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, 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, 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;
504 if (room > 0 && !tcp_under_memory_pressure(sk)) {
505 unsigned int truesize = truesize_adjust(adjust, skb);
508 /* Check #2. Increase window, if skb with such overhead
509 * will fit to rcvbuf in future.
511 if (tcp_win_from_space(sk, truesize) <= skb->len)
512 incr = 2 * tp->advmss;
514 incr = __tcp_grow_window(sk, skb, truesize);
517 incr = max_t(int, incr, 2 * skb->len);
518 tp->rcv_ssthresh += min(room, incr);
519 inet_csk(sk)->icsk_ack.quick |= 1;
524 /* 3. Try to fixup all. It is made immediately after connection enters
527 static void tcp_init_buffer_space(struct sock *sk)
529 int tcp_app_win = sock_net(sk)->ipv4.sysctl_tcp_app_win;
530 struct tcp_sock *tp = tcp_sk(sk);
533 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
534 tcp_sndbuf_expand(sk);
536 tcp_mstamp_refresh(tp);
537 tp->rcvq_space.time = tp->tcp_mstamp;
538 tp->rcvq_space.seq = tp->copied_seq;
540 maxwin = tcp_full_space(sk);
542 if (tp->window_clamp >= maxwin) {
543 tp->window_clamp = maxwin;
545 if (tcp_app_win && maxwin > 4 * tp->advmss)
546 tp->window_clamp = max(maxwin -
547 (maxwin >> tcp_app_win),
551 /* Force reservation of one segment. */
553 tp->window_clamp > 2 * tp->advmss &&
554 tp->window_clamp + tp->advmss > maxwin)
555 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
557 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
558 tp->snd_cwnd_stamp = tcp_jiffies32;
559 tp->rcvq_space.space = min3(tp->rcv_ssthresh, tp->rcv_wnd,
560 (u32)TCP_INIT_CWND * tp->advmss);
563 /* 4. Recalculate window clamp after socket hit its memory bounds. */
564 static void tcp_clamp_window(struct sock *sk)
566 struct tcp_sock *tp = tcp_sk(sk);
567 struct inet_connection_sock *icsk = inet_csk(sk);
568 struct net *net = sock_net(sk);
570 icsk->icsk_ack.quick = 0;
572 if (sk->sk_rcvbuf < net->ipv4.sysctl_tcp_rmem[2] &&
573 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
574 !tcp_under_memory_pressure(sk) &&
575 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
576 WRITE_ONCE(sk->sk_rcvbuf,
577 min(atomic_read(&sk->sk_rmem_alloc),
578 net->ipv4.sysctl_tcp_rmem[2]));
580 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
581 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
584 /* Initialize RCV_MSS value.
585 * RCV_MSS is an our guess about MSS used by the peer.
586 * We haven't any direct information about the MSS.
587 * It's better to underestimate the RCV_MSS rather than overestimate.
588 * Overestimations make us ACKing less frequently than needed.
589 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
591 void tcp_initialize_rcv_mss(struct sock *sk)
593 const struct tcp_sock *tp = tcp_sk(sk);
594 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
596 hint = min(hint, tp->rcv_wnd / 2);
597 hint = min(hint, TCP_MSS_DEFAULT);
598 hint = max(hint, TCP_MIN_MSS);
600 inet_csk(sk)->icsk_ack.rcv_mss = hint;
602 EXPORT_SYMBOL(tcp_initialize_rcv_mss);
604 /* Receiver "autotuning" code.
606 * The algorithm for RTT estimation w/o timestamps is based on
607 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
608 * <https://public.lanl.gov/radiant/pubs.html#DRS>
610 * More detail on this code can be found at
611 * <http://staff.psc.edu/jheffner/>,
612 * though this reference is out of date. A new paper
615 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
617 u32 new_sample = tp->rcv_rtt_est.rtt_us;
620 if (new_sample != 0) {
621 /* If we sample in larger samples in the non-timestamp
622 * case, we could grossly overestimate the RTT especially
623 * with chatty applications or bulk transfer apps which
624 * are stalled on filesystem I/O.
626 * Also, since we are only going for a minimum in the
627 * non-timestamp case, we do not smooth things out
628 * else with timestamps disabled convergence takes too
632 m -= (new_sample >> 3);
640 /* No previous measure. */
644 tp->rcv_rtt_est.rtt_us = new_sample;
647 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
651 if (tp->rcv_rtt_est.time == 0)
653 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
655 delta_us = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcv_rtt_est.time);
658 tcp_rcv_rtt_update(tp, delta_us, 1);
661 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
662 tp->rcv_rtt_est.time = tp->tcp_mstamp;
665 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
666 const struct sk_buff *skb)
668 struct tcp_sock *tp = tcp_sk(sk);
670 if (tp->rx_opt.rcv_tsecr == tp->rcv_rtt_last_tsecr)
672 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr;
674 if (TCP_SKB_CB(skb)->end_seq -
675 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss) {
676 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr;
679 if (likely(delta < INT_MAX / (USEC_PER_SEC / TCP_TS_HZ))) {
682 delta_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
683 tcp_rcv_rtt_update(tp, delta_us, 0);
689 * This function should be called every time data is copied to user space.
690 * It calculates the appropriate TCP receive buffer space.
692 void tcp_rcv_space_adjust(struct sock *sk)
694 struct tcp_sock *tp = tcp_sk(sk);
698 trace_tcp_rcv_space_adjust(sk);
700 tcp_mstamp_refresh(tp);
701 time = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcvq_space.time);
702 if (time < (tp->rcv_rtt_est.rtt_us >> 3) || tp->rcv_rtt_est.rtt_us == 0)
705 /* Number of bytes copied to user in last RTT */
706 copied = tp->copied_seq - tp->rcvq_space.seq;
707 if (copied <= tp->rcvq_space.space)
711 * copied = bytes received in previous RTT, our base window
712 * To cope with packet losses, we need a 2x factor
713 * To cope with slow start, and sender growing its cwin by 100 %
714 * every RTT, we need a 4x factor, because the ACK we are sending
715 * now is for the next RTT, not the current one :
716 * <prev RTT . ><current RTT .. ><next RTT .... >
719 if (sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf &&
720 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
724 /* minimal window to cope with packet losses, assuming
725 * steady state. Add some cushion because of small variations.
727 rcvwin = ((u64)copied << 1) + 16 * tp->advmss;
729 /* Accommodate for sender rate increase (eg. slow start) */
730 grow = rcvwin * (copied - tp->rcvq_space.space);
731 do_div(grow, tp->rcvq_space.space);
732 rcvwin += (grow << 1);
734 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
735 while (tcp_win_from_space(sk, rcvmem) < tp->advmss)
738 do_div(rcvwin, tp->advmss);
739 rcvbuf = min_t(u64, rcvwin * rcvmem,
740 sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
741 if (rcvbuf > sk->sk_rcvbuf) {
742 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
744 /* Make the window clamp follow along. */
745 tp->window_clamp = tcp_win_from_space(sk, rcvbuf);
748 tp->rcvq_space.space = copied;
751 tp->rcvq_space.seq = tp->copied_seq;
752 tp->rcvq_space.time = tp->tcp_mstamp;
755 /* There is something which you must keep in mind when you analyze the
756 * behavior of the tp->ato delayed ack timeout interval. When a
757 * connection starts up, we want to ack as quickly as possible. The
758 * problem is that "good" TCP's do slow start at the beginning of data
759 * transmission. The means that until we send the first few ACK's the
760 * sender will sit on his end and only queue most of his data, because
761 * he can only send snd_cwnd unacked packets at any given time. For
762 * each ACK we send, he increments snd_cwnd and transmits more of his
765 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
767 struct tcp_sock *tp = tcp_sk(sk);
768 struct inet_connection_sock *icsk = inet_csk(sk);
771 inet_csk_schedule_ack(sk);
773 tcp_measure_rcv_mss(sk, skb);
775 tcp_rcv_rtt_measure(tp);
779 if (!icsk->icsk_ack.ato) {
780 /* The _first_ data packet received, initialize
781 * delayed ACK engine.
783 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
784 icsk->icsk_ack.ato = TCP_ATO_MIN;
786 int m = now - icsk->icsk_ack.lrcvtime;
788 if (m <= TCP_ATO_MIN / 2) {
789 /* The fastest case is the first. */
790 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
791 } else if (m < icsk->icsk_ack.ato) {
792 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
793 if (icsk->icsk_ack.ato > icsk->icsk_rto)
794 icsk->icsk_ack.ato = icsk->icsk_rto;
795 } else if (m > icsk->icsk_rto) {
796 /* Too long gap. Apparently sender failed to
797 * restart window, so that we send ACKs quickly.
799 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
803 icsk->icsk_ack.lrcvtime = now;
805 tcp_ecn_check_ce(sk, skb);
808 tcp_grow_window(sk, skb, true);
811 /* Called to compute a smoothed rtt estimate. The data fed to this
812 * routine either comes from timestamps, or from segments that were
813 * known _not_ to have been retransmitted [see Karn/Partridge
814 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
815 * piece by Van Jacobson.
816 * NOTE: the next three routines used to be one big routine.
817 * To save cycles in the RFC 1323 implementation it was better to break
818 * it up into three procedures. -- erics
820 static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
822 struct tcp_sock *tp = tcp_sk(sk);
823 long m = mrtt_us; /* RTT */
824 u32 srtt = tp->srtt_us;
826 /* The following amusing code comes from Jacobson's
827 * article in SIGCOMM '88. Note that rtt and mdev
828 * are scaled versions of rtt and mean deviation.
829 * This is designed to be as fast as possible
830 * m stands for "measurement".
832 * On a 1990 paper the rto value is changed to:
833 * RTO = rtt + 4 * mdev
835 * Funny. This algorithm seems to be very broken.
836 * These formulae increase RTO, when it should be decreased, increase
837 * too slowly, when it should be increased quickly, decrease too quickly
838 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
839 * does not matter how to _calculate_ it. Seems, it was trap
840 * that VJ failed to avoid. 8)
843 m -= (srtt >> 3); /* m is now error in rtt est */
844 srtt += m; /* rtt = 7/8 rtt + 1/8 new */
846 m = -m; /* m is now abs(error) */
847 m -= (tp->mdev_us >> 2); /* similar update on mdev */
848 /* This is similar to one of Eifel findings.
849 * Eifel blocks mdev updates when rtt decreases.
850 * This solution is a bit different: we use finer gain
851 * for mdev in this case (alpha*beta).
852 * Like Eifel it also prevents growth of rto,
853 * but also it limits too fast rto decreases,
854 * happening in pure Eifel.
859 m -= (tp->mdev_us >> 2); /* similar update on mdev */
861 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */
862 if (tp->mdev_us > tp->mdev_max_us) {
863 tp->mdev_max_us = tp->mdev_us;
864 if (tp->mdev_max_us > tp->rttvar_us)
865 tp->rttvar_us = tp->mdev_max_us;
867 if (after(tp->snd_una, tp->rtt_seq)) {
868 if (tp->mdev_max_us < tp->rttvar_us)
869 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
870 tp->rtt_seq = tp->snd_nxt;
871 tp->mdev_max_us = tcp_rto_min_us(sk);
876 /* no previous measure. */
877 srtt = m << 3; /* take the measured time to be rtt */
878 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */
879 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
880 tp->mdev_max_us = tp->rttvar_us;
881 tp->rtt_seq = tp->snd_nxt;
885 tp->srtt_us = max(1U, srtt);
888 static void tcp_update_pacing_rate(struct sock *sk)
890 const struct tcp_sock *tp = tcp_sk(sk);
893 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
894 rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3);
896 /* current rate is (cwnd * mss) / srtt
897 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
898 * In Congestion Avoidance phase, set it to 120 % the current rate.
900 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
901 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
902 * end of slow start and should slow down.
904 if (tcp_snd_cwnd(tp) < tp->snd_ssthresh / 2)
905 rate *= sock_net(sk)->ipv4.sysctl_tcp_pacing_ss_ratio;
907 rate *= sock_net(sk)->ipv4.sysctl_tcp_pacing_ca_ratio;
909 rate *= max(tcp_snd_cwnd(tp), tp->packets_out);
911 if (likely(tp->srtt_us))
912 do_div(rate, tp->srtt_us);
914 /* WRITE_ONCE() is needed because sch_fq fetches sk_pacing_rate
915 * without any lock. We want to make sure compiler wont store
916 * intermediate values in this location.
918 WRITE_ONCE(sk->sk_pacing_rate, min_t(u64, rate,
919 sk->sk_max_pacing_rate));
922 /* Calculate rto without backoff. This is the second half of Van Jacobson's
923 * routine referred to above.
925 static void tcp_set_rto(struct sock *sk)
927 const struct tcp_sock *tp = tcp_sk(sk);
928 /* Old crap is replaced with new one. 8)
931 * 1. If rtt variance happened to be less 50msec, it is hallucination.
932 * It cannot be less due to utterly erratic ACK generation made
933 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
934 * to do with delayed acks, because at cwnd>2 true delack timeout
935 * is invisible. Actually, Linux-2.4 also generates erratic
936 * ACKs in some circumstances.
938 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
940 /* 2. Fixups made earlier cannot be right.
941 * If we do not estimate RTO correctly without them,
942 * all the algo is pure shit and should be replaced
943 * with correct one. It is exactly, which we pretend to do.
946 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
947 * guarantees that rto is higher.
952 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
954 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
957 cwnd = TCP_INIT_CWND;
958 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
961 struct tcp_sacktag_state {
962 /* Timestamps for earliest and latest never-retransmitted segment
963 * that was SACKed. RTO needs the earliest RTT to stay conservative,
964 * but congestion control should still get an accurate delay signal.
971 unsigned int mss_now;
972 struct rate_sample *rate;
975 /* Take a notice that peer is sending D-SACKs. Skip update of data delivery
976 * and spurious retransmission information if this DSACK is unlikely caused by
978 * - DSACKed sequence range is larger than maximum receiver's window.
979 * - Total no. of DSACKed segments exceed the total no. of retransmitted segs.
981 static u32 tcp_dsack_seen(struct tcp_sock *tp, u32 start_seq,
982 u32 end_seq, struct tcp_sacktag_state *state)
984 u32 seq_len, dup_segs = 1;
986 if (!before(start_seq, end_seq))
989 seq_len = end_seq - start_seq;
990 /* Dubious DSACK: DSACKed range greater than maximum advertised rwnd */
991 if (seq_len > tp->max_window)
993 if (seq_len > tp->mss_cache)
994 dup_segs = DIV_ROUND_UP(seq_len, tp->mss_cache);
995 else if (tp->tlp_high_seq && tp->tlp_high_seq == end_seq)
996 state->flag |= FLAG_DSACK_TLP;
998 tp->dsack_dups += dup_segs;
999 /* Skip the DSACK if dup segs weren't retransmitted by sender */
1000 if (tp->dsack_dups > tp->total_retrans)
1003 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
1004 /* We increase the RACK ordering window in rounds where we receive
1005 * DSACKs that may have been due to reordering causing RACK to trigger
1006 * a spurious fast recovery. Thus RACK ignores DSACKs that happen
1007 * without having seen reordering, or that match TLP probes (TLP
1008 * is timer-driven, not triggered by RACK).
1010 if (tp->reord_seen && !(state->flag & FLAG_DSACK_TLP))
1011 tp->rack.dsack_seen = 1;
1013 state->flag |= FLAG_DSACKING_ACK;
1014 /* A spurious retransmission is delivered */
1015 state->sack_delivered += dup_segs;
1020 /* It's reordering when higher sequence was delivered (i.e. sacked) before
1021 * some lower never-retransmitted sequence ("low_seq"). The maximum reordering
1022 * distance is approximated in full-mss packet distance ("reordering").
1024 static void tcp_check_sack_reordering(struct sock *sk, const u32 low_seq,
1027 struct tcp_sock *tp = tcp_sk(sk);
1028 const u32 mss = tp->mss_cache;
1031 fack = tcp_highest_sack_seq(tp);
1032 if (!before(low_seq, fack))
1035 metric = fack - low_seq;
1036 if ((metric > tp->reordering * mss) && mss) {
1037 #if FASTRETRANS_DEBUG > 1
1038 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
1039 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
1043 tp->undo_marker ? tp->undo_retrans : 0);
1045 tp->reordering = min_t(u32, (metric + mss - 1) / mss,
1046 sock_net(sk)->ipv4.sysctl_tcp_max_reordering);
1049 /* This exciting event is worth to be remembered. 8) */
1051 NET_INC_STATS(sock_net(sk),
1052 ts ? LINUX_MIB_TCPTSREORDER : LINUX_MIB_TCPSACKREORDER);
1055 /* This must be called before lost_out or retrans_out are updated
1056 * on a new loss, because we want to know if all skbs previously
1057 * known to be lost have already been retransmitted, indicating
1058 * that this newly lost skb is our next skb to retransmit.
1060 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
1062 if ((!tp->retransmit_skb_hint && tp->retrans_out >= tp->lost_out) ||
1063 (tp->retransmit_skb_hint &&
1064 before(TCP_SKB_CB(skb)->seq,
1065 TCP_SKB_CB(tp->retransmit_skb_hint)->seq)))
1066 tp->retransmit_skb_hint = skb;
1069 /* Sum the number of packets on the wire we have marked as lost, and
1070 * notify the congestion control module that the given skb was marked lost.
1072 static void tcp_notify_skb_loss_event(struct tcp_sock *tp, const struct sk_buff *skb)
1074 tp->lost += tcp_skb_pcount(skb);
1077 void tcp_mark_skb_lost(struct sock *sk, struct sk_buff *skb)
1079 __u8 sacked = TCP_SKB_CB(skb)->sacked;
1080 struct tcp_sock *tp = tcp_sk(sk);
1082 if (sacked & TCPCB_SACKED_ACKED)
1085 tcp_verify_retransmit_hint(tp, skb);
1086 if (sacked & TCPCB_LOST) {
1087 if (sacked & TCPCB_SACKED_RETRANS) {
1088 /* Account for retransmits that are lost again */
1089 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1090 tp->retrans_out -= tcp_skb_pcount(skb);
1091 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT,
1092 tcp_skb_pcount(skb));
1093 tcp_notify_skb_loss_event(tp, skb);
1096 tp->lost_out += tcp_skb_pcount(skb);
1097 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1098 tcp_notify_skb_loss_event(tp, skb);
1102 /* Updates the delivered and delivered_ce counts */
1103 static void tcp_count_delivered(struct tcp_sock *tp, u32 delivered,
1106 tp->delivered += delivered;
1108 tp->delivered_ce += delivered;
1111 /* This procedure tags the retransmission queue when SACKs arrive.
1113 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
1114 * Packets in queue with these bits set are counted in variables
1115 * sacked_out, retrans_out and lost_out, correspondingly.
1117 * Valid combinations are:
1118 * Tag InFlight Description
1119 * 0 1 - orig segment is in flight.
1120 * S 0 - nothing flies, orig reached receiver.
1121 * L 0 - nothing flies, orig lost by net.
1122 * R 2 - both orig and retransmit are in flight.
1123 * L|R 1 - orig is lost, retransmit is in flight.
1124 * S|R 1 - orig reached receiver, retrans is still in flight.
1125 * (L|S|R is logically valid, it could occur when L|R is sacked,
1126 * but it is equivalent to plain S and code short-curcuits it to S.
1127 * L|S is logically invalid, it would mean -1 packet in flight 8))
1129 * These 6 states form finite state machine, controlled by the following events:
1130 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
1131 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
1132 * 3. Loss detection event of two flavors:
1133 * A. Scoreboard estimator decided the packet is lost.
1134 * A'. Reno "three dupacks" marks head of queue lost.
1135 * B. SACK arrives sacking SND.NXT at the moment, when the
1136 * segment was retransmitted.
1137 * 4. D-SACK added new rule: D-SACK changes any tag to S.
1139 * It is pleasant to note, that state diagram turns out to be commutative,
1140 * so that we are allowed not to be bothered by order of our actions,
1141 * when multiple events arrive simultaneously. (see the function below).
1143 * Reordering detection.
1144 * --------------------
1145 * Reordering metric is maximal distance, which a packet can be displaced
1146 * in packet stream. With SACKs we can estimate it:
1148 * 1. SACK fills old hole and the corresponding segment was not
1149 * ever retransmitted -> reordering. Alas, we cannot use it
1150 * when segment was retransmitted.
1151 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1152 * for retransmitted and already SACKed segment -> reordering..
1153 * Both of these heuristics are not used in Loss state, when we cannot
1154 * account for retransmits accurately.
1156 * SACK block validation.
1157 * ----------------------
1159 * SACK block range validation checks that the received SACK block fits to
1160 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1161 * Note that SND.UNA is not included to the range though being valid because
1162 * it means that the receiver is rather inconsistent with itself reporting
1163 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1164 * perfectly valid, however, in light of RFC2018 which explicitly states
1165 * that "SACK block MUST reflect the newest segment. Even if the newest
1166 * segment is going to be discarded ...", not that it looks very clever
1167 * in case of head skb. Due to potentional receiver driven attacks, we
1168 * choose to avoid immediate execution of a walk in write queue due to
1169 * reneging and defer head skb's loss recovery to standard loss recovery
1170 * procedure that will eventually trigger (nothing forbids us doing this).
1172 * Implements also blockage to start_seq wrap-around. Problem lies in the
1173 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1174 * there's no guarantee that it will be before snd_nxt (n). The problem
1175 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1178 * <- outs wnd -> <- wrapzone ->
1179 * u e n u_w e_w s n_w
1181 * |<------------+------+----- TCP seqno space --------------+---------->|
1182 * ...-- <2^31 ->| |<--------...
1183 * ...---- >2^31 ------>| |<--------...
1185 * Current code wouldn't be vulnerable but it's better still to discard such
1186 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1187 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1188 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1189 * equal to the ideal case (infinite seqno space without wrap caused issues).
1191 * With D-SACK the lower bound is extended to cover sequence space below
1192 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1193 * again, D-SACK block must not to go across snd_una (for the same reason as
1194 * for the normal SACK blocks, explained above). But there all simplicity
1195 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1196 * fully below undo_marker they do not affect behavior in anyway and can
1197 * therefore be safely ignored. In rare cases (which are more or less
1198 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1199 * fragmentation and packet reordering past skb's retransmission. To consider
1200 * them correctly, the acceptable range must be extended even more though
1201 * the exact amount is rather hard to quantify. However, tp->max_window can
1202 * be used as an exaggerated estimate.
1204 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
1205 u32 start_seq, u32 end_seq)
1207 /* Too far in future, or reversed (interpretation is ambiguous) */
1208 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1211 /* Nasty start_seq wrap-around check (see comments above) */
1212 if (!before(start_seq, tp->snd_nxt))
1215 /* In outstanding window? ...This is valid exit for D-SACKs too.
1216 * start_seq == snd_una is non-sensical (see comments above)
1218 if (after(start_seq, tp->snd_una))
1221 if (!is_dsack || !tp->undo_marker)
1224 /* ...Then it's D-SACK, and must reside below snd_una completely */
1225 if (after(end_seq, tp->snd_una))
1228 if (!before(start_seq, tp->undo_marker))
1232 if (!after(end_seq, tp->undo_marker))
1235 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1236 * start_seq < undo_marker and end_seq >= undo_marker.
1238 return !before(start_seq, end_seq - tp->max_window);
1241 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1242 struct tcp_sack_block_wire *sp, int num_sacks,
1243 u32 prior_snd_una, struct tcp_sacktag_state *state)
1245 struct tcp_sock *tp = tcp_sk(sk);
1246 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1247 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1250 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1251 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1252 } else if (num_sacks > 1) {
1253 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1254 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1256 if (after(end_seq_0, end_seq_1) || before(start_seq_0, start_seq_1))
1258 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKOFORECV);
1263 dup_segs = tcp_dsack_seen(tp, start_seq_0, end_seq_0, state);
1264 if (!dup_segs) { /* Skip dubious DSACK */
1265 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKIGNOREDDUBIOUS);
1269 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECVSEGS, dup_segs);
1271 /* D-SACK for already forgotten data... Do dumb counting. */
1272 if (tp->undo_marker && tp->undo_retrans > 0 &&
1273 !after(end_seq_0, prior_snd_una) &&
1274 after(end_seq_0, tp->undo_marker))
1275 tp->undo_retrans = max_t(int, 0, tp->undo_retrans - dup_segs);
1280 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1281 * the incoming SACK may not exactly match but we can find smaller MSS
1282 * aligned portion of it that matches. Therefore we might need to fragment
1283 * which may fail and creates some hassle (caller must handle error case
1286 * FIXME: this could be merged to shift decision code
1288 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1289 u32 start_seq, u32 end_seq)
1293 unsigned int pkt_len;
1296 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1297 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1299 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1300 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1301 mss = tcp_skb_mss(skb);
1302 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1305 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1309 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1314 /* Round if necessary so that SACKs cover only full MSSes
1315 * and/or the remaining small portion (if present)
1317 if (pkt_len > mss) {
1318 unsigned int new_len = (pkt_len / mss) * mss;
1319 if (!in_sack && new_len < pkt_len)
1324 if (pkt_len >= skb->len && !in_sack)
1327 err = tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
1328 pkt_len, mss, GFP_ATOMIC);
1336 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1337 static u8 tcp_sacktag_one(struct sock *sk,
1338 struct tcp_sacktag_state *state, u8 sacked,
1339 u32 start_seq, u32 end_seq,
1340 int dup_sack, int pcount,
1343 struct tcp_sock *tp = tcp_sk(sk);
1345 /* Account D-SACK for retransmitted packet. */
1346 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1347 if (tp->undo_marker && tp->undo_retrans > 0 &&
1348 after(end_seq, tp->undo_marker))
1349 tp->undo_retrans = max_t(int, 0, tp->undo_retrans - pcount);
1350 if ((sacked & TCPCB_SACKED_ACKED) &&
1351 before(start_seq, state->reord))
1352 state->reord = start_seq;
1355 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1356 if (!after(end_seq, tp->snd_una))
1359 if (!(sacked & TCPCB_SACKED_ACKED)) {
1360 tcp_rack_advance(tp, sacked, end_seq, xmit_time);
1362 if (sacked & TCPCB_SACKED_RETRANS) {
1363 /* If the segment is not tagged as lost,
1364 * we do not clear RETRANS, believing
1365 * that retransmission is still in flight.
1367 if (sacked & TCPCB_LOST) {
1368 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1369 tp->lost_out -= pcount;
1370 tp->retrans_out -= pcount;
1373 if (!(sacked & TCPCB_RETRANS)) {
1374 /* New sack for not retransmitted frame,
1375 * which was in hole. It is reordering.
1377 if (before(start_seq,
1378 tcp_highest_sack_seq(tp)) &&
1379 before(start_seq, state->reord))
1380 state->reord = start_seq;
1382 if (!after(end_seq, tp->high_seq))
1383 state->flag |= FLAG_ORIG_SACK_ACKED;
1384 if (state->first_sackt == 0)
1385 state->first_sackt = xmit_time;
1386 state->last_sackt = xmit_time;
1389 if (sacked & TCPCB_LOST) {
1390 sacked &= ~TCPCB_LOST;
1391 tp->lost_out -= pcount;
1395 sacked |= TCPCB_SACKED_ACKED;
1396 state->flag |= FLAG_DATA_SACKED;
1397 tp->sacked_out += pcount;
1398 /* Out-of-order packets delivered */
1399 state->sack_delivered += pcount;
1401 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1402 if (tp->lost_skb_hint &&
1403 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1404 tp->lost_cnt_hint += pcount;
1407 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1408 * frames and clear it. undo_retrans is decreased above, L|R frames
1409 * are accounted above as well.
1411 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1412 sacked &= ~TCPCB_SACKED_RETRANS;
1413 tp->retrans_out -= pcount;
1419 /* Shift newly-SACKed bytes from this skb to the immediately previous
1420 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1422 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *prev,
1423 struct sk_buff *skb,
1424 struct tcp_sacktag_state *state,
1425 unsigned int pcount, int shifted, int mss,
1428 struct tcp_sock *tp = tcp_sk(sk);
1429 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1430 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
1434 /* Adjust counters and hints for the newly sacked sequence
1435 * range but discard the return value since prev is already
1436 * marked. We must tag the range first because the seq
1437 * advancement below implicitly advances
1438 * tcp_highest_sack_seq() when skb is highest_sack.
1440 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1441 start_seq, end_seq, dup_sack, pcount,
1442 tcp_skb_timestamp_us(skb));
1443 tcp_rate_skb_delivered(sk, skb, state->rate);
1445 if (skb == tp->lost_skb_hint)
1446 tp->lost_cnt_hint += pcount;
1448 TCP_SKB_CB(prev)->end_seq += shifted;
1449 TCP_SKB_CB(skb)->seq += shifted;
1451 tcp_skb_pcount_add(prev, pcount);
1452 WARN_ON_ONCE(tcp_skb_pcount(skb) < pcount);
1453 tcp_skb_pcount_add(skb, -pcount);
1455 /* When we're adding to gso_segs == 1, gso_size will be zero,
1456 * in theory this shouldn't be necessary but as long as DSACK
1457 * code can come after this skb later on it's better to keep
1458 * setting gso_size to something.
1460 if (!TCP_SKB_CB(prev)->tcp_gso_size)
1461 TCP_SKB_CB(prev)->tcp_gso_size = mss;
1463 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1464 if (tcp_skb_pcount(skb) <= 1)
1465 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1467 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1468 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1471 BUG_ON(!tcp_skb_pcount(skb));
1472 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1476 /* Whole SKB was eaten :-) */
1478 if (skb == tp->retransmit_skb_hint)
1479 tp->retransmit_skb_hint = prev;
1480 if (skb == tp->lost_skb_hint) {
1481 tp->lost_skb_hint = prev;
1482 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1485 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1486 TCP_SKB_CB(prev)->eor = TCP_SKB_CB(skb)->eor;
1487 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1488 TCP_SKB_CB(prev)->end_seq++;
1490 if (skb == tcp_highest_sack(sk))
1491 tcp_advance_highest_sack(sk, skb);
1493 tcp_skb_collapse_tstamp(prev, skb);
1494 if (unlikely(TCP_SKB_CB(prev)->tx.delivered_mstamp))
1495 TCP_SKB_CB(prev)->tx.delivered_mstamp = 0;
1497 tcp_rtx_queue_unlink_and_free(skb, sk);
1499 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKMERGED);
1504 /* I wish gso_size would have a bit more sane initialization than
1505 * something-or-zero which complicates things
1507 static int tcp_skb_seglen(const struct sk_buff *skb)
1509 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1512 /* Shifting pages past head area doesn't work */
1513 static int skb_can_shift(const struct sk_buff *skb)
1515 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1518 int tcp_skb_shift(struct sk_buff *to, struct sk_buff *from,
1519 int pcount, int shiftlen)
1521 /* TCP min gso_size is 8 bytes (TCP_MIN_GSO_SIZE)
1522 * Since TCP_SKB_CB(skb)->tcp_gso_segs is 16 bits, we need
1523 * to make sure not storing more than 65535 * 8 bytes per skb,
1524 * even if current MSS is bigger.
1526 if (unlikely(to->len + shiftlen >= 65535 * TCP_MIN_GSO_SIZE))
1528 if (unlikely(tcp_skb_pcount(to) + pcount > 65535))
1530 return skb_shift(to, from, shiftlen);
1533 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1536 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1537 struct tcp_sacktag_state *state,
1538 u32 start_seq, u32 end_seq,
1541 struct tcp_sock *tp = tcp_sk(sk);
1542 struct sk_buff *prev;
1548 /* Normally R but no L won't result in plain S */
1550 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1552 if (!skb_can_shift(skb))
1554 /* This frame is about to be dropped (was ACKed). */
1555 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1558 /* Can only happen with delayed DSACK + discard craziness */
1559 prev = skb_rb_prev(skb);
1563 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1566 if (!tcp_skb_can_collapse(prev, skb))
1569 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1570 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1574 pcount = tcp_skb_pcount(skb);
1575 mss = tcp_skb_seglen(skb);
1577 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1578 * drop this restriction as unnecessary
1580 if (mss != tcp_skb_seglen(prev))
1583 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1585 /* CHECKME: This is non-MSS split case only?, this will
1586 * cause skipped skbs due to advancing loop btw, original
1587 * has that feature too
1589 if (tcp_skb_pcount(skb) <= 1)
1592 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1594 /* TODO: head merge to next could be attempted here
1595 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1596 * though it might not be worth of the additional hassle
1598 * ...we can probably just fallback to what was done
1599 * previously. We could try merging non-SACKed ones
1600 * as well but it probably isn't going to buy off
1601 * because later SACKs might again split them, and
1602 * it would make skb timestamp tracking considerably
1608 len = end_seq - TCP_SKB_CB(skb)->seq;
1610 BUG_ON(len > skb->len);
1612 /* MSS boundaries should be honoured or else pcount will
1613 * severely break even though it makes things bit trickier.
1614 * Optimize common case to avoid most of the divides
1616 mss = tcp_skb_mss(skb);
1618 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1619 * drop this restriction as unnecessary
1621 if (mss != tcp_skb_seglen(prev))
1626 } else if (len < mss) {
1634 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1635 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1638 if (!tcp_skb_shift(prev, skb, pcount, len))
1640 if (!tcp_shifted_skb(sk, prev, skb, state, pcount, len, mss, dup_sack))
1643 /* Hole filled allows collapsing with the next as well, this is very
1644 * useful when hole on every nth skb pattern happens
1646 skb = skb_rb_next(prev);
1650 if (!skb_can_shift(skb) ||
1651 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1652 (mss != tcp_skb_seglen(skb)))
1655 if (!tcp_skb_can_collapse(prev, skb))
1658 pcount = tcp_skb_pcount(skb);
1659 if (tcp_skb_shift(prev, skb, pcount, len))
1660 tcp_shifted_skb(sk, prev, skb, state, pcount,
1670 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1674 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1675 struct tcp_sack_block *next_dup,
1676 struct tcp_sacktag_state *state,
1677 u32 start_seq, u32 end_seq,
1680 struct tcp_sock *tp = tcp_sk(sk);
1681 struct sk_buff *tmp;
1683 skb_rbtree_walk_from(skb) {
1685 bool dup_sack = dup_sack_in;
1687 /* queue is in-order => we can short-circuit the walk early */
1688 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1692 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1693 in_sack = tcp_match_skb_to_sack(sk, skb,
1694 next_dup->start_seq,
1700 /* skb reference here is a bit tricky to get right, since
1701 * shifting can eat and free both this skb and the next,
1702 * so not even _safe variant of the loop is enough.
1705 tmp = tcp_shift_skb_data(sk, skb, state,
1706 start_seq, end_seq, dup_sack);
1715 in_sack = tcp_match_skb_to_sack(sk, skb,
1721 if (unlikely(in_sack < 0))
1725 TCP_SKB_CB(skb)->sacked =
1728 TCP_SKB_CB(skb)->sacked,
1729 TCP_SKB_CB(skb)->seq,
1730 TCP_SKB_CB(skb)->end_seq,
1732 tcp_skb_pcount(skb),
1733 tcp_skb_timestamp_us(skb));
1734 tcp_rate_skb_delivered(sk, skb, state->rate);
1735 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1736 list_del_init(&skb->tcp_tsorted_anchor);
1738 if (!before(TCP_SKB_CB(skb)->seq,
1739 tcp_highest_sack_seq(tp)))
1740 tcp_advance_highest_sack(sk, skb);
1746 static struct sk_buff *tcp_sacktag_bsearch(struct sock *sk, u32 seq)
1748 struct rb_node *parent, **p = &sk->tcp_rtx_queue.rb_node;
1749 struct sk_buff *skb;
1753 skb = rb_to_skb(parent);
1754 if (before(seq, TCP_SKB_CB(skb)->seq)) {
1755 p = &parent->rb_left;
1758 if (!before(seq, TCP_SKB_CB(skb)->end_seq)) {
1759 p = &parent->rb_right;
1767 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1770 if (skb && after(TCP_SKB_CB(skb)->seq, skip_to_seq))
1773 return tcp_sacktag_bsearch(sk, skip_to_seq);
1776 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1778 struct tcp_sack_block *next_dup,
1779 struct tcp_sacktag_state *state,
1785 if (before(next_dup->start_seq, skip_to_seq)) {
1786 skb = tcp_sacktag_skip(skb, sk, next_dup->start_seq);
1787 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1788 next_dup->start_seq, next_dup->end_seq,
1795 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1797 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1801 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1802 u32 prior_snd_una, struct tcp_sacktag_state *state)
1804 struct tcp_sock *tp = tcp_sk(sk);
1805 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1806 TCP_SKB_CB(ack_skb)->sacked);
1807 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1808 struct tcp_sack_block sp[TCP_NUM_SACKS];
1809 struct tcp_sack_block *cache;
1810 struct sk_buff *skb;
1811 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1813 bool found_dup_sack = false;
1815 int first_sack_index;
1818 state->reord = tp->snd_nxt;
1820 if (!tp->sacked_out)
1821 tcp_highest_sack_reset(sk);
1823 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1824 num_sacks, prior_snd_una, state);
1826 /* Eliminate too old ACKs, but take into
1827 * account more or less fresh ones, they can
1828 * contain valid SACK info.
1830 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1833 if (!tp->packets_out)
1837 first_sack_index = 0;
1838 for (i = 0; i < num_sacks; i++) {
1839 bool dup_sack = !i && found_dup_sack;
1841 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1842 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1844 if (!tcp_is_sackblock_valid(tp, dup_sack,
1845 sp[used_sacks].start_seq,
1846 sp[used_sacks].end_seq)) {
1850 if (!tp->undo_marker)
1851 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1853 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1855 /* Don't count olds caused by ACK reordering */
1856 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1857 !after(sp[used_sacks].end_seq, tp->snd_una))
1859 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1862 NET_INC_STATS(sock_net(sk), mib_idx);
1864 first_sack_index = -1;
1868 /* Ignore very old stuff early */
1869 if (!after(sp[used_sacks].end_seq, prior_snd_una)) {
1871 first_sack_index = -1;
1878 /* order SACK blocks to allow in order walk of the retrans queue */
1879 for (i = used_sacks - 1; i > 0; i--) {
1880 for (j = 0; j < i; j++) {
1881 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1882 swap(sp[j], sp[j + 1]);
1884 /* Track where the first SACK block goes to */
1885 if (j == first_sack_index)
1886 first_sack_index = j + 1;
1891 state->mss_now = tcp_current_mss(sk);
1895 if (!tp->sacked_out) {
1896 /* It's already past, so skip checking against it */
1897 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1899 cache = tp->recv_sack_cache;
1900 /* Skip empty blocks in at head of the cache */
1901 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1906 while (i < used_sacks) {
1907 u32 start_seq = sp[i].start_seq;
1908 u32 end_seq = sp[i].end_seq;
1909 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1910 struct tcp_sack_block *next_dup = NULL;
1912 if (found_dup_sack && ((i + 1) == first_sack_index))
1913 next_dup = &sp[i + 1];
1915 /* Skip too early cached blocks */
1916 while (tcp_sack_cache_ok(tp, cache) &&
1917 !before(start_seq, cache->end_seq))
1920 /* Can skip some work by looking recv_sack_cache? */
1921 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1922 after(end_seq, cache->start_seq)) {
1925 if (before(start_seq, cache->start_seq)) {
1926 skb = tcp_sacktag_skip(skb, sk, start_seq);
1927 skb = tcp_sacktag_walk(skb, sk, next_dup,
1934 /* Rest of the block already fully processed? */
1935 if (!after(end_seq, cache->end_seq))
1938 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1942 /* ...tail remains todo... */
1943 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1944 /* ...but better entrypoint exists! */
1945 skb = tcp_highest_sack(sk);
1952 skb = tcp_sacktag_skip(skb, sk, cache->end_seq);
1953 /* Check overlap against next cached too (past this one already) */
1958 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1959 skb = tcp_highest_sack(sk);
1963 skb = tcp_sacktag_skip(skb, sk, start_seq);
1966 skb = tcp_sacktag_walk(skb, sk, next_dup, state,
1967 start_seq, end_seq, dup_sack);
1973 /* Clear the head of the cache sack blocks so we can skip it next time */
1974 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1975 tp->recv_sack_cache[i].start_seq = 0;
1976 tp->recv_sack_cache[i].end_seq = 0;
1978 for (j = 0; j < used_sacks; j++)
1979 tp->recv_sack_cache[i++] = sp[j];
1981 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Loss || tp->undo_marker)
1982 tcp_check_sack_reordering(sk, state->reord, 0);
1984 tcp_verify_left_out(tp);
1987 #if FASTRETRANS_DEBUG > 0
1988 WARN_ON((int)tp->sacked_out < 0);
1989 WARN_ON((int)tp->lost_out < 0);
1990 WARN_ON((int)tp->retrans_out < 0);
1991 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1996 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1997 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1999 static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
2003 holes = max(tp->lost_out, 1U);
2004 holes = min(holes, tp->packets_out);
2006 if ((tp->sacked_out + holes) > tp->packets_out) {
2007 tp->sacked_out = tp->packets_out - holes;
2013 /* If we receive more dupacks than we expected counting segments
2014 * in assumption of absent reordering, interpret this as reordering.
2015 * The only another reason could be bug in receiver TCP.
2017 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
2019 struct tcp_sock *tp = tcp_sk(sk);
2021 if (!tcp_limit_reno_sacked(tp))
2024 tp->reordering = min_t(u32, tp->packets_out + addend,
2025 sock_net(sk)->ipv4.sysctl_tcp_max_reordering);
2027 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRENOREORDER);
2030 /* Emulate SACKs for SACKless connection: account for a new dupack. */
2032 static void tcp_add_reno_sack(struct sock *sk, int num_dupack, bool ece_ack)
2035 struct tcp_sock *tp = tcp_sk(sk);
2036 u32 prior_sacked = tp->sacked_out;
2039 tp->sacked_out += num_dupack;
2040 tcp_check_reno_reordering(sk, 0);
2041 delivered = tp->sacked_out - prior_sacked;
2043 tcp_count_delivered(tp, delivered, ece_ack);
2044 tcp_verify_left_out(tp);
2048 /* Account for ACK, ACKing some data in Reno Recovery phase. */
2050 static void tcp_remove_reno_sacks(struct sock *sk, int acked, bool ece_ack)
2052 struct tcp_sock *tp = tcp_sk(sk);
2055 /* One ACK acked hole. The rest eat duplicate ACKs. */
2056 tcp_count_delivered(tp, max_t(int, acked - tp->sacked_out, 1),
2058 if (acked - 1 >= tp->sacked_out)
2061 tp->sacked_out -= acked - 1;
2063 tcp_check_reno_reordering(sk, acked);
2064 tcp_verify_left_out(tp);
2067 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
2072 void tcp_clear_retrans(struct tcp_sock *tp)
2074 tp->retrans_out = 0;
2076 tp->undo_marker = 0;
2077 tp->undo_retrans = -1;
2081 static inline void tcp_init_undo(struct tcp_sock *tp)
2083 tp->undo_marker = tp->snd_una;
2084 /* Retransmission still in flight may cause DSACKs later. */
2085 tp->undo_retrans = tp->retrans_out ? : -1;
2088 static bool tcp_is_rack(const struct sock *sk)
2090 return sock_net(sk)->ipv4.sysctl_tcp_recovery & TCP_RACK_LOSS_DETECTION;
2093 /* If we detect SACK reneging, forget all SACK information
2094 * and reset tags completely, otherwise preserve SACKs. If receiver
2095 * dropped its ofo queue, we will know this due to reneging detection.
2097 static void tcp_timeout_mark_lost(struct sock *sk)
2099 struct tcp_sock *tp = tcp_sk(sk);
2100 struct sk_buff *skb, *head;
2101 bool is_reneg; /* is receiver reneging on SACKs? */
2103 head = tcp_rtx_queue_head(sk);
2104 is_reneg = head && (TCP_SKB_CB(head)->sacked & TCPCB_SACKED_ACKED);
2106 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
2108 /* Mark SACK reneging until we recover from this loss event. */
2109 tp->is_sack_reneg = 1;
2110 } else if (tcp_is_reno(tp)) {
2111 tcp_reset_reno_sack(tp);
2115 skb_rbtree_walk_from(skb) {
2117 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
2118 else if (tcp_is_rack(sk) && skb != head &&
2119 tcp_rack_skb_timeout(tp, skb, 0) > 0)
2120 continue; /* Don't mark recently sent ones lost yet */
2121 tcp_mark_skb_lost(sk, skb);
2123 tcp_verify_left_out(tp);
2124 tcp_clear_all_retrans_hints(tp);
2127 /* Enter Loss state. */
2128 void tcp_enter_loss(struct sock *sk)
2130 const struct inet_connection_sock *icsk = inet_csk(sk);
2131 struct tcp_sock *tp = tcp_sk(sk);
2132 struct net *net = sock_net(sk);
2133 bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery;
2135 tcp_timeout_mark_lost(sk);
2137 /* Reduce ssthresh if it has not yet been made inside this window. */
2138 if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
2139 !after(tp->high_seq, tp->snd_una) ||
2140 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
2141 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2142 tp->prior_cwnd = tcp_snd_cwnd(tp);
2143 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2144 tcp_ca_event(sk, CA_EVENT_LOSS);
2147 tcp_snd_cwnd_set(tp, tcp_packets_in_flight(tp) + 1);
2148 tp->snd_cwnd_cnt = 0;
2149 tp->snd_cwnd_stamp = tcp_jiffies32;
2151 /* Timeout in disordered state after receiving substantial DUPACKs
2152 * suggests that the degree of reordering is over-estimated.
2154 if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
2155 tp->sacked_out >= net->ipv4.sysctl_tcp_reordering)
2156 tp->reordering = min_t(unsigned int, tp->reordering,
2157 net->ipv4.sysctl_tcp_reordering);
2158 tcp_set_ca_state(sk, TCP_CA_Loss);
2159 tp->high_seq = tp->snd_nxt;
2160 tcp_ecn_queue_cwr(tp);
2162 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
2163 * loss recovery is underway except recurring timeout(s) on
2164 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
2166 tp->frto = net->ipv4.sysctl_tcp_frto &&
2167 (new_recovery || icsk->icsk_retransmits) &&
2168 !inet_csk(sk)->icsk_mtup.probe_size;
2171 /* If ACK arrived pointing to a remembered SACK, it means that our
2172 * remembered SACKs do not reflect real state of receiver i.e.
2173 * receiver _host_ is heavily congested (or buggy).
2175 * To avoid big spurious retransmission bursts due to transient SACK
2176 * scoreboard oddities that look like reneging, we give the receiver a
2177 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
2178 * restore sanity to the SACK scoreboard. If the apparent reneging
2179 * persists until this RTO then we'll clear the SACK scoreboard.
2181 static bool tcp_check_sack_reneging(struct sock *sk, int flag)
2183 if (flag & FLAG_SACK_RENEGING) {
2184 struct tcp_sock *tp = tcp_sk(sk);
2185 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4),
2186 msecs_to_jiffies(10));
2188 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2189 delay, TCP_RTO_MAX);
2195 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2196 * counter when SACK is enabled (without SACK, sacked_out is used for
2199 * With reordering, holes may still be in flight, so RFC3517 recovery
2200 * uses pure sacked_out (total number of SACKed segments) even though
2201 * it violates the RFC that uses duplicate ACKs, often these are equal
2202 * but when e.g. out-of-window ACKs or packet duplication occurs,
2203 * they differ. Since neither occurs due to loss, TCP should really
2206 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2208 return tp->sacked_out + 1;
2211 /* Linux NewReno/SACK/ECN state machine.
2212 * --------------------------------------
2214 * "Open" Normal state, no dubious events, fast path.
2215 * "Disorder" In all the respects it is "Open",
2216 * but requires a bit more attention. It is entered when
2217 * we see some SACKs or dupacks. It is split of "Open"
2218 * mainly to move some processing from fast path to slow one.
2219 * "CWR" CWND was reduced due to some Congestion Notification event.
2220 * It can be ECN, ICMP source quench, local device congestion.
2221 * "Recovery" CWND was reduced, we are fast-retransmitting.
2222 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2224 * tcp_fastretrans_alert() is entered:
2225 * - each incoming ACK, if state is not "Open"
2226 * - when arrived ACK is unusual, namely:
2231 * Counting packets in flight is pretty simple.
2233 * in_flight = packets_out - left_out + retrans_out
2235 * packets_out is SND.NXT-SND.UNA counted in packets.
2237 * retrans_out is number of retransmitted segments.
2239 * left_out is number of segments left network, but not ACKed yet.
2241 * left_out = sacked_out + lost_out
2243 * sacked_out: Packets, which arrived to receiver out of order
2244 * and hence not ACKed. With SACKs this number is simply
2245 * amount of SACKed data. Even without SACKs
2246 * it is easy to give pretty reliable estimate of this number,
2247 * counting duplicate ACKs.
2249 * lost_out: Packets lost by network. TCP has no explicit
2250 * "loss notification" feedback from network (for now).
2251 * It means that this number can be only _guessed_.
2252 * Actually, it is the heuristics to predict lossage that
2253 * distinguishes different algorithms.
2255 * F.e. after RTO, when all the queue is considered as lost,
2256 * lost_out = packets_out and in_flight = retrans_out.
2258 * Essentially, we have now a few algorithms detecting
2261 * If the receiver supports SACK:
2263 * RFC6675/3517: It is the conventional algorithm. A packet is
2264 * considered lost if the number of higher sequence packets
2265 * SACKed is greater than or equal the DUPACK thoreshold
2266 * (reordering). This is implemented in tcp_mark_head_lost and
2267 * tcp_update_scoreboard.
2269 * RACK (draft-ietf-tcpm-rack-01): it is a newer algorithm
2270 * (2017-) that checks timing instead of counting DUPACKs.
2271 * Essentially a packet is considered lost if it's not S/ACKed
2272 * after RTT + reordering_window, where both metrics are
2273 * dynamically measured and adjusted. This is implemented in
2274 * tcp_rack_mark_lost.
2276 * If the receiver does not support SACK:
2278 * NewReno (RFC6582): in Recovery we assume that one segment
2279 * is lost (classic Reno). While we are in Recovery and
2280 * a partial ACK arrives, we assume that one more packet
2281 * is lost (NewReno). This heuristics are the same in NewReno
2284 * Really tricky (and requiring careful tuning) part of algorithm
2285 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2286 * The first determines the moment _when_ we should reduce CWND and,
2287 * hence, slow down forward transmission. In fact, it determines the moment
2288 * when we decide that hole is caused by loss, rather than by a reorder.
2290 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2291 * holes, caused by lost packets.
2293 * And the most logically complicated part of algorithm is undo
2294 * heuristics. We detect false retransmits due to both too early
2295 * fast retransmit (reordering) and underestimated RTO, analyzing
2296 * timestamps and D-SACKs. When we detect that some segments were
2297 * retransmitted by mistake and CWND reduction was wrong, we undo
2298 * window reduction and abort recovery phase. This logic is hidden
2299 * inside several functions named tcp_try_undo_<something>.
2302 /* This function decides, when we should leave Disordered state
2303 * and enter Recovery phase, reducing congestion window.
2305 * Main question: may we further continue forward transmission
2306 * with the same cwnd?
2308 static bool tcp_time_to_recover(struct sock *sk, int flag)
2310 struct tcp_sock *tp = tcp_sk(sk);
2312 /* Trick#1: The loss is proven. */
2316 /* Not-A-Trick#2 : Classic rule... */
2317 if (!tcp_is_rack(sk) && tcp_dupack_heuristics(tp) > tp->reordering)
2323 /* Detect loss in event "A" above by marking head of queue up as lost.
2324 * For RFC3517 SACK, a segment is considered lost if it
2325 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2326 * the maximum SACKed segments to pass before reaching this limit.
2328 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2330 struct tcp_sock *tp = tcp_sk(sk);
2331 struct sk_buff *skb;
2333 /* Use SACK to deduce losses of new sequences sent during recovery */
2334 const u32 loss_high = tp->snd_nxt;
2336 WARN_ON(packets > tp->packets_out);
2337 skb = tp->lost_skb_hint;
2339 /* Head already handled? */
2340 if (mark_head && after(TCP_SKB_CB(skb)->seq, tp->snd_una))
2342 cnt = tp->lost_cnt_hint;
2344 skb = tcp_rtx_queue_head(sk);
2348 skb_rbtree_walk_from(skb) {
2349 /* TODO: do this better */
2350 /* this is not the most efficient way to do this... */
2351 tp->lost_skb_hint = skb;
2352 tp->lost_cnt_hint = cnt;
2354 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2357 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2358 cnt += tcp_skb_pcount(skb);
2363 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_LOST))
2364 tcp_mark_skb_lost(sk, skb);
2369 tcp_verify_left_out(tp);
2372 /* Account newly detected lost packet(s) */
2374 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2376 struct tcp_sock *tp = tcp_sk(sk);
2378 if (tcp_is_sack(tp)) {
2379 int sacked_upto = tp->sacked_out - tp->reordering;
2380 if (sacked_upto >= 0)
2381 tcp_mark_head_lost(sk, sacked_upto, 0);
2382 else if (fast_rexmit)
2383 tcp_mark_head_lost(sk, 1, 1);
2387 static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when)
2389 return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2390 before(tp->rx_opt.rcv_tsecr, when);
2393 /* skb is spurious retransmitted if the returned timestamp echo
2394 * reply is prior to the skb transmission time
2396 static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp,
2397 const struct sk_buff *skb)
2399 return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) &&
2400 tcp_tsopt_ecr_before(tp, tcp_skb_timestamp(skb));
2403 /* Nothing was retransmitted or returned timestamp is less
2404 * than timestamp of the first retransmission.
2406 static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2408 return tp->retrans_stamp &&
2409 tcp_tsopt_ecr_before(tp, tp->retrans_stamp);
2412 /* Undo procedures. */
2414 /* We can clear retrans_stamp when there are no retransmissions in the
2415 * window. It would seem that it is trivially available for us in
2416 * tp->retrans_out, however, that kind of assumptions doesn't consider
2417 * what will happen if errors occur when sending retransmission for the
2418 * second time. ...It could the that such segment has only
2419 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2420 * the head skb is enough except for some reneging corner cases that
2421 * are not worth the effort.
2423 * Main reason for all this complexity is the fact that connection dying
2424 * time now depends on the validity of the retrans_stamp, in particular,
2425 * that successive retransmissions of a segment must not advance
2426 * retrans_stamp under any conditions.
2428 static bool tcp_any_retrans_done(const struct sock *sk)
2430 const struct tcp_sock *tp = tcp_sk(sk);
2431 struct sk_buff *skb;
2433 if (tp->retrans_out)
2436 skb = tcp_rtx_queue_head(sk);
2437 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2443 static void DBGUNDO(struct sock *sk, const char *msg)
2445 #if FASTRETRANS_DEBUG > 1
2446 struct tcp_sock *tp = tcp_sk(sk);
2447 struct inet_sock *inet = inet_sk(sk);
2449 if (sk->sk_family == AF_INET) {
2450 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2452 &inet->inet_daddr, ntohs(inet->inet_dport),
2453 tcp_snd_cwnd(tp), tcp_left_out(tp),
2454 tp->snd_ssthresh, tp->prior_ssthresh,
2457 #if IS_ENABLED(CONFIG_IPV6)
2458 else if (sk->sk_family == AF_INET6) {
2459 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2461 &sk->sk_v6_daddr, ntohs(inet->inet_dport),
2462 tcp_snd_cwnd(tp), tcp_left_out(tp),
2463 tp->snd_ssthresh, tp->prior_ssthresh,
2470 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
2472 struct tcp_sock *tp = tcp_sk(sk);
2475 struct sk_buff *skb;
2477 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2478 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2481 tcp_clear_all_retrans_hints(tp);
2484 if (tp->prior_ssthresh) {
2485 const struct inet_connection_sock *icsk = inet_csk(sk);
2487 tcp_snd_cwnd_set(tp, icsk->icsk_ca_ops->undo_cwnd(sk));
2489 if (tp->prior_ssthresh > tp->snd_ssthresh) {
2490 tp->snd_ssthresh = tp->prior_ssthresh;
2491 tcp_ecn_withdraw_cwr(tp);
2494 tp->snd_cwnd_stamp = tcp_jiffies32;
2495 tp->undo_marker = 0;
2496 tp->rack.advanced = 1; /* Force RACK to re-exam losses */
2499 static inline bool tcp_may_undo(const struct tcp_sock *tp)
2501 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2504 /* People celebrate: "We love our President!" */
2505 static bool tcp_try_undo_recovery(struct sock *sk)
2507 struct tcp_sock *tp = tcp_sk(sk);
2509 if (tcp_may_undo(tp)) {
2512 /* Happy end! We did not retransmit anything
2513 * or our original transmission succeeded.
2515 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2516 tcp_undo_cwnd_reduction(sk, false);
2517 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2518 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2520 mib_idx = LINUX_MIB_TCPFULLUNDO;
2522 NET_INC_STATS(sock_net(sk), mib_idx);
2523 } else if (tp->rack.reo_wnd_persist) {
2524 tp->rack.reo_wnd_persist--;
2526 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2527 /* Hold old state until something *above* high_seq
2528 * is ACKed. For Reno it is MUST to prevent false
2529 * fast retransmits (RFC2582). SACK TCP is safe. */
2530 if (!tcp_any_retrans_done(sk))
2531 tp->retrans_stamp = 0;
2534 tcp_set_ca_state(sk, TCP_CA_Open);
2535 tp->is_sack_reneg = 0;
2539 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2540 static bool tcp_try_undo_dsack(struct sock *sk)
2542 struct tcp_sock *tp = tcp_sk(sk);
2544 if (tp->undo_marker && !tp->undo_retrans) {
2545 tp->rack.reo_wnd_persist = min(TCP_RACK_RECOVERY_THRESH,
2546 tp->rack.reo_wnd_persist + 1);
2547 DBGUNDO(sk, "D-SACK");
2548 tcp_undo_cwnd_reduction(sk, false);
2549 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2555 /* Undo during loss recovery after partial ACK or using F-RTO. */
2556 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2558 struct tcp_sock *tp = tcp_sk(sk);
2560 if (frto_undo || tcp_may_undo(tp)) {
2561 tcp_undo_cwnd_reduction(sk, true);
2563 DBGUNDO(sk, "partial loss");
2564 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2566 NET_INC_STATS(sock_net(sk),
2567 LINUX_MIB_TCPSPURIOUSRTOS);
2568 inet_csk(sk)->icsk_retransmits = 0;
2569 if (frto_undo || tcp_is_sack(tp)) {
2570 tcp_set_ca_state(sk, TCP_CA_Open);
2571 tp->is_sack_reneg = 0;
2578 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
2579 * It computes the number of packets to send (sndcnt) based on packets newly
2581 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2582 * cwnd reductions across a full RTT.
2583 * 2) Otherwise PRR uses packet conservation to send as much as delivered.
2584 * But when SND_UNA is acked without further losses,
2585 * slow starts cwnd up to ssthresh to speed up the recovery.
2587 static void tcp_init_cwnd_reduction(struct sock *sk)
2589 struct tcp_sock *tp = tcp_sk(sk);
2591 tp->high_seq = tp->snd_nxt;
2592 tp->tlp_high_seq = 0;
2593 tp->snd_cwnd_cnt = 0;
2594 tp->prior_cwnd = tcp_snd_cwnd(tp);
2595 tp->prr_delivered = 0;
2597 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2598 tcp_ecn_queue_cwr(tp);
2601 void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked, int newly_lost, int flag)
2603 struct tcp_sock *tp = tcp_sk(sk);
2605 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2607 if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd))
2610 tp->prr_delivered += newly_acked_sacked;
2612 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2614 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2615 } else if (flag & FLAG_SND_UNA_ADVANCED && !newly_lost) {
2616 sndcnt = min_t(int, delta,
2617 max_t(int, tp->prr_delivered - tp->prr_out,
2618 newly_acked_sacked) + 1);
2620 sndcnt = min(delta, newly_acked_sacked);
2622 /* Force a fast retransmit upon entering fast recovery */
2623 sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1));
2624 tcp_snd_cwnd_set(tp, tcp_packets_in_flight(tp) + sndcnt);
2627 static inline void tcp_end_cwnd_reduction(struct sock *sk)
2629 struct tcp_sock *tp = tcp_sk(sk);
2631 if (inet_csk(sk)->icsk_ca_ops->cong_control)
2634 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2635 if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH &&
2636 (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR || tp->undo_marker)) {
2637 tcp_snd_cwnd_set(tp, tp->snd_ssthresh);
2638 tp->snd_cwnd_stamp = tcp_jiffies32;
2640 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2643 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2644 void tcp_enter_cwr(struct sock *sk)
2646 struct tcp_sock *tp = tcp_sk(sk);
2648 tp->prior_ssthresh = 0;
2649 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2650 tp->undo_marker = 0;
2651 tcp_init_cwnd_reduction(sk);
2652 tcp_set_ca_state(sk, TCP_CA_CWR);
2655 EXPORT_SYMBOL(tcp_enter_cwr);
2657 static void tcp_try_keep_open(struct sock *sk)
2659 struct tcp_sock *tp = tcp_sk(sk);
2660 int state = TCP_CA_Open;
2662 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2663 state = TCP_CA_Disorder;
2665 if (inet_csk(sk)->icsk_ca_state != state) {
2666 tcp_set_ca_state(sk, state);
2667 tp->high_seq = tp->snd_nxt;
2671 static void tcp_try_to_open(struct sock *sk, int flag)
2673 struct tcp_sock *tp = tcp_sk(sk);
2675 tcp_verify_left_out(tp);
2677 if (!tcp_any_retrans_done(sk))
2678 tp->retrans_stamp = 0;
2680 if (flag & FLAG_ECE)
2683 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2684 tcp_try_keep_open(sk);
2688 static void tcp_mtup_probe_failed(struct sock *sk)
2690 struct inet_connection_sock *icsk = inet_csk(sk);
2692 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2693 icsk->icsk_mtup.probe_size = 0;
2694 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPFAIL);
2697 static void tcp_mtup_probe_success(struct sock *sk)
2699 struct tcp_sock *tp = tcp_sk(sk);
2700 struct inet_connection_sock *icsk = inet_csk(sk);
2702 /* FIXME: breaks with very large cwnd */
2703 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2704 tcp_snd_cwnd_set(tp, tcp_snd_cwnd(tp) *
2705 tcp_mss_to_mtu(sk, tp->mss_cache) /
2706 icsk->icsk_mtup.probe_size);
2707 tp->snd_cwnd_cnt = 0;
2708 tp->snd_cwnd_stamp = tcp_jiffies32;
2709 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2711 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2712 icsk->icsk_mtup.probe_size = 0;
2713 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2714 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS);
2717 /* Do a simple retransmit without using the backoff mechanisms in
2718 * tcp_timer. This is used for path mtu discovery.
2719 * The socket is already locked here.
2721 void tcp_simple_retransmit(struct sock *sk)
2723 const struct inet_connection_sock *icsk = inet_csk(sk);
2724 struct tcp_sock *tp = tcp_sk(sk);
2725 struct sk_buff *skb;
2728 /* A fastopen SYN request is stored as two separate packets within
2729 * the retransmit queue, this is done by tcp_send_syn_data().
2730 * As a result simply checking the MSS of the frames in the queue
2731 * will not work for the SYN packet.
2733 * Us being here is an indication of a path MTU issue so we can
2734 * assume that the fastopen SYN was lost and just mark all the
2735 * frames in the retransmit queue as lost. We will use an MSS of
2736 * -1 to mark all frames as lost, otherwise compute the current MSS.
2738 if (tp->syn_data && sk->sk_state == TCP_SYN_SENT)
2741 mss = tcp_current_mss(sk);
2743 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2744 if (tcp_skb_seglen(skb) > mss)
2745 tcp_mark_skb_lost(sk, skb);
2748 tcp_clear_retrans_hints_partial(tp);
2753 if (tcp_is_reno(tp))
2754 tcp_limit_reno_sacked(tp);
2756 tcp_verify_left_out(tp);
2758 /* Don't muck with the congestion window here.
2759 * Reason is that we do not increase amount of _data_
2760 * in network, but units changed and effective
2761 * cwnd/ssthresh really reduced now.
2763 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2764 tp->high_seq = tp->snd_nxt;
2765 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2766 tp->prior_ssthresh = 0;
2767 tp->undo_marker = 0;
2768 tcp_set_ca_state(sk, TCP_CA_Loss);
2770 tcp_xmit_retransmit_queue(sk);
2772 EXPORT_SYMBOL(tcp_simple_retransmit);
2774 void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2776 struct tcp_sock *tp = tcp_sk(sk);
2779 if (tcp_is_reno(tp))
2780 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2782 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2784 NET_INC_STATS(sock_net(sk), mib_idx);
2786 tp->prior_ssthresh = 0;
2789 if (!tcp_in_cwnd_reduction(sk)) {
2791 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2792 tcp_init_cwnd_reduction(sk);
2794 tcp_set_ca_state(sk, TCP_CA_Recovery);
2797 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2798 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2800 static void tcp_process_loss(struct sock *sk, int flag, int num_dupack,
2803 struct tcp_sock *tp = tcp_sk(sk);
2804 bool recovered = !before(tp->snd_una, tp->high_seq);
2806 if ((flag & FLAG_SND_UNA_ADVANCED || rcu_access_pointer(tp->fastopen_rsk)) &&
2807 tcp_try_undo_loss(sk, false))
2810 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2811 /* Step 3.b. A timeout is spurious if not all data are
2812 * lost, i.e., never-retransmitted data are (s)acked.
2814 if ((flag & FLAG_ORIG_SACK_ACKED) &&
2815 tcp_try_undo_loss(sk, true))
2818 if (after(tp->snd_nxt, tp->high_seq)) {
2819 if (flag & FLAG_DATA_SACKED || num_dupack)
2820 tp->frto = 0; /* Step 3.a. loss was real */
2821 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2822 tp->high_seq = tp->snd_nxt;
2823 /* Step 2.b. Try send new data (but deferred until cwnd
2824 * is updated in tcp_ack()). Otherwise fall back to
2825 * the conventional recovery.
2827 if (!tcp_write_queue_empty(sk) &&
2828 after(tcp_wnd_end(tp), tp->snd_nxt)) {
2829 *rexmit = REXMIT_NEW;
2837 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2838 tcp_try_undo_recovery(sk);
2841 if (tcp_is_reno(tp)) {
2842 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2843 * delivered. Lower inflight to clock out (re)tranmissions.
2845 if (after(tp->snd_nxt, tp->high_seq) && num_dupack)
2846 tcp_add_reno_sack(sk, num_dupack, flag & FLAG_ECE);
2847 else if (flag & FLAG_SND_UNA_ADVANCED)
2848 tcp_reset_reno_sack(tp);
2850 *rexmit = REXMIT_LOST;
2853 static bool tcp_force_fast_retransmit(struct sock *sk)
2855 struct tcp_sock *tp = tcp_sk(sk);
2857 return after(tcp_highest_sack_seq(tp),
2858 tp->snd_una + tp->reordering * tp->mss_cache);
2861 /* Undo during fast recovery after partial ACK. */
2862 static bool tcp_try_undo_partial(struct sock *sk, u32 prior_snd_una,
2865 struct tcp_sock *tp = tcp_sk(sk);
2867 if (tp->undo_marker && tcp_packet_delayed(tp)) {
2868 /* Plain luck! Hole if filled with delayed
2869 * packet, rather than with a retransmit. Check reordering.
2871 tcp_check_sack_reordering(sk, prior_snd_una, 1);
2873 /* We are getting evidence that the reordering degree is higher
2874 * than we realized. If there are no retransmits out then we
2875 * can undo. Otherwise we clock out new packets but do not
2876 * mark more packets lost or retransmit more.
2878 if (tp->retrans_out)
2881 if (!tcp_any_retrans_done(sk))
2882 tp->retrans_stamp = 0;
2884 DBGUNDO(sk, "partial recovery");
2885 tcp_undo_cwnd_reduction(sk, true);
2886 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2887 tcp_try_keep_open(sk);
2889 /* Partial ACK arrived. Force fast retransmit. */
2890 *do_lost = tcp_force_fast_retransmit(sk);
2895 static void tcp_identify_packet_loss(struct sock *sk, int *ack_flag)
2897 struct tcp_sock *tp = tcp_sk(sk);
2899 if (tcp_rtx_queue_empty(sk))
2902 if (unlikely(tcp_is_reno(tp))) {
2903 tcp_newreno_mark_lost(sk, *ack_flag & FLAG_SND_UNA_ADVANCED);
2904 } else if (tcp_is_rack(sk)) {
2905 u32 prior_retrans = tp->retrans_out;
2907 if (tcp_rack_mark_lost(sk))
2908 *ack_flag &= ~FLAG_SET_XMIT_TIMER;
2909 if (prior_retrans > tp->retrans_out)
2910 *ack_flag |= FLAG_LOST_RETRANS;
2914 /* Process an event, which can update packets-in-flight not trivially.
2915 * Main goal of this function is to calculate new estimate for left_out,
2916 * taking into account both packets sitting in receiver's buffer and
2917 * packets lost by network.
2919 * Besides that it updates the congestion state when packet loss or ECN
2920 * is detected. But it does not reduce the cwnd, it is done by the
2921 * congestion control later.
2923 * It does _not_ decide what to send, it is made in function
2924 * tcp_xmit_retransmit_queue().
2926 static void tcp_fastretrans_alert(struct sock *sk, const u32 prior_snd_una,
2927 int num_dupack, int *ack_flag, int *rexmit)
2929 struct inet_connection_sock *icsk = inet_csk(sk);
2930 struct tcp_sock *tp = tcp_sk(sk);
2931 int fast_rexmit = 0, flag = *ack_flag;
2932 bool ece_ack = flag & FLAG_ECE;
2933 bool do_lost = num_dupack || ((flag & FLAG_DATA_SACKED) &&
2934 tcp_force_fast_retransmit(sk));
2936 if (!tp->packets_out && tp->sacked_out)
2939 /* Now state machine starts.
2940 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2942 tp->prior_ssthresh = 0;
2944 /* B. In all the states check for reneging SACKs. */
2945 if (tcp_check_sack_reneging(sk, flag))
2948 /* C. Check consistency of the current state. */
2949 tcp_verify_left_out(tp);
2951 /* D. Check state exit conditions. State can be terminated
2952 * when high_seq is ACKed. */
2953 if (icsk->icsk_ca_state == TCP_CA_Open) {
2954 WARN_ON(tp->retrans_out != 0 && !tp->syn_data);
2955 tp->retrans_stamp = 0;
2956 } else if (!before(tp->snd_una, tp->high_seq)) {
2957 switch (icsk->icsk_ca_state) {
2959 /* CWR is to be held something *above* high_seq
2960 * is ACKed for CWR bit to reach receiver. */
2961 if (tp->snd_una != tp->high_seq) {
2962 tcp_end_cwnd_reduction(sk);
2963 tcp_set_ca_state(sk, TCP_CA_Open);
2967 case TCP_CA_Recovery:
2968 if (tcp_is_reno(tp))
2969 tcp_reset_reno_sack(tp);
2970 if (tcp_try_undo_recovery(sk))
2972 tcp_end_cwnd_reduction(sk);
2977 /* E. Process state. */
2978 switch (icsk->icsk_ca_state) {
2979 case TCP_CA_Recovery:
2980 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2981 if (tcp_is_reno(tp))
2982 tcp_add_reno_sack(sk, num_dupack, ece_ack);
2983 } else if (tcp_try_undo_partial(sk, prior_snd_una, &do_lost))
2986 if (tcp_try_undo_dsack(sk))
2987 tcp_try_keep_open(sk);
2989 tcp_identify_packet_loss(sk, ack_flag);
2990 if (icsk->icsk_ca_state != TCP_CA_Recovery) {
2991 if (!tcp_time_to_recover(sk, flag))
2993 /* Undo reverts the recovery state. If loss is evident,
2994 * starts a new recovery (e.g. reordering then loss);
2996 tcp_enter_recovery(sk, ece_ack);
3000 tcp_process_loss(sk, flag, num_dupack, rexmit);
3001 tcp_identify_packet_loss(sk, ack_flag);
3002 if (!(icsk->icsk_ca_state == TCP_CA_Open ||
3003 (*ack_flag & FLAG_LOST_RETRANS)))
3005 /* Change state if cwnd is undone or retransmits are lost */
3008 if (tcp_is_reno(tp)) {
3009 if (flag & FLAG_SND_UNA_ADVANCED)
3010 tcp_reset_reno_sack(tp);
3011 tcp_add_reno_sack(sk, num_dupack, ece_ack);
3014 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
3015 tcp_try_undo_dsack(sk);
3017 tcp_identify_packet_loss(sk, ack_flag);
3018 if (!tcp_time_to_recover(sk, flag)) {
3019 tcp_try_to_open(sk, flag);
3023 /* MTU probe failure: don't reduce cwnd */
3024 if (icsk->icsk_ca_state < TCP_CA_CWR &&
3025 icsk->icsk_mtup.probe_size &&
3026 tp->snd_una == tp->mtu_probe.probe_seq_start) {
3027 tcp_mtup_probe_failed(sk);
3028 /* Restores the reduction we did in tcp_mtup_probe() */
3029 tcp_snd_cwnd_set(tp, tcp_snd_cwnd(tp) + 1);
3030 tcp_simple_retransmit(sk);
3034 /* Otherwise enter Recovery state */
3035 tcp_enter_recovery(sk, ece_ack);
3039 if (!tcp_is_rack(sk) && do_lost)
3040 tcp_update_scoreboard(sk, fast_rexmit);
3041 *rexmit = REXMIT_LOST;
3044 static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us, const int flag)
3046 u32 wlen = sock_net(sk)->ipv4.sysctl_tcp_min_rtt_wlen * HZ;
3047 struct tcp_sock *tp = tcp_sk(sk);
3049 if ((flag & FLAG_ACK_MAYBE_DELAYED) && rtt_us > tcp_min_rtt(tp)) {
3050 /* If the remote keeps returning delayed ACKs, eventually
3051 * the min filter would pick it up and overestimate the
3052 * prop. delay when it expires. Skip suspected delayed ACKs.
3056 minmax_running_min(&tp->rtt_min, wlen, tcp_jiffies32,
3057 rtt_us ? : jiffies_to_usecs(1));
3060 static bool tcp_ack_update_rtt(struct sock *sk, const int flag,
3061 long seq_rtt_us, long sack_rtt_us,
3062 long ca_rtt_us, struct rate_sample *rs)
3064 const struct tcp_sock *tp = tcp_sk(sk);
3066 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
3067 * broken middle-boxes or peers may corrupt TS-ECR fields. But
3068 * Karn's algorithm forbids taking RTT if some retransmitted data
3069 * is acked (RFC6298).
3072 seq_rtt_us = sack_rtt_us;
3074 /* RTTM Rule: A TSecr value received in a segment is used to
3075 * update the averaged RTT measurement only if the segment
3076 * acknowledges some new data, i.e., only if it advances the
3077 * left edge of the send window.
3078 * See draft-ietf-tcplw-high-performance-00, section 3.3.
3080 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
3081 flag & FLAG_ACKED) {
3082 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr;
3084 if (likely(delta < INT_MAX / (USEC_PER_SEC / TCP_TS_HZ))) {
3087 seq_rtt_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
3088 ca_rtt_us = seq_rtt_us;
3091 rs->rtt_us = ca_rtt_us; /* RTT of last (S)ACKed packet (or -1) */
3095 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
3096 * always taken together with ACK, SACK, or TS-opts. Any negative
3097 * values will be skipped with the seq_rtt_us < 0 check above.
3099 tcp_update_rtt_min(sk, ca_rtt_us, flag);
3100 tcp_rtt_estimator(sk, seq_rtt_us);
3103 /* RFC6298: only reset backoff on valid RTT measurement. */
3104 inet_csk(sk)->icsk_backoff = 0;
3108 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
3109 void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req)
3111 struct rate_sample rs;
3114 if (req && !req->num_retrans && tcp_rsk(req)->snt_synack)
3115 rtt_us = tcp_stamp_us_delta(tcp_clock_us(), tcp_rsk(req)->snt_synack);
3117 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us, &rs);
3121 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
3123 const struct inet_connection_sock *icsk = inet_csk(sk);
3125 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
3126 tcp_sk(sk)->snd_cwnd_stamp = tcp_jiffies32;
3129 /* Restart timer after forward progress on connection.
3130 * RFC2988 recommends to restart timer to now+rto.
3132 void tcp_rearm_rto(struct sock *sk)
3134 const struct inet_connection_sock *icsk = inet_csk(sk);
3135 struct tcp_sock *tp = tcp_sk(sk);
3137 /* If the retrans timer is currently being used by Fast Open
3138 * for SYN-ACK retrans purpose, stay put.
3140 if (rcu_access_pointer(tp->fastopen_rsk))
3143 if (!tp->packets_out) {
3144 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
3146 u32 rto = inet_csk(sk)->icsk_rto;
3147 /* Offset the time elapsed after installing regular RTO */
3148 if (icsk->icsk_pending == ICSK_TIME_REO_TIMEOUT ||
3149 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
3150 s64 delta_us = tcp_rto_delta_us(sk);
3151 /* delta_us may not be positive if the socket is locked
3152 * when the retrans timer fires and is rescheduled.
3154 rto = usecs_to_jiffies(max_t(int, delta_us, 1));
3156 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
3161 /* Try to schedule a loss probe; if that doesn't work, then schedule an RTO. */
3162 static void tcp_set_xmit_timer(struct sock *sk)
3164 if (!tcp_schedule_loss_probe(sk, true))
3168 /* If we get here, the whole TSO packet has not been acked. */
3169 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3171 struct tcp_sock *tp = tcp_sk(sk);
3174 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3176 packets_acked = tcp_skb_pcount(skb);
3177 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3179 packets_acked -= tcp_skb_pcount(skb);
3181 if (packets_acked) {
3182 BUG_ON(tcp_skb_pcount(skb) == 0);
3183 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3186 return packets_acked;
3189 static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
3190 const struct sk_buff *ack_skb, u32 prior_snd_una)
3192 const struct skb_shared_info *shinfo;
3194 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3195 if (likely(!TCP_SKB_CB(skb)->txstamp_ack))
3198 shinfo = skb_shinfo(skb);
3199 if (!before(shinfo->tskey, prior_snd_una) &&
3200 before(shinfo->tskey, tcp_sk(sk)->snd_una)) {
3201 tcp_skb_tsorted_save(skb) {
3202 __skb_tstamp_tx(skb, ack_skb, NULL, sk, SCM_TSTAMP_ACK);
3203 } tcp_skb_tsorted_restore(skb);
3207 /* Remove acknowledged frames from the retransmission queue. If our packet
3208 * is before the ack sequence we can discard it as it's confirmed to have
3209 * arrived at the other end.
3211 static int tcp_clean_rtx_queue(struct sock *sk, const struct sk_buff *ack_skb,
3212 u32 prior_fack, u32 prior_snd_una,
3213 struct tcp_sacktag_state *sack, bool ece_ack)
3215 const struct inet_connection_sock *icsk = inet_csk(sk);
3216 u64 first_ackt, last_ackt;
3217 struct tcp_sock *tp = tcp_sk(sk);
3218 u32 prior_sacked = tp->sacked_out;
3219 u32 reord = tp->snd_nxt; /* lowest acked un-retx un-sacked seq */
3220 struct sk_buff *skb, *next;
3221 bool fully_acked = true;
3222 long sack_rtt_us = -1L;
3223 long seq_rtt_us = -1L;
3224 long ca_rtt_us = -1L;
3226 u32 last_in_flight = 0;
3232 for (skb = skb_rb_first(&sk->tcp_rtx_queue); skb; skb = next) {
3233 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3234 const u32 start_seq = scb->seq;
3235 u8 sacked = scb->sacked;
3238 /* Determine how many packets and what bytes were acked, tso and else */
3239 if (after(scb->end_seq, tp->snd_una)) {
3240 if (tcp_skb_pcount(skb) == 1 ||
3241 !after(tp->snd_una, scb->seq))
3244 acked_pcount = tcp_tso_acked(sk, skb);
3247 fully_acked = false;
3249 acked_pcount = tcp_skb_pcount(skb);
3252 if (unlikely(sacked & TCPCB_RETRANS)) {
3253 if (sacked & TCPCB_SACKED_RETRANS)
3254 tp->retrans_out -= acked_pcount;
3255 flag |= FLAG_RETRANS_DATA_ACKED;
3256 } else if (!(sacked & TCPCB_SACKED_ACKED)) {
3257 last_ackt = tcp_skb_timestamp_us(skb);
3258 WARN_ON_ONCE(last_ackt == 0);
3260 first_ackt = last_ackt;
3262 last_in_flight = TCP_SKB_CB(skb)->tx.in_flight;
3263 if (before(start_seq, reord))
3265 if (!after(scb->end_seq, tp->high_seq))
3266 flag |= FLAG_ORIG_SACK_ACKED;
3269 if (sacked & TCPCB_SACKED_ACKED) {
3270 tp->sacked_out -= acked_pcount;
3271 } else if (tcp_is_sack(tp)) {
3272 tcp_count_delivered(tp, acked_pcount, ece_ack);
3273 if (!tcp_skb_spurious_retrans(tp, skb))
3274 tcp_rack_advance(tp, sacked, scb->end_seq,
3275 tcp_skb_timestamp_us(skb));
3277 if (sacked & TCPCB_LOST)
3278 tp->lost_out -= acked_pcount;
3280 tp->packets_out -= acked_pcount;
3281 pkts_acked += acked_pcount;
3282 tcp_rate_skb_delivered(sk, skb, sack->rate);
3284 /* Initial outgoing SYN's get put onto the write_queue
3285 * just like anything else we transmit. It is not
3286 * true data, and if we misinform our callers that
3287 * this ACK acks real data, we will erroneously exit
3288 * connection startup slow start one packet too
3289 * quickly. This is severely frowned upon behavior.
3291 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
3292 flag |= FLAG_DATA_ACKED;
3294 flag |= FLAG_SYN_ACKED;
3295 tp->retrans_stamp = 0;
3301 tcp_ack_tstamp(sk, skb, ack_skb, prior_snd_una);
3303 next = skb_rb_next(skb);
3304 if (unlikely(skb == tp->retransmit_skb_hint))
3305 tp->retransmit_skb_hint = NULL;
3306 if (unlikely(skb == tp->lost_skb_hint))
3307 tp->lost_skb_hint = NULL;
3308 tcp_highest_sack_replace(sk, skb, next);
3309 tcp_rtx_queue_unlink_and_free(skb, sk);
3313 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
3315 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3316 tp->snd_up = tp->snd_una;
3319 tcp_ack_tstamp(sk, skb, ack_skb, prior_snd_una);
3320 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
3321 flag |= FLAG_SACK_RENEGING;
3324 if (likely(first_ackt) && !(flag & FLAG_RETRANS_DATA_ACKED)) {
3325 seq_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, first_ackt);
3326 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, last_ackt);
3328 if (pkts_acked == 1 && last_in_flight < tp->mss_cache &&
3329 last_in_flight && !prior_sacked && fully_acked &&
3330 sack->rate->prior_delivered + 1 == tp->delivered &&
3331 !(flag & (FLAG_CA_ALERT | FLAG_SYN_ACKED))) {
3332 /* Conservatively mark a delayed ACK. It's typically
3333 * from a lone runt packet over the round trip to
3334 * a receiver w/o out-of-order or CE events.
3336 flag |= FLAG_ACK_MAYBE_DELAYED;
3339 if (sack->first_sackt) {
3340 sack_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->first_sackt);
3341 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->last_sackt);
3343 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us,
3344 ca_rtt_us, sack->rate);
3346 if (flag & FLAG_ACKED) {
3347 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3348 if (unlikely(icsk->icsk_mtup.probe_size &&
3349 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3350 tcp_mtup_probe_success(sk);
3353 if (tcp_is_reno(tp)) {
3354 tcp_remove_reno_sacks(sk, pkts_acked, ece_ack);
3356 /* If any of the cumulatively ACKed segments was
3357 * retransmitted, non-SACK case cannot confirm that
3358 * progress was due to original transmission due to
3359 * lack of TCPCB_SACKED_ACKED bits even if some of
3360 * the packets may have been never retransmitted.
3362 if (flag & FLAG_RETRANS_DATA_ACKED)
3363 flag &= ~FLAG_ORIG_SACK_ACKED;
3367 /* Non-retransmitted hole got filled? That's reordering */
3368 if (before(reord, prior_fack))
3369 tcp_check_sack_reordering(sk, reord, 0);
3371 delta = prior_sacked - tp->sacked_out;
3372 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3374 } else if (skb && rtt_update && sack_rtt_us >= 0 &&
3375 sack_rtt_us > tcp_stamp_us_delta(tp->tcp_mstamp,
3376 tcp_skb_timestamp_us(skb))) {
3377 /* Do not re-arm RTO if the sack RTT is measured from data sent
3378 * after when the head was last (re)transmitted. Otherwise the
3379 * timeout may continue to extend in loss recovery.
3381 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3384 if (icsk->icsk_ca_ops->pkts_acked) {
3385 struct ack_sample sample = { .pkts_acked = pkts_acked,
3386 .rtt_us = sack->rate->rtt_us,
3387 .in_flight = last_in_flight };
3389 icsk->icsk_ca_ops->pkts_acked(sk, &sample);
3392 #if FASTRETRANS_DEBUG > 0
3393 WARN_ON((int)tp->sacked_out < 0);
3394 WARN_ON((int)tp->lost_out < 0);
3395 WARN_ON((int)tp->retrans_out < 0);
3396 if (!tp->packets_out && tcp_is_sack(tp)) {
3397 icsk = inet_csk(sk);
3399 pr_debug("Leak l=%u %d\n",
3400 tp->lost_out, icsk->icsk_ca_state);
3403 if (tp->sacked_out) {
3404 pr_debug("Leak s=%u %d\n",
3405 tp->sacked_out, icsk->icsk_ca_state);
3408 if (tp->retrans_out) {
3409 pr_debug("Leak r=%u %d\n",
3410 tp->retrans_out, icsk->icsk_ca_state);
3411 tp->retrans_out = 0;
3418 static void tcp_ack_probe(struct sock *sk)
3420 struct inet_connection_sock *icsk = inet_csk(sk);
3421 struct sk_buff *head = tcp_send_head(sk);
3422 const struct tcp_sock *tp = tcp_sk(sk);
3424 /* Was it a usable window open? */
3427 if (!after(TCP_SKB_CB(head)->end_seq, tcp_wnd_end(tp))) {
3428 icsk->icsk_backoff = 0;
3429 icsk->icsk_probes_tstamp = 0;
3430 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3431 /* Socket must be waked up by subsequent tcp_data_snd_check().
3432 * This function is not for random using!
3435 unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX);
3437 when = tcp_clamp_probe0_to_user_timeout(sk, when);
3438 tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, when, TCP_RTO_MAX);
3442 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3444 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3445 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3448 /* Decide wheather to run the increase function of congestion control. */
3449 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3451 /* If reordering is high then always grow cwnd whenever data is
3452 * delivered regardless of its ordering. Otherwise stay conservative
3453 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3454 * new SACK or ECE mark may first advance cwnd here and later reduce
3455 * cwnd in tcp_fastretrans_alert() based on more states.
3457 if (tcp_sk(sk)->reordering > sock_net(sk)->ipv4.sysctl_tcp_reordering)
3458 return flag & FLAG_FORWARD_PROGRESS;
3460 return flag & FLAG_DATA_ACKED;
3463 /* The "ultimate" congestion control function that aims to replace the rigid
3464 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction).
3465 * It's called toward the end of processing an ACK with precise rate
3466 * information. All transmission or retransmission are delayed afterwards.
3468 static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked,
3469 int flag, const struct rate_sample *rs)
3471 const struct inet_connection_sock *icsk = inet_csk(sk);
3473 if (icsk->icsk_ca_ops->cong_control) {
3474 icsk->icsk_ca_ops->cong_control(sk, rs);
3478 if (tcp_in_cwnd_reduction(sk)) {
3479 /* Reduce cwnd if state mandates */
3480 tcp_cwnd_reduction(sk, acked_sacked, rs->losses, flag);
3481 } else if (tcp_may_raise_cwnd(sk, flag)) {
3482 /* Advance cwnd if state allows */
3483 tcp_cong_avoid(sk, ack, acked_sacked);
3485 tcp_update_pacing_rate(sk);
3488 /* Check that window update is acceptable.
3489 * The function assumes that snd_una<=ack<=snd_next.
3491 static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3492 const u32 ack, const u32 ack_seq,
3495 return after(ack, tp->snd_una) ||
3496 after(ack_seq, tp->snd_wl1) ||
3497 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3500 /* If we update tp->snd_una, also update tp->bytes_acked */
3501 static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack)
3503 u32 delta = ack - tp->snd_una;
3505 sock_owned_by_me((struct sock *)tp);
3506 tp->bytes_acked += delta;
3510 /* If we update tp->rcv_nxt, also update tp->bytes_received */
3511 static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq)
3513 u32 delta = seq - tp->rcv_nxt;
3515 sock_owned_by_me((struct sock *)tp);
3516 tp->bytes_received += delta;
3517 WRITE_ONCE(tp->rcv_nxt, seq);
3520 /* Update our send window.
3522 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3523 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3525 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3528 struct tcp_sock *tp = tcp_sk(sk);
3530 u32 nwin = ntohs(tcp_hdr(skb)->window);
3532 if (likely(!tcp_hdr(skb)->syn))
3533 nwin <<= tp->rx_opt.snd_wscale;
3535 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3536 flag |= FLAG_WIN_UPDATE;
3537 tcp_update_wl(tp, ack_seq);
3539 if (tp->snd_wnd != nwin) {
3542 /* Note, it is the only place, where
3543 * fast path is recovered for sending TCP.
3546 tcp_fast_path_check(sk);
3548 if (!tcp_write_queue_empty(sk))
3549 tcp_slow_start_after_idle_check(sk);
3551 if (nwin > tp->max_window) {
3552 tp->max_window = nwin;
3553 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3558 tcp_snd_una_update(tp, ack);
3563 static bool __tcp_oow_rate_limited(struct net *net, int mib_idx,
3564 u32 *last_oow_ack_time)
3566 if (*last_oow_ack_time) {
3567 s32 elapsed = (s32)(tcp_jiffies32 - *last_oow_ack_time);
3569 if (0 <= elapsed && elapsed < net->ipv4.sysctl_tcp_invalid_ratelimit) {
3570 NET_INC_STATS(net, mib_idx);
3571 return true; /* rate-limited: don't send yet! */
3575 *last_oow_ack_time = tcp_jiffies32;
3577 return false; /* not rate-limited: go ahead, send dupack now! */
3580 /* Return true if we're currently rate-limiting out-of-window ACKs and
3581 * thus shouldn't send a dupack right now. We rate-limit dupacks in
3582 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3583 * attacks that send repeated SYNs or ACKs for the same connection. To
3584 * do this, we do not send a duplicate SYNACK or ACK if the remote
3585 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3587 bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb,
3588 int mib_idx, u32 *last_oow_ack_time)
3590 /* Data packets without SYNs are not likely part of an ACK loop. */
3591 if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) &&
3595 return __tcp_oow_rate_limited(net, mib_idx, last_oow_ack_time);
3598 /* RFC 5961 7 [ACK Throttling] */
3599 static void tcp_send_challenge_ack(struct sock *sk, const struct sk_buff *skb)
3601 /* unprotected vars, we dont care of overwrites */
3602 static u32 challenge_timestamp;
3603 static unsigned int challenge_count;
3604 struct tcp_sock *tp = tcp_sk(sk);
3605 struct net *net = sock_net(sk);
3608 /* First check our per-socket dupack rate limit. */
3609 if (__tcp_oow_rate_limited(net,
3610 LINUX_MIB_TCPACKSKIPPEDCHALLENGE,
3611 &tp->last_oow_ack_time))
3614 /* Then check host-wide RFC 5961 rate limit. */
3616 if (now != challenge_timestamp) {
3617 u32 ack_limit = net->ipv4.sysctl_tcp_challenge_ack_limit;
3618 u32 half = (ack_limit + 1) >> 1;
3620 challenge_timestamp = now;
3621 WRITE_ONCE(challenge_count, half + prandom_u32_max(ack_limit));
3623 count = READ_ONCE(challenge_count);
3625 WRITE_ONCE(challenge_count, count - 1);
3626 NET_INC_STATS(net, LINUX_MIB_TCPCHALLENGEACK);
3631 static void tcp_store_ts_recent(struct tcp_sock *tp)
3633 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3634 tp->rx_opt.ts_recent_stamp = ktime_get_seconds();
3637 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3639 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3640 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3641 * extra check below makes sure this can only happen
3642 * for pure ACK frames. -DaveM
3644 * Not only, also it occurs for expired timestamps.
3647 if (tcp_paws_check(&tp->rx_opt, 0))
3648 tcp_store_ts_recent(tp);
3652 /* This routine deals with acks during a TLP episode and ends an episode by
3653 * resetting tlp_high_seq. Ref: TLP algorithm in draft-ietf-tcpm-rack
3655 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3657 struct tcp_sock *tp = tcp_sk(sk);
3659 if (before(ack, tp->tlp_high_seq))
3662 if (!tp->tlp_retrans) {
3663 /* TLP of new data has been acknowledged */
3664 tp->tlp_high_seq = 0;
3665 } else if (flag & FLAG_DSACK_TLP) {
3666 /* This DSACK means original and TLP probe arrived; no loss */
3667 tp->tlp_high_seq = 0;
3668 } else if (after(ack, tp->tlp_high_seq)) {
3669 /* ACK advances: there was a loss, so reduce cwnd. Reset
3670 * tlp_high_seq in tcp_init_cwnd_reduction()
3672 tcp_init_cwnd_reduction(sk);
3673 tcp_set_ca_state(sk, TCP_CA_CWR);
3674 tcp_end_cwnd_reduction(sk);
3675 tcp_try_keep_open(sk);
3676 NET_INC_STATS(sock_net(sk),
3677 LINUX_MIB_TCPLOSSPROBERECOVERY);
3678 } else if (!(flag & (FLAG_SND_UNA_ADVANCED |
3679 FLAG_NOT_DUP | FLAG_DATA_SACKED))) {
3680 /* Pure dupack: original and TLP probe arrived; no loss */
3681 tp->tlp_high_seq = 0;
3685 static inline void tcp_in_ack_event(struct sock *sk, u32 flags)
3687 const struct inet_connection_sock *icsk = inet_csk(sk);
3689 if (icsk->icsk_ca_ops->in_ack_event)
3690 icsk->icsk_ca_ops->in_ack_event(sk, flags);
3693 /* Congestion control has updated the cwnd already. So if we're in
3694 * loss recovery then now we do any new sends (for FRTO) or
3695 * retransmits (for CA_Loss or CA_recovery) that make sense.
3697 static void tcp_xmit_recovery(struct sock *sk, int rexmit)
3699 struct tcp_sock *tp = tcp_sk(sk);
3701 if (rexmit == REXMIT_NONE || sk->sk_state == TCP_SYN_SENT)
3704 if (unlikely(rexmit == REXMIT_NEW)) {
3705 __tcp_push_pending_frames(sk, tcp_current_mss(sk),
3707 if (after(tp->snd_nxt, tp->high_seq))
3711 tcp_xmit_retransmit_queue(sk);
3714 /* Returns the number of packets newly acked or sacked by the current ACK */
3715 static u32 tcp_newly_delivered(struct sock *sk, u32 prior_delivered, int flag)
3717 const struct net *net = sock_net(sk);
3718 struct tcp_sock *tp = tcp_sk(sk);
3721 delivered = tp->delivered - prior_delivered;
3722 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVERED, delivered);
3723 if (flag & FLAG_ECE)
3724 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVEREDCE, delivered);
3729 /* This routine deals with incoming acks, but not outgoing ones. */
3730 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3732 struct inet_connection_sock *icsk = inet_csk(sk);
3733 struct tcp_sock *tp = tcp_sk(sk);
3734 struct tcp_sacktag_state sack_state;
3735 struct rate_sample rs = { .prior_delivered = 0 };
3736 u32 prior_snd_una = tp->snd_una;
3737 bool is_sack_reneg = tp->is_sack_reneg;
3738 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3739 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3741 int prior_packets = tp->packets_out;
3742 u32 delivered = tp->delivered;
3743 u32 lost = tp->lost;
3744 int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */
3747 sack_state.first_sackt = 0;
3748 sack_state.rate = &rs;
3749 sack_state.sack_delivered = 0;
3751 /* We very likely will need to access rtx queue. */
3752 prefetch(sk->tcp_rtx_queue.rb_node);
3754 /* If the ack is older than previous acks
3755 * then we can probably ignore it.
3757 if (before(ack, prior_snd_una)) {
3758 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3759 if (before(ack, prior_snd_una - tp->max_window)) {
3760 if (!(flag & FLAG_NO_CHALLENGE_ACK))
3761 tcp_send_challenge_ack(sk, skb);
3767 /* If the ack includes data we haven't sent yet, discard
3768 * this segment (RFC793 Section 3.9).
3770 if (after(ack, tp->snd_nxt))
3773 if (after(ack, prior_snd_una)) {
3774 flag |= FLAG_SND_UNA_ADVANCED;
3775 icsk->icsk_retransmits = 0;
3777 #if IS_ENABLED(CONFIG_TLS_DEVICE)
3778 if (static_branch_unlikely(&clean_acked_data_enabled.key))
3779 if (icsk->icsk_clean_acked)
3780 icsk->icsk_clean_acked(sk, ack);
3784 prior_fack = tcp_is_sack(tp) ? tcp_highest_sack_seq(tp) : tp->snd_una;
3785 rs.prior_in_flight = tcp_packets_in_flight(tp);
3787 /* ts_recent update must be made after we are sure that the packet
3790 if (flag & FLAG_UPDATE_TS_RECENT)
3791 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3793 if ((flag & (FLAG_SLOWPATH | FLAG_SND_UNA_ADVANCED)) ==
3794 FLAG_SND_UNA_ADVANCED) {
3795 /* Window is constant, pure forward advance.
3796 * No more checks are required.
3797 * Note, we use the fact that SND.UNA>=SND.WL2.
3799 tcp_update_wl(tp, ack_seq);
3800 tcp_snd_una_update(tp, ack);
3801 flag |= FLAG_WIN_UPDATE;
3803 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
3805 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPACKS);
3807 u32 ack_ev_flags = CA_ACK_SLOWPATH;
3809 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3812 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3814 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3816 if (TCP_SKB_CB(skb)->sacked)
3817 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3820 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) {
3822 ack_ev_flags |= CA_ACK_ECE;
3825 if (sack_state.sack_delivered)
3826 tcp_count_delivered(tp, sack_state.sack_delivered,
3829 if (flag & FLAG_WIN_UPDATE)
3830 ack_ev_flags |= CA_ACK_WIN_UPDATE;
3832 tcp_in_ack_event(sk, ack_ev_flags);
3835 /* This is a deviation from RFC3168 since it states that:
3836 * "When the TCP data sender is ready to set the CWR bit after reducing
3837 * the congestion window, it SHOULD set the CWR bit only on the first
3838 * new data packet that it transmits."
3839 * We accept CWR on pure ACKs to be more robust
3840 * with widely-deployed TCP implementations that do this.
3842 tcp_ecn_accept_cwr(sk, skb);
3844 /* We passed data and got it acked, remove any soft error
3845 * log. Something worked...
3847 sk->sk_err_soft = 0;
3848 icsk->icsk_probes_out = 0;
3849 tp->rcv_tstamp = tcp_jiffies32;
3853 /* See if we can take anything off of the retransmit queue. */
3854 flag |= tcp_clean_rtx_queue(sk, skb, prior_fack, prior_snd_una,
3855 &sack_state, flag & FLAG_ECE);
3857 tcp_rack_update_reo_wnd(sk, &rs);
3859 if (tp->tlp_high_seq)
3860 tcp_process_tlp_ack(sk, ack, flag);
3862 if (tcp_ack_is_dubious(sk, flag)) {
3863 if (!(flag & (FLAG_SND_UNA_ADVANCED |
3864 FLAG_NOT_DUP | FLAG_DSACKING_ACK))) {
3866 /* Consider if pure acks were aggregated in tcp_add_backlog() */
3867 if (!(flag & FLAG_DATA))
3868 num_dupack = max_t(u16, 1, skb_shinfo(skb)->gso_segs);
3870 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
3874 /* If needed, reset TLP/RTO timer when RACK doesn't set. */
3875 if (flag & FLAG_SET_XMIT_TIMER)
3876 tcp_set_xmit_timer(sk);
3878 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
3881 delivered = tcp_newly_delivered(sk, delivered, flag);
3882 lost = tp->lost - lost; /* freshly marked lost */
3883 rs.is_ack_delayed = !!(flag & FLAG_ACK_MAYBE_DELAYED);
3884 tcp_rate_gen(sk, delivered, lost, is_sack_reneg, sack_state.rate);
3885 tcp_cong_control(sk, ack, delivered, flag, sack_state.rate);
3886 tcp_xmit_recovery(sk, rexmit);
3890 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3891 if (flag & FLAG_DSACKING_ACK) {
3892 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
3894 tcp_newly_delivered(sk, delivered, flag);
3896 /* If this ack opens up a zero window, clear backoff. It was
3897 * being used to time the probes, and is probably far higher than
3898 * it needs to be for normal retransmission.
3902 if (tp->tlp_high_seq)
3903 tcp_process_tlp_ack(sk, ack, flag);
3907 /* If data was SACKed, tag it and see if we should send more data.
3908 * If data was DSACKed, see if we can undo a cwnd reduction.
3910 if (TCP_SKB_CB(skb)->sacked) {
3911 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3913 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
3915 tcp_newly_delivered(sk, delivered, flag);
3916 tcp_xmit_recovery(sk, rexmit);
3922 static void tcp_parse_fastopen_option(int len, const unsigned char *cookie,
3923 bool syn, struct tcp_fastopen_cookie *foc,
3926 /* Valid only in SYN or SYN-ACK with an even length. */
3927 if (!foc || !syn || len < 0 || (len & 1))
3930 if (len >= TCP_FASTOPEN_COOKIE_MIN &&
3931 len <= TCP_FASTOPEN_COOKIE_MAX)
3932 memcpy(foc->val, cookie, len);
3939 static bool smc_parse_options(const struct tcphdr *th,
3940 struct tcp_options_received *opt_rx,
3941 const unsigned char *ptr,
3944 #if IS_ENABLED(CONFIG_SMC)
3945 if (static_branch_unlikely(&tcp_have_smc)) {
3946 if (th->syn && !(opsize & 1) &&
3947 opsize >= TCPOLEN_EXP_SMC_BASE &&
3948 get_unaligned_be32(ptr) == TCPOPT_SMC_MAGIC) {
3957 /* Try to parse the MSS option from the TCP header. Return 0 on failure, clamped
3960 static u16 tcp_parse_mss_option(const struct tcphdr *th, u16 user_mss)
3962 const unsigned char *ptr = (const unsigned char *)(th + 1);
3963 int length = (th->doff * 4) - sizeof(struct tcphdr);
3966 while (length > 0) {
3967 int opcode = *ptr++;
3973 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3980 if (opsize < 2) /* "silly options" */
3982 if (opsize > length)
3983 return mss; /* fail on partial options */
3984 if (opcode == TCPOPT_MSS && opsize == TCPOLEN_MSS) {
3985 u16 in_mss = get_unaligned_be16(ptr);
3988 if (user_mss && user_mss < in_mss)
4000 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
4001 * But, this can also be called on packets in the established flow when
4002 * the fast version below fails.
4004 void tcp_parse_options(const struct net *net,
4005 const struct sk_buff *skb,
4006 struct tcp_options_received *opt_rx, int estab,
4007 struct tcp_fastopen_cookie *foc)
4009 const unsigned char *ptr;
4010 const struct tcphdr *th = tcp_hdr(skb);
4011 int length = (th->doff * 4) - sizeof(struct tcphdr);
4013 ptr = (const unsigned char *)(th + 1);
4014 opt_rx->saw_tstamp = 0;
4015 opt_rx->saw_unknown = 0;
4017 while (length > 0) {
4018 int opcode = *ptr++;
4024 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
4031 if (opsize < 2) /* "silly options" */
4033 if (opsize > length)
4034 return; /* don't parse partial options */
4037 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
4038 u16 in_mss = get_unaligned_be16(ptr);
4040 if (opt_rx->user_mss &&
4041 opt_rx->user_mss < in_mss)
4042 in_mss = opt_rx->user_mss;
4043 opt_rx->mss_clamp = in_mss;
4048 if (opsize == TCPOLEN_WINDOW && th->syn &&
4049 !estab && net->ipv4.sysctl_tcp_window_scaling) {
4050 __u8 snd_wscale = *(__u8 *)ptr;
4051 opt_rx->wscale_ok = 1;
4052 if (snd_wscale > TCP_MAX_WSCALE) {
4053 net_info_ratelimited("%s: Illegal window scaling value %d > %u received\n",
4057 snd_wscale = TCP_MAX_WSCALE;
4059 opt_rx->snd_wscale = snd_wscale;
4062 case TCPOPT_TIMESTAMP:
4063 if ((opsize == TCPOLEN_TIMESTAMP) &&
4064 ((estab && opt_rx->tstamp_ok) ||
4065 (!estab && net->ipv4.sysctl_tcp_timestamps))) {
4066 opt_rx->saw_tstamp = 1;
4067 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
4068 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
4071 case TCPOPT_SACK_PERM:
4072 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
4073 !estab && net->ipv4.sysctl_tcp_sack) {
4074 opt_rx->sack_ok = TCP_SACK_SEEN;
4075 tcp_sack_reset(opt_rx);
4080 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
4081 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
4083 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
4086 #ifdef CONFIG_TCP_MD5SIG
4089 * The MD5 Hash has already been
4090 * checked (see tcp_v{4,6}_do_rcv()).
4094 case TCPOPT_FASTOPEN:
4095 tcp_parse_fastopen_option(
4096 opsize - TCPOLEN_FASTOPEN_BASE,
4097 ptr, th->syn, foc, false);
4101 /* Fast Open option shares code 254 using a
4102 * 16 bits magic number.
4104 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE &&
4105 get_unaligned_be16(ptr) ==
4106 TCPOPT_FASTOPEN_MAGIC) {
4107 tcp_parse_fastopen_option(opsize -
4108 TCPOLEN_EXP_FASTOPEN_BASE,
4109 ptr + 2, th->syn, foc, true);
4113 if (smc_parse_options(th, opt_rx, ptr, opsize))
4116 opt_rx->saw_unknown = 1;
4120 opt_rx->saw_unknown = 1;
4127 EXPORT_SYMBOL(tcp_parse_options);
4129 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
4131 const __be32 *ptr = (const __be32 *)(th + 1);
4133 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
4134 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
4135 tp->rx_opt.saw_tstamp = 1;
4137 tp->rx_opt.rcv_tsval = ntohl(*ptr);
4140 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
4142 tp->rx_opt.rcv_tsecr = 0;
4148 /* Fast parse options. This hopes to only see timestamps.
4149 * If it is wrong it falls back on tcp_parse_options().
4151 static bool tcp_fast_parse_options(const struct net *net,
4152 const struct sk_buff *skb,
4153 const struct tcphdr *th, struct tcp_sock *tp)
4155 /* In the spirit of fast parsing, compare doff directly to constant
4156 * values. Because equality is used, short doff can be ignored here.
4158 if (th->doff == (sizeof(*th) / 4)) {
4159 tp->rx_opt.saw_tstamp = 0;
4161 } else if (tp->rx_opt.tstamp_ok &&
4162 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
4163 if (tcp_parse_aligned_timestamp(tp, th))
4167 tcp_parse_options(net, skb, &tp->rx_opt, 1, NULL);
4168 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
4169 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
4174 #ifdef CONFIG_TCP_MD5SIG
4176 * Parse MD5 Signature option
4178 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
4180 int length = (th->doff << 2) - sizeof(*th);
4181 const u8 *ptr = (const u8 *)(th + 1);
4183 /* If not enough data remaining, we can short cut */
4184 while (length >= TCPOLEN_MD5SIG) {
4185 int opcode = *ptr++;
4196 if (opsize < 2 || opsize > length)
4198 if (opcode == TCPOPT_MD5SIG)
4199 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
4206 EXPORT_SYMBOL(tcp_parse_md5sig_option);
4209 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
4211 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
4212 * it can pass through stack. So, the following predicate verifies that
4213 * this segment is not used for anything but congestion avoidance or
4214 * fast retransmit. Moreover, we even are able to eliminate most of such
4215 * second order effects, if we apply some small "replay" window (~RTO)
4216 * to timestamp space.
4218 * All these measures still do not guarantee that we reject wrapped ACKs
4219 * on networks with high bandwidth, when sequence space is recycled fastly,
4220 * but it guarantees that such events will be very rare and do not affect
4221 * connection seriously. This doesn't look nice, but alas, PAWS is really
4224 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
4225 * states that events when retransmit arrives after original data are rare.
4226 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
4227 * the biggest problem on large power networks even with minor reordering.
4228 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
4229 * up to bandwidth of 18Gigabit/sec. 8) ]
4232 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
4234 const struct tcp_sock *tp = tcp_sk(sk);
4235 const struct tcphdr *th = tcp_hdr(skb);
4236 u32 seq = TCP_SKB_CB(skb)->seq;
4237 u32 ack = TCP_SKB_CB(skb)->ack_seq;
4239 return (/* 1. Pure ACK with correct sequence number. */
4240 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
4242 /* 2. ... and duplicate ACK. */
4243 ack == tp->snd_una &&
4245 /* 3. ... and does not update window. */
4246 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
4248 /* 4. ... and sits in replay window. */
4249 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
4252 static inline bool tcp_paws_discard(const struct sock *sk,
4253 const struct sk_buff *skb)
4255 const struct tcp_sock *tp = tcp_sk(sk);
4257 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
4258 !tcp_disordered_ack(sk, skb);
4261 /* Check segment sequence number for validity.
4263 * Segment controls are considered valid, if the segment
4264 * fits to the window after truncation to the window. Acceptability
4265 * of data (and SYN, FIN, of course) is checked separately.
4266 * See tcp_data_queue(), for example.
4268 * Also, controls (RST is main one) are accepted using RCV.WUP instead
4269 * of RCV.NXT. Peer still did not advance his SND.UNA when we
4270 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4271 * (borrowed from freebsd)
4274 static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
4276 return !before(end_seq, tp->rcv_wup) &&
4277 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
4280 /* When we get a reset we do this. */
4281 void tcp_reset(struct sock *sk, struct sk_buff *skb)
4283 trace_tcp_receive_reset(sk);
4285 /* mptcp can't tell us to ignore reset pkts,
4286 * so just ignore the return value of mptcp_incoming_options().
4288 if (sk_is_mptcp(sk))
4289 mptcp_incoming_options(sk, skb);
4291 /* We want the right error as BSD sees it (and indeed as we do). */
4292 switch (sk->sk_state) {
4294 sk->sk_err = ECONNREFUSED;
4296 case TCP_CLOSE_WAIT:
4302 sk->sk_err = ECONNRESET;
4304 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4307 tcp_write_queue_purge(sk);
4310 if (!sock_flag(sk, SOCK_DEAD))
4311 sk_error_report(sk);
4315 * Process the FIN bit. This now behaves as it is supposed to work
4316 * and the FIN takes effect when it is validly part of sequence
4317 * space. Not before when we get holes.
4319 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4320 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4323 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4324 * close and we go into CLOSING (and later onto TIME-WAIT)
4326 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4328 void tcp_fin(struct sock *sk)
4330 struct tcp_sock *tp = tcp_sk(sk);
4332 inet_csk_schedule_ack(sk);
4334 sk->sk_shutdown |= RCV_SHUTDOWN;
4335 sock_set_flag(sk, SOCK_DONE);
4337 switch (sk->sk_state) {
4339 case TCP_ESTABLISHED:
4340 /* Move to CLOSE_WAIT */
4341 tcp_set_state(sk, TCP_CLOSE_WAIT);
4342 inet_csk_enter_pingpong_mode(sk);
4345 case TCP_CLOSE_WAIT:
4347 /* Received a retransmission of the FIN, do
4352 /* RFC793: Remain in the LAST-ACK state. */
4356 /* This case occurs when a simultaneous close
4357 * happens, we must ack the received FIN and
4358 * enter the CLOSING state.
4361 tcp_set_state(sk, TCP_CLOSING);
4364 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4366 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4369 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4370 * cases we should never reach this piece of code.
4372 pr_err("%s: Impossible, sk->sk_state=%d\n",
4373 __func__, sk->sk_state);
4377 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4378 * Probably, we should reset in this case. For now drop them.
4380 skb_rbtree_purge(&tp->out_of_order_queue);
4381 if (tcp_is_sack(tp))
4382 tcp_sack_reset(&tp->rx_opt);
4385 if (!sock_flag(sk, SOCK_DEAD)) {
4386 sk->sk_state_change(sk);
4388 /* Do not send POLL_HUP for half duplex close. */
4389 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4390 sk->sk_state == TCP_CLOSE)
4391 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4393 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4397 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4400 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4401 if (before(seq, sp->start_seq))
4402 sp->start_seq = seq;
4403 if (after(end_seq, sp->end_seq))
4404 sp->end_seq = end_seq;
4410 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4412 struct tcp_sock *tp = tcp_sk(sk);
4414 if (tcp_is_sack(tp) && sock_net(sk)->ipv4.sysctl_tcp_dsack) {
4417 if (before(seq, tp->rcv_nxt))
4418 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4420 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4422 NET_INC_STATS(sock_net(sk), mib_idx);
4424 tp->rx_opt.dsack = 1;
4425 tp->duplicate_sack[0].start_seq = seq;
4426 tp->duplicate_sack[0].end_seq = end_seq;
4430 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4432 struct tcp_sock *tp = tcp_sk(sk);
4434 if (!tp->rx_opt.dsack)
4435 tcp_dsack_set(sk, seq, end_seq);
4437 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4440 static void tcp_rcv_spurious_retrans(struct sock *sk, const struct sk_buff *skb)
4442 /* When the ACK path fails or drops most ACKs, the sender would
4443 * timeout and spuriously retransmit the same segment repeatedly.
4444 * The receiver remembers and reflects via DSACKs. Leverage the
4445 * DSACK state and change the txhash to re-route speculatively.
4447 if (TCP_SKB_CB(skb)->seq == tcp_sk(sk)->duplicate_sack[0].start_seq &&
4448 sk_rethink_txhash(sk))
4449 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDUPLICATEDATAREHASH);
4452 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4454 struct tcp_sock *tp = tcp_sk(sk);
4456 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4457 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4458 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4459 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
4461 if (tcp_is_sack(tp) && sock_net(sk)->ipv4.sysctl_tcp_dsack) {
4462 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4464 tcp_rcv_spurious_retrans(sk, skb);
4465 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4466 end_seq = tp->rcv_nxt;
4467 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4474 /* These routines update the SACK block as out-of-order packets arrive or
4475 * in-order packets close up the sequence space.
4477 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4480 struct tcp_sack_block *sp = &tp->selective_acks[0];
4481 struct tcp_sack_block *swalk = sp + 1;
4483 /* See if the recent change to the first SACK eats into
4484 * or hits the sequence space of other SACK blocks, if so coalesce.
4486 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4487 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4490 /* Zap SWALK, by moving every further SACK up by one slot.
4491 * Decrease num_sacks.
4493 tp->rx_opt.num_sacks--;
4494 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4503 static void tcp_sack_compress_send_ack(struct sock *sk)
4505 struct tcp_sock *tp = tcp_sk(sk);
4507 if (!tp->compressed_ack)
4510 if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1)
4513 /* Since we have to send one ack finally,
4514 * substract one from tp->compressed_ack to keep
4515 * LINUX_MIB_TCPACKCOMPRESSED accurate.
4517 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED,
4518 tp->compressed_ack - 1);
4520 tp->compressed_ack = 0;
4524 /* Reasonable amount of sack blocks included in TCP SACK option
4525 * The max is 4, but this becomes 3 if TCP timestamps are there.
4526 * Given that SACK packets might be lost, be conservative and use 2.
4528 #define TCP_SACK_BLOCKS_EXPECTED 2
4530 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4532 struct tcp_sock *tp = tcp_sk(sk);
4533 struct tcp_sack_block *sp = &tp->selective_acks[0];
4534 int cur_sacks = tp->rx_opt.num_sacks;
4540 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4541 if (tcp_sack_extend(sp, seq, end_seq)) {
4542 if (this_sack >= TCP_SACK_BLOCKS_EXPECTED)
4543 tcp_sack_compress_send_ack(sk);
4544 /* Rotate this_sack to the first one. */
4545 for (; this_sack > 0; this_sack--, sp--)
4546 swap(*sp, *(sp - 1));
4548 tcp_sack_maybe_coalesce(tp);
4553 if (this_sack >= TCP_SACK_BLOCKS_EXPECTED)
4554 tcp_sack_compress_send_ack(sk);
4556 /* Could not find an adjacent existing SACK, build a new one,
4557 * put it at the front, and shift everyone else down. We
4558 * always know there is at least one SACK present already here.
4560 * If the sack array is full, forget about the last one.
4562 if (this_sack >= TCP_NUM_SACKS) {
4564 tp->rx_opt.num_sacks--;
4567 for (; this_sack > 0; this_sack--, sp--)
4571 /* Build the new head SACK, and we're done. */
4572 sp->start_seq = seq;
4573 sp->end_seq = end_seq;
4574 tp->rx_opt.num_sacks++;
4577 /* RCV.NXT advances, some SACKs should be eaten. */
4579 static void tcp_sack_remove(struct tcp_sock *tp)
4581 struct tcp_sack_block *sp = &tp->selective_acks[0];
4582 int num_sacks = tp->rx_opt.num_sacks;
4585 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4586 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4587 tp->rx_opt.num_sacks = 0;
4591 for (this_sack = 0; this_sack < num_sacks;) {
4592 /* Check if the start of the sack is covered by RCV.NXT. */
4593 if (!before(tp->rcv_nxt, sp->start_seq)) {
4596 /* RCV.NXT must cover all the block! */
4597 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4599 /* Zap this SACK, by moving forward any other SACKS. */
4600 for (i = this_sack+1; i < num_sacks; i++)
4601 tp->selective_acks[i-1] = tp->selective_acks[i];
4608 tp->rx_opt.num_sacks = num_sacks;
4612 * tcp_try_coalesce - try to merge skb to prior one
4615 * @from: buffer to add in queue
4616 * @fragstolen: pointer to boolean
4618 * Before queueing skb @from after @to, try to merge them
4619 * to reduce overall memory use and queue lengths, if cost is small.
4620 * Packets in ofo or receive queues can stay a long time.
4621 * Better try to coalesce them right now to avoid future collapses.
4622 * Returns true if caller should free @from instead of queueing it
4624 static bool tcp_try_coalesce(struct sock *sk,
4626 struct sk_buff *from,
4631 *fragstolen = false;
4633 /* Its possible this segment overlaps with prior segment in queue */
4634 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4637 if (!mptcp_skb_can_collapse(to, from))
4640 #ifdef CONFIG_TLS_DEVICE
4641 if (from->decrypted != to->decrypted)
4645 if (!skb_try_coalesce(to, from, fragstolen, &delta))
4648 atomic_add(delta, &sk->sk_rmem_alloc);
4649 sk_mem_charge(sk, delta);
4650 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4651 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4652 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4653 TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
4655 if (TCP_SKB_CB(from)->has_rxtstamp) {
4656 TCP_SKB_CB(to)->has_rxtstamp = true;
4657 to->tstamp = from->tstamp;
4658 skb_hwtstamps(to)->hwtstamp = skb_hwtstamps(from)->hwtstamp;
4664 static bool tcp_ooo_try_coalesce(struct sock *sk,
4666 struct sk_buff *from,
4669 bool res = tcp_try_coalesce(sk, to, from, fragstolen);
4671 /* In case tcp_drop() is called later, update to->gso_segs */
4673 u32 gso_segs = max_t(u16, 1, skb_shinfo(to)->gso_segs) +
4674 max_t(u16, 1, skb_shinfo(from)->gso_segs);
4676 skb_shinfo(to)->gso_segs = min_t(u32, gso_segs, 0xFFFF);
4681 static void tcp_drop(struct sock *sk, struct sk_buff *skb)
4683 sk_drops_add(sk, skb);
4687 /* This one checks to see if we can put data from the
4688 * out_of_order queue into the receive_queue.
4690 static void tcp_ofo_queue(struct sock *sk)
4692 struct tcp_sock *tp = tcp_sk(sk);
4693 __u32 dsack_high = tp->rcv_nxt;
4694 bool fin, fragstolen, eaten;
4695 struct sk_buff *skb, *tail;
4698 p = rb_first(&tp->out_of_order_queue);
4701 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4704 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4705 __u32 dsack = dsack_high;
4706 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4707 dsack_high = TCP_SKB_CB(skb)->end_seq;
4708 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4711 rb_erase(&skb->rbnode, &tp->out_of_order_queue);
4713 if (unlikely(!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))) {
4718 tail = skb_peek_tail(&sk->sk_receive_queue);
4719 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen);
4720 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4721 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
4723 __skb_queue_tail(&sk->sk_receive_queue, skb);
4725 kfree_skb_partial(skb, fragstolen);
4727 if (unlikely(fin)) {
4729 /* tcp_fin() purges tp->out_of_order_queue,
4730 * so we must end this loop right now.
4737 static bool tcp_prune_ofo_queue(struct sock *sk);
4738 static int tcp_prune_queue(struct sock *sk);
4740 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4743 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4744 !sk_rmem_schedule(sk, skb, size)) {
4746 if (tcp_prune_queue(sk) < 0)
4749 while (!sk_rmem_schedule(sk, skb, size)) {
4750 if (!tcp_prune_ofo_queue(sk))
4757 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4759 struct tcp_sock *tp = tcp_sk(sk);
4760 struct rb_node **p, *parent;
4761 struct sk_buff *skb1;
4765 tcp_ecn_check_ce(sk, skb);
4767 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4768 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFODROP);
4769 sk->sk_data_ready(sk);
4774 /* Disable header prediction. */
4776 inet_csk_schedule_ack(sk);
4778 tp->rcv_ooopack += max_t(u16, 1, skb_shinfo(skb)->gso_segs);
4779 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4780 seq = TCP_SKB_CB(skb)->seq;
4781 end_seq = TCP_SKB_CB(skb)->end_seq;
4783 p = &tp->out_of_order_queue.rb_node;
4784 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4785 /* Initial out of order segment, build 1 SACK. */
4786 if (tcp_is_sack(tp)) {
4787 tp->rx_opt.num_sacks = 1;
4788 tp->selective_acks[0].start_seq = seq;
4789 tp->selective_acks[0].end_seq = end_seq;
4791 rb_link_node(&skb->rbnode, NULL, p);
4792 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4793 tp->ooo_last_skb = skb;
4797 /* In the typical case, we are adding an skb to the end of the list.
4798 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
4800 if (tcp_ooo_try_coalesce(sk, tp->ooo_last_skb,
4801 skb, &fragstolen)) {
4803 /* For non sack flows, do not grow window to force DUPACK
4804 * and trigger fast retransmit.
4806 if (tcp_is_sack(tp))
4807 tcp_grow_window(sk, skb, true);
4808 kfree_skb_partial(skb, fragstolen);
4812 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
4813 if (!before(seq, TCP_SKB_CB(tp->ooo_last_skb)->end_seq)) {
4814 parent = &tp->ooo_last_skb->rbnode;
4815 p = &parent->rb_right;
4819 /* Find place to insert this segment. Handle overlaps on the way. */
4823 skb1 = rb_to_skb(parent);
4824 if (before(seq, TCP_SKB_CB(skb1)->seq)) {
4825 p = &parent->rb_left;
4828 if (before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4829 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4830 /* All the bits are present. Drop. */
4831 NET_INC_STATS(sock_net(sk),
4832 LINUX_MIB_TCPOFOMERGE);
4835 tcp_dsack_set(sk, seq, end_seq);
4838 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4839 /* Partial overlap. */
4840 tcp_dsack_set(sk, seq, TCP_SKB_CB(skb1)->end_seq);
4842 /* skb's seq == skb1's seq and skb covers skb1.
4843 * Replace skb1 with skb.
4845 rb_replace_node(&skb1->rbnode, &skb->rbnode,
4846 &tp->out_of_order_queue);
4847 tcp_dsack_extend(sk,
4848 TCP_SKB_CB(skb1)->seq,
4849 TCP_SKB_CB(skb1)->end_seq);
4850 NET_INC_STATS(sock_net(sk),
4851 LINUX_MIB_TCPOFOMERGE);
4855 } else if (tcp_ooo_try_coalesce(sk, skb1,
4856 skb, &fragstolen)) {
4859 p = &parent->rb_right;
4862 /* Insert segment into RB tree. */
4863 rb_link_node(&skb->rbnode, parent, p);
4864 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4867 /* Remove other segments covered by skb. */
4868 while ((skb1 = skb_rb_next(skb)) != NULL) {
4869 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4871 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4872 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4876 rb_erase(&skb1->rbnode, &tp->out_of_order_queue);
4877 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4878 TCP_SKB_CB(skb1)->end_seq);
4879 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4882 /* If there is no skb after us, we are the last_skb ! */
4884 tp->ooo_last_skb = skb;
4887 if (tcp_is_sack(tp))
4888 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4891 /* For non sack flows, do not grow window to force DUPACK
4892 * and trigger fast retransmit.
4894 if (tcp_is_sack(tp))
4895 tcp_grow_window(sk, skb, false);
4897 skb_set_owner_r(skb, sk);
4901 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb,
4905 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4908 tcp_try_coalesce(sk, tail,
4909 skb, fragstolen)) ? 1 : 0;
4910 tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq);
4912 __skb_queue_tail(&sk->sk_receive_queue, skb);
4913 skb_set_owner_r(skb, sk);
4918 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4920 struct sk_buff *skb;
4928 if (size > PAGE_SIZE) {
4929 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS);
4931 data_len = npages << PAGE_SHIFT;
4932 size = data_len + (size & ~PAGE_MASK);
4934 skb = alloc_skb_with_frags(size - data_len, data_len,
4935 PAGE_ALLOC_COSTLY_ORDER,
4936 &err, sk->sk_allocation);
4940 skb_put(skb, size - data_len);
4941 skb->data_len = data_len;
4944 if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) {
4945 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP);
4949 err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size);
4953 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4954 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4955 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4957 if (tcp_queue_rcv(sk, skb, &fragstolen)) {
4958 WARN_ON_ONCE(fragstolen); /* should not happen */
4970 void tcp_data_ready(struct sock *sk)
4972 if (tcp_epollin_ready(sk, sk->sk_rcvlowat) || sock_flag(sk, SOCK_DONE))
4973 sk->sk_data_ready(sk);
4976 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4978 struct tcp_sock *tp = tcp_sk(sk);
4982 /* If a subflow has been reset, the packet should not continue
4983 * to be processed, drop the packet.
4985 if (sk_is_mptcp(sk) && !mptcp_incoming_options(sk, skb)) {
4990 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
4995 __skb_pull(skb, tcp_hdr(skb)->doff * 4);
4997 tp->rx_opt.dsack = 0;
4999 /* Queue data for delivery to the user.
5000 * Packets in sequence go to the receive queue.
5001 * Out of sequence packets to the out_of_order_queue.
5003 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
5004 if (tcp_receive_window(tp) == 0) {
5005 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP);
5009 /* Ok. In sequence. In window. */
5011 if (skb_queue_len(&sk->sk_receive_queue) == 0)
5012 sk_forced_mem_schedule(sk, skb->truesize);
5013 else if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) {
5014 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP);
5015 sk->sk_data_ready(sk);
5019 eaten = tcp_queue_rcv(sk, skb, &fragstolen);
5021 tcp_event_data_recv(sk, skb);
5022 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
5025 if (!RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
5028 /* RFC5681. 4.2. SHOULD send immediate ACK, when
5029 * gap in queue is filled.
5031 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
5032 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW;
5035 if (tp->rx_opt.num_sacks)
5036 tcp_sack_remove(tp);
5038 tcp_fast_path_check(sk);
5041 kfree_skb_partial(skb, fragstolen);
5042 if (!sock_flag(sk, SOCK_DEAD))
5047 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
5048 tcp_rcv_spurious_retrans(sk, skb);
5049 /* A retransmit, 2nd most common case. Force an immediate ack. */
5050 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
5051 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
5054 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
5055 inet_csk_schedule_ack(sk);
5061 /* Out of window. F.e. zero window probe. */
5062 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
5065 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
5066 /* Partial packet, seq < rcv_next < end_seq */
5067 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
5069 /* If window is closed, drop tail of packet. But after
5070 * remembering D-SACK for its head made in previous line.
5072 if (!tcp_receive_window(tp)) {
5073 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP);
5079 tcp_data_queue_ofo(sk, skb);
5082 static struct sk_buff *tcp_skb_next(struct sk_buff *skb, struct sk_buff_head *list)
5085 return !skb_queue_is_last(list, skb) ? skb->next : NULL;
5087 return skb_rb_next(skb);
5090 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
5091 struct sk_buff_head *list,
5092 struct rb_root *root)
5094 struct sk_buff *next = tcp_skb_next(skb, list);
5097 __skb_unlink(skb, list);
5099 rb_erase(&skb->rbnode, root);
5102 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
5107 /* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */
5108 void tcp_rbtree_insert(struct rb_root *root, struct sk_buff *skb)
5110 struct rb_node **p = &root->rb_node;
5111 struct rb_node *parent = NULL;
5112 struct sk_buff *skb1;
5116 skb1 = rb_to_skb(parent);
5117 if (before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb1)->seq))
5118 p = &parent->rb_left;
5120 p = &parent->rb_right;
5122 rb_link_node(&skb->rbnode, parent, p);
5123 rb_insert_color(&skb->rbnode, root);
5126 /* Collapse contiguous sequence of skbs head..tail with
5127 * sequence numbers start..end.
5129 * If tail is NULL, this means until the end of the queue.
5131 * Segments with FIN/SYN are not collapsed (only because this
5135 tcp_collapse(struct sock *sk, struct sk_buff_head *list, struct rb_root *root,
5136 struct sk_buff *head, struct sk_buff *tail, u32 start, u32 end)
5138 struct sk_buff *skb = head, *n;
5139 struct sk_buff_head tmp;
5142 /* First, check that queue is collapsible and find
5143 * the point where collapsing can be useful.
5146 for (end_of_skbs = true; skb != NULL && skb != tail; skb = n) {
5147 n = tcp_skb_next(skb, list);
5149 /* No new bits? It is possible on ofo queue. */
5150 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
5151 skb = tcp_collapse_one(sk, skb, list, root);
5157 /* The first skb to collapse is:
5159 * - bloated or contains data before "start" or
5160 * overlaps to the next one and mptcp allow collapsing.
5162 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
5163 (tcp_win_from_space(sk, skb->truesize) > skb->len ||
5164 before(TCP_SKB_CB(skb)->seq, start))) {
5165 end_of_skbs = false;
5169 if (n && n != tail && mptcp_skb_can_collapse(skb, n) &&
5170 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(n)->seq) {
5171 end_of_skbs = false;
5175 /* Decided to skip this, advance start seq. */
5176 start = TCP_SKB_CB(skb)->end_seq;
5179 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
5182 __skb_queue_head_init(&tmp);
5184 while (before(start, end)) {
5185 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
5186 struct sk_buff *nskb;
5188 nskb = alloc_skb(copy, GFP_ATOMIC);
5192 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
5193 #ifdef CONFIG_TLS_DEVICE
5194 nskb->decrypted = skb->decrypted;
5196 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
5198 __skb_queue_before(list, skb, nskb);
5200 __skb_queue_tail(&tmp, nskb); /* defer rbtree insertion */
5201 skb_set_owner_r(nskb, sk);
5202 mptcp_skb_ext_move(nskb, skb);
5204 /* Copy data, releasing collapsed skbs. */
5206 int offset = start - TCP_SKB_CB(skb)->seq;
5207 int size = TCP_SKB_CB(skb)->end_seq - start;
5211 size = min(copy, size);
5212 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
5214 TCP_SKB_CB(nskb)->end_seq += size;
5218 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
5219 skb = tcp_collapse_one(sk, skb, list, root);
5222 !mptcp_skb_can_collapse(nskb, skb) ||
5223 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
5225 #ifdef CONFIG_TLS_DEVICE
5226 if (skb->decrypted != nskb->decrypted)
5233 skb_queue_walk_safe(&tmp, skb, n)
5234 tcp_rbtree_insert(root, skb);
5237 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
5238 * and tcp_collapse() them until all the queue is collapsed.
5240 static void tcp_collapse_ofo_queue(struct sock *sk)
5242 struct tcp_sock *tp = tcp_sk(sk);
5243 u32 range_truesize, sum_tiny = 0;
5244 struct sk_buff *skb, *head;
5247 skb = skb_rb_first(&tp->out_of_order_queue);
5250 tp->ooo_last_skb = skb_rb_last(&tp->out_of_order_queue);
5253 start = TCP_SKB_CB(skb)->seq;
5254 end = TCP_SKB_CB(skb)->end_seq;
5255 range_truesize = skb->truesize;
5257 for (head = skb;;) {
5258 skb = skb_rb_next(skb);
5260 /* Range is terminated when we see a gap or when
5261 * we are at the queue end.
5264 after(TCP_SKB_CB(skb)->seq, end) ||
5265 before(TCP_SKB_CB(skb)->end_seq, start)) {
5266 /* Do not attempt collapsing tiny skbs */
5267 if (range_truesize != head->truesize ||
5268 end - start >= SKB_WITH_OVERHEAD(SK_MEM_QUANTUM)) {
5269 tcp_collapse(sk, NULL, &tp->out_of_order_queue,
5270 head, skb, start, end);
5272 sum_tiny += range_truesize;
5273 if (sum_tiny > sk->sk_rcvbuf >> 3)
5279 range_truesize += skb->truesize;
5280 if (unlikely(before(TCP_SKB_CB(skb)->seq, start)))
5281 start = TCP_SKB_CB(skb)->seq;
5282 if (after(TCP_SKB_CB(skb)->end_seq, end))
5283 end = TCP_SKB_CB(skb)->end_seq;
5288 * Clean the out-of-order queue to make room.
5289 * We drop high sequences packets to :
5290 * 1) Let a chance for holes to be filled.
5291 * 2) not add too big latencies if thousands of packets sit there.
5292 * (But if application shrinks SO_RCVBUF, we could still end up
5293 * freeing whole queue here)
5294 * 3) Drop at least 12.5 % of sk_rcvbuf to avoid malicious attacks.
5296 * Return true if queue has shrunk.
5298 static bool tcp_prune_ofo_queue(struct sock *sk)
5300 struct tcp_sock *tp = tcp_sk(sk);
5301 struct rb_node *node, *prev;
5304 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
5307 NET_INC_STATS(sock_net(sk), LINUX_MIB_OFOPRUNED);
5308 goal = sk->sk_rcvbuf >> 3;
5309 node = &tp->ooo_last_skb->rbnode;
5311 prev = rb_prev(node);
5312 rb_erase(node, &tp->out_of_order_queue);
5313 goal -= rb_to_skb(node)->truesize;
5314 tcp_drop(sk, rb_to_skb(node));
5315 if (!prev || goal <= 0) {
5317 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf &&
5318 !tcp_under_memory_pressure(sk))
5320 goal = sk->sk_rcvbuf >> 3;
5324 tp->ooo_last_skb = rb_to_skb(prev);
5326 /* Reset SACK state. A conforming SACK implementation will
5327 * do the same at a timeout based retransmit. When a connection
5328 * is in a sad state like this, we care only about integrity
5329 * of the connection not performance.
5331 if (tp->rx_opt.sack_ok)
5332 tcp_sack_reset(&tp->rx_opt);
5336 /* Reduce allocated memory if we can, trying to get
5337 * the socket within its memory limits again.
5339 * Return less than zero if we should start dropping frames
5340 * until the socket owning process reads some of the data
5341 * to stabilize the situation.
5343 static int tcp_prune_queue(struct sock *sk)
5345 struct tcp_sock *tp = tcp_sk(sk);
5347 NET_INC_STATS(sock_net(sk), LINUX_MIB_PRUNECALLED);
5349 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
5350 tcp_clamp_window(sk);
5351 else if (tcp_under_memory_pressure(sk))
5352 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
5354 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5357 tcp_collapse_ofo_queue(sk);
5358 if (!skb_queue_empty(&sk->sk_receive_queue))
5359 tcp_collapse(sk, &sk->sk_receive_queue, NULL,
5360 skb_peek(&sk->sk_receive_queue),
5362 tp->copied_seq, tp->rcv_nxt);
5365 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5368 /* Collapsing did not help, destructive actions follow.
5369 * This must not ever occur. */
5371 tcp_prune_ofo_queue(sk);
5373 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5376 /* If we are really being abused, tell the caller to silently
5377 * drop receive data on the floor. It will get retransmitted
5378 * and hopefully then we'll have sufficient space.
5380 NET_INC_STATS(sock_net(sk), LINUX_MIB_RCVPRUNED);
5382 /* Massive buffer overcommit. */
5387 static bool tcp_should_expand_sndbuf(const struct sock *sk)
5389 const struct tcp_sock *tp = tcp_sk(sk);
5391 /* If the user specified a specific send buffer setting, do
5394 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
5397 /* If we are under global TCP memory pressure, do not expand. */
5398 if (tcp_under_memory_pressure(sk))
5401 /* If we are under soft global TCP memory pressure, do not expand. */
5402 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
5405 /* If we filled the congestion window, do not expand. */
5406 if (tcp_packets_in_flight(tp) >= tcp_snd_cwnd(tp))
5412 static void tcp_new_space(struct sock *sk)
5414 struct tcp_sock *tp = tcp_sk(sk);
5416 if (tcp_should_expand_sndbuf(sk)) {
5417 tcp_sndbuf_expand(sk);
5418 tp->snd_cwnd_stamp = tcp_jiffies32;
5421 INDIRECT_CALL_1(sk->sk_write_space, sk_stream_write_space, sk);
5424 /* Caller made space either from:
5425 * 1) Freeing skbs in rtx queues (after tp->snd_una has advanced)
5426 * 2) Sent skbs from output queue (and thus advancing tp->snd_nxt)
5428 * We might be able to generate EPOLLOUT to the application if:
5429 * 1) Space consumed in output/rtx queues is below sk->sk_sndbuf/2
5430 * 2) notsent amount (tp->write_seq - tp->snd_nxt) became
5431 * small enough that tcp_stream_memory_free() decides it
5432 * is time to generate EPOLLOUT.
5434 void tcp_check_space(struct sock *sk)
5436 /* pairs with tcp_poll() */
5438 if (sk->sk_socket &&
5439 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
5441 if (!test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
5442 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
5446 static inline void tcp_data_snd_check(struct sock *sk)
5448 tcp_push_pending_frames(sk);
5449 tcp_check_space(sk);
5453 * Check if sending an ack is needed.
5455 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
5457 struct tcp_sock *tp = tcp_sk(sk);
5458 unsigned long rtt, delay;
5460 /* More than one full frame received... */
5461 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
5462 /* ... and right edge of window advances far enough.
5463 * (tcp_recvmsg() will send ACK otherwise).
5464 * If application uses SO_RCVLOWAT, we want send ack now if
5465 * we have not received enough bytes to satisfy the condition.
5467 (tp->rcv_nxt - tp->copied_seq < sk->sk_rcvlowat ||
5468 __tcp_select_window(sk) >= tp->rcv_wnd)) ||
5469 /* We ACK each frame or... */
5470 tcp_in_quickack_mode(sk) ||
5471 /* Protocol state mandates a one-time immediate ACK */
5472 inet_csk(sk)->icsk_ack.pending & ICSK_ACK_NOW) {
5478 if (!ofo_possible || RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
5479 tcp_send_delayed_ack(sk);
5483 if (!tcp_is_sack(tp) ||
5484 tp->compressed_ack >= sock_net(sk)->ipv4.sysctl_tcp_comp_sack_nr)
5487 if (tp->compressed_ack_rcv_nxt != tp->rcv_nxt) {
5488 tp->compressed_ack_rcv_nxt = tp->rcv_nxt;
5489 tp->dup_ack_counter = 0;
5491 if (tp->dup_ack_counter < TCP_FASTRETRANS_THRESH) {
5492 tp->dup_ack_counter++;
5495 tp->compressed_ack++;
5496 if (hrtimer_is_queued(&tp->compressed_ack_timer))
5499 /* compress ack timer : 5 % of rtt, but no more than tcp_comp_sack_delay_ns */
5501 rtt = tp->rcv_rtt_est.rtt_us;
5502 if (tp->srtt_us && tp->srtt_us < rtt)
5505 delay = min_t(unsigned long, sock_net(sk)->ipv4.sysctl_tcp_comp_sack_delay_ns,
5506 rtt * (NSEC_PER_USEC >> 3)/20);
5508 hrtimer_start_range_ns(&tp->compressed_ack_timer, ns_to_ktime(delay),
5509 sock_net(sk)->ipv4.sysctl_tcp_comp_sack_slack_ns,
5510 HRTIMER_MODE_REL_PINNED_SOFT);
5513 static inline void tcp_ack_snd_check(struct sock *sk)
5515 if (!inet_csk_ack_scheduled(sk)) {
5516 /* We sent a data segment already. */
5519 __tcp_ack_snd_check(sk, 1);
5523 * This routine is only called when we have urgent data
5524 * signaled. Its the 'slow' part of tcp_urg. It could be
5525 * moved inline now as tcp_urg is only called from one
5526 * place. We handle URGent data wrong. We have to - as
5527 * BSD still doesn't use the correction from RFC961.
5528 * For 1003.1g we should support a new option TCP_STDURG to permit
5529 * either form (or just set the sysctl tcp_stdurg).
5532 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
5534 struct tcp_sock *tp = tcp_sk(sk);
5535 u32 ptr = ntohs(th->urg_ptr);
5537 if (ptr && !sock_net(sk)->ipv4.sysctl_tcp_stdurg)
5539 ptr += ntohl(th->seq);
5541 /* Ignore urgent data that we've already seen and read. */
5542 if (after(tp->copied_seq, ptr))
5545 /* Do not replay urg ptr.
5547 * NOTE: interesting situation not covered by specs.
5548 * Misbehaving sender may send urg ptr, pointing to segment,
5549 * which we already have in ofo queue. We are not able to fetch
5550 * such data and will stay in TCP_URG_NOTYET until will be eaten
5551 * by recvmsg(). Seems, we are not obliged to handle such wicked
5552 * situations. But it is worth to think about possibility of some
5553 * DoSes using some hypothetical application level deadlock.
5555 if (before(ptr, tp->rcv_nxt))
5558 /* Do we already have a newer (or duplicate) urgent pointer? */
5559 if (tp->urg_data && !after(ptr, tp->urg_seq))
5562 /* Tell the world about our new urgent pointer. */
5565 /* We may be adding urgent data when the last byte read was
5566 * urgent. To do this requires some care. We cannot just ignore
5567 * tp->copied_seq since we would read the last urgent byte again
5568 * as data, nor can we alter copied_seq until this data arrives
5569 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5571 * NOTE. Double Dutch. Rendering to plain English: author of comment
5572 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5573 * and expect that both A and B disappear from stream. This is _wrong_.
5574 * Though this happens in BSD with high probability, this is occasional.
5575 * Any application relying on this is buggy. Note also, that fix "works"
5576 * only in this artificial test. Insert some normal data between A and B and we will
5577 * decline of BSD again. Verdict: it is better to remove to trap
5580 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5581 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5582 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5584 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5585 __skb_unlink(skb, &sk->sk_receive_queue);
5590 tp->urg_data = TCP_URG_NOTYET;
5591 WRITE_ONCE(tp->urg_seq, ptr);
5593 /* Disable header prediction. */
5597 /* This is the 'fast' part of urgent handling. */
5598 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
5600 struct tcp_sock *tp = tcp_sk(sk);
5602 /* Check if we get a new urgent pointer - normally not. */
5604 tcp_check_urg(sk, th);
5606 /* Do we wait for any urgent data? - normally not... */
5607 if (tp->urg_data == TCP_URG_NOTYET) {
5608 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5611 /* Is the urgent pointer pointing into this packet? */
5612 if (ptr < skb->len) {
5614 if (skb_copy_bits(skb, ptr, &tmp, 1))
5616 tp->urg_data = TCP_URG_VALID | tmp;
5617 if (!sock_flag(sk, SOCK_DEAD))
5618 sk->sk_data_ready(sk);
5623 /* Accept RST for rcv_nxt - 1 after a FIN.
5624 * When tcp connections are abruptly terminated from Mac OSX (via ^C), a
5625 * FIN is sent followed by a RST packet. The RST is sent with the same
5626 * sequence number as the FIN, and thus according to RFC 5961 a challenge
5627 * ACK should be sent. However, Mac OSX rate limits replies to challenge
5628 * ACKs on the closed socket. In addition middleboxes can drop either the
5629 * challenge ACK or a subsequent RST.
5631 static bool tcp_reset_check(const struct sock *sk, const struct sk_buff *skb)
5633 struct tcp_sock *tp = tcp_sk(sk);
5635 return unlikely(TCP_SKB_CB(skb)->seq == (tp->rcv_nxt - 1) &&
5636 (1 << sk->sk_state) & (TCPF_CLOSE_WAIT | TCPF_LAST_ACK |
5640 /* Does PAWS and seqno based validation of an incoming segment, flags will
5641 * play significant role here.
5643 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5644 const struct tcphdr *th, int syn_inerr)
5646 struct tcp_sock *tp = tcp_sk(sk);
5647 bool rst_seq_match = false;
5649 /* RFC1323: H1. Apply PAWS check first. */
5650 if (tcp_fast_parse_options(sock_net(sk), skb, th, tp) &&
5651 tp->rx_opt.saw_tstamp &&
5652 tcp_paws_discard(sk, skb)) {
5654 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5655 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5656 LINUX_MIB_TCPACKSKIPPEDPAWS,
5657 &tp->last_oow_ack_time))
5658 tcp_send_dupack(sk, skb);
5661 /* Reset is accepted even if it did not pass PAWS. */
5664 /* Step 1: check sequence number */
5665 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5666 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5667 * (RST) segments are validated by checking their SEQ-fields."
5668 * And page 69: "If an incoming segment is not acceptable,
5669 * an acknowledgment should be sent in reply (unless the RST
5670 * bit is set, if so drop the segment and return)".
5675 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5676 LINUX_MIB_TCPACKSKIPPEDSEQ,
5677 &tp->last_oow_ack_time))
5678 tcp_send_dupack(sk, skb);
5679 } else if (tcp_reset_check(sk, skb)) {
5685 /* Step 2: check RST bit */
5687 /* RFC 5961 3.2 (extend to match against (RCV.NXT - 1) after a
5688 * FIN and SACK too if available):
5689 * If seq num matches RCV.NXT or (RCV.NXT - 1) after a FIN, or
5690 * the right-most SACK block,
5692 * RESET the connection
5694 * Send a challenge ACK
5696 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt ||
5697 tcp_reset_check(sk, skb)) {
5698 rst_seq_match = true;
5699 } else if (tcp_is_sack(tp) && tp->rx_opt.num_sacks > 0) {
5700 struct tcp_sack_block *sp = &tp->selective_acks[0];
5701 int max_sack = sp[0].end_seq;
5704 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;
5706 max_sack = after(sp[this_sack].end_seq,
5708 sp[this_sack].end_seq : max_sack;
5711 if (TCP_SKB_CB(skb)->seq == max_sack)
5712 rst_seq_match = true;
5718 /* Disable TFO if RST is out-of-order
5719 * and no data has been received
5720 * for current active TFO socket
5722 if (tp->syn_fastopen && !tp->data_segs_in &&
5723 sk->sk_state == TCP_ESTABLISHED)
5724 tcp_fastopen_active_disable(sk);
5725 tcp_send_challenge_ack(sk, skb);
5730 /* step 3: check security and precedence [ignored] */
5732 /* step 4: Check for a SYN
5733 * RFC 5961 4.2 : Send a challenge ack
5738 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5739 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5740 tcp_send_challenge_ack(sk, skb);
5744 bpf_skops_parse_hdr(sk, skb);
5754 * TCP receive function for the ESTABLISHED state.
5756 * It is split into a fast path and a slow path. The fast path is
5758 * - A zero window was announced from us - zero window probing
5759 * is only handled properly in the slow path.
5760 * - Out of order segments arrived.
5761 * - Urgent data is expected.
5762 * - There is no buffer space left
5763 * - Unexpected TCP flags/window values/header lengths are received
5764 * (detected by checking the TCP header against pred_flags)
5765 * - Data is sent in both directions. Fast path only supports pure senders
5766 * or pure receivers (this means either the sequence number or the ack
5767 * value must stay constant)
5768 * - Unexpected TCP option.
5770 * When these conditions are not satisfied it drops into a standard
5771 * receive procedure patterned after RFC793 to handle all cases.
5772 * The first three cases are guaranteed by proper pred_flags setting,
5773 * the rest is checked inline. Fast processing is turned on in
5774 * tcp_data_queue when everything is OK.
5776 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb)
5778 const struct tcphdr *th = (const struct tcphdr *)skb->data;
5779 struct tcp_sock *tp = tcp_sk(sk);
5780 unsigned int len = skb->len;
5782 /* TCP congestion window tracking */
5783 trace_tcp_probe(sk, skb);
5785 tcp_mstamp_refresh(tp);
5786 if (unlikely(!rcu_access_pointer(sk->sk_rx_dst)))
5787 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
5789 * Header prediction.
5790 * The code loosely follows the one in the famous
5791 * "30 instruction TCP receive" Van Jacobson mail.
5793 * Van's trick is to deposit buffers into socket queue
5794 * on a device interrupt, to call tcp_recv function
5795 * on the receive process context and checksum and copy
5796 * the buffer to user space. smart...
5798 * Our current scheme is not silly either but we take the
5799 * extra cost of the net_bh soft interrupt processing...
5800 * We do checksum and copy also but from device to kernel.
5803 tp->rx_opt.saw_tstamp = 0;
5805 /* pred_flags is 0xS?10 << 16 + snd_wnd
5806 * if header_prediction is to be made
5807 * 'S' will always be tp->tcp_header_len >> 2
5808 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5809 * turn it off (when there are holes in the receive
5810 * space for instance)
5811 * PSH flag is ignored.
5814 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5815 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5816 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5817 int tcp_header_len = tp->tcp_header_len;
5819 /* Timestamp header prediction: tcp_header_len
5820 * is automatically equal to th->doff*4 due to pred_flags
5824 /* Check timestamp */
5825 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5826 /* No? Slow path! */
5827 if (!tcp_parse_aligned_timestamp(tp, th))
5830 /* If PAWS failed, check it more carefully in slow path */
5831 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5834 /* DO NOT update ts_recent here, if checksum fails
5835 * and timestamp was corrupted part, it will result
5836 * in a hung connection since we will drop all
5837 * future packets due to the PAWS test.
5841 if (len <= tcp_header_len) {
5842 /* Bulk data transfer: sender */
5843 if (len == tcp_header_len) {
5844 /* Predicted packet is in window by definition.
5845 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5846 * Hence, check seq<=rcv_wup reduces to:
5848 if (tcp_header_len ==
5849 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5850 tp->rcv_nxt == tp->rcv_wup)
5851 tcp_store_ts_recent(tp);
5853 /* We know that such packets are checksummed
5856 tcp_ack(sk, skb, 0);
5858 tcp_data_snd_check(sk);
5859 /* When receiving pure ack in fast path, update
5860 * last ts ecr directly instead of calling
5861 * tcp_rcv_rtt_measure_ts()
5863 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr;
5865 } else { /* Header too small */
5866 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5871 bool fragstolen = false;
5873 if (tcp_checksum_complete(skb))
5876 if ((int)skb->truesize > sk->sk_forward_alloc)
5879 /* Predicted packet is in window by definition.
5880 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5881 * Hence, check seq<=rcv_wup reduces to:
5883 if (tcp_header_len ==
5884 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5885 tp->rcv_nxt == tp->rcv_wup)
5886 tcp_store_ts_recent(tp);
5888 tcp_rcv_rtt_measure_ts(sk, skb);
5890 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPHITS);
5892 /* Bulk data transfer: receiver */
5893 __skb_pull(skb, tcp_header_len);
5894 eaten = tcp_queue_rcv(sk, skb, &fragstolen);
5896 tcp_event_data_recv(sk, skb);
5898 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5899 /* Well, only one small jumplet in fast path... */
5900 tcp_ack(sk, skb, FLAG_DATA);
5901 tcp_data_snd_check(sk);
5902 if (!inet_csk_ack_scheduled(sk))
5905 tcp_update_wl(tp, TCP_SKB_CB(skb)->seq);
5908 __tcp_ack_snd_check(sk, 0);
5911 kfree_skb_partial(skb, fragstolen);
5918 if (len < (th->doff << 2) || tcp_checksum_complete(skb))
5921 if (!th->ack && !th->rst && !th->syn)
5925 * Standard slow path.
5928 if (!tcp_validate_incoming(sk, skb, th, 1))
5932 if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0)
5935 tcp_rcv_rtt_measure_ts(sk, skb);
5937 /* Process urgent data. */
5938 tcp_urg(sk, skb, th);
5940 /* step 7: process the segment text */
5941 tcp_data_queue(sk, skb);
5943 tcp_data_snd_check(sk);
5944 tcp_ack_snd_check(sk);
5948 trace_tcp_bad_csum(skb);
5949 TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS);
5950 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5955 EXPORT_SYMBOL(tcp_rcv_established);
5957 void tcp_init_transfer(struct sock *sk, int bpf_op, struct sk_buff *skb)
5959 struct inet_connection_sock *icsk = inet_csk(sk);
5960 struct tcp_sock *tp = tcp_sk(sk);
5963 icsk->icsk_af_ops->rebuild_header(sk);
5964 tcp_init_metrics(sk);
5966 /* Initialize the congestion window to start the transfer.
5967 * Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
5968 * retransmitted. In light of RFC6298 more aggressive 1sec
5969 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
5970 * retransmission has occurred.
5972 if (tp->total_retrans > 1 && tp->undo_marker)
5973 tcp_snd_cwnd_set(tp, 1);
5975 tcp_snd_cwnd_set(tp, tcp_init_cwnd(tp, __sk_dst_get(sk)));
5976 tp->snd_cwnd_stamp = tcp_jiffies32;
5978 bpf_skops_established(sk, bpf_op, skb);
5979 /* Initialize congestion control unless BPF initialized it already: */
5980 if (!icsk->icsk_ca_initialized)
5981 tcp_init_congestion_control(sk);
5982 tcp_init_buffer_space(sk);
5985 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5987 struct tcp_sock *tp = tcp_sk(sk);
5988 struct inet_connection_sock *icsk = inet_csk(sk);
5990 tcp_set_state(sk, TCP_ESTABLISHED);
5991 icsk->icsk_ack.lrcvtime = tcp_jiffies32;
5994 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
5995 security_inet_conn_established(sk, skb);
5996 sk_mark_napi_id(sk, skb);
5999 tcp_init_transfer(sk, BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB, skb);
6001 /* Prevent spurious tcp_cwnd_restart() on first data
6004 tp->lsndtime = tcp_jiffies32;
6006 if (sock_flag(sk, SOCK_KEEPOPEN))
6007 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
6009 if (!tp->rx_opt.snd_wscale)
6010 __tcp_fast_path_on(tp, tp->snd_wnd);
6015 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
6016 struct tcp_fastopen_cookie *cookie)
6018 struct tcp_sock *tp = tcp_sk(sk);
6019 struct sk_buff *data = tp->syn_data ? tcp_rtx_queue_head(sk) : NULL;
6020 u16 mss = tp->rx_opt.mss_clamp, try_exp = 0;
6021 bool syn_drop = false;
6023 if (mss == tp->rx_opt.user_mss) {
6024 struct tcp_options_received opt;
6026 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
6027 tcp_clear_options(&opt);
6028 opt.user_mss = opt.mss_clamp = 0;
6029 tcp_parse_options(sock_net(sk), synack, &opt, 0, NULL);
6030 mss = opt.mss_clamp;
6033 if (!tp->syn_fastopen) {
6034 /* Ignore an unsolicited cookie */
6036 } else if (tp->total_retrans) {
6037 /* SYN timed out and the SYN-ACK neither has a cookie nor
6038 * acknowledges data. Presumably the remote received only
6039 * the retransmitted (regular) SYNs: either the original
6040 * SYN-data or the corresponding SYN-ACK was dropped.
6042 syn_drop = (cookie->len < 0 && data);
6043 } else if (cookie->len < 0 && !tp->syn_data) {
6044 /* We requested a cookie but didn't get it. If we did not use
6045 * the (old) exp opt format then try so next time (try_exp=1).
6046 * Otherwise we go back to use the RFC7413 opt (try_exp=2).
6048 try_exp = tp->syn_fastopen_exp ? 2 : 1;
6051 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp);
6053 if (data) { /* Retransmit unacked data in SYN */
6054 if (tp->total_retrans)
6055 tp->fastopen_client_fail = TFO_SYN_RETRANSMITTED;
6057 tp->fastopen_client_fail = TFO_DATA_NOT_ACKED;
6058 skb_rbtree_walk_from(data)
6059 tcp_mark_skb_lost(sk, data);
6060 tcp_xmit_retransmit_queue(sk);
6061 NET_INC_STATS(sock_net(sk),
6062 LINUX_MIB_TCPFASTOPENACTIVEFAIL);
6065 tp->syn_data_acked = tp->syn_data;
6066 if (tp->syn_data_acked) {
6067 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
6068 /* SYN-data is counted as two separate packets in tcp_ack() */
6069 if (tp->delivered > 1)
6073 tcp_fastopen_add_skb(sk, synack);
6078 static void smc_check_reset_syn(struct tcp_sock *tp)
6080 #if IS_ENABLED(CONFIG_SMC)
6081 if (static_branch_unlikely(&tcp_have_smc)) {
6082 if (tp->syn_smc && !tp->rx_opt.smc_ok)
6088 static void tcp_try_undo_spurious_syn(struct sock *sk)
6090 struct tcp_sock *tp = tcp_sk(sk);
6093 /* undo_marker is set when SYN or SYNACK times out. The timeout is
6094 * spurious if the ACK's timestamp option echo value matches the
6095 * original SYN timestamp.
6097 syn_stamp = tp->retrans_stamp;
6098 if (tp->undo_marker && syn_stamp && tp->rx_opt.saw_tstamp &&
6099 syn_stamp == tp->rx_opt.rcv_tsecr)
6100 tp->undo_marker = 0;
6103 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
6104 const struct tcphdr *th)
6106 struct inet_connection_sock *icsk = inet_csk(sk);
6107 struct tcp_sock *tp = tcp_sk(sk);
6108 struct tcp_fastopen_cookie foc = { .len = -1 };
6109 int saved_clamp = tp->rx_opt.mss_clamp;
6112 tcp_parse_options(sock_net(sk), skb, &tp->rx_opt, 0, &foc);
6113 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
6114 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
6118 * "If the state is SYN-SENT then
6119 * first check the ACK bit
6120 * If the ACK bit is set
6121 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
6122 * a reset (unless the RST bit is set, if so drop
6123 * the segment and return)"
6125 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
6126 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
6127 /* Previous FIN/ACK or RST/ACK might be ignored. */
6128 if (icsk->icsk_retransmits == 0)
6129 inet_csk_reset_xmit_timer(sk,
6131 TCP_TIMEOUT_MIN, TCP_RTO_MAX);
6132 goto reset_and_undo;
6135 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
6136 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
6137 tcp_time_stamp(tp))) {
6138 NET_INC_STATS(sock_net(sk),
6139 LINUX_MIB_PAWSACTIVEREJECTED);
6140 goto reset_and_undo;
6143 /* Now ACK is acceptable.
6145 * "If the RST bit is set
6146 * If the ACK was acceptable then signal the user "error:
6147 * connection reset", drop the segment, enter CLOSED state,
6148 * delete TCB, and return."
6157 * "fifth, if neither of the SYN or RST bits is set then
6158 * drop the segment and return."
6164 goto discard_and_undo;
6167 * "If the SYN bit is on ...
6168 * are acceptable then ...
6169 * (our SYN has been ACKed), change the connection
6170 * state to ESTABLISHED..."
6173 tcp_ecn_rcv_synack(tp, th);
6175 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
6176 tcp_try_undo_spurious_syn(sk);
6177 tcp_ack(sk, skb, FLAG_SLOWPATH);
6179 /* Ok.. it's good. Set up sequence numbers and
6180 * move to established.
6182 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1);
6183 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
6185 /* RFC1323: The window in SYN & SYN/ACK segments is
6188 tp->snd_wnd = ntohs(th->window);
6190 if (!tp->rx_opt.wscale_ok) {
6191 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
6192 tp->window_clamp = min(tp->window_clamp, 65535U);
6195 if (tp->rx_opt.saw_tstamp) {
6196 tp->rx_opt.tstamp_ok = 1;
6197 tp->tcp_header_len =
6198 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
6199 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
6200 tcp_store_ts_recent(tp);
6202 tp->tcp_header_len = sizeof(struct tcphdr);
6205 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
6206 tcp_initialize_rcv_mss(sk);
6208 /* Remember, tcp_poll() does not lock socket!
6209 * Change state from SYN-SENT only after copied_seq
6210 * is initialized. */
6211 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6213 smc_check_reset_syn(tp);
6217 tcp_finish_connect(sk, skb);
6219 fastopen_fail = (tp->syn_fastopen || tp->syn_data) &&
6220 tcp_rcv_fastopen_synack(sk, skb, &foc);
6222 if (!sock_flag(sk, SOCK_DEAD)) {
6223 sk->sk_state_change(sk);
6224 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
6228 if (sk->sk_write_pending ||
6229 icsk->icsk_accept_queue.rskq_defer_accept ||
6230 inet_csk_in_pingpong_mode(sk)) {
6231 /* Save one ACK. Data will be ready after
6232 * several ticks, if write_pending is set.
6234 * It may be deleted, but with this feature tcpdumps
6235 * look so _wonderfully_ clever, that I was not able
6236 * to stand against the temptation 8) --ANK
6238 inet_csk_schedule_ack(sk);
6239 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
6240 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
6241 TCP_DELACK_MAX, TCP_RTO_MAX);
6252 /* No ACK in the segment */
6256 * "If the RST bit is set
6258 * Otherwise (no ACK) drop the segment and return."
6261 goto discard_and_undo;
6265 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
6266 tcp_paws_reject(&tp->rx_opt, 0))
6267 goto discard_and_undo;
6270 /* We see SYN without ACK. It is attempt of
6271 * simultaneous connect with crossed SYNs.
6272 * Particularly, it can be connect to self.
6274 tcp_set_state(sk, TCP_SYN_RECV);
6276 if (tp->rx_opt.saw_tstamp) {
6277 tp->rx_opt.tstamp_ok = 1;
6278 tcp_store_ts_recent(tp);
6279 tp->tcp_header_len =
6280 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
6282 tp->tcp_header_len = sizeof(struct tcphdr);
6285 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1);
6286 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6287 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
6289 /* RFC1323: The window in SYN & SYN/ACK segments is
6292 tp->snd_wnd = ntohs(th->window);
6293 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
6294 tp->max_window = tp->snd_wnd;
6296 tcp_ecn_rcv_syn(tp, th);
6299 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
6300 tcp_initialize_rcv_mss(sk);
6302 tcp_send_synack(sk);
6304 /* Note, we could accept data and URG from this segment.
6305 * There are no obstacles to make this (except that we must
6306 * either change tcp_recvmsg() to prevent it from returning data
6307 * before 3WHS completes per RFC793, or employ TCP Fast Open).
6309 * However, if we ignore data in ACKless segments sometimes,
6310 * we have no reasons to accept it sometimes.
6311 * Also, seems the code doing it in step6 of tcp_rcv_state_process
6312 * is not flawless. So, discard packet for sanity.
6313 * Uncomment this return to process the data.
6320 /* "fifth, if neither of the SYN or RST bits is set then
6321 * drop the segment and return."
6325 tcp_clear_options(&tp->rx_opt);
6326 tp->rx_opt.mss_clamp = saved_clamp;
6330 tcp_clear_options(&tp->rx_opt);
6331 tp->rx_opt.mss_clamp = saved_clamp;
6335 static void tcp_rcv_synrecv_state_fastopen(struct sock *sk)
6337 struct request_sock *req;
6339 /* If we are still handling the SYNACK RTO, see if timestamp ECR allows
6340 * undo. If peer SACKs triggered fast recovery, we can't undo here.
6342 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
6343 tcp_try_undo_loss(sk, false);
6345 /* Reset rtx states to prevent spurious retransmits_timed_out() */
6346 tcp_sk(sk)->retrans_stamp = 0;
6347 inet_csk(sk)->icsk_retransmits = 0;
6349 /* Once we leave TCP_SYN_RECV or TCP_FIN_WAIT_1,
6350 * we no longer need req so release it.
6352 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk,
6353 lockdep_sock_is_held(sk));
6354 reqsk_fastopen_remove(sk, req, false);
6356 /* Re-arm the timer because data may have been sent out.
6357 * This is similar to the regular data transmission case
6358 * when new data has just been ack'ed.
6360 * (TFO) - we could try to be more aggressive and
6361 * retransmitting any data sooner based on when they
6368 * This function implements the receiving procedure of RFC 793 for
6369 * all states except ESTABLISHED and TIME_WAIT.
6370 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
6371 * address independent.
6374 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb)
6376 struct tcp_sock *tp = tcp_sk(sk);
6377 struct inet_connection_sock *icsk = inet_csk(sk);
6378 const struct tcphdr *th = tcp_hdr(skb);
6379 struct request_sock *req;
6383 switch (sk->sk_state) {
6397 /* It is possible that we process SYN packets from backlog,
6398 * so we need to make sure to disable BH and RCU right there.
6402 acceptable = icsk->icsk_af_ops->conn_request(sk, skb) >= 0;
6414 tp->rx_opt.saw_tstamp = 0;
6415 tcp_mstamp_refresh(tp);
6416 queued = tcp_rcv_synsent_state_process(sk, skb, th);
6420 /* Do step6 onward by hand. */
6421 tcp_urg(sk, skb, th);
6423 tcp_data_snd_check(sk);
6427 tcp_mstamp_refresh(tp);
6428 tp->rx_opt.saw_tstamp = 0;
6429 req = rcu_dereference_protected(tp->fastopen_rsk,
6430 lockdep_sock_is_held(sk));
6434 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
6435 sk->sk_state != TCP_FIN_WAIT1);
6437 if (!tcp_check_req(sk, skb, req, true, &req_stolen))
6441 if (!th->ack && !th->rst && !th->syn)
6444 if (!tcp_validate_incoming(sk, skb, th, 0))
6447 /* step 5: check the ACK field */
6448 acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
6449 FLAG_UPDATE_TS_RECENT |
6450 FLAG_NO_CHALLENGE_ACK) > 0;
6453 if (sk->sk_state == TCP_SYN_RECV)
6454 return 1; /* send one RST */
6455 tcp_send_challenge_ack(sk, skb);
6458 switch (sk->sk_state) {
6460 tp->delivered++; /* SYN-ACK delivery isn't tracked in tcp_ack */
6462 tcp_synack_rtt_meas(sk, req);
6465 tcp_rcv_synrecv_state_fastopen(sk);
6467 tcp_try_undo_spurious_syn(sk);
6468 tp->retrans_stamp = 0;
6469 tcp_init_transfer(sk, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB,
6471 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6474 tcp_set_state(sk, TCP_ESTABLISHED);
6475 sk->sk_state_change(sk);
6477 /* Note, that this wakeup is only for marginal crossed SYN case.
6478 * Passively open sockets are not waked up, because
6479 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
6482 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
6484 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
6485 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
6486 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
6488 if (tp->rx_opt.tstamp_ok)
6489 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
6491 if (!inet_csk(sk)->icsk_ca_ops->cong_control)
6492 tcp_update_pacing_rate(sk);
6494 /* Prevent spurious tcp_cwnd_restart() on first data packet */
6495 tp->lsndtime = tcp_jiffies32;
6497 tcp_initialize_rcv_mss(sk);
6498 tcp_fast_path_on(tp);
6501 case TCP_FIN_WAIT1: {
6505 tcp_rcv_synrecv_state_fastopen(sk);
6507 if (tp->snd_una != tp->write_seq)
6510 tcp_set_state(sk, TCP_FIN_WAIT2);
6511 sk->sk_shutdown |= SEND_SHUTDOWN;
6515 if (!sock_flag(sk, SOCK_DEAD)) {
6516 /* Wake up lingering close() */
6517 sk->sk_state_change(sk);
6521 if (tp->linger2 < 0) {
6523 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6526 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6527 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6528 /* Receive out of order FIN after close() */
6529 if (tp->syn_fastopen && th->fin)
6530 tcp_fastopen_active_disable(sk);
6532 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6536 tmo = tcp_fin_time(sk);
6537 if (tmo > TCP_TIMEWAIT_LEN) {
6538 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
6539 } else if (th->fin || sock_owned_by_user(sk)) {
6540 /* Bad case. We could lose such FIN otherwise.
6541 * It is not a big problem, but it looks confusing
6542 * and not so rare event. We still can lose it now,
6543 * if it spins in bh_lock_sock(), but it is really
6546 inet_csk_reset_keepalive_timer(sk, tmo);
6548 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
6555 if (tp->snd_una == tp->write_seq) {
6556 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
6562 if (tp->snd_una == tp->write_seq) {
6563 tcp_update_metrics(sk);
6570 /* step 6: check the URG bit */
6571 tcp_urg(sk, skb, th);
6573 /* step 7: process the segment text */
6574 switch (sk->sk_state) {
6575 case TCP_CLOSE_WAIT:
6578 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
6579 /* If a subflow has been reset, the packet should not
6580 * continue to be processed, drop the packet.
6582 if (sk_is_mptcp(sk) && !mptcp_incoming_options(sk, skb))
6589 /* RFC 793 says to queue data in these states,
6590 * RFC 1122 says we MUST send a reset.
6591 * BSD 4.4 also does reset.
6593 if (sk->sk_shutdown & RCV_SHUTDOWN) {
6594 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6595 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6596 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6602 case TCP_ESTABLISHED:
6603 tcp_data_queue(sk, skb);
6608 /* tcp_data could move socket to TIME-WAIT */
6609 if (sk->sk_state != TCP_CLOSE) {
6610 tcp_data_snd_check(sk);
6611 tcp_ack_snd_check(sk);
6620 EXPORT_SYMBOL(tcp_rcv_state_process);
6622 static inline void pr_drop_req(struct request_sock *req, __u16 port, int family)
6624 struct inet_request_sock *ireq = inet_rsk(req);
6626 if (family == AF_INET)
6627 net_dbg_ratelimited("drop open request from %pI4/%u\n",
6628 &ireq->ir_rmt_addr, port);
6629 #if IS_ENABLED(CONFIG_IPV6)
6630 else if (family == AF_INET6)
6631 net_dbg_ratelimited("drop open request from %pI6/%u\n",
6632 &ireq->ir_v6_rmt_addr, port);
6636 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
6638 * If we receive a SYN packet with these bits set, it means a
6639 * network is playing bad games with TOS bits. In order to
6640 * avoid possible false congestion notifications, we disable
6641 * TCP ECN negotiation.
6643 * Exception: tcp_ca wants ECN. This is required for DCTCP
6644 * congestion control: Linux DCTCP asserts ECT on all packets,
6645 * including SYN, which is most optimal solution; however,
6646 * others, such as FreeBSD do not.
6648 * Exception: At least one of the reserved bits of the TCP header (th->res1) is
6649 * set, indicating the use of a future TCP extension (such as AccECN). See
6650 * RFC8311 §4.3 which updates RFC3168 to allow the development of such
6653 static void tcp_ecn_create_request(struct request_sock *req,
6654 const struct sk_buff *skb,
6655 const struct sock *listen_sk,
6656 const struct dst_entry *dst)
6658 const struct tcphdr *th = tcp_hdr(skb);
6659 const struct net *net = sock_net(listen_sk);
6660 bool th_ecn = th->ece && th->cwr;
6667 ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield);
6668 ecn_ok_dst = dst_feature(dst, DST_FEATURE_ECN_MASK);
6669 ecn_ok = net->ipv4.sysctl_tcp_ecn || ecn_ok_dst;
6671 if (((!ect || th->res1) && ecn_ok) || tcp_ca_needs_ecn(listen_sk) ||
6672 (ecn_ok_dst & DST_FEATURE_ECN_CA) ||
6673 tcp_bpf_ca_needs_ecn((struct sock *)req))
6674 inet_rsk(req)->ecn_ok = 1;
6677 static void tcp_openreq_init(struct request_sock *req,
6678 const struct tcp_options_received *rx_opt,
6679 struct sk_buff *skb, const struct sock *sk)
6681 struct inet_request_sock *ireq = inet_rsk(req);
6683 req->rsk_rcv_wnd = 0; /* So that tcp_send_synack() knows! */
6684 tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq;
6685 tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
6686 tcp_rsk(req)->snt_synack = 0;
6687 tcp_rsk(req)->last_oow_ack_time = 0;
6688 req->mss = rx_opt->mss_clamp;
6689 req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0;
6690 ireq->tstamp_ok = rx_opt->tstamp_ok;
6691 ireq->sack_ok = rx_opt->sack_ok;
6692 ireq->snd_wscale = rx_opt->snd_wscale;
6693 ireq->wscale_ok = rx_opt->wscale_ok;
6696 ireq->ir_rmt_port = tcp_hdr(skb)->source;
6697 ireq->ir_num = ntohs(tcp_hdr(skb)->dest);
6698 ireq->ir_mark = inet_request_mark(sk, skb);
6699 #if IS_ENABLED(CONFIG_SMC)
6700 ireq->smc_ok = rx_opt->smc_ok;
6704 struct request_sock *inet_reqsk_alloc(const struct request_sock_ops *ops,
6705 struct sock *sk_listener,
6706 bool attach_listener)
6708 struct request_sock *req = reqsk_alloc(ops, sk_listener,
6712 struct inet_request_sock *ireq = inet_rsk(req);
6714 ireq->ireq_opt = NULL;
6715 #if IS_ENABLED(CONFIG_IPV6)
6716 ireq->pktopts = NULL;
6718 atomic64_set(&ireq->ir_cookie, 0);
6719 ireq->ireq_state = TCP_NEW_SYN_RECV;
6720 write_pnet(&ireq->ireq_net, sock_net(sk_listener));
6721 ireq->ireq_family = sk_listener->sk_family;
6726 EXPORT_SYMBOL(inet_reqsk_alloc);
6729 * Return true if a syncookie should be sent
6731 static bool tcp_syn_flood_action(const struct sock *sk, const char *proto)
6733 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
6734 const char *msg = "Dropping request";
6735 bool want_cookie = false;
6736 struct net *net = sock_net(sk);
6738 #ifdef CONFIG_SYN_COOKIES
6739 if (net->ipv4.sysctl_tcp_syncookies) {
6740 msg = "Sending cookies";
6742 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES);
6745 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP);
6747 if (!queue->synflood_warned &&
6748 net->ipv4.sysctl_tcp_syncookies != 2 &&
6749 xchg(&queue->synflood_warned, 1) == 0)
6750 net_info_ratelimited("%s: Possible SYN flooding on port %d. %s. Check SNMP counters.\n",
6751 proto, sk->sk_num, msg);
6756 static void tcp_reqsk_record_syn(const struct sock *sk,
6757 struct request_sock *req,
6758 const struct sk_buff *skb)
6760 if (tcp_sk(sk)->save_syn) {
6761 u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb);
6762 struct saved_syn *saved_syn;
6766 if (tcp_sk(sk)->save_syn == 2) { /* Save full header. */
6767 base = skb_mac_header(skb);
6768 mac_hdrlen = skb_mac_header_len(skb);
6771 base = skb_network_header(skb);
6775 saved_syn = kmalloc(struct_size(saved_syn, data, len),
6778 saved_syn->mac_hdrlen = mac_hdrlen;
6779 saved_syn->network_hdrlen = skb_network_header_len(skb);
6780 saved_syn->tcp_hdrlen = tcp_hdrlen(skb);
6781 memcpy(saved_syn->data, base, len);
6782 req->saved_syn = saved_syn;
6787 /* If a SYN cookie is required and supported, returns a clamped MSS value to be
6788 * used for SYN cookie generation.
6790 u16 tcp_get_syncookie_mss(struct request_sock_ops *rsk_ops,
6791 const struct tcp_request_sock_ops *af_ops,
6792 struct sock *sk, struct tcphdr *th)
6794 struct tcp_sock *tp = tcp_sk(sk);
6797 if (sock_net(sk)->ipv4.sysctl_tcp_syncookies != 2 &&
6798 !inet_csk_reqsk_queue_is_full(sk))
6801 if (!tcp_syn_flood_action(sk, rsk_ops->slab_name))
6804 if (sk_acceptq_is_full(sk)) {
6805 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
6809 mss = tcp_parse_mss_option(th, tp->rx_opt.user_mss);
6811 mss = af_ops->mss_clamp;
6815 EXPORT_SYMBOL_GPL(tcp_get_syncookie_mss);
6817 int tcp_conn_request(struct request_sock_ops *rsk_ops,
6818 const struct tcp_request_sock_ops *af_ops,
6819 struct sock *sk, struct sk_buff *skb)
6821 struct tcp_fastopen_cookie foc = { .len = -1 };
6822 __u32 isn = TCP_SKB_CB(skb)->tcp_tw_isn;
6823 struct tcp_options_received tmp_opt;
6824 struct tcp_sock *tp = tcp_sk(sk);
6825 struct net *net = sock_net(sk);
6826 struct sock *fastopen_sk = NULL;
6827 struct request_sock *req;
6828 bool want_cookie = false;
6829 struct dst_entry *dst;
6832 /* TW buckets are converted to open requests without
6833 * limitations, they conserve resources and peer is
6834 * evidently real one.
6836 if ((net->ipv4.sysctl_tcp_syncookies == 2 ||
6837 inet_csk_reqsk_queue_is_full(sk)) && !isn) {
6838 want_cookie = tcp_syn_flood_action(sk, rsk_ops->slab_name);
6843 if (sk_acceptq_is_full(sk)) {
6844 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
6848 req = inet_reqsk_alloc(rsk_ops, sk, !want_cookie);
6852 req->syncookie = want_cookie;
6853 tcp_rsk(req)->af_specific = af_ops;
6854 tcp_rsk(req)->ts_off = 0;
6855 #if IS_ENABLED(CONFIG_MPTCP)
6856 tcp_rsk(req)->is_mptcp = 0;
6859 tcp_clear_options(&tmp_opt);
6860 tmp_opt.mss_clamp = af_ops->mss_clamp;
6861 tmp_opt.user_mss = tp->rx_opt.user_mss;
6862 tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0,
6863 want_cookie ? NULL : &foc);
6865 if (want_cookie && !tmp_opt.saw_tstamp)
6866 tcp_clear_options(&tmp_opt);
6868 if (IS_ENABLED(CONFIG_SMC) && want_cookie)
6871 tmp_opt.tstamp_ok = tmp_opt.saw_tstamp;
6872 tcp_openreq_init(req, &tmp_opt, skb, sk);
6873 inet_rsk(req)->no_srccheck = inet_sk(sk)->transparent;
6875 /* Note: tcp_v6_init_req() might override ir_iif for link locals */
6876 inet_rsk(req)->ir_iif = inet_request_bound_dev_if(sk, skb);
6878 dst = af_ops->route_req(sk, skb, &fl, req);
6882 if (tmp_opt.tstamp_ok)
6883 tcp_rsk(req)->ts_off = af_ops->init_ts_off(net, skb);
6885 if (!want_cookie && !isn) {
6886 /* Kill the following clause, if you dislike this way. */
6887 if (!net->ipv4.sysctl_tcp_syncookies &&
6888 (net->ipv4.sysctl_max_syn_backlog - inet_csk_reqsk_queue_len(sk) <
6889 (net->ipv4.sysctl_max_syn_backlog >> 2)) &&
6890 !tcp_peer_is_proven(req, dst)) {
6891 /* Without syncookies last quarter of
6892 * backlog is filled with destinations,
6893 * proven to be alive.
6894 * It means that we continue to communicate
6895 * to destinations, already remembered
6896 * to the moment of synflood.
6898 pr_drop_req(req, ntohs(tcp_hdr(skb)->source),
6900 goto drop_and_release;
6903 isn = af_ops->init_seq(skb);
6906 tcp_ecn_create_request(req, skb, sk, dst);
6909 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss);
6910 if (!tmp_opt.tstamp_ok)
6911 inet_rsk(req)->ecn_ok = 0;
6914 tcp_rsk(req)->snt_isn = isn;
6915 tcp_rsk(req)->txhash = net_tx_rndhash();
6916 tcp_rsk(req)->syn_tos = TCP_SKB_CB(skb)->ip_dsfield;
6917 tcp_openreq_init_rwin(req, sk, dst);
6918 sk_rx_queue_set(req_to_sk(req), skb);
6920 tcp_reqsk_record_syn(sk, req, skb);
6921 fastopen_sk = tcp_try_fastopen(sk, skb, req, &foc, dst);
6924 af_ops->send_synack(fastopen_sk, dst, &fl, req,
6925 &foc, TCP_SYNACK_FASTOPEN, skb);
6926 /* Add the child socket directly into the accept queue */
6927 if (!inet_csk_reqsk_queue_add(sk, req, fastopen_sk)) {
6928 reqsk_fastopen_remove(fastopen_sk, req, false);
6929 bh_unlock_sock(fastopen_sk);
6930 sock_put(fastopen_sk);
6933 sk->sk_data_ready(sk);
6934 bh_unlock_sock(fastopen_sk);
6935 sock_put(fastopen_sk);
6937 tcp_rsk(req)->tfo_listener = false;
6939 inet_csk_reqsk_queue_hash_add(sk, req,
6940 tcp_timeout_init((struct sock *)req));
6941 af_ops->send_synack(sk, dst, &fl, req, &foc,
6942 !want_cookie ? TCP_SYNACK_NORMAL :
6961 EXPORT_SYMBOL(tcp_conn_request);