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>
82 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
84 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
85 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
86 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
87 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
88 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
89 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
90 #define FLAG_ECE 0x40 /* ECE in this ACK */
91 #define FLAG_LOST_RETRANS 0x80 /* This ACK marks some retransmission lost */
92 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
93 #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
94 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
95 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
96 #define FLAG_SET_XMIT_TIMER 0x1000 /* Set TLP or RTO timer */
97 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
98 #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
99 #define FLAG_NO_CHALLENGE_ACK 0x8000 /* do not call tcp_send_challenge_ack() */
101 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
102 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
103 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE|FLAG_DSACKING_ACK)
104 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
106 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
107 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
109 #define REXMIT_NONE 0 /* no loss recovery to do */
110 #define REXMIT_LOST 1 /* retransmit packets marked lost */
111 #define REXMIT_NEW 2 /* FRTO-style transmit of unsent/new packets */
113 static void tcp_gro_dev_warn(struct sock *sk, const struct sk_buff *skb,
116 static bool __once __read_mostly;
119 struct net_device *dev;
124 dev = dev_get_by_index_rcu(sock_net(sk), skb->skb_iif);
125 if (!dev || len >= dev->mtu)
126 pr_warn("%s: Driver has suspect GRO implementation, TCP performance may be compromised.\n",
127 dev ? dev->name : "Unknown driver");
132 /* Adapt the MSS value used to make delayed ack decision to the
135 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
137 struct inet_connection_sock *icsk = inet_csk(sk);
138 const unsigned int lss = icsk->icsk_ack.last_seg_size;
141 icsk->icsk_ack.last_seg_size = 0;
143 /* skb->len may jitter because of SACKs, even if peer
144 * sends good full-sized frames.
146 len = skb_shinfo(skb)->gso_size ? : skb->len;
147 if (len >= icsk->icsk_ack.rcv_mss) {
148 icsk->icsk_ack.rcv_mss = min_t(unsigned int, len,
150 /* Account for possibly-removed options */
151 if (unlikely(len > icsk->icsk_ack.rcv_mss +
152 MAX_TCP_OPTION_SPACE))
153 tcp_gro_dev_warn(sk, skb, len);
155 /* Otherwise, we make more careful check taking into account,
156 * that SACKs block is variable.
158 * "len" is invariant segment length, including TCP header.
160 len += skb->data - skb_transport_header(skb);
161 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
162 /* If PSH is not set, packet should be
163 * full sized, provided peer TCP is not badly broken.
164 * This observation (if it is correct 8)) allows
165 * to handle super-low mtu links fairly.
167 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
168 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
169 /* Subtract also invariant (if peer is RFC compliant),
170 * tcp header plus fixed timestamp option length.
171 * Resulting "len" is MSS free of SACK jitter.
173 len -= tcp_sk(sk)->tcp_header_len;
174 icsk->icsk_ack.last_seg_size = len;
176 icsk->icsk_ack.rcv_mss = len;
180 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
181 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
182 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
186 static void tcp_incr_quickack(struct sock *sk)
188 struct inet_connection_sock *icsk = inet_csk(sk);
189 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
193 if (quickacks > icsk->icsk_ack.quick)
194 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
197 static void tcp_enter_quickack_mode(struct sock *sk)
199 struct inet_connection_sock *icsk = inet_csk(sk);
200 tcp_incr_quickack(sk);
201 icsk->icsk_ack.pingpong = 0;
202 icsk->icsk_ack.ato = TCP_ATO_MIN;
205 /* Send ACKs quickly, if "quick" count is not exhausted
206 * and the session is not interactive.
209 static bool tcp_in_quickack_mode(struct sock *sk)
211 const struct inet_connection_sock *icsk = inet_csk(sk);
212 const struct dst_entry *dst = __sk_dst_get(sk);
214 return (dst && dst_metric(dst, RTAX_QUICKACK)) ||
215 (icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong);
218 static void tcp_ecn_queue_cwr(struct tcp_sock *tp)
220 if (tp->ecn_flags & TCP_ECN_OK)
221 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
224 static void tcp_ecn_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb)
226 if (tcp_hdr(skb)->cwr)
227 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
230 static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp)
232 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
235 static void __tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
237 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
238 case INET_ECN_NOT_ECT:
239 /* Funny extension: if ECT is not set on a segment,
240 * and we already seen ECT on a previous segment,
241 * it is probably a retransmit.
243 if (tp->ecn_flags & TCP_ECN_SEEN)
244 tcp_enter_quickack_mode((struct sock *)tp);
247 if (tcp_ca_needs_ecn((struct sock *)tp))
248 tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_IS_CE);
250 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
251 /* Better not delay acks, sender can have a very low cwnd */
252 tcp_enter_quickack_mode((struct sock *)tp);
253 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
255 tp->ecn_flags |= TCP_ECN_SEEN;
258 if (tcp_ca_needs_ecn((struct sock *)tp))
259 tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_NO_CE);
260 tp->ecn_flags |= TCP_ECN_SEEN;
265 static void tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
267 if (tp->ecn_flags & TCP_ECN_OK)
268 __tcp_ecn_check_ce(tp, skb);
271 static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
273 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
274 tp->ecn_flags &= ~TCP_ECN_OK;
277 static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
279 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
280 tp->ecn_flags &= ~TCP_ECN_OK;
283 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
285 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
290 /* Buffer size and advertised window tuning.
292 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
295 static void tcp_sndbuf_expand(struct sock *sk)
297 const struct tcp_sock *tp = tcp_sk(sk);
298 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
302 /* Worst case is non GSO/TSO : each frame consumes one skb
303 * and skb->head is kmalloced using power of two area of memory
305 per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
307 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
309 per_mss = roundup_pow_of_two(per_mss) +
310 SKB_DATA_ALIGN(sizeof(struct sk_buff));
312 nr_segs = max_t(u32, TCP_INIT_CWND, tp->snd_cwnd);
313 nr_segs = max_t(u32, nr_segs, tp->reordering + 1);
315 /* Fast Recovery (RFC 5681 3.2) :
316 * Cubic needs 1.7 factor, rounded to 2 to include
317 * extra cushion (application might react slowly to POLLOUT)
319 sndmem = ca_ops->sndbuf_expand ? ca_ops->sndbuf_expand(sk) : 2;
320 sndmem *= nr_segs * per_mss;
322 if (sk->sk_sndbuf < sndmem)
323 sk->sk_sndbuf = min(sndmem, sock_net(sk)->ipv4.sysctl_tcp_wmem[2]);
326 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
328 * All tcp_full_space() is split to two parts: "network" buffer, allocated
329 * forward and advertised in receiver window (tp->rcv_wnd) and
330 * "application buffer", required to isolate scheduling/application
331 * latencies from network.
332 * window_clamp is maximal advertised window. It can be less than
333 * tcp_full_space(), in this case tcp_full_space() - window_clamp
334 * is reserved for "application" buffer. The less window_clamp is
335 * the smoother our behaviour from viewpoint of network, but the lower
336 * throughput and the higher sensitivity of the connection to losses. 8)
338 * rcv_ssthresh is more strict window_clamp used at "slow start"
339 * phase to predict further behaviour of this connection.
340 * It is used for two goals:
341 * - to enforce header prediction at sender, even when application
342 * requires some significant "application buffer". It is check #1.
343 * - to prevent pruning of receive queue because of misprediction
344 * of receiver window. Check #2.
346 * The scheme does not work when sender sends good segments opening
347 * window and then starts to feed us spaghetti. But it should work
348 * in common situations. Otherwise, we have to rely on queue collapsing.
351 /* Slow part of check#2. */
352 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
354 struct tcp_sock *tp = tcp_sk(sk);
356 int truesize = tcp_win_from_space(sk, skb->truesize) >> 1;
357 int window = tcp_win_from_space(sk, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]) >> 1;
359 while (tp->rcv_ssthresh <= window) {
360 if (truesize <= skb->len)
361 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
369 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
371 struct tcp_sock *tp = tcp_sk(sk);
374 if (tp->rcv_ssthresh < tp->window_clamp &&
375 (int)tp->rcv_ssthresh < tcp_space(sk) &&
376 !tcp_under_memory_pressure(sk)) {
379 /* Check #2. Increase window, if skb with such overhead
380 * will fit to rcvbuf in future.
382 if (tcp_win_from_space(sk, skb->truesize) <= skb->len)
383 incr = 2 * tp->advmss;
385 incr = __tcp_grow_window(sk, skb);
388 incr = max_t(int, incr, 2 * skb->len);
389 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
391 inet_csk(sk)->icsk_ack.quick |= 1;
396 /* 3. Tuning rcvbuf, when connection enters established state. */
397 static void tcp_fixup_rcvbuf(struct sock *sk)
399 u32 mss = tcp_sk(sk)->advmss;
402 rcvmem = 2 * SKB_TRUESIZE(mss + MAX_TCP_HEADER) *
403 tcp_default_init_rwnd(mss);
405 /* Dynamic Right Sizing (DRS) has 2 to 3 RTT latency
406 * Allow enough cushion so that sender is not limited by our window
408 if (sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf)
411 if (sk->sk_rcvbuf < rcvmem)
412 sk->sk_rcvbuf = min(rcvmem, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
415 /* 4. Try to fixup all. It is made immediately after connection enters
418 void tcp_init_buffer_space(struct sock *sk)
420 int tcp_app_win = sock_net(sk)->ipv4.sysctl_tcp_app_win;
421 struct tcp_sock *tp = tcp_sk(sk);
424 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
425 tcp_fixup_rcvbuf(sk);
426 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
427 tcp_sndbuf_expand(sk);
429 tp->rcvq_space.space = tp->rcv_wnd;
430 tcp_mstamp_refresh(tp);
431 tp->rcvq_space.time = tp->tcp_mstamp;
432 tp->rcvq_space.seq = tp->copied_seq;
434 maxwin = tcp_full_space(sk);
436 if (tp->window_clamp >= maxwin) {
437 tp->window_clamp = maxwin;
439 if (tcp_app_win && maxwin > 4 * tp->advmss)
440 tp->window_clamp = max(maxwin -
441 (maxwin >> tcp_app_win),
445 /* Force reservation of one segment. */
447 tp->window_clamp > 2 * tp->advmss &&
448 tp->window_clamp + tp->advmss > maxwin)
449 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
451 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
452 tp->snd_cwnd_stamp = tcp_jiffies32;
455 /* 5. Recalculate window clamp after socket hit its memory bounds. */
456 static void tcp_clamp_window(struct sock *sk)
458 struct tcp_sock *tp = tcp_sk(sk);
459 struct inet_connection_sock *icsk = inet_csk(sk);
460 struct net *net = sock_net(sk);
462 icsk->icsk_ack.quick = 0;
464 if (sk->sk_rcvbuf < net->ipv4.sysctl_tcp_rmem[2] &&
465 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
466 !tcp_under_memory_pressure(sk) &&
467 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
468 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
469 net->ipv4.sysctl_tcp_rmem[2]);
471 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
472 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
475 /* Initialize RCV_MSS value.
476 * RCV_MSS is an our guess about MSS used by the peer.
477 * We haven't any direct information about the MSS.
478 * It's better to underestimate the RCV_MSS rather than overestimate.
479 * Overestimations make us ACKing less frequently than needed.
480 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
482 void tcp_initialize_rcv_mss(struct sock *sk)
484 const struct tcp_sock *tp = tcp_sk(sk);
485 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
487 hint = min(hint, tp->rcv_wnd / 2);
488 hint = min(hint, TCP_MSS_DEFAULT);
489 hint = max(hint, TCP_MIN_MSS);
491 inet_csk(sk)->icsk_ack.rcv_mss = hint;
493 EXPORT_SYMBOL(tcp_initialize_rcv_mss);
495 /* Receiver "autotuning" code.
497 * The algorithm for RTT estimation w/o timestamps is based on
498 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
499 * <http://public.lanl.gov/radiant/pubs.html#DRS>
501 * More detail on this code can be found at
502 * <http://staff.psc.edu/jheffner/>,
503 * though this reference is out of date. A new paper
506 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
508 u32 new_sample = tp->rcv_rtt_est.rtt_us;
511 if (new_sample != 0) {
512 /* If we sample in larger samples in the non-timestamp
513 * case, we could grossly overestimate the RTT especially
514 * with chatty applications or bulk transfer apps which
515 * are stalled on filesystem I/O.
517 * Also, since we are only going for a minimum in the
518 * non-timestamp case, we do not smooth things out
519 * else with timestamps disabled convergence takes too
523 m -= (new_sample >> 3);
531 /* No previous measure. */
535 tp->rcv_rtt_est.rtt_us = new_sample;
538 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
542 if (tp->rcv_rtt_est.time == 0)
544 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
546 delta_us = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcv_rtt_est.time);
549 tcp_rcv_rtt_update(tp, delta_us, 1);
552 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
553 tp->rcv_rtt_est.time = tp->tcp_mstamp;
556 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
557 const struct sk_buff *skb)
559 struct tcp_sock *tp = tcp_sk(sk);
561 if (tp->rx_opt.rcv_tsecr &&
562 (TCP_SKB_CB(skb)->end_seq -
563 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss)) {
564 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr;
569 delta_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
570 tcp_rcv_rtt_update(tp, delta_us, 0);
575 * This function should be called every time data is copied to user space.
576 * It calculates the appropriate TCP receive buffer space.
578 void tcp_rcv_space_adjust(struct sock *sk)
580 struct tcp_sock *tp = tcp_sk(sk);
584 tcp_mstamp_refresh(tp);
585 time = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcvq_space.time);
586 if (time < (tp->rcv_rtt_est.rtt_us >> 3) || tp->rcv_rtt_est.rtt_us == 0)
589 /* Number of bytes copied to user in last RTT */
590 copied = tp->copied_seq - tp->rcvq_space.seq;
591 if (copied <= tp->rcvq_space.space)
595 * copied = bytes received in previous RTT, our base window
596 * To cope with packet losses, we need a 2x factor
597 * To cope with slow start, and sender growing its cwin by 100 %
598 * every RTT, we need a 4x factor, because the ACK we are sending
599 * now is for the next RTT, not the current one :
600 * <prev RTT . ><current RTT .. ><next RTT .... >
603 if (sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf &&
604 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
605 int rcvwin, rcvmem, rcvbuf;
607 /* minimal window to cope with packet losses, assuming
608 * steady state. Add some cushion because of small variations.
610 rcvwin = (copied << 1) + 16 * tp->advmss;
612 /* If rate increased by 25%,
613 * assume slow start, rcvwin = 3 * copied
614 * If rate increased by 50%,
615 * assume sender can use 2x growth, rcvwin = 4 * copied
618 tp->rcvq_space.space + (tp->rcvq_space.space >> 2)) {
620 tp->rcvq_space.space + (tp->rcvq_space.space >> 1))
623 rcvwin += (rcvwin >> 1);
626 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
627 while (tcp_win_from_space(sk, rcvmem) < tp->advmss)
630 rcvbuf = min(rcvwin / tp->advmss * rcvmem,
631 sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
632 if (rcvbuf > sk->sk_rcvbuf) {
633 sk->sk_rcvbuf = rcvbuf;
635 /* Make the window clamp follow along. */
636 tp->window_clamp = rcvwin;
639 tp->rcvq_space.space = copied;
642 tp->rcvq_space.seq = tp->copied_seq;
643 tp->rcvq_space.time = tp->tcp_mstamp;
646 /* There is something which you must keep in mind when you analyze the
647 * behavior of the tp->ato delayed ack timeout interval. When a
648 * connection starts up, we want to ack as quickly as possible. The
649 * problem is that "good" TCP's do slow start at the beginning of data
650 * transmission. The means that until we send the first few ACK's the
651 * sender will sit on his end and only queue most of his data, because
652 * he can only send snd_cwnd unacked packets at any given time. For
653 * each ACK we send, he increments snd_cwnd and transmits more of his
656 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
658 struct tcp_sock *tp = tcp_sk(sk);
659 struct inet_connection_sock *icsk = inet_csk(sk);
662 inet_csk_schedule_ack(sk);
664 tcp_measure_rcv_mss(sk, skb);
666 tcp_rcv_rtt_measure(tp);
670 if (!icsk->icsk_ack.ato) {
671 /* The _first_ data packet received, initialize
672 * delayed ACK engine.
674 tcp_incr_quickack(sk);
675 icsk->icsk_ack.ato = TCP_ATO_MIN;
677 int m = now - icsk->icsk_ack.lrcvtime;
679 if (m <= TCP_ATO_MIN / 2) {
680 /* The fastest case is the first. */
681 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
682 } else if (m < icsk->icsk_ack.ato) {
683 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
684 if (icsk->icsk_ack.ato > icsk->icsk_rto)
685 icsk->icsk_ack.ato = icsk->icsk_rto;
686 } else if (m > icsk->icsk_rto) {
687 /* Too long gap. Apparently sender failed to
688 * restart window, so that we send ACKs quickly.
690 tcp_incr_quickack(sk);
694 icsk->icsk_ack.lrcvtime = now;
696 tcp_ecn_check_ce(tp, skb);
699 tcp_grow_window(sk, skb);
702 /* Called to compute a smoothed rtt estimate. The data fed to this
703 * routine either comes from timestamps, or from segments that were
704 * known _not_ to have been retransmitted [see Karn/Partridge
705 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
706 * piece by Van Jacobson.
707 * NOTE: the next three routines used to be one big routine.
708 * To save cycles in the RFC 1323 implementation it was better to break
709 * it up into three procedures. -- erics
711 static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
713 struct tcp_sock *tp = tcp_sk(sk);
714 long m = mrtt_us; /* RTT */
715 u32 srtt = tp->srtt_us;
717 /* The following amusing code comes from Jacobson's
718 * article in SIGCOMM '88. Note that rtt and mdev
719 * are scaled versions of rtt and mean deviation.
720 * This is designed to be as fast as possible
721 * m stands for "measurement".
723 * On a 1990 paper the rto value is changed to:
724 * RTO = rtt + 4 * mdev
726 * Funny. This algorithm seems to be very broken.
727 * These formulae increase RTO, when it should be decreased, increase
728 * too slowly, when it should be increased quickly, decrease too quickly
729 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
730 * does not matter how to _calculate_ it. Seems, it was trap
731 * that VJ failed to avoid. 8)
734 m -= (srtt >> 3); /* m is now error in rtt est */
735 srtt += m; /* rtt = 7/8 rtt + 1/8 new */
737 m = -m; /* m is now abs(error) */
738 m -= (tp->mdev_us >> 2); /* similar update on mdev */
739 /* This is similar to one of Eifel findings.
740 * Eifel blocks mdev updates when rtt decreases.
741 * This solution is a bit different: we use finer gain
742 * for mdev in this case (alpha*beta).
743 * Like Eifel it also prevents growth of rto,
744 * but also it limits too fast rto decreases,
745 * happening in pure Eifel.
750 m -= (tp->mdev_us >> 2); /* similar update on mdev */
752 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */
753 if (tp->mdev_us > tp->mdev_max_us) {
754 tp->mdev_max_us = tp->mdev_us;
755 if (tp->mdev_max_us > tp->rttvar_us)
756 tp->rttvar_us = tp->mdev_max_us;
758 if (after(tp->snd_una, tp->rtt_seq)) {
759 if (tp->mdev_max_us < tp->rttvar_us)
760 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
761 tp->rtt_seq = tp->snd_nxt;
762 tp->mdev_max_us = tcp_rto_min_us(sk);
765 /* no previous measure. */
766 srtt = m << 3; /* take the measured time to be rtt */
767 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */
768 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
769 tp->mdev_max_us = tp->rttvar_us;
770 tp->rtt_seq = tp->snd_nxt;
772 tp->srtt_us = max(1U, srtt);
775 static void tcp_update_pacing_rate(struct sock *sk)
777 const struct tcp_sock *tp = tcp_sk(sk);
780 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
781 rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3);
783 /* current rate is (cwnd * mss) / srtt
784 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
785 * In Congestion Avoidance phase, set it to 120 % the current rate.
787 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
788 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
789 * end of slow start and should slow down.
791 if (tp->snd_cwnd < tp->snd_ssthresh / 2)
792 rate *= sock_net(sk)->ipv4.sysctl_tcp_pacing_ss_ratio;
794 rate *= sock_net(sk)->ipv4.sysctl_tcp_pacing_ca_ratio;
796 rate *= max(tp->snd_cwnd, tp->packets_out);
798 if (likely(tp->srtt_us))
799 do_div(rate, tp->srtt_us);
801 /* WRITE_ONCE() is needed because sch_fq fetches sk_pacing_rate
802 * without any lock. We want to make sure compiler wont store
803 * intermediate values in this location.
805 WRITE_ONCE(sk->sk_pacing_rate, min_t(u64, rate,
806 sk->sk_max_pacing_rate));
809 /* Calculate rto without backoff. This is the second half of Van Jacobson's
810 * routine referred to above.
812 static void tcp_set_rto(struct sock *sk)
814 const struct tcp_sock *tp = tcp_sk(sk);
815 /* Old crap is replaced with new one. 8)
818 * 1. If rtt variance happened to be less 50msec, it is hallucination.
819 * It cannot be less due to utterly erratic ACK generation made
820 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
821 * to do with delayed acks, because at cwnd>2 true delack timeout
822 * is invisible. Actually, Linux-2.4 also generates erratic
823 * ACKs in some circumstances.
825 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
827 /* 2. Fixups made earlier cannot be right.
828 * If we do not estimate RTO correctly without them,
829 * all the algo is pure shit and should be replaced
830 * with correct one. It is exactly, which we pretend to do.
833 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
834 * guarantees that rto is higher.
839 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
841 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
844 cwnd = TCP_INIT_CWND;
845 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
848 /* Take a notice that peer is sending D-SACKs */
849 static void tcp_dsack_seen(struct tcp_sock *tp)
851 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
852 tp->rack.dsack_seen = 1;
855 /* It's reordering when higher sequence was delivered (i.e. sacked) before
856 * some lower never-retransmitted sequence ("low_seq"). The maximum reordering
857 * distance is approximated in full-mss packet distance ("reordering").
859 static void tcp_check_sack_reordering(struct sock *sk, const u32 low_seq,
862 struct tcp_sock *tp = tcp_sk(sk);
863 const u32 mss = tp->mss_cache;
866 fack = tcp_highest_sack_seq(tp);
867 if (!before(low_seq, fack))
870 metric = fack - low_seq;
871 if ((metric > tp->reordering * mss) && mss) {
872 #if FASTRETRANS_DEBUG > 1
873 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
874 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
878 tp->undo_marker ? tp->undo_retrans : 0);
880 tp->reordering = min_t(u32, (metric + mss - 1) / mss,
881 sock_net(sk)->ipv4.sysctl_tcp_max_reordering);
885 /* This exciting event is worth to be remembered. 8) */
886 NET_INC_STATS(sock_net(sk),
887 ts ? LINUX_MIB_TCPTSREORDER : LINUX_MIB_TCPSACKREORDER);
890 /* This must be called before lost_out is incremented */
891 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
893 if (!tp->retransmit_skb_hint ||
894 before(TCP_SKB_CB(skb)->seq,
895 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
896 tp->retransmit_skb_hint = skb;
899 /* Sum the number of packets on the wire we have marked as lost.
900 * There are two cases we care about here:
901 * a) Packet hasn't been marked lost (nor retransmitted),
902 * and this is the first loss.
903 * b) Packet has been marked both lost and retransmitted,
904 * and this means we think it was lost again.
906 static void tcp_sum_lost(struct tcp_sock *tp, struct sk_buff *skb)
908 __u8 sacked = TCP_SKB_CB(skb)->sacked;
910 if (!(sacked & TCPCB_LOST) ||
911 ((sacked & TCPCB_LOST) && (sacked & TCPCB_SACKED_RETRANS)))
912 tp->lost += tcp_skb_pcount(skb);
915 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
917 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
918 tcp_verify_retransmit_hint(tp, skb);
920 tp->lost_out += tcp_skb_pcount(skb);
921 tcp_sum_lost(tp, skb);
922 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
926 void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp, struct sk_buff *skb)
928 tcp_verify_retransmit_hint(tp, skb);
930 tcp_sum_lost(tp, skb);
931 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
932 tp->lost_out += tcp_skb_pcount(skb);
933 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
937 /* This procedure tags the retransmission queue when SACKs arrive.
939 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
940 * Packets in queue with these bits set are counted in variables
941 * sacked_out, retrans_out and lost_out, correspondingly.
943 * Valid combinations are:
944 * Tag InFlight Description
945 * 0 1 - orig segment is in flight.
946 * S 0 - nothing flies, orig reached receiver.
947 * L 0 - nothing flies, orig lost by net.
948 * R 2 - both orig and retransmit are in flight.
949 * L|R 1 - orig is lost, retransmit is in flight.
950 * S|R 1 - orig reached receiver, retrans is still in flight.
951 * (L|S|R is logically valid, it could occur when L|R is sacked,
952 * but it is equivalent to plain S and code short-curcuits it to S.
953 * L|S is logically invalid, it would mean -1 packet in flight 8))
955 * These 6 states form finite state machine, controlled by the following events:
956 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
957 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
958 * 3. Loss detection event of two flavors:
959 * A. Scoreboard estimator decided the packet is lost.
960 * A'. Reno "three dupacks" marks head of queue lost.
961 * B. SACK arrives sacking SND.NXT at the moment, when the
962 * segment was retransmitted.
963 * 4. D-SACK added new rule: D-SACK changes any tag to S.
965 * It is pleasant to note, that state diagram turns out to be commutative,
966 * so that we are allowed not to be bothered by order of our actions,
967 * when multiple events arrive simultaneously. (see the function below).
969 * Reordering detection.
970 * --------------------
971 * Reordering metric is maximal distance, which a packet can be displaced
972 * in packet stream. With SACKs we can estimate it:
974 * 1. SACK fills old hole and the corresponding segment was not
975 * ever retransmitted -> reordering. Alas, we cannot use it
976 * when segment was retransmitted.
977 * 2. The last flaw is solved with D-SACK. D-SACK arrives
978 * for retransmitted and already SACKed segment -> reordering..
979 * Both of these heuristics are not used in Loss state, when we cannot
980 * account for retransmits accurately.
982 * SACK block validation.
983 * ----------------------
985 * SACK block range validation checks that the received SACK block fits to
986 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
987 * Note that SND.UNA is not included to the range though being valid because
988 * it means that the receiver is rather inconsistent with itself reporting
989 * SACK reneging when it should advance SND.UNA. Such SACK block this is
990 * perfectly valid, however, in light of RFC2018 which explicitly states
991 * that "SACK block MUST reflect the newest segment. Even if the newest
992 * segment is going to be discarded ...", not that it looks very clever
993 * in case of head skb. Due to potentional receiver driven attacks, we
994 * choose to avoid immediate execution of a walk in write queue due to
995 * reneging and defer head skb's loss recovery to standard loss recovery
996 * procedure that will eventually trigger (nothing forbids us doing this).
998 * Implements also blockage to start_seq wrap-around. Problem lies in the
999 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1000 * there's no guarantee that it will be before snd_nxt (n). The problem
1001 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1004 * <- outs wnd -> <- wrapzone ->
1005 * u e n u_w e_w s n_w
1007 * |<------------+------+----- TCP seqno space --------------+---------->|
1008 * ...-- <2^31 ->| |<--------...
1009 * ...---- >2^31 ------>| |<--------...
1011 * Current code wouldn't be vulnerable but it's better still to discard such
1012 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1013 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1014 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1015 * equal to the ideal case (infinite seqno space without wrap caused issues).
1017 * With D-SACK the lower bound is extended to cover sequence space below
1018 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1019 * again, D-SACK block must not to go across snd_una (for the same reason as
1020 * for the normal SACK blocks, explained above). But there all simplicity
1021 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1022 * fully below undo_marker they do not affect behavior in anyway and can
1023 * therefore be safely ignored. In rare cases (which are more or less
1024 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1025 * fragmentation and packet reordering past skb's retransmission. To consider
1026 * them correctly, the acceptable range must be extended even more though
1027 * the exact amount is rather hard to quantify. However, tp->max_window can
1028 * be used as an exaggerated estimate.
1030 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
1031 u32 start_seq, u32 end_seq)
1033 /* Too far in future, or reversed (interpretation is ambiguous) */
1034 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1037 /* Nasty start_seq wrap-around check (see comments above) */
1038 if (!before(start_seq, tp->snd_nxt))
1041 /* In outstanding window? ...This is valid exit for D-SACKs too.
1042 * start_seq == snd_una is non-sensical (see comments above)
1044 if (after(start_seq, tp->snd_una))
1047 if (!is_dsack || !tp->undo_marker)
1050 /* ...Then it's D-SACK, and must reside below snd_una completely */
1051 if (after(end_seq, tp->snd_una))
1054 if (!before(start_seq, tp->undo_marker))
1058 if (!after(end_seq, tp->undo_marker))
1061 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1062 * start_seq < undo_marker and end_seq >= undo_marker.
1064 return !before(start_seq, end_seq - tp->max_window);
1067 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1068 struct tcp_sack_block_wire *sp, int num_sacks,
1071 struct tcp_sock *tp = tcp_sk(sk);
1072 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1073 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1074 bool dup_sack = false;
1076 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1079 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1080 } else if (num_sacks > 1) {
1081 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1082 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1084 if (!after(end_seq_0, end_seq_1) &&
1085 !before(start_seq_0, start_seq_1)) {
1088 NET_INC_STATS(sock_net(sk),
1089 LINUX_MIB_TCPDSACKOFORECV);
1093 /* D-SACK for already forgotten data... Do dumb counting. */
1094 if (dup_sack && tp->undo_marker && tp->undo_retrans > 0 &&
1095 !after(end_seq_0, prior_snd_una) &&
1096 after(end_seq_0, tp->undo_marker))
1102 struct tcp_sacktag_state {
1104 /* Timestamps for earliest and latest never-retransmitted segment
1105 * that was SACKed. RTO needs the earliest RTT to stay conservative,
1106 * but congestion control should still get an accurate delay signal.
1110 struct rate_sample *rate;
1112 unsigned int mss_now;
1115 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1116 * the incoming SACK may not exactly match but we can find smaller MSS
1117 * aligned portion of it that matches. Therefore we might need to fragment
1118 * which may fail and creates some hassle (caller must handle error case
1121 * FIXME: this could be merged to shift decision code
1123 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1124 u32 start_seq, u32 end_seq)
1128 unsigned int pkt_len;
1131 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1132 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1134 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1135 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1136 mss = tcp_skb_mss(skb);
1137 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1140 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1144 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1149 /* Round if necessary so that SACKs cover only full MSSes
1150 * and/or the remaining small portion (if present)
1152 if (pkt_len > mss) {
1153 unsigned int new_len = (pkt_len / mss) * mss;
1154 if (!in_sack && new_len < pkt_len)
1159 if (pkt_len >= skb->len && !in_sack)
1162 err = tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
1163 pkt_len, mss, GFP_ATOMIC);
1171 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1172 static u8 tcp_sacktag_one(struct sock *sk,
1173 struct tcp_sacktag_state *state, u8 sacked,
1174 u32 start_seq, u32 end_seq,
1175 int dup_sack, int pcount,
1178 struct tcp_sock *tp = tcp_sk(sk);
1180 /* Account D-SACK for retransmitted packet. */
1181 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1182 if (tp->undo_marker && tp->undo_retrans > 0 &&
1183 after(end_seq, tp->undo_marker))
1185 if ((sacked & TCPCB_SACKED_ACKED) &&
1186 before(start_seq, state->reord))
1187 state->reord = start_seq;
1190 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1191 if (!after(end_seq, tp->snd_una))
1194 if (!(sacked & TCPCB_SACKED_ACKED)) {
1195 tcp_rack_advance(tp, sacked, end_seq, xmit_time);
1197 if (sacked & TCPCB_SACKED_RETRANS) {
1198 /* If the segment is not tagged as lost,
1199 * we do not clear RETRANS, believing
1200 * that retransmission is still in flight.
1202 if (sacked & TCPCB_LOST) {
1203 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1204 tp->lost_out -= pcount;
1205 tp->retrans_out -= pcount;
1208 if (!(sacked & TCPCB_RETRANS)) {
1209 /* New sack for not retransmitted frame,
1210 * which was in hole. It is reordering.
1212 if (before(start_seq,
1213 tcp_highest_sack_seq(tp)) &&
1214 before(start_seq, state->reord))
1215 state->reord = start_seq;
1217 if (!after(end_seq, tp->high_seq))
1218 state->flag |= FLAG_ORIG_SACK_ACKED;
1219 if (state->first_sackt == 0)
1220 state->first_sackt = xmit_time;
1221 state->last_sackt = xmit_time;
1224 if (sacked & TCPCB_LOST) {
1225 sacked &= ~TCPCB_LOST;
1226 tp->lost_out -= pcount;
1230 sacked |= TCPCB_SACKED_ACKED;
1231 state->flag |= FLAG_DATA_SACKED;
1232 tp->sacked_out += pcount;
1233 tp->delivered += pcount; /* Out-of-order packets delivered */
1235 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1236 if (tp->lost_skb_hint &&
1237 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1238 tp->lost_cnt_hint += pcount;
1241 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1242 * frames and clear it. undo_retrans is decreased above, L|R frames
1243 * are accounted above as well.
1245 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1246 sacked &= ~TCPCB_SACKED_RETRANS;
1247 tp->retrans_out -= pcount;
1253 /* Shift newly-SACKed bytes from this skb to the immediately previous
1254 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1256 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *prev,
1257 struct sk_buff *skb,
1258 struct tcp_sacktag_state *state,
1259 unsigned int pcount, int shifted, int mss,
1262 struct tcp_sock *tp = tcp_sk(sk);
1263 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1264 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
1268 /* Adjust counters and hints for the newly sacked sequence
1269 * range but discard the return value since prev is already
1270 * marked. We must tag the range first because the seq
1271 * advancement below implicitly advances
1272 * tcp_highest_sack_seq() when skb is highest_sack.
1274 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1275 start_seq, end_seq, dup_sack, pcount,
1277 tcp_rate_skb_delivered(sk, skb, state->rate);
1279 if (skb == tp->lost_skb_hint)
1280 tp->lost_cnt_hint += pcount;
1282 TCP_SKB_CB(prev)->end_seq += shifted;
1283 TCP_SKB_CB(skb)->seq += shifted;
1285 tcp_skb_pcount_add(prev, pcount);
1286 BUG_ON(tcp_skb_pcount(skb) < pcount);
1287 tcp_skb_pcount_add(skb, -pcount);
1289 /* When we're adding to gso_segs == 1, gso_size will be zero,
1290 * in theory this shouldn't be necessary but as long as DSACK
1291 * code can come after this skb later on it's better to keep
1292 * setting gso_size to something.
1294 if (!TCP_SKB_CB(prev)->tcp_gso_size)
1295 TCP_SKB_CB(prev)->tcp_gso_size = mss;
1297 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1298 if (tcp_skb_pcount(skb) <= 1)
1299 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1301 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1302 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1305 BUG_ON(!tcp_skb_pcount(skb));
1306 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1310 /* Whole SKB was eaten :-) */
1312 if (skb == tp->retransmit_skb_hint)
1313 tp->retransmit_skb_hint = prev;
1314 if (skb == tp->lost_skb_hint) {
1315 tp->lost_skb_hint = prev;
1316 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1319 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1320 TCP_SKB_CB(prev)->eor = TCP_SKB_CB(skb)->eor;
1321 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1322 TCP_SKB_CB(prev)->end_seq++;
1324 if (skb == tcp_highest_sack(sk))
1325 tcp_advance_highest_sack(sk, skb);
1327 tcp_skb_collapse_tstamp(prev, skb);
1328 if (unlikely(TCP_SKB_CB(prev)->tx.delivered_mstamp))
1329 TCP_SKB_CB(prev)->tx.delivered_mstamp = 0;
1331 tcp_rtx_queue_unlink_and_free(skb, sk);
1333 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKMERGED);
1338 /* I wish gso_size would have a bit more sane initialization than
1339 * something-or-zero which complicates things
1341 static int tcp_skb_seglen(const struct sk_buff *skb)
1343 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1346 /* Shifting pages past head area doesn't work */
1347 static int skb_can_shift(const struct sk_buff *skb)
1349 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1352 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1355 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1356 struct tcp_sacktag_state *state,
1357 u32 start_seq, u32 end_seq,
1360 struct tcp_sock *tp = tcp_sk(sk);
1361 struct sk_buff *prev;
1367 if (!sk_can_gso(sk))
1370 /* Normally R but no L won't result in plain S */
1372 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1374 if (!skb_can_shift(skb))
1376 /* This frame is about to be dropped (was ACKed). */
1377 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1380 /* Can only happen with delayed DSACK + discard craziness */
1381 prev = skb_rb_prev(skb);
1385 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1388 if (!tcp_skb_can_collapse_to(prev))
1391 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1392 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1396 pcount = tcp_skb_pcount(skb);
1397 mss = tcp_skb_seglen(skb);
1399 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1400 * drop this restriction as unnecessary
1402 if (mss != tcp_skb_seglen(prev))
1405 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1407 /* CHECKME: This is non-MSS split case only?, this will
1408 * cause skipped skbs due to advancing loop btw, original
1409 * has that feature too
1411 if (tcp_skb_pcount(skb) <= 1)
1414 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1416 /* TODO: head merge to next could be attempted here
1417 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1418 * though it might not be worth of the additional hassle
1420 * ...we can probably just fallback to what was done
1421 * previously. We could try merging non-SACKed ones
1422 * as well but it probably isn't going to buy off
1423 * because later SACKs might again split them, and
1424 * it would make skb timestamp tracking considerably
1430 len = end_seq - TCP_SKB_CB(skb)->seq;
1432 BUG_ON(len > skb->len);
1434 /* MSS boundaries should be honoured or else pcount will
1435 * severely break even though it makes things bit trickier.
1436 * Optimize common case to avoid most of the divides
1438 mss = tcp_skb_mss(skb);
1440 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1441 * drop this restriction as unnecessary
1443 if (mss != tcp_skb_seglen(prev))
1448 } else if (len < mss) {
1456 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1457 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1460 if (!skb_shift(prev, skb, len))
1462 if (!tcp_shifted_skb(sk, prev, skb, state, pcount, len, mss, dup_sack))
1465 /* Hole filled allows collapsing with the next as well, this is very
1466 * useful when hole on every nth skb pattern happens
1468 skb = skb_rb_next(prev);
1472 if (!skb_can_shift(skb) ||
1473 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1474 (mss != tcp_skb_seglen(skb)))
1478 if (skb_shift(prev, skb, len)) {
1479 pcount += tcp_skb_pcount(skb);
1480 tcp_shifted_skb(sk, prev, skb, state, tcp_skb_pcount(skb),
1491 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1495 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1496 struct tcp_sack_block *next_dup,
1497 struct tcp_sacktag_state *state,
1498 u32 start_seq, u32 end_seq,
1501 struct tcp_sock *tp = tcp_sk(sk);
1502 struct sk_buff *tmp;
1504 skb_rbtree_walk_from(skb) {
1506 bool dup_sack = dup_sack_in;
1508 /* queue is in-order => we can short-circuit the walk early */
1509 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1513 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1514 in_sack = tcp_match_skb_to_sack(sk, skb,
1515 next_dup->start_seq,
1521 /* skb reference here is a bit tricky to get right, since
1522 * shifting can eat and free both this skb and the next,
1523 * so not even _safe variant of the loop is enough.
1526 tmp = tcp_shift_skb_data(sk, skb, state,
1527 start_seq, end_seq, dup_sack);
1536 in_sack = tcp_match_skb_to_sack(sk, skb,
1542 if (unlikely(in_sack < 0))
1546 TCP_SKB_CB(skb)->sacked =
1549 TCP_SKB_CB(skb)->sacked,
1550 TCP_SKB_CB(skb)->seq,
1551 TCP_SKB_CB(skb)->end_seq,
1553 tcp_skb_pcount(skb),
1555 tcp_rate_skb_delivered(sk, skb, state->rate);
1556 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1557 list_del_init(&skb->tcp_tsorted_anchor);
1559 if (!before(TCP_SKB_CB(skb)->seq,
1560 tcp_highest_sack_seq(tp)))
1561 tcp_advance_highest_sack(sk, skb);
1567 static struct sk_buff *tcp_sacktag_bsearch(struct sock *sk,
1568 struct tcp_sacktag_state *state,
1571 struct rb_node *parent, **p = &sk->tcp_rtx_queue.rb_node;
1572 struct sk_buff *skb;
1576 skb = rb_to_skb(parent);
1577 if (before(seq, TCP_SKB_CB(skb)->seq)) {
1578 p = &parent->rb_left;
1581 if (!before(seq, TCP_SKB_CB(skb)->end_seq)) {
1582 p = &parent->rb_right;
1590 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1591 struct tcp_sacktag_state *state,
1594 if (skb && after(TCP_SKB_CB(skb)->seq, skip_to_seq))
1597 return tcp_sacktag_bsearch(sk, state, skip_to_seq);
1600 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1602 struct tcp_sack_block *next_dup,
1603 struct tcp_sacktag_state *state,
1609 if (before(next_dup->start_seq, skip_to_seq)) {
1610 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1611 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1612 next_dup->start_seq, next_dup->end_seq,
1619 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1621 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1625 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1626 u32 prior_snd_una, struct tcp_sacktag_state *state)
1628 struct tcp_sock *tp = tcp_sk(sk);
1629 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1630 TCP_SKB_CB(ack_skb)->sacked);
1631 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1632 struct tcp_sack_block sp[TCP_NUM_SACKS];
1633 struct tcp_sack_block *cache;
1634 struct sk_buff *skb;
1635 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1637 bool found_dup_sack = false;
1639 int first_sack_index;
1642 state->reord = tp->snd_nxt;
1644 if (!tp->sacked_out)
1645 tcp_highest_sack_reset(sk);
1647 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1648 num_sacks, prior_snd_una);
1649 if (found_dup_sack) {
1650 state->flag |= FLAG_DSACKING_ACK;
1651 tp->delivered++; /* A spurious retransmission is delivered */
1654 /* Eliminate too old ACKs, but take into
1655 * account more or less fresh ones, they can
1656 * contain valid SACK info.
1658 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1661 if (!tp->packets_out)
1665 first_sack_index = 0;
1666 for (i = 0; i < num_sacks; i++) {
1667 bool dup_sack = !i && found_dup_sack;
1669 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1670 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1672 if (!tcp_is_sackblock_valid(tp, dup_sack,
1673 sp[used_sacks].start_seq,
1674 sp[used_sacks].end_seq)) {
1678 if (!tp->undo_marker)
1679 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1681 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1683 /* Don't count olds caused by ACK reordering */
1684 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1685 !after(sp[used_sacks].end_seq, tp->snd_una))
1687 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1690 NET_INC_STATS(sock_net(sk), mib_idx);
1692 first_sack_index = -1;
1696 /* Ignore very old stuff early */
1697 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1703 /* order SACK blocks to allow in order walk of the retrans queue */
1704 for (i = used_sacks - 1; i > 0; i--) {
1705 for (j = 0; j < i; j++) {
1706 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1707 swap(sp[j], sp[j + 1]);
1709 /* Track where the first SACK block goes to */
1710 if (j == first_sack_index)
1711 first_sack_index = j + 1;
1716 state->mss_now = tcp_current_mss(sk);
1720 if (!tp->sacked_out) {
1721 /* It's already past, so skip checking against it */
1722 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1724 cache = tp->recv_sack_cache;
1725 /* Skip empty blocks in at head of the cache */
1726 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1731 while (i < used_sacks) {
1732 u32 start_seq = sp[i].start_seq;
1733 u32 end_seq = sp[i].end_seq;
1734 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1735 struct tcp_sack_block *next_dup = NULL;
1737 if (found_dup_sack && ((i + 1) == first_sack_index))
1738 next_dup = &sp[i + 1];
1740 /* Skip too early cached blocks */
1741 while (tcp_sack_cache_ok(tp, cache) &&
1742 !before(start_seq, cache->end_seq))
1745 /* Can skip some work by looking recv_sack_cache? */
1746 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1747 after(end_seq, cache->start_seq)) {
1750 if (before(start_seq, cache->start_seq)) {
1751 skb = tcp_sacktag_skip(skb, sk, state,
1753 skb = tcp_sacktag_walk(skb, sk, next_dup,
1760 /* Rest of the block already fully processed? */
1761 if (!after(end_seq, cache->end_seq))
1764 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1768 /* ...tail remains todo... */
1769 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1770 /* ...but better entrypoint exists! */
1771 skb = tcp_highest_sack(sk);
1778 skb = tcp_sacktag_skip(skb, sk, state, cache->end_seq);
1779 /* Check overlap against next cached too (past this one already) */
1784 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1785 skb = tcp_highest_sack(sk);
1789 skb = tcp_sacktag_skip(skb, sk, state, start_seq);
1792 skb = tcp_sacktag_walk(skb, sk, next_dup, state,
1793 start_seq, end_seq, dup_sack);
1799 /* Clear the head of the cache sack blocks so we can skip it next time */
1800 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1801 tp->recv_sack_cache[i].start_seq = 0;
1802 tp->recv_sack_cache[i].end_seq = 0;
1804 for (j = 0; j < used_sacks; j++)
1805 tp->recv_sack_cache[i++] = sp[j];
1807 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Loss || tp->undo_marker)
1808 tcp_check_sack_reordering(sk, state->reord, 0);
1810 tcp_verify_left_out(tp);
1813 #if FASTRETRANS_DEBUG > 0
1814 WARN_ON((int)tp->sacked_out < 0);
1815 WARN_ON((int)tp->lost_out < 0);
1816 WARN_ON((int)tp->retrans_out < 0);
1817 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1822 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1823 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1825 static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
1829 holes = max(tp->lost_out, 1U);
1830 holes = min(holes, tp->packets_out);
1832 if ((tp->sacked_out + holes) > tp->packets_out) {
1833 tp->sacked_out = tp->packets_out - holes;
1839 /* If we receive more dupacks than we expected counting segments
1840 * in assumption of absent reordering, interpret this as reordering.
1841 * The only another reason could be bug in receiver TCP.
1843 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1845 struct tcp_sock *tp = tcp_sk(sk);
1847 if (!tcp_limit_reno_sacked(tp))
1850 tp->reordering = min_t(u32, tp->packets_out + addend,
1851 sock_net(sk)->ipv4.sysctl_tcp_max_reordering);
1852 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRENOREORDER);
1855 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1857 static void tcp_add_reno_sack(struct sock *sk)
1859 struct tcp_sock *tp = tcp_sk(sk);
1860 u32 prior_sacked = tp->sacked_out;
1863 tcp_check_reno_reordering(sk, 0);
1864 if (tp->sacked_out > prior_sacked)
1865 tp->delivered++; /* Some out-of-order packet is delivered */
1866 tcp_verify_left_out(tp);
1869 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1871 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1873 struct tcp_sock *tp = tcp_sk(sk);
1876 /* One ACK acked hole. The rest eat duplicate ACKs. */
1877 tp->delivered += max_t(int, acked - tp->sacked_out, 1);
1878 if (acked - 1 >= tp->sacked_out)
1881 tp->sacked_out -= acked - 1;
1883 tcp_check_reno_reordering(sk, acked);
1884 tcp_verify_left_out(tp);
1887 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1892 void tcp_clear_retrans(struct tcp_sock *tp)
1894 tp->retrans_out = 0;
1896 tp->undo_marker = 0;
1897 tp->undo_retrans = -1;
1901 static inline void tcp_init_undo(struct tcp_sock *tp)
1903 tp->undo_marker = tp->snd_una;
1904 /* Retransmission still in flight may cause DSACKs later. */
1905 tp->undo_retrans = tp->retrans_out ? : -1;
1908 /* Enter Loss state. If we detect SACK reneging, forget all SACK information
1909 * and reset tags completely, otherwise preserve SACKs. If receiver
1910 * dropped its ofo queue, we will know this due to reneging detection.
1912 void tcp_enter_loss(struct sock *sk)
1914 const struct inet_connection_sock *icsk = inet_csk(sk);
1915 struct tcp_sock *tp = tcp_sk(sk);
1916 struct net *net = sock_net(sk);
1917 struct sk_buff *skb;
1918 bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery;
1919 bool is_reneg; /* is receiver reneging on SACKs? */
1922 /* Reduce ssthresh if it has not yet been made inside this window. */
1923 if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
1924 !after(tp->high_seq, tp->snd_una) ||
1925 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1926 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1927 tp->prior_cwnd = tp->snd_cwnd;
1928 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1929 tcp_ca_event(sk, CA_EVENT_LOSS);
1933 tp->snd_cwnd_cnt = 0;
1934 tp->snd_cwnd_stamp = tcp_jiffies32;
1936 tp->retrans_out = 0;
1939 if (tcp_is_reno(tp))
1940 tcp_reset_reno_sack(tp);
1942 skb = tcp_rtx_queue_head(sk);
1943 is_reneg = skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED);
1945 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
1947 /* Mark SACK reneging until we recover from this loss event. */
1948 tp->is_sack_reneg = 1;
1950 tcp_clear_all_retrans_hints(tp);
1952 skb_rbtree_walk_from(skb) {
1953 mark_lost = (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
1956 tcp_sum_lost(tp, skb);
1957 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1959 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1960 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1961 tp->lost_out += tcp_skb_pcount(skb);
1964 tcp_verify_left_out(tp);
1966 /* Timeout in disordered state after receiving substantial DUPACKs
1967 * suggests that the degree of reordering is over-estimated.
1969 if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
1970 tp->sacked_out >= net->ipv4.sysctl_tcp_reordering)
1971 tp->reordering = min_t(unsigned int, tp->reordering,
1972 net->ipv4.sysctl_tcp_reordering);
1973 tcp_set_ca_state(sk, TCP_CA_Loss);
1974 tp->high_seq = tp->snd_nxt;
1975 tcp_ecn_queue_cwr(tp);
1977 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
1978 * loss recovery is underway except recurring timeout(s) on
1979 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
1981 * In theory F-RTO can be used repeatedly during loss recovery.
1982 * In practice this interacts badly with broken middle-boxes that
1983 * falsely raise the receive window, which results in repeated
1984 * timeouts and stop-and-go behavior.
1986 tp->frto = net->ipv4.sysctl_tcp_frto &&
1987 (new_recovery || icsk->icsk_retransmits) &&
1988 !inet_csk(sk)->icsk_mtup.probe_size;
1991 /* If ACK arrived pointing to a remembered SACK, it means that our
1992 * remembered SACKs do not reflect real state of receiver i.e.
1993 * receiver _host_ is heavily congested (or buggy).
1995 * To avoid big spurious retransmission bursts due to transient SACK
1996 * scoreboard oddities that look like reneging, we give the receiver a
1997 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
1998 * restore sanity to the SACK scoreboard. If the apparent reneging
1999 * persists until this RTO then we'll clear the SACK scoreboard.
2001 static bool tcp_check_sack_reneging(struct sock *sk, int flag)
2003 if (flag & FLAG_SACK_RENEGING) {
2004 struct tcp_sock *tp = tcp_sk(sk);
2005 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4),
2006 msecs_to_jiffies(10));
2008 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2009 delay, TCP_RTO_MAX);
2015 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2016 * counter when SACK is enabled (without SACK, sacked_out is used for
2019 * With reordering, holes may still be in flight, so RFC3517 recovery
2020 * uses pure sacked_out (total number of SACKed segments) even though
2021 * it violates the RFC that uses duplicate ACKs, often these are equal
2022 * but when e.g. out-of-window ACKs or packet duplication occurs,
2023 * they differ. Since neither occurs due to loss, TCP should really
2026 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2028 return tp->sacked_out + 1;
2031 /* Linux NewReno/SACK/ECN state machine.
2032 * --------------------------------------
2034 * "Open" Normal state, no dubious events, fast path.
2035 * "Disorder" In all the respects it is "Open",
2036 * but requires a bit more attention. It is entered when
2037 * we see some SACKs or dupacks. It is split of "Open"
2038 * mainly to move some processing from fast path to slow one.
2039 * "CWR" CWND was reduced due to some Congestion Notification event.
2040 * It can be ECN, ICMP source quench, local device congestion.
2041 * "Recovery" CWND was reduced, we are fast-retransmitting.
2042 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2044 * tcp_fastretrans_alert() is entered:
2045 * - each incoming ACK, if state is not "Open"
2046 * - when arrived ACK is unusual, namely:
2051 * Counting packets in flight is pretty simple.
2053 * in_flight = packets_out - left_out + retrans_out
2055 * packets_out is SND.NXT-SND.UNA counted in packets.
2057 * retrans_out is number of retransmitted segments.
2059 * left_out is number of segments left network, but not ACKed yet.
2061 * left_out = sacked_out + lost_out
2063 * sacked_out: Packets, which arrived to receiver out of order
2064 * and hence not ACKed. With SACKs this number is simply
2065 * amount of SACKed data. Even without SACKs
2066 * it is easy to give pretty reliable estimate of this number,
2067 * counting duplicate ACKs.
2069 * lost_out: Packets lost by network. TCP has no explicit
2070 * "loss notification" feedback from network (for now).
2071 * It means that this number can be only _guessed_.
2072 * Actually, it is the heuristics to predict lossage that
2073 * distinguishes different algorithms.
2075 * F.e. after RTO, when all the queue is considered as lost,
2076 * lost_out = packets_out and in_flight = retrans_out.
2078 * Essentially, we have now a few algorithms detecting
2081 * If the receiver supports SACK:
2083 * RFC6675/3517: It is the conventional algorithm. A packet is
2084 * considered lost if the number of higher sequence packets
2085 * SACKed is greater than or equal the DUPACK thoreshold
2086 * (reordering). This is implemented in tcp_mark_head_lost and
2087 * tcp_update_scoreboard.
2089 * RACK (draft-ietf-tcpm-rack-01): it is a newer algorithm
2090 * (2017-) that checks timing instead of counting DUPACKs.
2091 * Essentially a packet is considered lost if it's not S/ACKed
2092 * after RTT + reordering_window, where both metrics are
2093 * dynamically measured and adjusted. This is implemented in
2094 * tcp_rack_mark_lost.
2096 * If the receiver does not support SACK:
2098 * NewReno (RFC6582): in Recovery we assume that one segment
2099 * is lost (classic Reno). While we are in Recovery and
2100 * a partial ACK arrives, we assume that one more packet
2101 * is lost (NewReno). This heuristics are the same in NewReno
2104 * Really tricky (and requiring careful tuning) part of algorithm
2105 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2106 * The first determines the moment _when_ we should reduce CWND and,
2107 * hence, slow down forward transmission. In fact, it determines the moment
2108 * when we decide that hole is caused by loss, rather than by a reorder.
2110 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2111 * holes, caused by lost packets.
2113 * And the most logically complicated part of algorithm is undo
2114 * heuristics. We detect false retransmits due to both too early
2115 * fast retransmit (reordering) and underestimated RTO, analyzing
2116 * timestamps and D-SACKs. When we detect that some segments were
2117 * retransmitted by mistake and CWND reduction was wrong, we undo
2118 * window reduction and abort recovery phase. This logic is hidden
2119 * inside several functions named tcp_try_undo_<something>.
2122 /* This function decides, when we should leave Disordered state
2123 * and enter Recovery phase, reducing congestion window.
2125 * Main question: may we further continue forward transmission
2126 * with the same cwnd?
2128 static bool tcp_time_to_recover(struct sock *sk, int flag)
2130 struct tcp_sock *tp = tcp_sk(sk);
2132 /* Trick#1: The loss is proven. */
2136 /* Not-A-Trick#2 : Classic rule... */
2137 if (tcp_dupack_heuristics(tp) > tp->reordering)
2143 /* Detect loss in event "A" above by marking head of queue up as lost.
2144 * For non-SACK(Reno) senders, the first "packets" number of segments
2145 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2146 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2147 * the maximum SACKed segments to pass before reaching this limit.
2149 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2151 struct tcp_sock *tp = tcp_sk(sk);
2152 struct sk_buff *skb;
2153 int cnt, oldcnt, lost;
2155 /* Use SACK to deduce losses of new sequences sent during recovery */
2156 const u32 loss_high = tcp_is_sack(tp) ? tp->snd_nxt : tp->high_seq;
2158 WARN_ON(packets > tp->packets_out);
2159 skb = tp->lost_skb_hint;
2161 /* Head already handled? */
2162 if (mark_head && after(TCP_SKB_CB(skb)->seq, tp->snd_una))
2164 cnt = tp->lost_cnt_hint;
2166 skb = tcp_rtx_queue_head(sk);
2170 skb_rbtree_walk_from(skb) {
2171 /* TODO: do this better */
2172 /* this is not the most efficient way to do this... */
2173 tp->lost_skb_hint = skb;
2174 tp->lost_cnt_hint = cnt;
2176 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2180 if (tcp_is_reno(tp) ||
2181 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2182 cnt += tcp_skb_pcount(skb);
2184 if (cnt > packets) {
2185 if (tcp_is_sack(tp) ||
2186 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
2187 (oldcnt >= packets))
2190 mss = tcp_skb_mss(skb);
2191 /* If needed, chop off the prefix to mark as lost. */
2192 lost = (packets - oldcnt) * mss;
2193 if (lost < skb->len &&
2194 tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
2195 lost, mss, GFP_ATOMIC) < 0)
2200 tcp_skb_mark_lost(tp, skb);
2205 tcp_verify_left_out(tp);
2208 /* Account newly detected lost packet(s) */
2210 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2212 struct tcp_sock *tp = tcp_sk(sk);
2214 if (tcp_is_reno(tp)) {
2215 tcp_mark_head_lost(sk, 1, 1);
2217 int sacked_upto = tp->sacked_out - tp->reordering;
2218 if (sacked_upto >= 0)
2219 tcp_mark_head_lost(sk, sacked_upto, 0);
2220 else if (fast_rexmit)
2221 tcp_mark_head_lost(sk, 1, 1);
2225 static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when)
2227 return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2228 before(tp->rx_opt.rcv_tsecr, when);
2231 /* skb is spurious retransmitted if the returned timestamp echo
2232 * reply is prior to the skb transmission time
2234 static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp,
2235 const struct sk_buff *skb)
2237 return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) &&
2238 tcp_tsopt_ecr_before(tp, tcp_skb_timestamp(skb));
2241 /* Nothing was retransmitted or returned timestamp is less
2242 * than timestamp of the first retransmission.
2244 static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2246 return !tp->retrans_stamp ||
2247 tcp_tsopt_ecr_before(tp, tp->retrans_stamp);
2250 /* Undo procedures. */
2252 /* We can clear retrans_stamp when there are no retransmissions in the
2253 * window. It would seem that it is trivially available for us in
2254 * tp->retrans_out, however, that kind of assumptions doesn't consider
2255 * what will happen if errors occur when sending retransmission for the
2256 * second time. ...It could the that such segment has only
2257 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2258 * the head skb is enough except for some reneging corner cases that
2259 * are not worth the effort.
2261 * Main reason for all this complexity is the fact that connection dying
2262 * time now depends on the validity of the retrans_stamp, in particular,
2263 * that successive retransmissions of a segment must not advance
2264 * retrans_stamp under any conditions.
2266 static bool tcp_any_retrans_done(const struct sock *sk)
2268 const struct tcp_sock *tp = tcp_sk(sk);
2269 struct sk_buff *skb;
2271 if (tp->retrans_out)
2274 skb = tcp_rtx_queue_head(sk);
2275 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2281 static void DBGUNDO(struct sock *sk, const char *msg)
2283 #if FASTRETRANS_DEBUG > 1
2284 struct tcp_sock *tp = tcp_sk(sk);
2285 struct inet_sock *inet = inet_sk(sk);
2287 if (sk->sk_family == AF_INET) {
2288 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2290 &inet->inet_daddr, ntohs(inet->inet_dport),
2291 tp->snd_cwnd, tcp_left_out(tp),
2292 tp->snd_ssthresh, tp->prior_ssthresh,
2295 #if IS_ENABLED(CONFIG_IPV6)
2296 else if (sk->sk_family == AF_INET6) {
2297 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2299 &sk->sk_v6_daddr, ntohs(inet->inet_dport),
2300 tp->snd_cwnd, tcp_left_out(tp),
2301 tp->snd_ssthresh, tp->prior_ssthresh,
2308 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
2310 struct tcp_sock *tp = tcp_sk(sk);
2313 struct sk_buff *skb;
2315 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2316 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2319 tcp_clear_all_retrans_hints(tp);
2322 if (tp->prior_ssthresh) {
2323 const struct inet_connection_sock *icsk = inet_csk(sk);
2325 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2327 if (tp->prior_ssthresh > tp->snd_ssthresh) {
2328 tp->snd_ssthresh = tp->prior_ssthresh;
2329 tcp_ecn_withdraw_cwr(tp);
2332 tp->snd_cwnd_stamp = tcp_jiffies32;
2333 tp->undo_marker = 0;
2334 tp->rack.advanced = 1; /* Force RACK to re-exam losses */
2337 static inline bool tcp_may_undo(const struct tcp_sock *tp)
2339 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2342 /* People celebrate: "We love our President!" */
2343 static bool tcp_try_undo_recovery(struct sock *sk)
2345 struct tcp_sock *tp = tcp_sk(sk);
2347 if (tcp_may_undo(tp)) {
2350 /* Happy end! We did not retransmit anything
2351 * or our original transmission succeeded.
2353 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2354 tcp_undo_cwnd_reduction(sk, false);
2355 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2356 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2358 mib_idx = LINUX_MIB_TCPFULLUNDO;
2360 NET_INC_STATS(sock_net(sk), mib_idx);
2361 } else if (tp->rack.reo_wnd_persist) {
2362 tp->rack.reo_wnd_persist--;
2364 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2365 /* Hold old state until something *above* high_seq
2366 * is ACKed. For Reno it is MUST to prevent false
2367 * fast retransmits (RFC2582). SACK TCP is safe. */
2368 if (!tcp_any_retrans_done(sk))
2369 tp->retrans_stamp = 0;
2372 tcp_set_ca_state(sk, TCP_CA_Open);
2373 tp->is_sack_reneg = 0;
2377 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2378 static bool tcp_try_undo_dsack(struct sock *sk)
2380 struct tcp_sock *tp = tcp_sk(sk);
2382 if (tp->undo_marker && !tp->undo_retrans) {
2383 tp->rack.reo_wnd_persist = min(TCP_RACK_RECOVERY_THRESH,
2384 tp->rack.reo_wnd_persist + 1);
2385 DBGUNDO(sk, "D-SACK");
2386 tcp_undo_cwnd_reduction(sk, false);
2387 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2393 /* Undo during loss recovery after partial ACK or using F-RTO. */
2394 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2396 struct tcp_sock *tp = tcp_sk(sk);
2398 if (frto_undo || tcp_may_undo(tp)) {
2399 tcp_undo_cwnd_reduction(sk, true);
2401 DBGUNDO(sk, "partial loss");
2402 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2404 NET_INC_STATS(sock_net(sk),
2405 LINUX_MIB_TCPSPURIOUSRTOS);
2406 inet_csk(sk)->icsk_retransmits = 0;
2407 if (frto_undo || tcp_is_sack(tp)) {
2408 tcp_set_ca_state(sk, TCP_CA_Open);
2409 tp->is_sack_reneg = 0;
2416 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
2417 * It computes the number of packets to send (sndcnt) based on packets newly
2419 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2420 * cwnd reductions across a full RTT.
2421 * 2) Otherwise PRR uses packet conservation to send as much as delivered.
2422 * But when the retransmits are acked without further losses, PRR
2423 * slow starts cwnd up to ssthresh to speed up the recovery.
2425 static void tcp_init_cwnd_reduction(struct sock *sk)
2427 struct tcp_sock *tp = tcp_sk(sk);
2429 tp->high_seq = tp->snd_nxt;
2430 tp->tlp_high_seq = 0;
2431 tp->snd_cwnd_cnt = 0;
2432 tp->prior_cwnd = tp->snd_cwnd;
2433 tp->prr_delivered = 0;
2435 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2436 tcp_ecn_queue_cwr(tp);
2439 void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked, int flag)
2441 struct tcp_sock *tp = tcp_sk(sk);
2443 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2445 if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd))
2448 tp->prr_delivered += newly_acked_sacked;
2450 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2452 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2453 } else if ((flag & FLAG_RETRANS_DATA_ACKED) &&
2454 !(flag & FLAG_LOST_RETRANS)) {
2455 sndcnt = min_t(int, delta,
2456 max_t(int, tp->prr_delivered - tp->prr_out,
2457 newly_acked_sacked) + 1);
2459 sndcnt = min(delta, newly_acked_sacked);
2461 /* Force a fast retransmit upon entering fast recovery */
2462 sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1));
2463 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2466 static inline void tcp_end_cwnd_reduction(struct sock *sk)
2468 struct tcp_sock *tp = tcp_sk(sk);
2470 if (inet_csk(sk)->icsk_ca_ops->cong_control)
2473 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2474 if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH &&
2475 (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR || tp->undo_marker)) {
2476 tp->snd_cwnd = tp->snd_ssthresh;
2477 tp->snd_cwnd_stamp = tcp_jiffies32;
2479 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2482 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2483 void tcp_enter_cwr(struct sock *sk)
2485 struct tcp_sock *tp = tcp_sk(sk);
2487 tp->prior_ssthresh = 0;
2488 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2489 tp->undo_marker = 0;
2490 tcp_init_cwnd_reduction(sk);
2491 tcp_set_ca_state(sk, TCP_CA_CWR);
2494 EXPORT_SYMBOL(tcp_enter_cwr);
2496 static void tcp_try_keep_open(struct sock *sk)
2498 struct tcp_sock *tp = tcp_sk(sk);
2499 int state = TCP_CA_Open;
2501 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2502 state = TCP_CA_Disorder;
2504 if (inet_csk(sk)->icsk_ca_state != state) {
2505 tcp_set_ca_state(sk, state);
2506 tp->high_seq = tp->snd_nxt;
2510 static void tcp_try_to_open(struct sock *sk, int flag)
2512 struct tcp_sock *tp = tcp_sk(sk);
2514 tcp_verify_left_out(tp);
2516 if (!tcp_any_retrans_done(sk))
2517 tp->retrans_stamp = 0;
2519 if (flag & FLAG_ECE)
2522 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2523 tcp_try_keep_open(sk);
2527 static void tcp_mtup_probe_failed(struct sock *sk)
2529 struct inet_connection_sock *icsk = inet_csk(sk);
2531 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2532 icsk->icsk_mtup.probe_size = 0;
2533 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPFAIL);
2536 static void tcp_mtup_probe_success(struct sock *sk)
2538 struct tcp_sock *tp = tcp_sk(sk);
2539 struct inet_connection_sock *icsk = inet_csk(sk);
2541 /* FIXME: breaks with very large cwnd */
2542 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2543 tp->snd_cwnd = tp->snd_cwnd *
2544 tcp_mss_to_mtu(sk, tp->mss_cache) /
2545 icsk->icsk_mtup.probe_size;
2546 tp->snd_cwnd_cnt = 0;
2547 tp->snd_cwnd_stamp = tcp_jiffies32;
2548 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2550 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2551 icsk->icsk_mtup.probe_size = 0;
2552 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2553 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS);
2556 /* Do a simple retransmit without using the backoff mechanisms in
2557 * tcp_timer. This is used for path mtu discovery.
2558 * The socket is already locked here.
2560 void tcp_simple_retransmit(struct sock *sk)
2562 const struct inet_connection_sock *icsk = inet_csk(sk);
2563 struct tcp_sock *tp = tcp_sk(sk);
2564 struct sk_buff *skb;
2565 unsigned int mss = tcp_current_mss(sk);
2567 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2568 if (tcp_skb_seglen(skb) > mss &&
2569 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2570 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2571 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2572 tp->retrans_out -= tcp_skb_pcount(skb);
2574 tcp_skb_mark_lost_uncond_verify(tp, skb);
2578 tcp_clear_retrans_hints_partial(tp);
2583 if (tcp_is_reno(tp))
2584 tcp_limit_reno_sacked(tp);
2586 tcp_verify_left_out(tp);
2588 /* Don't muck with the congestion window here.
2589 * Reason is that we do not increase amount of _data_
2590 * in network, but units changed and effective
2591 * cwnd/ssthresh really reduced now.
2593 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2594 tp->high_seq = tp->snd_nxt;
2595 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2596 tp->prior_ssthresh = 0;
2597 tp->undo_marker = 0;
2598 tcp_set_ca_state(sk, TCP_CA_Loss);
2600 tcp_xmit_retransmit_queue(sk);
2602 EXPORT_SYMBOL(tcp_simple_retransmit);
2604 void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2606 struct tcp_sock *tp = tcp_sk(sk);
2609 if (tcp_is_reno(tp))
2610 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2612 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2614 NET_INC_STATS(sock_net(sk), mib_idx);
2616 tp->prior_ssthresh = 0;
2619 if (!tcp_in_cwnd_reduction(sk)) {
2621 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2622 tcp_init_cwnd_reduction(sk);
2624 tcp_set_ca_state(sk, TCP_CA_Recovery);
2627 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2628 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2630 static void tcp_process_loss(struct sock *sk, int flag, bool is_dupack,
2633 struct tcp_sock *tp = tcp_sk(sk);
2634 bool recovered = !before(tp->snd_una, tp->high_seq);
2636 if ((flag & FLAG_SND_UNA_ADVANCED) &&
2637 tcp_try_undo_loss(sk, false))
2640 /* The ACK (s)acks some never-retransmitted data meaning not all
2641 * the data packets before the timeout were lost. Therefore we
2642 * undo the congestion window and state. This is essentially
2643 * the operation in F-RTO (RFC5682 section 3.1 step 3.b). Since
2644 * a retransmitted skb is permantly marked, we can apply such an
2645 * operation even if F-RTO was not used.
2647 if ((flag & FLAG_ORIG_SACK_ACKED) &&
2648 tcp_try_undo_loss(sk, tp->undo_marker))
2651 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2652 if (after(tp->snd_nxt, tp->high_seq)) {
2653 if (flag & FLAG_DATA_SACKED || is_dupack)
2654 tp->frto = 0; /* Step 3.a. loss was real */
2655 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2656 tp->high_seq = tp->snd_nxt;
2657 /* Step 2.b. Try send new data (but deferred until cwnd
2658 * is updated in tcp_ack()). Otherwise fall back to
2659 * the conventional recovery.
2661 if (!tcp_write_queue_empty(sk) &&
2662 after(tcp_wnd_end(tp), tp->snd_nxt)) {
2663 *rexmit = REXMIT_NEW;
2671 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2672 tcp_try_undo_recovery(sk);
2675 if (tcp_is_reno(tp)) {
2676 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2677 * delivered. Lower inflight to clock out (re)tranmissions.
2679 if (after(tp->snd_nxt, tp->high_seq) && is_dupack)
2680 tcp_add_reno_sack(sk);
2681 else if (flag & FLAG_SND_UNA_ADVANCED)
2682 tcp_reset_reno_sack(tp);
2684 *rexmit = REXMIT_LOST;
2687 /* Undo during fast recovery after partial ACK. */
2688 static bool tcp_try_undo_partial(struct sock *sk, u32 prior_snd_una)
2690 struct tcp_sock *tp = tcp_sk(sk);
2692 if (tp->undo_marker && tcp_packet_delayed(tp)) {
2693 /* Plain luck! Hole if filled with delayed
2694 * packet, rather than with a retransmit. Check reordering.
2696 tcp_check_sack_reordering(sk, prior_snd_una, 1);
2698 /* We are getting evidence that the reordering degree is higher
2699 * than we realized. If there are no retransmits out then we
2700 * can undo. Otherwise we clock out new packets but do not
2701 * mark more packets lost or retransmit more.
2703 if (tp->retrans_out)
2706 if (!tcp_any_retrans_done(sk))
2707 tp->retrans_stamp = 0;
2709 DBGUNDO(sk, "partial recovery");
2710 tcp_undo_cwnd_reduction(sk, true);
2711 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2712 tcp_try_keep_open(sk);
2718 static void tcp_rack_identify_loss(struct sock *sk, int *ack_flag)
2720 struct tcp_sock *tp = tcp_sk(sk);
2722 /* Use RACK to detect loss */
2723 if (sock_net(sk)->ipv4.sysctl_tcp_recovery & TCP_RACK_LOSS_DETECTION) {
2724 u32 prior_retrans = tp->retrans_out;
2726 tcp_rack_mark_lost(sk);
2727 if (prior_retrans > tp->retrans_out)
2728 *ack_flag |= FLAG_LOST_RETRANS;
2732 static bool tcp_force_fast_retransmit(struct sock *sk)
2734 struct tcp_sock *tp = tcp_sk(sk);
2736 return after(tcp_highest_sack_seq(tp),
2737 tp->snd_una + tp->reordering * tp->mss_cache);
2740 /* Process an event, which can update packets-in-flight not trivially.
2741 * Main goal of this function is to calculate new estimate for left_out,
2742 * taking into account both packets sitting in receiver's buffer and
2743 * packets lost by network.
2745 * Besides that it updates the congestion state when packet loss or ECN
2746 * is detected. But it does not reduce the cwnd, it is done by the
2747 * congestion control later.
2749 * It does _not_ decide what to send, it is made in function
2750 * tcp_xmit_retransmit_queue().
2752 static void tcp_fastretrans_alert(struct sock *sk, const u32 prior_snd_una,
2753 bool is_dupack, int *ack_flag, int *rexmit)
2755 struct inet_connection_sock *icsk = inet_csk(sk);
2756 struct tcp_sock *tp = tcp_sk(sk);
2757 int fast_rexmit = 0, flag = *ack_flag;
2758 bool do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2759 tcp_force_fast_retransmit(sk));
2761 if (!tp->packets_out && tp->sacked_out)
2764 /* Now state machine starts.
2765 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2766 if (flag & FLAG_ECE)
2767 tp->prior_ssthresh = 0;
2769 /* B. In all the states check for reneging SACKs. */
2770 if (tcp_check_sack_reneging(sk, flag))
2773 /* C. Check consistency of the current state. */
2774 tcp_verify_left_out(tp);
2776 /* D. Check state exit conditions. State can be terminated
2777 * when high_seq is ACKed. */
2778 if (icsk->icsk_ca_state == TCP_CA_Open) {
2779 WARN_ON(tp->retrans_out != 0);
2780 tp->retrans_stamp = 0;
2781 } else if (!before(tp->snd_una, tp->high_seq)) {
2782 switch (icsk->icsk_ca_state) {
2784 /* CWR is to be held something *above* high_seq
2785 * is ACKed for CWR bit to reach receiver. */
2786 if (tp->snd_una != tp->high_seq) {
2787 tcp_end_cwnd_reduction(sk);
2788 tcp_set_ca_state(sk, TCP_CA_Open);
2792 case TCP_CA_Recovery:
2793 if (tcp_is_reno(tp))
2794 tcp_reset_reno_sack(tp);
2795 if (tcp_try_undo_recovery(sk))
2797 tcp_end_cwnd_reduction(sk);
2802 /* E. Process state. */
2803 switch (icsk->icsk_ca_state) {
2804 case TCP_CA_Recovery:
2805 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2806 if (tcp_is_reno(tp) && is_dupack)
2807 tcp_add_reno_sack(sk);
2809 if (tcp_try_undo_partial(sk, prior_snd_una))
2811 /* Partial ACK arrived. Force fast retransmit. */
2812 do_lost = tcp_is_reno(tp) ||
2813 tcp_force_fast_retransmit(sk);
2815 if (tcp_try_undo_dsack(sk)) {
2816 tcp_try_keep_open(sk);
2819 tcp_rack_identify_loss(sk, ack_flag);
2822 tcp_process_loss(sk, flag, is_dupack, rexmit);
2823 tcp_rack_identify_loss(sk, ack_flag);
2824 if (!(icsk->icsk_ca_state == TCP_CA_Open ||
2825 (*ack_flag & FLAG_LOST_RETRANS)))
2827 /* Change state if cwnd is undone or retransmits are lost */
2830 if (tcp_is_reno(tp)) {
2831 if (flag & FLAG_SND_UNA_ADVANCED)
2832 tcp_reset_reno_sack(tp);
2834 tcp_add_reno_sack(sk);
2837 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
2838 tcp_try_undo_dsack(sk);
2840 tcp_rack_identify_loss(sk, ack_flag);
2841 if (!tcp_time_to_recover(sk, flag)) {
2842 tcp_try_to_open(sk, flag);
2846 /* MTU probe failure: don't reduce cwnd */
2847 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2848 icsk->icsk_mtup.probe_size &&
2849 tp->snd_una == tp->mtu_probe.probe_seq_start) {
2850 tcp_mtup_probe_failed(sk);
2851 /* Restores the reduction we did in tcp_mtup_probe() */
2853 tcp_simple_retransmit(sk);
2857 /* Otherwise enter Recovery state */
2858 tcp_enter_recovery(sk, (flag & FLAG_ECE));
2863 tcp_update_scoreboard(sk, fast_rexmit);
2864 *rexmit = REXMIT_LOST;
2867 static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us)
2869 u32 wlen = sock_net(sk)->ipv4.sysctl_tcp_min_rtt_wlen * HZ;
2870 struct tcp_sock *tp = tcp_sk(sk);
2872 minmax_running_min(&tp->rtt_min, wlen, tcp_jiffies32,
2873 rtt_us ? : jiffies_to_usecs(1));
2876 static bool tcp_ack_update_rtt(struct sock *sk, const int flag,
2877 long seq_rtt_us, long sack_rtt_us,
2878 long ca_rtt_us, struct rate_sample *rs)
2880 const struct tcp_sock *tp = tcp_sk(sk);
2882 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2883 * broken middle-boxes or peers may corrupt TS-ECR fields. But
2884 * Karn's algorithm forbids taking RTT if some retransmitted data
2885 * is acked (RFC6298).
2888 seq_rtt_us = sack_rtt_us;
2890 /* RTTM Rule: A TSecr value received in a segment is used to
2891 * update the averaged RTT measurement only if the segment
2892 * acknowledges some new data, i.e., only if it advances the
2893 * left edge of the send window.
2894 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2896 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2897 flag & FLAG_ACKED) {
2898 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr;
2899 u32 delta_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
2901 seq_rtt_us = ca_rtt_us = delta_us;
2903 rs->rtt_us = ca_rtt_us; /* RTT of last (S)ACKed packet (or -1) */
2907 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
2908 * always taken together with ACK, SACK, or TS-opts. Any negative
2909 * values will be skipped with the seq_rtt_us < 0 check above.
2911 tcp_update_rtt_min(sk, ca_rtt_us);
2912 tcp_rtt_estimator(sk, seq_rtt_us);
2915 /* RFC6298: only reset backoff on valid RTT measurement. */
2916 inet_csk(sk)->icsk_backoff = 0;
2920 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
2921 void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req)
2923 struct rate_sample rs;
2926 if (req && !req->num_retrans && tcp_rsk(req)->snt_synack)
2927 rtt_us = tcp_stamp_us_delta(tcp_clock_us(), tcp_rsk(req)->snt_synack);
2929 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us, &rs);
2933 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
2935 const struct inet_connection_sock *icsk = inet_csk(sk);
2937 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
2938 tcp_sk(sk)->snd_cwnd_stamp = tcp_jiffies32;
2941 /* Restart timer after forward progress on connection.
2942 * RFC2988 recommends to restart timer to now+rto.
2944 void tcp_rearm_rto(struct sock *sk)
2946 const struct inet_connection_sock *icsk = inet_csk(sk);
2947 struct tcp_sock *tp = tcp_sk(sk);
2949 /* If the retrans timer is currently being used by Fast Open
2950 * for SYN-ACK retrans purpose, stay put.
2952 if (tp->fastopen_rsk)
2955 if (!tp->packets_out) {
2956 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
2958 u32 rto = inet_csk(sk)->icsk_rto;
2959 /* Offset the time elapsed after installing regular RTO */
2960 if (icsk->icsk_pending == ICSK_TIME_REO_TIMEOUT ||
2961 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
2962 s64 delta_us = tcp_rto_delta_us(sk);
2963 /* delta_us may not be positive if the socket is locked
2964 * when the retrans timer fires and is rescheduled.
2966 rto = usecs_to_jiffies(max_t(int, delta_us, 1));
2968 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
2973 /* Try to schedule a loss probe; if that doesn't work, then schedule an RTO. */
2974 static void tcp_set_xmit_timer(struct sock *sk)
2976 if (!tcp_schedule_loss_probe(sk, true))
2980 /* If we get here, the whole TSO packet has not been acked. */
2981 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
2983 struct tcp_sock *tp = tcp_sk(sk);
2986 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
2988 packets_acked = tcp_skb_pcount(skb);
2989 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2991 packets_acked -= tcp_skb_pcount(skb);
2993 if (packets_acked) {
2994 BUG_ON(tcp_skb_pcount(skb) == 0);
2995 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
2998 return packets_acked;
3001 static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
3004 const struct skb_shared_info *shinfo;
3006 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3007 if (likely(!TCP_SKB_CB(skb)->txstamp_ack))
3010 shinfo = skb_shinfo(skb);
3011 if (!before(shinfo->tskey, prior_snd_una) &&
3012 before(shinfo->tskey, tcp_sk(sk)->snd_una)) {
3013 tcp_skb_tsorted_save(skb) {
3014 __skb_tstamp_tx(skb, NULL, sk, SCM_TSTAMP_ACK);
3015 } tcp_skb_tsorted_restore(skb);
3019 /* Remove acknowledged frames from the retransmission queue. If our packet
3020 * is before the ack sequence we can discard it as it's confirmed to have
3021 * arrived at the other end.
3023 static int tcp_clean_rtx_queue(struct sock *sk, u32 prior_fack,
3025 struct tcp_sacktag_state *sack)
3027 const struct inet_connection_sock *icsk = inet_csk(sk);
3028 u64 first_ackt, last_ackt;
3029 struct tcp_sock *tp = tcp_sk(sk);
3030 u32 prior_sacked = tp->sacked_out;
3031 u32 reord = tp->snd_nxt; /* lowest acked un-retx un-sacked seq */
3032 struct sk_buff *skb, *next;
3033 bool fully_acked = true;
3034 long sack_rtt_us = -1L;
3035 long seq_rtt_us = -1L;
3036 long ca_rtt_us = -1L;
3038 u32 last_in_flight = 0;
3044 for (skb = skb_rb_first(&sk->tcp_rtx_queue); skb; skb = next) {
3045 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3046 const u32 start_seq = scb->seq;
3047 u8 sacked = scb->sacked;
3050 tcp_ack_tstamp(sk, skb, prior_snd_una);
3052 /* Determine how many packets and what bytes were acked, tso and else */
3053 if (after(scb->end_seq, tp->snd_una)) {
3054 if (tcp_skb_pcount(skb) == 1 ||
3055 !after(tp->snd_una, scb->seq))
3058 acked_pcount = tcp_tso_acked(sk, skb);
3061 fully_acked = false;
3063 acked_pcount = tcp_skb_pcount(skb);
3066 if (unlikely(sacked & TCPCB_RETRANS)) {
3067 if (sacked & TCPCB_SACKED_RETRANS)
3068 tp->retrans_out -= acked_pcount;
3069 flag |= FLAG_RETRANS_DATA_ACKED;
3070 } else if (!(sacked & TCPCB_SACKED_ACKED)) {
3071 last_ackt = skb->skb_mstamp;
3072 WARN_ON_ONCE(last_ackt == 0);
3074 first_ackt = last_ackt;
3076 last_in_flight = TCP_SKB_CB(skb)->tx.in_flight;
3077 if (before(start_seq, reord))
3079 if (!after(scb->end_seq, tp->high_seq))
3080 flag |= FLAG_ORIG_SACK_ACKED;
3083 if (sacked & TCPCB_SACKED_ACKED) {
3084 tp->sacked_out -= acked_pcount;
3085 } else if (tcp_is_sack(tp)) {
3086 tp->delivered += acked_pcount;
3087 if (!tcp_skb_spurious_retrans(tp, skb))
3088 tcp_rack_advance(tp, sacked, scb->end_seq,
3091 if (sacked & TCPCB_LOST)
3092 tp->lost_out -= acked_pcount;
3094 tp->packets_out -= acked_pcount;
3095 pkts_acked += acked_pcount;
3096 tcp_rate_skb_delivered(sk, skb, sack->rate);
3098 /* Initial outgoing SYN's get put onto the write_queue
3099 * just like anything else we transmit. It is not
3100 * true data, and if we misinform our callers that
3101 * this ACK acks real data, we will erroneously exit
3102 * connection startup slow start one packet too
3103 * quickly. This is severely frowned upon behavior.
3105 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
3106 flag |= FLAG_DATA_ACKED;
3108 flag |= FLAG_SYN_ACKED;
3109 tp->retrans_stamp = 0;
3115 next = skb_rb_next(skb);
3116 if (unlikely(skb == tp->retransmit_skb_hint))
3117 tp->retransmit_skb_hint = NULL;
3118 if (unlikely(skb == tp->lost_skb_hint))
3119 tp->lost_skb_hint = NULL;
3120 tcp_rtx_queue_unlink_and_free(skb, sk);
3124 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
3126 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3127 tp->snd_up = tp->snd_una;
3129 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3130 flag |= FLAG_SACK_RENEGING;
3132 if (likely(first_ackt) && !(flag & FLAG_RETRANS_DATA_ACKED)) {
3133 seq_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, first_ackt);
3134 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, last_ackt);
3136 if (sack->first_sackt) {
3137 sack_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->first_sackt);
3138 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->last_sackt);
3140 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us,
3141 ca_rtt_us, sack->rate);
3143 if (flag & FLAG_ACKED) {
3144 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3145 if (unlikely(icsk->icsk_mtup.probe_size &&
3146 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3147 tcp_mtup_probe_success(sk);
3150 if (tcp_is_reno(tp)) {
3151 tcp_remove_reno_sacks(sk, pkts_acked);
3155 /* Non-retransmitted hole got filled? That's reordering */
3156 if (before(reord, prior_fack))
3157 tcp_check_sack_reordering(sk, reord, 0);
3159 delta = prior_sacked - tp->sacked_out;
3160 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3162 } else if (skb && rtt_update && sack_rtt_us >= 0 &&
3163 sack_rtt_us > tcp_stamp_us_delta(tp->tcp_mstamp, skb->skb_mstamp)) {
3164 /* Do not re-arm RTO if the sack RTT is measured from data sent
3165 * after when the head was last (re)transmitted. Otherwise the
3166 * timeout may continue to extend in loss recovery.
3168 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3171 if (icsk->icsk_ca_ops->pkts_acked) {
3172 struct ack_sample sample = { .pkts_acked = pkts_acked,
3173 .rtt_us = sack->rate->rtt_us,
3174 .in_flight = last_in_flight };
3176 icsk->icsk_ca_ops->pkts_acked(sk, &sample);
3179 #if FASTRETRANS_DEBUG > 0
3180 WARN_ON((int)tp->sacked_out < 0);
3181 WARN_ON((int)tp->lost_out < 0);
3182 WARN_ON((int)tp->retrans_out < 0);
3183 if (!tp->packets_out && tcp_is_sack(tp)) {
3184 icsk = inet_csk(sk);
3186 pr_debug("Leak l=%u %d\n",
3187 tp->lost_out, icsk->icsk_ca_state);
3190 if (tp->sacked_out) {
3191 pr_debug("Leak s=%u %d\n",
3192 tp->sacked_out, icsk->icsk_ca_state);
3195 if (tp->retrans_out) {
3196 pr_debug("Leak r=%u %d\n",
3197 tp->retrans_out, icsk->icsk_ca_state);
3198 tp->retrans_out = 0;
3205 static void tcp_ack_probe(struct sock *sk)
3207 struct inet_connection_sock *icsk = inet_csk(sk);
3208 struct sk_buff *head = tcp_send_head(sk);
3209 const struct tcp_sock *tp = tcp_sk(sk);
3211 /* Was it a usable window open? */
3214 if (!after(TCP_SKB_CB(head)->end_seq, tcp_wnd_end(tp))) {
3215 icsk->icsk_backoff = 0;
3216 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3217 /* Socket must be waked up by subsequent tcp_data_snd_check().
3218 * This function is not for random using!
3221 unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX);
3223 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3228 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3230 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3231 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3234 /* Decide wheather to run the increase function of congestion control. */
3235 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3237 /* If reordering is high then always grow cwnd whenever data is
3238 * delivered regardless of its ordering. Otherwise stay conservative
3239 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3240 * new SACK or ECE mark may first advance cwnd here and later reduce
3241 * cwnd in tcp_fastretrans_alert() based on more states.
3243 if (tcp_sk(sk)->reordering > sock_net(sk)->ipv4.sysctl_tcp_reordering)
3244 return flag & FLAG_FORWARD_PROGRESS;
3246 return flag & FLAG_DATA_ACKED;
3249 /* The "ultimate" congestion control function that aims to replace the rigid
3250 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction).
3251 * It's called toward the end of processing an ACK with precise rate
3252 * information. All transmission or retransmission are delayed afterwards.
3254 static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked,
3255 int flag, const struct rate_sample *rs)
3257 const struct inet_connection_sock *icsk = inet_csk(sk);
3259 if (icsk->icsk_ca_ops->cong_control) {
3260 icsk->icsk_ca_ops->cong_control(sk, rs);
3264 if (tcp_in_cwnd_reduction(sk)) {
3265 /* Reduce cwnd if state mandates */
3266 tcp_cwnd_reduction(sk, acked_sacked, flag);
3267 } else if (tcp_may_raise_cwnd(sk, flag)) {
3268 /* Advance cwnd if state allows */
3269 tcp_cong_avoid(sk, ack, acked_sacked);
3271 tcp_update_pacing_rate(sk);
3274 /* Check that window update is acceptable.
3275 * The function assumes that snd_una<=ack<=snd_next.
3277 static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3278 const u32 ack, const u32 ack_seq,
3281 return after(ack, tp->snd_una) ||
3282 after(ack_seq, tp->snd_wl1) ||
3283 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3286 /* If we update tp->snd_una, also update tp->bytes_acked */
3287 static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack)
3289 u32 delta = ack - tp->snd_una;
3291 sock_owned_by_me((struct sock *)tp);
3292 tp->bytes_acked += delta;
3296 /* If we update tp->rcv_nxt, also update tp->bytes_received */
3297 static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq)
3299 u32 delta = seq - tp->rcv_nxt;
3301 sock_owned_by_me((struct sock *)tp);
3302 tp->bytes_received += delta;
3306 /* Update our send window.
3308 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3309 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3311 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3314 struct tcp_sock *tp = tcp_sk(sk);
3316 u32 nwin = ntohs(tcp_hdr(skb)->window);
3318 if (likely(!tcp_hdr(skb)->syn))
3319 nwin <<= tp->rx_opt.snd_wscale;
3321 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3322 flag |= FLAG_WIN_UPDATE;
3323 tcp_update_wl(tp, ack_seq);
3325 if (tp->snd_wnd != nwin) {
3328 /* Note, it is the only place, where
3329 * fast path is recovered for sending TCP.
3332 tcp_fast_path_check(sk);
3334 if (!tcp_write_queue_empty(sk))
3335 tcp_slow_start_after_idle_check(sk);
3337 if (nwin > tp->max_window) {
3338 tp->max_window = nwin;
3339 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3344 tcp_snd_una_update(tp, ack);
3349 static bool __tcp_oow_rate_limited(struct net *net, int mib_idx,
3350 u32 *last_oow_ack_time)
3352 if (*last_oow_ack_time) {
3353 s32 elapsed = (s32)(tcp_jiffies32 - *last_oow_ack_time);
3355 if (0 <= elapsed && elapsed < net->ipv4.sysctl_tcp_invalid_ratelimit) {
3356 NET_INC_STATS(net, mib_idx);
3357 return true; /* rate-limited: don't send yet! */
3361 *last_oow_ack_time = tcp_jiffies32;
3363 return false; /* not rate-limited: go ahead, send dupack now! */
3366 /* Return true if we're currently rate-limiting out-of-window ACKs and
3367 * thus shouldn't send a dupack right now. We rate-limit dupacks in
3368 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3369 * attacks that send repeated SYNs or ACKs for the same connection. To
3370 * do this, we do not send a duplicate SYNACK or ACK if the remote
3371 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3373 bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb,
3374 int mib_idx, u32 *last_oow_ack_time)
3376 /* Data packets without SYNs are not likely part of an ACK loop. */
3377 if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) &&
3381 return __tcp_oow_rate_limited(net, mib_idx, last_oow_ack_time);
3384 /* RFC 5961 7 [ACK Throttling] */
3385 static void tcp_send_challenge_ack(struct sock *sk, const struct sk_buff *skb)
3387 /* unprotected vars, we dont care of overwrites */
3388 static u32 challenge_timestamp;
3389 static unsigned int challenge_count;
3390 struct tcp_sock *tp = tcp_sk(sk);
3391 struct net *net = sock_net(sk);
3394 /* First check our per-socket dupack rate limit. */
3395 if (__tcp_oow_rate_limited(net,
3396 LINUX_MIB_TCPACKSKIPPEDCHALLENGE,
3397 &tp->last_oow_ack_time))
3400 /* Then check host-wide RFC 5961 rate limit. */
3402 if (now != challenge_timestamp) {
3403 u32 ack_limit = net->ipv4.sysctl_tcp_challenge_ack_limit;
3404 u32 half = (ack_limit + 1) >> 1;
3406 challenge_timestamp = now;
3407 WRITE_ONCE(challenge_count, half + prandom_u32_max(ack_limit));
3409 count = READ_ONCE(challenge_count);
3411 WRITE_ONCE(challenge_count, count - 1);
3412 NET_INC_STATS(net, LINUX_MIB_TCPCHALLENGEACK);
3417 static void tcp_store_ts_recent(struct tcp_sock *tp)
3419 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3420 tp->rx_opt.ts_recent_stamp = get_seconds();
3423 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3425 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3426 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3427 * extra check below makes sure this can only happen
3428 * for pure ACK frames. -DaveM
3430 * Not only, also it occurs for expired timestamps.
3433 if (tcp_paws_check(&tp->rx_opt, 0))
3434 tcp_store_ts_recent(tp);
3438 /* This routine deals with acks during a TLP episode.
3439 * We mark the end of a TLP episode on receiving TLP dupack or when
3440 * ack is after tlp_high_seq.
3441 * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3443 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3445 struct tcp_sock *tp = tcp_sk(sk);
3447 if (before(ack, tp->tlp_high_seq))
3450 if (flag & FLAG_DSACKING_ACK) {
3451 /* This DSACK means original and TLP probe arrived; no loss */
3452 tp->tlp_high_seq = 0;
3453 } else if (after(ack, tp->tlp_high_seq)) {
3454 /* ACK advances: there was a loss, so reduce cwnd. Reset
3455 * tlp_high_seq in tcp_init_cwnd_reduction()
3457 tcp_init_cwnd_reduction(sk);
3458 tcp_set_ca_state(sk, TCP_CA_CWR);
3459 tcp_end_cwnd_reduction(sk);
3460 tcp_try_keep_open(sk);
3461 NET_INC_STATS(sock_net(sk),
3462 LINUX_MIB_TCPLOSSPROBERECOVERY);
3463 } else if (!(flag & (FLAG_SND_UNA_ADVANCED |
3464 FLAG_NOT_DUP | FLAG_DATA_SACKED))) {
3465 /* Pure dupack: original and TLP probe arrived; no loss */
3466 tp->tlp_high_seq = 0;
3470 static inline void tcp_in_ack_event(struct sock *sk, u32 flags)
3472 const struct inet_connection_sock *icsk = inet_csk(sk);
3474 if (icsk->icsk_ca_ops->in_ack_event)
3475 icsk->icsk_ca_ops->in_ack_event(sk, flags);
3478 /* Congestion control has updated the cwnd already. So if we're in
3479 * loss recovery then now we do any new sends (for FRTO) or
3480 * retransmits (for CA_Loss or CA_recovery) that make sense.
3482 static void tcp_xmit_recovery(struct sock *sk, int rexmit)
3484 struct tcp_sock *tp = tcp_sk(sk);
3486 if (rexmit == REXMIT_NONE)
3489 if (unlikely(rexmit == 2)) {
3490 __tcp_push_pending_frames(sk, tcp_current_mss(sk),
3492 if (after(tp->snd_nxt, tp->high_seq))
3496 tcp_xmit_retransmit_queue(sk);
3499 /* This routine deals with incoming acks, but not outgoing ones. */
3500 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3502 struct inet_connection_sock *icsk = inet_csk(sk);
3503 struct tcp_sock *tp = tcp_sk(sk);
3504 struct tcp_sacktag_state sack_state;
3505 struct rate_sample rs = { .prior_delivered = 0 };
3506 u32 prior_snd_una = tp->snd_una;
3507 bool is_sack_reneg = tp->is_sack_reneg;
3508 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3509 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3510 bool is_dupack = false;
3511 int prior_packets = tp->packets_out;
3512 u32 delivered = tp->delivered;
3513 u32 lost = tp->lost;
3514 int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */
3517 sack_state.first_sackt = 0;
3518 sack_state.rate = &rs;
3520 /* We very likely will need to access rtx queue. */
3521 prefetch(sk->tcp_rtx_queue.rb_node);
3523 /* If the ack is older than previous acks
3524 * then we can probably ignore it.
3526 if (before(ack, prior_snd_una)) {
3527 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3528 if (before(ack, prior_snd_una - tp->max_window)) {
3529 if (!(flag & FLAG_NO_CHALLENGE_ACK))
3530 tcp_send_challenge_ack(sk, skb);
3536 /* If the ack includes data we haven't sent yet, discard
3537 * this segment (RFC793 Section 3.9).
3539 if (after(ack, tp->snd_nxt))
3542 if (after(ack, prior_snd_una)) {
3543 flag |= FLAG_SND_UNA_ADVANCED;
3544 icsk->icsk_retransmits = 0;
3547 prior_fack = tcp_is_sack(tp) ? tcp_highest_sack_seq(tp) : tp->snd_una;
3548 rs.prior_in_flight = tcp_packets_in_flight(tp);
3550 /* ts_recent update must be made after we are sure that the packet
3553 if (flag & FLAG_UPDATE_TS_RECENT)
3554 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3556 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3557 /* Window is constant, pure forward advance.
3558 * No more checks are required.
3559 * Note, we use the fact that SND.UNA>=SND.WL2.
3561 tcp_update_wl(tp, ack_seq);
3562 tcp_snd_una_update(tp, ack);
3563 flag |= FLAG_WIN_UPDATE;
3565 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
3567 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPACKS);
3569 u32 ack_ev_flags = CA_ACK_SLOWPATH;
3571 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3574 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3576 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3578 if (TCP_SKB_CB(skb)->sacked)
3579 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3582 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) {
3584 ack_ev_flags |= CA_ACK_ECE;
3587 if (flag & FLAG_WIN_UPDATE)
3588 ack_ev_flags |= CA_ACK_WIN_UPDATE;
3590 tcp_in_ack_event(sk, ack_ev_flags);
3593 /* We passed data and got it acked, remove any soft error
3594 * log. Something worked...
3596 sk->sk_err_soft = 0;
3597 icsk->icsk_probes_out = 0;
3598 tp->rcv_tstamp = tcp_jiffies32;
3602 /* See if we can take anything off of the retransmit queue. */
3603 flag |= tcp_clean_rtx_queue(sk, prior_fack, prior_snd_una, &sack_state);
3605 tcp_rack_update_reo_wnd(sk, &rs);
3607 if (tp->tlp_high_seq)
3608 tcp_process_tlp_ack(sk, ack, flag);
3609 /* If needed, reset TLP/RTO timer; RACK may later override this. */
3610 if (flag & FLAG_SET_XMIT_TIMER)
3611 tcp_set_xmit_timer(sk);
3613 if (tcp_ack_is_dubious(sk, flag)) {
3614 is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
3615 tcp_fastretrans_alert(sk, prior_snd_una, is_dupack, &flag,
3619 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
3622 delivered = tp->delivered - delivered; /* freshly ACKed or SACKed */
3623 lost = tp->lost - lost; /* freshly marked lost */
3624 tcp_rate_gen(sk, delivered, lost, is_sack_reneg, sack_state.rate);
3625 tcp_cong_control(sk, ack, delivered, flag, sack_state.rate);
3626 tcp_xmit_recovery(sk, rexmit);
3630 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3631 if (flag & FLAG_DSACKING_ACK)
3632 tcp_fastretrans_alert(sk, prior_snd_una, is_dupack, &flag,
3634 /* If this ack opens up a zero window, clear backoff. It was
3635 * being used to time the probes, and is probably far higher than
3636 * it needs to be for normal retransmission.
3640 if (tp->tlp_high_seq)
3641 tcp_process_tlp_ack(sk, ack, flag);
3645 SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3649 /* If data was SACKed, tag it and see if we should send more data.
3650 * If data was DSACKed, see if we can undo a cwnd reduction.
3652 if (TCP_SKB_CB(skb)->sacked) {
3653 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3655 tcp_fastretrans_alert(sk, prior_snd_una, is_dupack, &flag,
3657 tcp_xmit_recovery(sk, rexmit);
3660 SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3664 static void tcp_parse_fastopen_option(int len, const unsigned char *cookie,
3665 bool syn, struct tcp_fastopen_cookie *foc,
3668 /* Valid only in SYN or SYN-ACK with an even length. */
3669 if (!foc || !syn || len < 0 || (len & 1))
3672 if (len >= TCP_FASTOPEN_COOKIE_MIN &&
3673 len <= TCP_FASTOPEN_COOKIE_MAX)
3674 memcpy(foc->val, cookie, len);
3681 static void smc_parse_options(const struct tcphdr *th,
3682 struct tcp_options_received *opt_rx,
3683 const unsigned char *ptr,
3686 #if IS_ENABLED(CONFIG_SMC)
3687 if (static_branch_unlikely(&tcp_have_smc)) {
3688 if (th->syn && !(opsize & 1) &&
3689 opsize >= TCPOLEN_EXP_SMC_BASE &&
3690 get_unaligned_be32(ptr) == TCPOPT_SMC_MAGIC)
3696 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3697 * But, this can also be called on packets in the established flow when
3698 * the fast version below fails.
3700 void tcp_parse_options(const struct net *net,
3701 const struct sk_buff *skb,
3702 struct tcp_options_received *opt_rx, int estab,
3703 struct tcp_fastopen_cookie *foc)
3705 const unsigned char *ptr;
3706 const struct tcphdr *th = tcp_hdr(skb);
3707 int length = (th->doff * 4) - sizeof(struct tcphdr);
3709 ptr = (const unsigned char *)(th + 1);
3710 opt_rx->saw_tstamp = 0;
3712 while (length > 0) {
3713 int opcode = *ptr++;
3719 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3724 if (opsize < 2) /* "silly options" */
3726 if (opsize > length)
3727 return; /* don't parse partial options */
3730 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3731 u16 in_mss = get_unaligned_be16(ptr);
3733 if (opt_rx->user_mss &&
3734 opt_rx->user_mss < in_mss)
3735 in_mss = opt_rx->user_mss;
3736 opt_rx->mss_clamp = in_mss;
3741 if (opsize == TCPOLEN_WINDOW && th->syn &&
3742 !estab && net->ipv4.sysctl_tcp_window_scaling) {
3743 __u8 snd_wscale = *(__u8 *)ptr;
3744 opt_rx->wscale_ok = 1;
3745 if (snd_wscale > TCP_MAX_WSCALE) {
3746 net_info_ratelimited("%s: Illegal window scaling value %d > %u received\n",
3750 snd_wscale = TCP_MAX_WSCALE;
3752 opt_rx->snd_wscale = snd_wscale;
3755 case TCPOPT_TIMESTAMP:
3756 if ((opsize == TCPOLEN_TIMESTAMP) &&
3757 ((estab && opt_rx->tstamp_ok) ||
3758 (!estab && net->ipv4.sysctl_tcp_timestamps))) {
3759 opt_rx->saw_tstamp = 1;
3760 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3761 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3764 case TCPOPT_SACK_PERM:
3765 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3766 !estab && net->ipv4.sysctl_tcp_sack) {
3767 opt_rx->sack_ok = TCP_SACK_SEEN;
3768 tcp_sack_reset(opt_rx);
3773 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3774 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3776 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3779 #ifdef CONFIG_TCP_MD5SIG
3782 * The MD5 Hash has already been
3783 * checked (see tcp_v{4,6}_do_rcv()).
3787 case TCPOPT_FASTOPEN:
3788 tcp_parse_fastopen_option(
3789 opsize - TCPOLEN_FASTOPEN_BASE,
3790 ptr, th->syn, foc, false);
3794 /* Fast Open option shares code 254 using a
3795 * 16 bits magic number.
3797 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE &&
3798 get_unaligned_be16(ptr) ==
3799 TCPOPT_FASTOPEN_MAGIC)
3800 tcp_parse_fastopen_option(opsize -
3801 TCPOLEN_EXP_FASTOPEN_BASE,
3802 ptr + 2, th->syn, foc, true);
3804 smc_parse_options(th, opt_rx, ptr,
3814 EXPORT_SYMBOL(tcp_parse_options);
3816 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
3818 const __be32 *ptr = (const __be32 *)(th + 1);
3820 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3821 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3822 tp->rx_opt.saw_tstamp = 1;
3824 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3827 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
3829 tp->rx_opt.rcv_tsecr = 0;
3835 /* Fast parse options. This hopes to only see timestamps.
3836 * If it is wrong it falls back on tcp_parse_options().
3838 static bool tcp_fast_parse_options(const struct net *net,
3839 const struct sk_buff *skb,
3840 const struct tcphdr *th, struct tcp_sock *tp)
3842 /* In the spirit of fast parsing, compare doff directly to constant
3843 * values. Because equality is used, short doff can be ignored here.
3845 if (th->doff == (sizeof(*th) / 4)) {
3846 tp->rx_opt.saw_tstamp = 0;
3848 } else if (tp->rx_opt.tstamp_ok &&
3849 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3850 if (tcp_parse_aligned_timestamp(tp, th))
3854 tcp_parse_options(net, skb, &tp->rx_opt, 1, NULL);
3855 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3856 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
3861 #ifdef CONFIG_TCP_MD5SIG
3863 * Parse MD5 Signature option
3865 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
3867 int length = (th->doff << 2) - sizeof(*th);
3868 const u8 *ptr = (const u8 *)(th + 1);
3870 /* If the TCP option is too short, we can short cut */
3871 if (length < TCPOLEN_MD5SIG)
3874 while (length > 0) {
3875 int opcode = *ptr++;
3886 if (opsize < 2 || opsize > length)
3888 if (opcode == TCPOPT_MD5SIG)
3889 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
3896 EXPORT_SYMBOL(tcp_parse_md5sig_option);
3899 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3901 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3902 * it can pass through stack. So, the following predicate verifies that
3903 * this segment is not used for anything but congestion avoidance or
3904 * fast retransmit. Moreover, we even are able to eliminate most of such
3905 * second order effects, if we apply some small "replay" window (~RTO)
3906 * to timestamp space.
3908 * All these measures still do not guarantee that we reject wrapped ACKs
3909 * on networks with high bandwidth, when sequence space is recycled fastly,
3910 * but it guarantees that such events will be very rare and do not affect
3911 * connection seriously. This doesn't look nice, but alas, PAWS is really
3914 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3915 * states that events when retransmit arrives after original data are rare.
3916 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3917 * the biggest problem on large power networks even with minor reordering.
3918 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3919 * up to bandwidth of 18Gigabit/sec. 8) ]
3922 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3924 const struct tcp_sock *tp = tcp_sk(sk);
3925 const struct tcphdr *th = tcp_hdr(skb);
3926 u32 seq = TCP_SKB_CB(skb)->seq;
3927 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3929 return (/* 1. Pure ACK with correct sequence number. */
3930 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3932 /* 2. ... and duplicate ACK. */
3933 ack == tp->snd_una &&
3935 /* 3. ... and does not update window. */
3936 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3938 /* 4. ... and sits in replay window. */
3939 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3942 static inline bool tcp_paws_discard(const struct sock *sk,
3943 const struct sk_buff *skb)
3945 const struct tcp_sock *tp = tcp_sk(sk);
3947 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
3948 !tcp_disordered_ack(sk, skb);
3951 /* Check segment sequence number for validity.
3953 * Segment controls are considered valid, if the segment
3954 * fits to the window after truncation to the window. Acceptability
3955 * of data (and SYN, FIN, of course) is checked separately.
3956 * See tcp_data_queue(), for example.
3958 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3959 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3960 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3961 * (borrowed from freebsd)
3964 static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
3966 return !before(end_seq, tp->rcv_wup) &&
3967 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3970 /* When we get a reset we do this. */
3971 void tcp_reset(struct sock *sk)
3973 trace_tcp_receive_reset(sk);
3975 /* We want the right error as BSD sees it (and indeed as we do). */
3976 switch (sk->sk_state) {
3978 sk->sk_err = ECONNREFUSED;
3980 case TCP_CLOSE_WAIT:
3986 sk->sk_err = ECONNRESET;
3988 /* This barrier is coupled with smp_rmb() in tcp_poll() */
3993 if (!sock_flag(sk, SOCK_DEAD))
3994 sk->sk_error_report(sk);
3998 * Process the FIN bit. This now behaves as it is supposed to work
3999 * and the FIN takes effect when it is validly part of sequence
4000 * space. Not before when we get holes.
4002 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4003 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4006 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4007 * close and we go into CLOSING (and later onto TIME-WAIT)
4009 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4011 void tcp_fin(struct sock *sk)
4013 struct tcp_sock *tp = tcp_sk(sk);
4015 inet_csk_schedule_ack(sk);
4017 sk->sk_shutdown |= RCV_SHUTDOWN;
4018 sock_set_flag(sk, SOCK_DONE);
4020 switch (sk->sk_state) {
4022 case TCP_ESTABLISHED:
4023 /* Move to CLOSE_WAIT */
4024 tcp_set_state(sk, TCP_CLOSE_WAIT);
4025 inet_csk(sk)->icsk_ack.pingpong = 1;
4028 case TCP_CLOSE_WAIT:
4030 /* Received a retransmission of the FIN, do
4035 /* RFC793: Remain in the LAST-ACK state. */
4039 /* This case occurs when a simultaneous close
4040 * happens, we must ack the received FIN and
4041 * enter the CLOSING state.
4044 tcp_set_state(sk, TCP_CLOSING);
4047 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4049 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4052 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4053 * cases we should never reach this piece of code.
4055 pr_err("%s: Impossible, sk->sk_state=%d\n",
4056 __func__, sk->sk_state);
4060 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4061 * Probably, we should reset in this case. For now drop them.
4063 skb_rbtree_purge(&tp->out_of_order_queue);
4064 if (tcp_is_sack(tp))
4065 tcp_sack_reset(&tp->rx_opt);
4068 if (!sock_flag(sk, SOCK_DEAD)) {
4069 sk->sk_state_change(sk);
4071 /* Do not send POLL_HUP for half duplex close. */
4072 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4073 sk->sk_state == TCP_CLOSE)
4074 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4076 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4080 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4083 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4084 if (before(seq, sp->start_seq))
4085 sp->start_seq = seq;
4086 if (after(end_seq, sp->end_seq))
4087 sp->end_seq = end_seq;
4093 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4095 struct tcp_sock *tp = tcp_sk(sk);
4097 if (tcp_is_sack(tp) && sock_net(sk)->ipv4.sysctl_tcp_dsack) {
4100 if (before(seq, tp->rcv_nxt))
4101 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4103 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4105 NET_INC_STATS(sock_net(sk), mib_idx);
4107 tp->rx_opt.dsack = 1;
4108 tp->duplicate_sack[0].start_seq = seq;
4109 tp->duplicate_sack[0].end_seq = end_seq;
4113 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4115 struct tcp_sock *tp = tcp_sk(sk);
4117 if (!tp->rx_opt.dsack)
4118 tcp_dsack_set(sk, seq, end_seq);
4120 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4123 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4125 struct tcp_sock *tp = tcp_sk(sk);
4127 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4128 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4129 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4130 tcp_enter_quickack_mode(sk);
4132 if (tcp_is_sack(tp) && sock_net(sk)->ipv4.sysctl_tcp_dsack) {
4133 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4135 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4136 end_seq = tp->rcv_nxt;
4137 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4144 /* These routines update the SACK block as out-of-order packets arrive or
4145 * in-order packets close up the sequence space.
4147 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4150 struct tcp_sack_block *sp = &tp->selective_acks[0];
4151 struct tcp_sack_block *swalk = sp + 1;
4153 /* See if the recent change to the first SACK eats into
4154 * or hits the sequence space of other SACK blocks, if so coalesce.
4156 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4157 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4160 /* Zap SWALK, by moving every further SACK up by one slot.
4161 * Decrease num_sacks.
4163 tp->rx_opt.num_sacks--;
4164 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4168 this_sack++, swalk++;
4172 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4174 struct tcp_sock *tp = tcp_sk(sk);
4175 struct tcp_sack_block *sp = &tp->selective_acks[0];
4176 int cur_sacks = tp->rx_opt.num_sacks;
4182 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4183 if (tcp_sack_extend(sp, seq, end_seq)) {
4184 /* Rotate this_sack to the first one. */
4185 for (; this_sack > 0; this_sack--, sp--)
4186 swap(*sp, *(sp - 1));
4188 tcp_sack_maybe_coalesce(tp);
4193 /* Could not find an adjacent existing SACK, build a new one,
4194 * put it at the front, and shift everyone else down. We
4195 * always know there is at least one SACK present already here.
4197 * If the sack array is full, forget about the last one.
4199 if (this_sack >= TCP_NUM_SACKS) {
4201 tp->rx_opt.num_sacks--;
4204 for (; this_sack > 0; this_sack--, sp--)
4208 /* Build the new head SACK, and we're done. */
4209 sp->start_seq = seq;
4210 sp->end_seq = end_seq;
4211 tp->rx_opt.num_sacks++;
4214 /* RCV.NXT advances, some SACKs should be eaten. */
4216 static void tcp_sack_remove(struct tcp_sock *tp)
4218 struct tcp_sack_block *sp = &tp->selective_acks[0];
4219 int num_sacks = tp->rx_opt.num_sacks;
4222 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4223 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4224 tp->rx_opt.num_sacks = 0;
4228 for (this_sack = 0; this_sack < num_sacks;) {
4229 /* Check if the start of the sack is covered by RCV.NXT. */
4230 if (!before(tp->rcv_nxt, sp->start_seq)) {
4233 /* RCV.NXT must cover all the block! */
4234 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4236 /* Zap this SACK, by moving forward any other SACKS. */
4237 for (i = this_sack+1; i < num_sacks; i++)
4238 tp->selective_acks[i-1] = tp->selective_acks[i];
4245 tp->rx_opt.num_sacks = num_sacks;
4249 * tcp_try_coalesce - try to merge skb to prior one
4251 * @dest: destination queue
4253 * @from: buffer to add in queue
4254 * @fragstolen: pointer to boolean
4256 * Before queueing skb @from after @to, try to merge them
4257 * to reduce overall memory use and queue lengths, if cost is small.
4258 * Packets in ofo or receive queues can stay a long time.
4259 * Better try to coalesce them right now to avoid future collapses.
4260 * Returns true if caller should free @from instead of queueing it
4262 static bool tcp_try_coalesce(struct sock *sk,
4264 struct sk_buff *from,
4269 *fragstolen = false;
4271 /* Its possible this segment overlaps with prior segment in queue */
4272 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4275 if (!skb_try_coalesce(to, from, fragstolen, &delta))
4278 atomic_add(delta, &sk->sk_rmem_alloc);
4279 sk_mem_charge(sk, delta);
4280 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4281 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4282 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4283 TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
4285 if (TCP_SKB_CB(from)->has_rxtstamp) {
4286 TCP_SKB_CB(to)->has_rxtstamp = true;
4287 to->tstamp = from->tstamp;
4293 static void tcp_drop(struct sock *sk, struct sk_buff *skb)
4295 sk_drops_add(sk, skb);
4299 /* This one checks to see if we can put data from the
4300 * out_of_order queue into the receive_queue.
4302 static void tcp_ofo_queue(struct sock *sk)
4304 struct tcp_sock *tp = tcp_sk(sk);
4305 __u32 dsack_high = tp->rcv_nxt;
4306 bool fin, fragstolen, eaten;
4307 struct sk_buff *skb, *tail;
4310 p = rb_first(&tp->out_of_order_queue);
4313 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4316 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4317 __u32 dsack = dsack_high;
4318 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4319 dsack_high = TCP_SKB_CB(skb)->end_seq;
4320 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4323 rb_erase(&skb->rbnode, &tp->out_of_order_queue);
4325 if (unlikely(!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))) {
4326 SOCK_DEBUG(sk, "ofo packet was already received\n");
4330 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4331 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4332 TCP_SKB_CB(skb)->end_seq);
4334 tail = skb_peek_tail(&sk->sk_receive_queue);
4335 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen);
4336 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4337 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
4339 __skb_queue_tail(&sk->sk_receive_queue, skb);
4341 kfree_skb_partial(skb, fragstolen);
4343 if (unlikely(fin)) {
4345 /* tcp_fin() purges tp->out_of_order_queue,
4346 * so we must end this loop right now.
4353 static bool tcp_prune_ofo_queue(struct sock *sk);
4354 static int tcp_prune_queue(struct sock *sk);
4356 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4359 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4360 !sk_rmem_schedule(sk, skb, size)) {
4362 if (tcp_prune_queue(sk) < 0)
4365 while (!sk_rmem_schedule(sk, skb, size)) {
4366 if (!tcp_prune_ofo_queue(sk))
4373 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4375 struct tcp_sock *tp = tcp_sk(sk);
4376 struct rb_node **p, *parent;
4377 struct sk_buff *skb1;
4381 tcp_ecn_check_ce(tp, skb);
4383 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4384 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFODROP);
4389 /* Disable header prediction. */
4391 inet_csk_schedule_ack(sk);
4393 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4394 seq = TCP_SKB_CB(skb)->seq;
4395 end_seq = TCP_SKB_CB(skb)->end_seq;
4396 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4397 tp->rcv_nxt, seq, end_seq);
4399 p = &tp->out_of_order_queue.rb_node;
4400 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4401 /* Initial out of order segment, build 1 SACK. */
4402 if (tcp_is_sack(tp)) {
4403 tp->rx_opt.num_sacks = 1;
4404 tp->selective_acks[0].start_seq = seq;
4405 tp->selective_acks[0].end_seq = end_seq;
4407 rb_link_node(&skb->rbnode, NULL, p);
4408 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4409 tp->ooo_last_skb = skb;
4413 /* In the typical case, we are adding an skb to the end of the list.
4414 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
4416 if (tcp_try_coalesce(sk, tp->ooo_last_skb,
4417 skb, &fragstolen)) {
4419 tcp_grow_window(sk, skb);
4420 kfree_skb_partial(skb, fragstolen);
4424 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
4425 if (!before(seq, TCP_SKB_CB(tp->ooo_last_skb)->end_seq)) {
4426 parent = &tp->ooo_last_skb->rbnode;
4427 p = &parent->rb_right;
4431 /* Find place to insert this segment. Handle overlaps on the way. */
4435 skb1 = rb_to_skb(parent);
4436 if (before(seq, TCP_SKB_CB(skb1)->seq)) {
4437 p = &parent->rb_left;
4440 if (before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4441 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4442 /* All the bits are present. Drop. */
4443 NET_INC_STATS(sock_net(sk),
4444 LINUX_MIB_TCPOFOMERGE);
4447 tcp_dsack_set(sk, seq, end_seq);
4450 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4451 /* Partial overlap. */
4452 tcp_dsack_set(sk, seq, TCP_SKB_CB(skb1)->end_seq);
4454 /* skb's seq == skb1's seq and skb covers skb1.
4455 * Replace skb1 with skb.
4457 rb_replace_node(&skb1->rbnode, &skb->rbnode,
4458 &tp->out_of_order_queue);
4459 tcp_dsack_extend(sk,
4460 TCP_SKB_CB(skb1)->seq,
4461 TCP_SKB_CB(skb1)->end_seq);
4462 NET_INC_STATS(sock_net(sk),
4463 LINUX_MIB_TCPOFOMERGE);
4467 } else if (tcp_try_coalesce(sk, skb1,
4468 skb, &fragstolen)) {
4471 p = &parent->rb_right;
4474 /* Insert segment into RB tree. */
4475 rb_link_node(&skb->rbnode, parent, p);
4476 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4479 /* Remove other segments covered by skb. */
4480 while ((skb1 = skb_rb_next(skb)) != NULL) {
4481 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4483 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4484 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4488 rb_erase(&skb1->rbnode, &tp->out_of_order_queue);
4489 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4490 TCP_SKB_CB(skb1)->end_seq);
4491 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4494 /* If there is no skb after us, we are the last_skb ! */
4496 tp->ooo_last_skb = skb;
4499 if (tcp_is_sack(tp))
4500 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4503 tcp_grow_window(sk, skb);
4505 skb_set_owner_r(skb, sk);
4509 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, int hdrlen,
4513 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4515 __skb_pull(skb, hdrlen);
4517 tcp_try_coalesce(sk, tail,
4518 skb, fragstolen)) ? 1 : 0;
4519 tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq);
4521 __skb_queue_tail(&sk->sk_receive_queue, skb);
4522 skb_set_owner_r(skb, sk);
4527 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4529 struct sk_buff *skb;
4537 if (size > PAGE_SIZE) {
4538 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS);
4540 data_len = npages << PAGE_SHIFT;
4541 size = data_len + (size & ~PAGE_MASK);
4543 skb = alloc_skb_with_frags(size - data_len, data_len,
4544 PAGE_ALLOC_COSTLY_ORDER,
4545 &err, sk->sk_allocation);
4549 skb_put(skb, size - data_len);
4550 skb->data_len = data_len;
4553 if (tcp_try_rmem_schedule(sk, skb, skb->truesize))
4556 err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size);
4560 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4561 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4562 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4564 if (tcp_queue_rcv(sk, skb, 0, &fragstolen)) {
4565 WARN_ON_ONCE(fragstolen); /* should not happen */
4577 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4579 struct tcp_sock *tp = tcp_sk(sk);
4583 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
4588 __skb_pull(skb, tcp_hdr(skb)->doff * 4);
4590 tcp_ecn_accept_cwr(tp, skb);
4592 tp->rx_opt.dsack = 0;
4594 /* Queue data for delivery to the user.
4595 * Packets in sequence go to the receive queue.
4596 * Out of sequence packets to the out_of_order_queue.
4598 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4599 if (tcp_receive_window(tp) == 0)
4602 /* Ok. In sequence. In window. */
4604 if (skb_queue_len(&sk->sk_receive_queue) == 0)
4605 sk_forced_mem_schedule(sk, skb->truesize);
4606 else if (tcp_try_rmem_schedule(sk, skb, skb->truesize))
4609 eaten = tcp_queue_rcv(sk, skb, 0, &fragstolen);
4610 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4612 tcp_event_data_recv(sk, skb);
4613 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
4616 if (!RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4619 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4620 * gap in queue is filled.
4622 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
4623 inet_csk(sk)->icsk_ack.pingpong = 0;
4626 if (tp->rx_opt.num_sacks)
4627 tcp_sack_remove(tp);
4629 tcp_fast_path_check(sk);
4632 kfree_skb_partial(skb, fragstolen);
4633 if (!sock_flag(sk, SOCK_DEAD))
4634 sk->sk_data_ready(sk);
4638 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4639 /* A retransmit, 2nd most common case. Force an immediate ack. */
4640 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4641 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4644 tcp_enter_quickack_mode(sk);
4645 inet_csk_schedule_ack(sk);
4651 /* Out of window. F.e. zero window probe. */
4652 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4655 tcp_enter_quickack_mode(sk);
4657 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4658 /* Partial packet, seq < rcv_next < end_seq */
4659 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4660 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4661 TCP_SKB_CB(skb)->end_seq);
4663 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4665 /* If window is closed, drop tail of packet. But after
4666 * remembering D-SACK for its head made in previous line.
4668 if (!tcp_receive_window(tp))
4673 tcp_data_queue_ofo(sk, skb);
4676 static struct sk_buff *tcp_skb_next(struct sk_buff *skb, struct sk_buff_head *list)
4679 return !skb_queue_is_last(list, skb) ? skb->next : NULL;
4681 return skb_rb_next(skb);
4684 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4685 struct sk_buff_head *list,
4686 struct rb_root *root)
4688 struct sk_buff *next = tcp_skb_next(skb, list);
4691 __skb_unlink(skb, list);
4693 rb_erase(&skb->rbnode, root);
4696 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4701 /* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */
4702 void tcp_rbtree_insert(struct rb_root *root, struct sk_buff *skb)
4704 struct rb_node **p = &root->rb_node;
4705 struct rb_node *parent = NULL;
4706 struct sk_buff *skb1;
4710 skb1 = rb_to_skb(parent);
4711 if (before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb1)->seq))
4712 p = &parent->rb_left;
4714 p = &parent->rb_right;
4716 rb_link_node(&skb->rbnode, parent, p);
4717 rb_insert_color(&skb->rbnode, root);
4720 /* Collapse contiguous sequence of skbs head..tail with
4721 * sequence numbers start..end.
4723 * If tail is NULL, this means until the end of the queue.
4725 * Segments with FIN/SYN are not collapsed (only because this
4729 tcp_collapse(struct sock *sk, struct sk_buff_head *list, struct rb_root *root,
4730 struct sk_buff *head, struct sk_buff *tail, u32 start, u32 end)
4732 struct sk_buff *skb = head, *n;
4733 struct sk_buff_head tmp;
4736 /* First, check that queue is collapsible and find
4737 * the point where collapsing can be useful.
4740 for (end_of_skbs = true; skb != NULL && skb != tail; skb = n) {
4741 n = tcp_skb_next(skb, list);
4743 /* No new bits? It is possible on ofo queue. */
4744 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4745 skb = tcp_collapse_one(sk, skb, list, root);
4751 /* The first skb to collapse is:
4753 * - bloated or contains data before "start" or
4754 * overlaps to the next one.
4756 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
4757 (tcp_win_from_space(sk, skb->truesize) > skb->len ||
4758 before(TCP_SKB_CB(skb)->seq, start))) {
4759 end_of_skbs = false;
4763 if (n && n != tail &&
4764 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(n)->seq) {
4765 end_of_skbs = false;
4769 /* Decided to skip this, advance start seq. */
4770 start = TCP_SKB_CB(skb)->end_seq;
4773 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4776 __skb_queue_head_init(&tmp);
4778 while (before(start, end)) {
4779 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
4780 struct sk_buff *nskb;
4782 nskb = alloc_skb(copy, GFP_ATOMIC);
4786 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4787 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4789 __skb_queue_before(list, skb, nskb);
4791 __skb_queue_tail(&tmp, nskb); /* defer rbtree insertion */
4792 skb_set_owner_r(nskb, sk);
4794 /* Copy data, releasing collapsed skbs. */
4796 int offset = start - TCP_SKB_CB(skb)->seq;
4797 int size = TCP_SKB_CB(skb)->end_seq - start;
4801 size = min(copy, size);
4802 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4804 TCP_SKB_CB(nskb)->end_seq += size;
4808 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4809 skb = tcp_collapse_one(sk, skb, list, root);
4812 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4818 skb_queue_walk_safe(&tmp, skb, n)
4819 tcp_rbtree_insert(root, skb);
4822 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4823 * and tcp_collapse() them until all the queue is collapsed.
4825 static void tcp_collapse_ofo_queue(struct sock *sk)
4827 struct tcp_sock *tp = tcp_sk(sk);
4828 struct sk_buff *skb, *head;
4831 skb = skb_rb_first(&tp->out_of_order_queue);
4834 tp->ooo_last_skb = skb_rb_last(&tp->out_of_order_queue);
4837 start = TCP_SKB_CB(skb)->seq;
4838 end = TCP_SKB_CB(skb)->end_seq;
4840 for (head = skb;;) {
4841 skb = skb_rb_next(skb);
4843 /* Range is terminated when we see a gap or when
4844 * we are at the queue end.
4847 after(TCP_SKB_CB(skb)->seq, end) ||
4848 before(TCP_SKB_CB(skb)->end_seq, start)) {
4849 tcp_collapse(sk, NULL, &tp->out_of_order_queue,
4850 head, skb, start, end);
4854 if (unlikely(before(TCP_SKB_CB(skb)->seq, start)))
4855 start = TCP_SKB_CB(skb)->seq;
4856 if (after(TCP_SKB_CB(skb)->end_seq, end))
4857 end = TCP_SKB_CB(skb)->end_seq;
4862 * Clean the out-of-order queue to make room.
4863 * We drop high sequences packets to :
4864 * 1) Let a chance for holes to be filled.
4865 * 2) not add too big latencies if thousands of packets sit there.
4866 * (But if application shrinks SO_RCVBUF, we could still end up
4867 * freeing whole queue here)
4869 * Return true if queue has shrunk.
4871 static bool tcp_prune_ofo_queue(struct sock *sk)
4873 struct tcp_sock *tp = tcp_sk(sk);
4874 struct rb_node *node, *prev;
4876 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
4879 NET_INC_STATS(sock_net(sk), LINUX_MIB_OFOPRUNED);
4880 node = &tp->ooo_last_skb->rbnode;
4882 prev = rb_prev(node);
4883 rb_erase(node, &tp->out_of_order_queue);
4884 tcp_drop(sk, rb_to_skb(node));
4886 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf &&
4887 !tcp_under_memory_pressure(sk))
4891 tp->ooo_last_skb = rb_to_skb(prev);
4893 /* Reset SACK state. A conforming SACK implementation will
4894 * do the same at a timeout based retransmit. When a connection
4895 * is in a sad state like this, we care only about integrity
4896 * of the connection not performance.
4898 if (tp->rx_opt.sack_ok)
4899 tcp_sack_reset(&tp->rx_opt);
4903 /* Reduce allocated memory if we can, trying to get
4904 * the socket within its memory limits again.
4906 * Return less than zero if we should start dropping frames
4907 * until the socket owning process reads some of the data
4908 * to stabilize the situation.
4910 static int tcp_prune_queue(struct sock *sk)
4912 struct tcp_sock *tp = tcp_sk(sk);
4914 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4916 NET_INC_STATS(sock_net(sk), LINUX_MIB_PRUNECALLED);
4918 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4919 tcp_clamp_window(sk);
4920 else if (tcp_under_memory_pressure(sk))
4921 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4923 tcp_collapse_ofo_queue(sk);
4924 if (!skb_queue_empty(&sk->sk_receive_queue))
4925 tcp_collapse(sk, &sk->sk_receive_queue, NULL,
4926 skb_peek(&sk->sk_receive_queue),
4928 tp->copied_seq, tp->rcv_nxt);
4931 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4934 /* Collapsing did not help, destructive actions follow.
4935 * This must not ever occur. */
4937 tcp_prune_ofo_queue(sk);
4939 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4942 /* If we are really being abused, tell the caller to silently
4943 * drop receive data on the floor. It will get retransmitted
4944 * and hopefully then we'll have sufficient space.
4946 NET_INC_STATS(sock_net(sk), LINUX_MIB_RCVPRUNED);
4948 /* Massive buffer overcommit. */
4953 static bool tcp_should_expand_sndbuf(const struct sock *sk)
4955 const struct tcp_sock *tp = tcp_sk(sk);
4957 /* If the user specified a specific send buffer setting, do
4960 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4963 /* If we are under global TCP memory pressure, do not expand. */
4964 if (tcp_under_memory_pressure(sk))
4967 /* If we are under soft global TCP memory pressure, do not expand. */
4968 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
4971 /* If we filled the congestion window, do not expand. */
4972 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
4978 /* When incoming ACK allowed to free some skb from write_queue,
4979 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4980 * on the exit from tcp input handler.
4982 * PROBLEM: sndbuf expansion does not work well with largesend.
4984 static void tcp_new_space(struct sock *sk)
4986 struct tcp_sock *tp = tcp_sk(sk);
4988 if (tcp_should_expand_sndbuf(sk)) {
4989 tcp_sndbuf_expand(sk);
4990 tp->snd_cwnd_stamp = tcp_jiffies32;
4993 sk->sk_write_space(sk);
4996 static void tcp_check_space(struct sock *sk)
4998 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
4999 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
5000 /* pairs with tcp_poll() */
5002 if (sk->sk_socket &&
5003 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
5005 if (!test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
5006 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
5011 static inline void tcp_data_snd_check(struct sock *sk)
5013 tcp_push_pending_frames(sk);
5014 tcp_check_space(sk);
5018 * Check if sending an ack is needed.
5020 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
5022 struct tcp_sock *tp = tcp_sk(sk);
5024 /* More than one full frame received... */
5025 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
5026 /* ... and right edge of window advances far enough.
5027 * (tcp_recvmsg() will send ACK otherwise). Or...
5029 __tcp_select_window(sk) >= tp->rcv_wnd) ||
5030 /* We ACK each frame or... */
5031 tcp_in_quickack_mode(sk) ||
5032 /* We have out of order data. */
5033 (ofo_possible && !RB_EMPTY_ROOT(&tp->out_of_order_queue))) {
5034 /* Then ack it now */
5037 /* Else, send delayed ack. */
5038 tcp_send_delayed_ack(sk);
5042 static inline void tcp_ack_snd_check(struct sock *sk)
5044 if (!inet_csk_ack_scheduled(sk)) {
5045 /* We sent a data segment already. */
5048 __tcp_ack_snd_check(sk, 1);
5052 * This routine is only called when we have urgent data
5053 * signaled. Its the 'slow' part of tcp_urg. It could be
5054 * moved inline now as tcp_urg is only called from one
5055 * place. We handle URGent data wrong. We have to - as
5056 * BSD still doesn't use the correction from RFC961.
5057 * For 1003.1g we should support a new option TCP_STDURG to permit
5058 * either form (or just set the sysctl tcp_stdurg).
5061 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
5063 struct tcp_sock *tp = tcp_sk(sk);
5064 u32 ptr = ntohs(th->urg_ptr);
5066 if (ptr && !sock_net(sk)->ipv4.sysctl_tcp_stdurg)
5068 ptr += ntohl(th->seq);
5070 /* Ignore urgent data that we've already seen and read. */
5071 if (after(tp->copied_seq, ptr))
5074 /* Do not replay urg ptr.
5076 * NOTE: interesting situation not covered by specs.
5077 * Misbehaving sender may send urg ptr, pointing to segment,
5078 * which we already have in ofo queue. We are not able to fetch
5079 * such data and will stay in TCP_URG_NOTYET until will be eaten
5080 * by recvmsg(). Seems, we are not obliged to handle such wicked
5081 * situations. But it is worth to think about possibility of some
5082 * DoSes using some hypothetical application level deadlock.
5084 if (before(ptr, tp->rcv_nxt))
5087 /* Do we already have a newer (or duplicate) urgent pointer? */
5088 if (tp->urg_data && !after(ptr, tp->urg_seq))
5091 /* Tell the world about our new urgent pointer. */
5094 /* We may be adding urgent data when the last byte read was
5095 * urgent. To do this requires some care. We cannot just ignore
5096 * tp->copied_seq since we would read the last urgent byte again
5097 * as data, nor can we alter copied_seq until this data arrives
5098 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5100 * NOTE. Double Dutch. Rendering to plain English: author of comment
5101 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5102 * and expect that both A and B disappear from stream. This is _wrong_.
5103 * Though this happens in BSD with high probability, this is occasional.
5104 * Any application relying on this is buggy. Note also, that fix "works"
5105 * only in this artificial test. Insert some normal data between A and B and we will
5106 * decline of BSD again. Verdict: it is better to remove to trap
5109 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5110 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5111 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5113 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5114 __skb_unlink(skb, &sk->sk_receive_queue);
5119 tp->urg_data = TCP_URG_NOTYET;
5122 /* Disable header prediction. */
5126 /* This is the 'fast' part of urgent handling. */
5127 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
5129 struct tcp_sock *tp = tcp_sk(sk);
5131 /* Check if we get a new urgent pointer - normally not. */
5133 tcp_check_urg(sk, th);
5135 /* Do we wait for any urgent data? - normally not... */
5136 if (tp->urg_data == TCP_URG_NOTYET) {
5137 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5140 /* Is the urgent pointer pointing into this packet? */
5141 if (ptr < skb->len) {
5143 if (skb_copy_bits(skb, ptr, &tmp, 1))
5145 tp->urg_data = TCP_URG_VALID | tmp;
5146 if (!sock_flag(sk, SOCK_DEAD))
5147 sk->sk_data_ready(sk);
5152 /* Accept RST for rcv_nxt - 1 after a FIN.
5153 * When tcp connections are abruptly terminated from Mac OSX (via ^C), a
5154 * FIN is sent followed by a RST packet. The RST is sent with the same
5155 * sequence number as the FIN, and thus according to RFC 5961 a challenge
5156 * ACK should be sent. However, Mac OSX rate limits replies to challenge
5157 * ACKs on the closed socket. In addition middleboxes can drop either the
5158 * challenge ACK or a subsequent RST.
5160 static bool tcp_reset_check(const struct sock *sk, const struct sk_buff *skb)
5162 struct tcp_sock *tp = tcp_sk(sk);
5164 return unlikely(TCP_SKB_CB(skb)->seq == (tp->rcv_nxt - 1) &&
5165 (1 << sk->sk_state) & (TCPF_CLOSE_WAIT | TCPF_LAST_ACK |
5169 /* Does PAWS and seqno based validation of an incoming segment, flags will
5170 * play significant role here.
5172 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5173 const struct tcphdr *th, int syn_inerr)
5175 struct tcp_sock *tp = tcp_sk(sk);
5176 bool rst_seq_match = false;
5178 /* RFC1323: H1. Apply PAWS check first. */
5179 if (tcp_fast_parse_options(sock_net(sk), skb, th, tp) &&
5180 tp->rx_opt.saw_tstamp &&
5181 tcp_paws_discard(sk, skb)) {
5183 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5184 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5185 LINUX_MIB_TCPACKSKIPPEDPAWS,
5186 &tp->last_oow_ack_time))
5187 tcp_send_dupack(sk, skb);
5190 /* Reset is accepted even if it did not pass PAWS. */
5193 /* Step 1: check sequence number */
5194 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5195 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5196 * (RST) segments are validated by checking their SEQ-fields."
5197 * And page 69: "If an incoming segment is not acceptable,
5198 * an acknowledgment should be sent in reply (unless the RST
5199 * bit is set, if so drop the segment and return)".
5204 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5205 LINUX_MIB_TCPACKSKIPPEDSEQ,
5206 &tp->last_oow_ack_time))
5207 tcp_send_dupack(sk, skb);
5208 } else if (tcp_reset_check(sk, skb)) {
5214 /* Step 2: check RST bit */
5216 /* RFC 5961 3.2 (extend to match against (RCV.NXT - 1) after a
5217 * FIN and SACK too if available):
5218 * If seq num matches RCV.NXT or (RCV.NXT - 1) after a FIN, or
5219 * the right-most SACK block,
5221 * RESET the connection
5223 * Send a challenge ACK
5225 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt ||
5226 tcp_reset_check(sk, skb)) {
5227 rst_seq_match = true;
5228 } else if (tcp_is_sack(tp) && tp->rx_opt.num_sacks > 0) {
5229 struct tcp_sack_block *sp = &tp->selective_acks[0];
5230 int max_sack = sp[0].end_seq;
5233 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;
5235 max_sack = after(sp[this_sack].end_seq,
5237 sp[this_sack].end_seq : max_sack;
5240 if (TCP_SKB_CB(skb)->seq == max_sack)
5241 rst_seq_match = true;
5247 /* Disable TFO if RST is out-of-order
5248 * and no data has been received
5249 * for current active TFO socket
5251 if (tp->syn_fastopen && !tp->data_segs_in &&
5252 sk->sk_state == TCP_ESTABLISHED)
5253 tcp_fastopen_active_disable(sk);
5254 tcp_send_challenge_ack(sk, skb);
5259 /* step 3: check security and precedence [ignored] */
5261 /* step 4: Check for a SYN
5262 * RFC 5961 4.2 : Send a challenge ack
5267 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5268 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5269 tcp_send_challenge_ack(sk, skb);
5281 * TCP receive function for the ESTABLISHED state.
5283 * It is split into a fast path and a slow path. The fast path is
5285 * - A zero window was announced from us - zero window probing
5286 * is only handled properly in the slow path.
5287 * - Out of order segments arrived.
5288 * - Urgent data is expected.
5289 * - There is no buffer space left
5290 * - Unexpected TCP flags/window values/header lengths are received
5291 * (detected by checking the TCP header against pred_flags)
5292 * - Data is sent in both directions. Fast path only supports pure senders
5293 * or pure receivers (this means either the sequence number or the ack
5294 * value must stay constant)
5295 * - Unexpected TCP option.
5297 * When these conditions are not satisfied it drops into a standard
5298 * receive procedure patterned after RFC793 to handle all cases.
5299 * The first three cases are guaranteed by proper pred_flags setting,
5300 * the rest is checked inline. Fast processing is turned on in
5301 * tcp_data_queue when everything is OK.
5303 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5304 const struct tcphdr *th)
5306 unsigned int len = skb->len;
5307 struct tcp_sock *tp = tcp_sk(sk);
5309 tcp_mstamp_refresh(tp);
5310 if (unlikely(!sk->sk_rx_dst))
5311 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
5313 * Header prediction.
5314 * The code loosely follows the one in the famous
5315 * "30 instruction TCP receive" Van Jacobson mail.
5317 * Van's trick is to deposit buffers into socket queue
5318 * on a device interrupt, to call tcp_recv function
5319 * on the receive process context and checksum and copy
5320 * the buffer to user space. smart...
5322 * Our current scheme is not silly either but we take the
5323 * extra cost of the net_bh soft interrupt processing...
5324 * We do checksum and copy also but from device to kernel.
5327 tp->rx_opt.saw_tstamp = 0;
5329 /* pred_flags is 0xS?10 << 16 + snd_wnd
5330 * if header_prediction is to be made
5331 * 'S' will always be tp->tcp_header_len >> 2
5332 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5333 * turn it off (when there are holes in the receive
5334 * space for instance)
5335 * PSH flag is ignored.
5338 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5339 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5340 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5341 int tcp_header_len = tp->tcp_header_len;
5343 /* Timestamp header prediction: tcp_header_len
5344 * is automatically equal to th->doff*4 due to pred_flags
5348 /* Check timestamp */
5349 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5350 /* No? Slow path! */
5351 if (!tcp_parse_aligned_timestamp(tp, th))
5354 /* If PAWS failed, check it more carefully in slow path */
5355 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5358 /* DO NOT update ts_recent here, if checksum fails
5359 * and timestamp was corrupted part, it will result
5360 * in a hung connection since we will drop all
5361 * future packets due to the PAWS test.
5365 if (len <= tcp_header_len) {
5366 /* Bulk data transfer: sender */
5367 if (len == tcp_header_len) {
5368 /* Predicted packet is in window by definition.
5369 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5370 * Hence, check seq<=rcv_wup reduces to:
5372 if (tcp_header_len ==
5373 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5374 tp->rcv_nxt == tp->rcv_wup)
5375 tcp_store_ts_recent(tp);
5377 /* We know that such packets are checksummed
5380 tcp_ack(sk, skb, 0);
5382 tcp_data_snd_check(sk);
5384 } else { /* Header too small */
5385 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5390 bool fragstolen = false;
5392 if (tcp_checksum_complete(skb))
5395 if ((int)skb->truesize > sk->sk_forward_alloc)
5398 /* Predicted packet is in window by definition.
5399 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5400 * Hence, check seq<=rcv_wup reduces to:
5402 if (tcp_header_len ==
5403 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5404 tp->rcv_nxt == tp->rcv_wup)
5405 tcp_store_ts_recent(tp);
5407 tcp_rcv_rtt_measure_ts(sk, skb);
5409 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPHITS);
5411 /* Bulk data transfer: receiver */
5412 eaten = tcp_queue_rcv(sk, skb, tcp_header_len,
5415 tcp_event_data_recv(sk, skb);
5417 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5418 /* Well, only one small jumplet in fast path... */
5419 tcp_ack(sk, skb, FLAG_DATA);
5420 tcp_data_snd_check(sk);
5421 if (!inet_csk_ack_scheduled(sk))
5425 __tcp_ack_snd_check(sk, 0);
5428 kfree_skb_partial(skb, fragstolen);
5429 sk->sk_data_ready(sk);
5435 if (len < (th->doff << 2) || tcp_checksum_complete(skb))
5438 if (!th->ack && !th->rst && !th->syn)
5442 * Standard slow path.
5445 if (!tcp_validate_incoming(sk, skb, th, 1))
5449 if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0)
5452 tcp_rcv_rtt_measure_ts(sk, skb);
5454 /* Process urgent data. */
5455 tcp_urg(sk, skb, th);
5457 /* step 7: process the segment text */
5458 tcp_data_queue(sk, skb);
5460 tcp_data_snd_check(sk);
5461 tcp_ack_snd_check(sk);
5465 TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS);
5466 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5471 EXPORT_SYMBOL(tcp_rcv_established);
5473 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5475 struct tcp_sock *tp = tcp_sk(sk);
5476 struct inet_connection_sock *icsk = inet_csk(sk);
5478 tcp_set_state(sk, TCP_ESTABLISHED);
5479 icsk->icsk_ack.lrcvtime = tcp_jiffies32;
5482 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
5483 security_inet_conn_established(sk, skb);
5486 tcp_init_transfer(sk, BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB);
5488 /* Prevent spurious tcp_cwnd_restart() on first data
5491 tp->lsndtime = tcp_jiffies32;
5493 if (sock_flag(sk, SOCK_KEEPOPEN))
5494 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5496 if (!tp->rx_opt.snd_wscale)
5497 __tcp_fast_path_on(tp, tp->snd_wnd);
5502 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
5503 struct tcp_fastopen_cookie *cookie)
5505 struct tcp_sock *tp = tcp_sk(sk);
5506 struct sk_buff *data = tp->syn_data ? tcp_rtx_queue_head(sk) : NULL;
5507 u16 mss = tp->rx_opt.mss_clamp, try_exp = 0;
5508 bool syn_drop = false;
5510 if (mss == tp->rx_opt.user_mss) {
5511 struct tcp_options_received opt;
5513 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5514 tcp_clear_options(&opt);
5515 opt.user_mss = opt.mss_clamp = 0;
5516 tcp_parse_options(sock_net(sk), synack, &opt, 0, NULL);
5517 mss = opt.mss_clamp;
5520 if (!tp->syn_fastopen) {
5521 /* Ignore an unsolicited cookie */
5523 } else if (tp->total_retrans) {
5524 /* SYN timed out and the SYN-ACK neither has a cookie nor
5525 * acknowledges data. Presumably the remote received only
5526 * the retransmitted (regular) SYNs: either the original
5527 * SYN-data or the corresponding SYN-ACK was dropped.
5529 syn_drop = (cookie->len < 0 && data);
5530 } else if (cookie->len < 0 && !tp->syn_data) {
5531 /* We requested a cookie but didn't get it. If we did not use
5532 * the (old) exp opt format then try so next time (try_exp=1).
5533 * Otherwise we go back to use the RFC7413 opt (try_exp=2).
5535 try_exp = tp->syn_fastopen_exp ? 2 : 1;
5538 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp);
5540 if (data) { /* Retransmit unacked data in SYN */
5541 skb_rbtree_walk_from(data) {
5542 if (__tcp_retransmit_skb(sk, data, 1))
5546 NET_INC_STATS(sock_net(sk),
5547 LINUX_MIB_TCPFASTOPENACTIVEFAIL);
5550 tp->syn_data_acked = tp->syn_data;
5551 if (tp->syn_data_acked)
5552 NET_INC_STATS(sock_net(sk),
5553 LINUX_MIB_TCPFASTOPENACTIVE);
5555 tcp_fastopen_add_skb(sk, synack);
5560 static void smc_check_reset_syn(struct tcp_sock *tp)
5562 #if IS_ENABLED(CONFIG_SMC)
5563 if (static_branch_unlikely(&tcp_have_smc)) {
5564 if (tp->syn_smc && !tp->rx_opt.smc_ok)
5570 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5571 const struct tcphdr *th)
5573 struct inet_connection_sock *icsk = inet_csk(sk);
5574 struct tcp_sock *tp = tcp_sk(sk);
5575 struct tcp_fastopen_cookie foc = { .len = -1 };
5576 int saved_clamp = tp->rx_opt.mss_clamp;
5579 tcp_parse_options(sock_net(sk), skb, &tp->rx_opt, 0, &foc);
5580 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
5581 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
5585 * "If the state is SYN-SENT then
5586 * first check the ACK bit
5587 * If the ACK bit is set
5588 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5589 * a reset (unless the RST bit is set, if so drop
5590 * the segment and return)"
5592 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
5593 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt))
5594 goto reset_and_undo;
5596 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5597 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5598 tcp_time_stamp(tp))) {
5599 NET_INC_STATS(sock_net(sk),
5600 LINUX_MIB_PAWSACTIVEREJECTED);
5601 goto reset_and_undo;
5604 /* Now ACK is acceptable.
5606 * "If the RST bit is set
5607 * If the ACK was acceptable then signal the user "error:
5608 * connection reset", drop the segment, enter CLOSED state,
5609 * delete TCB, and return."
5618 * "fifth, if neither of the SYN or RST bits is set then
5619 * drop the segment and return."
5625 goto discard_and_undo;
5628 * "If the SYN bit is on ...
5629 * are acceptable then ...
5630 * (our SYN has been ACKed), change the connection
5631 * state to ESTABLISHED..."
5634 tcp_ecn_rcv_synack(tp, th);
5636 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5637 tcp_ack(sk, skb, FLAG_SLOWPATH);
5639 /* Ok.. it's good. Set up sequence numbers and
5640 * move to established.
5642 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5643 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5645 /* RFC1323: The window in SYN & SYN/ACK segments is
5648 tp->snd_wnd = ntohs(th->window);
5650 if (!tp->rx_opt.wscale_ok) {
5651 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5652 tp->window_clamp = min(tp->window_clamp, 65535U);
5655 if (tp->rx_opt.saw_tstamp) {
5656 tp->rx_opt.tstamp_ok = 1;
5657 tp->tcp_header_len =
5658 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5659 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5660 tcp_store_ts_recent(tp);
5662 tp->tcp_header_len = sizeof(struct tcphdr);
5665 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5666 tcp_initialize_rcv_mss(sk);
5668 /* Remember, tcp_poll() does not lock socket!
5669 * Change state from SYN-SENT only after copied_seq
5670 * is initialized. */
5671 tp->copied_seq = tp->rcv_nxt;
5673 smc_check_reset_syn(tp);
5677 tcp_finish_connect(sk, skb);
5679 fastopen_fail = (tp->syn_fastopen || tp->syn_data) &&
5680 tcp_rcv_fastopen_synack(sk, skb, &foc);
5682 if (!sock_flag(sk, SOCK_DEAD)) {
5683 sk->sk_state_change(sk);
5684 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5688 if (sk->sk_write_pending ||
5689 icsk->icsk_accept_queue.rskq_defer_accept ||
5690 icsk->icsk_ack.pingpong) {
5691 /* Save one ACK. Data will be ready after
5692 * several ticks, if write_pending is set.
5694 * It may be deleted, but with this feature tcpdumps
5695 * look so _wonderfully_ clever, that I was not able
5696 * to stand against the temptation 8) --ANK
5698 inet_csk_schedule_ack(sk);
5699 tcp_enter_quickack_mode(sk);
5700 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5701 TCP_DELACK_MAX, TCP_RTO_MAX);
5712 /* No ACK in the segment */
5716 * "If the RST bit is set
5718 * Otherwise (no ACK) drop the segment and return."
5721 goto discard_and_undo;
5725 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5726 tcp_paws_reject(&tp->rx_opt, 0))
5727 goto discard_and_undo;
5730 /* We see SYN without ACK. It is attempt of
5731 * simultaneous connect with crossed SYNs.
5732 * Particularly, it can be connect to self.
5734 tcp_set_state(sk, TCP_SYN_RECV);
5736 if (tp->rx_opt.saw_tstamp) {
5737 tp->rx_opt.tstamp_ok = 1;
5738 tcp_store_ts_recent(tp);
5739 tp->tcp_header_len =
5740 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5742 tp->tcp_header_len = sizeof(struct tcphdr);
5745 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5746 tp->copied_seq = tp->rcv_nxt;
5747 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5749 /* RFC1323: The window in SYN & SYN/ACK segments is
5752 tp->snd_wnd = ntohs(th->window);
5753 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5754 tp->max_window = tp->snd_wnd;
5756 tcp_ecn_rcv_syn(tp, th);
5759 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5760 tcp_initialize_rcv_mss(sk);
5762 tcp_send_synack(sk);
5764 /* Note, we could accept data and URG from this segment.
5765 * There are no obstacles to make this (except that we must
5766 * either change tcp_recvmsg() to prevent it from returning data
5767 * before 3WHS completes per RFC793, or employ TCP Fast Open).
5769 * However, if we ignore data in ACKless segments sometimes,
5770 * we have no reasons to accept it sometimes.
5771 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5772 * is not flawless. So, discard packet for sanity.
5773 * Uncomment this return to process the data.
5780 /* "fifth, if neither of the SYN or RST bits is set then
5781 * drop the segment and return."
5785 tcp_clear_options(&tp->rx_opt);
5786 tp->rx_opt.mss_clamp = saved_clamp;
5790 tcp_clear_options(&tp->rx_opt);
5791 tp->rx_opt.mss_clamp = saved_clamp;
5796 * This function implements the receiving procedure of RFC 793 for
5797 * all states except ESTABLISHED and TIME_WAIT.
5798 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5799 * address independent.
5802 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb)
5804 struct tcp_sock *tp = tcp_sk(sk);
5805 struct inet_connection_sock *icsk = inet_csk(sk);
5806 const struct tcphdr *th = tcp_hdr(skb);
5807 struct request_sock *req;
5811 switch (sk->sk_state) {
5825 /* It is possible that we process SYN packets from backlog,
5826 * so we need to make sure to disable BH right there.
5829 acceptable = icsk->icsk_af_ops->conn_request(sk, skb) >= 0;
5840 tp->rx_opt.saw_tstamp = 0;
5841 tcp_mstamp_refresh(tp);
5842 queued = tcp_rcv_synsent_state_process(sk, skb, th);
5846 /* Do step6 onward by hand. */
5847 tcp_urg(sk, skb, th);
5849 tcp_data_snd_check(sk);
5853 tcp_mstamp_refresh(tp);
5854 tp->rx_opt.saw_tstamp = 0;
5855 req = tp->fastopen_rsk;
5857 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
5858 sk->sk_state != TCP_FIN_WAIT1);
5860 if (!tcp_check_req(sk, skb, req, true))
5864 if (!th->ack && !th->rst && !th->syn)
5867 if (!tcp_validate_incoming(sk, skb, th, 0))
5870 /* step 5: check the ACK field */
5871 acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
5872 FLAG_UPDATE_TS_RECENT |
5873 FLAG_NO_CHALLENGE_ACK) > 0;
5876 if (sk->sk_state == TCP_SYN_RECV)
5877 return 1; /* send one RST */
5878 tcp_send_challenge_ack(sk, skb);
5881 switch (sk->sk_state) {
5884 tcp_synack_rtt_meas(sk, req);
5886 /* Once we leave TCP_SYN_RECV, we no longer need req
5890 inet_csk(sk)->icsk_retransmits = 0;
5891 reqsk_fastopen_remove(sk, req, false);
5892 /* Re-arm the timer because data may have been sent out.
5893 * This is similar to the regular data transmission case
5894 * when new data has just been ack'ed.
5896 * (TFO) - we could try to be more aggressive and
5897 * retransmitting any data sooner based on when they
5902 tcp_init_transfer(sk, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB);
5903 tp->copied_seq = tp->rcv_nxt;
5906 tcp_set_state(sk, TCP_ESTABLISHED);
5907 sk->sk_state_change(sk);
5909 /* Note, that this wakeup is only for marginal crossed SYN case.
5910 * Passively open sockets are not waked up, because
5911 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
5914 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5916 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5917 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
5918 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5920 if (tp->rx_opt.tstamp_ok)
5921 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5923 if (!inet_csk(sk)->icsk_ca_ops->cong_control)
5924 tcp_update_pacing_rate(sk);
5926 /* Prevent spurious tcp_cwnd_restart() on first data packet */
5927 tp->lsndtime = tcp_jiffies32;
5929 tcp_initialize_rcv_mss(sk);
5930 tcp_fast_path_on(tp);
5933 case TCP_FIN_WAIT1: {
5936 /* If we enter the TCP_FIN_WAIT1 state and we are a
5937 * Fast Open socket and this is the first acceptable
5938 * ACK we have received, this would have acknowledged
5939 * our SYNACK so stop the SYNACK timer.
5942 /* We no longer need the request sock. */
5943 reqsk_fastopen_remove(sk, req, false);
5946 if (tp->snd_una != tp->write_seq)
5949 tcp_set_state(sk, TCP_FIN_WAIT2);
5950 sk->sk_shutdown |= SEND_SHUTDOWN;
5954 if (!sock_flag(sk, SOCK_DEAD)) {
5955 /* Wake up lingering close() */
5956 sk->sk_state_change(sk);
5960 if (tp->linger2 < 0) {
5962 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5965 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5966 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
5967 /* Receive out of order FIN after close() */
5968 if (tp->syn_fastopen && th->fin)
5969 tcp_fastopen_active_disable(sk);
5971 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5975 tmo = tcp_fin_time(sk);
5976 if (tmo > TCP_TIMEWAIT_LEN) {
5977 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5978 } else if (th->fin || sock_owned_by_user(sk)) {
5979 /* Bad case. We could lose such FIN otherwise.
5980 * It is not a big problem, but it looks confusing
5981 * and not so rare event. We still can lose it now,
5982 * if it spins in bh_lock_sock(), but it is really
5985 inet_csk_reset_keepalive_timer(sk, tmo);
5987 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5994 if (tp->snd_una == tp->write_seq) {
5995 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
6001 if (tp->snd_una == tp->write_seq) {
6002 tcp_update_metrics(sk);
6009 /* step 6: check the URG bit */
6010 tcp_urg(sk, skb, th);
6012 /* step 7: process the segment text */
6013 switch (sk->sk_state) {
6014 case TCP_CLOSE_WAIT:
6017 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
6022 /* RFC 793 says to queue data in these states,
6023 * RFC 1122 says we MUST send a reset.
6024 * BSD 4.4 also does reset.
6026 if (sk->sk_shutdown & RCV_SHUTDOWN) {
6027 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6028 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6029 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6035 case TCP_ESTABLISHED:
6036 tcp_data_queue(sk, skb);
6041 /* tcp_data could move socket to TIME-WAIT */
6042 if (sk->sk_state != TCP_CLOSE) {
6043 tcp_data_snd_check(sk);
6044 tcp_ack_snd_check(sk);
6053 EXPORT_SYMBOL(tcp_rcv_state_process);
6055 static inline void pr_drop_req(struct request_sock *req, __u16 port, int family)
6057 struct inet_request_sock *ireq = inet_rsk(req);
6059 if (family == AF_INET)
6060 net_dbg_ratelimited("drop open request from %pI4/%u\n",
6061 &ireq->ir_rmt_addr, port);
6062 #if IS_ENABLED(CONFIG_IPV6)
6063 else if (family == AF_INET6)
6064 net_dbg_ratelimited("drop open request from %pI6/%u\n",
6065 &ireq->ir_v6_rmt_addr, port);
6069 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
6071 * If we receive a SYN packet with these bits set, it means a
6072 * network is playing bad games with TOS bits. In order to
6073 * avoid possible false congestion notifications, we disable
6074 * TCP ECN negotiation.
6076 * Exception: tcp_ca wants ECN. This is required for DCTCP
6077 * congestion control: Linux DCTCP asserts ECT on all packets,
6078 * including SYN, which is most optimal solution; however,
6079 * others, such as FreeBSD do not.
6081 static void tcp_ecn_create_request(struct request_sock *req,
6082 const struct sk_buff *skb,
6083 const struct sock *listen_sk,
6084 const struct dst_entry *dst)
6086 const struct tcphdr *th = tcp_hdr(skb);
6087 const struct net *net = sock_net(listen_sk);
6088 bool th_ecn = th->ece && th->cwr;
6095 ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield);
6096 ecn_ok_dst = dst_feature(dst, DST_FEATURE_ECN_MASK);
6097 ecn_ok = net->ipv4.sysctl_tcp_ecn || ecn_ok_dst;
6099 if ((!ect && ecn_ok) || tcp_ca_needs_ecn(listen_sk) ||
6100 (ecn_ok_dst & DST_FEATURE_ECN_CA) ||
6101 tcp_bpf_ca_needs_ecn((struct sock *)req))
6102 inet_rsk(req)->ecn_ok = 1;
6105 static void tcp_openreq_init(struct request_sock *req,
6106 const struct tcp_options_received *rx_opt,
6107 struct sk_buff *skb, const struct sock *sk)
6109 struct inet_request_sock *ireq = inet_rsk(req);
6111 req->rsk_rcv_wnd = 0; /* So that tcp_send_synack() knows! */
6113 tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq;
6114 tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
6115 tcp_rsk(req)->snt_synack = tcp_clock_us();
6116 tcp_rsk(req)->last_oow_ack_time = 0;
6117 req->mss = rx_opt->mss_clamp;
6118 req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0;
6119 ireq->tstamp_ok = rx_opt->tstamp_ok;
6120 ireq->sack_ok = rx_opt->sack_ok;
6121 ireq->snd_wscale = rx_opt->snd_wscale;
6122 ireq->wscale_ok = rx_opt->wscale_ok;
6125 ireq->ir_rmt_port = tcp_hdr(skb)->source;
6126 ireq->ir_num = ntohs(tcp_hdr(skb)->dest);
6127 ireq->ir_mark = inet_request_mark(sk, skb);
6128 #if IS_ENABLED(CONFIG_SMC)
6129 ireq->smc_ok = rx_opt->smc_ok;
6133 struct request_sock *inet_reqsk_alloc(const struct request_sock_ops *ops,
6134 struct sock *sk_listener,
6135 bool attach_listener)
6137 struct request_sock *req = reqsk_alloc(ops, sk_listener,
6141 struct inet_request_sock *ireq = inet_rsk(req);
6143 ireq->ireq_opt = NULL;
6144 #if IS_ENABLED(CONFIG_IPV6)
6145 ireq->pktopts = NULL;
6147 atomic64_set(&ireq->ir_cookie, 0);
6148 ireq->ireq_state = TCP_NEW_SYN_RECV;
6149 write_pnet(&ireq->ireq_net, sock_net(sk_listener));
6150 ireq->ireq_family = sk_listener->sk_family;
6155 EXPORT_SYMBOL(inet_reqsk_alloc);
6158 * Return true if a syncookie should be sent
6160 static bool tcp_syn_flood_action(const struct sock *sk,
6161 const struct sk_buff *skb,
6164 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
6165 const char *msg = "Dropping request";
6166 bool want_cookie = false;
6167 struct net *net = sock_net(sk);
6169 #ifdef CONFIG_SYN_COOKIES
6170 if (net->ipv4.sysctl_tcp_syncookies) {
6171 msg = "Sending cookies";
6173 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES);
6176 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP);
6178 if (!queue->synflood_warned &&
6179 net->ipv4.sysctl_tcp_syncookies != 2 &&
6180 xchg(&queue->synflood_warned, 1) == 0)
6181 pr_info("%s: Possible SYN flooding on port %d. %s. Check SNMP counters.\n",
6182 proto, ntohs(tcp_hdr(skb)->dest), msg);
6187 static void tcp_reqsk_record_syn(const struct sock *sk,
6188 struct request_sock *req,
6189 const struct sk_buff *skb)
6191 if (tcp_sk(sk)->save_syn) {
6192 u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb);
6195 copy = kmalloc(len + sizeof(u32), GFP_ATOMIC);
6198 memcpy(©[1], skb_network_header(skb), len);
6199 req->saved_syn = copy;
6204 int tcp_conn_request(struct request_sock_ops *rsk_ops,
6205 const struct tcp_request_sock_ops *af_ops,
6206 struct sock *sk, struct sk_buff *skb)
6208 struct tcp_fastopen_cookie foc = { .len = -1 };
6209 __u32 isn = TCP_SKB_CB(skb)->tcp_tw_isn;
6210 struct tcp_options_received tmp_opt;
6211 struct tcp_sock *tp = tcp_sk(sk);
6212 struct net *net = sock_net(sk);
6213 struct sock *fastopen_sk = NULL;
6214 struct request_sock *req;
6215 bool want_cookie = false;
6216 struct dst_entry *dst;
6219 /* TW buckets are converted to open requests without
6220 * limitations, they conserve resources and peer is
6221 * evidently real one.
6223 if ((net->ipv4.sysctl_tcp_syncookies == 2 ||
6224 inet_csk_reqsk_queue_is_full(sk)) && !isn) {
6225 want_cookie = tcp_syn_flood_action(sk, skb, rsk_ops->slab_name);
6230 if (sk_acceptq_is_full(sk)) {
6231 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
6235 req = inet_reqsk_alloc(rsk_ops, sk, !want_cookie);
6239 tcp_rsk(req)->af_specific = af_ops;
6240 tcp_rsk(req)->ts_off = 0;
6242 tcp_clear_options(&tmp_opt);
6243 tmp_opt.mss_clamp = af_ops->mss_clamp;
6244 tmp_opt.user_mss = tp->rx_opt.user_mss;
6245 tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0,
6246 want_cookie ? NULL : &foc);
6248 if (want_cookie && !tmp_opt.saw_tstamp)
6249 tcp_clear_options(&tmp_opt);
6251 tmp_opt.tstamp_ok = tmp_opt.saw_tstamp;
6252 tcp_openreq_init(req, &tmp_opt, skb, sk);
6253 inet_rsk(req)->no_srccheck = inet_sk(sk)->transparent;
6255 /* Note: tcp_v6_init_req() might override ir_iif for link locals */
6256 inet_rsk(req)->ir_iif = inet_request_bound_dev_if(sk, skb);
6258 af_ops->init_req(req, sk, skb);
6260 if (security_inet_conn_request(sk, skb, req))
6263 if (tmp_opt.tstamp_ok)
6264 tcp_rsk(req)->ts_off = af_ops->init_ts_off(net, skb);
6266 dst = af_ops->route_req(sk, &fl, req);
6270 if (!want_cookie && !isn) {
6271 /* Kill the following clause, if you dislike this way. */
6272 if (!net->ipv4.sysctl_tcp_syncookies &&
6273 (net->ipv4.sysctl_max_syn_backlog - inet_csk_reqsk_queue_len(sk) <
6274 (net->ipv4.sysctl_max_syn_backlog >> 2)) &&
6275 !tcp_peer_is_proven(req, dst)) {
6276 /* Without syncookies last quarter of
6277 * backlog is filled with destinations,
6278 * proven to be alive.
6279 * It means that we continue to communicate
6280 * to destinations, already remembered
6281 * to the moment of synflood.
6283 pr_drop_req(req, ntohs(tcp_hdr(skb)->source),
6285 goto drop_and_release;
6288 isn = af_ops->init_seq(skb);
6291 tcp_ecn_create_request(req, skb, sk, dst);
6294 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss);
6295 req->cookie_ts = tmp_opt.tstamp_ok;
6296 if (!tmp_opt.tstamp_ok)
6297 inet_rsk(req)->ecn_ok = 0;
6300 tcp_rsk(req)->snt_isn = isn;
6301 tcp_rsk(req)->txhash = net_tx_rndhash();
6302 tcp_openreq_init_rwin(req, sk, dst);
6304 tcp_reqsk_record_syn(sk, req, skb);
6305 fastopen_sk = tcp_try_fastopen(sk, skb, req, &foc, dst);
6308 af_ops->send_synack(fastopen_sk, dst, &fl, req,
6309 &foc, TCP_SYNACK_FASTOPEN);
6310 /* Add the child socket directly into the accept queue */
6311 inet_csk_reqsk_queue_add(sk, req, fastopen_sk);
6312 sk->sk_data_ready(sk);
6313 bh_unlock_sock(fastopen_sk);
6314 sock_put(fastopen_sk);
6316 tcp_rsk(req)->tfo_listener = false;
6318 inet_csk_reqsk_queue_hash_add(sk, req,
6319 tcp_timeout_init((struct sock *)req));
6320 af_ops->send_synack(sk, dst, &fl, req, &foc,
6321 !want_cookie ? TCP_SYNACK_NORMAL :
6339 EXPORT_SYMBOL(tcp_conn_request);