1 // SPDX-License-Identifier: GPL-2.0
3 * INET An implementation of the TCP/IP protocol suite for the LINUX
4 * operating system. INET is implemented using the BSD Socket
5 * interface as the means of communication with the user level.
7 * Implementation of the Transmission Control Protocol(TCP).
10 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
11 * Mark Evans, <evansmp@uhura.aston.ac.uk>
12 * Corey Minyard <wf-rch!minyard@relay.EU.net>
13 * Florian La Roche, <flla@stud.uni-sb.de>
14 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
15 * Linus Torvalds, <torvalds@cs.helsinki.fi>
16 * Alan Cox, <gw4pts@gw4pts.ampr.org>
17 * Matthew Dillon, <dillon@apollo.west.oic.com>
18 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
19 * Jorge Cwik, <jorge@laser.satlink.net>
24 * Pedro Roque : Fast Retransmit/Recovery.
26 * Retransmit queue handled by TCP.
27 * Better retransmit timer handling.
28 * New congestion avoidance.
32 * Eric : Fast Retransmit.
33 * Randy Scott : MSS option defines.
34 * Eric Schenk : Fixes to slow start algorithm.
35 * Eric Schenk : Yet another double ACK bug.
36 * Eric Schenk : Delayed ACK bug fixes.
37 * Eric Schenk : Floyd style fast retrans war avoidance.
38 * David S. Miller : Don't allow zero congestion window.
39 * Eric Schenk : Fix retransmitter so that it sends
40 * next packet on ack of previous packet.
41 * Andi Kleen : Moved open_request checking here
42 * and process RSTs for open_requests.
43 * Andi Kleen : Better prune_queue, and other fixes.
44 * Andrey Savochkin: Fix RTT measurements in the presence of
46 * Andrey Savochkin: Check sequence numbers correctly when
47 * removing SACKs due to in sequence incoming
49 * Andi Kleen: Make sure we never ack data there is not
50 * enough room for. Also make this condition
51 * a fatal error if it might still happen.
52 * Andi Kleen: Add tcp_measure_rcv_mss to make
53 * connections with MSS<min(MTU,ann. MSS)
54 * work without delayed acks.
55 * Andi Kleen: Process packets with PSH set in the
57 * J Hadi Salim: ECN support
60 * Panu Kuhlberg: Experimental audit of TCP (re)transmission
61 * engine. Lots of bugs are found.
62 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs
65 #define pr_fmt(fmt) "TCP: " fmt
68 #include <linux/slab.h>
69 #include <linux/module.h>
70 #include <linux/sysctl.h>
71 #include <linux/kernel.h>
72 #include <linux/prefetch.h>
75 #include <net/inet_common.h>
76 #include <linux/ipsec.h>
77 #include <asm/unaligned.h>
78 #include <linux/errqueue.h>
79 #include <trace/events/tcp.h>
80 #include <linux/jump_label_ratelimit.h>
81 #include <net/busy_poll.h>
82 #include <net/mptcp.h>
84 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
86 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
87 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
88 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
89 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
90 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
91 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
92 #define FLAG_ECE 0x40 /* ECE in this ACK */
93 #define FLAG_LOST_RETRANS 0x80 /* This ACK marks some retransmission lost */
94 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
95 #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
96 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
97 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
98 #define FLAG_SET_XMIT_TIMER 0x1000 /* Set TLP or RTO timer */
99 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
100 #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
101 #define FLAG_NO_CHALLENGE_ACK 0x8000 /* do not call tcp_send_challenge_ack() */
102 #define FLAG_ACK_MAYBE_DELAYED 0x10000 /* Likely a delayed ACK */
104 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
105 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
106 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE|FLAG_DSACKING_ACK)
107 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
109 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
110 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
112 #define REXMIT_NONE 0 /* no loss recovery to do */
113 #define REXMIT_LOST 1 /* retransmit packets marked lost */
114 #define REXMIT_NEW 2 /* FRTO-style transmit of unsent/new packets */
116 #if IS_ENABLED(CONFIG_TLS_DEVICE)
117 static DEFINE_STATIC_KEY_DEFERRED_FALSE(clean_acked_data_enabled, HZ);
119 void clean_acked_data_enable(struct inet_connection_sock *icsk,
120 void (*cad)(struct sock *sk, u32 ack_seq))
122 icsk->icsk_clean_acked = cad;
123 static_branch_deferred_inc(&clean_acked_data_enabled);
125 EXPORT_SYMBOL_GPL(clean_acked_data_enable);
127 void clean_acked_data_disable(struct inet_connection_sock *icsk)
129 static_branch_slow_dec_deferred(&clean_acked_data_enabled);
130 icsk->icsk_clean_acked = NULL;
132 EXPORT_SYMBOL_GPL(clean_acked_data_disable);
134 void clean_acked_data_flush(void)
136 static_key_deferred_flush(&clean_acked_data_enabled);
138 EXPORT_SYMBOL_GPL(clean_acked_data_flush);
141 #ifdef CONFIG_CGROUP_BPF
142 static void bpf_skops_parse_hdr(struct sock *sk, struct sk_buff *skb)
144 bool unknown_opt = tcp_sk(sk)->rx_opt.saw_unknown &&
145 BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk),
146 BPF_SOCK_OPS_PARSE_UNKNOWN_HDR_OPT_CB_FLAG);
147 bool parse_all_opt = BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk),
148 BPF_SOCK_OPS_PARSE_ALL_HDR_OPT_CB_FLAG);
149 struct bpf_sock_ops_kern sock_ops;
151 if (likely(!unknown_opt && !parse_all_opt))
154 /* The skb will be handled in the
155 * bpf_skops_established() or
156 * bpf_skops_write_hdr_opt().
158 switch (sk->sk_state) {
165 sock_owned_by_me(sk);
167 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp));
168 sock_ops.op = BPF_SOCK_OPS_PARSE_HDR_OPT_CB;
169 sock_ops.is_fullsock = 1;
171 bpf_skops_init_skb(&sock_ops, skb, tcp_hdrlen(skb));
173 BPF_CGROUP_RUN_PROG_SOCK_OPS(&sock_ops);
176 static void bpf_skops_established(struct sock *sk, int bpf_op,
179 struct bpf_sock_ops_kern sock_ops;
181 sock_owned_by_me(sk);
183 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp));
184 sock_ops.op = bpf_op;
185 sock_ops.is_fullsock = 1;
187 /* sk with TCP_REPAIR_ON does not have skb in tcp_finish_connect */
189 bpf_skops_init_skb(&sock_ops, skb, tcp_hdrlen(skb));
191 BPF_CGROUP_RUN_PROG_SOCK_OPS(&sock_ops);
194 static void bpf_skops_parse_hdr(struct sock *sk, struct sk_buff *skb)
198 static void bpf_skops_established(struct sock *sk, int bpf_op,
204 static void tcp_gro_dev_warn(struct sock *sk, const struct sk_buff *skb,
207 static bool __once __read_mostly;
210 struct net_device *dev;
215 dev = dev_get_by_index_rcu(sock_net(sk), skb->skb_iif);
216 if (!dev || len >= dev->mtu)
217 pr_warn("%s: Driver has suspect GRO implementation, TCP performance may be compromised.\n",
218 dev ? dev->name : "Unknown driver");
223 /* Adapt the MSS value used to make delayed ack decision to the
226 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
228 struct inet_connection_sock *icsk = inet_csk(sk);
229 const unsigned int lss = icsk->icsk_ack.last_seg_size;
232 icsk->icsk_ack.last_seg_size = 0;
234 /* skb->len may jitter because of SACKs, even if peer
235 * sends good full-sized frames.
237 len = skb_shinfo(skb)->gso_size ? : skb->len;
238 if (len >= icsk->icsk_ack.rcv_mss) {
239 icsk->icsk_ack.rcv_mss = min_t(unsigned int, len,
241 /* Account for possibly-removed options */
242 if (unlikely(len > icsk->icsk_ack.rcv_mss +
243 MAX_TCP_OPTION_SPACE))
244 tcp_gro_dev_warn(sk, skb, len);
246 /* Otherwise, we make more careful check taking into account,
247 * that SACKs block is variable.
249 * "len" is invariant segment length, including TCP header.
251 len += skb->data - skb_transport_header(skb);
252 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
253 /* If PSH is not set, packet should be
254 * full sized, provided peer TCP is not badly broken.
255 * This observation (if it is correct 8)) allows
256 * to handle super-low mtu links fairly.
258 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
259 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
260 /* Subtract also invariant (if peer is RFC compliant),
261 * tcp header plus fixed timestamp option length.
262 * Resulting "len" is MSS free of SACK jitter.
264 len -= tcp_sk(sk)->tcp_header_len;
265 icsk->icsk_ack.last_seg_size = len;
267 icsk->icsk_ack.rcv_mss = len;
271 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
272 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
273 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
277 static void tcp_incr_quickack(struct sock *sk, unsigned int max_quickacks)
279 struct inet_connection_sock *icsk = inet_csk(sk);
280 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
284 quickacks = min(quickacks, max_quickacks);
285 if (quickacks > icsk->icsk_ack.quick)
286 icsk->icsk_ack.quick = quickacks;
289 void tcp_enter_quickack_mode(struct sock *sk, unsigned int max_quickacks)
291 struct inet_connection_sock *icsk = inet_csk(sk);
293 tcp_incr_quickack(sk, max_quickacks);
294 inet_csk_exit_pingpong_mode(sk);
295 icsk->icsk_ack.ato = TCP_ATO_MIN;
297 EXPORT_SYMBOL(tcp_enter_quickack_mode);
299 /* Send ACKs quickly, if "quick" count is not exhausted
300 * and the session is not interactive.
303 static bool tcp_in_quickack_mode(struct sock *sk)
305 const struct inet_connection_sock *icsk = inet_csk(sk);
306 const struct dst_entry *dst = __sk_dst_get(sk);
308 return (dst && dst_metric(dst, RTAX_QUICKACK)) ||
309 (icsk->icsk_ack.quick && !inet_csk_in_pingpong_mode(sk));
312 static void tcp_ecn_queue_cwr(struct tcp_sock *tp)
314 if (tp->ecn_flags & TCP_ECN_OK)
315 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
318 static void tcp_ecn_accept_cwr(struct sock *sk, const struct sk_buff *skb)
320 if (tcp_hdr(skb)->cwr) {
321 tcp_sk(sk)->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
323 /* If the sender is telling us it has entered CWR, then its
324 * cwnd may be very low (even just 1 packet), so we should ACK
327 if (TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq)
328 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW;
332 static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp)
334 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
337 static void __tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb)
339 struct tcp_sock *tp = tcp_sk(sk);
341 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
342 case INET_ECN_NOT_ECT:
343 /* Funny extension: if ECT is not set on a segment,
344 * and we already seen ECT on a previous segment,
345 * it is probably a retransmit.
347 if (tp->ecn_flags & TCP_ECN_SEEN)
348 tcp_enter_quickack_mode(sk, 2);
351 if (tcp_ca_needs_ecn(sk))
352 tcp_ca_event(sk, CA_EVENT_ECN_IS_CE);
354 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
355 /* Better not delay acks, sender can have a very low cwnd */
356 tcp_enter_quickack_mode(sk, 2);
357 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
359 tp->ecn_flags |= TCP_ECN_SEEN;
362 if (tcp_ca_needs_ecn(sk))
363 tcp_ca_event(sk, CA_EVENT_ECN_NO_CE);
364 tp->ecn_flags |= TCP_ECN_SEEN;
369 static void tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb)
371 if (tcp_sk(sk)->ecn_flags & TCP_ECN_OK)
372 __tcp_ecn_check_ce(sk, skb);
375 static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
377 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
378 tp->ecn_flags &= ~TCP_ECN_OK;
381 static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
383 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
384 tp->ecn_flags &= ~TCP_ECN_OK;
387 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
389 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
394 /* Buffer size and advertised window tuning.
396 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
399 static void tcp_sndbuf_expand(struct sock *sk)
401 const struct tcp_sock *tp = tcp_sk(sk);
402 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
406 /* Worst case is non GSO/TSO : each frame consumes one skb
407 * and skb->head is kmalloced using power of two area of memory
409 per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
411 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
413 per_mss = roundup_pow_of_two(per_mss) +
414 SKB_DATA_ALIGN(sizeof(struct sk_buff));
416 nr_segs = max_t(u32, TCP_INIT_CWND, tp->snd_cwnd);
417 nr_segs = max_t(u32, nr_segs, tp->reordering + 1);
419 /* Fast Recovery (RFC 5681 3.2) :
420 * Cubic needs 1.7 factor, rounded to 2 to include
421 * extra cushion (application might react slowly to EPOLLOUT)
423 sndmem = ca_ops->sndbuf_expand ? ca_ops->sndbuf_expand(sk) : 2;
424 sndmem *= nr_segs * per_mss;
426 if (sk->sk_sndbuf < sndmem)
427 WRITE_ONCE(sk->sk_sndbuf,
428 min(sndmem, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_wmem[2])));
431 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
433 * All tcp_full_space() is split to two parts: "network" buffer, allocated
434 * forward and advertised in receiver window (tp->rcv_wnd) and
435 * "application buffer", required to isolate scheduling/application
436 * latencies from network.
437 * window_clamp is maximal advertised window. It can be less than
438 * tcp_full_space(), in this case tcp_full_space() - window_clamp
439 * is reserved for "application" buffer. The less window_clamp is
440 * the smoother our behaviour from viewpoint of network, but the lower
441 * throughput and the higher sensitivity of the connection to losses. 8)
443 * rcv_ssthresh is more strict window_clamp used at "slow start"
444 * phase to predict further behaviour of this connection.
445 * It is used for two goals:
446 * - to enforce header prediction at sender, even when application
447 * requires some significant "application buffer". It is check #1.
448 * - to prevent pruning of receive queue because of misprediction
449 * of receiver window. Check #2.
451 * The scheme does not work when sender sends good segments opening
452 * window and then starts to feed us spaghetti. But it should work
453 * in common situations. Otherwise, we have to rely on queue collapsing.
456 /* Slow part of check#2. */
457 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb,
458 unsigned int skbtruesize)
460 struct tcp_sock *tp = tcp_sk(sk);
462 int truesize = tcp_win_from_space(sk, skbtruesize) >> 1;
463 int window = tcp_win_from_space(sk, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2])) >> 1;
465 while (tp->rcv_ssthresh <= window) {
466 if (truesize <= skb->len)
467 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
475 /* Even if skb appears to have a bad len/truesize ratio, TCP coalescing
476 * can play nice with us, as sk_buff and skb->head might be either
477 * freed or shared with up to MAX_SKB_FRAGS segments.
478 * Only give a boost to drivers using page frag(s) to hold the frame(s),
479 * and if no payload was pulled in skb->head before reaching us.
481 static u32 truesize_adjust(bool adjust, const struct sk_buff *skb)
483 u32 truesize = skb->truesize;
485 if (adjust && !skb_headlen(skb)) {
486 truesize -= SKB_TRUESIZE(skb_end_offset(skb));
487 /* paranoid check, some drivers might be buggy */
488 if (unlikely((int)truesize < (int)skb->len))
489 truesize = skb->truesize;
494 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb,
497 struct tcp_sock *tp = tcp_sk(sk);
500 room = min_t(int, tp->window_clamp, tcp_space(sk)) - tp->rcv_ssthresh;
503 if (room > 0 && !tcp_under_memory_pressure(sk)) {
504 unsigned int truesize = truesize_adjust(adjust, skb);
507 /* Check #2. Increase window, if skb with such overhead
508 * will fit to rcvbuf in future.
510 if (tcp_win_from_space(sk, truesize) <= skb->len)
511 incr = 2 * tp->advmss;
513 incr = __tcp_grow_window(sk, skb, truesize);
516 incr = max_t(int, incr, 2 * skb->len);
517 tp->rcv_ssthresh += min(room, incr);
518 inet_csk(sk)->icsk_ack.quick |= 1;
523 /* 3. Try to fixup all. It is made immediately after connection enters
526 static void tcp_init_buffer_space(struct sock *sk)
528 int tcp_app_win = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_app_win);
529 struct tcp_sock *tp = tcp_sk(sk);
532 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
533 tcp_sndbuf_expand(sk);
535 tcp_mstamp_refresh(tp);
536 tp->rcvq_space.time = tp->tcp_mstamp;
537 tp->rcvq_space.seq = tp->copied_seq;
539 maxwin = tcp_full_space(sk);
541 if (tp->window_clamp >= maxwin) {
542 tp->window_clamp = maxwin;
544 if (tcp_app_win && maxwin > 4 * tp->advmss)
545 tp->window_clamp = max(maxwin -
546 (maxwin >> tcp_app_win),
550 /* Force reservation of one segment. */
552 tp->window_clamp > 2 * tp->advmss &&
553 tp->window_clamp + tp->advmss > maxwin)
554 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
556 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
557 tp->snd_cwnd_stamp = tcp_jiffies32;
558 tp->rcvq_space.space = min3(tp->rcv_ssthresh, tp->rcv_wnd,
559 (u32)TCP_INIT_CWND * tp->advmss);
562 /* 4. Recalculate window clamp after socket hit its memory bounds. */
563 static void tcp_clamp_window(struct sock *sk)
565 struct tcp_sock *tp = tcp_sk(sk);
566 struct inet_connection_sock *icsk = inet_csk(sk);
567 struct net *net = sock_net(sk);
570 icsk->icsk_ack.quick = 0;
571 rmem2 = READ_ONCE(net->ipv4.sysctl_tcp_rmem[2]);
573 if (sk->sk_rcvbuf < rmem2 &&
574 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
575 !tcp_under_memory_pressure(sk) &&
576 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
577 WRITE_ONCE(sk->sk_rcvbuf,
578 min(atomic_read(&sk->sk_rmem_alloc), rmem2));
580 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
581 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
584 /* Initialize RCV_MSS value.
585 * RCV_MSS is an our guess about MSS used by the peer.
586 * We haven't any direct information about the MSS.
587 * It's better to underestimate the RCV_MSS rather than overestimate.
588 * Overestimations make us ACKing less frequently than needed.
589 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
591 void tcp_initialize_rcv_mss(struct sock *sk)
593 const struct tcp_sock *tp = tcp_sk(sk);
594 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
596 hint = min(hint, tp->rcv_wnd / 2);
597 hint = min(hint, TCP_MSS_DEFAULT);
598 hint = max(hint, TCP_MIN_MSS);
600 inet_csk(sk)->icsk_ack.rcv_mss = hint;
602 EXPORT_SYMBOL(tcp_initialize_rcv_mss);
604 /* Receiver "autotuning" code.
606 * The algorithm for RTT estimation w/o timestamps is based on
607 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
608 * <https://public.lanl.gov/radiant/pubs.html#DRS>
610 * More detail on this code can be found at
611 * <http://staff.psc.edu/jheffner/>,
612 * though this reference is out of date. A new paper
615 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
617 u32 new_sample = tp->rcv_rtt_est.rtt_us;
620 if (new_sample != 0) {
621 /* If we sample in larger samples in the non-timestamp
622 * case, we could grossly overestimate the RTT especially
623 * with chatty applications or bulk transfer apps which
624 * are stalled on filesystem I/O.
626 * Also, since we are only going for a minimum in the
627 * non-timestamp case, we do not smooth things out
628 * else with timestamps disabled convergence takes too
632 m -= (new_sample >> 3);
640 /* No previous measure. */
644 tp->rcv_rtt_est.rtt_us = new_sample;
647 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
651 if (tp->rcv_rtt_est.time == 0)
653 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
655 delta_us = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcv_rtt_est.time);
658 tcp_rcv_rtt_update(tp, delta_us, 1);
661 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
662 tp->rcv_rtt_est.time = tp->tcp_mstamp;
665 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
666 const struct sk_buff *skb)
668 struct tcp_sock *tp = tcp_sk(sk);
670 if (tp->rx_opt.rcv_tsecr == tp->rcv_rtt_last_tsecr)
672 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr;
674 if (TCP_SKB_CB(skb)->end_seq -
675 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss) {
676 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr;
679 if (likely(delta < INT_MAX / (USEC_PER_SEC / TCP_TS_HZ))) {
682 delta_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
683 tcp_rcv_rtt_update(tp, delta_us, 0);
689 * This function should be called every time data is copied to user space.
690 * It calculates the appropriate TCP receive buffer space.
692 void tcp_rcv_space_adjust(struct sock *sk)
694 struct tcp_sock *tp = tcp_sk(sk);
698 trace_tcp_rcv_space_adjust(sk);
700 tcp_mstamp_refresh(tp);
701 time = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcvq_space.time);
702 if (time < (tp->rcv_rtt_est.rtt_us >> 3) || tp->rcv_rtt_est.rtt_us == 0)
705 /* Number of bytes copied to user in last RTT */
706 copied = tp->copied_seq - tp->rcvq_space.seq;
707 if (copied <= tp->rcvq_space.space)
711 * copied = bytes received in previous RTT, our base window
712 * To cope with packet losses, we need a 2x factor
713 * To cope with slow start, and sender growing its cwin by 100 %
714 * every RTT, we need a 4x factor, because the ACK we are sending
715 * now is for the next RTT, not the current one :
716 * <prev RTT . ><current RTT .. ><next RTT .... >
719 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf) &&
720 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
724 /* minimal window to cope with packet losses, assuming
725 * steady state. Add some cushion because of small variations.
727 rcvwin = ((u64)copied << 1) + 16 * tp->advmss;
729 /* Accommodate for sender rate increase (eg. slow start) */
730 grow = rcvwin * (copied - tp->rcvq_space.space);
731 do_div(grow, tp->rcvq_space.space);
732 rcvwin += (grow << 1);
734 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
735 while (tcp_win_from_space(sk, rcvmem) < tp->advmss)
738 do_div(rcvwin, tp->advmss);
739 rcvbuf = min_t(u64, rcvwin * rcvmem,
740 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]));
741 if (rcvbuf > sk->sk_rcvbuf) {
742 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
744 /* Make the window clamp follow along. */
745 tp->window_clamp = tcp_win_from_space(sk, rcvbuf);
748 tp->rcvq_space.space = copied;
751 tp->rcvq_space.seq = tp->copied_seq;
752 tp->rcvq_space.time = tp->tcp_mstamp;
755 /* There is something which you must keep in mind when you analyze the
756 * behavior of the tp->ato delayed ack timeout interval. When a
757 * connection starts up, we want to ack as quickly as possible. The
758 * problem is that "good" TCP's do slow start at the beginning of data
759 * transmission. The means that until we send the first few ACK's the
760 * sender will sit on his end and only queue most of his data, because
761 * he can only send snd_cwnd unacked packets at any given time. For
762 * each ACK we send, he increments snd_cwnd and transmits more of his
765 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
767 struct tcp_sock *tp = tcp_sk(sk);
768 struct inet_connection_sock *icsk = inet_csk(sk);
771 inet_csk_schedule_ack(sk);
773 tcp_measure_rcv_mss(sk, skb);
775 tcp_rcv_rtt_measure(tp);
779 if (!icsk->icsk_ack.ato) {
780 /* The _first_ data packet received, initialize
781 * delayed ACK engine.
783 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
784 icsk->icsk_ack.ato = TCP_ATO_MIN;
786 int m = now - icsk->icsk_ack.lrcvtime;
788 if (m <= TCP_ATO_MIN / 2) {
789 /* The fastest case is the first. */
790 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
791 } else if (m < icsk->icsk_ack.ato) {
792 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
793 if (icsk->icsk_ack.ato > icsk->icsk_rto)
794 icsk->icsk_ack.ato = icsk->icsk_rto;
795 } else if (m > icsk->icsk_rto) {
796 /* Too long gap. Apparently sender failed to
797 * restart window, so that we send ACKs quickly.
799 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
803 icsk->icsk_ack.lrcvtime = now;
805 tcp_ecn_check_ce(sk, skb);
808 tcp_grow_window(sk, skb, true);
811 /* Called to compute a smoothed rtt estimate. The data fed to this
812 * routine either comes from timestamps, or from segments that were
813 * known _not_ to have been retransmitted [see Karn/Partridge
814 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
815 * piece by Van Jacobson.
816 * NOTE: the next three routines used to be one big routine.
817 * To save cycles in the RFC 1323 implementation it was better to break
818 * it up into three procedures. -- erics
820 static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
822 struct tcp_sock *tp = tcp_sk(sk);
823 long m = mrtt_us; /* RTT */
824 u32 srtt = tp->srtt_us;
826 /* The following amusing code comes from Jacobson's
827 * article in SIGCOMM '88. Note that rtt and mdev
828 * are scaled versions of rtt and mean deviation.
829 * This is designed to be as fast as possible
830 * m stands for "measurement".
832 * On a 1990 paper the rto value is changed to:
833 * RTO = rtt + 4 * mdev
835 * Funny. This algorithm seems to be very broken.
836 * These formulae increase RTO, when it should be decreased, increase
837 * too slowly, when it should be increased quickly, decrease too quickly
838 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
839 * does not matter how to _calculate_ it. Seems, it was trap
840 * that VJ failed to avoid. 8)
843 m -= (srtt >> 3); /* m is now error in rtt est */
844 srtt += m; /* rtt = 7/8 rtt + 1/8 new */
846 m = -m; /* m is now abs(error) */
847 m -= (tp->mdev_us >> 2); /* similar update on mdev */
848 /* This is similar to one of Eifel findings.
849 * Eifel blocks mdev updates when rtt decreases.
850 * This solution is a bit different: we use finer gain
851 * for mdev in this case (alpha*beta).
852 * Like Eifel it also prevents growth of rto,
853 * but also it limits too fast rto decreases,
854 * happening in pure Eifel.
859 m -= (tp->mdev_us >> 2); /* similar update on mdev */
861 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */
862 if (tp->mdev_us > tp->mdev_max_us) {
863 tp->mdev_max_us = tp->mdev_us;
864 if (tp->mdev_max_us > tp->rttvar_us)
865 tp->rttvar_us = tp->mdev_max_us;
867 if (after(tp->snd_una, tp->rtt_seq)) {
868 if (tp->mdev_max_us < tp->rttvar_us)
869 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
870 tp->rtt_seq = tp->snd_nxt;
871 tp->mdev_max_us = tcp_rto_min_us(sk);
876 /* no previous measure. */
877 srtt = m << 3; /* take the measured time to be rtt */
878 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */
879 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
880 tp->mdev_max_us = tp->rttvar_us;
881 tp->rtt_seq = tp->snd_nxt;
885 tp->srtt_us = max(1U, srtt);
888 static void tcp_update_pacing_rate(struct sock *sk)
890 const struct tcp_sock *tp = tcp_sk(sk);
893 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
894 rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3);
896 /* current rate is (cwnd * mss) / srtt
897 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
898 * In Congestion Avoidance phase, set it to 120 % the current rate.
900 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
901 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
902 * end of slow start and should slow down.
904 if (tp->snd_cwnd < tp->snd_ssthresh / 2)
905 rate *= sock_net(sk)->ipv4.sysctl_tcp_pacing_ss_ratio;
907 rate *= sock_net(sk)->ipv4.sysctl_tcp_pacing_ca_ratio;
909 rate *= max(tp->snd_cwnd, tp->packets_out);
911 if (likely(tp->srtt_us))
912 do_div(rate, tp->srtt_us);
914 /* WRITE_ONCE() is needed because sch_fq fetches sk_pacing_rate
915 * without any lock. We want to make sure compiler wont store
916 * intermediate values in this location.
918 WRITE_ONCE(sk->sk_pacing_rate, min_t(u64, rate,
919 sk->sk_max_pacing_rate));
922 /* Calculate rto without backoff. This is the second half of Van Jacobson's
923 * routine referred to above.
925 static void tcp_set_rto(struct sock *sk)
927 const struct tcp_sock *tp = tcp_sk(sk);
928 /* Old crap is replaced with new one. 8)
931 * 1. If rtt variance happened to be less 50msec, it is hallucination.
932 * It cannot be less due to utterly erratic ACK generation made
933 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
934 * to do with delayed acks, because at cwnd>2 true delack timeout
935 * is invisible. Actually, Linux-2.4 also generates erratic
936 * ACKs in some circumstances.
938 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
940 /* 2. Fixups made earlier cannot be right.
941 * If we do not estimate RTO correctly without them,
942 * all the algo is pure shit and should be replaced
943 * with correct one. It is exactly, which we pretend to do.
946 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
947 * guarantees that rto is higher.
952 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
954 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
957 cwnd = TCP_INIT_CWND;
958 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
961 struct tcp_sacktag_state {
962 /* Timestamps for earliest and latest never-retransmitted segment
963 * that was SACKed. RTO needs the earliest RTT to stay conservative,
964 * but congestion control should still get an accurate delay signal.
971 unsigned int mss_now;
972 struct rate_sample *rate;
975 /* Take a notice that peer is sending D-SACKs. Skip update of data delivery
976 * and spurious retransmission information if this DSACK is unlikely caused by
978 * - DSACKed sequence range is larger than maximum receiver's window.
979 * - Total no. of DSACKed segments exceed the total no. of retransmitted segs.
981 static u32 tcp_dsack_seen(struct tcp_sock *tp, u32 start_seq,
982 u32 end_seq, struct tcp_sacktag_state *state)
984 u32 seq_len, dup_segs = 1;
986 if (!before(start_seq, end_seq))
989 seq_len = end_seq - start_seq;
990 /* Dubious DSACK: DSACKed range greater than maximum advertised rwnd */
991 if (seq_len > tp->max_window)
993 if (seq_len > tp->mss_cache)
994 dup_segs = DIV_ROUND_UP(seq_len, tp->mss_cache);
996 tp->dsack_dups += dup_segs;
997 /* Skip the DSACK if dup segs weren't retransmitted by sender */
998 if (tp->dsack_dups > tp->total_retrans)
1001 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
1002 tp->rack.dsack_seen = 1;
1004 state->flag |= FLAG_DSACKING_ACK;
1005 /* A spurious retransmission is delivered */
1006 state->sack_delivered += dup_segs;
1011 /* It's reordering when higher sequence was delivered (i.e. sacked) before
1012 * some lower never-retransmitted sequence ("low_seq"). The maximum reordering
1013 * distance is approximated in full-mss packet distance ("reordering").
1015 static void tcp_check_sack_reordering(struct sock *sk, const u32 low_seq,
1018 struct tcp_sock *tp = tcp_sk(sk);
1019 const u32 mss = tp->mss_cache;
1022 fack = tcp_highest_sack_seq(tp);
1023 if (!before(low_seq, fack))
1026 metric = fack - low_seq;
1027 if ((metric > tp->reordering * mss) && mss) {
1028 #if FASTRETRANS_DEBUG > 1
1029 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
1030 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
1034 tp->undo_marker ? tp->undo_retrans : 0);
1036 tp->reordering = min_t(u32, (metric + mss - 1) / mss,
1037 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_max_reordering));
1040 /* This exciting event is worth to be remembered. 8) */
1042 NET_INC_STATS(sock_net(sk),
1043 ts ? LINUX_MIB_TCPTSREORDER : LINUX_MIB_TCPSACKREORDER);
1046 /* This must be called before lost_out or retrans_out are updated
1047 * on a new loss, because we want to know if all skbs previously
1048 * known to be lost have already been retransmitted, indicating
1049 * that this newly lost skb is our next skb to retransmit.
1051 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
1053 if ((!tp->retransmit_skb_hint && tp->retrans_out >= tp->lost_out) ||
1054 (tp->retransmit_skb_hint &&
1055 before(TCP_SKB_CB(skb)->seq,
1056 TCP_SKB_CB(tp->retransmit_skb_hint)->seq)))
1057 tp->retransmit_skb_hint = skb;
1060 /* Sum the number of packets on the wire we have marked as lost, and
1061 * notify the congestion control module that the given skb was marked lost.
1063 static void tcp_notify_skb_loss_event(struct tcp_sock *tp, const struct sk_buff *skb)
1065 tp->lost += tcp_skb_pcount(skb);
1068 void tcp_mark_skb_lost(struct sock *sk, struct sk_buff *skb)
1070 __u8 sacked = TCP_SKB_CB(skb)->sacked;
1071 struct tcp_sock *tp = tcp_sk(sk);
1073 if (sacked & TCPCB_SACKED_ACKED)
1076 tcp_verify_retransmit_hint(tp, skb);
1077 if (sacked & TCPCB_LOST) {
1078 if (sacked & TCPCB_SACKED_RETRANS) {
1079 /* Account for retransmits that are lost again */
1080 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1081 tp->retrans_out -= tcp_skb_pcount(skb);
1082 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT,
1083 tcp_skb_pcount(skb));
1084 tcp_notify_skb_loss_event(tp, skb);
1087 tp->lost_out += tcp_skb_pcount(skb);
1088 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1089 tcp_notify_skb_loss_event(tp, skb);
1093 /* Updates the delivered and delivered_ce counts */
1094 static void tcp_count_delivered(struct tcp_sock *tp, u32 delivered,
1097 tp->delivered += delivered;
1099 tp->delivered_ce += delivered;
1102 /* This procedure tags the retransmission queue when SACKs arrive.
1104 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
1105 * Packets in queue with these bits set are counted in variables
1106 * sacked_out, retrans_out and lost_out, correspondingly.
1108 * Valid combinations are:
1109 * Tag InFlight Description
1110 * 0 1 - orig segment is in flight.
1111 * S 0 - nothing flies, orig reached receiver.
1112 * L 0 - nothing flies, orig lost by net.
1113 * R 2 - both orig and retransmit are in flight.
1114 * L|R 1 - orig is lost, retransmit is in flight.
1115 * S|R 1 - orig reached receiver, retrans is still in flight.
1116 * (L|S|R is logically valid, it could occur when L|R is sacked,
1117 * but it is equivalent to plain S and code short-curcuits it to S.
1118 * L|S is logically invalid, it would mean -1 packet in flight 8))
1120 * These 6 states form finite state machine, controlled by the following events:
1121 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
1122 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
1123 * 3. Loss detection event of two flavors:
1124 * A. Scoreboard estimator decided the packet is lost.
1125 * A'. Reno "three dupacks" marks head of queue lost.
1126 * B. SACK arrives sacking SND.NXT at the moment, when the
1127 * segment was retransmitted.
1128 * 4. D-SACK added new rule: D-SACK changes any tag to S.
1130 * It is pleasant to note, that state diagram turns out to be commutative,
1131 * so that we are allowed not to be bothered by order of our actions,
1132 * when multiple events arrive simultaneously. (see the function below).
1134 * Reordering detection.
1135 * --------------------
1136 * Reordering metric is maximal distance, which a packet can be displaced
1137 * in packet stream. With SACKs we can estimate it:
1139 * 1. SACK fills old hole and the corresponding segment was not
1140 * ever retransmitted -> reordering. Alas, we cannot use it
1141 * when segment was retransmitted.
1142 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1143 * for retransmitted and already SACKed segment -> reordering..
1144 * Both of these heuristics are not used in Loss state, when we cannot
1145 * account for retransmits accurately.
1147 * SACK block validation.
1148 * ----------------------
1150 * SACK block range validation checks that the received SACK block fits to
1151 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1152 * Note that SND.UNA is not included to the range though being valid because
1153 * it means that the receiver is rather inconsistent with itself reporting
1154 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1155 * perfectly valid, however, in light of RFC2018 which explicitly states
1156 * that "SACK block MUST reflect the newest segment. Even if the newest
1157 * segment is going to be discarded ...", not that it looks very clever
1158 * in case of head skb. Due to potentional receiver driven attacks, we
1159 * choose to avoid immediate execution of a walk in write queue due to
1160 * reneging and defer head skb's loss recovery to standard loss recovery
1161 * procedure that will eventually trigger (nothing forbids us doing this).
1163 * Implements also blockage to start_seq wrap-around. Problem lies in the
1164 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1165 * there's no guarantee that it will be before snd_nxt (n). The problem
1166 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1169 * <- outs wnd -> <- wrapzone ->
1170 * u e n u_w e_w s n_w
1172 * |<------------+------+----- TCP seqno space --------------+---------->|
1173 * ...-- <2^31 ->| |<--------...
1174 * ...---- >2^31 ------>| |<--------...
1176 * Current code wouldn't be vulnerable but it's better still to discard such
1177 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1178 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1179 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1180 * equal to the ideal case (infinite seqno space without wrap caused issues).
1182 * With D-SACK the lower bound is extended to cover sequence space below
1183 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1184 * again, D-SACK block must not to go across snd_una (for the same reason as
1185 * for the normal SACK blocks, explained above). But there all simplicity
1186 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1187 * fully below undo_marker they do not affect behavior in anyway and can
1188 * therefore be safely ignored. In rare cases (which are more or less
1189 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1190 * fragmentation and packet reordering past skb's retransmission. To consider
1191 * them correctly, the acceptable range must be extended even more though
1192 * the exact amount is rather hard to quantify. However, tp->max_window can
1193 * be used as an exaggerated estimate.
1195 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
1196 u32 start_seq, u32 end_seq)
1198 /* Too far in future, or reversed (interpretation is ambiguous) */
1199 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1202 /* Nasty start_seq wrap-around check (see comments above) */
1203 if (!before(start_seq, tp->snd_nxt))
1206 /* In outstanding window? ...This is valid exit for D-SACKs too.
1207 * start_seq == snd_una is non-sensical (see comments above)
1209 if (after(start_seq, tp->snd_una))
1212 if (!is_dsack || !tp->undo_marker)
1215 /* ...Then it's D-SACK, and must reside below snd_una completely */
1216 if (after(end_seq, tp->snd_una))
1219 if (!before(start_seq, tp->undo_marker))
1223 if (!after(end_seq, tp->undo_marker))
1226 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1227 * start_seq < undo_marker and end_seq >= undo_marker.
1229 return !before(start_seq, end_seq - tp->max_window);
1232 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1233 struct tcp_sack_block_wire *sp, int num_sacks,
1234 u32 prior_snd_una, struct tcp_sacktag_state *state)
1236 struct tcp_sock *tp = tcp_sk(sk);
1237 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1238 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1241 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1242 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1243 } else if (num_sacks > 1) {
1244 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1245 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1247 if (after(end_seq_0, end_seq_1) || before(start_seq_0, start_seq_1))
1249 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKOFORECV);
1254 dup_segs = tcp_dsack_seen(tp, start_seq_0, end_seq_0, state);
1255 if (!dup_segs) { /* Skip dubious DSACK */
1256 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKIGNOREDDUBIOUS);
1260 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECVSEGS, dup_segs);
1262 /* D-SACK for already forgotten data... Do dumb counting. */
1263 if (tp->undo_marker && tp->undo_retrans > 0 &&
1264 !after(end_seq_0, prior_snd_una) &&
1265 after(end_seq_0, tp->undo_marker))
1266 tp->undo_retrans = max_t(int, 0, tp->undo_retrans - dup_segs);
1271 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1272 * the incoming SACK may not exactly match but we can find smaller MSS
1273 * aligned portion of it that matches. Therefore we might need to fragment
1274 * which may fail and creates some hassle (caller must handle error case
1277 * FIXME: this could be merged to shift decision code
1279 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1280 u32 start_seq, u32 end_seq)
1284 unsigned int pkt_len;
1287 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1288 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1290 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1291 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1292 mss = tcp_skb_mss(skb);
1293 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1296 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1300 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1305 /* Round if necessary so that SACKs cover only full MSSes
1306 * and/or the remaining small portion (if present)
1308 if (pkt_len > mss) {
1309 unsigned int new_len = (pkt_len / mss) * mss;
1310 if (!in_sack && new_len < pkt_len)
1315 if (pkt_len >= skb->len && !in_sack)
1318 err = tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
1319 pkt_len, mss, GFP_ATOMIC);
1327 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1328 static u8 tcp_sacktag_one(struct sock *sk,
1329 struct tcp_sacktag_state *state, u8 sacked,
1330 u32 start_seq, u32 end_seq,
1331 int dup_sack, int pcount,
1334 struct tcp_sock *tp = tcp_sk(sk);
1336 /* Account D-SACK for retransmitted packet. */
1337 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1338 if (tp->undo_marker && tp->undo_retrans > 0 &&
1339 after(end_seq, tp->undo_marker))
1340 tp->undo_retrans = max_t(int, 0, tp->undo_retrans - pcount);
1341 if ((sacked & TCPCB_SACKED_ACKED) &&
1342 before(start_seq, state->reord))
1343 state->reord = start_seq;
1346 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1347 if (!after(end_seq, tp->snd_una))
1350 if (!(sacked & TCPCB_SACKED_ACKED)) {
1351 tcp_rack_advance(tp, sacked, end_seq, xmit_time);
1353 if (sacked & TCPCB_SACKED_RETRANS) {
1354 /* If the segment is not tagged as lost,
1355 * we do not clear RETRANS, believing
1356 * that retransmission is still in flight.
1358 if (sacked & TCPCB_LOST) {
1359 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1360 tp->lost_out -= pcount;
1361 tp->retrans_out -= pcount;
1364 if (!(sacked & TCPCB_RETRANS)) {
1365 /* New sack for not retransmitted frame,
1366 * which was in hole. It is reordering.
1368 if (before(start_seq,
1369 tcp_highest_sack_seq(tp)) &&
1370 before(start_seq, state->reord))
1371 state->reord = start_seq;
1373 if (!after(end_seq, tp->high_seq))
1374 state->flag |= FLAG_ORIG_SACK_ACKED;
1375 if (state->first_sackt == 0)
1376 state->first_sackt = xmit_time;
1377 state->last_sackt = xmit_time;
1380 if (sacked & TCPCB_LOST) {
1381 sacked &= ~TCPCB_LOST;
1382 tp->lost_out -= pcount;
1386 sacked |= TCPCB_SACKED_ACKED;
1387 state->flag |= FLAG_DATA_SACKED;
1388 tp->sacked_out += pcount;
1389 /* Out-of-order packets delivered */
1390 state->sack_delivered += pcount;
1392 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1393 if (tp->lost_skb_hint &&
1394 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1395 tp->lost_cnt_hint += pcount;
1398 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1399 * frames and clear it. undo_retrans is decreased above, L|R frames
1400 * are accounted above as well.
1402 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1403 sacked &= ~TCPCB_SACKED_RETRANS;
1404 tp->retrans_out -= pcount;
1410 /* Shift newly-SACKed bytes from this skb to the immediately previous
1411 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1413 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *prev,
1414 struct sk_buff *skb,
1415 struct tcp_sacktag_state *state,
1416 unsigned int pcount, int shifted, int mss,
1419 struct tcp_sock *tp = tcp_sk(sk);
1420 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1421 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
1425 /* Adjust counters and hints for the newly sacked sequence
1426 * range but discard the return value since prev is already
1427 * marked. We must tag the range first because the seq
1428 * advancement below implicitly advances
1429 * tcp_highest_sack_seq() when skb is highest_sack.
1431 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1432 start_seq, end_seq, dup_sack, pcount,
1433 tcp_skb_timestamp_us(skb));
1434 tcp_rate_skb_delivered(sk, skb, state->rate);
1436 if (skb == tp->lost_skb_hint)
1437 tp->lost_cnt_hint += pcount;
1439 TCP_SKB_CB(prev)->end_seq += shifted;
1440 TCP_SKB_CB(skb)->seq += shifted;
1442 tcp_skb_pcount_add(prev, pcount);
1443 WARN_ON_ONCE(tcp_skb_pcount(skb) < pcount);
1444 tcp_skb_pcount_add(skb, -pcount);
1446 /* When we're adding to gso_segs == 1, gso_size will be zero,
1447 * in theory this shouldn't be necessary but as long as DSACK
1448 * code can come after this skb later on it's better to keep
1449 * setting gso_size to something.
1451 if (!TCP_SKB_CB(prev)->tcp_gso_size)
1452 TCP_SKB_CB(prev)->tcp_gso_size = mss;
1454 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1455 if (tcp_skb_pcount(skb) <= 1)
1456 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1458 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1459 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1462 BUG_ON(!tcp_skb_pcount(skb));
1463 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1467 /* Whole SKB was eaten :-) */
1469 if (skb == tp->retransmit_skb_hint)
1470 tp->retransmit_skb_hint = prev;
1471 if (skb == tp->lost_skb_hint) {
1472 tp->lost_skb_hint = prev;
1473 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1476 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1477 TCP_SKB_CB(prev)->eor = TCP_SKB_CB(skb)->eor;
1478 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1479 TCP_SKB_CB(prev)->end_seq++;
1481 if (skb == tcp_highest_sack(sk))
1482 tcp_advance_highest_sack(sk, skb);
1484 tcp_skb_collapse_tstamp(prev, skb);
1485 if (unlikely(TCP_SKB_CB(prev)->tx.delivered_mstamp))
1486 TCP_SKB_CB(prev)->tx.delivered_mstamp = 0;
1488 tcp_rtx_queue_unlink_and_free(skb, sk);
1490 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKMERGED);
1495 /* I wish gso_size would have a bit more sane initialization than
1496 * something-or-zero which complicates things
1498 static int tcp_skb_seglen(const struct sk_buff *skb)
1500 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1503 /* Shifting pages past head area doesn't work */
1504 static int skb_can_shift(const struct sk_buff *skb)
1506 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1509 int tcp_skb_shift(struct sk_buff *to, struct sk_buff *from,
1510 int pcount, int shiftlen)
1512 /* TCP min gso_size is 8 bytes (TCP_MIN_GSO_SIZE)
1513 * Since TCP_SKB_CB(skb)->tcp_gso_segs is 16 bits, we need
1514 * to make sure not storing more than 65535 * 8 bytes per skb,
1515 * even if current MSS is bigger.
1517 if (unlikely(to->len + shiftlen >= 65535 * TCP_MIN_GSO_SIZE))
1519 if (unlikely(tcp_skb_pcount(to) + pcount > 65535))
1521 return skb_shift(to, from, shiftlen);
1524 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1527 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1528 struct tcp_sacktag_state *state,
1529 u32 start_seq, u32 end_seq,
1532 struct tcp_sock *tp = tcp_sk(sk);
1533 struct sk_buff *prev;
1539 /* Normally R but no L won't result in plain S */
1541 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1543 if (!skb_can_shift(skb))
1545 /* This frame is about to be dropped (was ACKed). */
1546 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1549 /* Can only happen with delayed DSACK + discard craziness */
1550 prev = skb_rb_prev(skb);
1554 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1557 if (!tcp_skb_can_collapse(prev, skb))
1560 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1561 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1565 pcount = tcp_skb_pcount(skb);
1566 mss = tcp_skb_seglen(skb);
1568 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1569 * drop this restriction as unnecessary
1571 if (mss != tcp_skb_seglen(prev))
1574 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1576 /* CHECKME: This is non-MSS split case only?, this will
1577 * cause skipped skbs due to advancing loop btw, original
1578 * has that feature too
1580 if (tcp_skb_pcount(skb) <= 1)
1583 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1585 /* TODO: head merge to next could be attempted here
1586 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1587 * though it might not be worth of the additional hassle
1589 * ...we can probably just fallback to what was done
1590 * previously. We could try merging non-SACKed ones
1591 * as well but it probably isn't going to buy off
1592 * because later SACKs might again split them, and
1593 * it would make skb timestamp tracking considerably
1599 len = end_seq - TCP_SKB_CB(skb)->seq;
1601 BUG_ON(len > skb->len);
1603 /* MSS boundaries should be honoured or else pcount will
1604 * severely break even though it makes things bit trickier.
1605 * Optimize common case to avoid most of the divides
1607 mss = tcp_skb_mss(skb);
1609 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1610 * drop this restriction as unnecessary
1612 if (mss != tcp_skb_seglen(prev))
1617 } else if (len < mss) {
1625 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1626 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1629 if (!tcp_skb_shift(prev, skb, pcount, len))
1631 if (!tcp_shifted_skb(sk, prev, skb, state, pcount, len, mss, dup_sack))
1634 /* Hole filled allows collapsing with the next as well, this is very
1635 * useful when hole on every nth skb pattern happens
1637 skb = skb_rb_next(prev);
1641 if (!skb_can_shift(skb) ||
1642 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1643 (mss != tcp_skb_seglen(skb)))
1646 if (!tcp_skb_can_collapse(prev, skb))
1649 pcount = tcp_skb_pcount(skb);
1650 if (tcp_skb_shift(prev, skb, pcount, len))
1651 tcp_shifted_skb(sk, prev, skb, state, pcount,
1661 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1665 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1666 struct tcp_sack_block *next_dup,
1667 struct tcp_sacktag_state *state,
1668 u32 start_seq, u32 end_seq,
1671 struct tcp_sock *tp = tcp_sk(sk);
1672 struct sk_buff *tmp;
1674 skb_rbtree_walk_from(skb) {
1676 bool dup_sack = dup_sack_in;
1678 /* queue is in-order => we can short-circuit the walk early */
1679 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1683 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1684 in_sack = tcp_match_skb_to_sack(sk, skb,
1685 next_dup->start_seq,
1691 /* skb reference here is a bit tricky to get right, since
1692 * shifting can eat and free both this skb and the next,
1693 * so not even _safe variant of the loop is enough.
1696 tmp = tcp_shift_skb_data(sk, skb, state,
1697 start_seq, end_seq, dup_sack);
1706 in_sack = tcp_match_skb_to_sack(sk, skb,
1712 if (unlikely(in_sack < 0))
1716 TCP_SKB_CB(skb)->sacked =
1719 TCP_SKB_CB(skb)->sacked,
1720 TCP_SKB_CB(skb)->seq,
1721 TCP_SKB_CB(skb)->end_seq,
1723 tcp_skb_pcount(skb),
1724 tcp_skb_timestamp_us(skb));
1725 tcp_rate_skb_delivered(sk, skb, state->rate);
1726 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1727 list_del_init(&skb->tcp_tsorted_anchor);
1729 if (!before(TCP_SKB_CB(skb)->seq,
1730 tcp_highest_sack_seq(tp)))
1731 tcp_advance_highest_sack(sk, skb);
1737 static struct sk_buff *tcp_sacktag_bsearch(struct sock *sk, u32 seq)
1739 struct rb_node *parent, **p = &sk->tcp_rtx_queue.rb_node;
1740 struct sk_buff *skb;
1744 skb = rb_to_skb(parent);
1745 if (before(seq, TCP_SKB_CB(skb)->seq)) {
1746 p = &parent->rb_left;
1749 if (!before(seq, TCP_SKB_CB(skb)->end_seq)) {
1750 p = &parent->rb_right;
1758 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1761 if (skb && after(TCP_SKB_CB(skb)->seq, skip_to_seq))
1764 return tcp_sacktag_bsearch(sk, skip_to_seq);
1767 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1769 struct tcp_sack_block *next_dup,
1770 struct tcp_sacktag_state *state,
1776 if (before(next_dup->start_seq, skip_to_seq)) {
1777 skb = tcp_sacktag_skip(skb, sk, next_dup->start_seq);
1778 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1779 next_dup->start_seq, next_dup->end_seq,
1786 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1788 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1792 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1793 u32 prior_snd_una, struct tcp_sacktag_state *state)
1795 struct tcp_sock *tp = tcp_sk(sk);
1796 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1797 TCP_SKB_CB(ack_skb)->sacked);
1798 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1799 struct tcp_sack_block sp[TCP_NUM_SACKS];
1800 struct tcp_sack_block *cache;
1801 struct sk_buff *skb;
1802 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1804 bool found_dup_sack = false;
1806 int first_sack_index;
1809 state->reord = tp->snd_nxt;
1811 if (!tp->sacked_out)
1812 tcp_highest_sack_reset(sk);
1814 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1815 num_sacks, prior_snd_una, state);
1817 /* Eliminate too old ACKs, but take into
1818 * account more or less fresh ones, they can
1819 * contain valid SACK info.
1821 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1824 if (!tp->packets_out)
1828 first_sack_index = 0;
1829 for (i = 0; i < num_sacks; i++) {
1830 bool dup_sack = !i && found_dup_sack;
1832 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1833 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1835 if (!tcp_is_sackblock_valid(tp, dup_sack,
1836 sp[used_sacks].start_seq,
1837 sp[used_sacks].end_seq)) {
1841 if (!tp->undo_marker)
1842 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1844 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1846 /* Don't count olds caused by ACK reordering */
1847 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1848 !after(sp[used_sacks].end_seq, tp->snd_una))
1850 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1853 NET_INC_STATS(sock_net(sk), mib_idx);
1855 first_sack_index = -1;
1859 /* Ignore very old stuff early */
1860 if (!after(sp[used_sacks].end_seq, prior_snd_una)) {
1862 first_sack_index = -1;
1869 /* order SACK blocks to allow in order walk of the retrans queue */
1870 for (i = used_sacks - 1; i > 0; i--) {
1871 for (j = 0; j < i; j++) {
1872 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1873 swap(sp[j], sp[j + 1]);
1875 /* Track where the first SACK block goes to */
1876 if (j == first_sack_index)
1877 first_sack_index = j + 1;
1882 state->mss_now = tcp_current_mss(sk);
1886 if (!tp->sacked_out) {
1887 /* It's already past, so skip checking against it */
1888 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1890 cache = tp->recv_sack_cache;
1891 /* Skip empty blocks in at head of the cache */
1892 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1897 while (i < used_sacks) {
1898 u32 start_seq = sp[i].start_seq;
1899 u32 end_seq = sp[i].end_seq;
1900 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1901 struct tcp_sack_block *next_dup = NULL;
1903 if (found_dup_sack && ((i + 1) == first_sack_index))
1904 next_dup = &sp[i + 1];
1906 /* Skip too early cached blocks */
1907 while (tcp_sack_cache_ok(tp, cache) &&
1908 !before(start_seq, cache->end_seq))
1911 /* Can skip some work by looking recv_sack_cache? */
1912 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1913 after(end_seq, cache->start_seq)) {
1916 if (before(start_seq, cache->start_seq)) {
1917 skb = tcp_sacktag_skip(skb, sk, start_seq);
1918 skb = tcp_sacktag_walk(skb, sk, next_dup,
1925 /* Rest of the block already fully processed? */
1926 if (!after(end_seq, cache->end_seq))
1929 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1933 /* ...tail remains todo... */
1934 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1935 /* ...but better entrypoint exists! */
1936 skb = tcp_highest_sack(sk);
1943 skb = tcp_sacktag_skip(skb, sk, cache->end_seq);
1944 /* Check overlap against next cached too (past this one already) */
1949 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1950 skb = tcp_highest_sack(sk);
1954 skb = tcp_sacktag_skip(skb, sk, start_seq);
1957 skb = tcp_sacktag_walk(skb, sk, next_dup, state,
1958 start_seq, end_seq, dup_sack);
1964 /* Clear the head of the cache sack blocks so we can skip it next time */
1965 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1966 tp->recv_sack_cache[i].start_seq = 0;
1967 tp->recv_sack_cache[i].end_seq = 0;
1969 for (j = 0; j < used_sacks; j++)
1970 tp->recv_sack_cache[i++] = sp[j];
1972 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Loss || tp->undo_marker)
1973 tcp_check_sack_reordering(sk, state->reord, 0);
1975 tcp_verify_left_out(tp);
1978 #if FASTRETRANS_DEBUG > 0
1979 WARN_ON((int)tp->sacked_out < 0);
1980 WARN_ON((int)tp->lost_out < 0);
1981 WARN_ON((int)tp->retrans_out < 0);
1982 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1987 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1988 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1990 static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
1994 holes = max(tp->lost_out, 1U);
1995 holes = min(holes, tp->packets_out);
1997 if ((tp->sacked_out + holes) > tp->packets_out) {
1998 tp->sacked_out = tp->packets_out - holes;
2004 /* If we receive more dupacks than we expected counting segments
2005 * in assumption of absent reordering, interpret this as reordering.
2006 * The only another reason could be bug in receiver TCP.
2008 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
2010 struct tcp_sock *tp = tcp_sk(sk);
2012 if (!tcp_limit_reno_sacked(tp))
2015 tp->reordering = min_t(u32, tp->packets_out + addend,
2016 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_max_reordering));
2018 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRENOREORDER);
2021 /* Emulate SACKs for SACKless connection: account for a new dupack. */
2023 static void tcp_add_reno_sack(struct sock *sk, int num_dupack, bool ece_ack)
2026 struct tcp_sock *tp = tcp_sk(sk);
2027 u32 prior_sacked = tp->sacked_out;
2030 tp->sacked_out += num_dupack;
2031 tcp_check_reno_reordering(sk, 0);
2032 delivered = tp->sacked_out - prior_sacked;
2034 tcp_count_delivered(tp, delivered, ece_ack);
2035 tcp_verify_left_out(tp);
2039 /* Account for ACK, ACKing some data in Reno Recovery phase. */
2041 static void tcp_remove_reno_sacks(struct sock *sk, int acked, bool ece_ack)
2043 struct tcp_sock *tp = tcp_sk(sk);
2046 /* One ACK acked hole. The rest eat duplicate ACKs. */
2047 tcp_count_delivered(tp, max_t(int, acked - tp->sacked_out, 1),
2049 if (acked - 1 >= tp->sacked_out)
2052 tp->sacked_out -= acked - 1;
2054 tcp_check_reno_reordering(sk, acked);
2055 tcp_verify_left_out(tp);
2058 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
2063 void tcp_clear_retrans(struct tcp_sock *tp)
2065 tp->retrans_out = 0;
2067 tp->undo_marker = 0;
2068 tp->undo_retrans = -1;
2072 static inline void tcp_init_undo(struct tcp_sock *tp)
2074 tp->undo_marker = tp->snd_una;
2075 /* Retransmission still in flight may cause DSACKs later. */
2076 tp->undo_retrans = tp->retrans_out ? : -1;
2079 static bool tcp_is_rack(const struct sock *sk)
2081 return READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_recovery) &
2082 TCP_RACK_LOSS_DETECTION;
2085 /* If we detect SACK reneging, forget all SACK information
2086 * and reset tags completely, otherwise preserve SACKs. If receiver
2087 * dropped its ofo queue, we will know this due to reneging detection.
2089 static void tcp_timeout_mark_lost(struct sock *sk)
2091 struct tcp_sock *tp = tcp_sk(sk);
2092 struct sk_buff *skb, *head;
2093 bool is_reneg; /* is receiver reneging on SACKs? */
2095 head = tcp_rtx_queue_head(sk);
2096 is_reneg = head && (TCP_SKB_CB(head)->sacked & TCPCB_SACKED_ACKED);
2098 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
2100 /* Mark SACK reneging until we recover from this loss event. */
2101 tp->is_sack_reneg = 1;
2102 } else if (tcp_is_reno(tp)) {
2103 tcp_reset_reno_sack(tp);
2107 skb_rbtree_walk_from(skb) {
2109 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
2110 else if (tcp_is_rack(sk) && skb != head &&
2111 tcp_rack_skb_timeout(tp, skb, 0) > 0)
2112 continue; /* Don't mark recently sent ones lost yet */
2113 tcp_mark_skb_lost(sk, skb);
2115 tcp_verify_left_out(tp);
2116 tcp_clear_all_retrans_hints(tp);
2119 /* Enter Loss state. */
2120 void tcp_enter_loss(struct sock *sk)
2122 const struct inet_connection_sock *icsk = inet_csk(sk);
2123 struct tcp_sock *tp = tcp_sk(sk);
2124 struct net *net = sock_net(sk);
2125 bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery;
2128 tcp_timeout_mark_lost(sk);
2130 /* Reduce ssthresh if it has not yet been made inside this window. */
2131 if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
2132 !after(tp->high_seq, tp->snd_una) ||
2133 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
2134 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2135 tp->prior_cwnd = tp->snd_cwnd;
2136 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2137 tcp_ca_event(sk, CA_EVENT_LOSS);
2140 tp->snd_cwnd = tcp_packets_in_flight(tp) + 1;
2141 tp->snd_cwnd_cnt = 0;
2142 tp->snd_cwnd_stamp = tcp_jiffies32;
2144 /* Timeout in disordered state after receiving substantial DUPACKs
2145 * suggests that the degree of reordering is over-estimated.
2147 reordering = READ_ONCE(net->ipv4.sysctl_tcp_reordering);
2148 if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
2149 tp->sacked_out >= reordering)
2150 tp->reordering = min_t(unsigned int, tp->reordering,
2153 tcp_set_ca_state(sk, TCP_CA_Loss);
2154 tp->high_seq = tp->snd_nxt;
2155 tcp_ecn_queue_cwr(tp);
2157 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
2158 * loss recovery is underway except recurring timeout(s) on
2159 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
2161 tp->frto = READ_ONCE(net->ipv4.sysctl_tcp_frto) &&
2162 (new_recovery || icsk->icsk_retransmits) &&
2163 !inet_csk(sk)->icsk_mtup.probe_size;
2166 /* If ACK arrived pointing to a remembered SACK, it means that our
2167 * remembered SACKs do not reflect real state of receiver i.e.
2168 * receiver _host_ is heavily congested (or buggy).
2170 * To avoid big spurious retransmission bursts due to transient SACK
2171 * scoreboard oddities that look like reneging, we give the receiver a
2172 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
2173 * restore sanity to the SACK scoreboard. If the apparent reneging
2174 * persists until this RTO then we'll clear the SACK scoreboard.
2176 static bool tcp_check_sack_reneging(struct sock *sk, int flag)
2178 if (flag & FLAG_SACK_RENEGING &&
2179 flag & FLAG_SND_UNA_ADVANCED) {
2180 struct tcp_sock *tp = tcp_sk(sk);
2181 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4),
2182 msecs_to_jiffies(10));
2184 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2185 delay, TCP_RTO_MAX);
2191 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2192 * counter when SACK is enabled (without SACK, sacked_out is used for
2195 * With reordering, holes may still be in flight, so RFC3517 recovery
2196 * uses pure sacked_out (total number of SACKed segments) even though
2197 * it violates the RFC that uses duplicate ACKs, often these are equal
2198 * but when e.g. out-of-window ACKs or packet duplication occurs,
2199 * they differ. Since neither occurs due to loss, TCP should really
2202 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2204 return tp->sacked_out + 1;
2207 /* Linux NewReno/SACK/ECN state machine.
2208 * --------------------------------------
2210 * "Open" Normal state, no dubious events, fast path.
2211 * "Disorder" In all the respects it is "Open",
2212 * but requires a bit more attention. It is entered when
2213 * we see some SACKs or dupacks. It is split of "Open"
2214 * mainly to move some processing from fast path to slow one.
2215 * "CWR" CWND was reduced due to some Congestion Notification event.
2216 * It can be ECN, ICMP source quench, local device congestion.
2217 * "Recovery" CWND was reduced, we are fast-retransmitting.
2218 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2220 * tcp_fastretrans_alert() is entered:
2221 * - each incoming ACK, if state is not "Open"
2222 * - when arrived ACK is unusual, namely:
2227 * Counting packets in flight is pretty simple.
2229 * in_flight = packets_out - left_out + retrans_out
2231 * packets_out is SND.NXT-SND.UNA counted in packets.
2233 * retrans_out is number of retransmitted segments.
2235 * left_out is number of segments left network, but not ACKed yet.
2237 * left_out = sacked_out + lost_out
2239 * sacked_out: Packets, which arrived to receiver out of order
2240 * and hence not ACKed. With SACKs this number is simply
2241 * amount of SACKed data. Even without SACKs
2242 * it is easy to give pretty reliable estimate of this number,
2243 * counting duplicate ACKs.
2245 * lost_out: Packets lost by network. TCP has no explicit
2246 * "loss notification" feedback from network (for now).
2247 * It means that this number can be only _guessed_.
2248 * Actually, it is the heuristics to predict lossage that
2249 * distinguishes different algorithms.
2251 * F.e. after RTO, when all the queue is considered as lost,
2252 * lost_out = packets_out and in_flight = retrans_out.
2254 * Essentially, we have now a few algorithms detecting
2257 * If the receiver supports SACK:
2259 * RFC6675/3517: It is the conventional algorithm. A packet is
2260 * considered lost if the number of higher sequence packets
2261 * SACKed is greater than or equal the DUPACK thoreshold
2262 * (reordering). This is implemented in tcp_mark_head_lost and
2263 * tcp_update_scoreboard.
2265 * RACK (draft-ietf-tcpm-rack-01): it is a newer algorithm
2266 * (2017-) that checks timing instead of counting DUPACKs.
2267 * Essentially a packet is considered lost if it's not S/ACKed
2268 * after RTT + reordering_window, where both metrics are
2269 * dynamically measured and adjusted. This is implemented in
2270 * tcp_rack_mark_lost.
2272 * If the receiver does not support SACK:
2274 * NewReno (RFC6582): in Recovery we assume that one segment
2275 * is lost (classic Reno). While we are in Recovery and
2276 * a partial ACK arrives, we assume that one more packet
2277 * is lost (NewReno). This heuristics are the same in NewReno
2280 * Really tricky (and requiring careful tuning) part of algorithm
2281 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2282 * The first determines the moment _when_ we should reduce CWND and,
2283 * hence, slow down forward transmission. In fact, it determines the moment
2284 * when we decide that hole is caused by loss, rather than by a reorder.
2286 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2287 * holes, caused by lost packets.
2289 * And the most logically complicated part of algorithm is undo
2290 * heuristics. We detect false retransmits due to both too early
2291 * fast retransmit (reordering) and underestimated RTO, analyzing
2292 * timestamps and D-SACKs. When we detect that some segments were
2293 * retransmitted by mistake and CWND reduction was wrong, we undo
2294 * window reduction and abort recovery phase. This logic is hidden
2295 * inside several functions named tcp_try_undo_<something>.
2298 /* This function decides, when we should leave Disordered state
2299 * and enter Recovery phase, reducing congestion window.
2301 * Main question: may we further continue forward transmission
2302 * with the same cwnd?
2304 static bool tcp_time_to_recover(struct sock *sk, int flag)
2306 struct tcp_sock *tp = tcp_sk(sk);
2308 /* Trick#1: The loss is proven. */
2312 /* Not-A-Trick#2 : Classic rule... */
2313 if (!tcp_is_rack(sk) && tcp_dupack_heuristics(tp) > tp->reordering)
2319 /* Detect loss in event "A" above by marking head of queue up as lost.
2320 * For RFC3517 SACK, a segment is considered lost if it
2321 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2322 * the maximum SACKed segments to pass before reaching this limit.
2324 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2326 struct tcp_sock *tp = tcp_sk(sk);
2327 struct sk_buff *skb;
2329 /* Use SACK to deduce losses of new sequences sent during recovery */
2330 const u32 loss_high = tp->snd_nxt;
2332 WARN_ON(packets > tp->packets_out);
2333 skb = tp->lost_skb_hint;
2335 /* Head already handled? */
2336 if (mark_head && after(TCP_SKB_CB(skb)->seq, tp->snd_una))
2338 cnt = tp->lost_cnt_hint;
2340 skb = tcp_rtx_queue_head(sk);
2344 skb_rbtree_walk_from(skb) {
2345 /* TODO: do this better */
2346 /* this is not the most efficient way to do this... */
2347 tp->lost_skb_hint = skb;
2348 tp->lost_cnt_hint = cnt;
2350 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2353 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2354 cnt += tcp_skb_pcount(skb);
2359 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_LOST))
2360 tcp_mark_skb_lost(sk, skb);
2365 tcp_verify_left_out(tp);
2368 /* Account newly detected lost packet(s) */
2370 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2372 struct tcp_sock *tp = tcp_sk(sk);
2374 if (tcp_is_sack(tp)) {
2375 int sacked_upto = tp->sacked_out - tp->reordering;
2376 if (sacked_upto >= 0)
2377 tcp_mark_head_lost(sk, sacked_upto, 0);
2378 else if (fast_rexmit)
2379 tcp_mark_head_lost(sk, 1, 1);
2383 static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when)
2385 return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2386 before(tp->rx_opt.rcv_tsecr, when);
2389 /* skb is spurious retransmitted if the returned timestamp echo
2390 * reply is prior to the skb transmission time
2392 static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp,
2393 const struct sk_buff *skb)
2395 return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) &&
2396 tcp_tsopt_ecr_before(tp, tcp_skb_timestamp(skb));
2399 /* Nothing was retransmitted or returned timestamp is less
2400 * than timestamp of the first retransmission.
2402 static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2404 return tp->retrans_stamp &&
2405 tcp_tsopt_ecr_before(tp, tp->retrans_stamp);
2408 /* Undo procedures. */
2410 /* We can clear retrans_stamp when there are no retransmissions in the
2411 * window. It would seem that it is trivially available for us in
2412 * tp->retrans_out, however, that kind of assumptions doesn't consider
2413 * what will happen if errors occur when sending retransmission for the
2414 * second time. ...It could the that such segment has only
2415 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2416 * the head skb is enough except for some reneging corner cases that
2417 * are not worth the effort.
2419 * Main reason for all this complexity is the fact that connection dying
2420 * time now depends on the validity of the retrans_stamp, in particular,
2421 * that successive retransmissions of a segment must not advance
2422 * retrans_stamp under any conditions.
2424 static bool tcp_any_retrans_done(const struct sock *sk)
2426 const struct tcp_sock *tp = tcp_sk(sk);
2427 struct sk_buff *skb;
2429 if (tp->retrans_out)
2432 skb = tcp_rtx_queue_head(sk);
2433 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2439 static void DBGUNDO(struct sock *sk, const char *msg)
2441 #if FASTRETRANS_DEBUG > 1
2442 struct tcp_sock *tp = tcp_sk(sk);
2443 struct inet_sock *inet = inet_sk(sk);
2445 if (sk->sk_family == AF_INET) {
2446 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2448 &inet->inet_daddr, ntohs(inet->inet_dport),
2449 tp->snd_cwnd, tcp_left_out(tp),
2450 tp->snd_ssthresh, tp->prior_ssthresh,
2453 #if IS_ENABLED(CONFIG_IPV6)
2454 else if (sk->sk_family == AF_INET6) {
2455 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2457 &sk->sk_v6_daddr, ntohs(inet->inet_dport),
2458 tp->snd_cwnd, tcp_left_out(tp),
2459 tp->snd_ssthresh, tp->prior_ssthresh,
2466 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
2468 struct tcp_sock *tp = tcp_sk(sk);
2471 struct sk_buff *skb;
2473 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2474 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2477 tcp_clear_all_retrans_hints(tp);
2480 if (tp->prior_ssthresh) {
2481 const struct inet_connection_sock *icsk = inet_csk(sk);
2483 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2485 if (tp->prior_ssthresh > tp->snd_ssthresh) {
2486 tp->snd_ssthresh = tp->prior_ssthresh;
2487 tcp_ecn_withdraw_cwr(tp);
2490 tp->snd_cwnd_stamp = tcp_jiffies32;
2491 tp->undo_marker = 0;
2492 tp->rack.advanced = 1; /* Force RACK to re-exam losses */
2495 static inline bool tcp_may_undo(const struct tcp_sock *tp)
2497 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2500 static bool tcp_is_non_sack_preventing_reopen(struct sock *sk)
2502 struct tcp_sock *tp = tcp_sk(sk);
2504 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2505 /* Hold old state until something *above* high_seq
2506 * is ACKed. For Reno it is MUST to prevent false
2507 * fast retransmits (RFC2582). SACK TCP is safe. */
2508 if (!tcp_any_retrans_done(sk))
2509 tp->retrans_stamp = 0;
2515 /* People celebrate: "We love our President!" */
2516 static bool tcp_try_undo_recovery(struct sock *sk)
2518 struct tcp_sock *tp = tcp_sk(sk);
2520 if (tcp_may_undo(tp)) {
2523 /* Happy end! We did not retransmit anything
2524 * or our original transmission succeeded.
2526 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2527 tcp_undo_cwnd_reduction(sk, false);
2528 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2529 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2531 mib_idx = LINUX_MIB_TCPFULLUNDO;
2533 NET_INC_STATS(sock_net(sk), mib_idx);
2534 } else if (tp->rack.reo_wnd_persist) {
2535 tp->rack.reo_wnd_persist--;
2537 if (tcp_is_non_sack_preventing_reopen(sk))
2539 tcp_set_ca_state(sk, TCP_CA_Open);
2540 tp->is_sack_reneg = 0;
2544 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2545 static bool tcp_try_undo_dsack(struct sock *sk)
2547 struct tcp_sock *tp = tcp_sk(sk);
2549 if (tp->undo_marker && !tp->undo_retrans) {
2550 tp->rack.reo_wnd_persist = min(TCP_RACK_RECOVERY_THRESH,
2551 tp->rack.reo_wnd_persist + 1);
2552 DBGUNDO(sk, "D-SACK");
2553 tcp_undo_cwnd_reduction(sk, false);
2554 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2560 /* Undo during loss recovery after partial ACK or using F-RTO. */
2561 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2563 struct tcp_sock *tp = tcp_sk(sk);
2565 if (frto_undo || tcp_may_undo(tp)) {
2566 tcp_undo_cwnd_reduction(sk, true);
2568 DBGUNDO(sk, "partial loss");
2569 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2571 NET_INC_STATS(sock_net(sk),
2572 LINUX_MIB_TCPSPURIOUSRTOS);
2573 inet_csk(sk)->icsk_retransmits = 0;
2574 if (tcp_is_non_sack_preventing_reopen(sk))
2576 if (frto_undo || tcp_is_sack(tp)) {
2577 tcp_set_ca_state(sk, TCP_CA_Open);
2578 tp->is_sack_reneg = 0;
2585 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
2586 * It computes the number of packets to send (sndcnt) based on packets newly
2588 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2589 * cwnd reductions across a full RTT.
2590 * 2) Otherwise PRR uses packet conservation to send as much as delivered.
2591 * But when the retransmits are acked without further losses, PRR
2592 * slow starts cwnd up to ssthresh to speed up the recovery.
2594 static void tcp_init_cwnd_reduction(struct sock *sk)
2596 struct tcp_sock *tp = tcp_sk(sk);
2598 tp->high_seq = tp->snd_nxt;
2599 tp->tlp_high_seq = 0;
2600 tp->snd_cwnd_cnt = 0;
2601 tp->prior_cwnd = tp->snd_cwnd;
2602 tp->prr_delivered = 0;
2604 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2605 tcp_ecn_queue_cwr(tp);
2608 void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked, int flag)
2610 struct tcp_sock *tp = tcp_sk(sk);
2612 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2614 if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd))
2617 tp->prr_delivered += newly_acked_sacked;
2619 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2621 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2622 } else if ((flag & (FLAG_RETRANS_DATA_ACKED | FLAG_LOST_RETRANS)) ==
2623 FLAG_RETRANS_DATA_ACKED) {
2624 sndcnt = min_t(int, delta,
2625 max_t(int, tp->prr_delivered - tp->prr_out,
2626 newly_acked_sacked) + 1);
2628 sndcnt = min(delta, newly_acked_sacked);
2630 /* Force a fast retransmit upon entering fast recovery */
2631 sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1));
2632 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2635 static inline void tcp_end_cwnd_reduction(struct sock *sk)
2637 struct tcp_sock *tp = tcp_sk(sk);
2639 if (inet_csk(sk)->icsk_ca_ops->cong_control)
2642 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2643 if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH &&
2644 (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR || tp->undo_marker)) {
2645 tp->snd_cwnd = tp->snd_ssthresh;
2646 tp->snd_cwnd_stamp = tcp_jiffies32;
2648 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2651 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2652 void tcp_enter_cwr(struct sock *sk)
2654 struct tcp_sock *tp = tcp_sk(sk);
2656 tp->prior_ssthresh = 0;
2657 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2658 tp->undo_marker = 0;
2659 tcp_init_cwnd_reduction(sk);
2660 tcp_set_ca_state(sk, TCP_CA_CWR);
2663 EXPORT_SYMBOL(tcp_enter_cwr);
2665 static void tcp_try_keep_open(struct sock *sk)
2667 struct tcp_sock *tp = tcp_sk(sk);
2668 int state = TCP_CA_Open;
2670 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2671 state = TCP_CA_Disorder;
2673 if (inet_csk(sk)->icsk_ca_state != state) {
2674 tcp_set_ca_state(sk, state);
2675 tp->high_seq = tp->snd_nxt;
2679 static void tcp_try_to_open(struct sock *sk, int flag)
2681 struct tcp_sock *tp = tcp_sk(sk);
2683 tcp_verify_left_out(tp);
2685 if (!tcp_any_retrans_done(sk))
2686 tp->retrans_stamp = 0;
2688 if (flag & FLAG_ECE)
2691 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2692 tcp_try_keep_open(sk);
2696 static void tcp_mtup_probe_failed(struct sock *sk)
2698 struct inet_connection_sock *icsk = inet_csk(sk);
2700 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2701 icsk->icsk_mtup.probe_size = 0;
2702 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPFAIL);
2705 static void tcp_mtup_probe_success(struct sock *sk)
2707 struct tcp_sock *tp = tcp_sk(sk);
2708 struct inet_connection_sock *icsk = inet_csk(sk);
2711 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2713 val = (u64)tp->snd_cwnd * tcp_mss_to_mtu(sk, tp->mss_cache);
2714 do_div(val, icsk->icsk_mtup.probe_size);
2715 WARN_ON_ONCE((u32)val != val);
2716 tp->snd_cwnd = max_t(u32, 1U, val);
2718 tp->snd_cwnd_cnt = 0;
2719 tp->snd_cwnd_stamp = tcp_jiffies32;
2720 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2722 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2723 icsk->icsk_mtup.probe_size = 0;
2724 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2725 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS);
2728 /* Do a simple retransmit without using the backoff mechanisms in
2729 * tcp_timer. This is used for path mtu discovery.
2730 * The socket is already locked here.
2732 void tcp_simple_retransmit(struct sock *sk)
2734 const struct inet_connection_sock *icsk = inet_csk(sk);
2735 struct tcp_sock *tp = tcp_sk(sk);
2736 struct sk_buff *skb;
2737 unsigned int mss = tcp_current_mss(sk);
2739 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2740 if (tcp_skb_seglen(skb) > mss)
2741 tcp_mark_skb_lost(sk, skb);
2744 tcp_clear_retrans_hints_partial(tp);
2749 if (tcp_is_reno(tp))
2750 tcp_limit_reno_sacked(tp);
2752 tcp_verify_left_out(tp);
2754 /* Don't muck with the congestion window here.
2755 * Reason is that we do not increase amount of _data_
2756 * in network, but units changed and effective
2757 * cwnd/ssthresh really reduced now.
2759 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2760 tp->high_seq = tp->snd_nxt;
2761 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2762 tp->prior_ssthresh = 0;
2763 tp->undo_marker = 0;
2764 tcp_set_ca_state(sk, TCP_CA_Loss);
2766 tcp_xmit_retransmit_queue(sk);
2768 EXPORT_SYMBOL(tcp_simple_retransmit);
2770 void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2772 struct tcp_sock *tp = tcp_sk(sk);
2775 if (tcp_is_reno(tp))
2776 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2778 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2780 NET_INC_STATS(sock_net(sk), mib_idx);
2782 tp->prior_ssthresh = 0;
2785 if (!tcp_in_cwnd_reduction(sk)) {
2787 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2788 tcp_init_cwnd_reduction(sk);
2790 tcp_set_ca_state(sk, TCP_CA_Recovery);
2793 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2794 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2796 static void tcp_process_loss(struct sock *sk, int flag, int num_dupack,
2799 struct tcp_sock *tp = tcp_sk(sk);
2800 bool recovered = !before(tp->snd_una, tp->high_seq);
2802 if ((flag & FLAG_SND_UNA_ADVANCED || rcu_access_pointer(tp->fastopen_rsk)) &&
2803 tcp_try_undo_loss(sk, false))
2806 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2807 /* Step 3.b. A timeout is spurious if not all data are
2808 * lost, i.e., never-retransmitted data are (s)acked.
2810 if ((flag & FLAG_ORIG_SACK_ACKED) &&
2811 tcp_try_undo_loss(sk, true))
2814 if (after(tp->snd_nxt, tp->high_seq)) {
2815 if (flag & FLAG_DATA_SACKED || num_dupack)
2816 tp->frto = 0; /* Step 3.a. loss was real */
2817 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2818 tp->high_seq = tp->snd_nxt;
2819 /* Step 2.b. Try send new data (but deferred until cwnd
2820 * is updated in tcp_ack()). Otherwise fall back to
2821 * the conventional recovery.
2823 if (!tcp_write_queue_empty(sk) &&
2824 after(tcp_wnd_end(tp), tp->snd_nxt)) {
2825 *rexmit = REXMIT_NEW;
2833 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2834 tcp_try_undo_recovery(sk);
2837 if (tcp_is_reno(tp)) {
2838 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2839 * delivered. Lower inflight to clock out (re)tranmissions.
2841 if (after(tp->snd_nxt, tp->high_seq) && num_dupack)
2842 tcp_add_reno_sack(sk, num_dupack, flag & FLAG_ECE);
2843 else if (flag & FLAG_SND_UNA_ADVANCED)
2844 tcp_reset_reno_sack(tp);
2846 *rexmit = REXMIT_LOST;
2849 static bool tcp_force_fast_retransmit(struct sock *sk)
2851 struct tcp_sock *tp = tcp_sk(sk);
2853 return after(tcp_highest_sack_seq(tp),
2854 tp->snd_una + tp->reordering * tp->mss_cache);
2857 /* Undo during fast recovery after partial ACK. */
2858 static bool tcp_try_undo_partial(struct sock *sk, u32 prior_snd_una,
2861 struct tcp_sock *tp = tcp_sk(sk);
2863 if (tp->undo_marker && tcp_packet_delayed(tp)) {
2864 /* Plain luck! Hole if filled with delayed
2865 * packet, rather than with a retransmit. Check reordering.
2867 tcp_check_sack_reordering(sk, prior_snd_una, 1);
2869 /* We are getting evidence that the reordering degree is higher
2870 * than we realized. If there are no retransmits out then we
2871 * can undo. Otherwise we clock out new packets but do not
2872 * mark more packets lost or retransmit more.
2874 if (tp->retrans_out)
2877 if (!tcp_any_retrans_done(sk))
2878 tp->retrans_stamp = 0;
2880 DBGUNDO(sk, "partial recovery");
2881 tcp_undo_cwnd_reduction(sk, true);
2882 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2883 tcp_try_keep_open(sk);
2885 /* Partial ACK arrived. Force fast retransmit. */
2886 *do_lost = tcp_force_fast_retransmit(sk);
2891 static void tcp_identify_packet_loss(struct sock *sk, int *ack_flag)
2893 struct tcp_sock *tp = tcp_sk(sk);
2895 if (tcp_rtx_queue_empty(sk))
2898 if (unlikely(tcp_is_reno(tp))) {
2899 tcp_newreno_mark_lost(sk, *ack_flag & FLAG_SND_UNA_ADVANCED);
2900 } else if (tcp_is_rack(sk)) {
2901 u32 prior_retrans = tp->retrans_out;
2903 if (tcp_rack_mark_lost(sk))
2904 *ack_flag &= ~FLAG_SET_XMIT_TIMER;
2905 if (prior_retrans > tp->retrans_out)
2906 *ack_flag |= FLAG_LOST_RETRANS;
2910 /* Process an event, which can update packets-in-flight not trivially.
2911 * Main goal of this function is to calculate new estimate for left_out,
2912 * taking into account both packets sitting in receiver's buffer and
2913 * packets lost by network.
2915 * Besides that it updates the congestion state when packet loss or ECN
2916 * is detected. But it does not reduce the cwnd, it is done by the
2917 * congestion control later.
2919 * It does _not_ decide what to send, it is made in function
2920 * tcp_xmit_retransmit_queue().
2922 static void tcp_fastretrans_alert(struct sock *sk, const u32 prior_snd_una,
2923 int num_dupack, int *ack_flag, int *rexmit)
2925 struct inet_connection_sock *icsk = inet_csk(sk);
2926 struct tcp_sock *tp = tcp_sk(sk);
2927 int fast_rexmit = 0, flag = *ack_flag;
2928 bool ece_ack = flag & FLAG_ECE;
2929 bool do_lost = num_dupack || ((flag & FLAG_DATA_SACKED) &&
2930 tcp_force_fast_retransmit(sk));
2932 if (!tp->packets_out && tp->sacked_out)
2935 /* Now state machine starts.
2936 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2938 tp->prior_ssthresh = 0;
2940 /* B. In all the states check for reneging SACKs. */
2941 if (tcp_check_sack_reneging(sk, flag))
2944 /* C. Check consistency of the current state. */
2945 tcp_verify_left_out(tp);
2947 /* D. Check state exit conditions. State can be terminated
2948 * when high_seq is ACKed. */
2949 if (icsk->icsk_ca_state == TCP_CA_Open) {
2950 WARN_ON(tp->retrans_out != 0);
2951 tp->retrans_stamp = 0;
2952 } else if (!before(tp->snd_una, tp->high_seq)) {
2953 switch (icsk->icsk_ca_state) {
2955 /* CWR is to be held something *above* high_seq
2956 * is ACKed for CWR bit to reach receiver. */
2957 if (tp->snd_una != tp->high_seq) {
2958 tcp_end_cwnd_reduction(sk);
2959 tcp_set_ca_state(sk, TCP_CA_Open);
2963 case TCP_CA_Recovery:
2964 if (tcp_is_reno(tp))
2965 tcp_reset_reno_sack(tp);
2966 if (tcp_try_undo_recovery(sk))
2968 tcp_end_cwnd_reduction(sk);
2973 /* E. Process state. */
2974 switch (icsk->icsk_ca_state) {
2975 case TCP_CA_Recovery:
2976 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2977 if (tcp_is_reno(tp))
2978 tcp_add_reno_sack(sk, num_dupack, ece_ack);
2979 } else if (tcp_try_undo_partial(sk, prior_snd_una, &do_lost))
2982 if (tcp_try_undo_dsack(sk))
2983 tcp_try_keep_open(sk);
2985 tcp_identify_packet_loss(sk, ack_flag);
2986 if (icsk->icsk_ca_state != TCP_CA_Recovery) {
2987 if (!tcp_time_to_recover(sk, flag))
2989 /* Undo reverts the recovery state. If loss is evident,
2990 * starts a new recovery (e.g. reordering then loss);
2992 tcp_enter_recovery(sk, ece_ack);
2996 tcp_process_loss(sk, flag, num_dupack, rexmit);
2997 tcp_identify_packet_loss(sk, ack_flag);
2998 if (!(icsk->icsk_ca_state == TCP_CA_Open ||
2999 (*ack_flag & FLAG_LOST_RETRANS)))
3001 /* Change state if cwnd is undone or retransmits are lost */
3004 if (tcp_is_reno(tp)) {
3005 if (flag & FLAG_SND_UNA_ADVANCED)
3006 tcp_reset_reno_sack(tp);
3007 tcp_add_reno_sack(sk, num_dupack, ece_ack);
3010 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
3011 tcp_try_undo_dsack(sk);
3013 tcp_identify_packet_loss(sk, ack_flag);
3014 if (!tcp_time_to_recover(sk, flag)) {
3015 tcp_try_to_open(sk, flag);
3019 /* MTU probe failure: don't reduce cwnd */
3020 if (icsk->icsk_ca_state < TCP_CA_CWR &&
3021 icsk->icsk_mtup.probe_size &&
3022 tp->snd_una == tp->mtu_probe.probe_seq_start) {
3023 tcp_mtup_probe_failed(sk);
3024 /* Restores the reduction we did in tcp_mtup_probe() */
3026 tcp_simple_retransmit(sk);
3030 /* Otherwise enter Recovery state */
3031 tcp_enter_recovery(sk, ece_ack);
3035 if (!tcp_is_rack(sk) && do_lost)
3036 tcp_update_scoreboard(sk, fast_rexmit);
3037 *rexmit = REXMIT_LOST;
3040 static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us, const int flag)
3042 u32 wlen = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_rtt_wlen) * HZ;
3043 struct tcp_sock *tp = tcp_sk(sk);
3045 if ((flag & FLAG_ACK_MAYBE_DELAYED) && rtt_us > tcp_min_rtt(tp)) {
3046 /* If the remote keeps returning delayed ACKs, eventually
3047 * the min filter would pick it up and overestimate the
3048 * prop. delay when it expires. Skip suspected delayed ACKs.
3052 minmax_running_min(&tp->rtt_min, wlen, tcp_jiffies32,
3053 rtt_us ? : jiffies_to_usecs(1));
3056 static bool tcp_ack_update_rtt(struct sock *sk, const int flag,
3057 long seq_rtt_us, long sack_rtt_us,
3058 long ca_rtt_us, struct rate_sample *rs)
3060 const struct tcp_sock *tp = tcp_sk(sk);
3062 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
3063 * broken middle-boxes or peers may corrupt TS-ECR fields. But
3064 * Karn's algorithm forbids taking RTT if some retransmitted data
3065 * is acked (RFC6298).
3068 seq_rtt_us = sack_rtt_us;
3070 /* RTTM Rule: A TSecr value received in a segment is used to
3071 * update the averaged RTT measurement only if the segment
3072 * acknowledges some new data, i.e., only if it advances the
3073 * left edge of the send window.
3074 * See draft-ietf-tcplw-high-performance-00, section 3.3.
3076 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
3077 flag & FLAG_ACKED) {
3078 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr;
3080 if (likely(delta < INT_MAX / (USEC_PER_SEC / TCP_TS_HZ))) {
3083 seq_rtt_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
3084 ca_rtt_us = seq_rtt_us;
3087 rs->rtt_us = ca_rtt_us; /* RTT of last (S)ACKed packet (or -1) */
3091 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
3092 * always taken together with ACK, SACK, or TS-opts. Any negative
3093 * values will be skipped with the seq_rtt_us < 0 check above.
3095 tcp_update_rtt_min(sk, ca_rtt_us, flag);
3096 tcp_rtt_estimator(sk, seq_rtt_us);
3099 /* RFC6298: only reset backoff on valid RTT measurement. */
3100 inet_csk(sk)->icsk_backoff = 0;
3104 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
3105 void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req)
3107 struct rate_sample rs;
3110 if (req && !req->num_retrans && tcp_rsk(req)->snt_synack)
3111 rtt_us = tcp_stamp_us_delta(tcp_clock_us(), tcp_rsk(req)->snt_synack);
3113 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us, &rs);
3117 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
3119 const struct inet_connection_sock *icsk = inet_csk(sk);
3121 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
3122 tcp_sk(sk)->snd_cwnd_stamp = tcp_jiffies32;
3125 /* Restart timer after forward progress on connection.
3126 * RFC2988 recommends to restart timer to now+rto.
3128 void tcp_rearm_rto(struct sock *sk)
3130 const struct inet_connection_sock *icsk = inet_csk(sk);
3131 struct tcp_sock *tp = tcp_sk(sk);
3133 /* If the retrans timer is currently being used by Fast Open
3134 * for SYN-ACK retrans purpose, stay put.
3136 if (rcu_access_pointer(tp->fastopen_rsk))
3139 if (!tp->packets_out) {
3140 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
3142 u32 rto = inet_csk(sk)->icsk_rto;
3143 /* Offset the time elapsed after installing regular RTO */
3144 if (icsk->icsk_pending == ICSK_TIME_REO_TIMEOUT ||
3145 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
3146 s64 delta_us = tcp_rto_delta_us(sk);
3147 /* delta_us may not be positive if the socket is locked
3148 * when the retrans timer fires and is rescheduled.
3150 rto = usecs_to_jiffies(max_t(int, delta_us, 1));
3152 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
3157 /* Try to schedule a loss probe; if that doesn't work, then schedule an RTO. */
3158 static void tcp_set_xmit_timer(struct sock *sk)
3160 if (!tcp_schedule_loss_probe(sk, true))
3164 /* If we get here, the whole TSO packet has not been acked. */
3165 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3167 struct tcp_sock *tp = tcp_sk(sk);
3170 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3172 packets_acked = tcp_skb_pcount(skb);
3173 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3175 packets_acked -= tcp_skb_pcount(skb);
3177 if (packets_acked) {
3178 BUG_ON(tcp_skb_pcount(skb) == 0);
3179 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3182 return packets_acked;
3185 static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
3188 const struct skb_shared_info *shinfo;
3190 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3191 if (likely(!TCP_SKB_CB(skb)->txstamp_ack))
3194 shinfo = skb_shinfo(skb);
3195 if (!before(shinfo->tskey, prior_snd_una) &&
3196 before(shinfo->tskey, tcp_sk(sk)->snd_una)) {
3197 tcp_skb_tsorted_save(skb) {
3198 __skb_tstamp_tx(skb, NULL, sk, SCM_TSTAMP_ACK);
3199 } tcp_skb_tsorted_restore(skb);
3203 /* Remove acknowledged frames from the retransmission queue. If our packet
3204 * is before the ack sequence we can discard it as it's confirmed to have
3205 * arrived at the other end.
3207 static int tcp_clean_rtx_queue(struct sock *sk, u32 prior_fack,
3209 struct tcp_sacktag_state *sack, bool ece_ack)
3211 const struct inet_connection_sock *icsk = inet_csk(sk);
3212 u64 first_ackt, last_ackt;
3213 struct tcp_sock *tp = tcp_sk(sk);
3214 u32 prior_sacked = tp->sacked_out;
3215 u32 reord = tp->snd_nxt; /* lowest acked un-retx un-sacked seq */
3216 struct sk_buff *skb, *next;
3217 bool fully_acked = true;
3218 long sack_rtt_us = -1L;
3219 long seq_rtt_us = -1L;
3220 long ca_rtt_us = -1L;
3222 u32 last_in_flight = 0;
3228 for (skb = skb_rb_first(&sk->tcp_rtx_queue); skb; skb = next) {
3229 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3230 const u32 start_seq = scb->seq;
3231 u8 sacked = scb->sacked;
3234 /* Determine how many packets and what bytes were acked, tso and else */
3235 if (after(scb->end_seq, tp->snd_una)) {
3236 if (tcp_skb_pcount(skb) == 1 ||
3237 !after(tp->snd_una, scb->seq))
3240 acked_pcount = tcp_tso_acked(sk, skb);
3243 fully_acked = false;
3245 acked_pcount = tcp_skb_pcount(skb);
3248 if (unlikely(sacked & TCPCB_RETRANS)) {
3249 if (sacked & TCPCB_SACKED_RETRANS)
3250 tp->retrans_out -= acked_pcount;
3251 flag |= FLAG_RETRANS_DATA_ACKED;
3252 } else if (!(sacked & TCPCB_SACKED_ACKED)) {
3253 last_ackt = tcp_skb_timestamp_us(skb);
3254 WARN_ON_ONCE(last_ackt == 0);
3256 first_ackt = last_ackt;
3258 last_in_flight = TCP_SKB_CB(skb)->tx.in_flight;
3259 if (before(start_seq, reord))
3261 if (!after(scb->end_seq, tp->high_seq))
3262 flag |= FLAG_ORIG_SACK_ACKED;
3265 if (sacked & TCPCB_SACKED_ACKED) {
3266 tp->sacked_out -= acked_pcount;
3267 } else if (tcp_is_sack(tp)) {
3268 tcp_count_delivered(tp, acked_pcount, ece_ack);
3269 if (!tcp_skb_spurious_retrans(tp, skb))
3270 tcp_rack_advance(tp, sacked, scb->end_seq,
3271 tcp_skb_timestamp_us(skb));
3273 if (sacked & TCPCB_LOST)
3274 tp->lost_out -= acked_pcount;
3276 tp->packets_out -= acked_pcount;
3277 pkts_acked += acked_pcount;
3278 tcp_rate_skb_delivered(sk, skb, sack->rate);
3280 /* Initial outgoing SYN's get put onto the write_queue
3281 * just like anything else we transmit. It is not
3282 * true data, and if we misinform our callers that
3283 * this ACK acks real data, we will erroneously exit
3284 * connection startup slow start one packet too
3285 * quickly. This is severely frowned upon behavior.
3287 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
3288 flag |= FLAG_DATA_ACKED;
3290 flag |= FLAG_SYN_ACKED;
3291 tp->retrans_stamp = 0;
3297 tcp_ack_tstamp(sk, skb, prior_snd_una);
3299 next = skb_rb_next(skb);
3300 if (unlikely(skb == tp->retransmit_skb_hint))
3301 tp->retransmit_skb_hint = NULL;
3302 if (unlikely(skb == tp->lost_skb_hint))
3303 tp->lost_skb_hint = NULL;
3304 tcp_highest_sack_replace(sk, skb, next);
3305 tcp_rtx_queue_unlink_and_free(skb, sk);
3309 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
3311 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3312 tp->snd_up = tp->snd_una;
3315 tcp_ack_tstamp(sk, skb, prior_snd_una);
3316 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
3317 flag |= FLAG_SACK_RENEGING;
3320 if (likely(first_ackt) && !(flag & FLAG_RETRANS_DATA_ACKED)) {
3321 seq_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, first_ackt);
3322 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, last_ackt);
3324 if (pkts_acked == 1 && last_in_flight < tp->mss_cache &&
3325 last_in_flight && !prior_sacked && fully_acked &&
3326 sack->rate->prior_delivered + 1 == tp->delivered &&
3327 !(flag & (FLAG_CA_ALERT | FLAG_SYN_ACKED))) {
3328 /* Conservatively mark a delayed ACK. It's typically
3329 * from a lone runt packet over the round trip to
3330 * a receiver w/o out-of-order or CE events.
3332 flag |= FLAG_ACK_MAYBE_DELAYED;
3335 if (sack->first_sackt) {
3336 sack_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->first_sackt);
3337 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->last_sackt);
3339 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us,
3340 ca_rtt_us, sack->rate);
3342 if (flag & FLAG_ACKED) {
3343 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3344 if (unlikely(icsk->icsk_mtup.probe_size &&
3345 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3346 tcp_mtup_probe_success(sk);
3349 if (tcp_is_reno(tp)) {
3350 tcp_remove_reno_sacks(sk, pkts_acked, ece_ack);
3352 /* If any of the cumulatively ACKed segments was
3353 * retransmitted, non-SACK case cannot confirm that
3354 * progress was due to original transmission due to
3355 * lack of TCPCB_SACKED_ACKED bits even if some of
3356 * the packets may have been never retransmitted.
3358 if (flag & FLAG_RETRANS_DATA_ACKED)
3359 flag &= ~FLAG_ORIG_SACK_ACKED;
3363 /* Non-retransmitted hole got filled? That's reordering */
3364 if (before(reord, prior_fack))
3365 tcp_check_sack_reordering(sk, reord, 0);
3367 delta = prior_sacked - tp->sacked_out;
3368 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3370 } else if (skb && rtt_update && sack_rtt_us >= 0 &&
3371 sack_rtt_us > tcp_stamp_us_delta(tp->tcp_mstamp,
3372 tcp_skb_timestamp_us(skb))) {
3373 /* Do not re-arm RTO if the sack RTT is measured from data sent
3374 * after when the head was last (re)transmitted. Otherwise the
3375 * timeout may continue to extend in loss recovery.
3377 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3380 if (icsk->icsk_ca_ops->pkts_acked) {
3381 struct ack_sample sample = { .pkts_acked = pkts_acked,
3382 .rtt_us = sack->rate->rtt_us,
3383 .in_flight = last_in_flight };
3385 icsk->icsk_ca_ops->pkts_acked(sk, &sample);
3388 #if FASTRETRANS_DEBUG > 0
3389 WARN_ON((int)tp->sacked_out < 0);
3390 WARN_ON((int)tp->lost_out < 0);
3391 WARN_ON((int)tp->retrans_out < 0);
3392 if (!tp->packets_out && tcp_is_sack(tp)) {
3393 icsk = inet_csk(sk);
3395 pr_debug("Leak l=%u %d\n",
3396 tp->lost_out, icsk->icsk_ca_state);
3399 if (tp->sacked_out) {
3400 pr_debug("Leak s=%u %d\n",
3401 tp->sacked_out, icsk->icsk_ca_state);
3404 if (tp->retrans_out) {
3405 pr_debug("Leak r=%u %d\n",
3406 tp->retrans_out, icsk->icsk_ca_state);
3407 tp->retrans_out = 0;
3414 static void tcp_ack_probe(struct sock *sk)
3416 struct inet_connection_sock *icsk = inet_csk(sk);
3417 struct sk_buff *head = tcp_send_head(sk);
3418 const struct tcp_sock *tp = tcp_sk(sk);
3420 /* Was it a usable window open? */
3423 if (!after(TCP_SKB_CB(head)->end_seq, tcp_wnd_end(tp))) {
3424 icsk->icsk_backoff = 0;
3425 icsk->icsk_probes_tstamp = 0;
3426 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3427 /* Socket must be waked up by subsequent tcp_data_snd_check().
3428 * This function is not for random using!
3431 unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX);
3433 when = tcp_clamp_probe0_to_user_timeout(sk, when);
3434 tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, when, TCP_RTO_MAX);
3438 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3440 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3441 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3444 /* Decide wheather to run the increase function of congestion control. */
3445 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3447 /* If reordering is high then always grow cwnd whenever data is
3448 * delivered regardless of its ordering. Otherwise stay conservative
3449 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3450 * new SACK or ECE mark may first advance cwnd here and later reduce
3451 * cwnd in tcp_fastretrans_alert() based on more states.
3453 if (tcp_sk(sk)->reordering >
3454 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_reordering))
3455 return flag & FLAG_FORWARD_PROGRESS;
3457 return flag & FLAG_DATA_ACKED;
3460 /* The "ultimate" congestion control function that aims to replace the rigid
3461 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction).
3462 * It's called toward the end of processing an ACK with precise rate
3463 * information. All transmission or retransmission are delayed afterwards.
3465 static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked,
3466 int flag, const struct rate_sample *rs)
3468 const struct inet_connection_sock *icsk = inet_csk(sk);
3470 if (icsk->icsk_ca_ops->cong_control) {
3471 icsk->icsk_ca_ops->cong_control(sk, rs);
3475 if (tcp_in_cwnd_reduction(sk)) {
3476 /* Reduce cwnd if state mandates */
3477 tcp_cwnd_reduction(sk, acked_sacked, flag);
3478 } else if (tcp_may_raise_cwnd(sk, flag)) {
3479 /* Advance cwnd if state allows */
3480 tcp_cong_avoid(sk, ack, acked_sacked);
3482 tcp_update_pacing_rate(sk);
3485 /* Check that window update is acceptable.
3486 * The function assumes that snd_una<=ack<=snd_next.
3488 static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3489 const u32 ack, const u32 ack_seq,
3492 return after(ack, tp->snd_una) ||
3493 after(ack_seq, tp->snd_wl1) ||
3494 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3497 /* If we update tp->snd_una, also update tp->bytes_acked */
3498 static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack)
3500 u32 delta = ack - tp->snd_una;
3502 sock_owned_by_me((struct sock *)tp);
3503 tp->bytes_acked += delta;
3507 /* If we update tp->rcv_nxt, also update tp->bytes_received */
3508 static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq)
3510 u32 delta = seq - tp->rcv_nxt;
3512 sock_owned_by_me((struct sock *)tp);
3513 tp->bytes_received += delta;
3514 WRITE_ONCE(tp->rcv_nxt, seq);
3517 /* Update our send window.
3519 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3520 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3522 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3525 struct tcp_sock *tp = tcp_sk(sk);
3527 u32 nwin = ntohs(tcp_hdr(skb)->window);
3529 if (likely(!tcp_hdr(skb)->syn))
3530 nwin <<= tp->rx_opt.snd_wscale;
3532 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3533 flag |= FLAG_WIN_UPDATE;
3534 tcp_update_wl(tp, ack_seq);
3536 if (tp->snd_wnd != nwin) {
3539 /* Note, it is the only place, where
3540 * fast path is recovered for sending TCP.
3543 tcp_fast_path_check(sk);
3545 if (!tcp_write_queue_empty(sk))
3546 tcp_slow_start_after_idle_check(sk);
3548 if (nwin > tp->max_window) {
3549 tp->max_window = nwin;
3550 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3555 tcp_snd_una_update(tp, ack);
3560 static bool __tcp_oow_rate_limited(struct net *net, int mib_idx,
3561 u32 *last_oow_ack_time)
3563 if (*last_oow_ack_time) {
3564 s32 elapsed = (s32)(tcp_jiffies32 - *last_oow_ack_time);
3567 elapsed < READ_ONCE(net->ipv4.sysctl_tcp_invalid_ratelimit)) {
3568 NET_INC_STATS(net, mib_idx);
3569 return true; /* rate-limited: don't send yet! */
3573 *last_oow_ack_time = tcp_jiffies32;
3575 return false; /* not rate-limited: go ahead, send dupack now! */
3578 /* Return true if we're currently rate-limiting out-of-window ACKs and
3579 * thus shouldn't send a dupack right now. We rate-limit dupacks in
3580 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3581 * attacks that send repeated SYNs or ACKs for the same connection. To
3582 * do this, we do not send a duplicate SYNACK or ACK if the remote
3583 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3585 bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb,
3586 int mib_idx, u32 *last_oow_ack_time)
3588 /* Data packets without SYNs are not likely part of an ACK loop. */
3589 if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) &&
3593 return __tcp_oow_rate_limited(net, mib_idx, last_oow_ack_time);
3596 /* RFC 5961 7 [ACK Throttling] */
3597 static void tcp_send_challenge_ack(struct sock *sk, const struct sk_buff *skb)
3599 /* unprotected vars, we dont care of overwrites */
3600 static u32 challenge_timestamp;
3601 static unsigned int challenge_count;
3602 struct tcp_sock *tp = tcp_sk(sk);
3603 struct net *net = sock_net(sk);
3606 /* First check our per-socket dupack rate limit. */
3607 if (__tcp_oow_rate_limited(net,
3608 LINUX_MIB_TCPACKSKIPPEDCHALLENGE,
3609 &tp->last_oow_ack_time))
3612 /* Then check host-wide RFC 5961 rate limit. */
3614 if (now != READ_ONCE(challenge_timestamp)) {
3615 u32 ack_limit = READ_ONCE(net->ipv4.sysctl_tcp_challenge_ack_limit);
3616 u32 half = (ack_limit + 1) >> 1;
3618 WRITE_ONCE(challenge_timestamp, now);
3619 WRITE_ONCE(challenge_count, half + prandom_u32_max(ack_limit));
3621 count = READ_ONCE(challenge_count);
3623 WRITE_ONCE(challenge_count, count - 1);
3624 NET_INC_STATS(net, LINUX_MIB_TCPCHALLENGEACK);
3629 static void tcp_store_ts_recent(struct tcp_sock *tp)
3631 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3632 tp->rx_opt.ts_recent_stamp = ktime_get_seconds();
3635 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3637 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3638 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3639 * extra check below makes sure this can only happen
3640 * for pure ACK frames. -DaveM
3642 * Not only, also it occurs for expired timestamps.
3645 if (tcp_paws_check(&tp->rx_opt, 0))
3646 tcp_store_ts_recent(tp);
3650 /* This routine deals with acks during a TLP episode and ends an episode by
3651 * resetting tlp_high_seq. Ref: TLP algorithm in draft-ietf-tcpm-rack
3653 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3655 struct tcp_sock *tp = tcp_sk(sk);
3657 if (before(ack, tp->tlp_high_seq))
3660 if (!tp->tlp_retrans) {
3661 /* TLP of new data has been acknowledged */
3662 tp->tlp_high_seq = 0;
3663 } else if (flag & FLAG_DSACKING_ACK) {
3664 /* This DSACK means original and TLP probe arrived; no loss */
3665 tp->tlp_high_seq = 0;
3666 } else if (after(ack, tp->tlp_high_seq)) {
3667 /* ACK advances: there was a loss, so reduce cwnd. Reset
3668 * tlp_high_seq in tcp_init_cwnd_reduction()
3670 tcp_init_cwnd_reduction(sk);
3671 tcp_set_ca_state(sk, TCP_CA_CWR);
3672 tcp_end_cwnd_reduction(sk);
3673 tcp_try_keep_open(sk);
3674 NET_INC_STATS(sock_net(sk),
3675 LINUX_MIB_TCPLOSSPROBERECOVERY);
3676 } else if (!(flag & (FLAG_SND_UNA_ADVANCED |
3677 FLAG_NOT_DUP | FLAG_DATA_SACKED))) {
3678 /* Pure dupack: original and TLP probe arrived; no loss */
3679 tp->tlp_high_seq = 0;
3683 static inline void tcp_in_ack_event(struct sock *sk, u32 flags)
3685 const struct inet_connection_sock *icsk = inet_csk(sk);
3687 if (icsk->icsk_ca_ops->in_ack_event)
3688 icsk->icsk_ca_ops->in_ack_event(sk, flags);
3691 /* Congestion control has updated the cwnd already. So if we're in
3692 * loss recovery then now we do any new sends (for FRTO) or
3693 * retransmits (for CA_Loss or CA_recovery) that make sense.
3695 static void tcp_xmit_recovery(struct sock *sk, int rexmit)
3697 struct tcp_sock *tp = tcp_sk(sk);
3699 if (rexmit == REXMIT_NONE || sk->sk_state == TCP_SYN_SENT)
3702 if (unlikely(rexmit == REXMIT_NEW)) {
3703 __tcp_push_pending_frames(sk, tcp_current_mss(sk),
3705 if (after(tp->snd_nxt, tp->high_seq))
3709 tcp_xmit_retransmit_queue(sk);
3712 /* Returns the number of packets newly acked or sacked by the current ACK */
3713 static u32 tcp_newly_delivered(struct sock *sk, u32 prior_delivered, int flag)
3715 const struct net *net = sock_net(sk);
3716 struct tcp_sock *tp = tcp_sk(sk);
3719 delivered = tp->delivered - prior_delivered;
3720 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVERED, delivered);
3721 if (flag & FLAG_ECE)
3722 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVEREDCE, delivered);
3727 /* This routine deals with incoming acks, but not outgoing ones. */
3728 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3730 struct inet_connection_sock *icsk = inet_csk(sk);
3731 struct tcp_sock *tp = tcp_sk(sk);
3732 struct tcp_sacktag_state sack_state;
3733 struct rate_sample rs = { .prior_delivered = 0 };
3734 u32 prior_snd_una = tp->snd_una;
3735 bool is_sack_reneg = tp->is_sack_reneg;
3736 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3737 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3739 int prior_packets = tp->packets_out;
3740 u32 delivered = tp->delivered;
3741 u32 lost = tp->lost;
3742 int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */
3745 sack_state.first_sackt = 0;
3746 sack_state.rate = &rs;
3747 sack_state.sack_delivered = 0;
3749 /* We very likely will need to access rtx queue. */
3750 prefetch(sk->tcp_rtx_queue.rb_node);
3752 /* If the ack is older than previous acks
3753 * then we can probably ignore it.
3755 if (before(ack, prior_snd_una)) {
3756 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3757 if (before(ack, prior_snd_una - tp->max_window)) {
3758 if (!(flag & FLAG_NO_CHALLENGE_ACK))
3759 tcp_send_challenge_ack(sk, skb);
3765 /* If the ack includes data we haven't sent yet, discard
3766 * this segment (RFC793 Section 3.9).
3768 if (after(ack, tp->snd_nxt))
3771 if (after(ack, prior_snd_una)) {
3772 flag |= FLAG_SND_UNA_ADVANCED;
3773 icsk->icsk_retransmits = 0;
3775 #if IS_ENABLED(CONFIG_TLS_DEVICE)
3776 if (static_branch_unlikely(&clean_acked_data_enabled.key))
3777 if (icsk->icsk_clean_acked)
3778 icsk->icsk_clean_acked(sk, ack);
3782 prior_fack = tcp_is_sack(tp) ? tcp_highest_sack_seq(tp) : tp->snd_una;
3783 rs.prior_in_flight = tcp_packets_in_flight(tp);
3785 /* ts_recent update must be made after we are sure that the packet
3788 if (flag & FLAG_UPDATE_TS_RECENT)
3789 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3791 if ((flag & (FLAG_SLOWPATH | FLAG_SND_UNA_ADVANCED)) ==
3792 FLAG_SND_UNA_ADVANCED) {
3793 /* Window is constant, pure forward advance.
3794 * No more checks are required.
3795 * Note, we use the fact that SND.UNA>=SND.WL2.
3797 tcp_update_wl(tp, ack_seq);
3798 tcp_snd_una_update(tp, ack);
3799 flag |= FLAG_WIN_UPDATE;
3801 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
3803 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPACKS);
3805 u32 ack_ev_flags = CA_ACK_SLOWPATH;
3807 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3810 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3812 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3814 if (TCP_SKB_CB(skb)->sacked)
3815 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3818 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) {
3820 ack_ev_flags |= CA_ACK_ECE;
3823 if (sack_state.sack_delivered)
3824 tcp_count_delivered(tp, sack_state.sack_delivered,
3827 if (flag & FLAG_WIN_UPDATE)
3828 ack_ev_flags |= CA_ACK_WIN_UPDATE;
3830 tcp_in_ack_event(sk, ack_ev_flags);
3833 /* This is a deviation from RFC3168 since it states that:
3834 * "When the TCP data sender is ready to set the CWR bit after reducing
3835 * the congestion window, it SHOULD set the CWR bit only on the first
3836 * new data packet that it transmits."
3837 * We accept CWR on pure ACKs to be more robust
3838 * with widely-deployed TCP implementations that do this.
3840 tcp_ecn_accept_cwr(sk, skb);
3842 /* We passed data and got it acked, remove any soft error
3843 * log. Something worked...
3845 sk->sk_err_soft = 0;
3846 icsk->icsk_probes_out = 0;
3847 tp->rcv_tstamp = tcp_jiffies32;
3851 /* See if we can take anything off of the retransmit queue. */
3852 flag |= tcp_clean_rtx_queue(sk, prior_fack, prior_snd_una, &sack_state,
3855 tcp_rack_update_reo_wnd(sk, &rs);
3857 if (tp->tlp_high_seq)
3858 tcp_process_tlp_ack(sk, ack, flag);
3860 if (tcp_ack_is_dubious(sk, flag)) {
3861 if (!(flag & (FLAG_SND_UNA_ADVANCED |
3862 FLAG_NOT_DUP | FLAG_DSACKING_ACK))) {
3864 /* Consider if pure acks were aggregated in tcp_add_backlog() */
3865 if (!(flag & FLAG_DATA))
3866 num_dupack = max_t(u16, 1, skb_shinfo(skb)->gso_segs);
3868 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
3872 /* If needed, reset TLP/RTO timer when RACK doesn't set. */
3873 if (flag & FLAG_SET_XMIT_TIMER)
3874 tcp_set_xmit_timer(sk);
3876 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
3879 delivered = tcp_newly_delivered(sk, delivered, flag);
3880 lost = tp->lost - lost; /* freshly marked lost */
3881 rs.is_ack_delayed = !!(flag & FLAG_ACK_MAYBE_DELAYED);
3882 tcp_rate_gen(sk, delivered, lost, is_sack_reneg, sack_state.rate);
3883 tcp_cong_control(sk, ack, delivered, flag, sack_state.rate);
3884 tcp_xmit_recovery(sk, rexmit);
3888 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3889 if (flag & FLAG_DSACKING_ACK) {
3890 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
3892 tcp_newly_delivered(sk, delivered, flag);
3894 /* If this ack opens up a zero window, clear backoff. It was
3895 * being used to time the probes, and is probably far higher than
3896 * it needs to be for normal retransmission.
3900 if (tp->tlp_high_seq)
3901 tcp_process_tlp_ack(sk, ack, flag);
3905 /* If data was SACKed, tag it and see if we should send more data.
3906 * If data was DSACKed, see if we can undo a cwnd reduction.
3908 if (TCP_SKB_CB(skb)->sacked) {
3909 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3911 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
3913 tcp_newly_delivered(sk, delivered, flag);
3914 tcp_xmit_recovery(sk, rexmit);
3920 static void tcp_parse_fastopen_option(int len, const unsigned char *cookie,
3921 bool syn, struct tcp_fastopen_cookie *foc,
3924 /* Valid only in SYN or SYN-ACK with an even length. */
3925 if (!foc || !syn || len < 0 || (len & 1))
3928 if (len >= TCP_FASTOPEN_COOKIE_MIN &&
3929 len <= TCP_FASTOPEN_COOKIE_MAX)
3930 memcpy(foc->val, cookie, len);
3937 static bool smc_parse_options(const struct tcphdr *th,
3938 struct tcp_options_received *opt_rx,
3939 const unsigned char *ptr,
3942 #if IS_ENABLED(CONFIG_SMC)
3943 if (static_branch_unlikely(&tcp_have_smc)) {
3944 if (th->syn && !(opsize & 1) &&
3945 opsize >= TCPOLEN_EXP_SMC_BASE &&
3946 get_unaligned_be32(ptr) == TCPOPT_SMC_MAGIC) {
3955 /* Try to parse the MSS option from the TCP header. Return 0 on failure, clamped
3958 static u16 tcp_parse_mss_option(const struct tcphdr *th, u16 user_mss)
3960 const unsigned char *ptr = (const unsigned char *)(th + 1);
3961 int length = (th->doff * 4) - sizeof(struct tcphdr);
3964 while (length > 0) {
3965 int opcode = *ptr++;
3971 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3978 if (opsize < 2) /* "silly options" */
3980 if (opsize > length)
3981 return mss; /* fail on partial options */
3982 if (opcode == TCPOPT_MSS && opsize == TCPOLEN_MSS) {
3983 u16 in_mss = get_unaligned_be16(ptr);
3986 if (user_mss && user_mss < in_mss)
3998 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3999 * But, this can also be called on packets in the established flow when
4000 * the fast version below fails.
4002 void tcp_parse_options(const struct net *net,
4003 const struct sk_buff *skb,
4004 struct tcp_options_received *opt_rx, int estab,
4005 struct tcp_fastopen_cookie *foc)
4007 const unsigned char *ptr;
4008 const struct tcphdr *th = tcp_hdr(skb);
4009 int length = (th->doff * 4) - sizeof(struct tcphdr);
4011 ptr = (const unsigned char *)(th + 1);
4012 opt_rx->saw_tstamp = 0;
4013 opt_rx->saw_unknown = 0;
4015 while (length > 0) {
4016 int opcode = *ptr++;
4022 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
4029 if (opsize < 2) /* "silly options" */
4031 if (opsize > length)
4032 return; /* don't parse partial options */
4035 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
4036 u16 in_mss = get_unaligned_be16(ptr);
4038 if (opt_rx->user_mss &&
4039 opt_rx->user_mss < in_mss)
4040 in_mss = opt_rx->user_mss;
4041 opt_rx->mss_clamp = in_mss;
4046 if (opsize == TCPOLEN_WINDOW && th->syn &&
4047 !estab && READ_ONCE(net->ipv4.sysctl_tcp_window_scaling)) {
4048 __u8 snd_wscale = *(__u8 *)ptr;
4049 opt_rx->wscale_ok = 1;
4050 if (snd_wscale > TCP_MAX_WSCALE) {
4051 net_info_ratelimited("%s: Illegal window scaling value %d > %u received\n",
4055 snd_wscale = TCP_MAX_WSCALE;
4057 opt_rx->snd_wscale = snd_wscale;
4060 case TCPOPT_TIMESTAMP:
4061 if ((opsize == TCPOLEN_TIMESTAMP) &&
4062 ((estab && opt_rx->tstamp_ok) ||
4063 (!estab && READ_ONCE(net->ipv4.sysctl_tcp_timestamps)))) {
4064 opt_rx->saw_tstamp = 1;
4065 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
4066 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
4069 case TCPOPT_SACK_PERM:
4070 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
4071 !estab && READ_ONCE(net->ipv4.sysctl_tcp_sack)) {
4072 opt_rx->sack_ok = TCP_SACK_SEEN;
4073 tcp_sack_reset(opt_rx);
4078 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
4079 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
4081 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
4084 #ifdef CONFIG_TCP_MD5SIG
4087 * The MD5 Hash has already been
4088 * checked (see tcp_v{4,6}_do_rcv()).
4092 case TCPOPT_FASTOPEN:
4093 tcp_parse_fastopen_option(
4094 opsize - TCPOLEN_FASTOPEN_BASE,
4095 ptr, th->syn, foc, false);
4099 /* Fast Open option shares code 254 using a
4100 * 16 bits magic number.
4102 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE &&
4103 get_unaligned_be16(ptr) ==
4104 TCPOPT_FASTOPEN_MAGIC) {
4105 tcp_parse_fastopen_option(opsize -
4106 TCPOLEN_EXP_FASTOPEN_BASE,
4107 ptr + 2, th->syn, foc, true);
4111 if (smc_parse_options(th, opt_rx, ptr, opsize))
4114 opt_rx->saw_unknown = 1;
4118 opt_rx->saw_unknown = 1;
4125 EXPORT_SYMBOL(tcp_parse_options);
4127 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
4129 const __be32 *ptr = (const __be32 *)(th + 1);
4131 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
4132 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
4133 tp->rx_opt.saw_tstamp = 1;
4135 tp->rx_opt.rcv_tsval = ntohl(*ptr);
4138 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
4140 tp->rx_opt.rcv_tsecr = 0;
4146 /* Fast parse options. This hopes to only see timestamps.
4147 * If it is wrong it falls back on tcp_parse_options().
4149 static bool tcp_fast_parse_options(const struct net *net,
4150 const struct sk_buff *skb,
4151 const struct tcphdr *th, struct tcp_sock *tp)
4153 /* In the spirit of fast parsing, compare doff directly to constant
4154 * values. Because equality is used, short doff can be ignored here.
4156 if (th->doff == (sizeof(*th) / 4)) {
4157 tp->rx_opt.saw_tstamp = 0;
4159 } else if (tp->rx_opt.tstamp_ok &&
4160 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
4161 if (tcp_parse_aligned_timestamp(tp, th))
4165 tcp_parse_options(net, skb, &tp->rx_opt, 1, NULL);
4166 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
4167 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
4172 #ifdef CONFIG_TCP_MD5SIG
4174 * Parse MD5 Signature option
4176 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
4178 int length = (th->doff << 2) - sizeof(*th);
4179 const u8 *ptr = (const u8 *)(th + 1);
4181 /* If not enough data remaining, we can short cut */
4182 while (length >= TCPOLEN_MD5SIG) {
4183 int opcode = *ptr++;
4194 if (opsize < 2 || opsize > length)
4196 if (opcode == TCPOPT_MD5SIG)
4197 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
4204 EXPORT_SYMBOL(tcp_parse_md5sig_option);
4207 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
4209 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
4210 * it can pass through stack. So, the following predicate verifies that
4211 * this segment is not used for anything but congestion avoidance or
4212 * fast retransmit. Moreover, we even are able to eliminate most of such
4213 * second order effects, if we apply some small "replay" window (~RTO)
4214 * to timestamp space.
4216 * All these measures still do not guarantee that we reject wrapped ACKs
4217 * on networks with high bandwidth, when sequence space is recycled fastly,
4218 * but it guarantees that such events will be very rare and do not affect
4219 * connection seriously. This doesn't look nice, but alas, PAWS is really
4222 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
4223 * states that events when retransmit arrives after original data are rare.
4224 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
4225 * the biggest problem on large power networks even with minor reordering.
4226 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
4227 * up to bandwidth of 18Gigabit/sec. 8) ]
4230 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
4232 const struct tcp_sock *tp = tcp_sk(sk);
4233 const struct tcphdr *th = tcp_hdr(skb);
4234 u32 seq = TCP_SKB_CB(skb)->seq;
4235 u32 ack = TCP_SKB_CB(skb)->ack_seq;
4237 return (/* 1. Pure ACK with correct sequence number. */
4238 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
4240 /* 2. ... and duplicate ACK. */
4241 ack == tp->snd_una &&
4243 /* 3. ... and does not update window. */
4244 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
4246 /* 4. ... and sits in replay window. */
4247 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
4250 static inline bool tcp_paws_discard(const struct sock *sk,
4251 const struct sk_buff *skb)
4253 const struct tcp_sock *tp = tcp_sk(sk);
4255 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
4256 !tcp_disordered_ack(sk, skb);
4259 /* Check segment sequence number for validity.
4261 * Segment controls are considered valid, if the segment
4262 * fits to the window after truncation to the window. Acceptability
4263 * of data (and SYN, FIN, of course) is checked separately.
4264 * See tcp_data_queue(), for example.
4266 * Also, controls (RST is main one) are accepted using RCV.WUP instead
4267 * of RCV.NXT. Peer still did not advance his SND.UNA when we
4268 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4269 * (borrowed from freebsd)
4272 static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
4274 return !before(end_seq, tp->rcv_wup) &&
4275 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
4278 /* When we get a reset we do this. */
4279 void tcp_reset(struct sock *sk)
4281 trace_tcp_receive_reset(sk);
4283 /* We want the right error as BSD sees it (and indeed as we do). */
4284 switch (sk->sk_state) {
4286 sk->sk_err = ECONNREFUSED;
4288 case TCP_CLOSE_WAIT:
4294 sk->sk_err = ECONNRESET;
4296 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4299 tcp_write_queue_purge(sk);
4302 if (!sock_flag(sk, SOCK_DEAD))
4303 sk->sk_error_report(sk);
4307 * Process the FIN bit. This now behaves as it is supposed to work
4308 * and the FIN takes effect when it is validly part of sequence
4309 * space. Not before when we get holes.
4311 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4312 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4315 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4316 * close and we go into CLOSING (and later onto TIME-WAIT)
4318 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4320 void tcp_fin(struct sock *sk)
4322 struct tcp_sock *tp = tcp_sk(sk);
4324 inet_csk_schedule_ack(sk);
4326 sk->sk_shutdown |= RCV_SHUTDOWN;
4327 sock_set_flag(sk, SOCK_DONE);
4329 switch (sk->sk_state) {
4331 case TCP_ESTABLISHED:
4332 /* Move to CLOSE_WAIT */
4333 tcp_set_state(sk, TCP_CLOSE_WAIT);
4334 inet_csk_enter_pingpong_mode(sk);
4337 case TCP_CLOSE_WAIT:
4339 /* Received a retransmission of the FIN, do
4344 /* RFC793: Remain in the LAST-ACK state. */
4348 /* This case occurs when a simultaneous close
4349 * happens, we must ack the received FIN and
4350 * enter the CLOSING state.
4353 tcp_set_state(sk, TCP_CLOSING);
4356 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4358 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4361 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4362 * cases we should never reach this piece of code.
4364 pr_err("%s: Impossible, sk->sk_state=%d\n",
4365 __func__, sk->sk_state);
4369 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4370 * Probably, we should reset in this case. For now drop them.
4372 skb_rbtree_purge(&tp->out_of_order_queue);
4373 if (tcp_is_sack(tp))
4374 tcp_sack_reset(&tp->rx_opt);
4377 if (!sock_flag(sk, SOCK_DEAD)) {
4378 sk->sk_state_change(sk);
4380 /* Do not send POLL_HUP for half duplex close. */
4381 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4382 sk->sk_state == TCP_CLOSE)
4383 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4385 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4389 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4392 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4393 if (before(seq, sp->start_seq))
4394 sp->start_seq = seq;
4395 if (after(end_seq, sp->end_seq))
4396 sp->end_seq = end_seq;
4402 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4404 struct tcp_sock *tp = tcp_sk(sk);
4406 if (tcp_is_sack(tp) && READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_dsack)) {
4409 if (before(seq, tp->rcv_nxt))
4410 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4412 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4414 NET_INC_STATS(sock_net(sk), mib_idx);
4416 tp->rx_opt.dsack = 1;
4417 tp->duplicate_sack[0].start_seq = seq;
4418 tp->duplicate_sack[0].end_seq = end_seq;
4422 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4424 struct tcp_sock *tp = tcp_sk(sk);
4426 if (!tp->rx_opt.dsack)
4427 tcp_dsack_set(sk, seq, end_seq);
4429 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4432 static void tcp_rcv_spurious_retrans(struct sock *sk, const struct sk_buff *skb)
4434 /* When the ACK path fails or drops most ACKs, the sender would
4435 * timeout and spuriously retransmit the same segment repeatedly.
4436 * The receiver remembers and reflects via DSACKs. Leverage the
4437 * DSACK state and change the txhash to re-route speculatively.
4439 if (TCP_SKB_CB(skb)->seq == tcp_sk(sk)->duplicate_sack[0].start_seq &&
4440 sk_rethink_txhash(sk))
4441 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDUPLICATEDATAREHASH);
4444 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4446 struct tcp_sock *tp = tcp_sk(sk);
4448 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4449 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4450 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4451 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
4453 if (tcp_is_sack(tp) && READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_dsack)) {
4454 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4456 tcp_rcv_spurious_retrans(sk, skb);
4457 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4458 end_seq = tp->rcv_nxt;
4459 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4466 /* These routines update the SACK block as out-of-order packets arrive or
4467 * in-order packets close up the sequence space.
4469 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4472 struct tcp_sack_block *sp = &tp->selective_acks[0];
4473 struct tcp_sack_block *swalk = sp + 1;
4475 /* See if the recent change to the first SACK eats into
4476 * or hits the sequence space of other SACK blocks, if so coalesce.
4478 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4479 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4482 /* Zap SWALK, by moving every further SACK up by one slot.
4483 * Decrease num_sacks.
4485 tp->rx_opt.num_sacks--;
4486 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4495 static void tcp_sack_compress_send_ack(struct sock *sk)
4497 struct tcp_sock *tp = tcp_sk(sk);
4499 if (!tp->compressed_ack)
4502 if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1)
4505 /* Since we have to send one ack finally,
4506 * substract one from tp->compressed_ack to keep
4507 * LINUX_MIB_TCPACKCOMPRESSED accurate.
4509 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED,
4510 tp->compressed_ack - 1);
4512 tp->compressed_ack = 0;
4516 /* Reasonable amount of sack blocks included in TCP SACK option
4517 * The max is 4, but this becomes 3 if TCP timestamps are there.
4518 * Given that SACK packets might be lost, be conservative and use 2.
4520 #define TCP_SACK_BLOCKS_EXPECTED 2
4522 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4524 struct tcp_sock *tp = tcp_sk(sk);
4525 struct tcp_sack_block *sp = &tp->selective_acks[0];
4526 int cur_sacks = tp->rx_opt.num_sacks;
4532 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4533 if (tcp_sack_extend(sp, seq, end_seq)) {
4534 if (this_sack >= TCP_SACK_BLOCKS_EXPECTED)
4535 tcp_sack_compress_send_ack(sk);
4536 /* Rotate this_sack to the first one. */
4537 for (; this_sack > 0; this_sack--, sp--)
4538 swap(*sp, *(sp - 1));
4540 tcp_sack_maybe_coalesce(tp);
4545 if (this_sack >= TCP_SACK_BLOCKS_EXPECTED)
4546 tcp_sack_compress_send_ack(sk);
4548 /* Could not find an adjacent existing SACK, build a new one,
4549 * put it at the front, and shift everyone else down. We
4550 * always know there is at least one SACK present already here.
4552 * If the sack array is full, forget about the last one.
4554 if (this_sack >= TCP_NUM_SACKS) {
4556 tp->rx_opt.num_sacks--;
4559 for (; this_sack > 0; this_sack--, sp--)
4563 /* Build the new head SACK, and we're done. */
4564 sp->start_seq = seq;
4565 sp->end_seq = end_seq;
4566 tp->rx_opt.num_sacks++;
4569 /* RCV.NXT advances, some SACKs should be eaten. */
4571 static void tcp_sack_remove(struct tcp_sock *tp)
4573 struct tcp_sack_block *sp = &tp->selective_acks[0];
4574 int num_sacks = tp->rx_opt.num_sacks;
4577 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4578 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4579 tp->rx_opt.num_sacks = 0;
4583 for (this_sack = 0; this_sack < num_sacks;) {
4584 /* Check if the start of the sack is covered by RCV.NXT. */
4585 if (!before(tp->rcv_nxt, sp->start_seq)) {
4588 /* RCV.NXT must cover all the block! */
4589 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4591 /* Zap this SACK, by moving forward any other SACKS. */
4592 for (i = this_sack+1; i < num_sacks; i++)
4593 tp->selective_acks[i-1] = tp->selective_acks[i];
4600 tp->rx_opt.num_sacks = num_sacks;
4604 * tcp_try_coalesce - try to merge skb to prior one
4607 * @from: buffer to add in queue
4608 * @fragstolen: pointer to boolean
4610 * Before queueing skb @from after @to, try to merge them
4611 * to reduce overall memory use and queue lengths, if cost is small.
4612 * Packets in ofo or receive queues can stay a long time.
4613 * Better try to coalesce them right now to avoid future collapses.
4614 * Returns true if caller should free @from instead of queueing it
4616 static bool tcp_try_coalesce(struct sock *sk,
4618 struct sk_buff *from,
4623 *fragstolen = false;
4625 /* Its possible this segment overlaps with prior segment in queue */
4626 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4629 if (!mptcp_skb_can_collapse(to, from))
4632 #ifdef CONFIG_TLS_DEVICE
4633 if (from->decrypted != to->decrypted)
4637 if (!skb_try_coalesce(to, from, fragstolen, &delta))
4640 atomic_add(delta, &sk->sk_rmem_alloc);
4641 sk_mem_charge(sk, delta);
4642 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4643 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4644 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4645 TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
4647 if (TCP_SKB_CB(from)->has_rxtstamp) {
4648 TCP_SKB_CB(to)->has_rxtstamp = true;
4649 to->tstamp = from->tstamp;
4650 skb_hwtstamps(to)->hwtstamp = skb_hwtstamps(from)->hwtstamp;
4656 static bool tcp_ooo_try_coalesce(struct sock *sk,
4658 struct sk_buff *from,
4661 bool res = tcp_try_coalesce(sk, to, from, fragstolen);
4663 /* In case tcp_drop() is called later, update to->gso_segs */
4665 u32 gso_segs = max_t(u16, 1, skb_shinfo(to)->gso_segs) +
4666 max_t(u16, 1, skb_shinfo(from)->gso_segs);
4668 skb_shinfo(to)->gso_segs = min_t(u32, gso_segs, 0xFFFF);
4673 static void tcp_drop(struct sock *sk, struct sk_buff *skb)
4675 sk_drops_add(sk, skb);
4679 /* This one checks to see if we can put data from the
4680 * out_of_order queue into the receive_queue.
4682 static void tcp_ofo_queue(struct sock *sk)
4684 struct tcp_sock *tp = tcp_sk(sk);
4685 __u32 dsack_high = tp->rcv_nxt;
4686 bool fin, fragstolen, eaten;
4687 struct sk_buff *skb, *tail;
4690 p = rb_first(&tp->out_of_order_queue);
4693 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4696 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4697 __u32 dsack = dsack_high;
4698 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4699 dsack_high = TCP_SKB_CB(skb)->end_seq;
4700 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4703 rb_erase(&skb->rbnode, &tp->out_of_order_queue);
4705 if (unlikely(!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))) {
4710 tail = skb_peek_tail(&sk->sk_receive_queue);
4711 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen);
4712 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4713 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
4715 __skb_queue_tail(&sk->sk_receive_queue, skb);
4717 kfree_skb_partial(skb, fragstolen);
4719 if (unlikely(fin)) {
4721 /* tcp_fin() purges tp->out_of_order_queue,
4722 * so we must end this loop right now.
4729 static bool tcp_prune_ofo_queue(struct sock *sk);
4730 static int tcp_prune_queue(struct sock *sk);
4732 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4735 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4736 !sk_rmem_schedule(sk, skb, size)) {
4738 if (tcp_prune_queue(sk) < 0)
4741 while (!sk_rmem_schedule(sk, skb, size)) {
4742 if (!tcp_prune_ofo_queue(sk))
4749 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4751 struct tcp_sock *tp = tcp_sk(sk);
4752 struct rb_node **p, *parent;
4753 struct sk_buff *skb1;
4757 tcp_ecn_check_ce(sk, skb);
4759 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4760 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFODROP);
4761 sk->sk_data_ready(sk);
4766 /* Disable header prediction. */
4768 inet_csk_schedule_ack(sk);
4770 tp->rcv_ooopack += max_t(u16, 1, skb_shinfo(skb)->gso_segs);
4771 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4772 seq = TCP_SKB_CB(skb)->seq;
4773 end_seq = TCP_SKB_CB(skb)->end_seq;
4775 p = &tp->out_of_order_queue.rb_node;
4776 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4777 /* Initial out of order segment, build 1 SACK. */
4778 if (tcp_is_sack(tp)) {
4779 tp->rx_opt.num_sacks = 1;
4780 tp->selective_acks[0].start_seq = seq;
4781 tp->selective_acks[0].end_seq = end_seq;
4783 rb_link_node(&skb->rbnode, NULL, p);
4784 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4785 tp->ooo_last_skb = skb;
4789 /* In the typical case, we are adding an skb to the end of the list.
4790 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
4792 if (tcp_ooo_try_coalesce(sk, tp->ooo_last_skb,
4793 skb, &fragstolen)) {
4795 /* For non sack flows, do not grow window to force DUPACK
4796 * and trigger fast retransmit.
4798 if (tcp_is_sack(tp))
4799 tcp_grow_window(sk, skb, true);
4800 kfree_skb_partial(skb, fragstolen);
4804 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
4805 if (!before(seq, TCP_SKB_CB(tp->ooo_last_skb)->end_seq)) {
4806 parent = &tp->ooo_last_skb->rbnode;
4807 p = &parent->rb_right;
4811 /* Find place to insert this segment. Handle overlaps on the way. */
4815 skb1 = rb_to_skb(parent);
4816 if (before(seq, TCP_SKB_CB(skb1)->seq)) {
4817 p = &parent->rb_left;
4820 if (before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4821 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4822 /* All the bits are present. Drop. */
4823 NET_INC_STATS(sock_net(sk),
4824 LINUX_MIB_TCPOFOMERGE);
4827 tcp_dsack_set(sk, seq, end_seq);
4830 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4831 /* Partial overlap. */
4832 tcp_dsack_set(sk, seq, TCP_SKB_CB(skb1)->end_seq);
4834 /* skb's seq == skb1's seq and skb covers skb1.
4835 * Replace skb1 with skb.
4837 rb_replace_node(&skb1->rbnode, &skb->rbnode,
4838 &tp->out_of_order_queue);
4839 tcp_dsack_extend(sk,
4840 TCP_SKB_CB(skb1)->seq,
4841 TCP_SKB_CB(skb1)->end_seq);
4842 NET_INC_STATS(sock_net(sk),
4843 LINUX_MIB_TCPOFOMERGE);
4847 } else if (tcp_ooo_try_coalesce(sk, skb1,
4848 skb, &fragstolen)) {
4851 p = &parent->rb_right;
4854 /* Insert segment into RB tree. */
4855 rb_link_node(&skb->rbnode, parent, p);
4856 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4859 /* Remove other segments covered by skb. */
4860 while ((skb1 = skb_rb_next(skb)) != NULL) {
4861 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4863 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4864 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4868 rb_erase(&skb1->rbnode, &tp->out_of_order_queue);
4869 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4870 TCP_SKB_CB(skb1)->end_seq);
4871 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4874 /* If there is no skb after us, we are the last_skb ! */
4876 tp->ooo_last_skb = skb;
4879 if (tcp_is_sack(tp))
4880 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4883 /* For non sack flows, do not grow window to force DUPACK
4884 * and trigger fast retransmit.
4886 if (tcp_is_sack(tp))
4887 tcp_grow_window(sk, skb, false);
4889 skb_set_owner_r(skb, sk);
4893 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb,
4897 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4900 tcp_try_coalesce(sk, tail,
4901 skb, fragstolen)) ? 1 : 0;
4902 tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq);
4904 __skb_queue_tail(&sk->sk_receive_queue, skb);
4905 skb_set_owner_r(skb, sk);
4910 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4912 struct sk_buff *skb;
4920 if (size > PAGE_SIZE) {
4921 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS);
4923 data_len = npages << PAGE_SHIFT;
4924 size = data_len + (size & ~PAGE_MASK);
4926 skb = alloc_skb_with_frags(size - data_len, data_len,
4927 PAGE_ALLOC_COSTLY_ORDER,
4928 &err, sk->sk_allocation);
4932 skb_put(skb, size - data_len);
4933 skb->data_len = data_len;
4936 if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) {
4937 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP);
4941 err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size);
4945 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4946 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4947 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4949 if (tcp_queue_rcv(sk, skb, &fragstolen)) {
4950 WARN_ON_ONCE(fragstolen); /* should not happen */
4962 void tcp_data_ready(struct sock *sk)
4964 const struct tcp_sock *tp = tcp_sk(sk);
4965 int avail = tp->rcv_nxt - tp->copied_seq;
4967 if (avail < sk->sk_rcvlowat && !tcp_rmem_pressure(sk) &&
4968 !sock_flag(sk, SOCK_DONE) &&
4969 tcp_receive_window(tp) > inet_csk(sk)->icsk_ack.rcv_mss)
4972 sk->sk_data_ready(sk);
4975 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4977 struct tcp_sock *tp = tcp_sk(sk);
4981 if (sk_is_mptcp(sk))
4982 mptcp_incoming_options(sk, skb);
4984 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
4989 __skb_pull(skb, tcp_hdr(skb)->doff * 4);
4991 tp->rx_opt.dsack = 0;
4993 /* Queue data for delivery to the user.
4994 * Packets in sequence go to the receive queue.
4995 * Out of sequence packets to the out_of_order_queue.
4997 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4998 if (tcp_receive_window(tp) == 0) {
4999 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP);
5003 /* Ok. In sequence. In window. */
5005 if (skb_queue_len(&sk->sk_receive_queue) == 0)
5006 sk_forced_mem_schedule(sk, skb->truesize);
5007 else if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) {
5008 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP);
5009 sk->sk_data_ready(sk);
5013 eaten = tcp_queue_rcv(sk, skb, &fragstolen);
5015 tcp_event_data_recv(sk, skb);
5016 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
5019 if (!RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
5022 /* RFC5681. 4.2. SHOULD send immediate ACK, when
5023 * gap in queue is filled.
5025 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
5026 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW;
5029 if (tp->rx_opt.num_sacks)
5030 tcp_sack_remove(tp);
5032 tcp_fast_path_check(sk);
5035 kfree_skb_partial(skb, fragstolen);
5036 if (!sock_flag(sk, SOCK_DEAD))
5041 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
5042 tcp_rcv_spurious_retrans(sk, skb);
5043 /* A retransmit, 2nd most common case. Force an immediate ack. */
5044 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
5045 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
5048 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
5049 inet_csk_schedule_ack(sk);
5055 /* Out of window. F.e. zero window probe. */
5056 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
5059 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
5060 /* Partial packet, seq < rcv_next < end_seq */
5061 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
5063 /* If window is closed, drop tail of packet. But after
5064 * remembering D-SACK for its head made in previous line.
5066 if (!tcp_receive_window(tp)) {
5067 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP);
5073 tcp_data_queue_ofo(sk, skb);
5076 static struct sk_buff *tcp_skb_next(struct sk_buff *skb, struct sk_buff_head *list)
5079 return !skb_queue_is_last(list, skb) ? skb->next : NULL;
5081 return skb_rb_next(skb);
5084 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
5085 struct sk_buff_head *list,
5086 struct rb_root *root)
5088 struct sk_buff *next = tcp_skb_next(skb, list);
5091 __skb_unlink(skb, list);
5093 rb_erase(&skb->rbnode, root);
5096 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
5101 /* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */
5102 void tcp_rbtree_insert(struct rb_root *root, struct sk_buff *skb)
5104 struct rb_node **p = &root->rb_node;
5105 struct rb_node *parent = NULL;
5106 struct sk_buff *skb1;
5110 skb1 = rb_to_skb(parent);
5111 if (before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb1)->seq))
5112 p = &parent->rb_left;
5114 p = &parent->rb_right;
5116 rb_link_node(&skb->rbnode, parent, p);
5117 rb_insert_color(&skb->rbnode, root);
5120 /* Collapse contiguous sequence of skbs head..tail with
5121 * sequence numbers start..end.
5123 * If tail is NULL, this means until the end of the queue.
5125 * Segments with FIN/SYN are not collapsed (only because this
5129 tcp_collapse(struct sock *sk, struct sk_buff_head *list, struct rb_root *root,
5130 struct sk_buff *head, struct sk_buff *tail, u32 start, u32 end)
5132 struct sk_buff *skb = head, *n;
5133 struct sk_buff_head tmp;
5136 /* First, check that queue is collapsible and find
5137 * the point where collapsing can be useful.
5140 for (end_of_skbs = true; skb != NULL && skb != tail; skb = n) {
5141 n = tcp_skb_next(skb, list);
5143 /* No new bits? It is possible on ofo queue. */
5144 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
5145 skb = tcp_collapse_one(sk, skb, list, root);
5151 /* The first skb to collapse is:
5153 * - bloated or contains data before "start" or
5154 * overlaps to the next one and mptcp allow collapsing.
5156 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
5157 (tcp_win_from_space(sk, skb->truesize) > skb->len ||
5158 before(TCP_SKB_CB(skb)->seq, start))) {
5159 end_of_skbs = false;
5163 if (n && n != tail && mptcp_skb_can_collapse(skb, n) &&
5164 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(n)->seq) {
5165 end_of_skbs = false;
5169 /* Decided to skip this, advance start seq. */
5170 start = TCP_SKB_CB(skb)->end_seq;
5173 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
5176 __skb_queue_head_init(&tmp);
5178 while (before(start, end)) {
5179 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
5180 struct sk_buff *nskb;
5182 nskb = alloc_skb(copy, GFP_ATOMIC);
5186 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
5187 #ifdef CONFIG_TLS_DEVICE
5188 nskb->decrypted = skb->decrypted;
5190 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
5192 __skb_queue_before(list, skb, nskb);
5194 __skb_queue_tail(&tmp, nskb); /* defer rbtree insertion */
5195 skb_set_owner_r(nskb, sk);
5196 mptcp_skb_ext_move(nskb, skb);
5198 /* Copy data, releasing collapsed skbs. */
5200 int offset = start - TCP_SKB_CB(skb)->seq;
5201 int size = TCP_SKB_CB(skb)->end_seq - start;
5205 size = min(copy, size);
5206 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
5208 TCP_SKB_CB(nskb)->end_seq += size;
5212 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
5213 skb = tcp_collapse_one(sk, skb, list, root);
5216 !mptcp_skb_can_collapse(nskb, skb) ||
5217 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
5219 #ifdef CONFIG_TLS_DEVICE
5220 if (skb->decrypted != nskb->decrypted)
5227 skb_queue_walk_safe(&tmp, skb, n)
5228 tcp_rbtree_insert(root, skb);
5231 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
5232 * and tcp_collapse() them until all the queue is collapsed.
5234 static void tcp_collapse_ofo_queue(struct sock *sk)
5236 struct tcp_sock *tp = tcp_sk(sk);
5237 u32 range_truesize, sum_tiny = 0;
5238 struct sk_buff *skb, *head;
5241 skb = skb_rb_first(&tp->out_of_order_queue);
5244 tp->ooo_last_skb = skb_rb_last(&tp->out_of_order_queue);
5247 start = TCP_SKB_CB(skb)->seq;
5248 end = TCP_SKB_CB(skb)->end_seq;
5249 range_truesize = skb->truesize;
5251 for (head = skb;;) {
5252 skb = skb_rb_next(skb);
5254 /* Range is terminated when we see a gap or when
5255 * we are at the queue end.
5258 after(TCP_SKB_CB(skb)->seq, end) ||
5259 before(TCP_SKB_CB(skb)->end_seq, start)) {
5260 /* Do not attempt collapsing tiny skbs */
5261 if (range_truesize != head->truesize ||
5262 end - start >= SKB_WITH_OVERHEAD(SK_MEM_QUANTUM)) {
5263 tcp_collapse(sk, NULL, &tp->out_of_order_queue,
5264 head, skb, start, end);
5266 sum_tiny += range_truesize;
5267 if (sum_tiny > sk->sk_rcvbuf >> 3)
5273 range_truesize += skb->truesize;
5274 if (unlikely(before(TCP_SKB_CB(skb)->seq, start)))
5275 start = TCP_SKB_CB(skb)->seq;
5276 if (after(TCP_SKB_CB(skb)->end_seq, end))
5277 end = TCP_SKB_CB(skb)->end_seq;
5282 * Clean the out-of-order queue to make room.
5283 * We drop high sequences packets to :
5284 * 1) Let a chance for holes to be filled.
5285 * 2) not add too big latencies if thousands of packets sit there.
5286 * (But if application shrinks SO_RCVBUF, we could still end up
5287 * freeing whole queue here)
5288 * 3) Drop at least 12.5 % of sk_rcvbuf to avoid malicious attacks.
5290 * Return true if queue has shrunk.
5292 static bool tcp_prune_ofo_queue(struct sock *sk)
5294 struct tcp_sock *tp = tcp_sk(sk);
5295 struct rb_node *node, *prev;
5298 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
5301 NET_INC_STATS(sock_net(sk), LINUX_MIB_OFOPRUNED);
5302 goal = sk->sk_rcvbuf >> 3;
5303 node = &tp->ooo_last_skb->rbnode;
5305 prev = rb_prev(node);
5306 rb_erase(node, &tp->out_of_order_queue);
5307 goal -= rb_to_skb(node)->truesize;
5308 tcp_drop(sk, rb_to_skb(node));
5309 if (!prev || goal <= 0) {
5311 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf &&
5312 !tcp_under_memory_pressure(sk))
5314 goal = sk->sk_rcvbuf >> 3;
5318 tp->ooo_last_skb = rb_to_skb(prev);
5320 /* Reset SACK state. A conforming SACK implementation will
5321 * do the same at a timeout based retransmit. When a connection
5322 * is in a sad state like this, we care only about integrity
5323 * of the connection not performance.
5325 if (tp->rx_opt.sack_ok)
5326 tcp_sack_reset(&tp->rx_opt);
5330 /* Reduce allocated memory if we can, trying to get
5331 * the socket within its memory limits again.
5333 * Return less than zero if we should start dropping frames
5334 * until the socket owning process reads some of the data
5335 * to stabilize the situation.
5337 static int tcp_prune_queue(struct sock *sk)
5339 struct tcp_sock *tp = tcp_sk(sk);
5341 NET_INC_STATS(sock_net(sk), LINUX_MIB_PRUNECALLED);
5343 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
5344 tcp_clamp_window(sk);
5345 else if (tcp_under_memory_pressure(sk))
5346 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
5348 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5351 tcp_collapse_ofo_queue(sk);
5352 if (!skb_queue_empty(&sk->sk_receive_queue))
5353 tcp_collapse(sk, &sk->sk_receive_queue, NULL,
5354 skb_peek(&sk->sk_receive_queue),
5356 tp->copied_seq, tp->rcv_nxt);
5359 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5362 /* Collapsing did not help, destructive actions follow.
5363 * This must not ever occur. */
5365 tcp_prune_ofo_queue(sk);
5367 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5370 /* If we are really being abused, tell the caller to silently
5371 * drop receive data on the floor. It will get retransmitted
5372 * and hopefully then we'll have sufficient space.
5374 NET_INC_STATS(sock_net(sk), LINUX_MIB_RCVPRUNED);
5376 /* Massive buffer overcommit. */
5381 static bool tcp_should_expand_sndbuf(const struct sock *sk)
5383 const struct tcp_sock *tp = tcp_sk(sk);
5385 /* If the user specified a specific send buffer setting, do
5388 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
5391 /* If we are under global TCP memory pressure, do not expand. */
5392 if (tcp_under_memory_pressure(sk))
5395 /* If we are under soft global TCP memory pressure, do not expand. */
5396 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
5399 /* If we filled the congestion window, do not expand. */
5400 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
5406 static void tcp_new_space(struct sock *sk)
5408 struct tcp_sock *tp = tcp_sk(sk);
5410 if (tcp_should_expand_sndbuf(sk)) {
5411 tcp_sndbuf_expand(sk);
5412 tp->snd_cwnd_stamp = tcp_jiffies32;
5415 sk->sk_write_space(sk);
5418 /* Caller made space either from:
5419 * 1) Freeing skbs in rtx queues (after tp->snd_una has advanced)
5420 * 2) Sent skbs from output queue (and thus advancing tp->snd_nxt)
5422 * We might be able to generate EPOLLOUT to the application if:
5423 * 1) Space consumed in output/rtx queues is below sk->sk_sndbuf/2
5424 * 2) notsent amount (tp->write_seq - tp->snd_nxt) became
5425 * small enough that tcp_stream_memory_free() decides it
5426 * is time to generate EPOLLOUT.
5428 void tcp_check_space(struct sock *sk)
5430 /* pairs with tcp_poll() */
5432 if (sk->sk_socket &&
5433 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
5435 if (!test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
5436 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
5440 static inline void tcp_data_snd_check(struct sock *sk)
5442 tcp_push_pending_frames(sk);
5443 tcp_check_space(sk);
5447 * Check if sending an ack is needed.
5449 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
5451 struct tcp_sock *tp = tcp_sk(sk);
5452 unsigned long rtt, delay;
5454 /* More than one full frame received... */
5455 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
5456 /* ... and right edge of window advances far enough.
5457 * (tcp_recvmsg() will send ACK otherwise).
5458 * If application uses SO_RCVLOWAT, we want send ack now if
5459 * we have not received enough bytes to satisfy the condition.
5461 (tp->rcv_nxt - tp->copied_seq < sk->sk_rcvlowat ||
5462 __tcp_select_window(sk) >= tp->rcv_wnd)) ||
5463 /* We ACK each frame or... */
5464 tcp_in_quickack_mode(sk) ||
5465 /* Protocol state mandates a one-time immediate ACK */
5466 inet_csk(sk)->icsk_ack.pending & ICSK_ACK_NOW) {
5472 if (!ofo_possible || RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
5473 tcp_send_delayed_ack(sk);
5477 if (!tcp_is_sack(tp) ||
5478 tp->compressed_ack >= READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_nr))
5481 if (tp->compressed_ack_rcv_nxt != tp->rcv_nxt) {
5482 tp->compressed_ack_rcv_nxt = tp->rcv_nxt;
5483 tp->dup_ack_counter = 0;
5485 if (tp->dup_ack_counter < TCP_FASTRETRANS_THRESH) {
5486 tp->dup_ack_counter++;
5489 tp->compressed_ack++;
5490 if (hrtimer_is_queued(&tp->compressed_ack_timer))
5493 /* compress ack timer : 5 % of rtt, but no more than tcp_comp_sack_delay_ns */
5495 rtt = tp->rcv_rtt_est.rtt_us;
5496 if (tp->srtt_us && tp->srtt_us < rtt)
5499 delay = min_t(unsigned long,
5500 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_delay_ns),
5501 rtt * (NSEC_PER_USEC >> 3)/20);
5503 hrtimer_start_range_ns(&tp->compressed_ack_timer, ns_to_ktime(delay),
5504 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_slack_ns),
5505 HRTIMER_MODE_REL_PINNED_SOFT);
5508 static inline void tcp_ack_snd_check(struct sock *sk)
5510 if (!inet_csk_ack_scheduled(sk)) {
5511 /* We sent a data segment already. */
5514 __tcp_ack_snd_check(sk, 1);
5518 * This routine is only called when we have urgent data
5519 * signaled. Its the 'slow' part of tcp_urg. It could be
5520 * moved inline now as tcp_urg is only called from one
5521 * place. We handle URGent data wrong. We have to - as
5522 * BSD still doesn't use the correction from RFC961.
5523 * For 1003.1g we should support a new option TCP_STDURG to permit
5524 * either form (or just set the sysctl tcp_stdurg).
5527 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
5529 struct tcp_sock *tp = tcp_sk(sk);
5530 u32 ptr = ntohs(th->urg_ptr);
5532 if (ptr && !READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_stdurg))
5534 ptr += ntohl(th->seq);
5536 /* Ignore urgent data that we've already seen and read. */
5537 if (after(tp->copied_seq, ptr))
5540 /* Do not replay urg ptr.
5542 * NOTE: interesting situation not covered by specs.
5543 * Misbehaving sender may send urg ptr, pointing to segment,
5544 * which we already have in ofo queue. We are not able to fetch
5545 * such data and will stay in TCP_URG_NOTYET until will be eaten
5546 * by recvmsg(). Seems, we are not obliged to handle such wicked
5547 * situations. But it is worth to think about possibility of some
5548 * DoSes using some hypothetical application level deadlock.
5550 if (before(ptr, tp->rcv_nxt))
5553 /* Do we already have a newer (or duplicate) urgent pointer? */
5554 if (tp->urg_data && !after(ptr, tp->urg_seq))
5557 /* Tell the world about our new urgent pointer. */
5560 /* We may be adding urgent data when the last byte read was
5561 * urgent. To do this requires some care. We cannot just ignore
5562 * tp->copied_seq since we would read the last urgent byte again
5563 * as data, nor can we alter copied_seq until this data arrives
5564 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5566 * NOTE. Double Dutch. Rendering to plain English: author of comment
5567 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5568 * and expect that both A and B disappear from stream. This is _wrong_.
5569 * Though this happens in BSD with high probability, this is occasional.
5570 * Any application relying on this is buggy. Note also, that fix "works"
5571 * only in this artificial test. Insert some normal data between A and B and we will
5572 * decline of BSD again. Verdict: it is better to remove to trap
5575 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5576 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5577 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5579 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5580 __skb_unlink(skb, &sk->sk_receive_queue);
5585 tp->urg_data = TCP_URG_NOTYET;
5586 WRITE_ONCE(tp->urg_seq, ptr);
5588 /* Disable header prediction. */
5592 /* This is the 'fast' part of urgent handling. */
5593 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
5595 struct tcp_sock *tp = tcp_sk(sk);
5597 /* Check if we get a new urgent pointer - normally not. */
5599 tcp_check_urg(sk, th);
5601 /* Do we wait for any urgent data? - normally not... */
5602 if (tp->urg_data == TCP_URG_NOTYET) {
5603 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5606 /* Is the urgent pointer pointing into this packet? */
5607 if (ptr < skb->len) {
5609 if (skb_copy_bits(skb, ptr, &tmp, 1))
5611 tp->urg_data = TCP_URG_VALID | tmp;
5612 if (!sock_flag(sk, SOCK_DEAD))
5613 sk->sk_data_ready(sk);
5618 /* Accept RST for rcv_nxt - 1 after a FIN.
5619 * When tcp connections are abruptly terminated from Mac OSX (via ^C), a
5620 * FIN is sent followed by a RST packet. The RST is sent with the same
5621 * sequence number as the FIN, and thus according to RFC 5961 a challenge
5622 * ACK should be sent. However, Mac OSX rate limits replies to challenge
5623 * ACKs on the closed socket. In addition middleboxes can drop either the
5624 * challenge ACK or a subsequent RST.
5626 static bool tcp_reset_check(const struct sock *sk, const struct sk_buff *skb)
5628 struct tcp_sock *tp = tcp_sk(sk);
5630 return unlikely(TCP_SKB_CB(skb)->seq == (tp->rcv_nxt - 1) &&
5631 (1 << sk->sk_state) & (TCPF_CLOSE_WAIT | TCPF_LAST_ACK |
5635 /* Does PAWS and seqno based validation of an incoming segment, flags will
5636 * play significant role here.
5638 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5639 const struct tcphdr *th, int syn_inerr)
5641 struct tcp_sock *tp = tcp_sk(sk);
5642 bool rst_seq_match = false;
5644 /* RFC1323: H1. Apply PAWS check first. */
5645 if (tcp_fast_parse_options(sock_net(sk), skb, th, tp) &&
5646 tp->rx_opt.saw_tstamp &&
5647 tcp_paws_discard(sk, skb)) {
5649 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5650 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5651 LINUX_MIB_TCPACKSKIPPEDPAWS,
5652 &tp->last_oow_ack_time))
5653 tcp_send_dupack(sk, skb);
5656 /* Reset is accepted even if it did not pass PAWS. */
5659 /* Step 1: check sequence number */
5660 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5661 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5662 * (RST) segments are validated by checking their SEQ-fields."
5663 * And page 69: "If an incoming segment is not acceptable,
5664 * an acknowledgment should be sent in reply (unless the RST
5665 * bit is set, if so drop the segment and return)".
5670 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5671 LINUX_MIB_TCPACKSKIPPEDSEQ,
5672 &tp->last_oow_ack_time))
5673 tcp_send_dupack(sk, skb);
5674 } else if (tcp_reset_check(sk, skb)) {
5680 /* Step 2: check RST bit */
5682 /* RFC 5961 3.2 (extend to match against (RCV.NXT - 1) after a
5683 * FIN and SACK too if available):
5684 * If seq num matches RCV.NXT or (RCV.NXT - 1) after a FIN, or
5685 * the right-most SACK block,
5687 * RESET the connection
5689 * Send a challenge ACK
5691 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt ||
5692 tcp_reset_check(sk, skb)) {
5693 rst_seq_match = true;
5694 } else if (tcp_is_sack(tp) && tp->rx_opt.num_sacks > 0) {
5695 struct tcp_sack_block *sp = &tp->selective_acks[0];
5696 int max_sack = sp[0].end_seq;
5699 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;
5701 max_sack = after(sp[this_sack].end_seq,
5703 sp[this_sack].end_seq : max_sack;
5706 if (TCP_SKB_CB(skb)->seq == max_sack)
5707 rst_seq_match = true;
5713 /* Disable TFO if RST is out-of-order
5714 * and no data has been received
5715 * for current active TFO socket
5717 if (tp->syn_fastopen && !tp->data_segs_in &&
5718 sk->sk_state == TCP_ESTABLISHED)
5719 tcp_fastopen_active_disable(sk);
5720 tcp_send_challenge_ack(sk, skb);
5725 /* step 3: check security and precedence [ignored] */
5727 /* step 4: Check for a SYN
5728 * RFC 5961 4.2 : Send a challenge ack
5733 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5734 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5735 tcp_send_challenge_ack(sk, skb);
5739 bpf_skops_parse_hdr(sk, skb);
5749 * TCP receive function for the ESTABLISHED state.
5751 * It is split into a fast path and a slow path. The fast path is
5753 * - A zero window was announced from us - zero window probing
5754 * is only handled properly in the slow path.
5755 * - Out of order segments arrived.
5756 * - Urgent data is expected.
5757 * - There is no buffer space left
5758 * - Unexpected TCP flags/window values/header lengths are received
5759 * (detected by checking the TCP header against pred_flags)
5760 * - Data is sent in both directions. Fast path only supports pure senders
5761 * or pure receivers (this means either the sequence number or the ack
5762 * value must stay constant)
5763 * - Unexpected TCP option.
5765 * When these conditions are not satisfied it drops into a standard
5766 * receive procedure patterned after RFC793 to handle all cases.
5767 * The first three cases are guaranteed by proper pred_flags setting,
5768 * the rest is checked inline. Fast processing is turned on in
5769 * tcp_data_queue when everything is OK.
5771 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb)
5773 const struct tcphdr *th = (const struct tcphdr *)skb->data;
5774 struct tcp_sock *tp = tcp_sk(sk);
5775 unsigned int len = skb->len;
5777 /* TCP congestion window tracking */
5778 trace_tcp_probe(sk, skb);
5780 tcp_mstamp_refresh(tp);
5781 if (unlikely(!rcu_access_pointer(sk->sk_rx_dst)))
5782 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
5784 * Header prediction.
5785 * The code loosely follows the one in the famous
5786 * "30 instruction TCP receive" Van Jacobson mail.
5788 * Van's trick is to deposit buffers into socket queue
5789 * on a device interrupt, to call tcp_recv function
5790 * on the receive process context and checksum and copy
5791 * the buffer to user space. smart...
5793 * Our current scheme is not silly either but we take the
5794 * extra cost of the net_bh soft interrupt processing...
5795 * We do checksum and copy also but from device to kernel.
5798 tp->rx_opt.saw_tstamp = 0;
5800 /* pred_flags is 0xS?10 << 16 + snd_wnd
5801 * if header_prediction is to be made
5802 * 'S' will always be tp->tcp_header_len >> 2
5803 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5804 * turn it off (when there are holes in the receive
5805 * space for instance)
5806 * PSH flag is ignored.
5809 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5810 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5811 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5812 int tcp_header_len = tp->tcp_header_len;
5814 /* Timestamp header prediction: tcp_header_len
5815 * is automatically equal to th->doff*4 due to pred_flags
5819 /* Check timestamp */
5820 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5821 /* No? Slow path! */
5822 if (!tcp_parse_aligned_timestamp(tp, th))
5825 /* If PAWS failed, check it more carefully in slow path */
5826 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5829 /* DO NOT update ts_recent here, if checksum fails
5830 * and timestamp was corrupted part, it will result
5831 * in a hung connection since we will drop all
5832 * future packets due to the PAWS test.
5836 if (len <= tcp_header_len) {
5837 /* Bulk data transfer: sender */
5838 if (len == tcp_header_len) {
5839 /* Predicted packet is in window by definition.
5840 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5841 * Hence, check seq<=rcv_wup reduces to:
5843 if (tcp_header_len ==
5844 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5845 tp->rcv_nxt == tp->rcv_wup)
5846 tcp_store_ts_recent(tp);
5848 /* We know that such packets are checksummed
5851 tcp_ack(sk, skb, 0);
5853 tcp_data_snd_check(sk);
5854 /* When receiving pure ack in fast path, update
5855 * last ts ecr directly instead of calling
5856 * tcp_rcv_rtt_measure_ts()
5858 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr;
5860 } else { /* Header too small */
5861 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5866 bool fragstolen = false;
5868 if (tcp_checksum_complete(skb))
5871 if ((int)skb->truesize > sk->sk_forward_alloc)
5874 /* Predicted packet is in window by definition.
5875 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5876 * Hence, check seq<=rcv_wup reduces to:
5878 if (tcp_header_len ==
5879 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5880 tp->rcv_nxt == tp->rcv_wup)
5881 tcp_store_ts_recent(tp);
5883 tcp_rcv_rtt_measure_ts(sk, skb);
5885 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPHITS);
5887 /* Bulk data transfer: receiver */
5888 __skb_pull(skb, tcp_header_len);
5889 eaten = tcp_queue_rcv(sk, skb, &fragstolen);
5891 tcp_event_data_recv(sk, skb);
5893 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5894 /* Well, only one small jumplet in fast path... */
5895 tcp_ack(sk, skb, FLAG_DATA);
5896 tcp_data_snd_check(sk);
5897 if (!inet_csk_ack_scheduled(sk))
5900 tcp_update_wl(tp, TCP_SKB_CB(skb)->seq);
5903 __tcp_ack_snd_check(sk, 0);
5906 kfree_skb_partial(skb, fragstolen);
5913 if (len < (th->doff << 2) || tcp_checksum_complete(skb))
5916 if (!th->ack && !th->rst && !th->syn)
5920 * Standard slow path.
5923 if (!tcp_validate_incoming(sk, skb, th, 1))
5927 if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0)
5930 tcp_rcv_rtt_measure_ts(sk, skb);
5932 /* Process urgent data. */
5933 tcp_urg(sk, skb, th);
5935 /* step 7: process the segment text */
5936 tcp_data_queue(sk, skb);
5938 tcp_data_snd_check(sk);
5939 tcp_ack_snd_check(sk);
5943 TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS);
5944 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5949 EXPORT_SYMBOL(tcp_rcv_established);
5951 void tcp_init_transfer(struct sock *sk, int bpf_op, struct sk_buff *skb)
5953 struct inet_connection_sock *icsk = inet_csk(sk);
5954 struct tcp_sock *tp = tcp_sk(sk);
5957 icsk->icsk_af_ops->rebuild_header(sk);
5958 tcp_init_metrics(sk);
5960 /* Initialize the congestion window to start the transfer.
5961 * Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
5962 * retransmitted. In light of RFC6298 more aggressive 1sec
5963 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
5964 * retransmission has occurred.
5966 if (tp->total_retrans > 1 && tp->undo_marker)
5969 tp->snd_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk));
5970 tp->snd_cwnd_stamp = tcp_jiffies32;
5972 bpf_skops_established(sk, bpf_op, skb);
5973 /* Initialize congestion control unless BPF initialized it already: */
5974 if (!icsk->icsk_ca_initialized)
5975 tcp_init_congestion_control(sk);
5976 tcp_init_buffer_space(sk);
5979 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5981 struct tcp_sock *tp = tcp_sk(sk);
5982 struct inet_connection_sock *icsk = inet_csk(sk);
5984 tcp_set_state(sk, TCP_ESTABLISHED);
5985 icsk->icsk_ack.lrcvtime = tcp_jiffies32;
5988 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
5989 security_inet_conn_established(sk, skb);
5990 sk_mark_napi_id(sk, skb);
5993 tcp_init_transfer(sk, BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB, skb);
5995 /* Prevent spurious tcp_cwnd_restart() on first data
5998 tp->lsndtime = tcp_jiffies32;
6000 if (sock_flag(sk, SOCK_KEEPOPEN))
6001 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
6003 if (!tp->rx_opt.snd_wscale)
6004 __tcp_fast_path_on(tp, tp->snd_wnd);
6009 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
6010 struct tcp_fastopen_cookie *cookie)
6012 struct tcp_sock *tp = tcp_sk(sk);
6013 struct sk_buff *data = tp->syn_data ? tcp_rtx_queue_head(sk) : NULL;
6014 u16 mss = tp->rx_opt.mss_clamp, try_exp = 0;
6015 bool syn_drop = false;
6017 if (mss == tp->rx_opt.user_mss) {
6018 struct tcp_options_received opt;
6020 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
6021 tcp_clear_options(&opt);
6022 opt.user_mss = opt.mss_clamp = 0;
6023 tcp_parse_options(sock_net(sk), synack, &opt, 0, NULL);
6024 mss = opt.mss_clamp;
6027 if (!tp->syn_fastopen) {
6028 /* Ignore an unsolicited cookie */
6030 } else if (tp->total_retrans) {
6031 /* SYN timed out and the SYN-ACK neither has a cookie nor
6032 * acknowledges data. Presumably the remote received only
6033 * the retransmitted (regular) SYNs: either the original
6034 * SYN-data or the corresponding SYN-ACK was dropped.
6036 syn_drop = (cookie->len < 0 && data);
6037 } else if (cookie->len < 0 && !tp->syn_data) {
6038 /* We requested a cookie but didn't get it. If we did not use
6039 * the (old) exp opt format then try so next time (try_exp=1).
6040 * Otherwise we go back to use the RFC7413 opt (try_exp=2).
6042 try_exp = tp->syn_fastopen_exp ? 2 : 1;
6045 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp);
6047 if (data) { /* Retransmit unacked data in SYN */
6048 if (tp->total_retrans)
6049 tp->fastopen_client_fail = TFO_SYN_RETRANSMITTED;
6051 tp->fastopen_client_fail = TFO_DATA_NOT_ACKED;
6052 skb_rbtree_walk_from(data) {
6053 if (__tcp_retransmit_skb(sk, data, 1))
6057 NET_INC_STATS(sock_net(sk),
6058 LINUX_MIB_TCPFASTOPENACTIVEFAIL);
6061 tp->syn_data_acked = tp->syn_data;
6062 if (tp->syn_data_acked) {
6063 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
6064 /* SYN-data is counted as two separate packets in tcp_ack() */
6065 if (tp->delivered > 1)
6069 tcp_fastopen_add_skb(sk, synack);
6074 static void smc_check_reset_syn(struct tcp_sock *tp)
6076 #if IS_ENABLED(CONFIG_SMC)
6077 if (static_branch_unlikely(&tcp_have_smc)) {
6078 if (tp->syn_smc && !tp->rx_opt.smc_ok)
6084 static void tcp_try_undo_spurious_syn(struct sock *sk)
6086 struct tcp_sock *tp = tcp_sk(sk);
6089 /* undo_marker is set when SYN or SYNACK times out. The timeout is
6090 * spurious if the ACK's timestamp option echo value matches the
6091 * original SYN timestamp.
6093 syn_stamp = tp->retrans_stamp;
6094 if (tp->undo_marker && syn_stamp && tp->rx_opt.saw_tstamp &&
6095 syn_stamp == tp->rx_opt.rcv_tsecr)
6096 tp->undo_marker = 0;
6099 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
6100 const struct tcphdr *th)
6102 struct inet_connection_sock *icsk = inet_csk(sk);
6103 struct tcp_sock *tp = tcp_sk(sk);
6104 struct tcp_fastopen_cookie foc = { .len = -1 };
6105 int saved_clamp = tp->rx_opt.mss_clamp;
6108 tcp_parse_options(sock_net(sk), skb, &tp->rx_opt, 0, &foc);
6109 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
6110 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
6114 * "If the state is SYN-SENT then
6115 * first check the ACK bit
6116 * If the ACK bit is set
6117 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
6118 * a reset (unless the RST bit is set, if so drop
6119 * the segment and return)"
6121 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
6122 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
6123 /* Previous FIN/ACK or RST/ACK might be ignored. */
6124 if (icsk->icsk_retransmits == 0)
6125 inet_csk_reset_xmit_timer(sk,
6127 TCP_TIMEOUT_MIN, TCP_RTO_MAX);
6128 goto reset_and_undo;
6131 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
6132 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
6133 tcp_time_stamp(tp))) {
6134 NET_INC_STATS(sock_net(sk),
6135 LINUX_MIB_PAWSACTIVEREJECTED);
6136 goto reset_and_undo;
6139 /* Now ACK is acceptable.
6141 * "If the RST bit is set
6142 * If the ACK was acceptable then signal the user "error:
6143 * connection reset", drop the segment, enter CLOSED state,
6144 * delete TCB, and return."
6153 * "fifth, if neither of the SYN or RST bits is set then
6154 * drop the segment and return."
6160 goto discard_and_undo;
6163 * "If the SYN bit is on ...
6164 * are acceptable then ...
6165 * (our SYN has been ACKed), change the connection
6166 * state to ESTABLISHED..."
6169 tcp_ecn_rcv_synack(tp, th);
6171 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
6172 tcp_try_undo_spurious_syn(sk);
6173 tcp_ack(sk, skb, FLAG_SLOWPATH);
6175 /* Ok.. it's good. Set up sequence numbers and
6176 * move to established.
6178 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1);
6179 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
6181 /* RFC1323: The window in SYN & SYN/ACK segments is
6184 tp->snd_wnd = ntohs(th->window);
6186 if (!tp->rx_opt.wscale_ok) {
6187 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
6188 tp->window_clamp = min(tp->window_clamp, 65535U);
6191 if (tp->rx_opt.saw_tstamp) {
6192 tp->rx_opt.tstamp_ok = 1;
6193 tp->tcp_header_len =
6194 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
6195 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
6196 tcp_store_ts_recent(tp);
6198 tp->tcp_header_len = sizeof(struct tcphdr);
6201 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
6202 tcp_initialize_rcv_mss(sk);
6204 /* Remember, tcp_poll() does not lock socket!
6205 * Change state from SYN-SENT only after copied_seq
6206 * is initialized. */
6207 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6209 smc_check_reset_syn(tp);
6213 tcp_finish_connect(sk, skb);
6215 fastopen_fail = (tp->syn_fastopen || tp->syn_data) &&
6216 tcp_rcv_fastopen_synack(sk, skb, &foc);
6218 if (!sock_flag(sk, SOCK_DEAD)) {
6219 sk->sk_state_change(sk);
6220 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
6224 if (sk->sk_write_pending ||
6225 icsk->icsk_accept_queue.rskq_defer_accept ||
6226 inet_csk_in_pingpong_mode(sk)) {
6227 /* Save one ACK. Data will be ready after
6228 * several ticks, if write_pending is set.
6230 * It may be deleted, but with this feature tcpdumps
6231 * look so _wonderfully_ clever, that I was not able
6232 * to stand against the temptation 8) --ANK
6234 inet_csk_schedule_ack(sk);
6235 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
6236 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
6237 TCP_DELACK_MAX, TCP_RTO_MAX);
6248 /* No ACK in the segment */
6252 * "If the RST bit is set
6254 * Otherwise (no ACK) drop the segment and return."
6257 goto discard_and_undo;
6261 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
6262 tcp_paws_reject(&tp->rx_opt, 0))
6263 goto discard_and_undo;
6266 /* We see SYN without ACK. It is attempt of
6267 * simultaneous connect with crossed SYNs.
6268 * Particularly, it can be connect to self.
6270 tcp_set_state(sk, TCP_SYN_RECV);
6272 if (tp->rx_opt.saw_tstamp) {
6273 tp->rx_opt.tstamp_ok = 1;
6274 tcp_store_ts_recent(tp);
6275 tp->tcp_header_len =
6276 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
6278 tp->tcp_header_len = sizeof(struct tcphdr);
6281 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1);
6282 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6283 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
6285 /* RFC1323: The window in SYN & SYN/ACK segments is
6288 tp->snd_wnd = ntohs(th->window);
6289 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
6290 tp->max_window = tp->snd_wnd;
6292 tcp_ecn_rcv_syn(tp, th);
6295 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
6296 tcp_initialize_rcv_mss(sk);
6298 tcp_send_synack(sk);
6300 /* Note, we could accept data and URG from this segment.
6301 * There are no obstacles to make this (except that we must
6302 * either change tcp_recvmsg() to prevent it from returning data
6303 * before 3WHS completes per RFC793, or employ TCP Fast Open).
6305 * However, if we ignore data in ACKless segments sometimes,
6306 * we have no reasons to accept it sometimes.
6307 * Also, seems the code doing it in step6 of tcp_rcv_state_process
6308 * is not flawless. So, discard packet for sanity.
6309 * Uncomment this return to process the data.
6316 /* "fifth, if neither of the SYN or RST bits is set then
6317 * drop the segment and return."
6321 tcp_clear_options(&tp->rx_opt);
6322 tp->rx_opt.mss_clamp = saved_clamp;
6326 tcp_clear_options(&tp->rx_opt);
6327 tp->rx_opt.mss_clamp = saved_clamp;
6331 static void tcp_rcv_synrecv_state_fastopen(struct sock *sk)
6333 struct request_sock *req;
6335 /* If we are still handling the SYNACK RTO, see if timestamp ECR allows
6336 * undo. If peer SACKs triggered fast recovery, we can't undo here.
6338 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
6339 tcp_try_undo_loss(sk, false);
6341 /* Reset rtx states to prevent spurious retransmits_timed_out() */
6342 tcp_sk(sk)->retrans_stamp = 0;
6343 inet_csk(sk)->icsk_retransmits = 0;
6345 /* Once we leave TCP_SYN_RECV or TCP_FIN_WAIT_1,
6346 * we no longer need req so release it.
6348 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk,
6349 lockdep_sock_is_held(sk));
6350 reqsk_fastopen_remove(sk, req, false);
6352 /* Re-arm the timer because data may have been sent out.
6353 * This is similar to the regular data transmission case
6354 * when new data has just been ack'ed.
6356 * (TFO) - we could try to be more aggressive and
6357 * retransmitting any data sooner based on when they
6364 * This function implements the receiving procedure of RFC 793 for
6365 * all states except ESTABLISHED and TIME_WAIT.
6366 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
6367 * address independent.
6370 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb)
6372 struct tcp_sock *tp = tcp_sk(sk);
6373 struct inet_connection_sock *icsk = inet_csk(sk);
6374 const struct tcphdr *th = tcp_hdr(skb);
6375 struct request_sock *req;
6379 switch (sk->sk_state) {
6393 /* It is possible that we process SYN packets from backlog,
6394 * so we need to make sure to disable BH and RCU right there.
6398 acceptable = icsk->icsk_af_ops->conn_request(sk, skb) >= 0;
6410 tp->rx_opt.saw_tstamp = 0;
6411 tcp_mstamp_refresh(tp);
6412 queued = tcp_rcv_synsent_state_process(sk, skb, th);
6416 /* Do step6 onward by hand. */
6417 tcp_urg(sk, skb, th);
6419 tcp_data_snd_check(sk);
6423 tcp_mstamp_refresh(tp);
6424 tp->rx_opt.saw_tstamp = 0;
6425 req = rcu_dereference_protected(tp->fastopen_rsk,
6426 lockdep_sock_is_held(sk));
6430 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
6431 sk->sk_state != TCP_FIN_WAIT1);
6433 if (!tcp_check_req(sk, skb, req, true, &req_stolen))
6437 if (!th->ack && !th->rst && !th->syn)
6440 if (!tcp_validate_incoming(sk, skb, th, 0))
6443 /* step 5: check the ACK field */
6444 acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
6445 FLAG_UPDATE_TS_RECENT |
6446 FLAG_NO_CHALLENGE_ACK) > 0;
6449 if (sk->sk_state == TCP_SYN_RECV)
6450 return 1; /* send one RST */
6451 tcp_send_challenge_ack(sk, skb);
6454 switch (sk->sk_state) {
6456 tp->delivered++; /* SYN-ACK delivery isn't tracked in tcp_ack */
6458 tcp_synack_rtt_meas(sk, req);
6461 tcp_rcv_synrecv_state_fastopen(sk);
6463 tcp_try_undo_spurious_syn(sk);
6464 tp->retrans_stamp = 0;
6465 tcp_init_transfer(sk, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB,
6467 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6470 tcp_set_state(sk, TCP_ESTABLISHED);
6471 sk->sk_state_change(sk);
6473 /* Note, that this wakeup is only for marginal crossed SYN case.
6474 * Passively open sockets are not waked up, because
6475 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
6478 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
6480 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
6481 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
6482 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
6484 if (tp->rx_opt.tstamp_ok)
6485 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
6487 if (!inet_csk(sk)->icsk_ca_ops->cong_control)
6488 tcp_update_pacing_rate(sk);
6490 /* Prevent spurious tcp_cwnd_restart() on first data packet */
6491 tp->lsndtime = tcp_jiffies32;
6493 tcp_initialize_rcv_mss(sk);
6494 tcp_fast_path_on(tp);
6497 case TCP_FIN_WAIT1: {
6501 tcp_rcv_synrecv_state_fastopen(sk);
6503 if (tp->snd_una != tp->write_seq)
6506 tcp_set_state(sk, TCP_FIN_WAIT2);
6507 sk->sk_shutdown |= SEND_SHUTDOWN;
6511 if (!sock_flag(sk, SOCK_DEAD)) {
6512 /* Wake up lingering close() */
6513 sk->sk_state_change(sk);
6517 if (tp->linger2 < 0) {
6519 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6522 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6523 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6524 /* Receive out of order FIN after close() */
6525 if (tp->syn_fastopen && th->fin)
6526 tcp_fastopen_active_disable(sk);
6528 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6532 tmo = tcp_fin_time(sk);
6533 if (tmo > TCP_TIMEWAIT_LEN) {
6534 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
6535 } else if (th->fin || sock_owned_by_user(sk)) {
6536 /* Bad case. We could lose such FIN otherwise.
6537 * It is not a big problem, but it looks confusing
6538 * and not so rare event. We still can lose it now,
6539 * if it spins in bh_lock_sock(), but it is really
6542 inet_csk_reset_keepalive_timer(sk, tmo);
6544 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
6551 if (tp->snd_una == tp->write_seq) {
6552 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
6558 if (tp->snd_una == tp->write_seq) {
6559 tcp_update_metrics(sk);
6566 /* step 6: check the URG bit */
6567 tcp_urg(sk, skb, th);
6569 /* step 7: process the segment text */
6570 switch (sk->sk_state) {
6571 case TCP_CLOSE_WAIT:
6574 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
6575 if (sk_is_mptcp(sk))
6576 mptcp_incoming_options(sk, skb);
6582 /* RFC 793 says to queue data in these states,
6583 * RFC 1122 says we MUST send a reset.
6584 * BSD 4.4 also does reset.
6586 if (sk->sk_shutdown & RCV_SHUTDOWN) {
6587 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6588 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6589 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6595 case TCP_ESTABLISHED:
6596 tcp_data_queue(sk, skb);
6601 /* tcp_data could move socket to TIME-WAIT */
6602 if (sk->sk_state != TCP_CLOSE) {
6603 tcp_data_snd_check(sk);
6604 tcp_ack_snd_check(sk);
6613 EXPORT_SYMBOL(tcp_rcv_state_process);
6615 static inline void pr_drop_req(struct request_sock *req, __u16 port, int family)
6617 struct inet_request_sock *ireq = inet_rsk(req);
6619 if (family == AF_INET)
6620 net_dbg_ratelimited("drop open request from %pI4/%u\n",
6621 &ireq->ir_rmt_addr, port);
6622 #if IS_ENABLED(CONFIG_IPV6)
6623 else if (family == AF_INET6)
6624 net_dbg_ratelimited("drop open request from %pI6/%u\n",
6625 &ireq->ir_v6_rmt_addr, port);
6629 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
6631 * If we receive a SYN packet with these bits set, it means a
6632 * network is playing bad games with TOS bits. In order to
6633 * avoid possible false congestion notifications, we disable
6634 * TCP ECN negotiation.
6636 * Exception: tcp_ca wants ECN. This is required for DCTCP
6637 * congestion control: Linux DCTCP asserts ECT on all packets,
6638 * including SYN, which is most optimal solution; however,
6639 * others, such as FreeBSD do not.
6641 * Exception: At least one of the reserved bits of the TCP header (th->res1) is
6642 * set, indicating the use of a future TCP extension (such as AccECN). See
6643 * RFC8311 §4.3 which updates RFC3168 to allow the development of such
6646 static void tcp_ecn_create_request(struct request_sock *req,
6647 const struct sk_buff *skb,
6648 const struct sock *listen_sk,
6649 const struct dst_entry *dst)
6651 const struct tcphdr *th = tcp_hdr(skb);
6652 const struct net *net = sock_net(listen_sk);
6653 bool th_ecn = th->ece && th->cwr;
6660 ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield);
6661 ecn_ok_dst = dst_feature(dst, DST_FEATURE_ECN_MASK);
6662 ecn_ok = net->ipv4.sysctl_tcp_ecn || ecn_ok_dst;
6664 if (((!ect || th->res1) && ecn_ok) || tcp_ca_needs_ecn(listen_sk) ||
6665 (ecn_ok_dst & DST_FEATURE_ECN_CA) ||
6666 tcp_bpf_ca_needs_ecn((struct sock *)req))
6667 inet_rsk(req)->ecn_ok = 1;
6670 static void tcp_openreq_init(struct request_sock *req,
6671 const struct tcp_options_received *rx_opt,
6672 struct sk_buff *skb, const struct sock *sk)
6674 struct inet_request_sock *ireq = inet_rsk(req);
6676 req->rsk_rcv_wnd = 0; /* So that tcp_send_synack() knows! */
6677 tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq;
6678 tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
6679 tcp_rsk(req)->snt_synack = 0;
6680 tcp_rsk(req)->last_oow_ack_time = 0;
6681 req->mss = rx_opt->mss_clamp;
6682 req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0;
6683 ireq->tstamp_ok = rx_opt->tstamp_ok;
6684 ireq->sack_ok = rx_opt->sack_ok;
6685 ireq->snd_wscale = rx_opt->snd_wscale;
6686 ireq->wscale_ok = rx_opt->wscale_ok;
6689 ireq->ir_rmt_port = tcp_hdr(skb)->source;
6690 ireq->ir_num = ntohs(tcp_hdr(skb)->dest);
6691 ireq->ir_mark = inet_request_mark(sk, skb);
6692 #if IS_ENABLED(CONFIG_SMC)
6693 ireq->smc_ok = rx_opt->smc_ok;
6697 struct request_sock *inet_reqsk_alloc(const struct request_sock_ops *ops,
6698 struct sock *sk_listener,
6699 bool attach_listener)
6701 struct request_sock *req = reqsk_alloc(ops, sk_listener,
6705 struct inet_request_sock *ireq = inet_rsk(req);
6707 ireq->ireq_opt = NULL;
6708 #if IS_ENABLED(CONFIG_IPV6)
6709 ireq->pktopts = NULL;
6711 atomic64_set(&ireq->ir_cookie, 0);
6712 ireq->ireq_state = TCP_NEW_SYN_RECV;
6713 write_pnet(&ireq->ireq_net, sock_net(sk_listener));
6714 ireq->ireq_family = sk_listener->sk_family;
6719 EXPORT_SYMBOL(inet_reqsk_alloc);
6722 * Return true if a syncookie should be sent
6724 static bool tcp_syn_flood_action(const struct sock *sk, const char *proto)
6726 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
6727 const char *msg = "Dropping request";
6728 struct net *net = sock_net(sk);
6729 bool want_cookie = false;
6732 syncookies = READ_ONCE(net->ipv4.sysctl_tcp_syncookies);
6734 #ifdef CONFIG_SYN_COOKIES
6736 msg = "Sending cookies";
6738 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES);
6741 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP);
6743 if (!queue->synflood_warned && syncookies != 2 &&
6744 xchg(&queue->synflood_warned, 1) == 0)
6745 net_info_ratelimited("%s: Possible SYN flooding on port %d. %s. Check SNMP counters.\n",
6746 proto, sk->sk_num, msg);
6751 static void tcp_reqsk_record_syn(const struct sock *sk,
6752 struct request_sock *req,
6753 const struct sk_buff *skb)
6755 if (tcp_sk(sk)->save_syn) {
6756 u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb);
6757 struct saved_syn *saved_syn;
6761 if (tcp_sk(sk)->save_syn == 2) { /* Save full header. */
6762 base = skb_mac_header(skb);
6763 mac_hdrlen = skb_mac_header_len(skb);
6766 base = skb_network_header(skb);
6770 saved_syn = kmalloc(struct_size(saved_syn, data, len),
6773 saved_syn->mac_hdrlen = mac_hdrlen;
6774 saved_syn->network_hdrlen = skb_network_header_len(skb);
6775 saved_syn->tcp_hdrlen = tcp_hdrlen(skb);
6776 memcpy(saved_syn->data, base, len);
6777 req->saved_syn = saved_syn;
6782 /* If a SYN cookie is required and supported, returns a clamped MSS value to be
6783 * used for SYN cookie generation.
6785 u16 tcp_get_syncookie_mss(struct request_sock_ops *rsk_ops,
6786 const struct tcp_request_sock_ops *af_ops,
6787 struct sock *sk, struct tcphdr *th)
6789 struct tcp_sock *tp = tcp_sk(sk);
6792 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_syncookies) != 2 &&
6793 !inet_csk_reqsk_queue_is_full(sk))
6796 if (!tcp_syn_flood_action(sk, rsk_ops->slab_name))
6799 if (sk_acceptq_is_full(sk)) {
6800 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
6804 mss = tcp_parse_mss_option(th, tp->rx_opt.user_mss);
6806 mss = af_ops->mss_clamp;
6810 EXPORT_SYMBOL_GPL(tcp_get_syncookie_mss);
6812 int tcp_conn_request(struct request_sock_ops *rsk_ops,
6813 const struct tcp_request_sock_ops *af_ops,
6814 struct sock *sk, struct sk_buff *skb)
6816 struct tcp_fastopen_cookie foc = { .len = -1 };
6817 __u32 isn = TCP_SKB_CB(skb)->tcp_tw_isn;
6818 struct tcp_options_received tmp_opt;
6819 struct tcp_sock *tp = tcp_sk(sk);
6820 struct net *net = sock_net(sk);
6821 struct sock *fastopen_sk = NULL;
6822 struct request_sock *req;
6823 bool want_cookie = false;
6824 struct dst_entry *dst;
6828 syncookies = READ_ONCE(net->ipv4.sysctl_tcp_syncookies);
6830 /* TW buckets are converted to open requests without
6831 * limitations, they conserve resources and peer is
6832 * evidently real one.
6834 if ((syncookies == 2 || inet_csk_reqsk_queue_is_full(sk)) && !isn) {
6835 want_cookie = tcp_syn_flood_action(sk, rsk_ops->slab_name);
6840 if (sk_acceptq_is_full(sk)) {
6841 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
6845 req = inet_reqsk_alloc(rsk_ops, sk, !want_cookie);
6849 req->syncookie = want_cookie;
6850 tcp_rsk(req)->af_specific = af_ops;
6851 tcp_rsk(req)->ts_off = 0;
6852 #if IS_ENABLED(CONFIG_MPTCP)
6853 tcp_rsk(req)->is_mptcp = 0;
6856 tcp_clear_options(&tmp_opt);
6857 tmp_opt.mss_clamp = af_ops->mss_clamp;
6858 tmp_opt.user_mss = tp->rx_opt.user_mss;
6859 tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0,
6860 want_cookie ? NULL : &foc);
6862 if (want_cookie && !tmp_opt.saw_tstamp)
6863 tcp_clear_options(&tmp_opt);
6865 if (IS_ENABLED(CONFIG_SMC) && want_cookie)
6868 tmp_opt.tstamp_ok = tmp_opt.saw_tstamp;
6869 tcp_openreq_init(req, &tmp_opt, skb, sk);
6870 inet_rsk(req)->no_srccheck = inet_sk(sk)->transparent;
6872 /* Note: tcp_v6_init_req() might override ir_iif for link locals */
6873 inet_rsk(req)->ir_iif = inet_request_bound_dev_if(sk, skb);
6875 af_ops->init_req(req, sk, skb);
6877 if (security_inet_conn_request(sk, skb, req))
6880 if (tmp_opt.tstamp_ok)
6881 tcp_rsk(req)->ts_off = af_ops->init_ts_off(net, skb);
6883 dst = af_ops->route_req(sk, &fl, req);
6887 if (!want_cookie && !isn) {
6888 int max_syn_backlog = READ_ONCE(net->ipv4.sysctl_max_syn_backlog);
6890 /* Kill the following clause, if you dislike this way. */
6892 (max_syn_backlog - inet_csk_reqsk_queue_len(sk) <
6893 (max_syn_backlog >> 2)) &&
6894 !tcp_peer_is_proven(req, dst)) {
6895 /* Without syncookies last quarter of
6896 * backlog is filled with destinations,
6897 * proven to be alive.
6898 * It means that we continue to communicate
6899 * to destinations, already remembered
6900 * to the moment of synflood.
6902 pr_drop_req(req, ntohs(tcp_hdr(skb)->source),
6904 goto drop_and_release;
6907 isn = af_ops->init_seq(skb);
6910 tcp_ecn_create_request(req, skb, sk, dst);
6913 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss);
6914 if (!tmp_opt.tstamp_ok)
6915 inet_rsk(req)->ecn_ok = 0;
6918 tcp_rsk(req)->snt_isn = isn;
6919 tcp_rsk(req)->txhash = net_tx_rndhash();
6920 tcp_rsk(req)->syn_tos = TCP_SKB_CB(skb)->ip_dsfield;
6921 tcp_openreq_init_rwin(req, sk, dst);
6922 sk_rx_queue_set(req_to_sk(req), skb);
6924 tcp_reqsk_record_syn(sk, req, skb);
6925 fastopen_sk = tcp_try_fastopen(sk, skb, req, &foc, dst);
6928 af_ops->send_synack(fastopen_sk, dst, &fl, req,
6929 &foc, TCP_SYNACK_FASTOPEN, skb);
6930 /* Add the child socket directly into the accept queue */
6931 if (!inet_csk_reqsk_queue_add(sk, req, fastopen_sk)) {
6932 reqsk_fastopen_remove(fastopen_sk, req, false);
6933 bh_unlock_sock(fastopen_sk);
6934 sock_put(fastopen_sk);
6937 sk->sk_data_ready(sk);
6938 bh_unlock_sock(fastopen_sk);
6939 sock_put(fastopen_sk);
6941 tcp_rsk(req)->tfo_listener = false;
6943 inet_csk_reqsk_queue_hash_add(sk, req,
6944 tcp_timeout_init((struct sock *)req));
6945 af_ops->send_synack(sk, dst, &fl, req, &foc,
6946 !want_cookie ? TCP_SYNACK_NORMAL :
6965 EXPORT_SYMBOL(tcp_conn_request);