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 icsk->icsk_ack.rcv_mss = min_t(unsigned int, len,
242 /* Account for possibly-removed options */
243 if (unlikely(len > icsk->icsk_ack.rcv_mss +
244 MAX_TCP_OPTION_SPACE))
245 tcp_gro_dev_warn(sk, skb, len);
246 /* If the skb has a len of exactly 1*MSS and has the PSH bit
247 * set then it is likely the end of an application write. So
248 * more data may not be arriving soon, and yet the data sender
249 * may be waiting for an ACK if cwnd-bound or using TX zero
250 * copy. So we set ICSK_ACK_PUSHED here so that
251 * tcp_cleanup_rbuf() will send an ACK immediately if the app
252 * reads all of the data and is not ping-pong. If len > MSS
253 * then this logic does not matter (and does not hurt) because
254 * tcp_cleanup_rbuf() will always ACK immediately if the app
255 * reads data and there is more than an MSS of unACKed data.
257 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_PSH)
258 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
260 /* Otherwise, we make more careful check taking into account,
261 * that SACKs block is variable.
263 * "len" is invariant segment length, including TCP header.
265 len += skb->data - skb_transport_header(skb);
266 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
267 /* If PSH is not set, packet should be
268 * full sized, provided peer TCP is not badly broken.
269 * This observation (if it is correct 8)) allows
270 * to handle super-low mtu links fairly.
272 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
273 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
274 /* Subtract also invariant (if peer is RFC compliant),
275 * tcp header plus fixed timestamp option length.
276 * Resulting "len" is MSS free of SACK jitter.
278 len -= tcp_sk(sk)->tcp_header_len;
279 icsk->icsk_ack.last_seg_size = len;
281 icsk->icsk_ack.rcv_mss = len;
285 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
286 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
287 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
291 static void tcp_incr_quickack(struct sock *sk, unsigned int max_quickacks)
293 struct inet_connection_sock *icsk = inet_csk(sk);
294 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
298 quickacks = min(quickacks, max_quickacks);
299 if (quickacks > icsk->icsk_ack.quick)
300 icsk->icsk_ack.quick = quickacks;
303 static void tcp_enter_quickack_mode(struct sock *sk, unsigned int max_quickacks)
305 struct inet_connection_sock *icsk = inet_csk(sk);
307 tcp_incr_quickack(sk, max_quickacks);
308 inet_csk_exit_pingpong_mode(sk);
309 icsk->icsk_ack.ato = TCP_ATO_MIN;
312 /* Send ACKs quickly, if "quick" count is not exhausted
313 * and the session is not interactive.
316 static bool tcp_in_quickack_mode(struct sock *sk)
318 const struct inet_connection_sock *icsk = inet_csk(sk);
319 const struct dst_entry *dst = __sk_dst_get(sk);
321 return (dst && dst_metric(dst, RTAX_QUICKACK)) ||
322 (icsk->icsk_ack.quick && !inet_csk_in_pingpong_mode(sk));
325 static void tcp_ecn_queue_cwr(struct tcp_sock *tp)
327 if (tp->ecn_flags & TCP_ECN_OK)
328 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
331 static void tcp_ecn_accept_cwr(struct sock *sk, const struct sk_buff *skb)
333 if (tcp_hdr(skb)->cwr) {
334 tcp_sk(sk)->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
336 /* If the sender is telling us it has entered CWR, then its
337 * cwnd may be very low (even just 1 packet), so we should ACK
340 if (TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq)
341 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW;
345 static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp)
347 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
350 static void __tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb)
352 struct tcp_sock *tp = tcp_sk(sk);
354 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
355 case INET_ECN_NOT_ECT:
356 /* Funny extension: if ECT is not set on a segment,
357 * and we already seen ECT on a previous segment,
358 * it is probably a retransmit.
360 if (tp->ecn_flags & TCP_ECN_SEEN)
361 tcp_enter_quickack_mode(sk, 2);
364 if (tcp_ca_needs_ecn(sk))
365 tcp_ca_event(sk, CA_EVENT_ECN_IS_CE);
367 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
368 /* Better not delay acks, sender can have a very low cwnd */
369 tcp_enter_quickack_mode(sk, 2);
370 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
372 tp->ecn_flags |= TCP_ECN_SEEN;
375 if (tcp_ca_needs_ecn(sk))
376 tcp_ca_event(sk, CA_EVENT_ECN_NO_CE);
377 tp->ecn_flags |= TCP_ECN_SEEN;
382 static void tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb)
384 if (tcp_sk(sk)->ecn_flags & TCP_ECN_OK)
385 __tcp_ecn_check_ce(sk, skb);
388 static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
390 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
391 tp->ecn_flags &= ~TCP_ECN_OK;
394 static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
396 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
397 tp->ecn_flags &= ~TCP_ECN_OK;
400 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
402 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
407 /* Buffer size and advertised window tuning.
409 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
412 static void tcp_sndbuf_expand(struct sock *sk)
414 const struct tcp_sock *tp = tcp_sk(sk);
415 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
419 /* Worst case is non GSO/TSO : each frame consumes one skb
420 * and skb->head is kmalloced using power of two area of memory
422 per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
424 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
426 per_mss = roundup_pow_of_two(per_mss) +
427 SKB_DATA_ALIGN(sizeof(struct sk_buff));
429 nr_segs = max_t(u32, TCP_INIT_CWND, tcp_snd_cwnd(tp));
430 nr_segs = max_t(u32, nr_segs, tp->reordering + 1);
432 /* Fast Recovery (RFC 5681 3.2) :
433 * Cubic needs 1.7 factor, rounded to 2 to include
434 * extra cushion (application might react slowly to EPOLLOUT)
436 sndmem = ca_ops->sndbuf_expand ? ca_ops->sndbuf_expand(sk) : 2;
437 sndmem *= nr_segs * per_mss;
439 if (sk->sk_sndbuf < sndmem)
440 WRITE_ONCE(sk->sk_sndbuf,
441 min(sndmem, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_wmem[2])));
444 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
446 * All tcp_full_space() is split to two parts: "network" buffer, allocated
447 * forward and advertised in receiver window (tp->rcv_wnd) and
448 * "application buffer", required to isolate scheduling/application
449 * latencies from network.
450 * window_clamp is maximal advertised window. It can be less than
451 * tcp_full_space(), in this case tcp_full_space() - window_clamp
452 * is reserved for "application" buffer. The less window_clamp is
453 * the smoother our behaviour from viewpoint of network, but the lower
454 * throughput and the higher sensitivity of the connection to losses. 8)
456 * rcv_ssthresh is more strict window_clamp used at "slow start"
457 * phase to predict further behaviour of this connection.
458 * It is used for two goals:
459 * - to enforce header prediction at sender, even when application
460 * requires some significant "application buffer". It is check #1.
461 * - to prevent pruning of receive queue because of misprediction
462 * of receiver window. Check #2.
464 * The scheme does not work when sender sends good segments opening
465 * window and then starts to feed us spaghetti. But it should work
466 * in common situations. Otherwise, we have to rely on queue collapsing.
469 /* Slow part of check#2. */
470 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb,
471 unsigned int skbtruesize)
473 struct tcp_sock *tp = tcp_sk(sk);
475 int truesize = tcp_win_from_space(sk, skbtruesize) >> 1;
476 int window = tcp_win_from_space(sk, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2])) >> 1;
478 while (tp->rcv_ssthresh <= window) {
479 if (truesize <= skb->len)
480 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
488 /* Even if skb appears to have a bad len/truesize ratio, TCP coalescing
489 * can play nice with us, as sk_buff and skb->head might be either
490 * freed or shared with up to MAX_SKB_FRAGS segments.
491 * Only give a boost to drivers using page frag(s) to hold the frame(s),
492 * and if no payload was pulled in skb->head before reaching us.
494 static u32 truesize_adjust(bool adjust, const struct sk_buff *skb)
496 u32 truesize = skb->truesize;
498 if (adjust && !skb_headlen(skb)) {
499 truesize -= SKB_TRUESIZE(skb_end_offset(skb));
500 /* paranoid check, some drivers might be buggy */
501 if (unlikely((int)truesize < (int)skb->len))
502 truesize = skb->truesize;
507 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb,
510 struct tcp_sock *tp = tcp_sk(sk);
513 room = min_t(int, tp->window_clamp, tcp_space(sk)) - tp->rcv_ssthresh;
519 if (!tcp_under_memory_pressure(sk)) {
520 unsigned int truesize = truesize_adjust(adjust, skb);
523 /* Check #2. Increase window, if skb with such overhead
524 * will fit to rcvbuf in future.
526 if (tcp_win_from_space(sk, truesize) <= skb->len)
527 incr = 2 * tp->advmss;
529 incr = __tcp_grow_window(sk, skb, truesize);
532 incr = max_t(int, incr, 2 * skb->len);
533 tp->rcv_ssthresh += min(room, incr);
534 inet_csk(sk)->icsk_ack.quick |= 1;
538 * Adjust rcv_ssthresh according to reserved mem
540 tcp_adjust_rcv_ssthresh(sk);
544 /* 3. Try to fixup all. It is made immediately after connection enters
547 static void tcp_init_buffer_space(struct sock *sk)
549 int tcp_app_win = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_app_win);
550 struct tcp_sock *tp = tcp_sk(sk);
553 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
554 tcp_sndbuf_expand(sk);
556 tcp_mstamp_refresh(tp);
557 tp->rcvq_space.time = tp->tcp_mstamp;
558 tp->rcvq_space.seq = tp->copied_seq;
560 maxwin = tcp_full_space(sk);
562 if (tp->window_clamp >= maxwin) {
563 tp->window_clamp = maxwin;
565 if (tcp_app_win && maxwin > 4 * tp->advmss)
566 tp->window_clamp = max(maxwin -
567 (maxwin >> tcp_app_win),
571 /* Force reservation of one segment. */
573 tp->window_clamp > 2 * tp->advmss &&
574 tp->window_clamp + tp->advmss > maxwin)
575 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
577 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
578 tp->snd_cwnd_stamp = tcp_jiffies32;
579 tp->rcvq_space.space = min3(tp->rcv_ssthresh, tp->rcv_wnd,
580 (u32)TCP_INIT_CWND * tp->advmss);
583 /* 4. Recalculate window clamp after socket hit its memory bounds. */
584 static void tcp_clamp_window(struct sock *sk)
586 struct tcp_sock *tp = tcp_sk(sk);
587 struct inet_connection_sock *icsk = inet_csk(sk);
588 struct net *net = sock_net(sk);
591 icsk->icsk_ack.quick = 0;
592 rmem2 = READ_ONCE(net->ipv4.sysctl_tcp_rmem[2]);
594 if (sk->sk_rcvbuf < rmem2 &&
595 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
596 !tcp_under_memory_pressure(sk) &&
597 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
598 WRITE_ONCE(sk->sk_rcvbuf,
599 min(atomic_read(&sk->sk_rmem_alloc), rmem2));
601 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
602 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
605 /* Initialize RCV_MSS value.
606 * RCV_MSS is an our guess about MSS used by the peer.
607 * We haven't any direct information about the MSS.
608 * It's better to underestimate the RCV_MSS rather than overestimate.
609 * Overestimations make us ACKing less frequently than needed.
610 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
612 void tcp_initialize_rcv_mss(struct sock *sk)
614 const struct tcp_sock *tp = tcp_sk(sk);
615 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
617 hint = min(hint, tp->rcv_wnd / 2);
618 hint = min(hint, TCP_MSS_DEFAULT);
619 hint = max(hint, TCP_MIN_MSS);
621 inet_csk(sk)->icsk_ack.rcv_mss = hint;
623 EXPORT_SYMBOL(tcp_initialize_rcv_mss);
625 /* Receiver "autotuning" code.
627 * The algorithm for RTT estimation w/o timestamps is based on
628 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
629 * <https://public.lanl.gov/radiant/pubs.html#DRS>
631 * More detail on this code can be found at
632 * <http://staff.psc.edu/jheffner/>,
633 * though this reference is out of date. A new paper
636 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
638 u32 new_sample = tp->rcv_rtt_est.rtt_us;
641 if (new_sample != 0) {
642 /* If we sample in larger samples in the non-timestamp
643 * case, we could grossly overestimate the RTT especially
644 * with chatty applications or bulk transfer apps which
645 * are stalled on filesystem I/O.
647 * Also, since we are only going for a minimum in the
648 * non-timestamp case, we do not smooth things out
649 * else with timestamps disabled convergence takes too
653 m -= (new_sample >> 3);
661 /* No previous measure. */
665 tp->rcv_rtt_est.rtt_us = new_sample;
668 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
672 if (tp->rcv_rtt_est.time == 0)
674 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
676 delta_us = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcv_rtt_est.time);
679 tcp_rcv_rtt_update(tp, delta_us, 1);
682 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
683 tp->rcv_rtt_est.time = tp->tcp_mstamp;
686 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
687 const struct sk_buff *skb)
689 struct tcp_sock *tp = tcp_sk(sk);
691 if (tp->rx_opt.rcv_tsecr == tp->rcv_rtt_last_tsecr)
693 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr;
695 if (TCP_SKB_CB(skb)->end_seq -
696 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss) {
697 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr;
700 if (likely(delta < INT_MAX / (USEC_PER_SEC / TCP_TS_HZ))) {
703 delta_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
704 tcp_rcv_rtt_update(tp, delta_us, 0);
710 * This function should be called every time data is copied to user space.
711 * It calculates the appropriate TCP receive buffer space.
713 void tcp_rcv_space_adjust(struct sock *sk)
715 struct tcp_sock *tp = tcp_sk(sk);
719 trace_tcp_rcv_space_adjust(sk);
721 tcp_mstamp_refresh(tp);
722 time = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcvq_space.time);
723 if (time < (tp->rcv_rtt_est.rtt_us >> 3) || tp->rcv_rtt_est.rtt_us == 0)
726 /* Number of bytes copied to user in last RTT */
727 copied = tp->copied_seq - tp->rcvq_space.seq;
728 if (copied <= tp->rcvq_space.space)
732 * copied = bytes received in previous RTT, our base window
733 * To cope with packet losses, we need a 2x factor
734 * To cope with slow start, and sender growing its cwin by 100 %
735 * every RTT, we need a 4x factor, because the ACK we are sending
736 * now is for the next RTT, not the current one :
737 * <prev RTT . ><current RTT .. ><next RTT .... >
740 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf) &&
741 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
745 /* minimal window to cope with packet losses, assuming
746 * steady state. Add some cushion because of small variations.
748 rcvwin = ((u64)copied << 1) + 16 * tp->advmss;
750 /* Accommodate for sender rate increase (eg. slow start) */
751 grow = rcvwin * (copied - tp->rcvq_space.space);
752 do_div(grow, tp->rcvq_space.space);
753 rcvwin += (grow << 1);
755 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
756 while (tcp_win_from_space(sk, rcvmem) < tp->advmss)
759 do_div(rcvwin, tp->advmss);
760 rcvbuf = min_t(u64, rcvwin * rcvmem,
761 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]));
762 if (rcvbuf > sk->sk_rcvbuf) {
763 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
765 /* Make the window clamp follow along. */
766 tp->window_clamp = tcp_win_from_space(sk, rcvbuf);
769 tp->rcvq_space.space = copied;
772 tp->rcvq_space.seq = tp->copied_seq;
773 tp->rcvq_space.time = tp->tcp_mstamp;
776 /* There is something which you must keep in mind when you analyze the
777 * behavior of the tp->ato delayed ack timeout interval. When a
778 * connection starts up, we want to ack as quickly as possible. The
779 * problem is that "good" TCP's do slow start at the beginning of data
780 * transmission. The means that until we send the first few ACK's the
781 * sender will sit on his end and only queue most of his data, because
782 * he can only send snd_cwnd unacked packets at any given time. For
783 * each ACK we send, he increments snd_cwnd and transmits more of his
786 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
788 struct tcp_sock *tp = tcp_sk(sk);
789 struct inet_connection_sock *icsk = inet_csk(sk);
792 inet_csk_schedule_ack(sk);
794 tcp_measure_rcv_mss(sk, skb);
796 tcp_rcv_rtt_measure(tp);
800 if (!icsk->icsk_ack.ato) {
801 /* The _first_ data packet received, initialize
802 * delayed ACK engine.
804 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
805 icsk->icsk_ack.ato = TCP_ATO_MIN;
807 int m = now - icsk->icsk_ack.lrcvtime;
809 if (m <= TCP_ATO_MIN / 2) {
810 /* The fastest case is the first. */
811 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
812 } else if (m < icsk->icsk_ack.ato) {
813 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
814 if (icsk->icsk_ack.ato > icsk->icsk_rto)
815 icsk->icsk_ack.ato = icsk->icsk_rto;
816 } else if (m > icsk->icsk_rto) {
817 /* Too long gap. Apparently sender failed to
818 * restart window, so that we send ACKs quickly.
820 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
823 icsk->icsk_ack.lrcvtime = now;
825 tcp_ecn_check_ce(sk, skb);
828 tcp_grow_window(sk, skb, true);
831 /* Called to compute a smoothed rtt estimate. The data fed to this
832 * routine either comes from timestamps, or from segments that were
833 * known _not_ to have been retransmitted [see Karn/Partridge
834 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
835 * piece by Van Jacobson.
836 * NOTE: the next three routines used to be one big routine.
837 * To save cycles in the RFC 1323 implementation it was better to break
838 * it up into three procedures. -- erics
840 static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
842 struct tcp_sock *tp = tcp_sk(sk);
843 long m = mrtt_us; /* RTT */
844 u32 srtt = tp->srtt_us;
846 /* The following amusing code comes from Jacobson's
847 * article in SIGCOMM '88. Note that rtt and mdev
848 * are scaled versions of rtt and mean deviation.
849 * This is designed to be as fast as possible
850 * m stands for "measurement".
852 * On a 1990 paper the rto value is changed to:
853 * RTO = rtt + 4 * mdev
855 * Funny. This algorithm seems to be very broken.
856 * These formulae increase RTO, when it should be decreased, increase
857 * too slowly, when it should be increased quickly, decrease too quickly
858 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
859 * does not matter how to _calculate_ it. Seems, it was trap
860 * that VJ failed to avoid. 8)
863 m -= (srtt >> 3); /* m is now error in rtt est */
864 srtt += m; /* rtt = 7/8 rtt + 1/8 new */
866 m = -m; /* m is now abs(error) */
867 m -= (tp->mdev_us >> 2); /* similar update on mdev */
868 /* This is similar to one of Eifel findings.
869 * Eifel blocks mdev updates when rtt decreases.
870 * This solution is a bit different: we use finer gain
871 * for mdev in this case (alpha*beta).
872 * Like Eifel it also prevents growth of rto,
873 * but also it limits too fast rto decreases,
874 * happening in pure Eifel.
879 m -= (tp->mdev_us >> 2); /* similar update on mdev */
881 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */
882 if (tp->mdev_us > tp->mdev_max_us) {
883 tp->mdev_max_us = tp->mdev_us;
884 if (tp->mdev_max_us > tp->rttvar_us)
885 tp->rttvar_us = tp->mdev_max_us;
887 if (after(tp->snd_una, tp->rtt_seq)) {
888 if (tp->mdev_max_us < tp->rttvar_us)
889 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
890 tp->rtt_seq = tp->snd_nxt;
891 tp->mdev_max_us = tcp_rto_min_us(sk);
896 /* no previous measure. */
897 srtt = m << 3; /* take the measured time to be rtt */
898 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */
899 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
900 tp->mdev_max_us = tp->rttvar_us;
901 tp->rtt_seq = tp->snd_nxt;
905 tp->srtt_us = max(1U, srtt);
908 static void tcp_update_pacing_rate(struct sock *sk)
910 const struct tcp_sock *tp = tcp_sk(sk);
913 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
914 rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3);
916 /* current rate is (cwnd * mss) / srtt
917 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
918 * In Congestion Avoidance phase, set it to 120 % the current rate.
920 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
921 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
922 * end of slow start and should slow down.
924 if (tcp_snd_cwnd(tp) < tp->snd_ssthresh / 2)
925 rate *= READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_pacing_ss_ratio);
927 rate *= READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_pacing_ca_ratio);
929 rate *= max(tcp_snd_cwnd(tp), tp->packets_out);
931 if (likely(tp->srtt_us))
932 do_div(rate, tp->srtt_us);
934 /* WRITE_ONCE() is needed because sch_fq fetches sk_pacing_rate
935 * without any lock. We want to make sure compiler wont store
936 * intermediate values in this location.
938 WRITE_ONCE(sk->sk_pacing_rate, min_t(u64, rate,
939 sk->sk_max_pacing_rate));
942 /* Calculate rto without backoff. This is the second half of Van Jacobson's
943 * routine referred to above.
945 static void tcp_set_rto(struct sock *sk)
947 const struct tcp_sock *tp = tcp_sk(sk);
948 /* Old crap is replaced with new one. 8)
951 * 1. If rtt variance happened to be less 50msec, it is hallucination.
952 * It cannot be less due to utterly erratic ACK generation made
953 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
954 * to do with delayed acks, because at cwnd>2 true delack timeout
955 * is invisible. Actually, Linux-2.4 also generates erratic
956 * ACKs in some circumstances.
958 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
960 /* 2. Fixups made earlier cannot be right.
961 * If we do not estimate RTO correctly without them,
962 * all the algo is pure shit and should be replaced
963 * with correct one. It is exactly, which we pretend to do.
966 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
967 * guarantees that rto is higher.
972 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
974 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
977 cwnd = TCP_INIT_CWND;
978 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
981 struct tcp_sacktag_state {
982 /* Timestamps for earliest and latest never-retransmitted segment
983 * that was SACKed. RTO needs the earliest RTT to stay conservative,
984 * but congestion control should still get an accurate delay signal.
991 unsigned int mss_now;
992 struct rate_sample *rate;
995 /* Take a notice that peer is sending D-SACKs. Skip update of data delivery
996 * and spurious retransmission information if this DSACK is unlikely caused by
998 * - DSACKed sequence range is larger than maximum receiver's window.
999 * - Total no. of DSACKed segments exceed the total no. of retransmitted segs.
1001 static u32 tcp_dsack_seen(struct tcp_sock *tp, u32 start_seq,
1002 u32 end_seq, struct tcp_sacktag_state *state)
1004 u32 seq_len, dup_segs = 1;
1006 if (!before(start_seq, end_seq))
1009 seq_len = end_seq - start_seq;
1010 /* Dubious DSACK: DSACKed range greater than maximum advertised rwnd */
1011 if (seq_len > tp->max_window)
1013 if (seq_len > tp->mss_cache)
1014 dup_segs = DIV_ROUND_UP(seq_len, tp->mss_cache);
1015 else if (tp->tlp_high_seq && tp->tlp_high_seq == end_seq)
1016 state->flag |= FLAG_DSACK_TLP;
1018 tp->dsack_dups += dup_segs;
1019 /* Skip the DSACK if dup segs weren't retransmitted by sender */
1020 if (tp->dsack_dups > tp->total_retrans)
1023 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
1024 /* We increase the RACK ordering window in rounds where we receive
1025 * DSACKs that may have been due to reordering causing RACK to trigger
1026 * a spurious fast recovery. Thus RACK ignores DSACKs that happen
1027 * without having seen reordering, or that match TLP probes (TLP
1028 * is timer-driven, not triggered by RACK).
1030 if (tp->reord_seen && !(state->flag & FLAG_DSACK_TLP))
1031 tp->rack.dsack_seen = 1;
1033 state->flag |= FLAG_DSACKING_ACK;
1034 /* A spurious retransmission is delivered */
1035 state->sack_delivered += dup_segs;
1040 /* It's reordering when higher sequence was delivered (i.e. sacked) before
1041 * some lower never-retransmitted sequence ("low_seq"). The maximum reordering
1042 * distance is approximated in full-mss packet distance ("reordering").
1044 static void tcp_check_sack_reordering(struct sock *sk, const u32 low_seq,
1047 struct tcp_sock *tp = tcp_sk(sk);
1048 const u32 mss = tp->mss_cache;
1051 fack = tcp_highest_sack_seq(tp);
1052 if (!before(low_seq, fack))
1055 metric = fack - low_seq;
1056 if ((metric > tp->reordering * mss) && mss) {
1057 #if FASTRETRANS_DEBUG > 1
1058 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
1059 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
1063 tp->undo_marker ? tp->undo_retrans : 0);
1065 tp->reordering = min_t(u32, (metric + mss - 1) / mss,
1066 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_max_reordering));
1069 /* This exciting event is worth to be remembered. 8) */
1071 NET_INC_STATS(sock_net(sk),
1072 ts ? LINUX_MIB_TCPTSREORDER : LINUX_MIB_TCPSACKREORDER);
1075 /* This must be called before lost_out or retrans_out are updated
1076 * on a new loss, because we want to know if all skbs previously
1077 * known to be lost have already been retransmitted, indicating
1078 * that this newly lost skb is our next skb to retransmit.
1080 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
1082 if ((!tp->retransmit_skb_hint && tp->retrans_out >= tp->lost_out) ||
1083 (tp->retransmit_skb_hint &&
1084 before(TCP_SKB_CB(skb)->seq,
1085 TCP_SKB_CB(tp->retransmit_skb_hint)->seq)))
1086 tp->retransmit_skb_hint = skb;
1089 /* Sum the number of packets on the wire we have marked as lost, and
1090 * notify the congestion control module that the given skb was marked lost.
1092 static void tcp_notify_skb_loss_event(struct tcp_sock *tp, const struct sk_buff *skb)
1094 tp->lost += tcp_skb_pcount(skb);
1097 void tcp_mark_skb_lost(struct sock *sk, struct sk_buff *skb)
1099 __u8 sacked = TCP_SKB_CB(skb)->sacked;
1100 struct tcp_sock *tp = tcp_sk(sk);
1102 if (sacked & TCPCB_SACKED_ACKED)
1105 tcp_verify_retransmit_hint(tp, skb);
1106 if (sacked & TCPCB_LOST) {
1107 if (sacked & TCPCB_SACKED_RETRANS) {
1108 /* Account for retransmits that are lost again */
1109 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1110 tp->retrans_out -= tcp_skb_pcount(skb);
1111 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT,
1112 tcp_skb_pcount(skb));
1113 tcp_notify_skb_loss_event(tp, skb);
1116 tp->lost_out += tcp_skb_pcount(skb);
1117 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1118 tcp_notify_skb_loss_event(tp, skb);
1122 /* Updates the delivered and delivered_ce counts */
1123 static void tcp_count_delivered(struct tcp_sock *tp, u32 delivered,
1126 tp->delivered += delivered;
1128 tp->delivered_ce += delivered;
1131 /* This procedure tags the retransmission queue when SACKs arrive.
1133 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
1134 * Packets in queue with these bits set are counted in variables
1135 * sacked_out, retrans_out and lost_out, correspondingly.
1137 * Valid combinations are:
1138 * Tag InFlight Description
1139 * 0 1 - orig segment is in flight.
1140 * S 0 - nothing flies, orig reached receiver.
1141 * L 0 - nothing flies, orig lost by net.
1142 * R 2 - both orig and retransmit are in flight.
1143 * L|R 1 - orig is lost, retransmit is in flight.
1144 * S|R 1 - orig reached receiver, retrans is still in flight.
1145 * (L|S|R is logically valid, it could occur when L|R is sacked,
1146 * but it is equivalent to plain S and code short-curcuits it to S.
1147 * L|S is logically invalid, it would mean -1 packet in flight 8))
1149 * These 6 states form finite state machine, controlled by the following events:
1150 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
1151 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
1152 * 3. Loss detection event of two flavors:
1153 * A. Scoreboard estimator decided the packet is lost.
1154 * A'. Reno "three dupacks" marks head of queue lost.
1155 * B. SACK arrives sacking SND.NXT at the moment, when the
1156 * segment was retransmitted.
1157 * 4. D-SACK added new rule: D-SACK changes any tag to S.
1159 * It is pleasant to note, that state diagram turns out to be commutative,
1160 * so that we are allowed not to be bothered by order of our actions,
1161 * when multiple events arrive simultaneously. (see the function below).
1163 * Reordering detection.
1164 * --------------------
1165 * Reordering metric is maximal distance, which a packet can be displaced
1166 * in packet stream. With SACKs we can estimate it:
1168 * 1. SACK fills old hole and the corresponding segment was not
1169 * ever retransmitted -> reordering. Alas, we cannot use it
1170 * when segment was retransmitted.
1171 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1172 * for retransmitted and already SACKed segment -> reordering..
1173 * Both of these heuristics are not used in Loss state, when we cannot
1174 * account for retransmits accurately.
1176 * SACK block validation.
1177 * ----------------------
1179 * SACK block range validation checks that the received SACK block fits to
1180 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1181 * Note that SND.UNA is not included to the range though being valid because
1182 * it means that the receiver is rather inconsistent with itself reporting
1183 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1184 * perfectly valid, however, in light of RFC2018 which explicitly states
1185 * that "SACK block MUST reflect the newest segment. Even if the newest
1186 * segment is going to be discarded ...", not that it looks very clever
1187 * in case of head skb. Due to potentional receiver driven attacks, we
1188 * choose to avoid immediate execution of a walk in write queue due to
1189 * reneging and defer head skb's loss recovery to standard loss recovery
1190 * procedure that will eventually trigger (nothing forbids us doing this).
1192 * Implements also blockage to start_seq wrap-around. Problem lies in the
1193 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1194 * there's no guarantee that it will be before snd_nxt (n). The problem
1195 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1198 * <- outs wnd -> <- wrapzone ->
1199 * u e n u_w e_w s n_w
1201 * |<------------+------+----- TCP seqno space --------------+---------->|
1202 * ...-- <2^31 ->| |<--------...
1203 * ...---- >2^31 ------>| |<--------...
1205 * Current code wouldn't be vulnerable but it's better still to discard such
1206 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1207 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1208 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1209 * equal to the ideal case (infinite seqno space without wrap caused issues).
1211 * With D-SACK the lower bound is extended to cover sequence space below
1212 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1213 * again, D-SACK block must not to go across snd_una (for the same reason as
1214 * for the normal SACK blocks, explained above). But there all simplicity
1215 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1216 * fully below undo_marker they do not affect behavior in anyway and can
1217 * therefore be safely ignored. In rare cases (which are more or less
1218 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1219 * fragmentation and packet reordering past skb's retransmission. To consider
1220 * them correctly, the acceptable range must be extended even more though
1221 * the exact amount is rather hard to quantify. However, tp->max_window can
1222 * be used as an exaggerated estimate.
1224 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
1225 u32 start_seq, u32 end_seq)
1227 /* Too far in future, or reversed (interpretation is ambiguous) */
1228 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1231 /* Nasty start_seq wrap-around check (see comments above) */
1232 if (!before(start_seq, tp->snd_nxt))
1235 /* In outstanding window? ...This is valid exit for D-SACKs too.
1236 * start_seq == snd_una is non-sensical (see comments above)
1238 if (after(start_seq, tp->snd_una))
1241 if (!is_dsack || !tp->undo_marker)
1244 /* ...Then it's D-SACK, and must reside below snd_una completely */
1245 if (after(end_seq, tp->snd_una))
1248 if (!before(start_seq, tp->undo_marker))
1252 if (!after(end_seq, tp->undo_marker))
1255 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1256 * start_seq < undo_marker and end_seq >= undo_marker.
1258 return !before(start_seq, end_seq - tp->max_window);
1261 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1262 struct tcp_sack_block_wire *sp, int num_sacks,
1263 u32 prior_snd_una, struct tcp_sacktag_state *state)
1265 struct tcp_sock *tp = tcp_sk(sk);
1266 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1267 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1270 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1271 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1272 } else if (num_sacks > 1) {
1273 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1274 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1276 if (after(end_seq_0, end_seq_1) || before(start_seq_0, start_seq_1))
1278 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKOFORECV);
1283 dup_segs = tcp_dsack_seen(tp, start_seq_0, end_seq_0, state);
1284 if (!dup_segs) { /* Skip dubious DSACK */
1285 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKIGNOREDDUBIOUS);
1289 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECVSEGS, dup_segs);
1291 /* D-SACK for already forgotten data... Do dumb counting. */
1292 if (tp->undo_marker && tp->undo_retrans > 0 &&
1293 !after(end_seq_0, prior_snd_una) &&
1294 after(end_seq_0, tp->undo_marker))
1295 tp->undo_retrans = max_t(int, 0, tp->undo_retrans - dup_segs);
1300 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1301 * the incoming SACK may not exactly match but we can find smaller MSS
1302 * aligned portion of it that matches. Therefore we might need to fragment
1303 * which may fail and creates some hassle (caller must handle error case
1306 * FIXME: this could be merged to shift decision code
1308 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1309 u32 start_seq, u32 end_seq)
1313 unsigned int pkt_len;
1316 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1317 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1319 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1320 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1321 mss = tcp_skb_mss(skb);
1322 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1325 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1329 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1334 /* Round if necessary so that SACKs cover only full MSSes
1335 * and/or the remaining small portion (if present)
1337 if (pkt_len > mss) {
1338 unsigned int new_len = (pkt_len / mss) * mss;
1339 if (!in_sack && new_len < pkt_len)
1344 if (pkt_len >= skb->len && !in_sack)
1347 err = tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
1348 pkt_len, mss, GFP_ATOMIC);
1356 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1357 static u8 tcp_sacktag_one(struct sock *sk,
1358 struct tcp_sacktag_state *state, u8 sacked,
1359 u32 start_seq, u32 end_seq,
1360 int dup_sack, int pcount,
1363 struct tcp_sock *tp = tcp_sk(sk);
1365 /* Account D-SACK for retransmitted packet. */
1366 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1367 if (tp->undo_marker && tp->undo_retrans > 0 &&
1368 after(end_seq, tp->undo_marker))
1369 tp->undo_retrans = max_t(int, 0, tp->undo_retrans - pcount);
1370 if ((sacked & TCPCB_SACKED_ACKED) &&
1371 before(start_seq, state->reord))
1372 state->reord = start_seq;
1375 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1376 if (!after(end_seq, tp->snd_una))
1379 if (!(sacked & TCPCB_SACKED_ACKED)) {
1380 tcp_rack_advance(tp, sacked, end_seq, xmit_time);
1382 if (sacked & TCPCB_SACKED_RETRANS) {
1383 /* If the segment is not tagged as lost,
1384 * we do not clear RETRANS, believing
1385 * that retransmission is still in flight.
1387 if (sacked & TCPCB_LOST) {
1388 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1389 tp->lost_out -= pcount;
1390 tp->retrans_out -= pcount;
1393 if (!(sacked & TCPCB_RETRANS)) {
1394 /* New sack for not retransmitted frame,
1395 * which was in hole. It is reordering.
1397 if (before(start_seq,
1398 tcp_highest_sack_seq(tp)) &&
1399 before(start_seq, state->reord))
1400 state->reord = start_seq;
1402 if (!after(end_seq, tp->high_seq))
1403 state->flag |= FLAG_ORIG_SACK_ACKED;
1404 if (state->first_sackt == 0)
1405 state->first_sackt = xmit_time;
1406 state->last_sackt = xmit_time;
1409 if (sacked & TCPCB_LOST) {
1410 sacked &= ~TCPCB_LOST;
1411 tp->lost_out -= pcount;
1415 sacked |= TCPCB_SACKED_ACKED;
1416 state->flag |= FLAG_DATA_SACKED;
1417 tp->sacked_out += pcount;
1418 /* Out-of-order packets delivered */
1419 state->sack_delivered += pcount;
1421 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1422 if (tp->lost_skb_hint &&
1423 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1424 tp->lost_cnt_hint += pcount;
1427 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1428 * frames and clear it. undo_retrans is decreased above, L|R frames
1429 * are accounted above as well.
1431 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1432 sacked &= ~TCPCB_SACKED_RETRANS;
1433 tp->retrans_out -= pcount;
1439 /* Shift newly-SACKed bytes from this skb to the immediately previous
1440 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1442 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *prev,
1443 struct sk_buff *skb,
1444 struct tcp_sacktag_state *state,
1445 unsigned int pcount, int shifted, int mss,
1448 struct tcp_sock *tp = tcp_sk(sk);
1449 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1450 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
1454 /* Adjust counters and hints for the newly sacked sequence
1455 * range but discard the return value since prev is already
1456 * marked. We must tag the range first because the seq
1457 * advancement below implicitly advances
1458 * tcp_highest_sack_seq() when skb is highest_sack.
1460 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1461 start_seq, end_seq, dup_sack, pcount,
1462 tcp_skb_timestamp_us(skb));
1463 tcp_rate_skb_delivered(sk, skb, state->rate);
1465 if (skb == tp->lost_skb_hint)
1466 tp->lost_cnt_hint += pcount;
1468 TCP_SKB_CB(prev)->end_seq += shifted;
1469 TCP_SKB_CB(skb)->seq += shifted;
1471 tcp_skb_pcount_add(prev, pcount);
1472 WARN_ON_ONCE(tcp_skb_pcount(skb) < pcount);
1473 tcp_skb_pcount_add(skb, -pcount);
1475 /* When we're adding to gso_segs == 1, gso_size will be zero,
1476 * in theory this shouldn't be necessary but as long as DSACK
1477 * code can come after this skb later on it's better to keep
1478 * setting gso_size to something.
1480 if (!TCP_SKB_CB(prev)->tcp_gso_size)
1481 TCP_SKB_CB(prev)->tcp_gso_size = mss;
1483 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1484 if (tcp_skb_pcount(skb) <= 1)
1485 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1487 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1488 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1491 BUG_ON(!tcp_skb_pcount(skb));
1492 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1496 /* Whole SKB was eaten :-) */
1498 if (skb == tp->retransmit_skb_hint)
1499 tp->retransmit_skb_hint = prev;
1500 if (skb == tp->lost_skb_hint) {
1501 tp->lost_skb_hint = prev;
1502 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1505 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1506 TCP_SKB_CB(prev)->eor = TCP_SKB_CB(skb)->eor;
1507 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1508 TCP_SKB_CB(prev)->end_seq++;
1510 if (skb == tcp_highest_sack(sk))
1511 tcp_advance_highest_sack(sk, skb);
1513 tcp_skb_collapse_tstamp(prev, skb);
1514 if (unlikely(TCP_SKB_CB(prev)->tx.delivered_mstamp))
1515 TCP_SKB_CB(prev)->tx.delivered_mstamp = 0;
1517 tcp_rtx_queue_unlink_and_free(skb, sk);
1519 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKMERGED);
1524 /* I wish gso_size would have a bit more sane initialization than
1525 * something-or-zero which complicates things
1527 static int tcp_skb_seglen(const struct sk_buff *skb)
1529 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1532 /* Shifting pages past head area doesn't work */
1533 static int skb_can_shift(const struct sk_buff *skb)
1535 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1538 int tcp_skb_shift(struct sk_buff *to, struct sk_buff *from,
1539 int pcount, int shiftlen)
1541 /* TCP min gso_size is 8 bytes (TCP_MIN_GSO_SIZE)
1542 * Since TCP_SKB_CB(skb)->tcp_gso_segs is 16 bits, we need
1543 * to make sure not storing more than 65535 * 8 bytes per skb,
1544 * even if current MSS is bigger.
1546 if (unlikely(to->len + shiftlen >= 65535 * TCP_MIN_GSO_SIZE))
1548 if (unlikely(tcp_skb_pcount(to) + pcount > 65535))
1550 return skb_shift(to, from, shiftlen);
1553 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1556 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1557 struct tcp_sacktag_state *state,
1558 u32 start_seq, u32 end_seq,
1561 struct tcp_sock *tp = tcp_sk(sk);
1562 struct sk_buff *prev;
1568 /* Normally R but no L won't result in plain S */
1570 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1572 if (!skb_can_shift(skb))
1574 /* This frame is about to be dropped (was ACKed). */
1575 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1578 /* Can only happen with delayed DSACK + discard craziness */
1579 prev = skb_rb_prev(skb);
1583 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1586 if (!tcp_skb_can_collapse(prev, skb))
1589 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1590 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1594 pcount = tcp_skb_pcount(skb);
1595 mss = tcp_skb_seglen(skb);
1597 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1598 * drop this restriction as unnecessary
1600 if (mss != tcp_skb_seglen(prev))
1603 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1605 /* CHECKME: This is non-MSS split case only?, this will
1606 * cause skipped skbs due to advancing loop btw, original
1607 * has that feature too
1609 if (tcp_skb_pcount(skb) <= 1)
1612 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1614 /* TODO: head merge to next could be attempted here
1615 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1616 * though it might not be worth of the additional hassle
1618 * ...we can probably just fallback to what was done
1619 * previously. We could try merging non-SACKed ones
1620 * as well but it probably isn't going to buy off
1621 * because later SACKs might again split them, and
1622 * it would make skb timestamp tracking considerably
1628 len = end_seq - TCP_SKB_CB(skb)->seq;
1630 BUG_ON(len > skb->len);
1632 /* MSS boundaries should be honoured or else pcount will
1633 * severely break even though it makes things bit trickier.
1634 * Optimize common case to avoid most of the divides
1636 mss = tcp_skb_mss(skb);
1638 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1639 * drop this restriction as unnecessary
1641 if (mss != tcp_skb_seglen(prev))
1646 } else if (len < mss) {
1654 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1655 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1658 if (!tcp_skb_shift(prev, skb, pcount, len))
1660 if (!tcp_shifted_skb(sk, prev, skb, state, pcount, len, mss, dup_sack))
1663 /* Hole filled allows collapsing with the next as well, this is very
1664 * useful when hole on every nth skb pattern happens
1666 skb = skb_rb_next(prev);
1670 if (!skb_can_shift(skb) ||
1671 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1672 (mss != tcp_skb_seglen(skb)))
1675 if (!tcp_skb_can_collapse(prev, skb))
1678 pcount = tcp_skb_pcount(skb);
1679 if (tcp_skb_shift(prev, skb, pcount, len))
1680 tcp_shifted_skb(sk, prev, skb, state, pcount,
1690 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1694 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1695 struct tcp_sack_block *next_dup,
1696 struct tcp_sacktag_state *state,
1697 u32 start_seq, u32 end_seq,
1700 struct tcp_sock *tp = tcp_sk(sk);
1701 struct sk_buff *tmp;
1703 skb_rbtree_walk_from(skb) {
1705 bool dup_sack = dup_sack_in;
1707 /* queue is in-order => we can short-circuit the walk early */
1708 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1712 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1713 in_sack = tcp_match_skb_to_sack(sk, skb,
1714 next_dup->start_seq,
1720 /* skb reference here is a bit tricky to get right, since
1721 * shifting can eat and free both this skb and the next,
1722 * so not even _safe variant of the loop is enough.
1725 tmp = tcp_shift_skb_data(sk, skb, state,
1726 start_seq, end_seq, dup_sack);
1735 in_sack = tcp_match_skb_to_sack(sk, skb,
1741 if (unlikely(in_sack < 0))
1745 TCP_SKB_CB(skb)->sacked =
1748 TCP_SKB_CB(skb)->sacked,
1749 TCP_SKB_CB(skb)->seq,
1750 TCP_SKB_CB(skb)->end_seq,
1752 tcp_skb_pcount(skb),
1753 tcp_skb_timestamp_us(skb));
1754 tcp_rate_skb_delivered(sk, skb, state->rate);
1755 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1756 list_del_init(&skb->tcp_tsorted_anchor);
1758 if (!before(TCP_SKB_CB(skb)->seq,
1759 tcp_highest_sack_seq(tp)))
1760 tcp_advance_highest_sack(sk, skb);
1766 static struct sk_buff *tcp_sacktag_bsearch(struct sock *sk, u32 seq)
1768 struct rb_node *parent, **p = &sk->tcp_rtx_queue.rb_node;
1769 struct sk_buff *skb;
1773 skb = rb_to_skb(parent);
1774 if (before(seq, TCP_SKB_CB(skb)->seq)) {
1775 p = &parent->rb_left;
1778 if (!before(seq, TCP_SKB_CB(skb)->end_seq)) {
1779 p = &parent->rb_right;
1787 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1790 if (skb && after(TCP_SKB_CB(skb)->seq, skip_to_seq))
1793 return tcp_sacktag_bsearch(sk, skip_to_seq);
1796 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1798 struct tcp_sack_block *next_dup,
1799 struct tcp_sacktag_state *state,
1805 if (before(next_dup->start_seq, skip_to_seq)) {
1806 skb = tcp_sacktag_skip(skb, sk, next_dup->start_seq);
1807 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1808 next_dup->start_seq, next_dup->end_seq,
1815 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1817 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1821 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1822 u32 prior_snd_una, struct tcp_sacktag_state *state)
1824 struct tcp_sock *tp = tcp_sk(sk);
1825 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1826 TCP_SKB_CB(ack_skb)->sacked);
1827 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1828 struct tcp_sack_block sp[TCP_NUM_SACKS];
1829 struct tcp_sack_block *cache;
1830 struct sk_buff *skb;
1831 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1833 bool found_dup_sack = false;
1835 int first_sack_index;
1838 state->reord = tp->snd_nxt;
1840 if (!tp->sacked_out)
1841 tcp_highest_sack_reset(sk);
1843 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1844 num_sacks, prior_snd_una, state);
1846 /* Eliminate too old ACKs, but take into
1847 * account more or less fresh ones, they can
1848 * contain valid SACK info.
1850 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1853 if (!tp->packets_out)
1857 first_sack_index = 0;
1858 for (i = 0; i < num_sacks; i++) {
1859 bool dup_sack = !i && found_dup_sack;
1861 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1862 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1864 if (!tcp_is_sackblock_valid(tp, dup_sack,
1865 sp[used_sacks].start_seq,
1866 sp[used_sacks].end_seq)) {
1870 if (!tp->undo_marker)
1871 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1873 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1875 /* Don't count olds caused by ACK reordering */
1876 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1877 !after(sp[used_sacks].end_seq, tp->snd_una))
1879 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1882 NET_INC_STATS(sock_net(sk), mib_idx);
1884 first_sack_index = -1;
1888 /* Ignore very old stuff early */
1889 if (!after(sp[used_sacks].end_seq, prior_snd_una)) {
1891 first_sack_index = -1;
1898 /* order SACK blocks to allow in order walk of the retrans queue */
1899 for (i = used_sacks - 1; i > 0; i--) {
1900 for (j = 0; j < i; j++) {
1901 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1902 swap(sp[j], sp[j + 1]);
1904 /* Track where the first SACK block goes to */
1905 if (j == first_sack_index)
1906 first_sack_index = j + 1;
1911 state->mss_now = tcp_current_mss(sk);
1915 if (!tp->sacked_out) {
1916 /* It's already past, so skip checking against it */
1917 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1919 cache = tp->recv_sack_cache;
1920 /* Skip empty blocks in at head of the cache */
1921 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1926 while (i < used_sacks) {
1927 u32 start_seq = sp[i].start_seq;
1928 u32 end_seq = sp[i].end_seq;
1929 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1930 struct tcp_sack_block *next_dup = NULL;
1932 if (found_dup_sack && ((i + 1) == first_sack_index))
1933 next_dup = &sp[i + 1];
1935 /* Skip too early cached blocks */
1936 while (tcp_sack_cache_ok(tp, cache) &&
1937 !before(start_seq, cache->end_seq))
1940 /* Can skip some work by looking recv_sack_cache? */
1941 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1942 after(end_seq, cache->start_seq)) {
1945 if (before(start_seq, cache->start_seq)) {
1946 skb = tcp_sacktag_skip(skb, sk, start_seq);
1947 skb = tcp_sacktag_walk(skb, sk, next_dup,
1954 /* Rest of the block already fully processed? */
1955 if (!after(end_seq, cache->end_seq))
1958 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1962 /* ...tail remains todo... */
1963 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1964 /* ...but better entrypoint exists! */
1965 skb = tcp_highest_sack(sk);
1972 skb = tcp_sacktag_skip(skb, sk, cache->end_seq);
1973 /* Check overlap against next cached too (past this one already) */
1978 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1979 skb = tcp_highest_sack(sk);
1983 skb = tcp_sacktag_skip(skb, sk, start_seq);
1986 skb = tcp_sacktag_walk(skb, sk, next_dup, state,
1987 start_seq, end_seq, dup_sack);
1993 /* Clear the head of the cache sack blocks so we can skip it next time */
1994 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1995 tp->recv_sack_cache[i].start_seq = 0;
1996 tp->recv_sack_cache[i].end_seq = 0;
1998 for (j = 0; j < used_sacks; j++)
1999 tp->recv_sack_cache[i++] = sp[j];
2001 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Loss || tp->undo_marker)
2002 tcp_check_sack_reordering(sk, state->reord, 0);
2004 tcp_verify_left_out(tp);
2007 #if FASTRETRANS_DEBUG > 0
2008 WARN_ON((int)tp->sacked_out < 0);
2009 WARN_ON((int)tp->lost_out < 0);
2010 WARN_ON((int)tp->retrans_out < 0);
2011 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
2016 /* Limits sacked_out so that sum with lost_out isn't ever larger than
2017 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
2019 static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
2023 holes = max(tp->lost_out, 1U);
2024 holes = min(holes, tp->packets_out);
2026 if ((tp->sacked_out + holes) > tp->packets_out) {
2027 tp->sacked_out = tp->packets_out - holes;
2033 /* If we receive more dupacks than we expected counting segments
2034 * in assumption of absent reordering, interpret this as reordering.
2035 * The only another reason could be bug in receiver TCP.
2037 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
2039 struct tcp_sock *tp = tcp_sk(sk);
2041 if (!tcp_limit_reno_sacked(tp))
2044 tp->reordering = min_t(u32, tp->packets_out + addend,
2045 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_max_reordering));
2047 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRENOREORDER);
2050 /* Emulate SACKs for SACKless connection: account for a new dupack. */
2052 static void tcp_add_reno_sack(struct sock *sk, int num_dupack, bool ece_ack)
2055 struct tcp_sock *tp = tcp_sk(sk);
2056 u32 prior_sacked = tp->sacked_out;
2059 tp->sacked_out += num_dupack;
2060 tcp_check_reno_reordering(sk, 0);
2061 delivered = tp->sacked_out - prior_sacked;
2063 tcp_count_delivered(tp, delivered, ece_ack);
2064 tcp_verify_left_out(tp);
2068 /* Account for ACK, ACKing some data in Reno Recovery phase. */
2070 static void tcp_remove_reno_sacks(struct sock *sk, int acked, bool ece_ack)
2072 struct tcp_sock *tp = tcp_sk(sk);
2075 /* One ACK acked hole. The rest eat duplicate ACKs. */
2076 tcp_count_delivered(tp, max_t(int, acked - tp->sacked_out, 1),
2078 if (acked - 1 >= tp->sacked_out)
2081 tp->sacked_out -= acked - 1;
2083 tcp_check_reno_reordering(sk, acked);
2084 tcp_verify_left_out(tp);
2087 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
2092 void tcp_clear_retrans(struct tcp_sock *tp)
2094 tp->retrans_out = 0;
2096 tp->undo_marker = 0;
2097 tp->undo_retrans = -1;
2101 static inline void tcp_init_undo(struct tcp_sock *tp)
2103 tp->undo_marker = tp->snd_una;
2104 /* Retransmission still in flight may cause DSACKs later. */
2105 tp->undo_retrans = tp->retrans_out ? : -1;
2108 static bool tcp_is_rack(const struct sock *sk)
2110 return READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_recovery) &
2111 TCP_RACK_LOSS_DETECTION;
2114 /* If we detect SACK reneging, forget all SACK information
2115 * and reset tags completely, otherwise preserve SACKs. If receiver
2116 * dropped its ofo queue, we will know this due to reneging detection.
2118 static void tcp_timeout_mark_lost(struct sock *sk)
2120 struct tcp_sock *tp = tcp_sk(sk);
2121 struct sk_buff *skb, *head;
2122 bool is_reneg; /* is receiver reneging on SACKs? */
2124 head = tcp_rtx_queue_head(sk);
2125 is_reneg = head && (TCP_SKB_CB(head)->sacked & TCPCB_SACKED_ACKED);
2127 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
2129 /* Mark SACK reneging until we recover from this loss event. */
2130 tp->is_sack_reneg = 1;
2131 } else if (tcp_is_reno(tp)) {
2132 tcp_reset_reno_sack(tp);
2136 skb_rbtree_walk_from(skb) {
2138 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
2139 else if (tcp_is_rack(sk) && skb != head &&
2140 tcp_rack_skb_timeout(tp, skb, 0) > 0)
2141 continue; /* Don't mark recently sent ones lost yet */
2142 tcp_mark_skb_lost(sk, skb);
2144 tcp_verify_left_out(tp);
2145 tcp_clear_all_retrans_hints(tp);
2148 /* Enter Loss state. */
2149 void tcp_enter_loss(struct sock *sk)
2151 const struct inet_connection_sock *icsk = inet_csk(sk);
2152 struct tcp_sock *tp = tcp_sk(sk);
2153 struct net *net = sock_net(sk);
2154 bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery;
2157 tcp_timeout_mark_lost(sk);
2159 /* Reduce ssthresh if it has not yet been made inside this window. */
2160 if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
2161 !after(tp->high_seq, tp->snd_una) ||
2162 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
2163 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2164 tp->prior_cwnd = tcp_snd_cwnd(tp);
2165 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2166 tcp_ca_event(sk, CA_EVENT_LOSS);
2169 tcp_snd_cwnd_set(tp, tcp_packets_in_flight(tp) + 1);
2170 tp->snd_cwnd_cnt = 0;
2171 tp->snd_cwnd_stamp = tcp_jiffies32;
2173 /* Timeout in disordered state after receiving substantial DUPACKs
2174 * suggests that the degree of reordering is over-estimated.
2176 reordering = READ_ONCE(net->ipv4.sysctl_tcp_reordering);
2177 if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
2178 tp->sacked_out >= reordering)
2179 tp->reordering = min_t(unsigned int, tp->reordering,
2182 tcp_set_ca_state(sk, TCP_CA_Loss);
2183 tp->high_seq = tp->snd_nxt;
2184 tcp_ecn_queue_cwr(tp);
2186 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
2187 * loss recovery is underway except recurring timeout(s) on
2188 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
2190 tp->frto = READ_ONCE(net->ipv4.sysctl_tcp_frto) &&
2191 (new_recovery || icsk->icsk_retransmits) &&
2192 !inet_csk(sk)->icsk_mtup.probe_size;
2195 /* If ACK arrived pointing to a remembered SACK, it means that our
2196 * remembered SACKs do not reflect real state of receiver i.e.
2197 * receiver _host_ is heavily congested (or buggy).
2199 * To avoid big spurious retransmission bursts due to transient SACK
2200 * scoreboard oddities that look like reneging, we give the receiver a
2201 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
2202 * restore sanity to the SACK scoreboard. If the apparent reneging
2203 * persists until this RTO then we'll clear the SACK scoreboard.
2205 static bool tcp_check_sack_reneging(struct sock *sk, int *ack_flag)
2207 if (*ack_flag & FLAG_SACK_RENEGING &&
2208 *ack_flag & FLAG_SND_UNA_ADVANCED) {
2209 struct tcp_sock *tp = tcp_sk(sk);
2210 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4),
2211 msecs_to_jiffies(10));
2213 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2214 delay, TCP_RTO_MAX);
2215 *ack_flag &= ~FLAG_SET_XMIT_TIMER;
2221 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2222 * counter when SACK is enabled (without SACK, sacked_out is used for
2225 * With reordering, holes may still be in flight, so RFC3517 recovery
2226 * uses pure sacked_out (total number of SACKed segments) even though
2227 * it violates the RFC that uses duplicate ACKs, often these are equal
2228 * but when e.g. out-of-window ACKs or packet duplication occurs,
2229 * they differ. Since neither occurs due to loss, TCP should really
2232 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2234 return tp->sacked_out + 1;
2237 /* Linux NewReno/SACK/ECN state machine.
2238 * --------------------------------------
2240 * "Open" Normal state, no dubious events, fast path.
2241 * "Disorder" In all the respects it is "Open",
2242 * but requires a bit more attention. It is entered when
2243 * we see some SACKs or dupacks. It is split of "Open"
2244 * mainly to move some processing from fast path to slow one.
2245 * "CWR" CWND was reduced due to some Congestion Notification event.
2246 * It can be ECN, ICMP source quench, local device congestion.
2247 * "Recovery" CWND was reduced, we are fast-retransmitting.
2248 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2250 * tcp_fastretrans_alert() is entered:
2251 * - each incoming ACK, if state is not "Open"
2252 * - when arrived ACK is unusual, namely:
2257 * Counting packets in flight is pretty simple.
2259 * in_flight = packets_out - left_out + retrans_out
2261 * packets_out is SND.NXT-SND.UNA counted in packets.
2263 * retrans_out is number of retransmitted segments.
2265 * left_out is number of segments left network, but not ACKed yet.
2267 * left_out = sacked_out + lost_out
2269 * sacked_out: Packets, which arrived to receiver out of order
2270 * and hence not ACKed. With SACKs this number is simply
2271 * amount of SACKed data. Even without SACKs
2272 * it is easy to give pretty reliable estimate of this number,
2273 * counting duplicate ACKs.
2275 * lost_out: Packets lost by network. TCP has no explicit
2276 * "loss notification" feedback from network (for now).
2277 * It means that this number can be only _guessed_.
2278 * Actually, it is the heuristics to predict lossage that
2279 * distinguishes different algorithms.
2281 * F.e. after RTO, when all the queue is considered as lost,
2282 * lost_out = packets_out and in_flight = retrans_out.
2284 * Essentially, we have now a few algorithms detecting
2287 * If the receiver supports SACK:
2289 * RFC6675/3517: It is the conventional algorithm. A packet is
2290 * considered lost if the number of higher sequence packets
2291 * SACKed is greater than or equal the DUPACK thoreshold
2292 * (reordering). This is implemented in tcp_mark_head_lost and
2293 * tcp_update_scoreboard.
2295 * RACK (draft-ietf-tcpm-rack-01): it is a newer algorithm
2296 * (2017-) that checks timing instead of counting DUPACKs.
2297 * Essentially a packet is considered lost if it's not S/ACKed
2298 * after RTT + reordering_window, where both metrics are
2299 * dynamically measured and adjusted. This is implemented in
2300 * tcp_rack_mark_lost.
2302 * If the receiver does not support SACK:
2304 * NewReno (RFC6582): in Recovery we assume that one segment
2305 * is lost (classic Reno). While we are in Recovery and
2306 * a partial ACK arrives, we assume that one more packet
2307 * is lost (NewReno). This heuristics are the same in NewReno
2310 * Really tricky (and requiring careful tuning) part of algorithm
2311 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2312 * The first determines the moment _when_ we should reduce CWND and,
2313 * hence, slow down forward transmission. In fact, it determines the moment
2314 * when we decide that hole is caused by loss, rather than by a reorder.
2316 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2317 * holes, caused by lost packets.
2319 * And the most logically complicated part of algorithm is undo
2320 * heuristics. We detect false retransmits due to both too early
2321 * fast retransmit (reordering) and underestimated RTO, analyzing
2322 * timestamps and D-SACKs. When we detect that some segments were
2323 * retransmitted by mistake and CWND reduction was wrong, we undo
2324 * window reduction and abort recovery phase. This logic is hidden
2325 * inside several functions named tcp_try_undo_<something>.
2328 /* This function decides, when we should leave Disordered state
2329 * and enter Recovery phase, reducing congestion window.
2331 * Main question: may we further continue forward transmission
2332 * with the same cwnd?
2334 static bool tcp_time_to_recover(struct sock *sk, int flag)
2336 struct tcp_sock *tp = tcp_sk(sk);
2338 /* Trick#1: The loss is proven. */
2342 /* Not-A-Trick#2 : Classic rule... */
2343 if (!tcp_is_rack(sk) && tcp_dupack_heuristics(tp) > tp->reordering)
2349 /* Detect loss in event "A" above by marking head of queue up as lost.
2350 * For RFC3517 SACK, a segment is considered lost if it
2351 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2352 * the maximum SACKed segments to pass before reaching this limit.
2354 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2356 struct tcp_sock *tp = tcp_sk(sk);
2357 struct sk_buff *skb;
2359 /* Use SACK to deduce losses of new sequences sent during recovery */
2360 const u32 loss_high = tp->snd_nxt;
2362 WARN_ON(packets > tp->packets_out);
2363 skb = tp->lost_skb_hint;
2365 /* Head already handled? */
2366 if (mark_head && after(TCP_SKB_CB(skb)->seq, tp->snd_una))
2368 cnt = tp->lost_cnt_hint;
2370 skb = tcp_rtx_queue_head(sk);
2374 skb_rbtree_walk_from(skb) {
2375 /* TODO: do this better */
2376 /* this is not the most efficient way to do this... */
2377 tp->lost_skb_hint = skb;
2378 tp->lost_cnt_hint = cnt;
2380 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2383 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2384 cnt += tcp_skb_pcount(skb);
2389 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_LOST))
2390 tcp_mark_skb_lost(sk, skb);
2395 tcp_verify_left_out(tp);
2398 /* Account newly detected lost packet(s) */
2400 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2402 struct tcp_sock *tp = tcp_sk(sk);
2404 if (tcp_is_sack(tp)) {
2405 int sacked_upto = tp->sacked_out - tp->reordering;
2406 if (sacked_upto >= 0)
2407 tcp_mark_head_lost(sk, sacked_upto, 0);
2408 else if (fast_rexmit)
2409 tcp_mark_head_lost(sk, 1, 1);
2413 static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when)
2415 return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2416 before(tp->rx_opt.rcv_tsecr, when);
2419 /* skb is spurious retransmitted if the returned timestamp echo
2420 * reply is prior to the skb transmission time
2422 static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp,
2423 const struct sk_buff *skb)
2425 return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) &&
2426 tcp_tsopt_ecr_before(tp, tcp_skb_timestamp(skb));
2429 /* Nothing was retransmitted or returned timestamp is less
2430 * than timestamp of the first retransmission.
2432 static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2434 return tp->retrans_stamp &&
2435 tcp_tsopt_ecr_before(tp, tp->retrans_stamp);
2438 /* Undo procedures. */
2440 /* We can clear retrans_stamp when there are no retransmissions in the
2441 * window. It would seem that it is trivially available for us in
2442 * tp->retrans_out, however, that kind of assumptions doesn't consider
2443 * what will happen if errors occur when sending retransmission for the
2444 * second time. ...It could the that such segment has only
2445 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2446 * the head skb is enough except for some reneging corner cases that
2447 * are not worth the effort.
2449 * Main reason for all this complexity is the fact that connection dying
2450 * time now depends on the validity of the retrans_stamp, in particular,
2451 * that successive retransmissions of a segment must not advance
2452 * retrans_stamp under any conditions.
2454 static bool tcp_any_retrans_done(const struct sock *sk)
2456 const struct tcp_sock *tp = tcp_sk(sk);
2457 struct sk_buff *skb;
2459 if (tp->retrans_out)
2462 skb = tcp_rtx_queue_head(sk);
2463 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2469 static void DBGUNDO(struct sock *sk, const char *msg)
2471 #if FASTRETRANS_DEBUG > 1
2472 struct tcp_sock *tp = tcp_sk(sk);
2473 struct inet_sock *inet = inet_sk(sk);
2475 if (sk->sk_family == AF_INET) {
2476 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2478 &inet->inet_daddr, ntohs(inet->inet_dport),
2479 tcp_snd_cwnd(tp), tcp_left_out(tp),
2480 tp->snd_ssthresh, tp->prior_ssthresh,
2483 #if IS_ENABLED(CONFIG_IPV6)
2484 else if (sk->sk_family == AF_INET6) {
2485 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2487 &sk->sk_v6_daddr, ntohs(inet->inet_dport),
2488 tcp_snd_cwnd(tp), tcp_left_out(tp),
2489 tp->snd_ssthresh, tp->prior_ssthresh,
2496 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
2498 struct tcp_sock *tp = tcp_sk(sk);
2501 struct sk_buff *skb;
2503 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2504 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2507 tcp_clear_all_retrans_hints(tp);
2510 if (tp->prior_ssthresh) {
2511 const struct inet_connection_sock *icsk = inet_csk(sk);
2513 tcp_snd_cwnd_set(tp, icsk->icsk_ca_ops->undo_cwnd(sk));
2515 if (tp->prior_ssthresh > tp->snd_ssthresh) {
2516 tp->snd_ssthresh = tp->prior_ssthresh;
2517 tcp_ecn_withdraw_cwr(tp);
2520 tp->snd_cwnd_stamp = tcp_jiffies32;
2521 tp->undo_marker = 0;
2522 tp->rack.advanced = 1; /* Force RACK to re-exam losses */
2525 static inline bool tcp_may_undo(const struct tcp_sock *tp)
2527 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2530 static bool tcp_is_non_sack_preventing_reopen(struct sock *sk)
2532 struct tcp_sock *tp = tcp_sk(sk);
2534 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2535 /* Hold old state until something *above* high_seq
2536 * is ACKed. For Reno it is MUST to prevent false
2537 * fast retransmits (RFC2582). SACK TCP is safe. */
2538 if (!tcp_any_retrans_done(sk))
2539 tp->retrans_stamp = 0;
2545 /* People celebrate: "We love our President!" */
2546 static bool tcp_try_undo_recovery(struct sock *sk)
2548 struct tcp_sock *tp = tcp_sk(sk);
2550 if (tcp_may_undo(tp)) {
2553 /* Happy end! We did not retransmit anything
2554 * or our original transmission succeeded.
2556 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2557 tcp_undo_cwnd_reduction(sk, false);
2558 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2559 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2561 mib_idx = LINUX_MIB_TCPFULLUNDO;
2563 NET_INC_STATS(sock_net(sk), mib_idx);
2564 } else if (tp->rack.reo_wnd_persist) {
2565 tp->rack.reo_wnd_persist--;
2567 if (tcp_is_non_sack_preventing_reopen(sk))
2569 tcp_set_ca_state(sk, TCP_CA_Open);
2570 tp->is_sack_reneg = 0;
2574 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2575 static bool tcp_try_undo_dsack(struct sock *sk)
2577 struct tcp_sock *tp = tcp_sk(sk);
2579 if (tp->undo_marker && !tp->undo_retrans) {
2580 tp->rack.reo_wnd_persist = min(TCP_RACK_RECOVERY_THRESH,
2581 tp->rack.reo_wnd_persist + 1);
2582 DBGUNDO(sk, "D-SACK");
2583 tcp_undo_cwnd_reduction(sk, false);
2584 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2590 /* Undo during loss recovery after partial ACK or using F-RTO. */
2591 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2593 struct tcp_sock *tp = tcp_sk(sk);
2595 if (frto_undo || tcp_may_undo(tp)) {
2596 tcp_undo_cwnd_reduction(sk, true);
2598 DBGUNDO(sk, "partial loss");
2599 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2601 NET_INC_STATS(sock_net(sk),
2602 LINUX_MIB_TCPSPURIOUSRTOS);
2603 inet_csk(sk)->icsk_retransmits = 0;
2604 if (tcp_is_non_sack_preventing_reopen(sk))
2606 if (frto_undo || tcp_is_sack(tp)) {
2607 tcp_set_ca_state(sk, TCP_CA_Open);
2608 tp->is_sack_reneg = 0;
2615 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
2616 * It computes the number of packets to send (sndcnt) based on packets newly
2618 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2619 * cwnd reductions across a full RTT.
2620 * 2) Otherwise PRR uses packet conservation to send as much as delivered.
2621 * But when SND_UNA is acked without further losses,
2622 * slow starts cwnd up to ssthresh to speed up the recovery.
2624 static void tcp_init_cwnd_reduction(struct sock *sk)
2626 struct tcp_sock *tp = tcp_sk(sk);
2628 tp->high_seq = tp->snd_nxt;
2629 tp->tlp_high_seq = 0;
2630 tp->snd_cwnd_cnt = 0;
2631 tp->prior_cwnd = tcp_snd_cwnd(tp);
2632 tp->prr_delivered = 0;
2634 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2635 tcp_ecn_queue_cwr(tp);
2638 void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked, int newly_lost, int flag)
2640 struct tcp_sock *tp = tcp_sk(sk);
2642 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2644 if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd))
2647 tp->prr_delivered += newly_acked_sacked;
2649 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2651 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2653 sndcnt = max_t(int, tp->prr_delivered - tp->prr_out,
2654 newly_acked_sacked);
2655 if (flag & FLAG_SND_UNA_ADVANCED && !newly_lost)
2657 sndcnt = min(delta, sndcnt);
2659 /* Force a fast retransmit upon entering fast recovery */
2660 sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1));
2661 tcp_snd_cwnd_set(tp, tcp_packets_in_flight(tp) + sndcnt);
2664 static inline void tcp_end_cwnd_reduction(struct sock *sk)
2666 struct tcp_sock *tp = tcp_sk(sk);
2668 if (inet_csk(sk)->icsk_ca_ops->cong_control)
2671 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2672 if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH &&
2673 (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR || tp->undo_marker)) {
2674 tcp_snd_cwnd_set(tp, tp->snd_ssthresh);
2675 tp->snd_cwnd_stamp = tcp_jiffies32;
2677 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2680 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2681 void tcp_enter_cwr(struct sock *sk)
2683 struct tcp_sock *tp = tcp_sk(sk);
2685 tp->prior_ssthresh = 0;
2686 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2687 tp->undo_marker = 0;
2688 tcp_init_cwnd_reduction(sk);
2689 tcp_set_ca_state(sk, TCP_CA_CWR);
2692 EXPORT_SYMBOL(tcp_enter_cwr);
2694 static void tcp_try_keep_open(struct sock *sk)
2696 struct tcp_sock *tp = tcp_sk(sk);
2697 int state = TCP_CA_Open;
2699 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2700 state = TCP_CA_Disorder;
2702 if (inet_csk(sk)->icsk_ca_state != state) {
2703 tcp_set_ca_state(sk, state);
2704 tp->high_seq = tp->snd_nxt;
2708 static void tcp_try_to_open(struct sock *sk, int flag)
2710 struct tcp_sock *tp = tcp_sk(sk);
2712 tcp_verify_left_out(tp);
2714 if (!tcp_any_retrans_done(sk))
2715 tp->retrans_stamp = 0;
2717 if (flag & FLAG_ECE)
2720 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2721 tcp_try_keep_open(sk);
2725 static void tcp_mtup_probe_failed(struct sock *sk)
2727 struct inet_connection_sock *icsk = inet_csk(sk);
2729 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2730 icsk->icsk_mtup.probe_size = 0;
2731 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPFAIL);
2734 static void tcp_mtup_probe_success(struct sock *sk)
2736 struct tcp_sock *tp = tcp_sk(sk);
2737 struct inet_connection_sock *icsk = inet_csk(sk);
2740 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2742 val = (u64)tcp_snd_cwnd(tp) * tcp_mss_to_mtu(sk, tp->mss_cache);
2743 do_div(val, icsk->icsk_mtup.probe_size);
2744 DEBUG_NET_WARN_ON_ONCE((u32)val != val);
2745 tcp_snd_cwnd_set(tp, max_t(u32, 1U, val));
2747 tp->snd_cwnd_cnt = 0;
2748 tp->snd_cwnd_stamp = tcp_jiffies32;
2749 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2751 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2752 icsk->icsk_mtup.probe_size = 0;
2753 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2754 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS);
2757 /* Do a simple retransmit without using the backoff mechanisms in
2758 * tcp_timer. This is used for path mtu discovery.
2759 * The socket is already locked here.
2761 void tcp_simple_retransmit(struct sock *sk)
2763 const struct inet_connection_sock *icsk = inet_csk(sk);
2764 struct tcp_sock *tp = tcp_sk(sk);
2765 struct sk_buff *skb;
2768 /* A fastopen SYN request is stored as two separate packets within
2769 * the retransmit queue, this is done by tcp_send_syn_data().
2770 * As a result simply checking the MSS of the frames in the queue
2771 * will not work for the SYN packet.
2773 * Us being here is an indication of a path MTU issue so we can
2774 * assume that the fastopen SYN was lost and just mark all the
2775 * frames in the retransmit queue as lost. We will use an MSS of
2776 * -1 to mark all frames as lost, otherwise compute the current MSS.
2778 if (tp->syn_data && sk->sk_state == TCP_SYN_SENT)
2781 mss = tcp_current_mss(sk);
2783 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2784 if (tcp_skb_seglen(skb) > mss)
2785 tcp_mark_skb_lost(sk, skb);
2788 tcp_clear_retrans_hints_partial(tp);
2793 if (tcp_is_reno(tp))
2794 tcp_limit_reno_sacked(tp);
2796 tcp_verify_left_out(tp);
2798 /* Don't muck with the congestion window here.
2799 * Reason is that we do not increase amount of _data_
2800 * in network, but units changed and effective
2801 * cwnd/ssthresh really reduced now.
2803 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2804 tp->high_seq = tp->snd_nxt;
2805 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2806 tp->prior_ssthresh = 0;
2807 tp->undo_marker = 0;
2808 tcp_set_ca_state(sk, TCP_CA_Loss);
2810 tcp_xmit_retransmit_queue(sk);
2812 EXPORT_SYMBOL(tcp_simple_retransmit);
2814 void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2816 struct tcp_sock *tp = tcp_sk(sk);
2819 if (tcp_is_reno(tp))
2820 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2822 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2824 NET_INC_STATS(sock_net(sk), mib_idx);
2826 tp->prior_ssthresh = 0;
2829 if (!tcp_in_cwnd_reduction(sk)) {
2831 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2832 tcp_init_cwnd_reduction(sk);
2834 tcp_set_ca_state(sk, TCP_CA_Recovery);
2837 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2838 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2840 static void tcp_process_loss(struct sock *sk, int flag, int num_dupack,
2843 struct tcp_sock *tp = tcp_sk(sk);
2844 bool recovered = !before(tp->snd_una, tp->high_seq);
2846 if ((flag & FLAG_SND_UNA_ADVANCED || rcu_access_pointer(tp->fastopen_rsk)) &&
2847 tcp_try_undo_loss(sk, false))
2850 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2851 /* Step 3.b. A timeout is spurious if not all data are
2852 * lost, i.e., never-retransmitted data are (s)acked.
2854 if ((flag & FLAG_ORIG_SACK_ACKED) &&
2855 tcp_try_undo_loss(sk, true))
2858 if (after(tp->snd_nxt, tp->high_seq)) {
2859 if (flag & FLAG_DATA_SACKED || num_dupack)
2860 tp->frto = 0; /* Step 3.a. loss was real */
2861 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2862 tp->high_seq = tp->snd_nxt;
2863 /* Step 2.b. Try send new data (but deferred until cwnd
2864 * is updated in tcp_ack()). Otherwise fall back to
2865 * the conventional recovery.
2867 if (!tcp_write_queue_empty(sk) &&
2868 after(tcp_wnd_end(tp), tp->snd_nxt)) {
2869 *rexmit = REXMIT_NEW;
2877 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2878 tcp_try_undo_recovery(sk);
2881 if (tcp_is_reno(tp)) {
2882 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2883 * delivered. Lower inflight to clock out (re)tranmissions.
2885 if (after(tp->snd_nxt, tp->high_seq) && num_dupack)
2886 tcp_add_reno_sack(sk, num_dupack, flag & FLAG_ECE);
2887 else if (flag & FLAG_SND_UNA_ADVANCED)
2888 tcp_reset_reno_sack(tp);
2890 *rexmit = REXMIT_LOST;
2893 static bool tcp_force_fast_retransmit(struct sock *sk)
2895 struct tcp_sock *tp = tcp_sk(sk);
2897 return after(tcp_highest_sack_seq(tp),
2898 tp->snd_una + tp->reordering * tp->mss_cache);
2901 /* Undo during fast recovery after partial ACK. */
2902 static bool tcp_try_undo_partial(struct sock *sk, u32 prior_snd_una,
2905 struct tcp_sock *tp = tcp_sk(sk);
2907 if (tp->undo_marker && tcp_packet_delayed(tp)) {
2908 /* Plain luck! Hole if filled with delayed
2909 * packet, rather than with a retransmit. Check reordering.
2911 tcp_check_sack_reordering(sk, prior_snd_una, 1);
2913 /* We are getting evidence that the reordering degree is higher
2914 * than we realized. If there are no retransmits out then we
2915 * can undo. Otherwise we clock out new packets but do not
2916 * mark more packets lost or retransmit more.
2918 if (tp->retrans_out)
2921 if (!tcp_any_retrans_done(sk))
2922 tp->retrans_stamp = 0;
2924 DBGUNDO(sk, "partial recovery");
2925 tcp_undo_cwnd_reduction(sk, true);
2926 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2927 tcp_try_keep_open(sk);
2929 /* Partial ACK arrived. Force fast retransmit. */
2930 *do_lost = tcp_force_fast_retransmit(sk);
2935 static void tcp_identify_packet_loss(struct sock *sk, int *ack_flag)
2937 struct tcp_sock *tp = tcp_sk(sk);
2939 if (tcp_rtx_queue_empty(sk))
2942 if (unlikely(tcp_is_reno(tp))) {
2943 tcp_newreno_mark_lost(sk, *ack_flag & FLAG_SND_UNA_ADVANCED);
2944 } else if (tcp_is_rack(sk)) {
2945 u32 prior_retrans = tp->retrans_out;
2947 if (tcp_rack_mark_lost(sk))
2948 *ack_flag &= ~FLAG_SET_XMIT_TIMER;
2949 if (prior_retrans > tp->retrans_out)
2950 *ack_flag |= FLAG_LOST_RETRANS;
2954 /* Process an event, which can update packets-in-flight not trivially.
2955 * Main goal of this function is to calculate new estimate for left_out,
2956 * taking into account both packets sitting in receiver's buffer and
2957 * packets lost by network.
2959 * Besides that it updates the congestion state when packet loss or ECN
2960 * is detected. But it does not reduce the cwnd, it is done by the
2961 * congestion control later.
2963 * It does _not_ decide what to send, it is made in function
2964 * tcp_xmit_retransmit_queue().
2966 static void tcp_fastretrans_alert(struct sock *sk, const u32 prior_snd_una,
2967 int num_dupack, int *ack_flag, int *rexmit)
2969 struct inet_connection_sock *icsk = inet_csk(sk);
2970 struct tcp_sock *tp = tcp_sk(sk);
2971 int fast_rexmit = 0, flag = *ack_flag;
2972 bool ece_ack = flag & FLAG_ECE;
2973 bool do_lost = num_dupack || ((flag & FLAG_DATA_SACKED) &&
2974 tcp_force_fast_retransmit(sk));
2976 if (!tp->packets_out && tp->sacked_out)
2979 /* Now state machine starts.
2980 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2982 tp->prior_ssthresh = 0;
2984 /* B. In all the states check for reneging SACKs. */
2985 if (tcp_check_sack_reneging(sk, ack_flag))
2988 /* C. Check consistency of the current state. */
2989 tcp_verify_left_out(tp);
2991 /* D. Check state exit conditions. State can be terminated
2992 * when high_seq is ACKed. */
2993 if (icsk->icsk_ca_state == TCP_CA_Open) {
2994 WARN_ON(tp->retrans_out != 0 && !tp->syn_data);
2995 tp->retrans_stamp = 0;
2996 } else if (!before(tp->snd_una, tp->high_seq)) {
2997 switch (icsk->icsk_ca_state) {
2999 /* CWR is to be held something *above* high_seq
3000 * is ACKed for CWR bit to reach receiver. */
3001 if (tp->snd_una != tp->high_seq) {
3002 tcp_end_cwnd_reduction(sk);
3003 tcp_set_ca_state(sk, TCP_CA_Open);
3007 case TCP_CA_Recovery:
3008 if (tcp_is_reno(tp))
3009 tcp_reset_reno_sack(tp);
3010 if (tcp_try_undo_recovery(sk))
3012 tcp_end_cwnd_reduction(sk);
3017 /* E. Process state. */
3018 switch (icsk->icsk_ca_state) {
3019 case TCP_CA_Recovery:
3020 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
3021 if (tcp_is_reno(tp))
3022 tcp_add_reno_sack(sk, num_dupack, ece_ack);
3023 } else if (tcp_try_undo_partial(sk, prior_snd_una, &do_lost))
3026 if (tcp_try_undo_dsack(sk))
3027 tcp_try_keep_open(sk);
3029 tcp_identify_packet_loss(sk, ack_flag);
3030 if (icsk->icsk_ca_state != TCP_CA_Recovery) {
3031 if (!tcp_time_to_recover(sk, flag))
3033 /* Undo reverts the recovery state. If loss is evident,
3034 * starts a new recovery (e.g. reordering then loss);
3036 tcp_enter_recovery(sk, ece_ack);
3040 tcp_process_loss(sk, flag, num_dupack, rexmit);
3041 tcp_identify_packet_loss(sk, ack_flag);
3042 if (!(icsk->icsk_ca_state == TCP_CA_Open ||
3043 (*ack_flag & FLAG_LOST_RETRANS)))
3045 /* Change state if cwnd is undone or retransmits are lost */
3048 if (tcp_is_reno(tp)) {
3049 if (flag & FLAG_SND_UNA_ADVANCED)
3050 tcp_reset_reno_sack(tp);
3051 tcp_add_reno_sack(sk, num_dupack, ece_ack);
3054 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
3055 tcp_try_undo_dsack(sk);
3057 tcp_identify_packet_loss(sk, ack_flag);
3058 if (!tcp_time_to_recover(sk, flag)) {
3059 tcp_try_to_open(sk, flag);
3063 /* MTU probe failure: don't reduce cwnd */
3064 if (icsk->icsk_ca_state < TCP_CA_CWR &&
3065 icsk->icsk_mtup.probe_size &&
3066 tp->snd_una == tp->mtu_probe.probe_seq_start) {
3067 tcp_mtup_probe_failed(sk);
3068 /* Restores the reduction we did in tcp_mtup_probe() */
3069 tcp_snd_cwnd_set(tp, tcp_snd_cwnd(tp) + 1);
3070 tcp_simple_retransmit(sk);
3074 /* Otherwise enter Recovery state */
3075 tcp_enter_recovery(sk, ece_ack);
3079 if (!tcp_is_rack(sk) && do_lost)
3080 tcp_update_scoreboard(sk, fast_rexmit);
3081 *rexmit = REXMIT_LOST;
3084 static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us, const int flag)
3086 u32 wlen = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_rtt_wlen) * HZ;
3087 struct tcp_sock *tp = tcp_sk(sk);
3089 if ((flag & FLAG_ACK_MAYBE_DELAYED) && rtt_us > tcp_min_rtt(tp)) {
3090 /* If the remote keeps returning delayed ACKs, eventually
3091 * the min filter would pick it up and overestimate the
3092 * prop. delay when it expires. Skip suspected delayed ACKs.
3096 minmax_running_min(&tp->rtt_min, wlen, tcp_jiffies32,
3097 rtt_us ? : jiffies_to_usecs(1));
3100 static bool tcp_ack_update_rtt(struct sock *sk, const int flag,
3101 long seq_rtt_us, long sack_rtt_us,
3102 long ca_rtt_us, struct rate_sample *rs)
3104 const struct tcp_sock *tp = tcp_sk(sk);
3106 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
3107 * broken middle-boxes or peers may corrupt TS-ECR fields. But
3108 * Karn's algorithm forbids taking RTT if some retransmitted data
3109 * is acked (RFC6298).
3112 seq_rtt_us = sack_rtt_us;
3114 /* RTTM Rule: A TSecr value received in a segment is used to
3115 * update the averaged RTT measurement only if the segment
3116 * acknowledges some new data, i.e., only if it advances the
3117 * left edge of the send window.
3118 * See draft-ietf-tcplw-high-performance-00, section 3.3.
3120 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
3121 flag & FLAG_ACKED) {
3122 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr;
3124 if (likely(delta < INT_MAX / (USEC_PER_SEC / TCP_TS_HZ))) {
3127 seq_rtt_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
3128 ca_rtt_us = seq_rtt_us;
3131 rs->rtt_us = ca_rtt_us; /* RTT of last (S)ACKed packet (or -1) */
3135 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
3136 * always taken together with ACK, SACK, or TS-opts. Any negative
3137 * values will be skipped with the seq_rtt_us < 0 check above.
3139 tcp_update_rtt_min(sk, ca_rtt_us, flag);
3140 tcp_rtt_estimator(sk, seq_rtt_us);
3143 /* RFC6298: only reset backoff on valid RTT measurement. */
3144 inet_csk(sk)->icsk_backoff = 0;
3148 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
3149 void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req)
3151 struct rate_sample rs;
3154 if (req && !req->num_retrans && tcp_rsk(req)->snt_synack)
3155 rtt_us = tcp_stamp_us_delta(tcp_clock_us(), tcp_rsk(req)->snt_synack);
3157 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us, &rs);
3161 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
3163 const struct inet_connection_sock *icsk = inet_csk(sk);
3165 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
3166 tcp_sk(sk)->snd_cwnd_stamp = tcp_jiffies32;
3169 /* Restart timer after forward progress on connection.
3170 * RFC2988 recommends to restart timer to now+rto.
3172 void tcp_rearm_rto(struct sock *sk)
3174 const struct inet_connection_sock *icsk = inet_csk(sk);
3175 struct tcp_sock *tp = tcp_sk(sk);
3177 /* If the retrans timer is currently being used by Fast Open
3178 * for SYN-ACK retrans purpose, stay put.
3180 if (rcu_access_pointer(tp->fastopen_rsk))
3183 if (!tp->packets_out) {
3184 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
3186 u32 rto = inet_csk(sk)->icsk_rto;
3187 /* Offset the time elapsed after installing regular RTO */
3188 if (icsk->icsk_pending == ICSK_TIME_REO_TIMEOUT ||
3189 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
3190 s64 delta_us = tcp_rto_delta_us(sk);
3191 /* delta_us may not be positive if the socket is locked
3192 * when the retrans timer fires and is rescheduled.
3194 rto = usecs_to_jiffies(max_t(int, delta_us, 1));
3196 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
3201 /* Try to schedule a loss probe; if that doesn't work, then schedule an RTO. */
3202 static void tcp_set_xmit_timer(struct sock *sk)
3204 if (!tcp_schedule_loss_probe(sk, true))
3208 /* If we get here, the whole TSO packet has not been acked. */
3209 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3211 struct tcp_sock *tp = tcp_sk(sk);
3214 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3216 packets_acked = tcp_skb_pcount(skb);
3217 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3219 packets_acked -= tcp_skb_pcount(skb);
3221 if (packets_acked) {
3222 BUG_ON(tcp_skb_pcount(skb) == 0);
3223 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3226 return packets_acked;
3229 static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
3230 const struct sk_buff *ack_skb, u32 prior_snd_una)
3232 const struct skb_shared_info *shinfo;
3234 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3235 if (likely(!TCP_SKB_CB(skb)->txstamp_ack))
3238 shinfo = skb_shinfo(skb);
3239 if (!before(shinfo->tskey, prior_snd_una) &&
3240 before(shinfo->tskey, tcp_sk(sk)->snd_una)) {
3241 tcp_skb_tsorted_save(skb) {
3242 __skb_tstamp_tx(skb, ack_skb, NULL, sk, SCM_TSTAMP_ACK);
3243 } tcp_skb_tsorted_restore(skb);
3247 /* Remove acknowledged frames from the retransmission queue. If our packet
3248 * is before the ack sequence we can discard it as it's confirmed to have
3249 * arrived at the other end.
3251 static int tcp_clean_rtx_queue(struct sock *sk, const struct sk_buff *ack_skb,
3252 u32 prior_fack, u32 prior_snd_una,
3253 struct tcp_sacktag_state *sack, bool ece_ack)
3255 const struct inet_connection_sock *icsk = inet_csk(sk);
3256 u64 first_ackt, last_ackt;
3257 struct tcp_sock *tp = tcp_sk(sk);
3258 u32 prior_sacked = tp->sacked_out;
3259 u32 reord = tp->snd_nxt; /* lowest acked un-retx un-sacked seq */
3260 struct sk_buff *skb, *next;
3261 bool fully_acked = true;
3262 long sack_rtt_us = -1L;
3263 long seq_rtt_us = -1L;
3264 long ca_rtt_us = -1L;
3271 for (skb = skb_rb_first(&sk->tcp_rtx_queue); skb; skb = next) {
3272 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3273 const u32 start_seq = scb->seq;
3274 u8 sacked = scb->sacked;
3277 /* Determine how many packets and what bytes were acked, tso and else */
3278 if (after(scb->end_seq, tp->snd_una)) {
3279 if (tcp_skb_pcount(skb) == 1 ||
3280 !after(tp->snd_una, scb->seq))
3283 acked_pcount = tcp_tso_acked(sk, skb);
3286 fully_acked = false;
3288 acked_pcount = tcp_skb_pcount(skb);
3291 if (unlikely(sacked & TCPCB_RETRANS)) {
3292 if (sacked & TCPCB_SACKED_RETRANS)
3293 tp->retrans_out -= acked_pcount;
3294 flag |= FLAG_RETRANS_DATA_ACKED;
3295 } else if (!(sacked & TCPCB_SACKED_ACKED)) {
3296 last_ackt = tcp_skb_timestamp_us(skb);
3297 WARN_ON_ONCE(last_ackt == 0);
3299 first_ackt = last_ackt;
3301 if (before(start_seq, reord))
3303 if (!after(scb->end_seq, tp->high_seq))
3304 flag |= FLAG_ORIG_SACK_ACKED;
3307 if (sacked & TCPCB_SACKED_ACKED) {
3308 tp->sacked_out -= acked_pcount;
3309 } else if (tcp_is_sack(tp)) {
3310 tcp_count_delivered(tp, acked_pcount, ece_ack);
3311 if (!tcp_skb_spurious_retrans(tp, skb))
3312 tcp_rack_advance(tp, sacked, scb->end_seq,
3313 tcp_skb_timestamp_us(skb));
3315 if (sacked & TCPCB_LOST)
3316 tp->lost_out -= acked_pcount;
3318 tp->packets_out -= acked_pcount;
3319 pkts_acked += acked_pcount;
3320 tcp_rate_skb_delivered(sk, skb, sack->rate);
3322 /* Initial outgoing SYN's get put onto the write_queue
3323 * just like anything else we transmit. It is not
3324 * true data, and if we misinform our callers that
3325 * this ACK acks real data, we will erroneously exit
3326 * connection startup slow start one packet too
3327 * quickly. This is severely frowned upon behavior.
3329 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
3330 flag |= FLAG_DATA_ACKED;
3332 flag |= FLAG_SYN_ACKED;
3333 tp->retrans_stamp = 0;
3339 tcp_ack_tstamp(sk, skb, ack_skb, prior_snd_una);
3341 next = skb_rb_next(skb);
3342 if (unlikely(skb == tp->retransmit_skb_hint))
3343 tp->retransmit_skb_hint = NULL;
3344 if (unlikely(skb == tp->lost_skb_hint))
3345 tp->lost_skb_hint = NULL;
3346 tcp_highest_sack_replace(sk, skb, next);
3347 tcp_rtx_queue_unlink_and_free(skb, sk);
3351 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
3353 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3354 tp->snd_up = tp->snd_una;
3357 tcp_ack_tstamp(sk, skb, ack_skb, prior_snd_una);
3358 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
3359 flag |= FLAG_SACK_RENEGING;
3362 if (likely(first_ackt) && !(flag & FLAG_RETRANS_DATA_ACKED)) {
3363 seq_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, first_ackt);
3364 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, last_ackt);
3366 if (pkts_acked == 1 && fully_acked && !prior_sacked &&
3367 (tp->snd_una - prior_snd_una) < tp->mss_cache &&
3368 sack->rate->prior_delivered + 1 == tp->delivered &&
3369 !(flag & (FLAG_CA_ALERT | FLAG_SYN_ACKED))) {
3370 /* Conservatively mark a delayed ACK. It's typically
3371 * from a lone runt packet over the round trip to
3372 * a receiver w/o out-of-order or CE events.
3374 flag |= FLAG_ACK_MAYBE_DELAYED;
3377 if (sack->first_sackt) {
3378 sack_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->first_sackt);
3379 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->last_sackt);
3381 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us,
3382 ca_rtt_us, sack->rate);
3384 if (flag & FLAG_ACKED) {
3385 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3386 if (unlikely(icsk->icsk_mtup.probe_size &&
3387 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3388 tcp_mtup_probe_success(sk);
3391 if (tcp_is_reno(tp)) {
3392 tcp_remove_reno_sacks(sk, pkts_acked, ece_ack);
3394 /* If any of the cumulatively ACKed segments was
3395 * retransmitted, non-SACK case cannot confirm that
3396 * progress was due to original transmission due to
3397 * lack of TCPCB_SACKED_ACKED bits even if some of
3398 * the packets may have been never retransmitted.
3400 if (flag & FLAG_RETRANS_DATA_ACKED)
3401 flag &= ~FLAG_ORIG_SACK_ACKED;
3405 /* Non-retransmitted hole got filled? That's reordering */
3406 if (before(reord, prior_fack))
3407 tcp_check_sack_reordering(sk, reord, 0);
3409 delta = prior_sacked - tp->sacked_out;
3410 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3412 } else if (skb && rtt_update && sack_rtt_us >= 0 &&
3413 sack_rtt_us > tcp_stamp_us_delta(tp->tcp_mstamp,
3414 tcp_skb_timestamp_us(skb))) {
3415 /* Do not re-arm RTO if the sack RTT is measured from data sent
3416 * after when the head was last (re)transmitted. Otherwise the
3417 * timeout may continue to extend in loss recovery.
3419 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3422 if (icsk->icsk_ca_ops->pkts_acked) {
3423 struct ack_sample sample = { .pkts_acked = pkts_acked,
3424 .rtt_us = sack->rate->rtt_us };
3426 sample.in_flight = tp->mss_cache *
3427 (tp->delivered - sack->rate->prior_delivered);
3428 icsk->icsk_ca_ops->pkts_acked(sk, &sample);
3431 #if FASTRETRANS_DEBUG > 0
3432 WARN_ON((int)tp->sacked_out < 0);
3433 WARN_ON((int)tp->lost_out < 0);
3434 WARN_ON((int)tp->retrans_out < 0);
3435 if (!tp->packets_out && tcp_is_sack(tp)) {
3436 icsk = inet_csk(sk);
3438 pr_debug("Leak l=%u %d\n",
3439 tp->lost_out, icsk->icsk_ca_state);
3442 if (tp->sacked_out) {
3443 pr_debug("Leak s=%u %d\n",
3444 tp->sacked_out, icsk->icsk_ca_state);
3447 if (tp->retrans_out) {
3448 pr_debug("Leak r=%u %d\n",
3449 tp->retrans_out, icsk->icsk_ca_state);
3450 tp->retrans_out = 0;
3457 static void tcp_ack_probe(struct sock *sk)
3459 struct inet_connection_sock *icsk = inet_csk(sk);
3460 struct sk_buff *head = tcp_send_head(sk);
3461 const struct tcp_sock *tp = tcp_sk(sk);
3463 /* Was it a usable window open? */
3466 if (!after(TCP_SKB_CB(head)->end_seq, tcp_wnd_end(tp))) {
3467 icsk->icsk_backoff = 0;
3468 icsk->icsk_probes_tstamp = 0;
3469 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3470 /* Socket must be waked up by subsequent tcp_data_snd_check().
3471 * This function is not for random using!
3474 unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX);
3476 when = tcp_clamp_probe0_to_user_timeout(sk, when);
3477 tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, when, TCP_RTO_MAX);
3481 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3483 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3484 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3487 /* Decide wheather to run the increase function of congestion control. */
3488 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3490 /* If reordering is high then always grow cwnd whenever data is
3491 * delivered regardless of its ordering. Otherwise stay conservative
3492 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3493 * new SACK or ECE mark may first advance cwnd here and later reduce
3494 * cwnd in tcp_fastretrans_alert() based on more states.
3496 if (tcp_sk(sk)->reordering >
3497 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_reordering))
3498 return flag & FLAG_FORWARD_PROGRESS;
3500 return flag & FLAG_DATA_ACKED;
3503 /* The "ultimate" congestion control function that aims to replace the rigid
3504 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction).
3505 * It's called toward the end of processing an ACK with precise rate
3506 * information. All transmission or retransmission are delayed afterwards.
3508 static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked,
3509 int flag, const struct rate_sample *rs)
3511 const struct inet_connection_sock *icsk = inet_csk(sk);
3513 if (icsk->icsk_ca_ops->cong_control) {
3514 icsk->icsk_ca_ops->cong_control(sk, rs);
3518 if (tcp_in_cwnd_reduction(sk)) {
3519 /* Reduce cwnd if state mandates */
3520 tcp_cwnd_reduction(sk, acked_sacked, rs->losses, flag);
3521 } else if (tcp_may_raise_cwnd(sk, flag)) {
3522 /* Advance cwnd if state allows */
3523 tcp_cong_avoid(sk, ack, acked_sacked);
3525 tcp_update_pacing_rate(sk);
3528 /* Check that window update is acceptable.
3529 * The function assumes that snd_una<=ack<=snd_next.
3531 static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3532 const u32 ack, const u32 ack_seq,
3535 return after(ack, tp->snd_una) ||
3536 after(ack_seq, tp->snd_wl1) ||
3537 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3540 /* If we update tp->snd_una, also update tp->bytes_acked */
3541 static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack)
3543 u32 delta = ack - tp->snd_una;
3545 sock_owned_by_me((struct sock *)tp);
3546 tp->bytes_acked += delta;
3550 /* If we update tp->rcv_nxt, also update tp->bytes_received */
3551 static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq)
3553 u32 delta = seq - tp->rcv_nxt;
3555 sock_owned_by_me((struct sock *)tp);
3556 tp->bytes_received += delta;
3557 WRITE_ONCE(tp->rcv_nxt, seq);
3560 /* Update our send window.
3562 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3563 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3565 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3568 struct tcp_sock *tp = tcp_sk(sk);
3570 u32 nwin = ntohs(tcp_hdr(skb)->window);
3572 if (likely(!tcp_hdr(skb)->syn))
3573 nwin <<= tp->rx_opt.snd_wscale;
3575 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3576 flag |= FLAG_WIN_UPDATE;
3577 tcp_update_wl(tp, ack_seq);
3579 if (tp->snd_wnd != nwin) {
3582 /* Note, it is the only place, where
3583 * fast path is recovered for sending TCP.
3586 tcp_fast_path_check(sk);
3588 if (!tcp_write_queue_empty(sk))
3589 tcp_slow_start_after_idle_check(sk);
3591 if (nwin > tp->max_window) {
3592 tp->max_window = nwin;
3593 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3598 tcp_snd_una_update(tp, ack);
3603 static bool __tcp_oow_rate_limited(struct net *net, int mib_idx,
3604 u32 *last_oow_ack_time)
3606 /* Paired with the WRITE_ONCE() in this function. */
3607 u32 val = READ_ONCE(*last_oow_ack_time);
3610 s32 elapsed = (s32)(tcp_jiffies32 - val);
3613 elapsed < READ_ONCE(net->ipv4.sysctl_tcp_invalid_ratelimit)) {
3614 NET_INC_STATS(net, mib_idx);
3615 return true; /* rate-limited: don't send yet! */
3619 /* Paired with the prior READ_ONCE() and with itself,
3620 * as we might be lockless.
3622 WRITE_ONCE(*last_oow_ack_time, tcp_jiffies32);
3624 return false; /* not rate-limited: go ahead, send dupack now! */
3627 /* Return true if we're currently rate-limiting out-of-window ACKs and
3628 * thus shouldn't send a dupack right now. We rate-limit dupacks in
3629 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3630 * attacks that send repeated SYNs or ACKs for the same connection. To
3631 * do this, we do not send a duplicate SYNACK or ACK if the remote
3632 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3634 bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb,
3635 int mib_idx, u32 *last_oow_ack_time)
3637 /* Data packets without SYNs are not likely part of an ACK loop. */
3638 if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) &&
3642 return __tcp_oow_rate_limited(net, mib_idx, last_oow_ack_time);
3645 /* RFC 5961 7 [ACK Throttling] */
3646 static void tcp_send_challenge_ack(struct sock *sk)
3648 struct tcp_sock *tp = tcp_sk(sk);
3649 struct net *net = sock_net(sk);
3650 u32 count, now, ack_limit;
3652 /* First check our per-socket dupack rate limit. */
3653 if (__tcp_oow_rate_limited(net,
3654 LINUX_MIB_TCPACKSKIPPEDCHALLENGE,
3655 &tp->last_oow_ack_time))
3658 ack_limit = READ_ONCE(net->ipv4.sysctl_tcp_challenge_ack_limit);
3659 if (ack_limit == INT_MAX)
3662 /* Then check host-wide RFC 5961 rate limit. */
3664 if (now != READ_ONCE(net->ipv4.tcp_challenge_timestamp)) {
3665 u32 half = (ack_limit + 1) >> 1;
3667 WRITE_ONCE(net->ipv4.tcp_challenge_timestamp, now);
3668 WRITE_ONCE(net->ipv4.tcp_challenge_count, half + prandom_u32_max(ack_limit));
3670 count = READ_ONCE(net->ipv4.tcp_challenge_count);
3672 WRITE_ONCE(net->ipv4.tcp_challenge_count, count - 1);
3674 NET_INC_STATS(net, LINUX_MIB_TCPCHALLENGEACK);
3679 static void tcp_store_ts_recent(struct tcp_sock *tp)
3681 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3682 tp->rx_opt.ts_recent_stamp = ktime_get_seconds();
3685 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3687 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3688 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3689 * extra check below makes sure this can only happen
3690 * for pure ACK frames. -DaveM
3692 * Not only, also it occurs for expired timestamps.
3695 if (tcp_paws_check(&tp->rx_opt, 0))
3696 tcp_store_ts_recent(tp);
3700 /* This routine deals with acks during a TLP episode and ends an episode by
3701 * resetting tlp_high_seq. Ref: TLP algorithm in draft-ietf-tcpm-rack
3703 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3705 struct tcp_sock *tp = tcp_sk(sk);
3707 if (before(ack, tp->tlp_high_seq))
3710 if (!tp->tlp_retrans) {
3711 /* TLP of new data has been acknowledged */
3712 tp->tlp_high_seq = 0;
3713 } else if (flag & FLAG_DSACK_TLP) {
3714 /* This DSACK means original and TLP probe arrived; no loss */
3715 tp->tlp_high_seq = 0;
3716 } else if (after(ack, tp->tlp_high_seq)) {
3717 /* ACK advances: there was a loss, so reduce cwnd. Reset
3718 * tlp_high_seq in tcp_init_cwnd_reduction()
3720 tcp_init_cwnd_reduction(sk);
3721 tcp_set_ca_state(sk, TCP_CA_CWR);
3722 tcp_end_cwnd_reduction(sk);
3723 tcp_try_keep_open(sk);
3724 NET_INC_STATS(sock_net(sk),
3725 LINUX_MIB_TCPLOSSPROBERECOVERY);
3726 } else if (!(flag & (FLAG_SND_UNA_ADVANCED |
3727 FLAG_NOT_DUP | FLAG_DATA_SACKED))) {
3728 /* Pure dupack: original and TLP probe arrived; no loss */
3729 tp->tlp_high_seq = 0;
3733 static inline void tcp_in_ack_event(struct sock *sk, u32 flags)
3735 const struct inet_connection_sock *icsk = inet_csk(sk);
3737 if (icsk->icsk_ca_ops->in_ack_event)
3738 icsk->icsk_ca_ops->in_ack_event(sk, flags);
3741 /* Congestion control has updated the cwnd already. So if we're in
3742 * loss recovery then now we do any new sends (for FRTO) or
3743 * retransmits (for CA_Loss or CA_recovery) that make sense.
3745 static void tcp_xmit_recovery(struct sock *sk, int rexmit)
3747 struct tcp_sock *tp = tcp_sk(sk);
3749 if (rexmit == REXMIT_NONE || sk->sk_state == TCP_SYN_SENT)
3752 if (unlikely(rexmit == REXMIT_NEW)) {
3753 __tcp_push_pending_frames(sk, tcp_current_mss(sk),
3755 if (after(tp->snd_nxt, tp->high_seq))
3759 tcp_xmit_retransmit_queue(sk);
3762 /* Returns the number of packets newly acked or sacked by the current ACK */
3763 static u32 tcp_newly_delivered(struct sock *sk, u32 prior_delivered, int flag)
3765 const struct net *net = sock_net(sk);
3766 struct tcp_sock *tp = tcp_sk(sk);
3769 delivered = tp->delivered - prior_delivered;
3770 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVERED, delivered);
3771 if (flag & FLAG_ECE)
3772 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVEREDCE, delivered);
3777 /* This routine deals with incoming acks, but not outgoing ones. */
3778 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3780 struct inet_connection_sock *icsk = inet_csk(sk);
3781 struct tcp_sock *tp = tcp_sk(sk);
3782 struct tcp_sacktag_state sack_state;
3783 struct rate_sample rs = { .prior_delivered = 0 };
3784 u32 prior_snd_una = tp->snd_una;
3785 bool is_sack_reneg = tp->is_sack_reneg;
3786 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3787 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3789 int prior_packets = tp->packets_out;
3790 u32 delivered = tp->delivered;
3791 u32 lost = tp->lost;
3792 int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */
3795 sack_state.first_sackt = 0;
3796 sack_state.rate = &rs;
3797 sack_state.sack_delivered = 0;
3799 /* We very likely will need to access rtx queue. */
3800 prefetch(sk->tcp_rtx_queue.rb_node);
3802 /* If the ack is older than previous acks
3803 * then we can probably ignore it.
3805 if (before(ack, prior_snd_una)) {
3808 /* do not accept ACK for bytes we never sent. */
3809 max_window = min_t(u64, tp->max_window, tp->bytes_acked);
3810 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3811 if (before(ack, prior_snd_una - max_window)) {
3812 if (!(flag & FLAG_NO_CHALLENGE_ACK))
3813 tcp_send_challenge_ack(sk);
3814 return -SKB_DROP_REASON_TCP_TOO_OLD_ACK;
3819 /* If the ack includes data we haven't sent yet, discard
3820 * this segment (RFC793 Section 3.9).
3822 if (after(ack, tp->snd_nxt))
3823 return -SKB_DROP_REASON_TCP_ACK_UNSENT_DATA;
3825 if (after(ack, prior_snd_una)) {
3826 flag |= FLAG_SND_UNA_ADVANCED;
3827 icsk->icsk_retransmits = 0;
3829 #if IS_ENABLED(CONFIG_TLS_DEVICE)
3830 if (static_branch_unlikely(&clean_acked_data_enabled.key))
3831 if (icsk->icsk_clean_acked)
3832 icsk->icsk_clean_acked(sk, ack);
3836 prior_fack = tcp_is_sack(tp) ? tcp_highest_sack_seq(tp) : tp->snd_una;
3837 rs.prior_in_flight = tcp_packets_in_flight(tp);
3839 /* ts_recent update must be made after we are sure that the packet
3842 if (flag & FLAG_UPDATE_TS_RECENT)
3843 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3845 if ((flag & (FLAG_SLOWPATH | FLAG_SND_UNA_ADVANCED)) ==
3846 FLAG_SND_UNA_ADVANCED) {
3847 /* Window is constant, pure forward advance.
3848 * No more checks are required.
3849 * Note, we use the fact that SND.UNA>=SND.WL2.
3851 tcp_update_wl(tp, ack_seq);
3852 tcp_snd_una_update(tp, ack);
3853 flag |= FLAG_WIN_UPDATE;
3855 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
3857 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPACKS);
3859 u32 ack_ev_flags = CA_ACK_SLOWPATH;
3861 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3864 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3866 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3868 if (TCP_SKB_CB(skb)->sacked)
3869 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3872 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) {
3874 ack_ev_flags |= CA_ACK_ECE;
3877 if (sack_state.sack_delivered)
3878 tcp_count_delivered(tp, sack_state.sack_delivered,
3881 if (flag & FLAG_WIN_UPDATE)
3882 ack_ev_flags |= CA_ACK_WIN_UPDATE;
3884 tcp_in_ack_event(sk, ack_ev_flags);
3887 /* This is a deviation from RFC3168 since it states that:
3888 * "When the TCP data sender is ready to set the CWR bit after reducing
3889 * the congestion window, it SHOULD set the CWR bit only on the first
3890 * new data packet that it transmits."
3891 * We accept CWR on pure ACKs to be more robust
3892 * with widely-deployed TCP implementations that do this.
3894 tcp_ecn_accept_cwr(sk, skb);
3896 /* We passed data and got it acked, remove any soft error
3897 * log. Something worked...
3899 sk->sk_err_soft = 0;
3900 icsk->icsk_probes_out = 0;
3901 tp->rcv_tstamp = tcp_jiffies32;
3905 /* See if we can take anything off of the retransmit queue. */
3906 flag |= tcp_clean_rtx_queue(sk, skb, prior_fack, prior_snd_una,
3907 &sack_state, flag & FLAG_ECE);
3909 tcp_rack_update_reo_wnd(sk, &rs);
3911 if (tp->tlp_high_seq)
3912 tcp_process_tlp_ack(sk, ack, flag);
3914 if (tcp_ack_is_dubious(sk, flag)) {
3915 if (!(flag & (FLAG_SND_UNA_ADVANCED |
3916 FLAG_NOT_DUP | FLAG_DSACKING_ACK))) {
3918 /* Consider if pure acks were aggregated in tcp_add_backlog() */
3919 if (!(flag & FLAG_DATA))
3920 num_dupack = max_t(u16, 1, skb_shinfo(skb)->gso_segs);
3922 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
3926 /* If needed, reset TLP/RTO timer when RACK doesn't set. */
3927 if (flag & FLAG_SET_XMIT_TIMER)
3928 tcp_set_xmit_timer(sk);
3930 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
3933 delivered = tcp_newly_delivered(sk, delivered, flag);
3934 lost = tp->lost - lost; /* freshly marked lost */
3935 rs.is_ack_delayed = !!(flag & FLAG_ACK_MAYBE_DELAYED);
3936 tcp_rate_gen(sk, delivered, lost, is_sack_reneg, sack_state.rate);
3937 tcp_cong_control(sk, ack, delivered, flag, sack_state.rate);
3938 tcp_xmit_recovery(sk, rexmit);
3942 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3943 if (flag & FLAG_DSACKING_ACK) {
3944 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
3946 tcp_newly_delivered(sk, delivered, flag);
3948 /* If this ack opens up a zero window, clear backoff. It was
3949 * being used to time the probes, and is probably far higher than
3950 * it needs to be for normal retransmission.
3954 if (tp->tlp_high_seq)
3955 tcp_process_tlp_ack(sk, ack, flag);
3959 /* If data was SACKed, tag it and see if we should send more data.
3960 * If data was DSACKed, see if we can undo a cwnd reduction.
3962 if (TCP_SKB_CB(skb)->sacked) {
3963 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3965 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
3967 tcp_newly_delivered(sk, delivered, flag);
3968 tcp_xmit_recovery(sk, rexmit);
3974 static void tcp_parse_fastopen_option(int len, const unsigned char *cookie,
3975 bool syn, struct tcp_fastopen_cookie *foc,
3978 /* Valid only in SYN or SYN-ACK with an even length. */
3979 if (!foc || !syn || len < 0 || (len & 1))
3982 if (len >= TCP_FASTOPEN_COOKIE_MIN &&
3983 len <= TCP_FASTOPEN_COOKIE_MAX)
3984 memcpy(foc->val, cookie, len);
3991 static bool smc_parse_options(const struct tcphdr *th,
3992 struct tcp_options_received *opt_rx,
3993 const unsigned char *ptr,
3996 #if IS_ENABLED(CONFIG_SMC)
3997 if (static_branch_unlikely(&tcp_have_smc)) {
3998 if (th->syn && !(opsize & 1) &&
3999 opsize >= TCPOLEN_EXP_SMC_BASE &&
4000 get_unaligned_be32(ptr) == TCPOPT_SMC_MAGIC) {
4009 /* Try to parse the MSS option from the TCP header. Return 0 on failure, clamped
4012 u16 tcp_parse_mss_option(const struct tcphdr *th, u16 user_mss)
4014 const unsigned char *ptr = (const unsigned char *)(th + 1);
4015 int length = (th->doff * 4) - sizeof(struct tcphdr);
4018 while (length > 0) {
4019 int opcode = *ptr++;
4025 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
4032 if (opsize < 2) /* "silly options" */
4034 if (opsize > length)
4035 return mss; /* fail on partial options */
4036 if (opcode == TCPOPT_MSS && opsize == TCPOLEN_MSS) {
4037 u16 in_mss = get_unaligned_be16(ptr);
4040 if (user_mss && user_mss < in_mss)
4051 EXPORT_SYMBOL_GPL(tcp_parse_mss_option);
4053 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
4054 * But, this can also be called on packets in the established flow when
4055 * the fast version below fails.
4057 void tcp_parse_options(const struct net *net,
4058 const struct sk_buff *skb,
4059 struct tcp_options_received *opt_rx, int estab,
4060 struct tcp_fastopen_cookie *foc)
4062 const unsigned char *ptr;
4063 const struct tcphdr *th = tcp_hdr(skb);
4064 int length = (th->doff * 4) - sizeof(struct tcphdr);
4066 ptr = (const unsigned char *)(th + 1);
4067 opt_rx->saw_tstamp = 0;
4068 opt_rx->saw_unknown = 0;
4070 while (length > 0) {
4071 int opcode = *ptr++;
4077 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
4084 if (opsize < 2) /* "silly options" */
4086 if (opsize > length)
4087 return; /* don't parse partial options */
4090 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
4091 u16 in_mss = get_unaligned_be16(ptr);
4093 if (opt_rx->user_mss &&
4094 opt_rx->user_mss < in_mss)
4095 in_mss = opt_rx->user_mss;
4096 opt_rx->mss_clamp = in_mss;
4101 if (opsize == TCPOLEN_WINDOW && th->syn &&
4102 !estab && READ_ONCE(net->ipv4.sysctl_tcp_window_scaling)) {
4103 __u8 snd_wscale = *(__u8 *)ptr;
4104 opt_rx->wscale_ok = 1;
4105 if (snd_wscale > TCP_MAX_WSCALE) {
4106 net_info_ratelimited("%s: Illegal window scaling value %d > %u received\n",
4110 snd_wscale = TCP_MAX_WSCALE;
4112 opt_rx->snd_wscale = snd_wscale;
4115 case TCPOPT_TIMESTAMP:
4116 if ((opsize == TCPOLEN_TIMESTAMP) &&
4117 ((estab && opt_rx->tstamp_ok) ||
4118 (!estab && READ_ONCE(net->ipv4.sysctl_tcp_timestamps)))) {
4119 opt_rx->saw_tstamp = 1;
4120 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
4121 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
4124 case TCPOPT_SACK_PERM:
4125 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
4126 !estab && READ_ONCE(net->ipv4.sysctl_tcp_sack)) {
4127 opt_rx->sack_ok = TCP_SACK_SEEN;
4128 tcp_sack_reset(opt_rx);
4133 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
4134 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
4136 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
4139 #ifdef CONFIG_TCP_MD5SIG
4142 * The MD5 Hash has already been
4143 * checked (see tcp_v{4,6}_do_rcv()).
4147 case TCPOPT_FASTOPEN:
4148 tcp_parse_fastopen_option(
4149 opsize - TCPOLEN_FASTOPEN_BASE,
4150 ptr, th->syn, foc, false);
4154 /* Fast Open option shares code 254 using a
4155 * 16 bits magic number.
4157 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE &&
4158 get_unaligned_be16(ptr) ==
4159 TCPOPT_FASTOPEN_MAGIC) {
4160 tcp_parse_fastopen_option(opsize -
4161 TCPOLEN_EXP_FASTOPEN_BASE,
4162 ptr + 2, th->syn, foc, true);
4166 if (smc_parse_options(th, opt_rx, ptr, opsize))
4169 opt_rx->saw_unknown = 1;
4173 opt_rx->saw_unknown = 1;
4180 EXPORT_SYMBOL(tcp_parse_options);
4182 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
4184 const __be32 *ptr = (const __be32 *)(th + 1);
4186 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
4187 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
4188 tp->rx_opt.saw_tstamp = 1;
4190 tp->rx_opt.rcv_tsval = ntohl(*ptr);
4193 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
4195 tp->rx_opt.rcv_tsecr = 0;
4201 /* Fast parse options. This hopes to only see timestamps.
4202 * If it is wrong it falls back on tcp_parse_options().
4204 static bool tcp_fast_parse_options(const struct net *net,
4205 const struct sk_buff *skb,
4206 const struct tcphdr *th, struct tcp_sock *tp)
4208 /* In the spirit of fast parsing, compare doff directly to constant
4209 * values. Because equality is used, short doff can be ignored here.
4211 if (th->doff == (sizeof(*th) / 4)) {
4212 tp->rx_opt.saw_tstamp = 0;
4214 } else if (tp->rx_opt.tstamp_ok &&
4215 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
4216 if (tcp_parse_aligned_timestamp(tp, th))
4220 tcp_parse_options(net, skb, &tp->rx_opt, 1, NULL);
4221 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
4222 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
4227 #ifdef CONFIG_TCP_MD5SIG
4229 * Parse MD5 Signature option
4231 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
4233 int length = (th->doff << 2) - sizeof(*th);
4234 const u8 *ptr = (const u8 *)(th + 1);
4236 /* If not enough data remaining, we can short cut */
4237 while (length >= TCPOLEN_MD5SIG) {
4238 int opcode = *ptr++;
4249 if (opsize < 2 || opsize > length)
4251 if (opcode == TCPOPT_MD5SIG)
4252 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
4259 EXPORT_SYMBOL(tcp_parse_md5sig_option);
4262 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
4264 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
4265 * it can pass through stack. So, the following predicate verifies that
4266 * this segment is not used for anything but congestion avoidance or
4267 * fast retransmit. Moreover, we even are able to eliminate most of such
4268 * second order effects, if we apply some small "replay" window (~RTO)
4269 * to timestamp space.
4271 * All these measures still do not guarantee that we reject wrapped ACKs
4272 * on networks with high bandwidth, when sequence space is recycled fastly,
4273 * but it guarantees that such events will be very rare and do not affect
4274 * connection seriously. This doesn't look nice, but alas, PAWS is really
4277 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
4278 * states that events when retransmit arrives after original data are rare.
4279 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
4280 * the biggest problem on large power networks even with minor reordering.
4281 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
4282 * up to bandwidth of 18Gigabit/sec. 8) ]
4285 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
4287 const struct tcp_sock *tp = tcp_sk(sk);
4288 const struct tcphdr *th = tcp_hdr(skb);
4289 u32 seq = TCP_SKB_CB(skb)->seq;
4290 u32 ack = TCP_SKB_CB(skb)->ack_seq;
4292 return (/* 1. Pure ACK with correct sequence number. */
4293 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
4295 /* 2. ... and duplicate ACK. */
4296 ack == tp->snd_una &&
4298 /* 3. ... and does not update window. */
4299 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
4301 /* 4. ... and sits in replay window. */
4302 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
4305 static inline bool tcp_paws_discard(const struct sock *sk,
4306 const struct sk_buff *skb)
4308 const struct tcp_sock *tp = tcp_sk(sk);
4310 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
4311 !tcp_disordered_ack(sk, skb);
4314 /* Check segment sequence number for validity.
4316 * Segment controls are considered valid, if the segment
4317 * fits to the window after truncation to the window. Acceptability
4318 * of data (and SYN, FIN, of course) is checked separately.
4319 * See tcp_data_queue(), for example.
4321 * Also, controls (RST is main one) are accepted using RCV.WUP instead
4322 * of RCV.NXT. Peer still did not advance his SND.UNA when we
4323 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4324 * (borrowed from freebsd)
4327 static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
4329 return !before(end_seq, tp->rcv_wup) &&
4330 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
4333 /* When we get a reset we do this. */
4334 void tcp_reset(struct sock *sk, struct sk_buff *skb)
4336 trace_tcp_receive_reset(sk);
4338 /* mptcp can't tell us to ignore reset pkts,
4339 * so just ignore the return value of mptcp_incoming_options().
4341 if (sk_is_mptcp(sk))
4342 mptcp_incoming_options(sk, skb);
4344 /* We want the right error as BSD sees it (and indeed as we do). */
4345 switch (sk->sk_state) {
4347 sk->sk_err = ECONNREFUSED;
4349 case TCP_CLOSE_WAIT:
4355 sk->sk_err = ECONNRESET;
4357 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4360 tcp_write_queue_purge(sk);
4363 if (!sock_flag(sk, SOCK_DEAD))
4364 sk_error_report(sk);
4368 * Process the FIN bit. This now behaves as it is supposed to work
4369 * and the FIN takes effect when it is validly part of sequence
4370 * space. Not before when we get holes.
4372 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4373 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4376 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4377 * close and we go into CLOSING (and later onto TIME-WAIT)
4379 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4381 void tcp_fin(struct sock *sk)
4383 struct tcp_sock *tp = tcp_sk(sk);
4385 inet_csk_schedule_ack(sk);
4387 WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | RCV_SHUTDOWN);
4388 sock_set_flag(sk, SOCK_DONE);
4390 switch (sk->sk_state) {
4392 case TCP_ESTABLISHED:
4393 /* Move to CLOSE_WAIT */
4394 tcp_set_state(sk, TCP_CLOSE_WAIT);
4395 inet_csk_enter_pingpong_mode(sk);
4398 case TCP_CLOSE_WAIT:
4400 /* Received a retransmission of the FIN, do
4405 /* RFC793: Remain in the LAST-ACK state. */
4409 /* This case occurs when a simultaneous close
4410 * happens, we must ack the received FIN and
4411 * enter the CLOSING state.
4414 tcp_set_state(sk, TCP_CLOSING);
4417 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4419 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4422 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4423 * cases we should never reach this piece of code.
4425 pr_err("%s: Impossible, sk->sk_state=%d\n",
4426 __func__, sk->sk_state);
4430 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4431 * Probably, we should reset in this case. For now drop them.
4433 skb_rbtree_purge(&tp->out_of_order_queue);
4434 if (tcp_is_sack(tp))
4435 tcp_sack_reset(&tp->rx_opt);
4437 if (!sock_flag(sk, SOCK_DEAD)) {
4438 sk->sk_state_change(sk);
4440 /* Do not send POLL_HUP for half duplex close. */
4441 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4442 sk->sk_state == TCP_CLOSE)
4443 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4445 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4449 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4452 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4453 if (before(seq, sp->start_seq))
4454 sp->start_seq = seq;
4455 if (after(end_seq, sp->end_seq))
4456 sp->end_seq = end_seq;
4462 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4464 struct tcp_sock *tp = tcp_sk(sk);
4466 if (tcp_is_sack(tp) && READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_dsack)) {
4469 if (before(seq, tp->rcv_nxt))
4470 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4472 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4474 NET_INC_STATS(sock_net(sk), mib_idx);
4476 tp->rx_opt.dsack = 1;
4477 tp->duplicate_sack[0].start_seq = seq;
4478 tp->duplicate_sack[0].end_seq = end_seq;
4482 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4484 struct tcp_sock *tp = tcp_sk(sk);
4486 if (!tp->rx_opt.dsack)
4487 tcp_dsack_set(sk, seq, end_seq);
4489 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4492 static void tcp_rcv_spurious_retrans(struct sock *sk, const struct sk_buff *skb)
4494 /* When the ACK path fails or drops most ACKs, the sender would
4495 * timeout and spuriously retransmit the same segment repeatedly.
4496 * The receiver remembers and reflects via DSACKs. Leverage the
4497 * DSACK state and change the txhash to re-route speculatively.
4499 if (TCP_SKB_CB(skb)->seq == tcp_sk(sk)->duplicate_sack[0].start_seq &&
4500 sk_rethink_txhash(sk))
4501 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDUPLICATEDATAREHASH);
4504 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4506 struct tcp_sock *tp = tcp_sk(sk);
4508 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4509 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4510 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4511 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
4513 if (tcp_is_sack(tp) && READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_dsack)) {
4514 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4516 tcp_rcv_spurious_retrans(sk, skb);
4517 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4518 end_seq = tp->rcv_nxt;
4519 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4526 /* These routines update the SACK block as out-of-order packets arrive or
4527 * in-order packets close up the sequence space.
4529 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4532 struct tcp_sack_block *sp = &tp->selective_acks[0];
4533 struct tcp_sack_block *swalk = sp + 1;
4535 /* See if the recent change to the first SACK eats into
4536 * or hits the sequence space of other SACK blocks, if so coalesce.
4538 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4539 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4542 /* Zap SWALK, by moving every further SACK up by one slot.
4543 * Decrease num_sacks.
4545 tp->rx_opt.num_sacks--;
4546 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4555 void tcp_sack_compress_send_ack(struct sock *sk)
4557 struct tcp_sock *tp = tcp_sk(sk);
4559 if (!tp->compressed_ack)
4562 if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1)
4565 /* Since we have to send one ack finally,
4566 * substract one from tp->compressed_ack to keep
4567 * LINUX_MIB_TCPACKCOMPRESSED accurate.
4569 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED,
4570 tp->compressed_ack - 1);
4572 tp->compressed_ack = 0;
4576 /* Reasonable amount of sack blocks included in TCP SACK option
4577 * The max is 4, but this becomes 3 if TCP timestamps are there.
4578 * Given that SACK packets might be lost, be conservative and use 2.
4580 #define TCP_SACK_BLOCKS_EXPECTED 2
4582 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4584 struct tcp_sock *tp = tcp_sk(sk);
4585 struct tcp_sack_block *sp = &tp->selective_acks[0];
4586 int cur_sacks = tp->rx_opt.num_sacks;
4592 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4593 if (tcp_sack_extend(sp, seq, end_seq)) {
4594 if (this_sack >= TCP_SACK_BLOCKS_EXPECTED)
4595 tcp_sack_compress_send_ack(sk);
4596 /* Rotate this_sack to the first one. */
4597 for (; this_sack > 0; this_sack--, sp--)
4598 swap(*sp, *(sp - 1));
4600 tcp_sack_maybe_coalesce(tp);
4605 if (this_sack >= TCP_SACK_BLOCKS_EXPECTED)
4606 tcp_sack_compress_send_ack(sk);
4608 /* Could not find an adjacent existing SACK, build a new one,
4609 * put it at the front, and shift everyone else down. We
4610 * always know there is at least one SACK present already here.
4612 * If the sack array is full, forget about the last one.
4614 if (this_sack >= TCP_NUM_SACKS) {
4616 tp->rx_opt.num_sacks--;
4619 for (; this_sack > 0; this_sack--, sp--)
4623 /* Build the new head SACK, and we're done. */
4624 sp->start_seq = seq;
4625 sp->end_seq = end_seq;
4626 tp->rx_opt.num_sacks++;
4629 /* RCV.NXT advances, some SACKs should be eaten. */
4631 static void tcp_sack_remove(struct tcp_sock *tp)
4633 struct tcp_sack_block *sp = &tp->selective_acks[0];
4634 int num_sacks = tp->rx_opt.num_sacks;
4637 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4638 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4639 tp->rx_opt.num_sacks = 0;
4643 for (this_sack = 0; this_sack < num_sacks;) {
4644 /* Check if the start of the sack is covered by RCV.NXT. */
4645 if (!before(tp->rcv_nxt, sp->start_seq)) {
4648 /* RCV.NXT must cover all the block! */
4649 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4651 /* Zap this SACK, by moving forward any other SACKS. */
4652 for (i = this_sack+1; i < num_sacks; i++)
4653 tp->selective_acks[i-1] = tp->selective_acks[i];
4660 tp->rx_opt.num_sacks = num_sacks;
4664 * tcp_try_coalesce - try to merge skb to prior one
4667 * @from: buffer to add in queue
4668 * @fragstolen: pointer to boolean
4670 * Before queueing skb @from after @to, try to merge them
4671 * to reduce overall memory use and queue lengths, if cost is small.
4672 * Packets in ofo or receive queues can stay a long time.
4673 * Better try to coalesce them right now to avoid future collapses.
4674 * Returns true if caller should free @from instead of queueing it
4676 static bool tcp_try_coalesce(struct sock *sk,
4678 struct sk_buff *from,
4683 *fragstolen = false;
4685 /* Its possible this segment overlaps with prior segment in queue */
4686 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4689 if (!mptcp_skb_can_collapse(to, from))
4692 #ifdef CONFIG_TLS_DEVICE
4693 if (from->decrypted != to->decrypted)
4697 if (!skb_try_coalesce(to, from, fragstolen, &delta))
4700 atomic_add(delta, &sk->sk_rmem_alloc);
4701 sk_mem_charge(sk, delta);
4702 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4703 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4704 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4705 TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
4707 if (TCP_SKB_CB(from)->has_rxtstamp) {
4708 TCP_SKB_CB(to)->has_rxtstamp = true;
4709 to->tstamp = from->tstamp;
4710 skb_hwtstamps(to)->hwtstamp = skb_hwtstamps(from)->hwtstamp;
4716 static bool tcp_ooo_try_coalesce(struct sock *sk,
4718 struct sk_buff *from,
4721 bool res = tcp_try_coalesce(sk, to, from, fragstolen);
4723 /* In case tcp_drop_reason() is called later, update to->gso_segs */
4725 u32 gso_segs = max_t(u16, 1, skb_shinfo(to)->gso_segs) +
4726 max_t(u16, 1, skb_shinfo(from)->gso_segs);
4728 skb_shinfo(to)->gso_segs = min_t(u32, gso_segs, 0xFFFF);
4733 static void tcp_drop_reason(struct sock *sk, struct sk_buff *skb,
4734 enum skb_drop_reason reason)
4736 sk_drops_add(sk, skb);
4737 kfree_skb_reason(skb, reason);
4740 /* This one checks to see if we can put data from the
4741 * out_of_order queue into the receive_queue.
4743 static void tcp_ofo_queue(struct sock *sk)
4745 struct tcp_sock *tp = tcp_sk(sk);
4746 __u32 dsack_high = tp->rcv_nxt;
4747 bool fin, fragstolen, eaten;
4748 struct sk_buff *skb, *tail;
4751 p = rb_first(&tp->out_of_order_queue);
4754 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4757 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4758 __u32 dsack = dsack_high;
4759 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4760 dsack_high = TCP_SKB_CB(skb)->end_seq;
4761 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4764 rb_erase(&skb->rbnode, &tp->out_of_order_queue);
4766 if (unlikely(!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))) {
4767 tcp_drop_reason(sk, skb, SKB_DROP_REASON_TCP_OFO_DROP);
4771 tail = skb_peek_tail(&sk->sk_receive_queue);
4772 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen);
4773 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4774 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
4776 __skb_queue_tail(&sk->sk_receive_queue, skb);
4778 kfree_skb_partial(skb, fragstolen);
4780 if (unlikely(fin)) {
4782 /* tcp_fin() purges tp->out_of_order_queue,
4783 * so we must end this loop right now.
4790 static bool tcp_prune_ofo_queue(struct sock *sk);
4791 static int tcp_prune_queue(struct sock *sk);
4793 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4796 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4797 !sk_rmem_schedule(sk, skb, size)) {
4799 if (tcp_prune_queue(sk) < 0)
4802 while (!sk_rmem_schedule(sk, skb, size)) {
4803 if (!tcp_prune_ofo_queue(sk))
4810 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4812 struct tcp_sock *tp = tcp_sk(sk);
4813 struct rb_node **p, *parent;
4814 struct sk_buff *skb1;
4818 tcp_ecn_check_ce(sk, skb);
4820 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4821 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFODROP);
4822 sk->sk_data_ready(sk);
4823 tcp_drop_reason(sk, skb, SKB_DROP_REASON_PROTO_MEM);
4827 /* Disable header prediction. */
4829 inet_csk_schedule_ack(sk);
4831 tp->rcv_ooopack += max_t(u16, 1, skb_shinfo(skb)->gso_segs);
4832 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4833 seq = TCP_SKB_CB(skb)->seq;
4834 end_seq = TCP_SKB_CB(skb)->end_seq;
4836 p = &tp->out_of_order_queue.rb_node;
4837 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4838 /* Initial out of order segment, build 1 SACK. */
4839 if (tcp_is_sack(tp)) {
4840 tp->rx_opt.num_sacks = 1;
4841 tp->selective_acks[0].start_seq = seq;
4842 tp->selective_acks[0].end_seq = end_seq;
4844 rb_link_node(&skb->rbnode, NULL, p);
4845 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4846 tp->ooo_last_skb = skb;
4850 /* In the typical case, we are adding an skb to the end of the list.
4851 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
4853 if (tcp_ooo_try_coalesce(sk, tp->ooo_last_skb,
4854 skb, &fragstolen)) {
4856 /* For non sack flows, do not grow window to force DUPACK
4857 * and trigger fast retransmit.
4859 if (tcp_is_sack(tp))
4860 tcp_grow_window(sk, skb, true);
4861 kfree_skb_partial(skb, fragstolen);
4865 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
4866 if (!before(seq, TCP_SKB_CB(tp->ooo_last_skb)->end_seq)) {
4867 parent = &tp->ooo_last_skb->rbnode;
4868 p = &parent->rb_right;
4872 /* Find place to insert this segment. Handle overlaps on the way. */
4876 skb1 = rb_to_skb(parent);
4877 if (before(seq, TCP_SKB_CB(skb1)->seq)) {
4878 p = &parent->rb_left;
4881 if (before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4882 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4883 /* All the bits are present. Drop. */
4884 NET_INC_STATS(sock_net(sk),
4885 LINUX_MIB_TCPOFOMERGE);
4886 tcp_drop_reason(sk, skb,
4887 SKB_DROP_REASON_TCP_OFOMERGE);
4889 tcp_dsack_set(sk, seq, end_seq);
4892 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4893 /* Partial overlap. */
4894 tcp_dsack_set(sk, seq, TCP_SKB_CB(skb1)->end_seq);
4896 /* skb's seq == skb1's seq and skb covers skb1.
4897 * Replace skb1 with skb.
4899 rb_replace_node(&skb1->rbnode, &skb->rbnode,
4900 &tp->out_of_order_queue);
4901 tcp_dsack_extend(sk,
4902 TCP_SKB_CB(skb1)->seq,
4903 TCP_SKB_CB(skb1)->end_seq);
4904 NET_INC_STATS(sock_net(sk),
4905 LINUX_MIB_TCPOFOMERGE);
4906 tcp_drop_reason(sk, skb1,
4907 SKB_DROP_REASON_TCP_OFOMERGE);
4910 } else if (tcp_ooo_try_coalesce(sk, skb1,
4911 skb, &fragstolen)) {
4914 p = &parent->rb_right;
4917 /* Insert segment into RB tree. */
4918 rb_link_node(&skb->rbnode, parent, p);
4919 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4922 /* Remove other segments covered by skb. */
4923 while ((skb1 = skb_rb_next(skb)) != NULL) {
4924 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4926 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4927 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4931 rb_erase(&skb1->rbnode, &tp->out_of_order_queue);
4932 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4933 TCP_SKB_CB(skb1)->end_seq);
4934 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4935 tcp_drop_reason(sk, skb1, SKB_DROP_REASON_TCP_OFOMERGE);
4937 /* If there is no skb after us, we are the last_skb ! */
4939 tp->ooo_last_skb = skb;
4942 if (tcp_is_sack(tp))
4943 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4946 /* For non sack flows, do not grow window to force DUPACK
4947 * and trigger fast retransmit.
4949 if (tcp_is_sack(tp))
4950 tcp_grow_window(sk, skb, false);
4952 skb_set_owner_r(skb, sk);
4956 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb,
4960 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4963 tcp_try_coalesce(sk, tail,
4964 skb, fragstolen)) ? 1 : 0;
4965 tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq);
4967 __skb_queue_tail(&sk->sk_receive_queue, skb);
4968 skb_set_owner_r(skb, sk);
4973 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4975 struct sk_buff *skb;
4983 if (size > PAGE_SIZE) {
4984 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS);
4986 data_len = npages << PAGE_SHIFT;
4987 size = data_len + (size & ~PAGE_MASK);
4989 skb = alloc_skb_with_frags(size - data_len, data_len,
4990 PAGE_ALLOC_COSTLY_ORDER,
4991 &err, sk->sk_allocation);
4995 skb_put(skb, size - data_len);
4996 skb->data_len = data_len;
4999 if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) {
5000 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP);
5004 err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size);
5008 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
5009 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
5010 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
5012 if (tcp_queue_rcv(sk, skb, &fragstolen)) {
5013 WARN_ON_ONCE(fragstolen); /* should not happen */
5025 void tcp_data_ready(struct sock *sk)
5027 if (tcp_epollin_ready(sk, sk->sk_rcvlowat) || sock_flag(sk, SOCK_DONE))
5028 sk->sk_data_ready(sk);
5031 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
5033 struct tcp_sock *tp = tcp_sk(sk);
5034 enum skb_drop_reason reason;
5038 /* If a subflow has been reset, the packet should not continue
5039 * to be processed, drop the packet.
5041 if (sk_is_mptcp(sk) && !mptcp_incoming_options(sk, skb)) {
5046 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
5051 __skb_pull(skb, tcp_hdr(skb)->doff * 4);
5053 reason = SKB_DROP_REASON_NOT_SPECIFIED;
5054 tp->rx_opt.dsack = 0;
5056 /* Queue data for delivery to the user.
5057 * Packets in sequence go to the receive queue.
5058 * Out of sequence packets to the out_of_order_queue.
5060 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
5061 if (tcp_receive_window(tp) == 0) {
5062 reason = SKB_DROP_REASON_TCP_ZEROWINDOW;
5063 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP);
5067 /* Ok. In sequence. In window. */
5069 if (skb_queue_len(&sk->sk_receive_queue) == 0)
5070 sk_forced_mem_schedule(sk, skb->truesize);
5071 else if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) {
5072 reason = SKB_DROP_REASON_PROTO_MEM;
5073 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP);
5074 sk->sk_data_ready(sk);
5078 eaten = tcp_queue_rcv(sk, skb, &fragstolen);
5080 tcp_event_data_recv(sk, skb);
5081 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
5084 if (!RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
5087 /* RFC5681. 4.2. SHOULD send immediate ACK, when
5088 * gap in queue is filled.
5090 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
5091 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW;
5094 if (tp->rx_opt.num_sacks)
5095 tcp_sack_remove(tp);
5097 tcp_fast_path_check(sk);
5100 kfree_skb_partial(skb, fragstolen);
5101 if (!sock_flag(sk, SOCK_DEAD))
5106 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
5107 tcp_rcv_spurious_retrans(sk, skb);
5108 /* A retransmit, 2nd most common case. Force an immediate ack. */
5109 reason = SKB_DROP_REASON_TCP_OLD_DATA;
5110 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
5111 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
5114 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
5115 inet_csk_schedule_ack(sk);
5117 tcp_drop_reason(sk, skb, reason);
5121 /* Out of window. F.e. zero window probe. */
5122 if (!before(TCP_SKB_CB(skb)->seq,
5123 tp->rcv_nxt + tcp_receive_window(tp))) {
5124 reason = SKB_DROP_REASON_TCP_OVERWINDOW;
5128 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
5129 /* Partial packet, seq < rcv_next < end_seq */
5130 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
5132 /* If window is closed, drop tail of packet. But after
5133 * remembering D-SACK for its head made in previous line.
5135 if (!tcp_receive_window(tp)) {
5136 reason = SKB_DROP_REASON_TCP_ZEROWINDOW;
5137 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP);
5143 tcp_data_queue_ofo(sk, skb);
5146 static struct sk_buff *tcp_skb_next(struct sk_buff *skb, struct sk_buff_head *list)
5149 return !skb_queue_is_last(list, skb) ? skb->next : NULL;
5151 return skb_rb_next(skb);
5154 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
5155 struct sk_buff_head *list,
5156 struct rb_root *root)
5158 struct sk_buff *next = tcp_skb_next(skb, list);
5161 __skb_unlink(skb, list);
5163 rb_erase(&skb->rbnode, root);
5166 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
5171 /* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */
5172 void tcp_rbtree_insert(struct rb_root *root, struct sk_buff *skb)
5174 struct rb_node **p = &root->rb_node;
5175 struct rb_node *parent = NULL;
5176 struct sk_buff *skb1;
5180 skb1 = rb_to_skb(parent);
5181 if (before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb1)->seq))
5182 p = &parent->rb_left;
5184 p = &parent->rb_right;
5186 rb_link_node(&skb->rbnode, parent, p);
5187 rb_insert_color(&skb->rbnode, root);
5190 /* Collapse contiguous sequence of skbs head..tail with
5191 * sequence numbers start..end.
5193 * If tail is NULL, this means until the end of the queue.
5195 * Segments with FIN/SYN are not collapsed (only because this
5199 tcp_collapse(struct sock *sk, struct sk_buff_head *list, struct rb_root *root,
5200 struct sk_buff *head, struct sk_buff *tail, u32 start, u32 end)
5202 struct sk_buff *skb = head, *n;
5203 struct sk_buff_head tmp;
5206 /* First, check that queue is collapsible and find
5207 * the point where collapsing can be useful.
5210 for (end_of_skbs = true; skb != NULL && skb != tail; skb = n) {
5211 n = tcp_skb_next(skb, list);
5213 /* No new bits? It is possible on ofo queue. */
5214 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
5215 skb = tcp_collapse_one(sk, skb, list, root);
5221 /* The first skb to collapse is:
5223 * - bloated or contains data before "start" or
5224 * overlaps to the next one and mptcp allow collapsing.
5226 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
5227 (tcp_win_from_space(sk, skb->truesize) > skb->len ||
5228 before(TCP_SKB_CB(skb)->seq, start))) {
5229 end_of_skbs = false;
5233 if (n && n != tail && mptcp_skb_can_collapse(skb, n) &&
5234 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(n)->seq) {
5235 end_of_skbs = false;
5239 /* Decided to skip this, advance start seq. */
5240 start = TCP_SKB_CB(skb)->end_seq;
5243 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
5246 __skb_queue_head_init(&tmp);
5248 while (before(start, end)) {
5249 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
5250 struct sk_buff *nskb;
5252 nskb = alloc_skb(copy, GFP_ATOMIC);
5256 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
5257 #ifdef CONFIG_TLS_DEVICE
5258 nskb->decrypted = skb->decrypted;
5260 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
5262 __skb_queue_before(list, skb, nskb);
5264 __skb_queue_tail(&tmp, nskb); /* defer rbtree insertion */
5265 skb_set_owner_r(nskb, sk);
5266 mptcp_skb_ext_move(nskb, skb);
5268 /* Copy data, releasing collapsed skbs. */
5270 int offset = start - TCP_SKB_CB(skb)->seq;
5271 int size = TCP_SKB_CB(skb)->end_seq - start;
5275 size = min(copy, size);
5276 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
5278 TCP_SKB_CB(nskb)->end_seq += size;
5282 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
5283 skb = tcp_collapse_one(sk, skb, list, root);
5286 !mptcp_skb_can_collapse(nskb, skb) ||
5287 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
5289 #ifdef CONFIG_TLS_DEVICE
5290 if (skb->decrypted != nskb->decrypted)
5297 skb_queue_walk_safe(&tmp, skb, n)
5298 tcp_rbtree_insert(root, skb);
5301 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
5302 * and tcp_collapse() them until all the queue is collapsed.
5304 static void tcp_collapse_ofo_queue(struct sock *sk)
5306 struct tcp_sock *tp = tcp_sk(sk);
5307 u32 range_truesize, sum_tiny = 0;
5308 struct sk_buff *skb, *head;
5311 skb = skb_rb_first(&tp->out_of_order_queue);
5314 tp->ooo_last_skb = skb_rb_last(&tp->out_of_order_queue);
5317 start = TCP_SKB_CB(skb)->seq;
5318 end = TCP_SKB_CB(skb)->end_seq;
5319 range_truesize = skb->truesize;
5321 for (head = skb;;) {
5322 skb = skb_rb_next(skb);
5324 /* Range is terminated when we see a gap or when
5325 * we are at the queue end.
5328 after(TCP_SKB_CB(skb)->seq, end) ||
5329 before(TCP_SKB_CB(skb)->end_seq, start)) {
5330 /* Do not attempt collapsing tiny skbs */
5331 if (range_truesize != head->truesize ||
5332 end - start >= SKB_WITH_OVERHEAD(PAGE_SIZE)) {
5333 tcp_collapse(sk, NULL, &tp->out_of_order_queue,
5334 head, skb, start, end);
5336 sum_tiny += range_truesize;
5337 if (sum_tiny > sk->sk_rcvbuf >> 3)
5343 range_truesize += skb->truesize;
5344 if (unlikely(before(TCP_SKB_CB(skb)->seq, start)))
5345 start = TCP_SKB_CB(skb)->seq;
5346 if (after(TCP_SKB_CB(skb)->end_seq, end))
5347 end = TCP_SKB_CB(skb)->end_seq;
5352 * Clean the out-of-order queue to make room.
5353 * We drop high sequences packets to :
5354 * 1) Let a chance for holes to be filled.
5355 * 2) not add too big latencies if thousands of packets sit there.
5356 * (But if application shrinks SO_RCVBUF, we could still end up
5357 * freeing whole queue here)
5358 * 3) Drop at least 12.5 % of sk_rcvbuf to avoid malicious attacks.
5360 * Return true if queue has shrunk.
5362 static bool tcp_prune_ofo_queue(struct sock *sk)
5364 struct tcp_sock *tp = tcp_sk(sk);
5365 struct rb_node *node, *prev;
5368 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
5371 NET_INC_STATS(sock_net(sk), LINUX_MIB_OFOPRUNED);
5372 goal = sk->sk_rcvbuf >> 3;
5373 node = &tp->ooo_last_skb->rbnode;
5375 prev = rb_prev(node);
5376 rb_erase(node, &tp->out_of_order_queue);
5377 goal -= rb_to_skb(node)->truesize;
5378 tcp_drop_reason(sk, rb_to_skb(node),
5379 SKB_DROP_REASON_TCP_OFO_QUEUE_PRUNE);
5380 if (!prev || goal <= 0) {
5381 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf &&
5382 !tcp_under_memory_pressure(sk))
5384 goal = sk->sk_rcvbuf >> 3;
5388 tp->ooo_last_skb = rb_to_skb(prev);
5390 /* Reset SACK state. A conforming SACK implementation will
5391 * do the same at a timeout based retransmit. When a connection
5392 * is in a sad state like this, we care only about integrity
5393 * of the connection not performance.
5395 if (tp->rx_opt.sack_ok)
5396 tcp_sack_reset(&tp->rx_opt);
5400 /* Reduce allocated memory if we can, trying to get
5401 * the socket within its memory limits again.
5403 * Return less than zero if we should start dropping frames
5404 * until the socket owning process reads some of the data
5405 * to stabilize the situation.
5407 static int tcp_prune_queue(struct sock *sk)
5409 struct tcp_sock *tp = tcp_sk(sk);
5411 NET_INC_STATS(sock_net(sk), LINUX_MIB_PRUNECALLED);
5413 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
5414 tcp_clamp_window(sk);
5415 else if (tcp_under_memory_pressure(sk))
5416 tcp_adjust_rcv_ssthresh(sk);
5418 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5421 tcp_collapse_ofo_queue(sk);
5422 if (!skb_queue_empty(&sk->sk_receive_queue))
5423 tcp_collapse(sk, &sk->sk_receive_queue, NULL,
5424 skb_peek(&sk->sk_receive_queue),
5426 tp->copied_seq, tp->rcv_nxt);
5428 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5431 /* Collapsing did not help, destructive actions follow.
5432 * This must not ever occur. */
5434 tcp_prune_ofo_queue(sk);
5436 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5439 /* If we are really being abused, tell the caller to silently
5440 * drop receive data on the floor. It will get retransmitted
5441 * and hopefully then we'll have sufficient space.
5443 NET_INC_STATS(sock_net(sk), LINUX_MIB_RCVPRUNED);
5445 /* Massive buffer overcommit. */
5450 static bool tcp_should_expand_sndbuf(struct sock *sk)
5452 const struct tcp_sock *tp = tcp_sk(sk);
5454 /* If the user specified a specific send buffer setting, do
5457 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
5460 /* If we are under global TCP memory pressure, do not expand. */
5461 if (tcp_under_memory_pressure(sk)) {
5462 int unused_mem = sk_unused_reserved_mem(sk);
5464 /* Adjust sndbuf according to reserved mem. But make sure
5465 * it never goes below SOCK_MIN_SNDBUF.
5466 * See sk_stream_moderate_sndbuf() for more details.
5468 if (unused_mem > SOCK_MIN_SNDBUF)
5469 WRITE_ONCE(sk->sk_sndbuf, unused_mem);
5474 /* If we are under soft global TCP memory pressure, do not expand. */
5475 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
5478 /* If we filled the congestion window, do not expand. */
5479 if (tcp_packets_in_flight(tp) >= tcp_snd_cwnd(tp))
5485 static void tcp_new_space(struct sock *sk)
5487 struct tcp_sock *tp = tcp_sk(sk);
5489 if (tcp_should_expand_sndbuf(sk)) {
5490 tcp_sndbuf_expand(sk);
5491 tp->snd_cwnd_stamp = tcp_jiffies32;
5494 INDIRECT_CALL_1(sk->sk_write_space, sk_stream_write_space, sk);
5497 /* Caller made space either from:
5498 * 1) Freeing skbs in rtx queues (after tp->snd_una has advanced)
5499 * 2) Sent skbs from output queue (and thus advancing tp->snd_nxt)
5501 * We might be able to generate EPOLLOUT to the application if:
5502 * 1) Space consumed in output/rtx queues is below sk->sk_sndbuf/2
5503 * 2) notsent amount (tp->write_seq - tp->snd_nxt) became
5504 * small enough that tcp_stream_memory_free() decides it
5505 * is time to generate EPOLLOUT.
5507 void tcp_check_space(struct sock *sk)
5509 /* pairs with tcp_poll() */
5511 if (sk->sk_socket &&
5512 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
5514 if (!test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
5515 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
5519 static inline void tcp_data_snd_check(struct sock *sk)
5521 tcp_push_pending_frames(sk);
5522 tcp_check_space(sk);
5526 * Check if sending an ack is needed.
5528 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
5530 struct tcp_sock *tp = tcp_sk(sk);
5531 unsigned long rtt, delay;
5533 /* More than one full frame received... */
5534 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
5535 /* ... and right edge of window advances far enough.
5536 * (tcp_recvmsg() will send ACK otherwise).
5537 * If application uses SO_RCVLOWAT, we want send ack now if
5538 * we have not received enough bytes to satisfy the condition.
5540 (tp->rcv_nxt - tp->copied_seq < sk->sk_rcvlowat ||
5541 __tcp_select_window(sk) >= tp->rcv_wnd)) ||
5542 /* We ACK each frame or... */
5543 tcp_in_quickack_mode(sk) ||
5544 /* Protocol state mandates a one-time immediate ACK */
5545 inet_csk(sk)->icsk_ack.pending & ICSK_ACK_NOW) {
5551 if (!ofo_possible || RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
5552 tcp_send_delayed_ack(sk);
5556 if (!tcp_is_sack(tp) ||
5557 tp->compressed_ack >= READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_nr))
5560 if (tp->compressed_ack_rcv_nxt != tp->rcv_nxt) {
5561 tp->compressed_ack_rcv_nxt = tp->rcv_nxt;
5562 tp->dup_ack_counter = 0;
5564 if (tp->dup_ack_counter < TCP_FASTRETRANS_THRESH) {
5565 tp->dup_ack_counter++;
5568 tp->compressed_ack++;
5569 if (hrtimer_is_queued(&tp->compressed_ack_timer))
5572 /* compress ack timer : 5 % of rtt, but no more than tcp_comp_sack_delay_ns */
5574 rtt = tp->rcv_rtt_est.rtt_us;
5575 if (tp->srtt_us && tp->srtt_us < rtt)
5578 delay = min_t(unsigned long,
5579 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_delay_ns),
5580 rtt * (NSEC_PER_USEC >> 3)/20);
5582 hrtimer_start_range_ns(&tp->compressed_ack_timer, ns_to_ktime(delay),
5583 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_slack_ns),
5584 HRTIMER_MODE_REL_PINNED_SOFT);
5587 static inline void tcp_ack_snd_check(struct sock *sk)
5589 if (!inet_csk_ack_scheduled(sk)) {
5590 /* We sent a data segment already. */
5593 __tcp_ack_snd_check(sk, 1);
5597 * This routine is only called when we have urgent data
5598 * signaled. Its the 'slow' part of tcp_urg. It could be
5599 * moved inline now as tcp_urg is only called from one
5600 * place. We handle URGent data wrong. We have to - as
5601 * BSD still doesn't use the correction from RFC961.
5602 * For 1003.1g we should support a new option TCP_STDURG to permit
5603 * either form (or just set the sysctl tcp_stdurg).
5606 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
5608 struct tcp_sock *tp = tcp_sk(sk);
5609 u32 ptr = ntohs(th->urg_ptr);
5611 if (ptr && !READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_stdurg))
5613 ptr += ntohl(th->seq);
5615 /* Ignore urgent data that we've already seen and read. */
5616 if (after(tp->copied_seq, ptr))
5619 /* Do not replay urg ptr.
5621 * NOTE: interesting situation not covered by specs.
5622 * Misbehaving sender may send urg ptr, pointing to segment,
5623 * which we already have in ofo queue. We are not able to fetch
5624 * such data and will stay in TCP_URG_NOTYET until will be eaten
5625 * by recvmsg(). Seems, we are not obliged to handle such wicked
5626 * situations. But it is worth to think about possibility of some
5627 * DoSes using some hypothetical application level deadlock.
5629 if (before(ptr, tp->rcv_nxt))
5632 /* Do we already have a newer (or duplicate) urgent pointer? */
5633 if (tp->urg_data && !after(ptr, tp->urg_seq))
5636 /* Tell the world about our new urgent pointer. */
5639 /* We may be adding urgent data when the last byte read was
5640 * urgent. To do this requires some care. We cannot just ignore
5641 * tp->copied_seq since we would read the last urgent byte again
5642 * as data, nor can we alter copied_seq until this data arrives
5643 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5645 * NOTE. Double Dutch. Rendering to plain English: author of comment
5646 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5647 * and expect that both A and B disappear from stream. This is _wrong_.
5648 * Though this happens in BSD with high probability, this is occasional.
5649 * Any application relying on this is buggy. Note also, that fix "works"
5650 * only in this artificial test. Insert some normal data between A and B and we will
5651 * decline of BSD again. Verdict: it is better to remove to trap
5654 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5655 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5656 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5658 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5659 __skb_unlink(skb, &sk->sk_receive_queue);
5664 WRITE_ONCE(tp->urg_data, TCP_URG_NOTYET);
5665 WRITE_ONCE(tp->urg_seq, ptr);
5667 /* Disable header prediction. */
5671 /* This is the 'fast' part of urgent handling. */
5672 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
5674 struct tcp_sock *tp = tcp_sk(sk);
5676 /* Check if we get a new urgent pointer - normally not. */
5677 if (unlikely(th->urg))
5678 tcp_check_urg(sk, th);
5680 /* Do we wait for any urgent data? - normally not... */
5681 if (unlikely(tp->urg_data == TCP_URG_NOTYET)) {
5682 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5685 /* Is the urgent pointer pointing into this packet? */
5686 if (ptr < skb->len) {
5688 if (skb_copy_bits(skb, ptr, &tmp, 1))
5690 WRITE_ONCE(tp->urg_data, TCP_URG_VALID | tmp);
5691 if (!sock_flag(sk, SOCK_DEAD))
5692 sk->sk_data_ready(sk);
5697 /* Accept RST for rcv_nxt - 1 after a FIN.
5698 * When tcp connections are abruptly terminated from Mac OSX (via ^C), a
5699 * FIN is sent followed by a RST packet. The RST is sent with the same
5700 * sequence number as the FIN, and thus according to RFC 5961 a challenge
5701 * ACK should be sent. However, Mac OSX rate limits replies to challenge
5702 * ACKs on the closed socket. In addition middleboxes can drop either the
5703 * challenge ACK or a subsequent RST.
5705 static bool tcp_reset_check(const struct sock *sk, const struct sk_buff *skb)
5707 struct tcp_sock *tp = tcp_sk(sk);
5709 return unlikely(TCP_SKB_CB(skb)->seq == (tp->rcv_nxt - 1) &&
5710 (1 << sk->sk_state) & (TCPF_CLOSE_WAIT | TCPF_LAST_ACK |
5714 /* Does PAWS and seqno based validation of an incoming segment, flags will
5715 * play significant role here.
5717 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5718 const struct tcphdr *th, int syn_inerr)
5720 struct tcp_sock *tp = tcp_sk(sk);
5723 /* RFC1323: H1. Apply PAWS check first. */
5724 if (tcp_fast_parse_options(sock_net(sk), skb, th, tp) &&
5725 tp->rx_opt.saw_tstamp &&
5726 tcp_paws_discard(sk, skb)) {
5728 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5729 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5730 LINUX_MIB_TCPACKSKIPPEDPAWS,
5731 &tp->last_oow_ack_time))
5732 tcp_send_dupack(sk, skb);
5733 SKB_DR_SET(reason, TCP_RFC7323_PAWS);
5736 /* Reset is accepted even if it did not pass PAWS. */
5739 /* Step 1: check sequence number */
5740 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5741 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5742 * (RST) segments are validated by checking their SEQ-fields."
5743 * And page 69: "If an incoming segment is not acceptable,
5744 * an acknowledgment should be sent in reply (unless the RST
5745 * bit is set, if so drop the segment and return)".
5750 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5751 LINUX_MIB_TCPACKSKIPPEDSEQ,
5752 &tp->last_oow_ack_time))
5753 tcp_send_dupack(sk, skb);
5754 } else if (tcp_reset_check(sk, skb)) {
5757 SKB_DR_SET(reason, TCP_INVALID_SEQUENCE);
5761 /* Step 2: check RST bit */
5763 /* RFC 5961 3.2 (extend to match against (RCV.NXT - 1) after a
5764 * FIN and SACK too if available):
5765 * If seq num matches RCV.NXT or (RCV.NXT - 1) after a FIN, or
5766 * the right-most SACK block,
5768 * RESET the connection
5770 * Send a challenge ACK
5772 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt ||
5773 tcp_reset_check(sk, skb))
5776 if (tcp_is_sack(tp) && tp->rx_opt.num_sacks > 0) {
5777 struct tcp_sack_block *sp = &tp->selective_acks[0];
5778 int max_sack = sp[0].end_seq;
5781 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;
5783 max_sack = after(sp[this_sack].end_seq,
5785 sp[this_sack].end_seq : max_sack;
5788 if (TCP_SKB_CB(skb)->seq == max_sack)
5792 /* Disable TFO if RST is out-of-order
5793 * and no data has been received
5794 * for current active TFO socket
5796 if (tp->syn_fastopen && !tp->data_segs_in &&
5797 sk->sk_state == TCP_ESTABLISHED)
5798 tcp_fastopen_active_disable(sk);
5799 tcp_send_challenge_ack(sk);
5800 SKB_DR_SET(reason, TCP_RESET);
5804 /* step 3: check security and precedence [ignored] */
5806 /* step 4: Check for a SYN
5807 * RFC 5961 4.2 : Send a challenge ack
5812 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5813 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5814 tcp_send_challenge_ack(sk);
5815 SKB_DR_SET(reason, TCP_INVALID_SYN);
5819 bpf_skops_parse_hdr(sk, skb);
5824 tcp_drop_reason(sk, skb, reason);
5834 * TCP receive function for the ESTABLISHED state.
5836 * It is split into a fast path and a slow path. The fast path is
5838 * - A zero window was announced from us - zero window probing
5839 * is only handled properly in the slow path.
5840 * - Out of order segments arrived.
5841 * - Urgent data is expected.
5842 * - There is no buffer space left
5843 * - Unexpected TCP flags/window values/header lengths are received
5844 * (detected by checking the TCP header against pred_flags)
5845 * - Data is sent in both directions. Fast path only supports pure senders
5846 * or pure receivers (this means either the sequence number or the ack
5847 * value must stay constant)
5848 * - Unexpected TCP option.
5850 * When these conditions are not satisfied it drops into a standard
5851 * receive procedure patterned after RFC793 to handle all cases.
5852 * The first three cases are guaranteed by proper pred_flags setting,
5853 * the rest is checked inline. Fast processing is turned on in
5854 * tcp_data_queue when everything is OK.
5856 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb)
5858 enum skb_drop_reason reason = SKB_DROP_REASON_NOT_SPECIFIED;
5859 const struct tcphdr *th = (const struct tcphdr *)skb->data;
5860 struct tcp_sock *tp = tcp_sk(sk);
5861 unsigned int len = skb->len;
5863 /* TCP congestion window tracking */
5864 trace_tcp_probe(sk, skb);
5866 tcp_mstamp_refresh(tp);
5867 if (unlikely(!rcu_access_pointer(sk->sk_rx_dst)))
5868 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
5870 * Header prediction.
5871 * The code loosely follows the one in the famous
5872 * "30 instruction TCP receive" Van Jacobson mail.
5874 * Van's trick is to deposit buffers into socket queue
5875 * on a device interrupt, to call tcp_recv function
5876 * on the receive process context and checksum and copy
5877 * the buffer to user space. smart...
5879 * Our current scheme is not silly either but we take the
5880 * extra cost of the net_bh soft interrupt processing...
5881 * We do checksum and copy also but from device to kernel.
5884 tp->rx_opt.saw_tstamp = 0;
5886 /* pred_flags is 0xS?10 << 16 + snd_wnd
5887 * if header_prediction is to be made
5888 * 'S' will always be tp->tcp_header_len >> 2
5889 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5890 * turn it off (when there are holes in the receive
5891 * space for instance)
5892 * PSH flag is ignored.
5895 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5896 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5897 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5898 int tcp_header_len = tp->tcp_header_len;
5900 /* Timestamp header prediction: tcp_header_len
5901 * is automatically equal to th->doff*4 due to pred_flags
5905 /* Check timestamp */
5906 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5907 /* No? Slow path! */
5908 if (!tcp_parse_aligned_timestamp(tp, th))
5911 /* If PAWS failed, check it more carefully in slow path */
5912 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5915 /* DO NOT update ts_recent here, if checksum fails
5916 * and timestamp was corrupted part, it will result
5917 * in a hung connection since we will drop all
5918 * future packets due to the PAWS test.
5922 if (len <= tcp_header_len) {
5923 /* Bulk data transfer: sender */
5924 if (len == tcp_header_len) {
5925 /* Predicted packet is in window by definition.
5926 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5927 * Hence, check seq<=rcv_wup reduces to:
5929 if (tcp_header_len ==
5930 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5931 tp->rcv_nxt == tp->rcv_wup)
5932 tcp_store_ts_recent(tp);
5934 /* We know that such packets are checksummed
5937 tcp_ack(sk, skb, 0);
5939 tcp_data_snd_check(sk);
5940 /* When receiving pure ack in fast path, update
5941 * last ts ecr directly instead of calling
5942 * tcp_rcv_rtt_measure_ts()
5944 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr;
5946 } else { /* Header too small */
5947 reason = SKB_DROP_REASON_PKT_TOO_SMALL;
5948 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5953 bool fragstolen = false;
5955 if (tcp_checksum_complete(skb))
5958 if ((int)skb->truesize > sk->sk_forward_alloc)
5961 /* Predicted packet is in window by definition.
5962 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5963 * Hence, check seq<=rcv_wup reduces to:
5965 if (tcp_header_len ==
5966 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5967 tp->rcv_nxt == tp->rcv_wup)
5968 tcp_store_ts_recent(tp);
5970 tcp_rcv_rtt_measure_ts(sk, skb);
5972 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPHITS);
5974 /* Bulk data transfer: receiver */
5976 __skb_pull(skb, tcp_header_len);
5977 eaten = tcp_queue_rcv(sk, skb, &fragstolen);
5979 tcp_event_data_recv(sk, skb);
5981 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5982 /* Well, only one small jumplet in fast path... */
5983 tcp_ack(sk, skb, FLAG_DATA);
5984 tcp_data_snd_check(sk);
5985 if (!inet_csk_ack_scheduled(sk))
5988 tcp_update_wl(tp, TCP_SKB_CB(skb)->seq);
5991 __tcp_ack_snd_check(sk, 0);
5994 kfree_skb_partial(skb, fragstolen);
6001 if (len < (th->doff << 2) || tcp_checksum_complete(skb))
6004 if (!th->ack && !th->rst && !th->syn) {
6005 reason = SKB_DROP_REASON_TCP_FLAGS;
6010 * Standard slow path.
6013 if (!tcp_validate_incoming(sk, skb, th, 1))
6017 reason = tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT);
6018 if ((int)reason < 0) {
6022 tcp_rcv_rtt_measure_ts(sk, skb);
6024 /* Process urgent data. */
6025 tcp_urg(sk, skb, th);
6027 /* step 7: process the segment text */
6028 tcp_data_queue(sk, skb);
6030 tcp_data_snd_check(sk);
6031 tcp_ack_snd_check(sk);
6035 reason = SKB_DROP_REASON_TCP_CSUM;
6036 trace_tcp_bad_csum(skb);
6037 TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS);
6038 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
6041 tcp_drop_reason(sk, skb, reason);
6043 EXPORT_SYMBOL(tcp_rcv_established);
6045 void tcp_init_transfer(struct sock *sk, int bpf_op, struct sk_buff *skb)
6047 struct inet_connection_sock *icsk = inet_csk(sk);
6048 struct tcp_sock *tp = tcp_sk(sk);
6051 icsk->icsk_af_ops->rebuild_header(sk);
6052 tcp_init_metrics(sk);
6054 /* Initialize the congestion window to start the transfer.
6055 * Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
6056 * retransmitted. In light of RFC6298 more aggressive 1sec
6057 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
6058 * retransmission has occurred.
6060 if (tp->total_retrans > 1 && tp->undo_marker)
6061 tcp_snd_cwnd_set(tp, 1);
6063 tcp_snd_cwnd_set(tp, tcp_init_cwnd(tp, __sk_dst_get(sk)));
6064 tp->snd_cwnd_stamp = tcp_jiffies32;
6066 bpf_skops_established(sk, bpf_op, skb);
6067 /* Initialize congestion control unless BPF initialized it already: */
6068 if (!icsk->icsk_ca_initialized)
6069 tcp_init_congestion_control(sk);
6070 tcp_init_buffer_space(sk);
6073 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
6075 struct tcp_sock *tp = tcp_sk(sk);
6076 struct inet_connection_sock *icsk = inet_csk(sk);
6078 tcp_set_state(sk, TCP_ESTABLISHED);
6079 icsk->icsk_ack.lrcvtime = tcp_jiffies32;
6082 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
6083 security_inet_conn_established(sk, skb);
6084 sk_mark_napi_id(sk, skb);
6087 tcp_init_transfer(sk, BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB, skb);
6089 /* Prevent spurious tcp_cwnd_restart() on first data
6092 tp->lsndtime = tcp_jiffies32;
6094 if (sock_flag(sk, SOCK_KEEPOPEN))
6095 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
6097 if (!tp->rx_opt.snd_wscale)
6098 __tcp_fast_path_on(tp, tp->snd_wnd);
6103 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
6104 struct tcp_fastopen_cookie *cookie)
6106 struct tcp_sock *tp = tcp_sk(sk);
6107 struct sk_buff *data = tp->syn_data ? tcp_rtx_queue_head(sk) : NULL;
6108 u16 mss = tp->rx_opt.mss_clamp, try_exp = 0;
6109 bool syn_drop = false;
6111 if (mss == tp->rx_opt.user_mss) {
6112 struct tcp_options_received opt;
6114 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
6115 tcp_clear_options(&opt);
6116 opt.user_mss = opt.mss_clamp = 0;
6117 tcp_parse_options(sock_net(sk), synack, &opt, 0, NULL);
6118 mss = opt.mss_clamp;
6121 if (!tp->syn_fastopen) {
6122 /* Ignore an unsolicited cookie */
6124 } else if (tp->total_retrans) {
6125 /* SYN timed out and the SYN-ACK neither has a cookie nor
6126 * acknowledges data. Presumably the remote received only
6127 * the retransmitted (regular) SYNs: either the original
6128 * SYN-data or the corresponding SYN-ACK was dropped.
6130 syn_drop = (cookie->len < 0 && data);
6131 } else if (cookie->len < 0 && !tp->syn_data) {
6132 /* We requested a cookie but didn't get it. If we did not use
6133 * the (old) exp opt format then try so next time (try_exp=1).
6134 * Otherwise we go back to use the RFC7413 opt (try_exp=2).
6136 try_exp = tp->syn_fastopen_exp ? 2 : 1;
6139 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp);
6141 if (data) { /* Retransmit unacked data in SYN */
6142 if (tp->total_retrans)
6143 tp->fastopen_client_fail = TFO_SYN_RETRANSMITTED;
6145 tp->fastopen_client_fail = TFO_DATA_NOT_ACKED;
6146 skb_rbtree_walk_from(data)
6147 tcp_mark_skb_lost(sk, data);
6148 tcp_xmit_retransmit_queue(sk);
6149 NET_INC_STATS(sock_net(sk),
6150 LINUX_MIB_TCPFASTOPENACTIVEFAIL);
6153 tp->syn_data_acked = tp->syn_data;
6154 if (tp->syn_data_acked) {
6155 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
6156 /* SYN-data is counted as two separate packets in tcp_ack() */
6157 if (tp->delivered > 1)
6161 tcp_fastopen_add_skb(sk, synack);
6166 static void smc_check_reset_syn(struct tcp_sock *tp)
6168 #if IS_ENABLED(CONFIG_SMC)
6169 if (static_branch_unlikely(&tcp_have_smc)) {
6170 if (tp->syn_smc && !tp->rx_opt.smc_ok)
6176 static void tcp_try_undo_spurious_syn(struct sock *sk)
6178 struct tcp_sock *tp = tcp_sk(sk);
6181 /* undo_marker is set when SYN or SYNACK times out. The timeout is
6182 * spurious if the ACK's timestamp option echo value matches the
6183 * original SYN timestamp.
6185 syn_stamp = tp->retrans_stamp;
6186 if (tp->undo_marker && syn_stamp && tp->rx_opt.saw_tstamp &&
6187 syn_stamp == tp->rx_opt.rcv_tsecr)
6188 tp->undo_marker = 0;
6191 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
6192 const struct tcphdr *th)
6194 struct inet_connection_sock *icsk = inet_csk(sk);
6195 struct tcp_sock *tp = tcp_sk(sk);
6196 struct tcp_fastopen_cookie foc = { .len = -1 };
6197 int saved_clamp = tp->rx_opt.mss_clamp;
6201 tcp_parse_options(sock_net(sk), skb, &tp->rx_opt, 0, &foc);
6202 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
6203 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
6207 * "If the state is SYN-SENT then
6208 * first check the ACK bit
6209 * If the ACK bit is set
6210 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
6211 * a reset (unless the RST bit is set, if so drop
6212 * the segment and return)"
6214 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
6215 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
6216 /* Previous FIN/ACK or RST/ACK might be ignored. */
6217 if (icsk->icsk_retransmits == 0)
6218 inet_csk_reset_xmit_timer(sk,
6220 TCP_TIMEOUT_MIN, TCP_RTO_MAX);
6221 goto reset_and_undo;
6224 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
6225 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
6226 tcp_time_stamp(tp))) {
6227 NET_INC_STATS(sock_net(sk),
6228 LINUX_MIB_PAWSACTIVEREJECTED);
6229 goto reset_and_undo;
6232 /* Now ACK is acceptable.
6234 * "If the RST bit is set
6235 * If the ACK was acceptable then signal the user "error:
6236 * connection reset", drop the segment, enter CLOSED state,
6237 * delete TCB, and return."
6248 * "fifth, if neither of the SYN or RST bits is set then
6249 * drop the segment and return."
6255 SKB_DR_SET(reason, TCP_FLAGS);
6256 goto discard_and_undo;
6259 * "If the SYN bit is on ...
6260 * are acceptable then ...
6261 * (our SYN has been ACKed), change the connection
6262 * state to ESTABLISHED..."
6265 tcp_ecn_rcv_synack(tp, th);
6267 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
6268 tcp_try_undo_spurious_syn(sk);
6269 tcp_ack(sk, skb, FLAG_SLOWPATH);
6271 /* Ok.. it's good. Set up sequence numbers and
6272 * move to established.
6274 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1);
6275 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
6277 /* RFC1323: The window in SYN & SYN/ACK segments is
6280 tp->snd_wnd = ntohs(th->window);
6282 if (!tp->rx_opt.wscale_ok) {
6283 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
6284 tp->window_clamp = min(tp->window_clamp, 65535U);
6287 if (tp->rx_opt.saw_tstamp) {
6288 tp->rx_opt.tstamp_ok = 1;
6289 tp->tcp_header_len =
6290 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
6291 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
6292 tcp_store_ts_recent(tp);
6294 tp->tcp_header_len = sizeof(struct tcphdr);
6297 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
6298 tcp_initialize_rcv_mss(sk);
6300 /* Remember, tcp_poll() does not lock socket!
6301 * Change state from SYN-SENT only after copied_seq
6302 * is initialized. */
6303 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6305 smc_check_reset_syn(tp);
6309 tcp_finish_connect(sk, skb);
6311 fastopen_fail = (tp->syn_fastopen || tp->syn_data) &&
6312 tcp_rcv_fastopen_synack(sk, skb, &foc);
6314 if (!sock_flag(sk, SOCK_DEAD)) {
6315 sk->sk_state_change(sk);
6316 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
6320 if (sk->sk_write_pending ||
6321 icsk->icsk_accept_queue.rskq_defer_accept ||
6322 inet_csk_in_pingpong_mode(sk)) {
6323 /* Save one ACK. Data will be ready after
6324 * several ticks, if write_pending is set.
6326 * It may be deleted, but with this feature tcpdumps
6327 * look so _wonderfully_ clever, that I was not able
6328 * to stand against the temptation 8) --ANK
6330 inet_csk_schedule_ack(sk);
6331 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
6332 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
6333 TCP_DELACK_MAX, TCP_RTO_MAX);
6340 /* No ACK in the segment */
6344 * "If the RST bit is set
6346 * Otherwise (no ACK) drop the segment and return."
6348 SKB_DR_SET(reason, TCP_RESET);
6349 goto discard_and_undo;
6353 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
6354 tcp_paws_reject(&tp->rx_opt, 0)) {
6355 SKB_DR_SET(reason, TCP_RFC7323_PAWS);
6356 goto discard_and_undo;
6359 /* We see SYN without ACK. It is attempt of
6360 * simultaneous connect with crossed SYNs.
6361 * Particularly, it can be connect to self.
6363 tcp_set_state(sk, TCP_SYN_RECV);
6365 if (tp->rx_opt.saw_tstamp) {
6366 tp->rx_opt.tstamp_ok = 1;
6367 tcp_store_ts_recent(tp);
6368 tp->tcp_header_len =
6369 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
6371 tp->tcp_header_len = sizeof(struct tcphdr);
6374 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1);
6375 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6376 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
6378 /* RFC1323: The window in SYN & SYN/ACK segments is
6381 tp->snd_wnd = ntohs(th->window);
6382 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
6383 tp->max_window = tp->snd_wnd;
6385 tcp_ecn_rcv_syn(tp, th);
6388 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
6389 tcp_initialize_rcv_mss(sk);
6391 tcp_send_synack(sk);
6393 /* Note, we could accept data and URG from this segment.
6394 * There are no obstacles to make this (except that we must
6395 * either change tcp_recvmsg() to prevent it from returning data
6396 * before 3WHS completes per RFC793, or employ TCP Fast Open).
6398 * However, if we ignore data in ACKless segments sometimes,
6399 * we have no reasons to accept it sometimes.
6400 * Also, seems the code doing it in step6 of tcp_rcv_state_process
6401 * is not flawless. So, discard packet for sanity.
6402 * Uncomment this return to process the data.
6409 /* "fifth, if neither of the SYN or RST bits is set then
6410 * drop the segment and return."
6414 tcp_clear_options(&tp->rx_opt);
6415 tp->rx_opt.mss_clamp = saved_clamp;
6416 tcp_drop_reason(sk, skb, reason);
6420 tcp_clear_options(&tp->rx_opt);
6421 tp->rx_opt.mss_clamp = saved_clamp;
6425 static void tcp_rcv_synrecv_state_fastopen(struct sock *sk)
6427 struct tcp_sock *tp = tcp_sk(sk);
6428 struct request_sock *req;
6430 /* If we are still handling the SYNACK RTO, see if timestamp ECR allows
6431 * undo. If peer SACKs triggered fast recovery, we can't undo here.
6433 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss && !tp->packets_out)
6434 tcp_try_undo_recovery(sk);
6436 /* Reset rtx states to prevent spurious retransmits_timed_out() */
6437 tp->retrans_stamp = 0;
6438 inet_csk(sk)->icsk_retransmits = 0;
6440 /* Once we leave TCP_SYN_RECV or TCP_FIN_WAIT_1,
6441 * we no longer need req so release it.
6443 req = rcu_dereference_protected(tp->fastopen_rsk,
6444 lockdep_sock_is_held(sk));
6445 reqsk_fastopen_remove(sk, req, false);
6447 /* Re-arm the timer because data may have been sent out.
6448 * This is similar to the regular data transmission case
6449 * when new data has just been ack'ed.
6451 * (TFO) - we could try to be more aggressive and
6452 * retransmitting any data sooner based on when they
6459 * This function implements the receiving procedure of RFC 793 for
6460 * all states except ESTABLISHED and TIME_WAIT.
6461 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
6462 * address independent.
6465 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb)
6467 struct tcp_sock *tp = tcp_sk(sk);
6468 struct inet_connection_sock *icsk = inet_csk(sk);
6469 const struct tcphdr *th = tcp_hdr(skb);
6470 struct request_sock *req;
6475 switch (sk->sk_state) {
6477 SKB_DR_SET(reason, TCP_CLOSE);
6485 SKB_DR_SET(reason, TCP_RESET);
6490 SKB_DR_SET(reason, TCP_FLAGS);
6493 /* It is possible that we process SYN packets from backlog,
6494 * so we need to make sure to disable BH and RCU right there.
6498 acceptable = icsk->icsk_af_ops->conn_request(sk, skb) >= 0;
6507 SKB_DR_SET(reason, TCP_FLAGS);
6511 tp->rx_opt.saw_tstamp = 0;
6512 tcp_mstamp_refresh(tp);
6513 queued = tcp_rcv_synsent_state_process(sk, skb, th);
6517 /* Do step6 onward by hand. */
6518 tcp_urg(sk, skb, th);
6520 tcp_data_snd_check(sk);
6524 tcp_mstamp_refresh(tp);
6525 tp->rx_opt.saw_tstamp = 0;
6526 req = rcu_dereference_protected(tp->fastopen_rsk,
6527 lockdep_sock_is_held(sk));
6531 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
6532 sk->sk_state != TCP_FIN_WAIT1);
6534 if (!tcp_check_req(sk, skb, req, true, &req_stolen)) {
6535 SKB_DR_SET(reason, TCP_FASTOPEN);
6540 if (!th->ack && !th->rst && !th->syn) {
6541 SKB_DR_SET(reason, TCP_FLAGS);
6544 if (!tcp_validate_incoming(sk, skb, th, 0))
6547 /* step 5: check the ACK field */
6548 acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
6549 FLAG_UPDATE_TS_RECENT |
6550 FLAG_NO_CHALLENGE_ACK) > 0;
6553 if (sk->sk_state == TCP_SYN_RECV)
6554 return 1; /* send one RST */
6555 tcp_send_challenge_ack(sk);
6556 SKB_DR_SET(reason, TCP_OLD_ACK);
6559 switch (sk->sk_state) {
6561 tp->delivered++; /* SYN-ACK delivery isn't tracked in tcp_ack */
6563 tcp_synack_rtt_meas(sk, req);
6566 tcp_rcv_synrecv_state_fastopen(sk);
6568 tcp_try_undo_spurious_syn(sk);
6569 tp->retrans_stamp = 0;
6570 tcp_init_transfer(sk, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB,
6572 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6575 tcp_set_state(sk, TCP_ESTABLISHED);
6576 sk->sk_state_change(sk);
6578 /* Note, that this wakeup is only for marginal crossed SYN case.
6579 * Passively open sockets are not waked up, because
6580 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
6583 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
6585 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
6586 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
6587 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
6589 if (tp->rx_opt.tstamp_ok)
6590 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
6592 if (!inet_csk(sk)->icsk_ca_ops->cong_control)
6593 tcp_update_pacing_rate(sk);
6595 /* Prevent spurious tcp_cwnd_restart() on first data packet */
6596 tp->lsndtime = tcp_jiffies32;
6598 tcp_initialize_rcv_mss(sk);
6599 tcp_fast_path_on(tp);
6602 case TCP_FIN_WAIT1: {
6606 tcp_rcv_synrecv_state_fastopen(sk);
6608 if (tp->snd_una != tp->write_seq)
6611 tcp_set_state(sk, TCP_FIN_WAIT2);
6612 WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | SEND_SHUTDOWN);
6616 if (!sock_flag(sk, SOCK_DEAD)) {
6617 /* Wake up lingering close() */
6618 sk->sk_state_change(sk);
6622 if (tp->linger2 < 0) {
6624 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6627 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6628 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6629 /* Receive out of order FIN after close() */
6630 if (tp->syn_fastopen && th->fin)
6631 tcp_fastopen_active_disable(sk);
6633 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6637 tmo = tcp_fin_time(sk);
6638 if (tmo > TCP_TIMEWAIT_LEN) {
6639 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
6640 } else if (th->fin || sock_owned_by_user(sk)) {
6641 /* Bad case. We could lose such FIN otherwise.
6642 * It is not a big problem, but it looks confusing
6643 * and not so rare event. We still can lose it now,
6644 * if it spins in bh_lock_sock(), but it is really
6647 inet_csk_reset_keepalive_timer(sk, tmo);
6649 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
6656 if (tp->snd_una == tp->write_seq) {
6657 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
6663 if (tp->snd_una == tp->write_seq) {
6664 tcp_update_metrics(sk);
6671 /* step 6: check the URG bit */
6672 tcp_urg(sk, skb, th);
6674 /* step 7: process the segment text */
6675 switch (sk->sk_state) {
6676 case TCP_CLOSE_WAIT:
6679 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
6680 /* If a subflow has been reset, the packet should not
6681 * continue to be processed, drop the packet.
6683 if (sk_is_mptcp(sk) && !mptcp_incoming_options(sk, skb))
6690 /* RFC 793 says to queue data in these states,
6691 * RFC 1122 says we MUST send a reset.
6692 * BSD 4.4 also does reset.
6694 if (sk->sk_shutdown & RCV_SHUTDOWN) {
6695 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6696 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6697 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6703 case TCP_ESTABLISHED:
6704 tcp_data_queue(sk, skb);
6709 /* tcp_data could move socket to TIME-WAIT */
6710 if (sk->sk_state != TCP_CLOSE) {
6711 tcp_data_snd_check(sk);
6712 tcp_ack_snd_check(sk);
6717 tcp_drop_reason(sk, skb, reason);
6725 EXPORT_SYMBOL(tcp_rcv_state_process);
6727 static inline void pr_drop_req(struct request_sock *req, __u16 port, int family)
6729 struct inet_request_sock *ireq = inet_rsk(req);
6731 if (family == AF_INET)
6732 net_dbg_ratelimited("drop open request from %pI4/%u\n",
6733 &ireq->ir_rmt_addr, port);
6734 #if IS_ENABLED(CONFIG_IPV6)
6735 else if (family == AF_INET6)
6736 net_dbg_ratelimited("drop open request from %pI6/%u\n",
6737 &ireq->ir_v6_rmt_addr, port);
6741 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
6743 * If we receive a SYN packet with these bits set, it means a
6744 * network is playing bad games with TOS bits. In order to
6745 * avoid possible false congestion notifications, we disable
6746 * TCP ECN negotiation.
6748 * Exception: tcp_ca wants ECN. This is required for DCTCP
6749 * congestion control: Linux DCTCP asserts ECT on all packets,
6750 * including SYN, which is most optimal solution; however,
6751 * others, such as FreeBSD do not.
6753 * Exception: At least one of the reserved bits of the TCP header (th->res1) is
6754 * set, indicating the use of a future TCP extension (such as AccECN). See
6755 * RFC8311 §4.3 which updates RFC3168 to allow the development of such
6758 static void tcp_ecn_create_request(struct request_sock *req,
6759 const struct sk_buff *skb,
6760 const struct sock *listen_sk,
6761 const struct dst_entry *dst)
6763 const struct tcphdr *th = tcp_hdr(skb);
6764 const struct net *net = sock_net(listen_sk);
6765 bool th_ecn = th->ece && th->cwr;
6772 ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield);
6773 ecn_ok_dst = dst_feature(dst, DST_FEATURE_ECN_MASK);
6774 ecn_ok = READ_ONCE(net->ipv4.sysctl_tcp_ecn) || ecn_ok_dst;
6776 if (((!ect || th->res1) && ecn_ok) || tcp_ca_needs_ecn(listen_sk) ||
6777 (ecn_ok_dst & DST_FEATURE_ECN_CA) ||
6778 tcp_bpf_ca_needs_ecn((struct sock *)req))
6779 inet_rsk(req)->ecn_ok = 1;
6782 static void tcp_openreq_init(struct request_sock *req,
6783 const struct tcp_options_received *rx_opt,
6784 struct sk_buff *skb, const struct sock *sk)
6786 struct inet_request_sock *ireq = inet_rsk(req);
6788 req->rsk_rcv_wnd = 0; /* So that tcp_send_synack() knows! */
6789 tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq;
6790 tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
6791 tcp_rsk(req)->snt_synack = 0;
6792 tcp_rsk(req)->last_oow_ack_time = 0;
6793 req->mss = rx_opt->mss_clamp;
6794 req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0;
6795 ireq->tstamp_ok = rx_opt->tstamp_ok;
6796 ireq->sack_ok = rx_opt->sack_ok;
6797 ireq->snd_wscale = rx_opt->snd_wscale;
6798 ireq->wscale_ok = rx_opt->wscale_ok;
6801 ireq->ir_rmt_port = tcp_hdr(skb)->source;
6802 ireq->ir_num = ntohs(tcp_hdr(skb)->dest);
6803 ireq->ir_mark = inet_request_mark(sk, skb);
6804 #if IS_ENABLED(CONFIG_SMC)
6805 ireq->smc_ok = rx_opt->smc_ok && !(tcp_sk(sk)->smc_hs_congested &&
6806 tcp_sk(sk)->smc_hs_congested(sk));
6810 struct request_sock *inet_reqsk_alloc(const struct request_sock_ops *ops,
6811 struct sock *sk_listener,
6812 bool attach_listener)
6814 struct request_sock *req = reqsk_alloc(ops, sk_listener,
6818 struct inet_request_sock *ireq = inet_rsk(req);
6820 ireq->ireq_opt = NULL;
6821 #if IS_ENABLED(CONFIG_IPV6)
6822 ireq->pktopts = NULL;
6824 atomic64_set(&ireq->ir_cookie, 0);
6825 ireq->ireq_state = TCP_NEW_SYN_RECV;
6826 write_pnet(&ireq->ireq_net, sock_net(sk_listener));
6827 ireq->ireq_family = sk_listener->sk_family;
6828 req->timeout = TCP_TIMEOUT_INIT;
6833 EXPORT_SYMBOL(inet_reqsk_alloc);
6836 * Return true if a syncookie should be sent
6838 static bool tcp_syn_flood_action(const struct sock *sk, const char *proto)
6840 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
6841 const char *msg = "Dropping request";
6842 struct net *net = sock_net(sk);
6843 bool want_cookie = false;
6846 syncookies = READ_ONCE(net->ipv4.sysctl_tcp_syncookies);
6848 #ifdef CONFIG_SYN_COOKIES
6850 msg = "Sending cookies";
6852 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES);
6855 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP);
6857 if (!queue->synflood_warned && syncookies != 2 &&
6858 xchg(&queue->synflood_warned, 1) == 0)
6859 net_info_ratelimited("%s: Possible SYN flooding on port %d. %s. Check SNMP counters.\n",
6860 proto, sk->sk_num, msg);
6865 static void tcp_reqsk_record_syn(const struct sock *sk,
6866 struct request_sock *req,
6867 const struct sk_buff *skb)
6869 if (tcp_sk(sk)->save_syn) {
6870 u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb);
6871 struct saved_syn *saved_syn;
6875 if (tcp_sk(sk)->save_syn == 2) { /* Save full header. */
6876 base = skb_mac_header(skb);
6877 mac_hdrlen = skb_mac_header_len(skb);
6880 base = skb_network_header(skb);
6884 saved_syn = kmalloc(struct_size(saved_syn, data, len),
6887 saved_syn->mac_hdrlen = mac_hdrlen;
6888 saved_syn->network_hdrlen = skb_network_header_len(skb);
6889 saved_syn->tcp_hdrlen = tcp_hdrlen(skb);
6890 memcpy(saved_syn->data, base, len);
6891 req->saved_syn = saved_syn;
6896 /* If a SYN cookie is required and supported, returns a clamped MSS value to be
6897 * used for SYN cookie generation.
6899 u16 tcp_get_syncookie_mss(struct request_sock_ops *rsk_ops,
6900 const struct tcp_request_sock_ops *af_ops,
6901 struct sock *sk, struct tcphdr *th)
6903 struct tcp_sock *tp = tcp_sk(sk);
6906 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_syncookies) != 2 &&
6907 !inet_csk_reqsk_queue_is_full(sk))
6910 if (!tcp_syn_flood_action(sk, rsk_ops->slab_name))
6913 if (sk_acceptq_is_full(sk)) {
6914 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
6918 mss = tcp_parse_mss_option(th, tp->rx_opt.user_mss);
6920 mss = af_ops->mss_clamp;
6924 EXPORT_SYMBOL_GPL(tcp_get_syncookie_mss);
6926 int tcp_conn_request(struct request_sock_ops *rsk_ops,
6927 const struct tcp_request_sock_ops *af_ops,
6928 struct sock *sk, struct sk_buff *skb)
6930 struct tcp_fastopen_cookie foc = { .len = -1 };
6931 __u32 isn = TCP_SKB_CB(skb)->tcp_tw_isn;
6932 struct tcp_options_received tmp_opt;
6933 struct tcp_sock *tp = tcp_sk(sk);
6934 struct net *net = sock_net(sk);
6935 struct sock *fastopen_sk = NULL;
6936 struct request_sock *req;
6937 bool want_cookie = false;
6938 struct dst_entry *dst;
6942 syncookies = READ_ONCE(net->ipv4.sysctl_tcp_syncookies);
6944 /* TW buckets are converted to open requests without
6945 * limitations, they conserve resources and peer is
6946 * evidently real one.
6948 if ((syncookies == 2 || inet_csk_reqsk_queue_is_full(sk)) && !isn) {
6949 want_cookie = tcp_syn_flood_action(sk, rsk_ops->slab_name);
6954 if (sk_acceptq_is_full(sk)) {
6955 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
6959 req = inet_reqsk_alloc(rsk_ops, sk, !want_cookie);
6963 req->syncookie = want_cookie;
6964 tcp_rsk(req)->af_specific = af_ops;
6965 tcp_rsk(req)->ts_off = 0;
6966 #if IS_ENABLED(CONFIG_MPTCP)
6967 tcp_rsk(req)->is_mptcp = 0;
6970 tcp_clear_options(&tmp_opt);
6971 tmp_opt.mss_clamp = af_ops->mss_clamp;
6972 tmp_opt.user_mss = tp->rx_opt.user_mss;
6973 tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0,
6974 want_cookie ? NULL : &foc);
6976 if (want_cookie && !tmp_opt.saw_tstamp)
6977 tcp_clear_options(&tmp_opt);
6979 if (IS_ENABLED(CONFIG_SMC) && want_cookie)
6982 tmp_opt.tstamp_ok = tmp_opt.saw_tstamp;
6983 tcp_openreq_init(req, &tmp_opt, skb, sk);
6984 inet_rsk(req)->no_srccheck = inet_sk(sk)->transparent;
6986 /* Note: tcp_v6_init_req() might override ir_iif for link locals */
6987 inet_rsk(req)->ir_iif = inet_request_bound_dev_if(sk, skb);
6989 dst = af_ops->route_req(sk, skb, &fl, req);
6993 if (tmp_opt.tstamp_ok)
6994 tcp_rsk(req)->ts_off = af_ops->init_ts_off(net, skb);
6996 if (!want_cookie && !isn) {
6997 int max_syn_backlog = READ_ONCE(net->ipv4.sysctl_max_syn_backlog);
6999 /* Kill the following clause, if you dislike this way. */
7001 (max_syn_backlog - inet_csk_reqsk_queue_len(sk) <
7002 (max_syn_backlog >> 2)) &&
7003 !tcp_peer_is_proven(req, dst)) {
7004 /* Without syncookies last quarter of
7005 * backlog is filled with destinations,
7006 * proven to be alive.
7007 * It means that we continue to communicate
7008 * to destinations, already remembered
7009 * to the moment of synflood.
7011 pr_drop_req(req, ntohs(tcp_hdr(skb)->source),
7013 goto drop_and_release;
7016 isn = af_ops->init_seq(skb);
7019 tcp_ecn_create_request(req, skb, sk, dst);
7022 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss);
7023 if (!tmp_opt.tstamp_ok)
7024 inet_rsk(req)->ecn_ok = 0;
7027 tcp_rsk(req)->snt_isn = isn;
7028 tcp_rsk(req)->txhash = net_tx_rndhash();
7029 tcp_rsk(req)->syn_tos = TCP_SKB_CB(skb)->ip_dsfield;
7030 tcp_openreq_init_rwin(req, sk, dst);
7031 sk_rx_queue_set(req_to_sk(req), skb);
7033 tcp_reqsk_record_syn(sk, req, skb);
7034 fastopen_sk = tcp_try_fastopen(sk, skb, req, &foc, dst);
7037 af_ops->send_synack(fastopen_sk, dst, &fl, req,
7038 &foc, TCP_SYNACK_FASTOPEN, skb);
7039 /* Add the child socket directly into the accept queue */
7040 if (!inet_csk_reqsk_queue_add(sk, req, fastopen_sk)) {
7041 reqsk_fastopen_remove(fastopen_sk, req, false);
7042 bh_unlock_sock(fastopen_sk);
7043 sock_put(fastopen_sk);
7046 sk->sk_data_ready(sk);
7047 bh_unlock_sock(fastopen_sk);
7048 sock_put(fastopen_sk);
7050 tcp_rsk(req)->tfo_listener = false;
7052 req->timeout = tcp_timeout_init((struct sock *)req);
7053 inet_csk_reqsk_queue_hash_add(sk, req, req->timeout);
7055 af_ops->send_synack(sk, dst, &fl, req, &foc,
7056 !want_cookie ? TCP_SYNACK_NORMAL :
7075 EXPORT_SYMBOL(tcp_conn_request);