csk(sk);
165 const unsigned int lss = icsk->icsk_ack.last_seg_size;
168 icsk->icsk_ack.last_seg_size = 0;
170 /* skb->len may jitter because of SACKs, even if peer
171 * sends good full-sized frames.
173 len = skb_shinfo(skb)->gso_size ? : skb->len;
174 if (len >= icsk->icsk_ack.rcv_mss) {
175 icsk->icsk_ack.rcv_mss = min_t(unsigned int, len,
177 /* Account for possibly-removed options */
178 if (unlikely(len > icsk->icsk_ack.rcv_mss +
179 MAX_TCP_OPTION_SPACE))
180 tcp_gro_dev_warn(sk, skb, len);
182 /* Otherwise, we make more careful check taking into account,
183 * that SACKs block is variable.
185 * "len" is invariant segment length, including TCP header.
187 len += skb->data - skb_transport_header(skb);
188 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
189 /* If PSH is not set, packet should be
190 * full sized, provided peer TCP is not badly broken.
191 * This observation (if it is correct 8)) allows
192 * to handle super-low mtu links fairly.
194 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
195 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
196 /* Subtract also invariant (if peer is RFC compliant),
197 * tcp header plus fixed timestamp option length.
198 * Resulting "len" is MSS free of SACK jitter.
200 len -= tcp_sk(sk)->tcp_header_len;
201 icsk->icsk_ack.last_seg_size = len;
203 icsk->icsk_ack.rcv_mss = len;
207 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
208 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
209 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
213 static void tcp_incr_quickack(struct sock *sk, unsigned int max_quickacks)
215 struct inet_connection_sock *icsk = inet_csk(sk);
216 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
220 quickacks = min(quickacks, max_quickacks);
221 if (quickacks > icsk->icsk_ack.quick)
222 icsk->icsk_ack.quick = quickacks;
225 void tcp_enter_quickack_mode(struct sock *sk, unsigned int max_quickacks)
227 struct inet_connection_sock *icsk = inet_csk(sk);
229 tcp_incr_quickack(sk, max_quickacks);
230 inet_csk_exit_pingpong_mode(sk);
231 icsk->icsk_ack.ato = TCP_ATO_MIN;
233 EXPORT_SYMBOL(tcp_enter_quickack_mode);
235 /* Send ACKs quickly, if "quick" count is not exhausted
236 * and the session is not interactive.
239 static bool tcp_in_quickack_mode(struct sock *sk)
241 const struct inet_connection_sock *icsk = inet_csk(sk);
242 const struct dst_entry *dst = __sk_dst_get(sk);
244 return (dst && dst_metric(dst, RTAX_QUICKACK)) ||
245 (icsk->icsk_ack.quick && !inet_csk_in_pingpong_mode(sk));
248 static void tcp_ecn_queue_cwr(struct tcp_sock *tp)
250 if (tp->ecn_flags & TCP_ECN_OK)
251 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
254 static void tcp_ecn_accept_cwr(struct sock *sk, const struct sk_buff *skb)
256 if (tcp_hdr(skb)->cwr) {
257 tcp_sk(sk)->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
259 /* If the sender is telling us it has entered CWR, then its
260 * cwnd may be very low (even just 1 packet), so we should ACK
263 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW;
267 static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp)
269 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
272 static void __tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb)
274 struct tcp_sock *tp = tcp_sk(sk);
276 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
277 case INET_ECN_NOT_ECT:
278 /* Funny extension: if ECT is not set on a segment,
279 * and we already seen ECT on a previous segment,
280 * it is probably a retransmit.
282 if (tp->ecn_flags & TCP_ECN_SEEN)
283 tcp_enter_quickack_mode(sk, 2);
286 if (tcp_ca_needs_ecn(sk))
287 tcp_ca_event(sk, CA_EVENT_ECN_IS_CE);
289 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
290 /* Better not delay acks, sender can have a very low cwnd */
291 tcp_enter_quickack_mode(sk, 2);
292 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
294 tp->ecn_flags |= TCP_ECN_SEEN;
297 if (tcp_ca_needs_ecn(sk))
298 tcp_ca_event(sk, CA_EVENT_ECN_NO_CE);
299 tp->ecn_flags |= TCP_ECN_SEEN;
304 static void tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb)
306 if (tcp_sk(sk)->ecn_flags & TCP_ECN_OK)
307 __tcp_ecn_check_ce(sk, skb);
310 static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
312 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
313 tp->ecn_flags &= ~TCP_ECN_OK;
316 static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
318 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
319 tp->ecn_flags &= ~TCP_ECN_OK;
322 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
324 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
329 /* Buffer size and advertised window tuning.
331 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
334 static void tcp_sndbuf_expand(struct sock *sk)
336 const struct tcp_sock *tp = tcp_sk(sk);
337 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
341 /* Worst case is non GSO/TSO : each frame consumes one skb
342 * and skb->head is kmalloced using power of two area of memory
344 per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
346 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
348 per_mss = roundup_pow_of_two(per_mss) +
349 SKB_DATA_ALIGN(sizeof(struct sk_buff));
351 nr_segs = max_t(u32, TCP_INIT_CWND, tp->snd_cwnd);
352 nr_segs = max_t(u32, nr_segs, tp->reordering + 1);
354 /* Fast Recovery (RFC 5681 3.2) :
355 * Cubic needs 1.7 factor, rounded to 2 to include
356 * extra cushion (application might react slowly to EPOLLOUT)
358 sndmem = ca_ops->sndbuf_expand ? ca_ops->sndbuf_expand(sk) : 2;
359 sndmem *= nr_segs * per_mss;
361 if (sk->sk_sndbuf < sndmem)
362 WRITE_ONCE(sk->sk_sndbuf,
363 min(sndmem, sock_net(sk)->ipv4.sysctl_tcp_wmem[2]));
366 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
368 * All tcp_full_space() is split to two parts: "network" buffer, allocated
369 * forward and advertised in receiver window (tp->rcv_wnd) and
370 * "application buffer", required to isolate scheduling/application
371 * latencies from network.
372 * window_clamp is maximal advertised window. It can be less than
373 * tcp_full_space(), in this case tcp_full_space() - window_clamp
374 * is reserved for "application" buffer. The less window_clamp is
375 * the smoother our behaviour from viewpoint of network, but the lower
376 * throughput and the higher sensitivity of the connection to losses. 8)
378 * rcv_ssthresh is more strict window_clamp used at "slow start"
379 * phase to predict further behaviour of this connection.
380 * It is used for two goals:
381 * - to enforce header prediction at sender, even when application
382 * requires some significant "application buffer". It is check #1.
383 * - to prevent pruning of receive queue because of misprediction
384 * of receiver window. Check #2.
386 * The scheme does not work when sender sends good segments opening
387 * window and then starts to feed us spaghetti. But it should work
388 * in common situations. Otherwise, we have to rely on queue collapsing.
391 /* Slow part of check#2. */
392 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
394 struct tcp_sock *tp = tcp_sk(sk);
396 int truesize = tcp_win_from_space(sk, skb->truesize) >> 1;
397 int window = tcp_win_from_space(sk, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]) >> 1;
399 while (tp->rcv_ssthresh <= window) {
400 if (truesize <= skb->len)
401 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
409 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
411 struct tcp_sock *tp = tcp_sk(sk);
414 room = min_t(int, tp->window_clamp, tcp_space(sk)) - tp->rcv_ssthresh;
417 if (room > 0 && !tcp_under_memory_pressure(sk)) {
420 /* Check #2. Increase window, if skb with such overhead
421 * will fit to rcvbuf in future.
423 if (tcp_win_from_space(sk, skb->truesize) <= skb->len)
424 incr = 2 * tp->advmss;
426 incr = __tcp_grow_window(sk, skb);
429 incr = max_t(int, incr, 2 * skb->len);
430 tp->rcv_ssthresh += min(room, incr);
431 inet_csk(sk)->icsk_ack.quick |= 1;
436 /* 3. Try to fixup all. It is made immediately after connection enters
439 void tcp_init_buffer_space(struct sock *sk)
441 int tcp_app_win = sock_net(sk)->ipv4.sysctl_tcp_app_win;
442 struct tcp_sock *tp = tcp_sk(sk);
445 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
446 tcp_sndbuf_expand(sk);
448 tp->rcvq_space.space = min_t(u32, tp->rcv_wnd, TCP_INIT_CWND * tp->advmss);
449 tcp_mstamp_refresh(tp);
450 tp->rcvq_space.time = tp->tcp_mstamp;
451 tp->rcvq_space.seq = tp->copied_seq;
453 maxwin = tcp_full_space(sk);
455 if (tp->window_clamp >= maxwin) {
456 tp->window_clamp = maxwin;
458 if (tcp_app_win && maxwin > 4 * tp->advmss)
459 tp->window_clamp = max(maxwin -
460 (maxwin >> tcp_app_win),
464 /* Force reservation of one segment. */
466 tp->window_clamp > 2 * tp->advmss &&
467 tp->window_clamp + tp->advmss > maxwin)
468 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
470 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
471 tp->snd_cwnd_stamp = tcp_jiffies32;
474 /* 4. Recalculate window clamp after socket hit its memory bounds. */
475 static void tcp_clamp_window(struct sock *sk)
477 struct tcp_sock *tp = tcp_sk(sk);
478 struct inet_connection_sock *icsk = inet_csk(sk);
479 struct net *net = sock_net(sk);
481 icsk->icsk_ack.quick = 0;
483 if (sk->sk_rcvbuf < net->ipv4.sysctl_tcp_rmem[2] &&
484 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
485 !tcp_under_memory_pressure(sk) &&
486 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
487 WRITE_ONCE(sk->sk_rcvbuf,
488 min(atomic_read(&sk->sk_rmem_alloc),
489 net->ipv4.sysctl_tcp_rmem[2]));
491 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
492 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
495 /* Initialize RCV_MSS value.
496 * RCV_MSS is an our guess about MSS used by the peer.
497 * We haven't any direct information about the MSS.
498 * It's better to underestimate the RCV_MSS rather than overestimate.
499 * Overestimations make us ACKing less frequently than needed.
500 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
502 void tcp_initialize_rcv_mss(struct sock *sk)
504 const struct tcp_sock *tp = tcp_sk(sk);
505 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
507 hint = min(hint, tp->rcv_wnd / 2);
508 hint = min(hint, TCP_MSS_DEFAULT);
509 hint = max(hint, TCP_MIN_MSS);
511 inet_csk(sk)->icsk_ack.rcv_mss = hint;
513 EXPORT_SYMBOL(tcp_initialize_rcv_mss);
515 /* Receiver "autotuning" code.
517 * The algorithm for RTT estimation w/o timestamps is based on
518 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
519 * <http://public.lanl.gov/radiant/pubs.html#DRS>
521 * More detail on this code can be found at
522 * <http://staff.psc.edu/jheffner/>,
523 * though this reference is out of date. A new paper
526 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
528 u32 new_sample = tp->rcv_rtt_est.rtt_us;
531 if (new_sample != 0) {
532 /* If we sample in larger samples in the non-timestamp
533 * case, we could grossly overestimate the RTT especially
534 * with chatty applications or bulk transfer apps which
535 * are stalled on filesystem I/O.
537 * Also, since we are only going for a minimum in the
538 * non-timestamp case, we do not smooth things out
539 * else with timestamps disabled convergence takes too
543 m -= (new_sample >> 3);
551 /* No previous measure. */
555 tp->rcv_rtt_est.rtt_us = new_sample;
558 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
562 if (tp->rcv_rtt_est.time == 0)
564 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
566 delta_us = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcv_rtt_est.time);
569 tcp_rcv_rtt_update(tp, delta_us, 1);
572 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
573 tp->rcv_rtt_est.time = tp->tcp_mstamp;
576 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
577 const struct sk_buff *skb)
579 struct tcp_sock *tp = tcp_sk(sk);
581 if (tp->rx_opt.rcv_tsecr == tp->rcv_rtt_last_tsecr)
583 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr;
585 if (TCP_SKB_CB(skb)->end_seq -
586 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss) {
587 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr;
590 if (likely(delta < INT_MAX / (USEC_PER_SEC / TCP_TS_HZ))) {
593 delta_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
594 tcp_rcv_rtt_update(tp, delta_us, 0);
600 * This function should be called every time data is copied to user space.
601 * It calculates the appropriate TCP receive buffer space.
603 void tcp_rcv_space_adjust(struct sock *sk)
605 struct tcp_sock *tp = tcp_sk(sk);
609 trace_tcp_rcv_space_adjust(sk);
611 tcp_mstamp_refresh(tp);
612 time = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcvq_space.time);
613 if (time < (tp->rcv_rtt_est.rtt_us >> 3) || tp->rcv_rtt_est.rtt_us == 0)
616 /* Number of bytes copied to user in last RTT */
617 copied = tp->copied_seq - tp->rcvq_space.seq;
618 if (copied <= tp->rcvq_space.space)
622 * copied = bytes received in previous RTT, our base window
623 * To cope with packet losses, we need a 2x factor
624 * To cope with slow start, and sender growing its cwin by 100 %
625 * every RTT, we need a 4x factor, because the ACK we are sending
626 * now is for the next RTT, not the current one :
627 * <prev RTT . ><current RTT .. ><next RTT .... >
630 if (sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf &&
631 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
635 /* minimal window to cope with packet losses, assuming
636 * steady state. Add some cushion because of small variations.
638 rcvwin = ((u64)copied << 1) + 16 * tp->advmss;
640 /* Accommodate for sender rate increase (eg. slow start) */
641 grow = rcvwin * (copied - tp->rcvq_space.space);
642 do_div(grow, tp->rcvq_space.space);
643 rcvwin += (grow << 1);
645 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
646 while (tcp_win_from_space(sk, rcvmem) < tp->advmss)
649 do_div(rcvwin, tp->advmss);
650 rcvbuf = min_t(u64, rcvwin * rcvmem,
651 sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
652 if (rcvbuf > sk->sk_rcvbuf) {
653 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
655 /* Make the window clamp follow along. */
656 tp->window_clamp = tcp_win_from_space(sk, rcvbuf);
659 tp->rcvq_space.space = copied;
662 tp->rcvq_space.seq = tp->copied_seq;
663 tp->rcvq_space.time = tp->tcp_mstamp;
666 /* There is something which you must keep in mind when you analyze the
667 * behavior of the tp->ato delayed ack timeout interval. When a
668 * connection starts up, we want to ack as quickly as possible. The
669 * problem is that "good" TCP's do slow start at the beginning of data
670 * transmission. The means that until we send the first few ACK's the
671 * sender will sit on his end and only queue most of his data, because
672 * he can only send snd_cwnd unacked packets at any given time. For
673 * each ACK we send, he increments snd_cwnd and transmits more of his
676 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
678 struct tcp_sock *tp = tcp_sk(sk);
679 struct inet_connection_sock *icsk = inet_csk(sk);
682 inet_csk_schedule_ack(sk);
684 tcp_measure_rcv_mss(sk, skb);
686 tcp_rcv_rtt_measure(tp);
690 if (!icsk->icsk_ack.ato) {
691 /* The _first_ data packet received, initialize
692 * delayed ACK engine.
694 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
695 icsk->icsk_ack.ato = TCP_ATO_MIN;
697 int m = now - icsk->icsk_ack.lrcvtime;
699 if (m <= TCP_ATO_MIN / 2) {
700 /* The fastest case is the first. */
701 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
702 } else if (m < icsk->icsk_ack.ato) {
703 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
704 if (icsk->icsk_ack.ato > icsk->icsk_rto)
705 icsk->icsk_ack.ato = icsk->icsk_rto;
706 } else if (m > icsk->icsk_rto) {
707 /* Too long gap. Apparently sender failed to
708 * restart window, so that we send ACKs quickly.
710 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
714 icsk->icsk_ack.lrcvtime = now;
716 tcp_ecn_check_ce(sk, skb);
719 tcp_grow_window(sk, skb);
722 /* Called to compute a smoothed rtt estimate. The data fed to this
723 * routine either comes from timestamps, or from segments that were
724 * known _not_ to have been retransmitted [see Karn/Partridge
725 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
726 * piece by Van Jacobson.
727 * NOTE: the next three routines used to be one big routine.
728 * To save cycles in the RFC 1323 implementation it was better to break
729 * it up into three procedures. -- erics
731 static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
733 struct tcp_sock *tp = tcp_sk(sk);
734 long m = mrtt_us; /* RTT */
735 u32 srtt = tp->srtt_us;
737 /* The following amusing code comes from Jacobson's
738 * article in SIGCOMM '88. Note that rtt and mdev
739 * are scaled versions of rtt and mean deviation.
740 * This is designed to be as fast as possible
741 * m stands for "measurement".
743 * On a 1990 paper the rto value is changed to:
744 * RTO = rtt + 4 * mdev
746 * Funny. This algorithm seems to be very broken.
747 * These formulae increase RTO, when it should be decreased, increase
748 * too slowly, when it should be increased quickly, decrease too quickly
749 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
750 * does not matter how to _calculate_ it. Seems, it was trap
751 * that VJ failed to avoid. 8)
754 m -= (srtt >> 3); /* m is now error in rtt est */
755 srtt += m; /* rtt = 7/8 rtt + 1/8 new */
757 m = -m; /* m is now abs(error) */
758 m -= (tp->mdev_us >> 2); /* similar update on mdev */
759 /* This is similar to one of Eifel findings.
760 * Eifel blocks mdev updates when rtt decreases.
761 * This solution is a bit different: we use finer gain
762 * for mdev in this case (alpha*beta).
763 * Like Eifel it also prevents growth of rto,
764 * but also it limits too fast rto decreases,
765 * happening in pure Eifel.
770 m -= (tp->mdev_us >> 2); /* similar update on mdev */
772 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */
773 if (tp->mdev_us > tp->mdev_max_us) {
774 tp->mdev_max_us = tp->mdev_us;
775 if (tp->mdev_max_us > tp->rttvar_us)
776 tp->rttvar_us = tp->mdev_max_us;
778 if (after(tp->snd_una, tp->rtt_seq)) {
779 if (tp->mdev_max_us < tp->rttvar_us)
780 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
781 tp->rtt_seq = tp->snd_nxt;
782 tp->mdev_max_us = tcp_rto_min_us(sk);
787 /* no previous measure. */
788 srtt = m << 3; /* take the measured time to be rtt */
789 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */
790 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
791 tp->mdev_max_us = tp->rttvar_us;
792 tp->rtt_seq = tp->snd_nxt;
796 tp->srtt_us = max(1U, srtt);
799 static void tcp_update_pacing_rate(struct sock *sk)
801 const struct tcp_sock *tp = tcp_sk(sk);
804 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
805 rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3);
807 /* current rate is (cwnd * mss) / srtt
808 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
809 * In Congestion Avoidance phase, set it to 120 % the current rate.
811 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
812 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
813 * end of slow start and should slow down.
815 if (tp->snd_cwnd < tp->snd_ssthresh / 2)
816 rate *= sock_net(sk)->ipv4.sysctl_tcp_pacing_ss_ratio;
818 rate *= sock_net(sk)->ipv4.sysctl_tcp_pacing_ca_ratio;
820 rate *= max(tp->snd_cwnd, tp->packets_out);
822 if (likely(tp->srtt_us))
823 do_div(rate, tp->srtt_us);
825 /* WRITE_ONCE() is needed because sch_fq fetches sk_pacing_rate
826 * without any lock. We want to make sure compiler wont store
827 * intermediate values in this location.
829 WRITE_ONCE(sk->sk_pacing_rate, min_t(u64, rate,
830 sk->sk_max_pacing_rate));
833 /* Calculate rto without backoff. This is the second half of Van Jacobson's
834 * routine referred to above.
836 static void tcp_set_rto(struct sock *sk)
838 const struct tcp_sock *tp = tcp_sk(sk);
839 /* Old crap is replaced with new one. 8)
842 * 1. If rtt variance happened to be less 50msec, it is hallucination.
843 * It cannot be less due to utterly erratic ACK generation made
844 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
845 * to do with delayed acks, because at cwnd>2 true delack timeout
846 * is invisible. Actually, Linux-2.4 also generates erratic
847 * ACKs in some circumstances.
849 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
851 /* 2. Fixups made earlier cannot be right.
852 * If we do not estimate RTO correctly without them,
853 * all the algo is pure shit and should be replaced
854 * with correct one. It is exactly, which we pretend to do.
857 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
858 * guarantees that rto is higher.
863 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
865 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
868 cwnd = TCP_INIT_CWND;
869 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
872 /* Take a notice that peer is sending D-SACKs */
873 static void tcp_dsack_seen(struct tcp_sock *tp)
875 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
876 tp->rack.dsack_seen = 1;
880 /* It's reordering when higher sequence was delivered (i.e. sacked) before
881 * some lower never-retransmitted sequence ("low_seq"). The maximum reordering
882 * distance is approximated in full-mss packet distance ("reordering").
884 static void tcp_check_sack_reordering(struct sock *sk, const u32 low_seq,
887 struct tcp_sock *tp = tcp_sk(sk);
888 const u32 mss = tp->mss_cache;
891 fack = tcp_highest_sack_seq(tp);
892 if (!before(low_seq, fack))
895 metric = fack - low_seq;
896 if ((metric > tp->reordering * mss) && mss) {
897 #if FASTRETRANS_DEBUG > 1
898 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
899 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
903 tp->undo_marker ? tp->undo_retrans : 0);
905 tp->reordering = min_t(u32, (metric + mss - 1) / mss,
906 sock_net(sk)->ipv4.sysctl_tcp_max_reordering);
909 /* This exciting event is worth to be remembered. 8) */
911 NET_INC_STATS(sock_net(sk),
912 ts ? LINUX_MIB_TCPTSREORDER : LINUX_MIB_TCPSACKREORDER);
915 /* This must be called before lost_out is incremented */
916 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
918 if ((!tp->retransmit_skb_hint && tp->retrans_out >= tp->lost_out) ||
919 (tp->retransmit_skb_hint &&
920 before(TCP_SKB_CB(skb)->seq,
921 TCP_SKB_CB(tp->retransmit_skb_hint)->seq)))
922 tp->retransmit_skb_hint = skb;
925 /* Sum the number of packets on the wire we have marked as lost.
926 * There are two cases we care about here:
927 * a) Packet hasn't been marked lost (nor retransmitted),
928 * and this is the first loss.
929 * b) Packet has been marked both lost and retransmitted,
930 * and this means we think it was lost again.
932 static void tcp_sum_lost(struct tcp_sock *tp, struct sk_buff *skb)
934 __u8 sacked = TCP_SKB_CB(skb)->sacked;
936 if (!(sacked & TCPCB_LOST) ||
937 ((sacked & TCPCB_LOST) && (sacked & TCPCB_SACKED_RETRANS)))
938 tp->lost += tcp_skb_pcount(skb);
941 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
943 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
944 tcp_verify_retransmit_hint(tp, skb);
946 tp->lost_out += tcp_skb_pcount(skb);
947 tcp_sum_lost(tp, skb);
948 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
952 void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp, struct sk_buff *skb)
954 tcp_verify_retransmit_hint(tp, skb);
956 tcp_sum_lost(tp, skb);
957 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
958 tp->lost_out += tcp_skb_pcount(skb);
959 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
963 /* This procedure tags the retransmission queue when SACKs arrive.
965 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
966 * Packets in queue with these bits set are counted in variables
967 * sacked_out, retrans_out and lost_out, correspondingly.
969 * Valid combinations are:
970 * Tag InFlight Description
971 * 0 1 - orig segment is in flight.
972 * S 0 - nothing flies, orig reached receiver.
973 * L 0 - nothing flies, orig lost by net.
974 * R 2 - both orig and retransmit are in flight.
975 * L|R 1 - orig is lost, retransmit is in flight.
976 * S|R 1 - orig reached receiver, retrans is still in flight.
977 * (L|S|R is logically valid, it could occur when L|R is sacked,
978 * but it is equivalent to plain S and code short-curcuits it to S.
979 * L|S is logically invalid, it would mean -1 packet in flight 8))
981 * These 6 states form finite state machine, controlled by the following events:
982 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
983 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
984 * 3. Loss detection event of two flavors:
985 * A. Scoreboard estimator decided the packet is lost.
986 * A'. Reno "three dupacks" marks head of queue lost.
987 * B. SACK arrives sacking SND.NXT at the moment, when the
988 * segment was retransmitted.
989 * 4. D-SACK added new rule: D-SACK changes any tag to S.
991 * It is pleasant to note, that state diagram turns out to be commutative,
992 * so that we are allowed not to be bothered by order of our actions,
993 * when multiple events arrive simultaneously. (see the function below).
995 * Reordering detection.
996 * --------------------
997 * Reordering metric is maximal distance, which a packet can be displaced
998 * in packet stream. With SACKs we can estimate it:
1000 * 1. SACK fills old hole and the corresponding segment was not
1001 * ever retransmitted -> reordering. Alas, we cannot use it
1002 * when segment was retransmitted.
1003 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1004 * for retransmitted and already SACKed segment -> reordering..
1005 * Both of these heuristics are not used in Loss state, when we cannot
1006 * account for retransmits accurately.
1008 * SACK block validation.
1009 * ----------------------
1011 * SACK block range validation checks that the received SACK block fits to
1012 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1013 * Note that SND.UNA is not included to the range though being valid because
1014 * it means that the receiver is rather inconsistent with itself reporting
1015 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1016 * perfectly valid, however, in light of RFC2018 which explicitly states
1017 * that "SACK block MUST reflect the newest segment. Even if the newest
1018 * segment is going to be discarded ...", not that it looks very clever
1019 * in case of head skb. Due to potentional receiver driven attacks, we
1020 * choose to avoid immediate execution of a walk in write queue due to
1021 * reneging and defer head skb's loss recovery to standard loss recovery
1022 * procedure that will eventually trigger (nothing forbids us doing this).
1024 * Implements also blockage to start_seq wrap-around. Problem lies in the
1025 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1026 * there's no guarantee that it will be before snd_nxt (n). The problem
1027 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1030 * <- outs wnd -> <- wrapzone ->
1031 * u e n u_w e_w s n_w
1033 * |<------------+------+----- TCP seqno space --------------+---------->|
1034 * ...-- <2^31 ->| |<--------...
1035 * ...---- >2^31 ------>| |<--------...
1037 * Current code wouldn't be vulnerable but it's better still to discard such
1038 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1039 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1040 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1041 * equal to the ideal case (infinite seqno space without wrap caused issues).
1043 * With D-SACK the lower bound is extended to cover sequence space below
1044 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1045 * again, D-SACK block must not to go across snd_una (for the same reason as
1046 * for the normal SACK blocks, explained above). But there all simplicity
1047 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1048 * fully below undo_marker they do not affect behavior in anyway and can
1049 * therefore be safely ignored. In rare cases (which are more or less
1050 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1051 * fragmentation and packet reordering past skb's retransmission. To consider
1052 * them correctly, the acceptable range must be extended even more though
1053 * the exact amount is rather hard to quantify. However, tp->max_window can
1054 * be used as an exaggerated estimate.
1056 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
1057 u32 start_seq, u32 end_seq)
1059 /* Too far in future, or reversed (interpretation is ambiguous) */
1060 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1063 /* Nasty start_seq wrap-around check (see comments above) */
1064 if (!before(start_seq, tp->snd_nxt))
1067 /* In outstanding window? ...This is valid exit for D-SACKs too.
1068 * start_seq == snd_una is non-sensical (see comments above)
1070 if (after(start_seq, tp->snd_una))
1073 if (!is_dsack || !tp->undo_marker)
1076 /* ...Then it's D-SACK, and must reside below snd_una completely */
1077 if (after(end_seq, tp->snd_una))
1080 if (!before(start_seq, tp->undo_marker))
1084 if (!after(end_seq, tp->undo_marker))
1087 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1088 * start_seq < undo_marker and end_seq >= undo_marker.
1090 return !before(start_seq, end_seq - tp->max_window);
1093 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1094 struct tcp_sack_block_wire *sp, int num_sacks,
1097 struct tcp_sock *tp = tcp_sk(sk);
1098 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1099 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1100 bool dup_sack = false;
1102 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1105 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1106 } else if (num_sacks > 1) {
1107 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1108 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1110 if (!after(end_seq_0, end_seq_1) &&
1111 !before(start_seq_0, start_seq_1)) {
1114 NET_INC_STATS(sock_net(sk),
1115 LINUX_MIB_TCPDSACKOFORECV);
1119 /* D-SACK for already forgotten data... Do dumb counting. */
1120 if (dup_sack && tp->undo_marker && tp->undo_retrans > 0 &&
1121 !after(end_seq_0, prior_snd_una) &&
1122 after(end_seq_0, tp->undo_marker))
1128 struct tcp_sacktag_state {
1130 /* Timestamps for earliest and latest never-retransmitted segment
1131 * that was SACKed. RTO needs the earliest RTT to stay conservative,
1132 * but congestion control should still get an accurate delay signal.
1136 struct rate_sample *rate;
1138 unsigned int mss_now;
1141 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1142 * the incoming SACK may not exactly match but we can find smaller MSS
1143 * aligned portion of it that matches. Therefore we might need to fragment
1144 * which may fail and creates some hassle (caller must handle error case
1147 * FIXME: this could be merged to shift decision code
1149 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1150 u32 start_seq, u32 end_seq)
1154 unsigned int pkt_len;
1157 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1158 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1160 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1161 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1162 mss = tcp_skb_mss(skb);
1163 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1166 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1170 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1175 /* Round if necessary so that SACKs cover only full MSSes
1176 * and/or the remaining small portion (if present)
1178 if (pkt_len > mss) {
1179 unsigned int new_len = (pkt_len / mss) * mss;
1180 if (!in_sack && new_len < pkt_len)
1185 if (pkt_len >= skb->len && !in_sack)
1188 err = tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
1189 pkt_len, mss, GFP_ATOMIC);
1197 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1198 static u8 tcp_sacktag_one(struct sock *sk,
1199 struct tcp_sacktag_state *state, u8 sacked,
1200 u32 start_seq, u32 end_seq,
1201 int dup_sack, int pcount,
1204 struct tcp_sock *tp = tcp_sk(sk);
1206 /* Account D-SACK for retransmitted packet. */
1207 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1208 if (tp->undo_marker && tp->undo_retrans > 0 &&
1209 after(end_seq, tp->undo_marker))
1211 if ((sacked & TCPCB_SACKED_ACKED) &&
1212 before(start_seq, state->reord))
1213 state->reord = start_seq;
1216 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1217 if (!after(end_seq, tp->snd_una))
1220 if (!(sacked & TCPCB_SACKED_ACKED)) {
1221 tcp_rack_advance(tp, sacked, end_seq, xmit_time);
1223 if (sacked & TCPCB_SACKED_RETRANS) {
1224 /* If the segment is not tagged as lost,
1225 * we do not clear RETRANS, believing
1226 * that retransmission is still in flight.
1228 if (sacked & TCPCB_LOST) {
1229 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1230 tp->lost_out -= pcount;
1231 tp->retrans_out -= pcount;
1234 if (!(sacked & TCPCB_RETRANS)) {
1235 /* New sack for not retransmitted frame,
1236 * which was in hole. It is reordering.
1238 if (before(start_seq,
1239 tcp_highest_sack_seq(tp)) &&
1240 before(start_seq, state->reord))
1241 state->reord = start_seq;
1243 if (!after(end_seq, tp->high_seq))
1244 state->flag |= FLAG_ORIG_SACK_ACKED;
1245 if (state->first_sackt == 0)
1246 state->first_sackt = xmit_time;
1247 state->last_sackt = xmit_time;
1250 if (sacked & TCPCB_LOST) {
1251 sacked &= ~TCPCB_LOST;
1252 tp->lost_out -= pcount;
1256 sacked |= TCPCB_SACKED_ACKED;
1257 state->flag |= FLAG_DATA_SACKED;
1258 tp->sacked_out += pcount;
1259 tp->delivered += pcount; /* Out-of-order packets delivered */
1261 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1262 if (tp->lost_skb_hint &&
1263 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1264 tp->lost_cnt_hint += pcount;
1267 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1268 * frames and clear it. undo_retrans is decreased above, L|R frames
1269 * are accounted above as well.
1271 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1272 sacked &= ~TCPCB_SACKED_RETRANS;
1273 tp->retrans_out -= pcount;
1279 /* Shift newly-SACKed bytes from this skb to the immediately previous
1280 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1282 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *prev,
1283 struct sk_buff *skb,
1284 struct tcp_sacktag_state *state,
1285 unsigned int pcount, int shifted, int mss,
1288 struct tcp_sock *tp = tcp_sk(sk);
1289 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1290 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
1294 /* Adjust counters and hints for the newly sacked sequence
1295 * range but discard the return value since prev is already
1296 * marked. We must tag the range first because the seq
1297 * advancement below implicitly advances
1298 * tcp_highest_sack_seq() when skb is highest_sack.
1300 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1301 start_seq, end_seq, dup_sack, pcount,
1302 tcp_skb_timestamp_us(skb));
1303 tcp_rate_skb_delivered(sk, skb, state->rate);
1305 if (skb == tp->lost_skb_hint)
1306 tp->lost_cnt_hint += pcount;
1308 TCP_SKB_CB(prev)->end_seq += shifted;
1309 TCP_SKB_CB(skb)->seq += shifted;
1311 tcp_skb_pcount_add(prev, pcount);
1312 WARN_ON_ONCE(tcp_skb_pcount(skb) < pcount);
1313 tcp_skb_pcount_add(skb, -pcount);
1315 /* When we're adding to gso_segs == 1, gso_size will be zero,
1316 * in theory this shouldn't be necessary but as long as DSACK
1317 * code can come after this skb later on it's better to keep
1318 * setting gso_size to something.
1320 if (!TCP_SKB_CB(prev)->tcp_gso_size)
1321 TCP_SKB_CB(prev)->tcp_gso_size = mss;
1323 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1324 if (tcp_skb_pcount(skb) <= 1)
1325 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1327 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1328 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1331 BUG_ON(!tcp_skb_pcount(skb));
1332 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1336 /* Whole SKB was eaten :-) */
1338 if (skb == tp->retransmit_skb_hint)
1339 tp->retransmit_skb_hint = prev;
1340 if (skb == tp->lost_skb_hint) {
1341 tp->lost_skb_hint = prev;
1342 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1345 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1346 TCP_SKB_CB(prev)->eor = TCP_SKB_CB(skb)->eor;
1347 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1348 TCP_SKB_CB(prev)->end_seq++;
1350 if (skb == tcp_highest_sack(sk))
1351 tcp_advance_highest_sack(sk, skb);
1353 tcp_skb_collapse_tstamp(prev, skb);
1354 if (unlikely(TCP_SKB_CB(prev)->tx.delivered_mstamp))
1355 TCP_SKB_CB(prev)->tx.delivered_mstamp = 0;
1357 tcp_rtx_queue_unlink_and_free(skb, sk);
1359 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKMERGED);
1364 /* I wish gso_size would have a bit more sane initialization than
1365 * something-or-zero which complicates things
1367 static int tcp_skb_seglen(const struct sk_buff *skb)
1369 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1372 /* Shifting pages past head area doesn't work */
1373 static int skb_can_shift(const struct sk_buff *skb)
1375 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1378 int tcp_skb_shift(struct sk_buff *to, struct sk_buff *from,
1379 int pcount, int shiftlen)
1381 /* TCP min gso_size is 8 bytes (TCP_MIN_GSO_SIZE)
1382 * Since TCP_SKB_CB(skb)->tcp_gso_segs is 16 bits, we need
1383 * to make sure not storing more than 65535 * 8 bytes per skb,
1384 * even if current MSS is bigger.
1386 if (unlikely(to->len + shiftlen >= 65535 * TCP_MIN_GSO_SIZE))
1388 if (unlikely(tcp_skb_pcount(to) + pcount > 65535))
1390 return skb_shift(to, from, shiftlen);
1393 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1396 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1397 struct tcp_sacktag_state *state,
1398 u32 start_seq, u32 end_seq,
1401 struct tcp_sock *tp = tcp_sk(sk);
1402 struct sk_buff *prev;
1408 /* Normally R but no L won't result in plain S */
1410 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1412 if (!skb_can_shift(skb))
1414 /* This frame is about to be dropped (was ACKed). */
1415 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1418 /* Can only happen with delayed DSACK + discard craziness */
1419 prev = skb_rb_prev(skb);
1423 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1426 if (!tcp_skb_can_collapse_to(prev))
1429 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1430 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1434 pcount = tcp_skb_pcount(skb);
1435 mss = tcp_skb_seglen(skb);
1437 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1438 * drop this restriction as unnecessary
1440 if (mss != tcp_skb_seglen(prev))
1443 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1445 /* CHECKME: This is non-MSS split case only?, this will
1446 * cause skipped skbs due to advancing loop btw, original
1447 * has that feature too
1449 if (tcp_skb_pcount(skb) <= 1)
1452 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1454 /* TODO: head merge to next could be attempted here
1455 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1456 * though it might not be worth of the additional hassle
1458 * ...we can probably just fallback to what was done
1459 * previously. We could try merging non-SACKed ones
1460 * as well but it probably isn't going to buy off
1461 * because later SACKs might again split them, and
1462 * it would make skb timestamp tracking considerably
1468 len = end_seq - TCP_SKB_CB(skb)->seq;
1470 BUG_ON(len > skb->len);
1472 /* MSS boundaries should be honoured or else pcount will
1473 * severely break even though it makes things bit trickier.
1474 * Optimize common case to avoid most of the divides
1476 mss = tcp_skb_mss(skb);
1478 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1479 * drop this restriction as unnecessary
1481 if (mss != tcp_skb_seglen(prev))
1486 } else if (len < mss) {
1494 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1495 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1498 if (!tcp_skb_shift(prev, skb, pcount, len))
1500 if (!tcp_shifted_skb(sk, prev, skb, state, pcount, len, mss, dup_sack))
1503 /* Hole filled allows collapsing with the next as well, this is very
1504 * useful when hole on every nth skb pattern happens
1506 skb = skb_rb_next(prev);
1510 if (!skb_can_shift(skb) ||
1511 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1512 (mss != tcp_skb_seglen(skb)))
1516 pcount = tcp_skb_pcount(skb);
1517 if (tcp_skb_shift(prev, skb, pcount, len))
1518 tcp_shifted_skb(sk, prev, skb, state, pcount,
1528 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1532 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1533 struct tcp_sack_block *next_dup,
1534 struct tcp_sacktag_state *state,
1535 u32 start_seq, u32 end_seq,
1538 struct tcp_sock *tp = tcp_sk(sk);
1539 struct sk_buff *tmp;
1541 skb_rbtree_walk_from(skb) {
1543 bool dup_sack = dup_sack_in;
1545 /* queue is in-order => we can short-circuit the walk early */
1546 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1550 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1551 in_sack = tcp_match_skb_to_sack(sk, skb,
1552 next_dup->start_seq,
1558 /* skb reference here is a bit tricky to get right, since
1559 * shifting can eat and free both this skb and the next,
1560 * so not even _safe variant of the loop is enough.
1563 tmp = tcp_shift_skb_data(sk, skb, state,
1564 start_seq, end_seq, dup_sack);
1573 in_sack = tcp_match_skb_to_sack(sk, skb,
1579 if (unlikely(in_sack < 0))
1583 TCP_SKB_CB(skb)->sacked =
1586 TCP_SKB_CB(skb)->sacked,
1587 TCP_SKB_CB(skb)->seq,
1588 TCP_SKB_CB(skb)->end_seq,
1590 tcp_skb_pcount(skb),
1591 tcp_skb_timestamp_us(skb));
1592 tcp_rate_skb_delivered(sk, skb, state->rate);
1593 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1594 list_del_init(&skb->tcp_tsorted_anchor);
1596 if (!before(TCP_SKB_CB(skb)->seq,
1597 tcp_highest_sack_seq(tp)))
1598 tcp_advance_highest_sack(sk, skb);
1604 static struct sk_buff *tcp_sacktag_bsearch(struct sock *sk, u32 seq)
1606 struct rb_node *parent, **p = &sk->tcp_rtx_queue.rb_node;
1607 struct sk_buff *skb;
1611 skb = rb_to_skb(parent);
1612 if (before(seq, TCP_SKB_CB(skb)->seq)) {
1613 p = &parent->rb_left;
1616 if (!before(seq, TCP_SKB_CB(skb)->end_seq)) {
1617 p = &parent->rb_right;
1625 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1628 if (skb && after(TCP_SKB_CB(skb)->seq, skip_to_seq))
1631 return tcp_sacktag_bsearch(sk, skip_to_seq);
1634 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1636 struct tcp_sack_block *next_dup,
1637 struct tcp_sacktag_state *state,
1643 if (before(next_dup->start_seq, skip_to_seq)) {
1644 skb = tcp_sacktag_skip(skb, sk, next_dup->start_seq);
1645 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1646 next_dup->start_seq, next_dup->end_seq,
1653 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1655 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1659 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1660 u32 prior_snd_una, struct tcp_sacktag_state *state)
1662 struct tcp_sock *tp = tcp_sk(sk);
1663 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1664 TCP_SKB_CB(ack_skb)->sacked);
1665 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1666 struct tcp_sack_block sp[TCP_NUM_SACKS];
1667 struct tcp_sack_block *cache;
1668 struct sk_buff *skb;
1669 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1671 bool found_dup_sack = false;
1673 int first_sack_index;
1676 state->reord = tp->snd_nxt;
1678 if (!tp->sacked_out)
1679 tcp_highest_sack_reset(sk);
1681 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1682 num_sacks, prior_snd_una);
1683 if (found_dup_sack) {
1684 state->flag |= FLAG_DSACKING_ACK;
1685 tp->delivered++; /* A spurious retransmission is delivered */
1688 /* Eliminate too old ACKs, but take into
1689 * account more or less fresh ones, they can
1690 * contain valid SACK info.
1692 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1695 if (!tp->packets_out)
1699 first_sack_index = 0;
1700 for (i = 0; i < num_sacks; i++) {
1701 bool dup_sack = !i && found_dup_sack;
1703 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1704 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1706 if (!tcp_is_sackblock_valid(tp, dup_sack,
1707 sp[used_sacks].start_seq,
1708 sp[used_sacks].end_seq)) {
1712 if (!tp->undo_marker)
1713 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1715 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1717 /* Don't count olds caused by ACK reordering */
1718 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1719 !after(sp[used_sacks].end_seq, tp->snd_una))
1721 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1724 NET_INC_STATS(sock_net(sk), mib_idx);
1726 first_sack_index = -1;
1730 /* Ignore very old stuff early */
1731 if (!after(sp[used_sacks].end_seq, prior_snd_una)) {
1733 first_sack_index = -1;
1740 /* order SACK blocks to allow in order walk of the retrans queue */
1741 for (i = used_sacks - 1; i > 0; i--) {
1742 for (j = 0; j < i; j++) {
1743 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1744 swap(sp[j], sp[j + 1]);
1746 /* Track where the first SACK block goes to */
1747 if (j == first_sack_index)
1748 first_sack_index = j + 1;
1753 state->mss_now = tcp_current_mss(sk);
1757 if (!tp->sacked_out) {
1758 /* It's already past, so skip checking against it */
1759 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1761 cache = tp->recv_sack_cache;
1762 /* Skip empty blocks in at head of the cache */
1763 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1768 while (i < used_sacks) {
1769 u32 start_seq = sp[i].start_seq;
1770 u32 end_seq = sp[i].end_seq;
1771 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1772 struct tcp_sack_block *next_dup = NULL;
1774 if (found_dup_sack && ((i + 1) == first_sack_index))
1775 next_dup = &sp[i + 1];
1777 /* Skip too early cached blocks */
1778 while (tcp_sack_cache_ok(tp, cache) &&
1779 !before(start_seq, cache->end_seq))
1782 /* Can skip some work by looking recv_sack_cache? */
1783 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1784 after(end_seq, cache->start_seq)) {
1787 if (before(start_seq, cache->start_seq)) {
1788 skb = tcp_sacktag_skip(skb, sk, start_seq);
1789 skb = tcp_sacktag_walk(skb, sk, next_dup,
1796 /* Rest of the block already fully processed? */
1797 if (!after(end_seq, cache->end_seq))
1800 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1804 /* ...tail remains todo... */
1805 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1806 /* ...but better entrypoint exists! */
1807 skb = tcp_highest_sack(sk);
1814 skb = tcp_sacktag_skip(skb, sk, cache->end_seq);
1815 /* Check overlap against next cached too (past this one already) */
1820 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1821 skb = tcp_highest_sack(sk);
1825 skb = tcp_sacktag_skip(skb, sk, start_seq);
1828 skb = tcp_sacktag_walk(skb, sk, next_dup, state,
1829 start_seq, end_seq, dup_sack);
1835 /* Clear the head of the cache sack blocks so we can skip it next time */
1836 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1837 tp->recv_sack_cache[i].start_seq = 0;
1838 tp->recv_sack_cache[i].end_seq = 0;
1840 for (j = 0; j < used_sacks; j++)
1841 tp->recv_sack_cache[i++] = sp[j];
1843 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Loss || tp->undo_marker)
1844 tcp_check_sack_reordering(sk, state->reord, 0);
1846 tcp_verify_left_out(tp);
1849 #if FASTRETRANS_DEBUG > 0
1850 WARN_ON((int)tp->sacked_out < 0);
1851 WARN_ON((int)tp->lost_out < 0);
1852 WARN_ON((int)tp->retrans_out < 0);
1853 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1858 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1859 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1861 static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
1865 holes = max(tp->lost_out, 1U);
1866 holes = min(holes, tp->packets_out);
1868 if ((tp->sacked_out + holes) > tp->packets_out) {
1869 tp->sacked_out = tp->packets_out - holes;
1875 /* If we receive more dupacks than we expected counting segments
1876 * in assumption of absent reordering, interpret this as reordering.
1877 * The only another reason could be bug in receiver TCP.
1879 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1881 struct tcp_sock *tp = tcp_sk(sk);
1883 if (!tcp_limit_reno_sacked(tp))
1886 tp->reordering = min_t(u32, tp->packets_out + addend,
1887 sock_net(sk)->ipv4.sysctl_tcp_max_reordering);
1889 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRENOREORDER);
1892 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1894 static void tcp_add_reno_sack(struct sock *sk, int num_dupack)
1897 struct tcp_sock *tp = tcp_sk(sk);
1898 u32 prior_sacked = tp->sacked_out;
1901 tp->sacked_out += num_dupack;
1902 tcp_check_reno_reordering(sk, 0);
1903 delivered = tp->sacked_out - prior_sacked;
1905 tp->delivered += delivered;
1906 tcp_verify_left_out(tp);
1910 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1912 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1914 struct tcp_sock *tp = tcp_sk(sk);
1917 /* One ACK acked hole. The rest eat duplicate ACKs. */
1918 tp->delivered += max_t(int, acked - tp->sacked_out, 1);
1919 if (acked - 1 >= tp->sacked_out)
1922 tp->sacked_out -= acked - 1;
1924 tcp_check_reno_reordering(sk, acked);
1925 tcp_verify_left_out(tp);
1928 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1933 void tcp_clear_retrans(struct tcp_sock *tp)
1935 tp->retrans_out = 0;
1937 tp->undo_marker = 0;
1938 tp->undo_retrans = -1;
1942 static inline void tcp_init_undo(struct tcp_sock *tp)
1944 tp->undo_marker = tp->snd_una;
1945 /* Retransmission still in flight may cause DSACKs later. */
1946 tp->undo_retrans = tp->retrans_out ? : -1;
1949 static bool tcp_is_rack(const struct sock *sk)
1951 return sock_net(sk)->ipv4.sysctl_tcp_recovery & TCP_RACK_LOSS_DETECTION;
1954 /* If we detect SACK reneging, forget all SACK information
1955 * and reset tags completely, otherwise preserve SACKs. If receiver
1956 * dropped its ofo queue, we will know this due to reneging detection.
1958 static void tcp_timeout_mark_lost(struct sock *sk)
1960 struct tcp_sock *tp = tcp_sk(sk);
1961 struct sk_buff *skb, *head;
1962 bool is_reneg; /* is receiver reneging on SACKs? */
1964 head = tcp_rtx_queue_head(sk);
1965 is_reneg = head && (TCP_SKB_CB(head)->sacked & TCPCB_SACKED_ACKED);
1967 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
1969 /* Mark SACK reneging until we recover from this loss event. */
1970 tp->is_sack_reneg = 1;
1971 } else if (tcp_is_reno(tp)) {
1972 tcp_reset_reno_sack(tp);
1976 skb_rbtree_walk_from(skb) {
1978 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1979 else if (tcp_is_rack(sk) && skb != head &&
1980 tcp_rack_skb_timeout(tp, skb, 0) > 0)
1981 continue; /* Don't mark recently sent ones lost yet */
1982 tcp_mark_skb_lost(sk, skb);
1984 tcp_verify_left_out(tp);
1985 tcp_clear_all_retrans_hints(tp);
1988 /* Enter Loss state. */
1989 void tcp_enter_loss(struct sock *sk)
1991 const struct inet_connection_sock *icsk = inet_csk(sk);
1992 struct tcp_sock *tp = tcp_sk(sk);
1993 struct net *net = sock_net(sk);
1994 bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery;
1996 tcp_timeout_mark_lost(sk);
1998 /* Reduce ssthresh if it has not yet been made inside this window. */
1999 if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
2000 !after(tp->high_seq, tp->snd_una) ||
2001 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
2002 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2003 tp->prior_cwnd = tp->snd_cwnd;
2004 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2005 tcp_ca_event(sk, CA_EVENT_LOSS);
2008 tp->snd_cwnd = tcp_packets_in_flight(tp) + 1;
2009 tp->snd_cwnd_cnt = 0;
2010 tp->snd_cwnd_stamp = tcp_jiffies32;
2012 /* Timeout in disordered state after receiving substantial DUPACKs
2013 * suggests that the degree of reordering is over-estimated.
2015 if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
2016 tp->sacked_out >= net->ipv4.sysctl_tcp_reordering)
2017 tp->reordering = min_t(unsigned int, tp->reordering,
2018 net->ipv4.sysctl_tcp_reordering);
2019 tcp_set_ca_state(sk, TCP_CA_Loss);
2020 tp->high_seq = tp->snd_nxt;
2021 tcp_ecn_queue_cwr(tp);
2023 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
2024 * loss recovery is underway except recurring timeout(s) on
2025 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
2027 tp->frto = net->ipv4.sysctl_tcp_frto &&
2028 (new_recovery || icsk->icsk_retransmits) &&
2029 !inet_csk(sk)->icsk_mtup.probe_size;
2032 /* If ACK arrived pointing to a remembered SACK, it means that our
2033 * remembered SACKs do not reflect real state of receiver i.e.
2034 * receiver _host_ is heavily congested (or buggy).
2036 * To avoid big spurious retransmission bursts due to transient SACK
2037 * scoreboard oddities that look like reneging, we give the receiver a
2038 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
2039 * restore sanity to the SACK scoreboard. If the apparent reneging
2040 * persists until this RTO then we'll clear the SACK scoreboard.
2042 static bool tcp_check_sack_reneging(struct sock *sk, int flag)
2044 if (flag & FLAG_SACK_RENEGING) {
2045 struct tcp_sock *tp = tcp_sk(sk);
2046 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4),
2047 msecs_to_jiffies(10));
2049 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2050 delay, TCP_RTO_MAX);
2056 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2057 * counter when SACK is enabled (without SACK, sacked_out is used for
2060 * With reordering, holes may still be in flight, so RFC3517 recovery
2061 * uses pure sacked_out (total number of SACKed segments) even though
2062 * it violates the RFC that uses duplicate ACKs, often these are equal
2063 * but when e.g. out-of-window ACKs or packet duplication occurs,
2064 * they differ. Since neither occurs due to loss, TCP should really
2067 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2069 return tp->sacked_out + 1;
2072 /* Linux NewReno/SACK/ECN state machine.
2073 * --------------------------------------
2075 * "Open" Normal state, no dubious events, fast path.
2076 * "Disorder" In all the respects it is "Open",
2077 * but requires a bit more attention. It is entered when
2078 * we see some SACKs or dupacks. It is split of "Open"
2079 * mainly to move some processing from fast path to slow one.
2080 * "CWR" CWND was reduced due to some Congestion Notification event.
2081 * It can be ECN, ICMP source quench, local device congestion.
2082 * "Recovery" CWND was reduced, we are fast-retransmitting.
2083 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2085 * tcp_fastretrans_alert() is entered:
2086 * - each incoming ACK, if state is not "Open"
2087 * - when arrived ACK is unusual, namely:
2092 * Counting packets in flight is pretty simple.
2094 * in_flight = packets_out - left_out + retrans_out
2096 * packets_out is SND.NXT-SND.UNA counted in packets.
2098 * retrans_out is number of retransmitted segments.
2100 * left_out is number of segments left network, but not ACKed yet.
2102 * left_out = sacked_out + lost_out
2104 * sacked_out: Packets, which arrived to receiver out of order
2105 * and hence not ACKed. With SACKs this number is simply
2106 * amount of SACKed data. Even without SACKs
2107 * it is easy to give pretty reliable estimate of this number,
2108 * counting duplicate ACKs.
2110 * lost_out: Packets lost by network. TCP has no explicit
2111 * "loss notification" feedback from network (for now).
2112 * It means that this number can be only _guessed_.
2113 * Actually, it is the heuristics to predict lossage that
2114 * distinguishes different algorithms.
2116 * F.e. after RTO, when all the queue is considered as lost,
2117 * lost_out = packets_out and in_flight = retrans_out.
2119 * Essentially, we have now a few algorithms detecting
2122 * If the receiver supports SACK:
2124 * RFC6675/3517: It is the conventional algorithm. A packet is
2125 * considered lost if the number of higher sequence packets
2126 * SACKed is greater than or equal the DUPACK thoreshold
2127 * (reordering). This is implemented in tcp_mark_head_lost and
2128 * tcp_update_scoreboard.
2130 * RACK (draft-ietf-tcpm-rack-01): it is a newer algorithm
2131 * (2017-) that checks timing instead of counting DUPACKs.
2132 * Essentially a packet is considered lost if it's not S/ACKed
2133 * after RTT + reordering_window, where both metrics are
2134 * dynamically measured and adjusted. This is implemented in
2135 * tcp_rack_mark_lost.
2137 * If the receiver does not support SACK:
2139 * NewReno (RFC6582): in Recovery we assume that one segment
2140 * is lost (classic Reno). While we are in Recovery and
2141 * a partial ACK arrives, we assume that one more packet
2142 * is lost (NewReno). This heuristics are the same in NewReno
2145 * Really tricky (and requiring careful tuning) part of algorithm
2146 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2147 * The first determines the moment _when_ we should reduce CWND and,
2148 * hence, slow down forward transmission. In fact, it determines the moment
2149 * when we decide that hole is caused by loss, rather than by a reorder.
2151 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2152 * holes, caused by lost packets.
2154 * And the most logically complicated part of algorithm is undo
2155 * heuristics. We detect false retransmits due to both too early
2156 * fast retransmit (reordering) and underestimated RTO, analyzing
2157 * timestamps and D-SACKs. When we detect that some segments were
2158 * retransmitted by mistake and CWND reduction was wrong, we undo
2159 * window reduction and abort recovery phase. This logic is hidden
2160 * inside several functions named tcp_try_undo_<something>.
2163 /* This function decides, when we should leave Disordered state
2164 * and enter Recovery phase, reducing congestion window.
2166 * Main question: may we further continue forward transmission
2167 * with the same cwnd?
2169 static bool tcp_time_to_recover(struct sock *sk, int flag)
2171 struct tcp_sock *tp = tcp_sk(sk);
2173 /* Trick#1: The loss is proven. */
2177 /* Not-A-Trick#2 : Classic rule... */
2178 if (!tcp_is_rack(sk) && tcp_dupack_heuristics(tp) > tp->reordering)
2184 /* Detect loss in event "A" above by marking head of queue up as lost.
2185 * For non-SACK(Reno) senders, the first "packets" number of segments
2186 * are considered lost. 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;
2194 int cnt, oldcnt, lost;
2196 /* Use SACK to deduce losses of new sequences sent during recovery */
2197 const u32 loss_high = tcp_is_sack(tp) ? tp->snd_nxt : tp->high_seq;
2199 WARN_ON(packets > tp->packets_out);
2200 skb = tp->lost_skb_hint;
2202 /* Head already handled? */
2203 if (mark_head && after(TCP_SKB_CB(skb)->seq, tp->snd_una))
2205 cnt = tp->lost_cnt_hint;
2207 skb = tcp_rtx_queue_head(sk);
2211 skb_rbtree_walk_from(skb) {
2212 /* TODO: do this better */
2213 /* this is not the most efficient way to do this... */
2214 tp->lost_skb_hint = skb;
2215 tp->lost_cnt_hint = cnt;
2217 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2221 if (tcp_is_reno(tp) ||
2222 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2223 cnt += tcp_skb_pcount(skb);
2225 if (cnt > packets) {
2226 if (tcp_is_sack(tp) ||
2227 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
2228 (oldcnt >= packets))
2231 mss = tcp_skb_mss(skb);
2232 /* If needed, chop off the prefix to mark as lost. */
2233 lost = (packets - oldcnt) * mss;
2234 if (lost < skb->len &&
2235 tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
2236 lost, mss, GFP_ATOMIC) < 0)
2241 tcp_skb_mark_lost(tp, skb);
2246 tcp_verify_left_out(tp);
2249 /* Account newly detected lost packet(s) */
2251 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2253 struct tcp_sock *tp = tcp_sk(sk);
2255 if (tcp_is_sack(tp)) {
2256 int sacked_upto = tp->sacked_out - tp->reordering;
2257 if (sacked_upto >= 0)
2258 tcp_mark_head_lost(sk, sacked_upto, 0);
2259 else if (fast_rexmit)
2260 tcp_mark_head_lost(sk, 1, 1);
2264 static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when)
2266 return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2267 before(tp->rx_opt.rcv_tsecr, when);
2270 /* skb is spurious retransmitted if the returned timestamp echo
2271 * reply is prior to the skb transmission time
2273 static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp,
2274 const struct sk_buff *skb)
2276 return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) &&
2277 tcp_tsopt_ecr_before(tp, tcp_skb_timestamp(skb));
2280 /* Nothing was retransmitted or returned timestamp is less
2281 * than timestamp of the first retransmission.
2283 static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2285 return tp->retrans_stamp &&
2286 tcp_tsopt_ecr_before(tp, tp->retrans_stamp);
2289 /* Undo procedures. */
2291 /* We can clear retrans_stamp when there are no retransmissions in the
2292 * window. It would seem that it is trivially available for us in
2293 * tp->retrans_out, however, that kind of assumptions doesn't consider
2294 * what will happen if errors occur when sending retransmission for the
2295 * second time. ...It could the that such segment has only
2296 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2297 * the head skb is enough except for some reneging corner cases that
2298 * are not worth the effort.
2300 * Main reason for all this complexity is the fact that connection dying
2301 * time now depends on the validity of the retrans_stamp, in particular,
2302 * that successive retransmissions of a segment must not advance
2303 * retrans_stamp under any conditions.
2305 static bool tcp_any_retrans_done(const struct sock *sk)
2307 const struct tcp_sock *tp = tcp_sk(sk);
2308 struct sk_buff *skb;
2310 if (tp->retrans_out)
2313 skb = tcp_rtx_queue_head(sk);
2314 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2320 static void DBGUNDO(struct sock *sk, const char *msg)
2322 #if FASTRETRANS_DEBUG > 1
2323 struct tcp_sock *tp = tcp_sk(sk);
2324 struct inet_sock *inet = inet_sk(sk);
2326 if (sk->sk_family == AF_INET) {
2327 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2329 &inet->inet_daddr, ntohs(inet->inet_dport),
2330 tp->snd_cwnd, tcp_left_out(tp),
2331 tp->snd_ssthresh, tp->prior_ssthresh,
2334 #if IS_ENABLED(CONFIG_IPV6)
2335 else if (sk->sk_family == AF_INET6) {
2336 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2338 &sk->sk_v6_daddr, ntohs(inet->inet_dport),
2339 tp->snd_cwnd, tcp_left_out(tp),
2340 tp->snd_ssthresh, tp->prior_ssthresh,
2347 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
2349 struct tcp_sock *tp = tcp_sk(sk);
2352 struct sk_buff *skb;
2354 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2355 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2358 tcp_clear_all_retrans_hints(tp);
2361 if (tp->prior_ssthresh) {
2362 const struct inet_connection_sock *icsk = inet_csk(sk);
2364 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2366 if (tp->prior_ssthresh > tp->snd_ssthresh) {
2367 tp->snd_ssthresh = tp->prior_ssthresh;
2368 tcp_ecn_withdraw_cwr(tp);
2371 tp->snd_cwnd_stamp = tcp_jiffies32;
2372 tp->undo_marker = 0;
2373 tp->rack.advanced = 1; /* Force RACK to re-exam losses */
2376 static inline bool tcp_may_undo(const struct tcp_sock *tp)
2378 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2381 /* People celebrate: "We love our President!" */
2382 static bool tcp_try_undo_recovery(struct sock *sk)
2384 struct tcp_sock *tp = tcp_sk(sk);
2386 if (tcp_may_undo(tp)) {
2389 /* Happy end! We did not retransmit anything
2390 * or our original transmission succeeded.
2392 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2393 tcp_undo_cwnd_reduction(sk, false);
2394 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2395 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2397 mib_idx = LINUX_MIB_TCPFULLUNDO;
2399 NET_INC_STATS(sock_net(sk), mib_idx);
2400 } else if (tp->rack.reo_wnd_persist) {
2401 tp->rack.reo_wnd_persist--;
2403 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2404 /* Hold old state until something *above* high_seq
2405 * is ACKed. For Reno it is MUST to prevent false
2406 * fast retransmits (RFC2582). SACK TCP is safe. */
2407 if (!tcp_any_retrans_done(sk))
2408 tp->retrans_stamp = 0;
2411 tcp_set_ca_state(sk, TCP_CA_Open);
2412 tp->is_sack_reneg = 0;
2416 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2417 static bool tcp_try_undo_dsack(struct sock *sk)
2419 struct tcp_sock *tp = tcp_sk(sk);
2421 if (tp->undo_marker && !tp->undo_retrans) {
2422 tp->rack.reo_wnd_persist = min(TCP_RACK_RECOVERY_THRESH,
2423 tp->rack.reo_wnd_persist + 1);
2424 DBGUNDO(sk, "D-SACK");
2425 tcp_undo_cwnd_reduction(sk, false);
2426 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2432 /* Undo during loss recovery after partial ACK or using F-RTO. */
2433 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2435 struct tcp_sock *tp = tcp_sk(sk);
2437 if (frto_undo || tcp_may_undo(tp)) {
2438 tcp_undo_cwnd_reduction(sk, true);
2440 DBGUNDO(sk, "partial loss");
2441 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2443 NET_INC_STATS(sock_net(sk),
2444 LINUX_MIB_TCPSPURIOUSRTOS);
2445 inet_csk(sk)->icsk_retransmits = 0;
2446 if (frto_undo || tcp_is_sack(tp)) {
2447 tcp_set_ca_state(sk, TCP_CA_Open);
2448 tp->is_sack_reneg = 0;
2455 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
2456 * It computes the number of packets to send (sndcnt) based on packets newly
2458 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2459 * cwnd reductions across a full RTT.
2460 * 2) Otherwise PRR uses packet conservation to send as much as delivered.
2461 * But when the retransmits are acked without further losses, PRR
2462 * slow starts cwnd up to ssthresh to speed up the recovery.
2464 static void tcp_init_cwnd_reduction(struct sock *sk)
2466 struct tcp_sock *tp = tcp_sk(sk);
2468 tp->high_seq = tp->snd_nxt;
2469 tp->tlp_high_seq = 0;
2470 tp->snd_cwnd_cnt = 0;
2471 tp->prior_cwnd = tp->snd_cwnd;
2472 tp->prr_delivered = 0;
2474 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2475 tcp_ecn_queue_cwr(tp);
2478 void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked, int flag)
2480 struct tcp_sock *tp = tcp_sk(sk);
2482 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2484 if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd))
2487 tp->prr_delivered += newly_acked_sacked;
2489 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2491 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2492 } else if ((flag & (FLAG_RETRANS_DATA_ACKED | FLAG_LOST_RETRANS)) ==
2493 FLAG_RETRANS_DATA_ACKED) {
2494 sndcnt = min_t(int, delta,
2495 max_t(int, tp->prr_delivered - tp->prr_out,
2496 newly_acked_sacked) + 1);
2498 sndcnt = min(delta, newly_acked_sacked);
2500 /* Force a fast retransmit upon entering fast recovery */
2501 sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1));
2502 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2505 static inline void tcp_end_cwnd_reduction(struct sock *sk)
2507 struct tcp_sock *tp = tcp_sk(sk);
2509 if (inet_csk(sk)->icsk_ca_ops->cong_control)
2512 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2513 if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH &&
2514 (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR || tp->undo_marker)) {
2515 tp->snd_cwnd = tp->snd_ssthresh;
2516 tp->snd_cwnd_stamp = tcp_jiffies32;
2518 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2521 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2522 void tcp_enter_cwr(struct sock *sk)
2524 struct tcp_sock *tp = tcp_sk(sk);
2526 tp->prior_ssthresh = 0;
2527 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2528 tp->undo_marker = 0;
2529 tcp_init_cwnd_reduction(sk);
2530 tcp_set_ca_state(sk, TCP_CA_CWR);
2533 EXPORT_SYMBOL(tcp_enter_cwr);
2535 static void tcp_try_keep_open(struct sock *sk)
2537 struct tcp_sock *tp = tcp_sk(sk);
2538 int state = TCP_CA_Open;
2540 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2541 state = TCP_CA_Disorder;
2543 if (inet_csk(sk)->icsk_ca_state != state) {
2544 tcp_set_ca_state(sk, state);
2545 tp->high_seq = tp->snd_nxt;
2549 static void tcp_try_to_open(struct sock *sk, int flag)
2551 struct tcp_sock *tp = tcp_sk(sk);
2553 tcp_verify_left_out(tp);
2555 if (!tcp_any_retrans_done(sk))
2556 tp->retrans_stamp = 0;
2558 if (flag & FLAG_ECE)
2561 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2562 tcp_try_keep_open(sk);
2566 static void tcp_mtup_probe_failed(struct sock *sk)
2568 struct inet_connection_sock *icsk = inet_csk(sk);
2570 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2571 icsk->icsk_mtup.probe_size = 0;
2572 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPFAIL);
2575 static void tcp_mtup_probe_success(struct sock *sk)
2577 struct tcp_sock *tp = tcp_sk(sk);
2578 struct inet_connection_sock *icsk = inet_csk(sk);
2580 /* FIXME: breaks with very large cwnd */
2581 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2582 tp->snd_cwnd = tp->snd_cwnd *
2583 tcp_mss_to_mtu(sk, tp->mss_cache) /
2584 icsk->icsk_mtup.probe_size;
2585 tp->snd_cwnd_cnt = 0;
2586 tp->snd_cwnd_stamp = tcp_jiffies32;
2587 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2589 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2590 icsk->icsk_mtup.probe_size = 0;
2591 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2592 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS);
2595 /* Do a simple retransmit without using the backoff mechanisms in
2596 * tcp_timer. This is used for path mtu discovery.
2597 * The socket is already locked here.
2599 void tcp_simple_retransmit(struct sock *sk)
2601 const struct inet_connection_sock *icsk = inet_csk(sk);
2602 struct tcp_sock *tp = tcp_sk(sk);
2603 struct sk_buff *skb;
2604 unsigned int mss = tcp_current_mss(sk);
2606 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2607 if (tcp_skb_seglen(skb) > mss &&
2608 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2609 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2610 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2611 tp->retrans_out -= tcp_skb_pcount(skb);
2613 tcp_skb_mark_lost_uncond_verify(tp, skb);
2617 tcp_clear_retrans_hints_partial(tp);
2622 if (tcp_is_reno(tp))
2623 tcp_limit_reno_sacked(tp);
2625 tcp_verify_left_out(tp);
2627 /* Don't muck with the congestion window here.
2628 * Reason is that we do not increase amount of _data_
2629 * in network, but units changed and effective
2630 * cwnd/ssthresh really reduced now.
2632 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2633 tp->high_seq = tp->snd_nxt;
2634 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2635 tp->prior_ssthresh = 0;
2636 tp->undo_marker = 0;
2637 tcp_set_ca_state(sk, TCP_CA_Loss);
2639 tcp_xmit_retransmit_queue(sk);
2641 EXPORT_SYMBOL(tcp_simple_retransmit);
2643 void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2645 struct tcp_sock *tp = tcp_sk(sk);
2648 if (tcp_is_reno(tp))
2649 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2651 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2653 NET_INC_STATS(sock_net(sk), mib_idx);
2655 tp->prior_ssthresh = 0;
2658 if (!tcp_in_cwnd_reduction(sk)) {
2660 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2661 tcp_init_cwnd_reduction(sk);
2663 tcp_set_ca_state(sk, TCP_CA_Recovery);
2666 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2667 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2669 static void tcp_process_loss(struct sock *sk, int flag, int num_dupack,
2672 struct tcp_sock *tp = tcp_sk(sk);
2673 bool recovered = !before(tp->snd_una, tp->high_seq);
2675 if ((flag & FLAG_SND_UNA_ADVANCED || rcu_access_pointer(tp->fastopen_rsk)) &&
2676 tcp_try_undo_loss(sk, false))
2679 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2680 /* Step 3.b. A timeout is spurious if not all data are
2681 * lost, i.e., never-retransmitted data are (s)acked.
2683 if ((flag & FLAG_ORIG_SACK_ACKED) &&
2684 tcp_try_undo_loss(sk, true))
2687 if (after(tp->snd_nxt, tp->high_seq)) {
2688 if (flag & FLAG_DATA_SACKED || num_dupack)
2689 tp->frto = 0; /* Step 3.a. loss was real */
2690 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2691 tp->high_seq = tp->snd_nxt;
2692 /* Step 2.b. Try send new data (but deferred until cwnd
2693 * is updated in tcp_ack()). Otherwise fall back to
2694 * the conventional recovery.
2696 if (!tcp_write_queue_empty(sk) &&
2697 after(tcp_wnd_end(tp), tp->snd_nxt)) {
2698 *rexmit = REXMIT_NEW;
2706 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2707 tcp_try_undo_recovery(sk);
2710 if (tcp_is_reno(tp)) {
2711 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2712 * delivered. Lower inflight to clock out (re)tranmissions.
2714 if (after(tp->snd_nxt, tp->high_seq) && num_dupack)
2715 tcp_add_reno_sack(sk, num_dupack);
2716 else if (flag & FLAG_SND_UNA_ADVANCED)
2717 tcp_reset_reno_sack(tp);
2719 *rexmit = REXMIT_LOST;
2722 /* Undo during fast recovery after partial ACK. */
2723 static bool tcp_try_undo_partial(struct sock *sk, u32 prior_snd_una)
2725 struct tcp_sock *tp = tcp_sk(sk);
2727 if (tp->undo_marker && tcp_packet_delayed(tp)) {
2728 /* Plain luck! Hole if filled with delayed
2729 * packet, rather than with a retransmit. Check reordering.
2731 tcp_check_sack_reordering(sk, prior_snd_una, 1);
2733 /* We are getting evidence that the reordering degree is higher
2734 * than we realized. If there are no retransmits out then we
2735 * can undo. Otherwise we clock out new packets but do not
2736 * mark more packets lost or retransmit more.
2738 if (tp->retrans_out)
2741 if (!tcp_any_retrans_done(sk))
2742 tp->retrans_stamp = 0;
2744 DBGUNDO(sk, "partial recovery");
2745 tcp_undo_cwnd_reduction(sk, true);
2746 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2747 tcp_try_keep_open(sk);
2753 static void tcp_identify_packet_loss(struct sock *sk, int *ack_flag)
2755 struct tcp_sock *tp = tcp_sk(sk);
2757 if (tcp_rtx_queue_empty(sk))
2760 if (unlikely(tcp_is_reno(tp))) {
2761 tcp_newreno_mark_lost(sk, *ack_flag & FLAG_SND_UNA_ADVANCED);
2762 } else if (tcp_is_rack(sk)) {
2763 u32 prior_retrans = tp->retrans_out;
2765 tcp_rack_mark_lost(sk);
2766 if (prior_retrans > tp->retrans_out)
2767 *ack_flag |= FLAG_LOST_RETRANS;
2771 static bool tcp_force_fast_retransmit(struct sock *sk)
2773 struct tcp_sock *tp = tcp_sk(sk);
2775 return after(tcp_highest_sack_seq(tp),
2776 tp->snd_una + tp->reordering * tp->mss_cache);
2779 /* Process an event, which can update packets-in-flight not trivially.
2780 * Main goal of this function is to calculate new estimate for left_out,
2781 * taking into account both packets sitting in receiver's buffer and
2782 * packets lost by network.
2784 * Besides that it updates the congestion state when packet loss or ECN
2785 * is detected. But it does not reduce the cwnd, it is done by the
2786 * congestion control later.
2788 * It does _not_ decide what to send, it is made in function
2789 * tcp_xmit_retransmit_queue().
2791 static void tcp_fastretrans_alert(struct sock *sk, const u32 prior_snd_una,
2792 int num_dupack, int *ack_flag, int *rexmit)
2794 struct inet_connection_sock *icsk = inet_csk(sk);
2795 struct tcp_sock *tp = tcp_sk(sk);
2796 int fast_rexmit = 0, flag = *ack_flag;
2797 bool do_lost = num_dupack || ((flag & FLAG_DATA_SACKED) &&
2798 tcp_force_fast_retransmit(sk));
2800 if (!tp->packets_out && tp->sacked_out)
2803 /* Now state machine starts.
2804 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2805 if (flag & FLAG_ECE)
2806 tp->prior_ssthresh = 0;
2808 /* B. In all the states check for reneging SACKs. */
2809 if (tcp_check_sack_reneging(sk, flag))
2812 /* C. Check consistency of the current state. */
2813 tcp_verify_left_out(tp);
2815 /* D. Check state exit conditions. State can be terminated
2816 * when high_seq is ACKed. */
2817 if (icsk->icsk_ca_state == TCP_CA_Open) {
2818 WARN_ON(tp->retrans_out != 0);
2819 tp->retrans_stamp = 0;
2820 } else if (!before(tp->snd_una, tp->high_seq)) {
2821 switch (icsk->icsk_ca_state) {
2823 /* CWR is to be held something *above* high_seq
2824 * is ACKed for CWR bit to reach receiver. */
2825 if (tp->snd_una != tp->high_seq) {
2826 tcp_end_cwnd_reduction(sk);
2827 tcp_set_ca_state(sk, TCP_CA_Open);
2831 case TCP_CA_Recovery:
2832 if (tcp_is_reno(tp))
2833 tcp_reset_reno_sack(tp);
2834 if (tcp_try_undo_recovery(sk))
2836 tcp_end_cwnd_reduction(sk);
2841 /* E. Process state. */
2842 switch (icsk->icsk_ca_state) {
2843 case TCP_CA_Recovery:
2844 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2845 if (tcp_is_reno(tp))
2846 tcp_add_reno_sack(sk, num_dupack);
2848 if (tcp_try_undo_partial(sk, prior_snd_una))
2850 /* Partial ACK arrived. Force fast retransmit. */
2851 do_lost = tcp_is_reno(tp) ||
2852 tcp_force_fast_retransmit(sk);
2854 if (tcp_try_undo_dsack(sk)) {
2855 tcp_try_keep_open(sk);
2858 tcp_identify_packet_loss(sk, ack_flag);
2861 tcp_process_loss(sk, flag, num_dupack, rexmit);
2862 tcp_identify_packet_loss(sk, ack_flag);
2863 if (!(icsk->icsk_ca_state == TCP_CA_Open ||
2864 (*ack_flag & FLAG_LOST_RETRANS)))
2866 /* Change state if cwnd is undone or retransmits are lost */
2869 if (tcp_is_reno(tp)) {
2870 if (flag & FLAG_SND_UNA_ADVANCED)
2871 tcp_reset_reno_sack(tp);
2872 tcp_add_reno_sack(sk, num_dupack);
2875 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
2876 tcp_try_undo_dsack(sk);
2878 tcp_identify_packet_loss(sk, ack_flag);
2879 if (!tcp_time_to_recover(sk, flag)) {
2880 tcp_try_to_open(sk, flag);
2884 /* MTU probe failure: don't reduce cwnd */
2885 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2886 icsk->icsk_mtup.probe_size &&
2887 tp->snd_una == tp->mtu_probe.probe_seq_start) {
2888 tcp_mtup_probe_failed(sk);
2889 /* Restores the reduction we did in tcp_mtup_probe() */
2891 tcp_simple_retransmit(sk);
2895 /* Otherwise enter Recovery state */
2896 tcp_enter_recovery(sk, (flag & FLAG_ECE));
2900 if (!tcp_is_rack(sk) && do_lost)
2901 tcp_update_scoreboard(sk, fast_rexmit);
2902 *rexmit = REXMIT_LOST;
2905 static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us, const int flag)
2907 u32 wlen = sock_net(sk)->ipv4.sysctl_tcp_min_rtt_wlen * HZ;
2908 struct tcp_sock *tp = tcp_sk(sk);
2910 if ((flag & FLAG_ACK_MAYBE_DELAYED) && rtt_us > tcp_min_rtt(tp)) {
2911 /* If the remote keeps returning delayed ACKs, eventually
2912 * the min filter would pick it up and overestimate the
2913 * prop. delay when it expires. Skip suspected delayed ACKs.
2917 minmax_running_min(&tp->rtt_min, wlen, tcp_jiffies32,
2918 rtt_us ? : jiffies_to_usecs(1));
2921 static bool tcp_ack_update_rtt(struct sock *sk, const int flag,
2922 long seq_rtt_us, long sack_rtt_us,
2923 long ca_rtt_us, struct rate_sample *rs)
2925 const struct tcp_sock *tp = tcp_sk(sk);
2927 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2928 * broken middle-boxes or peers may corrupt TS-ECR fields. But
2929 * Karn's algorithm forbids taking RTT if some retransmitted data
2930 * is acked (RFC6298).
2933 seq_rtt_us = sack_rtt_us;
2935 /* RTTM Rule: A TSecr value received in a segment is used to
2936 * update the averaged RTT measurement only if the segment
2937 * acknowledges some new data, i.e., only if it advances the
2938 * left edge of the send window.
2939 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2941 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2942 flag & FLAG_ACKED) {
2943 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr;
2945 if (likely(delta < INT_MAX / (USEC_PER_SEC / TCP_TS_HZ))) {
2946 seq_rtt_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
2947 ca_rtt_us = seq_rtt_us;
2950 rs->rtt_us = ca_rtt_us; /* RTT of last (S)ACKed packet (or -1) */
2954 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
2955 * always taken together with ACK, SACK, or TS-opts. Any negative
2956 * values will be skipped with the seq_rtt_us < 0 check above.
2958 tcp_update_rtt_min(sk, ca_rtt_us, flag);
2959 tcp_rtt_estimator(sk, seq_rtt_us);
2962 /* RFC6298: only reset backoff on valid RTT measurement. */
2963 inet_csk(sk)->icsk_backoff = 0;
2967 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
2968 void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req)
2970 struct rate_sample rs;
2973 if (req && !req->num_retrans && tcp_rsk(req)->snt_synack)
2974 rtt_us = tcp_stamp_us_delta(tcp_clock_us(), tcp_rsk(req)->snt_synack);
2976 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us, &rs);
2980 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
2982 const struct inet_connection_sock *icsk = inet_csk(sk);
2984 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
2985 tcp_sk(sk)->snd_cwnd_stamp = tcp_jiffies32;
2988 /* Restart timer after forward progress on connection.
2989 * RFC2988 recommends to restart timer to now+rto.
2991 void tcp_rearm_rto(struct sock *sk)
2993 const struct inet_connection_sock *icsk = inet_csk(sk);
2994 struct tcp_sock *tp = tcp_sk(sk);
2996 /* If the retrans timer is currently being used by Fast Open
2997 * for SYN-ACK retrans purpose, stay put.
2999 if (rcu_access_pointer(tp->fastopen_rsk))
3002 if (!tp->packets_out) {
3003 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
3005 u32 rto = inet_csk(sk)->icsk_rto;
3006 /* Offset the time elapsed after installing regular RTO */
3007 if (icsk->icsk_pending == ICSK_TIME_REO_TIMEOUT ||
3008 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
3009 s64 delta_us = tcp_rto_delta_us(sk);
3010 /* delta_us may not be positive if the socket is locked
3011 * when the retrans timer fires and is rescheduled.
3013 rto = usecs_to_jiffies(max_t(int, delta_us, 1));
3015 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
3016 TCP_RTO_MAX, tcp_rtx_queue_head(sk));
3020 /* Try to schedule a loss probe; if that doesn't work, then schedule an RTO. */
3021 static void tcp_set_xmit_timer(struct sock *sk)
3023 if (!tcp_schedule_loss_probe(sk, true))
3027 /* If we get here, the whole TSO packet has not been acked. */
3028 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3030 struct tcp_sock *tp = tcp_sk(sk);
3033 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3035 packets_acked = tcp_skb_pcount(skb);
3036 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3038 packets_acked -= tcp_skb_pcount(skb);
3040 if (packets_acked) {
3041 BUG_ON(tcp_skb_pcount(skb) == 0);
3042 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3045 return packets_acked;
3048 static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
3051 const struct skb_shared_info *shinfo;
3053 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3054 if (likely(!TCP_SKB_CB(skb)->txstamp_ack))
3057 shinfo = skb_shinfo(skb);
3058 if (!before(shinfo->tskey, prior_snd_una) &&
3059 before(shinfo->tskey, tcp_sk(sk)->snd_una)) {
3060 tcp_skb_tsorted_save(skb) {
3061 __skb_tstamp_tx(skb, NULL, sk, SCM_TSTAMP_ACK);
3062 } tcp_skb_tsorted_restore(skb);
3066 /* Remove acknowledged frames from the retransmission queue. If our packet
3067 * is before the ack sequence we can discard it as it's confirmed to have
3068 * arrived at the other end.
3070 static int tcp_clean_rtx_queue(struct sock *sk, u32 prior_fack,
3072 struct tcp_sacktag_state *sack)
3074 const struct inet_connection_sock *icsk = inet_csk(sk);
3075 u64 first_ackt, last_ackt;
3076 struct tcp_sock *tp = tcp_sk(sk);
3077 u32 prior_sacked = tp->sacked_out;
3078 u32 reord = tp->snd_nxt; /* lowest acked un-retx un-sacked seq */
3079 struct sk_buff *skb, *next;
3080 bool fully_acked = true;
3081 long sack_rtt_us = -1L;
3082 long seq_rtt_us = -1L;
3083 long ca_rtt_us = -1L;
3085 u32 last_in_flight = 0;
3091 for (skb = skb_rb_first(&sk->tcp_rtx_queue); skb; skb = next) {
3092 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3093 const u32 start_seq = scb->seq;
3094 u8 sacked = scb->sacked;
3097 tcp_ack_tstamp(sk, skb, prior_snd_una);
3099 /* Determine how many packets and what bytes were acked, tso and else */
3100 if (after(scb->end_seq, tp->snd_una)) {
3101 if (tcp_skb_pcount(skb) == 1 ||
3102 !after(tp->snd_una, scb->seq))
3105 acked_pcount = tcp_tso_acked(sk, skb);
3108 fully_acked = false;
3110 acked_pcount = tcp_skb_pcount(skb);
3113 if (unlikely(sacked & TCPCB_RETRANS)) {
3114 if (sacked & TCPCB_SACKED_RETRANS)
3115 tp->retrans_out -= acked_pcount;
3116 flag |= FLAG_RETRANS_DATA_ACKED;
3117 } else if (!(sacked & TCPCB_SACKED_ACKED)) {
3118 last_ackt = tcp_skb_timestamp_us(skb);
3119 WARN_ON_ONCE(last_ackt == 0);
3121 first_ackt = last_ackt;
3123 last_in_flight = TCP_SKB_CB(skb)->tx.in_flight;
3124 if (before(start_seq, reord))
3126 if (!after(scb->end_seq, tp->high_seq))
3127 flag |= FLAG_ORIG_SACK_ACKED;
3130 if (sacked & TCPCB_SACKED_ACKED) {
3131 tp->sacked_out -= acked_pcount;
3132 } else if (tcp_is_sack(tp)) {
3133 tp->delivered += acked_pcount;
3134 if (!tcp_skb_spurious_retrans(tp, skb))
3135 tcp_rack_advance(tp, sacked, scb->end_seq,
3136 tcp_skb_timestamp_us(skb));
3138 if (sacked & TCPCB_LOST)
3139 tp->lost_out -= acked_pcount;
3141 tp->packets_out -= acked_pcount;
3142 pkts_acked += acked_pcount;
3143 tcp_rate_skb_delivered(sk, skb, sack->rate);
3145 /* Initial outgoing SYN's get put onto the write_queue
3146 * just like anything else we transmit. It is not
3147 * true data, and if we misinform our callers that
3148 * this ACK acks real data, we will erroneously exit
3149 * connection startup slow start one packet too
3150 * quickly. This is severely frowned upon behavior.
3152 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
3153 flag |= FLAG_DATA_ACKED;
3155 flag |= FLAG_SYN_ACKED;
3156 tp->retrans_stamp = 0;
3162 next = skb_rb_next(skb);
3163 if (unlikely(skb == tp->retransmit_skb_hint))
3164 tp->retransmit_skb_hint = NULL;
3165 if (unlikely(skb == tp->lost_skb_hint))
3166 tp->lost_skb_hint = NULL;
3167 tcp_rtx_queue_unlink_and_free(skb, sk);
3171 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
3173 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3174 tp->snd_up = tp->snd_una;
3176 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3177 flag |= FLAG_SACK_RENEGING;
3179 if (likely(first_ackt) && !(flag & FLAG_RETRANS_DATA_ACKED)) {
3180 seq_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, first_ackt);
3181 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, last_ackt);
3183 if (pkts_acked == 1 && last_in_flight < tp->mss_cache &&
3184 last_in_flight && !prior_sacked && fully_acked &&
3185 sack->rate->prior_delivered + 1 == tp->delivered &&
3186 !(flag & (FLAG_CA_ALERT | FLAG_SYN_ACKED))) {
3187 /* Conservatively mark a delayed ACK. It's typically
3188 * from a lone runt packet over the round trip to
3189 * a receiver w/o out-of-order or CE events.
3191 flag |= FLAG_ACK_MAYBE_DELAYED;
3194 if (sack->first_sackt) {
3195 sack_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->first_sackt);
3196 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->last_sackt);
3198 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us,
3199 ca_rtt_us, sack->rate);
3201 if (flag & FLAG_ACKED) {
3202 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3203 if (unlikely(icsk->icsk_mtup.probe_size &&
3204 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3205 tcp_mtup_probe_success(sk);
3208 if (tcp_is_reno(tp)) {
3209 tcp_remove_reno_sacks(sk, pkts_acked);
3211 /* If any of the cumulatively ACKed segments was
3212 * retransmitted, non-SACK case cannot confirm that
3213 * progress was due to original transmission due to
3214 * lack of TCPCB_SACKED_ACKED bits even if some of
3215 * the packets may have been never retransmitted.
3217 if (flag & FLAG_RETRANS_DATA_ACKED)
3218 flag &= ~FLAG_ORIG_SACK_ACKED;
3222 /* Non-retransmitted hole got filled? That's reordering */
3223 if (before(reord, prior_fack))
3224 tcp_check_sack_reordering(sk, reord, 0);
3226 delta = prior_sacked - tp->sacked_out;
3227 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3229 } else if (skb && rtt_update && sack_rtt_us >= 0 &&
3230 sack_rtt_us > tcp_stamp_us_delta(tp->tcp_mstamp,
3231 tcp_skb_timestamp_us(skb))) {
3232 /* Do not re-arm RTO if the sack RTT is measured from data sent
3233 * after when the head was last (re)transmitted. Otherwise the
3234 * timeout may continue to extend in loss recovery.
3236 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3239 if (icsk->icsk_ca_ops->pkts_acked) {
3240 struct ack_sample sample = { .pkts_acked = pkts_acked,
3241 .rtt_us = sack->rate->rtt_us,
3242 .in_flight = last_in_flight };
3244 icsk->icsk_ca_ops->pkts_acked(sk, &sample);
3247 #if FASTRETRANS_DEBUG > 0
3248 WARN_ON((int)tp->sacked_out < 0);
3249 WARN_ON((int)tp->lost_out < 0);
3250 WARN_ON((int)tp->retrans_out < 0);
3251 if (!tp->packets_out && tcp_is_sack(tp)) {
3252 icsk = inet_csk(sk);
3254 pr_debug("Leak l=%u %d\n",
3255 tp->lost_out, icsk->icsk_ca_state);
3258 if (tp->sacked_out) {
3259 pr_debug("Leak s=%u %d\n",
3260 tp->sacked_out, icsk->icsk_ca_state);
3263 if (tp->retrans_out) {
3264 pr_debug("Leak r=%u %d\n",
3265 tp->retrans_out, icsk->icsk_ca_state);
3266 tp->retrans_out = 0;
3273 static void tcp_ack_probe(struct sock *sk)
3275 struct inet_connection_sock *icsk = inet_csk(sk);
3276 struct sk_buff *head = tcp_send_head(sk);
3277 const struct tcp_sock *tp = tcp_sk(sk);
3279 /* Was it a usable window open? */
3282 if (!after(TCP_SKB_CB(head)->end_seq, tcp_wnd_end(tp))) {
3283 icsk->icsk_backoff = 0;
3284 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3285 /* Socket must be waked up by subsequent tcp_data_snd_check().
3286 * This function is not for random using!
3289 unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX);
3291 tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3292 when, TCP_RTO_MAX, NULL);
3296 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3298 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3299 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3302 /* Decide wheather to run the increase function of congestion control. */
3303 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3305 /* If reordering is high then always grow cwnd whenever data is
3306 * delivered regardless of its ordering. Otherwise stay conservative
3307 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3308 * new SACK or ECE mark may first advance cwnd here and later reduce
3309 * cwnd in tcp_fastretrans_alert() based on more states.
3311 if (tcp_sk(sk)->reordering > sock_net(sk)->ipv4.sysctl_tcp_reordering)
3312 return flag & FLAG_FORWARD_PROGRESS;
3314 return flag & FLAG_DATA_ACKED;
3317 /* The "ultimate" congestion control function that aims to replace the rigid
3318 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction).
3319 * It's called toward the end of processing an ACK with precise rate
3320 * information. All transmission or retransmission are delayed afterwards.
3322 static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked,
3323 int flag, const struct rate_sample *rs)
3325 const struct inet_connection_sock *icsk = inet_csk(sk);
3327 if (icsk->icsk_ca_ops->cong_control) {
3328 icsk->icsk_ca_ops->cong_control(sk, rs);
3332 if (tcp_in_cwnd_reduction(sk)) {
3333 /* Reduce cwnd if state mandates */
3334 tcp_cwnd_reduction(sk, acked_sacked, flag);
3335 } else if (tcp_may_raise_cwnd(sk, flag)) {
3336 /* Advance cwnd if state allows */
3337 tcp_cong_avoid(sk, ack, acked_sacked);
3339 tcp_update_pacing_rate(sk);
3342 /* Check that window update is acceptable.
3343 * The function assumes that snd_una<=ack<=snd_next.
3345 static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3346 const u32 ack, const u32 ack_seq,
3349 return after(ack, tp->snd_una) ||
3350 after(ack_seq, tp->snd_wl1) ||
3351 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3354 /* If we update tp->snd_una, also update tp->bytes_acked */
3355 static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack)
3357 u32 delta = ack - tp->snd_una;
3359 sock_owned_by_me((struct sock *)tp);
3360 tp->bytes_acked += delta;
3364 /* If we update tp->rcv_nxt, also update tp->bytes_received */
3365 static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq)
3367 u32 delta = seq - tp->rcv_nxt;
3369 sock_owned_by_me((struct sock *)tp);
3370 tp->bytes_received += delta;
3371 WRITE_ONCE(tp->rcv_nxt, seq);
3374 /* Update our send window.
3376 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3377 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3379 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3382 struct tcp_sock *tp = tcp_sk(sk);
3384 u32 nwin = ntohs(tcp_hdr(skb)->window);
3386 if (likely(!tcp_hdr(skb)->syn))
3387 nwin <<= tp->rx_opt.snd_wscale;
3389 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3390 flag |= FLAG_WIN_UPDATE;
3391 tcp_update_wl(tp, ack_seq);
3393 if (tp->snd_wnd != nwin) {
3396 /* Note, it is the only place, where
3397 * fast path is recovered for sending TCP.
3400 tcp_fast_path_check(sk);
3402 if (!tcp_write_queue_empty(sk))
3403 tcp_slow_start_after_idle_check(sk);
3405 if (nwin > tp->max_window) {
3406 tp->max_window = nwin;
3407 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3412 tcp_snd_una_update(tp, ack);
3417 static bool __tcp_oow_rate_limited(struct net *net, int mib_idx,
3418 u32 *last_oow_ack_time)
3420 if (*last_oow_ack_time) {
3421 s32 elapsed = (s32)(tcp_jiffies32 - *last_oow_ack_time);
3423 if (0 <= elapsed && elapsed < net->ipv4.sysctl_tcp_invalid_ratelimit) {
3424 NET_INC_STATS(net, mib_idx);
3425 return true; /* rate-limited: don't send yet! */
3429 *last_oow_ack_time = tcp_jiffies32;
3431 return false; /* not rate-limited: go ahead, send dupack now! */
3434 /* Return true if we're currently rate-limiting out-of-window ACKs and
3435 * thus shouldn't send a dupack right now. We rate-limit dupacks in
3436 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3437 * attacks that send repeated SYNs or ACKs for the same connection. To
3438 * do this, we do not send a duplicate SYNACK or ACK if the remote
3439 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3441 bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb,
3442 int mib_idx, u32 *last_oow_ack_time)
3444 /* Data packets without SYNs are not likely part of an ACK loop. */
3445 if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) &&
3449 return __tcp_oow_rate_limited(net, mib_idx, last_oow_ack_time);
3452 /* RFC 5961 7 [ACK Throttling] */
3453 static void tcp_send_challenge_ack(struct sock *sk, const struct sk_buff *skb)
3455 /* unprotected vars, we dont care of overwrites */
3456 static u32 challenge_timestamp;
3457 static unsigned int challenge_count;
3458 struct tcp_sock *tp = tcp_sk(sk);
3459 struct net *net = sock_net(sk);
3462 /* First check our per-socket dupack rate limit. */
3463 if (__tcp_oow_rate_limited(net,
3464 LINUX_MIB_TCPACKSKIPPEDCHALLENGE,
3465 &tp->last_oow_ack_time))
3468 /* Then check host-wide RFC 5961 rate limit. */
3470 if (now != challenge_timestamp) {
3471 u32 ack_limit = net->ipv4.sysctl_tcp_challenge_ack_limit;
3472 u32 half = (ack_limit + 1) >> 1;
3474 challenge_timestamp = now;
3475 WRITE_ONCE(challenge_count, half + prandom_u32_max(ack_limit));
3477 count = READ_ONCE(challenge_count);
3479 WRITE_ONCE(challenge_count, count - 1);
3480 NET_INC_STATS(net, LINUX_MIB_TCPCHALLENGEACK);
3485 static void tcp_store_ts_recent(struct tcp_sock *tp)
3487 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3488 tp->rx_opt.ts_recent_stamp = ktime_get_seconds();
3491 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3493 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3494 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3495 * extra check below makes sure this can only happen
3496 * for pure ACK frames. -DaveM
3498 * Not only, also it occurs for expired timestamps.
3501 if (tcp_paws_check(&tp->rx_opt, 0))
3502 tcp_store_ts_recent(tp);
3506 /* This routine deals with acks during a TLP episode.
3507 * We mark the end of a TLP episode on receiving TLP dupack or when
3508 * ack is after tlp_high_seq.
3509 * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3511 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3513 struct tcp_sock *tp = tcp_sk(sk);
3515 if (before(ack, tp->tlp_high_seq))
3518 if (flag & FLAG_DSACKING_ACK) {
3519 /* This DSACK means original and TLP probe arrived; no loss */
3520 tp->tlp_high_seq = 0;
3521 } else if (after(ack, tp->tlp_high_seq)) {
3522 /* ACK advances: there was a loss, so reduce cwnd. Reset
3523 * tlp_high_seq in tcp_init_cwnd_reduction()
3525 tcp_init_cwnd_reduction(sk);
3526 tcp_set_ca_state(sk, TCP_CA_CWR);
3527 tcp_end_cwnd_reduction(sk);
3528 tcp_try_keep_open(sk);
3529 NET_INC_STATS(sock_net(sk),
3530 LINUX_MIB_TCPLOSSPROBERECOVERY);
3531 } else if (!(flag & (FLAG_SND_UNA_ADVANCED |
3532 FLAG_NOT_DUP | FLAG_DATA_SACKED))) {
3533 /* Pure dupack: original and TLP probe arrived; no loss */
3534 tp->tlp_high_seq = 0;
3538 static inline void tcp_in_ack_event(struct sock *sk, u32 flags)
3540 const struct inet_connection_sock *icsk = inet_csk(sk);
3542 if (icsk->icsk_ca_ops->in_ack_event)
3543 icsk->icsk_ca_ops->in_ack_event(sk, flags);
3546 /* Congestion control has updated the cwnd already. So if we're in
3547 * loss recovery then now we do any new sends (for FRTO) or
3548 * retransmits (for CA_Loss or CA_recovery) that make sense.
3550 static void tcp_xmit_recovery(struct sock *sk, int rexmit)
3552 struct tcp_sock *tp = tcp_sk(sk);
3554 if (rexmit == REXMIT_NONE || sk->sk_state == TCP_SYN_SENT)
3557 if (unlikely(rexmit == 2)) {
3558 __tcp_push_pending_frames(sk, tcp_current_mss(sk),
3560 if (after(tp->snd_nxt, tp->high_seq))
3564 tcp_xmit_retransmit_queue(sk);
3567 /* Returns the number of packets newly acked or sacked by the current ACK */
3568 static u32 tcp_newly_delivered(struct sock *sk, u32 prior_delivered, int flag)
3570 const struct net *net = sock_net(sk);
3571 struct tcp_sock *tp = tcp_sk(sk);
3574 delivered = tp->delivered - prior_delivered;
3575 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVERED, delivered);
3576 if (flag & FLAG_ECE) {
3577 tp->delivered_ce += delivered;
3578 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVEREDCE, delivered);
3583 /* This routine deals with incoming acks, but not outgoing ones. */
3584 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3586 struct inet_connection_sock *icsk = inet_csk(sk);
3587 struct tcp_sock *tp = tcp_sk(sk);
3588 struct tcp_sacktag_state sack_state;
3589 struct rate_sample rs = { .prior_delivered = 0 };
3590 u32 prior_snd_una = tp->snd_una;
3591 bool is_sack_reneg = tp->is_sack_reneg;
3592 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3593 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3595 int prior_packets = tp->packets_out;
3596 u32 delivered = tp->delivered;
3597 u32 lost = tp->lost;
3598 int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */
3601 sack_state.first_sackt = 0;
3602 sack_state.rate = &rs;
3604 /* We very likely will need to access rtx queue. */
3605 prefetch(sk->tcp_rtx_queue.rb_node);
3607 /* If the ack is older than previous acks
3608 * then we can probably ignore it.
3610 if (before(ack, prior_snd_una)) {
3611 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3612 if (before(ack, prior_snd_una - tp->max_window)) {
3613 if (!(flag & FLAG_NO_CHALLENGE_ACK))
3614 tcp_send_challenge_ack(sk, skb);
3620 /* If the ack includes data we haven't sent yet, discard
3621 * this segment (RFC793 Section 3.9).
3623 if (after(ack, tp->snd_nxt))
3626 if (after(ack, prior_snd_una)) {
3627 flag |= FLAG_SND_UNA_ADVANCED;
3628 icsk->icsk_retransmits = 0;
3630 #if IS_ENABLED(CONFIG_TLS_DEVICE)
3631 if (static_branch_unlikely(&clean_acked_data_enabled.key))
3632 if (icsk->icsk_clean_acked)
3633 icsk->icsk_clean_acked(sk, ack);
3637 prior_fack = tcp_is_sack(tp) ? tcp_highest_sack_seq(tp) : tp->snd_una;
3638 rs.prior_in_flight = tcp_packets_in_flight(tp);
3640 /* ts_recent update must be made after we are sure that the packet
3643 if (flag & FLAG_UPDATE_TS_RECENT)
3644 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3646 if ((flag & (FLAG_SLOWPATH | FLAG_SND_UNA_ADVANCED)) ==
3647 FLAG_SND_UNA_ADVANCED) {
3648 /* Window is constant, pure forward advance.
3649 * No more checks are required.
3650 * Note, we use the fact that SND.UNA>=SND.WL2.
3652 tcp_update_wl(tp, ack_seq);
3653 tcp_snd_una_update(tp, ack);
3654 flag |= FLAG_WIN_UPDATE;
3656 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
3658 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPACKS);
3660 u32 ack_ev_flags = CA_ACK_SLOWPATH;
3662 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3665 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3667 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3669 if (TCP_SKB_CB(skb)->sacked)
3670 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3673 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) {
3675 ack_ev_flags |= CA_ACK_ECE;
3678 if (flag & FLAG_WIN_UPDATE)
3679 ack_ev_flags |= CA_ACK_WIN_UPDATE;
3681 tcp_in_ack_event(sk, ack_ev_flags);
3684 /* We passed data and got it acked, remove any soft error
3685 * log. Something worked...
3687 sk->sk_err_soft = 0;
3688 icsk->icsk_probes_out = 0;
3689 tp->rcv_tstamp = tcp_jiffies32;
3693 /* See if we can take anything off of the retransmit queue. */
3694 flag |= tcp_clean_rtx_queue(sk, prior_fack, prior_snd_una, &sack_state);
3696 tcp_rack_update_reo_wnd(sk, &rs);
3698 if (tp->tlp_high_seq)
3699 tcp_process_tlp_ack(sk, ack, flag);
3700 /* If needed, reset TLP/RTO timer; RACK may later override this. */
3701 if (flag & FLAG_SET_XMIT_TIMER)
3702 tcp_set_xmit_timer(sk);
3704 if (tcp_ack_is_dubious(sk, flag)) {
3705 if (!(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP))) {
3707 /* Consider if pure acks were aggregated in tcp_add_backlog() */
3708 if (!(flag & FLAG_DATA))
3709 num_dupack = max_t(u16, 1, skb_shinfo(skb)->gso_segs);
3711 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
3715 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
3718 delivered = tcp_newly_delivered(sk, delivered, flag);
3719 lost = tp->lost - lost; /* freshly marked lost */
3720 rs.is_ack_delayed = !!(flag & FLAG_ACK_MAYBE_DELAYED);
3721 tcp_rate_gen(sk, delivered, lost, is_sack_reneg, sack_state.rate);
3722 tcp_cong_control(sk, ack, delivered, flag, sack_state.rate);
3723 tcp_xmit_recovery(sk, rexmit);
3727 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3728 if (flag & FLAG_DSACKING_ACK) {
3729 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
3731 tcp_newly_delivered(sk, delivered, flag);
3733 /* If this ack opens up a zero window, clear backoff. It was
3734 * being used to time the probes, and is probably far higher than
3735 * it needs to be for normal retransmission.
3739 if (tp->tlp_high_seq)
3740 tcp_process_tlp_ack(sk, ack, flag);
3744 /* If data was SACKed, tag it and see if we should send more data.
3745 * If data was DSACKed, see if we can undo a cwnd reduction.
3747 if (TCP_SKB_CB(skb)->sacked) {
3748 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3750 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
3752 tcp_newly_delivered(sk, delivered, flag);
3753 tcp_xmit_recovery(sk, rexmit);
3759 static void tcp_parse_fastopen_option(int len, const unsigned char *cookie,
3760 bool syn, struct tcp_fastopen_cookie *foc,
3763 /* Valid only in SYN or SYN-ACK with an even length. */
3764 if (!foc || !syn || len < 0 || (len & 1))
3767 if (len >= TCP_FASTOPEN_COOKIE_MIN &&
3768 len <= TCP_FASTOPEN_COOKIE_MAX)
3769 memcpy(foc->val, cookie, len);
3776 static void smc_parse_options(const struct tcphdr *th,
3777 struct tcp_options_received *opt_rx,
3778 const unsigned char *ptr,
3781 #if IS_ENABLED(CONFIG_SMC)
3782 if (static_branch_unlikely(&tcp_have_smc)) {
3783 if (th->syn && !(opsize & 1) &&
3784 opsize >= TCPOLEN_EXP_SMC_BASE &&
3785 get_unaligned_be32(ptr) == TCPOPT_SMC_MAGIC)
3791 /* Try to parse the MSS option from the TCP header. Return 0 on failure, clamped
3794 static u16 tcp_parse_mss_option(const struct tcphdr *th, u16 user_mss)
3796 const unsigned char *ptr = (const unsigned char *)(th + 1);
3797 int length = (th->doff * 4) - sizeof(struct tcphdr);
3800 while (length > 0) {
3801 int opcode = *ptr++;
3807 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3814 if (opsize < 2) /* "silly options" */
3816 if (opsize > length)
3817 return mss; /* fail on partial options */
3818 if (opcode == TCPOPT_MSS && opsize == TCPOLEN_MSS) {
3819 u16 in_mss = get_unaligned_be16(ptr);
3822 if (user_mss && user_mss < in_mss)
3834 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3835 * But, this can also be called on packets in the established flow when
3836 * the fast version below fails.
3838 void tcp_parse_options(const struct net *net,
3839 const struct sk_buff *skb,
3840 struct tcp_options_received *opt_rx, int estab,
3841 struct tcp_fastopen_cookie *foc)
3843 const unsigned char *ptr;
3844 const struct tcphdr *th = tcp_hdr(skb);
3845 int length = (th->doff * 4) - sizeof(struct tcphdr);
3847 ptr = (const unsigned char *)(th + 1);
3848 opt_rx->saw_tstamp = 0;
3850 while (length > 0) {
3851 int opcode = *ptr++;
3857 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3864 if (opsize < 2) /* "silly options" */
3866 if (opsize > length)
3867 return; /* don't parse partial options */
3870 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3871 u16 in_mss = get_unaligned_be16(ptr);
3873 if (opt_rx->user_mss &&
3874 opt_rx->user_mss < in_mss)
3875 in_mss = opt_rx->user_mss;
3876 opt_rx->mss_clamp = in_mss;
3881 if (opsize == TCPOLEN_WINDOW && th->syn &&
3882 !estab && net->ipv4.sysctl_tcp_window_scaling) {
3883 __u8 snd_wscale = *(__u8 *)ptr;
3884 opt_rx->wscale_ok = 1;
3885 if (snd_wscale > TCP_MAX_WSCALE) {
3886 net_info_ratelimited("%s: Illegal window scaling value %d > %u received\n",
3890 snd_wscale = TCP_MAX_WSCALE;
3892 opt_rx->snd_wscale = snd_wscale;
3895 case TCPOPT_TIMESTAMP:
3896 if ((opsize == TCPOLEN_TIMESTAMP) &&
3897 ((estab && opt_rx->tstamp_ok) ||
3898 (!estab && net->ipv4.sysctl_tcp_timestamps))) {
3899 opt_rx->saw_tstamp = 1;
3900 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3901 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3904 case TCPOPT_SACK_PERM:
3905 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3906 !estab && net->ipv4.sysctl_tcp_sack) {
3907 opt_rx->sack_ok = TCP_SACK_SEEN;
3908 tcp_sack_reset(opt_rx);
3913 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3914 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3916 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3919 #ifdef CONFIG_TCP_MD5SIG
3922 * The MD5 Hash has already been
3923 * checked (see tcp_v{4,6}_do_rcv()).
3927 case TCPOPT_FASTOPEN:
3928 tcp_parse_fastopen_option(
3929 opsize - TCPOLEN_FASTOPEN_BASE,
3930 ptr, th->syn, foc, false);
3934 /* Fast Open option shares code 254 using a
3935 * 16 bits magic number.
3937 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE &&
3938 get_unaligned_be16(ptr) ==
3939 TCPOPT_FASTOPEN_MAGIC)
3940 tcp_parse_fastopen_option(opsize -
3941 TCPOLEN_EXP_FASTOPEN_BASE,
3942 ptr + 2, th->syn, foc, true);
3944 smc_parse_options(th, opt_rx, ptr,
3954 EXPORT_SYMBOL(tcp_parse_options);
3956 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
3958 const __be32 *ptr = (const __be32 *)(th + 1);
3960 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3961 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3962 tp->rx_opt.saw_tstamp = 1;
3964 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3967 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
3969 tp->rx_opt.rcv_tsecr = 0;
3975 /* Fast parse options. This hopes to only see timestamps.
3976 * If it is wrong it falls back on tcp_parse_options().
3978 static bool tcp_fast_parse_options(const struct net *net,
3979 const struct sk_buff *skb,
3980 const struct tcphdr *th, struct tcp_sock *tp)
3982 /* In the spirit of fast parsing, compare doff directly to constant
3983 * values. Because equality is used, short doff can be ignored here.
3985 if (th->doff == (sizeof(*th) / 4)) {
3986 tp->rx_opt.saw_tstamp = 0;
3988 } else if (tp->rx_opt.tstamp_ok &&
3989 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3990 if (tcp_parse_aligned_timestamp(tp, th))
3994 tcp_parse_options(net, skb, &tp->rx_opt, 1, NULL);
3995 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3996 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
4001 #ifdef CONFIG_TCP_MD5SIG
4003 * Parse MD5 Signature option
4005 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
4007 int length = (th->doff << 2) - sizeof(*th);
4008 const u8 *ptr = (const u8 *)(th + 1);
4010 /* If not enough data remaining, we can short cut */
4011 while (length >= TCPOLEN_MD5SIG) {
4012 int opcode = *ptr++;
4023 if (opsize < 2 || opsize > length)
4025 if (opcode == TCPOPT_MD5SIG)
4026 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
4033 EXPORT_SYMBOL(tcp_parse_md5sig_option);
4036 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
4038 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
4039 * it can pass through stack. So, the following predicate verifies that
4040 * this segment is not used for anything but congestion avoidance or
4041 * fast retransmit. Moreover, we even are able to eliminate most of such
4042 * second order effects, if we apply some small "replay" window (~RTO)
4043 * to timestamp space.
4045 * All these measures still do not guarantee that we reject wrapped ACKs
4046 * on networks with high bandwidth, when sequence space is recycled fastly,
4047 * but it guarantees that such events will be very rare and do not affect
4048 * connection seriously. This doesn't look nice, but alas, PAWS is really
4051 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
4052 * states that events when retransmit arrives after original data are rare.
4053 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
4054 * the biggest problem on large power networks even with minor reordering.
4055 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
4056 * up to bandwidth of 18Gigabit/sec. 8) ]
4059 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
4061 const struct tcp_sock *tp = tcp_sk(sk);
4062 const struct tcphdr *th = tcp_hdr(skb);
4063 u32 seq = TCP_SKB_CB(skb)->seq;
4064 u32 ack = TCP_SKB_CB(skb)->ack_seq;
4066 return (/* 1. Pure ACK with correct sequence number. */
4067 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
4069 /* 2. ... and duplicate ACK. */
4070 ack == tp->snd_una &&
4072 /* 3. ... and does not update window. */
4073 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
4075 /* 4. ... and sits in replay window. */
4076 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
4079 static inline bool tcp_paws_discard(const struct sock *sk,
4080 const struct sk_buff *skb)
4082 const struct tcp_sock *tp = tcp_sk(sk);
4084 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
4085 !tcp_disordered_ack(sk, skb);
4088 /* Check segment sequence number for validity.
4090 * Segment controls are considered valid, if the segment
4091 * fits to the window after truncation to the window. Acceptability
4092 * of data (and SYN, FIN, of course) is checked separately.
4093 * See tcp_data_queue(), for example.
4095 * Also, controls (RST is main one) are accepted using RCV.WUP instead
4096 * of RCV.NXT. Peer still did not advance his SND.UNA when we
4097 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4098 * (borrowed from freebsd)
4101 static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
4103 return !before(end_seq, tp->rcv_wup) &&
4104 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
4107 /* When we get a reset we do this. */
4108 void tcp_reset(struct sock *sk)
4110 trace_tcp_receive_reset(sk);
4112 /* We want the right error as BSD sees it (and indeed as we do). */
4113 switch (sk->sk_state) {
4115 sk->sk_err = ECONNREFUSED;
4117 case TCP_CLOSE_WAIT:
4123 sk->sk_err = ECONNRESET;
4125 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4128 tcp_write_queue_purge(sk);
4131 if (!sock_flag(sk, SOCK_DEAD))
4132 sk->sk_error_report(sk);
4136 * Process the FIN bit. This now behaves as it is supposed to work
4137 * and the FIN takes effect when it is validly part of sequence
4138 * space. Not before when we get holes.
4140 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4141 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4144 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4145 * close and we go into CLOSING (and later onto TIME-WAIT)
4147 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4149 void tcp_fin(struct sock *sk)
4151 struct tcp_sock *tp = tcp_sk(sk);
4153 inet_csk_schedule_ack(sk);
4155 sk->sk_shutdown |= RCV_SHUTDOWN;
4156 sock_set_flag(sk, SOCK_DONE);
4158 switch (sk->sk_state) {
4160 case TCP_ESTABLISHED:
4161 /* Move to CLOSE_WAIT */
4162 tcp_set_state(sk, TCP_CLOSE_WAIT);
4163 inet_csk_enter_pingpong_mode(sk);
4166 case TCP_CLOSE_WAIT:
4168 /* Received a retransmission of the FIN, do
4173 /* RFC793: Remain in the LAST-ACK state. */
4177 /* This case occurs when a simultaneous close
4178 * happens, we must ack the received FIN and
4179 * enter the CLOSING state.
4182 tcp_set_state(sk, TCP_CLOSING);
4185 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4187 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4190 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4191 * cases we should never reach this piece of code.
4193 pr_err("%s: Impossible, sk->sk_state=%d\n",
4194 __func__, sk->sk_state);
4198 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4199 * Probably, we should reset in this case. For now drop them.
4201 skb_rbtree_purge(&tp->out_of_order_queue);
4202 if (tcp_is_sack(tp))
4203 tcp_sack_reset(&tp->rx_opt);
4206 if (!sock_flag(sk, SOCK_DEAD)) {
4207 sk->sk_state_change(sk);
4209 /* Do not send POLL_HUP for half duplex close. */
4210 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4211 sk->sk_state == TCP_CLOSE)
4212 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4214 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4218 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4221 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4222 if (before(seq, sp->start_seq))
4223 sp->start_seq = seq;
4224 if (after(end_seq, sp->end_seq))
4225 sp->end_seq = end_seq;
4231 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4233 struct tcp_sock *tp = tcp_sk(sk);
4235 if (tcp_is_sack(tp) && sock_net(sk)->ipv4.sysctl_tcp_dsack) {
4238 if (before(seq, tp->rcv_nxt))
4239 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4241 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4243 NET_INC_STATS(sock_net(sk), mib_idx);
4245 tp->rx_opt.dsack = 1;
4246 tp->duplicate_sack[0].start_seq = seq;
4247 tp->duplicate_sack[0].end_seq = end_seq;
4251 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4253 struct tcp_sock *tp = tcp_sk(sk);
4255 if (!tp->rx_opt.dsack)
4256 tcp_dsack_set(sk, seq, end_seq);
4258 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4261 static void tcp_rcv_spurious_retrans(struct sock *sk, const struct sk_buff *skb)
4263 /* When the ACK path fails or drops most ACKs, the sender would
4264 * timeout and spuriously retransmit the same segment repeatedly.
4265 * The receiver remembers and reflects via DSACKs. Leverage the
4266 * DSACK state and change the txhash to re-route speculatively.
4268 if (TCP_SKB_CB(skb)->seq == tcp_sk(sk)->duplicate_sack[0].start_seq)
4269 sk_rethink_txhash(sk);
4272 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4274 struct tcp_sock *tp = tcp_sk(sk);
4276 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4277 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4278 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4279 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
4281 if (tcp_is_sack(tp) && sock_net(sk)->ipv4.sysctl_tcp_dsack) {
4282 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4284 tcp_rcv_spurious_retrans(sk, skb);
4285 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4286 end_seq = tp->rcv_nxt;
4287 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4294 /* These routines update the SACK block as out-of-order packets arrive or
4295 * in-order packets close up the sequence space.
4297 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4300 struct tcp_sack_block *sp = &tp->selective_acks[0];
4301 struct tcp_sack_block *swalk = sp + 1;
4303 /* See if the recent change to the first SACK eats into
4304 * or hits the sequence space of other SACK blocks, if so coalesce.
4306 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4307 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4310 /* Zap SWALK, by moving every further SACK up by one slot.
4311 * Decrease num_sacks.
4313 tp->rx_opt.num_sacks--;
4314 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4318 this_sack++, swalk++;
4322 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4324 struct tcp_sock *tp = tcp_sk(sk);
4325 struct tcp_sack_block *sp = &tp->selective_acks[0];
4326 int cur_sacks = tp->rx_opt.num_sacks;
4332 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4333 if (tcp_sack_extend(sp, seq, end_seq)) {
4334 /* Rotate this_sack to the first one. */
4335 for (; this_sack > 0; this_sack--, sp--)
4336 swap(*sp, *(sp - 1));
4338 tcp_sack_maybe_coalesce(tp);
4343 /* Could not find an adjacent existing SACK, build a new one,
4344 * put it at the front, and shift everyone else down. We
4345 * always know there is at least one SACK present already here.
4347 * If the sack array is full, forget about the last one.
4349 if (this_sack >= TCP_NUM_SACKS) {
4350 if (tp->compressed_ack > TCP_FASTRETRANS_THRESH)
4353 tp->rx_opt.num_sacks--;
4356 for (; this_sack > 0; this_sack--, sp--)
4360 /* Build the new head SACK, and we're done. */
4361 sp->start_seq = seq;
4362 sp->end_seq = end_seq;
4363 tp->rx_opt.num_sacks++;
4366 /* RCV.NXT advances, some SACKs should be eaten. */
4368 static void tcp_sack_remove(struct tcp_sock *tp)
4370 struct tcp_sack_block *sp = &tp->selective_acks[0];
4371 int num_sacks = tp->rx_opt.num_sacks;
4374 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4375 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4376 tp->rx_opt.num_sacks = 0;
4380 for (this_sack = 0; this_sack < num_sacks;) {
4381 /* Check if the start of the sack is covered by RCV.NXT. */
4382 if (!before(tp->rcv_nxt, sp->start_seq)) {
4385 /* RCV.NXT must cover all the block! */
4386 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4388 /* Zap this SACK, by moving forward any other SACKS. */
4389 for (i = this_sack+1; i < num_sacks; i++)
4390 tp->selective_acks[i-1] = tp->selective_acks[i];
4397 tp->rx_opt.num_sacks = num_sacks;
4401 * tcp_try_coalesce - try to merge skb to prior one
4403 * @dest: destination queue
4405 * @from: buffer to add in queue
4406 * @fragstolen: pointer to boolean
4408 * Before queueing skb @from after @to, try to merge them
4409 * to reduce overall memory use and queue lengths, if cost is small.
4410 * Packets in ofo or receive queues can stay a long time.
4411 * Better try to coalesce them right now to avoid future collapses.
4412 * Returns true if caller should free @from instead of queueing it
4414 static bool tcp_try_coalesce(struct sock *sk,
4416 struct sk_buff *from,
4421 *fragstolen = false;
4423 /* Its possible this segment overlaps with prior segment in queue */
4424 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4427 #ifdef CONFIG_TLS_DEVICE
4428 if (from->decrypted != to->decrypted)
4432 if (!skb_try_coalesce(to, from, fragstolen, &delta))
4435 atomic_add(delta, &sk->sk_rmem_alloc);
4436 sk_mem_charge(sk, delta);
4437 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4438 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4439 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4440 TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
4442 if (TCP_SKB_CB(from)->has_rxtstamp) {
4443 TCP_SKB_CB(to)->has_rxtstamp = true;
4444 to->tstamp = from->tstamp;
4445 skb_hwtstamps(to)->hwtstamp = skb_hwtstamps(from)->hwtstamp;
4451 static bool tcp_ooo_try_coalesce(struct sock *sk,
4453 struct sk_buff *from,
4456 bool res = tcp_try_coalesce(sk, to, from, fragstolen);
4458 /* In case tcp_drop() is called later, update to->gso_segs */
4460 u32 gso_segs = max_t(u16, 1, skb_shinfo(to)->gso_segs) +
4461 max_t(u16, 1, skb_shinfo(from)->gso_segs);
4463 skb_shinfo(to)->gso_segs = min_t(u32, gso_segs, 0xFFFF);
4468 static void tcp_drop(struct sock *sk, struct sk_buff *skb)
4470 sk_drops_add(sk, skb);
4474 /* This one checks to see if we can put data from the
4475 * out_of_order queue into the receive_queue.
4477 static void tcp_ofo_queue(struct sock *sk)
4479 struct tcp_sock *tp = tcp_sk(sk);
4480 __u32 dsack_high = tp->rcv_nxt;
4481 bool fin, fragstolen, eaten;
4482 struct sk_buff *skb, *tail;
4485 p = rb_first(&tp->out_of_order_queue);
4488 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4491 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4492 __u32 dsack = dsack_high;
4493 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4494 dsack_high = TCP_SKB_CB(skb)->end_seq;
4495 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4498 rb_erase(&skb->rbnode, &tp->out_of_order_queue);
4500 if (unlikely(!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))) {
4505 tail = skb_peek_tail(&sk->sk_receive_queue);
4506 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen);
4507 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4508 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
4510 __skb_queue_tail(&sk->sk_receive_queue, skb);
4512 kfree_skb_partial(skb, fragstolen);
4514 if (unlikely(fin)) {
4516 /* tcp_fin() purges tp->out_of_order_queue,
4517 * so we must end this loop right now.
4524 static bool tcp_prune_ofo_queue(struct sock *sk);
4525 static int tcp_prune_queue(struct sock *sk);
4527 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4530 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4531 !sk_rmem_schedule(sk, skb, size)) {
4533 if (tcp_prune_queue(sk) < 0)
4536 while (!sk_rmem_schedule(sk, skb, size)) {
4537 if (!tcp_prune_ofo_queue(sk))
4544 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4546 struct tcp_sock *tp = tcp_sk(sk);
4547 struct rb_node **p, *parent;
4548 struct sk_buff *skb1;
4552 tcp_ecn_check_ce(sk, skb);
4554 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4555 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFODROP);
4560 /* Disable header prediction. */
4562 inet_csk_schedule_ack(sk);
4564 tp->rcv_ooopack += max_t(u16, 1, skb_shinfo(skb)->gso_segs);
4565 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4566 seq = TCP_SKB_CB(skb)->seq;
4567 end_seq = TCP_SKB_CB(skb)->end_seq;
4569 p = &tp->out_of_order_queue.rb_node;
4570 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4571 /* Initial out of order segment, build 1 SACK. */
4572 if (tcp_is_sack(tp)) {
4573 tp->rx_opt.num_sacks = 1;
4574 tp->selective_acks[0].start_seq = seq;
4575 tp->selective_acks[0].end_seq = end_seq;
4577 rb_link_node(&skb->rbnode, NULL, p);
4578 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4579 tp->ooo_last_skb = skb;
4583 /* In the typical case, we are adding an skb to the end of the list.
4584 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
4586 if (tcp_ooo_try_coalesce(sk, tp->ooo_last_skb,
4587 skb, &fragstolen)) {
4589 tcp_grow_window(sk, skb);
4590 kfree_skb_partial(skb, fragstolen);
4594 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
4595 if (!before(seq, TCP_SKB_CB(tp->ooo_last_skb)->end_seq)) {
4596 parent = &tp->ooo_last_skb->rbnode;
4597 p = &parent->rb_right;
4601 /* Find place to insert this segment. Handle overlaps on the way. */
4605 skb1 = rb_to_skb(parent);
4606 if (before(seq, TCP_SKB_CB(skb1)->seq)) {
4607 p = &parent->rb_left;
4610 if (before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4611 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4612 /* All the bits are present. Drop. */
4613 NET_INC_STATS(sock_net(sk),
4614 LINUX_MIB_TCPOFOMERGE);
4617 tcp_dsack_set(sk, seq, end_seq);
4620 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4621 /* Partial overlap. */
4622 tcp_dsack_set(sk, seq, TCP_SKB_CB(skb1)->end_seq);
4624 /* skb's seq == skb1's seq and skb covers skb1.
4625 * Replace skb1 with skb.
4627 rb_replace_node(&skb1->rbnode, &skb->rbnode,
4628 &tp->out_of_order_queue);
4629 tcp_dsack_extend(sk,
4630 TCP_SKB_CB(skb1)->seq,
4631 TCP_SKB_CB(skb1)->end_seq);
4632 NET_INC_STATS(sock_net(sk),
4633 LINUX_MIB_TCPOFOMERGE);
4637 } else if (tcp_ooo_try_coalesce(sk, skb1,
4638 skb, &fragstolen)) {
4641 p = &parent->rb_right;
4644 /* Insert segment into RB tree. */
4645 rb_link_node(&skb->rbnode, parent, p);
4646 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4649 /* Remove other segments covered by skb. */
4650 while ((skb1 = skb_rb_next(skb)) != NULL) {
4651 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4653 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4654 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4658 rb_erase(&skb1->rbnode, &tp->out_of_order_queue);
4659 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4660 TCP_SKB_CB(skb1)->end_seq);
4661 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4664 /* If there is no skb after us, we are the last_skb ! */
4666 tp->ooo_last_skb = skb;
4669 if (tcp_is_sack(tp))
4670 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4673 tcp_grow_window(sk, skb);
4675 skb_set_owner_r(skb, sk);
4679 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb,
4683 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4686 tcp_try_coalesce(sk, tail,
4687 skb, fragstolen)) ? 1 : 0;
4688 tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq);
4690 __skb_queue_tail(&sk->sk_receive_queue, skb);
4691 skb_set_owner_r(skb, sk);
4696 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4698 struct sk_buff *skb;
4706 if (size > PAGE_SIZE) {
4707 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS);
4709 data_len = npages << PAGE_SHIFT;
4710 size = data_len + (size & ~PAGE_MASK);
4712 skb = alloc_skb_with_frags(size - data_len, data_len,
4713 PAGE_ALLOC_COSTLY_ORDER,
4714 &err, sk->sk_allocation);
4718 skb_put(skb, size - data_len);
4719 skb->data_len = data_len;
4722 if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) {
4723 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP);
4727 err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size);
4731 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4732 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4733 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4735 if (tcp_queue_rcv(sk, skb, &fragstolen)) {
4736 WARN_ON_ONCE(fragstolen); /* should not happen */
4748 void tcp_data_ready(struct sock *sk)
4750 const struct tcp_sock *tp = tcp_sk(sk);
4751 int avail = tp->rcv_nxt - tp->copied_seq;
4753 if (avail < sk->sk_rcvlowat && !sock_flag(sk, SOCK_DONE))
4756 sk->sk_data_ready(sk);
4759 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4761 struct tcp_sock *tp = tcp_sk(sk);
4765 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
4770 __skb_pull(skb, tcp_hdr(skb)->doff * 4);
4772 tcp_ecn_accept_cwr(sk, skb);
4774 tp->rx_opt.dsack = 0;
4776 /* Queue data for delivery to the user.
4777 * Packets in sequence go to the receive queue.
4778 * Out of sequence packets to the out_of_order_queue.
4780 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4781 if (tcp_receive_window(tp) == 0) {
4782 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP);
4786 /* Ok. In sequence. In window. */
4788 if (skb_queue_len(&sk->sk_receive_queue) == 0)
4789 sk_forced_mem_schedule(sk, skb->truesize);
4790 else if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) {
4791 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP);
4795 eaten = tcp_queue_rcv(sk, skb, &fragstolen);
4797 tcp_event_data_recv(sk, skb);
4798 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
4801 if (!RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4804 /* RFC5681. 4.2. SHOULD send immediate ACK, when
4805 * gap in queue is filled.
4807 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
4808 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW;
4811 if (tp->rx_opt.num_sacks)
4812 tcp_sack_remove(tp);
4814 tcp_fast_path_check(sk);
4817 kfree_skb_partial(skb, fragstolen);
4818 if (!sock_flag(sk, SOCK_DEAD))
4823 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4824 tcp_rcv_spurious_retrans(sk, skb);
4825 /* A retransmit, 2nd most common case. Force an immediate ack. */
4826 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4827 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4830 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
4831 inet_csk_schedule_ack(sk);
4837 /* Out of window. F.e. zero window probe. */
4838 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4841 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4842 /* Partial packet, seq < rcv_next < end_seq */
4843 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4845 /* If window is closed, drop tail of packet. But after
4846 * remembering D-SACK for its head made in previous line.
4848 if (!tcp_receive_window(tp)) {
4849 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP);
4855 tcp_data_queue_ofo(sk, skb);
4858 static struct sk_buff *tcp_skb_next(struct sk_buff *skb, struct sk_buff_head *list)
4861 return !skb_queue_is_last(list, skb) ? skb->next : NULL;
4863 return skb_rb_next(skb);
4866 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4867 struct sk_buff_head *list,
4868 struct rb_root *root)
4870 struct sk_buff *next = tcp_skb_next(skb, list);
4873 __skb_unlink(skb, list);
4875 rb_erase(&skb->rbnode, root);
4878 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4883 /* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */
4884 void tcp_rbtree_insert(struct rb_root *root, struct sk_buff *skb)
4886 struct rb_node **p = &root->rb_node;
4887 struct rb_node *parent = NULL;
4888 struct sk_buff *skb1;
4892 skb1 = rb_to_skb(parent);
4893 if (before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb1)->seq))
4894 p = &parent->rb_left;
4896 p = &parent->rb_right;
4898 rb_link_node(&skb->rbnode, parent, p);
4899 rb_insert_color(&skb->rbnode, root);
4902 /* Collapse contiguous sequence of skbs head..tail with
4903 * sequence numbers start..end.
4905 * If tail is NULL, this means until the end of the queue.
4907 * Segments with FIN/SYN are not collapsed (only because this
4911 tcp_collapse(struct sock *sk, struct sk_buff_head *list, struct rb_root *root,
4912 struct sk_buff *head, struct sk_buff *tail, u32 start, u32 end)
4914 struct sk_buff *skb = head, *n;
4915 struct sk_buff_head tmp;
4918 /* First, check that queue is collapsible and find
4919 * the point where collapsing can be useful.
4922 for (end_of_skbs = true; skb != NULL && skb != tail; skb = n) {
4923 n = tcp_skb_next(skb, list);
4925 /* No new bits? It is possible on ofo queue. */
4926 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4927 skb = tcp_collapse_one(sk, skb, list, root);
4933 /* The first skb to collapse is:
4935 * - bloated or contains data before "start" or
4936 * overlaps to the next one.
4938 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
4939 (tcp_win_from_space(sk, skb->truesize) > skb->len ||
4940 before(TCP_SKB_CB(skb)->seq, start))) {
4941 end_of_skbs = false;
4945 if (n && n != tail &&
4946 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(n)->seq) {
4947 end_of_skbs = false;
4951 /* Decided to skip this, advance start seq. */
4952 start = TCP_SKB_CB(skb)->end_seq;
4955 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4958 __skb_queue_head_init(&tmp);
4960 while (before(start, end)) {
4961 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
4962 struct sk_buff *nskb;
4964 nskb = alloc_skb(copy, GFP_ATOMIC);
4968 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4969 #ifdef CONFIG_TLS_DEVICE
4970 nskb->decrypted = skb->decrypted;
4972 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4974 __skb_queue_before(list, skb, nskb);
4976 __skb_queue_tail(&tmp, nskb); /* defer rbtree insertion */
4977 skb_set_owner_r(nskb, sk);
4979 /* Copy data, releasing collapsed skbs. */
4981 int offset = start - TCP_SKB_CB(skb)->seq;
4982 int size = TCP_SKB_CB(skb)->end_seq - start;
4986 size = min(copy, size);
4987 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4989 TCP_SKB_CB(nskb)->end_seq += size;
4993 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4994 skb = tcp_collapse_one(sk, skb, list, root);
4997 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4999 #ifdef CONFIG_TLS_DEVICE
5000 if (skb->decrypted != nskb->decrypted)
5007 skb_queue_walk_safe(&tmp, skb, n)
5008 tcp_rbtree_insert(root, skb);
5011 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
5012 * and tcp_collapse() them until all the queue is collapsed.
5014 static void tcp_collapse_ofo_queue(struct sock *sk)
5016 struct tcp_sock *tp = tcp_sk(sk);
5017 u32 range_truesize, sum_tiny = 0;
5018 struct sk_buff *skb, *head;
5021 skb = skb_rb_first(&tp->out_of_order_queue);
5024 tp->ooo_last_skb = skb_rb_last(&tp->out_of_order_queue);
5027 start = TCP_SKB_CB(skb)->seq;
5028 end = TCP_SKB_CB(skb)->end_seq;
5029 range_truesize = skb->truesize;
5031 for (head = skb;;) {
5032 skb = skb_rb_next(skb);
5034 /* Range is terminated when we see a gap or when
5035 * we are at the queue end.
5038 after(TCP_SKB_CB(skb)->seq, end) ||
5039 before(TCP_SKB_CB(skb)->end_seq, start)) {
5040 /* Do not attempt collapsing tiny skbs */
5041 if (range_truesize != head->truesize ||
5042 end - start >= SKB_WITH_OVERHEAD(SK_MEM_QUANTUM)) {
5043 tcp_collapse(sk, NULL, &tp->out_of_order_queue,
5044 head, skb, start, end);
5046 sum_tiny += range_truesize;
5047 if (sum_tiny > sk->sk_rcvbuf >> 3)
5053 range_truesize += skb->truesize;
5054 if (unlikely(before(TCP_SKB_CB(skb)->seq, start)))
5055 start = TCP_SKB_CB(skb)->seq;
5056 if (after(TCP_SKB_CB(skb)->end_seq, end))
5057 end = TCP_SKB_CB(skb)->end_seq;
5062 * Clean the out-of-order queue to make room.
5063 * We drop high sequences packets to :
5064 * 1) Let a chance for holes to be filled.
5065 * 2) not add too big latencies if thousands of packets sit there.
5066 * (But if application shrinks SO_RCVBUF, we could still end up
5067 * freeing whole queue here)
5068 * 3) Drop at least 12.5 % of sk_rcvbuf to avoid malicious attacks.
5070 * Return true if queue has shrunk.
5072 static bool tcp_prune_ofo_queue(struct sock *sk)
5074 struct tcp_sock *tp = tcp_sk(sk);
5075 struct rb_node *node, *prev;
5078 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
5081 NET_INC_STATS(sock_net(sk), LINUX_MIB_OFOPRUNED);
5082 goal = sk->sk_rcvbuf >> 3;
5083 node = &tp->ooo_last_skb->rbnode;
5085 prev = rb_prev(node);
5086 rb_erase(node, &tp->out_of_order_queue);
5087 goal -= rb_to_skb(node)->truesize;
5088 tcp_drop(sk, rb_to_skb(node));
5089 if (!prev || goal <= 0) {
5091 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf &&
5092 !tcp_under_memory_pressure(sk))
5094 goal = sk->sk_rcvbuf >> 3;
5098 tp->ooo_last_skb = rb_to_skb(prev);
5100 /* Reset SACK state. A conforming SACK implementation will
5101 * do the same at a timeout based retransmit. When a connection
5102 * is in a sad state like this, we care only about integrity
5103 * of the connection not performance.
5105 if (tp->rx_opt.sack_ok)
5106 tcp_sack_reset(&tp->rx_opt);
5110 /* Reduce allocated memory if we can, trying to get
5111 * the socket within its memory limits again.
5113 * Return less than zero if we should start dropping frames
5114 * until the socket owning process reads some of the data
5115 * to stabilize the situation.
5117 static int tcp_prune_queue(struct sock *sk)
5119 struct tcp_sock *tp = tcp_sk(sk);
5121 NET_INC_STATS(sock_net(sk), LINUX_MIB_PRUNECALLED);
5123 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
5124 tcp_clamp_window(sk);
5125 else if (tcp_under_memory_pressure(sk))
5126 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
5128 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5131 tcp_collapse_ofo_queue(sk);
5132 if (!skb_queue_empty(&sk->sk_receive_queue))
5133 tcp_collapse(sk, &sk->sk_receive_queue, NULL,
5134 skb_peek(&sk->sk_receive_queue),
5136 tp->copied_seq, tp->rcv_nxt);
5139 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5142 /* Collapsing did not help, destructive actions follow.
5143 * This must not ever occur. */
5145 tcp_prune_ofo_queue(sk);
5147 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5150 /* If we are really being abused, tell the caller to silently
5151 * drop receive data on the floor. It will get retransmitted
5152 * and hopefully then we'll have sufficient space.
5154 NET_INC_STATS(sock_net(sk), LINUX_MIB_RCVPRUNED);
5156 /* Massive buffer overcommit. */
5161 static bool tcp_should_expand_sndbuf(const struct sock *sk)
5163 const struct tcp_sock *tp = tcp_sk(sk);
5165 /* If the user specified a specific send buffer setting, do
5168 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
5171 /* If we are under global TCP memory pressure, do not expand. */
5172 if (tcp_under_memory_pressure(sk))
5175 /* If we are under soft global TCP memory pressure, do not expand. */
5176 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
5179 /* If we filled the congestion window, do not expand. */
5180 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
5186 /* When incoming ACK allowed to free some skb from write_queue,
5187 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
5188 * on the exit from tcp input handler.
5190 * PROBLEM: sndbuf expansion does not work well with largesend.
5192 static void tcp_new_space(struct sock *sk)
5194 struct tcp_sock *tp = tcp_sk(sk);
5196 if (tcp_should_expand_sndbuf(sk)) {
5197 tcp_sndbuf_expand(sk);
5198 tp->snd_cwnd_stamp = tcp_jiffies32;
5201 sk->sk_write_space(sk);
5204 static void tcp_check_space(struct sock *sk)
5206 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
5207 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
5208 /* pairs with tcp_poll() */
5210 if (sk->sk_socket &&
5211 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
5213 if (!test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
5214 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
5219 static inline void tcp_data_snd_check(struct sock *sk)
5221 tcp_push_pending_frames(sk);
5222 tcp_check_space(sk);
5226 * Check if sending an ack is needed.
5228 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
5230 struct tcp_sock *tp = tcp_sk(sk);
5231 unsigned long rtt, delay;
5233 /* More than one full frame received... */
5234 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
5235 /* ... and right edge of window advances far enough.
5236 * (tcp_recvmsg() will send ACK otherwise).
5237 * If application uses SO_RCVLOWAT, we want send ack now if
5238 * we have not received enough bytes to satisfy the condition.
5240 (tp->rcv_nxt - tp->copied_seq < sk->sk_rcvlowat ||
5241 __tcp_select_window(sk) >= tp->rcv_wnd)) ||
5242 /* We ACK each frame or... */
5243 tcp_in_quickack_mode(sk) ||
5244 /* Protocol state mandates a one-time immediate ACK */
5245 inet_csk(sk)->icsk_ack.pending & ICSK_ACK_NOW) {
5251 if (!ofo_possible || RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
5252 tcp_send_delayed_ack(sk);
5256 if (!tcp_is_sack(tp) ||
5257 tp->compressed_ack >= sock_net(sk)->ipv4.sysctl_tcp_comp_sack_nr)
5260 if (tp->compressed_ack_rcv_nxt != tp->rcv_nxt) {
5261 tp->compressed_ack_rcv_nxt = tp->rcv_nxt;
5262 if (tp->compressed_ack > TCP_FASTRETRANS_THRESH)
5263 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED,
5264 tp->compressed_ack - TCP_FASTRETRANS_THRESH);
5265 tp->compressed_ack = 0;
5268 if (++tp->compressed_ack <= TCP_FASTRETRANS_THRESH)
5271 if (hrtimer_is_queued(&tp->compressed_ack_timer))
5274 /* compress ack timer : 5 % of rtt, but no more than tcp_comp_sack_delay_ns */
5276 rtt = tp->rcv_rtt_est.rtt_us;
5277 if (tp->srtt_us && tp->srtt_us < rtt)
5280 delay = min_t(unsigned long, sock_net(sk)->ipv4.sysctl_tcp_comp_sack_delay_ns,
5281 rtt * (NSEC_PER_USEC >> 3)/20);
5283 hrtimer_start(&tp->compressed_ack_timer, ns_to_ktime(delay),
5284 HRTIMER_MODE_REL_PINNED_SOFT);
5287 static inline void tcp_ack_snd_check(struct sock *sk)
5289 if (!inet_csk_ack_scheduled(sk)) {
5290 /* We sent a data segment already. */
5293 __tcp_ack_snd_check(sk, 1);
5297 * This routine is only called when we have urgent data
5298 * signaled. Its the 'slow' part of tcp_urg. It could be
5299 * moved inline now as tcp_urg is only called from one
5300 * place. We handle URGent data wrong. We have to - as
5301 * BSD still doesn't use the correction from RFC961.
5302 * For 1003.1g we should support a new option TCP_STDURG to permit
5303 * either form (or just set the sysctl tcp_stdurg).
5306 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
5308 struct tcp_sock *tp = tcp_sk(sk);
5309 u32 ptr = ntohs(th->urg_ptr);
5311 if (ptr && !sock_net(sk)->ipv4.sysctl_tcp_stdurg)
5313 ptr += ntohl(th->seq);
5315 /* Ignore urgent data that we've already seen and read. */
5316 if (after(tp->copied_seq, ptr))
5319 /* Do not replay urg ptr.
5321 * NOTE: interesting situation not covered by specs.
5322 * Misbehaving sender may send urg ptr, pointing to segment,
5323 * which we already have in ofo queue. We are not able to fetch
5324 * such data and will stay in TCP_URG_NOTYET until will be eaten
5325 * by recvmsg(). Seems, we are not obliged to handle such wicked
5326 * situations. But it is worth to think about possibility of some
5327 * DoSes using some hypothetical application level deadlock.
5329 if (before(ptr, tp->rcv_nxt))
5332 /* Do we already have a newer (or duplicate) urgent pointer? */
5333 if (tp->urg_data && !after(ptr, tp->urg_seq))
5336 /* Tell the world about our new urgent pointer. */
5339 /* We may be adding urgent data when the last byte read was
5340 * urgent. To do this requires some care. We cannot just ignore
5341 * tp->copied_seq since we would read the last urgent byte again
5342 * as data, nor can we alter copied_seq until this data arrives
5343 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5345 * NOTE. Double Dutch. Rendering to plain English: author of comment
5346 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5347 * and expect that both A and B disappear from stream. This is _wrong_.
5348 * Though this happens in BSD with high probability, this is occasional.
5349 * Any application relying on this is buggy. Note also, that fix "works"
5350 * only in this artificial test. Insert some normal data between A and B and we will
5351 * decline of BSD again. Verdict: it is better to remove to trap
5354 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5355 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5356 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5358 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5359 __skb_unlink(skb, &sk->sk_receive_queue);
5364 tp->urg_data = TCP_URG_NOTYET;
5365 WRITE_ONCE(tp->urg_seq, ptr);
5367 /* Disable header prediction. */
5371 /* This is the 'fast' part of urgent handling. */
5372 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
5374 struct tcp_sock *tp = tcp_sk(sk);
5376 /* Check if we get a new urgent pointer - normally not. */
5378 tcp_check_urg(sk, th);
5380 /* Do we wait for any urgent data? - normally not... */
5381 if (tp->urg_data == TCP_URG_NOTYET) {
5382 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5385 /* Is the urgent pointer pointing into this packet? */
5386 if (ptr < skb->len) {
5388 if (skb_copy_bits(skb, ptr, &tmp, 1))
5390 tp->urg_data = TCP_URG_VALID | tmp;
5391 if (!sock_flag(sk, SOCK_DEAD))
5392 sk->sk_data_ready(sk);
5397 /* Accept RST for rcv_nxt - 1 after a FIN.
5398 * When tcp connections are abruptly terminated from Mac OSX (via ^C), a
5399 * FIN is sent followed by a RST packet. The RST is sent with the same
5400 * sequence number as the FIN, and thus according to RFC 5961 a challenge
5401 * ACK should be sent. However, Mac OSX rate limits replies to challenge
5402 * ACKs on the closed socket. In addition middleboxes can drop either the
5403 * challenge ACK or a subsequent RST.
5405 static bool tcp_reset_check(const struct sock *sk, const struct sk_buff *skb)
5407 struct tcp_sock *tp = tcp_sk(sk);
5409 return unlikely(TCP_SKB_CB(skb)->seq == (tp->rcv_nxt - 1) &&
5410 (1 << sk->sk_state) & (TCPF_CLOSE_WAIT | TCPF_LAST_ACK |
5414 /* Does PAWS and seqno based validation of an incoming segment, flags will
5415 * play significant role here.
5417 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5418 const struct tcphdr *th, int syn_inerr)
5420 struct tcp_sock *tp = tcp_sk(sk);
5421 bool rst_seq_match = false;
5423 /* RFC1323: H1. Apply PAWS check first. */
5424 if (tcp_fast_parse_options(sock_net(sk), skb, th, tp) &&
5425 tp->rx_opt.saw_tstamp &&
5426 tcp_paws_discard(sk, skb)) {
5428 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5429 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5430 LINUX_MIB_TCPACKSKIPPEDPAWS,
5431 &tp->last_oow_ack_time))
5432 tcp_send_dupack(sk, skb);
5435 /* Reset is accepted even if it did not pass PAWS. */
5438 /* Step 1: check sequence number */
5439 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5440 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5441 * (RST) segments are validated by checking their SEQ-fields."
5442 * And page 69: "If an incoming segment is not acceptable,
5443 * an acknowledgment should be sent in reply (unless the RST
5444 * bit is set, if so drop the segment and return)".
5449 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5450 LINUX_MIB_TCPACKSKIPPEDSEQ,
5451 &tp->last_oow_ack_time))
5452 tcp_send_dupack(sk, skb);
5453 } else if (tcp_reset_check(sk, skb)) {
5459 /* Step 2: check RST bit */
5461 /* RFC 5961 3.2 (extend to match against (RCV.NXT - 1) after a
5462 * FIN and SACK too if available):
5463 * If seq num matches RCV.NXT or (RCV.NXT - 1) after a FIN, or
5464 * the right-most SACK block,
5466 * RESET the connection
5468 * Send a challenge ACK
5470 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt ||
5471 tcp_reset_check(sk, skb)) {
5472 rst_seq_match = true;
5473 } else if (tcp_is_sack(tp) && tp->rx_opt.num_sacks > 0) {
5474 struct tcp_sack_block *sp = &tp->selective_acks[0];
5475 int max_sack = sp[0].end_seq;
5478 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;
5480 max_sack = after(sp[this_sack].end_seq,
5482 sp[this_sack].end_seq : max_sack;
5485 if (TCP_SKB_CB(skb)->seq == max_sack)
5486 rst_seq_match = true;
5492 /* Disable TFO if RST is out-of-order
5493 * and no data has been received
5494 * for current active TFO socket
5496 if (tp->syn_fastopen && !tp->data_segs_in &&
5497 sk->sk_state == TCP_ESTABLISHED)
5498 tcp_fastopen_active_disable(sk);
5499 tcp_send_challenge_ack(sk, skb);
5504 /* step 3: check security and precedence [ignored] */
5506 /* step 4: Check for a SYN
5507 * RFC 5961 4.2 : Send a challenge ack
5512 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5513 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5514 tcp_send_challenge_ack(sk, skb);
5526 * TCP receive function for the ESTABLISHED state.
5528 * It is split into a fast path and a slow path. The fast path is
5530 * - A zero window was announced from us - zero window probing
5531 * is only handled properly in the slow path.
5532 * - Out of order segments arrived.
5533 * - Urgent data is expected.
5534 * - There is no buffer space left
5535 * - Unexpected TCP flags/window values/header lengths are received
5536 * (detected by checking the TCP header against pred_flags)
5537 * - Data is sent in both directions. Fast path only supports pure senders
5538 * or pure receivers (this means either the sequence number or the ack
5539 * value must stay constant)
5540 * - Unexpected TCP option.
5542 * When these conditions are not satisfied it drops into a standard
5543 * receive procedure patterned after RFC793 to handle all cases.
5544 * The first three cases are guaranteed by proper pred_flags setting,
5545 * the rest is checked inline. Fast processing is turned on in
5546 * tcp_data_queue when everything is OK.
5548 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb)
5550 const struct tcphdr *th = (const struct tcphdr *)skb->data;
5551 struct tcp_sock *tp = tcp_sk(sk);
5552 unsigned int len = skb->len;
5554 /* TCP congestion window tracking */
5555 trace_tcp_probe(sk, skb);
5557 tcp_mstamp_refresh(tp);
5558 if (unlikely(!sk->sk_rx_dst))
5559 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
5561 * Header prediction.
5562 * The code loosely follows the one in the famous
5563 * "30 instruction TCP receive" Van Jacobson mail.
5565 * Van's trick is to deposit buffers into socket queue
5566 * on a device interrupt, to call tcp_recv function
5567 * on the receive process context and checksum and copy
5568 * the buffer to user space. smart...
5570 * Our current scheme is not silly either but we take the
5571 * extra cost of the net_bh soft interrupt processing...
5572 * We do checksum and copy also but from device to kernel.
5575 tp->rx_opt.saw_tstamp = 0;
5577 /* pred_flags is 0xS?10 << 16 + snd_wnd
5578 * if header_prediction is to be made
5579 * 'S' will always be tp->tcp_header_len >> 2
5580 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5581 * turn it off (when there are holes in the receive
5582 * space for instance)
5583 * PSH flag is ignored.
5586 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5587 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5588 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5589 int tcp_header_len = tp->tcp_header_len;
5591 /* Timestamp header prediction: tcp_header_len
5592 * is automatically equal to th->doff*4 due to pred_flags
5596 /* Check timestamp */
5597 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5598 /* No? Slow path! */
5599 if (!tcp_parse_aligned_timestamp(tp, th))
5602 /* If PAWS failed, check it more carefully in slow path */
5603 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5606 /* DO NOT update ts_recent here, if checksum fails
5607 * and timestamp was corrupted part, it will result
5608 * in a hung connection since we will drop all
5609 * future packets due to the PAWS test.
5613 if (len <= tcp_header_len) {
5614 /* Bulk data transfer: sender */
5615 if (len == tcp_header_len) {
5616 /* Predicted packet is in window by definition.
5617 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5618 * Hence, check seq<=rcv_wup reduces to:
5620 if (tcp_header_len ==
5621 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5622 tp->rcv_nxt == tp->rcv_wup)
5623 tcp_store_ts_recent(tp);
5625 /* We know that such packets are checksummed
5628 tcp_ack(sk, skb, 0);
5630 tcp_data_snd_check(sk);
5631 /* When receiving pure ack in fast path, update
5632 * last ts ecr directly instead of calling
5633 * tcp_rcv_rtt_measure_ts()
5635 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr;
5637 } else { /* Header too small */
5638 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5643 bool fragstolen = false;
5645 if (tcp_checksum_complete(skb))
5648 if ((int)skb->truesize > sk->sk_forward_alloc)
5651 /* Predicted packet is in window by definition.
5652 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5653 * Hence, check seq<=rcv_wup reduces to:
5655 if (tcp_header_len ==
5656 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5657 tp->rcv_nxt == tp->rcv_wup)
5658 tcp_store_ts_recent(tp);
5660 tcp_rcv_rtt_measure_ts(sk, skb);
5662 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPHITS);
5664 /* Bulk data transfer: receiver */
5665 __skb_pull(skb, tcp_header_len);
5666 eaten = tcp_queue_rcv(sk, skb, &fragstolen);
5668 tcp_event_data_recv(sk, skb);
5670 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5671 /* Well, only one small jumplet in fast path... */
5672 tcp_ack(sk, skb, FLAG_DATA);
5673 tcp_data_snd_check(sk);
5674 if (!inet_csk_ack_scheduled(sk))
5678 __tcp_ack_snd_check(sk, 0);
5681 kfree_skb_partial(skb, fragstolen);
5688 if (len < (th->doff << 2) || tcp_checksum_complete(skb))
5691 if (!th->ack && !th->rst && !th->syn)
5695 * Standard slow path.
5698 if (!tcp_validate_incoming(sk, skb, th, 1))
5702 if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0)
5705 tcp_rcv_rtt_measure_ts(sk, skb);
5707 /* Process urgent data. */
5708 tcp_urg(sk, skb, th);
5710 /* step 7: process the segment text */
5711 tcp_data_queue(sk, skb);
5713 tcp_data_snd_check(sk);
5714 tcp_ack_snd_check(sk);
5718 TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS);
5719 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5724 EXPORT_SYMBOL(tcp_rcv_established);
5726 void tcp_init_transfer(struct sock *sk, int bpf_op)
5728 struct inet_connection_sock *icsk = inet_csk(sk);
5729 struct tcp_sock *tp = tcp_sk(sk);
5732 icsk->icsk_af_ops->rebuild_header(sk);
5733 tcp_init_metrics(sk);
5735 /* Initialize the congestion window to start the transfer.
5736 * Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
5737 * retransmitted. In light of RFC6298 more aggressive 1sec
5738 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
5739 * retransmission has occurred.
5741 if (tp->total_retrans > 1 && tp->undo_marker)
5744 tp->snd_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk));
5745 tp->snd_cwnd_stamp = tcp_jiffies32;
5747 tcp_call_bpf(sk, bpf_op, 0, NULL);
5748 tcp_init_congestion_control(sk);
5749 tcp_init_buffer_space(sk);
5752 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5754 struct tcp_sock *tp = tcp_sk(sk);
5755 struct inet_connection_sock *icsk = inet_csk(sk);
5757 tcp_set_state(sk, TCP_ESTABLISHED);
5758 icsk->icsk_ack.lrcvtime = tcp_jiffies32;
5761 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
5762 security_inet_conn_established(sk, skb);
5763 sk_mark_napi_id(sk, skb);
5766 tcp_init_transfer(sk, BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB);
5768 /* Prevent spurious tcp_cwnd_restart() on first data
5771 tp->lsndtime = tcp_jiffies32;
5773 if (sock_flag(sk, SOCK_KEEPOPEN))
5774 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5776 if (!tp->rx_opt.snd_wscale)
5777 __tcp_fast_path_on(tp, tp->snd_wnd);
5782 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
5783 struct tcp_fastopen_cookie *cookie)
5785 struct tcp_sock *tp = tcp_sk(sk);
5786 struct sk_buff *data = tp->syn_data ? tcp_rtx_queue_head(sk) : NULL;
5787 u16 mss = tp->rx_opt.mss_clamp, try_exp = 0;
5788 bool syn_drop = false;
5790 if (mss == tp->rx_opt.user_mss) {
5791 struct tcp_options_received opt;
5793 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5794 tcp_clear_options(&opt);
5795 opt.user_mss = opt.mss_clamp = 0;
5796 tcp_parse_options(sock_net(sk), synack, &opt, 0, NULL);
5797 mss = opt.mss_clamp;
5800 if (!tp->syn_fastopen) {
5801 /* Ignore an unsolicited cookie */
5803 } else if (tp->total_retrans) {
5804 /* SYN timed out and the SYN-ACK neither has a cookie nor
5805 * acknowledges data. Presumably the remote received only
5806 * the retransmitted (regular) SYNs: either the original
5807 * SYN-data or the corresponding SYN-ACK was dropped.
5809 syn_drop = (cookie->len < 0 && data);
5810 } else if (cookie->len < 0 && !tp->syn_data) {
5811 /* We requested a cookie but didn't get it. If we did not use
5812 * the (old) exp opt format then try so next time (try_exp=1).
5813 * Otherwise we go back to use the RFC7413 opt (try_exp=2).
5815 try_exp = tp->syn_fastopen_exp ? 2 : 1;
5818 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp);
5820 if (data) { /* Retransmit unacked data in SYN */
5821 if (tp->total_retrans)
5822 tp->fastopen_client_fail = TFO_SYN_RETRANSMITTED;
5824 tp->fastopen_client_fail = TFO_DATA_NOT_ACKED;
5825 skb_rbtree_walk_from(data) {
5826 if (__tcp_retransmit_skb(sk, data, 1))
5830 NET_INC_STATS(sock_net(sk),
5831 LINUX_MIB_TCPFASTOPENACTIVEFAIL);
5834 tp->syn_data_acked = tp->syn_data;
5835 if (tp->syn_data_acked) {
5836 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
5837 /* SYN-data is counted as two separate packets in tcp_ack() */
5838 if (tp->delivered > 1)
5842 tcp_fastopen_add_skb(sk, synack);
5847 static void smc_check_reset_syn(struct tcp_sock *tp)
5849 #if IS_ENABLED(CONFIG_SMC)
5850 if (static_branch_unlikely(&tcp_have_smc)) {
5851 if (tp->syn_smc && !tp->rx_opt.smc_ok)
5857 static void tcp_try_undo_spurious_syn(struct sock *sk)
5859 struct tcp_sock *tp = tcp_sk(sk);
5862 /* undo_marker is set when SYN or SYNACK times out. The timeout is
5863 * spurious if the ACK's timestamp option echo value matches the
5864 * original SYN timestamp.
5866 syn_stamp = tp->retrans_stamp;
5867 if (tp->undo_marker && syn_stamp && tp->rx_opt.saw_tstamp &&
5868 syn_stamp == tp->rx_opt.rcv_tsecr)
5869 tp->undo_marker = 0;
5872 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5873 const struct tcphdr *th)
5875 struct inet_connection_sock *icsk = inet_csk(sk);
5876 struct tcp_sock *tp = tcp_sk(sk);
5877 struct tcp_fastopen_cookie foc = { .len = -1 };
5878 int saved_clamp = tp->rx_opt.mss_clamp;
5881 tcp_parse_options(sock_net(sk), skb, &tp->rx_opt, 0, &foc);
5882 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
5883 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
5887 * "If the state is SYN-SENT then
5888 * first check the ACK bit
5889 * If the ACK bit is set
5890 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5891 * a reset (unless the RST bit is set, if so drop
5892 * the segment and return)"
5894 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
5895 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt))
5896 goto reset_and_undo;
5898 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5899 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5900 tcp_time_stamp(tp))) {
5901 NET_INC_STATS(sock_net(sk),
5902 LINUX_MIB_PAWSACTIVEREJECTED);
5903 goto reset_and_undo;
5906 /* Now ACK is acceptable.
5908 * "If the RST bit is set
5909 * If the ACK was acceptable then signal the user "error:
5910 * connection reset", drop the segment, enter CLOSED state,
5911 * delete TCB, and return."
5920 * "fifth, if neither of the SYN or RST bits is set then
5921 * drop the segment and return."
5927 goto discard_and_undo;
5930 * "If the SYN bit is on ...
5931 * are acceptable then ...
5932 * (our SYN has been ACKed), change the connection
5933 * state to ESTABLISHED..."
5936 tcp_ecn_rcv_synack(tp, th);
5938 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5939 tcp_try_undo_spurious_syn(sk);
5940 tcp_ack(sk, skb, FLAG_SLOWPATH);
5942 /* Ok.. it's good. Set up sequence numbers and
5943 * move to established.
5945 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1);
5946 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5948 /* RFC1323: The window in SYN & SYN/ACK segments is
5951 tp->snd_wnd = ntohs(th->window);
5953 if (!tp->rx_opt.wscale_ok) {
5954 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5955 tp->window_clamp = min(tp->window_clamp, 65535U);
5958 if (tp->rx_opt.saw_tstamp) {
5959 tp->rx_opt.tstamp_ok = 1;
5960 tp->tcp_header_len =
5961 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5962 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5963 tcp_store_ts_recent(tp);
5965 tp->tcp_header_len = sizeof(struct tcphdr);
5968 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5969 tcp_initialize_rcv_mss(sk);
5971 /* Remember, tcp_poll() does not lock socket!
5972 * Change state from SYN-SENT only after copied_seq
5973 * is initialized. */
5974 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
5976 smc_check_reset_syn(tp);
5980 tcp_finish_connect(sk, skb);
5982 fastopen_fail = (tp->syn_fastopen || tp->syn_data) &&
5983 tcp_rcv_fastopen_synack(sk, skb, &foc);
5985 if (!sock_flag(sk, SOCK_DEAD)) {
5986 sk->sk_state_change(sk);
5987 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5991 if (sk->sk_write_pending ||
5992 icsk->icsk_accept_queue.rskq_defer_accept ||
5993 inet_csk_in_pingpong_mode(sk)) {
5994 /* Save one ACK. Data will be ready after
5995 * several ticks, if write_pending is set.
5997 * It may be deleted, but with this feature tcpdumps
5998 * look so _wonderfully_ clever, that I was not able
5999 * to stand against the temptation 8) --ANK
6001 inet_csk_schedule_ack(sk);
6002 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
6003 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
6004 TCP_DELACK_MAX, TCP_RTO_MAX);
6015 /* No ACK in the segment */
6019 * "If the RST bit is set
6021 * Otherwise (no ACK) drop the segment and return."
6024 goto discard_and_undo;
6028 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
6029 tcp_paws_reject(&tp->rx_opt, 0))
6030 goto discard_and_undo;
6033 /* We see SYN without ACK. It is attempt of
6034 * simultaneous connect with crossed SYNs.
6035 * Particularly, it can be connect to self.
6037 tcp_set_state(sk, TCP_SYN_RECV);
6039 if (tp->rx_opt.saw_tstamp) {
6040 tp->rx_opt.tstamp_ok = 1;
6041 tcp_store_ts_recent(tp);
6042 tp->tcp_header_len =
6043 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
6045 tp->tcp_header_len = sizeof(struct tcphdr);
6048 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1);
6049 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6050 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
6052 /* RFC1323: The window in SYN & SYN/ACK segments is
6055 tp->snd_wnd = ntohs(th->window);
6056 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
6057 tp->max_window = tp->snd_wnd;
6059 tcp_ecn_rcv_syn(tp, th);
6062 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
6063 tcp_initialize_rcv_mss(sk);
6065 tcp_send_synack(sk);
6067 /* Note, we could accept data and URG from this segment.
6068 * There are no obstacles to make this (except that we must
6069 * either change tcp_recvmsg() to prevent it from returning data
6070 * before 3WHS completes per RFC793, or employ TCP Fast Open).
6072 * However, if we ignore data in ACKless segments sometimes,
6073 * we have no reasons to accept it sometimes.
6074 * Also, seems the code doing it in step6 of tcp_rcv_state_process
6075 * is not flawless. So, discard packet for sanity.
6076 * Uncomment this return to process the data.
6083 /* "fifth, if neither of the SYN or RST bits is set then
6084 * drop the segment and return."
6088 tcp_clear_options(&tp->rx_opt);
6089 tp->rx_opt.mss_clamp = saved_clamp;
6093 tcp_clear_options(&tp->rx_opt);
6094 tp->rx_opt.mss_clamp = saved_clamp;
6098 static void tcp_rcv_synrecv_state_fastopen(struct sock *sk)
6100 struct request_sock *req;
6102 tcp_try_undo_loss(sk, false);
6104 /* Reset rtx states to prevent spurious retransmits_timed_out() */
6105 tcp_sk(sk)->retrans_stamp = 0;
6106 inet_csk(sk)->icsk_retransmits = 0;
6108 /* Once we leave TCP_SYN_RECV or TCP_FIN_WAIT_1,
6109 * we no longer need req so release it.
6111 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk,
6112 lockdep_sock_is_held(sk));
6113 reqsk_fastopen_remove(sk, req, false);
6115 /* Re-arm the timer because data may have been sent out.
6116 * This is similar to the regular data transmission case
6117 * when new data has just been ack'ed.
6119 * (TFO) - we could try to be more aggressive and
6120 * retransmitting any data sooner based on when they
6127 * This function implements the receiving procedure of RFC 793 for
6128 * all states except ESTABLISHED and TIME_WAIT.
6129 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
6130 * address independent.
6133 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb)
6135 struct tcp_sock *tp = tcp_sk(sk);
6136 struct inet_connection_sock *icsk = inet_csk(sk);
6137 const struct tcphdr *th = tcp_hdr(skb);
6138 struct request_sock *req;
6142 switch (sk->sk_state) {
6156 /* It is possible that we process SYN packets from backlog,
6157 * so we need to make sure to disable BH and RCU right there.
6161 acceptable = icsk->icsk_af_ops->conn_request(sk, skb) >= 0;
6173 tp->rx_opt.saw_tstamp = 0;
6174 tcp_mstamp_refresh(tp);
6175 queued = tcp_rcv_synsent_state_process(sk, skb, th);
6179 /* Do step6 onward by hand. */
6180 tcp_urg(sk, skb, th);
6182 tcp_data_snd_check(sk);
6186 tcp_mstamp_refresh(tp);
6187 tp->rx_opt.saw_tstamp = 0;
6188 req = rcu_dereference_protected(tp->fastopen_rsk,
6189 lockdep_sock_is_held(sk));
6193 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
6194 sk->sk_state != TCP_FIN_WAIT1);
6196 if (!tcp_check_req(sk, skb, req, true, &req_stolen))
6200 if (!th->ack && !th->rst && !th->syn)
6203 if (!tcp_validate_incoming(sk, skb, th, 0))
6206 /* step 5: check the ACK field */
6207 acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
6208 FLAG_UPDATE_TS_RECENT |
6209 FLAG_NO_CHALLENGE_ACK) > 0;
6212 if (sk->sk_state == TCP_SYN_RECV)
6213 return 1; /* send one RST */
6214 tcp_send_challenge_ack(sk, skb);
6217 switch (sk->sk_state) {
6219 tp->delivered++; /* SYN-ACK delivery isn't tracked in tcp_ack */
6221 tcp_synack_rtt_meas(sk, req);
6224 tcp_rcv_synrecv_state_fastopen(sk);
6226 tcp_try_undo_spurious_syn(sk);
6227 tp->retrans_stamp = 0;
6228 tcp_init_transfer(sk, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB);
6229 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6232 tcp_set_state(sk, TCP_ESTABLISHED);
6233 sk->sk_state_change(sk);
6235 /* Note, that this wakeup is only for marginal crossed SYN case.
6236 * Passively open sockets are not waked up, because
6237 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
6240 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
6242 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
6243 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
6244 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
6246 if (tp->rx_opt.tstamp_ok)
6247 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
6249 if (!inet_csk(sk)->icsk_ca_ops->cong_control)
6250 tcp_update_pacing_rate(sk);
6252 /* Prevent spurious tcp_cwnd_restart() on first data packet */
6253 tp->lsndtime = tcp_jiffies32;
6255 tcp_initialize_rcv_mss(sk);
6256 tcp_fast_path_on(tp);
6259 case TCP_FIN_WAIT1: {
6263 tcp_rcv_synrecv_state_fastopen(sk);
6265 if (tp->snd_una != tp->write_seq)
6268 tcp_set_state(sk, TCP_FIN_WAIT2);
6269 sk->sk_shutdown |= SEND_SHUTDOWN;
6273 if (!sock_flag(sk, SOCK_DEAD)) {
6274 /* Wake up lingering close() */
6275 sk->sk_state_change(sk);
6279 if (tp->linger2 < 0) {
6281 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6284 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6285 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6286 /* Receive out of order FIN after close() */
6287 if (tp->syn_fastopen && th->fin)
6288 tcp_fastopen_active_disable(sk);
6290 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6294 tmo = tcp_fin_time(sk);
6295 if (tmo > TCP_TIMEWAIT_LEN) {
6296 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
6297 } else if (th->fin || sock_owned_by_user(sk)) {
6298 /* Bad case. We could lose such FIN otherwise.
6299 * It is not a big problem, but it looks confusing
6300 * and not so rare event. We still can lose it now,
6301 * if it spins in bh_lock_sock(), but it is really
6304 inet_csk_reset_keepalive_timer(sk, tmo);
6306 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
6313 if (tp->snd_una == tp->write_seq) {
6314 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
6320 if (tp->snd_una == tp->write_seq) {
6321 tcp_update_metrics(sk);
6328 /* step 6: check the URG bit */
6329 tcp_urg(sk, skb, th);
6331 /* step 7: process the segment text */
6332 switch (sk->sk_state) {
6333 case TCP_CLOSE_WAIT:
6336 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
6341 /* RFC 793 says to queue data in these states,
6342 * RFC 1122 says we MUST send a reset.
6343 * BSD 4.4 also does reset.
6345 if (sk->sk_shutdown & RCV_SHUTDOWN) {
6346 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6347 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6348 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6354 case TCP_ESTABLISHED:
6355 tcp_data_queue(sk, skb);
6360 /* tcp_data could move socket to TIME-WAIT */
6361 if (sk->sk_state != TCP_CLOSE) {
6362 tcp_data_snd_check(sk);
6363 tcp_ack_snd_check(sk);
6372 EXPORT_SYMBOL(tcp_rcv_state_process);
6374 static inline void pr_drop_req(struct request_sock *req, __u16 port, int family)
6376 struct inet_request_sock *ireq = inet_rsk(req);
6378 if (family == AF_INET)
6379 net_dbg_ratelimited("drop open request from %pI4/%u\n",
6380 &ireq->ir_rmt_addr, port);
6381 #if IS_ENABLED(CONFIG_IPV6)
6382 else if (family == AF_INET6)
6383 net_dbg_ratelimited("drop open request from %pI6/%u\n",
6384 &ireq->ir_v6_rmt_addr, port);
6388 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
6390 * If we receive a SYN packet with these bits set, it means a
6391 * network is playing bad games with TOS bits. In order to
6392 * avoid possible false congestion notifications, we disable
6393 * TCP ECN negotiation.
6395 * Exception: tcp_ca wants ECN. This is required for DCTCP
6396 * congestion control: Linux DCTCP asserts ECT on all packets,
6397 * including SYN, which is most optimal solution; however,
6398 * others, such as FreeBSD do not.
6400 * Exception: At least one of the reserved bits of the TCP header (th->res1) is
6401 * set, indicating the use of a future TCP extension (such as AccECN). See
6402 * RFC8311 §4.3 which updates RFC3168 to allow the development of such
6405 static void tcp_ecn_create_request(struct request_sock *req,
6406 const struct sk_buff *skb,
6407 const struct sock *listen_sk,
6408 const struct dst_entry *dst)
6410 const struct tcphdr *th = tcp_hdr(skb);
6411 const struct net *net = sock_net(listen_sk);
6412 bool th_ecn = th->ece && th->cwr;
6419 ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield);
6420 ecn_ok_dst = dst_feature(dst, DST_FEATURE_ECN_MASK);
6421 ecn_ok = net->ipv4.sysctl_tcp_ecn || ecn_ok_dst;
6423 if (((!ect || th->res1) && ecn_ok) || tcp_ca_needs_ecn(listen_sk) ||
6424 (ecn_ok_dst & DST_FEATURE_ECN_CA) ||
6425 tcp_bpf_ca_needs_ecn((struct sock *)req))
6426 inet_rsk(req)->ecn_ok = 1;
6429 static void tcp_openreq_init(struct request_sock *req,
6430 const struct tcp_options_received *rx_opt,
6431 struct sk_buff *skb, const struct sock *sk)
6433 struct inet_request_sock *ireq = inet_rsk(req);
6435 req->rsk_rcv_wnd = 0; /* So that tcp_send_synack() knows! */
6437 tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq;
6438 tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
6439 tcp_rsk(req)->snt_synack = 0;
6440 tcp_rsk(req)->last_oow_ack_time = 0;
6441 req->mss = rx_opt->mss_clamp;
6442 req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0;
6443 ireq->tstamp_ok = rx_opt->tstamp_ok;
6444 ireq->sack_ok = rx_opt->sack_ok;
6445 ireq->snd_wscale = rx_opt->snd_wscale;
6446 ireq->wscale_ok = rx_opt->wscale_ok;
6449 ireq->ir_rmt_port = tcp_hdr(skb)->source;
6450 ireq->ir_num = ntohs(tcp_hdr(skb)->dest);
6451 ireq->ir_mark = inet_request_mark(sk, skb);
6452 #if IS_ENABLED(CONFIG_SMC)
6453 ireq->smc_ok = rx_opt->smc_ok;
6457 struct request_sock *inet_reqsk_alloc(const struct request_sock_ops *ops,
6458 struct sock *sk_listener,
6459 bool attach_listener)
6461 struct request_sock *req = reqsk_alloc(ops, sk_listener,
6465 struct inet_request_sock *ireq = inet_rsk(req);
6467 ireq->ireq_opt = NULL;
6468 #if IS_ENABLED(CONFIG_IPV6)
6469 ireq->pktopts = NULL;
6471 atomic64_set(&ireq->ir_cookie, 0);
6472 ireq->ireq_state = TCP_NEW_SYN_RECV;
6473 write_pnet(&ireq->ireq_net, sock_net(sk_listener));
6474 ireq->ireq_family = sk_listener->sk_family;
6479 EXPORT_SYMBOL(inet_reqsk_alloc);
6482 * Return true if a syncookie should be sent
6484 static bool tcp_syn_flood_action(const struct sock *sk, const char *proto)
6486 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
6487 const char *msg = "Dropping request";
6488 bool want_cookie = false;
6489 struct net *net = sock_net(sk);
6491 #ifdef CONFIG_SYN_COOKIES
6492 if (net->ipv4.sysctl_tcp_syncookies) {
6493 msg = "Sending cookies";
6495 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES);
6498 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP);
6500 if (!queue->synflood_warned &&
6501 net->ipv4.sysctl_tcp_syncookies != 2 &&
6502 xchg(&queue->synflood_warned, 1) == 0)
6503 net_info_ratelimited("%s: Possible SYN flooding on port %d. %s. Check SNMP counters.\n",
6504 proto, sk->sk_num, msg);
6509 static void tcp_reqsk_record_syn(const struct sock *sk,
6510 struct request_sock *req,
6511 const struct sk_buff *skb)
6513 if (tcp_sk(sk)->save_syn) {
6514 u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb);
6517 copy = kmalloc(len + sizeof(u32), GFP_ATOMIC);
6520 memcpy(©[1], skb_network_header(skb), len);
6521 req->saved_syn = copy;
6526 /* If a SYN cookie is required and supported, returns a clamped MSS value to be
6527 * used for SYN cookie generation.
6529 u16 tcp_get_syncookie_mss(struct request_sock_ops *rsk_ops,
6530 const struct tcp_request_sock_ops *af_ops,
6531 struct sock *sk, struct tcphdr *th)
6533 struct tcp_sock *tp = tcp_sk(sk);
6536 if (sock_net(sk)->ipv4.sysctl_tcp_syncookies != 2 &&
6537 !inet_csk_reqsk_queue_is_full(sk))
6540 if (!tcp_syn_flood_action(sk, rsk_ops->slab_name))
6543 if (sk_acceptq_is_full(sk)) {
6544 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
6548 mss = tcp_parse_mss_option(th, tp->rx_opt.user_mss);
6550 mss = af_ops->mss_clamp;
6554 EXPORT_SYMBOL_GPL(tcp_get_syncookie_mss);
6556 int tcp_conn_request(struct request_sock_ops *rsk_ops,
6557 const struct tcp_request_sock_ops *af_ops,
6558 struct sock *sk, struct sk_buff *skb)
6560 struct tcp_fastopen_cookie foc = { .len = -1 };
6561 __u32 isn = TCP_SKB_CB(skb)->tcp_tw_isn;
6562 struct tcp_options_received tmp_opt;
6563 struct tcp_sock *tp = tcp_sk(sk);
6564 struct net *net = sock_net(sk);
6565 struct sock *fastopen_sk = NULL;
6566 struct request_sock *req;
6567 bool want_cookie = false;
6568 struct dst_entry *dst;
6571 /* TW buckets are converted to open requests without
6572 * limitations, they conserve resources and peer is
6573 * evidently real one.
6575 if ((net->ipv4.sysctl_tcp_syncookies == 2 ||
6576 inet_csk_reqsk_queue_is_full(sk)) && !isn) {
6577 want_cookie = tcp_syn_flood_action(sk, rsk_ops->slab_name);
6582 if (sk_acceptq_is_full(sk)) {
6583 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
6587 req = inet_reqsk_alloc(rsk_ops, sk, !want_cookie);
6591 tcp_rsk(req)->af_specific = af_ops;
6592 tcp_rsk(req)->ts_off = 0;
6594 tcp_clear_options(&tmp_opt);
6595 tmp_opt.mss_clamp = af_ops->mss_clamp;
6596 tmp_opt.user_mss = tp->rx_opt.user_mss;
6597 tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0,
6598 want_cookie ? NULL : &foc);
6600 if (want_cookie && !tmp_opt.saw_tstamp)
6601 tcp_clear_options(&tmp_opt);
6603 if (IS_ENABLED(CONFIG_SMC) && want_cookie)
6606 tmp_opt.tstamp_ok = tmp_opt.saw_tstamp;
6607 tcp_openreq_init(req, &tmp_opt, skb, sk);
6608 inet_rsk(req)->no_srccheck = inet_sk(sk)->transparent;
6610 /* Note: tcp_v6_init_req() might override ir_iif for link locals */
6611 inet_rsk(req)->ir_iif = inet_request_bound_dev_if(sk, skb);
6613 af_ops->init_req(req, sk, skb);
6615 if (security_inet_conn_request(sk, skb, req))
6618 if (tmp_opt.tstamp_ok)
6619 tcp_rsk(req)->ts_off = af_ops->init_ts_off(net, skb);
6621 dst = af_ops->route_req(sk, &fl, req);
6625 if (!want_cookie && !isn) {
6626 /* Kill the following clause, if you dislike this way. */
6627 if (!net->ipv4.sysctl_tcp_syncookies &&
6628 (net->ipv4.sysctl_max_syn_backlog - inet_csk_reqsk_queue_len(sk) <
6629 (net->ipv4.sysctl_max_syn_backlog >> 2)) &&
6630 !tcp_peer_is_proven(req, dst)) {
6631 /* Without syncookies last quarter of
6632 * backlog is filled with destinations,
6633 * proven to be alive.
6634 * It means that we continue to communicate
6635 * to destinations, already remembered
6636 * to the moment of synflood.
6638 pr_drop_req(req, ntohs(tcp_hdr(skb)->source),
6640 goto drop_and_release;
6643 isn = af_ops->init_seq(skb);
6646 tcp_ecn_create_request(req, skb, sk, dst);
6649 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss);
6650 req->cookie_ts = tmp_opt.tstamp_ok;
6651 if (!tmp_opt.tstamp_ok)
6652 inet_rsk(req)->ecn_ok = 0;
6655 tcp_rsk(req)->snt_isn = isn;
6656 tcp_rsk(req)->txhash = net_tx_rndhash();
6657 tcp_openreq_init_rwin(req, sk, dst);
6658 sk_rx_queue_set(req_to_sk(req), skb);
6660 tcp_reqsk_record_syn(sk, req, skb);
6661 fastopen_sk = tcp_try_fastopen(sk, skb, req, &foc, dst);
6664 af_ops->send_synack(fastopen_sk, dst, &fl, req,
6665 &foc, TCP_SYNACK_FASTOPEN);
6666 /* Add the child socket directly into the accept queue */
6667 if (!inet_csk_reqsk_queue_add(sk, req, fastopen_sk)) {
6668 reqsk_fastopen_remove(fastopen_sk, req, false);
6669 bh_unlock_sock(fastopen_sk);
6670 sock_put(fastopen_sk);
6673 sk->sk_data_ready(sk);
6674 bh_unlock_sock(fastopen_sk);
6675 sock_put(fastopen_sk);
6677 tcp_rsk(req)->tfo_listener = false;
6679 inet_csk_reqsk_queue_hash_add(sk, req,
6680 tcp_timeout_init((struct sock *)req));
6681 af_ops->send_synack(sk, dst, &fl, req, &foc,
6682 !want_cookie ? TCP_SYNACK_NORMAL :
6700 EXPORT_SYMBOL(tcp_conn_request);