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 static 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;
234 /* Send ACKs quickly, if "quick" count is not exhausted
235 * and the session is not interactive.
238 static bool tcp_in_quickack_mode(struct sock *sk)
240 const struct inet_connection_sock *icsk = inet_csk(sk);
241 const struct dst_entry *dst = __sk_dst_get(sk);
243 return (dst && dst_metric(dst, RTAX_QUICKACK)) ||
244 (icsk->icsk_ack.quick && !inet_csk_in_pingpong_mode(sk));
247 static void tcp_ecn_queue_cwr(struct tcp_sock *tp)
249 if (tp->ecn_flags & TCP_ECN_OK)
250 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
253 static void tcp_ecn_accept_cwr(struct sock *sk, const struct sk_buff *skb)
255 if (tcp_hdr(skb)->cwr) {
256 tcp_sk(sk)->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
258 /* If the sender is telling us it has entered CWR, then its
259 * cwnd may be very low (even just 1 packet), so we should ACK
262 if (TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq)
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 WRITE_ONCE(sk->sk_sndbuf,
363 min(sndmem, sock_net(sk)->ipv4.sysctl_tcp_wmem[2]));
366 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
368 * All tcp_full_space() is split to two parts: "network" buffer, allocated
369 * forward and advertised in receiver window (tp->rcv_wnd) and
370 * "application buffer", required to isolate scheduling/application
371 * latencies from network.
372 * window_clamp is maximal advertised window. It can be less than
373 * tcp_full_space(), in this case tcp_full_space() - window_clamp
374 * is reserved for "application" buffer. The less window_clamp is
375 * the smoother our behaviour from viewpoint of network, but the lower
376 * throughput and the higher sensitivity of the connection to losses. 8)
378 * rcv_ssthresh is more strict window_clamp used at "slow start"
379 * phase to predict further behaviour of this connection.
380 * It is used for two goals:
381 * - to enforce header prediction at sender, even when application
382 * requires some significant "application buffer". It is check #1.
383 * - to prevent pruning of receive queue because of misprediction
384 * of receiver window. Check #2.
386 * The scheme does not work when sender sends good segments opening
387 * window and then starts to feed us spaghetti. But it should work
388 * in common situations. Otherwise, we have to rely on queue collapsing.
391 /* Slow part of check#2. */
392 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
394 struct tcp_sock *tp = tcp_sk(sk);
396 int truesize = tcp_win_from_space(sk, skb->truesize) >> 1;
397 int window = tcp_win_from_space(sk, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]) >> 1;
399 while (tp->rcv_ssthresh <= window) {
400 if (truesize <= skb->len)
401 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
409 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
411 struct tcp_sock *tp = tcp_sk(sk);
414 room = min_t(int, tp->window_clamp, tcp_space(sk)) - tp->rcv_ssthresh;
417 if (room > 0 && !tcp_under_memory_pressure(sk)) {
420 /* Check #2. Increase window, if skb with such overhead
421 * will fit to rcvbuf in future.
423 if (tcp_win_from_space(sk, skb->truesize) <= skb->len)
424 incr = 2 * tp->advmss;
426 incr = __tcp_grow_window(sk, skb);
429 incr = max_t(int, incr, 2 * skb->len);
430 tp->rcv_ssthresh += min(room, incr);
431 inet_csk(sk)->icsk_ack.quick |= 1;
436 /* 3. Try to fixup all. It is made immediately after connection enters
439 void tcp_init_buffer_space(struct sock *sk)
441 int tcp_app_win = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_app_win);
442 struct tcp_sock *tp = tcp_sk(sk);
445 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
446 tcp_sndbuf_expand(sk);
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;
471 tp->rcvq_space.space = min3(tp->rcv_ssthresh, tp->rcv_wnd,
472 (u32)TCP_INIT_CWND * tp->advmss);
475 /* 4. Recalculate window clamp after socket hit its memory bounds. */
476 static void tcp_clamp_window(struct sock *sk)
478 struct tcp_sock *tp = tcp_sk(sk);
479 struct inet_connection_sock *icsk = inet_csk(sk);
480 struct net *net = sock_net(sk);
482 icsk->icsk_ack.quick = 0;
484 if (sk->sk_rcvbuf < net->ipv4.sysctl_tcp_rmem[2] &&
485 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
486 !tcp_under_memory_pressure(sk) &&
487 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
488 WRITE_ONCE(sk->sk_rcvbuf,
489 min(atomic_read(&sk->sk_rmem_alloc),
490 net->ipv4.sysctl_tcp_rmem[2]));
492 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
493 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
496 /* Initialize RCV_MSS value.
497 * RCV_MSS is an our guess about MSS used by the peer.
498 * We haven't any direct information about the MSS.
499 * It's better to underestimate the RCV_MSS rather than overestimate.
500 * Overestimations make us ACKing less frequently than needed.
501 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
503 void tcp_initialize_rcv_mss(struct sock *sk)
505 const struct tcp_sock *tp = tcp_sk(sk);
506 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
508 hint = min(hint, tp->rcv_wnd / 2);
509 hint = min(hint, TCP_MSS_DEFAULT);
510 hint = max(hint, TCP_MIN_MSS);
512 inet_csk(sk)->icsk_ack.rcv_mss = hint;
514 EXPORT_SYMBOL(tcp_initialize_rcv_mss);
516 /* Receiver "autotuning" code.
518 * The algorithm for RTT estimation w/o timestamps is based on
519 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
520 * <http://public.lanl.gov/radiant/pubs.html#DRS>
522 * More detail on this code can be found at
523 * <http://staff.psc.edu/jheffner/>,
524 * though this reference is out of date. A new paper
527 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
529 u32 new_sample = tp->rcv_rtt_est.rtt_us;
532 if (new_sample != 0) {
533 /* If we sample in larger samples in the non-timestamp
534 * case, we could grossly overestimate the RTT especially
535 * with chatty applications or bulk transfer apps which
536 * are stalled on filesystem I/O.
538 * Also, since we are only going for a minimum in the
539 * non-timestamp case, we do not smooth things out
540 * else with timestamps disabled convergence takes too
544 m -= (new_sample >> 3);
552 /* No previous measure. */
556 tp->rcv_rtt_est.rtt_us = new_sample;
559 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
563 if (tp->rcv_rtt_est.time == 0)
565 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
567 delta_us = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcv_rtt_est.time);
570 tcp_rcv_rtt_update(tp, delta_us, 1);
573 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
574 tp->rcv_rtt_est.time = tp->tcp_mstamp;
577 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
578 const struct sk_buff *skb)
580 struct tcp_sock *tp = tcp_sk(sk);
582 if (tp->rx_opt.rcv_tsecr == tp->rcv_rtt_last_tsecr)
584 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr;
586 if (TCP_SKB_CB(skb)->end_seq -
587 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss) {
588 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr;
591 if (likely(delta < INT_MAX / (USEC_PER_SEC / TCP_TS_HZ))) {
594 delta_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
595 tcp_rcv_rtt_update(tp, delta_us, 0);
601 * This function should be called every time data is copied to user space.
602 * It calculates the appropriate TCP receive buffer space.
604 void tcp_rcv_space_adjust(struct sock *sk)
606 struct tcp_sock *tp = tcp_sk(sk);
610 trace_tcp_rcv_space_adjust(sk);
612 tcp_mstamp_refresh(tp);
613 time = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcvq_space.time);
614 if (time < (tp->rcv_rtt_est.rtt_us >> 3) || tp->rcv_rtt_est.rtt_us == 0)
617 /* Number of bytes copied to user in last RTT */
618 copied = tp->copied_seq - tp->rcvq_space.seq;
619 if (copied <= tp->rcvq_space.space)
623 * copied = bytes received in previous RTT, our base window
624 * To cope with packet losses, we need a 2x factor
625 * To cope with slow start, and sender growing its cwin by 100 %
626 * every RTT, we need a 4x factor, because the ACK we are sending
627 * now is for the next RTT, not the current one :
628 * <prev RTT . ><current RTT .. ><next RTT .... >
631 if (sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf &&
632 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
636 /* minimal window to cope with packet losses, assuming
637 * steady state. Add some cushion because of small variations.
639 rcvwin = ((u64)copied << 1) + 16 * tp->advmss;
641 /* Accommodate for sender rate increase (eg. slow start) */
642 grow = rcvwin * (copied - tp->rcvq_space.space);
643 do_div(grow, tp->rcvq_space.space);
644 rcvwin += (grow << 1);
646 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
647 while (tcp_win_from_space(sk, rcvmem) < tp->advmss)
650 do_div(rcvwin, tp->advmss);
651 rcvbuf = min_t(u64, rcvwin * rcvmem,
652 sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
653 if (rcvbuf > sk->sk_rcvbuf) {
654 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
656 /* Make the window clamp follow along. */
657 tp->window_clamp = tcp_win_from_space(sk, rcvbuf);
660 tp->rcvq_space.space = copied;
663 tp->rcvq_space.seq = tp->copied_seq;
664 tp->rcvq_space.time = tp->tcp_mstamp;
667 /* There is something which you must keep in mind when you analyze the
668 * behavior of the tp->ato delayed ack timeout interval. When a
669 * connection starts up, we want to ack as quickly as possible. The
670 * problem is that "good" TCP's do slow start at the beginning of data
671 * transmission. The means that until we send the first few ACK's the
672 * sender will sit on his end and only queue most of his data, because
673 * he can only send snd_cwnd unacked packets at any given time. For
674 * each ACK we send, he increments snd_cwnd and transmits more of his
677 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
679 struct tcp_sock *tp = tcp_sk(sk);
680 struct inet_connection_sock *icsk = inet_csk(sk);
683 inet_csk_schedule_ack(sk);
685 tcp_measure_rcv_mss(sk, skb);
687 tcp_rcv_rtt_measure(tp);
691 if (!icsk->icsk_ack.ato) {
692 /* The _first_ data packet received, initialize
693 * delayed ACK engine.
695 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
696 icsk->icsk_ack.ato = TCP_ATO_MIN;
698 int m = now - icsk->icsk_ack.lrcvtime;
700 if (m <= TCP_ATO_MIN / 2) {
701 /* The fastest case is the first. */
702 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
703 } else if (m < icsk->icsk_ack.ato) {
704 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
705 if (icsk->icsk_ack.ato > icsk->icsk_rto)
706 icsk->icsk_ack.ato = icsk->icsk_rto;
707 } else if (m > icsk->icsk_rto) {
708 /* Too long gap. Apparently sender failed to
709 * restart window, so that we send ACKs quickly.
711 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
715 icsk->icsk_ack.lrcvtime = now;
717 tcp_ecn_check_ce(sk, skb);
720 tcp_grow_window(sk, skb);
723 /* Called to compute a smoothed rtt estimate. The data fed to this
724 * routine either comes from timestamps, or from segments that were
725 * known _not_ to have been retransmitted [see Karn/Partridge
726 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
727 * piece by Van Jacobson.
728 * NOTE: the next three routines used to be one big routine.
729 * To save cycles in the RFC 1323 implementation it was better to break
730 * it up into three procedures. -- erics
732 static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
734 struct tcp_sock *tp = tcp_sk(sk);
735 long m = mrtt_us; /* RTT */
736 u32 srtt = tp->srtt_us;
738 /* The following amusing code comes from Jacobson's
739 * article in SIGCOMM '88. Note that rtt and mdev
740 * are scaled versions of rtt and mean deviation.
741 * This is designed to be as fast as possible
742 * m stands for "measurement".
744 * On a 1990 paper the rto value is changed to:
745 * RTO = rtt + 4 * mdev
747 * Funny. This algorithm seems to be very broken.
748 * These formulae increase RTO, when it should be decreased, increase
749 * too slowly, when it should be increased quickly, decrease too quickly
750 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
751 * does not matter how to _calculate_ it. Seems, it was trap
752 * that VJ failed to avoid. 8)
755 m -= (srtt >> 3); /* m is now error in rtt est */
756 srtt += m; /* rtt = 7/8 rtt + 1/8 new */
758 m = -m; /* m is now abs(error) */
759 m -= (tp->mdev_us >> 2); /* similar update on mdev */
760 /* This is similar to one of Eifel findings.
761 * Eifel blocks mdev updates when rtt decreases.
762 * This solution is a bit different: we use finer gain
763 * for mdev in this case (alpha*beta).
764 * Like Eifel it also prevents growth of rto,
765 * but also it limits too fast rto decreases,
766 * happening in pure Eifel.
771 m -= (tp->mdev_us >> 2); /* similar update on mdev */
773 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */
774 if (tp->mdev_us > tp->mdev_max_us) {
775 tp->mdev_max_us = tp->mdev_us;
776 if (tp->mdev_max_us > tp->rttvar_us)
777 tp->rttvar_us = tp->mdev_max_us;
779 if (after(tp->snd_una, tp->rtt_seq)) {
780 if (tp->mdev_max_us < tp->rttvar_us)
781 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
782 tp->rtt_seq = tp->snd_nxt;
783 tp->mdev_max_us = tcp_rto_min_us(sk);
788 /* no previous measure. */
789 srtt = m << 3; /* take the measured time to be rtt */
790 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */
791 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
792 tp->mdev_max_us = tp->rttvar_us;
793 tp->rtt_seq = tp->snd_nxt;
797 tp->srtt_us = max(1U, srtt);
800 static void tcp_update_pacing_rate(struct sock *sk)
802 const struct tcp_sock *tp = tcp_sk(sk);
805 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
806 rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3);
808 /* current rate is (cwnd * mss) / srtt
809 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
810 * In Congestion Avoidance phase, set it to 120 % the current rate.
812 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
813 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
814 * end of slow start and should slow down.
816 if (tp->snd_cwnd < tp->snd_ssthresh / 2)
817 rate *= sock_net(sk)->ipv4.sysctl_tcp_pacing_ss_ratio;
819 rate *= sock_net(sk)->ipv4.sysctl_tcp_pacing_ca_ratio;
821 rate *= max(tp->snd_cwnd, tp->packets_out);
823 if (likely(tp->srtt_us))
824 do_div(rate, tp->srtt_us);
826 /* WRITE_ONCE() is needed because sch_fq fetches sk_pacing_rate
827 * without any lock. We want to make sure compiler wont store
828 * intermediate values in this location.
830 WRITE_ONCE(sk->sk_pacing_rate, min_t(u64, rate,
831 sk->sk_max_pacing_rate));
834 /* Calculate rto without backoff. This is the second half of Van Jacobson's
835 * routine referred to above.
837 static void tcp_set_rto(struct sock *sk)
839 const struct tcp_sock *tp = tcp_sk(sk);
840 /* Old crap is replaced with new one. 8)
843 * 1. If rtt variance happened to be less 50msec, it is hallucination.
844 * It cannot be less due to utterly erratic ACK generation made
845 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
846 * to do with delayed acks, because at cwnd>2 true delack timeout
847 * is invisible. Actually, Linux-2.4 also generates erratic
848 * ACKs in some circumstances.
850 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
852 /* 2. Fixups made earlier cannot be right.
853 * If we do not estimate RTO correctly without them,
854 * all the algo is pure shit and should be replaced
855 * with correct one. It is exactly, which we pretend to do.
858 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
859 * guarantees that rto is higher.
864 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
866 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
869 cwnd = TCP_INIT_CWND;
870 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
873 /* Take a notice that peer is sending D-SACKs */
874 static void tcp_dsack_seen(struct tcp_sock *tp)
876 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
877 tp->rack.dsack_seen = 1;
881 /* It's reordering when higher sequence was delivered (i.e. sacked) before
882 * some lower never-retransmitted sequence ("low_seq"). The maximum reordering
883 * distance is approximated in full-mss packet distance ("reordering").
885 static void tcp_check_sack_reordering(struct sock *sk, const u32 low_seq,
888 struct tcp_sock *tp = tcp_sk(sk);
889 const u32 mss = tp->mss_cache;
892 fack = tcp_highest_sack_seq(tp);
893 if (!before(low_seq, fack))
896 metric = fack - low_seq;
897 if ((metric > tp->reordering * mss) && mss) {
898 #if FASTRETRANS_DEBUG > 1
899 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
900 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
904 tp->undo_marker ? tp->undo_retrans : 0);
906 tp->reordering = min_t(u32, (metric + mss - 1) / mss,
907 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_max_reordering));
910 /* This exciting event is worth to be remembered. 8) */
912 NET_INC_STATS(sock_net(sk),
913 ts ? LINUX_MIB_TCPTSREORDER : LINUX_MIB_TCPSACKREORDER);
916 /* This must be called before lost_out is incremented */
917 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
919 if ((!tp->retransmit_skb_hint && tp->retrans_out >= tp->lost_out) ||
920 (tp->retransmit_skb_hint &&
921 before(TCP_SKB_CB(skb)->seq,
922 TCP_SKB_CB(tp->retransmit_skb_hint)->seq)))
923 tp->retransmit_skb_hint = skb;
926 /* Sum the number of packets on the wire we have marked as lost.
927 * There are two cases we care about here:
928 * a) Packet hasn't been marked lost (nor retransmitted),
929 * and this is the first loss.
930 * b) Packet has been marked both lost and retransmitted,
931 * and this means we think it was lost again.
933 static void tcp_sum_lost(struct tcp_sock *tp, struct sk_buff *skb)
935 __u8 sacked = TCP_SKB_CB(skb)->sacked;
937 if (!(sacked & TCPCB_LOST) ||
938 ((sacked & TCPCB_LOST) && (sacked & TCPCB_SACKED_RETRANS)))
939 tp->lost += tcp_skb_pcount(skb);
942 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
944 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
945 tcp_verify_retransmit_hint(tp, skb);
947 tp->lost_out += tcp_skb_pcount(skb);
948 tcp_sum_lost(tp, skb);
949 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
953 void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp, struct sk_buff *skb)
955 tcp_verify_retransmit_hint(tp, skb);
957 tcp_sum_lost(tp, skb);
958 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
959 tp->lost_out += tcp_skb_pcount(skb);
960 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
964 /* This procedure tags the retransmission queue when SACKs arrive.
966 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
967 * Packets in queue with these bits set are counted in variables
968 * sacked_out, retrans_out and lost_out, correspondingly.
970 * Valid combinations are:
971 * Tag InFlight Description
972 * 0 1 - orig segment is in flight.
973 * S 0 - nothing flies, orig reached receiver.
974 * L 0 - nothing flies, orig lost by net.
975 * R 2 - both orig and retransmit are in flight.
976 * L|R 1 - orig is lost, retransmit is in flight.
977 * S|R 1 - orig reached receiver, retrans is still in flight.
978 * (L|S|R is logically valid, it could occur when L|R is sacked,
979 * but it is equivalent to plain S and code short-curcuits it to S.
980 * L|S is logically invalid, it would mean -1 packet in flight 8))
982 * These 6 states form finite state machine, controlled by the following events:
983 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
984 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
985 * 3. Loss detection event of two flavors:
986 * A. Scoreboard estimator decided the packet is lost.
987 * A'. Reno "three dupacks" marks head of queue lost.
988 * B. SACK arrives sacking SND.NXT at the moment, when the
989 * segment was retransmitted.
990 * 4. D-SACK added new rule: D-SACK changes any tag to S.
992 * It is pleasant to note, that state diagram turns out to be commutative,
993 * so that we are allowed not to be bothered by order of our actions,
994 * when multiple events arrive simultaneously. (see the function below).
996 * Reordering detection.
997 * --------------------
998 * Reordering metric is maximal distance, which a packet can be displaced
999 * in packet stream. With SACKs we can estimate it:
1001 * 1. SACK fills old hole and the corresponding segment was not
1002 * ever retransmitted -> reordering. Alas, we cannot use it
1003 * when segment was retransmitted.
1004 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1005 * for retransmitted and already SACKed segment -> reordering..
1006 * Both of these heuristics are not used in Loss state, when we cannot
1007 * account for retransmits accurately.
1009 * SACK block validation.
1010 * ----------------------
1012 * SACK block range validation checks that the received SACK block fits to
1013 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1014 * Note that SND.UNA is not included to the range though being valid because
1015 * it means that the receiver is rather inconsistent with itself reporting
1016 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1017 * perfectly valid, however, in light of RFC2018 which explicitly states
1018 * that "SACK block MUST reflect the newest segment. Even if the newest
1019 * segment is going to be discarded ...", not that it looks very clever
1020 * in case of head skb. Due to potentional receiver driven attacks, we
1021 * choose to avoid immediate execution of a walk in write queue due to
1022 * reneging and defer head skb's loss recovery to standard loss recovery
1023 * procedure that will eventually trigger (nothing forbids us doing this).
1025 * Implements also blockage to start_seq wrap-around. Problem lies in the
1026 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1027 * there's no guarantee that it will be before snd_nxt (n). The problem
1028 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1031 * <- outs wnd -> <- wrapzone ->
1032 * u e n u_w e_w s n_w
1034 * |<------------+------+----- TCP seqno space --------------+---------->|
1035 * ...-- <2^31 ->| |<--------...
1036 * ...---- >2^31 ------>| |<--------...
1038 * Current code wouldn't be vulnerable but it's better still to discard such
1039 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1040 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1041 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1042 * equal to the ideal case (infinite seqno space without wrap caused issues).
1044 * With D-SACK the lower bound is extended to cover sequence space below
1045 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1046 * again, D-SACK block must not to go across snd_una (for the same reason as
1047 * for the normal SACK blocks, explained above). But there all simplicity
1048 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1049 * fully below undo_marker they do not affect behavior in anyway and can
1050 * therefore be safely ignored. In rare cases (which are more or less
1051 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1052 * fragmentation and packet reordering past skb's retransmission. To consider
1053 * them correctly, the acceptable range must be extended even more though
1054 * the exact amount is rather hard to quantify. However, tp->max_window can
1055 * be used as an exaggerated estimate.
1057 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
1058 u32 start_seq, u32 end_seq)
1060 /* Too far in future, or reversed (interpretation is ambiguous) */
1061 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1064 /* Nasty start_seq wrap-around check (see comments above) */
1065 if (!before(start_seq, tp->snd_nxt))
1068 /* In outstanding window? ...This is valid exit for D-SACKs too.
1069 * start_seq == snd_una is non-sensical (see comments above)
1071 if (after(start_seq, tp->snd_una))
1074 if (!is_dsack || !tp->undo_marker)
1077 /* ...Then it's D-SACK, and must reside below snd_una completely */
1078 if (after(end_seq, tp->snd_una))
1081 if (!before(start_seq, tp->undo_marker))
1085 if (!after(end_seq, tp->undo_marker))
1088 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1089 * start_seq < undo_marker and end_seq >= undo_marker.
1091 return !before(start_seq, end_seq - tp->max_window);
1094 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1095 struct tcp_sack_block_wire *sp, int num_sacks,
1098 struct tcp_sock *tp = tcp_sk(sk);
1099 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1100 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1101 bool dup_sack = false;
1103 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1106 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1107 } else if (num_sacks > 1) {
1108 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1109 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1111 if (!after(end_seq_0, end_seq_1) &&
1112 !before(start_seq_0, start_seq_1)) {
1115 NET_INC_STATS(sock_net(sk),
1116 LINUX_MIB_TCPDSACKOFORECV);
1120 /* D-SACK for already forgotten data... Do dumb counting. */
1121 if (dup_sack && tp->undo_marker && tp->undo_retrans > 0 &&
1122 !after(end_seq_0, prior_snd_una) &&
1123 after(end_seq_0, tp->undo_marker))
1129 struct tcp_sacktag_state {
1131 /* Timestamps for earliest and latest never-retransmitted segment
1132 * that was SACKed. RTO needs the earliest RTT to stay conservative,
1133 * but congestion control should still get an accurate delay signal.
1137 struct rate_sample *rate;
1139 unsigned int mss_now;
1142 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1143 * the incoming SACK may not exactly match but we can find smaller MSS
1144 * aligned portion of it that matches. Therefore we might need to fragment
1145 * which may fail and creates some hassle (caller must handle error case
1148 * FIXME: this could be merged to shift decision code
1150 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1151 u32 start_seq, u32 end_seq)
1155 unsigned int pkt_len;
1158 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1159 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1161 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1162 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1163 mss = tcp_skb_mss(skb);
1164 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1167 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1171 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1176 /* Round if necessary so that SACKs cover only full MSSes
1177 * and/or the remaining small portion (if present)
1179 if (pkt_len > mss) {
1180 unsigned int new_len = (pkt_len / mss) * mss;
1181 if (!in_sack && new_len < pkt_len)
1186 if (pkt_len >= skb->len && !in_sack)
1189 err = tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
1190 pkt_len, mss, GFP_ATOMIC);
1198 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1199 static u8 tcp_sacktag_one(struct sock *sk,
1200 struct tcp_sacktag_state *state, u8 sacked,
1201 u32 start_seq, u32 end_seq,
1202 int dup_sack, int pcount,
1205 struct tcp_sock *tp = tcp_sk(sk);
1207 /* Account D-SACK for retransmitted packet. */
1208 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1209 if (tp->undo_marker && tp->undo_retrans > 0 &&
1210 after(end_seq, tp->undo_marker))
1211 tp->undo_retrans = max_t(int, 0, tp->undo_retrans - pcount);
1212 if ((sacked & TCPCB_SACKED_ACKED) &&
1213 before(start_seq, state->reord))
1214 state->reord = start_seq;
1217 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1218 if (!after(end_seq, tp->snd_una))
1221 if (!(sacked & TCPCB_SACKED_ACKED)) {
1222 tcp_rack_advance(tp, sacked, end_seq, xmit_time);
1224 if (sacked & TCPCB_SACKED_RETRANS) {
1225 /* If the segment is not tagged as lost,
1226 * we do not clear RETRANS, believing
1227 * that retransmission is still in flight.
1229 if (sacked & TCPCB_LOST) {
1230 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1231 tp->lost_out -= pcount;
1232 tp->retrans_out -= pcount;
1235 if (!(sacked & TCPCB_RETRANS)) {
1236 /* New sack for not retransmitted frame,
1237 * which was in hole. It is reordering.
1239 if (before(start_seq,
1240 tcp_highest_sack_seq(tp)) &&
1241 before(start_seq, state->reord))
1242 state->reord = start_seq;
1244 if (!after(end_seq, tp->high_seq))
1245 state->flag |= FLAG_ORIG_SACK_ACKED;
1246 if (state->first_sackt == 0)
1247 state->first_sackt = xmit_time;
1248 state->last_sackt = xmit_time;
1251 if (sacked & TCPCB_LOST) {
1252 sacked &= ~TCPCB_LOST;
1253 tp->lost_out -= pcount;
1257 sacked |= TCPCB_SACKED_ACKED;
1258 state->flag |= FLAG_DATA_SACKED;
1259 tp->sacked_out += pcount;
1260 tp->delivered += pcount; /* Out-of-order packets delivered */
1262 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1263 if (tp->lost_skb_hint &&
1264 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1265 tp->lost_cnt_hint += pcount;
1268 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1269 * frames and clear it. undo_retrans is decreased above, L|R frames
1270 * are accounted above as well.
1272 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1273 sacked &= ~TCPCB_SACKED_RETRANS;
1274 tp->retrans_out -= pcount;
1280 /* Shift newly-SACKed bytes from this skb to the immediately previous
1281 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1283 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *prev,
1284 struct sk_buff *skb,
1285 struct tcp_sacktag_state *state,
1286 unsigned int pcount, int shifted, int mss,
1289 struct tcp_sock *tp = tcp_sk(sk);
1290 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1291 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
1295 /* Adjust counters and hints for the newly sacked sequence
1296 * range but discard the return value since prev is already
1297 * marked. We must tag the range first because the seq
1298 * advancement below implicitly advances
1299 * tcp_highest_sack_seq() when skb is highest_sack.
1301 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1302 start_seq, end_seq, dup_sack, pcount,
1303 tcp_skb_timestamp_us(skb));
1304 tcp_rate_skb_delivered(sk, skb, state->rate);
1306 if (skb == tp->lost_skb_hint)
1307 tp->lost_cnt_hint += pcount;
1309 TCP_SKB_CB(prev)->end_seq += shifted;
1310 TCP_SKB_CB(skb)->seq += shifted;
1312 tcp_skb_pcount_add(prev, pcount);
1313 WARN_ON_ONCE(tcp_skb_pcount(skb) < pcount);
1314 tcp_skb_pcount_add(skb, -pcount);
1316 /* When we're adding to gso_segs == 1, gso_size will be zero,
1317 * in theory this shouldn't be necessary but as long as DSACK
1318 * code can come after this skb later on it's better to keep
1319 * setting gso_size to something.
1321 if (!TCP_SKB_CB(prev)->tcp_gso_size)
1322 TCP_SKB_CB(prev)->tcp_gso_size = mss;
1324 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1325 if (tcp_skb_pcount(skb) <= 1)
1326 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1328 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1329 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1332 BUG_ON(!tcp_skb_pcount(skb));
1333 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1337 /* Whole SKB was eaten :-) */
1339 if (skb == tp->retransmit_skb_hint)
1340 tp->retransmit_skb_hint = prev;
1341 if (skb == tp->lost_skb_hint) {
1342 tp->lost_skb_hint = prev;
1343 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1346 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1347 TCP_SKB_CB(prev)->eor = TCP_SKB_CB(skb)->eor;
1348 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1349 TCP_SKB_CB(prev)->end_seq++;
1351 if (skb == tcp_highest_sack(sk))
1352 tcp_advance_highest_sack(sk, skb);
1354 tcp_skb_collapse_tstamp(prev, skb);
1355 if (unlikely(TCP_SKB_CB(prev)->tx.delivered_mstamp))
1356 TCP_SKB_CB(prev)->tx.delivered_mstamp = 0;
1358 tcp_rtx_queue_unlink_and_free(skb, sk);
1360 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKMERGED);
1365 /* I wish gso_size would have a bit more sane initialization than
1366 * something-or-zero which complicates things
1368 static int tcp_skb_seglen(const struct sk_buff *skb)
1370 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1373 /* Shifting pages past head area doesn't work */
1374 static int skb_can_shift(const struct sk_buff *skb)
1376 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1379 int tcp_skb_shift(struct sk_buff *to, struct sk_buff *from,
1380 int pcount, int shiftlen)
1382 /* TCP min gso_size is 8 bytes (TCP_MIN_GSO_SIZE)
1383 * Since TCP_SKB_CB(skb)->tcp_gso_segs is 16 bits, we need
1384 * to make sure not storing more than 65535 * 8 bytes per skb,
1385 * even if current MSS is bigger.
1387 if (unlikely(to->len + shiftlen >= 65535 * TCP_MIN_GSO_SIZE))
1389 if (unlikely(tcp_skb_pcount(to) + pcount > 65535))
1391 return skb_shift(to, from, shiftlen);
1394 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1397 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1398 struct tcp_sacktag_state *state,
1399 u32 start_seq, u32 end_seq,
1402 struct tcp_sock *tp = tcp_sk(sk);
1403 struct sk_buff *prev;
1409 /* Normally R but no L won't result in plain S */
1411 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1413 if (!skb_can_shift(skb))
1415 /* This frame is about to be dropped (was ACKed). */
1416 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1419 /* Can only happen with delayed DSACK + discard craziness */
1420 prev = skb_rb_prev(skb);
1424 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1427 if (!tcp_skb_can_collapse_to(prev))
1430 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1431 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1435 pcount = tcp_skb_pcount(skb);
1436 mss = tcp_skb_seglen(skb);
1438 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1439 * drop this restriction as unnecessary
1441 if (mss != tcp_skb_seglen(prev))
1444 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1446 /* CHECKME: This is non-MSS split case only?, this will
1447 * cause skipped skbs due to advancing loop btw, original
1448 * has that feature too
1450 if (tcp_skb_pcount(skb) <= 1)
1453 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1455 /* TODO: head merge to next could be attempted here
1456 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1457 * though it might not be worth of the additional hassle
1459 * ...we can probably just fallback to what was done
1460 * previously. We could try merging non-SACKed ones
1461 * as well but it probably isn't going to buy off
1462 * because later SACKs might again split them, and
1463 * it would make skb timestamp tracking considerably
1469 len = end_seq - TCP_SKB_CB(skb)->seq;
1471 BUG_ON(len > skb->len);
1473 /* MSS boundaries should be honoured or else pcount will
1474 * severely break even though it makes things bit trickier.
1475 * Optimize common case to avoid most of the divides
1477 mss = tcp_skb_mss(skb);
1479 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1480 * drop this restriction as unnecessary
1482 if (mss != tcp_skb_seglen(prev))
1487 } else if (len < mss) {
1495 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1496 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1499 if (!tcp_skb_shift(prev, skb, pcount, len))
1501 if (!tcp_shifted_skb(sk, prev, skb, state, pcount, len, mss, dup_sack))
1504 /* Hole filled allows collapsing with the next as well, this is very
1505 * useful when hole on every nth skb pattern happens
1507 skb = skb_rb_next(prev);
1511 if (!skb_can_shift(skb) ||
1512 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1513 (mss != tcp_skb_seglen(skb)))
1517 pcount = tcp_skb_pcount(skb);
1518 if (tcp_skb_shift(prev, skb, pcount, len))
1519 tcp_shifted_skb(sk, prev, skb, state, pcount,
1529 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1533 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1534 struct tcp_sack_block *next_dup,
1535 struct tcp_sacktag_state *state,
1536 u32 start_seq, u32 end_seq,
1539 struct tcp_sock *tp = tcp_sk(sk);
1540 struct sk_buff *tmp;
1542 skb_rbtree_walk_from(skb) {
1544 bool dup_sack = dup_sack_in;
1546 /* queue is in-order => we can short-circuit the walk early */
1547 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1551 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1552 in_sack = tcp_match_skb_to_sack(sk, skb,
1553 next_dup->start_seq,
1559 /* skb reference here is a bit tricky to get right, since
1560 * shifting can eat and free both this skb and the next,
1561 * so not even _safe variant of the loop is enough.
1564 tmp = tcp_shift_skb_data(sk, skb, state,
1565 start_seq, end_seq, dup_sack);
1574 in_sack = tcp_match_skb_to_sack(sk, skb,
1580 if (unlikely(in_sack < 0))
1584 TCP_SKB_CB(skb)->sacked =
1587 TCP_SKB_CB(skb)->sacked,
1588 TCP_SKB_CB(skb)->seq,
1589 TCP_SKB_CB(skb)->end_seq,
1591 tcp_skb_pcount(skb),
1592 tcp_skb_timestamp_us(skb));
1593 tcp_rate_skb_delivered(sk, skb, state->rate);
1594 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1595 list_del_init(&skb->tcp_tsorted_anchor);
1597 if (!before(TCP_SKB_CB(skb)->seq,
1598 tcp_highest_sack_seq(tp)))
1599 tcp_advance_highest_sack(sk, skb);
1605 static struct sk_buff *tcp_sacktag_bsearch(struct sock *sk, u32 seq)
1607 struct rb_node *parent, **p = &sk->tcp_rtx_queue.rb_node;
1608 struct sk_buff *skb;
1612 skb = rb_to_skb(parent);
1613 if (before(seq, TCP_SKB_CB(skb)->seq)) {
1614 p = &parent->rb_left;
1617 if (!before(seq, TCP_SKB_CB(skb)->end_seq)) {
1618 p = &parent->rb_right;
1626 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1629 if (skb && after(TCP_SKB_CB(skb)->seq, skip_to_seq))
1632 return tcp_sacktag_bsearch(sk, skip_to_seq);
1635 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1637 struct tcp_sack_block *next_dup,
1638 struct tcp_sacktag_state *state,
1644 if (before(next_dup->start_seq, skip_to_seq)) {
1645 skb = tcp_sacktag_skip(skb, sk, next_dup->start_seq);
1646 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1647 next_dup->start_seq, next_dup->end_seq,
1654 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1656 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1660 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1661 u32 prior_snd_una, struct tcp_sacktag_state *state)
1663 struct tcp_sock *tp = tcp_sk(sk);
1664 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1665 TCP_SKB_CB(ack_skb)->sacked);
1666 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1667 struct tcp_sack_block sp[TCP_NUM_SACKS];
1668 struct tcp_sack_block *cache;
1669 struct sk_buff *skb;
1670 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1672 bool found_dup_sack = false;
1674 int first_sack_index;
1677 state->reord = tp->snd_nxt;
1679 if (!tp->sacked_out)
1680 tcp_highest_sack_reset(sk);
1682 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1683 num_sacks, prior_snd_una);
1684 if (found_dup_sack) {
1685 state->flag |= FLAG_DSACKING_ACK;
1686 tp->delivered++; /* A spurious retransmission is delivered */
1689 /* Eliminate too old ACKs, but take into
1690 * account more or less fresh ones, they can
1691 * contain valid SACK info.
1693 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1696 if (!tp->packets_out)
1700 first_sack_index = 0;
1701 for (i = 0; i < num_sacks; i++) {
1702 bool dup_sack = !i && found_dup_sack;
1704 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1705 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1707 if (!tcp_is_sackblock_valid(tp, dup_sack,
1708 sp[used_sacks].start_seq,
1709 sp[used_sacks].end_seq)) {
1713 if (!tp->undo_marker)
1714 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1716 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1718 /* Don't count olds caused by ACK reordering */
1719 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1720 !after(sp[used_sacks].end_seq, tp->snd_una))
1722 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1725 NET_INC_STATS(sock_net(sk), mib_idx);
1727 first_sack_index = -1;
1731 /* Ignore very old stuff early */
1732 if (!after(sp[used_sacks].end_seq, prior_snd_una)) {
1734 first_sack_index = -1;
1741 /* order SACK blocks to allow in order walk of the retrans queue */
1742 for (i = used_sacks - 1; i > 0; i--) {
1743 for (j = 0; j < i; j++) {
1744 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1745 swap(sp[j], sp[j + 1]);
1747 /* Track where the first SACK block goes to */
1748 if (j == first_sack_index)
1749 first_sack_index = j + 1;
1754 state->mss_now = tcp_current_mss(sk);
1758 if (!tp->sacked_out) {
1759 /* It's already past, so skip checking against it */
1760 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1762 cache = tp->recv_sack_cache;
1763 /* Skip empty blocks in at head of the cache */
1764 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1769 while (i < used_sacks) {
1770 u32 start_seq = sp[i].start_seq;
1771 u32 end_seq = sp[i].end_seq;
1772 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1773 struct tcp_sack_block *next_dup = NULL;
1775 if (found_dup_sack && ((i + 1) == first_sack_index))
1776 next_dup = &sp[i + 1];
1778 /* Skip too early cached blocks */
1779 while (tcp_sack_cache_ok(tp, cache) &&
1780 !before(start_seq, cache->end_seq))
1783 /* Can skip some work by looking recv_sack_cache? */
1784 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1785 after(end_seq, cache->start_seq)) {
1788 if (before(start_seq, cache->start_seq)) {
1789 skb = tcp_sacktag_skip(skb, sk, start_seq);
1790 skb = tcp_sacktag_walk(skb, sk, next_dup,
1797 /* Rest of the block already fully processed? */
1798 if (!after(end_seq, cache->end_seq))
1801 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1805 /* ...tail remains todo... */
1806 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1807 /* ...but better entrypoint exists! */
1808 skb = tcp_highest_sack(sk);
1815 skb = tcp_sacktag_skip(skb, sk, cache->end_seq);
1816 /* Check overlap against next cached too (past this one already) */
1821 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1822 skb = tcp_highest_sack(sk);
1826 skb = tcp_sacktag_skip(skb, sk, start_seq);
1829 skb = tcp_sacktag_walk(skb, sk, next_dup, state,
1830 start_seq, end_seq, dup_sack);
1836 /* Clear the head of the cache sack blocks so we can skip it next time */
1837 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1838 tp->recv_sack_cache[i].start_seq = 0;
1839 tp->recv_sack_cache[i].end_seq = 0;
1841 for (j = 0; j < used_sacks; j++)
1842 tp->recv_sack_cache[i++] = sp[j];
1844 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Loss || tp->undo_marker)
1845 tcp_check_sack_reordering(sk, state->reord, 0);
1847 tcp_verify_left_out(tp);
1850 #if FASTRETRANS_DEBUG > 0
1851 WARN_ON((int)tp->sacked_out < 0);
1852 WARN_ON((int)tp->lost_out < 0);
1853 WARN_ON((int)tp->retrans_out < 0);
1854 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1859 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1860 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1862 static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
1866 holes = max(tp->lost_out, 1U);
1867 holes = min(holes, tp->packets_out);
1869 if ((tp->sacked_out + holes) > tp->packets_out) {
1870 tp->sacked_out = tp->packets_out - holes;
1876 /* If we receive more dupacks than we expected counting segments
1877 * in assumption of absent reordering, interpret this as reordering.
1878 * The only another reason could be bug in receiver TCP.
1880 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1882 struct tcp_sock *tp = tcp_sk(sk);
1884 if (!tcp_limit_reno_sacked(tp))
1887 tp->reordering = min_t(u32, tp->packets_out + addend,
1888 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_max_reordering));
1890 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRENOREORDER);
1893 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1895 static void tcp_add_reno_sack(struct sock *sk, int num_dupack)
1898 struct tcp_sock *tp = tcp_sk(sk);
1899 u32 prior_sacked = tp->sacked_out;
1902 tp->sacked_out += num_dupack;
1903 tcp_check_reno_reordering(sk, 0);
1904 delivered = tp->sacked_out - prior_sacked;
1906 tp->delivered += delivered;
1907 tcp_verify_left_out(tp);
1911 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1913 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1915 struct tcp_sock *tp = tcp_sk(sk);
1918 /* One ACK acked hole. The rest eat duplicate ACKs. */
1919 tp->delivered += max_t(int, acked - tp->sacked_out, 1);
1920 if (acked - 1 >= tp->sacked_out)
1923 tp->sacked_out -= acked - 1;
1925 tcp_check_reno_reordering(sk, acked);
1926 tcp_verify_left_out(tp);
1929 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1934 void tcp_clear_retrans(struct tcp_sock *tp)
1936 tp->retrans_out = 0;
1938 tp->undo_marker = 0;
1939 tp->undo_retrans = -1;
1943 static inline void tcp_init_undo(struct tcp_sock *tp)
1945 tp->undo_marker = tp->snd_una;
1946 /* Retransmission still in flight may cause DSACKs later. */
1947 tp->undo_retrans = tp->retrans_out ? : -1;
1950 static bool tcp_is_rack(const struct sock *sk)
1952 return READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_recovery) &
1953 TCP_RACK_LOSS_DETECTION;
1956 /* If we detect SACK reneging, forget all SACK information
1957 * and reset tags completely, otherwise preserve SACKs. If receiver
1958 * dropped its ofo queue, we will know this due to reneging detection.
1960 static void tcp_timeout_mark_lost(struct sock *sk)
1962 struct tcp_sock *tp = tcp_sk(sk);
1963 struct sk_buff *skb, *head;
1964 bool is_reneg; /* is receiver reneging on SACKs? */
1966 head = tcp_rtx_queue_head(sk);
1967 is_reneg = head && (TCP_SKB_CB(head)->sacked & TCPCB_SACKED_ACKED);
1969 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
1971 /* Mark SACK reneging until we recover from this loss event. */
1972 tp->is_sack_reneg = 1;
1973 } else if (tcp_is_reno(tp)) {
1974 tcp_reset_reno_sack(tp);
1978 skb_rbtree_walk_from(skb) {
1980 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1981 else if (tcp_is_rack(sk) && skb != head &&
1982 tcp_rack_skb_timeout(tp, skb, 0) > 0)
1983 continue; /* Don't mark recently sent ones lost yet */
1984 tcp_mark_skb_lost(sk, skb);
1986 tcp_verify_left_out(tp);
1987 tcp_clear_all_retrans_hints(tp);
1990 /* Enter Loss state. */
1991 void tcp_enter_loss(struct sock *sk)
1993 const struct inet_connection_sock *icsk = inet_csk(sk);
1994 struct tcp_sock *tp = tcp_sk(sk);
1995 struct net *net = sock_net(sk);
1996 bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery;
1999 tcp_timeout_mark_lost(sk);
2001 /* Reduce ssthresh if it has not yet been made inside this window. */
2002 if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
2003 !after(tp->high_seq, tp->snd_una) ||
2004 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
2005 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2006 tp->prior_cwnd = tp->snd_cwnd;
2007 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2008 tcp_ca_event(sk, CA_EVENT_LOSS);
2011 tp->snd_cwnd = tcp_packets_in_flight(tp) + 1;
2012 tp->snd_cwnd_cnt = 0;
2013 tp->snd_cwnd_stamp = tcp_jiffies32;
2015 /* Timeout in disordered state after receiving substantial DUPACKs
2016 * suggests that the degree of reordering is over-estimated.
2018 reordering = READ_ONCE(net->ipv4.sysctl_tcp_reordering);
2019 if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
2020 tp->sacked_out >= reordering)
2021 tp->reordering = min_t(unsigned int, tp->reordering,
2024 tcp_set_ca_state(sk, TCP_CA_Loss);
2025 tp->high_seq = tp->snd_nxt;
2026 tcp_ecn_queue_cwr(tp);
2028 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
2029 * loss recovery is underway except recurring timeout(s) on
2030 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
2032 tp->frto = READ_ONCE(net->ipv4.sysctl_tcp_frto) &&
2033 (new_recovery || icsk->icsk_retransmits) &&
2034 !inet_csk(sk)->icsk_mtup.probe_size;
2037 /* If ACK arrived pointing to a remembered SACK, it means that our
2038 * remembered SACKs do not reflect real state of receiver i.e.
2039 * receiver _host_ is heavily congested (or buggy).
2041 * To avoid big spurious retransmission bursts due to transient SACK
2042 * scoreboard oddities that look like reneging, we give the receiver a
2043 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
2044 * restore sanity to the SACK scoreboard. If the apparent reneging
2045 * persists until this RTO then we'll clear the SACK scoreboard.
2047 static bool tcp_check_sack_reneging(struct sock *sk, int flag)
2049 if (flag & FLAG_SACK_RENEGING &&
2050 flag & FLAG_SND_UNA_ADVANCED) {
2051 struct tcp_sock *tp = tcp_sk(sk);
2052 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4),
2053 msecs_to_jiffies(10));
2055 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2056 delay, TCP_RTO_MAX);
2062 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2063 * counter when SACK is enabled (without SACK, sacked_out is used for
2066 * With reordering, holes may still be in flight, so RFC3517 recovery
2067 * uses pure sacked_out (total number of SACKed segments) even though
2068 * it violates the RFC that uses duplicate ACKs, often these are equal
2069 * but when e.g. out-of-window ACKs or packet duplication occurs,
2070 * they differ. Since neither occurs due to loss, TCP should really
2073 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2075 return tp->sacked_out + 1;
2078 /* Linux NewReno/SACK/ECN state machine.
2079 * --------------------------------------
2081 * "Open" Normal state, no dubious events, fast path.
2082 * "Disorder" In all the respects it is "Open",
2083 * but requires a bit more attention. It is entered when
2084 * we see some SACKs or dupacks. It is split of "Open"
2085 * mainly to move some processing from fast path to slow one.
2086 * "CWR" CWND was reduced due to some Congestion Notification event.
2087 * It can be ECN, ICMP source quench, local device congestion.
2088 * "Recovery" CWND was reduced, we are fast-retransmitting.
2089 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2091 * tcp_fastretrans_alert() is entered:
2092 * - each incoming ACK, if state is not "Open"
2093 * - when arrived ACK is unusual, namely:
2098 * Counting packets in flight is pretty simple.
2100 * in_flight = packets_out - left_out + retrans_out
2102 * packets_out is SND.NXT-SND.UNA counted in packets.
2104 * retrans_out is number of retransmitted segments.
2106 * left_out is number of segments left network, but not ACKed yet.
2108 * left_out = sacked_out + lost_out
2110 * sacked_out: Packets, which arrived to receiver out of order
2111 * and hence not ACKed. With SACKs this number is simply
2112 * amount of SACKed data. Even without SACKs
2113 * it is easy to give pretty reliable estimate of this number,
2114 * counting duplicate ACKs.
2116 * lost_out: Packets lost by network. TCP has no explicit
2117 * "loss notification" feedback from network (for now).
2118 * It means that this number can be only _guessed_.
2119 * Actually, it is the heuristics to predict lossage that
2120 * distinguishes different algorithms.
2122 * F.e. after RTO, when all the queue is considered as lost,
2123 * lost_out = packets_out and in_flight = retrans_out.
2125 * Essentially, we have now a few algorithms detecting
2128 * If the receiver supports SACK:
2130 * RFC6675/3517: It is the conventional algorithm. A packet is
2131 * considered lost if the number of higher sequence packets
2132 * SACKed is greater than or equal the DUPACK thoreshold
2133 * (reordering). This is implemented in tcp_mark_head_lost and
2134 * tcp_update_scoreboard.
2136 * RACK (draft-ietf-tcpm-rack-01): it is a newer algorithm
2137 * (2017-) that checks timing instead of counting DUPACKs.
2138 * Essentially a packet is considered lost if it's not S/ACKed
2139 * after RTT + reordering_window, where both metrics are
2140 * dynamically measured and adjusted. This is implemented in
2141 * tcp_rack_mark_lost.
2143 * If the receiver does not support SACK:
2145 * NewReno (RFC6582): in Recovery we assume that one segment
2146 * is lost (classic Reno). While we are in Recovery and
2147 * a partial ACK arrives, we assume that one more packet
2148 * is lost (NewReno). This heuristics are the same in NewReno
2151 * Really tricky (and requiring careful tuning) part of algorithm
2152 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2153 * The first determines the moment _when_ we should reduce CWND and,
2154 * hence, slow down forward transmission. In fact, it determines the moment
2155 * when we decide that hole is caused by loss, rather than by a reorder.
2157 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2158 * holes, caused by lost packets.
2160 * And the most logically complicated part of algorithm is undo
2161 * heuristics. We detect false retransmits due to both too early
2162 * fast retransmit (reordering) and underestimated RTO, analyzing
2163 * timestamps and D-SACKs. When we detect that some segments were
2164 * retransmitted by mistake and CWND reduction was wrong, we undo
2165 * window reduction and abort recovery phase. This logic is hidden
2166 * inside several functions named tcp_try_undo_<something>.
2169 /* This function decides, when we should leave Disordered state
2170 * and enter Recovery phase, reducing congestion window.
2172 * Main question: may we further continue forward transmission
2173 * with the same cwnd?
2175 static bool tcp_time_to_recover(struct sock *sk, int flag)
2177 struct tcp_sock *tp = tcp_sk(sk);
2179 /* Trick#1: The loss is proven. */
2183 /* Not-A-Trick#2 : Classic rule... */
2184 if (!tcp_is_rack(sk) && tcp_dupack_heuristics(tp) > tp->reordering)
2190 /* Detect loss in event "A" above by marking head of queue up as lost.
2191 * For non-SACK(Reno) senders, the first "packets" number of segments
2192 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2193 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2194 * the maximum SACKed segments to pass before reaching this limit.
2196 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2198 struct tcp_sock *tp = tcp_sk(sk);
2199 struct sk_buff *skb;
2200 int cnt, oldcnt, lost;
2202 /* Use SACK to deduce losses of new sequences sent during recovery */
2203 const u32 loss_high = tcp_is_sack(tp) ? tp->snd_nxt : tp->high_seq;
2205 WARN_ON(packets > tp->packets_out);
2206 skb = tp->lost_skb_hint;
2208 /* Head already handled? */
2209 if (mark_head && after(TCP_SKB_CB(skb)->seq, tp->snd_una))
2211 cnt = tp->lost_cnt_hint;
2213 skb = tcp_rtx_queue_head(sk);
2217 skb_rbtree_walk_from(skb) {
2218 /* TODO: do this better */
2219 /* this is not the most efficient way to do this... */
2220 tp->lost_skb_hint = skb;
2221 tp->lost_cnt_hint = cnt;
2223 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2227 if (tcp_is_reno(tp) ||
2228 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2229 cnt += tcp_skb_pcount(skb);
2231 if (cnt > packets) {
2232 if (tcp_is_sack(tp) ||
2233 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
2234 (oldcnt >= packets))
2237 mss = tcp_skb_mss(skb);
2238 /* If needed, chop off the prefix to mark as lost. */
2239 lost = (packets - oldcnt) * mss;
2240 if (lost < skb->len &&
2241 tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
2242 lost, mss, GFP_ATOMIC) < 0)
2247 tcp_skb_mark_lost(tp, skb);
2252 tcp_verify_left_out(tp);
2255 /* Account newly detected lost packet(s) */
2257 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2259 struct tcp_sock *tp = tcp_sk(sk);
2261 if (tcp_is_sack(tp)) {
2262 int sacked_upto = tp->sacked_out - tp->reordering;
2263 if (sacked_upto >= 0)
2264 tcp_mark_head_lost(sk, sacked_upto, 0);
2265 else if (fast_rexmit)
2266 tcp_mark_head_lost(sk, 1, 1);
2270 static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when)
2272 return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2273 before(tp->rx_opt.rcv_tsecr, when);
2276 /* skb is spurious retransmitted if the returned timestamp echo
2277 * reply is prior to the skb transmission time
2279 static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp,
2280 const struct sk_buff *skb)
2282 return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) &&
2283 tcp_tsopt_ecr_before(tp, tcp_skb_timestamp(skb));
2286 /* Nothing was retransmitted or returned timestamp is less
2287 * than timestamp of the first retransmission.
2289 static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2291 return tp->retrans_stamp &&
2292 tcp_tsopt_ecr_before(tp, tp->retrans_stamp);
2295 /* Undo procedures. */
2297 /* We can clear retrans_stamp when there are no retransmissions in the
2298 * window. It would seem that it is trivially available for us in
2299 * tp->retrans_out, however, that kind of assumptions doesn't consider
2300 * what will happen if errors occur when sending retransmission for the
2301 * second time. ...It could the that such segment has only
2302 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2303 * the head skb is enough except for some reneging corner cases that
2304 * are not worth the effort.
2306 * Main reason for all this complexity is the fact that connection dying
2307 * time now depends on the validity of the retrans_stamp, in particular,
2308 * that successive retransmissions of a segment must not advance
2309 * retrans_stamp under any conditions.
2311 static bool tcp_any_retrans_done(const struct sock *sk)
2313 const struct tcp_sock *tp = tcp_sk(sk);
2314 struct sk_buff *skb;
2316 if (tp->retrans_out)
2319 skb = tcp_rtx_queue_head(sk);
2320 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2326 static void DBGUNDO(struct sock *sk, const char *msg)
2328 #if FASTRETRANS_DEBUG > 1
2329 struct tcp_sock *tp = tcp_sk(sk);
2330 struct inet_sock *inet = inet_sk(sk);
2332 if (sk->sk_family == AF_INET) {
2333 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2335 &inet->inet_daddr, ntohs(inet->inet_dport),
2336 tp->snd_cwnd, tcp_left_out(tp),
2337 tp->snd_ssthresh, tp->prior_ssthresh,
2340 #if IS_ENABLED(CONFIG_IPV6)
2341 else if (sk->sk_family == AF_INET6) {
2342 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2344 &sk->sk_v6_daddr, ntohs(inet->inet_dport),
2345 tp->snd_cwnd, tcp_left_out(tp),
2346 tp->snd_ssthresh, tp->prior_ssthresh,
2353 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
2355 struct tcp_sock *tp = tcp_sk(sk);
2358 struct sk_buff *skb;
2360 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2361 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2364 tcp_clear_all_retrans_hints(tp);
2367 if (tp->prior_ssthresh) {
2368 const struct inet_connection_sock *icsk = inet_csk(sk);
2370 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2372 if (tp->prior_ssthresh > tp->snd_ssthresh) {
2373 tp->snd_ssthresh = tp->prior_ssthresh;
2374 tcp_ecn_withdraw_cwr(tp);
2377 tp->snd_cwnd_stamp = tcp_jiffies32;
2378 tp->undo_marker = 0;
2379 tp->rack.advanced = 1; /* Force RACK to re-exam losses */
2382 static inline bool tcp_may_undo(const struct tcp_sock *tp)
2384 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2387 static bool tcp_is_non_sack_preventing_reopen(struct sock *sk)
2389 struct tcp_sock *tp = tcp_sk(sk);
2391 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2392 /* Hold old state until something *above* high_seq
2393 * is ACKed. For Reno it is MUST to prevent false
2394 * fast retransmits (RFC2582). SACK TCP is safe. */
2395 if (!tcp_any_retrans_done(sk))
2396 tp->retrans_stamp = 0;
2402 /* People celebrate: "We love our President!" */
2403 static bool tcp_try_undo_recovery(struct sock *sk)
2405 struct tcp_sock *tp = tcp_sk(sk);
2407 if (tcp_may_undo(tp)) {
2410 /* Happy end! We did not retransmit anything
2411 * or our original transmission succeeded.
2413 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2414 tcp_undo_cwnd_reduction(sk, false);
2415 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2416 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2418 mib_idx = LINUX_MIB_TCPFULLUNDO;
2420 NET_INC_STATS(sock_net(sk), mib_idx);
2421 } else if (tp->rack.reo_wnd_persist) {
2422 tp->rack.reo_wnd_persist--;
2424 if (tcp_is_non_sack_preventing_reopen(sk))
2426 tcp_set_ca_state(sk, TCP_CA_Open);
2427 tp->is_sack_reneg = 0;
2431 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2432 static bool tcp_try_undo_dsack(struct sock *sk)
2434 struct tcp_sock *tp = tcp_sk(sk);
2436 if (tp->undo_marker && !tp->undo_retrans) {
2437 tp->rack.reo_wnd_persist = min(TCP_RACK_RECOVERY_THRESH,
2438 tp->rack.reo_wnd_persist + 1);
2439 DBGUNDO(sk, "D-SACK");
2440 tcp_undo_cwnd_reduction(sk, false);
2441 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2447 /* Undo during loss recovery after partial ACK or using F-RTO. */
2448 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2450 struct tcp_sock *tp = tcp_sk(sk);
2452 if (frto_undo || tcp_may_undo(tp)) {
2453 tcp_undo_cwnd_reduction(sk, true);
2455 DBGUNDO(sk, "partial loss");
2456 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2458 NET_INC_STATS(sock_net(sk),
2459 LINUX_MIB_TCPSPURIOUSRTOS);
2460 inet_csk(sk)->icsk_retransmits = 0;
2461 if (tcp_is_non_sack_preventing_reopen(sk))
2463 if (frto_undo || tcp_is_sack(tp)) {
2464 tcp_set_ca_state(sk, TCP_CA_Open);
2465 tp->is_sack_reneg = 0;
2472 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
2473 * It computes the number of packets to send (sndcnt) based on packets newly
2475 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2476 * cwnd reductions across a full RTT.
2477 * 2) Otherwise PRR uses packet conservation to send as much as delivered.
2478 * But when the retransmits are acked without further losses, PRR
2479 * slow starts cwnd up to ssthresh to speed up the recovery.
2481 static void tcp_init_cwnd_reduction(struct sock *sk)
2483 struct tcp_sock *tp = tcp_sk(sk);
2485 tp->high_seq = tp->snd_nxt;
2486 tp->tlp_high_seq = 0;
2487 tp->snd_cwnd_cnt = 0;
2488 tp->prior_cwnd = tp->snd_cwnd;
2489 tp->prr_delivered = 0;
2491 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2492 tcp_ecn_queue_cwr(tp);
2495 void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked, int flag)
2497 struct tcp_sock *tp = tcp_sk(sk);
2499 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2501 if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd))
2504 tp->prr_delivered += newly_acked_sacked;
2506 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2508 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2509 } else if ((flag & (FLAG_RETRANS_DATA_ACKED | FLAG_LOST_RETRANS)) ==
2510 FLAG_RETRANS_DATA_ACKED) {
2511 sndcnt = min_t(int, delta,
2512 max_t(int, tp->prr_delivered - tp->prr_out,
2513 newly_acked_sacked) + 1);
2515 sndcnt = min(delta, newly_acked_sacked);
2517 /* Force a fast retransmit upon entering fast recovery */
2518 sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1));
2519 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2522 static inline void tcp_end_cwnd_reduction(struct sock *sk)
2524 struct tcp_sock *tp = tcp_sk(sk);
2526 if (inet_csk(sk)->icsk_ca_ops->cong_control)
2529 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2530 if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH &&
2531 (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR || tp->undo_marker)) {
2532 tp->snd_cwnd = tp->snd_ssthresh;
2533 tp->snd_cwnd_stamp = tcp_jiffies32;
2535 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2538 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2539 void tcp_enter_cwr(struct sock *sk)
2541 struct tcp_sock *tp = tcp_sk(sk);
2543 tp->prior_ssthresh = 0;
2544 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2545 tp->undo_marker = 0;
2546 tcp_init_cwnd_reduction(sk);
2547 tcp_set_ca_state(sk, TCP_CA_CWR);
2550 EXPORT_SYMBOL(tcp_enter_cwr);
2552 static void tcp_try_keep_open(struct sock *sk)
2554 struct tcp_sock *tp = tcp_sk(sk);
2555 int state = TCP_CA_Open;
2557 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2558 state = TCP_CA_Disorder;
2560 if (inet_csk(sk)->icsk_ca_state != state) {
2561 tcp_set_ca_state(sk, state);
2562 tp->high_seq = tp->snd_nxt;
2566 static void tcp_try_to_open(struct sock *sk, int flag)
2568 struct tcp_sock *tp = tcp_sk(sk);
2570 tcp_verify_left_out(tp);
2572 if (!tcp_any_retrans_done(sk))
2573 tp->retrans_stamp = 0;
2575 if (flag & FLAG_ECE)
2578 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2579 tcp_try_keep_open(sk);
2583 static void tcp_mtup_probe_failed(struct sock *sk)
2585 struct inet_connection_sock *icsk = inet_csk(sk);
2587 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2588 icsk->icsk_mtup.probe_size = 0;
2589 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPFAIL);
2592 static void tcp_mtup_probe_success(struct sock *sk)
2594 struct tcp_sock *tp = tcp_sk(sk);
2595 struct inet_connection_sock *icsk = inet_csk(sk);
2598 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2600 val = (u64)tp->snd_cwnd * tcp_mss_to_mtu(sk, tp->mss_cache);
2601 do_div(val, icsk->icsk_mtup.probe_size);
2602 WARN_ON_ONCE((u32)val != val);
2603 tp->snd_cwnd = max_t(u32, 1U, val);
2605 tp->snd_cwnd_cnt = 0;
2606 tp->snd_cwnd_stamp = tcp_jiffies32;
2607 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2609 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2610 icsk->icsk_mtup.probe_size = 0;
2611 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2612 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS);
2615 /* Do a simple retransmit without using the backoff mechanisms in
2616 * tcp_timer. This is used for path mtu discovery.
2617 * The socket is already locked here.
2619 void tcp_simple_retransmit(struct sock *sk)
2621 const struct inet_connection_sock *icsk = inet_csk(sk);
2622 struct tcp_sock *tp = tcp_sk(sk);
2623 struct sk_buff *skb;
2624 unsigned int mss = tcp_current_mss(sk);
2626 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2627 if (tcp_skb_seglen(skb) > mss &&
2628 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2629 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2630 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2631 tp->retrans_out -= tcp_skb_pcount(skb);
2633 tcp_skb_mark_lost_uncond_verify(tp, skb);
2637 tcp_clear_retrans_hints_partial(tp);
2642 if (tcp_is_reno(tp))
2643 tcp_limit_reno_sacked(tp);
2645 tcp_verify_left_out(tp);
2647 /* Don't muck with the congestion window here.
2648 * Reason is that we do not increase amount of _data_
2649 * in network, but units changed and effective
2650 * cwnd/ssthresh really reduced now.
2652 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2653 tp->high_seq = tp->snd_nxt;
2654 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2655 tp->prior_ssthresh = 0;
2656 tp->undo_marker = 0;
2657 tcp_set_ca_state(sk, TCP_CA_Loss);
2659 tcp_xmit_retransmit_queue(sk);
2661 EXPORT_SYMBOL(tcp_simple_retransmit);
2663 void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2665 struct tcp_sock *tp = tcp_sk(sk);
2668 if (tcp_is_reno(tp))
2669 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2671 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2673 NET_INC_STATS(sock_net(sk), mib_idx);
2675 tp->prior_ssthresh = 0;
2678 if (!tcp_in_cwnd_reduction(sk)) {
2680 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2681 tcp_init_cwnd_reduction(sk);
2683 tcp_set_ca_state(sk, TCP_CA_Recovery);
2686 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2687 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2689 static void tcp_process_loss(struct sock *sk, int flag, int num_dupack,
2692 struct tcp_sock *tp = tcp_sk(sk);
2693 bool recovered = !before(tp->snd_una, tp->high_seq);
2695 if ((flag & FLAG_SND_UNA_ADVANCED || rcu_access_pointer(tp->fastopen_rsk)) &&
2696 tcp_try_undo_loss(sk, false))
2699 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2700 /* Step 3.b. A timeout is spurious if not all data are
2701 * lost, i.e., never-retransmitted data are (s)acked.
2703 if ((flag & FLAG_ORIG_SACK_ACKED) &&
2704 tcp_try_undo_loss(sk, true))
2707 if (after(tp->snd_nxt, tp->high_seq)) {
2708 if (flag & FLAG_DATA_SACKED || num_dupack)
2709 tp->frto = 0; /* Step 3.a. loss was real */
2710 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2711 tp->high_seq = tp->snd_nxt;
2712 /* Step 2.b. Try send new data (but deferred until cwnd
2713 * is updated in tcp_ack()). Otherwise fall back to
2714 * the conventional recovery.
2716 if (!tcp_write_queue_empty(sk) &&
2717 after(tcp_wnd_end(tp), tp->snd_nxt)) {
2718 *rexmit = REXMIT_NEW;
2726 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2727 tcp_try_undo_recovery(sk);
2730 if (tcp_is_reno(tp)) {
2731 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2732 * delivered. Lower inflight to clock out (re)tranmissions.
2734 if (after(tp->snd_nxt, tp->high_seq) && num_dupack)
2735 tcp_add_reno_sack(sk, num_dupack);
2736 else if (flag & FLAG_SND_UNA_ADVANCED)
2737 tcp_reset_reno_sack(tp);
2739 *rexmit = REXMIT_LOST;
2742 /* Undo during fast recovery after partial ACK. */
2743 static bool tcp_try_undo_partial(struct sock *sk, u32 prior_snd_una)
2745 struct tcp_sock *tp = tcp_sk(sk);
2747 if (tp->undo_marker && tcp_packet_delayed(tp)) {
2748 /* Plain luck! Hole if filled with delayed
2749 * packet, rather than with a retransmit. Check reordering.
2751 tcp_check_sack_reordering(sk, prior_snd_una, 1);
2753 /* We are getting evidence that the reordering degree is higher
2754 * than we realized. If there are no retransmits out then we
2755 * can undo. Otherwise we clock out new packets but do not
2756 * mark more packets lost or retransmit more.
2758 if (tp->retrans_out)
2761 if (!tcp_any_retrans_done(sk))
2762 tp->retrans_stamp = 0;
2764 DBGUNDO(sk, "partial recovery");
2765 tcp_undo_cwnd_reduction(sk, true);
2766 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2767 tcp_try_keep_open(sk);
2773 static void tcp_identify_packet_loss(struct sock *sk, int *ack_flag)
2775 struct tcp_sock *tp = tcp_sk(sk);
2777 if (tcp_rtx_queue_empty(sk))
2780 if (unlikely(tcp_is_reno(tp))) {
2781 tcp_newreno_mark_lost(sk, *ack_flag & FLAG_SND_UNA_ADVANCED);
2782 } else if (tcp_is_rack(sk)) {
2783 u32 prior_retrans = tp->retrans_out;
2785 if (tcp_rack_mark_lost(sk))
2786 *ack_flag &= ~FLAG_SET_XMIT_TIMER;
2787 if (prior_retrans > tp->retrans_out)
2788 *ack_flag |= FLAG_LOST_RETRANS;
2792 static bool tcp_force_fast_retransmit(struct sock *sk)
2794 struct tcp_sock *tp = tcp_sk(sk);
2796 return after(tcp_highest_sack_seq(tp),
2797 tp->snd_una + tp->reordering * tp->mss_cache);
2800 /* Process an event, which can update packets-in-flight not trivially.
2801 * Main goal of this function is to calculate new estimate for left_out,
2802 * taking into account both packets sitting in receiver's buffer and
2803 * packets lost by network.
2805 * Besides that it updates the congestion state when packet loss or ECN
2806 * is detected. But it does not reduce the cwnd, it is done by the
2807 * congestion control later.
2809 * It does _not_ decide what to send, it is made in function
2810 * tcp_xmit_retransmit_queue().
2812 static void tcp_fastretrans_alert(struct sock *sk, const u32 prior_snd_una,
2813 int num_dupack, int *ack_flag, int *rexmit)
2815 struct inet_connection_sock *icsk = inet_csk(sk);
2816 struct tcp_sock *tp = tcp_sk(sk);
2817 int fast_rexmit = 0, flag = *ack_flag;
2818 bool do_lost = num_dupack || ((flag & FLAG_DATA_SACKED) &&
2819 tcp_force_fast_retransmit(sk));
2821 if (!tp->packets_out && tp->sacked_out)
2824 /* Now state machine starts.
2825 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2826 if (flag & FLAG_ECE)
2827 tp->prior_ssthresh = 0;
2829 /* B. In all the states check for reneging SACKs. */
2830 if (tcp_check_sack_reneging(sk, flag))
2833 /* C. Check consistency of the current state. */
2834 tcp_verify_left_out(tp);
2836 /* D. Check state exit conditions. State can be terminated
2837 * when high_seq is ACKed. */
2838 if (icsk->icsk_ca_state == TCP_CA_Open) {
2839 WARN_ON(tp->retrans_out != 0);
2840 tp->retrans_stamp = 0;
2841 } else if (!before(tp->snd_una, tp->high_seq)) {
2842 switch (icsk->icsk_ca_state) {
2844 /* CWR is to be held something *above* high_seq
2845 * is ACKed for CWR bit to reach receiver. */
2846 if (tp->snd_una != tp->high_seq) {
2847 tcp_end_cwnd_reduction(sk);
2848 tcp_set_ca_state(sk, TCP_CA_Open);
2852 case TCP_CA_Recovery:
2853 if (tcp_is_reno(tp))
2854 tcp_reset_reno_sack(tp);
2855 if (tcp_try_undo_recovery(sk))
2857 tcp_end_cwnd_reduction(sk);
2862 /* E. Process state. */
2863 switch (icsk->icsk_ca_state) {
2864 case TCP_CA_Recovery:
2865 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2866 if (tcp_is_reno(tp))
2867 tcp_add_reno_sack(sk, num_dupack);
2869 if (tcp_try_undo_partial(sk, prior_snd_una))
2871 /* Partial ACK arrived. Force fast retransmit. */
2872 do_lost = tcp_is_reno(tp) ||
2873 tcp_force_fast_retransmit(sk);
2875 if (tcp_try_undo_dsack(sk)) {
2876 tcp_try_keep_open(sk);
2879 tcp_identify_packet_loss(sk, ack_flag);
2882 tcp_process_loss(sk, flag, num_dupack, rexmit);
2883 tcp_identify_packet_loss(sk, ack_flag);
2884 if (!(icsk->icsk_ca_state == TCP_CA_Open ||
2885 (*ack_flag & FLAG_LOST_RETRANS)))
2887 /* Change state if cwnd is undone or retransmits are lost */
2890 if (tcp_is_reno(tp)) {
2891 if (flag & FLAG_SND_UNA_ADVANCED)
2892 tcp_reset_reno_sack(tp);
2893 tcp_add_reno_sack(sk, num_dupack);
2896 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
2897 tcp_try_undo_dsack(sk);
2899 tcp_identify_packet_loss(sk, ack_flag);
2900 if (!tcp_time_to_recover(sk, flag)) {
2901 tcp_try_to_open(sk, flag);
2905 /* MTU probe failure: don't reduce cwnd */
2906 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2907 icsk->icsk_mtup.probe_size &&
2908 tp->snd_una == tp->mtu_probe.probe_seq_start) {
2909 tcp_mtup_probe_failed(sk);
2910 /* Restores the reduction we did in tcp_mtup_probe() */
2912 tcp_simple_retransmit(sk);
2916 /* Otherwise enter Recovery state */
2917 tcp_enter_recovery(sk, (flag & FLAG_ECE));
2921 if (!tcp_is_rack(sk) && do_lost)
2922 tcp_update_scoreboard(sk, fast_rexmit);
2923 *rexmit = REXMIT_LOST;
2926 static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us, const int flag)
2928 u32 wlen = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_rtt_wlen) * HZ;
2929 struct tcp_sock *tp = tcp_sk(sk);
2931 if ((flag & FLAG_ACK_MAYBE_DELAYED) && rtt_us > tcp_min_rtt(tp)) {
2932 /* If the remote keeps returning delayed ACKs, eventually
2933 * the min filter would pick it up and overestimate the
2934 * prop. delay when it expires. Skip suspected delayed ACKs.
2938 minmax_running_min(&tp->rtt_min, wlen, tcp_jiffies32,
2939 rtt_us ? : jiffies_to_usecs(1));
2942 static bool tcp_ack_update_rtt(struct sock *sk, const int flag,
2943 long seq_rtt_us, long sack_rtt_us,
2944 long ca_rtt_us, struct rate_sample *rs)
2946 const struct tcp_sock *tp = tcp_sk(sk);
2948 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2949 * broken middle-boxes or peers may corrupt TS-ECR fields. But
2950 * Karn's algorithm forbids taking RTT if some retransmitted data
2951 * is acked (RFC6298).
2954 seq_rtt_us = sack_rtt_us;
2956 /* RTTM Rule: A TSecr value received in a segment is used to
2957 * update the averaged RTT measurement only if the segment
2958 * acknowledges some new data, i.e., only if it advances the
2959 * left edge of the send window.
2960 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2962 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2963 flag & FLAG_ACKED) {
2964 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr;
2966 if (likely(delta < INT_MAX / (USEC_PER_SEC / TCP_TS_HZ))) {
2969 seq_rtt_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
2970 ca_rtt_us = seq_rtt_us;
2973 rs->rtt_us = ca_rtt_us; /* RTT of last (S)ACKed packet (or -1) */
2977 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
2978 * always taken together with ACK, SACK, or TS-opts. Any negative
2979 * values will be skipped with the seq_rtt_us < 0 check above.
2981 tcp_update_rtt_min(sk, ca_rtt_us, flag);
2982 tcp_rtt_estimator(sk, seq_rtt_us);
2985 /* RFC6298: only reset backoff on valid RTT measurement. */
2986 inet_csk(sk)->icsk_backoff = 0;
2990 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
2991 void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req)
2993 struct rate_sample rs;
2996 if (req && !req->num_retrans && tcp_rsk(req)->snt_synack)
2997 rtt_us = tcp_stamp_us_delta(tcp_clock_us(), tcp_rsk(req)->snt_synack);
2999 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us, &rs);
3003 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
3005 const struct inet_connection_sock *icsk = inet_csk(sk);
3007 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
3008 tcp_sk(sk)->snd_cwnd_stamp = tcp_jiffies32;
3011 /* Restart timer after forward progress on connection.
3012 * RFC2988 recommends to restart timer to now+rto.
3014 void tcp_rearm_rto(struct sock *sk)
3016 const struct inet_connection_sock *icsk = inet_csk(sk);
3017 struct tcp_sock *tp = tcp_sk(sk);
3019 /* If the retrans timer is currently being used by Fast Open
3020 * for SYN-ACK retrans purpose, stay put.
3022 if (rcu_access_pointer(tp->fastopen_rsk))
3025 if (!tp->packets_out) {
3026 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
3028 u32 rto = inet_csk(sk)->icsk_rto;
3029 /* Offset the time elapsed after installing regular RTO */
3030 if (icsk->icsk_pending == ICSK_TIME_REO_TIMEOUT ||
3031 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
3032 s64 delta_us = tcp_rto_delta_us(sk);
3033 /* delta_us may not be positive if the socket is locked
3034 * when the retrans timer fires and is rescheduled.
3036 rto = usecs_to_jiffies(max_t(int, delta_us, 1));
3038 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
3039 TCP_RTO_MAX, tcp_rtx_queue_head(sk));
3043 /* Try to schedule a loss probe; if that doesn't work, then schedule an RTO. */
3044 static void tcp_set_xmit_timer(struct sock *sk)
3046 if (!tcp_schedule_loss_probe(sk, true))
3050 /* If we get here, the whole TSO packet has not been acked. */
3051 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3053 struct tcp_sock *tp = tcp_sk(sk);
3056 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3058 packets_acked = tcp_skb_pcount(skb);
3059 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3061 packets_acked -= tcp_skb_pcount(skb);
3063 if (packets_acked) {
3064 BUG_ON(tcp_skb_pcount(skb) == 0);
3065 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3068 return packets_acked;
3071 static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
3074 const struct skb_shared_info *shinfo;
3076 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3077 if (likely(!TCP_SKB_CB(skb)->txstamp_ack))
3080 shinfo = skb_shinfo(skb);
3081 if (!before(shinfo->tskey, prior_snd_una) &&
3082 before(shinfo->tskey, tcp_sk(sk)->snd_una)) {
3083 tcp_skb_tsorted_save(skb) {
3084 __skb_tstamp_tx(skb, NULL, sk, SCM_TSTAMP_ACK);
3085 } tcp_skb_tsorted_restore(skb);
3089 /* Remove acknowledged frames from the retransmission queue. If our packet
3090 * is before the ack sequence we can discard it as it's confirmed to have
3091 * arrived at the other end.
3093 static int tcp_clean_rtx_queue(struct sock *sk, u32 prior_fack,
3095 struct tcp_sacktag_state *sack)
3097 const struct inet_connection_sock *icsk = inet_csk(sk);
3098 u64 first_ackt, last_ackt;
3099 struct tcp_sock *tp = tcp_sk(sk);
3100 u32 prior_sacked = tp->sacked_out;
3101 u32 reord = tp->snd_nxt; /* lowest acked un-retx un-sacked seq */
3102 struct sk_buff *skb, *next;
3103 bool fully_acked = true;
3104 long sack_rtt_us = -1L;
3105 long seq_rtt_us = -1L;
3106 long ca_rtt_us = -1L;
3108 u32 last_in_flight = 0;
3114 for (skb = skb_rb_first(&sk->tcp_rtx_queue); skb; skb = next) {
3115 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3116 const u32 start_seq = scb->seq;
3117 u8 sacked = scb->sacked;
3120 tcp_ack_tstamp(sk, skb, prior_snd_una);
3122 /* Determine how many packets and what bytes were acked, tso and else */
3123 if (after(scb->end_seq, tp->snd_una)) {
3124 if (tcp_skb_pcount(skb) == 1 ||
3125 !after(tp->snd_una, scb->seq))
3128 acked_pcount = tcp_tso_acked(sk, skb);
3131 fully_acked = false;
3133 acked_pcount = tcp_skb_pcount(skb);
3136 if (unlikely(sacked & TCPCB_RETRANS)) {
3137 if (sacked & TCPCB_SACKED_RETRANS)
3138 tp->retrans_out -= acked_pcount;
3139 flag |= FLAG_RETRANS_DATA_ACKED;
3140 } else if (!(sacked & TCPCB_SACKED_ACKED)) {
3141 last_ackt = tcp_skb_timestamp_us(skb);
3142 WARN_ON_ONCE(last_ackt == 0);
3144 first_ackt = last_ackt;
3146 last_in_flight = TCP_SKB_CB(skb)->tx.in_flight;
3147 if (before(start_seq, reord))
3149 if (!after(scb->end_seq, tp->high_seq))
3150 flag |= FLAG_ORIG_SACK_ACKED;
3153 if (sacked & TCPCB_SACKED_ACKED) {
3154 tp->sacked_out -= acked_pcount;
3155 } else if (tcp_is_sack(tp)) {
3156 tp->delivered += acked_pcount;
3157 if (!tcp_skb_spurious_retrans(tp, skb))
3158 tcp_rack_advance(tp, sacked, scb->end_seq,
3159 tcp_skb_timestamp_us(skb));
3161 if (sacked & TCPCB_LOST)
3162 tp->lost_out -= acked_pcount;
3164 tp->packets_out -= acked_pcount;
3165 pkts_acked += acked_pcount;
3166 tcp_rate_skb_delivered(sk, skb, sack->rate);
3168 /* Initial outgoing SYN's get put onto the write_queue
3169 * just like anything else we transmit. It is not
3170 * true data, and if we misinform our callers that
3171 * this ACK acks real data, we will erroneously exit
3172 * connection startup slow start one packet too
3173 * quickly. This is severely frowned upon behavior.
3175 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
3176 flag |= FLAG_DATA_ACKED;
3178 flag |= FLAG_SYN_ACKED;
3179 tp->retrans_stamp = 0;
3185 next = skb_rb_next(skb);
3186 if (unlikely(skb == tp->retransmit_skb_hint))
3187 tp->retransmit_skb_hint = NULL;
3188 if (unlikely(skb == tp->lost_skb_hint))
3189 tp->lost_skb_hint = NULL;
3190 tcp_highest_sack_replace(sk, skb, next);
3191 tcp_rtx_queue_unlink_and_free(skb, sk);
3195 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
3197 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3198 tp->snd_up = tp->snd_una;
3200 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3201 flag |= FLAG_SACK_RENEGING;
3203 if (likely(first_ackt) && !(flag & FLAG_RETRANS_DATA_ACKED)) {
3204 seq_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, first_ackt);
3205 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, last_ackt);
3207 if (pkts_acked == 1 && last_in_flight < tp->mss_cache &&
3208 last_in_flight && !prior_sacked && fully_acked &&
3209 sack->rate->prior_delivered + 1 == tp->delivered &&
3210 !(flag & (FLAG_CA_ALERT | FLAG_SYN_ACKED))) {
3211 /* Conservatively mark a delayed ACK. It's typically
3212 * from a lone runt packet over the round trip to
3213 * a receiver w/o out-of-order or CE events.
3215 flag |= FLAG_ACK_MAYBE_DELAYED;
3218 if (sack->first_sackt) {
3219 sack_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->first_sackt);
3220 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->last_sackt);
3222 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us,
3223 ca_rtt_us, sack->rate);
3225 if (flag & FLAG_ACKED) {
3226 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3227 if (unlikely(icsk->icsk_mtup.probe_size &&
3228 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3229 tcp_mtup_probe_success(sk);
3232 if (tcp_is_reno(tp)) {
3233 tcp_remove_reno_sacks(sk, pkts_acked);
3235 /* If any of the cumulatively ACKed segments was
3236 * retransmitted, non-SACK case cannot confirm that
3237 * progress was due to original transmission due to
3238 * lack of TCPCB_SACKED_ACKED bits even if some of
3239 * the packets may have been never retransmitted.
3241 if (flag & FLAG_RETRANS_DATA_ACKED)
3242 flag &= ~FLAG_ORIG_SACK_ACKED;
3246 /* Non-retransmitted hole got filled? That's reordering */
3247 if (before(reord, prior_fack))
3248 tcp_check_sack_reordering(sk, reord, 0);
3250 delta = prior_sacked - tp->sacked_out;
3251 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3253 } else if (skb && rtt_update && sack_rtt_us >= 0 &&
3254 sack_rtt_us > tcp_stamp_us_delta(tp->tcp_mstamp,
3255 tcp_skb_timestamp_us(skb))) {
3256 /* Do not re-arm RTO if the sack RTT is measured from data sent
3257 * after when the head was last (re)transmitted. Otherwise the
3258 * timeout may continue to extend in loss recovery.
3260 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3263 if (icsk->icsk_ca_ops->pkts_acked) {
3264 struct ack_sample sample = { .pkts_acked = pkts_acked,
3265 .rtt_us = sack->rate->rtt_us,
3266 .in_flight = last_in_flight };
3268 icsk->icsk_ca_ops->pkts_acked(sk, &sample);
3271 #if FASTRETRANS_DEBUG > 0
3272 WARN_ON((int)tp->sacked_out < 0);
3273 WARN_ON((int)tp->lost_out < 0);
3274 WARN_ON((int)tp->retrans_out < 0);
3275 if (!tp->packets_out && tcp_is_sack(tp)) {
3276 icsk = inet_csk(sk);
3278 pr_debug("Leak l=%u %d\n",
3279 tp->lost_out, icsk->icsk_ca_state);
3282 if (tp->sacked_out) {
3283 pr_debug("Leak s=%u %d\n",
3284 tp->sacked_out, icsk->icsk_ca_state);
3287 if (tp->retrans_out) {
3288 pr_debug("Leak r=%u %d\n",
3289 tp->retrans_out, icsk->icsk_ca_state);
3290 tp->retrans_out = 0;
3297 static void tcp_ack_probe(struct sock *sk)
3299 struct inet_connection_sock *icsk = inet_csk(sk);
3300 struct sk_buff *head = tcp_send_head(sk);
3301 const struct tcp_sock *tp = tcp_sk(sk);
3303 /* Was it a usable window open? */
3306 if (!after(TCP_SKB_CB(head)->end_seq, tcp_wnd_end(tp))) {
3307 icsk->icsk_backoff = 0;
3308 icsk->icsk_probes_tstamp = 0;
3309 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3310 /* Socket must be waked up by subsequent tcp_data_snd_check().
3311 * This function is not for random using!
3314 unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX);
3316 when = tcp_clamp_probe0_to_user_timeout(sk, when);
3317 tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3318 when, TCP_RTO_MAX, NULL);
3322 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3324 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3325 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3328 /* Decide wheather to run the increase function of congestion control. */
3329 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3331 /* If reordering is high then always grow cwnd whenever data is
3332 * delivered regardless of its ordering. Otherwise stay conservative
3333 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3334 * new SACK or ECE mark may first advance cwnd here and later reduce
3335 * cwnd in tcp_fastretrans_alert() based on more states.
3337 if (tcp_sk(sk)->reordering >
3338 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_reordering))
3339 return flag & FLAG_FORWARD_PROGRESS;
3341 return flag & FLAG_DATA_ACKED;
3344 /* The "ultimate" congestion control function that aims to replace the rigid
3345 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction).
3346 * It's called toward the end of processing an ACK with precise rate
3347 * information. All transmission or retransmission are delayed afterwards.
3349 static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked,
3350 int flag, const struct rate_sample *rs)
3352 const struct inet_connection_sock *icsk = inet_csk(sk);
3354 if (icsk->icsk_ca_ops->cong_control) {
3355 icsk->icsk_ca_ops->cong_control(sk, rs);
3359 if (tcp_in_cwnd_reduction(sk)) {
3360 /* Reduce cwnd if state mandates */
3361 tcp_cwnd_reduction(sk, acked_sacked, flag);
3362 } else if (tcp_may_raise_cwnd(sk, flag)) {
3363 /* Advance cwnd if state allows */
3364 tcp_cong_avoid(sk, ack, acked_sacked);
3366 tcp_update_pacing_rate(sk);
3369 /* Check that window update is acceptable.
3370 * The function assumes that snd_una<=ack<=snd_next.
3372 static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3373 const u32 ack, const u32 ack_seq,
3376 return after(ack, tp->snd_una) ||
3377 after(ack_seq, tp->snd_wl1) ||
3378 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3381 /* If we update tp->snd_una, also update tp->bytes_acked */
3382 static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack)
3384 u32 delta = ack - tp->snd_una;
3386 sock_owned_by_me((struct sock *)tp);
3387 tp->bytes_acked += delta;
3391 /* If we update tp->rcv_nxt, also update tp->bytes_received */
3392 static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq)
3394 u32 delta = seq - tp->rcv_nxt;
3396 sock_owned_by_me((struct sock *)tp);
3397 tp->bytes_received += delta;
3398 WRITE_ONCE(tp->rcv_nxt, seq);
3401 /* Update our send window.
3403 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3404 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3406 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3409 struct tcp_sock *tp = tcp_sk(sk);
3411 u32 nwin = ntohs(tcp_hdr(skb)->window);
3413 if (likely(!tcp_hdr(skb)->syn))
3414 nwin <<= tp->rx_opt.snd_wscale;
3416 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3417 flag |= FLAG_WIN_UPDATE;
3418 tcp_update_wl(tp, ack_seq);
3420 if (tp->snd_wnd != nwin) {
3423 /* Note, it is the only place, where
3424 * fast path is recovered for sending TCP.
3427 tcp_fast_path_check(sk);
3429 if (!tcp_write_queue_empty(sk))
3430 tcp_slow_start_after_idle_check(sk);
3432 if (nwin > tp->max_window) {
3433 tp->max_window = nwin;
3434 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3439 tcp_snd_una_update(tp, ack);
3444 static bool __tcp_oow_rate_limited(struct net *net, int mib_idx,
3445 u32 *last_oow_ack_time)
3447 /* Paired with the WRITE_ONCE() in this function. */
3448 u32 val = READ_ONCE(*last_oow_ack_time);
3451 s32 elapsed = (s32)(tcp_jiffies32 - val);
3454 elapsed < READ_ONCE(net->ipv4.sysctl_tcp_invalid_ratelimit)) {
3455 NET_INC_STATS(net, mib_idx);
3456 return true; /* rate-limited: don't send yet! */
3460 /* Paired with the prior READ_ONCE() and with itself,
3461 * as we might be lockless.
3463 WRITE_ONCE(*last_oow_ack_time, tcp_jiffies32);
3465 return false; /* not rate-limited: go ahead, send dupack now! */
3468 /* Return true if we're currently rate-limiting out-of-window ACKs and
3469 * thus shouldn't send a dupack right now. We rate-limit dupacks in
3470 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3471 * attacks that send repeated SYNs or ACKs for the same connection. To
3472 * do this, we do not send a duplicate SYNACK or ACK if the remote
3473 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3475 bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb,
3476 int mib_idx, u32 *last_oow_ack_time)
3478 /* Data packets without SYNs are not likely part of an ACK loop. */
3479 if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) &&
3483 return __tcp_oow_rate_limited(net, mib_idx, last_oow_ack_time);
3486 /* RFC 5961 7 [ACK Throttling] */
3487 static void tcp_send_challenge_ack(struct sock *sk, const struct sk_buff *skb)
3489 /* unprotected vars, we dont care of overwrites */
3490 static u32 challenge_timestamp;
3491 static unsigned int challenge_count;
3492 struct tcp_sock *tp = tcp_sk(sk);
3493 struct net *net = sock_net(sk);
3496 /* First check our per-socket dupack rate limit. */
3497 if (__tcp_oow_rate_limited(net,
3498 LINUX_MIB_TCPACKSKIPPEDCHALLENGE,
3499 &tp->last_oow_ack_time))
3502 /* Then check host-wide RFC 5961 rate limit. */
3504 if (now != READ_ONCE(challenge_timestamp)) {
3505 u32 ack_limit = READ_ONCE(net->ipv4.sysctl_tcp_challenge_ack_limit);
3506 u32 half = (ack_limit + 1) >> 1;
3508 WRITE_ONCE(challenge_timestamp, now);
3509 WRITE_ONCE(challenge_count, half + prandom_u32_max(ack_limit));
3511 count = READ_ONCE(challenge_count);
3513 WRITE_ONCE(challenge_count, count - 1);
3514 NET_INC_STATS(net, LINUX_MIB_TCPCHALLENGEACK);
3519 static void tcp_store_ts_recent(struct tcp_sock *tp)
3521 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3522 tp->rx_opt.ts_recent_stamp = ktime_get_seconds();
3525 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3527 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3528 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3529 * extra check below makes sure this can only happen
3530 * for pure ACK frames. -DaveM
3532 * Not only, also it occurs for expired timestamps.
3535 if (tcp_paws_check(&tp->rx_opt, 0))
3536 tcp_store_ts_recent(tp);
3540 /* This routine deals with acks during a TLP episode and ends an episode by
3541 * resetting tlp_high_seq. Ref: TLP algorithm in draft-ietf-tcpm-rack
3543 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3545 struct tcp_sock *tp = tcp_sk(sk);
3547 if (before(ack, tp->tlp_high_seq))
3550 if (!tp->tlp_retrans) {
3551 /* TLP of new data has been acknowledged */
3552 tp->tlp_high_seq = 0;
3553 } else if (flag & FLAG_DSACKING_ACK) {
3554 /* This DSACK means original and TLP probe arrived; no loss */
3555 tp->tlp_high_seq = 0;
3556 } else if (after(ack, tp->tlp_high_seq)) {
3557 /* ACK advances: there was a loss, so reduce cwnd. Reset
3558 * tlp_high_seq in tcp_init_cwnd_reduction()
3560 tcp_init_cwnd_reduction(sk);
3561 tcp_set_ca_state(sk, TCP_CA_CWR);
3562 tcp_end_cwnd_reduction(sk);
3563 tcp_try_keep_open(sk);
3564 NET_INC_STATS(sock_net(sk),
3565 LINUX_MIB_TCPLOSSPROBERECOVERY);
3566 } else if (!(flag & (FLAG_SND_UNA_ADVANCED |
3567 FLAG_NOT_DUP | FLAG_DATA_SACKED))) {
3568 /* Pure dupack: original and TLP probe arrived; no loss */
3569 tp->tlp_high_seq = 0;
3573 static inline void tcp_in_ack_event(struct sock *sk, u32 flags)
3575 const struct inet_connection_sock *icsk = inet_csk(sk);
3577 if (icsk->icsk_ca_ops->in_ack_event)
3578 icsk->icsk_ca_ops->in_ack_event(sk, flags);
3581 /* Congestion control has updated the cwnd already. So if we're in
3582 * loss recovery then now we do any new sends (for FRTO) or
3583 * retransmits (for CA_Loss or CA_recovery) that make sense.
3585 static void tcp_xmit_recovery(struct sock *sk, int rexmit)
3587 struct tcp_sock *tp = tcp_sk(sk);
3589 if (rexmit == REXMIT_NONE || sk->sk_state == TCP_SYN_SENT)
3592 if (unlikely(rexmit == 2)) {
3593 __tcp_push_pending_frames(sk, tcp_current_mss(sk),
3595 if (after(tp->snd_nxt, tp->high_seq))
3599 tcp_xmit_retransmit_queue(sk);
3602 /* Returns the number of packets newly acked or sacked by the current ACK */
3603 static u32 tcp_newly_delivered(struct sock *sk, u32 prior_delivered, int flag)
3605 const struct net *net = sock_net(sk);
3606 struct tcp_sock *tp = tcp_sk(sk);
3609 delivered = tp->delivered - prior_delivered;
3610 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVERED, delivered);
3611 if (flag & FLAG_ECE) {
3612 tp->delivered_ce += delivered;
3613 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVEREDCE, delivered);
3618 /* This routine deals with incoming acks, but not outgoing ones. */
3619 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3621 struct inet_connection_sock *icsk = inet_csk(sk);
3622 struct tcp_sock *tp = tcp_sk(sk);
3623 struct tcp_sacktag_state sack_state;
3624 struct rate_sample rs = { .prior_delivered = 0 };
3625 u32 prior_snd_una = tp->snd_una;
3626 bool is_sack_reneg = tp->is_sack_reneg;
3627 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3628 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3630 int prior_packets = tp->packets_out;
3631 u32 delivered = tp->delivered;
3632 u32 lost = tp->lost;
3633 int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */
3636 sack_state.first_sackt = 0;
3637 sack_state.rate = &rs;
3639 /* We very likely will need to access rtx queue. */
3640 prefetch(sk->tcp_rtx_queue.rb_node);
3642 /* If the ack is older than previous acks
3643 * then we can probably ignore it.
3645 if (before(ack, prior_snd_una)) {
3646 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3647 if (before(ack, prior_snd_una - tp->max_window)) {
3648 if (!(flag & FLAG_NO_CHALLENGE_ACK))
3649 tcp_send_challenge_ack(sk, skb);
3655 /* If the ack includes data we haven't sent yet, discard
3656 * this segment (RFC793 Section 3.9).
3658 if (after(ack, tp->snd_nxt))
3661 if (after(ack, prior_snd_una)) {
3662 flag |= FLAG_SND_UNA_ADVANCED;
3663 icsk->icsk_retransmits = 0;
3665 #if IS_ENABLED(CONFIG_TLS_DEVICE)
3666 if (static_branch_unlikely(&clean_acked_data_enabled.key))
3667 if (icsk->icsk_clean_acked)
3668 icsk->icsk_clean_acked(sk, ack);
3672 prior_fack = tcp_is_sack(tp) ? tcp_highest_sack_seq(tp) : tp->snd_una;
3673 rs.prior_in_flight = tcp_packets_in_flight(tp);
3675 /* ts_recent update must be made after we are sure that the packet
3678 if (flag & FLAG_UPDATE_TS_RECENT)
3679 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3681 if ((flag & (FLAG_SLOWPATH | FLAG_SND_UNA_ADVANCED)) ==
3682 FLAG_SND_UNA_ADVANCED) {
3683 /* Window is constant, pure forward advance.
3684 * No more checks are required.
3685 * Note, we use the fact that SND.UNA>=SND.WL2.
3687 tcp_update_wl(tp, ack_seq);
3688 tcp_snd_una_update(tp, ack);
3689 flag |= FLAG_WIN_UPDATE;
3691 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
3693 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPACKS);
3695 u32 ack_ev_flags = CA_ACK_SLOWPATH;
3697 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3700 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3702 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3704 if (TCP_SKB_CB(skb)->sacked)
3705 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3708 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) {
3710 ack_ev_flags |= CA_ACK_ECE;
3713 if (flag & FLAG_WIN_UPDATE)
3714 ack_ev_flags |= CA_ACK_WIN_UPDATE;
3716 tcp_in_ack_event(sk, ack_ev_flags);
3719 /* This is a deviation from RFC3168 since it states that:
3720 * "When the TCP data sender is ready to set the CWR bit after reducing
3721 * the congestion window, it SHOULD set the CWR bit only on the first
3722 * new data packet that it transmits."
3723 * We accept CWR on pure ACKs to be more robust
3724 * with widely-deployed TCP implementations that do this.
3726 tcp_ecn_accept_cwr(sk, skb);
3728 /* We passed data and got it acked, remove any soft error
3729 * log. Something worked...
3731 sk->sk_err_soft = 0;
3732 icsk->icsk_probes_out = 0;
3733 tp->rcv_tstamp = tcp_jiffies32;
3737 /* See if we can take anything off of the retransmit queue. */
3738 flag |= tcp_clean_rtx_queue(sk, prior_fack, prior_snd_una, &sack_state);
3740 tcp_rack_update_reo_wnd(sk, &rs);
3742 if (tp->tlp_high_seq)
3743 tcp_process_tlp_ack(sk, ack, flag);
3745 if (tcp_ack_is_dubious(sk, flag)) {
3746 if (!(flag & (FLAG_SND_UNA_ADVANCED |
3747 FLAG_NOT_DUP | FLAG_DSACKING_ACK))) {
3749 /* Consider if pure acks were aggregated in tcp_add_backlog() */
3750 if (!(flag & FLAG_DATA))
3751 num_dupack = max_t(u16, 1, skb_shinfo(skb)->gso_segs);
3753 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
3757 /* If needed, reset TLP/RTO timer when RACK doesn't set. */
3758 if (flag & FLAG_SET_XMIT_TIMER)
3759 tcp_set_xmit_timer(sk);
3761 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
3764 delivered = tcp_newly_delivered(sk, delivered, flag);
3765 lost = tp->lost - lost; /* freshly marked lost */
3766 rs.is_ack_delayed = !!(flag & FLAG_ACK_MAYBE_DELAYED);
3767 tcp_rate_gen(sk, delivered, lost, is_sack_reneg, sack_state.rate);
3768 tcp_cong_control(sk, ack, delivered, flag, sack_state.rate);
3769 tcp_xmit_recovery(sk, rexmit);
3773 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3774 if (flag & FLAG_DSACKING_ACK) {
3775 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
3777 tcp_newly_delivered(sk, delivered, flag);
3779 /* If this ack opens up a zero window, clear backoff. It was
3780 * being used to time the probes, and is probably far higher than
3781 * it needs to be for normal retransmission.
3785 if (tp->tlp_high_seq)
3786 tcp_process_tlp_ack(sk, ack, flag);
3790 /* If data was SACKed, tag it and see if we should send more data.
3791 * If data was DSACKed, see if we can undo a cwnd reduction.
3793 if (TCP_SKB_CB(skb)->sacked) {
3794 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3796 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
3798 tcp_newly_delivered(sk, delivered, flag);
3799 tcp_xmit_recovery(sk, rexmit);
3805 static void tcp_parse_fastopen_option(int len, const unsigned char *cookie,
3806 bool syn, struct tcp_fastopen_cookie *foc,
3809 /* Valid only in SYN or SYN-ACK with an even length. */
3810 if (!foc || !syn || len < 0 || (len & 1))
3813 if (len >= TCP_FASTOPEN_COOKIE_MIN &&
3814 len <= TCP_FASTOPEN_COOKIE_MAX)
3815 memcpy(foc->val, cookie, len);
3822 static void smc_parse_options(const struct tcphdr *th,
3823 struct tcp_options_received *opt_rx,
3824 const unsigned char *ptr,
3827 #if IS_ENABLED(CONFIG_SMC)
3828 if (static_branch_unlikely(&tcp_have_smc)) {
3829 if (th->syn && !(opsize & 1) &&
3830 opsize >= TCPOLEN_EXP_SMC_BASE &&
3831 get_unaligned_be32(ptr) == TCPOPT_SMC_MAGIC)
3837 /* Try to parse the MSS option from the TCP header. Return 0 on failure, clamped
3840 static u16 tcp_parse_mss_option(const struct tcphdr *th, u16 user_mss)
3842 const unsigned char *ptr = (const unsigned char *)(th + 1);
3843 int length = (th->doff * 4) - sizeof(struct tcphdr);
3846 while (length > 0) {
3847 int opcode = *ptr++;
3853 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3860 if (opsize < 2) /* "silly options" */
3862 if (opsize > length)
3863 return mss; /* fail on partial options */
3864 if (opcode == TCPOPT_MSS && opsize == TCPOLEN_MSS) {
3865 u16 in_mss = get_unaligned_be16(ptr);
3868 if (user_mss && user_mss < in_mss)
3880 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3881 * But, this can also be called on packets in the established flow when
3882 * the fast version below fails.
3884 void tcp_parse_options(const struct net *net,
3885 const struct sk_buff *skb,
3886 struct tcp_options_received *opt_rx, int estab,
3887 struct tcp_fastopen_cookie *foc)
3889 const unsigned char *ptr;
3890 const struct tcphdr *th = tcp_hdr(skb);
3891 int length = (th->doff * 4) - sizeof(struct tcphdr);
3893 ptr = (const unsigned char *)(th + 1);
3894 opt_rx->saw_tstamp = 0;
3896 while (length > 0) {
3897 int opcode = *ptr++;
3903 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3910 if (opsize < 2) /* "silly options" */
3912 if (opsize > length)
3913 return; /* don't parse partial options */
3916 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3917 u16 in_mss = get_unaligned_be16(ptr);
3919 if (opt_rx->user_mss &&
3920 opt_rx->user_mss < in_mss)
3921 in_mss = opt_rx->user_mss;
3922 opt_rx->mss_clamp = in_mss;
3927 if (opsize == TCPOLEN_WINDOW && th->syn &&
3928 !estab && READ_ONCE(net->ipv4.sysctl_tcp_window_scaling)) {
3929 __u8 snd_wscale = *(__u8 *)ptr;
3930 opt_rx->wscale_ok = 1;
3931 if (snd_wscale > TCP_MAX_WSCALE) {
3932 net_info_ratelimited("%s: Illegal window scaling value %d > %u received\n",
3936 snd_wscale = TCP_MAX_WSCALE;
3938 opt_rx->snd_wscale = snd_wscale;
3941 case TCPOPT_TIMESTAMP:
3942 if ((opsize == TCPOLEN_TIMESTAMP) &&
3943 ((estab && opt_rx->tstamp_ok) ||
3944 (!estab && READ_ONCE(net->ipv4.sysctl_tcp_timestamps)))) {
3945 opt_rx->saw_tstamp = 1;
3946 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3947 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3950 case TCPOPT_SACK_PERM:
3951 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3952 !estab && READ_ONCE(net->ipv4.sysctl_tcp_sack)) {
3953 opt_rx->sack_ok = TCP_SACK_SEEN;
3954 tcp_sack_reset(opt_rx);
3959 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3960 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3962 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3965 #ifdef CONFIG_TCP_MD5SIG
3968 * The MD5 Hash has already been
3969 * checked (see tcp_v{4,6}_do_rcv()).
3973 case TCPOPT_FASTOPEN:
3974 tcp_parse_fastopen_option(
3975 opsize - TCPOLEN_FASTOPEN_BASE,
3976 ptr, th->syn, foc, false);
3980 /* Fast Open option shares code 254 using a
3981 * 16 bits magic number.
3983 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE &&
3984 get_unaligned_be16(ptr) ==
3985 TCPOPT_FASTOPEN_MAGIC)
3986 tcp_parse_fastopen_option(opsize -
3987 TCPOLEN_EXP_FASTOPEN_BASE,
3988 ptr + 2, th->syn, foc, true);
3990 smc_parse_options(th, opt_rx, ptr,
4000 EXPORT_SYMBOL(tcp_parse_options);
4002 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
4004 const __be32 *ptr = (const __be32 *)(th + 1);
4006 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
4007 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
4008 tp->rx_opt.saw_tstamp = 1;
4010 tp->rx_opt.rcv_tsval = ntohl(*ptr);
4013 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
4015 tp->rx_opt.rcv_tsecr = 0;
4021 /* Fast parse options. This hopes to only see timestamps.
4022 * If it is wrong it falls back on tcp_parse_options().
4024 static bool tcp_fast_parse_options(const struct net *net,
4025 const struct sk_buff *skb,
4026 const struct tcphdr *th, struct tcp_sock *tp)
4028 /* In the spirit of fast parsing, compare doff directly to constant
4029 * values. Because equality is used, short doff can be ignored here.
4031 if (th->doff == (sizeof(*th) / 4)) {
4032 tp->rx_opt.saw_tstamp = 0;
4034 } else if (tp->rx_opt.tstamp_ok &&
4035 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
4036 if (tcp_parse_aligned_timestamp(tp, th))
4040 tcp_parse_options(net, skb, &tp->rx_opt, 1, NULL);
4041 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
4042 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
4047 #ifdef CONFIG_TCP_MD5SIG
4049 * Parse MD5 Signature option
4051 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
4053 int length = (th->doff << 2) - sizeof(*th);
4054 const u8 *ptr = (const u8 *)(th + 1);
4056 /* If not enough data remaining, we can short cut */
4057 while (length >= TCPOLEN_MD5SIG) {
4058 int opcode = *ptr++;
4069 if (opsize < 2 || opsize > length)
4071 if (opcode == TCPOPT_MD5SIG)
4072 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
4079 EXPORT_SYMBOL(tcp_parse_md5sig_option);
4082 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
4084 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
4085 * it can pass through stack. So, the following predicate verifies that
4086 * this segment is not used for anything but congestion avoidance or
4087 * fast retransmit. Moreover, we even are able to eliminate most of such
4088 * second order effects, if we apply some small "replay" window (~RTO)
4089 * to timestamp space.
4091 * All these measures still do not guarantee that we reject wrapped ACKs
4092 * on networks with high bandwidth, when sequence space is recycled fastly,
4093 * but it guarantees that such events will be very rare and do not affect
4094 * connection seriously. This doesn't look nice, but alas, PAWS is really
4097 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
4098 * states that events when retransmit arrives after original data are rare.
4099 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
4100 * the biggest problem on large power networks even with minor reordering.
4101 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
4102 * up to bandwidth of 18Gigabit/sec. 8) ]
4105 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
4107 const struct tcp_sock *tp = tcp_sk(sk);
4108 const struct tcphdr *th = tcp_hdr(skb);
4109 u32 seq = TCP_SKB_CB(skb)->seq;
4110 u32 ack = TCP_SKB_CB(skb)->ack_seq;
4112 return (/* 1. Pure ACK with correct sequence number. */
4113 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
4115 /* 2. ... and duplicate ACK. */
4116 ack == tp->snd_una &&
4118 /* 3. ... and does not update window. */
4119 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
4121 /* 4. ... and sits in replay window. */
4122 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
4125 static inline bool tcp_paws_discard(const struct sock *sk,
4126 const struct sk_buff *skb)
4128 const struct tcp_sock *tp = tcp_sk(sk);
4130 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
4131 !tcp_disordered_ack(sk, skb);
4134 /* Check segment sequence number for validity.
4136 * Segment controls are considered valid, if the segment
4137 * fits to the window after truncation to the window. Acceptability
4138 * of data (and SYN, FIN, of course) is checked separately.
4139 * See tcp_data_queue(), for example.
4141 * Also, controls (RST is main one) are accepted using RCV.WUP instead
4142 * of RCV.NXT. Peer still did not advance his SND.UNA when we
4143 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4144 * (borrowed from freebsd)
4147 static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
4149 return !before(end_seq, tp->rcv_wup) &&
4150 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
4153 /* When we get a reset we do this. */
4154 void tcp_reset(struct sock *sk)
4156 trace_tcp_receive_reset(sk);
4158 /* We want the right error as BSD sees it (and indeed as we do). */
4159 switch (sk->sk_state) {
4161 sk->sk_err = ECONNREFUSED;
4163 case TCP_CLOSE_WAIT:
4169 sk->sk_err = ECONNRESET;
4171 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4174 tcp_write_queue_purge(sk);
4177 if (!sock_flag(sk, SOCK_DEAD))
4178 sk->sk_error_report(sk);
4182 * Process the FIN bit. This now behaves as it is supposed to work
4183 * and the FIN takes effect when it is validly part of sequence
4184 * space. Not before when we get holes.
4186 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4187 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4190 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4191 * close and we go into CLOSING (and later onto TIME-WAIT)
4193 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4195 void tcp_fin(struct sock *sk)
4197 struct tcp_sock *tp = tcp_sk(sk);
4199 inet_csk_schedule_ack(sk);
4201 sk->sk_shutdown |= RCV_SHUTDOWN;
4202 sock_set_flag(sk, SOCK_DONE);
4204 switch (sk->sk_state) {
4206 case TCP_ESTABLISHED:
4207 /* Move to CLOSE_WAIT */
4208 tcp_set_state(sk, TCP_CLOSE_WAIT);
4209 inet_csk_enter_pingpong_mode(sk);
4212 case TCP_CLOSE_WAIT:
4214 /* Received a retransmission of the FIN, do
4219 /* RFC793: Remain in the LAST-ACK state. */
4223 /* This case occurs when a simultaneous close
4224 * happens, we must ack the received FIN and
4225 * enter the CLOSING state.
4228 tcp_set_state(sk, TCP_CLOSING);
4231 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4233 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4236 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4237 * cases we should never reach this piece of code.
4239 pr_err("%s: Impossible, sk->sk_state=%d\n",
4240 __func__, sk->sk_state);
4244 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4245 * Probably, we should reset in this case. For now drop them.
4247 skb_rbtree_purge(&tp->out_of_order_queue);
4248 if (tcp_is_sack(tp))
4249 tcp_sack_reset(&tp->rx_opt);
4252 if (!sock_flag(sk, SOCK_DEAD)) {
4253 sk->sk_state_change(sk);
4255 /* Do not send POLL_HUP for half duplex close. */
4256 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4257 sk->sk_state == TCP_CLOSE)
4258 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4260 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4264 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4267 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4268 if (before(seq, sp->start_seq))
4269 sp->start_seq = seq;
4270 if (after(end_seq, sp->end_seq))
4271 sp->end_seq = end_seq;
4277 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4279 struct tcp_sock *tp = tcp_sk(sk);
4281 if (tcp_is_sack(tp) && READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_dsack)) {
4284 if (before(seq, tp->rcv_nxt))
4285 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4287 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4289 NET_INC_STATS(sock_net(sk), mib_idx);
4291 tp->rx_opt.dsack = 1;
4292 tp->duplicate_sack[0].start_seq = seq;
4293 tp->duplicate_sack[0].end_seq = end_seq;
4297 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4299 struct tcp_sock *tp = tcp_sk(sk);
4301 if (!tp->rx_opt.dsack)
4302 tcp_dsack_set(sk, seq, end_seq);
4304 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4307 static void tcp_rcv_spurious_retrans(struct sock *sk, const struct sk_buff *skb)
4309 /* When the ACK path fails or drops most ACKs, the sender would
4310 * timeout and spuriously retransmit the same segment repeatedly.
4311 * The receiver remembers and reflects via DSACKs. Leverage the
4312 * DSACK state and change the txhash to re-route speculatively.
4314 if (TCP_SKB_CB(skb)->seq == tcp_sk(sk)->duplicate_sack[0].start_seq)
4315 sk_rethink_txhash(sk);
4318 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4320 struct tcp_sock *tp = tcp_sk(sk);
4322 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4323 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4324 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4325 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
4327 if (tcp_is_sack(tp) && READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_dsack)) {
4328 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4330 tcp_rcv_spurious_retrans(sk, skb);
4331 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4332 end_seq = tp->rcv_nxt;
4333 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4340 /* These routines update the SACK block as out-of-order packets arrive or
4341 * in-order packets close up the sequence space.
4343 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4346 struct tcp_sack_block *sp = &tp->selective_acks[0];
4347 struct tcp_sack_block *swalk = sp + 1;
4349 /* See if the recent change to the first SACK eats into
4350 * or hits the sequence space of other SACK blocks, if so coalesce.
4352 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4353 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4356 /* Zap SWALK, by moving every further SACK up by one slot.
4357 * Decrease num_sacks.
4359 tp->rx_opt.num_sacks--;
4360 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4364 this_sack++, swalk++;
4368 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4370 struct tcp_sock *tp = tcp_sk(sk);
4371 struct tcp_sack_block *sp = &tp->selective_acks[0];
4372 int cur_sacks = tp->rx_opt.num_sacks;
4378 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4379 if (tcp_sack_extend(sp, seq, end_seq)) {
4380 /* Rotate this_sack to the first one. */
4381 for (; this_sack > 0; this_sack--, sp--)
4382 swap(*sp, *(sp - 1));
4384 tcp_sack_maybe_coalesce(tp);
4389 /* Could not find an adjacent existing SACK, build a new one,
4390 * put it at the front, and shift everyone else down. We
4391 * always know there is at least one SACK present already here.
4393 * If the sack array is full, forget about the last one.
4395 if (this_sack >= TCP_NUM_SACKS) {
4396 if (tp->compressed_ack > TCP_FASTRETRANS_THRESH)
4399 tp->rx_opt.num_sacks--;
4402 for (; this_sack > 0; this_sack--, sp--)
4406 /* Build the new head SACK, and we're done. */
4407 sp->start_seq = seq;
4408 sp->end_seq = end_seq;
4409 tp->rx_opt.num_sacks++;
4412 /* RCV.NXT advances, some SACKs should be eaten. */
4414 static void tcp_sack_remove(struct tcp_sock *tp)
4416 struct tcp_sack_block *sp = &tp->selective_acks[0];
4417 int num_sacks = tp->rx_opt.num_sacks;
4420 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4421 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4422 tp->rx_opt.num_sacks = 0;
4426 for (this_sack = 0; this_sack < num_sacks;) {
4427 /* Check if the start of the sack is covered by RCV.NXT. */
4428 if (!before(tp->rcv_nxt, sp->start_seq)) {
4431 /* RCV.NXT must cover all the block! */
4432 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4434 /* Zap this SACK, by moving forward any other SACKS. */
4435 for (i = this_sack+1; i < num_sacks; i++)
4436 tp->selective_acks[i-1] = tp->selective_acks[i];
4443 tp->rx_opt.num_sacks = num_sacks;
4447 * tcp_try_coalesce - try to merge skb to prior one
4449 * @dest: destination queue
4451 * @from: buffer to add in queue
4452 * @fragstolen: pointer to boolean
4454 * Before queueing skb @from after @to, try to merge them
4455 * to reduce overall memory use and queue lengths, if cost is small.
4456 * Packets in ofo or receive queues can stay a long time.
4457 * Better try to coalesce them right now to avoid future collapses.
4458 * Returns true if caller should free @from instead of queueing it
4460 static bool tcp_try_coalesce(struct sock *sk,
4462 struct sk_buff *from,
4467 *fragstolen = false;
4469 /* Its possible this segment overlaps with prior segment in queue */
4470 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4473 #ifdef CONFIG_TLS_DEVICE
4474 if (from->decrypted != to->decrypted)
4478 if (!skb_try_coalesce(to, from, fragstolen, &delta))
4481 atomic_add(delta, &sk->sk_rmem_alloc);
4482 sk_mem_charge(sk, delta);
4483 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4484 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4485 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4486 TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
4488 if (TCP_SKB_CB(from)->has_rxtstamp) {
4489 TCP_SKB_CB(to)->has_rxtstamp = true;
4490 to->tstamp = from->tstamp;
4491 skb_hwtstamps(to)->hwtstamp = skb_hwtstamps(from)->hwtstamp;
4497 static bool tcp_ooo_try_coalesce(struct sock *sk,
4499 struct sk_buff *from,
4502 bool res = tcp_try_coalesce(sk, to, from, fragstolen);
4504 /* In case tcp_drop() is called later, update to->gso_segs */
4506 u32 gso_segs = max_t(u16, 1, skb_shinfo(to)->gso_segs) +
4507 max_t(u16, 1, skb_shinfo(from)->gso_segs);
4509 skb_shinfo(to)->gso_segs = min_t(u32, gso_segs, 0xFFFF);
4514 static void tcp_drop(struct sock *sk, struct sk_buff *skb)
4516 sk_drops_add(sk, skb);
4520 /* This one checks to see if we can put data from the
4521 * out_of_order queue into the receive_queue.
4523 static void tcp_ofo_queue(struct sock *sk)
4525 struct tcp_sock *tp = tcp_sk(sk);
4526 __u32 dsack_high = tp->rcv_nxt;
4527 bool fin, fragstolen, eaten;
4528 struct sk_buff *skb, *tail;
4531 p = rb_first(&tp->out_of_order_queue);
4534 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4537 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4538 __u32 dsack = dsack_high;
4539 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4540 dsack_high = TCP_SKB_CB(skb)->end_seq;
4541 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4544 rb_erase(&skb->rbnode, &tp->out_of_order_queue);
4546 if (unlikely(!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))) {
4551 tail = skb_peek_tail(&sk->sk_receive_queue);
4552 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen);
4553 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4554 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
4556 __skb_queue_tail(&sk->sk_receive_queue, skb);
4558 kfree_skb_partial(skb, fragstolen);
4560 if (unlikely(fin)) {
4562 /* tcp_fin() purges tp->out_of_order_queue,
4563 * so we must end this loop right now.
4570 static bool tcp_prune_ofo_queue(struct sock *sk);
4571 static int tcp_prune_queue(struct sock *sk);
4573 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4576 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4577 !sk_rmem_schedule(sk, skb, size)) {
4579 if (tcp_prune_queue(sk) < 0)
4582 while (!sk_rmem_schedule(sk, skb, size)) {
4583 if (!tcp_prune_ofo_queue(sk))
4590 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4592 struct tcp_sock *tp = tcp_sk(sk);
4593 struct rb_node **p, *parent;
4594 struct sk_buff *skb1;
4598 tcp_ecn_check_ce(sk, skb);
4600 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4601 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFODROP);
4602 sk->sk_data_ready(sk);
4607 /* Disable header prediction. */
4609 inet_csk_schedule_ack(sk);
4611 tp->rcv_ooopack += max_t(u16, 1, skb_shinfo(skb)->gso_segs);
4612 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4613 seq = TCP_SKB_CB(skb)->seq;
4614 end_seq = TCP_SKB_CB(skb)->end_seq;
4616 p = &tp->out_of_order_queue.rb_node;
4617 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4618 /* Initial out of order segment, build 1 SACK. */
4619 if (tcp_is_sack(tp)) {
4620 tp->rx_opt.num_sacks = 1;
4621 tp->selective_acks[0].start_seq = seq;
4622 tp->selective_acks[0].end_seq = end_seq;
4624 rb_link_node(&skb->rbnode, NULL, p);
4625 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4626 tp->ooo_last_skb = skb;
4630 /* In the typical case, we are adding an skb to the end of the list.
4631 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
4633 if (tcp_ooo_try_coalesce(sk, tp->ooo_last_skb,
4634 skb, &fragstolen)) {
4636 /* For non sack flows, do not grow window to force DUPACK
4637 * and trigger fast retransmit.
4639 if (tcp_is_sack(tp))
4640 tcp_grow_window(sk, skb);
4641 kfree_skb_partial(skb, fragstolen);
4645 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
4646 if (!before(seq, TCP_SKB_CB(tp->ooo_last_skb)->end_seq)) {
4647 parent = &tp->ooo_last_skb->rbnode;
4648 p = &parent->rb_right;
4652 /* Find place to insert this segment. Handle overlaps on the way. */
4656 skb1 = rb_to_skb(parent);
4657 if (before(seq, TCP_SKB_CB(skb1)->seq)) {
4658 p = &parent->rb_left;
4661 if (before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4662 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4663 /* All the bits are present. Drop. */
4664 NET_INC_STATS(sock_net(sk),
4665 LINUX_MIB_TCPOFOMERGE);
4668 tcp_dsack_set(sk, seq, end_seq);
4671 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4672 /* Partial overlap. */
4673 tcp_dsack_set(sk, seq, TCP_SKB_CB(skb1)->end_seq);
4675 /* skb's seq == skb1's seq and skb covers skb1.
4676 * Replace skb1 with skb.
4678 rb_replace_node(&skb1->rbnode, &skb->rbnode,
4679 &tp->out_of_order_queue);
4680 tcp_dsack_extend(sk,
4681 TCP_SKB_CB(skb1)->seq,
4682 TCP_SKB_CB(skb1)->end_seq);
4683 NET_INC_STATS(sock_net(sk),
4684 LINUX_MIB_TCPOFOMERGE);
4688 } else if (tcp_ooo_try_coalesce(sk, skb1,
4689 skb, &fragstolen)) {
4692 p = &parent->rb_right;
4695 /* Insert segment into RB tree. */
4696 rb_link_node(&skb->rbnode, parent, p);
4697 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4700 /* Remove other segments covered by skb. */
4701 while ((skb1 = skb_rb_next(skb)) != NULL) {
4702 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4704 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4705 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4709 rb_erase(&skb1->rbnode, &tp->out_of_order_queue);
4710 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4711 TCP_SKB_CB(skb1)->end_seq);
4712 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4715 /* If there is no skb after us, we are the last_skb ! */
4717 tp->ooo_last_skb = skb;
4720 if (tcp_is_sack(tp))
4721 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4724 /* For non sack flows, do not grow window to force DUPACK
4725 * and trigger fast retransmit.
4727 if (tcp_is_sack(tp))
4728 tcp_grow_window(sk, skb);
4730 skb_set_owner_r(skb, sk);
4734 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb,
4738 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4741 tcp_try_coalesce(sk, tail,
4742 skb, fragstolen)) ? 1 : 0;
4743 tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq);
4745 __skb_queue_tail(&sk->sk_receive_queue, skb);
4746 skb_set_owner_r(skb, sk);
4751 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4753 struct sk_buff *skb;
4761 if (size > PAGE_SIZE) {
4762 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS);
4764 data_len = npages << PAGE_SHIFT;
4765 size = data_len + (size & ~PAGE_MASK);
4767 skb = alloc_skb_with_frags(size - data_len, data_len,
4768 PAGE_ALLOC_COSTLY_ORDER,
4769 &err, sk->sk_allocation);
4773 skb_put(skb, size - data_len);
4774 skb->data_len = data_len;
4777 if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) {
4778 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP);
4782 err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size);
4786 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4787 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4788 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4790 if (tcp_queue_rcv(sk, skb, &fragstolen)) {
4791 WARN_ON_ONCE(fragstolen); /* should not happen */
4803 void tcp_data_ready(struct sock *sk)
4805 const struct tcp_sock *tp = tcp_sk(sk);
4806 int avail = tp->rcv_nxt - tp->copied_seq;
4808 if (avail < sk->sk_rcvlowat && !tcp_rmem_pressure(sk) &&
4809 !sock_flag(sk, SOCK_DONE) &&
4810 tcp_receive_window(tp) > inet_csk(sk)->icsk_ack.rcv_mss)
4813 sk->sk_data_ready(sk);
4816 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4818 struct tcp_sock *tp = tcp_sk(sk);
4822 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
4827 __skb_pull(skb, tcp_hdr(skb)->doff * 4);
4829 tp->rx_opt.dsack = 0;
4831 /* Queue data for delivery to the user.
4832 * Packets in sequence go to the receive queue.
4833 * Out of sequence packets to the out_of_order_queue.
4835 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4836 if (tcp_receive_window(tp) == 0) {
4837 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP);
4841 /* Ok. In sequence. In window. */
4843 if (skb_queue_len(&sk->sk_receive_queue) == 0)
4844 sk_forced_mem_schedule(sk, skb->truesize);
4845 else if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) {
4846 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP);
4847 sk->sk_data_ready(sk);
4851 eaten = tcp_queue_rcv(sk, skb, &fragstolen);
4853 tcp_event_data_recv(sk, skb);
4854 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
4857 if (!RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4860 /* RFC5681. 4.2. SHOULD send immediate ACK, when
4861 * gap in queue is filled.
4863 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
4864 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW;
4867 if (tp->rx_opt.num_sacks)
4868 tcp_sack_remove(tp);
4870 tcp_fast_path_check(sk);
4873 kfree_skb_partial(skb, fragstolen);
4874 if (!sock_flag(sk, SOCK_DEAD))
4879 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4880 tcp_rcv_spurious_retrans(sk, skb);
4881 /* A retransmit, 2nd most common case. Force an immediate ack. */
4882 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4883 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4886 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
4887 inet_csk_schedule_ack(sk);
4893 /* Out of window. F.e. zero window probe. */
4894 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4897 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4898 /* Partial packet, seq < rcv_next < end_seq */
4899 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4901 /* If window is closed, drop tail of packet. But after
4902 * remembering D-SACK for its head made in previous line.
4904 if (!tcp_receive_window(tp)) {
4905 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP);
4911 tcp_data_queue_ofo(sk, skb);
4914 static struct sk_buff *tcp_skb_next(struct sk_buff *skb, struct sk_buff_head *list)
4917 return !skb_queue_is_last(list, skb) ? skb->next : NULL;
4919 return skb_rb_next(skb);
4922 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4923 struct sk_buff_head *list,
4924 struct rb_root *root)
4926 struct sk_buff *next = tcp_skb_next(skb, list);
4929 __skb_unlink(skb, list);
4931 rb_erase(&skb->rbnode, root);
4934 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4939 /* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */
4940 void tcp_rbtree_insert(struct rb_root *root, struct sk_buff *skb)
4942 struct rb_node **p = &root->rb_node;
4943 struct rb_node *parent = NULL;
4944 struct sk_buff *skb1;
4948 skb1 = rb_to_skb(parent);
4949 if (before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb1)->seq))
4950 p = &parent->rb_left;
4952 p = &parent->rb_right;
4954 rb_link_node(&skb->rbnode, parent, p);
4955 rb_insert_color(&skb->rbnode, root);
4958 /* Collapse contiguous sequence of skbs head..tail with
4959 * sequence numbers start..end.
4961 * If tail is NULL, this means until the end of the queue.
4963 * Segments with FIN/SYN are not collapsed (only because this
4967 tcp_collapse(struct sock *sk, struct sk_buff_head *list, struct rb_root *root,
4968 struct sk_buff *head, struct sk_buff *tail, u32 start, u32 end)
4970 struct sk_buff *skb = head, *n;
4971 struct sk_buff_head tmp;
4974 /* First, check that queue is collapsible and find
4975 * the point where collapsing can be useful.
4978 for (end_of_skbs = true; skb != NULL && skb != tail; skb = n) {
4979 n = tcp_skb_next(skb, list);
4981 /* No new bits? It is possible on ofo queue. */
4982 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4983 skb = tcp_collapse_one(sk, skb, list, root);
4989 /* The first skb to collapse is:
4991 * - bloated or contains data before "start" or
4992 * overlaps to the next one.
4994 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
4995 (tcp_win_from_space(sk, skb->truesize) > skb->len ||
4996 before(TCP_SKB_CB(skb)->seq, start))) {
4997 end_of_skbs = false;
5001 if (n && n != tail &&
5002 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(n)->seq) {
5003 end_of_skbs = false;
5007 /* Decided to skip this, advance start seq. */
5008 start = TCP_SKB_CB(skb)->end_seq;
5011 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
5014 __skb_queue_head_init(&tmp);
5016 while (before(start, end)) {
5017 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
5018 struct sk_buff *nskb;
5020 nskb = alloc_skb(copy, GFP_ATOMIC);
5024 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
5025 #ifdef CONFIG_TLS_DEVICE
5026 nskb->decrypted = skb->decrypted;
5028 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
5030 __skb_queue_before(list, skb, nskb);
5032 __skb_queue_tail(&tmp, nskb); /* defer rbtree insertion */
5033 skb_set_owner_r(nskb, sk);
5035 /* Copy data, releasing collapsed skbs. */
5037 int offset = start - TCP_SKB_CB(skb)->seq;
5038 int size = TCP_SKB_CB(skb)->end_seq - start;
5042 size = min(copy, size);
5043 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
5045 TCP_SKB_CB(nskb)->end_seq += size;
5049 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
5050 skb = tcp_collapse_one(sk, skb, list, root);
5053 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
5055 #ifdef CONFIG_TLS_DEVICE
5056 if (skb->decrypted != nskb->decrypted)
5063 skb_queue_walk_safe(&tmp, skb, n)
5064 tcp_rbtree_insert(root, skb);
5067 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
5068 * and tcp_collapse() them until all the queue is collapsed.
5070 static void tcp_collapse_ofo_queue(struct sock *sk)
5072 struct tcp_sock *tp = tcp_sk(sk);
5073 u32 range_truesize, sum_tiny = 0;
5074 struct sk_buff *skb, *head;
5077 skb = skb_rb_first(&tp->out_of_order_queue);
5080 tp->ooo_last_skb = skb_rb_last(&tp->out_of_order_queue);
5083 start = TCP_SKB_CB(skb)->seq;
5084 end = TCP_SKB_CB(skb)->end_seq;
5085 range_truesize = skb->truesize;
5087 for (head = skb;;) {
5088 skb = skb_rb_next(skb);
5090 /* Range is terminated when we see a gap or when
5091 * we are at the queue end.
5094 after(TCP_SKB_CB(skb)->seq, end) ||
5095 before(TCP_SKB_CB(skb)->end_seq, start)) {
5096 /* Do not attempt collapsing tiny skbs */
5097 if (range_truesize != head->truesize ||
5098 end - start >= SKB_WITH_OVERHEAD(SK_MEM_QUANTUM)) {
5099 tcp_collapse(sk, NULL, &tp->out_of_order_queue,
5100 head, skb, start, end);
5102 sum_tiny += range_truesize;
5103 if (sum_tiny > sk->sk_rcvbuf >> 3)
5109 range_truesize += skb->truesize;
5110 if (unlikely(before(TCP_SKB_CB(skb)->seq, start)))
5111 start = TCP_SKB_CB(skb)->seq;
5112 if (after(TCP_SKB_CB(skb)->end_seq, end))
5113 end = TCP_SKB_CB(skb)->end_seq;
5118 * Clean the out-of-order queue to make room.
5119 * We drop high sequences packets to :
5120 * 1) Let a chance for holes to be filled.
5121 * 2) not add too big latencies if thousands of packets sit there.
5122 * (But if application shrinks SO_RCVBUF, we could still end up
5123 * freeing whole queue here)
5124 * 3) Drop at least 12.5 % of sk_rcvbuf to avoid malicious attacks.
5126 * Return true if queue has shrunk.
5128 static bool tcp_prune_ofo_queue(struct sock *sk)
5130 struct tcp_sock *tp = tcp_sk(sk);
5131 struct rb_node *node, *prev;
5134 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
5137 NET_INC_STATS(sock_net(sk), LINUX_MIB_OFOPRUNED);
5138 goal = sk->sk_rcvbuf >> 3;
5139 node = &tp->ooo_last_skb->rbnode;
5141 prev = rb_prev(node);
5142 rb_erase(node, &tp->out_of_order_queue);
5143 goal -= rb_to_skb(node)->truesize;
5144 tcp_drop(sk, rb_to_skb(node));
5145 if (!prev || goal <= 0) {
5147 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf &&
5148 !tcp_under_memory_pressure(sk))
5150 goal = sk->sk_rcvbuf >> 3;
5154 tp->ooo_last_skb = rb_to_skb(prev);
5156 /* Reset SACK state. A conforming SACK implementation will
5157 * do the same at a timeout based retransmit. When a connection
5158 * is in a sad state like this, we care only about integrity
5159 * of the connection not performance.
5161 if (tp->rx_opt.sack_ok)
5162 tcp_sack_reset(&tp->rx_opt);
5166 /* Reduce allocated memory if we can, trying to get
5167 * the socket within its memory limits again.
5169 * Return less than zero if we should start dropping frames
5170 * until the socket owning process reads some of the data
5171 * to stabilize the situation.
5173 static int tcp_prune_queue(struct sock *sk)
5175 struct tcp_sock *tp = tcp_sk(sk);
5177 NET_INC_STATS(sock_net(sk), LINUX_MIB_PRUNECALLED);
5179 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
5180 tcp_clamp_window(sk);
5181 else if (tcp_under_memory_pressure(sk))
5182 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
5184 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5187 tcp_collapse_ofo_queue(sk);
5188 if (!skb_queue_empty(&sk->sk_receive_queue))
5189 tcp_collapse(sk, &sk->sk_receive_queue, NULL,
5190 skb_peek(&sk->sk_receive_queue),
5192 tp->copied_seq, tp->rcv_nxt);
5195 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5198 /* Collapsing did not help, destructive actions follow.
5199 * This must not ever occur. */
5201 tcp_prune_ofo_queue(sk);
5203 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5206 /* If we are really being abused, tell the caller to silently
5207 * drop receive data on the floor. It will get retransmitted
5208 * and hopefully then we'll have sufficient space.
5210 NET_INC_STATS(sock_net(sk), LINUX_MIB_RCVPRUNED);
5212 /* Massive buffer overcommit. */
5217 static bool tcp_should_expand_sndbuf(const struct sock *sk)
5219 const struct tcp_sock *tp = tcp_sk(sk);
5221 /* If the user specified a specific send buffer setting, do
5224 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
5227 /* If we are under global TCP memory pressure, do not expand. */
5228 if (tcp_under_memory_pressure(sk))
5231 /* If we are under soft global TCP memory pressure, do not expand. */
5232 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
5235 /* If we filled the congestion window, do not expand. */
5236 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
5242 /* When incoming ACK allowed to free some skb from write_queue,
5243 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
5244 * on the exit from tcp input handler.
5246 * PROBLEM: sndbuf expansion does not work well with largesend.
5248 static void tcp_new_space(struct sock *sk)
5250 struct tcp_sock *tp = tcp_sk(sk);
5252 if (tcp_should_expand_sndbuf(sk)) {
5253 tcp_sndbuf_expand(sk);
5254 tp->snd_cwnd_stamp = tcp_jiffies32;
5257 sk->sk_write_space(sk);
5260 /* Caller made space either from:
5261 * 1) Freeing skbs in rtx queues (after tp->snd_una has advanced)
5262 * 2) Sent skbs from output queue (and thus advancing tp->snd_nxt)
5264 * We might be able to generate EPOLLOUT to the application if:
5265 * 1) Space consumed in output/rtx queues is below sk->sk_sndbuf/2
5266 * 2) notsent amount (tp->write_seq - tp->snd_nxt) became
5267 * small enough that tcp_stream_memory_free() decides it
5268 * is time to generate EPOLLOUT.
5270 void tcp_check_space(struct sock *sk)
5272 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
5273 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
5274 /* pairs with tcp_poll() */
5276 if (sk->sk_socket &&
5277 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
5279 if (!test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
5280 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
5285 static inline void tcp_data_snd_check(struct sock *sk)
5287 tcp_push_pending_frames(sk);
5288 tcp_check_space(sk);
5292 * Check if sending an ack is needed.
5294 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
5296 struct tcp_sock *tp = tcp_sk(sk);
5297 unsigned long rtt, delay;
5299 /* More than one full frame received... */
5300 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
5301 /* ... and right edge of window advances far enough.
5302 * (tcp_recvmsg() will send ACK otherwise).
5303 * If application uses SO_RCVLOWAT, we want send ack now if
5304 * we have not received enough bytes to satisfy the condition.
5306 (tp->rcv_nxt - tp->copied_seq < sk->sk_rcvlowat ||
5307 __tcp_select_window(sk) >= tp->rcv_wnd)) ||
5308 /* We ACK each frame or... */
5309 tcp_in_quickack_mode(sk) ||
5310 /* Protocol state mandates a one-time immediate ACK */
5311 inet_csk(sk)->icsk_ack.pending & ICSK_ACK_NOW) {
5317 if (!ofo_possible || RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
5318 tcp_send_delayed_ack(sk);
5322 if (!tcp_is_sack(tp) ||
5323 tp->compressed_ack >= READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_nr))
5326 if (tp->compressed_ack_rcv_nxt != tp->rcv_nxt) {
5327 tp->compressed_ack_rcv_nxt = tp->rcv_nxt;
5328 if (tp->compressed_ack > TCP_FASTRETRANS_THRESH)
5329 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED,
5330 tp->compressed_ack - TCP_FASTRETRANS_THRESH);
5331 tp->compressed_ack = 0;
5334 if (++tp->compressed_ack <= TCP_FASTRETRANS_THRESH)
5337 if (hrtimer_is_queued(&tp->compressed_ack_timer))
5340 /* compress ack timer : 5 % of rtt, but no more than tcp_comp_sack_delay_ns */
5342 rtt = tp->rcv_rtt_est.rtt_us;
5343 if (tp->srtt_us && tp->srtt_us < rtt)
5346 delay = min_t(unsigned long,
5347 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_delay_ns),
5348 rtt * (NSEC_PER_USEC >> 3)/20);
5350 hrtimer_start(&tp->compressed_ack_timer, ns_to_ktime(delay),
5351 HRTIMER_MODE_REL_PINNED_SOFT);
5354 static inline void tcp_ack_snd_check(struct sock *sk)
5356 if (!inet_csk_ack_scheduled(sk)) {
5357 /* We sent a data segment already. */
5360 __tcp_ack_snd_check(sk, 1);
5364 * This routine is only called when we have urgent data
5365 * signaled. Its the 'slow' part of tcp_urg. It could be
5366 * moved inline now as tcp_urg is only called from one
5367 * place. We handle URGent data wrong. We have to - as
5368 * BSD still doesn't use the correction from RFC961.
5369 * For 1003.1g we should support a new option TCP_STDURG to permit
5370 * either form (or just set the sysctl tcp_stdurg).
5373 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
5375 struct tcp_sock *tp = tcp_sk(sk);
5376 u32 ptr = ntohs(th->urg_ptr);
5378 if (ptr && !READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_stdurg))
5380 ptr += ntohl(th->seq);
5382 /* Ignore urgent data that we've already seen and read. */
5383 if (after(tp->copied_seq, ptr))
5386 /* Do not replay urg ptr.
5388 * NOTE: interesting situation not covered by specs.
5389 * Misbehaving sender may send urg ptr, pointing to segment,
5390 * which we already have in ofo queue. We are not able to fetch
5391 * such data and will stay in TCP_URG_NOTYET until will be eaten
5392 * by recvmsg(). Seems, we are not obliged to handle such wicked
5393 * situations. But it is worth to think about possibility of some
5394 * DoSes using some hypothetical application level deadlock.
5396 if (before(ptr, tp->rcv_nxt))
5399 /* Do we already have a newer (or duplicate) urgent pointer? */
5400 if (tp->urg_data && !after(ptr, tp->urg_seq))
5403 /* Tell the world about our new urgent pointer. */
5406 /* We may be adding urgent data when the last byte read was
5407 * urgent. To do this requires some care. We cannot just ignore
5408 * tp->copied_seq since we would read the last urgent byte again
5409 * as data, nor can we alter copied_seq until this data arrives
5410 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5412 * NOTE. Double Dutch. Rendering to plain English: author of comment
5413 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5414 * and expect that both A and B disappear from stream. This is _wrong_.
5415 * Though this happens in BSD with high probability, this is occasional.
5416 * Any application relying on this is buggy. Note also, that fix "works"
5417 * only in this artificial test. Insert some normal data between A and B and we will
5418 * decline of BSD again. Verdict: it is better to remove to trap
5421 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5422 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5423 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5425 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5426 __skb_unlink(skb, &sk->sk_receive_queue);
5431 tp->urg_data = TCP_URG_NOTYET;
5432 WRITE_ONCE(tp->urg_seq, ptr);
5434 /* Disable header prediction. */
5438 /* This is the 'fast' part of urgent handling. */
5439 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
5441 struct tcp_sock *tp = tcp_sk(sk);
5443 /* Check if we get a new urgent pointer - normally not. */
5445 tcp_check_urg(sk, th);
5447 /* Do we wait for any urgent data? - normally not... */
5448 if (tp->urg_data == TCP_URG_NOTYET) {
5449 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5452 /* Is the urgent pointer pointing into this packet? */
5453 if (ptr < skb->len) {
5455 if (skb_copy_bits(skb, ptr, &tmp, 1))
5457 tp->urg_data = TCP_URG_VALID | tmp;
5458 if (!sock_flag(sk, SOCK_DEAD))
5459 sk->sk_data_ready(sk);
5464 /* Accept RST for rcv_nxt - 1 after a FIN.
5465 * When tcp connections are abruptly terminated from Mac OSX (via ^C), a
5466 * FIN is sent followed by a RST packet. The RST is sent with the same
5467 * sequence number as the FIN, and thus according to RFC 5961 a challenge
5468 * ACK should be sent. However, Mac OSX rate limits replies to challenge
5469 * ACKs on the closed socket. In addition middleboxes can drop either the
5470 * challenge ACK or a subsequent RST.
5472 static bool tcp_reset_check(const struct sock *sk, const struct sk_buff *skb)
5474 struct tcp_sock *tp = tcp_sk(sk);
5476 return unlikely(TCP_SKB_CB(skb)->seq == (tp->rcv_nxt - 1) &&
5477 (1 << sk->sk_state) & (TCPF_CLOSE_WAIT | TCPF_LAST_ACK |
5481 /* Does PAWS and seqno based validation of an incoming segment, flags will
5482 * play significant role here.
5484 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5485 const struct tcphdr *th, int syn_inerr)
5487 struct tcp_sock *tp = tcp_sk(sk);
5488 bool rst_seq_match = false;
5490 /* RFC1323: H1. Apply PAWS check first. */
5491 if (tcp_fast_parse_options(sock_net(sk), skb, th, tp) &&
5492 tp->rx_opt.saw_tstamp &&
5493 tcp_paws_discard(sk, skb)) {
5495 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5496 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5497 LINUX_MIB_TCPACKSKIPPEDPAWS,
5498 &tp->last_oow_ack_time))
5499 tcp_send_dupack(sk, skb);
5502 /* Reset is accepted even if it did not pass PAWS. */
5505 /* Step 1: check sequence number */
5506 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5507 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5508 * (RST) segments are validated by checking their SEQ-fields."
5509 * And page 69: "If an incoming segment is not acceptable,
5510 * an acknowledgment should be sent in reply (unless the RST
5511 * bit is set, if so drop the segment and return)".
5516 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5517 LINUX_MIB_TCPACKSKIPPEDSEQ,
5518 &tp->last_oow_ack_time))
5519 tcp_send_dupack(sk, skb);
5520 } else if (tcp_reset_check(sk, skb)) {
5526 /* Step 2: check RST bit */
5528 /* RFC 5961 3.2 (extend to match against (RCV.NXT - 1) after a
5529 * FIN and SACK too if available):
5530 * If seq num matches RCV.NXT or (RCV.NXT - 1) after a FIN, or
5531 * the right-most SACK block,
5533 * RESET the connection
5535 * Send a challenge ACK
5537 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt ||
5538 tcp_reset_check(sk, skb)) {
5539 rst_seq_match = true;
5540 } else if (tcp_is_sack(tp) && tp->rx_opt.num_sacks > 0) {
5541 struct tcp_sack_block *sp = &tp->selective_acks[0];
5542 int max_sack = sp[0].end_seq;
5545 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;
5547 max_sack = after(sp[this_sack].end_seq,
5549 sp[this_sack].end_seq : max_sack;
5552 if (TCP_SKB_CB(skb)->seq == max_sack)
5553 rst_seq_match = true;
5559 /* Disable TFO if RST is out-of-order
5560 * and no data has been received
5561 * for current active TFO socket
5563 if (tp->syn_fastopen && !tp->data_segs_in &&
5564 sk->sk_state == TCP_ESTABLISHED)
5565 tcp_fastopen_active_disable(sk);
5566 tcp_send_challenge_ack(sk, skb);
5571 /* step 3: check security and precedence [ignored] */
5573 /* step 4: Check for a SYN
5574 * RFC 5961 4.2 : Send a challenge ack
5579 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5580 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5581 tcp_send_challenge_ack(sk, skb);
5593 * TCP receive function for the ESTABLISHED state.
5595 * It is split into a fast path and a slow path. The fast path is
5597 * - A zero window was announced from us - zero window probing
5598 * is only handled properly in the slow path.
5599 * - Out of order segments arrived.
5600 * - Urgent data is expected.
5601 * - There is no buffer space left
5602 * - Unexpected TCP flags/window values/header lengths are received
5603 * (detected by checking the TCP header against pred_flags)
5604 * - Data is sent in both directions. Fast path only supports pure senders
5605 * or pure receivers (this means either the sequence number or the ack
5606 * value must stay constant)
5607 * - Unexpected TCP option.
5609 * When these conditions are not satisfied it drops into a standard
5610 * receive procedure patterned after RFC793 to handle all cases.
5611 * The first three cases are guaranteed by proper pred_flags setting,
5612 * the rest is checked inline. Fast processing is turned on in
5613 * tcp_data_queue when everything is OK.
5615 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb)
5617 const struct tcphdr *th = (const struct tcphdr *)skb->data;
5618 struct tcp_sock *tp = tcp_sk(sk);
5619 unsigned int len = skb->len;
5621 /* TCP congestion window tracking */
5622 trace_tcp_probe(sk, skb);
5624 tcp_mstamp_refresh(tp);
5625 if (unlikely(!rcu_access_pointer(sk->sk_rx_dst)))
5626 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
5628 * Header prediction.
5629 * The code loosely follows the one in the famous
5630 * "30 instruction TCP receive" Van Jacobson mail.
5632 * Van's trick is to deposit buffers into socket queue
5633 * on a device interrupt, to call tcp_recv function
5634 * on the receive process context and checksum and copy
5635 * the buffer to user space. smart...
5637 * Our current scheme is not silly either but we take the
5638 * extra cost of the net_bh soft interrupt processing...
5639 * We do checksum and copy also but from device to kernel.
5642 tp->rx_opt.saw_tstamp = 0;
5644 /* pred_flags is 0xS?10 << 16 + snd_wnd
5645 * if header_prediction is to be made
5646 * 'S' will always be tp->tcp_header_len >> 2
5647 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5648 * turn it off (when there are holes in the receive
5649 * space for instance)
5650 * PSH flag is ignored.
5653 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5654 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5655 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5656 int tcp_header_len = tp->tcp_header_len;
5658 /* Timestamp header prediction: tcp_header_len
5659 * is automatically equal to th->doff*4 due to pred_flags
5663 /* Check timestamp */
5664 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5665 /* No? Slow path! */
5666 if (!tcp_parse_aligned_timestamp(tp, th))
5669 /* If PAWS failed, check it more carefully in slow path */
5670 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5673 /* DO NOT update ts_recent here, if checksum fails
5674 * and timestamp was corrupted part, it will result
5675 * in a hung connection since we will drop all
5676 * future packets due to the PAWS test.
5680 if (len <= tcp_header_len) {
5681 /* Bulk data transfer: sender */
5682 if (len == tcp_header_len) {
5683 /* Predicted packet is in window by definition.
5684 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5685 * Hence, check seq<=rcv_wup reduces to:
5687 if (tcp_header_len ==
5688 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5689 tp->rcv_nxt == tp->rcv_wup)
5690 tcp_store_ts_recent(tp);
5692 /* We know that such packets are checksummed
5695 tcp_ack(sk, skb, 0);
5697 tcp_data_snd_check(sk);
5698 /* When receiving pure ack in fast path, update
5699 * last ts ecr directly instead of calling
5700 * tcp_rcv_rtt_measure_ts()
5702 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr;
5704 } else { /* Header too small */
5705 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5710 bool fragstolen = false;
5712 if (tcp_checksum_complete(skb))
5715 if ((int)skb->truesize > sk->sk_forward_alloc)
5718 /* Predicted packet is in window by definition.
5719 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5720 * Hence, check seq<=rcv_wup reduces to:
5722 if (tcp_header_len ==
5723 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5724 tp->rcv_nxt == tp->rcv_wup)
5725 tcp_store_ts_recent(tp);
5727 tcp_rcv_rtt_measure_ts(sk, skb);
5729 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPHITS);
5731 /* Bulk data transfer: receiver */
5732 __skb_pull(skb, tcp_header_len);
5733 eaten = tcp_queue_rcv(sk, skb, &fragstolen);
5735 tcp_event_data_recv(sk, skb);
5737 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5738 /* Well, only one small jumplet in fast path... */
5739 tcp_ack(sk, skb, FLAG_DATA);
5740 tcp_data_snd_check(sk);
5741 if (!inet_csk_ack_scheduled(sk))
5744 tcp_update_wl(tp, TCP_SKB_CB(skb)->seq);
5747 __tcp_ack_snd_check(sk, 0);
5750 kfree_skb_partial(skb, fragstolen);
5757 if (len < (th->doff << 2) || tcp_checksum_complete(skb))
5760 if (!th->ack && !th->rst && !th->syn)
5764 * Standard slow path.
5767 if (!tcp_validate_incoming(sk, skb, th, 1))
5771 if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0)
5774 tcp_rcv_rtt_measure_ts(sk, skb);
5776 /* Process urgent data. */
5777 tcp_urg(sk, skb, th);
5779 /* step 7: process the segment text */
5780 tcp_data_queue(sk, skb);
5782 tcp_data_snd_check(sk);
5783 tcp_ack_snd_check(sk);
5787 TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS);
5788 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5793 EXPORT_SYMBOL(tcp_rcv_established);
5795 void tcp_init_transfer(struct sock *sk, int bpf_op)
5797 struct inet_connection_sock *icsk = inet_csk(sk);
5798 struct tcp_sock *tp = tcp_sk(sk);
5801 icsk->icsk_af_ops->rebuild_header(sk);
5802 tcp_init_metrics(sk);
5804 /* Initialize the congestion window to start the transfer.
5805 * Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
5806 * retransmitted. In light of RFC6298 more aggressive 1sec
5807 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
5808 * retransmission has occurred.
5810 if (tp->total_retrans > 1 && tp->undo_marker)
5813 tp->snd_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk));
5814 tp->snd_cwnd_stamp = tcp_jiffies32;
5816 tcp_call_bpf(sk, bpf_op, 0, NULL);
5817 tcp_init_congestion_control(sk);
5818 tcp_init_buffer_space(sk);
5821 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5823 struct tcp_sock *tp = tcp_sk(sk);
5824 struct inet_connection_sock *icsk = inet_csk(sk);
5826 tcp_set_state(sk, TCP_ESTABLISHED);
5827 icsk->icsk_ack.lrcvtime = tcp_jiffies32;
5830 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
5831 security_inet_conn_established(sk, skb);
5832 sk_mark_napi_id(sk, skb);
5835 tcp_init_transfer(sk, BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB);
5837 /* Prevent spurious tcp_cwnd_restart() on first data
5840 tp->lsndtime = tcp_jiffies32;
5842 if (sock_flag(sk, SOCK_KEEPOPEN))
5843 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5845 if (!tp->rx_opt.snd_wscale)
5846 __tcp_fast_path_on(tp, tp->snd_wnd);
5851 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
5852 struct tcp_fastopen_cookie *cookie)
5854 struct tcp_sock *tp = tcp_sk(sk);
5855 struct sk_buff *data = tp->syn_data ? tcp_rtx_queue_head(sk) : NULL;
5856 u16 mss = tp->rx_opt.mss_clamp, try_exp = 0;
5857 bool syn_drop = false;
5859 if (mss == tp->rx_opt.user_mss) {
5860 struct tcp_options_received opt;
5862 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5863 tcp_clear_options(&opt);
5864 opt.user_mss = opt.mss_clamp = 0;
5865 tcp_parse_options(sock_net(sk), synack, &opt, 0, NULL);
5866 mss = opt.mss_clamp;
5869 if (!tp->syn_fastopen) {
5870 /* Ignore an unsolicited cookie */
5872 } else if (tp->total_retrans) {
5873 /* SYN timed out and the SYN-ACK neither has a cookie nor
5874 * acknowledges data. Presumably the remote received only
5875 * the retransmitted (regular) SYNs: either the original
5876 * SYN-data or the corresponding SYN-ACK was dropped.
5878 syn_drop = (cookie->len < 0 && data);
5879 } else if (cookie->len < 0 && !tp->syn_data) {
5880 /* We requested a cookie but didn't get it. If we did not use
5881 * the (old) exp opt format then try so next time (try_exp=1).
5882 * Otherwise we go back to use the RFC7413 opt (try_exp=2).
5884 try_exp = tp->syn_fastopen_exp ? 2 : 1;
5887 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp);
5889 if (data) { /* Retransmit unacked data in SYN */
5890 skb_rbtree_walk_from(data) {
5891 if (__tcp_retransmit_skb(sk, data, 1))
5895 NET_INC_STATS(sock_net(sk),
5896 LINUX_MIB_TCPFASTOPENACTIVEFAIL);
5899 tp->syn_data_acked = tp->syn_data;
5900 if (tp->syn_data_acked) {
5901 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
5902 /* SYN-data is counted as two separate packets in tcp_ack() */
5903 if (tp->delivered > 1)
5907 tcp_fastopen_add_skb(sk, synack);
5912 static void smc_check_reset_syn(struct tcp_sock *tp)
5914 #if IS_ENABLED(CONFIG_SMC)
5915 if (static_branch_unlikely(&tcp_have_smc)) {
5916 if (tp->syn_smc && !tp->rx_opt.smc_ok)
5922 static void tcp_try_undo_spurious_syn(struct sock *sk)
5924 struct tcp_sock *tp = tcp_sk(sk);
5927 /* undo_marker is set when SYN or SYNACK times out. The timeout is
5928 * spurious if the ACK's timestamp option echo value matches the
5929 * original SYN timestamp.
5931 syn_stamp = tp->retrans_stamp;
5932 if (tp->undo_marker && syn_stamp && tp->rx_opt.saw_tstamp &&
5933 syn_stamp == tp->rx_opt.rcv_tsecr)
5934 tp->undo_marker = 0;
5937 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5938 const struct tcphdr *th)
5940 struct inet_connection_sock *icsk = inet_csk(sk);
5941 struct tcp_sock *tp = tcp_sk(sk);
5942 struct tcp_fastopen_cookie foc = { .len = -1 };
5943 int saved_clamp = tp->rx_opt.mss_clamp;
5946 tcp_parse_options(sock_net(sk), skb, &tp->rx_opt, 0, &foc);
5947 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
5948 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
5952 * "If the state is SYN-SENT then
5953 * first check the ACK bit
5954 * If the ACK bit is set
5955 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5956 * a reset (unless the RST bit is set, if so drop
5957 * the segment and return)"
5959 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
5960 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt))
5961 goto reset_and_undo;
5963 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5964 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5965 tcp_time_stamp(tp))) {
5966 NET_INC_STATS(sock_net(sk),
5967 LINUX_MIB_PAWSACTIVEREJECTED);
5968 goto reset_and_undo;
5971 /* Now ACK is acceptable.
5973 * "If the RST bit is set
5974 * If the ACK was acceptable then signal the user "error:
5975 * connection reset", drop the segment, enter CLOSED state,
5976 * delete TCB, and return."
5985 * "fifth, if neither of the SYN or RST bits is set then
5986 * drop the segment and return."
5992 goto discard_and_undo;
5995 * "If the SYN bit is on ...
5996 * are acceptable then ...
5997 * (our SYN has been ACKed), change the connection
5998 * state to ESTABLISHED..."
6001 tcp_ecn_rcv_synack(tp, th);
6003 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
6004 tcp_try_undo_spurious_syn(sk);
6005 tcp_ack(sk, skb, FLAG_SLOWPATH);
6007 /* Ok.. it's good. Set up sequence numbers and
6008 * move to established.
6010 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1);
6011 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
6013 /* RFC1323: The window in SYN & SYN/ACK segments is
6016 tp->snd_wnd = ntohs(th->window);
6018 if (!tp->rx_opt.wscale_ok) {
6019 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
6020 tp->window_clamp = min(tp->window_clamp, 65535U);
6023 if (tp->rx_opt.saw_tstamp) {
6024 tp->rx_opt.tstamp_ok = 1;
6025 tp->tcp_header_len =
6026 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
6027 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
6028 tcp_store_ts_recent(tp);
6030 tp->tcp_header_len = sizeof(struct tcphdr);
6033 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
6034 tcp_initialize_rcv_mss(sk);
6036 /* Remember, tcp_poll() does not lock socket!
6037 * Change state from SYN-SENT only after copied_seq
6038 * is initialized. */
6039 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6041 smc_check_reset_syn(tp);
6045 tcp_finish_connect(sk, skb);
6047 fastopen_fail = (tp->syn_fastopen || tp->syn_data) &&
6048 tcp_rcv_fastopen_synack(sk, skb, &foc);
6050 if (!sock_flag(sk, SOCK_DEAD)) {
6051 sk->sk_state_change(sk);
6052 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
6056 if (sk->sk_write_pending ||
6057 icsk->icsk_accept_queue.rskq_defer_accept ||
6058 inet_csk_in_pingpong_mode(sk)) {
6059 /* Save one ACK. Data will be ready after
6060 * several ticks, if write_pending is set.
6062 * It may be deleted, but with this feature tcpdumps
6063 * look so _wonderfully_ clever, that I was not able
6064 * to stand against the temptation 8) --ANK
6066 inet_csk_schedule_ack(sk);
6067 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
6068 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
6069 TCP_DELACK_MAX, TCP_RTO_MAX);
6080 /* No ACK in the segment */
6084 * "If the RST bit is set
6086 * Otherwise (no ACK) drop the segment and return."
6089 goto discard_and_undo;
6093 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
6094 tcp_paws_reject(&tp->rx_opt, 0))
6095 goto discard_and_undo;
6098 /* We see SYN without ACK. It is attempt of
6099 * simultaneous connect with crossed SYNs.
6100 * Particularly, it can be connect to self.
6102 tcp_set_state(sk, TCP_SYN_RECV);
6104 if (tp->rx_opt.saw_tstamp) {
6105 tp->rx_opt.tstamp_ok = 1;
6106 tcp_store_ts_recent(tp);
6107 tp->tcp_header_len =
6108 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
6110 tp->tcp_header_len = sizeof(struct tcphdr);
6113 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1);
6114 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6115 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
6117 /* RFC1323: The window in SYN & SYN/ACK segments is
6120 tp->snd_wnd = ntohs(th->window);
6121 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
6122 tp->max_window = tp->snd_wnd;
6124 tcp_ecn_rcv_syn(tp, th);
6127 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
6128 tcp_initialize_rcv_mss(sk);
6130 tcp_send_synack(sk);
6132 /* Note, we could accept data and URG from this segment.
6133 * There are no obstacles to make this (except that we must
6134 * either change tcp_recvmsg() to prevent it from returning data
6135 * before 3WHS completes per RFC793, or employ TCP Fast Open).
6137 * However, if we ignore data in ACKless segments sometimes,
6138 * we have no reasons to accept it sometimes.
6139 * Also, seems the code doing it in step6 of tcp_rcv_state_process
6140 * is not flawless. So, discard packet for sanity.
6141 * Uncomment this return to process the data.
6148 /* "fifth, if neither of the SYN or RST bits is set then
6149 * drop the segment and return."
6153 tcp_clear_options(&tp->rx_opt);
6154 tp->rx_opt.mss_clamp = saved_clamp;
6158 tcp_clear_options(&tp->rx_opt);
6159 tp->rx_opt.mss_clamp = saved_clamp;
6163 static void tcp_rcv_synrecv_state_fastopen(struct sock *sk)
6165 struct request_sock *req;
6167 /* If we are still handling the SYNACK RTO, see if timestamp ECR allows
6168 * undo. If peer SACKs triggered fast recovery, we can't undo here.
6170 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
6171 tcp_try_undo_loss(sk, false);
6173 /* Reset rtx states to prevent spurious retransmits_timed_out() */
6174 tcp_sk(sk)->retrans_stamp = 0;
6175 inet_csk(sk)->icsk_retransmits = 0;
6177 /* Once we leave TCP_SYN_RECV or TCP_FIN_WAIT_1,
6178 * we no longer need req so release it.
6180 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk,
6181 lockdep_sock_is_held(sk));
6182 reqsk_fastopen_remove(sk, req, false);
6184 /* Re-arm the timer because data may have been sent out.
6185 * This is similar to the regular data transmission case
6186 * when new data has just been ack'ed.
6188 * (TFO) - we could try to be more aggressive and
6189 * retransmitting any data sooner based on when they
6196 * This function implements the receiving procedure of RFC 793 for
6197 * all states except ESTABLISHED and TIME_WAIT.
6198 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
6199 * address independent.
6202 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb)
6204 struct tcp_sock *tp = tcp_sk(sk);
6205 struct inet_connection_sock *icsk = inet_csk(sk);
6206 const struct tcphdr *th = tcp_hdr(skb);
6207 struct request_sock *req;
6211 switch (sk->sk_state) {
6225 /* It is possible that we process SYN packets from backlog,
6226 * so we need to make sure to disable BH and RCU right there.
6230 acceptable = icsk->icsk_af_ops->conn_request(sk, skb) >= 0;
6242 tp->rx_opt.saw_tstamp = 0;
6243 tcp_mstamp_refresh(tp);
6244 queued = tcp_rcv_synsent_state_process(sk, skb, th);
6248 /* Do step6 onward by hand. */
6249 tcp_urg(sk, skb, th);
6251 tcp_data_snd_check(sk);
6255 tcp_mstamp_refresh(tp);
6256 tp->rx_opt.saw_tstamp = 0;
6257 req = rcu_dereference_protected(tp->fastopen_rsk,
6258 lockdep_sock_is_held(sk));
6262 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
6263 sk->sk_state != TCP_FIN_WAIT1);
6265 if (!tcp_check_req(sk, skb, req, true, &req_stolen))
6269 if (!th->ack && !th->rst && !th->syn)
6272 if (!tcp_validate_incoming(sk, skb, th, 0))
6275 /* step 5: check the ACK field */
6276 acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
6277 FLAG_UPDATE_TS_RECENT |
6278 FLAG_NO_CHALLENGE_ACK) > 0;
6281 if (sk->sk_state == TCP_SYN_RECV)
6282 return 1; /* send one RST */
6283 tcp_send_challenge_ack(sk, skb);
6286 switch (sk->sk_state) {
6288 tp->delivered++; /* SYN-ACK delivery isn't tracked in tcp_ack */
6290 tcp_synack_rtt_meas(sk, req);
6293 tcp_rcv_synrecv_state_fastopen(sk);
6295 tcp_try_undo_spurious_syn(sk);
6296 tp->retrans_stamp = 0;
6297 tcp_init_transfer(sk, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB);
6298 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6301 tcp_set_state(sk, TCP_ESTABLISHED);
6302 sk->sk_state_change(sk);
6304 /* Note, that this wakeup is only for marginal crossed SYN case.
6305 * Passively open sockets are not waked up, because
6306 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
6309 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
6311 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
6312 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
6313 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
6315 if (tp->rx_opt.tstamp_ok)
6316 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
6318 if (!inet_csk(sk)->icsk_ca_ops->cong_control)
6319 tcp_update_pacing_rate(sk);
6321 /* Prevent spurious tcp_cwnd_restart() on first data packet */
6322 tp->lsndtime = tcp_jiffies32;
6324 tcp_initialize_rcv_mss(sk);
6325 tcp_fast_path_on(tp);
6328 case TCP_FIN_WAIT1: {
6332 tcp_rcv_synrecv_state_fastopen(sk);
6334 if (tp->snd_una != tp->write_seq)
6337 tcp_set_state(sk, TCP_FIN_WAIT2);
6338 sk->sk_shutdown |= SEND_SHUTDOWN;
6342 if (!sock_flag(sk, SOCK_DEAD)) {
6343 /* Wake up lingering close() */
6344 sk->sk_state_change(sk);
6348 if (tp->linger2 < 0) {
6350 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6353 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6354 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6355 /* Receive out of order FIN after close() */
6356 if (tp->syn_fastopen && th->fin)
6357 tcp_fastopen_active_disable(sk);
6359 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6363 tmo = tcp_fin_time(sk);
6364 if (tmo > TCP_TIMEWAIT_LEN) {
6365 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
6366 } else if (th->fin || sock_owned_by_user(sk)) {
6367 /* Bad case. We could lose such FIN otherwise.
6368 * It is not a big problem, but it looks confusing
6369 * and not so rare event. We still can lose it now,
6370 * if it spins in bh_lock_sock(), but it is really
6373 inet_csk_reset_keepalive_timer(sk, tmo);
6375 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
6382 if (tp->snd_una == tp->write_seq) {
6383 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
6389 if (tp->snd_una == tp->write_seq) {
6390 tcp_update_metrics(sk);
6397 /* step 6: check the URG bit */
6398 tcp_urg(sk, skb, th);
6400 /* step 7: process the segment text */
6401 switch (sk->sk_state) {
6402 case TCP_CLOSE_WAIT:
6405 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
6410 /* RFC 793 says to queue data in these states,
6411 * RFC 1122 says we MUST send a reset.
6412 * BSD 4.4 also does reset.
6414 if (sk->sk_shutdown & RCV_SHUTDOWN) {
6415 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6416 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6417 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6423 case TCP_ESTABLISHED:
6424 tcp_data_queue(sk, skb);
6429 /* tcp_data could move socket to TIME-WAIT */
6430 if (sk->sk_state != TCP_CLOSE) {
6431 tcp_data_snd_check(sk);
6432 tcp_ack_snd_check(sk);
6441 EXPORT_SYMBOL(tcp_rcv_state_process);
6443 static inline void pr_drop_req(struct request_sock *req, __u16 port, int family)
6445 struct inet_request_sock *ireq = inet_rsk(req);
6447 if (family == AF_INET)
6448 net_dbg_ratelimited("drop open request from %pI4/%u\n",
6449 &ireq->ir_rmt_addr, port);
6450 #if IS_ENABLED(CONFIG_IPV6)
6451 else if (family == AF_INET6)
6452 net_dbg_ratelimited("drop open request from %pI6/%u\n",
6453 &ireq->ir_v6_rmt_addr, port);
6457 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
6459 * If we receive a SYN packet with these bits set, it means a
6460 * network is playing bad games with TOS bits. In order to
6461 * avoid possible false congestion notifications, we disable
6462 * TCP ECN negotiation.
6464 * Exception: tcp_ca wants ECN. This is required for DCTCP
6465 * congestion control: Linux DCTCP asserts ECT on all packets,
6466 * including SYN, which is most optimal solution; however,
6467 * others, such as FreeBSD do not.
6469 * Exception: At least one of the reserved bits of the TCP header (th->res1) is
6470 * set, indicating the use of a future TCP extension (such as AccECN). See
6471 * RFC8311 §4.3 which updates RFC3168 to allow the development of such
6474 static void tcp_ecn_create_request(struct request_sock *req,
6475 const struct sk_buff *skb,
6476 const struct sock *listen_sk,
6477 const struct dst_entry *dst)
6479 const struct tcphdr *th = tcp_hdr(skb);
6480 const struct net *net = sock_net(listen_sk);
6481 bool th_ecn = th->ece && th->cwr;
6488 ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield);
6489 ecn_ok_dst = dst_feature(dst, DST_FEATURE_ECN_MASK);
6490 ecn_ok = net->ipv4.sysctl_tcp_ecn || ecn_ok_dst;
6492 if (((!ect || th->res1) && ecn_ok) || tcp_ca_needs_ecn(listen_sk) ||
6493 (ecn_ok_dst & DST_FEATURE_ECN_CA) ||
6494 tcp_bpf_ca_needs_ecn((struct sock *)req))
6495 inet_rsk(req)->ecn_ok = 1;
6498 static void tcp_openreq_init(struct request_sock *req,
6499 const struct tcp_options_received *rx_opt,
6500 struct sk_buff *skb, const struct sock *sk)
6502 struct inet_request_sock *ireq = inet_rsk(req);
6504 req->rsk_rcv_wnd = 0; /* So that tcp_send_synack() knows! */
6506 tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq;
6507 tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
6508 tcp_rsk(req)->snt_synack = 0;
6509 tcp_rsk(req)->last_oow_ack_time = 0;
6510 req->mss = rx_opt->mss_clamp;
6511 req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0;
6512 ireq->tstamp_ok = rx_opt->tstamp_ok;
6513 ireq->sack_ok = rx_opt->sack_ok;
6514 ireq->snd_wscale = rx_opt->snd_wscale;
6515 ireq->wscale_ok = rx_opt->wscale_ok;
6518 ireq->ir_rmt_port = tcp_hdr(skb)->source;
6519 ireq->ir_num = ntohs(tcp_hdr(skb)->dest);
6520 ireq->ir_mark = inet_request_mark(sk, skb);
6521 #if IS_ENABLED(CONFIG_SMC)
6522 ireq->smc_ok = rx_opt->smc_ok;
6526 struct request_sock *inet_reqsk_alloc(const struct request_sock_ops *ops,
6527 struct sock *sk_listener,
6528 bool attach_listener)
6530 struct request_sock *req = reqsk_alloc(ops, sk_listener,
6534 struct inet_request_sock *ireq = inet_rsk(req);
6536 ireq->ireq_opt = NULL;
6537 #if IS_ENABLED(CONFIG_IPV6)
6538 ireq->pktopts = NULL;
6540 atomic64_set(&ireq->ir_cookie, 0);
6541 ireq->ireq_state = TCP_NEW_SYN_RECV;
6542 write_pnet(&ireq->ireq_net, sock_net(sk_listener));
6543 ireq->ireq_family = sk_listener->sk_family;
6548 EXPORT_SYMBOL(inet_reqsk_alloc);
6551 * Return true if a syncookie should be sent
6553 static bool tcp_syn_flood_action(const struct sock *sk, const char *proto)
6555 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
6556 const char *msg = "Dropping request";
6557 struct net *net = sock_net(sk);
6558 bool want_cookie = false;
6561 syncookies = READ_ONCE(net->ipv4.sysctl_tcp_syncookies);
6563 #ifdef CONFIG_SYN_COOKIES
6565 msg = "Sending cookies";
6567 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES);
6570 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP);
6572 if (!queue->synflood_warned && syncookies != 2 &&
6573 xchg(&queue->synflood_warned, 1) == 0)
6574 net_info_ratelimited("%s: Possible SYN flooding on port %d. %s. Check SNMP counters.\n",
6575 proto, sk->sk_num, msg);
6580 static void tcp_reqsk_record_syn(const struct sock *sk,
6581 struct request_sock *req,
6582 const struct sk_buff *skb)
6584 if (tcp_sk(sk)->save_syn) {
6585 u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb);
6588 copy = kmalloc(len + sizeof(u32), GFP_ATOMIC);
6591 memcpy(©[1], skb_network_header(skb), len);
6592 req->saved_syn = copy;
6597 /* If a SYN cookie is required and supported, returns a clamped MSS value to be
6598 * used for SYN cookie generation.
6600 u16 tcp_get_syncookie_mss(struct request_sock_ops *rsk_ops,
6601 const struct tcp_request_sock_ops *af_ops,
6602 struct sock *sk, struct tcphdr *th)
6604 struct tcp_sock *tp = tcp_sk(sk);
6607 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_syncookies) != 2 &&
6608 !inet_csk_reqsk_queue_is_full(sk))
6611 if (!tcp_syn_flood_action(sk, rsk_ops->slab_name))
6614 if (sk_acceptq_is_full(sk)) {
6615 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
6619 mss = tcp_parse_mss_option(th, tp->rx_opt.user_mss);
6621 mss = af_ops->mss_clamp;
6625 EXPORT_SYMBOL_GPL(tcp_get_syncookie_mss);
6627 int tcp_conn_request(struct request_sock_ops *rsk_ops,
6628 const struct tcp_request_sock_ops *af_ops,
6629 struct sock *sk, struct sk_buff *skb)
6631 struct tcp_fastopen_cookie foc = { .len = -1 };
6632 __u32 isn = TCP_SKB_CB(skb)->tcp_tw_isn;
6633 struct tcp_options_received tmp_opt;
6634 struct tcp_sock *tp = tcp_sk(sk);
6635 struct net *net = sock_net(sk);
6636 struct sock *fastopen_sk = NULL;
6637 struct request_sock *req;
6638 bool want_cookie = false;
6639 struct dst_entry *dst;
6643 syncookies = READ_ONCE(net->ipv4.sysctl_tcp_syncookies);
6645 /* TW buckets are converted to open requests without
6646 * limitations, they conserve resources and peer is
6647 * evidently real one.
6649 if ((syncookies == 2 || inet_csk_reqsk_queue_is_full(sk)) && !isn) {
6650 want_cookie = tcp_syn_flood_action(sk, rsk_ops->slab_name);
6655 if (sk_acceptq_is_full(sk)) {
6656 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
6660 req = inet_reqsk_alloc(rsk_ops, sk, !want_cookie);
6664 tcp_rsk(req)->af_specific = af_ops;
6665 tcp_rsk(req)->ts_off = 0;
6667 tcp_clear_options(&tmp_opt);
6668 tmp_opt.mss_clamp = af_ops->mss_clamp;
6669 tmp_opt.user_mss = tp->rx_opt.user_mss;
6670 tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0,
6671 want_cookie ? NULL : &foc);
6673 if (want_cookie && !tmp_opt.saw_tstamp)
6674 tcp_clear_options(&tmp_opt);
6676 if (IS_ENABLED(CONFIG_SMC) && want_cookie)
6679 tmp_opt.tstamp_ok = tmp_opt.saw_tstamp;
6680 tcp_openreq_init(req, &tmp_opt, skb, sk);
6681 inet_rsk(req)->no_srccheck = inet_sk(sk)->transparent;
6683 /* Note: tcp_v6_init_req() might override ir_iif for link locals */
6684 inet_rsk(req)->ir_iif = inet_request_bound_dev_if(sk, skb);
6686 af_ops->init_req(req, sk, skb);
6688 if (security_inet_conn_request(sk, skb, req))
6691 if (tmp_opt.tstamp_ok)
6692 tcp_rsk(req)->ts_off = af_ops->init_ts_off(net, skb);
6694 dst = af_ops->route_req(sk, &fl, req);
6698 if (!want_cookie && !isn) {
6699 int max_syn_backlog = READ_ONCE(net->ipv4.sysctl_max_syn_backlog);
6701 /* Kill the following clause, if you dislike this way. */
6703 (max_syn_backlog - inet_csk_reqsk_queue_len(sk) <
6704 (max_syn_backlog >> 2)) &&
6705 !tcp_peer_is_proven(req, dst)) {
6706 /* Without syncookies last quarter of
6707 * backlog is filled with destinations,
6708 * proven to be alive.
6709 * It means that we continue to communicate
6710 * to destinations, already remembered
6711 * to the moment of synflood.
6713 pr_drop_req(req, ntohs(tcp_hdr(skb)->source),
6715 goto drop_and_release;
6718 isn = af_ops->init_seq(skb);
6721 tcp_ecn_create_request(req, skb, sk, dst);
6724 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss);
6725 req->cookie_ts = tmp_opt.tstamp_ok;
6726 if (!tmp_opt.tstamp_ok)
6727 inet_rsk(req)->ecn_ok = 0;
6730 tcp_rsk(req)->snt_isn = isn;
6731 tcp_rsk(req)->txhash = net_tx_rndhash();
6732 tcp_openreq_init_rwin(req, sk, dst);
6733 sk_rx_queue_set(req_to_sk(req), skb);
6735 tcp_reqsk_record_syn(sk, req, skb);
6736 fastopen_sk = tcp_try_fastopen(sk, skb, req, &foc, dst);
6739 af_ops->send_synack(fastopen_sk, dst, &fl, req,
6740 &foc, TCP_SYNACK_FASTOPEN);
6741 /* Add the child socket directly into the accept queue */
6742 if (!inet_csk_reqsk_queue_add(sk, req, fastopen_sk)) {
6743 reqsk_fastopen_remove(fastopen_sk, req, false);
6744 bh_unlock_sock(fastopen_sk);
6745 sock_put(fastopen_sk);
6748 sk->sk_data_ready(sk);
6749 bh_unlock_sock(fastopen_sk);
6750 sock_put(fastopen_sk);
6752 tcp_rsk(req)->tfo_listener = false;
6754 inet_csk_reqsk_queue_hash_add(sk, req,
6755 tcp_timeout_init((struct sock *)req));
6756 af_ops->send_synack(sk, dst, &fl, req, &foc,
6757 !want_cookie ? TCP_SYNACK_NORMAL :
6775 EXPORT_SYMBOL(tcp_conn_request);