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() */
100 #define FLAG_ACK_MAYBE_DELAYED 0x10000 /* Likely a delayed ACK */
102 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
103 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
104 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE|FLAG_DSACKING_ACK)
105 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
107 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
108 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
110 #define REXMIT_NONE 0 /* no loss recovery to do */
111 #define REXMIT_LOST 1 /* retransmit packets marked lost */
112 #define REXMIT_NEW 2 /* FRTO-style transmit of unsent/new packets */
114 static void tcp_gro_dev_warn(struct sock *sk, const struct sk_buff *skb,
117 static bool __once __read_mostly;
120 struct net_device *dev;
125 dev = dev_get_by_index_rcu(sock_net(sk), skb->skb_iif);
126 if (!dev || len >= dev->mtu)
127 pr_warn("%s: Driver has suspect GRO implementation, TCP performance may be compromised.\n",
128 dev ? dev->name : "Unknown driver");
133 /* Adapt the MSS value used to make delayed ack decision to the
136 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
138 struct inet_connection_sock *icsk = inet_csk(sk);
139 const unsigned int lss = icsk->icsk_ack.last_seg_size;
142 icsk->icsk_ack.last_seg_size = 0;
144 /* skb->len may jitter because of SACKs, even if peer
145 * sends good full-sized frames.
147 len = skb_shinfo(skb)->gso_size ? : skb->len;
148 if (len >= icsk->icsk_ack.rcv_mss) {
149 icsk->icsk_ack.rcv_mss = min_t(unsigned int, len,
151 /* Account for possibly-removed options */
152 if (unlikely(len > icsk->icsk_ack.rcv_mss +
153 MAX_TCP_OPTION_SPACE))
154 tcp_gro_dev_warn(sk, skb, len);
156 /* Otherwise, we make more careful check taking into account,
157 * that SACKs block is variable.
159 * "len" is invariant segment length, including TCP header.
161 len += skb->data - skb_transport_header(skb);
162 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
163 /* If PSH is not set, packet should be
164 * full sized, provided peer TCP is not badly broken.
165 * This observation (if it is correct 8)) allows
166 * to handle super-low mtu links fairly.
168 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
169 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
170 /* Subtract also invariant (if peer is RFC compliant),
171 * tcp header plus fixed timestamp option length.
172 * Resulting "len" is MSS free of SACK jitter.
174 len -= tcp_sk(sk)->tcp_header_len;
175 icsk->icsk_ack.last_seg_size = len;
177 icsk->icsk_ack.rcv_mss = len;
181 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
182 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
183 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
187 static void tcp_incr_quickack(struct sock *sk)
189 struct inet_connection_sock *icsk = inet_csk(sk);
190 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
194 if (quickacks > icsk->icsk_ack.quick)
195 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
198 static void tcp_enter_quickack_mode(struct sock *sk)
200 struct inet_connection_sock *icsk = inet_csk(sk);
201 tcp_incr_quickack(sk);
202 icsk->icsk_ack.pingpong = 0;
203 icsk->icsk_ack.ato = TCP_ATO_MIN;
206 /* Send ACKs quickly, if "quick" count is not exhausted
207 * and the session is not interactive.
210 static bool tcp_in_quickack_mode(struct sock *sk)
212 const struct inet_connection_sock *icsk = inet_csk(sk);
213 const struct dst_entry *dst = __sk_dst_get(sk);
215 return (dst && dst_metric(dst, RTAX_QUICKACK)) ||
216 (icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong);
219 static void tcp_ecn_queue_cwr(struct tcp_sock *tp)
221 if (tp->ecn_flags & TCP_ECN_OK)
222 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
225 static void tcp_ecn_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb)
227 if (tcp_hdr(skb)->cwr)
228 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
231 static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp)
233 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
236 static void __tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
238 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
239 case INET_ECN_NOT_ECT:
240 /* Funny extension: if ECT is not set on a segment,
241 * and we already seen ECT on a previous segment,
242 * it is probably a retransmit.
244 if (tp->ecn_flags & TCP_ECN_SEEN)
245 tcp_enter_quickack_mode((struct sock *)tp);
248 if (tcp_ca_needs_ecn((struct sock *)tp))
249 tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_IS_CE);
251 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
252 /* Better not delay acks, sender can have a very low cwnd */
253 tcp_enter_quickack_mode((struct sock *)tp);
254 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
256 tp->ecn_flags |= TCP_ECN_SEEN;
259 if (tcp_ca_needs_ecn((struct sock *)tp))
260 tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_NO_CE);
261 tp->ecn_flags |= TCP_ECN_SEEN;
266 static void tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
268 if (tp->ecn_flags & TCP_ECN_OK)
269 __tcp_ecn_check_ce(tp, skb);
272 static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
274 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
275 tp->ecn_flags &= ~TCP_ECN_OK;
278 static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
280 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
281 tp->ecn_flags &= ~TCP_ECN_OK;
284 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
286 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
291 /* Buffer size and advertised window tuning.
293 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
296 static void tcp_sndbuf_expand(struct sock *sk)
298 const struct tcp_sock *tp = tcp_sk(sk);
299 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
303 /* Worst case is non GSO/TSO : each frame consumes one skb
304 * and skb->head is kmalloced using power of two area of memory
306 per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
308 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
310 per_mss = roundup_pow_of_two(per_mss) +
311 SKB_DATA_ALIGN(sizeof(struct sk_buff));
313 nr_segs = max_t(u32, TCP_INIT_CWND, tp->snd_cwnd);
314 nr_segs = max_t(u32, nr_segs, tp->reordering + 1);
316 /* Fast Recovery (RFC 5681 3.2) :
317 * Cubic needs 1.7 factor, rounded to 2 to include
318 * extra cushion (application might react slowly to EPOLLOUT)
320 sndmem = ca_ops->sndbuf_expand ? ca_ops->sndbuf_expand(sk) : 2;
321 sndmem *= nr_segs * per_mss;
323 if (sk->sk_sndbuf < sndmem)
324 sk->sk_sndbuf = min(sndmem, sock_net(sk)->ipv4.sysctl_tcp_wmem[2]);
327 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
329 * All tcp_full_space() is split to two parts: "network" buffer, allocated
330 * forward and advertised in receiver window (tp->rcv_wnd) and
331 * "application buffer", required to isolate scheduling/application
332 * latencies from network.
333 * window_clamp is maximal advertised window. It can be less than
334 * tcp_full_space(), in this case tcp_full_space() - window_clamp
335 * is reserved for "application" buffer. The less window_clamp is
336 * the smoother our behaviour from viewpoint of network, but the lower
337 * throughput and the higher sensitivity of the connection to losses. 8)
339 * rcv_ssthresh is more strict window_clamp used at "slow start"
340 * phase to predict further behaviour of this connection.
341 * It is used for two goals:
342 * - to enforce header prediction at sender, even when application
343 * requires some significant "application buffer". It is check #1.
344 * - to prevent pruning of receive queue because of misprediction
345 * of receiver window. Check #2.
347 * The scheme does not work when sender sends good segments opening
348 * window and then starts to feed us spaghetti. But it should work
349 * in common situations. Otherwise, we have to rely on queue collapsing.
352 /* Slow part of check#2. */
353 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
355 struct tcp_sock *tp = tcp_sk(sk);
357 int truesize = tcp_win_from_space(sk, skb->truesize) >> 1;
358 int window = tcp_win_from_space(sk, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]) >> 1;
360 while (tp->rcv_ssthresh <= window) {
361 if (truesize <= skb->len)
362 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
370 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
372 struct tcp_sock *tp = tcp_sk(sk);
375 if (tp->rcv_ssthresh < tp->window_clamp &&
376 (int)tp->rcv_ssthresh < tcp_space(sk) &&
377 !tcp_under_memory_pressure(sk)) {
380 /* Check #2. Increase window, if skb with such overhead
381 * will fit to rcvbuf in future.
383 if (tcp_win_from_space(sk, skb->truesize) <= skb->len)
384 incr = 2 * tp->advmss;
386 incr = __tcp_grow_window(sk, skb);
389 incr = max_t(int, incr, 2 * skb->len);
390 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
392 inet_csk(sk)->icsk_ack.quick |= 1;
397 /* 3. Tuning rcvbuf, when connection enters established state. */
398 static void tcp_fixup_rcvbuf(struct sock *sk)
400 u32 mss = tcp_sk(sk)->advmss;
403 rcvmem = 2 * SKB_TRUESIZE(mss + MAX_TCP_HEADER) *
404 tcp_default_init_rwnd(mss);
406 /* Dynamic Right Sizing (DRS) has 2 to 3 RTT latency
407 * Allow enough cushion so that sender is not limited by our window
409 if (sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf)
412 if (sk->sk_rcvbuf < rcvmem)
413 sk->sk_rcvbuf = min(rcvmem, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
416 /* 4. Try to fixup all. It is made immediately after connection enters
419 void tcp_init_buffer_space(struct sock *sk)
421 int tcp_app_win = sock_net(sk)->ipv4.sysctl_tcp_app_win;
422 struct tcp_sock *tp = tcp_sk(sk);
425 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
426 tcp_fixup_rcvbuf(sk);
427 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
428 tcp_sndbuf_expand(sk);
430 tp->rcvq_space.space = tp->rcv_wnd;
431 tcp_mstamp_refresh(tp);
432 tp->rcvq_space.time = tp->tcp_mstamp;
433 tp->rcvq_space.seq = tp->copied_seq;
435 maxwin = tcp_full_space(sk);
437 if (tp->window_clamp >= maxwin) {
438 tp->window_clamp = maxwin;
440 if (tcp_app_win && maxwin > 4 * tp->advmss)
441 tp->window_clamp = max(maxwin -
442 (maxwin >> tcp_app_win),
446 /* Force reservation of one segment. */
448 tp->window_clamp > 2 * tp->advmss &&
449 tp->window_clamp + tp->advmss > maxwin)
450 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
452 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
453 tp->snd_cwnd_stamp = tcp_jiffies32;
456 /* 5. Recalculate window clamp after socket hit its memory bounds. */
457 static void tcp_clamp_window(struct sock *sk)
459 struct tcp_sock *tp = tcp_sk(sk);
460 struct inet_connection_sock *icsk = inet_csk(sk);
461 struct net *net = sock_net(sk);
463 icsk->icsk_ack.quick = 0;
465 if (sk->sk_rcvbuf < net->ipv4.sysctl_tcp_rmem[2] &&
466 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
467 !tcp_under_memory_pressure(sk) &&
468 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
469 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
470 net->ipv4.sysctl_tcp_rmem[2]);
472 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
473 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
476 /* Initialize RCV_MSS value.
477 * RCV_MSS is an our guess about MSS used by the peer.
478 * We haven't any direct information about the MSS.
479 * It's better to underestimate the RCV_MSS rather than overestimate.
480 * Overestimations make us ACKing less frequently than needed.
481 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
483 void tcp_initialize_rcv_mss(struct sock *sk)
485 const struct tcp_sock *tp = tcp_sk(sk);
486 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
488 hint = min(hint, tp->rcv_wnd / 2);
489 hint = min(hint, TCP_MSS_DEFAULT);
490 hint = max(hint, TCP_MIN_MSS);
492 inet_csk(sk)->icsk_ack.rcv_mss = hint;
494 EXPORT_SYMBOL(tcp_initialize_rcv_mss);
496 /* Receiver "autotuning" code.
498 * The algorithm for RTT estimation w/o timestamps is based on
499 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
500 * <http://public.lanl.gov/radiant/pubs.html#DRS>
502 * More detail on this code can be found at
503 * <http://staff.psc.edu/jheffner/>,
504 * though this reference is out of date. A new paper
507 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
509 u32 new_sample = tp->rcv_rtt_est.rtt_us;
512 if (new_sample != 0) {
513 /* If we sample in larger samples in the non-timestamp
514 * case, we could grossly overestimate the RTT especially
515 * with chatty applications or bulk transfer apps which
516 * are stalled on filesystem I/O.
518 * Also, since we are only going for a minimum in the
519 * non-timestamp case, we do not smooth things out
520 * else with timestamps disabled convergence takes too
524 m -= (new_sample >> 3);
532 /* No previous measure. */
536 tp->rcv_rtt_est.rtt_us = new_sample;
539 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
543 if (tp->rcv_rtt_est.time == 0)
545 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
547 delta_us = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcv_rtt_est.time);
550 tcp_rcv_rtt_update(tp, delta_us, 1);
553 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
554 tp->rcv_rtt_est.time = tp->tcp_mstamp;
557 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
558 const struct sk_buff *skb)
560 struct tcp_sock *tp = tcp_sk(sk);
562 if (tp->rx_opt.rcv_tsecr &&
563 (TCP_SKB_CB(skb)->end_seq -
564 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss)) {
565 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr;
570 delta_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
571 tcp_rcv_rtt_update(tp, delta_us, 0);
576 * This function should be called every time data is copied to user space.
577 * It calculates the appropriate TCP receive buffer space.
579 void tcp_rcv_space_adjust(struct sock *sk)
581 struct tcp_sock *tp = tcp_sk(sk);
585 tcp_mstamp_refresh(tp);
586 time = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcvq_space.time);
587 if (time < (tp->rcv_rtt_est.rtt_us >> 3) || tp->rcv_rtt_est.rtt_us == 0)
590 /* Number of bytes copied to user in last RTT */
591 copied = tp->copied_seq - tp->rcvq_space.seq;
592 if (copied <= tp->rcvq_space.space)
596 * copied = bytes received in previous RTT, our base window
597 * To cope with packet losses, we need a 2x factor
598 * To cope with slow start, and sender growing its cwin by 100 %
599 * every RTT, we need a 4x factor, because the ACK we are sending
600 * now is for the next RTT, not the current one :
601 * <prev RTT . ><current RTT .. ><next RTT .... >
604 if (sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf &&
605 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
609 /* minimal window to cope with packet losses, assuming
610 * steady state. Add some cushion because of small variations.
612 rcvwin = ((u64)copied << 1) + 16 * tp->advmss;
614 /* Accommodate for sender rate increase (eg. slow start) */
615 grow = rcvwin * (copied - tp->rcvq_space.space);
616 do_div(grow, tp->rcvq_space.space);
617 rcvwin += (grow << 1);
619 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
620 while (tcp_win_from_space(sk, rcvmem) < tp->advmss)
623 do_div(rcvwin, tp->advmss);
624 rcvbuf = min_t(u64, rcvwin * rcvmem,
625 sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
626 if (rcvbuf > sk->sk_rcvbuf) {
627 sk->sk_rcvbuf = rcvbuf;
629 /* Make the window clamp follow along. */
630 tp->window_clamp = tcp_win_from_space(sk, rcvbuf);
633 tp->rcvq_space.space = copied;
636 tp->rcvq_space.seq = tp->copied_seq;
637 tp->rcvq_space.time = tp->tcp_mstamp;
640 /* There is something which you must keep in mind when you analyze the
641 * behavior of the tp->ato delayed ack timeout interval. When a
642 * connection starts up, we want to ack as quickly as possible. The
643 * problem is that "good" TCP's do slow start at the beginning of data
644 * transmission. The means that until we send the first few ACK's the
645 * sender will sit on his end and only queue most of his data, because
646 * he can only send snd_cwnd unacked packets at any given time. For
647 * each ACK we send, he increments snd_cwnd and transmits more of his
650 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
652 struct tcp_sock *tp = tcp_sk(sk);
653 struct inet_connection_sock *icsk = inet_csk(sk);
656 inet_csk_schedule_ack(sk);
658 tcp_measure_rcv_mss(sk, skb);
660 tcp_rcv_rtt_measure(tp);
664 if (!icsk->icsk_ack.ato) {
665 /* The _first_ data packet received, initialize
666 * delayed ACK engine.
668 tcp_incr_quickack(sk);
669 icsk->icsk_ack.ato = TCP_ATO_MIN;
671 int m = now - icsk->icsk_ack.lrcvtime;
673 if (m <= TCP_ATO_MIN / 2) {
674 /* The fastest case is the first. */
675 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
676 } else if (m < icsk->icsk_ack.ato) {
677 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
678 if (icsk->icsk_ack.ato > icsk->icsk_rto)
679 icsk->icsk_ack.ato = icsk->icsk_rto;
680 } else if (m > icsk->icsk_rto) {
681 /* Too long gap. Apparently sender failed to
682 * restart window, so that we send ACKs quickly.
684 tcp_incr_quickack(sk);
688 icsk->icsk_ack.lrcvtime = now;
690 tcp_ecn_check_ce(tp, skb);
693 tcp_grow_window(sk, skb);
696 /* Called to compute a smoothed rtt estimate. The data fed to this
697 * routine either comes from timestamps, or from segments that were
698 * known _not_ to have been retransmitted [see Karn/Partridge
699 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
700 * piece by Van Jacobson.
701 * NOTE: the next three routines used to be one big routine.
702 * To save cycles in the RFC 1323 implementation it was better to break
703 * it up into three procedures. -- erics
705 static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
707 struct tcp_sock *tp = tcp_sk(sk);
708 long m = mrtt_us; /* RTT */
709 u32 srtt = tp->srtt_us;
711 /* The following amusing code comes from Jacobson's
712 * article in SIGCOMM '88. Note that rtt and mdev
713 * are scaled versions of rtt and mean deviation.
714 * This is designed to be as fast as possible
715 * m stands for "measurement".
717 * On a 1990 paper the rto value is changed to:
718 * RTO = rtt + 4 * mdev
720 * Funny. This algorithm seems to be very broken.
721 * These formulae increase RTO, when it should be decreased, increase
722 * too slowly, when it should be increased quickly, decrease too quickly
723 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
724 * does not matter how to _calculate_ it. Seems, it was trap
725 * that VJ failed to avoid. 8)
728 m -= (srtt >> 3); /* m is now error in rtt est */
729 srtt += m; /* rtt = 7/8 rtt + 1/8 new */
731 m = -m; /* m is now abs(error) */
732 m -= (tp->mdev_us >> 2); /* similar update on mdev */
733 /* This is similar to one of Eifel findings.
734 * Eifel blocks mdev updates when rtt decreases.
735 * This solution is a bit different: we use finer gain
736 * for mdev in this case (alpha*beta).
737 * Like Eifel it also prevents growth of rto,
738 * but also it limits too fast rto decreases,
739 * happening in pure Eifel.
744 m -= (tp->mdev_us >> 2); /* similar update on mdev */
746 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */
747 if (tp->mdev_us > tp->mdev_max_us) {
748 tp->mdev_max_us = tp->mdev_us;
749 if (tp->mdev_max_us > tp->rttvar_us)
750 tp->rttvar_us = tp->mdev_max_us;
752 if (after(tp->snd_una, tp->rtt_seq)) {
753 if (tp->mdev_max_us < tp->rttvar_us)
754 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
755 tp->rtt_seq = tp->snd_nxt;
756 tp->mdev_max_us = tcp_rto_min_us(sk);
759 /* no previous measure. */
760 srtt = m << 3; /* take the measured time to be rtt */
761 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */
762 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
763 tp->mdev_max_us = tp->rttvar_us;
764 tp->rtt_seq = tp->snd_nxt;
766 tp->srtt_us = max(1U, srtt);
769 static void tcp_update_pacing_rate(struct sock *sk)
771 const struct tcp_sock *tp = tcp_sk(sk);
774 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
775 rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3);
777 /* current rate is (cwnd * mss) / srtt
778 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
779 * In Congestion Avoidance phase, set it to 120 % the current rate.
781 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
782 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
783 * end of slow start and should slow down.
785 if (tp->snd_cwnd < tp->snd_ssthresh / 2)
786 rate *= sock_net(sk)->ipv4.sysctl_tcp_pacing_ss_ratio;
788 rate *= sock_net(sk)->ipv4.sysctl_tcp_pacing_ca_ratio;
790 rate *= max(tp->snd_cwnd, tp->packets_out);
792 if (likely(tp->srtt_us))
793 do_div(rate, tp->srtt_us);
795 /* WRITE_ONCE() is needed because sch_fq fetches sk_pacing_rate
796 * without any lock. We want to make sure compiler wont store
797 * intermediate values in this location.
799 WRITE_ONCE(sk->sk_pacing_rate, min_t(u64, rate,
800 sk->sk_max_pacing_rate));
803 /* Calculate rto without backoff. This is the second half of Van Jacobson's
804 * routine referred to above.
806 static void tcp_set_rto(struct sock *sk)
808 const struct tcp_sock *tp = tcp_sk(sk);
809 /* Old crap is replaced with new one. 8)
812 * 1. If rtt variance happened to be less 50msec, it is hallucination.
813 * It cannot be less due to utterly erratic ACK generation made
814 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
815 * to do with delayed acks, because at cwnd>2 true delack timeout
816 * is invisible. Actually, Linux-2.4 also generates erratic
817 * ACKs in some circumstances.
819 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
821 /* 2. Fixups made earlier cannot be right.
822 * If we do not estimate RTO correctly without them,
823 * all the algo is pure shit and should be replaced
824 * with correct one. It is exactly, which we pretend to do.
827 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
828 * guarantees that rto is higher.
833 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
835 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
838 cwnd = TCP_INIT_CWND;
839 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
842 /* Take a notice that peer is sending D-SACKs */
843 static void tcp_dsack_seen(struct tcp_sock *tp)
845 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
846 tp->rack.dsack_seen = 1;
849 /* It's reordering when higher sequence was delivered (i.e. sacked) before
850 * some lower never-retransmitted sequence ("low_seq"). The maximum reordering
851 * distance is approximated in full-mss packet distance ("reordering").
853 static void tcp_check_sack_reordering(struct sock *sk, const u32 low_seq,
856 struct tcp_sock *tp = tcp_sk(sk);
857 const u32 mss = tp->mss_cache;
860 fack = tcp_highest_sack_seq(tp);
861 if (!before(low_seq, fack))
864 metric = fack - low_seq;
865 if ((metric > tp->reordering * mss) && mss) {
866 #if FASTRETRANS_DEBUG > 1
867 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
868 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
872 tp->undo_marker ? tp->undo_retrans : 0);
874 tp->reordering = min_t(u32, (metric + mss - 1) / mss,
875 sock_net(sk)->ipv4.sysctl_tcp_max_reordering);
879 /* This exciting event is worth to be remembered. 8) */
880 NET_INC_STATS(sock_net(sk),
881 ts ? LINUX_MIB_TCPTSREORDER : LINUX_MIB_TCPSACKREORDER);
884 /* This must be called before lost_out is incremented */
885 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
887 if (!tp->retransmit_skb_hint ||
888 before(TCP_SKB_CB(skb)->seq,
889 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
890 tp->retransmit_skb_hint = skb;
893 /* Sum the number of packets on the wire we have marked as lost.
894 * There are two cases we care about here:
895 * a) Packet hasn't been marked lost (nor retransmitted),
896 * and this is the first loss.
897 * b) Packet has been marked both lost and retransmitted,
898 * and this means we think it was lost again.
900 static void tcp_sum_lost(struct tcp_sock *tp, struct sk_buff *skb)
902 __u8 sacked = TCP_SKB_CB(skb)->sacked;
904 if (!(sacked & TCPCB_LOST) ||
905 ((sacked & TCPCB_LOST) && (sacked & TCPCB_SACKED_RETRANS)))
906 tp->lost += tcp_skb_pcount(skb);
909 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
911 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
912 tcp_verify_retransmit_hint(tp, skb);
914 tp->lost_out += tcp_skb_pcount(skb);
915 tcp_sum_lost(tp, skb);
916 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
920 void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp, struct sk_buff *skb)
922 tcp_verify_retransmit_hint(tp, skb);
924 tcp_sum_lost(tp, skb);
925 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
926 tp->lost_out += tcp_skb_pcount(skb);
927 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
931 /* This procedure tags the retransmission queue when SACKs arrive.
933 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
934 * Packets in queue with these bits set are counted in variables
935 * sacked_out, retrans_out and lost_out, correspondingly.
937 * Valid combinations are:
938 * Tag InFlight Description
939 * 0 1 - orig segment is in flight.
940 * S 0 - nothing flies, orig reached receiver.
941 * L 0 - nothing flies, orig lost by net.
942 * R 2 - both orig and retransmit are in flight.
943 * L|R 1 - orig is lost, retransmit is in flight.
944 * S|R 1 - orig reached receiver, retrans is still in flight.
945 * (L|S|R is logically valid, it could occur when L|R is sacked,
946 * but it is equivalent to plain S and code short-curcuits it to S.
947 * L|S is logically invalid, it would mean -1 packet in flight 8))
949 * These 6 states form finite state machine, controlled by the following events:
950 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
951 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
952 * 3. Loss detection event of two flavors:
953 * A. Scoreboard estimator decided the packet is lost.
954 * A'. Reno "three dupacks" marks head of queue lost.
955 * B. SACK arrives sacking SND.NXT at the moment, when the
956 * segment was retransmitted.
957 * 4. D-SACK added new rule: D-SACK changes any tag to S.
959 * It is pleasant to note, that state diagram turns out to be commutative,
960 * so that we are allowed not to be bothered by order of our actions,
961 * when multiple events arrive simultaneously. (see the function below).
963 * Reordering detection.
964 * --------------------
965 * Reordering metric is maximal distance, which a packet can be displaced
966 * in packet stream. With SACKs we can estimate it:
968 * 1. SACK fills old hole and the corresponding segment was not
969 * ever retransmitted -> reordering. Alas, we cannot use it
970 * when segment was retransmitted.
971 * 2. The last flaw is solved with D-SACK. D-SACK arrives
972 * for retransmitted and already SACKed segment -> reordering..
973 * Both of these heuristics are not used in Loss state, when we cannot
974 * account for retransmits accurately.
976 * SACK block validation.
977 * ----------------------
979 * SACK block range validation checks that the received SACK block fits to
980 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
981 * Note that SND.UNA is not included to the range though being valid because
982 * it means that the receiver is rather inconsistent with itself reporting
983 * SACK reneging when it should advance SND.UNA. Such SACK block this is
984 * perfectly valid, however, in light of RFC2018 which explicitly states
985 * that "SACK block MUST reflect the newest segment. Even if the newest
986 * segment is going to be discarded ...", not that it looks very clever
987 * in case of head skb. Due to potentional receiver driven attacks, we
988 * choose to avoid immediate execution of a walk in write queue due to
989 * reneging and defer head skb's loss recovery to standard loss recovery
990 * procedure that will eventually trigger (nothing forbids us doing this).
992 * Implements also blockage to start_seq wrap-around. Problem lies in the
993 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
994 * there's no guarantee that it will be before snd_nxt (n). The problem
995 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
998 * <- outs wnd -> <- wrapzone ->
999 * u e n u_w e_w s n_w
1001 * |<------------+------+----- TCP seqno space --------------+---------->|
1002 * ...-- <2^31 ->| |<--------...
1003 * ...---- >2^31 ------>| |<--------...
1005 * Current code wouldn't be vulnerable but it's better still to discard such
1006 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1007 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1008 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1009 * equal to the ideal case (infinite seqno space without wrap caused issues).
1011 * With D-SACK the lower bound is extended to cover sequence space below
1012 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1013 * again, D-SACK block must not to go across snd_una (for the same reason as
1014 * for the normal SACK blocks, explained above). But there all simplicity
1015 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1016 * fully below undo_marker they do not affect behavior in anyway and can
1017 * therefore be safely ignored. In rare cases (which are more or less
1018 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1019 * fragmentation and packet reordering past skb's retransmission. To consider
1020 * them correctly, the acceptable range must be extended even more though
1021 * the exact amount is rather hard to quantify. However, tp->max_window can
1022 * be used as an exaggerated estimate.
1024 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
1025 u32 start_seq, u32 end_seq)
1027 /* Too far in future, or reversed (interpretation is ambiguous) */
1028 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1031 /* Nasty start_seq wrap-around check (see comments above) */
1032 if (!before(start_seq, tp->snd_nxt))
1035 /* In outstanding window? ...This is valid exit for D-SACKs too.
1036 * start_seq == snd_una is non-sensical (see comments above)
1038 if (after(start_seq, tp->snd_una))
1041 if (!is_dsack || !tp->undo_marker)
1044 /* ...Then it's D-SACK, and must reside below snd_una completely */
1045 if (after(end_seq, tp->snd_una))
1048 if (!before(start_seq, tp->undo_marker))
1052 if (!after(end_seq, tp->undo_marker))
1055 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1056 * start_seq < undo_marker and end_seq >= undo_marker.
1058 return !before(start_seq, end_seq - tp->max_window);
1061 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1062 struct tcp_sack_block_wire *sp, int num_sacks,
1065 struct tcp_sock *tp = tcp_sk(sk);
1066 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1067 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1068 bool dup_sack = false;
1070 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1073 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1074 } else if (num_sacks > 1) {
1075 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1076 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1078 if (!after(end_seq_0, end_seq_1) &&
1079 !before(start_seq_0, start_seq_1)) {
1082 NET_INC_STATS(sock_net(sk),
1083 LINUX_MIB_TCPDSACKOFORECV);
1087 /* D-SACK for already forgotten data... Do dumb counting. */
1088 if (dup_sack && tp->undo_marker && tp->undo_retrans > 0 &&
1089 !after(end_seq_0, prior_snd_una) &&
1090 after(end_seq_0, tp->undo_marker))
1096 struct tcp_sacktag_state {
1098 /* Timestamps for earliest and latest never-retransmitted segment
1099 * that was SACKed. RTO needs the earliest RTT to stay conservative,
1100 * but congestion control should still get an accurate delay signal.
1104 struct rate_sample *rate;
1106 unsigned int mss_now;
1109 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1110 * the incoming SACK may not exactly match but we can find smaller MSS
1111 * aligned portion of it that matches. Therefore we might need to fragment
1112 * which may fail and creates some hassle (caller must handle error case
1115 * FIXME: this could be merged to shift decision code
1117 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1118 u32 start_seq, u32 end_seq)
1122 unsigned int pkt_len;
1125 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1126 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1128 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1129 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1130 mss = tcp_skb_mss(skb);
1131 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1134 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1138 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1143 /* Round if necessary so that SACKs cover only full MSSes
1144 * and/or the remaining small portion (if present)
1146 if (pkt_len > mss) {
1147 unsigned int new_len = (pkt_len / mss) * mss;
1148 if (!in_sack && new_len < pkt_len)
1153 if (pkt_len >= skb->len && !in_sack)
1156 err = tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
1157 pkt_len, mss, GFP_ATOMIC);
1165 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1166 static u8 tcp_sacktag_one(struct sock *sk,
1167 struct tcp_sacktag_state *state, u8 sacked,
1168 u32 start_seq, u32 end_seq,
1169 int dup_sack, int pcount,
1172 struct tcp_sock *tp = tcp_sk(sk);
1174 /* Account D-SACK for retransmitted packet. */
1175 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1176 if (tp->undo_marker && tp->undo_retrans > 0 &&
1177 after(end_seq, tp->undo_marker))
1179 if ((sacked & TCPCB_SACKED_ACKED) &&
1180 before(start_seq, state->reord))
1181 state->reord = start_seq;
1184 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1185 if (!after(end_seq, tp->snd_una))
1188 if (!(sacked & TCPCB_SACKED_ACKED)) {
1189 tcp_rack_advance(tp, sacked, end_seq, xmit_time);
1191 if (sacked & TCPCB_SACKED_RETRANS) {
1192 /* If the segment is not tagged as lost,
1193 * we do not clear RETRANS, believing
1194 * that retransmission is still in flight.
1196 if (sacked & TCPCB_LOST) {
1197 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1198 tp->lost_out -= pcount;
1199 tp->retrans_out -= pcount;
1202 if (!(sacked & TCPCB_RETRANS)) {
1203 /* New sack for not retransmitted frame,
1204 * which was in hole. It is reordering.
1206 if (before(start_seq,
1207 tcp_highest_sack_seq(tp)) &&
1208 before(start_seq, state->reord))
1209 state->reord = start_seq;
1211 if (!after(end_seq, tp->high_seq))
1212 state->flag |= FLAG_ORIG_SACK_ACKED;
1213 if (state->first_sackt == 0)
1214 state->first_sackt = xmit_time;
1215 state->last_sackt = xmit_time;
1218 if (sacked & TCPCB_LOST) {
1219 sacked &= ~TCPCB_LOST;
1220 tp->lost_out -= pcount;
1224 sacked |= TCPCB_SACKED_ACKED;
1225 state->flag |= FLAG_DATA_SACKED;
1226 tp->sacked_out += pcount;
1227 tp->delivered += pcount; /* Out-of-order packets delivered */
1229 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1230 if (tp->lost_skb_hint &&
1231 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1232 tp->lost_cnt_hint += pcount;
1235 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1236 * frames and clear it. undo_retrans is decreased above, L|R frames
1237 * are accounted above as well.
1239 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1240 sacked &= ~TCPCB_SACKED_RETRANS;
1241 tp->retrans_out -= pcount;
1247 /* Shift newly-SACKed bytes from this skb to the immediately previous
1248 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1250 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *prev,
1251 struct sk_buff *skb,
1252 struct tcp_sacktag_state *state,
1253 unsigned int pcount, int shifted, int mss,
1256 struct tcp_sock *tp = tcp_sk(sk);
1257 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1258 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
1262 /* Adjust counters and hints for the newly sacked sequence
1263 * range but discard the return value since prev is already
1264 * marked. We must tag the range first because the seq
1265 * advancement below implicitly advances
1266 * tcp_highest_sack_seq() when skb is highest_sack.
1268 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1269 start_seq, end_seq, dup_sack, pcount,
1271 tcp_rate_skb_delivered(sk, skb, state->rate);
1273 if (skb == tp->lost_skb_hint)
1274 tp->lost_cnt_hint += pcount;
1276 TCP_SKB_CB(prev)->end_seq += shifted;
1277 TCP_SKB_CB(skb)->seq += shifted;
1279 tcp_skb_pcount_add(prev, pcount);
1280 BUG_ON(tcp_skb_pcount(skb) < pcount);
1281 tcp_skb_pcount_add(skb, -pcount);
1283 /* When we're adding to gso_segs == 1, gso_size will be zero,
1284 * in theory this shouldn't be necessary but as long as DSACK
1285 * code can come after this skb later on it's better to keep
1286 * setting gso_size to something.
1288 if (!TCP_SKB_CB(prev)->tcp_gso_size)
1289 TCP_SKB_CB(prev)->tcp_gso_size = mss;
1291 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1292 if (tcp_skb_pcount(skb) <= 1)
1293 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1295 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1296 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1299 BUG_ON(!tcp_skb_pcount(skb));
1300 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1304 /* Whole SKB was eaten :-) */
1306 if (skb == tp->retransmit_skb_hint)
1307 tp->retransmit_skb_hint = prev;
1308 if (skb == tp->lost_skb_hint) {
1309 tp->lost_skb_hint = prev;
1310 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1313 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1314 TCP_SKB_CB(prev)->eor = TCP_SKB_CB(skb)->eor;
1315 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1316 TCP_SKB_CB(prev)->end_seq++;
1318 if (skb == tcp_highest_sack(sk))
1319 tcp_advance_highest_sack(sk, skb);
1321 tcp_skb_collapse_tstamp(prev, skb);
1322 if (unlikely(TCP_SKB_CB(prev)->tx.delivered_mstamp))
1323 TCP_SKB_CB(prev)->tx.delivered_mstamp = 0;
1325 tcp_rtx_queue_unlink_and_free(skb, sk);
1327 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKMERGED);
1332 /* I wish gso_size would have a bit more sane initialization than
1333 * something-or-zero which complicates things
1335 static int tcp_skb_seglen(const struct sk_buff *skb)
1337 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1340 /* Shifting pages past head area doesn't work */
1341 static int skb_can_shift(const struct sk_buff *skb)
1343 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1346 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1349 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1350 struct tcp_sacktag_state *state,
1351 u32 start_seq, u32 end_seq,
1354 struct tcp_sock *tp = tcp_sk(sk);
1355 struct sk_buff *prev;
1361 if (!sk_can_gso(sk))
1364 /* Normally R but no L won't result in plain S */
1366 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1368 if (!skb_can_shift(skb))
1370 /* This frame is about to be dropped (was ACKed). */
1371 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1374 /* Can only happen with delayed DSACK + discard craziness */
1375 prev = skb_rb_prev(skb);
1379 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1382 if (!tcp_skb_can_collapse_to(prev))
1385 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1386 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1390 pcount = tcp_skb_pcount(skb);
1391 mss = tcp_skb_seglen(skb);
1393 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1394 * drop this restriction as unnecessary
1396 if (mss != tcp_skb_seglen(prev))
1399 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1401 /* CHECKME: This is non-MSS split case only?, this will
1402 * cause skipped skbs due to advancing loop btw, original
1403 * has that feature too
1405 if (tcp_skb_pcount(skb) <= 1)
1408 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1410 /* TODO: head merge to next could be attempted here
1411 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1412 * though it might not be worth of the additional hassle
1414 * ...we can probably just fallback to what was done
1415 * previously. We could try merging non-SACKed ones
1416 * as well but it probably isn't going to buy off
1417 * because later SACKs might again split them, and
1418 * it would make skb timestamp tracking considerably
1424 len = end_seq - TCP_SKB_CB(skb)->seq;
1426 BUG_ON(len > skb->len);
1428 /* MSS boundaries should be honoured or else pcount will
1429 * severely break even though it makes things bit trickier.
1430 * Optimize common case to avoid most of the divides
1432 mss = tcp_skb_mss(skb);
1434 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1435 * drop this restriction as unnecessary
1437 if (mss != tcp_skb_seglen(prev))
1442 } else if (len < mss) {
1450 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1451 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1454 if (!skb_shift(prev, skb, len))
1456 if (!tcp_shifted_skb(sk, prev, skb, state, pcount, len, mss, dup_sack))
1459 /* Hole filled allows collapsing with the next as well, this is very
1460 * useful when hole on every nth skb pattern happens
1462 skb = skb_rb_next(prev);
1466 if (!skb_can_shift(skb) ||
1467 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1468 (mss != tcp_skb_seglen(skb)))
1472 if (skb_shift(prev, skb, len)) {
1473 pcount += tcp_skb_pcount(skb);
1474 tcp_shifted_skb(sk, prev, skb, state, tcp_skb_pcount(skb),
1485 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1489 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1490 struct tcp_sack_block *next_dup,
1491 struct tcp_sacktag_state *state,
1492 u32 start_seq, u32 end_seq,
1495 struct tcp_sock *tp = tcp_sk(sk);
1496 struct sk_buff *tmp;
1498 skb_rbtree_walk_from(skb) {
1500 bool dup_sack = dup_sack_in;
1502 /* queue is in-order => we can short-circuit the walk early */
1503 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1507 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1508 in_sack = tcp_match_skb_to_sack(sk, skb,
1509 next_dup->start_seq,
1515 /* skb reference here is a bit tricky to get right, since
1516 * shifting can eat and free both this skb and the next,
1517 * so not even _safe variant of the loop is enough.
1520 tmp = tcp_shift_skb_data(sk, skb, state,
1521 start_seq, end_seq, dup_sack);
1530 in_sack = tcp_match_skb_to_sack(sk, skb,
1536 if (unlikely(in_sack < 0))
1540 TCP_SKB_CB(skb)->sacked =
1543 TCP_SKB_CB(skb)->sacked,
1544 TCP_SKB_CB(skb)->seq,
1545 TCP_SKB_CB(skb)->end_seq,
1547 tcp_skb_pcount(skb),
1549 tcp_rate_skb_delivered(sk, skb, state->rate);
1550 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1551 list_del_init(&skb->tcp_tsorted_anchor);
1553 if (!before(TCP_SKB_CB(skb)->seq,
1554 tcp_highest_sack_seq(tp)))
1555 tcp_advance_highest_sack(sk, skb);
1561 static struct sk_buff *tcp_sacktag_bsearch(struct sock *sk,
1562 struct tcp_sacktag_state *state,
1565 struct rb_node *parent, **p = &sk->tcp_rtx_queue.rb_node;
1566 struct sk_buff *skb;
1570 skb = rb_to_skb(parent);
1571 if (before(seq, TCP_SKB_CB(skb)->seq)) {
1572 p = &parent->rb_left;
1575 if (!before(seq, TCP_SKB_CB(skb)->end_seq)) {
1576 p = &parent->rb_right;
1584 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1585 struct tcp_sacktag_state *state,
1588 if (skb && after(TCP_SKB_CB(skb)->seq, skip_to_seq))
1591 return tcp_sacktag_bsearch(sk, state, skip_to_seq);
1594 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1596 struct tcp_sack_block *next_dup,
1597 struct tcp_sacktag_state *state,
1603 if (before(next_dup->start_seq, skip_to_seq)) {
1604 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1605 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1606 next_dup->start_seq, next_dup->end_seq,
1613 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1615 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1619 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1620 u32 prior_snd_una, struct tcp_sacktag_state *state)
1622 struct tcp_sock *tp = tcp_sk(sk);
1623 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1624 TCP_SKB_CB(ack_skb)->sacked);
1625 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1626 struct tcp_sack_block sp[TCP_NUM_SACKS];
1627 struct tcp_sack_block *cache;
1628 struct sk_buff *skb;
1629 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1631 bool found_dup_sack = false;
1633 int first_sack_index;
1636 state->reord = tp->snd_nxt;
1638 if (!tp->sacked_out)
1639 tcp_highest_sack_reset(sk);
1641 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1642 num_sacks, prior_snd_una);
1643 if (found_dup_sack) {
1644 state->flag |= FLAG_DSACKING_ACK;
1645 tp->delivered++; /* A spurious retransmission is delivered */
1648 /* Eliminate too old ACKs, but take into
1649 * account more or less fresh ones, they can
1650 * contain valid SACK info.
1652 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1655 if (!tp->packets_out)
1659 first_sack_index = 0;
1660 for (i = 0; i < num_sacks; i++) {
1661 bool dup_sack = !i && found_dup_sack;
1663 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1664 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1666 if (!tcp_is_sackblock_valid(tp, dup_sack,
1667 sp[used_sacks].start_seq,
1668 sp[used_sacks].end_seq)) {
1672 if (!tp->undo_marker)
1673 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1675 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1677 /* Don't count olds caused by ACK reordering */
1678 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1679 !after(sp[used_sacks].end_seq, tp->snd_una))
1681 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1684 NET_INC_STATS(sock_net(sk), mib_idx);
1686 first_sack_index = -1;
1690 /* Ignore very old stuff early */
1691 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1697 /* order SACK blocks to allow in order walk of the retrans queue */
1698 for (i = used_sacks - 1; i > 0; i--) {
1699 for (j = 0; j < i; j++) {
1700 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1701 swap(sp[j], sp[j + 1]);
1703 /* Track where the first SACK block goes to */
1704 if (j == first_sack_index)
1705 first_sack_index = j + 1;
1710 state->mss_now = tcp_current_mss(sk);
1714 if (!tp->sacked_out) {
1715 /* It's already past, so skip checking against it */
1716 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1718 cache = tp->recv_sack_cache;
1719 /* Skip empty blocks in at head of the cache */
1720 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1725 while (i < used_sacks) {
1726 u32 start_seq = sp[i].start_seq;
1727 u32 end_seq = sp[i].end_seq;
1728 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1729 struct tcp_sack_block *next_dup = NULL;
1731 if (found_dup_sack && ((i + 1) == first_sack_index))
1732 next_dup = &sp[i + 1];
1734 /* Skip too early cached blocks */
1735 while (tcp_sack_cache_ok(tp, cache) &&
1736 !before(start_seq, cache->end_seq))
1739 /* Can skip some work by looking recv_sack_cache? */
1740 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1741 after(end_seq, cache->start_seq)) {
1744 if (before(start_seq, cache->start_seq)) {
1745 skb = tcp_sacktag_skip(skb, sk, state,
1747 skb = tcp_sacktag_walk(skb, sk, next_dup,
1754 /* Rest of the block already fully processed? */
1755 if (!after(end_seq, cache->end_seq))
1758 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1762 /* ...tail remains todo... */
1763 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1764 /* ...but better entrypoint exists! */
1765 skb = tcp_highest_sack(sk);
1772 skb = tcp_sacktag_skip(skb, sk, state, cache->end_seq);
1773 /* Check overlap against next cached too (past this one already) */
1778 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1779 skb = tcp_highest_sack(sk);
1783 skb = tcp_sacktag_skip(skb, sk, state, start_seq);
1786 skb = tcp_sacktag_walk(skb, sk, next_dup, state,
1787 start_seq, end_seq, dup_sack);
1793 /* Clear the head of the cache sack blocks so we can skip it next time */
1794 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1795 tp->recv_sack_cache[i].start_seq = 0;
1796 tp->recv_sack_cache[i].end_seq = 0;
1798 for (j = 0; j < used_sacks; j++)
1799 tp->recv_sack_cache[i++] = sp[j];
1801 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Loss || tp->undo_marker)
1802 tcp_check_sack_reordering(sk, state->reord, 0);
1804 tcp_verify_left_out(tp);
1807 #if FASTRETRANS_DEBUG > 0
1808 WARN_ON((int)tp->sacked_out < 0);
1809 WARN_ON((int)tp->lost_out < 0);
1810 WARN_ON((int)tp->retrans_out < 0);
1811 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1816 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1817 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1819 static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
1823 holes = max(tp->lost_out, 1U);
1824 holes = min(holes, tp->packets_out);
1826 if ((tp->sacked_out + holes) > tp->packets_out) {
1827 tp->sacked_out = tp->packets_out - holes;
1833 /* If we receive more dupacks than we expected counting segments
1834 * in assumption of absent reordering, interpret this as reordering.
1835 * The only another reason could be bug in receiver TCP.
1837 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1839 struct tcp_sock *tp = tcp_sk(sk);
1841 if (!tcp_limit_reno_sacked(tp))
1844 tp->reordering = min_t(u32, tp->packets_out + addend,
1845 sock_net(sk)->ipv4.sysctl_tcp_max_reordering);
1846 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRENOREORDER);
1849 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1851 static void tcp_add_reno_sack(struct sock *sk)
1853 struct tcp_sock *tp = tcp_sk(sk);
1854 u32 prior_sacked = tp->sacked_out;
1857 tcp_check_reno_reordering(sk, 0);
1858 if (tp->sacked_out > prior_sacked)
1859 tp->delivered++; /* Some out-of-order packet is delivered */
1860 tcp_verify_left_out(tp);
1863 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1865 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1867 struct tcp_sock *tp = tcp_sk(sk);
1870 /* One ACK acked hole. The rest eat duplicate ACKs. */
1871 tp->delivered += max_t(int, acked - tp->sacked_out, 1);
1872 if (acked - 1 >= tp->sacked_out)
1875 tp->sacked_out -= acked - 1;
1877 tcp_check_reno_reordering(sk, acked);
1878 tcp_verify_left_out(tp);
1881 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1886 void tcp_clear_retrans(struct tcp_sock *tp)
1888 tp->retrans_out = 0;
1890 tp->undo_marker = 0;
1891 tp->undo_retrans = -1;
1895 static inline void tcp_init_undo(struct tcp_sock *tp)
1897 tp->undo_marker = tp->snd_una;
1898 /* Retransmission still in flight may cause DSACKs later. */
1899 tp->undo_retrans = tp->retrans_out ? : -1;
1902 /* Enter Loss state. If we detect SACK reneging, forget all SACK information
1903 * and reset tags completely, otherwise preserve SACKs. If receiver
1904 * dropped its ofo queue, we will know this due to reneging detection.
1906 void tcp_enter_loss(struct sock *sk)
1908 const struct inet_connection_sock *icsk = inet_csk(sk);
1909 struct tcp_sock *tp = tcp_sk(sk);
1910 struct net *net = sock_net(sk);
1911 struct sk_buff *skb;
1912 bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery;
1913 bool is_reneg; /* is receiver reneging on SACKs? */
1916 /* Reduce ssthresh if it has not yet been made inside this window. */
1917 if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
1918 !after(tp->high_seq, tp->snd_una) ||
1919 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1920 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1921 tp->prior_cwnd = tp->snd_cwnd;
1922 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1923 tcp_ca_event(sk, CA_EVENT_LOSS);
1927 tp->snd_cwnd_cnt = 0;
1928 tp->snd_cwnd_stamp = tcp_jiffies32;
1930 tp->retrans_out = 0;
1933 if (tcp_is_reno(tp))
1934 tcp_reset_reno_sack(tp);
1936 skb = tcp_rtx_queue_head(sk);
1937 is_reneg = skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED);
1939 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
1941 /* Mark SACK reneging until we recover from this loss event. */
1942 tp->is_sack_reneg = 1;
1944 tcp_clear_all_retrans_hints(tp);
1946 skb_rbtree_walk_from(skb) {
1947 mark_lost = (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
1950 tcp_sum_lost(tp, skb);
1951 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1953 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1954 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1955 tp->lost_out += tcp_skb_pcount(skb);
1958 tcp_verify_left_out(tp);
1960 /* Timeout in disordered state after receiving substantial DUPACKs
1961 * suggests that the degree of reordering is over-estimated.
1963 if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
1964 tp->sacked_out >= net->ipv4.sysctl_tcp_reordering)
1965 tp->reordering = min_t(unsigned int, tp->reordering,
1966 net->ipv4.sysctl_tcp_reordering);
1967 tcp_set_ca_state(sk, TCP_CA_Loss);
1968 tp->high_seq = tp->snd_nxt;
1969 tcp_ecn_queue_cwr(tp);
1971 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
1972 * loss recovery is underway except recurring timeout(s) on
1973 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
1975 tp->frto = net->ipv4.sysctl_tcp_frto &&
1976 (new_recovery || icsk->icsk_retransmits) &&
1977 !inet_csk(sk)->icsk_mtup.probe_size;
1980 /* If ACK arrived pointing to a remembered SACK, it means that our
1981 * remembered SACKs do not reflect real state of receiver i.e.
1982 * receiver _host_ is heavily congested (or buggy).
1984 * To avoid big spurious retransmission bursts due to transient SACK
1985 * scoreboard oddities that look like reneging, we give the receiver a
1986 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
1987 * restore sanity to the SACK scoreboard. If the apparent reneging
1988 * persists until this RTO then we'll clear the SACK scoreboard.
1990 static bool tcp_check_sack_reneging(struct sock *sk, int flag)
1992 if (flag & FLAG_SACK_RENEGING) {
1993 struct tcp_sock *tp = tcp_sk(sk);
1994 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4),
1995 msecs_to_jiffies(10));
1997 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1998 delay, TCP_RTO_MAX);
2004 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2005 * counter when SACK is enabled (without SACK, sacked_out is used for
2008 * With reordering, holes may still be in flight, so RFC3517 recovery
2009 * uses pure sacked_out (total number of SACKed segments) even though
2010 * it violates the RFC that uses duplicate ACKs, often these are equal
2011 * but when e.g. out-of-window ACKs or packet duplication occurs,
2012 * they differ. Since neither occurs due to loss, TCP should really
2015 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2017 return tp->sacked_out + 1;
2020 /* Linux NewReno/SACK/ECN state machine.
2021 * --------------------------------------
2023 * "Open" Normal state, no dubious events, fast path.
2024 * "Disorder" In all the respects it is "Open",
2025 * but requires a bit more attention. It is entered when
2026 * we see some SACKs or dupacks. It is split of "Open"
2027 * mainly to move some processing from fast path to slow one.
2028 * "CWR" CWND was reduced due to some Congestion Notification event.
2029 * It can be ECN, ICMP source quench, local device congestion.
2030 * "Recovery" CWND was reduced, we are fast-retransmitting.
2031 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2033 * tcp_fastretrans_alert() is entered:
2034 * - each incoming ACK, if state is not "Open"
2035 * - when arrived ACK is unusual, namely:
2040 * Counting packets in flight is pretty simple.
2042 * in_flight = packets_out - left_out + retrans_out
2044 * packets_out is SND.NXT-SND.UNA counted in packets.
2046 * retrans_out is number of retransmitted segments.
2048 * left_out is number of segments left network, but not ACKed yet.
2050 * left_out = sacked_out + lost_out
2052 * sacked_out: Packets, which arrived to receiver out of order
2053 * and hence not ACKed. With SACKs this number is simply
2054 * amount of SACKed data. Even without SACKs
2055 * it is easy to give pretty reliable estimate of this number,
2056 * counting duplicate ACKs.
2058 * lost_out: Packets lost by network. TCP has no explicit
2059 * "loss notification" feedback from network (for now).
2060 * It means that this number can be only _guessed_.
2061 * Actually, it is the heuristics to predict lossage that
2062 * distinguishes different algorithms.
2064 * F.e. after RTO, when all the queue is considered as lost,
2065 * lost_out = packets_out and in_flight = retrans_out.
2067 * Essentially, we have now a few algorithms detecting
2070 * If the receiver supports SACK:
2072 * RFC6675/3517: It is the conventional algorithm. A packet is
2073 * considered lost if the number of higher sequence packets
2074 * SACKed is greater than or equal the DUPACK thoreshold
2075 * (reordering). This is implemented in tcp_mark_head_lost and
2076 * tcp_update_scoreboard.
2078 * RACK (draft-ietf-tcpm-rack-01): it is a newer algorithm
2079 * (2017-) that checks timing instead of counting DUPACKs.
2080 * Essentially a packet is considered lost if it's not S/ACKed
2081 * after RTT + reordering_window, where both metrics are
2082 * dynamically measured and adjusted. This is implemented in
2083 * tcp_rack_mark_lost.
2085 * If the receiver does not support SACK:
2087 * NewReno (RFC6582): in Recovery we assume that one segment
2088 * is lost (classic Reno). While we are in Recovery and
2089 * a partial ACK arrives, we assume that one more packet
2090 * is lost (NewReno). This heuristics are the same in NewReno
2093 * Really tricky (and requiring careful tuning) part of algorithm
2094 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2095 * The first determines the moment _when_ we should reduce CWND and,
2096 * hence, slow down forward transmission. In fact, it determines the moment
2097 * when we decide that hole is caused by loss, rather than by a reorder.
2099 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2100 * holes, caused by lost packets.
2102 * And the most logically complicated part of algorithm is undo
2103 * heuristics. We detect false retransmits due to both too early
2104 * fast retransmit (reordering) and underestimated RTO, analyzing
2105 * timestamps and D-SACKs. When we detect that some segments were
2106 * retransmitted by mistake and CWND reduction was wrong, we undo
2107 * window reduction and abort recovery phase. This logic is hidden
2108 * inside several functions named tcp_try_undo_<something>.
2111 /* This function decides, when we should leave Disordered state
2112 * and enter Recovery phase, reducing congestion window.
2114 * Main question: may we further continue forward transmission
2115 * with the same cwnd?
2117 static bool tcp_time_to_recover(struct sock *sk, int flag)
2119 struct tcp_sock *tp = tcp_sk(sk);
2121 /* Trick#1: The loss is proven. */
2125 /* Not-A-Trick#2 : Classic rule... */
2126 if (tcp_dupack_heuristics(tp) > tp->reordering)
2132 /* Detect loss in event "A" above by marking head of queue up as lost.
2133 * For non-SACK(Reno) senders, the first "packets" number of segments
2134 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2135 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2136 * the maximum SACKed segments to pass before reaching this limit.
2138 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2140 struct tcp_sock *tp = tcp_sk(sk);
2141 struct sk_buff *skb;
2142 int cnt, oldcnt, lost;
2144 /* Use SACK to deduce losses of new sequences sent during recovery */
2145 const u32 loss_high = tcp_is_sack(tp) ? tp->snd_nxt : tp->high_seq;
2147 WARN_ON(packets > tp->packets_out);
2148 skb = tp->lost_skb_hint;
2150 /* Head already handled? */
2151 if (mark_head && after(TCP_SKB_CB(skb)->seq, tp->snd_una))
2153 cnt = tp->lost_cnt_hint;
2155 skb = tcp_rtx_queue_head(sk);
2159 skb_rbtree_walk_from(skb) {
2160 /* TODO: do this better */
2161 /* this is not the most efficient way to do this... */
2162 tp->lost_skb_hint = skb;
2163 tp->lost_cnt_hint = cnt;
2165 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2169 if (tcp_is_reno(tp) ||
2170 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2171 cnt += tcp_skb_pcount(skb);
2173 if (cnt > packets) {
2174 if (tcp_is_sack(tp) ||
2175 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
2176 (oldcnt >= packets))
2179 mss = tcp_skb_mss(skb);
2180 /* If needed, chop off the prefix to mark as lost. */
2181 lost = (packets - oldcnt) * mss;
2182 if (lost < skb->len &&
2183 tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
2184 lost, mss, GFP_ATOMIC) < 0)
2189 tcp_skb_mark_lost(tp, skb);
2194 tcp_verify_left_out(tp);
2197 /* Account newly detected lost packet(s) */
2199 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2201 struct tcp_sock *tp = tcp_sk(sk);
2203 if (tcp_is_reno(tp)) {
2204 tcp_mark_head_lost(sk, 1, 1);
2206 int sacked_upto = tp->sacked_out - tp->reordering;
2207 if (sacked_upto >= 0)
2208 tcp_mark_head_lost(sk, sacked_upto, 0);
2209 else if (fast_rexmit)
2210 tcp_mark_head_lost(sk, 1, 1);
2214 static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when)
2216 return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2217 before(tp->rx_opt.rcv_tsecr, when);
2220 /* skb is spurious retransmitted if the returned timestamp echo
2221 * reply is prior to the skb transmission time
2223 static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp,
2224 const struct sk_buff *skb)
2226 return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) &&
2227 tcp_tsopt_ecr_before(tp, tcp_skb_timestamp(skb));
2230 /* Nothing was retransmitted or returned timestamp is less
2231 * than timestamp of the first retransmission.
2233 static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2235 return !tp->retrans_stamp ||
2236 tcp_tsopt_ecr_before(tp, tp->retrans_stamp);
2239 /* Undo procedures. */
2241 /* We can clear retrans_stamp when there are no retransmissions in the
2242 * window. It would seem that it is trivially available for us in
2243 * tp->retrans_out, however, that kind of assumptions doesn't consider
2244 * what will happen if errors occur when sending retransmission for the
2245 * second time. ...It could the that such segment has only
2246 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2247 * the head skb is enough except for some reneging corner cases that
2248 * are not worth the effort.
2250 * Main reason for all this complexity is the fact that connection dying
2251 * time now depends on the validity of the retrans_stamp, in particular,
2252 * that successive retransmissions of a segment must not advance
2253 * retrans_stamp under any conditions.
2255 static bool tcp_any_retrans_done(const struct sock *sk)
2257 const struct tcp_sock *tp = tcp_sk(sk);
2258 struct sk_buff *skb;
2260 if (tp->retrans_out)
2263 skb = tcp_rtx_queue_head(sk);
2264 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2270 static void DBGUNDO(struct sock *sk, const char *msg)
2272 #if FASTRETRANS_DEBUG > 1
2273 struct tcp_sock *tp = tcp_sk(sk);
2274 struct inet_sock *inet = inet_sk(sk);
2276 if (sk->sk_family == AF_INET) {
2277 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2279 &inet->inet_daddr, ntohs(inet->inet_dport),
2280 tp->snd_cwnd, tcp_left_out(tp),
2281 tp->snd_ssthresh, tp->prior_ssthresh,
2284 #if IS_ENABLED(CONFIG_IPV6)
2285 else if (sk->sk_family == AF_INET6) {
2286 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2288 &sk->sk_v6_daddr, ntohs(inet->inet_dport),
2289 tp->snd_cwnd, tcp_left_out(tp),
2290 tp->snd_ssthresh, tp->prior_ssthresh,
2297 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
2299 struct tcp_sock *tp = tcp_sk(sk);
2302 struct sk_buff *skb;
2304 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2305 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2308 tcp_clear_all_retrans_hints(tp);
2311 if (tp->prior_ssthresh) {
2312 const struct inet_connection_sock *icsk = inet_csk(sk);
2314 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2316 if (tp->prior_ssthresh > tp->snd_ssthresh) {
2317 tp->snd_ssthresh = tp->prior_ssthresh;
2318 tcp_ecn_withdraw_cwr(tp);
2321 tp->snd_cwnd_stamp = tcp_jiffies32;
2322 tp->undo_marker = 0;
2323 tp->rack.advanced = 1; /* Force RACK to re-exam losses */
2326 static inline bool tcp_may_undo(const struct tcp_sock *tp)
2328 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2331 /* People celebrate: "We love our President!" */
2332 static bool tcp_try_undo_recovery(struct sock *sk)
2334 struct tcp_sock *tp = tcp_sk(sk);
2336 if (tcp_may_undo(tp)) {
2339 /* Happy end! We did not retransmit anything
2340 * or our original transmission succeeded.
2342 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2343 tcp_undo_cwnd_reduction(sk, false);
2344 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2345 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2347 mib_idx = LINUX_MIB_TCPFULLUNDO;
2349 NET_INC_STATS(sock_net(sk), mib_idx);
2350 } else if (tp->rack.reo_wnd_persist) {
2351 tp->rack.reo_wnd_persist--;
2353 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2354 /* Hold old state until something *above* high_seq
2355 * is ACKed. For Reno it is MUST to prevent false
2356 * fast retransmits (RFC2582). SACK TCP is safe. */
2357 if (!tcp_any_retrans_done(sk))
2358 tp->retrans_stamp = 0;
2361 tcp_set_ca_state(sk, TCP_CA_Open);
2362 tp->is_sack_reneg = 0;
2366 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2367 static bool tcp_try_undo_dsack(struct sock *sk)
2369 struct tcp_sock *tp = tcp_sk(sk);
2371 if (tp->undo_marker && !tp->undo_retrans) {
2372 tp->rack.reo_wnd_persist = min(TCP_RACK_RECOVERY_THRESH,
2373 tp->rack.reo_wnd_persist + 1);
2374 DBGUNDO(sk, "D-SACK");
2375 tcp_undo_cwnd_reduction(sk, false);
2376 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2382 /* Undo during loss recovery after partial ACK or using F-RTO. */
2383 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2385 struct tcp_sock *tp = tcp_sk(sk);
2387 if (frto_undo || tcp_may_undo(tp)) {
2388 tcp_undo_cwnd_reduction(sk, true);
2390 DBGUNDO(sk, "partial loss");
2391 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2393 NET_INC_STATS(sock_net(sk),
2394 LINUX_MIB_TCPSPURIOUSRTOS);
2395 inet_csk(sk)->icsk_retransmits = 0;
2396 if (frto_undo || tcp_is_sack(tp)) {
2397 tcp_set_ca_state(sk, TCP_CA_Open);
2398 tp->is_sack_reneg = 0;
2405 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
2406 * It computes the number of packets to send (sndcnt) based on packets newly
2408 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2409 * cwnd reductions across a full RTT.
2410 * 2) Otherwise PRR uses packet conservation to send as much as delivered.
2411 * But when the retransmits are acked without further losses, PRR
2412 * slow starts cwnd up to ssthresh to speed up the recovery.
2414 static void tcp_init_cwnd_reduction(struct sock *sk)
2416 struct tcp_sock *tp = tcp_sk(sk);
2418 tp->high_seq = tp->snd_nxt;
2419 tp->tlp_high_seq = 0;
2420 tp->snd_cwnd_cnt = 0;
2421 tp->prior_cwnd = tp->snd_cwnd;
2422 tp->prr_delivered = 0;
2424 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2425 tcp_ecn_queue_cwr(tp);
2428 void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked, int flag)
2430 struct tcp_sock *tp = tcp_sk(sk);
2432 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2434 if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd))
2437 tp->prr_delivered += newly_acked_sacked;
2439 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2441 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2442 } else if ((flag & FLAG_RETRANS_DATA_ACKED) &&
2443 !(flag & FLAG_LOST_RETRANS)) {
2444 sndcnt = min_t(int, delta,
2445 max_t(int, tp->prr_delivered - tp->prr_out,
2446 newly_acked_sacked) + 1);
2448 sndcnt = min(delta, newly_acked_sacked);
2450 /* Force a fast retransmit upon entering fast recovery */
2451 sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1));
2452 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2455 static inline void tcp_end_cwnd_reduction(struct sock *sk)
2457 struct tcp_sock *tp = tcp_sk(sk);
2459 if (inet_csk(sk)->icsk_ca_ops->cong_control)
2462 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2463 if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH &&
2464 (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR || tp->undo_marker)) {
2465 tp->snd_cwnd = tp->snd_ssthresh;
2466 tp->snd_cwnd_stamp = tcp_jiffies32;
2468 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2471 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2472 void tcp_enter_cwr(struct sock *sk)
2474 struct tcp_sock *tp = tcp_sk(sk);
2476 tp->prior_ssthresh = 0;
2477 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2478 tp->undo_marker = 0;
2479 tcp_init_cwnd_reduction(sk);
2480 tcp_set_ca_state(sk, TCP_CA_CWR);
2483 EXPORT_SYMBOL(tcp_enter_cwr);
2485 static void tcp_try_keep_open(struct sock *sk)
2487 struct tcp_sock *tp = tcp_sk(sk);
2488 int state = TCP_CA_Open;
2490 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2491 state = TCP_CA_Disorder;
2493 if (inet_csk(sk)->icsk_ca_state != state) {
2494 tcp_set_ca_state(sk, state);
2495 tp->high_seq = tp->snd_nxt;
2499 static void tcp_try_to_open(struct sock *sk, int flag)
2501 struct tcp_sock *tp = tcp_sk(sk);
2503 tcp_verify_left_out(tp);
2505 if (!tcp_any_retrans_done(sk))
2506 tp->retrans_stamp = 0;
2508 if (flag & FLAG_ECE)
2511 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2512 tcp_try_keep_open(sk);
2516 static void tcp_mtup_probe_failed(struct sock *sk)
2518 struct inet_connection_sock *icsk = inet_csk(sk);
2520 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2521 icsk->icsk_mtup.probe_size = 0;
2522 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPFAIL);
2525 static void tcp_mtup_probe_success(struct sock *sk)
2527 struct tcp_sock *tp = tcp_sk(sk);
2528 struct inet_connection_sock *icsk = inet_csk(sk);
2530 /* FIXME: breaks with very large cwnd */
2531 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2532 tp->snd_cwnd = tp->snd_cwnd *
2533 tcp_mss_to_mtu(sk, tp->mss_cache) /
2534 icsk->icsk_mtup.probe_size;
2535 tp->snd_cwnd_cnt = 0;
2536 tp->snd_cwnd_stamp = tcp_jiffies32;
2537 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2539 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2540 icsk->icsk_mtup.probe_size = 0;
2541 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2542 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS);
2545 /* Do a simple retransmit without using the backoff mechanisms in
2546 * tcp_timer. This is used for path mtu discovery.
2547 * The socket is already locked here.
2549 void tcp_simple_retransmit(struct sock *sk)
2551 const struct inet_connection_sock *icsk = inet_csk(sk);
2552 struct tcp_sock *tp = tcp_sk(sk);
2553 struct sk_buff *skb;
2554 unsigned int mss = tcp_current_mss(sk);
2556 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2557 if (tcp_skb_seglen(skb) > mss &&
2558 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2559 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2560 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2561 tp->retrans_out -= tcp_skb_pcount(skb);
2563 tcp_skb_mark_lost_uncond_verify(tp, skb);
2567 tcp_clear_retrans_hints_partial(tp);
2572 if (tcp_is_reno(tp))
2573 tcp_limit_reno_sacked(tp);
2575 tcp_verify_left_out(tp);
2577 /* Don't muck with the congestion window here.
2578 * Reason is that we do not increase amount of _data_
2579 * in network, but units changed and effective
2580 * cwnd/ssthresh really reduced now.
2582 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2583 tp->high_seq = tp->snd_nxt;
2584 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2585 tp->prior_ssthresh = 0;
2586 tp->undo_marker = 0;
2587 tcp_set_ca_state(sk, TCP_CA_Loss);
2589 tcp_xmit_retransmit_queue(sk);
2591 EXPORT_SYMBOL(tcp_simple_retransmit);
2593 void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2595 struct tcp_sock *tp = tcp_sk(sk);
2598 if (tcp_is_reno(tp))
2599 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2601 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2603 NET_INC_STATS(sock_net(sk), mib_idx);
2605 tp->prior_ssthresh = 0;
2608 if (!tcp_in_cwnd_reduction(sk)) {
2610 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2611 tcp_init_cwnd_reduction(sk);
2613 tcp_set_ca_state(sk, TCP_CA_Recovery);
2616 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2617 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2619 static void tcp_process_loss(struct sock *sk, int flag, bool is_dupack,
2622 struct tcp_sock *tp = tcp_sk(sk);
2623 bool recovered = !before(tp->snd_una, tp->high_seq);
2625 if ((flag & FLAG_SND_UNA_ADVANCED) &&
2626 tcp_try_undo_loss(sk, false))
2629 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2630 /* Step 3.b. A timeout is spurious if not all data are
2631 * lost, i.e., never-retransmitted data are (s)acked.
2633 if ((flag & FLAG_ORIG_SACK_ACKED) &&
2634 tcp_try_undo_loss(sk, true))
2637 if (after(tp->snd_nxt, tp->high_seq)) {
2638 if (flag & FLAG_DATA_SACKED || is_dupack)
2639 tp->frto = 0; /* Step 3.a. loss was real */
2640 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2641 tp->high_seq = tp->snd_nxt;
2642 /* Step 2.b. Try send new data (but deferred until cwnd
2643 * is updated in tcp_ack()). Otherwise fall back to
2644 * the conventional recovery.
2646 if (!tcp_write_queue_empty(sk) &&
2647 after(tcp_wnd_end(tp), tp->snd_nxt)) {
2648 *rexmit = REXMIT_NEW;
2656 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2657 tcp_try_undo_recovery(sk);
2660 if (tcp_is_reno(tp)) {
2661 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2662 * delivered. Lower inflight to clock out (re)tranmissions.
2664 if (after(tp->snd_nxt, tp->high_seq) && is_dupack)
2665 tcp_add_reno_sack(sk);
2666 else if (flag & FLAG_SND_UNA_ADVANCED)
2667 tcp_reset_reno_sack(tp);
2669 *rexmit = REXMIT_LOST;
2672 /* Undo during fast recovery after partial ACK. */
2673 static bool tcp_try_undo_partial(struct sock *sk, u32 prior_snd_una)
2675 struct tcp_sock *tp = tcp_sk(sk);
2677 if (tp->undo_marker && tcp_packet_delayed(tp)) {
2678 /* Plain luck! Hole if filled with delayed
2679 * packet, rather than with a retransmit. Check reordering.
2681 tcp_check_sack_reordering(sk, prior_snd_una, 1);
2683 /* We are getting evidence that the reordering degree is higher
2684 * than we realized. If there are no retransmits out then we
2685 * can undo. Otherwise we clock out new packets but do not
2686 * mark more packets lost or retransmit more.
2688 if (tp->retrans_out)
2691 if (!tcp_any_retrans_done(sk))
2692 tp->retrans_stamp = 0;
2694 DBGUNDO(sk, "partial recovery");
2695 tcp_undo_cwnd_reduction(sk, true);
2696 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2697 tcp_try_keep_open(sk);
2703 static void tcp_rack_identify_loss(struct sock *sk, int *ack_flag)
2705 struct tcp_sock *tp = tcp_sk(sk);
2707 /* Use RACK to detect loss */
2708 if (sock_net(sk)->ipv4.sysctl_tcp_recovery & TCP_RACK_LOSS_DETECTION) {
2709 u32 prior_retrans = tp->retrans_out;
2711 tcp_rack_mark_lost(sk);
2712 if (prior_retrans > tp->retrans_out)
2713 *ack_flag |= FLAG_LOST_RETRANS;
2717 static bool tcp_force_fast_retransmit(struct sock *sk)
2719 struct tcp_sock *tp = tcp_sk(sk);
2721 return after(tcp_highest_sack_seq(tp),
2722 tp->snd_una + tp->reordering * tp->mss_cache);
2725 /* Process an event, which can update packets-in-flight not trivially.
2726 * Main goal of this function is to calculate new estimate for left_out,
2727 * taking into account both packets sitting in receiver's buffer and
2728 * packets lost by network.
2730 * Besides that it updates the congestion state when packet loss or ECN
2731 * is detected. But it does not reduce the cwnd, it is done by the
2732 * congestion control later.
2734 * It does _not_ decide what to send, it is made in function
2735 * tcp_xmit_retransmit_queue().
2737 static void tcp_fastretrans_alert(struct sock *sk, const u32 prior_snd_una,
2738 bool is_dupack, int *ack_flag, int *rexmit)
2740 struct inet_connection_sock *icsk = inet_csk(sk);
2741 struct tcp_sock *tp = tcp_sk(sk);
2742 int fast_rexmit = 0, flag = *ack_flag;
2743 bool do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2744 tcp_force_fast_retransmit(sk));
2746 if (!tp->packets_out && tp->sacked_out)
2749 /* Now state machine starts.
2750 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2751 if (flag & FLAG_ECE)
2752 tp->prior_ssthresh = 0;
2754 /* B. In all the states check for reneging SACKs. */
2755 if (tcp_check_sack_reneging(sk, flag))
2758 /* C. Check consistency of the current state. */
2759 tcp_verify_left_out(tp);
2761 /* D. Check state exit conditions. State can be terminated
2762 * when high_seq is ACKed. */
2763 if (icsk->icsk_ca_state == TCP_CA_Open) {
2764 WARN_ON(tp->retrans_out != 0);
2765 tp->retrans_stamp = 0;
2766 } else if (!before(tp->snd_una, tp->high_seq)) {
2767 switch (icsk->icsk_ca_state) {
2769 /* CWR is to be held something *above* high_seq
2770 * is ACKed for CWR bit to reach receiver. */
2771 if (tp->snd_una != tp->high_seq) {
2772 tcp_end_cwnd_reduction(sk);
2773 tcp_set_ca_state(sk, TCP_CA_Open);
2777 case TCP_CA_Recovery:
2778 if (tcp_is_reno(tp))
2779 tcp_reset_reno_sack(tp);
2780 if (tcp_try_undo_recovery(sk))
2782 tcp_end_cwnd_reduction(sk);
2787 /* E. Process state. */
2788 switch (icsk->icsk_ca_state) {
2789 case TCP_CA_Recovery:
2790 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2791 if (tcp_is_reno(tp) && is_dupack)
2792 tcp_add_reno_sack(sk);
2794 if (tcp_try_undo_partial(sk, prior_snd_una))
2796 /* Partial ACK arrived. Force fast retransmit. */
2797 do_lost = tcp_is_reno(tp) ||
2798 tcp_force_fast_retransmit(sk);
2800 if (tcp_try_undo_dsack(sk)) {
2801 tcp_try_keep_open(sk);
2804 tcp_rack_identify_loss(sk, ack_flag);
2807 tcp_process_loss(sk, flag, is_dupack, rexmit);
2808 tcp_rack_identify_loss(sk, ack_flag);
2809 if (!(icsk->icsk_ca_state == TCP_CA_Open ||
2810 (*ack_flag & FLAG_LOST_RETRANS)))
2812 /* Change state if cwnd is undone or retransmits are lost */
2815 if (tcp_is_reno(tp)) {
2816 if (flag & FLAG_SND_UNA_ADVANCED)
2817 tcp_reset_reno_sack(tp);
2819 tcp_add_reno_sack(sk);
2822 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
2823 tcp_try_undo_dsack(sk);
2825 tcp_rack_identify_loss(sk, ack_flag);
2826 if (!tcp_time_to_recover(sk, flag)) {
2827 tcp_try_to_open(sk, flag);
2831 /* MTU probe failure: don't reduce cwnd */
2832 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2833 icsk->icsk_mtup.probe_size &&
2834 tp->snd_una == tp->mtu_probe.probe_seq_start) {
2835 tcp_mtup_probe_failed(sk);
2836 /* Restores the reduction we did in tcp_mtup_probe() */
2838 tcp_simple_retransmit(sk);
2842 /* Otherwise enter Recovery state */
2843 tcp_enter_recovery(sk, (flag & FLAG_ECE));
2848 tcp_update_scoreboard(sk, fast_rexmit);
2849 *rexmit = REXMIT_LOST;
2852 static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us, const int flag)
2854 u32 wlen = sock_net(sk)->ipv4.sysctl_tcp_min_rtt_wlen * HZ;
2855 struct tcp_sock *tp = tcp_sk(sk);
2857 if ((flag & FLAG_ACK_MAYBE_DELAYED) && rtt_us > tcp_min_rtt(tp)) {
2858 /* If the remote keeps returning delayed ACKs, eventually
2859 * the min filter would pick it up and overestimate the
2860 * prop. delay when it expires. Skip suspected delayed ACKs.
2864 minmax_running_min(&tp->rtt_min, wlen, tcp_jiffies32,
2865 rtt_us ? : jiffies_to_usecs(1));
2868 static bool tcp_ack_update_rtt(struct sock *sk, const int flag,
2869 long seq_rtt_us, long sack_rtt_us,
2870 long ca_rtt_us, struct rate_sample *rs)
2872 const struct tcp_sock *tp = tcp_sk(sk);
2874 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2875 * broken middle-boxes or peers may corrupt TS-ECR fields. But
2876 * Karn's algorithm forbids taking RTT if some retransmitted data
2877 * is acked (RFC6298).
2880 seq_rtt_us = sack_rtt_us;
2882 /* RTTM Rule: A TSecr value received in a segment is used to
2883 * update the averaged RTT measurement only if the segment
2884 * acknowledges some new data, i.e., only if it advances the
2885 * left edge of the send window.
2886 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2888 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2889 flag & FLAG_ACKED) {
2890 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr;
2891 u32 delta_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
2893 seq_rtt_us = ca_rtt_us = delta_us;
2895 rs->rtt_us = ca_rtt_us; /* RTT of last (S)ACKed packet (or -1) */
2899 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
2900 * always taken together with ACK, SACK, or TS-opts. Any negative
2901 * values will be skipped with the seq_rtt_us < 0 check above.
2903 tcp_update_rtt_min(sk, ca_rtt_us, flag);
2904 tcp_rtt_estimator(sk, seq_rtt_us);
2907 /* RFC6298: only reset backoff on valid RTT measurement. */
2908 inet_csk(sk)->icsk_backoff = 0;
2912 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
2913 void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req)
2915 struct rate_sample rs;
2918 if (req && !req->num_retrans && tcp_rsk(req)->snt_synack)
2919 rtt_us = tcp_stamp_us_delta(tcp_clock_us(), tcp_rsk(req)->snt_synack);
2921 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us, &rs);
2925 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
2927 const struct inet_connection_sock *icsk = inet_csk(sk);
2929 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
2930 tcp_sk(sk)->snd_cwnd_stamp = tcp_jiffies32;
2933 /* Restart timer after forward progress on connection.
2934 * RFC2988 recommends to restart timer to now+rto.
2936 void tcp_rearm_rto(struct sock *sk)
2938 const struct inet_connection_sock *icsk = inet_csk(sk);
2939 struct tcp_sock *tp = tcp_sk(sk);
2941 /* If the retrans timer is currently being used by Fast Open
2942 * for SYN-ACK retrans purpose, stay put.
2944 if (tp->fastopen_rsk)
2947 if (!tp->packets_out) {
2948 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
2950 u32 rto = inet_csk(sk)->icsk_rto;
2951 /* Offset the time elapsed after installing regular RTO */
2952 if (icsk->icsk_pending == ICSK_TIME_REO_TIMEOUT ||
2953 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
2954 s64 delta_us = tcp_rto_delta_us(sk);
2955 /* delta_us may not be positive if the socket is locked
2956 * when the retrans timer fires and is rescheduled.
2958 rto = usecs_to_jiffies(max_t(int, delta_us, 1));
2960 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
2965 /* Try to schedule a loss probe; if that doesn't work, then schedule an RTO. */
2966 static void tcp_set_xmit_timer(struct sock *sk)
2968 if (!tcp_schedule_loss_probe(sk, true))
2972 /* If we get here, the whole TSO packet has not been acked. */
2973 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
2975 struct tcp_sock *tp = tcp_sk(sk);
2978 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
2980 packets_acked = tcp_skb_pcount(skb);
2981 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2983 packets_acked -= tcp_skb_pcount(skb);
2985 if (packets_acked) {
2986 BUG_ON(tcp_skb_pcount(skb) == 0);
2987 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
2990 return packets_acked;
2993 static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
2996 const struct skb_shared_info *shinfo;
2998 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
2999 if (likely(!TCP_SKB_CB(skb)->txstamp_ack))
3002 shinfo = skb_shinfo(skb);
3003 if (!before(shinfo->tskey, prior_snd_una) &&
3004 before(shinfo->tskey, tcp_sk(sk)->snd_una)) {
3005 tcp_skb_tsorted_save(skb) {
3006 __skb_tstamp_tx(skb, NULL, sk, SCM_TSTAMP_ACK);
3007 } tcp_skb_tsorted_restore(skb);
3011 /* Remove acknowledged frames from the retransmission queue. If our packet
3012 * is before the ack sequence we can discard it as it's confirmed to have
3013 * arrived at the other end.
3015 static int tcp_clean_rtx_queue(struct sock *sk, u32 prior_fack,
3017 struct tcp_sacktag_state *sack)
3019 const struct inet_connection_sock *icsk = inet_csk(sk);
3020 u64 first_ackt, last_ackt;
3021 struct tcp_sock *tp = tcp_sk(sk);
3022 u32 prior_sacked = tp->sacked_out;
3023 u32 reord = tp->snd_nxt; /* lowest acked un-retx un-sacked seq */
3024 struct sk_buff *skb, *next;
3025 bool fully_acked = true;
3026 long sack_rtt_us = -1L;
3027 long seq_rtt_us = -1L;
3028 long ca_rtt_us = -1L;
3030 u32 last_in_flight = 0;
3036 for (skb = skb_rb_first(&sk->tcp_rtx_queue); skb; skb = next) {
3037 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3038 const u32 start_seq = scb->seq;
3039 u8 sacked = scb->sacked;
3042 tcp_ack_tstamp(sk, skb, prior_snd_una);
3044 /* Determine how many packets and what bytes were acked, tso and else */
3045 if (after(scb->end_seq, tp->snd_una)) {
3046 if (tcp_skb_pcount(skb) == 1 ||
3047 !after(tp->snd_una, scb->seq))
3050 acked_pcount = tcp_tso_acked(sk, skb);
3053 fully_acked = false;
3055 acked_pcount = tcp_skb_pcount(skb);
3058 if (unlikely(sacked & TCPCB_RETRANS)) {
3059 if (sacked & TCPCB_SACKED_RETRANS)
3060 tp->retrans_out -= acked_pcount;
3061 flag |= FLAG_RETRANS_DATA_ACKED;
3062 } else if (!(sacked & TCPCB_SACKED_ACKED)) {
3063 last_ackt = skb->skb_mstamp;
3064 WARN_ON_ONCE(last_ackt == 0);
3066 first_ackt = last_ackt;
3068 last_in_flight = TCP_SKB_CB(skb)->tx.in_flight;
3069 if (before(start_seq, reord))
3071 if (!after(scb->end_seq, tp->high_seq))
3072 flag |= FLAG_ORIG_SACK_ACKED;
3075 if (sacked & TCPCB_SACKED_ACKED) {
3076 tp->sacked_out -= acked_pcount;
3077 } else if (tcp_is_sack(tp)) {
3078 tp->delivered += acked_pcount;
3079 if (!tcp_skb_spurious_retrans(tp, skb))
3080 tcp_rack_advance(tp, sacked, scb->end_seq,
3083 if (sacked & TCPCB_LOST)
3084 tp->lost_out -= acked_pcount;
3086 tp->packets_out -= acked_pcount;
3087 pkts_acked += acked_pcount;
3088 tcp_rate_skb_delivered(sk, skb, sack->rate);
3090 /* Initial outgoing SYN's get put onto the write_queue
3091 * just like anything else we transmit. It is not
3092 * true data, and if we misinform our callers that
3093 * this ACK acks real data, we will erroneously exit
3094 * connection startup slow start one packet too
3095 * quickly. This is severely frowned upon behavior.
3097 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
3098 flag |= FLAG_DATA_ACKED;
3100 flag |= FLAG_SYN_ACKED;
3101 tp->retrans_stamp = 0;
3107 next = skb_rb_next(skb);
3108 if (unlikely(skb == tp->retransmit_skb_hint))
3109 tp->retransmit_skb_hint = NULL;
3110 if (unlikely(skb == tp->lost_skb_hint))
3111 tp->lost_skb_hint = NULL;
3112 tcp_rtx_queue_unlink_and_free(skb, sk);
3116 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
3118 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3119 tp->snd_up = tp->snd_una;
3121 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3122 flag |= FLAG_SACK_RENEGING;
3124 if (likely(first_ackt) && !(flag & FLAG_RETRANS_DATA_ACKED)) {
3125 seq_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, first_ackt);
3126 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, last_ackt);
3128 if (pkts_acked == 1 && last_in_flight < tp->mss_cache &&
3129 last_in_flight && !prior_sacked && fully_acked &&
3130 sack->rate->prior_delivered + 1 == tp->delivered &&
3131 !(flag & (FLAG_CA_ALERT | FLAG_SYN_ACKED))) {
3132 /* Conservatively mark a delayed ACK. It's typically
3133 * from a lone runt packet over the round trip to
3134 * a receiver w/o out-of-order or CE events.
3136 flag |= FLAG_ACK_MAYBE_DELAYED;
3139 if (sack->first_sackt) {
3140 sack_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->first_sackt);
3141 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->last_sackt);
3143 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us,
3144 ca_rtt_us, sack->rate);
3146 if (flag & FLAG_ACKED) {
3147 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3148 if (unlikely(icsk->icsk_mtup.probe_size &&
3149 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3150 tcp_mtup_probe_success(sk);
3153 if (tcp_is_reno(tp)) {
3154 tcp_remove_reno_sacks(sk, pkts_acked);
3158 /* Non-retransmitted hole got filled? That's reordering */
3159 if (before(reord, prior_fack))
3160 tcp_check_sack_reordering(sk, reord, 0);
3162 delta = prior_sacked - tp->sacked_out;
3163 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3165 } else if (skb && rtt_update && sack_rtt_us >= 0 &&
3166 sack_rtt_us > tcp_stamp_us_delta(tp->tcp_mstamp, skb->skb_mstamp)) {
3167 /* Do not re-arm RTO if the sack RTT is measured from data sent
3168 * after when the head was last (re)transmitted. Otherwise the
3169 * timeout may continue to extend in loss recovery.
3171 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3174 if (icsk->icsk_ca_ops->pkts_acked) {
3175 struct ack_sample sample = { .pkts_acked = pkts_acked,
3176 .rtt_us = sack->rate->rtt_us,
3177 .in_flight = last_in_flight };
3179 icsk->icsk_ca_ops->pkts_acked(sk, &sample);
3182 #if FASTRETRANS_DEBUG > 0
3183 WARN_ON((int)tp->sacked_out < 0);
3184 WARN_ON((int)tp->lost_out < 0);
3185 WARN_ON((int)tp->retrans_out < 0);
3186 if (!tp->packets_out && tcp_is_sack(tp)) {
3187 icsk = inet_csk(sk);
3189 pr_debug("Leak l=%u %d\n",
3190 tp->lost_out, icsk->icsk_ca_state);
3193 if (tp->sacked_out) {
3194 pr_debug("Leak s=%u %d\n",
3195 tp->sacked_out, icsk->icsk_ca_state);
3198 if (tp->retrans_out) {
3199 pr_debug("Leak r=%u %d\n",
3200 tp->retrans_out, icsk->icsk_ca_state);
3201 tp->retrans_out = 0;
3208 static void tcp_ack_probe(struct sock *sk)
3210 struct inet_connection_sock *icsk = inet_csk(sk);
3211 struct sk_buff *head = tcp_send_head(sk);
3212 const struct tcp_sock *tp = tcp_sk(sk);
3214 /* Was it a usable window open? */
3217 if (!after(TCP_SKB_CB(head)->end_seq, tcp_wnd_end(tp))) {
3218 icsk->icsk_backoff = 0;
3219 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3220 /* Socket must be waked up by subsequent tcp_data_snd_check().
3221 * This function is not for random using!
3224 unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX);
3226 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3231 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3233 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3234 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3237 /* Decide wheather to run the increase function of congestion control. */
3238 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3240 /* If reordering is high then always grow cwnd whenever data is
3241 * delivered regardless of its ordering. Otherwise stay conservative
3242 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3243 * new SACK or ECE mark may first advance cwnd here and later reduce
3244 * cwnd in tcp_fastretrans_alert() based on more states.
3246 if (tcp_sk(sk)->reordering > sock_net(sk)->ipv4.sysctl_tcp_reordering)
3247 return flag & FLAG_FORWARD_PROGRESS;
3249 return flag & FLAG_DATA_ACKED;
3252 /* The "ultimate" congestion control function that aims to replace the rigid
3253 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction).
3254 * It's called toward the end of processing an ACK with precise rate
3255 * information. All transmission or retransmission are delayed afterwards.
3257 static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked,
3258 int flag, const struct rate_sample *rs)
3260 const struct inet_connection_sock *icsk = inet_csk(sk);
3262 if (icsk->icsk_ca_ops->cong_control) {
3263 icsk->icsk_ca_ops->cong_control(sk, rs);
3267 if (tcp_in_cwnd_reduction(sk)) {
3268 /* Reduce cwnd if state mandates */
3269 tcp_cwnd_reduction(sk, acked_sacked, flag);
3270 } else if (tcp_may_raise_cwnd(sk, flag)) {
3271 /* Advance cwnd if state allows */
3272 tcp_cong_avoid(sk, ack, acked_sacked);
3274 tcp_update_pacing_rate(sk);
3277 /* Check that window update is acceptable.
3278 * The function assumes that snd_una<=ack<=snd_next.
3280 static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3281 const u32 ack, const u32 ack_seq,
3284 return after(ack, tp->snd_una) ||
3285 after(ack_seq, tp->snd_wl1) ||
3286 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3289 /* If we update tp->snd_una, also update tp->bytes_acked */
3290 static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack)
3292 u32 delta = ack - tp->snd_una;
3294 sock_owned_by_me((struct sock *)tp);
3295 tp->bytes_acked += delta;
3299 /* If we update tp->rcv_nxt, also update tp->bytes_received */
3300 static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq)
3302 u32 delta = seq - tp->rcv_nxt;
3304 sock_owned_by_me((struct sock *)tp);
3305 tp->bytes_received += delta;
3309 /* Update our send window.
3311 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3312 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3314 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3317 struct tcp_sock *tp = tcp_sk(sk);
3319 u32 nwin = ntohs(tcp_hdr(skb)->window);
3321 if (likely(!tcp_hdr(skb)->syn))
3322 nwin <<= tp->rx_opt.snd_wscale;
3324 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3325 flag |= FLAG_WIN_UPDATE;
3326 tcp_update_wl(tp, ack_seq);
3328 if (tp->snd_wnd != nwin) {
3331 /* Note, it is the only place, where
3332 * fast path is recovered for sending TCP.
3335 tcp_fast_path_check(sk);
3337 if (!tcp_write_queue_empty(sk))
3338 tcp_slow_start_after_idle_check(sk);
3340 if (nwin > tp->max_window) {
3341 tp->max_window = nwin;
3342 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3347 tcp_snd_una_update(tp, ack);
3352 static bool __tcp_oow_rate_limited(struct net *net, int mib_idx,
3353 u32 *last_oow_ack_time)
3355 if (*last_oow_ack_time) {
3356 s32 elapsed = (s32)(tcp_jiffies32 - *last_oow_ack_time);
3358 if (0 <= elapsed && elapsed < net->ipv4.sysctl_tcp_invalid_ratelimit) {
3359 NET_INC_STATS(net, mib_idx);
3360 return true; /* rate-limited: don't send yet! */
3364 *last_oow_ack_time = tcp_jiffies32;
3366 return false; /* not rate-limited: go ahead, send dupack now! */
3369 /* Return true if we're currently rate-limiting out-of-window ACKs and
3370 * thus shouldn't send a dupack right now. We rate-limit dupacks in
3371 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3372 * attacks that send repeated SYNs or ACKs for the same connection. To
3373 * do this, we do not send a duplicate SYNACK or ACK if the remote
3374 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3376 bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb,
3377 int mib_idx, u32 *last_oow_ack_time)
3379 /* Data packets without SYNs are not likely part of an ACK loop. */
3380 if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) &&
3384 return __tcp_oow_rate_limited(net, mib_idx, last_oow_ack_time);
3387 /* RFC 5961 7 [ACK Throttling] */
3388 static void tcp_send_challenge_ack(struct sock *sk, const struct sk_buff *skb)
3390 /* unprotected vars, we dont care of overwrites */
3391 static u32 challenge_timestamp;
3392 static unsigned int challenge_count;
3393 struct tcp_sock *tp = tcp_sk(sk);
3394 struct net *net = sock_net(sk);
3397 /* First check our per-socket dupack rate limit. */
3398 if (__tcp_oow_rate_limited(net,
3399 LINUX_MIB_TCPACKSKIPPEDCHALLENGE,
3400 &tp->last_oow_ack_time))
3403 /* Then check host-wide RFC 5961 rate limit. */
3405 if (now != challenge_timestamp) {
3406 u32 ack_limit = net->ipv4.sysctl_tcp_challenge_ack_limit;
3407 u32 half = (ack_limit + 1) >> 1;
3409 challenge_timestamp = now;
3410 WRITE_ONCE(challenge_count, half + prandom_u32_max(ack_limit));
3412 count = READ_ONCE(challenge_count);
3414 WRITE_ONCE(challenge_count, count - 1);
3415 NET_INC_STATS(net, LINUX_MIB_TCPCHALLENGEACK);
3420 static void tcp_store_ts_recent(struct tcp_sock *tp)
3422 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3423 tp->rx_opt.ts_recent_stamp = get_seconds();
3426 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3428 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3429 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3430 * extra check below makes sure this can only happen
3431 * for pure ACK frames. -DaveM
3433 * Not only, also it occurs for expired timestamps.
3436 if (tcp_paws_check(&tp->rx_opt, 0))
3437 tcp_store_ts_recent(tp);
3441 /* This routine deals with acks during a TLP episode.
3442 * We mark the end of a TLP episode on receiving TLP dupack or when
3443 * ack is after tlp_high_seq.
3444 * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3446 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3448 struct tcp_sock *tp = tcp_sk(sk);
3450 if (before(ack, tp->tlp_high_seq))
3453 if (flag & FLAG_DSACKING_ACK) {
3454 /* This DSACK means original and TLP probe arrived; no loss */
3455 tp->tlp_high_seq = 0;
3456 } else if (after(ack, tp->tlp_high_seq)) {
3457 /* ACK advances: there was a loss, so reduce cwnd. Reset
3458 * tlp_high_seq in tcp_init_cwnd_reduction()
3460 tcp_init_cwnd_reduction(sk);
3461 tcp_set_ca_state(sk, TCP_CA_CWR);
3462 tcp_end_cwnd_reduction(sk);
3463 tcp_try_keep_open(sk);
3464 NET_INC_STATS(sock_net(sk),
3465 LINUX_MIB_TCPLOSSPROBERECOVERY);
3466 } else if (!(flag & (FLAG_SND_UNA_ADVANCED |
3467 FLAG_NOT_DUP | FLAG_DATA_SACKED))) {
3468 /* Pure dupack: original and TLP probe arrived; no loss */
3469 tp->tlp_high_seq = 0;
3473 static inline void tcp_in_ack_event(struct sock *sk, u32 flags)
3475 const struct inet_connection_sock *icsk = inet_csk(sk);
3477 if (icsk->icsk_ca_ops->in_ack_event)
3478 icsk->icsk_ca_ops->in_ack_event(sk, flags);
3481 /* Congestion control has updated the cwnd already. So if we're in
3482 * loss recovery then now we do any new sends (for FRTO) or
3483 * retransmits (for CA_Loss or CA_recovery) that make sense.
3485 static void tcp_xmit_recovery(struct sock *sk, int rexmit)
3487 struct tcp_sock *tp = tcp_sk(sk);
3489 if (rexmit == REXMIT_NONE)
3492 if (unlikely(rexmit == 2)) {
3493 __tcp_push_pending_frames(sk, tcp_current_mss(sk),
3495 if (after(tp->snd_nxt, tp->high_seq))
3499 tcp_xmit_retransmit_queue(sk);
3502 /* This routine deals with incoming acks, but not outgoing ones. */
3503 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3505 struct inet_connection_sock *icsk = inet_csk(sk);
3506 struct tcp_sock *tp = tcp_sk(sk);
3507 struct tcp_sacktag_state sack_state;
3508 struct rate_sample rs = { .prior_delivered = 0 };
3509 u32 prior_snd_una = tp->snd_una;
3510 bool is_sack_reneg = tp->is_sack_reneg;
3511 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3512 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3513 bool is_dupack = false;
3514 int prior_packets = tp->packets_out;
3515 u32 delivered = tp->delivered;
3516 u32 lost = tp->lost;
3517 int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */
3520 sack_state.first_sackt = 0;
3521 sack_state.rate = &rs;
3523 /* We very likely will need to access rtx queue. */
3524 prefetch(sk->tcp_rtx_queue.rb_node);
3526 /* If the ack is older than previous acks
3527 * then we can probably ignore it.
3529 if (before(ack, prior_snd_una)) {
3530 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3531 if (before(ack, prior_snd_una - tp->max_window)) {
3532 if (!(flag & FLAG_NO_CHALLENGE_ACK))
3533 tcp_send_challenge_ack(sk, skb);
3539 /* If the ack includes data we haven't sent yet, discard
3540 * this segment (RFC793 Section 3.9).
3542 if (after(ack, tp->snd_nxt))
3545 if (after(ack, prior_snd_una)) {
3546 flag |= FLAG_SND_UNA_ADVANCED;
3547 icsk->icsk_retransmits = 0;
3550 prior_fack = tcp_is_sack(tp) ? tcp_highest_sack_seq(tp) : tp->snd_una;
3551 rs.prior_in_flight = tcp_packets_in_flight(tp);
3553 /* ts_recent update must be made after we are sure that the packet
3556 if (flag & FLAG_UPDATE_TS_RECENT)
3557 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3559 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3560 /* Window is constant, pure forward advance.
3561 * No more checks are required.
3562 * Note, we use the fact that SND.UNA>=SND.WL2.
3564 tcp_update_wl(tp, ack_seq);
3565 tcp_snd_una_update(tp, ack);
3566 flag |= FLAG_WIN_UPDATE;
3568 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
3570 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPACKS);
3572 u32 ack_ev_flags = CA_ACK_SLOWPATH;
3574 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3577 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3579 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3581 if (TCP_SKB_CB(skb)->sacked)
3582 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3585 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) {
3587 ack_ev_flags |= CA_ACK_ECE;
3590 if (flag & FLAG_WIN_UPDATE)
3591 ack_ev_flags |= CA_ACK_WIN_UPDATE;
3593 tcp_in_ack_event(sk, ack_ev_flags);
3596 /* We passed data and got it acked, remove any soft error
3597 * log. Something worked...
3599 sk->sk_err_soft = 0;
3600 icsk->icsk_probes_out = 0;
3601 tp->rcv_tstamp = tcp_jiffies32;
3605 /* See if we can take anything off of the retransmit queue. */
3606 flag |= tcp_clean_rtx_queue(sk, prior_fack, prior_snd_una, &sack_state);
3608 tcp_rack_update_reo_wnd(sk, &rs);
3610 if (tp->tlp_high_seq)
3611 tcp_process_tlp_ack(sk, ack, flag);
3612 /* If needed, reset TLP/RTO timer; RACK may later override this. */
3613 if (flag & FLAG_SET_XMIT_TIMER)
3614 tcp_set_xmit_timer(sk);
3616 if (tcp_ack_is_dubious(sk, flag)) {
3617 is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
3618 tcp_fastretrans_alert(sk, prior_snd_una, is_dupack, &flag,
3622 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
3625 delivered = tp->delivered - delivered; /* freshly ACKed or SACKed */
3626 lost = tp->lost - lost; /* freshly marked lost */
3627 rs.is_ack_delayed = !!(flag & FLAG_ACK_MAYBE_DELAYED);
3628 tcp_rate_gen(sk, delivered, lost, is_sack_reneg, sack_state.rate);
3629 tcp_cong_control(sk, ack, delivered, flag, sack_state.rate);
3630 tcp_xmit_recovery(sk, rexmit);
3634 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3635 if (flag & FLAG_DSACKING_ACK)
3636 tcp_fastretrans_alert(sk, prior_snd_una, is_dupack, &flag,
3638 /* If this ack opens up a zero window, clear backoff. It was
3639 * being used to time the probes, and is probably far higher than
3640 * it needs to be for normal retransmission.
3644 if (tp->tlp_high_seq)
3645 tcp_process_tlp_ack(sk, ack, flag);
3649 SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3653 /* If data was SACKed, tag it and see if we should send more data.
3654 * If data was DSACKed, see if we can undo a cwnd reduction.
3656 if (TCP_SKB_CB(skb)->sacked) {
3657 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3659 tcp_fastretrans_alert(sk, prior_snd_una, is_dupack, &flag,
3661 tcp_xmit_recovery(sk, rexmit);
3664 SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3668 static void tcp_parse_fastopen_option(int len, const unsigned char *cookie,
3669 bool syn, struct tcp_fastopen_cookie *foc,
3672 /* Valid only in SYN or SYN-ACK with an even length. */
3673 if (!foc || !syn || len < 0 || (len & 1))
3676 if (len >= TCP_FASTOPEN_COOKIE_MIN &&
3677 len <= TCP_FASTOPEN_COOKIE_MAX)
3678 memcpy(foc->val, cookie, len);
3685 static void smc_parse_options(const struct tcphdr *th,
3686 struct tcp_options_received *opt_rx,
3687 const unsigned char *ptr,
3690 #if IS_ENABLED(CONFIG_SMC)
3691 if (static_branch_unlikely(&tcp_have_smc)) {
3692 if (th->syn && !(opsize & 1) &&
3693 opsize >= TCPOLEN_EXP_SMC_BASE &&
3694 get_unaligned_be32(ptr) == TCPOPT_SMC_MAGIC)
3700 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3701 * But, this can also be called on packets in the established flow when
3702 * the fast version below fails.
3704 void tcp_parse_options(const struct net *net,
3705 const struct sk_buff *skb,
3706 struct tcp_options_received *opt_rx, int estab,
3707 struct tcp_fastopen_cookie *foc)
3709 const unsigned char *ptr;
3710 const struct tcphdr *th = tcp_hdr(skb);
3711 int length = (th->doff * 4) - sizeof(struct tcphdr);
3713 ptr = (const unsigned char *)(th + 1);
3714 opt_rx->saw_tstamp = 0;
3716 while (length > 0) {
3717 int opcode = *ptr++;
3723 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3728 if (opsize < 2) /* "silly options" */
3730 if (opsize > length)
3731 return; /* don't parse partial options */
3734 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3735 u16 in_mss = get_unaligned_be16(ptr);
3737 if (opt_rx->user_mss &&
3738 opt_rx->user_mss < in_mss)
3739 in_mss = opt_rx->user_mss;
3740 opt_rx->mss_clamp = in_mss;
3745 if (opsize == TCPOLEN_WINDOW && th->syn &&
3746 !estab && net->ipv4.sysctl_tcp_window_scaling) {
3747 __u8 snd_wscale = *(__u8 *)ptr;
3748 opt_rx->wscale_ok = 1;
3749 if (snd_wscale > TCP_MAX_WSCALE) {
3750 net_info_ratelimited("%s: Illegal window scaling value %d > %u received\n",
3754 snd_wscale = TCP_MAX_WSCALE;
3756 opt_rx->snd_wscale = snd_wscale;
3759 case TCPOPT_TIMESTAMP:
3760 if ((opsize == TCPOLEN_TIMESTAMP) &&
3761 ((estab && opt_rx->tstamp_ok) ||
3762 (!estab && net->ipv4.sysctl_tcp_timestamps))) {
3763 opt_rx->saw_tstamp = 1;
3764 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3765 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3768 case TCPOPT_SACK_PERM:
3769 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3770 !estab && net->ipv4.sysctl_tcp_sack) {
3771 opt_rx->sack_ok = TCP_SACK_SEEN;
3772 tcp_sack_reset(opt_rx);
3777 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3778 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3780 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3783 #ifdef CONFIG_TCP_MD5SIG
3786 * The MD5 Hash has already been
3787 * checked (see tcp_v{4,6}_do_rcv()).
3791 case TCPOPT_FASTOPEN:
3792 tcp_parse_fastopen_option(
3793 opsize - TCPOLEN_FASTOPEN_BASE,
3794 ptr, th->syn, foc, false);
3798 /* Fast Open option shares code 254 using a
3799 * 16 bits magic number.
3801 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE &&
3802 get_unaligned_be16(ptr) ==
3803 TCPOPT_FASTOPEN_MAGIC)
3804 tcp_parse_fastopen_option(opsize -
3805 TCPOLEN_EXP_FASTOPEN_BASE,
3806 ptr + 2, th->syn, foc, true);
3808 smc_parse_options(th, opt_rx, ptr,
3818 EXPORT_SYMBOL(tcp_parse_options);
3820 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
3822 const __be32 *ptr = (const __be32 *)(th + 1);
3824 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3825 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3826 tp->rx_opt.saw_tstamp = 1;
3828 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3831 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
3833 tp->rx_opt.rcv_tsecr = 0;
3839 /* Fast parse options. This hopes to only see timestamps.
3840 * If it is wrong it falls back on tcp_parse_options().
3842 static bool tcp_fast_parse_options(const struct net *net,
3843 const struct sk_buff *skb,
3844 const struct tcphdr *th, struct tcp_sock *tp)
3846 /* In the spirit of fast parsing, compare doff directly to constant
3847 * values. Because equality is used, short doff can be ignored here.
3849 if (th->doff == (sizeof(*th) / 4)) {
3850 tp->rx_opt.saw_tstamp = 0;
3852 } else if (tp->rx_opt.tstamp_ok &&
3853 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3854 if (tcp_parse_aligned_timestamp(tp, th))
3858 tcp_parse_options(net, skb, &tp->rx_opt, 1, NULL);
3859 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3860 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
3865 #ifdef CONFIG_TCP_MD5SIG
3867 * Parse MD5 Signature option
3869 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
3871 int length = (th->doff << 2) - sizeof(*th);
3872 const u8 *ptr = (const u8 *)(th + 1);
3874 /* If the TCP option is too short, we can short cut */
3875 if (length < TCPOLEN_MD5SIG)
3878 while (length > 0) {
3879 int opcode = *ptr++;
3890 if (opsize < 2 || opsize > length)
3892 if (opcode == TCPOPT_MD5SIG)
3893 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
3900 EXPORT_SYMBOL(tcp_parse_md5sig_option);
3903 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3905 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3906 * it can pass through stack. So, the following predicate verifies that
3907 * this segment is not used for anything but congestion avoidance or
3908 * fast retransmit. Moreover, we even are able to eliminate most of such
3909 * second order effects, if we apply some small "replay" window (~RTO)
3910 * to timestamp space.
3912 * All these measures still do not guarantee that we reject wrapped ACKs
3913 * on networks with high bandwidth, when sequence space is recycled fastly,
3914 * but it guarantees that such events will be very rare and do not affect
3915 * connection seriously. This doesn't look nice, but alas, PAWS is really
3918 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3919 * states that events when retransmit arrives after original data are rare.
3920 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3921 * the biggest problem on large power networks even with minor reordering.
3922 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3923 * up to bandwidth of 18Gigabit/sec. 8) ]
3926 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3928 const struct tcp_sock *tp = tcp_sk(sk);
3929 const struct tcphdr *th = tcp_hdr(skb);
3930 u32 seq = TCP_SKB_CB(skb)->seq;
3931 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3933 return (/* 1. Pure ACK with correct sequence number. */
3934 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3936 /* 2. ... and duplicate ACK. */
3937 ack == tp->snd_una &&
3939 /* 3. ... and does not update window. */
3940 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3942 /* 4. ... and sits in replay window. */
3943 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3946 static inline bool tcp_paws_discard(const struct sock *sk,
3947 const struct sk_buff *skb)
3949 const struct tcp_sock *tp = tcp_sk(sk);
3951 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
3952 !tcp_disordered_ack(sk, skb);
3955 /* Check segment sequence number for validity.
3957 * Segment controls are considered valid, if the segment
3958 * fits to the window after truncation to the window. Acceptability
3959 * of data (and SYN, FIN, of course) is checked separately.
3960 * See tcp_data_queue(), for example.
3962 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3963 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3964 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3965 * (borrowed from freebsd)
3968 static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
3970 return !before(end_seq, tp->rcv_wup) &&
3971 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3974 /* When we get a reset we do this. */
3975 void tcp_reset(struct sock *sk)
3977 trace_tcp_receive_reset(sk);
3979 /* We want the right error as BSD sees it (and indeed as we do). */
3980 switch (sk->sk_state) {
3982 sk->sk_err = ECONNREFUSED;
3984 case TCP_CLOSE_WAIT:
3990 sk->sk_err = ECONNRESET;
3992 /* This barrier is coupled with smp_rmb() in tcp_poll() */
3995 tcp_write_queue_purge(sk);
3998 if (!sock_flag(sk, SOCK_DEAD))
3999 sk->sk_error_report(sk);
4003 * Process the FIN bit. This now behaves as it is supposed to work
4004 * and the FIN takes effect when it is validly part of sequence
4005 * space. Not before when we get holes.
4007 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4008 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4011 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4012 * close and we go into CLOSING (and later onto TIME-WAIT)
4014 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4016 void tcp_fin(struct sock *sk)
4018 struct tcp_sock *tp = tcp_sk(sk);
4020 inet_csk_schedule_ack(sk);
4022 sk->sk_shutdown |= RCV_SHUTDOWN;
4023 sock_set_flag(sk, SOCK_DONE);
4025 switch (sk->sk_state) {
4027 case TCP_ESTABLISHED:
4028 /* Move to CLOSE_WAIT */
4029 tcp_set_state(sk, TCP_CLOSE_WAIT);
4030 inet_csk(sk)->icsk_ack.pingpong = 1;
4033 case TCP_CLOSE_WAIT:
4035 /* Received a retransmission of the FIN, do
4040 /* RFC793: Remain in the LAST-ACK state. */
4044 /* This case occurs when a simultaneous close
4045 * happens, we must ack the received FIN and
4046 * enter the CLOSING state.
4049 tcp_set_state(sk, TCP_CLOSING);
4052 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4054 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4057 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4058 * cases we should never reach this piece of code.
4060 pr_err("%s: Impossible, sk->sk_state=%d\n",
4061 __func__, sk->sk_state);
4065 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4066 * Probably, we should reset in this case. For now drop them.
4068 skb_rbtree_purge(&tp->out_of_order_queue);
4069 if (tcp_is_sack(tp))
4070 tcp_sack_reset(&tp->rx_opt);
4073 if (!sock_flag(sk, SOCK_DEAD)) {
4074 sk->sk_state_change(sk);
4076 /* Do not send POLL_HUP for half duplex close. */
4077 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4078 sk->sk_state == TCP_CLOSE)
4079 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4081 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4085 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4088 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4089 if (before(seq, sp->start_seq))
4090 sp->start_seq = seq;
4091 if (after(end_seq, sp->end_seq))
4092 sp->end_seq = end_seq;
4098 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4100 struct tcp_sock *tp = tcp_sk(sk);
4102 if (tcp_is_sack(tp) && sock_net(sk)->ipv4.sysctl_tcp_dsack) {
4105 if (before(seq, tp->rcv_nxt))
4106 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4108 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4110 NET_INC_STATS(sock_net(sk), mib_idx);
4112 tp->rx_opt.dsack = 1;
4113 tp->duplicate_sack[0].start_seq = seq;
4114 tp->duplicate_sack[0].end_seq = end_seq;
4118 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4120 struct tcp_sock *tp = tcp_sk(sk);
4122 if (!tp->rx_opt.dsack)
4123 tcp_dsack_set(sk, seq, end_seq);
4125 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4128 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4130 struct tcp_sock *tp = tcp_sk(sk);
4132 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4133 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4134 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4135 tcp_enter_quickack_mode(sk);
4137 if (tcp_is_sack(tp) && sock_net(sk)->ipv4.sysctl_tcp_dsack) {
4138 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4140 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4141 end_seq = tp->rcv_nxt;
4142 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4149 /* These routines update the SACK block as out-of-order packets arrive or
4150 * in-order packets close up the sequence space.
4152 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4155 struct tcp_sack_block *sp = &tp->selective_acks[0];
4156 struct tcp_sack_block *swalk = sp + 1;
4158 /* See if the recent change to the first SACK eats into
4159 * or hits the sequence space of other SACK blocks, if so coalesce.
4161 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4162 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4165 /* Zap SWALK, by moving every further SACK up by one slot.
4166 * Decrease num_sacks.
4168 tp->rx_opt.num_sacks--;
4169 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4173 this_sack++, swalk++;
4177 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4179 struct tcp_sock *tp = tcp_sk(sk);
4180 struct tcp_sack_block *sp = &tp->selective_acks[0];
4181 int cur_sacks = tp->rx_opt.num_sacks;
4187 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4188 if (tcp_sack_extend(sp, seq, end_seq)) {
4189 /* Rotate this_sack to the first one. */
4190 for (; this_sack > 0; this_sack--, sp--)
4191 swap(*sp, *(sp - 1));
4193 tcp_sack_maybe_coalesce(tp);
4198 /* Could not find an adjacent existing SACK, build a new one,
4199 * put it at the front, and shift everyone else down. We
4200 * always know there is at least one SACK present already here.
4202 * If the sack array is full, forget about the last one.
4204 if (this_sack >= TCP_NUM_SACKS) {
4206 tp->rx_opt.num_sacks--;
4209 for (; this_sack > 0; this_sack--, sp--)
4213 /* Build the new head SACK, and we're done. */
4214 sp->start_seq = seq;
4215 sp->end_seq = end_seq;
4216 tp->rx_opt.num_sacks++;
4219 /* RCV.NXT advances, some SACKs should be eaten. */
4221 static void tcp_sack_remove(struct tcp_sock *tp)
4223 struct tcp_sack_block *sp = &tp->selective_acks[0];
4224 int num_sacks = tp->rx_opt.num_sacks;
4227 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4228 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4229 tp->rx_opt.num_sacks = 0;
4233 for (this_sack = 0; this_sack < num_sacks;) {
4234 /* Check if the start of the sack is covered by RCV.NXT. */
4235 if (!before(tp->rcv_nxt, sp->start_seq)) {
4238 /* RCV.NXT must cover all the block! */
4239 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4241 /* Zap this SACK, by moving forward any other SACKS. */
4242 for (i = this_sack+1; i < num_sacks; i++)
4243 tp->selective_acks[i-1] = tp->selective_acks[i];
4250 tp->rx_opt.num_sacks = num_sacks;
4254 * tcp_try_coalesce - try to merge skb to prior one
4256 * @dest: destination queue
4258 * @from: buffer to add in queue
4259 * @fragstolen: pointer to boolean
4261 * Before queueing skb @from after @to, try to merge them
4262 * to reduce overall memory use and queue lengths, if cost is small.
4263 * Packets in ofo or receive queues can stay a long time.
4264 * Better try to coalesce them right now to avoid future collapses.
4265 * Returns true if caller should free @from instead of queueing it
4267 static bool tcp_try_coalesce(struct sock *sk,
4269 struct sk_buff *from,
4274 *fragstolen = false;
4276 /* Its possible this segment overlaps with prior segment in queue */
4277 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4280 if (!skb_try_coalesce(to, from, fragstolen, &delta))
4283 atomic_add(delta, &sk->sk_rmem_alloc);
4284 sk_mem_charge(sk, delta);
4285 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4286 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4287 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4288 TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
4290 if (TCP_SKB_CB(from)->has_rxtstamp) {
4291 TCP_SKB_CB(to)->has_rxtstamp = true;
4292 to->tstamp = from->tstamp;
4298 static void tcp_drop(struct sock *sk, struct sk_buff *skb)
4300 sk_drops_add(sk, skb);
4304 /* This one checks to see if we can put data from the
4305 * out_of_order queue into the receive_queue.
4307 static void tcp_ofo_queue(struct sock *sk)
4309 struct tcp_sock *tp = tcp_sk(sk);
4310 __u32 dsack_high = tp->rcv_nxt;
4311 bool fin, fragstolen, eaten;
4312 struct sk_buff *skb, *tail;
4315 p = rb_first(&tp->out_of_order_queue);
4318 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4321 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4322 __u32 dsack = dsack_high;
4323 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4324 dsack_high = TCP_SKB_CB(skb)->end_seq;
4325 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4328 rb_erase(&skb->rbnode, &tp->out_of_order_queue);
4330 if (unlikely(!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))) {
4331 SOCK_DEBUG(sk, "ofo packet was already received\n");
4335 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4336 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4337 TCP_SKB_CB(skb)->end_seq);
4339 tail = skb_peek_tail(&sk->sk_receive_queue);
4340 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen);
4341 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4342 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
4344 __skb_queue_tail(&sk->sk_receive_queue, skb);
4346 kfree_skb_partial(skb, fragstolen);
4348 if (unlikely(fin)) {
4350 /* tcp_fin() purges tp->out_of_order_queue,
4351 * so we must end this loop right now.
4358 static bool tcp_prune_ofo_queue(struct sock *sk);
4359 static int tcp_prune_queue(struct sock *sk);
4361 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4364 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4365 !sk_rmem_schedule(sk, skb, size)) {
4367 if (tcp_prune_queue(sk) < 0)
4370 while (!sk_rmem_schedule(sk, skb, size)) {
4371 if (!tcp_prune_ofo_queue(sk))
4378 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4380 struct tcp_sock *tp = tcp_sk(sk);
4381 struct rb_node **p, *parent;
4382 struct sk_buff *skb1;
4386 tcp_ecn_check_ce(tp, skb);
4388 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4389 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFODROP);
4394 /* Disable header prediction. */
4396 inet_csk_schedule_ack(sk);
4398 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4399 seq = TCP_SKB_CB(skb)->seq;
4400 end_seq = TCP_SKB_CB(skb)->end_seq;
4401 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4402 tp->rcv_nxt, seq, end_seq);
4404 p = &tp->out_of_order_queue.rb_node;
4405 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4406 /* Initial out of order segment, build 1 SACK. */
4407 if (tcp_is_sack(tp)) {
4408 tp->rx_opt.num_sacks = 1;
4409 tp->selective_acks[0].start_seq = seq;
4410 tp->selective_acks[0].end_seq = end_seq;
4412 rb_link_node(&skb->rbnode, NULL, p);
4413 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4414 tp->ooo_last_skb = skb;
4418 /* In the typical case, we are adding an skb to the end of the list.
4419 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
4421 if (tcp_try_coalesce(sk, tp->ooo_last_skb,
4422 skb, &fragstolen)) {
4424 tcp_grow_window(sk, skb);
4425 kfree_skb_partial(skb, fragstolen);
4429 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
4430 if (!before(seq, TCP_SKB_CB(tp->ooo_last_skb)->end_seq)) {
4431 parent = &tp->ooo_last_skb->rbnode;
4432 p = &parent->rb_right;
4436 /* Find place to insert this segment. Handle overlaps on the way. */
4440 skb1 = rb_to_skb(parent);
4441 if (before(seq, TCP_SKB_CB(skb1)->seq)) {
4442 p = &parent->rb_left;
4445 if (before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4446 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4447 /* All the bits are present. Drop. */
4448 NET_INC_STATS(sock_net(sk),
4449 LINUX_MIB_TCPOFOMERGE);
4452 tcp_dsack_set(sk, seq, end_seq);
4455 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4456 /* Partial overlap. */
4457 tcp_dsack_set(sk, seq, TCP_SKB_CB(skb1)->end_seq);
4459 /* skb's seq == skb1's seq and skb covers skb1.
4460 * Replace skb1 with skb.
4462 rb_replace_node(&skb1->rbnode, &skb->rbnode,
4463 &tp->out_of_order_queue);
4464 tcp_dsack_extend(sk,
4465 TCP_SKB_CB(skb1)->seq,
4466 TCP_SKB_CB(skb1)->end_seq);
4467 NET_INC_STATS(sock_net(sk),
4468 LINUX_MIB_TCPOFOMERGE);
4472 } else if (tcp_try_coalesce(sk, skb1,
4473 skb, &fragstolen)) {
4476 p = &parent->rb_right;
4479 /* Insert segment into RB tree. */
4480 rb_link_node(&skb->rbnode, parent, p);
4481 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4484 /* Remove other segments covered by skb. */
4485 while ((skb1 = skb_rb_next(skb)) != NULL) {
4486 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4488 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4489 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4493 rb_erase(&skb1->rbnode, &tp->out_of_order_queue);
4494 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4495 TCP_SKB_CB(skb1)->end_seq);
4496 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4499 /* If there is no skb after us, we are the last_skb ! */
4501 tp->ooo_last_skb = skb;
4504 if (tcp_is_sack(tp))
4505 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4508 tcp_grow_window(sk, skb);
4510 skb_set_owner_r(skb, sk);
4514 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, int hdrlen,
4518 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4520 __skb_pull(skb, hdrlen);
4522 tcp_try_coalesce(sk, tail,
4523 skb, fragstolen)) ? 1 : 0;
4524 tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq);
4526 __skb_queue_tail(&sk->sk_receive_queue, skb);
4527 skb_set_owner_r(skb, sk);
4532 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4534 struct sk_buff *skb;
4542 if (size > PAGE_SIZE) {
4543 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS);
4545 data_len = npages << PAGE_SHIFT;
4546 size = data_len + (size & ~PAGE_MASK);
4548 skb = alloc_skb_with_frags(size - data_len, data_len,
4549 PAGE_ALLOC_COSTLY_ORDER,
4550 &err, sk->sk_allocation);
4554 skb_put(skb, size - data_len);
4555 skb->data_len = data_len;
4558 if (tcp_try_rmem_schedule(sk, skb, skb->truesize))
4561 err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size);
4565 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4566 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4567 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4569 if (tcp_queue_rcv(sk, skb, 0, &fragstolen)) {
4570 WARN_ON_ONCE(fragstolen); /* should not happen */
4582 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4584 struct tcp_sock *tp = tcp_sk(sk);
4588 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
4593 __skb_pull(skb, tcp_hdr(skb)->doff * 4);
4595 tcp_ecn_accept_cwr(tp, skb);
4597 tp->rx_opt.dsack = 0;
4599 /* Queue data for delivery to the user.
4600 * Packets in sequence go to the receive queue.
4601 * Out of sequence packets to the out_of_order_queue.
4603 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4604 if (tcp_receive_window(tp) == 0)
4607 /* Ok. In sequence. In window. */
4609 if (skb_queue_len(&sk->sk_receive_queue) == 0)
4610 sk_forced_mem_schedule(sk, skb->truesize);
4611 else if (tcp_try_rmem_schedule(sk, skb, skb->truesize))
4614 eaten = tcp_queue_rcv(sk, skb, 0, &fragstolen);
4615 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4617 tcp_event_data_recv(sk, skb);
4618 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
4621 if (!RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4624 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4625 * gap in queue is filled.
4627 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
4628 inet_csk(sk)->icsk_ack.pingpong = 0;
4631 if (tp->rx_opt.num_sacks)
4632 tcp_sack_remove(tp);
4634 tcp_fast_path_check(sk);
4637 kfree_skb_partial(skb, fragstolen);
4638 if (!sock_flag(sk, SOCK_DEAD))
4639 sk->sk_data_ready(sk);
4643 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4644 /* A retransmit, 2nd most common case. Force an immediate ack. */
4645 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4646 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4649 tcp_enter_quickack_mode(sk);
4650 inet_csk_schedule_ack(sk);
4656 /* Out of window. F.e. zero window probe. */
4657 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4660 tcp_enter_quickack_mode(sk);
4662 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4663 /* Partial packet, seq < rcv_next < end_seq */
4664 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4665 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4666 TCP_SKB_CB(skb)->end_seq);
4668 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4670 /* If window is closed, drop tail of packet. But after
4671 * remembering D-SACK for its head made in previous line.
4673 if (!tcp_receive_window(tp))
4678 tcp_data_queue_ofo(sk, skb);
4681 static struct sk_buff *tcp_skb_next(struct sk_buff *skb, struct sk_buff_head *list)
4684 return !skb_queue_is_last(list, skb) ? skb->next : NULL;
4686 return skb_rb_next(skb);
4689 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4690 struct sk_buff_head *list,
4691 struct rb_root *root)
4693 struct sk_buff *next = tcp_skb_next(skb, list);
4696 __skb_unlink(skb, list);
4698 rb_erase(&skb->rbnode, root);
4701 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4706 /* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */
4707 void tcp_rbtree_insert(struct rb_root *root, struct sk_buff *skb)
4709 struct rb_node **p = &root->rb_node;
4710 struct rb_node *parent = NULL;
4711 struct sk_buff *skb1;
4715 skb1 = rb_to_skb(parent);
4716 if (before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb1)->seq))
4717 p = &parent->rb_left;
4719 p = &parent->rb_right;
4721 rb_link_node(&skb->rbnode, parent, p);
4722 rb_insert_color(&skb->rbnode, root);
4725 /* Collapse contiguous sequence of skbs head..tail with
4726 * sequence numbers start..end.
4728 * If tail is NULL, this means until the end of the queue.
4730 * Segments with FIN/SYN are not collapsed (only because this
4734 tcp_collapse(struct sock *sk, struct sk_buff_head *list, struct rb_root *root,
4735 struct sk_buff *head, struct sk_buff *tail, u32 start, u32 end)
4737 struct sk_buff *skb = head, *n;
4738 struct sk_buff_head tmp;
4741 /* First, check that queue is collapsible and find
4742 * the point where collapsing can be useful.
4745 for (end_of_skbs = true; skb != NULL && skb != tail; skb = n) {
4746 n = tcp_skb_next(skb, list);
4748 /* No new bits? It is possible on ofo queue. */
4749 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4750 skb = tcp_collapse_one(sk, skb, list, root);
4756 /* The first skb to collapse is:
4758 * - bloated or contains data before "start" or
4759 * overlaps to the next one.
4761 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
4762 (tcp_win_from_space(sk, skb->truesize) > skb->len ||
4763 before(TCP_SKB_CB(skb)->seq, start))) {
4764 end_of_skbs = false;
4768 if (n && n != tail &&
4769 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(n)->seq) {
4770 end_of_skbs = false;
4774 /* Decided to skip this, advance start seq. */
4775 start = TCP_SKB_CB(skb)->end_seq;
4778 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4781 __skb_queue_head_init(&tmp);
4783 while (before(start, end)) {
4784 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
4785 struct sk_buff *nskb;
4787 nskb = alloc_skb(copy, GFP_ATOMIC);
4791 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4792 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4794 __skb_queue_before(list, skb, nskb);
4796 __skb_queue_tail(&tmp, nskb); /* defer rbtree insertion */
4797 skb_set_owner_r(nskb, sk);
4799 /* Copy data, releasing collapsed skbs. */
4801 int offset = start - TCP_SKB_CB(skb)->seq;
4802 int size = TCP_SKB_CB(skb)->end_seq - start;
4806 size = min(copy, size);
4807 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4809 TCP_SKB_CB(nskb)->end_seq += size;
4813 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4814 skb = tcp_collapse_one(sk, skb, list, root);
4817 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4823 skb_queue_walk_safe(&tmp, skb, n)
4824 tcp_rbtree_insert(root, skb);
4827 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4828 * and tcp_collapse() them until all the queue is collapsed.
4830 static void tcp_collapse_ofo_queue(struct sock *sk)
4832 struct tcp_sock *tp = tcp_sk(sk);
4833 struct sk_buff *skb, *head;
4836 skb = skb_rb_first(&tp->out_of_order_queue);
4839 tp->ooo_last_skb = skb_rb_last(&tp->out_of_order_queue);
4842 start = TCP_SKB_CB(skb)->seq;
4843 end = TCP_SKB_CB(skb)->end_seq;
4845 for (head = skb;;) {
4846 skb = skb_rb_next(skb);
4848 /* Range is terminated when we see a gap or when
4849 * we are at the queue end.
4852 after(TCP_SKB_CB(skb)->seq, end) ||
4853 before(TCP_SKB_CB(skb)->end_seq, start)) {
4854 tcp_collapse(sk, NULL, &tp->out_of_order_queue,
4855 head, skb, start, end);
4859 if (unlikely(before(TCP_SKB_CB(skb)->seq, start)))
4860 start = TCP_SKB_CB(skb)->seq;
4861 if (after(TCP_SKB_CB(skb)->end_seq, end))
4862 end = TCP_SKB_CB(skb)->end_seq;
4867 * Clean the out-of-order queue to make room.
4868 * We drop high sequences packets to :
4869 * 1) Let a chance for holes to be filled.
4870 * 2) not add too big latencies if thousands of packets sit there.
4871 * (But if application shrinks SO_RCVBUF, we could still end up
4872 * freeing whole queue here)
4874 * Return true if queue has shrunk.
4876 static bool tcp_prune_ofo_queue(struct sock *sk)
4878 struct tcp_sock *tp = tcp_sk(sk);
4879 struct rb_node *node, *prev;
4881 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
4884 NET_INC_STATS(sock_net(sk), LINUX_MIB_OFOPRUNED);
4885 node = &tp->ooo_last_skb->rbnode;
4887 prev = rb_prev(node);
4888 rb_erase(node, &tp->out_of_order_queue);
4889 tcp_drop(sk, rb_to_skb(node));
4891 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf &&
4892 !tcp_under_memory_pressure(sk))
4896 tp->ooo_last_skb = rb_to_skb(prev);
4898 /* Reset SACK state. A conforming SACK implementation will
4899 * do the same at a timeout based retransmit. When a connection
4900 * is in a sad state like this, we care only about integrity
4901 * of the connection not performance.
4903 if (tp->rx_opt.sack_ok)
4904 tcp_sack_reset(&tp->rx_opt);
4908 /* Reduce allocated memory if we can, trying to get
4909 * the socket within its memory limits again.
4911 * Return less than zero if we should start dropping frames
4912 * until the socket owning process reads some of the data
4913 * to stabilize the situation.
4915 static int tcp_prune_queue(struct sock *sk)
4917 struct tcp_sock *tp = tcp_sk(sk);
4919 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4921 NET_INC_STATS(sock_net(sk), LINUX_MIB_PRUNECALLED);
4923 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4924 tcp_clamp_window(sk);
4925 else if (tcp_under_memory_pressure(sk))
4926 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4928 tcp_collapse_ofo_queue(sk);
4929 if (!skb_queue_empty(&sk->sk_receive_queue))
4930 tcp_collapse(sk, &sk->sk_receive_queue, NULL,
4931 skb_peek(&sk->sk_receive_queue),
4933 tp->copied_seq, tp->rcv_nxt);
4936 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4939 /* Collapsing did not help, destructive actions follow.
4940 * This must not ever occur. */
4942 tcp_prune_ofo_queue(sk);
4944 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4947 /* If we are really being abused, tell the caller to silently
4948 * drop receive data on the floor. It will get retransmitted
4949 * and hopefully then we'll have sufficient space.
4951 NET_INC_STATS(sock_net(sk), LINUX_MIB_RCVPRUNED);
4953 /* Massive buffer overcommit. */
4958 static bool tcp_should_expand_sndbuf(const struct sock *sk)
4960 const struct tcp_sock *tp = tcp_sk(sk);
4962 /* If the user specified a specific send buffer setting, do
4965 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4968 /* If we are under global TCP memory pressure, do not expand. */
4969 if (tcp_under_memory_pressure(sk))
4972 /* If we are under soft global TCP memory pressure, do not expand. */
4973 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
4976 /* If we filled the congestion window, do not expand. */
4977 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
4983 /* When incoming ACK allowed to free some skb from write_queue,
4984 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4985 * on the exit from tcp input handler.
4987 * PROBLEM: sndbuf expansion does not work well with largesend.
4989 static void tcp_new_space(struct sock *sk)
4991 struct tcp_sock *tp = tcp_sk(sk);
4993 if (tcp_should_expand_sndbuf(sk)) {
4994 tcp_sndbuf_expand(sk);
4995 tp->snd_cwnd_stamp = tcp_jiffies32;
4998 sk->sk_write_space(sk);
5001 static void tcp_check_space(struct sock *sk)
5003 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
5004 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
5005 /* pairs with tcp_poll() */
5007 if (sk->sk_socket &&
5008 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
5010 if (!test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
5011 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
5016 static inline void tcp_data_snd_check(struct sock *sk)
5018 tcp_push_pending_frames(sk);
5019 tcp_check_space(sk);
5023 * Check if sending an ack is needed.
5025 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
5027 struct tcp_sock *tp = tcp_sk(sk);
5029 /* More than one full frame received... */
5030 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
5031 /* ... and right edge of window advances far enough.
5032 * (tcp_recvmsg() will send ACK otherwise). Or...
5034 __tcp_select_window(sk) >= tp->rcv_wnd) ||
5035 /* We ACK each frame or... */
5036 tcp_in_quickack_mode(sk) ||
5037 /* We have out of order data. */
5038 (ofo_possible && !RB_EMPTY_ROOT(&tp->out_of_order_queue))) {
5039 /* Then ack it now */
5042 /* Else, send delayed ack. */
5043 tcp_send_delayed_ack(sk);
5047 static inline void tcp_ack_snd_check(struct sock *sk)
5049 if (!inet_csk_ack_scheduled(sk)) {
5050 /* We sent a data segment already. */
5053 __tcp_ack_snd_check(sk, 1);
5057 * This routine is only called when we have urgent data
5058 * signaled. Its the 'slow' part of tcp_urg. It could be
5059 * moved inline now as tcp_urg is only called from one
5060 * place. We handle URGent data wrong. We have to - as
5061 * BSD still doesn't use the correction from RFC961.
5062 * For 1003.1g we should support a new option TCP_STDURG to permit
5063 * either form (or just set the sysctl tcp_stdurg).
5066 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
5068 struct tcp_sock *tp = tcp_sk(sk);
5069 u32 ptr = ntohs(th->urg_ptr);
5071 if (ptr && !sock_net(sk)->ipv4.sysctl_tcp_stdurg)
5073 ptr += ntohl(th->seq);
5075 /* Ignore urgent data that we've already seen and read. */
5076 if (after(tp->copied_seq, ptr))
5079 /* Do not replay urg ptr.
5081 * NOTE: interesting situation not covered by specs.
5082 * Misbehaving sender may send urg ptr, pointing to segment,
5083 * which we already have in ofo queue. We are not able to fetch
5084 * such data and will stay in TCP_URG_NOTYET until will be eaten
5085 * by recvmsg(). Seems, we are not obliged to handle such wicked
5086 * situations. But it is worth to think about possibility of some
5087 * DoSes using some hypothetical application level deadlock.
5089 if (before(ptr, tp->rcv_nxt))
5092 /* Do we already have a newer (or duplicate) urgent pointer? */
5093 if (tp->urg_data && !after(ptr, tp->urg_seq))
5096 /* Tell the world about our new urgent pointer. */
5099 /* We may be adding urgent data when the last byte read was
5100 * urgent. To do this requires some care. We cannot just ignore
5101 * tp->copied_seq since we would read the last urgent byte again
5102 * as data, nor can we alter copied_seq until this data arrives
5103 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5105 * NOTE. Double Dutch. Rendering to plain English: author of comment
5106 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5107 * and expect that both A and B disappear from stream. This is _wrong_.
5108 * Though this happens in BSD with high probability, this is occasional.
5109 * Any application relying on this is buggy. Note also, that fix "works"
5110 * only in this artificial test. Insert some normal data between A and B and we will
5111 * decline of BSD again. Verdict: it is better to remove to trap
5114 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5115 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5116 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5118 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5119 __skb_unlink(skb, &sk->sk_receive_queue);
5124 tp->urg_data = TCP_URG_NOTYET;
5127 /* Disable header prediction. */
5131 /* This is the 'fast' part of urgent handling. */
5132 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
5134 struct tcp_sock *tp = tcp_sk(sk);
5136 /* Check if we get a new urgent pointer - normally not. */
5138 tcp_check_urg(sk, th);
5140 /* Do we wait for any urgent data? - normally not... */
5141 if (tp->urg_data == TCP_URG_NOTYET) {
5142 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5145 /* Is the urgent pointer pointing into this packet? */
5146 if (ptr < skb->len) {
5148 if (skb_copy_bits(skb, ptr, &tmp, 1))
5150 tp->urg_data = TCP_URG_VALID | tmp;
5151 if (!sock_flag(sk, SOCK_DEAD))
5152 sk->sk_data_ready(sk);
5157 /* Accept RST for rcv_nxt - 1 after a FIN.
5158 * When tcp connections are abruptly terminated from Mac OSX (via ^C), a
5159 * FIN is sent followed by a RST packet. The RST is sent with the same
5160 * sequence number as the FIN, and thus according to RFC 5961 a challenge
5161 * ACK should be sent. However, Mac OSX rate limits replies to challenge
5162 * ACKs on the closed socket. In addition middleboxes can drop either the
5163 * challenge ACK or a subsequent RST.
5165 static bool tcp_reset_check(const struct sock *sk, const struct sk_buff *skb)
5167 struct tcp_sock *tp = tcp_sk(sk);
5169 return unlikely(TCP_SKB_CB(skb)->seq == (tp->rcv_nxt - 1) &&
5170 (1 << sk->sk_state) & (TCPF_CLOSE_WAIT | TCPF_LAST_ACK |
5174 /* Does PAWS and seqno based validation of an incoming segment, flags will
5175 * play significant role here.
5177 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5178 const struct tcphdr *th, int syn_inerr)
5180 struct tcp_sock *tp = tcp_sk(sk);
5181 bool rst_seq_match = false;
5183 /* RFC1323: H1. Apply PAWS check first. */
5184 if (tcp_fast_parse_options(sock_net(sk), skb, th, tp) &&
5185 tp->rx_opt.saw_tstamp &&
5186 tcp_paws_discard(sk, skb)) {
5188 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5189 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5190 LINUX_MIB_TCPACKSKIPPEDPAWS,
5191 &tp->last_oow_ack_time))
5192 tcp_send_dupack(sk, skb);
5195 /* Reset is accepted even if it did not pass PAWS. */
5198 /* Step 1: check sequence number */
5199 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5200 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5201 * (RST) segments are validated by checking their SEQ-fields."
5202 * And page 69: "If an incoming segment is not acceptable,
5203 * an acknowledgment should be sent in reply (unless the RST
5204 * bit is set, if so drop the segment and return)".
5209 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5210 LINUX_MIB_TCPACKSKIPPEDSEQ,
5211 &tp->last_oow_ack_time))
5212 tcp_send_dupack(sk, skb);
5213 } else if (tcp_reset_check(sk, skb)) {
5219 /* Step 2: check RST bit */
5221 /* RFC 5961 3.2 (extend to match against (RCV.NXT - 1) after a
5222 * FIN and SACK too if available):
5223 * If seq num matches RCV.NXT or (RCV.NXT - 1) after a FIN, or
5224 * the right-most SACK block,
5226 * RESET the connection
5228 * Send a challenge ACK
5230 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt ||
5231 tcp_reset_check(sk, skb)) {
5232 rst_seq_match = true;
5233 } else if (tcp_is_sack(tp) && tp->rx_opt.num_sacks > 0) {
5234 struct tcp_sack_block *sp = &tp->selective_acks[0];
5235 int max_sack = sp[0].end_seq;
5238 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;
5240 max_sack = after(sp[this_sack].end_seq,
5242 sp[this_sack].end_seq : max_sack;
5245 if (TCP_SKB_CB(skb)->seq == max_sack)
5246 rst_seq_match = true;
5252 /* Disable TFO if RST is out-of-order
5253 * and no data has been received
5254 * for current active TFO socket
5256 if (tp->syn_fastopen && !tp->data_segs_in &&
5257 sk->sk_state == TCP_ESTABLISHED)
5258 tcp_fastopen_active_disable(sk);
5259 tcp_send_challenge_ack(sk, skb);
5264 /* step 3: check security and precedence [ignored] */
5266 /* step 4: Check for a SYN
5267 * RFC 5961 4.2 : Send a challenge ack
5272 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5273 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5274 tcp_send_challenge_ack(sk, skb);
5286 * TCP receive function for the ESTABLISHED state.
5288 * It is split into a fast path and a slow path. The fast path is
5290 * - A zero window was announced from us - zero window probing
5291 * is only handled properly in the slow path.
5292 * - Out of order segments arrived.
5293 * - Urgent data is expected.
5294 * - There is no buffer space left
5295 * - Unexpected TCP flags/window values/header lengths are received
5296 * (detected by checking the TCP header against pred_flags)
5297 * - Data is sent in both directions. Fast path only supports pure senders
5298 * or pure receivers (this means either the sequence number or the ack
5299 * value must stay constant)
5300 * - Unexpected TCP option.
5302 * When these conditions are not satisfied it drops into a standard
5303 * receive procedure patterned after RFC793 to handle all cases.
5304 * The first three cases are guaranteed by proper pred_flags setting,
5305 * the rest is checked inline. Fast processing is turned on in
5306 * tcp_data_queue when everything is OK.
5308 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5309 const struct tcphdr *th)
5311 unsigned int len = skb->len;
5312 struct tcp_sock *tp = tcp_sk(sk);
5314 /* TCP congestion window tracking */
5315 trace_tcp_probe(sk, skb);
5317 tcp_mstamp_refresh(tp);
5318 if (unlikely(!sk->sk_rx_dst))
5319 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
5321 * Header prediction.
5322 * The code loosely follows the one in the famous
5323 * "30 instruction TCP receive" Van Jacobson mail.
5325 * Van's trick is to deposit buffers into socket queue
5326 * on a device interrupt, to call tcp_recv function
5327 * on the receive process context and checksum and copy
5328 * the buffer to user space. smart...
5330 * Our current scheme is not silly either but we take the
5331 * extra cost of the net_bh soft interrupt processing...
5332 * We do checksum and copy also but from device to kernel.
5335 tp->rx_opt.saw_tstamp = 0;
5337 /* pred_flags is 0xS?10 << 16 + snd_wnd
5338 * if header_prediction is to be made
5339 * 'S' will always be tp->tcp_header_len >> 2
5340 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5341 * turn it off (when there are holes in the receive
5342 * space for instance)
5343 * PSH flag is ignored.
5346 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5347 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5348 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5349 int tcp_header_len = tp->tcp_header_len;
5351 /* Timestamp header prediction: tcp_header_len
5352 * is automatically equal to th->doff*4 due to pred_flags
5356 /* Check timestamp */
5357 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5358 /* No? Slow path! */
5359 if (!tcp_parse_aligned_timestamp(tp, th))
5362 /* If PAWS failed, check it more carefully in slow path */
5363 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5366 /* DO NOT update ts_recent here, if checksum fails
5367 * and timestamp was corrupted part, it will result
5368 * in a hung connection since we will drop all
5369 * future packets due to the PAWS test.
5373 if (len <= tcp_header_len) {
5374 /* Bulk data transfer: sender */
5375 if (len == tcp_header_len) {
5376 /* Predicted packet is in window by definition.
5377 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5378 * Hence, check seq<=rcv_wup reduces to:
5380 if (tcp_header_len ==
5381 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5382 tp->rcv_nxt == tp->rcv_wup)
5383 tcp_store_ts_recent(tp);
5385 /* We know that such packets are checksummed
5388 tcp_ack(sk, skb, 0);
5390 tcp_data_snd_check(sk);
5392 } else { /* Header too small */
5393 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5398 bool fragstolen = false;
5400 if (tcp_checksum_complete(skb))
5403 if ((int)skb->truesize > sk->sk_forward_alloc)
5406 /* Predicted packet is in window by definition.
5407 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5408 * Hence, check seq<=rcv_wup reduces to:
5410 if (tcp_header_len ==
5411 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5412 tp->rcv_nxt == tp->rcv_wup)
5413 tcp_store_ts_recent(tp);
5415 tcp_rcv_rtt_measure_ts(sk, skb);
5417 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPHITS);
5419 /* Bulk data transfer: receiver */
5420 eaten = tcp_queue_rcv(sk, skb, tcp_header_len,
5423 tcp_event_data_recv(sk, skb);
5425 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5426 /* Well, only one small jumplet in fast path... */
5427 tcp_ack(sk, skb, FLAG_DATA);
5428 tcp_data_snd_check(sk);
5429 if (!inet_csk_ack_scheduled(sk))
5433 __tcp_ack_snd_check(sk, 0);
5436 kfree_skb_partial(skb, fragstolen);
5437 sk->sk_data_ready(sk);
5443 if (len < (th->doff << 2) || tcp_checksum_complete(skb))
5446 if (!th->ack && !th->rst && !th->syn)
5450 * Standard slow path.
5453 if (!tcp_validate_incoming(sk, skb, th, 1))
5457 if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0)
5460 tcp_rcv_rtt_measure_ts(sk, skb);
5462 /* Process urgent data. */
5463 tcp_urg(sk, skb, th);
5465 /* step 7: process the segment text */
5466 tcp_data_queue(sk, skb);
5468 tcp_data_snd_check(sk);
5469 tcp_ack_snd_check(sk);
5473 TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS);
5474 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5479 EXPORT_SYMBOL(tcp_rcv_established);
5481 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5483 struct tcp_sock *tp = tcp_sk(sk);
5484 struct inet_connection_sock *icsk = inet_csk(sk);
5486 tcp_set_state(sk, TCP_ESTABLISHED);
5487 icsk->icsk_ack.lrcvtime = tcp_jiffies32;
5490 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
5491 security_inet_conn_established(sk, skb);
5494 tcp_init_transfer(sk, BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB);
5496 /* Prevent spurious tcp_cwnd_restart() on first data
5499 tp->lsndtime = tcp_jiffies32;
5501 if (sock_flag(sk, SOCK_KEEPOPEN))
5502 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5504 if (!tp->rx_opt.snd_wscale)
5505 __tcp_fast_path_on(tp, tp->snd_wnd);
5510 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
5511 struct tcp_fastopen_cookie *cookie)
5513 struct tcp_sock *tp = tcp_sk(sk);
5514 struct sk_buff *data = tp->syn_data ? tcp_rtx_queue_head(sk) : NULL;
5515 u16 mss = tp->rx_opt.mss_clamp, try_exp = 0;
5516 bool syn_drop = false;
5518 if (mss == tp->rx_opt.user_mss) {
5519 struct tcp_options_received opt;
5521 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5522 tcp_clear_options(&opt);
5523 opt.user_mss = opt.mss_clamp = 0;
5524 tcp_parse_options(sock_net(sk), synack, &opt, 0, NULL);
5525 mss = opt.mss_clamp;
5528 if (!tp->syn_fastopen) {
5529 /* Ignore an unsolicited cookie */
5531 } else if (tp->total_retrans) {
5532 /* SYN timed out and the SYN-ACK neither has a cookie nor
5533 * acknowledges data. Presumably the remote received only
5534 * the retransmitted (regular) SYNs: either the original
5535 * SYN-data or the corresponding SYN-ACK was dropped.
5537 syn_drop = (cookie->len < 0 && data);
5538 } else if (cookie->len < 0 && !tp->syn_data) {
5539 /* We requested a cookie but didn't get it. If we did not use
5540 * the (old) exp opt format then try so next time (try_exp=1).
5541 * Otherwise we go back to use the RFC7413 opt (try_exp=2).
5543 try_exp = tp->syn_fastopen_exp ? 2 : 1;
5546 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp);
5548 if (data) { /* Retransmit unacked data in SYN */
5549 skb_rbtree_walk_from(data) {
5550 if (__tcp_retransmit_skb(sk, data, 1))
5554 NET_INC_STATS(sock_net(sk),
5555 LINUX_MIB_TCPFASTOPENACTIVEFAIL);
5558 tp->syn_data_acked = tp->syn_data;
5559 if (tp->syn_data_acked)
5560 NET_INC_STATS(sock_net(sk),
5561 LINUX_MIB_TCPFASTOPENACTIVE);
5563 tcp_fastopen_add_skb(sk, synack);
5568 static void smc_check_reset_syn(struct tcp_sock *tp)
5570 #if IS_ENABLED(CONFIG_SMC)
5571 if (static_branch_unlikely(&tcp_have_smc)) {
5572 if (tp->syn_smc && !tp->rx_opt.smc_ok)
5578 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5579 const struct tcphdr *th)
5581 struct inet_connection_sock *icsk = inet_csk(sk);
5582 struct tcp_sock *tp = tcp_sk(sk);
5583 struct tcp_fastopen_cookie foc = { .len = -1 };
5584 int saved_clamp = tp->rx_opt.mss_clamp;
5587 tcp_parse_options(sock_net(sk), skb, &tp->rx_opt, 0, &foc);
5588 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
5589 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
5593 * "If the state is SYN-SENT then
5594 * first check the ACK bit
5595 * If the ACK bit is set
5596 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5597 * a reset (unless the RST bit is set, if so drop
5598 * the segment and return)"
5600 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
5601 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt))
5602 goto reset_and_undo;
5604 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5605 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5606 tcp_time_stamp(tp))) {
5607 NET_INC_STATS(sock_net(sk),
5608 LINUX_MIB_PAWSACTIVEREJECTED);
5609 goto reset_and_undo;
5612 /* Now ACK is acceptable.
5614 * "If the RST bit is set
5615 * If the ACK was acceptable then signal the user "error:
5616 * connection reset", drop the segment, enter CLOSED state,
5617 * delete TCB, and return."
5626 * "fifth, if neither of the SYN or RST bits is set then
5627 * drop the segment and return."
5633 goto discard_and_undo;
5636 * "If the SYN bit is on ...
5637 * are acceptable then ...
5638 * (our SYN has been ACKed), change the connection
5639 * state to ESTABLISHED..."
5642 tcp_ecn_rcv_synack(tp, th);
5644 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5645 tcp_ack(sk, skb, FLAG_SLOWPATH);
5647 /* Ok.. it's good. Set up sequence numbers and
5648 * move to established.
5650 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5651 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5653 /* RFC1323: The window in SYN & SYN/ACK segments is
5656 tp->snd_wnd = ntohs(th->window);
5658 if (!tp->rx_opt.wscale_ok) {
5659 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5660 tp->window_clamp = min(tp->window_clamp, 65535U);
5663 if (tp->rx_opt.saw_tstamp) {
5664 tp->rx_opt.tstamp_ok = 1;
5665 tp->tcp_header_len =
5666 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5667 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5668 tcp_store_ts_recent(tp);
5670 tp->tcp_header_len = sizeof(struct tcphdr);
5673 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5674 tcp_initialize_rcv_mss(sk);
5676 /* Remember, tcp_poll() does not lock socket!
5677 * Change state from SYN-SENT only after copied_seq
5678 * is initialized. */
5679 tp->copied_seq = tp->rcv_nxt;
5681 smc_check_reset_syn(tp);
5685 tcp_finish_connect(sk, skb);
5687 fastopen_fail = (tp->syn_fastopen || tp->syn_data) &&
5688 tcp_rcv_fastopen_synack(sk, skb, &foc);
5690 if (!sock_flag(sk, SOCK_DEAD)) {
5691 sk->sk_state_change(sk);
5692 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5696 if (sk->sk_write_pending ||
5697 icsk->icsk_accept_queue.rskq_defer_accept ||
5698 icsk->icsk_ack.pingpong) {
5699 /* Save one ACK. Data will be ready after
5700 * several ticks, if write_pending is set.
5702 * It may be deleted, but with this feature tcpdumps
5703 * look so _wonderfully_ clever, that I was not able
5704 * to stand against the temptation 8) --ANK
5706 inet_csk_schedule_ack(sk);
5707 tcp_enter_quickack_mode(sk);
5708 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5709 TCP_DELACK_MAX, TCP_RTO_MAX);
5720 /* No ACK in the segment */
5724 * "If the RST bit is set
5726 * Otherwise (no ACK) drop the segment and return."
5729 goto discard_and_undo;
5733 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5734 tcp_paws_reject(&tp->rx_opt, 0))
5735 goto discard_and_undo;
5738 /* We see SYN without ACK. It is attempt of
5739 * simultaneous connect with crossed SYNs.
5740 * Particularly, it can be connect to self.
5742 tcp_set_state(sk, TCP_SYN_RECV);
5744 if (tp->rx_opt.saw_tstamp) {
5745 tp->rx_opt.tstamp_ok = 1;
5746 tcp_store_ts_recent(tp);
5747 tp->tcp_header_len =
5748 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5750 tp->tcp_header_len = sizeof(struct tcphdr);
5753 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5754 tp->copied_seq = tp->rcv_nxt;
5755 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5757 /* RFC1323: The window in SYN & SYN/ACK segments is
5760 tp->snd_wnd = ntohs(th->window);
5761 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5762 tp->max_window = tp->snd_wnd;
5764 tcp_ecn_rcv_syn(tp, th);
5767 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5768 tcp_initialize_rcv_mss(sk);
5770 tcp_send_synack(sk);
5772 /* Note, we could accept data and URG from this segment.
5773 * There are no obstacles to make this (except that we must
5774 * either change tcp_recvmsg() to prevent it from returning data
5775 * before 3WHS completes per RFC793, or employ TCP Fast Open).
5777 * However, if we ignore data in ACKless segments sometimes,
5778 * we have no reasons to accept it sometimes.
5779 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5780 * is not flawless. So, discard packet for sanity.
5781 * Uncomment this return to process the data.
5788 /* "fifth, if neither of the SYN or RST bits is set then
5789 * drop the segment and return."
5793 tcp_clear_options(&tp->rx_opt);
5794 tp->rx_opt.mss_clamp = saved_clamp;
5798 tcp_clear_options(&tp->rx_opt);
5799 tp->rx_opt.mss_clamp = saved_clamp;
5804 * This function implements the receiving procedure of RFC 793 for
5805 * all states except ESTABLISHED and TIME_WAIT.
5806 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5807 * address independent.
5810 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb)
5812 struct tcp_sock *tp = tcp_sk(sk);
5813 struct inet_connection_sock *icsk = inet_csk(sk);
5814 const struct tcphdr *th = tcp_hdr(skb);
5815 struct request_sock *req;
5819 switch (sk->sk_state) {
5833 /* It is possible that we process SYN packets from backlog,
5834 * so we need to make sure to disable BH right there.
5837 acceptable = icsk->icsk_af_ops->conn_request(sk, skb) >= 0;
5848 tp->rx_opt.saw_tstamp = 0;
5849 tcp_mstamp_refresh(tp);
5850 queued = tcp_rcv_synsent_state_process(sk, skb, th);
5854 /* Do step6 onward by hand. */
5855 tcp_urg(sk, skb, th);
5857 tcp_data_snd_check(sk);
5861 tcp_mstamp_refresh(tp);
5862 tp->rx_opt.saw_tstamp = 0;
5863 req = tp->fastopen_rsk;
5865 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
5866 sk->sk_state != TCP_FIN_WAIT1);
5868 if (!tcp_check_req(sk, skb, req, true))
5872 if (!th->ack && !th->rst && !th->syn)
5875 if (!tcp_validate_incoming(sk, skb, th, 0))
5878 /* step 5: check the ACK field */
5879 acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
5880 FLAG_UPDATE_TS_RECENT |
5881 FLAG_NO_CHALLENGE_ACK) > 0;
5884 if (sk->sk_state == TCP_SYN_RECV)
5885 return 1; /* send one RST */
5886 tcp_send_challenge_ack(sk, skb);
5889 switch (sk->sk_state) {
5892 tcp_synack_rtt_meas(sk, req);
5894 /* Once we leave TCP_SYN_RECV, we no longer need req
5898 inet_csk(sk)->icsk_retransmits = 0;
5899 reqsk_fastopen_remove(sk, req, false);
5900 /* Re-arm the timer because data may have been sent out.
5901 * This is similar to the regular data transmission case
5902 * when new data has just been ack'ed.
5904 * (TFO) - we could try to be more aggressive and
5905 * retransmitting any data sooner based on when they
5910 tcp_init_transfer(sk, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB);
5911 tp->copied_seq = tp->rcv_nxt;
5914 tcp_set_state(sk, TCP_ESTABLISHED);
5915 sk->sk_state_change(sk);
5917 /* Note, that this wakeup is only for marginal crossed SYN case.
5918 * Passively open sockets are not waked up, because
5919 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
5922 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5924 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5925 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
5926 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5928 if (tp->rx_opt.tstamp_ok)
5929 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5931 if (!inet_csk(sk)->icsk_ca_ops->cong_control)
5932 tcp_update_pacing_rate(sk);
5934 /* Prevent spurious tcp_cwnd_restart() on first data packet */
5935 tp->lsndtime = tcp_jiffies32;
5937 tcp_initialize_rcv_mss(sk);
5938 tcp_fast_path_on(tp);
5941 case TCP_FIN_WAIT1: {
5944 /* If we enter the TCP_FIN_WAIT1 state and we are a
5945 * Fast Open socket and this is the first acceptable
5946 * ACK we have received, this would have acknowledged
5947 * our SYNACK so stop the SYNACK timer.
5950 /* We no longer need the request sock. */
5951 reqsk_fastopen_remove(sk, req, false);
5954 if (tp->snd_una != tp->write_seq)
5957 tcp_set_state(sk, TCP_FIN_WAIT2);
5958 sk->sk_shutdown |= SEND_SHUTDOWN;
5962 if (!sock_flag(sk, SOCK_DEAD)) {
5963 /* Wake up lingering close() */
5964 sk->sk_state_change(sk);
5968 if (tp->linger2 < 0) {
5970 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5973 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5974 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
5975 /* Receive out of order FIN after close() */
5976 if (tp->syn_fastopen && th->fin)
5977 tcp_fastopen_active_disable(sk);
5979 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5983 tmo = tcp_fin_time(sk);
5984 if (tmo > TCP_TIMEWAIT_LEN) {
5985 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5986 } else if (th->fin || sock_owned_by_user(sk)) {
5987 /* Bad case. We could lose such FIN otherwise.
5988 * It is not a big problem, but it looks confusing
5989 * and not so rare event. We still can lose it now,
5990 * if it spins in bh_lock_sock(), but it is really
5993 inet_csk_reset_keepalive_timer(sk, tmo);
5995 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
6002 if (tp->snd_una == tp->write_seq) {
6003 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
6009 if (tp->snd_una == tp->write_seq) {
6010 tcp_update_metrics(sk);
6017 /* step 6: check the URG bit */
6018 tcp_urg(sk, skb, th);
6020 /* step 7: process the segment text */
6021 switch (sk->sk_state) {
6022 case TCP_CLOSE_WAIT:
6025 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
6030 /* RFC 793 says to queue data in these states,
6031 * RFC 1122 says we MUST send a reset.
6032 * BSD 4.4 also does reset.
6034 if (sk->sk_shutdown & RCV_SHUTDOWN) {
6035 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6036 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6037 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6043 case TCP_ESTABLISHED:
6044 tcp_data_queue(sk, skb);
6049 /* tcp_data could move socket to TIME-WAIT */
6050 if (sk->sk_state != TCP_CLOSE) {
6051 tcp_data_snd_check(sk);
6052 tcp_ack_snd_check(sk);
6061 EXPORT_SYMBOL(tcp_rcv_state_process);
6063 static inline void pr_drop_req(struct request_sock *req, __u16 port, int family)
6065 struct inet_request_sock *ireq = inet_rsk(req);
6067 if (family == AF_INET)
6068 net_dbg_ratelimited("drop open request from %pI4/%u\n",
6069 &ireq->ir_rmt_addr, port);
6070 #if IS_ENABLED(CONFIG_IPV6)
6071 else if (family == AF_INET6)
6072 net_dbg_ratelimited("drop open request from %pI6/%u\n",
6073 &ireq->ir_v6_rmt_addr, port);
6077 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
6079 * If we receive a SYN packet with these bits set, it means a
6080 * network is playing bad games with TOS bits. In order to
6081 * avoid possible false congestion notifications, we disable
6082 * TCP ECN negotiation.
6084 * Exception: tcp_ca wants ECN. This is required for DCTCP
6085 * congestion control: Linux DCTCP asserts ECT on all packets,
6086 * including SYN, which is most optimal solution; however,
6087 * others, such as FreeBSD do not.
6089 static void tcp_ecn_create_request(struct request_sock *req,
6090 const struct sk_buff *skb,
6091 const struct sock *listen_sk,
6092 const struct dst_entry *dst)
6094 const struct tcphdr *th = tcp_hdr(skb);
6095 const struct net *net = sock_net(listen_sk);
6096 bool th_ecn = th->ece && th->cwr;
6103 ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield);
6104 ecn_ok_dst = dst_feature(dst, DST_FEATURE_ECN_MASK);
6105 ecn_ok = net->ipv4.sysctl_tcp_ecn || ecn_ok_dst;
6107 if ((!ect && ecn_ok) || tcp_ca_needs_ecn(listen_sk) ||
6108 (ecn_ok_dst & DST_FEATURE_ECN_CA) ||
6109 tcp_bpf_ca_needs_ecn((struct sock *)req))
6110 inet_rsk(req)->ecn_ok = 1;
6113 static void tcp_openreq_init(struct request_sock *req,
6114 const struct tcp_options_received *rx_opt,
6115 struct sk_buff *skb, const struct sock *sk)
6117 struct inet_request_sock *ireq = inet_rsk(req);
6119 req->rsk_rcv_wnd = 0; /* So that tcp_send_synack() knows! */
6121 tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq;
6122 tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
6123 tcp_rsk(req)->snt_synack = tcp_clock_us();
6124 tcp_rsk(req)->last_oow_ack_time = 0;
6125 req->mss = rx_opt->mss_clamp;
6126 req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0;
6127 ireq->tstamp_ok = rx_opt->tstamp_ok;
6128 ireq->sack_ok = rx_opt->sack_ok;
6129 ireq->snd_wscale = rx_opt->snd_wscale;
6130 ireq->wscale_ok = rx_opt->wscale_ok;
6133 ireq->ir_rmt_port = tcp_hdr(skb)->source;
6134 ireq->ir_num = ntohs(tcp_hdr(skb)->dest);
6135 ireq->ir_mark = inet_request_mark(sk, skb);
6136 #if IS_ENABLED(CONFIG_SMC)
6137 ireq->smc_ok = rx_opt->smc_ok;
6141 struct request_sock *inet_reqsk_alloc(const struct request_sock_ops *ops,
6142 struct sock *sk_listener,
6143 bool attach_listener)
6145 struct request_sock *req = reqsk_alloc(ops, sk_listener,
6149 struct inet_request_sock *ireq = inet_rsk(req);
6151 ireq->ireq_opt = NULL;
6152 #if IS_ENABLED(CONFIG_IPV6)
6153 ireq->pktopts = NULL;
6155 atomic64_set(&ireq->ir_cookie, 0);
6156 ireq->ireq_state = TCP_NEW_SYN_RECV;
6157 write_pnet(&ireq->ireq_net, sock_net(sk_listener));
6158 ireq->ireq_family = sk_listener->sk_family;
6163 EXPORT_SYMBOL(inet_reqsk_alloc);
6166 * Return true if a syncookie should be sent
6168 static bool tcp_syn_flood_action(const struct sock *sk,
6169 const struct sk_buff *skb,
6172 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
6173 const char *msg = "Dropping request";
6174 bool want_cookie = false;
6175 struct net *net = sock_net(sk);
6177 #ifdef CONFIG_SYN_COOKIES
6178 if (net->ipv4.sysctl_tcp_syncookies) {
6179 msg = "Sending cookies";
6181 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES);
6184 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP);
6186 if (!queue->synflood_warned &&
6187 net->ipv4.sysctl_tcp_syncookies != 2 &&
6188 xchg(&queue->synflood_warned, 1) == 0)
6189 pr_info("%s: Possible SYN flooding on port %d. %s. Check SNMP counters.\n",
6190 proto, ntohs(tcp_hdr(skb)->dest), msg);
6195 static void tcp_reqsk_record_syn(const struct sock *sk,
6196 struct request_sock *req,
6197 const struct sk_buff *skb)
6199 if (tcp_sk(sk)->save_syn) {
6200 u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb);
6203 copy = kmalloc(len + sizeof(u32), GFP_ATOMIC);
6206 memcpy(©[1], skb_network_header(skb), len);
6207 req->saved_syn = copy;
6212 int tcp_conn_request(struct request_sock_ops *rsk_ops,
6213 const struct tcp_request_sock_ops *af_ops,
6214 struct sock *sk, struct sk_buff *skb)
6216 struct tcp_fastopen_cookie foc = { .len = -1 };
6217 __u32 isn = TCP_SKB_CB(skb)->tcp_tw_isn;
6218 struct tcp_options_received tmp_opt;
6219 struct tcp_sock *tp = tcp_sk(sk);
6220 struct net *net = sock_net(sk);
6221 struct sock *fastopen_sk = NULL;
6222 struct request_sock *req;
6223 bool want_cookie = false;
6224 struct dst_entry *dst;
6227 /* TW buckets are converted to open requests without
6228 * limitations, they conserve resources and peer is
6229 * evidently real one.
6231 if ((net->ipv4.sysctl_tcp_syncookies == 2 ||
6232 inet_csk_reqsk_queue_is_full(sk)) && !isn) {
6233 want_cookie = tcp_syn_flood_action(sk, skb, rsk_ops->slab_name);
6238 if (sk_acceptq_is_full(sk)) {
6239 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
6243 req = inet_reqsk_alloc(rsk_ops, sk, !want_cookie);
6247 tcp_rsk(req)->af_specific = af_ops;
6248 tcp_rsk(req)->ts_off = 0;
6250 tcp_clear_options(&tmp_opt);
6251 tmp_opt.mss_clamp = af_ops->mss_clamp;
6252 tmp_opt.user_mss = tp->rx_opt.user_mss;
6253 tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0,
6254 want_cookie ? NULL : &foc);
6256 if (want_cookie && !tmp_opt.saw_tstamp)
6257 tcp_clear_options(&tmp_opt);
6259 tmp_opt.tstamp_ok = tmp_opt.saw_tstamp;
6260 tcp_openreq_init(req, &tmp_opt, skb, sk);
6261 inet_rsk(req)->no_srccheck = inet_sk(sk)->transparent;
6263 /* Note: tcp_v6_init_req() might override ir_iif for link locals */
6264 inet_rsk(req)->ir_iif = inet_request_bound_dev_if(sk, skb);
6266 af_ops->init_req(req, sk, skb);
6268 if (security_inet_conn_request(sk, skb, req))
6271 if (tmp_opt.tstamp_ok)
6272 tcp_rsk(req)->ts_off = af_ops->init_ts_off(net, skb);
6274 dst = af_ops->route_req(sk, &fl, req);
6278 if (!want_cookie && !isn) {
6279 /* Kill the following clause, if you dislike this way. */
6280 if (!net->ipv4.sysctl_tcp_syncookies &&
6281 (net->ipv4.sysctl_max_syn_backlog - inet_csk_reqsk_queue_len(sk) <
6282 (net->ipv4.sysctl_max_syn_backlog >> 2)) &&
6283 !tcp_peer_is_proven(req, dst)) {
6284 /* Without syncookies last quarter of
6285 * backlog is filled with destinations,
6286 * proven to be alive.
6287 * It means that we continue to communicate
6288 * to destinations, already remembered
6289 * to the moment of synflood.
6291 pr_drop_req(req, ntohs(tcp_hdr(skb)->source),
6293 goto drop_and_release;
6296 isn = af_ops->init_seq(skb);
6299 tcp_ecn_create_request(req, skb, sk, dst);
6302 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss);
6303 req->cookie_ts = tmp_opt.tstamp_ok;
6304 if (!tmp_opt.tstamp_ok)
6305 inet_rsk(req)->ecn_ok = 0;
6308 tcp_rsk(req)->snt_isn = isn;
6309 tcp_rsk(req)->txhash = net_tx_rndhash();
6310 tcp_openreq_init_rwin(req, sk, dst);
6312 tcp_reqsk_record_syn(sk, req, skb);
6313 fastopen_sk = tcp_try_fastopen(sk, skb, req, &foc, dst);
6316 af_ops->send_synack(fastopen_sk, dst, &fl, req,
6317 &foc, TCP_SYNACK_FASTOPEN);
6318 /* Add the child socket directly into the accept queue */
6319 inet_csk_reqsk_queue_add(sk, req, fastopen_sk);
6320 sk->sk_data_ready(sk);
6321 bh_unlock_sock(fastopen_sk);
6322 sock_put(fastopen_sk);
6324 tcp_rsk(req)->tfo_listener = false;
6326 inet_csk_reqsk_queue_hash_add(sk, req,
6327 tcp_timeout_init((struct sock *)req));
6328 af_ops->send_synack(sk, dst, &fl, req, &foc,
6329 !want_cookie ? TCP_SYNACK_NORMAL :
6347 EXPORT_SYMBOL(tcp_conn_request);