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 void tcp_gro_dev_warn(struct sock *sk, const struct sk_buff *skb,
208 static bool __once __read_mostly;
211 struct net_device *dev;
216 dev = dev_get_by_index_rcu(sock_net(sk), skb->skb_iif);
217 if (!dev || len >= dev->mtu)
218 pr_warn("%s: Driver has suspect GRO implementation, TCP performance may be compromised.\n",
219 dev ? dev->name : "Unknown driver");
224 /* Adapt the MSS value used to make delayed ack decision to the
227 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
229 struct inet_connection_sock *icsk = inet_csk(sk);
230 const unsigned int lss = icsk->icsk_ack.last_seg_size;
233 icsk->icsk_ack.last_seg_size = 0;
235 /* skb->len may jitter because of SACKs, even if peer
236 * sends good full-sized frames.
238 len = skb_shinfo(skb)->gso_size ? : skb->len;
239 if (len >= icsk->icsk_ack.rcv_mss) {
240 /* Note: divides are still a bit expensive.
241 * For the moment, only adjust scaling_ratio
242 * when we update icsk_ack.rcv_mss.
244 if (unlikely(len != icsk->icsk_ack.rcv_mss)) {
245 u64 val = (u64)skb->len << TCP_RMEM_TO_WIN_SCALE;
247 do_div(val, skb->truesize);
248 tcp_sk(sk)->scaling_ratio = val ? val : 1;
250 icsk->icsk_ack.rcv_mss = min_t(unsigned int, len,
252 /* Account for possibly-removed options */
253 if (unlikely(len > icsk->icsk_ack.rcv_mss +
254 MAX_TCP_OPTION_SPACE))
255 tcp_gro_dev_warn(sk, skb, len);
256 /* If the skb has a len of exactly 1*MSS and has the PSH bit
257 * set then it is likely the end of an application write. So
258 * more data may not be arriving soon, and yet the data sender
259 * may be waiting for an ACK if cwnd-bound or using TX zero
260 * copy. So we set ICSK_ACK_PUSHED here so that
261 * tcp_cleanup_rbuf() will send an ACK immediately if the app
262 * reads all of the data and is not ping-pong. If len > MSS
263 * then this logic does not matter (and does not hurt) because
264 * tcp_cleanup_rbuf() will always ACK immediately if the app
265 * reads data and there is more than an MSS of unACKed data.
267 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_PSH)
268 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
270 /* Otherwise, we make more careful check taking into account,
271 * that SACKs block is variable.
273 * "len" is invariant segment length, including TCP header.
275 len += skb->data - skb_transport_header(skb);
276 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
277 /* If PSH is not set, packet should be
278 * full sized, provided peer TCP is not badly broken.
279 * This observation (if it is correct 8)) allows
280 * to handle super-low mtu links fairly.
282 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
283 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
284 /* Subtract also invariant (if peer is RFC compliant),
285 * tcp header plus fixed timestamp option length.
286 * Resulting "len" is MSS free of SACK jitter.
288 len -= tcp_sk(sk)->tcp_header_len;
289 icsk->icsk_ack.last_seg_size = len;
291 icsk->icsk_ack.rcv_mss = len;
295 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
296 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
297 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
301 static void tcp_incr_quickack(struct sock *sk, unsigned int max_quickacks)
303 struct inet_connection_sock *icsk = inet_csk(sk);
304 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
308 quickacks = min(quickacks, max_quickacks);
309 if (quickacks > icsk->icsk_ack.quick)
310 icsk->icsk_ack.quick = quickacks;
313 static void tcp_enter_quickack_mode(struct sock *sk, unsigned int max_quickacks)
315 struct inet_connection_sock *icsk = inet_csk(sk);
317 tcp_incr_quickack(sk, max_quickacks);
318 inet_csk_exit_pingpong_mode(sk);
319 icsk->icsk_ack.ato = TCP_ATO_MIN;
322 /* Send ACKs quickly, if "quick" count is not exhausted
323 * and the session is not interactive.
326 static bool tcp_in_quickack_mode(struct sock *sk)
328 const struct inet_connection_sock *icsk = inet_csk(sk);
329 const struct dst_entry *dst = __sk_dst_get(sk);
331 return (dst && dst_metric(dst, RTAX_QUICKACK)) ||
332 (icsk->icsk_ack.quick && !inet_csk_in_pingpong_mode(sk));
335 static void tcp_ecn_queue_cwr(struct tcp_sock *tp)
337 if (tp->ecn_flags & TCP_ECN_OK)
338 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
341 static void tcp_ecn_accept_cwr(struct sock *sk, const struct sk_buff *skb)
343 if (tcp_hdr(skb)->cwr) {
344 tcp_sk(sk)->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
346 /* If the sender is telling us it has entered CWR, then its
347 * cwnd may be very low (even just 1 packet), so we should ACK
350 if (TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq)
351 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW;
355 static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp)
357 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
360 static void __tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb)
362 struct tcp_sock *tp = tcp_sk(sk);
364 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
365 case INET_ECN_NOT_ECT:
366 /* Funny extension: if ECT is not set on a segment,
367 * and we already seen ECT on a previous segment,
368 * it is probably a retransmit.
370 if (tp->ecn_flags & TCP_ECN_SEEN)
371 tcp_enter_quickack_mode(sk, 2);
374 if (tcp_ca_needs_ecn(sk))
375 tcp_ca_event(sk, CA_EVENT_ECN_IS_CE);
377 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
378 /* Better not delay acks, sender can have a very low cwnd */
379 tcp_enter_quickack_mode(sk, 2);
380 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
382 tp->ecn_flags |= TCP_ECN_SEEN;
385 if (tcp_ca_needs_ecn(sk))
386 tcp_ca_event(sk, CA_EVENT_ECN_NO_CE);
387 tp->ecn_flags |= TCP_ECN_SEEN;
392 static void tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb)
394 if (tcp_sk(sk)->ecn_flags & TCP_ECN_OK)
395 __tcp_ecn_check_ce(sk, skb);
398 static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
400 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
401 tp->ecn_flags &= ~TCP_ECN_OK;
404 static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
406 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
407 tp->ecn_flags &= ~TCP_ECN_OK;
410 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
412 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
417 /* Buffer size and advertised window tuning.
419 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
422 static void tcp_sndbuf_expand(struct sock *sk)
424 const struct tcp_sock *tp = tcp_sk(sk);
425 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
429 /* Worst case is non GSO/TSO : each frame consumes one skb
430 * and skb->head is kmalloced using power of two area of memory
432 per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
434 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
436 per_mss = roundup_pow_of_two(per_mss) +
437 SKB_DATA_ALIGN(sizeof(struct sk_buff));
439 nr_segs = max_t(u32, TCP_INIT_CWND, tcp_snd_cwnd(tp));
440 nr_segs = max_t(u32, nr_segs, tp->reordering + 1);
442 /* Fast Recovery (RFC 5681 3.2) :
443 * Cubic needs 1.7 factor, rounded to 2 to include
444 * extra cushion (application might react slowly to EPOLLOUT)
446 sndmem = ca_ops->sndbuf_expand ? ca_ops->sndbuf_expand(sk) : 2;
447 sndmem *= nr_segs * per_mss;
449 if (sk->sk_sndbuf < sndmem)
450 WRITE_ONCE(sk->sk_sndbuf,
451 min(sndmem, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_wmem[2])));
454 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
456 * All tcp_full_space() is split to two parts: "network" buffer, allocated
457 * forward and advertised in receiver window (tp->rcv_wnd) and
458 * "application buffer", required to isolate scheduling/application
459 * latencies from network.
460 * window_clamp is maximal advertised window. It can be less than
461 * tcp_full_space(), in this case tcp_full_space() - window_clamp
462 * is reserved for "application" buffer. The less window_clamp is
463 * the smoother our behaviour from viewpoint of network, but the lower
464 * throughput and the higher sensitivity of the connection to losses. 8)
466 * rcv_ssthresh is more strict window_clamp used at "slow start"
467 * phase to predict further behaviour of this connection.
468 * It is used for two goals:
469 * - to enforce header prediction at sender, even when application
470 * requires some significant "application buffer". It is check #1.
471 * - to prevent pruning of receive queue because of misprediction
472 * of receiver window. Check #2.
474 * The scheme does not work when sender sends good segments opening
475 * window and then starts to feed us spaghetti. But it should work
476 * in common situations. Otherwise, we have to rely on queue collapsing.
479 /* Slow part of check#2. */
480 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb,
481 unsigned int skbtruesize)
483 const struct tcp_sock *tp = tcp_sk(sk);
485 int truesize = tcp_win_from_space(sk, skbtruesize) >> 1;
486 int window = tcp_win_from_space(sk, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2])) >> 1;
488 while (tp->rcv_ssthresh <= window) {
489 if (truesize <= skb->len)
490 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
498 /* Even if skb appears to have a bad len/truesize ratio, TCP coalescing
499 * can play nice with us, as sk_buff and skb->head might be either
500 * freed or shared with up to MAX_SKB_FRAGS segments.
501 * Only give a boost to drivers using page frag(s) to hold the frame(s),
502 * and if no payload was pulled in skb->head before reaching us.
504 static u32 truesize_adjust(bool adjust, const struct sk_buff *skb)
506 u32 truesize = skb->truesize;
508 if (adjust && !skb_headlen(skb)) {
509 truesize -= SKB_TRUESIZE(skb_end_offset(skb));
510 /* paranoid check, some drivers might be buggy */
511 if (unlikely((int)truesize < (int)skb->len))
512 truesize = skb->truesize;
517 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb,
520 struct tcp_sock *tp = tcp_sk(sk);
523 room = min_t(int, tp->window_clamp, tcp_space(sk)) - tp->rcv_ssthresh;
529 if (!tcp_under_memory_pressure(sk)) {
530 unsigned int truesize = truesize_adjust(adjust, skb);
533 /* Check #2. Increase window, if skb with such overhead
534 * will fit to rcvbuf in future.
536 if (tcp_win_from_space(sk, truesize) <= skb->len)
537 incr = 2 * tp->advmss;
539 incr = __tcp_grow_window(sk, skb, truesize);
542 incr = max_t(int, incr, 2 * skb->len);
543 tp->rcv_ssthresh += min(room, incr);
544 inet_csk(sk)->icsk_ack.quick |= 1;
548 * Adjust rcv_ssthresh according to reserved mem
550 tcp_adjust_rcv_ssthresh(sk);
554 /* 3. Try to fixup all. It is made immediately after connection enters
557 static void tcp_init_buffer_space(struct sock *sk)
559 int tcp_app_win = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_app_win);
560 struct tcp_sock *tp = tcp_sk(sk);
563 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
564 tcp_sndbuf_expand(sk);
566 tcp_mstamp_refresh(tp);
567 tp->rcvq_space.time = tp->tcp_mstamp;
568 tp->rcvq_space.seq = tp->copied_seq;
570 maxwin = tcp_full_space(sk);
572 if (tp->window_clamp >= maxwin) {
573 tp->window_clamp = maxwin;
575 if (tcp_app_win && maxwin > 4 * tp->advmss)
576 tp->window_clamp = max(maxwin -
577 (maxwin >> tcp_app_win),
581 /* Force reservation of one segment. */
583 tp->window_clamp > 2 * tp->advmss &&
584 tp->window_clamp + tp->advmss > maxwin)
585 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
587 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
588 tp->snd_cwnd_stamp = tcp_jiffies32;
589 tp->rcvq_space.space = min3(tp->rcv_ssthresh, tp->rcv_wnd,
590 (u32)TCP_INIT_CWND * tp->advmss);
593 /* 4. Recalculate window clamp after socket hit its memory bounds. */
594 static void tcp_clamp_window(struct sock *sk)
596 struct tcp_sock *tp = tcp_sk(sk);
597 struct inet_connection_sock *icsk = inet_csk(sk);
598 struct net *net = sock_net(sk);
601 icsk->icsk_ack.quick = 0;
602 rmem2 = READ_ONCE(net->ipv4.sysctl_tcp_rmem[2]);
604 if (sk->sk_rcvbuf < rmem2 &&
605 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
606 !tcp_under_memory_pressure(sk) &&
607 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
608 WRITE_ONCE(sk->sk_rcvbuf,
609 min(atomic_read(&sk->sk_rmem_alloc), rmem2));
611 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
612 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
615 /* Initialize RCV_MSS value.
616 * RCV_MSS is an our guess about MSS used by the peer.
617 * We haven't any direct information about the MSS.
618 * It's better to underestimate the RCV_MSS rather than overestimate.
619 * Overestimations make us ACKing less frequently than needed.
620 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
622 void tcp_initialize_rcv_mss(struct sock *sk)
624 const struct tcp_sock *tp = tcp_sk(sk);
625 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
627 hint = min(hint, tp->rcv_wnd / 2);
628 hint = min(hint, TCP_MSS_DEFAULT);
629 hint = max(hint, TCP_MIN_MSS);
631 inet_csk(sk)->icsk_ack.rcv_mss = hint;
633 EXPORT_SYMBOL(tcp_initialize_rcv_mss);
635 /* Receiver "autotuning" code.
637 * The algorithm for RTT estimation w/o timestamps is based on
638 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
639 * <https://public.lanl.gov/radiant/pubs.html#DRS>
641 * More detail on this code can be found at
642 * <http://staff.psc.edu/jheffner/>,
643 * though this reference is out of date. A new paper
646 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
648 u32 new_sample = tp->rcv_rtt_est.rtt_us;
651 if (new_sample != 0) {
652 /* If we sample in larger samples in the non-timestamp
653 * case, we could grossly overestimate the RTT especially
654 * with chatty applications or bulk transfer apps which
655 * are stalled on filesystem I/O.
657 * Also, since we are only going for a minimum in the
658 * non-timestamp case, we do not smooth things out
659 * else with timestamps disabled convergence takes too
663 m -= (new_sample >> 3);
671 /* No previous measure. */
675 tp->rcv_rtt_est.rtt_us = new_sample;
678 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
682 if (tp->rcv_rtt_est.time == 0)
684 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
686 delta_us = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcv_rtt_est.time);
689 tcp_rcv_rtt_update(tp, delta_us, 1);
692 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
693 tp->rcv_rtt_est.time = tp->tcp_mstamp;
696 static s32 tcp_rtt_tsopt_us(const struct tcp_sock *tp)
700 delta = tcp_time_stamp_ts(tp) - tp->rx_opt.rcv_tsecr;
704 if (likely(delta < INT_MAX / (USEC_PER_SEC / TCP_TS_HZ))) {
707 delta_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
713 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
714 const struct sk_buff *skb)
716 struct tcp_sock *tp = tcp_sk(sk);
718 if (tp->rx_opt.rcv_tsecr == tp->rcv_rtt_last_tsecr)
720 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr;
722 if (TCP_SKB_CB(skb)->end_seq -
723 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss) {
724 s32 delta = tcp_rtt_tsopt_us(tp);
727 tcp_rcv_rtt_update(tp, delta, 0);
732 * This function should be called every time data is copied to user space.
733 * It calculates the appropriate TCP receive buffer space.
735 void tcp_rcv_space_adjust(struct sock *sk)
737 struct tcp_sock *tp = tcp_sk(sk);
741 trace_tcp_rcv_space_adjust(sk);
743 tcp_mstamp_refresh(tp);
744 time = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcvq_space.time);
745 if (time < (tp->rcv_rtt_est.rtt_us >> 3) || tp->rcv_rtt_est.rtt_us == 0)
748 /* Number of bytes copied to user in last RTT */
749 copied = tp->copied_seq - tp->rcvq_space.seq;
750 if (copied <= tp->rcvq_space.space)
754 * copied = bytes received in previous RTT, our base window
755 * To cope with packet losses, we need a 2x factor
756 * To cope with slow start, and sender growing its cwin by 100 %
757 * every RTT, we need a 4x factor, because the ACK we are sending
758 * now is for the next RTT, not the current one :
759 * <prev RTT . ><current RTT .. ><next RTT .... >
762 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf) &&
763 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
767 /* minimal window to cope with packet losses, assuming
768 * steady state. Add some cushion because of small variations.
770 rcvwin = ((u64)copied << 1) + 16 * tp->advmss;
772 /* Accommodate for sender rate increase (eg. slow start) */
773 grow = rcvwin * (copied - tp->rcvq_space.space);
774 do_div(grow, tp->rcvq_space.space);
775 rcvwin += (grow << 1);
777 rcvbuf = min_t(u64, tcp_space_from_win(sk, rcvwin),
778 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]));
779 if (rcvbuf > sk->sk_rcvbuf) {
780 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
782 /* Make the window clamp follow along. */
783 tp->window_clamp = tcp_win_from_space(sk, rcvbuf);
786 tp->rcvq_space.space = copied;
789 tp->rcvq_space.seq = tp->copied_seq;
790 tp->rcvq_space.time = tp->tcp_mstamp;
793 static void tcp_save_lrcv_flowlabel(struct sock *sk, const struct sk_buff *skb)
795 #if IS_ENABLED(CONFIG_IPV6)
796 struct inet_connection_sock *icsk = inet_csk(sk);
798 if (skb->protocol == htons(ETH_P_IPV6))
799 icsk->icsk_ack.lrcv_flowlabel = ntohl(ip6_flowlabel(ipv6_hdr(skb)));
803 /* There is something which you must keep in mind when you analyze the
804 * behavior of the tp->ato delayed ack timeout interval. When a
805 * connection starts up, we want to ack as quickly as possible. The
806 * problem is that "good" TCP's do slow start at the beginning of data
807 * transmission. The means that until we send the first few ACK's the
808 * sender will sit on his end and only queue most of his data, because
809 * he can only send snd_cwnd unacked packets at any given time. For
810 * each ACK we send, he increments snd_cwnd and transmits more of his
813 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
815 struct tcp_sock *tp = tcp_sk(sk);
816 struct inet_connection_sock *icsk = inet_csk(sk);
819 inet_csk_schedule_ack(sk);
821 tcp_measure_rcv_mss(sk, skb);
823 tcp_rcv_rtt_measure(tp);
827 if (!icsk->icsk_ack.ato) {
828 /* The _first_ data packet received, initialize
829 * delayed ACK engine.
831 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
832 icsk->icsk_ack.ato = TCP_ATO_MIN;
834 int m = now - icsk->icsk_ack.lrcvtime;
836 if (m <= TCP_ATO_MIN / 2) {
837 /* The fastest case is the first. */
838 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
839 } else if (m < icsk->icsk_ack.ato) {
840 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
841 if (icsk->icsk_ack.ato > icsk->icsk_rto)
842 icsk->icsk_ack.ato = icsk->icsk_rto;
843 } else if (m > icsk->icsk_rto) {
844 /* Too long gap. Apparently sender failed to
845 * restart window, so that we send ACKs quickly.
847 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
850 icsk->icsk_ack.lrcvtime = now;
851 tcp_save_lrcv_flowlabel(sk, skb);
853 tcp_ecn_check_ce(sk, skb);
856 tcp_grow_window(sk, skb, true);
859 /* Called to compute a smoothed rtt estimate. The data fed to this
860 * routine either comes from timestamps, or from segments that were
861 * known _not_ to have been retransmitted [see Karn/Partridge
862 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
863 * piece by Van Jacobson.
864 * NOTE: the next three routines used to be one big routine.
865 * To save cycles in the RFC 1323 implementation it was better to break
866 * it up into three procedures. -- erics
868 static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
870 struct tcp_sock *tp = tcp_sk(sk);
871 long m = mrtt_us; /* RTT */
872 u32 srtt = tp->srtt_us;
874 /* The following amusing code comes from Jacobson's
875 * article in SIGCOMM '88. Note that rtt and mdev
876 * are scaled versions of rtt and mean deviation.
877 * This is designed to be as fast as possible
878 * m stands for "measurement".
880 * On a 1990 paper the rto value is changed to:
881 * RTO = rtt + 4 * mdev
883 * Funny. This algorithm seems to be very broken.
884 * These formulae increase RTO, when it should be decreased, increase
885 * too slowly, when it should be increased quickly, decrease too quickly
886 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
887 * does not matter how to _calculate_ it. Seems, it was trap
888 * that VJ failed to avoid. 8)
891 m -= (srtt >> 3); /* m is now error in rtt est */
892 srtt += m; /* rtt = 7/8 rtt + 1/8 new */
894 m = -m; /* m is now abs(error) */
895 m -= (tp->mdev_us >> 2); /* similar update on mdev */
896 /* This is similar to one of Eifel findings.
897 * Eifel blocks mdev updates when rtt decreases.
898 * This solution is a bit different: we use finer gain
899 * for mdev in this case (alpha*beta).
900 * Like Eifel it also prevents growth of rto,
901 * but also it limits too fast rto decreases,
902 * happening in pure Eifel.
907 m -= (tp->mdev_us >> 2); /* similar update on mdev */
909 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */
910 if (tp->mdev_us > tp->mdev_max_us) {
911 tp->mdev_max_us = tp->mdev_us;
912 if (tp->mdev_max_us > tp->rttvar_us)
913 tp->rttvar_us = tp->mdev_max_us;
915 if (after(tp->snd_una, tp->rtt_seq)) {
916 if (tp->mdev_max_us < tp->rttvar_us)
917 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
918 tp->rtt_seq = tp->snd_nxt;
919 tp->mdev_max_us = tcp_rto_min_us(sk);
924 /* no previous measure. */
925 srtt = m << 3; /* take the measured time to be rtt */
926 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */
927 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
928 tp->mdev_max_us = tp->rttvar_us;
929 tp->rtt_seq = tp->snd_nxt;
933 tp->srtt_us = max(1U, srtt);
936 static void tcp_update_pacing_rate(struct sock *sk)
938 const struct tcp_sock *tp = tcp_sk(sk);
941 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
942 rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3);
944 /* current rate is (cwnd * mss) / srtt
945 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
946 * In Congestion Avoidance phase, set it to 120 % the current rate.
948 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
949 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
950 * end of slow start and should slow down.
952 if (tcp_snd_cwnd(tp) < tp->snd_ssthresh / 2)
953 rate *= READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_pacing_ss_ratio);
955 rate *= READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_pacing_ca_ratio);
957 rate *= max(tcp_snd_cwnd(tp), tp->packets_out);
959 if (likely(tp->srtt_us))
960 do_div(rate, tp->srtt_us);
962 /* WRITE_ONCE() is needed because sch_fq fetches sk_pacing_rate
963 * without any lock. We want to make sure compiler wont store
964 * intermediate values in this location.
966 WRITE_ONCE(sk->sk_pacing_rate,
967 min_t(u64, rate, READ_ONCE(sk->sk_max_pacing_rate)));
970 /* Calculate rto without backoff. This is the second half of Van Jacobson's
971 * routine referred to above.
973 static void tcp_set_rto(struct sock *sk)
975 const struct tcp_sock *tp = tcp_sk(sk);
976 /* Old crap is replaced with new one. 8)
979 * 1. If rtt variance happened to be less 50msec, it is hallucination.
980 * It cannot be less due to utterly erratic ACK generation made
981 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
982 * to do with delayed acks, because at cwnd>2 true delack timeout
983 * is invisible. Actually, Linux-2.4 also generates erratic
984 * ACKs in some circumstances.
986 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
988 /* 2. Fixups made earlier cannot be right.
989 * If we do not estimate RTO correctly without them,
990 * all the algo is pure shit and should be replaced
991 * with correct one. It is exactly, which we pretend to do.
994 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
995 * guarantees that rto is higher.
1000 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
1002 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
1005 cwnd = TCP_INIT_CWND;
1006 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
1009 struct tcp_sacktag_state {
1010 /* Timestamps for earliest and latest never-retransmitted segment
1011 * that was SACKed. RTO needs the earliest RTT to stay conservative,
1012 * but congestion control should still get an accurate delay signal.
1019 unsigned int mss_now;
1020 struct rate_sample *rate;
1023 /* Take a notice that peer is sending D-SACKs. Skip update of data delivery
1024 * and spurious retransmission information if this DSACK is unlikely caused by
1026 * - DSACKed sequence range is larger than maximum receiver's window.
1027 * - Total no. of DSACKed segments exceed the total no. of retransmitted segs.
1029 static u32 tcp_dsack_seen(struct tcp_sock *tp, u32 start_seq,
1030 u32 end_seq, struct tcp_sacktag_state *state)
1032 u32 seq_len, dup_segs = 1;
1034 if (!before(start_seq, end_seq))
1037 seq_len = end_seq - start_seq;
1038 /* Dubious DSACK: DSACKed range greater than maximum advertised rwnd */
1039 if (seq_len > tp->max_window)
1041 if (seq_len > tp->mss_cache)
1042 dup_segs = DIV_ROUND_UP(seq_len, tp->mss_cache);
1043 else if (tp->tlp_high_seq && tp->tlp_high_seq == end_seq)
1044 state->flag |= FLAG_DSACK_TLP;
1046 tp->dsack_dups += dup_segs;
1047 /* Skip the DSACK if dup segs weren't retransmitted by sender */
1048 if (tp->dsack_dups > tp->total_retrans)
1051 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
1052 /* We increase the RACK ordering window in rounds where we receive
1053 * DSACKs that may have been due to reordering causing RACK to trigger
1054 * a spurious fast recovery. Thus RACK ignores DSACKs that happen
1055 * without having seen reordering, or that match TLP probes (TLP
1056 * is timer-driven, not triggered by RACK).
1058 if (tp->reord_seen && !(state->flag & FLAG_DSACK_TLP))
1059 tp->rack.dsack_seen = 1;
1061 state->flag |= FLAG_DSACKING_ACK;
1062 /* A spurious retransmission is delivered */
1063 state->sack_delivered += dup_segs;
1068 /* It's reordering when higher sequence was delivered (i.e. sacked) before
1069 * some lower never-retransmitted sequence ("low_seq"). The maximum reordering
1070 * distance is approximated in full-mss packet distance ("reordering").
1072 static void tcp_check_sack_reordering(struct sock *sk, const u32 low_seq,
1075 struct tcp_sock *tp = tcp_sk(sk);
1076 const u32 mss = tp->mss_cache;
1079 fack = tcp_highest_sack_seq(tp);
1080 if (!before(low_seq, fack))
1083 metric = fack - low_seq;
1084 if ((metric > tp->reordering * mss) && mss) {
1085 #if FASTRETRANS_DEBUG > 1
1086 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
1087 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
1091 tp->undo_marker ? tp->undo_retrans : 0);
1093 tp->reordering = min_t(u32, (metric + mss - 1) / mss,
1094 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_max_reordering));
1097 /* This exciting event is worth to be remembered. 8) */
1099 NET_INC_STATS(sock_net(sk),
1100 ts ? LINUX_MIB_TCPTSREORDER : LINUX_MIB_TCPSACKREORDER);
1103 /* This must be called before lost_out or retrans_out are updated
1104 * on a new loss, because we want to know if all skbs previously
1105 * known to be lost have already been retransmitted, indicating
1106 * that this newly lost skb is our next skb to retransmit.
1108 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
1110 if ((!tp->retransmit_skb_hint && tp->retrans_out >= tp->lost_out) ||
1111 (tp->retransmit_skb_hint &&
1112 before(TCP_SKB_CB(skb)->seq,
1113 TCP_SKB_CB(tp->retransmit_skb_hint)->seq)))
1114 tp->retransmit_skb_hint = skb;
1117 /* Sum the number of packets on the wire we have marked as lost, and
1118 * notify the congestion control module that the given skb was marked lost.
1120 static void tcp_notify_skb_loss_event(struct tcp_sock *tp, const struct sk_buff *skb)
1122 tp->lost += tcp_skb_pcount(skb);
1125 void tcp_mark_skb_lost(struct sock *sk, struct sk_buff *skb)
1127 __u8 sacked = TCP_SKB_CB(skb)->sacked;
1128 struct tcp_sock *tp = tcp_sk(sk);
1130 if (sacked & TCPCB_SACKED_ACKED)
1133 tcp_verify_retransmit_hint(tp, skb);
1134 if (sacked & TCPCB_LOST) {
1135 if (sacked & TCPCB_SACKED_RETRANS) {
1136 /* Account for retransmits that are lost again */
1137 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1138 tp->retrans_out -= tcp_skb_pcount(skb);
1139 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT,
1140 tcp_skb_pcount(skb));
1141 tcp_notify_skb_loss_event(tp, skb);
1144 tp->lost_out += tcp_skb_pcount(skb);
1145 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1146 tcp_notify_skb_loss_event(tp, skb);
1150 /* Updates the delivered and delivered_ce counts */
1151 static void tcp_count_delivered(struct tcp_sock *tp, u32 delivered,
1154 tp->delivered += delivered;
1156 tp->delivered_ce += delivered;
1159 /* This procedure tags the retransmission queue when SACKs arrive.
1161 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
1162 * Packets in queue with these bits set are counted in variables
1163 * sacked_out, retrans_out and lost_out, correspondingly.
1165 * Valid combinations are:
1166 * Tag InFlight Description
1167 * 0 1 - orig segment is in flight.
1168 * S 0 - nothing flies, orig reached receiver.
1169 * L 0 - nothing flies, orig lost by net.
1170 * R 2 - both orig and retransmit are in flight.
1171 * L|R 1 - orig is lost, retransmit is in flight.
1172 * S|R 1 - orig reached receiver, retrans is still in flight.
1173 * (L|S|R is logically valid, it could occur when L|R is sacked,
1174 * but it is equivalent to plain S and code short-curcuits it to S.
1175 * L|S is logically invalid, it would mean -1 packet in flight 8))
1177 * These 6 states form finite state machine, controlled by the following events:
1178 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
1179 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
1180 * 3. Loss detection event of two flavors:
1181 * A. Scoreboard estimator decided the packet is lost.
1182 * A'. Reno "three dupacks" marks head of queue lost.
1183 * B. SACK arrives sacking SND.NXT at the moment, when the
1184 * segment was retransmitted.
1185 * 4. D-SACK added new rule: D-SACK changes any tag to S.
1187 * It is pleasant to note, that state diagram turns out to be commutative,
1188 * so that we are allowed not to be bothered by order of our actions,
1189 * when multiple events arrive simultaneously. (see the function below).
1191 * Reordering detection.
1192 * --------------------
1193 * Reordering metric is maximal distance, which a packet can be displaced
1194 * in packet stream. With SACKs we can estimate it:
1196 * 1. SACK fills old hole and the corresponding segment was not
1197 * ever retransmitted -> reordering. Alas, we cannot use it
1198 * when segment was retransmitted.
1199 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1200 * for retransmitted and already SACKed segment -> reordering..
1201 * Both of these heuristics are not used in Loss state, when we cannot
1202 * account for retransmits accurately.
1204 * SACK block validation.
1205 * ----------------------
1207 * SACK block range validation checks that the received SACK block fits to
1208 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1209 * Note that SND.UNA is not included to the range though being valid because
1210 * it means that the receiver is rather inconsistent with itself reporting
1211 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1212 * perfectly valid, however, in light of RFC2018 which explicitly states
1213 * that "SACK block MUST reflect the newest segment. Even if the newest
1214 * segment is going to be discarded ...", not that it looks very clever
1215 * in case of head skb. Due to potentional receiver driven attacks, we
1216 * choose to avoid immediate execution of a walk in write queue due to
1217 * reneging and defer head skb's loss recovery to standard loss recovery
1218 * procedure that will eventually trigger (nothing forbids us doing this).
1220 * Implements also blockage to start_seq wrap-around. Problem lies in the
1221 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1222 * there's no guarantee that it will be before snd_nxt (n). The problem
1223 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1226 * <- outs wnd -> <- wrapzone ->
1227 * u e n u_w e_w s n_w
1229 * |<------------+------+----- TCP seqno space --------------+---------->|
1230 * ...-- <2^31 ->| |<--------...
1231 * ...---- >2^31 ------>| |<--------...
1233 * Current code wouldn't be vulnerable but it's better still to discard such
1234 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1235 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1236 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1237 * equal to the ideal case (infinite seqno space without wrap caused issues).
1239 * With D-SACK the lower bound is extended to cover sequence space below
1240 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1241 * again, D-SACK block must not to go across snd_una (for the same reason as
1242 * for the normal SACK blocks, explained above). But there all simplicity
1243 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1244 * fully below undo_marker they do not affect behavior in anyway and can
1245 * therefore be safely ignored. In rare cases (which are more or less
1246 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1247 * fragmentation and packet reordering past skb's retransmission. To consider
1248 * them correctly, the acceptable range must be extended even more though
1249 * the exact amount is rather hard to quantify. However, tp->max_window can
1250 * be used as an exaggerated estimate.
1252 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
1253 u32 start_seq, u32 end_seq)
1255 /* Too far in future, or reversed (interpretation is ambiguous) */
1256 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1259 /* Nasty start_seq wrap-around check (see comments above) */
1260 if (!before(start_seq, tp->snd_nxt))
1263 /* In outstanding window? ...This is valid exit for D-SACKs too.
1264 * start_seq == snd_una is non-sensical (see comments above)
1266 if (after(start_seq, tp->snd_una))
1269 if (!is_dsack || !tp->undo_marker)
1272 /* ...Then it's D-SACK, and must reside below snd_una completely */
1273 if (after(end_seq, tp->snd_una))
1276 if (!before(start_seq, tp->undo_marker))
1280 if (!after(end_seq, tp->undo_marker))
1283 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1284 * start_seq < undo_marker and end_seq >= undo_marker.
1286 return !before(start_seq, end_seq - tp->max_window);
1289 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1290 struct tcp_sack_block_wire *sp, int num_sacks,
1291 u32 prior_snd_una, struct tcp_sacktag_state *state)
1293 struct tcp_sock *tp = tcp_sk(sk);
1294 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1295 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1298 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1299 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1300 } else if (num_sacks > 1) {
1301 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1302 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1304 if (after(end_seq_0, end_seq_1) || before(start_seq_0, start_seq_1))
1306 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKOFORECV);
1311 dup_segs = tcp_dsack_seen(tp, start_seq_0, end_seq_0, state);
1312 if (!dup_segs) { /* Skip dubious DSACK */
1313 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKIGNOREDDUBIOUS);
1317 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECVSEGS, dup_segs);
1319 /* D-SACK for already forgotten data... Do dumb counting. */
1320 if (tp->undo_marker && tp->undo_retrans > 0 &&
1321 !after(end_seq_0, prior_snd_una) &&
1322 after(end_seq_0, tp->undo_marker))
1323 tp->undo_retrans = max_t(int, 0, tp->undo_retrans - dup_segs);
1328 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1329 * the incoming SACK may not exactly match but we can find smaller MSS
1330 * aligned portion of it that matches. Therefore we might need to fragment
1331 * which may fail and creates some hassle (caller must handle error case
1334 * FIXME: this could be merged to shift decision code
1336 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1337 u32 start_seq, u32 end_seq)
1341 unsigned int pkt_len;
1344 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1345 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1347 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1348 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1349 mss = tcp_skb_mss(skb);
1350 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1353 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1357 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1362 /* Round if necessary so that SACKs cover only full MSSes
1363 * and/or the remaining small portion (if present)
1365 if (pkt_len > mss) {
1366 unsigned int new_len = (pkt_len / mss) * mss;
1367 if (!in_sack && new_len < pkt_len)
1372 if (pkt_len >= skb->len && !in_sack)
1375 err = tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
1376 pkt_len, mss, GFP_ATOMIC);
1384 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1385 static u8 tcp_sacktag_one(struct sock *sk,
1386 struct tcp_sacktag_state *state, u8 sacked,
1387 u32 start_seq, u32 end_seq,
1388 int dup_sack, int pcount,
1391 struct tcp_sock *tp = tcp_sk(sk);
1393 /* Account D-SACK for retransmitted packet. */
1394 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1395 if (tp->undo_marker && tp->undo_retrans > 0 &&
1396 after(end_seq, tp->undo_marker))
1397 tp->undo_retrans = max_t(int, 0, tp->undo_retrans - pcount);
1398 if ((sacked & TCPCB_SACKED_ACKED) &&
1399 before(start_seq, state->reord))
1400 state->reord = start_seq;
1403 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1404 if (!after(end_seq, tp->snd_una))
1407 if (!(sacked & TCPCB_SACKED_ACKED)) {
1408 tcp_rack_advance(tp, sacked, end_seq, xmit_time);
1410 if (sacked & TCPCB_SACKED_RETRANS) {
1411 /* If the segment is not tagged as lost,
1412 * we do not clear RETRANS, believing
1413 * that retransmission is still in flight.
1415 if (sacked & TCPCB_LOST) {
1416 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1417 tp->lost_out -= pcount;
1418 tp->retrans_out -= pcount;
1421 if (!(sacked & TCPCB_RETRANS)) {
1422 /* New sack for not retransmitted frame,
1423 * which was in hole. It is reordering.
1425 if (before(start_seq,
1426 tcp_highest_sack_seq(tp)) &&
1427 before(start_seq, state->reord))
1428 state->reord = start_seq;
1430 if (!after(end_seq, tp->high_seq))
1431 state->flag |= FLAG_ORIG_SACK_ACKED;
1432 if (state->first_sackt == 0)
1433 state->first_sackt = xmit_time;
1434 state->last_sackt = xmit_time;
1437 if (sacked & TCPCB_LOST) {
1438 sacked &= ~TCPCB_LOST;
1439 tp->lost_out -= pcount;
1443 sacked |= TCPCB_SACKED_ACKED;
1444 state->flag |= FLAG_DATA_SACKED;
1445 tp->sacked_out += pcount;
1446 /* Out-of-order packets delivered */
1447 state->sack_delivered += pcount;
1449 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1450 if (tp->lost_skb_hint &&
1451 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1452 tp->lost_cnt_hint += pcount;
1455 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1456 * frames and clear it. undo_retrans is decreased above, L|R frames
1457 * are accounted above as well.
1459 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1460 sacked &= ~TCPCB_SACKED_RETRANS;
1461 tp->retrans_out -= pcount;
1467 /* Shift newly-SACKed bytes from this skb to the immediately previous
1468 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1470 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *prev,
1471 struct sk_buff *skb,
1472 struct tcp_sacktag_state *state,
1473 unsigned int pcount, int shifted, int mss,
1476 struct tcp_sock *tp = tcp_sk(sk);
1477 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1478 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
1482 /* Adjust counters and hints for the newly sacked sequence
1483 * range but discard the return value since prev is already
1484 * marked. We must tag the range first because the seq
1485 * advancement below implicitly advances
1486 * tcp_highest_sack_seq() when skb is highest_sack.
1488 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1489 start_seq, end_seq, dup_sack, pcount,
1490 tcp_skb_timestamp_us(skb));
1491 tcp_rate_skb_delivered(sk, skb, state->rate);
1493 if (skb == tp->lost_skb_hint)
1494 tp->lost_cnt_hint += pcount;
1496 TCP_SKB_CB(prev)->end_seq += shifted;
1497 TCP_SKB_CB(skb)->seq += shifted;
1499 tcp_skb_pcount_add(prev, pcount);
1500 WARN_ON_ONCE(tcp_skb_pcount(skb) < pcount);
1501 tcp_skb_pcount_add(skb, -pcount);
1503 /* When we're adding to gso_segs == 1, gso_size will be zero,
1504 * in theory this shouldn't be necessary but as long as DSACK
1505 * code can come after this skb later on it's better to keep
1506 * setting gso_size to something.
1508 if (!TCP_SKB_CB(prev)->tcp_gso_size)
1509 TCP_SKB_CB(prev)->tcp_gso_size = mss;
1511 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1512 if (tcp_skb_pcount(skb) <= 1)
1513 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1515 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1516 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1519 BUG_ON(!tcp_skb_pcount(skb));
1520 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1524 /* Whole SKB was eaten :-) */
1526 if (skb == tp->retransmit_skb_hint)
1527 tp->retransmit_skb_hint = prev;
1528 if (skb == tp->lost_skb_hint) {
1529 tp->lost_skb_hint = prev;
1530 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1533 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1534 TCP_SKB_CB(prev)->eor = TCP_SKB_CB(skb)->eor;
1535 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1536 TCP_SKB_CB(prev)->end_seq++;
1538 if (skb == tcp_highest_sack(sk))
1539 tcp_advance_highest_sack(sk, skb);
1541 tcp_skb_collapse_tstamp(prev, skb);
1542 if (unlikely(TCP_SKB_CB(prev)->tx.delivered_mstamp))
1543 TCP_SKB_CB(prev)->tx.delivered_mstamp = 0;
1545 tcp_rtx_queue_unlink_and_free(skb, sk);
1547 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKMERGED);
1552 /* I wish gso_size would have a bit more sane initialization than
1553 * something-or-zero which complicates things
1555 static int tcp_skb_seglen(const struct sk_buff *skb)
1557 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1560 /* Shifting pages past head area doesn't work */
1561 static int skb_can_shift(const struct sk_buff *skb)
1563 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1566 int tcp_skb_shift(struct sk_buff *to, struct sk_buff *from,
1567 int pcount, int shiftlen)
1569 /* TCP min gso_size is 8 bytes (TCP_MIN_GSO_SIZE)
1570 * Since TCP_SKB_CB(skb)->tcp_gso_segs is 16 bits, we need
1571 * to make sure not storing more than 65535 * 8 bytes per skb,
1572 * even if current MSS is bigger.
1574 if (unlikely(to->len + shiftlen >= 65535 * TCP_MIN_GSO_SIZE))
1576 if (unlikely(tcp_skb_pcount(to) + pcount > 65535))
1578 return skb_shift(to, from, shiftlen);
1581 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1584 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1585 struct tcp_sacktag_state *state,
1586 u32 start_seq, u32 end_seq,
1589 struct tcp_sock *tp = tcp_sk(sk);
1590 struct sk_buff *prev;
1596 /* Normally R but no L won't result in plain S */
1598 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1600 if (!skb_can_shift(skb))
1602 /* This frame is about to be dropped (was ACKed). */
1603 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1606 /* Can only happen with delayed DSACK + discard craziness */
1607 prev = skb_rb_prev(skb);
1611 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1614 if (!tcp_skb_can_collapse(prev, skb))
1617 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1618 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1622 pcount = tcp_skb_pcount(skb);
1623 mss = tcp_skb_seglen(skb);
1625 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1626 * drop this restriction as unnecessary
1628 if (mss != tcp_skb_seglen(prev))
1631 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1633 /* CHECKME: This is non-MSS split case only?, this will
1634 * cause skipped skbs due to advancing loop btw, original
1635 * has that feature too
1637 if (tcp_skb_pcount(skb) <= 1)
1640 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1642 /* TODO: head merge to next could be attempted here
1643 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1644 * though it might not be worth of the additional hassle
1646 * ...we can probably just fallback to what was done
1647 * previously. We could try merging non-SACKed ones
1648 * as well but it probably isn't going to buy off
1649 * because later SACKs might again split them, and
1650 * it would make skb timestamp tracking considerably
1656 len = end_seq - TCP_SKB_CB(skb)->seq;
1658 BUG_ON(len > skb->len);
1660 /* MSS boundaries should be honoured or else pcount will
1661 * severely break even though it makes things bit trickier.
1662 * Optimize common case to avoid most of the divides
1664 mss = tcp_skb_mss(skb);
1666 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1667 * drop this restriction as unnecessary
1669 if (mss != tcp_skb_seglen(prev))
1674 } else if (len < mss) {
1682 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1683 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1686 if (!tcp_skb_shift(prev, skb, pcount, len))
1688 if (!tcp_shifted_skb(sk, prev, skb, state, pcount, len, mss, dup_sack))
1691 /* Hole filled allows collapsing with the next as well, this is very
1692 * useful when hole on every nth skb pattern happens
1694 skb = skb_rb_next(prev);
1698 if (!skb_can_shift(skb) ||
1699 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1700 (mss != tcp_skb_seglen(skb)))
1703 if (!tcp_skb_can_collapse(prev, skb))
1706 pcount = tcp_skb_pcount(skb);
1707 if (tcp_skb_shift(prev, skb, pcount, len))
1708 tcp_shifted_skb(sk, prev, skb, state, pcount,
1718 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1722 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1723 struct tcp_sack_block *next_dup,
1724 struct tcp_sacktag_state *state,
1725 u32 start_seq, u32 end_seq,
1728 struct tcp_sock *tp = tcp_sk(sk);
1729 struct sk_buff *tmp;
1731 skb_rbtree_walk_from(skb) {
1733 bool dup_sack = dup_sack_in;
1735 /* queue is in-order => we can short-circuit the walk early */
1736 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1740 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1741 in_sack = tcp_match_skb_to_sack(sk, skb,
1742 next_dup->start_seq,
1748 /* skb reference here is a bit tricky to get right, since
1749 * shifting can eat and free both this skb and the next,
1750 * so not even _safe variant of the loop is enough.
1753 tmp = tcp_shift_skb_data(sk, skb, state,
1754 start_seq, end_seq, dup_sack);
1763 in_sack = tcp_match_skb_to_sack(sk, skb,
1769 if (unlikely(in_sack < 0))
1773 TCP_SKB_CB(skb)->sacked =
1776 TCP_SKB_CB(skb)->sacked,
1777 TCP_SKB_CB(skb)->seq,
1778 TCP_SKB_CB(skb)->end_seq,
1780 tcp_skb_pcount(skb),
1781 tcp_skb_timestamp_us(skb));
1782 tcp_rate_skb_delivered(sk, skb, state->rate);
1783 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1784 list_del_init(&skb->tcp_tsorted_anchor);
1786 if (!before(TCP_SKB_CB(skb)->seq,
1787 tcp_highest_sack_seq(tp)))
1788 tcp_advance_highest_sack(sk, skb);
1794 static struct sk_buff *tcp_sacktag_bsearch(struct sock *sk, u32 seq)
1796 struct rb_node *parent, **p = &sk->tcp_rtx_queue.rb_node;
1797 struct sk_buff *skb;
1801 skb = rb_to_skb(parent);
1802 if (before(seq, TCP_SKB_CB(skb)->seq)) {
1803 p = &parent->rb_left;
1806 if (!before(seq, TCP_SKB_CB(skb)->end_seq)) {
1807 p = &parent->rb_right;
1815 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1818 if (skb && after(TCP_SKB_CB(skb)->seq, skip_to_seq))
1821 return tcp_sacktag_bsearch(sk, skip_to_seq);
1824 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1826 struct tcp_sack_block *next_dup,
1827 struct tcp_sacktag_state *state,
1833 if (before(next_dup->start_seq, skip_to_seq)) {
1834 skb = tcp_sacktag_skip(skb, sk, next_dup->start_seq);
1835 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1836 next_dup->start_seq, next_dup->end_seq,
1843 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1845 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1849 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1850 u32 prior_snd_una, struct tcp_sacktag_state *state)
1852 struct tcp_sock *tp = tcp_sk(sk);
1853 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1854 TCP_SKB_CB(ack_skb)->sacked);
1855 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1856 struct tcp_sack_block sp[TCP_NUM_SACKS];
1857 struct tcp_sack_block *cache;
1858 struct sk_buff *skb;
1859 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1861 bool found_dup_sack = false;
1863 int first_sack_index;
1866 state->reord = tp->snd_nxt;
1868 if (!tp->sacked_out)
1869 tcp_highest_sack_reset(sk);
1871 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1872 num_sacks, prior_snd_una, state);
1874 /* Eliminate too old ACKs, but take into
1875 * account more or less fresh ones, they can
1876 * contain valid SACK info.
1878 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1881 if (!tp->packets_out)
1885 first_sack_index = 0;
1886 for (i = 0; i < num_sacks; i++) {
1887 bool dup_sack = !i && found_dup_sack;
1889 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1890 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1892 if (!tcp_is_sackblock_valid(tp, dup_sack,
1893 sp[used_sacks].start_seq,
1894 sp[used_sacks].end_seq)) {
1898 if (!tp->undo_marker)
1899 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1901 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1903 /* Don't count olds caused by ACK reordering */
1904 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1905 !after(sp[used_sacks].end_seq, tp->snd_una))
1907 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1910 NET_INC_STATS(sock_net(sk), mib_idx);
1912 first_sack_index = -1;
1916 /* Ignore very old stuff early */
1917 if (!after(sp[used_sacks].end_seq, prior_snd_una)) {
1919 first_sack_index = -1;
1926 /* order SACK blocks to allow in order walk of the retrans queue */
1927 for (i = used_sacks - 1; i > 0; i--) {
1928 for (j = 0; j < i; j++) {
1929 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1930 swap(sp[j], sp[j + 1]);
1932 /* Track where the first SACK block goes to */
1933 if (j == first_sack_index)
1934 first_sack_index = j + 1;
1939 state->mss_now = tcp_current_mss(sk);
1943 if (!tp->sacked_out) {
1944 /* It's already past, so skip checking against it */
1945 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1947 cache = tp->recv_sack_cache;
1948 /* Skip empty blocks in at head of the cache */
1949 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1954 while (i < used_sacks) {
1955 u32 start_seq = sp[i].start_seq;
1956 u32 end_seq = sp[i].end_seq;
1957 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1958 struct tcp_sack_block *next_dup = NULL;
1960 if (found_dup_sack && ((i + 1) == first_sack_index))
1961 next_dup = &sp[i + 1];
1963 /* Skip too early cached blocks */
1964 while (tcp_sack_cache_ok(tp, cache) &&
1965 !before(start_seq, cache->end_seq))
1968 /* Can skip some work by looking recv_sack_cache? */
1969 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1970 after(end_seq, cache->start_seq)) {
1973 if (before(start_seq, cache->start_seq)) {
1974 skb = tcp_sacktag_skip(skb, sk, start_seq);
1975 skb = tcp_sacktag_walk(skb, sk, next_dup,
1982 /* Rest of the block already fully processed? */
1983 if (!after(end_seq, cache->end_seq))
1986 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1990 /* ...tail remains todo... */
1991 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1992 /* ...but better entrypoint exists! */
1993 skb = tcp_highest_sack(sk);
2000 skb = tcp_sacktag_skip(skb, sk, cache->end_seq);
2001 /* Check overlap against next cached too (past this one already) */
2006 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
2007 skb = tcp_highest_sack(sk);
2011 skb = tcp_sacktag_skip(skb, sk, start_seq);
2014 skb = tcp_sacktag_walk(skb, sk, next_dup, state,
2015 start_seq, end_seq, dup_sack);
2021 /* Clear the head of the cache sack blocks so we can skip it next time */
2022 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
2023 tp->recv_sack_cache[i].start_seq = 0;
2024 tp->recv_sack_cache[i].end_seq = 0;
2026 for (j = 0; j < used_sacks; j++)
2027 tp->recv_sack_cache[i++] = sp[j];
2029 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Loss || tp->undo_marker)
2030 tcp_check_sack_reordering(sk, state->reord, 0);
2032 tcp_verify_left_out(tp);
2035 #if FASTRETRANS_DEBUG > 0
2036 WARN_ON((int)tp->sacked_out < 0);
2037 WARN_ON((int)tp->lost_out < 0);
2038 WARN_ON((int)tp->retrans_out < 0);
2039 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
2044 /* Limits sacked_out so that sum with lost_out isn't ever larger than
2045 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
2047 static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
2051 holes = max(tp->lost_out, 1U);
2052 holes = min(holes, tp->packets_out);
2054 if ((tp->sacked_out + holes) > tp->packets_out) {
2055 tp->sacked_out = tp->packets_out - holes;
2061 /* If we receive more dupacks than we expected counting segments
2062 * in assumption of absent reordering, interpret this as reordering.
2063 * The only another reason could be bug in receiver TCP.
2065 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
2067 struct tcp_sock *tp = tcp_sk(sk);
2069 if (!tcp_limit_reno_sacked(tp))
2072 tp->reordering = min_t(u32, tp->packets_out + addend,
2073 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_max_reordering));
2075 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRENOREORDER);
2078 /* Emulate SACKs for SACKless connection: account for a new dupack. */
2080 static void tcp_add_reno_sack(struct sock *sk, int num_dupack, bool ece_ack)
2083 struct tcp_sock *tp = tcp_sk(sk);
2084 u32 prior_sacked = tp->sacked_out;
2087 tp->sacked_out += num_dupack;
2088 tcp_check_reno_reordering(sk, 0);
2089 delivered = tp->sacked_out - prior_sacked;
2091 tcp_count_delivered(tp, delivered, ece_ack);
2092 tcp_verify_left_out(tp);
2096 /* Account for ACK, ACKing some data in Reno Recovery phase. */
2098 static void tcp_remove_reno_sacks(struct sock *sk, int acked, bool ece_ack)
2100 struct tcp_sock *tp = tcp_sk(sk);
2103 /* One ACK acked hole. The rest eat duplicate ACKs. */
2104 tcp_count_delivered(tp, max_t(int, acked - tp->sacked_out, 1),
2106 if (acked - 1 >= tp->sacked_out)
2109 tp->sacked_out -= acked - 1;
2111 tcp_check_reno_reordering(sk, acked);
2112 tcp_verify_left_out(tp);
2115 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
2120 void tcp_clear_retrans(struct tcp_sock *tp)
2122 tp->retrans_out = 0;
2124 tp->undo_marker = 0;
2125 tp->undo_retrans = -1;
2129 tp->total_rto_recoveries = 0;
2130 tp->total_rto_time = 0;
2133 static inline void tcp_init_undo(struct tcp_sock *tp)
2135 tp->undo_marker = tp->snd_una;
2136 /* Retransmission still in flight may cause DSACKs later. */
2137 tp->undo_retrans = tp->retrans_out ? : -1;
2140 static bool tcp_is_rack(const struct sock *sk)
2142 return READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_recovery) &
2143 TCP_RACK_LOSS_DETECTION;
2146 /* If we detect SACK reneging, forget all SACK information
2147 * and reset tags completely, otherwise preserve SACKs. If receiver
2148 * dropped its ofo queue, we will know this due to reneging detection.
2150 static void tcp_timeout_mark_lost(struct sock *sk)
2152 struct tcp_sock *tp = tcp_sk(sk);
2153 struct sk_buff *skb, *head;
2154 bool is_reneg; /* is receiver reneging on SACKs? */
2156 head = tcp_rtx_queue_head(sk);
2157 is_reneg = head && (TCP_SKB_CB(head)->sacked & TCPCB_SACKED_ACKED);
2159 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
2161 /* Mark SACK reneging until we recover from this loss event. */
2162 tp->is_sack_reneg = 1;
2163 } else if (tcp_is_reno(tp)) {
2164 tcp_reset_reno_sack(tp);
2168 skb_rbtree_walk_from(skb) {
2170 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
2171 else if (tcp_is_rack(sk) && skb != head &&
2172 tcp_rack_skb_timeout(tp, skb, 0) > 0)
2173 continue; /* Don't mark recently sent ones lost yet */
2174 tcp_mark_skb_lost(sk, skb);
2176 tcp_verify_left_out(tp);
2177 tcp_clear_all_retrans_hints(tp);
2180 /* Enter Loss state. */
2181 void tcp_enter_loss(struct sock *sk)
2183 const struct inet_connection_sock *icsk = inet_csk(sk);
2184 struct tcp_sock *tp = tcp_sk(sk);
2185 struct net *net = sock_net(sk);
2186 bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery;
2189 tcp_timeout_mark_lost(sk);
2191 /* Reduce ssthresh if it has not yet been made inside this window. */
2192 if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
2193 !after(tp->high_seq, tp->snd_una) ||
2194 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
2195 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2196 tp->prior_cwnd = tcp_snd_cwnd(tp);
2197 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2198 tcp_ca_event(sk, CA_EVENT_LOSS);
2201 tcp_snd_cwnd_set(tp, tcp_packets_in_flight(tp) + 1);
2202 tp->snd_cwnd_cnt = 0;
2203 tp->snd_cwnd_stamp = tcp_jiffies32;
2205 /* Timeout in disordered state after receiving substantial DUPACKs
2206 * suggests that the degree of reordering is over-estimated.
2208 reordering = READ_ONCE(net->ipv4.sysctl_tcp_reordering);
2209 if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
2210 tp->sacked_out >= reordering)
2211 tp->reordering = min_t(unsigned int, tp->reordering,
2214 tcp_set_ca_state(sk, TCP_CA_Loss);
2215 tp->high_seq = tp->snd_nxt;
2216 tcp_ecn_queue_cwr(tp);
2218 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
2219 * loss recovery is underway except recurring timeout(s) on
2220 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
2222 tp->frto = READ_ONCE(net->ipv4.sysctl_tcp_frto) &&
2223 (new_recovery || icsk->icsk_retransmits) &&
2224 !inet_csk(sk)->icsk_mtup.probe_size;
2227 /* If ACK arrived pointing to a remembered SACK, it means that our
2228 * remembered SACKs do not reflect real state of receiver i.e.
2229 * receiver _host_ is heavily congested (or buggy).
2231 * To avoid big spurious retransmission bursts due to transient SACK
2232 * scoreboard oddities that look like reneging, we give the receiver a
2233 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
2234 * restore sanity to the SACK scoreboard. If the apparent reneging
2235 * persists until this RTO then we'll clear the SACK scoreboard.
2237 static bool tcp_check_sack_reneging(struct sock *sk, int *ack_flag)
2239 if (*ack_flag & FLAG_SACK_RENEGING &&
2240 *ack_flag & FLAG_SND_UNA_ADVANCED) {
2241 struct tcp_sock *tp = tcp_sk(sk);
2242 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4),
2243 msecs_to_jiffies(10));
2245 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2246 delay, TCP_RTO_MAX);
2247 *ack_flag &= ~FLAG_SET_XMIT_TIMER;
2253 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2254 * counter when SACK is enabled (without SACK, sacked_out is used for
2257 * With reordering, holes may still be in flight, so RFC3517 recovery
2258 * uses pure sacked_out (total number of SACKed segments) even though
2259 * it violates the RFC that uses duplicate ACKs, often these are equal
2260 * but when e.g. out-of-window ACKs or packet duplication occurs,
2261 * they differ. Since neither occurs due to loss, TCP should really
2264 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2266 return tp->sacked_out + 1;
2269 /* Linux NewReno/SACK/ECN state machine.
2270 * --------------------------------------
2272 * "Open" Normal state, no dubious events, fast path.
2273 * "Disorder" In all the respects it is "Open",
2274 * but requires a bit more attention. It is entered when
2275 * we see some SACKs or dupacks. It is split of "Open"
2276 * mainly to move some processing from fast path to slow one.
2277 * "CWR" CWND was reduced due to some Congestion Notification event.
2278 * It can be ECN, ICMP source quench, local device congestion.
2279 * "Recovery" CWND was reduced, we are fast-retransmitting.
2280 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2282 * tcp_fastretrans_alert() is entered:
2283 * - each incoming ACK, if state is not "Open"
2284 * - when arrived ACK is unusual, namely:
2289 * Counting packets in flight is pretty simple.
2291 * in_flight = packets_out - left_out + retrans_out
2293 * packets_out is SND.NXT-SND.UNA counted in packets.
2295 * retrans_out is number of retransmitted segments.
2297 * left_out is number of segments left network, but not ACKed yet.
2299 * left_out = sacked_out + lost_out
2301 * sacked_out: Packets, which arrived to receiver out of order
2302 * and hence not ACKed. With SACKs this number is simply
2303 * amount of SACKed data. Even without SACKs
2304 * it is easy to give pretty reliable estimate of this number,
2305 * counting duplicate ACKs.
2307 * lost_out: Packets lost by network. TCP has no explicit
2308 * "loss notification" feedback from network (for now).
2309 * It means that this number can be only _guessed_.
2310 * Actually, it is the heuristics to predict lossage that
2311 * distinguishes different algorithms.
2313 * F.e. after RTO, when all the queue is considered as lost,
2314 * lost_out = packets_out and in_flight = retrans_out.
2316 * Essentially, we have now a few algorithms detecting
2319 * If the receiver supports SACK:
2321 * RFC6675/3517: It is the conventional algorithm. A packet is
2322 * considered lost if the number of higher sequence packets
2323 * SACKed is greater than or equal the DUPACK thoreshold
2324 * (reordering). This is implemented in tcp_mark_head_lost and
2325 * tcp_update_scoreboard.
2327 * RACK (draft-ietf-tcpm-rack-01): it is a newer algorithm
2328 * (2017-) that checks timing instead of counting DUPACKs.
2329 * Essentially a packet is considered lost if it's not S/ACKed
2330 * after RTT + reordering_window, where both metrics are
2331 * dynamically measured and adjusted. This is implemented in
2332 * tcp_rack_mark_lost.
2334 * If the receiver does not support SACK:
2336 * NewReno (RFC6582): in Recovery we assume that one segment
2337 * is lost (classic Reno). While we are in Recovery and
2338 * a partial ACK arrives, we assume that one more packet
2339 * is lost (NewReno). This heuristics are the same in NewReno
2342 * Really tricky (and requiring careful tuning) part of algorithm
2343 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2344 * The first determines the moment _when_ we should reduce CWND and,
2345 * hence, slow down forward transmission. In fact, it determines the moment
2346 * when we decide that hole is caused by loss, rather than by a reorder.
2348 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2349 * holes, caused by lost packets.
2351 * And the most logically complicated part of algorithm is undo
2352 * heuristics. We detect false retransmits due to both too early
2353 * fast retransmit (reordering) and underestimated RTO, analyzing
2354 * timestamps and D-SACKs. When we detect that some segments were
2355 * retransmitted by mistake and CWND reduction was wrong, we undo
2356 * window reduction and abort recovery phase. This logic is hidden
2357 * inside several functions named tcp_try_undo_<something>.
2360 /* This function decides, when we should leave Disordered state
2361 * and enter Recovery phase, reducing congestion window.
2363 * Main question: may we further continue forward transmission
2364 * with the same cwnd?
2366 static bool tcp_time_to_recover(struct sock *sk, int flag)
2368 struct tcp_sock *tp = tcp_sk(sk);
2370 /* Trick#1: The loss is proven. */
2374 /* Not-A-Trick#2 : Classic rule... */
2375 if (!tcp_is_rack(sk) && tcp_dupack_heuristics(tp) > tp->reordering)
2381 /* Detect loss in event "A" above by marking head of queue up as lost.
2382 * For RFC3517 SACK, a segment is considered lost if it
2383 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2384 * the maximum SACKed segments to pass before reaching this limit.
2386 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2388 struct tcp_sock *tp = tcp_sk(sk);
2389 struct sk_buff *skb;
2391 /* Use SACK to deduce losses of new sequences sent during recovery */
2392 const u32 loss_high = tp->snd_nxt;
2394 WARN_ON(packets > tp->packets_out);
2395 skb = tp->lost_skb_hint;
2397 /* Head already handled? */
2398 if (mark_head && after(TCP_SKB_CB(skb)->seq, tp->snd_una))
2400 cnt = tp->lost_cnt_hint;
2402 skb = tcp_rtx_queue_head(sk);
2406 skb_rbtree_walk_from(skb) {
2407 /* TODO: do this better */
2408 /* this is not the most efficient way to do this... */
2409 tp->lost_skb_hint = skb;
2410 tp->lost_cnt_hint = cnt;
2412 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2415 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2416 cnt += tcp_skb_pcount(skb);
2421 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_LOST))
2422 tcp_mark_skb_lost(sk, skb);
2427 tcp_verify_left_out(tp);
2430 /* Account newly detected lost packet(s) */
2432 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2434 struct tcp_sock *tp = tcp_sk(sk);
2436 if (tcp_is_sack(tp)) {
2437 int sacked_upto = tp->sacked_out - tp->reordering;
2438 if (sacked_upto >= 0)
2439 tcp_mark_head_lost(sk, sacked_upto, 0);
2440 else if (fast_rexmit)
2441 tcp_mark_head_lost(sk, 1, 1);
2445 static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when)
2447 return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2448 before(tp->rx_opt.rcv_tsecr, when);
2451 /* skb is spurious retransmitted if the returned timestamp echo
2452 * reply is prior to the skb transmission time
2454 static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp,
2455 const struct sk_buff *skb)
2457 return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) &&
2458 tcp_tsopt_ecr_before(tp, tcp_skb_timestamp_ts(tp->tcp_usec_ts, skb));
2461 /* Nothing was retransmitted or returned timestamp is less
2462 * than timestamp of the first retransmission.
2464 static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2466 return tp->retrans_stamp &&
2467 tcp_tsopt_ecr_before(tp, tp->retrans_stamp);
2470 /* Undo procedures. */
2472 /* We can clear retrans_stamp when there are no retransmissions in the
2473 * window. It would seem that it is trivially available for us in
2474 * tp->retrans_out, however, that kind of assumptions doesn't consider
2475 * what will happen if errors occur when sending retransmission for the
2476 * second time. ...It could the that such segment has only
2477 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2478 * the head skb is enough except for some reneging corner cases that
2479 * are not worth the effort.
2481 * Main reason for all this complexity is the fact that connection dying
2482 * time now depends on the validity of the retrans_stamp, in particular,
2483 * that successive retransmissions of a segment must not advance
2484 * retrans_stamp under any conditions.
2486 static bool tcp_any_retrans_done(const struct sock *sk)
2488 const struct tcp_sock *tp = tcp_sk(sk);
2489 struct sk_buff *skb;
2491 if (tp->retrans_out)
2494 skb = tcp_rtx_queue_head(sk);
2495 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2501 static void DBGUNDO(struct sock *sk, const char *msg)
2503 #if FASTRETRANS_DEBUG > 1
2504 struct tcp_sock *tp = tcp_sk(sk);
2505 struct inet_sock *inet = inet_sk(sk);
2507 if (sk->sk_family == AF_INET) {
2508 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2510 &inet->inet_daddr, ntohs(inet->inet_dport),
2511 tcp_snd_cwnd(tp), tcp_left_out(tp),
2512 tp->snd_ssthresh, tp->prior_ssthresh,
2515 #if IS_ENABLED(CONFIG_IPV6)
2516 else if (sk->sk_family == AF_INET6) {
2517 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2519 &sk->sk_v6_daddr, ntohs(inet->inet_dport),
2520 tcp_snd_cwnd(tp), tcp_left_out(tp),
2521 tp->snd_ssthresh, tp->prior_ssthresh,
2528 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
2530 struct tcp_sock *tp = tcp_sk(sk);
2533 struct sk_buff *skb;
2535 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2536 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2539 tcp_clear_all_retrans_hints(tp);
2542 if (tp->prior_ssthresh) {
2543 const struct inet_connection_sock *icsk = inet_csk(sk);
2545 tcp_snd_cwnd_set(tp, icsk->icsk_ca_ops->undo_cwnd(sk));
2547 if (tp->prior_ssthresh > tp->snd_ssthresh) {
2548 tp->snd_ssthresh = tp->prior_ssthresh;
2549 tcp_ecn_withdraw_cwr(tp);
2552 tp->snd_cwnd_stamp = tcp_jiffies32;
2553 tp->undo_marker = 0;
2554 tp->rack.advanced = 1; /* Force RACK to re-exam losses */
2557 static inline bool tcp_may_undo(const struct tcp_sock *tp)
2559 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2562 static bool tcp_is_non_sack_preventing_reopen(struct sock *sk)
2564 struct tcp_sock *tp = tcp_sk(sk);
2566 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2567 /* Hold old state until something *above* high_seq
2568 * is ACKed. For Reno it is MUST to prevent false
2569 * fast retransmits (RFC2582). SACK TCP is safe. */
2570 if (!tcp_any_retrans_done(sk))
2571 tp->retrans_stamp = 0;
2577 /* People celebrate: "We love our President!" */
2578 static bool tcp_try_undo_recovery(struct sock *sk)
2580 struct tcp_sock *tp = tcp_sk(sk);
2582 if (tcp_may_undo(tp)) {
2585 /* Happy end! We did not retransmit anything
2586 * or our original transmission succeeded.
2588 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2589 tcp_undo_cwnd_reduction(sk, false);
2590 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2591 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2593 mib_idx = LINUX_MIB_TCPFULLUNDO;
2595 NET_INC_STATS(sock_net(sk), mib_idx);
2596 } else if (tp->rack.reo_wnd_persist) {
2597 tp->rack.reo_wnd_persist--;
2599 if (tcp_is_non_sack_preventing_reopen(sk))
2601 tcp_set_ca_state(sk, TCP_CA_Open);
2602 tp->is_sack_reneg = 0;
2606 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2607 static bool tcp_try_undo_dsack(struct sock *sk)
2609 struct tcp_sock *tp = tcp_sk(sk);
2611 if (tp->undo_marker && !tp->undo_retrans) {
2612 tp->rack.reo_wnd_persist = min(TCP_RACK_RECOVERY_THRESH,
2613 tp->rack.reo_wnd_persist + 1);
2614 DBGUNDO(sk, "D-SACK");
2615 tcp_undo_cwnd_reduction(sk, false);
2616 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2622 /* Undo during loss recovery after partial ACK or using F-RTO. */
2623 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2625 struct tcp_sock *tp = tcp_sk(sk);
2627 if (frto_undo || tcp_may_undo(tp)) {
2628 tcp_undo_cwnd_reduction(sk, true);
2630 DBGUNDO(sk, "partial loss");
2631 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2633 NET_INC_STATS(sock_net(sk),
2634 LINUX_MIB_TCPSPURIOUSRTOS);
2635 inet_csk(sk)->icsk_retransmits = 0;
2636 if (tcp_is_non_sack_preventing_reopen(sk))
2638 if (frto_undo || tcp_is_sack(tp)) {
2639 tcp_set_ca_state(sk, TCP_CA_Open);
2640 tp->is_sack_reneg = 0;
2647 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
2648 * It computes the number of packets to send (sndcnt) based on packets newly
2650 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2651 * cwnd reductions across a full RTT.
2652 * 2) Otherwise PRR uses packet conservation to send as much as delivered.
2653 * But when SND_UNA is acked without further losses,
2654 * slow starts cwnd up to ssthresh to speed up the recovery.
2656 static void tcp_init_cwnd_reduction(struct sock *sk)
2658 struct tcp_sock *tp = tcp_sk(sk);
2660 tp->high_seq = tp->snd_nxt;
2661 tp->tlp_high_seq = 0;
2662 tp->snd_cwnd_cnt = 0;
2663 tp->prior_cwnd = tcp_snd_cwnd(tp);
2664 tp->prr_delivered = 0;
2666 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2667 tcp_ecn_queue_cwr(tp);
2670 void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked, int newly_lost, int flag)
2672 struct tcp_sock *tp = tcp_sk(sk);
2674 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2676 if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd))
2679 tp->prr_delivered += newly_acked_sacked;
2681 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2683 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2685 sndcnt = max_t(int, tp->prr_delivered - tp->prr_out,
2686 newly_acked_sacked);
2687 if (flag & FLAG_SND_UNA_ADVANCED && !newly_lost)
2689 sndcnt = min(delta, sndcnt);
2691 /* Force a fast retransmit upon entering fast recovery */
2692 sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1));
2693 tcp_snd_cwnd_set(tp, tcp_packets_in_flight(tp) + sndcnt);
2696 static inline void tcp_end_cwnd_reduction(struct sock *sk)
2698 struct tcp_sock *tp = tcp_sk(sk);
2700 if (inet_csk(sk)->icsk_ca_ops->cong_control)
2703 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2704 if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH &&
2705 (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR || tp->undo_marker)) {
2706 tcp_snd_cwnd_set(tp, tp->snd_ssthresh);
2707 tp->snd_cwnd_stamp = tcp_jiffies32;
2709 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2712 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2713 void tcp_enter_cwr(struct sock *sk)
2715 struct tcp_sock *tp = tcp_sk(sk);
2717 tp->prior_ssthresh = 0;
2718 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2719 tp->undo_marker = 0;
2720 tcp_init_cwnd_reduction(sk);
2721 tcp_set_ca_state(sk, TCP_CA_CWR);
2724 EXPORT_SYMBOL(tcp_enter_cwr);
2726 static void tcp_try_keep_open(struct sock *sk)
2728 struct tcp_sock *tp = tcp_sk(sk);
2729 int state = TCP_CA_Open;
2731 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2732 state = TCP_CA_Disorder;
2734 if (inet_csk(sk)->icsk_ca_state != state) {
2735 tcp_set_ca_state(sk, state);
2736 tp->high_seq = tp->snd_nxt;
2740 static void tcp_try_to_open(struct sock *sk, int flag)
2742 struct tcp_sock *tp = tcp_sk(sk);
2744 tcp_verify_left_out(tp);
2746 if (!tcp_any_retrans_done(sk))
2747 tp->retrans_stamp = 0;
2749 if (flag & FLAG_ECE)
2752 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2753 tcp_try_keep_open(sk);
2757 static void tcp_mtup_probe_failed(struct sock *sk)
2759 struct inet_connection_sock *icsk = inet_csk(sk);
2761 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2762 icsk->icsk_mtup.probe_size = 0;
2763 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPFAIL);
2766 static void tcp_mtup_probe_success(struct sock *sk)
2768 struct tcp_sock *tp = tcp_sk(sk);
2769 struct inet_connection_sock *icsk = inet_csk(sk);
2772 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2774 val = (u64)tcp_snd_cwnd(tp) * tcp_mss_to_mtu(sk, tp->mss_cache);
2775 do_div(val, icsk->icsk_mtup.probe_size);
2776 DEBUG_NET_WARN_ON_ONCE((u32)val != val);
2777 tcp_snd_cwnd_set(tp, max_t(u32, 1U, val));
2779 tp->snd_cwnd_cnt = 0;
2780 tp->snd_cwnd_stamp = tcp_jiffies32;
2781 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2783 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2784 icsk->icsk_mtup.probe_size = 0;
2785 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2786 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS);
2789 /* Do a simple retransmit without using the backoff mechanisms in
2790 * tcp_timer. This is used for path mtu discovery.
2791 * The socket is already locked here.
2793 void tcp_simple_retransmit(struct sock *sk)
2795 const struct inet_connection_sock *icsk = inet_csk(sk);
2796 struct tcp_sock *tp = tcp_sk(sk);
2797 struct sk_buff *skb;
2800 /* A fastopen SYN request is stored as two separate packets within
2801 * the retransmit queue, this is done by tcp_send_syn_data().
2802 * As a result simply checking the MSS of the frames in the queue
2803 * will not work for the SYN packet.
2805 * Us being here is an indication of a path MTU issue so we can
2806 * assume that the fastopen SYN was lost and just mark all the
2807 * frames in the retransmit queue as lost. We will use an MSS of
2808 * -1 to mark all frames as lost, otherwise compute the current MSS.
2810 if (tp->syn_data && sk->sk_state == TCP_SYN_SENT)
2813 mss = tcp_current_mss(sk);
2815 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2816 if (tcp_skb_seglen(skb) > mss)
2817 tcp_mark_skb_lost(sk, skb);
2820 tcp_clear_retrans_hints_partial(tp);
2825 if (tcp_is_reno(tp))
2826 tcp_limit_reno_sacked(tp);
2828 tcp_verify_left_out(tp);
2830 /* Don't muck with the congestion window here.
2831 * Reason is that we do not increase amount of _data_
2832 * in network, but units changed and effective
2833 * cwnd/ssthresh really reduced now.
2835 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2836 tp->high_seq = tp->snd_nxt;
2837 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2838 tp->prior_ssthresh = 0;
2839 tp->undo_marker = 0;
2840 tcp_set_ca_state(sk, TCP_CA_Loss);
2842 tcp_xmit_retransmit_queue(sk);
2844 EXPORT_SYMBOL(tcp_simple_retransmit);
2846 void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2848 struct tcp_sock *tp = tcp_sk(sk);
2851 if (tcp_is_reno(tp))
2852 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2854 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2856 NET_INC_STATS(sock_net(sk), mib_idx);
2858 tp->prior_ssthresh = 0;
2861 if (!tcp_in_cwnd_reduction(sk)) {
2863 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2864 tcp_init_cwnd_reduction(sk);
2866 tcp_set_ca_state(sk, TCP_CA_Recovery);
2869 static void tcp_update_rto_time(struct tcp_sock *tp)
2871 if (tp->rto_stamp) {
2872 tp->total_rto_time += tcp_time_stamp_ms(tp) - tp->rto_stamp;
2877 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2878 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2880 static void tcp_process_loss(struct sock *sk, int flag, int num_dupack,
2883 struct tcp_sock *tp = tcp_sk(sk);
2884 bool recovered = !before(tp->snd_una, tp->high_seq);
2886 if ((flag & FLAG_SND_UNA_ADVANCED || rcu_access_pointer(tp->fastopen_rsk)) &&
2887 tcp_try_undo_loss(sk, false))
2890 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2891 /* Step 3.b. A timeout is spurious if not all data are
2892 * lost, i.e., never-retransmitted data are (s)acked.
2894 if ((flag & FLAG_ORIG_SACK_ACKED) &&
2895 tcp_try_undo_loss(sk, true))
2898 if (after(tp->snd_nxt, tp->high_seq)) {
2899 if (flag & FLAG_DATA_SACKED || num_dupack)
2900 tp->frto = 0; /* Step 3.a. loss was real */
2901 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2902 tp->high_seq = tp->snd_nxt;
2903 /* Step 2.b. Try send new data (but deferred until cwnd
2904 * is updated in tcp_ack()). Otherwise fall back to
2905 * the conventional recovery.
2907 if (!tcp_write_queue_empty(sk) &&
2908 after(tcp_wnd_end(tp), tp->snd_nxt)) {
2909 *rexmit = REXMIT_NEW;
2917 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2918 tcp_try_undo_recovery(sk);
2921 if (tcp_is_reno(tp)) {
2922 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2923 * delivered. Lower inflight to clock out (re)transmissions.
2925 if (after(tp->snd_nxt, tp->high_seq) && num_dupack)
2926 tcp_add_reno_sack(sk, num_dupack, flag & FLAG_ECE);
2927 else if (flag & FLAG_SND_UNA_ADVANCED)
2928 tcp_reset_reno_sack(tp);
2930 *rexmit = REXMIT_LOST;
2933 static bool tcp_force_fast_retransmit(struct sock *sk)
2935 struct tcp_sock *tp = tcp_sk(sk);
2937 return after(tcp_highest_sack_seq(tp),
2938 tp->snd_una + tp->reordering * tp->mss_cache);
2941 /* Undo during fast recovery after partial ACK. */
2942 static bool tcp_try_undo_partial(struct sock *sk, u32 prior_snd_una,
2945 struct tcp_sock *tp = tcp_sk(sk);
2947 if (tp->undo_marker && tcp_packet_delayed(tp)) {
2948 /* Plain luck! Hole if filled with delayed
2949 * packet, rather than with a retransmit. Check reordering.
2951 tcp_check_sack_reordering(sk, prior_snd_una, 1);
2953 /* We are getting evidence that the reordering degree is higher
2954 * than we realized. If there are no retransmits out then we
2955 * can undo. Otherwise we clock out new packets but do not
2956 * mark more packets lost or retransmit more.
2958 if (tp->retrans_out)
2961 if (!tcp_any_retrans_done(sk))
2962 tp->retrans_stamp = 0;
2964 DBGUNDO(sk, "partial recovery");
2965 tcp_undo_cwnd_reduction(sk, true);
2966 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2967 tcp_try_keep_open(sk);
2969 /* Partial ACK arrived. Force fast retransmit. */
2970 *do_lost = tcp_force_fast_retransmit(sk);
2975 static void tcp_identify_packet_loss(struct sock *sk, int *ack_flag)
2977 struct tcp_sock *tp = tcp_sk(sk);
2979 if (tcp_rtx_queue_empty(sk))
2982 if (unlikely(tcp_is_reno(tp))) {
2983 tcp_newreno_mark_lost(sk, *ack_flag & FLAG_SND_UNA_ADVANCED);
2984 } else if (tcp_is_rack(sk)) {
2985 u32 prior_retrans = tp->retrans_out;
2987 if (tcp_rack_mark_lost(sk))
2988 *ack_flag &= ~FLAG_SET_XMIT_TIMER;
2989 if (prior_retrans > tp->retrans_out)
2990 *ack_flag |= FLAG_LOST_RETRANS;
2994 /* Process an event, which can update packets-in-flight not trivially.
2995 * Main goal of this function is to calculate new estimate for left_out,
2996 * taking into account both packets sitting in receiver's buffer and
2997 * packets lost by network.
2999 * Besides that it updates the congestion state when packet loss or ECN
3000 * is detected. But it does not reduce the cwnd, it is done by the
3001 * congestion control later.
3003 * It does _not_ decide what to send, it is made in function
3004 * tcp_xmit_retransmit_queue().
3006 static void tcp_fastretrans_alert(struct sock *sk, const u32 prior_snd_una,
3007 int num_dupack, int *ack_flag, int *rexmit)
3009 struct inet_connection_sock *icsk = inet_csk(sk);
3010 struct tcp_sock *tp = tcp_sk(sk);
3011 int fast_rexmit = 0, flag = *ack_flag;
3012 bool ece_ack = flag & FLAG_ECE;
3013 bool do_lost = num_dupack || ((flag & FLAG_DATA_SACKED) &&
3014 tcp_force_fast_retransmit(sk));
3016 if (!tp->packets_out && tp->sacked_out)
3019 /* Now state machine starts.
3020 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
3022 tp->prior_ssthresh = 0;
3024 /* B. In all the states check for reneging SACKs. */
3025 if (tcp_check_sack_reneging(sk, ack_flag))
3028 /* C. Check consistency of the current state. */
3029 tcp_verify_left_out(tp);
3031 /* D. Check state exit conditions. State can be terminated
3032 * when high_seq is ACKed. */
3033 if (icsk->icsk_ca_state == TCP_CA_Open) {
3034 WARN_ON(tp->retrans_out != 0 && !tp->syn_data);
3035 tp->retrans_stamp = 0;
3036 } else if (!before(tp->snd_una, tp->high_seq)) {
3037 switch (icsk->icsk_ca_state) {
3039 /* CWR is to be held something *above* high_seq
3040 * is ACKed for CWR bit to reach receiver. */
3041 if (tp->snd_una != tp->high_seq) {
3042 tcp_end_cwnd_reduction(sk);
3043 tcp_set_ca_state(sk, TCP_CA_Open);
3047 case TCP_CA_Recovery:
3048 if (tcp_is_reno(tp))
3049 tcp_reset_reno_sack(tp);
3050 if (tcp_try_undo_recovery(sk))
3052 tcp_end_cwnd_reduction(sk);
3057 /* E. Process state. */
3058 switch (icsk->icsk_ca_state) {
3059 case TCP_CA_Recovery:
3060 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
3061 if (tcp_is_reno(tp))
3062 tcp_add_reno_sack(sk, num_dupack, ece_ack);
3063 } else if (tcp_try_undo_partial(sk, prior_snd_una, &do_lost))
3066 if (tcp_try_undo_dsack(sk))
3067 tcp_try_keep_open(sk);
3069 tcp_identify_packet_loss(sk, ack_flag);
3070 if (icsk->icsk_ca_state != TCP_CA_Recovery) {
3071 if (!tcp_time_to_recover(sk, flag))
3073 /* Undo reverts the recovery state. If loss is evident,
3074 * starts a new recovery (e.g. reordering then loss);
3076 tcp_enter_recovery(sk, ece_ack);
3080 tcp_process_loss(sk, flag, num_dupack, rexmit);
3081 if (icsk->icsk_ca_state != TCP_CA_Loss)
3082 tcp_update_rto_time(tp);
3083 tcp_identify_packet_loss(sk, ack_flag);
3084 if (!(icsk->icsk_ca_state == TCP_CA_Open ||
3085 (*ack_flag & FLAG_LOST_RETRANS)))
3087 /* Change state if cwnd is undone or retransmits are lost */
3090 if (tcp_is_reno(tp)) {
3091 if (flag & FLAG_SND_UNA_ADVANCED)
3092 tcp_reset_reno_sack(tp);
3093 tcp_add_reno_sack(sk, num_dupack, ece_ack);
3096 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
3097 tcp_try_undo_dsack(sk);
3099 tcp_identify_packet_loss(sk, ack_flag);
3100 if (!tcp_time_to_recover(sk, flag)) {
3101 tcp_try_to_open(sk, flag);
3105 /* MTU probe failure: don't reduce cwnd */
3106 if (icsk->icsk_ca_state < TCP_CA_CWR &&
3107 icsk->icsk_mtup.probe_size &&
3108 tp->snd_una == tp->mtu_probe.probe_seq_start) {
3109 tcp_mtup_probe_failed(sk);
3110 /* Restores the reduction we did in tcp_mtup_probe() */
3111 tcp_snd_cwnd_set(tp, tcp_snd_cwnd(tp) + 1);
3112 tcp_simple_retransmit(sk);
3116 /* Otherwise enter Recovery state */
3117 tcp_enter_recovery(sk, ece_ack);
3121 if (!tcp_is_rack(sk) && do_lost)
3122 tcp_update_scoreboard(sk, fast_rexmit);
3123 *rexmit = REXMIT_LOST;
3126 static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us, const int flag)
3128 u32 wlen = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_rtt_wlen) * HZ;
3129 struct tcp_sock *tp = tcp_sk(sk);
3131 if ((flag & FLAG_ACK_MAYBE_DELAYED) && rtt_us > tcp_min_rtt(tp)) {
3132 /* If the remote keeps returning delayed ACKs, eventually
3133 * the min filter would pick it up and overestimate the
3134 * prop. delay when it expires. Skip suspected delayed ACKs.
3138 minmax_running_min(&tp->rtt_min, wlen, tcp_jiffies32,
3139 rtt_us ? : jiffies_to_usecs(1));
3142 static bool tcp_ack_update_rtt(struct sock *sk, const int flag,
3143 long seq_rtt_us, long sack_rtt_us,
3144 long ca_rtt_us, struct rate_sample *rs)
3146 const struct tcp_sock *tp = tcp_sk(sk);
3148 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
3149 * broken middle-boxes or peers may corrupt TS-ECR fields. But
3150 * Karn's algorithm forbids taking RTT if some retransmitted data
3151 * is acked (RFC6298).
3154 seq_rtt_us = sack_rtt_us;
3156 /* RTTM Rule: A TSecr value received in a segment is used to
3157 * update the averaged RTT measurement only if the segment
3158 * acknowledges some new data, i.e., only if it advances the
3159 * left edge of the send window.
3160 * See draft-ietf-tcplw-high-performance-00, section 3.3.
3162 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp &&
3163 tp->rx_opt.rcv_tsecr && flag & FLAG_ACKED)
3164 seq_rtt_us = ca_rtt_us = tcp_rtt_tsopt_us(tp);
3166 rs->rtt_us = ca_rtt_us; /* RTT of last (S)ACKed packet (or -1) */
3170 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
3171 * always taken together with ACK, SACK, or TS-opts. Any negative
3172 * values will be skipped with the seq_rtt_us < 0 check above.
3174 tcp_update_rtt_min(sk, ca_rtt_us, flag);
3175 tcp_rtt_estimator(sk, seq_rtt_us);
3178 /* RFC6298: only reset backoff on valid RTT measurement. */
3179 inet_csk(sk)->icsk_backoff = 0;
3183 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
3184 void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req)
3186 struct rate_sample rs;
3189 if (req && !req->num_retrans && tcp_rsk(req)->snt_synack)
3190 rtt_us = tcp_stamp_us_delta(tcp_clock_us(), tcp_rsk(req)->snt_synack);
3192 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us, &rs);
3196 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
3198 const struct inet_connection_sock *icsk = inet_csk(sk);
3200 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
3201 tcp_sk(sk)->snd_cwnd_stamp = tcp_jiffies32;
3204 /* Restart timer after forward progress on connection.
3205 * RFC2988 recommends to restart timer to now+rto.
3207 void tcp_rearm_rto(struct sock *sk)
3209 const struct inet_connection_sock *icsk = inet_csk(sk);
3210 struct tcp_sock *tp = tcp_sk(sk);
3212 /* If the retrans timer is currently being used by Fast Open
3213 * for SYN-ACK retrans purpose, stay put.
3215 if (rcu_access_pointer(tp->fastopen_rsk))
3218 if (!tp->packets_out) {
3219 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
3221 u32 rto = inet_csk(sk)->icsk_rto;
3222 /* Offset the time elapsed after installing regular RTO */
3223 if (icsk->icsk_pending == ICSK_TIME_REO_TIMEOUT ||
3224 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
3225 s64 delta_us = tcp_rto_delta_us(sk);
3226 /* delta_us may not be positive if the socket is locked
3227 * when the retrans timer fires and is rescheduled.
3229 rto = usecs_to_jiffies(max_t(int, delta_us, 1));
3231 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
3236 /* Try to schedule a loss probe; if that doesn't work, then schedule an RTO. */
3237 static void tcp_set_xmit_timer(struct sock *sk)
3239 if (!tcp_schedule_loss_probe(sk, true))
3243 /* If we get here, the whole TSO packet has not been acked. */
3244 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3246 struct tcp_sock *tp = tcp_sk(sk);
3249 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3251 packets_acked = tcp_skb_pcount(skb);
3252 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3254 packets_acked -= tcp_skb_pcount(skb);
3256 if (packets_acked) {
3257 BUG_ON(tcp_skb_pcount(skb) == 0);
3258 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3261 return packets_acked;
3264 static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
3265 const struct sk_buff *ack_skb, u32 prior_snd_una)
3267 const struct skb_shared_info *shinfo;
3269 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3270 if (likely(!TCP_SKB_CB(skb)->txstamp_ack))
3273 shinfo = skb_shinfo(skb);
3274 if (!before(shinfo->tskey, prior_snd_una) &&
3275 before(shinfo->tskey, tcp_sk(sk)->snd_una)) {
3276 tcp_skb_tsorted_save(skb) {
3277 __skb_tstamp_tx(skb, ack_skb, NULL, sk, SCM_TSTAMP_ACK);
3278 } tcp_skb_tsorted_restore(skb);
3282 /* Remove acknowledged frames from the retransmission queue. If our packet
3283 * is before the ack sequence we can discard it as it's confirmed to have
3284 * arrived at the other end.
3286 static int tcp_clean_rtx_queue(struct sock *sk, const struct sk_buff *ack_skb,
3287 u32 prior_fack, u32 prior_snd_una,
3288 struct tcp_sacktag_state *sack, bool ece_ack)
3290 const struct inet_connection_sock *icsk = inet_csk(sk);
3291 u64 first_ackt, last_ackt;
3292 struct tcp_sock *tp = tcp_sk(sk);
3293 u32 prior_sacked = tp->sacked_out;
3294 u32 reord = tp->snd_nxt; /* lowest acked un-retx un-sacked seq */
3295 struct sk_buff *skb, *next;
3296 bool fully_acked = true;
3297 long sack_rtt_us = -1L;
3298 long seq_rtt_us = -1L;
3299 long ca_rtt_us = -1L;
3306 for (skb = skb_rb_first(&sk->tcp_rtx_queue); skb; skb = next) {
3307 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3308 const u32 start_seq = scb->seq;
3309 u8 sacked = scb->sacked;
3312 /* Determine how many packets and what bytes were acked, tso and else */
3313 if (after(scb->end_seq, tp->snd_una)) {
3314 if (tcp_skb_pcount(skb) == 1 ||
3315 !after(tp->snd_una, scb->seq))
3318 acked_pcount = tcp_tso_acked(sk, skb);
3321 fully_acked = false;
3323 acked_pcount = tcp_skb_pcount(skb);
3326 if (unlikely(sacked & TCPCB_RETRANS)) {
3327 if (sacked & TCPCB_SACKED_RETRANS)
3328 tp->retrans_out -= acked_pcount;
3329 flag |= FLAG_RETRANS_DATA_ACKED;
3330 } else if (!(sacked & TCPCB_SACKED_ACKED)) {
3331 last_ackt = tcp_skb_timestamp_us(skb);
3332 WARN_ON_ONCE(last_ackt == 0);
3334 first_ackt = last_ackt;
3336 if (before(start_seq, reord))
3338 if (!after(scb->end_seq, tp->high_seq))
3339 flag |= FLAG_ORIG_SACK_ACKED;
3342 if (sacked & TCPCB_SACKED_ACKED) {
3343 tp->sacked_out -= acked_pcount;
3344 } else if (tcp_is_sack(tp)) {
3345 tcp_count_delivered(tp, acked_pcount, ece_ack);
3346 if (!tcp_skb_spurious_retrans(tp, skb))
3347 tcp_rack_advance(tp, sacked, scb->end_seq,
3348 tcp_skb_timestamp_us(skb));
3350 if (sacked & TCPCB_LOST)
3351 tp->lost_out -= acked_pcount;
3353 tp->packets_out -= acked_pcount;
3354 pkts_acked += acked_pcount;
3355 tcp_rate_skb_delivered(sk, skb, sack->rate);
3357 /* Initial outgoing SYN's get put onto the write_queue
3358 * just like anything else we transmit. It is not
3359 * true data, and if we misinform our callers that
3360 * this ACK acks real data, we will erroneously exit
3361 * connection startup slow start one packet too
3362 * quickly. This is severely frowned upon behavior.
3364 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
3365 flag |= FLAG_DATA_ACKED;
3367 flag |= FLAG_SYN_ACKED;
3368 tp->retrans_stamp = 0;
3374 tcp_ack_tstamp(sk, skb, ack_skb, prior_snd_una);
3376 next = skb_rb_next(skb);
3377 if (unlikely(skb == tp->retransmit_skb_hint))
3378 tp->retransmit_skb_hint = NULL;
3379 if (unlikely(skb == tp->lost_skb_hint))
3380 tp->lost_skb_hint = NULL;
3381 tcp_highest_sack_replace(sk, skb, next);
3382 tcp_rtx_queue_unlink_and_free(skb, sk);
3386 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
3388 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3389 tp->snd_up = tp->snd_una;
3392 tcp_ack_tstamp(sk, skb, ack_skb, prior_snd_una);
3393 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
3394 flag |= FLAG_SACK_RENEGING;
3397 if (likely(first_ackt) && !(flag & FLAG_RETRANS_DATA_ACKED)) {
3398 seq_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, first_ackt);
3399 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, last_ackt);
3401 if (pkts_acked == 1 && fully_acked && !prior_sacked &&
3402 (tp->snd_una - prior_snd_una) < tp->mss_cache &&
3403 sack->rate->prior_delivered + 1 == tp->delivered &&
3404 !(flag & (FLAG_CA_ALERT | FLAG_SYN_ACKED))) {
3405 /* Conservatively mark a delayed ACK. It's typically
3406 * from a lone runt packet over the round trip to
3407 * a receiver w/o out-of-order or CE events.
3409 flag |= FLAG_ACK_MAYBE_DELAYED;
3412 if (sack->first_sackt) {
3413 sack_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->first_sackt);
3414 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->last_sackt);
3416 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us,
3417 ca_rtt_us, sack->rate);
3419 if (flag & FLAG_ACKED) {
3420 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3421 if (unlikely(icsk->icsk_mtup.probe_size &&
3422 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3423 tcp_mtup_probe_success(sk);
3426 if (tcp_is_reno(tp)) {
3427 tcp_remove_reno_sacks(sk, pkts_acked, ece_ack);
3429 /* If any of the cumulatively ACKed segments was
3430 * retransmitted, non-SACK case cannot confirm that
3431 * progress was due to original transmission due to
3432 * lack of TCPCB_SACKED_ACKED bits even if some of
3433 * the packets may have been never retransmitted.
3435 if (flag & FLAG_RETRANS_DATA_ACKED)
3436 flag &= ~FLAG_ORIG_SACK_ACKED;
3440 /* Non-retransmitted hole got filled? That's reordering */
3441 if (before(reord, prior_fack))
3442 tcp_check_sack_reordering(sk, reord, 0);
3444 delta = prior_sacked - tp->sacked_out;
3445 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3447 } else if (skb && rtt_update && sack_rtt_us >= 0 &&
3448 sack_rtt_us > tcp_stamp_us_delta(tp->tcp_mstamp,
3449 tcp_skb_timestamp_us(skb))) {
3450 /* Do not re-arm RTO if the sack RTT is measured from data sent
3451 * after when the head was last (re)transmitted. Otherwise the
3452 * timeout may continue to extend in loss recovery.
3454 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3457 if (icsk->icsk_ca_ops->pkts_acked) {
3458 struct ack_sample sample = { .pkts_acked = pkts_acked,
3459 .rtt_us = sack->rate->rtt_us };
3461 sample.in_flight = tp->mss_cache *
3462 (tp->delivered - sack->rate->prior_delivered);
3463 icsk->icsk_ca_ops->pkts_acked(sk, &sample);
3466 #if FASTRETRANS_DEBUG > 0
3467 WARN_ON((int)tp->sacked_out < 0);
3468 WARN_ON((int)tp->lost_out < 0);
3469 WARN_ON((int)tp->retrans_out < 0);
3470 if (!tp->packets_out && tcp_is_sack(tp)) {
3471 icsk = inet_csk(sk);
3473 pr_debug("Leak l=%u %d\n",
3474 tp->lost_out, icsk->icsk_ca_state);
3477 if (tp->sacked_out) {
3478 pr_debug("Leak s=%u %d\n",
3479 tp->sacked_out, icsk->icsk_ca_state);
3482 if (tp->retrans_out) {
3483 pr_debug("Leak r=%u %d\n",
3484 tp->retrans_out, icsk->icsk_ca_state);
3485 tp->retrans_out = 0;
3492 static void tcp_ack_probe(struct sock *sk)
3494 struct inet_connection_sock *icsk = inet_csk(sk);
3495 struct sk_buff *head = tcp_send_head(sk);
3496 const struct tcp_sock *tp = tcp_sk(sk);
3498 /* Was it a usable window open? */
3501 if (!after(TCP_SKB_CB(head)->end_seq, tcp_wnd_end(tp))) {
3502 icsk->icsk_backoff = 0;
3503 icsk->icsk_probes_tstamp = 0;
3504 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3505 /* Socket must be waked up by subsequent tcp_data_snd_check().
3506 * This function is not for random using!
3509 unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX);
3511 when = tcp_clamp_probe0_to_user_timeout(sk, when);
3512 tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, when, TCP_RTO_MAX);
3516 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3518 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3519 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3522 /* Decide wheather to run the increase function of congestion control. */
3523 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3525 /* If reordering is high then always grow cwnd whenever data is
3526 * delivered regardless of its ordering. Otherwise stay conservative
3527 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3528 * new SACK or ECE mark may first advance cwnd here and later reduce
3529 * cwnd in tcp_fastretrans_alert() based on more states.
3531 if (tcp_sk(sk)->reordering >
3532 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_reordering))
3533 return flag & FLAG_FORWARD_PROGRESS;
3535 return flag & FLAG_DATA_ACKED;
3538 /* The "ultimate" congestion control function that aims to replace the rigid
3539 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction).
3540 * It's called toward the end of processing an ACK with precise rate
3541 * information. All transmission or retransmission are delayed afterwards.
3543 static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked,
3544 int flag, const struct rate_sample *rs)
3546 const struct inet_connection_sock *icsk = inet_csk(sk);
3548 if (icsk->icsk_ca_ops->cong_control) {
3549 icsk->icsk_ca_ops->cong_control(sk, rs);
3553 if (tcp_in_cwnd_reduction(sk)) {
3554 /* Reduce cwnd if state mandates */
3555 tcp_cwnd_reduction(sk, acked_sacked, rs->losses, flag);
3556 } else if (tcp_may_raise_cwnd(sk, flag)) {
3557 /* Advance cwnd if state allows */
3558 tcp_cong_avoid(sk, ack, acked_sacked);
3560 tcp_update_pacing_rate(sk);
3563 /* Check that window update is acceptable.
3564 * The function assumes that snd_una<=ack<=snd_next.
3566 static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3567 const u32 ack, const u32 ack_seq,
3570 return after(ack, tp->snd_una) ||
3571 after(ack_seq, tp->snd_wl1) ||
3572 (ack_seq == tp->snd_wl1 && (nwin > tp->snd_wnd || !nwin));
3575 static void tcp_snd_sne_update(struct tcp_sock *tp, u32 ack)
3577 #ifdef CONFIG_TCP_AO
3578 struct tcp_ao_info *ao;
3580 if (!static_branch_unlikely(&tcp_ao_needed.key))
3583 ao = rcu_dereference_protected(tp->ao_info,
3584 lockdep_sock_is_held((struct sock *)tp));
3585 if (ao && ack < tp->snd_una)
3590 /* If we update tp->snd_una, also update tp->bytes_acked */
3591 static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack)
3593 u32 delta = ack - tp->snd_una;
3595 sock_owned_by_me((struct sock *)tp);
3596 tp->bytes_acked += delta;
3597 tcp_snd_sne_update(tp, ack);
3601 static void tcp_rcv_sne_update(struct tcp_sock *tp, u32 seq)
3603 #ifdef CONFIG_TCP_AO
3604 struct tcp_ao_info *ao;
3606 if (!static_branch_unlikely(&tcp_ao_needed.key))
3609 ao = rcu_dereference_protected(tp->ao_info,
3610 lockdep_sock_is_held((struct sock *)tp));
3611 if (ao && seq < tp->rcv_nxt)
3616 /* If we update tp->rcv_nxt, also update tp->bytes_received */
3617 static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq)
3619 u32 delta = seq - tp->rcv_nxt;
3621 sock_owned_by_me((struct sock *)tp);
3622 tp->bytes_received += delta;
3623 tcp_rcv_sne_update(tp, seq);
3624 WRITE_ONCE(tp->rcv_nxt, seq);
3627 /* Update our send window.
3629 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3630 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3632 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3635 struct tcp_sock *tp = tcp_sk(sk);
3637 u32 nwin = ntohs(tcp_hdr(skb)->window);
3639 if (likely(!tcp_hdr(skb)->syn))
3640 nwin <<= tp->rx_opt.snd_wscale;
3642 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3643 flag |= FLAG_WIN_UPDATE;
3644 tcp_update_wl(tp, ack_seq);
3646 if (tp->snd_wnd != nwin) {
3649 /* Note, it is the only place, where
3650 * fast path is recovered for sending TCP.
3653 tcp_fast_path_check(sk);
3655 if (!tcp_write_queue_empty(sk))
3656 tcp_slow_start_after_idle_check(sk);
3658 if (nwin > tp->max_window) {
3659 tp->max_window = nwin;
3660 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3665 tcp_snd_una_update(tp, ack);
3670 static bool __tcp_oow_rate_limited(struct net *net, int mib_idx,
3671 u32 *last_oow_ack_time)
3673 /* Paired with the WRITE_ONCE() in this function. */
3674 u32 val = READ_ONCE(*last_oow_ack_time);
3677 s32 elapsed = (s32)(tcp_jiffies32 - val);
3680 elapsed < READ_ONCE(net->ipv4.sysctl_tcp_invalid_ratelimit)) {
3681 NET_INC_STATS(net, mib_idx);
3682 return true; /* rate-limited: don't send yet! */
3686 /* Paired with the prior READ_ONCE() and with itself,
3687 * as we might be lockless.
3689 WRITE_ONCE(*last_oow_ack_time, tcp_jiffies32);
3691 return false; /* not rate-limited: go ahead, send dupack now! */
3694 /* Return true if we're currently rate-limiting out-of-window ACKs and
3695 * thus shouldn't send a dupack right now. We rate-limit dupacks in
3696 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3697 * attacks that send repeated SYNs or ACKs for the same connection. To
3698 * do this, we do not send a duplicate SYNACK or ACK if the remote
3699 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3701 bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb,
3702 int mib_idx, u32 *last_oow_ack_time)
3704 /* Data packets without SYNs are not likely part of an ACK loop. */
3705 if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) &&
3709 return __tcp_oow_rate_limited(net, mib_idx, last_oow_ack_time);
3712 /* RFC 5961 7 [ACK Throttling] */
3713 static void tcp_send_challenge_ack(struct sock *sk)
3715 struct tcp_sock *tp = tcp_sk(sk);
3716 struct net *net = sock_net(sk);
3717 u32 count, now, ack_limit;
3719 /* First check our per-socket dupack rate limit. */
3720 if (__tcp_oow_rate_limited(net,
3721 LINUX_MIB_TCPACKSKIPPEDCHALLENGE,
3722 &tp->last_oow_ack_time))
3725 ack_limit = READ_ONCE(net->ipv4.sysctl_tcp_challenge_ack_limit);
3726 if (ack_limit == INT_MAX)
3729 /* Then check host-wide RFC 5961 rate limit. */
3731 if (now != READ_ONCE(net->ipv4.tcp_challenge_timestamp)) {
3732 u32 half = (ack_limit + 1) >> 1;
3734 WRITE_ONCE(net->ipv4.tcp_challenge_timestamp, now);
3735 WRITE_ONCE(net->ipv4.tcp_challenge_count,
3736 get_random_u32_inclusive(half, ack_limit + half - 1));
3738 count = READ_ONCE(net->ipv4.tcp_challenge_count);
3740 WRITE_ONCE(net->ipv4.tcp_challenge_count, count - 1);
3742 NET_INC_STATS(net, LINUX_MIB_TCPCHALLENGEACK);
3747 static void tcp_store_ts_recent(struct tcp_sock *tp)
3749 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3750 tp->rx_opt.ts_recent_stamp = ktime_get_seconds();
3753 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3755 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3756 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3757 * extra check below makes sure this can only happen
3758 * for pure ACK frames. -DaveM
3760 * Not only, also it occurs for expired timestamps.
3763 if (tcp_paws_check(&tp->rx_opt, 0))
3764 tcp_store_ts_recent(tp);
3768 /* This routine deals with acks during a TLP episode and ends an episode by
3769 * resetting tlp_high_seq. Ref: TLP algorithm in draft-ietf-tcpm-rack
3771 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3773 struct tcp_sock *tp = tcp_sk(sk);
3775 if (before(ack, tp->tlp_high_seq))
3778 if (!tp->tlp_retrans) {
3779 /* TLP of new data has been acknowledged */
3780 tp->tlp_high_seq = 0;
3781 } else if (flag & FLAG_DSACK_TLP) {
3782 /* This DSACK means original and TLP probe arrived; no loss */
3783 tp->tlp_high_seq = 0;
3784 } else if (after(ack, tp->tlp_high_seq)) {
3785 /* ACK advances: there was a loss, so reduce cwnd. Reset
3786 * tlp_high_seq in tcp_init_cwnd_reduction()
3788 tcp_init_cwnd_reduction(sk);
3789 tcp_set_ca_state(sk, TCP_CA_CWR);
3790 tcp_end_cwnd_reduction(sk);
3791 tcp_try_keep_open(sk);
3792 NET_INC_STATS(sock_net(sk),
3793 LINUX_MIB_TCPLOSSPROBERECOVERY);
3794 } else if (!(flag & (FLAG_SND_UNA_ADVANCED |
3795 FLAG_NOT_DUP | FLAG_DATA_SACKED))) {
3796 /* Pure dupack: original and TLP probe arrived; no loss */
3797 tp->tlp_high_seq = 0;
3801 static inline void tcp_in_ack_event(struct sock *sk, u32 flags)
3803 const struct inet_connection_sock *icsk = inet_csk(sk);
3805 if (icsk->icsk_ca_ops->in_ack_event)
3806 icsk->icsk_ca_ops->in_ack_event(sk, flags);
3809 /* Congestion control has updated the cwnd already. So if we're in
3810 * loss recovery then now we do any new sends (for FRTO) or
3811 * retransmits (for CA_Loss or CA_recovery) that make sense.
3813 static void tcp_xmit_recovery(struct sock *sk, int rexmit)
3815 struct tcp_sock *tp = tcp_sk(sk);
3817 if (rexmit == REXMIT_NONE || sk->sk_state == TCP_SYN_SENT)
3820 if (unlikely(rexmit == REXMIT_NEW)) {
3821 __tcp_push_pending_frames(sk, tcp_current_mss(sk),
3823 if (after(tp->snd_nxt, tp->high_seq))
3827 tcp_xmit_retransmit_queue(sk);
3830 /* Returns the number of packets newly acked or sacked by the current ACK */
3831 static u32 tcp_newly_delivered(struct sock *sk, u32 prior_delivered, int flag)
3833 const struct net *net = sock_net(sk);
3834 struct tcp_sock *tp = tcp_sk(sk);
3837 delivered = tp->delivered - prior_delivered;
3838 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVERED, delivered);
3839 if (flag & FLAG_ECE)
3840 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVEREDCE, delivered);
3845 /* This routine deals with incoming acks, but not outgoing ones. */
3846 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3848 struct inet_connection_sock *icsk = inet_csk(sk);
3849 struct tcp_sock *tp = tcp_sk(sk);
3850 struct tcp_sacktag_state sack_state;
3851 struct rate_sample rs = { .prior_delivered = 0 };
3852 u32 prior_snd_una = tp->snd_una;
3853 bool is_sack_reneg = tp->is_sack_reneg;
3854 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3855 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3857 int prior_packets = tp->packets_out;
3858 u32 delivered = tp->delivered;
3859 u32 lost = tp->lost;
3860 int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */
3863 sack_state.first_sackt = 0;
3864 sack_state.rate = &rs;
3865 sack_state.sack_delivered = 0;
3867 /* We very likely will need to access rtx queue. */
3868 prefetch(sk->tcp_rtx_queue.rb_node);
3870 /* If the ack is older than previous acks
3871 * then we can probably ignore it.
3873 if (before(ack, prior_snd_una)) {
3876 /* do not accept ACK for bytes we never sent. */
3877 max_window = min_t(u64, tp->max_window, tp->bytes_acked);
3878 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3879 if (before(ack, prior_snd_una - max_window)) {
3880 if (!(flag & FLAG_NO_CHALLENGE_ACK))
3881 tcp_send_challenge_ack(sk);
3882 return -SKB_DROP_REASON_TCP_TOO_OLD_ACK;
3887 /* If the ack includes data we haven't sent yet, discard
3888 * this segment (RFC793 Section 3.9).
3890 if (after(ack, tp->snd_nxt))
3891 return -SKB_DROP_REASON_TCP_ACK_UNSENT_DATA;
3893 if (after(ack, prior_snd_una)) {
3894 flag |= FLAG_SND_UNA_ADVANCED;
3895 icsk->icsk_retransmits = 0;
3897 #if IS_ENABLED(CONFIG_TLS_DEVICE)
3898 if (static_branch_unlikely(&clean_acked_data_enabled.key))
3899 if (icsk->icsk_clean_acked)
3900 icsk->icsk_clean_acked(sk, ack);
3904 prior_fack = tcp_is_sack(tp) ? tcp_highest_sack_seq(tp) : tp->snd_una;
3905 rs.prior_in_flight = tcp_packets_in_flight(tp);
3907 /* ts_recent update must be made after we are sure that the packet
3910 if (flag & FLAG_UPDATE_TS_RECENT)
3911 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3913 if ((flag & (FLAG_SLOWPATH | FLAG_SND_UNA_ADVANCED)) ==
3914 FLAG_SND_UNA_ADVANCED) {
3915 /* Window is constant, pure forward advance.
3916 * No more checks are required.
3917 * Note, we use the fact that SND.UNA>=SND.WL2.
3919 tcp_update_wl(tp, ack_seq);
3920 tcp_snd_una_update(tp, ack);
3921 flag |= FLAG_WIN_UPDATE;
3923 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
3925 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPACKS);
3927 u32 ack_ev_flags = CA_ACK_SLOWPATH;
3929 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3932 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3934 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3936 if (TCP_SKB_CB(skb)->sacked)
3937 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3940 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) {
3942 ack_ev_flags |= CA_ACK_ECE;
3945 if (sack_state.sack_delivered)
3946 tcp_count_delivered(tp, sack_state.sack_delivered,
3949 if (flag & FLAG_WIN_UPDATE)
3950 ack_ev_flags |= CA_ACK_WIN_UPDATE;
3952 tcp_in_ack_event(sk, ack_ev_flags);
3955 /* This is a deviation from RFC3168 since it states that:
3956 * "When the TCP data sender is ready to set the CWR bit after reducing
3957 * the congestion window, it SHOULD set the CWR bit only on the first
3958 * new data packet that it transmits."
3959 * We accept CWR on pure ACKs to be more robust
3960 * with widely-deployed TCP implementations that do this.
3962 tcp_ecn_accept_cwr(sk, skb);
3964 /* We passed data and got it acked, remove any soft error
3965 * log. Something worked...
3967 WRITE_ONCE(sk->sk_err_soft, 0);
3968 icsk->icsk_probes_out = 0;
3969 tp->rcv_tstamp = tcp_jiffies32;
3973 /* See if we can take anything off of the retransmit queue. */
3974 flag |= tcp_clean_rtx_queue(sk, skb, prior_fack, prior_snd_una,
3975 &sack_state, flag & FLAG_ECE);
3977 tcp_rack_update_reo_wnd(sk, &rs);
3979 if (tp->tlp_high_seq)
3980 tcp_process_tlp_ack(sk, ack, flag);
3982 if (tcp_ack_is_dubious(sk, flag)) {
3983 if (!(flag & (FLAG_SND_UNA_ADVANCED |
3984 FLAG_NOT_DUP | FLAG_DSACKING_ACK))) {
3986 /* Consider if pure acks were aggregated in tcp_add_backlog() */
3987 if (!(flag & FLAG_DATA))
3988 num_dupack = max_t(u16, 1, skb_shinfo(skb)->gso_segs);
3990 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
3994 /* If needed, reset TLP/RTO timer when RACK doesn't set. */
3995 if (flag & FLAG_SET_XMIT_TIMER)
3996 tcp_set_xmit_timer(sk);
3998 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
4001 delivered = tcp_newly_delivered(sk, delivered, flag);
4002 lost = tp->lost - lost; /* freshly marked lost */
4003 rs.is_ack_delayed = !!(flag & FLAG_ACK_MAYBE_DELAYED);
4004 tcp_rate_gen(sk, delivered, lost, is_sack_reneg, sack_state.rate);
4005 tcp_cong_control(sk, ack, delivered, flag, sack_state.rate);
4006 tcp_xmit_recovery(sk, rexmit);
4010 /* If data was DSACKed, see if we can undo a cwnd reduction. */
4011 if (flag & FLAG_DSACKING_ACK) {
4012 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
4014 tcp_newly_delivered(sk, delivered, flag);
4016 /* If this ack opens up a zero window, clear backoff. It was
4017 * being used to time the probes, and is probably far higher than
4018 * it needs to be for normal retransmission.
4022 if (tp->tlp_high_seq)
4023 tcp_process_tlp_ack(sk, ack, flag);
4027 /* If data was SACKed, tag it and see if we should send more data.
4028 * If data was DSACKed, see if we can undo a cwnd reduction.
4030 if (TCP_SKB_CB(skb)->sacked) {
4031 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
4033 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
4035 tcp_newly_delivered(sk, delivered, flag);
4036 tcp_xmit_recovery(sk, rexmit);
4042 static void tcp_parse_fastopen_option(int len, const unsigned char *cookie,
4043 bool syn, struct tcp_fastopen_cookie *foc,
4046 /* Valid only in SYN or SYN-ACK with an even length. */
4047 if (!foc || !syn || len < 0 || (len & 1))
4050 if (len >= TCP_FASTOPEN_COOKIE_MIN &&
4051 len <= TCP_FASTOPEN_COOKIE_MAX)
4052 memcpy(foc->val, cookie, len);
4059 static bool smc_parse_options(const struct tcphdr *th,
4060 struct tcp_options_received *opt_rx,
4061 const unsigned char *ptr,
4064 #if IS_ENABLED(CONFIG_SMC)
4065 if (static_branch_unlikely(&tcp_have_smc)) {
4066 if (th->syn && !(opsize & 1) &&
4067 opsize >= TCPOLEN_EXP_SMC_BASE &&
4068 get_unaligned_be32(ptr) == TCPOPT_SMC_MAGIC) {
4077 /* Try to parse the MSS option from the TCP header. Return 0 on failure, clamped
4080 u16 tcp_parse_mss_option(const struct tcphdr *th, u16 user_mss)
4082 const unsigned char *ptr = (const unsigned char *)(th + 1);
4083 int length = (th->doff * 4) - sizeof(struct tcphdr);
4086 while (length > 0) {
4087 int opcode = *ptr++;
4093 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
4100 if (opsize < 2) /* "silly options" */
4102 if (opsize > length)
4103 return mss; /* fail on partial options */
4104 if (opcode == TCPOPT_MSS && opsize == TCPOLEN_MSS) {
4105 u16 in_mss = get_unaligned_be16(ptr);
4108 if (user_mss && user_mss < in_mss)
4119 EXPORT_SYMBOL_GPL(tcp_parse_mss_option);
4121 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
4122 * But, this can also be called on packets in the established flow when
4123 * the fast version below fails.
4125 void tcp_parse_options(const struct net *net,
4126 const struct sk_buff *skb,
4127 struct tcp_options_received *opt_rx, int estab,
4128 struct tcp_fastopen_cookie *foc)
4130 const unsigned char *ptr;
4131 const struct tcphdr *th = tcp_hdr(skb);
4132 int length = (th->doff * 4) - sizeof(struct tcphdr);
4134 ptr = (const unsigned char *)(th + 1);
4135 opt_rx->saw_tstamp = 0;
4136 opt_rx->saw_unknown = 0;
4138 while (length > 0) {
4139 int opcode = *ptr++;
4145 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
4152 if (opsize < 2) /* "silly options" */
4154 if (opsize > length)
4155 return; /* don't parse partial options */
4158 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
4159 u16 in_mss = get_unaligned_be16(ptr);
4161 if (opt_rx->user_mss &&
4162 opt_rx->user_mss < in_mss)
4163 in_mss = opt_rx->user_mss;
4164 opt_rx->mss_clamp = in_mss;
4169 if (opsize == TCPOLEN_WINDOW && th->syn &&
4170 !estab && READ_ONCE(net->ipv4.sysctl_tcp_window_scaling)) {
4171 __u8 snd_wscale = *(__u8 *)ptr;
4172 opt_rx->wscale_ok = 1;
4173 if (snd_wscale > TCP_MAX_WSCALE) {
4174 net_info_ratelimited("%s: Illegal window scaling value %d > %u received\n",
4178 snd_wscale = TCP_MAX_WSCALE;
4180 opt_rx->snd_wscale = snd_wscale;
4183 case TCPOPT_TIMESTAMP:
4184 if ((opsize == TCPOLEN_TIMESTAMP) &&
4185 ((estab && opt_rx->tstamp_ok) ||
4186 (!estab && READ_ONCE(net->ipv4.sysctl_tcp_timestamps)))) {
4187 opt_rx->saw_tstamp = 1;
4188 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
4189 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
4192 case TCPOPT_SACK_PERM:
4193 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
4194 !estab && READ_ONCE(net->ipv4.sysctl_tcp_sack)) {
4195 opt_rx->sack_ok = TCP_SACK_SEEN;
4196 tcp_sack_reset(opt_rx);
4201 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
4202 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
4204 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
4207 #ifdef CONFIG_TCP_MD5SIG
4209 /* The MD5 Hash has already been
4210 * checked (see tcp_v{4,6}_rcv()).
4214 case TCPOPT_FASTOPEN:
4215 tcp_parse_fastopen_option(
4216 opsize - TCPOLEN_FASTOPEN_BASE,
4217 ptr, th->syn, foc, false);
4221 /* Fast Open option shares code 254 using a
4222 * 16 bits magic number.
4224 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE &&
4225 get_unaligned_be16(ptr) ==
4226 TCPOPT_FASTOPEN_MAGIC) {
4227 tcp_parse_fastopen_option(opsize -
4228 TCPOLEN_EXP_FASTOPEN_BASE,
4229 ptr + 2, th->syn, foc, true);
4233 if (smc_parse_options(th, opt_rx, ptr, opsize))
4236 opt_rx->saw_unknown = 1;
4240 opt_rx->saw_unknown = 1;
4247 EXPORT_SYMBOL(tcp_parse_options);
4249 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
4251 const __be32 *ptr = (const __be32 *)(th + 1);
4253 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
4254 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
4255 tp->rx_opt.saw_tstamp = 1;
4257 tp->rx_opt.rcv_tsval = ntohl(*ptr);
4260 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
4262 tp->rx_opt.rcv_tsecr = 0;
4268 /* Fast parse options. This hopes to only see timestamps.
4269 * If it is wrong it falls back on tcp_parse_options().
4271 static bool tcp_fast_parse_options(const struct net *net,
4272 const struct sk_buff *skb,
4273 const struct tcphdr *th, struct tcp_sock *tp)
4275 /* In the spirit of fast parsing, compare doff directly to constant
4276 * values. Because equality is used, short doff can be ignored here.
4278 if (th->doff == (sizeof(*th) / 4)) {
4279 tp->rx_opt.saw_tstamp = 0;
4281 } else if (tp->rx_opt.tstamp_ok &&
4282 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
4283 if (tcp_parse_aligned_timestamp(tp, th))
4287 tcp_parse_options(net, skb, &tp->rx_opt, 1, NULL);
4288 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
4289 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
4294 #if defined(CONFIG_TCP_MD5SIG) || defined(CONFIG_TCP_AO)
4296 * Parse Signature options
4298 int tcp_do_parse_auth_options(const struct tcphdr *th,
4299 const u8 **md5_hash, const u8 **ao_hash)
4301 int length = (th->doff << 2) - sizeof(*th);
4302 const u8 *ptr = (const u8 *)(th + 1);
4303 unsigned int minlen = TCPOLEN_MD5SIG;
4305 if (IS_ENABLED(CONFIG_TCP_AO))
4306 minlen = sizeof(struct tcp_ao_hdr) + 1;
4311 /* If not enough data remaining, we can short cut */
4312 while (length >= minlen) {
4313 int opcode = *ptr++;
4324 if (opsize < 2 || opsize > length)
4326 if (opcode == TCPOPT_MD5SIG) {
4327 if (opsize != TCPOLEN_MD5SIG)
4329 if (unlikely(*md5_hash || *ao_hash))
4332 } else if (opcode == TCPOPT_AO) {
4333 if (opsize <= sizeof(struct tcp_ao_hdr))
4335 if (unlikely(*md5_hash || *ao_hash))
4345 EXPORT_SYMBOL(tcp_do_parse_auth_options);
4348 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
4350 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
4351 * it can pass through stack. So, the following predicate verifies that
4352 * this segment is not used for anything but congestion avoidance or
4353 * fast retransmit. Moreover, we even are able to eliminate most of such
4354 * second order effects, if we apply some small "replay" window (~RTO)
4355 * to timestamp space.
4357 * All these measures still do not guarantee that we reject wrapped ACKs
4358 * on networks with high bandwidth, when sequence space is recycled fastly,
4359 * but it guarantees that such events will be very rare and do not affect
4360 * connection seriously. This doesn't look nice, but alas, PAWS is really
4363 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
4364 * states that events when retransmit arrives after original data are rare.
4365 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
4366 * the biggest problem on large power networks even with minor reordering.
4367 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
4368 * up to bandwidth of 18Gigabit/sec. 8) ]
4371 /* Estimates max number of increments of remote peer TSval in
4372 * a replay window (based on our current RTO estimation).
4374 static u32 tcp_tsval_replay(const struct sock *sk)
4376 /* If we use usec TS resolution,
4377 * then expect the remote peer to use the same resolution.
4379 if (tcp_sk(sk)->tcp_usec_ts)
4380 return inet_csk(sk)->icsk_rto * (USEC_PER_SEC / HZ);
4382 /* RFC 7323 recommends a TSval clock between 1ms and 1sec.
4383 * We know that some OS (including old linux) can use 1200 Hz.
4385 return inet_csk(sk)->icsk_rto * 1200 / HZ;
4388 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
4390 const struct tcp_sock *tp = tcp_sk(sk);
4391 const struct tcphdr *th = tcp_hdr(skb);
4392 u32 seq = TCP_SKB_CB(skb)->seq;
4393 u32 ack = TCP_SKB_CB(skb)->ack_seq;
4395 return /* 1. Pure ACK with correct sequence number. */
4396 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
4398 /* 2. ... and duplicate ACK. */
4399 ack == tp->snd_una &&
4401 /* 3. ... and does not update window. */
4402 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
4404 /* 4. ... and sits in replay window. */
4405 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <=
4406 tcp_tsval_replay(sk);
4409 static inline bool tcp_paws_discard(const struct sock *sk,
4410 const struct sk_buff *skb)
4412 const struct tcp_sock *tp = tcp_sk(sk);
4414 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
4415 !tcp_disordered_ack(sk, skb);
4418 /* Check segment sequence number for validity.
4420 * Segment controls are considered valid, if the segment
4421 * fits to the window after truncation to the window. Acceptability
4422 * of data (and SYN, FIN, of course) is checked separately.
4423 * See tcp_data_queue(), for example.
4425 * Also, controls (RST is main one) are accepted using RCV.WUP instead
4426 * of RCV.NXT. Peer still did not advance his SND.UNA when we
4427 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4428 * (borrowed from freebsd)
4431 static enum skb_drop_reason tcp_sequence(const struct tcp_sock *tp,
4432 u32 seq, u32 end_seq)
4434 if (before(end_seq, tp->rcv_wup))
4435 return SKB_DROP_REASON_TCP_OLD_SEQUENCE;
4437 if (after(seq, tp->rcv_nxt + tcp_receive_window(tp)))
4438 return SKB_DROP_REASON_TCP_INVALID_SEQUENCE;
4440 return SKB_NOT_DROPPED_YET;
4443 /* When we get a reset we do this. */
4444 void tcp_reset(struct sock *sk, struct sk_buff *skb)
4446 trace_tcp_receive_reset(sk);
4448 /* mptcp can't tell us to ignore reset pkts,
4449 * so just ignore the return value of mptcp_incoming_options().
4451 if (sk_is_mptcp(sk))
4452 mptcp_incoming_options(sk, skb);
4454 /* We want the right error as BSD sees it (and indeed as we do). */
4455 switch (sk->sk_state) {
4457 WRITE_ONCE(sk->sk_err, ECONNREFUSED);
4459 case TCP_CLOSE_WAIT:
4460 WRITE_ONCE(sk->sk_err, EPIPE);
4465 WRITE_ONCE(sk->sk_err, ECONNRESET);
4467 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4470 tcp_write_queue_purge(sk);
4473 if (!sock_flag(sk, SOCK_DEAD))
4474 sk_error_report(sk);
4478 * Process the FIN bit. This now behaves as it is supposed to work
4479 * and the FIN takes effect when it is validly part of sequence
4480 * space. Not before when we get holes.
4482 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4483 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4486 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4487 * close and we go into CLOSING (and later onto TIME-WAIT)
4489 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4491 void tcp_fin(struct sock *sk)
4493 struct tcp_sock *tp = tcp_sk(sk);
4495 inet_csk_schedule_ack(sk);
4497 WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | RCV_SHUTDOWN);
4498 sock_set_flag(sk, SOCK_DONE);
4500 switch (sk->sk_state) {
4502 case TCP_ESTABLISHED:
4503 /* Move to CLOSE_WAIT */
4504 tcp_set_state(sk, TCP_CLOSE_WAIT);
4505 inet_csk_enter_pingpong_mode(sk);
4508 case TCP_CLOSE_WAIT:
4510 /* Received a retransmission of the FIN, do
4515 /* RFC793: Remain in the LAST-ACK state. */
4519 /* This case occurs when a simultaneous close
4520 * happens, we must ack the received FIN and
4521 * enter the CLOSING state.
4524 tcp_set_state(sk, TCP_CLOSING);
4527 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4529 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4532 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4533 * cases we should never reach this piece of code.
4535 pr_err("%s: Impossible, sk->sk_state=%d\n",
4536 __func__, sk->sk_state);
4540 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4541 * Probably, we should reset in this case. For now drop them.
4543 skb_rbtree_purge(&tp->out_of_order_queue);
4544 if (tcp_is_sack(tp))
4545 tcp_sack_reset(&tp->rx_opt);
4547 if (!sock_flag(sk, SOCK_DEAD)) {
4548 sk->sk_state_change(sk);
4550 /* Do not send POLL_HUP for half duplex close. */
4551 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4552 sk->sk_state == TCP_CLOSE)
4553 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4555 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4559 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4562 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4563 if (before(seq, sp->start_seq))
4564 sp->start_seq = seq;
4565 if (after(end_seq, sp->end_seq))
4566 sp->end_seq = end_seq;
4572 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4574 struct tcp_sock *tp = tcp_sk(sk);
4576 if (tcp_is_sack(tp) && READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_dsack)) {
4579 if (before(seq, tp->rcv_nxt))
4580 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4582 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4584 NET_INC_STATS(sock_net(sk), mib_idx);
4586 tp->rx_opt.dsack = 1;
4587 tp->duplicate_sack[0].start_seq = seq;
4588 tp->duplicate_sack[0].end_seq = end_seq;
4592 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4594 struct tcp_sock *tp = tcp_sk(sk);
4596 if (!tp->rx_opt.dsack)
4597 tcp_dsack_set(sk, seq, end_seq);
4599 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4602 static void tcp_rcv_spurious_retrans(struct sock *sk, const struct sk_buff *skb)
4604 /* When the ACK path fails or drops most ACKs, the sender would
4605 * timeout and spuriously retransmit the same segment repeatedly.
4606 * If it seems our ACKs are not reaching the other side,
4607 * based on receiving a duplicate data segment with new flowlabel
4608 * (suggesting the sender suffered an RTO), and we are not already
4609 * repathing due to our own RTO, then rehash the socket to repath our
4612 #if IS_ENABLED(CONFIG_IPV6)
4613 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Loss &&
4614 skb->protocol == htons(ETH_P_IPV6) &&
4615 (tcp_sk(sk)->inet_conn.icsk_ack.lrcv_flowlabel !=
4616 ntohl(ip6_flowlabel(ipv6_hdr(skb)))) &&
4617 sk_rethink_txhash(sk))
4618 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDUPLICATEDATAREHASH);
4620 /* Save last flowlabel after a spurious retrans. */
4621 tcp_save_lrcv_flowlabel(sk, skb);
4625 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4627 struct tcp_sock *tp = tcp_sk(sk);
4629 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4630 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4631 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4632 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
4634 if (tcp_is_sack(tp) && READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_dsack)) {
4635 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4637 tcp_rcv_spurious_retrans(sk, skb);
4638 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4639 end_seq = tp->rcv_nxt;
4640 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4647 /* These routines update the SACK block as out-of-order packets arrive or
4648 * in-order packets close up the sequence space.
4650 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4653 struct tcp_sack_block *sp = &tp->selective_acks[0];
4654 struct tcp_sack_block *swalk = sp + 1;
4656 /* See if the recent change to the first SACK eats into
4657 * or hits the sequence space of other SACK blocks, if so coalesce.
4659 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4660 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4663 /* Zap SWALK, by moving every further SACK up by one slot.
4664 * Decrease num_sacks.
4666 tp->rx_opt.num_sacks--;
4667 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4676 void tcp_sack_compress_send_ack(struct sock *sk)
4678 struct tcp_sock *tp = tcp_sk(sk);
4680 if (!tp->compressed_ack)
4683 if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1)
4686 /* Since we have to send one ack finally,
4687 * substract one from tp->compressed_ack to keep
4688 * LINUX_MIB_TCPACKCOMPRESSED accurate.
4690 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED,
4691 tp->compressed_ack - 1);
4693 tp->compressed_ack = 0;
4697 /* Reasonable amount of sack blocks included in TCP SACK option
4698 * The max is 4, but this becomes 3 if TCP timestamps are there.
4699 * Given that SACK packets might be lost, be conservative and use 2.
4701 #define TCP_SACK_BLOCKS_EXPECTED 2
4703 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4705 struct tcp_sock *tp = tcp_sk(sk);
4706 struct tcp_sack_block *sp = &tp->selective_acks[0];
4707 int cur_sacks = tp->rx_opt.num_sacks;
4713 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4714 if (tcp_sack_extend(sp, seq, end_seq)) {
4715 if (this_sack >= TCP_SACK_BLOCKS_EXPECTED)
4716 tcp_sack_compress_send_ack(sk);
4717 /* Rotate this_sack to the first one. */
4718 for (; this_sack > 0; this_sack--, sp--)
4719 swap(*sp, *(sp - 1));
4721 tcp_sack_maybe_coalesce(tp);
4726 if (this_sack >= TCP_SACK_BLOCKS_EXPECTED)
4727 tcp_sack_compress_send_ack(sk);
4729 /* Could not find an adjacent existing SACK, build a new one,
4730 * put it at the front, and shift everyone else down. We
4731 * always know there is at least one SACK present already here.
4733 * If the sack array is full, forget about the last one.
4735 if (this_sack >= TCP_NUM_SACKS) {
4737 tp->rx_opt.num_sacks--;
4740 for (; this_sack > 0; this_sack--, sp--)
4744 /* Build the new head SACK, and we're done. */
4745 sp->start_seq = seq;
4746 sp->end_seq = end_seq;
4747 tp->rx_opt.num_sacks++;
4750 /* RCV.NXT advances, some SACKs should be eaten. */
4752 static void tcp_sack_remove(struct tcp_sock *tp)
4754 struct tcp_sack_block *sp = &tp->selective_acks[0];
4755 int num_sacks = tp->rx_opt.num_sacks;
4758 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4759 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4760 tp->rx_opt.num_sacks = 0;
4764 for (this_sack = 0; this_sack < num_sacks;) {
4765 /* Check if the start of the sack is covered by RCV.NXT. */
4766 if (!before(tp->rcv_nxt, sp->start_seq)) {
4769 /* RCV.NXT must cover all the block! */
4770 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4772 /* Zap this SACK, by moving forward any other SACKS. */
4773 for (i = this_sack+1; i < num_sacks; i++)
4774 tp->selective_acks[i-1] = tp->selective_acks[i];
4781 tp->rx_opt.num_sacks = num_sacks;
4785 * tcp_try_coalesce - try to merge skb to prior one
4788 * @from: buffer to add in queue
4789 * @fragstolen: pointer to boolean
4791 * Before queueing skb @from after @to, try to merge them
4792 * to reduce overall memory use and queue lengths, if cost is small.
4793 * Packets in ofo or receive queues can stay a long time.
4794 * Better try to coalesce them right now to avoid future collapses.
4795 * Returns true if caller should free @from instead of queueing it
4797 static bool tcp_try_coalesce(struct sock *sk,
4799 struct sk_buff *from,
4804 *fragstolen = false;
4806 /* Its possible this segment overlaps with prior segment in queue */
4807 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4810 if (!mptcp_skb_can_collapse(to, from))
4813 #ifdef CONFIG_TLS_DEVICE
4814 if (from->decrypted != to->decrypted)
4818 if (!skb_try_coalesce(to, from, fragstolen, &delta))
4821 atomic_add(delta, &sk->sk_rmem_alloc);
4822 sk_mem_charge(sk, delta);
4823 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4824 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4825 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4826 TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
4828 if (TCP_SKB_CB(from)->has_rxtstamp) {
4829 TCP_SKB_CB(to)->has_rxtstamp = true;
4830 to->tstamp = from->tstamp;
4831 skb_hwtstamps(to)->hwtstamp = skb_hwtstamps(from)->hwtstamp;
4837 static bool tcp_ooo_try_coalesce(struct sock *sk,
4839 struct sk_buff *from,
4842 bool res = tcp_try_coalesce(sk, to, from, fragstolen);
4844 /* In case tcp_drop_reason() is called later, update to->gso_segs */
4846 u32 gso_segs = max_t(u16, 1, skb_shinfo(to)->gso_segs) +
4847 max_t(u16, 1, skb_shinfo(from)->gso_segs);
4849 skb_shinfo(to)->gso_segs = min_t(u32, gso_segs, 0xFFFF);
4854 static void tcp_drop_reason(struct sock *sk, struct sk_buff *skb,
4855 enum skb_drop_reason reason)
4857 sk_drops_add(sk, skb);
4858 kfree_skb_reason(skb, reason);
4861 /* This one checks to see if we can put data from the
4862 * out_of_order queue into the receive_queue.
4864 static void tcp_ofo_queue(struct sock *sk)
4866 struct tcp_sock *tp = tcp_sk(sk);
4867 __u32 dsack_high = tp->rcv_nxt;
4868 bool fin, fragstolen, eaten;
4869 struct sk_buff *skb, *tail;
4872 p = rb_first(&tp->out_of_order_queue);
4875 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4878 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4879 __u32 dsack = dsack_high;
4880 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4881 dsack_high = TCP_SKB_CB(skb)->end_seq;
4882 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4885 rb_erase(&skb->rbnode, &tp->out_of_order_queue);
4887 if (unlikely(!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))) {
4888 tcp_drop_reason(sk, skb, SKB_DROP_REASON_TCP_OFO_DROP);
4892 tail = skb_peek_tail(&sk->sk_receive_queue);
4893 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen);
4894 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4895 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
4897 __skb_queue_tail(&sk->sk_receive_queue, skb);
4899 kfree_skb_partial(skb, fragstolen);
4901 if (unlikely(fin)) {
4903 /* tcp_fin() purges tp->out_of_order_queue,
4904 * so we must end this loop right now.
4911 static bool tcp_prune_ofo_queue(struct sock *sk, const struct sk_buff *in_skb);
4912 static int tcp_prune_queue(struct sock *sk, const struct sk_buff *in_skb);
4914 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4917 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4918 !sk_rmem_schedule(sk, skb, size)) {
4920 if (tcp_prune_queue(sk, skb) < 0)
4923 while (!sk_rmem_schedule(sk, skb, size)) {
4924 if (!tcp_prune_ofo_queue(sk, skb))
4931 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4933 struct tcp_sock *tp = tcp_sk(sk);
4934 struct rb_node **p, *parent;
4935 struct sk_buff *skb1;
4939 tcp_save_lrcv_flowlabel(sk, skb);
4940 tcp_ecn_check_ce(sk, skb);
4942 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4943 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFODROP);
4944 sk->sk_data_ready(sk);
4945 tcp_drop_reason(sk, skb, SKB_DROP_REASON_PROTO_MEM);
4949 /* Disable header prediction. */
4951 inet_csk_schedule_ack(sk);
4953 tp->rcv_ooopack += max_t(u16, 1, skb_shinfo(skb)->gso_segs);
4954 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4955 seq = TCP_SKB_CB(skb)->seq;
4956 end_seq = TCP_SKB_CB(skb)->end_seq;
4958 p = &tp->out_of_order_queue.rb_node;
4959 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4960 /* Initial out of order segment, build 1 SACK. */
4961 if (tcp_is_sack(tp)) {
4962 tp->rx_opt.num_sacks = 1;
4963 tp->selective_acks[0].start_seq = seq;
4964 tp->selective_acks[0].end_seq = end_seq;
4966 rb_link_node(&skb->rbnode, NULL, p);
4967 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4968 tp->ooo_last_skb = skb;
4972 /* In the typical case, we are adding an skb to the end of the list.
4973 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
4975 if (tcp_ooo_try_coalesce(sk, tp->ooo_last_skb,
4976 skb, &fragstolen)) {
4978 /* For non sack flows, do not grow window to force DUPACK
4979 * and trigger fast retransmit.
4981 if (tcp_is_sack(tp))
4982 tcp_grow_window(sk, skb, true);
4983 kfree_skb_partial(skb, fragstolen);
4987 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
4988 if (!before(seq, TCP_SKB_CB(tp->ooo_last_skb)->end_seq)) {
4989 parent = &tp->ooo_last_skb->rbnode;
4990 p = &parent->rb_right;
4994 /* Find place to insert this segment. Handle overlaps on the way. */
4998 skb1 = rb_to_skb(parent);
4999 if (before(seq, TCP_SKB_CB(skb1)->seq)) {
5000 p = &parent->rb_left;
5003 if (before(seq, TCP_SKB_CB(skb1)->end_seq)) {
5004 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
5005 /* All the bits are present. Drop. */
5006 NET_INC_STATS(sock_net(sk),
5007 LINUX_MIB_TCPOFOMERGE);
5008 tcp_drop_reason(sk, skb,
5009 SKB_DROP_REASON_TCP_OFOMERGE);
5011 tcp_dsack_set(sk, seq, end_seq);
5014 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
5015 /* Partial overlap. */
5016 tcp_dsack_set(sk, seq, TCP_SKB_CB(skb1)->end_seq);
5018 /* skb's seq == skb1's seq and skb covers skb1.
5019 * Replace skb1 with skb.
5021 rb_replace_node(&skb1->rbnode, &skb->rbnode,
5022 &tp->out_of_order_queue);
5023 tcp_dsack_extend(sk,
5024 TCP_SKB_CB(skb1)->seq,
5025 TCP_SKB_CB(skb1)->end_seq);
5026 NET_INC_STATS(sock_net(sk),
5027 LINUX_MIB_TCPOFOMERGE);
5028 tcp_drop_reason(sk, skb1,
5029 SKB_DROP_REASON_TCP_OFOMERGE);
5032 } else if (tcp_ooo_try_coalesce(sk, skb1,
5033 skb, &fragstolen)) {
5036 p = &parent->rb_right;
5039 /* Insert segment into RB tree. */
5040 rb_link_node(&skb->rbnode, parent, p);
5041 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
5044 /* Remove other segments covered by skb. */
5045 while ((skb1 = skb_rb_next(skb)) != NULL) {
5046 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
5048 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
5049 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
5053 rb_erase(&skb1->rbnode, &tp->out_of_order_queue);
5054 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
5055 TCP_SKB_CB(skb1)->end_seq);
5056 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
5057 tcp_drop_reason(sk, skb1, SKB_DROP_REASON_TCP_OFOMERGE);
5059 /* If there is no skb after us, we are the last_skb ! */
5061 tp->ooo_last_skb = skb;
5064 if (tcp_is_sack(tp))
5065 tcp_sack_new_ofo_skb(sk, seq, end_seq);
5068 /* For non sack flows, do not grow window to force DUPACK
5069 * and trigger fast retransmit.
5071 if (tcp_is_sack(tp))
5072 tcp_grow_window(sk, skb, false);
5074 skb_set_owner_r(skb, sk);
5078 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb,
5082 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
5085 tcp_try_coalesce(sk, tail,
5086 skb, fragstolen)) ? 1 : 0;
5087 tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq);
5089 __skb_queue_tail(&sk->sk_receive_queue, skb);
5090 skb_set_owner_r(skb, sk);
5095 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
5097 struct sk_buff *skb;
5105 if (size > PAGE_SIZE) {
5106 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS);
5108 data_len = npages << PAGE_SHIFT;
5109 size = data_len + (size & ~PAGE_MASK);
5111 skb = alloc_skb_with_frags(size - data_len, data_len,
5112 PAGE_ALLOC_COSTLY_ORDER,
5113 &err, sk->sk_allocation);
5117 skb_put(skb, size - data_len);
5118 skb->data_len = data_len;
5121 if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) {
5122 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP);
5126 err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size);
5130 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
5131 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
5132 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
5134 if (tcp_queue_rcv(sk, skb, &fragstolen)) {
5135 WARN_ON_ONCE(fragstolen); /* should not happen */
5147 void tcp_data_ready(struct sock *sk)
5149 if (tcp_epollin_ready(sk, sk->sk_rcvlowat) || sock_flag(sk, SOCK_DONE))
5150 sk->sk_data_ready(sk);
5153 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
5155 struct tcp_sock *tp = tcp_sk(sk);
5156 enum skb_drop_reason reason;
5160 /* If a subflow has been reset, the packet should not continue
5161 * to be processed, drop the packet.
5163 if (sk_is_mptcp(sk) && !mptcp_incoming_options(sk, skb)) {
5168 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
5173 __skb_pull(skb, tcp_hdr(skb)->doff * 4);
5175 reason = SKB_DROP_REASON_NOT_SPECIFIED;
5176 tp->rx_opt.dsack = 0;
5178 /* Queue data for delivery to the user.
5179 * Packets in sequence go to the receive queue.
5180 * Out of sequence packets to the out_of_order_queue.
5182 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
5183 if (tcp_receive_window(tp) == 0) {
5184 reason = SKB_DROP_REASON_TCP_ZEROWINDOW;
5185 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP);
5189 /* Ok. In sequence. In window. */
5191 if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) {
5192 /* TODO: maybe ratelimit these WIN 0 ACK ? */
5193 inet_csk(sk)->icsk_ack.pending |=
5194 (ICSK_ACK_NOMEM | ICSK_ACK_NOW);
5195 inet_csk_schedule_ack(sk);
5196 sk->sk_data_ready(sk);
5198 if (skb_queue_len(&sk->sk_receive_queue)) {
5199 reason = SKB_DROP_REASON_PROTO_MEM;
5200 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP);
5203 sk_forced_mem_schedule(sk, skb->truesize);
5206 eaten = tcp_queue_rcv(sk, skb, &fragstolen);
5208 tcp_event_data_recv(sk, skb);
5209 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
5212 if (!RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
5215 /* RFC5681. 4.2. SHOULD send immediate ACK, when
5216 * gap in queue is filled.
5218 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
5219 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW;
5222 if (tp->rx_opt.num_sacks)
5223 tcp_sack_remove(tp);
5225 tcp_fast_path_check(sk);
5228 kfree_skb_partial(skb, fragstolen);
5229 if (!sock_flag(sk, SOCK_DEAD))
5234 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
5235 tcp_rcv_spurious_retrans(sk, skb);
5236 /* A retransmit, 2nd most common case. Force an immediate ack. */
5237 reason = SKB_DROP_REASON_TCP_OLD_DATA;
5238 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
5239 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
5242 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
5243 inet_csk_schedule_ack(sk);
5245 tcp_drop_reason(sk, skb, reason);
5249 /* Out of window. F.e. zero window probe. */
5250 if (!before(TCP_SKB_CB(skb)->seq,
5251 tp->rcv_nxt + tcp_receive_window(tp))) {
5252 reason = SKB_DROP_REASON_TCP_OVERWINDOW;
5256 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
5257 /* Partial packet, seq < rcv_next < end_seq */
5258 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
5260 /* If window is closed, drop tail of packet. But after
5261 * remembering D-SACK for its head made in previous line.
5263 if (!tcp_receive_window(tp)) {
5264 reason = SKB_DROP_REASON_TCP_ZEROWINDOW;
5265 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP);
5271 tcp_data_queue_ofo(sk, skb);
5274 static struct sk_buff *tcp_skb_next(struct sk_buff *skb, struct sk_buff_head *list)
5277 return !skb_queue_is_last(list, skb) ? skb->next : NULL;
5279 return skb_rb_next(skb);
5282 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
5283 struct sk_buff_head *list,
5284 struct rb_root *root)
5286 struct sk_buff *next = tcp_skb_next(skb, list);
5289 __skb_unlink(skb, list);
5291 rb_erase(&skb->rbnode, root);
5294 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
5299 /* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */
5300 void tcp_rbtree_insert(struct rb_root *root, struct sk_buff *skb)
5302 struct rb_node **p = &root->rb_node;
5303 struct rb_node *parent = NULL;
5304 struct sk_buff *skb1;
5308 skb1 = rb_to_skb(parent);
5309 if (before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb1)->seq))
5310 p = &parent->rb_left;
5312 p = &parent->rb_right;
5314 rb_link_node(&skb->rbnode, parent, p);
5315 rb_insert_color(&skb->rbnode, root);
5318 /* Collapse contiguous sequence of skbs head..tail with
5319 * sequence numbers start..end.
5321 * If tail is NULL, this means until the end of the queue.
5323 * Segments with FIN/SYN are not collapsed (only because this
5327 tcp_collapse(struct sock *sk, struct sk_buff_head *list, struct rb_root *root,
5328 struct sk_buff *head, struct sk_buff *tail, u32 start, u32 end)
5330 struct sk_buff *skb = head, *n;
5331 struct sk_buff_head tmp;
5334 /* First, check that queue is collapsible and find
5335 * the point where collapsing can be useful.
5338 for (end_of_skbs = true; skb != NULL && skb != tail; skb = n) {
5339 n = tcp_skb_next(skb, list);
5341 /* No new bits? It is possible on ofo queue. */
5342 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
5343 skb = tcp_collapse_one(sk, skb, list, root);
5349 /* The first skb to collapse is:
5351 * - bloated or contains data before "start" or
5352 * overlaps to the next one and mptcp allow collapsing.
5354 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
5355 (tcp_win_from_space(sk, skb->truesize) > skb->len ||
5356 before(TCP_SKB_CB(skb)->seq, start))) {
5357 end_of_skbs = false;
5361 if (n && n != tail && mptcp_skb_can_collapse(skb, n) &&
5362 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(n)->seq) {
5363 end_of_skbs = false;
5367 /* Decided to skip this, advance start seq. */
5368 start = TCP_SKB_CB(skb)->end_seq;
5371 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
5374 __skb_queue_head_init(&tmp);
5376 while (before(start, end)) {
5377 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
5378 struct sk_buff *nskb;
5380 nskb = alloc_skb(copy, GFP_ATOMIC);
5384 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
5385 #ifdef CONFIG_TLS_DEVICE
5386 nskb->decrypted = skb->decrypted;
5388 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
5390 __skb_queue_before(list, skb, nskb);
5392 __skb_queue_tail(&tmp, nskb); /* defer rbtree insertion */
5393 skb_set_owner_r(nskb, sk);
5394 mptcp_skb_ext_move(nskb, skb);
5396 /* Copy data, releasing collapsed skbs. */
5398 int offset = start - TCP_SKB_CB(skb)->seq;
5399 int size = TCP_SKB_CB(skb)->end_seq - start;
5403 size = min(copy, size);
5404 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
5406 TCP_SKB_CB(nskb)->end_seq += size;
5410 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
5411 skb = tcp_collapse_one(sk, skb, list, root);
5414 !mptcp_skb_can_collapse(nskb, skb) ||
5415 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
5417 #ifdef CONFIG_TLS_DEVICE
5418 if (skb->decrypted != nskb->decrypted)
5425 skb_queue_walk_safe(&tmp, skb, n)
5426 tcp_rbtree_insert(root, skb);
5429 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
5430 * and tcp_collapse() them until all the queue is collapsed.
5432 static void tcp_collapse_ofo_queue(struct sock *sk)
5434 struct tcp_sock *tp = tcp_sk(sk);
5435 u32 range_truesize, sum_tiny = 0;
5436 struct sk_buff *skb, *head;
5439 skb = skb_rb_first(&tp->out_of_order_queue);
5442 tp->ooo_last_skb = skb_rb_last(&tp->out_of_order_queue);
5445 start = TCP_SKB_CB(skb)->seq;
5446 end = TCP_SKB_CB(skb)->end_seq;
5447 range_truesize = skb->truesize;
5449 for (head = skb;;) {
5450 skb = skb_rb_next(skb);
5452 /* Range is terminated when we see a gap or when
5453 * we are at the queue end.
5456 after(TCP_SKB_CB(skb)->seq, end) ||
5457 before(TCP_SKB_CB(skb)->end_seq, start)) {
5458 /* Do not attempt collapsing tiny skbs */
5459 if (range_truesize != head->truesize ||
5460 end - start >= SKB_WITH_OVERHEAD(PAGE_SIZE)) {
5461 tcp_collapse(sk, NULL, &tp->out_of_order_queue,
5462 head, skb, start, end);
5464 sum_tiny += range_truesize;
5465 if (sum_tiny > sk->sk_rcvbuf >> 3)
5471 range_truesize += skb->truesize;
5472 if (unlikely(before(TCP_SKB_CB(skb)->seq, start)))
5473 start = TCP_SKB_CB(skb)->seq;
5474 if (after(TCP_SKB_CB(skb)->end_seq, end))
5475 end = TCP_SKB_CB(skb)->end_seq;
5480 * Clean the out-of-order queue to make room.
5481 * We drop high sequences packets to :
5482 * 1) Let a chance for holes to be filled.
5483 * This means we do not drop packets from ooo queue if their sequence
5484 * is before incoming packet sequence.
5485 * 2) not add too big latencies if thousands of packets sit there.
5486 * (But if application shrinks SO_RCVBUF, we could still end up
5487 * freeing whole queue here)
5488 * 3) Drop at least 12.5 % of sk_rcvbuf to avoid malicious attacks.
5490 * Return true if queue has shrunk.
5492 static bool tcp_prune_ofo_queue(struct sock *sk, const struct sk_buff *in_skb)
5494 struct tcp_sock *tp = tcp_sk(sk);
5495 struct rb_node *node, *prev;
5496 bool pruned = false;
5499 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
5502 goal = sk->sk_rcvbuf >> 3;
5503 node = &tp->ooo_last_skb->rbnode;
5506 struct sk_buff *skb = rb_to_skb(node);
5508 /* If incoming skb would land last in ofo queue, stop pruning. */
5509 if (after(TCP_SKB_CB(in_skb)->seq, TCP_SKB_CB(skb)->seq))
5512 prev = rb_prev(node);
5513 rb_erase(node, &tp->out_of_order_queue);
5514 goal -= skb->truesize;
5515 tcp_drop_reason(sk, skb, SKB_DROP_REASON_TCP_OFO_QUEUE_PRUNE);
5516 tp->ooo_last_skb = rb_to_skb(prev);
5517 if (!prev || goal <= 0) {
5518 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf &&
5519 !tcp_under_memory_pressure(sk))
5521 goal = sk->sk_rcvbuf >> 3;
5527 NET_INC_STATS(sock_net(sk), LINUX_MIB_OFOPRUNED);
5528 /* Reset SACK state. A conforming SACK implementation will
5529 * do the same at a timeout based retransmit. When a connection
5530 * is in a sad state like this, we care only about integrity
5531 * of the connection not performance.
5533 if (tp->rx_opt.sack_ok)
5534 tcp_sack_reset(&tp->rx_opt);
5539 /* Reduce allocated memory if we can, trying to get
5540 * the socket within its memory limits again.
5542 * Return less than zero if we should start dropping frames
5543 * until the socket owning process reads some of the data
5544 * to stabilize the situation.
5546 static int tcp_prune_queue(struct sock *sk, const struct sk_buff *in_skb)
5548 struct tcp_sock *tp = tcp_sk(sk);
5550 NET_INC_STATS(sock_net(sk), LINUX_MIB_PRUNECALLED);
5552 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
5553 tcp_clamp_window(sk);
5554 else if (tcp_under_memory_pressure(sk))
5555 tcp_adjust_rcv_ssthresh(sk);
5557 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5560 tcp_collapse_ofo_queue(sk);
5561 if (!skb_queue_empty(&sk->sk_receive_queue))
5562 tcp_collapse(sk, &sk->sk_receive_queue, NULL,
5563 skb_peek(&sk->sk_receive_queue),
5565 tp->copied_seq, tp->rcv_nxt);
5567 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5570 /* Collapsing did not help, destructive actions follow.
5571 * This must not ever occur. */
5573 tcp_prune_ofo_queue(sk, in_skb);
5575 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5578 /* If we are really being abused, tell the caller to silently
5579 * drop receive data on the floor. It will get retransmitted
5580 * and hopefully then we'll have sufficient space.
5582 NET_INC_STATS(sock_net(sk), LINUX_MIB_RCVPRUNED);
5584 /* Massive buffer overcommit. */
5589 static bool tcp_should_expand_sndbuf(struct sock *sk)
5591 const struct tcp_sock *tp = tcp_sk(sk);
5593 /* If the user specified a specific send buffer setting, do
5596 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
5599 /* If we are under global TCP memory pressure, do not expand. */
5600 if (tcp_under_memory_pressure(sk)) {
5601 int unused_mem = sk_unused_reserved_mem(sk);
5603 /* Adjust sndbuf according to reserved mem. But make sure
5604 * it never goes below SOCK_MIN_SNDBUF.
5605 * See sk_stream_moderate_sndbuf() for more details.
5607 if (unused_mem > SOCK_MIN_SNDBUF)
5608 WRITE_ONCE(sk->sk_sndbuf, unused_mem);
5613 /* If we are under soft global TCP memory pressure, do not expand. */
5614 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
5617 /* If we filled the congestion window, do not expand. */
5618 if (tcp_packets_in_flight(tp) >= tcp_snd_cwnd(tp))
5624 static void tcp_new_space(struct sock *sk)
5626 struct tcp_sock *tp = tcp_sk(sk);
5628 if (tcp_should_expand_sndbuf(sk)) {
5629 tcp_sndbuf_expand(sk);
5630 tp->snd_cwnd_stamp = tcp_jiffies32;
5633 INDIRECT_CALL_1(sk->sk_write_space, sk_stream_write_space, sk);
5636 /* Caller made space either from:
5637 * 1) Freeing skbs in rtx queues (after tp->snd_una has advanced)
5638 * 2) Sent skbs from output queue (and thus advancing tp->snd_nxt)
5640 * We might be able to generate EPOLLOUT to the application if:
5641 * 1) Space consumed in output/rtx queues is below sk->sk_sndbuf/2
5642 * 2) notsent amount (tp->write_seq - tp->snd_nxt) became
5643 * small enough that tcp_stream_memory_free() decides it
5644 * is time to generate EPOLLOUT.
5646 void tcp_check_space(struct sock *sk)
5648 /* pairs with tcp_poll() */
5650 if (sk->sk_socket &&
5651 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
5653 if (!test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
5654 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
5658 static inline void tcp_data_snd_check(struct sock *sk)
5660 tcp_push_pending_frames(sk);
5661 tcp_check_space(sk);
5665 * Check if sending an ack is needed.
5667 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
5669 struct tcp_sock *tp = tcp_sk(sk);
5670 unsigned long rtt, delay;
5672 /* More than one full frame received... */
5673 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
5674 /* ... and right edge of window advances far enough.
5675 * (tcp_recvmsg() will send ACK otherwise).
5676 * If application uses SO_RCVLOWAT, we want send ack now if
5677 * we have not received enough bytes to satisfy the condition.
5679 (tp->rcv_nxt - tp->copied_seq < sk->sk_rcvlowat ||
5680 __tcp_select_window(sk) >= tp->rcv_wnd)) ||
5681 /* We ACK each frame or... */
5682 tcp_in_quickack_mode(sk) ||
5683 /* Protocol state mandates a one-time immediate ACK */
5684 inet_csk(sk)->icsk_ack.pending & ICSK_ACK_NOW) {
5685 /* If we are running from __release_sock() in user context,
5686 * Defer the ack until tcp_release_cb().
5688 if (sock_owned_by_user_nocheck(sk) &&
5689 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_backlog_ack_defer)) {
5690 set_bit(TCP_ACK_DEFERRED, &sk->sk_tsq_flags);
5698 if (!ofo_possible || RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
5699 tcp_send_delayed_ack(sk);
5703 if (!tcp_is_sack(tp) ||
5704 tp->compressed_ack >= READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_nr))
5707 if (tp->compressed_ack_rcv_nxt != tp->rcv_nxt) {
5708 tp->compressed_ack_rcv_nxt = tp->rcv_nxt;
5709 tp->dup_ack_counter = 0;
5711 if (tp->dup_ack_counter < TCP_FASTRETRANS_THRESH) {
5712 tp->dup_ack_counter++;
5715 tp->compressed_ack++;
5716 if (hrtimer_is_queued(&tp->compressed_ack_timer))
5719 /* compress ack timer : 5 % of rtt, but no more than tcp_comp_sack_delay_ns */
5721 rtt = tp->rcv_rtt_est.rtt_us;
5722 if (tp->srtt_us && tp->srtt_us < rtt)
5725 delay = min_t(unsigned long,
5726 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_delay_ns),
5727 rtt * (NSEC_PER_USEC >> 3)/20);
5729 hrtimer_start_range_ns(&tp->compressed_ack_timer, ns_to_ktime(delay),
5730 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_slack_ns),
5731 HRTIMER_MODE_REL_PINNED_SOFT);
5734 static inline void tcp_ack_snd_check(struct sock *sk)
5736 if (!inet_csk_ack_scheduled(sk)) {
5737 /* We sent a data segment already. */
5740 __tcp_ack_snd_check(sk, 1);
5744 * This routine is only called when we have urgent data
5745 * signaled. Its the 'slow' part of tcp_urg. It could be
5746 * moved inline now as tcp_urg is only called from one
5747 * place. We handle URGent data wrong. We have to - as
5748 * BSD still doesn't use the correction from RFC961.
5749 * For 1003.1g we should support a new option TCP_STDURG to permit
5750 * either form (or just set the sysctl tcp_stdurg).
5753 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
5755 struct tcp_sock *tp = tcp_sk(sk);
5756 u32 ptr = ntohs(th->urg_ptr);
5758 if (ptr && !READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_stdurg))
5760 ptr += ntohl(th->seq);
5762 /* Ignore urgent data that we've already seen and read. */
5763 if (after(tp->copied_seq, ptr))
5766 /* Do not replay urg ptr.
5768 * NOTE: interesting situation not covered by specs.
5769 * Misbehaving sender may send urg ptr, pointing to segment,
5770 * which we already have in ofo queue. We are not able to fetch
5771 * such data and will stay in TCP_URG_NOTYET until will be eaten
5772 * by recvmsg(). Seems, we are not obliged to handle such wicked
5773 * situations. But it is worth to think about possibility of some
5774 * DoSes using some hypothetical application level deadlock.
5776 if (before(ptr, tp->rcv_nxt))
5779 /* Do we already have a newer (or duplicate) urgent pointer? */
5780 if (tp->urg_data && !after(ptr, tp->urg_seq))
5783 /* Tell the world about our new urgent pointer. */
5786 /* We may be adding urgent data when the last byte read was
5787 * urgent. To do this requires some care. We cannot just ignore
5788 * tp->copied_seq since we would read the last urgent byte again
5789 * as data, nor can we alter copied_seq until this data arrives
5790 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5792 * NOTE. Double Dutch. Rendering to plain English: author of comment
5793 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5794 * and expect that both A and B disappear from stream. This is _wrong_.
5795 * Though this happens in BSD with high probability, this is occasional.
5796 * Any application relying on this is buggy. Note also, that fix "works"
5797 * only in this artificial test. Insert some normal data between A and B and we will
5798 * decline of BSD again. Verdict: it is better to remove to trap
5801 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5802 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5803 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5805 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5806 __skb_unlink(skb, &sk->sk_receive_queue);
5811 WRITE_ONCE(tp->urg_data, TCP_URG_NOTYET);
5812 WRITE_ONCE(tp->urg_seq, ptr);
5814 /* Disable header prediction. */
5818 /* This is the 'fast' part of urgent handling. */
5819 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
5821 struct tcp_sock *tp = tcp_sk(sk);
5823 /* Check if we get a new urgent pointer - normally not. */
5824 if (unlikely(th->urg))
5825 tcp_check_urg(sk, th);
5827 /* Do we wait for any urgent data? - normally not... */
5828 if (unlikely(tp->urg_data == TCP_URG_NOTYET)) {
5829 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5832 /* Is the urgent pointer pointing into this packet? */
5833 if (ptr < skb->len) {
5835 if (skb_copy_bits(skb, ptr, &tmp, 1))
5837 WRITE_ONCE(tp->urg_data, TCP_URG_VALID | tmp);
5838 if (!sock_flag(sk, SOCK_DEAD))
5839 sk->sk_data_ready(sk);
5844 /* Accept RST for rcv_nxt - 1 after a FIN.
5845 * When tcp connections are abruptly terminated from Mac OSX (via ^C), a
5846 * FIN is sent followed by a RST packet. The RST is sent with the same
5847 * sequence number as the FIN, and thus according to RFC 5961 a challenge
5848 * ACK should be sent. However, Mac OSX rate limits replies to challenge
5849 * ACKs on the closed socket. In addition middleboxes can drop either the
5850 * challenge ACK or a subsequent RST.
5852 static bool tcp_reset_check(const struct sock *sk, const struct sk_buff *skb)
5854 const struct tcp_sock *tp = tcp_sk(sk);
5856 return unlikely(TCP_SKB_CB(skb)->seq == (tp->rcv_nxt - 1) &&
5857 (1 << sk->sk_state) & (TCPF_CLOSE_WAIT | TCPF_LAST_ACK |
5861 /* Does PAWS and seqno based validation of an incoming segment, flags will
5862 * play significant role here.
5864 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5865 const struct tcphdr *th, int syn_inerr)
5867 struct tcp_sock *tp = tcp_sk(sk);
5870 /* RFC1323: H1. Apply PAWS check first. */
5871 if (tcp_fast_parse_options(sock_net(sk), skb, th, tp) &&
5872 tp->rx_opt.saw_tstamp &&
5873 tcp_paws_discard(sk, skb)) {
5875 if (unlikely(th->syn))
5877 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5878 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5879 LINUX_MIB_TCPACKSKIPPEDPAWS,
5880 &tp->last_oow_ack_time))
5881 tcp_send_dupack(sk, skb);
5882 SKB_DR_SET(reason, TCP_RFC7323_PAWS);
5885 /* Reset is accepted even if it did not pass PAWS. */
5888 /* Step 1: check sequence number */
5889 reason = tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
5891 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5892 * (RST) segments are validated by checking their SEQ-fields."
5893 * And page 69: "If an incoming segment is not acceptable,
5894 * an acknowledgment should be sent in reply (unless the RST
5895 * bit is set, if so drop the segment and return)".
5900 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5901 LINUX_MIB_TCPACKSKIPPEDSEQ,
5902 &tp->last_oow_ack_time))
5903 tcp_send_dupack(sk, skb);
5904 } else if (tcp_reset_check(sk, skb)) {
5910 /* Step 2: check RST bit */
5912 /* RFC 5961 3.2 (extend to match against (RCV.NXT - 1) after a
5913 * FIN and SACK too if available):
5914 * If seq num matches RCV.NXT or (RCV.NXT - 1) after a FIN, or
5915 * the right-most SACK block,
5917 * RESET the connection
5919 * Send a challenge ACK
5921 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt ||
5922 tcp_reset_check(sk, skb))
5925 if (tcp_is_sack(tp) && tp->rx_opt.num_sacks > 0) {
5926 struct tcp_sack_block *sp = &tp->selective_acks[0];
5927 int max_sack = sp[0].end_seq;
5930 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;
5932 max_sack = after(sp[this_sack].end_seq,
5934 sp[this_sack].end_seq : max_sack;
5937 if (TCP_SKB_CB(skb)->seq == max_sack)
5941 /* Disable TFO if RST is out-of-order
5942 * and no data has been received
5943 * for current active TFO socket
5945 if (tp->syn_fastopen && !tp->data_segs_in &&
5946 sk->sk_state == TCP_ESTABLISHED)
5947 tcp_fastopen_active_disable(sk);
5948 tcp_send_challenge_ack(sk);
5949 SKB_DR_SET(reason, TCP_RESET);
5953 /* step 3: check security and precedence [ignored] */
5955 /* step 4: Check for a SYN
5956 * RFC 5961 4.2 : Send a challenge ack
5961 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5962 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5963 tcp_send_challenge_ack(sk);
5964 SKB_DR_SET(reason, TCP_INVALID_SYN);
5968 bpf_skops_parse_hdr(sk, skb);
5973 tcp_drop_reason(sk, skb, reason);
5983 * TCP receive function for the ESTABLISHED state.
5985 * It is split into a fast path and a slow path. The fast path is
5987 * - A zero window was announced from us - zero window probing
5988 * is only handled properly in the slow path.
5989 * - Out of order segments arrived.
5990 * - Urgent data is expected.
5991 * - There is no buffer space left
5992 * - Unexpected TCP flags/window values/header lengths are received
5993 * (detected by checking the TCP header against pred_flags)
5994 * - Data is sent in both directions. Fast path only supports pure senders
5995 * or pure receivers (this means either the sequence number or the ack
5996 * value must stay constant)
5997 * - Unexpected TCP option.
5999 * When these conditions are not satisfied it drops into a standard
6000 * receive procedure patterned after RFC793 to handle all cases.
6001 * The first three cases are guaranteed by proper pred_flags setting,
6002 * the rest is checked inline. Fast processing is turned on in
6003 * tcp_data_queue when everything is OK.
6005 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb)
6007 enum skb_drop_reason reason = SKB_DROP_REASON_NOT_SPECIFIED;
6008 const struct tcphdr *th = (const struct tcphdr *)skb->data;
6009 struct tcp_sock *tp = tcp_sk(sk);
6010 unsigned int len = skb->len;
6012 /* TCP congestion window tracking */
6013 trace_tcp_probe(sk, skb);
6015 tcp_mstamp_refresh(tp);
6016 if (unlikely(!rcu_access_pointer(sk->sk_rx_dst)))
6017 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
6019 * Header prediction.
6020 * The code loosely follows the one in the famous
6021 * "30 instruction TCP receive" Van Jacobson mail.
6023 * Van's trick is to deposit buffers into socket queue
6024 * on a device interrupt, to call tcp_recv function
6025 * on the receive process context and checksum and copy
6026 * the buffer to user space. smart...
6028 * Our current scheme is not silly either but we take the
6029 * extra cost of the net_bh soft interrupt processing...
6030 * We do checksum and copy also but from device to kernel.
6033 tp->rx_opt.saw_tstamp = 0;
6035 /* pred_flags is 0xS?10 << 16 + snd_wnd
6036 * if header_prediction is to be made
6037 * 'S' will always be tp->tcp_header_len >> 2
6038 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
6039 * turn it off (when there are holes in the receive
6040 * space for instance)
6041 * PSH flag is ignored.
6044 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
6045 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
6046 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
6047 int tcp_header_len = tp->tcp_header_len;
6049 /* Timestamp header prediction: tcp_header_len
6050 * is automatically equal to th->doff*4 due to pred_flags
6054 /* Check timestamp */
6055 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
6056 /* No? Slow path! */
6057 if (!tcp_parse_aligned_timestamp(tp, th))
6060 /* If PAWS failed, check it more carefully in slow path */
6061 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
6064 /* DO NOT update ts_recent here, if checksum fails
6065 * and timestamp was corrupted part, it will result
6066 * in a hung connection since we will drop all
6067 * future packets due to the PAWS test.
6071 if (len <= tcp_header_len) {
6072 /* Bulk data transfer: sender */
6073 if (len == tcp_header_len) {
6074 /* Predicted packet is in window by definition.
6075 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
6076 * Hence, check seq<=rcv_wup reduces to:
6078 if (tcp_header_len ==
6079 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
6080 tp->rcv_nxt == tp->rcv_wup)
6081 tcp_store_ts_recent(tp);
6083 /* We know that such packets are checksummed
6086 tcp_ack(sk, skb, 0);
6088 tcp_data_snd_check(sk);
6089 /* When receiving pure ack in fast path, update
6090 * last ts ecr directly instead of calling
6091 * tcp_rcv_rtt_measure_ts()
6093 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr;
6095 } else { /* Header too small */
6096 reason = SKB_DROP_REASON_PKT_TOO_SMALL;
6097 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
6102 bool fragstolen = false;
6104 if (tcp_checksum_complete(skb))
6107 if ((int)skb->truesize > sk->sk_forward_alloc)
6110 /* Predicted packet is in window by definition.
6111 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
6112 * Hence, check seq<=rcv_wup reduces to:
6114 if (tcp_header_len ==
6115 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
6116 tp->rcv_nxt == tp->rcv_wup)
6117 tcp_store_ts_recent(tp);
6119 tcp_rcv_rtt_measure_ts(sk, skb);
6121 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPHITS);
6123 /* Bulk data transfer: receiver */
6125 __skb_pull(skb, tcp_header_len);
6126 eaten = tcp_queue_rcv(sk, skb, &fragstolen);
6128 tcp_event_data_recv(sk, skb);
6130 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
6131 /* Well, only one small jumplet in fast path... */
6132 tcp_ack(sk, skb, FLAG_DATA);
6133 tcp_data_snd_check(sk);
6134 if (!inet_csk_ack_scheduled(sk))
6137 tcp_update_wl(tp, TCP_SKB_CB(skb)->seq);
6140 __tcp_ack_snd_check(sk, 0);
6143 kfree_skb_partial(skb, fragstolen);
6150 if (len < (th->doff << 2) || tcp_checksum_complete(skb))
6153 if (!th->ack && !th->rst && !th->syn) {
6154 reason = SKB_DROP_REASON_TCP_FLAGS;
6159 * Standard slow path.
6162 if (!tcp_validate_incoming(sk, skb, th, 1))
6166 reason = tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT);
6167 if ((int)reason < 0) {
6171 tcp_rcv_rtt_measure_ts(sk, skb);
6173 /* Process urgent data. */
6174 tcp_urg(sk, skb, th);
6176 /* step 7: process the segment text */
6177 tcp_data_queue(sk, skb);
6179 tcp_data_snd_check(sk);
6180 tcp_ack_snd_check(sk);
6184 reason = SKB_DROP_REASON_TCP_CSUM;
6185 trace_tcp_bad_csum(skb);
6186 TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS);
6187 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
6190 tcp_drop_reason(sk, skb, reason);
6192 EXPORT_SYMBOL(tcp_rcv_established);
6194 void tcp_init_transfer(struct sock *sk, int bpf_op, struct sk_buff *skb)
6196 struct inet_connection_sock *icsk = inet_csk(sk);
6197 struct tcp_sock *tp = tcp_sk(sk);
6200 icsk->icsk_af_ops->rebuild_header(sk);
6201 tcp_init_metrics(sk);
6203 /* Initialize the congestion window to start the transfer.
6204 * Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
6205 * retransmitted. In light of RFC6298 more aggressive 1sec
6206 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
6207 * retransmission has occurred.
6209 if (tp->total_retrans > 1 && tp->undo_marker)
6210 tcp_snd_cwnd_set(tp, 1);
6212 tcp_snd_cwnd_set(tp, tcp_init_cwnd(tp, __sk_dst_get(sk)));
6213 tp->snd_cwnd_stamp = tcp_jiffies32;
6215 bpf_skops_established(sk, bpf_op, skb);
6216 /* Initialize congestion control unless BPF initialized it already: */
6217 if (!icsk->icsk_ca_initialized)
6218 tcp_init_congestion_control(sk);
6219 tcp_init_buffer_space(sk);
6222 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
6224 struct tcp_sock *tp = tcp_sk(sk);
6225 struct inet_connection_sock *icsk = inet_csk(sk);
6227 tcp_ao_finish_connect(sk, skb);
6228 tcp_set_state(sk, TCP_ESTABLISHED);
6229 icsk->icsk_ack.lrcvtime = tcp_jiffies32;
6232 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
6233 security_inet_conn_established(sk, skb);
6234 sk_mark_napi_id(sk, skb);
6237 tcp_init_transfer(sk, BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB, skb);
6239 /* Prevent spurious tcp_cwnd_restart() on first data
6242 tp->lsndtime = tcp_jiffies32;
6244 if (sock_flag(sk, SOCK_KEEPOPEN))
6245 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
6247 if (!tp->rx_opt.snd_wscale)
6248 __tcp_fast_path_on(tp, tp->snd_wnd);
6253 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
6254 struct tcp_fastopen_cookie *cookie)
6256 struct tcp_sock *tp = tcp_sk(sk);
6257 struct sk_buff *data = tp->syn_data ? tcp_rtx_queue_head(sk) : NULL;
6258 u16 mss = tp->rx_opt.mss_clamp, try_exp = 0;
6259 bool syn_drop = false;
6261 if (mss == tp->rx_opt.user_mss) {
6262 struct tcp_options_received opt;
6264 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
6265 tcp_clear_options(&opt);
6266 opt.user_mss = opt.mss_clamp = 0;
6267 tcp_parse_options(sock_net(sk), synack, &opt, 0, NULL);
6268 mss = opt.mss_clamp;
6271 if (!tp->syn_fastopen) {
6272 /* Ignore an unsolicited cookie */
6274 } else if (tp->total_retrans) {
6275 /* SYN timed out and the SYN-ACK neither has a cookie nor
6276 * acknowledges data. Presumably the remote received only
6277 * the retransmitted (regular) SYNs: either the original
6278 * SYN-data or the corresponding SYN-ACK was dropped.
6280 syn_drop = (cookie->len < 0 && data);
6281 } else if (cookie->len < 0 && !tp->syn_data) {
6282 /* We requested a cookie but didn't get it. If we did not use
6283 * the (old) exp opt format then try so next time (try_exp=1).
6284 * Otherwise we go back to use the RFC7413 opt (try_exp=2).
6286 try_exp = tp->syn_fastopen_exp ? 2 : 1;
6289 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp);
6291 if (data) { /* Retransmit unacked data in SYN */
6292 if (tp->total_retrans)
6293 tp->fastopen_client_fail = TFO_SYN_RETRANSMITTED;
6295 tp->fastopen_client_fail = TFO_DATA_NOT_ACKED;
6296 skb_rbtree_walk_from(data)
6297 tcp_mark_skb_lost(sk, data);
6298 tcp_xmit_retransmit_queue(sk);
6299 NET_INC_STATS(sock_net(sk),
6300 LINUX_MIB_TCPFASTOPENACTIVEFAIL);
6303 tp->syn_data_acked = tp->syn_data;
6304 if (tp->syn_data_acked) {
6305 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
6306 /* SYN-data is counted as two separate packets in tcp_ack() */
6307 if (tp->delivered > 1)
6311 tcp_fastopen_add_skb(sk, synack);
6316 static void smc_check_reset_syn(struct tcp_sock *tp)
6318 #if IS_ENABLED(CONFIG_SMC)
6319 if (static_branch_unlikely(&tcp_have_smc)) {
6320 if (tp->syn_smc && !tp->rx_opt.smc_ok)
6326 static void tcp_try_undo_spurious_syn(struct sock *sk)
6328 struct tcp_sock *tp = tcp_sk(sk);
6331 /* undo_marker is set when SYN or SYNACK times out. The timeout is
6332 * spurious if the ACK's timestamp option echo value matches the
6333 * original SYN timestamp.
6335 syn_stamp = tp->retrans_stamp;
6336 if (tp->undo_marker && syn_stamp && tp->rx_opt.saw_tstamp &&
6337 syn_stamp == tp->rx_opt.rcv_tsecr)
6338 tp->undo_marker = 0;
6341 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
6342 const struct tcphdr *th)
6344 struct inet_connection_sock *icsk = inet_csk(sk);
6345 struct tcp_sock *tp = tcp_sk(sk);
6346 struct tcp_fastopen_cookie foc = { .len = -1 };
6347 int saved_clamp = tp->rx_opt.mss_clamp;
6351 tcp_parse_options(sock_net(sk), skb, &tp->rx_opt, 0, &foc);
6352 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
6353 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
6357 * "If the state is SYN-SENT then
6358 * first check the ACK bit
6359 * If the ACK bit is set
6360 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
6361 * a reset (unless the RST bit is set, if so drop
6362 * the segment and return)"
6364 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
6365 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
6366 /* Previous FIN/ACK or RST/ACK might be ignored. */
6367 if (icsk->icsk_retransmits == 0)
6368 inet_csk_reset_xmit_timer(sk,
6370 TCP_TIMEOUT_MIN, TCP_RTO_MAX);
6371 goto reset_and_undo;
6374 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
6375 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
6376 tcp_time_stamp_ts(tp))) {
6377 NET_INC_STATS(sock_net(sk),
6378 LINUX_MIB_PAWSACTIVEREJECTED);
6379 goto reset_and_undo;
6382 /* Now ACK is acceptable.
6384 * "If the RST bit is set
6385 * If the ACK was acceptable then signal the user "error:
6386 * connection reset", drop the segment, enter CLOSED state,
6387 * delete TCB, and return."
6398 * "fifth, if neither of the SYN or RST bits is set then
6399 * drop the segment and return."
6405 SKB_DR_SET(reason, TCP_FLAGS);
6406 goto discard_and_undo;
6409 * "If the SYN bit is on ...
6410 * are acceptable then ...
6411 * (our SYN has been ACKed), change the connection
6412 * state to ESTABLISHED..."
6415 tcp_ecn_rcv_synack(tp, th);
6417 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
6418 tcp_try_undo_spurious_syn(sk);
6419 tcp_ack(sk, skb, FLAG_SLOWPATH);
6421 /* Ok.. it's good. Set up sequence numbers and
6422 * move to established.
6424 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1);
6425 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
6427 /* RFC1323: The window in SYN & SYN/ACK segments is
6430 tp->snd_wnd = ntohs(th->window);
6432 if (!tp->rx_opt.wscale_ok) {
6433 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
6434 tp->window_clamp = min(tp->window_clamp, 65535U);
6437 if (tp->rx_opt.saw_tstamp) {
6438 tp->rx_opt.tstamp_ok = 1;
6439 tp->tcp_header_len =
6440 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
6441 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
6442 tcp_store_ts_recent(tp);
6444 tp->tcp_header_len = sizeof(struct tcphdr);
6447 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
6448 tcp_initialize_rcv_mss(sk);
6450 /* Remember, tcp_poll() does not lock socket!
6451 * Change state from SYN-SENT only after copied_seq
6452 * is initialized. */
6453 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6455 smc_check_reset_syn(tp);
6459 tcp_finish_connect(sk, skb);
6461 fastopen_fail = (tp->syn_fastopen || tp->syn_data) &&
6462 tcp_rcv_fastopen_synack(sk, skb, &foc);
6464 if (!sock_flag(sk, SOCK_DEAD)) {
6465 sk->sk_state_change(sk);
6466 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
6470 if (sk->sk_write_pending ||
6471 READ_ONCE(icsk->icsk_accept_queue.rskq_defer_accept) ||
6472 inet_csk_in_pingpong_mode(sk)) {
6473 /* Save one ACK. Data will be ready after
6474 * several ticks, if write_pending is set.
6476 * It may be deleted, but with this feature tcpdumps
6477 * look so _wonderfully_ clever, that I was not able
6478 * to stand against the temptation 8) --ANK
6480 inet_csk_schedule_ack(sk);
6481 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
6482 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
6483 TCP_DELACK_MAX, TCP_RTO_MAX);
6490 /* No ACK in the segment */
6494 * "If the RST bit is set
6496 * Otherwise (no ACK) drop the segment and return."
6498 SKB_DR_SET(reason, TCP_RESET);
6499 goto discard_and_undo;
6503 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
6504 tcp_paws_reject(&tp->rx_opt, 0)) {
6505 SKB_DR_SET(reason, TCP_RFC7323_PAWS);
6506 goto discard_and_undo;
6509 /* We see SYN without ACK. It is attempt of
6510 * simultaneous connect with crossed SYNs.
6511 * Particularly, it can be connect to self.
6513 #ifdef CONFIG_TCP_AO
6514 struct tcp_ao_info *ao;
6516 ao = rcu_dereference_protected(tp->ao_info,
6517 lockdep_sock_is_held(sk));
6519 WRITE_ONCE(ao->risn, th->seq);
6523 tcp_set_state(sk, TCP_SYN_RECV);
6525 if (tp->rx_opt.saw_tstamp) {
6526 tp->rx_opt.tstamp_ok = 1;
6527 tcp_store_ts_recent(tp);
6528 tp->tcp_header_len =
6529 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
6531 tp->tcp_header_len = sizeof(struct tcphdr);
6534 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1);
6535 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6536 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
6538 /* RFC1323: The window in SYN & SYN/ACK segments is
6541 tp->snd_wnd = ntohs(th->window);
6542 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
6543 tp->max_window = tp->snd_wnd;
6545 tcp_ecn_rcv_syn(tp, th);
6548 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
6549 tcp_initialize_rcv_mss(sk);
6551 tcp_send_synack(sk);
6553 /* Note, we could accept data and URG from this segment.
6554 * There are no obstacles to make this (except that we must
6555 * either change tcp_recvmsg() to prevent it from returning data
6556 * before 3WHS completes per RFC793, or employ TCP Fast Open).
6558 * However, if we ignore data in ACKless segments sometimes,
6559 * we have no reasons to accept it sometimes.
6560 * Also, seems the code doing it in step6 of tcp_rcv_state_process
6561 * is not flawless. So, discard packet for sanity.
6562 * Uncomment this return to process the data.
6569 /* "fifth, if neither of the SYN or RST bits is set then
6570 * drop the segment and return."
6574 tcp_clear_options(&tp->rx_opt);
6575 tp->rx_opt.mss_clamp = saved_clamp;
6576 tcp_drop_reason(sk, skb, reason);
6580 tcp_clear_options(&tp->rx_opt);
6581 tp->rx_opt.mss_clamp = saved_clamp;
6585 static void tcp_rcv_synrecv_state_fastopen(struct sock *sk)
6587 struct tcp_sock *tp = tcp_sk(sk);
6588 struct request_sock *req;
6590 /* If we are still handling the SYNACK RTO, see if timestamp ECR allows
6591 * undo. If peer SACKs triggered fast recovery, we can't undo here.
6593 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss && !tp->packets_out)
6594 tcp_try_undo_recovery(sk);
6596 /* Reset rtx states to prevent spurious retransmits_timed_out() */
6597 tcp_update_rto_time(tp);
6598 tp->retrans_stamp = 0;
6599 inet_csk(sk)->icsk_retransmits = 0;
6601 /* Once we leave TCP_SYN_RECV or TCP_FIN_WAIT_1,
6602 * we no longer need req so release it.
6604 req = rcu_dereference_protected(tp->fastopen_rsk,
6605 lockdep_sock_is_held(sk));
6606 reqsk_fastopen_remove(sk, req, false);
6608 /* Re-arm the timer because data may have been sent out.
6609 * This is similar to the regular data transmission case
6610 * when new data has just been ack'ed.
6612 * (TFO) - we could try to be more aggressive and
6613 * retransmitting any data sooner based on when they
6620 * This function implements the receiving procedure of RFC 793 for
6621 * all states except ESTABLISHED and TIME_WAIT.
6622 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
6623 * address independent.
6626 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb)
6628 struct tcp_sock *tp = tcp_sk(sk);
6629 struct inet_connection_sock *icsk = inet_csk(sk);
6630 const struct tcphdr *th = tcp_hdr(skb);
6631 struct request_sock *req;
6636 switch (sk->sk_state) {
6638 SKB_DR_SET(reason, TCP_CLOSE);
6646 SKB_DR_SET(reason, TCP_RESET);
6651 SKB_DR_SET(reason, TCP_FLAGS);
6654 /* It is possible that we process SYN packets from backlog,
6655 * so we need to make sure to disable BH and RCU right there.
6659 acceptable = icsk->icsk_af_ops->conn_request(sk, skb) >= 0;
6668 SKB_DR_SET(reason, TCP_FLAGS);
6672 tp->rx_opt.saw_tstamp = 0;
6673 tcp_mstamp_refresh(tp);
6674 queued = tcp_rcv_synsent_state_process(sk, skb, th);
6678 /* Do step6 onward by hand. */
6679 tcp_urg(sk, skb, th);
6681 tcp_data_snd_check(sk);
6685 tcp_mstamp_refresh(tp);
6686 tp->rx_opt.saw_tstamp = 0;
6687 req = rcu_dereference_protected(tp->fastopen_rsk,
6688 lockdep_sock_is_held(sk));
6692 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
6693 sk->sk_state != TCP_FIN_WAIT1);
6695 if (!tcp_check_req(sk, skb, req, true, &req_stolen)) {
6696 SKB_DR_SET(reason, TCP_FASTOPEN);
6701 if (!th->ack && !th->rst && !th->syn) {
6702 SKB_DR_SET(reason, TCP_FLAGS);
6705 if (!tcp_validate_incoming(sk, skb, th, 0))
6708 /* step 5: check the ACK field */
6709 acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
6710 FLAG_UPDATE_TS_RECENT |
6711 FLAG_NO_CHALLENGE_ACK) > 0;
6714 if (sk->sk_state == TCP_SYN_RECV)
6715 return 1; /* send one RST */
6716 tcp_send_challenge_ack(sk);
6717 SKB_DR_SET(reason, TCP_OLD_ACK);
6720 switch (sk->sk_state) {
6722 tp->delivered++; /* SYN-ACK delivery isn't tracked in tcp_ack */
6724 tcp_synack_rtt_meas(sk, req);
6727 tcp_rcv_synrecv_state_fastopen(sk);
6729 tcp_try_undo_spurious_syn(sk);
6730 tp->retrans_stamp = 0;
6731 tcp_init_transfer(sk, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB,
6733 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6735 tcp_ao_established(sk);
6737 tcp_set_state(sk, TCP_ESTABLISHED);
6738 sk->sk_state_change(sk);
6740 /* Note, that this wakeup is only for marginal crossed SYN case.
6741 * Passively open sockets are not waked up, because
6742 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
6745 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
6747 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
6748 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
6749 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
6751 if (tp->rx_opt.tstamp_ok)
6752 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
6754 if (!inet_csk(sk)->icsk_ca_ops->cong_control)
6755 tcp_update_pacing_rate(sk);
6757 /* Prevent spurious tcp_cwnd_restart() on first data packet */
6758 tp->lsndtime = tcp_jiffies32;
6760 tcp_initialize_rcv_mss(sk);
6761 tcp_fast_path_on(tp);
6764 case TCP_FIN_WAIT1: {
6768 tcp_rcv_synrecv_state_fastopen(sk);
6770 if (tp->snd_una != tp->write_seq)
6773 tcp_set_state(sk, TCP_FIN_WAIT2);
6774 WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | SEND_SHUTDOWN);
6778 if (!sock_flag(sk, SOCK_DEAD)) {
6779 /* Wake up lingering close() */
6780 sk->sk_state_change(sk);
6784 if (READ_ONCE(tp->linger2) < 0) {
6786 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6789 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6790 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6791 /* Receive out of order FIN after close() */
6792 if (tp->syn_fastopen && th->fin)
6793 tcp_fastopen_active_disable(sk);
6795 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6799 tmo = tcp_fin_time(sk);
6800 if (tmo > TCP_TIMEWAIT_LEN) {
6801 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
6802 } else if (th->fin || sock_owned_by_user(sk)) {
6803 /* Bad case. We could lose such FIN otherwise.
6804 * It is not a big problem, but it looks confusing
6805 * and not so rare event. We still can lose it now,
6806 * if it spins in bh_lock_sock(), but it is really
6809 inet_csk_reset_keepalive_timer(sk, tmo);
6811 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
6818 if (tp->snd_una == tp->write_seq) {
6819 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
6825 if (tp->snd_una == tp->write_seq) {
6826 tcp_update_metrics(sk);
6833 /* step 6: check the URG bit */
6834 tcp_urg(sk, skb, th);
6836 /* step 7: process the segment text */
6837 switch (sk->sk_state) {
6838 case TCP_CLOSE_WAIT:
6841 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
6842 /* If a subflow has been reset, the packet should not
6843 * continue to be processed, drop the packet.
6845 if (sk_is_mptcp(sk) && !mptcp_incoming_options(sk, skb))
6852 /* RFC 793 says to queue data in these states,
6853 * RFC 1122 says we MUST send a reset.
6854 * BSD 4.4 also does reset.
6856 if (sk->sk_shutdown & RCV_SHUTDOWN) {
6857 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6858 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6859 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6865 case TCP_ESTABLISHED:
6866 tcp_data_queue(sk, skb);
6871 /* tcp_data could move socket to TIME-WAIT */
6872 if (sk->sk_state != TCP_CLOSE) {
6873 tcp_data_snd_check(sk);
6874 tcp_ack_snd_check(sk);
6879 tcp_drop_reason(sk, skb, reason);
6887 EXPORT_SYMBOL(tcp_rcv_state_process);
6889 static inline void pr_drop_req(struct request_sock *req, __u16 port, int family)
6891 struct inet_request_sock *ireq = inet_rsk(req);
6893 if (family == AF_INET)
6894 net_dbg_ratelimited("drop open request from %pI4/%u\n",
6895 &ireq->ir_rmt_addr, port);
6896 #if IS_ENABLED(CONFIG_IPV6)
6897 else if (family == AF_INET6)
6898 net_dbg_ratelimited("drop open request from %pI6/%u\n",
6899 &ireq->ir_v6_rmt_addr, port);
6903 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
6905 * If we receive a SYN packet with these bits set, it means a
6906 * network is playing bad games with TOS bits. In order to
6907 * avoid possible false congestion notifications, we disable
6908 * TCP ECN negotiation.
6910 * Exception: tcp_ca wants ECN. This is required for DCTCP
6911 * congestion control: Linux DCTCP asserts ECT on all packets,
6912 * including SYN, which is most optimal solution; however,
6913 * others, such as FreeBSD do not.
6915 * Exception: At least one of the reserved bits of the TCP header (th->res1) is
6916 * set, indicating the use of a future TCP extension (such as AccECN). See
6917 * RFC8311 §4.3 which updates RFC3168 to allow the development of such
6920 static void tcp_ecn_create_request(struct request_sock *req,
6921 const struct sk_buff *skb,
6922 const struct sock *listen_sk,
6923 const struct dst_entry *dst)
6925 const struct tcphdr *th = tcp_hdr(skb);
6926 const struct net *net = sock_net(listen_sk);
6927 bool th_ecn = th->ece && th->cwr;
6934 ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield);
6935 ecn_ok_dst = dst_feature(dst, DST_FEATURE_ECN_MASK);
6936 ecn_ok = READ_ONCE(net->ipv4.sysctl_tcp_ecn) || ecn_ok_dst;
6938 if (((!ect || th->res1) && ecn_ok) || tcp_ca_needs_ecn(listen_sk) ||
6939 (ecn_ok_dst & DST_FEATURE_ECN_CA) ||
6940 tcp_bpf_ca_needs_ecn((struct sock *)req))
6941 inet_rsk(req)->ecn_ok = 1;
6944 static void tcp_openreq_init(struct request_sock *req,
6945 const struct tcp_options_received *rx_opt,
6946 struct sk_buff *skb, const struct sock *sk)
6948 struct inet_request_sock *ireq = inet_rsk(req);
6950 req->rsk_rcv_wnd = 0; /* So that tcp_send_synack() knows! */
6951 tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq;
6952 tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
6953 tcp_rsk(req)->snt_synack = 0;
6954 tcp_rsk(req)->last_oow_ack_time = 0;
6955 req->mss = rx_opt->mss_clamp;
6956 req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0;
6957 ireq->tstamp_ok = rx_opt->tstamp_ok;
6958 ireq->sack_ok = rx_opt->sack_ok;
6959 ireq->snd_wscale = rx_opt->snd_wscale;
6960 ireq->wscale_ok = rx_opt->wscale_ok;
6963 ireq->ir_rmt_port = tcp_hdr(skb)->source;
6964 ireq->ir_num = ntohs(tcp_hdr(skb)->dest);
6965 ireq->ir_mark = inet_request_mark(sk, skb);
6966 #if IS_ENABLED(CONFIG_SMC)
6967 ireq->smc_ok = rx_opt->smc_ok && !(tcp_sk(sk)->smc_hs_congested &&
6968 tcp_sk(sk)->smc_hs_congested(sk));
6972 struct request_sock *inet_reqsk_alloc(const struct request_sock_ops *ops,
6973 struct sock *sk_listener,
6974 bool attach_listener)
6976 struct request_sock *req = reqsk_alloc(ops, sk_listener,
6980 struct inet_request_sock *ireq = inet_rsk(req);
6982 ireq->ireq_opt = NULL;
6983 #if IS_ENABLED(CONFIG_IPV6)
6984 ireq->pktopts = NULL;
6986 atomic64_set(&ireq->ir_cookie, 0);
6987 ireq->ireq_state = TCP_NEW_SYN_RECV;
6988 write_pnet(&ireq->ireq_net, sock_net(sk_listener));
6989 ireq->ireq_family = sk_listener->sk_family;
6990 req->timeout = TCP_TIMEOUT_INIT;
6995 EXPORT_SYMBOL(inet_reqsk_alloc);
6998 * Return true if a syncookie should be sent
7000 static bool tcp_syn_flood_action(const struct sock *sk, const char *proto)
7002 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
7003 const char *msg = "Dropping request";
7004 struct net *net = sock_net(sk);
7005 bool want_cookie = false;
7008 syncookies = READ_ONCE(net->ipv4.sysctl_tcp_syncookies);
7010 #ifdef CONFIG_SYN_COOKIES
7012 msg = "Sending cookies";
7014 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES);
7017 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP);
7019 if (!READ_ONCE(queue->synflood_warned) && syncookies != 2 &&
7020 xchg(&queue->synflood_warned, 1) == 0) {
7021 if (IS_ENABLED(CONFIG_IPV6) && sk->sk_family == AF_INET6) {
7022 net_info_ratelimited("%s: Possible SYN flooding on port [%pI6c]:%u. %s.\n",
7023 proto, inet6_rcv_saddr(sk),
7026 net_info_ratelimited("%s: Possible SYN flooding on port %pI4:%u. %s.\n",
7027 proto, &sk->sk_rcv_saddr,
7035 static void tcp_reqsk_record_syn(const struct sock *sk,
7036 struct request_sock *req,
7037 const struct sk_buff *skb)
7039 if (tcp_sk(sk)->save_syn) {
7040 u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb);
7041 struct saved_syn *saved_syn;
7045 if (tcp_sk(sk)->save_syn == 2) { /* Save full header. */
7046 base = skb_mac_header(skb);
7047 mac_hdrlen = skb_mac_header_len(skb);
7050 base = skb_network_header(skb);
7054 saved_syn = kmalloc(struct_size(saved_syn, data, len),
7057 saved_syn->mac_hdrlen = mac_hdrlen;
7058 saved_syn->network_hdrlen = skb_network_header_len(skb);
7059 saved_syn->tcp_hdrlen = tcp_hdrlen(skb);
7060 memcpy(saved_syn->data, base, len);
7061 req->saved_syn = saved_syn;
7066 /* If a SYN cookie is required and supported, returns a clamped MSS value to be
7067 * used for SYN cookie generation.
7069 u16 tcp_get_syncookie_mss(struct request_sock_ops *rsk_ops,
7070 const struct tcp_request_sock_ops *af_ops,
7071 struct sock *sk, struct tcphdr *th)
7073 struct tcp_sock *tp = tcp_sk(sk);
7076 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_syncookies) != 2 &&
7077 !inet_csk_reqsk_queue_is_full(sk))
7080 if (!tcp_syn_flood_action(sk, rsk_ops->slab_name))
7083 if (sk_acceptq_is_full(sk)) {
7084 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
7088 mss = tcp_parse_mss_option(th, tp->rx_opt.user_mss);
7090 mss = af_ops->mss_clamp;
7094 EXPORT_SYMBOL_GPL(tcp_get_syncookie_mss);
7096 int tcp_conn_request(struct request_sock_ops *rsk_ops,
7097 const struct tcp_request_sock_ops *af_ops,
7098 struct sock *sk, struct sk_buff *skb)
7100 struct tcp_fastopen_cookie foc = { .len = -1 };
7101 __u32 isn = TCP_SKB_CB(skb)->tcp_tw_isn;
7102 struct tcp_options_received tmp_opt;
7103 struct tcp_sock *tp = tcp_sk(sk);
7104 struct net *net = sock_net(sk);
7105 struct sock *fastopen_sk = NULL;
7106 struct request_sock *req;
7107 bool want_cookie = false;
7108 struct dst_entry *dst;
7112 #ifdef CONFIG_TCP_AO
7113 const struct tcp_ao_hdr *aoh;
7116 syncookies = READ_ONCE(net->ipv4.sysctl_tcp_syncookies);
7118 /* TW buckets are converted to open requests without
7119 * limitations, they conserve resources and peer is
7120 * evidently real one.
7122 if ((syncookies == 2 || inet_csk_reqsk_queue_is_full(sk)) && !isn) {
7123 want_cookie = tcp_syn_flood_action(sk, rsk_ops->slab_name);
7128 if (sk_acceptq_is_full(sk)) {
7129 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
7133 req = inet_reqsk_alloc(rsk_ops, sk, !want_cookie);
7137 req->syncookie = want_cookie;
7138 tcp_rsk(req)->af_specific = af_ops;
7139 tcp_rsk(req)->ts_off = 0;
7140 tcp_rsk(req)->req_usec_ts = false;
7141 #if IS_ENABLED(CONFIG_MPTCP)
7142 tcp_rsk(req)->is_mptcp = 0;
7145 tcp_clear_options(&tmp_opt);
7146 tmp_opt.mss_clamp = af_ops->mss_clamp;
7147 tmp_opt.user_mss = tp->rx_opt.user_mss;
7148 tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0,
7149 want_cookie ? NULL : &foc);
7151 if (want_cookie && !tmp_opt.saw_tstamp)
7152 tcp_clear_options(&tmp_opt);
7154 if (IS_ENABLED(CONFIG_SMC) && want_cookie)
7157 tmp_opt.tstamp_ok = tmp_opt.saw_tstamp;
7158 tcp_openreq_init(req, &tmp_opt, skb, sk);
7159 inet_rsk(req)->no_srccheck = inet_test_bit(TRANSPARENT, sk);
7161 /* Note: tcp_v6_init_req() might override ir_iif for link locals */
7162 inet_rsk(req)->ir_iif = inet_request_bound_dev_if(sk, skb);
7164 dst = af_ops->route_req(sk, skb, &fl, req);
7168 if (tmp_opt.tstamp_ok) {
7169 tcp_rsk(req)->req_usec_ts = dst_tcp_usec_ts(dst);
7170 tcp_rsk(req)->ts_off = af_ops->init_ts_off(net, skb);
7172 if (!want_cookie && !isn) {
7173 int max_syn_backlog = READ_ONCE(net->ipv4.sysctl_max_syn_backlog);
7175 /* Kill the following clause, if you dislike this way. */
7177 (max_syn_backlog - inet_csk_reqsk_queue_len(sk) <
7178 (max_syn_backlog >> 2)) &&
7179 !tcp_peer_is_proven(req, dst)) {
7180 /* Without syncookies last quarter of
7181 * backlog is filled with destinations,
7182 * proven to be alive.
7183 * It means that we continue to communicate
7184 * to destinations, already remembered
7185 * to the moment of synflood.
7187 pr_drop_req(req, ntohs(tcp_hdr(skb)->source),
7189 goto drop_and_release;
7192 isn = af_ops->init_seq(skb);
7195 tcp_ecn_create_request(req, skb, sk, dst);
7198 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss);
7199 if (!tmp_opt.tstamp_ok)
7200 inet_rsk(req)->ecn_ok = 0;
7203 #ifdef CONFIG_TCP_AO
7204 if (tcp_parse_auth_options(tcp_hdr(skb), NULL, &aoh))
7205 goto drop_and_release; /* Invalid TCP options */
7207 tcp_rsk(req)->used_tcp_ao = true;
7208 tcp_rsk(req)->ao_rcv_next = aoh->keyid;
7209 tcp_rsk(req)->ao_keyid = aoh->rnext_keyid;
7212 tcp_rsk(req)->used_tcp_ao = false;
7215 tcp_rsk(req)->snt_isn = isn;
7216 tcp_rsk(req)->txhash = net_tx_rndhash();
7217 tcp_rsk(req)->syn_tos = TCP_SKB_CB(skb)->ip_dsfield;
7218 tcp_openreq_init_rwin(req, sk, dst);
7219 sk_rx_queue_set(req_to_sk(req), skb);
7221 tcp_reqsk_record_syn(sk, req, skb);
7222 fastopen_sk = tcp_try_fastopen(sk, skb, req, &foc, dst);
7225 af_ops->send_synack(fastopen_sk, dst, &fl, req,
7226 &foc, TCP_SYNACK_FASTOPEN, skb);
7227 /* Add the child socket directly into the accept queue */
7228 if (!inet_csk_reqsk_queue_add(sk, req, fastopen_sk)) {
7229 reqsk_fastopen_remove(fastopen_sk, req, false);
7230 bh_unlock_sock(fastopen_sk);
7231 sock_put(fastopen_sk);
7234 sk->sk_data_ready(sk);
7235 bh_unlock_sock(fastopen_sk);
7236 sock_put(fastopen_sk);
7238 tcp_rsk(req)->tfo_listener = false;
7240 req->timeout = tcp_timeout_init((struct sock *)req);
7241 inet_csk_reqsk_queue_hash_add(sk, req, req->timeout);
7243 af_ops->send_synack(sk, dst, &fl, req, &foc,
7244 !want_cookie ? TCP_SYNACK_NORMAL :
7263 EXPORT_SYMBOL(tcp_conn_request);