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);
811 icsk->icsk_ack.lrcvtime = now;
813 tcp_ecn_check_ce(sk, skb);
816 tcp_grow_window(sk, skb, true);
819 /* Called to compute a smoothed rtt estimate. The data fed to this
820 * routine either comes from timestamps, or from segments that were
821 * known _not_ to have been retransmitted [see Karn/Partridge
822 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
823 * piece by Van Jacobson.
824 * NOTE: the next three routines used to be one big routine.
825 * To save cycles in the RFC 1323 implementation it was better to break
826 * it up into three procedures. -- erics
828 static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
830 struct tcp_sock *tp = tcp_sk(sk);
831 long m = mrtt_us; /* RTT */
832 u32 srtt = tp->srtt_us;
834 /* The following amusing code comes from Jacobson's
835 * article in SIGCOMM '88. Note that rtt and mdev
836 * are scaled versions of rtt and mean deviation.
837 * This is designed to be as fast as possible
838 * m stands for "measurement".
840 * On a 1990 paper the rto value is changed to:
841 * RTO = rtt + 4 * mdev
843 * Funny. This algorithm seems to be very broken.
844 * These formulae increase RTO, when it should be decreased, increase
845 * too slowly, when it should be increased quickly, decrease too quickly
846 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
847 * does not matter how to _calculate_ it. Seems, it was trap
848 * that VJ failed to avoid. 8)
851 m -= (srtt >> 3); /* m is now error in rtt est */
852 srtt += m; /* rtt = 7/8 rtt + 1/8 new */
854 m = -m; /* m is now abs(error) */
855 m -= (tp->mdev_us >> 2); /* similar update on mdev */
856 /* This is similar to one of Eifel findings.
857 * Eifel blocks mdev updates when rtt decreases.
858 * This solution is a bit different: we use finer gain
859 * for mdev in this case (alpha*beta).
860 * Like Eifel it also prevents growth of rto,
861 * but also it limits too fast rto decreases,
862 * happening in pure Eifel.
867 m -= (tp->mdev_us >> 2); /* similar update on mdev */
869 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */
870 if (tp->mdev_us > tp->mdev_max_us) {
871 tp->mdev_max_us = tp->mdev_us;
872 if (tp->mdev_max_us > tp->rttvar_us)
873 tp->rttvar_us = tp->mdev_max_us;
875 if (after(tp->snd_una, tp->rtt_seq)) {
876 if (tp->mdev_max_us < tp->rttvar_us)
877 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
878 tp->rtt_seq = tp->snd_nxt;
879 tp->mdev_max_us = tcp_rto_min_us(sk);
884 /* no previous measure. */
885 srtt = m << 3; /* take the measured time to be rtt */
886 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */
887 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
888 tp->mdev_max_us = tp->rttvar_us;
889 tp->rtt_seq = tp->snd_nxt;
893 tp->srtt_us = max(1U, srtt);
896 static void tcp_update_pacing_rate(struct sock *sk)
898 const struct tcp_sock *tp = tcp_sk(sk);
901 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
902 rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3);
904 /* current rate is (cwnd * mss) / srtt
905 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
906 * In Congestion Avoidance phase, set it to 120 % the current rate.
908 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
909 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
910 * end of slow start and should slow down.
912 if (tcp_snd_cwnd(tp) < tp->snd_ssthresh / 2)
913 rate *= READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_pacing_ss_ratio);
915 rate *= READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_pacing_ca_ratio);
917 rate *= max(tcp_snd_cwnd(tp), tp->packets_out);
919 if (likely(tp->srtt_us))
920 do_div(rate, tp->srtt_us);
922 /* WRITE_ONCE() is needed because sch_fq fetches sk_pacing_rate
923 * without any lock. We want to make sure compiler wont store
924 * intermediate values in this location.
926 WRITE_ONCE(sk->sk_pacing_rate, min_t(u64, rate,
927 sk->sk_max_pacing_rate));
930 /* Calculate rto without backoff. This is the second half of Van Jacobson's
931 * routine referred to above.
933 static void tcp_set_rto(struct sock *sk)
935 const struct tcp_sock *tp = tcp_sk(sk);
936 /* Old crap is replaced with new one. 8)
939 * 1. If rtt variance happened to be less 50msec, it is hallucination.
940 * It cannot be less due to utterly erratic ACK generation made
941 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
942 * to do with delayed acks, because at cwnd>2 true delack timeout
943 * is invisible. Actually, Linux-2.4 also generates erratic
944 * ACKs in some circumstances.
946 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
948 /* 2. Fixups made earlier cannot be right.
949 * If we do not estimate RTO correctly without them,
950 * all the algo is pure shit and should be replaced
951 * with correct one. It is exactly, which we pretend to do.
954 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
955 * guarantees that rto is higher.
960 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
962 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
965 cwnd = TCP_INIT_CWND;
966 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
969 struct tcp_sacktag_state {
970 /* Timestamps for earliest and latest never-retransmitted segment
971 * that was SACKed. RTO needs the earliest RTT to stay conservative,
972 * but congestion control should still get an accurate delay signal.
979 unsigned int mss_now;
980 struct rate_sample *rate;
983 /* Take a notice that peer is sending D-SACKs. Skip update of data delivery
984 * and spurious retransmission information if this DSACK is unlikely caused by
986 * - DSACKed sequence range is larger than maximum receiver's window.
987 * - Total no. of DSACKed segments exceed the total no. of retransmitted segs.
989 static u32 tcp_dsack_seen(struct tcp_sock *tp, u32 start_seq,
990 u32 end_seq, struct tcp_sacktag_state *state)
992 u32 seq_len, dup_segs = 1;
994 if (!before(start_seq, end_seq))
997 seq_len = end_seq - start_seq;
998 /* Dubious DSACK: DSACKed range greater than maximum advertised rwnd */
999 if (seq_len > tp->max_window)
1001 if (seq_len > tp->mss_cache)
1002 dup_segs = DIV_ROUND_UP(seq_len, tp->mss_cache);
1003 else if (tp->tlp_high_seq && tp->tlp_high_seq == end_seq)
1004 state->flag |= FLAG_DSACK_TLP;
1006 tp->dsack_dups += dup_segs;
1007 /* Skip the DSACK if dup segs weren't retransmitted by sender */
1008 if (tp->dsack_dups > tp->total_retrans)
1011 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
1012 /* We increase the RACK ordering window in rounds where we receive
1013 * DSACKs that may have been due to reordering causing RACK to trigger
1014 * a spurious fast recovery. Thus RACK ignores DSACKs that happen
1015 * without having seen reordering, or that match TLP probes (TLP
1016 * is timer-driven, not triggered by RACK).
1018 if (tp->reord_seen && !(state->flag & FLAG_DSACK_TLP))
1019 tp->rack.dsack_seen = 1;
1021 state->flag |= FLAG_DSACKING_ACK;
1022 /* A spurious retransmission is delivered */
1023 state->sack_delivered += dup_segs;
1028 /* It's reordering when higher sequence was delivered (i.e. sacked) before
1029 * some lower never-retransmitted sequence ("low_seq"). The maximum reordering
1030 * distance is approximated in full-mss packet distance ("reordering").
1032 static void tcp_check_sack_reordering(struct sock *sk, const u32 low_seq,
1035 struct tcp_sock *tp = tcp_sk(sk);
1036 const u32 mss = tp->mss_cache;
1039 fack = tcp_highest_sack_seq(tp);
1040 if (!before(low_seq, fack))
1043 metric = fack - low_seq;
1044 if ((metric > tp->reordering * mss) && mss) {
1045 #if FASTRETRANS_DEBUG > 1
1046 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
1047 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
1051 tp->undo_marker ? tp->undo_retrans : 0);
1053 tp->reordering = min_t(u32, (metric + mss - 1) / mss,
1054 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_max_reordering));
1057 /* This exciting event is worth to be remembered. 8) */
1059 NET_INC_STATS(sock_net(sk),
1060 ts ? LINUX_MIB_TCPTSREORDER : LINUX_MIB_TCPSACKREORDER);
1063 /* This must be called before lost_out or retrans_out are updated
1064 * on a new loss, because we want to know if all skbs previously
1065 * known to be lost have already been retransmitted, indicating
1066 * that this newly lost skb is our next skb to retransmit.
1068 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
1070 if ((!tp->retransmit_skb_hint && tp->retrans_out >= tp->lost_out) ||
1071 (tp->retransmit_skb_hint &&
1072 before(TCP_SKB_CB(skb)->seq,
1073 TCP_SKB_CB(tp->retransmit_skb_hint)->seq)))
1074 tp->retransmit_skb_hint = skb;
1077 /* Sum the number of packets on the wire we have marked as lost, and
1078 * notify the congestion control module that the given skb was marked lost.
1080 static void tcp_notify_skb_loss_event(struct tcp_sock *tp, const struct sk_buff *skb)
1082 tp->lost += tcp_skb_pcount(skb);
1085 void tcp_mark_skb_lost(struct sock *sk, struct sk_buff *skb)
1087 __u8 sacked = TCP_SKB_CB(skb)->sacked;
1088 struct tcp_sock *tp = tcp_sk(sk);
1090 if (sacked & TCPCB_SACKED_ACKED)
1093 tcp_verify_retransmit_hint(tp, skb);
1094 if (sacked & TCPCB_LOST) {
1095 if (sacked & TCPCB_SACKED_RETRANS) {
1096 /* Account for retransmits that are lost again */
1097 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1098 tp->retrans_out -= tcp_skb_pcount(skb);
1099 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT,
1100 tcp_skb_pcount(skb));
1101 tcp_notify_skb_loss_event(tp, skb);
1104 tp->lost_out += tcp_skb_pcount(skb);
1105 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1106 tcp_notify_skb_loss_event(tp, skb);
1110 /* Updates the delivered and delivered_ce counts */
1111 static void tcp_count_delivered(struct tcp_sock *tp, u32 delivered,
1114 tp->delivered += delivered;
1116 tp->delivered_ce += delivered;
1119 /* This procedure tags the retransmission queue when SACKs arrive.
1121 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
1122 * Packets in queue with these bits set are counted in variables
1123 * sacked_out, retrans_out and lost_out, correspondingly.
1125 * Valid combinations are:
1126 * Tag InFlight Description
1127 * 0 1 - orig segment is in flight.
1128 * S 0 - nothing flies, orig reached receiver.
1129 * L 0 - nothing flies, orig lost by net.
1130 * R 2 - both orig and retransmit are in flight.
1131 * L|R 1 - orig is lost, retransmit is in flight.
1132 * S|R 1 - orig reached receiver, retrans is still in flight.
1133 * (L|S|R is logically valid, it could occur when L|R is sacked,
1134 * but it is equivalent to plain S and code short-curcuits it to S.
1135 * L|S is logically invalid, it would mean -1 packet in flight 8))
1137 * These 6 states form finite state machine, controlled by the following events:
1138 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
1139 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
1140 * 3. Loss detection event of two flavors:
1141 * A. Scoreboard estimator decided the packet is lost.
1142 * A'. Reno "three dupacks" marks head of queue lost.
1143 * B. SACK arrives sacking SND.NXT at the moment, when the
1144 * segment was retransmitted.
1145 * 4. D-SACK added new rule: D-SACK changes any tag to S.
1147 * It is pleasant to note, that state diagram turns out to be commutative,
1148 * so that we are allowed not to be bothered by order of our actions,
1149 * when multiple events arrive simultaneously. (see the function below).
1151 * Reordering detection.
1152 * --------------------
1153 * Reordering metric is maximal distance, which a packet can be displaced
1154 * in packet stream. With SACKs we can estimate it:
1156 * 1. SACK fills old hole and the corresponding segment was not
1157 * ever retransmitted -> reordering. Alas, we cannot use it
1158 * when segment was retransmitted.
1159 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1160 * for retransmitted and already SACKed segment -> reordering..
1161 * Both of these heuristics are not used in Loss state, when we cannot
1162 * account for retransmits accurately.
1164 * SACK block validation.
1165 * ----------------------
1167 * SACK block range validation checks that the received SACK block fits to
1168 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1169 * Note that SND.UNA is not included to the range though being valid because
1170 * it means that the receiver is rather inconsistent with itself reporting
1171 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1172 * perfectly valid, however, in light of RFC2018 which explicitly states
1173 * that "SACK block MUST reflect the newest segment. Even if the newest
1174 * segment is going to be discarded ...", not that it looks very clever
1175 * in case of head skb. Due to potentional receiver driven attacks, we
1176 * choose to avoid immediate execution of a walk in write queue due to
1177 * reneging and defer head skb's loss recovery to standard loss recovery
1178 * procedure that will eventually trigger (nothing forbids us doing this).
1180 * Implements also blockage to start_seq wrap-around. Problem lies in the
1181 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1182 * there's no guarantee that it will be before snd_nxt (n). The problem
1183 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1186 * <- outs wnd -> <- wrapzone ->
1187 * u e n u_w e_w s n_w
1189 * |<------------+------+----- TCP seqno space --------------+---------->|
1190 * ...-- <2^31 ->| |<--------...
1191 * ...---- >2^31 ------>| |<--------...
1193 * Current code wouldn't be vulnerable but it's better still to discard such
1194 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1195 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1196 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1197 * equal to the ideal case (infinite seqno space without wrap caused issues).
1199 * With D-SACK the lower bound is extended to cover sequence space below
1200 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1201 * again, D-SACK block must not to go across snd_una (for the same reason as
1202 * for the normal SACK blocks, explained above). But there all simplicity
1203 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1204 * fully below undo_marker they do not affect behavior in anyway and can
1205 * therefore be safely ignored. In rare cases (which are more or less
1206 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1207 * fragmentation and packet reordering past skb's retransmission. To consider
1208 * them correctly, the acceptable range must be extended even more though
1209 * the exact amount is rather hard to quantify. However, tp->max_window can
1210 * be used as an exaggerated estimate.
1212 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
1213 u32 start_seq, u32 end_seq)
1215 /* Too far in future, or reversed (interpretation is ambiguous) */
1216 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1219 /* Nasty start_seq wrap-around check (see comments above) */
1220 if (!before(start_seq, tp->snd_nxt))
1223 /* In outstanding window? ...This is valid exit for D-SACKs too.
1224 * start_seq == snd_una is non-sensical (see comments above)
1226 if (after(start_seq, tp->snd_una))
1229 if (!is_dsack || !tp->undo_marker)
1232 /* ...Then it's D-SACK, and must reside below snd_una completely */
1233 if (after(end_seq, tp->snd_una))
1236 if (!before(start_seq, tp->undo_marker))
1240 if (!after(end_seq, tp->undo_marker))
1243 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1244 * start_seq < undo_marker and end_seq >= undo_marker.
1246 return !before(start_seq, end_seq - tp->max_window);
1249 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1250 struct tcp_sack_block_wire *sp, int num_sacks,
1251 u32 prior_snd_una, struct tcp_sacktag_state *state)
1253 struct tcp_sock *tp = tcp_sk(sk);
1254 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1255 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1258 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1259 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1260 } else if (num_sacks > 1) {
1261 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1262 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1264 if (after(end_seq_0, end_seq_1) || before(start_seq_0, start_seq_1))
1266 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKOFORECV);
1271 dup_segs = tcp_dsack_seen(tp, start_seq_0, end_seq_0, state);
1272 if (!dup_segs) { /* Skip dubious DSACK */
1273 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKIGNOREDDUBIOUS);
1277 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECVSEGS, dup_segs);
1279 /* D-SACK for already forgotten data... Do dumb counting. */
1280 if (tp->undo_marker && tp->undo_retrans > 0 &&
1281 !after(end_seq_0, prior_snd_una) &&
1282 after(end_seq_0, tp->undo_marker))
1283 tp->undo_retrans = max_t(int, 0, tp->undo_retrans - dup_segs);
1288 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1289 * the incoming SACK may not exactly match but we can find smaller MSS
1290 * aligned portion of it that matches. Therefore we might need to fragment
1291 * which may fail and creates some hassle (caller must handle error case
1294 * FIXME: this could be merged to shift decision code
1296 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1297 u32 start_seq, u32 end_seq)
1301 unsigned int pkt_len;
1304 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1305 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1307 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1308 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1309 mss = tcp_skb_mss(skb);
1310 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1313 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1317 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1322 /* Round if necessary so that SACKs cover only full MSSes
1323 * and/or the remaining small portion (if present)
1325 if (pkt_len > mss) {
1326 unsigned int new_len = (pkt_len / mss) * mss;
1327 if (!in_sack && new_len < pkt_len)
1332 if (pkt_len >= skb->len && !in_sack)
1335 err = tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
1336 pkt_len, mss, GFP_ATOMIC);
1344 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1345 static u8 tcp_sacktag_one(struct sock *sk,
1346 struct tcp_sacktag_state *state, u8 sacked,
1347 u32 start_seq, u32 end_seq,
1348 int dup_sack, int pcount,
1351 struct tcp_sock *tp = tcp_sk(sk);
1353 /* Account D-SACK for retransmitted packet. */
1354 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1355 if (tp->undo_marker && tp->undo_retrans > 0 &&
1356 after(end_seq, tp->undo_marker))
1357 tp->undo_retrans = max_t(int, 0, tp->undo_retrans - pcount);
1358 if ((sacked & TCPCB_SACKED_ACKED) &&
1359 before(start_seq, state->reord))
1360 state->reord = start_seq;
1363 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1364 if (!after(end_seq, tp->snd_una))
1367 if (!(sacked & TCPCB_SACKED_ACKED)) {
1368 tcp_rack_advance(tp, sacked, end_seq, xmit_time);
1370 if (sacked & TCPCB_SACKED_RETRANS) {
1371 /* If the segment is not tagged as lost,
1372 * we do not clear RETRANS, believing
1373 * that retransmission is still in flight.
1375 if (sacked & TCPCB_LOST) {
1376 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1377 tp->lost_out -= pcount;
1378 tp->retrans_out -= pcount;
1381 if (!(sacked & TCPCB_RETRANS)) {
1382 /* New sack for not retransmitted frame,
1383 * which was in hole. It is reordering.
1385 if (before(start_seq,
1386 tcp_highest_sack_seq(tp)) &&
1387 before(start_seq, state->reord))
1388 state->reord = start_seq;
1390 if (!after(end_seq, tp->high_seq))
1391 state->flag |= FLAG_ORIG_SACK_ACKED;
1392 if (state->first_sackt == 0)
1393 state->first_sackt = xmit_time;
1394 state->last_sackt = xmit_time;
1397 if (sacked & TCPCB_LOST) {
1398 sacked &= ~TCPCB_LOST;
1399 tp->lost_out -= pcount;
1403 sacked |= TCPCB_SACKED_ACKED;
1404 state->flag |= FLAG_DATA_SACKED;
1405 tp->sacked_out += pcount;
1406 /* Out-of-order packets delivered */
1407 state->sack_delivered += pcount;
1409 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1410 if (tp->lost_skb_hint &&
1411 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1412 tp->lost_cnt_hint += pcount;
1415 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1416 * frames and clear it. undo_retrans is decreased above, L|R frames
1417 * are accounted above as well.
1419 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1420 sacked &= ~TCPCB_SACKED_RETRANS;
1421 tp->retrans_out -= pcount;
1427 /* Shift newly-SACKed bytes from this skb to the immediately previous
1428 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1430 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *prev,
1431 struct sk_buff *skb,
1432 struct tcp_sacktag_state *state,
1433 unsigned int pcount, int shifted, int mss,
1436 struct tcp_sock *tp = tcp_sk(sk);
1437 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1438 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
1442 /* Adjust counters and hints for the newly sacked sequence
1443 * range but discard the return value since prev is already
1444 * marked. We must tag the range first because the seq
1445 * advancement below implicitly advances
1446 * tcp_highest_sack_seq() when skb is highest_sack.
1448 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1449 start_seq, end_seq, dup_sack, pcount,
1450 tcp_skb_timestamp_us(skb));
1451 tcp_rate_skb_delivered(sk, skb, state->rate);
1453 if (skb == tp->lost_skb_hint)
1454 tp->lost_cnt_hint += pcount;
1456 TCP_SKB_CB(prev)->end_seq += shifted;
1457 TCP_SKB_CB(skb)->seq += shifted;
1459 tcp_skb_pcount_add(prev, pcount);
1460 WARN_ON_ONCE(tcp_skb_pcount(skb) < pcount);
1461 tcp_skb_pcount_add(skb, -pcount);
1463 /* When we're adding to gso_segs == 1, gso_size will be zero,
1464 * in theory this shouldn't be necessary but as long as DSACK
1465 * code can come after this skb later on it's better to keep
1466 * setting gso_size to something.
1468 if (!TCP_SKB_CB(prev)->tcp_gso_size)
1469 TCP_SKB_CB(prev)->tcp_gso_size = mss;
1471 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1472 if (tcp_skb_pcount(skb) <= 1)
1473 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1475 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1476 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1479 BUG_ON(!tcp_skb_pcount(skb));
1480 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1484 /* Whole SKB was eaten :-) */
1486 if (skb == tp->retransmit_skb_hint)
1487 tp->retransmit_skb_hint = prev;
1488 if (skb == tp->lost_skb_hint) {
1489 tp->lost_skb_hint = prev;
1490 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1493 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1494 TCP_SKB_CB(prev)->eor = TCP_SKB_CB(skb)->eor;
1495 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1496 TCP_SKB_CB(prev)->end_seq++;
1498 if (skb == tcp_highest_sack(sk))
1499 tcp_advance_highest_sack(sk, skb);
1501 tcp_skb_collapse_tstamp(prev, skb);
1502 if (unlikely(TCP_SKB_CB(prev)->tx.delivered_mstamp))
1503 TCP_SKB_CB(prev)->tx.delivered_mstamp = 0;
1505 tcp_rtx_queue_unlink_and_free(skb, sk);
1507 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKMERGED);
1512 /* I wish gso_size would have a bit more sane initialization than
1513 * something-or-zero which complicates things
1515 static int tcp_skb_seglen(const struct sk_buff *skb)
1517 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1520 /* Shifting pages past head area doesn't work */
1521 static int skb_can_shift(const struct sk_buff *skb)
1523 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1526 int tcp_skb_shift(struct sk_buff *to, struct sk_buff *from,
1527 int pcount, int shiftlen)
1529 /* TCP min gso_size is 8 bytes (TCP_MIN_GSO_SIZE)
1530 * Since TCP_SKB_CB(skb)->tcp_gso_segs is 16 bits, we need
1531 * to make sure not storing more than 65535 * 8 bytes per skb,
1532 * even if current MSS is bigger.
1534 if (unlikely(to->len + shiftlen >= 65535 * TCP_MIN_GSO_SIZE))
1536 if (unlikely(tcp_skb_pcount(to) + pcount > 65535))
1538 return skb_shift(to, from, shiftlen);
1541 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1544 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1545 struct tcp_sacktag_state *state,
1546 u32 start_seq, u32 end_seq,
1549 struct tcp_sock *tp = tcp_sk(sk);
1550 struct sk_buff *prev;
1556 /* Normally R but no L won't result in plain S */
1558 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1560 if (!skb_can_shift(skb))
1562 /* This frame is about to be dropped (was ACKed). */
1563 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1566 /* Can only happen with delayed DSACK + discard craziness */
1567 prev = skb_rb_prev(skb);
1571 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1574 if (!tcp_skb_can_collapse(prev, skb))
1577 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1578 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1582 pcount = tcp_skb_pcount(skb);
1583 mss = tcp_skb_seglen(skb);
1585 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1586 * drop this restriction as unnecessary
1588 if (mss != tcp_skb_seglen(prev))
1591 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1593 /* CHECKME: This is non-MSS split case only?, this will
1594 * cause skipped skbs due to advancing loop btw, original
1595 * has that feature too
1597 if (tcp_skb_pcount(skb) <= 1)
1600 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1602 /* TODO: head merge to next could be attempted here
1603 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1604 * though it might not be worth of the additional hassle
1606 * ...we can probably just fallback to what was done
1607 * previously. We could try merging non-SACKed ones
1608 * as well but it probably isn't going to buy off
1609 * because later SACKs might again split them, and
1610 * it would make skb timestamp tracking considerably
1616 len = end_seq - TCP_SKB_CB(skb)->seq;
1618 BUG_ON(len > skb->len);
1620 /* MSS boundaries should be honoured or else pcount will
1621 * severely break even though it makes things bit trickier.
1622 * Optimize common case to avoid most of the divides
1624 mss = tcp_skb_mss(skb);
1626 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1627 * drop this restriction as unnecessary
1629 if (mss != tcp_skb_seglen(prev))
1634 } else if (len < mss) {
1642 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1643 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1646 if (!tcp_skb_shift(prev, skb, pcount, len))
1648 if (!tcp_shifted_skb(sk, prev, skb, state, pcount, len, mss, dup_sack))
1651 /* Hole filled allows collapsing with the next as well, this is very
1652 * useful when hole on every nth skb pattern happens
1654 skb = skb_rb_next(prev);
1658 if (!skb_can_shift(skb) ||
1659 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1660 (mss != tcp_skb_seglen(skb)))
1663 if (!tcp_skb_can_collapse(prev, skb))
1666 pcount = tcp_skb_pcount(skb);
1667 if (tcp_skb_shift(prev, skb, pcount, len))
1668 tcp_shifted_skb(sk, prev, skb, state, pcount,
1678 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1682 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1683 struct tcp_sack_block *next_dup,
1684 struct tcp_sacktag_state *state,
1685 u32 start_seq, u32 end_seq,
1688 struct tcp_sock *tp = tcp_sk(sk);
1689 struct sk_buff *tmp;
1691 skb_rbtree_walk_from(skb) {
1693 bool dup_sack = dup_sack_in;
1695 /* queue is in-order => we can short-circuit the walk early */
1696 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1700 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1701 in_sack = tcp_match_skb_to_sack(sk, skb,
1702 next_dup->start_seq,
1708 /* skb reference here is a bit tricky to get right, since
1709 * shifting can eat and free both this skb and the next,
1710 * so not even _safe variant of the loop is enough.
1713 tmp = tcp_shift_skb_data(sk, skb, state,
1714 start_seq, end_seq, dup_sack);
1723 in_sack = tcp_match_skb_to_sack(sk, skb,
1729 if (unlikely(in_sack < 0))
1733 TCP_SKB_CB(skb)->sacked =
1736 TCP_SKB_CB(skb)->sacked,
1737 TCP_SKB_CB(skb)->seq,
1738 TCP_SKB_CB(skb)->end_seq,
1740 tcp_skb_pcount(skb),
1741 tcp_skb_timestamp_us(skb));
1742 tcp_rate_skb_delivered(sk, skb, state->rate);
1743 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1744 list_del_init(&skb->tcp_tsorted_anchor);
1746 if (!before(TCP_SKB_CB(skb)->seq,
1747 tcp_highest_sack_seq(tp)))
1748 tcp_advance_highest_sack(sk, skb);
1754 static struct sk_buff *tcp_sacktag_bsearch(struct sock *sk, u32 seq)
1756 struct rb_node *parent, **p = &sk->tcp_rtx_queue.rb_node;
1757 struct sk_buff *skb;
1761 skb = rb_to_skb(parent);
1762 if (before(seq, TCP_SKB_CB(skb)->seq)) {
1763 p = &parent->rb_left;
1766 if (!before(seq, TCP_SKB_CB(skb)->end_seq)) {
1767 p = &parent->rb_right;
1775 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1778 if (skb && after(TCP_SKB_CB(skb)->seq, skip_to_seq))
1781 return tcp_sacktag_bsearch(sk, skip_to_seq);
1784 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1786 struct tcp_sack_block *next_dup,
1787 struct tcp_sacktag_state *state,
1793 if (before(next_dup->start_seq, skip_to_seq)) {
1794 skb = tcp_sacktag_skip(skb, sk, next_dup->start_seq);
1795 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1796 next_dup->start_seq, next_dup->end_seq,
1803 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1805 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1809 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1810 u32 prior_snd_una, struct tcp_sacktag_state *state)
1812 struct tcp_sock *tp = tcp_sk(sk);
1813 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1814 TCP_SKB_CB(ack_skb)->sacked);
1815 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1816 struct tcp_sack_block sp[TCP_NUM_SACKS];
1817 struct tcp_sack_block *cache;
1818 struct sk_buff *skb;
1819 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1821 bool found_dup_sack = false;
1823 int first_sack_index;
1826 state->reord = tp->snd_nxt;
1828 if (!tp->sacked_out)
1829 tcp_highest_sack_reset(sk);
1831 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1832 num_sacks, prior_snd_una, state);
1834 /* Eliminate too old ACKs, but take into
1835 * account more or less fresh ones, they can
1836 * contain valid SACK info.
1838 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1841 if (!tp->packets_out)
1845 first_sack_index = 0;
1846 for (i = 0; i < num_sacks; i++) {
1847 bool dup_sack = !i && found_dup_sack;
1849 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1850 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1852 if (!tcp_is_sackblock_valid(tp, dup_sack,
1853 sp[used_sacks].start_seq,
1854 sp[used_sacks].end_seq)) {
1858 if (!tp->undo_marker)
1859 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1861 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1863 /* Don't count olds caused by ACK reordering */
1864 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1865 !after(sp[used_sacks].end_seq, tp->snd_una))
1867 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1870 NET_INC_STATS(sock_net(sk), mib_idx);
1872 first_sack_index = -1;
1876 /* Ignore very old stuff early */
1877 if (!after(sp[used_sacks].end_seq, prior_snd_una)) {
1879 first_sack_index = -1;
1886 /* order SACK blocks to allow in order walk of the retrans queue */
1887 for (i = used_sacks - 1; i > 0; i--) {
1888 for (j = 0; j < i; j++) {
1889 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1890 swap(sp[j], sp[j + 1]);
1892 /* Track where the first SACK block goes to */
1893 if (j == first_sack_index)
1894 first_sack_index = j + 1;
1899 state->mss_now = tcp_current_mss(sk);
1903 if (!tp->sacked_out) {
1904 /* It's already past, so skip checking against it */
1905 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1907 cache = tp->recv_sack_cache;
1908 /* Skip empty blocks in at head of the cache */
1909 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1914 while (i < used_sacks) {
1915 u32 start_seq = sp[i].start_seq;
1916 u32 end_seq = sp[i].end_seq;
1917 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1918 struct tcp_sack_block *next_dup = NULL;
1920 if (found_dup_sack && ((i + 1) == first_sack_index))
1921 next_dup = &sp[i + 1];
1923 /* Skip too early cached blocks */
1924 while (tcp_sack_cache_ok(tp, cache) &&
1925 !before(start_seq, cache->end_seq))
1928 /* Can skip some work by looking recv_sack_cache? */
1929 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1930 after(end_seq, cache->start_seq)) {
1933 if (before(start_seq, cache->start_seq)) {
1934 skb = tcp_sacktag_skip(skb, sk, start_seq);
1935 skb = tcp_sacktag_walk(skb, sk, next_dup,
1942 /* Rest of the block already fully processed? */
1943 if (!after(end_seq, cache->end_seq))
1946 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1950 /* ...tail remains todo... */
1951 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1952 /* ...but better entrypoint exists! */
1953 skb = tcp_highest_sack(sk);
1960 skb = tcp_sacktag_skip(skb, sk, cache->end_seq);
1961 /* Check overlap against next cached too (past this one already) */
1966 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1967 skb = tcp_highest_sack(sk);
1971 skb = tcp_sacktag_skip(skb, sk, start_seq);
1974 skb = tcp_sacktag_walk(skb, sk, next_dup, state,
1975 start_seq, end_seq, dup_sack);
1981 /* Clear the head of the cache sack blocks so we can skip it next time */
1982 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1983 tp->recv_sack_cache[i].start_seq = 0;
1984 tp->recv_sack_cache[i].end_seq = 0;
1986 for (j = 0; j < used_sacks; j++)
1987 tp->recv_sack_cache[i++] = sp[j];
1989 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Loss || tp->undo_marker)
1990 tcp_check_sack_reordering(sk, state->reord, 0);
1992 tcp_verify_left_out(tp);
1995 #if FASTRETRANS_DEBUG > 0
1996 WARN_ON((int)tp->sacked_out < 0);
1997 WARN_ON((int)tp->lost_out < 0);
1998 WARN_ON((int)tp->retrans_out < 0);
1999 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
2004 /* Limits sacked_out so that sum with lost_out isn't ever larger than
2005 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
2007 static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
2011 holes = max(tp->lost_out, 1U);
2012 holes = min(holes, tp->packets_out);
2014 if ((tp->sacked_out + holes) > tp->packets_out) {
2015 tp->sacked_out = tp->packets_out - holes;
2021 /* If we receive more dupacks than we expected counting segments
2022 * in assumption of absent reordering, interpret this as reordering.
2023 * The only another reason could be bug in receiver TCP.
2025 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
2027 struct tcp_sock *tp = tcp_sk(sk);
2029 if (!tcp_limit_reno_sacked(tp))
2032 tp->reordering = min_t(u32, tp->packets_out + addend,
2033 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_max_reordering));
2035 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRENOREORDER);
2038 /* Emulate SACKs for SACKless connection: account for a new dupack. */
2040 static void tcp_add_reno_sack(struct sock *sk, int num_dupack, bool ece_ack)
2043 struct tcp_sock *tp = tcp_sk(sk);
2044 u32 prior_sacked = tp->sacked_out;
2047 tp->sacked_out += num_dupack;
2048 tcp_check_reno_reordering(sk, 0);
2049 delivered = tp->sacked_out - prior_sacked;
2051 tcp_count_delivered(tp, delivered, ece_ack);
2052 tcp_verify_left_out(tp);
2056 /* Account for ACK, ACKing some data in Reno Recovery phase. */
2058 static void tcp_remove_reno_sacks(struct sock *sk, int acked, bool ece_ack)
2060 struct tcp_sock *tp = tcp_sk(sk);
2063 /* One ACK acked hole. The rest eat duplicate ACKs. */
2064 tcp_count_delivered(tp, max_t(int, acked - tp->sacked_out, 1),
2066 if (acked - 1 >= tp->sacked_out)
2069 tp->sacked_out -= acked - 1;
2071 tcp_check_reno_reordering(sk, acked);
2072 tcp_verify_left_out(tp);
2075 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
2080 void tcp_clear_retrans(struct tcp_sock *tp)
2082 tp->retrans_out = 0;
2084 tp->undo_marker = 0;
2085 tp->undo_retrans = -1;
2089 static inline void tcp_init_undo(struct tcp_sock *tp)
2091 tp->undo_marker = tp->snd_una;
2092 /* Retransmission still in flight may cause DSACKs later. */
2093 tp->undo_retrans = tp->retrans_out ? : -1;
2096 static bool tcp_is_rack(const struct sock *sk)
2098 return READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_recovery) &
2099 TCP_RACK_LOSS_DETECTION;
2102 /* If we detect SACK reneging, forget all SACK information
2103 * and reset tags completely, otherwise preserve SACKs. If receiver
2104 * dropped its ofo queue, we will know this due to reneging detection.
2106 static void tcp_timeout_mark_lost(struct sock *sk)
2108 struct tcp_sock *tp = tcp_sk(sk);
2109 struct sk_buff *skb, *head;
2110 bool is_reneg; /* is receiver reneging on SACKs? */
2112 head = tcp_rtx_queue_head(sk);
2113 is_reneg = head && (TCP_SKB_CB(head)->sacked & TCPCB_SACKED_ACKED);
2115 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
2117 /* Mark SACK reneging until we recover from this loss event. */
2118 tp->is_sack_reneg = 1;
2119 } else if (tcp_is_reno(tp)) {
2120 tcp_reset_reno_sack(tp);
2124 skb_rbtree_walk_from(skb) {
2126 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
2127 else if (tcp_is_rack(sk) && skb != head &&
2128 tcp_rack_skb_timeout(tp, skb, 0) > 0)
2129 continue; /* Don't mark recently sent ones lost yet */
2130 tcp_mark_skb_lost(sk, skb);
2132 tcp_verify_left_out(tp);
2133 tcp_clear_all_retrans_hints(tp);
2136 /* Enter Loss state. */
2137 void tcp_enter_loss(struct sock *sk)
2139 const struct inet_connection_sock *icsk = inet_csk(sk);
2140 struct tcp_sock *tp = tcp_sk(sk);
2141 struct net *net = sock_net(sk);
2142 bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery;
2145 tcp_timeout_mark_lost(sk);
2147 /* Reduce ssthresh if it has not yet been made inside this window. */
2148 if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
2149 !after(tp->high_seq, tp->snd_una) ||
2150 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
2151 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2152 tp->prior_cwnd = tcp_snd_cwnd(tp);
2153 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2154 tcp_ca_event(sk, CA_EVENT_LOSS);
2157 tcp_snd_cwnd_set(tp, tcp_packets_in_flight(tp) + 1);
2158 tp->snd_cwnd_cnt = 0;
2159 tp->snd_cwnd_stamp = tcp_jiffies32;
2161 /* Timeout in disordered state after receiving substantial DUPACKs
2162 * suggests that the degree of reordering is over-estimated.
2164 reordering = READ_ONCE(net->ipv4.sysctl_tcp_reordering);
2165 if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
2166 tp->sacked_out >= reordering)
2167 tp->reordering = min_t(unsigned int, tp->reordering,
2170 tcp_set_ca_state(sk, TCP_CA_Loss);
2171 tp->high_seq = tp->snd_nxt;
2172 tcp_ecn_queue_cwr(tp);
2174 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
2175 * loss recovery is underway except recurring timeout(s) on
2176 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
2178 tp->frto = READ_ONCE(net->ipv4.sysctl_tcp_frto) &&
2179 (new_recovery || icsk->icsk_retransmits) &&
2180 !inet_csk(sk)->icsk_mtup.probe_size;
2183 /* If ACK arrived pointing to a remembered SACK, it means that our
2184 * remembered SACKs do not reflect real state of receiver i.e.
2185 * receiver _host_ is heavily congested (or buggy).
2187 * To avoid big spurious retransmission bursts due to transient SACK
2188 * scoreboard oddities that look like reneging, we give the receiver a
2189 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
2190 * restore sanity to the SACK scoreboard. If the apparent reneging
2191 * persists until this RTO then we'll clear the SACK scoreboard.
2193 static bool tcp_check_sack_reneging(struct sock *sk, int flag)
2195 if (flag & FLAG_SACK_RENEGING &&
2196 flag & FLAG_SND_UNA_ADVANCED) {
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 static bool tcp_is_non_sack_preventing_reopen(struct sock *sk)
2519 struct tcp_sock *tp = tcp_sk(sk);
2521 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2522 /* Hold old state until something *above* high_seq
2523 * is ACKed. For Reno it is MUST to prevent false
2524 * fast retransmits (RFC2582). SACK TCP is safe. */
2525 if (!tcp_any_retrans_done(sk))
2526 tp->retrans_stamp = 0;
2532 /* People celebrate: "We love our President!" */
2533 static bool tcp_try_undo_recovery(struct sock *sk)
2535 struct tcp_sock *tp = tcp_sk(sk);
2537 if (tcp_may_undo(tp)) {
2540 /* Happy end! We did not retransmit anything
2541 * or our original transmission succeeded.
2543 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2544 tcp_undo_cwnd_reduction(sk, false);
2545 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2546 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2548 mib_idx = LINUX_MIB_TCPFULLUNDO;
2550 NET_INC_STATS(sock_net(sk), mib_idx);
2551 } else if (tp->rack.reo_wnd_persist) {
2552 tp->rack.reo_wnd_persist--;
2554 if (tcp_is_non_sack_preventing_reopen(sk))
2556 tcp_set_ca_state(sk, TCP_CA_Open);
2557 tp->is_sack_reneg = 0;
2561 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2562 static bool tcp_try_undo_dsack(struct sock *sk)
2564 struct tcp_sock *tp = tcp_sk(sk);
2566 if (tp->undo_marker && !tp->undo_retrans) {
2567 tp->rack.reo_wnd_persist = min(TCP_RACK_RECOVERY_THRESH,
2568 tp->rack.reo_wnd_persist + 1);
2569 DBGUNDO(sk, "D-SACK");
2570 tcp_undo_cwnd_reduction(sk, false);
2571 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2577 /* Undo during loss recovery after partial ACK or using F-RTO. */
2578 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2580 struct tcp_sock *tp = tcp_sk(sk);
2582 if (frto_undo || tcp_may_undo(tp)) {
2583 tcp_undo_cwnd_reduction(sk, true);
2585 DBGUNDO(sk, "partial loss");
2586 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2588 NET_INC_STATS(sock_net(sk),
2589 LINUX_MIB_TCPSPURIOUSRTOS);
2590 inet_csk(sk)->icsk_retransmits = 0;
2591 if (tcp_is_non_sack_preventing_reopen(sk))
2593 if (frto_undo || tcp_is_sack(tp)) {
2594 tcp_set_ca_state(sk, TCP_CA_Open);
2595 tp->is_sack_reneg = 0;
2602 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
2603 * It computes the number of packets to send (sndcnt) based on packets newly
2605 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2606 * cwnd reductions across a full RTT.
2607 * 2) Otherwise PRR uses packet conservation to send as much as delivered.
2608 * But when SND_UNA is acked without further losses,
2609 * slow starts cwnd up to ssthresh to speed up the recovery.
2611 static void tcp_init_cwnd_reduction(struct sock *sk)
2613 struct tcp_sock *tp = tcp_sk(sk);
2615 tp->high_seq = tp->snd_nxt;
2616 tp->tlp_high_seq = 0;
2617 tp->snd_cwnd_cnt = 0;
2618 tp->prior_cwnd = tcp_snd_cwnd(tp);
2619 tp->prr_delivered = 0;
2621 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2622 tcp_ecn_queue_cwr(tp);
2625 void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked, int newly_lost, int flag)
2627 struct tcp_sock *tp = tcp_sk(sk);
2629 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2631 if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd))
2634 tp->prr_delivered += newly_acked_sacked;
2636 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2638 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2640 sndcnt = max_t(int, tp->prr_delivered - tp->prr_out,
2641 newly_acked_sacked);
2642 if (flag & FLAG_SND_UNA_ADVANCED && !newly_lost)
2644 sndcnt = min(delta, sndcnt);
2646 /* Force a fast retransmit upon entering fast recovery */
2647 sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1));
2648 tcp_snd_cwnd_set(tp, tcp_packets_in_flight(tp) + sndcnt);
2651 static inline void tcp_end_cwnd_reduction(struct sock *sk)
2653 struct tcp_sock *tp = tcp_sk(sk);
2655 if (inet_csk(sk)->icsk_ca_ops->cong_control)
2658 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2659 if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH &&
2660 (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR || tp->undo_marker)) {
2661 tcp_snd_cwnd_set(tp, tp->snd_ssthresh);
2662 tp->snd_cwnd_stamp = tcp_jiffies32;
2664 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2667 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2668 void tcp_enter_cwr(struct sock *sk)
2670 struct tcp_sock *tp = tcp_sk(sk);
2672 tp->prior_ssthresh = 0;
2673 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2674 tp->undo_marker = 0;
2675 tcp_init_cwnd_reduction(sk);
2676 tcp_set_ca_state(sk, TCP_CA_CWR);
2679 EXPORT_SYMBOL(tcp_enter_cwr);
2681 static void tcp_try_keep_open(struct sock *sk)
2683 struct tcp_sock *tp = tcp_sk(sk);
2684 int state = TCP_CA_Open;
2686 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2687 state = TCP_CA_Disorder;
2689 if (inet_csk(sk)->icsk_ca_state != state) {
2690 tcp_set_ca_state(sk, state);
2691 tp->high_seq = tp->snd_nxt;
2695 static void tcp_try_to_open(struct sock *sk, int flag)
2697 struct tcp_sock *tp = tcp_sk(sk);
2699 tcp_verify_left_out(tp);
2701 if (!tcp_any_retrans_done(sk))
2702 tp->retrans_stamp = 0;
2704 if (flag & FLAG_ECE)
2707 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2708 tcp_try_keep_open(sk);
2712 static void tcp_mtup_probe_failed(struct sock *sk)
2714 struct inet_connection_sock *icsk = inet_csk(sk);
2716 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2717 icsk->icsk_mtup.probe_size = 0;
2718 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPFAIL);
2721 static void tcp_mtup_probe_success(struct sock *sk)
2723 struct tcp_sock *tp = tcp_sk(sk);
2724 struct inet_connection_sock *icsk = inet_csk(sk);
2727 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2729 val = (u64)tcp_snd_cwnd(tp) * tcp_mss_to_mtu(sk, tp->mss_cache);
2730 do_div(val, icsk->icsk_mtup.probe_size);
2731 DEBUG_NET_WARN_ON_ONCE((u32)val != val);
2732 tcp_snd_cwnd_set(tp, max_t(u32, 1U, val));
2734 tp->snd_cwnd_cnt = 0;
2735 tp->snd_cwnd_stamp = tcp_jiffies32;
2736 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2738 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2739 icsk->icsk_mtup.probe_size = 0;
2740 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2741 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS);
2744 /* Do a simple retransmit without using the backoff mechanisms in
2745 * tcp_timer. This is used for path mtu discovery.
2746 * The socket is already locked here.
2748 void tcp_simple_retransmit(struct sock *sk)
2750 const struct inet_connection_sock *icsk = inet_csk(sk);
2751 struct tcp_sock *tp = tcp_sk(sk);
2752 struct sk_buff *skb;
2755 /* A fastopen SYN request is stored as two separate packets within
2756 * the retransmit queue, this is done by tcp_send_syn_data().
2757 * As a result simply checking the MSS of the frames in the queue
2758 * will not work for the SYN packet.
2760 * Us being here is an indication of a path MTU issue so we can
2761 * assume that the fastopen SYN was lost and just mark all the
2762 * frames in the retransmit queue as lost. We will use an MSS of
2763 * -1 to mark all frames as lost, otherwise compute the current MSS.
2765 if (tp->syn_data && sk->sk_state == TCP_SYN_SENT)
2768 mss = tcp_current_mss(sk);
2770 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2771 if (tcp_skb_seglen(skb) > mss)
2772 tcp_mark_skb_lost(sk, skb);
2775 tcp_clear_retrans_hints_partial(tp);
2780 if (tcp_is_reno(tp))
2781 tcp_limit_reno_sacked(tp);
2783 tcp_verify_left_out(tp);
2785 /* Don't muck with the congestion window here.
2786 * Reason is that we do not increase amount of _data_
2787 * in network, but units changed and effective
2788 * cwnd/ssthresh really reduced now.
2790 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2791 tp->high_seq = tp->snd_nxt;
2792 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2793 tp->prior_ssthresh = 0;
2794 tp->undo_marker = 0;
2795 tcp_set_ca_state(sk, TCP_CA_Loss);
2797 tcp_xmit_retransmit_queue(sk);
2799 EXPORT_SYMBOL(tcp_simple_retransmit);
2801 void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2803 struct tcp_sock *tp = tcp_sk(sk);
2806 if (tcp_is_reno(tp))
2807 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2809 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2811 NET_INC_STATS(sock_net(sk), mib_idx);
2813 tp->prior_ssthresh = 0;
2816 if (!tcp_in_cwnd_reduction(sk)) {
2818 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2819 tcp_init_cwnd_reduction(sk);
2821 tcp_set_ca_state(sk, TCP_CA_Recovery);
2824 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2825 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2827 static void tcp_process_loss(struct sock *sk, int flag, int num_dupack,
2830 struct tcp_sock *tp = tcp_sk(sk);
2831 bool recovered = !before(tp->snd_una, tp->high_seq);
2833 if ((flag & FLAG_SND_UNA_ADVANCED || rcu_access_pointer(tp->fastopen_rsk)) &&
2834 tcp_try_undo_loss(sk, false))
2837 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2838 /* Step 3.b. A timeout is spurious if not all data are
2839 * lost, i.e., never-retransmitted data are (s)acked.
2841 if ((flag & FLAG_ORIG_SACK_ACKED) &&
2842 tcp_try_undo_loss(sk, true))
2845 if (after(tp->snd_nxt, tp->high_seq)) {
2846 if (flag & FLAG_DATA_SACKED || num_dupack)
2847 tp->frto = 0; /* Step 3.a. loss was real */
2848 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2849 tp->high_seq = tp->snd_nxt;
2850 /* Step 2.b. Try send new data (but deferred until cwnd
2851 * is updated in tcp_ack()). Otherwise fall back to
2852 * the conventional recovery.
2854 if (!tcp_write_queue_empty(sk) &&
2855 after(tcp_wnd_end(tp), tp->snd_nxt)) {
2856 *rexmit = REXMIT_NEW;
2864 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2865 tcp_try_undo_recovery(sk);
2868 if (tcp_is_reno(tp)) {
2869 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2870 * delivered. Lower inflight to clock out (re)tranmissions.
2872 if (after(tp->snd_nxt, tp->high_seq) && num_dupack)
2873 tcp_add_reno_sack(sk, num_dupack, flag & FLAG_ECE);
2874 else if (flag & FLAG_SND_UNA_ADVANCED)
2875 tcp_reset_reno_sack(tp);
2877 *rexmit = REXMIT_LOST;
2880 static bool tcp_force_fast_retransmit(struct sock *sk)
2882 struct tcp_sock *tp = tcp_sk(sk);
2884 return after(tcp_highest_sack_seq(tp),
2885 tp->snd_una + tp->reordering * tp->mss_cache);
2888 /* Undo during fast recovery after partial ACK. */
2889 static bool tcp_try_undo_partial(struct sock *sk, u32 prior_snd_una,
2892 struct tcp_sock *tp = tcp_sk(sk);
2894 if (tp->undo_marker && tcp_packet_delayed(tp)) {
2895 /* Plain luck! Hole if filled with delayed
2896 * packet, rather than with a retransmit. Check reordering.
2898 tcp_check_sack_reordering(sk, prior_snd_una, 1);
2900 /* We are getting evidence that the reordering degree is higher
2901 * than we realized. If there are no retransmits out then we
2902 * can undo. Otherwise we clock out new packets but do not
2903 * mark more packets lost or retransmit more.
2905 if (tp->retrans_out)
2908 if (!tcp_any_retrans_done(sk))
2909 tp->retrans_stamp = 0;
2911 DBGUNDO(sk, "partial recovery");
2912 tcp_undo_cwnd_reduction(sk, true);
2913 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2914 tcp_try_keep_open(sk);
2916 /* Partial ACK arrived. Force fast retransmit. */
2917 *do_lost = tcp_force_fast_retransmit(sk);
2922 static void tcp_identify_packet_loss(struct sock *sk, int *ack_flag)
2924 struct tcp_sock *tp = tcp_sk(sk);
2926 if (tcp_rtx_queue_empty(sk))
2929 if (unlikely(tcp_is_reno(tp))) {
2930 tcp_newreno_mark_lost(sk, *ack_flag & FLAG_SND_UNA_ADVANCED);
2931 } else if (tcp_is_rack(sk)) {
2932 u32 prior_retrans = tp->retrans_out;
2934 if (tcp_rack_mark_lost(sk))
2935 *ack_flag &= ~FLAG_SET_XMIT_TIMER;
2936 if (prior_retrans > tp->retrans_out)
2937 *ack_flag |= FLAG_LOST_RETRANS;
2941 /* Process an event, which can update packets-in-flight not trivially.
2942 * Main goal of this function is to calculate new estimate for left_out,
2943 * taking into account both packets sitting in receiver's buffer and
2944 * packets lost by network.
2946 * Besides that it updates the congestion state when packet loss or ECN
2947 * is detected. But it does not reduce the cwnd, it is done by the
2948 * congestion control later.
2950 * It does _not_ decide what to send, it is made in function
2951 * tcp_xmit_retransmit_queue().
2953 static void tcp_fastretrans_alert(struct sock *sk, const u32 prior_snd_una,
2954 int num_dupack, int *ack_flag, int *rexmit)
2956 struct inet_connection_sock *icsk = inet_csk(sk);
2957 struct tcp_sock *tp = tcp_sk(sk);
2958 int fast_rexmit = 0, flag = *ack_flag;
2959 bool ece_ack = flag & FLAG_ECE;
2960 bool do_lost = num_dupack || ((flag & FLAG_DATA_SACKED) &&
2961 tcp_force_fast_retransmit(sk));
2963 if (!tp->packets_out && tp->sacked_out)
2966 /* Now state machine starts.
2967 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2969 tp->prior_ssthresh = 0;
2971 /* B. In all the states check for reneging SACKs. */
2972 if (tcp_check_sack_reneging(sk, flag))
2975 /* C. Check consistency of the current state. */
2976 tcp_verify_left_out(tp);
2978 /* D. Check state exit conditions. State can be terminated
2979 * when high_seq is ACKed. */
2980 if (icsk->icsk_ca_state == TCP_CA_Open) {
2981 WARN_ON(tp->retrans_out != 0 && !tp->syn_data);
2982 tp->retrans_stamp = 0;
2983 } else if (!before(tp->snd_una, tp->high_seq)) {
2984 switch (icsk->icsk_ca_state) {
2986 /* CWR is to be held something *above* high_seq
2987 * is ACKed for CWR bit to reach receiver. */
2988 if (tp->snd_una != tp->high_seq) {
2989 tcp_end_cwnd_reduction(sk);
2990 tcp_set_ca_state(sk, TCP_CA_Open);
2994 case TCP_CA_Recovery:
2995 if (tcp_is_reno(tp))
2996 tcp_reset_reno_sack(tp);
2997 if (tcp_try_undo_recovery(sk))
2999 tcp_end_cwnd_reduction(sk);
3004 /* E. Process state. */
3005 switch (icsk->icsk_ca_state) {
3006 case TCP_CA_Recovery:
3007 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
3008 if (tcp_is_reno(tp))
3009 tcp_add_reno_sack(sk, num_dupack, ece_ack);
3010 } else if (tcp_try_undo_partial(sk, prior_snd_una, &do_lost))
3013 if (tcp_try_undo_dsack(sk))
3014 tcp_try_keep_open(sk);
3016 tcp_identify_packet_loss(sk, ack_flag);
3017 if (icsk->icsk_ca_state != TCP_CA_Recovery) {
3018 if (!tcp_time_to_recover(sk, flag))
3020 /* Undo reverts the recovery state. If loss is evident,
3021 * starts a new recovery (e.g. reordering then loss);
3023 tcp_enter_recovery(sk, ece_ack);
3027 tcp_process_loss(sk, flag, num_dupack, rexmit);
3028 tcp_identify_packet_loss(sk, ack_flag);
3029 if (!(icsk->icsk_ca_state == TCP_CA_Open ||
3030 (*ack_flag & FLAG_LOST_RETRANS)))
3032 /* Change state if cwnd is undone or retransmits are lost */
3035 if (tcp_is_reno(tp)) {
3036 if (flag & FLAG_SND_UNA_ADVANCED)
3037 tcp_reset_reno_sack(tp);
3038 tcp_add_reno_sack(sk, num_dupack, ece_ack);
3041 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
3042 tcp_try_undo_dsack(sk);
3044 tcp_identify_packet_loss(sk, ack_flag);
3045 if (!tcp_time_to_recover(sk, flag)) {
3046 tcp_try_to_open(sk, flag);
3050 /* MTU probe failure: don't reduce cwnd */
3051 if (icsk->icsk_ca_state < TCP_CA_CWR &&
3052 icsk->icsk_mtup.probe_size &&
3053 tp->snd_una == tp->mtu_probe.probe_seq_start) {
3054 tcp_mtup_probe_failed(sk);
3055 /* Restores the reduction we did in tcp_mtup_probe() */
3056 tcp_snd_cwnd_set(tp, tcp_snd_cwnd(tp) + 1);
3057 tcp_simple_retransmit(sk);
3061 /* Otherwise enter Recovery state */
3062 tcp_enter_recovery(sk, ece_ack);
3066 if (!tcp_is_rack(sk) && do_lost)
3067 tcp_update_scoreboard(sk, fast_rexmit);
3068 *rexmit = REXMIT_LOST;
3071 static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us, const int flag)
3073 u32 wlen = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_rtt_wlen) * HZ;
3074 struct tcp_sock *tp = tcp_sk(sk);
3076 if ((flag & FLAG_ACK_MAYBE_DELAYED) && rtt_us > tcp_min_rtt(tp)) {
3077 /* If the remote keeps returning delayed ACKs, eventually
3078 * the min filter would pick it up and overestimate the
3079 * prop. delay when it expires. Skip suspected delayed ACKs.
3083 minmax_running_min(&tp->rtt_min, wlen, tcp_jiffies32,
3084 rtt_us ? : jiffies_to_usecs(1));
3087 static bool tcp_ack_update_rtt(struct sock *sk, const int flag,
3088 long seq_rtt_us, long sack_rtt_us,
3089 long ca_rtt_us, struct rate_sample *rs)
3091 const struct tcp_sock *tp = tcp_sk(sk);
3093 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
3094 * broken middle-boxes or peers may corrupt TS-ECR fields. But
3095 * Karn's algorithm forbids taking RTT if some retransmitted data
3096 * is acked (RFC6298).
3099 seq_rtt_us = sack_rtt_us;
3101 /* RTTM Rule: A TSecr value received in a segment is used to
3102 * update the averaged RTT measurement only if the segment
3103 * acknowledges some new data, i.e., only if it advances the
3104 * left edge of the send window.
3105 * See draft-ietf-tcplw-high-performance-00, section 3.3.
3107 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
3108 flag & FLAG_ACKED) {
3109 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr;
3111 if (likely(delta < INT_MAX / (USEC_PER_SEC / TCP_TS_HZ))) {
3114 seq_rtt_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
3115 ca_rtt_us = seq_rtt_us;
3118 rs->rtt_us = ca_rtt_us; /* RTT of last (S)ACKed packet (or -1) */
3122 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
3123 * always taken together with ACK, SACK, or TS-opts. Any negative
3124 * values will be skipped with the seq_rtt_us < 0 check above.
3126 tcp_update_rtt_min(sk, ca_rtt_us, flag);
3127 tcp_rtt_estimator(sk, seq_rtt_us);
3130 /* RFC6298: only reset backoff on valid RTT measurement. */
3131 inet_csk(sk)->icsk_backoff = 0;
3135 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
3136 void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req)
3138 struct rate_sample rs;
3141 if (req && !req->num_retrans && tcp_rsk(req)->snt_synack)
3142 rtt_us = tcp_stamp_us_delta(tcp_clock_us(), tcp_rsk(req)->snt_synack);
3144 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us, &rs);
3148 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
3150 const struct inet_connection_sock *icsk = inet_csk(sk);
3152 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
3153 tcp_sk(sk)->snd_cwnd_stamp = tcp_jiffies32;
3156 /* Restart timer after forward progress on connection.
3157 * RFC2988 recommends to restart timer to now+rto.
3159 void tcp_rearm_rto(struct sock *sk)
3161 const struct inet_connection_sock *icsk = inet_csk(sk);
3162 struct tcp_sock *tp = tcp_sk(sk);
3164 /* If the retrans timer is currently being used by Fast Open
3165 * for SYN-ACK retrans purpose, stay put.
3167 if (rcu_access_pointer(tp->fastopen_rsk))
3170 if (!tp->packets_out) {
3171 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
3173 u32 rto = inet_csk(sk)->icsk_rto;
3174 /* Offset the time elapsed after installing regular RTO */
3175 if (icsk->icsk_pending == ICSK_TIME_REO_TIMEOUT ||
3176 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
3177 s64 delta_us = tcp_rto_delta_us(sk);
3178 /* delta_us may not be positive if the socket is locked
3179 * when the retrans timer fires and is rescheduled.
3181 rto = usecs_to_jiffies(max_t(int, delta_us, 1));
3183 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
3188 /* Try to schedule a loss probe; if that doesn't work, then schedule an RTO. */
3189 static void tcp_set_xmit_timer(struct sock *sk)
3191 if (!tcp_schedule_loss_probe(sk, true))
3195 /* If we get here, the whole TSO packet has not been acked. */
3196 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3198 struct tcp_sock *tp = tcp_sk(sk);
3201 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3203 packets_acked = tcp_skb_pcount(skb);
3204 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3206 packets_acked -= tcp_skb_pcount(skb);
3208 if (packets_acked) {
3209 BUG_ON(tcp_skb_pcount(skb) == 0);
3210 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3213 return packets_acked;
3216 static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
3217 const struct sk_buff *ack_skb, u32 prior_snd_una)
3219 const struct skb_shared_info *shinfo;
3221 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3222 if (likely(!TCP_SKB_CB(skb)->txstamp_ack))
3225 shinfo = skb_shinfo(skb);
3226 if (!before(shinfo->tskey, prior_snd_una) &&
3227 before(shinfo->tskey, tcp_sk(sk)->snd_una)) {
3228 tcp_skb_tsorted_save(skb) {
3229 __skb_tstamp_tx(skb, ack_skb, NULL, sk, SCM_TSTAMP_ACK);
3230 } tcp_skb_tsorted_restore(skb);
3234 /* Remove acknowledged frames from the retransmission queue. If our packet
3235 * is before the ack sequence we can discard it as it's confirmed to have
3236 * arrived at the other end.
3238 static int tcp_clean_rtx_queue(struct sock *sk, const struct sk_buff *ack_skb,
3239 u32 prior_fack, u32 prior_snd_una,
3240 struct tcp_sacktag_state *sack, bool ece_ack)
3242 const struct inet_connection_sock *icsk = inet_csk(sk);
3243 u64 first_ackt, last_ackt;
3244 struct tcp_sock *tp = tcp_sk(sk);
3245 u32 prior_sacked = tp->sacked_out;
3246 u32 reord = tp->snd_nxt; /* lowest acked un-retx un-sacked seq */
3247 struct sk_buff *skb, *next;
3248 bool fully_acked = true;
3249 long sack_rtt_us = -1L;
3250 long seq_rtt_us = -1L;
3251 long ca_rtt_us = -1L;
3258 for (skb = skb_rb_first(&sk->tcp_rtx_queue); skb; skb = next) {
3259 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3260 const u32 start_seq = scb->seq;
3261 u8 sacked = scb->sacked;
3264 /* Determine how many packets and what bytes were acked, tso and else */
3265 if (after(scb->end_seq, tp->snd_una)) {
3266 if (tcp_skb_pcount(skb) == 1 ||
3267 !after(tp->snd_una, scb->seq))
3270 acked_pcount = tcp_tso_acked(sk, skb);
3273 fully_acked = false;
3275 acked_pcount = tcp_skb_pcount(skb);
3278 if (unlikely(sacked & TCPCB_RETRANS)) {
3279 if (sacked & TCPCB_SACKED_RETRANS)
3280 tp->retrans_out -= acked_pcount;
3281 flag |= FLAG_RETRANS_DATA_ACKED;
3282 } else if (!(sacked & TCPCB_SACKED_ACKED)) {
3283 last_ackt = tcp_skb_timestamp_us(skb);
3284 WARN_ON_ONCE(last_ackt == 0);
3286 first_ackt = last_ackt;
3288 if (before(start_seq, reord))
3290 if (!after(scb->end_seq, tp->high_seq))
3291 flag |= FLAG_ORIG_SACK_ACKED;
3294 if (sacked & TCPCB_SACKED_ACKED) {
3295 tp->sacked_out -= acked_pcount;
3296 } else if (tcp_is_sack(tp)) {
3297 tcp_count_delivered(tp, acked_pcount, ece_ack);
3298 if (!tcp_skb_spurious_retrans(tp, skb))
3299 tcp_rack_advance(tp, sacked, scb->end_seq,
3300 tcp_skb_timestamp_us(skb));
3302 if (sacked & TCPCB_LOST)
3303 tp->lost_out -= acked_pcount;
3305 tp->packets_out -= acked_pcount;
3306 pkts_acked += acked_pcount;
3307 tcp_rate_skb_delivered(sk, skb, sack->rate);
3309 /* Initial outgoing SYN's get put onto the write_queue
3310 * just like anything else we transmit. It is not
3311 * true data, and if we misinform our callers that
3312 * this ACK acks real data, we will erroneously exit
3313 * connection startup slow start one packet too
3314 * quickly. This is severely frowned upon behavior.
3316 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
3317 flag |= FLAG_DATA_ACKED;
3319 flag |= FLAG_SYN_ACKED;
3320 tp->retrans_stamp = 0;
3326 tcp_ack_tstamp(sk, skb, ack_skb, prior_snd_una);
3328 next = skb_rb_next(skb);
3329 if (unlikely(skb == tp->retransmit_skb_hint))
3330 tp->retransmit_skb_hint = NULL;
3331 if (unlikely(skb == tp->lost_skb_hint))
3332 tp->lost_skb_hint = NULL;
3333 tcp_highest_sack_replace(sk, skb, next);
3334 tcp_rtx_queue_unlink_and_free(skb, sk);
3338 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
3340 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3341 tp->snd_up = tp->snd_una;
3344 tcp_ack_tstamp(sk, skb, ack_skb, prior_snd_una);
3345 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
3346 flag |= FLAG_SACK_RENEGING;
3349 if (likely(first_ackt) && !(flag & FLAG_RETRANS_DATA_ACKED)) {
3350 seq_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, first_ackt);
3351 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, last_ackt);
3353 if (pkts_acked == 1 && fully_acked && !prior_sacked &&
3354 (tp->snd_una - prior_snd_una) < tp->mss_cache &&
3355 sack->rate->prior_delivered + 1 == tp->delivered &&
3356 !(flag & (FLAG_CA_ALERT | FLAG_SYN_ACKED))) {
3357 /* Conservatively mark a delayed ACK. It's typically
3358 * from a lone runt packet over the round trip to
3359 * a receiver w/o out-of-order or CE events.
3361 flag |= FLAG_ACK_MAYBE_DELAYED;
3364 if (sack->first_sackt) {
3365 sack_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->first_sackt);
3366 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->last_sackt);
3368 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us,
3369 ca_rtt_us, sack->rate);
3371 if (flag & FLAG_ACKED) {
3372 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3373 if (unlikely(icsk->icsk_mtup.probe_size &&
3374 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3375 tcp_mtup_probe_success(sk);
3378 if (tcp_is_reno(tp)) {
3379 tcp_remove_reno_sacks(sk, pkts_acked, ece_ack);
3381 /* If any of the cumulatively ACKed segments was
3382 * retransmitted, non-SACK case cannot confirm that
3383 * progress was due to original transmission due to
3384 * lack of TCPCB_SACKED_ACKED bits even if some of
3385 * the packets may have been never retransmitted.
3387 if (flag & FLAG_RETRANS_DATA_ACKED)
3388 flag &= ~FLAG_ORIG_SACK_ACKED;
3392 /* Non-retransmitted hole got filled? That's reordering */
3393 if (before(reord, prior_fack))
3394 tcp_check_sack_reordering(sk, reord, 0);
3396 delta = prior_sacked - tp->sacked_out;
3397 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3399 } else if (skb && rtt_update && sack_rtt_us >= 0 &&
3400 sack_rtt_us > tcp_stamp_us_delta(tp->tcp_mstamp,
3401 tcp_skb_timestamp_us(skb))) {
3402 /* Do not re-arm RTO if the sack RTT is measured from data sent
3403 * after when the head was last (re)transmitted. Otherwise the
3404 * timeout may continue to extend in loss recovery.
3406 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3409 if (icsk->icsk_ca_ops->pkts_acked) {
3410 struct ack_sample sample = { .pkts_acked = pkts_acked,
3411 .rtt_us = sack->rate->rtt_us };
3413 sample.in_flight = tp->mss_cache *
3414 (tp->delivered - sack->rate->prior_delivered);
3415 icsk->icsk_ca_ops->pkts_acked(sk, &sample);
3418 #if FASTRETRANS_DEBUG > 0
3419 WARN_ON((int)tp->sacked_out < 0);
3420 WARN_ON((int)tp->lost_out < 0);
3421 WARN_ON((int)tp->retrans_out < 0);
3422 if (!tp->packets_out && tcp_is_sack(tp)) {
3423 icsk = inet_csk(sk);
3425 pr_debug("Leak l=%u %d\n",
3426 tp->lost_out, icsk->icsk_ca_state);
3429 if (tp->sacked_out) {
3430 pr_debug("Leak s=%u %d\n",
3431 tp->sacked_out, icsk->icsk_ca_state);
3434 if (tp->retrans_out) {
3435 pr_debug("Leak r=%u %d\n",
3436 tp->retrans_out, icsk->icsk_ca_state);
3437 tp->retrans_out = 0;
3444 static void tcp_ack_probe(struct sock *sk)
3446 struct inet_connection_sock *icsk = inet_csk(sk);
3447 struct sk_buff *head = tcp_send_head(sk);
3448 const struct tcp_sock *tp = tcp_sk(sk);
3450 /* Was it a usable window open? */
3453 if (!after(TCP_SKB_CB(head)->end_seq, tcp_wnd_end(tp))) {
3454 icsk->icsk_backoff = 0;
3455 icsk->icsk_probes_tstamp = 0;
3456 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3457 /* Socket must be waked up by subsequent tcp_data_snd_check().
3458 * This function is not for random using!
3461 unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX);
3463 when = tcp_clamp_probe0_to_user_timeout(sk, when);
3464 tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, when, TCP_RTO_MAX);
3468 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3470 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3471 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3474 /* Decide wheather to run the increase function of congestion control. */
3475 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3477 /* If reordering is high then always grow cwnd whenever data is
3478 * delivered regardless of its ordering. Otherwise stay conservative
3479 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3480 * new SACK or ECE mark may first advance cwnd here and later reduce
3481 * cwnd in tcp_fastretrans_alert() based on more states.
3483 if (tcp_sk(sk)->reordering >
3484 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_reordering))
3485 return flag & FLAG_FORWARD_PROGRESS;
3487 return flag & FLAG_DATA_ACKED;
3490 /* The "ultimate" congestion control function that aims to replace the rigid
3491 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction).
3492 * It's called toward the end of processing an ACK with precise rate
3493 * information. All transmission or retransmission are delayed afterwards.
3495 static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked,
3496 int flag, const struct rate_sample *rs)
3498 const struct inet_connection_sock *icsk = inet_csk(sk);
3500 if (icsk->icsk_ca_ops->cong_control) {
3501 icsk->icsk_ca_ops->cong_control(sk, rs);
3505 if (tcp_in_cwnd_reduction(sk)) {
3506 /* Reduce cwnd if state mandates */
3507 tcp_cwnd_reduction(sk, acked_sacked, rs->losses, flag);
3508 } else if (tcp_may_raise_cwnd(sk, flag)) {
3509 /* Advance cwnd if state allows */
3510 tcp_cong_avoid(sk, ack, acked_sacked);
3512 tcp_update_pacing_rate(sk);
3515 /* Check that window update is acceptable.
3516 * The function assumes that snd_una<=ack<=snd_next.
3518 static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3519 const u32 ack, const u32 ack_seq,
3522 return after(ack, tp->snd_una) ||
3523 after(ack_seq, tp->snd_wl1) ||
3524 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3527 /* If we update tp->snd_una, also update tp->bytes_acked */
3528 static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack)
3530 u32 delta = ack - tp->snd_una;
3532 sock_owned_by_me((struct sock *)tp);
3533 tp->bytes_acked += delta;
3537 /* If we update tp->rcv_nxt, also update tp->bytes_received */
3538 static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq)
3540 u32 delta = seq - tp->rcv_nxt;
3542 sock_owned_by_me((struct sock *)tp);
3543 tp->bytes_received += delta;
3544 WRITE_ONCE(tp->rcv_nxt, seq);
3547 /* Update our send window.
3549 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3550 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3552 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3555 struct tcp_sock *tp = tcp_sk(sk);
3557 u32 nwin = ntohs(tcp_hdr(skb)->window);
3559 if (likely(!tcp_hdr(skb)->syn))
3560 nwin <<= tp->rx_opt.snd_wscale;
3562 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3563 flag |= FLAG_WIN_UPDATE;
3564 tcp_update_wl(tp, ack_seq);
3566 if (tp->snd_wnd != nwin) {
3569 /* Note, it is the only place, where
3570 * fast path is recovered for sending TCP.
3573 tcp_fast_path_check(sk);
3575 if (!tcp_write_queue_empty(sk))
3576 tcp_slow_start_after_idle_check(sk);
3578 if (nwin > tp->max_window) {
3579 tp->max_window = nwin;
3580 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3585 tcp_snd_una_update(tp, ack);
3590 static bool __tcp_oow_rate_limited(struct net *net, int mib_idx,
3591 u32 *last_oow_ack_time)
3593 if (*last_oow_ack_time) {
3594 s32 elapsed = (s32)(tcp_jiffies32 - *last_oow_ack_time);
3597 elapsed < READ_ONCE(net->ipv4.sysctl_tcp_invalid_ratelimit)) {
3598 NET_INC_STATS(net, mib_idx);
3599 return true; /* rate-limited: don't send yet! */
3603 *last_oow_ack_time = tcp_jiffies32;
3605 return false; /* not rate-limited: go ahead, send dupack now! */
3608 /* Return true if we're currently rate-limiting out-of-window ACKs and
3609 * thus shouldn't send a dupack right now. We rate-limit dupacks in
3610 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3611 * attacks that send repeated SYNs or ACKs for the same connection. To
3612 * do this, we do not send a duplicate SYNACK or ACK if the remote
3613 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3615 bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb,
3616 int mib_idx, u32 *last_oow_ack_time)
3618 /* Data packets without SYNs are not likely part of an ACK loop. */
3619 if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) &&
3623 return __tcp_oow_rate_limited(net, mib_idx, last_oow_ack_time);
3626 /* RFC 5961 7 [ACK Throttling] */
3627 static void tcp_send_challenge_ack(struct sock *sk)
3629 struct tcp_sock *tp = tcp_sk(sk);
3630 struct net *net = sock_net(sk);
3631 u32 count, now, ack_limit;
3633 /* First check our per-socket dupack rate limit. */
3634 if (__tcp_oow_rate_limited(net,
3635 LINUX_MIB_TCPACKSKIPPEDCHALLENGE,
3636 &tp->last_oow_ack_time))
3639 ack_limit = READ_ONCE(net->ipv4.sysctl_tcp_challenge_ack_limit);
3640 if (ack_limit == INT_MAX)
3643 /* Then check host-wide RFC 5961 rate limit. */
3645 if (now != READ_ONCE(net->ipv4.tcp_challenge_timestamp)) {
3646 u32 half = (ack_limit + 1) >> 1;
3648 WRITE_ONCE(net->ipv4.tcp_challenge_timestamp, now);
3649 WRITE_ONCE(net->ipv4.tcp_challenge_count, half + prandom_u32_max(ack_limit));
3651 count = READ_ONCE(net->ipv4.tcp_challenge_count);
3653 WRITE_ONCE(net->ipv4.tcp_challenge_count, count - 1);
3655 NET_INC_STATS(net, LINUX_MIB_TCPCHALLENGEACK);
3660 static void tcp_store_ts_recent(struct tcp_sock *tp)
3662 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3663 tp->rx_opt.ts_recent_stamp = ktime_get_seconds();
3666 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3668 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3669 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3670 * extra check below makes sure this can only happen
3671 * for pure ACK frames. -DaveM
3673 * Not only, also it occurs for expired timestamps.
3676 if (tcp_paws_check(&tp->rx_opt, 0))
3677 tcp_store_ts_recent(tp);
3681 /* This routine deals with acks during a TLP episode and ends an episode by
3682 * resetting tlp_high_seq. Ref: TLP algorithm in draft-ietf-tcpm-rack
3684 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3686 struct tcp_sock *tp = tcp_sk(sk);
3688 if (before(ack, tp->tlp_high_seq))
3691 if (!tp->tlp_retrans) {
3692 /* TLP of new data has been acknowledged */
3693 tp->tlp_high_seq = 0;
3694 } else if (flag & FLAG_DSACK_TLP) {
3695 /* This DSACK means original and TLP probe arrived; no loss */
3696 tp->tlp_high_seq = 0;
3697 } else if (after(ack, tp->tlp_high_seq)) {
3698 /* ACK advances: there was a loss, so reduce cwnd. Reset
3699 * tlp_high_seq in tcp_init_cwnd_reduction()
3701 tcp_init_cwnd_reduction(sk);
3702 tcp_set_ca_state(sk, TCP_CA_CWR);
3703 tcp_end_cwnd_reduction(sk);
3704 tcp_try_keep_open(sk);
3705 NET_INC_STATS(sock_net(sk),
3706 LINUX_MIB_TCPLOSSPROBERECOVERY);
3707 } else if (!(flag & (FLAG_SND_UNA_ADVANCED |
3708 FLAG_NOT_DUP | FLAG_DATA_SACKED))) {
3709 /* Pure dupack: original and TLP probe arrived; no loss */
3710 tp->tlp_high_seq = 0;
3714 static inline void tcp_in_ack_event(struct sock *sk, u32 flags)
3716 const struct inet_connection_sock *icsk = inet_csk(sk);
3718 if (icsk->icsk_ca_ops->in_ack_event)
3719 icsk->icsk_ca_ops->in_ack_event(sk, flags);
3722 /* Congestion control has updated the cwnd already. So if we're in
3723 * loss recovery then now we do any new sends (for FRTO) or
3724 * retransmits (for CA_Loss or CA_recovery) that make sense.
3726 static void tcp_xmit_recovery(struct sock *sk, int rexmit)
3728 struct tcp_sock *tp = tcp_sk(sk);
3730 if (rexmit == REXMIT_NONE || sk->sk_state == TCP_SYN_SENT)
3733 if (unlikely(rexmit == REXMIT_NEW)) {
3734 __tcp_push_pending_frames(sk, tcp_current_mss(sk),
3736 if (after(tp->snd_nxt, tp->high_seq))
3740 tcp_xmit_retransmit_queue(sk);
3743 /* Returns the number of packets newly acked or sacked by the current ACK */
3744 static u32 tcp_newly_delivered(struct sock *sk, u32 prior_delivered, int flag)
3746 const struct net *net = sock_net(sk);
3747 struct tcp_sock *tp = tcp_sk(sk);
3750 delivered = tp->delivered - prior_delivered;
3751 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVERED, delivered);
3752 if (flag & FLAG_ECE)
3753 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVEREDCE, delivered);
3758 /* This routine deals with incoming acks, but not outgoing ones. */
3759 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3761 struct inet_connection_sock *icsk = inet_csk(sk);
3762 struct tcp_sock *tp = tcp_sk(sk);
3763 struct tcp_sacktag_state sack_state;
3764 struct rate_sample rs = { .prior_delivered = 0 };
3765 u32 prior_snd_una = tp->snd_una;
3766 bool is_sack_reneg = tp->is_sack_reneg;
3767 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3768 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3770 int prior_packets = tp->packets_out;
3771 u32 delivered = tp->delivered;
3772 u32 lost = tp->lost;
3773 int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */
3776 sack_state.first_sackt = 0;
3777 sack_state.rate = &rs;
3778 sack_state.sack_delivered = 0;
3780 /* We very likely will need to access rtx queue. */
3781 prefetch(sk->tcp_rtx_queue.rb_node);
3783 /* If the ack is older than previous acks
3784 * then we can probably ignore it.
3786 if (before(ack, prior_snd_una)) {
3787 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3788 if (before(ack, prior_snd_una - tp->max_window)) {
3789 if (!(flag & FLAG_NO_CHALLENGE_ACK))
3790 tcp_send_challenge_ack(sk);
3791 return -SKB_DROP_REASON_TCP_TOO_OLD_ACK;
3796 /* If the ack includes data we haven't sent yet, discard
3797 * this segment (RFC793 Section 3.9).
3799 if (after(ack, tp->snd_nxt))
3800 return -SKB_DROP_REASON_TCP_ACK_UNSENT_DATA;
3802 if (after(ack, prior_snd_una)) {
3803 flag |= FLAG_SND_UNA_ADVANCED;
3804 icsk->icsk_retransmits = 0;
3806 #if IS_ENABLED(CONFIG_TLS_DEVICE)
3807 if (static_branch_unlikely(&clean_acked_data_enabled.key))
3808 if (icsk->icsk_clean_acked)
3809 icsk->icsk_clean_acked(sk, ack);
3813 prior_fack = tcp_is_sack(tp) ? tcp_highest_sack_seq(tp) : tp->snd_una;
3814 rs.prior_in_flight = tcp_packets_in_flight(tp);
3816 /* ts_recent update must be made after we are sure that the packet
3819 if (flag & FLAG_UPDATE_TS_RECENT)
3820 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3822 if ((flag & (FLAG_SLOWPATH | FLAG_SND_UNA_ADVANCED)) ==
3823 FLAG_SND_UNA_ADVANCED) {
3824 /* Window is constant, pure forward advance.
3825 * No more checks are required.
3826 * Note, we use the fact that SND.UNA>=SND.WL2.
3828 tcp_update_wl(tp, ack_seq);
3829 tcp_snd_una_update(tp, ack);
3830 flag |= FLAG_WIN_UPDATE;
3832 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
3834 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPACKS);
3836 u32 ack_ev_flags = CA_ACK_SLOWPATH;
3838 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3841 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3843 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3845 if (TCP_SKB_CB(skb)->sacked)
3846 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3849 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) {
3851 ack_ev_flags |= CA_ACK_ECE;
3854 if (sack_state.sack_delivered)
3855 tcp_count_delivered(tp, sack_state.sack_delivered,
3858 if (flag & FLAG_WIN_UPDATE)
3859 ack_ev_flags |= CA_ACK_WIN_UPDATE;
3861 tcp_in_ack_event(sk, ack_ev_flags);
3864 /* This is a deviation from RFC3168 since it states that:
3865 * "When the TCP data sender is ready to set the CWR bit after reducing
3866 * the congestion window, it SHOULD set the CWR bit only on the first
3867 * new data packet that it transmits."
3868 * We accept CWR on pure ACKs to be more robust
3869 * with widely-deployed TCP implementations that do this.
3871 tcp_ecn_accept_cwr(sk, skb);
3873 /* We passed data and got it acked, remove any soft error
3874 * log. Something worked...
3876 sk->sk_err_soft = 0;
3877 icsk->icsk_probes_out = 0;
3878 tp->rcv_tstamp = tcp_jiffies32;
3882 /* See if we can take anything off of the retransmit queue. */
3883 flag |= tcp_clean_rtx_queue(sk, skb, prior_fack, prior_snd_una,
3884 &sack_state, flag & FLAG_ECE);
3886 tcp_rack_update_reo_wnd(sk, &rs);
3888 if (tp->tlp_high_seq)
3889 tcp_process_tlp_ack(sk, ack, flag);
3891 if (tcp_ack_is_dubious(sk, flag)) {
3892 if (!(flag & (FLAG_SND_UNA_ADVANCED |
3893 FLAG_NOT_DUP | FLAG_DSACKING_ACK))) {
3895 /* Consider if pure acks were aggregated in tcp_add_backlog() */
3896 if (!(flag & FLAG_DATA))
3897 num_dupack = max_t(u16, 1, skb_shinfo(skb)->gso_segs);
3899 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
3903 /* If needed, reset TLP/RTO timer when RACK doesn't set. */
3904 if (flag & FLAG_SET_XMIT_TIMER)
3905 tcp_set_xmit_timer(sk);
3907 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
3910 delivered = tcp_newly_delivered(sk, delivered, flag);
3911 lost = tp->lost - lost; /* freshly marked lost */
3912 rs.is_ack_delayed = !!(flag & FLAG_ACK_MAYBE_DELAYED);
3913 tcp_rate_gen(sk, delivered, lost, is_sack_reneg, sack_state.rate);
3914 tcp_cong_control(sk, ack, delivered, flag, sack_state.rate);
3915 tcp_xmit_recovery(sk, rexmit);
3919 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3920 if (flag & FLAG_DSACKING_ACK) {
3921 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
3923 tcp_newly_delivered(sk, delivered, flag);
3925 /* If this ack opens up a zero window, clear backoff. It was
3926 * being used to time the probes, and is probably far higher than
3927 * it needs to be for normal retransmission.
3931 if (tp->tlp_high_seq)
3932 tcp_process_tlp_ack(sk, ack, flag);
3936 /* If data was SACKed, tag it and see if we should send more data.
3937 * If data was DSACKed, see if we can undo a cwnd reduction.
3939 if (TCP_SKB_CB(skb)->sacked) {
3940 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3942 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
3944 tcp_newly_delivered(sk, delivered, flag);
3945 tcp_xmit_recovery(sk, rexmit);
3951 static void tcp_parse_fastopen_option(int len, const unsigned char *cookie,
3952 bool syn, struct tcp_fastopen_cookie *foc,
3955 /* Valid only in SYN or SYN-ACK with an even length. */
3956 if (!foc || !syn || len < 0 || (len & 1))
3959 if (len >= TCP_FASTOPEN_COOKIE_MIN &&
3960 len <= TCP_FASTOPEN_COOKIE_MAX)
3961 memcpy(foc->val, cookie, len);
3968 static bool smc_parse_options(const struct tcphdr *th,
3969 struct tcp_options_received *opt_rx,
3970 const unsigned char *ptr,
3973 #if IS_ENABLED(CONFIG_SMC)
3974 if (static_branch_unlikely(&tcp_have_smc)) {
3975 if (th->syn && !(opsize & 1) &&
3976 opsize >= TCPOLEN_EXP_SMC_BASE &&
3977 get_unaligned_be32(ptr) == TCPOPT_SMC_MAGIC) {
3986 /* Try to parse the MSS option from the TCP header. Return 0 on failure, clamped
3989 u16 tcp_parse_mss_option(const struct tcphdr *th, u16 user_mss)
3991 const unsigned char *ptr = (const unsigned char *)(th + 1);
3992 int length = (th->doff * 4) - sizeof(struct tcphdr);
3995 while (length > 0) {
3996 int opcode = *ptr++;
4002 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
4009 if (opsize < 2) /* "silly options" */
4011 if (opsize > length)
4012 return mss; /* fail on partial options */
4013 if (opcode == TCPOPT_MSS && opsize == TCPOLEN_MSS) {
4014 u16 in_mss = get_unaligned_be16(ptr);
4017 if (user_mss && user_mss < in_mss)
4028 EXPORT_SYMBOL_GPL(tcp_parse_mss_option);
4030 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
4031 * But, this can also be called on packets in the established flow when
4032 * the fast version below fails.
4034 void tcp_parse_options(const struct net *net,
4035 const struct sk_buff *skb,
4036 struct tcp_options_received *opt_rx, int estab,
4037 struct tcp_fastopen_cookie *foc)
4039 const unsigned char *ptr;
4040 const struct tcphdr *th = tcp_hdr(skb);
4041 int length = (th->doff * 4) - sizeof(struct tcphdr);
4043 ptr = (const unsigned char *)(th + 1);
4044 opt_rx->saw_tstamp = 0;
4045 opt_rx->saw_unknown = 0;
4047 while (length > 0) {
4048 int opcode = *ptr++;
4054 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
4061 if (opsize < 2) /* "silly options" */
4063 if (opsize > length)
4064 return; /* don't parse partial options */
4067 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
4068 u16 in_mss = get_unaligned_be16(ptr);
4070 if (opt_rx->user_mss &&
4071 opt_rx->user_mss < in_mss)
4072 in_mss = opt_rx->user_mss;
4073 opt_rx->mss_clamp = in_mss;
4078 if (opsize == TCPOLEN_WINDOW && th->syn &&
4079 !estab && READ_ONCE(net->ipv4.sysctl_tcp_window_scaling)) {
4080 __u8 snd_wscale = *(__u8 *)ptr;
4081 opt_rx->wscale_ok = 1;
4082 if (snd_wscale > TCP_MAX_WSCALE) {
4083 net_info_ratelimited("%s: Illegal window scaling value %d > %u received\n",
4087 snd_wscale = TCP_MAX_WSCALE;
4089 opt_rx->snd_wscale = snd_wscale;
4092 case TCPOPT_TIMESTAMP:
4093 if ((opsize == TCPOLEN_TIMESTAMP) &&
4094 ((estab && opt_rx->tstamp_ok) ||
4095 (!estab && READ_ONCE(net->ipv4.sysctl_tcp_timestamps)))) {
4096 opt_rx->saw_tstamp = 1;
4097 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
4098 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
4101 case TCPOPT_SACK_PERM:
4102 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
4103 !estab && READ_ONCE(net->ipv4.sysctl_tcp_sack)) {
4104 opt_rx->sack_ok = TCP_SACK_SEEN;
4105 tcp_sack_reset(opt_rx);
4110 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
4111 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
4113 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
4116 #ifdef CONFIG_TCP_MD5SIG
4119 * The MD5 Hash has already been
4120 * checked (see tcp_v{4,6}_do_rcv()).
4124 case TCPOPT_FASTOPEN:
4125 tcp_parse_fastopen_option(
4126 opsize - TCPOLEN_FASTOPEN_BASE,
4127 ptr, th->syn, foc, false);
4131 /* Fast Open option shares code 254 using a
4132 * 16 bits magic number.
4134 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE &&
4135 get_unaligned_be16(ptr) ==
4136 TCPOPT_FASTOPEN_MAGIC) {
4137 tcp_parse_fastopen_option(opsize -
4138 TCPOLEN_EXP_FASTOPEN_BASE,
4139 ptr + 2, th->syn, foc, true);
4143 if (smc_parse_options(th, opt_rx, ptr, opsize))
4146 opt_rx->saw_unknown = 1;
4150 opt_rx->saw_unknown = 1;
4157 EXPORT_SYMBOL(tcp_parse_options);
4159 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
4161 const __be32 *ptr = (const __be32 *)(th + 1);
4163 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
4164 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
4165 tp->rx_opt.saw_tstamp = 1;
4167 tp->rx_opt.rcv_tsval = ntohl(*ptr);
4170 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
4172 tp->rx_opt.rcv_tsecr = 0;
4178 /* Fast parse options. This hopes to only see timestamps.
4179 * If it is wrong it falls back on tcp_parse_options().
4181 static bool tcp_fast_parse_options(const struct net *net,
4182 const struct sk_buff *skb,
4183 const struct tcphdr *th, struct tcp_sock *tp)
4185 /* In the spirit of fast parsing, compare doff directly to constant
4186 * values. Because equality is used, short doff can be ignored here.
4188 if (th->doff == (sizeof(*th) / 4)) {
4189 tp->rx_opt.saw_tstamp = 0;
4191 } else if (tp->rx_opt.tstamp_ok &&
4192 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
4193 if (tcp_parse_aligned_timestamp(tp, th))
4197 tcp_parse_options(net, skb, &tp->rx_opt, 1, NULL);
4198 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
4199 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
4204 #ifdef CONFIG_TCP_MD5SIG
4206 * Parse MD5 Signature option
4208 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
4210 int length = (th->doff << 2) - sizeof(*th);
4211 const u8 *ptr = (const u8 *)(th + 1);
4213 /* If not enough data remaining, we can short cut */
4214 while (length >= TCPOLEN_MD5SIG) {
4215 int opcode = *ptr++;
4226 if (opsize < 2 || opsize > length)
4228 if (opcode == TCPOPT_MD5SIG)
4229 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
4236 EXPORT_SYMBOL(tcp_parse_md5sig_option);
4239 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
4241 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
4242 * it can pass through stack. So, the following predicate verifies that
4243 * this segment is not used for anything but congestion avoidance or
4244 * fast retransmit. Moreover, we even are able to eliminate most of such
4245 * second order effects, if we apply some small "replay" window (~RTO)
4246 * to timestamp space.
4248 * All these measures still do not guarantee that we reject wrapped ACKs
4249 * on networks with high bandwidth, when sequence space is recycled fastly,
4250 * but it guarantees that such events will be very rare and do not affect
4251 * connection seriously. This doesn't look nice, but alas, PAWS is really
4254 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
4255 * states that events when retransmit arrives after original data are rare.
4256 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
4257 * the biggest problem on large power networks even with minor reordering.
4258 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
4259 * up to bandwidth of 18Gigabit/sec. 8) ]
4262 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
4264 const struct tcp_sock *tp = tcp_sk(sk);
4265 const struct tcphdr *th = tcp_hdr(skb);
4266 u32 seq = TCP_SKB_CB(skb)->seq;
4267 u32 ack = TCP_SKB_CB(skb)->ack_seq;
4269 return (/* 1. Pure ACK with correct sequence number. */
4270 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
4272 /* 2. ... and duplicate ACK. */
4273 ack == tp->snd_una &&
4275 /* 3. ... and does not update window. */
4276 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
4278 /* 4. ... and sits in replay window. */
4279 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
4282 static inline bool tcp_paws_discard(const struct sock *sk,
4283 const struct sk_buff *skb)
4285 const struct tcp_sock *tp = tcp_sk(sk);
4287 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
4288 !tcp_disordered_ack(sk, skb);
4291 /* Check segment sequence number for validity.
4293 * Segment controls are considered valid, if the segment
4294 * fits to the window after truncation to the window. Acceptability
4295 * of data (and SYN, FIN, of course) is checked separately.
4296 * See tcp_data_queue(), for example.
4298 * Also, controls (RST is main one) are accepted using RCV.WUP instead
4299 * of RCV.NXT. Peer still did not advance his SND.UNA when we
4300 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4301 * (borrowed from freebsd)
4304 static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
4306 return !before(end_seq, tp->rcv_wup) &&
4307 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
4310 /* When we get a reset we do this. */
4311 void tcp_reset(struct sock *sk, struct sk_buff *skb)
4313 trace_tcp_receive_reset(sk);
4315 /* mptcp can't tell us to ignore reset pkts,
4316 * so just ignore the return value of mptcp_incoming_options().
4318 if (sk_is_mptcp(sk))
4319 mptcp_incoming_options(sk, skb);
4321 /* We want the right error as BSD sees it (and indeed as we do). */
4322 switch (sk->sk_state) {
4324 sk->sk_err = ECONNREFUSED;
4326 case TCP_CLOSE_WAIT:
4332 sk->sk_err = ECONNRESET;
4334 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4337 tcp_write_queue_purge(sk);
4340 if (!sock_flag(sk, SOCK_DEAD))
4341 sk_error_report(sk);
4345 * Process the FIN bit. This now behaves as it is supposed to work
4346 * and the FIN takes effect when it is validly part of sequence
4347 * space. Not before when we get holes.
4349 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4350 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4353 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4354 * close and we go into CLOSING (and later onto TIME-WAIT)
4356 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4358 void tcp_fin(struct sock *sk)
4360 struct tcp_sock *tp = tcp_sk(sk);
4362 inet_csk_schedule_ack(sk);
4364 WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | RCV_SHUTDOWN);
4365 sock_set_flag(sk, SOCK_DONE);
4367 switch (sk->sk_state) {
4369 case TCP_ESTABLISHED:
4370 /* Move to CLOSE_WAIT */
4371 tcp_set_state(sk, TCP_CLOSE_WAIT);
4372 inet_csk_enter_pingpong_mode(sk);
4375 case TCP_CLOSE_WAIT:
4377 /* Received a retransmission of the FIN, do
4382 /* RFC793: Remain in the LAST-ACK state. */
4386 /* This case occurs when a simultaneous close
4387 * happens, we must ack the received FIN and
4388 * enter the CLOSING state.
4391 tcp_set_state(sk, TCP_CLOSING);
4394 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4396 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4399 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4400 * cases we should never reach this piece of code.
4402 pr_err("%s: Impossible, sk->sk_state=%d\n",
4403 __func__, sk->sk_state);
4407 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4408 * Probably, we should reset in this case. For now drop them.
4410 skb_rbtree_purge(&tp->out_of_order_queue);
4411 if (tcp_is_sack(tp))
4412 tcp_sack_reset(&tp->rx_opt);
4414 if (!sock_flag(sk, SOCK_DEAD)) {
4415 sk->sk_state_change(sk);
4417 /* Do not send POLL_HUP for half duplex close. */
4418 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4419 sk->sk_state == TCP_CLOSE)
4420 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4422 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4426 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4429 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4430 if (before(seq, sp->start_seq))
4431 sp->start_seq = seq;
4432 if (after(end_seq, sp->end_seq))
4433 sp->end_seq = end_seq;
4439 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4441 struct tcp_sock *tp = tcp_sk(sk);
4443 if (tcp_is_sack(tp) && READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_dsack)) {
4446 if (before(seq, tp->rcv_nxt))
4447 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4449 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4451 NET_INC_STATS(sock_net(sk), mib_idx);
4453 tp->rx_opt.dsack = 1;
4454 tp->duplicate_sack[0].start_seq = seq;
4455 tp->duplicate_sack[0].end_seq = end_seq;
4459 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4461 struct tcp_sock *tp = tcp_sk(sk);
4463 if (!tp->rx_opt.dsack)
4464 tcp_dsack_set(sk, seq, end_seq);
4466 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4469 static void tcp_rcv_spurious_retrans(struct sock *sk, const struct sk_buff *skb)
4471 /* When the ACK path fails or drops most ACKs, the sender would
4472 * timeout and spuriously retransmit the same segment repeatedly.
4473 * The receiver remembers and reflects via DSACKs. Leverage the
4474 * DSACK state and change the txhash to re-route speculatively.
4476 if (TCP_SKB_CB(skb)->seq == tcp_sk(sk)->duplicate_sack[0].start_seq &&
4477 sk_rethink_txhash(sk))
4478 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDUPLICATEDATAREHASH);
4481 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4483 struct tcp_sock *tp = tcp_sk(sk);
4485 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4486 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4487 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4488 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
4490 if (tcp_is_sack(tp) && READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_dsack)) {
4491 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4493 tcp_rcv_spurious_retrans(sk, skb);
4494 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4495 end_seq = tp->rcv_nxt;
4496 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4503 /* These routines update the SACK block as out-of-order packets arrive or
4504 * in-order packets close up the sequence space.
4506 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4509 struct tcp_sack_block *sp = &tp->selective_acks[0];
4510 struct tcp_sack_block *swalk = sp + 1;
4512 /* See if the recent change to the first SACK eats into
4513 * or hits the sequence space of other SACK blocks, if so coalesce.
4515 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4516 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4519 /* Zap SWALK, by moving every further SACK up by one slot.
4520 * Decrease num_sacks.
4522 tp->rx_opt.num_sacks--;
4523 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4532 void tcp_sack_compress_send_ack(struct sock *sk)
4534 struct tcp_sock *tp = tcp_sk(sk);
4536 if (!tp->compressed_ack)
4539 if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1)
4542 /* Since we have to send one ack finally,
4543 * substract one from tp->compressed_ack to keep
4544 * LINUX_MIB_TCPACKCOMPRESSED accurate.
4546 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED,
4547 tp->compressed_ack - 1);
4549 tp->compressed_ack = 0;
4553 /* Reasonable amount of sack blocks included in TCP SACK option
4554 * The max is 4, but this becomes 3 if TCP timestamps are there.
4555 * Given that SACK packets might be lost, be conservative and use 2.
4557 #define TCP_SACK_BLOCKS_EXPECTED 2
4559 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4561 struct tcp_sock *tp = tcp_sk(sk);
4562 struct tcp_sack_block *sp = &tp->selective_acks[0];
4563 int cur_sacks = tp->rx_opt.num_sacks;
4569 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4570 if (tcp_sack_extend(sp, seq, end_seq)) {
4571 if (this_sack >= TCP_SACK_BLOCKS_EXPECTED)
4572 tcp_sack_compress_send_ack(sk);
4573 /* Rotate this_sack to the first one. */
4574 for (; this_sack > 0; this_sack--, sp--)
4575 swap(*sp, *(sp - 1));
4577 tcp_sack_maybe_coalesce(tp);
4582 if (this_sack >= TCP_SACK_BLOCKS_EXPECTED)
4583 tcp_sack_compress_send_ack(sk);
4585 /* Could not find an adjacent existing SACK, build a new one,
4586 * put it at the front, and shift everyone else down. We
4587 * always know there is at least one SACK present already here.
4589 * If the sack array is full, forget about the last one.
4591 if (this_sack >= TCP_NUM_SACKS) {
4593 tp->rx_opt.num_sacks--;
4596 for (; this_sack > 0; this_sack--, sp--)
4600 /* Build the new head SACK, and we're done. */
4601 sp->start_seq = seq;
4602 sp->end_seq = end_seq;
4603 tp->rx_opt.num_sacks++;
4606 /* RCV.NXT advances, some SACKs should be eaten. */
4608 static void tcp_sack_remove(struct tcp_sock *tp)
4610 struct tcp_sack_block *sp = &tp->selective_acks[0];
4611 int num_sacks = tp->rx_opt.num_sacks;
4614 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4615 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4616 tp->rx_opt.num_sacks = 0;
4620 for (this_sack = 0; this_sack < num_sacks;) {
4621 /* Check if the start of the sack is covered by RCV.NXT. */
4622 if (!before(tp->rcv_nxt, sp->start_seq)) {
4625 /* RCV.NXT must cover all the block! */
4626 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4628 /* Zap this SACK, by moving forward any other SACKS. */
4629 for (i = this_sack+1; i < num_sacks; i++)
4630 tp->selective_acks[i-1] = tp->selective_acks[i];
4637 tp->rx_opt.num_sacks = num_sacks;
4641 * tcp_try_coalesce - try to merge skb to prior one
4644 * @from: buffer to add in queue
4645 * @fragstolen: pointer to boolean
4647 * Before queueing skb @from after @to, try to merge them
4648 * to reduce overall memory use and queue lengths, if cost is small.
4649 * Packets in ofo or receive queues can stay a long time.
4650 * Better try to coalesce them right now to avoid future collapses.
4651 * Returns true if caller should free @from instead of queueing it
4653 static bool tcp_try_coalesce(struct sock *sk,
4655 struct sk_buff *from,
4660 *fragstolen = false;
4662 /* Its possible this segment overlaps with prior segment in queue */
4663 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4666 if (!mptcp_skb_can_collapse(to, from))
4669 #ifdef CONFIG_TLS_DEVICE
4670 if (from->decrypted != to->decrypted)
4674 if (!skb_try_coalesce(to, from, fragstolen, &delta))
4677 atomic_add(delta, &sk->sk_rmem_alloc);
4678 sk_mem_charge(sk, delta);
4679 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4680 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4681 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4682 TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
4684 if (TCP_SKB_CB(from)->has_rxtstamp) {
4685 TCP_SKB_CB(to)->has_rxtstamp = true;
4686 to->tstamp = from->tstamp;
4687 skb_hwtstamps(to)->hwtstamp = skb_hwtstamps(from)->hwtstamp;
4693 static bool tcp_ooo_try_coalesce(struct sock *sk,
4695 struct sk_buff *from,
4698 bool res = tcp_try_coalesce(sk, to, from, fragstolen);
4700 /* In case tcp_drop_reason() is called later, update to->gso_segs */
4702 u32 gso_segs = max_t(u16, 1, skb_shinfo(to)->gso_segs) +
4703 max_t(u16, 1, skb_shinfo(from)->gso_segs);
4705 skb_shinfo(to)->gso_segs = min_t(u32, gso_segs, 0xFFFF);
4710 static void tcp_drop_reason(struct sock *sk, struct sk_buff *skb,
4711 enum skb_drop_reason reason)
4713 sk_drops_add(sk, skb);
4714 kfree_skb_reason(skb, reason);
4717 /* This one checks to see if we can put data from the
4718 * out_of_order queue into the receive_queue.
4720 static void tcp_ofo_queue(struct sock *sk)
4722 struct tcp_sock *tp = tcp_sk(sk);
4723 __u32 dsack_high = tp->rcv_nxt;
4724 bool fin, fragstolen, eaten;
4725 struct sk_buff *skb, *tail;
4728 p = rb_first(&tp->out_of_order_queue);
4731 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4734 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4735 __u32 dsack = dsack_high;
4736 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4737 dsack_high = TCP_SKB_CB(skb)->end_seq;
4738 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4741 rb_erase(&skb->rbnode, &tp->out_of_order_queue);
4743 if (unlikely(!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))) {
4744 tcp_drop_reason(sk, skb, SKB_DROP_REASON_TCP_OFO_DROP);
4748 tail = skb_peek_tail(&sk->sk_receive_queue);
4749 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen);
4750 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4751 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
4753 __skb_queue_tail(&sk->sk_receive_queue, skb);
4755 kfree_skb_partial(skb, fragstolen);
4757 if (unlikely(fin)) {
4759 /* tcp_fin() purges tp->out_of_order_queue,
4760 * so we must end this loop right now.
4767 static bool tcp_prune_ofo_queue(struct sock *sk);
4768 static int tcp_prune_queue(struct sock *sk);
4770 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4773 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4774 !sk_rmem_schedule(sk, skb, size)) {
4776 if (tcp_prune_queue(sk) < 0)
4779 while (!sk_rmem_schedule(sk, skb, size)) {
4780 if (!tcp_prune_ofo_queue(sk))
4787 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4789 struct tcp_sock *tp = tcp_sk(sk);
4790 struct rb_node **p, *parent;
4791 struct sk_buff *skb1;
4795 tcp_ecn_check_ce(sk, skb);
4797 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4798 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFODROP);
4799 sk->sk_data_ready(sk);
4800 tcp_drop_reason(sk, skb, SKB_DROP_REASON_PROTO_MEM);
4804 /* Disable header prediction. */
4806 inet_csk_schedule_ack(sk);
4808 tp->rcv_ooopack += max_t(u16, 1, skb_shinfo(skb)->gso_segs);
4809 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4810 seq = TCP_SKB_CB(skb)->seq;
4811 end_seq = TCP_SKB_CB(skb)->end_seq;
4813 p = &tp->out_of_order_queue.rb_node;
4814 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4815 /* Initial out of order segment, build 1 SACK. */
4816 if (tcp_is_sack(tp)) {
4817 tp->rx_opt.num_sacks = 1;
4818 tp->selective_acks[0].start_seq = seq;
4819 tp->selective_acks[0].end_seq = end_seq;
4821 rb_link_node(&skb->rbnode, NULL, p);
4822 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4823 tp->ooo_last_skb = skb;
4827 /* In the typical case, we are adding an skb to the end of the list.
4828 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
4830 if (tcp_ooo_try_coalesce(sk, tp->ooo_last_skb,
4831 skb, &fragstolen)) {
4833 /* For non sack flows, do not grow window to force DUPACK
4834 * and trigger fast retransmit.
4836 if (tcp_is_sack(tp))
4837 tcp_grow_window(sk, skb, true);
4838 kfree_skb_partial(skb, fragstolen);
4842 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
4843 if (!before(seq, TCP_SKB_CB(tp->ooo_last_skb)->end_seq)) {
4844 parent = &tp->ooo_last_skb->rbnode;
4845 p = &parent->rb_right;
4849 /* Find place to insert this segment. Handle overlaps on the way. */
4853 skb1 = rb_to_skb(parent);
4854 if (before(seq, TCP_SKB_CB(skb1)->seq)) {
4855 p = &parent->rb_left;
4858 if (before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4859 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4860 /* All the bits are present. Drop. */
4861 NET_INC_STATS(sock_net(sk),
4862 LINUX_MIB_TCPOFOMERGE);
4863 tcp_drop_reason(sk, skb,
4864 SKB_DROP_REASON_TCP_OFOMERGE);
4866 tcp_dsack_set(sk, seq, end_seq);
4869 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4870 /* Partial overlap. */
4871 tcp_dsack_set(sk, seq, TCP_SKB_CB(skb1)->end_seq);
4873 /* skb's seq == skb1's seq and skb covers skb1.
4874 * Replace skb1 with skb.
4876 rb_replace_node(&skb1->rbnode, &skb->rbnode,
4877 &tp->out_of_order_queue);
4878 tcp_dsack_extend(sk,
4879 TCP_SKB_CB(skb1)->seq,
4880 TCP_SKB_CB(skb1)->end_seq);
4881 NET_INC_STATS(sock_net(sk),
4882 LINUX_MIB_TCPOFOMERGE);
4883 tcp_drop_reason(sk, skb1,
4884 SKB_DROP_REASON_TCP_OFOMERGE);
4887 } else if (tcp_ooo_try_coalesce(sk, skb1,
4888 skb, &fragstolen)) {
4891 p = &parent->rb_right;
4894 /* Insert segment into RB tree. */
4895 rb_link_node(&skb->rbnode, parent, p);
4896 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4899 /* Remove other segments covered by skb. */
4900 while ((skb1 = skb_rb_next(skb)) != NULL) {
4901 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4903 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4904 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4908 rb_erase(&skb1->rbnode, &tp->out_of_order_queue);
4909 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4910 TCP_SKB_CB(skb1)->end_seq);
4911 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4912 tcp_drop_reason(sk, skb1, SKB_DROP_REASON_TCP_OFOMERGE);
4914 /* If there is no skb after us, we are the last_skb ! */
4916 tp->ooo_last_skb = skb;
4919 if (tcp_is_sack(tp))
4920 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4923 /* For non sack flows, do not grow window to force DUPACK
4924 * and trigger fast retransmit.
4926 if (tcp_is_sack(tp))
4927 tcp_grow_window(sk, skb, false);
4929 skb_set_owner_r(skb, sk);
4933 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb,
4937 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4940 tcp_try_coalesce(sk, tail,
4941 skb, fragstolen)) ? 1 : 0;
4942 tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq);
4944 __skb_queue_tail(&sk->sk_receive_queue, skb);
4945 skb_set_owner_r(skb, sk);
4950 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4952 struct sk_buff *skb;
4960 if (size > PAGE_SIZE) {
4961 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS);
4963 data_len = npages << PAGE_SHIFT;
4964 size = data_len + (size & ~PAGE_MASK);
4966 skb = alloc_skb_with_frags(size - data_len, data_len,
4967 PAGE_ALLOC_COSTLY_ORDER,
4968 &err, sk->sk_allocation);
4972 skb_put(skb, size - data_len);
4973 skb->data_len = data_len;
4976 if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) {
4977 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP);
4981 err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size);
4985 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4986 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4987 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4989 if (tcp_queue_rcv(sk, skb, &fragstolen)) {
4990 WARN_ON_ONCE(fragstolen); /* should not happen */
5002 void tcp_data_ready(struct sock *sk)
5004 if (tcp_epollin_ready(sk, sk->sk_rcvlowat) || sock_flag(sk, SOCK_DONE))
5005 sk->sk_data_ready(sk);
5008 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
5010 struct tcp_sock *tp = tcp_sk(sk);
5011 enum skb_drop_reason reason;
5015 /* If a subflow has been reset, the packet should not continue
5016 * to be processed, drop the packet.
5018 if (sk_is_mptcp(sk) && !mptcp_incoming_options(sk, skb)) {
5023 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
5028 __skb_pull(skb, tcp_hdr(skb)->doff * 4);
5030 reason = SKB_DROP_REASON_NOT_SPECIFIED;
5031 tp->rx_opt.dsack = 0;
5033 /* Queue data for delivery to the user.
5034 * Packets in sequence go to the receive queue.
5035 * Out of sequence packets to the out_of_order_queue.
5037 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
5038 if (tcp_receive_window(tp) == 0) {
5039 reason = SKB_DROP_REASON_TCP_ZEROWINDOW;
5040 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP);
5044 /* Ok. In sequence. In window. */
5046 if (skb_queue_len(&sk->sk_receive_queue) == 0)
5047 sk_forced_mem_schedule(sk, skb->truesize);
5048 else if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) {
5049 reason = SKB_DROP_REASON_PROTO_MEM;
5050 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP);
5051 sk->sk_data_ready(sk);
5055 eaten = tcp_queue_rcv(sk, skb, &fragstolen);
5057 tcp_event_data_recv(sk, skb);
5058 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
5061 if (!RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
5064 /* RFC5681. 4.2. SHOULD send immediate ACK, when
5065 * gap in queue is filled.
5067 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
5068 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW;
5071 if (tp->rx_opt.num_sacks)
5072 tcp_sack_remove(tp);
5074 tcp_fast_path_check(sk);
5077 kfree_skb_partial(skb, fragstolen);
5078 if (!sock_flag(sk, SOCK_DEAD))
5083 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
5084 tcp_rcv_spurious_retrans(sk, skb);
5085 /* A retransmit, 2nd most common case. Force an immediate ack. */
5086 reason = SKB_DROP_REASON_TCP_OLD_DATA;
5087 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
5088 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
5091 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
5092 inet_csk_schedule_ack(sk);
5094 tcp_drop_reason(sk, skb, reason);
5098 /* Out of window. F.e. zero window probe. */
5099 if (!before(TCP_SKB_CB(skb)->seq,
5100 tp->rcv_nxt + tcp_receive_window(tp))) {
5101 reason = SKB_DROP_REASON_TCP_OVERWINDOW;
5105 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
5106 /* Partial packet, seq < rcv_next < end_seq */
5107 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
5109 /* If window is closed, drop tail of packet. But after
5110 * remembering D-SACK for its head made in previous line.
5112 if (!tcp_receive_window(tp)) {
5113 reason = SKB_DROP_REASON_TCP_ZEROWINDOW;
5114 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP);
5120 tcp_data_queue_ofo(sk, skb);
5123 static struct sk_buff *tcp_skb_next(struct sk_buff *skb, struct sk_buff_head *list)
5126 return !skb_queue_is_last(list, skb) ? skb->next : NULL;
5128 return skb_rb_next(skb);
5131 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
5132 struct sk_buff_head *list,
5133 struct rb_root *root)
5135 struct sk_buff *next = tcp_skb_next(skb, list);
5138 __skb_unlink(skb, list);
5140 rb_erase(&skb->rbnode, root);
5143 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
5148 /* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */
5149 void tcp_rbtree_insert(struct rb_root *root, struct sk_buff *skb)
5151 struct rb_node **p = &root->rb_node;
5152 struct rb_node *parent = NULL;
5153 struct sk_buff *skb1;
5157 skb1 = rb_to_skb(parent);
5158 if (before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb1)->seq))
5159 p = &parent->rb_left;
5161 p = &parent->rb_right;
5163 rb_link_node(&skb->rbnode, parent, p);
5164 rb_insert_color(&skb->rbnode, root);
5167 /* Collapse contiguous sequence of skbs head..tail with
5168 * sequence numbers start..end.
5170 * If tail is NULL, this means until the end of the queue.
5172 * Segments with FIN/SYN are not collapsed (only because this
5176 tcp_collapse(struct sock *sk, struct sk_buff_head *list, struct rb_root *root,
5177 struct sk_buff *head, struct sk_buff *tail, u32 start, u32 end)
5179 struct sk_buff *skb = head, *n;
5180 struct sk_buff_head tmp;
5183 /* First, check that queue is collapsible and find
5184 * the point where collapsing can be useful.
5187 for (end_of_skbs = true; skb != NULL && skb != tail; skb = n) {
5188 n = tcp_skb_next(skb, list);
5190 /* No new bits? It is possible on ofo queue. */
5191 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
5192 skb = tcp_collapse_one(sk, skb, list, root);
5198 /* The first skb to collapse is:
5200 * - bloated or contains data before "start" or
5201 * overlaps to the next one and mptcp allow collapsing.
5203 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
5204 (tcp_win_from_space(sk, skb->truesize) > skb->len ||
5205 before(TCP_SKB_CB(skb)->seq, start))) {
5206 end_of_skbs = false;
5210 if (n && n != tail && mptcp_skb_can_collapse(skb, n) &&
5211 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(n)->seq) {
5212 end_of_skbs = false;
5216 /* Decided to skip this, advance start seq. */
5217 start = TCP_SKB_CB(skb)->end_seq;
5220 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
5223 __skb_queue_head_init(&tmp);
5225 while (before(start, end)) {
5226 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
5227 struct sk_buff *nskb;
5229 nskb = alloc_skb(copy, GFP_ATOMIC);
5233 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
5234 #ifdef CONFIG_TLS_DEVICE
5235 nskb->decrypted = skb->decrypted;
5237 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
5239 __skb_queue_before(list, skb, nskb);
5241 __skb_queue_tail(&tmp, nskb); /* defer rbtree insertion */
5242 skb_set_owner_r(nskb, sk);
5243 mptcp_skb_ext_move(nskb, skb);
5245 /* Copy data, releasing collapsed skbs. */
5247 int offset = start - TCP_SKB_CB(skb)->seq;
5248 int size = TCP_SKB_CB(skb)->end_seq - start;
5252 size = min(copy, size);
5253 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
5255 TCP_SKB_CB(nskb)->end_seq += size;
5259 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
5260 skb = tcp_collapse_one(sk, skb, list, root);
5263 !mptcp_skb_can_collapse(nskb, skb) ||
5264 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
5266 #ifdef CONFIG_TLS_DEVICE
5267 if (skb->decrypted != nskb->decrypted)
5274 skb_queue_walk_safe(&tmp, skb, n)
5275 tcp_rbtree_insert(root, skb);
5278 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
5279 * and tcp_collapse() them until all the queue is collapsed.
5281 static void tcp_collapse_ofo_queue(struct sock *sk)
5283 struct tcp_sock *tp = tcp_sk(sk);
5284 u32 range_truesize, sum_tiny = 0;
5285 struct sk_buff *skb, *head;
5288 skb = skb_rb_first(&tp->out_of_order_queue);
5291 tp->ooo_last_skb = skb_rb_last(&tp->out_of_order_queue);
5294 start = TCP_SKB_CB(skb)->seq;
5295 end = TCP_SKB_CB(skb)->end_seq;
5296 range_truesize = skb->truesize;
5298 for (head = skb;;) {
5299 skb = skb_rb_next(skb);
5301 /* Range is terminated when we see a gap or when
5302 * we are at the queue end.
5305 after(TCP_SKB_CB(skb)->seq, end) ||
5306 before(TCP_SKB_CB(skb)->end_seq, start)) {
5307 /* Do not attempt collapsing tiny skbs */
5308 if (range_truesize != head->truesize ||
5309 end - start >= SKB_WITH_OVERHEAD(PAGE_SIZE)) {
5310 tcp_collapse(sk, NULL, &tp->out_of_order_queue,
5311 head, skb, start, end);
5313 sum_tiny += range_truesize;
5314 if (sum_tiny > sk->sk_rcvbuf >> 3)
5320 range_truesize += skb->truesize;
5321 if (unlikely(before(TCP_SKB_CB(skb)->seq, start)))
5322 start = TCP_SKB_CB(skb)->seq;
5323 if (after(TCP_SKB_CB(skb)->end_seq, end))
5324 end = TCP_SKB_CB(skb)->end_seq;
5329 * Clean the out-of-order queue to make room.
5330 * We drop high sequences packets to :
5331 * 1) Let a chance for holes to be filled.
5332 * 2) not add too big latencies if thousands of packets sit there.
5333 * (But if application shrinks SO_RCVBUF, we could still end up
5334 * freeing whole queue here)
5335 * 3) Drop at least 12.5 % of sk_rcvbuf to avoid malicious attacks.
5337 * Return true if queue has shrunk.
5339 static bool tcp_prune_ofo_queue(struct sock *sk)
5341 struct tcp_sock *tp = tcp_sk(sk);
5342 struct rb_node *node, *prev;
5345 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
5348 NET_INC_STATS(sock_net(sk), LINUX_MIB_OFOPRUNED);
5349 goal = sk->sk_rcvbuf >> 3;
5350 node = &tp->ooo_last_skb->rbnode;
5352 prev = rb_prev(node);
5353 rb_erase(node, &tp->out_of_order_queue);
5354 goal -= rb_to_skb(node)->truesize;
5355 tcp_drop_reason(sk, rb_to_skb(node),
5356 SKB_DROP_REASON_TCP_OFO_QUEUE_PRUNE);
5357 if (!prev || goal <= 0) {
5358 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf &&
5359 !tcp_under_memory_pressure(sk))
5361 goal = sk->sk_rcvbuf >> 3;
5365 tp->ooo_last_skb = rb_to_skb(prev);
5367 /* Reset SACK state. A conforming SACK implementation will
5368 * do the same at a timeout based retransmit. When a connection
5369 * is in a sad state like this, we care only about integrity
5370 * of the connection not performance.
5372 if (tp->rx_opt.sack_ok)
5373 tcp_sack_reset(&tp->rx_opt);
5377 /* Reduce allocated memory if we can, trying to get
5378 * the socket within its memory limits again.
5380 * Return less than zero if we should start dropping frames
5381 * until the socket owning process reads some of the data
5382 * to stabilize the situation.
5384 static int tcp_prune_queue(struct sock *sk)
5386 struct tcp_sock *tp = tcp_sk(sk);
5388 NET_INC_STATS(sock_net(sk), LINUX_MIB_PRUNECALLED);
5390 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
5391 tcp_clamp_window(sk);
5392 else if (tcp_under_memory_pressure(sk))
5393 tcp_adjust_rcv_ssthresh(sk);
5395 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5398 tcp_collapse_ofo_queue(sk);
5399 if (!skb_queue_empty(&sk->sk_receive_queue))
5400 tcp_collapse(sk, &sk->sk_receive_queue, NULL,
5401 skb_peek(&sk->sk_receive_queue),
5403 tp->copied_seq, tp->rcv_nxt);
5405 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5408 /* Collapsing did not help, destructive actions follow.
5409 * This must not ever occur. */
5411 tcp_prune_ofo_queue(sk);
5413 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5416 /* If we are really being abused, tell the caller to silently
5417 * drop receive data on the floor. It will get retransmitted
5418 * and hopefully then we'll have sufficient space.
5420 NET_INC_STATS(sock_net(sk), LINUX_MIB_RCVPRUNED);
5422 /* Massive buffer overcommit. */
5427 static bool tcp_should_expand_sndbuf(struct sock *sk)
5429 const struct tcp_sock *tp = tcp_sk(sk);
5431 /* If the user specified a specific send buffer setting, do
5434 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
5437 /* If we are under global TCP memory pressure, do not expand. */
5438 if (tcp_under_memory_pressure(sk)) {
5439 int unused_mem = sk_unused_reserved_mem(sk);
5441 /* Adjust sndbuf according to reserved mem. But make sure
5442 * it never goes below SOCK_MIN_SNDBUF.
5443 * See sk_stream_moderate_sndbuf() for more details.
5445 if (unused_mem > SOCK_MIN_SNDBUF)
5446 WRITE_ONCE(sk->sk_sndbuf, unused_mem);
5451 /* If we are under soft global TCP memory pressure, do not expand. */
5452 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
5455 /* If we filled the congestion window, do not expand. */
5456 if (tcp_packets_in_flight(tp) >= tcp_snd_cwnd(tp))
5462 static void tcp_new_space(struct sock *sk)
5464 struct tcp_sock *tp = tcp_sk(sk);
5466 if (tcp_should_expand_sndbuf(sk)) {
5467 tcp_sndbuf_expand(sk);
5468 tp->snd_cwnd_stamp = tcp_jiffies32;
5471 INDIRECT_CALL_1(sk->sk_write_space, sk_stream_write_space, sk);
5474 /* Caller made space either from:
5475 * 1) Freeing skbs in rtx queues (after tp->snd_una has advanced)
5476 * 2) Sent skbs from output queue (and thus advancing tp->snd_nxt)
5478 * We might be able to generate EPOLLOUT to the application if:
5479 * 1) Space consumed in output/rtx queues is below sk->sk_sndbuf/2
5480 * 2) notsent amount (tp->write_seq - tp->snd_nxt) became
5481 * small enough that tcp_stream_memory_free() decides it
5482 * is time to generate EPOLLOUT.
5484 void tcp_check_space(struct sock *sk)
5486 /* pairs with tcp_poll() */
5488 if (sk->sk_socket &&
5489 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
5491 if (!test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
5492 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
5496 static inline void tcp_data_snd_check(struct sock *sk)
5498 tcp_push_pending_frames(sk);
5499 tcp_check_space(sk);
5503 * Check if sending an ack is needed.
5505 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
5507 struct tcp_sock *tp = tcp_sk(sk);
5508 unsigned long rtt, delay;
5510 /* More than one full frame received... */
5511 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
5512 /* ... and right edge of window advances far enough.
5513 * (tcp_recvmsg() will send ACK otherwise).
5514 * If application uses SO_RCVLOWAT, we want send ack now if
5515 * we have not received enough bytes to satisfy the condition.
5517 (tp->rcv_nxt - tp->copied_seq < sk->sk_rcvlowat ||
5518 __tcp_select_window(sk) >= tp->rcv_wnd)) ||
5519 /* We ACK each frame or... */
5520 tcp_in_quickack_mode(sk) ||
5521 /* Protocol state mandates a one-time immediate ACK */
5522 inet_csk(sk)->icsk_ack.pending & ICSK_ACK_NOW) {
5528 if (!ofo_possible || RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
5529 tcp_send_delayed_ack(sk);
5533 if (!tcp_is_sack(tp) ||
5534 tp->compressed_ack >= READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_nr))
5537 if (tp->compressed_ack_rcv_nxt != tp->rcv_nxt) {
5538 tp->compressed_ack_rcv_nxt = tp->rcv_nxt;
5539 tp->dup_ack_counter = 0;
5541 if (tp->dup_ack_counter < TCP_FASTRETRANS_THRESH) {
5542 tp->dup_ack_counter++;
5545 tp->compressed_ack++;
5546 if (hrtimer_is_queued(&tp->compressed_ack_timer))
5549 /* compress ack timer : 5 % of rtt, but no more than tcp_comp_sack_delay_ns */
5551 rtt = tp->rcv_rtt_est.rtt_us;
5552 if (tp->srtt_us && tp->srtt_us < rtt)
5555 delay = min_t(unsigned long,
5556 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_delay_ns),
5557 rtt * (NSEC_PER_USEC >> 3)/20);
5559 hrtimer_start_range_ns(&tp->compressed_ack_timer, ns_to_ktime(delay),
5560 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_slack_ns),
5561 HRTIMER_MODE_REL_PINNED_SOFT);
5564 static inline void tcp_ack_snd_check(struct sock *sk)
5566 if (!inet_csk_ack_scheduled(sk)) {
5567 /* We sent a data segment already. */
5570 __tcp_ack_snd_check(sk, 1);
5574 * This routine is only called when we have urgent data
5575 * signaled. Its the 'slow' part of tcp_urg. It could be
5576 * moved inline now as tcp_urg is only called from one
5577 * place. We handle URGent data wrong. We have to - as
5578 * BSD still doesn't use the correction from RFC961.
5579 * For 1003.1g we should support a new option TCP_STDURG to permit
5580 * either form (or just set the sysctl tcp_stdurg).
5583 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
5585 struct tcp_sock *tp = tcp_sk(sk);
5586 u32 ptr = ntohs(th->urg_ptr);
5588 if (ptr && !READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_stdurg))
5590 ptr += ntohl(th->seq);
5592 /* Ignore urgent data that we've already seen and read. */
5593 if (after(tp->copied_seq, ptr))
5596 /* Do not replay urg ptr.
5598 * NOTE: interesting situation not covered by specs.
5599 * Misbehaving sender may send urg ptr, pointing to segment,
5600 * which we already have in ofo queue. We are not able to fetch
5601 * such data and will stay in TCP_URG_NOTYET until will be eaten
5602 * by recvmsg(). Seems, we are not obliged to handle such wicked
5603 * situations. But it is worth to think about possibility of some
5604 * DoSes using some hypothetical application level deadlock.
5606 if (before(ptr, tp->rcv_nxt))
5609 /* Do we already have a newer (or duplicate) urgent pointer? */
5610 if (tp->urg_data && !after(ptr, tp->urg_seq))
5613 /* Tell the world about our new urgent pointer. */
5616 /* We may be adding urgent data when the last byte read was
5617 * urgent. To do this requires some care. We cannot just ignore
5618 * tp->copied_seq since we would read the last urgent byte again
5619 * as data, nor can we alter copied_seq until this data arrives
5620 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5622 * NOTE. Double Dutch. Rendering to plain English: author of comment
5623 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5624 * and expect that both A and B disappear from stream. This is _wrong_.
5625 * Though this happens in BSD with high probability, this is occasional.
5626 * Any application relying on this is buggy. Note also, that fix "works"
5627 * only in this artificial test. Insert some normal data between A and B and we will
5628 * decline of BSD again. Verdict: it is better to remove to trap
5631 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5632 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5633 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5635 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5636 __skb_unlink(skb, &sk->sk_receive_queue);
5641 WRITE_ONCE(tp->urg_data, TCP_URG_NOTYET);
5642 WRITE_ONCE(tp->urg_seq, ptr);
5644 /* Disable header prediction. */
5648 /* This is the 'fast' part of urgent handling. */
5649 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
5651 struct tcp_sock *tp = tcp_sk(sk);
5653 /* Check if we get a new urgent pointer - normally not. */
5654 if (unlikely(th->urg))
5655 tcp_check_urg(sk, th);
5657 /* Do we wait for any urgent data? - normally not... */
5658 if (unlikely(tp->urg_data == TCP_URG_NOTYET)) {
5659 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5662 /* Is the urgent pointer pointing into this packet? */
5663 if (ptr < skb->len) {
5665 if (skb_copy_bits(skb, ptr, &tmp, 1))
5667 WRITE_ONCE(tp->urg_data, TCP_URG_VALID | tmp);
5668 if (!sock_flag(sk, SOCK_DEAD))
5669 sk->sk_data_ready(sk);
5674 /* Accept RST for rcv_nxt - 1 after a FIN.
5675 * When tcp connections are abruptly terminated from Mac OSX (via ^C), a
5676 * FIN is sent followed by a RST packet. The RST is sent with the same
5677 * sequence number as the FIN, and thus according to RFC 5961 a challenge
5678 * ACK should be sent. However, Mac OSX rate limits replies to challenge
5679 * ACKs on the closed socket. In addition middleboxes can drop either the
5680 * challenge ACK or a subsequent RST.
5682 static bool tcp_reset_check(const struct sock *sk, const struct sk_buff *skb)
5684 struct tcp_sock *tp = tcp_sk(sk);
5686 return unlikely(TCP_SKB_CB(skb)->seq == (tp->rcv_nxt - 1) &&
5687 (1 << sk->sk_state) & (TCPF_CLOSE_WAIT | TCPF_LAST_ACK |
5691 /* Does PAWS and seqno based validation of an incoming segment, flags will
5692 * play significant role here.
5694 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5695 const struct tcphdr *th, int syn_inerr)
5697 struct tcp_sock *tp = tcp_sk(sk);
5700 /* RFC1323: H1. Apply PAWS check first. */
5701 if (tcp_fast_parse_options(sock_net(sk), skb, th, tp) &&
5702 tp->rx_opt.saw_tstamp &&
5703 tcp_paws_discard(sk, skb)) {
5705 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5706 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5707 LINUX_MIB_TCPACKSKIPPEDPAWS,
5708 &tp->last_oow_ack_time))
5709 tcp_send_dupack(sk, skb);
5710 SKB_DR_SET(reason, TCP_RFC7323_PAWS);
5713 /* Reset is accepted even if it did not pass PAWS. */
5716 /* Step 1: check sequence number */
5717 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5718 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5719 * (RST) segments are validated by checking their SEQ-fields."
5720 * And page 69: "If an incoming segment is not acceptable,
5721 * an acknowledgment should be sent in reply (unless the RST
5722 * bit is set, if so drop the segment and return)".
5727 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5728 LINUX_MIB_TCPACKSKIPPEDSEQ,
5729 &tp->last_oow_ack_time))
5730 tcp_send_dupack(sk, skb);
5731 } else if (tcp_reset_check(sk, skb)) {
5734 SKB_DR_SET(reason, TCP_INVALID_SEQUENCE);
5738 /* Step 2: check RST bit */
5740 /* RFC 5961 3.2 (extend to match against (RCV.NXT - 1) after a
5741 * FIN and SACK too if available):
5742 * If seq num matches RCV.NXT or (RCV.NXT - 1) after a FIN, or
5743 * the right-most SACK block,
5745 * RESET the connection
5747 * Send a challenge ACK
5749 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt ||
5750 tcp_reset_check(sk, skb))
5753 if (tcp_is_sack(tp) && tp->rx_opt.num_sacks > 0) {
5754 struct tcp_sack_block *sp = &tp->selective_acks[0];
5755 int max_sack = sp[0].end_seq;
5758 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;
5760 max_sack = after(sp[this_sack].end_seq,
5762 sp[this_sack].end_seq : max_sack;
5765 if (TCP_SKB_CB(skb)->seq == max_sack)
5769 /* Disable TFO if RST is out-of-order
5770 * and no data has been received
5771 * for current active TFO socket
5773 if (tp->syn_fastopen && !tp->data_segs_in &&
5774 sk->sk_state == TCP_ESTABLISHED)
5775 tcp_fastopen_active_disable(sk);
5776 tcp_send_challenge_ack(sk);
5777 SKB_DR_SET(reason, TCP_RESET);
5781 /* step 3: check security and precedence [ignored] */
5783 /* step 4: Check for a SYN
5784 * RFC 5961 4.2 : Send a challenge ack
5789 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5790 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5791 tcp_send_challenge_ack(sk);
5792 SKB_DR_SET(reason, TCP_INVALID_SYN);
5796 bpf_skops_parse_hdr(sk, skb);
5801 tcp_drop_reason(sk, skb, reason);
5811 * TCP receive function for the ESTABLISHED state.
5813 * It is split into a fast path and a slow path. The fast path is
5815 * - A zero window was announced from us - zero window probing
5816 * is only handled properly in the slow path.
5817 * - Out of order segments arrived.
5818 * - Urgent data is expected.
5819 * - There is no buffer space left
5820 * - Unexpected TCP flags/window values/header lengths are received
5821 * (detected by checking the TCP header against pred_flags)
5822 * - Data is sent in both directions. Fast path only supports pure senders
5823 * or pure receivers (this means either the sequence number or the ack
5824 * value must stay constant)
5825 * - Unexpected TCP option.
5827 * When these conditions are not satisfied it drops into a standard
5828 * receive procedure patterned after RFC793 to handle all cases.
5829 * The first three cases are guaranteed by proper pred_flags setting,
5830 * the rest is checked inline. Fast processing is turned on in
5831 * tcp_data_queue when everything is OK.
5833 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb)
5835 enum skb_drop_reason reason = SKB_DROP_REASON_NOT_SPECIFIED;
5836 const struct tcphdr *th = (const struct tcphdr *)skb->data;
5837 struct tcp_sock *tp = tcp_sk(sk);
5838 unsigned int len = skb->len;
5840 /* TCP congestion window tracking */
5841 trace_tcp_probe(sk, skb);
5843 tcp_mstamp_refresh(tp);
5844 if (unlikely(!rcu_access_pointer(sk->sk_rx_dst)))
5845 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
5847 * Header prediction.
5848 * The code loosely follows the one in the famous
5849 * "30 instruction TCP receive" Van Jacobson mail.
5851 * Van's trick is to deposit buffers into socket queue
5852 * on a device interrupt, to call tcp_recv function
5853 * on the receive process context and checksum and copy
5854 * the buffer to user space. smart...
5856 * Our current scheme is not silly either but we take the
5857 * extra cost of the net_bh soft interrupt processing...
5858 * We do checksum and copy also but from device to kernel.
5861 tp->rx_opt.saw_tstamp = 0;
5863 /* pred_flags is 0xS?10 << 16 + snd_wnd
5864 * if header_prediction is to be made
5865 * 'S' will always be tp->tcp_header_len >> 2
5866 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5867 * turn it off (when there are holes in the receive
5868 * space for instance)
5869 * PSH flag is ignored.
5872 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5873 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5874 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5875 int tcp_header_len = tp->tcp_header_len;
5877 /* Timestamp header prediction: tcp_header_len
5878 * is automatically equal to th->doff*4 due to pred_flags
5882 /* Check timestamp */
5883 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5884 /* No? Slow path! */
5885 if (!tcp_parse_aligned_timestamp(tp, th))
5888 /* If PAWS failed, check it more carefully in slow path */
5889 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5892 /* DO NOT update ts_recent here, if checksum fails
5893 * and timestamp was corrupted part, it will result
5894 * in a hung connection since we will drop all
5895 * future packets due to the PAWS test.
5899 if (len <= tcp_header_len) {
5900 /* Bulk data transfer: sender */
5901 if (len == tcp_header_len) {
5902 /* Predicted packet is in window by definition.
5903 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5904 * Hence, check seq<=rcv_wup reduces to:
5906 if (tcp_header_len ==
5907 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5908 tp->rcv_nxt == tp->rcv_wup)
5909 tcp_store_ts_recent(tp);
5911 /* We know that such packets are checksummed
5914 tcp_ack(sk, skb, 0);
5916 tcp_data_snd_check(sk);
5917 /* When receiving pure ack in fast path, update
5918 * last ts ecr directly instead of calling
5919 * tcp_rcv_rtt_measure_ts()
5921 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr;
5923 } else { /* Header too small */
5924 reason = SKB_DROP_REASON_PKT_TOO_SMALL;
5925 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5930 bool fragstolen = false;
5932 if (tcp_checksum_complete(skb))
5935 if ((int)skb->truesize > sk->sk_forward_alloc)
5938 /* Predicted packet is in window by definition.
5939 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5940 * Hence, check seq<=rcv_wup reduces to:
5942 if (tcp_header_len ==
5943 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5944 tp->rcv_nxt == tp->rcv_wup)
5945 tcp_store_ts_recent(tp);
5947 tcp_rcv_rtt_measure_ts(sk, skb);
5949 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPHITS);
5951 /* Bulk data transfer: receiver */
5953 __skb_pull(skb, tcp_header_len);
5954 eaten = tcp_queue_rcv(sk, skb, &fragstolen);
5956 tcp_event_data_recv(sk, skb);
5958 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5959 /* Well, only one small jumplet in fast path... */
5960 tcp_ack(sk, skb, FLAG_DATA);
5961 tcp_data_snd_check(sk);
5962 if (!inet_csk_ack_scheduled(sk))
5965 tcp_update_wl(tp, TCP_SKB_CB(skb)->seq);
5968 __tcp_ack_snd_check(sk, 0);
5971 kfree_skb_partial(skb, fragstolen);
5978 if (len < (th->doff << 2) || tcp_checksum_complete(skb))
5981 if (!th->ack && !th->rst && !th->syn) {
5982 reason = SKB_DROP_REASON_TCP_FLAGS;
5987 * Standard slow path.
5990 if (!tcp_validate_incoming(sk, skb, th, 1))
5994 reason = tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT);
5995 if ((int)reason < 0) {
5999 tcp_rcv_rtt_measure_ts(sk, skb);
6001 /* Process urgent data. */
6002 tcp_urg(sk, skb, th);
6004 /* step 7: process the segment text */
6005 tcp_data_queue(sk, skb);
6007 tcp_data_snd_check(sk);
6008 tcp_ack_snd_check(sk);
6012 reason = SKB_DROP_REASON_TCP_CSUM;
6013 trace_tcp_bad_csum(skb);
6014 TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS);
6015 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
6018 tcp_drop_reason(sk, skb, reason);
6020 EXPORT_SYMBOL(tcp_rcv_established);
6022 void tcp_init_transfer(struct sock *sk, int bpf_op, struct sk_buff *skb)
6024 struct inet_connection_sock *icsk = inet_csk(sk);
6025 struct tcp_sock *tp = tcp_sk(sk);
6028 icsk->icsk_af_ops->rebuild_header(sk);
6029 tcp_init_metrics(sk);
6031 /* Initialize the congestion window to start the transfer.
6032 * Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
6033 * retransmitted. In light of RFC6298 more aggressive 1sec
6034 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
6035 * retransmission has occurred.
6037 if (tp->total_retrans > 1 && tp->undo_marker)
6038 tcp_snd_cwnd_set(tp, 1);
6040 tcp_snd_cwnd_set(tp, tcp_init_cwnd(tp, __sk_dst_get(sk)));
6041 tp->snd_cwnd_stamp = tcp_jiffies32;
6043 bpf_skops_established(sk, bpf_op, skb);
6044 /* Initialize congestion control unless BPF initialized it already: */
6045 if (!icsk->icsk_ca_initialized)
6046 tcp_init_congestion_control(sk);
6047 tcp_init_buffer_space(sk);
6050 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
6052 struct tcp_sock *tp = tcp_sk(sk);
6053 struct inet_connection_sock *icsk = inet_csk(sk);
6055 tcp_set_state(sk, TCP_ESTABLISHED);
6056 icsk->icsk_ack.lrcvtime = tcp_jiffies32;
6059 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
6060 security_inet_conn_established(sk, skb);
6061 sk_mark_napi_id(sk, skb);
6064 tcp_init_transfer(sk, BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB, skb);
6066 /* Prevent spurious tcp_cwnd_restart() on first data
6069 tp->lsndtime = tcp_jiffies32;
6071 if (sock_flag(sk, SOCK_KEEPOPEN))
6072 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
6074 if (!tp->rx_opt.snd_wscale)
6075 __tcp_fast_path_on(tp, tp->snd_wnd);
6080 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
6081 struct tcp_fastopen_cookie *cookie)
6083 struct tcp_sock *tp = tcp_sk(sk);
6084 struct sk_buff *data = tp->syn_data ? tcp_rtx_queue_head(sk) : NULL;
6085 u16 mss = tp->rx_opt.mss_clamp, try_exp = 0;
6086 bool syn_drop = false;
6088 if (mss == tp->rx_opt.user_mss) {
6089 struct tcp_options_received opt;
6091 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
6092 tcp_clear_options(&opt);
6093 opt.user_mss = opt.mss_clamp = 0;
6094 tcp_parse_options(sock_net(sk), synack, &opt, 0, NULL);
6095 mss = opt.mss_clamp;
6098 if (!tp->syn_fastopen) {
6099 /* Ignore an unsolicited cookie */
6101 } else if (tp->total_retrans) {
6102 /* SYN timed out and the SYN-ACK neither has a cookie nor
6103 * acknowledges data. Presumably the remote received only
6104 * the retransmitted (regular) SYNs: either the original
6105 * SYN-data or the corresponding SYN-ACK was dropped.
6107 syn_drop = (cookie->len < 0 && data);
6108 } else if (cookie->len < 0 && !tp->syn_data) {
6109 /* We requested a cookie but didn't get it. If we did not use
6110 * the (old) exp opt format then try so next time (try_exp=1).
6111 * Otherwise we go back to use the RFC7413 opt (try_exp=2).
6113 try_exp = tp->syn_fastopen_exp ? 2 : 1;
6116 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp);
6118 if (data) { /* Retransmit unacked data in SYN */
6119 if (tp->total_retrans)
6120 tp->fastopen_client_fail = TFO_SYN_RETRANSMITTED;
6122 tp->fastopen_client_fail = TFO_DATA_NOT_ACKED;
6123 skb_rbtree_walk_from(data)
6124 tcp_mark_skb_lost(sk, data);
6125 tcp_xmit_retransmit_queue(sk);
6126 NET_INC_STATS(sock_net(sk),
6127 LINUX_MIB_TCPFASTOPENACTIVEFAIL);
6130 tp->syn_data_acked = tp->syn_data;
6131 if (tp->syn_data_acked) {
6132 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
6133 /* SYN-data is counted as two separate packets in tcp_ack() */
6134 if (tp->delivered > 1)
6138 tcp_fastopen_add_skb(sk, synack);
6143 static void smc_check_reset_syn(struct tcp_sock *tp)
6145 #if IS_ENABLED(CONFIG_SMC)
6146 if (static_branch_unlikely(&tcp_have_smc)) {
6147 if (tp->syn_smc && !tp->rx_opt.smc_ok)
6153 static void tcp_try_undo_spurious_syn(struct sock *sk)
6155 struct tcp_sock *tp = tcp_sk(sk);
6158 /* undo_marker is set when SYN or SYNACK times out. The timeout is
6159 * spurious if the ACK's timestamp option echo value matches the
6160 * original SYN timestamp.
6162 syn_stamp = tp->retrans_stamp;
6163 if (tp->undo_marker && syn_stamp && tp->rx_opt.saw_tstamp &&
6164 syn_stamp == tp->rx_opt.rcv_tsecr)
6165 tp->undo_marker = 0;
6168 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
6169 const struct tcphdr *th)
6171 struct inet_connection_sock *icsk = inet_csk(sk);
6172 struct tcp_sock *tp = tcp_sk(sk);
6173 struct tcp_fastopen_cookie foc = { .len = -1 };
6174 int saved_clamp = tp->rx_opt.mss_clamp;
6178 tcp_parse_options(sock_net(sk), skb, &tp->rx_opt, 0, &foc);
6179 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
6180 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
6184 * "If the state is SYN-SENT then
6185 * first check the ACK bit
6186 * If the ACK bit is set
6187 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
6188 * a reset (unless the RST bit is set, if so drop
6189 * the segment and return)"
6191 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
6192 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
6193 /* Previous FIN/ACK or RST/ACK might be ignored. */
6194 if (icsk->icsk_retransmits == 0)
6195 inet_csk_reset_xmit_timer(sk,
6197 TCP_TIMEOUT_MIN, TCP_RTO_MAX);
6198 goto reset_and_undo;
6201 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
6202 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
6203 tcp_time_stamp(tp))) {
6204 NET_INC_STATS(sock_net(sk),
6205 LINUX_MIB_PAWSACTIVEREJECTED);
6206 goto reset_and_undo;
6209 /* Now ACK is acceptable.
6211 * "If the RST bit is set
6212 * If the ACK was acceptable then signal the user "error:
6213 * connection reset", drop the segment, enter CLOSED state,
6214 * delete TCB, and return."
6225 * "fifth, if neither of the SYN or RST bits is set then
6226 * drop the segment and return."
6232 SKB_DR_SET(reason, TCP_FLAGS);
6233 goto discard_and_undo;
6236 * "If the SYN bit is on ...
6237 * are acceptable then ...
6238 * (our SYN has been ACKed), change the connection
6239 * state to ESTABLISHED..."
6242 tcp_ecn_rcv_synack(tp, th);
6244 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
6245 tcp_try_undo_spurious_syn(sk);
6246 tcp_ack(sk, skb, FLAG_SLOWPATH);
6248 /* Ok.. it's good. Set up sequence numbers and
6249 * move to established.
6251 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1);
6252 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
6254 /* RFC1323: The window in SYN & SYN/ACK segments is
6257 tp->snd_wnd = ntohs(th->window);
6259 if (!tp->rx_opt.wscale_ok) {
6260 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
6261 tp->window_clamp = min(tp->window_clamp, 65535U);
6264 if (tp->rx_opt.saw_tstamp) {
6265 tp->rx_opt.tstamp_ok = 1;
6266 tp->tcp_header_len =
6267 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
6268 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
6269 tcp_store_ts_recent(tp);
6271 tp->tcp_header_len = sizeof(struct tcphdr);
6274 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
6275 tcp_initialize_rcv_mss(sk);
6277 /* Remember, tcp_poll() does not lock socket!
6278 * Change state from SYN-SENT only after copied_seq
6279 * is initialized. */
6280 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6282 smc_check_reset_syn(tp);
6286 tcp_finish_connect(sk, skb);
6288 fastopen_fail = (tp->syn_fastopen || tp->syn_data) &&
6289 tcp_rcv_fastopen_synack(sk, skb, &foc);
6291 if (!sock_flag(sk, SOCK_DEAD)) {
6292 sk->sk_state_change(sk);
6293 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
6297 if (sk->sk_write_pending ||
6298 icsk->icsk_accept_queue.rskq_defer_accept ||
6299 inet_csk_in_pingpong_mode(sk)) {
6300 /* Save one ACK. Data will be ready after
6301 * several ticks, if write_pending is set.
6303 * It may be deleted, but with this feature tcpdumps
6304 * look so _wonderfully_ clever, that I was not able
6305 * to stand against the temptation 8) --ANK
6307 inet_csk_schedule_ack(sk);
6308 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
6309 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
6310 TCP_DELACK_MAX, TCP_RTO_MAX);
6317 /* No ACK in the segment */
6321 * "If the RST bit is set
6323 * Otherwise (no ACK) drop the segment and return."
6325 SKB_DR_SET(reason, TCP_RESET);
6326 goto discard_and_undo;
6330 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
6331 tcp_paws_reject(&tp->rx_opt, 0)) {
6332 SKB_DR_SET(reason, TCP_RFC7323_PAWS);
6333 goto discard_and_undo;
6336 /* We see SYN without ACK. It is attempt of
6337 * simultaneous connect with crossed SYNs.
6338 * Particularly, it can be connect to self.
6340 tcp_set_state(sk, TCP_SYN_RECV);
6342 if (tp->rx_opt.saw_tstamp) {
6343 tp->rx_opt.tstamp_ok = 1;
6344 tcp_store_ts_recent(tp);
6345 tp->tcp_header_len =
6346 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
6348 tp->tcp_header_len = sizeof(struct tcphdr);
6351 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1);
6352 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6353 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
6355 /* RFC1323: The window in SYN & SYN/ACK segments is
6358 tp->snd_wnd = ntohs(th->window);
6359 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
6360 tp->max_window = tp->snd_wnd;
6362 tcp_ecn_rcv_syn(tp, th);
6365 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
6366 tcp_initialize_rcv_mss(sk);
6368 tcp_send_synack(sk);
6370 /* Note, we could accept data and URG from this segment.
6371 * There are no obstacles to make this (except that we must
6372 * either change tcp_recvmsg() to prevent it from returning data
6373 * before 3WHS completes per RFC793, or employ TCP Fast Open).
6375 * However, if we ignore data in ACKless segments sometimes,
6376 * we have no reasons to accept it sometimes.
6377 * Also, seems the code doing it in step6 of tcp_rcv_state_process
6378 * is not flawless. So, discard packet for sanity.
6379 * Uncomment this return to process the data.
6386 /* "fifth, if neither of the SYN or RST bits is set then
6387 * drop the segment and return."
6391 tcp_clear_options(&tp->rx_opt);
6392 tp->rx_opt.mss_clamp = saved_clamp;
6393 tcp_drop_reason(sk, skb, reason);
6397 tcp_clear_options(&tp->rx_opt);
6398 tp->rx_opt.mss_clamp = saved_clamp;
6402 static void tcp_rcv_synrecv_state_fastopen(struct sock *sk)
6404 struct request_sock *req;
6406 /* If we are still handling the SYNACK RTO, see if timestamp ECR allows
6407 * undo. If peer SACKs triggered fast recovery, we can't undo here.
6409 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
6410 tcp_try_undo_loss(sk, false);
6412 /* Reset rtx states to prevent spurious retransmits_timed_out() */
6413 tcp_sk(sk)->retrans_stamp = 0;
6414 inet_csk(sk)->icsk_retransmits = 0;
6416 /* Once we leave TCP_SYN_RECV or TCP_FIN_WAIT_1,
6417 * we no longer need req so release it.
6419 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk,
6420 lockdep_sock_is_held(sk));
6421 reqsk_fastopen_remove(sk, req, false);
6423 /* Re-arm the timer because data may have been sent out.
6424 * This is similar to the regular data transmission case
6425 * when new data has just been ack'ed.
6427 * (TFO) - we could try to be more aggressive and
6428 * retransmitting any data sooner based on when they
6435 * This function implements the receiving procedure of RFC 793 for
6436 * all states except ESTABLISHED and TIME_WAIT.
6437 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
6438 * address independent.
6441 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb)
6443 struct tcp_sock *tp = tcp_sk(sk);
6444 struct inet_connection_sock *icsk = inet_csk(sk);
6445 const struct tcphdr *th = tcp_hdr(skb);
6446 struct request_sock *req;
6451 switch (sk->sk_state) {
6453 SKB_DR_SET(reason, TCP_CLOSE);
6461 SKB_DR_SET(reason, TCP_RESET);
6466 SKB_DR_SET(reason, TCP_FLAGS);
6469 /* It is possible that we process SYN packets from backlog,
6470 * so we need to make sure to disable BH and RCU right there.
6474 acceptable = icsk->icsk_af_ops->conn_request(sk, skb) >= 0;
6483 SKB_DR_SET(reason, TCP_FLAGS);
6487 tp->rx_opt.saw_tstamp = 0;
6488 tcp_mstamp_refresh(tp);
6489 queued = tcp_rcv_synsent_state_process(sk, skb, th);
6493 /* Do step6 onward by hand. */
6494 tcp_urg(sk, skb, th);
6496 tcp_data_snd_check(sk);
6500 tcp_mstamp_refresh(tp);
6501 tp->rx_opt.saw_tstamp = 0;
6502 req = rcu_dereference_protected(tp->fastopen_rsk,
6503 lockdep_sock_is_held(sk));
6507 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
6508 sk->sk_state != TCP_FIN_WAIT1);
6510 if (!tcp_check_req(sk, skb, req, true, &req_stolen)) {
6511 SKB_DR_SET(reason, TCP_FASTOPEN);
6516 if (!th->ack && !th->rst && !th->syn) {
6517 SKB_DR_SET(reason, TCP_FLAGS);
6520 if (!tcp_validate_incoming(sk, skb, th, 0))
6523 /* step 5: check the ACK field */
6524 acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
6525 FLAG_UPDATE_TS_RECENT |
6526 FLAG_NO_CHALLENGE_ACK) > 0;
6529 if (sk->sk_state == TCP_SYN_RECV)
6530 return 1; /* send one RST */
6531 tcp_send_challenge_ack(sk);
6532 SKB_DR_SET(reason, TCP_OLD_ACK);
6535 switch (sk->sk_state) {
6537 tp->delivered++; /* SYN-ACK delivery isn't tracked in tcp_ack */
6539 tcp_synack_rtt_meas(sk, req);
6542 tcp_rcv_synrecv_state_fastopen(sk);
6544 tcp_try_undo_spurious_syn(sk);
6545 tp->retrans_stamp = 0;
6546 tcp_init_transfer(sk, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB,
6548 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6551 tcp_set_state(sk, TCP_ESTABLISHED);
6552 sk->sk_state_change(sk);
6554 /* Note, that this wakeup is only for marginal crossed SYN case.
6555 * Passively open sockets are not waked up, because
6556 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
6559 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
6561 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
6562 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
6563 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
6565 if (tp->rx_opt.tstamp_ok)
6566 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
6568 if (!inet_csk(sk)->icsk_ca_ops->cong_control)
6569 tcp_update_pacing_rate(sk);
6571 /* Prevent spurious tcp_cwnd_restart() on first data packet */
6572 tp->lsndtime = tcp_jiffies32;
6574 tcp_initialize_rcv_mss(sk);
6575 tcp_fast_path_on(tp);
6578 case TCP_FIN_WAIT1: {
6582 tcp_rcv_synrecv_state_fastopen(sk);
6584 if (tp->snd_una != tp->write_seq)
6587 tcp_set_state(sk, TCP_FIN_WAIT2);
6588 WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | SEND_SHUTDOWN);
6592 if (!sock_flag(sk, SOCK_DEAD)) {
6593 /* Wake up lingering close() */
6594 sk->sk_state_change(sk);
6598 if (tp->linger2 < 0) {
6600 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6603 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6604 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6605 /* Receive out of order FIN after close() */
6606 if (tp->syn_fastopen && th->fin)
6607 tcp_fastopen_active_disable(sk);
6609 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6613 tmo = tcp_fin_time(sk);
6614 if (tmo > TCP_TIMEWAIT_LEN) {
6615 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
6616 } else if (th->fin || sock_owned_by_user(sk)) {
6617 /* Bad case. We could lose such FIN otherwise.
6618 * It is not a big problem, but it looks confusing
6619 * and not so rare event. We still can lose it now,
6620 * if it spins in bh_lock_sock(), but it is really
6623 inet_csk_reset_keepalive_timer(sk, tmo);
6625 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
6632 if (tp->snd_una == tp->write_seq) {
6633 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
6639 if (tp->snd_una == tp->write_seq) {
6640 tcp_update_metrics(sk);
6647 /* step 6: check the URG bit */
6648 tcp_urg(sk, skb, th);
6650 /* step 7: process the segment text */
6651 switch (sk->sk_state) {
6652 case TCP_CLOSE_WAIT:
6655 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
6656 /* If a subflow has been reset, the packet should not
6657 * continue to be processed, drop the packet.
6659 if (sk_is_mptcp(sk) && !mptcp_incoming_options(sk, skb))
6666 /* RFC 793 says to queue data in these states,
6667 * RFC 1122 says we MUST send a reset.
6668 * BSD 4.4 also does reset.
6670 if (sk->sk_shutdown & RCV_SHUTDOWN) {
6671 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6672 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6673 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6679 case TCP_ESTABLISHED:
6680 tcp_data_queue(sk, skb);
6685 /* tcp_data could move socket to TIME-WAIT */
6686 if (sk->sk_state != TCP_CLOSE) {
6687 tcp_data_snd_check(sk);
6688 tcp_ack_snd_check(sk);
6693 tcp_drop_reason(sk, skb, reason);
6701 EXPORT_SYMBOL(tcp_rcv_state_process);
6703 static inline void pr_drop_req(struct request_sock *req, __u16 port, int family)
6705 struct inet_request_sock *ireq = inet_rsk(req);
6707 if (family == AF_INET)
6708 net_dbg_ratelimited("drop open request from %pI4/%u\n",
6709 &ireq->ir_rmt_addr, port);
6710 #if IS_ENABLED(CONFIG_IPV6)
6711 else if (family == AF_INET6)
6712 net_dbg_ratelimited("drop open request from %pI6/%u\n",
6713 &ireq->ir_v6_rmt_addr, port);
6717 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
6719 * If we receive a SYN packet with these bits set, it means a
6720 * network is playing bad games with TOS bits. In order to
6721 * avoid possible false congestion notifications, we disable
6722 * TCP ECN negotiation.
6724 * Exception: tcp_ca wants ECN. This is required for DCTCP
6725 * congestion control: Linux DCTCP asserts ECT on all packets,
6726 * including SYN, which is most optimal solution; however,
6727 * others, such as FreeBSD do not.
6729 * Exception: At least one of the reserved bits of the TCP header (th->res1) is
6730 * set, indicating the use of a future TCP extension (such as AccECN). See
6731 * RFC8311 §4.3 which updates RFC3168 to allow the development of such
6734 static void tcp_ecn_create_request(struct request_sock *req,
6735 const struct sk_buff *skb,
6736 const struct sock *listen_sk,
6737 const struct dst_entry *dst)
6739 const struct tcphdr *th = tcp_hdr(skb);
6740 const struct net *net = sock_net(listen_sk);
6741 bool th_ecn = th->ece && th->cwr;
6748 ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield);
6749 ecn_ok_dst = dst_feature(dst, DST_FEATURE_ECN_MASK);
6750 ecn_ok = READ_ONCE(net->ipv4.sysctl_tcp_ecn) || ecn_ok_dst;
6752 if (((!ect || th->res1) && ecn_ok) || tcp_ca_needs_ecn(listen_sk) ||
6753 (ecn_ok_dst & DST_FEATURE_ECN_CA) ||
6754 tcp_bpf_ca_needs_ecn((struct sock *)req))
6755 inet_rsk(req)->ecn_ok = 1;
6758 static void tcp_openreq_init(struct request_sock *req,
6759 const struct tcp_options_received *rx_opt,
6760 struct sk_buff *skb, const struct sock *sk)
6762 struct inet_request_sock *ireq = inet_rsk(req);
6764 req->rsk_rcv_wnd = 0; /* So that tcp_send_synack() knows! */
6765 tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq;
6766 tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
6767 tcp_rsk(req)->snt_synack = 0;
6768 tcp_rsk(req)->last_oow_ack_time = 0;
6769 req->mss = rx_opt->mss_clamp;
6770 req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0;
6771 ireq->tstamp_ok = rx_opt->tstamp_ok;
6772 ireq->sack_ok = rx_opt->sack_ok;
6773 ireq->snd_wscale = rx_opt->snd_wscale;
6774 ireq->wscale_ok = rx_opt->wscale_ok;
6777 ireq->ir_rmt_port = tcp_hdr(skb)->source;
6778 ireq->ir_num = ntohs(tcp_hdr(skb)->dest);
6779 ireq->ir_mark = inet_request_mark(sk, skb);
6780 #if IS_ENABLED(CONFIG_SMC)
6781 ireq->smc_ok = rx_opt->smc_ok && !(tcp_sk(sk)->smc_hs_congested &&
6782 tcp_sk(sk)->smc_hs_congested(sk));
6786 struct request_sock *inet_reqsk_alloc(const struct request_sock_ops *ops,
6787 struct sock *sk_listener,
6788 bool attach_listener)
6790 struct request_sock *req = reqsk_alloc(ops, sk_listener,
6794 struct inet_request_sock *ireq = inet_rsk(req);
6796 ireq->ireq_opt = NULL;
6797 #if IS_ENABLED(CONFIG_IPV6)
6798 ireq->pktopts = NULL;
6800 atomic64_set(&ireq->ir_cookie, 0);
6801 ireq->ireq_state = TCP_NEW_SYN_RECV;
6802 write_pnet(&ireq->ireq_net, sock_net(sk_listener));
6803 ireq->ireq_family = sk_listener->sk_family;
6804 req->timeout = TCP_TIMEOUT_INIT;
6809 EXPORT_SYMBOL(inet_reqsk_alloc);
6812 * Return true if a syncookie should be sent
6814 static bool tcp_syn_flood_action(const struct sock *sk, const char *proto)
6816 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
6817 const char *msg = "Dropping request";
6818 struct net *net = sock_net(sk);
6819 bool want_cookie = false;
6822 syncookies = READ_ONCE(net->ipv4.sysctl_tcp_syncookies);
6824 #ifdef CONFIG_SYN_COOKIES
6826 msg = "Sending cookies";
6828 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES);
6831 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP);
6833 if (!queue->synflood_warned && syncookies != 2 &&
6834 xchg(&queue->synflood_warned, 1) == 0)
6835 net_info_ratelimited("%s: Possible SYN flooding on port %d. %s. Check SNMP counters.\n",
6836 proto, sk->sk_num, msg);
6841 static void tcp_reqsk_record_syn(const struct sock *sk,
6842 struct request_sock *req,
6843 const struct sk_buff *skb)
6845 if (tcp_sk(sk)->save_syn) {
6846 u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb);
6847 struct saved_syn *saved_syn;
6851 if (tcp_sk(sk)->save_syn == 2) { /* Save full header. */
6852 base = skb_mac_header(skb);
6853 mac_hdrlen = skb_mac_header_len(skb);
6856 base = skb_network_header(skb);
6860 saved_syn = kmalloc(struct_size(saved_syn, data, len),
6863 saved_syn->mac_hdrlen = mac_hdrlen;
6864 saved_syn->network_hdrlen = skb_network_header_len(skb);
6865 saved_syn->tcp_hdrlen = tcp_hdrlen(skb);
6866 memcpy(saved_syn->data, base, len);
6867 req->saved_syn = saved_syn;
6872 /* If a SYN cookie is required and supported, returns a clamped MSS value to be
6873 * used for SYN cookie generation.
6875 u16 tcp_get_syncookie_mss(struct request_sock_ops *rsk_ops,
6876 const struct tcp_request_sock_ops *af_ops,
6877 struct sock *sk, struct tcphdr *th)
6879 struct tcp_sock *tp = tcp_sk(sk);
6882 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_syncookies) != 2 &&
6883 !inet_csk_reqsk_queue_is_full(sk))
6886 if (!tcp_syn_flood_action(sk, rsk_ops->slab_name))
6889 if (sk_acceptq_is_full(sk)) {
6890 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
6894 mss = tcp_parse_mss_option(th, tp->rx_opt.user_mss);
6896 mss = af_ops->mss_clamp;
6900 EXPORT_SYMBOL_GPL(tcp_get_syncookie_mss);
6902 int tcp_conn_request(struct request_sock_ops *rsk_ops,
6903 const struct tcp_request_sock_ops *af_ops,
6904 struct sock *sk, struct sk_buff *skb)
6906 struct tcp_fastopen_cookie foc = { .len = -1 };
6907 __u32 isn = TCP_SKB_CB(skb)->tcp_tw_isn;
6908 struct tcp_options_received tmp_opt;
6909 struct tcp_sock *tp = tcp_sk(sk);
6910 struct net *net = sock_net(sk);
6911 struct sock *fastopen_sk = NULL;
6912 struct request_sock *req;
6913 bool want_cookie = false;
6914 struct dst_entry *dst;
6918 syncookies = READ_ONCE(net->ipv4.sysctl_tcp_syncookies);
6920 /* TW buckets are converted to open requests without
6921 * limitations, they conserve resources and peer is
6922 * evidently real one.
6924 if ((syncookies == 2 || inet_csk_reqsk_queue_is_full(sk)) && !isn) {
6925 want_cookie = tcp_syn_flood_action(sk, rsk_ops->slab_name);
6930 if (sk_acceptq_is_full(sk)) {
6931 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
6935 req = inet_reqsk_alloc(rsk_ops, sk, !want_cookie);
6939 req->syncookie = want_cookie;
6940 tcp_rsk(req)->af_specific = af_ops;
6941 tcp_rsk(req)->ts_off = 0;
6942 #if IS_ENABLED(CONFIG_MPTCP)
6943 tcp_rsk(req)->is_mptcp = 0;
6946 tcp_clear_options(&tmp_opt);
6947 tmp_opt.mss_clamp = af_ops->mss_clamp;
6948 tmp_opt.user_mss = tp->rx_opt.user_mss;
6949 tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0,
6950 want_cookie ? NULL : &foc);
6952 if (want_cookie && !tmp_opt.saw_tstamp)
6953 tcp_clear_options(&tmp_opt);
6955 if (IS_ENABLED(CONFIG_SMC) && want_cookie)
6958 tmp_opt.tstamp_ok = tmp_opt.saw_tstamp;
6959 tcp_openreq_init(req, &tmp_opt, skb, sk);
6960 inet_rsk(req)->no_srccheck = inet_sk(sk)->transparent;
6962 /* Note: tcp_v6_init_req() might override ir_iif for link locals */
6963 inet_rsk(req)->ir_iif = inet_request_bound_dev_if(sk, skb);
6965 dst = af_ops->route_req(sk, skb, &fl, req);
6969 if (tmp_opt.tstamp_ok)
6970 tcp_rsk(req)->ts_off = af_ops->init_ts_off(net, skb);
6972 if (!want_cookie && !isn) {
6973 int max_syn_backlog = READ_ONCE(net->ipv4.sysctl_max_syn_backlog);
6975 /* Kill the following clause, if you dislike this way. */
6977 (max_syn_backlog - inet_csk_reqsk_queue_len(sk) <
6978 (max_syn_backlog >> 2)) &&
6979 !tcp_peer_is_proven(req, dst)) {
6980 /* Without syncookies last quarter of
6981 * backlog is filled with destinations,
6982 * proven to be alive.
6983 * It means that we continue to communicate
6984 * to destinations, already remembered
6985 * to the moment of synflood.
6987 pr_drop_req(req, ntohs(tcp_hdr(skb)->source),
6989 goto drop_and_release;
6992 isn = af_ops->init_seq(skb);
6995 tcp_ecn_create_request(req, skb, sk, dst);
6998 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss);
6999 if (!tmp_opt.tstamp_ok)
7000 inet_rsk(req)->ecn_ok = 0;
7003 tcp_rsk(req)->snt_isn = isn;
7004 tcp_rsk(req)->txhash = net_tx_rndhash();
7005 tcp_rsk(req)->syn_tos = TCP_SKB_CB(skb)->ip_dsfield;
7006 tcp_openreq_init_rwin(req, sk, dst);
7007 sk_rx_queue_set(req_to_sk(req), skb);
7009 tcp_reqsk_record_syn(sk, req, skb);
7010 fastopen_sk = tcp_try_fastopen(sk, skb, req, &foc, dst);
7013 af_ops->send_synack(fastopen_sk, dst, &fl, req,
7014 &foc, TCP_SYNACK_FASTOPEN, skb);
7015 /* Add the child socket directly into the accept queue */
7016 if (!inet_csk_reqsk_queue_add(sk, req, fastopen_sk)) {
7017 reqsk_fastopen_remove(fastopen_sk, req, false);
7018 bh_unlock_sock(fastopen_sk);
7019 sock_put(fastopen_sk);
7022 sk->sk_data_ready(sk);
7023 bh_unlock_sock(fastopen_sk);
7024 sock_put(fastopen_sk);
7026 tcp_rsk(req)->tfo_listener = false;
7028 req->timeout = tcp_timeout_init((struct sock *)req);
7029 inet_csk_reqsk_queue_hash_add(sk, req, req->timeout);
7031 af_ops->send_synack(sk, dst, &fl, req, &foc,
7032 !want_cookie ? TCP_SYNACK_NORMAL :
7051 EXPORT_SYMBOL(tcp_conn_request);