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);
256 /* If the skb has a len of exactly 1*MSS and has the PSH bit
257 * set then it is likely the end of an application write. So
258 * more data may not be arriving soon, and yet the data sender
259 * may be waiting for an ACK if cwnd-bound or using TX zero
260 * copy. So we set ICSK_ACK_PUSHED here so that
261 * tcp_cleanup_rbuf() will send an ACK immediately if the app
262 * reads all of the data and is not ping-pong. If len > MSS
263 * then this logic does not matter (and does not hurt) because
264 * tcp_cleanup_rbuf() will always ACK immediately if the app
265 * reads data and there is more than an MSS of unACKed data.
267 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_PSH)
268 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
270 /* Otherwise, we make more careful check taking into account,
271 * that SACKs block is variable.
273 * "len" is invariant segment length, including TCP header.
275 len += skb->data - skb_transport_header(skb);
276 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
277 /* If PSH is not set, packet should be
278 * full sized, provided peer TCP is not badly broken.
279 * This observation (if it is correct 8)) allows
280 * to handle super-low mtu links fairly.
282 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
283 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
284 /* Subtract also invariant (if peer is RFC compliant),
285 * tcp header plus fixed timestamp option length.
286 * Resulting "len" is MSS free of SACK jitter.
288 len -= tcp_sk(sk)->tcp_header_len;
289 icsk->icsk_ack.last_seg_size = len;
291 icsk->icsk_ack.rcv_mss = len;
295 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
296 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
297 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
301 static void tcp_incr_quickack(struct sock *sk, unsigned int max_quickacks)
303 struct inet_connection_sock *icsk = inet_csk(sk);
304 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
308 quickacks = min(quickacks, max_quickacks);
309 if (quickacks > icsk->icsk_ack.quick)
310 icsk->icsk_ack.quick = quickacks;
313 static void tcp_enter_quickack_mode(struct sock *sk, unsigned int max_quickacks)
315 struct inet_connection_sock *icsk = inet_csk(sk);
317 tcp_incr_quickack(sk, max_quickacks);
318 inet_csk_exit_pingpong_mode(sk);
319 icsk->icsk_ack.ato = TCP_ATO_MIN;
322 /* Send ACKs quickly, if "quick" count is not exhausted
323 * and the session is not interactive.
326 static bool tcp_in_quickack_mode(struct sock *sk)
328 const struct inet_connection_sock *icsk = inet_csk(sk);
329 const struct dst_entry *dst = __sk_dst_get(sk);
331 return (dst && dst_metric(dst, RTAX_QUICKACK)) ||
332 (icsk->icsk_ack.quick && !inet_csk_in_pingpong_mode(sk));
335 static void tcp_ecn_queue_cwr(struct tcp_sock *tp)
337 if (tp->ecn_flags & TCP_ECN_OK)
338 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
341 static void tcp_ecn_accept_cwr(struct sock *sk, const struct sk_buff *skb)
343 if (tcp_hdr(skb)->cwr) {
344 tcp_sk(sk)->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
346 /* If the sender is telling us it has entered CWR, then its
347 * cwnd may be very low (even just 1 packet), so we should ACK
350 if (TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq)
351 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW;
355 static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp)
357 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
360 static void __tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb)
362 struct tcp_sock *tp = tcp_sk(sk);
364 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
365 case INET_ECN_NOT_ECT:
366 /* Funny extension: if ECT is not set on a segment,
367 * and we already seen ECT on a previous segment,
368 * it is probably a retransmit.
370 if (tp->ecn_flags & TCP_ECN_SEEN)
371 tcp_enter_quickack_mode(sk, 2);
374 if (tcp_ca_needs_ecn(sk))
375 tcp_ca_event(sk, CA_EVENT_ECN_IS_CE);
377 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
378 /* Better not delay acks, sender can have a very low cwnd */
379 tcp_enter_quickack_mode(sk, 2);
380 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
382 tp->ecn_flags |= TCP_ECN_SEEN;
385 if (tcp_ca_needs_ecn(sk))
386 tcp_ca_event(sk, CA_EVENT_ECN_NO_CE);
387 tp->ecn_flags |= TCP_ECN_SEEN;
392 static void tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb)
394 if (tcp_sk(sk)->ecn_flags & TCP_ECN_OK)
395 __tcp_ecn_check_ce(sk, skb);
398 static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
400 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
401 tp->ecn_flags &= ~TCP_ECN_OK;
404 static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
406 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
407 tp->ecn_flags &= ~TCP_ECN_OK;
410 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
412 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
417 /* Buffer size and advertised window tuning.
419 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
422 static void tcp_sndbuf_expand(struct sock *sk)
424 const struct tcp_sock *tp = tcp_sk(sk);
425 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
429 /* Worst case is non GSO/TSO : each frame consumes one skb
430 * and skb->head is kmalloced using power of two area of memory
432 per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
434 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
436 per_mss = roundup_pow_of_two(per_mss) +
437 SKB_DATA_ALIGN(sizeof(struct sk_buff));
439 nr_segs = max_t(u32, TCP_INIT_CWND, tcp_snd_cwnd(tp));
440 nr_segs = max_t(u32, nr_segs, tp->reordering + 1);
442 /* Fast Recovery (RFC 5681 3.2) :
443 * Cubic needs 1.7 factor, rounded to 2 to include
444 * extra cushion (application might react slowly to EPOLLOUT)
446 sndmem = ca_ops->sndbuf_expand ? ca_ops->sndbuf_expand(sk) : 2;
447 sndmem *= nr_segs * per_mss;
449 if (sk->sk_sndbuf < sndmem)
450 WRITE_ONCE(sk->sk_sndbuf,
451 min(sndmem, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_wmem[2])));
454 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
456 * All tcp_full_space() is split to two parts: "network" buffer, allocated
457 * forward and advertised in receiver window (tp->rcv_wnd) and
458 * "application buffer", required to isolate scheduling/application
459 * latencies from network.
460 * window_clamp is maximal advertised window. It can be less than
461 * tcp_full_space(), in this case tcp_full_space() - window_clamp
462 * is reserved for "application" buffer. The less window_clamp is
463 * the smoother our behaviour from viewpoint of network, but the lower
464 * throughput and the higher sensitivity of the connection to losses. 8)
466 * rcv_ssthresh is more strict window_clamp used at "slow start"
467 * phase to predict further behaviour of this connection.
468 * It is used for two goals:
469 * - to enforce header prediction at sender, even when application
470 * requires some significant "application buffer". It is check #1.
471 * - to prevent pruning of receive queue because of misprediction
472 * of receiver window. Check #2.
474 * The scheme does not work when sender sends good segments opening
475 * window and then starts to feed us spaghetti. But it should work
476 * in common situations. Otherwise, we have to rely on queue collapsing.
479 /* Slow part of check#2. */
480 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb,
481 unsigned int skbtruesize)
483 const struct tcp_sock *tp = tcp_sk(sk);
485 int truesize = tcp_win_from_space(sk, skbtruesize) >> 1;
486 int window = tcp_win_from_space(sk, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2])) >> 1;
488 while (tp->rcv_ssthresh <= window) {
489 if (truesize <= skb->len)
490 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
498 /* Even if skb appears to have a bad len/truesize ratio, TCP coalescing
499 * can play nice with us, as sk_buff and skb->head might be either
500 * freed or shared with up to MAX_SKB_FRAGS segments.
501 * Only give a boost to drivers using page frag(s) to hold the frame(s),
502 * and if no payload was pulled in skb->head before reaching us.
504 static u32 truesize_adjust(bool adjust, const struct sk_buff *skb)
506 u32 truesize = skb->truesize;
508 if (adjust && !skb_headlen(skb)) {
509 truesize -= SKB_TRUESIZE(skb_end_offset(skb));
510 /* paranoid check, some drivers might be buggy */
511 if (unlikely((int)truesize < (int)skb->len))
512 truesize = skb->truesize;
517 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb,
520 struct tcp_sock *tp = tcp_sk(sk);
523 room = min_t(int, tp->window_clamp, tcp_space(sk)) - tp->rcv_ssthresh;
529 if (!tcp_under_memory_pressure(sk)) {
530 unsigned int truesize = truesize_adjust(adjust, skb);
533 /* Check #2. Increase window, if skb with such overhead
534 * will fit to rcvbuf in future.
536 if (tcp_win_from_space(sk, truesize) <= skb->len)
537 incr = 2 * tp->advmss;
539 incr = __tcp_grow_window(sk, skb, truesize);
542 incr = max_t(int, incr, 2 * skb->len);
543 tp->rcv_ssthresh += min(room, incr);
544 inet_csk(sk)->icsk_ack.quick |= 1;
548 * Adjust rcv_ssthresh according to reserved mem
550 tcp_adjust_rcv_ssthresh(sk);
554 /* 3. Try to fixup all. It is made immediately after connection enters
557 static void tcp_init_buffer_space(struct sock *sk)
559 int tcp_app_win = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_app_win);
560 struct tcp_sock *tp = tcp_sk(sk);
563 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
564 tcp_sndbuf_expand(sk);
566 tcp_mstamp_refresh(tp);
567 tp->rcvq_space.time = tp->tcp_mstamp;
568 tp->rcvq_space.seq = tp->copied_seq;
570 maxwin = tcp_full_space(sk);
572 if (tp->window_clamp >= maxwin) {
573 tp->window_clamp = maxwin;
575 if (tcp_app_win && maxwin > 4 * tp->advmss)
576 tp->window_clamp = max(maxwin -
577 (maxwin >> tcp_app_win),
581 /* Force reservation of one segment. */
583 tp->window_clamp > 2 * tp->advmss &&
584 tp->window_clamp + tp->advmss > maxwin)
585 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
587 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
588 tp->snd_cwnd_stamp = tcp_jiffies32;
589 tp->rcvq_space.space = min3(tp->rcv_ssthresh, tp->rcv_wnd,
590 (u32)TCP_INIT_CWND * tp->advmss);
593 /* 4. Recalculate window clamp after socket hit its memory bounds. */
594 static void tcp_clamp_window(struct sock *sk)
596 struct tcp_sock *tp = tcp_sk(sk);
597 struct inet_connection_sock *icsk = inet_csk(sk);
598 struct net *net = sock_net(sk);
601 icsk->icsk_ack.quick = 0;
602 rmem2 = READ_ONCE(net->ipv4.sysctl_tcp_rmem[2]);
604 if (sk->sk_rcvbuf < rmem2 &&
605 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
606 !tcp_under_memory_pressure(sk) &&
607 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
608 WRITE_ONCE(sk->sk_rcvbuf,
609 min(atomic_read(&sk->sk_rmem_alloc), rmem2));
611 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
612 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
615 /* Initialize RCV_MSS value.
616 * RCV_MSS is an our guess about MSS used by the peer.
617 * We haven't any direct information about the MSS.
618 * It's better to underestimate the RCV_MSS rather than overestimate.
619 * Overestimations make us ACKing less frequently than needed.
620 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
622 void tcp_initialize_rcv_mss(struct sock *sk)
624 const struct tcp_sock *tp = tcp_sk(sk);
625 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
627 hint = min(hint, tp->rcv_wnd / 2);
628 hint = min(hint, TCP_MSS_DEFAULT);
629 hint = max(hint, TCP_MIN_MSS);
631 inet_csk(sk)->icsk_ack.rcv_mss = hint;
633 EXPORT_SYMBOL(tcp_initialize_rcv_mss);
635 /* Receiver "autotuning" code.
637 * The algorithm for RTT estimation w/o timestamps is based on
638 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
639 * <https://public.lanl.gov/radiant/pubs.html#DRS>
641 * More detail on this code can be found at
642 * <http://staff.psc.edu/jheffner/>,
643 * though this reference is out of date. A new paper
646 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
648 u32 new_sample = tp->rcv_rtt_est.rtt_us;
651 if (new_sample != 0) {
652 /* If we sample in larger samples in the non-timestamp
653 * case, we could grossly overestimate the RTT especially
654 * with chatty applications or bulk transfer apps which
655 * are stalled on filesystem I/O.
657 * Also, since we are only going for a minimum in the
658 * non-timestamp case, we do not smooth things out
659 * else with timestamps disabled convergence takes too
663 m -= (new_sample >> 3);
671 /* No previous measure. */
675 tp->rcv_rtt_est.rtt_us = new_sample;
678 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
682 if (tp->rcv_rtt_est.time == 0)
684 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
686 delta_us = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcv_rtt_est.time);
689 tcp_rcv_rtt_update(tp, delta_us, 1);
692 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
693 tp->rcv_rtt_est.time = tp->tcp_mstamp;
696 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
697 const struct sk_buff *skb)
699 struct tcp_sock *tp = tcp_sk(sk);
701 if (tp->rx_opt.rcv_tsecr == tp->rcv_rtt_last_tsecr)
703 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr;
705 if (TCP_SKB_CB(skb)->end_seq -
706 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss) {
707 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr;
710 if (likely(delta < INT_MAX / (USEC_PER_SEC / TCP_TS_HZ))) {
713 delta_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
714 tcp_rcv_rtt_update(tp, delta_us, 0);
720 * This function should be called every time data is copied to user space.
721 * It calculates the appropriate TCP receive buffer space.
723 void tcp_rcv_space_adjust(struct sock *sk)
725 struct tcp_sock *tp = tcp_sk(sk);
729 trace_tcp_rcv_space_adjust(sk);
731 tcp_mstamp_refresh(tp);
732 time = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcvq_space.time);
733 if (time < (tp->rcv_rtt_est.rtt_us >> 3) || tp->rcv_rtt_est.rtt_us == 0)
736 /* Number of bytes copied to user in last RTT */
737 copied = tp->copied_seq - tp->rcvq_space.seq;
738 if (copied <= tp->rcvq_space.space)
742 * copied = bytes received in previous RTT, our base window
743 * To cope with packet losses, we need a 2x factor
744 * To cope with slow start, and sender growing its cwin by 100 %
745 * every RTT, we need a 4x factor, because the ACK we are sending
746 * now is for the next RTT, not the current one :
747 * <prev RTT . ><current RTT .. ><next RTT .... >
750 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf) &&
751 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
755 /* minimal window to cope with packet losses, assuming
756 * steady state. Add some cushion because of small variations.
758 rcvwin = ((u64)copied << 1) + 16 * tp->advmss;
760 /* Accommodate for sender rate increase (eg. slow start) */
761 grow = rcvwin * (copied - tp->rcvq_space.space);
762 do_div(grow, tp->rcvq_space.space);
763 rcvwin += (grow << 1);
765 rcvbuf = min_t(u64, tcp_space_from_win(sk, rcvwin),
766 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]));
767 if (rcvbuf > sk->sk_rcvbuf) {
768 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
770 /* Make the window clamp follow along. */
771 tp->window_clamp = tcp_win_from_space(sk, rcvbuf);
774 tp->rcvq_space.space = copied;
777 tp->rcvq_space.seq = tp->copied_seq;
778 tp->rcvq_space.time = tp->tcp_mstamp;
781 /* There is something which you must keep in mind when you analyze the
782 * behavior of the tp->ato delayed ack timeout interval. When a
783 * connection starts up, we want to ack as quickly as possible. The
784 * problem is that "good" TCP's do slow start at the beginning of data
785 * transmission. The means that until we send the first few ACK's the
786 * sender will sit on his end and only queue most of his data, because
787 * he can only send snd_cwnd unacked packets at any given time. For
788 * each ACK we send, he increments snd_cwnd and transmits more of his
791 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
793 struct tcp_sock *tp = tcp_sk(sk);
794 struct inet_connection_sock *icsk = inet_csk(sk);
797 inet_csk_schedule_ack(sk);
799 tcp_measure_rcv_mss(sk, skb);
801 tcp_rcv_rtt_measure(tp);
805 if (!icsk->icsk_ack.ato) {
806 /* The _first_ data packet received, initialize
807 * delayed ACK engine.
809 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
810 icsk->icsk_ack.ato = TCP_ATO_MIN;
812 int m = now - icsk->icsk_ack.lrcvtime;
814 if (m <= TCP_ATO_MIN / 2) {
815 /* The fastest case is the first. */
816 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
817 } else if (m < icsk->icsk_ack.ato) {
818 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
819 if (icsk->icsk_ack.ato > icsk->icsk_rto)
820 icsk->icsk_ack.ato = icsk->icsk_rto;
821 } else if (m > icsk->icsk_rto) {
822 /* Too long gap. Apparently sender failed to
823 * restart window, so that we send ACKs quickly.
825 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
828 icsk->icsk_ack.lrcvtime = now;
830 tcp_ecn_check_ce(sk, skb);
833 tcp_grow_window(sk, skb, true);
836 /* Called to compute a smoothed rtt estimate. The data fed to this
837 * routine either comes from timestamps, or from segments that were
838 * known _not_ to have been retransmitted [see Karn/Partridge
839 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
840 * piece by Van Jacobson.
841 * NOTE: the next three routines used to be one big routine.
842 * To save cycles in the RFC 1323 implementation it was better to break
843 * it up into three procedures. -- erics
845 static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
847 struct tcp_sock *tp = tcp_sk(sk);
848 long m = mrtt_us; /* RTT */
849 u32 srtt = tp->srtt_us;
851 /* The following amusing code comes from Jacobson's
852 * article in SIGCOMM '88. Note that rtt and mdev
853 * are scaled versions of rtt and mean deviation.
854 * This is designed to be as fast as possible
855 * m stands for "measurement".
857 * On a 1990 paper the rto value is changed to:
858 * RTO = rtt + 4 * mdev
860 * Funny. This algorithm seems to be very broken.
861 * These formulae increase RTO, when it should be decreased, increase
862 * too slowly, when it should be increased quickly, decrease too quickly
863 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
864 * does not matter how to _calculate_ it. Seems, it was trap
865 * that VJ failed to avoid. 8)
868 m -= (srtt >> 3); /* m is now error in rtt est */
869 srtt += m; /* rtt = 7/8 rtt + 1/8 new */
871 m = -m; /* m is now abs(error) */
872 m -= (tp->mdev_us >> 2); /* similar update on mdev */
873 /* This is similar to one of Eifel findings.
874 * Eifel blocks mdev updates when rtt decreases.
875 * This solution is a bit different: we use finer gain
876 * for mdev in this case (alpha*beta).
877 * Like Eifel it also prevents growth of rto,
878 * but also it limits too fast rto decreases,
879 * happening in pure Eifel.
884 m -= (tp->mdev_us >> 2); /* similar update on mdev */
886 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */
887 if (tp->mdev_us > tp->mdev_max_us) {
888 tp->mdev_max_us = tp->mdev_us;
889 if (tp->mdev_max_us > tp->rttvar_us)
890 tp->rttvar_us = tp->mdev_max_us;
892 if (after(tp->snd_una, tp->rtt_seq)) {
893 if (tp->mdev_max_us < tp->rttvar_us)
894 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
895 tp->rtt_seq = tp->snd_nxt;
896 tp->mdev_max_us = tcp_rto_min_us(sk);
901 /* no previous measure. */
902 srtt = m << 3; /* take the measured time to be rtt */
903 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */
904 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
905 tp->mdev_max_us = tp->rttvar_us;
906 tp->rtt_seq = tp->snd_nxt;
910 tp->srtt_us = max(1U, srtt);
913 static void tcp_update_pacing_rate(struct sock *sk)
915 const struct tcp_sock *tp = tcp_sk(sk);
918 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
919 rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3);
921 /* current rate is (cwnd * mss) / srtt
922 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
923 * In Congestion Avoidance phase, set it to 120 % the current rate.
925 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
926 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
927 * end of slow start and should slow down.
929 if (tcp_snd_cwnd(tp) < tp->snd_ssthresh / 2)
930 rate *= READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_pacing_ss_ratio);
932 rate *= READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_pacing_ca_ratio);
934 rate *= max(tcp_snd_cwnd(tp), tp->packets_out);
936 if (likely(tp->srtt_us))
937 do_div(rate, tp->srtt_us);
939 /* WRITE_ONCE() is needed because sch_fq fetches sk_pacing_rate
940 * without any lock. We want to make sure compiler wont store
941 * intermediate values in this location.
943 WRITE_ONCE(sk->sk_pacing_rate, min_t(u64, rate,
944 sk->sk_max_pacing_rate));
947 /* Calculate rto without backoff. This is the second half of Van Jacobson's
948 * routine referred to above.
950 static void tcp_set_rto(struct sock *sk)
952 const struct tcp_sock *tp = tcp_sk(sk);
953 /* Old crap is replaced with new one. 8)
956 * 1. If rtt variance happened to be less 50msec, it is hallucination.
957 * It cannot be less due to utterly erratic ACK generation made
958 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
959 * to do with delayed acks, because at cwnd>2 true delack timeout
960 * is invisible. Actually, Linux-2.4 also generates erratic
961 * ACKs in some circumstances.
963 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
965 /* 2. Fixups made earlier cannot be right.
966 * If we do not estimate RTO correctly without them,
967 * all the algo is pure shit and should be replaced
968 * with correct one. It is exactly, which we pretend to do.
971 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
972 * guarantees that rto is higher.
977 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
979 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
982 cwnd = TCP_INIT_CWND;
983 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
986 struct tcp_sacktag_state {
987 /* Timestamps for earliest and latest never-retransmitted segment
988 * that was SACKed. RTO needs the earliest RTT to stay conservative,
989 * but congestion control should still get an accurate delay signal.
996 unsigned int mss_now;
997 struct rate_sample *rate;
1000 /* Take a notice that peer is sending D-SACKs. Skip update of data delivery
1001 * and spurious retransmission information if this DSACK is unlikely caused by
1003 * - DSACKed sequence range is larger than maximum receiver's window.
1004 * - Total no. of DSACKed segments exceed the total no. of retransmitted segs.
1006 static u32 tcp_dsack_seen(struct tcp_sock *tp, u32 start_seq,
1007 u32 end_seq, struct tcp_sacktag_state *state)
1009 u32 seq_len, dup_segs = 1;
1011 if (!before(start_seq, end_seq))
1014 seq_len = end_seq - start_seq;
1015 /* Dubious DSACK: DSACKed range greater than maximum advertised rwnd */
1016 if (seq_len > tp->max_window)
1018 if (seq_len > tp->mss_cache)
1019 dup_segs = DIV_ROUND_UP(seq_len, tp->mss_cache);
1020 else if (tp->tlp_high_seq && tp->tlp_high_seq == end_seq)
1021 state->flag |= FLAG_DSACK_TLP;
1023 tp->dsack_dups += dup_segs;
1024 /* Skip the DSACK if dup segs weren't retransmitted by sender */
1025 if (tp->dsack_dups > tp->total_retrans)
1028 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
1029 /* We increase the RACK ordering window in rounds where we receive
1030 * DSACKs that may have been due to reordering causing RACK to trigger
1031 * a spurious fast recovery. Thus RACK ignores DSACKs that happen
1032 * without having seen reordering, or that match TLP probes (TLP
1033 * is timer-driven, not triggered by RACK).
1035 if (tp->reord_seen && !(state->flag & FLAG_DSACK_TLP))
1036 tp->rack.dsack_seen = 1;
1038 state->flag |= FLAG_DSACKING_ACK;
1039 /* A spurious retransmission is delivered */
1040 state->sack_delivered += dup_segs;
1045 /* It's reordering when higher sequence was delivered (i.e. sacked) before
1046 * some lower never-retransmitted sequence ("low_seq"). The maximum reordering
1047 * distance is approximated in full-mss packet distance ("reordering").
1049 static void tcp_check_sack_reordering(struct sock *sk, const u32 low_seq,
1052 struct tcp_sock *tp = tcp_sk(sk);
1053 const u32 mss = tp->mss_cache;
1056 fack = tcp_highest_sack_seq(tp);
1057 if (!before(low_seq, fack))
1060 metric = fack - low_seq;
1061 if ((metric > tp->reordering * mss) && mss) {
1062 #if FASTRETRANS_DEBUG > 1
1063 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
1064 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
1068 tp->undo_marker ? tp->undo_retrans : 0);
1070 tp->reordering = min_t(u32, (metric + mss - 1) / mss,
1071 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_max_reordering));
1074 /* This exciting event is worth to be remembered. 8) */
1076 NET_INC_STATS(sock_net(sk),
1077 ts ? LINUX_MIB_TCPTSREORDER : LINUX_MIB_TCPSACKREORDER);
1080 /* This must be called before lost_out or retrans_out are updated
1081 * on a new loss, because we want to know if all skbs previously
1082 * known to be lost have already been retransmitted, indicating
1083 * that this newly lost skb is our next skb to retransmit.
1085 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
1087 if ((!tp->retransmit_skb_hint && tp->retrans_out >= tp->lost_out) ||
1088 (tp->retransmit_skb_hint &&
1089 before(TCP_SKB_CB(skb)->seq,
1090 TCP_SKB_CB(tp->retransmit_skb_hint)->seq)))
1091 tp->retransmit_skb_hint = skb;
1094 /* Sum the number of packets on the wire we have marked as lost, and
1095 * notify the congestion control module that the given skb was marked lost.
1097 static void tcp_notify_skb_loss_event(struct tcp_sock *tp, const struct sk_buff *skb)
1099 tp->lost += tcp_skb_pcount(skb);
1102 void tcp_mark_skb_lost(struct sock *sk, struct sk_buff *skb)
1104 __u8 sacked = TCP_SKB_CB(skb)->sacked;
1105 struct tcp_sock *tp = tcp_sk(sk);
1107 if (sacked & TCPCB_SACKED_ACKED)
1110 tcp_verify_retransmit_hint(tp, skb);
1111 if (sacked & TCPCB_LOST) {
1112 if (sacked & TCPCB_SACKED_RETRANS) {
1113 /* Account for retransmits that are lost again */
1114 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1115 tp->retrans_out -= tcp_skb_pcount(skb);
1116 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT,
1117 tcp_skb_pcount(skb));
1118 tcp_notify_skb_loss_event(tp, skb);
1121 tp->lost_out += tcp_skb_pcount(skb);
1122 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1123 tcp_notify_skb_loss_event(tp, skb);
1127 /* Updates the delivered and delivered_ce counts */
1128 static void tcp_count_delivered(struct tcp_sock *tp, u32 delivered,
1131 tp->delivered += delivered;
1133 tp->delivered_ce += delivered;
1136 /* This procedure tags the retransmission queue when SACKs arrive.
1138 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
1139 * Packets in queue with these bits set are counted in variables
1140 * sacked_out, retrans_out and lost_out, correspondingly.
1142 * Valid combinations are:
1143 * Tag InFlight Description
1144 * 0 1 - orig segment is in flight.
1145 * S 0 - nothing flies, orig reached receiver.
1146 * L 0 - nothing flies, orig lost by net.
1147 * R 2 - both orig and retransmit are in flight.
1148 * L|R 1 - orig is lost, retransmit is in flight.
1149 * S|R 1 - orig reached receiver, retrans is still in flight.
1150 * (L|S|R is logically valid, it could occur when L|R is sacked,
1151 * but it is equivalent to plain S and code short-curcuits it to S.
1152 * L|S is logically invalid, it would mean -1 packet in flight 8))
1154 * These 6 states form finite state machine, controlled by the following events:
1155 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
1156 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
1157 * 3. Loss detection event of two flavors:
1158 * A. Scoreboard estimator decided the packet is lost.
1159 * A'. Reno "three dupacks" marks head of queue lost.
1160 * B. SACK arrives sacking SND.NXT at the moment, when the
1161 * segment was retransmitted.
1162 * 4. D-SACK added new rule: D-SACK changes any tag to S.
1164 * It is pleasant to note, that state diagram turns out to be commutative,
1165 * so that we are allowed not to be bothered by order of our actions,
1166 * when multiple events arrive simultaneously. (see the function below).
1168 * Reordering detection.
1169 * --------------------
1170 * Reordering metric is maximal distance, which a packet can be displaced
1171 * in packet stream. With SACKs we can estimate it:
1173 * 1. SACK fills old hole and the corresponding segment was not
1174 * ever retransmitted -> reordering. Alas, we cannot use it
1175 * when segment was retransmitted.
1176 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1177 * for retransmitted and already SACKed segment -> reordering..
1178 * Both of these heuristics are not used in Loss state, when we cannot
1179 * account for retransmits accurately.
1181 * SACK block validation.
1182 * ----------------------
1184 * SACK block range validation checks that the received SACK block fits to
1185 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1186 * Note that SND.UNA is not included to the range though being valid because
1187 * it means that the receiver is rather inconsistent with itself reporting
1188 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1189 * perfectly valid, however, in light of RFC2018 which explicitly states
1190 * that "SACK block MUST reflect the newest segment. Even if the newest
1191 * segment is going to be discarded ...", not that it looks very clever
1192 * in case of head skb. Due to potentional receiver driven attacks, we
1193 * choose to avoid immediate execution of a walk in write queue due to
1194 * reneging and defer head skb's loss recovery to standard loss recovery
1195 * procedure that will eventually trigger (nothing forbids us doing this).
1197 * Implements also blockage to start_seq wrap-around. Problem lies in the
1198 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1199 * there's no guarantee that it will be before snd_nxt (n). The problem
1200 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1203 * <- outs wnd -> <- wrapzone ->
1204 * u e n u_w e_w s n_w
1206 * |<------------+------+----- TCP seqno space --------------+---------->|
1207 * ...-- <2^31 ->| |<--------...
1208 * ...---- >2^31 ------>| |<--------...
1210 * Current code wouldn't be vulnerable but it's better still to discard such
1211 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1212 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1213 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1214 * equal to the ideal case (infinite seqno space without wrap caused issues).
1216 * With D-SACK the lower bound is extended to cover sequence space below
1217 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1218 * again, D-SACK block must not to go across snd_una (for the same reason as
1219 * for the normal SACK blocks, explained above). But there all simplicity
1220 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1221 * fully below undo_marker they do not affect behavior in anyway and can
1222 * therefore be safely ignored. In rare cases (which are more or less
1223 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1224 * fragmentation and packet reordering past skb's retransmission. To consider
1225 * them correctly, the acceptable range must be extended even more though
1226 * the exact amount is rather hard to quantify. However, tp->max_window can
1227 * be used as an exaggerated estimate.
1229 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
1230 u32 start_seq, u32 end_seq)
1232 /* Too far in future, or reversed (interpretation is ambiguous) */
1233 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1236 /* Nasty start_seq wrap-around check (see comments above) */
1237 if (!before(start_seq, tp->snd_nxt))
1240 /* In outstanding window? ...This is valid exit for D-SACKs too.
1241 * start_seq == snd_una is non-sensical (see comments above)
1243 if (after(start_seq, tp->snd_una))
1246 if (!is_dsack || !tp->undo_marker)
1249 /* ...Then it's D-SACK, and must reside below snd_una completely */
1250 if (after(end_seq, tp->snd_una))
1253 if (!before(start_seq, tp->undo_marker))
1257 if (!after(end_seq, tp->undo_marker))
1260 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1261 * start_seq < undo_marker and end_seq >= undo_marker.
1263 return !before(start_seq, end_seq - tp->max_window);
1266 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1267 struct tcp_sack_block_wire *sp, int num_sacks,
1268 u32 prior_snd_una, struct tcp_sacktag_state *state)
1270 struct tcp_sock *tp = tcp_sk(sk);
1271 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1272 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1275 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1276 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1277 } else if (num_sacks > 1) {
1278 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1279 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1281 if (after(end_seq_0, end_seq_1) || before(start_seq_0, start_seq_1))
1283 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKOFORECV);
1288 dup_segs = tcp_dsack_seen(tp, start_seq_0, end_seq_0, state);
1289 if (!dup_segs) { /* Skip dubious DSACK */
1290 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKIGNOREDDUBIOUS);
1294 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECVSEGS, dup_segs);
1296 /* D-SACK for already forgotten data... Do dumb counting. */
1297 if (tp->undo_marker && tp->undo_retrans > 0 &&
1298 !after(end_seq_0, prior_snd_una) &&
1299 after(end_seq_0, tp->undo_marker))
1300 tp->undo_retrans = max_t(int, 0, tp->undo_retrans - dup_segs);
1305 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1306 * the incoming SACK may not exactly match but we can find smaller MSS
1307 * aligned portion of it that matches. Therefore we might need to fragment
1308 * which may fail and creates some hassle (caller must handle error case
1311 * FIXME: this could be merged to shift decision code
1313 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1314 u32 start_seq, u32 end_seq)
1318 unsigned int pkt_len;
1321 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1322 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1324 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1325 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1326 mss = tcp_skb_mss(skb);
1327 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1330 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1334 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1339 /* Round if necessary so that SACKs cover only full MSSes
1340 * and/or the remaining small portion (if present)
1342 if (pkt_len > mss) {
1343 unsigned int new_len = (pkt_len / mss) * mss;
1344 if (!in_sack && new_len < pkt_len)
1349 if (pkt_len >= skb->len && !in_sack)
1352 err = tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
1353 pkt_len, mss, GFP_ATOMIC);
1361 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1362 static u8 tcp_sacktag_one(struct sock *sk,
1363 struct tcp_sacktag_state *state, u8 sacked,
1364 u32 start_seq, u32 end_seq,
1365 int dup_sack, int pcount,
1368 struct tcp_sock *tp = tcp_sk(sk);
1370 /* Account D-SACK for retransmitted packet. */
1371 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1372 if (tp->undo_marker && tp->undo_retrans > 0 &&
1373 after(end_seq, tp->undo_marker))
1374 tp->undo_retrans = max_t(int, 0, tp->undo_retrans - pcount);
1375 if ((sacked & TCPCB_SACKED_ACKED) &&
1376 before(start_seq, state->reord))
1377 state->reord = start_seq;
1380 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1381 if (!after(end_seq, tp->snd_una))
1384 if (!(sacked & TCPCB_SACKED_ACKED)) {
1385 tcp_rack_advance(tp, sacked, end_seq, xmit_time);
1387 if (sacked & TCPCB_SACKED_RETRANS) {
1388 /* If the segment is not tagged as lost,
1389 * we do not clear RETRANS, believing
1390 * that retransmission is still in flight.
1392 if (sacked & TCPCB_LOST) {
1393 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1394 tp->lost_out -= pcount;
1395 tp->retrans_out -= pcount;
1398 if (!(sacked & TCPCB_RETRANS)) {
1399 /* New sack for not retransmitted frame,
1400 * which was in hole. It is reordering.
1402 if (before(start_seq,
1403 tcp_highest_sack_seq(tp)) &&
1404 before(start_seq, state->reord))
1405 state->reord = start_seq;
1407 if (!after(end_seq, tp->high_seq))
1408 state->flag |= FLAG_ORIG_SACK_ACKED;
1409 if (state->first_sackt == 0)
1410 state->first_sackt = xmit_time;
1411 state->last_sackt = xmit_time;
1414 if (sacked & TCPCB_LOST) {
1415 sacked &= ~TCPCB_LOST;
1416 tp->lost_out -= pcount;
1420 sacked |= TCPCB_SACKED_ACKED;
1421 state->flag |= FLAG_DATA_SACKED;
1422 tp->sacked_out += pcount;
1423 /* Out-of-order packets delivered */
1424 state->sack_delivered += pcount;
1426 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1427 if (tp->lost_skb_hint &&
1428 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1429 tp->lost_cnt_hint += pcount;
1432 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1433 * frames and clear it. undo_retrans is decreased above, L|R frames
1434 * are accounted above as well.
1436 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1437 sacked &= ~TCPCB_SACKED_RETRANS;
1438 tp->retrans_out -= pcount;
1444 /* Shift newly-SACKed bytes from this skb to the immediately previous
1445 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1447 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *prev,
1448 struct sk_buff *skb,
1449 struct tcp_sacktag_state *state,
1450 unsigned int pcount, int shifted, int mss,
1453 struct tcp_sock *tp = tcp_sk(sk);
1454 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1455 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
1459 /* Adjust counters and hints for the newly sacked sequence
1460 * range but discard the return value since prev is already
1461 * marked. We must tag the range first because the seq
1462 * advancement below implicitly advances
1463 * tcp_highest_sack_seq() when skb is highest_sack.
1465 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1466 start_seq, end_seq, dup_sack, pcount,
1467 tcp_skb_timestamp_us(skb));
1468 tcp_rate_skb_delivered(sk, skb, state->rate);
1470 if (skb == tp->lost_skb_hint)
1471 tp->lost_cnt_hint += pcount;
1473 TCP_SKB_CB(prev)->end_seq += shifted;
1474 TCP_SKB_CB(skb)->seq += shifted;
1476 tcp_skb_pcount_add(prev, pcount);
1477 WARN_ON_ONCE(tcp_skb_pcount(skb) < pcount);
1478 tcp_skb_pcount_add(skb, -pcount);
1480 /* When we're adding to gso_segs == 1, gso_size will be zero,
1481 * in theory this shouldn't be necessary but as long as DSACK
1482 * code can come after this skb later on it's better to keep
1483 * setting gso_size to something.
1485 if (!TCP_SKB_CB(prev)->tcp_gso_size)
1486 TCP_SKB_CB(prev)->tcp_gso_size = mss;
1488 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1489 if (tcp_skb_pcount(skb) <= 1)
1490 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1492 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1493 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1496 BUG_ON(!tcp_skb_pcount(skb));
1497 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1501 /* Whole SKB was eaten :-) */
1503 if (skb == tp->retransmit_skb_hint)
1504 tp->retransmit_skb_hint = prev;
1505 if (skb == tp->lost_skb_hint) {
1506 tp->lost_skb_hint = prev;
1507 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1510 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1511 TCP_SKB_CB(prev)->eor = TCP_SKB_CB(skb)->eor;
1512 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1513 TCP_SKB_CB(prev)->end_seq++;
1515 if (skb == tcp_highest_sack(sk))
1516 tcp_advance_highest_sack(sk, skb);
1518 tcp_skb_collapse_tstamp(prev, skb);
1519 if (unlikely(TCP_SKB_CB(prev)->tx.delivered_mstamp))
1520 TCP_SKB_CB(prev)->tx.delivered_mstamp = 0;
1522 tcp_rtx_queue_unlink_and_free(skb, sk);
1524 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKMERGED);
1529 /* I wish gso_size would have a bit more sane initialization than
1530 * something-or-zero which complicates things
1532 static int tcp_skb_seglen(const struct sk_buff *skb)
1534 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1537 /* Shifting pages past head area doesn't work */
1538 static int skb_can_shift(const struct sk_buff *skb)
1540 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1543 int tcp_skb_shift(struct sk_buff *to, struct sk_buff *from,
1544 int pcount, int shiftlen)
1546 /* TCP min gso_size is 8 bytes (TCP_MIN_GSO_SIZE)
1547 * Since TCP_SKB_CB(skb)->tcp_gso_segs is 16 bits, we need
1548 * to make sure not storing more than 65535 * 8 bytes per skb,
1549 * even if current MSS is bigger.
1551 if (unlikely(to->len + shiftlen >= 65535 * TCP_MIN_GSO_SIZE))
1553 if (unlikely(tcp_skb_pcount(to) + pcount > 65535))
1555 return skb_shift(to, from, shiftlen);
1558 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1561 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1562 struct tcp_sacktag_state *state,
1563 u32 start_seq, u32 end_seq,
1566 struct tcp_sock *tp = tcp_sk(sk);
1567 struct sk_buff *prev;
1573 /* Normally R but no L won't result in plain S */
1575 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1577 if (!skb_can_shift(skb))
1579 /* This frame is about to be dropped (was ACKed). */
1580 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1583 /* Can only happen with delayed DSACK + discard craziness */
1584 prev = skb_rb_prev(skb);
1588 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1591 if (!tcp_skb_can_collapse(prev, skb))
1594 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1595 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1599 pcount = tcp_skb_pcount(skb);
1600 mss = tcp_skb_seglen(skb);
1602 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1603 * drop this restriction as unnecessary
1605 if (mss != tcp_skb_seglen(prev))
1608 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1610 /* CHECKME: This is non-MSS split case only?, this will
1611 * cause skipped skbs due to advancing loop btw, original
1612 * has that feature too
1614 if (tcp_skb_pcount(skb) <= 1)
1617 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1619 /* TODO: head merge to next could be attempted here
1620 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1621 * though it might not be worth of the additional hassle
1623 * ...we can probably just fallback to what was done
1624 * previously. We could try merging non-SACKed ones
1625 * as well but it probably isn't going to buy off
1626 * because later SACKs might again split them, and
1627 * it would make skb timestamp tracking considerably
1633 len = end_seq - TCP_SKB_CB(skb)->seq;
1635 BUG_ON(len > skb->len);
1637 /* MSS boundaries should be honoured or else pcount will
1638 * severely break even though it makes things bit trickier.
1639 * Optimize common case to avoid most of the divides
1641 mss = tcp_skb_mss(skb);
1643 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1644 * drop this restriction as unnecessary
1646 if (mss != tcp_skb_seglen(prev))
1651 } else if (len < mss) {
1659 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1660 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1663 if (!tcp_skb_shift(prev, skb, pcount, len))
1665 if (!tcp_shifted_skb(sk, prev, skb, state, pcount, len, mss, dup_sack))
1668 /* Hole filled allows collapsing with the next as well, this is very
1669 * useful when hole on every nth skb pattern happens
1671 skb = skb_rb_next(prev);
1675 if (!skb_can_shift(skb) ||
1676 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1677 (mss != tcp_skb_seglen(skb)))
1680 if (!tcp_skb_can_collapse(prev, skb))
1683 pcount = tcp_skb_pcount(skb);
1684 if (tcp_skb_shift(prev, skb, pcount, len))
1685 tcp_shifted_skb(sk, prev, skb, state, pcount,
1695 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1699 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1700 struct tcp_sack_block *next_dup,
1701 struct tcp_sacktag_state *state,
1702 u32 start_seq, u32 end_seq,
1705 struct tcp_sock *tp = tcp_sk(sk);
1706 struct sk_buff *tmp;
1708 skb_rbtree_walk_from(skb) {
1710 bool dup_sack = dup_sack_in;
1712 /* queue is in-order => we can short-circuit the walk early */
1713 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1717 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1718 in_sack = tcp_match_skb_to_sack(sk, skb,
1719 next_dup->start_seq,
1725 /* skb reference here is a bit tricky to get right, since
1726 * shifting can eat and free both this skb and the next,
1727 * so not even _safe variant of the loop is enough.
1730 tmp = tcp_shift_skb_data(sk, skb, state,
1731 start_seq, end_seq, dup_sack);
1740 in_sack = tcp_match_skb_to_sack(sk, skb,
1746 if (unlikely(in_sack < 0))
1750 TCP_SKB_CB(skb)->sacked =
1753 TCP_SKB_CB(skb)->sacked,
1754 TCP_SKB_CB(skb)->seq,
1755 TCP_SKB_CB(skb)->end_seq,
1757 tcp_skb_pcount(skb),
1758 tcp_skb_timestamp_us(skb));
1759 tcp_rate_skb_delivered(sk, skb, state->rate);
1760 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1761 list_del_init(&skb->tcp_tsorted_anchor);
1763 if (!before(TCP_SKB_CB(skb)->seq,
1764 tcp_highest_sack_seq(tp)))
1765 tcp_advance_highest_sack(sk, skb);
1771 static struct sk_buff *tcp_sacktag_bsearch(struct sock *sk, u32 seq)
1773 struct rb_node *parent, **p = &sk->tcp_rtx_queue.rb_node;
1774 struct sk_buff *skb;
1778 skb = rb_to_skb(parent);
1779 if (before(seq, TCP_SKB_CB(skb)->seq)) {
1780 p = &parent->rb_left;
1783 if (!before(seq, TCP_SKB_CB(skb)->end_seq)) {
1784 p = &parent->rb_right;
1792 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1795 if (skb && after(TCP_SKB_CB(skb)->seq, skip_to_seq))
1798 return tcp_sacktag_bsearch(sk, skip_to_seq);
1801 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1803 struct tcp_sack_block *next_dup,
1804 struct tcp_sacktag_state *state,
1810 if (before(next_dup->start_seq, skip_to_seq)) {
1811 skb = tcp_sacktag_skip(skb, sk, next_dup->start_seq);
1812 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1813 next_dup->start_seq, next_dup->end_seq,
1820 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1822 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1826 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1827 u32 prior_snd_una, struct tcp_sacktag_state *state)
1829 struct tcp_sock *tp = tcp_sk(sk);
1830 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1831 TCP_SKB_CB(ack_skb)->sacked);
1832 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1833 struct tcp_sack_block sp[TCP_NUM_SACKS];
1834 struct tcp_sack_block *cache;
1835 struct sk_buff *skb;
1836 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1838 bool found_dup_sack = false;
1840 int first_sack_index;
1843 state->reord = tp->snd_nxt;
1845 if (!tp->sacked_out)
1846 tcp_highest_sack_reset(sk);
1848 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1849 num_sacks, prior_snd_una, state);
1851 /* Eliminate too old ACKs, but take into
1852 * account more or less fresh ones, they can
1853 * contain valid SACK info.
1855 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1858 if (!tp->packets_out)
1862 first_sack_index = 0;
1863 for (i = 0; i < num_sacks; i++) {
1864 bool dup_sack = !i && found_dup_sack;
1866 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1867 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1869 if (!tcp_is_sackblock_valid(tp, dup_sack,
1870 sp[used_sacks].start_seq,
1871 sp[used_sacks].end_seq)) {
1875 if (!tp->undo_marker)
1876 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1878 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1880 /* Don't count olds caused by ACK reordering */
1881 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1882 !after(sp[used_sacks].end_seq, tp->snd_una))
1884 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1887 NET_INC_STATS(sock_net(sk), mib_idx);
1889 first_sack_index = -1;
1893 /* Ignore very old stuff early */
1894 if (!after(sp[used_sacks].end_seq, prior_snd_una)) {
1896 first_sack_index = -1;
1903 /* order SACK blocks to allow in order walk of the retrans queue */
1904 for (i = used_sacks - 1; i > 0; i--) {
1905 for (j = 0; j < i; j++) {
1906 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1907 swap(sp[j], sp[j + 1]);
1909 /* Track where the first SACK block goes to */
1910 if (j == first_sack_index)
1911 first_sack_index = j + 1;
1916 state->mss_now = tcp_current_mss(sk);
1920 if (!tp->sacked_out) {
1921 /* It's already past, so skip checking against it */
1922 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1924 cache = tp->recv_sack_cache;
1925 /* Skip empty blocks in at head of the cache */
1926 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1931 while (i < used_sacks) {
1932 u32 start_seq = sp[i].start_seq;
1933 u32 end_seq = sp[i].end_seq;
1934 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1935 struct tcp_sack_block *next_dup = NULL;
1937 if (found_dup_sack && ((i + 1) == first_sack_index))
1938 next_dup = &sp[i + 1];
1940 /* Skip too early cached blocks */
1941 while (tcp_sack_cache_ok(tp, cache) &&
1942 !before(start_seq, cache->end_seq))
1945 /* Can skip some work by looking recv_sack_cache? */
1946 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1947 after(end_seq, cache->start_seq)) {
1950 if (before(start_seq, cache->start_seq)) {
1951 skb = tcp_sacktag_skip(skb, sk, start_seq);
1952 skb = tcp_sacktag_walk(skb, sk, next_dup,
1959 /* Rest of the block already fully processed? */
1960 if (!after(end_seq, cache->end_seq))
1963 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1967 /* ...tail remains todo... */
1968 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1969 /* ...but better entrypoint exists! */
1970 skb = tcp_highest_sack(sk);
1977 skb = tcp_sacktag_skip(skb, sk, cache->end_seq);
1978 /* Check overlap against next cached too (past this one already) */
1983 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1984 skb = tcp_highest_sack(sk);
1988 skb = tcp_sacktag_skip(skb, sk, start_seq);
1991 skb = tcp_sacktag_walk(skb, sk, next_dup, state,
1992 start_seq, end_seq, dup_sack);
1998 /* Clear the head of the cache sack blocks so we can skip it next time */
1999 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
2000 tp->recv_sack_cache[i].start_seq = 0;
2001 tp->recv_sack_cache[i].end_seq = 0;
2003 for (j = 0; j < used_sacks; j++)
2004 tp->recv_sack_cache[i++] = sp[j];
2006 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Loss || tp->undo_marker)
2007 tcp_check_sack_reordering(sk, state->reord, 0);
2009 tcp_verify_left_out(tp);
2012 #if FASTRETRANS_DEBUG > 0
2013 WARN_ON((int)tp->sacked_out < 0);
2014 WARN_ON((int)tp->lost_out < 0);
2015 WARN_ON((int)tp->retrans_out < 0);
2016 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
2021 /* Limits sacked_out so that sum with lost_out isn't ever larger than
2022 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
2024 static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
2028 holes = max(tp->lost_out, 1U);
2029 holes = min(holes, tp->packets_out);
2031 if ((tp->sacked_out + holes) > tp->packets_out) {
2032 tp->sacked_out = tp->packets_out - holes;
2038 /* If we receive more dupacks than we expected counting segments
2039 * in assumption of absent reordering, interpret this as reordering.
2040 * The only another reason could be bug in receiver TCP.
2042 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
2044 struct tcp_sock *tp = tcp_sk(sk);
2046 if (!tcp_limit_reno_sacked(tp))
2049 tp->reordering = min_t(u32, tp->packets_out + addend,
2050 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_max_reordering));
2052 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRENOREORDER);
2055 /* Emulate SACKs for SACKless connection: account for a new dupack. */
2057 static void tcp_add_reno_sack(struct sock *sk, int num_dupack, bool ece_ack)
2060 struct tcp_sock *tp = tcp_sk(sk);
2061 u32 prior_sacked = tp->sacked_out;
2064 tp->sacked_out += num_dupack;
2065 tcp_check_reno_reordering(sk, 0);
2066 delivered = tp->sacked_out - prior_sacked;
2068 tcp_count_delivered(tp, delivered, ece_ack);
2069 tcp_verify_left_out(tp);
2073 /* Account for ACK, ACKing some data in Reno Recovery phase. */
2075 static void tcp_remove_reno_sacks(struct sock *sk, int acked, bool ece_ack)
2077 struct tcp_sock *tp = tcp_sk(sk);
2080 /* One ACK acked hole. The rest eat duplicate ACKs. */
2081 tcp_count_delivered(tp, max_t(int, acked - tp->sacked_out, 1),
2083 if (acked - 1 >= tp->sacked_out)
2086 tp->sacked_out -= acked - 1;
2088 tcp_check_reno_reordering(sk, acked);
2089 tcp_verify_left_out(tp);
2092 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
2097 void tcp_clear_retrans(struct tcp_sock *tp)
2099 tp->retrans_out = 0;
2101 tp->undo_marker = 0;
2102 tp->undo_retrans = -1;
2106 static inline void tcp_init_undo(struct tcp_sock *tp)
2108 tp->undo_marker = tp->snd_una;
2109 /* Retransmission still in flight may cause DSACKs later. */
2110 tp->undo_retrans = tp->retrans_out ? : -1;
2113 static bool tcp_is_rack(const struct sock *sk)
2115 return READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_recovery) &
2116 TCP_RACK_LOSS_DETECTION;
2119 /* If we detect SACK reneging, forget all SACK information
2120 * and reset tags completely, otherwise preserve SACKs. If receiver
2121 * dropped its ofo queue, we will know this due to reneging detection.
2123 static void tcp_timeout_mark_lost(struct sock *sk)
2125 struct tcp_sock *tp = tcp_sk(sk);
2126 struct sk_buff *skb, *head;
2127 bool is_reneg; /* is receiver reneging on SACKs? */
2129 head = tcp_rtx_queue_head(sk);
2130 is_reneg = head && (TCP_SKB_CB(head)->sacked & TCPCB_SACKED_ACKED);
2132 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
2134 /* Mark SACK reneging until we recover from this loss event. */
2135 tp->is_sack_reneg = 1;
2136 } else if (tcp_is_reno(tp)) {
2137 tcp_reset_reno_sack(tp);
2141 skb_rbtree_walk_from(skb) {
2143 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
2144 else if (tcp_is_rack(sk) && skb != head &&
2145 tcp_rack_skb_timeout(tp, skb, 0) > 0)
2146 continue; /* Don't mark recently sent ones lost yet */
2147 tcp_mark_skb_lost(sk, skb);
2149 tcp_verify_left_out(tp);
2150 tcp_clear_all_retrans_hints(tp);
2153 /* Enter Loss state. */
2154 void tcp_enter_loss(struct sock *sk)
2156 const struct inet_connection_sock *icsk = inet_csk(sk);
2157 struct tcp_sock *tp = tcp_sk(sk);
2158 struct net *net = sock_net(sk);
2159 bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery;
2162 tcp_timeout_mark_lost(sk);
2164 /* Reduce ssthresh if it has not yet been made inside this window. */
2165 if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
2166 !after(tp->high_seq, tp->snd_una) ||
2167 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
2168 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2169 tp->prior_cwnd = tcp_snd_cwnd(tp);
2170 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2171 tcp_ca_event(sk, CA_EVENT_LOSS);
2174 tcp_snd_cwnd_set(tp, tcp_packets_in_flight(tp) + 1);
2175 tp->snd_cwnd_cnt = 0;
2176 tp->snd_cwnd_stamp = tcp_jiffies32;
2178 /* Timeout in disordered state after receiving substantial DUPACKs
2179 * suggests that the degree of reordering is over-estimated.
2181 reordering = READ_ONCE(net->ipv4.sysctl_tcp_reordering);
2182 if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
2183 tp->sacked_out >= reordering)
2184 tp->reordering = min_t(unsigned int, tp->reordering,
2187 tcp_set_ca_state(sk, TCP_CA_Loss);
2188 tp->high_seq = tp->snd_nxt;
2189 tcp_ecn_queue_cwr(tp);
2191 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
2192 * loss recovery is underway except recurring timeout(s) on
2193 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
2195 tp->frto = READ_ONCE(net->ipv4.sysctl_tcp_frto) &&
2196 (new_recovery || icsk->icsk_retransmits) &&
2197 !inet_csk(sk)->icsk_mtup.probe_size;
2200 /* If ACK arrived pointing to a remembered SACK, it means that our
2201 * remembered SACKs do not reflect real state of receiver i.e.
2202 * receiver _host_ is heavily congested (or buggy).
2204 * To avoid big spurious retransmission bursts due to transient SACK
2205 * scoreboard oddities that look like reneging, we give the receiver a
2206 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
2207 * restore sanity to the SACK scoreboard. If the apparent reneging
2208 * persists until this RTO then we'll clear the SACK scoreboard.
2210 static bool tcp_check_sack_reneging(struct sock *sk, int *ack_flag)
2212 if (*ack_flag & FLAG_SACK_RENEGING &&
2213 *ack_flag & FLAG_SND_UNA_ADVANCED) {
2214 struct tcp_sock *tp = tcp_sk(sk);
2215 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4),
2216 msecs_to_jiffies(10));
2218 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2219 delay, TCP_RTO_MAX);
2220 *ack_flag &= ~FLAG_SET_XMIT_TIMER;
2226 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2227 * counter when SACK is enabled (without SACK, sacked_out is used for
2230 * With reordering, holes may still be in flight, so RFC3517 recovery
2231 * uses pure sacked_out (total number of SACKed segments) even though
2232 * it violates the RFC that uses duplicate ACKs, often these are equal
2233 * but when e.g. out-of-window ACKs or packet duplication occurs,
2234 * they differ. Since neither occurs due to loss, TCP should really
2237 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2239 return tp->sacked_out + 1;
2242 /* Linux NewReno/SACK/ECN state machine.
2243 * --------------------------------------
2245 * "Open" Normal state, no dubious events, fast path.
2246 * "Disorder" In all the respects it is "Open",
2247 * but requires a bit more attention. It is entered when
2248 * we see some SACKs or dupacks. It is split of "Open"
2249 * mainly to move some processing from fast path to slow one.
2250 * "CWR" CWND was reduced due to some Congestion Notification event.
2251 * It can be ECN, ICMP source quench, local device congestion.
2252 * "Recovery" CWND was reduced, we are fast-retransmitting.
2253 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2255 * tcp_fastretrans_alert() is entered:
2256 * - each incoming ACK, if state is not "Open"
2257 * - when arrived ACK is unusual, namely:
2262 * Counting packets in flight is pretty simple.
2264 * in_flight = packets_out - left_out + retrans_out
2266 * packets_out is SND.NXT-SND.UNA counted in packets.
2268 * retrans_out is number of retransmitted segments.
2270 * left_out is number of segments left network, but not ACKed yet.
2272 * left_out = sacked_out + lost_out
2274 * sacked_out: Packets, which arrived to receiver out of order
2275 * and hence not ACKed. With SACKs this number is simply
2276 * amount of SACKed data. Even without SACKs
2277 * it is easy to give pretty reliable estimate of this number,
2278 * counting duplicate ACKs.
2280 * lost_out: Packets lost by network. TCP has no explicit
2281 * "loss notification" feedback from network (for now).
2282 * It means that this number can be only _guessed_.
2283 * Actually, it is the heuristics to predict lossage that
2284 * distinguishes different algorithms.
2286 * F.e. after RTO, when all the queue is considered as lost,
2287 * lost_out = packets_out and in_flight = retrans_out.
2289 * Essentially, we have now a few algorithms detecting
2292 * If the receiver supports SACK:
2294 * RFC6675/3517: It is the conventional algorithm. A packet is
2295 * considered lost if the number of higher sequence packets
2296 * SACKed is greater than or equal the DUPACK thoreshold
2297 * (reordering). This is implemented in tcp_mark_head_lost and
2298 * tcp_update_scoreboard.
2300 * RACK (draft-ietf-tcpm-rack-01): it is a newer algorithm
2301 * (2017-) that checks timing instead of counting DUPACKs.
2302 * Essentially a packet is considered lost if it's not S/ACKed
2303 * after RTT + reordering_window, where both metrics are
2304 * dynamically measured and adjusted. This is implemented in
2305 * tcp_rack_mark_lost.
2307 * If the receiver does not support SACK:
2309 * NewReno (RFC6582): in Recovery we assume that one segment
2310 * is lost (classic Reno). While we are in Recovery and
2311 * a partial ACK arrives, we assume that one more packet
2312 * is lost (NewReno). This heuristics are the same in NewReno
2315 * Really tricky (and requiring careful tuning) part of algorithm
2316 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2317 * The first determines the moment _when_ we should reduce CWND and,
2318 * hence, slow down forward transmission. In fact, it determines the moment
2319 * when we decide that hole is caused by loss, rather than by a reorder.
2321 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2322 * holes, caused by lost packets.
2324 * And the most logically complicated part of algorithm is undo
2325 * heuristics. We detect false retransmits due to both too early
2326 * fast retransmit (reordering) and underestimated RTO, analyzing
2327 * timestamps and D-SACKs. When we detect that some segments were
2328 * retransmitted by mistake and CWND reduction was wrong, we undo
2329 * window reduction and abort recovery phase. This logic is hidden
2330 * inside several functions named tcp_try_undo_<something>.
2333 /* This function decides, when we should leave Disordered state
2334 * and enter Recovery phase, reducing congestion window.
2336 * Main question: may we further continue forward transmission
2337 * with the same cwnd?
2339 static bool tcp_time_to_recover(struct sock *sk, int flag)
2341 struct tcp_sock *tp = tcp_sk(sk);
2343 /* Trick#1: The loss is proven. */
2347 /* Not-A-Trick#2 : Classic rule... */
2348 if (!tcp_is_rack(sk) && tcp_dupack_heuristics(tp) > tp->reordering)
2354 /* Detect loss in event "A" above by marking head of queue up as lost.
2355 * For RFC3517 SACK, a segment is considered lost if it
2356 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2357 * the maximum SACKed segments to pass before reaching this limit.
2359 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2361 struct tcp_sock *tp = tcp_sk(sk);
2362 struct sk_buff *skb;
2364 /* Use SACK to deduce losses of new sequences sent during recovery */
2365 const u32 loss_high = tp->snd_nxt;
2367 WARN_ON(packets > tp->packets_out);
2368 skb = tp->lost_skb_hint;
2370 /* Head already handled? */
2371 if (mark_head && after(TCP_SKB_CB(skb)->seq, tp->snd_una))
2373 cnt = tp->lost_cnt_hint;
2375 skb = tcp_rtx_queue_head(sk);
2379 skb_rbtree_walk_from(skb) {
2380 /* TODO: do this better */
2381 /* this is not the most efficient way to do this... */
2382 tp->lost_skb_hint = skb;
2383 tp->lost_cnt_hint = cnt;
2385 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2388 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2389 cnt += tcp_skb_pcount(skb);
2394 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_LOST))
2395 tcp_mark_skb_lost(sk, skb);
2400 tcp_verify_left_out(tp);
2403 /* Account newly detected lost packet(s) */
2405 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2407 struct tcp_sock *tp = tcp_sk(sk);
2409 if (tcp_is_sack(tp)) {
2410 int sacked_upto = tp->sacked_out - tp->reordering;
2411 if (sacked_upto >= 0)
2412 tcp_mark_head_lost(sk, sacked_upto, 0);
2413 else if (fast_rexmit)
2414 tcp_mark_head_lost(sk, 1, 1);
2418 static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when)
2420 return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2421 before(tp->rx_opt.rcv_tsecr, when);
2424 /* skb is spurious retransmitted if the returned timestamp echo
2425 * reply is prior to the skb transmission time
2427 static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp,
2428 const struct sk_buff *skb)
2430 return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) &&
2431 tcp_tsopt_ecr_before(tp, tcp_skb_timestamp(skb));
2434 /* Nothing was retransmitted or returned timestamp is less
2435 * than timestamp of the first retransmission.
2437 static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2439 return tp->retrans_stamp &&
2440 tcp_tsopt_ecr_before(tp, tp->retrans_stamp);
2443 /* Undo procedures. */
2445 /* We can clear retrans_stamp when there are no retransmissions in the
2446 * window. It would seem that it is trivially available for us in
2447 * tp->retrans_out, however, that kind of assumptions doesn't consider
2448 * what will happen if errors occur when sending retransmission for the
2449 * second time. ...It could the that such segment has only
2450 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2451 * the head skb is enough except for some reneging corner cases that
2452 * are not worth the effort.
2454 * Main reason for all this complexity is the fact that connection dying
2455 * time now depends on the validity of the retrans_stamp, in particular,
2456 * that successive retransmissions of a segment must not advance
2457 * retrans_stamp under any conditions.
2459 static bool tcp_any_retrans_done(const struct sock *sk)
2461 const struct tcp_sock *tp = tcp_sk(sk);
2462 struct sk_buff *skb;
2464 if (tp->retrans_out)
2467 skb = tcp_rtx_queue_head(sk);
2468 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2474 static void DBGUNDO(struct sock *sk, const char *msg)
2476 #if FASTRETRANS_DEBUG > 1
2477 struct tcp_sock *tp = tcp_sk(sk);
2478 struct inet_sock *inet = inet_sk(sk);
2480 if (sk->sk_family == AF_INET) {
2481 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2483 &inet->inet_daddr, ntohs(inet->inet_dport),
2484 tcp_snd_cwnd(tp), tcp_left_out(tp),
2485 tp->snd_ssthresh, tp->prior_ssthresh,
2488 #if IS_ENABLED(CONFIG_IPV6)
2489 else if (sk->sk_family == AF_INET6) {
2490 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2492 &sk->sk_v6_daddr, ntohs(inet->inet_dport),
2493 tcp_snd_cwnd(tp), tcp_left_out(tp),
2494 tp->snd_ssthresh, tp->prior_ssthresh,
2501 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
2503 struct tcp_sock *tp = tcp_sk(sk);
2506 struct sk_buff *skb;
2508 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2509 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2512 tcp_clear_all_retrans_hints(tp);
2515 if (tp->prior_ssthresh) {
2516 const struct inet_connection_sock *icsk = inet_csk(sk);
2518 tcp_snd_cwnd_set(tp, icsk->icsk_ca_ops->undo_cwnd(sk));
2520 if (tp->prior_ssthresh > tp->snd_ssthresh) {
2521 tp->snd_ssthresh = tp->prior_ssthresh;
2522 tcp_ecn_withdraw_cwr(tp);
2525 tp->snd_cwnd_stamp = tcp_jiffies32;
2526 tp->undo_marker = 0;
2527 tp->rack.advanced = 1; /* Force RACK to re-exam losses */
2530 static inline bool tcp_may_undo(const struct tcp_sock *tp)
2532 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2535 static bool tcp_is_non_sack_preventing_reopen(struct sock *sk)
2537 struct tcp_sock *tp = tcp_sk(sk);
2539 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2540 /* Hold old state until something *above* high_seq
2541 * is ACKed. For Reno it is MUST to prevent false
2542 * fast retransmits (RFC2582). SACK TCP is safe. */
2543 if (!tcp_any_retrans_done(sk))
2544 tp->retrans_stamp = 0;
2550 /* People celebrate: "We love our President!" */
2551 static bool tcp_try_undo_recovery(struct sock *sk)
2553 struct tcp_sock *tp = tcp_sk(sk);
2555 if (tcp_may_undo(tp)) {
2558 /* Happy end! We did not retransmit anything
2559 * or our original transmission succeeded.
2561 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2562 tcp_undo_cwnd_reduction(sk, false);
2563 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2564 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2566 mib_idx = LINUX_MIB_TCPFULLUNDO;
2568 NET_INC_STATS(sock_net(sk), mib_idx);
2569 } else if (tp->rack.reo_wnd_persist) {
2570 tp->rack.reo_wnd_persist--;
2572 if (tcp_is_non_sack_preventing_reopen(sk))
2574 tcp_set_ca_state(sk, TCP_CA_Open);
2575 tp->is_sack_reneg = 0;
2579 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2580 static bool tcp_try_undo_dsack(struct sock *sk)
2582 struct tcp_sock *tp = tcp_sk(sk);
2584 if (tp->undo_marker && !tp->undo_retrans) {
2585 tp->rack.reo_wnd_persist = min(TCP_RACK_RECOVERY_THRESH,
2586 tp->rack.reo_wnd_persist + 1);
2587 DBGUNDO(sk, "D-SACK");
2588 tcp_undo_cwnd_reduction(sk, false);
2589 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2595 /* Undo during loss recovery after partial ACK or using F-RTO. */
2596 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2598 struct tcp_sock *tp = tcp_sk(sk);
2600 if (frto_undo || tcp_may_undo(tp)) {
2601 tcp_undo_cwnd_reduction(sk, true);
2603 DBGUNDO(sk, "partial loss");
2604 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2606 NET_INC_STATS(sock_net(sk),
2607 LINUX_MIB_TCPSPURIOUSRTOS);
2608 inet_csk(sk)->icsk_retransmits = 0;
2609 if (tcp_is_non_sack_preventing_reopen(sk))
2611 if (frto_undo || tcp_is_sack(tp)) {
2612 tcp_set_ca_state(sk, TCP_CA_Open);
2613 tp->is_sack_reneg = 0;
2620 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
2621 * It computes the number of packets to send (sndcnt) based on packets newly
2623 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2624 * cwnd reductions across a full RTT.
2625 * 2) Otherwise PRR uses packet conservation to send as much as delivered.
2626 * But when SND_UNA is acked without further losses,
2627 * slow starts cwnd up to ssthresh to speed up the recovery.
2629 static void tcp_init_cwnd_reduction(struct sock *sk)
2631 struct tcp_sock *tp = tcp_sk(sk);
2633 tp->high_seq = tp->snd_nxt;
2634 tp->tlp_high_seq = 0;
2635 tp->snd_cwnd_cnt = 0;
2636 tp->prior_cwnd = tcp_snd_cwnd(tp);
2637 tp->prr_delivered = 0;
2639 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2640 tcp_ecn_queue_cwr(tp);
2643 void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked, int newly_lost, int flag)
2645 struct tcp_sock *tp = tcp_sk(sk);
2647 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2649 if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd))
2652 tp->prr_delivered += newly_acked_sacked;
2654 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2656 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2658 sndcnt = max_t(int, tp->prr_delivered - tp->prr_out,
2659 newly_acked_sacked);
2660 if (flag & FLAG_SND_UNA_ADVANCED && !newly_lost)
2662 sndcnt = min(delta, sndcnt);
2664 /* Force a fast retransmit upon entering fast recovery */
2665 sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1));
2666 tcp_snd_cwnd_set(tp, tcp_packets_in_flight(tp) + sndcnt);
2669 static inline void tcp_end_cwnd_reduction(struct sock *sk)
2671 struct tcp_sock *tp = tcp_sk(sk);
2673 if (inet_csk(sk)->icsk_ca_ops->cong_control)
2676 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2677 if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH &&
2678 (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR || tp->undo_marker)) {
2679 tcp_snd_cwnd_set(tp, tp->snd_ssthresh);
2680 tp->snd_cwnd_stamp = tcp_jiffies32;
2682 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2685 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2686 void tcp_enter_cwr(struct sock *sk)
2688 struct tcp_sock *tp = tcp_sk(sk);
2690 tp->prior_ssthresh = 0;
2691 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2692 tp->undo_marker = 0;
2693 tcp_init_cwnd_reduction(sk);
2694 tcp_set_ca_state(sk, TCP_CA_CWR);
2697 EXPORT_SYMBOL(tcp_enter_cwr);
2699 static void tcp_try_keep_open(struct sock *sk)
2701 struct tcp_sock *tp = tcp_sk(sk);
2702 int state = TCP_CA_Open;
2704 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2705 state = TCP_CA_Disorder;
2707 if (inet_csk(sk)->icsk_ca_state != state) {
2708 tcp_set_ca_state(sk, state);
2709 tp->high_seq = tp->snd_nxt;
2713 static void tcp_try_to_open(struct sock *sk, int flag)
2715 struct tcp_sock *tp = tcp_sk(sk);
2717 tcp_verify_left_out(tp);
2719 if (!tcp_any_retrans_done(sk))
2720 tp->retrans_stamp = 0;
2722 if (flag & FLAG_ECE)
2725 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2726 tcp_try_keep_open(sk);
2730 static void tcp_mtup_probe_failed(struct sock *sk)
2732 struct inet_connection_sock *icsk = inet_csk(sk);
2734 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2735 icsk->icsk_mtup.probe_size = 0;
2736 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPFAIL);
2739 static void tcp_mtup_probe_success(struct sock *sk)
2741 struct tcp_sock *tp = tcp_sk(sk);
2742 struct inet_connection_sock *icsk = inet_csk(sk);
2745 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2747 val = (u64)tcp_snd_cwnd(tp) * tcp_mss_to_mtu(sk, tp->mss_cache);
2748 do_div(val, icsk->icsk_mtup.probe_size);
2749 DEBUG_NET_WARN_ON_ONCE((u32)val != val);
2750 tcp_snd_cwnd_set(tp, max_t(u32, 1U, val));
2752 tp->snd_cwnd_cnt = 0;
2753 tp->snd_cwnd_stamp = tcp_jiffies32;
2754 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2756 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2757 icsk->icsk_mtup.probe_size = 0;
2758 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2759 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS);
2762 /* Do a simple retransmit without using the backoff mechanisms in
2763 * tcp_timer. This is used for path mtu discovery.
2764 * The socket is already locked here.
2766 void tcp_simple_retransmit(struct sock *sk)
2768 const struct inet_connection_sock *icsk = inet_csk(sk);
2769 struct tcp_sock *tp = tcp_sk(sk);
2770 struct sk_buff *skb;
2773 /* A fastopen SYN request is stored as two separate packets within
2774 * the retransmit queue, this is done by tcp_send_syn_data().
2775 * As a result simply checking the MSS of the frames in the queue
2776 * will not work for the SYN packet.
2778 * Us being here is an indication of a path MTU issue so we can
2779 * assume that the fastopen SYN was lost and just mark all the
2780 * frames in the retransmit queue as lost. We will use an MSS of
2781 * -1 to mark all frames as lost, otherwise compute the current MSS.
2783 if (tp->syn_data && sk->sk_state == TCP_SYN_SENT)
2786 mss = tcp_current_mss(sk);
2788 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2789 if (tcp_skb_seglen(skb) > mss)
2790 tcp_mark_skb_lost(sk, skb);
2793 tcp_clear_retrans_hints_partial(tp);
2798 if (tcp_is_reno(tp))
2799 tcp_limit_reno_sacked(tp);
2801 tcp_verify_left_out(tp);
2803 /* Don't muck with the congestion window here.
2804 * Reason is that we do not increase amount of _data_
2805 * in network, but units changed and effective
2806 * cwnd/ssthresh really reduced now.
2808 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2809 tp->high_seq = tp->snd_nxt;
2810 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2811 tp->prior_ssthresh = 0;
2812 tp->undo_marker = 0;
2813 tcp_set_ca_state(sk, TCP_CA_Loss);
2815 tcp_xmit_retransmit_queue(sk);
2817 EXPORT_SYMBOL(tcp_simple_retransmit);
2819 void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2821 struct tcp_sock *tp = tcp_sk(sk);
2824 if (tcp_is_reno(tp))
2825 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2827 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2829 NET_INC_STATS(sock_net(sk), mib_idx);
2831 tp->prior_ssthresh = 0;
2834 if (!tcp_in_cwnd_reduction(sk)) {
2836 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2837 tcp_init_cwnd_reduction(sk);
2839 tcp_set_ca_state(sk, TCP_CA_Recovery);
2842 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2843 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2845 static void tcp_process_loss(struct sock *sk, int flag, int num_dupack,
2848 struct tcp_sock *tp = tcp_sk(sk);
2849 bool recovered = !before(tp->snd_una, tp->high_seq);
2851 if ((flag & FLAG_SND_UNA_ADVANCED || rcu_access_pointer(tp->fastopen_rsk)) &&
2852 tcp_try_undo_loss(sk, false))
2855 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2856 /* Step 3.b. A timeout is spurious if not all data are
2857 * lost, i.e., never-retransmitted data are (s)acked.
2859 if ((flag & FLAG_ORIG_SACK_ACKED) &&
2860 tcp_try_undo_loss(sk, true))
2863 if (after(tp->snd_nxt, tp->high_seq)) {
2864 if (flag & FLAG_DATA_SACKED || num_dupack)
2865 tp->frto = 0; /* Step 3.a. loss was real */
2866 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2867 tp->high_seq = tp->snd_nxt;
2868 /* Step 2.b. Try send new data (but deferred until cwnd
2869 * is updated in tcp_ack()). Otherwise fall back to
2870 * the conventional recovery.
2872 if (!tcp_write_queue_empty(sk) &&
2873 after(tcp_wnd_end(tp), tp->snd_nxt)) {
2874 *rexmit = REXMIT_NEW;
2882 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2883 tcp_try_undo_recovery(sk);
2886 if (tcp_is_reno(tp)) {
2887 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2888 * delivered. Lower inflight to clock out (re)transmissions.
2890 if (after(tp->snd_nxt, tp->high_seq) && num_dupack)
2891 tcp_add_reno_sack(sk, num_dupack, flag & FLAG_ECE);
2892 else if (flag & FLAG_SND_UNA_ADVANCED)
2893 tcp_reset_reno_sack(tp);
2895 *rexmit = REXMIT_LOST;
2898 static bool tcp_force_fast_retransmit(struct sock *sk)
2900 struct tcp_sock *tp = tcp_sk(sk);
2902 return after(tcp_highest_sack_seq(tp),
2903 tp->snd_una + tp->reordering * tp->mss_cache);
2906 /* Undo during fast recovery after partial ACK. */
2907 static bool tcp_try_undo_partial(struct sock *sk, u32 prior_snd_una,
2910 struct tcp_sock *tp = tcp_sk(sk);
2912 if (tp->undo_marker && tcp_packet_delayed(tp)) {
2913 /* Plain luck! Hole if filled with delayed
2914 * packet, rather than with a retransmit. Check reordering.
2916 tcp_check_sack_reordering(sk, prior_snd_una, 1);
2918 /* We are getting evidence that the reordering degree is higher
2919 * than we realized. If there are no retransmits out then we
2920 * can undo. Otherwise we clock out new packets but do not
2921 * mark more packets lost or retransmit more.
2923 if (tp->retrans_out)
2926 if (!tcp_any_retrans_done(sk))
2927 tp->retrans_stamp = 0;
2929 DBGUNDO(sk, "partial recovery");
2930 tcp_undo_cwnd_reduction(sk, true);
2931 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2932 tcp_try_keep_open(sk);
2934 /* Partial ACK arrived. Force fast retransmit. */
2935 *do_lost = tcp_force_fast_retransmit(sk);
2940 static void tcp_identify_packet_loss(struct sock *sk, int *ack_flag)
2942 struct tcp_sock *tp = tcp_sk(sk);
2944 if (tcp_rtx_queue_empty(sk))
2947 if (unlikely(tcp_is_reno(tp))) {
2948 tcp_newreno_mark_lost(sk, *ack_flag & FLAG_SND_UNA_ADVANCED);
2949 } else if (tcp_is_rack(sk)) {
2950 u32 prior_retrans = tp->retrans_out;
2952 if (tcp_rack_mark_lost(sk))
2953 *ack_flag &= ~FLAG_SET_XMIT_TIMER;
2954 if (prior_retrans > tp->retrans_out)
2955 *ack_flag |= FLAG_LOST_RETRANS;
2959 /* Process an event, which can update packets-in-flight not trivially.
2960 * Main goal of this function is to calculate new estimate for left_out,
2961 * taking into account both packets sitting in receiver's buffer and
2962 * packets lost by network.
2964 * Besides that it updates the congestion state when packet loss or ECN
2965 * is detected. But it does not reduce the cwnd, it is done by the
2966 * congestion control later.
2968 * It does _not_ decide what to send, it is made in function
2969 * tcp_xmit_retransmit_queue().
2971 static void tcp_fastretrans_alert(struct sock *sk, const u32 prior_snd_una,
2972 int num_dupack, int *ack_flag, int *rexmit)
2974 struct inet_connection_sock *icsk = inet_csk(sk);
2975 struct tcp_sock *tp = tcp_sk(sk);
2976 int fast_rexmit = 0, flag = *ack_flag;
2977 bool ece_ack = flag & FLAG_ECE;
2978 bool do_lost = num_dupack || ((flag & FLAG_DATA_SACKED) &&
2979 tcp_force_fast_retransmit(sk));
2981 if (!tp->packets_out && tp->sacked_out)
2984 /* Now state machine starts.
2985 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2987 tp->prior_ssthresh = 0;
2989 /* B. In all the states check for reneging SACKs. */
2990 if (tcp_check_sack_reneging(sk, ack_flag))
2993 /* C. Check consistency of the current state. */
2994 tcp_verify_left_out(tp);
2996 /* D. Check state exit conditions. State can be terminated
2997 * when high_seq is ACKed. */
2998 if (icsk->icsk_ca_state == TCP_CA_Open) {
2999 WARN_ON(tp->retrans_out != 0 && !tp->syn_data);
3000 tp->retrans_stamp = 0;
3001 } else if (!before(tp->snd_una, tp->high_seq)) {
3002 switch (icsk->icsk_ca_state) {
3004 /* CWR is to be held something *above* high_seq
3005 * is ACKed for CWR bit to reach receiver. */
3006 if (tp->snd_una != tp->high_seq) {
3007 tcp_end_cwnd_reduction(sk);
3008 tcp_set_ca_state(sk, TCP_CA_Open);
3012 case TCP_CA_Recovery:
3013 if (tcp_is_reno(tp))
3014 tcp_reset_reno_sack(tp);
3015 if (tcp_try_undo_recovery(sk))
3017 tcp_end_cwnd_reduction(sk);
3022 /* E. Process state. */
3023 switch (icsk->icsk_ca_state) {
3024 case TCP_CA_Recovery:
3025 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
3026 if (tcp_is_reno(tp))
3027 tcp_add_reno_sack(sk, num_dupack, ece_ack);
3028 } else if (tcp_try_undo_partial(sk, prior_snd_una, &do_lost))
3031 if (tcp_try_undo_dsack(sk))
3032 tcp_try_keep_open(sk);
3034 tcp_identify_packet_loss(sk, ack_flag);
3035 if (icsk->icsk_ca_state != TCP_CA_Recovery) {
3036 if (!tcp_time_to_recover(sk, flag))
3038 /* Undo reverts the recovery state. If loss is evident,
3039 * starts a new recovery (e.g. reordering then loss);
3041 tcp_enter_recovery(sk, ece_ack);
3045 tcp_process_loss(sk, flag, num_dupack, rexmit);
3046 tcp_identify_packet_loss(sk, ack_flag);
3047 if (!(icsk->icsk_ca_state == TCP_CA_Open ||
3048 (*ack_flag & FLAG_LOST_RETRANS)))
3050 /* Change state if cwnd is undone or retransmits are lost */
3053 if (tcp_is_reno(tp)) {
3054 if (flag & FLAG_SND_UNA_ADVANCED)
3055 tcp_reset_reno_sack(tp);
3056 tcp_add_reno_sack(sk, num_dupack, ece_ack);
3059 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
3060 tcp_try_undo_dsack(sk);
3062 tcp_identify_packet_loss(sk, ack_flag);
3063 if (!tcp_time_to_recover(sk, flag)) {
3064 tcp_try_to_open(sk, flag);
3068 /* MTU probe failure: don't reduce cwnd */
3069 if (icsk->icsk_ca_state < TCP_CA_CWR &&
3070 icsk->icsk_mtup.probe_size &&
3071 tp->snd_una == tp->mtu_probe.probe_seq_start) {
3072 tcp_mtup_probe_failed(sk);
3073 /* Restores the reduction we did in tcp_mtup_probe() */
3074 tcp_snd_cwnd_set(tp, tcp_snd_cwnd(tp) + 1);
3075 tcp_simple_retransmit(sk);
3079 /* Otherwise enter Recovery state */
3080 tcp_enter_recovery(sk, ece_ack);
3084 if (!tcp_is_rack(sk) && do_lost)
3085 tcp_update_scoreboard(sk, fast_rexmit);
3086 *rexmit = REXMIT_LOST;
3089 static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us, const int flag)
3091 u32 wlen = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_rtt_wlen) * HZ;
3092 struct tcp_sock *tp = tcp_sk(sk);
3094 if ((flag & FLAG_ACK_MAYBE_DELAYED) && rtt_us > tcp_min_rtt(tp)) {
3095 /* If the remote keeps returning delayed ACKs, eventually
3096 * the min filter would pick it up and overestimate the
3097 * prop. delay when it expires. Skip suspected delayed ACKs.
3101 minmax_running_min(&tp->rtt_min, wlen, tcp_jiffies32,
3102 rtt_us ? : jiffies_to_usecs(1));
3105 static bool tcp_ack_update_rtt(struct sock *sk, const int flag,
3106 long seq_rtt_us, long sack_rtt_us,
3107 long ca_rtt_us, struct rate_sample *rs)
3109 const struct tcp_sock *tp = tcp_sk(sk);
3111 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
3112 * broken middle-boxes or peers may corrupt TS-ECR fields. But
3113 * Karn's algorithm forbids taking RTT if some retransmitted data
3114 * is acked (RFC6298).
3117 seq_rtt_us = sack_rtt_us;
3119 /* RTTM Rule: A TSecr value received in a segment is used to
3120 * update the averaged RTT measurement only if the segment
3121 * acknowledges some new data, i.e., only if it advances the
3122 * left edge of the send window.
3123 * See draft-ietf-tcplw-high-performance-00, section 3.3.
3125 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
3126 flag & FLAG_ACKED) {
3127 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr;
3129 if (likely(delta < INT_MAX / (USEC_PER_SEC / TCP_TS_HZ))) {
3132 seq_rtt_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
3133 ca_rtt_us = seq_rtt_us;
3136 rs->rtt_us = ca_rtt_us; /* RTT of last (S)ACKed packet (or -1) */
3140 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
3141 * always taken together with ACK, SACK, or TS-opts. Any negative
3142 * values will be skipped with the seq_rtt_us < 0 check above.
3144 tcp_update_rtt_min(sk, ca_rtt_us, flag);
3145 tcp_rtt_estimator(sk, seq_rtt_us);
3148 /* RFC6298: only reset backoff on valid RTT measurement. */
3149 inet_csk(sk)->icsk_backoff = 0;
3153 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
3154 void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req)
3156 struct rate_sample rs;
3159 if (req && !req->num_retrans && tcp_rsk(req)->snt_synack)
3160 rtt_us = tcp_stamp_us_delta(tcp_clock_us(), tcp_rsk(req)->snt_synack);
3162 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us, &rs);
3166 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
3168 const struct inet_connection_sock *icsk = inet_csk(sk);
3170 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
3171 tcp_sk(sk)->snd_cwnd_stamp = tcp_jiffies32;
3174 /* Restart timer after forward progress on connection.
3175 * RFC2988 recommends to restart timer to now+rto.
3177 void tcp_rearm_rto(struct sock *sk)
3179 const struct inet_connection_sock *icsk = inet_csk(sk);
3180 struct tcp_sock *tp = tcp_sk(sk);
3182 /* If the retrans timer is currently being used by Fast Open
3183 * for SYN-ACK retrans purpose, stay put.
3185 if (rcu_access_pointer(tp->fastopen_rsk))
3188 if (!tp->packets_out) {
3189 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
3191 u32 rto = inet_csk(sk)->icsk_rto;
3192 /* Offset the time elapsed after installing regular RTO */
3193 if (icsk->icsk_pending == ICSK_TIME_REO_TIMEOUT ||
3194 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
3195 s64 delta_us = tcp_rto_delta_us(sk);
3196 /* delta_us may not be positive if the socket is locked
3197 * when the retrans timer fires and is rescheduled.
3199 rto = usecs_to_jiffies(max_t(int, delta_us, 1));
3201 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
3206 /* Try to schedule a loss probe; if that doesn't work, then schedule an RTO. */
3207 static void tcp_set_xmit_timer(struct sock *sk)
3209 if (!tcp_schedule_loss_probe(sk, true))
3213 /* If we get here, the whole TSO packet has not been acked. */
3214 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3216 struct tcp_sock *tp = tcp_sk(sk);
3219 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3221 packets_acked = tcp_skb_pcount(skb);
3222 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3224 packets_acked -= tcp_skb_pcount(skb);
3226 if (packets_acked) {
3227 BUG_ON(tcp_skb_pcount(skb) == 0);
3228 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3231 return packets_acked;
3234 static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
3235 const struct sk_buff *ack_skb, u32 prior_snd_una)
3237 const struct skb_shared_info *shinfo;
3239 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3240 if (likely(!TCP_SKB_CB(skb)->txstamp_ack))
3243 shinfo = skb_shinfo(skb);
3244 if (!before(shinfo->tskey, prior_snd_una) &&
3245 before(shinfo->tskey, tcp_sk(sk)->snd_una)) {
3246 tcp_skb_tsorted_save(skb) {
3247 __skb_tstamp_tx(skb, ack_skb, NULL, sk, SCM_TSTAMP_ACK);
3248 } tcp_skb_tsorted_restore(skb);
3252 /* Remove acknowledged frames from the retransmission queue. If our packet
3253 * is before the ack sequence we can discard it as it's confirmed to have
3254 * arrived at the other end.
3256 static int tcp_clean_rtx_queue(struct sock *sk, const struct sk_buff *ack_skb,
3257 u32 prior_fack, u32 prior_snd_una,
3258 struct tcp_sacktag_state *sack, bool ece_ack)
3260 const struct inet_connection_sock *icsk = inet_csk(sk);
3261 u64 first_ackt, last_ackt;
3262 struct tcp_sock *tp = tcp_sk(sk);
3263 u32 prior_sacked = tp->sacked_out;
3264 u32 reord = tp->snd_nxt; /* lowest acked un-retx un-sacked seq */
3265 struct sk_buff *skb, *next;
3266 bool fully_acked = true;
3267 long sack_rtt_us = -1L;
3268 long seq_rtt_us = -1L;
3269 long ca_rtt_us = -1L;
3276 for (skb = skb_rb_first(&sk->tcp_rtx_queue); skb; skb = next) {
3277 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3278 const u32 start_seq = scb->seq;
3279 u8 sacked = scb->sacked;
3282 /* Determine how many packets and what bytes were acked, tso and else */
3283 if (after(scb->end_seq, tp->snd_una)) {
3284 if (tcp_skb_pcount(skb) == 1 ||
3285 !after(tp->snd_una, scb->seq))
3288 acked_pcount = tcp_tso_acked(sk, skb);
3291 fully_acked = false;
3293 acked_pcount = tcp_skb_pcount(skb);
3296 if (unlikely(sacked & TCPCB_RETRANS)) {
3297 if (sacked & TCPCB_SACKED_RETRANS)
3298 tp->retrans_out -= acked_pcount;
3299 flag |= FLAG_RETRANS_DATA_ACKED;
3300 } else if (!(sacked & TCPCB_SACKED_ACKED)) {
3301 last_ackt = tcp_skb_timestamp_us(skb);
3302 WARN_ON_ONCE(last_ackt == 0);
3304 first_ackt = last_ackt;
3306 if (before(start_seq, reord))
3308 if (!after(scb->end_seq, tp->high_seq))
3309 flag |= FLAG_ORIG_SACK_ACKED;
3312 if (sacked & TCPCB_SACKED_ACKED) {
3313 tp->sacked_out -= acked_pcount;
3314 } else if (tcp_is_sack(tp)) {
3315 tcp_count_delivered(tp, acked_pcount, ece_ack);
3316 if (!tcp_skb_spurious_retrans(tp, skb))
3317 tcp_rack_advance(tp, sacked, scb->end_seq,
3318 tcp_skb_timestamp_us(skb));
3320 if (sacked & TCPCB_LOST)
3321 tp->lost_out -= acked_pcount;
3323 tp->packets_out -= acked_pcount;
3324 pkts_acked += acked_pcount;
3325 tcp_rate_skb_delivered(sk, skb, sack->rate);
3327 /* Initial outgoing SYN's get put onto the write_queue
3328 * just like anything else we transmit. It is not
3329 * true data, and if we misinform our callers that
3330 * this ACK acks real data, we will erroneously exit
3331 * connection startup slow start one packet too
3332 * quickly. This is severely frowned upon behavior.
3334 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
3335 flag |= FLAG_DATA_ACKED;
3337 flag |= FLAG_SYN_ACKED;
3338 tp->retrans_stamp = 0;
3344 tcp_ack_tstamp(sk, skb, ack_skb, prior_snd_una);
3346 next = skb_rb_next(skb);
3347 if (unlikely(skb == tp->retransmit_skb_hint))
3348 tp->retransmit_skb_hint = NULL;
3349 if (unlikely(skb == tp->lost_skb_hint))
3350 tp->lost_skb_hint = NULL;
3351 tcp_highest_sack_replace(sk, skb, next);
3352 tcp_rtx_queue_unlink_and_free(skb, sk);
3356 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
3358 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3359 tp->snd_up = tp->snd_una;
3362 tcp_ack_tstamp(sk, skb, ack_skb, prior_snd_una);
3363 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
3364 flag |= FLAG_SACK_RENEGING;
3367 if (likely(first_ackt) && !(flag & FLAG_RETRANS_DATA_ACKED)) {
3368 seq_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, first_ackt);
3369 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, last_ackt);
3371 if (pkts_acked == 1 && fully_acked && !prior_sacked &&
3372 (tp->snd_una - prior_snd_una) < tp->mss_cache &&
3373 sack->rate->prior_delivered + 1 == tp->delivered &&
3374 !(flag & (FLAG_CA_ALERT | FLAG_SYN_ACKED))) {
3375 /* Conservatively mark a delayed ACK. It's typically
3376 * from a lone runt packet over the round trip to
3377 * a receiver w/o out-of-order or CE events.
3379 flag |= FLAG_ACK_MAYBE_DELAYED;
3382 if (sack->first_sackt) {
3383 sack_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->first_sackt);
3384 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->last_sackt);
3386 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us,
3387 ca_rtt_us, sack->rate);
3389 if (flag & FLAG_ACKED) {
3390 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3391 if (unlikely(icsk->icsk_mtup.probe_size &&
3392 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3393 tcp_mtup_probe_success(sk);
3396 if (tcp_is_reno(tp)) {
3397 tcp_remove_reno_sacks(sk, pkts_acked, ece_ack);
3399 /* If any of the cumulatively ACKed segments was
3400 * retransmitted, non-SACK case cannot confirm that
3401 * progress was due to original transmission due to
3402 * lack of TCPCB_SACKED_ACKED bits even if some of
3403 * the packets may have been never retransmitted.
3405 if (flag & FLAG_RETRANS_DATA_ACKED)
3406 flag &= ~FLAG_ORIG_SACK_ACKED;
3410 /* Non-retransmitted hole got filled? That's reordering */
3411 if (before(reord, prior_fack))
3412 tcp_check_sack_reordering(sk, reord, 0);
3414 delta = prior_sacked - tp->sacked_out;
3415 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3417 } else if (skb && rtt_update && sack_rtt_us >= 0 &&
3418 sack_rtt_us > tcp_stamp_us_delta(tp->tcp_mstamp,
3419 tcp_skb_timestamp_us(skb))) {
3420 /* Do not re-arm RTO if the sack RTT is measured from data sent
3421 * after when the head was last (re)transmitted. Otherwise the
3422 * timeout may continue to extend in loss recovery.
3424 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3427 if (icsk->icsk_ca_ops->pkts_acked) {
3428 struct ack_sample sample = { .pkts_acked = pkts_acked,
3429 .rtt_us = sack->rate->rtt_us };
3431 sample.in_flight = tp->mss_cache *
3432 (tp->delivered - sack->rate->prior_delivered);
3433 icsk->icsk_ca_ops->pkts_acked(sk, &sample);
3436 #if FASTRETRANS_DEBUG > 0
3437 WARN_ON((int)tp->sacked_out < 0);
3438 WARN_ON((int)tp->lost_out < 0);
3439 WARN_ON((int)tp->retrans_out < 0);
3440 if (!tp->packets_out && tcp_is_sack(tp)) {
3441 icsk = inet_csk(sk);
3443 pr_debug("Leak l=%u %d\n",
3444 tp->lost_out, icsk->icsk_ca_state);
3447 if (tp->sacked_out) {
3448 pr_debug("Leak s=%u %d\n",
3449 tp->sacked_out, icsk->icsk_ca_state);
3452 if (tp->retrans_out) {
3453 pr_debug("Leak r=%u %d\n",
3454 tp->retrans_out, icsk->icsk_ca_state);
3455 tp->retrans_out = 0;
3462 static void tcp_ack_probe(struct sock *sk)
3464 struct inet_connection_sock *icsk = inet_csk(sk);
3465 struct sk_buff *head = tcp_send_head(sk);
3466 const struct tcp_sock *tp = tcp_sk(sk);
3468 /* Was it a usable window open? */
3471 if (!after(TCP_SKB_CB(head)->end_seq, tcp_wnd_end(tp))) {
3472 icsk->icsk_backoff = 0;
3473 icsk->icsk_probes_tstamp = 0;
3474 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3475 /* Socket must be waked up by subsequent tcp_data_snd_check().
3476 * This function is not for random using!
3479 unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX);
3481 when = tcp_clamp_probe0_to_user_timeout(sk, when);
3482 tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, when, TCP_RTO_MAX);
3486 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3488 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3489 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3492 /* Decide wheather to run the increase function of congestion control. */
3493 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3495 /* If reordering is high then always grow cwnd whenever data is
3496 * delivered regardless of its ordering. Otherwise stay conservative
3497 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3498 * new SACK or ECE mark may first advance cwnd here and later reduce
3499 * cwnd in tcp_fastretrans_alert() based on more states.
3501 if (tcp_sk(sk)->reordering >
3502 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_reordering))
3503 return flag & FLAG_FORWARD_PROGRESS;
3505 return flag & FLAG_DATA_ACKED;
3508 /* The "ultimate" congestion control function that aims to replace the rigid
3509 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction).
3510 * It's called toward the end of processing an ACK with precise rate
3511 * information. All transmission or retransmission are delayed afterwards.
3513 static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked,
3514 int flag, const struct rate_sample *rs)
3516 const struct inet_connection_sock *icsk = inet_csk(sk);
3518 if (icsk->icsk_ca_ops->cong_control) {
3519 icsk->icsk_ca_ops->cong_control(sk, rs);
3523 if (tcp_in_cwnd_reduction(sk)) {
3524 /* Reduce cwnd if state mandates */
3525 tcp_cwnd_reduction(sk, acked_sacked, rs->losses, flag);
3526 } else if (tcp_may_raise_cwnd(sk, flag)) {
3527 /* Advance cwnd if state allows */
3528 tcp_cong_avoid(sk, ack, acked_sacked);
3530 tcp_update_pacing_rate(sk);
3533 /* Check that window update is acceptable.
3534 * The function assumes that snd_una<=ack<=snd_next.
3536 static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3537 const u32 ack, const u32 ack_seq,
3540 return after(ack, tp->snd_una) ||
3541 after(ack_seq, tp->snd_wl1) ||
3542 (ack_seq == tp->snd_wl1 && (nwin > tp->snd_wnd || !nwin));
3545 /* If we update tp->snd_una, also update tp->bytes_acked */
3546 static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack)
3548 u32 delta = ack - tp->snd_una;
3550 sock_owned_by_me((struct sock *)tp);
3551 tp->bytes_acked += delta;
3555 /* If we update tp->rcv_nxt, also update tp->bytes_received */
3556 static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq)
3558 u32 delta = seq - tp->rcv_nxt;
3560 sock_owned_by_me((struct sock *)tp);
3561 tp->bytes_received += delta;
3562 WRITE_ONCE(tp->rcv_nxt, seq);
3565 /* Update our send window.
3567 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3568 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3570 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3573 struct tcp_sock *tp = tcp_sk(sk);
3575 u32 nwin = ntohs(tcp_hdr(skb)->window);
3577 if (likely(!tcp_hdr(skb)->syn))
3578 nwin <<= tp->rx_opt.snd_wscale;
3580 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3581 flag |= FLAG_WIN_UPDATE;
3582 tcp_update_wl(tp, ack_seq);
3584 if (tp->snd_wnd != nwin) {
3587 /* Note, it is the only place, where
3588 * fast path is recovered for sending TCP.
3591 tcp_fast_path_check(sk);
3593 if (!tcp_write_queue_empty(sk))
3594 tcp_slow_start_after_idle_check(sk);
3596 if (nwin > tp->max_window) {
3597 tp->max_window = nwin;
3598 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3603 tcp_snd_una_update(tp, ack);
3608 static bool __tcp_oow_rate_limited(struct net *net, int mib_idx,
3609 u32 *last_oow_ack_time)
3611 /* Paired with the WRITE_ONCE() in this function. */
3612 u32 val = READ_ONCE(*last_oow_ack_time);
3615 s32 elapsed = (s32)(tcp_jiffies32 - val);
3618 elapsed < READ_ONCE(net->ipv4.sysctl_tcp_invalid_ratelimit)) {
3619 NET_INC_STATS(net, mib_idx);
3620 return true; /* rate-limited: don't send yet! */
3624 /* Paired with the prior READ_ONCE() and with itself,
3625 * as we might be lockless.
3627 WRITE_ONCE(*last_oow_ack_time, tcp_jiffies32);
3629 return false; /* not rate-limited: go ahead, send dupack now! */
3632 /* Return true if we're currently rate-limiting out-of-window ACKs and
3633 * thus shouldn't send a dupack right now. We rate-limit dupacks in
3634 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3635 * attacks that send repeated SYNs or ACKs for the same connection. To
3636 * do this, we do not send a duplicate SYNACK or ACK if the remote
3637 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3639 bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb,
3640 int mib_idx, u32 *last_oow_ack_time)
3642 /* Data packets without SYNs are not likely part of an ACK loop. */
3643 if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) &&
3647 return __tcp_oow_rate_limited(net, mib_idx, last_oow_ack_time);
3650 /* RFC 5961 7 [ACK Throttling] */
3651 static void tcp_send_challenge_ack(struct sock *sk)
3653 struct tcp_sock *tp = tcp_sk(sk);
3654 struct net *net = sock_net(sk);
3655 u32 count, now, ack_limit;
3657 /* First check our per-socket dupack rate limit. */
3658 if (__tcp_oow_rate_limited(net,
3659 LINUX_MIB_TCPACKSKIPPEDCHALLENGE,
3660 &tp->last_oow_ack_time))
3663 ack_limit = READ_ONCE(net->ipv4.sysctl_tcp_challenge_ack_limit);
3664 if (ack_limit == INT_MAX)
3667 /* Then check host-wide RFC 5961 rate limit. */
3669 if (now != READ_ONCE(net->ipv4.tcp_challenge_timestamp)) {
3670 u32 half = (ack_limit + 1) >> 1;
3672 WRITE_ONCE(net->ipv4.tcp_challenge_timestamp, now);
3673 WRITE_ONCE(net->ipv4.tcp_challenge_count,
3674 get_random_u32_inclusive(half, ack_limit + half - 1));
3676 count = READ_ONCE(net->ipv4.tcp_challenge_count);
3678 WRITE_ONCE(net->ipv4.tcp_challenge_count, count - 1);
3680 NET_INC_STATS(net, LINUX_MIB_TCPCHALLENGEACK);
3685 static void tcp_store_ts_recent(struct tcp_sock *tp)
3687 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3688 tp->rx_opt.ts_recent_stamp = ktime_get_seconds();
3691 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3693 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3694 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3695 * extra check below makes sure this can only happen
3696 * for pure ACK frames. -DaveM
3698 * Not only, also it occurs for expired timestamps.
3701 if (tcp_paws_check(&tp->rx_opt, 0))
3702 tcp_store_ts_recent(tp);
3706 /* This routine deals with acks during a TLP episode and ends an episode by
3707 * resetting tlp_high_seq. Ref: TLP algorithm in draft-ietf-tcpm-rack
3709 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3711 struct tcp_sock *tp = tcp_sk(sk);
3713 if (before(ack, tp->tlp_high_seq))
3716 if (!tp->tlp_retrans) {
3717 /* TLP of new data has been acknowledged */
3718 tp->tlp_high_seq = 0;
3719 } else if (flag & FLAG_DSACK_TLP) {
3720 /* This DSACK means original and TLP probe arrived; no loss */
3721 tp->tlp_high_seq = 0;
3722 } else if (after(ack, tp->tlp_high_seq)) {
3723 /* ACK advances: there was a loss, so reduce cwnd. Reset
3724 * tlp_high_seq in tcp_init_cwnd_reduction()
3726 tcp_init_cwnd_reduction(sk);
3727 tcp_set_ca_state(sk, TCP_CA_CWR);
3728 tcp_end_cwnd_reduction(sk);
3729 tcp_try_keep_open(sk);
3730 NET_INC_STATS(sock_net(sk),
3731 LINUX_MIB_TCPLOSSPROBERECOVERY);
3732 } else if (!(flag & (FLAG_SND_UNA_ADVANCED |
3733 FLAG_NOT_DUP | FLAG_DATA_SACKED))) {
3734 /* Pure dupack: original and TLP probe arrived; no loss */
3735 tp->tlp_high_seq = 0;
3739 static inline void tcp_in_ack_event(struct sock *sk, u32 flags)
3741 const struct inet_connection_sock *icsk = inet_csk(sk);
3743 if (icsk->icsk_ca_ops->in_ack_event)
3744 icsk->icsk_ca_ops->in_ack_event(sk, flags);
3747 /* Congestion control has updated the cwnd already. So if we're in
3748 * loss recovery then now we do any new sends (for FRTO) or
3749 * retransmits (for CA_Loss or CA_recovery) that make sense.
3751 static void tcp_xmit_recovery(struct sock *sk, int rexmit)
3753 struct tcp_sock *tp = tcp_sk(sk);
3755 if (rexmit == REXMIT_NONE || sk->sk_state == TCP_SYN_SENT)
3758 if (unlikely(rexmit == REXMIT_NEW)) {
3759 __tcp_push_pending_frames(sk, tcp_current_mss(sk),
3761 if (after(tp->snd_nxt, tp->high_seq))
3765 tcp_xmit_retransmit_queue(sk);
3768 /* Returns the number of packets newly acked or sacked by the current ACK */
3769 static u32 tcp_newly_delivered(struct sock *sk, u32 prior_delivered, int flag)
3771 const struct net *net = sock_net(sk);
3772 struct tcp_sock *tp = tcp_sk(sk);
3775 delivered = tp->delivered - prior_delivered;
3776 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVERED, delivered);
3777 if (flag & FLAG_ECE)
3778 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVEREDCE, delivered);
3783 /* This routine deals with incoming acks, but not outgoing ones. */
3784 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3786 struct inet_connection_sock *icsk = inet_csk(sk);
3787 struct tcp_sock *tp = tcp_sk(sk);
3788 struct tcp_sacktag_state sack_state;
3789 struct rate_sample rs = { .prior_delivered = 0 };
3790 u32 prior_snd_una = tp->snd_una;
3791 bool is_sack_reneg = tp->is_sack_reneg;
3792 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3793 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3795 int prior_packets = tp->packets_out;
3796 u32 delivered = tp->delivered;
3797 u32 lost = tp->lost;
3798 int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */
3801 sack_state.first_sackt = 0;
3802 sack_state.rate = &rs;
3803 sack_state.sack_delivered = 0;
3805 /* We very likely will need to access rtx queue. */
3806 prefetch(sk->tcp_rtx_queue.rb_node);
3808 /* If the ack is older than previous acks
3809 * then we can probably ignore it.
3811 if (before(ack, prior_snd_una)) {
3814 /* do not accept ACK for bytes we never sent. */
3815 max_window = min_t(u64, tp->max_window, tp->bytes_acked);
3816 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3817 if (before(ack, prior_snd_una - max_window)) {
3818 if (!(flag & FLAG_NO_CHALLENGE_ACK))
3819 tcp_send_challenge_ack(sk);
3820 return -SKB_DROP_REASON_TCP_TOO_OLD_ACK;
3825 /* If the ack includes data we haven't sent yet, discard
3826 * this segment (RFC793 Section 3.9).
3828 if (after(ack, tp->snd_nxt))
3829 return -SKB_DROP_REASON_TCP_ACK_UNSENT_DATA;
3831 if (after(ack, prior_snd_una)) {
3832 flag |= FLAG_SND_UNA_ADVANCED;
3833 icsk->icsk_retransmits = 0;
3835 #if IS_ENABLED(CONFIG_TLS_DEVICE)
3836 if (static_branch_unlikely(&clean_acked_data_enabled.key))
3837 if (icsk->icsk_clean_acked)
3838 icsk->icsk_clean_acked(sk, ack);
3842 prior_fack = tcp_is_sack(tp) ? tcp_highest_sack_seq(tp) : tp->snd_una;
3843 rs.prior_in_flight = tcp_packets_in_flight(tp);
3845 /* ts_recent update must be made after we are sure that the packet
3848 if (flag & FLAG_UPDATE_TS_RECENT)
3849 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3851 if ((flag & (FLAG_SLOWPATH | FLAG_SND_UNA_ADVANCED)) ==
3852 FLAG_SND_UNA_ADVANCED) {
3853 /* Window is constant, pure forward advance.
3854 * No more checks are required.
3855 * Note, we use the fact that SND.UNA>=SND.WL2.
3857 tcp_update_wl(tp, ack_seq);
3858 tcp_snd_una_update(tp, ack);
3859 flag |= FLAG_WIN_UPDATE;
3861 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
3863 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPACKS);
3865 u32 ack_ev_flags = CA_ACK_SLOWPATH;
3867 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3870 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3872 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3874 if (TCP_SKB_CB(skb)->sacked)
3875 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3878 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) {
3880 ack_ev_flags |= CA_ACK_ECE;
3883 if (sack_state.sack_delivered)
3884 tcp_count_delivered(tp, sack_state.sack_delivered,
3887 if (flag & FLAG_WIN_UPDATE)
3888 ack_ev_flags |= CA_ACK_WIN_UPDATE;
3890 tcp_in_ack_event(sk, ack_ev_flags);
3893 /* This is a deviation from RFC3168 since it states that:
3894 * "When the TCP data sender is ready to set the CWR bit after reducing
3895 * the congestion window, it SHOULD set the CWR bit only on the first
3896 * new data packet that it transmits."
3897 * We accept CWR on pure ACKs to be more robust
3898 * with widely-deployed TCP implementations that do this.
3900 tcp_ecn_accept_cwr(sk, skb);
3902 /* We passed data and got it acked, remove any soft error
3903 * log. Something worked...
3905 WRITE_ONCE(sk->sk_err_soft, 0);
3906 icsk->icsk_probes_out = 0;
3907 tp->rcv_tstamp = tcp_jiffies32;
3911 /* See if we can take anything off of the retransmit queue. */
3912 flag |= tcp_clean_rtx_queue(sk, skb, prior_fack, prior_snd_una,
3913 &sack_state, flag & FLAG_ECE);
3915 tcp_rack_update_reo_wnd(sk, &rs);
3917 if (tp->tlp_high_seq)
3918 tcp_process_tlp_ack(sk, ack, flag);
3920 if (tcp_ack_is_dubious(sk, flag)) {
3921 if (!(flag & (FLAG_SND_UNA_ADVANCED |
3922 FLAG_NOT_DUP | FLAG_DSACKING_ACK))) {
3924 /* Consider if pure acks were aggregated in tcp_add_backlog() */
3925 if (!(flag & FLAG_DATA))
3926 num_dupack = max_t(u16, 1, skb_shinfo(skb)->gso_segs);
3928 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
3932 /* If needed, reset TLP/RTO timer when RACK doesn't set. */
3933 if (flag & FLAG_SET_XMIT_TIMER)
3934 tcp_set_xmit_timer(sk);
3936 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
3939 delivered = tcp_newly_delivered(sk, delivered, flag);
3940 lost = tp->lost - lost; /* freshly marked lost */
3941 rs.is_ack_delayed = !!(flag & FLAG_ACK_MAYBE_DELAYED);
3942 tcp_rate_gen(sk, delivered, lost, is_sack_reneg, sack_state.rate);
3943 tcp_cong_control(sk, ack, delivered, flag, sack_state.rate);
3944 tcp_xmit_recovery(sk, rexmit);
3948 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3949 if (flag & FLAG_DSACKING_ACK) {
3950 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
3952 tcp_newly_delivered(sk, delivered, flag);
3954 /* If this ack opens up a zero window, clear backoff. It was
3955 * being used to time the probes, and is probably far higher than
3956 * it needs to be for normal retransmission.
3960 if (tp->tlp_high_seq)
3961 tcp_process_tlp_ack(sk, ack, flag);
3965 /* If data was SACKed, tag it and see if we should send more data.
3966 * If data was DSACKed, see if we can undo a cwnd reduction.
3968 if (TCP_SKB_CB(skb)->sacked) {
3969 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3971 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
3973 tcp_newly_delivered(sk, delivered, flag);
3974 tcp_xmit_recovery(sk, rexmit);
3980 static void tcp_parse_fastopen_option(int len, const unsigned char *cookie,
3981 bool syn, struct tcp_fastopen_cookie *foc,
3984 /* Valid only in SYN or SYN-ACK with an even length. */
3985 if (!foc || !syn || len < 0 || (len & 1))
3988 if (len >= TCP_FASTOPEN_COOKIE_MIN &&
3989 len <= TCP_FASTOPEN_COOKIE_MAX)
3990 memcpy(foc->val, cookie, len);
3997 static bool smc_parse_options(const struct tcphdr *th,
3998 struct tcp_options_received *opt_rx,
3999 const unsigned char *ptr,
4002 #if IS_ENABLED(CONFIG_SMC)
4003 if (static_branch_unlikely(&tcp_have_smc)) {
4004 if (th->syn && !(opsize & 1) &&
4005 opsize >= TCPOLEN_EXP_SMC_BASE &&
4006 get_unaligned_be32(ptr) == TCPOPT_SMC_MAGIC) {
4015 /* Try to parse the MSS option from the TCP header. Return 0 on failure, clamped
4018 u16 tcp_parse_mss_option(const struct tcphdr *th, u16 user_mss)
4020 const unsigned char *ptr = (const unsigned char *)(th + 1);
4021 int length = (th->doff * 4) - sizeof(struct tcphdr);
4024 while (length > 0) {
4025 int opcode = *ptr++;
4031 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
4038 if (opsize < 2) /* "silly options" */
4040 if (opsize > length)
4041 return mss; /* fail on partial options */
4042 if (opcode == TCPOPT_MSS && opsize == TCPOLEN_MSS) {
4043 u16 in_mss = get_unaligned_be16(ptr);
4046 if (user_mss && user_mss < in_mss)
4057 EXPORT_SYMBOL_GPL(tcp_parse_mss_option);
4059 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
4060 * But, this can also be called on packets in the established flow when
4061 * the fast version below fails.
4063 void tcp_parse_options(const struct net *net,
4064 const struct sk_buff *skb,
4065 struct tcp_options_received *opt_rx, int estab,
4066 struct tcp_fastopen_cookie *foc)
4068 const unsigned char *ptr;
4069 const struct tcphdr *th = tcp_hdr(skb);
4070 int length = (th->doff * 4) - sizeof(struct tcphdr);
4072 ptr = (const unsigned char *)(th + 1);
4073 opt_rx->saw_tstamp = 0;
4074 opt_rx->saw_unknown = 0;
4076 while (length > 0) {
4077 int opcode = *ptr++;
4083 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
4090 if (opsize < 2) /* "silly options" */
4092 if (opsize > length)
4093 return; /* don't parse partial options */
4096 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
4097 u16 in_mss = get_unaligned_be16(ptr);
4099 if (opt_rx->user_mss &&
4100 opt_rx->user_mss < in_mss)
4101 in_mss = opt_rx->user_mss;
4102 opt_rx->mss_clamp = in_mss;
4107 if (opsize == TCPOLEN_WINDOW && th->syn &&
4108 !estab && READ_ONCE(net->ipv4.sysctl_tcp_window_scaling)) {
4109 __u8 snd_wscale = *(__u8 *)ptr;
4110 opt_rx->wscale_ok = 1;
4111 if (snd_wscale > TCP_MAX_WSCALE) {
4112 net_info_ratelimited("%s: Illegal window scaling value %d > %u received\n",
4116 snd_wscale = TCP_MAX_WSCALE;
4118 opt_rx->snd_wscale = snd_wscale;
4121 case TCPOPT_TIMESTAMP:
4122 if ((opsize == TCPOLEN_TIMESTAMP) &&
4123 ((estab && opt_rx->tstamp_ok) ||
4124 (!estab && READ_ONCE(net->ipv4.sysctl_tcp_timestamps)))) {
4125 opt_rx->saw_tstamp = 1;
4126 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
4127 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
4130 case TCPOPT_SACK_PERM:
4131 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
4132 !estab && READ_ONCE(net->ipv4.sysctl_tcp_sack)) {
4133 opt_rx->sack_ok = TCP_SACK_SEEN;
4134 tcp_sack_reset(opt_rx);
4139 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
4140 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
4142 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
4145 #ifdef CONFIG_TCP_MD5SIG
4147 /* The MD5 Hash has already been
4148 * checked (see tcp_v{4,6}_rcv()).
4152 case TCPOPT_FASTOPEN:
4153 tcp_parse_fastopen_option(
4154 opsize - TCPOLEN_FASTOPEN_BASE,
4155 ptr, th->syn, foc, false);
4159 /* Fast Open option shares code 254 using a
4160 * 16 bits magic number.
4162 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE &&
4163 get_unaligned_be16(ptr) ==
4164 TCPOPT_FASTOPEN_MAGIC) {
4165 tcp_parse_fastopen_option(opsize -
4166 TCPOLEN_EXP_FASTOPEN_BASE,
4167 ptr + 2, th->syn, foc, true);
4171 if (smc_parse_options(th, opt_rx, ptr, opsize))
4174 opt_rx->saw_unknown = 1;
4178 opt_rx->saw_unknown = 1;
4185 EXPORT_SYMBOL(tcp_parse_options);
4187 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
4189 const __be32 *ptr = (const __be32 *)(th + 1);
4191 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
4192 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
4193 tp->rx_opt.saw_tstamp = 1;
4195 tp->rx_opt.rcv_tsval = ntohl(*ptr);
4198 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
4200 tp->rx_opt.rcv_tsecr = 0;
4206 /* Fast parse options. This hopes to only see timestamps.
4207 * If it is wrong it falls back on tcp_parse_options().
4209 static bool tcp_fast_parse_options(const struct net *net,
4210 const struct sk_buff *skb,
4211 const struct tcphdr *th, struct tcp_sock *tp)
4213 /* In the spirit of fast parsing, compare doff directly to constant
4214 * values. Because equality is used, short doff can be ignored here.
4216 if (th->doff == (sizeof(*th) / 4)) {
4217 tp->rx_opt.saw_tstamp = 0;
4219 } else if (tp->rx_opt.tstamp_ok &&
4220 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
4221 if (tcp_parse_aligned_timestamp(tp, th))
4225 tcp_parse_options(net, skb, &tp->rx_opt, 1, NULL);
4226 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
4227 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
4232 #ifdef CONFIG_TCP_MD5SIG
4234 * Parse MD5 Signature option
4236 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
4238 int length = (th->doff << 2) - sizeof(*th);
4239 const u8 *ptr = (const u8 *)(th + 1);
4241 /* If not enough data remaining, we can short cut */
4242 while (length >= TCPOLEN_MD5SIG) {
4243 int opcode = *ptr++;
4254 if (opsize < 2 || opsize > length)
4256 if (opcode == TCPOPT_MD5SIG)
4257 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
4264 EXPORT_SYMBOL(tcp_parse_md5sig_option);
4267 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
4269 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
4270 * it can pass through stack. So, the following predicate verifies that
4271 * this segment is not used for anything but congestion avoidance or
4272 * fast retransmit. Moreover, we even are able to eliminate most of such
4273 * second order effects, if we apply some small "replay" window (~RTO)
4274 * to timestamp space.
4276 * All these measures still do not guarantee that we reject wrapped ACKs
4277 * on networks with high bandwidth, when sequence space is recycled fastly,
4278 * but it guarantees that such events will be very rare and do not affect
4279 * connection seriously. This doesn't look nice, but alas, PAWS is really
4282 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
4283 * states that events when retransmit arrives after original data are rare.
4284 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
4285 * the biggest problem on large power networks even with minor reordering.
4286 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
4287 * up to bandwidth of 18Gigabit/sec. 8) ]
4290 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
4292 const struct tcp_sock *tp = tcp_sk(sk);
4293 const struct tcphdr *th = tcp_hdr(skb);
4294 u32 seq = TCP_SKB_CB(skb)->seq;
4295 u32 ack = TCP_SKB_CB(skb)->ack_seq;
4297 return (/* 1. Pure ACK with correct sequence number. */
4298 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
4300 /* 2. ... and duplicate ACK. */
4301 ack == tp->snd_una &&
4303 /* 3. ... and does not update window. */
4304 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
4306 /* 4. ... and sits in replay window. */
4307 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
4310 static inline bool tcp_paws_discard(const struct sock *sk,
4311 const struct sk_buff *skb)
4313 const struct tcp_sock *tp = tcp_sk(sk);
4315 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
4316 !tcp_disordered_ack(sk, skb);
4319 /* Check segment sequence number for validity.
4321 * Segment controls are considered valid, if the segment
4322 * fits to the window after truncation to the window. Acceptability
4323 * of data (and SYN, FIN, of course) is checked separately.
4324 * See tcp_data_queue(), for example.
4326 * Also, controls (RST is main one) are accepted using RCV.WUP instead
4327 * of RCV.NXT. Peer still did not advance his SND.UNA when we
4328 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4329 * (borrowed from freebsd)
4332 static enum skb_drop_reason tcp_sequence(const struct tcp_sock *tp,
4333 u32 seq, u32 end_seq)
4335 if (before(end_seq, tp->rcv_wup))
4336 return SKB_DROP_REASON_TCP_OLD_SEQUENCE;
4338 if (after(seq, tp->rcv_nxt + tcp_receive_window(tp)))
4339 return SKB_DROP_REASON_TCP_INVALID_SEQUENCE;
4341 return SKB_NOT_DROPPED_YET;
4344 /* When we get a reset we do this. */
4345 void tcp_reset(struct sock *sk, struct sk_buff *skb)
4347 trace_tcp_receive_reset(sk);
4349 /* mptcp can't tell us to ignore reset pkts,
4350 * so just ignore the return value of mptcp_incoming_options().
4352 if (sk_is_mptcp(sk))
4353 mptcp_incoming_options(sk, skb);
4355 /* We want the right error as BSD sees it (and indeed as we do). */
4356 switch (sk->sk_state) {
4358 WRITE_ONCE(sk->sk_err, ECONNREFUSED);
4360 case TCP_CLOSE_WAIT:
4361 WRITE_ONCE(sk->sk_err, EPIPE);
4366 WRITE_ONCE(sk->sk_err, ECONNRESET);
4368 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4371 tcp_write_queue_purge(sk);
4374 if (!sock_flag(sk, SOCK_DEAD))
4375 sk_error_report(sk);
4379 * Process the FIN bit. This now behaves as it is supposed to work
4380 * and the FIN takes effect when it is validly part of sequence
4381 * space. Not before when we get holes.
4383 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4384 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4387 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4388 * close and we go into CLOSING (and later onto TIME-WAIT)
4390 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4392 void tcp_fin(struct sock *sk)
4394 struct tcp_sock *tp = tcp_sk(sk);
4396 inet_csk_schedule_ack(sk);
4398 WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | RCV_SHUTDOWN);
4399 sock_set_flag(sk, SOCK_DONE);
4401 switch (sk->sk_state) {
4403 case TCP_ESTABLISHED:
4404 /* Move to CLOSE_WAIT */
4405 tcp_set_state(sk, TCP_CLOSE_WAIT);
4406 inet_csk_enter_pingpong_mode(sk);
4409 case TCP_CLOSE_WAIT:
4411 /* Received a retransmission of the FIN, do
4416 /* RFC793: Remain in the LAST-ACK state. */
4420 /* This case occurs when a simultaneous close
4421 * happens, we must ack the received FIN and
4422 * enter the CLOSING state.
4425 tcp_set_state(sk, TCP_CLOSING);
4428 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4430 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4433 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4434 * cases we should never reach this piece of code.
4436 pr_err("%s: Impossible, sk->sk_state=%d\n",
4437 __func__, sk->sk_state);
4441 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4442 * Probably, we should reset in this case. For now drop them.
4444 skb_rbtree_purge(&tp->out_of_order_queue);
4445 if (tcp_is_sack(tp))
4446 tcp_sack_reset(&tp->rx_opt);
4448 if (!sock_flag(sk, SOCK_DEAD)) {
4449 sk->sk_state_change(sk);
4451 /* Do not send POLL_HUP for half duplex close. */
4452 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4453 sk->sk_state == TCP_CLOSE)
4454 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4456 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4460 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4463 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4464 if (before(seq, sp->start_seq))
4465 sp->start_seq = seq;
4466 if (after(end_seq, sp->end_seq))
4467 sp->end_seq = end_seq;
4473 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4475 struct tcp_sock *tp = tcp_sk(sk);
4477 if (tcp_is_sack(tp) && READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_dsack)) {
4480 if (before(seq, tp->rcv_nxt))
4481 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4483 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4485 NET_INC_STATS(sock_net(sk), mib_idx);
4487 tp->rx_opt.dsack = 1;
4488 tp->duplicate_sack[0].start_seq = seq;
4489 tp->duplicate_sack[0].end_seq = end_seq;
4493 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4495 struct tcp_sock *tp = tcp_sk(sk);
4497 if (!tp->rx_opt.dsack)
4498 tcp_dsack_set(sk, seq, end_seq);
4500 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4503 static void tcp_rcv_spurious_retrans(struct sock *sk, const struct sk_buff *skb)
4505 /* When the ACK path fails or drops most ACKs, the sender would
4506 * timeout and spuriously retransmit the same segment repeatedly.
4507 * The receiver remembers and reflects via DSACKs. Leverage the
4508 * DSACK state and change the txhash to re-route speculatively.
4510 if (TCP_SKB_CB(skb)->seq == tcp_sk(sk)->duplicate_sack[0].start_seq &&
4511 sk_rethink_txhash(sk))
4512 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDUPLICATEDATAREHASH);
4515 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4517 struct tcp_sock *tp = tcp_sk(sk);
4519 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4520 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4521 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4522 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
4524 if (tcp_is_sack(tp) && READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_dsack)) {
4525 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4527 tcp_rcv_spurious_retrans(sk, skb);
4528 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4529 end_seq = tp->rcv_nxt;
4530 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4537 /* These routines update the SACK block as out-of-order packets arrive or
4538 * in-order packets close up the sequence space.
4540 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4543 struct tcp_sack_block *sp = &tp->selective_acks[0];
4544 struct tcp_sack_block *swalk = sp + 1;
4546 /* See if the recent change to the first SACK eats into
4547 * or hits the sequence space of other SACK blocks, if so coalesce.
4549 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4550 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4553 /* Zap SWALK, by moving every further SACK up by one slot.
4554 * Decrease num_sacks.
4556 tp->rx_opt.num_sacks--;
4557 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4566 void tcp_sack_compress_send_ack(struct sock *sk)
4568 struct tcp_sock *tp = tcp_sk(sk);
4570 if (!tp->compressed_ack)
4573 if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1)
4576 /* Since we have to send one ack finally,
4577 * substract one from tp->compressed_ack to keep
4578 * LINUX_MIB_TCPACKCOMPRESSED accurate.
4580 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED,
4581 tp->compressed_ack - 1);
4583 tp->compressed_ack = 0;
4587 /* Reasonable amount of sack blocks included in TCP SACK option
4588 * The max is 4, but this becomes 3 if TCP timestamps are there.
4589 * Given that SACK packets might be lost, be conservative and use 2.
4591 #define TCP_SACK_BLOCKS_EXPECTED 2
4593 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4595 struct tcp_sock *tp = tcp_sk(sk);
4596 struct tcp_sack_block *sp = &tp->selective_acks[0];
4597 int cur_sacks = tp->rx_opt.num_sacks;
4603 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4604 if (tcp_sack_extend(sp, seq, end_seq)) {
4605 if (this_sack >= TCP_SACK_BLOCKS_EXPECTED)
4606 tcp_sack_compress_send_ack(sk);
4607 /* Rotate this_sack to the first one. */
4608 for (; this_sack > 0; this_sack--, sp--)
4609 swap(*sp, *(sp - 1));
4611 tcp_sack_maybe_coalesce(tp);
4616 if (this_sack >= TCP_SACK_BLOCKS_EXPECTED)
4617 tcp_sack_compress_send_ack(sk);
4619 /* Could not find an adjacent existing SACK, build a new one,
4620 * put it at the front, and shift everyone else down. We
4621 * always know there is at least one SACK present already here.
4623 * If the sack array is full, forget about the last one.
4625 if (this_sack >= TCP_NUM_SACKS) {
4627 tp->rx_opt.num_sacks--;
4630 for (; this_sack > 0; this_sack--, sp--)
4634 /* Build the new head SACK, and we're done. */
4635 sp->start_seq = seq;
4636 sp->end_seq = end_seq;
4637 tp->rx_opt.num_sacks++;
4640 /* RCV.NXT advances, some SACKs should be eaten. */
4642 static void tcp_sack_remove(struct tcp_sock *tp)
4644 struct tcp_sack_block *sp = &tp->selective_acks[0];
4645 int num_sacks = tp->rx_opt.num_sacks;
4648 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4649 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4650 tp->rx_opt.num_sacks = 0;
4654 for (this_sack = 0; this_sack < num_sacks;) {
4655 /* Check if the start of the sack is covered by RCV.NXT. */
4656 if (!before(tp->rcv_nxt, sp->start_seq)) {
4659 /* RCV.NXT must cover all the block! */
4660 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4662 /* Zap this SACK, by moving forward any other SACKS. */
4663 for (i = this_sack+1; i < num_sacks; i++)
4664 tp->selective_acks[i-1] = tp->selective_acks[i];
4671 tp->rx_opt.num_sacks = num_sacks;
4675 * tcp_try_coalesce - try to merge skb to prior one
4678 * @from: buffer to add in queue
4679 * @fragstolen: pointer to boolean
4681 * Before queueing skb @from after @to, try to merge them
4682 * to reduce overall memory use and queue lengths, if cost is small.
4683 * Packets in ofo or receive queues can stay a long time.
4684 * Better try to coalesce them right now to avoid future collapses.
4685 * Returns true if caller should free @from instead of queueing it
4687 static bool tcp_try_coalesce(struct sock *sk,
4689 struct sk_buff *from,
4694 *fragstolen = false;
4696 /* Its possible this segment overlaps with prior segment in queue */
4697 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4700 if (!mptcp_skb_can_collapse(to, from))
4703 #ifdef CONFIG_TLS_DEVICE
4704 if (from->decrypted != to->decrypted)
4708 if (!skb_try_coalesce(to, from, fragstolen, &delta))
4711 atomic_add(delta, &sk->sk_rmem_alloc);
4712 sk_mem_charge(sk, delta);
4713 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4714 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4715 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4716 TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
4718 if (TCP_SKB_CB(from)->has_rxtstamp) {
4719 TCP_SKB_CB(to)->has_rxtstamp = true;
4720 to->tstamp = from->tstamp;
4721 skb_hwtstamps(to)->hwtstamp = skb_hwtstamps(from)->hwtstamp;
4727 static bool tcp_ooo_try_coalesce(struct sock *sk,
4729 struct sk_buff *from,
4732 bool res = tcp_try_coalesce(sk, to, from, fragstolen);
4734 /* In case tcp_drop_reason() is called later, update to->gso_segs */
4736 u32 gso_segs = max_t(u16, 1, skb_shinfo(to)->gso_segs) +
4737 max_t(u16, 1, skb_shinfo(from)->gso_segs);
4739 skb_shinfo(to)->gso_segs = min_t(u32, gso_segs, 0xFFFF);
4744 static void tcp_drop_reason(struct sock *sk, struct sk_buff *skb,
4745 enum skb_drop_reason reason)
4747 sk_drops_add(sk, skb);
4748 kfree_skb_reason(skb, reason);
4751 /* This one checks to see if we can put data from the
4752 * out_of_order queue into the receive_queue.
4754 static void tcp_ofo_queue(struct sock *sk)
4756 struct tcp_sock *tp = tcp_sk(sk);
4757 __u32 dsack_high = tp->rcv_nxt;
4758 bool fin, fragstolen, eaten;
4759 struct sk_buff *skb, *tail;
4762 p = rb_first(&tp->out_of_order_queue);
4765 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4768 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4769 __u32 dsack = dsack_high;
4770 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4771 dsack_high = TCP_SKB_CB(skb)->end_seq;
4772 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4775 rb_erase(&skb->rbnode, &tp->out_of_order_queue);
4777 if (unlikely(!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))) {
4778 tcp_drop_reason(sk, skb, SKB_DROP_REASON_TCP_OFO_DROP);
4782 tail = skb_peek_tail(&sk->sk_receive_queue);
4783 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen);
4784 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4785 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
4787 __skb_queue_tail(&sk->sk_receive_queue, skb);
4789 kfree_skb_partial(skb, fragstolen);
4791 if (unlikely(fin)) {
4793 /* tcp_fin() purges tp->out_of_order_queue,
4794 * so we must end this loop right now.
4801 static bool tcp_prune_ofo_queue(struct sock *sk, const struct sk_buff *in_skb);
4802 static int tcp_prune_queue(struct sock *sk, const struct sk_buff *in_skb);
4804 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4807 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4808 !sk_rmem_schedule(sk, skb, size)) {
4810 if (tcp_prune_queue(sk, skb) < 0)
4813 while (!sk_rmem_schedule(sk, skb, size)) {
4814 if (!tcp_prune_ofo_queue(sk, skb))
4821 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4823 struct tcp_sock *tp = tcp_sk(sk);
4824 struct rb_node **p, *parent;
4825 struct sk_buff *skb1;
4829 tcp_ecn_check_ce(sk, skb);
4831 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4832 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFODROP);
4833 sk->sk_data_ready(sk);
4834 tcp_drop_reason(sk, skb, SKB_DROP_REASON_PROTO_MEM);
4838 /* Disable header prediction. */
4840 inet_csk_schedule_ack(sk);
4842 tp->rcv_ooopack += max_t(u16, 1, skb_shinfo(skb)->gso_segs);
4843 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4844 seq = TCP_SKB_CB(skb)->seq;
4845 end_seq = TCP_SKB_CB(skb)->end_seq;
4847 p = &tp->out_of_order_queue.rb_node;
4848 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4849 /* Initial out of order segment, build 1 SACK. */
4850 if (tcp_is_sack(tp)) {
4851 tp->rx_opt.num_sacks = 1;
4852 tp->selective_acks[0].start_seq = seq;
4853 tp->selective_acks[0].end_seq = end_seq;
4855 rb_link_node(&skb->rbnode, NULL, p);
4856 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4857 tp->ooo_last_skb = skb;
4861 /* In the typical case, we are adding an skb to the end of the list.
4862 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
4864 if (tcp_ooo_try_coalesce(sk, tp->ooo_last_skb,
4865 skb, &fragstolen)) {
4867 /* For non sack flows, do not grow window to force DUPACK
4868 * and trigger fast retransmit.
4870 if (tcp_is_sack(tp))
4871 tcp_grow_window(sk, skb, true);
4872 kfree_skb_partial(skb, fragstolen);
4876 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
4877 if (!before(seq, TCP_SKB_CB(tp->ooo_last_skb)->end_seq)) {
4878 parent = &tp->ooo_last_skb->rbnode;
4879 p = &parent->rb_right;
4883 /* Find place to insert this segment. Handle overlaps on the way. */
4887 skb1 = rb_to_skb(parent);
4888 if (before(seq, TCP_SKB_CB(skb1)->seq)) {
4889 p = &parent->rb_left;
4892 if (before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4893 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4894 /* All the bits are present. Drop. */
4895 NET_INC_STATS(sock_net(sk),
4896 LINUX_MIB_TCPOFOMERGE);
4897 tcp_drop_reason(sk, skb,
4898 SKB_DROP_REASON_TCP_OFOMERGE);
4900 tcp_dsack_set(sk, seq, end_seq);
4903 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4904 /* Partial overlap. */
4905 tcp_dsack_set(sk, seq, TCP_SKB_CB(skb1)->end_seq);
4907 /* skb's seq == skb1's seq and skb covers skb1.
4908 * Replace skb1 with skb.
4910 rb_replace_node(&skb1->rbnode, &skb->rbnode,
4911 &tp->out_of_order_queue);
4912 tcp_dsack_extend(sk,
4913 TCP_SKB_CB(skb1)->seq,
4914 TCP_SKB_CB(skb1)->end_seq);
4915 NET_INC_STATS(sock_net(sk),
4916 LINUX_MIB_TCPOFOMERGE);
4917 tcp_drop_reason(sk, skb1,
4918 SKB_DROP_REASON_TCP_OFOMERGE);
4921 } else if (tcp_ooo_try_coalesce(sk, skb1,
4922 skb, &fragstolen)) {
4925 p = &parent->rb_right;
4928 /* Insert segment into RB tree. */
4929 rb_link_node(&skb->rbnode, parent, p);
4930 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4933 /* Remove other segments covered by skb. */
4934 while ((skb1 = skb_rb_next(skb)) != NULL) {
4935 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4937 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4938 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4942 rb_erase(&skb1->rbnode, &tp->out_of_order_queue);
4943 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4944 TCP_SKB_CB(skb1)->end_seq);
4945 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4946 tcp_drop_reason(sk, skb1, SKB_DROP_REASON_TCP_OFOMERGE);
4948 /* If there is no skb after us, we are the last_skb ! */
4950 tp->ooo_last_skb = skb;
4953 if (tcp_is_sack(tp))
4954 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4957 /* For non sack flows, do not grow window to force DUPACK
4958 * and trigger fast retransmit.
4960 if (tcp_is_sack(tp))
4961 tcp_grow_window(sk, skb, false);
4963 skb_set_owner_r(skb, sk);
4967 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb,
4971 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4974 tcp_try_coalesce(sk, tail,
4975 skb, fragstolen)) ? 1 : 0;
4976 tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq);
4978 __skb_queue_tail(&sk->sk_receive_queue, skb);
4979 skb_set_owner_r(skb, sk);
4984 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4986 struct sk_buff *skb;
4994 if (size > PAGE_SIZE) {
4995 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS);
4997 data_len = npages << PAGE_SHIFT;
4998 size = data_len + (size & ~PAGE_MASK);
5000 skb = alloc_skb_with_frags(size - data_len, data_len,
5001 PAGE_ALLOC_COSTLY_ORDER,
5002 &err, sk->sk_allocation);
5006 skb_put(skb, size - data_len);
5007 skb->data_len = data_len;
5010 if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) {
5011 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP);
5015 err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size);
5019 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
5020 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
5021 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
5023 if (tcp_queue_rcv(sk, skb, &fragstolen)) {
5024 WARN_ON_ONCE(fragstolen); /* should not happen */
5036 void tcp_data_ready(struct sock *sk)
5038 if (tcp_epollin_ready(sk, sk->sk_rcvlowat) || sock_flag(sk, SOCK_DONE))
5039 sk->sk_data_ready(sk);
5042 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
5044 struct tcp_sock *tp = tcp_sk(sk);
5045 enum skb_drop_reason reason;
5049 /* If a subflow has been reset, the packet should not continue
5050 * to be processed, drop the packet.
5052 if (sk_is_mptcp(sk) && !mptcp_incoming_options(sk, skb)) {
5057 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
5062 __skb_pull(skb, tcp_hdr(skb)->doff * 4);
5064 reason = SKB_DROP_REASON_NOT_SPECIFIED;
5065 tp->rx_opt.dsack = 0;
5067 /* Queue data for delivery to the user.
5068 * Packets in sequence go to the receive queue.
5069 * Out of sequence packets to the out_of_order_queue.
5071 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
5072 if (tcp_receive_window(tp) == 0) {
5073 reason = SKB_DROP_REASON_TCP_ZEROWINDOW;
5074 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP);
5078 /* Ok. In sequence. In window. */
5080 if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) {
5081 /* TODO: maybe ratelimit these WIN 0 ACK ? */
5082 inet_csk(sk)->icsk_ack.pending |=
5083 (ICSK_ACK_NOMEM | ICSK_ACK_NOW);
5084 inet_csk_schedule_ack(sk);
5085 sk->sk_data_ready(sk);
5087 if (skb_queue_len(&sk->sk_receive_queue)) {
5088 reason = SKB_DROP_REASON_PROTO_MEM;
5089 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP);
5092 sk_forced_mem_schedule(sk, skb->truesize);
5095 eaten = tcp_queue_rcv(sk, skb, &fragstolen);
5097 tcp_event_data_recv(sk, skb);
5098 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
5101 if (!RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
5104 /* RFC5681. 4.2. SHOULD send immediate ACK, when
5105 * gap in queue is filled.
5107 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
5108 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW;
5111 if (tp->rx_opt.num_sacks)
5112 tcp_sack_remove(tp);
5114 tcp_fast_path_check(sk);
5117 kfree_skb_partial(skb, fragstolen);
5118 if (!sock_flag(sk, SOCK_DEAD))
5123 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
5124 tcp_rcv_spurious_retrans(sk, skb);
5125 /* A retransmit, 2nd most common case. Force an immediate ack. */
5126 reason = SKB_DROP_REASON_TCP_OLD_DATA;
5127 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
5128 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
5131 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
5132 inet_csk_schedule_ack(sk);
5134 tcp_drop_reason(sk, skb, reason);
5138 /* Out of window. F.e. zero window probe. */
5139 if (!before(TCP_SKB_CB(skb)->seq,
5140 tp->rcv_nxt + tcp_receive_window(tp))) {
5141 reason = SKB_DROP_REASON_TCP_OVERWINDOW;
5145 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
5146 /* Partial packet, seq < rcv_next < end_seq */
5147 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
5149 /* If window is closed, drop tail of packet. But after
5150 * remembering D-SACK for its head made in previous line.
5152 if (!tcp_receive_window(tp)) {
5153 reason = SKB_DROP_REASON_TCP_ZEROWINDOW;
5154 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP);
5160 tcp_data_queue_ofo(sk, skb);
5163 static struct sk_buff *tcp_skb_next(struct sk_buff *skb, struct sk_buff_head *list)
5166 return !skb_queue_is_last(list, skb) ? skb->next : NULL;
5168 return skb_rb_next(skb);
5171 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
5172 struct sk_buff_head *list,
5173 struct rb_root *root)
5175 struct sk_buff *next = tcp_skb_next(skb, list);
5178 __skb_unlink(skb, list);
5180 rb_erase(&skb->rbnode, root);
5183 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
5188 /* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */
5189 void tcp_rbtree_insert(struct rb_root *root, struct sk_buff *skb)
5191 struct rb_node **p = &root->rb_node;
5192 struct rb_node *parent = NULL;
5193 struct sk_buff *skb1;
5197 skb1 = rb_to_skb(parent);
5198 if (before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb1)->seq))
5199 p = &parent->rb_left;
5201 p = &parent->rb_right;
5203 rb_link_node(&skb->rbnode, parent, p);
5204 rb_insert_color(&skb->rbnode, root);
5207 /* Collapse contiguous sequence of skbs head..tail with
5208 * sequence numbers start..end.
5210 * If tail is NULL, this means until the end of the queue.
5212 * Segments with FIN/SYN are not collapsed (only because this
5216 tcp_collapse(struct sock *sk, struct sk_buff_head *list, struct rb_root *root,
5217 struct sk_buff *head, struct sk_buff *tail, u32 start, u32 end)
5219 struct sk_buff *skb = head, *n;
5220 struct sk_buff_head tmp;
5223 /* First, check that queue is collapsible and find
5224 * the point where collapsing can be useful.
5227 for (end_of_skbs = true; skb != NULL && skb != tail; skb = n) {
5228 n = tcp_skb_next(skb, list);
5230 /* No new bits? It is possible on ofo queue. */
5231 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
5232 skb = tcp_collapse_one(sk, skb, list, root);
5238 /* The first skb to collapse is:
5240 * - bloated or contains data before "start" or
5241 * overlaps to the next one and mptcp allow collapsing.
5243 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
5244 (tcp_win_from_space(sk, skb->truesize) > skb->len ||
5245 before(TCP_SKB_CB(skb)->seq, start))) {
5246 end_of_skbs = false;
5250 if (n && n != tail && mptcp_skb_can_collapse(skb, n) &&
5251 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(n)->seq) {
5252 end_of_skbs = false;
5256 /* Decided to skip this, advance start seq. */
5257 start = TCP_SKB_CB(skb)->end_seq;
5260 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
5263 __skb_queue_head_init(&tmp);
5265 while (before(start, end)) {
5266 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
5267 struct sk_buff *nskb;
5269 nskb = alloc_skb(copy, GFP_ATOMIC);
5273 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
5274 #ifdef CONFIG_TLS_DEVICE
5275 nskb->decrypted = skb->decrypted;
5277 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
5279 __skb_queue_before(list, skb, nskb);
5281 __skb_queue_tail(&tmp, nskb); /* defer rbtree insertion */
5282 skb_set_owner_r(nskb, sk);
5283 mptcp_skb_ext_move(nskb, skb);
5285 /* Copy data, releasing collapsed skbs. */
5287 int offset = start - TCP_SKB_CB(skb)->seq;
5288 int size = TCP_SKB_CB(skb)->end_seq - start;
5292 size = min(copy, size);
5293 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
5295 TCP_SKB_CB(nskb)->end_seq += size;
5299 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
5300 skb = tcp_collapse_one(sk, skb, list, root);
5303 !mptcp_skb_can_collapse(nskb, skb) ||
5304 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
5306 #ifdef CONFIG_TLS_DEVICE
5307 if (skb->decrypted != nskb->decrypted)
5314 skb_queue_walk_safe(&tmp, skb, n)
5315 tcp_rbtree_insert(root, skb);
5318 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
5319 * and tcp_collapse() them until all the queue is collapsed.
5321 static void tcp_collapse_ofo_queue(struct sock *sk)
5323 struct tcp_sock *tp = tcp_sk(sk);
5324 u32 range_truesize, sum_tiny = 0;
5325 struct sk_buff *skb, *head;
5328 skb = skb_rb_first(&tp->out_of_order_queue);
5331 tp->ooo_last_skb = skb_rb_last(&tp->out_of_order_queue);
5334 start = TCP_SKB_CB(skb)->seq;
5335 end = TCP_SKB_CB(skb)->end_seq;
5336 range_truesize = skb->truesize;
5338 for (head = skb;;) {
5339 skb = skb_rb_next(skb);
5341 /* Range is terminated when we see a gap or when
5342 * we are at the queue end.
5345 after(TCP_SKB_CB(skb)->seq, end) ||
5346 before(TCP_SKB_CB(skb)->end_seq, start)) {
5347 /* Do not attempt collapsing tiny skbs */
5348 if (range_truesize != head->truesize ||
5349 end - start >= SKB_WITH_OVERHEAD(PAGE_SIZE)) {
5350 tcp_collapse(sk, NULL, &tp->out_of_order_queue,
5351 head, skb, start, end);
5353 sum_tiny += range_truesize;
5354 if (sum_tiny > sk->sk_rcvbuf >> 3)
5360 range_truesize += skb->truesize;
5361 if (unlikely(before(TCP_SKB_CB(skb)->seq, start)))
5362 start = TCP_SKB_CB(skb)->seq;
5363 if (after(TCP_SKB_CB(skb)->end_seq, end))
5364 end = TCP_SKB_CB(skb)->end_seq;
5369 * Clean the out-of-order queue to make room.
5370 * We drop high sequences packets to :
5371 * 1) Let a chance for holes to be filled.
5372 * This means we do not drop packets from ooo queue if their sequence
5373 * is before incoming packet sequence.
5374 * 2) not add too big latencies if thousands of packets sit there.
5375 * (But if application shrinks SO_RCVBUF, we could still end up
5376 * freeing whole queue here)
5377 * 3) Drop at least 12.5 % of sk_rcvbuf to avoid malicious attacks.
5379 * Return true if queue has shrunk.
5381 static bool tcp_prune_ofo_queue(struct sock *sk, const struct sk_buff *in_skb)
5383 struct tcp_sock *tp = tcp_sk(sk);
5384 struct rb_node *node, *prev;
5385 bool pruned = false;
5388 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
5391 goal = sk->sk_rcvbuf >> 3;
5392 node = &tp->ooo_last_skb->rbnode;
5395 struct sk_buff *skb = rb_to_skb(node);
5397 /* If incoming skb would land last in ofo queue, stop pruning. */
5398 if (after(TCP_SKB_CB(in_skb)->seq, TCP_SKB_CB(skb)->seq))
5401 prev = rb_prev(node);
5402 rb_erase(node, &tp->out_of_order_queue);
5403 goal -= skb->truesize;
5404 tcp_drop_reason(sk, skb, SKB_DROP_REASON_TCP_OFO_QUEUE_PRUNE);
5405 tp->ooo_last_skb = rb_to_skb(prev);
5406 if (!prev || goal <= 0) {
5407 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf &&
5408 !tcp_under_memory_pressure(sk))
5410 goal = sk->sk_rcvbuf >> 3;
5416 NET_INC_STATS(sock_net(sk), LINUX_MIB_OFOPRUNED);
5417 /* Reset SACK state. A conforming SACK implementation will
5418 * do the same at a timeout based retransmit. When a connection
5419 * is in a sad state like this, we care only about integrity
5420 * of the connection not performance.
5422 if (tp->rx_opt.sack_ok)
5423 tcp_sack_reset(&tp->rx_opt);
5428 /* Reduce allocated memory if we can, trying to get
5429 * the socket within its memory limits again.
5431 * Return less than zero if we should start dropping frames
5432 * until the socket owning process reads some of the data
5433 * to stabilize the situation.
5435 static int tcp_prune_queue(struct sock *sk, const struct sk_buff *in_skb)
5437 struct tcp_sock *tp = tcp_sk(sk);
5439 NET_INC_STATS(sock_net(sk), LINUX_MIB_PRUNECALLED);
5441 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
5442 tcp_clamp_window(sk);
5443 else if (tcp_under_memory_pressure(sk))
5444 tcp_adjust_rcv_ssthresh(sk);
5446 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5449 tcp_collapse_ofo_queue(sk);
5450 if (!skb_queue_empty(&sk->sk_receive_queue))
5451 tcp_collapse(sk, &sk->sk_receive_queue, NULL,
5452 skb_peek(&sk->sk_receive_queue),
5454 tp->copied_seq, tp->rcv_nxt);
5456 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5459 /* Collapsing did not help, destructive actions follow.
5460 * This must not ever occur. */
5462 tcp_prune_ofo_queue(sk, in_skb);
5464 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5467 /* If we are really being abused, tell the caller to silently
5468 * drop receive data on the floor. It will get retransmitted
5469 * and hopefully then we'll have sufficient space.
5471 NET_INC_STATS(sock_net(sk), LINUX_MIB_RCVPRUNED);
5473 /* Massive buffer overcommit. */
5478 static bool tcp_should_expand_sndbuf(struct sock *sk)
5480 const struct tcp_sock *tp = tcp_sk(sk);
5482 /* If the user specified a specific send buffer setting, do
5485 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
5488 /* If we are under global TCP memory pressure, do not expand. */
5489 if (tcp_under_memory_pressure(sk)) {
5490 int unused_mem = sk_unused_reserved_mem(sk);
5492 /* Adjust sndbuf according to reserved mem. But make sure
5493 * it never goes below SOCK_MIN_SNDBUF.
5494 * See sk_stream_moderate_sndbuf() for more details.
5496 if (unused_mem > SOCK_MIN_SNDBUF)
5497 WRITE_ONCE(sk->sk_sndbuf, unused_mem);
5502 /* If we are under soft global TCP memory pressure, do not expand. */
5503 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
5506 /* If we filled the congestion window, do not expand. */
5507 if (tcp_packets_in_flight(tp) >= tcp_snd_cwnd(tp))
5513 static void tcp_new_space(struct sock *sk)
5515 struct tcp_sock *tp = tcp_sk(sk);
5517 if (tcp_should_expand_sndbuf(sk)) {
5518 tcp_sndbuf_expand(sk);
5519 tp->snd_cwnd_stamp = tcp_jiffies32;
5522 INDIRECT_CALL_1(sk->sk_write_space, sk_stream_write_space, sk);
5525 /* Caller made space either from:
5526 * 1) Freeing skbs in rtx queues (after tp->snd_una has advanced)
5527 * 2) Sent skbs from output queue (and thus advancing tp->snd_nxt)
5529 * We might be able to generate EPOLLOUT to the application if:
5530 * 1) Space consumed in output/rtx queues is below sk->sk_sndbuf/2
5531 * 2) notsent amount (tp->write_seq - tp->snd_nxt) became
5532 * small enough that tcp_stream_memory_free() decides it
5533 * is time to generate EPOLLOUT.
5535 void tcp_check_space(struct sock *sk)
5537 /* pairs with tcp_poll() */
5539 if (sk->sk_socket &&
5540 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
5542 if (!test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
5543 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
5547 static inline void tcp_data_snd_check(struct sock *sk)
5549 tcp_push_pending_frames(sk);
5550 tcp_check_space(sk);
5554 * Check if sending an ack is needed.
5556 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
5558 struct tcp_sock *tp = tcp_sk(sk);
5559 unsigned long rtt, delay;
5561 /* More than one full frame received... */
5562 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
5563 /* ... and right edge of window advances far enough.
5564 * (tcp_recvmsg() will send ACK otherwise).
5565 * If application uses SO_RCVLOWAT, we want send ack now if
5566 * we have not received enough bytes to satisfy the condition.
5568 (tp->rcv_nxt - tp->copied_seq < sk->sk_rcvlowat ||
5569 __tcp_select_window(sk) >= tp->rcv_wnd)) ||
5570 /* We ACK each frame or... */
5571 tcp_in_quickack_mode(sk) ||
5572 /* Protocol state mandates a one-time immediate ACK */
5573 inet_csk(sk)->icsk_ack.pending & ICSK_ACK_NOW) {
5579 if (!ofo_possible || RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
5580 tcp_send_delayed_ack(sk);
5584 if (!tcp_is_sack(tp) ||
5585 tp->compressed_ack >= READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_nr))
5588 if (tp->compressed_ack_rcv_nxt != tp->rcv_nxt) {
5589 tp->compressed_ack_rcv_nxt = tp->rcv_nxt;
5590 tp->dup_ack_counter = 0;
5592 if (tp->dup_ack_counter < TCP_FASTRETRANS_THRESH) {
5593 tp->dup_ack_counter++;
5596 tp->compressed_ack++;
5597 if (hrtimer_is_queued(&tp->compressed_ack_timer))
5600 /* compress ack timer : 5 % of rtt, but no more than tcp_comp_sack_delay_ns */
5602 rtt = tp->rcv_rtt_est.rtt_us;
5603 if (tp->srtt_us && tp->srtt_us < rtt)
5606 delay = min_t(unsigned long,
5607 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_delay_ns),
5608 rtt * (NSEC_PER_USEC >> 3)/20);
5610 hrtimer_start_range_ns(&tp->compressed_ack_timer, ns_to_ktime(delay),
5611 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_slack_ns),
5612 HRTIMER_MODE_REL_PINNED_SOFT);
5615 static inline void tcp_ack_snd_check(struct sock *sk)
5617 if (!inet_csk_ack_scheduled(sk)) {
5618 /* We sent a data segment already. */
5621 __tcp_ack_snd_check(sk, 1);
5625 * This routine is only called when we have urgent data
5626 * signaled. Its the 'slow' part of tcp_urg. It could be
5627 * moved inline now as tcp_urg is only called from one
5628 * place. We handle URGent data wrong. We have to - as
5629 * BSD still doesn't use the correction from RFC961.
5630 * For 1003.1g we should support a new option TCP_STDURG to permit
5631 * either form (or just set the sysctl tcp_stdurg).
5634 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
5636 struct tcp_sock *tp = tcp_sk(sk);
5637 u32 ptr = ntohs(th->urg_ptr);
5639 if (ptr && !READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_stdurg))
5641 ptr += ntohl(th->seq);
5643 /* Ignore urgent data that we've already seen and read. */
5644 if (after(tp->copied_seq, ptr))
5647 /* Do not replay urg ptr.
5649 * NOTE: interesting situation not covered by specs.
5650 * Misbehaving sender may send urg ptr, pointing to segment,
5651 * which we already have in ofo queue. We are not able to fetch
5652 * such data and will stay in TCP_URG_NOTYET until will be eaten
5653 * by recvmsg(). Seems, we are not obliged to handle such wicked
5654 * situations. But it is worth to think about possibility of some
5655 * DoSes using some hypothetical application level deadlock.
5657 if (before(ptr, tp->rcv_nxt))
5660 /* Do we already have a newer (or duplicate) urgent pointer? */
5661 if (tp->urg_data && !after(ptr, tp->urg_seq))
5664 /* Tell the world about our new urgent pointer. */
5667 /* We may be adding urgent data when the last byte read was
5668 * urgent. To do this requires some care. We cannot just ignore
5669 * tp->copied_seq since we would read the last urgent byte again
5670 * as data, nor can we alter copied_seq until this data arrives
5671 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5673 * NOTE. Double Dutch. Rendering to plain English: author of comment
5674 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5675 * and expect that both A and B disappear from stream. This is _wrong_.
5676 * Though this happens in BSD with high probability, this is occasional.
5677 * Any application relying on this is buggy. Note also, that fix "works"
5678 * only in this artificial test. Insert some normal data between A and B and we will
5679 * decline of BSD again. Verdict: it is better to remove to trap
5682 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5683 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5684 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5686 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5687 __skb_unlink(skb, &sk->sk_receive_queue);
5692 WRITE_ONCE(tp->urg_data, TCP_URG_NOTYET);
5693 WRITE_ONCE(tp->urg_seq, ptr);
5695 /* Disable header prediction. */
5699 /* This is the 'fast' part of urgent handling. */
5700 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
5702 struct tcp_sock *tp = tcp_sk(sk);
5704 /* Check if we get a new urgent pointer - normally not. */
5705 if (unlikely(th->urg))
5706 tcp_check_urg(sk, th);
5708 /* Do we wait for any urgent data? - normally not... */
5709 if (unlikely(tp->urg_data == TCP_URG_NOTYET)) {
5710 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5713 /* Is the urgent pointer pointing into this packet? */
5714 if (ptr < skb->len) {
5716 if (skb_copy_bits(skb, ptr, &tmp, 1))
5718 WRITE_ONCE(tp->urg_data, TCP_URG_VALID | tmp);
5719 if (!sock_flag(sk, SOCK_DEAD))
5720 sk->sk_data_ready(sk);
5725 /* Accept RST for rcv_nxt - 1 after a FIN.
5726 * When tcp connections are abruptly terminated from Mac OSX (via ^C), a
5727 * FIN is sent followed by a RST packet. The RST is sent with the same
5728 * sequence number as the FIN, and thus according to RFC 5961 a challenge
5729 * ACK should be sent. However, Mac OSX rate limits replies to challenge
5730 * ACKs on the closed socket. In addition middleboxes can drop either the
5731 * challenge ACK or a subsequent RST.
5733 static bool tcp_reset_check(const struct sock *sk, const struct sk_buff *skb)
5735 const struct tcp_sock *tp = tcp_sk(sk);
5737 return unlikely(TCP_SKB_CB(skb)->seq == (tp->rcv_nxt - 1) &&
5738 (1 << sk->sk_state) & (TCPF_CLOSE_WAIT | TCPF_LAST_ACK |
5742 /* Does PAWS and seqno based validation of an incoming segment, flags will
5743 * play significant role here.
5745 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5746 const struct tcphdr *th, int syn_inerr)
5748 struct tcp_sock *tp = tcp_sk(sk);
5751 /* RFC1323: H1. Apply PAWS check first. */
5752 if (tcp_fast_parse_options(sock_net(sk), skb, th, tp) &&
5753 tp->rx_opt.saw_tstamp &&
5754 tcp_paws_discard(sk, skb)) {
5756 if (unlikely(th->syn))
5758 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5759 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5760 LINUX_MIB_TCPACKSKIPPEDPAWS,
5761 &tp->last_oow_ack_time))
5762 tcp_send_dupack(sk, skb);
5763 SKB_DR_SET(reason, TCP_RFC7323_PAWS);
5766 /* Reset is accepted even if it did not pass PAWS. */
5769 /* Step 1: check sequence number */
5770 reason = tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
5772 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5773 * (RST) segments are validated by checking their SEQ-fields."
5774 * And page 69: "If an incoming segment is not acceptable,
5775 * an acknowledgment should be sent in reply (unless the RST
5776 * bit is set, if so drop the segment and return)".
5781 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5782 LINUX_MIB_TCPACKSKIPPEDSEQ,
5783 &tp->last_oow_ack_time))
5784 tcp_send_dupack(sk, skb);
5785 } else if (tcp_reset_check(sk, skb)) {
5791 /* Step 2: check RST bit */
5793 /* RFC 5961 3.2 (extend to match against (RCV.NXT - 1) after a
5794 * FIN and SACK too if available):
5795 * If seq num matches RCV.NXT or (RCV.NXT - 1) after a FIN, or
5796 * the right-most SACK block,
5798 * RESET the connection
5800 * Send a challenge ACK
5802 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt ||
5803 tcp_reset_check(sk, skb))
5806 if (tcp_is_sack(tp) && tp->rx_opt.num_sacks > 0) {
5807 struct tcp_sack_block *sp = &tp->selective_acks[0];
5808 int max_sack = sp[0].end_seq;
5811 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;
5813 max_sack = after(sp[this_sack].end_seq,
5815 sp[this_sack].end_seq : max_sack;
5818 if (TCP_SKB_CB(skb)->seq == max_sack)
5822 /* Disable TFO if RST is out-of-order
5823 * and no data has been received
5824 * for current active TFO socket
5826 if (tp->syn_fastopen && !tp->data_segs_in &&
5827 sk->sk_state == TCP_ESTABLISHED)
5828 tcp_fastopen_active_disable(sk);
5829 tcp_send_challenge_ack(sk);
5830 SKB_DR_SET(reason, TCP_RESET);
5834 /* step 3: check security and precedence [ignored] */
5836 /* step 4: Check for a SYN
5837 * RFC 5961 4.2 : Send a challenge ack
5842 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5843 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5844 tcp_send_challenge_ack(sk);
5845 SKB_DR_SET(reason, TCP_INVALID_SYN);
5849 bpf_skops_parse_hdr(sk, skb);
5854 tcp_drop_reason(sk, skb, reason);
5864 * TCP receive function for the ESTABLISHED state.
5866 * It is split into a fast path and a slow path. The fast path is
5868 * - A zero window was announced from us - zero window probing
5869 * is only handled properly in the slow path.
5870 * - Out of order segments arrived.
5871 * - Urgent data is expected.
5872 * - There is no buffer space left
5873 * - Unexpected TCP flags/window values/header lengths are received
5874 * (detected by checking the TCP header against pred_flags)
5875 * - Data is sent in both directions. Fast path only supports pure senders
5876 * or pure receivers (this means either the sequence number or the ack
5877 * value must stay constant)
5878 * - Unexpected TCP option.
5880 * When these conditions are not satisfied it drops into a standard
5881 * receive procedure patterned after RFC793 to handle all cases.
5882 * The first three cases are guaranteed by proper pred_flags setting,
5883 * the rest is checked inline. Fast processing is turned on in
5884 * tcp_data_queue when everything is OK.
5886 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb)
5888 enum skb_drop_reason reason = SKB_DROP_REASON_NOT_SPECIFIED;
5889 const struct tcphdr *th = (const struct tcphdr *)skb->data;
5890 struct tcp_sock *tp = tcp_sk(sk);
5891 unsigned int len = skb->len;
5893 /* TCP congestion window tracking */
5894 trace_tcp_probe(sk, skb);
5896 tcp_mstamp_refresh(tp);
5897 if (unlikely(!rcu_access_pointer(sk->sk_rx_dst)))
5898 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
5900 * Header prediction.
5901 * The code loosely follows the one in the famous
5902 * "30 instruction TCP receive" Van Jacobson mail.
5904 * Van's trick is to deposit buffers into socket queue
5905 * on a device interrupt, to call tcp_recv function
5906 * on the receive process context and checksum and copy
5907 * the buffer to user space. smart...
5909 * Our current scheme is not silly either but we take the
5910 * extra cost of the net_bh soft interrupt processing...
5911 * We do checksum and copy also but from device to kernel.
5914 tp->rx_opt.saw_tstamp = 0;
5916 /* pred_flags is 0xS?10 << 16 + snd_wnd
5917 * if header_prediction is to be made
5918 * 'S' will always be tp->tcp_header_len >> 2
5919 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5920 * turn it off (when there are holes in the receive
5921 * space for instance)
5922 * PSH flag is ignored.
5925 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5926 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5927 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5928 int tcp_header_len = tp->tcp_header_len;
5930 /* Timestamp header prediction: tcp_header_len
5931 * is automatically equal to th->doff*4 due to pred_flags
5935 /* Check timestamp */
5936 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5937 /* No? Slow path! */
5938 if (!tcp_parse_aligned_timestamp(tp, th))
5941 /* If PAWS failed, check it more carefully in slow path */
5942 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5945 /* DO NOT update ts_recent here, if checksum fails
5946 * and timestamp was corrupted part, it will result
5947 * in a hung connection since we will drop all
5948 * future packets due to the PAWS test.
5952 if (len <= tcp_header_len) {
5953 /* Bulk data transfer: sender */
5954 if (len == tcp_header_len) {
5955 /* Predicted packet is in window by definition.
5956 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5957 * Hence, check seq<=rcv_wup reduces to:
5959 if (tcp_header_len ==
5960 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5961 tp->rcv_nxt == tp->rcv_wup)
5962 tcp_store_ts_recent(tp);
5964 /* We know that such packets are checksummed
5967 tcp_ack(sk, skb, 0);
5969 tcp_data_snd_check(sk);
5970 /* When receiving pure ack in fast path, update
5971 * last ts ecr directly instead of calling
5972 * tcp_rcv_rtt_measure_ts()
5974 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr;
5976 } else { /* Header too small */
5977 reason = SKB_DROP_REASON_PKT_TOO_SMALL;
5978 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5983 bool fragstolen = false;
5985 if (tcp_checksum_complete(skb))
5988 if ((int)skb->truesize > sk->sk_forward_alloc)
5991 /* Predicted packet is in window by definition.
5992 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5993 * Hence, check seq<=rcv_wup reduces to:
5995 if (tcp_header_len ==
5996 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5997 tp->rcv_nxt == tp->rcv_wup)
5998 tcp_store_ts_recent(tp);
6000 tcp_rcv_rtt_measure_ts(sk, skb);
6002 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPHITS);
6004 /* Bulk data transfer: receiver */
6006 __skb_pull(skb, tcp_header_len);
6007 eaten = tcp_queue_rcv(sk, skb, &fragstolen);
6009 tcp_event_data_recv(sk, skb);
6011 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
6012 /* Well, only one small jumplet in fast path... */
6013 tcp_ack(sk, skb, FLAG_DATA);
6014 tcp_data_snd_check(sk);
6015 if (!inet_csk_ack_scheduled(sk))
6018 tcp_update_wl(tp, TCP_SKB_CB(skb)->seq);
6021 __tcp_ack_snd_check(sk, 0);
6024 kfree_skb_partial(skb, fragstolen);
6031 if (len < (th->doff << 2) || tcp_checksum_complete(skb))
6034 if (!th->ack && !th->rst && !th->syn) {
6035 reason = SKB_DROP_REASON_TCP_FLAGS;
6040 * Standard slow path.
6043 if (!tcp_validate_incoming(sk, skb, th, 1))
6047 reason = tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT);
6048 if ((int)reason < 0) {
6052 tcp_rcv_rtt_measure_ts(sk, skb);
6054 /* Process urgent data. */
6055 tcp_urg(sk, skb, th);
6057 /* step 7: process the segment text */
6058 tcp_data_queue(sk, skb);
6060 tcp_data_snd_check(sk);
6061 tcp_ack_snd_check(sk);
6065 reason = SKB_DROP_REASON_TCP_CSUM;
6066 trace_tcp_bad_csum(skb);
6067 TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS);
6068 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
6071 tcp_drop_reason(sk, skb, reason);
6073 EXPORT_SYMBOL(tcp_rcv_established);
6075 void tcp_init_transfer(struct sock *sk, int bpf_op, struct sk_buff *skb)
6077 struct inet_connection_sock *icsk = inet_csk(sk);
6078 struct tcp_sock *tp = tcp_sk(sk);
6081 icsk->icsk_af_ops->rebuild_header(sk);
6082 tcp_init_metrics(sk);
6084 /* Initialize the congestion window to start the transfer.
6085 * Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
6086 * retransmitted. In light of RFC6298 more aggressive 1sec
6087 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
6088 * retransmission has occurred.
6090 if (tp->total_retrans > 1 && tp->undo_marker)
6091 tcp_snd_cwnd_set(tp, 1);
6093 tcp_snd_cwnd_set(tp, tcp_init_cwnd(tp, __sk_dst_get(sk)));
6094 tp->snd_cwnd_stamp = tcp_jiffies32;
6096 bpf_skops_established(sk, bpf_op, skb);
6097 /* Initialize congestion control unless BPF initialized it already: */
6098 if (!icsk->icsk_ca_initialized)
6099 tcp_init_congestion_control(sk);
6100 tcp_init_buffer_space(sk);
6103 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
6105 struct tcp_sock *tp = tcp_sk(sk);
6106 struct inet_connection_sock *icsk = inet_csk(sk);
6108 tcp_set_state(sk, TCP_ESTABLISHED);
6109 icsk->icsk_ack.lrcvtime = tcp_jiffies32;
6112 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
6113 security_inet_conn_established(sk, skb);
6114 sk_mark_napi_id(sk, skb);
6117 tcp_init_transfer(sk, BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB, skb);
6119 /* Prevent spurious tcp_cwnd_restart() on first data
6122 tp->lsndtime = tcp_jiffies32;
6124 if (sock_flag(sk, SOCK_KEEPOPEN))
6125 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
6127 if (!tp->rx_opt.snd_wscale)
6128 __tcp_fast_path_on(tp, tp->snd_wnd);
6133 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
6134 struct tcp_fastopen_cookie *cookie)
6136 struct tcp_sock *tp = tcp_sk(sk);
6137 struct sk_buff *data = tp->syn_data ? tcp_rtx_queue_head(sk) : NULL;
6138 u16 mss = tp->rx_opt.mss_clamp, try_exp = 0;
6139 bool syn_drop = false;
6141 if (mss == tp->rx_opt.user_mss) {
6142 struct tcp_options_received opt;
6144 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
6145 tcp_clear_options(&opt);
6146 opt.user_mss = opt.mss_clamp = 0;
6147 tcp_parse_options(sock_net(sk), synack, &opt, 0, NULL);
6148 mss = opt.mss_clamp;
6151 if (!tp->syn_fastopen) {
6152 /* Ignore an unsolicited cookie */
6154 } else if (tp->total_retrans) {
6155 /* SYN timed out and the SYN-ACK neither has a cookie nor
6156 * acknowledges data. Presumably the remote received only
6157 * the retransmitted (regular) SYNs: either the original
6158 * SYN-data or the corresponding SYN-ACK was dropped.
6160 syn_drop = (cookie->len < 0 && data);
6161 } else if (cookie->len < 0 && !tp->syn_data) {
6162 /* We requested a cookie but didn't get it. If we did not use
6163 * the (old) exp opt format then try so next time (try_exp=1).
6164 * Otherwise we go back to use the RFC7413 opt (try_exp=2).
6166 try_exp = tp->syn_fastopen_exp ? 2 : 1;
6169 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp);
6171 if (data) { /* Retransmit unacked data in SYN */
6172 if (tp->total_retrans)
6173 tp->fastopen_client_fail = TFO_SYN_RETRANSMITTED;
6175 tp->fastopen_client_fail = TFO_DATA_NOT_ACKED;
6176 skb_rbtree_walk_from(data)
6177 tcp_mark_skb_lost(sk, data);
6178 tcp_xmit_retransmit_queue(sk);
6179 NET_INC_STATS(sock_net(sk),
6180 LINUX_MIB_TCPFASTOPENACTIVEFAIL);
6183 tp->syn_data_acked = tp->syn_data;
6184 if (tp->syn_data_acked) {
6185 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
6186 /* SYN-data is counted as two separate packets in tcp_ack() */
6187 if (tp->delivered > 1)
6191 tcp_fastopen_add_skb(sk, synack);
6196 static void smc_check_reset_syn(struct tcp_sock *tp)
6198 #if IS_ENABLED(CONFIG_SMC)
6199 if (static_branch_unlikely(&tcp_have_smc)) {
6200 if (tp->syn_smc && !tp->rx_opt.smc_ok)
6206 static void tcp_try_undo_spurious_syn(struct sock *sk)
6208 struct tcp_sock *tp = tcp_sk(sk);
6211 /* undo_marker is set when SYN or SYNACK times out. The timeout is
6212 * spurious if the ACK's timestamp option echo value matches the
6213 * original SYN timestamp.
6215 syn_stamp = tp->retrans_stamp;
6216 if (tp->undo_marker && syn_stamp && tp->rx_opt.saw_tstamp &&
6217 syn_stamp == tp->rx_opt.rcv_tsecr)
6218 tp->undo_marker = 0;
6221 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
6222 const struct tcphdr *th)
6224 struct inet_connection_sock *icsk = inet_csk(sk);
6225 struct tcp_sock *tp = tcp_sk(sk);
6226 struct tcp_fastopen_cookie foc = { .len = -1 };
6227 int saved_clamp = tp->rx_opt.mss_clamp;
6231 tcp_parse_options(sock_net(sk), skb, &tp->rx_opt, 0, &foc);
6232 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
6233 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
6237 * "If the state is SYN-SENT then
6238 * first check the ACK bit
6239 * If the ACK bit is set
6240 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
6241 * a reset (unless the RST bit is set, if so drop
6242 * the segment and return)"
6244 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
6245 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
6246 /* Previous FIN/ACK or RST/ACK might be ignored. */
6247 if (icsk->icsk_retransmits == 0)
6248 inet_csk_reset_xmit_timer(sk,
6250 TCP_TIMEOUT_MIN, TCP_RTO_MAX);
6251 goto reset_and_undo;
6254 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
6255 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
6256 tcp_time_stamp(tp))) {
6257 NET_INC_STATS(sock_net(sk),
6258 LINUX_MIB_PAWSACTIVEREJECTED);
6259 goto reset_and_undo;
6262 /* Now ACK is acceptable.
6264 * "If the RST bit is set
6265 * If the ACK was acceptable then signal the user "error:
6266 * connection reset", drop the segment, enter CLOSED state,
6267 * delete TCB, and return."
6278 * "fifth, if neither of the SYN or RST bits is set then
6279 * drop the segment and return."
6285 SKB_DR_SET(reason, TCP_FLAGS);
6286 goto discard_and_undo;
6289 * "If the SYN bit is on ...
6290 * are acceptable then ...
6291 * (our SYN has been ACKed), change the connection
6292 * state to ESTABLISHED..."
6295 tcp_ecn_rcv_synack(tp, th);
6297 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
6298 tcp_try_undo_spurious_syn(sk);
6299 tcp_ack(sk, skb, FLAG_SLOWPATH);
6301 /* Ok.. it's good. Set up sequence numbers and
6302 * move to established.
6304 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1);
6305 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
6307 /* RFC1323: The window in SYN & SYN/ACK segments is
6310 tp->snd_wnd = ntohs(th->window);
6312 if (!tp->rx_opt.wscale_ok) {
6313 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
6314 tp->window_clamp = min(tp->window_clamp, 65535U);
6317 if (tp->rx_opt.saw_tstamp) {
6318 tp->rx_opt.tstamp_ok = 1;
6319 tp->tcp_header_len =
6320 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
6321 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
6322 tcp_store_ts_recent(tp);
6324 tp->tcp_header_len = sizeof(struct tcphdr);
6327 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
6328 tcp_initialize_rcv_mss(sk);
6330 /* Remember, tcp_poll() does not lock socket!
6331 * Change state from SYN-SENT only after copied_seq
6332 * is initialized. */
6333 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6335 smc_check_reset_syn(tp);
6339 tcp_finish_connect(sk, skb);
6341 fastopen_fail = (tp->syn_fastopen || tp->syn_data) &&
6342 tcp_rcv_fastopen_synack(sk, skb, &foc);
6344 if (!sock_flag(sk, SOCK_DEAD)) {
6345 sk->sk_state_change(sk);
6346 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
6350 if (sk->sk_write_pending ||
6351 READ_ONCE(icsk->icsk_accept_queue.rskq_defer_accept) ||
6352 inet_csk_in_pingpong_mode(sk)) {
6353 /* Save one ACK. Data will be ready after
6354 * several ticks, if write_pending is set.
6356 * It may be deleted, but with this feature tcpdumps
6357 * look so _wonderfully_ clever, that I was not able
6358 * to stand against the temptation 8) --ANK
6360 inet_csk_schedule_ack(sk);
6361 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
6362 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
6363 TCP_DELACK_MAX, TCP_RTO_MAX);
6370 /* No ACK in the segment */
6374 * "If the RST bit is set
6376 * Otherwise (no ACK) drop the segment and return."
6378 SKB_DR_SET(reason, TCP_RESET);
6379 goto discard_and_undo;
6383 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
6384 tcp_paws_reject(&tp->rx_opt, 0)) {
6385 SKB_DR_SET(reason, TCP_RFC7323_PAWS);
6386 goto discard_and_undo;
6389 /* We see SYN without ACK. It is attempt of
6390 * simultaneous connect with crossed SYNs.
6391 * Particularly, it can be connect to self.
6393 tcp_set_state(sk, TCP_SYN_RECV);
6395 if (tp->rx_opt.saw_tstamp) {
6396 tp->rx_opt.tstamp_ok = 1;
6397 tcp_store_ts_recent(tp);
6398 tp->tcp_header_len =
6399 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
6401 tp->tcp_header_len = sizeof(struct tcphdr);
6404 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1);
6405 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6406 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
6408 /* RFC1323: The window in SYN & SYN/ACK segments is
6411 tp->snd_wnd = ntohs(th->window);
6412 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
6413 tp->max_window = tp->snd_wnd;
6415 tcp_ecn_rcv_syn(tp, th);
6418 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
6419 tcp_initialize_rcv_mss(sk);
6421 tcp_send_synack(sk);
6423 /* Note, we could accept data and URG from this segment.
6424 * There are no obstacles to make this (except that we must
6425 * either change tcp_recvmsg() to prevent it from returning data
6426 * before 3WHS completes per RFC793, or employ TCP Fast Open).
6428 * However, if we ignore data in ACKless segments sometimes,
6429 * we have no reasons to accept it sometimes.
6430 * Also, seems the code doing it in step6 of tcp_rcv_state_process
6431 * is not flawless. So, discard packet for sanity.
6432 * Uncomment this return to process the data.
6439 /* "fifth, if neither of the SYN or RST bits is set then
6440 * drop the segment and return."
6444 tcp_clear_options(&tp->rx_opt);
6445 tp->rx_opt.mss_clamp = saved_clamp;
6446 tcp_drop_reason(sk, skb, reason);
6450 tcp_clear_options(&tp->rx_opt);
6451 tp->rx_opt.mss_clamp = saved_clamp;
6455 static void tcp_rcv_synrecv_state_fastopen(struct sock *sk)
6457 struct tcp_sock *tp = tcp_sk(sk);
6458 struct request_sock *req;
6460 /* If we are still handling the SYNACK RTO, see if timestamp ECR allows
6461 * undo. If peer SACKs triggered fast recovery, we can't undo here.
6463 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss && !tp->packets_out)
6464 tcp_try_undo_recovery(sk);
6466 /* Reset rtx states to prevent spurious retransmits_timed_out() */
6467 tp->retrans_stamp = 0;
6468 inet_csk(sk)->icsk_retransmits = 0;
6470 /* Once we leave TCP_SYN_RECV or TCP_FIN_WAIT_1,
6471 * we no longer need req so release it.
6473 req = rcu_dereference_protected(tp->fastopen_rsk,
6474 lockdep_sock_is_held(sk));
6475 reqsk_fastopen_remove(sk, req, false);
6477 /* Re-arm the timer because data may have been sent out.
6478 * This is similar to the regular data transmission case
6479 * when new data has just been ack'ed.
6481 * (TFO) - we could try to be more aggressive and
6482 * retransmitting any data sooner based on when they
6489 * This function implements the receiving procedure of RFC 793 for
6490 * all states except ESTABLISHED and TIME_WAIT.
6491 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
6492 * address independent.
6495 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb)
6497 struct tcp_sock *tp = tcp_sk(sk);
6498 struct inet_connection_sock *icsk = inet_csk(sk);
6499 const struct tcphdr *th = tcp_hdr(skb);
6500 struct request_sock *req;
6505 switch (sk->sk_state) {
6507 SKB_DR_SET(reason, TCP_CLOSE);
6515 SKB_DR_SET(reason, TCP_RESET);
6520 SKB_DR_SET(reason, TCP_FLAGS);
6523 /* It is possible that we process SYN packets from backlog,
6524 * so we need to make sure to disable BH and RCU right there.
6528 acceptable = icsk->icsk_af_ops->conn_request(sk, skb) >= 0;
6537 SKB_DR_SET(reason, TCP_FLAGS);
6541 tp->rx_opt.saw_tstamp = 0;
6542 tcp_mstamp_refresh(tp);
6543 queued = tcp_rcv_synsent_state_process(sk, skb, th);
6547 /* Do step6 onward by hand. */
6548 tcp_urg(sk, skb, th);
6550 tcp_data_snd_check(sk);
6554 tcp_mstamp_refresh(tp);
6555 tp->rx_opt.saw_tstamp = 0;
6556 req = rcu_dereference_protected(tp->fastopen_rsk,
6557 lockdep_sock_is_held(sk));
6561 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
6562 sk->sk_state != TCP_FIN_WAIT1);
6564 if (!tcp_check_req(sk, skb, req, true, &req_stolen)) {
6565 SKB_DR_SET(reason, TCP_FASTOPEN);
6570 if (!th->ack && !th->rst && !th->syn) {
6571 SKB_DR_SET(reason, TCP_FLAGS);
6574 if (!tcp_validate_incoming(sk, skb, th, 0))
6577 /* step 5: check the ACK field */
6578 acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
6579 FLAG_UPDATE_TS_RECENT |
6580 FLAG_NO_CHALLENGE_ACK) > 0;
6583 if (sk->sk_state == TCP_SYN_RECV)
6584 return 1; /* send one RST */
6585 tcp_send_challenge_ack(sk);
6586 SKB_DR_SET(reason, TCP_OLD_ACK);
6589 switch (sk->sk_state) {
6591 tp->delivered++; /* SYN-ACK delivery isn't tracked in tcp_ack */
6593 tcp_synack_rtt_meas(sk, req);
6596 tcp_rcv_synrecv_state_fastopen(sk);
6598 tcp_try_undo_spurious_syn(sk);
6599 tp->retrans_stamp = 0;
6600 tcp_init_transfer(sk, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB,
6602 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6605 tcp_set_state(sk, TCP_ESTABLISHED);
6606 sk->sk_state_change(sk);
6608 /* Note, that this wakeup is only for marginal crossed SYN case.
6609 * Passively open sockets are not waked up, because
6610 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
6613 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
6615 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
6616 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
6617 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
6619 if (tp->rx_opt.tstamp_ok)
6620 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
6622 if (!inet_csk(sk)->icsk_ca_ops->cong_control)
6623 tcp_update_pacing_rate(sk);
6625 /* Prevent spurious tcp_cwnd_restart() on first data packet */
6626 tp->lsndtime = tcp_jiffies32;
6628 tcp_initialize_rcv_mss(sk);
6629 tcp_fast_path_on(tp);
6632 case TCP_FIN_WAIT1: {
6636 tcp_rcv_synrecv_state_fastopen(sk);
6638 if (tp->snd_una != tp->write_seq)
6641 tcp_set_state(sk, TCP_FIN_WAIT2);
6642 WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | SEND_SHUTDOWN);
6646 if (!sock_flag(sk, SOCK_DEAD)) {
6647 /* Wake up lingering close() */
6648 sk->sk_state_change(sk);
6652 if (READ_ONCE(tp->linger2) < 0) {
6654 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6657 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6658 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6659 /* Receive out of order FIN after close() */
6660 if (tp->syn_fastopen && th->fin)
6661 tcp_fastopen_active_disable(sk);
6663 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6667 tmo = tcp_fin_time(sk);
6668 if (tmo > TCP_TIMEWAIT_LEN) {
6669 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
6670 } else if (th->fin || sock_owned_by_user(sk)) {
6671 /* Bad case. We could lose such FIN otherwise.
6672 * It is not a big problem, but it looks confusing
6673 * and not so rare event. We still can lose it now,
6674 * if it spins in bh_lock_sock(), but it is really
6677 inet_csk_reset_keepalive_timer(sk, tmo);
6679 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
6686 if (tp->snd_una == tp->write_seq) {
6687 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
6693 if (tp->snd_una == tp->write_seq) {
6694 tcp_update_metrics(sk);
6701 /* step 6: check the URG bit */
6702 tcp_urg(sk, skb, th);
6704 /* step 7: process the segment text */
6705 switch (sk->sk_state) {
6706 case TCP_CLOSE_WAIT:
6709 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
6710 /* If a subflow has been reset, the packet should not
6711 * continue to be processed, drop the packet.
6713 if (sk_is_mptcp(sk) && !mptcp_incoming_options(sk, skb))
6720 /* RFC 793 says to queue data in these states,
6721 * RFC 1122 says we MUST send a reset.
6722 * BSD 4.4 also does reset.
6724 if (sk->sk_shutdown & RCV_SHUTDOWN) {
6725 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6726 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6727 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6733 case TCP_ESTABLISHED:
6734 tcp_data_queue(sk, skb);
6739 /* tcp_data could move socket to TIME-WAIT */
6740 if (sk->sk_state != TCP_CLOSE) {
6741 tcp_data_snd_check(sk);
6742 tcp_ack_snd_check(sk);
6747 tcp_drop_reason(sk, skb, reason);
6755 EXPORT_SYMBOL(tcp_rcv_state_process);
6757 static inline void pr_drop_req(struct request_sock *req, __u16 port, int family)
6759 struct inet_request_sock *ireq = inet_rsk(req);
6761 if (family == AF_INET)
6762 net_dbg_ratelimited("drop open request from %pI4/%u\n",
6763 &ireq->ir_rmt_addr, port);
6764 #if IS_ENABLED(CONFIG_IPV6)
6765 else if (family == AF_INET6)
6766 net_dbg_ratelimited("drop open request from %pI6/%u\n",
6767 &ireq->ir_v6_rmt_addr, port);
6771 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
6773 * If we receive a SYN packet with these bits set, it means a
6774 * network is playing bad games with TOS bits. In order to
6775 * avoid possible false congestion notifications, we disable
6776 * TCP ECN negotiation.
6778 * Exception: tcp_ca wants ECN. This is required for DCTCP
6779 * congestion control: Linux DCTCP asserts ECT on all packets,
6780 * including SYN, which is most optimal solution; however,
6781 * others, such as FreeBSD do not.
6783 * Exception: At least one of the reserved bits of the TCP header (th->res1) is
6784 * set, indicating the use of a future TCP extension (such as AccECN). See
6785 * RFC8311 §4.3 which updates RFC3168 to allow the development of such
6788 static void tcp_ecn_create_request(struct request_sock *req,
6789 const struct sk_buff *skb,
6790 const struct sock *listen_sk,
6791 const struct dst_entry *dst)
6793 const struct tcphdr *th = tcp_hdr(skb);
6794 const struct net *net = sock_net(listen_sk);
6795 bool th_ecn = th->ece && th->cwr;
6802 ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield);
6803 ecn_ok_dst = dst_feature(dst, DST_FEATURE_ECN_MASK);
6804 ecn_ok = READ_ONCE(net->ipv4.sysctl_tcp_ecn) || ecn_ok_dst;
6806 if (((!ect || th->res1) && ecn_ok) || tcp_ca_needs_ecn(listen_sk) ||
6807 (ecn_ok_dst & DST_FEATURE_ECN_CA) ||
6808 tcp_bpf_ca_needs_ecn((struct sock *)req))
6809 inet_rsk(req)->ecn_ok = 1;
6812 static void tcp_openreq_init(struct request_sock *req,
6813 const struct tcp_options_received *rx_opt,
6814 struct sk_buff *skb, const struct sock *sk)
6816 struct inet_request_sock *ireq = inet_rsk(req);
6818 req->rsk_rcv_wnd = 0; /* So that tcp_send_synack() knows! */
6819 tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq;
6820 tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
6821 tcp_rsk(req)->snt_synack = 0;
6822 tcp_rsk(req)->last_oow_ack_time = 0;
6823 req->mss = rx_opt->mss_clamp;
6824 req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0;
6825 ireq->tstamp_ok = rx_opt->tstamp_ok;
6826 ireq->sack_ok = rx_opt->sack_ok;
6827 ireq->snd_wscale = rx_opt->snd_wscale;
6828 ireq->wscale_ok = rx_opt->wscale_ok;
6831 ireq->ir_rmt_port = tcp_hdr(skb)->source;
6832 ireq->ir_num = ntohs(tcp_hdr(skb)->dest);
6833 ireq->ir_mark = inet_request_mark(sk, skb);
6834 #if IS_ENABLED(CONFIG_SMC)
6835 ireq->smc_ok = rx_opt->smc_ok && !(tcp_sk(sk)->smc_hs_congested &&
6836 tcp_sk(sk)->smc_hs_congested(sk));
6840 struct request_sock *inet_reqsk_alloc(const struct request_sock_ops *ops,
6841 struct sock *sk_listener,
6842 bool attach_listener)
6844 struct request_sock *req = reqsk_alloc(ops, sk_listener,
6848 struct inet_request_sock *ireq = inet_rsk(req);
6850 ireq->ireq_opt = NULL;
6851 #if IS_ENABLED(CONFIG_IPV6)
6852 ireq->pktopts = NULL;
6854 atomic64_set(&ireq->ir_cookie, 0);
6855 ireq->ireq_state = TCP_NEW_SYN_RECV;
6856 write_pnet(&ireq->ireq_net, sock_net(sk_listener));
6857 ireq->ireq_family = sk_listener->sk_family;
6858 req->timeout = TCP_TIMEOUT_INIT;
6863 EXPORT_SYMBOL(inet_reqsk_alloc);
6866 * Return true if a syncookie should be sent
6868 static bool tcp_syn_flood_action(const struct sock *sk, const char *proto)
6870 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
6871 const char *msg = "Dropping request";
6872 struct net *net = sock_net(sk);
6873 bool want_cookie = false;
6876 syncookies = READ_ONCE(net->ipv4.sysctl_tcp_syncookies);
6878 #ifdef CONFIG_SYN_COOKIES
6880 msg = "Sending cookies";
6882 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES);
6885 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP);
6887 if (!READ_ONCE(queue->synflood_warned) && syncookies != 2 &&
6888 xchg(&queue->synflood_warned, 1) == 0) {
6889 if (IS_ENABLED(CONFIG_IPV6) && sk->sk_family == AF_INET6) {
6890 net_info_ratelimited("%s: Possible SYN flooding on port [%pI6c]:%u. %s.\n",
6891 proto, inet6_rcv_saddr(sk),
6894 net_info_ratelimited("%s: Possible SYN flooding on port %pI4:%u. %s.\n",
6895 proto, &sk->sk_rcv_saddr,
6903 static void tcp_reqsk_record_syn(const struct sock *sk,
6904 struct request_sock *req,
6905 const struct sk_buff *skb)
6907 if (tcp_sk(sk)->save_syn) {
6908 u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb);
6909 struct saved_syn *saved_syn;
6913 if (tcp_sk(sk)->save_syn == 2) { /* Save full header. */
6914 base = skb_mac_header(skb);
6915 mac_hdrlen = skb_mac_header_len(skb);
6918 base = skb_network_header(skb);
6922 saved_syn = kmalloc(struct_size(saved_syn, data, len),
6925 saved_syn->mac_hdrlen = mac_hdrlen;
6926 saved_syn->network_hdrlen = skb_network_header_len(skb);
6927 saved_syn->tcp_hdrlen = tcp_hdrlen(skb);
6928 memcpy(saved_syn->data, base, len);
6929 req->saved_syn = saved_syn;
6934 /* If a SYN cookie is required and supported, returns a clamped MSS value to be
6935 * used for SYN cookie generation.
6937 u16 tcp_get_syncookie_mss(struct request_sock_ops *rsk_ops,
6938 const struct tcp_request_sock_ops *af_ops,
6939 struct sock *sk, struct tcphdr *th)
6941 struct tcp_sock *tp = tcp_sk(sk);
6944 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_syncookies) != 2 &&
6945 !inet_csk_reqsk_queue_is_full(sk))
6948 if (!tcp_syn_flood_action(sk, rsk_ops->slab_name))
6951 if (sk_acceptq_is_full(sk)) {
6952 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
6956 mss = tcp_parse_mss_option(th, tp->rx_opt.user_mss);
6958 mss = af_ops->mss_clamp;
6962 EXPORT_SYMBOL_GPL(tcp_get_syncookie_mss);
6964 int tcp_conn_request(struct request_sock_ops *rsk_ops,
6965 const struct tcp_request_sock_ops *af_ops,
6966 struct sock *sk, struct sk_buff *skb)
6968 struct tcp_fastopen_cookie foc = { .len = -1 };
6969 __u32 isn = TCP_SKB_CB(skb)->tcp_tw_isn;
6970 struct tcp_options_received tmp_opt;
6971 struct tcp_sock *tp = tcp_sk(sk);
6972 struct net *net = sock_net(sk);
6973 struct sock *fastopen_sk = NULL;
6974 struct request_sock *req;
6975 bool want_cookie = false;
6976 struct dst_entry *dst;
6980 syncookies = READ_ONCE(net->ipv4.sysctl_tcp_syncookies);
6982 /* TW buckets are converted to open requests without
6983 * limitations, they conserve resources and peer is
6984 * evidently real one.
6986 if ((syncookies == 2 || inet_csk_reqsk_queue_is_full(sk)) && !isn) {
6987 want_cookie = tcp_syn_flood_action(sk, rsk_ops->slab_name);
6992 if (sk_acceptq_is_full(sk)) {
6993 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
6997 req = inet_reqsk_alloc(rsk_ops, sk, !want_cookie);
7001 req->syncookie = want_cookie;
7002 tcp_rsk(req)->af_specific = af_ops;
7003 tcp_rsk(req)->ts_off = 0;
7004 #if IS_ENABLED(CONFIG_MPTCP)
7005 tcp_rsk(req)->is_mptcp = 0;
7008 tcp_clear_options(&tmp_opt);
7009 tmp_opt.mss_clamp = af_ops->mss_clamp;
7010 tmp_opt.user_mss = tp->rx_opt.user_mss;
7011 tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0,
7012 want_cookie ? NULL : &foc);
7014 if (want_cookie && !tmp_opt.saw_tstamp)
7015 tcp_clear_options(&tmp_opt);
7017 if (IS_ENABLED(CONFIG_SMC) && want_cookie)
7020 tmp_opt.tstamp_ok = tmp_opt.saw_tstamp;
7021 tcp_openreq_init(req, &tmp_opt, skb, sk);
7022 inet_rsk(req)->no_srccheck = inet_test_bit(TRANSPARENT, sk);
7024 /* Note: tcp_v6_init_req() might override ir_iif for link locals */
7025 inet_rsk(req)->ir_iif = inet_request_bound_dev_if(sk, skb);
7027 dst = af_ops->route_req(sk, skb, &fl, req);
7031 if (tmp_opt.tstamp_ok)
7032 tcp_rsk(req)->ts_off = af_ops->init_ts_off(net, skb);
7034 if (!want_cookie && !isn) {
7035 int max_syn_backlog = READ_ONCE(net->ipv4.sysctl_max_syn_backlog);
7037 /* Kill the following clause, if you dislike this way. */
7039 (max_syn_backlog - inet_csk_reqsk_queue_len(sk) <
7040 (max_syn_backlog >> 2)) &&
7041 !tcp_peer_is_proven(req, dst)) {
7042 /* Without syncookies last quarter of
7043 * backlog is filled with destinations,
7044 * proven to be alive.
7045 * It means that we continue to communicate
7046 * to destinations, already remembered
7047 * to the moment of synflood.
7049 pr_drop_req(req, ntohs(tcp_hdr(skb)->source),
7051 goto drop_and_release;
7054 isn = af_ops->init_seq(skb);
7057 tcp_ecn_create_request(req, skb, sk, dst);
7060 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss);
7061 if (!tmp_opt.tstamp_ok)
7062 inet_rsk(req)->ecn_ok = 0;
7065 tcp_rsk(req)->snt_isn = isn;
7066 tcp_rsk(req)->txhash = net_tx_rndhash();
7067 tcp_rsk(req)->syn_tos = TCP_SKB_CB(skb)->ip_dsfield;
7068 tcp_openreq_init_rwin(req, sk, dst);
7069 sk_rx_queue_set(req_to_sk(req), skb);
7071 tcp_reqsk_record_syn(sk, req, skb);
7072 fastopen_sk = tcp_try_fastopen(sk, skb, req, &foc, dst);
7075 af_ops->send_synack(fastopen_sk, dst, &fl, req,
7076 &foc, TCP_SYNACK_FASTOPEN, skb);
7077 /* Add the child socket directly into the accept queue */
7078 if (!inet_csk_reqsk_queue_add(sk, req, fastopen_sk)) {
7079 reqsk_fastopen_remove(fastopen_sk, req, false);
7080 bh_unlock_sock(fastopen_sk);
7081 sock_put(fastopen_sk);
7084 sk->sk_data_ready(sk);
7085 bh_unlock_sock(fastopen_sk);
7086 sock_put(fastopen_sk);
7088 tcp_rsk(req)->tfo_listener = false;
7090 req->timeout = tcp_timeout_init((struct sock *)req);
7091 inet_csk_reqsk_queue_hash_add(sk, req, req->timeout);
7093 af_ops->send_synack(sk, dst, &fl, req, &foc,
7094 !want_cookie ? TCP_SYNACK_NORMAL :
7113 EXPORT_SYMBOL(tcp_conn_request);