1 // SPDX-License-Identifier: GPL-2.0
3 * INET An implementation of the TCP/IP protocol suite for the LINUX
4 * operating system. INET is implemented using the BSD Socket
5 * interface as the means of communication with the user level.
7 * Implementation of the Transmission Control Protocol(TCP).
10 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
11 * Mark Evans, <evansmp@uhura.aston.ac.uk>
12 * Corey Minyard <wf-rch!minyard@relay.EU.net>
13 * Florian La Roche, <flla@stud.uni-sb.de>
14 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
15 * Linus Torvalds, <torvalds@cs.helsinki.fi>
16 * Alan Cox, <gw4pts@gw4pts.ampr.org>
17 * Matthew Dillon, <dillon@apollo.west.oic.com>
18 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
19 * Jorge Cwik, <jorge@laser.satlink.net>
24 * Pedro Roque : Fast Retransmit/Recovery.
26 * Retransmit queue handled by TCP.
27 * Better retransmit timer handling.
28 * New congestion avoidance.
32 * Eric : Fast Retransmit.
33 * Randy Scott : MSS option defines.
34 * Eric Schenk : Fixes to slow start algorithm.
35 * Eric Schenk : Yet another double ACK bug.
36 * Eric Schenk : Delayed ACK bug fixes.
37 * Eric Schenk : Floyd style fast retrans war avoidance.
38 * David S. Miller : Don't allow zero congestion window.
39 * Eric Schenk : Fix retransmitter so that it sends
40 * next packet on ack of previous packet.
41 * Andi Kleen : Moved open_request checking here
42 * and process RSTs for open_requests.
43 * Andi Kleen : Better prune_queue, and other fixes.
44 * Andrey Savochkin: Fix RTT measurements in the presence of
46 * Andrey Savochkin: Check sequence numbers correctly when
47 * removing SACKs due to in sequence incoming
49 * Andi Kleen: Make sure we never ack data there is not
50 * enough room for. Also make this condition
51 * a fatal error if it might still happen.
52 * Andi Kleen: Add tcp_measure_rcv_mss to make
53 * connections with MSS<min(MTU,ann. MSS)
54 * work without delayed acks.
55 * Andi Kleen: Process packets with PSH set in the
57 * J Hadi Salim: ECN support
60 * Panu Kuhlberg: Experimental audit of TCP (re)transmission
61 * engine. Lots of bugs are found.
62 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs
65 #define pr_fmt(fmt) "TCP: " fmt
68 #include <linux/slab.h>
69 #include <linux/module.h>
70 #include <linux/sysctl.h>
71 #include <linux/kernel.h>
72 #include <linux/prefetch.h>
75 #include <net/inet_common.h>
76 #include <linux/ipsec.h>
77 #include <asm/unaligned.h>
78 #include <linux/errqueue.h>
79 #include <trace/events/tcp.h>
80 #include <linux/static_key.h>
81 #include <net/busy_poll.h>
83 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
85 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
86 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
87 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
88 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
89 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
90 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
91 #define FLAG_ECE 0x40 /* ECE in this ACK */
92 #define FLAG_LOST_RETRANS 0x80 /* This ACK marks some retransmission lost */
93 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
94 #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
95 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
96 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
97 #define FLAG_SET_XMIT_TIMER 0x1000 /* Set TLP or RTO timer */
98 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
99 #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
100 #define FLAG_NO_CHALLENGE_ACK 0x8000 /* do not call tcp_send_challenge_ack() */
101 #define FLAG_ACK_MAYBE_DELAYED 0x10000 /* Likely a delayed ACK */
103 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
104 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
105 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE|FLAG_DSACKING_ACK)
106 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
108 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
109 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
111 #define REXMIT_NONE 0 /* no loss recovery to do */
112 #define REXMIT_LOST 1 /* retransmit packets marked lost */
113 #define REXMIT_NEW 2 /* FRTO-style transmit of unsent/new packets */
115 #if IS_ENABLED(CONFIG_TLS_DEVICE)
116 static DEFINE_STATIC_KEY_FALSE(clean_acked_data_enabled);
118 void clean_acked_data_enable(struct inet_connection_sock *icsk,
119 void (*cad)(struct sock *sk, u32 ack_seq))
121 icsk->icsk_clean_acked = cad;
122 static_branch_inc(&clean_acked_data_enabled);
124 EXPORT_SYMBOL_GPL(clean_acked_data_enable);
126 void clean_acked_data_disable(struct inet_connection_sock *icsk)
128 static_branch_dec(&clean_acked_data_enabled);
129 icsk->icsk_clean_acked = NULL;
131 EXPORT_SYMBOL_GPL(clean_acked_data_disable);
134 static void tcp_gro_dev_warn(struct sock *sk, const struct sk_buff *skb,
137 static bool __once __read_mostly;
140 struct net_device *dev;
145 dev = dev_get_by_index_rcu(sock_net(sk), skb->skb_iif);
146 if (!dev || len >= dev->mtu)
147 pr_warn("%s: Driver has suspect GRO implementation, TCP performance may be compromised.\n",
148 dev ? dev->name : "Unknown driver");
153 /* Adapt the MSS value used to make delayed ack decision to the
156 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
158 struct inet_connection_sock *icsk = inet_csk(sk);
159 const unsigned int lss = icsk->icsk_ack.last_seg_size;
162 icsk->icsk_ack.last_seg_size = 0;
164 /* skb->len may jitter because of SACKs, even if peer
165 * sends good full-sized frames.
167 len = skb_shinfo(skb)->gso_size ? : skb->len;
168 if (len >= icsk->icsk_ack.rcv_mss) {
169 icsk->icsk_ack.rcv_mss = min_t(unsigned int, len,
171 /* Account for possibly-removed options */
172 if (unlikely(len > icsk->icsk_ack.rcv_mss +
173 MAX_TCP_OPTION_SPACE))
174 tcp_gro_dev_warn(sk, skb, len);
176 /* Otherwise, we make more careful check taking into account,
177 * that SACKs block is variable.
179 * "len" is invariant segment length, including TCP header.
181 len += skb->data - skb_transport_header(skb);
182 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
183 /* If PSH is not set, packet should be
184 * full sized, provided peer TCP is not badly broken.
185 * This observation (if it is correct 8)) allows
186 * to handle super-low mtu links fairly.
188 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
189 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
190 /* Subtract also invariant (if peer is RFC compliant),
191 * tcp header plus fixed timestamp option length.
192 * Resulting "len" is MSS free of SACK jitter.
194 len -= tcp_sk(sk)->tcp_header_len;
195 icsk->icsk_ack.last_seg_size = len;
197 icsk->icsk_ack.rcv_mss = len;
201 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
202 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
203 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
207 static void tcp_incr_quickack(struct sock *sk, unsigned int max_quickacks)
209 struct inet_connection_sock *icsk = inet_csk(sk);
210 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
214 quickacks = min(quickacks, max_quickacks);
215 if (quickacks > icsk->icsk_ack.quick)
216 icsk->icsk_ack.quick = quickacks;
219 void tcp_enter_quickack_mode(struct sock *sk, unsigned int max_quickacks)
221 struct inet_connection_sock *icsk = inet_csk(sk);
223 tcp_incr_quickack(sk, max_quickacks);
224 icsk->icsk_ack.pingpong = 0;
225 icsk->icsk_ack.ato = TCP_ATO_MIN;
227 EXPORT_SYMBOL(tcp_enter_quickack_mode);
229 /* Send ACKs quickly, if "quick" count is not exhausted
230 * and the session is not interactive.
233 static bool tcp_in_quickack_mode(struct sock *sk)
235 const struct inet_connection_sock *icsk = inet_csk(sk);
236 const struct dst_entry *dst = __sk_dst_get(sk);
238 return (dst && dst_metric(dst, RTAX_QUICKACK)) ||
239 (icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong);
242 static void tcp_ecn_queue_cwr(struct tcp_sock *tp)
244 if (tp->ecn_flags & TCP_ECN_OK)
245 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
248 static void tcp_ecn_accept_cwr(struct sock *sk, const struct sk_buff *skb)
250 if (tcp_hdr(skb)->cwr) {
251 tcp_sk(sk)->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
253 /* If the sender is telling us it has entered CWR, then its
254 * cwnd may be very low (even just 1 packet), so we should ACK
257 if (TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq)
258 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW;
262 static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp)
264 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
267 static void __tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb)
269 struct tcp_sock *tp = tcp_sk(sk);
271 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
272 case INET_ECN_NOT_ECT:
273 /* Funny extension: if ECT is not set on a segment,
274 * and we already seen ECT on a previous segment,
275 * it is probably a retransmit.
277 if (tp->ecn_flags & TCP_ECN_SEEN)
278 tcp_enter_quickack_mode(sk, 2);
281 if (tcp_ca_needs_ecn(sk))
282 tcp_ca_event(sk, CA_EVENT_ECN_IS_CE);
284 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
285 /* Better not delay acks, sender can have a very low cwnd */
286 tcp_enter_quickack_mode(sk, 2);
287 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
289 tp->ecn_flags |= TCP_ECN_SEEN;
292 if (tcp_ca_needs_ecn(sk))
293 tcp_ca_event(sk, CA_EVENT_ECN_NO_CE);
294 tp->ecn_flags |= TCP_ECN_SEEN;
299 static void tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb)
301 if (tcp_sk(sk)->ecn_flags & TCP_ECN_OK)
302 __tcp_ecn_check_ce(sk, skb);
305 static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
307 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
308 tp->ecn_flags &= ~TCP_ECN_OK;
311 static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
313 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
314 tp->ecn_flags &= ~TCP_ECN_OK;
317 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
319 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
324 /* Buffer size and advertised window tuning.
326 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
329 static void tcp_sndbuf_expand(struct sock *sk)
331 const struct tcp_sock *tp = tcp_sk(sk);
332 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
336 /* Worst case is non GSO/TSO : each frame consumes one skb
337 * and skb->head is kmalloced using power of two area of memory
339 per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
341 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
343 per_mss = roundup_pow_of_two(per_mss) +
344 SKB_DATA_ALIGN(sizeof(struct sk_buff));
346 nr_segs = max_t(u32, TCP_INIT_CWND, tp->snd_cwnd);
347 nr_segs = max_t(u32, nr_segs, tp->reordering + 1);
349 /* Fast Recovery (RFC 5681 3.2) :
350 * Cubic needs 1.7 factor, rounded to 2 to include
351 * extra cushion (application might react slowly to EPOLLOUT)
353 sndmem = ca_ops->sndbuf_expand ? ca_ops->sndbuf_expand(sk) : 2;
354 sndmem *= nr_segs * per_mss;
356 if (sk->sk_sndbuf < sndmem)
357 sk->sk_sndbuf = min(sndmem, sock_net(sk)->ipv4.sysctl_tcp_wmem[2]);
360 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
362 * All tcp_full_space() is split to two parts: "network" buffer, allocated
363 * forward and advertised in receiver window (tp->rcv_wnd) and
364 * "application buffer", required to isolate scheduling/application
365 * latencies from network.
366 * window_clamp is maximal advertised window. It can be less than
367 * tcp_full_space(), in this case tcp_full_space() - window_clamp
368 * is reserved for "application" buffer. The less window_clamp is
369 * the smoother our behaviour from viewpoint of network, but the lower
370 * throughput and the higher sensitivity of the connection to losses. 8)
372 * rcv_ssthresh is more strict window_clamp used at "slow start"
373 * phase to predict further behaviour of this connection.
374 * It is used for two goals:
375 * - to enforce header prediction at sender, even when application
376 * requires some significant "application buffer". It is check #1.
377 * - to prevent pruning of receive queue because of misprediction
378 * of receiver window. Check #2.
380 * The scheme does not work when sender sends good segments opening
381 * window and then starts to feed us spaghetti. But it should work
382 * in common situations. Otherwise, we have to rely on queue collapsing.
385 /* Slow part of check#2. */
386 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
388 struct tcp_sock *tp = tcp_sk(sk);
390 int truesize = tcp_win_from_space(sk, skb->truesize) >> 1;
391 int window = tcp_win_from_space(sk, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]) >> 1;
393 while (tp->rcv_ssthresh <= window) {
394 if (truesize <= skb->len)
395 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
403 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
405 struct tcp_sock *tp = tcp_sk(sk);
408 room = min_t(int, tp->window_clamp, tcp_space(sk)) - tp->rcv_ssthresh;
411 if (room > 0 && !tcp_under_memory_pressure(sk)) {
414 /* Check #2. Increase window, if skb with such overhead
415 * will fit to rcvbuf in future.
417 if (tcp_win_from_space(sk, skb->truesize) <= skb->len)
418 incr = 2 * tp->advmss;
420 incr = __tcp_grow_window(sk, skb);
423 incr = max_t(int, incr, 2 * skb->len);
424 tp->rcv_ssthresh += min(room, incr);
425 inet_csk(sk)->icsk_ack.quick |= 1;
430 /* 3. Try to fixup all. It is made immediately after connection enters
433 void tcp_init_buffer_space(struct sock *sk)
435 int tcp_app_win = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_app_win);
436 struct tcp_sock *tp = tcp_sk(sk);
439 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
440 tcp_sndbuf_expand(sk);
442 tcp_mstamp_refresh(tp);
443 tp->rcvq_space.time = tp->tcp_mstamp;
444 tp->rcvq_space.seq = tp->copied_seq;
446 maxwin = tcp_full_space(sk);
448 if (tp->window_clamp >= maxwin) {
449 tp->window_clamp = maxwin;
451 if (tcp_app_win && maxwin > 4 * tp->advmss)
452 tp->window_clamp = max(maxwin -
453 (maxwin >> tcp_app_win),
457 /* Force reservation of one segment. */
459 tp->window_clamp > 2 * tp->advmss &&
460 tp->window_clamp + tp->advmss > maxwin)
461 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
463 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
464 tp->snd_cwnd_stamp = tcp_jiffies32;
465 tp->rcvq_space.space = min3(tp->rcv_ssthresh, tp->rcv_wnd,
466 (u32)TCP_INIT_CWND * tp->advmss);
469 /* 4. Recalculate window clamp after socket hit its memory bounds. */
470 static void tcp_clamp_window(struct sock *sk)
472 struct tcp_sock *tp = tcp_sk(sk);
473 struct inet_connection_sock *icsk = inet_csk(sk);
474 struct net *net = sock_net(sk);
476 icsk->icsk_ack.quick = 0;
478 if (sk->sk_rcvbuf < net->ipv4.sysctl_tcp_rmem[2] &&
479 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
480 !tcp_under_memory_pressure(sk) &&
481 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
482 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
483 net->ipv4.sysctl_tcp_rmem[2]);
485 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
486 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
489 /* Initialize RCV_MSS value.
490 * RCV_MSS is an our guess about MSS used by the peer.
491 * We haven't any direct information about the MSS.
492 * It's better to underestimate the RCV_MSS rather than overestimate.
493 * Overestimations make us ACKing less frequently than needed.
494 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
496 void tcp_initialize_rcv_mss(struct sock *sk)
498 const struct tcp_sock *tp = tcp_sk(sk);
499 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
501 hint = min(hint, tp->rcv_wnd / 2);
502 hint = min(hint, TCP_MSS_DEFAULT);
503 hint = max(hint, TCP_MIN_MSS);
505 inet_csk(sk)->icsk_ack.rcv_mss = hint;
507 EXPORT_SYMBOL(tcp_initialize_rcv_mss);
509 /* Receiver "autotuning" code.
511 * The algorithm for RTT estimation w/o timestamps is based on
512 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
513 * <http://public.lanl.gov/radiant/pubs.html#DRS>
515 * More detail on this code can be found at
516 * <http://staff.psc.edu/jheffner/>,
517 * though this reference is out of date. A new paper
520 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
522 u32 new_sample = tp->rcv_rtt_est.rtt_us;
525 if (new_sample != 0) {
526 /* If we sample in larger samples in the non-timestamp
527 * case, we could grossly overestimate the RTT especially
528 * with chatty applications or bulk transfer apps which
529 * are stalled on filesystem I/O.
531 * Also, since we are only going for a minimum in the
532 * non-timestamp case, we do not smooth things out
533 * else with timestamps disabled convergence takes too
537 m -= (new_sample >> 3);
545 /* No previous measure. */
549 tp->rcv_rtt_est.rtt_us = new_sample;
552 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
556 if (tp->rcv_rtt_est.time == 0)
558 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
560 delta_us = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcv_rtt_est.time);
563 tcp_rcv_rtt_update(tp, delta_us, 1);
566 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
567 tp->rcv_rtt_est.time = tp->tcp_mstamp;
570 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
571 const struct sk_buff *skb)
573 struct tcp_sock *tp = tcp_sk(sk);
575 if (tp->rx_opt.rcv_tsecr == tp->rcv_rtt_last_tsecr)
577 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr;
579 if (TCP_SKB_CB(skb)->end_seq -
580 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss) {
581 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr;
584 if (likely(delta < INT_MAX / (USEC_PER_SEC / TCP_TS_HZ))) {
587 delta_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
588 tcp_rcv_rtt_update(tp, delta_us, 0);
594 * This function should be called every time data is copied to user space.
595 * It calculates the appropriate TCP receive buffer space.
597 void tcp_rcv_space_adjust(struct sock *sk)
599 struct tcp_sock *tp = tcp_sk(sk);
603 trace_tcp_rcv_space_adjust(sk);
605 tcp_mstamp_refresh(tp);
606 time = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcvq_space.time);
607 if (time < (tp->rcv_rtt_est.rtt_us >> 3) || tp->rcv_rtt_est.rtt_us == 0)
610 /* Number of bytes copied to user in last RTT */
611 copied = tp->copied_seq - tp->rcvq_space.seq;
612 if (copied <= tp->rcvq_space.space)
616 * copied = bytes received in previous RTT, our base window
617 * To cope with packet losses, we need a 2x factor
618 * To cope with slow start, and sender growing its cwin by 100 %
619 * every RTT, we need a 4x factor, because the ACK we are sending
620 * now is for the next RTT, not the current one :
621 * <prev RTT . ><current RTT .. ><next RTT .... >
624 if (sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf &&
625 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
629 /* minimal window to cope with packet losses, assuming
630 * steady state. Add some cushion because of small variations.
632 rcvwin = ((u64)copied << 1) + 16 * tp->advmss;
634 /* Accommodate for sender rate increase (eg. slow start) */
635 grow = rcvwin * (copied - tp->rcvq_space.space);
636 do_div(grow, tp->rcvq_space.space);
637 rcvwin += (grow << 1);
639 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
640 while (tcp_win_from_space(sk, rcvmem) < tp->advmss)
643 do_div(rcvwin, tp->advmss);
644 rcvbuf = min_t(u64, rcvwin * rcvmem,
645 sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
646 if (rcvbuf > sk->sk_rcvbuf) {
647 sk->sk_rcvbuf = rcvbuf;
649 /* Make the window clamp follow along. */
650 tp->window_clamp = tcp_win_from_space(sk, rcvbuf);
653 tp->rcvq_space.space = copied;
656 tp->rcvq_space.seq = tp->copied_seq;
657 tp->rcvq_space.time = tp->tcp_mstamp;
660 /* There is something which you must keep in mind when you analyze the
661 * behavior of the tp->ato delayed ack timeout interval. When a
662 * connection starts up, we want to ack as quickly as possible. The
663 * problem is that "good" TCP's do slow start at the beginning of data
664 * transmission. The means that until we send the first few ACK's the
665 * sender will sit on his end and only queue most of his data, because
666 * he can only send snd_cwnd unacked packets at any given time. For
667 * each ACK we send, he increments snd_cwnd and transmits more of his
670 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
672 struct tcp_sock *tp = tcp_sk(sk);
673 struct inet_connection_sock *icsk = inet_csk(sk);
676 inet_csk_schedule_ack(sk);
678 tcp_measure_rcv_mss(sk, skb);
680 tcp_rcv_rtt_measure(tp);
684 if (!icsk->icsk_ack.ato) {
685 /* The _first_ data packet received, initialize
686 * delayed ACK engine.
688 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
689 icsk->icsk_ack.ato = TCP_ATO_MIN;
691 int m = now - icsk->icsk_ack.lrcvtime;
693 if (m <= TCP_ATO_MIN / 2) {
694 /* The fastest case is the first. */
695 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
696 } else if (m < icsk->icsk_ack.ato) {
697 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
698 if (icsk->icsk_ack.ato > icsk->icsk_rto)
699 icsk->icsk_ack.ato = icsk->icsk_rto;
700 } else if (m > icsk->icsk_rto) {
701 /* Too long gap. Apparently sender failed to
702 * restart window, so that we send ACKs quickly.
704 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
708 icsk->icsk_ack.lrcvtime = now;
710 tcp_ecn_check_ce(sk, skb);
713 tcp_grow_window(sk, skb);
716 /* Called to compute a smoothed rtt estimate. The data fed to this
717 * routine either comes from timestamps, or from segments that were
718 * known _not_ to have been retransmitted [see Karn/Partridge
719 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
720 * piece by Van Jacobson.
721 * NOTE: the next three routines used to be one big routine.
722 * To save cycles in the RFC 1323 implementation it was better to break
723 * it up into three procedures. -- erics
725 static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
727 struct tcp_sock *tp = tcp_sk(sk);
728 long m = mrtt_us; /* RTT */
729 u32 srtt = tp->srtt_us;
731 /* The following amusing code comes from Jacobson's
732 * article in SIGCOMM '88. Note that rtt and mdev
733 * are scaled versions of rtt and mean deviation.
734 * This is designed to be as fast as possible
735 * m stands for "measurement".
737 * On a 1990 paper the rto value is changed to:
738 * RTO = rtt + 4 * mdev
740 * Funny. This algorithm seems to be very broken.
741 * These formulae increase RTO, when it should be decreased, increase
742 * too slowly, when it should be increased quickly, decrease too quickly
743 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
744 * does not matter how to _calculate_ it. Seems, it was trap
745 * that VJ failed to avoid. 8)
748 m -= (srtt >> 3); /* m is now error in rtt est */
749 srtt += m; /* rtt = 7/8 rtt + 1/8 new */
751 m = -m; /* m is now abs(error) */
752 m -= (tp->mdev_us >> 2); /* similar update on mdev */
753 /* This is similar to one of Eifel findings.
754 * Eifel blocks mdev updates when rtt decreases.
755 * This solution is a bit different: we use finer gain
756 * for mdev in this case (alpha*beta).
757 * Like Eifel it also prevents growth of rto,
758 * but also it limits too fast rto decreases,
759 * happening in pure Eifel.
764 m -= (tp->mdev_us >> 2); /* similar update on mdev */
766 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */
767 if (tp->mdev_us > tp->mdev_max_us) {
768 tp->mdev_max_us = tp->mdev_us;
769 if (tp->mdev_max_us > tp->rttvar_us)
770 tp->rttvar_us = tp->mdev_max_us;
772 if (after(tp->snd_una, tp->rtt_seq)) {
773 if (tp->mdev_max_us < tp->rttvar_us)
774 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
775 tp->rtt_seq = tp->snd_nxt;
776 tp->mdev_max_us = tcp_rto_min_us(sk);
779 /* no previous measure. */
780 srtt = m << 3; /* take the measured time to be rtt */
781 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */
782 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
783 tp->mdev_max_us = tp->rttvar_us;
784 tp->rtt_seq = tp->snd_nxt;
786 tp->srtt_us = max(1U, srtt);
789 static void tcp_update_pacing_rate(struct sock *sk)
791 const struct tcp_sock *tp = tcp_sk(sk);
794 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
795 rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3);
797 /* current rate is (cwnd * mss) / srtt
798 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
799 * In Congestion Avoidance phase, set it to 120 % the current rate.
801 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
802 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
803 * end of slow start and should slow down.
805 if (tp->snd_cwnd < tp->snd_ssthresh / 2)
806 rate *= sock_net(sk)->ipv4.sysctl_tcp_pacing_ss_ratio;
808 rate *= sock_net(sk)->ipv4.sysctl_tcp_pacing_ca_ratio;
810 rate *= max(tp->snd_cwnd, tp->packets_out);
812 if (likely(tp->srtt_us))
813 do_div(rate, tp->srtt_us);
815 /* WRITE_ONCE() is needed because sch_fq fetches sk_pacing_rate
816 * without any lock. We want to make sure compiler wont store
817 * intermediate values in this location.
819 WRITE_ONCE(sk->sk_pacing_rate, min_t(u64, rate,
820 sk->sk_max_pacing_rate));
823 /* Calculate rto without backoff. This is the second half of Van Jacobson's
824 * routine referred to above.
826 static void tcp_set_rto(struct sock *sk)
828 const struct tcp_sock *tp = tcp_sk(sk);
829 /* Old crap is replaced with new one. 8)
832 * 1. If rtt variance happened to be less 50msec, it is hallucination.
833 * It cannot be less due to utterly erratic ACK generation made
834 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
835 * to do with delayed acks, because at cwnd>2 true delack timeout
836 * is invisible. Actually, Linux-2.4 also generates erratic
837 * ACKs in some circumstances.
839 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
841 /* 2. Fixups made earlier cannot be right.
842 * If we do not estimate RTO correctly without them,
843 * all the algo is pure shit and should be replaced
844 * with correct one. It is exactly, which we pretend to do.
847 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
848 * guarantees that rto is higher.
853 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
855 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
858 cwnd = TCP_INIT_CWND;
859 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
862 /* Take a notice that peer is sending D-SACKs */
863 static void tcp_dsack_seen(struct tcp_sock *tp)
865 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
866 tp->rack.dsack_seen = 1;
870 /* It's reordering when higher sequence was delivered (i.e. sacked) before
871 * some lower never-retransmitted sequence ("low_seq"). The maximum reordering
872 * distance is approximated in full-mss packet distance ("reordering").
874 static void tcp_check_sack_reordering(struct sock *sk, const u32 low_seq,
877 struct tcp_sock *tp = tcp_sk(sk);
878 const u32 mss = tp->mss_cache;
881 fack = tcp_highest_sack_seq(tp);
882 if (!before(low_seq, fack))
885 metric = fack - low_seq;
886 if ((metric > tp->reordering * mss) && mss) {
887 #if FASTRETRANS_DEBUG > 1
888 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
889 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
893 tp->undo_marker ? tp->undo_retrans : 0);
895 tp->reordering = min_t(u32, (metric + mss - 1) / mss,
896 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_max_reordering));
899 /* This exciting event is worth to be remembered. 8) */
901 NET_INC_STATS(sock_net(sk),
902 ts ? LINUX_MIB_TCPTSREORDER : LINUX_MIB_TCPSACKREORDER);
905 /* This must be called before lost_out is incremented */
906 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
908 if ((!tp->retransmit_skb_hint && tp->retrans_out >= tp->lost_out) ||
909 (tp->retransmit_skb_hint &&
910 before(TCP_SKB_CB(skb)->seq,
911 TCP_SKB_CB(tp->retransmit_skb_hint)->seq)))
912 tp->retransmit_skb_hint = skb;
915 /* Sum the number of packets on the wire we have marked as lost.
916 * There are two cases we care about here:
917 * a) Packet hasn't been marked lost (nor retransmitted),
918 * and this is the first loss.
919 * b) Packet has been marked both lost and retransmitted,
920 * and this means we think it was lost again.
922 static void tcp_sum_lost(struct tcp_sock *tp, struct sk_buff *skb)
924 __u8 sacked = TCP_SKB_CB(skb)->sacked;
926 if (!(sacked & TCPCB_LOST) ||
927 ((sacked & TCPCB_LOST) && (sacked & TCPCB_SACKED_RETRANS)))
928 tp->lost += tcp_skb_pcount(skb);
931 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
933 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
934 tcp_verify_retransmit_hint(tp, skb);
936 tp->lost_out += tcp_skb_pcount(skb);
937 tcp_sum_lost(tp, skb);
938 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
942 void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp, struct sk_buff *skb)
944 tcp_verify_retransmit_hint(tp, skb);
946 tcp_sum_lost(tp, skb);
947 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
948 tp->lost_out += tcp_skb_pcount(skb);
949 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
953 /* This procedure tags the retransmission queue when SACKs arrive.
955 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
956 * Packets in queue with these bits set are counted in variables
957 * sacked_out, retrans_out and lost_out, correspondingly.
959 * Valid combinations are:
960 * Tag InFlight Description
961 * 0 1 - orig segment is in flight.
962 * S 0 - nothing flies, orig reached receiver.
963 * L 0 - nothing flies, orig lost by net.
964 * R 2 - both orig and retransmit are in flight.
965 * L|R 1 - orig is lost, retransmit is in flight.
966 * S|R 1 - orig reached receiver, retrans is still in flight.
967 * (L|S|R is logically valid, it could occur when L|R is sacked,
968 * but it is equivalent to plain S and code short-curcuits it to S.
969 * L|S is logically invalid, it would mean -1 packet in flight 8))
971 * These 6 states form finite state machine, controlled by the following events:
972 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
973 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
974 * 3. Loss detection event of two flavors:
975 * A. Scoreboard estimator decided the packet is lost.
976 * A'. Reno "three dupacks" marks head of queue lost.
977 * B. SACK arrives sacking SND.NXT at the moment, when the
978 * segment was retransmitted.
979 * 4. D-SACK added new rule: D-SACK changes any tag to S.
981 * It is pleasant to note, that state diagram turns out to be commutative,
982 * so that we are allowed not to be bothered by order of our actions,
983 * when multiple events arrive simultaneously. (see the function below).
985 * Reordering detection.
986 * --------------------
987 * Reordering metric is maximal distance, which a packet can be displaced
988 * in packet stream. With SACKs we can estimate it:
990 * 1. SACK fills old hole and the corresponding segment was not
991 * ever retransmitted -> reordering. Alas, we cannot use it
992 * when segment was retransmitted.
993 * 2. The last flaw is solved with D-SACK. D-SACK arrives
994 * for retransmitted and already SACKed segment -> reordering..
995 * Both of these heuristics are not used in Loss state, when we cannot
996 * account for retransmits accurately.
998 * SACK block validation.
999 * ----------------------
1001 * SACK block range validation checks that the received SACK block fits to
1002 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1003 * Note that SND.UNA is not included to the range though being valid because
1004 * it means that the receiver is rather inconsistent with itself reporting
1005 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1006 * perfectly valid, however, in light of RFC2018 which explicitly states
1007 * that "SACK block MUST reflect the newest segment. Even if the newest
1008 * segment is going to be discarded ...", not that it looks very clever
1009 * in case of head skb. Due to potentional receiver driven attacks, we
1010 * choose to avoid immediate execution of a walk in write queue due to
1011 * reneging and defer head skb's loss recovery to standard loss recovery
1012 * procedure that will eventually trigger (nothing forbids us doing this).
1014 * Implements also blockage to start_seq wrap-around. Problem lies in the
1015 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1016 * there's no guarantee that it will be before snd_nxt (n). The problem
1017 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1020 * <- outs wnd -> <- wrapzone ->
1021 * u e n u_w e_w s n_w
1023 * |<------------+------+----- TCP seqno space --------------+---------->|
1024 * ...-- <2^31 ->| |<--------...
1025 * ...---- >2^31 ------>| |<--------...
1027 * Current code wouldn't be vulnerable but it's better still to discard such
1028 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1029 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1030 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1031 * equal to the ideal case (infinite seqno space without wrap caused issues).
1033 * With D-SACK the lower bound is extended to cover sequence space below
1034 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1035 * again, D-SACK block must not to go across snd_una (for the same reason as
1036 * for the normal SACK blocks, explained above). But there all simplicity
1037 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1038 * fully below undo_marker they do not affect behavior in anyway and can
1039 * therefore be safely ignored. In rare cases (which are more or less
1040 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1041 * fragmentation and packet reordering past skb's retransmission. To consider
1042 * them correctly, the acceptable range must be extended even more though
1043 * the exact amount is rather hard to quantify. However, tp->max_window can
1044 * be used as an exaggerated estimate.
1046 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
1047 u32 start_seq, u32 end_seq)
1049 /* Too far in future, or reversed (interpretation is ambiguous) */
1050 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1053 /* Nasty start_seq wrap-around check (see comments above) */
1054 if (!before(start_seq, tp->snd_nxt))
1057 /* In outstanding window? ...This is valid exit for D-SACKs too.
1058 * start_seq == snd_una is non-sensical (see comments above)
1060 if (after(start_seq, tp->snd_una))
1063 if (!is_dsack || !tp->undo_marker)
1066 /* ...Then it's D-SACK, and must reside below snd_una completely */
1067 if (after(end_seq, tp->snd_una))
1070 if (!before(start_seq, tp->undo_marker))
1074 if (!after(end_seq, tp->undo_marker))
1077 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1078 * start_seq < undo_marker and end_seq >= undo_marker.
1080 return !before(start_seq, end_seq - tp->max_window);
1083 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1084 struct tcp_sack_block_wire *sp, int num_sacks,
1087 struct tcp_sock *tp = tcp_sk(sk);
1088 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1089 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1090 bool dup_sack = false;
1092 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1095 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1096 } else if (num_sacks > 1) {
1097 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1098 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1100 if (!after(end_seq_0, end_seq_1) &&
1101 !before(start_seq_0, start_seq_1)) {
1104 NET_INC_STATS(sock_net(sk),
1105 LINUX_MIB_TCPDSACKOFORECV);
1109 /* D-SACK for already forgotten data... Do dumb counting. */
1110 if (dup_sack && tp->undo_marker && tp->undo_retrans > 0 &&
1111 !after(end_seq_0, prior_snd_una) &&
1112 after(end_seq_0, tp->undo_marker))
1118 struct tcp_sacktag_state {
1120 /* Timestamps for earliest and latest never-retransmitted segment
1121 * that was SACKed. RTO needs the earliest RTT to stay conservative,
1122 * but congestion control should still get an accurate delay signal.
1126 struct rate_sample *rate;
1128 unsigned int mss_now;
1131 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1132 * the incoming SACK may not exactly match but we can find smaller MSS
1133 * aligned portion of it that matches. Therefore we might need to fragment
1134 * which may fail and creates some hassle (caller must handle error case
1137 * FIXME: this could be merged to shift decision code
1139 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1140 u32 start_seq, u32 end_seq)
1144 unsigned int pkt_len;
1147 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1148 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1150 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1151 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1152 mss = tcp_skb_mss(skb);
1153 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1156 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1160 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1165 /* Round if necessary so that SACKs cover only full MSSes
1166 * and/or the remaining small portion (if present)
1168 if (pkt_len > mss) {
1169 unsigned int new_len = (pkt_len / mss) * mss;
1170 if (!in_sack && new_len < pkt_len)
1175 if (pkt_len >= skb->len && !in_sack)
1178 err = tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
1179 pkt_len, mss, GFP_ATOMIC);
1187 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1188 static u8 tcp_sacktag_one(struct sock *sk,
1189 struct tcp_sacktag_state *state, u8 sacked,
1190 u32 start_seq, u32 end_seq,
1191 int dup_sack, int pcount,
1194 struct tcp_sock *tp = tcp_sk(sk);
1196 /* Account D-SACK for retransmitted packet. */
1197 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1198 if (tp->undo_marker && tp->undo_retrans > 0 &&
1199 after(end_seq, tp->undo_marker))
1200 tp->undo_retrans = max_t(int, 0, tp->undo_retrans - pcount);
1201 if ((sacked & TCPCB_SACKED_ACKED) &&
1202 before(start_seq, state->reord))
1203 state->reord = start_seq;
1206 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1207 if (!after(end_seq, tp->snd_una))
1210 if (!(sacked & TCPCB_SACKED_ACKED)) {
1211 tcp_rack_advance(tp, sacked, end_seq, xmit_time);
1213 if (sacked & TCPCB_SACKED_RETRANS) {
1214 /* If the segment is not tagged as lost,
1215 * we do not clear RETRANS, believing
1216 * that retransmission is still in flight.
1218 if (sacked & TCPCB_LOST) {
1219 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1220 tp->lost_out -= pcount;
1221 tp->retrans_out -= pcount;
1224 if (!(sacked & TCPCB_RETRANS)) {
1225 /* New sack for not retransmitted frame,
1226 * which was in hole. It is reordering.
1228 if (before(start_seq,
1229 tcp_highest_sack_seq(tp)) &&
1230 before(start_seq, state->reord))
1231 state->reord = start_seq;
1233 if (!after(end_seq, tp->high_seq))
1234 state->flag |= FLAG_ORIG_SACK_ACKED;
1235 if (state->first_sackt == 0)
1236 state->first_sackt = xmit_time;
1237 state->last_sackt = xmit_time;
1240 if (sacked & TCPCB_LOST) {
1241 sacked &= ~TCPCB_LOST;
1242 tp->lost_out -= pcount;
1246 sacked |= TCPCB_SACKED_ACKED;
1247 state->flag |= FLAG_DATA_SACKED;
1248 tp->sacked_out += pcount;
1249 tp->delivered += pcount; /* Out-of-order packets delivered */
1251 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1252 if (tp->lost_skb_hint &&
1253 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1254 tp->lost_cnt_hint += pcount;
1257 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1258 * frames and clear it. undo_retrans is decreased above, L|R frames
1259 * are accounted above as well.
1261 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1262 sacked &= ~TCPCB_SACKED_RETRANS;
1263 tp->retrans_out -= pcount;
1269 /* Shift newly-SACKed bytes from this skb to the immediately previous
1270 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1272 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *prev,
1273 struct sk_buff *skb,
1274 struct tcp_sacktag_state *state,
1275 unsigned int pcount, int shifted, int mss,
1278 struct tcp_sock *tp = tcp_sk(sk);
1279 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1280 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
1284 /* Adjust counters and hints for the newly sacked sequence
1285 * range but discard the return value since prev is already
1286 * marked. We must tag the range first because the seq
1287 * advancement below implicitly advances
1288 * tcp_highest_sack_seq() when skb is highest_sack.
1290 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1291 start_seq, end_seq, dup_sack, pcount,
1293 tcp_rate_skb_delivered(sk, skb, state->rate);
1295 if (skb == tp->lost_skb_hint)
1296 tp->lost_cnt_hint += pcount;
1298 TCP_SKB_CB(prev)->end_seq += shifted;
1299 TCP_SKB_CB(skb)->seq += shifted;
1301 tcp_skb_pcount_add(prev, pcount);
1302 WARN_ON_ONCE(tcp_skb_pcount(skb) < pcount);
1303 tcp_skb_pcount_add(skb, -pcount);
1305 /* When we're adding to gso_segs == 1, gso_size will be zero,
1306 * in theory this shouldn't be necessary but as long as DSACK
1307 * code can come after this skb later on it's better to keep
1308 * setting gso_size to something.
1310 if (!TCP_SKB_CB(prev)->tcp_gso_size)
1311 TCP_SKB_CB(prev)->tcp_gso_size = mss;
1313 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1314 if (tcp_skb_pcount(skb) <= 1)
1315 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1317 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1318 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1321 BUG_ON(!tcp_skb_pcount(skb));
1322 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1326 /* Whole SKB was eaten :-) */
1328 if (skb == tp->retransmit_skb_hint)
1329 tp->retransmit_skb_hint = prev;
1330 if (skb == tp->lost_skb_hint) {
1331 tp->lost_skb_hint = prev;
1332 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1335 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1336 TCP_SKB_CB(prev)->eor = TCP_SKB_CB(skb)->eor;
1337 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1338 TCP_SKB_CB(prev)->end_seq++;
1340 if (skb == tcp_highest_sack(sk))
1341 tcp_advance_highest_sack(sk, skb);
1343 tcp_skb_collapse_tstamp(prev, skb);
1344 if (unlikely(TCP_SKB_CB(prev)->tx.delivered_mstamp))
1345 TCP_SKB_CB(prev)->tx.delivered_mstamp = 0;
1347 tcp_rtx_queue_unlink_and_free(skb, sk);
1349 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKMERGED);
1354 /* I wish gso_size would have a bit more sane initialization than
1355 * something-or-zero which complicates things
1357 static int tcp_skb_seglen(const struct sk_buff *skb)
1359 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1362 /* Shifting pages past head area doesn't work */
1363 static int skb_can_shift(const struct sk_buff *skb)
1365 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1368 int tcp_skb_shift(struct sk_buff *to, struct sk_buff *from,
1369 int pcount, int shiftlen)
1371 /* TCP min gso_size is 8 bytes (TCP_MIN_GSO_SIZE)
1372 * Since TCP_SKB_CB(skb)->tcp_gso_segs is 16 bits, we need
1373 * to make sure not storing more than 65535 * 8 bytes per skb,
1374 * even if current MSS is bigger.
1376 if (unlikely(to->len + shiftlen >= 65535 * TCP_MIN_GSO_SIZE))
1378 if (unlikely(tcp_skb_pcount(to) + pcount > 65535))
1380 return skb_shift(to, from, shiftlen);
1383 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1386 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1387 struct tcp_sacktag_state *state,
1388 u32 start_seq, u32 end_seq,
1391 struct tcp_sock *tp = tcp_sk(sk);
1392 struct sk_buff *prev;
1398 /* Normally R but no L won't result in plain S */
1400 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1402 if (!skb_can_shift(skb))
1404 /* This frame is about to be dropped (was ACKed). */
1405 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1408 /* Can only happen with delayed DSACK + discard craziness */
1409 prev = skb_rb_prev(skb);
1413 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1416 if (!tcp_skb_can_collapse_to(prev))
1419 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1420 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1424 pcount = tcp_skb_pcount(skb);
1425 mss = tcp_skb_seglen(skb);
1427 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1428 * drop this restriction as unnecessary
1430 if (mss != tcp_skb_seglen(prev))
1433 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1435 /* CHECKME: This is non-MSS split case only?, this will
1436 * cause skipped skbs due to advancing loop btw, original
1437 * has that feature too
1439 if (tcp_skb_pcount(skb) <= 1)
1442 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1444 /* TODO: head merge to next could be attempted here
1445 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1446 * though it might not be worth of the additional hassle
1448 * ...we can probably just fallback to what was done
1449 * previously. We could try merging non-SACKed ones
1450 * as well but it probably isn't going to buy off
1451 * because later SACKs might again split them, and
1452 * it would make skb timestamp tracking considerably
1458 len = end_seq - TCP_SKB_CB(skb)->seq;
1460 BUG_ON(len > skb->len);
1462 /* MSS boundaries should be honoured or else pcount will
1463 * severely break even though it makes things bit trickier.
1464 * Optimize common case to avoid most of the divides
1466 mss = tcp_skb_mss(skb);
1468 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1469 * drop this restriction as unnecessary
1471 if (mss != tcp_skb_seglen(prev))
1476 } else if (len < mss) {
1484 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1485 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1488 if (!tcp_skb_shift(prev, skb, pcount, len))
1490 if (!tcp_shifted_skb(sk, prev, skb, state, pcount, len, mss, dup_sack))
1493 /* Hole filled allows collapsing with the next as well, this is very
1494 * useful when hole on every nth skb pattern happens
1496 skb = skb_rb_next(prev);
1500 if (!skb_can_shift(skb) ||
1501 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1502 (mss != tcp_skb_seglen(skb)))
1506 pcount = tcp_skb_pcount(skb);
1507 if (tcp_skb_shift(prev, skb, pcount, len))
1508 tcp_shifted_skb(sk, prev, skb, state, pcount,
1518 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1522 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1523 struct tcp_sack_block *next_dup,
1524 struct tcp_sacktag_state *state,
1525 u32 start_seq, u32 end_seq,
1528 struct tcp_sock *tp = tcp_sk(sk);
1529 struct sk_buff *tmp;
1531 skb_rbtree_walk_from(skb) {
1533 bool dup_sack = dup_sack_in;
1535 /* queue is in-order => we can short-circuit the walk early */
1536 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1540 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1541 in_sack = tcp_match_skb_to_sack(sk, skb,
1542 next_dup->start_seq,
1548 /* skb reference here is a bit tricky to get right, since
1549 * shifting can eat and free both this skb and the next,
1550 * so not even _safe variant of the loop is enough.
1553 tmp = tcp_shift_skb_data(sk, skb, state,
1554 start_seq, end_seq, dup_sack);
1563 in_sack = tcp_match_skb_to_sack(sk, skb,
1569 if (unlikely(in_sack < 0))
1573 TCP_SKB_CB(skb)->sacked =
1576 TCP_SKB_CB(skb)->sacked,
1577 TCP_SKB_CB(skb)->seq,
1578 TCP_SKB_CB(skb)->end_seq,
1580 tcp_skb_pcount(skb),
1582 tcp_rate_skb_delivered(sk, skb, state->rate);
1583 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1584 list_del_init(&skb->tcp_tsorted_anchor);
1586 if (!before(TCP_SKB_CB(skb)->seq,
1587 tcp_highest_sack_seq(tp)))
1588 tcp_advance_highest_sack(sk, skb);
1594 static struct sk_buff *tcp_sacktag_bsearch(struct sock *sk,
1595 struct tcp_sacktag_state *state,
1598 struct rb_node *parent, **p = &sk->tcp_rtx_queue.rb_node;
1599 struct sk_buff *skb;
1603 skb = rb_to_skb(parent);
1604 if (before(seq, TCP_SKB_CB(skb)->seq)) {
1605 p = &parent->rb_left;
1608 if (!before(seq, TCP_SKB_CB(skb)->end_seq)) {
1609 p = &parent->rb_right;
1617 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1618 struct tcp_sacktag_state *state,
1621 if (skb && after(TCP_SKB_CB(skb)->seq, skip_to_seq))
1624 return tcp_sacktag_bsearch(sk, state, skip_to_seq);
1627 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1629 struct tcp_sack_block *next_dup,
1630 struct tcp_sacktag_state *state,
1636 if (before(next_dup->start_seq, skip_to_seq)) {
1637 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1638 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1639 next_dup->start_seq, next_dup->end_seq,
1646 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1648 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1652 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1653 u32 prior_snd_una, struct tcp_sacktag_state *state)
1655 struct tcp_sock *tp = tcp_sk(sk);
1656 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1657 TCP_SKB_CB(ack_skb)->sacked);
1658 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1659 struct tcp_sack_block sp[TCP_NUM_SACKS];
1660 struct tcp_sack_block *cache;
1661 struct sk_buff *skb;
1662 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1664 bool found_dup_sack = false;
1666 int first_sack_index;
1669 state->reord = tp->snd_nxt;
1671 if (!tp->sacked_out)
1672 tcp_highest_sack_reset(sk);
1674 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1675 num_sacks, prior_snd_una);
1676 if (found_dup_sack) {
1677 state->flag |= FLAG_DSACKING_ACK;
1678 tp->delivered++; /* A spurious retransmission is delivered */
1681 /* Eliminate too old ACKs, but take into
1682 * account more or less fresh ones, they can
1683 * contain valid SACK info.
1685 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1688 if (!tp->packets_out)
1692 first_sack_index = 0;
1693 for (i = 0; i < num_sacks; i++) {
1694 bool dup_sack = !i && found_dup_sack;
1696 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1697 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1699 if (!tcp_is_sackblock_valid(tp, dup_sack,
1700 sp[used_sacks].start_seq,
1701 sp[used_sacks].end_seq)) {
1705 if (!tp->undo_marker)
1706 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1708 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1710 /* Don't count olds caused by ACK reordering */
1711 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1712 !after(sp[used_sacks].end_seq, tp->snd_una))
1714 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1717 NET_INC_STATS(sock_net(sk), mib_idx);
1719 first_sack_index = -1;
1723 /* Ignore very old stuff early */
1724 if (!after(sp[used_sacks].end_seq, prior_snd_una)) {
1726 first_sack_index = -1;
1733 /* order SACK blocks to allow in order walk of the retrans queue */
1734 for (i = used_sacks - 1; i > 0; i--) {
1735 for (j = 0; j < i; j++) {
1736 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1737 swap(sp[j], sp[j + 1]);
1739 /* Track where the first SACK block goes to */
1740 if (j == first_sack_index)
1741 first_sack_index = j + 1;
1746 state->mss_now = tcp_current_mss(sk);
1750 if (!tp->sacked_out) {
1751 /* It's already past, so skip checking against it */
1752 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1754 cache = tp->recv_sack_cache;
1755 /* Skip empty blocks in at head of the cache */
1756 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1761 while (i < used_sacks) {
1762 u32 start_seq = sp[i].start_seq;
1763 u32 end_seq = sp[i].end_seq;
1764 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1765 struct tcp_sack_block *next_dup = NULL;
1767 if (found_dup_sack && ((i + 1) == first_sack_index))
1768 next_dup = &sp[i + 1];
1770 /* Skip too early cached blocks */
1771 while (tcp_sack_cache_ok(tp, cache) &&
1772 !before(start_seq, cache->end_seq))
1775 /* Can skip some work by looking recv_sack_cache? */
1776 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1777 after(end_seq, cache->start_seq)) {
1780 if (before(start_seq, cache->start_seq)) {
1781 skb = tcp_sacktag_skip(skb, sk, state,
1783 skb = tcp_sacktag_walk(skb, sk, next_dup,
1790 /* Rest of the block already fully processed? */
1791 if (!after(end_seq, cache->end_seq))
1794 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1798 /* ...tail remains todo... */
1799 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1800 /* ...but better entrypoint exists! */
1801 skb = tcp_highest_sack(sk);
1808 skb = tcp_sacktag_skip(skb, sk, state, cache->end_seq);
1809 /* Check overlap against next cached too (past this one already) */
1814 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1815 skb = tcp_highest_sack(sk);
1819 skb = tcp_sacktag_skip(skb, sk, state, start_seq);
1822 skb = tcp_sacktag_walk(skb, sk, next_dup, state,
1823 start_seq, end_seq, dup_sack);
1829 /* Clear the head of the cache sack blocks so we can skip it next time */
1830 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1831 tp->recv_sack_cache[i].start_seq = 0;
1832 tp->recv_sack_cache[i].end_seq = 0;
1834 for (j = 0; j < used_sacks; j++)
1835 tp->recv_sack_cache[i++] = sp[j];
1837 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Loss || tp->undo_marker)
1838 tcp_check_sack_reordering(sk, state->reord, 0);
1840 tcp_verify_left_out(tp);
1843 #if FASTRETRANS_DEBUG > 0
1844 WARN_ON((int)tp->sacked_out < 0);
1845 WARN_ON((int)tp->lost_out < 0);
1846 WARN_ON((int)tp->retrans_out < 0);
1847 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1852 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1853 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1855 static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
1859 holes = max(tp->lost_out, 1U);
1860 holes = min(holes, tp->packets_out);
1862 if ((tp->sacked_out + holes) > tp->packets_out) {
1863 tp->sacked_out = tp->packets_out - holes;
1869 /* If we receive more dupacks than we expected counting segments
1870 * in assumption of absent reordering, interpret this as reordering.
1871 * The only another reason could be bug in receiver TCP.
1873 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1875 struct tcp_sock *tp = tcp_sk(sk);
1877 if (!tcp_limit_reno_sacked(tp))
1880 tp->reordering = min_t(u32, tp->packets_out + addend,
1881 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_max_reordering));
1883 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRENOREORDER);
1886 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1888 static void tcp_add_reno_sack(struct sock *sk)
1890 struct tcp_sock *tp = tcp_sk(sk);
1891 u32 prior_sacked = tp->sacked_out;
1894 tcp_check_reno_reordering(sk, 0);
1895 if (tp->sacked_out > prior_sacked)
1896 tp->delivered++; /* Some out-of-order packet is delivered */
1897 tcp_verify_left_out(tp);
1900 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1902 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1904 struct tcp_sock *tp = tcp_sk(sk);
1907 /* One ACK acked hole. The rest eat duplicate ACKs. */
1908 tp->delivered += max_t(int, acked - tp->sacked_out, 1);
1909 if (acked - 1 >= tp->sacked_out)
1912 tp->sacked_out -= acked - 1;
1914 tcp_check_reno_reordering(sk, acked);
1915 tcp_verify_left_out(tp);
1918 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1923 void tcp_clear_retrans(struct tcp_sock *tp)
1925 tp->retrans_out = 0;
1927 tp->undo_marker = 0;
1928 tp->undo_retrans = -1;
1932 static inline void tcp_init_undo(struct tcp_sock *tp)
1934 tp->undo_marker = tp->snd_una;
1935 /* Retransmission still in flight may cause DSACKs later. */
1936 tp->undo_retrans = tp->retrans_out ? : -1;
1939 static bool tcp_is_rack(const struct sock *sk)
1941 return READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_recovery) &
1942 TCP_RACK_LOSS_DETECTION;
1945 /* If we detect SACK reneging, forget all SACK information
1946 * and reset tags completely, otherwise preserve SACKs. If receiver
1947 * dropped its ofo queue, we will know this due to reneging detection.
1949 static void tcp_timeout_mark_lost(struct sock *sk)
1951 struct tcp_sock *tp = tcp_sk(sk);
1952 struct sk_buff *skb, *head;
1953 bool is_reneg; /* is receiver reneging on SACKs? */
1955 head = tcp_rtx_queue_head(sk);
1956 is_reneg = head && (TCP_SKB_CB(head)->sacked & TCPCB_SACKED_ACKED);
1958 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
1960 /* Mark SACK reneging until we recover from this loss event. */
1961 tp->is_sack_reneg = 1;
1962 } else if (tcp_is_reno(tp)) {
1963 tcp_reset_reno_sack(tp);
1967 skb_rbtree_walk_from(skb) {
1969 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1970 else if (tcp_is_rack(sk) && skb != head &&
1971 tcp_rack_skb_timeout(tp, skb, 0) > 0)
1972 continue; /* Don't mark recently sent ones lost yet */
1973 tcp_mark_skb_lost(sk, skb);
1975 tcp_verify_left_out(tp);
1976 tcp_clear_all_retrans_hints(tp);
1979 /* Enter Loss state. */
1980 void tcp_enter_loss(struct sock *sk)
1982 const struct inet_connection_sock *icsk = inet_csk(sk);
1983 struct tcp_sock *tp = tcp_sk(sk);
1984 struct net *net = sock_net(sk);
1985 bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery;
1988 tcp_timeout_mark_lost(sk);
1990 /* Reduce ssthresh if it has not yet been made inside this window. */
1991 if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
1992 !after(tp->high_seq, tp->snd_una) ||
1993 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1994 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1995 tp->prior_cwnd = tp->snd_cwnd;
1996 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1997 tcp_ca_event(sk, CA_EVENT_LOSS);
2000 tp->snd_cwnd = tcp_packets_in_flight(tp) + 1;
2001 tp->snd_cwnd_cnt = 0;
2002 tp->snd_cwnd_stamp = tcp_jiffies32;
2004 /* Timeout in disordered state after receiving substantial DUPACKs
2005 * suggests that the degree of reordering is over-estimated.
2007 reordering = READ_ONCE(net->ipv4.sysctl_tcp_reordering);
2008 if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
2009 tp->sacked_out >= reordering)
2010 tp->reordering = min_t(unsigned int, tp->reordering,
2013 tcp_set_ca_state(sk, TCP_CA_Loss);
2014 tp->high_seq = tp->snd_nxt;
2015 tcp_ecn_queue_cwr(tp);
2017 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
2018 * loss recovery is underway except recurring timeout(s) on
2019 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
2021 tp->frto = READ_ONCE(net->ipv4.sysctl_tcp_frto) &&
2022 (new_recovery || icsk->icsk_retransmits) &&
2023 !inet_csk(sk)->icsk_mtup.probe_size;
2026 /* If ACK arrived pointing to a remembered SACK, it means that our
2027 * remembered SACKs do not reflect real state of receiver i.e.
2028 * receiver _host_ is heavily congested (or buggy).
2030 * To avoid big spurious retransmission bursts due to transient SACK
2031 * scoreboard oddities that look like reneging, we give the receiver a
2032 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
2033 * restore sanity to the SACK scoreboard. If the apparent reneging
2034 * persists until this RTO then we'll clear the SACK scoreboard.
2036 static bool tcp_check_sack_reneging(struct sock *sk, int flag)
2038 if (flag & FLAG_SACK_RENEGING &&
2039 flag & FLAG_SND_UNA_ADVANCED) {
2040 struct tcp_sock *tp = tcp_sk(sk);
2041 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4),
2042 msecs_to_jiffies(10));
2044 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2045 delay, TCP_RTO_MAX);
2051 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2052 * counter when SACK is enabled (without SACK, sacked_out is used for
2055 * With reordering, holes may still be in flight, so RFC3517 recovery
2056 * uses pure sacked_out (total number of SACKed segments) even though
2057 * it violates the RFC that uses duplicate ACKs, often these are equal
2058 * but when e.g. out-of-window ACKs or packet duplication occurs,
2059 * they differ. Since neither occurs due to loss, TCP should really
2062 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2064 return tp->sacked_out + 1;
2067 /* Linux NewReno/SACK/ECN state machine.
2068 * --------------------------------------
2070 * "Open" Normal state, no dubious events, fast path.
2071 * "Disorder" In all the respects it is "Open",
2072 * but requires a bit more attention. It is entered when
2073 * we see some SACKs or dupacks. It is split of "Open"
2074 * mainly to move some processing from fast path to slow one.
2075 * "CWR" CWND was reduced due to some Congestion Notification event.
2076 * It can be ECN, ICMP source quench, local device congestion.
2077 * "Recovery" CWND was reduced, we are fast-retransmitting.
2078 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2080 * tcp_fastretrans_alert() is entered:
2081 * - each incoming ACK, if state is not "Open"
2082 * - when arrived ACK is unusual, namely:
2087 * Counting packets in flight is pretty simple.
2089 * in_flight = packets_out - left_out + retrans_out
2091 * packets_out is SND.NXT-SND.UNA counted in packets.
2093 * retrans_out is number of retransmitted segments.
2095 * left_out is number of segments left network, but not ACKed yet.
2097 * left_out = sacked_out + lost_out
2099 * sacked_out: Packets, which arrived to receiver out of order
2100 * and hence not ACKed. With SACKs this number is simply
2101 * amount of SACKed data. Even without SACKs
2102 * it is easy to give pretty reliable estimate of this number,
2103 * counting duplicate ACKs.
2105 * lost_out: Packets lost by network. TCP has no explicit
2106 * "loss notification" feedback from network (for now).
2107 * It means that this number can be only _guessed_.
2108 * Actually, it is the heuristics to predict lossage that
2109 * distinguishes different algorithms.
2111 * F.e. after RTO, when all the queue is considered as lost,
2112 * lost_out = packets_out and in_flight = retrans_out.
2114 * Essentially, we have now a few algorithms detecting
2117 * If the receiver supports SACK:
2119 * RFC6675/3517: It is the conventional algorithm. A packet is
2120 * considered lost if the number of higher sequence packets
2121 * SACKed is greater than or equal the DUPACK thoreshold
2122 * (reordering). This is implemented in tcp_mark_head_lost and
2123 * tcp_update_scoreboard.
2125 * RACK (draft-ietf-tcpm-rack-01): it is a newer algorithm
2126 * (2017-) that checks timing instead of counting DUPACKs.
2127 * Essentially a packet is considered lost if it's not S/ACKed
2128 * after RTT + reordering_window, where both metrics are
2129 * dynamically measured and adjusted. This is implemented in
2130 * tcp_rack_mark_lost.
2132 * If the receiver does not support SACK:
2134 * NewReno (RFC6582): in Recovery we assume that one segment
2135 * is lost (classic Reno). While we are in Recovery and
2136 * a partial ACK arrives, we assume that one more packet
2137 * is lost (NewReno). This heuristics are the same in NewReno
2140 * Really tricky (and requiring careful tuning) part of algorithm
2141 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2142 * The first determines the moment _when_ we should reduce CWND and,
2143 * hence, slow down forward transmission. In fact, it determines the moment
2144 * when we decide that hole is caused by loss, rather than by a reorder.
2146 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2147 * holes, caused by lost packets.
2149 * And the most logically complicated part of algorithm is undo
2150 * heuristics. We detect false retransmits due to both too early
2151 * fast retransmit (reordering) and underestimated RTO, analyzing
2152 * timestamps and D-SACKs. When we detect that some segments were
2153 * retransmitted by mistake and CWND reduction was wrong, we undo
2154 * window reduction and abort recovery phase. This logic is hidden
2155 * inside several functions named tcp_try_undo_<something>.
2158 /* This function decides, when we should leave Disordered state
2159 * and enter Recovery phase, reducing congestion window.
2161 * Main question: may we further continue forward transmission
2162 * with the same cwnd?
2164 static bool tcp_time_to_recover(struct sock *sk, int flag)
2166 struct tcp_sock *tp = tcp_sk(sk);
2168 /* Trick#1: The loss is proven. */
2172 /* Not-A-Trick#2 : Classic rule... */
2173 if (!tcp_is_rack(sk) && tcp_dupack_heuristics(tp) > tp->reordering)
2179 /* Detect loss in event "A" above by marking head of queue up as lost.
2180 * For non-SACK(Reno) senders, the first "packets" number of segments
2181 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2182 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2183 * the maximum SACKed segments to pass before reaching this limit.
2185 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2187 struct tcp_sock *tp = tcp_sk(sk);
2188 struct sk_buff *skb;
2189 int cnt, oldcnt, lost;
2191 /* Use SACK to deduce losses of new sequences sent during recovery */
2192 const u32 loss_high = tcp_is_sack(tp) ? tp->snd_nxt : tp->high_seq;
2194 WARN_ON(packets > tp->packets_out);
2195 skb = tp->lost_skb_hint;
2197 /* Head already handled? */
2198 if (mark_head && after(TCP_SKB_CB(skb)->seq, tp->snd_una))
2200 cnt = tp->lost_cnt_hint;
2202 skb = tcp_rtx_queue_head(sk);
2206 skb_rbtree_walk_from(skb) {
2207 /* TODO: do this better */
2208 /* this is not the most efficient way to do this... */
2209 tp->lost_skb_hint = skb;
2210 tp->lost_cnt_hint = cnt;
2212 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2216 if (tcp_is_reno(tp) ||
2217 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2218 cnt += tcp_skb_pcount(skb);
2220 if (cnt > packets) {
2221 if (tcp_is_sack(tp) ||
2222 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
2223 (oldcnt >= packets))
2226 mss = tcp_skb_mss(skb);
2227 /* If needed, chop off the prefix to mark as lost. */
2228 lost = (packets - oldcnt) * mss;
2229 if (lost < skb->len &&
2230 tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
2231 lost, mss, GFP_ATOMIC) < 0)
2236 tcp_skb_mark_lost(tp, skb);
2241 tcp_verify_left_out(tp);
2244 /* Account newly detected lost packet(s) */
2246 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2248 struct tcp_sock *tp = tcp_sk(sk);
2250 if (tcp_is_sack(tp)) {
2251 int sacked_upto = tp->sacked_out - tp->reordering;
2252 if (sacked_upto >= 0)
2253 tcp_mark_head_lost(sk, sacked_upto, 0);
2254 else if (fast_rexmit)
2255 tcp_mark_head_lost(sk, 1, 1);
2259 static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when)
2261 return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2262 before(tp->rx_opt.rcv_tsecr, when);
2265 /* skb is spurious retransmitted if the returned timestamp echo
2266 * reply is prior to the skb transmission time
2268 static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp,
2269 const struct sk_buff *skb)
2271 return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) &&
2272 tcp_tsopt_ecr_before(tp, tcp_skb_timestamp(skb));
2275 /* Nothing was retransmitted or returned timestamp is less
2276 * than timestamp of the first retransmission.
2278 static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2280 return !tp->retrans_stamp ||
2281 tcp_tsopt_ecr_before(tp, tp->retrans_stamp);
2284 /* Undo procedures. */
2286 /* We can clear retrans_stamp when there are no retransmissions in the
2287 * window. It would seem that it is trivially available for us in
2288 * tp->retrans_out, however, that kind of assumptions doesn't consider
2289 * what will happen if errors occur when sending retransmission for the
2290 * second time. ...It could the that such segment has only
2291 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2292 * the head skb is enough except for some reneging corner cases that
2293 * are not worth the effort.
2295 * Main reason for all this complexity is the fact that connection dying
2296 * time now depends on the validity of the retrans_stamp, in particular,
2297 * that successive retransmissions of a segment must not advance
2298 * retrans_stamp under any conditions.
2300 static bool tcp_any_retrans_done(const struct sock *sk)
2302 const struct tcp_sock *tp = tcp_sk(sk);
2303 struct sk_buff *skb;
2305 if (tp->retrans_out)
2308 skb = tcp_rtx_queue_head(sk);
2309 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2315 static void DBGUNDO(struct sock *sk, const char *msg)
2317 #if FASTRETRANS_DEBUG > 1
2318 struct tcp_sock *tp = tcp_sk(sk);
2319 struct inet_sock *inet = inet_sk(sk);
2321 if (sk->sk_family == AF_INET) {
2322 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2324 &inet->inet_daddr, ntohs(inet->inet_dport),
2325 tp->snd_cwnd, tcp_left_out(tp),
2326 tp->snd_ssthresh, tp->prior_ssthresh,
2329 #if IS_ENABLED(CONFIG_IPV6)
2330 else if (sk->sk_family == AF_INET6) {
2331 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2333 &sk->sk_v6_daddr, ntohs(inet->inet_dport),
2334 tp->snd_cwnd, tcp_left_out(tp),
2335 tp->snd_ssthresh, tp->prior_ssthresh,
2342 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
2344 struct tcp_sock *tp = tcp_sk(sk);
2347 struct sk_buff *skb;
2349 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2350 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2353 tcp_clear_all_retrans_hints(tp);
2356 if (tp->prior_ssthresh) {
2357 const struct inet_connection_sock *icsk = inet_csk(sk);
2359 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2361 if (tp->prior_ssthresh > tp->snd_ssthresh) {
2362 tp->snd_ssthresh = tp->prior_ssthresh;
2363 tcp_ecn_withdraw_cwr(tp);
2366 tp->snd_cwnd_stamp = tcp_jiffies32;
2367 tp->undo_marker = 0;
2368 tp->rack.advanced = 1; /* Force RACK to re-exam losses */
2371 static inline bool tcp_may_undo(const struct tcp_sock *tp)
2373 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2376 static bool tcp_is_non_sack_preventing_reopen(struct sock *sk)
2378 struct tcp_sock *tp = tcp_sk(sk);
2380 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2381 /* Hold old state until something *above* high_seq
2382 * is ACKed. For Reno it is MUST to prevent false
2383 * fast retransmits (RFC2582). SACK TCP is safe. */
2384 if (!tcp_any_retrans_done(sk))
2385 tp->retrans_stamp = 0;
2391 /* People celebrate: "We love our President!" */
2392 static bool tcp_try_undo_recovery(struct sock *sk)
2394 struct tcp_sock *tp = tcp_sk(sk);
2396 if (tcp_may_undo(tp)) {
2399 /* Happy end! We did not retransmit anything
2400 * or our original transmission succeeded.
2402 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2403 tcp_undo_cwnd_reduction(sk, false);
2404 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2405 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2407 mib_idx = LINUX_MIB_TCPFULLUNDO;
2409 NET_INC_STATS(sock_net(sk), mib_idx);
2410 } else if (tp->rack.reo_wnd_persist) {
2411 tp->rack.reo_wnd_persist--;
2413 if (tcp_is_non_sack_preventing_reopen(sk))
2415 tcp_set_ca_state(sk, TCP_CA_Open);
2416 tp->is_sack_reneg = 0;
2420 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2421 static bool tcp_try_undo_dsack(struct sock *sk)
2423 struct tcp_sock *tp = tcp_sk(sk);
2425 if (tp->undo_marker && !tp->undo_retrans) {
2426 tp->rack.reo_wnd_persist = min(TCP_RACK_RECOVERY_THRESH,
2427 tp->rack.reo_wnd_persist + 1);
2428 DBGUNDO(sk, "D-SACK");
2429 tcp_undo_cwnd_reduction(sk, false);
2430 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2436 /* Undo during loss recovery after partial ACK or using F-RTO. */
2437 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2439 struct tcp_sock *tp = tcp_sk(sk);
2441 if (frto_undo || tcp_may_undo(tp)) {
2442 tcp_undo_cwnd_reduction(sk, true);
2444 DBGUNDO(sk, "partial loss");
2445 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2447 NET_INC_STATS(sock_net(sk),
2448 LINUX_MIB_TCPSPURIOUSRTOS);
2449 inet_csk(sk)->icsk_retransmits = 0;
2450 if (tcp_is_non_sack_preventing_reopen(sk))
2452 if (frto_undo || tcp_is_sack(tp)) {
2453 tcp_set_ca_state(sk, TCP_CA_Open);
2454 tp->is_sack_reneg = 0;
2461 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
2462 * It computes the number of packets to send (sndcnt) based on packets newly
2464 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2465 * cwnd reductions across a full RTT.
2466 * 2) Otherwise PRR uses packet conservation to send as much as delivered.
2467 * But when the retransmits are acked without further losses, PRR
2468 * slow starts cwnd up to ssthresh to speed up the recovery.
2470 static void tcp_init_cwnd_reduction(struct sock *sk)
2472 struct tcp_sock *tp = tcp_sk(sk);
2474 tp->high_seq = tp->snd_nxt;
2475 tp->tlp_high_seq = 0;
2476 tp->snd_cwnd_cnt = 0;
2477 tp->prior_cwnd = tp->snd_cwnd;
2478 tp->prr_delivered = 0;
2480 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2481 tcp_ecn_queue_cwr(tp);
2484 void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked, int flag)
2486 struct tcp_sock *tp = tcp_sk(sk);
2488 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2490 if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd))
2493 tp->prr_delivered += newly_acked_sacked;
2495 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2497 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2498 } else if ((flag & FLAG_RETRANS_DATA_ACKED) &&
2499 !(flag & FLAG_LOST_RETRANS)) {
2500 sndcnt = min_t(int, delta,
2501 max_t(int, tp->prr_delivered - tp->prr_out,
2502 newly_acked_sacked) + 1);
2504 sndcnt = min(delta, newly_acked_sacked);
2506 /* Force a fast retransmit upon entering fast recovery */
2507 sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1));
2508 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2511 static inline void tcp_end_cwnd_reduction(struct sock *sk)
2513 struct tcp_sock *tp = tcp_sk(sk);
2515 if (inet_csk(sk)->icsk_ca_ops->cong_control)
2518 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2519 if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH &&
2520 (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR || tp->undo_marker)) {
2521 tp->snd_cwnd = tp->snd_ssthresh;
2522 tp->snd_cwnd_stamp = tcp_jiffies32;
2524 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2527 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2528 void tcp_enter_cwr(struct sock *sk)
2530 struct tcp_sock *tp = tcp_sk(sk);
2532 tp->prior_ssthresh = 0;
2533 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2534 tp->undo_marker = 0;
2535 tcp_init_cwnd_reduction(sk);
2536 tcp_set_ca_state(sk, TCP_CA_CWR);
2539 EXPORT_SYMBOL(tcp_enter_cwr);
2541 static void tcp_try_keep_open(struct sock *sk)
2543 struct tcp_sock *tp = tcp_sk(sk);
2544 int state = TCP_CA_Open;
2546 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2547 state = TCP_CA_Disorder;
2549 if (inet_csk(sk)->icsk_ca_state != state) {
2550 tcp_set_ca_state(sk, state);
2551 tp->high_seq = tp->snd_nxt;
2555 static void tcp_try_to_open(struct sock *sk, int flag)
2557 struct tcp_sock *tp = tcp_sk(sk);
2559 tcp_verify_left_out(tp);
2561 if (!tcp_any_retrans_done(sk))
2562 tp->retrans_stamp = 0;
2564 if (flag & FLAG_ECE)
2567 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2568 tcp_try_keep_open(sk);
2572 static void tcp_mtup_probe_failed(struct sock *sk)
2574 struct inet_connection_sock *icsk = inet_csk(sk);
2576 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2577 icsk->icsk_mtup.probe_size = 0;
2578 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPFAIL);
2581 static void tcp_mtup_probe_success(struct sock *sk)
2583 struct tcp_sock *tp = tcp_sk(sk);
2584 struct inet_connection_sock *icsk = inet_csk(sk);
2587 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2589 val = (u64)tp->snd_cwnd * tcp_mss_to_mtu(sk, tp->mss_cache);
2590 do_div(val, icsk->icsk_mtup.probe_size);
2591 WARN_ON_ONCE((u32)val != val);
2592 tp->snd_cwnd = max_t(u32, 1U, val);
2594 tp->snd_cwnd_cnt = 0;
2595 tp->snd_cwnd_stamp = tcp_jiffies32;
2596 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2598 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2599 icsk->icsk_mtup.probe_size = 0;
2600 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2601 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS);
2604 /* Do a simple retransmit without using the backoff mechanisms in
2605 * tcp_timer. This is used for path mtu discovery.
2606 * The socket is already locked here.
2608 void tcp_simple_retransmit(struct sock *sk)
2610 const struct inet_connection_sock *icsk = inet_csk(sk);
2611 struct tcp_sock *tp = tcp_sk(sk);
2612 struct sk_buff *skb;
2613 unsigned int mss = tcp_current_mss(sk);
2615 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2616 if (tcp_skb_seglen(skb) > mss &&
2617 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2618 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2619 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2620 tp->retrans_out -= tcp_skb_pcount(skb);
2622 tcp_skb_mark_lost_uncond_verify(tp, skb);
2626 tcp_clear_retrans_hints_partial(tp);
2631 if (tcp_is_reno(tp))
2632 tcp_limit_reno_sacked(tp);
2634 tcp_verify_left_out(tp);
2636 /* Don't muck with the congestion window here.
2637 * Reason is that we do not increase amount of _data_
2638 * in network, but units changed and effective
2639 * cwnd/ssthresh really reduced now.
2641 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2642 tp->high_seq = tp->snd_nxt;
2643 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2644 tp->prior_ssthresh = 0;
2645 tp->undo_marker = 0;
2646 tcp_set_ca_state(sk, TCP_CA_Loss);
2648 tcp_xmit_retransmit_queue(sk);
2650 EXPORT_SYMBOL(tcp_simple_retransmit);
2652 void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2654 struct tcp_sock *tp = tcp_sk(sk);
2657 if (tcp_is_reno(tp))
2658 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2660 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2662 NET_INC_STATS(sock_net(sk), mib_idx);
2664 tp->prior_ssthresh = 0;
2667 if (!tcp_in_cwnd_reduction(sk)) {
2669 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2670 tcp_init_cwnd_reduction(sk);
2672 tcp_set_ca_state(sk, TCP_CA_Recovery);
2675 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2676 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2678 static void tcp_process_loss(struct sock *sk, int flag, bool is_dupack,
2681 struct tcp_sock *tp = tcp_sk(sk);
2682 bool recovered = !before(tp->snd_una, tp->high_seq);
2684 if ((flag & FLAG_SND_UNA_ADVANCED) &&
2685 tcp_try_undo_loss(sk, false))
2688 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2689 /* Step 3.b. A timeout is spurious if not all data are
2690 * lost, i.e., never-retransmitted data are (s)acked.
2692 if ((flag & FLAG_ORIG_SACK_ACKED) &&
2693 tcp_try_undo_loss(sk, true))
2696 if (after(tp->snd_nxt, tp->high_seq)) {
2697 if (flag & FLAG_DATA_SACKED || is_dupack)
2698 tp->frto = 0; /* Step 3.a. loss was real */
2699 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2700 tp->high_seq = tp->snd_nxt;
2701 /* Step 2.b. Try send new data (but deferred until cwnd
2702 * is updated in tcp_ack()). Otherwise fall back to
2703 * the conventional recovery.
2705 if (!tcp_write_queue_empty(sk) &&
2706 after(tcp_wnd_end(tp), tp->snd_nxt)) {
2707 *rexmit = REXMIT_NEW;
2715 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2716 tcp_try_undo_recovery(sk);
2719 if (tcp_is_reno(tp)) {
2720 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2721 * delivered. Lower inflight to clock out (re)tranmissions.
2723 if (after(tp->snd_nxt, tp->high_seq) && is_dupack)
2724 tcp_add_reno_sack(sk);
2725 else if (flag & FLAG_SND_UNA_ADVANCED)
2726 tcp_reset_reno_sack(tp);
2728 *rexmit = REXMIT_LOST;
2731 /* Undo during fast recovery after partial ACK. */
2732 static bool tcp_try_undo_partial(struct sock *sk, u32 prior_snd_una)
2734 struct tcp_sock *tp = tcp_sk(sk);
2736 if (tp->undo_marker && tcp_packet_delayed(tp)) {
2737 /* Plain luck! Hole if filled with delayed
2738 * packet, rather than with a retransmit. Check reordering.
2740 tcp_check_sack_reordering(sk, prior_snd_una, 1);
2742 /* We are getting evidence that the reordering degree is higher
2743 * than we realized. If there are no retransmits out then we
2744 * can undo. Otherwise we clock out new packets but do not
2745 * mark more packets lost or retransmit more.
2747 if (tp->retrans_out)
2750 if (!tcp_any_retrans_done(sk))
2751 tp->retrans_stamp = 0;
2753 DBGUNDO(sk, "partial recovery");
2754 tcp_undo_cwnd_reduction(sk, true);
2755 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2756 tcp_try_keep_open(sk);
2762 static void tcp_identify_packet_loss(struct sock *sk, int *ack_flag)
2764 struct tcp_sock *tp = tcp_sk(sk);
2766 if (tcp_rtx_queue_empty(sk))
2769 if (unlikely(tcp_is_reno(tp))) {
2770 tcp_newreno_mark_lost(sk, *ack_flag & FLAG_SND_UNA_ADVANCED);
2771 } else if (tcp_is_rack(sk)) {
2772 u32 prior_retrans = tp->retrans_out;
2774 if (tcp_rack_mark_lost(sk))
2775 *ack_flag &= ~FLAG_SET_XMIT_TIMER;
2776 if (prior_retrans > tp->retrans_out)
2777 *ack_flag |= FLAG_LOST_RETRANS;
2781 static bool tcp_force_fast_retransmit(struct sock *sk)
2783 struct tcp_sock *tp = tcp_sk(sk);
2785 return after(tcp_highest_sack_seq(tp),
2786 tp->snd_una + tp->reordering * tp->mss_cache);
2789 /* Process an event, which can update packets-in-flight not trivially.
2790 * Main goal of this function is to calculate new estimate for left_out,
2791 * taking into account both packets sitting in receiver's buffer and
2792 * packets lost by network.
2794 * Besides that it updates the congestion state when packet loss or ECN
2795 * is detected. But it does not reduce the cwnd, it is done by the
2796 * congestion control later.
2798 * It does _not_ decide what to send, it is made in function
2799 * tcp_xmit_retransmit_queue().
2801 static void tcp_fastretrans_alert(struct sock *sk, const u32 prior_snd_una,
2802 bool is_dupack, int *ack_flag, int *rexmit)
2804 struct inet_connection_sock *icsk = inet_csk(sk);
2805 struct tcp_sock *tp = tcp_sk(sk);
2806 int fast_rexmit = 0, flag = *ack_flag;
2807 bool do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2808 tcp_force_fast_retransmit(sk));
2810 if (!tp->packets_out && tp->sacked_out)
2813 /* Now state machine starts.
2814 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2815 if (flag & FLAG_ECE)
2816 tp->prior_ssthresh = 0;
2818 /* B. In all the states check for reneging SACKs. */
2819 if (tcp_check_sack_reneging(sk, flag))
2822 /* C. Check consistency of the current state. */
2823 tcp_verify_left_out(tp);
2825 /* D. Check state exit conditions. State can be terminated
2826 * when high_seq is ACKed. */
2827 if (icsk->icsk_ca_state == TCP_CA_Open) {
2828 WARN_ON(tp->retrans_out != 0);
2829 tp->retrans_stamp = 0;
2830 } else if (!before(tp->snd_una, tp->high_seq)) {
2831 switch (icsk->icsk_ca_state) {
2833 /* CWR is to be held something *above* high_seq
2834 * is ACKed for CWR bit to reach receiver. */
2835 if (tp->snd_una != tp->high_seq) {
2836 tcp_end_cwnd_reduction(sk);
2837 tcp_set_ca_state(sk, TCP_CA_Open);
2841 case TCP_CA_Recovery:
2842 if (tcp_is_reno(tp))
2843 tcp_reset_reno_sack(tp);
2844 if (tcp_try_undo_recovery(sk))
2846 tcp_end_cwnd_reduction(sk);
2851 /* E. Process state. */
2852 switch (icsk->icsk_ca_state) {
2853 case TCP_CA_Recovery:
2854 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2855 if (tcp_is_reno(tp) && is_dupack)
2856 tcp_add_reno_sack(sk);
2858 if (tcp_try_undo_partial(sk, prior_snd_una))
2860 /* Partial ACK arrived. Force fast retransmit. */
2861 do_lost = tcp_is_reno(tp) ||
2862 tcp_force_fast_retransmit(sk);
2864 if (tcp_try_undo_dsack(sk)) {
2865 tcp_try_keep_open(sk);
2868 tcp_identify_packet_loss(sk, ack_flag);
2871 tcp_process_loss(sk, flag, is_dupack, rexmit);
2872 tcp_identify_packet_loss(sk, ack_flag);
2873 if (!(icsk->icsk_ca_state == TCP_CA_Open ||
2874 (*ack_flag & FLAG_LOST_RETRANS)))
2876 /* Change state if cwnd is undone or retransmits are lost */
2879 if (tcp_is_reno(tp)) {
2880 if (flag & FLAG_SND_UNA_ADVANCED)
2881 tcp_reset_reno_sack(tp);
2883 tcp_add_reno_sack(sk);
2886 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
2887 tcp_try_undo_dsack(sk);
2889 tcp_identify_packet_loss(sk, ack_flag);
2890 if (!tcp_time_to_recover(sk, flag)) {
2891 tcp_try_to_open(sk, flag);
2895 /* MTU probe failure: don't reduce cwnd */
2896 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2897 icsk->icsk_mtup.probe_size &&
2898 tp->snd_una == tp->mtu_probe.probe_seq_start) {
2899 tcp_mtup_probe_failed(sk);
2900 /* Restores the reduction we did in tcp_mtup_probe() */
2902 tcp_simple_retransmit(sk);
2906 /* Otherwise enter Recovery state */
2907 tcp_enter_recovery(sk, (flag & FLAG_ECE));
2911 if (!tcp_is_rack(sk) && do_lost)
2912 tcp_update_scoreboard(sk, fast_rexmit);
2913 *rexmit = REXMIT_LOST;
2916 static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us, const int flag)
2918 u32 wlen = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_rtt_wlen) * HZ;
2919 struct tcp_sock *tp = tcp_sk(sk);
2921 if ((flag & FLAG_ACK_MAYBE_DELAYED) && rtt_us > tcp_min_rtt(tp)) {
2922 /* If the remote keeps returning delayed ACKs, eventually
2923 * the min filter would pick it up and overestimate the
2924 * prop. delay when it expires. Skip suspected delayed ACKs.
2928 minmax_running_min(&tp->rtt_min, wlen, tcp_jiffies32,
2929 rtt_us ? : jiffies_to_usecs(1));
2932 static bool tcp_ack_update_rtt(struct sock *sk, const int flag,
2933 long seq_rtt_us, long sack_rtt_us,
2934 long ca_rtt_us, struct rate_sample *rs)
2936 const struct tcp_sock *tp = tcp_sk(sk);
2938 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2939 * broken middle-boxes or peers may corrupt TS-ECR fields. But
2940 * Karn's algorithm forbids taking RTT if some retransmitted data
2941 * is acked (RFC6298).
2944 seq_rtt_us = sack_rtt_us;
2946 /* RTTM Rule: A TSecr value received in a segment is used to
2947 * update the averaged RTT measurement only if the segment
2948 * acknowledges some new data, i.e., only if it advances the
2949 * left edge of the send window.
2950 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2952 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2953 flag & FLAG_ACKED) {
2954 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr;
2956 if (likely(delta < INT_MAX / (USEC_PER_SEC / TCP_TS_HZ))) {
2957 seq_rtt_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
2958 ca_rtt_us = seq_rtt_us;
2961 rs->rtt_us = ca_rtt_us; /* RTT of last (S)ACKed packet (or -1) */
2965 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
2966 * always taken together with ACK, SACK, or TS-opts. Any negative
2967 * values will be skipped with the seq_rtt_us < 0 check above.
2969 tcp_update_rtt_min(sk, ca_rtt_us, flag);
2970 tcp_rtt_estimator(sk, seq_rtt_us);
2973 /* RFC6298: only reset backoff on valid RTT measurement. */
2974 inet_csk(sk)->icsk_backoff = 0;
2978 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
2979 void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req)
2981 struct rate_sample rs;
2984 if (req && !req->num_retrans && tcp_rsk(req)->snt_synack)
2985 rtt_us = tcp_stamp_us_delta(tcp_clock_us(), tcp_rsk(req)->snt_synack);
2987 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us, &rs);
2991 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
2993 const struct inet_connection_sock *icsk = inet_csk(sk);
2995 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
2996 tcp_sk(sk)->snd_cwnd_stamp = tcp_jiffies32;
2999 /* Restart timer after forward progress on connection.
3000 * RFC2988 recommends to restart timer to now+rto.
3002 void tcp_rearm_rto(struct sock *sk)
3004 const struct inet_connection_sock *icsk = inet_csk(sk);
3005 struct tcp_sock *tp = tcp_sk(sk);
3007 /* If the retrans timer is currently being used by Fast Open
3008 * for SYN-ACK retrans purpose, stay put.
3010 if (tp->fastopen_rsk)
3013 if (!tp->packets_out) {
3014 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
3016 u32 rto = inet_csk(sk)->icsk_rto;
3017 /* Offset the time elapsed after installing regular RTO */
3018 if (icsk->icsk_pending == ICSK_TIME_REO_TIMEOUT ||
3019 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
3020 s64 delta_us = tcp_rto_delta_us(sk);
3021 /* delta_us may not be positive if the socket is locked
3022 * when the retrans timer fires and is rescheduled.
3024 rto = usecs_to_jiffies(max_t(int, delta_us, 1));
3026 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
3031 /* Try to schedule a loss probe; if that doesn't work, then schedule an RTO. */
3032 static void tcp_set_xmit_timer(struct sock *sk)
3034 if (!tcp_schedule_loss_probe(sk, true))
3038 /* If we get here, the whole TSO packet has not been acked. */
3039 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3041 struct tcp_sock *tp = tcp_sk(sk);
3044 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3046 packets_acked = tcp_skb_pcount(skb);
3047 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3049 packets_acked -= tcp_skb_pcount(skb);
3051 if (packets_acked) {
3052 BUG_ON(tcp_skb_pcount(skb) == 0);
3053 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3056 return packets_acked;
3059 static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
3062 const struct skb_shared_info *shinfo;
3064 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3065 if (likely(!TCP_SKB_CB(skb)->txstamp_ack))
3068 shinfo = skb_shinfo(skb);
3069 if (!before(shinfo->tskey, prior_snd_una) &&
3070 before(shinfo->tskey, tcp_sk(sk)->snd_una)) {
3071 tcp_skb_tsorted_save(skb) {
3072 __skb_tstamp_tx(skb, NULL, sk, SCM_TSTAMP_ACK);
3073 } tcp_skb_tsorted_restore(skb);
3077 /* Remove acknowledged frames from the retransmission queue. If our packet
3078 * is before the ack sequence we can discard it as it's confirmed to have
3079 * arrived at the other end.
3081 static int tcp_clean_rtx_queue(struct sock *sk, u32 prior_fack,
3083 struct tcp_sacktag_state *sack)
3085 const struct inet_connection_sock *icsk = inet_csk(sk);
3086 u64 first_ackt, last_ackt;
3087 struct tcp_sock *tp = tcp_sk(sk);
3088 u32 prior_sacked = tp->sacked_out;
3089 u32 reord = tp->snd_nxt; /* lowest acked un-retx un-sacked seq */
3090 struct sk_buff *skb, *next;
3091 bool fully_acked = true;
3092 long sack_rtt_us = -1L;
3093 long seq_rtt_us = -1L;
3094 long ca_rtt_us = -1L;
3096 u32 last_in_flight = 0;
3102 for (skb = skb_rb_first(&sk->tcp_rtx_queue); skb; skb = next) {
3103 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3104 const u32 start_seq = scb->seq;
3105 u8 sacked = scb->sacked;
3108 tcp_ack_tstamp(sk, skb, prior_snd_una);
3110 /* Determine how many packets and what bytes were acked, tso and else */
3111 if (after(scb->end_seq, tp->snd_una)) {
3112 if (tcp_skb_pcount(skb) == 1 ||
3113 !after(tp->snd_una, scb->seq))
3116 acked_pcount = tcp_tso_acked(sk, skb);
3119 fully_acked = false;
3121 acked_pcount = tcp_skb_pcount(skb);
3124 if (unlikely(sacked & TCPCB_RETRANS)) {
3125 if (sacked & TCPCB_SACKED_RETRANS)
3126 tp->retrans_out -= acked_pcount;
3127 flag |= FLAG_RETRANS_DATA_ACKED;
3128 } else if (!(sacked & TCPCB_SACKED_ACKED)) {
3129 last_ackt = skb->skb_mstamp;
3130 WARN_ON_ONCE(last_ackt == 0);
3132 first_ackt = last_ackt;
3134 last_in_flight = TCP_SKB_CB(skb)->tx.in_flight;
3135 if (before(start_seq, reord))
3137 if (!after(scb->end_seq, tp->high_seq))
3138 flag |= FLAG_ORIG_SACK_ACKED;
3141 if (sacked & TCPCB_SACKED_ACKED) {
3142 tp->sacked_out -= acked_pcount;
3143 } else if (tcp_is_sack(tp)) {
3144 tp->delivered += acked_pcount;
3145 if (!tcp_skb_spurious_retrans(tp, skb))
3146 tcp_rack_advance(tp, sacked, scb->end_seq,
3149 if (sacked & TCPCB_LOST)
3150 tp->lost_out -= acked_pcount;
3152 tp->packets_out -= acked_pcount;
3153 pkts_acked += acked_pcount;
3154 tcp_rate_skb_delivered(sk, skb, sack->rate);
3156 /* Initial outgoing SYN's get put onto the write_queue
3157 * just like anything else we transmit. It is not
3158 * true data, and if we misinform our callers that
3159 * this ACK acks real data, we will erroneously exit
3160 * connection startup slow start one packet too
3161 * quickly. This is severely frowned upon behavior.
3163 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
3164 flag |= FLAG_DATA_ACKED;
3166 flag |= FLAG_SYN_ACKED;
3167 tp->retrans_stamp = 0;
3173 next = skb_rb_next(skb);
3174 if (unlikely(skb == tp->retransmit_skb_hint))
3175 tp->retransmit_skb_hint = NULL;
3176 if (unlikely(skb == tp->lost_skb_hint))
3177 tp->lost_skb_hint = NULL;
3178 tcp_highest_sack_replace(sk, skb, next);
3179 tcp_rtx_queue_unlink_and_free(skb, sk);
3183 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
3185 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3186 tp->snd_up = tp->snd_una;
3188 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3189 flag |= FLAG_SACK_RENEGING;
3191 if (likely(first_ackt) && !(flag & FLAG_RETRANS_DATA_ACKED)) {
3192 seq_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, first_ackt);
3193 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, last_ackt);
3195 if (pkts_acked == 1 && last_in_flight < tp->mss_cache &&
3196 last_in_flight && !prior_sacked && fully_acked &&
3197 sack->rate->prior_delivered + 1 == tp->delivered &&
3198 !(flag & (FLAG_CA_ALERT | FLAG_SYN_ACKED))) {
3199 /* Conservatively mark a delayed ACK. It's typically
3200 * from a lone runt packet over the round trip to
3201 * a receiver w/o out-of-order or CE events.
3203 flag |= FLAG_ACK_MAYBE_DELAYED;
3206 if (sack->first_sackt) {
3207 sack_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->first_sackt);
3208 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->last_sackt);
3210 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us,
3211 ca_rtt_us, sack->rate);
3213 if (flag & FLAG_ACKED) {
3214 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3215 if (unlikely(icsk->icsk_mtup.probe_size &&
3216 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3217 tcp_mtup_probe_success(sk);
3220 if (tcp_is_reno(tp)) {
3221 tcp_remove_reno_sacks(sk, pkts_acked);
3223 /* If any of the cumulatively ACKed segments was
3224 * retransmitted, non-SACK case cannot confirm that
3225 * progress was due to original transmission due to
3226 * lack of TCPCB_SACKED_ACKED bits even if some of
3227 * the packets may have been never retransmitted.
3229 if (flag & FLAG_RETRANS_DATA_ACKED)
3230 flag &= ~FLAG_ORIG_SACK_ACKED;
3234 /* Non-retransmitted hole got filled? That's reordering */
3235 if (before(reord, prior_fack))
3236 tcp_check_sack_reordering(sk, reord, 0);
3238 delta = prior_sacked - tp->sacked_out;
3239 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3241 } else if (skb && rtt_update && sack_rtt_us >= 0 &&
3242 sack_rtt_us > tcp_stamp_us_delta(tp->tcp_mstamp, skb->skb_mstamp)) {
3243 /* Do not re-arm RTO if the sack RTT is measured from data sent
3244 * after when the head was last (re)transmitted. Otherwise the
3245 * timeout may continue to extend in loss recovery.
3247 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3250 if (icsk->icsk_ca_ops->pkts_acked) {
3251 struct ack_sample sample = { .pkts_acked = pkts_acked,
3252 .rtt_us = sack->rate->rtt_us,
3253 .in_flight = last_in_flight };
3255 icsk->icsk_ca_ops->pkts_acked(sk, &sample);
3258 #if FASTRETRANS_DEBUG > 0
3259 WARN_ON((int)tp->sacked_out < 0);
3260 WARN_ON((int)tp->lost_out < 0);
3261 WARN_ON((int)tp->retrans_out < 0);
3262 if (!tp->packets_out && tcp_is_sack(tp)) {
3263 icsk = inet_csk(sk);
3265 pr_debug("Leak l=%u %d\n",
3266 tp->lost_out, icsk->icsk_ca_state);
3269 if (tp->sacked_out) {
3270 pr_debug("Leak s=%u %d\n",
3271 tp->sacked_out, icsk->icsk_ca_state);
3274 if (tp->retrans_out) {
3275 pr_debug("Leak r=%u %d\n",
3276 tp->retrans_out, icsk->icsk_ca_state);
3277 tp->retrans_out = 0;
3284 static void tcp_ack_probe(struct sock *sk)
3286 struct inet_connection_sock *icsk = inet_csk(sk);
3287 struct sk_buff *head = tcp_send_head(sk);
3288 const struct tcp_sock *tp = tcp_sk(sk);
3290 /* Was it a usable window open? */
3293 if (!after(TCP_SKB_CB(head)->end_seq, tcp_wnd_end(tp))) {
3294 icsk->icsk_backoff = 0;
3295 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3296 /* Socket must be waked up by subsequent tcp_data_snd_check().
3297 * This function is not for random using!
3300 unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX);
3302 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3307 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3309 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3310 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3313 /* Decide wheather to run the increase function of congestion control. */
3314 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3316 /* If reordering is high then always grow cwnd whenever data is
3317 * delivered regardless of its ordering. Otherwise stay conservative
3318 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3319 * new SACK or ECE mark may first advance cwnd here and later reduce
3320 * cwnd in tcp_fastretrans_alert() based on more states.
3322 if (tcp_sk(sk)->reordering >
3323 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_reordering))
3324 return flag & FLAG_FORWARD_PROGRESS;
3326 return flag & FLAG_DATA_ACKED;
3329 /* The "ultimate" congestion control function that aims to replace the rigid
3330 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction).
3331 * It's called toward the end of processing an ACK with precise rate
3332 * information. All transmission or retransmission are delayed afterwards.
3334 static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked,
3335 int flag, const struct rate_sample *rs)
3337 const struct inet_connection_sock *icsk = inet_csk(sk);
3339 if (icsk->icsk_ca_ops->cong_control) {
3340 icsk->icsk_ca_ops->cong_control(sk, rs);
3344 if (tcp_in_cwnd_reduction(sk)) {
3345 /* Reduce cwnd if state mandates */
3346 tcp_cwnd_reduction(sk, acked_sacked, flag);
3347 } else if (tcp_may_raise_cwnd(sk, flag)) {
3348 /* Advance cwnd if state allows */
3349 tcp_cong_avoid(sk, ack, acked_sacked);
3351 tcp_update_pacing_rate(sk);
3354 /* Check that window update is acceptable.
3355 * The function assumes that snd_una<=ack<=snd_next.
3357 static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3358 const u32 ack, const u32 ack_seq,
3361 return after(ack, tp->snd_una) ||
3362 after(ack_seq, tp->snd_wl1) ||
3363 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3366 /* If we update tp->snd_una, also update tp->bytes_acked */
3367 static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack)
3369 u32 delta = ack - tp->snd_una;
3371 sock_owned_by_me((struct sock *)tp);
3372 tp->bytes_acked += delta;
3376 /* If we update tp->rcv_nxt, also update tp->bytes_received */
3377 static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq)
3379 u32 delta = seq - tp->rcv_nxt;
3381 sock_owned_by_me((struct sock *)tp);
3382 tp->bytes_received += delta;
3383 WRITE_ONCE(tp->rcv_nxt, seq);
3386 /* Update our send window.
3388 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3389 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3391 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3394 struct tcp_sock *tp = tcp_sk(sk);
3396 u32 nwin = ntohs(tcp_hdr(skb)->window);
3398 if (likely(!tcp_hdr(skb)->syn))
3399 nwin <<= tp->rx_opt.snd_wscale;
3401 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3402 flag |= FLAG_WIN_UPDATE;
3403 tcp_update_wl(tp, ack_seq);
3405 if (tp->snd_wnd != nwin) {
3408 /* Note, it is the only place, where
3409 * fast path is recovered for sending TCP.
3412 tcp_fast_path_check(sk);
3414 if (!tcp_write_queue_empty(sk))
3415 tcp_slow_start_after_idle_check(sk);
3417 if (nwin > tp->max_window) {
3418 tp->max_window = nwin;
3419 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3424 tcp_snd_una_update(tp, ack);
3429 static bool __tcp_oow_rate_limited(struct net *net, int mib_idx,
3430 u32 *last_oow_ack_time)
3432 if (*last_oow_ack_time) {
3433 s32 elapsed = (s32)(tcp_jiffies32 - *last_oow_ack_time);
3436 elapsed < READ_ONCE(net->ipv4.sysctl_tcp_invalid_ratelimit)) {
3437 NET_INC_STATS(net, mib_idx);
3438 return true; /* rate-limited: don't send yet! */
3442 *last_oow_ack_time = tcp_jiffies32;
3444 return false; /* not rate-limited: go ahead, send dupack now! */
3447 /* Return true if we're currently rate-limiting out-of-window ACKs and
3448 * thus shouldn't send a dupack right now. We rate-limit dupacks in
3449 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3450 * attacks that send repeated SYNs or ACKs for the same connection. To
3451 * do this, we do not send a duplicate SYNACK or ACK if the remote
3452 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3454 bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb,
3455 int mib_idx, u32 *last_oow_ack_time)
3457 /* Data packets without SYNs are not likely part of an ACK loop. */
3458 if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) &&
3462 return __tcp_oow_rate_limited(net, mib_idx, last_oow_ack_time);
3465 /* RFC 5961 7 [ACK Throttling] */
3466 static void tcp_send_challenge_ack(struct sock *sk, const struct sk_buff *skb)
3468 /* unprotected vars, we dont care of overwrites */
3469 static u32 challenge_timestamp;
3470 static unsigned int challenge_count;
3471 struct tcp_sock *tp = tcp_sk(sk);
3472 struct net *net = sock_net(sk);
3475 /* First check our per-socket dupack rate limit. */
3476 if (__tcp_oow_rate_limited(net,
3477 LINUX_MIB_TCPACKSKIPPEDCHALLENGE,
3478 &tp->last_oow_ack_time))
3481 /* Then check host-wide RFC 5961 rate limit. */
3483 if (now != READ_ONCE(challenge_timestamp)) {
3484 u32 ack_limit = READ_ONCE(net->ipv4.sysctl_tcp_challenge_ack_limit);
3485 u32 half = (ack_limit + 1) >> 1;
3487 WRITE_ONCE(challenge_timestamp, now);
3488 WRITE_ONCE(challenge_count, half + prandom_u32_max(ack_limit));
3490 count = READ_ONCE(challenge_count);
3492 WRITE_ONCE(challenge_count, count - 1);
3493 NET_INC_STATS(net, LINUX_MIB_TCPCHALLENGEACK);
3498 static void tcp_store_ts_recent(struct tcp_sock *tp)
3500 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3501 tp->rx_opt.ts_recent_stamp = ktime_get_seconds();
3504 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3506 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3507 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3508 * extra check below makes sure this can only happen
3509 * for pure ACK frames. -DaveM
3511 * Not only, also it occurs for expired timestamps.
3514 if (tcp_paws_check(&tp->rx_opt, 0))
3515 tcp_store_ts_recent(tp);
3519 /* This routine deals with acks during a TLP episode and ends an episode by
3520 * resetting tlp_high_seq. Ref: TLP algorithm in draft-ietf-tcpm-rack
3522 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3524 struct tcp_sock *tp = tcp_sk(sk);
3526 if (before(ack, tp->tlp_high_seq))
3529 if (!tp->tlp_retrans) {
3530 /* TLP of new data has been acknowledged */
3531 tp->tlp_high_seq = 0;
3532 } else if (flag & FLAG_DSACKING_ACK) {
3533 /* This DSACK means original and TLP probe arrived; no loss */
3534 tp->tlp_high_seq = 0;
3535 } else if (after(ack, tp->tlp_high_seq)) {
3536 /* ACK advances: there was a loss, so reduce cwnd. Reset
3537 * tlp_high_seq in tcp_init_cwnd_reduction()
3539 tcp_init_cwnd_reduction(sk);
3540 tcp_set_ca_state(sk, TCP_CA_CWR);
3541 tcp_end_cwnd_reduction(sk);
3542 tcp_try_keep_open(sk);
3543 NET_INC_STATS(sock_net(sk),
3544 LINUX_MIB_TCPLOSSPROBERECOVERY);
3545 } else if (!(flag & (FLAG_SND_UNA_ADVANCED |
3546 FLAG_NOT_DUP | FLAG_DATA_SACKED))) {
3547 /* Pure dupack: original and TLP probe arrived; no loss */
3548 tp->tlp_high_seq = 0;
3552 static inline void tcp_in_ack_event(struct sock *sk, u32 flags)
3554 const struct inet_connection_sock *icsk = inet_csk(sk);
3556 if (icsk->icsk_ca_ops->in_ack_event)
3557 icsk->icsk_ca_ops->in_ack_event(sk, flags);
3560 /* Congestion control has updated the cwnd already. So if we're in
3561 * loss recovery then now we do any new sends (for FRTO) or
3562 * retransmits (for CA_Loss or CA_recovery) that make sense.
3564 static void tcp_xmit_recovery(struct sock *sk, int rexmit)
3566 struct tcp_sock *tp = tcp_sk(sk);
3568 if (rexmit == REXMIT_NONE)
3571 if (unlikely(rexmit == 2)) {
3572 __tcp_push_pending_frames(sk, tcp_current_mss(sk),
3574 if (after(tp->snd_nxt, tp->high_seq))
3578 tcp_xmit_retransmit_queue(sk);
3581 /* Returns the number of packets newly acked or sacked by the current ACK */
3582 static u32 tcp_newly_delivered(struct sock *sk, u32 prior_delivered, int flag)
3584 const struct net *net = sock_net(sk);
3585 struct tcp_sock *tp = tcp_sk(sk);
3588 delivered = tp->delivered - prior_delivered;
3589 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVERED, delivered);
3590 if (flag & FLAG_ECE) {
3591 tp->delivered_ce += delivered;
3592 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVEREDCE, delivered);
3597 /* This routine deals with incoming acks, but not outgoing ones. */
3598 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3600 struct inet_connection_sock *icsk = inet_csk(sk);
3601 struct tcp_sock *tp = tcp_sk(sk);
3602 struct tcp_sacktag_state sack_state;
3603 struct rate_sample rs = { .prior_delivered = 0 };
3604 u32 prior_snd_una = tp->snd_una;
3605 bool is_sack_reneg = tp->is_sack_reneg;
3606 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3607 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3608 bool is_dupack = false;
3609 int prior_packets = tp->packets_out;
3610 u32 delivered = tp->delivered;
3611 u32 lost = tp->lost;
3612 int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */
3615 sack_state.first_sackt = 0;
3616 sack_state.rate = &rs;
3618 /* We very likely will need to access rtx queue. */
3619 prefetch(sk->tcp_rtx_queue.rb_node);
3621 /* If the ack is older than previous acks
3622 * then we can probably ignore it.
3624 if (before(ack, prior_snd_una)) {
3625 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3626 if (before(ack, prior_snd_una - tp->max_window)) {
3627 if (!(flag & FLAG_NO_CHALLENGE_ACK))
3628 tcp_send_challenge_ack(sk, skb);
3634 /* If the ack includes data we haven't sent yet, discard
3635 * this segment (RFC793 Section 3.9).
3637 if (after(ack, tp->snd_nxt))
3640 if (after(ack, prior_snd_una)) {
3641 flag |= FLAG_SND_UNA_ADVANCED;
3642 icsk->icsk_retransmits = 0;
3644 #if IS_ENABLED(CONFIG_TLS_DEVICE)
3645 if (static_branch_unlikely(&clean_acked_data_enabled))
3646 if (icsk->icsk_clean_acked)
3647 icsk->icsk_clean_acked(sk, ack);
3651 prior_fack = tcp_is_sack(tp) ? tcp_highest_sack_seq(tp) : tp->snd_una;
3652 rs.prior_in_flight = tcp_packets_in_flight(tp);
3654 /* ts_recent update must be made after we are sure that the packet
3657 if (flag & FLAG_UPDATE_TS_RECENT)
3658 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3660 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3661 /* Window is constant, pure forward advance.
3662 * No more checks are required.
3663 * Note, we use the fact that SND.UNA>=SND.WL2.
3665 tcp_update_wl(tp, ack_seq);
3666 tcp_snd_una_update(tp, ack);
3667 flag |= FLAG_WIN_UPDATE;
3669 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
3671 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPACKS);
3673 u32 ack_ev_flags = CA_ACK_SLOWPATH;
3675 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3678 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3680 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3682 if (TCP_SKB_CB(skb)->sacked)
3683 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3686 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) {
3688 ack_ev_flags |= CA_ACK_ECE;
3691 if (flag & FLAG_WIN_UPDATE)
3692 ack_ev_flags |= CA_ACK_WIN_UPDATE;
3694 tcp_in_ack_event(sk, ack_ev_flags);
3697 /* This is a deviation from RFC3168 since it states that:
3698 * "When the TCP data sender is ready to set the CWR bit after reducing
3699 * the congestion window, it SHOULD set the CWR bit only on the first
3700 * new data packet that it transmits."
3701 * We accept CWR on pure ACKs to be more robust
3702 * with widely-deployed TCP implementations that do this.
3704 tcp_ecn_accept_cwr(sk, skb);
3706 /* We passed data and got it acked, remove any soft error
3707 * log. Something worked...
3709 sk->sk_err_soft = 0;
3710 icsk->icsk_probes_out = 0;
3711 tp->rcv_tstamp = tcp_jiffies32;
3715 /* See if we can take anything off of the retransmit queue. */
3716 flag |= tcp_clean_rtx_queue(sk, prior_fack, prior_snd_una, &sack_state);
3718 tcp_rack_update_reo_wnd(sk, &rs);
3720 if (tp->tlp_high_seq)
3721 tcp_process_tlp_ack(sk, ack, flag);
3723 if (tcp_ack_is_dubious(sk, flag)) {
3724 is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
3725 tcp_fastretrans_alert(sk, prior_snd_una, is_dupack, &flag,
3729 /* If needed, reset TLP/RTO timer when RACK doesn't set. */
3730 if (flag & FLAG_SET_XMIT_TIMER)
3731 tcp_set_xmit_timer(sk);
3733 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
3736 delivered = tcp_newly_delivered(sk, delivered, flag);
3737 lost = tp->lost - lost; /* freshly marked lost */
3738 rs.is_ack_delayed = !!(flag & FLAG_ACK_MAYBE_DELAYED);
3739 tcp_rate_gen(sk, delivered, lost, is_sack_reneg, sack_state.rate);
3740 tcp_cong_control(sk, ack, delivered, flag, sack_state.rate);
3741 tcp_xmit_recovery(sk, rexmit);
3745 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3746 if (flag & FLAG_DSACKING_ACK) {
3747 tcp_fastretrans_alert(sk, prior_snd_una, is_dupack, &flag,
3749 tcp_newly_delivered(sk, delivered, flag);
3751 /* If this ack opens up a zero window, clear backoff. It was
3752 * being used to time the probes, and is probably far higher than
3753 * it needs to be for normal retransmission.
3757 if (tp->tlp_high_seq)
3758 tcp_process_tlp_ack(sk, ack, flag);
3762 SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3766 /* If data was SACKed, tag it and see if we should send more data.
3767 * If data was DSACKed, see if we can undo a cwnd reduction.
3769 if (TCP_SKB_CB(skb)->sacked) {
3770 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3772 tcp_fastretrans_alert(sk, prior_snd_una, is_dupack, &flag,
3774 tcp_newly_delivered(sk, delivered, flag);
3775 tcp_xmit_recovery(sk, rexmit);
3778 SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3782 static void tcp_parse_fastopen_option(int len, const unsigned char *cookie,
3783 bool syn, struct tcp_fastopen_cookie *foc,
3786 /* Valid only in SYN or SYN-ACK with an even length. */
3787 if (!foc || !syn || len < 0 || (len & 1))
3790 if (len >= TCP_FASTOPEN_COOKIE_MIN &&
3791 len <= TCP_FASTOPEN_COOKIE_MAX)
3792 memcpy(foc->val, cookie, len);
3799 static void smc_parse_options(const struct tcphdr *th,
3800 struct tcp_options_received *opt_rx,
3801 const unsigned char *ptr,
3804 #if IS_ENABLED(CONFIG_SMC)
3805 if (static_branch_unlikely(&tcp_have_smc)) {
3806 if (th->syn && !(opsize & 1) &&
3807 opsize >= TCPOLEN_EXP_SMC_BASE &&
3808 get_unaligned_be32(ptr) == TCPOPT_SMC_MAGIC)
3814 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3815 * But, this can also be called on packets in the established flow when
3816 * the fast version below fails.
3818 void tcp_parse_options(const struct net *net,
3819 const struct sk_buff *skb,
3820 struct tcp_options_received *opt_rx, int estab,
3821 struct tcp_fastopen_cookie *foc)
3823 const unsigned char *ptr;
3824 const struct tcphdr *th = tcp_hdr(skb);
3825 int length = (th->doff * 4) - sizeof(struct tcphdr);
3827 ptr = (const unsigned char *)(th + 1);
3828 opt_rx->saw_tstamp = 0;
3830 while (length > 0) {
3831 int opcode = *ptr++;
3837 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3842 if (opsize < 2) /* "silly options" */
3844 if (opsize > length)
3845 return; /* don't parse partial options */
3848 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3849 u16 in_mss = get_unaligned_be16(ptr);
3851 if (opt_rx->user_mss &&
3852 opt_rx->user_mss < in_mss)
3853 in_mss = opt_rx->user_mss;
3854 opt_rx->mss_clamp = in_mss;
3859 if (opsize == TCPOLEN_WINDOW && th->syn &&
3860 !estab && net->ipv4.sysctl_tcp_window_scaling) {
3861 __u8 snd_wscale = *(__u8 *)ptr;
3862 opt_rx->wscale_ok = 1;
3863 if (snd_wscale > TCP_MAX_WSCALE) {
3864 net_info_ratelimited("%s: Illegal window scaling value %d > %u received\n",
3868 snd_wscale = TCP_MAX_WSCALE;
3870 opt_rx->snd_wscale = snd_wscale;
3873 case TCPOPT_TIMESTAMP:
3874 if ((opsize == TCPOLEN_TIMESTAMP) &&
3875 ((estab && opt_rx->tstamp_ok) ||
3876 (!estab && net->ipv4.sysctl_tcp_timestamps))) {
3877 opt_rx->saw_tstamp = 1;
3878 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3879 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3882 case TCPOPT_SACK_PERM:
3883 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3884 !estab && net->ipv4.sysctl_tcp_sack) {
3885 opt_rx->sack_ok = TCP_SACK_SEEN;
3886 tcp_sack_reset(opt_rx);
3891 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3892 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3894 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3897 #ifdef CONFIG_TCP_MD5SIG
3900 * The MD5 Hash has already been
3901 * checked (see tcp_v{4,6}_do_rcv()).
3905 case TCPOPT_FASTOPEN:
3906 tcp_parse_fastopen_option(
3907 opsize - TCPOLEN_FASTOPEN_BASE,
3908 ptr, th->syn, foc, false);
3912 /* Fast Open option shares code 254 using a
3913 * 16 bits magic number.
3915 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE &&
3916 get_unaligned_be16(ptr) ==
3917 TCPOPT_FASTOPEN_MAGIC)
3918 tcp_parse_fastopen_option(opsize -
3919 TCPOLEN_EXP_FASTOPEN_BASE,
3920 ptr + 2, th->syn, foc, true);
3922 smc_parse_options(th, opt_rx, ptr,
3932 EXPORT_SYMBOL(tcp_parse_options);
3934 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
3936 const __be32 *ptr = (const __be32 *)(th + 1);
3938 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3939 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3940 tp->rx_opt.saw_tstamp = 1;
3942 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3945 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
3947 tp->rx_opt.rcv_tsecr = 0;
3953 /* Fast parse options. This hopes to only see timestamps.
3954 * If it is wrong it falls back on tcp_parse_options().
3956 static bool tcp_fast_parse_options(const struct net *net,
3957 const struct sk_buff *skb,
3958 const struct tcphdr *th, struct tcp_sock *tp)
3960 /* In the spirit of fast parsing, compare doff directly to constant
3961 * values. Because equality is used, short doff can be ignored here.
3963 if (th->doff == (sizeof(*th) / 4)) {
3964 tp->rx_opt.saw_tstamp = 0;
3966 } else if (tp->rx_opt.tstamp_ok &&
3967 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3968 if (tcp_parse_aligned_timestamp(tp, th))
3972 tcp_parse_options(net, skb, &tp->rx_opt, 1, NULL);
3973 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3974 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
3979 #ifdef CONFIG_TCP_MD5SIG
3981 * Parse MD5 Signature option
3983 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
3985 int length = (th->doff << 2) - sizeof(*th);
3986 const u8 *ptr = (const u8 *)(th + 1);
3988 /* If not enough data remaining, we can short cut */
3989 while (length >= TCPOLEN_MD5SIG) {
3990 int opcode = *ptr++;
4001 if (opsize < 2 || opsize > length)
4003 if (opcode == TCPOPT_MD5SIG)
4004 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
4011 EXPORT_SYMBOL(tcp_parse_md5sig_option);
4014 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
4016 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
4017 * it can pass through stack. So, the following predicate verifies that
4018 * this segment is not used for anything but congestion avoidance or
4019 * fast retransmit. Moreover, we even are able to eliminate most of such
4020 * second order effects, if we apply some small "replay" window (~RTO)
4021 * to timestamp space.
4023 * All these measures still do not guarantee that we reject wrapped ACKs
4024 * on networks with high bandwidth, when sequence space is recycled fastly,
4025 * but it guarantees that such events will be very rare and do not affect
4026 * connection seriously. This doesn't look nice, but alas, PAWS is really
4029 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
4030 * states that events when retransmit arrives after original data are rare.
4031 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
4032 * the biggest problem on large power networks even with minor reordering.
4033 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
4034 * up to bandwidth of 18Gigabit/sec. 8) ]
4037 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
4039 const struct tcp_sock *tp = tcp_sk(sk);
4040 const struct tcphdr *th = tcp_hdr(skb);
4041 u32 seq = TCP_SKB_CB(skb)->seq;
4042 u32 ack = TCP_SKB_CB(skb)->ack_seq;
4044 return (/* 1. Pure ACK with correct sequence number. */
4045 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
4047 /* 2. ... and duplicate ACK. */
4048 ack == tp->snd_una &&
4050 /* 3. ... and does not update window. */
4051 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
4053 /* 4. ... and sits in replay window. */
4054 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
4057 static inline bool tcp_paws_discard(const struct sock *sk,
4058 const struct sk_buff *skb)
4060 const struct tcp_sock *tp = tcp_sk(sk);
4062 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
4063 !tcp_disordered_ack(sk, skb);
4066 /* Check segment sequence number for validity.
4068 * Segment controls are considered valid, if the segment
4069 * fits to the window after truncation to the window. Acceptability
4070 * of data (and SYN, FIN, of course) is checked separately.
4071 * See tcp_data_queue(), for example.
4073 * Also, controls (RST is main one) are accepted using RCV.WUP instead
4074 * of RCV.NXT. Peer still did not advance his SND.UNA when we
4075 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4076 * (borrowed from freebsd)
4079 static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
4081 return !before(end_seq, tp->rcv_wup) &&
4082 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
4085 /* When we get a reset we do this. */
4086 void tcp_reset(struct sock *sk)
4088 trace_tcp_receive_reset(sk);
4090 /* We want the right error as BSD sees it (and indeed as we do). */
4091 switch (sk->sk_state) {
4093 sk->sk_err = ECONNREFUSED;
4095 case TCP_CLOSE_WAIT:
4101 sk->sk_err = ECONNRESET;
4103 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4106 tcp_write_queue_purge(sk);
4109 if (!sock_flag(sk, SOCK_DEAD))
4110 sk->sk_error_report(sk);
4114 * Process the FIN bit. This now behaves as it is supposed to work
4115 * and the FIN takes effect when it is validly part of sequence
4116 * space. Not before when we get holes.
4118 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4119 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4122 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4123 * close and we go into CLOSING (and later onto TIME-WAIT)
4125 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4127 void tcp_fin(struct sock *sk)
4129 struct tcp_sock *tp = tcp_sk(sk);
4131 inet_csk_schedule_ack(sk);
4133 WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | RCV_SHUTDOWN);
4134 sock_set_flag(sk, SOCK_DONE);
4136 switch (sk->sk_state) {
4138 case TCP_ESTABLISHED:
4139 /* Move to CLOSE_WAIT */
4140 tcp_set_state(sk, TCP_CLOSE_WAIT);
4141 inet_csk(sk)->icsk_ack.pingpong = 1;
4144 case TCP_CLOSE_WAIT:
4146 /* Received a retransmission of the FIN, do
4151 /* RFC793: Remain in the LAST-ACK state. */
4155 /* This case occurs when a simultaneous close
4156 * happens, we must ack the received FIN and
4157 * enter the CLOSING state.
4160 tcp_set_state(sk, TCP_CLOSING);
4163 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4165 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4168 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4169 * cases we should never reach this piece of code.
4171 pr_err("%s: Impossible, sk->sk_state=%d\n",
4172 __func__, sk->sk_state);
4176 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4177 * Probably, we should reset in this case. For now drop them.
4179 skb_rbtree_purge(&tp->out_of_order_queue);
4180 if (tcp_is_sack(tp))
4181 tcp_sack_reset(&tp->rx_opt);
4184 if (!sock_flag(sk, SOCK_DEAD)) {
4185 sk->sk_state_change(sk);
4187 /* Do not send POLL_HUP for half duplex close. */
4188 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4189 sk->sk_state == TCP_CLOSE)
4190 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4192 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4196 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4199 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4200 if (before(seq, sp->start_seq))
4201 sp->start_seq = seq;
4202 if (after(end_seq, sp->end_seq))
4203 sp->end_seq = end_seq;
4209 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4211 struct tcp_sock *tp = tcp_sk(sk);
4213 if (tcp_is_sack(tp) && READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_dsack)) {
4216 if (before(seq, tp->rcv_nxt))
4217 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4219 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4221 NET_INC_STATS(sock_net(sk), mib_idx);
4223 tp->rx_opt.dsack = 1;
4224 tp->duplicate_sack[0].start_seq = seq;
4225 tp->duplicate_sack[0].end_seq = end_seq;
4229 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4231 struct tcp_sock *tp = tcp_sk(sk);
4233 if (!tp->rx_opt.dsack)
4234 tcp_dsack_set(sk, seq, end_seq);
4236 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4239 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4241 struct tcp_sock *tp = tcp_sk(sk);
4243 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4244 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4245 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4246 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
4248 if (tcp_is_sack(tp) && READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_dsack)) {
4249 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4251 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4252 end_seq = tp->rcv_nxt;
4253 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4260 /* These routines update the SACK block as out-of-order packets arrive or
4261 * in-order packets close up the sequence space.
4263 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4266 struct tcp_sack_block *sp = &tp->selective_acks[0];
4267 struct tcp_sack_block *swalk = sp + 1;
4269 /* See if the recent change to the first SACK eats into
4270 * or hits the sequence space of other SACK blocks, if so coalesce.
4272 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4273 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4276 /* Zap SWALK, by moving every further SACK up by one slot.
4277 * Decrease num_sacks.
4279 tp->rx_opt.num_sacks--;
4280 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4284 this_sack++, swalk++;
4288 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4290 struct tcp_sock *tp = tcp_sk(sk);
4291 struct tcp_sack_block *sp = &tp->selective_acks[0];
4292 int cur_sacks = tp->rx_opt.num_sacks;
4298 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4299 if (tcp_sack_extend(sp, seq, end_seq)) {
4300 /* Rotate this_sack to the first one. */
4301 for (; this_sack > 0; this_sack--, sp--)
4302 swap(*sp, *(sp - 1));
4304 tcp_sack_maybe_coalesce(tp);
4309 /* Could not find an adjacent existing SACK, build a new one,
4310 * put it at the front, and shift everyone else down. We
4311 * always know there is at least one SACK present already here.
4313 * If the sack array is full, forget about the last one.
4315 if (this_sack >= TCP_NUM_SACKS) {
4316 if (tp->compressed_ack > TCP_FASTRETRANS_THRESH)
4319 tp->rx_opt.num_sacks--;
4322 for (; this_sack > 0; this_sack--, sp--)
4326 /* Build the new head SACK, and we're done. */
4327 sp->start_seq = seq;
4328 sp->end_seq = end_seq;
4329 tp->rx_opt.num_sacks++;
4332 /* RCV.NXT advances, some SACKs should be eaten. */
4334 static void tcp_sack_remove(struct tcp_sock *tp)
4336 struct tcp_sack_block *sp = &tp->selective_acks[0];
4337 int num_sacks = tp->rx_opt.num_sacks;
4340 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4341 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4342 tp->rx_opt.num_sacks = 0;
4346 for (this_sack = 0; this_sack < num_sacks;) {
4347 /* Check if the start of the sack is covered by RCV.NXT. */
4348 if (!before(tp->rcv_nxt, sp->start_seq)) {
4351 /* RCV.NXT must cover all the block! */
4352 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4354 /* Zap this SACK, by moving forward any other SACKS. */
4355 for (i = this_sack+1; i < num_sacks; i++)
4356 tp->selective_acks[i-1] = tp->selective_acks[i];
4363 tp->rx_opt.num_sacks = num_sacks;
4367 * tcp_try_coalesce - try to merge skb to prior one
4369 * @dest: destination queue
4371 * @from: buffer to add in queue
4372 * @fragstolen: pointer to boolean
4374 * Before queueing skb @from after @to, try to merge them
4375 * to reduce overall memory use and queue lengths, if cost is small.
4376 * Packets in ofo or receive queues can stay a long time.
4377 * Better try to coalesce them right now to avoid future collapses.
4378 * Returns true if caller should free @from instead of queueing it
4380 static bool tcp_try_coalesce(struct sock *sk,
4382 struct sk_buff *from,
4387 *fragstolen = false;
4389 /* Its possible this segment overlaps with prior segment in queue */
4390 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4393 #ifdef CONFIG_TLS_DEVICE
4394 if (from->decrypted != to->decrypted)
4398 if (!skb_try_coalesce(to, from, fragstolen, &delta))
4401 atomic_add(delta, &sk->sk_rmem_alloc);
4402 sk_mem_charge(sk, delta);
4403 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4404 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4405 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4406 TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
4408 if (TCP_SKB_CB(from)->has_rxtstamp) {
4409 TCP_SKB_CB(to)->has_rxtstamp = true;
4410 to->tstamp = from->tstamp;
4411 skb_hwtstamps(to)->hwtstamp = skb_hwtstamps(from)->hwtstamp;
4417 static bool tcp_ooo_try_coalesce(struct sock *sk,
4419 struct sk_buff *from,
4422 bool res = tcp_try_coalesce(sk, to, from, fragstolen);
4424 /* In case tcp_drop() is called later, update to->gso_segs */
4426 u32 gso_segs = max_t(u16, 1, skb_shinfo(to)->gso_segs) +
4427 max_t(u16, 1, skb_shinfo(from)->gso_segs);
4429 skb_shinfo(to)->gso_segs = min_t(u32, gso_segs, 0xFFFF);
4434 static void tcp_drop(struct sock *sk, struct sk_buff *skb)
4436 sk_drops_add(sk, skb);
4440 /* This one checks to see if we can put data from the
4441 * out_of_order queue into the receive_queue.
4443 static void tcp_ofo_queue(struct sock *sk)
4445 struct tcp_sock *tp = tcp_sk(sk);
4446 __u32 dsack_high = tp->rcv_nxt;
4447 bool fin, fragstolen, eaten;
4448 struct sk_buff *skb, *tail;
4451 p = rb_first(&tp->out_of_order_queue);
4454 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4457 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4458 __u32 dsack = dsack_high;
4459 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4460 dsack_high = TCP_SKB_CB(skb)->end_seq;
4461 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4464 rb_erase(&skb->rbnode, &tp->out_of_order_queue);
4466 if (unlikely(!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))) {
4467 SOCK_DEBUG(sk, "ofo packet was already received\n");
4471 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4472 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4473 TCP_SKB_CB(skb)->end_seq);
4475 tail = skb_peek_tail(&sk->sk_receive_queue);
4476 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen);
4477 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4478 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
4480 __skb_queue_tail(&sk->sk_receive_queue, skb);
4482 kfree_skb_partial(skb, fragstolen);
4484 if (unlikely(fin)) {
4486 /* tcp_fin() purges tp->out_of_order_queue,
4487 * so we must end this loop right now.
4494 static bool tcp_prune_ofo_queue(struct sock *sk);
4495 static int tcp_prune_queue(struct sock *sk);
4497 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4500 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4501 !sk_rmem_schedule(sk, skb, size)) {
4503 if (tcp_prune_queue(sk) < 0)
4506 while (!sk_rmem_schedule(sk, skb, size)) {
4507 if (!tcp_prune_ofo_queue(sk))
4514 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4516 struct tcp_sock *tp = tcp_sk(sk);
4517 struct rb_node **p, *parent;
4518 struct sk_buff *skb1;
4522 tcp_ecn_check_ce(sk, skb);
4524 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4525 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFODROP);
4526 sk->sk_data_ready(sk);
4531 /* Disable header prediction. */
4533 inet_csk_schedule_ack(sk);
4535 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4536 seq = TCP_SKB_CB(skb)->seq;
4537 end_seq = TCP_SKB_CB(skb)->end_seq;
4538 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4539 tp->rcv_nxt, seq, end_seq);
4541 p = &tp->out_of_order_queue.rb_node;
4542 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4543 /* Initial out of order segment, build 1 SACK. */
4544 if (tcp_is_sack(tp)) {
4545 tp->rx_opt.num_sacks = 1;
4546 tp->selective_acks[0].start_seq = seq;
4547 tp->selective_acks[0].end_seq = end_seq;
4549 rb_link_node(&skb->rbnode, NULL, p);
4550 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4551 tp->ooo_last_skb = skb;
4555 /* In the typical case, we are adding an skb to the end of the list.
4556 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
4558 if (tcp_ooo_try_coalesce(sk, tp->ooo_last_skb,
4559 skb, &fragstolen)) {
4561 /* For non sack flows, do not grow window to force DUPACK
4562 * and trigger fast retransmit.
4564 if (tcp_is_sack(tp))
4565 tcp_grow_window(sk, skb);
4566 kfree_skb_partial(skb, fragstolen);
4570 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
4571 if (!before(seq, TCP_SKB_CB(tp->ooo_last_skb)->end_seq)) {
4572 parent = &tp->ooo_last_skb->rbnode;
4573 p = &parent->rb_right;
4577 /* Find place to insert this segment. Handle overlaps on the way. */
4581 skb1 = rb_to_skb(parent);
4582 if (before(seq, TCP_SKB_CB(skb1)->seq)) {
4583 p = &parent->rb_left;
4586 if (before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4587 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4588 /* All the bits are present. Drop. */
4589 NET_INC_STATS(sock_net(sk),
4590 LINUX_MIB_TCPOFOMERGE);
4593 tcp_dsack_set(sk, seq, end_seq);
4596 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4597 /* Partial overlap. */
4598 tcp_dsack_set(sk, seq, TCP_SKB_CB(skb1)->end_seq);
4600 /* skb's seq == skb1's seq and skb covers skb1.
4601 * Replace skb1 with skb.
4603 rb_replace_node(&skb1->rbnode, &skb->rbnode,
4604 &tp->out_of_order_queue);
4605 tcp_dsack_extend(sk,
4606 TCP_SKB_CB(skb1)->seq,
4607 TCP_SKB_CB(skb1)->end_seq);
4608 NET_INC_STATS(sock_net(sk),
4609 LINUX_MIB_TCPOFOMERGE);
4613 } else if (tcp_ooo_try_coalesce(sk, skb1,
4614 skb, &fragstolen)) {
4617 p = &parent->rb_right;
4620 /* Insert segment into RB tree. */
4621 rb_link_node(&skb->rbnode, parent, p);
4622 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4625 /* Remove other segments covered by skb. */
4626 while ((skb1 = skb_rb_next(skb)) != NULL) {
4627 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4629 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4630 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4634 rb_erase(&skb1->rbnode, &tp->out_of_order_queue);
4635 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4636 TCP_SKB_CB(skb1)->end_seq);
4637 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4640 /* If there is no skb after us, we are the last_skb ! */
4642 tp->ooo_last_skb = skb;
4645 if (tcp_is_sack(tp))
4646 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4649 /* For non sack flows, do not grow window to force DUPACK
4650 * and trigger fast retransmit.
4652 if (tcp_is_sack(tp))
4653 tcp_grow_window(sk, skb);
4655 skb_set_owner_r(skb, sk);
4659 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, int hdrlen,
4663 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4665 __skb_pull(skb, hdrlen);
4667 tcp_try_coalesce(sk, tail,
4668 skb, fragstolen)) ? 1 : 0;
4669 tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq);
4671 __skb_queue_tail(&sk->sk_receive_queue, skb);
4672 skb_set_owner_r(skb, sk);
4677 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4679 struct sk_buff *skb;
4687 if (size > PAGE_SIZE) {
4688 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS);
4690 data_len = npages << PAGE_SHIFT;
4691 size = data_len + (size & ~PAGE_MASK);
4693 skb = alloc_skb_with_frags(size - data_len, data_len,
4694 PAGE_ALLOC_COSTLY_ORDER,
4695 &err, sk->sk_allocation);
4699 skb_put(skb, size - data_len);
4700 skb->data_len = data_len;
4703 if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) {
4704 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP);
4708 err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size);
4712 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4713 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4714 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4716 if (tcp_queue_rcv(sk, skb, 0, &fragstolen)) {
4717 WARN_ON_ONCE(fragstolen); /* should not happen */
4729 void tcp_data_ready(struct sock *sk)
4731 const struct tcp_sock *tp = tcp_sk(sk);
4732 int avail = tp->rcv_nxt - tp->copied_seq;
4734 if (avail < sk->sk_rcvlowat && !tcp_rmem_pressure(sk) &&
4735 !sock_flag(sk, SOCK_DONE) &&
4736 tcp_receive_window(tp) > inet_csk(sk)->icsk_ack.rcv_mss)
4739 sk->sk_data_ready(sk);
4742 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4744 struct tcp_sock *tp = tcp_sk(sk);
4748 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
4753 __skb_pull(skb, tcp_hdr(skb)->doff * 4);
4755 tp->rx_opt.dsack = 0;
4757 /* Queue data for delivery to the user.
4758 * Packets in sequence go to the receive queue.
4759 * Out of sequence packets to the out_of_order_queue.
4761 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4762 if (tcp_receive_window(tp) == 0) {
4763 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP);
4767 /* Ok. In sequence. In window. */
4769 if (skb_queue_len(&sk->sk_receive_queue) == 0)
4770 sk_forced_mem_schedule(sk, skb->truesize);
4771 else if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) {
4772 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP);
4773 sk->sk_data_ready(sk);
4777 eaten = tcp_queue_rcv(sk, skb, 0, &fragstolen);
4779 tcp_event_data_recv(sk, skb);
4780 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
4783 if (!RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4786 /* RFC5681. 4.2. SHOULD send immediate ACK, when
4787 * gap in queue is filled.
4789 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
4790 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW;
4793 if (tp->rx_opt.num_sacks)
4794 tcp_sack_remove(tp);
4796 tcp_fast_path_check(sk);
4799 kfree_skb_partial(skb, fragstolen);
4800 if (!sock_flag(sk, SOCK_DEAD))
4805 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4806 /* A retransmit, 2nd most common case. Force an immediate ack. */
4807 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4808 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4811 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
4812 inet_csk_schedule_ack(sk);
4818 /* Out of window. F.e. zero window probe. */
4819 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4822 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4823 /* Partial packet, seq < rcv_next < end_seq */
4824 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4825 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4826 TCP_SKB_CB(skb)->end_seq);
4828 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4830 /* If window is closed, drop tail of packet. But after
4831 * remembering D-SACK for its head made in previous line.
4833 if (!tcp_receive_window(tp)) {
4834 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP);
4840 tcp_data_queue_ofo(sk, skb);
4843 static struct sk_buff *tcp_skb_next(struct sk_buff *skb, struct sk_buff_head *list)
4846 return !skb_queue_is_last(list, skb) ? skb->next : NULL;
4848 return skb_rb_next(skb);
4851 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4852 struct sk_buff_head *list,
4853 struct rb_root *root)
4855 struct sk_buff *next = tcp_skb_next(skb, list);
4858 __skb_unlink(skb, list);
4860 rb_erase(&skb->rbnode, root);
4863 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4868 /* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */
4869 void tcp_rbtree_insert(struct rb_root *root, struct sk_buff *skb)
4871 struct rb_node **p = &root->rb_node;
4872 struct rb_node *parent = NULL;
4873 struct sk_buff *skb1;
4877 skb1 = rb_to_skb(parent);
4878 if (before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb1)->seq))
4879 p = &parent->rb_left;
4881 p = &parent->rb_right;
4883 rb_link_node(&skb->rbnode, parent, p);
4884 rb_insert_color(&skb->rbnode, root);
4887 /* Collapse contiguous sequence of skbs head..tail with
4888 * sequence numbers start..end.
4890 * If tail is NULL, this means until the end of the queue.
4892 * Segments with FIN/SYN are not collapsed (only because this
4896 tcp_collapse(struct sock *sk, struct sk_buff_head *list, struct rb_root *root,
4897 struct sk_buff *head, struct sk_buff *tail, u32 start, u32 end)
4899 struct sk_buff *skb = head, *n;
4900 struct sk_buff_head tmp;
4903 /* First, check that queue is collapsible and find
4904 * the point where collapsing can be useful.
4907 for (end_of_skbs = true; skb != NULL && skb != tail; skb = n) {
4908 n = tcp_skb_next(skb, list);
4910 /* No new bits? It is possible on ofo queue. */
4911 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4912 skb = tcp_collapse_one(sk, skb, list, root);
4918 /* The first skb to collapse is:
4920 * - bloated or contains data before "start" or
4921 * overlaps to the next one.
4923 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
4924 (tcp_win_from_space(sk, skb->truesize) > skb->len ||
4925 before(TCP_SKB_CB(skb)->seq, start))) {
4926 end_of_skbs = false;
4930 if (n && n != tail &&
4931 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(n)->seq) {
4932 end_of_skbs = false;
4936 /* Decided to skip this, advance start seq. */
4937 start = TCP_SKB_CB(skb)->end_seq;
4940 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4943 __skb_queue_head_init(&tmp);
4945 while (before(start, end)) {
4946 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
4947 struct sk_buff *nskb;
4949 nskb = alloc_skb(copy, GFP_ATOMIC);
4953 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4954 #ifdef CONFIG_TLS_DEVICE
4955 nskb->decrypted = skb->decrypted;
4957 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4959 __skb_queue_before(list, skb, nskb);
4961 __skb_queue_tail(&tmp, nskb); /* defer rbtree insertion */
4962 skb_set_owner_r(nskb, sk);
4964 /* Copy data, releasing collapsed skbs. */
4966 int offset = start - TCP_SKB_CB(skb)->seq;
4967 int size = TCP_SKB_CB(skb)->end_seq - start;
4971 size = min(copy, size);
4972 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4974 TCP_SKB_CB(nskb)->end_seq += size;
4978 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4979 skb = tcp_collapse_one(sk, skb, list, root);
4982 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4984 #ifdef CONFIG_TLS_DEVICE
4985 if (skb->decrypted != nskb->decrypted)
4992 skb_queue_walk_safe(&tmp, skb, n)
4993 tcp_rbtree_insert(root, skb);
4996 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4997 * and tcp_collapse() them until all the queue is collapsed.
4999 static void tcp_collapse_ofo_queue(struct sock *sk)
5001 struct tcp_sock *tp = tcp_sk(sk);
5002 u32 range_truesize, sum_tiny = 0;
5003 struct sk_buff *skb, *head;
5006 skb = skb_rb_first(&tp->out_of_order_queue);
5009 tp->ooo_last_skb = skb_rb_last(&tp->out_of_order_queue);
5012 start = TCP_SKB_CB(skb)->seq;
5013 end = TCP_SKB_CB(skb)->end_seq;
5014 range_truesize = skb->truesize;
5016 for (head = skb;;) {
5017 skb = skb_rb_next(skb);
5019 /* Range is terminated when we see a gap or when
5020 * we are at the queue end.
5023 after(TCP_SKB_CB(skb)->seq, end) ||
5024 before(TCP_SKB_CB(skb)->end_seq, start)) {
5025 /* Do not attempt collapsing tiny skbs */
5026 if (range_truesize != head->truesize ||
5027 end - start >= SKB_WITH_OVERHEAD(SK_MEM_QUANTUM)) {
5028 tcp_collapse(sk, NULL, &tp->out_of_order_queue,
5029 head, skb, start, end);
5031 sum_tiny += range_truesize;
5032 if (sum_tiny > sk->sk_rcvbuf >> 3)
5038 range_truesize += skb->truesize;
5039 if (unlikely(before(TCP_SKB_CB(skb)->seq, start)))
5040 start = TCP_SKB_CB(skb)->seq;
5041 if (after(TCP_SKB_CB(skb)->end_seq, end))
5042 end = TCP_SKB_CB(skb)->end_seq;
5047 * Clean the out-of-order queue to make room.
5048 * We drop high sequences packets to :
5049 * 1) Let a chance for holes to be filled.
5050 * 2) not add too big latencies if thousands of packets sit there.
5051 * (But if application shrinks SO_RCVBUF, we could still end up
5052 * freeing whole queue here)
5053 * 3) Drop at least 12.5 % of sk_rcvbuf to avoid malicious attacks.
5055 * Return true if queue has shrunk.
5057 static bool tcp_prune_ofo_queue(struct sock *sk)
5059 struct tcp_sock *tp = tcp_sk(sk);
5060 struct rb_node *node, *prev;
5063 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
5066 NET_INC_STATS(sock_net(sk), LINUX_MIB_OFOPRUNED);
5067 goal = sk->sk_rcvbuf >> 3;
5068 node = &tp->ooo_last_skb->rbnode;
5070 prev = rb_prev(node);
5071 rb_erase(node, &tp->out_of_order_queue);
5072 goal -= rb_to_skb(node)->truesize;
5073 tcp_drop(sk, rb_to_skb(node));
5074 if (!prev || goal <= 0) {
5076 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf &&
5077 !tcp_under_memory_pressure(sk))
5079 goal = sk->sk_rcvbuf >> 3;
5083 tp->ooo_last_skb = rb_to_skb(prev);
5085 /* Reset SACK state. A conforming SACK implementation will
5086 * do the same at a timeout based retransmit. When a connection
5087 * is in a sad state like this, we care only about integrity
5088 * of the connection not performance.
5090 if (tp->rx_opt.sack_ok)
5091 tcp_sack_reset(&tp->rx_opt);
5095 /* Reduce allocated memory if we can, trying to get
5096 * the socket within its memory limits again.
5098 * Return less than zero if we should start dropping frames
5099 * until the socket owning process reads some of the data
5100 * to stabilize the situation.
5102 static int tcp_prune_queue(struct sock *sk)
5104 struct tcp_sock *tp = tcp_sk(sk);
5106 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
5108 NET_INC_STATS(sock_net(sk), LINUX_MIB_PRUNECALLED);
5110 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
5111 tcp_clamp_window(sk);
5112 else if (tcp_under_memory_pressure(sk))
5113 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
5115 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5118 tcp_collapse_ofo_queue(sk);
5119 if (!skb_queue_empty(&sk->sk_receive_queue))
5120 tcp_collapse(sk, &sk->sk_receive_queue, NULL,
5121 skb_peek(&sk->sk_receive_queue),
5123 tp->copied_seq, tp->rcv_nxt);
5126 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5129 /* Collapsing did not help, destructive actions follow.
5130 * This must not ever occur. */
5132 tcp_prune_ofo_queue(sk);
5134 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5137 /* If we are really being abused, tell the caller to silently
5138 * drop receive data on the floor. It will get retransmitted
5139 * and hopefully then we'll have sufficient space.
5141 NET_INC_STATS(sock_net(sk), LINUX_MIB_RCVPRUNED);
5143 /* Massive buffer overcommit. */
5148 static bool tcp_should_expand_sndbuf(const struct sock *sk)
5150 const struct tcp_sock *tp = tcp_sk(sk);
5152 /* If the user specified a specific send buffer setting, do
5155 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
5158 /* If we are under global TCP memory pressure, do not expand. */
5159 if (tcp_under_memory_pressure(sk))
5162 /* If we are under soft global TCP memory pressure, do not expand. */
5163 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
5166 /* If we filled the congestion window, do not expand. */
5167 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
5173 /* When incoming ACK allowed to free some skb from write_queue,
5174 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
5175 * on the exit from tcp input handler.
5177 * PROBLEM: sndbuf expansion does not work well with largesend.
5179 static void tcp_new_space(struct sock *sk)
5181 struct tcp_sock *tp = tcp_sk(sk);
5183 if (tcp_should_expand_sndbuf(sk)) {
5184 tcp_sndbuf_expand(sk);
5185 tp->snd_cwnd_stamp = tcp_jiffies32;
5188 sk->sk_write_space(sk);
5191 /* Caller made space either from:
5192 * 1) Freeing skbs in rtx queues (after tp->snd_una has advanced)
5193 * 2) Sent skbs from output queue (and thus advancing tp->snd_nxt)
5195 * We might be able to generate EPOLLOUT to the application if:
5196 * 1) Space consumed in output/rtx queues is below sk->sk_sndbuf/2
5197 * 2) notsent amount (tp->write_seq - tp->snd_nxt) became
5198 * small enough that tcp_stream_memory_free() decides it
5199 * is time to generate EPOLLOUT.
5201 void tcp_check_space(struct sock *sk)
5203 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
5204 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
5205 /* pairs with tcp_poll() */
5207 if (sk->sk_socket &&
5208 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
5210 if (!test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
5211 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
5216 static inline void tcp_data_snd_check(struct sock *sk)
5218 tcp_push_pending_frames(sk);
5219 tcp_check_space(sk);
5223 * Check if sending an ack is needed.
5225 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
5227 struct tcp_sock *tp = tcp_sk(sk);
5228 unsigned long rtt, delay;
5230 /* More than one full frame received... */
5231 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
5232 /* ... and right edge of window advances far enough.
5233 * (tcp_recvmsg() will send ACK otherwise).
5234 * If application uses SO_RCVLOWAT, we want send ack now if
5235 * we have not received enough bytes to satisfy the condition.
5237 (tp->rcv_nxt - tp->copied_seq < sk->sk_rcvlowat ||
5238 __tcp_select_window(sk) >= tp->rcv_wnd)) ||
5239 /* We ACK each frame or... */
5240 tcp_in_quickack_mode(sk) ||
5241 /* Protocol state mandates a one-time immediate ACK */
5242 inet_csk(sk)->icsk_ack.pending & ICSK_ACK_NOW) {
5248 if (!ofo_possible || RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
5249 tcp_send_delayed_ack(sk);
5253 if (!tcp_is_sack(tp) ||
5254 tp->compressed_ack >= READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_nr))
5257 if (tp->compressed_ack_rcv_nxt != tp->rcv_nxt) {
5258 tp->compressed_ack_rcv_nxt = tp->rcv_nxt;
5259 if (tp->compressed_ack > TCP_FASTRETRANS_THRESH)
5260 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED,
5261 tp->compressed_ack - TCP_FASTRETRANS_THRESH);
5262 tp->compressed_ack = 0;
5265 if (++tp->compressed_ack <= TCP_FASTRETRANS_THRESH)
5268 if (hrtimer_is_queued(&tp->compressed_ack_timer))
5271 /* compress ack timer : 5 % of rtt, but no more than tcp_comp_sack_delay_ns */
5273 rtt = tp->rcv_rtt_est.rtt_us;
5274 if (tp->srtt_us && tp->srtt_us < rtt)
5277 delay = min_t(unsigned long,
5278 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_delay_ns),
5279 rtt * (NSEC_PER_USEC >> 3)/20);
5281 hrtimer_start(&tp->compressed_ack_timer, ns_to_ktime(delay),
5282 HRTIMER_MODE_REL_PINNED_SOFT);
5285 static inline void tcp_ack_snd_check(struct sock *sk)
5287 if (!inet_csk_ack_scheduled(sk)) {
5288 /* We sent a data segment already. */
5291 __tcp_ack_snd_check(sk, 1);
5295 * This routine is only called when we have urgent data
5296 * signaled. Its the 'slow' part of tcp_urg. It could be
5297 * moved inline now as tcp_urg is only called from one
5298 * place. We handle URGent data wrong. We have to - as
5299 * BSD still doesn't use the correction from RFC961.
5300 * For 1003.1g we should support a new option TCP_STDURG to permit
5301 * either form (or just set the sysctl tcp_stdurg).
5304 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
5306 struct tcp_sock *tp = tcp_sk(sk);
5307 u32 ptr = ntohs(th->urg_ptr);
5309 if (ptr && !READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_stdurg))
5311 ptr += ntohl(th->seq);
5313 /* Ignore urgent data that we've already seen and read. */
5314 if (after(tp->copied_seq, ptr))
5317 /* Do not replay urg ptr.
5319 * NOTE: interesting situation not covered by specs.
5320 * Misbehaving sender may send urg ptr, pointing to segment,
5321 * which we already have in ofo queue. We are not able to fetch
5322 * such data and will stay in TCP_URG_NOTYET until will be eaten
5323 * by recvmsg(). Seems, we are not obliged to handle such wicked
5324 * situations. But it is worth to think about possibility of some
5325 * DoSes using some hypothetical application level deadlock.
5327 if (before(ptr, tp->rcv_nxt))
5330 /* Do we already have a newer (or duplicate) urgent pointer? */
5331 if (tp->urg_data && !after(ptr, tp->urg_seq))
5334 /* Tell the world about our new urgent pointer. */
5337 /* We may be adding urgent data when the last byte read was
5338 * urgent. To do this requires some care. We cannot just ignore
5339 * tp->copied_seq since we would read the last urgent byte again
5340 * as data, nor can we alter copied_seq until this data arrives
5341 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5343 * NOTE. Double Dutch. Rendering to plain English: author of comment
5344 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5345 * and expect that both A and B disappear from stream. This is _wrong_.
5346 * Though this happens in BSD with high probability, this is occasional.
5347 * Any application relying on this is buggy. Note also, that fix "works"
5348 * only in this artificial test. Insert some normal data between A and B and we will
5349 * decline of BSD again. Verdict: it is better to remove to trap
5352 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5353 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5354 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5356 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5357 __skb_unlink(skb, &sk->sk_receive_queue);
5362 tp->urg_data = TCP_URG_NOTYET;
5365 /* Disable header prediction. */
5369 /* This is the 'fast' part of urgent handling. */
5370 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
5372 struct tcp_sock *tp = tcp_sk(sk);
5374 /* Check if we get a new urgent pointer - normally not. */
5376 tcp_check_urg(sk, th);
5378 /* Do we wait for any urgent data? - normally not... */
5379 if (tp->urg_data == TCP_URG_NOTYET) {
5380 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5383 /* Is the urgent pointer pointing into this packet? */
5384 if (ptr < skb->len) {
5386 if (skb_copy_bits(skb, ptr, &tmp, 1))
5388 tp->urg_data = TCP_URG_VALID | tmp;
5389 if (!sock_flag(sk, SOCK_DEAD))
5390 sk->sk_data_ready(sk);
5395 /* Accept RST for rcv_nxt - 1 after a FIN.
5396 * When tcp connections are abruptly terminated from Mac OSX (via ^C), a
5397 * FIN is sent followed by a RST packet. The RST is sent with the same
5398 * sequence number as the FIN, and thus according to RFC 5961 a challenge
5399 * ACK should be sent. However, Mac OSX rate limits replies to challenge
5400 * ACKs on the closed socket. In addition middleboxes can drop either the
5401 * challenge ACK or a subsequent RST.
5403 static bool tcp_reset_check(const struct sock *sk, const struct sk_buff *skb)
5405 struct tcp_sock *tp = tcp_sk(sk);
5407 return unlikely(TCP_SKB_CB(skb)->seq == (tp->rcv_nxt - 1) &&
5408 (1 << sk->sk_state) & (TCPF_CLOSE_WAIT | TCPF_LAST_ACK |
5412 /* Does PAWS and seqno based validation of an incoming segment, flags will
5413 * play significant role here.
5415 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5416 const struct tcphdr *th, int syn_inerr)
5418 struct tcp_sock *tp = tcp_sk(sk);
5419 bool rst_seq_match = false;
5421 /* RFC1323: H1. Apply PAWS check first. */
5422 if (tcp_fast_parse_options(sock_net(sk), skb, th, tp) &&
5423 tp->rx_opt.saw_tstamp &&
5424 tcp_paws_discard(sk, skb)) {
5426 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5427 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5428 LINUX_MIB_TCPACKSKIPPEDPAWS,
5429 &tp->last_oow_ack_time))
5430 tcp_send_dupack(sk, skb);
5433 /* Reset is accepted even if it did not pass PAWS. */
5436 /* Step 1: check sequence number */
5437 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5438 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5439 * (RST) segments are validated by checking their SEQ-fields."
5440 * And page 69: "If an incoming segment is not acceptable,
5441 * an acknowledgment should be sent in reply (unless the RST
5442 * bit is set, if so drop the segment and return)".
5447 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5448 LINUX_MIB_TCPACKSKIPPEDSEQ,
5449 &tp->last_oow_ack_time))
5450 tcp_send_dupack(sk, skb);
5451 } else if (tcp_reset_check(sk, skb)) {
5457 /* Step 2: check RST bit */
5459 /* RFC 5961 3.2 (extend to match against (RCV.NXT - 1) after a
5460 * FIN and SACK too if available):
5461 * If seq num matches RCV.NXT or (RCV.NXT - 1) after a FIN, or
5462 * the right-most SACK block,
5464 * RESET the connection
5466 * Send a challenge ACK
5468 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt ||
5469 tcp_reset_check(sk, skb)) {
5470 rst_seq_match = true;
5471 } else if (tcp_is_sack(tp) && tp->rx_opt.num_sacks > 0) {
5472 struct tcp_sack_block *sp = &tp->selective_acks[0];
5473 int max_sack = sp[0].end_seq;
5476 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;
5478 max_sack = after(sp[this_sack].end_seq,
5480 sp[this_sack].end_seq : max_sack;
5483 if (TCP_SKB_CB(skb)->seq == max_sack)
5484 rst_seq_match = true;
5490 /* Disable TFO if RST is out-of-order
5491 * and no data has been received
5492 * for current active TFO socket
5494 if (tp->syn_fastopen && !tp->data_segs_in &&
5495 sk->sk_state == TCP_ESTABLISHED)
5496 tcp_fastopen_active_disable(sk);
5497 tcp_send_challenge_ack(sk, skb);
5502 /* step 3: check security and precedence [ignored] */
5504 /* step 4: Check for a SYN
5505 * RFC 5961 4.2 : Send a challenge ack
5510 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5511 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5512 tcp_send_challenge_ack(sk, skb);
5524 * TCP receive function for the ESTABLISHED state.
5526 * It is split into a fast path and a slow path. The fast path is
5528 * - A zero window was announced from us - zero window probing
5529 * is only handled properly in the slow path.
5530 * - Out of order segments arrived.
5531 * - Urgent data is expected.
5532 * - There is no buffer space left
5533 * - Unexpected TCP flags/window values/header lengths are received
5534 * (detected by checking the TCP header against pred_flags)
5535 * - Data is sent in both directions. Fast path only supports pure senders
5536 * or pure receivers (this means either the sequence number or the ack
5537 * value must stay constant)
5538 * - Unexpected TCP option.
5540 * When these conditions are not satisfied it drops into a standard
5541 * receive procedure patterned after RFC793 to handle all cases.
5542 * The first three cases are guaranteed by proper pred_flags setting,
5543 * the rest is checked inline. Fast processing is turned on in
5544 * tcp_data_queue when everything is OK.
5546 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb)
5548 const struct tcphdr *th = (const struct tcphdr *)skb->data;
5549 struct tcp_sock *tp = tcp_sk(sk);
5550 unsigned int len = skb->len;
5552 /* TCP congestion window tracking */
5553 trace_tcp_probe(sk, skb);
5555 tcp_mstamp_refresh(tp);
5556 if (unlikely(!rcu_access_pointer(sk->sk_rx_dst)))
5557 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
5559 * Header prediction.
5560 * The code loosely follows the one in the famous
5561 * "30 instruction TCP receive" Van Jacobson mail.
5563 * Van's trick is to deposit buffers into socket queue
5564 * on a device interrupt, to call tcp_recv function
5565 * on the receive process context and checksum and copy
5566 * the buffer to user space. smart...
5568 * Our current scheme is not silly either but we take the
5569 * extra cost of the net_bh soft interrupt processing...
5570 * We do checksum and copy also but from device to kernel.
5573 tp->rx_opt.saw_tstamp = 0;
5575 /* pred_flags is 0xS?10 << 16 + snd_wnd
5576 * if header_prediction is to be made
5577 * 'S' will always be tp->tcp_header_len >> 2
5578 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5579 * turn it off (when there are holes in the receive
5580 * space for instance)
5581 * PSH flag is ignored.
5584 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5585 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5586 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5587 int tcp_header_len = tp->tcp_header_len;
5589 /* Timestamp header prediction: tcp_header_len
5590 * is automatically equal to th->doff*4 due to pred_flags
5594 /* Check timestamp */
5595 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5596 /* No? Slow path! */
5597 if (!tcp_parse_aligned_timestamp(tp, th))
5600 /* If PAWS failed, check it more carefully in slow path */
5601 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5604 /* DO NOT update ts_recent here, if checksum fails
5605 * and timestamp was corrupted part, it will result
5606 * in a hung connection since we will drop all
5607 * future packets due to the PAWS test.
5611 if (len <= tcp_header_len) {
5612 /* Bulk data transfer: sender */
5613 if (len == tcp_header_len) {
5614 /* Predicted packet is in window by definition.
5615 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5616 * Hence, check seq<=rcv_wup reduces to:
5618 if (tcp_header_len ==
5619 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5620 tp->rcv_nxt == tp->rcv_wup)
5621 tcp_store_ts_recent(tp);
5623 /* We know that such packets are checksummed
5626 tcp_ack(sk, skb, 0);
5628 tcp_data_snd_check(sk);
5629 /* When receiving pure ack in fast path, update
5630 * last ts ecr directly instead of calling
5631 * tcp_rcv_rtt_measure_ts()
5633 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr;
5635 } else { /* Header too small */
5636 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5641 bool fragstolen = false;
5643 if (tcp_checksum_complete(skb))
5646 if ((int)skb->truesize > sk->sk_forward_alloc)
5649 /* Predicted packet is in window by definition.
5650 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5651 * Hence, check seq<=rcv_wup reduces to:
5653 if (tcp_header_len ==
5654 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5655 tp->rcv_nxt == tp->rcv_wup)
5656 tcp_store_ts_recent(tp);
5658 tcp_rcv_rtt_measure_ts(sk, skb);
5660 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPHITS);
5662 /* Bulk data transfer: receiver */
5663 eaten = tcp_queue_rcv(sk, skb, tcp_header_len,
5666 tcp_event_data_recv(sk, skb);
5668 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5669 /* Well, only one small jumplet in fast path... */
5670 tcp_ack(sk, skb, FLAG_DATA);
5671 tcp_data_snd_check(sk);
5672 if (!inet_csk_ack_scheduled(sk))
5675 tcp_update_wl(tp, TCP_SKB_CB(skb)->seq);
5678 __tcp_ack_snd_check(sk, 0);
5681 kfree_skb_partial(skb, fragstolen);
5688 if (len < (th->doff << 2) || tcp_checksum_complete(skb))
5691 if (!th->ack && !th->rst && !th->syn)
5695 * Standard slow path.
5698 if (!tcp_validate_incoming(sk, skb, th, 1))
5702 if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0)
5705 tcp_rcv_rtt_measure_ts(sk, skb);
5707 /* Process urgent data. */
5708 tcp_urg(sk, skb, th);
5710 /* step 7: process the segment text */
5711 tcp_data_queue(sk, skb);
5713 tcp_data_snd_check(sk);
5714 tcp_ack_snd_check(sk);
5718 TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS);
5719 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5724 EXPORT_SYMBOL(tcp_rcv_established);
5726 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5728 struct tcp_sock *tp = tcp_sk(sk);
5729 struct inet_connection_sock *icsk = inet_csk(sk);
5731 tcp_set_state(sk, TCP_ESTABLISHED);
5732 icsk->icsk_ack.lrcvtime = tcp_jiffies32;
5735 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
5736 security_inet_conn_established(sk, skb);
5737 sk_mark_napi_id(sk, skb);
5740 tcp_init_transfer(sk, BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB);
5742 /* Prevent spurious tcp_cwnd_restart() on first data
5745 tp->lsndtime = tcp_jiffies32;
5747 if (sock_flag(sk, SOCK_KEEPOPEN))
5748 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5750 if (!tp->rx_opt.snd_wscale)
5751 __tcp_fast_path_on(tp, tp->snd_wnd);
5756 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
5757 struct tcp_fastopen_cookie *cookie)
5759 struct tcp_sock *tp = tcp_sk(sk);
5760 struct sk_buff *data = tp->syn_data ? tcp_rtx_queue_head(sk) : NULL;
5761 u16 mss = tp->rx_opt.mss_clamp, try_exp = 0;
5762 bool syn_drop = false;
5764 if (mss == tp->rx_opt.user_mss) {
5765 struct tcp_options_received opt;
5767 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5768 tcp_clear_options(&opt);
5769 opt.user_mss = opt.mss_clamp = 0;
5770 tcp_parse_options(sock_net(sk), synack, &opt, 0, NULL);
5771 mss = opt.mss_clamp;
5774 if (!tp->syn_fastopen) {
5775 /* Ignore an unsolicited cookie */
5777 } else if (tp->total_retrans) {
5778 /* SYN timed out and the SYN-ACK neither has a cookie nor
5779 * acknowledges data. Presumably the remote received only
5780 * the retransmitted (regular) SYNs: either the original
5781 * SYN-data or the corresponding SYN-ACK was dropped.
5783 syn_drop = (cookie->len < 0 && data);
5784 } else if (cookie->len < 0 && !tp->syn_data) {
5785 /* We requested a cookie but didn't get it. If we did not use
5786 * the (old) exp opt format then try so next time (try_exp=1).
5787 * Otherwise we go back to use the RFC7413 opt (try_exp=2).
5789 try_exp = tp->syn_fastopen_exp ? 2 : 1;
5792 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp);
5794 if (data) { /* Retransmit unacked data in SYN */
5795 skb_rbtree_walk_from(data) {
5796 if (__tcp_retransmit_skb(sk, data, 1))
5800 NET_INC_STATS(sock_net(sk),
5801 LINUX_MIB_TCPFASTOPENACTIVEFAIL);
5804 tp->syn_data_acked = tp->syn_data;
5805 if (tp->syn_data_acked) {
5806 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
5807 /* SYN-data is counted as two separate packets in tcp_ack() */
5808 if (tp->delivered > 1)
5812 tcp_fastopen_add_skb(sk, synack);
5817 static void smc_check_reset_syn(struct tcp_sock *tp)
5819 #if IS_ENABLED(CONFIG_SMC)
5820 if (static_branch_unlikely(&tcp_have_smc)) {
5821 if (tp->syn_smc && !tp->rx_opt.smc_ok)
5827 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5828 const struct tcphdr *th)
5830 struct inet_connection_sock *icsk = inet_csk(sk);
5831 struct tcp_sock *tp = tcp_sk(sk);
5832 struct tcp_fastopen_cookie foc = { .len = -1 };
5833 int saved_clamp = tp->rx_opt.mss_clamp;
5836 tcp_parse_options(sock_net(sk), skb, &tp->rx_opt, 0, &foc);
5837 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
5838 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
5842 * "If the state is SYN-SENT then
5843 * first check the ACK bit
5844 * If the ACK bit is set
5845 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5846 * a reset (unless the RST bit is set, if so drop
5847 * the segment and return)"
5849 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
5850 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt))
5851 goto reset_and_undo;
5853 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5854 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5855 tcp_time_stamp(tp))) {
5856 NET_INC_STATS(sock_net(sk),
5857 LINUX_MIB_PAWSACTIVEREJECTED);
5858 goto reset_and_undo;
5861 /* Now ACK is acceptable.
5863 * "If the RST bit is set
5864 * If the ACK was acceptable then signal the user "error:
5865 * connection reset", drop the segment, enter CLOSED state,
5866 * delete TCB, and return."
5875 * "fifth, if neither of the SYN or RST bits is set then
5876 * drop the segment and return."
5882 goto discard_and_undo;
5885 * "If the SYN bit is on ...
5886 * are acceptable then ...
5887 * (our SYN has been ACKed), change the connection
5888 * state to ESTABLISHED..."
5891 tcp_ecn_rcv_synack(tp, th);
5893 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5894 tcp_ack(sk, skb, FLAG_SLOWPATH);
5896 /* Ok.. it's good. Set up sequence numbers and
5897 * move to established.
5899 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1);
5900 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5902 /* RFC1323: The window in SYN & SYN/ACK segments is
5905 tp->snd_wnd = ntohs(th->window);
5907 if (!tp->rx_opt.wscale_ok) {
5908 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5909 tp->window_clamp = min(tp->window_clamp, 65535U);
5912 if (tp->rx_opt.saw_tstamp) {
5913 tp->rx_opt.tstamp_ok = 1;
5914 tp->tcp_header_len =
5915 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5916 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5917 tcp_store_ts_recent(tp);
5919 tp->tcp_header_len = sizeof(struct tcphdr);
5922 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5923 tcp_initialize_rcv_mss(sk);
5925 /* Remember, tcp_poll() does not lock socket!
5926 * Change state from SYN-SENT only after copied_seq
5927 * is initialized. */
5928 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
5930 smc_check_reset_syn(tp);
5934 tcp_finish_connect(sk, skb);
5936 fastopen_fail = (tp->syn_fastopen || tp->syn_data) &&
5937 tcp_rcv_fastopen_synack(sk, skb, &foc);
5939 if (!sock_flag(sk, SOCK_DEAD)) {
5940 sk->sk_state_change(sk);
5941 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5945 if (sk->sk_write_pending ||
5946 icsk->icsk_accept_queue.rskq_defer_accept ||
5947 icsk->icsk_ack.pingpong) {
5948 /* Save one ACK. Data will be ready after
5949 * several ticks, if write_pending is set.
5951 * It may be deleted, but with this feature tcpdumps
5952 * look so _wonderfully_ clever, that I was not able
5953 * to stand against the temptation 8) --ANK
5955 inet_csk_schedule_ack(sk);
5956 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
5957 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5958 TCP_DELACK_MAX, TCP_RTO_MAX);
5969 /* No ACK in the segment */
5973 * "If the RST bit is set
5975 * Otherwise (no ACK) drop the segment and return."
5978 goto discard_and_undo;
5982 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5983 tcp_paws_reject(&tp->rx_opt, 0))
5984 goto discard_and_undo;
5987 /* We see SYN without ACK. It is attempt of
5988 * simultaneous connect with crossed SYNs.
5989 * Particularly, it can be connect to self.
5991 tcp_set_state(sk, TCP_SYN_RECV);
5993 if (tp->rx_opt.saw_tstamp) {
5994 tp->rx_opt.tstamp_ok = 1;
5995 tcp_store_ts_recent(tp);
5996 tp->tcp_header_len =
5997 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5999 tp->tcp_header_len = sizeof(struct tcphdr);
6002 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1);
6003 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6004 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
6006 /* RFC1323: The window in SYN & SYN/ACK segments is
6009 tp->snd_wnd = ntohs(th->window);
6010 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
6011 tp->max_window = tp->snd_wnd;
6013 tcp_ecn_rcv_syn(tp, th);
6016 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
6017 tcp_initialize_rcv_mss(sk);
6019 tcp_send_synack(sk);
6021 /* Note, we could accept data and URG from this segment.
6022 * There are no obstacles to make this (except that we must
6023 * either change tcp_recvmsg() to prevent it from returning data
6024 * before 3WHS completes per RFC793, or employ TCP Fast Open).
6026 * However, if we ignore data in ACKless segments sometimes,
6027 * we have no reasons to accept it sometimes.
6028 * Also, seems the code doing it in step6 of tcp_rcv_state_process
6029 * is not flawless. So, discard packet for sanity.
6030 * Uncomment this return to process the data.
6037 /* "fifth, if neither of the SYN or RST bits is set then
6038 * drop the segment and return."
6042 tcp_clear_options(&tp->rx_opt);
6043 tp->rx_opt.mss_clamp = saved_clamp;
6047 tcp_clear_options(&tp->rx_opt);
6048 tp->rx_opt.mss_clamp = saved_clamp;
6053 * This function implements the receiving procedure of RFC 793 for
6054 * all states except ESTABLISHED and TIME_WAIT.
6055 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
6056 * address independent.
6059 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb)
6061 struct tcp_sock *tp = tcp_sk(sk);
6062 struct inet_connection_sock *icsk = inet_csk(sk);
6063 const struct tcphdr *th = tcp_hdr(skb);
6064 struct request_sock *req;
6068 switch (sk->sk_state) {
6082 /* It is possible that we process SYN packets from backlog,
6083 * so we need to make sure to disable BH and RCU right there.
6087 acceptable = icsk->icsk_af_ops->conn_request(sk, skb) >= 0;
6099 tp->rx_opt.saw_tstamp = 0;
6100 tcp_mstamp_refresh(tp);
6101 queued = tcp_rcv_synsent_state_process(sk, skb, th);
6105 /* Do step6 onward by hand. */
6106 tcp_urg(sk, skb, th);
6108 tcp_data_snd_check(sk);
6112 tcp_mstamp_refresh(tp);
6113 tp->rx_opt.saw_tstamp = 0;
6114 req = tp->fastopen_rsk;
6118 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
6119 sk->sk_state != TCP_FIN_WAIT1);
6121 if (!tcp_check_req(sk, skb, req, true, &req_stolen))
6125 if (!th->ack && !th->rst && !th->syn)
6128 if (!tcp_validate_incoming(sk, skb, th, 0))
6131 /* step 5: check the ACK field */
6132 acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
6133 FLAG_UPDATE_TS_RECENT |
6134 FLAG_NO_CHALLENGE_ACK) > 0;
6137 if (sk->sk_state == TCP_SYN_RECV)
6138 return 1; /* send one RST */
6139 tcp_send_challenge_ack(sk, skb);
6142 switch (sk->sk_state) {
6144 tp->delivered++; /* SYN-ACK delivery isn't tracked in tcp_ack */
6146 tcp_synack_rtt_meas(sk, req);
6148 /* Once we leave TCP_SYN_RECV, we no longer need req
6152 inet_csk(sk)->icsk_retransmits = 0;
6153 reqsk_fastopen_remove(sk, req, false);
6154 /* Re-arm the timer because data may have been sent out.
6155 * This is similar to the regular data transmission case
6156 * when new data has just been ack'ed.
6158 * (TFO) - we could try to be more aggressive and
6159 * retransmitting any data sooner based on when they
6164 tcp_init_transfer(sk, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB);
6165 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6168 tcp_set_state(sk, TCP_ESTABLISHED);
6169 sk->sk_state_change(sk);
6171 /* Note, that this wakeup is only for marginal crossed SYN case.
6172 * Passively open sockets are not waked up, because
6173 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
6176 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
6178 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
6179 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
6180 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
6182 if (tp->rx_opt.tstamp_ok)
6183 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
6185 if (!inet_csk(sk)->icsk_ca_ops->cong_control)
6186 tcp_update_pacing_rate(sk);
6188 /* Prevent spurious tcp_cwnd_restart() on first data packet */
6189 tp->lsndtime = tcp_jiffies32;
6191 tcp_initialize_rcv_mss(sk);
6192 tcp_fast_path_on(tp);
6195 case TCP_FIN_WAIT1: {
6198 /* If we enter the TCP_FIN_WAIT1 state and we are a
6199 * Fast Open socket and this is the first acceptable
6200 * ACK we have received, this would have acknowledged
6201 * our SYNACK so stop the SYNACK timer.
6204 /* We no longer need the request sock. */
6205 reqsk_fastopen_remove(sk, req, false);
6208 if (tp->snd_una != tp->write_seq)
6211 tcp_set_state(sk, TCP_FIN_WAIT2);
6212 WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | SEND_SHUTDOWN);
6216 if (!sock_flag(sk, SOCK_DEAD)) {
6217 /* Wake up lingering close() */
6218 sk->sk_state_change(sk);
6222 if (tp->linger2 < 0) {
6224 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6227 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6228 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6229 /* Receive out of order FIN after close() */
6230 if (tp->syn_fastopen && th->fin)
6231 tcp_fastopen_active_disable(sk);
6233 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6237 tmo = tcp_fin_time(sk);
6238 if (tmo > TCP_TIMEWAIT_LEN) {
6239 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
6240 } else if (th->fin || sock_owned_by_user(sk)) {
6241 /* Bad case. We could lose such FIN otherwise.
6242 * It is not a big problem, but it looks confusing
6243 * and not so rare event. We still can lose it now,
6244 * if it spins in bh_lock_sock(), but it is really
6247 inet_csk_reset_keepalive_timer(sk, tmo);
6249 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
6256 if (tp->snd_una == tp->write_seq) {
6257 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
6263 if (tp->snd_una == tp->write_seq) {
6264 tcp_update_metrics(sk);
6271 /* step 6: check the URG bit */
6272 tcp_urg(sk, skb, th);
6274 /* step 7: process the segment text */
6275 switch (sk->sk_state) {
6276 case TCP_CLOSE_WAIT:
6279 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
6284 /* RFC 793 says to queue data in these states,
6285 * RFC 1122 says we MUST send a reset.
6286 * BSD 4.4 also does reset.
6288 if (sk->sk_shutdown & RCV_SHUTDOWN) {
6289 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6290 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6291 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6297 case TCP_ESTABLISHED:
6298 tcp_data_queue(sk, skb);
6303 /* tcp_data could move socket to TIME-WAIT */
6304 if (sk->sk_state != TCP_CLOSE) {
6305 tcp_data_snd_check(sk);
6306 tcp_ack_snd_check(sk);
6315 EXPORT_SYMBOL(tcp_rcv_state_process);
6317 static inline void pr_drop_req(struct request_sock *req, __u16 port, int family)
6319 struct inet_request_sock *ireq = inet_rsk(req);
6321 if (family == AF_INET)
6322 net_dbg_ratelimited("drop open request from %pI4/%u\n",
6323 &ireq->ir_rmt_addr, port);
6324 #if IS_ENABLED(CONFIG_IPV6)
6325 else if (family == AF_INET6)
6326 net_dbg_ratelimited("drop open request from %pI6/%u\n",
6327 &ireq->ir_v6_rmt_addr, port);
6331 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
6333 * If we receive a SYN packet with these bits set, it means a
6334 * network is playing bad games with TOS bits. In order to
6335 * avoid possible false congestion notifications, we disable
6336 * TCP ECN negotiation.
6338 * Exception: tcp_ca wants ECN. This is required for DCTCP
6339 * congestion control: Linux DCTCP asserts ECT on all packets,
6340 * including SYN, which is most optimal solution; however,
6341 * others, such as FreeBSD do not.
6343 static void tcp_ecn_create_request(struct request_sock *req,
6344 const struct sk_buff *skb,
6345 const struct sock *listen_sk,
6346 const struct dst_entry *dst)
6348 const struct tcphdr *th = tcp_hdr(skb);
6349 const struct net *net = sock_net(listen_sk);
6350 bool th_ecn = th->ece && th->cwr;
6357 ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield);
6358 ecn_ok_dst = dst_feature(dst, DST_FEATURE_ECN_MASK);
6359 ecn_ok = net->ipv4.sysctl_tcp_ecn || ecn_ok_dst;
6361 if ((!ect && ecn_ok) || tcp_ca_needs_ecn(listen_sk) ||
6362 (ecn_ok_dst & DST_FEATURE_ECN_CA) ||
6363 tcp_bpf_ca_needs_ecn((struct sock *)req))
6364 inet_rsk(req)->ecn_ok = 1;
6367 static void tcp_openreq_init(struct request_sock *req,
6368 const struct tcp_options_received *rx_opt,
6369 struct sk_buff *skb, const struct sock *sk)
6371 struct inet_request_sock *ireq = inet_rsk(req);
6373 req->rsk_rcv_wnd = 0; /* So that tcp_send_synack() knows! */
6375 tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq;
6376 tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
6377 tcp_rsk(req)->snt_synack = tcp_clock_us();
6378 tcp_rsk(req)->last_oow_ack_time = 0;
6379 req->mss = rx_opt->mss_clamp;
6380 req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0;
6381 ireq->tstamp_ok = rx_opt->tstamp_ok;
6382 ireq->sack_ok = rx_opt->sack_ok;
6383 ireq->snd_wscale = rx_opt->snd_wscale;
6384 ireq->wscale_ok = rx_opt->wscale_ok;
6387 ireq->ir_rmt_port = tcp_hdr(skb)->source;
6388 ireq->ir_num = ntohs(tcp_hdr(skb)->dest);
6389 ireq->ir_mark = inet_request_mark(sk, skb);
6390 #if IS_ENABLED(CONFIG_SMC)
6391 ireq->smc_ok = rx_opt->smc_ok;
6395 struct request_sock *inet_reqsk_alloc(const struct request_sock_ops *ops,
6396 struct sock *sk_listener,
6397 bool attach_listener)
6399 struct request_sock *req = reqsk_alloc(ops, sk_listener,
6403 struct inet_request_sock *ireq = inet_rsk(req);
6405 ireq->ireq_opt = NULL;
6406 #if IS_ENABLED(CONFIG_IPV6)
6407 ireq->pktopts = NULL;
6409 atomic64_set(&ireq->ir_cookie, 0);
6410 ireq->ireq_state = TCP_NEW_SYN_RECV;
6411 write_pnet(&ireq->ireq_net, sock_net(sk_listener));
6412 ireq->ireq_family = sk_listener->sk_family;
6417 EXPORT_SYMBOL(inet_reqsk_alloc);
6420 * Return true if a syncookie should be sent
6422 static bool tcp_syn_flood_action(const struct sock *sk,
6423 const struct sk_buff *skb,
6426 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
6427 const char *msg = "Dropping request";
6428 bool want_cookie = false;
6429 struct net *net = sock_net(sk);
6431 #ifdef CONFIG_SYN_COOKIES
6432 if (net->ipv4.sysctl_tcp_syncookies) {
6433 msg = "Sending cookies";
6435 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES);
6438 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP);
6440 if (!queue->synflood_warned &&
6441 net->ipv4.sysctl_tcp_syncookies != 2 &&
6442 xchg(&queue->synflood_warned, 1) == 0)
6443 net_info_ratelimited("%s: Possible SYN flooding on port %d. %s. Check SNMP counters.\n",
6444 proto, ntohs(tcp_hdr(skb)->dest), msg);
6449 static void tcp_reqsk_record_syn(const struct sock *sk,
6450 struct request_sock *req,
6451 const struct sk_buff *skb)
6453 if (tcp_sk(sk)->save_syn) {
6454 u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb);
6457 copy = kmalloc(len + sizeof(u32), GFP_ATOMIC);
6460 memcpy(©[1], skb_network_header(skb), len);
6461 req->saved_syn = copy;
6466 int tcp_conn_request(struct request_sock_ops *rsk_ops,
6467 const struct tcp_request_sock_ops *af_ops,
6468 struct sock *sk, struct sk_buff *skb)
6470 struct tcp_fastopen_cookie foc = { .len = -1 };
6471 __u32 isn = TCP_SKB_CB(skb)->tcp_tw_isn;
6472 struct tcp_options_received tmp_opt;
6473 struct tcp_sock *tp = tcp_sk(sk);
6474 struct net *net = sock_net(sk);
6475 struct sock *fastopen_sk = NULL;
6476 struct request_sock *req;
6477 bool want_cookie = false;
6478 struct dst_entry *dst;
6481 /* TW buckets are converted to open requests without
6482 * limitations, they conserve resources and peer is
6483 * evidently real one.
6485 if ((net->ipv4.sysctl_tcp_syncookies == 2 ||
6486 inet_csk_reqsk_queue_is_full(sk)) && !isn) {
6487 want_cookie = tcp_syn_flood_action(sk, skb, rsk_ops->slab_name);
6492 if (sk_acceptq_is_full(sk)) {
6493 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
6497 req = inet_reqsk_alloc(rsk_ops, sk, !want_cookie);
6501 tcp_rsk(req)->af_specific = af_ops;
6502 tcp_rsk(req)->ts_off = 0;
6504 tcp_clear_options(&tmp_opt);
6505 tmp_opt.mss_clamp = af_ops->mss_clamp;
6506 tmp_opt.user_mss = tp->rx_opt.user_mss;
6507 tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0,
6508 want_cookie ? NULL : &foc);
6510 if (want_cookie && !tmp_opt.saw_tstamp)
6511 tcp_clear_options(&tmp_opt);
6513 if (IS_ENABLED(CONFIG_SMC) && want_cookie)
6516 tmp_opt.tstamp_ok = tmp_opt.saw_tstamp;
6517 tcp_openreq_init(req, &tmp_opt, skb, sk);
6518 inet_rsk(req)->no_srccheck = inet_sk(sk)->transparent;
6520 /* Note: tcp_v6_init_req() might override ir_iif for link locals */
6521 inet_rsk(req)->ir_iif = inet_request_bound_dev_if(sk, skb);
6523 af_ops->init_req(req, sk, skb);
6525 if (security_inet_conn_request(sk, skb, req))
6528 if (tmp_opt.tstamp_ok)
6529 tcp_rsk(req)->ts_off = af_ops->init_ts_off(net, skb);
6531 dst = af_ops->route_req(sk, &fl, req);
6535 if (!want_cookie && !isn) {
6536 /* Kill the following clause, if you dislike this way. */
6537 if (!net->ipv4.sysctl_tcp_syncookies &&
6538 (net->ipv4.sysctl_max_syn_backlog - inet_csk_reqsk_queue_len(sk) <
6539 (net->ipv4.sysctl_max_syn_backlog >> 2)) &&
6540 !tcp_peer_is_proven(req, dst)) {
6541 /* Without syncookies last quarter of
6542 * backlog is filled with destinations,
6543 * proven to be alive.
6544 * It means that we continue to communicate
6545 * to destinations, already remembered
6546 * to the moment of synflood.
6548 pr_drop_req(req, ntohs(tcp_hdr(skb)->source),
6550 goto drop_and_release;
6553 isn = af_ops->init_seq(skb);
6556 tcp_ecn_create_request(req, skb, sk, dst);
6559 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss);
6560 req->cookie_ts = tmp_opt.tstamp_ok;
6561 if (!tmp_opt.tstamp_ok)
6562 inet_rsk(req)->ecn_ok = 0;
6565 tcp_rsk(req)->snt_isn = isn;
6566 tcp_rsk(req)->txhash = net_tx_rndhash();
6567 tcp_openreq_init_rwin(req, sk, dst);
6568 sk_rx_queue_set(req_to_sk(req), skb);
6570 tcp_reqsk_record_syn(sk, req, skb);
6571 fastopen_sk = tcp_try_fastopen(sk, skb, req, &foc, dst);
6574 af_ops->send_synack(fastopen_sk, dst, &fl, req,
6575 &foc, TCP_SYNACK_FASTOPEN);
6576 /* Add the child socket directly into the accept queue */
6577 if (!inet_csk_reqsk_queue_add(sk, req, fastopen_sk)) {
6578 reqsk_fastopen_remove(fastopen_sk, req, false);
6579 bh_unlock_sock(fastopen_sk);
6580 sock_put(fastopen_sk);
6584 sk->sk_data_ready(sk);
6585 bh_unlock_sock(fastopen_sk);
6586 sock_put(fastopen_sk);
6588 tcp_rsk(req)->tfo_listener = false;
6590 inet_csk_reqsk_queue_hash_add(sk, req,
6591 tcp_timeout_init((struct sock *)req));
6592 af_ops->send_synack(sk, dst, &fl, req, &foc,
6593 !want_cookie ? TCP_SYNACK_NORMAL :
6611 EXPORT_SYMBOL(tcp_conn_request);