1 // SPDX-License-Identifier: GPL-2.0
3 * INET An implementation of the TCP/IP protocol suite for the LINUX
4 * operating system. INET is implemented using the BSD Socket
5 * interface as the means of communication with the user level.
7 * Implementation of the Transmission Control Protocol(TCP).
10 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
11 * Mark Evans, <evansmp@uhura.aston.ac.uk>
12 * Corey Minyard <wf-rch!minyard@relay.EU.net>
13 * Florian La Roche, <flla@stud.uni-sb.de>
14 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
15 * Linus Torvalds, <torvalds@cs.helsinki.fi>
16 * Alan Cox, <gw4pts@gw4pts.ampr.org>
17 * Matthew Dillon, <dillon@apollo.west.oic.com>
18 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
19 * Jorge Cwik, <jorge@laser.satlink.net>
24 * Pedro Roque : Fast Retransmit/Recovery.
26 * Retransmit queue handled by TCP.
27 * Better retransmit timer handling.
28 * New congestion avoidance.
32 * Eric : Fast Retransmit.
33 * Randy Scott : MSS option defines.
34 * Eric Schenk : Fixes to slow start algorithm.
35 * Eric Schenk : Yet another double ACK bug.
36 * Eric Schenk : Delayed ACK bug fixes.
37 * Eric Schenk : Floyd style fast retrans war avoidance.
38 * David S. Miller : Don't allow zero congestion window.
39 * Eric Schenk : Fix retransmitter so that it sends
40 * next packet on ack of previous packet.
41 * Andi Kleen : Moved open_request checking here
42 * and process RSTs for open_requests.
43 * Andi Kleen : Better prune_queue, and other fixes.
44 * Andrey Savochkin: Fix RTT measurements in the presence of
46 * Andrey Savochkin: Check sequence numbers correctly when
47 * removing SACKs due to in sequence incoming
49 * Andi Kleen: Make sure we never ack data there is not
50 * enough room for. Also make this condition
51 * a fatal error if it might still happen.
52 * Andi Kleen: Add tcp_measure_rcv_mss to make
53 * connections with MSS<min(MTU,ann. MSS)
54 * work without delayed acks.
55 * Andi Kleen: Process packets with PSH set in the
57 * J Hadi Salim: ECN support
60 * Panu Kuhlberg: Experimental audit of TCP (re)transmission
61 * engine. Lots of bugs are found.
62 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs
65 #define pr_fmt(fmt) "TCP: " fmt
68 #include <linux/slab.h>
69 #include <linux/module.h>
70 #include <linux/sysctl.h>
71 #include <linux/kernel.h>
72 #include <linux/prefetch.h>
75 #include <net/inet_common.h>
76 #include <linux/ipsec.h>
77 #include <asm/unaligned.h>
78 #include <linux/errqueue.h>
79 #include <trace/events/tcp.h>
80 #include <linux/jump_label_ratelimit.h>
81 #include <net/busy_poll.h>
82 #include <net/mptcp.h>
84 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
86 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
87 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
88 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
89 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
90 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
91 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
92 #define FLAG_ECE 0x40 /* ECE in this ACK */
93 #define FLAG_LOST_RETRANS 0x80 /* This ACK marks some retransmission lost */
94 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
95 #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
96 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
97 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
98 #define FLAG_SET_XMIT_TIMER 0x1000 /* Set TLP or RTO timer */
99 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
100 #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
101 #define FLAG_NO_CHALLENGE_ACK 0x8000 /* do not call tcp_send_challenge_ack() */
102 #define FLAG_ACK_MAYBE_DELAYED 0x10000 /* Likely a delayed ACK */
103 #define FLAG_DSACK_TLP 0x20000 /* DSACK for tail loss probe */
105 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
106 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
107 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE|FLAG_DSACKING_ACK)
108 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
110 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
111 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
113 #define REXMIT_NONE 0 /* no loss recovery to do */
114 #define REXMIT_LOST 1 /* retransmit packets marked lost */
115 #define REXMIT_NEW 2 /* FRTO-style transmit of unsent/new packets */
117 #if IS_ENABLED(CONFIG_TLS_DEVICE)
118 static DEFINE_STATIC_KEY_DEFERRED_FALSE(clean_acked_data_enabled, HZ);
120 void clean_acked_data_enable(struct inet_connection_sock *icsk,
121 void (*cad)(struct sock *sk, u32 ack_seq))
123 icsk->icsk_clean_acked = cad;
124 static_branch_deferred_inc(&clean_acked_data_enabled);
126 EXPORT_SYMBOL_GPL(clean_acked_data_enable);
128 void clean_acked_data_disable(struct inet_connection_sock *icsk)
130 static_branch_slow_dec_deferred(&clean_acked_data_enabled);
131 icsk->icsk_clean_acked = NULL;
133 EXPORT_SYMBOL_GPL(clean_acked_data_disable);
135 void clean_acked_data_flush(void)
137 static_key_deferred_flush(&clean_acked_data_enabled);
139 EXPORT_SYMBOL_GPL(clean_acked_data_flush);
142 #ifdef CONFIG_CGROUP_BPF
143 static void bpf_skops_parse_hdr(struct sock *sk, struct sk_buff *skb)
145 bool unknown_opt = tcp_sk(sk)->rx_opt.saw_unknown &&
146 BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk),
147 BPF_SOCK_OPS_PARSE_UNKNOWN_HDR_OPT_CB_FLAG);
148 bool parse_all_opt = BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk),
149 BPF_SOCK_OPS_PARSE_ALL_HDR_OPT_CB_FLAG);
150 struct bpf_sock_ops_kern sock_ops;
152 if (likely(!unknown_opt && !parse_all_opt))
155 /* The skb will be handled in the
156 * bpf_skops_established() or
157 * bpf_skops_write_hdr_opt().
159 switch (sk->sk_state) {
166 sock_owned_by_me(sk);
168 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp));
169 sock_ops.op = BPF_SOCK_OPS_PARSE_HDR_OPT_CB;
170 sock_ops.is_fullsock = 1;
172 bpf_skops_init_skb(&sock_ops, skb, tcp_hdrlen(skb));
174 BPF_CGROUP_RUN_PROG_SOCK_OPS(&sock_ops);
177 static void bpf_skops_established(struct sock *sk, int bpf_op,
180 struct bpf_sock_ops_kern sock_ops;
182 sock_owned_by_me(sk);
184 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp));
185 sock_ops.op = bpf_op;
186 sock_ops.is_fullsock = 1;
188 /* sk with TCP_REPAIR_ON does not have skb in tcp_finish_connect */
190 bpf_skops_init_skb(&sock_ops, skb, tcp_hdrlen(skb));
192 BPF_CGROUP_RUN_PROG_SOCK_OPS(&sock_ops);
195 static void bpf_skops_parse_hdr(struct sock *sk, struct sk_buff *skb)
199 static void bpf_skops_established(struct sock *sk, int bpf_op,
205 static void tcp_gro_dev_warn(struct sock *sk, const struct sk_buff *skb,
208 static bool __once __read_mostly;
211 struct net_device *dev;
216 dev = dev_get_by_index_rcu(sock_net(sk), skb->skb_iif);
217 if (!dev || len >= dev->mtu)
218 pr_warn("%s: Driver has suspect GRO implementation, TCP performance may be compromised.\n",
219 dev ? dev->name : "Unknown driver");
224 /* Adapt the MSS value used to make delayed ack decision to the
227 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
229 struct inet_connection_sock *icsk = inet_csk(sk);
230 const unsigned int lss = icsk->icsk_ack.last_seg_size;
233 icsk->icsk_ack.last_seg_size = 0;
235 /* skb->len may jitter because of SACKs, even if peer
236 * sends good full-sized frames.
238 len = skb_shinfo(skb)->gso_size ? : skb->len;
239 if (len >= icsk->icsk_ack.rcv_mss) {
240 icsk->icsk_ack.rcv_mss = min_t(unsigned int, len,
242 /* Account for possibly-removed options */
243 if (unlikely(len > icsk->icsk_ack.rcv_mss +
244 MAX_TCP_OPTION_SPACE))
245 tcp_gro_dev_warn(sk, skb, len);
247 /* Otherwise, we make more careful check taking into account,
248 * that SACKs block is variable.
250 * "len" is invariant segment length, including TCP header.
252 len += skb->data - skb_transport_header(skb);
253 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
254 /* If PSH is not set, packet should be
255 * full sized, provided peer TCP is not badly broken.
256 * This observation (if it is correct 8)) allows
257 * to handle super-low mtu links fairly.
259 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
260 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
261 /* Subtract also invariant (if peer is RFC compliant),
262 * tcp header plus fixed timestamp option length.
263 * Resulting "len" is MSS free of SACK jitter.
265 len -= tcp_sk(sk)->tcp_header_len;
266 icsk->icsk_ack.last_seg_size = len;
268 icsk->icsk_ack.rcv_mss = len;
272 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
273 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
274 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
278 static void tcp_incr_quickack(struct sock *sk, unsigned int max_quickacks)
280 struct inet_connection_sock *icsk = inet_csk(sk);
281 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
285 quickacks = min(quickacks, max_quickacks);
286 if (quickacks > icsk->icsk_ack.quick)
287 icsk->icsk_ack.quick = quickacks;
290 void tcp_enter_quickack_mode(struct sock *sk, unsigned int max_quickacks)
292 struct inet_connection_sock *icsk = inet_csk(sk);
294 tcp_incr_quickack(sk, max_quickacks);
295 inet_csk_exit_pingpong_mode(sk);
296 icsk->icsk_ack.ato = TCP_ATO_MIN;
298 EXPORT_SYMBOL(tcp_enter_quickack_mode);
300 /* Send ACKs quickly, if "quick" count is not exhausted
301 * and the session is not interactive.
304 static bool tcp_in_quickack_mode(struct sock *sk)
306 const struct inet_connection_sock *icsk = inet_csk(sk);
307 const struct dst_entry *dst = __sk_dst_get(sk);
309 return (dst && dst_metric(dst, RTAX_QUICKACK)) ||
310 (icsk->icsk_ack.quick && !inet_csk_in_pingpong_mode(sk));
313 static void tcp_ecn_queue_cwr(struct tcp_sock *tp)
315 if (tp->ecn_flags & TCP_ECN_OK)
316 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
319 static void tcp_ecn_accept_cwr(struct sock *sk, const struct sk_buff *skb)
321 if (tcp_hdr(skb)->cwr) {
322 tcp_sk(sk)->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
324 /* If the sender is telling us it has entered CWR, then its
325 * cwnd may be very low (even just 1 packet), so we should ACK
328 if (TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq)
329 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW;
333 static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp)
335 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
338 static void __tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb)
340 struct tcp_sock *tp = tcp_sk(sk);
342 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
343 case INET_ECN_NOT_ECT:
344 /* Funny extension: if ECT is not set on a segment,
345 * and we already seen ECT on a previous segment,
346 * it is probably a retransmit.
348 if (tp->ecn_flags & TCP_ECN_SEEN)
349 tcp_enter_quickack_mode(sk, 2);
352 if (tcp_ca_needs_ecn(sk))
353 tcp_ca_event(sk, CA_EVENT_ECN_IS_CE);
355 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
356 /* Better not delay acks, sender can have a very low cwnd */
357 tcp_enter_quickack_mode(sk, 2);
358 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
360 tp->ecn_flags |= TCP_ECN_SEEN;
363 if (tcp_ca_needs_ecn(sk))
364 tcp_ca_event(sk, CA_EVENT_ECN_NO_CE);
365 tp->ecn_flags |= TCP_ECN_SEEN;
370 static void tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb)
372 if (tcp_sk(sk)->ecn_flags & TCP_ECN_OK)
373 __tcp_ecn_check_ce(sk, skb);
376 static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
378 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
379 tp->ecn_flags &= ~TCP_ECN_OK;
382 static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
384 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
385 tp->ecn_flags &= ~TCP_ECN_OK;
388 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
390 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
395 /* Buffer size and advertised window tuning.
397 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
400 static void tcp_sndbuf_expand(struct sock *sk)
402 const struct tcp_sock *tp = tcp_sk(sk);
403 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
407 /* Worst case is non GSO/TSO : each frame consumes one skb
408 * and skb->head is kmalloced using power of two area of memory
410 per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
412 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
414 per_mss = roundup_pow_of_two(per_mss) +
415 SKB_DATA_ALIGN(sizeof(struct sk_buff));
417 nr_segs = max_t(u32, TCP_INIT_CWND, tcp_snd_cwnd(tp));
418 nr_segs = max_t(u32, nr_segs, tp->reordering + 1);
420 /* Fast Recovery (RFC 5681 3.2) :
421 * Cubic needs 1.7 factor, rounded to 2 to include
422 * extra cushion (application might react slowly to EPOLLOUT)
424 sndmem = ca_ops->sndbuf_expand ? ca_ops->sndbuf_expand(sk) : 2;
425 sndmem *= nr_segs * per_mss;
427 if (sk->sk_sndbuf < sndmem)
428 WRITE_ONCE(sk->sk_sndbuf,
429 min(sndmem, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_wmem[2])));
432 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
434 * All tcp_full_space() is split to two parts: "network" buffer, allocated
435 * forward and advertised in receiver window (tp->rcv_wnd) and
436 * "application buffer", required to isolate scheduling/application
437 * latencies from network.
438 * window_clamp is maximal advertised window. It can be less than
439 * tcp_full_space(), in this case tcp_full_space() - window_clamp
440 * is reserved for "application" buffer. The less window_clamp is
441 * the smoother our behaviour from viewpoint of network, but the lower
442 * throughput and the higher sensitivity of the connection to losses. 8)
444 * rcv_ssthresh is more strict window_clamp used at "slow start"
445 * phase to predict further behaviour of this connection.
446 * It is used for two goals:
447 * - to enforce header prediction at sender, even when application
448 * requires some significant "application buffer". It is check #1.
449 * - to prevent pruning of receive queue because of misprediction
450 * of receiver window. Check #2.
452 * The scheme does not work when sender sends good segments opening
453 * window and then starts to feed us spaghetti. But it should work
454 * in common situations. Otherwise, we have to rely on queue collapsing.
457 /* Slow part of check#2. */
458 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb,
459 unsigned int skbtruesize)
461 struct tcp_sock *tp = tcp_sk(sk);
463 int truesize = tcp_win_from_space(sk, skbtruesize) >> 1;
464 int window = tcp_win_from_space(sk, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2])) >> 1;
466 while (tp->rcv_ssthresh <= window) {
467 if (truesize <= skb->len)
468 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
476 /* Even if skb appears to have a bad len/truesize ratio, TCP coalescing
477 * can play nice with us, as sk_buff and skb->head might be either
478 * freed or shared with up to MAX_SKB_FRAGS segments.
479 * Only give a boost to drivers using page frag(s) to hold the frame(s),
480 * and if no payload was pulled in skb->head before reaching us.
482 static u32 truesize_adjust(bool adjust, const struct sk_buff *skb)
484 u32 truesize = skb->truesize;
486 if (adjust && !skb_headlen(skb)) {
487 truesize -= SKB_TRUESIZE(skb_end_offset(skb));
488 /* paranoid check, some drivers might be buggy */
489 if (unlikely((int)truesize < (int)skb->len))
490 truesize = skb->truesize;
495 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb,
498 struct tcp_sock *tp = tcp_sk(sk);
501 room = min_t(int, tp->window_clamp, tcp_space(sk)) - tp->rcv_ssthresh;
507 if (!tcp_under_memory_pressure(sk)) {
508 unsigned int truesize = truesize_adjust(adjust, skb);
511 /* Check #2. Increase window, if skb with such overhead
512 * will fit to rcvbuf in future.
514 if (tcp_win_from_space(sk, truesize) <= skb->len)
515 incr = 2 * tp->advmss;
517 incr = __tcp_grow_window(sk, skb, truesize);
520 incr = max_t(int, incr, 2 * skb->len);
521 tp->rcv_ssthresh += min(room, incr);
522 inet_csk(sk)->icsk_ack.quick |= 1;
526 * Adjust rcv_ssthresh according to reserved mem
528 tcp_adjust_rcv_ssthresh(sk);
532 /* 3. Try to fixup all. It is made immediately after connection enters
535 static void tcp_init_buffer_space(struct sock *sk)
537 int tcp_app_win = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_app_win);
538 struct tcp_sock *tp = tcp_sk(sk);
541 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
542 tcp_sndbuf_expand(sk);
544 tcp_mstamp_refresh(tp);
545 tp->rcvq_space.time = tp->tcp_mstamp;
546 tp->rcvq_space.seq = tp->copied_seq;
548 maxwin = tcp_full_space(sk);
550 if (tp->window_clamp >= maxwin) {
551 tp->window_clamp = maxwin;
553 if (tcp_app_win && maxwin > 4 * tp->advmss)
554 tp->window_clamp = max(maxwin -
555 (maxwin >> tcp_app_win),
559 /* Force reservation of one segment. */
561 tp->window_clamp > 2 * tp->advmss &&
562 tp->window_clamp + tp->advmss > maxwin)
563 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
565 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
566 tp->snd_cwnd_stamp = tcp_jiffies32;
567 tp->rcvq_space.space = min3(tp->rcv_ssthresh, tp->rcv_wnd,
568 (u32)TCP_INIT_CWND * tp->advmss);
571 /* 4. Recalculate window clamp after socket hit its memory bounds. */
572 static void tcp_clamp_window(struct sock *sk)
574 struct tcp_sock *tp = tcp_sk(sk);
575 struct inet_connection_sock *icsk = inet_csk(sk);
576 struct net *net = sock_net(sk);
579 icsk->icsk_ack.quick = 0;
580 rmem2 = READ_ONCE(net->ipv4.sysctl_tcp_rmem[2]);
582 if (sk->sk_rcvbuf < rmem2 &&
583 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
584 !tcp_under_memory_pressure(sk) &&
585 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
586 WRITE_ONCE(sk->sk_rcvbuf,
587 min(atomic_read(&sk->sk_rmem_alloc), rmem2));
589 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
590 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
593 /* Initialize RCV_MSS value.
594 * RCV_MSS is an our guess about MSS used by the peer.
595 * We haven't any direct information about the MSS.
596 * It's better to underestimate the RCV_MSS rather than overestimate.
597 * Overestimations make us ACKing less frequently than needed.
598 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
600 void tcp_initialize_rcv_mss(struct sock *sk)
602 const struct tcp_sock *tp = tcp_sk(sk);
603 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
605 hint = min(hint, tp->rcv_wnd / 2);
606 hint = min(hint, TCP_MSS_DEFAULT);
607 hint = max(hint, TCP_MIN_MSS);
609 inet_csk(sk)->icsk_ack.rcv_mss = hint;
611 EXPORT_SYMBOL(tcp_initialize_rcv_mss);
613 /* Receiver "autotuning" code.
615 * The algorithm for RTT estimation w/o timestamps is based on
616 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
617 * <https://public.lanl.gov/radiant/pubs.html#DRS>
619 * More detail on this code can be found at
620 * <http://staff.psc.edu/jheffner/>,
621 * though this reference is out of date. A new paper
624 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
626 u32 new_sample = tp->rcv_rtt_est.rtt_us;
629 if (new_sample != 0) {
630 /* If we sample in larger samples in the non-timestamp
631 * case, we could grossly overestimate the RTT especially
632 * with chatty applications or bulk transfer apps which
633 * are stalled on filesystem I/O.
635 * Also, since we are only going for a minimum in the
636 * non-timestamp case, we do not smooth things out
637 * else with timestamps disabled convergence takes too
641 m -= (new_sample >> 3);
649 /* No previous measure. */
653 tp->rcv_rtt_est.rtt_us = new_sample;
656 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
660 if (tp->rcv_rtt_est.time == 0)
662 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
664 delta_us = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcv_rtt_est.time);
667 tcp_rcv_rtt_update(tp, delta_us, 1);
670 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
671 tp->rcv_rtt_est.time = tp->tcp_mstamp;
674 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
675 const struct sk_buff *skb)
677 struct tcp_sock *tp = tcp_sk(sk);
679 if (tp->rx_opt.rcv_tsecr == tp->rcv_rtt_last_tsecr)
681 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr;
683 if (TCP_SKB_CB(skb)->end_seq -
684 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss) {
685 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr;
688 if (likely(delta < INT_MAX / (USEC_PER_SEC / TCP_TS_HZ))) {
691 delta_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
692 tcp_rcv_rtt_update(tp, delta_us, 0);
698 * This function should be called every time data is copied to user space.
699 * It calculates the appropriate TCP receive buffer space.
701 void tcp_rcv_space_adjust(struct sock *sk)
703 struct tcp_sock *tp = tcp_sk(sk);
707 trace_tcp_rcv_space_adjust(sk);
709 tcp_mstamp_refresh(tp);
710 time = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcvq_space.time);
711 if (time < (tp->rcv_rtt_est.rtt_us >> 3) || tp->rcv_rtt_est.rtt_us == 0)
714 /* Number of bytes copied to user in last RTT */
715 copied = tp->copied_seq - tp->rcvq_space.seq;
716 if (copied <= tp->rcvq_space.space)
720 * copied = bytes received in previous RTT, our base window
721 * To cope with packet losses, we need a 2x factor
722 * To cope with slow start, and sender growing its cwin by 100 %
723 * every RTT, we need a 4x factor, because the ACK we are sending
724 * now is for the next RTT, not the current one :
725 * <prev RTT . ><current RTT .. ><next RTT .... >
728 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf) &&
729 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
733 /* minimal window to cope with packet losses, assuming
734 * steady state. Add some cushion because of small variations.
736 rcvwin = ((u64)copied << 1) + 16 * tp->advmss;
738 /* Accommodate for sender rate increase (eg. slow start) */
739 grow = rcvwin * (copied - tp->rcvq_space.space);
740 do_div(grow, tp->rcvq_space.space);
741 rcvwin += (grow << 1);
743 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
744 while (tcp_win_from_space(sk, rcvmem) < tp->advmss)
747 do_div(rcvwin, tp->advmss);
748 rcvbuf = min_t(u64, rcvwin * rcvmem,
749 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]));
750 if (rcvbuf > sk->sk_rcvbuf) {
751 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
753 /* Make the window clamp follow along. */
754 tp->window_clamp = tcp_win_from_space(sk, rcvbuf);
757 tp->rcvq_space.space = copied;
760 tp->rcvq_space.seq = tp->copied_seq;
761 tp->rcvq_space.time = tp->tcp_mstamp;
764 /* There is something which you must keep in mind when you analyze the
765 * behavior of the tp->ato delayed ack timeout interval. When a
766 * connection starts up, we want to ack as quickly as possible. The
767 * problem is that "good" TCP's do slow start at the beginning of data
768 * transmission. The means that until we send the first few ACK's the
769 * sender will sit on his end and only queue most of his data, because
770 * he can only send snd_cwnd unacked packets at any given time. For
771 * each ACK we send, he increments snd_cwnd and transmits more of his
774 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
776 struct tcp_sock *tp = tcp_sk(sk);
777 struct inet_connection_sock *icsk = inet_csk(sk);
780 inet_csk_schedule_ack(sk);
782 tcp_measure_rcv_mss(sk, skb);
784 tcp_rcv_rtt_measure(tp);
788 if (!icsk->icsk_ack.ato) {
789 /* The _first_ data packet received, initialize
790 * delayed ACK engine.
792 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
793 icsk->icsk_ack.ato = TCP_ATO_MIN;
795 int m = now - icsk->icsk_ack.lrcvtime;
797 if (m <= TCP_ATO_MIN / 2) {
798 /* The fastest case is the first. */
799 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
800 } else if (m < icsk->icsk_ack.ato) {
801 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
802 if (icsk->icsk_ack.ato > icsk->icsk_rto)
803 icsk->icsk_ack.ato = icsk->icsk_rto;
804 } else if (m > icsk->icsk_rto) {
805 /* Too long gap. Apparently sender failed to
806 * restart window, so that we send ACKs quickly.
808 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
812 icsk->icsk_ack.lrcvtime = now;
814 tcp_ecn_check_ce(sk, skb);
817 tcp_grow_window(sk, skb, true);
820 /* Called to compute a smoothed rtt estimate. The data fed to this
821 * routine either comes from timestamps, or from segments that were
822 * known _not_ to have been retransmitted [see Karn/Partridge
823 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
824 * piece by Van Jacobson.
825 * NOTE: the next three routines used to be one big routine.
826 * To save cycles in the RFC 1323 implementation it was better to break
827 * it up into three procedures. -- erics
829 static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
831 struct tcp_sock *tp = tcp_sk(sk);
832 long m = mrtt_us; /* RTT */
833 u32 srtt = tp->srtt_us;
835 /* The following amusing code comes from Jacobson's
836 * article in SIGCOMM '88. Note that rtt and mdev
837 * are scaled versions of rtt and mean deviation.
838 * This is designed to be as fast as possible
839 * m stands for "measurement".
841 * On a 1990 paper the rto value is changed to:
842 * RTO = rtt + 4 * mdev
844 * Funny. This algorithm seems to be very broken.
845 * These formulae increase RTO, when it should be decreased, increase
846 * too slowly, when it should be increased quickly, decrease too quickly
847 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
848 * does not matter how to _calculate_ it. Seems, it was trap
849 * that VJ failed to avoid. 8)
852 m -= (srtt >> 3); /* m is now error in rtt est */
853 srtt += m; /* rtt = 7/8 rtt + 1/8 new */
855 m = -m; /* m is now abs(error) */
856 m -= (tp->mdev_us >> 2); /* similar update on mdev */
857 /* This is similar to one of Eifel findings.
858 * Eifel blocks mdev updates when rtt decreases.
859 * This solution is a bit different: we use finer gain
860 * for mdev in this case (alpha*beta).
861 * Like Eifel it also prevents growth of rto,
862 * but also it limits too fast rto decreases,
863 * happening in pure Eifel.
868 m -= (tp->mdev_us >> 2); /* similar update on mdev */
870 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */
871 if (tp->mdev_us > tp->mdev_max_us) {
872 tp->mdev_max_us = tp->mdev_us;
873 if (tp->mdev_max_us > tp->rttvar_us)
874 tp->rttvar_us = tp->mdev_max_us;
876 if (after(tp->snd_una, tp->rtt_seq)) {
877 if (tp->mdev_max_us < tp->rttvar_us)
878 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
879 tp->rtt_seq = tp->snd_nxt;
880 tp->mdev_max_us = tcp_rto_min_us(sk);
885 /* no previous measure. */
886 srtt = m << 3; /* take the measured time to be rtt */
887 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */
888 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
889 tp->mdev_max_us = tp->rttvar_us;
890 tp->rtt_seq = tp->snd_nxt;
894 tp->srtt_us = max(1U, srtt);
897 static void tcp_update_pacing_rate(struct sock *sk)
899 const struct tcp_sock *tp = tcp_sk(sk);
902 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
903 rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3);
905 /* current rate is (cwnd * mss) / srtt
906 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
907 * In Congestion Avoidance phase, set it to 120 % the current rate.
909 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
910 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
911 * end of slow start and should slow down.
913 if (tcp_snd_cwnd(tp) < tp->snd_ssthresh / 2)
914 rate *= READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_pacing_ss_ratio);
916 rate *= READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_pacing_ca_ratio);
918 rate *= max(tcp_snd_cwnd(tp), tp->packets_out);
920 if (likely(tp->srtt_us))
921 do_div(rate, tp->srtt_us);
923 /* WRITE_ONCE() is needed because sch_fq fetches sk_pacing_rate
924 * without any lock. We want to make sure compiler wont store
925 * intermediate values in this location.
927 WRITE_ONCE(sk->sk_pacing_rate, min_t(u64, rate,
928 sk->sk_max_pacing_rate));
931 /* Calculate rto without backoff. This is the second half of Van Jacobson's
932 * routine referred to above.
934 static void tcp_set_rto(struct sock *sk)
936 const struct tcp_sock *tp = tcp_sk(sk);
937 /* Old crap is replaced with new one. 8)
940 * 1. If rtt variance happened to be less 50msec, it is hallucination.
941 * It cannot be less due to utterly erratic ACK generation made
942 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
943 * to do with delayed acks, because at cwnd>2 true delack timeout
944 * is invisible. Actually, Linux-2.4 also generates erratic
945 * ACKs in some circumstances.
947 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
949 /* 2. Fixups made earlier cannot be right.
950 * If we do not estimate RTO correctly without them,
951 * all the algo is pure shit and should be replaced
952 * with correct one. It is exactly, which we pretend to do.
955 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
956 * guarantees that rto is higher.
961 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
963 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
966 cwnd = TCP_INIT_CWND;
967 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
970 struct tcp_sacktag_state {
971 /* Timestamps for earliest and latest never-retransmitted segment
972 * that was SACKed. RTO needs the earliest RTT to stay conservative,
973 * but congestion control should still get an accurate delay signal.
980 unsigned int mss_now;
981 struct rate_sample *rate;
984 /* Take a notice that peer is sending D-SACKs. Skip update of data delivery
985 * and spurious retransmission information if this DSACK is unlikely caused by
987 * - DSACKed sequence range is larger than maximum receiver's window.
988 * - Total no. of DSACKed segments exceed the total no. of retransmitted segs.
990 static u32 tcp_dsack_seen(struct tcp_sock *tp, u32 start_seq,
991 u32 end_seq, struct tcp_sacktag_state *state)
993 u32 seq_len, dup_segs = 1;
995 if (!before(start_seq, end_seq))
998 seq_len = end_seq - start_seq;
999 /* Dubious DSACK: DSACKed range greater than maximum advertised rwnd */
1000 if (seq_len > tp->max_window)
1002 if (seq_len > tp->mss_cache)
1003 dup_segs = DIV_ROUND_UP(seq_len, tp->mss_cache);
1004 else if (tp->tlp_high_seq && tp->tlp_high_seq == end_seq)
1005 state->flag |= FLAG_DSACK_TLP;
1007 tp->dsack_dups += dup_segs;
1008 /* Skip the DSACK if dup segs weren't retransmitted by sender */
1009 if (tp->dsack_dups > tp->total_retrans)
1012 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
1013 /* We increase the RACK ordering window in rounds where we receive
1014 * DSACKs that may have been due to reordering causing RACK to trigger
1015 * a spurious fast recovery. Thus RACK ignores DSACKs that happen
1016 * without having seen reordering, or that match TLP probes (TLP
1017 * is timer-driven, not triggered by RACK).
1019 if (tp->reord_seen && !(state->flag & FLAG_DSACK_TLP))
1020 tp->rack.dsack_seen = 1;
1022 state->flag |= FLAG_DSACKING_ACK;
1023 /* A spurious retransmission is delivered */
1024 state->sack_delivered += dup_segs;
1029 /* It's reordering when higher sequence was delivered (i.e. sacked) before
1030 * some lower never-retransmitted sequence ("low_seq"). The maximum reordering
1031 * distance is approximated in full-mss packet distance ("reordering").
1033 static void tcp_check_sack_reordering(struct sock *sk, const u32 low_seq,
1036 struct tcp_sock *tp = tcp_sk(sk);
1037 const u32 mss = tp->mss_cache;
1040 fack = tcp_highest_sack_seq(tp);
1041 if (!before(low_seq, fack))
1044 metric = fack - low_seq;
1045 if ((metric > tp->reordering * mss) && mss) {
1046 #if FASTRETRANS_DEBUG > 1
1047 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
1048 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
1052 tp->undo_marker ? tp->undo_retrans : 0);
1054 tp->reordering = min_t(u32, (metric + mss - 1) / mss,
1055 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_max_reordering));
1058 /* This exciting event is worth to be remembered. 8) */
1060 NET_INC_STATS(sock_net(sk),
1061 ts ? LINUX_MIB_TCPTSREORDER : LINUX_MIB_TCPSACKREORDER);
1064 /* This must be called before lost_out or retrans_out are updated
1065 * on a new loss, because we want to know if all skbs previously
1066 * known to be lost have already been retransmitted, indicating
1067 * that this newly lost skb is our next skb to retransmit.
1069 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
1071 if ((!tp->retransmit_skb_hint && tp->retrans_out >= tp->lost_out) ||
1072 (tp->retransmit_skb_hint &&
1073 before(TCP_SKB_CB(skb)->seq,
1074 TCP_SKB_CB(tp->retransmit_skb_hint)->seq)))
1075 tp->retransmit_skb_hint = skb;
1078 /* Sum the number of packets on the wire we have marked as lost, and
1079 * notify the congestion control module that the given skb was marked lost.
1081 static void tcp_notify_skb_loss_event(struct tcp_sock *tp, const struct sk_buff *skb)
1083 tp->lost += tcp_skb_pcount(skb);
1086 void tcp_mark_skb_lost(struct sock *sk, struct sk_buff *skb)
1088 __u8 sacked = TCP_SKB_CB(skb)->sacked;
1089 struct tcp_sock *tp = tcp_sk(sk);
1091 if (sacked & TCPCB_SACKED_ACKED)
1094 tcp_verify_retransmit_hint(tp, skb);
1095 if (sacked & TCPCB_LOST) {
1096 if (sacked & TCPCB_SACKED_RETRANS) {
1097 /* Account for retransmits that are lost again */
1098 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1099 tp->retrans_out -= tcp_skb_pcount(skb);
1100 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT,
1101 tcp_skb_pcount(skb));
1102 tcp_notify_skb_loss_event(tp, skb);
1105 tp->lost_out += tcp_skb_pcount(skb);
1106 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1107 tcp_notify_skb_loss_event(tp, skb);
1111 /* Updates the delivered and delivered_ce counts */
1112 static void tcp_count_delivered(struct tcp_sock *tp, u32 delivered,
1115 tp->delivered += delivered;
1117 tp->delivered_ce += delivered;
1120 /* This procedure tags the retransmission queue when SACKs arrive.
1122 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
1123 * Packets in queue with these bits set are counted in variables
1124 * sacked_out, retrans_out and lost_out, correspondingly.
1126 * Valid combinations are:
1127 * Tag InFlight Description
1128 * 0 1 - orig segment is in flight.
1129 * S 0 - nothing flies, orig reached receiver.
1130 * L 0 - nothing flies, orig lost by net.
1131 * R 2 - both orig and retransmit are in flight.
1132 * L|R 1 - orig is lost, retransmit is in flight.
1133 * S|R 1 - orig reached receiver, retrans is still in flight.
1134 * (L|S|R is logically valid, it could occur when L|R is sacked,
1135 * but it is equivalent to plain S and code short-curcuits it to S.
1136 * L|S is logically invalid, it would mean -1 packet in flight 8))
1138 * These 6 states form finite state machine, controlled by the following events:
1139 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
1140 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
1141 * 3. Loss detection event of two flavors:
1142 * A. Scoreboard estimator decided the packet is lost.
1143 * A'. Reno "three dupacks" marks head of queue lost.
1144 * B. SACK arrives sacking SND.NXT at the moment, when the
1145 * segment was retransmitted.
1146 * 4. D-SACK added new rule: D-SACK changes any tag to S.
1148 * It is pleasant to note, that state diagram turns out to be commutative,
1149 * so that we are allowed not to be bothered by order of our actions,
1150 * when multiple events arrive simultaneously. (see the function below).
1152 * Reordering detection.
1153 * --------------------
1154 * Reordering metric is maximal distance, which a packet can be displaced
1155 * in packet stream. With SACKs we can estimate it:
1157 * 1. SACK fills old hole and the corresponding segment was not
1158 * ever retransmitted -> reordering. Alas, we cannot use it
1159 * when segment was retransmitted.
1160 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1161 * for retransmitted and already SACKed segment -> reordering..
1162 * Both of these heuristics are not used in Loss state, when we cannot
1163 * account for retransmits accurately.
1165 * SACK block validation.
1166 * ----------------------
1168 * SACK block range validation checks that the received SACK block fits to
1169 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1170 * Note that SND.UNA is not included to the range though being valid because
1171 * it means that the receiver is rather inconsistent with itself reporting
1172 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1173 * perfectly valid, however, in light of RFC2018 which explicitly states
1174 * that "SACK block MUST reflect the newest segment. Even if the newest
1175 * segment is going to be discarded ...", not that it looks very clever
1176 * in case of head skb. Due to potentional receiver driven attacks, we
1177 * choose to avoid immediate execution of a walk in write queue due to
1178 * reneging and defer head skb's loss recovery to standard loss recovery
1179 * procedure that will eventually trigger (nothing forbids us doing this).
1181 * Implements also blockage to start_seq wrap-around. Problem lies in the
1182 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1183 * there's no guarantee that it will be before snd_nxt (n). The problem
1184 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1187 * <- outs wnd -> <- wrapzone ->
1188 * u e n u_w e_w s n_w
1190 * |<------------+------+----- TCP seqno space --------------+---------->|
1191 * ...-- <2^31 ->| |<--------...
1192 * ...---- >2^31 ------>| |<--------...
1194 * Current code wouldn't be vulnerable but it's better still to discard such
1195 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1196 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1197 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1198 * equal to the ideal case (infinite seqno space without wrap caused issues).
1200 * With D-SACK the lower bound is extended to cover sequence space below
1201 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1202 * again, D-SACK block must not to go across snd_una (for the same reason as
1203 * for the normal SACK blocks, explained above). But there all simplicity
1204 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1205 * fully below undo_marker they do not affect behavior in anyway and can
1206 * therefore be safely ignored. In rare cases (which are more or less
1207 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1208 * fragmentation and packet reordering past skb's retransmission. To consider
1209 * them correctly, the acceptable range must be extended even more though
1210 * the exact amount is rather hard to quantify. However, tp->max_window can
1211 * be used as an exaggerated estimate.
1213 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
1214 u32 start_seq, u32 end_seq)
1216 /* Too far in future, or reversed (interpretation is ambiguous) */
1217 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1220 /* Nasty start_seq wrap-around check (see comments above) */
1221 if (!before(start_seq, tp->snd_nxt))
1224 /* In outstanding window? ...This is valid exit for D-SACKs too.
1225 * start_seq == snd_una is non-sensical (see comments above)
1227 if (after(start_seq, tp->snd_una))
1230 if (!is_dsack || !tp->undo_marker)
1233 /* ...Then it's D-SACK, and must reside below snd_una completely */
1234 if (after(end_seq, tp->snd_una))
1237 if (!before(start_seq, tp->undo_marker))
1241 if (!after(end_seq, tp->undo_marker))
1244 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1245 * start_seq < undo_marker and end_seq >= undo_marker.
1247 return !before(start_seq, end_seq - tp->max_window);
1250 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1251 struct tcp_sack_block_wire *sp, int num_sacks,
1252 u32 prior_snd_una, struct tcp_sacktag_state *state)
1254 struct tcp_sock *tp = tcp_sk(sk);
1255 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1256 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1259 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1260 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1261 } else if (num_sacks > 1) {
1262 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1263 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1265 if (after(end_seq_0, end_seq_1) || before(start_seq_0, start_seq_1))
1267 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKOFORECV);
1272 dup_segs = tcp_dsack_seen(tp, start_seq_0, end_seq_0, state);
1273 if (!dup_segs) { /* Skip dubious DSACK */
1274 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKIGNOREDDUBIOUS);
1278 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECVSEGS, dup_segs);
1280 /* D-SACK for already forgotten data... Do dumb counting. */
1281 if (tp->undo_marker && tp->undo_retrans > 0 &&
1282 !after(end_seq_0, prior_snd_una) &&
1283 after(end_seq_0, tp->undo_marker))
1284 tp->undo_retrans = max_t(int, 0, tp->undo_retrans - dup_segs);
1289 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1290 * the incoming SACK may not exactly match but we can find smaller MSS
1291 * aligned portion of it that matches. Therefore we might need to fragment
1292 * which may fail and creates some hassle (caller must handle error case
1295 * FIXME: this could be merged to shift decision code
1297 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1298 u32 start_seq, u32 end_seq)
1302 unsigned int pkt_len;
1305 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1306 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1308 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1309 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1310 mss = tcp_skb_mss(skb);
1311 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1314 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1318 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1323 /* Round if necessary so that SACKs cover only full MSSes
1324 * and/or the remaining small portion (if present)
1326 if (pkt_len > mss) {
1327 unsigned int new_len = (pkt_len / mss) * mss;
1328 if (!in_sack && new_len < pkt_len)
1333 if (pkt_len >= skb->len && !in_sack)
1336 err = tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
1337 pkt_len, mss, GFP_ATOMIC);
1345 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1346 static u8 tcp_sacktag_one(struct sock *sk,
1347 struct tcp_sacktag_state *state, u8 sacked,
1348 u32 start_seq, u32 end_seq,
1349 int dup_sack, int pcount,
1352 struct tcp_sock *tp = tcp_sk(sk);
1354 /* Account D-SACK for retransmitted packet. */
1355 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1356 if (tp->undo_marker && tp->undo_retrans > 0 &&
1357 after(end_seq, tp->undo_marker))
1358 tp->undo_retrans = max_t(int, 0, tp->undo_retrans - pcount);
1359 if ((sacked & TCPCB_SACKED_ACKED) &&
1360 before(start_seq, state->reord))
1361 state->reord = start_seq;
1364 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1365 if (!after(end_seq, tp->snd_una))
1368 if (!(sacked & TCPCB_SACKED_ACKED)) {
1369 tcp_rack_advance(tp, sacked, end_seq, xmit_time);
1371 if (sacked & TCPCB_SACKED_RETRANS) {
1372 /* If the segment is not tagged as lost,
1373 * we do not clear RETRANS, believing
1374 * that retransmission is still in flight.
1376 if (sacked & TCPCB_LOST) {
1377 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1378 tp->lost_out -= pcount;
1379 tp->retrans_out -= pcount;
1382 if (!(sacked & TCPCB_RETRANS)) {
1383 /* New sack for not retransmitted frame,
1384 * which was in hole. It is reordering.
1386 if (before(start_seq,
1387 tcp_highest_sack_seq(tp)) &&
1388 before(start_seq, state->reord))
1389 state->reord = start_seq;
1391 if (!after(end_seq, tp->high_seq))
1392 state->flag |= FLAG_ORIG_SACK_ACKED;
1393 if (state->first_sackt == 0)
1394 state->first_sackt = xmit_time;
1395 state->last_sackt = xmit_time;
1398 if (sacked & TCPCB_LOST) {
1399 sacked &= ~TCPCB_LOST;
1400 tp->lost_out -= pcount;
1404 sacked |= TCPCB_SACKED_ACKED;
1405 state->flag |= FLAG_DATA_SACKED;
1406 tp->sacked_out += pcount;
1407 /* Out-of-order packets delivered */
1408 state->sack_delivered += pcount;
1410 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1411 if (tp->lost_skb_hint &&
1412 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1413 tp->lost_cnt_hint += pcount;
1416 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1417 * frames and clear it. undo_retrans is decreased above, L|R frames
1418 * are accounted above as well.
1420 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1421 sacked &= ~TCPCB_SACKED_RETRANS;
1422 tp->retrans_out -= pcount;
1428 /* Shift newly-SACKed bytes from this skb to the immediately previous
1429 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1431 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *prev,
1432 struct sk_buff *skb,
1433 struct tcp_sacktag_state *state,
1434 unsigned int pcount, int shifted, int mss,
1437 struct tcp_sock *tp = tcp_sk(sk);
1438 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1439 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
1443 /* Adjust counters and hints for the newly sacked sequence
1444 * range but discard the return value since prev is already
1445 * marked. We must tag the range first because the seq
1446 * advancement below implicitly advances
1447 * tcp_highest_sack_seq() when skb is highest_sack.
1449 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1450 start_seq, end_seq, dup_sack, pcount,
1451 tcp_skb_timestamp_us(skb));
1452 tcp_rate_skb_delivered(sk, skb, state->rate);
1454 if (skb == tp->lost_skb_hint)
1455 tp->lost_cnt_hint += pcount;
1457 TCP_SKB_CB(prev)->end_seq += shifted;
1458 TCP_SKB_CB(skb)->seq += shifted;
1460 tcp_skb_pcount_add(prev, pcount);
1461 WARN_ON_ONCE(tcp_skb_pcount(skb) < pcount);
1462 tcp_skb_pcount_add(skb, -pcount);
1464 /* When we're adding to gso_segs == 1, gso_size will be zero,
1465 * in theory this shouldn't be necessary but as long as DSACK
1466 * code can come after this skb later on it's better to keep
1467 * setting gso_size to something.
1469 if (!TCP_SKB_CB(prev)->tcp_gso_size)
1470 TCP_SKB_CB(prev)->tcp_gso_size = mss;
1472 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1473 if (tcp_skb_pcount(skb) <= 1)
1474 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1476 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1477 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1480 BUG_ON(!tcp_skb_pcount(skb));
1481 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1485 /* Whole SKB was eaten :-) */
1487 if (skb == tp->retransmit_skb_hint)
1488 tp->retransmit_skb_hint = prev;
1489 if (skb == tp->lost_skb_hint) {
1490 tp->lost_skb_hint = prev;
1491 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1494 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1495 TCP_SKB_CB(prev)->eor = TCP_SKB_CB(skb)->eor;
1496 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1497 TCP_SKB_CB(prev)->end_seq++;
1499 if (skb == tcp_highest_sack(sk))
1500 tcp_advance_highest_sack(sk, skb);
1502 tcp_skb_collapse_tstamp(prev, skb);
1503 if (unlikely(TCP_SKB_CB(prev)->tx.delivered_mstamp))
1504 TCP_SKB_CB(prev)->tx.delivered_mstamp = 0;
1506 tcp_rtx_queue_unlink_and_free(skb, sk);
1508 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKMERGED);
1513 /* I wish gso_size would have a bit more sane initialization than
1514 * something-or-zero which complicates things
1516 static int tcp_skb_seglen(const struct sk_buff *skb)
1518 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1521 /* Shifting pages past head area doesn't work */
1522 static int skb_can_shift(const struct sk_buff *skb)
1524 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1527 int tcp_skb_shift(struct sk_buff *to, struct sk_buff *from,
1528 int pcount, int shiftlen)
1530 /* TCP min gso_size is 8 bytes (TCP_MIN_GSO_SIZE)
1531 * Since TCP_SKB_CB(skb)->tcp_gso_segs is 16 bits, we need
1532 * to make sure not storing more than 65535 * 8 bytes per skb,
1533 * even if current MSS is bigger.
1535 if (unlikely(to->len + shiftlen >= 65535 * TCP_MIN_GSO_SIZE))
1537 if (unlikely(tcp_skb_pcount(to) + pcount > 65535))
1539 return skb_shift(to, from, shiftlen);
1542 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1545 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1546 struct tcp_sacktag_state *state,
1547 u32 start_seq, u32 end_seq,
1550 struct tcp_sock *tp = tcp_sk(sk);
1551 struct sk_buff *prev;
1557 /* Normally R but no L won't result in plain S */
1559 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1561 if (!skb_can_shift(skb))
1563 /* This frame is about to be dropped (was ACKed). */
1564 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1567 /* Can only happen with delayed DSACK + discard craziness */
1568 prev = skb_rb_prev(skb);
1572 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1575 if (!tcp_skb_can_collapse(prev, skb))
1578 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1579 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1583 pcount = tcp_skb_pcount(skb);
1584 mss = tcp_skb_seglen(skb);
1586 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1587 * drop this restriction as unnecessary
1589 if (mss != tcp_skb_seglen(prev))
1592 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1594 /* CHECKME: This is non-MSS split case only?, this will
1595 * cause skipped skbs due to advancing loop btw, original
1596 * has that feature too
1598 if (tcp_skb_pcount(skb) <= 1)
1601 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1603 /* TODO: head merge to next could be attempted here
1604 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1605 * though it might not be worth of the additional hassle
1607 * ...we can probably just fallback to what was done
1608 * previously. We could try merging non-SACKed ones
1609 * as well but it probably isn't going to buy off
1610 * because later SACKs might again split them, and
1611 * it would make skb timestamp tracking considerably
1617 len = end_seq - TCP_SKB_CB(skb)->seq;
1619 BUG_ON(len > skb->len);
1621 /* MSS boundaries should be honoured or else pcount will
1622 * severely break even though it makes things bit trickier.
1623 * Optimize common case to avoid most of the divides
1625 mss = tcp_skb_mss(skb);
1627 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1628 * drop this restriction as unnecessary
1630 if (mss != tcp_skb_seglen(prev))
1635 } else if (len < mss) {
1643 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1644 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1647 if (!tcp_skb_shift(prev, skb, pcount, len))
1649 if (!tcp_shifted_skb(sk, prev, skb, state, pcount, len, mss, dup_sack))
1652 /* Hole filled allows collapsing with the next as well, this is very
1653 * useful when hole on every nth skb pattern happens
1655 skb = skb_rb_next(prev);
1659 if (!skb_can_shift(skb) ||
1660 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1661 (mss != tcp_skb_seglen(skb)))
1664 if (!tcp_skb_can_collapse(prev, skb))
1667 pcount = tcp_skb_pcount(skb);
1668 if (tcp_skb_shift(prev, skb, pcount, len))
1669 tcp_shifted_skb(sk, prev, skb, state, pcount,
1679 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1683 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1684 struct tcp_sack_block *next_dup,
1685 struct tcp_sacktag_state *state,
1686 u32 start_seq, u32 end_seq,
1689 struct tcp_sock *tp = tcp_sk(sk);
1690 struct sk_buff *tmp;
1692 skb_rbtree_walk_from(skb) {
1694 bool dup_sack = dup_sack_in;
1696 /* queue is in-order => we can short-circuit the walk early */
1697 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1701 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1702 in_sack = tcp_match_skb_to_sack(sk, skb,
1703 next_dup->start_seq,
1709 /* skb reference here is a bit tricky to get right, since
1710 * shifting can eat and free both this skb and the next,
1711 * so not even _safe variant of the loop is enough.
1714 tmp = tcp_shift_skb_data(sk, skb, state,
1715 start_seq, end_seq, dup_sack);
1724 in_sack = tcp_match_skb_to_sack(sk, skb,
1730 if (unlikely(in_sack < 0))
1734 TCP_SKB_CB(skb)->sacked =
1737 TCP_SKB_CB(skb)->sacked,
1738 TCP_SKB_CB(skb)->seq,
1739 TCP_SKB_CB(skb)->end_seq,
1741 tcp_skb_pcount(skb),
1742 tcp_skb_timestamp_us(skb));
1743 tcp_rate_skb_delivered(sk, skb, state->rate);
1744 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1745 list_del_init(&skb->tcp_tsorted_anchor);
1747 if (!before(TCP_SKB_CB(skb)->seq,
1748 tcp_highest_sack_seq(tp)))
1749 tcp_advance_highest_sack(sk, skb);
1755 static struct sk_buff *tcp_sacktag_bsearch(struct sock *sk, u32 seq)
1757 struct rb_node *parent, **p = &sk->tcp_rtx_queue.rb_node;
1758 struct sk_buff *skb;
1762 skb = rb_to_skb(parent);
1763 if (before(seq, TCP_SKB_CB(skb)->seq)) {
1764 p = &parent->rb_left;
1767 if (!before(seq, TCP_SKB_CB(skb)->end_seq)) {
1768 p = &parent->rb_right;
1776 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1779 if (skb && after(TCP_SKB_CB(skb)->seq, skip_to_seq))
1782 return tcp_sacktag_bsearch(sk, skip_to_seq);
1785 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1787 struct tcp_sack_block *next_dup,
1788 struct tcp_sacktag_state *state,
1794 if (before(next_dup->start_seq, skip_to_seq)) {
1795 skb = tcp_sacktag_skip(skb, sk, next_dup->start_seq);
1796 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1797 next_dup->start_seq, next_dup->end_seq,
1804 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1806 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1810 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1811 u32 prior_snd_una, struct tcp_sacktag_state *state)
1813 struct tcp_sock *tp = tcp_sk(sk);
1814 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1815 TCP_SKB_CB(ack_skb)->sacked);
1816 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1817 struct tcp_sack_block sp[TCP_NUM_SACKS];
1818 struct tcp_sack_block *cache;
1819 struct sk_buff *skb;
1820 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1822 bool found_dup_sack = false;
1824 int first_sack_index;
1827 state->reord = tp->snd_nxt;
1829 if (!tp->sacked_out)
1830 tcp_highest_sack_reset(sk);
1832 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1833 num_sacks, prior_snd_una, state);
1835 /* Eliminate too old ACKs, but take into
1836 * account more or less fresh ones, they can
1837 * contain valid SACK info.
1839 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1842 if (!tp->packets_out)
1846 first_sack_index = 0;
1847 for (i = 0; i < num_sacks; i++) {
1848 bool dup_sack = !i && found_dup_sack;
1850 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1851 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1853 if (!tcp_is_sackblock_valid(tp, dup_sack,
1854 sp[used_sacks].start_seq,
1855 sp[used_sacks].end_seq)) {
1859 if (!tp->undo_marker)
1860 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1862 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1864 /* Don't count olds caused by ACK reordering */
1865 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1866 !after(sp[used_sacks].end_seq, tp->snd_una))
1868 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1871 NET_INC_STATS(sock_net(sk), mib_idx);
1873 first_sack_index = -1;
1877 /* Ignore very old stuff early */
1878 if (!after(sp[used_sacks].end_seq, prior_snd_una)) {
1880 first_sack_index = -1;
1887 /* order SACK blocks to allow in order walk of the retrans queue */
1888 for (i = used_sacks - 1; i > 0; i--) {
1889 for (j = 0; j < i; j++) {
1890 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1891 swap(sp[j], sp[j + 1]);
1893 /* Track where the first SACK block goes to */
1894 if (j == first_sack_index)
1895 first_sack_index = j + 1;
1900 state->mss_now = tcp_current_mss(sk);
1904 if (!tp->sacked_out) {
1905 /* It's already past, so skip checking against it */
1906 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1908 cache = tp->recv_sack_cache;
1909 /* Skip empty blocks in at head of the cache */
1910 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1915 while (i < used_sacks) {
1916 u32 start_seq = sp[i].start_seq;
1917 u32 end_seq = sp[i].end_seq;
1918 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1919 struct tcp_sack_block *next_dup = NULL;
1921 if (found_dup_sack && ((i + 1) == first_sack_index))
1922 next_dup = &sp[i + 1];
1924 /* Skip too early cached blocks */
1925 while (tcp_sack_cache_ok(tp, cache) &&
1926 !before(start_seq, cache->end_seq))
1929 /* Can skip some work by looking recv_sack_cache? */
1930 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1931 after(end_seq, cache->start_seq)) {
1934 if (before(start_seq, cache->start_seq)) {
1935 skb = tcp_sacktag_skip(skb, sk, start_seq);
1936 skb = tcp_sacktag_walk(skb, sk, next_dup,
1943 /* Rest of the block already fully processed? */
1944 if (!after(end_seq, cache->end_seq))
1947 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1951 /* ...tail remains todo... */
1952 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1953 /* ...but better entrypoint exists! */
1954 skb = tcp_highest_sack(sk);
1961 skb = tcp_sacktag_skip(skb, sk, cache->end_seq);
1962 /* Check overlap against next cached too (past this one already) */
1967 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1968 skb = tcp_highest_sack(sk);
1972 skb = tcp_sacktag_skip(skb, sk, start_seq);
1975 skb = tcp_sacktag_walk(skb, sk, next_dup, state,
1976 start_seq, end_seq, dup_sack);
1982 /* Clear the head of the cache sack blocks so we can skip it next time */
1983 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1984 tp->recv_sack_cache[i].start_seq = 0;
1985 tp->recv_sack_cache[i].end_seq = 0;
1987 for (j = 0; j < used_sacks; j++)
1988 tp->recv_sack_cache[i++] = sp[j];
1990 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Loss || tp->undo_marker)
1991 tcp_check_sack_reordering(sk, state->reord, 0);
1993 tcp_verify_left_out(tp);
1996 #if FASTRETRANS_DEBUG > 0
1997 WARN_ON((int)tp->sacked_out < 0);
1998 WARN_ON((int)tp->lost_out < 0);
1999 WARN_ON((int)tp->retrans_out < 0);
2000 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
2005 /* Limits sacked_out so that sum with lost_out isn't ever larger than
2006 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
2008 static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
2012 holes = max(tp->lost_out, 1U);
2013 holes = min(holes, tp->packets_out);
2015 if ((tp->sacked_out + holes) > tp->packets_out) {
2016 tp->sacked_out = tp->packets_out - holes;
2022 /* If we receive more dupacks than we expected counting segments
2023 * in assumption of absent reordering, interpret this as reordering.
2024 * The only another reason could be bug in receiver TCP.
2026 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
2028 struct tcp_sock *tp = tcp_sk(sk);
2030 if (!tcp_limit_reno_sacked(tp))
2033 tp->reordering = min_t(u32, tp->packets_out + addend,
2034 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_max_reordering));
2036 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRENOREORDER);
2039 /* Emulate SACKs for SACKless connection: account for a new dupack. */
2041 static void tcp_add_reno_sack(struct sock *sk, int num_dupack, bool ece_ack)
2044 struct tcp_sock *tp = tcp_sk(sk);
2045 u32 prior_sacked = tp->sacked_out;
2048 tp->sacked_out += num_dupack;
2049 tcp_check_reno_reordering(sk, 0);
2050 delivered = tp->sacked_out - prior_sacked;
2052 tcp_count_delivered(tp, delivered, ece_ack);
2053 tcp_verify_left_out(tp);
2057 /* Account for ACK, ACKing some data in Reno Recovery phase. */
2059 static void tcp_remove_reno_sacks(struct sock *sk, int acked, bool ece_ack)
2061 struct tcp_sock *tp = tcp_sk(sk);
2064 /* One ACK acked hole. The rest eat duplicate ACKs. */
2065 tcp_count_delivered(tp, max_t(int, acked - tp->sacked_out, 1),
2067 if (acked - 1 >= tp->sacked_out)
2070 tp->sacked_out -= acked - 1;
2072 tcp_check_reno_reordering(sk, acked);
2073 tcp_verify_left_out(tp);
2076 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
2081 void tcp_clear_retrans(struct tcp_sock *tp)
2083 tp->retrans_out = 0;
2085 tp->undo_marker = 0;
2086 tp->undo_retrans = -1;
2090 static inline void tcp_init_undo(struct tcp_sock *tp)
2092 tp->undo_marker = tp->snd_una;
2093 /* Retransmission still in flight may cause DSACKs later. */
2094 tp->undo_retrans = tp->retrans_out ? : -1;
2097 static bool tcp_is_rack(const struct sock *sk)
2099 return READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_recovery) &
2100 TCP_RACK_LOSS_DETECTION;
2103 /* If we detect SACK reneging, forget all SACK information
2104 * and reset tags completely, otherwise preserve SACKs. If receiver
2105 * dropped its ofo queue, we will know this due to reneging detection.
2107 static void tcp_timeout_mark_lost(struct sock *sk)
2109 struct tcp_sock *tp = tcp_sk(sk);
2110 struct sk_buff *skb, *head;
2111 bool is_reneg; /* is receiver reneging on SACKs? */
2113 head = tcp_rtx_queue_head(sk);
2114 is_reneg = head && (TCP_SKB_CB(head)->sacked & TCPCB_SACKED_ACKED);
2116 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
2118 /* Mark SACK reneging until we recover from this loss event. */
2119 tp->is_sack_reneg = 1;
2120 } else if (tcp_is_reno(tp)) {
2121 tcp_reset_reno_sack(tp);
2125 skb_rbtree_walk_from(skb) {
2127 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
2128 else if (tcp_is_rack(sk) && skb != head &&
2129 tcp_rack_skb_timeout(tp, skb, 0) > 0)
2130 continue; /* Don't mark recently sent ones lost yet */
2131 tcp_mark_skb_lost(sk, skb);
2133 tcp_verify_left_out(tp);
2134 tcp_clear_all_retrans_hints(tp);
2137 /* Enter Loss state. */
2138 void tcp_enter_loss(struct sock *sk)
2140 const struct inet_connection_sock *icsk = inet_csk(sk);
2141 struct tcp_sock *tp = tcp_sk(sk);
2142 struct net *net = sock_net(sk);
2143 bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery;
2146 tcp_timeout_mark_lost(sk);
2148 /* Reduce ssthresh if it has not yet been made inside this window. */
2149 if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
2150 !after(tp->high_seq, tp->snd_una) ||
2151 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
2152 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2153 tp->prior_cwnd = tcp_snd_cwnd(tp);
2154 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2155 tcp_ca_event(sk, CA_EVENT_LOSS);
2158 tcp_snd_cwnd_set(tp, tcp_packets_in_flight(tp) + 1);
2159 tp->snd_cwnd_cnt = 0;
2160 tp->snd_cwnd_stamp = tcp_jiffies32;
2162 /* Timeout in disordered state after receiving substantial DUPACKs
2163 * suggests that the degree of reordering is over-estimated.
2165 reordering = READ_ONCE(net->ipv4.sysctl_tcp_reordering);
2166 if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
2167 tp->sacked_out >= reordering)
2168 tp->reordering = min_t(unsigned int, tp->reordering,
2171 tcp_set_ca_state(sk, TCP_CA_Loss);
2172 tp->high_seq = tp->snd_nxt;
2173 tcp_ecn_queue_cwr(tp);
2175 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
2176 * loss recovery is underway except recurring timeout(s) on
2177 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
2179 tp->frto = READ_ONCE(net->ipv4.sysctl_tcp_frto) &&
2180 (new_recovery || icsk->icsk_retransmits) &&
2181 !inet_csk(sk)->icsk_mtup.probe_size;
2184 /* If ACK arrived pointing to a remembered SACK, it means that our
2185 * remembered SACKs do not reflect real state of receiver i.e.
2186 * receiver _host_ is heavily congested (or buggy).
2188 * To avoid big spurious retransmission bursts due to transient SACK
2189 * scoreboard oddities that look like reneging, we give the receiver a
2190 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
2191 * restore sanity to the SACK scoreboard. If the apparent reneging
2192 * persists until this RTO then we'll clear the SACK scoreboard.
2194 static bool tcp_check_sack_reneging(struct sock *sk, int flag)
2196 if (flag & FLAG_SACK_RENEGING) {
2197 struct tcp_sock *tp = tcp_sk(sk);
2198 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4),
2199 msecs_to_jiffies(10));
2201 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2202 delay, TCP_RTO_MAX);
2208 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2209 * counter when SACK is enabled (without SACK, sacked_out is used for
2212 * With reordering, holes may still be in flight, so RFC3517 recovery
2213 * uses pure sacked_out (total number of SACKed segments) even though
2214 * it violates the RFC that uses duplicate ACKs, often these are equal
2215 * but when e.g. out-of-window ACKs or packet duplication occurs,
2216 * they differ. Since neither occurs due to loss, TCP should really
2219 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2221 return tp->sacked_out + 1;
2224 /* Linux NewReno/SACK/ECN state machine.
2225 * --------------------------------------
2227 * "Open" Normal state, no dubious events, fast path.
2228 * "Disorder" In all the respects it is "Open",
2229 * but requires a bit more attention. It is entered when
2230 * we see some SACKs or dupacks. It is split of "Open"
2231 * mainly to move some processing from fast path to slow one.
2232 * "CWR" CWND was reduced due to some Congestion Notification event.
2233 * It can be ECN, ICMP source quench, local device congestion.
2234 * "Recovery" CWND was reduced, we are fast-retransmitting.
2235 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2237 * tcp_fastretrans_alert() is entered:
2238 * - each incoming ACK, if state is not "Open"
2239 * - when arrived ACK is unusual, namely:
2244 * Counting packets in flight is pretty simple.
2246 * in_flight = packets_out - left_out + retrans_out
2248 * packets_out is SND.NXT-SND.UNA counted in packets.
2250 * retrans_out is number of retransmitted segments.
2252 * left_out is number of segments left network, but not ACKed yet.
2254 * left_out = sacked_out + lost_out
2256 * sacked_out: Packets, which arrived to receiver out of order
2257 * and hence not ACKed. With SACKs this number is simply
2258 * amount of SACKed data. Even without SACKs
2259 * it is easy to give pretty reliable estimate of this number,
2260 * counting duplicate ACKs.
2262 * lost_out: Packets lost by network. TCP has no explicit
2263 * "loss notification" feedback from network (for now).
2264 * It means that this number can be only _guessed_.
2265 * Actually, it is the heuristics to predict lossage that
2266 * distinguishes different algorithms.
2268 * F.e. after RTO, when all the queue is considered as lost,
2269 * lost_out = packets_out and in_flight = retrans_out.
2271 * Essentially, we have now a few algorithms detecting
2274 * If the receiver supports SACK:
2276 * RFC6675/3517: It is the conventional algorithm. A packet is
2277 * considered lost if the number of higher sequence packets
2278 * SACKed is greater than or equal the DUPACK thoreshold
2279 * (reordering). This is implemented in tcp_mark_head_lost and
2280 * tcp_update_scoreboard.
2282 * RACK (draft-ietf-tcpm-rack-01): it is a newer algorithm
2283 * (2017-) that checks timing instead of counting DUPACKs.
2284 * Essentially a packet is considered lost if it's not S/ACKed
2285 * after RTT + reordering_window, where both metrics are
2286 * dynamically measured and adjusted. This is implemented in
2287 * tcp_rack_mark_lost.
2289 * If the receiver does not support SACK:
2291 * NewReno (RFC6582): in Recovery we assume that one segment
2292 * is lost (classic Reno). While we are in Recovery and
2293 * a partial ACK arrives, we assume that one more packet
2294 * is lost (NewReno). This heuristics are the same in NewReno
2297 * Really tricky (and requiring careful tuning) part of algorithm
2298 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2299 * The first determines the moment _when_ we should reduce CWND and,
2300 * hence, slow down forward transmission. In fact, it determines the moment
2301 * when we decide that hole is caused by loss, rather than by a reorder.
2303 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2304 * holes, caused by lost packets.
2306 * And the most logically complicated part of algorithm is undo
2307 * heuristics. We detect false retransmits due to both too early
2308 * fast retransmit (reordering) and underestimated RTO, analyzing
2309 * timestamps and D-SACKs. When we detect that some segments were
2310 * retransmitted by mistake and CWND reduction was wrong, we undo
2311 * window reduction and abort recovery phase. This logic is hidden
2312 * inside several functions named tcp_try_undo_<something>.
2315 /* This function decides, when we should leave Disordered state
2316 * and enter Recovery phase, reducing congestion window.
2318 * Main question: may we further continue forward transmission
2319 * with the same cwnd?
2321 static bool tcp_time_to_recover(struct sock *sk, int flag)
2323 struct tcp_sock *tp = tcp_sk(sk);
2325 /* Trick#1: The loss is proven. */
2329 /* Not-A-Trick#2 : Classic rule... */
2330 if (!tcp_is_rack(sk) && tcp_dupack_heuristics(tp) > tp->reordering)
2336 /* Detect loss in event "A" above by marking head of queue up as lost.
2337 * For RFC3517 SACK, a segment is considered lost if it
2338 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2339 * the maximum SACKed segments to pass before reaching this limit.
2341 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2343 struct tcp_sock *tp = tcp_sk(sk);
2344 struct sk_buff *skb;
2346 /* Use SACK to deduce losses of new sequences sent during recovery */
2347 const u32 loss_high = tp->snd_nxt;
2349 WARN_ON(packets > tp->packets_out);
2350 skb = tp->lost_skb_hint;
2352 /* Head already handled? */
2353 if (mark_head && after(TCP_SKB_CB(skb)->seq, tp->snd_una))
2355 cnt = tp->lost_cnt_hint;
2357 skb = tcp_rtx_queue_head(sk);
2361 skb_rbtree_walk_from(skb) {
2362 /* TODO: do this better */
2363 /* this is not the most efficient way to do this... */
2364 tp->lost_skb_hint = skb;
2365 tp->lost_cnt_hint = cnt;
2367 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2370 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2371 cnt += tcp_skb_pcount(skb);
2376 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_LOST))
2377 tcp_mark_skb_lost(sk, skb);
2382 tcp_verify_left_out(tp);
2385 /* Account newly detected lost packet(s) */
2387 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2389 struct tcp_sock *tp = tcp_sk(sk);
2391 if (tcp_is_sack(tp)) {
2392 int sacked_upto = tp->sacked_out - tp->reordering;
2393 if (sacked_upto >= 0)
2394 tcp_mark_head_lost(sk, sacked_upto, 0);
2395 else if (fast_rexmit)
2396 tcp_mark_head_lost(sk, 1, 1);
2400 static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when)
2402 return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2403 before(tp->rx_opt.rcv_tsecr, when);
2406 /* skb is spurious retransmitted if the returned timestamp echo
2407 * reply is prior to the skb transmission time
2409 static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp,
2410 const struct sk_buff *skb)
2412 return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) &&
2413 tcp_tsopt_ecr_before(tp, tcp_skb_timestamp(skb));
2416 /* Nothing was retransmitted or returned timestamp is less
2417 * than timestamp of the first retransmission.
2419 static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2421 return tp->retrans_stamp &&
2422 tcp_tsopt_ecr_before(tp, tp->retrans_stamp);
2425 /* Undo procedures. */
2427 /* We can clear retrans_stamp when there are no retransmissions in the
2428 * window. It would seem that it is trivially available for us in
2429 * tp->retrans_out, however, that kind of assumptions doesn't consider
2430 * what will happen if errors occur when sending retransmission for the
2431 * second time. ...It could the that such segment has only
2432 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2433 * the head skb is enough except for some reneging corner cases that
2434 * are not worth the effort.
2436 * Main reason for all this complexity is the fact that connection dying
2437 * time now depends on the validity of the retrans_stamp, in particular,
2438 * that successive retransmissions of a segment must not advance
2439 * retrans_stamp under any conditions.
2441 static bool tcp_any_retrans_done(const struct sock *sk)
2443 const struct tcp_sock *tp = tcp_sk(sk);
2444 struct sk_buff *skb;
2446 if (tp->retrans_out)
2449 skb = tcp_rtx_queue_head(sk);
2450 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2456 static void DBGUNDO(struct sock *sk, const char *msg)
2458 #if FASTRETRANS_DEBUG > 1
2459 struct tcp_sock *tp = tcp_sk(sk);
2460 struct inet_sock *inet = inet_sk(sk);
2462 if (sk->sk_family == AF_INET) {
2463 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2465 &inet->inet_daddr, ntohs(inet->inet_dport),
2466 tcp_snd_cwnd(tp), tcp_left_out(tp),
2467 tp->snd_ssthresh, tp->prior_ssthresh,
2470 #if IS_ENABLED(CONFIG_IPV6)
2471 else if (sk->sk_family == AF_INET6) {
2472 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2474 &sk->sk_v6_daddr, ntohs(inet->inet_dport),
2475 tcp_snd_cwnd(tp), tcp_left_out(tp),
2476 tp->snd_ssthresh, tp->prior_ssthresh,
2483 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
2485 struct tcp_sock *tp = tcp_sk(sk);
2488 struct sk_buff *skb;
2490 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2491 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2494 tcp_clear_all_retrans_hints(tp);
2497 if (tp->prior_ssthresh) {
2498 const struct inet_connection_sock *icsk = inet_csk(sk);
2500 tcp_snd_cwnd_set(tp, icsk->icsk_ca_ops->undo_cwnd(sk));
2502 if (tp->prior_ssthresh > tp->snd_ssthresh) {
2503 tp->snd_ssthresh = tp->prior_ssthresh;
2504 tcp_ecn_withdraw_cwr(tp);
2507 tp->snd_cwnd_stamp = tcp_jiffies32;
2508 tp->undo_marker = 0;
2509 tp->rack.advanced = 1; /* Force RACK to re-exam losses */
2512 static inline bool tcp_may_undo(const struct tcp_sock *tp)
2514 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2517 /* People celebrate: "We love our President!" */
2518 static bool tcp_try_undo_recovery(struct sock *sk)
2520 struct tcp_sock *tp = tcp_sk(sk);
2522 if (tcp_may_undo(tp)) {
2525 /* Happy end! We did not retransmit anything
2526 * or our original transmission succeeded.
2528 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2529 tcp_undo_cwnd_reduction(sk, false);
2530 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2531 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2533 mib_idx = LINUX_MIB_TCPFULLUNDO;
2535 NET_INC_STATS(sock_net(sk), mib_idx);
2536 } else if (tp->rack.reo_wnd_persist) {
2537 tp->rack.reo_wnd_persist--;
2539 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2540 /* Hold old state until something *above* high_seq
2541 * is ACKed. For Reno it is MUST to prevent false
2542 * fast retransmits (RFC2582). SACK TCP is safe. */
2543 if (!tcp_any_retrans_done(sk))
2544 tp->retrans_stamp = 0;
2547 tcp_set_ca_state(sk, TCP_CA_Open);
2548 tp->is_sack_reneg = 0;
2552 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2553 static bool tcp_try_undo_dsack(struct sock *sk)
2555 struct tcp_sock *tp = tcp_sk(sk);
2557 if (tp->undo_marker && !tp->undo_retrans) {
2558 tp->rack.reo_wnd_persist = min(TCP_RACK_RECOVERY_THRESH,
2559 tp->rack.reo_wnd_persist + 1);
2560 DBGUNDO(sk, "D-SACK");
2561 tcp_undo_cwnd_reduction(sk, false);
2562 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2568 /* Undo during loss recovery after partial ACK or using F-RTO. */
2569 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2571 struct tcp_sock *tp = tcp_sk(sk);
2573 if (frto_undo || tcp_may_undo(tp)) {
2574 tcp_undo_cwnd_reduction(sk, true);
2576 DBGUNDO(sk, "partial loss");
2577 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2579 NET_INC_STATS(sock_net(sk),
2580 LINUX_MIB_TCPSPURIOUSRTOS);
2581 inet_csk(sk)->icsk_retransmits = 0;
2582 if (frto_undo || tcp_is_sack(tp)) {
2583 tcp_set_ca_state(sk, TCP_CA_Open);
2584 tp->is_sack_reneg = 0;
2591 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
2592 * It computes the number of packets to send (sndcnt) based on packets newly
2594 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2595 * cwnd reductions across a full RTT.
2596 * 2) Otherwise PRR uses packet conservation to send as much as delivered.
2597 * But when SND_UNA is acked without further losses,
2598 * slow starts cwnd up to ssthresh to speed up the recovery.
2600 static void tcp_init_cwnd_reduction(struct sock *sk)
2602 struct tcp_sock *tp = tcp_sk(sk);
2604 tp->high_seq = tp->snd_nxt;
2605 tp->tlp_high_seq = 0;
2606 tp->snd_cwnd_cnt = 0;
2607 tp->prior_cwnd = tcp_snd_cwnd(tp);
2608 tp->prr_delivered = 0;
2610 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2611 tcp_ecn_queue_cwr(tp);
2614 void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked, int newly_lost, int flag)
2616 struct tcp_sock *tp = tcp_sk(sk);
2618 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2620 if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd))
2623 tp->prr_delivered += newly_acked_sacked;
2625 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2627 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2629 sndcnt = max_t(int, tp->prr_delivered - tp->prr_out,
2630 newly_acked_sacked);
2631 if (flag & FLAG_SND_UNA_ADVANCED && !newly_lost)
2633 sndcnt = min(delta, sndcnt);
2635 /* Force a fast retransmit upon entering fast recovery */
2636 sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1));
2637 tcp_snd_cwnd_set(tp, tcp_packets_in_flight(tp) + sndcnt);
2640 static inline void tcp_end_cwnd_reduction(struct sock *sk)
2642 struct tcp_sock *tp = tcp_sk(sk);
2644 if (inet_csk(sk)->icsk_ca_ops->cong_control)
2647 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2648 if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH &&
2649 (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR || tp->undo_marker)) {
2650 tcp_snd_cwnd_set(tp, tp->snd_ssthresh);
2651 tp->snd_cwnd_stamp = tcp_jiffies32;
2653 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2656 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2657 void tcp_enter_cwr(struct sock *sk)
2659 struct tcp_sock *tp = tcp_sk(sk);
2661 tp->prior_ssthresh = 0;
2662 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2663 tp->undo_marker = 0;
2664 tcp_init_cwnd_reduction(sk);
2665 tcp_set_ca_state(sk, TCP_CA_CWR);
2668 EXPORT_SYMBOL(tcp_enter_cwr);
2670 static void tcp_try_keep_open(struct sock *sk)
2672 struct tcp_sock *tp = tcp_sk(sk);
2673 int state = TCP_CA_Open;
2675 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2676 state = TCP_CA_Disorder;
2678 if (inet_csk(sk)->icsk_ca_state != state) {
2679 tcp_set_ca_state(sk, state);
2680 tp->high_seq = tp->snd_nxt;
2684 static void tcp_try_to_open(struct sock *sk, int flag)
2686 struct tcp_sock *tp = tcp_sk(sk);
2688 tcp_verify_left_out(tp);
2690 if (!tcp_any_retrans_done(sk))
2691 tp->retrans_stamp = 0;
2693 if (flag & FLAG_ECE)
2696 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2697 tcp_try_keep_open(sk);
2701 static void tcp_mtup_probe_failed(struct sock *sk)
2703 struct inet_connection_sock *icsk = inet_csk(sk);
2705 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2706 icsk->icsk_mtup.probe_size = 0;
2707 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPFAIL);
2710 static void tcp_mtup_probe_success(struct sock *sk)
2712 struct tcp_sock *tp = tcp_sk(sk);
2713 struct inet_connection_sock *icsk = inet_csk(sk);
2716 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2718 val = (u64)tcp_snd_cwnd(tp) * tcp_mss_to_mtu(sk, tp->mss_cache);
2719 do_div(val, icsk->icsk_mtup.probe_size);
2720 DEBUG_NET_WARN_ON_ONCE((u32)val != val);
2721 tcp_snd_cwnd_set(tp, max_t(u32, 1U, val));
2723 tp->snd_cwnd_cnt = 0;
2724 tp->snd_cwnd_stamp = tcp_jiffies32;
2725 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2727 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2728 icsk->icsk_mtup.probe_size = 0;
2729 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2730 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS);
2733 /* Do a simple retransmit without using the backoff mechanisms in
2734 * tcp_timer. This is used for path mtu discovery.
2735 * The socket is already locked here.
2737 void tcp_simple_retransmit(struct sock *sk)
2739 const struct inet_connection_sock *icsk = inet_csk(sk);
2740 struct tcp_sock *tp = tcp_sk(sk);
2741 struct sk_buff *skb;
2744 /* A fastopen SYN request is stored as two separate packets within
2745 * the retransmit queue, this is done by tcp_send_syn_data().
2746 * As a result simply checking the MSS of the frames in the queue
2747 * will not work for the SYN packet.
2749 * Us being here is an indication of a path MTU issue so we can
2750 * assume that the fastopen SYN was lost and just mark all the
2751 * frames in the retransmit queue as lost. We will use an MSS of
2752 * -1 to mark all frames as lost, otherwise compute the current MSS.
2754 if (tp->syn_data && sk->sk_state == TCP_SYN_SENT)
2757 mss = tcp_current_mss(sk);
2759 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2760 if (tcp_skb_seglen(skb) > mss)
2761 tcp_mark_skb_lost(sk, skb);
2764 tcp_clear_retrans_hints_partial(tp);
2769 if (tcp_is_reno(tp))
2770 tcp_limit_reno_sacked(tp);
2772 tcp_verify_left_out(tp);
2774 /* Don't muck with the congestion window here.
2775 * Reason is that we do not increase amount of _data_
2776 * in network, but units changed and effective
2777 * cwnd/ssthresh really reduced now.
2779 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2780 tp->high_seq = tp->snd_nxt;
2781 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2782 tp->prior_ssthresh = 0;
2783 tp->undo_marker = 0;
2784 tcp_set_ca_state(sk, TCP_CA_Loss);
2786 tcp_xmit_retransmit_queue(sk);
2788 EXPORT_SYMBOL(tcp_simple_retransmit);
2790 void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2792 struct tcp_sock *tp = tcp_sk(sk);
2795 if (tcp_is_reno(tp))
2796 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2798 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2800 NET_INC_STATS(sock_net(sk), mib_idx);
2802 tp->prior_ssthresh = 0;
2805 if (!tcp_in_cwnd_reduction(sk)) {
2807 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2808 tcp_init_cwnd_reduction(sk);
2810 tcp_set_ca_state(sk, TCP_CA_Recovery);
2813 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2814 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2816 static void tcp_process_loss(struct sock *sk, int flag, int num_dupack,
2819 struct tcp_sock *tp = tcp_sk(sk);
2820 bool recovered = !before(tp->snd_una, tp->high_seq);
2822 if ((flag & FLAG_SND_UNA_ADVANCED || rcu_access_pointer(tp->fastopen_rsk)) &&
2823 tcp_try_undo_loss(sk, false))
2826 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2827 /* Step 3.b. A timeout is spurious if not all data are
2828 * lost, i.e., never-retransmitted data are (s)acked.
2830 if ((flag & FLAG_ORIG_SACK_ACKED) &&
2831 tcp_try_undo_loss(sk, true))
2834 if (after(tp->snd_nxt, tp->high_seq)) {
2835 if (flag & FLAG_DATA_SACKED || num_dupack)
2836 tp->frto = 0; /* Step 3.a. loss was real */
2837 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2838 tp->high_seq = tp->snd_nxt;
2839 /* Step 2.b. Try send new data (but deferred until cwnd
2840 * is updated in tcp_ack()). Otherwise fall back to
2841 * the conventional recovery.
2843 if (!tcp_write_queue_empty(sk) &&
2844 after(tcp_wnd_end(tp), tp->snd_nxt)) {
2845 *rexmit = REXMIT_NEW;
2853 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2854 tcp_try_undo_recovery(sk);
2857 if (tcp_is_reno(tp)) {
2858 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2859 * delivered. Lower inflight to clock out (re)tranmissions.
2861 if (after(tp->snd_nxt, tp->high_seq) && num_dupack)
2862 tcp_add_reno_sack(sk, num_dupack, flag & FLAG_ECE);
2863 else if (flag & FLAG_SND_UNA_ADVANCED)
2864 tcp_reset_reno_sack(tp);
2866 *rexmit = REXMIT_LOST;
2869 static bool tcp_force_fast_retransmit(struct sock *sk)
2871 struct tcp_sock *tp = tcp_sk(sk);
2873 return after(tcp_highest_sack_seq(tp),
2874 tp->snd_una + tp->reordering * tp->mss_cache);
2877 /* Undo during fast recovery after partial ACK. */
2878 static bool tcp_try_undo_partial(struct sock *sk, u32 prior_snd_una,
2881 struct tcp_sock *tp = tcp_sk(sk);
2883 if (tp->undo_marker && tcp_packet_delayed(tp)) {
2884 /* Plain luck! Hole if filled with delayed
2885 * packet, rather than with a retransmit. Check reordering.
2887 tcp_check_sack_reordering(sk, prior_snd_una, 1);
2889 /* We are getting evidence that the reordering degree is higher
2890 * than we realized. If there are no retransmits out then we
2891 * can undo. Otherwise we clock out new packets but do not
2892 * mark more packets lost or retransmit more.
2894 if (tp->retrans_out)
2897 if (!tcp_any_retrans_done(sk))
2898 tp->retrans_stamp = 0;
2900 DBGUNDO(sk, "partial recovery");
2901 tcp_undo_cwnd_reduction(sk, true);
2902 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2903 tcp_try_keep_open(sk);
2905 /* Partial ACK arrived. Force fast retransmit. */
2906 *do_lost = tcp_force_fast_retransmit(sk);
2911 static void tcp_identify_packet_loss(struct sock *sk, int *ack_flag)
2913 struct tcp_sock *tp = tcp_sk(sk);
2915 if (tcp_rtx_queue_empty(sk))
2918 if (unlikely(tcp_is_reno(tp))) {
2919 tcp_newreno_mark_lost(sk, *ack_flag & FLAG_SND_UNA_ADVANCED);
2920 } else if (tcp_is_rack(sk)) {
2921 u32 prior_retrans = tp->retrans_out;
2923 if (tcp_rack_mark_lost(sk))
2924 *ack_flag &= ~FLAG_SET_XMIT_TIMER;
2925 if (prior_retrans > tp->retrans_out)
2926 *ack_flag |= FLAG_LOST_RETRANS;
2930 /* Process an event, which can update packets-in-flight not trivially.
2931 * Main goal of this function is to calculate new estimate for left_out,
2932 * taking into account both packets sitting in receiver's buffer and
2933 * packets lost by network.
2935 * Besides that it updates the congestion state when packet loss or ECN
2936 * is detected. But it does not reduce the cwnd, it is done by the
2937 * congestion control later.
2939 * It does _not_ decide what to send, it is made in function
2940 * tcp_xmit_retransmit_queue().
2942 static void tcp_fastretrans_alert(struct sock *sk, const u32 prior_snd_una,
2943 int num_dupack, int *ack_flag, int *rexmit)
2945 struct inet_connection_sock *icsk = inet_csk(sk);
2946 struct tcp_sock *tp = tcp_sk(sk);
2947 int fast_rexmit = 0, flag = *ack_flag;
2948 bool ece_ack = flag & FLAG_ECE;
2949 bool do_lost = num_dupack || ((flag & FLAG_DATA_SACKED) &&
2950 tcp_force_fast_retransmit(sk));
2952 if (!tp->packets_out && tp->sacked_out)
2955 /* Now state machine starts.
2956 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2958 tp->prior_ssthresh = 0;
2960 /* B. In all the states check for reneging SACKs. */
2961 if (tcp_check_sack_reneging(sk, flag))
2964 /* C. Check consistency of the current state. */
2965 tcp_verify_left_out(tp);
2967 /* D. Check state exit conditions. State can be terminated
2968 * when high_seq is ACKed. */
2969 if (icsk->icsk_ca_state == TCP_CA_Open) {
2970 WARN_ON(tp->retrans_out != 0 && !tp->syn_data);
2971 tp->retrans_stamp = 0;
2972 } else if (!before(tp->snd_una, tp->high_seq)) {
2973 switch (icsk->icsk_ca_state) {
2975 /* CWR is to be held something *above* high_seq
2976 * is ACKed for CWR bit to reach receiver. */
2977 if (tp->snd_una != tp->high_seq) {
2978 tcp_end_cwnd_reduction(sk);
2979 tcp_set_ca_state(sk, TCP_CA_Open);
2983 case TCP_CA_Recovery:
2984 if (tcp_is_reno(tp))
2985 tcp_reset_reno_sack(tp);
2986 if (tcp_try_undo_recovery(sk))
2988 tcp_end_cwnd_reduction(sk);
2993 /* E. Process state. */
2994 switch (icsk->icsk_ca_state) {
2995 case TCP_CA_Recovery:
2996 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2997 if (tcp_is_reno(tp))
2998 tcp_add_reno_sack(sk, num_dupack, ece_ack);
2999 } else if (tcp_try_undo_partial(sk, prior_snd_una, &do_lost))
3002 if (tcp_try_undo_dsack(sk))
3003 tcp_try_keep_open(sk);
3005 tcp_identify_packet_loss(sk, ack_flag);
3006 if (icsk->icsk_ca_state != TCP_CA_Recovery) {
3007 if (!tcp_time_to_recover(sk, flag))
3009 /* Undo reverts the recovery state. If loss is evident,
3010 * starts a new recovery (e.g. reordering then loss);
3012 tcp_enter_recovery(sk, ece_ack);
3016 tcp_process_loss(sk, flag, num_dupack, rexmit);
3017 tcp_identify_packet_loss(sk, ack_flag);
3018 if (!(icsk->icsk_ca_state == TCP_CA_Open ||
3019 (*ack_flag & FLAG_LOST_RETRANS)))
3021 /* Change state if cwnd is undone or retransmits are lost */
3024 if (tcp_is_reno(tp)) {
3025 if (flag & FLAG_SND_UNA_ADVANCED)
3026 tcp_reset_reno_sack(tp);
3027 tcp_add_reno_sack(sk, num_dupack, ece_ack);
3030 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
3031 tcp_try_undo_dsack(sk);
3033 tcp_identify_packet_loss(sk, ack_flag);
3034 if (!tcp_time_to_recover(sk, flag)) {
3035 tcp_try_to_open(sk, flag);
3039 /* MTU probe failure: don't reduce cwnd */
3040 if (icsk->icsk_ca_state < TCP_CA_CWR &&
3041 icsk->icsk_mtup.probe_size &&
3042 tp->snd_una == tp->mtu_probe.probe_seq_start) {
3043 tcp_mtup_probe_failed(sk);
3044 /* Restores the reduction we did in tcp_mtup_probe() */
3045 tcp_snd_cwnd_set(tp, tcp_snd_cwnd(tp) + 1);
3046 tcp_simple_retransmit(sk);
3050 /* Otherwise enter Recovery state */
3051 tcp_enter_recovery(sk, ece_ack);
3055 if (!tcp_is_rack(sk) && do_lost)
3056 tcp_update_scoreboard(sk, fast_rexmit);
3057 *rexmit = REXMIT_LOST;
3060 static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us, const int flag)
3062 u32 wlen = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_rtt_wlen) * HZ;
3063 struct tcp_sock *tp = tcp_sk(sk);
3065 if ((flag & FLAG_ACK_MAYBE_DELAYED) && rtt_us > tcp_min_rtt(tp)) {
3066 /* If the remote keeps returning delayed ACKs, eventually
3067 * the min filter would pick it up and overestimate the
3068 * prop. delay when it expires. Skip suspected delayed ACKs.
3072 minmax_running_min(&tp->rtt_min, wlen, tcp_jiffies32,
3073 rtt_us ? : jiffies_to_usecs(1));
3076 static bool tcp_ack_update_rtt(struct sock *sk, const int flag,
3077 long seq_rtt_us, long sack_rtt_us,
3078 long ca_rtt_us, struct rate_sample *rs)
3080 const struct tcp_sock *tp = tcp_sk(sk);
3082 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
3083 * broken middle-boxes or peers may corrupt TS-ECR fields. But
3084 * Karn's algorithm forbids taking RTT if some retransmitted data
3085 * is acked (RFC6298).
3088 seq_rtt_us = sack_rtt_us;
3090 /* RTTM Rule: A TSecr value received in a segment is used to
3091 * update the averaged RTT measurement only if the segment
3092 * acknowledges some new data, i.e., only if it advances the
3093 * left edge of the send window.
3094 * See draft-ietf-tcplw-high-performance-00, section 3.3.
3096 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
3097 flag & FLAG_ACKED) {
3098 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr;
3100 if (likely(delta < INT_MAX / (USEC_PER_SEC / TCP_TS_HZ))) {
3103 seq_rtt_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
3104 ca_rtt_us = seq_rtt_us;
3107 rs->rtt_us = ca_rtt_us; /* RTT of last (S)ACKed packet (or -1) */
3111 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
3112 * always taken together with ACK, SACK, or TS-opts. Any negative
3113 * values will be skipped with the seq_rtt_us < 0 check above.
3115 tcp_update_rtt_min(sk, ca_rtt_us, flag);
3116 tcp_rtt_estimator(sk, seq_rtt_us);
3119 /* RFC6298: only reset backoff on valid RTT measurement. */
3120 inet_csk(sk)->icsk_backoff = 0;
3124 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
3125 void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req)
3127 struct rate_sample rs;
3130 if (req && !req->num_retrans && tcp_rsk(req)->snt_synack)
3131 rtt_us = tcp_stamp_us_delta(tcp_clock_us(), tcp_rsk(req)->snt_synack);
3133 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us, &rs);
3137 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
3139 const struct inet_connection_sock *icsk = inet_csk(sk);
3141 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
3142 tcp_sk(sk)->snd_cwnd_stamp = tcp_jiffies32;
3145 /* Restart timer after forward progress on connection.
3146 * RFC2988 recommends to restart timer to now+rto.
3148 void tcp_rearm_rto(struct sock *sk)
3150 const struct inet_connection_sock *icsk = inet_csk(sk);
3151 struct tcp_sock *tp = tcp_sk(sk);
3153 /* If the retrans timer is currently being used by Fast Open
3154 * for SYN-ACK retrans purpose, stay put.
3156 if (rcu_access_pointer(tp->fastopen_rsk))
3159 if (!tp->packets_out) {
3160 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
3162 u32 rto = inet_csk(sk)->icsk_rto;
3163 /* Offset the time elapsed after installing regular RTO */
3164 if (icsk->icsk_pending == ICSK_TIME_REO_TIMEOUT ||
3165 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
3166 s64 delta_us = tcp_rto_delta_us(sk);
3167 /* delta_us may not be positive if the socket is locked
3168 * when the retrans timer fires and is rescheduled.
3170 rto = usecs_to_jiffies(max_t(int, delta_us, 1));
3172 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
3177 /* Try to schedule a loss probe; if that doesn't work, then schedule an RTO. */
3178 static void tcp_set_xmit_timer(struct sock *sk)
3180 if (!tcp_schedule_loss_probe(sk, true))
3184 /* If we get here, the whole TSO packet has not been acked. */
3185 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3187 struct tcp_sock *tp = tcp_sk(sk);
3190 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3192 packets_acked = tcp_skb_pcount(skb);
3193 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3195 packets_acked -= tcp_skb_pcount(skb);
3197 if (packets_acked) {
3198 BUG_ON(tcp_skb_pcount(skb) == 0);
3199 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3202 return packets_acked;
3205 static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
3206 const struct sk_buff *ack_skb, u32 prior_snd_una)
3208 const struct skb_shared_info *shinfo;
3210 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3211 if (likely(!TCP_SKB_CB(skb)->txstamp_ack))
3214 shinfo = skb_shinfo(skb);
3215 if (!before(shinfo->tskey, prior_snd_una) &&
3216 before(shinfo->tskey, tcp_sk(sk)->snd_una)) {
3217 tcp_skb_tsorted_save(skb) {
3218 __skb_tstamp_tx(skb, ack_skb, NULL, sk, SCM_TSTAMP_ACK);
3219 } tcp_skb_tsorted_restore(skb);
3223 /* Remove acknowledged frames from the retransmission queue. If our packet
3224 * is before the ack sequence we can discard it as it's confirmed to have
3225 * arrived at the other end.
3227 static int tcp_clean_rtx_queue(struct sock *sk, const struct sk_buff *ack_skb,
3228 u32 prior_fack, u32 prior_snd_una,
3229 struct tcp_sacktag_state *sack, bool ece_ack)
3231 const struct inet_connection_sock *icsk = inet_csk(sk);
3232 u64 first_ackt, last_ackt;
3233 struct tcp_sock *tp = tcp_sk(sk);
3234 u32 prior_sacked = tp->sacked_out;
3235 u32 reord = tp->snd_nxt; /* lowest acked un-retx un-sacked seq */
3236 struct sk_buff *skb, *next;
3237 bool fully_acked = true;
3238 long sack_rtt_us = -1L;
3239 long seq_rtt_us = -1L;
3240 long ca_rtt_us = -1L;
3247 for (skb = skb_rb_first(&sk->tcp_rtx_queue); skb; skb = next) {
3248 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3249 const u32 start_seq = scb->seq;
3250 u8 sacked = scb->sacked;
3253 /* Determine how many packets and what bytes were acked, tso and else */
3254 if (after(scb->end_seq, tp->snd_una)) {
3255 if (tcp_skb_pcount(skb) == 1 ||
3256 !after(tp->snd_una, scb->seq))
3259 acked_pcount = tcp_tso_acked(sk, skb);
3262 fully_acked = false;
3264 acked_pcount = tcp_skb_pcount(skb);
3267 if (unlikely(sacked & TCPCB_RETRANS)) {
3268 if (sacked & TCPCB_SACKED_RETRANS)
3269 tp->retrans_out -= acked_pcount;
3270 flag |= FLAG_RETRANS_DATA_ACKED;
3271 } else if (!(sacked & TCPCB_SACKED_ACKED)) {
3272 last_ackt = tcp_skb_timestamp_us(skb);
3273 WARN_ON_ONCE(last_ackt == 0);
3275 first_ackt = last_ackt;
3277 if (before(start_seq, reord))
3279 if (!after(scb->end_seq, tp->high_seq))
3280 flag |= FLAG_ORIG_SACK_ACKED;
3283 if (sacked & TCPCB_SACKED_ACKED) {
3284 tp->sacked_out -= acked_pcount;
3285 } else if (tcp_is_sack(tp)) {
3286 tcp_count_delivered(tp, acked_pcount, ece_ack);
3287 if (!tcp_skb_spurious_retrans(tp, skb))
3288 tcp_rack_advance(tp, sacked, scb->end_seq,
3289 tcp_skb_timestamp_us(skb));
3291 if (sacked & TCPCB_LOST)
3292 tp->lost_out -= acked_pcount;
3294 tp->packets_out -= acked_pcount;
3295 pkts_acked += acked_pcount;
3296 tcp_rate_skb_delivered(sk, skb, sack->rate);
3298 /* Initial outgoing SYN's get put onto the write_queue
3299 * just like anything else we transmit. It is not
3300 * true data, and if we misinform our callers that
3301 * this ACK acks real data, we will erroneously exit
3302 * connection startup slow start one packet too
3303 * quickly. This is severely frowned upon behavior.
3305 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
3306 flag |= FLAG_DATA_ACKED;
3308 flag |= FLAG_SYN_ACKED;
3309 tp->retrans_stamp = 0;
3315 tcp_ack_tstamp(sk, skb, ack_skb, prior_snd_una);
3317 next = skb_rb_next(skb);
3318 if (unlikely(skb == tp->retransmit_skb_hint))
3319 tp->retransmit_skb_hint = NULL;
3320 if (unlikely(skb == tp->lost_skb_hint))
3321 tp->lost_skb_hint = NULL;
3322 tcp_highest_sack_replace(sk, skb, next);
3323 tcp_rtx_queue_unlink_and_free(skb, sk);
3327 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
3329 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3330 tp->snd_up = tp->snd_una;
3333 tcp_ack_tstamp(sk, skb, ack_skb, prior_snd_una);
3334 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
3335 flag |= FLAG_SACK_RENEGING;
3338 if (likely(first_ackt) && !(flag & FLAG_RETRANS_DATA_ACKED)) {
3339 seq_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, first_ackt);
3340 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, last_ackt);
3342 if (pkts_acked == 1 && fully_acked && !prior_sacked &&
3343 (tp->snd_una - prior_snd_una) < tp->mss_cache &&
3344 sack->rate->prior_delivered + 1 == tp->delivered &&
3345 !(flag & (FLAG_CA_ALERT | FLAG_SYN_ACKED))) {
3346 /* Conservatively mark a delayed ACK. It's typically
3347 * from a lone runt packet over the round trip to
3348 * a receiver w/o out-of-order or CE events.
3350 flag |= FLAG_ACK_MAYBE_DELAYED;
3353 if (sack->first_sackt) {
3354 sack_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->first_sackt);
3355 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->last_sackt);
3357 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us,
3358 ca_rtt_us, sack->rate);
3360 if (flag & FLAG_ACKED) {
3361 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3362 if (unlikely(icsk->icsk_mtup.probe_size &&
3363 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3364 tcp_mtup_probe_success(sk);
3367 if (tcp_is_reno(tp)) {
3368 tcp_remove_reno_sacks(sk, pkts_acked, ece_ack);
3370 /* If any of the cumulatively ACKed segments was
3371 * retransmitted, non-SACK case cannot confirm that
3372 * progress was due to original transmission due to
3373 * lack of TCPCB_SACKED_ACKED bits even if some of
3374 * the packets may have been never retransmitted.
3376 if (flag & FLAG_RETRANS_DATA_ACKED)
3377 flag &= ~FLAG_ORIG_SACK_ACKED;
3381 /* Non-retransmitted hole got filled? That's reordering */
3382 if (before(reord, prior_fack))
3383 tcp_check_sack_reordering(sk, reord, 0);
3385 delta = prior_sacked - tp->sacked_out;
3386 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3388 } else if (skb && rtt_update && sack_rtt_us >= 0 &&
3389 sack_rtt_us > tcp_stamp_us_delta(tp->tcp_mstamp,
3390 tcp_skb_timestamp_us(skb))) {
3391 /* Do not re-arm RTO if the sack RTT is measured from data sent
3392 * after when the head was last (re)transmitted. Otherwise the
3393 * timeout may continue to extend in loss recovery.
3395 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3398 if (icsk->icsk_ca_ops->pkts_acked) {
3399 struct ack_sample sample = { .pkts_acked = pkts_acked,
3400 .rtt_us = sack->rate->rtt_us };
3402 sample.in_flight = tp->mss_cache *
3403 (tp->delivered - sack->rate->prior_delivered);
3404 icsk->icsk_ca_ops->pkts_acked(sk, &sample);
3407 #if FASTRETRANS_DEBUG > 0
3408 WARN_ON((int)tp->sacked_out < 0);
3409 WARN_ON((int)tp->lost_out < 0);
3410 WARN_ON((int)tp->retrans_out < 0);
3411 if (!tp->packets_out && tcp_is_sack(tp)) {
3412 icsk = inet_csk(sk);
3414 pr_debug("Leak l=%u %d\n",
3415 tp->lost_out, icsk->icsk_ca_state);
3418 if (tp->sacked_out) {
3419 pr_debug("Leak s=%u %d\n",
3420 tp->sacked_out, icsk->icsk_ca_state);
3423 if (tp->retrans_out) {
3424 pr_debug("Leak r=%u %d\n",
3425 tp->retrans_out, icsk->icsk_ca_state);
3426 tp->retrans_out = 0;
3433 static void tcp_ack_probe(struct sock *sk)
3435 struct inet_connection_sock *icsk = inet_csk(sk);
3436 struct sk_buff *head = tcp_send_head(sk);
3437 const struct tcp_sock *tp = tcp_sk(sk);
3439 /* Was it a usable window open? */
3442 if (!after(TCP_SKB_CB(head)->end_seq, tcp_wnd_end(tp))) {
3443 icsk->icsk_backoff = 0;
3444 icsk->icsk_probes_tstamp = 0;
3445 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3446 /* Socket must be waked up by subsequent tcp_data_snd_check().
3447 * This function is not for random using!
3450 unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX);
3452 when = tcp_clamp_probe0_to_user_timeout(sk, when);
3453 tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, when, TCP_RTO_MAX);
3457 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3459 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3460 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3463 /* Decide wheather to run the increase function of congestion control. */
3464 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3466 /* If reordering is high then always grow cwnd whenever data is
3467 * delivered regardless of its ordering. Otherwise stay conservative
3468 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3469 * new SACK or ECE mark may first advance cwnd here and later reduce
3470 * cwnd in tcp_fastretrans_alert() based on more states.
3472 if (tcp_sk(sk)->reordering >
3473 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_reordering))
3474 return flag & FLAG_FORWARD_PROGRESS;
3476 return flag & FLAG_DATA_ACKED;
3479 /* The "ultimate" congestion control function that aims to replace the rigid
3480 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction).
3481 * It's called toward the end of processing an ACK with precise rate
3482 * information. All transmission or retransmission are delayed afterwards.
3484 static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked,
3485 int flag, const struct rate_sample *rs)
3487 const struct inet_connection_sock *icsk = inet_csk(sk);
3489 if (icsk->icsk_ca_ops->cong_control) {
3490 icsk->icsk_ca_ops->cong_control(sk, rs);
3494 if (tcp_in_cwnd_reduction(sk)) {
3495 /* Reduce cwnd if state mandates */
3496 tcp_cwnd_reduction(sk, acked_sacked, rs->losses, flag);
3497 } else if (tcp_may_raise_cwnd(sk, flag)) {
3498 /* Advance cwnd if state allows */
3499 tcp_cong_avoid(sk, ack, acked_sacked);
3501 tcp_update_pacing_rate(sk);
3504 /* Check that window update is acceptable.
3505 * The function assumes that snd_una<=ack<=snd_next.
3507 static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3508 const u32 ack, const u32 ack_seq,
3511 return after(ack, tp->snd_una) ||
3512 after(ack_seq, tp->snd_wl1) ||
3513 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3516 /* If we update tp->snd_una, also update tp->bytes_acked */
3517 static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack)
3519 u32 delta = ack - tp->snd_una;
3521 sock_owned_by_me((struct sock *)tp);
3522 tp->bytes_acked += delta;
3526 /* If we update tp->rcv_nxt, also update tp->bytes_received */
3527 static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq)
3529 u32 delta = seq - tp->rcv_nxt;
3531 sock_owned_by_me((struct sock *)tp);
3532 tp->bytes_received += delta;
3533 WRITE_ONCE(tp->rcv_nxt, seq);
3536 /* Update our send window.
3538 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3539 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3541 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3544 struct tcp_sock *tp = tcp_sk(sk);
3546 u32 nwin = ntohs(tcp_hdr(skb)->window);
3548 if (likely(!tcp_hdr(skb)->syn))
3549 nwin <<= tp->rx_opt.snd_wscale;
3551 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3552 flag |= FLAG_WIN_UPDATE;
3553 tcp_update_wl(tp, ack_seq);
3555 if (tp->snd_wnd != nwin) {
3558 /* Note, it is the only place, where
3559 * fast path is recovered for sending TCP.
3562 tcp_fast_path_check(sk);
3564 if (!tcp_write_queue_empty(sk))
3565 tcp_slow_start_after_idle_check(sk);
3567 if (nwin > tp->max_window) {
3568 tp->max_window = nwin;
3569 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3574 tcp_snd_una_update(tp, ack);
3579 static bool __tcp_oow_rate_limited(struct net *net, int mib_idx,
3580 u32 *last_oow_ack_time)
3582 if (*last_oow_ack_time) {
3583 s32 elapsed = (s32)(tcp_jiffies32 - *last_oow_ack_time);
3586 elapsed < READ_ONCE(net->ipv4.sysctl_tcp_invalid_ratelimit)) {
3587 NET_INC_STATS(net, mib_idx);
3588 return true; /* rate-limited: don't send yet! */
3592 *last_oow_ack_time = tcp_jiffies32;
3594 return false; /* not rate-limited: go ahead, send dupack now! */
3597 /* Return true if we're currently rate-limiting out-of-window ACKs and
3598 * thus shouldn't send a dupack right now. We rate-limit dupacks in
3599 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3600 * attacks that send repeated SYNs or ACKs for the same connection. To
3601 * do this, we do not send a duplicate SYNACK or ACK if the remote
3602 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3604 bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb,
3605 int mib_idx, u32 *last_oow_ack_time)
3607 /* Data packets without SYNs are not likely part of an ACK loop. */
3608 if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) &&
3612 return __tcp_oow_rate_limited(net, mib_idx, last_oow_ack_time);
3615 /* RFC 5961 7 [ACK Throttling] */
3616 static void tcp_send_challenge_ack(struct sock *sk)
3618 /* unprotected vars, we dont care of overwrites */
3619 static u32 challenge_timestamp;
3620 static unsigned int challenge_count;
3621 struct tcp_sock *tp = tcp_sk(sk);
3622 struct net *net = sock_net(sk);
3625 /* First check our per-socket dupack rate limit. */
3626 if (__tcp_oow_rate_limited(net,
3627 LINUX_MIB_TCPACKSKIPPEDCHALLENGE,
3628 &tp->last_oow_ack_time))
3631 /* Then check host-wide RFC 5961 rate limit. */
3633 if (now != challenge_timestamp) {
3634 u32 ack_limit = READ_ONCE(net->ipv4.sysctl_tcp_challenge_ack_limit);
3635 u32 half = (ack_limit + 1) >> 1;
3637 challenge_timestamp = now;
3638 WRITE_ONCE(challenge_count, half + prandom_u32_max(ack_limit));
3640 count = READ_ONCE(challenge_count);
3642 WRITE_ONCE(challenge_count, count - 1);
3643 NET_INC_STATS(net, LINUX_MIB_TCPCHALLENGEACK);
3648 static void tcp_store_ts_recent(struct tcp_sock *tp)
3650 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3651 tp->rx_opt.ts_recent_stamp = ktime_get_seconds();
3654 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3656 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3657 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3658 * extra check below makes sure this can only happen
3659 * for pure ACK frames. -DaveM
3661 * Not only, also it occurs for expired timestamps.
3664 if (tcp_paws_check(&tp->rx_opt, 0))
3665 tcp_store_ts_recent(tp);
3669 /* This routine deals with acks during a TLP episode and ends an episode by
3670 * resetting tlp_high_seq. Ref: TLP algorithm in draft-ietf-tcpm-rack
3672 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3674 struct tcp_sock *tp = tcp_sk(sk);
3676 if (before(ack, tp->tlp_high_seq))
3679 if (!tp->tlp_retrans) {
3680 /* TLP of new data has been acknowledged */
3681 tp->tlp_high_seq = 0;
3682 } else if (flag & FLAG_DSACK_TLP) {
3683 /* This DSACK means original and TLP probe arrived; no loss */
3684 tp->tlp_high_seq = 0;
3685 } else if (after(ack, tp->tlp_high_seq)) {
3686 /* ACK advances: there was a loss, so reduce cwnd. Reset
3687 * tlp_high_seq in tcp_init_cwnd_reduction()
3689 tcp_init_cwnd_reduction(sk);
3690 tcp_set_ca_state(sk, TCP_CA_CWR);
3691 tcp_end_cwnd_reduction(sk);
3692 tcp_try_keep_open(sk);
3693 NET_INC_STATS(sock_net(sk),
3694 LINUX_MIB_TCPLOSSPROBERECOVERY);
3695 } else if (!(flag & (FLAG_SND_UNA_ADVANCED |
3696 FLAG_NOT_DUP | FLAG_DATA_SACKED))) {
3697 /* Pure dupack: original and TLP probe arrived; no loss */
3698 tp->tlp_high_seq = 0;
3702 static inline void tcp_in_ack_event(struct sock *sk, u32 flags)
3704 const struct inet_connection_sock *icsk = inet_csk(sk);
3706 if (icsk->icsk_ca_ops->in_ack_event)
3707 icsk->icsk_ca_ops->in_ack_event(sk, flags);
3710 /* Congestion control has updated the cwnd already. So if we're in
3711 * loss recovery then now we do any new sends (for FRTO) or
3712 * retransmits (for CA_Loss or CA_recovery) that make sense.
3714 static void tcp_xmit_recovery(struct sock *sk, int rexmit)
3716 struct tcp_sock *tp = tcp_sk(sk);
3718 if (rexmit == REXMIT_NONE || sk->sk_state == TCP_SYN_SENT)
3721 if (unlikely(rexmit == REXMIT_NEW)) {
3722 __tcp_push_pending_frames(sk, tcp_current_mss(sk),
3724 if (after(tp->snd_nxt, tp->high_seq))
3728 tcp_xmit_retransmit_queue(sk);
3731 /* Returns the number of packets newly acked or sacked by the current ACK */
3732 static u32 tcp_newly_delivered(struct sock *sk, u32 prior_delivered, int flag)
3734 const struct net *net = sock_net(sk);
3735 struct tcp_sock *tp = tcp_sk(sk);
3738 delivered = tp->delivered - prior_delivered;
3739 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVERED, delivered);
3740 if (flag & FLAG_ECE)
3741 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVEREDCE, delivered);
3746 /* This routine deals with incoming acks, but not outgoing ones. */
3747 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3749 struct inet_connection_sock *icsk = inet_csk(sk);
3750 struct tcp_sock *tp = tcp_sk(sk);
3751 struct tcp_sacktag_state sack_state;
3752 struct rate_sample rs = { .prior_delivered = 0 };
3753 u32 prior_snd_una = tp->snd_una;
3754 bool is_sack_reneg = tp->is_sack_reneg;
3755 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3756 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3758 int prior_packets = tp->packets_out;
3759 u32 delivered = tp->delivered;
3760 u32 lost = tp->lost;
3761 int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */
3764 sack_state.first_sackt = 0;
3765 sack_state.rate = &rs;
3766 sack_state.sack_delivered = 0;
3768 /* We very likely will need to access rtx queue. */
3769 prefetch(sk->tcp_rtx_queue.rb_node);
3771 /* If the ack is older than previous acks
3772 * then we can probably ignore it.
3774 if (before(ack, prior_snd_una)) {
3775 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3776 if (before(ack, prior_snd_una - tp->max_window)) {
3777 if (!(flag & FLAG_NO_CHALLENGE_ACK))
3778 tcp_send_challenge_ack(sk);
3779 return -SKB_DROP_REASON_TCP_TOO_OLD_ACK;
3784 /* If the ack includes data we haven't sent yet, discard
3785 * this segment (RFC793 Section 3.9).
3787 if (after(ack, tp->snd_nxt))
3788 return -SKB_DROP_REASON_TCP_ACK_UNSENT_DATA;
3790 if (after(ack, prior_snd_una)) {
3791 flag |= FLAG_SND_UNA_ADVANCED;
3792 icsk->icsk_retransmits = 0;
3794 #if IS_ENABLED(CONFIG_TLS_DEVICE)
3795 if (static_branch_unlikely(&clean_acked_data_enabled.key))
3796 if (icsk->icsk_clean_acked)
3797 icsk->icsk_clean_acked(sk, ack);
3801 prior_fack = tcp_is_sack(tp) ? tcp_highest_sack_seq(tp) : tp->snd_una;
3802 rs.prior_in_flight = tcp_packets_in_flight(tp);
3804 /* ts_recent update must be made after we are sure that the packet
3807 if (flag & FLAG_UPDATE_TS_RECENT)
3808 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3810 if ((flag & (FLAG_SLOWPATH | FLAG_SND_UNA_ADVANCED)) ==
3811 FLAG_SND_UNA_ADVANCED) {
3812 /* Window is constant, pure forward advance.
3813 * No more checks are required.
3814 * Note, we use the fact that SND.UNA>=SND.WL2.
3816 tcp_update_wl(tp, ack_seq);
3817 tcp_snd_una_update(tp, ack);
3818 flag |= FLAG_WIN_UPDATE;
3820 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
3822 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPACKS);
3824 u32 ack_ev_flags = CA_ACK_SLOWPATH;
3826 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3829 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3831 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3833 if (TCP_SKB_CB(skb)->sacked)
3834 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3837 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) {
3839 ack_ev_flags |= CA_ACK_ECE;
3842 if (sack_state.sack_delivered)
3843 tcp_count_delivered(tp, sack_state.sack_delivered,
3846 if (flag & FLAG_WIN_UPDATE)
3847 ack_ev_flags |= CA_ACK_WIN_UPDATE;
3849 tcp_in_ack_event(sk, ack_ev_flags);
3852 /* This is a deviation from RFC3168 since it states that:
3853 * "When the TCP data sender is ready to set the CWR bit after reducing
3854 * the congestion window, it SHOULD set the CWR bit only on the first
3855 * new data packet that it transmits."
3856 * We accept CWR on pure ACKs to be more robust
3857 * with widely-deployed TCP implementations that do this.
3859 tcp_ecn_accept_cwr(sk, skb);
3861 /* We passed data and got it acked, remove any soft error
3862 * log. Something worked...
3864 sk->sk_err_soft = 0;
3865 icsk->icsk_probes_out = 0;
3866 tp->rcv_tstamp = tcp_jiffies32;
3870 /* See if we can take anything off of the retransmit queue. */
3871 flag |= tcp_clean_rtx_queue(sk, skb, prior_fack, prior_snd_una,
3872 &sack_state, flag & FLAG_ECE);
3874 tcp_rack_update_reo_wnd(sk, &rs);
3876 if (tp->tlp_high_seq)
3877 tcp_process_tlp_ack(sk, ack, flag);
3879 if (tcp_ack_is_dubious(sk, flag)) {
3880 if (!(flag & (FLAG_SND_UNA_ADVANCED |
3881 FLAG_NOT_DUP | FLAG_DSACKING_ACK))) {
3883 /* Consider if pure acks were aggregated in tcp_add_backlog() */
3884 if (!(flag & FLAG_DATA))
3885 num_dupack = max_t(u16, 1, skb_shinfo(skb)->gso_segs);
3887 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
3891 /* If needed, reset TLP/RTO timer when RACK doesn't set. */
3892 if (flag & FLAG_SET_XMIT_TIMER)
3893 tcp_set_xmit_timer(sk);
3895 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
3898 delivered = tcp_newly_delivered(sk, delivered, flag);
3899 lost = tp->lost - lost; /* freshly marked lost */
3900 rs.is_ack_delayed = !!(flag & FLAG_ACK_MAYBE_DELAYED);
3901 tcp_rate_gen(sk, delivered, lost, is_sack_reneg, sack_state.rate);
3902 tcp_cong_control(sk, ack, delivered, flag, sack_state.rate);
3903 tcp_xmit_recovery(sk, rexmit);
3907 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3908 if (flag & FLAG_DSACKING_ACK) {
3909 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
3911 tcp_newly_delivered(sk, delivered, flag);
3913 /* If this ack opens up a zero window, clear backoff. It was
3914 * being used to time the probes, and is probably far higher than
3915 * it needs to be for normal retransmission.
3919 if (tp->tlp_high_seq)
3920 tcp_process_tlp_ack(sk, ack, flag);
3924 /* If data was SACKed, tag it and see if we should send more data.
3925 * If data was DSACKed, see if we can undo a cwnd reduction.
3927 if (TCP_SKB_CB(skb)->sacked) {
3928 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3930 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
3932 tcp_newly_delivered(sk, delivered, flag);
3933 tcp_xmit_recovery(sk, rexmit);
3939 static void tcp_parse_fastopen_option(int len, const unsigned char *cookie,
3940 bool syn, struct tcp_fastopen_cookie *foc,
3943 /* Valid only in SYN or SYN-ACK with an even length. */
3944 if (!foc || !syn || len < 0 || (len & 1))
3947 if (len >= TCP_FASTOPEN_COOKIE_MIN &&
3948 len <= TCP_FASTOPEN_COOKIE_MAX)
3949 memcpy(foc->val, cookie, len);
3956 static bool smc_parse_options(const struct tcphdr *th,
3957 struct tcp_options_received *opt_rx,
3958 const unsigned char *ptr,
3961 #if IS_ENABLED(CONFIG_SMC)
3962 if (static_branch_unlikely(&tcp_have_smc)) {
3963 if (th->syn && !(opsize & 1) &&
3964 opsize >= TCPOLEN_EXP_SMC_BASE &&
3965 get_unaligned_be32(ptr) == TCPOPT_SMC_MAGIC) {
3974 /* Try to parse the MSS option from the TCP header. Return 0 on failure, clamped
3977 static u16 tcp_parse_mss_option(const struct tcphdr *th, u16 user_mss)
3979 const unsigned char *ptr = (const unsigned char *)(th + 1);
3980 int length = (th->doff * 4) - sizeof(struct tcphdr);
3983 while (length > 0) {
3984 int opcode = *ptr++;
3990 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3997 if (opsize < 2) /* "silly options" */
3999 if (opsize > length)
4000 return mss; /* fail on partial options */
4001 if (opcode == TCPOPT_MSS && opsize == TCPOLEN_MSS) {
4002 u16 in_mss = get_unaligned_be16(ptr);
4005 if (user_mss && user_mss < in_mss)
4017 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
4018 * But, this can also be called on packets in the established flow when
4019 * the fast version below fails.
4021 void tcp_parse_options(const struct net *net,
4022 const struct sk_buff *skb,
4023 struct tcp_options_received *opt_rx, int estab,
4024 struct tcp_fastopen_cookie *foc)
4026 const unsigned char *ptr;
4027 const struct tcphdr *th = tcp_hdr(skb);
4028 int length = (th->doff * 4) - sizeof(struct tcphdr);
4030 ptr = (const unsigned char *)(th + 1);
4031 opt_rx->saw_tstamp = 0;
4032 opt_rx->saw_unknown = 0;
4034 while (length > 0) {
4035 int opcode = *ptr++;
4041 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
4048 if (opsize < 2) /* "silly options" */
4050 if (opsize > length)
4051 return; /* don't parse partial options */
4054 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
4055 u16 in_mss = get_unaligned_be16(ptr);
4057 if (opt_rx->user_mss &&
4058 opt_rx->user_mss < in_mss)
4059 in_mss = opt_rx->user_mss;
4060 opt_rx->mss_clamp = in_mss;
4065 if (opsize == TCPOLEN_WINDOW && th->syn &&
4066 !estab && READ_ONCE(net->ipv4.sysctl_tcp_window_scaling)) {
4067 __u8 snd_wscale = *(__u8 *)ptr;
4068 opt_rx->wscale_ok = 1;
4069 if (snd_wscale > TCP_MAX_WSCALE) {
4070 net_info_ratelimited("%s: Illegal window scaling value %d > %u received\n",
4074 snd_wscale = TCP_MAX_WSCALE;
4076 opt_rx->snd_wscale = snd_wscale;
4079 case TCPOPT_TIMESTAMP:
4080 if ((opsize == TCPOLEN_TIMESTAMP) &&
4081 ((estab && opt_rx->tstamp_ok) ||
4082 (!estab && READ_ONCE(net->ipv4.sysctl_tcp_timestamps)))) {
4083 opt_rx->saw_tstamp = 1;
4084 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
4085 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
4088 case TCPOPT_SACK_PERM:
4089 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
4090 !estab && READ_ONCE(net->ipv4.sysctl_tcp_sack)) {
4091 opt_rx->sack_ok = TCP_SACK_SEEN;
4092 tcp_sack_reset(opt_rx);
4097 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
4098 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
4100 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
4103 #ifdef CONFIG_TCP_MD5SIG
4106 * The MD5 Hash has already been
4107 * checked (see tcp_v{4,6}_do_rcv()).
4111 case TCPOPT_FASTOPEN:
4112 tcp_parse_fastopen_option(
4113 opsize - TCPOLEN_FASTOPEN_BASE,
4114 ptr, th->syn, foc, false);
4118 /* Fast Open option shares code 254 using a
4119 * 16 bits magic number.
4121 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE &&
4122 get_unaligned_be16(ptr) ==
4123 TCPOPT_FASTOPEN_MAGIC) {
4124 tcp_parse_fastopen_option(opsize -
4125 TCPOLEN_EXP_FASTOPEN_BASE,
4126 ptr + 2, th->syn, foc, true);
4130 if (smc_parse_options(th, opt_rx, ptr, opsize))
4133 opt_rx->saw_unknown = 1;
4137 opt_rx->saw_unknown = 1;
4144 EXPORT_SYMBOL(tcp_parse_options);
4146 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
4148 const __be32 *ptr = (const __be32 *)(th + 1);
4150 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
4151 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
4152 tp->rx_opt.saw_tstamp = 1;
4154 tp->rx_opt.rcv_tsval = ntohl(*ptr);
4157 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
4159 tp->rx_opt.rcv_tsecr = 0;
4165 /* Fast parse options. This hopes to only see timestamps.
4166 * If it is wrong it falls back on tcp_parse_options().
4168 static bool tcp_fast_parse_options(const struct net *net,
4169 const struct sk_buff *skb,
4170 const struct tcphdr *th, struct tcp_sock *tp)
4172 /* In the spirit of fast parsing, compare doff directly to constant
4173 * values. Because equality is used, short doff can be ignored here.
4175 if (th->doff == (sizeof(*th) / 4)) {
4176 tp->rx_opt.saw_tstamp = 0;
4178 } else if (tp->rx_opt.tstamp_ok &&
4179 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
4180 if (tcp_parse_aligned_timestamp(tp, th))
4184 tcp_parse_options(net, skb, &tp->rx_opt, 1, NULL);
4185 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
4186 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
4191 #ifdef CONFIG_TCP_MD5SIG
4193 * Parse MD5 Signature option
4195 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
4197 int length = (th->doff << 2) - sizeof(*th);
4198 const u8 *ptr = (const u8 *)(th + 1);
4200 /* If not enough data remaining, we can short cut */
4201 while (length >= TCPOLEN_MD5SIG) {
4202 int opcode = *ptr++;
4213 if (opsize < 2 || opsize > length)
4215 if (opcode == TCPOPT_MD5SIG)
4216 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
4223 EXPORT_SYMBOL(tcp_parse_md5sig_option);
4226 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
4228 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
4229 * it can pass through stack. So, the following predicate verifies that
4230 * this segment is not used for anything but congestion avoidance or
4231 * fast retransmit. Moreover, we even are able to eliminate most of such
4232 * second order effects, if we apply some small "replay" window (~RTO)
4233 * to timestamp space.
4235 * All these measures still do not guarantee that we reject wrapped ACKs
4236 * on networks with high bandwidth, when sequence space is recycled fastly,
4237 * but it guarantees that such events will be very rare and do not affect
4238 * connection seriously. This doesn't look nice, but alas, PAWS is really
4241 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
4242 * states that events when retransmit arrives after original data are rare.
4243 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
4244 * the biggest problem on large power networks even with minor reordering.
4245 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
4246 * up to bandwidth of 18Gigabit/sec. 8) ]
4249 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
4251 const struct tcp_sock *tp = tcp_sk(sk);
4252 const struct tcphdr *th = tcp_hdr(skb);
4253 u32 seq = TCP_SKB_CB(skb)->seq;
4254 u32 ack = TCP_SKB_CB(skb)->ack_seq;
4256 return (/* 1. Pure ACK with correct sequence number. */
4257 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
4259 /* 2. ... and duplicate ACK. */
4260 ack == tp->snd_una &&
4262 /* 3. ... and does not update window. */
4263 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
4265 /* 4. ... and sits in replay window. */
4266 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
4269 static inline bool tcp_paws_discard(const struct sock *sk,
4270 const struct sk_buff *skb)
4272 const struct tcp_sock *tp = tcp_sk(sk);
4274 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
4275 !tcp_disordered_ack(sk, skb);
4278 /* Check segment sequence number for validity.
4280 * Segment controls are considered valid, if the segment
4281 * fits to the window after truncation to the window. Acceptability
4282 * of data (and SYN, FIN, of course) is checked separately.
4283 * See tcp_data_queue(), for example.
4285 * Also, controls (RST is main one) are accepted using RCV.WUP instead
4286 * of RCV.NXT. Peer still did not advance his SND.UNA when we
4287 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4288 * (borrowed from freebsd)
4291 static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
4293 return !before(end_seq, tp->rcv_wup) &&
4294 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
4297 /* When we get a reset we do this. */
4298 void tcp_reset(struct sock *sk, struct sk_buff *skb)
4300 trace_tcp_receive_reset(sk);
4302 /* mptcp can't tell us to ignore reset pkts,
4303 * so just ignore the return value of mptcp_incoming_options().
4305 if (sk_is_mptcp(sk))
4306 mptcp_incoming_options(sk, skb);
4308 /* We want the right error as BSD sees it (and indeed as we do). */
4309 switch (sk->sk_state) {
4311 sk->sk_err = ECONNREFUSED;
4313 case TCP_CLOSE_WAIT:
4319 sk->sk_err = ECONNRESET;
4321 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4324 tcp_write_queue_purge(sk);
4327 if (!sock_flag(sk, SOCK_DEAD))
4328 sk_error_report(sk);
4332 * Process the FIN bit. This now behaves as it is supposed to work
4333 * and the FIN takes effect when it is validly part of sequence
4334 * space. Not before when we get holes.
4336 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4337 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4340 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4341 * close and we go into CLOSING (and later onto TIME-WAIT)
4343 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4345 void tcp_fin(struct sock *sk)
4347 struct tcp_sock *tp = tcp_sk(sk);
4349 inet_csk_schedule_ack(sk);
4351 sk->sk_shutdown |= RCV_SHUTDOWN;
4352 sock_set_flag(sk, SOCK_DONE);
4354 switch (sk->sk_state) {
4356 case TCP_ESTABLISHED:
4357 /* Move to CLOSE_WAIT */
4358 tcp_set_state(sk, TCP_CLOSE_WAIT);
4359 inet_csk_enter_pingpong_mode(sk);
4362 case TCP_CLOSE_WAIT:
4364 /* Received a retransmission of the FIN, do
4369 /* RFC793: Remain in the LAST-ACK state. */
4373 /* This case occurs when a simultaneous close
4374 * happens, we must ack the received FIN and
4375 * enter the CLOSING state.
4378 tcp_set_state(sk, TCP_CLOSING);
4381 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4383 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4386 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4387 * cases we should never reach this piece of code.
4389 pr_err("%s: Impossible, sk->sk_state=%d\n",
4390 __func__, sk->sk_state);
4394 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4395 * Probably, we should reset in this case. For now drop them.
4397 skb_rbtree_purge(&tp->out_of_order_queue);
4398 if (tcp_is_sack(tp))
4399 tcp_sack_reset(&tp->rx_opt);
4402 if (!sock_flag(sk, SOCK_DEAD)) {
4403 sk->sk_state_change(sk);
4405 /* Do not send POLL_HUP for half duplex close. */
4406 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4407 sk->sk_state == TCP_CLOSE)
4408 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4410 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4414 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4417 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4418 if (before(seq, sp->start_seq))
4419 sp->start_seq = seq;
4420 if (after(end_seq, sp->end_seq))
4421 sp->end_seq = end_seq;
4427 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4429 struct tcp_sock *tp = tcp_sk(sk);
4431 if (tcp_is_sack(tp) && READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_dsack)) {
4434 if (before(seq, tp->rcv_nxt))
4435 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4437 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4439 NET_INC_STATS(sock_net(sk), mib_idx);
4441 tp->rx_opt.dsack = 1;
4442 tp->duplicate_sack[0].start_seq = seq;
4443 tp->duplicate_sack[0].end_seq = end_seq;
4447 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4449 struct tcp_sock *tp = tcp_sk(sk);
4451 if (!tp->rx_opt.dsack)
4452 tcp_dsack_set(sk, seq, end_seq);
4454 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4457 static void tcp_rcv_spurious_retrans(struct sock *sk, const struct sk_buff *skb)
4459 /* When the ACK path fails or drops most ACKs, the sender would
4460 * timeout and spuriously retransmit the same segment repeatedly.
4461 * The receiver remembers and reflects via DSACKs. Leverage the
4462 * DSACK state and change the txhash to re-route speculatively.
4464 if (TCP_SKB_CB(skb)->seq == tcp_sk(sk)->duplicate_sack[0].start_seq &&
4465 sk_rethink_txhash(sk))
4466 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDUPLICATEDATAREHASH);
4469 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4471 struct tcp_sock *tp = tcp_sk(sk);
4473 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4474 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4475 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4476 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
4478 if (tcp_is_sack(tp) && READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_dsack)) {
4479 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4481 tcp_rcv_spurious_retrans(sk, skb);
4482 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4483 end_seq = tp->rcv_nxt;
4484 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4491 /* These routines update the SACK block as out-of-order packets arrive or
4492 * in-order packets close up the sequence space.
4494 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4497 struct tcp_sack_block *sp = &tp->selective_acks[0];
4498 struct tcp_sack_block *swalk = sp + 1;
4500 /* See if the recent change to the first SACK eats into
4501 * or hits the sequence space of other SACK blocks, if so coalesce.
4503 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4504 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4507 /* Zap SWALK, by moving every further SACK up by one slot.
4508 * Decrease num_sacks.
4510 tp->rx_opt.num_sacks--;
4511 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4520 static void tcp_sack_compress_send_ack(struct sock *sk)
4522 struct tcp_sock *tp = tcp_sk(sk);
4524 if (!tp->compressed_ack)
4527 if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1)
4530 /* Since we have to send one ack finally,
4531 * substract one from tp->compressed_ack to keep
4532 * LINUX_MIB_TCPACKCOMPRESSED accurate.
4534 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED,
4535 tp->compressed_ack - 1);
4537 tp->compressed_ack = 0;
4541 /* Reasonable amount of sack blocks included in TCP SACK option
4542 * The max is 4, but this becomes 3 if TCP timestamps are there.
4543 * Given that SACK packets might be lost, be conservative and use 2.
4545 #define TCP_SACK_BLOCKS_EXPECTED 2
4547 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4549 struct tcp_sock *tp = tcp_sk(sk);
4550 struct tcp_sack_block *sp = &tp->selective_acks[0];
4551 int cur_sacks = tp->rx_opt.num_sacks;
4557 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4558 if (tcp_sack_extend(sp, seq, end_seq)) {
4559 if (this_sack >= TCP_SACK_BLOCKS_EXPECTED)
4560 tcp_sack_compress_send_ack(sk);
4561 /* Rotate this_sack to the first one. */
4562 for (; this_sack > 0; this_sack--, sp--)
4563 swap(*sp, *(sp - 1));
4565 tcp_sack_maybe_coalesce(tp);
4570 if (this_sack >= TCP_SACK_BLOCKS_EXPECTED)
4571 tcp_sack_compress_send_ack(sk);
4573 /* Could not find an adjacent existing SACK, build a new one,
4574 * put it at the front, and shift everyone else down. We
4575 * always know there is at least one SACK present already here.
4577 * If the sack array is full, forget about the last one.
4579 if (this_sack >= TCP_NUM_SACKS) {
4581 tp->rx_opt.num_sacks--;
4584 for (; this_sack > 0; this_sack--, sp--)
4588 /* Build the new head SACK, and we're done. */
4589 sp->start_seq = seq;
4590 sp->end_seq = end_seq;
4591 tp->rx_opt.num_sacks++;
4594 /* RCV.NXT advances, some SACKs should be eaten. */
4596 static void tcp_sack_remove(struct tcp_sock *tp)
4598 struct tcp_sack_block *sp = &tp->selective_acks[0];
4599 int num_sacks = tp->rx_opt.num_sacks;
4602 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4603 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4604 tp->rx_opt.num_sacks = 0;
4608 for (this_sack = 0; this_sack < num_sacks;) {
4609 /* Check if the start of the sack is covered by RCV.NXT. */
4610 if (!before(tp->rcv_nxt, sp->start_seq)) {
4613 /* RCV.NXT must cover all the block! */
4614 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4616 /* Zap this SACK, by moving forward any other SACKS. */
4617 for (i = this_sack+1; i < num_sacks; i++)
4618 tp->selective_acks[i-1] = tp->selective_acks[i];
4625 tp->rx_opt.num_sacks = num_sacks;
4629 * tcp_try_coalesce - try to merge skb to prior one
4632 * @from: buffer to add in queue
4633 * @fragstolen: pointer to boolean
4635 * Before queueing skb @from after @to, try to merge them
4636 * to reduce overall memory use and queue lengths, if cost is small.
4637 * Packets in ofo or receive queues can stay a long time.
4638 * Better try to coalesce them right now to avoid future collapses.
4639 * Returns true if caller should free @from instead of queueing it
4641 static bool tcp_try_coalesce(struct sock *sk,
4643 struct sk_buff *from,
4648 *fragstolen = false;
4650 /* Its possible this segment overlaps with prior segment in queue */
4651 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4654 if (!mptcp_skb_can_collapse(to, from))
4657 #ifdef CONFIG_TLS_DEVICE
4658 if (from->decrypted != to->decrypted)
4662 if (!skb_try_coalesce(to, from, fragstolen, &delta))
4665 atomic_add(delta, &sk->sk_rmem_alloc);
4666 sk_mem_charge(sk, delta);
4667 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4668 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4669 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4670 TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
4672 if (TCP_SKB_CB(from)->has_rxtstamp) {
4673 TCP_SKB_CB(to)->has_rxtstamp = true;
4674 to->tstamp = from->tstamp;
4675 skb_hwtstamps(to)->hwtstamp = skb_hwtstamps(from)->hwtstamp;
4681 static bool tcp_ooo_try_coalesce(struct sock *sk,
4683 struct sk_buff *from,
4686 bool res = tcp_try_coalesce(sk, to, from, fragstolen);
4688 /* In case tcp_drop_reason() is called later, update to->gso_segs */
4690 u32 gso_segs = max_t(u16, 1, skb_shinfo(to)->gso_segs) +
4691 max_t(u16, 1, skb_shinfo(from)->gso_segs);
4693 skb_shinfo(to)->gso_segs = min_t(u32, gso_segs, 0xFFFF);
4698 static void tcp_drop_reason(struct sock *sk, struct sk_buff *skb,
4699 enum skb_drop_reason reason)
4701 sk_drops_add(sk, skb);
4702 kfree_skb_reason(skb, reason);
4705 /* This one checks to see if we can put data from the
4706 * out_of_order queue into the receive_queue.
4708 static void tcp_ofo_queue(struct sock *sk)
4710 struct tcp_sock *tp = tcp_sk(sk);
4711 __u32 dsack_high = tp->rcv_nxt;
4712 bool fin, fragstolen, eaten;
4713 struct sk_buff *skb, *tail;
4716 p = rb_first(&tp->out_of_order_queue);
4719 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4722 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4723 __u32 dsack = dsack_high;
4724 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4725 dsack_high = TCP_SKB_CB(skb)->end_seq;
4726 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4729 rb_erase(&skb->rbnode, &tp->out_of_order_queue);
4731 if (unlikely(!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))) {
4732 tcp_drop_reason(sk, skb, SKB_DROP_REASON_TCP_OFO_DROP);
4736 tail = skb_peek_tail(&sk->sk_receive_queue);
4737 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen);
4738 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4739 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
4741 __skb_queue_tail(&sk->sk_receive_queue, skb);
4743 kfree_skb_partial(skb, fragstolen);
4745 if (unlikely(fin)) {
4747 /* tcp_fin() purges tp->out_of_order_queue,
4748 * so we must end this loop right now.
4755 static bool tcp_prune_ofo_queue(struct sock *sk);
4756 static int tcp_prune_queue(struct sock *sk);
4758 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4761 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4762 !sk_rmem_schedule(sk, skb, size)) {
4764 if (tcp_prune_queue(sk) < 0)
4767 while (!sk_rmem_schedule(sk, skb, size)) {
4768 if (!tcp_prune_ofo_queue(sk))
4775 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4777 struct tcp_sock *tp = tcp_sk(sk);
4778 struct rb_node **p, *parent;
4779 struct sk_buff *skb1;
4783 tcp_ecn_check_ce(sk, skb);
4785 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4786 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFODROP);
4787 sk->sk_data_ready(sk);
4788 tcp_drop_reason(sk, skb, SKB_DROP_REASON_PROTO_MEM);
4792 /* Disable header prediction. */
4794 inet_csk_schedule_ack(sk);
4796 tp->rcv_ooopack += max_t(u16, 1, skb_shinfo(skb)->gso_segs);
4797 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4798 seq = TCP_SKB_CB(skb)->seq;
4799 end_seq = TCP_SKB_CB(skb)->end_seq;
4801 p = &tp->out_of_order_queue.rb_node;
4802 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4803 /* Initial out of order segment, build 1 SACK. */
4804 if (tcp_is_sack(tp)) {
4805 tp->rx_opt.num_sacks = 1;
4806 tp->selective_acks[0].start_seq = seq;
4807 tp->selective_acks[0].end_seq = end_seq;
4809 rb_link_node(&skb->rbnode, NULL, p);
4810 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4811 tp->ooo_last_skb = skb;
4815 /* In the typical case, we are adding an skb to the end of the list.
4816 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
4818 if (tcp_ooo_try_coalesce(sk, tp->ooo_last_skb,
4819 skb, &fragstolen)) {
4821 /* For non sack flows, do not grow window to force DUPACK
4822 * and trigger fast retransmit.
4824 if (tcp_is_sack(tp))
4825 tcp_grow_window(sk, skb, true);
4826 kfree_skb_partial(skb, fragstolen);
4830 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
4831 if (!before(seq, TCP_SKB_CB(tp->ooo_last_skb)->end_seq)) {
4832 parent = &tp->ooo_last_skb->rbnode;
4833 p = &parent->rb_right;
4837 /* Find place to insert this segment. Handle overlaps on the way. */
4841 skb1 = rb_to_skb(parent);
4842 if (before(seq, TCP_SKB_CB(skb1)->seq)) {
4843 p = &parent->rb_left;
4846 if (before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4847 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4848 /* All the bits are present. Drop. */
4849 NET_INC_STATS(sock_net(sk),
4850 LINUX_MIB_TCPOFOMERGE);
4851 tcp_drop_reason(sk, skb,
4852 SKB_DROP_REASON_TCP_OFOMERGE);
4854 tcp_dsack_set(sk, seq, end_seq);
4857 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4858 /* Partial overlap. */
4859 tcp_dsack_set(sk, seq, TCP_SKB_CB(skb1)->end_seq);
4861 /* skb's seq == skb1's seq and skb covers skb1.
4862 * Replace skb1 with skb.
4864 rb_replace_node(&skb1->rbnode, &skb->rbnode,
4865 &tp->out_of_order_queue);
4866 tcp_dsack_extend(sk,
4867 TCP_SKB_CB(skb1)->seq,
4868 TCP_SKB_CB(skb1)->end_seq);
4869 NET_INC_STATS(sock_net(sk),
4870 LINUX_MIB_TCPOFOMERGE);
4871 tcp_drop_reason(sk, skb1,
4872 SKB_DROP_REASON_TCP_OFOMERGE);
4875 } else if (tcp_ooo_try_coalesce(sk, skb1,
4876 skb, &fragstolen)) {
4879 p = &parent->rb_right;
4882 /* Insert segment into RB tree. */
4883 rb_link_node(&skb->rbnode, parent, p);
4884 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4887 /* Remove other segments covered by skb. */
4888 while ((skb1 = skb_rb_next(skb)) != NULL) {
4889 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4891 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4892 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4896 rb_erase(&skb1->rbnode, &tp->out_of_order_queue);
4897 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4898 TCP_SKB_CB(skb1)->end_seq);
4899 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4900 tcp_drop_reason(sk, skb1, SKB_DROP_REASON_TCP_OFOMERGE);
4902 /* If there is no skb after us, we are the last_skb ! */
4904 tp->ooo_last_skb = skb;
4907 if (tcp_is_sack(tp))
4908 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4911 /* For non sack flows, do not grow window to force DUPACK
4912 * and trigger fast retransmit.
4914 if (tcp_is_sack(tp))
4915 tcp_grow_window(sk, skb, false);
4917 skb_set_owner_r(skb, sk);
4921 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb,
4925 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4928 tcp_try_coalesce(sk, tail,
4929 skb, fragstolen)) ? 1 : 0;
4930 tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq);
4932 __skb_queue_tail(&sk->sk_receive_queue, skb);
4933 skb_set_owner_r(skb, sk);
4938 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4940 struct sk_buff *skb;
4948 if (size > PAGE_SIZE) {
4949 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS);
4951 data_len = npages << PAGE_SHIFT;
4952 size = data_len + (size & ~PAGE_MASK);
4954 skb = alloc_skb_with_frags(size - data_len, data_len,
4955 PAGE_ALLOC_COSTLY_ORDER,
4956 &err, sk->sk_allocation);
4960 skb_put(skb, size - data_len);
4961 skb->data_len = data_len;
4964 if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) {
4965 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP);
4969 err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size);
4973 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4974 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4975 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4977 if (tcp_queue_rcv(sk, skb, &fragstolen)) {
4978 WARN_ON_ONCE(fragstolen); /* should not happen */
4990 void tcp_data_ready(struct sock *sk)
4992 if (tcp_epollin_ready(sk, sk->sk_rcvlowat) || sock_flag(sk, SOCK_DONE))
4993 sk->sk_data_ready(sk);
4996 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4998 struct tcp_sock *tp = tcp_sk(sk);
4999 enum skb_drop_reason reason;
5003 /* If a subflow has been reset, the packet should not continue
5004 * to be processed, drop the packet.
5006 if (sk_is_mptcp(sk) && !mptcp_incoming_options(sk, skb)) {
5011 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
5016 __skb_pull(skb, tcp_hdr(skb)->doff * 4);
5018 reason = SKB_DROP_REASON_NOT_SPECIFIED;
5019 tp->rx_opt.dsack = 0;
5021 /* Queue data for delivery to the user.
5022 * Packets in sequence go to the receive queue.
5023 * Out of sequence packets to the out_of_order_queue.
5025 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
5026 if (tcp_receive_window(tp) == 0) {
5027 reason = SKB_DROP_REASON_TCP_ZEROWINDOW;
5028 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP);
5032 /* Ok. In sequence. In window. */
5034 if (skb_queue_len(&sk->sk_receive_queue) == 0)
5035 sk_forced_mem_schedule(sk, skb->truesize);
5036 else if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) {
5037 reason = SKB_DROP_REASON_PROTO_MEM;
5038 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP);
5039 sk->sk_data_ready(sk);
5043 eaten = tcp_queue_rcv(sk, skb, &fragstolen);
5045 tcp_event_data_recv(sk, skb);
5046 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
5049 if (!RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
5052 /* RFC5681. 4.2. SHOULD send immediate ACK, when
5053 * gap in queue is filled.
5055 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
5056 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW;
5059 if (tp->rx_opt.num_sacks)
5060 tcp_sack_remove(tp);
5062 tcp_fast_path_check(sk);
5065 kfree_skb_partial(skb, fragstolen);
5066 if (!sock_flag(sk, SOCK_DEAD))
5071 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
5072 tcp_rcv_spurious_retrans(sk, skb);
5073 /* A retransmit, 2nd most common case. Force an immediate ack. */
5074 reason = SKB_DROP_REASON_TCP_OLD_DATA;
5075 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
5076 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
5079 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
5080 inet_csk_schedule_ack(sk);
5082 tcp_drop_reason(sk, skb, reason);
5086 /* Out of window. F.e. zero window probe. */
5087 if (!before(TCP_SKB_CB(skb)->seq,
5088 tp->rcv_nxt + tcp_receive_window(tp))) {
5089 reason = SKB_DROP_REASON_TCP_OVERWINDOW;
5093 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
5094 /* Partial packet, seq < rcv_next < end_seq */
5095 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
5097 /* If window is closed, drop tail of packet. But after
5098 * remembering D-SACK for its head made in previous line.
5100 if (!tcp_receive_window(tp)) {
5101 reason = SKB_DROP_REASON_TCP_ZEROWINDOW;
5102 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP);
5108 tcp_data_queue_ofo(sk, skb);
5111 static struct sk_buff *tcp_skb_next(struct sk_buff *skb, struct sk_buff_head *list)
5114 return !skb_queue_is_last(list, skb) ? skb->next : NULL;
5116 return skb_rb_next(skb);
5119 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
5120 struct sk_buff_head *list,
5121 struct rb_root *root)
5123 struct sk_buff *next = tcp_skb_next(skb, list);
5126 __skb_unlink(skb, list);
5128 rb_erase(&skb->rbnode, root);
5131 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
5136 /* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */
5137 void tcp_rbtree_insert(struct rb_root *root, struct sk_buff *skb)
5139 struct rb_node **p = &root->rb_node;
5140 struct rb_node *parent = NULL;
5141 struct sk_buff *skb1;
5145 skb1 = rb_to_skb(parent);
5146 if (before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb1)->seq))
5147 p = &parent->rb_left;
5149 p = &parent->rb_right;
5151 rb_link_node(&skb->rbnode, parent, p);
5152 rb_insert_color(&skb->rbnode, root);
5155 /* Collapse contiguous sequence of skbs head..tail with
5156 * sequence numbers start..end.
5158 * If tail is NULL, this means until the end of the queue.
5160 * Segments with FIN/SYN are not collapsed (only because this
5164 tcp_collapse(struct sock *sk, struct sk_buff_head *list, struct rb_root *root,
5165 struct sk_buff *head, struct sk_buff *tail, u32 start, u32 end)
5167 struct sk_buff *skb = head, *n;
5168 struct sk_buff_head tmp;
5171 /* First, check that queue is collapsible and find
5172 * the point where collapsing can be useful.
5175 for (end_of_skbs = true; skb != NULL && skb != tail; skb = n) {
5176 n = tcp_skb_next(skb, list);
5178 /* No new bits? It is possible on ofo queue. */
5179 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
5180 skb = tcp_collapse_one(sk, skb, list, root);
5186 /* The first skb to collapse is:
5188 * - bloated or contains data before "start" or
5189 * overlaps to the next one and mptcp allow collapsing.
5191 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
5192 (tcp_win_from_space(sk, skb->truesize) > skb->len ||
5193 before(TCP_SKB_CB(skb)->seq, start))) {
5194 end_of_skbs = false;
5198 if (n && n != tail && mptcp_skb_can_collapse(skb, n) &&
5199 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(n)->seq) {
5200 end_of_skbs = false;
5204 /* Decided to skip this, advance start seq. */
5205 start = TCP_SKB_CB(skb)->end_seq;
5208 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
5211 __skb_queue_head_init(&tmp);
5213 while (before(start, end)) {
5214 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
5215 struct sk_buff *nskb;
5217 nskb = alloc_skb(copy, GFP_ATOMIC);
5221 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
5222 #ifdef CONFIG_TLS_DEVICE
5223 nskb->decrypted = skb->decrypted;
5225 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
5227 __skb_queue_before(list, skb, nskb);
5229 __skb_queue_tail(&tmp, nskb); /* defer rbtree insertion */
5230 skb_set_owner_r(nskb, sk);
5231 mptcp_skb_ext_move(nskb, skb);
5233 /* Copy data, releasing collapsed skbs. */
5235 int offset = start - TCP_SKB_CB(skb)->seq;
5236 int size = TCP_SKB_CB(skb)->end_seq - start;
5240 size = min(copy, size);
5241 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
5243 TCP_SKB_CB(nskb)->end_seq += size;
5247 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
5248 skb = tcp_collapse_one(sk, skb, list, root);
5251 !mptcp_skb_can_collapse(nskb, skb) ||
5252 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
5254 #ifdef CONFIG_TLS_DEVICE
5255 if (skb->decrypted != nskb->decrypted)
5262 skb_queue_walk_safe(&tmp, skb, n)
5263 tcp_rbtree_insert(root, skb);
5266 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
5267 * and tcp_collapse() them until all the queue is collapsed.
5269 static void tcp_collapse_ofo_queue(struct sock *sk)
5271 struct tcp_sock *tp = tcp_sk(sk);
5272 u32 range_truesize, sum_tiny = 0;
5273 struct sk_buff *skb, *head;
5276 skb = skb_rb_first(&tp->out_of_order_queue);
5279 tp->ooo_last_skb = skb_rb_last(&tp->out_of_order_queue);
5282 start = TCP_SKB_CB(skb)->seq;
5283 end = TCP_SKB_CB(skb)->end_seq;
5284 range_truesize = skb->truesize;
5286 for (head = skb;;) {
5287 skb = skb_rb_next(skb);
5289 /* Range is terminated when we see a gap or when
5290 * we are at the queue end.
5293 after(TCP_SKB_CB(skb)->seq, end) ||
5294 before(TCP_SKB_CB(skb)->end_seq, start)) {
5295 /* Do not attempt collapsing tiny skbs */
5296 if (range_truesize != head->truesize ||
5297 end - start >= SKB_WITH_OVERHEAD(SK_MEM_QUANTUM)) {
5298 tcp_collapse(sk, NULL, &tp->out_of_order_queue,
5299 head, skb, start, end);
5301 sum_tiny += range_truesize;
5302 if (sum_tiny > sk->sk_rcvbuf >> 3)
5308 range_truesize += skb->truesize;
5309 if (unlikely(before(TCP_SKB_CB(skb)->seq, start)))
5310 start = TCP_SKB_CB(skb)->seq;
5311 if (after(TCP_SKB_CB(skb)->end_seq, end))
5312 end = TCP_SKB_CB(skb)->end_seq;
5317 * Clean the out-of-order queue to make room.
5318 * We drop high sequences packets to :
5319 * 1) Let a chance for holes to be filled.
5320 * 2) not add too big latencies if thousands of packets sit there.
5321 * (But if application shrinks SO_RCVBUF, we could still end up
5322 * freeing whole queue here)
5323 * 3) Drop at least 12.5 % of sk_rcvbuf to avoid malicious attacks.
5325 * Return true if queue has shrunk.
5327 static bool tcp_prune_ofo_queue(struct sock *sk)
5329 struct tcp_sock *tp = tcp_sk(sk);
5330 struct rb_node *node, *prev;
5333 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
5336 NET_INC_STATS(sock_net(sk), LINUX_MIB_OFOPRUNED);
5337 goal = sk->sk_rcvbuf >> 3;
5338 node = &tp->ooo_last_skb->rbnode;
5340 prev = rb_prev(node);
5341 rb_erase(node, &tp->out_of_order_queue);
5342 goal -= rb_to_skb(node)->truesize;
5343 tcp_drop_reason(sk, rb_to_skb(node),
5344 SKB_DROP_REASON_TCP_OFO_QUEUE_PRUNE);
5345 if (!prev || goal <= 0) {
5347 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf &&
5348 !tcp_under_memory_pressure(sk))
5350 goal = sk->sk_rcvbuf >> 3;
5354 tp->ooo_last_skb = rb_to_skb(prev);
5356 /* Reset SACK state. A conforming SACK implementation will
5357 * do the same at a timeout based retransmit. When a connection
5358 * is in a sad state like this, we care only about integrity
5359 * of the connection not performance.
5361 if (tp->rx_opt.sack_ok)
5362 tcp_sack_reset(&tp->rx_opt);
5366 /* Reduce allocated memory if we can, trying to get
5367 * the socket within its memory limits again.
5369 * Return less than zero if we should start dropping frames
5370 * until the socket owning process reads some of the data
5371 * to stabilize the situation.
5373 static int tcp_prune_queue(struct sock *sk)
5375 struct tcp_sock *tp = tcp_sk(sk);
5377 NET_INC_STATS(sock_net(sk), LINUX_MIB_PRUNECALLED);
5379 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
5380 tcp_clamp_window(sk);
5381 else if (tcp_under_memory_pressure(sk))
5382 tcp_adjust_rcv_ssthresh(sk);
5384 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5387 tcp_collapse_ofo_queue(sk);
5388 if (!skb_queue_empty(&sk->sk_receive_queue))
5389 tcp_collapse(sk, &sk->sk_receive_queue, NULL,
5390 skb_peek(&sk->sk_receive_queue),
5392 tp->copied_seq, tp->rcv_nxt);
5395 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5398 /* Collapsing did not help, destructive actions follow.
5399 * This must not ever occur. */
5401 tcp_prune_ofo_queue(sk);
5403 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5406 /* If we are really being abused, tell the caller to silently
5407 * drop receive data on the floor. It will get retransmitted
5408 * and hopefully then we'll have sufficient space.
5410 NET_INC_STATS(sock_net(sk), LINUX_MIB_RCVPRUNED);
5412 /* Massive buffer overcommit. */
5417 static bool tcp_should_expand_sndbuf(struct sock *sk)
5419 const struct tcp_sock *tp = tcp_sk(sk);
5421 /* If the user specified a specific send buffer setting, do
5424 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
5427 /* If we are under global TCP memory pressure, do not expand. */
5428 if (tcp_under_memory_pressure(sk)) {
5429 int unused_mem = sk_unused_reserved_mem(sk);
5431 /* Adjust sndbuf according to reserved mem. But make sure
5432 * it never goes below SOCK_MIN_SNDBUF.
5433 * See sk_stream_moderate_sndbuf() for more details.
5435 if (unused_mem > SOCK_MIN_SNDBUF)
5436 WRITE_ONCE(sk->sk_sndbuf, unused_mem);
5441 /* If we are under soft global TCP memory pressure, do not expand. */
5442 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
5445 /* If we filled the congestion window, do not expand. */
5446 if (tcp_packets_in_flight(tp) >= tcp_snd_cwnd(tp))
5452 static void tcp_new_space(struct sock *sk)
5454 struct tcp_sock *tp = tcp_sk(sk);
5456 if (tcp_should_expand_sndbuf(sk)) {
5457 tcp_sndbuf_expand(sk);
5458 tp->snd_cwnd_stamp = tcp_jiffies32;
5461 INDIRECT_CALL_1(sk->sk_write_space, sk_stream_write_space, sk);
5464 /* Caller made space either from:
5465 * 1) Freeing skbs in rtx queues (after tp->snd_una has advanced)
5466 * 2) Sent skbs from output queue (and thus advancing tp->snd_nxt)
5468 * We might be able to generate EPOLLOUT to the application if:
5469 * 1) Space consumed in output/rtx queues is below sk->sk_sndbuf/2
5470 * 2) notsent amount (tp->write_seq - tp->snd_nxt) became
5471 * small enough that tcp_stream_memory_free() decides it
5472 * is time to generate EPOLLOUT.
5474 void tcp_check_space(struct sock *sk)
5476 /* pairs with tcp_poll() */
5478 if (sk->sk_socket &&
5479 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
5481 if (!test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
5482 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
5486 static inline void tcp_data_snd_check(struct sock *sk)
5488 tcp_push_pending_frames(sk);
5489 tcp_check_space(sk);
5493 * Check if sending an ack is needed.
5495 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
5497 struct tcp_sock *tp = tcp_sk(sk);
5498 unsigned long rtt, delay;
5500 /* More than one full frame received... */
5501 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
5502 /* ... and right edge of window advances far enough.
5503 * (tcp_recvmsg() will send ACK otherwise).
5504 * If application uses SO_RCVLOWAT, we want send ack now if
5505 * we have not received enough bytes to satisfy the condition.
5507 (tp->rcv_nxt - tp->copied_seq < sk->sk_rcvlowat ||
5508 __tcp_select_window(sk) >= tp->rcv_wnd)) ||
5509 /* We ACK each frame or... */
5510 tcp_in_quickack_mode(sk) ||
5511 /* Protocol state mandates a one-time immediate ACK */
5512 inet_csk(sk)->icsk_ack.pending & ICSK_ACK_NOW) {
5518 if (!ofo_possible || RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
5519 tcp_send_delayed_ack(sk);
5523 if (!tcp_is_sack(tp) ||
5524 tp->compressed_ack >= sock_net(sk)->ipv4.sysctl_tcp_comp_sack_nr)
5527 if (tp->compressed_ack_rcv_nxt != tp->rcv_nxt) {
5528 tp->compressed_ack_rcv_nxt = tp->rcv_nxt;
5529 tp->dup_ack_counter = 0;
5531 if (tp->dup_ack_counter < TCP_FASTRETRANS_THRESH) {
5532 tp->dup_ack_counter++;
5535 tp->compressed_ack++;
5536 if (hrtimer_is_queued(&tp->compressed_ack_timer))
5539 /* compress ack timer : 5 % of rtt, but no more than tcp_comp_sack_delay_ns */
5541 rtt = tp->rcv_rtt_est.rtt_us;
5542 if (tp->srtt_us && tp->srtt_us < rtt)
5545 delay = min_t(unsigned long,
5546 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_delay_ns),
5547 rtt * (NSEC_PER_USEC >> 3)/20);
5549 hrtimer_start_range_ns(&tp->compressed_ack_timer, ns_to_ktime(delay),
5550 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_slack_ns),
5551 HRTIMER_MODE_REL_PINNED_SOFT);
5554 static inline void tcp_ack_snd_check(struct sock *sk)
5556 if (!inet_csk_ack_scheduled(sk)) {
5557 /* We sent a data segment already. */
5560 __tcp_ack_snd_check(sk, 1);
5564 * This routine is only called when we have urgent data
5565 * signaled. Its the 'slow' part of tcp_urg. It could be
5566 * moved inline now as tcp_urg is only called from one
5567 * place. We handle URGent data wrong. We have to - as
5568 * BSD still doesn't use the correction from RFC961.
5569 * For 1003.1g we should support a new option TCP_STDURG to permit
5570 * either form (or just set the sysctl tcp_stdurg).
5573 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
5575 struct tcp_sock *tp = tcp_sk(sk);
5576 u32 ptr = ntohs(th->urg_ptr);
5578 if (ptr && !READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_stdurg))
5580 ptr += ntohl(th->seq);
5582 /* Ignore urgent data that we've already seen and read. */
5583 if (after(tp->copied_seq, ptr))
5586 /* Do not replay urg ptr.
5588 * NOTE: interesting situation not covered by specs.
5589 * Misbehaving sender may send urg ptr, pointing to segment,
5590 * which we already have in ofo queue. We are not able to fetch
5591 * such data and will stay in TCP_URG_NOTYET until will be eaten
5592 * by recvmsg(). Seems, we are not obliged to handle such wicked
5593 * situations. But it is worth to think about possibility of some
5594 * DoSes using some hypothetical application level deadlock.
5596 if (before(ptr, tp->rcv_nxt))
5599 /* Do we already have a newer (or duplicate) urgent pointer? */
5600 if (tp->urg_data && !after(ptr, tp->urg_seq))
5603 /* Tell the world about our new urgent pointer. */
5606 /* We may be adding urgent data when the last byte read was
5607 * urgent. To do this requires some care. We cannot just ignore
5608 * tp->copied_seq since we would read the last urgent byte again
5609 * as data, nor can we alter copied_seq until this data arrives
5610 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5612 * NOTE. Double Dutch. Rendering to plain English: author of comment
5613 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5614 * and expect that both A and B disappear from stream. This is _wrong_.
5615 * Though this happens in BSD with high probability, this is occasional.
5616 * Any application relying on this is buggy. Note also, that fix "works"
5617 * only in this artificial test. Insert some normal data between A and B and we will
5618 * decline of BSD again. Verdict: it is better to remove to trap
5621 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5622 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5623 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5625 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5626 __skb_unlink(skb, &sk->sk_receive_queue);
5631 WRITE_ONCE(tp->urg_data, TCP_URG_NOTYET);
5632 WRITE_ONCE(tp->urg_seq, ptr);
5634 /* Disable header prediction. */
5638 /* This is the 'fast' part of urgent handling. */
5639 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
5641 struct tcp_sock *tp = tcp_sk(sk);
5643 /* Check if we get a new urgent pointer - normally not. */
5644 if (unlikely(th->urg))
5645 tcp_check_urg(sk, th);
5647 /* Do we wait for any urgent data? - normally not... */
5648 if (unlikely(tp->urg_data == TCP_URG_NOTYET)) {
5649 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5652 /* Is the urgent pointer pointing into this packet? */
5653 if (ptr < skb->len) {
5655 if (skb_copy_bits(skb, ptr, &tmp, 1))
5657 WRITE_ONCE(tp->urg_data, TCP_URG_VALID | tmp);
5658 if (!sock_flag(sk, SOCK_DEAD))
5659 sk->sk_data_ready(sk);
5664 /* Accept RST for rcv_nxt - 1 after a FIN.
5665 * When tcp connections are abruptly terminated from Mac OSX (via ^C), a
5666 * FIN is sent followed by a RST packet. The RST is sent with the same
5667 * sequence number as the FIN, and thus according to RFC 5961 a challenge
5668 * ACK should be sent. However, Mac OSX rate limits replies to challenge
5669 * ACKs on the closed socket. In addition middleboxes can drop either the
5670 * challenge ACK or a subsequent RST.
5672 static bool tcp_reset_check(const struct sock *sk, const struct sk_buff *skb)
5674 struct tcp_sock *tp = tcp_sk(sk);
5676 return unlikely(TCP_SKB_CB(skb)->seq == (tp->rcv_nxt - 1) &&
5677 (1 << sk->sk_state) & (TCPF_CLOSE_WAIT | TCPF_LAST_ACK |
5681 /* Does PAWS and seqno based validation of an incoming segment, flags will
5682 * play significant role here.
5684 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5685 const struct tcphdr *th, int syn_inerr)
5687 struct tcp_sock *tp = tcp_sk(sk);
5690 /* RFC1323: H1. Apply PAWS check first. */
5691 if (tcp_fast_parse_options(sock_net(sk), skb, th, tp) &&
5692 tp->rx_opt.saw_tstamp &&
5693 tcp_paws_discard(sk, skb)) {
5695 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5696 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5697 LINUX_MIB_TCPACKSKIPPEDPAWS,
5698 &tp->last_oow_ack_time))
5699 tcp_send_dupack(sk, skb);
5700 SKB_DR_SET(reason, TCP_RFC7323_PAWS);
5703 /* Reset is accepted even if it did not pass PAWS. */
5706 /* Step 1: check sequence number */
5707 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5708 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5709 * (RST) segments are validated by checking their SEQ-fields."
5710 * And page 69: "If an incoming segment is not acceptable,
5711 * an acknowledgment should be sent in reply (unless the RST
5712 * bit is set, if so drop the segment and return)".
5717 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5718 LINUX_MIB_TCPACKSKIPPEDSEQ,
5719 &tp->last_oow_ack_time))
5720 tcp_send_dupack(sk, skb);
5721 } else if (tcp_reset_check(sk, skb)) {
5724 SKB_DR_SET(reason, TCP_INVALID_SEQUENCE);
5728 /* Step 2: check RST bit */
5730 /* RFC 5961 3.2 (extend to match against (RCV.NXT - 1) after a
5731 * FIN and SACK too if available):
5732 * If seq num matches RCV.NXT or (RCV.NXT - 1) after a FIN, or
5733 * the right-most SACK block,
5735 * RESET the connection
5737 * Send a challenge ACK
5739 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt ||
5740 tcp_reset_check(sk, skb))
5743 if (tcp_is_sack(tp) && tp->rx_opt.num_sacks > 0) {
5744 struct tcp_sack_block *sp = &tp->selective_acks[0];
5745 int max_sack = sp[0].end_seq;
5748 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;
5750 max_sack = after(sp[this_sack].end_seq,
5752 sp[this_sack].end_seq : max_sack;
5755 if (TCP_SKB_CB(skb)->seq == max_sack)
5759 /* Disable TFO if RST is out-of-order
5760 * and no data has been received
5761 * for current active TFO socket
5763 if (tp->syn_fastopen && !tp->data_segs_in &&
5764 sk->sk_state == TCP_ESTABLISHED)
5765 tcp_fastopen_active_disable(sk);
5766 tcp_send_challenge_ack(sk);
5767 SKB_DR_SET(reason, TCP_RESET);
5771 /* step 3: check security and precedence [ignored] */
5773 /* step 4: Check for a SYN
5774 * RFC 5961 4.2 : Send a challenge ack
5779 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5780 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5781 tcp_send_challenge_ack(sk);
5782 SKB_DR_SET(reason, TCP_INVALID_SYN);
5786 bpf_skops_parse_hdr(sk, skb);
5791 tcp_drop_reason(sk, skb, reason);
5801 * TCP receive function for the ESTABLISHED state.
5803 * It is split into a fast path and a slow path. The fast path is
5805 * - A zero window was announced from us - zero window probing
5806 * is only handled properly in the slow path.
5807 * - Out of order segments arrived.
5808 * - Urgent data is expected.
5809 * - There is no buffer space left
5810 * - Unexpected TCP flags/window values/header lengths are received
5811 * (detected by checking the TCP header against pred_flags)
5812 * - Data is sent in both directions. Fast path only supports pure senders
5813 * or pure receivers (this means either the sequence number or the ack
5814 * value must stay constant)
5815 * - Unexpected TCP option.
5817 * When these conditions are not satisfied it drops into a standard
5818 * receive procedure patterned after RFC793 to handle all cases.
5819 * The first three cases are guaranteed by proper pred_flags setting,
5820 * the rest is checked inline. Fast processing is turned on in
5821 * tcp_data_queue when everything is OK.
5823 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb)
5825 enum skb_drop_reason reason = SKB_DROP_REASON_NOT_SPECIFIED;
5826 const struct tcphdr *th = (const struct tcphdr *)skb->data;
5827 struct tcp_sock *tp = tcp_sk(sk);
5828 unsigned int len = skb->len;
5830 /* TCP congestion window tracking */
5831 trace_tcp_probe(sk, skb);
5833 tcp_mstamp_refresh(tp);
5834 if (unlikely(!rcu_access_pointer(sk->sk_rx_dst)))
5835 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
5837 * Header prediction.
5838 * The code loosely follows the one in the famous
5839 * "30 instruction TCP receive" Van Jacobson mail.
5841 * Van's trick is to deposit buffers into socket queue
5842 * on a device interrupt, to call tcp_recv function
5843 * on the receive process context and checksum and copy
5844 * the buffer to user space. smart...
5846 * Our current scheme is not silly either but we take the
5847 * extra cost of the net_bh soft interrupt processing...
5848 * We do checksum and copy also but from device to kernel.
5851 tp->rx_opt.saw_tstamp = 0;
5853 /* pred_flags is 0xS?10 << 16 + snd_wnd
5854 * if header_prediction is to be made
5855 * 'S' will always be tp->tcp_header_len >> 2
5856 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5857 * turn it off (when there are holes in the receive
5858 * space for instance)
5859 * PSH flag is ignored.
5862 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5863 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5864 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5865 int tcp_header_len = tp->tcp_header_len;
5867 /* Timestamp header prediction: tcp_header_len
5868 * is automatically equal to th->doff*4 due to pred_flags
5872 /* Check timestamp */
5873 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5874 /* No? Slow path! */
5875 if (!tcp_parse_aligned_timestamp(tp, th))
5878 /* If PAWS failed, check it more carefully in slow path */
5879 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5882 /* DO NOT update ts_recent here, if checksum fails
5883 * and timestamp was corrupted part, it will result
5884 * in a hung connection since we will drop all
5885 * future packets due to the PAWS test.
5889 if (len <= tcp_header_len) {
5890 /* Bulk data transfer: sender */
5891 if (len == tcp_header_len) {
5892 /* Predicted packet is in window by definition.
5893 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5894 * Hence, check seq<=rcv_wup reduces to:
5896 if (tcp_header_len ==
5897 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5898 tp->rcv_nxt == tp->rcv_wup)
5899 tcp_store_ts_recent(tp);
5901 /* We know that such packets are checksummed
5904 tcp_ack(sk, skb, 0);
5906 tcp_data_snd_check(sk);
5907 /* When receiving pure ack in fast path, update
5908 * last ts ecr directly instead of calling
5909 * tcp_rcv_rtt_measure_ts()
5911 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr;
5913 } else { /* Header too small */
5914 reason = SKB_DROP_REASON_PKT_TOO_SMALL;
5915 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5920 bool fragstolen = false;
5922 if (tcp_checksum_complete(skb))
5925 if ((int)skb->truesize > sk->sk_forward_alloc)
5928 /* Predicted packet is in window by definition.
5929 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5930 * Hence, check seq<=rcv_wup reduces to:
5932 if (tcp_header_len ==
5933 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5934 tp->rcv_nxt == tp->rcv_wup)
5935 tcp_store_ts_recent(tp);
5937 tcp_rcv_rtt_measure_ts(sk, skb);
5939 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPHITS);
5941 /* Bulk data transfer: receiver */
5943 __skb_pull(skb, tcp_header_len);
5944 eaten = tcp_queue_rcv(sk, skb, &fragstolen);
5946 tcp_event_data_recv(sk, skb);
5948 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5949 /* Well, only one small jumplet in fast path... */
5950 tcp_ack(sk, skb, FLAG_DATA);
5951 tcp_data_snd_check(sk);
5952 if (!inet_csk_ack_scheduled(sk))
5955 tcp_update_wl(tp, TCP_SKB_CB(skb)->seq);
5958 __tcp_ack_snd_check(sk, 0);
5961 kfree_skb_partial(skb, fragstolen);
5968 if (len < (th->doff << 2) || tcp_checksum_complete(skb))
5971 if (!th->ack && !th->rst && !th->syn) {
5972 reason = SKB_DROP_REASON_TCP_FLAGS;
5977 * Standard slow path.
5980 if (!tcp_validate_incoming(sk, skb, th, 1))
5984 reason = tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT);
5985 if ((int)reason < 0) {
5989 tcp_rcv_rtt_measure_ts(sk, skb);
5991 /* Process urgent data. */
5992 tcp_urg(sk, skb, th);
5994 /* step 7: process the segment text */
5995 tcp_data_queue(sk, skb);
5997 tcp_data_snd_check(sk);
5998 tcp_ack_snd_check(sk);
6002 reason = SKB_DROP_REASON_TCP_CSUM;
6003 trace_tcp_bad_csum(skb);
6004 TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS);
6005 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
6008 tcp_drop_reason(sk, skb, reason);
6010 EXPORT_SYMBOL(tcp_rcv_established);
6012 void tcp_init_transfer(struct sock *sk, int bpf_op, struct sk_buff *skb)
6014 struct inet_connection_sock *icsk = inet_csk(sk);
6015 struct tcp_sock *tp = tcp_sk(sk);
6018 icsk->icsk_af_ops->rebuild_header(sk);
6019 tcp_init_metrics(sk);
6021 /* Initialize the congestion window to start the transfer.
6022 * Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
6023 * retransmitted. In light of RFC6298 more aggressive 1sec
6024 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
6025 * retransmission has occurred.
6027 if (tp->total_retrans > 1 && tp->undo_marker)
6028 tcp_snd_cwnd_set(tp, 1);
6030 tcp_snd_cwnd_set(tp, tcp_init_cwnd(tp, __sk_dst_get(sk)));
6031 tp->snd_cwnd_stamp = tcp_jiffies32;
6033 bpf_skops_established(sk, bpf_op, skb);
6034 /* Initialize congestion control unless BPF initialized it already: */
6035 if (!icsk->icsk_ca_initialized)
6036 tcp_init_congestion_control(sk);
6037 tcp_init_buffer_space(sk);
6040 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
6042 struct tcp_sock *tp = tcp_sk(sk);
6043 struct inet_connection_sock *icsk = inet_csk(sk);
6045 tcp_set_state(sk, TCP_ESTABLISHED);
6046 icsk->icsk_ack.lrcvtime = tcp_jiffies32;
6049 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
6050 security_inet_conn_established(sk, skb);
6051 sk_mark_napi_id(sk, skb);
6054 tcp_init_transfer(sk, BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB, skb);
6056 /* Prevent spurious tcp_cwnd_restart() on first data
6059 tp->lsndtime = tcp_jiffies32;
6061 if (sock_flag(sk, SOCK_KEEPOPEN))
6062 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
6064 if (!tp->rx_opt.snd_wscale)
6065 __tcp_fast_path_on(tp, tp->snd_wnd);
6070 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
6071 struct tcp_fastopen_cookie *cookie)
6073 struct tcp_sock *tp = tcp_sk(sk);
6074 struct sk_buff *data = tp->syn_data ? tcp_rtx_queue_head(sk) : NULL;
6075 u16 mss = tp->rx_opt.mss_clamp, try_exp = 0;
6076 bool syn_drop = false;
6078 if (mss == tp->rx_opt.user_mss) {
6079 struct tcp_options_received opt;
6081 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
6082 tcp_clear_options(&opt);
6083 opt.user_mss = opt.mss_clamp = 0;
6084 tcp_parse_options(sock_net(sk), synack, &opt, 0, NULL);
6085 mss = opt.mss_clamp;
6088 if (!tp->syn_fastopen) {
6089 /* Ignore an unsolicited cookie */
6091 } else if (tp->total_retrans) {
6092 /* SYN timed out and the SYN-ACK neither has a cookie nor
6093 * acknowledges data. Presumably the remote received only
6094 * the retransmitted (regular) SYNs: either the original
6095 * SYN-data or the corresponding SYN-ACK was dropped.
6097 syn_drop = (cookie->len < 0 && data);
6098 } else if (cookie->len < 0 && !tp->syn_data) {
6099 /* We requested a cookie but didn't get it. If we did not use
6100 * the (old) exp opt format then try so next time (try_exp=1).
6101 * Otherwise we go back to use the RFC7413 opt (try_exp=2).
6103 try_exp = tp->syn_fastopen_exp ? 2 : 1;
6106 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp);
6108 if (data) { /* Retransmit unacked data in SYN */
6109 if (tp->total_retrans)
6110 tp->fastopen_client_fail = TFO_SYN_RETRANSMITTED;
6112 tp->fastopen_client_fail = TFO_DATA_NOT_ACKED;
6113 skb_rbtree_walk_from(data)
6114 tcp_mark_skb_lost(sk, data);
6115 tcp_xmit_retransmit_queue(sk);
6116 NET_INC_STATS(sock_net(sk),
6117 LINUX_MIB_TCPFASTOPENACTIVEFAIL);
6120 tp->syn_data_acked = tp->syn_data;
6121 if (tp->syn_data_acked) {
6122 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
6123 /* SYN-data is counted as two separate packets in tcp_ack() */
6124 if (tp->delivered > 1)
6128 tcp_fastopen_add_skb(sk, synack);
6133 static void smc_check_reset_syn(struct tcp_sock *tp)
6135 #if IS_ENABLED(CONFIG_SMC)
6136 if (static_branch_unlikely(&tcp_have_smc)) {
6137 if (tp->syn_smc && !tp->rx_opt.smc_ok)
6143 static void tcp_try_undo_spurious_syn(struct sock *sk)
6145 struct tcp_sock *tp = tcp_sk(sk);
6148 /* undo_marker is set when SYN or SYNACK times out. The timeout is
6149 * spurious if the ACK's timestamp option echo value matches the
6150 * original SYN timestamp.
6152 syn_stamp = tp->retrans_stamp;
6153 if (tp->undo_marker && syn_stamp && tp->rx_opt.saw_tstamp &&
6154 syn_stamp == tp->rx_opt.rcv_tsecr)
6155 tp->undo_marker = 0;
6158 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
6159 const struct tcphdr *th)
6161 struct inet_connection_sock *icsk = inet_csk(sk);
6162 struct tcp_sock *tp = tcp_sk(sk);
6163 struct tcp_fastopen_cookie foc = { .len = -1 };
6164 int saved_clamp = tp->rx_opt.mss_clamp;
6168 tcp_parse_options(sock_net(sk), skb, &tp->rx_opt, 0, &foc);
6169 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
6170 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
6174 * "If the state is SYN-SENT then
6175 * first check the ACK bit
6176 * If the ACK bit is set
6177 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
6178 * a reset (unless the RST bit is set, if so drop
6179 * the segment and return)"
6181 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
6182 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
6183 /* Previous FIN/ACK or RST/ACK might be ignored. */
6184 if (icsk->icsk_retransmits == 0)
6185 inet_csk_reset_xmit_timer(sk,
6187 TCP_TIMEOUT_MIN, TCP_RTO_MAX);
6188 goto reset_and_undo;
6191 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
6192 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
6193 tcp_time_stamp(tp))) {
6194 NET_INC_STATS(sock_net(sk),
6195 LINUX_MIB_PAWSACTIVEREJECTED);
6196 goto reset_and_undo;
6199 /* Now ACK is acceptable.
6201 * "If the RST bit is set
6202 * If the ACK was acceptable then signal the user "error:
6203 * connection reset", drop the segment, enter CLOSED state,
6204 * delete TCB, and return."
6215 * "fifth, if neither of the SYN or RST bits is set then
6216 * drop the segment and return."
6222 SKB_DR_SET(reason, TCP_FLAGS);
6223 goto discard_and_undo;
6226 * "If the SYN bit is on ...
6227 * are acceptable then ...
6228 * (our SYN has been ACKed), change the connection
6229 * state to ESTABLISHED..."
6232 tcp_ecn_rcv_synack(tp, th);
6234 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
6235 tcp_try_undo_spurious_syn(sk);
6236 tcp_ack(sk, skb, FLAG_SLOWPATH);
6238 /* Ok.. it's good. Set up sequence numbers and
6239 * move to established.
6241 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1);
6242 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
6244 /* RFC1323: The window in SYN & SYN/ACK segments is
6247 tp->snd_wnd = ntohs(th->window);
6249 if (!tp->rx_opt.wscale_ok) {
6250 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
6251 tp->window_clamp = min(tp->window_clamp, 65535U);
6254 if (tp->rx_opt.saw_tstamp) {
6255 tp->rx_opt.tstamp_ok = 1;
6256 tp->tcp_header_len =
6257 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
6258 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
6259 tcp_store_ts_recent(tp);
6261 tp->tcp_header_len = sizeof(struct tcphdr);
6264 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
6265 tcp_initialize_rcv_mss(sk);
6267 /* Remember, tcp_poll() does not lock socket!
6268 * Change state from SYN-SENT only after copied_seq
6269 * is initialized. */
6270 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6272 smc_check_reset_syn(tp);
6276 tcp_finish_connect(sk, skb);
6278 fastopen_fail = (tp->syn_fastopen || tp->syn_data) &&
6279 tcp_rcv_fastopen_synack(sk, skb, &foc);
6281 if (!sock_flag(sk, SOCK_DEAD)) {
6282 sk->sk_state_change(sk);
6283 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
6287 if (sk->sk_write_pending ||
6288 icsk->icsk_accept_queue.rskq_defer_accept ||
6289 inet_csk_in_pingpong_mode(sk)) {
6290 /* Save one ACK. Data will be ready after
6291 * several ticks, if write_pending is set.
6293 * It may be deleted, but with this feature tcpdumps
6294 * look so _wonderfully_ clever, that I was not able
6295 * to stand against the temptation 8) --ANK
6297 inet_csk_schedule_ack(sk);
6298 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
6299 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
6300 TCP_DELACK_MAX, TCP_RTO_MAX);
6307 /* No ACK in the segment */
6311 * "If the RST bit is set
6313 * Otherwise (no ACK) drop the segment and return."
6315 SKB_DR_SET(reason, TCP_RESET);
6316 goto discard_and_undo;
6320 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
6321 tcp_paws_reject(&tp->rx_opt, 0)) {
6322 SKB_DR_SET(reason, TCP_RFC7323_PAWS);
6323 goto discard_and_undo;
6326 /* We see SYN without ACK. It is attempt of
6327 * simultaneous connect with crossed SYNs.
6328 * Particularly, it can be connect to self.
6330 tcp_set_state(sk, TCP_SYN_RECV);
6332 if (tp->rx_opt.saw_tstamp) {
6333 tp->rx_opt.tstamp_ok = 1;
6334 tcp_store_ts_recent(tp);
6335 tp->tcp_header_len =
6336 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
6338 tp->tcp_header_len = sizeof(struct tcphdr);
6341 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1);
6342 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6343 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
6345 /* RFC1323: The window in SYN & SYN/ACK segments is
6348 tp->snd_wnd = ntohs(th->window);
6349 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
6350 tp->max_window = tp->snd_wnd;
6352 tcp_ecn_rcv_syn(tp, th);
6355 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
6356 tcp_initialize_rcv_mss(sk);
6358 tcp_send_synack(sk);
6360 /* Note, we could accept data and URG from this segment.
6361 * There are no obstacles to make this (except that we must
6362 * either change tcp_recvmsg() to prevent it from returning data
6363 * before 3WHS completes per RFC793, or employ TCP Fast Open).
6365 * However, if we ignore data in ACKless segments sometimes,
6366 * we have no reasons to accept it sometimes.
6367 * Also, seems the code doing it in step6 of tcp_rcv_state_process
6368 * is not flawless. So, discard packet for sanity.
6369 * Uncomment this return to process the data.
6376 /* "fifth, if neither of the SYN or RST bits is set then
6377 * drop the segment and return."
6381 tcp_clear_options(&tp->rx_opt);
6382 tp->rx_opt.mss_clamp = saved_clamp;
6383 tcp_drop_reason(sk, skb, reason);
6387 tcp_clear_options(&tp->rx_opt);
6388 tp->rx_opt.mss_clamp = saved_clamp;
6392 static void tcp_rcv_synrecv_state_fastopen(struct sock *sk)
6394 struct request_sock *req;
6396 /* If we are still handling the SYNACK RTO, see if timestamp ECR allows
6397 * undo. If peer SACKs triggered fast recovery, we can't undo here.
6399 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
6400 tcp_try_undo_loss(sk, false);
6402 /* Reset rtx states to prevent spurious retransmits_timed_out() */
6403 tcp_sk(sk)->retrans_stamp = 0;
6404 inet_csk(sk)->icsk_retransmits = 0;
6406 /* Once we leave TCP_SYN_RECV or TCP_FIN_WAIT_1,
6407 * we no longer need req so release it.
6409 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk,
6410 lockdep_sock_is_held(sk));
6411 reqsk_fastopen_remove(sk, req, false);
6413 /* Re-arm the timer because data may have been sent out.
6414 * This is similar to the regular data transmission case
6415 * when new data has just been ack'ed.
6417 * (TFO) - we could try to be more aggressive and
6418 * retransmitting any data sooner based on when they
6425 * This function implements the receiving procedure of RFC 793 for
6426 * all states except ESTABLISHED and TIME_WAIT.
6427 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
6428 * address independent.
6431 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb)
6433 struct tcp_sock *tp = tcp_sk(sk);
6434 struct inet_connection_sock *icsk = inet_csk(sk);
6435 const struct tcphdr *th = tcp_hdr(skb);
6436 struct request_sock *req;
6441 switch (sk->sk_state) {
6443 SKB_DR_SET(reason, TCP_CLOSE);
6451 SKB_DR_SET(reason, TCP_RESET);
6456 SKB_DR_SET(reason, TCP_FLAGS);
6459 /* It is possible that we process SYN packets from backlog,
6460 * so we need to make sure to disable BH and RCU right there.
6464 acceptable = icsk->icsk_af_ops->conn_request(sk, skb) >= 0;
6473 SKB_DR_SET(reason, TCP_FLAGS);
6477 tp->rx_opt.saw_tstamp = 0;
6478 tcp_mstamp_refresh(tp);
6479 queued = tcp_rcv_synsent_state_process(sk, skb, th);
6483 /* Do step6 onward by hand. */
6484 tcp_urg(sk, skb, th);
6486 tcp_data_snd_check(sk);
6490 tcp_mstamp_refresh(tp);
6491 tp->rx_opt.saw_tstamp = 0;
6492 req = rcu_dereference_protected(tp->fastopen_rsk,
6493 lockdep_sock_is_held(sk));
6497 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
6498 sk->sk_state != TCP_FIN_WAIT1);
6500 if (!tcp_check_req(sk, skb, req, true, &req_stolen)) {
6501 SKB_DR_SET(reason, TCP_FASTOPEN);
6506 if (!th->ack && !th->rst && !th->syn) {
6507 SKB_DR_SET(reason, TCP_FLAGS);
6510 if (!tcp_validate_incoming(sk, skb, th, 0))
6513 /* step 5: check the ACK field */
6514 acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
6515 FLAG_UPDATE_TS_RECENT |
6516 FLAG_NO_CHALLENGE_ACK) > 0;
6519 if (sk->sk_state == TCP_SYN_RECV)
6520 return 1; /* send one RST */
6521 tcp_send_challenge_ack(sk);
6522 SKB_DR_SET(reason, TCP_OLD_ACK);
6525 switch (sk->sk_state) {
6527 tp->delivered++; /* SYN-ACK delivery isn't tracked in tcp_ack */
6529 tcp_synack_rtt_meas(sk, req);
6532 tcp_rcv_synrecv_state_fastopen(sk);
6534 tcp_try_undo_spurious_syn(sk);
6535 tp->retrans_stamp = 0;
6536 tcp_init_transfer(sk, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB,
6538 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6541 tcp_set_state(sk, TCP_ESTABLISHED);
6542 sk->sk_state_change(sk);
6544 /* Note, that this wakeup is only for marginal crossed SYN case.
6545 * Passively open sockets are not waked up, because
6546 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
6549 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
6551 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
6552 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
6553 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
6555 if (tp->rx_opt.tstamp_ok)
6556 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
6558 if (!inet_csk(sk)->icsk_ca_ops->cong_control)
6559 tcp_update_pacing_rate(sk);
6561 /* Prevent spurious tcp_cwnd_restart() on first data packet */
6562 tp->lsndtime = tcp_jiffies32;
6564 tcp_initialize_rcv_mss(sk);
6565 tcp_fast_path_on(tp);
6568 case TCP_FIN_WAIT1: {
6572 tcp_rcv_synrecv_state_fastopen(sk);
6574 if (tp->snd_una != tp->write_seq)
6577 tcp_set_state(sk, TCP_FIN_WAIT2);
6578 sk->sk_shutdown |= SEND_SHUTDOWN;
6582 if (!sock_flag(sk, SOCK_DEAD)) {
6583 /* Wake up lingering close() */
6584 sk->sk_state_change(sk);
6588 if (tp->linger2 < 0) {
6590 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6593 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6594 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6595 /* Receive out of order FIN after close() */
6596 if (tp->syn_fastopen && th->fin)
6597 tcp_fastopen_active_disable(sk);
6599 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6603 tmo = tcp_fin_time(sk);
6604 if (tmo > TCP_TIMEWAIT_LEN) {
6605 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
6606 } else if (th->fin || sock_owned_by_user(sk)) {
6607 /* Bad case. We could lose such FIN otherwise.
6608 * It is not a big problem, but it looks confusing
6609 * and not so rare event. We still can lose it now,
6610 * if it spins in bh_lock_sock(), but it is really
6613 inet_csk_reset_keepalive_timer(sk, tmo);
6615 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
6622 if (tp->snd_una == tp->write_seq) {
6623 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
6629 if (tp->snd_una == tp->write_seq) {
6630 tcp_update_metrics(sk);
6637 /* step 6: check the URG bit */
6638 tcp_urg(sk, skb, th);
6640 /* step 7: process the segment text */
6641 switch (sk->sk_state) {
6642 case TCP_CLOSE_WAIT:
6645 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
6646 /* If a subflow has been reset, the packet should not
6647 * continue to be processed, drop the packet.
6649 if (sk_is_mptcp(sk) && !mptcp_incoming_options(sk, skb))
6656 /* RFC 793 says to queue data in these states,
6657 * RFC 1122 says we MUST send a reset.
6658 * BSD 4.4 also does reset.
6660 if (sk->sk_shutdown & RCV_SHUTDOWN) {
6661 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6662 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6663 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6669 case TCP_ESTABLISHED:
6670 tcp_data_queue(sk, skb);
6675 /* tcp_data could move socket to TIME-WAIT */
6676 if (sk->sk_state != TCP_CLOSE) {
6677 tcp_data_snd_check(sk);
6678 tcp_ack_snd_check(sk);
6683 tcp_drop_reason(sk, skb, reason);
6691 EXPORT_SYMBOL(tcp_rcv_state_process);
6693 static inline void pr_drop_req(struct request_sock *req, __u16 port, int family)
6695 struct inet_request_sock *ireq = inet_rsk(req);
6697 if (family == AF_INET)
6698 net_dbg_ratelimited("drop open request from %pI4/%u\n",
6699 &ireq->ir_rmt_addr, port);
6700 #if IS_ENABLED(CONFIG_IPV6)
6701 else if (family == AF_INET6)
6702 net_dbg_ratelimited("drop open request from %pI6/%u\n",
6703 &ireq->ir_v6_rmt_addr, port);
6707 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
6709 * If we receive a SYN packet with these bits set, it means a
6710 * network is playing bad games with TOS bits. In order to
6711 * avoid possible false congestion notifications, we disable
6712 * TCP ECN negotiation.
6714 * Exception: tcp_ca wants ECN. This is required for DCTCP
6715 * congestion control: Linux DCTCP asserts ECT on all packets,
6716 * including SYN, which is most optimal solution; however,
6717 * others, such as FreeBSD do not.
6719 * Exception: At least one of the reserved bits of the TCP header (th->res1) is
6720 * set, indicating the use of a future TCP extension (such as AccECN). See
6721 * RFC8311 §4.3 which updates RFC3168 to allow the development of such
6724 static void tcp_ecn_create_request(struct request_sock *req,
6725 const struct sk_buff *skb,
6726 const struct sock *listen_sk,
6727 const struct dst_entry *dst)
6729 const struct tcphdr *th = tcp_hdr(skb);
6730 const struct net *net = sock_net(listen_sk);
6731 bool th_ecn = th->ece && th->cwr;
6738 ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield);
6739 ecn_ok_dst = dst_feature(dst, DST_FEATURE_ECN_MASK);
6740 ecn_ok = READ_ONCE(net->ipv4.sysctl_tcp_ecn) || ecn_ok_dst;
6742 if (((!ect || th->res1) && ecn_ok) || tcp_ca_needs_ecn(listen_sk) ||
6743 (ecn_ok_dst & DST_FEATURE_ECN_CA) ||
6744 tcp_bpf_ca_needs_ecn((struct sock *)req))
6745 inet_rsk(req)->ecn_ok = 1;
6748 static void tcp_openreq_init(struct request_sock *req,
6749 const struct tcp_options_received *rx_opt,
6750 struct sk_buff *skb, const struct sock *sk)
6752 struct inet_request_sock *ireq = inet_rsk(req);
6754 req->rsk_rcv_wnd = 0; /* So that tcp_send_synack() knows! */
6755 tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq;
6756 tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
6757 tcp_rsk(req)->snt_synack = 0;
6758 tcp_rsk(req)->last_oow_ack_time = 0;
6759 req->mss = rx_opt->mss_clamp;
6760 req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0;
6761 ireq->tstamp_ok = rx_opt->tstamp_ok;
6762 ireq->sack_ok = rx_opt->sack_ok;
6763 ireq->snd_wscale = rx_opt->snd_wscale;
6764 ireq->wscale_ok = rx_opt->wscale_ok;
6767 ireq->ir_rmt_port = tcp_hdr(skb)->source;
6768 ireq->ir_num = ntohs(tcp_hdr(skb)->dest);
6769 ireq->ir_mark = inet_request_mark(sk, skb);
6770 #if IS_ENABLED(CONFIG_SMC)
6771 ireq->smc_ok = rx_opt->smc_ok && !(tcp_sk(sk)->smc_hs_congested &&
6772 tcp_sk(sk)->smc_hs_congested(sk));
6776 struct request_sock *inet_reqsk_alloc(const struct request_sock_ops *ops,
6777 struct sock *sk_listener,
6778 bool attach_listener)
6780 struct request_sock *req = reqsk_alloc(ops, sk_listener,
6784 struct inet_request_sock *ireq = inet_rsk(req);
6786 ireq->ireq_opt = NULL;
6787 #if IS_ENABLED(CONFIG_IPV6)
6788 ireq->pktopts = NULL;
6790 atomic64_set(&ireq->ir_cookie, 0);
6791 ireq->ireq_state = TCP_NEW_SYN_RECV;
6792 write_pnet(&ireq->ireq_net, sock_net(sk_listener));
6793 ireq->ireq_family = sk_listener->sk_family;
6794 req->timeout = TCP_TIMEOUT_INIT;
6799 EXPORT_SYMBOL(inet_reqsk_alloc);
6802 * Return true if a syncookie should be sent
6804 static bool tcp_syn_flood_action(const struct sock *sk, const char *proto)
6806 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
6807 const char *msg = "Dropping request";
6808 struct net *net = sock_net(sk);
6809 bool want_cookie = false;
6812 syncookies = READ_ONCE(net->ipv4.sysctl_tcp_syncookies);
6814 #ifdef CONFIG_SYN_COOKIES
6816 msg = "Sending cookies";
6818 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES);
6821 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP);
6823 if (!queue->synflood_warned && syncookies != 2 &&
6824 xchg(&queue->synflood_warned, 1) == 0)
6825 net_info_ratelimited("%s: Possible SYN flooding on port %d. %s. Check SNMP counters.\n",
6826 proto, sk->sk_num, msg);
6831 static void tcp_reqsk_record_syn(const struct sock *sk,
6832 struct request_sock *req,
6833 const struct sk_buff *skb)
6835 if (tcp_sk(sk)->save_syn) {
6836 u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb);
6837 struct saved_syn *saved_syn;
6841 if (tcp_sk(sk)->save_syn == 2) { /* Save full header. */
6842 base = skb_mac_header(skb);
6843 mac_hdrlen = skb_mac_header_len(skb);
6846 base = skb_network_header(skb);
6850 saved_syn = kmalloc(struct_size(saved_syn, data, len),
6853 saved_syn->mac_hdrlen = mac_hdrlen;
6854 saved_syn->network_hdrlen = skb_network_header_len(skb);
6855 saved_syn->tcp_hdrlen = tcp_hdrlen(skb);
6856 memcpy(saved_syn->data, base, len);
6857 req->saved_syn = saved_syn;
6862 /* If a SYN cookie is required and supported, returns a clamped MSS value to be
6863 * used for SYN cookie generation.
6865 u16 tcp_get_syncookie_mss(struct request_sock_ops *rsk_ops,
6866 const struct tcp_request_sock_ops *af_ops,
6867 struct sock *sk, struct tcphdr *th)
6869 struct tcp_sock *tp = tcp_sk(sk);
6872 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_syncookies) != 2 &&
6873 !inet_csk_reqsk_queue_is_full(sk))
6876 if (!tcp_syn_flood_action(sk, rsk_ops->slab_name))
6879 if (sk_acceptq_is_full(sk)) {
6880 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
6884 mss = tcp_parse_mss_option(th, tp->rx_opt.user_mss);
6886 mss = af_ops->mss_clamp;
6890 EXPORT_SYMBOL_GPL(tcp_get_syncookie_mss);
6892 int tcp_conn_request(struct request_sock_ops *rsk_ops,
6893 const struct tcp_request_sock_ops *af_ops,
6894 struct sock *sk, struct sk_buff *skb)
6896 struct tcp_fastopen_cookie foc = { .len = -1 };
6897 __u32 isn = TCP_SKB_CB(skb)->tcp_tw_isn;
6898 struct tcp_options_received tmp_opt;
6899 struct tcp_sock *tp = tcp_sk(sk);
6900 struct net *net = sock_net(sk);
6901 struct sock *fastopen_sk = NULL;
6902 struct request_sock *req;
6903 bool want_cookie = false;
6904 struct dst_entry *dst;
6908 syncookies = READ_ONCE(net->ipv4.sysctl_tcp_syncookies);
6910 /* TW buckets are converted to open requests without
6911 * limitations, they conserve resources and peer is
6912 * evidently real one.
6914 if ((syncookies == 2 || inet_csk_reqsk_queue_is_full(sk)) && !isn) {
6915 want_cookie = tcp_syn_flood_action(sk, rsk_ops->slab_name);
6920 if (sk_acceptq_is_full(sk)) {
6921 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
6925 req = inet_reqsk_alloc(rsk_ops, sk, !want_cookie);
6929 req->syncookie = want_cookie;
6930 tcp_rsk(req)->af_specific = af_ops;
6931 tcp_rsk(req)->ts_off = 0;
6932 #if IS_ENABLED(CONFIG_MPTCP)
6933 tcp_rsk(req)->is_mptcp = 0;
6936 tcp_clear_options(&tmp_opt);
6937 tmp_opt.mss_clamp = af_ops->mss_clamp;
6938 tmp_opt.user_mss = tp->rx_opt.user_mss;
6939 tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0,
6940 want_cookie ? NULL : &foc);
6942 if (want_cookie && !tmp_opt.saw_tstamp)
6943 tcp_clear_options(&tmp_opt);
6945 if (IS_ENABLED(CONFIG_SMC) && want_cookie)
6948 tmp_opt.tstamp_ok = tmp_opt.saw_tstamp;
6949 tcp_openreq_init(req, &tmp_opt, skb, sk);
6950 inet_rsk(req)->no_srccheck = inet_sk(sk)->transparent;
6952 /* Note: tcp_v6_init_req() might override ir_iif for link locals */
6953 inet_rsk(req)->ir_iif = inet_request_bound_dev_if(sk, skb);
6955 dst = af_ops->route_req(sk, skb, &fl, req);
6959 if (tmp_opt.tstamp_ok)
6960 tcp_rsk(req)->ts_off = af_ops->init_ts_off(net, skb);
6962 if (!want_cookie && !isn) {
6963 int max_syn_backlog = READ_ONCE(net->ipv4.sysctl_max_syn_backlog);
6965 /* Kill the following clause, if you dislike this way. */
6967 (max_syn_backlog - inet_csk_reqsk_queue_len(sk) <
6968 (max_syn_backlog >> 2)) &&
6969 !tcp_peer_is_proven(req, dst)) {
6970 /* Without syncookies last quarter of
6971 * backlog is filled with destinations,
6972 * proven to be alive.
6973 * It means that we continue to communicate
6974 * to destinations, already remembered
6975 * to the moment of synflood.
6977 pr_drop_req(req, ntohs(tcp_hdr(skb)->source),
6979 goto drop_and_release;
6982 isn = af_ops->init_seq(skb);
6985 tcp_ecn_create_request(req, skb, sk, dst);
6988 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss);
6989 if (!tmp_opt.tstamp_ok)
6990 inet_rsk(req)->ecn_ok = 0;
6993 tcp_rsk(req)->snt_isn = isn;
6994 tcp_rsk(req)->txhash = net_tx_rndhash();
6995 tcp_rsk(req)->syn_tos = TCP_SKB_CB(skb)->ip_dsfield;
6996 tcp_openreq_init_rwin(req, sk, dst);
6997 sk_rx_queue_set(req_to_sk(req), skb);
6999 tcp_reqsk_record_syn(sk, req, skb);
7000 fastopen_sk = tcp_try_fastopen(sk, skb, req, &foc, dst);
7003 af_ops->send_synack(fastopen_sk, dst, &fl, req,
7004 &foc, TCP_SYNACK_FASTOPEN, skb);
7005 /* Add the child socket directly into the accept queue */
7006 if (!inet_csk_reqsk_queue_add(sk, req, fastopen_sk)) {
7007 reqsk_fastopen_remove(fastopen_sk, req, false);
7008 bh_unlock_sock(fastopen_sk);
7009 sock_put(fastopen_sk);
7012 sk->sk_data_ready(sk);
7013 bh_unlock_sock(fastopen_sk);
7014 sock_put(fastopen_sk);
7016 tcp_rsk(req)->tfo_listener = false;
7018 req->timeout = tcp_timeout_init((struct sock *)req);
7019 inet_csk_reqsk_queue_hash_add(sk, req, req->timeout);
7021 af_ops->send_synack(sk, dst, &fl, req, &foc,
7022 !want_cookie ? TCP_SYNACK_NORMAL :
7041 EXPORT_SYMBOL(tcp_conn_request);