1 // SPDX-License-Identifier: GPL-2.0
3 * INET An implementation of the TCP/IP protocol suite for the LINUX
4 * operating system. INET is implemented using the BSD Socket
5 * interface as the means of communication with the user level.
7 * Implementation of the Transmission Control Protocol(TCP).
10 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
11 * Mark Evans, <evansmp@uhura.aston.ac.uk>
12 * Corey Minyard <wf-rch!minyard@relay.EU.net>
13 * Florian La Roche, <flla@stud.uni-sb.de>
14 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
15 * Linus Torvalds, <torvalds@cs.helsinki.fi>
16 * Alan Cox, <gw4pts@gw4pts.ampr.org>
17 * Matthew Dillon, <dillon@apollo.west.oic.com>
18 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
19 * Jorge Cwik, <jorge@laser.satlink.net>
24 * Pedro Roque : Fast Retransmit/Recovery.
26 * Retransmit queue handled by TCP.
27 * Better retransmit timer handling.
28 * New congestion avoidance.
32 * Eric : Fast Retransmit.
33 * Randy Scott : MSS option defines.
34 * Eric Schenk : Fixes to slow start algorithm.
35 * Eric Schenk : Yet another double ACK bug.
36 * Eric Schenk : Delayed ACK bug fixes.
37 * Eric Schenk : Floyd style fast retrans war avoidance.
38 * David S. Miller : Don't allow zero congestion window.
39 * Eric Schenk : Fix retransmitter so that it sends
40 * next packet on ack of previous packet.
41 * Andi Kleen : Moved open_request checking here
42 * and process RSTs for open_requests.
43 * Andi Kleen : Better prune_queue, and other fixes.
44 * Andrey Savochkin: Fix RTT measurements in the presence of
46 * Andrey Savochkin: Check sequence numbers correctly when
47 * removing SACKs due to in sequence incoming
49 * Andi Kleen: Make sure we never ack data there is not
50 * enough room for. Also make this condition
51 * a fatal error if it might still happen.
52 * Andi Kleen: Add tcp_measure_rcv_mss to make
53 * connections with MSS<min(MTU,ann. MSS)
54 * work without delayed acks.
55 * Andi Kleen: Process packets with PSH set in the
57 * J Hadi Salim: ECN support
60 * Panu Kuhlberg: Experimental audit of TCP (re)transmission
61 * engine. Lots of bugs are found.
62 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs
65 #define pr_fmt(fmt) "TCP: " fmt
68 #include <linux/slab.h>
69 #include <linux/module.h>
70 #include <linux/sysctl.h>
71 #include <linux/kernel.h>
72 #include <linux/prefetch.h>
75 #include <net/inet_common.h>
76 #include <linux/ipsec.h>
77 #include <asm/unaligned.h>
78 #include <linux/errqueue.h>
79 #include <trace/events/tcp.h>
80 #include <linux/static_key.h>
81 #include <net/busy_poll.h>
83 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
85 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
86 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
87 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
88 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
89 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
90 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
91 #define FLAG_ECE 0x40 /* ECE in this ACK */
92 #define FLAG_LOST_RETRANS 0x80 /* This ACK marks some retransmission lost */
93 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
94 #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
95 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
96 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
97 #define FLAG_SET_XMIT_TIMER 0x1000 /* Set TLP or RTO timer */
98 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
99 #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
100 #define FLAG_NO_CHALLENGE_ACK 0x8000 /* do not call tcp_send_challenge_ack() */
101 #define FLAG_ACK_MAYBE_DELAYED 0x10000 /* Likely a delayed ACK */
103 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
104 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
105 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE|FLAG_DSACKING_ACK)
106 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
108 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
109 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
111 #define REXMIT_NONE 0 /* no loss recovery to do */
112 #define REXMIT_LOST 1 /* retransmit packets marked lost */
113 #define REXMIT_NEW 2 /* FRTO-style transmit of unsent/new packets */
115 #if IS_ENABLED(CONFIG_TLS_DEVICE)
116 static DEFINE_STATIC_KEY_FALSE(clean_acked_data_enabled);
118 void clean_acked_data_enable(struct inet_connection_sock *icsk,
119 void (*cad)(struct sock *sk, u32 ack_seq))
121 icsk->icsk_clean_acked = cad;
122 static_branch_inc(&clean_acked_data_enabled);
124 EXPORT_SYMBOL_GPL(clean_acked_data_enable);
126 void clean_acked_data_disable(struct inet_connection_sock *icsk)
128 static_branch_dec(&clean_acked_data_enabled);
129 icsk->icsk_clean_acked = NULL;
131 EXPORT_SYMBOL_GPL(clean_acked_data_disable);
134 static void tcp_gro_dev_warn(struct sock *sk, const struct sk_buff *skb,
137 static bool __once __read_mostly;
140 struct net_device *dev;
145 dev = dev_get_by_index_rcu(sock_net(sk), skb->skb_iif);
146 if (!dev || len >= dev->mtu)
147 pr_warn("%s: Driver has suspect GRO implementation, TCP performance may be compromised.\n",
148 dev ? dev->name : "Unknown driver");
153 /* Adapt the MSS value used to make delayed ack decision to the
156 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
158 struct inet_connection_sock *icsk = inet_csk(sk);
159 const unsigned int lss = icsk->icsk_ack.last_seg_size;
162 icsk->icsk_ack.last_seg_size = 0;
164 /* skb->len may jitter because of SACKs, even if peer
165 * sends good full-sized frames.
167 len = skb_shinfo(skb)->gso_size ? : skb->len;
168 if (len >= icsk->icsk_ack.rcv_mss) {
169 icsk->icsk_ack.rcv_mss = min_t(unsigned int, len,
171 /* Account for possibly-removed options */
172 if (unlikely(len > icsk->icsk_ack.rcv_mss +
173 MAX_TCP_OPTION_SPACE))
174 tcp_gro_dev_warn(sk, skb, len);
176 /* Otherwise, we make more careful check taking into account,
177 * that SACKs block is variable.
179 * "len" is invariant segment length, including TCP header.
181 len += skb->data - skb_transport_header(skb);
182 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
183 /* If PSH is not set, packet should be
184 * full sized, provided peer TCP is not badly broken.
185 * This observation (if it is correct 8)) allows
186 * to handle super-low mtu links fairly.
188 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
189 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
190 /* Subtract also invariant (if peer is RFC compliant),
191 * tcp header plus fixed timestamp option length.
192 * Resulting "len" is MSS free of SACK jitter.
194 len -= tcp_sk(sk)->tcp_header_len;
195 icsk->icsk_ack.last_seg_size = len;
197 icsk->icsk_ack.rcv_mss = len;
201 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
202 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
203 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
207 static void tcp_incr_quickack(struct sock *sk, unsigned int max_quickacks)
209 struct inet_connection_sock *icsk = inet_csk(sk);
210 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
214 quickacks = min(quickacks, max_quickacks);
215 if (quickacks > icsk->icsk_ack.quick)
216 icsk->icsk_ack.quick = quickacks;
219 void tcp_enter_quickack_mode(struct sock *sk, unsigned int max_quickacks)
221 struct inet_connection_sock *icsk = inet_csk(sk);
223 tcp_incr_quickack(sk, max_quickacks);
224 icsk->icsk_ack.pingpong = 0;
225 icsk->icsk_ack.ato = TCP_ATO_MIN;
227 EXPORT_SYMBOL(tcp_enter_quickack_mode);
229 /* Send ACKs quickly, if "quick" count is not exhausted
230 * and the session is not interactive.
233 static bool tcp_in_quickack_mode(struct sock *sk)
235 const struct inet_connection_sock *icsk = inet_csk(sk);
236 const struct dst_entry *dst = __sk_dst_get(sk);
238 return (dst && dst_metric(dst, RTAX_QUICKACK)) ||
239 (icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong);
242 static void tcp_ecn_queue_cwr(struct tcp_sock *tp)
244 if (tp->ecn_flags & TCP_ECN_OK)
245 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
248 static void tcp_ecn_accept_cwr(struct sock *sk, const struct sk_buff *skb)
250 if (tcp_hdr(skb)->cwr) {
251 tcp_sk(sk)->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
253 /* If the sender is telling us it has entered CWR, then its
254 * cwnd may be very low (even just 1 packet), so we should ACK
257 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW;
261 static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp)
263 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
266 static void __tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb)
268 struct tcp_sock *tp = tcp_sk(sk);
270 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
271 case INET_ECN_NOT_ECT:
272 /* Funny extension: if ECT is not set on a segment,
273 * and we already seen ECT on a previous segment,
274 * it is probably a retransmit.
276 if (tp->ecn_flags & TCP_ECN_SEEN)
277 tcp_enter_quickack_mode(sk, 2);
280 if (tcp_ca_needs_ecn(sk))
281 tcp_ca_event(sk, CA_EVENT_ECN_IS_CE);
283 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
284 /* Better not delay acks, sender can have a very low cwnd */
285 tcp_enter_quickack_mode(sk, 2);
286 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
288 tp->ecn_flags |= TCP_ECN_SEEN;
291 if (tcp_ca_needs_ecn(sk))
292 tcp_ca_event(sk, CA_EVENT_ECN_NO_CE);
293 tp->ecn_flags |= TCP_ECN_SEEN;
298 static void tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb)
300 if (tcp_sk(sk)->ecn_flags & TCP_ECN_OK)
301 __tcp_ecn_check_ce(sk, skb);
304 static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
306 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
307 tp->ecn_flags &= ~TCP_ECN_OK;
310 static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
312 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
313 tp->ecn_flags &= ~TCP_ECN_OK;
316 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
318 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
323 /* Buffer size and advertised window tuning.
325 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
328 static void tcp_sndbuf_expand(struct sock *sk)
330 const struct tcp_sock *tp = tcp_sk(sk);
331 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
335 /* Worst case is non GSO/TSO : each frame consumes one skb
336 * and skb->head is kmalloced using power of two area of memory
338 per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
340 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
342 per_mss = roundup_pow_of_two(per_mss) +
343 SKB_DATA_ALIGN(sizeof(struct sk_buff));
345 nr_segs = max_t(u32, TCP_INIT_CWND, tp->snd_cwnd);
346 nr_segs = max_t(u32, nr_segs, tp->reordering + 1);
348 /* Fast Recovery (RFC 5681 3.2) :
349 * Cubic needs 1.7 factor, rounded to 2 to include
350 * extra cushion (application might react slowly to EPOLLOUT)
352 sndmem = ca_ops->sndbuf_expand ? ca_ops->sndbuf_expand(sk) : 2;
353 sndmem *= nr_segs * per_mss;
355 if (sk->sk_sndbuf < sndmem)
356 sk->sk_sndbuf = min(sndmem, sock_net(sk)->ipv4.sysctl_tcp_wmem[2]);
359 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
361 * All tcp_full_space() is split to two parts: "network" buffer, allocated
362 * forward and advertised in receiver window (tp->rcv_wnd) and
363 * "application buffer", required to isolate scheduling/application
364 * latencies from network.
365 * window_clamp is maximal advertised window. It can be less than
366 * tcp_full_space(), in this case tcp_full_space() - window_clamp
367 * is reserved for "application" buffer. The less window_clamp is
368 * the smoother our behaviour from viewpoint of network, but the lower
369 * throughput and the higher sensitivity of the connection to losses. 8)
371 * rcv_ssthresh is more strict window_clamp used at "slow start"
372 * phase to predict further behaviour of this connection.
373 * It is used for two goals:
374 * - to enforce header prediction at sender, even when application
375 * requires some significant "application buffer". It is check #1.
376 * - to prevent pruning of receive queue because of misprediction
377 * of receiver window. Check #2.
379 * The scheme does not work when sender sends good segments opening
380 * window and then starts to feed us spaghetti. But it should work
381 * in common situations. Otherwise, we have to rely on queue collapsing.
384 /* Slow part of check#2. */
385 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
387 struct tcp_sock *tp = tcp_sk(sk);
389 int truesize = tcp_win_from_space(sk, skb->truesize) >> 1;
390 int window = tcp_win_from_space(sk, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]) >> 1;
392 while (tp->rcv_ssthresh <= window) {
393 if (truesize <= skb->len)
394 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
402 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
404 struct tcp_sock *tp = tcp_sk(sk);
407 if (tp->rcv_ssthresh < tp->window_clamp &&
408 (int)tp->rcv_ssthresh < tcp_space(sk) &&
409 !tcp_under_memory_pressure(sk)) {
412 /* Check #2. Increase window, if skb with such overhead
413 * will fit to rcvbuf in future.
415 if (tcp_win_from_space(sk, skb->truesize) <= skb->len)
416 incr = 2 * tp->advmss;
418 incr = __tcp_grow_window(sk, skb);
421 incr = max_t(int, incr, 2 * skb->len);
422 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
424 inet_csk(sk)->icsk_ack.quick |= 1;
429 /* 3. Try to fixup all. It is made immediately after connection enters
432 void tcp_init_buffer_space(struct sock *sk)
434 int tcp_app_win = sock_net(sk)->ipv4.sysctl_tcp_app_win;
435 struct tcp_sock *tp = tcp_sk(sk);
438 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
439 tcp_sndbuf_expand(sk);
441 tp->rcvq_space.space = min_t(u32, tp->rcv_wnd, TCP_INIT_CWND * tp->advmss);
442 tcp_mstamp_refresh(tp);
443 tp->rcvq_space.time = tp->tcp_mstamp;
444 tp->rcvq_space.seq = tp->copied_seq;
446 maxwin = tcp_full_space(sk);
448 if (tp->window_clamp >= maxwin) {
449 tp->window_clamp = maxwin;
451 if (tcp_app_win && maxwin > 4 * tp->advmss)
452 tp->window_clamp = max(maxwin -
453 (maxwin >> tcp_app_win),
457 /* Force reservation of one segment. */
459 tp->window_clamp > 2 * tp->advmss &&
460 tp->window_clamp + tp->advmss > maxwin)
461 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
463 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
464 tp->snd_cwnd_stamp = tcp_jiffies32;
467 /* 4. Recalculate window clamp after socket hit its memory bounds. */
468 static void tcp_clamp_window(struct sock *sk)
470 struct tcp_sock *tp = tcp_sk(sk);
471 struct inet_connection_sock *icsk = inet_csk(sk);
472 struct net *net = sock_net(sk);
474 icsk->icsk_ack.quick = 0;
476 if (sk->sk_rcvbuf < net->ipv4.sysctl_tcp_rmem[2] &&
477 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
478 !tcp_under_memory_pressure(sk) &&
479 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
480 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
481 net->ipv4.sysctl_tcp_rmem[2]);
483 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
484 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
487 /* Initialize RCV_MSS value.
488 * RCV_MSS is an our guess about MSS used by the peer.
489 * We haven't any direct information about the MSS.
490 * It's better to underestimate the RCV_MSS rather than overestimate.
491 * Overestimations make us ACKing less frequently than needed.
492 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
494 void tcp_initialize_rcv_mss(struct sock *sk)
496 const struct tcp_sock *tp = tcp_sk(sk);
497 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
499 hint = min(hint, tp->rcv_wnd / 2);
500 hint = min(hint, TCP_MSS_DEFAULT);
501 hint = max(hint, TCP_MIN_MSS);
503 inet_csk(sk)->icsk_ack.rcv_mss = hint;
505 EXPORT_SYMBOL(tcp_initialize_rcv_mss);
507 /* Receiver "autotuning" code.
509 * The algorithm for RTT estimation w/o timestamps is based on
510 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
511 * <http://public.lanl.gov/radiant/pubs.html#DRS>
513 * More detail on this code can be found at
514 * <http://staff.psc.edu/jheffner/>,
515 * though this reference is out of date. A new paper
518 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
520 u32 new_sample = tp->rcv_rtt_est.rtt_us;
523 if (new_sample != 0) {
524 /* If we sample in larger samples in the non-timestamp
525 * case, we could grossly overestimate the RTT especially
526 * with chatty applications or bulk transfer apps which
527 * are stalled on filesystem I/O.
529 * Also, since we are only going for a minimum in the
530 * non-timestamp case, we do not smooth things out
531 * else with timestamps disabled convergence takes too
535 m -= (new_sample >> 3);
543 /* No previous measure. */
547 tp->rcv_rtt_est.rtt_us = new_sample;
550 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
554 if (tp->rcv_rtt_est.time == 0)
556 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
558 delta_us = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcv_rtt_est.time);
561 tcp_rcv_rtt_update(tp, delta_us, 1);
564 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
565 tp->rcv_rtt_est.time = tp->tcp_mstamp;
568 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
569 const struct sk_buff *skb)
571 struct tcp_sock *tp = tcp_sk(sk);
573 if (tp->rx_opt.rcv_tsecr == tp->rcv_rtt_last_tsecr)
575 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr;
577 if (TCP_SKB_CB(skb)->end_seq -
578 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss) {
579 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr;
582 if (likely(delta < INT_MAX / (USEC_PER_SEC / TCP_TS_HZ))) {
585 delta_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
586 tcp_rcv_rtt_update(tp, delta_us, 0);
592 * This function should be called every time data is copied to user space.
593 * It calculates the appropriate TCP receive buffer space.
595 void tcp_rcv_space_adjust(struct sock *sk)
597 struct tcp_sock *tp = tcp_sk(sk);
601 trace_tcp_rcv_space_adjust(sk);
603 tcp_mstamp_refresh(tp);
604 time = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcvq_space.time);
605 if (time < (tp->rcv_rtt_est.rtt_us >> 3) || tp->rcv_rtt_est.rtt_us == 0)
608 /* Number of bytes copied to user in last RTT */
609 copied = tp->copied_seq - tp->rcvq_space.seq;
610 if (copied <= tp->rcvq_space.space)
614 * copied = bytes received in previous RTT, our base window
615 * To cope with packet losses, we need a 2x factor
616 * To cope with slow start, and sender growing its cwin by 100 %
617 * every RTT, we need a 4x factor, because the ACK we are sending
618 * now is for the next RTT, not the current one :
619 * <prev RTT . ><current RTT .. ><next RTT .... >
622 if (sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf &&
623 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
627 /* minimal window to cope with packet losses, assuming
628 * steady state. Add some cushion because of small variations.
630 rcvwin = ((u64)copied << 1) + 16 * tp->advmss;
632 /* Accommodate for sender rate increase (eg. slow start) */
633 grow = rcvwin * (copied - tp->rcvq_space.space);
634 do_div(grow, tp->rcvq_space.space);
635 rcvwin += (grow << 1);
637 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
638 while (tcp_win_from_space(sk, rcvmem) < tp->advmss)
641 do_div(rcvwin, tp->advmss);
642 rcvbuf = min_t(u64, rcvwin * rcvmem,
643 sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
644 if (rcvbuf > sk->sk_rcvbuf) {
645 sk->sk_rcvbuf = rcvbuf;
647 /* Make the window clamp follow along. */
648 tp->window_clamp = tcp_win_from_space(sk, rcvbuf);
651 tp->rcvq_space.space = copied;
654 tp->rcvq_space.seq = tp->copied_seq;
655 tp->rcvq_space.time = tp->tcp_mstamp;
658 /* There is something which you must keep in mind when you analyze the
659 * behavior of the tp->ato delayed ack timeout interval. When a
660 * connection starts up, we want to ack as quickly as possible. The
661 * problem is that "good" TCP's do slow start at the beginning of data
662 * transmission. The means that until we send the first few ACK's the
663 * sender will sit on his end and only queue most of his data, because
664 * he can only send snd_cwnd unacked packets at any given time. For
665 * each ACK we send, he increments snd_cwnd and transmits more of his
668 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
670 struct tcp_sock *tp = tcp_sk(sk);
671 struct inet_connection_sock *icsk = inet_csk(sk);
674 inet_csk_schedule_ack(sk);
676 tcp_measure_rcv_mss(sk, skb);
678 tcp_rcv_rtt_measure(tp);
682 if (!icsk->icsk_ack.ato) {
683 /* The _first_ data packet received, initialize
684 * delayed ACK engine.
686 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
687 icsk->icsk_ack.ato = TCP_ATO_MIN;
689 int m = now - icsk->icsk_ack.lrcvtime;
691 if (m <= TCP_ATO_MIN / 2) {
692 /* The fastest case is the first. */
693 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
694 } else if (m < icsk->icsk_ack.ato) {
695 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
696 if (icsk->icsk_ack.ato > icsk->icsk_rto)
697 icsk->icsk_ack.ato = icsk->icsk_rto;
698 } else if (m > icsk->icsk_rto) {
699 /* Too long gap. Apparently sender failed to
700 * restart window, so that we send ACKs quickly.
702 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
706 icsk->icsk_ack.lrcvtime = now;
708 tcp_ecn_check_ce(sk, skb);
711 tcp_grow_window(sk, skb);
714 /* Called to compute a smoothed rtt estimate. The data fed to this
715 * routine either comes from timestamps, or from segments that were
716 * known _not_ to have been retransmitted [see Karn/Partridge
717 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
718 * piece by Van Jacobson.
719 * NOTE: the next three routines used to be one big routine.
720 * To save cycles in the RFC 1323 implementation it was better to break
721 * it up into three procedures. -- erics
723 static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
725 struct tcp_sock *tp = tcp_sk(sk);
726 long m = mrtt_us; /* RTT */
727 u32 srtt = tp->srtt_us;
729 /* The following amusing code comes from Jacobson's
730 * article in SIGCOMM '88. Note that rtt and mdev
731 * are scaled versions of rtt and mean deviation.
732 * This is designed to be as fast as possible
733 * m stands for "measurement".
735 * On a 1990 paper the rto value is changed to:
736 * RTO = rtt + 4 * mdev
738 * Funny. This algorithm seems to be very broken.
739 * These formulae increase RTO, when it should be decreased, increase
740 * too slowly, when it should be increased quickly, decrease too quickly
741 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
742 * does not matter how to _calculate_ it. Seems, it was trap
743 * that VJ failed to avoid. 8)
746 m -= (srtt >> 3); /* m is now error in rtt est */
747 srtt += m; /* rtt = 7/8 rtt + 1/8 new */
749 m = -m; /* m is now abs(error) */
750 m -= (tp->mdev_us >> 2); /* similar update on mdev */
751 /* This is similar to one of Eifel findings.
752 * Eifel blocks mdev updates when rtt decreases.
753 * This solution is a bit different: we use finer gain
754 * for mdev in this case (alpha*beta).
755 * Like Eifel it also prevents growth of rto,
756 * but also it limits too fast rto decreases,
757 * happening in pure Eifel.
762 m -= (tp->mdev_us >> 2); /* similar update on mdev */
764 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */
765 if (tp->mdev_us > tp->mdev_max_us) {
766 tp->mdev_max_us = tp->mdev_us;
767 if (tp->mdev_max_us > tp->rttvar_us)
768 tp->rttvar_us = tp->mdev_max_us;
770 if (after(tp->snd_una, tp->rtt_seq)) {
771 if (tp->mdev_max_us < tp->rttvar_us)
772 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
773 tp->rtt_seq = tp->snd_nxt;
774 tp->mdev_max_us = tcp_rto_min_us(sk);
777 /* no previous measure. */
778 srtt = m << 3; /* take the measured time to be rtt */
779 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */
780 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
781 tp->mdev_max_us = tp->rttvar_us;
782 tp->rtt_seq = tp->snd_nxt;
784 tp->srtt_us = max(1U, srtt);
787 static void tcp_update_pacing_rate(struct sock *sk)
789 const struct tcp_sock *tp = tcp_sk(sk);
792 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
793 rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3);
795 /* current rate is (cwnd * mss) / srtt
796 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
797 * In Congestion Avoidance phase, set it to 120 % the current rate.
799 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
800 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
801 * end of slow start and should slow down.
803 if (tp->snd_cwnd < tp->snd_ssthresh / 2)
804 rate *= sock_net(sk)->ipv4.sysctl_tcp_pacing_ss_ratio;
806 rate *= sock_net(sk)->ipv4.sysctl_tcp_pacing_ca_ratio;
808 rate *= max(tp->snd_cwnd, tp->packets_out);
810 if (likely(tp->srtt_us))
811 do_div(rate, tp->srtt_us);
813 /* WRITE_ONCE() is needed because sch_fq fetches sk_pacing_rate
814 * without any lock. We want to make sure compiler wont store
815 * intermediate values in this location.
817 WRITE_ONCE(sk->sk_pacing_rate, min_t(u64, rate,
818 sk->sk_max_pacing_rate));
821 /* Calculate rto without backoff. This is the second half of Van Jacobson's
822 * routine referred to above.
824 static void tcp_set_rto(struct sock *sk)
826 const struct tcp_sock *tp = tcp_sk(sk);
827 /* Old crap is replaced with new one. 8)
830 * 1. If rtt variance happened to be less 50msec, it is hallucination.
831 * It cannot be less due to utterly erratic ACK generation made
832 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
833 * to do with delayed acks, because at cwnd>2 true delack timeout
834 * is invisible. Actually, Linux-2.4 also generates erratic
835 * ACKs in some circumstances.
837 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
839 /* 2. Fixups made earlier cannot be right.
840 * If we do not estimate RTO correctly without them,
841 * all the algo is pure shit and should be replaced
842 * with correct one. It is exactly, which we pretend to do.
845 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
846 * guarantees that rto is higher.
851 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
853 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
856 cwnd = TCP_INIT_CWND;
857 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
860 /* Take a notice that peer is sending D-SACKs */
861 static void tcp_dsack_seen(struct tcp_sock *tp)
863 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
864 tp->rack.dsack_seen = 1;
868 /* It's reordering when higher sequence was delivered (i.e. sacked) before
869 * some lower never-retransmitted sequence ("low_seq"). The maximum reordering
870 * distance is approximated in full-mss packet distance ("reordering").
872 static void tcp_check_sack_reordering(struct sock *sk, const u32 low_seq,
875 struct tcp_sock *tp = tcp_sk(sk);
876 const u32 mss = tp->mss_cache;
879 fack = tcp_highest_sack_seq(tp);
880 if (!before(low_seq, fack))
883 metric = fack - low_seq;
884 if ((metric > tp->reordering * mss) && mss) {
885 #if FASTRETRANS_DEBUG > 1
886 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
887 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
891 tp->undo_marker ? tp->undo_retrans : 0);
893 tp->reordering = min_t(u32, (metric + mss - 1) / mss,
894 sock_net(sk)->ipv4.sysctl_tcp_max_reordering);
897 /* This exciting event is worth to be remembered. 8) */
899 NET_INC_STATS(sock_net(sk),
900 ts ? LINUX_MIB_TCPTSREORDER : LINUX_MIB_TCPSACKREORDER);
903 /* This must be called before lost_out is incremented */
904 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
906 if (!tp->retransmit_skb_hint ||
907 before(TCP_SKB_CB(skb)->seq,
908 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
909 tp->retransmit_skb_hint = skb;
912 /* Sum the number of packets on the wire we have marked as lost.
913 * There are two cases we care about here:
914 * a) Packet hasn't been marked lost (nor retransmitted),
915 * and this is the first loss.
916 * b) Packet has been marked both lost and retransmitted,
917 * and this means we think it was lost again.
919 static void tcp_sum_lost(struct tcp_sock *tp, struct sk_buff *skb)
921 __u8 sacked = TCP_SKB_CB(skb)->sacked;
923 if (!(sacked & TCPCB_LOST) ||
924 ((sacked & TCPCB_LOST) && (sacked & TCPCB_SACKED_RETRANS)))
925 tp->lost += tcp_skb_pcount(skb);
928 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
930 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
931 tcp_verify_retransmit_hint(tp, skb);
933 tp->lost_out += tcp_skb_pcount(skb);
934 tcp_sum_lost(tp, skb);
935 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
939 void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp, struct sk_buff *skb)
941 tcp_verify_retransmit_hint(tp, skb);
943 tcp_sum_lost(tp, skb);
944 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
945 tp->lost_out += tcp_skb_pcount(skb);
946 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
950 /* This procedure tags the retransmission queue when SACKs arrive.
952 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
953 * Packets in queue with these bits set are counted in variables
954 * sacked_out, retrans_out and lost_out, correspondingly.
956 * Valid combinations are:
957 * Tag InFlight Description
958 * 0 1 - orig segment is in flight.
959 * S 0 - nothing flies, orig reached receiver.
960 * L 0 - nothing flies, orig lost by net.
961 * R 2 - both orig and retransmit are in flight.
962 * L|R 1 - orig is lost, retransmit is in flight.
963 * S|R 1 - orig reached receiver, retrans is still in flight.
964 * (L|S|R is logically valid, it could occur when L|R is sacked,
965 * but it is equivalent to plain S and code short-curcuits it to S.
966 * L|S is logically invalid, it would mean -1 packet in flight 8))
968 * These 6 states form finite state machine, controlled by the following events:
969 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
970 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
971 * 3. Loss detection event of two flavors:
972 * A. Scoreboard estimator decided the packet is lost.
973 * A'. Reno "three dupacks" marks head of queue lost.
974 * B. SACK arrives sacking SND.NXT at the moment, when the
975 * segment was retransmitted.
976 * 4. D-SACK added new rule: D-SACK changes any tag to S.
978 * It is pleasant to note, that state diagram turns out to be commutative,
979 * so that we are allowed not to be bothered by order of our actions,
980 * when multiple events arrive simultaneously. (see the function below).
982 * Reordering detection.
983 * --------------------
984 * Reordering metric is maximal distance, which a packet can be displaced
985 * in packet stream. With SACKs we can estimate it:
987 * 1. SACK fills old hole and the corresponding segment was not
988 * ever retransmitted -> reordering. Alas, we cannot use it
989 * when segment was retransmitted.
990 * 2. The last flaw is solved with D-SACK. D-SACK arrives
991 * for retransmitted and already SACKed segment -> reordering..
992 * Both of these heuristics are not used in Loss state, when we cannot
993 * account for retransmits accurately.
995 * SACK block validation.
996 * ----------------------
998 * SACK block range validation checks that the received SACK block fits to
999 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1000 * Note that SND.UNA is not included to the range though being valid because
1001 * it means that the receiver is rather inconsistent with itself reporting
1002 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1003 * perfectly valid, however, in light of RFC2018 which explicitly states
1004 * that "SACK block MUST reflect the newest segment. Even if the newest
1005 * segment is going to be discarded ...", not that it looks very clever
1006 * in case of head skb. Due to potentional receiver driven attacks, we
1007 * choose to avoid immediate execution of a walk in write queue due to
1008 * reneging and defer head skb's loss recovery to standard loss recovery
1009 * procedure that will eventually trigger (nothing forbids us doing this).
1011 * Implements also blockage to start_seq wrap-around. Problem lies in the
1012 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1013 * there's no guarantee that it will be before snd_nxt (n). The problem
1014 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1017 * <- outs wnd -> <- wrapzone ->
1018 * u e n u_w e_w s n_w
1020 * |<------------+------+----- TCP seqno space --------------+---------->|
1021 * ...-- <2^31 ->| |<--------...
1022 * ...---- >2^31 ------>| |<--------...
1024 * Current code wouldn't be vulnerable but it's better still to discard such
1025 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1026 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1027 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1028 * equal to the ideal case (infinite seqno space without wrap caused issues).
1030 * With D-SACK the lower bound is extended to cover sequence space below
1031 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1032 * again, D-SACK block must not to go across snd_una (for the same reason as
1033 * for the normal SACK blocks, explained above). But there all simplicity
1034 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1035 * fully below undo_marker they do not affect behavior in anyway and can
1036 * therefore be safely ignored. In rare cases (which are more or less
1037 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1038 * fragmentation and packet reordering past skb's retransmission. To consider
1039 * them correctly, the acceptable range must be extended even more though
1040 * the exact amount is rather hard to quantify. However, tp->max_window can
1041 * be used as an exaggerated estimate.
1043 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
1044 u32 start_seq, u32 end_seq)
1046 /* Too far in future, or reversed (interpretation is ambiguous) */
1047 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1050 /* Nasty start_seq wrap-around check (see comments above) */
1051 if (!before(start_seq, tp->snd_nxt))
1054 /* In outstanding window? ...This is valid exit for D-SACKs too.
1055 * start_seq == snd_una is non-sensical (see comments above)
1057 if (after(start_seq, tp->snd_una))
1060 if (!is_dsack || !tp->undo_marker)
1063 /* ...Then it's D-SACK, and must reside below snd_una completely */
1064 if (after(end_seq, tp->snd_una))
1067 if (!before(start_seq, tp->undo_marker))
1071 if (!after(end_seq, tp->undo_marker))
1074 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1075 * start_seq < undo_marker and end_seq >= undo_marker.
1077 return !before(start_seq, end_seq - tp->max_window);
1080 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1081 struct tcp_sack_block_wire *sp, int num_sacks,
1084 struct tcp_sock *tp = tcp_sk(sk);
1085 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1086 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1087 bool dup_sack = false;
1089 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1092 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1093 } else if (num_sacks > 1) {
1094 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1095 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1097 if (!after(end_seq_0, end_seq_1) &&
1098 !before(start_seq_0, start_seq_1)) {
1101 NET_INC_STATS(sock_net(sk),
1102 LINUX_MIB_TCPDSACKOFORECV);
1106 /* D-SACK for already forgotten data... Do dumb counting. */
1107 if (dup_sack && tp->undo_marker && tp->undo_retrans > 0 &&
1108 !after(end_seq_0, prior_snd_una) &&
1109 after(end_seq_0, tp->undo_marker))
1115 struct tcp_sacktag_state {
1117 /* Timestamps for earliest and latest never-retransmitted segment
1118 * that was SACKed. RTO needs the earliest RTT to stay conservative,
1119 * but congestion control should still get an accurate delay signal.
1123 struct rate_sample *rate;
1125 unsigned int mss_now;
1128 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1129 * the incoming SACK may not exactly match but we can find smaller MSS
1130 * aligned portion of it that matches. Therefore we might need to fragment
1131 * which may fail and creates some hassle (caller must handle error case
1134 * FIXME: this could be merged to shift decision code
1136 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1137 u32 start_seq, u32 end_seq)
1141 unsigned int pkt_len;
1144 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1145 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1147 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1148 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1149 mss = tcp_skb_mss(skb);
1150 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1153 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1157 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1162 /* Round if necessary so that SACKs cover only full MSSes
1163 * and/or the remaining small portion (if present)
1165 if (pkt_len > mss) {
1166 unsigned int new_len = (pkt_len / mss) * mss;
1167 if (!in_sack && new_len < pkt_len)
1172 if (pkt_len >= skb->len && !in_sack)
1175 err = tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
1176 pkt_len, mss, GFP_ATOMIC);
1184 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1185 static u8 tcp_sacktag_one(struct sock *sk,
1186 struct tcp_sacktag_state *state, u8 sacked,
1187 u32 start_seq, u32 end_seq,
1188 int dup_sack, int pcount,
1191 struct tcp_sock *tp = tcp_sk(sk);
1193 /* Account D-SACK for retransmitted packet. */
1194 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1195 if (tp->undo_marker && tp->undo_retrans > 0 &&
1196 after(end_seq, tp->undo_marker))
1198 if ((sacked & TCPCB_SACKED_ACKED) &&
1199 before(start_seq, state->reord))
1200 state->reord = start_seq;
1203 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1204 if (!after(end_seq, tp->snd_una))
1207 if (!(sacked & TCPCB_SACKED_ACKED)) {
1208 tcp_rack_advance(tp, sacked, end_seq, xmit_time);
1210 if (sacked & TCPCB_SACKED_RETRANS) {
1211 /* If the segment is not tagged as lost,
1212 * we do not clear RETRANS, believing
1213 * that retransmission is still in flight.
1215 if (sacked & TCPCB_LOST) {
1216 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1217 tp->lost_out -= pcount;
1218 tp->retrans_out -= pcount;
1221 if (!(sacked & TCPCB_RETRANS)) {
1222 /* New sack for not retransmitted frame,
1223 * which was in hole. It is reordering.
1225 if (before(start_seq,
1226 tcp_highest_sack_seq(tp)) &&
1227 before(start_seq, state->reord))
1228 state->reord = start_seq;
1230 if (!after(end_seq, tp->high_seq))
1231 state->flag |= FLAG_ORIG_SACK_ACKED;
1232 if (state->first_sackt == 0)
1233 state->first_sackt = xmit_time;
1234 state->last_sackt = xmit_time;
1237 if (sacked & TCPCB_LOST) {
1238 sacked &= ~TCPCB_LOST;
1239 tp->lost_out -= pcount;
1243 sacked |= TCPCB_SACKED_ACKED;
1244 state->flag |= FLAG_DATA_SACKED;
1245 tp->sacked_out += pcount;
1246 tp->delivered += pcount; /* Out-of-order packets delivered */
1248 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1249 if (tp->lost_skb_hint &&
1250 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1251 tp->lost_cnt_hint += pcount;
1254 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1255 * frames and clear it. undo_retrans is decreased above, L|R frames
1256 * are accounted above as well.
1258 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1259 sacked &= ~TCPCB_SACKED_RETRANS;
1260 tp->retrans_out -= pcount;
1266 /* Shift newly-SACKed bytes from this skb to the immediately previous
1267 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1269 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *prev,
1270 struct sk_buff *skb,
1271 struct tcp_sacktag_state *state,
1272 unsigned int pcount, int shifted, int mss,
1275 struct tcp_sock *tp = tcp_sk(sk);
1276 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1277 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
1281 /* Adjust counters and hints for the newly sacked sequence
1282 * range but discard the return value since prev is already
1283 * marked. We must tag the range first because the seq
1284 * advancement below implicitly advances
1285 * tcp_highest_sack_seq() when skb is highest_sack.
1287 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1288 start_seq, end_seq, dup_sack, pcount,
1289 tcp_skb_timestamp_us(skb));
1290 tcp_rate_skb_delivered(sk, skb, state->rate);
1292 if (skb == tp->lost_skb_hint)
1293 tp->lost_cnt_hint += pcount;
1295 TCP_SKB_CB(prev)->end_seq += shifted;
1296 TCP_SKB_CB(skb)->seq += shifted;
1298 tcp_skb_pcount_add(prev, pcount);
1299 BUG_ON(tcp_skb_pcount(skb) < pcount);
1300 tcp_skb_pcount_add(skb, -pcount);
1302 /* When we're adding to gso_segs == 1, gso_size will be zero,
1303 * in theory this shouldn't be necessary but as long as DSACK
1304 * code can come after this skb later on it's better to keep
1305 * setting gso_size to something.
1307 if (!TCP_SKB_CB(prev)->tcp_gso_size)
1308 TCP_SKB_CB(prev)->tcp_gso_size = mss;
1310 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1311 if (tcp_skb_pcount(skb) <= 1)
1312 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1314 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1315 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1318 BUG_ON(!tcp_skb_pcount(skb));
1319 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1323 /* Whole SKB was eaten :-) */
1325 if (skb == tp->retransmit_skb_hint)
1326 tp->retransmit_skb_hint = prev;
1327 if (skb == tp->lost_skb_hint) {
1328 tp->lost_skb_hint = prev;
1329 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1332 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1333 TCP_SKB_CB(prev)->eor = TCP_SKB_CB(skb)->eor;
1334 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1335 TCP_SKB_CB(prev)->end_seq++;
1337 if (skb == tcp_highest_sack(sk))
1338 tcp_advance_highest_sack(sk, skb);
1340 tcp_skb_collapse_tstamp(prev, skb);
1341 if (unlikely(TCP_SKB_CB(prev)->tx.delivered_mstamp))
1342 TCP_SKB_CB(prev)->tx.delivered_mstamp = 0;
1344 tcp_rtx_queue_unlink_and_free(skb, sk);
1346 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKMERGED);
1351 /* I wish gso_size would have a bit more sane initialization than
1352 * something-or-zero which complicates things
1354 static int tcp_skb_seglen(const struct sk_buff *skb)
1356 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1359 /* Shifting pages past head area doesn't work */
1360 static int skb_can_shift(const struct sk_buff *skb)
1362 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1365 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1368 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1369 struct tcp_sacktag_state *state,
1370 u32 start_seq, u32 end_seq,
1373 struct tcp_sock *tp = tcp_sk(sk);
1374 struct sk_buff *prev;
1380 /* Normally R but no L won't result in plain S */
1382 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1384 if (!skb_can_shift(skb))
1386 /* This frame is about to be dropped (was ACKed). */
1387 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1390 /* Can only happen with delayed DSACK + discard craziness */
1391 prev = skb_rb_prev(skb);
1395 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1398 if (!tcp_skb_can_collapse_to(prev))
1401 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1402 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1406 pcount = tcp_skb_pcount(skb);
1407 mss = tcp_skb_seglen(skb);
1409 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1410 * drop this restriction as unnecessary
1412 if (mss != tcp_skb_seglen(prev))
1415 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1417 /* CHECKME: This is non-MSS split case only?, this will
1418 * cause skipped skbs due to advancing loop btw, original
1419 * has that feature too
1421 if (tcp_skb_pcount(skb) <= 1)
1424 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1426 /* TODO: head merge to next could be attempted here
1427 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1428 * though it might not be worth of the additional hassle
1430 * ...we can probably just fallback to what was done
1431 * previously. We could try merging non-SACKed ones
1432 * as well but it probably isn't going to buy off
1433 * because later SACKs might again split them, and
1434 * it would make skb timestamp tracking considerably
1440 len = end_seq - TCP_SKB_CB(skb)->seq;
1442 BUG_ON(len > skb->len);
1444 /* MSS boundaries should be honoured or else pcount will
1445 * severely break even though it makes things bit trickier.
1446 * Optimize common case to avoid most of the divides
1448 mss = tcp_skb_mss(skb);
1450 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1451 * drop this restriction as unnecessary
1453 if (mss != tcp_skb_seglen(prev))
1458 } else if (len < mss) {
1466 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1467 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1470 if (!skb_shift(prev, skb, len))
1472 if (!tcp_shifted_skb(sk, prev, skb, state, pcount, len, mss, dup_sack))
1475 /* Hole filled allows collapsing with the next as well, this is very
1476 * useful when hole on every nth skb pattern happens
1478 skb = skb_rb_next(prev);
1482 if (!skb_can_shift(skb) ||
1483 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1484 (mss != tcp_skb_seglen(skb)))
1488 if (skb_shift(prev, skb, len)) {
1489 pcount += tcp_skb_pcount(skb);
1490 tcp_shifted_skb(sk, prev, skb, state, tcp_skb_pcount(skb),
1501 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1505 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1506 struct tcp_sack_block *next_dup,
1507 struct tcp_sacktag_state *state,
1508 u32 start_seq, u32 end_seq,
1511 struct tcp_sock *tp = tcp_sk(sk);
1512 struct sk_buff *tmp;
1514 skb_rbtree_walk_from(skb) {
1516 bool dup_sack = dup_sack_in;
1518 /* queue is in-order => we can short-circuit the walk early */
1519 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1523 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1524 in_sack = tcp_match_skb_to_sack(sk, skb,
1525 next_dup->start_seq,
1531 /* skb reference here is a bit tricky to get right, since
1532 * shifting can eat and free both this skb and the next,
1533 * so not even _safe variant of the loop is enough.
1536 tmp = tcp_shift_skb_data(sk, skb, state,
1537 start_seq, end_seq, dup_sack);
1546 in_sack = tcp_match_skb_to_sack(sk, skb,
1552 if (unlikely(in_sack < 0))
1556 TCP_SKB_CB(skb)->sacked =
1559 TCP_SKB_CB(skb)->sacked,
1560 TCP_SKB_CB(skb)->seq,
1561 TCP_SKB_CB(skb)->end_seq,
1563 tcp_skb_pcount(skb),
1564 tcp_skb_timestamp_us(skb));
1565 tcp_rate_skb_delivered(sk, skb, state->rate);
1566 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1567 list_del_init(&skb->tcp_tsorted_anchor);
1569 if (!before(TCP_SKB_CB(skb)->seq,
1570 tcp_highest_sack_seq(tp)))
1571 tcp_advance_highest_sack(sk, skb);
1577 static struct sk_buff *tcp_sacktag_bsearch(struct sock *sk,
1578 struct tcp_sacktag_state *state,
1581 struct rb_node *parent, **p = &sk->tcp_rtx_queue.rb_node;
1582 struct sk_buff *skb;
1586 skb = rb_to_skb(parent);
1587 if (before(seq, TCP_SKB_CB(skb)->seq)) {
1588 p = &parent->rb_left;
1591 if (!before(seq, TCP_SKB_CB(skb)->end_seq)) {
1592 p = &parent->rb_right;
1600 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1601 struct tcp_sacktag_state *state,
1604 if (skb && after(TCP_SKB_CB(skb)->seq, skip_to_seq))
1607 return tcp_sacktag_bsearch(sk, state, skip_to_seq);
1610 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1612 struct tcp_sack_block *next_dup,
1613 struct tcp_sacktag_state *state,
1619 if (before(next_dup->start_seq, skip_to_seq)) {
1620 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1621 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1622 next_dup->start_seq, next_dup->end_seq,
1629 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1631 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1635 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1636 u32 prior_snd_una, struct tcp_sacktag_state *state)
1638 struct tcp_sock *tp = tcp_sk(sk);
1639 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1640 TCP_SKB_CB(ack_skb)->sacked);
1641 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1642 struct tcp_sack_block sp[TCP_NUM_SACKS];
1643 struct tcp_sack_block *cache;
1644 struct sk_buff *skb;
1645 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1647 bool found_dup_sack = false;
1649 int first_sack_index;
1652 state->reord = tp->snd_nxt;
1654 if (!tp->sacked_out)
1655 tcp_highest_sack_reset(sk);
1657 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1658 num_sacks, prior_snd_una);
1659 if (found_dup_sack) {
1660 state->flag |= FLAG_DSACKING_ACK;
1661 tp->delivered++; /* A spurious retransmission is delivered */
1664 /* Eliminate too old ACKs, but take into
1665 * account more or less fresh ones, they can
1666 * contain valid SACK info.
1668 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1671 if (!tp->packets_out)
1675 first_sack_index = 0;
1676 for (i = 0; i < num_sacks; i++) {
1677 bool dup_sack = !i && found_dup_sack;
1679 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1680 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1682 if (!tcp_is_sackblock_valid(tp, dup_sack,
1683 sp[used_sacks].start_seq,
1684 sp[used_sacks].end_seq)) {
1688 if (!tp->undo_marker)
1689 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1691 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1693 /* Don't count olds caused by ACK reordering */
1694 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1695 !after(sp[used_sacks].end_seq, tp->snd_una))
1697 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1700 NET_INC_STATS(sock_net(sk), mib_idx);
1702 first_sack_index = -1;
1706 /* Ignore very old stuff early */
1707 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1713 /* order SACK blocks to allow in order walk of the retrans queue */
1714 for (i = used_sacks - 1; i > 0; i--) {
1715 for (j = 0; j < i; j++) {
1716 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1717 swap(sp[j], sp[j + 1]);
1719 /* Track where the first SACK block goes to */
1720 if (j == first_sack_index)
1721 first_sack_index = j + 1;
1726 state->mss_now = tcp_current_mss(sk);
1730 if (!tp->sacked_out) {
1731 /* It's already past, so skip checking against it */
1732 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1734 cache = tp->recv_sack_cache;
1735 /* Skip empty blocks in at head of the cache */
1736 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1741 while (i < used_sacks) {
1742 u32 start_seq = sp[i].start_seq;
1743 u32 end_seq = sp[i].end_seq;
1744 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1745 struct tcp_sack_block *next_dup = NULL;
1747 if (found_dup_sack && ((i + 1) == first_sack_index))
1748 next_dup = &sp[i + 1];
1750 /* Skip too early cached blocks */
1751 while (tcp_sack_cache_ok(tp, cache) &&
1752 !before(start_seq, cache->end_seq))
1755 /* Can skip some work by looking recv_sack_cache? */
1756 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1757 after(end_seq, cache->start_seq)) {
1760 if (before(start_seq, cache->start_seq)) {
1761 skb = tcp_sacktag_skip(skb, sk, state,
1763 skb = tcp_sacktag_walk(skb, sk, next_dup,
1770 /* Rest of the block already fully processed? */
1771 if (!after(end_seq, cache->end_seq))
1774 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1778 /* ...tail remains todo... */
1779 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1780 /* ...but better entrypoint exists! */
1781 skb = tcp_highest_sack(sk);
1788 skb = tcp_sacktag_skip(skb, sk, state, cache->end_seq);
1789 /* Check overlap against next cached too (past this one already) */
1794 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1795 skb = tcp_highest_sack(sk);
1799 skb = tcp_sacktag_skip(skb, sk, state, start_seq);
1802 skb = tcp_sacktag_walk(skb, sk, next_dup, state,
1803 start_seq, end_seq, dup_sack);
1809 /* Clear the head of the cache sack blocks so we can skip it next time */
1810 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1811 tp->recv_sack_cache[i].start_seq = 0;
1812 tp->recv_sack_cache[i].end_seq = 0;
1814 for (j = 0; j < used_sacks; j++)
1815 tp->recv_sack_cache[i++] = sp[j];
1817 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Loss || tp->undo_marker)
1818 tcp_check_sack_reordering(sk, state->reord, 0);
1820 tcp_verify_left_out(tp);
1823 #if FASTRETRANS_DEBUG > 0
1824 WARN_ON((int)tp->sacked_out < 0);
1825 WARN_ON((int)tp->lost_out < 0);
1826 WARN_ON((int)tp->retrans_out < 0);
1827 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1832 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1833 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1835 static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
1839 holes = max(tp->lost_out, 1U);
1840 holes = min(holes, tp->packets_out);
1842 if ((tp->sacked_out + holes) > tp->packets_out) {
1843 tp->sacked_out = tp->packets_out - holes;
1849 /* If we receive more dupacks than we expected counting segments
1850 * in assumption of absent reordering, interpret this as reordering.
1851 * The only another reason could be bug in receiver TCP.
1853 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1855 struct tcp_sock *tp = tcp_sk(sk);
1857 if (!tcp_limit_reno_sacked(tp))
1860 tp->reordering = min_t(u32, tp->packets_out + addend,
1861 sock_net(sk)->ipv4.sysctl_tcp_max_reordering);
1863 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRENOREORDER);
1866 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1868 static void tcp_add_reno_sack(struct sock *sk, int num_dupack)
1871 struct tcp_sock *tp = tcp_sk(sk);
1872 u32 prior_sacked = tp->sacked_out;
1875 tp->sacked_out += num_dupack;
1876 tcp_check_reno_reordering(sk, 0);
1877 delivered = tp->sacked_out - prior_sacked;
1879 tp->delivered += delivered;
1880 tcp_verify_left_out(tp);
1884 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1886 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1888 struct tcp_sock *tp = tcp_sk(sk);
1891 /* One ACK acked hole. The rest eat duplicate ACKs. */
1892 tp->delivered += max_t(int, acked - tp->sacked_out, 1);
1893 if (acked - 1 >= tp->sacked_out)
1896 tp->sacked_out -= acked - 1;
1898 tcp_check_reno_reordering(sk, acked);
1899 tcp_verify_left_out(tp);
1902 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1907 void tcp_clear_retrans(struct tcp_sock *tp)
1909 tp->retrans_out = 0;
1911 tp->undo_marker = 0;
1912 tp->undo_retrans = -1;
1916 static inline void tcp_init_undo(struct tcp_sock *tp)
1918 tp->undo_marker = tp->snd_una;
1919 /* Retransmission still in flight may cause DSACKs later. */
1920 tp->undo_retrans = tp->retrans_out ? : -1;
1923 static bool tcp_is_rack(const struct sock *sk)
1925 return sock_net(sk)->ipv4.sysctl_tcp_recovery & TCP_RACK_LOSS_DETECTION;
1928 /* If we detect SACK reneging, forget all SACK information
1929 * and reset tags completely, otherwise preserve SACKs. If receiver
1930 * dropped its ofo queue, we will know this due to reneging detection.
1932 static void tcp_timeout_mark_lost(struct sock *sk)
1934 struct tcp_sock *tp = tcp_sk(sk);
1935 struct sk_buff *skb, *head;
1936 bool is_reneg; /* is receiver reneging on SACKs? */
1938 head = tcp_rtx_queue_head(sk);
1939 is_reneg = head && (TCP_SKB_CB(head)->sacked & TCPCB_SACKED_ACKED);
1941 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
1943 /* Mark SACK reneging until we recover from this loss event. */
1944 tp->is_sack_reneg = 1;
1945 } else if (tcp_is_reno(tp)) {
1946 tcp_reset_reno_sack(tp);
1950 skb_rbtree_walk_from(skb) {
1952 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1953 else if (tcp_is_rack(sk) && skb != head &&
1954 tcp_rack_skb_timeout(tp, skb, 0) > 0)
1955 continue; /* Don't mark recently sent ones lost yet */
1956 tcp_mark_skb_lost(sk, skb);
1958 tcp_verify_left_out(tp);
1959 tcp_clear_all_retrans_hints(tp);
1962 /* Enter Loss state. */
1963 void tcp_enter_loss(struct sock *sk)
1965 const struct inet_connection_sock *icsk = inet_csk(sk);
1966 struct tcp_sock *tp = tcp_sk(sk);
1967 struct net *net = sock_net(sk);
1968 bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery;
1970 tcp_timeout_mark_lost(sk);
1972 /* Reduce ssthresh if it has not yet been made inside this window. */
1973 if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
1974 !after(tp->high_seq, tp->snd_una) ||
1975 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1976 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1977 tp->prior_cwnd = tp->snd_cwnd;
1978 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1979 tcp_ca_event(sk, CA_EVENT_LOSS);
1982 tp->snd_cwnd = tcp_packets_in_flight(tp) + 1;
1983 tp->snd_cwnd_cnt = 0;
1984 tp->snd_cwnd_stamp = tcp_jiffies32;
1986 /* Timeout in disordered state after receiving substantial DUPACKs
1987 * suggests that the degree of reordering is over-estimated.
1989 if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
1990 tp->sacked_out >= net->ipv4.sysctl_tcp_reordering)
1991 tp->reordering = min_t(unsigned int, tp->reordering,
1992 net->ipv4.sysctl_tcp_reordering);
1993 tcp_set_ca_state(sk, TCP_CA_Loss);
1994 tp->high_seq = tp->snd_nxt;
1995 tcp_ecn_queue_cwr(tp);
1997 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
1998 * loss recovery is underway except recurring timeout(s) on
1999 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
2001 tp->frto = net->ipv4.sysctl_tcp_frto &&
2002 (new_recovery || icsk->icsk_retransmits) &&
2003 !inet_csk(sk)->icsk_mtup.probe_size;
2006 /* If ACK arrived pointing to a remembered SACK, it means that our
2007 * remembered SACKs do not reflect real state of receiver i.e.
2008 * receiver _host_ is heavily congested (or buggy).
2010 * To avoid big spurious retransmission bursts due to transient SACK
2011 * scoreboard oddities that look like reneging, we give the receiver a
2012 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
2013 * restore sanity to the SACK scoreboard. If the apparent reneging
2014 * persists until this RTO then we'll clear the SACK scoreboard.
2016 static bool tcp_check_sack_reneging(struct sock *sk, int flag)
2018 if (flag & FLAG_SACK_RENEGING) {
2019 struct tcp_sock *tp = tcp_sk(sk);
2020 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4),
2021 msecs_to_jiffies(10));
2023 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2024 delay, TCP_RTO_MAX);
2030 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2031 * counter when SACK is enabled (without SACK, sacked_out is used for
2034 * With reordering, holes may still be in flight, so RFC3517 recovery
2035 * uses pure sacked_out (total number of SACKed segments) even though
2036 * it violates the RFC that uses duplicate ACKs, often these are equal
2037 * but when e.g. out-of-window ACKs or packet duplication occurs,
2038 * they differ. Since neither occurs due to loss, TCP should really
2041 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2043 return tp->sacked_out + 1;
2046 /* Linux NewReno/SACK/ECN state machine.
2047 * --------------------------------------
2049 * "Open" Normal state, no dubious events, fast path.
2050 * "Disorder" In all the respects it is "Open",
2051 * but requires a bit more attention. It is entered when
2052 * we see some SACKs or dupacks. It is split of "Open"
2053 * mainly to move some processing from fast path to slow one.
2054 * "CWR" CWND was reduced due to some Congestion Notification event.
2055 * It can be ECN, ICMP source quench, local device congestion.
2056 * "Recovery" CWND was reduced, we are fast-retransmitting.
2057 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2059 * tcp_fastretrans_alert() is entered:
2060 * - each incoming ACK, if state is not "Open"
2061 * - when arrived ACK is unusual, namely:
2066 * Counting packets in flight is pretty simple.
2068 * in_flight = packets_out - left_out + retrans_out
2070 * packets_out is SND.NXT-SND.UNA counted in packets.
2072 * retrans_out is number of retransmitted segments.
2074 * left_out is number of segments left network, but not ACKed yet.
2076 * left_out = sacked_out + lost_out
2078 * sacked_out: Packets, which arrived to receiver out of order
2079 * and hence not ACKed. With SACKs this number is simply
2080 * amount of SACKed data. Even without SACKs
2081 * it is easy to give pretty reliable estimate of this number,
2082 * counting duplicate ACKs.
2084 * lost_out: Packets lost by network. TCP has no explicit
2085 * "loss notification" feedback from network (for now).
2086 * It means that this number can be only _guessed_.
2087 * Actually, it is the heuristics to predict lossage that
2088 * distinguishes different algorithms.
2090 * F.e. after RTO, when all the queue is considered as lost,
2091 * lost_out = packets_out and in_flight = retrans_out.
2093 * Essentially, we have now a few algorithms detecting
2096 * If the receiver supports SACK:
2098 * RFC6675/3517: It is the conventional algorithm. A packet is
2099 * considered lost if the number of higher sequence packets
2100 * SACKed is greater than or equal the DUPACK thoreshold
2101 * (reordering). This is implemented in tcp_mark_head_lost and
2102 * tcp_update_scoreboard.
2104 * RACK (draft-ietf-tcpm-rack-01): it is a newer algorithm
2105 * (2017-) that checks timing instead of counting DUPACKs.
2106 * Essentially a packet is considered lost if it's not S/ACKed
2107 * after RTT + reordering_window, where both metrics are
2108 * dynamically measured and adjusted. This is implemented in
2109 * tcp_rack_mark_lost.
2111 * If the receiver does not support SACK:
2113 * NewReno (RFC6582): in Recovery we assume that one segment
2114 * is lost (classic Reno). While we are in Recovery and
2115 * a partial ACK arrives, we assume that one more packet
2116 * is lost (NewReno). This heuristics are the same in NewReno
2119 * Really tricky (and requiring careful tuning) part of algorithm
2120 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2121 * The first determines the moment _when_ we should reduce CWND and,
2122 * hence, slow down forward transmission. In fact, it determines the moment
2123 * when we decide that hole is caused by loss, rather than by a reorder.
2125 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2126 * holes, caused by lost packets.
2128 * And the most logically complicated part of algorithm is undo
2129 * heuristics. We detect false retransmits due to both too early
2130 * fast retransmit (reordering) and underestimated RTO, analyzing
2131 * timestamps and D-SACKs. When we detect that some segments were
2132 * retransmitted by mistake and CWND reduction was wrong, we undo
2133 * window reduction and abort recovery phase. This logic is hidden
2134 * inside several functions named tcp_try_undo_<something>.
2137 /* This function decides, when we should leave Disordered state
2138 * and enter Recovery phase, reducing congestion window.
2140 * Main question: may we further continue forward transmission
2141 * with the same cwnd?
2143 static bool tcp_time_to_recover(struct sock *sk, int flag)
2145 struct tcp_sock *tp = tcp_sk(sk);
2147 /* Trick#1: The loss is proven. */
2151 /* Not-A-Trick#2 : Classic rule... */
2152 if (!tcp_is_rack(sk) && tcp_dupack_heuristics(tp) > tp->reordering)
2158 /* Detect loss in event "A" above by marking head of queue up as lost.
2159 * For non-SACK(Reno) senders, the first "packets" number of segments
2160 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2161 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2162 * the maximum SACKed segments to pass before reaching this limit.
2164 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2166 struct tcp_sock *tp = tcp_sk(sk);
2167 struct sk_buff *skb;
2168 int cnt, oldcnt, lost;
2170 /* Use SACK to deduce losses of new sequences sent during recovery */
2171 const u32 loss_high = tcp_is_sack(tp) ? tp->snd_nxt : tp->high_seq;
2173 WARN_ON(packets > tp->packets_out);
2174 skb = tp->lost_skb_hint;
2176 /* Head already handled? */
2177 if (mark_head && after(TCP_SKB_CB(skb)->seq, tp->snd_una))
2179 cnt = tp->lost_cnt_hint;
2181 skb = tcp_rtx_queue_head(sk);
2185 skb_rbtree_walk_from(skb) {
2186 /* TODO: do this better */
2187 /* this is not the most efficient way to do this... */
2188 tp->lost_skb_hint = skb;
2189 tp->lost_cnt_hint = cnt;
2191 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2195 if (tcp_is_reno(tp) ||
2196 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2197 cnt += tcp_skb_pcount(skb);
2199 if (cnt > packets) {
2200 if (tcp_is_sack(tp) ||
2201 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
2202 (oldcnt >= packets))
2205 mss = tcp_skb_mss(skb);
2206 /* If needed, chop off the prefix to mark as lost. */
2207 lost = (packets - oldcnt) * mss;
2208 if (lost < skb->len &&
2209 tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
2210 lost, mss, GFP_ATOMIC) < 0)
2215 tcp_skb_mark_lost(tp, skb);
2220 tcp_verify_left_out(tp);
2223 /* Account newly detected lost packet(s) */
2225 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2227 struct tcp_sock *tp = tcp_sk(sk);
2229 if (tcp_is_sack(tp)) {
2230 int sacked_upto = tp->sacked_out - tp->reordering;
2231 if (sacked_upto >= 0)
2232 tcp_mark_head_lost(sk, sacked_upto, 0);
2233 else if (fast_rexmit)
2234 tcp_mark_head_lost(sk, 1, 1);
2238 static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when)
2240 return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2241 before(tp->rx_opt.rcv_tsecr, when);
2244 /* skb is spurious retransmitted if the returned timestamp echo
2245 * reply is prior to the skb transmission time
2247 static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp,
2248 const struct sk_buff *skb)
2250 return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) &&
2251 tcp_tsopt_ecr_before(tp, tcp_skb_timestamp(skb));
2254 /* Nothing was retransmitted or returned timestamp is less
2255 * than timestamp of the first retransmission.
2257 static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2259 return !tp->retrans_stamp ||
2260 tcp_tsopt_ecr_before(tp, tp->retrans_stamp);
2263 /* Undo procedures. */
2265 /* We can clear retrans_stamp when there are no retransmissions in the
2266 * window. It would seem that it is trivially available for us in
2267 * tp->retrans_out, however, that kind of assumptions doesn't consider
2268 * what will happen if errors occur when sending retransmission for the
2269 * second time. ...It could the that such segment has only
2270 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2271 * the head skb is enough except for some reneging corner cases that
2272 * are not worth the effort.
2274 * Main reason for all this complexity is the fact that connection dying
2275 * time now depends on the validity of the retrans_stamp, in particular,
2276 * that successive retransmissions of a segment must not advance
2277 * retrans_stamp under any conditions.
2279 static bool tcp_any_retrans_done(const struct sock *sk)
2281 const struct tcp_sock *tp = tcp_sk(sk);
2282 struct sk_buff *skb;
2284 if (tp->retrans_out)
2287 skb = tcp_rtx_queue_head(sk);
2288 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2294 static void DBGUNDO(struct sock *sk, const char *msg)
2296 #if FASTRETRANS_DEBUG > 1
2297 struct tcp_sock *tp = tcp_sk(sk);
2298 struct inet_sock *inet = inet_sk(sk);
2300 if (sk->sk_family == AF_INET) {
2301 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2303 &inet->inet_daddr, ntohs(inet->inet_dport),
2304 tp->snd_cwnd, tcp_left_out(tp),
2305 tp->snd_ssthresh, tp->prior_ssthresh,
2308 #if IS_ENABLED(CONFIG_IPV6)
2309 else if (sk->sk_family == AF_INET6) {
2310 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2312 &sk->sk_v6_daddr, ntohs(inet->inet_dport),
2313 tp->snd_cwnd, tcp_left_out(tp),
2314 tp->snd_ssthresh, tp->prior_ssthresh,
2321 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
2323 struct tcp_sock *tp = tcp_sk(sk);
2326 struct sk_buff *skb;
2328 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2329 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2332 tcp_clear_all_retrans_hints(tp);
2335 if (tp->prior_ssthresh) {
2336 const struct inet_connection_sock *icsk = inet_csk(sk);
2338 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2340 if (tp->prior_ssthresh > tp->snd_ssthresh) {
2341 tp->snd_ssthresh = tp->prior_ssthresh;
2342 tcp_ecn_withdraw_cwr(tp);
2345 tp->snd_cwnd_stamp = tcp_jiffies32;
2346 tp->undo_marker = 0;
2347 tp->rack.advanced = 1; /* Force RACK to re-exam losses */
2350 static inline bool tcp_may_undo(const struct tcp_sock *tp)
2352 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2355 /* People celebrate: "We love our President!" */
2356 static bool tcp_try_undo_recovery(struct sock *sk)
2358 struct tcp_sock *tp = tcp_sk(sk);
2360 if (tcp_may_undo(tp)) {
2363 /* Happy end! We did not retransmit anything
2364 * or our original transmission succeeded.
2366 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2367 tcp_undo_cwnd_reduction(sk, false);
2368 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2369 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2371 mib_idx = LINUX_MIB_TCPFULLUNDO;
2373 NET_INC_STATS(sock_net(sk), mib_idx);
2374 } else if (tp->rack.reo_wnd_persist) {
2375 tp->rack.reo_wnd_persist--;
2377 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2378 /* Hold old state until something *above* high_seq
2379 * is ACKed. For Reno it is MUST to prevent false
2380 * fast retransmits (RFC2582). SACK TCP is safe. */
2381 if (!tcp_any_retrans_done(sk))
2382 tp->retrans_stamp = 0;
2385 tcp_set_ca_state(sk, TCP_CA_Open);
2386 tp->is_sack_reneg = 0;
2390 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2391 static bool tcp_try_undo_dsack(struct sock *sk)
2393 struct tcp_sock *tp = tcp_sk(sk);
2395 if (tp->undo_marker && !tp->undo_retrans) {
2396 tp->rack.reo_wnd_persist = min(TCP_RACK_RECOVERY_THRESH,
2397 tp->rack.reo_wnd_persist + 1);
2398 DBGUNDO(sk, "D-SACK");
2399 tcp_undo_cwnd_reduction(sk, false);
2400 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2406 /* Undo during loss recovery after partial ACK or using F-RTO. */
2407 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2409 struct tcp_sock *tp = tcp_sk(sk);
2411 if (frto_undo || tcp_may_undo(tp)) {
2412 tcp_undo_cwnd_reduction(sk, true);
2414 DBGUNDO(sk, "partial loss");
2415 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2417 NET_INC_STATS(sock_net(sk),
2418 LINUX_MIB_TCPSPURIOUSRTOS);
2419 inet_csk(sk)->icsk_retransmits = 0;
2420 if (frto_undo || tcp_is_sack(tp)) {
2421 tcp_set_ca_state(sk, TCP_CA_Open);
2422 tp->is_sack_reneg = 0;
2429 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
2430 * It computes the number of packets to send (sndcnt) based on packets newly
2432 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2433 * cwnd reductions across a full RTT.
2434 * 2) Otherwise PRR uses packet conservation to send as much as delivered.
2435 * But when the retransmits are acked without further losses, PRR
2436 * slow starts cwnd up to ssthresh to speed up the recovery.
2438 static void tcp_init_cwnd_reduction(struct sock *sk)
2440 struct tcp_sock *tp = tcp_sk(sk);
2442 tp->high_seq = tp->snd_nxt;
2443 tp->tlp_high_seq = 0;
2444 tp->snd_cwnd_cnt = 0;
2445 tp->prior_cwnd = tp->snd_cwnd;
2446 tp->prr_delivered = 0;
2448 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2449 tcp_ecn_queue_cwr(tp);
2452 void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked, int flag)
2454 struct tcp_sock *tp = tcp_sk(sk);
2456 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2458 if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd))
2461 tp->prr_delivered += newly_acked_sacked;
2463 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2465 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2466 } else if ((flag & (FLAG_RETRANS_DATA_ACKED | FLAG_LOST_RETRANS)) ==
2467 FLAG_RETRANS_DATA_ACKED) {
2468 sndcnt = min_t(int, delta,
2469 max_t(int, tp->prr_delivered - tp->prr_out,
2470 newly_acked_sacked) + 1);
2472 sndcnt = min(delta, newly_acked_sacked);
2474 /* Force a fast retransmit upon entering fast recovery */
2475 sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1));
2476 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2479 static inline void tcp_end_cwnd_reduction(struct sock *sk)
2481 struct tcp_sock *tp = tcp_sk(sk);
2483 if (inet_csk(sk)->icsk_ca_ops->cong_control)
2486 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2487 if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH &&
2488 (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR || tp->undo_marker)) {
2489 tp->snd_cwnd = tp->snd_ssthresh;
2490 tp->snd_cwnd_stamp = tcp_jiffies32;
2492 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2495 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2496 void tcp_enter_cwr(struct sock *sk)
2498 struct tcp_sock *tp = tcp_sk(sk);
2500 tp->prior_ssthresh = 0;
2501 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2502 tp->undo_marker = 0;
2503 tcp_init_cwnd_reduction(sk);
2504 tcp_set_ca_state(sk, TCP_CA_CWR);
2507 EXPORT_SYMBOL(tcp_enter_cwr);
2509 static void tcp_try_keep_open(struct sock *sk)
2511 struct tcp_sock *tp = tcp_sk(sk);
2512 int state = TCP_CA_Open;
2514 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2515 state = TCP_CA_Disorder;
2517 if (inet_csk(sk)->icsk_ca_state != state) {
2518 tcp_set_ca_state(sk, state);
2519 tp->high_seq = tp->snd_nxt;
2523 static void tcp_try_to_open(struct sock *sk, int flag)
2525 struct tcp_sock *tp = tcp_sk(sk);
2527 tcp_verify_left_out(tp);
2529 if (!tcp_any_retrans_done(sk))
2530 tp->retrans_stamp = 0;
2532 if (flag & FLAG_ECE)
2535 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2536 tcp_try_keep_open(sk);
2540 static void tcp_mtup_probe_failed(struct sock *sk)
2542 struct inet_connection_sock *icsk = inet_csk(sk);
2544 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2545 icsk->icsk_mtup.probe_size = 0;
2546 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPFAIL);
2549 static void tcp_mtup_probe_success(struct sock *sk)
2551 struct tcp_sock *tp = tcp_sk(sk);
2552 struct inet_connection_sock *icsk = inet_csk(sk);
2554 /* FIXME: breaks with very large cwnd */
2555 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2556 tp->snd_cwnd = tp->snd_cwnd *
2557 tcp_mss_to_mtu(sk, tp->mss_cache) /
2558 icsk->icsk_mtup.probe_size;
2559 tp->snd_cwnd_cnt = 0;
2560 tp->snd_cwnd_stamp = tcp_jiffies32;
2561 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2563 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2564 icsk->icsk_mtup.probe_size = 0;
2565 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2566 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS);
2569 /* Do a simple retransmit without using the backoff mechanisms in
2570 * tcp_timer. This is used for path mtu discovery.
2571 * The socket is already locked here.
2573 void tcp_simple_retransmit(struct sock *sk)
2575 const struct inet_connection_sock *icsk = inet_csk(sk);
2576 struct tcp_sock *tp = tcp_sk(sk);
2577 struct sk_buff *skb;
2578 unsigned int mss = tcp_current_mss(sk);
2580 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2581 if (tcp_skb_seglen(skb) > mss &&
2582 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2583 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2584 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2585 tp->retrans_out -= tcp_skb_pcount(skb);
2587 tcp_skb_mark_lost_uncond_verify(tp, skb);
2591 tcp_clear_retrans_hints_partial(tp);
2596 if (tcp_is_reno(tp))
2597 tcp_limit_reno_sacked(tp);
2599 tcp_verify_left_out(tp);
2601 /* Don't muck with the congestion window here.
2602 * Reason is that we do not increase amount of _data_
2603 * in network, but units changed and effective
2604 * cwnd/ssthresh really reduced now.
2606 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2607 tp->high_seq = tp->snd_nxt;
2608 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2609 tp->prior_ssthresh = 0;
2610 tp->undo_marker = 0;
2611 tcp_set_ca_state(sk, TCP_CA_Loss);
2613 tcp_xmit_retransmit_queue(sk);
2615 EXPORT_SYMBOL(tcp_simple_retransmit);
2617 void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2619 struct tcp_sock *tp = tcp_sk(sk);
2622 if (tcp_is_reno(tp))
2623 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2625 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2627 NET_INC_STATS(sock_net(sk), mib_idx);
2629 tp->prior_ssthresh = 0;
2632 if (!tcp_in_cwnd_reduction(sk)) {
2634 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2635 tcp_init_cwnd_reduction(sk);
2637 tcp_set_ca_state(sk, TCP_CA_Recovery);
2640 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2641 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2643 static void tcp_process_loss(struct sock *sk, int flag, int num_dupack,
2646 struct tcp_sock *tp = tcp_sk(sk);
2647 bool recovered = !before(tp->snd_una, tp->high_seq);
2649 if ((flag & FLAG_SND_UNA_ADVANCED) &&
2650 tcp_try_undo_loss(sk, false))
2653 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2654 /* Step 3.b. A timeout is spurious if not all data are
2655 * lost, i.e., never-retransmitted data are (s)acked.
2657 if ((flag & FLAG_ORIG_SACK_ACKED) &&
2658 tcp_try_undo_loss(sk, true))
2661 if (after(tp->snd_nxt, tp->high_seq)) {
2662 if (flag & FLAG_DATA_SACKED || num_dupack)
2663 tp->frto = 0; /* Step 3.a. loss was real */
2664 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2665 tp->high_seq = tp->snd_nxt;
2666 /* Step 2.b. Try send new data (but deferred until cwnd
2667 * is updated in tcp_ack()). Otherwise fall back to
2668 * the conventional recovery.
2670 if (!tcp_write_queue_empty(sk) &&
2671 after(tcp_wnd_end(tp), tp->snd_nxt)) {
2672 *rexmit = REXMIT_NEW;
2680 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2681 tcp_try_undo_recovery(sk);
2684 if (tcp_is_reno(tp)) {
2685 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2686 * delivered. Lower inflight to clock out (re)tranmissions.
2688 if (after(tp->snd_nxt, tp->high_seq) && num_dupack)
2689 tcp_add_reno_sack(sk, num_dupack);
2690 else if (flag & FLAG_SND_UNA_ADVANCED)
2691 tcp_reset_reno_sack(tp);
2693 *rexmit = REXMIT_LOST;
2696 /* Undo during fast recovery after partial ACK. */
2697 static bool tcp_try_undo_partial(struct sock *sk, u32 prior_snd_una)
2699 struct tcp_sock *tp = tcp_sk(sk);
2701 if (tp->undo_marker && tcp_packet_delayed(tp)) {
2702 /* Plain luck! Hole if filled with delayed
2703 * packet, rather than with a retransmit. Check reordering.
2705 tcp_check_sack_reordering(sk, prior_snd_una, 1);
2707 /* We are getting evidence that the reordering degree is higher
2708 * than we realized. If there are no retransmits out then we
2709 * can undo. Otherwise we clock out new packets but do not
2710 * mark more packets lost or retransmit more.
2712 if (tp->retrans_out)
2715 if (!tcp_any_retrans_done(sk))
2716 tp->retrans_stamp = 0;
2718 DBGUNDO(sk, "partial recovery");
2719 tcp_undo_cwnd_reduction(sk, true);
2720 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2721 tcp_try_keep_open(sk);
2727 static void tcp_identify_packet_loss(struct sock *sk, int *ack_flag)
2729 struct tcp_sock *tp = tcp_sk(sk);
2731 if (tcp_rtx_queue_empty(sk))
2734 if (unlikely(tcp_is_reno(tp))) {
2735 tcp_newreno_mark_lost(sk, *ack_flag & FLAG_SND_UNA_ADVANCED);
2736 } else if (tcp_is_rack(sk)) {
2737 u32 prior_retrans = tp->retrans_out;
2739 tcp_rack_mark_lost(sk);
2740 if (prior_retrans > tp->retrans_out)
2741 *ack_flag |= FLAG_LOST_RETRANS;
2745 static bool tcp_force_fast_retransmit(struct sock *sk)
2747 struct tcp_sock *tp = tcp_sk(sk);
2749 return after(tcp_highest_sack_seq(tp),
2750 tp->snd_una + tp->reordering * tp->mss_cache);
2753 /* Process an event, which can update packets-in-flight not trivially.
2754 * Main goal of this function is to calculate new estimate for left_out,
2755 * taking into account both packets sitting in receiver's buffer and
2756 * packets lost by network.
2758 * Besides that it updates the congestion state when packet loss or ECN
2759 * is detected. But it does not reduce the cwnd, it is done by the
2760 * congestion control later.
2762 * It does _not_ decide what to send, it is made in function
2763 * tcp_xmit_retransmit_queue().
2765 static void tcp_fastretrans_alert(struct sock *sk, const u32 prior_snd_una,
2766 int num_dupack, int *ack_flag, int *rexmit)
2768 struct inet_connection_sock *icsk = inet_csk(sk);
2769 struct tcp_sock *tp = tcp_sk(sk);
2770 int fast_rexmit = 0, flag = *ack_flag;
2771 bool do_lost = num_dupack || ((flag & FLAG_DATA_SACKED) &&
2772 tcp_force_fast_retransmit(sk));
2774 if (!tp->packets_out && tp->sacked_out)
2777 /* Now state machine starts.
2778 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2779 if (flag & FLAG_ECE)
2780 tp->prior_ssthresh = 0;
2782 /* B. In all the states check for reneging SACKs. */
2783 if (tcp_check_sack_reneging(sk, flag))
2786 /* C. Check consistency of the current state. */
2787 tcp_verify_left_out(tp);
2789 /* D. Check state exit conditions. State can be terminated
2790 * when high_seq is ACKed. */
2791 if (icsk->icsk_ca_state == TCP_CA_Open) {
2792 WARN_ON(tp->retrans_out != 0);
2793 tp->retrans_stamp = 0;
2794 } else if (!before(tp->snd_una, tp->high_seq)) {
2795 switch (icsk->icsk_ca_state) {
2797 /* CWR is to be held something *above* high_seq
2798 * is ACKed for CWR bit to reach receiver. */
2799 if (tp->snd_una != tp->high_seq) {
2800 tcp_end_cwnd_reduction(sk);
2801 tcp_set_ca_state(sk, TCP_CA_Open);
2805 case TCP_CA_Recovery:
2806 if (tcp_is_reno(tp))
2807 tcp_reset_reno_sack(tp);
2808 if (tcp_try_undo_recovery(sk))
2810 tcp_end_cwnd_reduction(sk);
2815 /* E. Process state. */
2816 switch (icsk->icsk_ca_state) {
2817 case TCP_CA_Recovery:
2818 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2819 if (tcp_is_reno(tp))
2820 tcp_add_reno_sack(sk, num_dupack);
2822 if (tcp_try_undo_partial(sk, prior_snd_una))
2824 /* Partial ACK arrived. Force fast retransmit. */
2825 do_lost = tcp_is_reno(tp) ||
2826 tcp_force_fast_retransmit(sk);
2828 if (tcp_try_undo_dsack(sk)) {
2829 tcp_try_keep_open(sk);
2832 tcp_identify_packet_loss(sk, ack_flag);
2835 tcp_process_loss(sk, flag, num_dupack, rexmit);
2836 tcp_identify_packet_loss(sk, ack_flag);
2837 if (!(icsk->icsk_ca_state == TCP_CA_Open ||
2838 (*ack_flag & FLAG_LOST_RETRANS)))
2840 /* Change state if cwnd is undone or retransmits are lost */
2843 if (tcp_is_reno(tp)) {
2844 if (flag & FLAG_SND_UNA_ADVANCED)
2845 tcp_reset_reno_sack(tp);
2846 tcp_add_reno_sack(sk, num_dupack);
2849 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
2850 tcp_try_undo_dsack(sk);
2852 tcp_identify_packet_loss(sk, ack_flag);
2853 if (!tcp_time_to_recover(sk, flag)) {
2854 tcp_try_to_open(sk, flag);
2858 /* MTU probe failure: don't reduce cwnd */
2859 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2860 icsk->icsk_mtup.probe_size &&
2861 tp->snd_una == tp->mtu_probe.probe_seq_start) {
2862 tcp_mtup_probe_failed(sk);
2863 /* Restores the reduction we did in tcp_mtup_probe() */
2865 tcp_simple_retransmit(sk);
2869 /* Otherwise enter Recovery state */
2870 tcp_enter_recovery(sk, (flag & FLAG_ECE));
2874 if (!tcp_is_rack(sk) && do_lost)
2875 tcp_update_scoreboard(sk, fast_rexmit);
2876 *rexmit = REXMIT_LOST;
2879 static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us, const int flag)
2881 u32 wlen = sock_net(sk)->ipv4.sysctl_tcp_min_rtt_wlen * HZ;
2882 struct tcp_sock *tp = tcp_sk(sk);
2884 if ((flag & FLAG_ACK_MAYBE_DELAYED) && rtt_us > tcp_min_rtt(tp)) {
2885 /* If the remote keeps returning delayed ACKs, eventually
2886 * the min filter would pick it up and overestimate the
2887 * prop. delay when it expires. Skip suspected delayed ACKs.
2891 minmax_running_min(&tp->rtt_min, wlen, tcp_jiffies32,
2892 rtt_us ? : jiffies_to_usecs(1));
2895 static bool tcp_ack_update_rtt(struct sock *sk, const int flag,
2896 long seq_rtt_us, long sack_rtt_us,
2897 long ca_rtt_us, struct rate_sample *rs)
2899 const struct tcp_sock *tp = tcp_sk(sk);
2901 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2902 * broken middle-boxes or peers may corrupt TS-ECR fields. But
2903 * Karn's algorithm forbids taking RTT if some retransmitted data
2904 * is acked (RFC6298).
2907 seq_rtt_us = sack_rtt_us;
2909 /* RTTM Rule: A TSecr value received in a segment is used to
2910 * update the averaged RTT measurement only if the segment
2911 * acknowledges some new data, i.e., only if it advances the
2912 * left edge of the send window.
2913 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2915 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2916 flag & FLAG_ACKED) {
2917 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr;
2919 if (likely(delta < INT_MAX / (USEC_PER_SEC / TCP_TS_HZ))) {
2920 seq_rtt_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
2921 ca_rtt_us = seq_rtt_us;
2924 rs->rtt_us = ca_rtt_us; /* RTT of last (S)ACKed packet (or -1) */
2928 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
2929 * always taken together with ACK, SACK, or TS-opts. Any negative
2930 * values will be skipped with the seq_rtt_us < 0 check above.
2932 tcp_update_rtt_min(sk, ca_rtt_us, flag);
2933 tcp_rtt_estimator(sk, seq_rtt_us);
2936 /* RFC6298: only reset backoff on valid RTT measurement. */
2937 inet_csk(sk)->icsk_backoff = 0;
2941 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
2942 void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req)
2944 struct rate_sample rs;
2947 if (req && !req->num_retrans && tcp_rsk(req)->snt_synack)
2948 rtt_us = tcp_stamp_us_delta(tcp_clock_us(), tcp_rsk(req)->snt_synack);
2950 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us, &rs);
2954 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
2956 const struct inet_connection_sock *icsk = inet_csk(sk);
2958 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
2959 tcp_sk(sk)->snd_cwnd_stamp = tcp_jiffies32;
2962 /* Restart timer after forward progress on connection.
2963 * RFC2988 recommends to restart timer to now+rto.
2965 void tcp_rearm_rto(struct sock *sk)
2967 const struct inet_connection_sock *icsk = inet_csk(sk);
2968 struct tcp_sock *tp = tcp_sk(sk);
2970 /* If the retrans timer is currently being used by Fast Open
2971 * for SYN-ACK retrans purpose, stay put.
2973 if (tp->fastopen_rsk)
2976 if (!tp->packets_out) {
2977 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
2979 u32 rto = inet_csk(sk)->icsk_rto;
2980 /* Offset the time elapsed after installing regular RTO */
2981 if (icsk->icsk_pending == ICSK_TIME_REO_TIMEOUT ||
2982 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
2983 s64 delta_us = tcp_rto_delta_us(sk);
2984 /* delta_us may not be positive if the socket is locked
2985 * when the retrans timer fires and is rescheduled.
2987 rto = usecs_to_jiffies(max_t(int, delta_us, 1));
2989 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
2990 TCP_RTO_MAX, tcp_rtx_queue_head(sk));
2994 /* Try to schedule a loss probe; if that doesn't work, then schedule an RTO. */
2995 static void tcp_set_xmit_timer(struct sock *sk)
2997 if (!tcp_schedule_loss_probe(sk, true))
3001 /* If we get here, the whole TSO packet has not been acked. */
3002 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3004 struct tcp_sock *tp = tcp_sk(sk);
3007 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3009 packets_acked = tcp_skb_pcount(skb);
3010 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3012 packets_acked -= tcp_skb_pcount(skb);
3014 if (packets_acked) {
3015 BUG_ON(tcp_skb_pcount(skb) == 0);
3016 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3019 return packets_acked;
3022 static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
3025 const struct skb_shared_info *shinfo;
3027 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3028 if (likely(!TCP_SKB_CB(skb)->txstamp_ack))
3031 shinfo = skb_shinfo(skb);
3032 if (!before(shinfo->tskey, prior_snd_una) &&
3033 before(shinfo->tskey, tcp_sk(sk)->snd_una)) {
3034 tcp_skb_tsorted_save(skb) {
3035 __skb_tstamp_tx(skb, NULL, sk, SCM_TSTAMP_ACK);
3036 } tcp_skb_tsorted_restore(skb);
3040 /* Remove acknowledged frames from the retransmission queue. If our packet
3041 * is before the ack sequence we can discard it as it's confirmed to have
3042 * arrived at the other end.
3044 static int tcp_clean_rtx_queue(struct sock *sk, u32 prior_fack,
3046 struct tcp_sacktag_state *sack)
3048 const struct inet_connection_sock *icsk = inet_csk(sk);
3049 u64 first_ackt, last_ackt;
3050 struct tcp_sock *tp = tcp_sk(sk);
3051 u32 prior_sacked = tp->sacked_out;
3052 u32 reord = tp->snd_nxt; /* lowest acked un-retx un-sacked seq */
3053 struct sk_buff *skb, *next;
3054 bool fully_acked = true;
3055 long sack_rtt_us = -1L;
3056 long seq_rtt_us = -1L;
3057 long ca_rtt_us = -1L;
3059 u32 last_in_flight = 0;
3065 for (skb = skb_rb_first(&sk->tcp_rtx_queue); skb; skb = next) {
3066 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3067 const u32 start_seq = scb->seq;
3068 u8 sacked = scb->sacked;
3071 tcp_ack_tstamp(sk, skb, prior_snd_una);
3073 /* Determine how many packets and what bytes were acked, tso and else */
3074 if (after(scb->end_seq, tp->snd_una)) {
3075 if (tcp_skb_pcount(skb) == 1 ||
3076 !after(tp->snd_una, scb->seq))
3079 acked_pcount = tcp_tso_acked(sk, skb);
3082 fully_acked = false;
3084 acked_pcount = tcp_skb_pcount(skb);
3087 if (unlikely(sacked & TCPCB_RETRANS)) {
3088 if (sacked & TCPCB_SACKED_RETRANS)
3089 tp->retrans_out -= acked_pcount;
3090 flag |= FLAG_RETRANS_DATA_ACKED;
3091 } else if (!(sacked & TCPCB_SACKED_ACKED)) {
3092 last_ackt = tcp_skb_timestamp_us(skb);
3093 WARN_ON_ONCE(last_ackt == 0);
3095 first_ackt = last_ackt;
3097 last_in_flight = TCP_SKB_CB(skb)->tx.in_flight;
3098 if (before(start_seq, reord))
3100 if (!after(scb->end_seq, tp->high_seq))
3101 flag |= FLAG_ORIG_SACK_ACKED;
3104 if (sacked & TCPCB_SACKED_ACKED) {
3105 tp->sacked_out -= acked_pcount;
3106 } else if (tcp_is_sack(tp)) {
3107 tp->delivered += acked_pcount;
3108 if (!tcp_skb_spurious_retrans(tp, skb))
3109 tcp_rack_advance(tp, sacked, scb->end_seq,
3110 tcp_skb_timestamp_us(skb));
3112 if (sacked & TCPCB_LOST)
3113 tp->lost_out -= acked_pcount;
3115 tp->packets_out -= acked_pcount;
3116 pkts_acked += acked_pcount;
3117 tcp_rate_skb_delivered(sk, skb, sack->rate);
3119 /* Initial outgoing SYN's get put onto the write_queue
3120 * just like anything else we transmit. It is not
3121 * true data, and if we misinform our callers that
3122 * this ACK acks real data, we will erroneously exit
3123 * connection startup slow start one packet too
3124 * quickly. This is severely frowned upon behavior.
3126 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
3127 flag |= FLAG_DATA_ACKED;
3129 flag |= FLAG_SYN_ACKED;
3130 tp->retrans_stamp = 0;
3136 next = skb_rb_next(skb);
3137 if (unlikely(skb == tp->retransmit_skb_hint))
3138 tp->retransmit_skb_hint = NULL;
3139 if (unlikely(skb == tp->lost_skb_hint))
3140 tp->lost_skb_hint = NULL;
3141 tcp_rtx_queue_unlink_and_free(skb, sk);
3145 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
3147 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3148 tp->snd_up = tp->snd_una;
3150 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3151 flag |= FLAG_SACK_RENEGING;
3153 if (likely(first_ackt) && !(flag & FLAG_RETRANS_DATA_ACKED)) {
3154 seq_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, first_ackt);
3155 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, last_ackt);
3157 if (pkts_acked == 1 && last_in_flight < tp->mss_cache &&
3158 last_in_flight && !prior_sacked && fully_acked &&
3159 sack->rate->prior_delivered + 1 == tp->delivered &&
3160 !(flag & (FLAG_CA_ALERT | FLAG_SYN_ACKED))) {
3161 /* Conservatively mark a delayed ACK. It's typically
3162 * from a lone runt packet over the round trip to
3163 * a receiver w/o out-of-order or CE events.
3165 flag |= FLAG_ACK_MAYBE_DELAYED;
3168 if (sack->first_sackt) {
3169 sack_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->first_sackt);
3170 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->last_sackt);
3172 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us,
3173 ca_rtt_us, sack->rate);
3175 if (flag & FLAG_ACKED) {
3176 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3177 if (unlikely(icsk->icsk_mtup.probe_size &&
3178 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3179 tcp_mtup_probe_success(sk);
3182 if (tcp_is_reno(tp)) {
3183 tcp_remove_reno_sacks(sk, pkts_acked);
3185 /* If any of the cumulatively ACKed segments was
3186 * retransmitted, non-SACK case cannot confirm that
3187 * progress was due to original transmission due to
3188 * lack of TCPCB_SACKED_ACKED bits even if some of
3189 * the packets may have been never retransmitted.
3191 if (flag & FLAG_RETRANS_DATA_ACKED)
3192 flag &= ~FLAG_ORIG_SACK_ACKED;
3196 /* Non-retransmitted hole got filled? That's reordering */
3197 if (before(reord, prior_fack))
3198 tcp_check_sack_reordering(sk, reord, 0);
3200 delta = prior_sacked - tp->sacked_out;
3201 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3203 } else if (skb && rtt_update && sack_rtt_us >= 0 &&
3204 sack_rtt_us > tcp_stamp_us_delta(tp->tcp_mstamp,
3205 tcp_skb_timestamp_us(skb))) {
3206 /* Do not re-arm RTO if the sack RTT is measured from data sent
3207 * after when the head was last (re)transmitted. Otherwise the
3208 * timeout may continue to extend in loss recovery.
3210 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3213 if (icsk->icsk_ca_ops->pkts_acked) {
3214 struct ack_sample sample = { .pkts_acked = pkts_acked,
3215 .rtt_us = sack->rate->rtt_us,
3216 .in_flight = last_in_flight };
3218 icsk->icsk_ca_ops->pkts_acked(sk, &sample);
3221 #if FASTRETRANS_DEBUG > 0
3222 WARN_ON((int)tp->sacked_out < 0);
3223 WARN_ON((int)tp->lost_out < 0);
3224 WARN_ON((int)tp->retrans_out < 0);
3225 if (!tp->packets_out && tcp_is_sack(tp)) {
3226 icsk = inet_csk(sk);
3228 pr_debug("Leak l=%u %d\n",
3229 tp->lost_out, icsk->icsk_ca_state);
3232 if (tp->sacked_out) {
3233 pr_debug("Leak s=%u %d\n",
3234 tp->sacked_out, icsk->icsk_ca_state);
3237 if (tp->retrans_out) {
3238 pr_debug("Leak r=%u %d\n",
3239 tp->retrans_out, icsk->icsk_ca_state);
3240 tp->retrans_out = 0;
3247 static void tcp_ack_probe(struct sock *sk)
3249 struct inet_connection_sock *icsk = inet_csk(sk);
3250 struct sk_buff *head = tcp_send_head(sk);
3251 const struct tcp_sock *tp = tcp_sk(sk);
3253 /* Was it a usable window open? */
3256 if (!after(TCP_SKB_CB(head)->end_seq, tcp_wnd_end(tp))) {
3257 icsk->icsk_backoff = 0;
3258 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3259 /* Socket must be waked up by subsequent tcp_data_snd_check().
3260 * This function is not for random using!
3263 unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX);
3265 tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3266 when, TCP_RTO_MAX, NULL);
3270 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3272 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3273 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3276 /* Decide wheather to run the increase function of congestion control. */
3277 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3279 /* If reordering is high then always grow cwnd whenever data is
3280 * delivered regardless of its ordering. Otherwise stay conservative
3281 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3282 * new SACK or ECE mark may first advance cwnd here and later reduce
3283 * cwnd in tcp_fastretrans_alert() based on more states.
3285 if (tcp_sk(sk)->reordering > sock_net(sk)->ipv4.sysctl_tcp_reordering)
3286 return flag & FLAG_FORWARD_PROGRESS;
3288 return flag & FLAG_DATA_ACKED;
3291 /* The "ultimate" congestion control function that aims to replace the rigid
3292 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction).
3293 * It's called toward the end of processing an ACK with precise rate
3294 * information. All transmission or retransmission are delayed afterwards.
3296 static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked,
3297 int flag, const struct rate_sample *rs)
3299 const struct inet_connection_sock *icsk = inet_csk(sk);
3301 if (icsk->icsk_ca_ops->cong_control) {
3302 icsk->icsk_ca_ops->cong_control(sk, rs);
3306 if (tcp_in_cwnd_reduction(sk)) {
3307 /* Reduce cwnd if state mandates */
3308 tcp_cwnd_reduction(sk, acked_sacked, flag);
3309 } else if (tcp_may_raise_cwnd(sk, flag)) {
3310 /* Advance cwnd if state allows */
3311 tcp_cong_avoid(sk, ack, acked_sacked);
3313 tcp_update_pacing_rate(sk);
3316 /* Check that window update is acceptable.
3317 * The function assumes that snd_una<=ack<=snd_next.
3319 static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3320 const u32 ack, const u32 ack_seq,
3323 return after(ack, tp->snd_una) ||
3324 after(ack_seq, tp->snd_wl1) ||
3325 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3328 /* If we update tp->snd_una, also update tp->bytes_acked */
3329 static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack)
3331 u32 delta = ack - tp->snd_una;
3333 sock_owned_by_me((struct sock *)tp);
3334 tp->bytes_acked += delta;
3338 /* If we update tp->rcv_nxt, also update tp->bytes_received */
3339 static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq)
3341 u32 delta = seq - tp->rcv_nxt;
3343 sock_owned_by_me((struct sock *)tp);
3344 tp->bytes_received += delta;
3348 /* Update our send window.
3350 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3351 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3353 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3356 struct tcp_sock *tp = tcp_sk(sk);
3358 u32 nwin = ntohs(tcp_hdr(skb)->window);
3360 if (likely(!tcp_hdr(skb)->syn))
3361 nwin <<= tp->rx_opt.snd_wscale;
3363 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3364 flag |= FLAG_WIN_UPDATE;
3365 tcp_update_wl(tp, ack_seq);
3367 if (tp->snd_wnd != nwin) {
3370 /* Note, it is the only place, where
3371 * fast path is recovered for sending TCP.
3374 tcp_fast_path_check(sk);
3376 if (!tcp_write_queue_empty(sk))
3377 tcp_slow_start_after_idle_check(sk);
3379 if (nwin > tp->max_window) {
3380 tp->max_window = nwin;
3381 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3386 tcp_snd_una_update(tp, ack);
3391 static bool __tcp_oow_rate_limited(struct net *net, int mib_idx,
3392 u32 *last_oow_ack_time)
3394 if (*last_oow_ack_time) {
3395 s32 elapsed = (s32)(tcp_jiffies32 - *last_oow_ack_time);
3397 if (0 <= elapsed && elapsed < net->ipv4.sysctl_tcp_invalid_ratelimit) {
3398 NET_INC_STATS(net, mib_idx);
3399 return true; /* rate-limited: don't send yet! */
3403 *last_oow_ack_time = tcp_jiffies32;
3405 return false; /* not rate-limited: go ahead, send dupack now! */
3408 /* Return true if we're currently rate-limiting out-of-window ACKs and
3409 * thus shouldn't send a dupack right now. We rate-limit dupacks in
3410 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3411 * attacks that send repeated SYNs or ACKs for the same connection. To
3412 * do this, we do not send a duplicate SYNACK or ACK if the remote
3413 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3415 bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb,
3416 int mib_idx, u32 *last_oow_ack_time)
3418 /* Data packets without SYNs are not likely part of an ACK loop. */
3419 if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) &&
3423 return __tcp_oow_rate_limited(net, mib_idx, last_oow_ack_time);
3426 /* RFC 5961 7 [ACK Throttling] */
3427 static void tcp_send_challenge_ack(struct sock *sk, const struct sk_buff *skb)
3429 /* unprotected vars, we dont care of overwrites */
3430 static u32 challenge_timestamp;
3431 static unsigned int challenge_count;
3432 struct tcp_sock *tp = tcp_sk(sk);
3433 struct net *net = sock_net(sk);
3436 /* First check our per-socket dupack rate limit. */
3437 if (__tcp_oow_rate_limited(net,
3438 LINUX_MIB_TCPACKSKIPPEDCHALLENGE,
3439 &tp->last_oow_ack_time))
3442 /* Then check host-wide RFC 5961 rate limit. */
3444 if (now != challenge_timestamp) {
3445 u32 ack_limit = net->ipv4.sysctl_tcp_challenge_ack_limit;
3446 u32 half = (ack_limit + 1) >> 1;
3448 challenge_timestamp = now;
3449 WRITE_ONCE(challenge_count, half + prandom_u32_max(ack_limit));
3451 count = READ_ONCE(challenge_count);
3453 WRITE_ONCE(challenge_count, count - 1);
3454 NET_INC_STATS(net, LINUX_MIB_TCPCHALLENGEACK);
3459 static void tcp_store_ts_recent(struct tcp_sock *tp)
3461 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3462 tp->rx_opt.ts_recent_stamp = ktime_get_seconds();
3465 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3467 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3468 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3469 * extra check below makes sure this can only happen
3470 * for pure ACK frames. -DaveM
3472 * Not only, also it occurs for expired timestamps.
3475 if (tcp_paws_check(&tp->rx_opt, 0))
3476 tcp_store_ts_recent(tp);
3480 /* This routine deals with acks during a TLP episode.
3481 * We mark the end of a TLP episode on receiving TLP dupack or when
3482 * ack is after tlp_high_seq.
3483 * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3485 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3487 struct tcp_sock *tp = tcp_sk(sk);
3489 if (before(ack, tp->tlp_high_seq))
3492 if (flag & FLAG_DSACKING_ACK) {
3493 /* This DSACK means original and TLP probe arrived; no loss */
3494 tp->tlp_high_seq = 0;
3495 } else if (after(ack, tp->tlp_high_seq)) {
3496 /* ACK advances: there was a loss, so reduce cwnd. Reset
3497 * tlp_high_seq in tcp_init_cwnd_reduction()
3499 tcp_init_cwnd_reduction(sk);
3500 tcp_set_ca_state(sk, TCP_CA_CWR);
3501 tcp_end_cwnd_reduction(sk);
3502 tcp_try_keep_open(sk);
3503 NET_INC_STATS(sock_net(sk),
3504 LINUX_MIB_TCPLOSSPROBERECOVERY);
3505 } else if (!(flag & (FLAG_SND_UNA_ADVANCED |
3506 FLAG_NOT_DUP | FLAG_DATA_SACKED))) {
3507 /* Pure dupack: original and TLP probe arrived; no loss */
3508 tp->tlp_high_seq = 0;
3512 static inline void tcp_in_ack_event(struct sock *sk, u32 flags)
3514 const struct inet_connection_sock *icsk = inet_csk(sk);
3516 if (icsk->icsk_ca_ops->in_ack_event)
3517 icsk->icsk_ca_ops->in_ack_event(sk, flags);
3520 /* Congestion control has updated the cwnd already. So if we're in
3521 * loss recovery then now we do any new sends (for FRTO) or
3522 * retransmits (for CA_Loss or CA_recovery) that make sense.
3524 static void tcp_xmit_recovery(struct sock *sk, int rexmit)
3526 struct tcp_sock *tp = tcp_sk(sk);
3528 if (rexmit == REXMIT_NONE)
3531 if (unlikely(rexmit == 2)) {
3532 __tcp_push_pending_frames(sk, tcp_current_mss(sk),
3534 if (after(tp->snd_nxt, tp->high_seq))
3538 tcp_xmit_retransmit_queue(sk);
3541 /* Returns the number of packets newly acked or sacked by the current ACK */
3542 static u32 tcp_newly_delivered(struct sock *sk, u32 prior_delivered, int flag)
3544 const struct net *net = sock_net(sk);
3545 struct tcp_sock *tp = tcp_sk(sk);
3548 delivered = tp->delivered - prior_delivered;
3549 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVERED, delivered);
3550 if (flag & FLAG_ECE) {
3551 tp->delivered_ce += delivered;
3552 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVEREDCE, delivered);
3557 /* This routine deals with incoming acks, but not outgoing ones. */
3558 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3560 struct inet_connection_sock *icsk = inet_csk(sk);
3561 struct tcp_sock *tp = tcp_sk(sk);
3562 struct tcp_sacktag_state sack_state;
3563 struct rate_sample rs = { .prior_delivered = 0 };
3564 u32 prior_snd_una = tp->snd_una;
3565 bool is_sack_reneg = tp->is_sack_reneg;
3566 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3567 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3569 int prior_packets = tp->packets_out;
3570 u32 delivered = tp->delivered;
3571 u32 lost = tp->lost;
3572 int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */
3575 sack_state.first_sackt = 0;
3576 sack_state.rate = &rs;
3578 /* We very likely will need to access rtx queue. */
3579 prefetch(sk->tcp_rtx_queue.rb_node);
3581 /* If the ack is older than previous acks
3582 * then we can probably ignore it.
3584 if (before(ack, prior_snd_una)) {
3585 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3586 if (before(ack, prior_snd_una - tp->max_window)) {
3587 if (!(flag & FLAG_NO_CHALLENGE_ACK))
3588 tcp_send_challenge_ack(sk, skb);
3594 /* If the ack includes data we haven't sent yet, discard
3595 * this segment (RFC793 Section 3.9).
3597 if (after(ack, tp->snd_nxt))
3600 if (after(ack, prior_snd_una)) {
3601 flag |= FLAG_SND_UNA_ADVANCED;
3602 icsk->icsk_retransmits = 0;
3604 #if IS_ENABLED(CONFIG_TLS_DEVICE)
3605 if (static_branch_unlikely(&clean_acked_data_enabled))
3606 if (icsk->icsk_clean_acked)
3607 icsk->icsk_clean_acked(sk, ack);
3611 prior_fack = tcp_is_sack(tp) ? tcp_highest_sack_seq(tp) : tp->snd_una;
3612 rs.prior_in_flight = tcp_packets_in_flight(tp);
3614 /* ts_recent update must be made after we are sure that the packet
3617 if (flag & FLAG_UPDATE_TS_RECENT)
3618 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3620 if ((flag & (FLAG_SLOWPATH | FLAG_SND_UNA_ADVANCED)) ==
3621 FLAG_SND_UNA_ADVANCED) {
3622 /* Window is constant, pure forward advance.
3623 * No more checks are required.
3624 * Note, we use the fact that SND.UNA>=SND.WL2.
3626 tcp_update_wl(tp, ack_seq);
3627 tcp_snd_una_update(tp, ack);
3628 flag |= FLAG_WIN_UPDATE;
3630 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
3632 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPACKS);
3634 u32 ack_ev_flags = CA_ACK_SLOWPATH;
3636 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3639 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3641 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3643 if (TCP_SKB_CB(skb)->sacked)
3644 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3647 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) {
3649 ack_ev_flags |= CA_ACK_ECE;
3652 if (flag & FLAG_WIN_UPDATE)
3653 ack_ev_flags |= CA_ACK_WIN_UPDATE;
3655 tcp_in_ack_event(sk, ack_ev_flags);
3658 /* We passed data and got it acked, remove any soft error
3659 * log. Something worked...
3661 sk->sk_err_soft = 0;
3662 icsk->icsk_probes_out = 0;
3663 tp->rcv_tstamp = tcp_jiffies32;
3667 /* See if we can take anything off of the retransmit queue. */
3668 flag |= tcp_clean_rtx_queue(sk, prior_fack, prior_snd_una, &sack_state);
3670 tcp_rack_update_reo_wnd(sk, &rs);
3672 if (tp->tlp_high_seq)
3673 tcp_process_tlp_ack(sk, ack, flag);
3674 /* If needed, reset TLP/RTO timer; RACK may later override this. */
3675 if (flag & FLAG_SET_XMIT_TIMER)
3676 tcp_set_xmit_timer(sk);
3678 if (tcp_ack_is_dubious(sk, flag)) {
3679 if (!(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP))) {
3681 /* Consider if pure acks were aggregated in tcp_add_backlog() */
3682 if (!(flag & FLAG_DATA))
3683 num_dupack = max_t(u16, 1, skb_shinfo(skb)->gso_segs);
3685 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
3689 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
3692 delivered = tcp_newly_delivered(sk, delivered, flag);
3693 lost = tp->lost - lost; /* freshly marked lost */
3694 rs.is_ack_delayed = !!(flag & FLAG_ACK_MAYBE_DELAYED);
3695 tcp_rate_gen(sk, delivered, lost, is_sack_reneg, sack_state.rate);
3696 tcp_cong_control(sk, ack, delivered, flag, sack_state.rate);
3697 tcp_xmit_recovery(sk, rexmit);
3701 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3702 if (flag & FLAG_DSACKING_ACK) {
3703 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
3705 tcp_newly_delivered(sk, delivered, flag);
3707 /* If this ack opens up a zero window, clear backoff. It was
3708 * being used to time the probes, and is probably far higher than
3709 * it needs to be for normal retransmission.
3713 if (tp->tlp_high_seq)
3714 tcp_process_tlp_ack(sk, ack, flag);
3718 SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3722 /* If data was SACKed, tag it and see if we should send more data.
3723 * If data was DSACKed, see if we can undo a cwnd reduction.
3725 if (TCP_SKB_CB(skb)->sacked) {
3726 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3728 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
3730 tcp_newly_delivered(sk, delivered, flag);
3731 tcp_xmit_recovery(sk, rexmit);
3734 SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3738 static void tcp_parse_fastopen_option(int len, const unsigned char *cookie,
3739 bool syn, struct tcp_fastopen_cookie *foc,
3742 /* Valid only in SYN or SYN-ACK with an even length. */
3743 if (!foc || !syn || len < 0 || (len & 1))
3746 if (len >= TCP_FASTOPEN_COOKIE_MIN &&
3747 len <= TCP_FASTOPEN_COOKIE_MAX)
3748 memcpy(foc->val, cookie, len);
3755 static void smc_parse_options(const struct tcphdr *th,
3756 struct tcp_options_received *opt_rx,
3757 const unsigned char *ptr,
3760 #if IS_ENABLED(CONFIG_SMC)
3761 if (static_branch_unlikely(&tcp_have_smc)) {
3762 if (th->syn && !(opsize & 1) &&
3763 opsize >= TCPOLEN_EXP_SMC_BASE &&
3764 get_unaligned_be32(ptr) == TCPOPT_SMC_MAGIC)
3770 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3771 * But, this can also be called on packets in the established flow when
3772 * the fast version below fails.
3774 void tcp_parse_options(const struct net *net,
3775 const struct sk_buff *skb,
3776 struct tcp_options_received *opt_rx, int estab,
3777 struct tcp_fastopen_cookie *foc)
3779 const unsigned char *ptr;
3780 const struct tcphdr *th = tcp_hdr(skb);
3781 int length = (th->doff * 4) - sizeof(struct tcphdr);
3783 ptr = (const unsigned char *)(th + 1);
3784 opt_rx->saw_tstamp = 0;
3786 while (length > 0) {
3787 int opcode = *ptr++;
3793 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3798 if (opsize < 2) /* "silly options" */
3800 if (opsize > length)
3801 return; /* don't parse partial options */
3804 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3805 u16 in_mss = get_unaligned_be16(ptr);
3807 if (opt_rx->user_mss &&
3808 opt_rx->user_mss < in_mss)
3809 in_mss = opt_rx->user_mss;
3810 opt_rx->mss_clamp = in_mss;
3815 if (opsize == TCPOLEN_WINDOW && th->syn &&
3816 !estab && net->ipv4.sysctl_tcp_window_scaling) {
3817 __u8 snd_wscale = *(__u8 *)ptr;
3818 opt_rx->wscale_ok = 1;
3819 if (snd_wscale > TCP_MAX_WSCALE) {
3820 net_info_ratelimited("%s: Illegal window scaling value %d > %u received\n",
3824 snd_wscale = TCP_MAX_WSCALE;
3826 opt_rx->snd_wscale = snd_wscale;
3829 case TCPOPT_TIMESTAMP:
3830 if ((opsize == TCPOLEN_TIMESTAMP) &&
3831 ((estab && opt_rx->tstamp_ok) ||
3832 (!estab && net->ipv4.sysctl_tcp_timestamps))) {
3833 opt_rx->saw_tstamp = 1;
3834 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3835 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3838 case TCPOPT_SACK_PERM:
3839 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3840 !estab && net->ipv4.sysctl_tcp_sack) {
3841 opt_rx->sack_ok = TCP_SACK_SEEN;
3842 tcp_sack_reset(opt_rx);
3847 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3848 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3850 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3853 #ifdef CONFIG_TCP_MD5SIG
3856 * The MD5 Hash has already been
3857 * checked (see tcp_v{4,6}_do_rcv()).
3861 case TCPOPT_FASTOPEN:
3862 tcp_parse_fastopen_option(
3863 opsize - TCPOLEN_FASTOPEN_BASE,
3864 ptr, th->syn, foc, false);
3868 /* Fast Open option shares code 254 using a
3869 * 16 bits magic number.
3871 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE &&
3872 get_unaligned_be16(ptr) ==
3873 TCPOPT_FASTOPEN_MAGIC)
3874 tcp_parse_fastopen_option(opsize -
3875 TCPOLEN_EXP_FASTOPEN_BASE,
3876 ptr + 2, th->syn, foc, true);
3878 smc_parse_options(th, opt_rx, ptr,
3888 EXPORT_SYMBOL(tcp_parse_options);
3890 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
3892 const __be32 *ptr = (const __be32 *)(th + 1);
3894 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3895 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3896 tp->rx_opt.saw_tstamp = 1;
3898 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3901 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
3903 tp->rx_opt.rcv_tsecr = 0;
3909 /* Fast parse options. This hopes to only see timestamps.
3910 * If it is wrong it falls back on tcp_parse_options().
3912 static bool tcp_fast_parse_options(const struct net *net,
3913 const struct sk_buff *skb,
3914 const struct tcphdr *th, struct tcp_sock *tp)
3916 /* In the spirit of fast parsing, compare doff directly to constant
3917 * values. Because equality is used, short doff can be ignored here.
3919 if (th->doff == (sizeof(*th) / 4)) {
3920 tp->rx_opt.saw_tstamp = 0;
3922 } else if (tp->rx_opt.tstamp_ok &&
3923 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3924 if (tcp_parse_aligned_timestamp(tp, th))
3928 tcp_parse_options(net, skb, &tp->rx_opt, 1, NULL);
3929 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3930 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
3935 #ifdef CONFIG_TCP_MD5SIG
3937 * Parse MD5 Signature option
3939 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
3941 int length = (th->doff << 2) - sizeof(*th);
3942 const u8 *ptr = (const u8 *)(th + 1);
3944 /* If not enough data remaining, we can short cut */
3945 while (length >= TCPOLEN_MD5SIG) {
3946 int opcode = *ptr++;
3957 if (opsize < 2 || opsize > length)
3959 if (opcode == TCPOPT_MD5SIG)
3960 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
3967 EXPORT_SYMBOL(tcp_parse_md5sig_option);
3970 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3972 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3973 * it can pass through stack. So, the following predicate verifies that
3974 * this segment is not used for anything but congestion avoidance or
3975 * fast retransmit. Moreover, we even are able to eliminate most of such
3976 * second order effects, if we apply some small "replay" window (~RTO)
3977 * to timestamp space.
3979 * All these measures still do not guarantee that we reject wrapped ACKs
3980 * on networks with high bandwidth, when sequence space is recycled fastly,
3981 * but it guarantees that such events will be very rare and do not affect
3982 * connection seriously. This doesn't look nice, but alas, PAWS is really
3985 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3986 * states that events when retransmit arrives after original data are rare.
3987 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3988 * the biggest problem on large power networks even with minor reordering.
3989 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3990 * up to bandwidth of 18Gigabit/sec. 8) ]
3993 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3995 const struct tcp_sock *tp = tcp_sk(sk);
3996 const struct tcphdr *th = tcp_hdr(skb);
3997 u32 seq = TCP_SKB_CB(skb)->seq;
3998 u32 ack = TCP_SKB_CB(skb)->ack_seq;
4000 return (/* 1. Pure ACK with correct sequence number. */
4001 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
4003 /* 2. ... and duplicate ACK. */
4004 ack == tp->snd_una &&
4006 /* 3. ... and does not update window. */
4007 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
4009 /* 4. ... and sits in replay window. */
4010 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
4013 static inline bool tcp_paws_discard(const struct sock *sk,
4014 const struct sk_buff *skb)
4016 const struct tcp_sock *tp = tcp_sk(sk);
4018 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
4019 !tcp_disordered_ack(sk, skb);
4022 /* Check segment sequence number for validity.
4024 * Segment controls are considered valid, if the segment
4025 * fits to the window after truncation to the window. Acceptability
4026 * of data (and SYN, FIN, of course) is checked separately.
4027 * See tcp_data_queue(), for example.
4029 * Also, controls (RST is main one) are accepted using RCV.WUP instead
4030 * of RCV.NXT. Peer still did not advance his SND.UNA when we
4031 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4032 * (borrowed from freebsd)
4035 static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
4037 return !before(end_seq, tp->rcv_wup) &&
4038 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
4041 /* When we get a reset we do this. */
4042 void tcp_reset(struct sock *sk)
4044 trace_tcp_receive_reset(sk);
4046 /* We want the right error as BSD sees it (and indeed as we do). */
4047 switch (sk->sk_state) {
4049 sk->sk_err = ECONNREFUSED;
4051 case TCP_CLOSE_WAIT:
4057 sk->sk_err = ECONNRESET;
4059 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4062 tcp_write_queue_purge(sk);
4065 if (!sock_flag(sk, SOCK_DEAD))
4066 sk->sk_error_report(sk);
4070 * Process the FIN bit. This now behaves as it is supposed to work
4071 * and the FIN takes effect when it is validly part of sequence
4072 * space. Not before when we get holes.
4074 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4075 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4078 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4079 * close and we go into CLOSING (and later onto TIME-WAIT)
4081 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4083 void tcp_fin(struct sock *sk)
4085 struct tcp_sock *tp = tcp_sk(sk);
4087 inet_csk_schedule_ack(sk);
4089 sk->sk_shutdown |= RCV_SHUTDOWN;
4090 sock_set_flag(sk, SOCK_DONE);
4092 switch (sk->sk_state) {
4094 case TCP_ESTABLISHED:
4095 /* Move to CLOSE_WAIT */
4096 tcp_set_state(sk, TCP_CLOSE_WAIT);
4097 inet_csk(sk)->icsk_ack.pingpong = 1;
4100 case TCP_CLOSE_WAIT:
4102 /* Received a retransmission of the FIN, do
4107 /* RFC793: Remain in the LAST-ACK state. */
4111 /* This case occurs when a simultaneous close
4112 * happens, we must ack the received FIN and
4113 * enter the CLOSING state.
4116 tcp_set_state(sk, TCP_CLOSING);
4119 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4121 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4124 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4125 * cases we should never reach this piece of code.
4127 pr_err("%s: Impossible, sk->sk_state=%d\n",
4128 __func__, sk->sk_state);
4132 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4133 * Probably, we should reset in this case. For now drop them.
4135 skb_rbtree_purge(&tp->out_of_order_queue);
4136 if (tcp_is_sack(tp))
4137 tcp_sack_reset(&tp->rx_opt);
4140 if (!sock_flag(sk, SOCK_DEAD)) {
4141 sk->sk_state_change(sk);
4143 /* Do not send POLL_HUP for half duplex close. */
4144 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4145 sk->sk_state == TCP_CLOSE)
4146 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4148 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4152 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4155 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4156 if (before(seq, sp->start_seq))
4157 sp->start_seq = seq;
4158 if (after(end_seq, sp->end_seq))
4159 sp->end_seq = end_seq;
4165 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4167 struct tcp_sock *tp = tcp_sk(sk);
4169 if (tcp_is_sack(tp) && sock_net(sk)->ipv4.sysctl_tcp_dsack) {
4172 if (before(seq, tp->rcv_nxt))
4173 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4175 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4177 NET_INC_STATS(sock_net(sk), mib_idx);
4179 tp->rx_opt.dsack = 1;
4180 tp->duplicate_sack[0].start_seq = seq;
4181 tp->duplicate_sack[0].end_seq = end_seq;
4185 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4187 struct tcp_sock *tp = tcp_sk(sk);
4189 if (!tp->rx_opt.dsack)
4190 tcp_dsack_set(sk, seq, end_seq);
4192 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4195 static void tcp_rcv_spurious_retrans(struct sock *sk, const struct sk_buff *skb)
4197 /* When the ACK path fails or drops most ACKs, the sender would
4198 * timeout and spuriously retransmit the same segment repeatedly.
4199 * The receiver remembers and reflects via DSACKs. Leverage the
4200 * DSACK state and change the txhash to re-route speculatively.
4202 if (TCP_SKB_CB(skb)->seq == tcp_sk(sk)->duplicate_sack[0].start_seq)
4203 sk_rethink_txhash(sk);
4206 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4208 struct tcp_sock *tp = tcp_sk(sk);
4210 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4211 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4212 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4213 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
4215 if (tcp_is_sack(tp) && sock_net(sk)->ipv4.sysctl_tcp_dsack) {
4216 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4218 tcp_rcv_spurious_retrans(sk, skb);
4219 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4220 end_seq = tp->rcv_nxt;
4221 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4228 /* These routines update the SACK block as out-of-order packets arrive or
4229 * in-order packets close up the sequence space.
4231 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4234 struct tcp_sack_block *sp = &tp->selective_acks[0];
4235 struct tcp_sack_block *swalk = sp + 1;
4237 /* See if the recent change to the first SACK eats into
4238 * or hits the sequence space of other SACK blocks, if so coalesce.
4240 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4241 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4244 /* Zap SWALK, by moving every further SACK up by one slot.
4245 * Decrease num_sacks.
4247 tp->rx_opt.num_sacks--;
4248 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4252 this_sack++, swalk++;
4256 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4258 struct tcp_sock *tp = tcp_sk(sk);
4259 struct tcp_sack_block *sp = &tp->selective_acks[0];
4260 int cur_sacks = tp->rx_opt.num_sacks;
4266 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4267 if (tcp_sack_extend(sp, seq, end_seq)) {
4268 /* Rotate this_sack to the first one. */
4269 for (; this_sack > 0; this_sack--, sp--)
4270 swap(*sp, *(sp - 1));
4272 tcp_sack_maybe_coalesce(tp);
4277 /* Could not find an adjacent existing SACK, build a new one,
4278 * put it at the front, and shift everyone else down. We
4279 * always know there is at least one SACK present already here.
4281 * If the sack array is full, forget about the last one.
4283 if (this_sack >= TCP_NUM_SACKS) {
4284 if (tp->compressed_ack > TCP_FASTRETRANS_THRESH)
4287 tp->rx_opt.num_sacks--;
4290 for (; this_sack > 0; this_sack--, sp--)
4294 /* Build the new head SACK, and we're done. */
4295 sp->start_seq = seq;
4296 sp->end_seq = end_seq;
4297 tp->rx_opt.num_sacks++;
4300 /* RCV.NXT advances, some SACKs should be eaten. */
4302 static void tcp_sack_remove(struct tcp_sock *tp)
4304 struct tcp_sack_block *sp = &tp->selective_acks[0];
4305 int num_sacks = tp->rx_opt.num_sacks;
4308 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4309 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4310 tp->rx_opt.num_sacks = 0;
4314 for (this_sack = 0; this_sack < num_sacks;) {
4315 /* Check if the start of the sack is covered by RCV.NXT. */
4316 if (!before(tp->rcv_nxt, sp->start_seq)) {
4319 /* RCV.NXT must cover all the block! */
4320 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4322 /* Zap this SACK, by moving forward any other SACKS. */
4323 for (i = this_sack+1; i < num_sacks; i++)
4324 tp->selective_acks[i-1] = tp->selective_acks[i];
4331 tp->rx_opt.num_sacks = num_sacks;
4335 * tcp_try_coalesce - try to merge skb to prior one
4337 * @dest: destination queue
4339 * @from: buffer to add in queue
4340 * @fragstolen: pointer to boolean
4342 * Before queueing skb @from after @to, try to merge them
4343 * to reduce overall memory use and queue lengths, if cost is small.
4344 * Packets in ofo or receive queues can stay a long time.
4345 * Better try to coalesce them right now to avoid future collapses.
4346 * Returns true if caller should free @from instead of queueing it
4348 static bool tcp_try_coalesce(struct sock *sk,
4350 struct sk_buff *from,
4355 *fragstolen = false;
4357 /* Its possible this segment overlaps with prior segment in queue */
4358 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4361 #ifdef CONFIG_TLS_DEVICE
4362 if (from->decrypted != to->decrypted)
4366 if (!skb_try_coalesce(to, from, fragstolen, &delta))
4369 atomic_add(delta, &sk->sk_rmem_alloc);
4370 sk_mem_charge(sk, delta);
4371 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4372 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4373 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4374 TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
4376 if (TCP_SKB_CB(from)->has_rxtstamp) {
4377 TCP_SKB_CB(to)->has_rxtstamp = true;
4378 to->tstamp = from->tstamp;
4379 skb_hwtstamps(to)->hwtstamp = skb_hwtstamps(from)->hwtstamp;
4385 static bool tcp_ooo_try_coalesce(struct sock *sk,
4387 struct sk_buff *from,
4390 bool res = tcp_try_coalesce(sk, to, from, fragstolen);
4392 /* In case tcp_drop() is called later, update to->gso_segs */
4394 u32 gso_segs = max_t(u16, 1, skb_shinfo(to)->gso_segs) +
4395 max_t(u16, 1, skb_shinfo(from)->gso_segs);
4397 skb_shinfo(to)->gso_segs = min_t(u32, gso_segs, 0xFFFF);
4402 static void tcp_drop(struct sock *sk, struct sk_buff *skb)
4404 sk_drops_add(sk, skb);
4408 /* This one checks to see if we can put data from the
4409 * out_of_order queue into the receive_queue.
4411 static void tcp_ofo_queue(struct sock *sk)
4413 struct tcp_sock *tp = tcp_sk(sk);
4414 __u32 dsack_high = tp->rcv_nxt;
4415 bool fin, fragstolen, eaten;
4416 struct sk_buff *skb, *tail;
4419 p = rb_first(&tp->out_of_order_queue);
4422 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4425 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4426 __u32 dsack = dsack_high;
4427 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4428 dsack_high = TCP_SKB_CB(skb)->end_seq;
4429 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4432 rb_erase(&skb->rbnode, &tp->out_of_order_queue);
4434 if (unlikely(!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))) {
4435 SOCK_DEBUG(sk, "ofo packet was already received\n");
4439 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4440 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4441 TCP_SKB_CB(skb)->end_seq);
4443 tail = skb_peek_tail(&sk->sk_receive_queue);
4444 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen);
4445 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4446 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
4448 __skb_queue_tail(&sk->sk_receive_queue, skb);
4450 kfree_skb_partial(skb, fragstolen);
4452 if (unlikely(fin)) {
4454 /* tcp_fin() purges tp->out_of_order_queue,
4455 * so we must end this loop right now.
4462 static bool tcp_prune_ofo_queue(struct sock *sk);
4463 static int tcp_prune_queue(struct sock *sk);
4465 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4468 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4469 !sk_rmem_schedule(sk, skb, size)) {
4471 if (tcp_prune_queue(sk) < 0)
4474 while (!sk_rmem_schedule(sk, skb, size)) {
4475 if (!tcp_prune_ofo_queue(sk))
4482 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4484 struct tcp_sock *tp = tcp_sk(sk);
4485 struct rb_node **p, *parent;
4486 struct sk_buff *skb1;
4490 tcp_ecn_check_ce(sk, skb);
4492 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4493 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFODROP);
4498 /* Disable header prediction. */
4500 inet_csk_schedule_ack(sk);
4502 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4503 seq = TCP_SKB_CB(skb)->seq;
4504 end_seq = TCP_SKB_CB(skb)->end_seq;
4505 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4506 tp->rcv_nxt, seq, end_seq);
4508 p = &tp->out_of_order_queue.rb_node;
4509 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4510 /* Initial out of order segment, build 1 SACK. */
4511 if (tcp_is_sack(tp)) {
4512 tp->rx_opt.num_sacks = 1;
4513 tp->selective_acks[0].start_seq = seq;
4514 tp->selective_acks[0].end_seq = end_seq;
4516 rb_link_node(&skb->rbnode, NULL, p);
4517 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4518 tp->ooo_last_skb = skb;
4522 /* In the typical case, we are adding an skb to the end of the list.
4523 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
4525 if (tcp_ooo_try_coalesce(sk, tp->ooo_last_skb,
4526 skb, &fragstolen)) {
4528 tcp_grow_window(sk, skb);
4529 kfree_skb_partial(skb, fragstolen);
4533 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
4534 if (!before(seq, TCP_SKB_CB(tp->ooo_last_skb)->end_seq)) {
4535 parent = &tp->ooo_last_skb->rbnode;
4536 p = &parent->rb_right;
4540 /* Find place to insert this segment. Handle overlaps on the way. */
4544 skb1 = rb_to_skb(parent);
4545 if (before(seq, TCP_SKB_CB(skb1)->seq)) {
4546 p = &parent->rb_left;
4549 if (before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4550 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4551 /* All the bits are present. Drop. */
4552 NET_INC_STATS(sock_net(sk),
4553 LINUX_MIB_TCPOFOMERGE);
4556 tcp_dsack_set(sk, seq, end_seq);
4559 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4560 /* Partial overlap. */
4561 tcp_dsack_set(sk, seq, TCP_SKB_CB(skb1)->end_seq);
4563 /* skb's seq == skb1's seq and skb covers skb1.
4564 * Replace skb1 with skb.
4566 rb_replace_node(&skb1->rbnode, &skb->rbnode,
4567 &tp->out_of_order_queue);
4568 tcp_dsack_extend(sk,
4569 TCP_SKB_CB(skb1)->seq,
4570 TCP_SKB_CB(skb1)->end_seq);
4571 NET_INC_STATS(sock_net(sk),
4572 LINUX_MIB_TCPOFOMERGE);
4576 } else if (tcp_ooo_try_coalesce(sk, skb1,
4577 skb, &fragstolen)) {
4580 p = &parent->rb_right;
4583 /* Insert segment into RB tree. */
4584 rb_link_node(&skb->rbnode, parent, p);
4585 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4588 /* Remove other segments covered by skb. */
4589 while ((skb1 = skb_rb_next(skb)) != NULL) {
4590 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4592 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4593 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4597 rb_erase(&skb1->rbnode, &tp->out_of_order_queue);
4598 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4599 TCP_SKB_CB(skb1)->end_seq);
4600 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4603 /* If there is no skb after us, we are the last_skb ! */
4605 tp->ooo_last_skb = skb;
4608 if (tcp_is_sack(tp))
4609 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4612 tcp_grow_window(sk, skb);
4614 skb_set_owner_r(skb, sk);
4618 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb,
4622 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4625 tcp_try_coalesce(sk, tail,
4626 skb, fragstolen)) ? 1 : 0;
4627 tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq);
4629 __skb_queue_tail(&sk->sk_receive_queue, skb);
4630 skb_set_owner_r(skb, sk);
4635 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4637 struct sk_buff *skb;
4645 if (size > PAGE_SIZE) {
4646 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS);
4648 data_len = npages << PAGE_SHIFT;
4649 size = data_len + (size & ~PAGE_MASK);
4651 skb = alloc_skb_with_frags(size - data_len, data_len,
4652 PAGE_ALLOC_COSTLY_ORDER,
4653 &err, sk->sk_allocation);
4657 skb_put(skb, size - data_len);
4658 skb->data_len = data_len;
4661 if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) {
4662 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP);
4666 err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size);
4670 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4671 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4672 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4674 if (tcp_queue_rcv(sk, skb, &fragstolen)) {
4675 WARN_ON_ONCE(fragstolen); /* should not happen */
4687 void tcp_data_ready(struct sock *sk)
4689 const struct tcp_sock *tp = tcp_sk(sk);
4690 int avail = tp->rcv_nxt - tp->copied_seq;
4692 if (avail < sk->sk_rcvlowat && !sock_flag(sk, SOCK_DONE))
4695 sk->sk_data_ready(sk);
4698 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4700 struct tcp_sock *tp = tcp_sk(sk);
4704 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
4709 __skb_pull(skb, tcp_hdr(skb)->doff * 4);
4711 tcp_ecn_accept_cwr(sk, skb);
4713 tp->rx_opt.dsack = 0;
4715 /* Queue data for delivery to the user.
4716 * Packets in sequence go to the receive queue.
4717 * Out of sequence packets to the out_of_order_queue.
4719 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4720 if (tcp_receive_window(tp) == 0) {
4721 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP);
4725 /* Ok. In sequence. In window. */
4727 if (skb_queue_len(&sk->sk_receive_queue) == 0)
4728 sk_forced_mem_schedule(sk, skb->truesize);
4729 else if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) {
4730 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP);
4734 eaten = tcp_queue_rcv(sk, skb, &fragstolen);
4736 tcp_event_data_recv(sk, skb);
4737 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
4740 if (!RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4743 /* RFC5681. 4.2. SHOULD send immediate ACK, when
4744 * gap in queue is filled.
4746 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
4747 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW;
4750 if (tp->rx_opt.num_sacks)
4751 tcp_sack_remove(tp);
4753 tcp_fast_path_check(sk);
4756 kfree_skb_partial(skb, fragstolen);
4757 if (!sock_flag(sk, SOCK_DEAD))
4762 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4763 tcp_rcv_spurious_retrans(sk, skb);
4764 /* A retransmit, 2nd most common case. Force an immediate ack. */
4765 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4766 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4769 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
4770 inet_csk_schedule_ack(sk);
4776 /* Out of window. F.e. zero window probe. */
4777 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4780 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4781 /* Partial packet, seq < rcv_next < end_seq */
4782 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4783 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4784 TCP_SKB_CB(skb)->end_seq);
4786 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4788 /* If window is closed, drop tail of packet. But after
4789 * remembering D-SACK for its head made in previous line.
4791 if (!tcp_receive_window(tp)) {
4792 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP);
4798 tcp_data_queue_ofo(sk, skb);
4801 static struct sk_buff *tcp_skb_next(struct sk_buff *skb, struct sk_buff_head *list)
4804 return !skb_queue_is_last(list, skb) ? skb->next : NULL;
4806 return skb_rb_next(skb);
4809 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4810 struct sk_buff_head *list,
4811 struct rb_root *root)
4813 struct sk_buff *next = tcp_skb_next(skb, list);
4816 __skb_unlink(skb, list);
4818 rb_erase(&skb->rbnode, root);
4821 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4826 /* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */
4827 void tcp_rbtree_insert(struct rb_root *root, struct sk_buff *skb)
4829 struct rb_node **p = &root->rb_node;
4830 struct rb_node *parent = NULL;
4831 struct sk_buff *skb1;
4835 skb1 = rb_to_skb(parent);
4836 if (before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb1)->seq))
4837 p = &parent->rb_left;
4839 p = &parent->rb_right;
4841 rb_link_node(&skb->rbnode, parent, p);
4842 rb_insert_color(&skb->rbnode, root);
4845 /* Collapse contiguous sequence of skbs head..tail with
4846 * sequence numbers start..end.
4848 * If tail is NULL, this means until the end of the queue.
4850 * Segments with FIN/SYN are not collapsed (only because this
4854 tcp_collapse(struct sock *sk, struct sk_buff_head *list, struct rb_root *root,
4855 struct sk_buff *head, struct sk_buff *tail, u32 start, u32 end)
4857 struct sk_buff *skb = head, *n;
4858 struct sk_buff_head tmp;
4861 /* First, check that queue is collapsible and find
4862 * the point where collapsing can be useful.
4865 for (end_of_skbs = true; skb != NULL && skb != tail; skb = n) {
4866 n = tcp_skb_next(skb, list);
4868 /* No new bits? It is possible on ofo queue. */
4869 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4870 skb = tcp_collapse_one(sk, skb, list, root);
4876 /* The first skb to collapse is:
4878 * - bloated or contains data before "start" or
4879 * overlaps to the next one.
4881 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
4882 (tcp_win_from_space(sk, skb->truesize) > skb->len ||
4883 before(TCP_SKB_CB(skb)->seq, start))) {
4884 end_of_skbs = false;
4888 if (n && n != tail &&
4889 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(n)->seq) {
4890 end_of_skbs = false;
4894 /* Decided to skip this, advance start seq. */
4895 start = TCP_SKB_CB(skb)->end_seq;
4898 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4901 __skb_queue_head_init(&tmp);
4903 while (before(start, end)) {
4904 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
4905 struct sk_buff *nskb;
4907 nskb = alloc_skb(copy, GFP_ATOMIC);
4911 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4912 #ifdef CONFIG_TLS_DEVICE
4913 nskb->decrypted = skb->decrypted;
4915 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4917 __skb_queue_before(list, skb, nskb);
4919 __skb_queue_tail(&tmp, nskb); /* defer rbtree insertion */
4920 skb_set_owner_r(nskb, sk);
4922 /* Copy data, releasing collapsed skbs. */
4924 int offset = start - TCP_SKB_CB(skb)->seq;
4925 int size = TCP_SKB_CB(skb)->end_seq - start;
4929 size = min(copy, size);
4930 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4932 TCP_SKB_CB(nskb)->end_seq += size;
4936 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4937 skb = tcp_collapse_one(sk, skb, list, root);
4940 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4942 #ifdef CONFIG_TLS_DEVICE
4943 if (skb->decrypted != nskb->decrypted)
4950 skb_queue_walk_safe(&tmp, skb, n)
4951 tcp_rbtree_insert(root, skb);
4954 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4955 * and tcp_collapse() them until all the queue is collapsed.
4957 static void tcp_collapse_ofo_queue(struct sock *sk)
4959 struct tcp_sock *tp = tcp_sk(sk);
4960 u32 range_truesize, sum_tiny = 0;
4961 struct sk_buff *skb, *head;
4964 skb = skb_rb_first(&tp->out_of_order_queue);
4967 tp->ooo_last_skb = skb_rb_last(&tp->out_of_order_queue);
4970 start = TCP_SKB_CB(skb)->seq;
4971 end = TCP_SKB_CB(skb)->end_seq;
4972 range_truesize = skb->truesize;
4974 for (head = skb;;) {
4975 skb = skb_rb_next(skb);
4977 /* Range is terminated when we see a gap or when
4978 * we are at the queue end.
4981 after(TCP_SKB_CB(skb)->seq, end) ||
4982 before(TCP_SKB_CB(skb)->end_seq, start)) {
4983 /* Do not attempt collapsing tiny skbs */
4984 if (range_truesize != head->truesize ||
4985 end - start >= SKB_WITH_OVERHEAD(SK_MEM_QUANTUM)) {
4986 tcp_collapse(sk, NULL, &tp->out_of_order_queue,
4987 head, skb, start, end);
4989 sum_tiny += range_truesize;
4990 if (sum_tiny > sk->sk_rcvbuf >> 3)
4996 range_truesize += skb->truesize;
4997 if (unlikely(before(TCP_SKB_CB(skb)->seq, start)))
4998 start = TCP_SKB_CB(skb)->seq;
4999 if (after(TCP_SKB_CB(skb)->end_seq, end))
5000 end = TCP_SKB_CB(skb)->end_seq;
5005 * Clean the out-of-order queue to make room.
5006 * We drop high sequences packets to :
5007 * 1) Let a chance for holes to be filled.
5008 * 2) not add too big latencies if thousands of packets sit there.
5009 * (But if application shrinks SO_RCVBUF, we could still end up
5010 * freeing whole queue here)
5011 * 3) Drop at least 12.5 % of sk_rcvbuf to avoid malicious attacks.
5013 * Return true if queue has shrunk.
5015 static bool tcp_prune_ofo_queue(struct sock *sk)
5017 struct tcp_sock *tp = tcp_sk(sk);
5018 struct rb_node *node, *prev;
5021 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
5024 NET_INC_STATS(sock_net(sk), LINUX_MIB_OFOPRUNED);
5025 goal = sk->sk_rcvbuf >> 3;
5026 node = &tp->ooo_last_skb->rbnode;
5028 prev = rb_prev(node);
5029 rb_erase(node, &tp->out_of_order_queue);
5030 goal -= rb_to_skb(node)->truesize;
5031 tcp_drop(sk, rb_to_skb(node));
5032 if (!prev || goal <= 0) {
5034 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf &&
5035 !tcp_under_memory_pressure(sk))
5037 goal = sk->sk_rcvbuf >> 3;
5041 tp->ooo_last_skb = rb_to_skb(prev);
5043 /* Reset SACK state. A conforming SACK implementation will
5044 * do the same at a timeout based retransmit. When a connection
5045 * is in a sad state like this, we care only about integrity
5046 * of the connection not performance.
5048 if (tp->rx_opt.sack_ok)
5049 tcp_sack_reset(&tp->rx_opt);
5053 /* Reduce allocated memory if we can, trying to get
5054 * the socket within its memory limits again.
5056 * Return less than zero if we should start dropping frames
5057 * until the socket owning process reads some of the data
5058 * to stabilize the situation.
5060 static int tcp_prune_queue(struct sock *sk)
5062 struct tcp_sock *tp = tcp_sk(sk);
5064 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
5066 NET_INC_STATS(sock_net(sk), LINUX_MIB_PRUNECALLED);
5068 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
5069 tcp_clamp_window(sk);
5070 else if (tcp_under_memory_pressure(sk))
5071 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
5073 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5076 tcp_collapse_ofo_queue(sk);
5077 if (!skb_queue_empty(&sk->sk_receive_queue))
5078 tcp_collapse(sk, &sk->sk_receive_queue, NULL,
5079 skb_peek(&sk->sk_receive_queue),
5081 tp->copied_seq, tp->rcv_nxt);
5084 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5087 /* Collapsing did not help, destructive actions follow.
5088 * This must not ever occur. */
5090 tcp_prune_ofo_queue(sk);
5092 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5095 /* If we are really being abused, tell the caller to silently
5096 * drop receive data on the floor. It will get retransmitted
5097 * and hopefully then we'll have sufficient space.
5099 NET_INC_STATS(sock_net(sk), LINUX_MIB_RCVPRUNED);
5101 /* Massive buffer overcommit. */
5106 static bool tcp_should_expand_sndbuf(const struct sock *sk)
5108 const struct tcp_sock *tp = tcp_sk(sk);
5110 /* If the user specified a specific send buffer setting, do
5113 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
5116 /* If we are under global TCP memory pressure, do not expand. */
5117 if (tcp_under_memory_pressure(sk))
5120 /* If we are under soft global TCP memory pressure, do not expand. */
5121 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
5124 /* If we filled the congestion window, do not expand. */
5125 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
5131 /* When incoming ACK allowed to free some skb from write_queue,
5132 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
5133 * on the exit from tcp input handler.
5135 * PROBLEM: sndbuf expansion does not work well with largesend.
5137 static void tcp_new_space(struct sock *sk)
5139 struct tcp_sock *tp = tcp_sk(sk);
5141 if (tcp_should_expand_sndbuf(sk)) {
5142 tcp_sndbuf_expand(sk);
5143 tp->snd_cwnd_stamp = tcp_jiffies32;
5146 sk->sk_write_space(sk);
5149 static void tcp_check_space(struct sock *sk)
5151 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
5152 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
5153 /* pairs with tcp_poll() */
5155 if (sk->sk_socket &&
5156 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
5158 if (!test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
5159 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
5164 static inline void tcp_data_snd_check(struct sock *sk)
5166 tcp_push_pending_frames(sk);
5167 tcp_check_space(sk);
5171 * Check if sending an ack is needed.
5173 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
5175 struct tcp_sock *tp = tcp_sk(sk);
5176 unsigned long rtt, delay;
5178 /* More than one full frame received... */
5179 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
5180 /* ... and right edge of window advances far enough.
5181 * (tcp_recvmsg() will send ACK otherwise).
5182 * If application uses SO_RCVLOWAT, we want send ack now if
5183 * we have not received enough bytes to satisfy the condition.
5185 (tp->rcv_nxt - tp->copied_seq < sk->sk_rcvlowat ||
5186 __tcp_select_window(sk) >= tp->rcv_wnd)) ||
5187 /* We ACK each frame or... */
5188 tcp_in_quickack_mode(sk) ||
5189 /* Protocol state mandates a one-time immediate ACK */
5190 inet_csk(sk)->icsk_ack.pending & ICSK_ACK_NOW) {
5196 if (!ofo_possible || RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
5197 tcp_send_delayed_ack(sk);
5201 if (!tcp_is_sack(tp) ||
5202 tp->compressed_ack >= sock_net(sk)->ipv4.sysctl_tcp_comp_sack_nr)
5205 if (tp->compressed_ack_rcv_nxt != tp->rcv_nxt) {
5206 tp->compressed_ack_rcv_nxt = tp->rcv_nxt;
5207 if (tp->compressed_ack > TCP_FASTRETRANS_THRESH)
5208 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED,
5209 tp->compressed_ack - TCP_FASTRETRANS_THRESH);
5210 tp->compressed_ack = 0;
5213 if (++tp->compressed_ack <= TCP_FASTRETRANS_THRESH)
5216 if (hrtimer_is_queued(&tp->compressed_ack_timer))
5219 /* compress ack timer : 5 % of rtt, but no more than tcp_comp_sack_delay_ns */
5221 rtt = tp->rcv_rtt_est.rtt_us;
5222 if (tp->srtt_us && tp->srtt_us < rtt)
5225 delay = min_t(unsigned long, sock_net(sk)->ipv4.sysctl_tcp_comp_sack_delay_ns,
5226 rtt * (NSEC_PER_USEC >> 3)/20);
5228 hrtimer_start(&tp->compressed_ack_timer, ns_to_ktime(delay),
5229 HRTIMER_MODE_REL_PINNED_SOFT);
5232 static inline void tcp_ack_snd_check(struct sock *sk)
5234 if (!inet_csk_ack_scheduled(sk)) {
5235 /* We sent a data segment already. */
5238 __tcp_ack_snd_check(sk, 1);
5242 * This routine is only called when we have urgent data
5243 * signaled. Its the 'slow' part of tcp_urg. It could be
5244 * moved inline now as tcp_urg is only called from one
5245 * place. We handle URGent data wrong. We have to - as
5246 * BSD still doesn't use the correction from RFC961.
5247 * For 1003.1g we should support a new option TCP_STDURG to permit
5248 * either form (or just set the sysctl tcp_stdurg).
5251 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
5253 struct tcp_sock *tp = tcp_sk(sk);
5254 u32 ptr = ntohs(th->urg_ptr);
5256 if (ptr && !sock_net(sk)->ipv4.sysctl_tcp_stdurg)
5258 ptr += ntohl(th->seq);
5260 /* Ignore urgent data that we've already seen and read. */
5261 if (after(tp->copied_seq, ptr))
5264 /* Do not replay urg ptr.
5266 * NOTE: interesting situation not covered by specs.
5267 * Misbehaving sender may send urg ptr, pointing to segment,
5268 * which we already have in ofo queue. We are not able to fetch
5269 * such data and will stay in TCP_URG_NOTYET until will be eaten
5270 * by recvmsg(). Seems, we are not obliged to handle such wicked
5271 * situations. But it is worth to think about possibility of some
5272 * DoSes using some hypothetical application level deadlock.
5274 if (before(ptr, tp->rcv_nxt))
5277 /* Do we already have a newer (or duplicate) urgent pointer? */
5278 if (tp->urg_data && !after(ptr, tp->urg_seq))
5281 /* Tell the world about our new urgent pointer. */
5284 /* We may be adding urgent data when the last byte read was
5285 * urgent. To do this requires some care. We cannot just ignore
5286 * tp->copied_seq since we would read the last urgent byte again
5287 * as data, nor can we alter copied_seq until this data arrives
5288 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5290 * NOTE. Double Dutch. Rendering to plain English: author of comment
5291 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5292 * and expect that both A and B disappear from stream. This is _wrong_.
5293 * Though this happens in BSD with high probability, this is occasional.
5294 * Any application relying on this is buggy. Note also, that fix "works"
5295 * only in this artificial test. Insert some normal data between A and B and we will
5296 * decline of BSD again. Verdict: it is better to remove to trap
5299 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5300 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5301 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5303 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5304 __skb_unlink(skb, &sk->sk_receive_queue);
5309 tp->urg_data = TCP_URG_NOTYET;
5312 /* Disable header prediction. */
5316 /* This is the 'fast' part of urgent handling. */
5317 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
5319 struct tcp_sock *tp = tcp_sk(sk);
5321 /* Check if we get a new urgent pointer - normally not. */
5323 tcp_check_urg(sk, th);
5325 /* Do we wait for any urgent data? - normally not... */
5326 if (tp->urg_data == TCP_URG_NOTYET) {
5327 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5330 /* Is the urgent pointer pointing into this packet? */
5331 if (ptr < skb->len) {
5333 if (skb_copy_bits(skb, ptr, &tmp, 1))
5335 tp->urg_data = TCP_URG_VALID | tmp;
5336 if (!sock_flag(sk, SOCK_DEAD))
5337 sk->sk_data_ready(sk);
5342 /* Accept RST for rcv_nxt - 1 after a FIN.
5343 * When tcp connections are abruptly terminated from Mac OSX (via ^C), a
5344 * FIN is sent followed by a RST packet. The RST is sent with the same
5345 * sequence number as the FIN, and thus according to RFC 5961 a challenge
5346 * ACK should be sent. However, Mac OSX rate limits replies to challenge
5347 * ACKs on the closed socket. In addition middleboxes can drop either the
5348 * challenge ACK or a subsequent RST.
5350 static bool tcp_reset_check(const struct sock *sk, const struct sk_buff *skb)
5352 struct tcp_sock *tp = tcp_sk(sk);
5354 return unlikely(TCP_SKB_CB(skb)->seq == (tp->rcv_nxt - 1) &&
5355 (1 << sk->sk_state) & (TCPF_CLOSE_WAIT | TCPF_LAST_ACK |
5359 /* Does PAWS and seqno based validation of an incoming segment, flags will
5360 * play significant role here.
5362 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5363 const struct tcphdr *th, int syn_inerr)
5365 struct tcp_sock *tp = tcp_sk(sk);
5366 bool rst_seq_match = false;
5368 /* RFC1323: H1. Apply PAWS check first. */
5369 if (tcp_fast_parse_options(sock_net(sk), skb, th, tp) &&
5370 tp->rx_opt.saw_tstamp &&
5371 tcp_paws_discard(sk, skb)) {
5373 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5374 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5375 LINUX_MIB_TCPACKSKIPPEDPAWS,
5376 &tp->last_oow_ack_time))
5377 tcp_send_dupack(sk, skb);
5380 /* Reset is accepted even if it did not pass PAWS. */
5383 /* Step 1: check sequence number */
5384 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5385 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5386 * (RST) segments are validated by checking their SEQ-fields."
5387 * And page 69: "If an incoming segment is not acceptable,
5388 * an acknowledgment should be sent in reply (unless the RST
5389 * bit is set, if so drop the segment and return)".
5394 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5395 LINUX_MIB_TCPACKSKIPPEDSEQ,
5396 &tp->last_oow_ack_time))
5397 tcp_send_dupack(sk, skb);
5398 } else if (tcp_reset_check(sk, skb)) {
5404 /* Step 2: check RST bit */
5406 /* RFC 5961 3.2 (extend to match against (RCV.NXT - 1) after a
5407 * FIN and SACK too if available):
5408 * If seq num matches RCV.NXT or (RCV.NXT - 1) after a FIN, or
5409 * the right-most SACK block,
5411 * RESET the connection
5413 * Send a challenge ACK
5415 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt ||
5416 tcp_reset_check(sk, skb)) {
5417 rst_seq_match = true;
5418 } else if (tcp_is_sack(tp) && tp->rx_opt.num_sacks > 0) {
5419 struct tcp_sack_block *sp = &tp->selective_acks[0];
5420 int max_sack = sp[0].end_seq;
5423 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;
5425 max_sack = after(sp[this_sack].end_seq,
5427 sp[this_sack].end_seq : max_sack;
5430 if (TCP_SKB_CB(skb)->seq == max_sack)
5431 rst_seq_match = true;
5437 /* Disable TFO if RST is out-of-order
5438 * and no data has been received
5439 * for current active TFO socket
5441 if (tp->syn_fastopen && !tp->data_segs_in &&
5442 sk->sk_state == TCP_ESTABLISHED)
5443 tcp_fastopen_active_disable(sk);
5444 tcp_send_challenge_ack(sk, skb);
5449 /* step 3: check security and precedence [ignored] */
5451 /* step 4: Check for a SYN
5452 * RFC 5961 4.2 : Send a challenge ack
5457 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5458 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5459 tcp_send_challenge_ack(sk, skb);
5471 * TCP receive function for the ESTABLISHED state.
5473 * It is split into a fast path and a slow path. The fast path is
5475 * - A zero window was announced from us - zero window probing
5476 * is only handled properly in the slow path.
5477 * - Out of order segments arrived.
5478 * - Urgent data is expected.
5479 * - There is no buffer space left
5480 * - Unexpected TCP flags/window values/header lengths are received
5481 * (detected by checking the TCP header against pred_flags)
5482 * - Data is sent in both directions. Fast path only supports pure senders
5483 * or pure receivers (this means either the sequence number or the ack
5484 * value must stay constant)
5485 * - Unexpected TCP option.
5487 * When these conditions are not satisfied it drops into a standard
5488 * receive procedure patterned after RFC793 to handle all cases.
5489 * The first three cases are guaranteed by proper pred_flags setting,
5490 * the rest is checked inline. Fast processing is turned on in
5491 * tcp_data_queue when everything is OK.
5493 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb)
5495 const struct tcphdr *th = (const struct tcphdr *)skb->data;
5496 struct tcp_sock *tp = tcp_sk(sk);
5497 unsigned int len = skb->len;
5499 /* TCP congestion window tracking */
5500 trace_tcp_probe(sk, skb);
5502 tcp_mstamp_refresh(tp);
5503 if (unlikely(!sk->sk_rx_dst))
5504 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
5506 * Header prediction.
5507 * The code loosely follows the one in the famous
5508 * "30 instruction TCP receive" Van Jacobson mail.
5510 * Van's trick is to deposit buffers into socket queue
5511 * on a device interrupt, to call tcp_recv function
5512 * on the receive process context and checksum and copy
5513 * the buffer to user space. smart...
5515 * Our current scheme is not silly either but we take the
5516 * extra cost of the net_bh soft interrupt processing...
5517 * We do checksum and copy also but from device to kernel.
5520 tp->rx_opt.saw_tstamp = 0;
5522 /* pred_flags is 0xS?10 << 16 + snd_wnd
5523 * if header_prediction is to be made
5524 * 'S' will always be tp->tcp_header_len >> 2
5525 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5526 * turn it off (when there are holes in the receive
5527 * space for instance)
5528 * PSH flag is ignored.
5531 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5532 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5533 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5534 int tcp_header_len = tp->tcp_header_len;
5536 /* Timestamp header prediction: tcp_header_len
5537 * is automatically equal to th->doff*4 due to pred_flags
5541 /* Check timestamp */
5542 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5543 /* No? Slow path! */
5544 if (!tcp_parse_aligned_timestamp(tp, th))
5547 /* If PAWS failed, check it more carefully in slow path */
5548 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5551 /* DO NOT update ts_recent here, if checksum fails
5552 * and timestamp was corrupted part, it will result
5553 * in a hung connection since we will drop all
5554 * future packets due to the PAWS test.
5558 if (len <= tcp_header_len) {
5559 /* Bulk data transfer: sender */
5560 if (len == tcp_header_len) {
5561 /* Predicted packet is in window by definition.
5562 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5563 * Hence, check seq<=rcv_wup reduces to:
5565 if (tcp_header_len ==
5566 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5567 tp->rcv_nxt == tp->rcv_wup)
5568 tcp_store_ts_recent(tp);
5570 /* We know that such packets are checksummed
5573 tcp_ack(sk, skb, 0);
5575 tcp_data_snd_check(sk);
5576 /* When receiving pure ack in fast path, update
5577 * last ts ecr directly instead of calling
5578 * tcp_rcv_rtt_measure_ts()
5580 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr;
5582 } else { /* Header too small */
5583 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5588 bool fragstolen = false;
5590 if (tcp_checksum_complete(skb))
5593 if ((int)skb->truesize > sk->sk_forward_alloc)
5596 /* Predicted packet is in window by definition.
5597 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5598 * Hence, check seq<=rcv_wup reduces to:
5600 if (tcp_header_len ==
5601 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5602 tp->rcv_nxt == tp->rcv_wup)
5603 tcp_store_ts_recent(tp);
5605 tcp_rcv_rtt_measure_ts(sk, skb);
5607 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPHITS);
5609 /* Bulk data transfer: receiver */
5610 __skb_pull(skb, tcp_header_len);
5611 eaten = tcp_queue_rcv(sk, skb, &fragstolen);
5613 tcp_event_data_recv(sk, skb);
5615 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5616 /* Well, only one small jumplet in fast path... */
5617 tcp_ack(sk, skb, FLAG_DATA);
5618 tcp_data_snd_check(sk);
5619 if (!inet_csk_ack_scheduled(sk))
5623 __tcp_ack_snd_check(sk, 0);
5626 kfree_skb_partial(skb, fragstolen);
5633 if (len < (th->doff << 2) || tcp_checksum_complete(skb))
5636 if (!th->ack && !th->rst && !th->syn)
5640 * Standard slow path.
5643 if (!tcp_validate_incoming(sk, skb, th, 1))
5647 if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0)
5650 tcp_rcv_rtt_measure_ts(sk, skb);
5652 /* Process urgent data. */
5653 tcp_urg(sk, skb, th);
5655 /* step 7: process the segment text */
5656 tcp_data_queue(sk, skb);
5658 tcp_data_snd_check(sk);
5659 tcp_ack_snd_check(sk);
5663 TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS);
5664 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5669 EXPORT_SYMBOL(tcp_rcv_established);
5671 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5673 struct tcp_sock *tp = tcp_sk(sk);
5674 struct inet_connection_sock *icsk = inet_csk(sk);
5676 tcp_set_state(sk, TCP_ESTABLISHED);
5677 icsk->icsk_ack.lrcvtime = tcp_jiffies32;
5680 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
5681 security_inet_conn_established(sk, skb);
5682 sk_mark_napi_id(sk, skb);
5685 tcp_init_transfer(sk, BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB);
5687 /* Prevent spurious tcp_cwnd_restart() on first data
5690 tp->lsndtime = tcp_jiffies32;
5692 if (sock_flag(sk, SOCK_KEEPOPEN))
5693 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5695 if (!tp->rx_opt.snd_wscale)
5696 __tcp_fast_path_on(tp, tp->snd_wnd);
5701 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
5702 struct tcp_fastopen_cookie *cookie)
5704 struct tcp_sock *tp = tcp_sk(sk);
5705 struct sk_buff *data = tp->syn_data ? tcp_rtx_queue_head(sk) : NULL;
5706 u16 mss = tp->rx_opt.mss_clamp, try_exp = 0;
5707 bool syn_drop = false;
5709 if (mss == tp->rx_opt.user_mss) {
5710 struct tcp_options_received opt;
5712 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5713 tcp_clear_options(&opt);
5714 opt.user_mss = opt.mss_clamp = 0;
5715 tcp_parse_options(sock_net(sk), synack, &opt, 0, NULL);
5716 mss = opt.mss_clamp;
5719 if (!tp->syn_fastopen) {
5720 /* Ignore an unsolicited cookie */
5722 } else if (tp->total_retrans) {
5723 /* SYN timed out and the SYN-ACK neither has a cookie nor
5724 * acknowledges data. Presumably the remote received only
5725 * the retransmitted (regular) SYNs: either the original
5726 * SYN-data or the corresponding SYN-ACK was dropped.
5728 syn_drop = (cookie->len < 0 && data);
5729 } else if (cookie->len < 0 && !tp->syn_data) {
5730 /* We requested a cookie but didn't get it. If we did not use
5731 * the (old) exp opt format then try so next time (try_exp=1).
5732 * Otherwise we go back to use the RFC7413 opt (try_exp=2).
5734 try_exp = tp->syn_fastopen_exp ? 2 : 1;
5737 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp);
5739 if (data) { /* Retransmit unacked data in SYN */
5740 skb_rbtree_walk_from(data) {
5741 if (__tcp_retransmit_skb(sk, data, 1))
5745 NET_INC_STATS(sock_net(sk),
5746 LINUX_MIB_TCPFASTOPENACTIVEFAIL);
5749 tp->syn_data_acked = tp->syn_data;
5750 if (tp->syn_data_acked) {
5751 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
5752 /* SYN-data is counted as two separate packets in tcp_ack() */
5753 if (tp->delivered > 1)
5757 tcp_fastopen_add_skb(sk, synack);
5762 static void smc_check_reset_syn(struct tcp_sock *tp)
5764 #if IS_ENABLED(CONFIG_SMC)
5765 if (static_branch_unlikely(&tcp_have_smc)) {
5766 if (tp->syn_smc && !tp->rx_opt.smc_ok)
5772 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5773 const struct tcphdr *th)
5775 struct inet_connection_sock *icsk = inet_csk(sk);
5776 struct tcp_sock *tp = tcp_sk(sk);
5777 struct tcp_fastopen_cookie foc = { .len = -1 };
5778 int saved_clamp = tp->rx_opt.mss_clamp;
5781 tcp_parse_options(sock_net(sk), skb, &tp->rx_opt, 0, &foc);
5782 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
5783 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
5787 * "If the state is SYN-SENT then
5788 * first check the ACK bit
5789 * If the ACK bit is set
5790 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5791 * a reset (unless the RST bit is set, if so drop
5792 * the segment and return)"
5794 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
5795 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt))
5796 goto reset_and_undo;
5798 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5799 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5800 tcp_time_stamp(tp))) {
5801 NET_INC_STATS(sock_net(sk),
5802 LINUX_MIB_PAWSACTIVEREJECTED);
5803 goto reset_and_undo;
5806 /* Now ACK is acceptable.
5808 * "If the RST bit is set
5809 * If the ACK was acceptable then signal the user "error:
5810 * connection reset", drop the segment, enter CLOSED state,
5811 * delete TCB, and return."
5820 * "fifth, if neither of the SYN or RST bits is set then
5821 * drop the segment and return."
5827 goto discard_and_undo;
5830 * "If the SYN bit is on ...
5831 * are acceptable then ...
5832 * (our SYN has been ACKed), change the connection
5833 * state to ESTABLISHED..."
5836 tcp_ecn_rcv_synack(tp, th);
5838 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5839 tcp_ack(sk, skb, FLAG_SLOWPATH);
5841 /* Ok.. it's good. Set up sequence numbers and
5842 * move to established.
5844 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5845 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5847 /* RFC1323: The window in SYN & SYN/ACK segments is
5850 tp->snd_wnd = ntohs(th->window);
5852 if (!tp->rx_opt.wscale_ok) {
5853 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5854 tp->window_clamp = min(tp->window_clamp, 65535U);
5857 if (tp->rx_opt.saw_tstamp) {
5858 tp->rx_opt.tstamp_ok = 1;
5859 tp->tcp_header_len =
5860 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5861 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5862 tcp_store_ts_recent(tp);
5864 tp->tcp_header_len = sizeof(struct tcphdr);
5867 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5868 tcp_initialize_rcv_mss(sk);
5870 /* Remember, tcp_poll() does not lock socket!
5871 * Change state from SYN-SENT only after copied_seq
5872 * is initialized. */
5873 tp->copied_seq = tp->rcv_nxt;
5875 smc_check_reset_syn(tp);
5879 tcp_finish_connect(sk, skb);
5881 fastopen_fail = (tp->syn_fastopen || tp->syn_data) &&
5882 tcp_rcv_fastopen_synack(sk, skb, &foc);
5884 if (!sock_flag(sk, SOCK_DEAD)) {
5885 sk->sk_state_change(sk);
5886 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5890 if (sk->sk_write_pending ||
5891 icsk->icsk_accept_queue.rskq_defer_accept ||
5892 icsk->icsk_ack.pingpong) {
5893 /* Save one ACK. Data will be ready after
5894 * several ticks, if write_pending is set.
5896 * It may be deleted, but with this feature tcpdumps
5897 * look so _wonderfully_ clever, that I was not able
5898 * to stand against the temptation 8) --ANK
5900 inet_csk_schedule_ack(sk);
5901 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
5902 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5903 TCP_DELACK_MAX, TCP_RTO_MAX);
5914 /* No ACK in the segment */
5918 * "If the RST bit is set
5920 * Otherwise (no ACK) drop the segment and return."
5923 goto discard_and_undo;
5927 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5928 tcp_paws_reject(&tp->rx_opt, 0))
5929 goto discard_and_undo;
5932 /* We see SYN without ACK. It is attempt of
5933 * simultaneous connect with crossed SYNs.
5934 * Particularly, it can be connect to self.
5936 tcp_set_state(sk, TCP_SYN_RECV);
5938 if (tp->rx_opt.saw_tstamp) {
5939 tp->rx_opt.tstamp_ok = 1;
5940 tcp_store_ts_recent(tp);
5941 tp->tcp_header_len =
5942 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5944 tp->tcp_header_len = sizeof(struct tcphdr);
5947 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5948 tp->copied_seq = tp->rcv_nxt;
5949 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5951 /* RFC1323: The window in SYN & SYN/ACK segments is
5954 tp->snd_wnd = ntohs(th->window);
5955 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5956 tp->max_window = tp->snd_wnd;
5958 tcp_ecn_rcv_syn(tp, th);
5961 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5962 tcp_initialize_rcv_mss(sk);
5964 tcp_send_synack(sk);
5966 /* Note, we could accept data and URG from this segment.
5967 * There are no obstacles to make this (except that we must
5968 * either change tcp_recvmsg() to prevent it from returning data
5969 * before 3WHS completes per RFC793, or employ TCP Fast Open).
5971 * However, if we ignore data in ACKless segments sometimes,
5972 * we have no reasons to accept it sometimes.
5973 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5974 * is not flawless. So, discard packet for sanity.
5975 * Uncomment this return to process the data.
5982 /* "fifth, if neither of the SYN or RST bits is set then
5983 * drop the segment and return."
5987 tcp_clear_options(&tp->rx_opt);
5988 tp->rx_opt.mss_clamp = saved_clamp;
5992 tcp_clear_options(&tp->rx_opt);
5993 tp->rx_opt.mss_clamp = saved_clamp;
5998 * This function implements the receiving procedure of RFC 793 for
5999 * all states except ESTABLISHED and TIME_WAIT.
6000 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
6001 * address independent.
6004 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb)
6006 struct tcp_sock *tp = tcp_sk(sk);
6007 struct inet_connection_sock *icsk = inet_csk(sk);
6008 const struct tcphdr *th = tcp_hdr(skb);
6009 struct request_sock *req;
6013 switch (sk->sk_state) {
6027 /* It is possible that we process SYN packets from backlog,
6028 * so we need to make sure to disable BH and RCU right there.
6032 acceptable = icsk->icsk_af_ops->conn_request(sk, skb) >= 0;
6044 tp->rx_opt.saw_tstamp = 0;
6045 tcp_mstamp_refresh(tp);
6046 queued = tcp_rcv_synsent_state_process(sk, skb, th);
6050 /* Do step6 onward by hand. */
6051 tcp_urg(sk, skb, th);
6053 tcp_data_snd_check(sk);
6057 tcp_mstamp_refresh(tp);
6058 tp->rx_opt.saw_tstamp = 0;
6059 req = tp->fastopen_rsk;
6063 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
6064 sk->sk_state != TCP_FIN_WAIT1);
6066 if (!tcp_check_req(sk, skb, req, true, &req_stolen))
6070 if (!th->ack && !th->rst && !th->syn)
6073 if (!tcp_validate_incoming(sk, skb, th, 0))
6076 /* step 5: check the ACK field */
6077 acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
6078 FLAG_UPDATE_TS_RECENT |
6079 FLAG_NO_CHALLENGE_ACK) > 0;
6082 if (sk->sk_state == TCP_SYN_RECV)
6083 return 1; /* send one RST */
6084 tcp_send_challenge_ack(sk, skb);
6087 switch (sk->sk_state) {
6089 tp->delivered++; /* SYN-ACK delivery isn't tracked in tcp_ack */
6091 tcp_synack_rtt_meas(sk, req);
6093 /* Once we leave TCP_SYN_RECV, we no longer need req
6097 inet_csk(sk)->icsk_retransmits = 0;
6098 reqsk_fastopen_remove(sk, req, false);
6099 /* Re-arm the timer because data may have been sent out.
6100 * This is similar to the regular data transmission case
6101 * when new data has just been ack'ed.
6103 * (TFO) - we could try to be more aggressive and
6104 * retransmitting any data sooner based on when they
6109 tcp_init_transfer(sk, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB);
6110 tp->copied_seq = tp->rcv_nxt;
6113 tcp_set_state(sk, TCP_ESTABLISHED);
6114 sk->sk_state_change(sk);
6116 /* Note, that this wakeup is only for marginal crossed SYN case.
6117 * Passively open sockets are not waked up, because
6118 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
6121 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
6123 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
6124 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
6125 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
6127 if (tp->rx_opt.tstamp_ok)
6128 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
6130 if (!inet_csk(sk)->icsk_ca_ops->cong_control)
6131 tcp_update_pacing_rate(sk);
6133 /* Prevent spurious tcp_cwnd_restart() on first data packet */
6134 tp->lsndtime = tcp_jiffies32;
6136 tcp_initialize_rcv_mss(sk);
6137 tcp_fast_path_on(tp);
6140 case TCP_FIN_WAIT1: {
6143 /* If we enter the TCP_FIN_WAIT1 state and we are a
6144 * Fast Open socket and this is the first acceptable
6145 * ACK we have received, this would have acknowledged
6146 * our SYNACK so stop the SYNACK timer.
6149 /* We no longer need the request sock. */
6150 reqsk_fastopen_remove(sk, req, false);
6153 if (tp->snd_una != tp->write_seq)
6156 tcp_set_state(sk, TCP_FIN_WAIT2);
6157 sk->sk_shutdown |= SEND_SHUTDOWN;
6161 if (!sock_flag(sk, SOCK_DEAD)) {
6162 /* Wake up lingering close() */
6163 sk->sk_state_change(sk);
6167 if (tp->linger2 < 0) {
6169 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6172 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6173 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6174 /* Receive out of order FIN after close() */
6175 if (tp->syn_fastopen && th->fin)
6176 tcp_fastopen_active_disable(sk);
6178 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6182 tmo = tcp_fin_time(sk);
6183 if (tmo > TCP_TIMEWAIT_LEN) {
6184 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
6185 } else if (th->fin || sock_owned_by_user(sk)) {
6186 /* Bad case. We could lose such FIN otherwise.
6187 * It is not a big problem, but it looks confusing
6188 * and not so rare event. We still can lose it now,
6189 * if it spins in bh_lock_sock(), but it is really
6192 inet_csk_reset_keepalive_timer(sk, tmo);
6194 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
6201 if (tp->snd_una == tp->write_seq) {
6202 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
6208 if (tp->snd_una == tp->write_seq) {
6209 tcp_update_metrics(sk);
6216 /* step 6: check the URG bit */
6217 tcp_urg(sk, skb, th);
6219 /* step 7: process the segment text */
6220 switch (sk->sk_state) {
6221 case TCP_CLOSE_WAIT:
6224 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
6229 /* RFC 793 says to queue data in these states,
6230 * RFC 1122 says we MUST send a reset.
6231 * BSD 4.4 also does reset.
6233 if (sk->sk_shutdown & RCV_SHUTDOWN) {
6234 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6235 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6236 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6242 case TCP_ESTABLISHED:
6243 tcp_data_queue(sk, skb);
6248 /* tcp_data could move socket to TIME-WAIT */
6249 if (sk->sk_state != TCP_CLOSE) {
6250 tcp_data_snd_check(sk);
6251 tcp_ack_snd_check(sk);
6260 EXPORT_SYMBOL(tcp_rcv_state_process);
6262 static inline void pr_drop_req(struct request_sock *req, __u16 port, int family)
6264 struct inet_request_sock *ireq = inet_rsk(req);
6266 if (family == AF_INET)
6267 net_dbg_ratelimited("drop open request from %pI4/%u\n",
6268 &ireq->ir_rmt_addr, port);
6269 #if IS_ENABLED(CONFIG_IPV6)
6270 else if (family == AF_INET6)
6271 net_dbg_ratelimited("drop open request from %pI6/%u\n",
6272 &ireq->ir_v6_rmt_addr, port);
6276 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
6278 * If we receive a SYN packet with these bits set, it means a
6279 * network is playing bad games with TOS bits. In order to
6280 * avoid possible false congestion notifications, we disable
6281 * TCP ECN negotiation.
6283 * Exception: tcp_ca wants ECN. This is required for DCTCP
6284 * congestion control: Linux DCTCP asserts ECT on all packets,
6285 * including SYN, which is most optimal solution; however,
6286 * others, such as FreeBSD do not.
6288 static void tcp_ecn_create_request(struct request_sock *req,
6289 const struct sk_buff *skb,
6290 const struct sock *listen_sk,
6291 const struct dst_entry *dst)
6293 const struct tcphdr *th = tcp_hdr(skb);
6294 const struct net *net = sock_net(listen_sk);
6295 bool th_ecn = th->ece && th->cwr;
6302 ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield);
6303 ecn_ok_dst = dst_feature(dst, DST_FEATURE_ECN_MASK);
6304 ecn_ok = net->ipv4.sysctl_tcp_ecn || ecn_ok_dst;
6306 if ((!ect && ecn_ok) || tcp_ca_needs_ecn(listen_sk) ||
6307 (ecn_ok_dst & DST_FEATURE_ECN_CA) ||
6308 tcp_bpf_ca_needs_ecn((struct sock *)req))
6309 inet_rsk(req)->ecn_ok = 1;
6312 static void tcp_openreq_init(struct request_sock *req,
6313 const struct tcp_options_received *rx_opt,
6314 struct sk_buff *skb, const struct sock *sk)
6316 struct inet_request_sock *ireq = inet_rsk(req);
6318 req->rsk_rcv_wnd = 0; /* So that tcp_send_synack() knows! */
6320 tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq;
6321 tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
6322 tcp_rsk(req)->snt_synack = tcp_clock_us();
6323 tcp_rsk(req)->last_oow_ack_time = 0;
6324 req->mss = rx_opt->mss_clamp;
6325 req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0;
6326 ireq->tstamp_ok = rx_opt->tstamp_ok;
6327 ireq->sack_ok = rx_opt->sack_ok;
6328 ireq->snd_wscale = rx_opt->snd_wscale;
6329 ireq->wscale_ok = rx_opt->wscale_ok;
6332 ireq->ir_rmt_port = tcp_hdr(skb)->source;
6333 ireq->ir_num = ntohs(tcp_hdr(skb)->dest);
6334 ireq->ir_mark = inet_request_mark(sk, skb);
6335 #if IS_ENABLED(CONFIG_SMC)
6336 ireq->smc_ok = rx_opt->smc_ok;
6340 struct request_sock *inet_reqsk_alloc(const struct request_sock_ops *ops,
6341 struct sock *sk_listener,
6342 bool attach_listener)
6344 struct request_sock *req = reqsk_alloc(ops, sk_listener,
6348 struct inet_request_sock *ireq = inet_rsk(req);
6350 ireq->ireq_opt = NULL;
6351 #if IS_ENABLED(CONFIG_IPV6)
6352 ireq->pktopts = NULL;
6354 atomic64_set(&ireq->ir_cookie, 0);
6355 ireq->ireq_state = TCP_NEW_SYN_RECV;
6356 write_pnet(&ireq->ireq_net, sock_net(sk_listener));
6357 ireq->ireq_family = sk_listener->sk_family;
6362 EXPORT_SYMBOL(inet_reqsk_alloc);
6365 * Return true if a syncookie should be sent
6367 static bool tcp_syn_flood_action(const struct sock *sk,
6368 const struct sk_buff *skb,
6371 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
6372 const char *msg = "Dropping request";
6373 bool want_cookie = false;
6374 struct net *net = sock_net(sk);
6376 #ifdef CONFIG_SYN_COOKIES
6377 if (net->ipv4.sysctl_tcp_syncookies) {
6378 msg = "Sending cookies";
6380 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES);
6383 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP);
6385 if (!queue->synflood_warned &&
6386 net->ipv4.sysctl_tcp_syncookies != 2 &&
6387 xchg(&queue->synflood_warned, 1) == 0)
6388 net_info_ratelimited("%s: Possible SYN flooding on port %d. %s. Check SNMP counters.\n",
6389 proto, ntohs(tcp_hdr(skb)->dest), msg);
6394 static void tcp_reqsk_record_syn(const struct sock *sk,
6395 struct request_sock *req,
6396 const struct sk_buff *skb)
6398 if (tcp_sk(sk)->save_syn) {
6399 u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb);
6402 copy = kmalloc(len + sizeof(u32), GFP_ATOMIC);
6405 memcpy(©[1], skb_network_header(skb), len);
6406 req->saved_syn = copy;
6411 int tcp_conn_request(struct request_sock_ops *rsk_ops,
6412 const struct tcp_request_sock_ops *af_ops,
6413 struct sock *sk, struct sk_buff *skb)
6415 struct tcp_fastopen_cookie foc = { .len = -1 };
6416 __u32 isn = TCP_SKB_CB(skb)->tcp_tw_isn;
6417 struct tcp_options_received tmp_opt;
6418 struct tcp_sock *tp = tcp_sk(sk);
6419 struct net *net = sock_net(sk);
6420 struct sock *fastopen_sk = NULL;
6421 struct request_sock *req;
6422 bool want_cookie = false;
6423 struct dst_entry *dst;
6426 /* TW buckets are converted to open requests without
6427 * limitations, they conserve resources and peer is
6428 * evidently real one.
6430 if ((net->ipv4.sysctl_tcp_syncookies == 2 ||
6431 inet_csk_reqsk_queue_is_full(sk)) && !isn) {
6432 want_cookie = tcp_syn_flood_action(sk, skb, rsk_ops->slab_name);
6437 if (sk_acceptq_is_full(sk)) {
6438 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
6442 req = inet_reqsk_alloc(rsk_ops, sk, !want_cookie);
6446 tcp_rsk(req)->af_specific = af_ops;
6447 tcp_rsk(req)->ts_off = 0;
6449 tcp_clear_options(&tmp_opt);
6450 tmp_opt.mss_clamp = af_ops->mss_clamp;
6451 tmp_opt.user_mss = tp->rx_opt.user_mss;
6452 tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0,
6453 want_cookie ? NULL : &foc);
6455 if (want_cookie && !tmp_opt.saw_tstamp)
6456 tcp_clear_options(&tmp_opt);
6458 if (IS_ENABLED(CONFIG_SMC) && want_cookie)
6461 tmp_opt.tstamp_ok = tmp_opt.saw_tstamp;
6462 tcp_openreq_init(req, &tmp_opt, skb, sk);
6463 inet_rsk(req)->no_srccheck = inet_sk(sk)->transparent;
6465 /* Note: tcp_v6_init_req() might override ir_iif for link locals */
6466 inet_rsk(req)->ir_iif = inet_request_bound_dev_if(sk, skb);
6468 af_ops->init_req(req, sk, skb);
6470 if (security_inet_conn_request(sk, skb, req))
6473 if (tmp_opt.tstamp_ok)
6474 tcp_rsk(req)->ts_off = af_ops->init_ts_off(net, skb);
6476 dst = af_ops->route_req(sk, &fl, req);
6480 if (!want_cookie && !isn) {
6481 /* Kill the following clause, if you dislike this way. */
6482 if (!net->ipv4.sysctl_tcp_syncookies &&
6483 (net->ipv4.sysctl_max_syn_backlog - inet_csk_reqsk_queue_len(sk) <
6484 (net->ipv4.sysctl_max_syn_backlog >> 2)) &&
6485 !tcp_peer_is_proven(req, dst)) {
6486 /* Without syncookies last quarter of
6487 * backlog is filled with destinations,
6488 * proven to be alive.
6489 * It means that we continue to communicate
6490 * to destinations, already remembered
6491 * to the moment of synflood.
6493 pr_drop_req(req, ntohs(tcp_hdr(skb)->source),
6495 goto drop_and_release;
6498 isn = af_ops->init_seq(skb);
6501 tcp_ecn_create_request(req, skb, sk, dst);
6504 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss);
6505 req->cookie_ts = tmp_opt.tstamp_ok;
6506 if (!tmp_opt.tstamp_ok)
6507 inet_rsk(req)->ecn_ok = 0;
6510 tcp_rsk(req)->snt_isn = isn;
6511 tcp_rsk(req)->txhash = net_tx_rndhash();
6512 tcp_openreq_init_rwin(req, sk, dst);
6513 sk_rx_queue_set(req_to_sk(req), skb);
6515 tcp_reqsk_record_syn(sk, req, skb);
6516 fastopen_sk = tcp_try_fastopen(sk, skb, req, &foc, dst);
6519 af_ops->send_synack(fastopen_sk, dst, &fl, req,
6520 &foc, TCP_SYNACK_FASTOPEN);
6521 /* Add the child socket directly into the accept queue */
6522 inet_csk_reqsk_queue_add(sk, req, fastopen_sk);
6523 sk->sk_data_ready(sk);
6524 bh_unlock_sock(fastopen_sk);
6525 sock_put(fastopen_sk);
6527 tcp_rsk(req)->tfo_listener = false;
6529 inet_csk_reqsk_queue_hash_add(sk, req,
6530 tcp_timeout_init((struct sock *)req));
6531 af_ops->send_synack(sk, dst, &fl, req, &foc,
6532 !want_cookie ? TCP_SYNACK_NORMAL :
6550 EXPORT_SYMBOL(tcp_conn_request);