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>
82 #include <net/mptcp.h>
84 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
86 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
87 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
88 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
89 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
90 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
91 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
92 #define FLAG_ECE 0x40 /* ECE in this ACK */
93 #define FLAG_LOST_RETRANS 0x80 /* This ACK marks some retransmission lost */
94 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
95 #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
96 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
97 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
98 #define FLAG_SET_XMIT_TIMER 0x1000 /* Set TLP or RTO timer */
99 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
100 #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
101 #define FLAG_NO_CHALLENGE_ACK 0x8000 /* do not call tcp_send_challenge_ack() */
102 #define FLAG_ACK_MAYBE_DELAYED 0x10000 /* Likely a delayed ACK */
104 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
105 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
106 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE|FLAG_DSACKING_ACK)
107 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
109 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
110 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
112 #define REXMIT_NONE 0 /* no loss recovery to do */
113 #define REXMIT_LOST 1 /* retransmit packets marked lost */
114 #define REXMIT_NEW 2 /* FRTO-style transmit of unsent/new packets */
116 #if IS_ENABLED(CONFIG_TLS_DEVICE)
117 static DEFINE_STATIC_KEY_DEFERRED_FALSE(clean_acked_data_enabled, HZ);
119 void clean_acked_data_enable(struct inet_connection_sock *icsk,
120 void (*cad)(struct sock *sk, u32 ack_seq))
122 icsk->icsk_clean_acked = cad;
123 static_branch_deferred_inc(&clean_acked_data_enabled);
125 EXPORT_SYMBOL_GPL(clean_acked_data_enable);
127 void clean_acked_data_disable(struct inet_connection_sock *icsk)
129 static_branch_slow_dec_deferred(&clean_acked_data_enabled);
130 icsk->icsk_clean_acked = NULL;
132 EXPORT_SYMBOL_GPL(clean_acked_data_disable);
134 void clean_acked_data_flush(void)
136 static_key_deferred_flush(&clean_acked_data_enabled);
138 EXPORT_SYMBOL_GPL(clean_acked_data_flush);
141 #ifdef CONFIG_CGROUP_BPF
142 static void bpf_skops_parse_hdr(struct sock *sk, struct sk_buff *skb)
144 bool unknown_opt = tcp_sk(sk)->rx_opt.saw_unknown &&
145 BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk),
146 BPF_SOCK_OPS_PARSE_UNKNOWN_HDR_OPT_CB_FLAG);
147 bool parse_all_opt = BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk),
148 BPF_SOCK_OPS_PARSE_ALL_HDR_OPT_CB_FLAG);
149 struct bpf_sock_ops_kern sock_ops;
151 if (likely(!unknown_opt && !parse_all_opt))
154 /* The skb will be handled in the
155 * bpf_skops_established() or
156 * bpf_skops_write_hdr_opt().
158 switch (sk->sk_state) {
165 sock_owned_by_me(sk);
167 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp));
168 sock_ops.op = BPF_SOCK_OPS_PARSE_HDR_OPT_CB;
169 sock_ops.is_fullsock = 1;
171 bpf_skops_init_skb(&sock_ops, skb, tcp_hdrlen(skb));
173 BPF_CGROUP_RUN_PROG_SOCK_OPS(&sock_ops);
176 static void bpf_skops_established(struct sock *sk, int bpf_op,
179 struct bpf_sock_ops_kern sock_ops;
181 sock_owned_by_me(sk);
183 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp));
184 sock_ops.op = bpf_op;
185 sock_ops.is_fullsock = 1;
187 /* sk with TCP_REPAIR_ON does not have skb in tcp_finish_connect */
189 bpf_skops_init_skb(&sock_ops, skb, tcp_hdrlen(skb));
191 BPF_CGROUP_RUN_PROG_SOCK_OPS(&sock_ops);
194 static void bpf_skops_parse_hdr(struct sock *sk, struct sk_buff *skb)
198 static void bpf_skops_established(struct sock *sk, int bpf_op,
204 static void tcp_gro_dev_warn(struct sock *sk, const struct sk_buff *skb,
207 static bool __once __read_mostly;
210 struct net_device *dev;
215 dev = dev_get_by_index_rcu(sock_net(sk), skb->skb_iif);
216 if (!dev || len >= dev->mtu)
217 pr_warn("%s: Driver has suspect GRO implementation, TCP performance may be compromised.\n",
218 dev ? dev->name : "Unknown driver");
223 /* Adapt the MSS value used to make delayed ack decision to the
226 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
228 struct inet_connection_sock *icsk = inet_csk(sk);
229 const unsigned int lss = icsk->icsk_ack.last_seg_size;
232 icsk->icsk_ack.last_seg_size = 0;
234 /* skb->len may jitter because of SACKs, even if peer
235 * sends good full-sized frames.
237 len = skb_shinfo(skb)->gso_size ? : skb->len;
238 if (len >= icsk->icsk_ack.rcv_mss) {
239 icsk->icsk_ack.rcv_mss = min_t(unsigned int, len,
241 /* Account for possibly-removed options */
242 if (unlikely(len > icsk->icsk_ack.rcv_mss +
243 MAX_TCP_OPTION_SPACE))
244 tcp_gro_dev_warn(sk, skb, len);
245 /* If the skb has a len of exactly 1*MSS and has the PSH bit
246 * set then it is likely the end of an application write. So
247 * more data may not be arriving soon, and yet the data sender
248 * may be waiting for an ACK if cwnd-bound or using TX zero
249 * copy. So we set ICSK_ACK_PUSHED here so that
250 * tcp_cleanup_rbuf() will send an ACK immediately if the app
251 * reads all of the data and is not ping-pong. If len > MSS
252 * then this logic does not matter (and does not hurt) because
253 * tcp_cleanup_rbuf() will always ACK immediately if the app
254 * reads data and there is more than an MSS of unACKed data.
256 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_PSH)
257 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
259 /* Otherwise, we make more careful check taking into account,
260 * that SACKs block is variable.
262 * "len" is invariant segment length, including TCP header.
264 len += skb->data - skb_transport_header(skb);
265 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
266 /* If PSH is not set, packet should be
267 * full sized, provided peer TCP is not badly broken.
268 * This observation (if it is correct 8)) allows
269 * to handle super-low mtu links fairly.
271 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
272 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
273 /* Subtract also invariant (if peer is RFC compliant),
274 * tcp header plus fixed timestamp option length.
275 * Resulting "len" is MSS free of SACK jitter.
277 len -= tcp_sk(sk)->tcp_header_len;
278 icsk->icsk_ack.last_seg_size = len;
280 icsk->icsk_ack.rcv_mss = len;
284 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
285 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
286 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
290 static void tcp_incr_quickack(struct sock *sk, unsigned int max_quickacks)
292 struct inet_connection_sock *icsk = inet_csk(sk);
293 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
297 quickacks = min(quickacks, max_quickacks);
298 if (quickacks > icsk->icsk_ack.quick)
299 icsk->icsk_ack.quick = quickacks;
302 static void tcp_enter_quickack_mode(struct sock *sk, unsigned int max_quickacks)
304 struct inet_connection_sock *icsk = inet_csk(sk);
306 tcp_incr_quickack(sk, max_quickacks);
307 inet_csk_exit_pingpong_mode(sk);
308 icsk->icsk_ack.ato = TCP_ATO_MIN;
311 /* Send ACKs quickly, if "quick" count is not exhausted
312 * and the session is not interactive.
315 static bool tcp_in_quickack_mode(struct sock *sk)
317 const struct inet_connection_sock *icsk = inet_csk(sk);
318 const struct dst_entry *dst = __sk_dst_get(sk);
320 return (dst && dst_metric(dst, RTAX_QUICKACK)) ||
321 (icsk->icsk_ack.quick && !inet_csk_in_pingpong_mode(sk));
324 static void tcp_ecn_queue_cwr(struct tcp_sock *tp)
326 if (tp->ecn_flags & TCP_ECN_OK)
327 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
330 static void tcp_ecn_accept_cwr(struct sock *sk, const struct sk_buff *skb)
332 if (tcp_hdr(skb)->cwr) {
333 tcp_sk(sk)->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
335 /* If the sender is telling us it has entered CWR, then its
336 * cwnd may be very low (even just 1 packet), so we should ACK
339 if (TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq)
340 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW;
344 static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp)
346 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
349 static void __tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb)
351 struct tcp_sock *tp = tcp_sk(sk);
353 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
354 case INET_ECN_NOT_ECT:
355 /* Funny extension: if ECT is not set on a segment,
356 * and we already seen ECT on a previous segment,
357 * it is probably a retransmit.
359 if (tp->ecn_flags & TCP_ECN_SEEN)
360 tcp_enter_quickack_mode(sk, 2);
363 if (tcp_ca_needs_ecn(sk))
364 tcp_ca_event(sk, CA_EVENT_ECN_IS_CE);
366 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
367 /* Better not delay acks, sender can have a very low cwnd */
368 tcp_enter_quickack_mode(sk, 2);
369 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
371 tp->ecn_flags |= TCP_ECN_SEEN;
374 if (tcp_ca_needs_ecn(sk))
375 tcp_ca_event(sk, CA_EVENT_ECN_NO_CE);
376 tp->ecn_flags |= TCP_ECN_SEEN;
381 static void tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb)
383 if (tcp_sk(sk)->ecn_flags & TCP_ECN_OK)
384 __tcp_ecn_check_ce(sk, skb);
387 static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
389 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
390 tp->ecn_flags &= ~TCP_ECN_OK;
393 static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
395 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
396 tp->ecn_flags &= ~TCP_ECN_OK;
399 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
401 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
406 /* Buffer size and advertised window tuning.
408 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
411 static void tcp_sndbuf_expand(struct sock *sk)
413 const struct tcp_sock *tp = tcp_sk(sk);
414 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
418 /* Worst case is non GSO/TSO : each frame consumes one skb
419 * and skb->head is kmalloced using power of two area of memory
421 per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
423 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
425 per_mss = roundup_pow_of_two(per_mss) +
426 SKB_DATA_ALIGN(sizeof(struct sk_buff));
428 nr_segs = max_t(u32, TCP_INIT_CWND, tp->snd_cwnd);
429 nr_segs = max_t(u32, nr_segs, tp->reordering + 1);
431 /* Fast Recovery (RFC 5681 3.2) :
432 * Cubic needs 1.7 factor, rounded to 2 to include
433 * extra cushion (application might react slowly to EPOLLOUT)
435 sndmem = ca_ops->sndbuf_expand ? ca_ops->sndbuf_expand(sk) : 2;
436 sndmem *= nr_segs * per_mss;
438 if (sk->sk_sndbuf < sndmem)
439 WRITE_ONCE(sk->sk_sndbuf,
440 min(sndmem, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_wmem[2])));
443 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
445 * All tcp_full_space() is split to two parts: "network" buffer, allocated
446 * forward and advertised in receiver window (tp->rcv_wnd) and
447 * "application buffer", required to isolate scheduling/application
448 * latencies from network.
449 * window_clamp is maximal advertised window. It can be less than
450 * tcp_full_space(), in this case tcp_full_space() - window_clamp
451 * is reserved for "application" buffer. The less window_clamp is
452 * the smoother our behaviour from viewpoint of network, but the lower
453 * throughput and the higher sensitivity of the connection to losses. 8)
455 * rcv_ssthresh is more strict window_clamp used at "slow start"
456 * phase to predict further behaviour of this connection.
457 * It is used for two goals:
458 * - to enforce header prediction at sender, even when application
459 * requires some significant "application buffer". It is check #1.
460 * - to prevent pruning of receive queue because of misprediction
461 * of receiver window. Check #2.
463 * The scheme does not work when sender sends good segments opening
464 * window and then starts to feed us spaghetti. But it should work
465 * in common situations. Otherwise, we have to rely on queue collapsing.
468 /* Slow part of check#2. */
469 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb,
470 unsigned int skbtruesize)
472 struct tcp_sock *tp = tcp_sk(sk);
474 int truesize = tcp_win_from_space(sk, skbtruesize) >> 1;
475 int window = tcp_win_from_space(sk, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2])) >> 1;
477 while (tp->rcv_ssthresh <= window) {
478 if (truesize <= skb->len)
479 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
487 /* Even if skb appears to have a bad len/truesize ratio, TCP coalescing
488 * can play nice with us, as sk_buff and skb->head might be either
489 * freed or shared with up to MAX_SKB_FRAGS segments.
490 * Only give a boost to drivers using page frag(s) to hold the frame(s),
491 * and if no payload was pulled in skb->head before reaching us.
493 static u32 truesize_adjust(bool adjust, const struct sk_buff *skb)
495 u32 truesize = skb->truesize;
497 if (adjust && !skb_headlen(skb)) {
498 truesize -= SKB_TRUESIZE(skb_end_offset(skb));
499 /* paranoid check, some drivers might be buggy */
500 if (unlikely((int)truesize < (int)skb->len))
501 truesize = skb->truesize;
506 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb,
509 struct tcp_sock *tp = tcp_sk(sk);
512 room = min_t(int, tp->window_clamp, tcp_space(sk)) - tp->rcv_ssthresh;
515 if (room > 0 && !tcp_under_memory_pressure(sk)) {
516 unsigned int truesize = truesize_adjust(adjust, skb);
519 /* Check #2. Increase window, if skb with such overhead
520 * will fit to rcvbuf in future.
522 if (tcp_win_from_space(sk, truesize) <= skb->len)
523 incr = 2 * tp->advmss;
525 incr = __tcp_grow_window(sk, skb, truesize);
528 incr = max_t(int, incr, 2 * skb->len);
529 tp->rcv_ssthresh += min(room, incr);
530 inet_csk(sk)->icsk_ack.quick |= 1;
535 /* 3. Try to fixup all. It is made immediately after connection enters
538 static void tcp_init_buffer_space(struct sock *sk)
540 int tcp_app_win = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_app_win);
541 struct tcp_sock *tp = tcp_sk(sk);
544 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
545 tcp_sndbuf_expand(sk);
547 tcp_mstamp_refresh(tp);
548 tp->rcvq_space.time = tp->tcp_mstamp;
549 tp->rcvq_space.seq = tp->copied_seq;
551 maxwin = tcp_full_space(sk);
553 if (tp->window_clamp >= maxwin) {
554 tp->window_clamp = maxwin;
556 if (tcp_app_win && maxwin > 4 * tp->advmss)
557 tp->window_clamp = max(maxwin -
558 (maxwin >> tcp_app_win),
562 /* Force reservation of one segment. */
564 tp->window_clamp > 2 * tp->advmss &&
565 tp->window_clamp + tp->advmss > maxwin)
566 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
568 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
569 tp->snd_cwnd_stamp = tcp_jiffies32;
570 tp->rcvq_space.space = min3(tp->rcv_ssthresh, tp->rcv_wnd,
571 (u32)TCP_INIT_CWND * tp->advmss);
574 /* 4. Recalculate window clamp after socket hit its memory bounds. */
575 static void tcp_clamp_window(struct sock *sk)
577 struct tcp_sock *tp = tcp_sk(sk);
578 struct inet_connection_sock *icsk = inet_csk(sk);
579 struct net *net = sock_net(sk);
582 icsk->icsk_ack.quick = 0;
583 rmem2 = READ_ONCE(net->ipv4.sysctl_tcp_rmem[2]);
585 if (sk->sk_rcvbuf < rmem2 &&
586 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
587 !tcp_under_memory_pressure(sk) &&
588 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
589 WRITE_ONCE(sk->sk_rcvbuf,
590 min(atomic_read(&sk->sk_rmem_alloc), rmem2));
592 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
593 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
596 /* Initialize RCV_MSS value.
597 * RCV_MSS is an our guess about MSS used by the peer.
598 * We haven't any direct information about the MSS.
599 * It's better to underestimate the RCV_MSS rather than overestimate.
600 * Overestimations make us ACKing less frequently than needed.
601 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
603 void tcp_initialize_rcv_mss(struct sock *sk)
605 const struct tcp_sock *tp = tcp_sk(sk);
606 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
608 hint = min(hint, tp->rcv_wnd / 2);
609 hint = min(hint, TCP_MSS_DEFAULT);
610 hint = max(hint, TCP_MIN_MSS);
612 inet_csk(sk)->icsk_ack.rcv_mss = hint;
614 EXPORT_SYMBOL(tcp_initialize_rcv_mss);
616 /* Receiver "autotuning" code.
618 * The algorithm for RTT estimation w/o timestamps is based on
619 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
620 * <https://public.lanl.gov/radiant/pubs.html#DRS>
622 * More detail on this code can be found at
623 * <http://staff.psc.edu/jheffner/>,
624 * though this reference is out of date. A new paper
627 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
629 u32 new_sample = tp->rcv_rtt_est.rtt_us;
632 if (new_sample != 0) {
633 /* If we sample in larger samples in the non-timestamp
634 * case, we could grossly overestimate the RTT especially
635 * with chatty applications or bulk transfer apps which
636 * are stalled on filesystem I/O.
638 * Also, since we are only going for a minimum in the
639 * non-timestamp case, we do not smooth things out
640 * else with timestamps disabled convergence takes too
644 m -= (new_sample >> 3);
652 /* No previous measure. */
656 tp->rcv_rtt_est.rtt_us = new_sample;
659 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
663 if (tp->rcv_rtt_est.time == 0)
665 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
667 delta_us = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcv_rtt_est.time);
670 tcp_rcv_rtt_update(tp, delta_us, 1);
673 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
674 tp->rcv_rtt_est.time = tp->tcp_mstamp;
677 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
678 const struct sk_buff *skb)
680 struct tcp_sock *tp = tcp_sk(sk);
682 if (tp->rx_opt.rcv_tsecr == tp->rcv_rtt_last_tsecr)
684 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr;
686 if (TCP_SKB_CB(skb)->end_seq -
687 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss) {
688 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr;
691 if (likely(delta < INT_MAX / (USEC_PER_SEC / TCP_TS_HZ))) {
694 delta_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
695 tcp_rcv_rtt_update(tp, delta_us, 0);
701 * This function should be called every time data is copied to user space.
702 * It calculates the appropriate TCP receive buffer space.
704 void tcp_rcv_space_adjust(struct sock *sk)
706 struct tcp_sock *tp = tcp_sk(sk);
710 trace_tcp_rcv_space_adjust(sk);
712 tcp_mstamp_refresh(tp);
713 time = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcvq_space.time);
714 if (time < (tp->rcv_rtt_est.rtt_us >> 3) || tp->rcv_rtt_est.rtt_us == 0)
717 /* Number of bytes copied to user in last RTT */
718 copied = tp->copied_seq - tp->rcvq_space.seq;
719 if (copied <= tp->rcvq_space.space)
723 * copied = bytes received in previous RTT, our base window
724 * To cope with packet losses, we need a 2x factor
725 * To cope with slow start, and sender growing its cwin by 100 %
726 * every RTT, we need a 4x factor, because the ACK we are sending
727 * now is for the next RTT, not the current one :
728 * <prev RTT . ><current RTT .. ><next RTT .... >
731 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf) &&
732 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
736 /* minimal window to cope with packet losses, assuming
737 * steady state. Add some cushion because of small variations.
739 rcvwin = ((u64)copied << 1) + 16 * tp->advmss;
741 /* Accommodate for sender rate increase (eg. slow start) */
742 grow = rcvwin * (copied - tp->rcvq_space.space);
743 do_div(grow, tp->rcvq_space.space);
744 rcvwin += (grow << 1);
746 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
747 while (tcp_win_from_space(sk, rcvmem) < tp->advmss)
750 do_div(rcvwin, tp->advmss);
751 rcvbuf = min_t(u64, rcvwin * rcvmem,
752 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]));
753 if (rcvbuf > sk->sk_rcvbuf) {
754 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
756 /* Make the window clamp follow along. */
757 tp->window_clamp = tcp_win_from_space(sk, rcvbuf);
760 tp->rcvq_space.space = copied;
763 tp->rcvq_space.seq = tp->copied_seq;
764 tp->rcvq_space.time = tp->tcp_mstamp;
767 /* There is something which you must keep in mind when you analyze the
768 * behavior of the tp->ato delayed ack timeout interval. When a
769 * connection starts up, we want to ack as quickly as possible. The
770 * problem is that "good" TCP's do slow start at the beginning of data
771 * transmission. The means that until we send the first few ACK's the
772 * sender will sit on his end and only queue most of his data, because
773 * he can only send snd_cwnd unacked packets at any given time. For
774 * each ACK we send, he increments snd_cwnd and transmits more of his
777 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
779 struct tcp_sock *tp = tcp_sk(sk);
780 struct inet_connection_sock *icsk = inet_csk(sk);
783 inet_csk_schedule_ack(sk);
785 tcp_measure_rcv_mss(sk, skb);
787 tcp_rcv_rtt_measure(tp);
791 if (!icsk->icsk_ack.ato) {
792 /* The _first_ data packet received, initialize
793 * delayed ACK engine.
795 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
796 icsk->icsk_ack.ato = TCP_ATO_MIN;
798 int m = now - icsk->icsk_ack.lrcvtime;
800 if (m <= TCP_ATO_MIN / 2) {
801 /* The fastest case is the first. */
802 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
803 } else if (m < icsk->icsk_ack.ato) {
804 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
805 if (icsk->icsk_ack.ato > icsk->icsk_rto)
806 icsk->icsk_ack.ato = icsk->icsk_rto;
807 } else if (m > icsk->icsk_rto) {
808 /* Too long gap. Apparently sender failed to
809 * restart window, so that we send ACKs quickly.
811 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
815 icsk->icsk_ack.lrcvtime = now;
817 tcp_ecn_check_ce(sk, skb);
820 tcp_grow_window(sk, skb, true);
823 /* Called to compute a smoothed rtt estimate. The data fed to this
824 * routine either comes from timestamps, or from segments that were
825 * known _not_ to have been retransmitted [see Karn/Partridge
826 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
827 * piece by Van Jacobson.
828 * NOTE: the next three routines used to be one big routine.
829 * To save cycles in the RFC 1323 implementation it was better to break
830 * it up into three procedures. -- erics
832 static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
834 struct tcp_sock *tp = tcp_sk(sk);
835 long m = mrtt_us; /* RTT */
836 u32 srtt = tp->srtt_us;
838 /* The following amusing code comes from Jacobson's
839 * article in SIGCOMM '88. Note that rtt and mdev
840 * are scaled versions of rtt and mean deviation.
841 * This is designed to be as fast as possible
842 * m stands for "measurement".
844 * On a 1990 paper the rto value is changed to:
845 * RTO = rtt + 4 * mdev
847 * Funny. This algorithm seems to be very broken.
848 * These formulae increase RTO, when it should be decreased, increase
849 * too slowly, when it should be increased quickly, decrease too quickly
850 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
851 * does not matter how to _calculate_ it. Seems, it was trap
852 * that VJ failed to avoid. 8)
855 m -= (srtt >> 3); /* m is now error in rtt est */
856 srtt += m; /* rtt = 7/8 rtt + 1/8 new */
858 m = -m; /* m is now abs(error) */
859 m -= (tp->mdev_us >> 2); /* similar update on mdev */
860 /* This is similar to one of Eifel findings.
861 * Eifel blocks mdev updates when rtt decreases.
862 * This solution is a bit different: we use finer gain
863 * for mdev in this case (alpha*beta).
864 * Like Eifel it also prevents growth of rto,
865 * but also it limits too fast rto decreases,
866 * happening in pure Eifel.
871 m -= (tp->mdev_us >> 2); /* similar update on mdev */
873 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */
874 if (tp->mdev_us > tp->mdev_max_us) {
875 tp->mdev_max_us = tp->mdev_us;
876 if (tp->mdev_max_us > tp->rttvar_us)
877 tp->rttvar_us = tp->mdev_max_us;
879 if (after(tp->snd_una, tp->rtt_seq)) {
880 if (tp->mdev_max_us < tp->rttvar_us)
881 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
882 tp->rtt_seq = tp->snd_nxt;
883 tp->mdev_max_us = tcp_rto_min_us(sk);
888 /* no previous measure. */
889 srtt = m << 3; /* take the measured time to be rtt */
890 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */
891 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
892 tp->mdev_max_us = tp->rttvar_us;
893 tp->rtt_seq = tp->snd_nxt;
897 tp->srtt_us = max(1U, srtt);
900 static void tcp_update_pacing_rate(struct sock *sk)
902 const struct tcp_sock *tp = tcp_sk(sk);
905 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
906 rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3);
908 /* current rate is (cwnd * mss) / srtt
909 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
910 * In Congestion Avoidance phase, set it to 120 % the current rate.
912 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
913 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
914 * end of slow start and should slow down.
916 if (tp->snd_cwnd < tp->snd_ssthresh / 2)
917 rate *= sock_net(sk)->ipv4.sysctl_tcp_pacing_ss_ratio;
919 rate *= sock_net(sk)->ipv4.sysctl_tcp_pacing_ca_ratio;
921 rate *= max(tp->snd_cwnd, tp->packets_out);
923 if (likely(tp->srtt_us))
924 do_div(rate, tp->srtt_us);
926 /* WRITE_ONCE() is needed because sch_fq fetches sk_pacing_rate
927 * without any lock. We want to make sure compiler wont store
928 * intermediate values in this location.
930 WRITE_ONCE(sk->sk_pacing_rate, min_t(u64, rate,
931 sk->sk_max_pacing_rate));
934 /* Calculate rto without backoff. This is the second half of Van Jacobson's
935 * routine referred to above.
937 static void tcp_set_rto(struct sock *sk)
939 const struct tcp_sock *tp = tcp_sk(sk);
940 /* Old crap is replaced with new one. 8)
943 * 1. If rtt variance happened to be less 50msec, it is hallucination.
944 * It cannot be less due to utterly erratic ACK generation made
945 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
946 * to do with delayed acks, because at cwnd>2 true delack timeout
947 * is invisible. Actually, Linux-2.4 also generates erratic
948 * ACKs in some circumstances.
950 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
952 /* 2. Fixups made earlier cannot be right.
953 * If we do not estimate RTO correctly without them,
954 * all the algo is pure shit and should be replaced
955 * with correct one. It is exactly, which we pretend to do.
958 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
959 * guarantees that rto is higher.
964 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
966 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
969 cwnd = TCP_INIT_CWND;
970 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
973 struct tcp_sacktag_state {
974 /* Timestamps for earliest and latest never-retransmitted segment
975 * that was SACKed. RTO needs the earliest RTT to stay conservative,
976 * but congestion control should still get an accurate delay signal.
983 unsigned int mss_now;
984 struct rate_sample *rate;
987 /* Take a notice that peer is sending D-SACKs. Skip update of data delivery
988 * and spurious retransmission information if this DSACK is unlikely caused by
990 * - DSACKed sequence range is larger than maximum receiver's window.
991 * - Total no. of DSACKed segments exceed the total no. of retransmitted segs.
993 static u32 tcp_dsack_seen(struct tcp_sock *tp, u32 start_seq,
994 u32 end_seq, struct tcp_sacktag_state *state)
996 u32 seq_len, dup_segs = 1;
998 if (!before(start_seq, end_seq))
1001 seq_len = end_seq - start_seq;
1002 /* Dubious DSACK: DSACKed range greater than maximum advertised rwnd */
1003 if (seq_len > tp->max_window)
1005 if (seq_len > tp->mss_cache)
1006 dup_segs = DIV_ROUND_UP(seq_len, tp->mss_cache);
1008 tp->dsack_dups += dup_segs;
1009 /* Skip the DSACK if dup segs weren't retransmitted by sender */
1010 if (tp->dsack_dups > tp->total_retrans)
1013 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
1014 tp->rack.dsack_seen = 1;
1016 state->flag |= FLAG_DSACKING_ACK;
1017 /* A spurious retransmission is delivered */
1018 state->sack_delivered += dup_segs;
1023 /* It's reordering when higher sequence was delivered (i.e. sacked) before
1024 * some lower never-retransmitted sequence ("low_seq"). The maximum reordering
1025 * distance is approximated in full-mss packet distance ("reordering").
1027 static void tcp_check_sack_reordering(struct sock *sk, const u32 low_seq,
1030 struct tcp_sock *tp = tcp_sk(sk);
1031 const u32 mss = tp->mss_cache;
1034 fack = tcp_highest_sack_seq(tp);
1035 if (!before(low_seq, fack))
1038 metric = fack - low_seq;
1039 if ((metric > tp->reordering * mss) && mss) {
1040 #if FASTRETRANS_DEBUG > 1
1041 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
1042 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
1046 tp->undo_marker ? tp->undo_retrans : 0);
1048 tp->reordering = min_t(u32, (metric + mss - 1) / mss,
1049 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_max_reordering));
1052 /* This exciting event is worth to be remembered. 8) */
1054 NET_INC_STATS(sock_net(sk),
1055 ts ? LINUX_MIB_TCPTSREORDER : LINUX_MIB_TCPSACKREORDER);
1058 /* This must be called before lost_out or retrans_out are updated
1059 * on a new loss, because we want to know if all skbs previously
1060 * known to be lost have already been retransmitted, indicating
1061 * that this newly lost skb is our next skb to retransmit.
1063 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
1065 if ((!tp->retransmit_skb_hint && tp->retrans_out >= tp->lost_out) ||
1066 (tp->retransmit_skb_hint &&
1067 before(TCP_SKB_CB(skb)->seq,
1068 TCP_SKB_CB(tp->retransmit_skb_hint)->seq)))
1069 tp->retransmit_skb_hint = skb;
1072 /* Sum the number of packets on the wire we have marked as lost, and
1073 * notify the congestion control module that the given skb was marked lost.
1075 static void tcp_notify_skb_loss_event(struct tcp_sock *tp, const struct sk_buff *skb)
1077 tp->lost += tcp_skb_pcount(skb);
1080 void tcp_mark_skb_lost(struct sock *sk, struct sk_buff *skb)
1082 __u8 sacked = TCP_SKB_CB(skb)->sacked;
1083 struct tcp_sock *tp = tcp_sk(sk);
1085 if (sacked & TCPCB_SACKED_ACKED)
1088 tcp_verify_retransmit_hint(tp, skb);
1089 if (sacked & TCPCB_LOST) {
1090 if (sacked & TCPCB_SACKED_RETRANS) {
1091 /* Account for retransmits that are lost again */
1092 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1093 tp->retrans_out -= tcp_skb_pcount(skb);
1094 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT,
1095 tcp_skb_pcount(skb));
1096 tcp_notify_skb_loss_event(tp, skb);
1099 tp->lost_out += tcp_skb_pcount(skb);
1100 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1101 tcp_notify_skb_loss_event(tp, skb);
1105 /* Updates the delivered and delivered_ce counts */
1106 static void tcp_count_delivered(struct tcp_sock *tp, u32 delivered,
1109 tp->delivered += delivered;
1111 tp->delivered_ce += delivered;
1114 /* This procedure tags the retransmission queue when SACKs arrive.
1116 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
1117 * Packets in queue with these bits set are counted in variables
1118 * sacked_out, retrans_out and lost_out, correspondingly.
1120 * Valid combinations are:
1121 * Tag InFlight Description
1122 * 0 1 - orig segment is in flight.
1123 * S 0 - nothing flies, orig reached receiver.
1124 * L 0 - nothing flies, orig lost by net.
1125 * R 2 - both orig and retransmit are in flight.
1126 * L|R 1 - orig is lost, retransmit is in flight.
1127 * S|R 1 - orig reached receiver, retrans is still in flight.
1128 * (L|S|R is logically valid, it could occur when L|R is sacked,
1129 * but it is equivalent to plain S and code short-curcuits it to S.
1130 * L|S is logically invalid, it would mean -1 packet in flight 8))
1132 * These 6 states form finite state machine, controlled by the following events:
1133 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
1134 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
1135 * 3. Loss detection event of two flavors:
1136 * A. Scoreboard estimator decided the packet is lost.
1137 * A'. Reno "three dupacks" marks head of queue lost.
1138 * B. SACK arrives sacking SND.NXT at the moment, when the
1139 * segment was retransmitted.
1140 * 4. D-SACK added new rule: D-SACK changes any tag to S.
1142 * It is pleasant to note, that state diagram turns out to be commutative,
1143 * so that we are allowed not to be bothered by order of our actions,
1144 * when multiple events arrive simultaneously. (see the function below).
1146 * Reordering detection.
1147 * --------------------
1148 * Reordering metric is maximal distance, which a packet can be displaced
1149 * in packet stream. With SACKs we can estimate it:
1151 * 1. SACK fills old hole and the corresponding segment was not
1152 * ever retransmitted -> reordering. Alas, we cannot use it
1153 * when segment was retransmitted.
1154 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1155 * for retransmitted and already SACKed segment -> reordering..
1156 * Both of these heuristics are not used in Loss state, when we cannot
1157 * account for retransmits accurately.
1159 * SACK block validation.
1160 * ----------------------
1162 * SACK block range validation checks that the received SACK block fits to
1163 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1164 * Note that SND.UNA is not included to the range though being valid because
1165 * it means that the receiver is rather inconsistent with itself reporting
1166 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1167 * perfectly valid, however, in light of RFC2018 which explicitly states
1168 * that "SACK block MUST reflect the newest segment. Even if the newest
1169 * segment is going to be discarded ...", not that it looks very clever
1170 * in case of head skb. Due to potentional receiver driven attacks, we
1171 * choose to avoid immediate execution of a walk in write queue due to
1172 * reneging and defer head skb's loss recovery to standard loss recovery
1173 * procedure that will eventually trigger (nothing forbids us doing this).
1175 * Implements also blockage to start_seq wrap-around. Problem lies in the
1176 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1177 * there's no guarantee that it will be before snd_nxt (n). The problem
1178 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1181 * <- outs wnd -> <- wrapzone ->
1182 * u e n u_w e_w s n_w
1184 * |<------------+------+----- TCP seqno space --------------+---------->|
1185 * ...-- <2^31 ->| |<--------...
1186 * ...---- >2^31 ------>| |<--------...
1188 * Current code wouldn't be vulnerable but it's better still to discard such
1189 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1190 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1191 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1192 * equal to the ideal case (infinite seqno space without wrap caused issues).
1194 * With D-SACK the lower bound is extended to cover sequence space below
1195 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1196 * again, D-SACK block must not to go across snd_una (for the same reason as
1197 * for the normal SACK blocks, explained above). But there all simplicity
1198 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1199 * fully below undo_marker they do not affect behavior in anyway and can
1200 * therefore be safely ignored. In rare cases (which are more or less
1201 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1202 * fragmentation and packet reordering past skb's retransmission. To consider
1203 * them correctly, the acceptable range must be extended even more though
1204 * the exact amount is rather hard to quantify. However, tp->max_window can
1205 * be used as an exaggerated estimate.
1207 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
1208 u32 start_seq, u32 end_seq)
1210 /* Too far in future, or reversed (interpretation is ambiguous) */
1211 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1214 /* Nasty start_seq wrap-around check (see comments above) */
1215 if (!before(start_seq, tp->snd_nxt))
1218 /* In outstanding window? ...This is valid exit for D-SACKs too.
1219 * start_seq == snd_una is non-sensical (see comments above)
1221 if (after(start_seq, tp->snd_una))
1224 if (!is_dsack || !tp->undo_marker)
1227 /* ...Then it's D-SACK, and must reside below snd_una completely */
1228 if (after(end_seq, tp->snd_una))
1231 if (!before(start_seq, tp->undo_marker))
1235 if (!after(end_seq, tp->undo_marker))
1238 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1239 * start_seq < undo_marker and end_seq >= undo_marker.
1241 return !before(start_seq, end_seq - tp->max_window);
1244 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1245 struct tcp_sack_block_wire *sp, int num_sacks,
1246 u32 prior_snd_una, struct tcp_sacktag_state *state)
1248 struct tcp_sock *tp = tcp_sk(sk);
1249 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1250 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1253 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1254 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1255 } else if (num_sacks > 1) {
1256 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1257 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1259 if (after(end_seq_0, end_seq_1) || before(start_seq_0, start_seq_1))
1261 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKOFORECV);
1266 dup_segs = tcp_dsack_seen(tp, start_seq_0, end_seq_0, state);
1267 if (!dup_segs) { /* Skip dubious DSACK */
1268 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKIGNOREDDUBIOUS);
1272 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECVSEGS, dup_segs);
1274 /* D-SACK for already forgotten data... Do dumb counting. */
1275 if (tp->undo_marker && tp->undo_retrans > 0 &&
1276 !after(end_seq_0, prior_snd_una) &&
1277 after(end_seq_0, tp->undo_marker))
1278 tp->undo_retrans = max_t(int, 0, tp->undo_retrans - dup_segs);
1283 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1284 * the incoming SACK may not exactly match but we can find smaller MSS
1285 * aligned portion of it that matches. Therefore we might need to fragment
1286 * which may fail and creates some hassle (caller must handle error case
1289 * FIXME: this could be merged to shift decision code
1291 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1292 u32 start_seq, u32 end_seq)
1296 unsigned int pkt_len;
1299 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1300 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1302 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1303 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1304 mss = tcp_skb_mss(skb);
1305 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1308 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1312 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1317 /* Round if necessary so that SACKs cover only full MSSes
1318 * and/or the remaining small portion (if present)
1320 if (pkt_len > mss) {
1321 unsigned int new_len = (pkt_len / mss) * mss;
1322 if (!in_sack && new_len < pkt_len)
1327 if (pkt_len >= skb->len && !in_sack)
1330 err = tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
1331 pkt_len, mss, GFP_ATOMIC);
1339 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1340 static u8 tcp_sacktag_one(struct sock *sk,
1341 struct tcp_sacktag_state *state, u8 sacked,
1342 u32 start_seq, u32 end_seq,
1343 int dup_sack, int pcount,
1346 struct tcp_sock *tp = tcp_sk(sk);
1348 /* Account D-SACK for retransmitted packet. */
1349 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1350 if (tp->undo_marker && tp->undo_retrans > 0 &&
1351 after(end_seq, tp->undo_marker))
1352 tp->undo_retrans = max_t(int, 0, tp->undo_retrans - pcount);
1353 if ((sacked & TCPCB_SACKED_ACKED) &&
1354 before(start_seq, state->reord))
1355 state->reord = start_seq;
1358 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1359 if (!after(end_seq, tp->snd_una))
1362 if (!(sacked & TCPCB_SACKED_ACKED)) {
1363 tcp_rack_advance(tp, sacked, end_seq, xmit_time);
1365 if (sacked & TCPCB_SACKED_RETRANS) {
1366 /* If the segment is not tagged as lost,
1367 * we do not clear RETRANS, believing
1368 * that retransmission is still in flight.
1370 if (sacked & TCPCB_LOST) {
1371 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1372 tp->lost_out -= pcount;
1373 tp->retrans_out -= pcount;
1376 if (!(sacked & TCPCB_RETRANS)) {
1377 /* New sack for not retransmitted frame,
1378 * which was in hole. It is reordering.
1380 if (before(start_seq,
1381 tcp_highest_sack_seq(tp)) &&
1382 before(start_seq, state->reord))
1383 state->reord = start_seq;
1385 if (!after(end_seq, tp->high_seq))
1386 state->flag |= FLAG_ORIG_SACK_ACKED;
1387 if (state->first_sackt == 0)
1388 state->first_sackt = xmit_time;
1389 state->last_sackt = xmit_time;
1392 if (sacked & TCPCB_LOST) {
1393 sacked &= ~TCPCB_LOST;
1394 tp->lost_out -= pcount;
1398 sacked |= TCPCB_SACKED_ACKED;
1399 state->flag |= FLAG_DATA_SACKED;
1400 tp->sacked_out += pcount;
1401 /* Out-of-order packets delivered */
1402 state->sack_delivered += pcount;
1404 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1405 if (tp->lost_skb_hint &&
1406 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1407 tp->lost_cnt_hint += pcount;
1410 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1411 * frames and clear it. undo_retrans is decreased above, L|R frames
1412 * are accounted above as well.
1414 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1415 sacked &= ~TCPCB_SACKED_RETRANS;
1416 tp->retrans_out -= pcount;
1422 /* Shift newly-SACKed bytes from this skb to the immediately previous
1423 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1425 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *prev,
1426 struct sk_buff *skb,
1427 struct tcp_sacktag_state *state,
1428 unsigned int pcount, int shifted, int mss,
1431 struct tcp_sock *tp = tcp_sk(sk);
1432 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1433 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
1437 /* Adjust counters and hints for the newly sacked sequence
1438 * range but discard the return value since prev is already
1439 * marked. We must tag the range first because the seq
1440 * advancement below implicitly advances
1441 * tcp_highest_sack_seq() when skb is highest_sack.
1443 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1444 start_seq, end_seq, dup_sack, pcount,
1445 tcp_skb_timestamp_us(skb));
1446 tcp_rate_skb_delivered(sk, skb, state->rate);
1448 if (skb == tp->lost_skb_hint)
1449 tp->lost_cnt_hint += pcount;
1451 TCP_SKB_CB(prev)->end_seq += shifted;
1452 TCP_SKB_CB(skb)->seq += shifted;
1454 tcp_skb_pcount_add(prev, pcount);
1455 WARN_ON_ONCE(tcp_skb_pcount(skb) < pcount);
1456 tcp_skb_pcount_add(skb, -pcount);
1458 /* When we're adding to gso_segs == 1, gso_size will be zero,
1459 * in theory this shouldn't be necessary but as long as DSACK
1460 * code can come after this skb later on it's better to keep
1461 * setting gso_size to something.
1463 if (!TCP_SKB_CB(prev)->tcp_gso_size)
1464 TCP_SKB_CB(prev)->tcp_gso_size = mss;
1466 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1467 if (tcp_skb_pcount(skb) <= 1)
1468 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1470 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1471 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1474 BUG_ON(!tcp_skb_pcount(skb));
1475 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1479 /* Whole SKB was eaten :-) */
1481 if (skb == tp->retransmit_skb_hint)
1482 tp->retransmit_skb_hint = prev;
1483 if (skb == tp->lost_skb_hint) {
1484 tp->lost_skb_hint = prev;
1485 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1488 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1489 TCP_SKB_CB(prev)->eor = TCP_SKB_CB(skb)->eor;
1490 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1491 TCP_SKB_CB(prev)->end_seq++;
1493 if (skb == tcp_highest_sack(sk))
1494 tcp_advance_highest_sack(sk, skb);
1496 tcp_skb_collapse_tstamp(prev, skb);
1497 if (unlikely(TCP_SKB_CB(prev)->tx.delivered_mstamp))
1498 TCP_SKB_CB(prev)->tx.delivered_mstamp = 0;
1500 tcp_rtx_queue_unlink_and_free(skb, sk);
1502 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKMERGED);
1507 /* I wish gso_size would have a bit more sane initialization than
1508 * something-or-zero which complicates things
1510 static int tcp_skb_seglen(const struct sk_buff *skb)
1512 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1515 /* Shifting pages past head area doesn't work */
1516 static int skb_can_shift(const struct sk_buff *skb)
1518 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1521 int tcp_skb_shift(struct sk_buff *to, struct sk_buff *from,
1522 int pcount, int shiftlen)
1524 /* TCP min gso_size is 8 bytes (TCP_MIN_GSO_SIZE)
1525 * Since TCP_SKB_CB(skb)->tcp_gso_segs is 16 bits, we need
1526 * to make sure not storing more than 65535 * 8 bytes per skb,
1527 * even if current MSS is bigger.
1529 if (unlikely(to->len + shiftlen >= 65535 * TCP_MIN_GSO_SIZE))
1531 if (unlikely(tcp_skb_pcount(to) + pcount > 65535))
1533 return skb_shift(to, from, shiftlen);
1536 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1539 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1540 struct tcp_sacktag_state *state,
1541 u32 start_seq, u32 end_seq,
1544 struct tcp_sock *tp = tcp_sk(sk);
1545 struct sk_buff *prev;
1551 /* Normally R but no L won't result in plain S */
1553 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1555 if (!skb_can_shift(skb))
1557 /* This frame is about to be dropped (was ACKed). */
1558 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1561 /* Can only happen with delayed DSACK + discard craziness */
1562 prev = skb_rb_prev(skb);
1566 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1569 if (!tcp_skb_can_collapse(prev, skb))
1572 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1573 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1577 pcount = tcp_skb_pcount(skb);
1578 mss = tcp_skb_seglen(skb);
1580 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1581 * drop this restriction as unnecessary
1583 if (mss != tcp_skb_seglen(prev))
1586 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1588 /* CHECKME: This is non-MSS split case only?, this will
1589 * cause skipped skbs due to advancing loop btw, original
1590 * has that feature too
1592 if (tcp_skb_pcount(skb) <= 1)
1595 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1597 /* TODO: head merge to next could be attempted here
1598 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1599 * though it might not be worth of the additional hassle
1601 * ...we can probably just fallback to what was done
1602 * previously. We could try merging non-SACKed ones
1603 * as well but it probably isn't going to buy off
1604 * because later SACKs might again split them, and
1605 * it would make skb timestamp tracking considerably
1611 len = end_seq - TCP_SKB_CB(skb)->seq;
1613 BUG_ON(len > skb->len);
1615 /* MSS boundaries should be honoured or else pcount will
1616 * severely break even though it makes things bit trickier.
1617 * Optimize common case to avoid most of the divides
1619 mss = tcp_skb_mss(skb);
1621 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1622 * drop this restriction as unnecessary
1624 if (mss != tcp_skb_seglen(prev))
1629 } else if (len < mss) {
1637 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1638 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1641 if (!tcp_skb_shift(prev, skb, pcount, len))
1643 if (!tcp_shifted_skb(sk, prev, skb, state, pcount, len, mss, dup_sack))
1646 /* Hole filled allows collapsing with the next as well, this is very
1647 * useful when hole on every nth skb pattern happens
1649 skb = skb_rb_next(prev);
1653 if (!skb_can_shift(skb) ||
1654 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1655 (mss != tcp_skb_seglen(skb)))
1658 if (!tcp_skb_can_collapse(prev, skb))
1661 pcount = tcp_skb_pcount(skb);
1662 if (tcp_skb_shift(prev, skb, pcount, len))
1663 tcp_shifted_skb(sk, prev, skb, state, pcount,
1673 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1677 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1678 struct tcp_sack_block *next_dup,
1679 struct tcp_sacktag_state *state,
1680 u32 start_seq, u32 end_seq,
1683 struct tcp_sock *tp = tcp_sk(sk);
1684 struct sk_buff *tmp;
1686 skb_rbtree_walk_from(skb) {
1688 bool dup_sack = dup_sack_in;
1690 /* queue is in-order => we can short-circuit the walk early */
1691 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1695 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1696 in_sack = tcp_match_skb_to_sack(sk, skb,
1697 next_dup->start_seq,
1703 /* skb reference here is a bit tricky to get right, since
1704 * shifting can eat and free both this skb and the next,
1705 * so not even _safe variant of the loop is enough.
1708 tmp = tcp_shift_skb_data(sk, skb, state,
1709 start_seq, end_seq, dup_sack);
1718 in_sack = tcp_match_skb_to_sack(sk, skb,
1724 if (unlikely(in_sack < 0))
1728 TCP_SKB_CB(skb)->sacked =
1731 TCP_SKB_CB(skb)->sacked,
1732 TCP_SKB_CB(skb)->seq,
1733 TCP_SKB_CB(skb)->end_seq,
1735 tcp_skb_pcount(skb),
1736 tcp_skb_timestamp_us(skb));
1737 tcp_rate_skb_delivered(sk, skb, state->rate);
1738 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1739 list_del_init(&skb->tcp_tsorted_anchor);
1741 if (!before(TCP_SKB_CB(skb)->seq,
1742 tcp_highest_sack_seq(tp)))
1743 tcp_advance_highest_sack(sk, skb);
1749 static struct sk_buff *tcp_sacktag_bsearch(struct sock *sk, u32 seq)
1751 struct rb_node *parent, **p = &sk->tcp_rtx_queue.rb_node;
1752 struct sk_buff *skb;
1756 skb = rb_to_skb(parent);
1757 if (before(seq, TCP_SKB_CB(skb)->seq)) {
1758 p = &parent->rb_left;
1761 if (!before(seq, TCP_SKB_CB(skb)->end_seq)) {
1762 p = &parent->rb_right;
1770 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1773 if (skb && after(TCP_SKB_CB(skb)->seq, skip_to_seq))
1776 return tcp_sacktag_bsearch(sk, skip_to_seq);
1779 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1781 struct tcp_sack_block *next_dup,
1782 struct tcp_sacktag_state *state,
1788 if (before(next_dup->start_seq, skip_to_seq)) {
1789 skb = tcp_sacktag_skip(skb, sk, next_dup->start_seq);
1790 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1791 next_dup->start_seq, next_dup->end_seq,
1798 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1800 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1804 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1805 u32 prior_snd_una, struct tcp_sacktag_state *state)
1807 struct tcp_sock *tp = tcp_sk(sk);
1808 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1809 TCP_SKB_CB(ack_skb)->sacked);
1810 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1811 struct tcp_sack_block sp[TCP_NUM_SACKS];
1812 struct tcp_sack_block *cache;
1813 struct sk_buff *skb;
1814 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1816 bool found_dup_sack = false;
1818 int first_sack_index;
1821 state->reord = tp->snd_nxt;
1823 if (!tp->sacked_out)
1824 tcp_highest_sack_reset(sk);
1826 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1827 num_sacks, prior_snd_una, state);
1829 /* Eliminate too old ACKs, but take into
1830 * account more or less fresh ones, they can
1831 * contain valid SACK info.
1833 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1836 if (!tp->packets_out)
1840 first_sack_index = 0;
1841 for (i = 0; i < num_sacks; i++) {
1842 bool dup_sack = !i && found_dup_sack;
1844 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1845 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1847 if (!tcp_is_sackblock_valid(tp, dup_sack,
1848 sp[used_sacks].start_seq,
1849 sp[used_sacks].end_seq)) {
1853 if (!tp->undo_marker)
1854 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1856 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1858 /* Don't count olds caused by ACK reordering */
1859 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1860 !after(sp[used_sacks].end_seq, tp->snd_una))
1862 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1865 NET_INC_STATS(sock_net(sk), mib_idx);
1867 first_sack_index = -1;
1871 /* Ignore very old stuff early */
1872 if (!after(sp[used_sacks].end_seq, prior_snd_una)) {
1874 first_sack_index = -1;
1881 /* order SACK blocks to allow in order walk of the retrans queue */
1882 for (i = used_sacks - 1; i > 0; i--) {
1883 for (j = 0; j < i; j++) {
1884 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1885 swap(sp[j], sp[j + 1]);
1887 /* Track where the first SACK block goes to */
1888 if (j == first_sack_index)
1889 first_sack_index = j + 1;
1894 state->mss_now = tcp_current_mss(sk);
1898 if (!tp->sacked_out) {
1899 /* It's already past, so skip checking against it */
1900 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1902 cache = tp->recv_sack_cache;
1903 /* Skip empty blocks in at head of the cache */
1904 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1909 while (i < used_sacks) {
1910 u32 start_seq = sp[i].start_seq;
1911 u32 end_seq = sp[i].end_seq;
1912 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1913 struct tcp_sack_block *next_dup = NULL;
1915 if (found_dup_sack && ((i + 1) == first_sack_index))
1916 next_dup = &sp[i + 1];
1918 /* Skip too early cached blocks */
1919 while (tcp_sack_cache_ok(tp, cache) &&
1920 !before(start_seq, cache->end_seq))
1923 /* Can skip some work by looking recv_sack_cache? */
1924 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1925 after(end_seq, cache->start_seq)) {
1928 if (before(start_seq, cache->start_seq)) {
1929 skb = tcp_sacktag_skip(skb, sk, start_seq);
1930 skb = tcp_sacktag_walk(skb, sk, next_dup,
1937 /* Rest of the block already fully processed? */
1938 if (!after(end_seq, cache->end_seq))
1941 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1945 /* ...tail remains todo... */
1946 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1947 /* ...but better entrypoint exists! */
1948 skb = tcp_highest_sack(sk);
1955 skb = tcp_sacktag_skip(skb, sk, cache->end_seq);
1956 /* Check overlap against next cached too (past this one already) */
1961 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1962 skb = tcp_highest_sack(sk);
1966 skb = tcp_sacktag_skip(skb, sk, start_seq);
1969 skb = tcp_sacktag_walk(skb, sk, next_dup, state,
1970 start_seq, end_seq, dup_sack);
1976 /* Clear the head of the cache sack blocks so we can skip it next time */
1977 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1978 tp->recv_sack_cache[i].start_seq = 0;
1979 tp->recv_sack_cache[i].end_seq = 0;
1981 for (j = 0; j < used_sacks; j++)
1982 tp->recv_sack_cache[i++] = sp[j];
1984 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Loss || tp->undo_marker)
1985 tcp_check_sack_reordering(sk, state->reord, 0);
1987 tcp_verify_left_out(tp);
1990 #if FASTRETRANS_DEBUG > 0
1991 WARN_ON((int)tp->sacked_out < 0);
1992 WARN_ON((int)tp->lost_out < 0);
1993 WARN_ON((int)tp->retrans_out < 0);
1994 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1999 /* Limits sacked_out so that sum with lost_out isn't ever larger than
2000 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
2002 static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
2006 holes = max(tp->lost_out, 1U);
2007 holes = min(holes, tp->packets_out);
2009 if ((tp->sacked_out + holes) > tp->packets_out) {
2010 tp->sacked_out = tp->packets_out - holes;
2016 /* If we receive more dupacks than we expected counting segments
2017 * in assumption of absent reordering, interpret this as reordering.
2018 * The only another reason could be bug in receiver TCP.
2020 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
2022 struct tcp_sock *tp = tcp_sk(sk);
2024 if (!tcp_limit_reno_sacked(tp))
2027 tp->reordering = min_t(u32, tp->packets_out + addend,
2028 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_max_reordering));
2030 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRENOREORDER);
2033 /* Emulate SACKs for SACKless connection: account for a new dupack. */
2035 static void tcp_add_reno_sack(struct sock *sk, int num_dupack, bool ece_ack)
2038 struct tcp_sock *tp = tcp_sk(sk);
2039 u32 prior_sacked = tp->sacked_out;
2042 tp->sacked_out += num_dupack;
2043 tcp_check_reno_reordering(sk, 0);
2044 delivered = tp->sacked_out - prior_sacked;
2046 tcp_count_delivered(tp, delivered, ece_ack);
2047 tcp_verify_left_out(tp);
2051 /* Account for ACK, ACKing some data in Reno Recovery phase. */
2053 static void tcp_remove_reno_sacks(struct sock *sk, int acked, bool ece_ack)
2055 struct tcp_sock *tp = tcp_sk(sk);
2058 /* One ACK acked hole. The rest eat duplicate ACKs. */
2059 tcp_count_delivered(tp, max_t(int, acked - tp->sacked_out, 1),
2061 if (acked - 1 >= tp->sacked_out)
2064 tp->sacked_out -= acked - 1;
2066 tcp_check_reno_reordering(sk, acked);
2067 tcp_verify_left_out(tp);
2070 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
2075 void tcp_clear_retrans(struct tcp_sock *tp)
2077 tp->retrans_out = 0;
2079 tp->undo_marker = 0;
2080 tp->undo_retrans = -1;
2084 static inline void tcp_init_undo(struct tcp_sock *tp)
2086 tp->undo_marker = tp->snd_una;
2087 /* Retransmission still in flight may cause DSACKs later. */
2088 tp->undo_retrans = tp->retrans_out ? : -1;
2091 static bool tcp_is_rack(const struct sock *sk)
2093 return READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_recovery) &
2094 TCP_RACK_LOSS_DETECTION;
2097 /* If we detect SACK reneging, forget all SACK information
2098 * and reset tags completely, otherwise preserve SACKs. If receiver
2099 * dropped its ofo queue, we will know this due to reneging detection.
2101 static void tcp_timeout_mark_lost(struct sock *sk)
2103 struct tcp_sock *tp = tcp_sk(sk);
2104 struct sk_buff *skb, *head;
2105 bool is_reneg; /* is receiver reneging on SACKs? */
2107 head = tcp_rtx_queue_head(sk);
2108 is_reneg = head && (TCP_SKB_CB(head)->sacked & TCPCB_SACKED_ACKED);
2110 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
2112 /* Mark SACK reneging until we recover from this loss event. */
2113 tp->is_sack_reneg = 1;
2114 } else if (tcp_is_reno(tp)) {
2115 tcp_reset_reno_sack(tp);
2119 skb_rbtree_walk_from(skb) {
2121 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
2122 else if (tcp_is_rack(sk) && skb != head &&
2123 tcp_rack_skb_timeout(tp, skb, 0) > 0)
2124 continue; /* Don't mark recently sent ones lost yet */
2125 tcp_mark_skb_lost(sk, skb);
2127 tcp_verify_left_out(tp);
2128 tcp_clear_all_retrans_hints(tp);
2131 /* Enter Loss state. */
2132 void tcp_enter_loss(struct sock *sk)
2134 const struct inet_connection_sock *icsk = inet_csk(sk);
2135 struct tcp_sock *tp = tcp_sk(sk);
2136 struct net *net = sock_net(sk);
2137 bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery;
2140 tcp_timeout_mark_lost(sk);
2142 /* Reduce ssthresh if it has not yet been made inside this window. */
2143 if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
2144 !after(tp->high_seq, tp->snd_una) ||
2145 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
2146 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2147 tp->prior_cwnd = tp->snd_cwnd;
2148 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2149 tcp_ca_event(sk, CA_EVENT_LOSS);
2152 tp->snd_cwnd = tcp_packets_in_flight(tp) + 1;
2153 tp->snd_cwnd_cnt = 0;
2154 tp->snd_cwnd_stamp = tcp_jiffies32;
2156 /* Timeout in disordered state after receiving substantial DUPACKs
2157 * suggests that the degree of reordering is over-estimated.
2159 reordering = READ_ONCE(net->ipv4.sysctl_tcp_reordering);
2160 if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
2161 tp->sacked_out >= reordering)
2162 tp->reordering = min_t(unsigned int, tp->reordering,
2165 tcp_set_ca_state(sk, TCP_CA_Loss);
2166 tp->high_seq = tp->snd_nxt;
2167 tcp_ecn_queue_cwr(tp);
2169 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
2170 * loss recovery is underway except recurring timeout(s) on
2171 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
2173 tp->frto = READ_ONCE(net->ipv4.sysctl_tcp_frto) &&
2174 (new_recovery || icsk->icsk_retransmits) &&
2175 !inet_csk(sk)->icsk_mtup.probe_size;
2178 /* If ACK arrived pointing to a remembered SACK, it means that our
2179 * remembered SACKs do not reflect real state of receiver i.e.
2180 * receiver _host_ is heavily congested (or buggy).
2182 * To avoid big spurious retransmission bursts due to transient SACK
2183 * scoreboard oddities that look like reneging, we give the receiver a
2184 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
2185 * restore sanity to the SACK scoreboard. If the apparent reneging
2186 * persists until this RTO then we'll clear the SACK scoreboard.
2188 static bool tcp_check_sack_reneging(struct sock *sk, int *ack_flag)
2190 if (*ack_flag & FLAG_SACK_RENEGING &&
2191 *ack_flag & FLAG_SND_UNA_ADVANCED) {
2192 struct tcp_sock *tp = tcp_sk(sk);
2193 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4),
2194 msecs_to_jiffies(10));
2196 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2197 delay, TCP_RTO_MAX);
2198 *ack_flag &= ~FLAG_SET_XMIT_TIMER;
2204 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2205 * counter when SACK is enabled (without SACK, sacked_out is used for
2208 * With reordering, holes may still be in flight, so RFC3517 recovery
2209 * uses pure sacked_out (total number of SACKed segments) even though
2210 * it violates the RFC that uses duplicate ACKs, often these are equal
2211 * but when e.g. out-of-window ACKs or packet duplication occurs,
2212 * they differ. Since neither occurs due to loss, TCP should really
2215 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2217 return tp->sacked_out + 1;
2220 /* Linux NewReno/SACK/ECN state machine.
2221 * --------------------------------------
2223 * "Open" Normal state, no dubious events, fast path.
2224 * "Disorder" In all the respects it is "Open",
2225 * but requires a bit more attention. It is entered when
2226 * we see some SACKs or dupacks. It is split of "Open"
2227 * mainly to move some processing from fast path to slow one.
2228 * "CWR" CWND was reduced due to some Congestion Notification event.
2229 * It can be ECN, ICMP source quench, local device congestion.
2230 * "Recovery" CWND was reduced, we are fast-retransmitting.
2231 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2233 * tcp_fastretrans_alert() is entered:
2234 * - each incoming ACK, if state is not "Open"
2235 * - when arrived ACK is unusual, namely:
2240 * Counting packets in flight is pretty simple.
2242 * in_flight = packets_out - left_out + retrans_out
2244 * packets_out is SND.NXT-SND.UNA counted in packets.
2246 * retrans_out is number of retransmitted segments.
2248 * left_out is number of segments left network, but not ACKed yet.
2250 * left_out = sacked_out + lost_out
2252 * sacked_out: Packets, which arrived to receiver out of order
2253 * and hence not ACKed. With SACKs this number is simply
2254 * amount of SACKed data. Even without SACKs
2255 * it is easy to give pretty reliable estimate of this number,
2256 * counting duplicate ACKs.
2258 * lost_out: Packets lost by network. TCP has no explicit
2259 * "loss notification" feedback from network (for now).
2260 * It means that this number can be only _guessed_.
2261 * Actually, it is the heuristics to predict lossage that
2262 * distinguishes different algorithms.
2264 * F.e. after RTO, when all the queue is considered as lost,
2265 * lost_out = packets_out and in_flight = retrans_out.
2267 * Essentially, we have now a few algorithms detecting
2270 * If the receiver supports SACK:
2272 * RFC6675/3517: It is the conventional algorithm. A packet is
2273 * considered lost if the number of higher sequence packets
2274 * SACKed is greater than or equal the DUPACK thoreshold
2275 * (reordering). This is implemented in tcp_mark_head_lost and
2276 * tcp_update_scoreboard.
2278 * RACK (draft-ietf-tcpm-rack-01): it is a newer algorithm
2279 * (2017-) that checks timing instead of counting DUPACKs.
2280 * Essentially a packet is considered lost if it's not S/ACKed
2281 * after RTT + reordering_window, where both metrics are
2282 * dynamically measured and adjusted. This is implemented in
2283 * tcp_rack_mark_lost.
2285 * If the receiver does not support SACK:
2287 * NewReno (RFC6582): in Recovery we assume that one segment
2288 * is lost (classic Reno). While we are in Recovery and
2289 * a partial ACK arrives, we assume that one more packet
2290 * is lost (NewReno). This heuristics are the same in NewReno
2293 * Really tricky (and requiring careful tuning) part of algorithm
2294 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2295 * The first determines the moment _when_ we should reduce CWND and,
2296 * hence, slow down forward transmission. In fact, it determines the moment
2297 * when we decide that hole is caused by loss, rather than by a reorder.
2299 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2300 * holes, caused by lost packets.
2302 * And the most logically complicated part of algorithm is undo
2303 * heuristics. We detect false retransmits due to both too early
2304 * fast retransmit (reordering) and underestimated RTO, analyzing
2305 * timestamps and D-SACKs. When we detect that some segments were
2306 * retransmitted by mistake and CWND reduction was wrong, we undo
2307 * window reduction and abort recovery phase. This logic is hidden
2308 * inside several functions named tcp_try_undo_<something>.
2311 /* This function decides, when we should leave Disordered state
2312 * and enter Recovery phase, reducing congestion window.
2314 * Main question: may we further continue forward transmission
2315 * with the same cwnd?
2317 static bool tcp_time_to_recover(struct sock *sk, int flag)
2319 struct tcp_sock *tp = tcp_sk(sk);
2321 /* Trick#1: The loss is proven. */
2325 /* Not-A-Trick#2 : Classic rule... */
2326 if (!tcp_is_rack(sk) && tcp_dupack_heuristics(tp) > tp->reordering)
2332 /* Detect loss in event "A" above by marking head of queue up as lost.
2333 * For RFC3517 SACK, a segment is considered lost if it
2334 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2335 * the maximum SACKed segments to pass before reaching this limit.
2337 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2339 struct tcp_sock *tp = tcp_sk(sk);
2340 struct sk_buff *skb;
2342 /* Use SACK to deduce losses of new sequences sent during recovery */
2343 const u32 loss_high = tp->snd_nxt;
2345 WARN_ON(packets > tp->packets_out);
2346 skb = tp->lost_skb_hint;
2348 /* Head already handled? */
2349 if (mark_head && after(TCP_SKB_CB(skb)->seq, tp->snd_una))
2351 cnt = tp->lost_cnt_hint;
2353 skb = tcp_rtx_queue_head(sk);
2357 skb_rbtree_walk_from(skb) {
2358 /* TODO: do this better */
2359 /* this is not the most efficient way to do this... */
2360 tp->lost_skb_hint = skb;
2361 tp->lost_cnt_hint = cnt;
2363 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2366 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2367 cnt += tcp_skb_pcount(skb);
2372 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_LOST))
2373 tcp_mark_skb_lost(sk, skb);
2378 tcp_verify_left_out(tp);
2381 /* Account newly detected lost packet(s) */
2383 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2385 struct tcp_sock *tp = tcp_sk(sk);
2387 if (tcp_is_sack(tp)) {
2388 int sacked_upto = tp->sacked_out - tp->reordering;
2389 if (sacked_upto >= 0)
2390 tcp_mark_head_lost(sk, sacked_upto, 0);
2391 else if (fast_rexmit)
2392 tcp_mark_head_lost(sk, 1, 1);
2396 static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when)
2398 return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2399 before(tp->rx_opt.rcv_tsecr, when);
2402 /* skb is spurious retransmitted if the returned timestamp echo
2403 * reply is prior to the skb transmission time
2405 static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp,
2406 const struct sk_buff *skb)
2408 return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) &&
2409 tcp_tsopt_ecr_before(tp, tcp_skb_timestamp(skb));
2412 /* Nothing was retransmitted or returned timestamp is less
2413 * than timestamp of the first retransmission.
2415 static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2417 return tp->retrans_stamp &&
2418 tcp_tsopt_ecr_before(tp, tp->retrans_stamp);
2421 /* Undo procedures. */
2423 /* We can clear retrans_stamp when there are no retransmissions in the
2424 * window. It would seem that it is trivially available for us in
2425 * tp->retrans_out, however, that kind of assumptions doesn't consider
2426 * what will happen if errors occur when sending retransmission for the
2427 * second time. ...It could the that such segment has only
2428 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2429 * the head skb is enough except for some reneging corner cases that
2430 * are not worth the effort.
2432 * Main reason for all this complexity is the fact that connection dying
2433 * time now depends on the validity of the retrans_stamp, in particular,
2434 * that successive retransmissions of a segment must not advance
2435 * retrans_stamp under any conditions.
2437 static bool tcp_any_retrans_done(const struct sock *sk)
2439 const struct tcp_sock *tp = tcp_sk(sk);
2440 struct sk_buff *skb;
2442 if (tp->retrans_out)
2445 skb = tcp_rtx_queue_head(sk);
2446 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2452 static void DBGUNDO(struct sock *sk, const char *msg)
2454 #if FASTRETRANS_DEBUG > 1
2455 struct tcp_sock *tp = tcp_sk(sk);
2456 struct inet_sock *inet = inet_sk(sk);
2458 if (sk->sk_family == AF_INET) {
2459 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2461 &inet->inet_daddr, ntohs(inet->inet_dport),
2462 tp->snd_cwnd, tcp_left_out(tp),
2463 tp->snd_ssthresh, tp->prior_ssthresh,
2466 #if IS_ENABLED(CONFIG_IPV6)
2467 else if (sk->sk_family == AF_INET6) {
2468 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2470 &sk->sk_v6_daddr, ntohs(inet->inet_dport),
2471 tp->snd_cwnd, tcp_left_out(tp),
2472 tp->snd_ssthresh, tp->prior_ssthresh,
2479 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
2481 struct tcp_sock *tp = tcp_sk(sk);
2484 struct sk_buff *skb;
2486 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2487 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2490 tcp_clear_all_retrans_hints(tp);
2493 if (tp->prior_ssthresh) {
2494 const struct inet_connection_sock *icsk = inet_csk(sk);
2496 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2498 if (tp->prior_ssthresh > tp->snd_ssthresh) {
2499 tp->snd_ssthresh = tp->prior_ssthresh;
2500 tcp_ecn_withdraw_cwr(tp);
2503 tp->snd_cwnd_stamp = tcp_jiffies32;
2504 tp->undo_marker = 0;
2505 tp->rack.advanced = 1; /* Force RACK to re-exam losses */
2508 static inline bool tcp_may_undo(const struct tcp_sock *tp)
2510 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2513 static bool tcp_is_non_sack_preventing_reopen(struct sock *sk)
2515 struct tcp_sock *tp = tcp_sk(sk);
2517 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2518 /* Hold old state until something *above* high_seq
2519 * is ACKed. For Reno it is MUST to prevent false
2520 * fast retransmits (RFC2582). SACK TCP is safe. */
2521 if (!tcp_any_retrans_done(sk))
2522 tp->retrans_stamp = 0;
2528 /* People celebrate: "We love our President!" */
2529 static bool tcp_try_undo_recovery(struct sock *sk)
2531 struct tcp_sock *tp = tcp_sk(sk);
2533 if (tcp_may_undo(tp)) {
2536 /* Happy end! We did not retransmit anything
2537 * or our original transmission succeeded.
2539 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2540 tcp_undo_cwnd_reduction(sk, false);
2541 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2542 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2544 mib_idx = LINUX_MIB_TCPFULLUNDO;
2546 NET_INC_STATS(sock_net(sk), mib_idx);
2547 } else if (tp->rack.reo_wnd_persist) {
2548 tp->rack.reo_wnd_persist--;
2550 if (tcp_is_non_sack_preventing_reopen(sk))
2552 tcp_set_ca_state(sk, TCP_CA_Open);
2553 tp->is_sack_reneg = 0;
2557 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2558 static bool tcp_try_undo_dsack(struct sock *sk)
2560 struct tcp_sock *tp = tcp_sk(sk);
2562 if (tp->undo_marker && !tp->undo_retrans) {
2563 tp->rack.reo_wnd_persist = min(TCP_RACK_RECOVERY_THRESH,
2564 tp->rack.reo_wnd_persist + 1);
2565 DBGUNDO(sk, "D-SACK");
2566 tcp_undo_cwnd_reduction(sk, false);
2567 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2573 /* Undo during loss recovery after partial ACK or using F-RTO. */
2574 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2576 struct tcp_sock *tp = tcp_sk(sk);
2578 if (frto_undo || tcp_may_undo(tp)) {
2579 tcp_undo_cwnd_reduction(sk, true);
2581 DBGUNDO(sk, "partial loss");
2582 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2584 NET_INC_STATS(sock_net(sk),
2585 LINUX_MIB_TCPSPURIOUSRTOS);
2586 inet_csk(sk)->icsk_retransmits = 0;
2587 if (tcp_is_non_sack_preventing_reopen(sk))
2589 if (frto_undo || tcp_is_sack(tp)) {
2590 tcp_set_ca_state(sk, TCP_CA_Open);
2591 tp->is_sack_reneg = 0;
2598 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
2599 * It computes the number of packets to send (sndcnt) based on packets newly
2601 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2602 * cwnd reductions across a full RTT.
2603 * 2) Otherwise PRR uses packet conservation to send as much as delivered.
2604 * But when the retransmits are acked without further losses, PRR
2605 * slow starts cwnd up to ssthresh to speed up the recovery.
2607 static void tcp_init_cwnd_reduction(struct sock *sk)
2609 struct tcp_sock *tp = tcp_sk(sk);
2611 tp->high_seq = tp->snd_nxt;
2612 tp->tlp_high_seq = 0;
2613 tp->snd_cwnd_cnt = 0;
2614 tp->prior_cwnd = tp->snd_cwnd;
2615 tp->prr_delivered = 0;
2617 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2618 tcp_ecn_queue_cwr(tp);
2621 void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked, int flag)
2623 struct tcp_sock *tp = tcp_sk(sk);
2625 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2627 if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd))
2630 tp->prr_delivered += newly_acked_sacked;
2632 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2634 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2635 } else if ((flag & (FLAG_RETRANS_DATA_ACKED | FLAG_LOST_RETRANS)) ==
2636 FLAG_RETRANS_DATA_ACKED) {
2637 sndcnt = min_t(int, delta,
2638 max_t(int, tp->prr_delivered - tp->prr_out,
2639 newly_acked_sacked) + 1);
2641 sndcnt = min(delta, newly_acked_sacked);
2643 /* Force a fast retransmit upon entering fast recovery */
2644 sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1));
2645 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2648 static inline void tcp_end_cwnd_reduction(struct sock *sk)
2650 struct tcp_sock *tp = tcp_sk(sk);
2652 if (inet_csk(sk)->icsk_ca_ops->cong_control)
2655 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2656 if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH &&
2657 (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR || tp->undo_marker)) {
2658 tp->snd_cwnd = tp->snd_ssthresh;
2659 tp->snd_cwnd_stamp = tcp_jiffies32;
2661 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2664 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2665 void tcp_enter_cwr(struct sock *sk)
2667 struct tcp_sock *tp = tcp_sk(sk);
2669 tp->prior_ssthresh = 0;
2670 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2671 tp->undo_marker = 0;
2672 tcp_init_cwnd_reduction(sk);
2673 tcp_set_ca_state(sk, TCP_CA_CWR);
2676 EXPORT_SYMBOL(tcp_enter_cwr);
2678 static void tcp_try_keep_open(struct sock *sk)
2680 struct tcp_sock *tp = tcp_sk(sk);
2681 int state = TCP_CA_Open;
2683 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2684 state = TCP_CA_Disorder;
2686 if (inet_csk(sk)->icsk_ca_state != state) {
2687 tcp_set_ca_state(sk, state);
2688 tp->high_seq = tp->snd_nxt;
2692 static void tcp_try_to_open(struct sock *sk, int flag)
2694 struct tcp_sock *tp = tcp_sk(sk);
2696 tcp_verify_left_out(tp);
2698 if (!tcp_any_retrans_done(sk))
2699 tp->retrans_stamp = 0;
2701 if (flag & FLAG_ECE)
2704 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2705 tcp_try_keep_open(sk);
2709 static void tcp_mtup_probe_failed(struct sock *sk)
2711 struct inet_connection_sock *icsk = inet_csk(sk);
2713 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2714 icsk->icsk_mtup.probe_size = 0;
2715 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPFAIL);
2718 static void tcp_mtup_probe_success(struct sock *sk)
2720 struct tcp_sock *tp = tcp_sk(sk);
2721 struct inet_connection_sock *icsk = inet_csk(sk);
2724 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2726 val = (u64)tp->snd_cwnd * tcp_mss_to_mtu(sk, tp->mss_cache);
2727 do_div(val, icsk->icsk_mtup.probe_size);
2728 WARN_ON_ONCE((u32)val != val);
2729 tp->snd_cwnd = max_t(u32, 1U, val);
2731 tp->snd_cwnd_cnt = 0;
2732 tp->snd_cwnd_stamp = tcp_jiffies32;
2733 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2735 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2736 icsk->icsk_mtup.probe_size = 0;
2737 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2738 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS);
2741 /* Do a simple retransmit without using the backoff mechanisms in
2742 * tcp_timer. This is used for path mtu discovery.
2743 * The socket is already locked here.
2745 void tcp_simple_retransmit(struct sock *sk)
2747 const struct inet_connection_sock *icsk = inet_csk(sk);
2748 struct tcp_sock *tp = tcp_sk(sk);
2749 struct sk_buff *skb;
2750 unsigned int mss = tcp_current_mss(sk);
2752 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2753 if (tcp_skb_seglen(skb) > mss)
2754 tcp_mark_skb_lost(sk, skb);
2757 tcp_clear_retrans_hints_partial(tp);
2762 if (tcp_is_reno(tp))
2763 tcp_limit_reno_sacked(tp);
2765 tcp_verify_left_out(tp);
2767 /* Don't muck with the congestion window here.
2768 * Reason is that we do not increase amount of _data_
2769 * in network, but units changed and effective
2770 * cwnd/ssthresh really reduced now.
2772 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2773 tp->high_seq = tp->snd_nxt;
2774 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2775 tp->prior_ssthresh = 0;
2776 tp->undo_marker = 0;
2777 tcp_set_ca_state(sk, TCP_CA_Loss);
2779 tcp_xmit_retransmit_queue(sk);
2781 EXPORT_SYMBOL(tcp_simple_retransmit);
2783 void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2785 struct tcp_sock *tp = tcp_sk(sk);
2788 if (tcp_is_reno(tp))
2789 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2791 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2793 NET_INC_STATS(sock_net(sk), mib_idx);
2795 tp->prior_ssthresh = 0;
2798 if (!tcp_in_cwnd_reduction(sk)) {
2800 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2801 tcp_init_cwnd_reduction(sk);
2803 tcp_set_ca_state(sk, TCP_CA_Recovery);
2806 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2807 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2809 static void tcp_process_loss(struct sock *sk, int flag, int num_dupack,
2812 struct tcp_sock *tp = tcp_sk(sk);
2813 bool recovered = !before(tp->snd_una, tp->high_seq);
2815 if ((flag & FLAG_SND_UNA_ADVANCED || rcu_access_pointer(tp->fastopen_rsk)) &&
2816 tcp_try_undo_loss(sk, false))
2819 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2820 /* Step 3.b. A timeout is spurious if not all data are
2821 * lost, i.e., never-retransmitted data are (s)acked.
2823 if ((flag & FLAG_ORIG_SACK_ACKED) &&
2824 tcp_try_undo_loss(sk, true))
2827 if (after(tp->snd_nxt, tp->high_seq)) {
2828 if (flag & FLAG_DATA_SACKED || num_dupack)
2829 tp->frto = 0; /* Step 3.a. loss was real */
2830 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2831 tp->high_seq = tp->snd_nxt;
2832 /* Step 2.b. Try send new data (but deferred until cwnd
2833 * is updated in tcp_ack()). Otherwise fall back to
2834 * the conventional recovery.
2836 if (!tcp_write_queue_empty(sk) &&
2837 after(tcp_wnd_end(tp), tp->snd_nxt)) {
2838 *rexmit = REXMIT_NEW;
2846 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2847 tcp_try_undo_recovery(sk);
2850 if (tcp_is_reno(tp)) {
2851 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2852 * delivered. Lower inflight to clock out (re)tranmissions.
2854 if (after(tp->snd_nxt, tp->high_seq) && num_dupack)
2855 tcp_add_reno_sack(sk, num_dupack, flag & FLAG_ECE);
2856 else if (flag & FLAG_SND_UNA_ADVANCED)
2857 tcp_reset_reno_sack(tp);
2859 *rexmit = REXMIT_LOST;
2862 static bool tcp_force_fast_retransmit(struct sock *sk)
2864 struct tcp_sock *tp = tcp_sk(sk);
2866 return after(tcp_highest_sack_seq(tp),
2867 tp->snd_una + tp->reordering * tp->mss_cache);
2870 /* Undo during fast recovery after partial ACK. */
2871 static bool tcp_try_undo_partial(struct sock *sk, u32 prior_snd_una,
2874 struct tcp_sock *tp = tcp_sk(sk);
2876 if (tp->undo_marker && tcp_packet_delayed(tp)) {
2877 /* Plain luck! Hole if filled with delayed
2878 * packet, rather than with a retransmit. Check reordering.
2880 tcp_check_sack_reordering(sk, prior_snd_una, 1);
2882 /* We are getting evidence that the reordering degree is higher
2883 * than we realized. If there are no retransmits out then we
2884 * can undo. Otherwise we clock out new packets but do not
2885 * mark more packets lost or retransmit more.
2887 if (tp->retrans_out)
2890 if (!tcp_any_retrans_done(sk))
2891 tp->retrans_stamp = 0;
2893 DBGUNDO(sk, "partial recovery");
2894 tcp_undo_cwnd_reduction(sk, true);
2895 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2896 tcp_try_keep_open(sk);
2898 /* Partial ACK arrived. Force fast retransmit. */
2899 *do_lost = tcp_force_fast_retransmit(sk);
2904 static void tcp_identify_packet_loss(struct sock *sk, int *ack_flag)
2906 struct tcp_sock *tp = tcp_sk(sk);
2908 if (tcp_rtx_queue_empty(sk))
2911 if (unlikely(tcp_is_reno(tp))) {
2912 tcp_newreno_mark_lost(sk, *ack_flag & FLAG_SND_UNA_ADVANCED);
2913 } else if (tcp_is_rack(sk)) {
2914 u32 prior_retrans = tp->retrans_out;
2916 if (tcp_rack_mark_lost(sk))
2917 *ack_flag &= ~FLAG_SET_XMIT_TIMER;
2918 if (prior_retrans > tp->retrans_out)
2919 *ack_flag |= FLAG_LOST_RETRANS;
2923 /* Process an event, which can update packets-in-flight not trivially.
2924 * Main goal of this function is to calculate new estimate for left_out,
2925 * taking into account both packets sitting in receiver's buffer and
2926 * packets lost by network.
2928 * Besides that it updates the congestion state when packet loss or ECN
2929 * is detected. But it does not reduce the cwnd, it is done by the
2930 * congestion control later.
2932 * It does _not_ decide what to send, it is made in function
2933 * tcp_xmit_retransmit_queue().
2935 static void tcp_fastretrans_alert(struct sock *sk, const u32 prior_snd_una,
2936 int num_dupack, int *ack_flag, int *rexmit)
2938 struct inet_connection_sock *icsk = inet_csk(sk);
2939 struct tcp_sock *tp = tcp_sk(sk);
2940 int fast_rexmit = 0, flag = *ack_flag;
2941 bool ece_ack = flag & FLAG_ECE;
2942 bool do_lost = num_dupack || ((flag & FLAG_DATA_SACKED) &&
2943 tcp_force_fast_retransmit(sk));
2945 if (!tp->packets_out && tp->sacked_out)
2948 /* Now state machine starts.
2949 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2951 tp->prior_ssthresh = 0;
2953 /* B. In all the states check for reneging SACKs. */
2954 if (tcp_check_sack_reneging(sk, ack_flag))
2957 /* C. Check consistency of the current state. */
2958 tcp_verify_left_out(tp);
2960 /* D. Check state exit conditions. State can be terminated
2961 * when high_seq is ACKed. */
2962 if (icsk->icsk_ca_state == TCP_CA_Open) {
2963 WARN_ON(tp->retrans_out != 0);
2964 tp->retrans_stamp = 0;
2965 } else if (!before(tp->snd_una, tp->high_seq)) {
2966 switch (icsk->icsk_ca_state) {
2968 /* CWR is to be held something *above* high_seq
2969 * is ACKed for CWR bit to reach receiver. */
2970 if (tp->snd_una != tp->high_seq) {
2971 tcp_end_cwnd_reduction(sk);
2972 tcp_set_ca_state(sk, TCP_CA_Open);
2976 case TCP_CA_Recovery:
2977 if (tcp_is_reno(tp))
2978 tcp_reset_reno_sack(tp);
2979 if (tcp_try_undo_recovery(sk))
2981 tcp_end_cwnd_reduction(sk);
2986 /* E. Process state. */
2987 switch (icsk->icsk_ca_state) {
2988 case TCP_CA_Recovery:
2989 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2990 if (tcp_is_reno(tp))
2991 tcp_add_reno_sack(sk, num_dupack, ece_ack);
2992 } else if (tcp_try_undo_partial(sk, prior_snd_una, &do_lost))
2995 if (tcp_try_undo_dsack(sk))
2996 tcp_try_keep_open(sk);
2998 tcp_identify_packet_loss(sk, ack_flag);
2999 if (icsk->icsk_ca_state != TCP_CA_Recovery) {
3000 if (!tcp_time_to_recover(sk, flag))
3002 /* Undo reverts the recovery state. If loss is evident,
3003 * starts a new recovery (e.g. reordering then loss);
3005 tcp_enter_recovery(sk, ece_ack);
3009 tcp_process_loss(sk, flag, num_dupack, rexmit);
3010 tcp_identify_packet_loss(sk, ack_flag);
3011 if (!(icsk->icsk_ca_state == TCP_CA_Open ||
3012 (*ack_flag & FLAG_LOST_RETRANS)))
3014 /* Change state if cwnd is undone or retransmits are lost */
3017 if (tcp_is_reno(tp)) {
3018 if (flag & FLAG_SND_UNA_ADVANCED)
3019 tcp_reset_reno_sack(tp);
3020 tcp_add_reno_sack(sk, num_dupack, ece_ack);
3023 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
3024 tcp_try_undo_dsack(sk);
3026 tcp_identify_packet_loss(sk, ack_flag);
3027 if (!tcp_time_to_recover(sk, flag)) {
3028 tcp_try_to_open(sk, flag);
3032 /* MTU probe failure: don't reduce cwnd */
3033 if (icsk->icsk_ca_state < TCP_CA_CWR &&
3034 icsk->icsk_mtup.probe_size &&
3035 tp->snd_una == tp->mtu_probe.probe_seq_start) {
3036 tcp_mtup_probe_failed(sk);
3037 /* Restores the reduction we did in tcp_mtup_probe() */
3039 tcp_simple_retransmit(sk);
3043 /* Otherwise enter Recovery state */
3044 tcp_enter_recovery(sk, ece_ack);
3048 if (!tcp_is_rack(sk) && do_lost)
3049 tcp_update_scoreboard(sk, fast_rexmit);
3050 *rexmit = REXMIT_LOST;
3053 static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us, const int flag)
3055 u32 wlen = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_rtt_wlen) * HZ;
3056 struct tcp_sock *tp = tcp_sk(sk);
3058 if ((flag & FLAG_ACK_MAYBE_DELAYED) && rtt_us > tcp_min_rtt(tp)) {
3059 /* If the remote keeps returning delayed ACKs, eventually
3060 * the min filter would pick it up and overestimate the
3061 * prop. delay when it expires. Skip suspected delayed ACKs.
3065 minmax_running_min(&tp->rtt_min, wlen, tcp_jiffies32,
3066 rtt_us ? : jiffies_to_usecs(1));
3069 static bool tcp_ack_update_rtt(struct sock *sk, const int flag,
3070 long seq_rtt_us, long sack_rtt_us,
3071 long ca_rtt_us, struct rate_sample *rs)
3073 const struct tcp_sock *tp = tcp_sk(sk);
3075 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
3076 * broken middle-boxes or peers may corrupt TS-ECR fields. But
3077 * Karn's algorithm forbids taking RTT if some retransmitted data
3078 * is acked (RFC6298).
3081 seq_rtt_us = sack_rtt_us;
3083 /* RTTM Rule: A TSecr value received in a segment is used to
3084 * update the averaged RTT measurement only if the segment
3085 * acknowledges some new data, i.e., only if it advances the
3086 * left edge of the send window.
3087 * See draft-ietf-tcplw-high-performance-00, section 3.3.
3089 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
3090 flag & FLAG_ACKED) {
3091 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr;
3093 if (likely(delta < INT_MAX / (USEC_PER_SEC / TCP_TS_HZ))) {
3096 seq_rtt_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
3097 ca_rtt_us = seq_rtt_us;
3100 rs->rtt_us = ca_rtt_us; /* RTT of last (S)ACKed packet (or -1) */
3104 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
3105 * always taken together with ACK, SACK, or TS-opts. Any negative
3106 * values will be skipped with the seq_rtt_us < 0 check above.
3108 tcp_update_rtt_min(sk, ca_rtt_us, flag);
3109 tcp_rtt_estimator(sk, seq_rtt_us);
3112 /* RFC6298: only reset backoff on valid RTT measurement. */
3113 inet_csk(sk)->icsk_backoff = 0;
3117 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
3118 void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req)
3120 struct rate_sample rs;
3123 if (req && !req->num_retrans && tcp_rsk(req)->snt_synack)
3124 rtt_us = tcp_stamp_us_delta(tcp_clock_us(), tcp_rsk(req)->snt_synack);
3126 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us, &rs);
3130 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
3132 const struct inet_connection_sock *icsk = inet_csk(sk);
3134 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
3135 tcp_sk(sk)->snd_cwnd_stamp = tcp_jiffies32;
3138 /* Restart timer after forward progress on connection.
3139 * RFC2988 recommends to restart timer to now+rto.
3141 void tcp_rearm_rto(struct sock *sk)
3143 const struct inet_connection_sock *icsk = inet_csk(sk);
3144 struct tcp_sock *tp = tcp_sk(sk);
3146 /* If the retrans timer is currently being used by Fast Open
3147 * for SYN-ACK retrans purpose, stay put.
3149 if (rcu_access_pointer(tp->fastopen_rsk))
3152 if (!tp->packets_out) {
3153 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
3155 u32 rto = inet_csk(sk)->icsk_rto;
3156 /* Offset the time elapsed after installing regular RTO */
3157 if (icsk->icsk_pending == ICSK_TIME_REO_TIMEOUT ||
3158 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
3159 s64 delta_us = tcp_rto_delta_us(sk);
3160 /* delta_us may not be positive if the socket is locked
3161 * when the retrans timer fires and is rescheduled.
3163 rto = usecs_to_jiffies(max_t(int, delta_us, 1));
3165 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
3170 /* Try to schedule a loss probe; if that doesn't work, then schedule an RTO. */
3171 static void tcp_set_xmit_timer(struct sock *sk)
3173 if (!tcp_schedule_loss_probe(sk, true))
3177 /* If we get here, the whole TSO packet has not been acked. */
3178 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3180 struct tcp_sock *tp = tcp_sk(sk);
3183 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3185 packets_acked = tcp_skb_pcount(skb);
3186 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3188 packets_acked -= tcp_skb_pcount(skb);
3190 if (packets_acked) {
3191 BUG_ON(tcp_skb_pcount(skb) == 0);
3192 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3195 return packets_acked;
3198 static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
3201 const struct skb_shared_info *shinfo;
3203 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3204 if (likely(!TCP_SKB_CB(skb)->txstamp_ack))
3207 shinfo = skb_shinfo(skb);
3208 if (!before(shinfo->tskey, prior_snd_una) &&
3209 before(shinfo->tskey, tcp_sk(sk)->snd_una)) {
3210 tcp_skb_tsorted_save(skb) {
3211 __skb_tstamp_tx(skb, NULL, sk, SCM_TSTAMP_ACK);
3212 } tcp_skb_tsorted_restore(skb);
3216 /* Remove acknowledged frames from the retransmission queue. If our packet
3217 * is before the ack sequence we can discard it as it's confirmed to have
3218 * arrived at the other end.
3220 static int tcp_clean_rtx_queue(struct sock *sk, u32 prior_fack,
3222 struct tcp_sacktag_state *sack, bool ece_ack)
3224 const struct inet_connection_sock *icsk = inet_csk(sk);
3225 u64 first_ackt, last_ackt;
3226 struct tcp_sock *tp = tcp_sk(sk);
3227 u32 prior_sacked = tp->sacked_out;
3228 u32 reord = tp->snd_nxt; /* lowest acked un-retx un-sacked seq */
3229 struct sk_buff *skb, *next;
3230 bool fully_acked = true;
3231 long sack_rtt_us = -1L;
3232 long seq_rtt_us = -1L;
3233 long ca_rtt_us = -1L;
3235 u32 last_in_flight = 0;
3241 for (skb = skb_rb_first(&sk->tcp_rtx_queue); skb; skb = next) {
3242 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3243 const u32 start_seq = scb->seq;
3244 u8 sacked = scb->sacked;
3247 /* Determine how many packets and what bytes were acked, tso and else */
3248 if (after(scb->end_seq, tp->snd_una)) {
3249 if (tcp_skb_pcount(skb) == 1 ||
3250 !after(tp->snd_una, scb->seq))
3253 acked_pcount = tcp_tso_acked(sk, skb);
3256 fully_acked = false;
3258 acked_pcount = tcp_skb_pcount(skb);
3261 if (unlikely(sacked & TCPCB_RETRANS)) {
3262 if (sacked & TCPCB_SACKED_RETRANS)
3263 tp->retrans_out -= acked_pcount;
3264 flag |= FLAG_RETRANS_DATA_ACKED;
3265 } else if (!(sacked & TCPCB_SACKED_ACKED)) {
3266 last_ackt = tcp_skb_timestamp_us(skb);
3267 WARN_ON_ONCE(last_ackt == 0);
3269 first_ackt = last_ackt;
3271 last_in_flight = TCP_SKB_CB(skb)->tx.in_flight;
3272 if (before(start_seq, reord))
3274 if (!after(scb->end_seq, tp->high_seq))
3275 flag |= FLAG_ORIG_SACK_ACKED;
3278 if (sacked & TCPCB_SACKED_ACKED) {
3279 tp->sacked_out -= acked_pcount;
3280 } else if (tcp_is_sack(tp)) {
3281 tcp_count_delivered(tp, acked_pcount, ece_ack);
3282 if (!tcp_skb_spurious_retrans(tp, skb))
3283 tcp_rack_advance(tp, sacked, scb->end_seq,
3284 tcp_skb_timestamp_us(skb));
3286 if (sacked & TCPCB_LOST)
3287 tp->lost_out -= acked_pcount;
3289 tp->packets_out -= acked_pcount;
3290 pkts_acked += acked_pcount;
3291 tcp_rate_skb_delivered(sk, skb, sack->rate);
3293 /* Initial outgoing SYN's get put onto the write_queue
3294 * just like anything else we transmit. It is not
3295 * true data, and if we misinform our callers that
3296 * this ACK acks real data, we will erroneously exit
3297 * connection startup slow start one packet too
3298 * quickly. This is severely frowned upon behavior.
3300 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
3301 flag |= FLAG_DATA_ACKED;
3303 flag |= FLAG_SYN_ACKED;
3304 tp->retrans_stamp = 0;
3310 tcp_ack_tstamp(sk, skb, prior_snd_una);
3312 next = skb_rb_next(skb);
3313 if (unlikely(skb == tp->retransmit_skb_hint))
3314 tp->retransmit_skb_hint = NULL;
3315 if (unlikely(skb == tp->lost_skb_hint))
3316 tp->lost_skb_hint = NULL;
3317 tcp_highest_sack_replace(sk, skb, next);
3318 tcp_rtx_queue_unlink_and_free(skb, sk);
3322 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
3324 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3325 tp->snd_up = tp->snd_una;
3328 tcp_ack_tstamp(sk, skb, prior_snd_una);
3329 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
3330 flag |= FLAG_SACK_RENEGING;
3333 if (likely(first_ackt) && !(flag & FLAG_RETRANS_DATA_ACKED)) {
3334 seq_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, first_ackt);
3335 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, last_ackt);
3337 if (pkts_acked == 1 && last_in_flight < tp->mss_cache &&
3338 last_in_flight && !prior_sacked && fully_acked &&
3339 sack->rate->prior_delivered + 1 == tp->delivered &&
3340 !(flag & (FLAG_CA_ALERT | FLAG_SYN_ACKED))) {
3341 /* Conservatively mark a delayed ACK. It's typically
3342 * from a lone runt packet over the round trip to
3343 * a receiver w/o out-of-order or CE events.
3345 flag |= FLAG_ACK_MAYBE_DELAYED;
3348 if (sack->first_sackt) {
3349 sack_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->first_sackt);
3350 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->last_sackt);
3352 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us,
3353 ca_rtt_us, sack->rate);
3355 if (flag & FLAG_ACKED) {
3356 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3357 if (unlikely(icsk->icsk_mtup.probe_size &&
3358 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3359 tcp_mtup_probe_success(sk);
3362 if (tcp_is_reno(tp)) {
3363 tcp_remove_reno_sacks(sk, pkts_acked, ece_ack);
3365 /* If any of the cumulatively ACKed segments was
3366 * retransmitted, non-SACK case cannot confirm that
3367 * progress was due to original transmission due to
3368 * lack of TCPCB_SACKED_ACKED bits even if some of
3369 * the packets may have been never retransmitted.
3371 if (flag & FLAG_RETRANS_DATA_ACKED)
3372 flag &= ~FLAG_ORIG_SACK_ACKED;
3376 /* Non-retransmitted hole got filled? That's reordering */
3377 if (before(reord, prior_fack))
3378 tcp_check_sack_reordering(sk, reord, 0);
3380 delta = prior_sacked - tp->sacked_out;
3381 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3383 } else if (skb && rtt_update && sack_rtt_us >= 0 &&
3384 sack_rtt_us > tcp_stamp_us_delta(tp->tcp_mstamp,
3385 tcp_skb_timestamp_us(skb))) {
3386 /* Do not re-arm RTO if the sack RTT is measured from data sent
3387 * after when the head was last (re)transmitted. Otherwise the
3388 * timeout may continue to extend in loss recovery.
3390 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3393 if (icsk->icsk_ca_ops->pkts_acked) {
3394 struct ack_sample sample = { .pkts_acked = pkts_acked,
3395 .rtt_us = sack->rate->rtt_us,
3396 .in_flight = last_in_flight };
3398 icsk->icsk_ca_ops->pkts_acked(sk, &sample);
3401 #if FASTRETRANS_DEBUG > 0
3402 WARN_ON((int)tp->sacked_out < 0);
3403 WARN_ON((int)tp->lost_out < 0);
3404 WARN_ON((int)tp->retrans_out < 0);
3405 if (!tp->packets_out && tcp_is_sack(tp)) {
3406 icsk = inet_csk(sk);
3408 pr_debug("Leak l=%u %d\n",
3409 tp->lost_out, icsk->icsk_ca_state);
3412 if (tp->sacked_out) {
3413 pr_debug("Leak s=%u %d\n",
3414 tp->sacked_out, icsk->icsk_ca_state);
3417 if (tp->retrans_out) {
3418 pr_debug("Leak r=%u %d\n",
3419 tp->retrans_out, icsk->icsk_ca_state);
3420 tp->retrans_out = 0;
3427 static void tcp_ack_probe(struct sock *sk)
3429 struct inet_connection_sock *icsk = inet_csk(sk);
3430 struct sk_buff *head = tcp_send_head(sk);
3431 const struct tcp_sock *tp = tcp_sk(sk);
3433 /* Was it a usable window open? */
3436 if (!after(TCP_SKB_CB(head)->end_seq, tcp_wnd_end(tp))) {
3437 icsk->icsk_backoff = 0;
3438 icsk->icsk_probes_tstamp = 0;
3439 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3440 /* Socket must be waked up by subsequent tcp_data_snd_check().
3441 * This function is not for random using!
3444 unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX);
3446 when = tcp_clamp_probe0_to_user_timeout(sk, when);
3447 tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, when, TCP_RTO_MAX);
3451 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3453 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3454 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3457 /* Decide wheather to run the increase function of congestion control. */
3458 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3460 /* If reordering is high then always grow cwnd whenever data is
3461 * delivered regardless of its ordering. Otherwise stay conservative
3462 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3463 * new SACK or ECE mark may first advance cwnd here and later reduce
3464 * cwnd in tcp_fastretrans_alert() based on more states.
3466 if (tcp_sk(sk)->reordering >
3467 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_reordering))
3468 return flag & FLAG_FORWARD_PROGRESS;
3470 return flag & FLAG_DATA_ACKED;
3473 /* The "ultimate" congestion control function that aims to replace the rigid
3474 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction).
3475 * It's called toward the end of processing an ACK with precise rate
3476 * information. All transmission or retransmission are delayed afterwards.
3478 static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked,
3479 int flag, const struct rate_sample *rs)
3481 const struct inet_connection_sock *icsk = inet_csk(sk);
3483 if (icsk->icsk_ca_ops->cong_control) {
3484 icsk->icsk_ca_ops->cong_control(sk, rs);
3488 if (tcp_in_cwnd_reduction(sk)) {
3489 /* Reduce cwnd if state mandates */
3490 tcp_cwnd_reduction(sk, acked_sacked, flag);
3491 } else if (tcp_may_raise_cwnd(sk, flag)) {
3492 /* Advance cwnd if state allows */
3493 tcp_cong_avoid(sk, ack, acked_sacked);
3495 tcp_update_pacing_rate(sk);
3498 /* Check that window update is acceptable.
3499 * The function assumes that snd_una<=ack<=snd_next.
3501 static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3502 const u32 ack, const u32 ack_seq,
3505 return after(ack, tp->snd_una) ||
3506 after(ack_seq, tp->snd_wl1) ||
3507 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3510 /* If we update tp->snd_una, also update tp->bytes_acked */
3511 static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack)
3513 u32 delta = ack - tp->snd_una;
3515 sock_owned_by_me((struct sock *)tp);
3516 tp->bytes_acked += delta;
3520 /* If we update tp->rcv_nxt, also update tp->bytes_received */
3521 static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq)
3523 u32 delta = seq - tp->rcv_nxt;
3525 sock_owned_by_me((struct sock *)tp);
3526 tp->bytes_received += delta;
3527 WRITE_ONCE(tp->rcv_nxt, seq);
3530 /* Update our send window.
3532 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3533 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3535 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3538 struct tcp_sock *tp = tcp_sk(sk);
3540 u32 nwin = ntohs(tcp_hdr(skb)->window);
3542 if (likely(!tcp_hdr(skb)->syn))
3543 nwin <<= tp->rx_opt.snd_wscale;
3545 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3546 flag |= FLAG_WIN_UPDATE;
3547 tcp_update_wl(tp, ack_seq);
3549 if (tp->snd_wnd != nwin) {
3552 /* Note, it is the only place, where
3553 * fast path is recovered for sending TCP.
3556 tcp_fast_path_check(sk);
3558 if (!tcp_write_queue_empty(sk))
3559 tcp_slow_start_after_idle_check(sk);
3561 if (nwin > tp->max_window) {
3562 tp->max_window = nwin;
3563 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3568 tcp_snd_una_update(tp, ack);
3573 static bool __tcp_oow_rate_limited(struct net *net, int mib_idx,
3574 u32 *last_oow_ack_time)
3576 /* Paired with the WRITE_ONCE() in this function. */
3577 u32 val = READ_ONCE(*last_oow_ack_time);
3580 s32 elapsed = (s32)(tcp_jiffies32 - val);
3583 elapsed < READ_ONCE(net->ipv4.sysctl_tcp_invalid_ratelimit)) {
3584 NET_INC_STATS(net, mib_idx);
3585 return true; /* rate-limited: don't send yet! */
3589 /* Paired with the prior READ_ONCE() and with itself,
3590 * as we might be lockless.
3592 WRITE_ONCE(*last_oow_ack_time, tcp_jiffies32);
3594 return false; /* not rate-limited: go ahead, send dupack now! */
3597 /* Return true if we're currently rate-limiting out-of-window ACKs and
3598 * thus shouldn't send a dupack right now. We rate-limit dupacks in
3599 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3600 * attacks that send repeated SYNs or ACKs for the same connection. To
3601 * do this, we do not send a duplicate SYNACK or ACK if the remote
3602 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3604 bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb,
3605 int mib_idx, u32 *last_oow_ack_time)
3607 /* Data packets without SYNs are not likely part of an ACK loop. */
3608 if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) &&
3612 return __tcp_oow_rate_limited(net, mib_idx, last_oow_ack_time);
3615 /* RFC 5961 7 [ACK Throttling] */
3616 static void tcp_send_challenge_ack(struct sock *sk, const struct sk_buff *skb)
3618 /* unprotected vars, we dont care of overwrites */
3619 static u32 challenge_timestamp;
3620 static unsigned int challenge_count;
3621 struct tcp_sock *tp = tcp_sk(sk);
3622 struct net *net = sock_net(sk);
3625 /* First check our per-socket dupack rate limit. */
3626 if (__tcp_oow_rate_limited(net,
3627 LINUX_MIB_TCPACKSKIPPEDCHALLENGE,
3628 &tp->last_oow_ack_time))
3631 /* Then check host-wide RFC 5961 rate limit. */
3633 if (now != READ_ONCE(challenge_timestamp)) {
3634 u32 ack_limit = READ_ONCE(net->ipv4.sysctl_tcp_challenge_ack_limit);
3635 u32 half = (ack_limit + 1) >> 1;
3637 WRITE_ONCE(challenge_timestamp, now);
3638 WRITE_ONCE(challenge_count, half + prandom_u32_max(ack_limit));
3640 count = READ_ONCE(challenge_count);
3642 WRITE_ONCE(challenge_count, count - 1);
3643 NET_INC_STATS(net, LINUX_MIB_TCPCHALLENGEACK);
3648 static void tcp_store_ts_recent(struct tcp_sock *tp)
3650 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3651 tp->rx_opt.ts_recent_stamp = ktime_get_seconds();
3654 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3656 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3657 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3658 * extra check below makes sure this can only happen
3659 * for pure ACK frames. -DaveM
3661 * Not only, also it occurs for expired timestamps.
3664 if (tcp_paws_check(&tp->rx_opt, 0))
3665 tcp_store_ts_recent(tp);
3669 /* This routine deals with acks during a TLP episode and ends an episode by
3670 * resetting tlp_high_seq. Ref: TLP algorithm in draft-ietf-tcpm-rack
3672 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3674 struct tcp_sock *tp = tcp_sk(sk);
3676 if (before(ack, tp->tlp_high_seq))
3679 if (!tp->tlp_retrans) {
3680 /* TLP of new data has been acknowledged */
3681 tp->tlp_high_seq = 0;
3682 } else if (flag & FLAG_DSACKING_ACK) {
3683 /* This DSACK means original and TLP probe arrived; no loss */
3684 tp->tlp_high_seq = 0;
3685 } else if (after(ack, tp->tlp_high_seq)) {
3686 /* ACK advances: there was a loss, so reduce cwnd. Reset
3687 * tlp_high_seq in tcp_init_cwnd_reduction()
3689 tcp_init_cwnd_reduction(sk);
3690 tcp_set_ca_state(sk, TCP_CA_CWR);
3691 tcp_end_cwnd_reduction(sk);
3692 tcp_try_keep_open(sk);
3693 NET_INC_STATS(sock_net(sk),
3694 LINUX_MIB_TCPLOSSPROBERECOVERY);
3695 } else if (!(flag & (FLAG_SND_UNA_ADVANCED |
3696 FLAG_NOT_DUP | FLAG_DATA_SACKED))) {
3697 /* Pure dupack: original and TLP probe arrived; no loss */
3698 tp->tlp_high_seq = 0;
3702 static inline void tcp_in_ack_event(struct sock *sk, u32 flags)
3704 const struct inet_connection_sock *icsk = inet_csk(sk);
3706 if (icsk->icsk_ca_ops->in_ack_event)
3707 icsk->icsk_ca_ops->in_ack_event(sk, flags);
3710 /* Congestion control has updated the cwnd already. So if we're in
3711 * loss recovery then now we do any new sends (for FRTO) or
3712 * retransmits (for CA_Loss or CA_recovery) that make sense.
3714 static void tcp_xmit_recovery(struct sock *sk, int rexmit)
3716 struct tcp_sock *tp = tcp_sk(sk);
3718 if (rexmit == REXMIT_NONE || sk->sk_state == TCP_SYN_SENT)
3721 if (unlikely(rexmit == REXMIT_NEW)) {
3722 __tcp_push_pending_frames(sk, tcp_current_mss(sk),
3724 if (after(tp->snd_nxt, tp->high_seq))
3728 tcp_xmit_retransmit_queue(sk);
3731 /* Returns the number of packets newly acked or sacked by the current ACK */
3732 static u32 tcp_newly_delivered(struct sock *sk, u32 prior_delivered, int flag)
3734 const struct net *net = sock_net(sk);
3735 struct tcp_sock *tp = tcp_sk(sk);
3738 delivered = tp->delivered - prior_delivered;
3739 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVERED, delivered);
3740 if (flag & FLAG_ECE)
3741 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVEREDCE, delivered);
3746 /* This routine deals with incoming acks, but not outgoing ones. */
3747 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3749 struct inet_connection_sock *icsk = inet_csk(sk);
3750 struct tcp_sock *tp = tcp_sk(sk);
3751 struct tcp_sacktag_state sack_state;
3752 struct rate_sample rs = { .prior_delivered = 0 };
3753 u32 prior_snd_una = tp->snd_una;
3754 bool is_sack_reneg = tp->is_sack_reneg;
3755 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3756 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3758 int prior_packets = tp->packets_out;
3759 u32 delivered = tp->delivered;
3760 u32 lost = tp->lost;
3761 int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */
3764 sack_state.first_sackt = 0;
3765 sack_state.rate = &rs;
3766 sack_state.sack_delivered = 0;
3768 /* We very likely will need to access rtx queue. */
3769 prefetch(sk->tcp_rtx_queue.rb_node);
3771 /* If the ack is older than previous acks
3772 * then we can probably ignore it.
3774 if (before(ack, prior_snd_una)) {
3777 /* do not accept ACK for bytes we never sent. */
3778 max_window = min_t(u64, tp->max_window, tp->bytes_acked);
3779 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3780 if (before(ack, prior_snd_una - max_window)) {
3781 if (!(flag & FLAG_NO_CHALLENGE_ACK))
3782 tcp_send_challenge_ack(sk, skb);
3788 /* If the ack includes data we haven't sent yet, discard
3789 * this segment (RFC793 Section 3.9).
3791 if (after(ack, tp->snd_nxt))
3794 if (after(ack, prior_snd_una)) {
3795 flag |= FLAG_SND_UNA_ADVANCED;
3796 icsk->icsk_retransmits = 0;
3798 #if IS_ENABLED(CONFIG_TLS_DEVICE)
3799 if (static_branch_unlikely(&clean_acked_data_enabled.key))
3800 if (icsk->icsk_clean_acked)
3801 icsk->icsk_clean_acked(sk, ack);
3805 prior_fack = tcp_is_sack(tp) ? tcp_highest_sack_seq(tp) : tp->snd_una;
3806 rs.prior_in_flight = tcp_packets_in_flight(tp);
3808 /* ts_recent update must be made after we are sure that the packet
3811 if (flag & FLAG_UPDATE_TS_RECENT)
3812 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3814 if ((flag & (FLAG_SLOWPATH | FLAG_SND_UNA_ADVANCED)) ==
3815 FLAG_SND_UNA_ADVANCED) {
3816 /* Window is constant, pure forward advance.
3817 * No more checks are required.
3818 * Note, we use the fact that SND.UNA>=SND.WL2.
3820 tcp_update_wl(tp, ack_seq);
3821 tcp_snd_una_update(tp, ack);
3822 flag |= FLAG_WIN_UPDATE;
3824 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
3826 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPACKS);
3828 u32 ack_ev_flags = CA_ACK_SLOWPATH;
3830 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3833 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3835 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3837 if (TCP_SKB_CB(skb)->sacked)
3838 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3841 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) {
3843 ack_ev_flags |= CA_ACK_ECE;
3846 if (sack_state.sack_delivered)
3847 tcp_count_delivered(tp, sack_state.sack_delivered,
3850 if (flag & FLAG_WIN_UPDATE)
3851 ack_ev_flags |= CA_ACK_WIN_UPDATE;
3853 tcp_in_ack_event(sk, ack_ev_flags);
3856 /* This is a deviation from RFC3168 since it states that:
3857 * "When the TCP data sender is ready to set the CWR bit after reducing
3858 * the congestion window, it SHOULD set the CWR bit only on the first
3859 * new data packet that it transmits."
3860 * We accept CWR on pure ACKs to be more robust
3861 * with widely-deployed TCP implementations that do this.
3863 tcp_ecn_accept_cwr(sk, skb);
3865 /* We passed data and got it acked, remove any soft error
3866 * log. Something worked...
3868 sk->sk_err_soft = 0;
3869 icsk->icsk_probes_out = 0;
3870 tp->rcv_tstamp = tcp_jiffies32;
3874 /* See if we can take anything off of the retransmit queue. */
3875 flag |= tcp_clean_rtx_queue(sk, prior_fack, prior_snd_una, &sack_state,
3878 tcp_rack_update_reo_wnd(sk, &rs);
3880 if (tp->tlp_high_seq)
3881 tcp_process_tlp_ack(sk, ack, flag);
3883 if (tcp_ack_is_dubious(sk, flag)) {
3884 if (!(flag & (FLAG_SND_UNA_ADVANCED |
3885 FLAG_NOT_DUP | FLAG_DSACKING_ACK))) {
3887 /* Consider if pure acks were aggregated in tcp_add_backlog() */
3888 if (!(flag & FLAG_DATA))
3889 num_dupack = max_t(u16, 1, skb_shinfo(skb)->gso_segs);
3891 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
3895 /* If needed, reset TLP/RTO timer when RACK doesn't set. */
3896 if (flag & FLAG_SET_XMIT_TIMER)
3897 tcp_set_xmit_timer(sk);
3899 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
3902 delivered = tcp_newly_delivered(sk, delivered, flag);
3903 lost = tp->lost - lost; /* freshly marked lost */
3904 rs.is_ack_delayed = !!(flag & FLAG_ACK_MAYBE_DELAYED);
3905 tcp_rate_gen(sk, delivered, lost, is_sack_reneg, sack_state.rate);
3906 tcp_cong_control(sk, ack, delivered, flag, sack_state.rate);
3907 tcp_xmit_recovery(sk, rexmit);
3911 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3912 if (flag & FLAG_DSACKING_ACK) {
3913 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
3915 tcp_newly_delivered(sk, delivered, flag);
3917 /* If this ack opens up a zero window, clear backoff. It was
3918 * being used to time the probes, and is probably far higher than
3919 * it needs to be for normal retransmission.
3923 if (tp->tlp_high_seq)
3924 tcp_process_tlp_ack(sk, ack, flag);
3928 /* If data was SACKed, tag it and see if we should send more data.
3929 * If data was DSACKed, see if we can undo a cwnd reduction.
3931 if (TCP_SKB_CB(skb)->sacked) {
3932 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3934 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
3936 tcp_newly_delivered(sk, delivered, flag);
3937 tcp_xmit_recovery(sk, rexmit);
3943 static void tcp_parse_fastopen_option(int len, const unsigned char *cookie,
3944 bool syn, struct tcp_fastopen_cookie *foc,
3947 /* Valid only in SYN or SYN-ACK with an even length. */
3948 if (!foc || !syn || len < 0 || (len & 1))
3951 if (len >= TCP_FASTOPEN_COOKIE_MIN &&
3952 len <= TCP_FASTOPEN_COOKIE_MAX)
3953 memcpy(foc->val, cookie, len);
3960 static bool smc_parse_options(const struct tcphdr *th,
3961 struct tcp_options_received *opt_rx,
3962 const unsigned char *ptr,
3965 #if IS_ENABLED(CONFIG_SMC)
3966 if (static_branch_unlikely(&tcp_have_smc)) {
3967 if (th->syn && !(opsize & 1) &&
3968 opsize >= TCPOLEN_EXP_SMC_BASE &&
3969 get_unaligned_be32(ptr) == TCPOPT_SMC_MAGIC) {
3978 /* Try to parse the MSS option from the TCP header. Return 0 on failure, clamped
3981 static u16 tcp_parse_mss_option(const struct tcphdr *th, u16 user_mss)
3983 const unsigned char *ptr = (const unsigned char *)(th + 1);
3984 int length = (th->doff * 4) - sizeof(struct tcphdr);
3987 while (length > 0) {
3988 int opcode = *ptr++;
3994 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
4001 if (opsize < 2) /* "silly options" */
4003 if (opsize > length)
4004 return mss; /* fail on partial options */
4005 if (opcode == TCPOPT_MSS && opsize == TCPOLEN_MSS) {
4006 u16 in_mss = get_unaligned_be16(ptr);
4009 if (user_mss && user_mss < in_mss)
4021 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
4022 * But, this can also be called on packets in the established flow when
4023 * the fast version below fails.
4025 void tcp_parse_options(const struct net *net,
4026 const struct sk_buff *skb,
4027 struct tcp_options_received *opt_rx, int estab,
4028 struct tcp_fastopen_cookie *foc)
4030 const unsigned char *ptr;
4031 const struct tcphdr *th = tcp_hdr(skb);
4032 int length = (th->doff * 4) - sizeof(struct tcphdr);
4034 ptr = (const unsigned char *)(th + 1);
4035 opt_rx->saw_tstamp = 0;
4036 opt_rx->saw_unknown = 0;
4038 while (length > 0) {
4039 int opcode = *ptr++;
4045 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
4052 if (opsize < 2) /* "silly options" */
4054 if (opsize > length)
4055 return; /* don't parse partial options */
4058 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
4059 u16 in_mss = get_unaligned_be16(ptr);
4061 if (opt_rx->user_mss &&
4062 opt_rx->user_mss < in_mss)
4063 in_mss = opt_rx->user_mss;
4064 opt_rx->mss_clamp = in_mss;
4069 if (opsize == TCPOLEN_WINDOW && th->syn &&
4070 !estab && READ_ONCE(net->ipv4.sysctl_tcp_window_scaling)) {
4071 __u8 snd_wscale = *(__u8 *)ptr;
4072 opt_rx->wscale_ok = 1;
4073 if (snd_wscale > TCP_MAX_WSCALE) {
4074 net_info_ratelimited("%s: Illegal window scaling value %d > %u received\n",
4078 snd_wscale = TCP_MAX_WSCALE;
4080 opt_rx->snd_wscale = snd_wscale;
4083 case TCPOPT_TIMESTAMP:
4084 if ((opsize == TCPOLEN_TIMESTAMP) &&
4085 ((estab && opt_rx->tstamp_ok) ||
4086 (!estab && READ_ONCE(net->ipv4.sysctl_tcp_timestamps)))) {
4087 opt_rx->saw_tstamp = 1;
4088 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
4089 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
4092 case TCPOPT_SACK_PERM:
4093 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
4094 !estab && READ_ONCE(net->ipv4.sysctl_tcp_sack)) {
4095 opt_rx->sack_ok = TCP_SACK_SEEN;
4096 tcp_sack_reset(opt_rx);
4101 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
4102 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
4104 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
4107 #ifdef CONFIG_TCP_MD5SIG
4110 * The MD5 Hash has already been
4111 * checked (see tcp_v{4,6}_do_rcv()).
4115 case TCPOPT_FASTOPEN:
4116 tcp_parse_fastopen_option(
4117 opsize - TCPOLEN_FASTOPEN_BASE,
4118 ptr, th->syn, foc, false);
4122 /* Fast Open option shares code 254 using a
4123 * 16 bits magic number.
4125 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE &&
4126 get_unaligned_be16(ptr) ==
4127 TCPOPT_FASTOPEN_MAGIC) {
4128 tcp_parse_fastopen_option(opsize -
4129 TCPOLEN_EXP_FASTOPEN_BASE,
4130 ptr + 2, th->syn, foc, true);
4134 if (smc_parse_options(th, opt_rx, ptr, opsize))
4137 opt_rx->saw_unknown = 1;
4141 opt_rx->saw_unknown = 1;
4148 EXPORT_SYMBOL(tcp_parse_options);
4150 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
4152 const __be32 *ptr = (const __be32 *)(th + 1);
4154 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
4155 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
4156 tp->rx_opt.saw_tstamp = 1;
4158 tp->rx_opt.rcv_tsval = ntohl(*ptr);
4161 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
4163 tp->rx_opt.rcv_tsecr = 0;
4169 /* Fast parse options. This hopes to only see timestamps.
4170 * If it is wrong it falls back on tcp_parse_options().
4172 static bool tcp_fast_parse_options(const struct net *net,
4173 const struct sk_buff *skb,
4174 const struct tcphdr *th, struct tcp_sock *tp)
4176 /* In the spirit of fast parsing, compare doff directly to constant
4177 * values. Because equality is used, short doff can be ignored here.
4179 if (th->doff == (sizeof(*th) / 4)) {
4180 tp->rx_opt.saw_tstamp = 0;
4182 } else if (tp->rx_opt.tstamp_ok &&
4183 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
4184 if (tcp_parse_aligned_timestamp(tp, th))
4188 tcp_parse_options(net, skb, &tp->rx_opt, 1, NULL);
4189 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
4190 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
4195 #ifdef CONFIG_TCP_MD5SIG
4197 * Parse MD5 Signature option
4199 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
4201 int length = (th->doff << 2) - sizeof(*th);
4202 const u8 *ptr = (const u8 *)(th + 1);
4204 /* If not enough data remaining, we can short cut */
4205 while (length >= TCPOLEN_MD5SIG) {
4206 int opcode = *ptr++;
4217 if (opsize < 2 || opsize > length)
4219 if (opcode == TCPOPT_MD5SIG)
4220 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
4227 EXPORT_SYMBOL(tcp_parse_md5sig_option);
4230 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
4232 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
4233 * it can pass through stack. So, the following predicate verifies that
4234 * this segment is not used for anything but congestion avoidance or
4235 * fast retransmit. Moreover, we even are able to eliminate most of such
4236 * second order effects, if we apply some small "replay" window (~RTO)
4237 * to timestamp space.
4239 * All these measures still do not guarantee that we reject wrapped ACKs
4240 * on networks with high bandwidth, when sequence space is recycled fastly,
4241 * but it guarantees that such events will be very rare and do not affect
4242 * connection seriously. This doesn't look nice, but alas, PAWS is really
4245 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
4246 * states that events when retransmit arrives after original data are rare.
4247 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
4248 * the biggest problem on large power networks even with minor reordering.
4249 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
4250 * up to bandwidth of 18Gigabit/sec. 8) ]
4253 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
4255 const struct tcp_sock *tp = tcp_sk(sk);
4256 const struct tcphdr *th = tcp_hdr(skb);
4257 u32 seq = TCP_SKB_CB(skb)->seq;
4258 u32 ack = TCP_SKB_CB(skb)->ack_seq;
4260 return (/* 1. Pure ACK with correct sequence number. */
4261 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
4263 /* 2. ... and duplicate ACK. */
4264 ack == tp->snd_una &&
4266 /* 3. ... and does not update window. */
4267 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
4269 /* 4. ... and sits in replay window. */
4270 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
4273 static inline bool tcp_paws_discard(const struct sock *sk,
4274 const struct sk_buff *skb)
4276 const struct tcp_sock *tp = tcp_sk(sk);
4278 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
4279 !tcp_disordered_ack(sk, skb);
4282 /* Check segment sequence number for validity.
4284 * Segment controls are considered valid, if the segment
4285 * fits to the window after truncation to the window. Acceptability
4286 * of data (and SYN, FIN, of course) is checked separately.
4287 * See tcp_data_queue(), for example.
4289 * Also, controls (RST is main one) are accepted using RCV.WUP instead
4290 * of RCV.NXT. Peer still did not advance his SND.UNA when we
4291 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4292 * (borrowed from freebsd)
4295 static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
4297 return !before(end_seq, tp->rcv_wup) &&
4298 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
4301 /* When we get a reset we do this. */
4302 void tcp_reset(struct sock *sk)
4304 trace_tcp_receive_reset(sk);
4306 /* We want the right error as BSD sees it (and indeed as we do). */
4307 switch (sk->sk_state) {
4309 sk->sk_err = ECONNREFUSED;
4311 case TCP_CLOSE_WAIT:
4317 sk->sk_err = ECONNRESET;
4319 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4322 tcp_write_queue_purge(sk);
4325 if (!sock_flag(sk, SOCK_DEAD))
4326 sk->sk_error_report(sk);
4330 * Process the FIN bit. This now behaves as it is supposed to work
4331 * and the FIN takes effect when it is validly part of sequence
4332 * space. Not before when we get holes.
4334 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4335 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4338 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4339 * close and we go into CLOSING (and later onto TIME-WAIT)
4341 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4343 void tcp_fin(struct sock *sk)
4345 struct tcp_sock *tp = tcp_sk(sk);
4347 inet_csk_schedule_ack(sk);
4349 WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | RCV_SHUTDOWN);
4350 sock_set_flag(sk, SOCK_DONE);
4352 switch (sk->sk_state) {
4354 case TCP_ESTABLISHED:
4355 /* Move to CLOSE_WAIT */
4356 tcp_set_state(sk, TCP_CLOSE_WAIT);
4357 inet_csk_enter_pingpong_mode(sk);
4360 case TCP_CLOSE_WAIT:
4362 /* Received a retransmission of the FIN, do
4367 /* RFC793: Remain in the LAST-ACK state. */
4371 /* This case occurs when a simultaneous close
4372 * happens, we must ack the received FIN and
4373 * enter the CLOSING state.
4376 tcp_set_state(sk, TCP_CLOSING);
4379 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4381 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4384 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4385 * cases we should never reach this piece of code.
4387 pr_err("%s: Impossible, sk->sk_state=%d\n",
4388 __func__, sk->sk_state);
4392 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4393 * Probably, we should reset in this case. For now drop them.
4395 skb_rbtree_purge(&tp->out_of_order_queue);
4396 if (tcp_is_sack(tp))
4397 tcp_sack_reset(&tp->rx_opt);
4400 if (!sock_flag(sk, SOCK_DEAD)) {
4401 sk->sk_state_change(sk);
4403 /* Do not send POLL_HUP for half duplex close. */
4404 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4405 sk->sk_state == TCP_CLOSE)
4406 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4408 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4412 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4415 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4416 if (before(seq, sp->start_seq))
4417 sp->start_seq = seq;
4418 if (after(end_seq, sp->end_seq))
4419 sp->end_seq = end_seq;
4425 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4427 struct tcp_sock *tp = tcp_sk(sk);
4429 if (tcp_is_sack(tp) && READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_dsack)) {
4432 if (before(seq, tp->rcv_nxt))
4433 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4435 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4437 NET_INC_STATS(sock_net(sk), mib_idx);
4439 tp->rx_opt.dsack = 1;
4440 tp->duplicate_sack[0].start_seq = seq;
4441 tp->duplicate_sack[0].end_seq = end_seq;
4445 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4447 struct tcp_sock *tp = tcp_sk(sk);
4449 if (!tp->rx_opt.dsack)
4450 tcp_dsack_set(sk, seq, end_seq);
4452 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4455 static void tcp_rcv_spurious_retrans(struct sock *sk, const struct sk_buff *skb)
4457 /* When the ACK path fails or drops most ACKs, the sender would
4458 * timeout and spuriously retransmit the same segment repeatedly.
4459 * The receiver remembers and reflects via DSACKs. Leverage the
4460 * DSACK state and change the txhash to re-route speculatively.
4462 if (TCP_SKB_CB(skb)->seq == tcp_sk(sk)->duplicate_sack[0].start_seq &&
4463 sk_rethink_txhash(sk))
4464 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDUPLICATEDATAREHASH);
4467 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4469 struct tcp_sock *tp = tcp_sk(sk);
4471 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4472 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4473 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4474 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
4476 if (tcp_is_sack(tp) && READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_dsack)) {
4477 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4479 tcp_rcv_spurious_retrans(sk, skb);
4480 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4481 end_seq = tp->rcv_nxt;
4482 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4489 /* These routines update the SACK block as out-of-order packets arrive or
4490 * in-order packets close up the sequence space.
4492 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4495 struct tcp_sack_block *sp = &tp->selective_acks[0];
4496 struct tcp_sack_block *swalk = sp + 1;
4498 /* See if the recent change to the first SACK eats into
4499 * or hits the sequence space of other SACK blocks, if so coalesce.
4501 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4502 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4505 /* Zap SWALK, by moving every further SACK up by one slot.
4506 * Decrease num_sacks.
4508 tp->rx_opt.num_sacks--;
4509 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4518 static void tcp_sack_compress_send_ack(struct sock *sk)
4520 struct tcp_sock *tp = tcp_sk(sk);
4522 if (!tp->compressed_ack)
4525 if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1)
4528 /* Since we have to send one ack finally,
4529 * substract one from tp->compressed_ack to keep
4530 * LINUX_MIB_TCPACKCOMPRESSED accurate.
4532 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED,
4533 tp->compressed_ack - 1);
4535 tp->compressed_ack = 0;
4539 /* Reasonable amount of sack blocks included in TCP SACK option
4540 * The max is 4, but this becomes 3 if TCP timestamps are there.
4541 * Given that SACK packets might be lost, be conservative and use 2.
4543 #define TCP_SACK_BLOCKS_EXPECTED 2
4545 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4547 struct tcp_sock *tp = tcp_sk(sk);
4548 struct tcp_sack_block *sp = &tp->selective_acks[0];
4549 int cur_sacks = tp->rx_opt.num_sacks;
4555 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4556 if (tcp_sack_extend(sp, seq, end_seq)) {
4557 if (this_sack >= TCP_SACK_BLOCKS_EXPECTED)
4558 tcp_sack_compress_send_ack(sk);
4559 /* Rotate this_sack to the first one. */
4560 for (; this_sack > 0; this_sack--, sp--)
4561 swap(*sp, *(sp - 1));
4563 tcp_sack_maybe_coalesce(tp);
4568 if (this_sack >= TCP_SACK_BLOCKS_EXPECTED)
4569 tcp_sack_compress_send_ack(sk);
4571 /* Could not find an adjacent existing SACK, build a new one,
4572 * put it at the front, and shift everyone else down. We
4573 * always know there is at least one SACK present already here.
4575 * If the sack array is full, forget about the last one.
4577 if (this_sack >= TCP_NUM_SACKS) {
4579 tp->rx_opt.num_sacks--;
4582 for (; this_sack > 0; this_sack--, sp--)
4586 /* Build the new head SACK, and we're done. */
4587 sp->start_seq = seq;
4588 sp->end_seq = end_seq;
4589 tp->rx_opt.num_sacks++;
4592 /* RCV.NXT advances, some SACKs should be eaten. */
4594 static void tcp_sack_remove(struct tcp_sock *tp)
4596 struct tcp_sack_block *sp = &tp->selective_acks[0];
4597 int num_sacks = tp->rx_opt.num_sacks;
4600 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4601 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4602 tp->rx_opt.num_sacks = 0;
4606 for (this_sack = 0; this_sack < num_sacks;) {
4607 /* Check if the start of the sack is covered by RCV.NXT. */
4608 if (!before(tp->rcv_nxt, sp->start_seq)) {
4611 /* RCV.NXT must cover all the block! */
4612 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4614 /* Zap this SACK, by moving forward any other SACKS. */
4615 for (i = this_sack+1; i < num_sacks; i++)
4616 tp->selective_acks[i-1] = tp->selective_acks[i];
4623 tp->rx_opt.num_sacks = num_sacks;
4627 * tcp_try_coalesce - try to merge skb to prior one
4630 * @from: buffer to add in queue
4631 * @fragstolen: pointer to boolean
4633 * Before queueing skb @from after @to, try to merge them
4634 * to reduce overall memory use and queue lengths, if cost is small.
4635 * Packets in ofo or receive queues can stay a long time.
4636 * Better try to coalesce them right now to avoid future collapses.
4637 * Returns true if caller should free @from instead of queueing it
4639 static bool tcp_try_coalesce(struct sock *sk,
4641 struct sk_buff *from,
4646 *fragstolen = false;
4648 /* Its possible this segment overlaps with prior segment in queue */
4649 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4652 if (!mptcp_skb_can_collapse(to, from))
4655 #ifdef CONFIG_TLS_DEVICE
4656 if (from->decrypted != to->decrypted)
4660 if (!skb_try_coalesce(to, from, fragstolen, &delta))
4663 atomic_add(delta, &sk->sk_rmem_alloc);
4664 sk_mem_charge(sk, delta);
4665 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4666 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4667 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4668 TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
4670 if (TCP_SKB_CB(from)->has_rxtstamp) {
4671 TCP_SKB_CB(to)->has_rxtstamp = true;
4672 to->tstamp = from->tstamp;
4673 skb_hwtstamps(to)->hwtstamp = skb_hwtstamps(from)->hwtstamp;
4679 static bool tcp_ooo_try_coalesce(struct sock *sk,
4681 struct sk_buff *from,
4684 bool res = tcp_try_coalesce(sk, to, from, fragstolen);
4686 /* In case tcp_drop() is called later, update to->gso_segs */
4688 u32 gso_segs = max_t(u16, 1, skb_shinfo(to)->gso_segs) +
4689 max_t(u16, 1, skb_shinfo(from)->gso_segs);
4691 skb_shinfo(to)->gso_segs = min_t(u32, gso_segs, 0xFFFF);
4696 static void tcp_drop(struct sock *sk, struct sk_buff *skb)
4698 sk_drops_add(sk, skb);
4702 /* This one checks to see if we can put data from the
4703 * out_of_order queue into the receive_queue.
4705 static void tcp_ofo_queue(struct sock *sk)
4707 struct tcp_sock *tp = tcp_sk(sk);
4708 __u32 dsack_high = tp->rcv_nxt;
4709 bool fin, fragstolen, eaten;
4710 struct sk_buff *skb, *tail;
4713 p = rb_first(&tp->out_of_order_queue);
4716 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4719 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4720 __u32 dsack = dsack_high;
4721 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4722 dsack_high = TCP_SKB_CB(skb)->end_seq;
4723 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4726 rb_erase(&skb->rbnode, &tp->out_of_order_queue);
4728 if (unlikely(!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))) {
4733 tail = skb_peek_tail(&sk->sk_receive_queue);
4734 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen);
4735 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4736 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
4738 __skb_queue_tail(&sk->sk_receive_queue, skb);
4740 kfree_skb_partial(skb, fragstolen);
4742 if (unlikely(fin)) {
4744 /* tcp_fin() purges tp->out_of_order_queue,
4745 * so we must end this loop right now.
4752 static bool tcp_prune_ofo_queue(struct sock *sk);
4753 static int tcp_prune_queue(struct sock *sk);
4755 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4758 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4759 !sk_rmem_schedule(sk, skb, size)) {
4761 if (tcp_prune_queue(sk) < 0)
4764 while (!sk_rmem_schedule(sk, skb, size)) {
4765 if (!tcp_prune_ofo_queue(sk))
4772 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4774 struct tcp_sock *tp = tcp_sk(sk);
4775 struct rb_node **p, *parent;
4776 struct sk_buff *skb1;
4780 tcp_ecn_check_ce(sk, skb);
4782 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4783 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFODROP);
4784 sk->sk_data_ready(sk);
4789 /* Disable header prediction. */
4791 inet_csk_schedule_ack(sk);
4793 tp->rcv_ooopack += max_t(u16, 1, skb_shinfo(skb)->gso_segs);
4794 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4795 seq = TCP_SKB_CB(skb)->seq;
4796 end_seq = TCP_SKB_CB(skb)->end_seq;
4798 p = &tp->out_of_order_queue.rb_node;
4799 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4800 /* Initial out of order segment, build 1 SACK. */
4801 if (tcp_is_sack(tp)) {
4802 tp->rx_opt.num_sacks = 1;
4803 tp->selective_acks[0].start_seq = seq;
4804 tp->selective_acks[0].end_seq = end_seq;
4806 rb_link_node(&skb->rbnode, NULL, p);
4807 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4808 tp->ooo_last_skb = skb;
4812 /* In the typical case, we are adding an skb to the end of the list.
4813 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
4815 if (tcp_ooo_try_coalesce(sk, tp->ooo_last_skb,
4816 skb, &fragstolen)) {
4818 /* For non sack flows, do not grow window to force DUPACK
4819 * and trigger fast retransmit.
4821 if (tcp_is_sack(tp))
4822 tcp_grow_window(sk, skb, true);
4823 kfree_skb_partial(skb, fragstolen);
4827 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
4828 if (!before(seq, TCP_SKB_CB(tp->ooo_last_skb)->end_seq)) {
4829 parent = &tp->ooo_last_skb->rbnode;
4830 p = &parent->rb_right;
4834 /* Find place to insert this segment. Handle overlaps on the way. */
4838 skb1 = rb_to_skb(parent);
4839 if (before(seq, TCP_SKB_CB(skb1)->seq)) {
4840 p = &parent->rb_left;
4843 if (before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4844 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4845 /* All the bits are present. Drop. */
4846 NET_INC_STATS(sock_net(sk),
4847 LINUX_MIB_TCPOFOMERGE);
4850 tcp_dsack_set(sk, seq, end_seq);
4853 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4854 /* Partial overlap. */
4855 tcp_dsack_set(sk, seq, TCP_SKB_CB(skb1)->end_seq);
4857 /* skb's seq == skb1's seq and skb covers skb1.
4858 * Replace skb1 with skb.
4860 rb_replace_node(&skb1->rbnode, &skb->rbnode,
4861 &tp->out_of_order_queue);
4862 tcp_dsack_extend(sk,
4863 TCP_SKB_CB(skb1)->seq,
4864 TCP_SKB_CB(skb1)->end_seq);
4865 NET_INC_STATS(sock_net(sk),
4866 LINUX_MIB_TCPOFOMERGE);
4870 } else if (tcp_ooo_try_coalesce(sk, skb1,
4871 skb, &fragstolen)) {
4874 p = &parent->rb_right;
4877 /* Insert segment into RB tree. */
4878 rb_link_node(&skb->rbnode, parent, p);
4879 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4882 /* Remove other segments covered by skb. */
4883 while ((skb1 = skb_rb_next(skb)) != NULL) {
4884 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4886 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4887 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4891 rb_erase(&skb1->rbnode, &tp->out_of_order_queue);
4892 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4893 TCP_SKB_CB(skb1)->end_seq);
4894 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4897 /* If there is no skb after us, we are the last_skb ! */
4899 tp->ooo_last_skb = skb;
4902 if (tcp_is_sack(tp))
4903 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4906 /* For non sack flows, do not grow window to force DUPACK
4907 * and trigger fast retransmit.
4909 if (tcp_is_sack(tp))
4910 tcp_grow_window(sk, skb, false);
4912 skb_set_owner_r(skb, sk);
4916 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb,
4920 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4923 tcp_try_coalesce(sk, tail,
4924 skb, fragstolen)) ? 1 : 0;
4925 tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq);
4927 __skb_queue_tail(&sk->sk_receive_queue, skb);
4928 skb_set_owner_r(skb, sk);
4933 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4935 struct sk_buff *skb;
4943 if (size > PAGE_SIZE) {
4944 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS);
4946 data_len = npages << PAGE_SHIFT;
4947 size = data_len + (size & ~PAGE_MASK);
4949 skb = alloc_skb_with_frags(size - data_len, data_len,
4950 PAGE_ALLOC_COSTLY_ORDER,
4951 &err, sk->sk_allocation);
4955 skb_put(skb, size - data_len);
4956 skb->data_len = data_len;
4959 if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) {
4960 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP);
4964 err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size);
4968 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4969 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4970 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4972 if (tcp_queue_rcv(sk, skb, &fragstolen)) {
4973 WARN_ON_ONCE(fragstolen); /* should not happen */
4985 void tcp_data_ready(struct sock *sk)
4987 const struct tcp_sock *tp = tcp_sk(sk);
4988 int avail = tp->rcv_nxt - tp->copied_seq;
4990 if (avail < sk->sk_rcvlowat && !tcp_rmem_pressure(sk) &&
4991 !sock_flag(sk, SOCK_DONE) &&
4992 tcp_receive_window(tp) > inet_csk(sk)->icsk_ack.rcv_mss)
4995 sk->sk_data_ready(sk);
4998 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
5000 struct tcp_sock *tp = tcp_sk(sk);
5004 if (sk_is_mptcp(sk))
5005 mptcp_incoming_options(sk, skb);
5007 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
5012 __skb_pull(skb, tcp_hdr(skb)->doff * 4);
5014 tp->rx_opt.dsack = 0;
5016 /* Queue data for delivery to the user.
5017 * Packets in sequence go to the receive queue.
5018 * Out of sequence packets to the out_of_order_queue.
5020 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
5021 if (tcp_receive_window(tp) == 0) {
5022 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP);
5026 /* Ok. In sequence. In window. */
5028 if (skb_queue_len(&sk->sk_receive_queue) == 0)
5029 sk_forced_mem_schedule(sk, skb->truesize);
5030 else if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) {
5031 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP);
5032 sk->sk_data_ready(sk);
5036 eaten = tcp_queue_rcv(sk, skb, &fragstolen);
5038 tcp_event_data_recv(sk, skb);
5039 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
5042 if (!RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
5045 /* RFC5681. 4.2. SHOULD send immediate ACK, when
5046 * gap in queue is filled.
5048 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
5049 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW;
5052 if (tp->rx_opt.num_sacks)
5053 tcp_sack_remove(tp);
5055 tcp_fast_path_check(sk);
5058 kfree_skb_partial(skb, fragstolen);
5059 if (!sock_flag(sk, SOCK_DEAD))
5064 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
5065 tcp_rcv_spurious_retrans(sk, skb);
5066 /* A retransmit, 2nd most common case. Force an immediate ack. */
5067 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
5068 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
5071 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
5072 inet_csk_schedule_ack(sk);
5078 /* Out of window. F.e. zero window probe. */
5079 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
5082 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
5083 /* Partial packet, seq < rcv_next < end_seq */
5084 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
5086 /* If window is closed, drop tail of packet. But after
5087 * remembering D-SACK for its head made in previous line.
5089 if (!tcp_receive_window(tp)) {
5090 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP);
5096 tcp_data_queue_ofo(sk, skb);
5099 static struct sk_buff *tcp_skb_next(struct sk_buff *skb, struct sk_buff_head *list)
5102 return !skb_queue_is_last(list, skb) ? skb->next : NULL;
5104 return skb_rb_next(skb);
5107 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
5108 struct sk_buff_head *list,
5109 struct rb_root *root)
5111 struct sk_buff *next = tcp_skb_next(skb, list);
5114 __skb_unlink(skb, list);
5116 rb_erase(&skb->rbnode, root);
5119 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
5124 /* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */
5125 void tcp_rbtree_insert(struct rb_root *root, struct sk_buff *skb)
5127 struct rb_node **p = &root->rb_node;
5128 struct rb_node *parent = NULL;
5129 struct sk_buff *skb1;
5133 skb1 = rb_to_skb(parent);
5134 if (before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb1)->seq))
5135 p = &parent->rb_left;
5137 p = &parent->rb_right;
5139 rb_link_node(&skb->rbnode, parent, p);
5140 rb_insert_color(&skb->rbnode, root);
5143 /* Collapse contiguous sequence of skbs head..tail with
5144 * sequence numbers start..end.
5146 * If tail is NULL, this means until the end of the queue.
5148 * Segments with FIN/SYN are not collapsed (only because this
5152 tcp_collapse(struct sock *sk, struct sk_buff_head *list, struct rb_root *root,
5153 struct sk_buff *head, struct sk_buff *tail, u32 start, u32 end)
5155 struct sk_buff *skb = head, *n;
5156 struct sk_buff_head tmp;
5159 /* First, check that queue is collapsible and find
5160 * the point where collapsing can be useful.
5163 for (end_of_skbs = true; skb != NULL && skb != tail; skb = n) {
5164 n = tcp_skb_next(skb, list);
5166 /* No new bits? It is possible on ofo queue. */
5167 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
5168 skb = tcp_collapse_one(sk, skb, list, root);
5174 /* The first skb to collapse is:
5176 * - bloated or contains data before "start" or
5177 * overlaps to the next one and mptcp allow collapsing.
5179 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
5180 (tcp_win_from_space(sk, skb->truesize) > skb->len ||
5181 before(TCP_SKB_CB(skb)->seq, start))) {
5182 end_of_skbs = false;
5186 if (n && n != tail && mptcp_skb_can_collapse(skb, n) &&
5187 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(n)->seq) {
5188 end_of_skbs = false;
5192 /* Decided to skip this, advance start seq. */
5193 start = TCP_SKB_CB(skb)->end_seq;
5196 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
5199 __skb_queue_head_init(&tmp);
5201 while (before(start, end)) {
5202 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
5203 struct sk_buff *nskb;
5205 nskb = alloc_skb(copy, GFP_ATOMIC);
5209 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
5210 #ifdef CONFIG_TLS_DEVICE
5211 nskb->decrypted = skb->decrypted;
5213 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
5215 __skb_queue_before(list, skb, nskb);
5217 __skb_queue_tail(&tmp, nskb); /* defer rbtree insertion */
5218 skb_set_owner_r(nskb, sk);
5219 mptcp_skb_ext_move(nskb, skb);
5221 /* Copy data, releasing collapsed skbs. */
5223 int offset = start - TCP_SKB_CB(skb)->seq;
5224 int size = TCP_SKB_CB(skb)->end_seq - start;
5228 size = min(copy, size);
5229 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
5231 TCP_SKB_CB(nskb)->end_seq += size;
5235 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
5236 skb = tcp_collapse_one(sk, skb, list, root);
5239 !mptcp_skb_can_collapse(nskb, skb) ||
5240 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
5242 #ifdef CONFIG_TLS_DEVICE
5243 if (skb->decrypted != nskb->decrypted)
5250 skb_queue_walk_safe(&tmp, skb, n)
5251 tcp_rbtree_insert(root, skb);
5254 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
5255 * and tcp_collapse() them until all the queue is collapsed.
5257 static void tcp_collapse_ofo_queue(struct sock *sk)
5259 struct tcp_sock *tp = tcp_sk(sk);
5260 u32 range_truesize, sum_tiny = 0;
5261 struct sk_buff *skb, *head;
5264 skb = skb_rb_first(&tp->out_of_order_queue);
5267 tp->ooo_last_skb = skb_rb_last(&tp->out_of_order_queue);
5270 start = TCP_SKB_CB(skb)->seq;
5271 end = TCP_SKB_CB(skb)->end_seq;
5272 range_truesize = skb->truesize;
5274 for (head = skb;;) {
5275 skb = skb_rb_next(skb);
5277 /* Range is terminated when we see a gap or when
5278 * we are at the queue end.
5281 after(TCP_SKB_CB(skb)->seq, end) ||
5282 before(TCP_SKB_CB(skb)->end_seq, start)) {
5283 /* Do not attempt collapsing tiny skbs */
5284 if (range_truesize != head->truesize ||
5285 end - start >= SKB_WITH_OVERHEAD(SK_MEM_QUANTUM)) {
5286 tcp_collapse(sk, NULL, &tp->out_of_order_queue,
5287 head, skb, start, end);
5289 sum_tiny += range_truesize;
5290 if (sum_tiny > sk->sk_rcvbuf >> 3)
5296 range_truesize += skb->truesize;
5297 if (unlikely(before(TCP_SKB_CB(skb)->seq, start)))
5298 start = TCP_SKB_CB(skb)->seq;
5299 if (after(TCP_SKB_CB(skb)->end_seq, end))
5300 end = TCP_SKB_CB(skb)->end_seq;
5305 * Clean the out-of-order queue to make room.
5306 * We drop high sequences packets to :
5307 * 1) Let a chance for holes to be filled.
5308 * 2) not add too big latencies if thousands of packets sit there.
5309 * (But if application shrinks SO_RCVBUF, we could still end up
5310 * freeing whole queue here)
5311 * 3) Drop at least 12.5 % of sk_rcvbuf to avoid malicious attacks.
5313 * Return true if queue has shrunk.
5315 static bool tcp_prune_ofo_queue(struct sock *sk)
5317 struct tcp_sock *tp = tcp_sk(sk);
5318 struct rb_node *node, *prev;
5321 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
5324 NET_INC_STATS(sock_net(sk), LINUX_MIB_OFOPRUNED);
5325 goal = sk->sk_rcvbuf >> 3;
5326 node = &tp->ooo_last_skb->rbnode;
5328 prev = rb_prev(node);
5329 rb_erase(node, &tp->out_of_order_queue);
5330 goal -= rb_to_skb(node)->truesize;
5331 tcp_drop(sk, rb_to_skb(node));
5332 if (!prev || goal <= 0) {
5334 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf &&
5335 !tcp_under_memory_pressure(sk))
5337 goal = sk->sk_rcvbuf >> 3;
5341 tp->ooo_last_skb = rb_to_skb(prev);
5343 /* Reset SACK state. A conforming SACK implementation will
5344 * do the same at a timeout based retransmit. When a connection
5345 * is in a sad state like this, we care only about integrity
5346 * of the connection not performance.
5348 if (tp->rx_opt.sack_ok)
5349 tcp_sack_reset(&tp->rx_opt);
5353 /* Reduce allocated memory if we can, trying to get
5354 * the socket within its memory limits again.
5356 * Return less than zero if we should start dropping frames
5357 * until the socket owning process reads some of the data
5358 * to stabilize the situation.
5360 static int tcp_prune_queue(struct sock *sk)
5362 struct tcp_sock *tp = tcp_sk(sk);
5364 NET_INC_STATS(sock_net(sk), LINUX_MIB_PRUNECALLED);
5366 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
5367 tcp_clamp_window(sk);
5368 else if (tcp_under_memory_pressure(sk))
5369 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
5371 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5374 tcp_collapse_ofo_queue(sk);
5375 if (!skb_queue_empty(&sk->sk_receive_queue))
5376 tcp_collapse(sk, &sk->sk_receive_queue, NULL,
5377 skb_peek(&sk->sk_receive_queue),
5379 tp->copied_seq, tp->rcv_nxt);
5382 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5385 /* Collapsing did not help, destructive actions follow.
5386 * This must not ever occur. */
5388 tcp_prune_ofo_queue(sk);
5390 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5393 /* If we are really being abused, tell the caller to silently
5394 * drop receive data on the floor. It will get retransmitted
5395 * and hopefully then we'll have sufficient space.
5397 NET_INC_STATS(sock_net(sk), LINUX_MIB_RCVPRUNED);
5399 /* Massive buffer overcommit. */
5404 static bool tcp_should_expand_sndbuf(const struct sock *sk)
5406 const struct tcp_sock *tp = tcp_sk(sk);
5408 /* If the user specified a specific send buffer setting, do
5411 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
5414 /* If we are under global TCP memory pressure, do not expand. */
5415 if (tcp_under_memory_pressure(sk))
5418 /* If we are under soft global TCP memory pressure, do not expand. */
5419 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
5422 /* If we filled the congestion window, do not expand. */
5423 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
5429 static void tcp_new_space(struct sock *sk)
5431 struct tcp_sock *tp = tcp_sk(sk);
5433 if (tcp_should_expand_sndbuf(sk)) {
5434 tcp_sndbuf_expand(sk);
5435 tp->snd_cwnd_stamp = tcp_jiffies32;
5438 sk->sk_write_space(sk);
5441 /* Caller made space either from:
5442 * 1) Freeing skbs in rtx queues (after tp->snd_una has advanced)
5443 * 2) Sent skbs from output queue (and thus advancing tp->snd_nxt)
5445 * We might be able to generate EPOLLOUT to the application if:
5446 * 1) Space consumed in output/rtx queues is below sk->sk_sndbuf/2
5447 * 2) notsent amount (tp->write_seq - tp->snd_nxt) became
5448 * small enough that tcp_stream_memory_free() decides it
5449 * is time to generate EPOLLOUT.
5451 void tcp_check_space(struct sock *sk)
5453 /* pairs with tcp_poll() */
5455 if (sk->sk_socket &&
5456 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
5458 if (!test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
5459 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
5463 static inline void tcp_data_snd_check(struct sock *sk)
5465 tcp_push_pending_frames(sk);
5466 tcp_check_space(sk);
5470 * Check if sending an ack is needed.
5472 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
5474 struct tcp_sock *tp = tcp_sk(sk);
5475 unsigned long rtt, delay;
5477 /* More than one full frame received... */
5478 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
5479 /* ... and right edge of window advances far enough.
5480 * (tcp_recvmsg() will send ACK otherwise).
5481 * If application uses SO_RCVLOWAT, we want send ack now if
5482 * we have not received enough bytes to satisfy the condition.
5484 (tp->rcv_nxt - tp->copied_seq < sk->sk_rcvlowat ||
5485 __tcp_select_window(sk) >= tp->rcv_wnd)) ||
5486 /* We ACK each frame or... */
5487 tcp_in_quickack_mode(sk) ||
5488 /* Protocol state mandates a one-time immediate ACK */
5489 inet_csk(sk)->icsk_ack.pending & ICSK_ACK_NOW) {
5495 if (!ofo_possible || RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
5496 tcp_send_delayed_ack(sk);
5500 if (!tcp_is_sack(tp) ||
5501 tp->compressed_ack >= READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_nr))
5504 if (tp->compressed_ack_rcv_nxt != tp->rcv_nxt) {
5505 tp->compressed_ack_rcv_nxt = tp->rcv_nxt;
5506 tp->dup_ack_counter = 0;
5508 if (tp->dup_ack_counter < TCP_FASTRETRANS_THRESH) {
5509 tp->dup_ack_counter++;
5512 tp->compressed_ack++;
5513 if (hrtimer_is_queued(&tp->compressed_ack_timer))
5516 /* compress ack timer : 5 % of rtt, but no more than tcp_comp_sack_delay_ns */
5518 rtt = tp->rcv_rtt_est.rtt_us;
5519 if (tp->srtt_us && tp->srtt_us < rtt)
5522 delay = min_t(unsigned long,
5523 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_delay_ns),
5524 rtt * (NSEC_PER_USEC >> 3)/20);
5526 hrtimer_start_range_ns(&tp->compressed_ack_timer, ns_to_ktime(delay),
5527 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_slack_ns),
5528 HRTIMER_MODE_REL_PINNED_SOFT);
5531 static inline void tcp_ack_snd_check(struct sock *sk)
5533 if (!inet_csk_ack_scheduled(sk)) {
5534 /* We sent a data segment already. */
5537 __tcp_ack_snd_check(sk, 1);
5541 * This routine is only called when we have urgent data
5542 * signaled. Its the 'slow' part of tcp_urg. It could be
5543 * moved inline now as tcp_urg is only called from one
5544 * place. We handle URGent data wrong. We have to - as
5545 * BSD still doesn't use the correction from RFC961.
5546 * For 1003.1g we should support a new option TCP_STDURG to permit
5547 * either form (or just set the sysctl tcp_stdurg).
5550 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
5552 struct tcp_sock *tp = tcp_sk(sk);
5553 u32 ptr = ntohs(th->urg_ptr);
5555 if (ptr && !READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_stdurg))
5557 ptr += ntohl(th->seq);
5559 /* Ignore urgent data that we've already seen and read. */
5560 if (after(tp->copied_seq, ptr))
5563 /* Do not replay urg ptr.
5565 * NOTE: interesting situation not covered by specs.
5566 * Misbehaving sender may send urg ptr, pointing to segment,
5567 * which we already have in ofo queue. We are not able to fetch
5568 * such data and will stay in TCP_URG_NOTYET until will be eaten
5569 * by recvmsg(). Seems, we are not obliged to handle such wicked
5570 * situations. But it is worth to think about possibility of some
5571 * DoSes using some hypothetical application level deadlock.
5573 if (before(ptr, tp->rcv_nxt))
5576 /* Do we already have a newer (or duplicate) urgent pointer? */
5577 if (tp->urg_data && !after(ptr, tp->urg_seq))
5580 /* Tell the world about our new urgent pointer. */
5583 /* We may be adding urgent data when the last byte read was
5584 * urgent. To do this requires some care. We cannot just ignore
5585 * tp->copied_seq since we would read the last urgent byte again
5586 * as data, nor can we alter copied_seq until this data arrives
5587 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5589 * NOTE. Double Dutch. Rendering to plain English: author of comment
5590 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5591 * and expect that both A and B disappear from stream. This is _wrong_.
5592 * Though this happens in BSD with high probability, this is occasional.
5593 * Any application relying on this is buggy. Note also, that fix "works"
5594 * only in this artificial test. Insert some normal data between A and B and we will
5595 * decline of BSD again. Verdict: it is better to remove to trap
5598 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5599 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5600 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5602 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5603 __skb_unlink(skb, &sk->sk_receive_queue);
5608 tp->urg_data = TCP_URG_NOTYET;
5609 WRITE_ONCE(tp->urg_seq, ptr);
5611 /* Disable header prediction. */
5615 /* This is the 'fast' part of urgent handling. */
5616 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
5618 struct tcp_sock *tp = tcp_sk(sk);
5620 /* Check if we get a new urgent pointer - normally not. */
5622 tcp_check_urg(sk, th);
5624 /* Do we wait for any urgent data? - normally not... */
5625 if (tp->urg_data == TCP_URG_NOTYET) {
5626 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5629 /* Is the urgent pointer pointing into this packet? */
5630 if (ptr < skb->len) {
5632 if (skb_copy_bits(skb, ptr, &tmp, 1))
5634 tp->urg_data = TCP_URG_VALID | tmp;
5635 if (!sock_flag(sk, SOCK_DEAD))
5636 sk->sk_data_ready(sk);
5641 /* Accept RST for rcv_nxt - 1 after a FIN.
5642 * When tcp connections are abruptly terminated from Mac OSX (via ^C), a
5643 * FIN is sent followed by a RST packet. The RST is sent with the same
5644 * sequence number as the FIN, and thus according to RFC 5961 a challenge
5645 * ACK should be sent. However, Mac OSX rate limits replies to challenge
5646 * ACKs on the closed socket. In addition middleboxes can drop either the
5647 * challenge ACK or a subsequent RST.
5649 static bool tcp_reset_check(const struct sock *sk, const struct sk_buff *skb)
5651 struct tcp_sock *tp = tcp_sk(sk);
5653 return unlikely(TCP_SKB_CB(skb)->seq == (tp->rcv_nxt - 1) &&
5654 (1 << sk->sk_state) & (TCPF_CLOSE_WAIT | TCPF_LAST_ACK |
5658 /* Does PAWS and seqno based validation of an incoming segment, flags will
5659 * play significant role here.
5661 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5662 const struct tcphdr *th, int syn_inerr)
5664 struct tcp_sock *tp = tcp_sk(sk);
5665 bool rst_seq_match = false;
5667 /* RFC1323: H1. Apply PAWS check first. */
5668 if (tcp_fast_parse_options(sock_net(sk), skb, th, tp) &&
5669 tp->rx_opt.saw_tstamp &&
5670 tcp_paws_discard(sk, skb)) {
5672 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5673 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5674 LINUX_MIB_TCPACKSKIPPEDPAWS,
5675 &tp->last_oow_ack_time))
5676 tcp_send_dupack(sk, skb);
5679 /* Reset is accepted even if it did not pass PAWS. */
5682 /* Step 1: check sequence number */
5683 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5684 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5685 * (RST) segments are validated by checking their SEQ-fields."
5686 * And page 69: "If an incoming segment is not acceptable,
5687 * an acknowledgment should be sent in reply (unless the RST
5688 * bit is set, if so drop the segment and return)".
5693 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5694 LINUX_MIB_TCPACKSKIPPEDSEQ,
5695 &tp->last_oow_ack_time))
5696 tcp_send_dupack(sk, skb);
5697 } else if (tcp_reset_check(sk, skb)) {
5703 /* Step 2: check RST bit */
5705 /* RFC 5961 3.2 (extend to match against (RCV.NXT - 1) after a
5706 * FIN and SACK too if available):
5707 * If seq num matches RCV.NXT or (RCV.NXT - 1) after a FIN, or
5708 * the right-most SACK block,
5710 * RESET the connection
5712 * Send a challenge ACK
5714 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt ||
5715 tcp_reset_check(sk, skb)) {
5716 rst_seq_match = true;
5717 } else if (tcp_is_sack(tp) && tp->rx_opt.num_sacks > 0) {
5718 struct tcp_sack_block *sp = &tp->selective_acks[0];
5719 int max_sack = sp[0].end_seq;
5722 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;
5724 max_sack = after(sp[this_sack].end_seq,
5726 sp[this_sack].end_seq : max_sack;
5729 if (TCP_SKB_CB(skb)->seq == max_sack)
5730 rst_seq_match = true;
5736 /* Disable TFO if RST is out-of-order
5737 * and no data has been received
5738 * for current active TFO socket
5740 if (tp->syn_fastopen && !tp->data_segs_in &&
5741 sk->sk_state == TCP_ESTABLISHED)
5742 tcp_fastopen_active_disable(sk);
5743 tcp_send_challenge_ack(sk, skb);
5748 /* step 3: check security and precedence [ignored] */
5750 /* step 4: Check for a SYN
5751 * RFC 5961 4.2 : Send a challenge ack
5756 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5757 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5758 tcp_send_challenge_ack(sk, skb);
5762 bpf_skops_parse_hdr(sk, skb);
5772 * TCP receive function for the ESTABLISHED state.
5774 * It is split into a fast path and a slow path. The fast path is
5776 * - A zero window was announced from us - zero window probing
5777 * is only handled properly in the slow path.
5778 * - Out of order segments arrived.
5779 * - Urgent data is expected.
5780 * - There is no buffer space left
5781 * - Unexpected TCP flags/window values/header lengths are received
5782 * (detected by checking the TCP header against pred_flags)
5783 * - Data is sent in both directions. Fast path only supports pure senders
5784 * or pure receivers (this means either the sequence number or the ack
5785 * value must stay constant)
5786 * - Unexpected TCP option.
5788 * When these conditions are not satisfied it drops into a standard
5789 * receive procedure patterned after RFC793 to handle all cases.
5790 * The first three cases are guaranteed by proper pred_flags setting,
5791 * the rest is checked inline. Fast processing is turned on in
5792 * tcp_data_queue when everything is OK.
5794 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb)
5796 const struct tcphdr *th = (const struct tcphdr *)skb->data;
5797 struct tcp_sock *tp = tcp_sk(sk);
5798 unsigned int len = skb->len;
5800 /* TCP congestion window tracking */
5801 trace_tcp_probe(sk, skb);
5803 tcp_mstamp_refresh(tp);
5804 if (unlikely(!rcu_access_pointer(sk->sk_rx_dst)))
5805 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
5807 * Header prediction.
5808 * The code loosely follows the one in the famous
5809 * "30 instruction TCP receive" Van Jacobson mail.
5811 * Van's trick is to deposit buffers into socket queue
5812 * on a device interrupt, to call tcp_recv function
5813 * on the receive process context and checksum and copy
5814 * the buffer to user space. smart...
5816 * Our current scheme is not silly either but we take the
5817 * extra cost of the net_bh soft interrupt processing...
5818 * We do checksum and copy also but from device to kernel.
5821 tp->rx_opt.saw_tstamp = 0;
5823 /* pred_flags is 0xS?10 << 16 + snd_wnd
5824 * if header_prediction is to be made
5825 * 'S' will always be tp->tcp_header_len >> 2
5826 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5827 * turn it off (when there are holes in the receive
5828 * space for instance)
5829 * PSH flag is ignored.
5832 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5833 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5834 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5835 int tcp_header_len = tp->tcp_header_len;
5837 /* Timestamp header prediction: tcp_header_len
5838 * is automatically equal to th->doff*4 due to pred_flags
5842 /* Check timestamp */
5843 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5844 /* No? Slow path! */
5845 if (!tcp_parse_aligned_timestamp(tp, th))
5848 /* If PAWS failed, check it more carefully in slow path */
5849 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5852 /* DO NOT update ts_recent here, if checksum fails
5853 * and timestamp was corrupted part, it will result
5854 * in a hung connection since we will drop all
5855 * future packets due to the PAWS test.
5859 if (len <= tcp_header_len) {
5860 /* Bulk data transfer: sender */
5861 if (len == tcp_header_len) {
5862 /* Predicted packet is in window by definition.
5863 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5864 * Hence, check seq<=rcv_wup reduces to:
5866 if (tcp_header_len ==
5867 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5868 tp->rcv_nxt == tp->rcv_wup)
5869 tcp_store_ts_recent(tp);
5871 /* We know that such packets are checksummed
5874 tcp_ack(sk, skb, 0);
5876 tcp_data_snd_check(sk);
5877 /* When receiving pure ack in fast path, update
5878 * last ts ecr directly instead of calling
5879 * tcp_rcv_rtt_measure_ts()
5881 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr;
5883 } else { /* Header too small */
5884 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5889 bool fragstolen = false;
5891 if (tcp_checksum_complete(skb))
5894 if ((int)skb->truesize > sk->sk_forward_alloc)
5897 /* Predicted packet is in window by definition.
5898 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5899 * Hence, check seq<=rcv_wup reduces to:
5901 if (tcp_header_len ==
5902 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5903 tp->rcv_nxt == tp->rcv_wup)
5904 tcp_store_ts_recent(tp);
5906 tcp_rcv_rtt_measure_ts(sk, skb);
5908 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPHITS);
5910 /* Bulk data transfer: receiver */
5911 __skb_pull(skb, tcp_header_len);
5912 eaten = tcp_queue_rcv(sk, skb, &fragstolen);
5914 tcp_event_data_recv(sk, skb);
5916 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5917 /* Well, only one small jumplet in fast path... */
5918 tcp_ack(sk, skb, FLAG_DATA);
5919 tcp_data_snd_check(sk);
5920 if (!inet_csk_ack_scheduled(sk))
5923 tcp_update_wl(tp, TCP_SKB_CB(skb)->seq);
5926 __tcp_ack_snd_check(sk, 0);
5929 kfree_skb_partial(skb, fragstolen);
5936 if (len < (th->doff << 2) || tcp_checksum_complete(skb))
5939 if (!th->ack && !th->rst && !th->syn)
5943 * Standard slow path.
5946 if (!tcp_validate_incoming(sk, skb, th, 1))
5950 if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0)
5953 tcp_rcv_rtt_measure_ts(sk, skb);
5955 /* Process urgent data. */
5956 tcp_urg(sk, skb, th);
5958 /* step 7: process the segment text */
5959 tcp_data_queue(sk, skb);
5961 tcp_data_snd_check(sk);
5962 tcp_ack_snd_check(sk);
5966 TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS);
5967 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5972 EXPORT_SYMBOL(tcp_rcv_established);
5974 void tcp_init_transfer(struct sock *sk, int bpf_op, struct sk_buff *skb)
5976 struct inet_connection_sock *icsk = inet_csk(sk);
5977 struct tcp_sock *tp = tcp_sk(sk);
5980 icsk->icsk_af_ops->rebuild_header(sk);
5981 tcp_init_metrics(sk);
5983 /* Initialize the congestion window to start the transfer.
5984 * Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
5985 * retransmitted. In light of RFC6298 more aggressive 1sec
5986 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
5987 * retransmission has occurred.
5989 if (tp->total_retrans > 1 && tp->undo_marker)
5992 tp->snd_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk));
5993 tp->snd_cwnd_stamp = tcp_jiffies32;
5995 bpf_skops_established(sk, bpf_op, skb);
5996 /* Initialize congestion control unless BPF initialized it already: */
5997 if (!icsk->icsk_ca_initialized)
5998 tcp_init_congestion_control(sk);
5999 tcp_init_buffer_space(sk);
6002 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
6004 struct tcp_sock *tp = tcp_sk(sk);
6005 struct inet_connection_sock *icsk = inet_csk(sk);
6007 tcp_set_state(sk, TCP_ESTABLISHED);
6008 icsk->icsk_ack.lrcvtime = tcp_jiffies32;
6011 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
6012 security_inet_conn_established(sk, skb);
6013 sk_mark_napi_id(sk, skb);
6016 tcp_init_transfer(sk, BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB, skb);
6018 /* Prevent spurious tcp_cwnd_restart() on first data
6021 tp->lsndtime = tcp_jiffies32;
6023 if (sock_flag(sk, SOCK_KEEPOPEN))
6024 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
6026 if (!tp->rx_opt.snd_wscale)
6027 __tcp_fast_path_on(tp, tp->snd_wnd);
6032 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
6033 struct tcp_fastopen_cookie *cookie)
6035 struct tcp_sock *tp = tcp_sk(sk);
6036 struct sk_buff *data = tp->syn_data ? tcp_rtx_queue_head(sk) : NULL;
6037 u16 mss = tp->rx_opt.mss_clamp, try_exp = 0;
6038 bool syn_drop = false;
6040 if (mss == tp->rx_opt.user_mss) {
6041 struct tcp_options_received opt;
6043 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
6044 tcp_clear_options(&opt);
6045 opt.user_mss = opt.mss_clamp = 0;
6046 tcp_parse_options(sock_net(sk), synack, &opt, 0, NULL);
6047 mss = opt.mss_clamp;
6050 if (!tp->syn_fastopen) {
6051 /* Ignore an unsolicited cookie */
6053 } else if (tp->total_retrans) {
6054 /* SYN timed out and the SYN-ACK neither has a cookie nor
6055 * acknowledges data. Presumably the remote received only
6056 * the retransmitted (regular) SYNs: either the original
6057 * SYN-data or the corresponding SYN-ACK was dropped.
6059 syn_drop = (cookie->len < 0 && data);
6060 } else if (cookie->len < 0 && !tp->syn_data) {
6061 /* We requested a cookie but didn't get it. If we did not use
6062 * the (old) exp opt format then try so next time (try_exp=1).
6063 * Otherwise we go back to use the RFC7413 opt (try_exp=2).
6065 try_exp = tp->syn_fastopen_exp ? 2 : 1;
6068 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp);
6070 if (data) { /* Retransmit unacked data in SYN */
6071 if (tp->total_retrans)
6072 tp->fastopen_client_fail = TFO_SYN_RETRANSMITTED;
6074 tp->fastopen_client_fail = TFO_DATA_NOT_ACKED;
6075 skb_rbtree_walk_from(data) {
6076 if (__tcp_retransmit_skb(sk, data, 1))
6080 NET_INC_STATS(sock_net(sk),
6081 LINUX_MIB_TCPFASTOPENACTIVEFAIL);
6084 tp->syn_data_acked = tp->syn_data;
6085 if (tp->syn_data_acked) {
6086 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
6087 /* SYN-data is counted as two separate packets in tcp_ack() */
6088 if (tp->delivered > 1)
6092 tcp_fastopen_add_skb(sk, synack);
6097 static void smc_check_reset_syn(struct tcp_sock *tp)
6099 #if IS_ENABLED(CONFIG_SMC)
6100 if (static_branch_unlikely(&tcp_have_smc)) {
6101 if (tp->syn_smc && !tp->rx_opt.smc_ok)
6107 static void tcp_try_undo_spurious_syn(struct sock *sk)
6109 struct tcp_sock *tp = tcp_sk(sk);
6112 /* undo_marker is set when SYN or SYNACK times out. The timeout is
6113 * spurious if the ACK's timestamp option echo value matches the
6114 * original SYN timestamp.
6116 syn_stamp = tp->retrans_stamp;
6117 if (tp->undo_marker && syn_stamp && tp->rx_opt.saw_tstamp &&
6118 syn_stamp == tp->rx_opt.rcv_tsecr)
6119 tp->undo_marker = 0;
6122 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
6123 const struct tcphdr *th)
6125 struct inet_connection_sock *icsk = inet_csk(sk);
6126 struct tcp_sock *tp = tcp_sk(sk);
6127 struct tcp_fastopen_cookie foc = { .len = -1 };
6128 int saved_clamp = tp->rx_opt.mss_clamp;
6131 tcp_parse_options(sock_net(sk), skb, &tp->rx_opt, 0, &foc);
6132 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
6133 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
6137 * "If the state is SYN-SENT then
6138 * first check the ACK bit
6139 * If the ACK bit is set
6140 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
6141 * a reset (unless the RST bit is set, if so drop
6142 * the segment and return)"
6144 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
6145 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
6146 /* Previous FIN/ACK or RST/ACK might be ignored. */
6147 if (icsk->icsk_retransmits == 0)
6148 inet_csk_reset_xmit_timer(sk,
6150 TCP_TIMEOUT_MIN, TCP_RTO_MAX);
6151 goto reset_and_undo;
6154 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
6155 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
6156 tcp_time_stamp(tp))) {
6157 NET_INC_STATS(sock_net(sk),
6158 LINUX_MIB_PAWSACTIVEREJECTED);
6159 goto reset_and_undo;
6162 /* Now ACK is acceptable.
6164 * "If the RST bit is set
6165 * If the ACK was acceptable then signal the user "error:
6166 * connection reset", drop the segment, enter CLOSED state,
6167 * delete TCB, and return."
6176 * "fifth, if neither of the SYN or RST bits is set then
6177 * drop the segment and return."
6183 goto discard_and_undo;
6186 * "If the SYN bit is on ...
6187 * are acceptable then ...
6188 * (our SYN has been ACKed), change the connection
6189 * state to ESTABLISHED..."
6192 tcp_ecn_rcv_synack(tp, th);
6194 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
6195 tcp_try_undo_spurious_syn(sk);
6196 tcp_ack(sk, skb, FLAG_SLOWPATH);
6198 /* Ok.. it's good. Set up sequence numbers and
6199 * move to established.
6201 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1);
6202 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
6204 /* RFC1323: The window in SYN & SYN/ACK segments is
6207 tp->snd_wnd = ntohs(th->window);
6209 if (!tp->rx_opt.wscale_ok) {
6210 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
6211 tp->window_clamp = min(tp->window_clamp, 65535U);
6214 if (tp->rx_opt.saw_tstamp) {
6215 tp->rx_opt.tstamp_ok = 1;
6216 tp->tcp_header_len =
6217 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
6218 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
6219 tcp_store_ts_recent(tp);
6221 tp->tcp_header_len = sizeof(struct tcphdr);
6224 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
6225 tcp_initialize_rcv_mss(sk);
6227 /* Remember, tcp_poll() does not lock socket!
6228 * Change state from SYN-SENT only after copied_seq
6229 * is initialized. */
6230 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6232 smc_check_reset_syn(tp);
6236 tcp_finish_connect(sk, skb);
6238 fastopen_fail = (tp->syn_fastopen || tp->syn_data) &&
6239 tcp_rcv_fastopen_synack(sk, skb, &foc);
6241 if (!sock_flag(sk, SOCK_DEAD)) {
6242 sk->sk_state_change(sk);
6243 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
6247 if (sk->sk_write_pending ||
6248 icsk->icsk_accept_queue.rskq_defer_accept ||
6249 inet_csk_in_pingpong_mode(sk)) {
6250 /* Save one ACK. Data will be ready after
6251 * several ticks, if write_pending is set.
6253 * It may be deleted, but with this feature tcpdumps
6254 * look so _wonderfully_ clever, that I was not able
6255 * to stand against the temptation 8) --ANK
6257 inet_csk_schedule_ack(sk);
6258 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
6259 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
6260 TCP_DELACK_MAX, TCP_RTO_MAX);
6271 /* No ACK in the segment */
6275 * "If the RST bit is set
6277 * Otherwise (no ACK) drop the segment and return."
6280 goto discard_and_undo;
6284 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
6285 tcp_paws_reject(&tp->rx_opt, 0))
6286 goto discard_and_undo;
6289 /* We see SYN without ACK. It is attempt of
6290 * simultaneous connect with crossed SYNs.
6291 * Particularly, it can be connect to self.
6293 tcp_set_state(sk, TCP_SYN_RECV);
6295 if (tp->rx_opt.saw_tstamp) {
6296 tp->rx_opt.tstamp_ok = 1;
6297 tcp_store_ts_recent(tp);
6298 tp->tcp_header_len =
6299 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
6301 tp->tcp_header_len = sizeof(struct tcphdr);
6304 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1);
6305 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6306 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
6308 /* RFC1323: The window in SYN & SYN/ACK segments is
6311 tp->snd_wnd = ntohs(th->window);
6312 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
6313 tp->max_window = tp->snd_wnd;
6315 tcp_ecn_rcv_syn(tp, th);
6318 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
6319 tcp_initialize_rcv_mss(sk);
6321 tcp_send_synack(sk);
6323 /* Note, we could accept data and URG from this segment.
6324 * There are no obstacles to make this (except that we must
6325 * either change tcp_recvmsg() to prevent it from returning data
6326 * before 3WHS completes per RFC793, or employ TCP Fast Open).
6328 * However, if we ignore data in ACKless segments sometimes,
6329 * we have no reasons to accept it sometimes.
6330 * Also, seems the code doing it in step6 of tcp_rcv_state_process
6331 * is not flawless. So, discard packet for sanity.
6332 * Uncomment this return to process the data.
6339 /* "fifth, if neither of the SYN or RST bits is set then
6340 * drop the segment and return."
6344 tcp_clear_options(&tp->rx_opt);
6345 tp->rx_opt.mss_clamp = saved_clamp;
6349 tcp_clear_options(&tp->rx_opt);
6350 tp->rx_opt.mss_clamp = saved_clamp;
6354 static void tcp_rcv_synrecv_state_fastopen(struct sock *sk)
6356 struct tcp_sock *tp = tcp_sk(sk);
6357 struct request_sock *req;
6359 /* If we are still handling the SYNACK RTO, see if timestamp ECR allows
6360 * undo. If peer SACKs triggered fast recovery, we can't undo here.
6362 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss && !tp->packets_out)
6363 tcp_try_undo_recovery(sk);
6365 /* Reset rtx states to prevent spurious retransmits_timed_out() */
6366 tp->retrans_stamp = 0;
6367 inet_csk(sk)->icsk_retransmits = 0;
6369 /* Once we leave TCP_SYN_RECV or TCP_FIN_WAIT_1,
6370 * we no longer need req so release it.
6372 req = rcu_dereference_protected(tp->fastopen_rsk,
6373 lockdep_sock_is_held(sk));
6374 reqsk_fastopen_remove(sk, req, false);
6376 /* Re-arm the timer because data may have been sent out.
6377 * This is similar to the regular data transmission case
6378 * when new data has just been ack'ed.
6380 * (TFO) - we could try to be more aggressive and
6381 * retransmitting any data sooner based on when they
6388 * This function implements the receiving procedure of RFC 793 for
6389 * all states except ESTABLISHED and TIME_WAIT.
6390 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
6391 * address independent.
6394 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb)
6396 struct tcp_sock *tp = tcp_sk(sk);
6397 struct inet_connection_sock *icsk = inet_csk(sk);
6398 const struct tcphdr *th = tcp_hdr(skb);
6399 struct request_sock *req;
6403 switch (sk->sk_state) {
6417 /* It is possible that we process SYN packets from backlog,
6418 * so we need to make sure to disable BH and RCU right there.
6422 acceptable = icsk->icsk_af_ops->conn_request(sk, skb) >= 0;
6434 tp->rx_opt.saw_tstamp = 0;
6435 tcp_mstamp_refresh(tp);
6436 queued = tcp_rcv_synsent_state_process(sk, skb, th);
6440 /* Do step6 onward by hand. */
6441 tcp_urg(sk, skb, th);
6443 tcp_data_snd_check(sk);
6447 tcp_mstamp_refresh(tp);
6448 tp->rx_opt.saw_tstamp = 0;
6449 req = rcu_dereference_protected(tp->fastopen_rsk,
6450 lockdep_sock_is_held(sk));
6454 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
6455 sk->sk_state != TCP_FIN_WAIT1);
6457 if (!tcp_check_req(sk, skb, req, true, &req_stolen))
6461 if (!th->ack && !th->rst && !th->syn)
6464 if (!tcp_validate_incoming(sk, skb, th, 0))
6467 /* step 5: check the ACK field */
6468 acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
6469 FLAG_UPDATE_TS_RECENT |
6470 FLAG_NO_CHALLENGE_ACK) > 0;
6473 if (sk->sk_state == TCP_SYN_RECV)
6474 return 1; /* send one RST */
6475 tcp_send_challenge_ack(sk, skb);
6478 switch (sk->sk_state) {
6480 tp->delivered++; /* SYN-ACK delivery isn't tracked in tcp_ack */
6482 tcp_synack_rtt_meas(sk, req);
6485 tcp_rcv_synrecv_state_fastopen(sk);
6487 tcp_try_undo_spurious_syn(sk);
6488 tp->retrans_stamp = 0;
6489 tcp_init_transfer(sk, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB,
6491 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6494 tcp_set_state(sk, TCP_ESTABLISHED);
6495 sk->sk_state_change(sk);
6497 /* Note, that this wakeup is only for marginal crossed SYN case.
6498 * Passively open sockets are not waked up, because
6499 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
6502 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
6504 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
6505 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
6506 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
6508 if (tp->rx_opt.tstamp_ok)
6509 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
6511 if (!inet_csk(sk)->icsk_ca_ops->cong_control)
6512 tcp_update_pacing_rate(sk);
6514 /* Prevent spurious tcp_cwnd_restart() on first data packet */
6515 tp->lsndtime = tcp_jiffies32;
6517 tcp_initialize_rcv_mss(sk);
6518 tcp_fast_path_on(tp);
6519 if (sk->sk_shutdown & SEND_SHUTDOWN)
6520 tcp_shutdown(sk, SEND_SHUTDOWN);
6523 case TCP_FIN_WAIT1: {
6527 tcp_rcv_synrecv_state_fastopen(sk);
6529 if (tp->snd_una != tp->write_seq)
6532 tcp_set_state(sk, TCP_FIN_WAIT2);
6533 WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | SEND_SHUTDOWN);
6537 if (!sock_flag(sk, SOCK_DEAD)) {
6538 /* Wake up lingering close() */
6539 sk->sk_state_change(sk);
6543 if (tp->linger2 < 0) {
6545 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6548 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6549 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6550 /* Receive out of order FIN after close() */
6551 if (tp->syn_fastopen && th->fin)
6552 tcp_fastopen_active_disable(sk);
6554 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6558 tmo = tcp_fin_time(sk);
6559 if (tmo > TCP_TIMEWAIT_LEN) {
6560 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
6561 } else if (th->fin || sock_owned_by_user(sk)) {
6562 /* Bad case. We could lose such FIN otherwise.
6563 * It is not a big problem, but it looks confusing
6564 * and not so rare event. We still can lose it now,
6565 * if it spins in bh_lock_sock(), but it is really
6568 inet_csk_reset_keepalive_timer(sk, tmo);
6570 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
6577 if (tp->snd_una == tp->write_seq) {
6578 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
6584 if (tp->snd_una == tp->write_seq) {
6585 tcp_update_metrics(sk);
6592 /* step 6: check the URG bit */
6593 tcp_urg(sk, skb, th);
6595 /* step 7: process the segment text */
6596 switch (sk->sk_state) {
6597 case TCP_CLOSE_WAIT:
6600 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
6601 if (sk_is_mptcp(sk))
6602 mptcp_incoming_options(sk, skb);
6608 /* RFC 793 says to queue data in these states,
6609 * RFC 1122 says we MUST send a reset.
6610 * BSD 4.4 also does reset.
6612 if (sk->sk_shutdown & RCV_SHUTDOWN) {
6613 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6614 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6615 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6621 case TCP_ESTABLISHED:
6622 tcp_data_queue(sk, skb);
6627 /* tcp_data could move socket to TIME-WAIT */
6628 if (sk->sk_state != TCP_CLOSE) {
6629 tcp_data_snd_check(sk);
6630 tcp_ack_snd_check(sk);
6639 EXPORT_SYMBOL(tcp_rcv_state_process);
6641 static inline void pr_drop_req(struct request_sock *req, __u16 port, int family)
6643 struct inet_request_sock *ireq = inet_rsk(req);
6645 if (family == AF_INET)
6646 net_dbg_ratelimited("drop open request from %pI4/%u\n",
6647 &ireq->ir_rmt_addr, port);
6648 #if IS_ENABLED(CONFIG_IPV6)
6649 else if (family == AF_INET6)
6650 net_dbg_ratelimited("drop open request from %pI6/%u\n",
6651 &ireq->ir_v6_rmt_addr, port);
6655 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
6657 * If we receive a SYN packet with these bits set, it means a
6658 * network is playing bad games with TOS bits. In order to
6659 * avoid possible false congestion notifications, we disable
6660 * TCP ECN negotiation.
6662 * Exception: tcp_ca wants ECN. This is required for DCTCP
6663 * congestion control: Linux DCTCP asserts ECT on all packets,
6664 * including SYN, which is most optimal solution; however,
6665 * others, such as FreeBSD do not.
6667 * Exception: At least one of the reserved bits of the TCP header (th->res1) is
6668 * set, indicating the use of a future TCP extension (such as AccECN). See
6669 * RFC8311 §4.3 which updates RFC3168 to allow the development of such
6672 static void tcp_ecn_create_request(struct request_sock *req,
6673 const struct sk_buff *skb,
6674 const struct sock *listen_sk,
6675 const struct dst_entry *dst)
6677 const struct tcphdr *th = tcp_hdr(skb);
6678 const struct net *net = sock_net(listen_sk);
6679 bool th_ecn = th->ece && th->cwr;
6686 ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield);
6687 ecn_ok_dst = dst_feature(dst, DST_FEATURE_ECN_MASK);
6688 ecn_ok = net->ipv4.sysctl_tcp_ecn || ecn_ok_dst;
6690 if (((!ect || th->res1) && ecn_ok) || tcp_ca_needs_ecn(listen_sk) ||
6691 (ecn_ok_dst & DST_FEATURE_ECN_CA) ||
6692 tcp_bpf_ca_needs_ecn((struct sock *)req))
6693 inet_rsk(req)->ecn_ok = 1;
6696 static void tcp_openreq_init(struct request_sock *req,
6697 const struct tcp_options_received *rx_opt,
6698 struct sk_buff *skb, const struct sock *sk)
6700 struct inet_request_sock *ireq = inet_rsk(req);
6702 req->rsk_rcv_wnd = 0; /* So that tcp_send_synack() knows! */
6703 tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq;
6704 tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
6705 tcp_rsk(req)->snt_synack = 0;
6706 tcp_rsk(req)->last_oow_ack_time = 0;
6707 req->mss = rx_opt->mss_clamp;
6708 req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0;
6709 ireq->tstamp_ok = rx_opt->tstamp_ok;
6710 ireq->sack_ok = rx_opt->sack_ok;
6711 ireq->snd_wscale = rx_opt->snd_wscale;
6712 ireq->wscale_ok = rx_opt->wscale_ok;
6715 ireq->ir_rmt_port = tcp_hdr(skb)->source;
6716 ireq->ir_num = ntohs(tcp_hdr(skb)->dest);
6717 ireq->ir_mark = inet_request_mark(sk, skb);
6718 #if IS_ENABLED(CONFIG_SMC)
6719 ireq->smc_ok = rx_opt->smc_ok;
6723 struct request_sock *inet_reqsk_alloc(const struct request_sock_ops *ops,
6724 struct sock *sk_listener,
6725 bool attach_listener)
6727 struct request_sock *req = reqsk_alloc(ops, sk_listener,
6731 struct inet_request_sock *ireq = inet_rsk(req);
6733 ireq->ireq_opt = NULL;
6734 #if IS_ENABLED(CONFIG_IPV6)
6735 ireq->pktopts = NULL;
6737 atomic64_set(&ireq->ir_cookie, 0);
6738 ireq->ireq_state = TCP_NEW_SYN_RECV;
6739 write_pnet(&ireq->ireq_net, sock_net(sk_listener));
6740 ireq->ireq_family = sk_listener->sk_family;
6745 EXPORT_SYMBOL(inet_reqsk_alloc);
6748 * Return true if a syncookie should be sent
6750 static bool tcp_syn_flood_action(const struct sock *sk, const char *proto)
6752 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
6753 const char *msg = "Dropping request";
6754 struct net *net = sock_net(sk);
6755 bool want_cookie = false;
6758 syncookies = READ_ONCE(net->ipv4.sysctl_tcp_syncookies);
6760 #ifdef CONFIG_SYN_COOKIES
6762 msg = "Sending cookies";
6764 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES);
6767 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP);
6769 if (!queue->synflood_warned && syncookies != 2 &&
6770 xchg(&queue->synflood_warned, 1) == 0)
6771 net_info_ratelimited("%s: Possible SYN flooding on port %d. %s. Check SNMP counters.\n",
6772 proto, sk->sk_num, msg);
6777 static void tcp_reqsk_record_syn(const struct sock *sk,
6778 struct request_sock *req,
6779 const struct sk_buff *skb)
6781 if (tcp_sk(sk)->save_syn) {
6782 u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb);
6783 struct saved_syn *saved_syn;
6787 if (tcp_sk(sk)->save_syn == 2) { /* Save full header. */
6788 base = skb_mac_header(skb);
6789 mac_hdrlen = skb_mac_header_len(skb);
6792 base = skb_network_header(skb);
6796 saved_syn = kmalloc(struct_size(saved_syn, data, len),
6799 saved_syn->mac_hdrlen = mac_hdrlen;
6800 saved_syn->network_hdrlen = skb_network_header_len(skb);
6801 saved_syn->tcp_hdrlen = tcp_hdrlen(skb);
6802 memcpy(saved_syn->data, base, len);
6803 req->saved_syn = saved_syn;
6808 /* If a SYN cookie is required and supported, returns a clamped MSS value to be
6809 * used for SYN cookie generation.
6811 u16 tcp_get_syncookie_mss(struct request_sock_ops *rsk_ops,
6812 const struct tcp_request_sock_ops *af_ops,
6813 struct sock *sk, struct tcphdr *th)
6815 struct tcp_sock *tp = tcp_sk(sk);
6818 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_syncookies) != 2 &&
6819 !inet_csk_reqsk_queue_is_full(sk))
6822 if (!tcp_syn_flood_action(sk, rsk_ops->slab_name))
6825 if (sk_acceptq_is_full(sk)) {
6826 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
6830 mss = tcp_parse_mss_option(th, tp->rx_opt.user_mss);
6832 mss = af_ops->mss_clamp;
6836 EXPORT_SYMBOL_GPL(tcp_get_syncookie_mss);
6838 int tcp_conn_request(struct request_sock_ops *rsk_ops,
6839 const struct tcp_request_sock_ops *af_ops,
6840 struct sock *sk, struct sk_buff *skb)
6842 struct tcp_fastopen_cookie foc = { .len = -1 };
6843 __u32 isn = TCP_SKB_CB(skb)->tcp_tw_isn;
6844 struct tcp_options_received tmp_opt;
6845 struct tcp_sock *tp = tcp_sk(sk);
6846 struct net *net = sock_net(sk);
6847 struct sock *fastopen_sk = NULL;
6848 struct request_sock *req;
6849 bool want_cookie = false;
6850 struct dst_entry *dst;
6854 syncookies = READ_ONCE(net->ipv4.sysctl_tcp_syncookies);
6856 /* TW buckets are converted to open requests without
6857 * limitations, they conserve resources and peer is
6858 * evidently real one.
6860 if ((syncookies == 2 || inet_csk_reqsk_queue_is_full(sk)) && !isn) {
6861 want_cookie = tcp_syn_flood_action(sk, rsk_ops->slab_name);
6866 if (sk_acceptq_is_full(sk)) {
6867 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
6871 req = inet_reqsk_alloc(rsk_ops, sk, !want_cookie);
6875 req->syncookie = want_cookie;
6876 tcp_rsk(req)->af_specific = af_ops;
6877 tcp_rsk(req)->ts_off = 0;
6878 #if IS_ENABLED(CONFIG_MPTCP)
6879 tcp_rsk(req)->is_mptcp = 0;
6882 tcp_clear_options(&tmp_opt);
6883 tmp_opt.mss_clamp = af_ops->mss_clamp;
6884 tmp_opt.user_mss = tp->rx_opt.user_mss;
6885 tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0,
6886 want_cookie ? NULL : &foc);
6888 if (want_cookie && !tmp_opt.saw_tstamp)
6889 tcp_clear_options(&tmp_opt);
6891 if (IS_ENABLED(CONFIG_SMC) && want_cookie)
6894 tmp_opt.tstamp_ok = tmp_opt.saw_tstamp;
6895 tcp_openreq_init(req, &tmp_opt, skb, sk);
6896 inet_rsk(req)->no_srccheck = inet_sk(sk)->transparent;
6898 /* Note: tcp_v6_init_req() might override ir_iif for link locals */
6899 inet_rsk(req)->ir_iif = inet_request_bound_dev_if(sk, skb);
6901 af_ops->init_req(req, sk, skb);
6903 if (security_inet_conn_request(sk, skb, req))
6906 if (tmp_opt.tstamp_ok)
6907 tcp_rsk(req)->ts_off = af_ops->init_ts_off(net, skb);
6909 dst = af_ops->route_req(sk, &fl, req);
6913 if (!want_cookie && !isn) {
6914 int max_syn_backlog = READ_ONCE(net->ipv4.sysctl_max_syn_backlog);
6916 /* Kill the following clause, if you dislike this way. */
6918 (max_syn_backlog - inet_csk_reqsk_queue_len(sk) <
6919 (max_syn_backlog >> 2)) &&
6920 !tcp_peer_is_proven(req, dst)) {
6921 /* Without syncookies last quarter of
6922 * backlog is filled with destinations,
6923 * proven to be alive.
6924 * It means that we continue to communicate
6925 * to destinations, already remembered
6926 * to the moment of synflood.
6928 pr_drop_req(req, ntohs(tcp_hdr(skb)->source),
6930 goto drop_and_release;
6933 isn = af_ops->init_seq(skb);
6936 tcp_ecn_create_request(req, skb, sk, dst);
6939 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss);
6940 if (!tmp_opt.tstamp_ok)
6941 inet_rsk(req)->ecn_ok = 0;
6944 tcp_rsk(req)->snt_isn = isn;
6945 tcp_rsk(req)->txhash = net_tx_rndhash();
6946 tcp_rsk(req)->syn_tos = TCP_SKB_CB(skb)->ip_dsfield;
6947 tcp_openreq_init_rwin(req, sk, dst);
6948 sk_rx_queue_set(req_to_sk(req), skb);
6950 tcp_reqsk_record_syn(sk, req, skb);
6951 fastopen_sk = tcp_try_fastopen(sk, skb, req, &foc, dst);
6954 af_ops->send_synack(fastopen_sk, dst, &fl, req,
6955 &foc, TCP_SYNACK_FASTOPEN, skb);
6956 /* Add the child socket directly into the accept queue */
6957 if (!inet_csk_reqsk_queue_add(sk, req, fastopen_sk)) {
6958 reqsk_fastopen_remove(fastopen_sk, req, false);
6959 bh_unlock_sock(fastopen_sk);
6960 sock_put(fastopen_sk);
6963 sk->sk_data_ready(sk);
6964 bh_unlock_sock(fastopen_sk);
6965 sock_put(fastopen_sk);
6967 tcp_rsk(req)->tfo_listener = false;
6969 inet_csk_reqsk_queue_hash_add(sk, req,
6970 tcp_timeout_init((struct sock *)req));
6971 af_ops->send_synack(sk, dst, &fl, req, &foc,
6972 !want_cookie ? TCP_SYNACK_NORMAL :
6991 EXPORT_SYMBOL(tcp_conn_request);