1 // SPDX-License-Identifier: GPL-2.0
3 * INET An implementation of the TCP/IP protocol suite for the LINUX
4 * operating system. INET is implemented using the BSD Socket
5 * interface as the means of communication with the user level.
7 * Implementation of the Transmission Control Protocol(TCP).
10 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
11 * Mark Evans, <evansmp@uhura.aston.ac.uk>
12 * Corey Minyard <wf-rch!minyard@relay.EU.net>
13 * Florian La Roche, <flla@stud.uni-sb.de>
14 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
15 * Linus Torvalds, <torvalds@cs.helsinki.fi>
16 * Alan Cox, <gw4pts@gw4pts.ampr.org>
17 * Matthew Dillon, <dillon@apollo.west.oic.com>
18 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
19 * Jorge Cwik, <jorge@laser.satlink.net>
24 * Pedro Roque : Fast Retransmit/Recovery.
26 * Retransmit queue handled by TCP.
27 * Better retransmit timer handling.
28 * New congestion avoidance.
32 * Eric : Fast Retransmit.
33 * Randy Scott : MSS option defines.
34 * Eric Schenk : Fixes to slow start algorithm.
35 * Eric Schenk : Yet another double ACK bug.
36 * Eric Schenk : Delayed ACK bug fixes.
37 * Eric Schenk : Floyd style fast retrans war avoidance.
38 * David S. Miller : Don't allow zero congestion window.
39 * Eric Schenk : Fix retransmitter so that it sends
40 * next packet on ack of previous packet.
41 * Andi Kleen : Moved open_request checking here
42 * and process RSTs for open_requests.
43 * Andi Kleen : Better prune_queue, and other fixes.
44 * Andrey Savochkin: Fix RTT measurements in the presence of
46 * Andrey Savochkin: Check sequence numbers correctly when
47 * removing SACKs due to in sequence incoming
49 * Andi Kleen: Make sure we never ack data there is not
50 * enough room for. Also make this condition
51 * a fatal error if it might still happen.
52 * Andi Kleen: Add tcp_measure_rcv_mss to make
53 * connections with MSS<min(MTU,ann. MSS)
54 * work without delayed acks.
55 * Andi Kleen: Process packets with PSH set in the
57 * J Hadi Salim: ECN support
60 * Panu Kuhlberg: Experimental audit of TCP (re)transmission
61 * engine. Lots of bugs are found.
62 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs
65 #define pr_fmt(fmt) "TCP: " fmt
68 #include <linux/slab.h>
69 #include <linux/module.h>
70 #include <linux/sysctl.h>
71 #include <linux/kernel.h>
72 #include <linux/prefetch.h>
75 #include <net/inet_common.h>
76 #include <linux/ipsec.h>
77 #include <asm/unaligned.h>
78 #include <linux/errqueue.h>
79 #include <trace/events/tcp.h>
80 #include <linux/jump_label_ratelimit.h>
81 #include <net/busy_poll.h>
83 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
85 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
86 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
87 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
88 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
89 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
90 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
91 #define FLAG_ECE 0x40 /* ECE in this ACK */
92 #define FLAG_LOST_RETRANS 0x80 /* This ACK marks some retransmission lost */
93 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
94 #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
95 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
96 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
97 #define FLAG_SET_XMIT_TIMER 0x1000 /* Set TLP or RTO timer */
98 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
99 #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
100 #define FLAG_NO_CHALLENGE_ACK 0x8000 /* do not call tcp_send_challenge_ack() */
101 #define FLAG_ACK_MAYBE_DELAYED 0x10000 /* Likely a delayed ACK */
103 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
104 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
105 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE|FLAG_DSACKING_ACK)
106 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
108 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
109 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
111 #define REXMIT_NONE 0 /* no loss recovery to do */
112 #define REXMIT_LOST 1 /* retransmit packets marked lost */
113 #define REXMIT_NEW 2 /* FRTO-style transmit of unsent/new packets */
115 #if IS_ENABLED(CONFIG_TLS_DEVICE)
116 static DEFINE_STATIC_KEY_DEFERRED_FALSE(clean_acked_data_enabled, HZ);
118 void clean_acked_data_enable(struct inet_connection_sock *icsk,
119 void (*cad)(struct sock *sk, u32 ack_seq))
121 icsk->icsk_clean_acked = cad;
122 static_branch_deferred_inc(&clean_acked_data_enabled);
124 EXPORT_SYMBOL_GPL(clean_acked_data_enable);
126 void clean_acked_data_disable(struct inet_connection_sock *icsk)
128 static_branch_slow_dec_deferred(&clean_acked_data_enabled);
129 icsk->icsk_clean_acked = NULL;
131 EXPORT_SYMBOL_GPL(clean_acked_data_disable);
133 void clean_acked_data_flush(void)
135 static_key_deferred_flush(&clean_acked_data_enabled);
137 EXPORT_SYMBOL_GPL(clean_acked_data_flush);
140 static void tcp_gro_dev_warn(struct sock *sk, const struct sk_buff *skb,
143 static bool __once __read_mostly;
146 struct net_device *dev;
151 dev = dev_get_by_index_rcu(sock_net(sk), skb->skb_iif);
152 if (!dev || len >= dev->mtu)
153 pr_warn("%s: Driver has suspect GRO implementation, TCP performance may be compromised.\n",
154 dev ? dev->name : "Unknown driver");
159 /* Adapt the MSS value used to make delayed ack decision to the
162 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
164 struct inet_connection_sock *icsk = inet_csk(sk);
165 const unsigned int lss = icsk->icsk_ack.last_seg_size;
168 icsk->icsk_ack.last_seg_size = 0;
170 /* skb->len may jitter because of SACKs, even if peer
171 * sends good full-sized frames.
173 len = skb_shinfo(skb)->gso_size ? : skb->len;
174 if (len >= icsk->icsk_ack.rcv_mss) {
175 icsk->icsk_ack.rcv_mss = min_t(unsigned int, len,
177 /* Account for possibly-removed options */
178 if (unlikely(len > icsk->icsk_ack.rcv_mss +
179 MAX_TCP_OPTION_SPACE))
180 tcp_gro_dev_warn(sk, skb, len);
182 /* Otherwise, we make more careful check taking into account,
183 * that SACKs block is variable.
185 * "len" is invariant segment length, including TCP header.
187 len += skb->data - skb_transport_header(skb);
188 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
189 /* If PSH is not set, packet should be
190 * full sized, provided peer TCP is not badly broken.
191 * This observation (if it is correct 8)) allows
192 * to handle super-low mtu links fairly.
194 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
195 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
196 /* Subtract also invariant (if peer is RFC compliant),
197 * tcp header plus fixed timestamp option length.
198 * Resulting "len" is MSS free of SACK jitter.
200 len -= tcp_sk(sk)->tcp_header_len;
201 icsk->icsk_ack.last_seg_size = len;
203 icsk->icsk_ack.rcv_mss = len;
207 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
208 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
209 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
213 static void tcp_incr_quickack(struct sock *sk, unsigned int max_quickacks)
215 struct inet_connection_sock *icsk = inet_csk(sk);
216 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
220 quickacks = min(quickacks, max_quickacks);
221 if (quickacks > icsk->icsk_ack.quick)
222 icsk->icsk_ack.quick = quickacks;
225 void tcp_enter_quickack_mode(struct sock *sk, unsigned int max_quickacks)
227 struct inet_connection_sock *icsk = inet_csk(sk);
229 tcp_incr_quickack(sk, max_quickacks);
230 inet_csk_exit_pingpong_mode(sk);
231 icsk->icsk_ack.ato = TCP_ATO_MIN;
233 EXPORT_SYMBOL(tcp_enter_quickack_mode);
235 /* Send ACKs quickly, if "quick" count is not exhausted
236 * and the session is not interactive.
239 static bool tcp_in_quickack_mode(struct sock *sk)
241 const struct inet_connection_sock *icsk = inet_csk(sk);
242 const struct dst_entry *dst = __sk_dst_get(sk);
244 return (dst && dst_metric(dst, RTAX_QUICKACK)) ||
245 (icsk->icsk_ack.quick && !inet_csk_in_pingpong_mode(sk));
248 static void tcp_ecn_queue_cwr(struct tcp_sock *tp)
250 if (tp->ecn_flags & TCP_ECN_OK)
251 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
254 static void tcp_ecn_accept_cwr(struct sock *sk, const struct sk_buff *skb)
256 if (tcp_hdr(skb)->cwr) {
257 tcp_sk(sk)->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
259 /* If the sender is telling us it has entered CWR, then its
260 * cwnd may be very low (even just 1 packet), so we should ACK
263 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW;
267 static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp)
269 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
272 static void __tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb)
274 struct tcp_sock *tp = tcp_sk(sk);
276 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
277 case INET_ECN_NOT_ECT:
278 /* Funny extension: if ECT is not set on a segment,
279 * and we already seen ECT on a previous segment,
280 * it is probably a retransmit.
282 if (tp->ecn_flags & TCP_ECN_SEEN)
283 tcp_enter_quickack_mode(sk, 2);
286 if (tcp_ca_needs_ecn(sk))
287 tcp_ca_event(sk, CA_EVENT_ECN_IS_CE);
289 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
290 /* Better not delay acks, sender can have a very low cwnd */
291 tcp_enter_quickack_mode(sk, 2);
292 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
294 tp->ecn_flags |= TCP_ECN_SEEN;
297 if (tcp_ca_needs_ecn(sk))
298 tcp_ca_event(sk, CA_EVENT_ECN_NO_CE);
299 tp->ecn_flags |= TCP_ECN_SEEN;
304 static void tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb)
306 if (tcp_sk(sk)->ecn_flags & TCP_ECN_OK)
307 __tcp_ecn_check_ce(sk, skb);
310 static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
312 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
313 tp->ecn_flags &= ~TCP_ECN_OK;
316 static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
318 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
319 tp->ecn_flags &= ~TCP_ECN_OK;
322 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
324 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
329 /* Buffer size and advertised window tuning.
331 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
334 static void tcp_sndbuf_expand(struct sock *sk)
336 const struct tcp_sock *tp = tcp_sk(sk);
337 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
341 /* Worst case is non GSO/TSO : each frame consumes one skb
342 * and skb->head is kmalloced using power of two area of memory
344 per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
346 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
348 per_mss = roundup_pow_of_two(per_mss) +
349 SKB_DATA_ALIGN(sizeof(struct sk_buff));
351 nr_segs = max_t(u32, TCP_INIT_CWND, tp->snd_cwnd);
352 nr_segs = max_t(u32, nr_segs, tp->reordering + 1);
354 /* Fast Recovery (RFC 5681 3.2) :
355 * Cubic needs 1.7 factor, rounded to 2 to include
356 * extra cushion (application might react slowly to EPOLLOUT)
358 sndmem = ca_ops->sndbuf_expand ? ca_ops->sndbuf_expand(sk) : 2;
359 sndmem *= nr_segs * per_mss;
361 if (sk->sk_sndbuf < sndmem)
362 sk->sk_sndbuf = min(sndmem, sock_net(sk)->ipv4.sysctl_tcp_wmem[2]);
365 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
367 * All tcp_full_space() is split to two parts: "network" buffer, allocated
368 * forward and advertised in receiver window (tp->rcv_wnd) and
369 * "application buffer", required to isolate scheduling/application
370 * latencies from network.
371 * window_clamp is maximal advertised window. It can be less than
372 * tcp_full_space(), in this case tcp_full_space() - window_clamp
373 * is reserved for "application" buffer. The less window_clamp is
374 * the smoother our behaviour from viewpoint of network, but the lower
375 * throughput and the higher sensitivity of the connection to losses. 8)
377 * rcv_ssthresh is more strict window_clamp used at "slow start"
378 * phase to predict further behaviour of this connection.
379 * It is used for two goals:
380 * - to enforce header prediction at sender, even when application
381 * requires some significant "application buffer". It is check #1.
382 * - to prevent pruning of receive queue because of misprediction
383 * of receiver window. Check #2.
385 * The scheme does not work when sender sends good segments opening
386 * window and then starts to feed us spaghetti. But it should work
387 * in common situations. Otherwise, we have to rely on queue collapsing.
390 /* Slow part of check#2. */
391 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
393 struct tcp_sock *tp = tcp_sk(sk);
395 int truesize = tcp_win_from_space(sk, skb->truesize) >> 1;
396 int window = tcp_win_from_space(sk, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]) >> 1;
398 while (tp->rcv_ssthresh <= window) {
399 if (truesize <= skb->len)
400 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
408 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
410 struct tcp_sock *tp = tcp_sk(sk);
413 room = min_t(int, tp->window_clamp, tcp_space(sk)) - tp->rcv_ssthresh;
416 if (room > 0 && !tcp_under_memory_pressure(sk)) {
419 /* Check #2. Increase window, if skb with such overhead
420 * will fit to rcvbuf in future.
422 if (tcp_win_from_space(sk, skb->truesize) <= skb->len)
423 incr = 2 * tp->advmss;
425 incr = __tcp_grow_window(sk, skb);
428 incr = max_t(int, incr, 2 * skb->len);
429 tp->rcv_ssthresh += min(room, incr);
430 inet_csk(sk)->icsk_ack.quick |= 1;
435 /* 3. Try to fixup all. It is made immediately after connection enters
438 void tcp_init_buffer_space(struct sock *sk)
440 int tcp_app_win = sock_net(sk)->ipv4.sysctl_tcp_app_win;
441 struct tcp_sock *tp = tcp_sk(sk);
444 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
445 tcp_sndbuf_expand(sk);
447 tp->rcvq_space.space = min_t(u32, tp->rcv_wnd, TCP_INIT_CWND * tp->advmss);
448 tcp_mstamp_refresh(tp);
449 tp->rcvq_space.time = tp->tcp_mstamp;
450 tp->rcvq_space.seq = tp->copied_seq;
452 maxwin = tcp_full_space(sk);
454 if (tp->window_clamp >= maxwin) {
455 tp->window_clamp = maxwin;
457 if (tcp_app_win && maxwin > 4 * tp->advmss)
458 tp->window_clamp = max(maxwin -
459 (maxwin >> tcp_app_win),
463 /* Force reservation of one segment. */
465 tp->window_clamp > 2 * tp->advmss &&
466 tp->window_clamp + tp->advmss > maxwin)
467 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
469 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
470 tp->snd_cwnd_stamp = tcp_jiffies32;
473 /* 4. Recalculate window clamp after socket hit its memory bounds. */
474 static void tcp_clamp_window(struct sock *sk)
476 struct tcp_sock *tp = tcp_sk(sk);
477 struct inet_connection_sock *icsk = inet_csk(sk);
478 struct net *net = sock_net(sk);
480 icsk->icsk_ack.quick = 0;
482 if (sk->sk_rcvbuf < net->ipv4.sysctl_tcp_rmem[2] &&
483 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
484 !tcp_under_memory_pressure(sk) &&
485 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
486 WRITE_ONCE(sk->sk_rcvbuf,
487 min(atomic_read(&sk->sk_rmem_alloc),
488 net->ipv4.sysctl_tcp_rmem[2]));
490 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
491 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
494 /* Initialize RCV_MSS value.
495 * RCV_MSS is an our guess about MSS used by the peer.
496 * We haven't any direct information about the MSS.
497 * It's better to underestimate the RCV_MSS rather than overestimate.
498 * Overestimations make us ACKing less frequently than needed.
499 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
501 void tcp_initialize_rcv_mss(struct sock *sk)
503 const struct tcp_sock *tp = tcp_sk(sk);
504 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
506 hint = min(hint, tp->rcv_wnd / 2);
507 hint = min(hint, TCP_MSS_DEFAULT);
508 hint = max(hint, TCP_MIN_MSS);
510 inet_csk(sk)->icsk_ack.rcv_mss = hint;
512 EXPORT_SYMBOL(tcp_initialize_rcv_mss);
514 /* Receiver "autotuning" code.
516 * The algorithm for RTT estimation w/o timestamps is based on
517 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
518 * <http://public.lanl.gov/radiant/pubs.html#DRS>
520 * More detail on this code can be found at
521 * <http://staff.psc.edu/jheffner/>,
522 * though this reference is out of date. A new paper
525 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
527 u32 new_sample = tp->rcv_rtt_est.rtt_us;
530 if (new_sample != 0) {
531 /* If we sample in larger samples in the non-timestamp
532 * case, we could grossly overestimate the RTT especially
533 * with chatty applications or bulk transfer apps which
534 * are stalled on filesystem I/O.
536 * Also, since we are only going for a minimum in the
537 * non-timestamp case, we do not smooth things out
538 * else with timestamps disabled convergence takes too
542 m -= (new_sample >> 3);
550 /* No previous measure. */
554 tp->rcv_rtt_est.rtt_us = new_sample;
557 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
561 if (tp->rcv_rtt_est.time == 0)
563 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
565 delta_us = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcv_rtt_est.time);
568 tcp_rcv_rtt_update(tp, delta_us, 1);
571 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
572 tp->rcv_rtt_est.time = tp->tcp_mstamp;
575 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
576 const struct sk_buff *skb)
578 struct tcp_sock *tp = tcp_sk(sk);
580 if (tp->rx_opt.rcv_tsecr == tp->rcv_rtt_last_tsecr)
582 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr;
584 if (TCP_SKB_CB(skb)->end_seq -
585 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss) {
586 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr;
589 if (likely(delta < INT_MAX / (USEC_PER_SEC / TCP_TS_HZ))) {
592 delta_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
593 tcp_rcv_rtt_update(tp, delta_us, 0);
599 * This function should be called every time data is copied to user space.
600 * It calculates the appropriate TCP receive buffer space.
602 void tcp_rcv_space_adjust(struct sock *sk)
604 struct tcp_sock *tp = tcp_sk(sk);
608 trace_tcp_rcv_space_adjust(sk);
610 tcp_mstamp_refresh(tp);
611 time = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcvq_space.time);
612 if (time < (tp->rcv_rtt_est.rtt_us >> 3) || tp->rcv_rtt_est.rtt_us == 0)
615 /* Number of bytes copied to user in last RTT */
616 copied = tp->copied_seq - tp->rcvq_space.seq;
617 if (copied <= tp->rcvq_space.space)
621 * copied = bytes received in previous RTT, our base window
622 * To cope with packet losses, we need a 2x factor
623 * To cope with slow start, and sender growing its cwin by 100 %
624 * every RTT, we need a 4x factor, because the ACK we are sending
625 * now is for the next RTT, not the current one :
626 * <prev RTT . ><current RTT .. ><next RTT .... >
629 if (sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf &&
630 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
634 /* minimal window to cope with packet losses, assuming
635 * steady state. Add some cushion because of small variations.
637 rcvwin = ((u64)copied << 1) + 16 * tp->advmss;
639 /* Accommodate for sender rate increase (eg. slow start) */
640 grow = rcvwin * (copied - tp->rcvq_space.space);
641 do_div(grow, tp->rcvq_space.space);
642 rcvwin += (grow << 1);
644 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
645 while (tcp_win_from_space(sk, rcvmem) < tp->advmss)
648 do_div(rcvwin, tp->advmss);
649 rcvbuf = min_t(u64, rcvwin * rcvmem,
650 sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
651 if (rcvbuf > sk->sk_rcvbuf) {
652 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
654 /* Make the window clamp follow along. */
655 tp->window_clamp = tcp_win_from_space(sk, rcvbuf);
658 tp->rcvq_space.space = copied;
661 tp->rcvq_space.seq = tp->copied_seq;
662 tp->rcvq_space.time = tp->tcp_mstamp;
665 /* There is something which you must keep in mind when you analyze the
666 * behavior of the tp->ato delayed ack timeout interval. When a
667 * connection starts up, we want to ack as quickly as possible. The
668 * problem is that "good" TCP's do slow start at the beginning of data
669 * transmission. The means that until we send the first few ACK's the
670 * sender will sit on his end and only queue most of his data, because
671 * he can only send snd_cwnd unacked packets at any given time. For
672 * each ACK we send, he increments snd_cwnd and transmits more of his
675 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
677 struct tcp_sock *tp = tcp_sk(sk);
678 struct inet_connection_sock *icsk = inet_csk(sk);
681 inet_csk_schedule_ack(sk);
683 tcp_measure_rcv_mss(sk, skb);
685 tcp_rcv_rtt_measure(tp);
689 if (!icsk->icsk_ack.ato) {
690 /* The _first_ data packet received, initialize
691 * delayed ACK engine.
693 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
694 icsk->icsk_ack.ato = TCP_ATO_MIN;
696 int m = now - icsk->icsk_ack.lrcvtime;
698 if (m <= TCP_ATO_MIN / 2) {
699 /* The fastest case is the first. */
700 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
701 } else if (m < icsk->icsk_ack.ato) {
702 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
703 if (icsk->icsk_ack.ato > icsk->icsk_rto)
704 icsk->icsk_ack.ato = icsk->icsk_rto;
705 } else if (m > icsk->icsk_rto) {
706 /* Too long gap. Apparently sender failed to
707 * restart window, so that we send ACKs quickly.
709 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
713 icsk->icsk_ack.lrcvtime = now;
715 tcp_ecn_check_ce(sk, skb);
718 tcp_grow_window(sk, skb);
721 /* Called to compute a smoothed rtt estimate. The data fed to this
722 * routine either comes from timestamps, or from segments that were
723 * known _not_ to have been retransmitted [see Karn/Partridge
724 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
725 * piece by Van Jacobson.
726 * NOTE: the next three routines used to be one big routine.
727 * To save cycles in the RFC 1323 implementation it was better to break
728 * it up into three procedures. -- erics
730 static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
732 struct tcp_sock *tp = tcp_sk(sk);
733 long m = mrtt_us; /* RTT */
734 u32 srtt = tp->srtt_us;
736 /* The following amusing code comes from Jacobson's
737 * article in SIGCOMM '88. Note that rtt and mdev
738 * are scaled versions of rtt and mean deviation.
739 * This is designed to be as fast as possible
740 * m stands for "measurement".
742 * On a 1990 paper the rto value is changed to:
743 * RTO = rtt + 4 * mdev
745 * Funny. This algorithm seems to be very broken.
746 * These formulae increase RTO, when it should be decreased, increase
747 * too slowly, when it should be increased quickly, decrease too quickly
748 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
749 * does not matter how to _calculate_ it. Seems, it was trap
750 * that VJ failed to avoid. 8)
753 m -= (srtt >> 3); /* m is now error in rtt est */
754 srtt += m; /* rtt = 7/8 rtt + 1/8 new */
756 m = -m; /* m is now abs(error) */
757 m -= (tp->mdev_us >> 2); /* similar update on mdev */
758 /* This is similar to one of Eifel findings.
759 * Eifel blocks mdev updates when rtt decreases.
760 * This solution is a bit different: we use finer gain
761 * for mdev in this case (alpha*beta).
762 * Like Eifel it also prevents growth of rto,
763 * but also it limits too fast rto decreases,
764 * happening in pure Eifel.
769 m -= (tp->mdev_us >> 2); /* similar update on mdev */
771 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */
772 if (tp->mdev_us > tp->mdev_max_us) {
773 tp->mdev_max_us = tp->mdev_us;
774 if (tp->mdev_max_us > tp->rttvar_us)
775 tp->rttvar_us = tp->mdev_max_us;
777 if (after(tp->snd_una, tp->rtt_seq)) {
778 if (tp->mdev_max_us < tp->rttvar_us)
779 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
780 tp->rtt_seq = tp->snd_nxt;
781 tp->mdev_max_us = tcp_rto_min_us(sk);
786 /* no previous measure. */
787 srtt = m << 3; /* take the measured time to be rtt */
788 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */
789 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
790 tp->mdev_max_us = tp->rttvar_us;
791 tp->rtt_seq = tp->snd_nxt;
795 tp->srtt_us = max(1U, srtt);
798 static void tcp_update_pacing_rate(struct sock *sk)
800 const struct tcp_sock *tp = tcp_sk(sk);
803 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
804 rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3);
806 /* current rate is (cwnd * mss) / srtt
807 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
808 * In Congestion Avoidance phase, set it to 120 % the current rate.
810 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
811 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
812 * end of slow start and should slow down.
814 if (tp->snd_cwnd < tp->snd_ssthresh / 2)
815 rate *= sock_net(sk)->ipv4.sysctl_tcp_pacing_ss_ratio;
817 rate *= sock_net(sk)->ipv4.sysctl_tcp_pacing_ca_ratio;
819 rate *= max(tp->snd_cwnd, tp->packets_out);
821 if (likely(tp->srtt_us))
822 do_div(rate, tp->srtt_us);
824 /* WRITE_ONCE() is needed because sch_fq fetches sk_pacing_rate
825 * without any lock. We want to make sure compiler wont store
826 * intermediate values in this location.
828 WRITE_ONCE(sk->sk_pacing_rate, min_t(u64, rate,
829 sk->sk_max_pacing_rate));
832 /* Calculate rto without backoff. This is the second half of Van Jacobson's
833 * routine referred to above.
835 static void tcp_set_rto(struct sock *sk)
837 const struct tcp_sock *tp = tcp_sk(sk);
838 /* Old crap is replaced with new one. 8)
841 * 1. If rtt variance happened to be less 50msec, it is hallucination.
842 * It cannot be less due to utterly erratic ACK generation made
843 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
844 * to do with delayed acks, because at cwnd>2 true delack timeout
845 * is invisible. Actually, Linux-2.4 also generates erratic
846 * ACKs in some circumstances.
848 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
850 /* 2. Fixups made earlier cannot be right.
851 * If we do not estimate RTO correctly without them,
852 * all the algo is pure shit and should be replaced
853 * with correct one. It is exactly, which we pretend to do.
856 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
857 * guarantees that rto is higher.
862 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
864 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
867 cwnd = TCP_INIT_CWND;
868 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
871 /* Take a notice that peer is sending D-SACKs */
872 static void tcp_dsack_seen(struct tcp_sock *tp)
874 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
875 tp->rack.dsack_seen = 1;
879 /* It's reordering when higher sequence was delivered (i.e. sacked) before
880 * some lower never-retransmitted sequence ("low_seq"). The maximum reordering
881 * distance is approximated in full-mss packet distance ("reordering").
883 static void tcp_check_sack_reordering(struct sock *sk, const u32 low_seq,
886 struct tcp_sock *tp = tcp_sk(sk);
887 const u32 mss = tp->mss_cache;
890 fack = tcp_highest_sack_seq(tp);
891 if (!before(low_seq, fack))
894 metric = fack - low_seq;
895 if ((metric > tp->reordering * mss) && mss) {
896 #if FASTRETRANS_DEBUG > 1
897 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
898 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
902 tp->undo_marker ? tp->undo_retrans : 0);
904 tp->reordering = min_t(u32, (metric + mss - 1) / mss,
905 sock_net(sk)->ipv4.sysctl_tcp_max_reordering);
908 /* This exciting event is worth to be remembered. 8) */
910 NET_INC_STATS(sock_net(sk),
911 ts ? LINUX_MIB_TCPTSREORDER : LINUX_MIB_TCPSACKREORDER);
914 /* This must be called before lost_out is incremented */
915 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
917 if (!tp->retransmit_skb_hint ||
918 before(TCP_SKB_CB(skb)->seq,
919 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
920 tp->retransmit_skb_hint = skb;
923 /* Sum the number of packets on the wire we have marked as lost.
924 * There are two cases we care about here:
925 * a) Packet hasn't been marked lost (nor retransmitted),
926 * and this is the first loss.
927 * b) Packet has been marked both lost and retransmitted,
928 * and this means we think it was lost again.
930 static void tcp_sum_lost(struct tcp_sock *tp, struct sk_buff *skb)
932 __u8 sacked = TCP_SKB_CB(skb)->sacked;
934 if (!(sacked & TCPCB_LOST) ||
935 ((sacked & TCPCB_LOST) && (sacked & TCPCB_SACKED_RETRANS)))
936 tp->lost += tcp_skb_pcount(skb);
939 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
941 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
942 tcp_verify_retransmit_hint(tp, skb);
944 tp->lost_out += tcp_skb_pcount(skb);
945 tcp_sum_lost(tp, skb);
946 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
950 void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp, struct sk_buff *skb)
952 tcp_verify_retransmit_hint(tp, skb);
954 tcp_sum_lost(tp, skb);
955 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
956 tp->lost_out += tcp_skb_pcount(skb);
957 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
961 /* This procedure tags the retransmission queue when SACKs arrive.
963 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
964 * Packets in queue with these bits set are counted in variables
965 * sacked_out, retrans_out and lost_out, correspondingly.
967 * Valid combinations are:
968 * Tag InFlight Description
969 * 0 1 - orig segment is in flight.
970 * S 0 - nothing flies, orig reached receiver.
971 * L 0 - nothing flies, orig lost by net.
972 * R 2 - both orig and retransmit are in flight.
973 * L|R 1 - orig is lost, retransmit is in flight.
974 * S|R 1 - orig reached receiver, retrans is still in flight.
975 * (L|S|R is logically valid, it could occur when L|R is sacked,
976 * but it is equivalent to plain S and code short-curcuits it to S.
977 * L|S is logically invalid, it would mean -1 packet in flight 8))
979 * These 6 states form finite state machine, controlled by the following events:
980 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
981 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
982 * 3. Loss detection event of two flavors:
983 * A. Scoreboard estimator decided the packet is lost.
984 * A'. Reno "three dupacks" marks head of queue lost.
985 * B. SACK arrives sacking SND.NXT at the moment, when the
986 * segment was retransmitted.
987 * 4. D-SACK added new rule: D-SACK changes any tag to S.
989 * It is pleasant to note, that state diagram turns out to be commutative,
990 * so that we are allowed not to be bothered by order of our actions,
991 * when multiple events arrive simultaneously. (see the function below).
993 * Reordering detection.
994 * --------------------
995 * Reordering metric is maximal distance, which a packet can be displaced
996 * in packet stream. With SACKs we can estimate it:
998 * 1. SACK fills old hole and the corresponding segment was not
999 * ever retransmitted -> reordering. Alas, we cannot use it
1000 * when segment was retransmitted.
1001 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1002 * for retransmitted and already SACKed segment -> reordering..
1003 * Both of these heuristics are not used in Loss state, when we cannot
1004 * account for retransmits accurately.
1006 * SACK block validation.
1007 * ----------------------
1009 * SACK block range validation checks that the received SACK block fits to
1010 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1011 * Note that SND.UNA is not included to the range though being valid because
1012 * it means that the receiver is rather inconsistent with itself reporting
1013 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1014 * perfectly valid, however, in light of RFC2018 which explicitly states
1015 * that "SACK block MUST reflect the newest segment. Even if the newest
1016 * segment is going to be discarded ...", not that it looks very clever
1017 * in case of head skb. Due to potentional receiver driven attacks, we
1018 * choose to avoid immediate execution of a walk in write queue due to
1019 * reneging and defer head skb's loss recovery to standard loss recovery
1020 * procedure that will eventually trigger (nothing forbids us doing this).
1022 * Implements also blockage to start_seq wrap-around. Problem lies in the
1023 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1024 * there's no guarantee that it will be before snd_nxt (n). The problem
1025 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1028 * <- outs wnd -> <- wrapzone ->
1029 * u e n u_w e_w s n_w
1031 * |<------------+------+----- TCP seqno space --------------+---------->|
1032 * ...-- <2^31 ->| |<--------...
1033 * ...---- >2^31 ------>| |<--------...
1035 * Current code wouldn't be vulnerable but it's better still to discard such
1036 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1037 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1038 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1039 * equal to the ideal case (infinite seqno space without wrap caused issues).
1041 * With D-SACK the lower bound is extended to cover sequence space below
1042 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1043 * again, D-SACK block must not to go across snd_una (for the same reason as
1044 * for the normal SACK blocks, explained above). But there all simplicity
1045 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1046 * fully below undo_marker they do not affect behavior in anyway and can
1047 * therefore be safely ignored. In rare cases (which are more or less
1048 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1049 * fragmentation and packet reordering past skb's retransmission. To consider
1050 * them correctly, the acceptable range must be extended even more though
1051 * the exact amount is rather hard to quantify. However, tp->max_window can
1052 * be used as an exaggerated estimate.
1054 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
1055 u32 start_seq, u32 end_seq)
1057 /* Too far in future, or reversed (interpretation is ambiguous) */
1058 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1061 /* Nasty start_seq wrap-around check (see comments above) */
1062 if (!before(start_seq, tp->snd_nxt))
1065 /* In outstanding window? ...This is valid exit for D-SACKs too.
1066 * start_seq == snd_una is non-sensical (see comments above)
1068 if (after(start_seq, tp->snd_una))
1071 if (!is_dsack || !tp->undo_marker)
1074 /* ...Then it's D-SACK, and must reside below snd_una completely */
1075 if (after(end_seq, tp->snd_una))
1078 if (!before(start_seq, tp->undo_marker))
1082 if (!after(end_seq, tp->undo_marker))
1085 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1086 * start_seq < undo_marker and end_seq >= undo_marker.
1088 return !before(start_seq, end_seq - tp->max_window);
1091 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1092 struct tcp_sack_block_wire *sp, int num_sacks,
1095 struct tcp_sock *tp = tcp_sk(sk);
1096 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1097 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1098 bool dup_sack = false;
1100 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1103 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1104 } else if (num_sacks > 1) {
1105 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1106 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1108 if (!after(end_seq_0, end_seq_1) &&
1109 !before(start_seq_0, start_seq_1)) {
1112 NET_INC_STATS(sock_net(sk),
1113 LINUX_MIB_TCPDSACKOFORECV);
1117 /* D-SACK for already forgotten data... Do dumb counting. */
1118 if (dup_sack && tp->undo_marker && tp->undo_retrans > 0 &&
1119 !after(end_seq_0, prior_snd_una) &&
1120 after(end_seq_0, tp->undo_marker))
1126 struct tcp_sacktag_state {
1128 /* Timestamps for earliest and latest never-retransmitted segment
1129 * that was SACKed. RTO needs the earliest RTT to stay conservative,
1130 * but congestion control should still get an accurate delay signal.
1134 struct rate_sample *rate;
1136 unsigned int mss_now;
1139 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1140 * the incoming SACK may not exactly match but we can find smaller MSS
1141 * aligned portion of it that matches. Therefore we might need to fragment
1142 * which may fail and creates some hassle (caller must handle error case
1145 * FIXME: this could be merged to shift decision code
1147 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1148 u32 start_seq, u32 end_seq)
1152 unsigned int pkt_len;
1155 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1156 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1158 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1159 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1160 mss = tcp_skb_mss(skb);
1161 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1164 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1168 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1173 /* Round if necessary so that SACKs cover only full MSSes
1174 * and/or the remaining small portion (if present)
1176 if (pkt_len > mss) {
1177 unsigned int new_len = (pkt_len / mss) * mss;
1178 if (!in_sack && new_len < pkt_len)
1183 if (pkt_len >= skb->len && !in_sack)
1186 err = tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
1187 pkt_len, mss, GFP_ATOMIC);
1195 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1196 static u8 tcp_sacktag_one(struct sock *sk,
1197 struct tcp_sacktag_state *state, u8 sacked,
1198 u32 start_seq, u32 end_seq,
1199 int dup_sack, int pcount,
1202 struct tcp_sock *tp = tcp_sk(sk);
1204 /* Account D-SACK for retransmitted packet. */
1205 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1206 if (tp->undo_marker && tp->undo_retrans > 0 &&
1207 after(end_seq, tp->undo_marker))
1209 if ((sacked & TCPCB_SACKED_ACKED) &&
1210 before(start_seq, state->reord))
1211 state->reord = start_seq;
1214 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1215 if (!after(end_seq, tp->snd_una))
1218 if (!(sacked & TCPCB_SACKED_ACKED)) {
1219 tcp_rack_advance(tp, sacked, end_seq, xmit_time);
1221 if (sacked & TCPCB_SACKED_RETRANS) {
1222 /* If the segment is not tagged as lost,
1223 * we do not clear RETRANS, believing
1224 * that retransmission is still in flight.
1226 if (sacked & TCPCB_LOST) {
1227 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1228 tp->lost_out -= pcount;
1229 tp->retrans_out -= pcount;
1232 if (!(sacked & TCPCB_RETRANS)) {
1233 /* New sack for not retransmitted frame,
1234 * which was in hole. It is reordering.
1236 if (before(start_seq,
1237 tcp_highest_sack_seq(tp)) &&
1238 before(start_seq, state->reord))
1239 state->reord = start_seq;
1241 if (!after(end_seq, tp->high_seq))
1242 state->flag |= FLAG_ORIG_SACK_ACKED;
1243 if (state->first_sackt == 0)
1244 state->first_sackt = xmit_time;
1245 state->last_sackt = xmit_time;
1248 if (sacked & TCPCB_LOST) {
1249 sacked &= ~TCPCB_LOST;
1250 tp->lost_out -= pcount;
1254 sacked |= TCPCB_SACKED_ACKED;
1255 state->flag |= FLAG_DATA_SACKED;
1256 tp->sacked_out += pcount;
1257 tp->delivered += pcount; /* Out-of-order packets delivered */
1259 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1260 if (tp->lost_skb_hint &&
1261 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1262 tp->lost_cnt_hint += pcount;
1265 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1266 * frames and clear it. undo_retrans is decreased above, L|R frames
1267 * are accounted above as well.
1269 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1270 sacked &= ~TCPCB_SACKED_RETRANS;
1271 tp->retrans_out -= pcount;
1277 /* Shift newly-SACKed bytes from this skb to the immediately previous
1278 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1280 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *prev,
1281 struct sk_buff *skb,
1282 struct tcp_sacktag_state *state,
1283 unsigned int pcount, int shifted, int mss,
1286 struct tcp_sock *tp = tcp_sk(sk);
1287 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1288 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
1292 /* Adjust counters and hints for the newly sacked sequence
1293 * range but discard the return value since prev is already
1294 * marked. We must tag the range first because the seq
1295 * advancement below implicitly advances
1296 * tcp_highest_sack_seq() when skb is highest_sack.
1298 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1299 start_seq, end_seq, dup_sack, pcount,
1300 tcp_skb_timestamp_us(skb));
1301 tcp_rate_skb_delivered(sk, skb, state->rate);
1303 if (skb == tp->lost_skb_hint)
1304 tp->lost_cnt_hint += pcount;
1306 TCP_SKB_CB(prev)->end_seq += shifted;
1307 TCP_SKB_CB(skb)->seq += shifted;
1309 tcp_skb_pcount_add(prev, pcount);
1310 WARN_ON_ONCE(tcp_skb_pcount(skb) < pcount);
1311 tcp_skb_pcount_add(skb, -pcount);
1313 /* When we're adding to gso_segs == 1, gso_size will be zero,
1314 * in theory this shouldn't be necessary but as long as DSACK
1315 * code can come after this skb later on it's better to keep
1316 * setting gso_size to something.
1318 if (!TCP_SKB_CB(prev)->tcp_gso_size)
1319 TCP_SKB_CB(prev)->tcp_gso_size = mss;
1321 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1322 if (tcp_skb_pcount(skb) <= 1)
1323 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1325 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1326 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1329 BUG_ON(!tcp_skb_pcount(skb));
1330 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1334 /* Whole SKB was eaten :-) */
1336 if (skb == tp->retransmit_skb_hint)
1337 tp->retransmit_skb_hint = prev;
1338 if (skb == tp->lost_skb_hint) {
1339 tp->lost_skb_hint = prev;
1340 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1343 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1344 TCP_SKB_CB(prev)->eor = TCP_SKB_CB(skb)->eor;
1345 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1346 TCP_SKB_CB(prev)->end_seq++;
1348 if (skb == tcp_highest_sack(sk))
1349 tcp_advance_highest_sack(sk, skb);
1351 tcp_skb_collapse_tstamp(prev, skb);
1352 if (unlikely(TCP_SKB_CB(prev)->tx.delivered_mstamp))
1353 TCP_SKB_CB(prev)->tx.delivered_mstamp = 0;
1355 tcp_rtx_queue_unlink_and_free(skb, sk);
1357 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKMERGED);
1362 /* I wish gso_size would have a bit more sane initialization than
1363 * something-or-zero which complicates things
1365 static int tcp_skb_seglen(const struct sk_buff *skb)
1367 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1370 /* Shifting pages past head area doesn't work */
1371 static int skb_can_shift(const struct sk_buff *skb)
1373 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1376 int tcp_skb_shift(struct sk_buff *to, struct sk_buff *from,
1377 int pcount, int shiftlen)
1379 /* TCP min gso_size is 8 bytes (TCP_MIN_GSO_SIZE)
1380 * Since TCP_SKB_CB(skb)->tcp_gso_segs is 16 bits, we need
1381 * to make sure not storing more than 65535 * 8 bytes per skb,
1382 * even if current MSS is bigger.
1384 if (unlikely(to->len + shiftlen >= 65535 * TCP_MIN_GSO_SIZE))
1386 if (unlikely(tcp_skb_pcount(to) + pcount > 65535))
1388 return skb_shift(to, from, shiftlen);
1391 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1394 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1395 struct tcp_sacktag_state *state,
1396 u32 start_seq, u32 end_seq,
1399 struct tcp_sock *tp = tcp_sk(sk);
1400 struct sk_buff *prev;
1406 /* Normally R but no L won't result in plain S */
1408 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1410 if (!skb_can_shift(skb))
1412 /* This frame is about to be dropped (was ACKed). */
1413 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1416 /* Can only happen with delayed DSACK + discard craziness */
1417 prev = skb_rb_prev(skb);
1421 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1424 if (!tcp_skb_can_collapse_to(prev))
1427 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1428 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1432 pcount = tcp_skb_pcount(skb);
1433 mss = tcp_skb_seglen(skb);
1435 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1436 * drop this restriction as unnecessary
1438 if (mss != tcp_skb_seglen(prev))
1441 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1443 /* CHECKME: This is non-MSS split case only?, this will
1444 * cause skipped skbs due to advancing loop btw, original
1445 * has that feature too
1447 if (tcp_skb_pcount(skb) <= 1)
1450 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1452 /* TODO: head merge to next could be attempted here
1453 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1454 * though it might not be worth of the additional hassle
1456 * ...we can probably just fallback to what was done
1457 * previously. We could try merging non-SACKed ones
1458 * as well but it probably isn't going to buy off
1459 * because later SACKs might again split them, and
1460 * it would make skb timestamp tracking considerably
1466 len = end_seq - TCP_SKB_CB(skb)->seq;
1468 BUG_ON(len > skb->len);
1470 /* MSS boundaries should be honoured or else pcount will
1471 * severely break even though it makes things bit trickier.
1472 * Optimize common case to avoid most of the divides
1474 mss = tcp_skb_mss(skb);
1476 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1477 * drop this restriction as unnecessary
1479 if (mss != tcp_skb_seglen(prev))
1484 } else if (len < mss) {
1492 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1493 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1496 if (!tcp_skb_shift(prev, skb, pcount, len))
1498 if (!tcp_shifted_skb(sk, prev, skb, state, pcount, len, mss, dup_sack))
1501 /* Hole filled allows collapsing with the next as well, this is very
1502 * useful when hole on every nth skb pattern happens
1504 skb = skb_rb_next(prev);
1508 if (!skb_can_shift(skb) ||
1509 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1510 (mss != tcp_skb_seglen(skb)))
1514 pcount = tcp_skb_pcount(skb);
1515 if (tcp_skb_shift(prev, skb, pcount, len))
1516 tcp_shifted_skb(sk, prev, skb, state, pcount,
1526 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1530 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1531 struct tcp_sack_block *next_dup,
1532 struct tcp_sacktag_state *state,
1533 u32 start_seq, u32 end_seq,
1536 struct tcp_sock *tp = tcp_sk(sk);
1537 struct sk_buff *tmp;
1539 skb_rbtree_walk_from(skb) {
1541 bool dup_sack = dup_sack_in;
1543 /* queue is in-order => we can short-circuit the walk early */
1544 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1548 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1549 in_sack = tcp_match_skb_to_sack(sk, skb,
1550 next_dup->start_seq,
1556 /* skb reference here is a bit tricky to get right, since
1557 * shifting can eat and free both this skb and the next,
1558 * so not even _safe variant of the loop is enough.
1561 tmp = tcp_shift_skb_data(sk, skb, state,
1562 start_seq, end_seq, dup_sack);
1571 in_sack = tcp_match_skb_to_sack(sk, skb,
1577 if (unlikely(in_sack < 0))
1581 TCP_SKB_CB(skb)->sacked =
1584 TCP_SKB_CB(skb)->sacked,
1585 TCP_SKB_CB(skb)->seq,
1586 TCP_SKB_CB(skb)->end_seq,
1588 tcp_skb_pcount(skb),
1589 tcp_skb_timestamp_us(skb));
1590 tcp_rate_skb_delivered(sk, skb, state->rate);
1591 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1592 list_del_init(&skb->tcp_tsorted_anchor);
1594 if (!before(TCP_SKB_CB(skb)->seq,
1595 tcp_highest_sack_seq(tp)))
1596 tcp_advance_highest_sack(sk, skb);
1602 static struct sk_buff *tcp_sacktag_bsearch(struct sock *sk, u32 seq)
1604 struct rb_node *parent, **p = &sk->tcp_rtx_queue.rb_node;
1605 struct sk_buff *skb;
1609 skb = rb_to_skb(parent);
1610 if (before(seq, TCP_SKB_CB(skb)->seq)) {
1611 p = &parent->rb_left;
1614 if (!before(seq, TCP_SKB_CB(skb)->end_seq)) {
1615 p = &parent->rb_right;
1623 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1626 if (skb && after(TCP_SKB_CB(skb)->seq, skip_to_seq))
1629 return tcp_sacktag_bsearch(sk, skip_to_seq);
1632 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1634 struct tcp_sack_block *next_dup,
1635 struct tcp_sacktag_state *state,
1641 if (before(next_dup->start_seq, skip_to_seq)) {
1642 skb = tcp_sacktag_skip(skb, sk, next_dup->start_seq);
1643 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1644 next_dup->start_seq, next_dup->end_seq,
1651 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1653 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1657 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1658 u32 prior_snd_una, struct tcp_sacktag_state *state)
1660 struct tcp_sock *tp = tcp_sk(sk);
1661 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1662 TCP_SKB_CB(ack_skb)->sacked);
1663 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1664 struct tcp_sack_block sp[TCP_NUM_SACKS];
1665 struct tcp_sack_block *cache;
1666 struct sk_buff *skb;
1667 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1669 bool found_dup_sack = false;
1671 int first_sack_index;
1674 state->reord = tp->snd_nxt;
1676 if (!tp->sacked_out)
1677 tcp_highest_sack_reset(sk);
1679 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1680 num_sacks, prior_snd_una);
1681 if (found_dup_sack) {
1682 state->flag |= FLAG_DSACKING_ACK;
1683 tp->delivered++; /* A spurious retransmission is delivered */
1686 /* Eliminate too old ACKs, but take into
1687 * account more or less fresh ones, they can
1688 * contain valid SACK info.
1690 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1693 if (!tp->packets_out)
1697 first_sack_index = 0;
1698 for (i = 0; i < num_sacks; i++) {
1699 bool dup_sack = !i && found_dup_sack;
1701 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1702 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1704 if (!tcp_is_sackblock_valid(tp, dup_sack,
1705 sp[used_sacks].start_seq,
1706 sp[used_sacks].end_seq)) {
1710 if (!tp->undo_marker)
1711 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1713 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1715 /* Don't count olds caused by ACK reordering */
1716 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1717 !after(sp[used_sacks].end_seq, tp->snd_una))
1719 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1722 NET_INC_STATS(sock_net(sk), mib_idx);
1724 first_sack_index = -1;
1728 /* Ignore very old stuff early */
1729 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1735 /* order SACK blocks to allow in order walk of the retrans queue */
1736 for (i = used_sacks - 1; i > 0; i--) {
1737 for (j = 0; j < i; j++) {
1738 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1739 swap(sp[j], sp[j + 1]);
1741 /* Track where the first SACK block goes to */
1742 if (j == first_sack_index)
1743 first_sack_index = j + 1;
1748 state->mss_now = tcp_current_mss(sk);
1752 if (!tp->sacked_out) {
1753 /* It's already past, so skip checking against it */
1754 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1756 cache = tp->recv_sack_cache;
1757 /* Skip empty blocks in at head of the cache */
1758 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1763 while (i < used_sacks) {
1764 u32 start_seq = sp[i].start_seq;
1765 u32 end_seq = sp[i].end_seq;
1766 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1767 struct tcp_sack_block *next_dup = NULL;
1769 if (found_dup_sack && ((i + 1) == first_sack_index))
1770 next_dup = &sp[i + 1];
1772 /* Skip too early cached blocks */
1773 while (tcp_sack_cache_ok(tp, cache) &&
1774 !before(start_seq, cache->end_seq))
1777 /* Can skip some work by looking recv_sack_cache? */
1778 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1779 after(end_seq, cache->start_seq)) {
1782 if (before(start_seq, cache->start_seq)) {
1783 skb = tcp_sacktag_skip(skb, sk, start_seq);
1784 skb = tcp_sacktag_walk(skb, sk, next_dup,
1791 /* Rest of the block already fully processed? */
1792 if (!after(end_seq, cache->end_seq))
1795 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1799 /* ...tail remains todo... */
1800 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1801 /* ...but better entrypoint exists! */
1802 skb = tcp_highest_sack(sk);
1809 skb = tcp_sacktag_skip(skb, sk, cache->end_seq);
1810 /* Check overlap against next cached too (past this one already) */
1815 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1816 skb = tcp_highest_sack(sk);
1820 skb = tcp_sacktag_skip(skb, sk, start_seq);
1823 skb = tcp_sacktag_walk(skb, sk, next_dup, state,
1824 start_seq, end_seq, dup_sack);
1830 /* Clear the head of the cache sack blocks so we can skip it next time */
1831 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1832 tp->recv_sack_cache[i].start_seq = 0;
1833 tp->recv_sack_cache[i].end_seq = 0;
1835 for (j = 0; j < used_sacks; j++)
1836 tp->recv_sack_cache[i++] = sp[j];
1838 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Loss || tp->undo_marker)
1839 tcp_check_sack_reordering(sk, state->reord, 0);
1841 tcp_verify_left_out(tp);
1844 #if FASTRETRANS_DEBUG > 0
1845 WARN_ON((int)tp->sacked_out < 0);
1846 WARN_ON((int)tp->lost_out < 0);
1847 WARN_ON((int)tp->retrans_out < 0);
1848 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1853 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1854 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1856 static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
1860 holes = max(tp->lost_out, 1U);
1861 holes = min(holes, tp->packets_out);
1863 if ((tp->sacked_out + holes) > tp->packets_out) {
1864 tp->sacked_out = tp->packets_out - holes;
1870 /* If we receive more dupacks than we expected counting segments
1871 * in assumption of absent reordering, interpret this as reordering.
1872 * The only another reason could be bug in receiver TCP.
1874 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1876 struct tcp_sock *tp = tcp_sk(sk);
1878 if (!tcp_limit_reno_sacked(tp))
1881 tp->reordering = min_t(u32, tp->packets_out + addend,
1882 sock_net(sk)->ipv4.sysctl_tcp_max_reordering);
1884 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRENOREORDER);
1887 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1889 static void tcp_add_reno_sack(struct sock *sk, int num_dupack)
1892 struct tcp_sock *tp = tcp_sk(sk);
1893 u32 prior_sacked = tp->sacked_out;
1896 tp->sacked_out += num_dupack;
1897 tcp_check_reno_reordering(sk, 0);
1898 delivered = tp->sacked_out - prior_sacked;
1900 tp->delivered += delivered;
1901 tcp_verify_left_out(tp);
1905 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1907 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1909 struct tcp_sock *tp = tcp_sk(sk);
1912 /* One ACK acked hole. The rest eat duplicate ACKs. */
1913 tp->delivered += max_t(int, acked - tp->sacked_out, 1);
1914 if (acked - 1 >= tp->sacked_out)
1917 tp->sacked_out -= acked - 1;
1919 tcp_check_reno_reordering(sk, acked);
1920 tcp_verify_left_out(tp);
1923 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1928 void tcp_clear_retrans(struct tcp_sock *tp)
1930 tp->retrans_out = 0;
1932 tp->undo_marker = 0;
1933 tp->undo_retrans = -1;
1937 static inline void tcp_init_undo(struct tcp_sock *tp)
1939 tp->undo_marker = tp->snd_una;
1940 /* Retransmission still in flight may cause DSACKs later. */
1941 tp->undo_retrans = tp->retrans_out ? : -1;
1944 static bool tcp_is_rack(const struct sock *sk)
1946 return sock_net(sk)->ipv4.sysctl_tcp_recovery & TCP_RACK_LOSS_DETECTION;
1949 /* If we detect SACK reneging, forget all SACK information
1950 * and reset tags completely, otherwise preserve SACKs. If receiver
1951 * dropped its ofo queue, we will know this due to reneging detection.
1953 static void tcp_timeout_mark_lost(struct sock *sk)
1955 struct tcp_sock *tp = tcp_sk(sk);
1956 struct sk_buff *skb, *head;
1957 bool is_reneg; /* is receiver reneging on SACKs? */
1959 head = tcp_rtx_queue_head(sk);
1960 is_reneg = head && (TCP_SKB_CB(head)->sacked & TCPCB_SACKED_ACKED);
1962 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
1964 /* Mark SACK reneging until we recover from this loss event. */
1965 tp->is_sack_reneg = 1;
1966 } else if (tcp_is_reno(tp)) {
1967 tcp_reset_reno_sack(tp);
1971 skb_rbtree_walk_from(skb) {
1973 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1974 else if (tcp_is_rack(sk) && skb != head &&
1975 tcp_rack_skb_timeout(tp, skb, 0) > 0)
1976 continue; /* Don't mark recently sent ones lost yet */
1977 tcp_mark_skb_lost(sk, skb);
1979 tcp_verify_left_out(tp);
1980 tcp_clear_all_retrans_hints(tp);
1983 /* Enter Loss state. */
1984 void tcp_enter_loss(struct sock *sk)
1986 const struct inet_connection_sock *icsk = inet_csk(sk);
1987 struct tcp_sock *tp = tcp_sk(sk);
1988 struct net *net = sock_net(sk);
1989 bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery;
1991 tcp_timeout_mark_lost(sk);
1993 /* Reduce ssthresh if it has not yet been made inside this window. */
1994 if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
1995 !after(tp->high_seq, tp->snd_una) ||
1996 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1997 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1998 tp->prior_cwnd = tp->snd_cwnd;
1999 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2000 tcp_ca_event(sk, CA_EVENT_LOSS);
2003 tp->snd_cwnd = tcp_packets_in_flight(tp) + 1;
2004 tp->snd_cwnd_cnt = 0;
2005 tp->snd_cwnd_stamp = tcp_jiffies32;
2007 /* Timeout in disordered state after receiving substantial DUPACKs
2008 * suggests that the degree of reordering is over-estimated.
2010 if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
2011 tp->sacked_out >= net->ipv4.sysctl_tcp_reordering)
2012 tp->reordering = min_t(unsigned int, tp->reordering,
2013 net->ipv4.sysctl_tcp_reordering);
2014 tcp_set_ca_state(sk, TCP_CA_Loss);
2015 tp->high_seq = tp->snd_nxt;
2016 tcp_ecn_queue_cwr(tp);
2018 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
2019 * loss recovery is underway except recurring timeout(s) on
2020 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
2022 tp->frto = net->ipv4.sysctl_tcp_frto &&
2023 (new_recovery || icsk->icsk_retransmits) &&
2024 !inet_csk(sk)->icsk_mtup.probe_size;
2027 /* If ACK arrived pointing to a remembered SACK, it means that our
2028 * remembered SACKs do not reflect real state of receiver i.e.
2029 * receiver _host_ is heavily congested (or buggy).
2031 * To avoid big spurious retransmission bursts due to transient SACK
2032 * scoreboard oddities that look like reneging, we give the receiver a
2033 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
2034 * restore sanity to the SACK scoreboard. If the apparent reneging
2035 * persists until this RTO then we'll clear the SACK scoreboard.
2037 static bool tcp_check_sack_reneging(struct sock *sk, int flag)
2039 if (flag & FLAG_SACK_RENEGING) {
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 /* People celebrate: "We love our President!" */
2377 static bool tcp_try_undo_recovery(struct sock *sk)
2379 struct tcp_sock *tp = tcp_sk(sk);
2381 if (tcp_may_undo(tp)) {
2384 /* Happy end! We did not retransmit anything
2385 * or our original transmission succeeded.
2387 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2388 tcp_undo_cwnd_reduction(sk, false);
2389 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2390 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2392 mib_idx = LINUX_MIB_TCPFULLUNDO;
2394 NET_INC_STATS(sock_net(sk), mib_idx);
2395 } else if (tp->rack.reo_wnd_persist) {
2396 tp->rack.reo_wnd_persist--;
2398 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2399 /* Hold old state until something *above* high_seq
2400 * is ACKed. For Reno it is MUST to prevent false
2401 * fast retransmits (RFC2582). SACK TCP is safe. */
2402 if (!tcp_any_retrans_done(sk))
2403 tp->retrans_stamp = 0;
2406 tcp_set_ca_state(sk, TCP_CA_Open);
2407 tp->is_sack_reneg = 0;
2411 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2412 static bool tcp_try_undo_dsack(struct sock *sk)
2414 struct tcp_sock *tp = tcp_sk(sk);
2416 if (tp->undo_marker && !tp->undo_retrans) {
2417 tp->rack.reo_wnd_persist = min(TCP_RACK_RECOVERY_THRESH,
2418 tp->rack.reo_wnd_persist + 1);
2419 DBGUNDO(sk, "D-SACK");
2420 tcp_undo_cwnd_reduction(sk, false);
2421 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2427 /* Undo during loss recovery after partial ACK or using F-RTO. */
2428 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2430 struct tcp_sock *tp = tcp_sk(sk);
2432 if (frto_undo || tcp_may_undo(tp)) {
2433 tcp_undo_cwnd_reduction(sk, true);
2435 DBGUNDO(sk, "partial loss");
2436 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2438 NET_INC_STATS(sock_net(sk),
2439 LINUX_MIB_TCPSPURIOUSRTOS);
2440 inet_csk(sk)->icsk_retransmits = 0;
2441 if (frto_undo || tcp_is_sack(tp)) {
2442 tcp_set_ca_state(sk, TCP_CA_Open);
2443 tp->is_sack_reneg = 0;
2450 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
2451 * It computes the number of packets to send (sndcnt) based on packets newly
2453 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2454 * cwnd reductions across a full RTT.
2455 * 2) Otherwise PRR uses packet conservation to send as much as delivered.
2456 * But when the retransmits are acked without further losses, PRR
2457 * slow starts cwnd up to ssthresh to speed up the recovery.
2459 static void tcp_init_cwnd_reduction(struct sock *sk)
2461 struct tcp_sock *tp = tcp_sk(sk);
2463 tp->high_seq = tp->snd_nxt;
2464 tp->tlp_high_seq = 0;
2465 tp->snd_cwnd_cnt = 0;
2466 tp->prior_cwnd = tp->snd_cwnd;
2467 tp->prr_delivered = 0;
2469 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2470 tcp_ecn_queue_cwr(tp);
2473 void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked, int flag)
2475 struct tcp_sock *tp = tcp_sk(sk);
2477 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2479 if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd))
2482 tp->prr_delivered += newly_acked_sacked;
2484 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2486 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2487 } else if ((flag & (FLAG_RETRANS_DATA_ACKED | FLAG_LOST_RETRANS)) ==
2488 FLAG_RETRANS_DATA_ACKED) {
2489 sndcnt = min_t(int, delta,
2490 max_t(int, tp->prr_delivered - tp->prr_out,
2491 newly_acked_sacked) + 1);
2493 sndcnt = min(delta, newly_acked_sacked);
2495 /* Force a fast retransmit upon entering fast recovery */
2496 sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1));
2497 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2500 static inline void tcp_end_cwnd_reduction(struct sock *sk)
2502 struct tcp_sock *tp = tcp_sk(sk);
2504 if (inet_csk(sk)->icsk_ca_ops->cong_control)
2507 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2508 if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH &&
2509 (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR || tp->undo_marker)) {
2510 tp->snd_cwnd = tp->snd_ssthresh;
2511 tp->snd_cwnd_stamp = tcp_jiffies32;
2513 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2516 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2517 void tcp_enter_cwr(struct sock *sk)
2519 struct tcp_sock *tp = tcp_sk(sk);
2521 tp->prior_ssthresh = 0;
2522 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2523 tp->undo_marker = 0;
2524 tcp_init_cwnd_reduction(sk);
2525 tcp_set_ca_state(sk, TCP_CA_CWR);
2528 EXPORT_SYMBOL(tcp_enter_cwr);
2530 static void tcp_try_keep_open(struct sock *sk)
2532 struct tcp_sock *tp = tcp_sk(sk);
2533 int state = TCP_CA_Open;
2535 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2536 state = TCP_CA_Disorder;
2538 if (inet_csk(sk)->icsk_ca_state != state) {
2539 tcp_set_ca_state(sk, state);
2540 tp->high_seq = tp->snd_nxt;
2544 static void tcp_try_to_open(struct sock *sk, int flag)
2546 struct tcp_sock *tp = tcp_sk(sk);
2548 tcp_verify_left_out(tp);
2550 if (!tcp_any_retrans_done(sk))
2551 tp->retrans_stamp = 0;
2553 if (flag & FLAG_ECE)
2556 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2557 tcp_try_keep_open(sk);
2561 static void tcp_mtup_probe_failed(struct sock *sk)
2563 struct inet_connection_sock *icsk = inet_csk(sk);
2565 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2566 icsk->icsk_mtup.probe_size = 0;
2567 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPFAIL);
2570 static void tcp_mtup_probe_success(struct sock *sk)
2572 struct tcp_sock *tp = tcp_sk(sk);
2573 struct inet_connection_sock *icsk = inet_csk(sk);
2575 /* FIXME: breaks with very large cwnd */
2576 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2577 tp->snd_cwnd = tp->snd_cwnd *
2578 tcp_mss_to_mtu(sk, tp->mss_cache) /
2579 icsk->icsk_mtup.probe_size;
2580 tp->snd_cwnd_cnt = 0;
2581 tp->snd_cwnd_stamp = tcp_jiffies32;
2582 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2584 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2585 icsk->icsk_mtup.probe_size = 0;
2586 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2587 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS);
2590 /* Do a simple retransmit without using the backoff mechanisms in
2591 * tcp_timer. This is used for path mtu discovery.
2592 * The socket is already locked here.
2594 void tcp_simple_retransmit(struct sock *sk)
2596 const struct inet_connection_sock *icsk = inet_csk(sk);
2597 struct tcp_sock *tp = tcp_sk(sk);
2598 struct sk_buff *skb;
2599 unsigned int mss = tcp_current_mss(sk);
2601 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2602 if (tcp_skb_seglen(skb) > mss &&
2603 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2604 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2605 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2606 tp->retrans_out -= tcp_skb_pcount(skb);
2608 tcp_skb_mark_lost_uncond_verify(tp, skb);
2612 tcp_clear_retrans_hints_partial(tp);
2617 if (tcp_is_reno(tp))
2618 tcp_limit_reno_sacked(tp);
2620 tcp_verify_left_out(tp);
2622 /* Don't muck with the congestion window here.
2623 * Reason is that we do not increase amount of _data_
2624 * in network, but units changed and effective
2625 * cwnd/ssthresh really reduced now.
2627 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2628 tp->high_seq = tp->snd_nxt;
2629 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2630 tp->prior_ssthresh = 0;
2631 tp->undo_marker = 0;
2632 tcp_set_ca_state(sk, TCP_CA_Loss);
2634 tcp_xmit_retransmit_queue(sk);
2636 EXPORT_SYMBOL(tcp_simple_retransmit);
2638 void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2640 struct tcp_sock *tp = tcp_sk(sk);
2643 if (tcp_is_reno(tp))
2644 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2646 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2648 NET_INC_STATS(sock_net(sk), mib_idx);
2650 tp->prior_ssthresh = 0;
2653 if (!tcp_in_cwnd_reduction(sk)) {
2655 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2656 tcp_init_cwnd_reduction(sk);
2658 tcp_set_ca_state(sk, TCP_CA_Recovery);
2661 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2662 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2664 static void tcp_process_loss(struct sock *sk, int flag, int num_dupack,
2667 struct tcp_sock *tp = tcp_sk(sk);
2668 bool recovered = !before(tp->snd_una, tp->high_seq);
2670 if ((flag & FLAG_SND_UNA_ADVANCED || rcu_access_pointer(tp->fastopen_rsk)) &&
2671 tcp_try_undo_loss(sk, false))
2674 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2675 /* Step 3.b. A timeout is spurious if not all data are
2676 * lost, i.e., never-retransmitted data are (s)acked.
2678 if ((flag & FLAG_ORIG_SACK_ACKED) &&
2679 tcp_try_undo_loss(sk, true))
2682 if (after(tp->snd_nxt, tp->high_seq)) {
2683 if (flag & FLAG_DATA_SACKED || num_dupack)
2684 tp->frto = 0; /* Step 3.a. loss was real */
2685 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2686 tp->high_seq = tp->snd_nxt;
2687 /* Step 2.b. Try send new data (but deferred until cwnd
2688 * is updated in tcp_ack()). Otherwise fall back to
2689 * the conventional recovery.
2691 if (!tcp_write_queue_empty(sk) &&
2692 after(tcp_wnd_end(tp), tp->snd_nxt)) {
2693 *rexmit = REXMIT_NEW;
2701 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2702 tcp_try_undo_recovery(sk);
2705 if (tcp_is_reno(tp)) {
2706 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2707 * delivered. Lower inflight to clock out (re)tranmissions.
2709 if (after(tp->snd_nxt, tp->high_seq) && num_dupack)
2710 tcp_add_reno_sack(sk, num_dupack);
2711 else if (flag & FLAG_SND_UNA_ADVANCED)
2712 tcp_reset_reno_sack(tp);
2714 *rexmit = REXMIT_LOST;
2717 /* Undo during fast recovery after partial ACK. */
2718 static bool tcp_try_undo_partial(struct sock *sk, u32 prior_snd_una)
2720 struct tcp_sock *tp = tcp_sk(sk);
2722 if (tp->undo_marker && tcp_packet_delayed(tp)) {
2723 /* Plain luck! Hole if filled with delayed
2724 * packet, rather than with a retransmit. Check reordering.
2726 tcp_check_sack_reordering(sk, prior_snd_una, 1);
2728 /* We are getting evidence that the reordering degree is higher
2729 * than we realized. If there are no retransmits out then we
2730 * can undo. Otherwise we clock out new packets but do not
2731 * mark more packets lost or retransmit more.
2733 if (tp->retrans_out)
2736 if (!tcp_any_retrans_done(sk))
2737 tp->retrans_stamp = 0;
2739 DBGUNDO(sk, "partial recovery");
2740 tcp_undo_cwnd_reduction(sk, true);
2741 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2742 tcp_try_keep_open(sk);
2748 static void tcp_identify_packet_loss(struct sock *sk, int *ack_flag)
2750 struct tcp_sock *tp = tcp_sk(sk);
2752 if (tcp_rtx_queue_empty(sk))
2755 if (unlikely(tcp_is_reno(tp))) {
2756 tcp_newreno_mark_lost(sk, *ack_flag & FLAG_SND_UNA_ADVANCED);
2757 } else if (tcp_is_rack(sk)) {
2758 u32 prior_retrans = tp->retrans_out;
2760 tcp_rack_mark_lost(sk);
2761 if (prior_retrans > tp->retrans_out)
2762 *ack_flag |= FLAG_LOST_RETRANS;
2766 static bool tcp_force_fast_retransmit(struct sock *sk)
2768 struct tcp_sock *tp = tcp_sk(sk);
2770 return after(tcp_highest_sack_seq(tp),
2771 tp->snd_una + tp->reordering * tp->mss_cache);
2774 /* Process an event, which can update packets-in-flight not trivially.
2775 * Main goal of this function is to calculate new estimate for left_out,
2776 * taking into account both packets sitting in receiver's buffer and
2777 * packets lost by network.
2779 * Besides that it updates the congestion state when packet loss or ECN
2780 * is detected. But it does not reduce the cwnd, it is done by the
2781 * congestion control later.
2783 * It does _not_ decide what to send, it is made in function
2784 * tcp_xmit_retransmit_queue().
2786 static void tcp_fastretrans_alert(struct sock *sk, const u32 prior_snd_una,
2787 int num_dupack, int *ack_flag, int *rexmit)
2789 struct inet_connection_sock *icsk = inet_csk(sk);
2790 struct tcp_sock *tp = tcp_sk(sk);
2791 int fast_rexmit = 0, flag = *ack_flag;
2792 bool do_lost = num_dupack || ((flag & FLAG_DATA_SACKED) &&
2793 tcp_force_fast_retransmit(sk));
2795 if (!tp->packets_out && tp->sacked_out)
2798 /* Now state machine starts.
2799 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2800 if (flag & FLAG_ECE)
2801 tp->prior_ssthresh = 0;
2803 /* B. In all the states check for reneging SACKs. */
2804 if (tcp_check_sack_reneging(sk, flag))
2807 /* C. Check consistency of the current state. */
2808 tcp_verify_left_out(tp);
2810 /* D. Check state exit conditions. State can be terminated
2811 * when high_seq is ACKed. */
2812 if (icsk->icsk_ca_state == TCP_CA_Open) {
2813 WARN_ON(tp->retrans_out != 0);
2814 tp->retrans_stamp = 0;
2815 } else if (!before(tp->snd_una, tp->high_seq)) {
2816 switch (icsk->icsk_ca_state) {
2818 /* CWR is to be held something *above* high_seq
2819 * is ACKed for CWR bit to reach receiver. */
2820 if (tp->snd_una != tp->high_seq) {
2821 tcp_end_cwnd_reduction(sk);
2822 tcp_set_ca_state(sk, TCP_CA_Open);
2826 case TCP_CA_Recovery:
2827 if (tcp_is_reno(tp))
2828 tcp_reset_reno_sack(tp);
2829 if (tcp_try_undo_recovery(sk))
2831 tcp_end_cwnd_reduction(sk);
2836 /* E. Process state. */
2837 switch (icsk->icsk_ca_state) {
2838 case TCP_CA_Recovery:
2839 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2840 if (tcp_is_reno(tp))
2841 tcp_add_reno_sack(sk, num_dupack);
2843 if (tcp_try_undo_partial(sk, prior_snd_una))
2845 /* Partial ACK arrived. Force fast retransmit. */
2846 do_lost = tcp_is_reno(tp) ||
2847 tcp_force_fast_retransmit(sk);
2849 if (tcp_try_undo_dsack(sk)) {
2850 tcp_try_keep_open(sk);
2853 tcp_identify_packet_loss(sk, ack_flag);
2856 tcp_process_loss(sk, flag, num_dupack, rexmit);
2857 tcp_identify_packet_loss(sk, ack_flag);
2858 if (!(icsk->icsk_ca_state == TCP_CA_Open ||
2859 (*ack_flag & FLAG_LOST_RETRANS)))
2861 /* Change state if cwnd is undone or retransmits are lost */
2864 if (tcp_is_reno(tp)) {
2865 if (flag & FLAG_SND_UNA_ADVANCED)
2866 tcp_reset_reno_sack(tp);
2867 tcp_add_reno_sack(sk, num_dupack);
2870 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
2871 tcp_try_undo_dsack(sk);
2873 tcp_identify_packet_loss(sk, ack_flag);
2874 if (!tcp_time_to_recover(sk, flag)) {
2875 tcp_try_to_open(sk, flag);
2879 /* MTU probe failure: don't reduce cwnd */
2880 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2881 icsk->icsk_mtup.probe_size &&
2882 tp->snd_una == tp->mtu_probe.probe_seq_start) {
2883 tcp_mtup_probe_failed(sk);
2884 /* Restores the reduction we did in tcp_mtup_probe() */
2886 tcp_simple_retransmit(sk);
2890 /* Otherwise enter Recovery state */
2891 tcp_enter_recovery(sk, (flag & FLAG_ECE));
2895 if (!tcp_is_rack(sk) && do_lost)
2896 tcp_update_scoreboard(sk, fast_rexmit);
2897 *rexmit = REXMIT_LOST;
2900 static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us, const int flag)
2902 u32 wlen = sock_net(sk)->ipv4.sysctl_tcp_min_rtt_wlen * HZ;
2903 struct tcp_sock *tp = tcp_sk(sk);
2905 if ((flag & FLAG_ACK_MAYBE_DELAYED) && rtt_us > tcp_min_rtt(tp)) {
2906 /* If the remote keeps returning delayed ACKs, eventually
2907 * the min filter would pick it up and overestimate the
2908 * prop. delay when it expires. Skip suspected delayed ACKs.
2912 minmax_running_min(&tp->rtt_min, wlen, tcp_jiffies32,
2913 rtt_us ? : jiffies_to_usecs(1));
2916 static bool tcp_ack_update_rtt(struct sock *sk, const int flag,
2917 long seq_rtt_us, long sack_rtt_us,
2918 long ca_rtt_us, struct rate_sample *rs)
2920 const struct tcp_sock *tp = tcp_sk(sk);
2922 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2923 * broken middle-boxes or peers may corrupt TS-ECR fields. But
2924 * Karn's algorithm forbids taking RTT if some retransmitted data
2925 * is acked (RFC6298).
2928 seq_rtt_us = sack_rtt_us;
2930 /* RTTM Rule: A TSecr value received in a segment is used to
2931 * update the averaged RTT measurement only if the segment
2932 * acknowledges some new data, i.e., only if it advances the
2933 * left edge of the send window.
2934 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2936 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2937 flag & FLAG_ACKED) {
2938 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr;
2940 if (likely(delta < INT_MAX / (USEC_PER_SEC / TCP_TS_HZ))) {
2941 seq_rtt_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
2942 ca_rtt_us = seq_rtt_us;
2945 rs->rtt_us = ca_rtt_us; /* RTT of last (S)ACKed packet (or -1) */
2949 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
2950 * always taken together with ACK, SACK, or TS-opts. Any negative
2951 * values will be skipped with the seq_rtt_us < 0 check above.
2953 tcp_update_rtt_min(sk, ca_rtt_us, flag);
2954 tcp_rtt_estimator(sk, seq_rtt_us);
2957 /* RFC6298: only reset backoff on valid RTT measurement. */
2958 inet_csk(sk)->icsk_backoff = 0;
2962 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
2963 void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req)
2965 struct rate_sample rs;
2968 if (req && !req->num_retrans && tcp_rsk(req)->snt_synack)
2969 rtt_us = tcp_stamp_us_delta(tcp_clock_us(), tcp_rsk(req)->snt_synack);
2971 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us, &rs);
2975 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
2977 const struct inet_connection_sock *icsk = inet_csk(sk);
2979 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
2980 tcp_sk(sk)->snd_cwnd_stamp = tcp_jiffies32;
2983 /* Restart timer after forward progress on connection.
2984 * RFC2988 recommends to restart timer to now+rto.
2986 void tcp_rearm_rto(struct sock *sk)
2988 const struct inet_connection_sock *icsk = inet_csk(sk);
2989 struct tcp_sock *tp = tcp_sk(sk);
2991 /* If the retrans timer is currently being used by Fast Open
2992 * for SYN-ACK retrans purpose, stay put.
2994 if (rcu_access_pointer(tp->fastopen_rsk))
2997 if (!tp->packets_out) {
2998 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
3000 u32 rto = inet_csk(sk)->icsk_rto;
3001 /* Offset the time elapsed after installing regular RTO */
3002 if (icsk->icsk_pending == ICSK_TIME_REO_TIMEOUT ||
3003 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
3004 s64 delta_us = tcp_rto_delta_us(sk);
3005 /* delta_us may not be positive if the socket is locked
3006 * when the retrans timer fires and is rescheduled.
3008 rto = usecs_to_jiffies(max_t(int, delta_us, 1));
3010 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
3011 TCP_RTO_MAX, tcp_rtx_queue_head(sk));
3015 /* Try to schedule a loss probe; if that doesn't work, then schedule an RTO. */
3016 static void tcp_set_xmit_timer(struct sock *sk)
3018 if (!tcp_schedule_loss_probe(sk, true))
3022 /* If we get here, the whole TSO packet has not been acked. */
3023 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3025 struct tcp_sock *tp = tcp_sk(sk);
3028 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3030 packets_acked = tcp_skb_pcount(skb);
3031 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3033 packets_acked -= tcp_skb_pcount(skb);
3035 if (packets_acked) {
3036 BUG_ON(tcp_skb_pcount(skb) == 0);
3037 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3040 return packets_acked;
3043 static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
3046 const struct skb_shared_info *shinfo;
3048 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3049 if (likely(!TCP_SKB_CB(skb)->txstamp_ack))
3052 shinfo = skb_shinfo(skb);
3053 if (!before(shinfo->tskey, prior_snd_una) &&
3054 before(shinfo->tskey, tcp_sk(sk)->snd_una)) {
3055 tcp_skb_tsorted_save(skb) {
3056 __skb_tstamp_tx(skb, NULL, sk, SCM_TSTAMP_ACK);
3057 } tcp_skb_tsorted_restore(skb);
3061 /* Remove acknowledged frames from the retransmission queue. If our packet
3062 * is before the ack sequence we can discard it as it's confirmed to have
3063 * arrived at the other end.
3065 static int tcp_clean_rtx_queue(struct sock *sk, u32 prior_fack,
3067 struct tcp_sacktag_state *sack)
3069 const struct inet_connection_sock *icsk = inet_csk(sk);
3070 u64 first_ackt, last_ackt;
3071 struct tcp_sock *tp = tcp_sk(sk);
3072 u32 prior_sacked = tp->sacked_out;
3073 u32 reord = tp->snd_nxt; /* lowest acked un-retx un-sacked seq */
3074 struct sk_buff *skb, *next;
3075 bool fully_acked = true;
3076 long sack_rtt_us = -1L;
3077 long seq_rtt_us = -1L;
3078 long ca_rtt_us = -1L;
3080 u32 last_in_flight = 0;
3086 for (skb = skb_rb_first(&sk->tcp_rtx_queue); skb; skb = next) {
3087 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3088 const u32 start_seq = scb->seq;
3089 u8 sacked = scb->sacked;
3092 tcp_ack_tstamp(sk, skb, prior_snd_una);
3094 /* Determine how many packets and what bytes were acked, tso and else */
3095 if (after(scb->end_seq, tp->snd_una)) {
3096 if (tcp_skb_pcount(skb) == 1 ||
3097 !after(tp->snd_una, scb->seq))
3100 acked_pcount = tcp_tso_acked(sk, skb);
3103 fully_acked = false;
3105 acked_pcount = tcp_skb_pcount(skb);
3108 if (unlikely(sacked & TCPCB_RETRANS)) {
3109 if (sacked & TCPCB_SACKED_RETRANS)
3110 tp->retrans_out -= acked_pcount;
3111 flag |= FLAG_RETRANS_DATA_ACKED;
3112 } else if (!(sacked & TCPCB_SACKED_ACKED)) {
3113 last_ackt = tcp_skb_timestamp_us(skb);
3114 WARN_ON_ONCE(last_ackt == 0);
3116 first_ackt = last_ackt;
3118 last_in_flight = TCP_SKB_CB(skb)->tx.in_flight;
3119 if (before(start_seq, reord))
3121 if (!after(scb->end_seq, tp->high_seq))
3122 flag |= FLAG_ORIG_SACK_ACKED;
3125 if (sacked & TCPCB_SACKED_ACKED) {
3126 tp->sacked_out -= acked_pcount;
3127 } else if (tcp_is_sack(tp)) {
3128 tp->delivered += acked_pcount;
3129 if (!tcp_skb_spurious_retrans(tp, skb))
3130 tcp_rack_advance(tp, sacked, scb->end_seq,
3131 tcp_skb_timestamp_us(skb));
3133 if (sacked & TCPCB_LOST)
3134 tp->lost_out -= acked_pcount;
3136 tp->packets_out -= acked_pcount;
3137 pkts_acked += acked_pcount;
3138 tcp_rate_skb_delivered(sk, skb, sack->rate);
3140 /* Initial outgoing SYN's get put onto the write_queue
3141 * just like anything else we transmit. It is not
3142 * true data, and if we misinform our callers that
3143 * this ACK acks real data, we will erroneously exit
3144 * connection startup slow start one packet too
3145 * quickly. This is severely frowned upon behavior.
3147 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
3148 flag |= FLAG_DATA_ACKED;
3150 flag |= FLAG_SYN_ACKED;
3151 tp->retrans_stamp = 0;
3157 next = skb_rb_next(skb);
3158 if (unlikely(skb == tp->retransmit_skb_hint))
3159 tp->retransmit_skb_hint = NULL;
3160 if (unlikely(skb == tp->lost_skb_hint))
3161 tp->lost_skb_hint = NULL;
3162 tcp_rtx_queue_unlink_and_free(skb, sk);
3166 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
3168 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3169 tp->snd_up = tp->snd_una;
3171 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3172 flag |= FLAG_SACK_RENEGING;
3174 if (likely(first_ackt) && !(flag & FLAG_RETRANS_DATA_ACKED)) {
3175 seq_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, first_ackt);
3176 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, last_ackt);
3178 if (pkts_acked == 1 && last_in_flight < tp->mss_cache &&
3179 last_in_flight && !prior_sacked && fully_acked &&
3180 sack->rate->prior_delivered + 1 == tp->delivered &&
3181 !(flag & (FLAG_CA_ALERT | FLAG_SYN_ACKED))) {
3182 /* Conservatively mark a delayed ACK. It's typically
3183 * from a lone runt packet over the round trip to
3184 * a receiver w/o out-of-order or CE events.
3186 flag |= FLAG_ACK_MAYBE_DELAYED;
3189 if (sack->first_sackt) {
3190 sack_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->first_sackt);
3191 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->last_sackt);
3193 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us,
3194 ca_rtt_us, sack->rate);
3196 if (flag & FLAG_ACKED) {
3197 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3198 if (unlikely(icsk->icsk_mtup.probe_size &&
3199 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3200 tcp_mtup_probe_success(sk);
3203 if (tcp_is_reno(tp)) {
3204 tcp_remove_reno_sacks(sk, pkts_acked);
3206 /* If any of the cumulatively ACKed segments was
3207 * retransmitted, non-SACK case cannot confirm that
3208 * progress was due to original transmission due to
3209 * lack of TCPCB_SACKED_ACKED bits even if some of
3210 * the packets may have been never retransmitted.
3212 if (flag & FLAG_RETRANS_DATA_ACKED)
3213 flag &= ~FLAG_ORIG_SACK_ACKED;
3217 /* Non-retransmitted hole got filled? That's reordering */
3218 if (before(reord, prior_fack))
3219 tcp_check_sack_reordering(sk, reord, 0);
3221 delta = prior_sacked - tp->sacked_out;
3222 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3224 } else if (skb && rtt_update && sack_rtt_us >= 0 &&
3225 sack_rtt_us > tcp_stamp_us_delta(tp->tcp_mstamp,
3226 tcp_skb_timestamp_us(skb))) {
3227 /* Do not re-arm RTO if the sack RTT is measured from data sent
3228 * after when the head was last (re)transmitted. Otherwise the
3229 * timeout may continue to extend in loss recovery.
3231 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3234 if (icsk->icsk_ca_ops->pkts_acked) {
3235 struct ack_sample sample = { .pkts_acked = pkts_acked,
3236 .rtt_us = sack->rate->rtt_us,
3237 .in_flight = last_in_flight };
3239 icsk->icsk_ca_ops->pkts_acked(sk, &sample);
3242 #if FASTRETRANS_DEBUG > 0
3243 WARN_ON((int)tp->sacked_out < 0);
3244 WARN_ON((int)tp->lost_out < 0);
3245 WARN_ON((int)tp->retrans_out < 0);
3246 if (!tp->packets_out && tcp_is_sack(tp)) {
3247 icsk = inet_csk(sk);
3249 pr_debug("Leak l=%u %d\n",
3250 tp->lost_out, icsk->icsk_ca_state);
3253 if (tp->sacked_out) {
3254 pr_debug("Leak s=%u %d\n",
3255 tp->sacked_out, icsk->icsk_ca_state);
3258 if (tp->retrans_out) {
3259 pr_debug("Leak r=%u %d\n",
3260 tp->retrans_out, icsk->icsk_ca_state);
3261 tp->retrans_out = 0;
3268 static void tcp_ack_probe(struct sock *sk)
3270 struct inet_connection_sock *icsk = inet_csk(sk);
3271 struct sk_buff *head = tcp_send_head(sk);
3272 const struct tcp_sock *tp = tcp_sk(sk);
3274 /* Was it a usable window open? */
3277 if (!after(TCP_SKB_CB(head)->end_seq, tcp_wnd_end(tp))) {
3278 icsk->icsk_backoff = 0;
3279 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3280 /* Socket must be waked up by subsequent tcp_data_snd_check().
3281 * This function is not for random using!
3284 unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX);
3286 tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3287 when, TCP_RTO_MAX, NULL);
3291 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3293 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3294 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3297 /* Decide wheather to run the increase function of congestion control. */
3298 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3300 /* If reordering is high then always grow cwnd whenever data is
3301 * delivered regardless of its ordering. Otherwise stay conservative
3302 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3303 * new SACK or ECE mark may first advance cwnd here and later reduce
3304 * cwnd in tcp_fastretrans_alert() based on more states.
3306 if (tcp_sk(sk)->reordering > sock_net(sk)->ipv4.sysctl_tcp_reordering)
3307 return flag & FLAG_FORWARD_PROGRESS;
3309 return flag & FLAG_DATA_ACKED;
3312 /* The "ultimate" congestion control function that aims to replace the rigid
3313 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction).
3314 * It's called toward the end of processing an ACK with precise rate
3315 * information. All transmission or retransmission are delayed afterwards.
3317 static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked,
3318 int flag, const struct rate_sample *rs)
3320 const struct inet_connection_sock *icsk = inet_csk(sk);
3322 if (icsk->icsk_ca_ops->cong_control) {
3323 icsk->icsk_ca_ops->cong_control(sk, rs);
3327 if (tcp_in_cwnd_reduction(sk)) {
3328 /* Reduce cwnd if state mandates */
3329 tcp_cwnd_reduction(sk, acked_sacked, flag);
3330 } else if (tcp_may_raise_cwnd(sk, flag)) {
3331 /* Advance cwnd if state allows */
3332 tcp_cong_avoid(sk, ack, acked_sacked);
3334 tcp_update_pacing_rate(sk);
3337 /* Check that window update is acceptable.
3338 * The function assumes that snd_una<=ack<=snd_next.
3340 static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3341 const u32 ack, const u32 ack_seq,
3344 return after(ack, tp->snd_una) ||
3345 after(ack_seq, tp->snd_wl1) ||
3346 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3349 /* If we update tp->snd_una, also update tp->bytes_acked */
3350 static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack)
3352 u32 delta = ack - tp->snd_una;
3354 sock_owned_by_me((struct sock *)tp);
3355 tp->bytes_acked += delta;
3359 /* If we update tp->rcv_nxt, also update tp->bytes_received */
3360 static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq)
3362 u32 delta = seq - tp->rcv_nxt;
3364 sock_owned_by_me((struct sock *)tp);
3365 tp->bytes_received += delta;
3366 WRITE_ONCE(tp->rcv_nxt, seq);
3369 /* Update our send window.
3371 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3372 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3374 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3377 struct tcp_sock *tp = tcp_sk(sk);
3379 u32 nwin = ntohs(tcp_hdr(skb)->window);
3381 if (likely(!tcp_hdr(skb)->syn))
3382 nwin <<= tp->rx_opt.snd_wscale;
3384 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3385 flag |= FLAG_WIN_UPDATE;
3386 tcp_update_wl(tp, ack_seq);
3388 if (tp->snd_wnd != nwin) {
3391 /* Note, it is the only place, where
3392 * fast path is recovered for sending TCP.
3395 tcp_fast_path_check(sk);
3397 if (!tcp_write_queue_empty(sk))
3398 tcp_slow_start_after_idle_check(sk);
3400 if (nwin > tp->max_window) {
3401 tp->max_window = nwin;
3402 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3407 tcp_snd_una_update(tp, ack);
3412 static bool __tcp_oow_rate_limited(struct net *net, int mib_idx,
3413 u32 *last_oow_ack_time)
3415 if (*last_oow_ack_time) {
3416 s32 elapsed = (s32)(tcp_jiffies32 - *last_oow_ack_time);
3418 if (0 <= elapsed && elapsed < net->ipv4.sysctl_tcp_invalid_ratelimit) {
3419 NET_INC_STATS(net, mib_idx);
3420 return true; /* rate-limited: don't send yet! */
3424 *last_oow_ack_time = tcp_jiffies32;
3426 return false; /* not rate-limited: go ahead, send dupack now! */
3429 /* Return true if we're currently rate-limiting out-of-window ACKs and
3430 * thus shouldn't send a dupack right now. We rate-limit dupacks in
3431 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3432 * attacks that send repeated SYNs or ACKs for the same connection. To
3433 * do this, we do not send a duplicate SYNACK or ACK if the remote
3434 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3436 bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb,
3437 int mib_idx, u32 *last_oow_ack_time)
3439 /* Data packets without SYNs are not likely part of an ACK loop. */
3440 if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) &&
3444 return __tcp_oow_rate_limited(net, mib_idx, last_oow_ack_time);
3447 /* RFC 5961 7 [ACK Throttling] */
3448 static void tcp_send_challenge_ack(struct sock *sk, const struct sk_buff *skb)
3450 /* unprotected vars, we dont care of overwrites */
3451 static u32 challenge_timestamp;
3452 static unsigned int challenge_count;
3453 struct tcp_sock *tp = tcp_sk(sk);
3454 struct net *net = sock_net(sk);
3457 /* First check our per-socket dupack rate limit. */
3458 if (__tcp_oow_rate_limited(net,
3459 LINUX_MIB_TCPACKSKIPPEDCHALLENGE,
3460 &tp->last_oow_ack_time))
3463 /* Then check host-wide RFC 5961 rate limit. */
3465 if (now != challenge_timestamp) {
3466 u32 ack_limit = net->ipv4.sysctl_tcp_challenge_ack_limit;
3467 u32 half = (ack_limit + 1) >> 1;
3469 challenge_timestamp = now;
3470 WRITE_ONCE(challenge_count, half + prandom_u32_max(ack_limit));
3472 count = READ_ONCE(challenge_count);
3474 WRITE_ONCE(challenge_count, count - 1);
3475 NET_INC_STATS(net, LINUX_MIB_TCPCHALLENGEACK);
3480 static void tcp_store_ts_recent(struct tcp_sock *tp)
3482 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3483 tp->rx_opt.ts_recent_stamp = ktime_get_seconds();
3486 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3488 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3489 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3490 * extra check below makes sure this can only happen
3491 * for pure ACK frames. -DaveM
3493 * Not only, also it occurs for expired timestamps.
3496 if (tcp_paws_check(&tp->rx_opt, 0))
3497 tcp_store_ts_recent(tp);
3501 /* This routine deals with acks during a TLP episode.
3502 * We mark the end of a TLP episode on receiving TLP dupack or when
3503 * ack is after tlp_high_seq.
3504 * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3506 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3508 struct tcp_sock *tp = tcp_sk(sk);
3510 if (before(ack, tp->tlp_high_seq))
3513 if (flag & FLAG_DSACKING_ACK) {
3514 /* This DSACK means original and TLP probe arrived; no loss */
3515 tp->tlp_high_seq = 0;
3516 } else if (after(ack, tp->tlp_high_seq)) {
3517 /* ACK advances: there was a loss, so reduce cwnd. Reset
3518 * tlp_high_seq in tcp_init_cwnd_reduction()
3520 tcp_init_cwnd_reduction(sk);
3521 tcp_set_ca_state(sk, TCP_CA_CWR);
3522 tcp_end_cwnd_reduction(sk);
3523 tcp_try_keep_open(sk);
3524 NET_INC_STATS(sock_net(sk),
3525 LINUX_MIB_TCPLOSSPROBERECOVERY);
3526 } else if (!(flag & (FLAG_SND_UNA_ADVANCED |
3527 FLAG_NOT_DUP | FLAG_DATA_SACKED))) {
3528 /* Pure dupack: original and TLP probe arrived; no loss */
3529 tp->tlp_high_seq = 0;
3533 static inline void tcp_in_ack_event(struct sock *sk, u32 flags)
3535 const struct inet_connection_sock *icsk = inet_csk(sk);
3537 if (icsk->icsk_ca_ops->in_ack_event)
3538 icsk->icsk_ca_ops->in_ack_event(sk, flags);
3541 /* Congestion control has updated the cwnd already. So if we're in
3542 * loss recovery then now we do any new sends (for FRTO) or
3543 * retransmits (for CA_Loss or CA_recovery) that make sense.
3545 static void tcp_xmit_recovery(struct sock *sk, int rexmit)
3547 struct tcp_sock *tp = tcp_sk(sk);
3549 if (rexmit == REXMIT_NONE || sk->sk_state == TCP_SYN_SENT)
3552 if (unlikely(rexmit == 2)) {
3553 __tcp_push_pending_frames(sk, tcp_current_mss(sk),
3555 if (after(tp->snd_nxt, tp->high_seq))
3559 tcp_xmit_retransmit_queue(sk);
3562 /* Returns the number of packets newly acked or sacked by the current ACK */
3563 static u32 tcp_newly_delivered(struct sock *sk, u32 prior_delivered, int flag)
3565 const struct net *net = sock_net(sk);
3566 struct tcp_sock *tp = tcp_sk(sk);
3569 delivered = tp->delivered - prior_delivered;
3570 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVERED, delivered);
3571 if (flag & FLAG_ECE) {
3572 tp->delivered_ce += delivered;
3573 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVEREDCE, delivered);
3578 /* This routine deals with incoming acks, but not outgoing ones. */
3579 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3581 struct inet_connection_sock *icsk = inet_csk(sk);
3582 struct tcp_sock *tp = tcp_sk(sk);
3583 struct tcp_sacktag_state sack_state;
3584 struct rate_sample rs = { .prior_delivered = 0 };
3585 u32 prior_snd_una = tp->snd_una;
3586 bool is_sack_reneg = tp->is_sack_reneg;
3587 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3588 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3590 int prior_packets = tp->packets_out;
3591 u32 delivered = tp->delivered;
3592 u32 lost = tp->lost;
3593 int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */
3596 sack_state.first_sackt = 0;
3597 sack_state.rate = &rs;
3599 /* We very likely will need to access rtx queue. */
3600 prefetch(sk->tcp_rtx_queue.rb_node);
3602 /* If the ack is older than previous acks
3603 * then we can probably ignore it.
3605 if (before(ack, prior_snd_una)) {
3606 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3607 if (before(ack, prior_snd_una - tp->max_window)) {
3608 if (!(flag & FLAG_NO_CHALLENGE_ACK))
3609 tcp_send_challenge_ack(sk, skb);
3615 /* If the ack includes data we haven't sent yet, discard
3616 * this segment (RFC793 Section 3.9).
3618 if (after(ack, tp->snd_nxt))
3621 if (after(ack, prior_snd_una)) {
3622 flag |= FLAG_SND_UNA_ADVANCED;
3623 icsk->icsk_retransmits = 0;
3625 #if IS_ENABLED(CONFIG_TLS_DEVICE)
3626 if (static_branch_unlikely(&clean_acked_data_enabled.key))
3627 if (icsk->icsk_clean_acked)
3628 icsk->icsk_clean_acked(sk, ack);
3632 prior_fack = tcp_is_sack(tp) ? tcp_highest_sack_seq(tp) : tp->snd_una;
3633 rs.prior_in_flight = tcp_packets_in_flight(tp);
3635 /* ts_recent update must be made after we are sure that the packet
3638 if (flag & FLAG_UPDATE_TS_RECENT)
3639 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3641 if ((flag & (FLAG_SLOWPATH | FLAG_SND_UNA_ADVANCED)) ==
3642 FLAG_SND_UNA_ADVANCED) {
3643 /* Window is constant, pure forward advance.
3644 * No more checks are required.
3645 * Note, we use the fact that SND.UNA>=SND.WL2.
3647 tcp_update_wl(tp, ack_seq);
3648 tcp_snd_una_update(tp, ack);
3649 flag |= FLAG_WIN_UPDATE;
3651 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
3653 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPACKS);
3655 u32 ack_ev_flags = CA_ACK_SLOWPATH;
3657 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3660 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3662 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3664 if (TCP_SKB_CB(skb)->sacked)
3665 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3668 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) {
3670 ack_ev_flags |= CA_ACK_ECE;
3673 if (flag & FLAG_WIN_UPDATE)
3674 ack_ev_flags |= CA_ACK_WIN_UPDATE;
3676 tcp_in_ack_event(sk, ack_ev_flags);
3679 /* We passed data and got it acked, remove any soft error
3680 * log. Something worked...
3682 sk->sk_err_soft = 0;
3683 icsk->icsk_probes_out = 0;
3684 tp->rcv_tstamp = tcp_jiffies32;
3688 /* See if we can take anything off of the retransmit queue. */
3689 flag |= tcp_clean_rtx_queue(sk, prior_fack, prior_snd_una, &sack_state);
3691 tcp_rack_update_reo_wnd(sk, &rs);
3693 if (tp->tlp_high_seq)
3694 tcp_process_tlp_ack(sk, ack, flag);
3695 /* If needed, reset TLP/RTO timer; RACK may later override this. */
3696 if (flag & FLAG_SET_XMIT_TIMER)
3697 tcp_set_xmit_timer(sk);
3699 if (tcp_ack_is_dubious(sk, flag)) {
3700 if (!(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP))) {
3702 /* Consider if pure acks were aggregated in tcp_add_backlog() */
3703 if (!(flag & FLAG_DATA))
3704 num_dupack = max_t(u16, 1, skb_shinfo(skb)->gso_segs);
3706 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
3710 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
3713 delivered = tcp_newly_delivered(sk, delivered, flag);
3714 lost = tp->lost - lost; /* freshly marked lost */
3715 rs.is_ack_delayed = !!(flag & FLAG_ACK_MAYBE_DELAYED);
3716 tcp_rate_gen(sk, delivered, lost, is_sack_reneg, sack_state.rate);
3717 tcp_cong_control(sk, ack, delivered, flag, sack_state.rate);
3718 tcp_xmit_recovery(sk, rexmit);
3722 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3723 if (flag & FLAG_DSACKING_ACK) {
3724 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
3726 tcp_newly_delivered(sk, delivered, flag);
3728 /* If this ack opens up a zero window, clear backoff. It was
3729 * being used to time the probes, and is probably far higher than
3730 * it needs to be for normal retransmission.
3734 if (tp->tlp_high_seq)
3735 tcp_process_tlp_ack(sk, ack, flag);
3739 /* If data was SACKed, tag it and see if we should send more data.
3740 * If data was DSACKed, see if we can undo a cwnd reduction.
3742 if (TCP_SKB_CB(skb)->sacked) {
3743 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3745 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
3747 tcp_newly_delivered(sk, delivered, flag);
3748 tcp_xmit_recovery(sk, rexmit);
3754 static void tcp_parse_fastopen_option(int len, const unsigned char *cookie,
3755 bool syn, struct tcp_fastopen_cookie *foc,
3758 /* Valid only in SYN or SYN-ACK with an even length. */
3759 if (!foc || !syn || len < 0 || (len & 1))
3762 if (len >= TCP_FASTOPEN_COOKIE_MIN &&
3763 len <= TCP_FASTOPEN_COOKIE_MAX)
3764 memcpy(foc->val, cookie, len);
3771 static void smc_parse_options(const struct tcphdr *th,
3772 struct tcp_options_received *opt_rx,
3773 const unsigned char *ptr,
3776 #if IS_ENABLED(CONFIG_SMC)
3777 if (static_branch_unlikely(&tcp_have_smc)) {
3778 if (th->syn && !(opsize & 1) &&
3779 opsize >= TCPOLEN_EXP_SMC_BASE &&
3780 get_unaligned_be32(ptr) == TCPOPT_SMC_MAGIC)
3786 /* Try to parse the MSS option from the TCP header. Return 0 on failure, clamped
3789 static u16 tcp_parse_mss_option(const struct tcphdr *th, u16 user_mss)
3791 const unsigned char *ptr = (const unsigned char *)(th + 1);
3792 int length = (th->doff * 4) - sizeof(struct tcphdr);
3795 while (length > 0) {
3796 int opcode = *ptr++;
3802 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3809 if (opsize < 2) /* "silly options" */
3811 if (opsize > length)
3812 return mss; /* fail on partial options */
3813 if (opcode == TCPOPT_MSS && opsize == TCPOLEN_MSS) {
3814 u16 in_mss = get_unaligned_be16(ptr);
3817 if (user_mss && user_mss < in_mss)
3829 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3830 * But, this can also be called on packets in the established flow when
3831 * the fast version below fails.
3833 void tcp_parse_options(const struct net *net,
3834 const struct sk_buff *skb,
3835 struct tcp_options_received *opt_rx, int estab,
3836 struct tcp_fastopen_cookie *foc)
3838 const unsigned char *ptr;
3839 const struct tcphdr *th = tcp_hdr(skb);
3840 int length = (th->doff * 4) - sizeof(struct tcphdr);
3842 ptr = (const unsigned char *)(th + 1);
3843 opt_rx->saw_tstamp = 0;
3845 while (length > 0) {
3846 int opcode = *ptr++;
3852 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3859 if (opsize < 2) /* "silly options" */
3861 if (opsize > length)
3862 return; /* don't parse partial options */
3865 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3866 u16 in_mss = get_unaligned_be16(ptr);
3868 if (opt_rx->user_mss &&
3869 opt_rx->user_mss < in_mss)
3870 in_mss = opt_rx->user_mss;
3871 opt_rx->mss_clamp = in_mss;
3876 if (opsize == TCPOLEN_WINDOW && th->syn &&
3877 !estab && net->ipv4.sysctl_tcp_window_scaling) {
3878 __u8 snd_wscale = *(__u8 *)ptr;
3879 opt_rx->wscale_ok = 1;
3880 if (snd_wscale > TCP_MAX_WSCALE) {
3881 net_info_ratelimited("%s: Illegal window scaling value %d > %u received\n",
3885 snd_wscale = TCP_MAX_WSCALE;
3887 opt_rx->snd_wscale = snd_wscale;
3890 case TCPOPT_TIMESTAMP:
3891 if ((opsize == TCPOLEN_TIMESTAMP) &&
3892 ((estab && opt_rx->tstamp_ok) ||
3893 (!estab && net->ipv4.sysctl_tcp_timestamps))) {
3894 opt_rx->saw_tstamp = 1;
3895 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3896 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3899 case TCPOPT_SACK_PERM:
3900 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3901 !estab && net->ipv4.sysctl_tcp_sack) {
3902 opt_rx->sack_ok = TCP_SACK_SEEN;
3903 tcp_sack_reset(opt_rx);
3908 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3909 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3911 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3914 #ifdef CONFIG_TCP_MD5SIG
3917 * The MD5 Hash has already been
3918 * checked (see tcp_v{4,6}_do_rcv()).
3922 case TCPOPT_FASTOPEN:
3923 tcp_parse_fastopen_option(
3924 opsize - TCPOLEN_FASTOPEN_BASE,
3925 ptr, th->syn, foc, false);
3929 /* Fast Open option shares code 254 using a
3930 * 16 bits magic number.
3932 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE &&
3933 get_unaligned_be16(ptr) ==
3934 TCPOPT_FASTOPEN_MAGIC)
3935 tcp_parse_fastopen_option(opsize -
3936 TCPOLEN_EXP_FASTOPEN_BASE,
3937 ptr + 2, th->syn, foc, true);
3939 smc_parse_options(th, opt_rx, ptr,
3949 EXPORT_SYMBOL(tcp_parse_options);
3951 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
3953 const __be32 *ptr = (const __be32 *)(th + 1);
3955 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3956 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3957 tp->rx_opt.saw_tstamp = 1;
3959 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3962 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
3964 tp->rx_opt.rcv_tsecr = 0;
3970 /* Fast parse options. This hopes to only see timestamps.
3971 * If it is wrong it falls back on tcp_parse_options().
3973 static bool tcp_fast_parse_options(const struct net *net,
3974 const struct sk_buff *skb,
3975 const struct tcphdr *th, struct tcp_sock *tp)
3977 /* In the spirit of fast parsing, compare doff directly to constant
3978 * values. Because equality is used, short doff can be ignored here.
3980 if (th->doff == (sizeof(*th) / 4)) {
3981 tp->rx_opt.saw_tstamp = 0;
3983 } else if (tp->rx_opt.tstamp_ok &&
3984 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3985 if (tcp_parse_aligned_timestamp(tp, th))
3989 tcp_parse_options(net, skb, &tp->rx_opt, 1, NULL);
3990 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3991 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
3996 #ifdef CONFIG_TCP_MD5SIG
3998 * Parse MD5 Signature option
4000 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
4002 int length = (th->doff << 2) - sizeof(*th);
4003 const u8 *ptr = (const u8 *)(th + 1);
4005 /* If not enough data remaining, we can short cut */
4006 while (length >= TCPOLEN_MD5SIG) {
4007 int opcode = *ptr++;
4018 if (opsize < 2 || opsize > length)
4020 if (opcode == TCPOPT_MD5SIG)
4021 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
4028 EXPORT_SYMBOL(tcp_parse_md5sig_option);
4031 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
4033 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
4034 * it can pass through stack. So, the following predicate verifies that
4035 * this segment is not used for anything but congestion avoidance or
4036 * fast retransmit. Moreover, we even are able to eliminate most of such
4037 * second order effects, if we apply some small "replay" window (~RTO)
4038 * to timestamp space.
4040 * All these measures still do not guarantee that we reject wrapped ACKs
4041 * on networks with high bandwidth, when sequence space is recycled fastly,
4042 * but it guarantees that such events will be very rare and do not affect
4043 * connection seriously. This doesn't look nice, but alas, PAWS is really
4046 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
4047 * states that events when retransmit arrives after original data are rare.
4048 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
4049 * the biggest problem on large power networks even with minor reordering.
4050 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
4051 * up to bandwidth of 18Gigabit/sec. 8) ]
4054 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
4056 const struct tcp_sock *tp = tcp_sk(sk);
4057 const struct tcphdr *th = tcp_hdr(skb);
4058 u32 seq = TCP_SKB_CB(skb)->seq;
4059 u32 ack = TCP_SKB_CB(skb)->ack_seq;
4061 return (/* 1. Pure ACK with correct sequence number. */
4062 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
4064 /* 2. ... and duplicate ACK. */
4065 ack == tp->snd_una &&
4067 /* 3. ... and does not update window. */
4068 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
4070 /* 4. ... and sits in replay window. */
4071 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
4074 static inline bool tcp_paws_discard(const struct sock *sk,
4075 const struct sk_buff *skb)
4077 const struct tcp_sock *tp = tcp_sk(sk);
4079 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
4080 !tcp_disordered_ack(sk, skb);
4083 /* Check segment sequence number for validity.
4085 * Segment controls are considered valid, if the segment
4086 * fits to the window after truncation to the window. Acceptability
4087 * of data (and SYN, FIN, of course) is checked separately.
4088 * See tcp_data_queue(), for example.
4090 * Also, controls (RST is main one) are accepted using RCV.WUP instead
4091 * of RCV.NXT. Peer still did not advance his SND.UNA when we
4092 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4093 * (borrowed from freebsd)
4096 static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
4098 return !before(end_seq, tp->rcv_wup) &&
4099 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
4102 /* When we get a reset we do this. */
4103 void tcp_reset(struct sock *sk)
4105 trace_tcp_receive_reset(sk);
4107 /* We want the right error as BSD sees it (and indeed as we do). */
4108 switch (sk->sk_state) {
4110 sk->sk_err = ECONNREFUSED;
4112 case TCP_CLOSE_WAIT:
4118 sk->sk_err = ECONNRESET;
4120 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4123 tcp_write_queue_purge(sk);
4126 if (!sock_flag(sk, SOCK_DEAD))
4127 sk->sk_error_report(sk);
4131 * Process the FIN bit. This now behaves as it is supposed to work
4132 * and the FIN takes effect when it is validly part of sequence
4133 * space. Not before when we get holes.
4135 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4136 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4139 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4140 * close and we go into CLOSING (and later onto TIME-WAIT)
4142 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4144 void tcp_fin(struct sock *sk)
4146 struct tcp_sock *tp = tcp_sk(sk);
4148 inet_csk_schedule_ack(sk);
4150 sk->sk_shutdown |= RCV_SHUTDOWN;
4151 sock_set_flag(sk, SOCK_DONE);
4153 switch (sk->sk_state) {
4155 case TCP_ESTABLISHED:
4156 /* Move to CLOSE_WAIT */
4157 tcp_set_state(sk, TCP_CLOSE_WAIT);
4158 inet_csk_enter_pingpong_mode(sk);
4161 case TCP_CLOSE_WAIT:
4163 /* Received a retransmission of the FIN, do
4168 /* RFC793: Remain in the LAST-ACK state. */
4172 /* This case occurs when a simultaneous close
4173 * happens, we must ack the received FIN and
4174 * enter the CLOSING state.
4177 tcp_set_state(sk, TCP_CLOSING);
4180 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4182 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4185 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4186 * cases we should never reach this piece of code.
4188 pr_err("%s: Impossible, sk->sk_state=%d\n",
4189 __func__, sk->sk_state);
4193 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4194 * Probably, we should reset in this case. For now drop them.
4196 skb_rbtree_purge(&tp->out_of_order_queue);
4197 if (tcp_is_sack(tp))
4198 tcp_sack_reset(&tp->rx_opt);
4201 if (!sock_flag(sk, SOCK_DEAD)) {
4202 sk->sk_state_change(sk);
4204 /* Do not send POLL_HUP for half duplex close. */
4205 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4206 sk->sk_state == TCP_CLOSE)
4207 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4209 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4213 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4216 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4217 if (before(seq, sp->start_seq))
4218 sp->start_seq = seq;
4219 if (after(end_seq, sp->end_seq))
4220 sp->end_seq = end_seq;
4226 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4228 struct tcp_sock *tp = tcp_sk(sk);
4230 if (tcp_is_sack(tp) && sock_net(sk)->ipv4.sysctl_tcp_dsack) {
4233 if (before(seq, tp->rcv_nxt))
4234 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4236 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4238 NET_INC_STATS(sock_net(sk), mib_idx);
4240 tp->rx_opt.dsack = 1;
4241 tp->duplicate_sack[0].start_seq = seq;
4242 tp->duplicate_sack[0].end_seq = end_seq;
4246 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4248 struct tcp_sock *tp = tcp_sk(sk);
4250 if (!tp->rx_opt.dsack)
4251 tcp_dsack_set(sk, seq, end_seq);
4253 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4256 static void tcp_rcv_spurious_retrans(struct sock *sk, const struct sk_buff *skb)
4258 /* When the ACK path fails or drops most ACKs, the sender would
4259 * timeout and spuriously retransmit the same segment repeatedly.
4260 * The receiver remembers and reflects via DSACKs. Leverage the
4261 * DSACK state and change the txhash to re-route speculatively.
4263 if (TCP_SKB_CB(skb)->seq == tcp_sk(sk)->duplicate_sack[0].start_seq)
4264 sk_rethink_txhash(sk);
4267 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4269 struct tcp_sock *tp = tcp_sk(sk);
4271 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4272 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4273 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4274 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
4276 if (tcp_is_sack(tp) && sock_net(sk)->ipv4.sysctl_tcp_dsack) {
4277 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4279 tcp_rcv_spurious_retrans(sk, skb);
4280 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4281 end_seq = tp->rcv_nxt;
4282 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4289 /* These routines update the SACK block as out-of-order packets arrive or
4290 * in-order packets close up the sequence space.
4292 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4295 struct tcp_sack_block *sp = &tp->selective_acks[0];
4296 struct tcp_sack_block *swalk = sp + 1;
4298 /* See if the recent change to the first SACK eats into
4299 * or hits the sequence space of other SACK blocks, if so coalesce.
4301 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4302 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4305 /* Zap SWALK, by moving every further SACK up by one slot.
4306 * Decrease num_sacks.
4308 tp->rx_opt.num_sacks--;
4309 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4313 this_sack++, swalk++;
4317 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4319 struct tcp_sock *tp = tcp_sk(sk);
4320 struct tcp_sack_block *sp = &tp->selective_acks[0];
4321 int cur_sacks = tp->rx_opt.num_sacks;
4327 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4328 if (tcp_sack_extend(sp, seq, end_seq)) {
4329 /* Rotate this_sack to the first one. */
4330 for (; this_sack > 0; this_sack--, sp--)
4331 swap(*sp, *(sp - 1));
4333 tcp_sack_maybe_coalesce(tp);
4338 /* Could not find an adjacent existing SACK, build a new one,
4339 * put it at the front, and shift everyone else down. We
4340 * always know there is at least one SACK present already here.
4342 * If the sack array is full, forget about the last one.
4344 if (this_sack >= TCP_NUM_SACKS) {
4345 if (tp->compressed_ack > TCP_FASTRETRANS_THRESH)
4348 tp->rx_opt.num_sacks--;
4351 for (; this_sack > 0; this_sack--, sp--)
4355 /* Build the new head SACK, and we're done. */
4356 sp->start_seq = seq;
4357 sp->end_seq = end_seq;
4358 tp->rx_opt.num_sacks++;
4361 /* RCV.NXT advances, some SACKs should be eaten. */
4363 static void tcp_sack_remove(struct tcp_sock *tp)
4365 struct tcp_sack_block *sp = &tp->selective_acks[0];
4366 int num_sacks = tp->rx_opt.num_sacks;
4369 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4370 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4371 tp->rx_opt.num_sacks = 0;
4375 for (this_sack = 0; this_sack < num_sacks;) {
4376 /* Check if the start of the sack is covered by RCV.NXT. */
4377 if (!before(tp->rcv_nxt, sp->start_seq)) {
4380 /* RCV.NXT must cover all the block! */
4381 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4383 /* Zap this SACK, by moving forward any other SACKS. */
4384 for (i = this_sack+1; i < num_sacks; i++)
4385 tp->selective_acks[i-1] = tp->selective_acks[i];
4392 tp->rx_opt.num_sacks = num_sacks;
4396 * tcp_try_coalesce - try to merge skb to prior one
4398 * @dest: destination queue
4400 * @from: buffer to add in queue
4401 * @fragstolen: pointer to boolean
4403 * Before queueing skb @from after @to, try to merge them
4404 * to reduce overall memory use and queue lengths, if cost is small.
4405 * Packets in ofo or receive queues can stay a long time.
4406 * Better try to coalesce them right now to avoid future collapses.
4407 * Returns true if caller should free @from instead of queueing it
4409 static bool tcp_try_coalesce(struct sock *sk,
4411 struct sk_buff *from,
4416 *fragstolen = false;
4418 /* Its possible this segment overlaps with prior segment in queue */
4419 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4422 #ifdef CONFIG_TLS_DEVICE
4423 if (from->decrypted != to->decrypted)
4427 if (!skb_try_coalesce(to, from, fragstolen, &delta))
4430 atomic_add(delta, &sk->sk_rmem_alloc);
4431 sk_mem_charge(sk, delta);
4432 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4433 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4434 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4435 TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
4437 if (TCP_SKB_CB(from)->has_rxtstamp) {
4438 TCP_SKB_CB(to)->has_rxtstamp = true;
4439 to->tstamp = from->tstamp;
4440 skb_hwtstamps(to)->hwtstamp = skb_hwtstamps(from)->hwtstamp;
4446 static bool tcp_ooo_try_coalesce(struct sock *sk,
4448 struct sk_buff *from,
4451 bool res = tcp_try_coalesce(sk, to, from, fragstolen);
4453 /* In case tcp_drop() is called later, update to->gso_segs */
4455 u32 gso_segs = max_t(u16, 1, skb_shinfo(to)->gso_segs) +
4456 max_t(u16, 1, skb_shinfo(from)->gso_segs);
4458 skb_shinfo(to)->gso_segs = min_t(u32, gso_segs, 0xFFFF);
4463 static void tcp_drop(struct sock *sk, struct sk_buff *skb)
4465 sk_drops_add(sk, skb);
4469 /* This one checks to see if we can put data from the
4470 * out_of_order queue into the receive_queue.
4472 static void tcp_ofo_queue(struct sock *sk)
4474 struct tcp_sock *tp = tcp_sk(sk);
4475 __u32 dsack_high = tp->rcv_nxt;
4476 bool fin, fragstolen, eaten;
4477 struct sk_buff *skb, *tail;
4480 p = rb_first(&tp->out_of_order_queue);
4483 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4486 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4487 __u32 dsack = dsack_high;
4488 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4489 dsack_high = TCP_SKB_CB(skb)->end_seq;
4490 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4493 rb_erase(&skb->rbnode, &tp->out_of_order_queue);
4495 if (unlikely(!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))) {
4500 tail = skb_peek_tail(&sk->sk_receive_queue);
4501 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen);
4502 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4503 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
4505 __skb_queue_tail(&sk->sk_receive_queue, skb);
4507 kfree_skb_partial(skb, fragstolen);
4509 if (unlikely(fin)) {
4511 /* tcp_fin() purges tp->out_of_order_queue,
4512 * so we must end this loop right now.
4519 static bool tcp_prune_ofo_queue(struct sock *sk);
4520 static int tcp_prune_queue(struct sock *sk);
4522 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4525 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4526 !sk_rmem_schedule(sk, skb, size)) {
4528 if (tcp_prune_queue(sk) < 0)
4531 while (!sk_rmem_schedule(sk, skb, size)) {
4532 if (!tcp_prune_ofo_queue(sk))
4539 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4541 struct tcp_sock *tp = tcp_sk(sk);
4542 struct rb_node **p, *parent;
4543 struct sk_buff *skb1;
4547 tcp_ecn_check_ce(sk, skb);
4549 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4550 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFODROP);
4555 /* Disable header prediction. */
4557 inet_csk_schedule_ack(sk);
4559 tp->rcv_ooopack += max_t(u16, 1, skb_shinfo(skb)->gso_segs);
4560 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4561 seq = TCP_SKB_CB(skb)->seq;
4562 end_seq = TCP_SKB_CB(skb)->end_seq;
4564 p = &tp->out_of_order_queue.rb_node;
4565 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4566 /* Initial out of order segment, build 1 SACK. */
4567 if (tcp_is_sack(tp)) {
4568 tp->rx_opt.num_sacks = 1;
4569 tp->selective_acks[0].start_seq = seq;
4570 tp->selective_acks[0].end_seq = end_seq;
4572 rb_link_node(&skb->rbnode, NULL, p);
4573 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4574 tp->ooo_last_skb = skb;
4578 /* In the typical case, we are adding an skb to the end of the list.
4579 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
4581 if (tcp_ooo_try_coalesce(sk, tp->ooo_last_skb,
4582 skb, &fragstolen)) {
4584 tcp_grow_window(sk, skb);
4585 kfree_skb_partial(skb, fragstolen);
4589 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
4590 if (!before(seq, TCP_SKB_CB(tp->ooo_last_skb)->end_seq)) {
4591 parent = &tp->ooo_last_skb->rbnode;
4592 p = &parent->rb_right;
4596 /* Find place to insert this segment. Handle overlaps on the way. */
4600 skb1 = rb_to_skb(parent);
4601 if (before(seq, TCP_SKB_CB(skb1)->seq)) {
4602 p = &parent->rb_left;
4605 if (before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4606 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4607 /* All the bits are present. Drop. */
4608 NET_INC_STATS(sock_net(sk),
4609 LINUX_MIB_TCPOFOMERGE);
4612 tcp_dsack_set(sk, seq, end_seq);
4615 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4616 /* Partial overlap. */
4617 tcp_dsack_set(sk, seq, TCP_SKB_CB(skb1)->end_seq);
4619 /* skb's seq == skb1's seq and skb covers skb1.
4620 * Replace skb1 with skb.
4622 rb_replace_node(&skb1->rbnode, &skb->rbnode,
4623 &tp->out_of_order_queue);
4624 tcp_dsack_extend(sk,
4625 TCP_SKB_CB(skb1)->seq,
4626 TCP_SKB_CB(skb1)->end_seq);
4627 NET_INC_STATS(sock_net(sk),
4628 LINUX_MIB_TCPOFOMERGE);
4632 } else if (tcp_ooo_try_coalesce(sk, skb1,
4633 skb, &fragstolen)) {
4636 p = &parent->rb_right;
4639 /* Insert segment into RB tree. */
4640 rb_link_node(&skb->rbnode, parent, p);
4641 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4644 /* Remove other segments covered by skb. */
4645 while ((skb1 = skb_rb_next(skb)) != NULL) {
4646 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4648 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4649 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4653 rb_erase(&skb1->rbnode, &tp->out_of_order_queue);
4654 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4655 TCP_SKB_CB(skb1)->end_seq);
4656 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4659 /* If there is no skb after us, we are the last_skb ! */
4661 tp->ooo_last_skb = skb;
4664 if (tcp_is_sack(tp))
4665 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4668 tcp_grow_window(sk, skb);
4670 skb_set_owner_r(skb, sk);
4674 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb,
4678 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4681 tcp_try_coalesce(sk, tail,
4682 skb, fragstolen)) ? 1 : 0;
4683 tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq);
4685 __skb_queue_tail(&sk->sk_receive_queue, skb);
4686 skb_set_owner_r(skb, sk);
4691 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4693 struct sk_buff *skb;
4701 if (size > PAGE_SIZE) {
4702 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS);
4704 data_len = npages << PAGE_SHIFT;
4705 size = data_len + (size & ~PAGE_MASK);
4707 skb = alloc_skb_with_frags(size - data_len, data_len,
4708 PAGE_ALLOC_COSTLY_ORDER,
4709 &err, sk->sk_allocation);
4713 skb_put(skb, size - data_len);
4714 skb->data_len = data_len;
4717 if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) {
4718 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP);
4722 err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size);
4726 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4727 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4728 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4730 if (tcp_queue_rcv(sk, skb, &fragstolen)) {
4731 WARN_ON_ONCE(fragstolen); /* should not happen */
4743 void tcp_data_ready(struct sock *sk)
4745 const struct tcp_sock *tp = tcp_sk(sk);
4746 int avail = tp->rcv_nxt - tp->copied_seq;
4748 if (avail < sk->sk_rcvlowat && !sock_flag(sk, SOCK_DONE))
4751 sk->sk_data_ready(sk);
4754 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4756 struct tcp_sock *tp = tcp_sk(sk);
4760 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
4765 __skb_pull(skb, tcp_hdr(skb)->doff * 4);
4767 tcp_ecn_accept_cwr(sk, skb);
4769 tp->rx_opt.dsack = 0;
4771 /* Queue data for delivery to the user.
4772 * Packets in sequence go to the receive queue.
4773 * Out of sequence packets to the out_of_order_queue.
4775 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4776 if (tcp_receive_window(tp) == 0) {
4777 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP);
4781 /* Ok. In sequence. In window. */
4783 if (skb_queue_len(&sk->sk_receive_queue) == 0)
4784 sk_forced_mem_schedule(sk, skb->truesize);
4785 else if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) {
4786 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP);
4790 eaten = tcp_queue_rcv(sk, skb, &fragstolen);
4792 tcp_event_data_recv(sk, skb);
4793 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
4796 if (!RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4799 /* RFC5681. 4.2. SHOULD send immediate ACK, when
4800 * gap in queue is filled.
4802 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
4803 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW;
4806 if (tp->rx_opt.num_sacks)
4807 tcp_sack_remove(tp);
4809 tcp_fast_path_check(sk);
4812 kfree_skb_partial(skb, fragstolen);
4813 if (!sock_flag(sk, SOCK_DEAD))
4818 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4819 tcp_rcv_spurious_retrans(sk, skb);
4820 /* A retransmit, 2nd most common case. Force an immediate ack. */
4821 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4822 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4825 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
4826 inet_csk_schedule_ack(sk);
4832 /* Out of window. F.e. zero window probe. */
4833 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4836 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4837 /* Partial packet, seq < rcv_next < end_seq */
4838 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4840 /* If window is closed, drop tail of packet. But after
4841 * remembering D-SACK for its head made in previous line.
4843 if (!tcp_receive_window(tp)) {
4844 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP);
4850 tcp_data_queue_ofo(sk, skb);
4853 static struct sk_buff *tcp_skb_next(struct sk_buff *skb, struct sk_buff_head *list)
4856 return !skb_queue_is_last(list, skb) ? skb->next : NULL;
4858 return skb_rb_next(skb);
4861 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4862 struct sk_buff_head *list,
4863 struct rb_root *root)
4865 struct sk_buff *next = tcp_skb_next(skb, list);
4868 __skb_unlink(skb, list);
4870 rb_erase(&skb->rbnode, root);
4873 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4878 /* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */
4879 void tcp_rbtree_insert(struct rb_root *root, struct sk_buff *skb)
4881 struct rb_node **p = &root->rb_node;
4882 struct rb_node *parent = NULL;
4883 struct sk_buff *skb1;
4887 skb1 = rb_to_skb(parent);
4888 if (before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb1)->seq))
4889 p = &parent->rb_left;
4891 p = &parent->rb_right;
4893 rb_link_node(&skb->rbnode, parent, p);
4894 rb_insert_color(&skb->rbnode, root);
4897 /* Collapse contiguous sequence of skbs head..tail with
4898 * sequence numbers start..end.
4900 * If tail is NULL, this means until the end of the queue.
4902 * Segments with FIN/SYN are not collapsed (only because this
4906 tcp_collapse(struct sock *sk, struct sk_buff_head *list, struct rb_root *root,
4907 struct sk_buff *head, struct sk_buff *tail, u32 start, u32 end)
4909 struct sk_buff *skb = head, *n;
4910 struct sk_buff_head tmp;
4913 /* First, check that queue is collapsible and find
4914 * the point where collapsing can be useful.
4917 for (end_of_skbs = true; skb != NULL && skb != tail; skb = n) {
4918 n = tcp_skb_next(skb, list);
4920 /* No new bits? It is possible on ofo queue. */
4921 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4922 skb = tcp_collapse_one(sk, skb, list, root);
4928 /* The first skb to collapse is:
4930 * - bloated or contains data before "start" or
4931 * overlaps to the next one.
4933 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
4934 (tcp_win_from_space(sk, skb->truesize) > skb->len ||
4935 before(TCP_SKB_CB(skb)->seq, start))) {
4936 end_of_skbs = false;
4940 if (n && n != tail &&
4941 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(n)->seq) {
4942 end_of_skbs = false;
4946 /* Decided to skip this, advance start seq. */
4947 start = TCP_SKB_CB(skb)->end_seq;
4950 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4953 __skb_queue_head_init(&tmp);
4955 while (before(start, end)) {
4956 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
4957 struct sk_buff *nskb;
4959 nskb = alloc_skb(copy, GFP_ATOMIC);
4963 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4964 #ifdef CONFIG_TLS_DEVICE
4965 nskb->decrypted = skb->decrypted;
4967 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4969 __skb_queue_before(list, skb, nskb);
4971 __skb_queue_tail(&tmp, nskb); /* defer rbtree insertion */
4972 skb_set_owner_r(nskb, sk);
4974 /* Copy data, releasing collapsed skbs. */
4976 int offset = start - TCP_SKB_CB(skb)->seq;
4977 int size = TCP_SKB_CB(skb)->end_seq - start;
4981 size = min(copy, size);
4982 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4984 TCP_SKB_CB(nskb)->end_seq += size;
4988 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4989 skb = tcp_collapse_one(sk, skb, list, root);
4992 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4994 #ifdef CONFIG_TLS_DEVICE
4995 if (skb->decrypted != nskb->decrypted)
5002 skb_queue_walk_safe(&tmp, skb, n)
5003 tcp_rbtree_insert(root, skb);
5006 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
5007 * and tcp_collapse() them until all the queue is collapsed.
5009 static void tcp_collapse_ofo_queue(struct sock *sk)
5011 struct tcp_sock *tp = tcp_sk(sk);
5012 u32 range_truesize, sum_tiny = 0;
5013 struct sk_buff *skb, *head;
5016 skb = skb_rb_first(&tp->out_of_order_queue);
5019 tp->ooo_last_skb = skb_rb_last(&tp->out_of_order_queue);
5022 start = TCP_SKB_CB(skb)->seq;
5023 end = TCP_SKB_CB(skb)->end_seq;
5024 range_truesize = skb->truesize;
5026 for (head = skb;;) {
5027 skb = skb_rb_next(skb);
5029 /* Range is terminated when we see a gap or when
5030 * we are at the queue end.
5033 after(TCP_SKB_CB(skb)->seq, end) ||
5034 before(TCP_SKB_CB(skb)->end_seq, start)) {
5035 /* Do not attempt collapsing tiny skbs */
5036 if (range_truesize != head->truesize ||
5037 end - start >= SKB_WITH_OVERHEAD(SK_MEM_QUANTUM)) {
5038 tcp_collapse(sk, NULL, &tp->out_of_order_queue,
5039 head, skb, start, end);
5041 sum_tiny += range_truesize;
5042 if (sum_tiny > sk->sk_rcvbuf >> 3)
5048 range_truesize += skb->truesize;
5049 if (unlikely(before(TCP_SKB_CB(skb)->seq, start)))
5050 start = TCP_SKB_CB(skb)->seq;
5051 if (after(TCP_SKB_CB(skb)->end_seq, end))
5052 end = TCP_SKB_CB(skb)->end_seq;
5057 * Clean the out-of-order queue to make room.
5058 * We drop high sequences packets to :
5059 * 1) Let a chance for holes to be filled.
5060 * 2) not add too big latencies if thousands of packets sit there.
5061 * (But if application shrinks SO_RCVBUF, we could still end up
5062 * freeing whole queue here)
5063 * 3) Drop at least 12.5 % of sk_rcvbuf to avoid malicious attacks.
5065 * Return true if queue has shrunk.
5067 static bool tcp_prune_ofo_queue(struct sock *sk)
5069 struct tcp_sock *tp = tcp_sk(sk);
5070 struct rb_node *node, *prev;
5073 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
5076 NET_INC_STATS(sock_net(sk), LINUX_MIB_OFOPRUNED);
5077 goal = sk->sk_rcvbuf >> 3;
5078 node = &tp->ooo_last_skb->rbnode;
5080 prev = rb_prev(node);
5081 rb_erase(node, &tp->out_of_order_queue);
5082 goal -= rb_to_skb(node)->truesize;
5083 tcp_drop(sk, rb_to_skb(node));
5084 if (!prev || goal <= 0) {
5086 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf &&
5087 !tcp_under_memory_pressure(sk))
5089 goal = sk->sk_rcvbuf >> 3;
5093 tp->ooo_last_skb = rb_to_skb(prev);
5095 /* Reset SACK state. A conforming SACK implementation will
5096 * do the same at a timeout based retransmit. When a connection
5097 * is in a sad state like this, we care only about integrity
5098 * of the connection not performance.
5100 if (tp->rx_opt.sack_ok)
5101 tcp_sack_reset(&tp->rx_opt);
5105 /* Reduce allocated memory if we can, trying to get
5106 * the socket within its memory limits again.
5108 * Return less than zero if we should start dropping frames
5109 * until the socket owning process reads some of the data
5110 * to stabilize the situation.
5112 static int tcp_prune_queue(struct sock *sk)
5114 struct tcp_sock *tp = tcp_sk(sk);
5116 NET_INC_STATS(sock_net(sk), LINUX_MIB_PRUNECALLED);
5118 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
5119 tcp_clamp_window(sk);
5120 else if (tcp_under_memory_pressure(sk))
5121 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
5123 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5126 tcp_collapse_ofo_queue(sk);
5127 if (!skb_queue_empty(&sk->sk_receive_queue))
5128 tcp_collapse(sk, &sk->sk_receive_queue, NULL,
5129 skb_peek(&sk->sk_receive_queue),
5131 tp->copied_seq, tp->rcv_nxt);
5134 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5137 /* Collapsing did not help, destructive actions follow.
5138 * This must not ever occur. */
5140 tcp_prune_ofo_queue(sk);
5142 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5145 /* If we are really being abused, tell the caller to silently
5146 * drop receive data on the floor. It will get retransmitted
5147 * and hopefully then we'll have sufficient space.
5149 NET_INC_STATS(sock_net(sk), LINUX_MIB_RCVPRUNED);
5151 /* Massive buffer overcommit. */
5156 static bool tcp_should_expand_sndbuf(const struct sock *sk)
5158 const struct tcp_sock *tp = tcp_sk(sk);
5160 /* If the user specified a specific send buffer setting, do
5163 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
5166 /* If we are under global TCP memory pressure, do not expand. */
5167 if (tcp_under_memory_pressure(sk))
5170 /* If we are under soft global TCP memory pressure, do not expand. */
5171 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
5174 /* If we filled the congestion window, do not expand. */
5175 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
5181 /* When incoming ACK allowed to free some skb from write_queue,
5182 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
5183 * on the exit from tcp input handler.
5185 * PROBLEM: sndbuf expansion does not work well with largesend.
5187 static void tcp_new_space(struct sock *sk)
5189 struct tcp_sock *tp = tcp_sk(sk);
5191 if (tcp_should_expand_sndbuf(sk)) {
5192 tcp_sndbuf_expand(sk);
5193 tp->snd_cwnd_stamp = tcp_jiffies32;
5196 sk->sk_write_space(sk);
5199 static void tcp_check_space(struct sock *sk)
5201 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
5202 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
5203 /* pairs with tcp_poll() */
5205 if (sk->sk_socket &&
5206 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
5208 if (!test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
5209 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
5214 static inline void tcp_data_snd_check(struct sock *sk)
5216 tcp_push_pending_frames(sk);
5217 tcp_check_space(sk);
5221 * Check if sending an ack is needed.
5223 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
5225 struct tcp_sock *tp = tcp_sk(sk);
5226 unsigned long rtt, delay;
5228 /* More than one full frame received... */
5229 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
5230 /* ... and right edge of window advances far enough.
5231 * (tcp_recvmsg() will send ACK otherwise).
5232 * If application uses SO_RCVLOWAT, we want send ack now if
5233 * we have not received enough bytes to satisfy the condition.
5235 (tp->rcv_nxt - tp->copied_seq < sk->sk_rcvlowat ||
5236 __tcp_select_window(sk) >= tp->rcv_wnd)) ||
5237 /* We ACK each frame or... */
5238 tcp_in_quickack_mode(sk) ||
5239 /* Protocol state mandates a one-time immediate ACK */
5240 inet_csk(sk)->icsk_ack.pending & ICSK_ACK_NOW) {
5246 if (!ofo_possible || RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
5247 tcp_send_delayed_ack(sk);
5251 if (!tcp_is_sack(tp) ||
5252 tp->compressed_ack >= sock_net(sk)->ipv4.sysctl_tcp_comp_sack_nr)
5255 if (tp->compressed_ack_rcv_nxt != tp->rcv_nxt) {
5256 tp->compressed_ack_rcv_nxt = tp->rcv_nxt;
5257 if (tp->compressed_ack > TCP_FASTRETRANS_THRESH)
5258 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED,
5259 tp->compressed_ack - TCP_FASTRETRANS_THRESH);
5260 tp->compressed_ack = 0;
5263 if (++tp->compressed_ack <= TCP_FASTRETRANS_THRESH)
5266 if (hrtimer_is_queued(&tp->compressed_ack_timer))
5269 /* compress ack timer : 5 % of rtt, but no more than tcp_comp_sack_delay_ns */
5271 rtt = tp->rcv_rtt_est.rtt_us;
5272 if (tp->srtt_us && tp->srtt_us < rtt)
5275 delay = min_t(unsigned long, sock_net(sk)->ipv4.sysctl_tcp_comp_sack_delay_ns,
5276 rtt * (NSEC_PER_USEC >> 3)/20);
5278 hrtimer_start(&tp->compressed_ack_timer, ns_to_ktime(delay),
5279 HRTIMER_MODE_REL_PINNED_SOFT);
5282 static inline void tcp_ack_snd_check(struct sock *sk)
5284 if (!inet_csk_ack_scheduled(sk)) {
5285 /* We sent a data segment already. */
5288 __tcp_ack_snd_check(sk, 1);
5292 * This routine is only called when we have urgent data
5293 * signaled. Its the 'slow' part of tcp_urg. It could be
5294 * moved inline now as tcp_urg is only called from one
5295 * place. We handle URGent data wrong. We have to - as
5296 * BSD still doesn't use the correction from RFC961.
5297 * For 1003.1g we should support a new option TCP_STDURG to permit
5298 * either form (or just set the sysctl tcp_stdurg).
5301 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
5303 struct tcp_sock *tp = tcp_sk(sk);
5304 u32 ptr = ntohs(th->urg_ptr);
5306 if (ptr && !sock_net(sk)->ipv4.sysctl_tcp_stdurg)
5308 ptr += ntohl(th->seq);
5310 /* Ignore urgent data that we've already seen and read. */
5311 if (after(tp->copied_seq, ptr))
5314 /* Do not replay urg ptr.
5316 * NOTE: interesting situation not covered by specs.
5317 * Misbehaving sender may send urg ptr, pointing to segment,
5318 * which we already have in ofo queue. We are not able to fetch
5319 * such data and will stay in TCP_URG_NOTYET until will be eaten
5320 * by recvmsg(). Seems, we are not obliged to handle such wicked
5321 * situations. But it is worth to think about possibility of some
5322 * DoSes using some hypothetical application level deadlock.
5324 if (before(ptr, tp->rcv_nxt))
5327 /* Do we already have a newer (or duplicate) urgent pointer? */
5328 if (tp->urg_data && !after(ptr, tp->urg_seq))
5331 /* Tell the world about our new urgent pointer. */
5334 /* We may be adding urgent data when the last byte read was
5335 * urgent. To do this requires some care. We cannot just ignore
5336 * tp->copied_seq since we would read the last urgent byte again
5337 * as data, nor can we alter copied_seq until this data arrives
5338 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5340 * NOTE. Double Dutch. Rendering to plain English: author of comment
5341 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5342 * and expect that both A and B disappear from stream. This is _wrong_.
5343 * Though this happens in BSD with high probability, this is occasional.
5344 * Any application relying on this is buggy. Note also, that fix "works"
5345 * only in this artificial test. Insert some normal data between A and B and we will
5346 * decline of BSD again. Verdict: it is better to remove to trap
5349 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5350 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5351 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5353 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5354 __skb_unlink(skb, &sk->sk_receive_queue);
5359 tp->urg_data = TCP_URG_NOTYET;
5360 WRITE_ONCE(tp->urg_seq, ptr);
5362 /* Disable header prediction. */
5366 /* This is the 'fast' part of urgent handling. */
5367 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
5369 struct tcp_sock *tp = tcp_sk(sk);
5371 /* Check if we get a new urgent pointer - normally not. */
5373 tcp_check_urg(sk, th);
5375 /* Do we wait for any urgent data? - normally not... */
5376 if (tp->urg_data == TCP_URG_NOTYET) {
5377 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5380 /* Is the urgent pointer pointing into this packet? */
5381 if (ptr < skb->len) {
5383 if (skb_copy_bits(skb, ptr, &tmp, 1))
5385 tp->urg_data = TCP_URG_VALID | tmp;
5386 if (!sock_flag(sk, SOCK_DEAD))
5387 sk->sk_data_ready(sk);
5392 /* Accept RST for rcv_nxt - 1 after a FIN.
5393 * When tcp connections are abruptly terminated from Mac OSX (via ^C), a
5394 * FIN is sent followed by a RST packet. The RST is sent with the same
5395 * sequence number as the FIN, and thus according to RFC 5961 a challenge
5396 * ACK should be sent. However, Mac OSX rate limits replies to challenge
5397 * ACKs on the closed socket. In addition middleboxes can drop either the
5398 * challenge ACK or a subsequent RST.
5400 static bool tcp_reset_check(const struct sock *sk, const struct sk_buff *skb)
5402 struct tcp_sock *tp = tcp_sk(sk);
5404 return unlikely(TCP_SKB_CB(skb)->seq == (tp->rcv_nxt - 1) &&
5405 (1 << sk->sk_state) & (TCPF_CLOSE_WAIT | TCPF_LAST_ACK |
5409 /* Does PAWS and seqno based validation of an incoming segment, flags will
5410 * play significant role here.
5412 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5413 const struct tcphdr *th, int syn_inerr)
5415 struct tcp_sock *tp = tcp_sk(sk);
5416 bool rst_seq_match = false;
5418 /* RFC1323: H1. Apply PAWS check first. */
5419 if (tcp_fast_parse_options(sock_net(sk), skb, th, tp) &&
5420 tp->rx_opt.saw_tstamp &&
5421 tcp_paws_discard(sk, skb)) {
5423 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5424 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5425 LINUX_MIB_TCPACKSKIPPEDPAWS,
5426 &tp->last_oow_ack_time))
5427 tcp_send_dupack(sk, skb);
5430 /* Reset is accepted even if it did not pass PAWS. */
5433 /* Step 1: check sequence number */
5434 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5435 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5436 * (RST) segments are validated by checking their SEQ-fields."
5437 * And page 69: "If an incoming segment is not acceptable,
5438 * an acknowledgment should be sent in reply (unless the RST
5439 * bit is set, if so drop the segment and return)".
5444 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5445 LINUX_MIB_TCPACKSKIPPEDSEQ,
5446 &tp->last_oow_ack_time))
5447 tcp_send_dupack(sk, skb);
5448 } else if (tcp_reset_check(sk, skb)) {
5454 /* Step 2: check RST bit */
5456 /* RFC 5961 3.2 (extend to match against (RCV.NXT - 1) after a
5457 * FIN and SACK too if available):
5458 * If seq num matches RCV.NXT or (RCV.NXT - 1) after a FIN, or
5459 * the right-most SACK block,
5461 * RESET the connection
5463 * Send a challenge ACK
5465 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt ||
5466 tcp_reset_check(sk, skb)) {
5467 rst_seq_match = true;
5468 } else if (tcp_is_sack(tp) && tp->rx_opt.num_sacks > 0) {
5469 struct tcp_sack_block *sp = &tp->selective_acks[0];
5470 int max_sack = sp[0].end_seq;
5473 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;
5475 max_sack = after(sp[this_sack].end_seq,
5477 sp[this_sack].end_seq : max_sack;
5480 if (TCP_SKB_CB(skb)->seq == max_sack)
5481 rst_seq_match = true;
5487 /* Disable TFO if RST is out-of-order
5488 * and no data has been received
5489 * for current active TFO socket
5491 if (tp->syn_fastopen && !tp->data_segs_in &&
5492 sk->sk_state == TCP_ESTABLISHED)
5493 tcp_fastopen_active_disable(sk);
5494 tcp_send_challenge_ack(sk, skb);
5499 /* step 3: check security and precedence [ignored] */
5501 /* step 4: Check for a SYN
5502 * RFC 5961 4.2 : Send a challenge ack
5507 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5508 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5509 tcp_send_challenge_ack(sk, skb);
5521 * TCP receive function for the ESTABLISHED state.
5523 * It is split into a fast path and a slow path. The fast path is
5525 * - A zero window was announced from us - zero window probing
5526 * is only handled properly in the slow path.
5527 * - Out of order segments arrived.
5528 * - Urgent data is expected.
5529 * - There is no buffer space left
5530 * - Unexpected TCP flags/window values/header lengths are received
5531 * (detected by checking the TCP header against pred_flags)
5532 * - Data is sent in both directions. Fast path only supports pure senders
5533 * or pure receivers (this means either the sequence number or the ack
5534 * value must stay constant)
5535 * - Unexpected TCP option.
5537 * When these conditions are not satisfied it drops into a standard
5538 * receive procedure patterned after RFC793 to handle all cases.
5539 * The first three cases are guaranteed by proper pred_flags setting,
5540 * the rest is checked inline. Fast processing is turned on in
5541 * tcp_data_queue when everything is OK.
5543 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb)
5545 const struct tcphdr *th = (const struct tcphdr *)skb->data;
5546 struct tcp_sock *tp = tcp_sk(sk);
5547 unsigned int len = skb->len;
5549 /* TCP congestion window tracking */
5550 trace_tcp_probe(sk, skb);
5552 tcp_mstamp_refresh(tp);
5553 if (unlikely(!sk->sk_rx_dst))
5554 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
5556 * Header prediction.
5557 * The code loosely follows the one in the famous
5558 * "30 instruction TCP receive" Van Jacobson mail.
5560 * Van's trick is to deposit buffers into socket queue
5561 * on a device interrupt, to call tcp_recv function
5562 * on the receive process context and checksum and copy
5563 * the buffer to user space. smart...
5565 * Our current scheme is not silly either but we take the
5566 * extra cost of the net_bh soft interrupt processing...
5567 * We do checksum and copy also but from device to kernel.
5570 tp->rx_opt.saw_tstamp = 0;
5572 /* pred_flags is 0xS?10 << 16 + snd_wnd
5573 * if header_prediction is to be made
5574 * 'S' will always be tp->tcp_header_len >> 2
5575 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5576 * turn it off (when there are holes in the receive
5577 * space for instance)
5578 * PSH flag is ignored.
5581 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5582 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5583 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5584 int tcp_header_len = tp->tcp_header_len;
5586 /* Timestamp header prediction: tcp_header_len
5587 * is automatically equal to th->doff*4 due to pred_flags
5591 /* Check timestamp */
5592 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5593 /* No? Slow path! */
5594 if (!tcp_parse_aligned_timestamp(tp, th))
5597 /* If PAWS failed, check it more carefully in slow path */
5598 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5601 /* DO NOT update ts_recent here, if checksum fails
5602 * and timestamp was corrupted part, it will result
5603 * in a hung connection since we will drop all
5604 * future packets due to the PAWS test.
5608 if (len <= tcp_header_len) {
5609 /* Bulk data transfer: sender */
5610 if (len == tcp_header_len) {
5611 /* Predicted packet is in window by definition.
5612 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5613 * Hence, check seq<=rcv_wup reduces to:
5615 if (tcp_header_len ==
5616 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5617 tp->rcv_nxt == tp->rcv_wup)
5618 tcp_store_ts_recent(tp);
5620 /* We know that such packets are checksummed
5623 tcp_ack(sk, skb, 0);
5625 tcp_data_snd_check(sk);
5626 /* When receiving pure ack in fast path, update
5627 * last ts ecr directly instead of calling
5628 * tcp_rcv_rtt_measure_ts()
5630 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr;
5632 } else { /* Header too small */
5633 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5638 bool fragstolen = false;
5640 if (tcp_checksum_complete(skb))
5643 if ((int)skb->truesize > sk->sk_forward_alloc)
5646 /* Predicted packet is in window by definition.
5647 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5648 * Hence, check seq<=rcv_wup reduces to:
5650 if (tcp_header_len ==
5651 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5652 tp->rcv_nxt == tp->rcv_wup)
5653 tcp_store_ts_recent(tp);
5655 tcp_rcv_rtt_measure_ts(sk, skb);
5657 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPHITS);
5659 /* Bulk data transfer: receiver */
5660 __skb_pull(skb, tcp_header_len);
5661 eaten = tcp_queue_rcv(sk, skb, &fragstolen);
5663 tcp_event_data_recv(sk, skb);
5665 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5666 /* Well, only one small jumplet in fast path... */
5667 tcp_ack(sk, skb, FLAG_DATA);
5668 tcp_data_snd_check(sk);
5669 if (!inet_csk_ack_scheduled(sk))
5673 __tcp_ack_snd_check(sk, 0);
5676 kfree_skb_partial(skb, fragstolen);
5683 if (len < (th->doff << 2) || tcp_checksum_complete(skb))
5686 if (!th->ack && !th->rst && !th->syn)
5690 * Standard slow path.
5693 if (!tcp_validate_incoming(sk, skb, th, 1))
5697 if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0)
5700 tcp_rcv_rtt_measure_ts(sk, skb);
5702 /* Process urgent data. */
5703 tcp_urg(sk, skb, th);
5705 /* step 7: process the segment text */
5706 tcp_data_queue(sk, skb);
5708 tcp_data_snd_check(sk);
5709 tcp_ack_snd_check(sk);
5713 TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS);
5714 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5719 EXPORT_SYMBOL(tcp_rcv_established);
5721 void tcp_init_transfer(struct sock *sk, int bpf_op)
5723 struct inet_connection_sock *icsk = inet_csk(sk);
5724 struct tcp_sock *tp = tcp_sk(sk);
5727 icsk->icsk_af_ops->rebuild_header(sk);
5728 tcp_init_metrics(sk);
5730 /* Initialize the congestion window to start the transfer.
5731 * Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
5732 * retransmitted. In light of RFC6298 more aggressive 1sec
5733 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
5734 * retransmission has occurred.
5736 if (tp->total_retrans > 1 && tp->undo_marker)
5739 tp->snd_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk));
5740 tp->snd_cwnd_stamp = tcp_jiffies32;
5742 tcp_call_bpf(sk, bpf_op, 0, NULL);
5743 tcp_init_congestion_control(sk);
5744 tcp_init_buffer_space(sk);
5747 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5749 struct tcp_sock *tp = tcp_sk(sk);
5750 struct inet_connection_sock *icsk = inet_csk(sk);
5752 tcp_set_state(sk, TCP_ESTABLISHED);
5753 icsk->icsk_ack.lrcvtime = tcp_jiffies32;
5756 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
5757 security_inet_conn_established(sk, skb);
5758 sk_mark_napi_id(sk, skb);
5761 tcp_init_transfer(sk, BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB);
5763 /* Prevent spurious tcp_cwnd_restart() on first data
5766 tp->lsndtime = tcp_jiffies32;
5768 if (sock_flag(sk, SOCK_KEEPOPEN))
5769 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5771 if (!tp->rx_opt.snd_wscale)
5772 __tcp_fast_path_on(tp, tp->snd_wnd);
5777 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
5778 struct tcp_fastopen_cookie *cookie)
5780 struct tcp_sock *tp = tcp_sk(sk);
5781 struct sk_buff *data = tp->syn_data ? tcp_rtx_queue_head(sk) : NULL;
5782 u16 mss = tp->rx_opt.mss_clamp, try_exp = 0;
5783 bool syn_drop = false;
5785 if (mss == tp->rx_opt.user_mss) {
5786 struct tcp_options_received opt;
5788 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5789 tcp_clear_options(&opt);
5790 opt.user_mss = opt.mss_clamp = 0;
5791 tcp_parse_options(sock_net(sk), synack, &opt, 0, NULL);
5792 mss = opt.mss_clamp;
5795 if (!tp->syn_fastopen) {
5796 /* Ignore an unsolicited cookie */
5798 } else if (tp->total_retrans) {
5799 /* SYN timed out and the SYN-ACK neither has a cookie nor
5800 * acknowledges data. Presumably the remote received only
5801 * the retransmitted (regular) SYNs: either the original
5802 * SYN-data or the corresponding SYN-ACK was dropped.
5804 syn_drop = (cookie->len < 0 && data);
5805 } else if (cookie->len < 0 && !tp->syn_data) {
5806 /* We requested a cookie but didn't get it. If we did not use
5807 * the (old) exp opt format then try so next time (try_exp=1).
5808 * Otherwise we go back to use the RFC7413 opt (try_exp=2).
5810 try_exp = tp->syn_fastopen_exp ? 2 : 1;
5813 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp);
5815 if (data) { /* Retransmit unacked data in SYN */
5816 skb_rbtree_walk_from(data) {
5817 if (__tcp_retransmit_skb(sk, data, 1))
5821 NET_INC_STATS(sock_net(sk),
5822 LINUX_MIB_TCPFASTOPENACTIVEFAIL);
5825 tp->syn_data_acked = tp->syn_data;
5826 if (tp->syn_data_acked) {
5827 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
5828 /* SYN-data is counted as two separate packets in tcp_ack() */
5829 if (tp->delivered > 1)
5833 tcp_fastopen_add_skb(sk, synack);
5838 static void smc_check_reset_syn(struct tcp_sock *tp)
5840 #if IS_ENABLED(CONFIG_SMC)
5841 if (static_branch_unlikely(&tcp_have_smc)) {
5842 if (tp->syn_smc && !tp->rx_opt.smc_ok)
5848 static void tcp_try_undo_spurious_syn(struct sock *sk)
5850 struct tcp_sock *tp = tcp_sk(sk);
5853 /* undo_marker is set when SYN or SYNACK times out. The timeout is
5854 * spurious if the ACK's timestamp option echo value matches the
5855 * original SYN timestamp.
5857 syn_stamp = tp->retrans_stamp;
5858 if (tp->undo_marker && syn_stamp && tp->rx_opt.saw_tstamp &&
5859 syn_stamp == tp->rx_opt.rcv_tsecr)
5860 tp->undo_marker = 0;
5863 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5864 const struct tcphdr *th)
5866 struct inet_connection_sock *icsk = inet_csk(sk);
5867 struct tcp_sock *tp = tcp_sk(sk);
5868 struct tcp_fastopen_cookie foc = { .len = -1 };
5869 int saved_clamp = tp->rx_opt.mss_clamp;
5872 tcp_parse_options(sock_net(sk), skb, &tp->rx_opt, 0, &foc);
5873 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
5874 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
5878 * "If the state is SYN-SENT then
5879 * first check the ACK bit
5880 * If the ACK bit is set
5881 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5882 * a reset (unless the RST bit is set, if so drop
5883 * the segment and return)"
5885 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
5886 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt))
5887 goto reset_and_undo;
5889 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5890 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5891 tcp_time_stamp(tp))) {
5892 NET_INC_STATS(sock_net(sk),
5893 LINUX_MIB_PAWSACTIVEREJECTED);
5894 goto reset_and_undo;
5897 /* Now ACK is acceptable.
5899 * "If the RST bit is set
5900 * If the ACK was acceptable then signal the user "error:
5901 * connection reset", drop the segment, enter CLOSED state,
5902 * delete TCB, and return."
5911 * "fifth, if neither of the SYN or RST bits is set then
5912 * drop the segment and return."
5918 goto discard_and_undo;
5921 * "If the SYN bit is on ...
5922 * are acceptable then ...
5923 * (our SYN has been ACKed), change the connection
5924 * state to ESTABLISHED..."
5927 tcp_ecn_rcv_synack(tp, th);
5929 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5930 tcp_try_undo_spurious_syn(sk);
5931 tcp_ack(sk, skb, FLAG_SLOWPATH);
5933 /* Ok.. it's good. Set up sequence numbers and
5934 * move to established.
5936 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1);
5937 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5939 /* RFC1323: The window in SYN & SYN/ACK segments is
5942 tp->snd_wnd = ntohs(th->window);
5944 if (!tp->rx_opt.wscale_ok) {
5945 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5946 tp->window_clamp = min(tp->window_clamp, 65535U);
5949 if (tp->rx_opt.saw_tstamp) {
5950 tp->rx_opt.tstamp_ok = 1;
5951 tp->tcp_header_len =
5952 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5953 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5954 tcp_store_ts_recent(tp);
5956 tp->tcp_header_len = sizeof(struct tcphdr);
5959 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5960 tcp_initialize_rcv_mss(sk);
5962 /* Remember, tcp_poll() does not lock socket!
5963 * Change state from SYN-SENT only after copied_seq
5964 * is initialized. */
5965 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
5967 smc_check_reset_syn(tp);
5971 tcp_finish_connect(sk, skb);
5973 fastopen_fail = (tp->syn_fastopen || tp->syn_data) &&
5974 tcp_rcv_fastopen_synack(sk, skb, &foc);
5976 if (!sock_flag(sk, SOCK_DEAD)) {
5977 sk->sk_state_change(sk);
5978 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5982 if (sk->sk_write_pending ||
5983 icsk->icsk_accept_queue.rskq_defer_accept ||
5984 inet_csk_in_pingpong_mode(sk)) {
5985 /* Save one ACK. Data will be ready after
5986 * several ticks, if write_pending is set.
5988 * It may be deleted, but with this feature tcpdumps
5989 * look so _wonderfully_ clever, that I was not able
5990 * to stand against the temptation 8) --ANK
5992 inet_csk_schedule_ack(sk);
5993 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
5994 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5995 TCP_DELACK_MAX, TCP_RTO_MAX);
6006 /* No ACK in the segment */
6010 * "If the RST bit is set
6012 * Otherwise (no ACK) drop the segment and return."
6015 goto discard_and_undo;
6019 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
6020 tcp_paws_reject(&tp->rx_opt, 0))
6021 goto discard_and_undo;
6024 /* We see SYN without ACK. It is attempt of
6025 * simultaneous connect with crossed SYNs.
6026 * Particularly, it can be connect to self.
6028 tcp_set_state(sk, TCP_SYN_RECV);
6030 if (tp->rx_opt.saw_tstamp) {
6031 tp->rx_opt.tstamp_ok = 1;
6032 tcp_store_ts_recent(tp);
6033 tp->tcp_header_len =
6034 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
6036 tp->tcp_header_len = sizeof(struct tcphdr);
6039 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1);
6040 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6041 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
6043 /* RFC1323: The window in SYN & SYN/ACK segments is
6046 tp->snd_wnd = ntohs(th->window);
6047 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
6048 tp->max_window = tp->snd_wnd;
6050 tcp_ecn_rcv_syn(tp, th);
6053 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
6054 tcp_initialize_rcv_mss(sk);
6056 tcp_send_synack(sk);
6058 /* Note, we could accept data and URG from this segment.
6059 * There are no obstacles to make this (except that we must
6060 * either change tcp_recvmsg() to prevent it from returning data
6061 * before 3WHS completes per RFC793, or employ TCP Fast Open).
6063 * However, if we ignore data in ACKless segments sometimes,
6064 * we have no reasons to accept it sometimes.
6065 * Also, seems the code doing it in step6 of tcp_rcv_state_process
6066 * is not flawless. So, discard packet for sanity.
6067 * Uncomment this return to process the data.
6074 /* "fifth, if neither of the SYN or RST bits is set then
6075 * drop the segment and return."
6079 tcp_clear_options(&tp->rx_opt);
6080 tp->rx_opt.mss_clamp = saved_clamp;
6084 tcp_clear_options(&tp->rx_opt);
6085 tp->rx_opt.mss_clamp = saved_clamp;
6089 static void tcp_rcv_synrecv_state_fastopen(struct sock *sk)
6091 struct request_sock *req;
6093 tcp_try_undo_loss(sk, false);
6095 /* Reset rtx states to prevent spurious retransmits_timed_out() */
6096 tcp_sk(sk)->retrans_stamp = 0;
6097 inet_csk(sk)->icsk_retransmits = 0;
6099 /* Once we leave TCP_SYN_RECV or TCP_FIN_WAIT_1,
6100 * we no longer need req so release it.
6102 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk,
6103 lockdep_sock_is_held(sk));
6104 reqsk_fastopen_remove(sk, req, false);
6106 /* Re-arm the timer because data may have been sent out.
6107 * This is similar to the regular data transmission case
6108 * when new data has just been ack'ed.
6110 * (TFO) - we could try to be more aggressive and
6111 * retransmitting any data sooner based on when they
6118 * This function implements the receiving procedure of RFC 793 for
6119 * all states except ESTABLISHED and TIME_WAIT.
6120 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
6121 * address independent.
6124 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb)
6126 struct tcp_sock *tp = tcp_sk(sk);
6127 struct inet_connection_sock *icsk = inet_csk(sk);
6128 const struct tcphdr *th = tcp_hdr(skb);
6129 struct request_sock *req;
6133 switch (sk->sk_state) {
6147 /* It is possible that we process SYN packets from backlog,
6148 * so we need to make sure to disable BH and RCU right there.
6152 acceptable = icsk->icsk_af_ops->conn_request(sk, skb) >= 0;
6164 tp->rx_opt.saw_tstamp = 0;
6165 tcp_mstamp_refresh(tp);
6166 queued = tcp_rcv_synsent_state_process(sk, skb, th);
6170 /* Do step6 onward by hand. */
6171 tcp_urg(sk, skb, th);
6173 tcp_data_snd_check(sk);
6177 tcp_mstamp_refresh(tp);
6178 tp->rx_opt.saw_tstamp = 0;
6179 req = rcu_dereference_protected(tp->fastopen_rsk,
6180 lockdep_sock_is_held(sk));
6184 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
6185 sk->sk_state != TCP_FIN_WAIT1);
6187 if (!tcp_check_req(sk, skb, req, true, &req_stolen))
6191 if (!th->ack && !th->rst && !th->syn)
6194 if (!tcp_validate_incoming(sk, skb, th, 0))
6197 /* step 5: check the ACK field */
6198 acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
6199 FLAG_UPDATE_TS_RECENT |
6200 FLAG_NO_CHALLENGE_ACK) > 0;
6203 if (sk->sk_state == TCP_SYN_RECV)
6204 return 1; /* send one RST */
6205 tcp_send_challenge_ack(sk, skb);
6208 switch (sk->sk_state) {
6210 tp->delivered++; /* SYN-ACK delivery isn't tracked in tcp_ack */
6212 tcp_synack_rtt_meas(sk, req);
6215 tcp_rcv_synrecv_state_fastopen(sk);
6217 tcp_try_undo_spurious_syn(sk);
6218 tp->retrans_stamp = 0;
6219 tcp_init_transfer(sk, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB);
6220 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6223 tcp_set_state(sk, TCP_ESTABLISHED);
6224 sk->sk_state_change(sk);
6226 /* Note, that this wakeup is only for marginal crossed SYN case.
6227 * Passively open sockets are not waked up, because
6228 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
6231 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
6233 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
6234 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
6235 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
6237 if (tp->rx_opt.tstamp_ok)
6238 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
6240 if (!inet_csk(sk)->icsk_ca_ops->cong_control)
6241 tcp_update_pacing_rate(sk);
6243 /* Prevent spurious tcp_cwnd_restart() on first data packet */
6244 tp->lsndtime = tcp_jiffies32;
6246 tcp_initialize_rcv_mss(sk);
6247 tcp_fast_path_on(tp);
6250 case TCP_FIN_WAIT1: {
6254 tcp_rcv_synrecv_state_fastopen(sk);
6256 if (tp->snd_una != tp->write_seq)
6259 tcp_set_state(sk, TCP_FIN_WAIT2);
6260 sk->sk_shutdown |= SEND_SHUTDOWN;
6264 if (!sock_flag(sk, SOCK_DEAD)) {
6265 /* Wake up lingering close() */
6266 sk->sk_state_change(sk);
6270 if (tp->linger2 < 0) {
6272 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6275 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6276 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6277 /* Receive out of order FIN after close() */
6278 if (tp->syn_fastopen && th->fin)
6279 tcp_fastopen_active_disable(sk);
6281 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6285 tmo = tcp_fin_time(sk);
6286 if (tmo > TCP_TIMEWAIT_LEN) {
6287 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
6288 } else if (th->fin || sock_owned_by_user(sk)) {
6289 /* Bad case. We could lose such FIN otherwise.
6290 * It is not a big problem, but it looks confusing
6291 * and not so rare event. We still can lose it now,
6292 * if it spins in bh_lock_sock(), but it is really
6295 inet_csk_reset_keepalive_timer(sk, tmo);
6297 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
6304 if (tp->snd_una == tp->write_seq) {
6305 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
6311 if (tp->snd_una == tp->write_seq) {
6312 tcp_update_metrics(sk);
6319 /* step 6: check the URG bit */
6320 tcp_urg(sk, skb, th);
6322 /* step 7: process the segment text */
6323 switch (sk->sk_state) {
6324 case TCP_CLOSE_WAIT:
6327 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
6332 /* RFC 793 says to queue data in these states,
6333 * RFC 1122 says we MUST send a reset.
6334 * BSD 4.4 also does reset.
6336 if (sk->sk_shutdown & RCV_SHUTDOWN) {
6337 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6338 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6339 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6345 case TCP_ESTABLISHED:
6346 tcp_data_queue(sk, skb);
6351 /* tcp_data could move socket to TIME-WAIT */
6352 if (sk->sk_state != TCP_CLOSE) {
6353 tcp_data_snd_check(sk);
6354 tcp_ack_snd_check(sk);
6363 EXPORT_SYMBOL(tcp_rcv_state_process);
6365 static inline void pr_drop_req(struct request_sock *req, __u16 port, int family)
6367 struct inet_request_sock *ireq = inet_rsk(req);
6369 if (family == AF_INET)
6370 net_dbg_ratelimited("drop open request from %pI4/%u\n",
6371 &ireq->ir_rmt_addr, port);
6372 #if IS_ENABLED(CONFIG_IPV6)
6373 else if (family == AF_INET6)
6374 net_dbg_ratelimited("drop open request from %pI6/%u\n",
6375 &ireq->ir_v6_rmt_addr, port);
6379 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
6381 * If we receive a SYN packet with these bits set, it means a
6382 * network is playing bad games with TOS bits. In order to
6383 * avoid possible false congestion notifications, we disable
6384 * TCP ECN negotiation.
6386 * Exception: tcp_ca wants ECN. This is required for DCTCP
6387 * congestion control: Linux DCTCP asserts ECT on all packets,
6388 * including SYN, which is most optimal solution; however,
6389 * others, such as FreeBSD do not.
6391 * Exception: At least one of the reserved bits of the TCP header (th->res1) is
6392 * set, indicating the use of a future TCP extension (such as AccECN). See
6393 * RFC8311 §4.3 which updates RFC3168 to allow the development of such
6396 static void tcp_ecn_create_request(struct request_sock *req,
6397 const struct sk_buff *skb,
6398 const struct sock *listen_sk,
6399 const struct dst_entry *dst)
6401 const struct tcphdr *th = tcp_hdr(skb);
6402 const struct net *net = sock_net(listen_sk);
6403 bool th_ecn = th->ece && th->cwr;
6410 ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield);
6411 ecn_ok_dst = dst_feature(dst, DST_FEATURE_ECN_MASK);
6412 ecn_ok = net->ipv4.sysctl_tcp_ecn || ecn_ok_dst;
6414 if (((!ect || th->res1) && ecn_ok) || tcp_ca_needs_ecn(listen_sk) ||
6415 (ecn_ok_dst & DST_FEATURE_ECN_CA) ||
6416 tcp_bpf_ca_needs_ecn((struct sock *)req))
6417 inet_rsk(req)->ecn_ok = 1;
6420 static void tcp_openreq_init(struct request_sock *req,
6421 const struct tcp_options_received *rx_opt,
6422 struct sk_buff *skb, const struct sock *sk)
6424 struct inet_request_sock *ireq = inet_rsk(req);
6426 req->rsk_rcv_wnd = 0; /* So that tcp_send_synack() knows! */
6428 tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq;
6429 tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
6430 tcp_rsk(req)->snt_synack = 0;
6431 tcp_rsk(req)->last_oow_ack_time = 0;
6432 req->mss = rx_opt->mss_clamp;
6433 req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0;
6434 ireq->tstamp_ok = rx_opt->tstamp_ok;
6435 ireq->sack_ok = rx_opt->sack_ok;
6436 ireq->snd_wscale = rx_opt->snd_wscale;
6437 ireq->wscale_ok = rx_opt->wscale_ok;
6440 ireq->ir_rmt_port = tcp_hdr(skb)->source;
6441 ireq->ir_num = ntohs(tcp_hdr(skb)->dest);
6442 ireq->ir_mark = inet_request_mark(sk, skb);
6443 #if IS_ENABLED(CONFIG_SMC)
6444 ireq->smc_ok = rx_opt->smc_ok;
6448 struct request_sock *inet_reqsk_alloc(const struct request_sock_ops *ops,
6449 struct sock *sk_listener,
6450 bool attach_listener)
6452 struct request_sock *req = reqsk_alloc(ops, sk_listener,
6456 struct inet_request_sock *ireq = inet_rsk(req);
6458 ireq->ireq_opt = NULL;
6459 #if IS_ENABLED(CONFIG_IPV6)
6460 ireq->pktopts = NULL;
6462 atomic64_set(&ireq->ir_cookie, 0);
6463 ireq->ireq_state = TCP_NEW_SYN_RECV;
6464 write_pnet(&ireq->ireq_net, sock_net(sk_listener));
6465 ireq->ireq_family = sk_listener->sk_family;
6470 EXPORT_SYMBOL(inet_reqsk_alloc);
6473 * Return true if a syncookie should be sent
6475 static bool tcp_syn_flood_action(const struct sock *sk, const char *proto)
6477 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
6478 const char *msg = "Dropping request";
6479 bool want_cookie = false;
6480 struct net *net = sock_net(sk);
6482 #ifdef CONFIG_SYN_COOKIES
6483 if (net->ipv4.sysctl_tcp_syncookies) {
6484 msg = "Sending cookies";
6486 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES);
6489 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP);
6491 if (!queue->synflood_warned &&
6492 net->ipv4.sysctl_tcp_syncookies != 2 &&
6493 xchg(&queue->synflood_warned, 1) == 0)
6494 net_info_ratelimited("%s: Possible SYN flooding on port %d. %s. Check SNMP counters.\n",
6495 proto, sk->sk_num, msg);
6500 static void tcp_reqsk_record_syn(const struct sock *sk,
6501 struct request_sock *req,
6502 const struct sk_buff *skb)
6504 if (tcp_sk(sk)->save_syn) {
6505 u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb);
6508 copy = kmalloc(len + sizeof(u32), GFP_ATOMIC);
6511 memcpy(©[1], skb_network_header(skb), len);
6512 req->saved_syn = copy;
6517 /* If a SYN cookie is required and supported, returns a clamped MSS value to be
6518 * used for SYN cookie generation.
6520 u16 tcp_get_syncookie_mss(struct request_sock_ops *rsk_ops,
6521 const struct tcp_request_sock_ops *af_ops,
6522 struct sock *sk, struct tcphdr *th)
6524 struct tcp_sock *tp = tcp_sk(sk);
6527 if (sock_net(sk)->ipv4.sysctl_tcp_syncookies != 2 &&
6528 !inet_csk_reqsk_queue_is_full(sk))
6531 if (!tcp_syn_flood_action(sk, rsk_ops->slab_name))
6534 if (sk_acceptq_is_full(sk)) {
6535 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
6539 mss = tcp_parse_mss_option(th, tp->rx_opt.user_mss);
6541 mss = af_ops->mss_clamp;
6545 EXPORT_SYMBOL_GPL(tcp_get_syncookie_mss);
6547 int tcp_conn_request(struct request_sock_ops *rsk_ops,
6548 const struct tcp_request_sock_ops *af_ops,
6549 struct sock *sk, struct sk_buff *skb)
6551 struct tcp_fastopen_cookie foc = { .len = -1 };
6552 __u32 isn = TCP_SKB_CB(skb)->tcp_tw_isn;
6553 struct tcp_options_received tmp_opt;
6554 struct tcp_sock *tp = tcp_sk(sk);
6555 struct net *net = sock_net(sk);
6556 struct sock *fastopen_sk = NULL;
6557 struct request_sock *req;
6558 bool want_cookie = false;
6559 struct dst_entry *dst;
6562 /* TW buckets are converted to open requests without
6563 * limitations, they conserve resources and peer is
6564 * evidently real one.
6566 if ((net->ipv4.sysctl_tcp_syncookies == 2 ||
6567 inet_csk_reqsk_queue_is_full(sk)) && !isn) {
6568 want_cookie = tcp_syn_flood_action(sk, rsk_ops->slab_name);
6573 if (sk_acceptq_is_full(sk)) {
6574 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
6578 req = inet_reqsk_alloc(rsk_ops, sk, !want_cookie);
6582 tcp_rsk(req)->af_specific = af_ops;
6583 tcp_rsk(req)->ts_off = 0;
6585 tcp_clear_options(&tmp_opt);
6586 tmp_opt.mss_clamp = af_ops->mss_clamp;
6587 tmp_opt.user_mss = tp->rx_opt.user_mss;
6588 tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0,
6589 want_cookie ? NULL : &foc);
6591 if (want_cookie && !tmp_opt.saw_tstamp)
6592 tcp_clear_options(&tmp_opt);
6594 if (IS_ENABLED(CONFIG_SMC) && want_cookie)
6597 tmp_opt.tstamp_ok = tmp_opt.saw_tstamp;
6598 tcp_openreq_init(req, &tmp_opt, skb, sk);
6599 inet_rsk(req)->no_srccheck = inet_sk(sk)->transparent;
6601 /* Note: tcp_v6_init_req() might override ir_iif for link locals */
6602 inet_rsk(req)->ir_iif = inet_request_bound_dev_if(sk, skb);
6604 af_ops->init_req(req, sk, skb);
6606 if (security_inet_conn_request(sk, skb, req))
6609 if (tmp_opt.tstamp_ok)
6610 tcp_rsk(req)->ts_off = af_ops->init_ts_off(net, skb);
6612 dst = af_ops->route_req(sk, &fl, req);
6616 if (!want_cookie && !isn) {
6617 /* Kill the following clause, if you dislike this way. */
6618 if (!net->ipv4.sysctl_tcp_syncookies &&
6619 (net->ipv4.sysctl_max_syn_backlog - inet_csk_reqsk_queue_len(sk) <
6620 (net->ipv4.sysctl_max_syn_backlog >> 2)) &&
6621 !tcp_peer_is_proven(req, dst)) {
6622 /* Without syncookies last quarter of
6623 * backlog is filled with destinations,
6624 * proven to be alive.
6625 * It means that we continue to communicate
6626 * to destinations, already remembered
6627 * to the moment of synflood.
6629 pr_drop_req(req, ntohs(tcp_hdr(skb)->source),
6631 goto drop_and_release;
6634 isn = af_ops->init_seq(skb);
6637 tcp_ecn_create_request(req, skb, sk, dst);
6640 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss);
6641 req->cookie_ts = tmp_opt.tstamp_ok;
6642 if (!tmp_opt.tstamp_ok)
6643 inet_rsk(req)->ecn_ok = 0;
6646 tcp_rsk(req)->snt_isn = isn;
6647 tcp_rsk(req)->txhash = net_tx_rndhash();
6648 tcp_openreq_init_rwin(req, sk, dst);
6649 sk_rx_queue_set(req_to_sk(req), skb);
6651 tcp_reqsk_record_syn(sk, req, skb);
6652 fastopen_sk = tcp_try_fastopen(sk, skb, req, &foc, dst);
6655 af_ops->send_synack(fastopen_sk, dst, &fl, req,
6656 &foc, TCP_SYNACK_FASTOPEN);
6657 /* Add the child socket directly into the accept queue */
6658 if (!inet_csk_reqsk_queue_add(sk, req, fastopen_sk)) {
6659 reqsk_fastopen_remove(fastopen_sk, req, false);
6660 bh_unlock_sock(fastopen_sk);
6661 sock_put(fastopen_sk);
6664 sk->sk_data_ready(sk);
6665 bh_unlock_sock(fastopen_sk);
6666 sock_put(fastopen_sk);
6668 tcp_rsk(req)->tfo_listener = false;
6670 inet_csk_reqsk_queue_hash_add(sk, req,
6671 tcp_timeout_init((struct sock *)req));
6672 af_ops->send_synack(sk, dst, &fl, req, &foc,
6673 !want_cookie ? TCP_SYNACK_NORMAL :
6691 EXPORT_SYMBOL(tcp_conn_request);