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 /* Note: divides are still a bit expensive.
241 * For the moment, only adjust scaling_ratio
242 * when we update icsk_ack.rcv_mss.
244 if (unlikely(len != icsk->icsk_ack.rcv_mss)) {
245 u64 val = (u64)skb->len << TCP_RMEM_TO_WIN_SCALE;
247 do_div(val, skb->truesize);
248 tcp_sk(sk)->scaling_ratio = val ? val : 1;
250 icsk->icsk_ack.rcv_mss = min_t(unsigned int, len,
252 /* Account for possibly-removed options */
253 if (unlikely(len > icsk->icsk_ack.rcv_mss +
254 MAX_TCP_OPTION_SPACE))
255 tcp_gro_dev_warn(sk, skb, len);
257 /* Otherwise, we make more careful check taking into account,
258 * that SACKs block is variable.
260 * "len" is invariant segment length, including TCP header.
262 len += skb->data - skb_transport_header(skb);
263 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
264 /* If PSH is not set, packet should be
265 * full sized, provided peer TCP is not badly broken.
266 * This observation (if it is correct 8)) allows
267 * to handle super-low mtu links fairly.
269 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
270 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
271 /* Subtract also invariant (if peer is RFC compliant),
272 * tcp header plus fixed timestamp option length.
273 * Resulting "len" is MSS free of SACK jitter.
275 len -= tcp_sk(sk)->tcp_header_len;
276 icsk->icsk_ack.last_seg_size = len;
278 icsk->icsk_ack.rcv_mss = len;
282 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
283 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
284 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
288 static void tcp_incr_quickack(struct sock *sk, unsigned int max_quickacks)
290 struct inet_connection_sock *icsk = inet_csk(sk);
291 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
295 quickacks = min(quickacks, max_quickacks);
296 if (quickacks > icsk->icsk_ack.quick)
297 icsk->icsk_ack.quick = quickacks;
300 static void tcp_enter_quickack_mode(struct sock *sk, unsigned int max_quickacks)
302 struct inet_connection_sock *icsk = inet_csk(sk);
304 tcp_incr_quickack(sk, max_quickacks);
305 inet_csk_exit_pingpong_mode(sk);
306 icsk->icsk_ack.ato = TCP_ATO_MIN;
309 /* Send ACKs quickly, if "quick" count is not exhausted
310 * and the session is not interactive.
313 static bool tcp_in_quickack_mode(struct sock *sk)
315 const struct inet_connection_sock *icsk = inet_csk(sk);
316 const struct dst_entry *dst = __sk_dst_get(sk);
318 return (dst && dst_metric(dst, RTAX_QUICKACK)) ||
319 (icsk->icsk_ack.quick && !inet_csk_in_pingpong_mode(sk));
322 static void tcp_ecn_queue_cwr(struct tcp_sock *tp)
324 if (tp->ecn_flags & TCP_ECN_OK)
325 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
328 static void tcp_ecn_accept_cwr(struct sock *sk, const struct sk_buff *skb)
330 if (tcp_hdr(skb)->cwr) {
331 tcp_sk(sk)->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
333 /* If the sender is telling us it has entered CWR, then its
334 * cwnd may be very low (even just 1 packet), so we should ACK
337 if (TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq)
338 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW;
342 static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp)
344 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
347 static void __tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb)
349 struct tcp_sock *tp = tcp_sk(sk);
351 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
352 case INET_ECN_NOT_ECT:
353 /* Funny extension: if ECT is not set on a segment,
354 * and we already seen ECT on a previous segment,
355 * it is probably a retransmit.
357 if (tp->ecn_flags & TCP_ECN_SEEN)
358 tcp_enter_quickack_mode(sk, 2);
361 if (tcp_ca_needs_ecn(sk))
362 tcp_ca_event(sk, CA_EVENT_ECN_IS_CE);
364 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
365 /* Better not delay acks, sender can have a very low cwnd */
366 tcp_enter_quickack_mode(sk, 2);
367 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
369 tp->ecn_flags |= TCP_ECN_SEEN;
372 if (tcp_ca_needs_ecn(sk))
373 tcp_ca_event(sk, CA_EVENT_ECN_NO_CE);
374 tp->ecn_flags |= TCP_ECN_SEEN;
379 static void tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb)
381 if (tcp_sk(sk)->ecn_flags & TCP_ECN_OK)
382 __tcp_ecn_check_ce(sk, skb);
385 static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
387 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
388 tp->ecn_flags &= ~TCP_ECN_OK;
391 static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
393 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
394 tp->ecn_flags &= ~TCP_ECN_OK;
397 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
399 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
404 /* Buffer size and advertised window tuning.
406 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
409 static void tcp_sndbuf_expand(struct sock *sk)
411 const struct tcp_sock *tp = tcp_sk(sk);
412 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
416 /* Worst case is non GSO/TSO : each frame consumes one skb
417 * and skb->head is kmalloced using power of two area of memory
419 per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
421 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
423 per_mss = roundup_pow_of_two(per_mss) +
424 SKB_DATA_ALIGN(sizeof(struct sk_buff));
426 nr_segs = max_t(u32, TCP_INIT_CWND, tcp_snd_cwnd(tp));
427 nr_segs = max_t(u32, nr_segs, tp->reordering + 1);
429 /* Fast Recovery (RFC 5681 3.2) :
430 * Cubic needs 1.7 factor, rounded to 2 to include
431 * extra cushion (application might react slowly to EPOLLOUT)
433 sndmem = ca_ops->sndbuf_expand ? ca_ops->sndbuf_expand(sk) : 2;
434 sndmem *= nr_segs * per_mss;
436 if (sk->sk_sndbuf < sndmem)
437 WRITE_ONCE(sk->sk_sndbuf,
438 min(sndmem, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_wmem[2])));
441 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
443 * All tcp_full_space() is split to two parts: "network" buffer, allocated
444 * forward and advertised in receiver window (tp->rcv_wnd) and
445 * "application buffer", required to isolate scheduling/application
446 * latencies from network.
447 * window_clamp is maximal advertised window. It can be less than
448 * tcp_full_space(), in this case tcp_full_space() - window_clamp
449 * is reserved for "application" buffer. The less window_clamp is
450 * the smoother our behaviour from viewpoint of network, but the lower
451 * throughput and the higher sensitivity of the connection to losses. 8)
453 * rcv_ssthresh is more strict window_clamp used at "slow start"
454 * phase to predict further behaviour of this connection.
455 * It is used for two goals:
456 * - to enforce header prediction at sender, even when application
457 * requires some significant "application buffer". It is check #1.
458 * - to prevent pruning of receive queue because of misprediction
459 * of receiver window. Check #2.
461 * The scheme does not work when sender sends good segments opening
462 * window and then starts to feed us spaghetti. But it should work
463 * in common situations. Otherwise, we have to rely on queue collapsing.
466 /* Slow part of check#2. */
467 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb,
468 unsigned int skbtruesize)
470 const struct tcp_sock *tp = tcp_sk(sk);
472 int truesize = tcp_win_from_space(sk, skbtruesize) >> 1;
473 int window = tcp_win_from_space(sk, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2])) >> 1;
475 while (tp->rcv_ssthresh <= window) {
476 if (truesize <= skb->len)
477 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
485 /* Even if skb appears to have a bad len/truesize ratio, TCP coalescing
486 * can play nice with us, as sk_buff and skb->head might be either
487 * freed or shared with up to MAX_SKB_FRAGS segments.
488 * Only give a boost to drivers using page frag(s) to hold the frame(s),
489 * and if no payload was pulled in skb->head before reaching us.
491 static u32 truesize_adjust(bool adjust, const struct sk_buff *skb)
493 u32 truesize = skb->truesize;
495 if (adjust && !skb_headlen(skb)) {
496 truesize -= SKB_TRUESIZE(skb_end_offset(skb));
497 /* paranoid check, some drivers might be buggy */
498 if (unlikely((int)truesize < (int)skb->len))
499 truesize = skb->truesize;
504 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb,
507 struct tcp_sock *tp = tcp_sk(sk);
510 room = min_t(int, tp->window_clamp, tcp_space(sk)) - tp->rcv_ssthresh;
516 if (!tcp_under_memory_pressure(sk)) {
517 unsigned int truesize = truesize_adjust(adjust, skb);
520 /* Check #2. Increase window, if skb with such overhead
521 * will fit to rcvbuf in future.
523 if (tcp_win_from_space(sk, truesize) <= skb->len)
524 incr = 2 * tp->advmss;
526 incr = __tcp_grow_window(sk, skb, truesize);
529 incr = max_t(int, incr, 2 * skb->len);
530 tp->rcv_ssthresh += min(room, incr);
531 inet_csk(sk)->icsk_ack.quick |= 1;
535 * Adjust rcv_ssthresh according to reserved mem
537 tcp_adjust_rcv_ssthresh(sk);
541 /* 3. Try to fixup all. It is made immediately after connection enters
544 static void tcp_init_buffer_space(struct sock *sk)
546 int tcp_app_win = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_app_win);
547 struct tcp_sock *tp = tcp_sk(sk);
550 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
551 tcp_sndbuf_expand(sk);
553 tcp_mstamp_refresh(tp);
554 tp->rcvq_space.time = tp->tcp_mstamp;
555 tp->rcvq_space.seq = tp->copied_seq;
557 maxwin = tcp_full_space(sk);
559 if (tp->window_clamp >= maxwin) {
560 tp->window_clamp = maxwin;
562 if (tcp_app_win && maxwin > 4 * tp->advmss)
563 tp->window_clamp = max(maxwin -
564 (maxwin >> tcp_app_win),
568 /* Force reservation of one segment. */
570 tp->window_clamp > 2 * tp->advmss &&
571 tp->window_clamp + tp->advmss > maxwin)
572 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
574 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
575 tp->snd_cwnd_stamp = tcp_jiffies32;
576 tp->rcvq_space.space = min3(tp->rcv_ssthresh, tp->rcv_wnd,
577 (u32)TCP_INIT_CWND * tp->advmss);
580 /* 4. Recalculate window clamp after socket hit its memory bounds. */
581 static void tcp_clamp_window(struct sock *sk)
583 struct tcp_sock *tp = tcp_sk(sk);
584 struct inet_connection_sock *icsk = inet_csk(sk);
585 struct net *net = sock_net(sk);
588 icsk->icsk_ack.quick = 0;
589 rmem2 = READ_ONCE(net->ipv4.sysctl_tcp_rmem[2]);
591 if (sk->sk_rcvbuf < rmem2 &&
592 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
593 !tcp_under_memory_pressure(sk) &&
594 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
595 WRITE_ONCE(sk->sk_rcvbuf,
596 min(atomic_read(&sk->sk_rmem_alloc), rmem2));
598 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
599 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
602 /* Initialize RCV_MSS value.
603 * RCV_MSS is an our guess about MSS used by the peer.
604 * We haven't any direct information about the MSS.
605 * It's better to underestimate the RCV_MSS rather than overestimate.
606 * Overestimations make us ACKing less frequently than needed.
607 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
609 void tcp_initialize_rcv_mss(struct sock *sk)
611 const struct tcp_sock *tp = tcp_sk(sk);
612 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
614 hint = min(hint, tp->rcv_wnd / 2);
615 hint = min(hint, TCP_MSS_DEFAULT);
616 hint = max(hint, TCP_MIN_MSS);
618 inet_csk(sk)->icsk_ack.rcv_mss = hint;
620 EXPORT_SYMBOL(tcp_initialize_rcv_mss);
622 /* Receiver "autotuning" code.
624 * The algorithm for RTT estimation w/o timestamps is based on
625 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
626 * <https://public.lanl.gov/radiant/pubs.html#DRS>
628 * More detail on this code can be found at
629 * <http://staff.psc.edu/jheffner/>,
630 * though this reference is out of date. A new paper
633 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
635 u32 new_sample = tp->rcv_rtt_est.rtt_us;
638 if (new_sample != 0) {
639 /* If we sample in larger samples in the non-timestamp
640 * case, we could grossly overestimate the RTT especially
641 * with chatty applications or bulk transfer apps which
642 * are stalled on filesystem I/O.
644 * Also, since we are only going for a minimum in the
645 * non-timestamp case, we do not smooth things out
646 * else with timestamps disabled convergence takes too
650 m -= (new_sample >> 3);
658 /* No previous measure. */
662 tp->rcv_rtt_est.rtt_us = new_sample;
665 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
669 if (tp->rcv_rtt_est.time == 0)
671 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
673 delta_us = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcv_rtt_est.time);
676 tcp_rcv_rtt_update(tp, delta_us, 1);
679 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
680 tp->rcv_rtt_est.time = tp->tcp_mstamp;
683 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
684 const struct sk_buff *skb)
686 struct tcp_sock *tp = tcp_sk(sk);
688 if (tp->rx_opt.rcv_tsecr == tp->rcv_rtt_last_tsecr)
690 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr;
692 if (TCP_SKB_CB(skb)->end_seq -
693 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss) {
694 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr;
697 if (likely(delta < INT_MAX / (USEC_PER_SEC / TCP_TS_HZ))) {
700 delta_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
701 tcp_rcv_rtt_update(tp, delta_us, 0);
707 * This function should be called every time data is copied to user space.
708 * It calculates the appropriate TCP receive buffer space.
710 void tcp_rcv_space_adjust(struct sock *sk)
712 struct tcp_sock *tp = tcp_sk(sk);
716 trace_tcp_rcv_space_adjust(sk);
718 tcp_mstamp_refresh(tp);
719 time = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcvq_space.time);
720 if (time < (tp->rcv_rtt_est.rtt_us >> 3) || tp->rcv_rtt_est.rtt_us == 0)
723 /* Number of bytes copied to user in last RTT */
724 copied = tp->copied_seq - tp->rcvq_space.seq;
725 if (copied <= tp->rcvq_space.space)
729 * copied = bytes received in previous RTT, our base window
730 * To cope with packet losses, we need a 2x factor
731 * To cope with slow start, and sender growing its cwin by 100 %
732 * every RTT, we need a 4x factor, because the ACK we are sending
733 * now is for the next RTT, not the current one :
734 * <prev RTT . ><current RTT .. ><next RTT .... >
737 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf) &&
738 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
742 /* minimal window to cope with packet losses, assuming
743 * steady state. Add some cushion because of small variations.
745 rcvwin = ((u64)copied << 1) + 16 * tp->advmss;
747 /* Accommodate for sender rate increase (eg. slow start) */
748 grow = rcvwin * (copied - tp->rcvq_space.space);
749 do_div(grow, tp->rcvq_space.space);
750 rcvwin += (grow << 1);
752 rcvbuf = min_t(u64, tcp_space_from_win(sk, rcvwin),
753 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]));
754 if (rcvbuf > sk->sk_rcvbuf) {
755 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
757 /* Make the window clamp follow along. */
758 tp->window_clamp = tcp_win_from_space(sk, rcvbuf);
761 tp->rcvq_space.space = copied;
764 tp->rcvq_space.seq = tp->copied_seq;
765 tp->rcvq_space.time = tp->tcp_mstamp;
768 /* There is something which you must keep in mind when you analyze the
769 * behavior of the tp->ato delayed ack timeout interval. When a
770 * connection starts up, we want to ack as quickly as possible. The
771 * problem is that "good" TCP's do slow start at the beginning of data
772 * transmission. The means that until we send the first few ACK's the
773 * sender will sit on his end and only queue most of his data, because
774 * he can only send snd_cwnd unacked packets at any given time. For
775 * each ACK we send, he increments snd_cwnd and transmits more of his
778 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
780 struct tcp_sock *tp = tcp_sk(sk);
781 struct inet_connection_sock *icsk = inet_csk(sk);
784 inet_csk_schedule_ack(sk);
786 tcp_measure_rcv_mss(sk, skb);
788 tcp_rcv_rtt_measure(tp);
792 if (!icsk->icsk_ack.ato) {
793 /* The _first_ data packet received, initialize
794 * delayed ACK engine.
796 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
797 icsk->icsk_ack.ato = TCP_ATO_MIN;
799 int m = now - icsk->icsk_ack.lrcvtime;
801 if (m <= TCP_ATO_MIN / 2) {
802 /* The fastest case is the first. */
803 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
804 } else if (m < icsk->icsk_ack.ato) {
805 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
806 if (icsk->icsk_ack.ato > icsk->icsk_rto)
807 icsk->icsk_ack.ato = icsk->icsk_rto;
808 } else if (m > icsk->icsk_rto) {
809 /* Too long gap. Apparently sender failed to
810 * restart window, so that we send ACKs quickly.
812 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
815 icsk->icsk_ack.lrcvtime = now;
817 tcp_ecn_check_ce(sk, skb);
820 tcp_grow_window(sk, skb, true);
823 /* Called to compute a smoothed rtt estimate. The data fed to this
824 * routine either comes from timestamps, or from segments that were
825 * known _not_ to have been retransmitted [see Karn/Partridge
826 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
827 * piece by Van Jacobson.
828 * NOTE: the next three routines used to be one big routine.
829 * To save cycles in the RFC 1323 implementation it was better to break
830 * it up into three procedures. -- erics
832 static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
834 struct tcp_sock *tp = tcp_sk(sk);
835 long m = mrtt_us; /* RTT */
836 u32 srtt = tp->srtt_us;
838 /* The following amusing code comes from Jacobson's
839 * article in SIGCOMM '88. Note that rtt and mdev
840 * are scaled versions of rtt and mean deviation.
841 * This is designed to be as fast as possible
842 * m stands for "measurement".
844 * On a 1990 paper the rto value is changed to:
845 * RTO = rtt + 4 * mdev
847 * Funny. This algorithm seems to be very broken.
848 * These formulae increase RTO, when it should be decreased, increase
849 * too slowly, when it should be increased quickly, decrease too quickly
850 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
851 * does not matter how to _calculate_ it. Seems, it was trap
852 * that VJ failed to avoid. 8)
855 m -= (srtt >> 3); /* m is now error in rtt est */
856 srtt += m; /* rtt = 7/8 rtt + 1/8 new */
858 m = -m; /* m is now abs(error) */
859 m -= (tp->mdev_us >> 2); /* similar update on mdev */
860 /* This is similar to one of Eifel findings.
861 * Eifel blocks mdev updates when rtt decreases.
862 * This solution is a bit different: we use finer gain
863 * for mdev in this case (alpha*beta).
864 * Like Eifel it also prevents growth of rto,
865 * but also it limits too fast rto decreases,
866 * happening in pure Eifel.
871 m -= (tp->mdev_us >> 2); /* similar update on mdev */
873 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */
874 if (tp->mdev_us > tp->mdev_max_us) {
875 tp->mdev_max_us = tp->mdev_us;
876 if (tp->mdev_max_us > tp->rttvar_us)
877 tp->rttvar_us = tp->mdev_max_us;
879 if (after(tp->snd_una, tp->rtt_seq)) {
880 if (tp->mdev_max_us < tp->rttvar_us)
881 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
882 tp->rtt_seq = tp->snd_nxt;
883 tp->mdev_max_us = tcp_rto_min_us(sk);
888 /* no previous measure. */
889 srtt = m << 3; /* take the measured time to be rtt */
890 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */
891 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
892 tp->mdev_max_us = tp->rttvar_us;
893 tp->rtt_seq = tp->snd_nxt;
897 tp->srtt_us = max(1U, srtt);
900 static void tcp_update_pacing_rate(struct sock *sk)
902 const struct tcp_sock *tp = tcp_sk(sk);
905 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
906 rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3);
908 /* current rate is (cwnd * mss) / srtt
909 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
910 * In Congestion Avoidance phase, set it to 120 % the current rate.
912 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
913 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
914 * end of slow start and should slow down.
916 if (tcp_snd_cwnd(tp) < tp->snd_ssthresh / 2)
917 rate *= READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_pacing_ss_ratio);
919 rate *= READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_pacing_ca_ratio);
921 rate *= max(tcp_snd_cwnd(tp), tp->packets_out);
923 if (likely(tp->srtt_us))
924 do_div(rate, tp->srtt_us);
926 /* WRITE_ONCE() is needed because sch_fq fetches sk_pacing_rate
927 * without any lock. We want to make sure compiler wont store
928 * intermediate values in this location.
930 WRITE_ONCE(sk->sk_pacing_rate, min_t(u64, rate,
931 sk->sk_max_pacing_rate));
934 /* Calculate rto without backoff. This is the second half of Van Jacobson's
935 * routine referred to above.
937 static void tcp_set_rto(struct sock *sk)
939 const struct tcp_sock *tp = tcp_sk(sk);
940 /* Old crap is replaced with new one. 8)
943 * 1. If rtt variance happened to be less 50msec, it is hallucination.
944 * It cannot be less due to utterly erratic ACK generation made
945 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
946 * to do with delayed acks, because at cwnd>2 true delack timeout
947 * is invisible. Actually, Linux-2.4 also generates erratic
948 * ACKs in some circumstances.
950 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
952 /* 2. Fixups made earlier cannot be right.
953 * If we do not estimate RTO correctly without them,
954 * all the algo is pure shit and should be replaced
955 * with correct one. It is exactly, which we pretend to do.
958 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
959 * guarantees that rto is higher.
964 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
966 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
969 cwnd = TCP_INIT_CWND;
970 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
973 struct tcp_sacktag_state {
974 /* Timestamps for earliest and latest never-retransmitted segment
975 * that was SACKed. RTO needs the earliest RTT to stay conservative,
976 * but congestion control should still get an accurate delay signal.
983 unsigned int mss_now;
984 struct rate_sample *rate;
987 /* Take a notice that peer is sending D-SACKs. Skip update of data delivery
988 * and spurious retransmission information if this DSACK is unlikely caused by
990 * - DSACKed sequence range is larger than maximum receiver's window.
991 * - Total no. of DSACKed segments exceed the total no. of retransmitted segs.
993 static u32 tcp_dsack_seen(struct tcp_sock *tp, u32 start_seq,
994 u32 end_seq, struct tcp_sacktag_state *state)
996 u32 seq_len, dup_segs = 1;
998 if (!before(start_seq, end_seq))
1001 seq_len = end_seq - start_seq;
1002 /* Dubious DSACK: DSACKed range greater than maximum advertised rwnd */
1003 if (seq_len > tp->max_window)
1005 if (seq_len > tp->mss_cache)
1006 dup_segs = DIV_ROUND_UP(seq_len, tp->mss_cache);
1007 else if (tp->tlp_high_seq && tp->tlp_high_seq == end_seq)
1008 state->flag |= FLAG_DSACK_TLP;
1010 tp->dsack_dups += dup_segs;
1011 /* Skip the DSACK if dup segs weren't retransmitted by sender */
1012 if (tp->dsack_dups > tp->total_retrans)
1015 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
1016 /* We increase the RACK ordering window in rounds where we receive
1017 * DSACKs that may have been due to reordering causing RACK to trigger
1018 * a spurious fast recovery. Thus RACK ignores DSACKs that happen
1019 * without having seen reordering, or that match TLP probes (TLP
1020 * is timer-driven, not triggered by RACK).
1022 if (tp->reord_seen && !(state->flag & FLAG_DSACK_TLP))
1023 tp->rack.dsack_seen = 1;
1025 state->flag |= FLAG_DSACKING_ACK;
1026 /* A spurious retransmission is delivered */
1027 state->sack_delivered += dup_segs;
1032 /* It's reordering when higher sequence was delivered (i.e. sacked) before
1033 * some lower never-retransmitted sequence ("low_seq"). The maximum reordering
1034 * distance is approximated in full-mss packet distance ("reordering").
1036 static void tcp_check_sack_reordering(struct sock *sk, const u32 low_seq,
1039 struct tcp_sock *tp = tcp_sk(sk);
1040 const u32 mss = tp->mss_cache;
1043 fack = tcp_highest_sack_seq(tp);
1044 if (!before(low_seq, fack))
1047 metric = fack - low_seq;
1048 if ((metric > tp->reordering * mss) && mss) {
1049 #if FASTRETRANS_DEBUG > 1
1050 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
1051 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
1055 tp->undo_marker ? tp->undo_retrans : 0);
1057 tp->reordering = min_t(u32, (metric + mss - 1) / mss,
1058 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_max_reordering));
1061 /* This exciting event is worth to be remembered. 8) */
1063 NET_INC_STATS(sock_net(sk),
1064 ts ? LINUX_MIB_TCPTSREORDER : LINUX_MIB_TCPSACKREORDER);
1067 /* This must be called before lost_out or retrans_out are updated
1068 * on a new loss, because we want to know if all skbs previously
1069 * known to be lost have already been retransmitted, indicating
1070 * that this newly lost skb is our next skb to retransmit.
1072 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
1074 if ((!tp->retransmit_skb_hint && tp->retrans_out >= tp->lost_out) ||
1075 (tp->retransmit_skb_hint &&
1076 before(TCP_SKB_CB(skb)->seq,
1077 TCP_SKB_CB(tp->retransmit_skb_hint)->seq)))
1078 tp->retransmit_skb_hint = skb;
1081 /* Sum the number of packets on the wire we have marked as lost, and
1082 * notify the congestion control module that the given skb was marked lost.
1084 static void tcp_notify_skb_loss_event(struct tcp_sock *tp, const struct sk_buff *skb)
1086 tp->lost += tcp_skb_pcount(skb);
1089 void tcp_mark_skb_lost(struct sock *sk, struct sk_buff *skb)
1091 __u8 sacked = TCP_SKB_CB(skb)->sacked;
1092 struct tcp_sock *tp = tcp_sk(sk);
1094 if (sacked & TCPCB_SACKED_ACKED)
1097 tcp_verify_retransmit_hint(tp, skb);
1098 if (sacked & TCPCB_LOST) {
1099 if (sacked & TCPCB_SACKED_RETRANS) {
1100 /* Account for retransmits that are lost again */
1101 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1102 tp->retrans_out -= tcp_skb_pcount(skb);
1103 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT,
1104 tcp_skb_pcount(skb));
1105 tcp_notify_skb_loss_event(tp, skb);
1108 tp->lost_out += tcp_skb_pcount(skb);
1109 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1110 tcp_notify_skb_loss_event(tp, skb);
1114 /* Updates the delivered and delivered_ce counts */
1115 static void tcp_count_delivered(struct tcp_sock *tp, u32 delivered,
1118 tp->delivered += delivered;
1120 tp->delivered_ce += delivered;
1123 /* This procedure tags the retransmission queue when SACKs arrive.
1125 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
1126 * Packets in queue with these bits set are counted in variables
1127 * sacked_out, retrans_out and lost_out, correspondingly.
1129 * Valid combinations are:
1130 * Tag InFlight Description
1131 * 0 1 - orig segment is in flight.
1132 * S 0 - nothing flies, orig reached receiver.
1133 * L 0 - nothing flies, orig lost by net.
1134 * R 2 - both orig and retransmit are in flight.
1135 * L|R 1 - orig is lost, retransmit is in flight.
1136 * S|R 1 - orig reached receiver, retrans is still in flight.
1137 * (L|S|R is logically valid, it could occur when L|R is sacked,
1138 * but it is equivalent to plain S and code short-curcuits it to S.
1139 * L|S is logically invalid, it would mean -1 packet in flight 8))
1141 * These 6 states form finite state machine, controlled by the following events:
1142 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
1143 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
1144 * 3. Loss detection event of two flavors:
1145 * A. Scoreboard estimator decided the packet is lost.
1146 * A'. Reno "three dupacks" marks head of queue lost.
1147 * B. SACK arrives sacking SND.NXT at the moment, when the
1148 * segment was retransmitted.
1149 * 4. D-SACK added new rule: D-SACK changes any tag to S.
1151 * It is pleasant to note, that state diagram turns out to be commutative,
1152 * so that we are allowed not to be bothered by order of our actions,
1153 * when multiple events arrive simultaneously. (see the function below).
1155 * Reordering detection.
1156 * --------------------
1157 * Reordering metric is maximal distance, which a packet can be displaced
1158 * in packet stream. With SACKs we can estimate it:
1160 * 1. SACK fills old hole and the corresponding segment was not
1161 * ever retransmitted -> reordering. Alas, we cannot use it
1162 * when segment was retransmitted.
1163 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1164 * for retransmitted and already SACKed segment -> reordering..
1165 * Both of these heuristics are not used in Loss state, when we cannot
1166 * account for retransmits accurately.
1168 * SACK block validation.
1169 * ----------------------
1171 * SACK block range validation checks that the received SACK block fits to
1172 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1173 * Note that SND.UNA is not included to the range though being valid because
1174 * it means that the receiver is rather inconsistent with itself reporting
1175 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1176 * perfectly valid, however, in light of RFC2018 which explicitly states
1177 * that "SACK block MUST reflect the newest segment. Even if the newest
1178 * segment is going to be discarded ...", not that it looks very clever
1179 * in case of head skb. Due to potentional receiver driven attacks, we
1180 * choose to avoid immediate execution of a walk in write queue due to
1181 * reneging and defer head skb's loss recovery to standard loss recovery
1182 * procedure that will eventually trigger (nothing forbids us doing this).
1184 * Implements also blockage to start_seq wrap-around. Problem lies in the
1185 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1186 * there's no guarantee that it will be before snd_nxt (n). The problem
1187 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1190 * <- outs wnd -> <- wrapzone ->
1191 * u e n u_w e_w s n_w
1193 * |<------------+------+----- TCP seqno space --------------+---------->|
1194 * ...-- <2^31 ->| |<--------...
1195 * ...---- >2^31 ------>| |<--------...
1197 * Current code wouldn't be vulnerable but it's better still to discard such
1198 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1199 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1200 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1201 * equal to the ideal case (infinite seqno space without wrap caused issues).
1203 * With D-SACK the lower bound is extended to cover sequence space below
1204 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1205 * again, D-SACK block must not to go across snd_una (for the same reason as
1206 * for the normal SACK blocks, explained above). But there all simplicity
1207 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1208 * fully below undo_marker they do not affect behavior in anyway and can
1209 * therefore be safely ignored. In rare cases (which are more or less
1210 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1211 * fragmentation and packet reordering past skb's retransmission. To consider
1212 * them correctly, the acceptable range must be extended even more though
1213 * the exact amount is rather hard to quantify. However, tp->max_window can
1214 * be used as an exaggerated estimate.
1216 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
1217 u32 start_seq, u32 end_seq)
1219 /* Too far in future, or reversed (interpretation is ambiguous) */
1220 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1223 /* Nasty start_seq wrap-around check (see comments above) */
1224 if (!before(start_seq, tp->snd_nxt))
1227 /* In outstanding window? ...This is valid exit for D-SACKs too.
1228 * start_seq == snd_una is non-sensical (see comments above)
1230 if (after(start_seq, tp->snd_una))
1233 if (!is_dsack || !tp->undo_marker)
1236 /* ...Then it's D-SACK, and must reside below snd_una completely */
1237 if (after(end_seq, tp->snd_una))
1240 if (!before(start_seq, tp->undo_marker))
1244 if (!after(end_seq, tp->undo_marker))
1247 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1248 * start_seq < undo_marker and end_seq >= undo_marker.
1250 return !before(start_seq, end_seq - tp->max_window);
1253 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1254 struct tcp_sack_block_wire *sp, int num_sacks,
1255 u32 prior_snd_una, struct tcp_sacktag_state *state)
1257 struct tcp_sock *tp = tcp_sk(sk);
1258 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1259 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1262 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1263 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1264 } else if (num_sacks > 1) {
1265 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1266 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1268 if (after(end_seq_0, end_seq_1) || before(start_seq_0, start_seq_1))
1270 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKOFORECV);
1275 dup_segs = tcp_dsack_seen(tp, start_seq_0, end_seq_0, state);
1276 if (!dup_segs) { /* Skip dubious DSACK */
1277 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKIGNOREDDUBIOUS);
1281 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECVSEGS, dup_segs);
1283 /* D-SACK for already forgotten data... Do dumb counting. */
1284 if (tp->undo_marker && tp->undo_retrans > 0 &&
1285 !after(end_seq_0, prior_snd_una) &&
1286 after(end_seq_0, tp->undo_marker))
1287 tp->undo_retrans = max_t(int, 0, tp->undo_retrans - dup_segs);
1292 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1293 * the incoming SACK may not exactly match but we can find smaller MSS
1294 * aligned portion of it that matches. Therefore we might need to fragment
1295 * which may fail and creates some hassle (caller must handle error case
1298 * FIXME: this could be merged to shift decision code
1300 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1301 u32 start_seq, u32 end_seq)
1305 unsigned int pkt_len;
1308 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1309 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1311 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1312 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1313 mss = tcp_skb_mss(skb);
1314 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1317 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1321 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1326 /* Round if necessary so that SACKs cover only full MSSes
1327 * and/or the remaining small portion (if present)
1329 if (pkt_len > mss) {
1330 unsigned int new_len = (pkt_len / mss) * mss;
1331 if (!in_sack && new_len < pkt_len)
1336 if (pkt_len >= skb->len && !in_sack)
1339 err = tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
1340 pkt_len, mss, GFP_ATOMIC);
1348 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1349 static u8 tcp_sacktag_one(struct sock *sk,
1350 struct tcp_sacktag_state *state, u8 sacked,
1351 u32 start_seq, u32 end_seq,
1352 int dup_sack, int pcount,
1355 struct tcp_sock *tp = tcp_sk(sk);
1357 /* Account D-SACK for retransmitted packet. */
1358 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1359 if (tp->undo_marker && tp->undo_retrans > 0 &&
1360 after(end_seq, tp->undo_marker))
1361 tp->undo_retrans = max_t(int, 0, tp->undo_retrans - pcount);
1362 if ((sacked & TCPCB_SACKED_ACKED) &&
1363 before(start_seq, state->reord))
1364 state->reord = start_seq;
1367 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1368 if (!after(end_seq, tp->snd_una))
1371 if (!(sacked & TCPCB_SACKED_ACKED)) {
1372 tcp_rack_advance(tp, sacked, end_seq, xmit_time);
1374 if (sacked & TCPCB_SACKED_RETRANS) {
1375 /* If the segment is not tagged as lost,
1376 * we do not clear RETRANS, believing
1377 * that retransmission is still in flight.
1379 if (sacked & TCPCB_LOST) {
1380 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1381 tp->lost_out -= pcount;
1382 tp->retrans_out -= pcount;
1385 if (!(sacked & TCPCB_RETRANS)) {
1386 /* New sack for not retransmitted frame,
1387 * which was in hole. It is reordering.
1389 if (before(start_seq,
1390 tcp_highest_sack_seq(tp)) &&
1391 before(start_seq, state->reord))
1392 state->reord = start_seq;
1394 if (!after(end_seq, tp->high_seq))
1395 state->flag |= FLAG_ORIG_SACK_ACKED;
1396 if (state->first_sackt == 0)
1397 state->first_sackt = xmit_time;
1398 state->last_sackt = xmit_time;
1401 if (sacked & TCPCB_LOST) {
1402 sacked &= ~TCPCB_LOST;
1403 tp->lost_out -= pcount;
1407 sacked |= TCPCB_SACKED_ACKED;
1408 state->flag |= FLAG_DATA_SACKED;
1409 tp->sacked_out += pcount;
1410 /* Out-of-order packets delivered */
1411 state->sack_delivered += pcount;
1413 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1414 if (tp->lost_skb_hint &&
1415 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1416 tp->lost_cnt_hint += pcount;
1419 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1420 * frames and clear it. undo_retrans is decreased above, L|R frames
1421 * are accounted above as well.
1423 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1424 sacked &= ~TCPCB_SACKED_RETRANS;
1425 tp->retrans_out -= pcount;
1431 /* Shift newly-SACKed bytes from this skb to the immediately previous
1432 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1434 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *prev,
1435 struct sk_buff *skb,
1436 struct tcp_sacktag_state *state,
1437 unsigned int pcount, int shifted, int mss,
1440 struct tcp_sock *tp = tcp_sk(sk);
1441 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1442 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
1446 /* Adjust counters and hints for the newly sacked sequence
1447 * range but discard the return value since prev is already
1448 * marked. We must tag the range first because the seq
1449 * advancement below implicitly advances
1450 * tcp_highest_sack_seq() when skb is highest_sack.
1452 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1453 start_seq, end_seq, dup_sack, pcount,
1454 tcp_skb_timestamp_us(skb));
1455 tcp_rate_skb_delivered(sk, skb, state->rate);
1457 if (skb == tp->lost_skb_hint)
1458 tp->lost_cnt_hint += pcount;
1460 TCP_SKB_CB(prev)->end_seq += shifted;
1461 TCP_SKB_CB(skb)->seq += shifted;
1463 tcp_skb_pcount_add(prev, pcount);
1464 WARN_ON_ONCE(tcp_skb_pcount(skb) < pcount);
1465 tcp_skb_pcount_add(skb, -pcount);
1467 /* When we're adding to gso_segs == 1, gso_size will be zero,
1468 * in theory this shouldn't be necessary but as long as DSACK
1469 * code can come after this skb later on it's better to keep
1470 * setting gso_size to something.
1472 if (!TCP_SKB_CB(prev)->tcp_gso_size)
1473 TCP_SKB_CB(prev)->tcp_gso_size = mss;
1475 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1476 if (tcp_skb_pcount(skb) <= 1)
1477 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1479 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1480 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1483 BUG_ON(!tcp_skb_pcount(skb));
1484 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1488 /* Whole SKB was eaten :-) */
1490 if (skb == tp->retransmit_skb_hint)
1491 tp->retransmit_skb_hint = prev;
1492 if (skb == tp->lost_skb_hint) {
1493 tp->lost_skb_hint = prev;
1494 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1497 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1498 TCP_SKB_CB(prev)->eor = TCP_SKB_CB(skb)->eor;
1499 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1500 TCP_SKB_CB(prev)->end_seq++;
1502 if (skb == tcp_highest_sack(sk))
1503 tcp_advance_highest_sack(sk, skb);
1505 tcp_skb_collapse_tstamp(prev, skb);
1506 if (unlikely(TCP_SKB_CB(prev)->tx.delivered_mstamp))
1507 TCP_SKB_CB(prev)->tx.delivered_mstamp = 0;
1509 tcp_rtx_queue_unlink_and_free(skb, sk);
1511 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKMERGED);
1516 /* I wish gso_size would have a bit more sane initialization than
1517 * something-or-zero which complicates things
1519 static int tcp_skb_seglen(const struct sk_buff *skb)
1521 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1524 /* Shifting pages past head area doesn't work */
1525 static int skb_can_shift(const struct sk_buff *skb)
1527 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1530 int tcp_skb_shift(struct sk_buff *to, struct sk_buff *from,
1531 int pcount, int shiftlen)
1533 /* TCP min gso_size is 8 bytes (TCP_MIN_GSO_SIZE)
1534 * Since TCP_SKB_CB(skb)->tcp_gso_segs is 16 bits, we need
1535 * to make sure not storing more than 65535 * 8 bytes per skb,
1536 * even if current MSS is bigger.
1538 if (unlikely(to->len + shiftlen >= 65535 * TCP_MIN_GSO_SIZE))
1540 if (unlikely(tcp_skb_pcount(to) + pcount > 65535))
1542 return skb_shift(to, from, shiftlen);
1545 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1548 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1549 struct tcp_sacktag_state *state,
1550 u32 start_seq, u32 end_seq,
1553 struct tcp_sock *tp = tcp_sk(sk);
1554 struct sk_buff *prev;
1560 /* Normally R but no L won't result in plain S */
1562 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1564 if (!skb_can_shift(skb))
1566 /* This frame is about to be dropped (was ACKed). */
1567 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1570 /* Can only happen with delayed DSACK + discard craziness */
1571 prev = skb_rb_prev(skb);
1575 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1578 if (!tcp_skb_can_collapse(prev, skb))
1581 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1582 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1586 pcount = tcp_skb_pcount(skb);
1587 mss = tcp_skb_seglen(skb);
1589 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1590 * drop this restriction as unnecessary
1592 if (mss != tcp_skb_seglen(prev))
1595 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1597 /* CHECKME: This is non-MSS split case only?, this will
1598 * cause skipped skbs due to advancing loop btw, original
1599 * has that feature too
1601 if (tcp_skb_pcount(skb) <= 1)
1604 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1606 /* TODO: head merge to next could be attempted here
1607 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1608 * though it might not be worth of the additional hassle
1610 * ...we can probably just fallback to what was done
1611 * previously. We could try merging non-SACKed ones
1612 * as well but it probably isn't going to buy off
1613 * because later SACKs might again split them, and
1614 * it would make skb timestamp tracking considerably
1620 len = end_seq - TCP_SKB_CB(skb)->seq;
1622 BUG_ON(len > skb->len);
1624 /* MSS boundaries should be honoured or else pcount will
1625 * severely break even though it makes things bit trickier.
1626 * Optimize common case to avoid most of the divides
1628 mss = tcp_skb_mss(skb);
1630 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1631 * drop this restriction as unnecessary
1633 if (mss != tcp_skb_seglen(prev))
1638 } else if (len < mss) {
1646 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1647 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1650 if (!tcp_skb_shift(prev, skb, pcount, len))
1652 if (!tcp_shifted_skb(sk, prev, skb, state, pcount, len, mss, dup_sack))
1655 /* Hole filled allows collapsing with the next as well, this is very
1656 * useful when hole on every nth skb pattern happens
1658 skb = skb_rb_next(prev);
1662 if (!skb_can_shift(skb) ||
1663 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1664 (mss != tcp_skb_seglen(skb)))
1667 if (!tcp_skb_can_collapse(prev, skb))
1670 pcount = tcp_skb_pcount(skb);
1671 if (tcp_skb_shift(prev, skb, pcount, len))
1672 tcp_shifted_skb(sk, prev, skb, state, pcount,
1682 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1686 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1687 struct tcp_sack_block *next_dup,
1688 struct tcp_sacktag_state *state,
1689 u32 start_seq, u32 end_seq,
1692 struct tcp_sock *tp = tcp_sk(sk);
1693 struct sk_buff *tmp;
1695 skb_rbtree_walk_from(skb) {
1697 bool dup_sack = dup_sack_in;
1699 /* queue is in-order => we can short-circuit the walk early */
1700 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1704 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1705 in_sack = tcp_match_skb_to_sack(sk, skb,
1706 next_dup->start_seq,
1712 /* skb reference here is a bit tricky to get right, since
1713 * shifting can eat and free both this skb and the next,
1714 * so not even _safe variant of the loop is enough.
1717 tmp = tcp_shift_skb_data(sk, skb, state,
1718 start_seq, end_seq, dup_sack);
1727 in_sack = tcp_match_skb_to_sack(sk, skb,
1733 if (unlikely(in_sack < 0))
1737 TCP_SKB_CB(skb)->sacked =
1740 TCP_SKB_CB(skb)->sacked,
1741 TCP_SKB_CB(skb)->seq,
1742 TCP_SKB_CB(skb)->end_seq,
1744 tcp_skb_pcount(skb),
1745 tcp_skb_timestamp_us(skb));
1746 tcp_rate_skb_delivered(sk, skb, state->rate);
1747 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1748 list_del_init(&skb->tcp_tsorted_anchor);
1750 if (!before(TCP_SKB_CB(skb)->seq,
1751 tcp_highest_sack_seq(tp)))
1752 tcp_advance_highest_sack(sk, skb);
1758 static struct sk_buff *tcp_sacktag_bsearch(struct sock *sk, u32 seq)
1760 struct rb_node *parent, **p = &sk->tcp_rtx_queue.rb_node;
1761 struct sk_buff *skb;
1765 skb = rb_to_skb(parent);
1766 if (before(seq, TCP_SKB_CB(skb)->seq)) {
1767 p = &parent->rb_left;
1770 if (!before(seq, TCP_SKB_CB(skb)->end_seq)) {
1771 p = &parent->rb_right;
1779 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1782 if (skb && after(TCP_SKB_CB(skb)->seq, skip_to_seq))
1785 return tcp_sacktag_bsearch(sk, skip_to_seq);
1788 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1790 struct tcp_sack_block *next_dup,
1791 struct tcp_sacktag_state *state,
1797 if (before(next_dup->start_seq, skip_to_seq)) {
1798 skb = tcp_sacktag_skip(skb, sk, next_dup->start_seq);
1799 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1800 next_dup->start_seq, next_dup->end_seq,
1807 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1809 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1813 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1814 u32 prior_snd_una, struct tcp_sacktag_state *state)
1816 struct tcp_sock *tp = tcp_sk(sk);
1817 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1818 TCP_SKB_CB(ack_skb)->sacked);
1819 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1820 struct tcp_sack_block sp[TCP_NUM_SACKS];
1821 struct tcp_sack_block *cache;
1822 struct sk_buff *skb;
1823 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1825 bool found_dup_sack = false;
1827 int first_sack_index;
1830 state->reord = tp->snd_nxt;
1832 if (!tp->sacked_out)
1833 tcp_highest_sack_reset(sk);
1835 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1836 num_sacks, prior_snd_una, state);
1838 /* Eliminate too old ACKs, but take into
1839 * account more or less fresh ones, they can
1840 * contain valid SACK info.
1842 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1845 if (!tp->packets_out)
1849 first_sack_index = 0;
1850 for (i = 0; i < num_sacks; i++) {
1851 bool dup_sack = !i && found_dup_sack;
1853 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1854 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1856 if (!tcp_is_sackblock_valid(tp, dup_sack,
1857 sp[used_sacks].start_seq,
1858 sp[used_sacks].end_seq)) {
1862 if (!tp->undo_marker)
1863 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1865 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1867 /* Don't count olds caused by ACK reordering */
1868 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1869 !after(sp[used_sacks].end_seq, tp->snd_una))
1871 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1874 NET_INC_STATS(sock_net(sk), mib_idx);
1876 first_sack_index = -1;
1880 /* Ignore very old stuff early */
1881 if (!after(sp[used_sacks].end_seq, prior_snd_una)) {
1883 first_sack_index = -1;
1890 /* order SACK blocks to allow in order walk of the retrans queue */
1891 for (i = used_sacks - 1; i > 0; i--) {
1892 for (j = 0; j < i; j++) {
1893 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1894 swap(sp[j], sp[j + 1]);
1896 /* Track where the first SACK block goes to */
1897 if (j == first_sack_index)
1898 first_sack_index = j + 1;
1903 state->mss_now = tcp_current_mss(sk);
1907 if (!tp->sacked_out) {
1908 /* It's already past, so skip checking against it */
1909 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1911 cache = tp->recv_sack_cache;
1912 /* Skip empty blocks in at head of the cache */
1913 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1918 while (i < used_sacks) {
1919 u32 start_seq = sp[i].start_seq;
1920 u32 end_seq = sp[i].end_seq;
1921 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1922 struct tcp_sack_block *next_dup = NULL;
1924 if (found_dup_sack && ((i + 1) == first_sack_index))
1925 next_dup = &sp[i + 1];
1927 /* Skip too early cached blocks */
1928 while (tcp_sack_cache_ok(tp, cache) &&
1929 !before(start_seq, cache->end_seq))
1932 /* Can skip some work by looking recv_sack_cache? */
1933 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1934 after(end_seq, cache->start_seq)) {
1937 if (before(start_seq, cache->start_seq)) {
1938 skb = tcp_sacktag_skip(skb, sk, start_seq);
1939 skb = tcp_sacktag_walk(skb, sk, next_dup,
1946 /* Rest of the block already fully processed? */
1947 if (!after(end_seq, cache->end_seq))
1950 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1954 /* ...tail remains todo... */
1955 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1956 /* ...but better entrypoint exists! */
1957 skb = tcp_highest_sack(sk);
1964 skb = tcp_sacktag_skip(skb, sk, cache->end_seq);
1965 /* Check overlap against next cached too (past this one already) */
1970 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1971 skb = tcp_highest_sack(sk);
1975 skb = tcp_sacktag_skip(skb, sk, start_seq);
1978 skb = tcp_sacktag_walk(skb, sk, next_dup, state,
1979 start_seq, end_seq, dup_sack);
1985 /* Clear the head of the cache sack blocks so we can skip it next time */
1986 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1987 tp->recv_sack_cache[i].start_seq = 0;
1988 tp->recv_sack_cache[i].end_seq = 0;
1990 for (j = 0; j < used_sacks; j++)
1991 tp->recv_sack_cache[i++] = sp[j];
1993 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Loss || tp->undo_marker)
1994 tcp_check_sack_reordering(sk, state->reord, 0);
1996 tcp_verify_left_out(tp);
1999 #if FASTRETRANS_DEBUG > 0
2000 WARN_ON((int)tp->sacked_out < 0);
2001 WARN_ON((int)tp->lost_out < 0);
2002 WARN_ON((int)tp->retrans_out < 0);
2003 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
2008 /* Limits sacked_out so that sum with lost_out isn't ever larger than
2009 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
2011 static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
2015 holes = max(tp->lost_out, 1U);
2016 holes = min(holes, tp->packets_out);
2018 if ((tp->sacked_out + holes) > tp->packets_out) {
2019 tp->sacked_out = tp->packets_out - holes;
2025 /* If we receive more dupacks than we expected counting segments
2026 * in assumption of absent reordering, interpret this as reordering.
2027 * The only another reason could be bug in receiver TCP.
2029 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
2031 struct tcp_sock *tp = tcp_sk(sk);
2033 if (!tcp_limit_reno_sacked(tp))
2036 tp->reordering = min_t(u32, tp->packets_out + addend,
2037 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_max_reordering));
2039 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRENOREORDER);
2042 /* Emulate SACKs for SACKless connection: account for a new dupack. */
2044 static void tcp_add_reno_sack(struct sock *sk, int num_dupack, bool ece_ack)
2047 struct tcp_sock *tp = tcp_sk(sk);
2048 u32 prior_sacked = tp->sacked_out;
2051 tp->sacked_out += num_dupack;
2052 tcp_check_reno_reordering(sk, 0);
2053 delivered = tp->sacked_out - prior_sacked;
2055 tcp_count_delivered(tp, delivered, ece_ack);
2056 tcp_verify_left_out(tp);
2060 /* Account for ACK, ACKing some data in Reno Recovery phase. */
2062 static void tcp_remove_reno_sacks(struct sock *sk, int acked, bool ece_ack)
2064 struct tcp_sock *tp = tcp_sk(sk);
2067 /* One ACK acked hole. The rest eat duplicate ACKs. */
2068 tcp_count_delivered(tp, max_t(int, acked - tp->sacked_out, 1),
2070 if (acked - 1 >= tp->sacked_out)
2073 tp->sacked_out -= acked - 1;
2075 tcp_check_reno_reordering(sk, acked);
2076 tcp_verify_left_out(tp);
2079 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
2084 void tcp_clear_retrans(struct tcp_sock *tp)
2086 tp->retrans_out = 0;
2088 tp->undo_marker = 0;
2089 tp->undo_retrans = -1;
2093 static inline void tcp_init_undo(struct tcp_sock *tp)
2095 tp->undo_marker = tp->snd_una;
2096 /* Retransmission still in flight may cause DSACKs later. */
2097 tp->undo_retrans = tp->retrans_out ? : -1;
2100 static bool tcp_is_rack(const struct sock *sk)
2102 return READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_recovery) &
2103 TCP_RACK_LOSS_DETECTION;
2106 /* If we detect SACK reneging, forget all SACK information
2107 * and reset tags completely, otherwise preserve SACKs. If receiver
2108 * dropped its ofo queue, we will know this due to reneging detection.
2110 static void tcp_timeout_mark_lost(struct sock *sk)
2112 struct tcp_sock *tp = tcp_sk(sk);
2113 struct sk_buff *skb, *head;
2114 bool is_reneg; /* is receiver reneging on SACKs? */
2116 head = tcp_rtx_queue_head(sk);
2117 is_reneg = head && (TCP_SKB_CB(head)->sacked & TCPCB_SACKED_ACKED);
2119 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
2121 /* Mark SACK reneging until we recover from this loss event. */
2122 tp->is_sack_reneg = 1;
2123 } else if (tcp_is_reno(tp)) {
2124 tcp_reset_reno_sack(tp);
2128 skb_rbtree_walk_from(skb) {
2130 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
2131 else if (tcp_is_rack(sk) && skb != head &&
2132 tcp_rack_skb_timeout(tp, skb, 0) > 0)
2133 continue; /* Don't mark recently sent ones lost yet */
2134 tcp_mark_skb_lost(sk, skb);
2136 tcp_verify_left_out(tp);
2137 tcp_clear_all_retrans_hints(tp);
2140 /* Enter Loss state. */
2141 void tcp_enter_loss(struct sock *sk)
2143 const struct inet_connection_sock *icsk = inet_csk(sk);
2144 struct tcp_sock *tp = tcp_sk(sk);
2145 struct net *net = sock_net(sk);
2146 bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery;
2149 tcp_timeout_mark_lost(sk);
2151 /* Reduce ssthresh if it has not yet been made inside this window. */
2152 if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
2153 !after(tp->high_seq, tp->snd_una) ||
2154 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
2155 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2156 tp->prior_cwnd = tcp_snd_cwnd(tp);
2157 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2158 tcp_ca_event(sk, CA_EVENT_LOSS);
2161 tcp_snd_cwnd_set(tp, tcp_packets_in_flight(tp) + 1);
2162 tp->snd_cwnd_cnt = 0;
2163 tp->snd_cwnd_stamp = tcp_jiffies32;
2165 /* Timeout in disordered state after receiving substantial DUPACKs
2166 * suggests that the degree of reordering is over-estimated.
2168 reordering = READ_ONCE(net->ipv4.sysctl_tcp_reordering);
2169 if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
2170 tp->sacked_out >= reordering)
2171 tp->reordering = min_t(unsigned int, tp->reordering,
2174 tcp_set_ca_state(sk, TCP_CA_Loss);
2175 tp->high_seq = tp->snd_nxt;
2176 tcp_ecn_queue_cwr(tp);
2178 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
2179 * loss recovery is underway except recurring timeout(s) on
2180 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
2182 tp->frto = READ_ONCE(net->ipv4.sysctl_tcp_frto) &&
2183 (new_recovery || icsk->icsk_retransmits) &&
2184 !inet_csk(sk)->icsk_mtup.probe_size;
2187 /* If ACK arrived pointing to a remembered SACK, it means that our
2188 * remembered SACKs do not reflect real state of receiver i.e.
2189 * receiver _host_ is heavily congested (or buggy).
2191 * To avoid big spurious retransmission bursts due to transient SACK
2192 * scoreboard oddities that look like reneging, we give the receiver a
2193 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
2194 * restore sanity to the SACK scoreboard. If the apparent reneging
2195 * persists until this RTO then we'll clear the SACK scoreboard.
2197 static bool tcp_check_sack_reneging(struct sock *sk, int flag)
2199 if (flag & FLAG_SACK_RENEGING &&
2200 flag & FLAG_SND_UNA_ADVANCED) {
2201 struct tcp_sock *tp = tcp_sk(sk);
2202 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4),
2203 msecs_to_jiffies(10));
2205 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2206 delay, TCP_RTO_MAX);
2212 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2213 * counter when SACK is enabled (without SACK, sacked_out is used for
2216 * With reordering, holes may still be in flight, so RFC3517 recovery
2217 * uses pure sacked_out (total number of SACKed segments) even though
2218 * it violates the RFC that uses duplicate ACKs, often these are equal
2219 * but when e.g. out-of-window ACKs or packet duplication occurs,
2220 * they differ. Since neither occurs due to loss, TCP should really
2223 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2225 return tp->sacked_out + 1;
2228 /* Linux NewReno/SACK/ECN state machine.
2229 * --------------------------------------
2231 * "Open" Normal state, no dubious events, fast path.
2232 * "Disorder" In all the respects it is "Open",
2233 * but requires a bit more attention. It is entered when
2234 * we see some SACKs or dupacks. It is split of "Open"
2235 * mainly to move some processing from fast path to slow one.
2236 * "CWR" CWND was reduced due to some Congestion Notification event.
2237 * It can be ECN, ICMP source quench, local device congestion.
2238 * "Recovery" CWND was reduced, we are fast-retransmitting.
2239 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2241 * tcp_fastretrans_alert() is entered:
2242 * - each incoming ACK, if state is not "Open"
2243 * - when arrived ACK is unusual, namely:
2248 * Counting packets in flight is pretty simple.
2250 * in_flight = packets_out - left_out + retrans_out
2252 * packets_out is SND.NXT-SND.UNA counted in packets.
2254 * retrans_out is number of retransmitted segments.
2256 * left_out is number of segments left network, but not ACKed yet.
2258 * left_out = sacked_out + lost_out
2260 * sacked_out: Packets, which arrived to receiver out of order
2261 * and hence not ACKed. With SACKs this number is simply
2262 * amount of SACKed data. Even without SACKs
2263 * it is easy to give pretty reliable estimate of this number,
2264 * counting duplicate ACKs.
2266 * lost_out: Packets lost by network. TCP has no explicit
2267 * "loss notification" feedback from network (for now).
2268 * It means that this number can be only _guessed_.
2269 * Actually, it is the heuristics to predict lossage that
2270 * distinguishes different algorithms.
2272 * F.e. after RTO, when all the queue is considered as lost,
2273 * lost_out = packets_out and in_flight = retrans_out.
2275 * Essentially, we have now a few algorithms detecting
2278 * If the receiver supports SACK:
2280 * RFC6675/3517: It is the conventional algorithm. A packet is
2281 * considered lost if the number of higher sequence packets
2282 * SACKed is greater than or equal the DUPACK thoreshold
2283 * (reordering). This is implemented in tcp_mark_head_lost and
2284 * tcp_update_scoreboard.
2286 * RACK (draft-ietf-tcpm-rack-01): it is a newer algorithm
2287 * (2017-) that checks timing instead of counting DUPACKs.
2288 * Essentially a packet is considered lost if it's not S/ACKed
2289 * after RTT + reordering_window, where both metrics are
2290 * dynamically measured and adjusted. This is implemented in
2291 * tcp_rack_mark_lost.
2293 * If the receiver does not support SACK:
2295 * NewReno (RFC6582): in Recovery we assume that one segment
2296 * is lost (classic Reno). While we are in Recovery and
2297 * a partial ACK arrives, we assume that one more packet
2298 * is lost (NewReno). This heuristics are the same in NewReno
2301 * Really tricky (and requiring careful tuning) part of algorithm
2302 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2303 * The first determines the moment _when_ we should reduce CWND and,
2304 * hence, slow down forward transmission. In fact, it determines the moment
2305 * when we decide that hole is caused by loss, rather than by a reorder.
2307 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2308 * holes, caused by lost packets.
2310 * And the most logically complicated part of algorithm is undo
2311 * heuristics. We detect false retransmits due to both too early
2312 * fast retransmit (reordering) and underestimated RTO, analyzing
2313 * timestamps and D-SACKs. When we detect that some segments were
2314 * retransmitted by mistake and CWND reduction was wrong, we undo
2315 * window reduction and abort recovery phase. This logic is hidden
2316 * inside several functions named tcp_try_undo_<something>.
2319 /* This function decides, when we should leave Disordered state
2320 * and enter Recovery phase, reducing congestion window.
2322 * Main question: may we further continue forward transmission
2323 * with the same cwnd?
2325 static bool tcp_time_to_recover(struct sock *sk, int flag)
2327 struct tcp_sock *tp = tcp_sk(sk);
2329 /* Trick#1: The loss is proven. */
2333 /* Not-A-Trick#2 : Classic rule... */
2334 if (!tcp_is_rack(sk) && tcp_dupack_heuristics(tp) > tp->reordering)
2340 /* Detect loss in event "A" above by marking head of queue up as lost.
2341 * For RFC3517 SACK, a segment is considered lost if it
2342 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2343 * the maximum SACKed segments to pass before reaching this limit.
2345 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2347 struct tcp_sock *tp = tcp_sk(sk);
2348 struct sk_buff *skb;
2350 /* Use SACK to deduce losses of new sequences sent during recovery */
2351 const u32 loss_high = tp->snd_nxt;
2353 WARN_ON(packets > tp->packets_out);
2354 skb = tp->lost_skb_hint;
2356 /* Head already handled? */
2357 if (mark_head && after(TCP_SKB_CB(skb)->seq, tp->snd_una))
2359 cnt = tp->lost_cnt_hint;
2361 skb = tcp_rtx_queue_head(sk);
2365 skb_rbtree_walk_from(skb) {
2366 /* TODO: do this better */
2367 /* this is not the most efficient way to do this... */
2368 tp->lost_skb_hint = skb;
2369 tp->lost_cnt_hint = cnt;
2371 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2374 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2375 cnt += tcp_skb_pcount(skb);
2380 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_LOST))
2381 tcp_mark_skb_lost(sk, skb);
2386 tcp_verify_left_out(tp);
2389 /* Account newly detected lost packet(s) */
2391 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2393 struct tcp_sock *tp = tcp_sk(sk);
2395 if (tcp_is_sack(tp)) {
2396 int sacked_upto = tp->sacked_out - tp->reordering;
2397 if (sacked_upto >= 0)
2398 tcp_mark_head_lost(sk, sacked_upto, 0);
2399 else if (fast_rexmit)
2400 tcp_mark_head_lost(sk, 1, 1);
2404 static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when)
2406 return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2407 before(tp->rx_opt.rcv_tsecr, when);
2410 /* skb is spurious retransmitted if the returned timestamp echo
2411 * reply is prior to the skb transmission time
2413 static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp,
2414 const struct sk_buff *skb)
2416 return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) &&
2417 tcp_tsopt_ecr_before(tp, tcp_skb_timestamp(skb));
2420 /* Nothing was retransmitted or returned timestamp is less
2421 * than timestamp of the first retransmission.
2423 static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2425 return tp->retrans_stamp &&
2426 tcp_tsopt_ecr_before(tp, tp->retrans_stamp);
2429 /* Undo procedures. */
2431 /* We can clear retrans_stamp when there are no retransmissions in the
2432 * window. It would seem that it is trivially available for us in
2433 * tp->retrans_out, however, that kind of assumptions doesn't consider
2434 * what will happen if errors occur when sending retransmission for the
2435 * second time. ...It could the that such segment has only
2436 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2437 * the head skb is enough except for some reneging corner cases that
2438 * are not worth the effort.
2440 * Main reason for all this complexity is the fact that connection dying
2441 * time now depends on the validity of the retrans_stamp, in particular,
2442 * that successive retransmissions of a segment must not advance
2443 * retrans_stamp under any conditions.
2445 static bool tcp_any_retrans_done(const struct sock *sk)
2447 const struct tcp_sock *tp = tcp_sk(sk);
2448 struct sk_buff *skb;
2450 if (tp->retrans_out)
2453 skb = tcp_rtx_queue_head(sk);
2454 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2460 static void DBGUNDO(struct sock *sk, const char *msg)
2462 #if FASTRETRANS_DEBUG > 1
2463 struct tcp_sock *tp = tcp_sk(sk);
2464 struct inet_sock *inet = inet_sk(sk);
2466 if (sk->sk_family == AF_INET) {
2467 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2469 &inet->inet_daddr, ntohs(inet->inet_dport),
2470 tcp_snd_cwnd(tp), tcp_left_out(tp),
2471 tp->snd_ssthresh, tp->prior_ssthresh,
2474 #if IS_ENABLED(CONFIG_IPV6)
2475 else if (sk->sk_family == AF_INET6) {
2476 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2478 &sk->sk_v6_daddr, ntohs(inet->inet_dport),
2479 tcp_snd_cwnd(tp), tcp_left_out(tp),
2480 tp->snd_ssthresh, tp->prior_ssthresh,
2487 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
2489 struct tcp_sock *tp = tcp_sk(sk);
2492 struct sk_buff *skb;
2494 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2495 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2498 tcp_clear_all_retrans_hints(tp);
2501 if (tp->prior_ssthresh) {
2502 const struct inet_connection_sock *icsk = inet_csk(sk);
2504 tcp_snd_cwnd_set(tp, icsk->icsk_ca_ops->undo_cwnd(sk));
2506 if (tp->prior_ssthresh > tp->snd_ssthresh) {
2507 tp->snd_ssthresh = tp->prior_ssthresh;
2508 tcp_ecn_withdraw_cwr(tp);
2511 tp->snd_cwnd_stamp = tcp_jiffies32;
2512 tp->undo_marker = 0;
2513 tp->rack.advanced = 1; /* Force RACK to re-exam losses */
2516 static inline bool tcp_may_undo(const struct tcp_sock *tp)
2518 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2521 static bool tcp_is_non_sack_preventing_reopen(struct sock *sk)
2523 struct tcp_sock *tp = tcp_sk(sk);
2525 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2526 /* Hold old state until something *above* high_seq
2527 * is ACKed. For Reno it is MUST to prevent false
2528 * fast retransmits (RFC2582). SACK TCP is safe. */
2529 if (!tcp_any_retrans_done(sk))
2530 tp->retrans_stamp = 0;
2536 /* People celebrate: "We love our President!" */
2537 static bool tcp_try_undo_recovery(struct sock *sk)
2539 struct tcp_sock *tp = tcp_sk(sk);
2541 if (tcp_may_undo(tp)) {
2544 /* Happy end! We did not retransmit anything
2545 * or our original transmission succeeded.
2547 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2548 tcp_undo_cwnd_reduction(sk, false);
2549 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2550 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2552 mib_idx = LINUX_MIB_TCPFULLUNDO;
2554 NET_INC_STATS(sock_net(sk), mib_idx);
2555 } else if (tp->rack.reo_wnd_persist) {
2556 tp->rack.reo_wnd_persist--;
2558 if (tcp_is_non_sack_preventing_reopen(sk))
2560 tcp_set_ca_state(sk, TCP_CA_Open);
2561 tp->is_sack_reneg = 0;
2565 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2566 static bool tcp_try_undo_dsack(struct sock *sk)
2568 struct tcp_sock *tp = tcp_sk(sk);
2570 if (tp->undo_marker && !tp->undo_retrans) {
2571 tp->rack.reo_wnd_persist = min(TCP_RACK_RECOVERY_THRESH,
2572 tp->rack.reo_wnd_persist + 1);
2573 DBGUNDO(sk, "D-SACK");
2574 tcp_undo_cwnd_reduction(sk, false);
2575 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2581 /* Undo during loss recovery after partial ACK or using F-RTO. */
2582 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2584 struct tcp_sock *tp = tcp_sk(sk);
2586 if (frto_undo || tcp_may_undo(tp)) {
2587 tcp_undo_cwnd_reduction(sk, true);
2589 DBGUNDO(sk, "partial loss");
2590 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2592 NET_INC_STATS(sock_net(sk),
2593 LINUX_MIB_TCPSPURIOUSRTOS);
2594 inet_csk(sk)->icsk_retransmits = 0;
2595 if (tcp_is_non_sack_preventing_reopen(sk))
2597 if (frto_undo || tcp_is_sack(tp)) {
2598 tcp_set_ca_state(sk, TCP_CA_Open);
2599 tp->is_sack_reneg = 0;
2606 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
2607 * It computes the number of packets to send (sndcnt) based on packets newly
2609 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2610 * cwnd reductions across a full RTT.
2611 * 2) Otherwise PRR uses packet conservation to send as much as delivered.
2612 * But when SND_UNA is acked without further losses,
2613 * slow starts cwnd up to ssthresh to speed up the recovery.
2615 static void tcp_init_cwnd_reduction(struct sock *sk)
2617 struct tcp_sock *tp = tcp_sk(sk);
2619 tp->high_seq = tp->snd_nxt;
2620 tp->tlp_high_seq = 0;
2621 tp->snd_cwnd_cnt = 0;
2622 tp->prior_cwnd = tcp_snd_cwnd(tp);
2623 tp->prr_delivered = 0;
2625 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2626 tcp_ecn_queue_cwr(tp);
2629 void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked, int newly_lost, int flag)
2631 struct tcp_sock *tp = tcp_sk(sk);
2633 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2635 if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd))
2638 tp->prr_delivered += newly_acked_sacked;
2640 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2642 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2644 sndcnt = max_t(int, tp->prr_delivered - tp->prr_out,
2645 newly_acked_sacked);
2646 if (flag & FLAG_SND_UNA_ADVANCED && !newly_lost)
2648 sndcnt = min(delta, sndcnt);
2650 /* Force a fast retransmit upon entering fast recovery */
2651 sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1));
2652 tcp_snd_cwnd_set(tp, tcp_packets_in_flight(tp) + sndcnt);
2655 static inline void tcp_end_cwnd_reduction(struct sock *sk)
2657 struct tcp_sock *tp = tcp_sk(sk);
2659 if (inet_csk(sk)->icsk_ca_ops->cong_control)
2662 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2663 if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH &&
2664 (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR || tp->undo_marker)) {
2665 tcp_snd_cwnd_set(tp, tp->snd_ssthresh);
2666 tp->snd_cwnd_stamp = tcp_jiffies32;
2668 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2671 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2672 void tcp_enter_cwr(struct sock *sk)
2674 struct tcp_sock *tp = tcp_sk(sk);
2676 tp->prior_ssthresh = 0;
2677 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2678 tp->undo_marker = 0;
2679 tcp_init_cwnd_reduction(sk);
2680 tcp_set_ca_state(sk, TCP_CA_CWR);
2683 EXPORT_SYMBOL(tcp_enter_cwr);
2685 static void tcp_try_keep_open(struct sock *sk)
2687 struct tcp_sock *tp = tcp_sk(sk);
2688 int state = TCP_CA_Open;
2690 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2691 state = TCP_CA_Disorder;
2693 if (inet_csk(sk)->icsk_ca_state != state) {
2694 tcp_set_ca_state(sk, state);
2695 tp->high_seq = tp->snd_nxt;
2699 static void tcp_try_to_open(struct sock *sk, int flag)
2701 struct tcp_sock *tp = tcp_sk(sk);
2703 tcp_verify_left_out(tp);
2705 if (!tcp_any_retrans_done(sk))
2706 tp->retrans_stamp = 0;
2708 if (flag & FLAG_ECE)
2711 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2712 tcp_try_keep_open(sk);
2716 static void tcp_mtup_probe_failed(struct sock *sk)
2718 struct inet_connection_sock *icsk = inet_csk(sk);
2720 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2721 icsk->icsk_mtup.probe_size = 0;
2722 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPFAIL);
2725 static void tcp_mtup_probe_success(struct sock *sk)
2727 struct tcp_sock *tp = tcp_sk(sk);
2728 struct inet_connection_sock *icsk = inet_csk(sk);
2731 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2733 val = (u64)tcp_snd_cwnd(tp) * tcp_mss_to_mtu(sk, tp->mss_cache);
2734 do_div(val, icsk->icsk_mtup.probe_size);
2735 DEBUG_NET_WARN_ON_ONCE((u32)val != val);
2736 tcp_snd_cwnd_set(tp, max_t(u32, 1U, val));
2738 tp->snd_cwnd_cnt = 0;
2739 tp->snd_cwnd_stamp = tcp_jiffies32;
2740 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2742 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2743 icsk->icsk_mtup.probe_size = 0;
2744 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2745 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS);
2748 /* Do a simple retransmit without using the backoff mechanisms in
2749 * tcp_timer. This is used for path mtu discovery.
2750 * The socket is already locked here.
2752 void tcp_simple_retransmit(struct sock *sk)
2754 const struct inet_connection_sock *icsk = inet_csk(sk);
2755 struct tcp_sock *tp = tcp_sk(sk);
2756 struct sk_buff *skb;
2759 /* A fastopen SYN request is stored as two separate packets within
2760 * the retransmit queue, this is done by tcp_send_syn_data().
2761 * As a result simply checking the MSS of the frames in the queue
2762 * will not work for the SYN packet.
2764 * Us being here is an indication of a path MTU issue so we can
2765 * assume that the fastopen SYN was lost and just mark all the
2766 * frames in the retransmit queue as lost. We will use an MSS of
2767 * -1 to mark all frames as lost, otherwise compute the current MSS.
2769 if (tp->syn_data && sk->sk_state == TCP_SYN_SENT)
2772 mss = tcp_current_mss(sk);
2774 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2775 if (tcp_skb_seglen(skb) > mss)
2776 tcp_mark_skb_lost(sk, skb);
2779 tcp_clear_retrans_hints_partial(tp);
2784 if (tcp_is_reno(tp))
2785 tcp_limit_reno_sacked(tp);
2787 tcp_verify_left_out(tp);
2789 /* Don't muck with the congestion window here.
2790 * Reason is that we do not increase amount of _data_
2791 * in network, but units changed and effective
2792 * cwnd/ssthresh really reduced now.
2794 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2795 tp->high_seq = tp->snd_nxt;
2796 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2797 tp->prior_ssthresh = 0;
2798 tp->undo_marker = 0;
2799 tcp_set_ca_state(sk, TCP_CA_Loss);
2801 tcp_xmit_retransmit_queue(sk);
2803 EXPORT_SYMBOL(tcp_simple_retransmit);
2805 void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2807 struct tcp_sock *tp = tcp_sk(sk);
2810 if (tcp_is_reno(tp))
2811 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2813 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2815 NET_INC_STATS(sock_net(sk), mib_idx);
2817 tp->prior_ssthresh = 0;
2820 if (!tcp_in_cwnd_reduction(sk)) {
2822 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2823 tcp_init_cwnd_reduction(sk);
2825 tcp_set_ca_state(sk, TCP_CA_Recovery);
2828 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2829 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2831 static void tcp_process_loss(struct sock *sk, int flag, int num_dupack,
2834 struct tcp_sock *tp = tcp_sk(sk);
2835 bool recovered = !before(tp->snd_una, tp->high_seq);
2837 if ((flag & FLAG_SND_UNA_ADVANCED || rcu_access_pointer(tp->fastopen_rsk)) &&
2838 tcp_try_undo_loss(sk, false))
2841 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2842 /* Step 3.b. A timeout is spurious if not all data are
2843 * lost, i.e., never-retransmitted data are (s)acked.
2845 if ((flag & FLAG_ORIG_SACK_ACKED) &&
2846 tcp_try_undo_loss(sk, true))
2849 if (after(tp->snd_nxt, tp->high_seq)) {
2850 if (flag & FLAG_DATA_SACKED || num_dupack)
2851 tp->frto = 0; /* Step 3.a. loss was real */
2852 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2853 tp->high_seq = tp->snd_nxt;
2854 /* Step 2.b. Try send new data (but deferred until cwnd
2855 * is updated in tcp_ack()). Otherwise fall back to
2856 * the conventional recovery.
2858 if (!tcp_write_queue_empty(sk) &&
2859 after(tcp_wnd_end(tp), tp->snd_nxt)) {
2860 *rexmit = REXMIT_NEW;
2868 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2869 tcp_try_undo_recovery(sk);
2872 if (tcp_is_reno(tp)) {
2873 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2874 * delivered. Lower inflight to clock out (re)transmissions.
2876 if (after(tp->snd_nxt, tp->high_seq) && num_dupack)
2877 tcp_add_reno_sack(sk, num_dupack, flag & FLAG_ECE);
2878 else if (flag & FLAG_SND_UNA_ADVANCED)
2879 tcp_reset_reno_sack(tp);
2881 *rexmit = REXMIT_LOST;
2884 static bool tcp_force_fast_retransmit(struct sock *sk)
2886 struct tcp_sock *tp = tcp_sk(sk);
2888 return after(tcp_highest_sack_seq(tp),
2889 tp->snd_una + tp->reordering * tp->mss_cache);
2892 /* Undo during fast recovery after partial ACK. */
2893 static bool tcp_try_undo_partial(struct sock *sk, u32 prior_snd_una,
2896 struct tcp_sock *tp = tcp_sk(sk);
2898 if (tp->undo_marker && tcp_packet_delayed(tp)) {
2899 /* Plain luck! Hole if filled with delayed
2900 * packet, rather than with a retransmit. Check reordering.
2902 tcp_check_sack_reordering(sk, prior_snd_una, 1);
2904 /* We are getting evidence that the reordering degree is higher
2905 * than we realized. If there are no retransmits out then we
2906 * can undo. Otherwise we clock out new packets but do not
2907 * mark more packets lost or retransmit more.
2909 if (tp->retrans_out)
2912 if (!tcp_any_retrans_done(sk))
2913 tp->retrans_stamp = 0;
2915 DBGUNDO(sk, "partial recovery");
2916 tcp_undo_cwnd_reduction(sk, true);
2917 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2918 tcp_try_keep_open(sk);
2920 /* Partial ACK arrived. Force fast retransmit. */
2921 *do_lost = tcp_force_fast_retransmit(sk);
2926 static void tcp_identify_packet_loss(struct sock *sk, int *ack_flag)
2928 struct tcp_sock *tp = tcp_sk(sk);
2930 if (tcp_rtx_queue_empty(sk))
2933 if (unlikely(tcp_is_reno(tp))) {
2934 tcp_newreno_mark_lost(sk, *ack_flag & FLAG_SND_UNA_ADVANCED);
2935 } else if (tcp_is_rack(sk)) {
2936 u32 prior_retrans = tp->retrans_out;
2938 if (tcp_rack_mark_lost(sk))
2939 *ack_flag &= ~FLAG_SET_XMIT_TIMER;
2940 if (prior_retrans > tp->retrans_out)
2941 *ack_flag |= FLAG_LOST_RETRANS;
2945 /* Process an event, which can update packets-in-flight not trivially.
2946 * Main goal of this function is to calculate new estimate for left_out,
2947 * taking into account both packets sitting in receiver's buffer and
2948 * packets lost by network.
2950 * Besides that it updates the congestion state when packet loss or ECN
2951 * is detected. But it does not reduce the cwnd, it is done by the
2952 * congestion control later.
2954 * It does _not_ decide what to send, it is made in function
2955 * tcp_xmit_retransmit_queue().
2957 static void tcp_fastretrans_alert(struct sock *sk, const u32 prior_snd_una,
2958 int num_dupack, int *ack_flag, int *rexmit)
2960 struct inet_connection_sock *icsk = inet_csk(sk);
2961 struct tcp_sock *tp = tcp_sk(sk);
2962 int fast_rexmit = 0, flag = *ack_flag;
2963 bool ece_ack = flag & FLAG_ECE;
2964 bool do_lost = num_dupack || ((flag & FLAG_DATA_SACKED) &&
2965 tcp_force_fast_retransmit(sk));
2967 if (!tp->packets_out && tp->sacked_out)
2970 /* Now state machine starts.
2971 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2973 tp->prior_ssthresh = 0;
2975 /* B. In all the states check for reneging SACKs. */
2976 if (tcp_check_sack_reneging(sk, flag))
2979 /* C. Check consistency of the current state. */
2980 tcp_verify_left_out(tp);
2982 /* D. Check state exit conditions. State can be terminated
2983 * when high_seq is ACKed. */
2984 if (icsk->icsk_ca_state == TCP_CA_Open) {
2985 WARN_ON(tp->retrans_out != 0 && !tp->syn_data);
2986 tp->retrans_stamp = 0;
2987 } else if (!before(tp->snd_una, tp->high_seq)) {
2988 switch (icsk->icsk_ca_state) {
2990 /* CWR is to be held something *above* high_seq
2991 * is ACKed for CWR bit to reach receiver. */
2992 if (tp->snd_una != tp->high_seq) {
2993 tcp_end_cwnd_reduction(sk);
2994 tcp_set_ca_state(sk, TCP_CA_Open);
2998 case TCP_CA_Recovery:
2999 if (tcp_is_reno(tp))
3000 tcp_reset_reno_sack(tp);
3001 if (tcp_try_undo_recovery(sk))
3003 tcp_end_cwnd_reduction(sk);
3008 /* E. Process state. */
3009 switch (icsk->icsk_ca_state) {
3010 case TCP_CA_Recovery:
3011 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
3012 if (tcp_is_reno(tp))
3013 tcp_add_reno_sack(sk, num_dupack, ece_ack);
3014 } else if (tcp_try_undo_partial(sk, prior_snd_una, &do_lost))
3017 if (tcp_try_undo_dsack(sk))
3018 tcp_try_keep_open(sk);
3020 tcp_identify_packet_loss(sk, ack_flag);
3021 if (icsk->icsk_ca_state != TCP_CA_Recovery) {
3022 if (!tcp_time_to_recover(sk, flag))
3024 /* Undo reverts the recovery state. If loss is evident,
3025 * starts a new recovery (e.g. reordering then loss);
3027 tcp_enter_recovery(sk, ece_ack);
3031 tcp_process_loss(sk, flag, num_dupack, rexmit);
3032 tcp_identify_packet_loss(sk, ack_flag);
3033 if (!(icsk->icsk_ca_state == TCP_CA_Open ||
3034 (*ack_flag & FLAG_LOST_RETRANS)))
3036 /* Change state if cwnd is undone or retransmits are lost */
3039 if (tcp_is_reno(tp)) {
3040 if (flag & FLAG_SND_UNA_ADVANCED)
3041 tcp_reset_reno_sack(tp);
3042 tcp_add_reno_sack(sk, num_dupack, ece_ack);
3045 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
3046 tcp_try_undo_dsack(sk);
3048 tcp_identify_packet_loss(sk, ack_flag);
3049 if (!tcp_time_to_recover(sk, flag)) {
3050 tcp_try_to_open(sk, flag);
3054 /* MTU probe failure: don't reduce cwnd */
3055 if (icsk->icsk_ca_state < TCP_CA_CWR &&
3056 icsk->icsk_mtup.probe_size &&
3057 tp->snd_una == tp->mtu_probe.probe_seq_start) {
3058 tcp_mtup_probe_failed(sk);
3059 /* Restores the reduction we did in tcp_mtup_probe() */
3060 tcp_snd_cwnd_set(tp, tcp_snd_cwnd(tp) + 1);
3061 tcp_simple_retransmit(sk);
3065 /* Otherwise enter Recovery state */
3066 tcp_enter_recovery(sk, ece_ack);
3070 if (!tcp_is_rack(sk) && do_lost)
3071 tcp_update_scoreboard(sk, fast_rexmit);
3072 *rexmit = REXMIT_LOST;
3075 static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us, const int flag)
3077 u32 wlen = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_rtt_wlen) * HZ;
3078 struct tcp_sock *tp = tcp_sk(sk);
3080 if ((flag & FLAG_ACK_MAYBE_DELAYED) && rtt_us > tcp_min_rtt(tp)) {
3081 /* If the remote keeps returning delayed ACKs, eventually
3082 * the min filter would pick it up and overestimate the
3083 * prop. delay when it expires. Skip suspected delayed ACKs.
3087 minmax_running_min(&tp->rtt_min, wlen, tcp_jiffies32,
3088 rtt_us ? : jiffies_to_usecs(1));
3091 static bool tcp_ack_update_rtt(struct sock *sk, const int flag,
3092 long seq_rtt_us, long sack_rtt_us,
3093 long ca_rtt_us, struct rate_sample *rs)
3095 const struct tcp_sock *tp = tcp_sk(sk);
3097 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
3098 * broken middle-boxes or peers may corrupt TS-ECR fields. But
3099 * Karn's algorithm forbids taking RTT if some retransmitted data
3100 * is acked (RFC6298).
3103 seq_rtt_us = sack_rtt_us;
3105 /* RTTM Rule: A TSecr value received in a segment is used to
3106 * update the averaged RTT measurement only if the segment
3107 * acknowledges some new data, i.e., only if it advances the
3108 * left edge of the send window.
3109 * See draft-ietf-tcplw-high-performance-00, section 3.3.
3111 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
3112 flag & FLAG_ACKED) {
3113 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr;
3115 if (likely(delta < INT_MAX / (USEC_PER_SEC / TCP_TS_HZ))) {
3118 seq_rtt_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
3119 ca_rtt_us = seq_rtt_us;
3122 rs->rtt_us = ca_rtt_us; /* RTT of last (S)ACKed packet (or -1) */
3126 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
3127 * always taken together with ACK, SACK, or TS-opts. Any negative
3128 * values will be skipped with the seq_rtt_us < 0 check above.
3130 tcp_update_rtt_min(sk, ca_rtt_us, flag);
3131 tcp_rtt_estimator(sk, seq_rtt_us);
3134 /* RFC6298: only reset backoff on valid RTT measurement. */
3135 inet_csk(sk)->icsk_backoff = 0;
3139 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
3140 void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req)
3142 struct rate_sample rs;
3145 if (req && !req->num_retrans && tcp_rsk(req)->snt_synack)
3146 rtt_us = tcp_stamp_us_delta(tcp_clock_us(), tcp_rsk(req)->snt_synack);
3148 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us, &rs);
3152 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
3154 const struct inet_connection_sock *icsk = inet_csk(sk);
3156 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
3157 tcp_sk(sk)->snd_cwnd_stamp = tcp_jiffies32;
3160 /* Restart timer after forward progress on connection.
3161 * RFC2988 recommends to restart timer to now+rto.
3163 void tcp_rearm_rto(struct sock *sk)
3165 const struct inet_connection_sock *icsk = inet_csk(sk);
3166 struct tcp_sock *tp = tcp_sk(sk);
3168 /* If the retrans timer is currently being used by Fast Open
3169 * for SYN-ACK retrans purpose, stay put.
3171 if (rcu_access_pointer(tp->fastopen_rsk))
3174 if (!tp->packets_out) {
3175 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
3177 u32 rto = inet_csk(sk)->icsk_rto;
3178 /* Offset the time elapsed after installing regular RTO */
3179 if (icsk->icsk_pending == ICSK_TIME_REO_TIMEOUT ||
3180 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
3181 s64 delta_us = tcp_rto_delta_us(sk);
3182 /* delta_us may not be positive if the socket is locked
3183 * when the retrans timer fires and is rescheduled.
3185 rto = usecs_to_jiffies(max_t(int, delta_us, 1));
3187 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
3192 /* Try to schedule a loss probe; if that doesn't work, then schedule an RTO. */
3193 static void tcp_set_xmit_timer(struct sock *sk)
3195 if (!tcp_schedule_loss_probe(sk, true))
3199 /* If we get here, the whole TSO packet has not been acked. */
3200 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3202 struct tcp_sock *tp = tcp_sk(sk);
3205 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3207 packets_acked = tcp_skb_pcount(skb);
3208 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3210 packets_acked -= tcp_skb_pcount(skb);
3212 if (packets_acked) {
3213 BUG_ON(tcp_skb_pcount(skb) == 0);
3214 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3217 return packets_acked;
3220 static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
3221 const struct sk_buff *ack_skb, u32 prior_snd_una)
3223 const struct skb_shared_info *shinfo;
3225 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3226 if (likely(!TCP_SKB_CB(skb)->txstamp_ack))
3229 shinfo = skb_shinfo(skb);
3230 if (!before(shinfo->tskey, prior_snd_una) &&
3231 before(shinfo->tskey, tcp_sk(sk)->snd_una)) {
3232 tcp_skb_tsorted_save(skb) {
3233 __skb_tstamp_tx(skb, ack_skb, NULL, sk, SCM_TSTAMP_ACK);
3234 } tcp_skb_tsorted_restore(skb);
3238 /* Remove acknowledged frames from the retransmission queue. If our packet
3239 * is before the ack sequence we can discard it as it's confirmed to have
3240 * arrived at the other end.
3242 static int tcp_clean_rtx_queue(struct sock *sk, const struct sk_buff *ack_skb,
3243 u32 prior_fack, u32 prior_snd_una,
3244 struct tcp_sacktag_state *sack, bool ece_ack)
3246 const struct inet_connection_sock *icsk = inet_csk(sk);
3247 u64 first_ackt, last_ackt;
3248 struct tcp_sock *tp = tcp_sk(sk);
3249 u32 prior_sacked = tp->sacked_out;
3250 u32 reord = tp->snd_nxt; /* lowest acked un-retx un-sacked seq */
3251 struct sk_buff *skb, *next;
3252 bool fully_acked = true;
3253 long sack_rtt_us = -1L;
3254 long seq_rtt_us = -1L;
3255 long ca_rtt_us = -1L;
3262 for (skb = skb_rb_first(&sk->tcp_rtx_queue); skb; skb = next) {
3263 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3264 const u32 start_seq = scb->seq;
3265 u8 sacked = scb->sacked;
3268 /* Determine how many packets and what bytes were acked, tso and else */
3269 if (after(scb->end_seq, tp->snd_una)) {
3270 if (tcp_skb_pcount(skb) == 1 ||
3271 !after(tp->snd_una, scb->seq))
3274 acked_pcount = tcp_tso_acked(sk, skb);
3277 fully_acked = false;
3279 acked_pcount = tcp_skb_pcount(skb);
3282 if (unlikely(sacked & TCPCB_RETRANS)) {
3283 if (sacked & TCPCB_SACKED_RETRANS)
3284 tp->retrans_out -= acked_pcount;
3285 flag |= FLAG_RETRANS_DATA_ACKED;
3286 } else if (!(sacked & TCPCB_SACKED_ACKED)) {
3287 last_ackt = tcp_skb_timestamp_us(skb);
3288 WARN_ON_ONCE(last_ackt == 0);
3290 first_ackt = last_ackt;
3292 if (before(start_seq, reord))
3294 if (!after(scb->end_seq, tp->high_seq))
3295 flag |= FLAG_ORIG_SACK_ACKED;
3298 if (sacked & TCPCB_SACKED_ACKED) {
3299 tp->sacked_out -= acked_pcount;
3300 } else if (tcp_is_sack(tp)) {
3301 tcp_count_delivered(tp, acked_pcount, ece_ack);
3302 if (!tcp_skb_spurious_retrans(tp, skb))
3303 tcp_rack_advance(tp, sacked, scb->end_seq,
3304 tcp_skb_timestamp_us(skb));
3306 if (sacked & TCPCB_LOST)
3307 tp->lost_out -= acked_pcount;
3309 tp->packets_out -= acked_pcount;
3310 pkts_acked += acked_pcount;
3311 tcp_rate_skb_delivered(sk, skb, sack->rate);
3313 /* Initial outgoing SYN's get put onto the write_queue
3314 * just like anything else we transmit. It is not
3315 * true data, and if we misinform our callers that
3316 * this ACK acks real data, we will erroneously exit
3317 * connection startup slow start one packet too
3318 * quickly. This is severely frowned upon behavior.
3320 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
3321 flag |= FLAG_DATA_ACKED;
3323 flag |= FLAG_SYN_ACKED;
3324 tp->retrans_stamp = 0;
3330 tcp_ack_tstamp(sk, skb, ack_skb, prior_snd_una);
3332 next = skb_rb_next(skb);
3333 if (unlikely(skb == tp->retransmit_skb_hint))
3334 tp->retransmit_skb_hint = NULL;
3335 if (unlikely(skb == tp->lost_skb_hint))
3336 tp->lost_skb_hint = NULL;
3337 tcp_highest_sack_replace(sk, skb, next);
3338 tcp_rtx_queue_unlink_and_free(skb, sk);
3342 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
3344 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3345 tp->snd_up = tp->snd_una;
3348 tcp_ack_tstamp(sk, skb, ack_skb, prior_snd_una);
3349 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
3350 flag |= FLAG_SACK_RENEGING;
3353 if (likely(first_ackt) && !(flag & FLAG_RETRANS_DATA_ACKED)) {
3354 seq_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, first_ackt);
3355 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, last_ackt);
3357 if (pkts_acked == 1 && fully_acked && !prior_sacked &&
3358 (tp->snd_una - prior_snd_una) < tp->mss_cache &&
3359 sack->rate->prior_delivered + 1 == tp->delivered &&
3360 !(flag & (FLAG_CA_ALERT | FLAG_SYN_ACKED))) {
3361 /* Conservatively mark a delayed ACK. It's typically
3362 * from a lone runt packet over the round trip to
3363 * a receiver w/o out-of-order or CE events.
3365 flag |= FLAG_ACK_MAYBE_DELAYED;
3368 if (sack->first_sackt) {
3369 sack_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->first_sackt);
3370 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->last_sackt);
3372 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us,
3373 ca_rtt_us, sack->rate);
3375 if (flag & FLAG_ACKED) {
3376 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3377 if (unlikely(icsk->icsk_mtup.probe_size &&
3378 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3379 tcp_mtup_probe_success(sk);
3382 if (tcp_is_reno(tp)) {
3383 tcp_remove_reno_sacks(sk, pkts_acked, ece_ack);
3385 /* If any of the cumulatively ACKed segments was
3386 * retransmitted, non-SACK case cannot confirm that
3387 * progress was due to original transmission due to
3388 * lack of TCPCB_SACKED_ACKED bits even if some of
3389 * the packets may have been never retransmitted.
3391 if (flag & FLAG_RETRANS_DATA_ACKED)
3392 flag &= ~FLAG_ORIG_SACK_ACKED;
3396 /* Non-retransmitted hole got filled? That's reordering */
3397 if (before(reord, prior_fack))
3398 tcp_check_sack_reordering(sk, reord, 0);
3400 delta = prior_sacked - tp->sacked_out;
3401 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3403 } else if (skb && rtt_update && sack_rtt_us >= 0 &&
3404 sack_rtt_us > tcp_stamp_us_delta(tp->tcp_mstamp,
3405 tcp_skb_timestamp_us(skb))) {
3406 /* Do not re-arm RTO if the sack RTT is measured from data sent
3407 * after when the head was last (re)transmitted. Otherwise the
3408 * timeout may continue to extend in loss recovery.
3410 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3413 if (icsk->icsk_ca_ops->pkts_acked) {
3414 struct ack_sample sample = { .pkts_acked = pkts_acked,
3415 .rtt_us = sack->rate->rtt_us };
3417 sample.in_flight = tp->mss_cache *
3418 (tp->delivered - sack->rate->prior_delivered);
3419 icsk->icsk_ca_ops->pkts_acked(sk, &sample);
3422 #if FASTRETRANS_DEBUG > 0
3423 WARN_ON((int)tp->sacked_out < 0);
3424 WARN_ON((int)tp->lost_out < 0);
3425 WARN_ON((int)tp->retrans_out < 0);
3426 if (!tp->packets_out && tcp_is_sack(tp)) {
3427 icsk = inet_csk(sk);
3429 pr_debug("Leak l=%u %d\n",
3430 tp->lost_out, icsk->icsk_ca_state);
3433 if (tp->sacked_out) {
3434 pr_debug("Leak s=%u %d\n",
3435 tp->sacked_out, icsk->icsk_ca_state);
3438 if (tp->retrans_out) {
3439 pr_debug("Leak r=%u %d\n",
3440 tp->retrans_out, icsk->icsk_ca_state);
3441 tp->retrans_out = 0;
3448 static void tcp_ack_probe(struct sock *sk)
3450 struct inet_connection_sock *icsk = inet_csk(sk);
3451 struct sk_buff *head = tcp_send_head(sk);
3452 const struct tcp_sock *tp = tcp_sk(sk);
3454 /* Was it a usable window open? */
3457 if (!after(TCP_SKB_CB(head)->end_seq, tcp_wnd_end(tp))) {
3458 icsk->icsk_backoff = 0;
3459 icsk->icsk_probes_tstamp = 0;
3460 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3461 /* Socket must be waked up by subsequent tcp_data_snd_check().
3462 * This function is not for random using!
3465 unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX);
3467 when = tcp_clamp_probe0_to_user_timeout(sk, when);
3468 tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, when, TCP_RTO_MAX);
3472 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3474 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3475 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3478 /* Decide wheather to run the increase function of congestion control. */
3479 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3481 /* If reordering is high then always grow cwnd whenever data is
3482 * delivered regardless of its ordering. Otherwise stay conservative
3483 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3484 * new SACK or ECE mark may first advance cwnd here and later reduce
3485 * cwnd in tcp_fastretrans_alert() based on more states.
3487 if (tcp_sk(sk)->reordering >
3488 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_reordering))
3489 return flag & FLAG_FORWARD_PROGRESS;
3491 return flag & FLAG_DATA_ACKED;
3494 /* The "ultimate" congestion control function that aims to replace the rigid
3495 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction).
3496 * It's called toward the end of processing an ACK with precise rate
3497 * information. All transmission or retransmission are delayed afterwards.
3499 static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked,
3500 int flag, const struct rate_sample *rs)
3502 const struct inet_connection_sock *icsk = inet_csk(sk);
3504 if (icsk->icsk_ca_ops->cong_control) {
3505 icsk->icsk_ca_ops->cong_control(sk, rs);
3509 if (tcp_in_cwnd_reduction(sk)) {
3510 /* Reduce cwnd if state mandates */
3511 tcp_cwnd_reduction(sk, acked_sacked, rs->losses, flag);
3512 } else if (tcp_may_raise_cwnd(sk, flag)) {
3513 /* Advance cwnd if state allows */
3514 tcp_cong_avoid(sk, ack, acked_sacked);
3516 tcp_update_pacing_rate(sk);
3519 /* Check that window update is acceptable.
3520 * The function assumes that snd_una<=ack<=snd_next.
3522 static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3523 const u32 ack, const u32 ack_seq,
3526 return after(ack, tp->snd_una) ||
3527 after(ack_seq, tp->snd_wl1) ||
3528 (ack_seq == tp->snd_wl1 && (nwin > tp->snd_wnd || !nwin));
3531 /* If we update tp->snd_una, also update tp->bytes_acked */
3532 static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack)
3534 u32 delta = ack - tp->snd_una;
3536 sock_owned_by_me((struct sock *)tp);
3537 tp->bytes_acked += delta;
3541 /* If we update tp->rcv_nxt, also update tp->bytes_received */
3542 static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq)
3544 u32 delta = seq - tp->rcv_nxt;
3546 sock_owned_by_me((struct sock *)tp);
3547 tp->bytes_received += delta;
3548 WRITE_ONCE(tp->rcv_nxt, seq);
3551 /* Update our send window.
3553 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3554 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3556 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3559 struct tcp_sock *tp = tcp_sk(sk);
3561 u32 nwin = ntohs(tcp_hdr(skb)->window);
3563 if (likely(!tcp_hdr(skb)->syn))
3564 nwin <<= tp->rx_opt.snd_wscale;
3566 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3567 flag |= FLAG_WIN_UPDATE;
3568 tcp_update_wl(tp, ack_seq);
3570 if (tp->snd_wnd != nwin) {
3573 /* Note, it is the only place, where
3574 * fast path is recovered for sending TCP.
3577 tcp_fast_path_check(sk);
3579 if (!tcp_write_queue_empty(sk))
3580 tcp_slow_start_after_idle_check(sk);
3582 if (nwin > tp->max_window) {
3583 tp->max_window = nwin;
3584 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3589 tcp_snd_una_update(tp, ack);
3594 static bool __tcp_oow_rate_limited(struct net *net, int mib_idx,
3595 u32 *last_oow_ack_time)
3597 /* Paired with the WRITE_ONCE() in this function. */
3598 u32 val = READ_ONCE(*last_oow_ack_time);
3601 s32 elapsed = (s32)(tcp_jiffies32 - val);
3604 elapsed < READ_ONCE(net->ipv4.sysctl_tcp_invalid_ratelimit)) {
3605 NET_INC_STATS(net, mib_idx);
3606 return true; /* rate-limited: don't send yet! */
3610 /* Paired with the prior READ_ONCE() and with itself,
3611 * as we might be lockless.
3613 WRITE_ONCE(*last_oow_ack_time, tcp_jiffies32);
3615 return false; /* not rate-limited: go ahead, send dupack now! */
3618 /* Return true if we're currently rate-limiting out-of-window ACKs and
3619 * thus shouldn't send a dupack right now. We rate-limit dupacks in
3620 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3621 * attacks that send repeated SYNs or ACKs for the same connection. To
3622 * do this, we do not send a duplicate SYNACK or ACK if the remote
3623 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3625 bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb,
3626 int mib_idx, u32 *last_oow_ack_time)
3628 /* Data packets without SYNs are not likely part of an ACK loop. */
3629 if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) &&
3633 return __tcp_oow_rate_limited(net, mib_idx, last_oow_ack_time);
3636 /* RFC 5961 7 [ACK Throttling] */
3637 static void tcp_send_challenge_ack(struct sock *sk)
3639 struct tcp_sock *tp = tcp_sk(sk);
3640 struct net *net = sock_net(sk);
3641 u32 count, now, ack_limit;
3643 /* First check our per-socket dupack rate limit. */
3644 if (__tcp_oow_rate_limited(net,
3645 LINUX_MIB_TCPACKSKIPPEDCHALLENGE,
3646 &tp->last_oow_ack_time))
3649 ack_limit = READ_ONCE(net->ipv4.sysctl_tcp_challenge_ack_limit);
3650 if (ack_limit == INT_MAX)
3653 /* Then check host-wide RFC 5961 rate limit. */
3655 if (now != READ_ONCE(net->ipv4.tcp_challenge_timestamp)) {
3656 u32 half = (ack_limit + 1) >> 1;
3658 WRITE_ONCE(net->ipv4.tcp_challenge_timestamp, now);
3659 WRITE_ONCE(net->ipv4.tcp_challenge_count,
3660 get_random_u32_inclusive(half, ack_limit + half - 1));
3662 count = READ_ONCE(net->ipv4.tcp_challenge_count);
3664 WRITE_ONCE(net->ipv4.tcp_challenge_count, count - 1);
3666 NET_INC_STATS(net, LINUX_MIB_TCPCHALLENGEACK);
3671 static void tcp_store_ts_recent(struct tcp_sock *tp)
3673 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3674 tp->rx_opt.ts_recent_stamp = ktime_get_seconds();
3677 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3679 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3680 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3681 * extra check below makes sure this can only happen
3682 * for pure ACK frames. -DaveM
3684 * Not only, also it occurs for expired timestamps.
3687 if (tcp_paws_check(&tp->rx_opt, 0))
3688 tcp_store_ts_recent(tp);
3692 /* This routine deals with acks during a TLP episode and ends an episode by
3693 * resetting tlp_high_seq. Ref: TLP algorithm in draft-ietf-tcpm-rack
3695 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3697 struct tcp_sock *tp = tcp_sk(sk);
3699 if (before(ack, tp->tlp_high_seq))
3702 if (!tp->tlp_retrans) {
3703 /* TLP of new data has been acknowledged */
3704 tp->tlp_high_seq = 0;
3705 } else if (flag & FLAG_DSACK_TLP) {
3706 /* This DSACK means original and TLP probe arrived; no loss */
3707 tp->tlp_high_seq = 0;
3708 } else if (after(ack, tp->tlp_high_seq)) {
3709 /* ACK advances: there was a loss, so reduce cwnd. Reset
3710 * tlp_high_seq in tcp_init_cwnd_reduction()
3712 tcp_init_cwnd_reduction(sk);
3713 tcp_set_ca_state(sk, TCP_CA_CWR);
3714 tcp_end_cwnd_reduction(sk);
3715 tcp_try_keep_open(sk);
3716 NET_INC_STATS(sock_net(sk),
3717 LINUX_MIB_TCPLOSSPROBERECOVERY);
3718 } else if (!(flag & (FLAG_SND_UNA_ADVANCED |
3719 FLAG_NOT_DUP | FLAG_DATA_SACKED))) {
3720 /* Pure dupack: original and TLP probe arrived; no loss */
3721 tp->tlp_high_seq = 0;
3725 static inline void tcp_in_ack_event(struct sock *sk, u32 flags)
3727 const struct inet_connection_sock *icsk = inet_csk(sk);
3729 if (icsk->icsk_ca_ops->in_ack_event)
3730 icsk->icsk_ca_ops->in_ack_event(sk, flags);
3733 /* Congestion control has updated the cwnd already. So if we're in
3734 * loss recovery then now we do any new sends (for FRTO) or
3735 * retransmits (for CA_Loss or CA_recovery) that make sense.
3737 static void tcp_xmit_recovery(struct sock *sk, int rexmit)
3739 struct tcp_sock *tp = tcp_sk(sk);
3741 if (rexmit == REXMIT_NONE || sk->sk_state == TCP_SYN_SENT)
3744 if (unlikely(rexmit == REXMIT_NEW)) {
3745 __tcp_push_pending_frames(sk, tcp_current_mss(sk),
3747 if (after(tp->snd_nxt, tp->high_seq))
3751 tcp_xmit_retransmit_queue(sk);
3754 /* Returns the number of packets newly acked or sacked by the current ACK */
3755 static u32 tcp_newly_delivered(struct sock *sk, u32 prior_delivered, int flag)
3757 const struct net *net = sock_net(sk);
3758 struct tcp_sock *tp = tcp_sk(sk);
3761 delivered = tp->delivered - prior_delivered;
3762 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVERED, delivered);
3763 if (flag & FLAG_ECE)
3764 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVEREDCE, delivered);
3769 /* This routine deals with incoming acks, but not outgoing ones. */
3770 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3772 struct inet_connection_sock *icsk = inet_csk(sk);
3773 struct tcp_sock *tp = tcp_sk(sk);
3774 struct tcp_sacktag_state sack_state;
3775 struct rate_sample rs = { .prior_delivered = 0 };
3776 u32 prior_snd_una = tp->snd_una;
3777 bool is_sack_reneg = tp->is_sack_reneg;
3778 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3779 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3781 int prior_packets = tp->packets_out;
3782 u32 delivered = tp->delivered;
3783 u32 lost = tp->lost;
3784 int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */
3787 sack_state.first_sackt = 0;
3788 sack_state.rate = &rs;
3789 sack_state.sack_delivered = 0;
3791 /* We very likely will need to access rtx queue. */
3792 prefetch(sk->tcp_rtx_queue.rb_node);
3794 /* If the ack is older than previous acks
3795 * then we can probably ignore it.
3797 if (before(ack, prior_snd_una)) {
3798 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3799 if (before(ack, prior_snd_una - tp->max_window)) {
3800 if (!(flag & FLAG_NO_CHALLENGE_ACK))
3801 tcp_send_challenge_ack(sk);
3802 return -SKB_DROP_REASON_TCP_TOO_OLD_ACK;
3807 /* If the ack includes data we haven't sent yet, discard
3808 * this segment (RFC793 Section 3.9).
3810 if (after(ack, tp->snd_nxt))
3811 return -SKB_DROP_REASON_TCP_ACK_UNSENT_DATA;
3813 if (after(ack, prior_snd_una)) {
3814 flag |= FLAG_SND_UNA_ADVANCED;
3815 icsk->icsk_retransmits = 0;
3817 #if IS_ENABLED(CONFIG_TLS_DEVICE)
3818 if (static_branch_unlikely(&clean_acked_data_enabled.key))
3819 if (icsk->icsk_clean_acked)
3820 icsk->icsk_clean_acked(sk, ack);
3824 prior_fack = tcp_is_sack(tp) ? tcp_highest_sack_seq(tp) : tp->snd_una;
3825 rs.prior_in_flight = tcp_packets_in_flight(tp);
3827 /* ts_recent update must be made after we are sure that the packet
3830 if (flag & FLAG_UPDATE_TS_RECENT)
3831 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3833 if ((flag & (FLAG_SLOWPATH | FLAG_SND_UNA_ADVANCED)) ==
3834 FLAG_SND_UNA_ADVANCED) {
3835 /* Window is constant, pure forward advance.
3836 * No more checks are required.
3837 * Note, we use the fact that SND.UNA>=SND.WL2.
3839 tcp_update_wl(tp, ack_seq);
3840 tcp_snd_una_update(tp, ack);
3841 flag |= FLAG_WIN_UPDATE;
3843 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
3845 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPACKS);
3847 u32 ack_ev_flags = CA_ACK_SLOWPATH;
3849 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3852 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3854 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3856 if (TCP_SKB_CB(skb)->sacked)
3857 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3860 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) {
3862 ack_ev_flags |= CA_ACK_ECE;
3865 if (sack_state.sack_delivered)
3866 tcp_count_delivered(tp, sack_state.sack_delivered,
3869 if (flag & FLAG_WIN_UPDATE)
3870 ack_ev_flags |= CA_ACK_WIN_UPDATE;
3872 tcp_in_ack_event(sk, ack_ev_flags);
3875 /* This is a deviation from RFC3168 since it states that:
3876 * "When the TCP data sender is ready to set the CWR bit after reducing
3877 * the congestion window, it SHOULD set the CWR bit only on the first
3878 * new data packet that it transmits."
3879 * We accept CWR on pure ACKs to be more robust
3880 * with widely-deployed TCP implementations that do this.
3882 tcp_ecn_accept_cwr(sk, skb);
3884 /* We passed data and got it acked, remove any soft error
3885 * log. Something worked...
3887 WRITE_ONCE(sk->sk_err_soft, 0);
3888 icsk->icsk_probes_out = 0;
3889 tp->rcv_tstamp = tcp_jiffies32;
3893 /* See if we can take anything off of the retransmit queue. */
3894 flag |= tcp_clean_rtx_queue(sk, skb, prior_fack, prior_snd_una,
3895 &sack_state, flag & FLAG_ECE);
3897 tcp_rack_update_reo_wnd(sk, &rs);
3899 if (tp->tlp_high_seq)
3900 tcp_process_tlp_ack(sk, ack, flag);
3902 if (tcp_ack_is_dubious(sk, flag)) {
3903 if (!(flag & (FLAG_SND_UNA_ADVANCED |
3904 FLAG_NOT_DUP | FLAG_DSACKING_ACK))) {
3906 /* Consider if pure acks were aggregated in tcp_add_backlog() */
3907 if (!(flag & FLAG_DATA))
3908 num_dupack = max_t(u16, 1, skb_shinfo(skb)->gso_segs);
3910 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
3914 /* If needed, reset TLP/RTO timer when RACK doesn't set. */
3915 if (flag & FLAG_SET_XMIT_TIMER)
3916 tcp_set_xmit_timer(sk);
3918 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
3921 delivered = tcp_newly_delivered(sk, delivered, flag);
3922 lost = tp->lost - lost; /* freshly marked lost */
3923 rs.is_ack_delayed = !!(flag & FLAG_ACK_MAYBE_DELAYED);
3924 tcp_rate_gen(sk, delivered, lost, is_sack_reneg, sack_state.rate);
3925 tcp_cong_control(sk, ack, delivered, flag, sack_state.rate);
3926 tcp_xmit_recovery(sk, rexmit);
3930 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3931 if (flag & FLAG_DSACKING_ACK) {
3932 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
3934 tcp_newly_delivered(sk, delivered, flag);
3936 /* If this ack opens up a zero window, clear backoff. It was
3937 * being used to time the probes, and is probably far higher than
3938 * it needs to be for normal retransmission.
3942 if (tp->tlp_high_seq)
3943 tcp_process_tlp_ack(sk, ack, flag);
3947 /* If data was SACKed, tag it and see if we should send more data.
3948 * If data was DSACKed, see if we can undo a cwnd reduction.
3950 if (TCP_SKB_CB(skb)->sacked) {
3951 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3953 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
3955 tcp_newly_delivered(sk, delivered, flag);
3956 tcp_xmit_recovery(sk, rexmit);
3962 static void tcp_parse_fastopen_option(int len, const unsigned char *cookie,
3963 bool syn, struct tcp_fastopen_cookie *foc,
3966 /* Valid only in SYN or SYN-ACK with an even length. */
3967 if (!foc || !syn || len < 0 || (len & 1))
3970 if (len >= TCP_FASTOPEN_COOKIE_MIN &&
3971 len <= TCP_FASTOPEN_COOKIE_MAX)
3972 memcpy(foc->val, cookie, len);
3979 static bool smc_parse_options(const struct tcphdr *th,
3980 struct tcp_options_received *opt_rx,
3981 const unsigned char *ptr,
3984 #if IS_ENABLED(CONFIG_SMC)
3985 if (static_branch_unlikely(&tcp_have_smc)) {
3986 if (th->syn && !(opsize & 1) &&
3987 opsize >= TCPOLEN_EXP_SMC_BASE &&
3988 get_unaligned_be32(ptr) == TCPOPT_SMC_MAGIC) {
3997 /* Try to parse the MSS option from the TCP header. Return 0 on failure, clamped
4000 u16 tcp_parse_mss_option(const struct tcphdr *th, u16 user_mss)
4002 const unsigned char *ptr = (const unsigned char *)(th + 1);
4003 int length = (th->doff * 4) - sizeof(struct tcphdr);
4006 while (length > 0) {
4007 int opcode = *ptr++;
4013 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
4020 if (opsize < 2) /* "silly options" */
4022 if (opsize > length)
4023 return mss; /* fail on partial options */
4024 if (opcode == TCPOPT_MSS && opsize == TCPOLEN_MSS) {
4025 u16 in_mss = get_unaligned_be16(ptr);
4028 if (user_mss && user_mss < in_mss)
4039 EXPORT_SYMBOL_GPL(tcp_parse_mss_option);
4041 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
4042 * But, this can also be called on packets in the established flow when
4043 * the fast version below fails.
4045 void tcp_parse_options(const struct net *net,
4046 const struct sk_buff *skb,
4047 struct tcp_options_received *opt_rx, int estab,
4048 struct tcp_fastopen_cookie *foc)
4050 const unsigned char *ptr;
4051 const struct tcphdr *th = tcp_hdr(skb);
4052 int length = (th->doff * 4) - sizeof(struct tcphdr);
4054 ptr = (const unsigned char *)(th + 1);
4055 opt_rx->saw_tstamp = 0;
4056 opt_rx->saw_unknown = 0;
4058 while (length > 0) {
4059 int opcode = *ptr++;
4065 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
4072 if (opsize < 2) /* "silly options" */
4074 if (opsize > length)
4075 return; /* don't parse partial options */
4078 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
4079 u16 in_mss = get_unaligned_be16(ptr);
4081 if (opt_rx->user_mss &&
4082 opt_rx->user_mss < in_mss)
4083 in_mss = opt_rx->user_mss;
4084 opt_rx->mss_clamp = in_mss;
4089 if (opsize == TCPOLEN_WINDOW && th->syn &&
4090 !estab && READ_ONCE(net->ipv4.sysctl_tcp_window_scaling)) {
4091 __u8 snd_wscale = *(__u8 *)ptr;
4092 opt_rx->wscale_ok = 1;
4093 if (snd_wscale > TCP_MAX_WSCALE) {
4094 net_info_ratelimited("%s: Illegal window scaling value %d > %u received\n",
4098 snd_wscale = TCP_MAX_WSCALE;
4100 opt_rx->snd_wscale = snd_wscale;
4103 case TCPOPT_TIMESTAMP:
4104 if ((opsize == TCPOLEN_TIMESTAMP) &&
4105 ((estab && opt_rx->tstamp_ok) ||
4106 (!estab && READ_ONCE(net->ipv4.sysctl_tcp_timestamps)))) {
4107 opt_rx->saw_tstamp = 1;
4108 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
4109 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
4112 case TCPOPT_SACK_PERM:
4113 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
4114 !estab && READ_ONCE(net->ipv4.sysctl_tcp_sack)) {
4115 opt_rx->sack_ok = TCP_SACK_SEEN;
4116 tcp_sack_reset(opt_rx);
4121 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
4122 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
4124 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
4127 #ifdef CONFIG_TCP_MD5SIG
4129 /* The MD5 Hash has already been
4130 * checked (see tcp_v{4,6}_rcv()).
4134 case TCPOPT_FASTOPEN:
4135 tcp_parse_fastopen_option(
4136 opsize - TCPOLEN_FASTOPEN_BASE,
4137 ptr, th->syn, foc, false);
4141 /* Fast Open option shares code 254 using a
4142 * 16 bits magic number.
4144 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE &&
4145 get_unaligned_be16(ptr) ==
4146 TCPOPT_FASTOPEN_MAGIC) {
4147 tcp_parse_fastopen_option(opsize -
4148 TCPOLEN_EXP_FASTOPEN_BASE,
4149 ptr + 2, th->syn, foc, true);
4153 if (smc_parse_options(th, opt_rx, ptr, opsize))
4156 opt_rx->saw_unknown = 1;
4160 opt_rx->saw_unknown = 1;
4167 EXPORT_SYMBOL(tcp_parse_options);
4169 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
4171 const __be32 *ptr = (const __be32 *)(th + 1);
4173 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
4174 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
4175 tp->rx_opt.saw_tstamp = 1;
4177 tp->rx_opt.rcv_tsval = ntohl(*ptr);
4180 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
4182 tp->rx_opt.rcv_tsecr = 0;
4188 /* Fast parse options. This hopes to only see timestamps.
4189 * If it is wrong it falls back on tcp_parse_options().
4191 static bool tcp_fast_parse_options(const struct net *net,
4192 const struct sk_buff *skb,
4193 const struct tcphdr *th, struct tcp_sock *tp)
4195 /* In the spirit of fast parsing, compare doff directly to constant
4196 * values. Because equality is used, short doff can be ignored here.
4198 if (th->doff == (sizeof(*th) / 4)) {
4199 tp->rx_opt.saw_tstamp = 0;
4201 } else if (tp->rx_opt.tstamp_ok &&
4202 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
4203 if (tcp_parse_aligned_timestamp(tp, th))
4207 tcp_parse_options(net, skb, &tp->rx_opt, 1, NULL);
4208 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
4209 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
4214 #ifdef CONFIG_TCP_MD5SIG
4216 * Parse MD5 Signature option
4218 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
4220 int length = (th->doff << 2) - sizeof(*th);
4221 const u8 *ptr = (const u8 *)(th + 1);
4223 /* If not enough data remaining, we can short cut */
4224 while (length >= TCPOLEN_MD5SIG) {
4225 int opcode = *ptr++;
4236 if (opsize < 2 || opsize > length)
4238 if (opcode == TCPOPT_MD5SIG)
4239 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
4246 EXPORT_SYMBOL(tcp_parse_md5sig_option);
4249 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
4251 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
4252 * it can pass through stack. So, the following predicate verifies that
4253 * this segment is not used for anything but congestion avoidance or
4254 * fast retransmit. Moreover, we even are able to eliminate most of such
4255 * second order effects, if we apply some small "replay" window (~RTO)
4256 * to timestamp space.
4258 * All these measures still do not guarantee that we reject wrapped ACKs
4259 * on networks with high bandwidth, when sequence space is recycled fastly,
4260 * but it guarantees that such events will be very rare and do not affect
4261 * connection seriously. This doesn't look nice, but alas, PAWS is really
4264 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
4265 * states that events when retransmit arrives after original data are rare.
4266 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
4267 * the biggest problem on large power networks even with minor reordering.
4268 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
4269 * up to bandwidth of 18Gigabit/sec. 8) ]
4272 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
4274 const struct tcp_sock *tp = tcp_sk(sk);
4275 const struct tcphdr *th = tcp_hdr(skb);
4276 u32 seq = TCP_SKB_CB(skb)->seq;
4277 u32 ack = TCP_SKB_CB(skb)->ack_seq;
4279 return (/* 1. Pure ACK with correct sequence number. */
4280 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
4282 /* 2. ... and duplicate ACK. */
4283 ack == tp->snd_una &&
4285 /* 3. ... and does not update window. */
4286 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
4288 /* 4. ... and sits in replay window. */
4289 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
4292 static inline bool tcp_paws_discard(const struct sock *sk,
4293 const struct sk_buff *skb)
4295 const struct tcp_sock *tp = tcp_sk(sk);
4297 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
4298 !tcp_disordered_ack(sk, skb);
4301 /* Check segment sequence number for validity.
4303 * Segment controls are considered valid, if the segment
4304 * fits to the window after truncation to the window. Acceptability
4305 * of data (and SYN, FIN, of course) is checked separately.
4306 * See tcp_data_queue(), for example.
4308 * Also, controls (RST is main one) are accepted using RCV.WUP instead
4309 * of RCV.NXT. Peer still did not advance his SND.UNA when we
4310 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4311 * (borrowed from freebsd)
4314 static enum skb_drop_reason tcp_sequence(const struct tcp_sock *tp,
4315 u32 seq, u32 end_seq)
4317 if (before(end_seq, tp->rcv_wup))
4318 return SKB_DROP_REASON_TCP_OLD_SEQUENCE;
4320 if (after(seq, tp->rcv_nxt + tcp_receive_window(tp)))
4321 return SKB_DROP_REASON_TCP_INVALID_SEQUENCE;
4323 return SKB_NOT_DROPPED_YET;
4326 /* When we get a reset we do this. */
4327 void tcp_reset(struct sock *sk, struct sk_buff *skb)
4329 trace_tcp_receive_reset(sk);
4331 /* mptcp can't tell us to ignore reset pkts,
4332 * so just ignore the return value of mptcp_incoming_options().
4334 if (sk_is_mptcp(sk))
4335 mptcp_incoming_options(sk, skb);
4337 /* We want the right error as BSD sees it (and indeed as we do). */
4338 switch (sk->sk_state) {
4340 WRITE_ONCE(sk->sk_err, ECONNREFUSED);
4342 case TCP_CLOSE_WAIT:
4343 WRITE_ONCE(sk->sk_err, EPIPE);
4348 WRITE_ONCE(sk->sk_err, ECONNRESET);
4350 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4353 tcp_write_queue_purge(sk);
4356 if (!sock_flag(sk, SOCK_DEAD))
4357 sk_error_report(sk);
4361 * Process the FIN bit. This now behaves as it is supposed to work
4362 * and the FIN takes effect when it is validly part of sequence
4363 * space. Not before when we get holes.
4365 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4366 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4369 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4370 * close and we go into CLOSING (and later onto TIME-WAIT)
4372 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4374 void tcp_fin(struct sock *sk)
4376 struct tcp_sock *tp = tcp_sk(sk);
4378 inet_csk_schedule_ack(sk);
4380 WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | RCV_SHUTDOWN);
4381 sock_set_flag(sk, SOCK_DONE);
4383 switch (sk->sk_state) {
4385 case TCP_ESTABLISHED:
4386 /* Move to CLOSE_WAIT */
4387 tcp_set_state(sk, TCP_CLOSE_WAIT);
4388 inet_csk_enter_pingpong_mode(sk);
4391 case TCP_CLOSE_WAIT:
4393 /* Received a retransmission of the FIN, do
4398 /* RFC793: Remain in the LAST-ACK state. */
4402 /* This case occurs when a simultaneous close
4403 * happens, we must ack the received FIN and
4404 * enter the CLOSING state.
4407 tcp_set_state(sk, TCP_CLOSING);
4410 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4412 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4415 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4416 * cases we should never reach this piece of code.
4418 pr_err("%s: Impossible, sk->sk_state=%d\n",
4419 __func__, sk->sk_state);
4423 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4424 * Probably, we should reset in this case. For now drop them.
4426 skb_rbtree_purge(&tp->out_of_order_queue);
4427 if (tcp_is_sack(tp))
4428 tcp_sack_reset(&tp->rx_opt);
4430 if (!sock_flag(sk, SOCK_DEAD)) {
4431 sk->sk_state_change(sk);
4433 /* Do not send POLL_HUP for half duplex close. */
4434 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4435 sk->sk_state == TCP_CLOSE)
4436 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4438 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4442 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4445 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4446 if (before(seq, sp->start_seq))
4447 sp->start_seq = seq;
4448 if (after(end_seq, sp->end_seq))
4449 sp->end_seq = end_seq;
4455 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4457 struct tcp_sock *tp = tcp_sk(sk);
4459 if (tcp_is_sack(tp) && READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_dsack)) {
4462 if (before(seq, tp->rcv_nxt))
4463 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4465 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4467 NET_INC_STATS(sock_net(sk), mib_idx);
4469 tp->rx_opt.dsack = 1;
4470 tp->duplicate_sack[0].start_seq = seq;
4471 tp->duplicate_sack[0].end_seq = end_seq;
4475 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4477 struct tcp_sock *tp = tcp_sk(sk);
4479 if (!tp->rx_opt.dsack)
4480 tcp_dsack_set(sk, seq, end_seq);
4482 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4485 static void tcp_rcv_spurious_retrans(struct sock *sk, const struct sk_buff *skb)
4487 /* When the ACK path fails or drops most ACKs, the sender would
4488 * timeout and spuriously retransmit the same segment repeatedly.
4489 * The receiver remembers and reflects via DSACKs. Leverage the
4490 * DSACK state and change the txhash to re-route speculatively.
4492 if (TCP_SKB_CB(skb)->seq == tcp_sk(sk)->duplicate_sack[0].start_seq &&
4493 sk_rethink_txhash(sk))
4494 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDUPLICATEDATAREHASH);
4497 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4499 struct tcp_sock *tp = tcp_sk(sk);
4501 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4502 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4503 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4504 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
4506 if (tcp_is_sack(tp) && READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_dsack)) {
4507 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4509 tcp_rcv_spurious_retrans(sk, skb);
4510 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4511 end_seq = tp->rcv_nxt;
4512 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4519 /* These routines update the SACK block as out-of-order packets arrive or
4520 * in-order packets close up the sequence space.
4522 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4525 struct tcp_sack_block *sp = &tp->selective_acks[0];
4526 struct tcp_sack_block *swalk = sp + 1;
4528 /* See if the recent change to the first SACK eats into
4529 * or hits the sequence space of other SACK blocks, if so coalesce.
4531 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4532 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4535 /* Zap SWALK, by moving every further SACK up by one slot.
4536 * Decrease num_sacks.
4538 tp->rx_opt.num_sacks--;
4539 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4548 void tcp_sack_compress_send_ack(struct sock *sk)
4550 struct tcp_sock *tp = tcp_sk(sk);
4552 if (!tp->compressed_ack)
4555 if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1)
4558 /* Since we have to send one ack finally,
4559 * substract one from tp->compressed_ack to keep
4560 * LINUX_MIB_TCPACKCOMPRESSED accurate.
4562 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED,
4563 tp->compressed_ack - 1);
4565 tp->compressed_ack = 0;
4569 /* Reasonable amount of sack blocks included in TCP SACK option
4570 * The max is 4, but this becomes 3 if TCP timestamps are there.
4571 * Given that SACK packets might be lost, be conservative and use 2.
4573 #define TCP_SACK_BLOCKS_EXPECTED 2
4575 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4577 struct tcp_sock *tp = tcp_sk(sk);
4578 struct tcp_sack_block *sp = &tp->selective_acks[0];
4579 int cur_sacks = tp->rx_opt.num_sacks;
4585 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4586 if (tcp_sack_extend(sp, seq, end_seq)) {
4587 if (this_sack >= TCP_SACK_BLOCKS_EXPECTED)
4588 tcp_sack_compress_send_ack(sk);
4589 /* Rotate this_sack to the first one. */
4590 for (; this_sack > 0; this_sack--, sp--)
4591 swap(*sp, *(sp - 1));
4593 tcp_sack_maybe_coalesce(tp);
4598 if (this_sack >= TCP_SACK_BLOCKS_EXPECTED)
4599 tcp_sack_compress_send_ack(sk);
4601 /* Could not find an adjacent existing SACK, build a new one,
4602 * put it at the front, and shift everyone else down. We
4603 * always know there is at least one SACK present already here.
4605 * If the sack array is full, forget about the last one.
4607 if (this_sack >= TCP_NUM_SACKS) {
4609 tp->rx_opt.num_sacks--;
4612 for (; this_sack > 0; this_sack--, sp--)
4616 /* Build the new head SACK, and we're done. */
4617 sp->start_seq = seq;
4618 sp->end_seq = end_seq;
4619 tp->rx_opt.num_sacks++;
4622 /* RCV.NXT advances, some SACKs should be eaten. */
4624 static void tcp_sack_remove(struct tcp_sock *tp)
4626 struct tcp_sack_block *sp = &tp->selective_acks[0];
4627 int num_sacks = tp->rx_opt.num_sacks;
4630 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4631 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4632 tp->rx_opt.num_sacks = 0;
4636 for (this_sack = 0; this_sack < num_sacks;) {
4637 /* Check if the start of the sack is covered by RCV.NXT. */
4638 if (!before(tp->rcv_nxt, sp->start_seq)) {
4641 /* RCV.NXT must cover all the block! */
4642 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4644 /* Zap this SACK, by moving forward any other SACKS. */
4645 for (i = this_sack+1; i < num_sacks; i++)
4646 tp->selective_acks[i-1] = tp->selective_acks[i];
4653 tp->rx_opt.num_sacks = num_sacks;
4657 * tcp_try_coalesce - try to merge skb to prior one
4660 * @from: buffer to add in queue
4661 * @fragstolen: pointer to boolean
4663 * Before queueing skb @from after @to, try to merge them
4664 * to reduce overall memory use and queue lengths, if cost is small.
4665 * Packets in ofo or receive queues can stay a long time.
4666 * Better try to coalesce them right now to avoid future collapses.
4667 * Returns true if caller should free @from instead of queueing it
4669 static bool tcp_try_coalesce(struct sock *sk,
4671 struct sk_buff *from,
4676 *fragstolen = false;
4678 /* Its possible this segment overlaps with prior segment in queue */
4679 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4682 if (!mptcp_skb_can_collapse(to, from))
4685 #ifdef CONFIG_TLS_DEVICE
4686 if (from->decrypted != to->decrypted)
4690 if (!skb_try_coalesce(to, from, fragstolen, &delta))
4693 atomic_add(delta, &sk->sk_rmem_alloc);
4694 sk_mem_charge(sk, delta);
4695 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4696 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4697 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4698 TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
4700 if (TCP_SKB_CB(from)->has_rxtstamp) {
4701 TCP_SKB_CB(to)->has_rxtstamp = true;
4702 to->tstamp = from->tstamp;
4703 skb_hwtstamps(to)->hwtstamp = skb_hwtstamps(from)->hwtstamp;
4709 static bool tcp_ooo_try_coalesce(struct sock *sk,
4711 struct sk_buff *from,
4714 bool res = tcp_try_coalesce(sk, to, from, fragstolen);
4716 /* In case tcp_drop_reason() is called later, update to->gso_segs */
4718 u32 gso_segs = max_t(u16, 1, skb_shinfo(to)->gso_segs) +
4719 max_t(u16, 1, skb_shinfo(from)->gso_segs);
4721 skb_shinfo(to)->gso_segs = min_t(u32, gso_segs, 0xFFFF);
4726 static void tcp_drop_reason(struct sock *sk, struct sk_buff *skb,
4727 enum skb_drop_reason reason)
4729 sk_drops_add(sk, skb);
4730 kfree_skb_reason(skb, reason);
4733 /* This one checks to see if we can put data from the
4734 * out_of_order queue into the receive_queue.
4736 static void tcp_ofo_queue(struct sock *sk)
4738 struct tcp_sock *tp = tcp_sk(sk);
4739 __u32 dsack_high = tp->rcv_nxt;
4740 bool fin, fragstolen, eaten;
4741 struct sk_buff *skb, *tail;
4744 p = rb_first(&tp->out_of_order_queue);
4747 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4750 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4751 __u32 dsack = dsack_high;
4752 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4753 dsack_high = TCP_SKB_CB(skb)->end_seq;
4754 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4757 rb_erase(&skb->rbnode, &tp->out_of_order_queue);
4759 if (unlikely(!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))) {
4760 tcp_drop_reason(sk, skb, SKB_DROP_REASON_TCP_OFO_DROP);
4764 tail = skb_peek_tail(&sk->sk_receive_queue);
4765 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen);
4766 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4767 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
4769 __skb_queue_tail(&sk->sk_receive_queue, skb);
4771 kfree_skb_partial(skb, fragstolen);
4773 if (unlikely(fin)) {
4775 /* tcp_fin() purges tp->out_of_order_queue,
4776 * so we must end this loop right now.
4783 static bool tcp_prune_ofo_queue(struct sock *sk, const struct sk_buff *in_skb);
4784 static int tcp_prune_queue(struct sock *sk, const struct sk_buff *in_skb);
4786 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4789 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4790 !sk_rmem_schedule(sk, skb, size)) {
4792 if (tcp_prune_queue(sk, skb) < 0)
4795 while (!sk_rmem_schedule(sk, skb, size)) {
4796 if (!tcp_prune_ofo_queue(sk, skb))
4803 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4805 struct tcp_sock *tp = tcp_sk(sk);
4806 struct rb_node **p, *parent;
4807 struct sk_buff *skb1;
4811 tcp_ecn_check_ce(sk, skb);
4813 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4814 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFODROP);
4815 sk->sk_data_ready(sk);
4816 tcp_drop_reason(sk, skb, SKB_DROP_REASON_PROTO_MEM);
4820 /* Disable header prediction. */
4822 inet_csk_schedule_ack(sk);
4824 tp->rcv_ooopack += max_t(u16, 1, skb_shinfo(skb)->gso_segs);
4825 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4826 seq = TCP_SKB_CB(skb)->seq;
4827 end_seq = TCP_SKB_CB(skb)->end_seq;
4829 p = &tp->out_of_order_queue.rb_node;
4830 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4831 /* Initial out of order segment, build 1 SACK. */
4832 if (tcp_is_sack(tp)) {
4833 tp->rx_opt.num_sacks = 1;
4834 tp->selective_acks[0].start_seq = seq;
4835 tp->selective_acks[0].end_seq = end_seq;
4837 rb_link_node(&skb->rbnode, NULL, p);
4838 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4839 tp->ooo_last_skb = skb;
4843 /* In the typical case, we are adding an skb to the end of the list.
4844 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
4846 if (tcp_ooo_try_coalesce(sk, tp->ooo_last_skb,
4847 skb, &fragstolen)) {
4849 /* For non sack flows, do not grow window to force DUPACK
4850 * and trigger fast retransmit.
4852 if (tcp_is_sack(tp))
4853 tcp_grow_window(sk, skb, true);
4854 kfree_skb_partial(skb, fragstolen);
4858 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
4859 if (!before(seq, TCP_SKB_CB(tp->ooo_last_skb)->end_seq)) {
4860 parent = &tp->ooo_last_skb->rbnode;
4861 p = &parent->rb_right;
4865 /* Find place to insert this segment. Handle overlaps on the way. */
4869 skb1 = rb_to_skb(parent);
4870 if (before(seq, TCP_SKB_CB(skb1)->seq)) {
4871 p = &parent->rb_left;
4874 if (before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4875 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4876 /* All the bits are present. Drop. */
4877 NET_INC_STATS(sock_net(sk),
4878 LINUX_MIB_TCPOFOMERGE);
4879 tcp_drop_reason(sk, skb,
4880 SKB_DROP_REASON_TCP_OFOMERGE);
4882 tcp_dsack_set(sk, seq, end_seq);
4885 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4886 /* Partial overlap. */
4887 tcp_dsack_set(sk, seq, TCP_SKB_CB(skb1)->end_seq);
4889 /* skb's seq == skb1's seq and skb covers skb1.
4890 * Replace skb1 with skb.
4892 rb_replace_node(&skb1->rbnode, &skb->rbnode,
4893 &tp->out_of_order_queue);
4894 tcp_dsack_extend(sk,
4895 TCP_SKB_CB(skb1)->seq,
4896 TCP_SKB_CB(skb1)->end_seq);
4897 NET_INC_STATS(sock_net(sk),
4898 LINUX_MIB_TCPOFOMERGE);
4899 tcp_drop_reason(sk, skb1,
4900 SKB_DROP_REASON_TCP_OFOMERGE);
4903 } else if (tcp_ooo_try_coalesce(sk, skb1,
4904 skb, &fragstolen)) {
4907 p = &parent->rb_right;
4910 /* Insert segment into RB tree. */
4911 rb_link_node(&skb->rbnode, parent, p);
4912 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4915 /* Remove other segments covered by skb. */
4916 while ((skb1 = skb_rb_next(skb)) != NULL) {
4917 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4919 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4920 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4924 rb_erase(&skb1->rbnode, &tp->out_of_order_queue);
4925 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4926 TCP_SKB_CB(skb1)->end_seq);
4927 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4928 tcp_drop_reason(sk, skb1, SKB_DROP_REASON_TCP_OFOMERGE);
4930 /* If there is no skb after us, we are the last_skb ! */
4932 tp->ooo_last_skb = skb;
4935 if (tcp_is_sack(tp))
4936 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4939 /* For non sack flows, do not grow window to force DUPACK
4940 * and trigger fast retransmit.
4942 if (tcp_is_sack(tp))
4943 tcp_grow_window(sk, skb, false);
4945 skb_set_owner_r(skb, sk);
4949 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb,
4953 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4956 tcp_try_coalesce(sk, tail,
4957 skb, fragstolen)) ? 1 : 0;
4958 tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq);
4960 __skb_queue_tail(&sk->sk_receive_queue, skb);
4961 skb_set_owner_r(skb, sk);
4966 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4968 struct sk_buff *skb;
4976 if (size > PAGE_SIZE) {
4977 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS);
4979 data_len = npages << PAGE_SHIFT;
4980 size = data_len + (size & ~PAGE_MASK);
4982 skb = alloc_skb_with_frags(size - data_len, data_len,
4983 PAGE_ALLOC_COSTLY_ORDER,
4984 &err, sk->sk_allocation);
4988 skb_put(skb, size - data_len);
4989 skb->data_len = data_len;
4992 if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) {
4993 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP);
4997 err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size);
5001 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
5002 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
5003 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
5005 if (tcp_queue_rcv(sk, skb, &fragstolen)) {
5006 WARN_ON_ONCE(fragstolen); /* should not happen */
5018 void tcp_data_ready(struct sock *sk)
5020 if (tcp_epollin_ready(sk, sk->sk_rcvlowat) || sock_flag(sk, SOCK_DONE))
5021 sk->sk_data_ready(sk);
5024 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
5026 struct tcp_sock *tp = tcp_sk(sk);
5027 enum skb_drop_reason reason;
5031 /* If a subflow has been reset, the packet should not continue
5032 * to be processed, drop the packet.
5034 if (sk_is_mptcp(sk) && !mptcp_incoming_options(sk, skb)) {
5039 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
5044 __skb_pull(skb, tcp_hdr(skb)->doff * 4);
5046 reason = SKB_DROP_REASON_NOT_SPECIFIED;
5047 tp->rx_opt.dsack = 0;
5049 /* Queue data for delivery to the user.
5050 * Packets in sequence go to the receive queue.
5051 * Out of sequence packets to the out_of_order_queue.
5053 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
5054 if (tcp_receive_window(tp) == 0) {
5055 reason = SKB_DROP_REASON_TCP_ZEROWINDOW;
5056 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP);
5060 /* Ok. In sequence. In window. */
5062 if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) {
5063 /* TODO: maybe ratelimit these WIN 0 ACK ? */
5064 inet_csk(sk)->icsk_ack.pending |=
5065 (ICSK_ACK_NOMEM | ICSK_ACK_NOW);
5066 inet_csk_schedule_ack(sk);
5067 sk->sk_data_ready(sk);
5069 if (skb_queue_len(&sk->sk_receive_queue)) {
5070 reason = SKB_DROP_REASON_PROTO_MEM;
5071 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP);
5074 sk_forced_mem_schedule(sk, skb->truesize);
5077 eaten = tcp_queue_rcv(sk, skb, &fragstolen);
5079 tcp_event_data_recv(sk, skb);
5080 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
5083 if (!RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
5086 /* RFC5681. 4.2. SHOULD send immediate ACK, when
5087 * gap in queue is filled.
5089 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
5090 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW;
5093 if (tp->rx_opt.num_sacks)
5094 tcp_sack_remove(tp);
5096 tcp_fast_path_check(sk);
5099 kfree_skb_partial(skb, fragstolen);
5100 if (!sock_flag(sk, SOCK_DEAD))
5105 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
5106 tcp_rcv_spurious_retrans(sk, skb);
5107 /* A retransmit, 2nd most common case. Force an immediate ack. */
5108 reason = SKB_DROP_REASON_TCP_OLD_DATA;
5109 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
5110 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
5113 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
5114 inet_csk_schedule_ack(sk);
5116 tcp_drop_reason(sk, skb, reason);
5120 /* Out of window. F.e. zero window probe. */
5121 if (!before(TCP_SKB_CB(skb)->seq,
5122 tp->rcv_nxt + tcp_receive_window(tp))) {
5123 reason = SKB_DROP_REASON_TCP_OVERWINDOW;
5127 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
5128 /* Partial packet, seq < rcv_next < end_seq */
5129 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
5131 /* If window is closed, drop tail of packet. But after
5132 * remembering D-SACK for its head made in previous line.
5134 if (!tcp_receive_window(tp)) {
5135 reason = SKB_DROP_REASON_TCP_ZEROWINDOW;
5136 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP);
5142 tcp_data_queue_ofo(sk, skb);
5145 static struct sk_buff *tcp_skb_next(struct sk_buff *skb, struct sk_buff_head *list)
5148 return !skb_queue_is_last(list, skb) ? skb->next : NULL;
5150 return skb_rb_next(skb);
5153 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
5154 struct sk_buff_head *list,
5155 struct rb_root *root)
5157 struct sk_buff *next = tcp_skb_next(skb, list);
5160 __skb_unlink(skb, list);
5162 rb_erase(&skb->rbnode, root);
5165 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
5170 /* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */
5171 void tcp_rbtree_insert(struct rb_root *root, struct sk_buff *skb)
5173 struct rb_node **p = &root->rb_node;
5174 struct rb_node *parent = NULL;
5175 struct sk_buff *skb1;
5179 skb1 = rb_to_skb(parent);
5180 if (before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb1)->seq))
5181 p = &parent->rb_left;
5183 p = &parent->rb_right;
5185 rb_link_node(&skb->rbnode, parent, p);
5186 rb_insert_color(&skb->rbnode, root);
5189 /* Collapse contiguous sequence of skbs head..tail with
5190 * sequence numbers start..end.
5192 * If tail is NULL, this means until the end of the queue.
5194 * Segments with FIN/SYN are not collapsed (only because this
5198 tcp_collapse(struct sock *sk, struct sk_buff_head *list, struct rb_root *root,
5199 struct sk_buff *head, struct sk_buff *tail, u32 start, u32 end)
5201 struct sk_buff *skb = head, *n;
5202 struct sk_buff_head tmp;
5205 /* First, check that queue is collapsible and find
5206 * the point where collapsing can be useful.
5209 for (end_of_skbs = true; skb != NULL && skb != tail; skb = n) {
5210 n = tcp_skb_next(skb, list);
5212 /* No new bits? It is possible on ofo queue. */
5213 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
5214 skb = tcp_collapse_one(sk, skb, list, root);
5220 /* The first skb to collapse is:
5222 * - bloated or contains data before "start" or
5223 * overlaps to the next one and mptcp allow collapsing.
5225 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
5226 (tcp_win_from_space(sk, skb->truesize) > skb->len ||
5227 before(TCP_SKB_CB(skb)->seq, start))) {
5228 end_of_skbs = false;
5232 if (n && n != tail && mptcp_skb_can_collapse(skb, n) &&
5233 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(n)->seq) {
5234 end_of_skbs = false;
5238 /* Decided to skip this, advance start seq. */
5239 start = TCP_SKB_CB(skb)->end_seq;
5242 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
5245 __skb_queue_head_init(&tmp);
5247 while (before(start, end)) {
5248 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
5249 struct sk_buff *nskb;
5251 nskb = alloc_skb(copy, GFP_ATOMIC);
5255 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
5256 #ifdef CONFIG_TLS_DEVICE
5257 nskb->decrypted = skb->decrypted;
5259 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
5261 __skb_queue_before(list, skb, nskb);
5263 __skb_queue_tail(&tmp, nskb); /* defer rbtree insertion */
5264 skb_set_owner_r(nskb, sk);
5265 mptcp_skb_ext_move(nskb, skb);
5267 /* Copy data, releasing collapsed skbs. */
5269 int offset = start - TCP_SKB_CB(skb)->seq;
5270 int size = TCP_SKB_CB(skb)->end_seq - start;
5274 size = min(copy, size);
5275 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
5277 TCP_SKB_CB(nskb)->end_seq += size;
5281 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
5282 skb = tcp_collapse_one(sk, skb, list, root);
5285 !mptcp_skb_can_collapse(nskb, skb) ||
5286 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
5288 #ifdef CONFIG_TLS_DEVICE
5289 if (skb->decrypted != nskb->decrypted)
5296 skb_queue_walk_safe(&tmp, skb, n)
5297 tcp_rbtree_insert(root, skb);
5300 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
5301 * and tcp_collapse() them until all the queue is collapsed.
5303 static void tcp_collapse_ofo_queue(struct sock *sk)
5305 struct tcp_sock *tp = tcp_sk(sk);
5306 u32 range_truesize, sum_tiny = 0;
5307 struct sk_buff *skb, *head;
5310 skb = skb_rb_first(&tp->out_of_order_queue);
5313 tp->ooo_last_skb = skb_rb_last(&tp->out_of_order_queue);
5316 start = TCP_SKB_CB(skb)->seq;
5317 end = TCP_SKB_CB(skb)->end_seq;
5318 range_truesize = skb->truesize;
5320 for (head = skb;;) {
5321 skb = skb_rb_next(skb);
5323 /* Range is terminated when we see a gap or when
5324 * we are at the queue end.
5327 after(TCP_SKB_CB(skb)->seq, end) ||
5328 before(TCP_SKB_CB(skb)->end_seq, start)) {
5329 /* Do not attempt collapsing tiny skbs */
5330 if (range_truesize != head->truesize ||
5331 end - start >= SKB_WITH_OVERHEAD(PAGE_SIZE)) {
5332 tcp_collapse(sk, NULL, &tp->out_of_order_queue,
5333 head, skb, start, end);
5335 sum_tiny += range_truesize;
5336 if (sum_tiny > sk->sk_rcvbuf >> 3)
5342 range_truesize += skb->truesize;
5343 if (unlikely(before(TCP_SKB_CB(skb)->seq, start)))
5344 start = TCP_SKB_CB(skb)->seq;
5345 if (after(TCP_SKB_CB(skb)->end_seq, end))
5346 end = TCP_SKB_CB(skb)->end_seq;
5351 * Clean the out-of-order queue to make room.
5352 * We drop high sequences packets to :
5353 * 1) Let a chance for holes to be filled.
5354 * This means we do not drop packets from ooo queue if their sequence
5355 * is before incoming packet sequence.
5356 * 2) not add too big latencies if thousands of packets sit there.
5357 * (But if application shrinks SO_RCVBUF, we could still end up
5358 * freeing whole queue here)
5359 * 3) Drop at least 12.5 % of sk_rcvbuf to avoid malicious attacks.
5361 * Return true if queue has shrunk.
5363 static bool tcp_prune_ofo_queue(struct sock *sk, const struct sk_buff *in_skb)
5365 struct tcp_sock *tp = tcp_sk(sk);
5366 struct rb_node *node, *prev;
5367 bool pruned = false;
5370 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
5373 goal = sk->sk_rcvbuf >> 3;
5374 node = &tp->ooo_last_skb->rbnode;
5377 struct sk_buff *skb = rb_to_skb(node);
5379 /* If incoming skb would land last in ofo queue, stop pruning. */
5380 if (after(TCP_SKB_CB(in_skb)->seq, TCP_SKB_CB(skb)->seq))
5383 prev = rb_prev(node);
5384 rb_erase(node, &tp->out_of_order_queue);
5385 goal -= skb->truesize;
5386 tcp_drop_reason(sk, skb, SKB_DROP_REASON_TCP_OFO_QUEUE_PRUNE);
5387 tp->ooo_last_skb = rb_to_skb(prev);
5388 if (!prev || goal <= 0) {
5389 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf &&
5390 !tcp_under_memory_pressure(sk))
5392 goal = sk->sk_rcvbuf >> 3;
5398 NET_INC_STATS(sock_net(sk), LINUX_MIB_OFOPRUNED);
5399 /* Reset SACK state. A conforming SACK implementation will
5400 * do the same at a timeout based retransmit. When a connection
5401 * is in a sad state like this, we care only about integrity
5402 * of the connection not performance.
5404 if (tp->rx_opt.sack_ok)
5405 tcp_sack_reset(&tp->rx_opt);
5410 /* Reduce allocated memory if we can, trying to get
5411 * the socket within its memory limits again.
5413 * Return less than zero if we should start dropping frames
5414 * until the socket owning process reads some of the data
5415 * to stabilize the situation.
5417 static int tcp_prune_queue(struct sock *sk, const struct sk_buff *in_skb)
5419 struct tcp_sock *tp = tcp_sk(sk);
5421 NET_INC_STATS(sock_net(sk), LINUX_MIB_PRUNECALLED);
5423 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
5424 tcp_clamp_window(sk);
5425 else if (tcp_under_memory_pressure(sk))
5426 tcp_adjust_rcv_ssthresh(sk);
5428 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5431 tcp_collapse_ofo_queue(sk);
5432 if (!skb_queue_empty(&sk->sk_receive_queue))
5433 tcp_collapse(sk, &sk->sk_receive_queue, NULL,
5434 skb_peek(&sk->sk_receive_queue),
5436 tp->copied_seq, tp->rcv_nxt);
5438 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5441 /* Collapsing did not help, destructive actions follow.
5442 * This must not ever occur. */
5444 tcp_prune_ofo_queue(sk, in_skb);
5446 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5449 /* If we are really being abused, tell the caller to silently
5450 * drop receive data on the floor. It will get retransmitted
5451 * and hopefully then we'll have sufficient space.
5453 NET_INC_STATS(sock_net(sk), LINUX_MIB_RCVPRUNED);
5455 /* Massive buffer overcommit. */
5460 static bool tcp_should_expand_sndbuf(struct sock *sk)
5462 const struct tcp_sock *tp = tcp_sk(sk);
5464 /* If the user specified a specific send buffer setting, do
5467 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
5470 /* If we are under global TCP memory pressure, do not expand. */
5471 if (tcp_under_memory_pressure(sk)) {
5472 int unused_mem = sk_unused_reserved_mem(sk);
5474 /* Adjust sndbuf according to reserved mem. But make sure
5475 * it never goes below SOCK_MIN_SNDBUF.
5476 * See sk_stream_moderate_sndbuf() for more details.
5478 if (unused_mem > SOCK_MIN_SNDBUF)
5479 WRITE_ONCE(sk->sk_sndbuf, unused_mem);
5484 /* If we are under soft global TCP memory pressure, do not expand. */
5485 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
5488 /* If we filled the congestion window, do not expand. */
5489 if (tcp_packets_in_flight(tp) >= tcp_snd_cwnd(tp))
5495 static void tcp_new_space(struct sock *sk)
5497 struct tcp_sock *tp = tcp_sk(sk);
5499 if (tcp_should_expand_sndbuf(sk)) {
5500 tcp_sndbuf_expand(sk);
5501 tp->snd_cwnd_stamp = tcp_jiffies32;
5504 INDIRECT_CALL_1(sk->sk_write_space, sk_stream_write_space, sk);
5507 /* Caller made space either from:
5508 * 1) Freeing skbs in rtx queues (after tp->snd_una has advanced)
5509 * 2) Sent skbs from output queue (and thus advancing tp->snd_nxt)
5511 * We might be able to generate EPOLLOUT to the application if:
5512 * 1) Space consumed in output/rtx queues is below sk->sk_sndbuf/2
5513 * 2) notsent amount (tp->write_seq - tp->snd_nxt) became
5514 * small enough that tcp_stream_memory_free() decides it
5515 * is time to generate EPOLLOUT.
5517 void tcp_check_space(struct sock *sk)
5519 /* pairs with tcp_poll() */
5521 if (sk->sk_socket &&
5522 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
5524 if (!test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
5525 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
5529 static inline void tcp_data_snd_check(struct sock *sk)
5531 tcp_push_pending_frames(sk);
5532 tcp_check_space(sk);
5536 * Check if sending an ack is needed.
5538 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
5540 struct tcp_sock *tp = tcp_sk(sk);
5541 unsigned long rtt, delay;
5543 /* More than one full frame received... */
5544 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
5545 /* ... and right edge of window advances far enough.
5546 * (tcp_recvmsg() will send ACK otherwise).
5547 * If application uses SO_RCVLOWAT, we want send ack now if
5548 * we have not received enough bytes to satisfy the condition.
5550 (tp->rcv_nxt - tp->copied_seq < sk->sk_rcvlowat ||
5551 __tcp_select_window(sk) >= tp->rcv_wnd)) ||
5552 /* We ACK each frame or... */
5553 tcp_in_quickack_mode(sk) ||
5554 /* Protocol state mandates a one-time immediate ACK */
5555 inet_csk(sk)->icsk_ack.pending & ICSK_ACK_NOW) {
5561 if (!ofo_possible || RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
5562 tcp_send_delayed_ack(sk);
5566 if (!tcp_is_sack(tp) ||
5567 tp->compressed_ack >= READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_nr))
5570 if (tp->compressed_ack_rcv_nxt != tp->rcv_nxt) {
5571 tp->compressed_ack_rcv_nxt = tp->rcv_nxt;
5572 tp->dup_ack_counter = 0;
5574 if (tp->dup_ack_counter < TCP_FASTRETRANS_THRESH) {
5575 tp->dup_ack_counter++;
5578 tp->compressed_ack++;
5579 if (hrtimer_is_queued(&tp->compressed_ack_timer))
5582 /* compress ack timer : 5 % of rtt, but no more than tcp_comp_sack_delay_ns */
5584 rtt = tp->rcv_rtt_est.rtt_us;
5585 if (tp->srtt_us && tp->srtt_us < rtt)
5588 delay = min_t(unsigned long,
5589 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_delay_ns),
5590 rtt * (NSEC_PER_USEC >> 3)/20);
5592 hrtimer_start_range_ns(&tp->compressed_ack_timer, ns_to_ktime(delay),
5593 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_slack_ns),
5594 HRTIMER_MODE_REL_PINNED_SOFT);
5597 static inline void tcp_ack_snd_check(struct sock *sk)
5599 if (!inet_csk_ack_scheduled(sk)) {
5600 /* We sent a data segment already. */
5603 __tcp_ack_snd_check(sk, 1);
5607 * This routine is only called when we have urgent data
5608 * signaled. Its the 'slow' part of tcp_urg. It could be
5609 * moved inline now as tcp_urg is only called from one
5610 * place. We handle URGent data wrong. We have to - as
5611 * BSD still doesn't use the correction from RFC961.
5612 * For 1003.1g we should support a new option TCP_STDURG to permit
5613 * either form (or just set the sysctl tcp_stdurg).
5616 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
5618 struct tcp_sock *tp = tcp_sk(sk);
5619 u32 ptr = ntohs(th->urg_ptr);
5621 if (ptr && !READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_stdurg))
5623 ptr += ntohl(th->seq);
5625 /* Ignore urgent data that we've already seen and read. */
5626 if (after(tp->copied_seq, ptr))
5629 /* Do not replay urg ptr.
5631 * NOTE: interesting situation not covered by specs.
5632 * Misbehaving sender may send urg ptr, pointing to segment,
5633 * which we already have in ofo queue. We are not able to fetch
5634 * such data and will stay in TCP_URG_NOTYET until will be eaten
5635 * by recvmsg(). Seems, we are not obliged to handle such wicked
5636 * situations. But it is worth to think about possibility of some
5637 * DoSes using some hypothetical application level deadlock.
5639 if (before(ptr, tp->rcv_nxt))
5642 /* Do we already have a newer (or duplicate) urgent pointer? */
5643 if (tp->urg_data && !after(ptr, tp->urg_seq))
5646 /* Tell the world about our new urgent pointer. */
5649 /* We may be adding urgent data when the last byte read was
5650 * urgent. To do this requires some care. We cannot just ignore
5651 * tp->copied_seq since we would read the last urgent byte again
5652 * as data, nor can we alter copied_seq until this data arrives
5653 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5655 * NOTE. Double Dutch. Rendering to plain English: author of comment
5656 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5657 * and expect that both A and B disappear from stream. This is _wrong_.
5658 * Though this happens in BSD with high probability, this is occasional.
5659 * Any application relying on this is buggy. Note also, that fix "works"
5660 * only in this artificial test. Insert some normal data between A and B and we will
5661 * decline of BSD again. Verdict: it is better to remove to trap
5664 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5665 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5666 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5668 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5669 __skb_unlink(skb, &sk->sk_receive_queue);
5674 WRITE_ONCE(tp->urg_data, TCP_URG_NOTYET);
5675 WRITE_ONCE(tp->urg_seq, ptr);
5677 /* Disable header prediction. */
5681 /* This is the 'fast' part of urgent handling. */
5682 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
5684 struct tcp_sock *tp = tcp_sk(sk);
5686 /* Check if we get a new urgent pointer - normally not. */
5687 if (unlikely(th->urg))
5688 tcp_check_urg(sk, th);
5690 /* Do we wait for any urgent data? - normally not... */
5691 if (unlikely(tp->urg_data == TCP_URG_NOTYET)) {
5692 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5695 /* Is the urgent pointer pointing into this packet? */
5696 if (ptr < skb->len) {
5698 if (skb_copy_bits(skb, ptr, &tmp, 1))
5700 WRITE_ONCE(tp->urg_data, TCP_URG_VALID | tmp);
5701 if (!sock_flag(sk, SOCK_DEAD))
5702 sk->sk_data_ready(sk);
5707 /* Accept RST for rcv_nxt - 1 after a FIN.
5708 * When tcp connections are abruptly terminated from Mac OSX (via ^C), a
5709 * FIN is sent followed by a RST packet. The RST is sent with the same
5710 * sequence number as the FIN, and thus according to RFC 5961 a challenge
5711 * ACK should be sent. However, Mac OSX rate limits replies to challenge
5712 * ACKs on the closed socket. In addition middleboxes can drop either the
5713 * challenge ACK or a subsequent RST.
5715 static bool tcp_reset_check(const struct sock *sk, const struct sk_buff *skb)
5717 const struct tcp_sock *tp = tcp_sk(sk);
5719 return unlikely(TCP_SKB_CB(skb)->seq == (tp->rcv_nxt - 1) &&
5720 (1 << sk->sk_state) & (TCPF_CLOSE_WAIT | TCPF_LAST_ACK |
5724 /* Does PAWS and seqno based validation of an incoming segment, flags will
5725 * play significant role here.
5727 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5728 const struct tcphdr *th, int syn_inerr)
5730 struct tcp_sock *tp = tcp_sk(sk);
5733 /* RFC1323: H1. Apply PAWS check first. */
5734 if (tcp_fast_parse_options(sock_net(sk), skb, th, tp) &&
5735 tp->rx_opt.saw_tstamp &&
5736 tcp_paws_discard(sk, skb)) {
5738 if (unlikely(th->syn))
5740 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5741 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5742 LINUX_MIB_TCPACKSKIPPEDPAWS,
5743 &tp->last_oow_ack_time))
5744 tcp_send_dupack(sk, skb);
5745 SKB_DR_SET(reason, TCP_RFC7323_PAWS);
5748 /* Reset is accepted even if it did not pass PAWS. */
5751 /* Step 1: check sequence number */
5752 reason = tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
5754 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5755 * (RST) segments are validated by checking their SEQ-fields."
5756 * And page 69: "If an incoming segment is not acceptable,
5757 * an acknowledgment should be sent in reply (unless the RST
5758 * bit is set, if so drop the segment and return)".
5763 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5764 LINUX_MIB_TCPACKSKIPPEDSEQ,
5765 &tp->last_oow_ack_time))
5766 tcp_send_dupack(sk, skb);
5767 } else if (tcp_reset_check(sk, skb)) {
5773 /* Step 2: check RST bit */
5775 /* RFC 5961 3.2 (extend to match against (RCV.NXT - 1) after a
5776 * FIN and SACK too if available):
5777 * If seq num matches RCV.NXT or (RCV.NXT - 1) after a FIN, or
5778 * the right-most SACK block,
5780 * RESET the connection
5782 * Send a challenge ACK
5784 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt ||
5785 tcp_reset_check(sk, skb))
5788 if (tcp_is_sack(tp) && tp->rx_opt.num_sacks > 0) {
5789 struct tcp_sack_block *sp = &tp->selective_acks[0];
5790 int max_sack = sp[0].end_seq;
5793 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;
5795 max_sack = after(sp[this_sack].end_seq,
5797 sp[this_sack].end_seq : max_sack;
5800 if (TCP_SKB_CB(skb)->seq == max_sack)
5804 /* Disable TFO if RST is out-of-order
5805 * and no data has been received
5806 * for current active TFO socket
5808 if (tp->syn_fastopen && !tp->data_segs_in &&
5809 sk->sk_state == TCP_ESTABLISHED)
5810 tcp_fastopen_active_disable(sk);
5811 tcp_send_challenge_ack(sk);
5812 SKB_DR_SET(reason, TCP_RESET);
5816 /* step 3: check security and precedence [ignored] */
5818 /* step 4: Check for a SYN
5819 * RFC 5961 4.2 : Send a challenge ack
5824 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5825 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5826 tcp_send_challenge_ack(sk);
5827 SKB_DR_SET(reason, TCP_INVALID_SYN);
5831 bpf_skops_parse_hdr(sk, skb);
5836 tcp_drop_reason(sk, skb, reason);
5846 * TCP receive function for the ESTABLISHED state.
5848 * It is split into a fast path and a slow path. The fast path is
5850 * - A zero window was announced from us - zero window probing
5851 * is only handled properly in the slow path.
5852 * - Out of order segments arrived.
5853 * - Urgent data is expected.
5854 * - There is no buffer space left
5855 * - Unexpected TCP flags/window values/header lengths are received
5856 * (detected by checking the TCP header against pred_flags)
5857 * - Data is sent in both directions. Fast path only supports pure senders
5858 * or pure receivers (this means either the sequence number or the ack
5859 * value must stay constant)
5860 * - Unexpected TCP option.
5862 * When these conditions are not satisfied it drops into a standard
5863 * receive procedure patterned after RFC793 to handle all cases.
5864 * The first three cases are guaranteed by proper pred_flags setting,
5865 * the rest is checked inline. Fast processing is turned on in
5866 * tcp_data_queue when everything is OK.
5868 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb)
5870 enum skb_drop_reason reason = SKB_DROP_REASON_NOT_SPECIFIED;
5871 const struct tcphdr *th = (const struct tcphdr *)skb->data;
5872 struct tcp_sock *tp = tcp_sk(sk);
5873 unsigned int len = skb->len;
5875 /* TCP congestion window tracking */
5876 trace_tcp_probe(sk, skb);
5878 tcp_mstamp_refresh(tp);
5879 if (unlikely(!rcu_access_pointer(sk->sk_rx_dst)))
5880 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
5882 * Header prediction.
5883 * The code loosely follows the one in the famous
5884 * "30 instruction TCP receive" Van Jacobson mail.
5886 * Van's trick is to deposit buffers into socket queue
5887 * on a device interrupt, to call tcp_recv function
5888 * on the receive process context and checksum and copy
5889 * the buffer to user space. smart...
5891 * Our current scheme is not silly either but we take the
5892 * extra cost of the net_bh soft interrupt processing...
5893 * We do checksum and copy also but from device to kernel.
5896 tp->rx_opt.saw_tstamp = 0;
5898 /* pred_flags is 0xS?10 << 16 + snd_wnd
5899 * if header_prediction is to be made
5900 * 'S' will always be tp->tcp_header_len >> 2
5901 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5902 * turn it off (when there are holes in the receive
5903 * space for instance)
5904 * PSH flag is ignored.
5907 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5908 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5909 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5910 int tcp_header_len = tp->tcp_header_len;
5912 /* Timestamp header prediction: tcp_header_len
5913 * is automatically equal to th->doff*4 due to pred_flags
5917 /* Check timestamp */
5918 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5919 /* No? Slow path! */
5920 if (!tcp_parse_aligned_timestamp(tp, th))
5923 /* If PAWS failed, check it more carefully in slow path */
5924 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5927 /* DO NOT update ts_recent here, if checksum fails
5928 * and timestamp was corrupted part, it will result
5929 * in a hung connection since we will drop all
5930 * future packets due to the PAWS test.
5934 if (len <= tcp_header_len) {
5935 /* Bulk data transfer: sender */
5936 if (len == tcp_header_len) {
5937 /* Predicted packet is in window by definition.
5938 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5939 * Hence, check seq<=rcv_wup reduces to:
5941 if (tcp_header_len ==
5942 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5943 tp->rcv_nxt == tp->rcv_wup)
5944 tcp_store_ts_recent(tp);
5946 /* We know that such packets are checksummed
5949 tcp_ack(sk, skb, 0);
5951 tcp_data_snd_check(sk);
5952 /* When receiving pure ack in fast path, update
5953 * last ts ecr directly instead of calling
5954 * tcp_rcv_rtt_measure_ts()
5956 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr;
5958 } else { /* Header too small */
5959 reason = SKB_DROP_REASON_PKT_TOO_SMALL;
5960 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5965 bool fragstolen = false;
5967 if (tcp_checksum_complete(skb))
5970 if ((int)skb->truesize > sk->sk_forward_alloc)
5973 /* Predicted packet is in window by definition.
5974 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5975 * Hence, check seq<=rcv_wup reduces to:
5977 if (tcp_header_len ==
5978 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5979 tp->rcv_nxt == tp->rcv_wup)
5980 tcp_store_ts_recent(tp);
5982 tcp_rcv_rtt_measure_ts(sk, skb);
5984 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPHITS);
5986 /* Bulk data transfer: receiver */
5988 __skb_pull(skb, tcp_header_len);
5989 eaten = tcp_queue_rcv(sk, skb, &fragstolen);
5991 tcp_event_data_recv(sk, skb);
5993 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5994 /* Well, only one small jumplet in fast path... */
5995 tcp_ack(sk, skb, FLAG_DATA);
5996 tcp_data_snd_check(sk);
5997 if (!inet_csk_ack_scheduled(sk))
6000 tcp_update_wl(tp, TCP_SKB_CB(skb)->seq);
6003 __tcp_ack_snd_check(sk, 0);
6006 kfree_skb_partial(skb, fragstolen);
6013 if (len < (th->doff << 2) || tcp_checksum_complete(skb))
6016 if (!th->ack && !th->rst && !th->syn) {
6017 reason = SKB_DROP_REASON_TCP_FLAGS;
6022 * Standard slow path.
6025 if (!tcp_validate_incoming(sk, skb, th, 1))
6029 reason = tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT);
6030 if ((int)reason < 0) {
6034 tcp_rcv_rtt_measure_ts(sk, skb);
6036 /* Process urgent data. */
6037 tcp_urg(sk, skb, th);
6039 /* step 7: process the segment text */
6040 tcp_data_queue(sk, skb);
6042 tcp_data_snd_check(sk);
6043 tcp_ack_snd_check(sk);
6047 reason = SKB_DROP_REASON_TCP_CSUM;
6048 trace_tcp_bad_csum(skb);
6049 TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS);
6050 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
6053 tcp_drop_reason(sk, skb, reason);
6055 EXPORT_SYMBOL(tcp_rcv_established);
6057 void tcp_init_transfer(struct sock *sk, int bpf_op, struct sk_buff *skb)
6059 struct inet_connection_sock *icsk = inet_csk(sk);
6060 struct tcp_sock *tp = tcp_sk(sk);
6063 icsk->icsk_af_ops->rebuild_header(sk);
6064 tcp_init_metrics(sk);
6066 /* Initialize the congestion window to start the transfer.
6067 * Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
6068 * retransmitted. In light of RFC6298 more aggressive 1sec
6069 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
6070 * retransmission has occurred.
6072 if (tp->total_retrans > 1 && tp->undo_marker)
6073 tcp_snd_cwnd_set(tp, 1);
6075 tcp_snd_cwnd_set(tp, tcp_init_cwnd(tp, __sk_dst_get(sk)));
6076 tp->snd_cwnd_stamp = tcp_jiffies32;
6078 bpf_skops_established(sk, bpf_op, skb);
6079 /* Initialize congestion control unless BPF initialized it already: */
6080 if (!icsk->icsk_ca_initialized)
6081 tcp_init_congestion_control(sk);
6082 tcp_init_buffer_space(sk);
6085 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
6087 struct tcp_sock *tp = tcp_sk(sk);
6088 struct inet_connection_sock *icsk = inet_csk(sk);
6090 tcp_set_state(sk, TCP_ESTABLISHED);
6091 icsk->icsk_ack.lrcvtime = tcp_jiffies32;
6094 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
6095 security_inet_conn_established(sk, skb);
6096 sk_mark_napi_id(sk, skb);
6099 tcp_init_transfer(sk, BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB, skb);
6101 /* Prevent spurious tcp_cwnd_restart() on first data
6104 tp->lsndtime = tcp_jiffies32;
6106 if (sock_flag(sk, SOCK_KEEPOPEN))
6107 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
6109 if (!tp->rx_opt.snd_wscale)
6110 __tcp_fast_path_on(tp, tp->snd_wnd);
6115 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
6116 struct tcp_fastopen_cookie *cookie)
6118 struct tcp_sock *tp = tcp_sk(sk);
6119 struct sk_buff *data = tp->syn_data ? tcp_rtx_queue_head(sk) : NULL;
6120 u16 mss = tp->rx_opt.mss_clamp, try_exp = 0;
6121 bool syn_drop = false;
6123 if (mss == tp->rx_opt.user_mss) {
6124 struct tcp_options_received opt;
6126 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
6127 tcp_clear_options(&opt);
6128 opt.user_mss = opt.mss_clamp = 0;
6129 tcp_parse_options(sock_net(sk), synack, &opt, 0, NULL);
6130 mss = opt.mss_clamp;
6133 if (!tp->syn_fastopen) {
6134 /* Ignore an unsolicited cookie */
6136 } else if (tp->total_retrans) {
6137 /* SYN timed out and the SYN-ACK neither has a cookie nor
6138 * acknowledges data. Presumably the remote received only
6139 * the retransmitted (regular) SYNs: either the original
6140 * SYN-data or the corresponding SYN-ACK was dropped.
6142 syn_drop = (cookie->len < 0 && data);
6143 } else if (cookie->len < 0 && !tp->syn_data) {
6144 /* We requested a cookie but didn't get it. If we did not use
6145 * the (old) exp opt format then try so next time (try_exp=1).
6146 * Otherwise we go back to use the RFC7413 opt (try_exp=2).
6148 try_exp = tp->syn_fastopen_exp ? 2 : 1;
6151 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp);
6153 if (data) { /* Retransmit unacked data in SYN */
6154 if (tp->total_retrans)
6155 tp->fastopen_client_fail = TFO_SYN_RETRANSMITTED;
6157 tp->fastopen_client_fail = TFO_DATA_NOT_ACKED;
6158 skb_rbtree_walk_from(data)
6159 tcp_mark_skb_lost(sk, data);
6160 tcp_xmit_retransmit_queue(sk);
6161 NET_INC_STATS(sock_net(sk),
6162 LINUX_MIB_TCPFASTOPENACTIVEFAIL);
6165 tp->syn_data_acked = tp->syn_data;
6166 if (tp->syn_data_acked) {
6167 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
6168 /* SYN-data is counted as two separate packets in tcp_ack() */
6169 if (tp->delivered > 1)
6173 tcp_fastopen_add_skb(sk, synack);
6178 static void smc_check_reset_syn(struct tcp_sock *tp)
6180 #if IS_ENABLED(CONFIG_SMC)
6181 if (static_branch_unlikely(&tcp_have_smc)) {
6182 if (tp->syn_smc && !tp->rx_opt.smc_ok)
6188 static void tcp_try_undo_spurious_syn(struct sock *sk)
6190 struct tcp_sock *tp = tcp_sk(sk);
6193 /* undo_marker is set when SYN or SYNACK times out. The timeout is
6194 * spurious if the ACK's timestamp option echo value matches the
6195 * original SYN timestamp.
6197 syn_stamp = tp->retrans_stamp;
6198 if (tp->undo_marker && syn_stamp && tp->rx_opt.saw_tstamp &&
6199 syn_stamp == tp->rx_opt.rcv_tsecr)
6200 tp->undo_marker = 0;
6203 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
6204 const struct tcphdr *th)
6206 struct inet_connection_sock *icsk = inet_csk(sk);
6207 struct tcp_sock *tp = tcp_sk(sk);
6208 struct tcp_fastopen_cookie foc = { .len = -1 };
6209 int saved_clamp = tp->rx_opt.mss_clamp;
6213 tcp_parse_options(sock_net(sk), skb, &tp->rx_opt, 0, &foc);
6214 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
6215 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
6219 * "If the state is SYN-SENT then
6220 * first check the ACK bit
6221 * If the ACK bit is set
6222 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
6223 * a reset (unless the RST bit is set, if so drop
6224 * the segment and return)"
6226 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
6227 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
6228 /* Previous FIN/ACK or RST/ACK might be ignored. */
6229 if (icsk->icsk_retransmits == 0)
6230 inet_csk_reset_xmit_timer(sk,
6232 TCP_TIMEOUT_MIN, TCP_RTO_MAX);
6233 goto reset_and_undo;
6236 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
6237 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
6238 tcp_time_stamp(tp))) {
6239 NET_INC_STATS(sock_net(sk),
6240 LINUX_MIB_PAWSACTIVEREJECTED);
6241 goto reset_and_undo;
6244 /* Now ACK is acceptable.
6246 * "If the RST bit is set
6247 * If the ACK was acceptable then signal the user "error:
6248 * connection reset", drop the segment, enter CLOSED state,
6249 * delete TCB, and return."
6260 * "fifth, if neither of the SYN or RST bits is set then
6261 * drop the segment and return."
6267 SKB_DR_SET(reason, TCP_FLAGS);
6268 goto discard_and_undo;
6271 * "If the SYN bit is on ...
6272 * are acceptable then ...
6273 * (our SYN has been ACKed), change the connection
6274 * state to ESTABLISHED..."
6277 tcp_ecn_rcv_synack(tp, th);
6279 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
6280 tcp_try_undo_spurious_syn(sk);
6281 tcp_ack(sk, skb, FLAG_SLOWPATH);
6283 /* Ok.. it's good. Set up sequence numbers and
6284 * move to established.
6286 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1);
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);
6294 if (!tp->rx_opt.wscale_ok) {
6295 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
6296 tp->window_clamp = min(tp->window_clamp, 65535U);
6299 if (tp->rx_opt.saw_tstamp) {
6300 tp->rx_opt.tstamp_ok = 1;
6301 tp->tcp_header_len =
6302 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
6303 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
6304 tcp_store_ts_recent(tp);
6306 tp->tcp_header_len = sizeof(struct tcphdr);
6309 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
6310 tcp_initialize_rcv_mss(sk);
6312 /* Remember, tcp_poll() does not lock socket!
6313 * Change state from SYN-SENT only after copied_seq
6314 * is initialized. */
6315 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6317 smc_check_reset_syn(tp);
6321 tcp_finish_connect(sk, skb);
6323 fastopen_fail = (tp->syn_fastopen || tp->syn_data) &&
6324 tcp_rcv_fastopen_synack(sk, skb, &foc);
6326 if (!sock_flag(sk, SOCK_DEAD)) {
6327 sk->sk_state_change(sk);
6328 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
6332 if (sk->sk_write_pending ||
6333 READ_ONCE(icsk->icsk_accept_queue.rskq_defer_accept) ||
6334 inet_csk_in_pingpong_mode(sk)) {
6335 /* Save one ACK. Data will be ready after
6336 * several ticks, if write_pending is set.
6338 * It may be deleted, but with this feature tcpdumps
6339 * look so _wonderfully_ clever, that I was not able
6340 * to stand against the temptation 8) --ANK
6342 inet_csk_schedule_ack(sk);
6343 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
6344 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
6345 TCP_DELACK_MAX, TCP_RTO_MAX);
6352 /* No ACK in the segment */
6356 * "If the RST bit is set
6358 * Otherwise (no ACK) drop the segment and return."
6360 SKB_DR_SET(reason, TCP_RESET);
6361 goto discard_and_undo;
6365 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
6366 tcp_paws_reject(&tp->rx_opt, 0)) {
6367 SKB_DR_SET(reason, TCP_RFC7323_PAWS);
6368 goto discard_and_undo;
6371 /* We see SYN without ACK. It is attempt of
6372 * simultaneous connect with crossed SYNs.
6373 * Particularly, it can be connect to self.
6375 tcp_set_state(sk, TCP_SYN_RECV);
6377 if (tp->rx_opt.saw_tstamp) {
6378 tp->rx_opt.tstamp_ok = 1;
6379 tcp_store_ts_recent(tp);
6380 tp->tcp_header_len =
6381 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
6383 tp->tcp_header_len = sizeof(struct tcphdr);
6386 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1);
6387 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6388 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
6390 /* RFC1323: The window in SYN & SYN/ACK segments is
6393 tp->snd_wnd = ntohs(th->window);
6394 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
6395 tp->max_window = tp->snd_wnd;
6397 tcp_ecn_rcv_syn(tp, th);
6400 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
6401 tcp_initialize_rcv_mss(sk);
6403 tcp_send_synack(sk);
6405 /* Note, we could accept data and URG from this segment.
6406 * There are no obstacles to make this (except that we must
6407 * either change tcp_recvmsg() to prevent it from returning data
6408 * before 3WHS completes per RFC793, or employ TCP Fast Open).
6410 * However, if we ignore data in ACKless segments sometimes,
6411 * we have no reasons to accept it sometimes.
6412 * Also, seems the code doing it in step6 of tcp_rcv_state_process
6413 * is not flawless. So, discard packet for sanity.
6414 * Uncomment this return to process the data.
6421 /* "fifth, if neither of the SYN or RST bits is set then
6422 * drop the segment and return."
6426 tcp_clear_options(&tp->rx_opt);
6427 tp->rx_opt.mss_clamp = saved_clamp;
6428 tcp_drop_reason(sk, skb, reason);
6432 tcp_clear_options(&tp->rx_opt);
6433 tp->rx_opt.mss_clamp = saved_clamp;
6437 static void tcp_rcv_synrecv_state_fastopen(struct sock *sk)
6439 struct request_sock *req;
6441 /* If we are still handling the SYNACK RTO, see if timestamp ECR allows
6442 * undo. If peer SACKs triggered fast recovery, we can't undo here.
6444 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
6445 tcp_try_undo_loss(sk, false);
6447 /* Reset rtx states to prevent spurious retransmits_timed_out() */
6448 tcp_sk(sk)->retrans_stamp = 0;
6449 inet_csk(sk)->icsk_retransmits = 0;
6451 /* Once we leave TCP_SYN_RECV or TCP_FIN_WAIT_1,
6452 * we no longer need req so release it.
6454 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk,
6455 lockdep_sock_is_held(sk));
6456 reqsk_fastopen_remove(sk, req, false);
6458 /* Re-arm the timer because data may have been sent out.
6459 * This is similar to the regular data transmission case
6460 * when new data has just been ack'ed.
6462 * (TFO) - we could try to be more aggressive and
6463 * retransmitting any data sooner based on when they
6470 * This function implements the receiving procedure of RFC 793 for
6471 * all states except ESTABLISHED and TIME_WAIT.
6472 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
6473 * address independent.
6476 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb)
6478 struct tcp_sock *tp = tcp_sk(sk);
6479 struct inet_connection_sock *icsk = inet_csk(sk);
6480 const struct tcphdr *th = tcp_hdr(skb);
6481 struct request_sock *req;
6486 switch (sk->sk_state) {
6488 SKB_DR_SET(reason, TCP_CLOSE);
6496 SKB_DR_SET(reason, TCP_RESET);
6501 SKB_DR_SET(reason, TCP_FLAGS);
6504 /* It is possible that we process SYN packets from backlog,
6505 * so we need to make sure to disable BH and RCU right there.
6509 acceptable = icsk->icsk_af_ops->conn_request(sk, skb) >= 0;
6518 SKB_DR_SET(reason, TCP_FLAGS);
6522 tp->rx_opt.saw_tstamp = 0;
6523 tcp_mstamp_refresh(tp);
6524 queued = tcp_rcv_synsent_state_process(sk, skb, th);
6528 /* Do step6 onward by hand. */
6529 tcp_urg(sk, skb, th);
6531 tcp_data_snd_check(sk);
6535 tcp_mstamp_refresh(tp);
6536 tp->rx_opt.saw_tstamp = 0;
6537 req = rcu_dereference_protected(tp->fastopen_rsk,
6538 lockdep_sock_is_held(sk));
6542 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
6543 sk->sk_state != TCP_FIN_WAIT1);
6545 if (!tcp_check_req(sk, skb, req, true, &req_stolen)) {
6546 SKB_DR_SET(reason, TCP_FASTOPEN);
6551 if (!th->ack && !th->rst && !th->syn) {
6552 SKB_DR_SET(reason, TCP_FLAGS);
6555 if (!tcp_validate_incoming(sk, skb, th, 0))
6558 /* step 5: check the ACK field */
6559 acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
6560 FLAG_UPDATE_TS_RECENT |
6561 FLAG_NO_CHALLENGE_ACK) > 0;
6564 if (sk->sk_state == TCP_SYN_RECV)
6565 return 1; /* send one RST */
6566 tcp_send_challenge_ack(sk);
6567 SKB_DR_SET(reason, TCP_OLD_ACK);
6570 switch (sk->sk_state) {
6572 tp->delivered++; /* SYN-ACK delivery isn't tracked in tcp_ack */
6574 tcp_synack_rtt_meas(sk, req);
6577 tcp_rcv_synrecv_state_fastopen(sk);
6579 tcp_try_undo_spurious_syn(sk);
6580 tp->retrans_stamp = 0;
6581 tcp_init_transfer(sk, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB,
6583 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6586 tcp_set_state(sk, TCP_ESTABLISHED);
6587 sk->sk_state_change(sk);
6589 /* Note, that this wakeup is only for marginal crossed SYN case.
6590 * Passively open sockets are not waked up, because
6591 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
6594 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
6596 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
6597 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
6598 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
6600 if (tp->rx_opt.tstamp_ok)
6601 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
6603 if (!inet_csk(sk)->icsk_ca_ops->cong_control)
6604 tcp_update_pacing_rate(sk);
6606 /* Prevent spurious tcp_cwnd_restart() on first data packet */
6607 tp->lsndtime = tcp_jiffies32;
6609 tcp_initialize_rcv_mss(sk);
6610 tcp_fast_path_on(tp);
6613 case TCP_FIN_WAIT1: {
6617 tcp_rcv_synrecv_state_fastopen(sk);
6619 if (tp->snd_una != tp->write_seq)
6622 tcp_set_state(sk, TCP_FIN_WAIT2);
6623 WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | SEND_SHUTDOWN);
6627 if (!sock_flag(sk, SOCK_DEAD)) {
6628 /* Wake up lingering close() */
6629 sk->sk_state_change(sk);
6633 if (READ_ONCE(tp->linger2) < 0) {
6635 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6638 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6639 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6640 /* Receive out of order FIN after close() */
6641 if (tp->syn_fastopen && th->fin)
6642 tcp_fastopen_active_disable(sk);
6644 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6648 tmo = tcp_fin_time(sk);
6649 if (tmo > TCP_TIMEWAIT_LEN) {
6650 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
6651 } else if (th->fin || sock_owned_by_user(sk)) {
6652 /* Bad case. We could lose such FIN otherwise.
6653 * It is not a big problem, but it looks confusing
6654 * and not so rare event. We still can lose it now,
6655 * if it spins in bh_lock_sock(), but it is really
6658 inet_csk_reset_keepalive_timer(sk, tmo);
6660 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
6667 if (tp->snd_una == tp->write_seq) {
6668 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
6674 if (tp->snd_una == tp->write_seq) {
6675 tcp_update_metrics(sk);
6682 /* step 6: check the URG bit */
6683 tcp_urg(sk, skb, th);
6685 /* step 7: process the segment text */
6686 switch (sk->sk_state) {
6687 case TCP_CLOSE_WAIT:
6690 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
6691 /* If a subflow has been reset, the packet should not
6692 * continue to be processed, drop the packet.
6694 if (sk_is_mptcp(sk) && !mptcp_incoming_options(sk, skb))
6701 /* RFC 793 says to queue data in these states,
6702 * RFC 1122 says we MUST send a reset.
6703 * BSD 4.4 also does reset.
6705 if (sk->sk_shutdown & RCV_SHUTDOWN) {
6706 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6707 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6708 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6714 case TCP_ESTABLISHED:
6715 tcp_data_queue(sk, skb);
6720 /* tcp_data could move socket to TIME-WAIT */
6721 if (sk->sk_state != TCP_CLOSE) {
6722 tcp_data_snd_check(sk);
6723 tcp_ack_snd_check(sk);
6728 tcp_drop_reason(sk, skb, reason);
6736 EXPORT_SYMBOL(tcp_rcv_state_process);
6738 static inline void pr_drop_req(struct request_sock *req, __u16 port, int family)
6740 struct inet_request_sock *ireq = inet_rsk(req);
6742 if (family == AF_INET)
6743 net_dbg_ratelimited("drop open request from %pI4/%u\n",
6744 &ireq->ir_rmt_addr, port);
6745 #if IS_ENABLED(CONFIG_IPV6)
6746 else if (family == AF_INET6)
6747 net_dbg_ratelimited("drop open request from %pI6/%u\n",
6748 &ireq->ir_v6_rmt_addr, port);
6752 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
6754 * If we receive a SYN packet with these bits set, it means a
6755 * network is playing bad games with TOS bits. In order to
6756 * avoid possible false congestion notifications, we disable
6757 * TCP ECN negotiation.
6759 * Exception: tcp_ca wants ECN. This is required for DCTCP
6760 * congestion control: Linux DCTCP asserts ECT on all packets,
6761 * including SYN, which is most optimal solution; however,
6762 * others, such as FreeBSD do not.
6764 * Exception: At least one of the reserved bits of the TCP header (th->res1) is
6765 * set, indicating the use of a future TCP extension (such as AccECN). See
6766 * RFC8311 §4.3 which updates RFC3168 to allow the development of such
6769 static void tcp_ecn_create_request(struct request_sock *req,
6770 const struct sk_buff *skb,
6771 const struct sock *listen_sk,
6772 const struct dst_entry *dst)
6774 const struct tcphdr *th = tcp_hdr(skb);
6775 const struct net *net = sock_net(listen_sk);
6776 bool th_ecn = th->ece && th->cwr;
6783 ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield);
6784 ecn_ok_dst = dst_feature(dst, DST_FEATURE_ECN_MASK);
6785 ecn_ok = READ_ONCE(net->ipv4.sysctl_tcp_ecn) || ecn_ok_dst;
6787 if (((!ect || th->res1) && ecn_ok) || tcp_ca_needs_ecn(listen_sk) ||
6788 (ecn_ok_dst & DST_FEATURE_ECN_CA) ||
6789 tcp_bpf_ca_needs_ecn((struct sock *)req))
6790 inet_rsk(req)->ecn_ok = 1;
6793 static void tcp_openreq_init(struct request_sock *req,
6794 const struct tcp_options_received *rx_opt,
6795 struct sk_buff *skb, const struct sock *sk)
6797 struct inet_request_sock *ireq = inet_rsk(req);
6799 req->rsk_rcv_wnd = 0; /* So that tcp_send_synack() knows! */
6800 tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq;
6801 tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
6802 tcp_rsk(req)->snt_synack = 0;
6803 tcp_rsk(req)->last_oow_ack_time = 0;
6804 req->mss = rx_opt->mss_clamp;
6805 req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0;
6806 ireq->tstamp_ok = rx_opt->tstamp_ok;
6807 ireq->sack_ok = rx_opt->sack_ok;
6808 ireq->snd_wscale = rx_opt->snd_wscale;
6809 ireq->wscale_ok = rx_opt->wscale_ok;
6812 ireq->ir_rmt_port = tcp_hdr(skb)->source;
6813 ireq->ir_num = ntohs(tcp_hdr(skb)->dest);
6814 ireq->ir_mark = inet_request_mark(sk, skb);
6815 #if IS_ENABLED(CONFIG_SMC)
6816 ireq->smc_ok = rx_opt->smc_ok && !(tcp_sk(sk)->smc_hs_congested &&
6817 tcp_sk(sk)->smc_hs_congested(sk));
6821 struct request_sock *inet_reqsk_alloc(const struct request_sock_ops *ops,
6822 struct sock *sk_listener,
6823 bool attach_listener)
6825 struct request_sock *req = reqsk_alloc(ops, sk_listener,
6829 struct inet_request_sock *ireq = inet_rsk(req);
6831 ireq->ireq_opt = NULL;
6832 #if IS_ENABLED(CONFIG_IPV6)
6833 ireq->pktopts = NULL;
6835 atomic64_set(&ireq->ir_cookie, 0);
6836 ireq->ireq_state = TCP_NEW_SYN_RECV;
6837 write_pnet(&ireq->ireq_net, sock_net(sk_listener));
6838 ireq->ireq_family = sk_listener->sk_family;
6839 req->timeout = TCP_TIMEOUT_INIT;
6844 EXPORT_SYMBOL(inet_reqsk_alloc);
6847 * Return true if a syncookie should be sent
6849 static bool tcp_syn_flood_action(const struct sock *sk, const char *proto)
6851 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
6852 const char *msg = "Dropping request";
6853 struct net *net = sock_net(sk);
6854 bool want_cookie = false;
6857 syncookies = READ_ONCE(net->ipv4.sysctl_tcp_syncookies);
6859 #ifdef CONFIG_SYN_COOKIES
6861 msg = "Sending cookies";
6863 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES);
6866 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP);
6868 if (!READ_ONCE(queue->synflood_warned) && syncookies != 2 &&
6869 xchg(&queue->synflood_warned, 1) == 0) {
6870 if (IS_ENABLED(CONFIG_IPV6) && sk->sk_family == AF_INET6) {
6871 net_info_ratelimited("%s: Possible SYN flooding on port [%pI6c]:%u. %s.\n",
6872 proto, inet6_rcv_saddr(sk),
6875 net_info_ratelimited("%s: Possible SYN flooding on port %pI4:%u. %s.\n",
6876 proto, &sk->sk_rcv_saddr,
6884 static void tcp_reqsk_record_syn(const struct sock *sk,
6885 struct request_sock *req,
6886 const struct sk_buff *skb)
6888 if (tcp_sk(sk)->save_syn) {
6889 u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb);
6890 struct saved_syn *saved_syn;
6894 if (tcp_sk(sk)->save_syn == 2) { /* Save full header. */
6895 base = skb_mac_header(skb);
6896 mac_hdrlen = skb_mac_header_len(skb);
6899 base = skb_network_header(skb);
6903 saved_syn = kmalloc(struct_size(saved_syn, data, len),
6906 saved_syn->mac_hdrlen = mac_hdrlen;
6907 saved_syn->network_hdrlen = skb_network_header_len(skb);
6908 saved_syn->tcp_hdrlen = tcp_hdrlen(skb);
6909 memcpy(saved_syn->data, base, len);
6910 req->saved_syn = saved_syn;
6915 /* If a SYN cookie is required and supported, returns a clamped MSS value to be
6916 * used for SYN cookie generation.
6918 u16 tcp_get_syncookie_mss(struct request_sock_ops *rsk_ops,
6919 const struct tcp_request_sock_ops *af_ops,
6920 struct sock *sk, struct tcphdr *th)
6922 struct tcp_sock *tp = tcp_sk(sk);
6925 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_syncookies) != 2 &&
6926 !inet_csk_reqsk_queue_is_full(sk))
6929 if (!tcp_syn_flood_action(sk, rsk_ops->slab_name))
6932 if (sk_acceptq_is_full(sk)) {
6933 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
6937 mss = tcp_parse_mss_option(th, tp->rx_opt.user_mss);
6939 mss = af_ops->mss_clamp;
6943 EXPORT_SYMBOL_GPL(tcp_get_syncookie_mss);
6945 int tcp_conn_request(struct request_sock_ops *rsk_ops,
6946 const struct tcp_request_sock_ops *af_ops,
6947 struct sock *sk, struct sk_buff *skb)
6949 struct tcp_fastopen_cookie foc = { .len = -1 };
6950 __u32 isn = TCP_SKB_CB(skb)->tcp_tw_isn;
6951 struct tcp_options_received tmp_opt;
6952 struct tcp_sock *tp = tcp_sk(sk);
6953 struct net *net = sock_net(sk);
6954 struct sock *fastopen_sk = NULL;
6955 struct request_sock *req;
6956 bool want_cookie = false;
6957 struct dst_entry *dst;
6961 syncookies = READ_ONCE(net->ipv4.sysctl_tcp_syncookies);
6963 /* TW buckets are converted to open requests without
6964 * limitations, they conserve resources and peer is
6965 * evidently real one.
6967 if ((syncookies == 2 || inet_csk_reqsk_queue_is_full(sk)) && !isn) {
6968 want_cookie = tcp_syn_flood_action(sk, rsk_ops->slab_name);
6973 if (sk_acceptq_is_full(sk)) {
6974 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
6978 req = inet_reqsk_alloc(rsk_ops, sk, !want_cookie);
6982 req->syncookie = want_cookie;
6983 tcp_rsk(req)->af_specific = af_ops;
6984 tcp_rsk(req)->ts_off = 0;
6985 #if IS_ENABLED(CONFIG_MPTCP)
6986 tcp_rsk(req)->is_mptcp = 0;
6989 tcp_clear_options(&tmp_opt);
6990 tmp_opt.mss_clamp = af_ops->mss_clamp;
6991 tmp_opt.user_mss = tp->rx_opt.user_mss;
6992 tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0,
6993 want_cookie ? NULL : &foc);
6995 if (want_cookie && !tmp_opt.saw_tstamp)
6996 tcp_clear_options(&tmp_opt);
6998 if (IS_ENABLED(CONFIG_SMC) && want_cookie)
7001 tmp_opt.tstamp_ok = tmp_opt.saw_tstamp;
7002 tcp_openreq_init(req, &tmp_opt, skb, sk);
7003 inet_rsk(req)->no_srccheck = inet_test_bit(TRANSPARENT, sk);
7005 /* Note: tcp_v6_init_req() might override ir_iif for link locals */
7006 inet_rsk(req)->ir_iif = inet_request_bound_dev_if(sk, skb);
7008 dst = af_ops->route_req(sk, skb, &fl, req);
7012 if (tmp_opt.tstamp_ok)
7013 tcp_rsk(req)->ts_off = af_ops->init_ts_off(net, skb);
7015 if (!want_cookie && !isn) {
7016 int max_syn_backlog = READ_ONCE(net->ipv4.sysctl_max_syn_backlog);
7018 /* Kill the following clause, if you dislike this way. */
7020 (max_syn_backlog - inet_csk_reqsk_queue_len(sk) <
7021 (max_syn_backlog >> 2)) &&
7022 !tcp_peer_is_proven(req, dst)) {
7023 /* Without syncookies last quarter of
7024 * backlog is filled with destinations,
7025 * proven to be alive.
7026 * It means that we continue to communicate
7027 * to destinations, already remembered
7028 * to the moment of synflood.
7030 pr_drop_req(req, ntohs(tcp_hdr(skb)->source),
7032 goto drop_and_release;
7035 isn = af_ops->init_seq(skb);
7038 tcp_ecn_create_request(req, skb, sk, dst);
7041 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss);
7042 if (!tmp_opt.tstamp_ok)
7043 inet_rsk(req)->ecn_ok = 0;
7046 tcp_rsk(req)->snt_isn = isn;
7047 tcp_rsk(req)->txhash = net_tx_rndhash();
7048 tcp_rsk(req)->syn_tos = TCP_SKB_CB(skb)->ip_dsfield;
7049 tcp_openreq_init_rwin(req, sk, dst);
7050 sk_rx_queue_set(req_to_sk(req), skb);
7052 tcp_reqsk_record_syn(sk, req, skb);
7053 fastopen_sk = tcp_try_fastopen(sk, skb, req, &foc, dst);
7056 af_ops->send_synack(fastopen_sk, dst, &fl, req,
7057 &foc, TCP_SYNACK_FASTOPEN, skb);
7058 /* Add the child socket directly into the accept queue */
7059 if (!inet_csk_reqsk_queue_add(sk, req, fastopen_sk)) {
7060 reqsk_fastopen_remove(fastopen_sk, req, false);
7061 bh_unlock_sock(fastopen_sk);
7062 sock_put(fastopen_sk);
7065 sk->sk_data_ready(sk);
7066 bh_unlock_sock(fastopen_sk);
7067 sock_put(fastopen_sk);
7069 tcp_rsk(req)->tfo_listener = false;
7071 req->timeout = tcp_timeout_init((struct sock *)req);
7072 inet_csk_reqsk_queue_hash_add(sk, req, req->timeout);
7074 af_ops->send_synack(sk, dst, &fl, req, &foc,
7075 !want_cookie ? TCP_SYNACK_NORMAL :
7094 EXPORT_SYMBOL(tcp_conn_request);