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 */
103 #define FLAG_DSACK_TLP 0x20000 /* DSACK for tail loss probe */
105 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
106 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
107 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE|FLAG_DSACKING_ACK)
108 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
110 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
111 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
113 #define REXMIT_NONE 0 /* no loss recovery to do */
114 #define REXMIT_LOST 1 /* retransmit packets marked lost */
115 #define REXMIT_NEW 2 /* FRTO-style transmit of unsent/new packets */
117 #if IS_ENABLED(CONFIG_TLS_DEVICE)
118 static DEFINE_STATIC_KEY_DEFERRED_FALSE(clean_acked_data_enabled, HZ);
120 void clean_acked_data_enable(struct inet_connection_sock *icsk,
121 void (*cad)(struct sock *sk, u32 ack_seq))
123 icsk->icsk_clean_acked = cad;
124 static_branch_deferred_inc(&clean_acked_data_enabled);
126 EXPORT_SYMBOL_GPL(clean_acked_data_enable);
128 void clean_acked_data_disable(struct inet_connection_sock *icsk)
130 static_branch_slow_dec_deferred(&clean_acked_data_enabled);
131 icsk->icsk_clean_acked = NULL;
133 EXPORT_SYMBOL_GPL(clean_acked_data_disable);
135 void clean_acked_data_flush(void)
137 static_key_deferred_flush(&clean_acked_data_enabled);
139 EXPORT_SYMBOL_GPL(clean_acked_data_flush);
142 #ifdef CONFIG_CGROUP_BPF
143 static void bpf_skops_parse_hdr(struct sock *sk, struct sk_buff *skb)
145 bool unknown_opt = tcp_sk(sk)->rx_opt.saw_unknown &&
146 BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk),
147 BPF_SOCK_OPS_PARSE_UNKNOWN_HDR_OPT_CB_FLAG);
148 bool parse_all_opt = BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk),
149 BPF_SOCK_OPS_PARSE_ALL_HDR_OPT_CB_FLAG);
150 struct bpf_sock_ops_kern sock_ops;
152 if (likely(!unknown_opt && !parse_all_opt))
155 /* The skb will be handled in the
156 * bpf_skops_established() or
157 * bpf_skops_write_hdr_opt().
159 switch (sk->sk_state) {
166 sock_owned_by_me(sk);
168 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp));
169 sock_ops.op = BPF_SOCK_OPS_PARSE_HDR_OPT_CB;
170 sock_ops.is_fullsock = 1;
172 bpf_skops_init_skb(&sock_ops, skb, tcp_hdrlen(skb));
174 BPF_CGROUP_RUN_PROG_SOCK_OPS(&sock_ops);
177 static void bpf_skops_established(struct sock *sk, int bpf_op,
180 struct bpf_sock_ops_kern sock_ops;
182 sock_owned_by_me(sk);
184 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp));
185 sock_ops.op = bpf_op;
186 sock_ops.is_fullsock = 1;
188 /* sk with TCP_REPAIR_ON does not have skb in tcp_finish_connect */
190 bpf_skops_init_skb(&sock_ops, skb, tcp_hdrlen(skb));
192 BPF_CGROUP_RUN_PROG_SOCK_OPS(&sock_ops);
195 static void bpf_skops_parse_hdr(struct sock *sk, struct sk_buff *skb)
199 static void bpf_skops_established(struct sock *sk, int bpf_op,
205 static __cold void tcp_gro_dev_warn(const struct sock *sk, const struct sk_buff *skb,
208 struct net_device *dev;
211 dev = dev_get_by_index_rcu(sock_net(sk), skb->skb_iif);
212 if (!dev || len >= READ_ONCE(dev->mtu))
213 pr_warn("%s: Driver has suspect GRO implementation, TCP performance may be compromised.\n",
214 dev ? dev->name : "Unknown driver");
218 /* Adapt the MSS value used to make delayed ack decision to the
221 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
223 struct inet_connection_sock *icsk = inet_csk(sk);
224 const unsigned int lss = icsk->icsk_ack.last_seg_size;
227 icsk->icsk_ack.last_seg_size = 0;
229 /* skb->len may jitter because of SACKs, even if peer
230 * sends good full-sized frames.
232 len = skb_shinfo(skb)->gso_size ? : skb->len;
233 if (len >= icsk->icsk_ack.rcv_mss) {
234 /* Note: divides are still a bit expensive.
235 * For the moment, only adjust scaling_ratio
236 * when we update icsk_ack.rcv_mss.
238 if (unlikely(len != icsk->icsk_ack.rcv_mss)) {
239 u64 val = (u64)skb->len << TCP_RMEM_TO_WIN_SCALE;
241 do_div(val, skb->truesize);
242 tcp_sk(sk)->scaling_ratio = val ? val : 1;
244 icsk->icsk_ack.rcv_mss = min_t(unsigned int, len,
246 /* Account for possibly-removed options */
247 DO_ONCE_LITE_IF(len > icsk->icsk_ack.rcv_mss + MAX_TCP_OPTION_SPACE,
248 tcp_gro_dev_warn, sk, skb, len);
249 /* If the skb has a len of exactly 1*MSS and has the PSH bit
250 * set then it is likely the end of an application write. So
251 * more data may not be arriving soon, and yet the data sender
252 * may be waiting for an ACK if cwnd-bound or using TX zero
253 * copy. So we set ICSK_ACK_PUSHED here so that
254 * tcp_cleanup_rbuf() will send an ACK immediately if the app
255 * reads all of the data and is not ping-pong. If len > MSS
256 * then this logic does not matter (and does not hurt) because
257 * tcp_cleanup_rbuf() will always ACK immediately if the app
258 * reads data and there is more than an MSS of unACKed data.
260 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_PSH)
261 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
263 /* Otherwise, we make more careful check taking into account,
264 * that SACKs block is variable.
266 * "len" is invariant segment length, including TCP header.
268 len += skb->data - skb_transport_header(skb);
269 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
270 /* If PSH is not set, packet should be
271 * full sized, provided peer TCP is not badly broken.
272 * This observation (if it is correct 8)) allows
273 * to handle super-low mtu links fairly.
275 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
276 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
277 /* Subtract also invariant (if peer is RFC compliant),
278 * tcp header plus fixed timestamp option length.
279 * Resulting "len" is MSS free of SACK jitter.
281 len -= tcp_sk(sk)->tcp_header_len;
282 icsk->icsk_ack.last_seg_size = len;
284 icsk->icsk_ack.rcv_mss = len;
288 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
289 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
290 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
294 static void tcp_incr_quickack(struct sock *sk, unsigned int max_quickacks)
296 struct inet_connection_sock *icsk = inet_csk(sk);
297 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
301 quickacks = min(quickacks, max_quickacks);
302 if (quickacks > icsk->icsk_ack.quick)
303 icsk->icsk_ack.quick = quickacks;
306 static void tcp_enter_quickack_mode(struct sock *sk, unsigned int max_quickacks)
308 struct inet_connection_sock *icsk = inet_csk(sk);
310 tcp_incr_quickack(sk, max_quickacks);
311 inet_csk_exit_pingpong_mode(sk);
312 icsk->icsk_ack.ato = TCP_ATO_MIN;
315 /* Send ACKs quickly, if "quick" count is not exhausted
316 * and the session is not interactive.
319 static bool tcp_in_quickack_mode(struct sock *sk)
321 const struct inet_connection_sock *icsk = inet_csk(sk);
322 const struct dst_entry *dst = __sk_dst_get(sk);
324 return (dst && dst_metric(dst, RTAX_QUICKACK)) ||
325 (icsk->icsk_ack.quick && !inet_csk_in_pingpong_mode(sk));
328 static void tcp_ecn_queue_cwr(struct tcp_sock *tp)
330 if (tp->ecn_flags & TCP_ECN_OK)
331 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
334 static void tcp_ecn_accept_cwr(struct sock *sk, const struct sk_buff *skb)
336 if (tcp_hdr(skb)->cwr) {
337 tcp_sk(sk)->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
339 /* If the sender is telling us it has entered CWR, then its
340 * cwnd may be very low (even just 1 packet), so we should ACK
343 if (TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq)
344 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW;
348 static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp)
350 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
353 static void __tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb)
355 struct tcp_sock *tp = tcp_sk(sk);
357 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
358 case INET_ECN_NOT_ECT:
359 /* Funny extension: if ECT is not set on a segment,
360 * and we already seen ECT on a previous segment,
361 * it is probably a retransmit.
363 if (tp->ecn_flags & TCP_ECN_SEEN)
364 tcp_enter_quickack_mode(sk, 2);
367 if (tcp_ca_needs_ecn(sk))
368 tcp_ca_event(sk, CA_EVENT_ECN_IS_CE);
370 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
371 /* Better not delay acks, sender can have a very low cwnd */
372 tcp_enter_quickack_mode(sk, 2);
373 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
375 tp->ecn_flags |= TCP_ECN_SEEN;
378 if (tcp_ca_needs_ecn(sk))
379 tcp_ca_event(sk, CA_EVENT_ECN_NO_CE);
380 tp->ecn_flags |= TCP_ECN_SEEN;
385 static void tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb)
387 if (tcp_sk(sk)->ecn_flags & TCP_ECN_OK)
388 __tcp_ecn_check_ce(sk, skb);
391 static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
393 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
394 tp->ecn_flags &= ~TCP_ECN_OK;
397 static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
399 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
400 tp->ecn_flags &= ~TCP_ECN_OK;
403 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
405 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
410 /* Buffer size and advertised window tuning.
412 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
415 static void tcp_sndbuf_expand(struct sock *sk)
417 const struct tcp_sock *tp = tcp_sk(sk);
418 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
422 /* Worst case is non GSO/TSO : each frame consumes one skb
423 * and skb->head is kmalloced using power of two area of memory
425 per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
427 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
429 per_mss = roundup_pow_of_two(per_mss) +
430 SKB_DATA_ALIGN(sizeof(struct sk_buff));
432 nr_segs = max_t(u32, TCP_INIT_CWND, tcp_snd_cwnd(tp));
433 nr_segs = max_t(u32, nr_segs, tp->reordering + 1);
435 /* Fast Recovery (RFC 5681 3.2) :
436 * Cubic needs 1.7 factor, rounded to 2 to include
437 * extra cushion (application might react slowly to EPOLLOUT)
439 sndmem = ca_ops->sndbuf_expand ? ca_ops->sndbuf_expand(sk) : 2;
440 sndmem *= nr_segs * per_mss;
442 if (sk->sk_sndbuf < sndmem)
443 WRITE_ONCE(sk->sk_sndbuf,
444 min(sndmem, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_wmem[2])));
447 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
449 * All tcp_full_space() is split to two parts: "network" buffer, allocated
450 * forward and advertised in receiver window (tp->rcv_wnd) and
451 * "application buffer", required to isolate scheduling/application
452 * latencies from network.
453 * window_clamp is maximal advertised window. It can be less than
454 * tcp_full_space(), in this case tcp_full_space() - window_clamp
455 * is reserved for "application" buffer. The less window_clamp is
456 * the smoother our behaviour from viewpoint of network, but the lower
457 * throughput and the higher sensitivity of the connection to losses. 8)
459 * rcv_ssthresh is more strict window_clamp used at "slow start"
460 * phase to predict further behaviour of this connection.
461 * It is used for two goals:
462 * - to enforce header prediction at sender, even when application
463 * requires some significant "application buffer". It is check #1.
464 * - to prevent pruning of receive queue because of misprediction
465 * of receiver window. Check #2.
467 * The scheme does not work when sender sends good segments opening
468 * window and then starts to feed us spaghetti. But it should work
469 * in common situations. Otherwise, we have to rely on queue collapsing.
472 /* Slow part of check#2. */
473 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb,
474 unsigned int skbtruesize)
476 const struct tcp_sock *tp = tcp_sk(sk);
478 int truesize = tcp_win_from_space(sk, skbtruesize) >> 1;
479 int window = tcp_win_from_space(sk, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2])) >> 1;
481 while (tp->rcv_ssthresh <= window) {
482 if (truesize <= skb->len)
483 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
491 /* Even if skb appears to have a bad len/truesize ratio, TCP coalescing
492 * can play nice with us, as sk_buff and skb->head might be either
493 * freed or shared with up to MAX_SKB_FRAGS segments.
494 * Only give a boost to drivers using page frag(s) to hold the frame(s),
495 * and if no payload was pulled in skb->head before reaching us.
497 static u32 truesize_adjust(bool adjust, const struct sk_buff *skb)
499 u32 truesize = skb->truesize;
501 if (adjust && !skb_headlen(skb)) {
502 truesize -= SKB_TRUESIZE(skb_end_offset(skb));
503 /* paranoid check, some drivers might be buggy */
504 if (unlikely((int)truesize < (int)skb->len))
505 truesize = skb->truesize;
510 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb,
513 struct tcp_sock *tp = tcp_sk(sk);
516 room = min_t(int, tp->window_clamp, tcp_space(sk)) - tp->rcv_ssthresh;
522 if (!tcp_under_memory_pressure(sk)) {
523 unsigned int truesize = truesize_adjust(adjust, skb);
526 /* Check #2. Increase window, if skb with such overhead
527 * will fit to rcvbuf in future.
529 if (tcp_win_from_space(sk, truesize) <= skb->len)
530 incr = 2 * tp->advmss;
532 incr = __tcp_grow_window(sk, skb, truesize);
535 incr = max_t(int, incr, 2 * skb->len);
536 tp->rcv_ssthresh += min(room, incr);
537 inet_csk(sk)->icsk_ack.quick |= 1;
541 * Adjust rcv_ssthresh according to reserved mem
543 tcp_adjust_rcv_ssthresh(sk);
547 /* 3. Try to fixup all. It is made immediately after connection enters
550 static void tcp_init_buffer_space(struct sock *sk)
552 int tcp_app_win = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_app_win);
553 struct tcp_sock *tp = tcp_sk(sk);
556 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
557 tcp_sndbuf_expand(sk);
559 tcp_mstamp_refresh(tp);
560 tp->rcvq_space.time = tp->tcp_mstamp;
561 tp->rcvq_space.seq = tp->copied_seq;
563 maxwin = tcp_full_space(sk);
565 if (tp->window_clamp >= maxwin) {
566 tp->window_clamp = maxwin;
568 if (tcp_app_win && maxwin > 4 * tp->advmss)
569 tp->window_clamp = max(maxwin -
570 (maxwin >> tcp_app_win),
574 /* Force reservation of one segment. */
576 tp->window_clamp > 2 * tp->advmss &&
577 tp->window_clamp + tp->advmss > maxwin)
578 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
580 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
581 tp->snd_cwnd_stamp = tcp_jiffies32;
582 tp->rcvq_space.space = min3(tp->rcv_ssthresh, tp->rcv_wnd,
583 (u32)TCP_INIT_CWND * tp->advmss);
586 /* 4. Recalculate window clamp after socket hit its memory bounds. */
587 static void tcp_clamp_window(struct sock *sk)
589 struct tcp_sock *tp = tcp_sk(sk);
590 struct inet_connection_sock *icsk = inet_csk(sk);
591 struct net *net = sock_net(sk);
594 icsk->icsk_ack.quick = 0;
595 rmem2 = READ_ONCE(net->ipv4.sysctl_tcp_rmem[2]);
597 if (sk->sk_rcvbuf < rmem2 &&
598 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
599 !tcp_under_memory_pressure(sk) &&
600 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
601 WRITE_ONCE(sk->sk_rcvbuf,
602 min(atomic_read(&sk->sk_rmem_alloc), rmem2));
604 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
605 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
608 /* Initialize RCV_MSS value.
609 * RCV_MSS is an our guess about MSS used by the peer.
610 * We haven't any direct information about the MSS.
611 * It's better to underestimate the RCV_MSS rather than overestimate.
612 * Overestimations make us ACKing less frequently than needed.
613 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
615 void tcp_initialize_rcv_mss(struct sock *sk)
617 const struct tcp_sock *tp = tcp_sk(sk);
618 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
620 hint = min(hint, tp->rcv_wnd / 2);
621 hint = min(hint, TCP_MSS_DEFAULT);
622 hint = max(hint, TCP_MIN_MSS);
624 inet_csk(sk)->icsk_ack.rcv_mss = hint;
626 EXPORT_SYMBOL(tcp_initialize_rcv_mss);
628 /* Receiver "autotuning" code.
630 * The algorithm for RTT estimation w/o timestamps is based on
631 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
632 * <https://public.lanl.gov/radiant/pubs.html#DRS>
634 * More detail on this code can be found at
635 * <http://staff.psc.edu/jheffner/>,
636 * though this reference is out of date. A new paper
639 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
641 u32 new_sample = tp->rcv_rtt_est.rtt_us;
644 if (new_sample != 0) {
645 /* If we sample in larger samples in the non-timestamp
646 * case, we could grossly overestimate the RTT especially
647 * with chatty applications or bulk transfer apps which
648 * are stalled on filesystem I/O.
650 * Also, since we are only going for a minimum in the
651 * non-timestamp case, we do not smooth things out
652 * else with timestamps disabled convergence takes too
656 m -= (new_sample >> 3);
664 /* No previous measure. */
668 tp->rcv_rtt_est.rtt_us = new_sample;
671 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
675 if (tp->rcv_rtt_est.time == 0)
677 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
679 delta_us = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcv_rtt_est.time);
682 tcp_rcv_rtt_update(tp, delta_us, 1);
685 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
686 tp->rcv_rtt_est.time = tp->tcp_mstamp;
689 static s32 tcp_rtt_tsopt_us(const struct tcp_sock *tp)
693 delta = tcp_time_stamp_ts(tp) - tp->rx_opt.rcv_tsecr;
697 if (likely(delta < INT_MAX / (USEC_PER_SEC / TCP_TS_HZ))) {
700 delta_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
706 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
707 const struct sk_buff *skb)
709 struct tcp_sock *tp = tcp_sk(sk);
711 if (tp->rx_opt.rcv_tsecr == tp->rcv_rtt_last_tsecr)
713 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr;
715 if (TCP_SKB_CB(skb)->end_seq -
716 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss) {
717 s32 delta = tcp_rtt_tsopt_us(tp);
720 tcp_rcv_rtt_update(tp, delta, 0);
725 * This function should be called every time data is copied to user space.
726 * It calculates the appropriate TCP receive buffer space.
728 void tcp_rcv_space_adjust(struct sock *sk)
730 struct tcp_sock *tp = tcp_sk(sk);
734 trace_tcp_rcv_space_adjust(sk);
736 tcp_mstamp_refresh(tp);
737 time = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcvq_space.time);
738 if (time < (tp->rcv_rtt_est.rtt_us >> 3) || tp->rcv_rtt_est.rtt_us == 0)
741 /* Number of bytes copied to user in last RTT */
742 copied = tp->copied_seq - tp->rcvq_space.seq;
743 if (copied <= tp->rcvq_space.space)
747 * copied = bytes received in previous RTT, our base window
748 * To cope with packet losses, we need a 2x factor
749 * To cope with slow start, and sender growing its cwin by 100 %
750 * every RTT, we need a 4x factor, because the ACK we are sending
751 * now is for the next RTT, not the current one :
752 * <prev RTT . ><current RTT .. ><next RTT .... >
755 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf) &&
756 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
760 /* minimal window to cope with packet losses, assuming
761 * steady state. Add some cushion because of small variations.
763 rcvwin = ((u64)copied << 1) + 16 * tp->advmss;
765 /* Accommodate for sender rate increase (eg. slow start) */
766 grow = rcvwin * (copied - tp->rcvq_space.space);
767 do_div(grow, tp->rcvq_space.space);
768 rcvwin += (grow << 1);
770 rcvbuf = min_t(u64, tcp_space_from_win(sk, rcvwin),
771 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]));
772 if (rcvbuf > sk->sk_rcvbuf) {
773 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
775 /* Make the window clamp follow along. */
776 tp->window_clamp = tcp_win_from_space(sk, rcvbuf);
779 tp->rcvq_space.space = copied;
782 tp->rcvq_space.seq = tp->copied_seq;
783 tp->rcvq_space.time = tp->tcp_mstamp;
786 static void tcp_save_lrcv_flowlabel(struct sock *sk, const struct sk_buff *skb)
788 #if IS_ENABLED(CONFIG_IPV6)
789 struct inet_connection_sock *icsk = inet_csk(sk);
791 if (skb->protocol == htons(ETH_P_IPV6))
792 icsk->icsk_ack.lrcv_flowlabel = ntohl(ip6_flowlabel(ipv6_hdr(skb)));
796 /* There is something which you must keep in mind when you analyze the
797 * behavior of the tp->ato delayed ack timeout interval. When a
798 * connection starts up, we want to ack as quickly as possible. The
799 * problem is that "good" TCP's do slow start at the beginning of data
800 * transmission. The means that until we send the first few ACK's the
801 * sender will sit on his end and only queue most of his data, because
802 * he can only send snd_cwnd unacked packets at any given time. For
803 * each ACK we send, he increments snd_cwnd and transmits more of his
806 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
808 struct tcp_sock *tp = tcp_sk(sk);
809 struct inet_connection_sock *icsk = inet_csk(sk);
812 inet_csk_schedule_ack(sk);
814 tcp_measure_rcv_mss(sk, skb);
816 tcp_rcv_rtt_measure(tp);
820 if (!icsk->icsk_ack.ato) {
821 /* The _first_ data packet received, initialize
822 * delayed ACK engine.
824 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
825 icsk->icsk_ack.ato = TCP_ATO_MIN;
827 int m = now - icsk->icsk_ack.lrcvtime;
829 if (m <= TCP_ATO_MIN / 2) {
830 /* The fastest case is the first. */
831 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
832 } else if (m < icsk->icsk_ack.ato) {
833 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
834 if (icsk->icsk_ack.ato > icsk->icsk_rto)
835 icsk->icsk_ack.ato = icsk->icsk_rto;
836 } else if (m > icsk->icsk_rto) {
837 /* Too long gap. Apparently sender failed to
838 * restart window, so that we send ACKs quickly.
840 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
843 icsk->icsk_ack.lrcvtime = now;
844 tcp_save_lrcv_flowlabel(sk, skb);
846 tcp_ecn_check_ce(sk, skb);
849 tcp_grow_window(sk, skb, true);
852 /* Called to compute a smoothed rtt estimate. The data fed to this
853 * routine either comes from timestamps, or from segments that were
854 * known _not_ to have been retransmitted [see Karn/Partridge
855 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
856 * piece by Van Jacobson.
857 * NOTE: the next three routines used to be one big routine.
858 * To save cycles in the RFC 1323 implementation it was better to break
859 * it up into three procedures. -- erics
861 static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
863 struct tcp_sock *tp = tcp_sk(sk);
864 long m = mrtt_us; /* RTT */
865 u32 srtt = tp->srtt_us;
867 /* The following amusing code comes from Jacobson's
868 * article in SIGCOMM '88. Note that rtt and mdev
869 * are scaled versions of rtt and mean deviation.
870 * This is designed to be as fast as possible
871 * m stands for "measurement".
873 * On a 1990 paper the rto value is changed to:
874 * RTO = rtt + 4 * mdev
876 * Funny. This algorithm seems to be very broken.
877 * These formulae increase RTO, when it should be decreased, increase
878 * too slowly, when it should be increased quickly, decrease too quickly
879 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
880 * does not matter how to _calculate_ it. Seems, it was trap
881 * that VJ failed to avoid. 8)
884 m -= (srtt >> 3); /* m is now error in rtt est */
885 srtt += m; /* rtt = 7/8 rtt + 1/8 new */
887 m = -m; /* m is now abs(error) */
888 m -= (tp->mdev_us >> 2); /* similar update on mdev */
889 /* This is similar to one of Eifel findings.
890 * Eifel blocks mdev updates when rtt decreases.
891 * This solution is a bit different: we use finer gain
892 * for mdev in this case (alpha*beta).
893 * Like Eifel it also prevents growth of rto,
894 * but also it limits too fast rto decreases,
895 * happening in pure Eifel.
900 m -= (tp->mdev_us >> 2); /* similar update on mdev */
902 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */
903 if (tp->mdev_us > tp->mdev_max_us) {
904 tp->mdev_max_us = tp->mdev_us;
905 if (tp->mdev_max_us > tp->rttvar_us)
906 tp->rttvar_us = tp->mdev_max_us;
908 if (after(tp->snd_una, tp->rtt_seq)) {
909 if (tp->mdev_max_us < tp->rttvar_us)
910 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
911 tp->rtt_seq = tp->snd_nxt;
912 tp->mdev_max_us = tcp_rto_min_us(sk);
917 /* no previous measure. */
918 srtt = m << 3; /* take the measured time to be rtt */
919 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */
920 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
921 tp->mdev_max_us = tp->rttvar_us;
922 tp->rtt_seq = tp->snd_nxt;
926 tp->srtt_us = max(1U, srtt);
929 static void tcp_update_pacing_rate(struct sock *sk)
931 const struct tcp_sock *tp = tcp_sk(sk);
934 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
935 rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3);
937 /* current rate is (cwnd * mss) / srtt
938 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
939 * In Congestion Avoidance phase, set it to 120 % the current rate.
941 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
942 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
943 * end of slow start and should slow down.
945 if (tcp_snd_cwnd(tp) < tp->snd_ssthresh / 2)
946 rate *= READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_pacing_ss_ratio);
948 rate *= READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_pacing_ca_ratio);
950 rate *= max(tcp_snd_cwnd(tp), tp->packets_out);
952 if (likely(tp->srtt_us))
953 do_div(rate, tp->srtt_us);
955 /* WRITE_ONCE() is needed because sch_fq fetches sk_pacing_rate
956 * without any lock. We want to make sure compiler wont store
957 * intermediate values in this location.
959 WRITE_ONCE(sk->sk_pacing_rate,
960 min_t(u64, rate, READ_ONCE(sk->sk_max_pacing_rate)));
963 /* Calculate rto without backoff. This is the second half of Van Jacobson's
964 * routine referred to above.
966 static void tcp_set_rto(struct sock *sk)
968 const struct tcp_sock *tp = tcp_sk(sk);
969 /* Old crap is replaced with new one. 8)
972 * 1. If rtt variance happened to be less 50msec, it is hallucination.
973 * It cannot be less due to utterly erratic ACK generation made
974 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
975 * to do with delayed acks, because at cwnd>2 true delack timeout
976 * is invisible. Actually, Linux-2.4 also generates erratic
977 * ACKs in some circumstances.
979 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
981 /* 2. Fixups made earlier cannot be right.
982 * If we do not estimate RTO correctly without them,
983 * all the algo is pure shit and should be replaced
984 * with correct one. It is exactly, which we pretend to do.
987 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
988 * guarantees that rto is higher.
993 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
995 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
998 cwnd = TCP_INIT_CWND;
999 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
1002 struct tcp_sacktag_state {
1003 /* Timestamps for earliest and latest never-retransmitted segment
1004 * that was SACKed. RTO needs the earliest RTT to stay conservative,
1005 * but congestion control should still get an accurate delay signal.
1012 unsigned int mss_now;
1013 struct rate_sample *rate;
1016 /* Take a notice that peer is sending D-SACKs. Skip update of data delivery
1017 * and spurious retransmission information if this DSACK is unlikely caused by
1019 * - DSACKed sequence range is larger than maximum receiver's window.
1020 * - Total no. of DSACKed segments exceed the total no. of retransmitted segs.
1022 static u32 tcp_dsack_seen(struct tcp_sock *tp, u32 start_seq,
1023 u32 end_seq, struct tcp_sacktag_state *state)
1025 u32 seq_len, dup_segs = 1;
1027 if (!before(start_seq, end_seq))
1030 seq_len = end_seq - start_seq;
1031 /* Dubious DSACK: DSACKed range greater than maximum advertised rwnd */
1032 if (seq_len > tp->max_window)
1034 if (seq_len > tp->mss_cache)
1035 dup_segs = DIV_ROUND_UP(seq_len, tp->mss_cache);
1036 else if (tp->tlp_high_seq && tp->tlp_high_seq == end_seq)
1037 state->flag |= FLAG_DSACK_TLP;
1039 tp->dsack_dups += dup_segs;
1040 /* Skip the DSACK if dup segs weren't retransmitted by sender */
1041 if (tp->dsack_dups > tp->total_retrans)
1044 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
1045 /* We increase the RACK ordering window in rounds where we receive
1046 * DSACKs that may have been due to reordering causing RACK to trigger
1047 * a spurious fast recovery. Thus RACK ignores DSACKs that happen
1048 * without having seen reordering, or that match TLP probes (TLP
1049 * is timer-driven, not triggered by RACK).
1051 if (tp->reord_seen && !(state->flag & FLAG_DSACK_TLP))
1052 tp->rack.dsack_seen = 1;
1054 state->flag |= FLAG_DSACKING_ACK;
1055 /* A spurious retransmission is delivered */
1056 state->sack_delivered += dup_segs;
1061 /* It's reordering when higher sequence was delivered (i.e. sacked) before
1062 * some lower never-retransmitted sequence ("low_seq"). The maximum reordering
1063 * distance is approximated in full-mss packet distance ("reordering").
1065 static void tcp_check_sack_reordering(struct sock *sk, const u32 low_seq,
1068 struct tcp_sock *tp = tcp_sk(sk);
1069 const u32 mss = tp->mss_cache;
1072 fack = tcp_highest_sack_seq(tp);
1073 if (!before(low_seq, fack))
1076 metric = fack - low_seq;
1077 if ((metric > tp->reordering * mss) && mss) {
1078 #if FASTRETRANS_DEBUG > 1
1079 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
1080 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
1084 tp->undo_marker ? tp->undo_retrans : 0);
1086 tp->reordering = min_t(u32, (metric + mss - 1) / mss,
1087 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_max_reordering));
1090 /* This exciting event is worth to be remembered. 8) */
1092 NET_INC_STATS(sock_net(sk),
1093 ts ? LINUX_MIB_TCPTSREORDER : LINUX_MIB_TCPSACKREORDER);
1096 /* This must be called before lost_out or retrans_out are updated
1097 * on a new loss, because we want to know if all skbs previously
1098 * known to be lost have already been retransmitted, indicating
1099 * that this newly lost skb is our next skb to retransmit.
1101 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
1103 if ((!tp->retransmit_skb_hint && tp->retrans_out >= tp->lost_out) ||
1104 (tp->retransmit_skb_hint &&
1105 before(TCP_SKB_CB(skb)->seq,
1106 TCP_SKB_CB(tp->retransmit_skb_hint)->seq)))
1107 tp->retransmit_skb_hint = skb;
1110 /* Sum the number of packets on the wire we have marked as lost, and
1111 * notify the congestion control module that the given skb was marked lost.
1113 static void tcp_notify_skb_loss_event(struct tcp_sock *tp, const struct sk_buff *skb)
1115 tp->lost += tcp_skb_pcount(skb);
1118 void tcp_mark_skb_lost(struct sock *sk, struct sk_buff *skb)
1120 __u8 sacked = TCP_SKB_CB(skb)->sacked;
1121 struct tcp_sock *tp = tcp_sk(sk);
1123 if (sacked & TCPCB_SACKED_ACKED)
1126 tcp_verify_retransmit_hint(tp, skb);
1127 if (sacked & TCPCB_LOST) {
1128 if (sacked & TCPCB_SACKED_RETRANS) {
1129 /* Account for retransmits that are lost again */
1130 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1131 tp->retrans_out -= tcp_skb_pcount(skb);
1132 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT,
1133 tcp_skb_pcount(skb));
1134 tcp_notify_skb_loss_event(tp, skb);
1137 tp->lost_out += tcp_skb_pcount(skb);
1138 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1139 tcp_notify_skb_loss_event(tp, skb);
1143 /* Updates the delivered and delivered_ce counts */
1144 static void tcp_count_delivered(struct tcp_sock *tp, u32 delivered,
1147 tp->delivered += delivered;
1149 tp->delivered_ce += delivered;
1152 /* This procedure tags the retransmission queue when SACKs arrive.
1154 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
1155 * Packets in queue with these bits set are counted in variables
1156 * sacked_out, retrans_out and lost_out, correspondingly.
1158 * Valid combinations are:
1159 * Tag InFlight Description
1160 * 0 1 - orig segment is in flight.
1161 * S 0 - nothing flies, orig reached receiver.
1162 * L 0 - nothing flies, orig lost by net.
1163 * R 2 - both orig and retransmit are in flight.
1164 * L|R 1 - orig is lost, retransmit is in flight.
1165 * S|R 1 - orig reached receiver, retrans is still in flight.
1166 * (L|S|R is logically valid, it could occur when L|R is sacked,
1167 * but it is equivalent to plain S and code short-curcuits it to S.
1168 * L|S is logically invalid, it would mean -1 packet in flight 8))
1170 * These 6 states form finite state machine, controlled by the following events:
1171 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
1172 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
1173 * 3. Loss detection event of two flavors:
1174 * A. Scoreboard estimator decided the packet is lost.
1175 * A'. Reno "three dupacks" marks head of queue lost.
1176 * B. SACK arrives sacking SND.NXT at the moment, when the
1177 * segment was retransmitted.
1178 * 4. D-SACK added new rule: D-SACK changes any tag to S.
1180 * It is pleasant to note, that state diagram turns out to be commutative,
1181 * so that we are allowed not to be bothered by order of our actions,
1182 * when multiple events arrive simultaneously. (see the function below).
1184 * Reordering detection.
1185 * --------------------
1186 * Reordering metric is maximal distance, which a packet can be displaced
1187 * in packet stream. With SACKs we can estimate it:
1189 * 1. SACK fills old hole and the corresponding segment was not
1190 * ever retransmitted -> reordering. Alas, we cannot use it
1191 * when segment was retransmitted.
1192 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1193 * for retransmitted and already SACKed segment -> reordering..
1194 * Both of these heuristics are not used in Loss state, when we cannot
1195 * account for retransmits accurately.
1197 * SACK block validation.
1198 * ----------------------
1200 * SACK block range validation checks that the received SACK block fits to
1201 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1202 * Note that SND.UNA is not included to the range though being valid because
1203 * it means that the receiver is rather inconsistent with itself reporting
1204 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1205 * perfectly valid, however, in light of RFC2018 which explicitly states
1206 * that "SACK block MUST reflect the newest segment. Even if the newest
1207 * segment is going to be discarded ...", not that it looks very clever
1208 * in case of head skb. Due to potentional receiver driven attacks, we
1209 * choose to avoid immediate execution of a walk in write queue due to
1210 * reneging and defer head skb's loss recovery to standard loss recovery
1211 * procedure that will eventually trigger (nothing forbids us doing this).
1213 * Implements also blockage to start_seq wrap-around. Problem lies in the
1214 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1215 * there's no guarantee that it will be before snd_nxt (n). The problem
1216 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1219 * <- outs wnd -> <- wrapzone ->
1220 * u e n u_w e_w s n_w
1222 * |<------------+------+----- TCP seqno space --------------+---------->|
1223 * ...-- <2^31 ->| |<--------...
1224 * ...---- >2^31 ------>| |<--------...
1226 * Current code wouldn't be vulnerable but it's better still to discard such
1227 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1228 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1229 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1230 * equal to the ideal case (infinite seqno space without wrap caused issues).
1232 * With D-SACK the lower bound is extended to cover sequence space below
1233 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1234 * again, D-SACK block must not to go across snd_una (for the same reason as
1235 * for the normal SACK blocks, explained above). But there all simplicity
1236 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1237 * fully below undo_marker they do not affect behavior in anyway and can
1238 * therefore be safely ignored. In rare cases (which are more or less
1239 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1240 * fragmentation and packet reordering past skb's retransmission. To consider
1241 * them correctly, the acceptable range must be extended even more though
1242 * the exact amount is rather hard to quantify. However, tp->max_window can
1243 * be used as an exaggerated estimate.
1245 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
1246 u32 start_seq, u32 end_seq)
1248 /* Too far in future, or reversed (interpretation is ambiguous) */
1249 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1252 /* Nasty start_seq wrap-around check (see comments above) */
1253 if (!before(start_seq, tp->snd_nxt))
1256 /* In outstanding window? ...This is valid exit for D-SACKs too.
1257 * start_seq == snd_una is non-sensical (see comments above)
1259 if (after(start_seq, tp->snd_una))
1262 if (!is_dsack || !tp->undo_marker)
1265 /* ...Then it's D-SACK, and must reside below snd_una completely */
1266 if (after(end_seq, tp->snd_una))
1269 if (!before(start_seq, tp->undo_marker))
1273 if (!after(end_seq, tp->undo_marker))
1276 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1277 * start_seq < undo_marker and end_seq >= undo_marker.
1279 return !before(start_seq, end_seq - tp->max_window);
1282 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1283 struct tcp_sack_block_wire *sp, int num_sacks,
1284 u32 prior_snd_una, struct tcp_sacktag_state *state)
1286 struct tcp_sock *tp = tcp_sk(sk);
1287 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1288 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1291 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1292 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1293 } else if (num_sacks > 1) {
1294 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1295 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1297 if (after(end_seq_0, end_seq_1) || before(start_seq_0, start_seq_1))
1299 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKOFORECV);
1304 dup_segs = tcp_dsack_seen(tp, start_seq_0, end_seq_0, state);
1305 if (!dup_segs) { /* Skip dubious DSACK */
1306 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKIGNOREDDUBIOUS);
1310 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECVSEGS, dup_segs);
1312 /* D-SACK for already forgotten data... Do dumb counting. */
1313 if (tp->undo_marker && tp->undo_retrans > 0 &&
1314 !after(end_seq_0, prior_snd_una) &&
1315 after(end_seq_0, tp->undo_marker))
1316 tp->undo_retrans = max_t(int, 0, tp->undo_retrans - dup_segs);
1321 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1322 * the incoming SACK may not exactly match but we can find smaller MSS
1323 * aligned portion of it that matches. Therefore we might need to fragment
1324 * which may fail and creates some hassle (caller must handle error case
1327 * FIXME: this could be merged to shift decision code
1329 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1330 u32 start_seq, u32 end_seq)
1334 unsigned int pkt_len;
1337 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1338 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1340 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1341 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1342 mss = tcp_skb_mss(skb);
1343 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1346 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1350 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1355 /* Round if necessary so that SACKs cover only full MSSes
1356 * and/or the remaining small portion (if present)
1358 if (pkt_len > mss) {
1359 unsigned int new_len = (pkt_len / mss) * mss;
1360 if (!in_sack && new_len < pkt_len)
1365 if (pkt_len >= skb->len && !in_sack)
1368 err = tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
1369 pkt_len, mss, GFP_ATOMIC);
1377 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1378 static u8 tcp_sacktag_one(struct sock *sk,
1379 struct tcp_sacktag_state *state, u8 sacked,
1380 u32 start_seq, u32 end_seq,
1381 int dup_sack, int pcount,
1384 struct tcp_sock *tp = tcp_sk(sk);
1386 /* Account D-SACK for retransmitted packet. */
1387 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1388 if (tp->undo_marker && tp->undo_retrans > 0 &&
1389 after(end_seq, tp->undo_marker))
1390 tp->undo_retrans = max_t(int, 0, tp->undo_retrans - pcount);
1391 if ((sacked & TCPCB_SACKED_ACKED) &&
1392 before(start_seq, state->reord))
1393 state->reord = start_seq;
1396 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1397 if (!after(end_seq, tp->snd_una))
1400 if (!(sacked & TCPCB_SACKED_ACKED)) {
1401 tcp_rack_advance(tp, sacked, end_seq, xmit_time);
1403 if (sacked & TCPCB_SACKED_RETRANS) {
1404 /* If the segment is not tagged as lost,
1405 * we do not clear RETRANS, believing
1406 * that retransmission is still in flight.
1408 if (sacked & TCPCB_LOST) {
1409 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1410 tp->lost_out -= pcount;
1411 tp->retrans_out -= pcount;
1414 if (!(sacked & TCPCB_RETRANS)) {
1415 /* New sack for not retransmitted frame,
1416 * which was in hole. It is reordering.
1418 if (before(start_seq,
1419 tcp_highest_sack_seq(tp)) &&
1420 before(start_seq, state->reord))
1421 state->reord = start_seq;
1423 if (!after(end_seq, tp->high_seq))
1424 state->flag |= FLAG_ORIG_SACK_ACKED;
1425 if (state->first_sackt == 0)
1426 state->first_sackt = xmit_time;
1427 state->last_sackt = xmit_time;
1430 if (sacked & TCPCB_LOST) {
1431 sacked &= ~TCPCB_LOST;
1432 tp->lost_out -= pcount;
1436 sacked |= TCPCB_SACKED_ACKED;
1437 state->flag |= FLAG_DATA_SACKED;
1438 tp->sacked_out += pcount;
1439 /* Out-of-order packets delivered */
1440 state->sack_delivered += pcount;
1442 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1443 if (tp->lost_skb_hint &&
1444 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1445 tp->lost_cnt_hint += pcount;
1448 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1449 * frames and clear it. undo_retrans is decreased above, L|R frames
1450 * are accounted above as well.
1452 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1453 sacked &= ~TCPCB_SACKED_RETRANS;
1454 tp->retrans_out -= pcount;
1460 /* Shift newly-SACKed bytes from this skb to the immediately previous
1461 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1463 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *prev,
1464 struct sk_buff *skb,
1465 struct tcp_sacktag_state *state,
1466 unsigned int pcount, int shifted, int mss,
1469 struct tcp_sock *tp = tcp_sk(sk);
1470 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1471 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
1475 /* Adjust counters and hints for the newly sacked sequence
1476 * range but discard the return value since prev is already
1477 * marked. We must tag the range first because the seq
1478 * advancement below implicitly advances
1479 * tcp_highest_sack_seq() when skb is highest_sack.
1481 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1482 start_seq, end_seq, dup_sack, pcount,
1483 tcp_skb_timestamp_us(skb));
1484 tcp_rate_skb_delivered(sk, skb, state->rate);
1486 if (skb == tp->lost_skb_hint)
1487 tp->lost_cnt_hint += pcount;
1489 TCP_SKB_CB(prev)->end_seq += shifted;
1490 TCP_SKB_CB(skb)->seq += shifted;
1492 tcp_skb_pcount_add(prev, pcount);
1493 WARN_ON_ONCE(tcp_skb_pcount(skb) < pcount);
1494 tcp_skb_pcount_add(skb, -pcount);
1496 /* When we're adding to gso_segs == 1, gso_size will be zero,
1497 * in theory this shouldn't be necessary but as long as DSACK
1498 * code can come after this skb later on it's better to keep
1499 * setting gso_size to something.
1501 if (!TCP_SKB_CB(prev)->tcp_gso_size)
1502 TCP_SKB_CB(prev)->tcp_gso_size = mss;
1504 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1505 if (tcp_skb_pcount(skb) <= 1)
1506 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1508 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1509 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1512 BUG_ON(!tcp_skb_pcount(skb));
1513 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1517 /* Whole SKB was eaten :-) */
1519 if (skb == tp->retransmit_skb_hint)
1520 tp->retransmit_skb_hint = prev;
1521 if (skb == tp->lost_skb_hint) {
1522 tp->lost_skb_hint = prev;
1523 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1526 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1527 TCP_SKB_CB(prev)->eor = TCP_SKB_CB(skb)->eor;
1528 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1529 TCP_SKB_CB(prev)->end_seq++;
1531 if (skb == tcp_highest_sack(sk))
1532 tcp_advance_highest_sack(sk, skb);
1534 tcp_skb_collapse_tstamp(prev, skb);
1535 if (unlikely(TCP_SKB_CB(prev)->tx.delivered_mstamp))
1536 TCP_SKB_CB(prev)->tx.delivered_mstamp = 0;
1538 tcp_rtx_queue_unlink_and_free(skb, sk);
1540 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKMERGED);
1545 /* I wish gso_size would have a bit more sane initialization than
1546 * something-or-zero which complicates things
1548 static int tcp_skb_seglen(const struct sk_buff *skb)
1550 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1553 /* Shifting pages past head area doesn't work */
1554 static int skb_can_shift(const struct sk_buff *skb)
1556 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1559 int tcp_skb_shift(struct sk_buff *to, struct sk_buff *from,
1560 int pcount, int shiftlen)
1562 /* TCP min gso_size is 8 bytes (TCP_MIN_GSO_SIZE)
1563 * Since TCP_SKB_CB(skb)->tcp_gso_segs is 16 bits, we need
1564 * to make sure not storing more than 65535 * 8 bytes per skb,
1565 * even if current MSS is bigger.
1567 if (unlikely(to->len + shiftlen >= 65535 * TCP_MIN_GSO_SIZE))
1569 if (unlikely(tcp_skb_pcount(to) + pcount > 65535))
1571 return skb_shift(to, from, shiftlen);
1574 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1577 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1578 struct tcp_sacktag_state *state,
1579 u32 start_seq, u32 end_seq,
1582 struct tcp_sock *tp = tcp_sk(sk);
1583 struct sk_buff *prev;
1589 /* Normally R but no L won't result in plain S */
1591 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1593 if (!skb_can_shift(skb))
1595 /* This frame is about to be dropped (was ACKed). */
1596 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1599 /* Can only happen with delayed DSACK + discard craziness */
1600 prev = skb_rb_prev(skb);
1604 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1607 if (!tcp_skb_can_collapse(prev, skb))
1610 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1611 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1615 pcount = tcp_skb_pcount(skb);
1616 mss = tcp_skb_seglen(skb);
1618 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1619 * drop this restriction as unnecessary
1621 if (mss != tcp_skb_seglen(prev))
1624 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1626 /* CHECKME: This is non-MSS split case only?, this will
1627 * cause skipped skbs due to advancing loop btw, original
1628 * has that feature too
1630 if (tcp_skb_pcount(skb) <= 1)
1633 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1635 /* TODO: head merge to next could be attempted here
1636 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1637 * though it might not be worth of the additional hassle
1639 * ...we can probably just fallback to what was done
1640 * previously. We could try merging non-SACKed ones
1641 * as well but it probably isn't going to buy off
1642 * because later SACKs might again split them, and
1643 * it would make skb timestamp tracking considerably
1649 len = end_seq - TCP_SKB_CB(skb)->seq;
1651 BUG_ON(len > skb->len);
1653 /* MSS boundaries should be honoured or else pcount will
1654 * severely break even though it makes things bit trickier.
1655 * Optimize common case to avoid most of the divides
1657 mss = tcp_skb_mss(skb);
1659 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1660 * drop this restriction as unnecessary
1662 if (mss != tcp_skb_seglen(prev))
1667 } else if (len < mss) {
1675 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1676 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1679 if (!tcp_skb_shift(prev, skb, pcount, len))
1681 if (!tcp_shifted_skb(sk, prev, skb, state, pcount, len, mss, dup_sack))
1684 /* Hole filled allows collapsing with the next as well, this is very
1685 * useful when hole on every nth skb pattern happens
1687 skb = skb_rb_next(prev);
1691 if (!skb_can_shift(skb) ||
1692 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1693 (mss != tcp_skb_seglen(skb)))
1696 if (!tcp_skb_can_collapse(prev, skb))
1699 pcount = tcp_skb_pcount(skb);
1700 if (tcp_skb_shift(prev, skb, pcount, len))
1701 tcp_shifted_skb(sk, prev, skb, state, pcount,
1711 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1715 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1716 struct tcp_sack_block *next_dup,
1717 struct tcp_sacktag_state *state,
1718 u32 start_seq, u32 end_seq,
1721 struct tcp_sock *tp = tcp_sk(sk);
1722 struct sk_buff *tmp;
1724 skb_rbtree_walk_from(skb) {
1726 bool dup_sack = dup_sack_in;
1728 /* queue is in-order => we can short-circuit the walk early */
1729 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1733 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1734 in_sack = tcp_match_skb_to_sack(sk, skb,
1735 next_dup->start_seq,
1741 /* skb reference here is a bit tricky to get right, since
1742 * shifting can eat and free both this skb and the next,
1743 * so not even _safe variant of the loop is enough.
1746 tmp = tcp_shift_skb_data(sk, skb, state,
1747 start_seq, end_seq, dup_sack);
1756 in_sack = tcp_match_skb_to_sack(sk, skb,
1762 if (unlikely(in_sack < 0))
1766 TCP_SKB_CB(skb)->sacked =
1769 TCP_SKB_CB(skb)->sacked,
1770 TCP_SKB_CB(skb)->seq,
1771 TCP_SKB_CB(skb)->end_seq,
1773 tcp_skb_pcount(skb),
1774 tcp_skb_timestamp_us(skb));
1775 tcp_rate_skb_delivered(sk, skb, state->rate);
1776 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1777 list_del_init(&skb->tcp_tsorted_anchor);
1779 if (!before(TCP_SKB_CB(skb)->seq,
1780 tcp_highest_sack_seq(tp)))
1781 tcp_advance_highest_sack(sk, skb);
1787 static struct sk_buff *tcp_sacktag_bsearch(struct sock *sk, u32 seq)
1789 struct rb_node *parent, **p = &sk->tcp_rtx_queue.rb_node;
1790 struct sk_buff *skb;
1794 skb = rb_to_skb(parent);
1795 if (before(seq, TCP_SKB_CB(skb)->seq)) {
1796 p = &parent->rb_left;
1799 if (!before(seq, TCP_SKB_CB(skb)->end_seq)) {
1800 p = &parent->rb_right;
1808 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1811 if (skb && after(TCP_SKB_CB(skb)->seq, skip_to_seq))
1814 return tcp_sacktag_bsearch(sk, skip_to_seq);
1817 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1819 struct tcp_sack_block *next_dup,
1820 struct tcp_sacktag_state *state,
1826 if (before(next_dup->start_seq, skip_to_seq)) {
1827 skb = tcp_sacktag_skip(skb, sk, next_dup->start_seq);
1828 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1829 next_dup->start_seq, next_dup->end_seq,
1836 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1838 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1842 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1843 u32 prior_snd_una, struct tcp_sacktag_state *state)
1845 struct tcp_sock *tp = tcp_sk(sk);
1846 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1847 TCP_SKB_CB(ack_skb)->sacked);
1848 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1849 struct tcp_sack_block sp[TCP_NUM_SACKS];
1850 struct tcp_sack_block *cache;
1851 struct sk_buff *skb;
1852 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1854 bool found_dup_sack = false;
1856 int first_sack_index;
1859 state->reord = tp->snd_nxt;
1861 if (!tp->sacked_out)
1862 tcp_highest_sack_reset(sk);
1864 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1865 num_sacks, prior_snd_una, state);
1867 /* Eliminate too old ACKs, but take into
1868 * account more or less fresh ones, they can
1869 * contain valid SACK info.
1871 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1874 if (!tp->packets_out)
1878 first_sack_index = 0;
1879 for (i = 0; i < num_sacks; i++) {
1880 bool dup_sack = !i && found_dup_sack;
1882 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1883 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1885 if (!tcp_is_sackblock_valid(tp, dup_sack,
1886 sp[used_sacks].start_seq,
1887 sp[used_sacks].end_seq)) {
1891 if (!tp->undo_marker)
1892 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1894 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1896 /* Don't count olds caused by ACK reordering */
1897 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1898 !after(sp[used_sacks].end_seq, tp->snd_una))
1900 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1903 NET_INC_STATS(sock_net(sk), mib_idx);
1905 first_sack_index = -1;
1909 /* Ignore very old stuff early */
1910 if (!after(sp[used_sacks].end_seq, prior_snd_una)) {
1912 first_sack_index = -1;
1919 /* order SACK blocks to allow in order walk of the retrans queue */
1920 for (i = used_sacks - 1; i > 0; i--) {
1921 for (j = 0; j < i; j++) {
1922 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1923 swap(sp[j], sp[j + 1]);
1925 /* Track where the first SACK block goes to */
1926 if (j == first_sack_index)
1927 first_sack_index = j + 1;
1932 state->mss_now = tcp_current_mss(sk);
1936 if (!tp->sacked_out) {
1937 /* It's already past, so skip checking against it */
1938 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1940 cache = tp->recv_sack_cache;
1941 /* Skip empty blocks in at head of the cache */
1942 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1947 while (i < used_sacks) {
1948 u32 start_seq = sp[i].start_seq;
1949 u32 end_seq = sp[i].end_seq;
1950 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1951 struct tcp_sack_block *next_dup = NULL;
1953 if (found_dup_sack && ((i + 1) == first_sack_index))
1954 next_dup = &sp[i + 1];
1956 /* Skip too early cached blocks */
1957 while (tcp_sack_cache_ok(tp, cache) &&
1958 !before(start_seq, cache->end_seq))
1961 /* Can skip some work by looking recv_sack_cache? */
1962 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1963 after(end_seq, cache->start_seq)) {
1966 if (before(start_seq, cache->start_seq)) {
1967 skb = tcp_sacktag_skip(skb, sk, start_seq);
1968 skb = tcp_sacktag_walk(skb, sk, next_dup,
1975 /* Rest of the block already fully processed? */
1976 if (!after(end_seq, cache->end_seq))
1979 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1983 /* ...tail remains todo... */
1984 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1985 /* ...but better entrypoint exists! */
1986 skb = tcp_highest_sack(sk);
1993 skb = tcp_sacktag_skip(skb, sk, cache->end_seq);
1994 /* Check overlap against next cached too (past this one already) */
1999 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
2000 skb = tcp_highest_sack(sk);
2004 skb = tcp_sacktag_skip(skb, sk, start_seq);
2007 skb = tcp_sacktag_walk(skb, sk, next_dup, state,
2008 start_seq, end_seq, dup_sack);
2014 /* Clear the head of the cache sack blocks so we can skip it next time */
2015 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
2016 tp->recv_sack_cache[i].start_seq = 0;
2017 tp->recv_sack_cache[i].end_seq = 0;
2019 for (j = 0; j < used_sacks; j++)
2020 tp->recv_sack_cache[i++] = sp[j];
2022 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Loss || tp->undo_marker)
2023 tcp_check_sack_reordering(sk, state->reord, 0);
2025 tcp_verify_left_out(tp);
2028 #if FASTRETRANS_DEBUG > 0
2029 WARN_ON((int)tp->sacked_out < 0);
2030 WARN_ON((int)tp->lost_out < 0);
2031 WARN_ON((int)tp->retrans_out < 0);
2032 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
2037 /* Limits sacked_out so that sum with lost_out isn't ever larger than
2038 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
2040 static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
2044 holes = max(tp->lost_out, 1U);
2045 holes = min(holes, tp->packets_out);
2047 if ((tp->sacked_out + holes) > tp->packets_out) {
2048 tp->sacked_out = tp->packets_out - holes;
2054 /* If we receive more dupacks than we expected counting segments
2055 * in assumption of absent reordering, interpret this as reordering.
2056 * The only another reason could be bug in receiver TCP.
2058 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
2060 struct tcp_sock *tp = tcp_sk(sk);
2062 if (!tcp_limit_reno_sacked(tp))
2065 tp->reordering = min_t(u32, tp->packets_out + addend,
2066 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_max_reordering));
2068 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRENOREORDER);
2071 /* Emulate SACKs for SACKless connection: account for a new dupack. */
2073 static void tcp_add_reno_sack(struct sock *sk, int num_dupack, bool ece_ack)
2076 struct tcp_sock *tp = tcp_sk(sk);
2077 u32 prior_sacked = tp->sacked_out;
2080 tp->sacked_out += num_dupack;
2081 tcp_check_reno_reordering(sk, 0);
2082 delivered = tp->sacked_out - prior_sacked;
2084 tcp_count_delivered(tp, delivered, ece_ack);
2085 tcp_verify_left_out(tp);
2089 /* Account for ACK, ACKing some data in Reno Recovery phase. */
2091 static void tcp_remove_reno_sacks(struct sock *sk, int acked, bool ece_ack)
2093 struct tcp_sock *tp = tcp_sk(sk);
2096 /* One ACK acked hole. The rest eat duplicate ACKs. */
2097 tcp_count_delivered(tp, max_t(int, acked - tp->sacked_out, 1),
2099 if (acked - 1 >= tp->sacked_out)
2102 tp->sacked_out -= acked - 1;
2104 tcp_check_reno_reordering(sk, acked);
2105 tcp_verify_left_out(tp);
2108 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
2113 void tcp_clear_retrans(struct tcp_sock *tp)
2115 tp->retrans_out = 0;
2117 tp->undo_marker = 0;
2118 tp->undo_retrans = -1;
2122 tp->total_rto_recoveries = 0;
2123 tp->total_rto_time = 0;
2126 static inline void tcp_init_undo(struct tcp_sock *tp)
2128 tp->undo_marker = tp->snd_una;
2129 /* Retransmission still in flight may cause DSACKs later. */
2130 tp->undo_retrans = tp->retrans_out ? : -1;
2133 static bool tcp_is_rack(const struct sock *sk)
2135 return READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_recovery) &
2136 TCP_RACK_LOSS_DETECTION;
2139 /* If we detect SACK reneging, forget all SACK information
2140 * and reset tags completely, otherwise preserve SACKs. If receiver
2141 * dropped its ofo queue, we will know this due to reneging detection.
2143 static void tcp_timeout_mark_lost(struct sock *sk)
2145 struct tcp_sock *tp = tcp_sk(sk);
2146 struct sk_buff *skb, *head;
2147 bool is_reneg; /* is receiver reneging on SACKs? */
2149 head = tcp_rtx_queue_head(sk);
2150 is_reneg = head && (TCP_SKB_CB(head)->sacked & TCPCB_SACKED_ACKED);
2152 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
2154 /* Mark SACK reneging until we recover from this loss event. */
2155 tp->is_sack_reneg = 1;
2156 } else if (tcp_is_reno(tp)) {
2157 tcp_reset_reno_sack(tp);
2161 skb_rbtree_walk_from(skb) {
2163 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
2164 else if (tcp_is_rack(sk) && skb != head &&
2165 tcp_rack_skb_timeout(tp, skb, 0) > 0)
2166 continue; /* Don't mark recently sent ones lost yet */
2167 tcp_mark_skb_lost(sk, skb);
2169 tcp_verify_left_out(tp);
2170 tcp_clear_all_retrans_hints(tp);
2173 /* Enter Loss state. */
2174 void tcp_enter_loss(struct sock *sk)
2176 const struct inet_connection_sock *icsk = inet_csk(sk);
2177 struct tcp_sock *tp = tcp_sk(sk);
2178 struct net *net = sock_net(sk);
2179 bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery;
2182 tcp_timeout_mark_lost(sk);
2184 /* Reduce ssthresh if it has not yet been made inside this window. */
2185 if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
2186 !after(tp->high_seq, tp->snd_una) ||
2187 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
2188 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2189 tp->prior_cwnd = tcp_snd_cwnd(tp);
2190 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2191 tcp_ca_event(sk, CA_EVENT_LOSS);
2194 tcp_snd_cwnd_set(tp, tcp_packets_in_flight(tp) + 1);
2195 tp->snd_cwnd_cnt = 0;
2196 tp->snd_cwnd_stamp = tcp_jiffies32;
2198 /* Timeout in disordered state after receiving substantial DUPACKs
2199 * suggests that the degree of reordering is over-estimated.
2201 reordering = READ_ONCE(net->ipv4.sysctl_tcp_reordering);
2202 if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
2203 tp->sacked_out >= reordering)
2204 tp->reordering = min_t(unsigned int, tp->reordering,
2207 tcp_set_ca_state(sk, TCP_CA_Loss);
2208 tp->high_seq = tp->snd_nxt;
2209 tcp_ecn_queue_cwr(tp);
2211 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
2212 * loss recovery is underway except recurring timeout(s) on
2213 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
2215 tp->frto = READ_ONCE(net->ipv4.sysctl_tcp_frto) &&
2216 (new_recovery || icsk->icsk_retransmits) &&
2217 !inet_csk(sk)->icsk_mtup.probe_size;
2220 /* If ACK arrived pointing to a remembered SACK, it means that our
2221 * remembered SACKs do not reflect real state of receiver i.e.
2222 * receiver _host_ is heavily congested (or buggy).
2224 * To avoid big spurious retransmission bursts due to transient SACK
2225 * scoreboard oddities that look like reneging, we give the receiver a
2226 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
2227 * restore sanity to the SACK scoreboard. If the apparent reneging
2228 * persists until this RTO then we'll clear the SACK scoreboard.
2230 static bool tcp_check_sack_reneging(struct sock *sk, int *ack_flag)
2232 if (*ack_flag & FLAG_SACK_RENEGING &&
2233 *ack_flag & FLAG_SND_UNA_ADVANCED) {
2234 struct tcp_sock *tp = tcp_sk(sk);
2235 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4),
2236 msecs_to_jiffies(10));
2238 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2239 delay, TCP_RTO_MAX);
2240 *ack_flag &= ~FLAG_SET_XMIT_TIMER;
2246 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2247 * counter when SACK is enabled (without SACK, sacked_out is used for
2250 * With reordering, holes may still be in flight, so RFC3517 recovery
2251 * uses pure sacked_out (total number of SACKed segments) even though
2252 * it violates the RFC that uses duplicate ACKs, often these are equal
2253 * but when e.g. out-of-window ACKs or packet duplication occurs,
2254 * they differ. Since neither occurs due to loss, TCP should really
2257 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2259 return tp->sacked_out + 1;
2262 /* Linux NewReno/SACK/ECN state machine.
2263 * --------------------------------------
2265 * "Open" Normal state, no dubious events, fast path.
2266 * "Disorder" In all the respects it is "Open",
2267 * but requires a bit more attention. It is entered when
2268 * we see some SACKs or dupacks. It is split of "Open"
2269 * mainly to move some processing from fast path to slow one.
2270 * "CWR" CWND was reduced due to some Congestion Notification event.
2271 * It can be ECN, ICMP source quench, local device congestion.
2272 * "Recovery" CWND was reduced, we are fast-retransmitting.
2273 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2275 * tcp_fastretrans_alert() is entered:
2276 * - each incoming ACK, if state is not "Open"
2277 * - when arrived ACK is unusual, namely:
2282 * Counting packets in flight is pretty simple.
2284 * in_flight = packets_out - left_out + retrans_out
2286 * packets_out is SND.NXT-SND.UNA counted in packets.
2288 * retrans_out is number of retransmitted segments.
2290 * left_out is number of segments left network, but not ACKed yet.
2292 * left_out = sacked_out + lost_out
2294 * sacked_out: Packets, which arrived to receiver out of order
2295 * and hence not ACKed. With SACKs this number is simply
2296 * amount of SACKed data. Even without SACKs
2297 * it is easy to give pretty reliable estimate of this number,
2298 * counting duplicate ACKs.
2300 * lost_out: Packets lost by network. TCP has no explicit
2301 * "loss notification" feedback from network (for now).
2302 * It means that this number can be only _guessed_.
2303 * Actually, it is the heuristics to predict lossage that
2304 * distinguishes different algorithms.
2306 * F.e. after RTO, when all the queue is considered as lost,
2307 * lost_out = packets_out and in_flight = retrans_out.
2309 * Essentially, we have now a few algorithms detecting
2312 * If the receiver supports SACK:
2314 * RFC6675/3517: It is the conventional algorithm. A packet is
2315 * considered lost if the number of higher sequence packets
2316 * SACKed is greater than or equal the DUPACK thoreshold
2317 * (reordering). This is implemented in tcp_mark_head_lost and
2318 * tcp_update_scoreboard.
2320 * RACK (draft-ietf-tcpm-rack-01): it is a newer algorithm
2321 * (2017-) that checks timing instead of counting DUPACKs.
2322 * Essentially a packet is considered lost if it's not S/ACKed
2323 * after RTT + reordering_window, where both metrics are
2324 * dynamically measured and adjusted. This is implemented in
2325 * tcp_rack_mark_lost.
2327 * If the receiver does not support SACK:
2329 * NewReno (RFC6582): in Recovery we assume that one segment
2330 * is lost (classic Reno). While we are in Recovery and
2331 * a partial ACK arrives, we assume that one more packet
2332 * is lost (NewReno). This heuristics are the same in NewReno
2335 * Really tricky (and requiring careful tuning) part of algorithm
2336 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2337 * The first determines the moment _when_ we should reduce CWND and,
2338 * hence, slow down forward transmission. In fact, it determines the moment
2339 * when we decide that hole is caused by loss, rather than by a reorder.
2341 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2342 * holes, caused by lost packets.
2344 * And the most logically complicated part of algorithm is undo
2345 * heuristics. We detect false retransmits due to both too early
2346 * fast retransmit (reordering) and underestimated RTO, analyzing
2347 * timestamps and D-SACKs. When we detect that some segments were
2348 * retransmitted by mistake and CWND reduction was wrong, we undo
2349 * window reduction and abort recovery phase. This logic is hidden
2350 * inside several functions named tcp_try_undo_<something>.
2353 /* This function decides, when we should leave Disordered state
2354 * and enter Recovery phase, reducing congestion window.
2356 * Main question: may we further continue forward transmission
2357 * with the same cwnd?
2359 static bool tcp_time_to_recover(struct sock *sk, int flag)
2361 struct tcp_sock *tp = tcp_sk(sk);
2363 /* Trick#1: The loss is proven. */
2367 /* Not-A-Trick#2 : Classic rule... */
2368 if (!tcp_is_rack(sk) && tcp_dupack_heuristics(tp) > tp->reordering)
2374 /* Detect loss in event "A" above by marking head of queue up as lost.
2375 * For RFC3517 SACK, a segment is considered lost if it
2376 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2377 * the maximum SACKed segments to pass before reaching this limit.
2379 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2381 struct tcp_sock *tp = tcp_sk(sk);
2382 struct sk_buff *skb;
2384 /* Use SACK to deduce losses of new sequences sent during recovery */
2385 const u32 loss_high = tp->snd_nxt;
2387 WARN_ON(packets > tp->packets_out);
2388 skb = tp->lost_skb_hint;
2390 /* Head already handled? */
2391 if (mark_head && after(TCP_SKB_CB(skb)->seq, tp->snd_una))
2393 cnt = tp->lost_cnt_hint;
2395 skb = tcp_rtx_queue_head(sk);
2399 skb_rbtree_walk_from(skb) {
2400 /* TODO: do this better */
2401 /* this is not the most efficient way to do this... */
2402 tp->lost_skb_hint = skb;
2403 tp->lost_cnt_hint = cnt;
2405 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2408 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2409 cnt += tcp_skb_pcount(skb);
2414 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_LOST))
2415 tcp_mark_skb_lost(sk, skb);
2420 tcp_verify_left_out(tp);
2423 /* Account newly detected lost packet(s) */
2425 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2427 struct tcp_sock *tp = tcp_sk(sk);
2429 if (tcp_is_sack(tp)) {
2430 int sacked_upto = tp->sacked_out - tp->reordering;
2431 if (sacked_upto >= 0)
2432 tcp_mark_head_lost(sk, sacked_upto, 0);
2433 else if (fast_rexmit)
2434 tcp_mark_head_lost(sk, 1, 1);
2438 static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when)
2440 return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2441 before(tp->rx_opt.rcv_tsecr, when);
2444 /* skb is spurious retransmitted if the returned timestamp echo
2445 * reply is prior to the skb transmission time
2447 static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp,
2448 const struct sk_buff *skb)
2450 return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) &&
2451 tcp_tsopt_ecr_before(tp, tcp_skb_timestamp_ts(tp->tcp_usec_ts, skb));
2454 /* Nothing was retransmitted or returned timestamp is less
2455 * than timestamp of the first retransmission.
2457 static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2459 return tp->retrans_stamp &&
2460 tcp_tsopt_ecr_before(tp, tp->retrans_stamp);
2463 /* Undo procedures. */
2465 /* We can clear retrans_stamp when there are no retransmissions in the
2466 * window. It would seem that it is trivially available for us in
2467 * tp->retrans_out, however, that kind of assumptions doesn't consider
2468 * what will happen if errors occur when sending retransmission for the
2469 * second time. ...It could the that such segment has only
2470 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2471 * the head skb is enough except for some reneging corner cases that
2472 * are not worth the effort.
2474 * Main reason for all this complexity is the fact that connection dying
2475 * time now depends on the validity of the retrans_stamp, in particular,
2476 * that successive retransmissions of a segment must not advance
2477 * retrans_stamp under any conditions.
2479 static bool tcp_any_retrans_done(const struct sock *sk)
2481 const struct tcp_sock *tp = tcp_sk(sk);
2482 struct sk_buff *skb;
2484 if (tp->retrans_out)
2487 skb = tcp_rtx_queue_head(sk);
2488 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2494 static void DBGUNDO(struct sock *sk, const char *msg)
2496 #if FASTRETRANS_DEBUG > 1
2497 struct tcp_sock *tp = tcp_sk(sk);
2498 struct inet_sock *inet = inet_sk(sk);
2500 if (sk->sk_family == AF_INET) {
2501 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2503 &inet->inet_daddr, ntohs(inet->inet_dport),
2504 tcp_snd_cwnd(tp), tcp_left_out(tp),
2505 tp->snd_ssthresh, tp->prior_ssthresh,
2508 #if IS_ENABLED(CONFIG_IPV6)
2509 else if (sk->sk_family == AF_INET6) {
2510 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2512 &sk->sk_v6_daddr, ntohs(inet->inet_dport),
2513 tcp_snd_cwnd(tp), tcp_left_out(tp),
2514 tp->snd_ssthresh, tp->prior_ssthresh,
2521 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
2523 struct tcp_sock *tp = tcp_sk(sk);
2526 struct sk_buff *skb;
2528 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2529 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2532 tcp_clear_all_retrans_hints(tp);
2535 if (tp->prior_ssthresh) {
2536 const struct inet_connection_sock *icsk = inet_csk(sk);
2538 tcp_snd_cwnd_set(tp, icsk->icsk_ca_ops->undo_cwnd(sk));
2540 if (tp->prior_ssthresh > tp->snd_ssthresh) {
2541 tp->snd_ssthresh = tp->prior_ssthresh;
2542 tcp_ecn_withdraw_cwr(tp);
2545 tp->snd_cwnd_stamp = tcp_jiffies32;
2546 tp->undo_marker = 0;
2547 tp->rack.advanced = 1; /* Force RACK to re-exam losses */
2550 static inline bool tcp_may_undo(const struct tcp_sock *tp)
2552 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2555 static bool tcp_is_non_sack_preventing_reopen(struct sock *sk)
2557 struct tcp_sock *tp = tcp_sk(sk);
2559 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2560 /* Hold old state until something *above* high_seq
2561 * is ACKed. For Reno it is MUST to prevent false
2562 * fast retransmits (RFC2582). SACK TCP is safe. */
2563 if (!tcp_any_retrans_done(sk))
2564 tp->retrans_stamp = 0;
2570 /* People celebrate: "We love our President!" */
2571 static bool tcp_try_undo_recovery(struct sock *sk)
2573 struct tcp_sock *tp = tcp_sk(sk);
2575 if (tcp_may_undo(tp)) {
2578 /* Happy end! We did not retransmit anything
2579 * or our original transmission succeeded.
2581 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2582 tcp_undo_cwnd_reduction(sk, false);
2583 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2584 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2586 mib_idx = LINUX_MIB_TCPFULLUNDO;
2588 NET_INC_STATS(sock_net(sk), mib_idx);
2589 } else if (tp->rack.reo_wnd_persist) {
2590 tp->rack.reo_wnd_persist--;
2592 if (tcp_is_non_sack_preventing_reopen(sk))
2594 tcp_set_ca_state(sk, TCP_CA_Open);
2595 tp->is_sack_reneg = 0;
2599 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2600 static bool tcp_try_undo_dsack(struct sock *sk)
2602 struct tcp_sock *tp = tcp_sk(sk);
2604 if (tp->undo_marker && !tp->undo_retrans) {
2605 tp->rack.reo_wnd_persist = min(TCP_RACK_RECOVERY_THRESH,
2606 tp->rack.reo_wnd_persist + 1);
2607 DBGUNDO(sk, "D-SACK");
2608 tcp_undo_cwnd_reduction(sk, false);
2609 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2615 /* Undo during loss recovery after partial ACK or using F-RTO. */
2616 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2618 struct tcp_sock *tp = tcp_sk(sk);
2620 if (frto_undo || tcp_may_undo(tp)) {
2621 tcp_undo_cwnd_reduction(sk, true);
2623 DBGUNDO(sk, "partial loss");
2624 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2626 NET_INC_STATS(sock_net(sk),
2627 LINUX_MIB_TCPSPURIOUSRTOS);
2628 inet_csk(sk)->icsk_retransmits = 0;
2629 if (tcp_is_non_sack_preventing_reopen(sk))
2631 if (frto_undo || tcp_is_sack(tp)) {
2632 tcp_set_ca_state(sk, TCP_CA_Open);
2633 tp->is_sack_reneg = 0;
2640 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
2641 * It computes the number of packets to send (sndcnt) based on packets newly
2643 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2644 * cwnd reductions across a full RTT.
2645 * 2) Otherwise PRR uses packet conservation to send as much as delivered.
2646 * But when SND_UNA is acked without further losses,
2647 * slow starts cwnd up to ssthresh to speed up the recovery.
2649 static void tcp_init_cwnd_reduction(struct sock *sk)
2651 struct tcp_sock *tp = tcp_sk(sk);
2653 tp->high_seq = tp->snd_nxt;
2654 tp->tlp_high_seq = 0;
2655 tp->snd_cwnd_cnt = 0;
2656 tp->prior_cwnd = tcp_snd_cwnd(tp);
2657 tp->prr_delivered = 0;
2659 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2660 tcp_ecn_queue_cwr(tp);
2663 void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked, int newly_lost, int flag)
2665 struct tcp_sock *tp = tcp_sk(sk);
2667 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2669 if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd))
2672 tp->prr_delivered += newly_acked_sacked;
2674 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2676 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2678 sndcnt = max_t(int, tp->prr_delivered - tp->prr_out,
2679 newly_acked_sacked);
2680 if (flag & FLAG_SND_UNA_ADVANCED && !newly_lost)
2682 sndcnt = min(delta, sndcnt);
2684 /* Force a fast retransmit upon entering fast recovery */
2685 sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1));
2686 tcp_snd_cwnd_set(tp, tcp_packets_in_flight(tp) + sndcnt);
2689 static inline void tcp_end_cwnd_reduction(struct sock *sk)
2691 struct tcp_sock *tp = tcp_sk(sk);
2693 if (inet_csk(sk)->icsk_ca_ops->cong_control)
2696 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2697 if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH &&
2698 (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR || tp->undo_marker)) {
2699 tcp_snd_cwnd_set(tp, tp->snd_ssthresh);
2700 tp->snd_cwnd_stamp = tcp_jiffies32;
2702 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2705 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2706 void tcp_enter_cwr(struct sock *sk)
2708 struct tcp_sock *tp = tcp_sk(sk);
2710 tp->prior_ssthresh = 0;
2711 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2712 tp->undo_marker = 0;
2713 tcp_init_cwnd_reduction(sk);
2714 tcp_set_ca_state(sk, TCP_CA_CWR);
2717 EXPORT_SYMBOL(tcp_enter_cwr);
2719 static void tcp_try_keep_open(struct sock *sk)
2721 struct tcp_sock *tp = tcp_sk(sk);
2722 int state = TCP_CA_Open;
2724 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2725 state = TCP_CA_Disorder;
2727 if (inet_csk(sk)->icsk_ca_state != state) {
2728 tcp_set_ca_state(sk, state);
2729 tp->high_seq = tp->snd_nxt;
2733 static void tcp_try_to_open(struct sock *sk, int flag)
2735 struct tcp_sock *tp = tcp_sk(sk);
2737 tcp_verify_left_out(tp);
2739 if (!tcp_any_retrans_done(sk))
2740 tp->retrans_stamp = 0;
2742 if (flag & FLAG_ECE)
2745 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2746 tcp_try_keep_open(sk);
2750 static void tcp_mtup_probe_failed(struct sock *sk)
2752 struct inet_connection_sock *icsk = inet_csk(sk);
2754 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2755 icsk->icsk_mtup.probe_size = 0;
2756 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPFAIL);
2759 static void tcp_mtup_probe_success(struct sock *sk)
2761 struct tcp_sock *tp = tcp_sk(sk);
2762 struct inet_connection_sock *icsk = inet_csk(sk);
2765 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2767 val = (u64)tcp_snd_cwnd(tp) * tcp_mss_to_mtu(sk, tp->mss_cache);
2768 do_div(val, icsk->icsk_mtup.probe_size);
2769 DEBUG_NET_WARN_ON_ONCE((u32)val != val);
2770 tcp_snd_cwnd_set(tp, max_t(u32, 1U, val));
2772 tp->snd_cwnd_cnt = 0;
2773 tp->snd_cwnd_stamp = tcp_jiffies32;
2774 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2776 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2777 icsk->icsk_mtup.probe_size = 0;
2778 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2779 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS);
2782 /* Do a simple retransmit without using the backoff mechanisms in
2783 * tcp_timer. This is used for path mtu discovery.
2784 * The socket is already locked here.
2786 void tcp_simple_retransmit(struct sock *sk)
2788 const struct inet_connection_sock *icsk = inet_csk(sk);
2789 struct tcp_sock *tp = tcp_sk(sk);
2790 struct sk_buff *skb;
2793 /* A fastopen SYN request is stored as two separate packets within
2794 * the retransmit queue, this is done by tcp_send_syn_data().
2795 * As a result simply checking the MSS of the frames in the queue
2796 * will not work for the SYN packet.
2798 * Us being here is an indication of a path MTU issue so we can
2799 * assume that the fastopen SYN was lost and just mark all the
2800 * frames in the retransmit queue as lost. We will use an MSS of
2801 * -1 to mark all frames as lost, otherwise compute the current MSS.
2803 if (tp->syn_data && sk->sk_state == TCP_SYN_SENT)
2806 mss = tcp_current_mss(sk);
2808 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2809 if (tcp_skb_seglen(skb) > mss)
2810 tcp_mark_skb_lost(sk, skb);
2813 tcp_clear_retrans_hints_partial(tp);
2818 if (tcp_is_reno(tp))
2819 tcp_limit_reno_sacked(tp);
2821 tcp_verify_left_out(tp);
2823 /* Don't muck with the congestion window here.
2824 * Reason is that we do not increase amount of _data_
2825 * in network, but units changed and effective
2826 * cwnd/ssthresh really reduced now.
2828 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2829 tp->high_seq = tp->snd_nxt;
2830 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2831 tp->prior_ssthresh = 0;
2832 tp->undo_marker = 0;
2833 tcp_set_ca_state(sk, TCP_CA_Loss);
2835 tcp_xmit_retransmit_queue(sk);
2837 EXPORT_SYMBOL(tcp_simple_retransmit);
2839 void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2841 struct tcp_sock *tp = tcp_sk(sk);
2844 if (tcp_is_reno(tp))
2845 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2847 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2849 NET_INC_STATS(sock_net(sk), mib_idx);
2851 tp->prior_ssthresh = 0;
2854 if (!tcp_in_cwnd_reduction(sk)) {
2856 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2857 tcp_init_cwnd_reduction(sk);
2859 tcp_set_ca_state(sk, TCP_CA_Recovery);
2862 static void tcp_update_rto_time(struct tcp_sock *tp)
2864 if (tp->rto_stamp) {
2865 tp->total_rto_time += tcp_time_stamp_ms(tp) - tp->rto_stamp;
2870 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2871 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2873 static void tcp_process_loss(struct sock *sk, int flag, int num_dupack,
2876 struct tcp_sock *tp = tcp_sk(sk);
2877 bool recovered = !before(tp->snd_una, tp->high_seq);
2879 if ((flag & FLAG_SND_UNA_ADVANCED || rcu_access_pointer(tp->fastopen_rsk)) &&
2880 tcp_try_undo_loss(sk, false))
2883 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2884 /* Step 3.b. A timeout is spurious if not all data are
2885 * lost, i.e., never-retransmitted data are (s)acked.
2887 if ((flag & FLAG_ORIG_SACK_ACKED) &&
2888 tcp_try_undo_loss(sk, true))
2891 if (after(tp->snd_nxt, tp->high_seq)) {
2892 if (flag & FLAG_DATA_SACKED || num_dupack)
2893 tp->frto = 0; /* Step 3.a. loss was real */
2894 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2895 tp->high_seq = tp->snd_nxt;
2896 /* Step 2.b. Try send new data (but deferred until cwnd
2897 * is updated in tcp_ack()). Otherwise fall back to
2898 * the conventional recovery.
2900 if (!tcp_write_queue_empty(sk) &&
2901 after(tcp_wnd_end(tp), tp->snd_nxt)) {
2902 *rexmit = REXMIT_NEW;
2910 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2911 tcp_try_undo_recovery(sk);
2914 if (tcp_is_reno(tp)) {
2915 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2916 * delivered. Lower inflight to clock out (re)transmissions.
2918 if (after(tp->snd_nxt, tp->high_seq) && num_dupack)
2919 tcp_add_reno_sack(sk, num_dupack, flag & FLAG_ECE);
2920 else if (flag & FLAG_SND_UNA_ADVANCED)
2921 tcp_reset_reno_sack(tp);
2923 *rexmit = REXMIT_LOST;
2926 static bool tcp_force_fast_retransmit(struct sock *sk)
2928 struct tcp_sock *tp = tcp_sk(sk);
2930 return after(tcp_highest_sack_seq(tp),
2931 tp->snd_una + tp->reordering * tp->mss_cache);
2934 /* Undo during fast recovery after partial ACK. */
2935 static bool tcp_try_undo_partial(struct sock *sk, u32 prior_snd_una,
2938 struct tcp_sock *tp = tcp_sk(sk);
2940 if (tp->undo_marker && tcp_packet_delayed(tp)) {
2941 /* Plain luck! Hole if filled with delayed
2942 * packet, rather than with a retransmit. Check reordering.
2944 tcp_check_sack_reordering(sk, prior_snd_una, 1);
2946 /* We are getting evidence that the reordering degree is higher
2947 * than we realized. If there are no retransmits out then we
2948 * can undo. Otherwise we clock out new packets but do not
2949 * mark more packets lost or retransmit more.
2951 if (tp->retrans_out)
2954 if (!tcp_any_retrans_done(sk))
2955 tp->retrans_stamp = 0;
2957 DBGUNDO(sk, "partial recovery");
2958 tcp_undo_cwnd_reduction(sk, true);
2959 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2960 tcp_try_keep_open(sk);
2962 /* Partial ACK arrived. Force fast retransmit. */
2963 *do_lost = tcp_force_fast_retransmit(sk);
2968 static void tcp_identify_packet_loss(struct sock *sk, int *ack_flag)
2970 struct tcp_sock *tp = tcp_sk(sk);
2972 if (tcp_rtx_queue_empty(sk))
2975 if (unlikely(tcp_is_reno(tp))) {
2976 tcp_newreno_mark_lost(sk, *ack_flag & FLAG_SND_UNA_ADVANCED);
2977 } else if (tcp_is_rack(sk)) {
2978 u32 prior_retrans = tp->retrans_out;
2980 if (tcp_rack_mark_lost(sk))
2981 *ack_flag &= ~FLAG_SET_XMIT_TIMER;
2982 if (prior_retrans > tp->retrans_out)
2983 *ack_flag |= FLAG_LOST_RETRANS;
2987 /* Process an event, which can update packets-in-flight not trivially.
2988 * Main goal of this function is to calculate new estimate for left_out,
2989 * taking into account both packets sitting in receiver's buffer and
2990 * packets lost by network.
2992 * Besides that it updates the congestion state when packet loss or ECN
2993 * is detected. But it does not reduce the cwnd, it is done by the
2994 * congestion control later.
2996 * It does _not_ decide what to send, it is made in function
2997 * tcp_xmit_retransmit_queue().
2999 static void tcp_fastretrans_alert(struct sock *sk, const u32 prior_snd_una,
3000 int num_dupack, int *ack_flag, int *rexmit)
3002 struct inet_connection_sock *icsk = inet_csk(sk);
3003 struct tcp_sock *tp = tcp_sk(sk);
3004 int fast_rexmit = 0, flag = *ack_flag;
3005 bool ece_ack = flag & FLAG_ECE;
3006 bool do_lost = num_dupack || ((flag & FLAG_DATA_SACKED) &&
3007 tcp_force_fast_retransmit(sk));
3009 if (!tp->packets_out && tp->sacked_out)
3012 /* Now state machine starts.
3013 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
3015 tp->prior_ssthresh = 0;
3017 /* B. In all the states check for reneging SACKs. */
3018 if (tcp_check_sack_reneging(sk, ack_flag))
3021 /* C. Check consistency of the current state. */
3022 tcp_verify_left_out(tp);
3024 /* D. Check state exit conditions. State can be terminated
3025 * when high_seq is ACKed. */
3026 if (icsk->icsk_ca_state == TCP_CA_Open) {
3027 WARN_ON(tp->retrans_out != 0 && !tp->syn_data);
3028 tp->retrans_stamp = 0;
3029 } else if (!before(tp->snd_una, tp->high_seq)) {
3030 switch (icsk->icsk_ca_state) {
3032 /* CWR is to be held something *above* high_seq
3033 * is ACKed for CWR bit to reach receiver. */
3034 if (tp->snd_una != tp->high_seq) {
3035 tcp_end_cwnd_reduction(sk);
3036 tcp_set_ca_state(sk, TCP_CA_Open);
3040 case TCP_CA_Recovery:
3041 if (tcp_is_reno(tp))
3042 tcp_reset_reno_sack(tp);
3043 if (tcp_try_undo_recovery(sk))
3045 tcp_end_cwnd_reduction(sk);
3050 /* E. Process state. */
3051 switch (icsk->icsk_ca_state) {
3052 case TCP_CA_Recovery:
3053 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
3054 if (tcp_is_reno(tp))
3055 tcp_add_reno_sack(sk, num_dupack, ece_ack);
3056 } else if (tcp_try_undo_partial(sk, prior_snd_una, &do_lost))
3059 if (tcp_try_undo_dsack(sk))
3060 tcp_try_keep_open(sk);
3062 tcp_identify_packet_loss(sk, ack_flag);
3063 if (icsk->icsk_ca_state != TCP_CA_Recovery) {
3064 if (!tcp_time_to_recover(sk, flag))
3066 /* Undo reverts the recovery state. If loss is evident,
3067 * starts a new recovery (e.g. reordering then loss);
3069 tcp_enter_recovery(sk, ece_ack);
3073 tcp_process_loss(sk, flag, num_dupack, rexmit);
3074 if (icsk->icsk_ca_state != TCP_CA_Loss)
3075 tcp_update_rto_time(tp);
3076 tcp_identify_packet_loss(sk, ack_flag);
3077 if (!(icsk->icsk_ca_state == TCP_CA_Open ||
3078 (*ack_flag & FLAG_LOST_RETRANS)))
3080 /* Change state if cwnd is undone or retransmits are lost */
3083 if (tcp_is_reno(tp)) {
3084 if (flag & FLAG_SND_UNA_ADVANCED)
3085 tcp_reset_reno_sack(tp);
3086 tcp_add_reno_sack(sk, num_dupack, ece_ack);
3089 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
3090 tcp_try_undo_dsack(sk);
3092 tcp_identify_packet_loss(sk, ack_flag);
3093 if (!tcp_time_to_recover(sk, flag)) {
3094 tcp_try_to_open(sk, flag);
3098 /* MTU probe failure: don't reduce cwnd */
3099 if (icsk->icsk_ca_state < TCP_CA_CWR &&
3100 icsk->icsk_mtup.probe_size &&
3101 tp->snd_una == tp->mtu_probe.probe_seq_start) {
3102 tcp_mtup_probe_failed(sk);
3103 /* Restores the reduction we did in tcp_mtup_probe() */
3104 tcp_snd_cwnd_set(tp, tcp_snd_cwnd(tp) + 1);
3105 tcp_simple_retransmit(sk);
3109 /* Otherwise enter Recovery state */
3110 tcp_enter_recovery(sk, ece_ack);
3114 if (!tcp_is_rack(sk) && do_lost)
3115 tcp_update_scoreboard(sk, fast_rexmit);
3116 *rexmit = REXMIT_LOST;
3119 static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us, const int flag)
3121 u32 wlen = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_rtt_wlen) * HZ;
3122 struct tcp_sock *tp = tcp_sk(sk);
3124 if ((flag & FLAG_ACK_MAYBE_DELAYED) && rtt_us > tcp_min_rtt(tp)) {
3125 /* If the remote keeps returning delayed ACKs, eventually
3126 * the min filter would pick it up and overestimate the
3127 * prop. delay when it expires. Skip suspected delayed ACKs.
3131 minmax_running_min(&tp->rtt_min, wlen, tcp_jiffies32,
3132 rtt_us ? : jiffies_to_usecs(1));
3135 static bool tcp_ack_update_rtt(struct sock *sk, const int flag,
3136 long seq_rtt_us, long sack_rtt_us,
3137 long ca_rtt_us, struct rate_sample *rs)
3139 const struct tcp_sock *tp = tcp_sk(sk);
3141 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
3142 * broken middle-boxes or peers may corrupt TS-ECR fields. But
3143 * Karn's algorithm forbids taking RTT if some retransmitted data
3144 * is acked (RFC6298).
3147 seq_rtt_us = sack_rtt_us;
3149 /* RTTM Rule: A TSecr value received in a segment is used to
3150 * update the averaged RTT measurement only if the segment
3151 * acknowledges some new data, i.e., only if it advances the
3152 * left edge of the send window.
3153 * See draft-ietf-tcplw-high-performance-00, section 3.3.
3155 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp &&
3156 tp->rx_opt.rcv_tsecr && flag & FLAG_ACKED)
3157 seq_rtt_us = ca_rtt_us = tcp_rtt_tsopt_us(tp);
3159 rs->rtt_us = ca_rtt_us; /* RTT of last (S)ACKed packet (or -1) */
3163 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
3164 * always taken together with ACK, SACK, or TS-opts. Any negative
3165 * values will be skipped with the seq_rtt_us < 0 check above.
3167 tcp_update_rtt_min(sk, ca_rtt_us, flag);
3168 tcp_rtt_estimator(sk, seq_rtt_us);
3171 /* RFC6298: only reset backoff on valid RTT measurement. */
3172 inet_csk(sk)->icsk_backoff = 0;
3176 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
3177 void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req)
3179 struct rate_sample rs;
3182 if (req && !req->num_retrans && tcp_rsk(req)->snt_synack)
3183 rtt_us = tcp_stamp_us_delta(tcp_clock_us(), tcp_rsk(req)->snt_synack);
3185 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us, &rs);
3189 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
3191 const struct inet_connection_sock *icsk = inet_csk(sk);
3193 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
3194 tcp_sk(sk)->snd_cwnd_stamp = tcp_jiffies32;
3197 /* Restart timer after forward progress on connection.
3198 * RFC2988 recommends to restart timer to now+rto.
3200 void tcp_rearm_rto(struct sock *sk)
3202 const struct inet_connection_sock *icsk = inet_csk(sk);
3203 struct tcp_sock *tp = tcp_sk(sk);
3205 /* If the retrans timer is currently being used by Fast Open
3206 * for SYN-ACK retrans purpose, stay put.
3208 if (rcu_access_pointer(tp->fastopen_rsk))
3211 if (!tp->packets_out) {
3212 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
3214 u32 rto = inet_csk(sk)->icsk_rto;
3215 /* Offset the time elapsed after installing regular RTO */
3216 if (icsk->icsk_pending == ICSK_TIME_REO_TIMEOUT ||
3217 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
3218 s64 delta_us = tcp_rto_delta_us(sk);
3219 /* delta_us may not be positive if the socket is locked
3220 * when the retrans timer fires and is rescheduled.
3222 rto = usecs_to_jiffies(max_t(int, delta_us, 1));
3224 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
3229 /* Try to schedule a loss probe; if that doesn't work, then schedule an RTO. */
3230 static void tcp_set_xmit_timer(struct sock *sk)
3232 if (!tcp_schedule_loss_probe(sk, true))
3236 /* If we get here, the whole TSO packet has not been acked. */
3237 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3239 struct tcp_sock *tp = tcp_sk(sk);
3242 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3244 packets_acked = tcp_skb_pcount(skb);
3245 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3247 packets_acked -= tcp_skb_pcount(skb);
3249 if (packets_acked) {
3250 BUG_ON(tcp_skb_pcount(skb) == 0);
3251 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3254 return packets_acked;
3257 static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
3258 const struct sk_buff *ack_skb, u32 prior_snd_una)
3260 const struct skb_shared_info *shinfo;
3262 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3263 if (likely(!TCP_SKB_CB(skb)->txstamp_ack))
3266 shinfo = skb_shinfo(skb);
3267 if (!before(shinfo->tskey, prior_snd_una) &&
3268 before(shinfo->tskey, tcp_sk(sk)->snd_una)) {
3269 tcp_skb_tsorted_save(skb) {
3270 __skb_tstamp_tx(skb, ack_skb, NULL, sk, SCM_TSTAMP_ACK);
3271 } tcp_skb_tsorted_restore(skb);
3275 /* Remove acknowledged frames from the retransmission queue. If our packet
3276 * is before the ack sequence we can discard it as it's confirmed to have
3277 * arrived at the other end.
3279 static int tcp_clean_rtx_queue(struct sock *sk, const struct sk_buff *ack_skb,
3280 u32 prior_fack, u32 prior_snd_una,
3281 struct tcp_sacktag_state *sack, bool ece_ack)
3283 const struct inet_connection_sock *icsk = inet_csk(sk);
3284 u64 first_ackt, last_ackt;
3285 struct tcp_sock *tp = tcp_sk(sk);
3286 u32 prior_sacked = tp->sacked_out;
3287 u32 reord = tp->snd_nxt; /* lowest acked un-retx un-sacked seq */
3288 struct sk_buff *skb, *next;
3289 bool fully_acked = true;
3290 long sack_rtt_us = -1L;
3291 long seq_rtt_us = -1L;
3292 long ca_rtt_us = -1L;
3299 for (skb = skb_rb_first(&sk->tcp_rtx_queue); skb; skb = next) {
3300 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3301 const u32 start_seq = scb->seq;
3302 u8 sacked = scb->sacked;
3305 /* Determine how many packets and what bytes were acked, tso and else */
3306 if (after(scb->end_seq, tp->snd_una)) {
3307 if (tcp_skb_pcount(skb) == 1 ||
3308 !after(tp->snd_una, scb->seq))
3311 acked_pcount = tcp_tso_acked(sk, skb);
3314 fully_acked = false;
3316 acked_pcount = tcp_skb_pcount(skb);
3319 if (unlikely(sacked & TCPCB_RETRANS)) {
3320 if (sacked & TCPCB_SACKED_RETRANS)
3321 tp->retrans_out -= acked_pcount;
3322 flag |= FLAG_RETRANS_DATA_ACKED;
3323 } else if (!(sacked & TCPCB_SACKED_ACKED)) {
3324 last_ackt = tcp_skb_timestamp_us(skb);
3325 WARN_ON_ONCE(last_ackt == 0);
3327 first_ackt = last_ackt;
3329 if (before(start_seq, reord))
3331 if (!after(scb->end_seq, tp->high_seq))
3332 flag |= FLAG_ORIG_SACK_ACKED;
3335 if (sacked & TCPCB_SACKED_ACKED) {
3336 tp->sacked_out -= acked_pcount;
3337 } else if (tcp_is_sack(tp)) {
3338 tcp_count_delivered(tp, acked_pcount, ece_ack);
3339 if (!tcp_skb_spurious_retrans(tp, skb))
3340 tcp_rack_advance(tp, sacked, scb->end_seq,
3341 tcp_skb_timestamp_us(skb));
3343 if (sacked & TCPCB_LOST)
3344 tp->lost_out -= acked_pcount;
3346 tp->packets_out -= acked_pcount;
3347 pkts_acked += acked_pcount;
3348 tcp_rate_skb_delivered(sk, skb, sack->rate);
3350 /* Initial outgoing SYN's get put onto the write_queue
3351 * just like anything else we transmit. It is not
3352 * true data, and if we misinform our callers that
3353 * this ACK acks real data, we will erroneously exit
3354 * connection startup slow start one packet too
3355 * quickly. This is severely frowned upon behavior.
3357 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
3358 flag |= FLAG_DATA_ACKED;
3360 flag |= FLAG_SYN_ACKED;
3361 tp->retrans_stamp = 0;
3367 tcp_ack_tstamp(sk, skb, ack_skb, prior_snd_una);
3369 next = skb_rb_next(skb);
3370 if (unlikely(skb == tp->retransmit_skb_hint))
3371 tp->retransmit_skb_hint = NULL;
3372 if (unlikely(skb == tp->lost_skb_hint))
3373 tp->lost_skb_hint = NULL;
3374 tcp_highest_sack_replace(sk, skb, next);
3375 tcp_rtx_queue_unlink_and_free(skb, sk);
3379 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
3381 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3382 tp->snd_up = tp->snd_una;
3385 tcp_ack_tstamp(sk, skb, ack_skb, prior_snd_una);
3386 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
3387 flag |= FLAG_SACK_RENEGING;
3390 if (likely(first_ackt) && !(flag & FLAG_RETRANS_DATA_ACKED)) {
3391 seq_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, first_ackt);
3392 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, last_ackt);
3394 if (pkts_acked == 1 && fully_acked && !prior_sacked &&
3395 (tp->snd_una - prior_snd_una) < tp->mss_cache &&
3396 sack->rate->prior_delivered + 1 == tp->delivered &&
3397 !(flag & (FLAG_CA_ALERT | FLAG_SYN_ACKED))) {
3398 /* Conservatively mark a delayed ACK. It's typically
3399 * from a lone runt packet over the round trip to
3400 * a receiver w/o out-of-order or CE events.
3402 flag |= FLAG_ACK_MAYBE_DELAYED;
3405 if (sack->first_sackt) {
3406 sack_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->first_sackt);
3407 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->last_sackt);
3409 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us,
3410 ca_rtt_us, sack->rate);
3412 if (flag & FLAG_ACKED) {
3413 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3414 if (unlikely(icsk->icsk_mtup.probe_size &&
3415 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3416 tcp_mtup_probe_success(sk);
3419 if (tcp_is_reno(tp)) {
3420 tcp_remove_reno_sacks(sk, pkts_acked, ece_ack);
3422 /* If any of the cumulatively ACKed segments was
3423 * retransmitted, non-SACK case cannot confirm that
3424 * progress was due to original transmission due to
3425 * lack of TCPCB_SACKED_ACKED bits even if some of
3426 * the packets may have been never retransmitted.
3428 if (flag & FLAG_RETRANS_DATA_ACKED)
3429 flag &= ~FLAG_ORIG_SACK_ACKED;
3433 /* Non-retransmitted hole got filled? That's reordering */
3434 if (before(reord, prior_fack))
3435 tcp_check_sack_reordering(sk, reord, 0);
3437 delta = prior_sacked - tp->sacked_out;
3438 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3440 } else if (skb && rtt_update && sack_rtt_us >= 0 &&
3441 sack_rtt_us > tcp_stamp_us_delta(tp->tcp_mstamp,
3442 tcp_skb_timestamp_us(skb))) {
3443 /* Do not re-arm RTO if the sack RTT is measured from data sent
3444 * after when the head was last (re)transmitted. Otherwise the
3445 * timeout may continue to extend in loss recovery.
3447 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3450 if (icsk->icsk_ca_ops->pkts_acked) {
3451 struct ack_sample sample = { .pkts_acked = pkts_acked,
3452 .rtt_us = sack->rate->rtt_us };
3454 sample.in_flight = tp->mss_cache *
3455 (tp->delivered - sack->rate->prior_delivered);
3456 icsk->icsk_ca_ops->pkts_acked(sk, &sample);
3459 #if FASTRETRANS_DEBUG > 0
3460 WARN_ON((int)tp->sacked_out < 0);
3461 WARN_ON((int)tp->lost_out < 0);
3462 WARN_ON((int)tp->retrans_out < 0);
3463 if (!tp->packets_out && tcp_is_sack(tp)) {
3464 icsk = inet_csk(sk);
3466 pr_debug("Leak l=%u %d\n",
3467 tp->lost_out, icsk->icsk_ca_state);
3470 if (tp->sacked_out) {
3471 pr_debug("Leak s=%u %d\n",
3472 tp->sacked_out, icsk->icsk_ca_state);
3475 if (tp->retrans_out) {
3476 pr_debug("Leak r=%u %d\n",
3477 tp->retrans_out, icsk->icsk_ca_state);
3478 tp->retrans_out = 0;
3485 static void tcp_ack_probe(struct sock *sk)
3487 struct inet_connection_sock *icsk = inet_csk(sk);
3488 struct sk_buff *head = tcp_send_head(sk);
3489 const struct tcp_sock *tp = tcp_sk(sk);
3491 /* Was it a usable window open? */
3494 if (!after(TCP_SKB_CB(head)->end_seq, tcp_wnd_end(tp))) {
3495 icsk->icsk_backoff = 0;
3496 icsk->icsk_probes_tstamp = 0;
3497 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3498 /* Socket must be waked up by subsequent tcp_data_snd_check().
3499 * This function is not for random using!
3502 unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX);
3504 when = tcp_clamp_probe0_to_user_timeout(sk, when);
3505 tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, when, TCP_RTO_MAX);
3509 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3511 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3512 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3515 /* Decide wheather to run the increase function of congestion control. */
3516 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3518 /* If reordering is high then always grow cwnd whenever data is
3519 * delivered regardless of its ordering. Otherwise stay conservative
3520 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3521 * new SACK or ECE mark may first advance cwnd here and later reduce
3522 * cwnd in tcp_fastretrans_alert() based on more states.
3524 if (tcp_sk(sk)->reordering >
3525 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_reordering))
3526 return flag & FLAG_FORWARD_PROGRESS;
3528 return flag & FLAG_DATA_ACKED;
3531 /* The "ultimate" congestion control function that aims to replace the rigid
3532 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction).
3533 * It's called toward the end of processing an ACK with precise rate
3534 * information. All transmission or retransmission are delayed afterwards.
3536 static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked,
3537 int flag, const struct rate_sample *rs)
3539 const struct inet_connection_sock *icsk = inet_csk(sk);
3541 if (icsk->icsk_ca_ops->cong_control) {
3542 icsk->icsk_ca_ops->cong_control(sk, rs);
3546 if (tcp_in_cwnd_reduction(sk)) {
3547 /* Reduce cwnd if state mandates */
3548 tcp_cwnd_reduction(sk, acked_sacked, rs->losses, flag);
3549 } else if (tcp_may_raise_cwnd(sk, flag)) {
3550 /* Advance cwnd if state allows */
3551 tcp_cong_avoid(sk, ack, acked_sacked);
3553 tcp_update_pacing_rate(sk);
3556 /* Check that window update is acceptable.
3557 * The function assumes that snd_una<=ack<=snd_next.
3559 static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3560 const u32 ack, const u32 ack_seq,
3563 return after(ack, tp->snd_una) ||
3564 after(ack_seq, tp->snd_wl1) ||
3565 (ack_seq == tp->snd_wl1 && (nwin > tp->snd_wnd || !nwin));
3568 static void tcp_snd_sne_update(struct tcp_sock *tp, u32 ack)
3570 #ifdef CONFIG_TCP_AO
3571 struct tcp_ao_info *ao;
3573 if (!static_branch_unlikely(&tcp_ao_needed.key))
3576 ao = rcu_dereference_protected(tp->ao_info,
3577 lockdep_sock_is_held((struct sock *)tp));
3578 if (ao && ack < tp->snd_una)
3583 /* If we update tp->snd_una, also update tp->bytes_acked */
3584 static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack)
3586 u32 delta = ack - tp->snd_una;
3588 sock_owned_by_me((struct sock *)tp);
3589 tp->bytes_acked += delta;
3590 tcp_snd_sne_update(tp, ack);
3594 static void tcp_rcv_sne_update(struct tcp_sock *tp, u32 seq)
3596 #ifdef CONFIG_TCP_AO
3597 struct tcp_ao_info *ao;
3599 if (!static_branch_unlikely(&tcp_ao_needed.key))
3602 ao = rcu_dereference_protected(tp->ao_info,
3603 lockdep_sock_is_held((struct sock *)tp));
3604 if (ao && seq < tp->rcv_nxt)
3609 /* If we update tp->rcv_nxt, also update tp->bytes_received */
3610 static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq)
3612 u32 delta = seq - tp->rcv_nxt;
3614 sock_owned_by_me((struct sock *)tp);
3615 tp->bytes_received += delta;
3616 tcp_rcv_sne_update(tp, seq);
3617 WRITE_ONCE(tp->rcv_nxt, seq);
3620 /* Update our send window.
3622 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3623 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3625 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3628 struct tcp_sock *tp = tcp_sk(sk);
3630 u32 nwin = ntohs(tcp_hdr(skb)->window);
3632 if (likely(!tcp_hdr(skb)->syn))
3633 nwin <<= tp->rx_opt.snd_wscale;
3635 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3636 flag |= FLAG_WIN_UPDATE;
3637 tcp_update_wl(tp, ack_seq);
3639 if (tp->snd_wnd != nwin) {
3642 /* Note, it is the only place, where
3643 * fast path is recovered for sending TCP.
3646 tcp_fast_path_check(sk);
3648 if (!tcp_write_queue_empty(sk))
3649 tcp_slow_start_after_idle_check(sk);
3651 if (nwin > tp->max_window) {
3652 tp->max_window = nwin;
3653 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3658 tcp_snd_una_update(tp, ack);
3663 static bool __tcp_oow_rate_limited(struct net *net, int mib_idx,
3664 u32 *last_oow_ack_time)
3666 /* Paired with the WRITE_ONCE() in this function. */
3667 u32 val = READ_ONCE(*last_oow_ack_time);
3670 s32 elapsed = (s32)(tcp_jiffies32 - val);
3673 elapsed < READ_ONCE(net->ipv4.sysctl_tcp_invalid_ratelimit)) {
3674 NET_INC_STATS(net, mib_idx);
3675 return true; /* rate-limited: don't send yet! */
3679 /* Paired with the prior READ_ONCE() and with itself,
3680 * as we might be lockless.
3682 WRITE_ONCE(*last_oow_ack_time, tcp_jiffies32);
3684 return false; /* not rate-limited: go ahead, send dupack now! */
3687 /* Return true if we're currently rate-limiting out-of-window ACKs and
3688 * thus shouldn't send a dupack right now. We rate-limit dupacks in
3689 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3690 * attacks that send repeated SYNs or ACKs for the same connection. To
3691 * do this, we do not send a duplicate SYNACK or ACK if the remote
3692 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3694 bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb,
3695 int mib_idx, u32 *last_oow_ack_time)
3697 /* Data packets without SYNs are not likely part of an ACK loop. */
3698 if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) &&
3702 return __tcp_oow_rate_limited(net, mib_idx, last_oow_ack_time);
3705 /* RFC 5961 7 [ACK Throttling] */
3706 static void tcp_send_challenge_ack(struct sock *sk)
3708 struct tcp_sock *tp = tcp_sk(sk);
3709 struct net *net = sock_net(sk);
3710 u32 count, now, ack_limit;
3712 /* First check our per-socket dupack rate limit. */
3713 if (__tcp_oow_rate_limited(net,
3714 LINUX_MIB_TCPACKSKIPPEDCHALLENGE,
3715 &tp->last_oow_ack_time))
3718 ack_limit = READ_ONCE(net->ipv4.sysctl_tcp_challenge_ack_limit);
3719 if (ack_limit == INT_MAX)
3722 /* Then check host-wide RFC 5961 rate limit. */
3724 if (now != READ_ONCE(net->ipv4.tcp_challenge_timestamp)) {
3725 u32 half = (ack_limit + 1) >> 1;
3727 WRITE_ONCE(net->ipv4.tcp_challenge_timestamp, now);
3728 WRITE_ONCE(net->ipv4.tcp_challenge_count,
3729 get_random_u32_inclusive(half, ack_limit + half - 1));
3731 count = READ_ONCE(net->ipv4.tcp_challenge_count);
3733 WRITE_ONCE(net->ipv4.tcp_challenge_count, count - 1);
3735 NET_INC_STATS(net, LINUX_MIB_TCPCHALLENGEACK);
3740 static void tcp_store_ts_recent(struct tcp_sock *tp)
3742 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3743 tp->rx_opt.ts_recent_stamp = ktime_get_seconds();
3746 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3748 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3749 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3750 * extra check below makes sure this can only happen
3751 * for pure ACK frames. -DaveM
3753 * Not only, also it occurs for expired timestamps.
3756 if (tcp_paws_check(&tp->rx_opt, 0))
3757 tcp_store_ts_recent(tp);
3761 /* This routine deals with acks during a TLP episode and ends an episode by
3762 * resetting tlp_high_seq. Ref: TLP algorithm in draft-ietf-tcpm-rack
3764 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3766 struct tcp_sock *tp = tcp_sk(sk);
3768 if (before(ack, tp->tlp_high_seq))
3771 if (!tp->tlp_retrans) {
3772 /* TLP of new data has been acknowledged */
3773 tp->tlp_high_seq = 0;
3774 } else if (flag & FLAG_DSACK_TLP) {
3775 /* This DSACK means original and TLP probe arrived; no loss */
3776 tp->tlp_high_seq = 0;
3777 } else if (after(ack, tp->tlp_high_seq)) {
3778 /* ACK advances: there was a loss, so reduce cwnd. Reset
3779 * tlp_high_seq in tcp_init_cwnd_reduction()
3781 tcp_init_cwnd_reduction(sk);
3782 tcp_set_ca_state(sk, TCP_CA_CWR);
3783 tcp_end_cwnd_reduction(sk);
3784 tcp_try_keep_open(sk);
3785 NET_INC_STATS(sock_net(sk),
3786 LINUX_MIB_TCPLOSSPROBERECOVERY);
3787 } else if (!(flag & (FLAG_SND_UNA_ADVANCED |
3788 FLAG_NOT_DUP | FLAG_DATA_SACKED))) {
3789 /* Pure dupack: original and TLP probe arrived; no loss */
3790 tp->tlp_high_seq = 0;
3794 static inline void tcp_in_ack_event(struct sock *sk, u32 flags)
3796 const struct inet_connection_sock *icsk = inet_csk(sk);
3798 if (icsk->icsk_ca_ops->in_ack_event)
3799 icsk->icsk_ca_ops->in_ack_event(sk, flags);
3802 /* Congestion control has updated the cwnd already. So if we're in
3803 * loss recovery then now we do any new sends (for FRTO) or
3804 * retransmits (for CA_Loss or CA_recovery) that make sense.
3806 static void tcp_xmit_recovery(struct sock *sk, int rexmit)
3808 struct tcp_sock *tp = tcp_sk(sk);
3810 if (rexmit == REXMIT_NONE || sk->sk_state == TCP_SYN_SENT)
3813 if (unlikely(rexmit == REXMIT_NEW)) {
3814 __tcp_push_pending_frames(sk, tcp_current_mss(sk),
3816 if (after(tp->snd_nxt, tp->high_seq))
3820 tcp_xmit_retransmit_queue(sk);
3823 /* Returns the number of packets newly acked or sacked by the current ACK */
3824 static u32 tcp_newly_delivered(struct sock *sk, u32 prior_delivered, int flag)
3826 const struct net *net = sock_net(sk);
3827 struct tcp_sock *tp = tcp_sk(sk);
3830 delivered = tp->delivered - prior_delivered;
3831 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVERED, delivered);
3832 if (flag & FLAG_ECE)
3833 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVEREDCE, delivered);
3838 /* This routine deals with incoming acks, but not outgoing ones. */
3839 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3841 struct inet_connection_sock *icsk = inet_csk(sk);
3842 struct tcp_sock *tp = tcp_sk(sk);
3843 struct tcp_sacktag_state sack_state;
3844 struct rate_sample rs = { .prior_delivered = 0 };
3845 u32 prior_snd_una = tp->snd_una;
3846 bool is_sack_reneg = tp->is_sack_reneg;
3847 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3848 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3850 int prior_packets = tp->packets_out;
3851 u32 delivered = tp->delivered;
3852 u32 lost = tp->lost;
3853 int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */
3856 sack_state.first_sackt = 0;
3857 sack_state.rate = &rs;
3858 sack_state.sack_delivered = 0;
3860 /* We very likely will need to access rtx queue. */
3861 prefetch(sk->tcp_rtx_queue.rb_node);
3863 /* If the ack is older than previous acks
3864 * then we can probably ignore it.
3866 if (before(ack, prior_snd_una)) {
3869 /* do not accept ACK for bytes we never sent. */
3870 max_window = min_t(u64, tp->max_window, tp->bytes_acked);
3871 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3872 if (before(ack, prior_snd_una - max_window)) {
3873 if (!(flag & FLAG_NO_CHALLENGE_ACK))
3874 tcp_send_challenge_ack(sk);
3875 return -SKB_DROP_REASON_TCP_TOO_OLD_ACK;
3880 /* If the ack includes data we haven't sent yet, discard
3881 * this segment (RFC793 Section 3.9).
3883 if (after(ack, tp->snd_nxt))
3884 return -SKB_DROP_REASON_TCP_ACK_UNSENT_DATA;
3886 if (after(ack, prior_snd_una)) {
3887 flag |= FLAG_SND_UNA_ADVANCED;
3888 icsk->icsk_retransmits = 0;
3890 #if IS_ENABLED(CONFIG_TLS_DEVICE)
3891 if (static_branch_unlikely(&clean_acked_data_enabled.key))
3892 if (icsk->icsk_clean_acked)
3893 icsk->icsk_clean_acked(sk, ack);
3897 prior_fack = tcp_is_sack(tp) ? tcp_highest_sack_seq(tp) : tp->snd_una;
3898 rs.prior_in_flight = tcp_packets_in_flight(tp);
3900 /* ts_recent update must be made after we are sure that the packet
3903 if (flag & FLAG_UPDATE_TS_RECENT)
3904 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3906 if ((flag & (FLAG_SLOWPATH | FLAG_SND_UNA_ADVANCED)) ==
3907 FLAG_SND_UNA_ADVANCED) {
3908 /* Window is constant, pure forward advance.
3909 * No more checks are required.
3910 * Note, we use the fact that SND.UNA>=SND.WL2.
3912 tcp_update_wl(tp, ack_seq);
3913 tcp_snd_una_update(tp, ack);
3914 flag |= FLAG_WIN_UPDATE;
3916 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
3918 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPACKS);
3920 u32 ack_ev_flags = CA_ACK_SLOWPATH;
3922 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3925 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3927 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3929 if (TCP_SKB_CB(skb)->sacked)
3930 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3933 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) {
3935 ack_ev_flags |= CA_ACK_ECE;
3938 if (sack_state.sack_delivered)
3939 tcp_count_delivered(tp, sack_state.sack_delivered,
3942 if (flag & FLAG_WIN_UPDATE)
3943 ack_ev_flags |= CA_ACK_WIN_UPDATE;
3945 tcp_in_ack_event(sk, ack_ev_flags);
3948 /* This is a deviation from RFC3168 since it states that:
3949 * "When the TCP data sender is ready to set the CWR bit after reducing
3950 * the congestion window, it SHOULD set the CWR bit only on the first
3951 * new data packet that it transmits."
3952 * We accept CWR on pure ACKs to be more robust
3953 * with widely-deployed TCP implementations that do this.
3955 tcp_ecn_accept_cwr(sk, skb);
3957 /* We passed data and got it acked, remove any soft error
3958 * log. Something worked...
3960 WRITE_ONCE(sk->sk_err_soft, 0);
3961 icsk->icsk_probes_out = 0;
3962 tp->rcv_tstamp = tcp_jiffies32;
3966 /* See if we can take anything off of the retransmit queue. */
3967 flag |= tcp_clean_rtx_queue(sk, skb, prior_fack, prior_snd_una,
3968 &sack_state, flag & FLAG_ECE);
3970 tcp_rack_update_reo_wnd(sk, &rs);
3972 if (tp->tlp_high_seq)
3973 tcp_process_tlp_ack(sk, ack, flag);
3975 if (tcp_ack_is_dubious(sk, flag)) {
3976 if (!(flag & (FLAG_SND_UNA_ADVANCED |
3977 FLAG_NOT_DUP | FLAG_DSACKING_ACK))) {
3979 /* Consider if pure acks were aggregated in tcp_add_backlog() */
3980 if (!(flag & FLAG_DATA))
3981 num_dupack = max_t(u16, 1, skb_shinfo(skb)->gso_segs);
3983 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
3987 /* If needed, reset TLP/RTO timer when RACK doesn't set. */
3988 if (flag & FLAG_SET_XMIT_TIMER)
3989 tcp_set_xmit_timer(sk);
3991 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
3994 delivered = tcp_newly_delivered(sk, delivered, flag);
3995 lost = tp->lost - lost; /* freshly marked lost */
3996 rs.is_ack_delayed = !!(flag & FLAG_ACK_MAYBE_DELAYED);
3997 tcp_rate_gen(sk, delivered, lost, is_sack_reneg, sack_state.rate);
3998 tcp_cong_control(sk, ack, delivered, flag, sack_state.rate);
3999 tcp_xmit_recovery(sk, rexmit);
4003 /* If data was DSACKed, see if we can undo a cwnd reduction. */
4004 if (flag & FLAG_DSACKING_ACK) {
4005 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
4007 tcp_newly_delivered(sk, delivered, flag);
4009 /* If this ack opens up a zero window, clear backoff. It was
4010 * being used to time the probes, and is probably far higher than
4011 * it needs to be for normal retransmission.
4015 if (tp->tlp_high_seq)
4016 tcp_process_tlp_ack(sk, ack, flag);
4020 /* If data was SACKed, tag it and see if we should send more data.
4021 * If data was DSACKed, see if we can undo a cwnd reduction.
4023 if (TCP_SKB_CB(skb)->sacked) {
4024 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
4026 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
4028 tcp_newly_delivered(sk, delivered, flag);
4029 tcp_xmit_recovery(sk, rexmit);
4035 static void tcp_parse_fastopen_option(int len, const unsigned char *cookie,
4036 bool syn, struct tcp_fastopen_cookie *foc,
4039 /* Valid only in SYN or SYN-ACK with an even length. */
4040 if (!foc || !syn || len < 0 || (len & 1))
4043 if (len >= TCP_FASTOPEN_COOKIE_MIN &&
4044 len <= TCP_FASTOPEN_COOKIE_MAX)
4045 memcpy(foc->val, cookie, len);
4052 static bool smc_parse_options(const struct tcphdr *th,
4053 struct tcp_options_received *opt_rx,
4054 const unsigned char *ptr,
4057 #if IS_ENABLED(CONFIG_SMC)
4058 if (static_branch_unlikely(&tcp_have_smc)) {
4059 if (th->syn && !(opsize & 1) &&
4060 opsize >= TCPOLEN_EXP_SMC_BASE &&
4061 get_unaligned_be32(ptr) == TCPOPT_SMC_MAGIC) {
4070 /* Try to parse the MSS option from the TCP header. Return 0 on failure, clamped
4073 u16 tcp_parse_mss_option(const struct tcphdr *th, u16 user_mss)
4075 const unsigned char *ptr = (const unsigned char *)(th + 1);
4076 int length = (th->doff * 4) - sizeof(struct tcphdr);
4079 while (length > 0) {
4080 int opcode = *ptr++;
4086 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
4093 if (opsize < 2) /* "silly options" */
4095 if (opsize > length)
4096 return mss; /* fail on partial options */
4097 if (opcode == TCPOPT_MSS && opsize == TCPOLEN_MSS) {
4098 u16 in_mss = get_unaligned_be16(ptr);
4101 if (user_mss && user_mss < in_mss)
4112 EXPORT_SYMBOL_GPL(tcp_parse_mss_option);
4114 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
4115 * But, this can also be called on packets in the established flow when
4116 * the fast version below fails.
4118 void tcp_parse_options(const struct net *net,
4119 const struct sk_buff *skb,
4120 struct tcp_options_received *opt_rx, int estab,
4121 struct tcp_fastopen_cookie *foc)
4123 const unsigned char *ptr;
4124 const struct tcphdr *th = tcp_hdr(skb);
4125 int length = (th->doff * 4) - sizeof(struct tcphdr);
4127 ptr = (const unsigned char *)(th + 1);
4128 opt_rx->saw_tstamp = 0;
4129 opt_rx->saw_unknown = 0;
4131 while (length > 0) {
4132 int opcode = *ptr++;
4138 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
4145 if (opsize < 2) /* "silly options" */
4147 if (opsize > length)
4148 return; /* don't parse partial options */
4151 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
4152 u16 in_mss = get_unaligned_be16(ptr);
4154 if (opt_rx->user_mss &&
4155 opt_rx->user_mss < in_mss)
4156 in_mss = opt_rx->user_mss;
4157 opt_rx->mss_clamp = in_mss;
4162 if (opsize == TCPOLEN_WINDOW && th->syn &&
4163 !estab && READ_ONCE(net->ipv4.sysctl_tcp_window_scaling)) {
4164 __u8 snd_wscale = *(__u8 *)ptr;
4165 opt_rx->wscale_ok = 1;
4166 if (snd_wscale > TCP_MAX_WSCALE) {
4167 net_info_ratelimited("%s: Illegal window scaling value %d > %u received\n",
4171 snd_wscale = TCP_MAX_WSCALE;
4173 opt_rx->snd_wscale = snd_wscale;
4176 case TCPOPT_TIMESTAMP:
4177 if ((opsize == TCPOLEN_TIMESTAMP) &&
4178 ((estab && opt_rx->tstamp_ok) ||
4179 (!estab && READ_ONCE(net->ipv4.sysctl_tcp_timestamps)))) {
4180 opt_rx->saw_tstamp = 1;
4181 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
4182 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
4185 case TCPOPT_SACK_PERM:
4186 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
4187 !estab && READ_ONCE(net->ipv4.sysctl_tcp_sack)) {
4188 opt_rx->sack_ok = TCP_SACK_SEEN;
4189 tcp_sack_reset(opt_rx);
4194 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
4195 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
4197 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
4200 #ifdef CONFIG_TCP_MD5SIG
4202 /* The MD5 Hash has already been
4203 * checked (see tcp_v{4,6}_rcv()).
4207 case TCPOPT_FASTOPEN:
4208 tcp_parse_fastopen_option(
4209 opsize - TCPOLEN_FASTOPEN_BASE,
4210 ptr, th->syn, foc, false);
4214 /* Fast Open option shares code 254 using a
4215 * 16 bits magic number.
4217 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE &&
4218 get_unaligned_be16(ptr) ==
4219 TCPOPT_FASTOPEN_MAGIC) {
4220 tcp_parse_fastopen_option(opsize -
4221 TCPOLEN_EXP_FASTOPEN_BASE,
4222 ptr + 2, th->syn, foc, true);
4226 if (smc_parse_options(th, opt_rx, ptr, opsize))
4229 opt_rx->saw_unknown = 1;
4233 opt_rx->saw_unknown = 1;
4240 EXPORT_SYMBOL(tcp_parse_options);
4242 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
4244 const __be32 *ptr = (const __be32 *)(th + 1);
4246 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
4247 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
4248 tp->rx_opt.saw_tstamp = 1;
4250 tp->rx_opt.rcv_tsval = ntohl(*ptr);
4253 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
4255 tp->rx_opt.rcv_tsecr = 0;
4261 /* Fast parse options. This hopes to only see timestamps.
4262 * If it is wrong it falls back on tcp_parse_options().
4264 static bool tcp_fast_parse_options(const struct net *net,
4265 const struct sk_buff *skb,
4266 const struct tcphdr *th, struct tcp_sock *tp)
4268 /* In the spirit of fast parsing, compare doff directly to constant
4269 * values. Because equality is used, short doff can be ignored here.
4271 if (th->doff == (sizeof(*th) / 4)) {
4272 tp->rx_opt.saw_tstamp = 0;
4274 } else if (tp->rx_opt.tstamp_ok &&
4275 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
4276 if (tcp_parse_aligned_timestamp(tp, th))
4280 tcp_parse_options(net, skb, &tp->rx_opt, 1, NULL);
4281 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
4282 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
4287 #if defined(CONFIG_TCP_MD5SIG) || defined(CONFIG_TCP_AO)
4289 * Parse Signature options
4291 int tcp_do_parse_auth_options(const struct tcphdr *th,
4292 const u8 **md5_hash, const u8 **ao_hash)
4294 int length = (th->doff << 2) - sizeof(*th);
4295 const u8 *ptr = (const u8 *)(th + 1);
4296 unsigned int minlen = TCPOLEN_MD5SIG;
4298 if (IS_ENABLED(CONFIG_TCP_AO))
4299 minlen = sizeof(struct tcp_ao_hdr) + 1;
4304 /* If not enough data remaining, we can short cut */
4305 while (length >= minlen) {
4306 int opcode = *ptr++;
4317 if (opsize < 2 || opsize > length)
4319 if (opcode == TCPOPT_MD5SIG) {
4320 if (opsize != TCPOLEN_MD5SIG)
4322 if (unlikely(*md5_hash || *ao_hash))
4325 } else if (opcode == TCPOPT_AO) {
4326 if (opsize <= sizeof(struct tcp_ao_hdr))
4328 if (unlikely(*md5_hash || *ao_hash))
4338 EXPORT_SYMBOL(tcp_do_parse_auth_options);
4341 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
4343 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
4344 * it can pass through stack. So, the following predicate verifies that
4345 * this segment is not used for anything but congestion avoidance or
4346 * fast retransmit. Moreover, we even are able to eliminate most of such
4347 * second order effects, if we apply some small "replay" window (~RTO)
4348 * to timestamp space.
4350 * All these measures still do not guarantee that we reject wrapped ACKs
4351 * on networks with high bandwidth, when sequence space is recycled fastly,
4352 * but it guarantees that such events will be very rare and do not affect
4353 * connection seriously. This doesn't look nice, but alas, PAWS is really
4356 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
4357 * states that events when retransmit arrives after original data are rare.
4358 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
4359 * the biggest problem on large power networks even with minor reordering.
4360 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
4361 * up to bandwidth of 18Gigabit/sec. 8) ]
4364 /* Estimates max number of increments of remote peer TSval in
4365 * a replay window (based on our current RTO estimation).
4367 static u32 tcp_tsval_replay(const struct sock *sk)
4369 /* If we use usec TS resolution,
4370 * then expect the remote peer to use the same resolution.
4372 if (tcp_sk(sk)->tcp_usec_ts)
4373 return inet_csk(sk)->icsk_rto * (USEC_PER_SEC / HZ);
4375 /* RFC 7323 recommends a TSval clock between 1ms and 1sec.
4376 * We know that some OS (including old linux) can use 1200 Hz.
4378 return inet_csk(sk)->icsk_rto * 1200 / HZ;
4381 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
4383 const struct tcp_sock *tp = tcp_sk(sk);
4384 const struct tcphdr *th = tcp_hdr(skb);
4385 u32 seq = TCP_SKB_CB(skb)->seq;
4386 u32 ack = TCP_SKB_CB(skb)->ack_seq;
4388 return /* 1. Pure ACK with correct sequence number. */
4389 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
4391 /* 2. ... and duplicate ACK. */
4392 ack == tp->snd_una &&
4394 /* 3. ... and does not update window. */
4395 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
4397 /* 4. ... and sits in replay window. */
4398 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <=
4399 tcp_tsval_replay(sk);
4402 static inline bool tcp_paws_discard(const struct sock *sk,
4403 const struct sk_buff *skb)
4405 const struct tcp_sock *tp = tcp_sk(sk);
4407 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
4408 !tcp_disordered_ack(sk, skb);
4411 /* Check segment sequence number for validity.
4413 * Segment controls are considered valid, if the segment
4414 * fits to the window after truncation to the window. Acceptability
4415 * of data (and SYN, FIN, of course) is checked separately.
4416 * See tcp_data_queue(), for example.
4418 * Also, controls (RST is main one) are accepted using RCV.WUP instead
4419 * of RCV.NXT. Peer still did not advance his SND.UNA when we
4420 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4421 * (borrowed from freebsd)
4424 static enum skb_drop_reason tcp_sequence(const struct tcp_sock *tp,
4425 u32 seq, u32 end_seq)
4427 if (before(end_seq, tp->rcv_wup))
4428 return SKB_DROP_REASON_TCP_OLD_SEQUENCE;
4430 if (after(seq, tp->rcv_nxt + tcp_receive_window(tp)))
4431 return SKB_DROP_REASON_TCP_INVALID_SEQUENCE;
4433 return SKB_NOT_DROPPED_YET;
4436 /* When we get a reset we do this. */
4437 void tcp_reset(struct sock *sk, struct sk_buff *skb)
4439 trace_tcp_receive_reset(sk);
4441 /* mptcp can't tell us to ignore reset pkts,
4442 * so just ignore the return value of mptcp_incoming_options().
4444 if (sk_is_mptcp(sk))
4445 mptcp_incoming_options(sk, skb);
4447 /* We want the right error as BSD sees it (and indeed as we do). */
4448 switch (sk->sk_state) {
4450 WRITE_ONCE(sk->sk_err, ECONNREFUSED);
4452 case TCP_CLOSE_WAIT:
4453 WRITE_ONCE(sk->sk_err, EPIPE);
4458 WRITE_ONCE(sk->sk_err, ECONNRESET);
4460 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4463 tcp_write_queue_purge(sk);
4466 if (!sock_flag(sk, SOCK_DEAD))
4467 sk_error_report(sk);
4471 * Process the FIN bit. This now behaves as it is supposed to work
4472 * and the FIN takes effect when it is validly part of sequence
4473 * space. Not before when we get holes.
4475 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4476 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4479 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4480 * close and we go into CLOSING (and later onto TIME-WAIT)
4482 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4484 void tcp_fin(struct sock *sk)
4486 struct tcp_sock *tp = tcp_sk(sk);
4488 inet_csk_schedule_ack(sk);
4490 WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | RCV_SHUTDOWN);
4491 sock_set_flag(sk, SOCK_DONE);
4493 switch (sk->sk_state) {
4495 case TCP_ESTABLISHED:
4496 /* Move to CLOSE_WAIT */
4497 tcp_set_state(sk, TCP_CLOSE_WAIT);
4498 inet_csk_enter_pingpong_mode(sk);
4501 case TCP_CLOSE_WAIT:
4503 /* Received a retransmission of the FIN, do
4508 /* RFC793: Remain in the LAST-ACK state. */
4512 /* This case occurs when a simultaneous close
4513 * happens, we must ack the received FIN and
4514 * enter the CLOSING state.
4517 tcp_set_state(sk, TCP_CLOSING);
4520 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4522 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4525 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4526 * cases we should never reach this piece of code.
4528 pr_err("%s: Impossible, sk->sk_state=%d\n",
4529 __func__, sk->sk_state);
4533 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4534 * Probably, we should reset in this case. For now drop them.
4536 skb_rbtree_purge(&tp->out_of_order_queue);
4537 if (tcp_is_sack(tp))
4538 tcp_sack_reset(&tp->rx_opt);
4540 if (!sock_flag(sk, SOCK_DEAD)) {
4541 sk->sk_state_change(sk);
4543 /* Do not send POLL_HUP for half duplex close. */
4544 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4545 sk->sk_state == TCP_CLOSE)
4546 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4548 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4552 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4555 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4556 if (before(seq, sp->start_seq))
4557 sp->start_seq = seq;
4558 if (after(end_seq, sp->end_seq))
4559 sp->end_seq = end_seq;
4565 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4567 struct tcp_sock *tp = tcp_sk(sk);
4569 if (tcp_is_sack(tp) && READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_dsack)) {
4572 if (before(seq, tp->rcv_nxt))
4573 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4575 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4577 NET_INC_STATS(sock_net(sk), mib_idx);
4579 tp->rx_opt.dsack = 1;
4580 tp->duplicate_sack[0].start_seq = seq;
4581 tp->duplicate_sack[0].end_seq = end_seq;
4585 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4587 struct tcp_sock *tp = tcp_sk(sk);
4589 if (!tp->rx_opt.dsack)
4590 tcp_dsack_set(sk, seq, end_seq);
4592 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4595 static void tcp_rcv_spurious_retrans(struct sock *sk, const struct sk_buff *skb)
4597 /* When the ACK path fails or drops most ACKs, the sender would
4598 * timeout and spuriously retransmit the same segment repeatedly.
4599 * If it seems our ACKs are not reaching the other side,
4600 * based on receiving a duplicate data segment with new flowlabel
4601 * (suggesting the sender suffered an RTO), and we are not already
4602 * repathing due to our own RTO, then rehash the socket to repath our
4605 #if IS_ENABLED(CONFIG_IPV6)
4606 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Loss &&
4607 skb->protocol == htons(ETH_P_IPV6) &&
4608 (tcp_sk(sk)->inet_conn.icsk_ack.lrcv_flowlabel !=
4609 ntohl(ip6_flowlabel(ipv6_hdr(skb)))) &&
4610 sk_rethink_txhash(sk))
4611 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDUPLICATEDATAREHASH);
4613 /* Save last flowlabel after a spurious retrans. */
4614 tcp_save_lrcv_flowlabel(sk, skb);
4618 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4620 struct tcp_sock *tp = tcp_sk(sk);
4622 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4623 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4624 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4625 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
4627 if (tcp_is_sack(tp) && READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_dsack)) {
4628 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4630 tcp_rcv_spurious_retrans(sk, skb);
4631 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4632 end_seq = tp->rcv_nxt;
4633 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4640 /* These routines update the SACK block as out-of-order packets arrive or
4641 * in-order packets close up the sequence space.
4643 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4646 struct tcp_sack_block *sp = &tp->selective_acks[0];
4647 struct tcp_sack_block *swalk = sp + 1;
4649 /* See if the recent change to the first SACK eats into
4650 * or hits the sequence space of other SACK blocks, if so coalesce.
4652 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4653 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4656 /* Zap SWALK, by moving every further SACK up by one slot.
4657 * Decrease num_sacks.
4659 tp->rx_opt.num_sacks--;
4660 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4669 void tcp_sack_compress_send_ack(struct sock *sk)
4671 struct tcp_sock *tp = tcp_sk(sk);
4673 if (!tp->compressed_ack)
4676 if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1)
4679 /* Since we have to send one ack finally,
4680 * substract one from tp->compressed_ack to keep
4681 * LINUX_MIB_TCPACKCOMPRESSED accurate.
4683 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED,
4684 tp->compressed_ack - 1);
4686 tp->compressed_ack = 0;
4690 /* Reasonable amount of sack blocks included in TCP SACK option
4691 * The max is 4, but this becomes 3 if TCP timestamps are there.
4692 * Given that SACK packets might be lost, be conservative and use 2.
4694 #define TCP_SACK_BLOCKS_EXPECTED 2
4696 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4698 struct tcp_sock *tp = tcp_sk(sk);
4699 struct tcp_sack_block *sp = &tp->selective_acks[0];
4700 int cur_sacks = tp->rx_opt.num_sacks;
4706 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4707 if (tcp_sack_extend(sp, seq, end_seq)) {
4708 if (this_sack >= TCP_SACK_BLOCKS_EXPECTED)
4709 tcp_sack_compress_send_ack(sk);
4710 /* Rotate this_sack to the first one. */
4711 for (; this_sack > 0; this_sack--, sp--)
4712 swap(*sp, *(sp - 1));
4714 tcp_sack_maybe_coalesce(tp);
4719 if (this_sack >= TCP_SACK_BLOCKS_EXPECTED)
4720 tcp_sack_compress_send_ack(sk);
4722 /* Could not find an adjacent existing SACK, build a new one,
4723 * put it at the front, and shift everyone else down. We
4724 * always know there is at least one SACK present already here.
4726 * If the sack array is full, forget about the last one.
4728 if (this_sack >= TCP_NUM_SACKS) {
4730 tp->rx_opt.num_sacks--;
4733 for (; this_sack > 0; this_sack--, sp--)
4737 /* Build the new head SACK, and we're done. */
4738 sp->start_seq = seq;
4739 sp->end_seq = end_seq;
4740 tp->rx_opt.num_sacks++;
4743 /* RCV.NXT advances, some SACKs should be eaten. */
4745 static void tcp_sack_remove(struct tcp_sock *tp)
4747 struct tcp_sack_block *sp = &tp->selective_acks[0];
4748 int num_sacks = tp->rx_opt.num_sacks;
4751 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4752 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4753 tp->rx_opt.num_sacks = 0;
4757 for (this_sack = 0; this_sack < num_sacks;) {
4758 /* Check if the start of the sack is covered by RCV.NXT. */
4759 if (!before(tp->rcv_nxt, sp->start_seq)) {
4762 /* RCV.NXT must cover all the block! */
4763 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4765 /* Zap this SACK, by moving forward any other SACKS. */
4766 for (i = this_sack+1; i < num_sacks; i++)
4767 tp->selective_acks[i-1] = tp->selective_acks[i];
4774 tp->rx_opt.num_sacks = num_sacks;
4778 * tcp_try_coalesce - try to merge skb to prior one
4781 * @from: buffer to add in queue
4782 * @fragstolen: pointer to boolean
4784 * Before queueing skb @from after @to, try to merge them
4785 * to reduce overall memory use and queue lengths, if cost is small.
4786 * Packets in ofo or receive queues can stay a long time.
4787 * Better try to coalesce them right now to avoid future collapses.
4788 * Returns true if caller should free @from instead of queueing it
4790 static bool tcp_try_coalesce(struct sock *sk,
4792 struct sk_buff *from,
4797 *fragstolen = false;
4799 /* Its possible this segment overlaps with prior segment in queue */
4800 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4803 if (!mptcp_skb_can_collapse(to, from))
4806 #ifdef CONFIG_TLS_DEVICE
4807 if (from->decrypted != to->decrypted)
4811 if (!skb_try_coalesce(to, from, fragstolen, &delta))
4814 atomic_add(delta, &sk->sk_rmem_alloc);
4815 sk_mem_charge(sk, delta);
4816 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4817 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4818 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4819 TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
4821 if (TCP_SKB_CB(from)->has_rxtstamp) {
4822 TCP_SKB_CB(to)->has_rxtstamp = true;
4823 to->tstamp = from->tstamp;
4824 skb_hwtstamps(to)->hwtstamp = skb_hwtstamps(from)->hwtstamp;
4830 static bool tcp_ooo_try_coalesce(struct sock *sk,
4832 struct sk_buff *from,
4835 bool res = tcp_try_coalesce(sk, to, from, fragstolen);
4837 /* In case tcp_drop_reason() is called later, update to->gso_segs */
4839 u32 gso_segs = max_t(u16, 1, skb_shinfo(to)->gso_segs) +
4840 max_t(u16, 1, skb_shinfo(from)->gso_segs);
4842 skb_shinfo(to)->gso_segs = min_t(u32, gso_segs, 0xFFFF);
4847 static void tcp_drop_reason(struct sock *sk, struct sk_buff *skb,
4848 enum skb_drop_reason reason)
4850 sk_drops_add(sk, skb);
4851 kfree_skb_reason(skb, reason);
4854 /* This one checks to see if we can put data from the
4855 * out_of_order queue into the receive_queue.
4857 static void tcp_ofo_queue(struct sock *sk)
4859 struct tcp_sock *tp = tcp_sk(sk);
4860 __u32 dsack_high = tp->rcv_nxt;
4861 bool fin, fragstolen, eaten;
4862 struct sk_buff *skb, *tail;
4865 p = rb_first(&tp->out_of_order_queue);
4868 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4871 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4872 __u32 dsack = dsack_high;
4873 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4874 dsack_high = TCP_SKB_CB(skb)->end_seq;
4875 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4878 rb_erase(&skb->rbnode, &tp->out_of_order_queue);
4880 if (unlikely(!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))) {
4881 tcp_drop_reason(sk, skb, SKB_DROP_REASON_TCP_OFO_DROP);
4885 tail = skb_peek_tail(&sk->sk_receive_queue);
4886 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen);
4887 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4888 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
4890 __skb_queue_tail(&sk->sk_receive_queue, skb);
4892 kfree_skb_partial(skb, fragstolen);
4894 if (unlikely(fin)) {
4896 /* tcp_fin() purges tp->out_of_order_queue,
4897 * so we must end this loop right now.
4904 static bool tcp_prune_ofo_queue(struct sock *sk, const struct sk_buff *in_skb);
4905 static int tcp_prune_queue(struct sock *sk, const struct sk_buff *in_skb);
4907 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4910 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4911 !sk_rmem_schedule(sk, skb, size)) {
4913 if (tcp_prune_queue(sk, skb) < 0)
4916 while (!sk_rmem_schedule(sk, skb, size)) {
4917 if (!tcp_prune_ofo_queue(sk, skb))
4924 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4926 struct tcp_sock *tp = tcp_sk(sk);
4927 struct rb_node **p, *parent;
4928 struct sk_buff *skb1;
4932 tcp_save_lrcv_flowlabel(sk, skb);
4933 tcp_ecn_check_ce(sk, skb);
4935 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4936 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFODROP);
4937 sk->sk_data_ready(sk);
4938 tcp_drop_reason(sk, skb, SKB_DROP_REASON_PROTO_MEM);
4942 /* Disable header prediction. */
4944 inet_csk_schedule_ack(sk);
4946 tp->rcv_ooopack += max_t(u16, 1, skb_shinfo(skb)->gso_segs);
4947 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4948 seq = TCP_SKB_CB(skb)->seq;
4949 end_seq = TCP_SKB_CB(skb)->end_seq;
4951 p = &tp->out_of_order_queue.rb_node;
4952 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4953 /* Initial out of order segment, build 1 SACK. */
4954 if (tcp_is_sack(tp)) {
4955 tp->rx_opt.num_sacks = 1;
4956 tp->selective_acks[0].start_seq = seq;
4957 tp->selective_acks[0].end_seq = end_seq;
4959 rb_link_node(&skb->rbnode, NULL, p);
4960 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4961 tp->ooo_last_skb = skb;
4965 /* In the typical case, we are adding an skb to the end of the list.
4966 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
4968 if (tcp_ooo_try_coalesce(sk, tp->ooo_last_skb,
4969 skb, &fragstolen)) {
4971 /* For non sack flows, do not grow window to force DUPACK
4972 * and trigger fast retransmit.
4974 if (tcp_is_sack(tp))
4975 tcp_grow_window(sk, skb, true);
4976 kfree_skb_partial(skb, fragstolen);
4980 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
4981 if (!before(seq, TCP_SKB_CB(tp->ooo_last_skb)->end_seq)) {
4982 parent = &tp->ooo_last_skb->rbnode;
4983 p = &parent->rb_right;
4987 /* Find place to insert this segment. Handle overlaps on the way. */
4991 skb1 = rb_to_skb(parent);
4992 if (before(seq, TCP_SKB_CB(skb1)->seq)) {
4993 p = &parent->rb_left;
4996 if (before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4997 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4998 /* All the bits are present. Drop. */
4999 NET_INC_STATS(sock_net(sk),
5000 LINUX_MIB_TCPOFOMERGE);
5001 tcp_drop_reason(sk, skb,
5002 SKB_DROP_REASON_TCP_OFOMERGE);
5004 tcp_dsack_set(sk, seq, end_seq);
5007 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
5008 /* Partial overlap. */
5009 tcp_dsack_set(sk, seq, TCP_SKB_CB(skb1)->end_seq);
5011 /* skb's seq == skb1's seq and skb covers skb1.
5012 * Replace skb1 with skb.
5014 rb_replace_node(&skb1->rbnode, &skb->rbnode,
5015 &tp->out_of_order_queue);
5016 tcp_dsack_extend(sk,
5017 TCP_SKB_CB(skb1)->seq,
5018 TCP_SKB_CB(skb1)->end_seq);
5019 NET_INC_STATS(sock_net(sk),
5020 LINUX_MIB_TCPOFOMERGE);
5021 tcp_drop_reason(sk, skb1,
5022 SKB_DROP_REASON_TCP_OFOMERGE);
5025 } else if (tcp_ooo_try_coalesce(sk, skb1,
5026 skb, &fragstolen)) {
5029 p = &parent->rb_right;
5032 /* Insert segment into RB tree. */
5033 rb_link_node(&skb->rbnode, parent, p);
5034 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
5037 /* Remove other segments covered by skb. */
5038 while ((skb1 = skb_rb_next(skb)) != NULL) {
5039 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
5041 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
5042 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
5046 rb_erase(&skb1->rbnode, &tp->out_of_order_queue);
5047 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
5048 TCP_SKB_CB(skb1)->end_seq);
5049 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
5050 tcp_drop_reason(sk, skb1, SKB_DROP_REASON_TCP_OFOMERGE);
5052 /* If there is no skb after us, we are the last_skb ! */
5054 tp->ooo_last_skb = skb;
5057 if (tcp_is_sack(tp))
5058 tcp_sack_new_ofo_skb(sk, seq, end_seq);
5061 /* For non sack flows, do not grow window to force DUPACK
5062 * and trigger fast retransmit.
5064 if (tcp_is_sack(tp))
5065 tcp_grow_window(sk, skb, false);
5067 skb_set_owner_r(skb, sk);
5071 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb,
5075 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
5078 tcp_try_coalesce(sk, tail,
5079 skb, fragstolen)) ? 1 : 0;
5080 tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq);
5082 __skb_queue_tail(&sk->sk_receive_queue, skb);
5083 skb_set_owner_r(skb, sk);
5088 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
5090 struct sk_buff *skb;
5098 if (size > PAGE_SIZE) {
5099 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS);
5101 data_len = npages << PAGE_SHIFT;
5102 size = data_len + (size & ~PAGE_MASK);
5104 skb = alloc_skb_with_frags(size - data_len, data_len,
5105 PAGE_ALLOC_COSTLY_ORDER,
5106 &err, sk->sk_allocation);
5110 skb_put(skb, size - data_len);
5111 skb->data_len = data_len;
5114 if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) {
5115 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP);
5119 err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size);
5123 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
5124 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
5125 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
5127 if (tcp_queue_rcv(sk, skb, &fragstolen)) {
5128 WARN_ON_ONCE(fragstolen); /* should not happen */
5140 void tcp_data_ready(struct sock *sk)
5142 if (tcp_epollin_ready(sk, sk->sk_rcvlowat) || sock_flag(sk, SOCK_DONE))
5143 sk->sk_data_ready(sk);
5146 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
5148 struct tcp_sock *tp = tcp_sk(sk);
5149 enum skb_drop_reason reason;
5153 /* If a subflow has been reset, the packet should not continue
5154 * to be processed, drop the packet.
5156 if (sk_is_mptcp(sk) && !mptcp_incoming_options(sk, skb)) {
5161 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
5166 __skb_pull(skb, tcp_hdr(skb)->doff * 4);
5168 reason = SKB_DROP_REASON_NOT_SPECIFIED;
5169 tp->rx_opt.dsack = 0;
5171 /* Queue data for delivery to the user.
5172 * Packets in sequence go to the receive queue.
5173 * Out of sequence packets to the out_of_order_queue.
5175 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
5176 if (tcp_receive_window(tp) == 0) {
5177 reason = SKB_DROP_REASON_TCP_ZEROWINDOW;
5178 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP);
5182 /* Ok. In sequence. In window. */
5184 if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) {
5185 /* TODO: maybe ratelimit these WIN 0 ACK ? */
5186 inet_csk(sk)->icsk_ack.pending |=
5187 (ICSK_ACK_NOMEM | ICSK_ACK_NOW);
5188 inet_csk_schedule_ack(sk);
5189 sk->sk_data_ready(sk);
5191 if (skb_queue_len(&sk->sk_receive_queue)) {
5192 reason = SKB_DROP_REASON_PROTO_MEM;
5193 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP);
5196 sk_forced_mem_schedule(sk, skb->truesize);
5199 eaten = tcp_queue_rcv(sk, skb, &fragstolen);
5201 tcp_event_data_recv(sk, skb);
5202 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
5205 if (!RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
5208 /* RFC5681. 4.2. SHOULD send immediate ACK, when
5209 * gap in queue is filled.
5211 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
5212 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW;
5215 if (tp->rx_opt.num_sacks)
5216 tcp_sack_remove(tp);
5218 tcp_fast_path_check(sk);
5221 kfree_skb_partial(skb, fragstolen);
5222 if (!sock_flag(sk, SOCK_DEAD))
5227 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
5228 tcp_rcv_spurious_retrans(sk, skb);
5229 /* A retransmit, 2nd most common case. Force an immediate ack. */
5230 reason = SKB_DROP_REASON_TCP_OLD_DATA;
5231 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
5232 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
5235 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
5236 inet_csk_schedule_ack(sk);
5238 tcp_drop_reason(sk, skb, reason);
5242 /* Out of window. F.e. zero window probe. */
5243 if (!before(TCP_SKB_CB(skb)->seq,
5244 tp->rcv_nxt + tcp_receive_window(tp))) {
5245 reason = SKB_DROP_REASON_TCP_OVERWINDOW;
5249 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
5250 /* Partial packet, seq < rcv_next < end_seq */
5251 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
5253 /* If window is closed, drop tail of packet. But after
5254 * remembering D-SACK for its head made in previous line.
5256 if (!tcp_receive_window(tp)) {
5257 reason = SKB_DROP_REASON_TCP_ZEROWINDOW;
5258 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP);
5264 tcp_data_queue_ofo(sk, skb);
5267 static struct sk_buff *tcp_skb_next(struct sk_buff *skb, struct sk_buff_head *list)
5270 return !skb_queue_is_last(list, skb) ? skb->next : NULL;
5272 return skb_rb_next(skb);
5275 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
5276 struct sk_buff_head *list,
5277 struct rb_root *root)
5279 struct sk_buff *next = tcp_skb_next(skb, list);
5282 __skb_unlink(skb, list);
5284 rb_erase(&skb->rbnode, root);
5287 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
5292 /* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */
5293 void tcp_rbtree_insert(struct rb_root *root, struct sk_buff *skb)
5295 struct rb_node **p = &root->rb_node;
5296 struct rb_node *parent = NULL;
5297 struct sk_buff *skb1;
5301 skb1 = rb_to_skb(parent);
5302 if (before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb1)->seq))
5303 p = &parent->rb_left;
5305 p = &parent->rb_right;
5307 rb_link_node(&skb->rbnode, parent, p);
5308 rb_insert_color(&skb->rbnode, root);
5311 /* Collapse contiguous sequence of skbs head..tail with
5312 * sequence numbers start..end.
5314 * If tail is NULL, this means until the end of the queue.
5316 * Segments with FIN/SYN are not collapsed (only because this
5320 tcp_collapse(struct sock *sk, struct sk_buff_head *list, struct rb_root *root,
5321 struct sk_buff *head, struct sk_buff *tail, u32 start, u32 end)
5323 struct sk_buff *skb = head, *n;
5324 struct sk_buff_head tmp;
5327 /* First, check that queue is collapsible and find
5328 * the point where collapsing can be useful.
5331 for (end_of_skbs = true; skb != NULL && skb != tail; skb = n) {
5332 n = tcp_skb_next(skb, list);
5334 /* No new bits? It is possible on ofo queue. */
5335 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
5336 skb = tcp_collapse_one(sk, skb, list, root);
5342 /* The first skb to collapse is:
5344 * - bloated or contains data before "start" or
5345 * overlaps to the next one and mptcp allow collapsing.
5347 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
5348 (tcp_win_from_space(sk, skb->truesize) > skb->len ||
5349 before(TCP_SKB_CB(skb)->seq, start))) {
5350 end_of_skbs = false;
5354 if (n && n != tail && mptcp_skb_can_collapse(skb, n) &&
5355 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(n)->seq) {
5356 end_of_skbs = false;
5360 /* Decided to skip this, advance start seq. */
5361 start = TCP_SKB_CB(skb)->end_seq;
5364 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
5367 __skb_queue_head_init(&tmp);
5369 while (before(start, end)) {
5370 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
5371 struct sk_buff *nskb;
5373 nskb = alloc_skb(copy, GFP_ATOMIC);
5377 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
5378 #ifdef CONFIG_TLS_DEVICE
5379 nskb->decrypted = skb->decrypted;
5381 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
5383 __skb_queue_before(list, skb, nskb);
5385 __skb_queue_tail(&tmp, nskb); /* defer rbtree insertion */
5386 skb_set_owner_r(nskb, sk);
5387 mptcp_skb_ext_move(nskb, skb);
5389 /* Copy data, releasing collapsed skbs. */
5391 int offset = start - TCP_SKB_CB(skb)->seq;
5392 int size = TCP_SKB_CB(skb)->end_seq - start;
5396 size = min(copy, size);
5397 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
5399 TCP_SKB_CB(nskb)->end_seq += size;
5403 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
5404 skb = tcp_collapse_one(sk, skb, list, root);
5407 !mptcp_skb_can_collapse(nskb, skb) ||
5408 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
5410 #ifdef CONFIG_TLS_DEVICE
5411 if (skb->decrypted != nskb->decrypted)
5418 skb_queue_walk_safe(&tmp, skb, n)
5419 tcp_rbtree_insert(root, skb);
5422 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
5423 * and tcp_collapse() them until all the queue is collapsed.
5425 static void tcp_collapse_ofo_queue(struct sock *sk)
5427 struct tcp_sock *tp = tcp_sk(sk);
5428 u32 range_truesize, sum_tiny = 0;
5429 struct sk_buff *skb, *head;
5432 skb = skb_rb_first(&tp->out_of_order_queue);
5435 tp->ooo_last_skb = skb_rb_last(&tp->out_of_order_queue);
5438 start = TCP_SKB_CB(skb)->seq;
5439 end = TCP_SKB_CB(skb)->end_seq;
5440 range_truesize = skb->truesize;
5442 for (head = skb;;) {
5443 skb = skb_rb_next(skb);
5445 /* Range is terminated when we see a gap or when
5446 * we are at the queue end.
5449 after(TCP_SKB_CB(skb)->seq, end) ||
5450 before(TCP_SKB_CB(skb)->end_seq, start)) {
5451 /* Do not attempt collapsing tiny skbs */
5452 if (range_truesize != head->truesize ||
5453 end - start >= SKB_WITH_OVERHEAD(PAGE_SIZE)) {
5454 tcp_collapse(sk, NULL, &tp->out_of_order_queue,
5455 head, skb, start, end);
5457 sum_tiny += range_truesize;
5458 if (sum_tiny > sk->sk_rcvbuf >> 3)
5464 range_truesize += skb->truesize;
5465 if (unlikely(before(TCP_SKB_CB(skb)->seq, start)))
5466 start = TCP_SKB_CB(skb)->seq;
5467 if (after(TCP_SKB_CB(skb)->end_seq, end))
5468 end = TCP_SKB_CB(skb)->end_seq;
5473 * Clean the out-of-order queue to make room.
5474 * We drop high sequences packets to :
5475 * 1) Let a chance for holes to be filled.
5476 * This means we do not drop packets from ooo queue if their sequence
5477 * is before incoming packet sequence.
5478 * 2) not add too big latencies if thousands of packets sit there.
5479 * (But if application shrinks SO_RCVBUF, we could still end up
5480 * freeing whole queue here)
5481 * 3) Drop at least 12.5 % of sk_rcvbuf to avoid malicious attacks.
5483 * Return true if queue has shrunk.
5485 static bool tcp_prune_ofo_queue(struct sock *sk, const struct sk_buff *in_skb)
5487 struct tcp_sock *tp = tcp_sk(sk);
5488 struct rb_node *node, *prev;
5489 bool pruned = false;
5492 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
5495 goal = sk->sk_rcvbuf >> 3;
5496 node = &tp->ooo_last_skb->rbnode;
5499 struct sk_buff *skb = rb_to_skb(node);
5501 /* If incoming skb would land last in ofo queue, stop pruning. */
5502 if (after(TCP_SKB_CB(in_skb)->seq, TCP_SKB_CB(skb)->seq))
5505 prev = rb_prev(node);
5506 rb_erase(node, &tp->out_of_order_queue);
5507 goal -= skb->truesize;
5508 tcp_drop_reason(sk, skb, SKB_DROP_REASON_TCP_OFO_QUEUE_PRUNE);
5509 tp->ooo_last_skb = rb_to_skb(prev);
5510 if (!prev || goal <= 0) {
5511 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf &&
5512 !tcp_under_memory_pressure(sk))
5514 goal = sk->sk_rcvbuf >> 3;
5520 NET_INC_STATS(sock_net(sk), LINUX_MIB_OFOPRUNED);
5521 /* Reset SACK state. A conforming SACK implementation will
5522 * do the same at a timeout based retransmit. When a connection
5523 * is in a sad state like this, we care only about integrity
5524 * of the connection not performance.
5526 if (tp->rx_opt.sack_ok)
5527 tcp_sack_reset(&tp->rx_opt);
5532 /* Reduce allocated memory if we can, trying to get
5533 * the socket within its memory limits again.
5535 * Return less than zero if we should start dropping frames
5536 * until the socket owning process reads some of the data
5537 * to stabilize the situation.
5539 static int tcp_prune_queue(struct sock *sk, const struct sk_buff *in_skb)
5541 struct tcp_sock *tp = tcp_sk(sk);
5543 NET_INC_STATS(sock_net(sk), LINUX_MIB_PRUNECALLED);
5545 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
5546 tcp_clamp_window(sk);
5547 else if (tcp_under_memory_pressure(sk))
5548 tcp_adjust_rcv_ssthresh(sk);
5550 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5553 tcp_collapse_ofo_queue(sk);
5554 if (!skb_queue_empty(&sk->sk_receive_queue))
5555 tcp_collapse(sk, &sk->sk_receive_queue, NULL,
5556 skb_peek(&sk->sk_receive_queue),
5558 tp->copied_seq, tp->rcv_nxt);
5560 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5563 /* Collapsing did not help, destructive actions follow.
5564 * This must not ever occur. */
5566 tcp_prune_ofo_queue(sk, in_skb);
5568 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5571 /* If we are really being abused, tell the caller to silently
5572 * drop receive data on the floor. It will get retransmitted
5573 * and hopefully then we'll have sufficient space.
5575 NET_INC_STATS(sock_net(sk), LINUX_MIB_RCVPRUNED);
5577 /* Massive buffer overcommit. */
5582 static bool tcp_should_expand_sndbuf(struct sock *sk)
5584 const struct tcp_sock *tp = tcp_sk(sk);
5586 /* If the user specified a specific send buffer setting, do
5589 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
5592 /* If we are under global TCP memory pressure, do not expand. */
5593 if (tcp_under_memory_pressure(sk)) {
5594 int unused_mem = sk_unused_reserved_mem(sk);
5596 /* Adjust sndbuf according to reserved mem. But make sure
5597 * it never goes below SOCK_MIN_SNDBUF.
5598 * See sk_stream_moderate_sndbuf() for more details.
5600 if (unused_mem > SOCK_MIN_SNDBUF)
5601 WRITE_ONCE(sk->sk_sndbuf, unused_mem);
5606 /* If we are under soft global TCP memory pressure, do not expand. */
5607 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
5610 /* If we filled the congestion window, do not expand. */
5611 if (tcp_packets_in_flight(tp) >= tcp_snd_cwnd(tp))
5617 static void tcp_new_space(struct sock *sk)
5619 struct tcp_sock *tp = tcp_sk(sk);
5621 if (tcp_should_expand_sndbuf(sk)) {
5622 tcp_sndbuf_expand(sk);
5623 tp->snd_cwnd_stamp = tcp_jiffies32;
5626 INDIRECT_CALL_1(sk->sk_write_space, sk_stream_write_space, sk);
5629 /* Caller made space either from:
5630 * 1) Freeing skbs in rtx queues (after tp->snd_una has advanced)
5631 * 2) Sent skbs from output queue (and thus advancing tp->snd_nxt)
5633 * We might be able to generate EPOLLOUT to the application if:
5634 * 1) Space consumed in output/rtx queues is below sk->sk_sndbuf/2
5635 * 2) notsent amount (tp->write_seq - tp->snd_nxt) became
5636 * small enough that tcp_stream_memory_free() decides it
5637 * is time to generate EPOLLOUT.
5639 void tcp_check_space(struct sock *sk)
5641 /* pairs with tcp_poll() */
5643 if (sk->sk_socket &&
5644 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
5646 if (!test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
5647 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
5651 static inline void tcp_data_snd_check(struct sock *sk)
5653 tcp_push_pending_frames(sk);
5654 tcp_check_space(sk);
5658 * Check if sending an ack is needed.
5660 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
5662 struct tcp_sock *tp = tcp_sk(sk);
5663 unsigned long rtt, delay;
5665 /* More than one full frame received... */
5666 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
5667 /* ... and right edge of window advances far enough.
5668 * (tcp_recvmsg() will send ACK otherwise).
5669 * If application uses SO_RCVLOWAT, we want send ack now if
5670 * we have not received enough bytes to satisfy the condition.
5672 (tp->rcv_nxt - tp->copied_seq < sk->sk_rcvlowat ||
5673 __tcp_select_window(sk) >= tp->rcv_wnd)) ||
5674 /* We ACK each frame or... */
5675 tcp_in_quickack_mode(sk) ||
5676 /* Protocol state mandates a one-time immediate ACK */
5677 inet_csk(sk)->icsk_ack.pending & ICSK_ACK_NOW) {
5678 /* If we are running from __release_sock() in user context,
5679 * Defer the ack until tcp_release_cb().
5681 if (sock_owned_by_user_nocheck(sk) &&
5682 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_backlog_ack_defer)) {
5683 set_bit(TCP_ACK_DEFERRED, &sk->sk_tsq_flags);
5691 if (!ofo_possible || RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
5692 tcp_send_delayed_ack(sk);
5696 if (!tcp_is_sack(tp) ||
5697 tp->compressed_ack >= READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_nr))
5700 if (tp->compressed_ack_rcv_nxt != tp->rcv_nxt) {
5701 tp->compressed_ack_rcv_nxt = tp->rcv_nxt;
5702 tp->dup_ack_counter = 0;
5704 if (tp->dup_ack_counter < TCP_FASTRETRANS_THRESH) {
5705 tp->dup_ack_counter++;
5708 tp->compressed_ack++;
5709 if (hrtimer_is_queued(&tp->compressed_ack_timer))
5712 /* compress ack timer : 5 % of rtt, but no more than tcp_comp_sack_delay_ns */
5714 rtt = tp->rcv_rtt_est.rtt_us;
5715 if (tp->srtt_us && tp->srtt_us < rtt)
5718 delay = min_t(unsigned long,
5719 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_delay_ns),
5720 rtt * (NSEC_PER_USEC >> 3)/20);
5722 hrtimer_start_range_ns(&tp->compressed_ack_timer, ns_to_ktime(delay),
5723 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_slack_ns),
5724 HRTIMER_MODE_REL_PINNED_SOFT);
5727 static inline void tcp_ack_snd_check(struct sock *sk)
5729 if (!inet_csk_ack_scheduled(sk)) {
5730 /* We sent a data segment already. */
5733 __tcp_ack_snd_check(sk, 1);
5737 * This routine is only called when we have urgent data
5738 * signaled. Its the 'slow' part of tcp_urg. It could be
5739 * moved inline now as tcp_urg is only called from one
5740 * place. We handle URGent data wrong. We have to - as
5741 * BSD still doesn't use the correction from RFC961.
5742 * For 1003.1g we should support a new option TCP_STDURG to permit
5743 * either form (or just set the sysctl tcp_stdurg).
5746 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
5748 struct tcp_sock *tp = tcp_sk(sk);
5749 u32 ptr = ntohs(th->urg_ptr);
5751 if (ptr && !READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_stdurg))
5753 ptr += ntohl(th->seq);
5755 /* Ignore urgent data that we've already seen and read. */
5756 if (after(tp->copied_seq, ptr))
5759 /* Do not replay urg ptr.
5761 * NOTE: interesting situation not covered by specs.
5762 * Misbehaving sender may send urg ptr, pointing to segment,
5763 * which we already have in ofo queue. We are not able to fetch
5764 * such data and will stay in TCP_URG_NOTYET until will be eaten
5765 * by recvmsg(). Seems, we are not obliged to handle such wicked
5766 * situations. But it is worth to think about possibility of some
5767 * DoSes using some hypothetical application level deadlock.
5769 if (before(ptr, tp->rcv_nxt))
5772 /* Do we already have a newer (or duplicate) urgent pointer? */
5773 if (tp->urg_data && !after(ptr, tp->urg_seq))
5776 /* Tell the world about our new urgent pointer. */
5779 /* We may be adding urgent data when the last byte read was
5780 * urgent. To do this requires some care. We cannot just ignore
5781 * tp->copied_seq since we would read the last urgent byte again
5782 * as data, nor can we alter copied_seq until this data arrives
5783 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5785 * NOTE. Double Dutch. Rendering to plain English: author of comment
5786 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5787 * and expect that both A and B disappear from stream. This is _wrong_.
5788 * Though this happens in BSD with high probability, this is occasional.
5789 * Any application relying on this is buggy. Note also, that fix "works"
5790 * only in this artificial test. Insert some normal data between A and B and we will
5791 * decline of BSD again. Verdict: it is better to remove to trap
5794 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5795 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5796 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5798 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5799 __skb_unlink(skb, &sk->sk_receive_queue);
5804 WRITE_ONCE(tp->urg_data, TCP_URG_NOTYET);
5805 WRITE_ONCE(tp->urg_seq, ptr);
5807 /* Disable header prediction. */
5811 /* This is the 'fast' part of urgent handling. */
5812 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
5814 struct tcp_sock *tp = tcp_sk(sk);
5816 /* Check if we get a new urgent pointer - normally not. */
5817 if (unlikely(th->urg))
5818 tcp_check_urg(sk, th);
5820 /* Do we wait for any urgent data? - normally not... */
5821 if (unlikely(tp->urg_data == TCP_URG_NOTYET)) {
5822 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5825 /* Is the urgent pointer pointing into this packet? */
5826 if (ptr < skb->len) {
5828 if (skb_copy_bits(skb, ptr, &tmp, 1))
5830 WRITE_ONCE(tp->urg_data, TCP_URG_VALID | tmp);
5831 if (!sock_flag(sk, SOCK_DEAD))
5832 sk->sk_data_ready(sk);
5837 /* Accept RST for rcv_nxt - 1 after a FIN.
5838 * When tcp connections are abruptly terminated from Mac OSX (via ^C), a
5839 * FIN is sent followed by a RST packet. The RST is sent with the same
5840 * sequence number as the FIN, and thus according to RFC 5961 a challenge
5841 * ACK should be sent. However, Mac OSX rate limits replies to challenge
5842 * ACKs on the closed socket. In addition middleboxes can drop either the
5843 * challenge ACK or a subsequent RST.
5845 static bool tcp_reset_check(const struct sock *sk, const struct sk_buff *skb)
5847 const struct tcp_sock *tp = tcp_sk(sk);
5849 return unlikely(TCP_SKB_CB(skb)->seq == (tp->rcv_nxt - 1) &&
5850 (1 << sk->sk_state) & (TCPF_CLOSE_WAIT | TCPF_LAST_ACK |
5854 /* Does PAWS and seqno based validation of an incoming segment, flags will
5855 * play significant role here.
5857 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5858 const struct tcphdr *th, int syn_inerr)
5860 struct tcp_sock *tp = tcp_sk(sk);
5863 /* RFC1323: H1. Apply PAWS check first. */
5864 if (tcp_fast_parse_options(sock_net(sk), skb, th, tp) &&
5865 tp->rx_opt.saw_tstamp &&
5866 tcp_paws_discard(sk, skb)) {
5868 if (unlikely(th->syn))
5870 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5871 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5872 LINUX_MIB_TCPACKSKIPPEDPAWS,
5873 &tp->last_oow_ack_time))
5874 tcp_send_dupack(sk, skb);
5875 SKB_DR_SET(reason, TCP_RFC7323_PAWS);
5878 /* Reset is accepted even if it did not pass PAWS. */
5881 /* Step 1: check sequence number */
5882 reason = tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
5884 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5885 * (RST) segments are validated by checking their SEQ-fields."
5886 * And page 69: "If an incoming segment is not acceptable,
5887 * an acknowledgment should be sent in reply (unless the RST
5888 * bit is set, if so drop the segment and return)".
5893 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5894 LINUX_MIB_TCPACKSKIPPEDSEQ,
5895 &tp->last_oow_ack_time))
5896 tcp_send_dupack(sk, skb);
5897 } else if (tcp_reset_check(sk, skb)) {
5903 /* Step 2: check RST bit */
5905 /* RFC 5961 3.2 (extend to match against (RCV.NXT - 1) after a
5906 * FIN and SACK too if available):
5907 * If seq num matches RCV.NXT or (RCV.NXT - 1) after a FIN, or
5908 * the right-most SACK block,
5910 * RESET the connection
5912 * Send a challenge ACK
5914 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt ||
5915 tcp_reset_check(sk, skb))
5918 if (tcp_is_sack(tp) && tp->rx_opt.num_sacks > 0) {
5919 struct tcp_sack_block *sp = &tp->selective_acks[0];
5920 int max_sack = sp[0].end_seq;
5923 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;
5925 max_sack = after(sp[this_sack].end_seq,
5927 sp[this_sack].end_seq : max_sack;
5930 if (TCP_SKB_CB(skb)->seq == max_sack)
5934 /* Disable TFO if RST is out-of-order
5935 * and no data has been received
5936 * for current active TFO socket
5938 if (tp->syn_fastopen && !tp->data_segs_in &&
5939 sk->sk_state == TCP_ESTABLISHED)
5940 tcp_fastopen_active_disable(sk);
5941 tcp_send_challenge_ack(sk);
5942 SKB_DR_SET(reason, TCP_RESET);
5946 /* step 3: check security and precedence [ignored] */
5948 /* step 4: Check for a SYN
5949 * RFC 5961 4.2 : Send a challenge ack
5954 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5955 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5956 tcp_send_challenge_ack(sk);
5957 SKB_DR_SET(reason, TCP_INVALID_SYN);
5961 bpf_skops_parse_hdr(sk, skb);
5966 tcp_drop_reason(sk, skb, reason);
5976 * TCP receive function for the ESTABLISHED state.
5978 * It is split into a fast path and a slow path. The fast path is
5980 * - A zero window was announced from us - zero window probing
5981 * is only handled properly in the slow path.
5982 * - Out of order segments arrived.
5983 * - Urgent data is expected.
5984 * - There is no buffer space left
5985 * - Unexpected TCP flags/window values/header lengths are received
5986 * (detected by checking the TCP header against pred_flags)
5987 * - Data is sent in both directions. Fast path only supports pure senders
5988 * or pure receivers (this means either the sequence number or the ack
5989 * value must stay constant)
5990 * - Unexpected TCP option.
5992 * When these conditions are not satisfied it drops into a standard
5993 * receive procedure patterned after RFC793 to handle all cases.
5994 * The first three cases are guaranteed by proper pred_flags setting,
5995 * the rest is checked inline. Fast processing is turned on in
5996 * tcp_data_queue when everything is OK.
5998 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb)
6000 enum skb_drop_reason reason = SKB_DROP_REASON_NOT_SPECIFIED;
6001 const struct tcphdr *th = (const struct tcphdr *)skb->data;
6002 struct tcp_sock *tp = tcp_sk(sk);
6003 unsigned int len = skb->len;
6005 /* TCP congestion window tracking */
6006 trace_tcp_probe(sk, skb);
6008 tcp_mstamp_refresh(tp);
6009 if (unlikely(!rcu_access_pointer(sk->sk_rx_dst)))
6010 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
6012 * Header prediction.
6013 * The code loosely follows the one in the famous
6014 * "30 instruction TCP receive" Van Jacobson mail.
6016 * Van's trick is to deposit buffers into socket queue
6017 * on a device interrupt, to call tcp_recv function
6018 * on the receive process context and checksum and copy
6019 * the buffer to user space. smart...
6021 * Our current scheme is not silly either but we take the
6022 * extra cost of the net_bh soft interrupt processing...
6023 * We do checksum and copy also but from device to kernel.
6026 tp->rx_opt.saw_tstamp = 0;
6028 /* pred_flags is 0xS?10 << 16 + snd_wnd
6029 * if header_prediction is to be made
6030 * 'S' will always be tp->tcp_header_len >> 2
6031 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
6032 * turn it off (when there are holes in the receive
6033 * space for instance)
6034 * PSH flag is ignored.
6037 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
6038 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
6039 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
6040 int tcp_header_len = tp->tcp_header_len;
6042 /* Timestamp header prediction: tcp_header_len
6043 * is automatically equal to th->doff*4 due to pred_flags
6047 /* Check timestamp */
6048 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
6049 /* No? Slow path! */
6050 if (!tcp_parse_aligned_timestamp(tp, th))
6053 /* If PAWS failed, check it more carefully in slow path */
6054 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
6057 /* DO NOT update ts_recent here, if checksum fails
6058 * and timestamp was corrupted part, it will result
6059 * in a hung connection since we will drop all
6060 * future packets due to the PAWS test.
6064 if (len <= tcp_header_len) {
6065 /* Bulk data transfer: sender */
6066 if (len == tcp_header_len) {
6067 /* Predicted packet is in window by definition.
6068 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
6069 * Hence, check seq<=rcv_wup reduces to:
6071 if (tcp_header_len ==
6072 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
6073 tp->rcv_nxt == tp->rcv_wup)
6074 tcp_store_ts_recent(tp);
6076 /* We know that such packets are checksummed
6079 tcp_ack(sk, skb, 0);
6081 tcp_data_snd_check(sk);
6082 /* When receiving pure ack in fast path, update
6083 * last ts ecr directly instead of calling
6084 * tcp_rcv_rtt_measure_ts()
6086 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr;
6088 } else { /* Header too small */
6089 reason = SKB_DROP_REASON_PKT_TOO_SMALL;
6090 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
6095 bool fragstolen = false;
6097 if (tcp_checksum_complete(skb))
6100 if ((int)skb->truesize > sk->sk_forward_alloc)
6103 /* Predicted packet is in window by definition.
6104 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
6105 * Hence, check seq<=rcv_wup reduces to:
6107 if (tcp_header_len ==
6108 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
6109 tp->rcv_nxt == tp->rcv_wup)
6110 tcp_store_ts_recent(tp);
6112 tcp_rcv_rtt_measure_ts(sk, skb);
6114 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPHITS);
6116 /* Bulk data transfer: receiver */
6118 __skb_pull(skb, tcp_header_len);
6119 eaten = tcp_queue_rcv(sk, skb, &fragstolen);
6121 tcp_event_data_recv(sk, skb);
6123 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
6124 /* Well, only one small jumplet in fast path... */
6125 tcp_ack(sk, skb, FLAG_DATA);
6126 tcp_data_snd_check(sk);
6127 if (!inet_csk_ack_scheduled(sk))
6130 tcp_update_wl(tp, TCP_SKB_CB(skb)->seq);
6133 __tcp_ack_snd_check(sk, 0);
6136 kfree_skb_partial(skb, fragstolen);
6143 if (len < (th->doff << 2) || tcp_checksum_complete(skb))
6146 if (!th->ack && !th->rst && !th->syn) {
6147 reason = SKB_DROP_REASON_TCP_FLAGS;
6152 * Standard slow path.
6155 if (!tcp_validate_incoming(sk, skb, th, 1))
6159 reason = tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT);
6160 if ((int)reason < 0) {
6164 tcp_rcv_rtt_measure_ts(sk, skb);
6166 /* Process urgent data. */
6167 tcp_urg(sk, skb, th);
6169 /* step 7: process the segment text */
6170 tcp_data_queue(sk, skb);
6172 tcp_data_snd_check(sk);
6173 tcp_ack_snd_check(sk);
6177 reason = SKB_DROP_REASON_TCP_CSUM;
6178 trace_tcp_bad_csum(skb);
6179 TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS);
6180 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
6183 tcp_drop_reason(sk, skb, reason);
6185 EXPORT_SYMBOL(tcp_rcv_established);
6187 void tcp_init_transfer(struct sock *sk, int bpf_op, struct sk_buff *skb)
6189 struct inet_connection_sock *icsk = inet_csk(sk);
6190 struct tcp_sock *tp = tcp_sk(sk);
6193 icsk->icsk_af_ops->rebuild_header(sk);
6194 tcp_init_metrics(sk);
6196 /* Initialize the congestion window to start the transfer.
6197 * Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
6198 * retransmitted. In light of RFC6298 more aggressive 1sec
6199 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
6200 * retransmission has occurred.
6202 if (tp->total_retrans > 1 && tp->undo_marker)
6203 tcp_snd_cwnd_set(tp, 1);
6205 tcp_snd_cwnd_set(tp, tcp_init_cwnd(tp, __sk_dst_get(sk)));
6206 tp->snd_cwnd_stamp = tcp_jiffies32;
6208 bpf_skops_established(sk, bpf_op, skb);
6209 /* Initialize congestion control unless BPF initialized it already: */
6210 if (!icsk->icsk_ca_initialized)
6211 tcp_init_congestion_control(sk);
6212 tcp_init_buffer_space(sk);
6215 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
6217 struct tcp_sock *tp = tcp_sk(sk);
6218 struct inet_connection_sock *icsk = inet_csk(sk);
6220 tcp_ao_finish_connect(sk, skb);
6221 tcp_set_state(sk, TCP_ESTABLISHED);
6222 icsk->icsk_ack.lrcvtime = tcp_jiffies32;
6225 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
6226 security_inet_conn_established(sk, skb);
6227 sk_mark_napi_id(sk, skb);
6230 tcp_init_transfer(sk, BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB, skb);
6232 /* Prevent spurious tcp_cwnd_restart() on first data
6235 tp->lsndtime = tcp_jiffies32;
6237 if (sock_flag(sk, SOCK_KEEPOPEN))
6238 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
6240 if (!tp->rx_opt.snd_wscale)
6241 __tcp_fast_path_on(tp, tp->snd_wnd);
6246 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
6247 struct tcp_fastopen_cookie *cookie)
6249 struct tcp_sock *tp = tcp_sk(sk);
6250 struct sk_buff *data = tp->syn_data ? tcp_rtx_queue_head(sk) : NULL;
6251 u16 mss = tp->rx_opt.mss_clamp, try_exp = 0;
6252 bool syn_drop = false;
6254 if (mss == tp->rx_opt.user_mss) {
6255 struct tcp_options_received opt;
6257 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
6258 tcp_clear_options(&opt);
6259 opt.user_mss = opt.mss_clamp = 0;
6260 tcp_parse_options(sock_net(sk), synack, &opt, 0, NULL);
6261 mss = opt.mss_clamp;
6264 if (!tp->syn_fastopen) {
6265 /* Ignore an unsolicited cookie */
6267 } else if (tp->total_retrans) {
6268 /* SYN timed out and the SYN-ACK neither has a cookie nor
6269 * acknowledges data. Presumably the remote received only
6270 * the retransmitted (regular) SYNs: either the original
6271 * SYN-data or the corresponding SYN-ACK was dropped.
6273 syn_drop = (cookie->len < 0 && data);
6274 } else if (cookie->len < 0 && !tp->syn_data) {
6275 /* We requested a cookie but didn't get it. If we did not use
6276 * the (old) exp opt format then try so next time (try_exp=1).
6277 * Otherwise we go back to use the RFC7413 opt (try_exp=2).
6279 try_exp = tp->syn_fastopen_exp ? 2 : 1;
6282 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp);
6284 if (data) { /* Retransmit unacked data in SYN */
6285 if (tp->total_retrans)
6286 tp->fastopen_client_fail = TFO_SYN_RETRANSMITTED;
6288 tp->fastopen_client_fail = TFO_DATA_NOT_ACKED;
6289 skb_rbtree_walk_from(data)
6290 tcp_mark_skb_lost(sk, data);
6291 tcp_xmit_retransmit_queue(sk);
6292 NET_INC_STATS(sock_net(sk),
6293 LINUX_MIB_TCPFASTOPENACTIVEFAIL);
6296 tp->syn_data_acked = tp->syn_data;
6297 if (tp->syn_data_acked) {
6298 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
6299 /* SYN-data is counted as two separate packets in tcp_ack() */
6300 if (tp->delivered > 1)
6304 tcp_fastopen_add_skb(sk, synack);
6309 static void smc_check_reset_syn(struct tcp_sock *tp)
6311 #if IS_ENABLED(CONFIG_SMC)
6312 if (static_branch_unlikely(&tcp_have_smc)) {
6313 if (tp->syn_smc && !tp->rx_opt.smc_ok)
6319 static void tcp_try_undo_spurious_syn(struct sock *sk)
6321 struct tcp_sock *tp = tcp_sk(sk);
6324 /* undo_marker is set when SYN or SYNACK times out. The timeout is
6325 * spurious if the ACK's timestamp option echo value matches the
6326 * original SYN timestamp.
6328 syn_stamp = tp->retrans_stamp;
6329 if (tp->undo_marker && syn_stamp && tp->rx_opt.saw_tstamp &&
6330 syn_stamp == tp->rx_opt.rcv_tsecr)
6331 tp->undo_marker = 0;
6334 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
6335 const struct tcphdr *th)
6337 struct inet_connection_sock *icsk = inet_csk(sk);
6338 struct tcp_sock *tp = tcp_sk(sk);
6339 struct tcp_fastopen_cookie foc = { .len = -1 };
6340 int saved_clamp = tp->rx_opt.mss_clamp;
6344 tcp_parse_options(sock_net(sk), skb, &tp->rx_opt, 0, &foc);
6345 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
6346 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
6350 * "If the state is SYN-SENT then
6351 * first check the ACK bit
6352 * If the ACK bit is set
6353 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
6354 * a reset (unless the RST bit is set, if so drop
6355 * the segment and return)"
6357 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
6358 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
6359 /* Previous FIN/ACK or RST/ACK might be ignored. */
6360 if (icsk->icsk_retransmits == 0)
6361 inet_csk_reset_xmit_timer(sk,
6363 TCP_TIMEOUT_MIN, TCP_RTO_MAX);
6364 goto reset_and_undo;
6367 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
6368 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
6369 tcp_time_stamp_ts(tp))) {
6370 NET_INC_STATS(sock_net(sk),
6371 LINUX_MIB_PAWSACTIVEREJECTED);
6372 goto reset_and_undo;
6375 /* Now ACK is acceptable.
6377 * "If the RST bit is set
6378 * If the ACK was acceptable then signal the user "error:
6379 * connection reset", drop the segment, enter CLOSED state,
6380 * delete TCB, and return."
6391 * "fifth, if neither of the SYN or RST bits is set then
6392 * drop the segment and return."
6398 SKB_DR_SET(reason, TCP_FLAGS);
6399 goto discard_and_undo;
6402 * "If the SYN bit is on ...
6403 * are acceptable then ...
6404 * (our SYN has been ACKed), change the connection
6405 * state to ESTABLISHED..."
6408 tcp_ecn_rcv_synack(tp, th);
6410 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
6411 tcp_try_undo_spurious_syn(sk);
6412 tcp_ack(sk, skb, FLAG_SLOWPATH);
6414 /* Ok.. it's good. Set up sequence numbers and
6415 * move to established.
6417 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1);
6418 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
6420 /* RFC1323: The window in SYN & SYN/ACK segments is
6423 tp->snd_wnd = ntohs(th->window);
6425 if (!tp->rx_opt.wscale_ok) {
6426 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
6427 tp->window_clamp = min(tp->window_clamp, 65535U);
6430 if (tp->rx_opt.saw_tstamp) {
6431 tp->rx_opt.tstamp_ok = 1;
6432 tp->tcp_header_len =
6433 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
6434 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
6435 tcp_store_ts_recent(tp);
6437 tp->tcp_header_len = sizeof(struct tcphdr);
6440 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
6441 tcp_initialize_rcv_mss(sk);
6443 /* Remember, tcp_poll() does not lock socket!
6444 * Change state from SYN-SENT only after copied_seq
6445 * is initialized. */
6446 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6448 smc_check_reset_syn(tp);
6452 tcp_finish_connect(sk, skb);
6454 fastopen_fail = (tp->syn_fastopen || tp->syn_data) &&
6455 tcp_rcv_fastopen_synack(sk, skb, &foc);
6457 if (!sock_flag(sk, SOCK_DEAD)) {
6458 sk->sk_state_change(sk);
6459 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
6463 if (sk->sk_write_pending ||
6464 READ_ONCE(icsk->icsk_accept_queue.rskq_defer_accept) ||
6465 inet_csk_in_pingpong_mode(sk)) {
6466 /* Save one ACK. Data will be ready after
6467 * several ticks, if write_pending is set.
6469 * It may be deleted, but with this feature tcpdumps
6470 * look so _wonderfully_ clever, that I was not able
6471 * to stand against the temptation 8) --ANK
6473 inet_csk_schedule_ack(sk);
6474 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
6475 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
6476 TCP_DELACK_MAX, TCP_RTO_MAX);
6483 /* No ACK in the segment */
6487 * "If the RST bit is set
6489 * Otherwise (no ACK) drop the segment and return."
6491 SKB_DR_SET(reason, TCP_RESET);
6492 goto discard_and_undo;
6496 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
6497 tcp_paws_reject(&tp->rx_opt, 0)) {
6498 SKB_DR_SET(reason, TCP_RFC7323_PAWS);
6499 goto discard_and_undo;
6502 /* We see SYN without ACK. It is attempt of
6503 * simultaneous connect with crossed SYNs.
6504 * Particularly, it can be connect to self.
6506 #ifdef CONFIG_TCP_AO
6507 struct tcp_ao_info *ao;
6509 ao = rcu_dereference_protected(tp->ao_info,
6510 lockdep_sock_is_held(sk));
6512 WRITE_ONCE(ao->risn, th->seq);
6516 tcp_set_state(sk, TCP_SYN_RECV);
6518 if (tp->rx_opt.saw_tstamp) {
6519 tp->rx_opt.tstamp_ok = 1;
6520 tcp_store_ts_recent(tp);
6521 tp->tcp_header_len =
6522 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
6524 tp->tcp_header_len = sizeof(struct tcphdr);
6527 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1);
6528 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6529 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
6531 /* RFC1323: The window in SYN & SYN/ACK segments is
6534 tp->snd_wnd = ntohs(th->window);
6535 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
6536 tp->max_window = tp->snd_wnd;
6538 tcp_ecn_rcv_syn(tp, th);
6541 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
6542 tcp_initialize_rcv_mss(sk);
6544 tcp_send_synack(sk);
6546 /* Note, we could accept data and URG from this segment.
6547 * There are no obstacles to make this (except that we must
6548 * either change tcp_recvmsg() to prevent it from returning data
6549 * before 3WHS completes per RFC793, or employ TCP Fast Open).
6551 * However, if we ignore data in ACKless segments sometimes,
6552 * we have no reasons to accept it sometimes.
6553 * Also, seems the code doing it in step6 of tcp_rcv_state_process
6554 * is not flawless. So, discard packet for sanity.
6555 * Uncomment this return to process the data.
6562 /* "fifth, if neither of the SYN or RST bits is set then
6563 * drop the segment and return."
6567 tcp_clear_options(&tp->rx_opt);
6568 tp->rx_opt.mss_clamp = saved_clamp;
6569 tcp_drop_reason(sk, skb, reason);
6573 tcp_clear_options(&tp->rx_opt);
6574 tp->rx_opt.mss_clamp = saved_clamp;
6578 static void tcp_rcv_synrecv_state_fastopen(struct sock *sk)
6580 struct tcp_sock *tp = tcp_sk(sk);
6581 struct request_sock *req;
6583 /* If we are still handling the SYNACK RTO, see if timestamp ECR allows
6584 * undo. If peer SACKs triggered fast recovery, we can't undo here.
6586 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss && !tp->packets_out)
6587 tcp_try_undo_recovery(sk);
6589 /* Reset rtx states to prevent spurious retransmits_timed_out() */
6590 tcp_update_rto_time(tp);
6591 tp->retrans_stamp = 0;
6592 inet_csk(sk)->icsk_retransmits = 0;
6594 /* Once we leave TCP_SYN_RECV or TCP_FIN_WAIT_1,
6595 * we no longer need req so release it.
6597 req = rcu_dereference_protected(tp->fastopen_rsk,
6598 lockdep_sock_is_held(sk));
6599 reqsk_fastopen_remove(sk, req, false);
6601 /* Re-arm the timer because data may have been sent out.
6602 * This is similar to the regular data transmission case
6603 * when new data has just been ack'ed.
6605 * (TFO) - we could try to be more aggressive and
6606 * retransmitting any data sooner based on when they
6613 * This function implements the receiving procedure of RFC 793 for
6614 * all states except ESTABLISHED and TIME_WAIT.
6615 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
6616 * address independent.
6619 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb)
6621 struct tcp_sock *tp = tcp_sk(sk);
6622 struct inet_connection_sock *icsk = inet_csk(sk);
6623 const struct tcphdr *th = tcp_hdr(skb);
6624 struct request_sock *req;
6629 switch (sk->sk_state) {
6631 SKB_DR_SET(reason, TCP_CLOSE);
6639 SKB_DR_SET(reason, TCP_RESET);
6644 SKB_DR_SET(reason, TCP_FLAGS);
6647 /* It is possible that we process SYN packets from backlog,
6648 * so we need to make sure to disable BH and RCU right there.
6652 acceptable = icsk->icsk_af_ops->conn_request(sk, skb) >= 0;
6661 SKB_DR_SET(reason, TCP_FLAGS);
6665 tp->rx_opt.saw_tstamp = 0;
6666 tcp_mstamp_refresh(tp);
6667 queued = tcp_rcv_synsent_state_process(sk, skb, th);
6671 /* Do step6 onward by hand. */
6672 tcp_urg(sk, skb, th);
6674 tcp_data_snd_check(sk);
6678 tcp_mstamp_refresh(tp);
6679 tp->rx_opt.saw_tstamp = 0;
6680 req = rcu_dereference_protected(tp->fastopen_rsk,
6681 lockdep_sock_is_held(sk));
6685 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
6686 sk->sk_state != TCP_FIN_WAIT1);
6688 if (!tcp_check_req(sk, skb, req, true, &req_stolen)) {
6689 SKB_DR_SET(reason, TCP_FASTOPEN);
6694 if (!th->ack && !th->rst && !th->syn) {
6695 SKB_DR_SET(reason, TCP_FLAGS);
6698 if (!tcp_validate_incoming(sk, skb, th, 0))
6701 /* step 5: check the ACK field */
6702 acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
6703 FLAG_UPDATE_TS_RECENT |
6704 FLAG_NO_CHALLENGE_ACK) > 0;
6707 if (sk->sk_state == TCP_SYN_RECV)
6708 return 1; /* send one RST */
6709 tcp_send_challenge_ack(sk);
6710 SKB_DR_SET(reason, TCP_OLD_ACK);
6713 switch (sk->sk_state) {
6715 tp->delivered++; /* SYN-ACK delivery isn't tracked in tcp_ack */
6717 tcp_synack_rtt_meas(sk, req);
6720 tcp_rcv_synrecv_state_fastopen(sk);
6722 tcp_try_undo_spurious_syn(sk);
6723 tp->retrans_stamp = 0;
6724 tcp_init_transfer(sk, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB,
6726 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6728 tcp_ao_established(sk);
6730 tcp_set_state(sk, TCP_ESTABLISHED);
6731 sk->sk_state_change(sk);
6733 /* Note, that this wakeup is only for marginal crossed SYN case.
6734 * Passively open sockets are not waked up, because
6735 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
6738 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
6740 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
6741 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
6742 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
6744 if (tp->rx_opt.tstamp_ok)
6745 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
6747 if (!inet_csk(sk)->icsk_ca_ops->cong_control)
6748 tcp_update_pacing_rate(sk);
6750 /* Prevent spurious tcp_cwnd_restart() on first data packet */
6751 tp->lsndtime = tcp_jiffies32;
6753 tcp_initialize_rcv_mss(sk);
6754 tcp_fast_path_on(tp);
6757 case TCP_FIN_WAIT1: {
6761 tcp_rcv_synrecv_state_fastopen(sk);
6763 if (tp->snd_una != tp->write_seq)
6766 tcp_set_state(sk, TCP_FIN_WAIT2);
6767 WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | SEND_SHUTDOWN);
6771 if (!sock_flag(sk, SOCK_DEAD)) {
6772 /* Wake up lingering close() */
6773 sk->sk_state_change(sk);
6777 if (READ_ONCE(tp->linger2) < 0) {
6779 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6782 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6783 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6784 /* Receive out of order FIN after close() */
6785 if (tp->syn_fastopen && th->fin)
6786 tcp_fastopen_active_disable(sk);
6788 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6792 tmo = tcp_fin_time(sk);
6793 if (tmo > TCP_TIMEWAIT_LEN) {
6794 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
6795 } else if (th->fin || sock_owned_by_user(sk)) {
6796 /* Bad case. We could lose such FIN otherwise.
6797 * It is not a big problem, but it looks confusing
6798 * and not so rare event. We still can lose it now,
6799 * if it spins in bh_lock_sock(), but it is really
6802 inet_csk_reset_keepalive_timer(sk, tmo);
6804 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
6811 if (tp->snd_una == tp->write_seq) {
6812 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
6818 if (tp->snd_una == tp->write_seq) {
6819 tcp_update_metrics(sk);
6826 /* step 6: check the URG bit */
6827 tcp_urg(sk, skb, th);
6829 /* step 7: process the segment text */
6830 switch (sk->sk_state) {
6831 case TCP_CLOSE_WAIT:
6834 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
6835 /* If a subflow has been reset, the packet should not
6836 * continue to be processed, drop the packet.
6838 if (sk_is_mptcp(sk) && !mptcp_incoming_options(sk, skb))
6845 /* RFC 793 says to queue data in these states,
6846 * RFC 1122 says we MUST send a reset.
6847 * BSD 4.4 also does reset.
6849 if (sk->sk_shutdown & RCV_SHUTDOWN) {
6850 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6851 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6852 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6858 case TCP_ESTABLISHED:
6859 tcp_data_queue(sk, skb);
6864 /* tcp_data could move socket to TIME-WAIT */
6865 if (sk->sk_state != TCP_CLOSE) {
6866 tcp_data_snd_check(sk);
6867 tcp_ack_snd_check(sk);
6872 tcp_drop_reason(sk, skb, reason);
6880 EXPORT_SYMBOL(tcp_rcv_state_process);
6882 static inline void pr_drop_req(struct request_sock *req, __u16 port, int family)
6884 struct inet_request_sock *ireq = inet_rsk(req);
6886 if (family == AF_INET)
6887 net_dbg_ratelimited("drop open request from %pI4/%u\n",
6888 &ireq->ir_rmt_addr, port);
6889 #if IS_ENABLED(CONFIG_IPV6)
6890 else if (family == AF_INET6)
6891 net_dbg_ratelimited("drop open request from %pI6/%u\n",
6892 &ireq->ir_v6_rmt_addr, port);
6896 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
6898 * If we receive a SYN packet with these bits set, it means a
6899 * network is playing bad games with TOS bits. In order to
6900 * avoid possible false congestion notifications, we disable
6901 * TCP ECN negotiation.
6903 * Exception: tcp_ca wants ECN. This is required for DCTCP
6904 * congestion control: Linux DCTCP asserts ECT on all packets,
6905 * including SYN, which is most optimal solution; however,
6906 * others, such as FreeBSD do not.
6908 * Exception: At least one of the reserved bits of the TCP header (th->res1) is
6909 * set, indicating the use of a future TCP extension (such as AccECN). See
6910 * RFC8311 §4.3 which updates RFC3168 to allow the development of such
6913 static void tcp_ecn_create_request(struct request_sock *req,
6914 const struct sk_buff *skb,
6915 const struct sock *listen_sk,
6916 const struct dst_entry *dst)
6918 const struct tcphdr *th = tcp_hdr(skb);
6919 const struct net *net = sock_net(listen_sk);
6920 bool th_ecn = th->ece && th->cwr;
6927 ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield);
6928 ecn_ok_dst = dst_feature(dst, DST_FEATURE_ECN_MASK);
6929 ecn_ok = READ_ONCE(net->ipv4.sysctl_tcp_ecn) || ecn_ok_dst;
6931 if (((!ect || th->res1) && ecn_ok) || tcp_ca_needs_ecn(listen_sk) ||
6932 (ecn_ok_dst & DST_FEATURE_ECN_CA) ||
6933 tcp_bpf_ca_needs_ecn((struct sock *)req))
6934 inet_rsk(req)->ecn_ok = 1;
6937 static void tcp_openreq_init(struct request_sock *req,
6938 const struct tcp_options_received *rx_opt,
6939 struct sk_buff *skb, const struct sock *sk)
6941 struct inet_request_sock *ireq = inet_rsk(req);
6943 req->rsk_rcv_wnd = 0; /* So that tcp_send_synack() knows! */
6944 tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq;
6945 tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
6946 tcp_rsk(req)->snt_synack = 0;
6947 tcp_rsk(req)->last_oow_ack_time = 0;
6948 req->mss = rx_opt->mss_clamp;
6949 req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0;
6950 ireq->tstamp_ok = rx_opt->tstamp_ok;
6951 ireq->sack_ok = rx_opt->sack_ok;
6952 ireq->snd_wscale = rx_opt->snd_wscale;
6953 ireq->wscale_ok = rx_opt->wscale_ok;
6956 ireq->ir_rmt_port = tcp_hdr(skb)->source;
6957 ireq->ir_num = ntohs(tcp_hdr(skb)->dest);
6958 ireq->ir_mark = inet_request_mark(sk, skb);
6959 #if IS_ENABLED(CONFIG_SMC)
6960 ireq->smc_ok = rx_opt->smc_ok && !(tcp_sk(sk)->smc_hs_congested &&
6961 tcp_sk(sk)->smc_hs_congested(sk));
6965 struct request_sock *inet_reqsk_alloc(const struct request_sock_ops *ops,
6966 struct sock *sk_listener,
6967 bool attach_listener)
6969 struct request_sock *req = reqsk_alloc(ops, sk_listener,
6973 struct inet_request_sock *ireq = inet_rsk(req);
6975 ireq->ireq_opt = NULL;
6976 #if IS_ENABLED(CONFIG_IPV6)
6977 ireq->pktopts = NULL;
6979 atomic64_set(&ireq->ir_cookie, 0);
6980 ireq->ireq_state = TCP_NEW_SYN_RECV;
6981 write_pnet(&ireq->ireq_net, sock_net(sk_listener));
6982 ireq->ireq_family = sk_listener->sk_family;
6983 req->timeout = TCP_TIMEOUT_INIT;
6988 EXPORT_SYMBOL(inet_reqsk_alloc);
6991 * Return true if a syncookie should be sent
6993 static bool tcp_syn_flood_action(const struct sock *sk, const char *proto)
6995 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
6996 const char *msg = "Dropping request";
6997 struct net *net = sock_net(sk);
6998 bool want_cookie = false;
7001 syncookies = READ_ONCE(net->ipv4.sysctl_tcp_syncookies);
7003 #ifdef CONFIG_SYN_COOKIES
7005 msg = "Sending cookies";
7007 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES);
7010 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP);
7012 if (!READ_ONCE(queue->synflood_warned) && syncookies != 2 &&
7013 xchg(&queue->synflood_warned, 1) == 0) {
7014 if (IS_ENABLED(CONFIG_IPV6) && sk->sk_family == AF_INET6) {
7015 net_info_ratelimited("%s: Possible SYN flooding on port [%pI6c]:%u. %s.\n",
7016 proto, inet6_rcv_saddr(sk),
7019 net_info_ratelimited("%s: Possible SYN flooding on port %pI4:%u. %s.\n",
7020 proto, &sk->sk_rcv_saddr,
7028 static void tcp_reqsk_record_syn(const struct sock *sk,
7029 struct request_sock *req,
7030 const struct sk_buff *skb)
7032 if (tcp_sk(sk)->save_syn) {
7033 u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb);
7034 struct saved_syn *saved_syn;
7038 if (tcp_sk(sk)->save_syn == 2) { /* Save full header. */
7039 base = skb_mac_header(skb);
7040 mac_hdrlen = skb_mac_header_len(skb);
7043 base = skb_network_header(skb);
7047 saved_syn = kmalloc(struct_size(saved_syn, data, len),
7050 saved_syn->mac_hdrlen = mac_hdrlen;
7051 saved_syn->network_hdrlen = skb_network_header_len(skb);
7052 saved_syn->tcp_hdrlen = tcp_hdrlen(skb);
7053 memcpy(saved_syn->data, base, len);
7054 req->saved_syn = saved_syn;
7059 /* If a SYN cookie is required and supported, returns a clamped MSS value to be
7060 * used for SYN cookie generation.
7062 u16 tcp_get_syncookie_mss(struct request_sock_ops *rsk_ops,
7063 const struct tcp_request_sock_ops *af_ops,
7064 struct sock *sk, struct tcphdr *th)
7066 struct tcp_sock *tp = tcp_sk(sk);
7069 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_syncookies) != 2 &&
7070 !inet_csk_reqsk_queue_is_full(sk))
7073 if (!tcp_syn_flood_action(sk, rsk_ops->slab_name))
7076 if (sk_acceptq_is_full(sk)) {
7077 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
7081 mss = tcp_parse_mss_option(th, tp->rx_opt.user_mss);
7083 mss = af_ops->mss_clamp;
7087 EXPORT_SYMBOL_GPL(tcp_get_syncookie_mss);
7089 int tcp_conn_request(struct request_sock_ops *rsk_ops,
7090 const struct tcp_request_sock_ops *af_ops,
7091 struct sock *sk, struct sk_buff *skb)
7093 struct tcp_fastopen_cookie foc = { .len = -1 };
7094 __u32 isn = TCP_SKB_CB(skb)->tcp_tw_isn;
7095 struct tcp_options_received tmp_opt;
7096 struct tcp_sock *tp = tcp_sk(sk);
7097 struct net *net = sock_net(sk);
7098 struct sock *fastopen_sk = NULL;
7099 struct request_sock *req;
7100 bool want_cookie = false;
7101 struct dst_entry *dst;
7105 #ifdef CONFIG_TCP_AO
7106 const struct tcp_ao_hdr *aoh;
7109 syncookies = READ_ONCE(net->ipv4.sysctl_tcp_syncookies);
7111 /* TW buckets are converted to open requests without
7112 * limitations, they conserve resources and peer is
7113 * evidently real one.
7115 if ((syncookies == 2 || inet_csk_reqsk_queue_is_full(sk)) && !isn) {
7116 want_cookie = tcp_syn_flood_action(sk, rsk_ops->slab_name);
7121 if (sk_acceptq_is_full(sk)) {
7122 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
7126 req = inet_reqsk_alloc(rsk_ops, sk, !want_cookie);
7130 req->syncookie = want_cookie;
7131 tcp_rsk(req)->af_specific = af_ops;
7132 tcp_rsk(req)->ts_off = 0;
7133 tcp_rsk(req)->req_usec_ts = false;
7134 #if IS_ENABLED(CONFIG_MPTCP)
7135 tcp_rsk(req)->is_mptcp = 0;
7138 tcp_clear_options(&tmp_opt);
7139 tmp_opt.mss_clamp = af_ops->mss_clamp;
7140 tmp_opt.user_mss = tp->rx_opt.user_mss;
7141 tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0,
7142 want_cookie ? NULL : &foc);
7144 if (want_cookie && !tmp_opt.saw_tstamp)
7145 tcp_clear_options(&tmp_opt);
7147 if (IS_ENABLED(CONFIG_SMC) && want_cookie)
7150 tmp_opt.tstamp_ok = tmp_opt.saw_tstamp;
7151 tcp_openreq_init(req, &tmp_opt, skb, sk);
7152 inet_rsk(req)->no_srccheck = inet_test_bit(TRANSPARENT, sk);
7154 /* Note: tcp_v6_init_req() might override ir_iif for link locals */
7155 inet_rsk(req)->ir_iif = inet_request_bound_dev_if(sk, skb);
7157 dst = af_ops->route_req(sk, skb, &fl, req);
7161 if (tmp_opt.tstamp_ok) {
7162 tcp_rsk(req)->req_usec_ts = dst_tcp_usec_ts(dst);
7163 tcp_rsk(req)->ts_off = af_ops->init_ts_off(net, skb);
7165 if (!want_cookie && !isn) {
7166 int max_syn_backlog = READ_ONCE(net->ipv4.sysctl_max_syn_backlog);
7168 /* Kill the following clause, if you dislike this way. */
7170 (max_syn_backlog - inet_csk_reqsk_queue_len(sk) <
7171 (max_syn_backlog >> 2)) &&
7172 !tcp_peer_is_proven(req, dst)) {
7173 /* Without syncookies last quarter of
7174 * backlog is filled with destinations,
7175 * proven to be alive.
7176 * It means that we continue to communicate
7177 * to destinations, already remembered
7178 * to the moment of synflood.
7180 pr_drop_req(req, ntohs(tcp_hdr(skb)->source),
7182 goto drop_and_release;
7185 isn = af_ops->init_seq(skb);
7188 tcp_ecn_create_request(req, skb, sk, dst);
7191 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss);
7192 if (!tmp_opt.tstamp_ok)
7193 inet_rsk(req)->ecn_ok = 0;
7196 #ifdef CONFIG_TCP_AO
7197 if (tcp_parse_auth_options(tcp_hdr(skb), NULL, &aoh))
7198 goto drop_and_release; /* Invalid TCP options */
7200 tcp_rsk(req)->used_tcp_ao = true;
7201 tcp_rsk(req)->ao_rcv_next = aoh->keyid;
7202 tcp_rsk(req)->ao_keyid = aoh->rnext_keyid;
7205 tcp_rsk(req)->used_tcp_ao = false;
7208 tcp_rsk(req)->snt_isn = isn;
7209 tcp_rsk(req)->txhash = net_tx_rndhash();
7210 tcp_rsk(req)->syn_tos = TCP_SKB_CB(skb)->ip_dsfield;
7211 tcp_openreq_init_rwin(req, sk, dst);
7212 sk_rx_queue_set(req_to_sk(req), skb);
7214 tcp_reqsk_record_syn(sk, req, skb);
7215 fastopen_sk = tcp_try_fastopen(sk, skb, req, &foc, dst);
7218 af_ops->send_synack(fastopen_sk, dst, &fl, req,
7219 &foc, TCP_SYNACK_FASTOPEN, skb);
7220 /* Add the child socket directly into the accept queue */
7221 if (!inet_csk_reqsk_queue_add(sk, req, fastopen_sk)) {
7222 reqsk_fastopen_remove(fastopen_sk, req, false);
7223 bh_unlock_sock(fastopen_sk);
7224 sock_put(fastopen_sk);
7227 sk->sk_data_ready(sk);
7228 bh_unlock_sock(fastopen_sk);
7229 sock_put(fastopen_sk);
7231 tcp_rsk(req)->tfo_listener = false;
7233 req->timeout = tcp_timeout_init((struct sock *)req);
7234 inet_csk_reqsk_queue_hash_add(sk, req, req->timeout);
7236 af_ops->send_synack(sk, dst, &fl, req, &foc,
7237 !want_cookie ? TCP_SYNACK_NORMAL :
7256 EXPORT_SYMBOL(tcp_conn_request);