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).
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/proto_memory.h>
76 #include <net/inet_common.h>
77 #include <linux/ipsec.h>
78 #include <asm/unaligned.h>
79 #include <linux/errqueue.h>
80 #include <trace/events/tcp.h>
81 #include <linux/jump_label_ratelimit.h>
82 #include <net/busy_poll.h>
83 #include <net/mptcp.h>
85 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
87 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
88 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
89 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
90 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
91 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
92 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
93 #define FLAG_ECE 0x40 /* ECE in this ACK */
94 #define FLAG_LOST_RETRANS 0x80 /* This ACK marks some retransmission lost */
95 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
96 #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
97 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
98 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
99 #define FLAG_SET_XMIT_TIMER 0x1000 /* Set TLP or RTO timer */
100 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
101 #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
102 #define FLAG_NO_CHALLENGE_ACK 0x8000 /* do not call tcp_send_challenge_ack() */
103 #define FLAG_ACK_MAYBE_DELAYED 0x10000 /* Likely a delayed ACK */
104 #define FLAG_DSACK_TLP 0x20000 /* DSACK for tail loss probe */
106 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
107 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
108 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE|FLAG_DSACKING_ACK)
109 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
111 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
112 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
114 #define REXMIT_NONE 0 /* no loss recovery to do */
115 #define REXMIT_LOST 1 /* retransmit packets marked lost */
116 #define REXMIT_NEW 2 /* FRTO-style transmit of unsent/new packets */
118 #if IS_ENABLED(CONFIG_TLS_DEVICE)
119 static DEFINE_STATIC_KEY_DEFERRED_FALSE(clean_acked_data_enabled, HZ);
121 void clean_acked_data_enable(struct inet_connection_sock *icsk,
122 void (*cad)(struct sock *sk, u32 ack_seq))
124 icsk->icsk_clean_acked = cad;
125 static_branch_deferred_inc(&clean_acked_data_enabled);
127 EXPORT_SYMBOL_GPL(clean_acked_data_enable);
129 void clean_acked_data_disable(struct inet_connection_sock *icsk)
131 static_branch_slow_dec_deferred(&clean_acked_data_enabled);
132 icsk->icsk_clean_acked = NULL;
134 EXPORT_SYMBOL_GPL(clean_acked_data_disable);
136 void clean_acked_data_flush(void)
138 static_key_deferred_flush(&clean_acked_data_enabled);
140 EXPORT_SYMBOL_GPL(clean_acked_data_flush);
143 #ifdef CONFIG_CGROUP_BPF
144 static void bpf_skops_parse_hdr(struct sock *sk, struct sk_buff *skb)
146 bool unknown_opt = tcp_sk(sk)->rx_opt.saw_unknown &&
147 BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk),
148 BPF_SOCK_OPS_PARSE_UNKNOWN_HDR_OPT_CB_FLAG);
149 bool parse_all_opt = BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk),
150 BPF_SOCK_OPS_PARSE_ALL_HDR_OPT_CB_FLAG);
151 struct bpf_sock_ops_kern sock_ops;
153 if (likely(!unknown_opt && !parse_all_opt))
156 /* The skb will be handled in the
157 * bpf_skops_established() or
158 * bpf_skops_write_hdr_opt().
160 switch (sk->sk_state) {
167 sock_owned_by_me(sk);
169 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp));
170 sock_ops.op = BPF_SOCK_OPS_PARSE_HDR_OPT_CB;
171 sock_ops.is_fullsock = 1;
173 bpf_skops_init_skb(&sock_ops, skb, tcp_hdrlen(skb));
175 BPF_CGROUP_RUN_PROG_SOCK_OPS(&sock_ops);
178 static void bpf_skops_established(struct sock *sk, int bpf_op,
181 struct bpf_sock_ops_kern sock_ops;
183 sock_owned_by_me(sk);
185 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp));
186 sock_ops.op = bpf_op;
187 sock_ops.is_fullsock = 1;
189 /* sk with TCP_REPAIR_ON does not have skb in tcp_finish_connect */
191 bpf_skops_init_skb(&sock_ops, skb, tcp_hdrlen(skb));
193 BPF_CGROUP_RUN_PROG_SOCK_OPS(&sock_ops);
196 static void bpf_skops_parse_hdr(struct sock *sk, struct sk_buff *skb)
200 static void bpf_skops_established(struct sock *sk, int bpf_op,
206 static __cold void tcp_gro_dev_warn(const struct sock *sk, const struct sk_buff *skb,
209 struct net_device *dev;
212 dev = dev_get_by_index_rcu(sock_net(sk), skb->skb_iif);
213 if (!dev || len >= READ_ONCE(dev->mtu))
214 pr_warn("%s: Driver has suspect GRO implementation, TCP performance may be compromised.\n",
215 dev ? dev->name : "Unknown driver");
219 /* Adapt the MSS value used to make delayed ack decision to the
222 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
224 struct inet_connection_sock *icsk = inet_csk(sk);
225 const unsigned int lss = icsk->icsk_ack.last_seg_size;
228 icsk->icsk_ack.last_seg_size = 0;
230 /* skb->len may jitter because of SACKs, even if peer
231 * sends good full-sized frames.
233 len = skb_shinfo(skb)->gso_size ? : skb->len;
234 if (len >= icsk->icsk_ack.rcv_mss) {
235 /* Note: divides are still a bit expensive.
236 * For the moment, only adjust scaling_ratio
237 * when we update icsk_ack.rcv_mss.
239 if (unlikely(len != icsk->icsk_ack.rcv_mss)) {
240 u64 val = (u64)skb->len << TCP_RMEM_TO_WIN_SCALE;
241 u8 old_ratio = tcp_sk(sk)->scaling_ratio;
243 do_div(val, skb->truesize);
244 tcp_sk(sk)->scaling_ratio = val ? val : 1;
246 if (old_ratio != tcp_sk(sk)->scaling_ratio)
247 WRITE_ONCE(tcp_sk(sk)->window_clamp,
248 tcp_win_from_space(sk, sk->sk_rcvbuf));
250 icsk->icsk_ack.rcv_mss = min_t(unsigned int, len,
252 /* Account for possibly-removed options */
253 DO_ONCE_LITE_IF(len > icsk->icsk_ack.rcv_mss + MAX_TCP_OPTION_SPACE,
254 tcp_gro_dev_warn, sk, skb, len);
255 /* If the skb has a len of exactly 1*MSS and has the PSH bit
256 * set then it is likely the end of an application write. So
257 * more data may not be arriving soon, and yet the data sender
258 * may be waiting for an ACK if cwnd-bound or using TX zero
259 * copy. So we set ICSK_ACK_PUSHED here so that
260 * tcp_cleanup_rbuf() will send an ACK immediately if the app
261 * reads all of the data and is not ping-pong. If len > MSS
262 * then this logic does not matter (and does not hurt) because
263 * tcp_cleanup_rbuf() will always ACK immediately if the app
264 * reads data and there is more than an MSS of unACKed data.
266 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_PSH)
267 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
269 /* Otherwise, we make more careful check taking into account,
270 * that SACKs block is variable.
272 * "len" is invariant segment length, including TCP header.
274 len += skb->data - skb_transport_header(skb);
275 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
276 /* If PSH is not set, packet should be
277 * full sized, provided peer TCP is not badly broken.
278 * This observation (if it is correct 8)) allows
279 * to handle super-low mtu links fairly.
281 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
282 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
283 /* Subtract also invariant (if peer is RFC compliant),
284 * tcp header plus fixed timestamp option length.
285 * Resulting "len" is MSS free of SACK jitter.
287 len -= tcp_sk(sk)->tcp_header_len;
288 icsk->icsk_ack.last_seg_size = len;
290 icsk->icsk_ack.rcv_mss = len;
294 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
295 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
296 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
300 static void tcp_incr_quickack(struct sock *sk, unsigned int max_quickacks)
302 struct inet_connection_sock *icsk = inet_csk(sk);
303 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
307 quickacks = min(quickacks, max_quickacks);
308 if (quickacks > icsk->icsk_ack.quick)
309 icsk->icsk_ack.quick = quickacks;
312 static void tcp_enter_quickack_mode(struct sock *sk, unsigned int max_quickacks)
314 struct inet_connection_sock *icsk = inet_csk(sk);
316 tcp_incr_quickack(sk, max_quickacks);
317 inet_csk_exit_pingpong_mode(sk);
318 icsk->icsk_ack.ato = TCP_ATO_MIN;
321 /* Send ACKs quickly, if "quick" count is not exhausted
322 * and the session is not interactive.
325 static bool tcp_in_quickack_mode(struct sock *sk)
327 const struct inet_connection_sock *icsk = inet_csk(sk);
328 const struct dst_entry *dst = __sk_dst_get(sk);
330 return (dst && dst_metric(dst, RTAX_QUICKACK)) ||
331 (icsk->icsk_ack.quick && !inet_csk_in_pingpong_mode(sk));
334 static void tcp_ecn_queue_cwr(struct tcp_sock *tp)
336 if (tp->ecn_flags & TCP_ECN_OK)
337 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
340 static void tcp_ecn_accept_cwr(struct sock *sk, const struct sk_buff *skb)
342 if (tcp_hdr(skb)->cwr) {
343 tcp_sk(sk)->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
345 /* If the sender is telling us it has entered CWR, then its
346 * cwnd may be very low (even just 1 packet), so we should ACK
349 if (TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq)
350 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW;
354 static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp)
356 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
359 static void __tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb)
361 struct tcp_sock *tp = tcp_sk(sk);
363 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
364 case INET_ECN_NOT_ECT:
365 /* Funny extension: if ECT is not set on a segment,
366 * and we already seen ECT on a previous segment,
367 * it is probably a retransmit.
369 if (tp->ecn_flags & TCP_ECN_SEEN)
370 tcp_enter_quickack_mode(sk, 2);
373 if (tcp_ca_needs_ecn(sk))
374 tcp_ca_event(sk, CA_EVENT_ECN_IS_CE);
376 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
377 /* Better not delay acks, sender can have a very low cwnd */
378 tcp_enter_quickack_mode(sk, 2);
379 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
381 tp->ecn_flags |= TCP_ECN_SEEN;
384 if (tcp_ca_needs_ecn(sk))
385 tcp_ca_event(sk, CA_EVENT_ECN_NO_CE);
386 tp->ecn_flags |= TCP_ECN_SEEN;
391 static void tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb)
393 if (tcp_sk(sk)->ecn_flags & TCP_ECN_OK)
394 __tcp_ecn_check_ce(sk, skb);
397 static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
399 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
400 tp->ecn_flags &= ~TCP_ECN_OK;
403 static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
405 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
406 tp->ecn_flags &= ~TCP_ECN_OK;
409 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
411 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
416 /* Buffer size and advertised window tuning.
418 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
421 static void tcp_sndbuf_expand(struct sock *sk)
423 const struct tcp_sock *tp = tcp_sk(sk);
424 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
428 /* Worst case is non GSO/TSO : each frame consumes one skb
429 * and skb->head is kmalloced using power of two area of memory
431 per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
433 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
435 per_mss = roundup_pow_of_two(per_mss) +
436 SKB_DATA_ALIGN(sizeof(struct sk_buff));
438 nr_segs = max_t(u32, TCP_INIT_CWND, tcp_snd_cwnd(tp));
439 nr_segs = max_t(u32, nr_segs, tp->reordering + 1);
441 /* Fast Recovery (RFC 5681 3.2) :
442 * Cubic needs 1.7 factor, rounded to 2 to include
443 * extra cushion (application might react slowly to EPOLLOUT)
445 sndmem = ca_ops->sndbuf_expand ? ca_ops->sndbuf_expand(sk) : 2;
446 sndmem *= nr_segs * per_mss;
448 if (sk->sk_sndbuf < sndmem)
449 WRITE_ONCE(sk->sk_sndbuf,
450 min(sndmem, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_wmem[2])));
453 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
455 * All tcp_full_space() is split to two parts: "network" buffer, allocated
456 * forward and advertised in receiver window (tp->rcv_wnd) and
457 * "application buffer", required to isolate scheduling/application
458 * latencies from network.
459 * window_clamp is maximal advertised window. It can be less than
460 * tcp_full_space(), in this case tcp_full_space() - window_clamp
461 * is reserved for "application" buffer. The less window_clamp is
462 * the smoother our behaviour from viewpoint of network, but the lower
463 * throughput and the higher sensitivity of the connection to losses. 8)
465 * rcv_ssthresh is more strict window_clamp used at "slow start"
466 * phase to predict further behaviour of this connection.
467 * It is used for two goals:
468 * - to enforce header prediction at sender, even when application
469 * requires some significant "application buffer". It is check #1.
470 * - to prevent pruning of receive queue because of misprediction
471 * of receiver window. Check #2.
473 * The scheme does not work when sender sends good segments opening
474 * window and then starts to feed us spaghetti. But it should work
475 * in common situations. Otherwise, we have to rely on queue collapsing.
478 /* Slow part of check#2. */
479 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb,
480 unsigned int skbtruesize)
482 const struct tcp_sock *tp = tcp_sk(sk);
484 int truesize = tcp_win_from_space(sk, skbtruesize) >> 1;
485 int window = tcp_win_from_space(sk, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2])) >> 1;
487 while (tp->rcv_ssthresh <= window) {
488 if (truesize <= skb->len)
489 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
497 /* Even if skb appears to have a bad len/truesize ratio, TCP coalescing
498 * can play nice with us, as sk_buff and skb->head might be either
499 * freed or shared with up to MAX_SKB_FRAGS segments.
500 * Only give a boost to drivers using page frag(s) to hold the frame(s),
501 * and if no payload was pulled in skb->head before reaching us.
503 static u32 truesize_adjust(bool adjust, const struct sk_buff *skb)
505 u32 truesize = skb->truesize;
507 if (adjust && !skb_headlen(skb)) {
508 truesize -= SKB_TRUESIZE(skb_end_offset(skb));
509 /* paranoid check, some drivers might be buggy */
510 if (unlikely((int)truesize < (int)skb->len))
511 truesize = skb->truesize;
516 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb,
519 struct tcp_sock *tp = tcp_sk(sk);
522 room = min_t(int, tp->window_clamp, tcp_space(sk)) - tp->rcv_ssthresh;
528 if (!tcp_under_memory_pressure(sk)) {
529 unsigned int truesize = truesize_adjust(adjust, skb);
532 /* Check #2. Increase window, if skb with such overhead
533 * will fit to rcvbuf in future.
535 if (tcp_win_from_space(sk, truesize) <= skb->len)
536 incr = 2 * tp->advmss;
538 incr = __tcp_grow_window(sk, skb, truesize);
541 incr = max_t(int, incr, 2 * skb->len);
542 tp->rcv_ssthresh += min(room, incr);
543 inet_csk(sk)->icsk_ack.quick |= 1;
547 * Adjust rcv_ssthresh according to reserved mem
549 tcp_adjust_rcv_ssthresh(sk);
553 /* 3. Try to fixup all. It is made immediately after connection enters
556 static void tcp_init_buffer_space(struct sock *sk)
558 int tcp_app_win = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_app_win);
559 struct tcp_sock *tp = tcp_sk(sk);
562 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
563 tcp_sndbuf_expand(sk);
565 tcp_mstamp_refresh(tp);
566 tp->rcvq_space.time = tp->tcp_mstamp;
567 tp->rcvq_space.seq = tp->copied_seq;
569 maxwin = tcp_full_space(sk);
571 if (tp->window_clamp >= maxwin) {
572 WRITE_ONCE(tp->window_clamp, maxwin);
574 if (tcp_app_win && maxwin > 4 * tp->advmss)
575 WRITE_ONCE(tp->window_clamp,
576 max(maxwin - (maxwin >> tcp_app_win),
580 /* Force reservation of one segment. */
582 tp->window_clamp > 2 * tp->advmss &&
583 tp->window_clamp + tp->advmss > maxwin)
584 WRITE_ONCE(tp->window_clamp,
585 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 s32 tcp_rtt_tsopt_us(const struct tcp_sock *tp)
700 delta = tcp_time_stamp_ts(tp) - tp->rx_opt.rcv_tsecr;
704 if (likely(delta < INT_MAX / (USEC_PER_SEC / TCP_TS_HZ))) {
707 delta_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
713 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
714 const struct sk_buff *skb)
716 struct tcp_sock *tp = tcp_sk(sk);
718 if (tp->rx_opt.rcv_tsecr == tp->rcv_rtt_last_tsecr)
720 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr;
722 if (TCP_SKB_CB(skb)->end_seq -
723 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss) {
724 s32 delta = tcp_rtt_tsopt_us(tp);
727 tcp_rcv_rtt_update(tp, delta, 0);
732 * This function should be called every time data is copied to user space.
733 * It calculates the appropriate TCP receive buffer space.
735 void tcp_rcv_space_adjust(struct sock *sk)
737 struct tcp_sock *tp = tcp_sk(sk);
741 trace_tcp_rcv_space_adjust(sk);
743 tcp_mstamp_refresh(tp);
744 time = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcvq_space.time);
745 if (time < (tp->rcv_rtt_est.rtt_us >> 3) || tp->rcv_rtt_est.rtt_us == 0)
748 /* Number of bytes copied to user in last RTT */
749 copied = tp->copied_seq - tp->rcvq_space.seq;
750 if (copied <= tp->rcvq_space.space)
754 * copied = bytes received in previous RTT, our base window
755 * To cope with packet losses, we need a 2x factor
756 * To cope with slow start, and sender growing its cwin by 100 %
757 * every RTT, we need a 4x factor, because the ACK we are sending
758 * now is for the next RTT, not the current one :
759 * <prev RTT . ><current RTT .. ><next RTT .... >
762 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf) &&
763 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
767 /* minimal window to cope with packet losses, assuming
768 * steady state. Add some cushion because of small variations.
770 rcvwin = ((u64)copied << 1) + 16 * tp->advmss;
772 /* Accommodate for sender rate increase (eg. slow start) */
773 grow = rcvwin * (copied - tp->rcvq_space.space);
774 do_div(grow, tp->rcvq_space.space);
775 rcvwin += (grow << 1);
777 rcvbuf = min_t(u64, tcp_space_from_win(sk, rcvwin),
778 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]));
779 if (rcvbuf > sk->sk_rcvbuf) {
780 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
782 /* Make the window clamp follow along. */
783 WRITE_ONCE(tp->window_clamp,
784 tcp_win_from_space(sk, rcvbuf));
787 tp->rcvq_space.space = copied;
790 tp->rcvq_space.seq = tp->copied_seq;
791 tp->rcvq_space.time = tp->tcp_mstamp;
794 static void tcp_save_lrcv_flowlabel(struct sock *sk, const struct sk_buff *skb)
796 #if IS_ENABLED(CONFIG_IPV6)
797 struct inet_connection_sock *icsk = inet_csk(sk);
799 if (skb->protocol == htons(ETH_P_IPV6))
800 icsk->icsk_ack.lrcv_flowlabel = ntohl(ip6_flowlabel(ipv6_hdr(skb)));
804 /* There is something which you must keep in mind when you analyze the
805 * behavior of the tp->ato delayed ack timeout interval. When a
806 * connection starts up, we want to ack as quickly as possible. The
807 * problem is that "good" TCP's do slow start at the beginning of data
808 * transmission. The means that until we send the first few ACK's the
809 * sender will sit on his end and only queue most of his data, because
810 * he can only send snd_cwnd unacked packets at any given time. For
811 * each ACK we send, he increments snd_cwnd and transmits more of his
814 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
816 struct tcp_sock *tp = tcp_sk(sk);
817 struct inet_connection_sock *icsk = inet_csk(sk);
820 inet_csk_schedule_ack(sk);
822 tcp_measure_rcv_mss(sk, skb);
824 tcp_rcv_rtt_measure(tp);
828 if (!icsk->icsk_ack.ato) {
829 /* The _first_ data packet received, initialize
830 * delayed ACK engine.
832 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
833 icsk->icsk_ack.ato = TCP_ATO_MIN;
835 int m = now - icsk->icsk_ack.lrcvtime;
837 if (m <= TCP_ATO_MIN / 2) {
838 /* The fastest case is the first. */
839 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
840 } else if (m < icsk->icsk_ack.ato) {
841 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
842 if (icsk->icsk_ack.ato > icsk->icsk_rto)
843 icsk->icsk_ack.ato = icsk->icsk_rto;
844 } else if (m > icsk->icsk_rto) {
845 /* Too long gap. Apparently sender failed to
846 * restart window, so that we send ACKs quickly.
848 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
851 icsk->icsk_ack.lrcvtime = now;
852 tcp_save_lrcv_flowlabel(sk, skb);
854 tcp_ecn_check_ce(sk, skb);
857 tcp_grow_window(sk, skb, true);
860 /* Called to compute a smoothed rtt estimate. The data fed to this
861 * routine either comes from timestamps, or from segments that were
862 * known _not_ to have been retransmitted [see Karn/Partridge
863 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
864 * piece by Van Jacobson.
865 * NOTE: the next three routines used to be one big routine.
866 * To save cycles in the RFC 1323 implementation it was better to break
867 * it up into three procedures. -- erics
869 static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
871 struct tcp_sock *tp = tcp_sk(sk);
872 long m = mrtt_us; /* RTT */
873 u32 srtt = tp->srtt_us;
875 /* The following amusing code comes from Jacobson's
876 * article in SIGCOMM '88. Note that rtt and mdev
877 * are scaled versions of rtt and mean deviation.
878 * This is designed to be as fast as possible
879 * m stands for "measurement".
881 * On a 1990 paper the rto value is changed to:
882 * RTO = rtt + 4 * mdev
884 * Funny. This algorithm seems to be very broken.
885 * These formulae increase RTO, when it should be decreased, increase
886 * too slowly, when it should be increased quickly, decrease too quickly
887 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
888 * does not matter how to _calculate_ it. Seems, it was trap
889 * that VJ failed to avoid. 8)
892 m -= (srtt >> 3); /* m is now error in rtt est */
893 srtt += m; /* rtt = 7/8 rtt + 1/8 new */
895 m = -m; /* m is now abs(error) */
896 m -= (tp->mdev_us >> 2); /* similar update on mdev */
897 /* This is similar to one of Eifel findings.
898 * Eifel blocks mdev updates when rtt decreases.
899 * This solution is a bit different: we use finer gain
900 * for mdev in this case (alpha*beta).
901 * Like Eifel it also prevents growth of rto,
902 * but also it limits too fast rto decreases,
903 * happening in pure Eifel.
908 m -= (tp->mdev_us >> 2); /* similar update on mdev */
910 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */
911 if (tp->mdev_us > tp->mdev_max_us) {
912 tp->mdev_max_us = tp->mdev_us;
913 if (tp->mdev_max_us > tp->rttvar_us)
914 tp->rttvar_us = tp->mdev_max_us;
916 if (after(tp->snd_una, tp->rtt_seq)) {
917 if (tp->mdev_max_us < tp->rttvar_us)
918 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
919 tp->rtt_seq = tp->snd_nxt;
920 tp->mdev_max_us = tcp_rto_min_us(sk);
922 tcp_bpf_rtt(sk, mrtt_us, srtt);
925 /* no previous measure. */
926 srtt = m << 3; /* take the measured time to be rtt */
927 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */
928 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
929 tp->mdev_max_us = tp->rttvar_us;
930 tp->rtt_seq = tp->snd_nxt;
932 tcp_bpf_rtt(sk, mrtt_us, srtt);
934 tp->srtt_us = max(1U, srtt);
937 static void tcp_update_pacing_rate(struct sock *sk)
939 const struct tcp_sock *tp = tcp_sk(sk);
942 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
943 rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3);
945 /* current rate is (cwnd * mss) / srtt
946 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
947 * In Congestion Avoidance phase, set it to 120 % the current rate.
949 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
950 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
951 * end of slow start and should slow down.
953 if (tcp_snd_cwnd(tp) < tp->snd_ssthresh / 2)
954 rate *= READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_pacing_ss_ratio);
956 rate *= READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_pacing_ca_ratio);
958 rate *= max(tcp_snd_cwnd(tp), tp->packets_out);
960 if (likely(tp->srtt_us))
961 do_div(rate, tp->srtt_us);
963 /* WRITE_ONCE() is needed because sch_fq fetches sk_pacing_rate
964 * without any lock. We want to make sure compiler wont store
965 * intermediate values in this location.
967 WRITE_ONCE(sk->sk_pacing_rate,
968 min_t(u64, rate, READ_ONCE(sk->sk_max_pacing_rate)));
971 /* Calculate rto without backoff. This is the second half of Van Jacobson's
972 * routine referred to above.
974 static void tcp_set_rto(struct sock *sk)
976 const struct tcp_sock *tp = tcp_sk(sk);
977 /* Old crap is replaced with new one. 8)
980 * 1. If rtt variance happened to be less 50msec, it is hallucination.
981 * It cannot be less due to utterly erratic ACK generation made
982 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
983 * to do with delayed acks, because at cwnd>2 true delack timeout
984 * is invisible. Actually, Linux-2.4 also generates erratic
985 * ACKs in some circumstances.
987 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
989 /* 2. Fixups made earlier cannot be right.
990 * If we do not estimate RTO correctly without them,
991 * all the algo is pure shit and should be replaced
992 * with correct one. It is exactly, which we pretend to do.
995 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
996 * guarantees that rto is higher.
1001 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
1003 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
1006 cwnd = TCP_INIT_CWND;
1007 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
1010 struct tcp_sacktag_state {
1011 /* Timestamps for earliest and latest never-retransmitted segment
1012 * that was SACKed. RTO needs the earliest RTT to stay conservative,
1013 * but congestion control should still get an accurate delay signal.
1020 unsigned int mss_now;
1021 struct rate_sample *rate;
1024 /* Take a notice that peer is sending D-SACKs. Skip update of data delivery
1025 * and spurious retransmission information if this DSACK is unlikely caused by
1027 * - DSACKed sequence range is larger than maximum receiver's window.
1028 * - Total no. of DSACKed segments exceed the total no. of retransmitted segs.
1030 static u32 tcp_dsack_seen(struct tcp_sock *tp, u32 start_seq,
1031 u32 end_seq, struct tcp_sacktag_state *state)
1033 u32 seq_len, dup_segs = 1;
1035 if (!before(start_seq, end_seq))
1038 seq_len = end_seq - start_seq;
1039 /* Dubious DSACK: DSACKed range greater than maximum advertised rwnd */
1040 if (seq_len > tp->max_window)
1042 if (seq_len > tp->mss_cache)
1043 dup_segs = DIV_ROUND_UP(seq_len, tp->mss_cache);
1044 else if (tp->tlp_high_seq && tp->tlp_high_seq == end_seq)
1045 state->flag |= FLAG_DSACK_TLP;
1047 tp->dsack_dups += dup_segs;
1048 /* Skip the DSACK if dup segs weren't retransmitted by sender */
1049 if (tp->dsack_dups > tp->total_retrans)
1052 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
1053 /* We increase the RACK ordering window in rounds where we receive
1054 * DSACKs that may have been due to reordering causing RACK to trigger
1055 * a spurious fast recovery. Thus RACK ignores DSACKs that happen
1056 * without having seen reordering, or that match TLP probes (TLP
1057 * is timer-driven, not triggered by RACK).
1059 if (tp->reord_seen && !(state->flag & FLAG_DSACK_TLP))
1060 tp->rack.dsack_seen = 1;
1062 state->flag |= FLAG_DSACKING_ACK;
1063 /* A spurious retransmission is delivered */
1064 state->sack_delivered += dup_segs;
1069 /* It's reordering when higher sequence was delivered (i.e. sacked) before
1070 * some lower never-retransmitted sequence ("low_seq"). The maximum reordering
1071 * distance is approximated in full-mss packet distance ("reordering").
1073 static void tcp_check_sack_reordering(struct sock *sk, const u32 low_seq,
1076 struct tcp_sock *tp = tcp_sk(sk);
1077 const u32 mss = tp->mss_cache;
1080 fack = tcp_highest_sack_seq(tp);
1081 if (!before(low_seq, fack))
1084 metric = fack - low_seq;
1085 if ((metric > tp->reordering * mss) && mss) {
1086 #if FASTRETRANS_DEBUG > 1
1087 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
1088 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
1092 tp->undo_marker ? tp->undo_retrans : 0);
1094 tp->reordering = min_t(u32, (metric + mss - 1) / mss,
1095 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_max_reordering));
1098 /* This exciting event is worth to be remembered. 8) */
1100 NET_INC_STATS(sock_net(sk),
1101 ts ? LINUX_MIB_TCPTSREORDER : LINUX_MIB_TCPSACKREORDER);
1104 /* This must be called before lost_out or retrans_out are updated
1105 * on a new loss, because we want to know if all skbs previously
1106 * known to be lost have already been retransmitted, indicating
1107 * that this newly lost skb is our next skb to retransmit.
1109 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
1111 if ((!tp->retransmit_skb_hint && tp->retrans_out >= tp->lost_out) ||
1112 (tp->retransmit_skb_hint &&
1113 before(TCP_SKB_CB(skb)->seq,
1114 TCP_SKB_CB(tp->retransmit_skb_hint)->seq)))
1115 tp->retransmit_skb_hint = skb;
1118 /* Sum the number of packets on the wire we have marked as lost, and
1119 * notify the congestion control module that the given skb was marked lost.
1121 static void tcp_notify_skb_loss_event(struct tcp_sock *tp, const struct sk_buff *skb)
1123 tp->lost += tcp_skb_pcount(skb);
1126 void tcp_mark_skb_lost(struct sock *sk, struct sk_buff *skb)
1128 __u8 sacked = TCP_SKB_CB(skb)->sacked;
1129 struct tcp_sock *tp = tcp_sk(sk);
1131 if (sacked & TCPCB_SACKED_ACKED)
1134 tcp_verify_retransmit_hint(tp, skb);
1135 if (sacked & TCPCB_LOST) {
1136 if (sacked & TCPCB_SACKED_RETRANS) {
1137 /* Account for retransmits that are lost again */
1138 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1139 tp->retrans_out -= tcp_skb_pcount(skb);
1140 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT,
1141 tcp_skb_pcount(skb));
1142 tcp_notify_skb_loss_event(tp, skb);
1145 tp->lost_out += tcp_skb_pcount(skb);
1146 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1147 tcp_notify_skb_loss_event(tp, skb);
1151 /* Updates the delivered and delivered_ce counts */
1152 static void tcp_count_delivered(struct tcp_sock *tp, u32 delivered,
1155 tp->delivered += delivered;
1157 tp->delivered_ce += delivered;
1160 /* This procedure tags the retransmission queue when SACKs arrive.
1162 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
1163 * Packets in queue with these bits set are counted in variables
1164 * sacked_out, retrans_out and lost_out, correspondingly.
1166 * Valid combinations are:
1167 * Tag InFlight Description
1168 * 0 1 - orig segment is in flight.
1169 * S 0 - nothing flies, orig reached receiver.
1170 * L 0 - nothing flies, orig lost by net.
1171 * R 2 - both orig and retransmit are in flight.
1172 * L|R 1 - orig is lost, retransmit is in flight.
1173 * S|R 1 - orig reached receiver, retrans is still in flight.
1174 * (L|S|R is logically valid, it could occur when L|R is sacked,
1175 * but it is equivalent to plain S and code short-circuits it to S.
1176 * L|S is logically invalid, it would mean -1 packet in flight 8))
1178 * These 6 states form finite state machine, controlled by the following events:
1179 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
1180 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
1181 * 3. Loss detection event of two flavors:
1182 * A. Scoreboard estimator decided the packet is lost.
1183 * A'. Reno "three dupacks" marks head of queue lost.
1184 * B. SACK arrives sacking SND.NXT at the moment, when the
1185 * segment was retransmitted.
1186 * 4. D-SACK added new rule: D-SACK changes any tag to S.
1188 * It is pleasant to note, that state diagram turns out to be commutative,
1189 * so that we are allowed not to be bothered by order of our actions,
1190 * when multiple events arrive simultaneously. (see the function below).
1192 * Reordering detection.
1193 * --------------------
1194 * Reordering metric is maximal distance, which a packet can be displaced
1195 * in packet stream. With SACKs we can estimate it:
1197 * 1. SACK fills old hole and the corresponding segment was not
1198 * ever retransmitted -> reordering. Alas, we cannot use it
1199 * when segment was retransmitted.
1200 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1201 * for retransmitted and already SACKed segment -> reordering..
1202 * Both of these heuristics are not used in Loss state, when we cannot
1203 * account for retransmits accurately.
1205 * SACK block validation.
1206 * ----------------------
1208 * SACK block range validation checks that the received SACK block fits to
1209 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1210 * Note that SND.UNA is not included to the range though being valid because
1211 * it means that the receiver is rather inconsistent with itself reporting
1212 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1213 * perfectly valid, however, in light of RFC2018 which explicitly states
1214 * that "SACK block MUST reflect the newest segment. Even if the newest
1215 * segment is going to be discarded ...", not that it looks very clever
1216 * in case of head skb. Due to potentional receiver driven attacks, we
1217 * choose to avoid immediate execution of a walk in write queue due to
1218 * reneging and defer head skb's loss recovery to standard loss recovery
1219 * procedure that will eventually trigger (nothing forbids us doing this).
1221 * Implements also blockage to start_seq wrap-around. Problem lies in the
1222 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1223 * there's no guarantee that it will be before snd_nxt (n). The problem
1224 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1227 * <- outs wnd -> <- wrapzone ->
1228 * u e n u_w e_w s n_w
1230 * |<------------+------+----- TCP seqno space --------------+---------->|
1231 * ...-- <2^31 ->| |<--------...
1232 * ...---- >2^31 ------>| |<--------...
1234 * Current code wouldn't be vulnerable but it's better still to discard such
1235 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1236 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1237 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1238 * equal to the ideal case (infinite seqno space without wrap caused issues).
1240 * With D-SACK the lower bound is extended to cover sequence space below
1241 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1242 * again, D-SACK block must not to go across snd_una (for the same reason as
1243 * for the normal SACK blocks, explained above). But there all simplicity
1244 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1245 * fully below undo_marker they do not affect behavior in anyway and can
1246 * therefore be safely ignored. In rare cases (which are more or less
1247 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1248 * fragmentation and packet reordering past skb's retransmission. To consider
1249 * them correctly, the acceptable range must be extended even more though
1250 * the exact amount is rather hard to quantify. However, tp->max_window can
1251 * be used as an exaggerated estimate.
1253 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
1254 u32 start_seq, u32 end_seq)
1256 /* Too far in future, or reversed (interpretation is ambiguous) */
1257 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1260 /* Nasty start_seq wrap-around check (see comments above) */
1261 if (!before(start_seq, tp->snd_nxt))
1264 /* In outstanding window? ...This is valid exit for D-SACKs too.
1265 * start_seq == snd_una is non-sensical (see comments above)
1267 if (after(start_seq, tp->snd_una))
1270 if (!is_dsack || !tp->undo_marker)
1273 /* ...Then it's D-SACK, and must reside below snd_una completely */
1274 if (after(end_seq, tp->snd_una))
1277 if (!before(start_seq, tp->undo_marker))
1281 if (!after(end_seq, tp->undo_marker))
1284 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1285 * start_seq < undo_marker and end_seq >= undo_marker.
1287 return !before(start_seq, end_seq - tp->max_window);
1290 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1291 struct tcp_sack_block_wire *sp, int num_sacks,
1292 u32 prior_snd_una, struct tcp_sacktag_state *state)
1294 struct tcp_sock *tp = tcp_sk(sk);
1295 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1296 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1299 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1300 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1301 } else if (num_sacks > 1) {
1302 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1303 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1305 if (after(end_seq_0, end_seq_1) || before(start_seq_0, start_seq_1))
1307 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKOFORECV);
1312 dup_segs = tcp_dsack_seen(tp, start_seq_0, end_seq_0, state);
1313 if (!dup_segs) { /* Skip dubious DSACK */
1314 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKIGNOREDDUBIOUS);
1318 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECVSEGS, dup_segs);
1320 /* D-SACK for already forgotten data... Do dumb counting. */
1321 if (tp->undo_marker && tp->undo_retrans > 0 &&
1322 !after(end_seq_0, prior_snd_una) &&
1323 after(end_seq_0, tp->undo_marker))
1324 tp->undo_retrans = max_t(int, 0, tp->undo_retrans - dup_segs);
1329 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1330 * the incoming SACK may not exactly match but we can find smaller MSS
1331 * aligned portion of it that matches. Therefore we might need to fragment
1332 * which may fail and creates some hassle (caller must handle error case
1335 * FIXME: this could be merged to shift decision code
1337 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1338 u32 start_seq, u32 end_seq)
1342 unsigned int pkt_len;
1345 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1346 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1348 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1349 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1350 mss = tcp_skb_mss(skb);
1351 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1354 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1358 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1363 /* Round if necessary so that SACKs cover only full MSSes
1364 * and/or the remaining small portion (if present)
1366 if (pkt_len > mss) {
1367 unsigned int new_len = (pkt_len / mss) * mss;
1368 if (!in_sack && new_len < pkt_len)
1373 if (pkt_len >= skb->len && !in_sack)
1376 err = tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
1377 pkt_len, mss, GFP_ATOMIC);
1385 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1386 static u8 tcp_sacktag_one(struct sock *sk,
1387 struct tcp_sacktag_state *state, u8 sacked,
1388 u32 start_seq, u32 end_seq,
1389 int dup_sack, int pcount,
1392 struct tcp_sock *tp = tcp_sk(sk);
1394 /* Account D-SACK for retransmitted packet. */
1395 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1396 if (tp->undo_marker && tp->undo_retrans > 0 &&
1397 after(end_seq, tp->undo_marker))
1398 tp->undo_retrans = max_t(int, 0, tp->undo_retrans - pcount);
1399 if ((sacked & TCPCB_SACKED_ACKED) &&
1400 before(start_seq, state->reord))
1401 state->reord = start_seq;
1404 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1405 if (!after(end_seq, tp->snd_una))
1408 if (!(sacked & TCPCB_SACKED_ACKED)) {
1409 tcp_rack_advance(tp, sacked, end_seq, xmit_time);
1411 if (sacked & TCPCB_SACKED_RETRANS) {
1412 /* If the segment is not tagged as lost,
1413 * we do not clear RETRANS, believing
1414 * that retransmission is still in flight.
1416 if (sacked & TCPCB_LOST) {
1417 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1418 tp->lost_out -= pcount;
1419 tp->retrans_out -= pcount;
1422 if (!(sacked & TCPCB_RETRANS)) {
1423 /* New sack for not retransmitted frame,
1424 * which was in hole. It is reordering.
1426 if (before(start_seq,
1427 tcp_highest_sack_seq(tp)) &&
1428 before(start_seq, state->reord))
1429 state->reord = start_seq;
1431 if (!after(end_seq, tp->high_seq))
1432 state->flag |= FLAG_ORIG_SACK_ACKED;
1433 if (state->first_sackt == 0)
1434 state->first_sackt = xmit_time;
1435 state->last_sackt = xmit_time;
1438 if (sacked & TCPCB_LOST) {
1439 sacked &= ~TCPCB_LOST;
1440 tp->lost_out -= pcount;
1444 sacked |= TCPCB_SACKED_ACKED;
1445 state->flag |= FLAG_DATA_SACKED;
1446 tp->sacked_out += pcount;
1447 /* Out-of-order packets delivered */
1448 state->sack_delivered += pcount;
1450 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1451 if (tp->lost_skb_hint &&
1452 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1453 tp->lost_cnt_hint += pcount;
1456 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1457 * frames and clear it. undo_retrans is decreased above, L|R frames
1458 * are accounted above as well.
1460 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1461 sacked &= ~TCPCB_SACKED_RETRANS;
1462 tp->retrans_out -= pcount;
1468 /* Shift newly-SACKed bytes from this skb to the immediately previous
1469 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1471 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *prev,
1472 struct sk_buff *skb,
1473 struct tcp_sacktag_state *state,
1474 unsigned int pcount, int shifted, int mss,
1477 struct tcp_sock *tp = tcp_sk(sk);
1478 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1479 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
1483 /* Adjust counters and hints for the newly sacked sequence
1484 * range but discard the return value since prev is already
1485 * marked. We must tag the range first because the seq
1486 * advancement below implicitly advances
1487 * tcp_highest_sack_seq() when skb is highest_sack.
1489 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1490 start_seq, end_seq, dup_sack, pcount,
1491 tcp_skb_timestamp_us(skb));
1492 tcp_rate_skb_delivered(sk, skb, state->rate);
1494 if (skb == tp->lost_skb_hint)
1495 tp->lost_cnt_hint += pcount;
1497 TCP_SKB_CB(prev)->end_seq += shifted;
1498 TCP_SKB_CB(skb)->seq += shifted;
1500 tcp_skb_pcount_add(prev, pcount);
1501 WARN_ON_ONCE(tcp_skb_pcount(skb) < pcount);
1502 tcp_skb_pcount_add(skb, -pcount);
1504 /* When we're adding to gso_segs == 1, gso_size will be zero,
1505 * in theory this shouldn't be necessary but as long as DSACK
1506 * code can come after this skb later on it's better to keep
1507 * setting gso_size to something.
1509 if (!TCP_SKB_CB(prev)->tcp_gso_size)
1510 TCP_SKB_CB(prev)->tcp_gso_size = mss;
1512 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1513 if (tcp_skb_pcount(skb) <= 1)
1514 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1516 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1517 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1520 BUG_ON(!tcp_skb_pcount(skb));
1521 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1525 /* Whole SKB was eaten :-) */
1527 if (skb == tp->retransmit_skb_hint)
1528 tp->retransmit_skb_hint = prev;
1529 if (skb == tp->lost_skb_hint) {
1530 tp->lost_skb_hint = prev;
1531 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1534 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1535 TCP_SKB_CB(prev)->eor = TCP_SKB_CB(skb)->eor;
1536 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1537 TCP_SKB_CB(prev)->end_seq++;
1539 if (skb == tcp_highest_sack(sk))
1540 tcp_advance_highest_sack(sk, skb);
1542 tcp_skb_collapse_tstamp(prev, skb);
1543 if (unlikely(TCP_SKB_CB(prev)->tx.delivered_mstamp))
1544 TCP_SKB_CB(prev)->tx.delivered_mstamp = 0;
1546 tcp_rtx_queue_unlink_and_free(skb, sk);
1548 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKMERGED);
1553 /* I wish gso_size would have a bit more sane initialization than
1554 * something-or-zero which complicates things
1556 static int tcp_skb_seglen(const struct sk_buff *skb)
1558 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1561 /* Shifting pages past head area doesn't work */
1562 static int skb_can_shift(const struct sk_buff *skb)
1564 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1567 int tcp_skb_shift(struct sk_buff *to, struct sk_buff *from,
1568 int pcount, int shiftlen)
1570 /* TCP min gso_size is 8 bytes (TCP_MIN_GSO_SIZE)
1571 * Since TCP_SKB_CB(skb)->tcp_gso_segs is 16 bits, we need
1572 * to make sure not storing more than 65535 * 8 bytes per skb,
1573 * even if current MSS is bigger.
1575 if (unlikely(to->len + shiftlen >= 65535 * TCP_MIN_GSO_SIZE))
1577 if (unlikely(tcp_skb_pcount(to) + pcount > 65535))
1579 return skb_shift(to, from, shiftlen);
1582 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1585 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1586 struct tcp_sacktag_state *state,
1587 u32 start_seq, u32 end_seq,
1590 struct tcp_sock *tp = tcp_sk(sk);
1591 struct sk_buff *prev;
1597 /* Normally R but no L won't result in plain S */
1599 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1601 if (!skb_can_shift(skb))
1603 /* This frame is about to be dropped (was ACKed). */
1604 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1607 /* Can only happen with delayed DSACK + discard craziness */
1608 prev = skb_rb_prev(skb);
1612 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1615 if (!tcp_skb_can_collapse(prev, skb))
1618 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1619 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1623 pcount = tcp_skb_pcount(skb);
1624 mss = tcp_skb_seglen(skb);
1626 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1627 * drop this restriction as unnecessary
1629 if (mss != tcp_skb_seglen(prev))
1632 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1634 /* CHECKME: This is non-MSS split case only?, this will
1635 * cause skipped skbs due to advancing loop btw, original
1636 * has that feature too
1638 if (tcp_skb_pcount(skb) <= 1)
1641 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1643 /* TODO: head merge to next could be attempted here
1644 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1645 * though it might not be worth of the additional hassle
1647 * ...we can probably just fallback to what was done
1648 * previously. We could try merging non-SACKed ones
1649 * as well but it probably isn't going to buy off
1650 * because later SACKs might again split them, and
1651 * it would make skb timestamp tracking considerably
1657 len = end_seq - TCP_SKB_CB(skb)->seq;
1659 BUG_ON(len > skb->len);
1661 /* MSS boundaries should be honoured or else pcount will
1662 * severely break even though it makes things bit trickier.
1663 * Optimize common case to avoid most of the divides
1665 mss = tcp_skb_mss(skb);
1667 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1668 * drop this restriction as unnecessary
1670 if (mss != tcp_skb_seglen(prev))
1675 } else if (len < mss) {
1683 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1684 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1687 if (!tcp_skb_shift(prev, skb, pcount, len))
1689 if (!tcp_shifted_skb(sk, prev, skb, state, pcount, len, mss, dup_sack))
1692 /* Hole filled allows collapsing with the next as well, this is very
1693 * useful when hole on every nth skb pattern happens
1695 skb = skb_rb_next(prev);
1699 if (!skb_can_shift(skb) ||
1700 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1701 (mss != tcp_skb_seglen(skb)))
1704 if (!tcp_skb_can_collapse(prev, skb))
1707 pcount = tcp_skb_pcount(skb);
1708 if (tcp_skb_shift(prev, skb, pcount, len))
1709 tcp_shifted_skb(sk, prev, skb, state, pcount,
1719 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1723 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1724 struct tcp_sack_block *next_dup,
1725 struct tcp_sacktag_state *state,
1726 u32 start_seq, u32 end_seq,
1729 struct tcp_sock *tp = tcp_sk(sk);
1730 struct sk_buff *tmp;
1732 skb_rbtree_walk_from(skb) {
1734 bool dup_sack = dup_sack_in;
1736 /* queue is in-order => we can short-circuit the walk early */
1737 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1741 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1742 in_sack = tcp_match_skb_to_sack(sk, skb,
1743 next_dup->start_seq,
1749 /* skb reference here is a bit tricky to get right, since
1750 * shifting can eat and free both this skb and the next,
1751 * so not even _safe variant of the loop is enough.
1754 tmp = tcp_shift_skb_data(sk, skb, state,
1755 start_seq, end_seq, dup_sack);
1764 in_sack = tcp_match_skb_to_sack(sk, skb,
1770 if (unlikely(in_sack < 0))
1774 TCP_SKB_CB(skb)->sacked =
1777 TCP_SKB_CB(skb)->sacked,
1778 TCP_SKB_CB(skb)->seq,
1779 TCP_SKB_CB(skb)->end_seq,
1781 tcp_skb_pcount(skb),
1782 tcp_skb_timestamp_us(skb));
1783 tcp_rate_skb_delivered(sk, skb, state->rate);
1784 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1785 list_del_init(&skb->tcp_tsorted_anchor);
1787 if (!before(TCP_SKB_CB(skb)->seq,
1788 tcp_highest_sack_seq(tp)))
1789 tcp_advance_highest_sack(sk, skb);
1795 static struct sk_buff *tcp_sacktag_bsearch(struct sock *sk, u32 seq)
1797 struct rb_node *parent, **p = &sk->tcp_rtx_queue.rb_node;
1798 struct sk_buff *skb;
1802 skb = rb_to_skb(parent);
1803 if (before(seq, TCP_SKB_CB(skb)->seq)) {
1804 p = &parent->rb_left;
1807 if (!before(seq, TCP_SKB_CB(skb)->end_seq)) {
1808 p = &parent->rb_right;
1816 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1819 if (skb && after(TCP_SKB_CB(skb)->seq, skip_to_seq))
1822 return tcp_sacktag_bsearch(sk, skip_to_seq);
1825 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1827 struct tcp_sack_block *next_dup,
1828 struct tcp_sacktag_state *state,
1834 if (before(next_dup->start_seq, skip_to_seq)) {
1835 skb = tcp_sacktag_skip(skb, sk, next_dup->start_seq);
1836 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1837 next_dup->start_seq, next_dup->end_seq,
1844 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1846 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1850 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1851 u32 prior_snd_una, struct tcp_sacktag_state *state)
1853 struct tcp_sock *tp = tcp_sk(sk);
1854 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1855 TCP_SKB_CB(ack_skb)->sacked);
1856 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1857 struct tcp_sack_block sp[TCP_NUM_SACKS];
1858 struct tcp_sack_block *cache;
1859 struct sk_buff *skb;
1860 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1862 bool found_dup_sack = false;
1864 int first_sack_index;
1867 state->reord = tp->snd_nxt;
1869 if (!tp->sacked_out)
1870 tcp_highest_sack_reset(sk);
1872 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1873 num_sacks, prior_snd_una, state);
1875 /* Eliminate too old ACKs, but take into
1876 * account more or less fresh ones, they can
1877 * contain valid SACK info.
1879 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1882 if (!tp->packets_out)
1886 first_sack_index = 0;
1887 for (i = 0; i < num_sacks; i++) {
1888 bool dup_sack = !i && found_dup_sack;
1890 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1891 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1893 if (!tcp_is_sackblock_valid(tp, dup_sack,
1894 sp[used_sacks].start_seq,
1895 sp[used_sacks].end_seq)) {
1899 if (!tp->undo_marker)
1900 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1902 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1904 /* Don't count olds caused by ACK reordering */
1905 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1906 !after(sp[used_sacks].end_seq, tp->snd_una))
1908 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1911 NET_INC_STATS(sock_net(sk), mib_idx);
1913 first_sack_index = -1;
1917 /* Ignore very old stuff early */
1918 if (!after(sp[used_sacks].end_seq, prior_snd_una)) {
1920 first_sack_index = -1;
1927 /* order SACK blocks to allow in order walk of the retrans queue */
1928 for (i = used_sacks - 1; i > 0; i--) {
1929 for (j = 0; j < i; j++) {
1930 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1931 swap(sp[j], sp[j + 1]);
1933 /* Track where the first SACK block goes to */
1934 if (j == first_sack_index)
1935 first_sack_index = j + 1;
1940 state->mss_now = tcp_current_mss(sk);
1944 if (!tp->sacked_out) {
1945 /* It's already past, so skip checking against it */
1946 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1948 cache = tp->recv_sack_cache;
1949 /* Skip empty blocks in at head of the cache */
1950 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1955 while (i < used_sacks) {
1956 u32 start_seq = sp[i].start_seq;
1957 u32 end_seq = sp[i].end_seq;
1958 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1959 struct tcp_sack_block *next_dup = NULL;
1961 if (found_dup_sack && ((i + 1) == first_sack_index))
1962 next_dup = &sp[i + 1];
1964 /* Skip too early cached blocks */
1965 while (tcp_sack_cache_ok(tp, cache) &&
1966 !before(start_seq, cache->end_seq))
1969 /* Can skip some work by looking recv_sack_cache? */
1970 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1971 after(end_seq, cache->start_seq)) {
1974 if (before(start_seq, cache->start_seq)) {
1975 skb = tcp_sacktag_skip(skb, sk, start_seq);
1976 skb = tcp_sacktag_walk(skb, sk, next_dup,
1983 /* Rest of the block already fully processed? */
1984 if (!after(end_seq, cache->end_seq))
1987 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1991 /* ...tail remains todo... */
1992 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1993 /* ...but better entrypoint exists! */
1994 skb = tcp_highest_sack(sk);
2001 skb = tcp_sacktag_skip(skb, sk, cache->end_seq);
2002 /* Check overlap against next cached too (past this one already) */
2007 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
2008 skb = tcp_highest_sack(sk);
2012 skb = tcp_sacktag_skip(skb, sk, start_seq);
2015 skb = tcp_sacktag_walk(skb, sk, next_dup, state,
2016 start_seq, end_seq, dup_sack);
2022 /* Clear the head of the cache sack blocks so we can skip it next time */
2023 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
2024 tp->recv_sack_cache[i].start_seq = 0;
2025 tp->recv_sack_cache[i].end_seq = 0;
2027 for (j = 0; j < used_sacks; j++)
2028 tp->recv_sack_cache[i++] = sp[j];
2030 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Loss || tp->undo_marker)
2031 tcp_check_sack_reordering(sk, state->reord, 0);
2033 tcp_verify_left_out(tp);
2036 #if FASTRETRANS_DEBUG > 0
2037 WARN_ON((int)tp->sacked_out < 0);
2038 WARN_ON((int)tp->lost_out < 0);
2039 WARN_ON((int)tp->retrans_out < 0);
2040 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
2045 /* Limits sacked_out so that sum with lost_out isn't ever larger than
2046 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
2048 static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
2052 holes = max(tp->lost_out, 1U);
2053 holes = min(holes, tp->packets_out);
2055 if ((tp->sacked_out + holes) > tp->packets_out) {
2056 tp->sacked_out = tp->packets_out - holes;
2062 /* If we receive more dupacks than we expected counting segments
2063 * in assumption of absent reordering, interpret this as reordering.
2064 * The only another reason could be bug in receiver TCP.
2066 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
2068 struct tcp_sock *tp = tcp_sk(sk);
2070 if (!tcp_limit_reno_sacked(tp))
2073 tp->reordering = min_t(u32, tp->packets_out + addend,
2074 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_max_reordering));
2076 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRENOREORDER);
2079 /* Emulate SACKs for SACKless connection: account for a new dupack. */
2081 static void tcp_add_reno_sack(struct sock *sk, int num_dupack, bool ece_ack)
2084 struct tcp_sock *tp = tcp_sk(sk);
2085 u32 prior_sacked = tp->sacked_out;
2088 tp->sacked_out += num_dupack;
2089 tcp_check_reno_reordering(sk, 0);
2090 delivered = tp->sacked_out - prior_sacked;
2092 tcp_count_delivered(tp, delivered, ece_ack);
2093 tcp_verify_left_out(tp);
2097 /* Account for ACK, ACKing some data in Reno Recovery phase. */
2099 static void tcp_remove_reno_sacks(struct sock *sk, int acked, bool ece_ack)
2101 struct tcp_sock *tp = tcp_sk(sk);
2104 /* One ACK acked hole. The rest eat duplicate ACKs. */
2105 tcp_count_delivered(tp, max_t(int, acked - tp->sacked_out, 1),
2107 if (acked - 1 >= tp->sacked_out)
2110 tp->sacked_out -= acked - 1;
2112 tcp_check_reno_reordering(sk, acked);
2113 tcp_verify_left_out(tp);
2116 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
2121 void tcp_clear_retrans(struct tcp_sock *tp)
2123 tp->retrans_out = 0;
2125 tp->undo_marker = 0;
2126 tp->undo_retrans = -1;
2130 tp->total_rto_recoveries = 0;
2131 tp->total_rto_time = 0;
2134 static inline void tcp_init_undo(struct tcp_sock *tp)
2136 tp->undo_marker = tp->snd_una;
2138 /* Retransmission still in flight may cause DSACKs later. */
2139 /* First, account for regular retransmits in flight: */
2140 tp->undo_retrans = tp->retrans_out;
2141 /* Next, account for TLP retransmits in flight: */
2142 if (tp->tlp_high_seq && tp->tlp_retrans)
2144 /* Finally, avoid 0, because undo_retrans==0 means "can undo now": */
2145 if (!tp->undo_retrans)
2146 tp->undo_retrans = -1;
2149 static bool tcp_is_rack(const struct sock *sk)
2151 return READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_recovery) &
2152 TCP_RACK_LOSS_DETECTION;
2155 /* If we detect SACK reneging, forget all SACK information
2156 * and reset tags completely, otherwise preserve SACKs. If receiver
2157 * dropped its ofo queue, we will know this due to reneging detection.
2159 static void tcp_timeout_mark_lost(struct sock *sk)
2161 struct tcp_sock *tp = tcp_sk(sk);
2162 struct sk_buff *skb, *head;
2163 bool is_reneg; /* is receiver reneging on SACKs? */
2165 head = tcp_rtx_queue_head(sk);
2166 is_reneg = head && (TCP_SKB_CB(head)->sacked & TCPCB_SACKED_ACKED);
2168 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
2170 /* Mark SACK reneging until we recover from this loss event. */
2171 tp->is_sack_reneg = 1;
2172 } else if (tcp_is_reno(tp)) {
2173 tcp_reset_reno_sack(tp);
2177 skb_rbtree_walk_from(skb) {
2179 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
2180 else if (tcp_is_rack(sk) && skb != head &&
2181 tcp_rack_skb_timeout(tp, skb, 0) > 0)
2182 continue; /* Don't mark recently sent ones lost yet */
2183 tcp_mark_skb_lost(sk, skb);
2185 tcp_verify_left_out(tp);
2186 tcp_clear_all_retrans_hints(tp);
2189 /* Enter Loss state. */
2190 void tcp_enter_loss(struct sock *sk)
2192 const struct inet_connection_sock *icsk = inet_csk(sk);
2193 struct tcp_sock *tp = tcp_sk(sk);
2194 struct net *net = sock_net(sk);
2195 bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery;
2198 tcp_timeout_mark_lost(sk);
2200 /* Reduce ssthresh if it has not yet been made inside this window. */
2201 if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
2202 !after(tp->high_seq, tp->snd_una) ||
2203 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
2204 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2205 tp->prior_cwnd = tcp_snd_cwnd(tp);
2206 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2207 tcp_ca_event(sk, CA_EVENT_LOSS);
2210 tcp_snd_cwnd_set(tp, tcp_packets_in_flight(tp) + 1);
2211 tp->snd_cwnd_cnt = 0;
2212 tp->snd_cwnd_stamp = tcp_jiffies32;
2214 /* Timeout in disordered state after receiving substantial DUPACKs
2215 * suggests that the degree of reordering is over-estimated.
2217 reordering = READ_ONCE(net->ipv4.sysctl_tcp_reordering);
2218 if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
2219 tp->sacked_out >= reordering)
2220 tp->reordering = min_t(unsigned int, tp->reordering,
2223 tcp_set_ca_state(sk, TCP_CA_Loss);
2224 tp->high_seq = tp->snd_nxt;
2225 tp->tlp_high_seq = 0;
2226 tcp_ecn_queue_cwr(tp);
2228 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
2229 * loss recovery is underway except recurring timeout(s) on
2230 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
2232 tp->frto = READ_ONCE(net->ipv4.sysctl_tcp_frto) &&
2233 (new_recovery || icsk->icsk_retransmits) &&
2234 !inet_csk(sk)->icsk_mtup.probe_size;
2237 /* If ACK arrived pointing to a remembered SACK, it means that our
2238 * remembered SACKs do not reflect real state of receiver i.e.
2239 * receiver _host_ is heavily congested (or buggy).
2241 * To avoid big spurious retransmission bursts due to transient SACK
2242 * scoreboard oddities that look like reneging, we give the receiver a
2243 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
2244 * restore sanity to the SACK scoreboard. If the apparent reneging
2245 * persists until this RTO then we'll clear the SACK scoreboard.
2247 static bool tcp_check_sack_reneging(struct sock *sk, int *ack_flag)
2249 if (*ack_flag & FLAG_SACK_RENEGING &&
2250 *ack_flag & FLAG_SND_UNA_ADVANCED) {
2251 struct tcp_sock *tp = tcp_sk(sk);
2252 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4),
2253 msecs_to_jiffies(10));
2255 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2256 delay, TCP_RTO_MAX);
2257 *ack_flag &= ~FLAG_SET_XMIT_TIMER;
2263 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2264 * counter when SACK is enabled (without SACK, sacked_out is used for
2267 * With reordering, holes may still be in flight, so RFC3517 recovery
2268 * uses pure sacked_out (total number of SACKed segments) even though
2269 * it violates the RFC that uses duplicate ACKs, often these are equal
2270 * but when e.g. out-of-window ACKs or packet duplication occurs,
2271 * they differ. Since neither occurs due to loss, TCP should really
2274 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2276 return tp->sacked_out + 1;
2279 /* Linux NewReno/SACK/ECN state machine.
2280 * --------------------------------------
2282 * "Open" Normal state, no dubious events, fast path.
2283 * "Disorder" In all the respects it is "Open",
2284 * but requires a bit more attention. It is entered when
2285 * we see some SACKs or dupacks. It is split of "Open"
2286 * mainly to move some processing from fast path to slow one.
2287 * "CWR" CWND was reduced due to some Congestion Notification event.
2288 * It can be ECN, ICMP source quench, local device congestion.
2289 * "Recovery" CWND was reduced, we are fast-retransmitting.
2290 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2292 * tcp_fastretrans_alert() is entered:
2293 * - each incoming ACK, if state is not "Open"
2294 * - when arrived ACK is unusual, namely:
2299 * Counting packets in flight is pretty simple.
2301 * in_flight = packets_out - left_out + retrans_out
2303 * packets_out is SND.NXT-SND.UNA counted in packets.
2305 * retrans_out is number of retransmitted segments.
2307 * left_out is number of segments left network, but not ACKed yet.
2309 * left_out = sacked_out + lost_out
2311 * sacked_out: Packets, which arrived to receiver out of order
2312 * and hence not ACKed. With SACKs this number is simply
2313 * amount of SACKed data. Even without SACKs
2314 * it is easy to give pretty reliable estimate of this number,
2315 * counting duplicate ACKs.
2317 * lost_out: Packets lost by network. TCP has no explicit
2318 * "loss notification" feedback from network (for now).
2319 * It means that this number can be only _guessed_.
2320 * Actually, it is the heuristics to predict lossage that
2321 * distinguishes different algorithms.
2323 * F.e. after RTO, when all the queue is considered as lost,
2324 * lost_out = packets_out and in_flight = retrans_out.
2326 * Essentially, we have now a few algorithms detecting
2329 * If the receiver supports SACK:
2331 * RFC6675/3517: It is the conventional algorithm. A packet is
2332 * considered lost if the number of higher sequence packets
2333 * SACKed is greater than or equal the DUPACK thoreshold
2334 * (reordering). This is implemented in tcp_mark_head_lost and
2335 * tcp_update_scoreboard.
2337 * RACK (draft-ietf-tcpm-rack-01): it is a newer algorithm
2338 * (2017-) that checks timing instead of counting DUPACKs.
2339 * Essentially a packet is considered lost if it's not S/ACKed
2340 * after RTT + reordering_window, where both metrics are
2341 * dynamically measured and adjusted. This is implemented in
2342 * tcp_rack_mark_lost.
2344 * If the receiver does not support SACK:
2346 * NewReno (RFC6582): in Recovery we assume that one segment
2347 * is lost (classic Reno). While we are in Recovery and
2348 * a partial ACK arrives, we assume that one more packet
2349 * is lost (NewReno). This heuristics are the same in NewReno
2352 * Really tricky (and requiring careful tuning) part of algorithm
2353 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2354 * The first determines the moment _when_ we should reduce CWND and,
2355 * hence, slow down forward transmission. In fact, it determines the moment
2356 * when we decide that hole is caused by loss, rather than by a reorder.
2358 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2359 * holes, caused by lost packets.
2361 * And the most logically complicated part of algorithm is undo
2362 * heuristics. We detect false retransmits due to both too early
2363 * fast retransmit (reordering) and underestimated RTO, analyzing
2364 * timestamps and D-SACKs. When we detect that some segments were
2365 * retransmitted by mistake and CWND reduction was wrong, we undo
2366 * window reduction and abort recovery phase. This logic is hidden
2367 * inside several functions named tcp_try_undo_<something>.
2370 /* This function decides, when we should leave Disordered state
2371 * and enter Recovery phase, reducing congestion window.
2373 * Main question: may we further continue forward transmission
2374 * with the same cwnd?
2376 static bool tcp_time_to_recover(struct sock *sk, int flag)
2378 struct tcp_sock *tp = tcp_sk(sk);
2380 /* Trick#1: The loss is proven. */
2384 /* Not-A-Trick#2 : Classic rule... */
2385 if (!tcp_is_rack(sk) && tcp_dupack_heuristics(tp) > tp->reordering)
2391 /* Detect loss in event "A" above by marking head of queue up as lost.
2392 * For RFC3517 SACK, a segment is considered lost if it
2393 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2394 * the maximum SACKed segments to pass before reaching this limit.
2396 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2398 struct tcp_sock *tp = tcp_sk(sk);
2399 struct sk_buff *skb;
2401 /* Use SACK to deduce losses of new sequences sent during recovery */
2402 const u32 loss_high = tp->snd_nxt;
2404 WARN_ON(packets > tp->packets_out);
2405 skb = tp->lost_skb_hint;
2407 /* Head already handled? */
2408 if (mark_head && after(TCP_SKB_CB(skb)->seq, tp->snd_una))
2410 cnt = tp->lost_cnt_hint;
2412 skb = tcp_rtx_queue_head(sk);
2416 skb_rbtree_walk_from(skb) {
2417 /* TODO: do this better */
2418 /* this is not the most efficient way to do this... */
2419 tp->lost_skb_hint = skb;
2420 tp->lost_cnt_hint = cnt;
2422 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2425 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2426 cnt += tcp_skb_pcount(skb);
2431 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_LOST))
2432 tcp_mark_skb_lost(sk, skb);
2437 tcp_verify_left_out(tp);
2440 /* Account newly detected lost packet(s) */
2442 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2444 struct tcp_sock *tp = tcp_sk(sk);
2446 if (tcp_is_sack(tp)) {
2447 int sacked_upto = tp->sacked_out - tp->reordering;
2448 if (sacked_upto >= 0)
2449 tcp_mark_head_lost(sk, sacked_upto, 0);
2450 else if (fast_rexmit)
2451 tcp_mark_head_lost(sk, 1, 1);
2455 static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when)
2457 return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2458 before(tp->rx_opt.rcv_tsecr, when);
2461 /* skb is spurious retransmitted if the returned timestamp echo
2462 * reply is prior to the skb transmission time
2464 static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp,
2465 const struct sk_buff *skb)
2467 return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) &&
2468 tcp_tsopt_ecr_before(tp, tcp_skb_timestamp_ts(tp->tcp_usec_ts, skb));
2471 /* Nothing was retransmitted or returned timestamp is less
2472 * than timestamp of the first retransmission.
2474 static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2476 return tp->retrans_stamp &&
2477 tcp_tsopt_ecr_before(tp, tp->retrans_stamp);
2480 /* Undo procedures. */
2482 /* We can clear retrans_stamp when there are no retransmissions in the
2483 * window. It would seem that it is trivially available for us in
2484 * tp->retrans_out, however, that kind of assumptions doesn't consider
2485 * what will happen if errors occur when sending retransmission for the
2486 * second time. ...It could the that such segment has only
2487 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2488 * the head skb is enough except for some reneging corner cases that
2489 * are not worth the effort.
2491 * Main reason for all this complexity is the fact that connection dying
2492 * time now depends on the validity of the retrans_stamp, in particular,
2493 * that successive retransmissions of a segment must not advance
2494 * retrans_stamp under any conditions.
2496 static bool tcp_any_retrans_done(const struct sock *sk)
2498 const struct tcp_sock *tp = tcp_sk(sk);
2499 struct sk_buff *skb;
2501 if (tp->retrans_out)
2504 skb = tcp_rtx_queue_head(sk);
2505 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2511 static void DBGUNDO(struct sock *sk, const char *msg)
2513 #if FASTRETRANS_DEBUG > 1
2514 struct tcp_sock *tp = tcp_sk(sk);
2515 struct inet_sock *inet = inet_sk(sk);
2517 if (sk->sk_family == AF_INET) {
2518 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2520 &inet->inet_daddr, ntohs(inet->inet_dport),
2521 tcp_snd_cwnd(tp), tcp_left_out(tp),
2522 tp->snd_ssthresh, tp->prior_ssthresh,
2525 #if IS_ENABLED(CONFIG_IPV6)
2526 else if (sk->sk_family == AF_INET6) {
2527 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2529 &sk->sk_v6_daddr, ntohs(inet->inet_dport),
2530 tcp_snd_cwnd(tp), tcp_left_out(tp),
2531 tp->snd_ssthresh, tp->prior_ssthresh,
2538 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
2540 struct tcp_sock *tp = tcp_sk(sk);
2543 struct sk_buff *skb;
2545 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2546 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2549 tcp_clear_all_retrans_hints(tp);
2552 if (tp->prior_ssthresh) {
2553 const struct inet_connection_sock *icsk = inet_csk(sk);
2555 tcp_snd_cwnd_set(tp, icsk->icsk_ca_ops->undo_cwnd(sk));
2557 if (tp->prior_ssthresh > tp->snd_ssthresh) {
2558 tp->snd_ssthresh = tp->prior_ssthresh;
2559 tcp_ecn_withdraw_cwr(tp);
2562 tp->snd_cwnd_stamp = tcp_jiffies32;
2563 tp->undo_marker = 0;
2564 tp->rack.advanced = 1; /* Force RACK to re-exam losses */
2567 static inline bool tcp_may_undo(const struct tcp_sock *tp)
2569 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2572 static bool tcp_is_non_sack_preventing_reopen(struct sock *sk)
2574 struct tcp_sock *tp = tcp_sk(sk);
2576 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2577 /* Hold old state until something *above* high_seq
2578 * is ACKed. For Reno it is MUST to prevent false
2579 * fast retransmits (RFC2582). SACK TCP is safe. */
2580 if (!tcp_any_retrans_done(sk))
2581 tp->retrans_stamp = 0;
2587 /* People celebrate: "We love our President!" */
2588 static bool tcp_try_undo_recovery(struct sock *sk)
2590 struct tcp_sock *tp = tcp_sk(sk);
2592 if (tcp_may_undo(tp)) {
2595 /* Happy end! We did not retransmit anything
2596 * or our original transmission succeeded.
2598 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2599 tcp_undo_cwnd_reduction(sk, false);
2600 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2601 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2603 mib_idx = LINUX_MIB_TCPFULLUNDO;
2605 NET_INC_STATS(sock_net(sk), mib_idx);
2606 } else if (tp->rack.reo_wnd_persist) {
2607 tp->rack.reo_wnd_persist--;
2609 if (tcp_is_non_sack_preventing_reopen(sk))
2611 tcp_set_ca_state(sk, TCP_CA_Open);
2612 tp->is_sack_reneg = 0;
2616 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2617 static bool tcp_try_undo_dsack(struct sock *sk)
2619 struct tcp_sock *tp = tcp_sk(sk);
2621 if (tp->undo_marker && !tp->undo_retrans) {
2622 tp->rack.reo_wnd_persist = min(TCP_RACK_RECOVERY_THRESH,
2623 tp->rack.reo_wnd_persist + 1);
2624 DBGUNDO(sk, "D-SACK");
2625 tcp_undo_cwnd_reduction(sk, false);
2626 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2632 /* Undo during loss recovery after partial ACK or using F-RTO. */
2633 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2635 struct tcp_sock *tp = tcp_sk(sk);
2637 if (frto_undo || tcp_may_undo(tp)) {
2638 tcp_undo_cwnd_reduction(sk, true);
2640 DBGUNDO(sk, "partial loss");
2641 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2643 NET_INC_STATS(sock_net(sk),
2644 LINUX_MIB_TCPSPURIOUSRTOS);
2645 inet_csk(sk)->icsk_retransmits = 0;
2646 if (tcp_is_non_sack_preventing_reopen(sk))
2648 if (frto_undo || tcp_is_sack(tp)) {
2649 tcp_set_ca_state(sk, TCP_CA_Open);
2650 tp->is_sack_reneg = 0;
2657 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
2658 * It computes the number of packets to send (sndcnt) based on packets newly
2660 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2661 * cwnd reductions across a full RTT.
2662 * 2) Otherwise PRR uses packet conservation to send as much as delivered.
2663 * But when SND_UNA is acked without further losses,
2664 * slow starts cwnd up to ssthresh to speed up the recovery.
2666 static void tcp_init_cwnd_reduction(struct sock *sk)
2668 struct tcp_sock *tp = tcp_sk(sk);
2670 tp->high_seq = tp->snd_nxt;
2671 tp->tlp_high_seq = 0;
2672 tp->snd_cwnd_cnt = 0;
2673 tp->prior_cwnd = tcp_snd_cwnd(tp);
2674 tp->prr_delivered = 0;
2676 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2677 tcp_ecn_queue_cwr(tp);
2680 void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked, int newly_lost, int flag)
2682 struct tcp_sock *tp = tcp_sk(sk);
2684 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2686 if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd))
2689 tp->prr_delivered += newly_acked_sacked;
2691 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2693 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2695 sndcnt = max_t(int, tp->prr_delivered - tp->prr_out,
2696 newly_acked_sacked);
2697 if (flag & FLAG_SND_UNA_ADVANCED && !newly_lost)
2699 sndcnt = min(delta, sndcnt);
2701 /* Force a fast retransmit upon entering fast recovery */
2702 sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1));
2703 tcp_snd_cwnd_set(tp, tcp_packets_in_flight(tp) + sndcnt);
2706 static inline void tcp_end_cwnd_reduction(struct sock *sk)
2708 struct tcp_sock *tp = tcp_sk(sk);
2710 if (inet_csk(sk)->icsk_ca_ops->cong_control)
2713 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2714 if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH &&
2715 (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR || tp->undo_marker)) {
2716 tcp_snd_cwnd_set(tp, tp->snd_ssthresh);
2717 tp->snd_cwnd_stamp = tcp_jiffies32;
2719 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2722 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2723 void tcp_enter_cwr(struct sock *sk)
2725 struct tcp_sock *tp = tcp_sk(sk);
2727 tp->prior_ssthresh = 0;
2728 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2729 tp->undo_marker = 0;
2730 tcp_init_cwnd_reduction(sk);
2731 tcp_set_ca_state(sk, TCP_CA_CWR);
2734 EXPORT_SYMBOL(tcp_enter_cwr);
2736 static void tcp_try_keep_open(struct sock *sk)
2738 struct tcp_sock *tp = tcp_sk(sk);
2739 int state = TCP_CA_Open;
2741 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2742 state = TCP_CA_Disorder;
2744 if (inet_csk(sk)->icsk_ca_state != state) {
2745 tcp_set_ca_state(sk, state);
2746 tp->high_seq = tp->snd_nxt;
2750 static void tcp_try_to_open(struct sock *sk, int flag)
2752 struct tcp_sock *tp = tcp_sk(sk);
2754 tcp_verify_left_out(tp);
2756 if (!tcp_any_retrans_done(sk))
2757 tp->retrans_stamp = 0;
2759 if (flag & FLAG_ECE)
2762 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2763 tcp_try_keep_open(sk);
2767 static void tcp_mtup_probe_failed(struct sock *sk)
2769 struct inet_connection_sock *icsk = inet_csk(sk);
2771 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2772 icsk->icsk_mtup.probe_size = 0;
2773 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPFAIL);
2776 static void tcp_mtup_probe_success(struct sock *sk)
2778 struct tcp_sock *tp = tcp_sk(sk);
2779 struct inet_connection_sock *icsk = inet_csk(sk);
2782 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2784 val = (u64)tcp_snd_cwnd(tp) * tcp_mss_to_mtu(sk, tp->mss_cache);
2785 do_div(val, icsk->icsk_mtup.probe_size);
2786 DEBUG_NET_WARN_ON_ONCE((u32)val != val);
2787 tcp_snd_cwnd_set(tp, max_t(u32, 1U, val));
2789 tp->snd_cwnd_cnt = 0;
2790 tp->snd_cwnd_stamp = tcp_jiffies32;
2791 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2793 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2794 icsk->icsk_mtup.probe_size = 0;
2795 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2796 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS);
2799 /* Sometimes we deduce that packets have been dropped due to reasons other than
2800 * congestion, like path MTU reductions or failed client TFO attempts. In these
2801 * cases we call this function to retransmit as many packets as cwnd allows,
2802 * without reducing cwnd. Given that retransmits will set retrans_stamp to a
2803 * non-zero value (and may do so in a later calling context due to TSQ), we
2804 * also enter CA_Loss so that we track when all retransmitted packets are ACKed
2805 * and clear retrans_stamp when that happens (to ensure later recurring RTOs
2806 * are using the correct retrans_stamp and don't declare ETIMEDOUT
2809 static void tcp_non_congestion_loss_retransmit(struct sock *sk)
2811 const struct inet_connection_sock *icsk = inet_csk(sk);
2812 struct tcp_sock *tp = tcp_sk(sk);
2814 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2815 tp->high_seq = tp->snd_nxt;
2816 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2817 tp->prior_ssthresh = 0;
2818 tp->undo_marker = 0;
2819 tcp_set_ca_state(sk, TCP_CA_Loss);
2821 tcp_xmit_retransmit_queue(sk);
2824 /* Do a simple retransmit without using the backoff mechanisms in
2825 * tcp_timer. This is used for path mtu discovery.
2826 * The socket is already locked here.
2828 void tcp_simple_retransmit(struct sock *sk)
2830 struct tcp_sock *tp = tcp_sk(sk);
2831 struct sk_buff *skb;
2834 /* A fastopen SYN request is stored as two separate packets within
2835 * the retransmit queue, this is done by tcp_send_syn_data().
2836 * As a result simply checking the MSS of the frames in the queue
2837 * will not work for the SYN packet.
2839 * Us being here is an indication of a path MTU issue so we can
2840 * assume that the fastopen SYN was lost and just mark all the
2841 * frames in the retransmit queue as lost. We will use an MSS of
2842 * -1 to mark all frames as lost, otherwise compute the current MSS.
2844 if (tp->syn_data && sk->sk_state == TCP_SYN_SENT)
2847 mss = tcp_current_mss(sk);
2849 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2850 if (tcp_skb_seglen(skb) > mss)
2851 tcp_mark_skb_lost(sk, skb);
2854 tcp_clear_retrans_hints_partial(tp);
2859 if (tcp_is_reno(tp))
2860 tcp_limit_reno_sacked(tp);
2862 tcp_verify_left_out(tp);
2864 /* Don't muck with the congestion window here.
2865 * Reason is that we do not increase amount of _data_
2866 * in network, but units changed and effective
2867 * cwnd/ssthresh really reduced now.
2869 tcp_non_congestion_loss_retransmit(sk);
2871 EXPORT_SYMBOL(tcp_simple_retransmit);
2873 void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2875 struct tcp_sock *tp = tcp_sk(sk);
2878 if (tcp_is_reno(tp))
2879 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2881 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2883 NET_INC_STATS(sock_net(sk), mib_idx);
2885 tp->prior_ssthresh = 0;
2888 if (!tcp_in_cwnd_reduction(sk)) {
2890 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2891 tcp_init_cwnd_reduction(sk);
2893 tcp_set_ca_state(sk, TCP_CA_Recovery);
2896 static void tcp_update_rto_time(struct tcp_sock *tp)
2898 if (tp->rto_stamp) {
2899 tp->total_rto_time += tcp_time_stamp_ms(tp) - tp->rto_stamp;
2904 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2905 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2907 static void tcp_process_loss(struct sock *sk, int flag, int num_dupack,
2910 struct tcp_sock *tp = tcp_sk(sk);
2911 bool recovered = !before(tp->snd_una, tp->high_seq);
2913 if ((flag & FLAG_SND_UNA_ADVANCED || rcu_access_pointer(tp->fastopen_rsk)) &&
2914 tcp_try_undo_loss(sk, false))
2917 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2918 /* Step 3.b. A timeout is spurious if not all data are
2919 * lost, i.e., never-retransmitted data are (s)acked.
2921 if ((flag & FLAG_ORIG_SACK_ACKED) &&
2922 tcp_try_undo_loss(sk, true))
2925 if (after(tp->snd_nxt, tp->high_seq)) {
2926 if (flag & FLAG_DATA_SACKED || num_dupack)
2927 tp->frto = 0; /* Step 3.a. loss was real */
2928 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2929 tp->high_seq = tp->snd_nxt;
2930 /* Step 2.b. Try send new data (but deferred until cwnd
2931 * is updated in tcp_ack()). Otherwise fall back to
2932 * the conventional recovery.
2934 if (!tcp_write_queue_empty(sk) &&
2935 after(tcp_wnd_end(tp), tp->snd_nxt)) {
2936 *rexmit = REXMIT_NEW;
2944 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2945 tcp_try_undo_recovery(sk);
2948 if (tcp_is_reno(tp)) {
2949 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2950 * delivered. Lower inflight to clock out (re)transmissions.
2952 if (after(tp->snd_nxt, tp->high_seq) && num_dupack)
2953 tcp_add_reno_sack(sk, num_dupack, flag & FLAG_ECE);
2954 else if (flag & FLAG_SND_UNA_ADVANCED)
2955 tcp_reset_reno_sack(tp);
2957 *rexmit = REXMIT_LOST;
2960 static bool tcp_force_fast_retransmit(struct sock *sk)
2962 struct tcp_sock *tp = tcp_sk(sk);
2964 return after(tcp_highest_sack_seq(tp),
2965 tp->snd_una + tp->reordering * tp->mss_cache);
2968 /* Undo during fast recovery after partial ACK. */
2969 static bool tcp_try_undo_partial(struct sock *sk, u32 prior_snd_una,
2972 struct tcp_sock *tp = tcp_sk(sk);
2974 if (tp->undo_marker && tcp_packet_delayed(tp)) {
2975 /* Plain luck! Hole if filled with delayed
2976 * packet, rather than with a retransmit. Check reordering.
2978 tcp_check_sack_reordering(sk, prior_snd_una, 1);
2980 /* We are getting evidence that the reordering degree is higher
2981 * than we realized. If there are no retransmits out then we
2982 * can undo. Otherwise we clock out new packets but do not
2983 * mark more packets lost or retransmit more.
2985 if (tp->retrans_out)
2988 if (!tcp_any_retrans_done(sk))
2989 tp->retrans_stamp = 0;
2991 DBGUNDO(sk, "partial recovery");
2992 tcp_undo_cwnd_reduction(sk, true);
2993 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2994 tcp_try_keep_open(sk);
2996 /* Partial ACK arrived. Force fast retransmit. */
2997 *do_lost = tcp_force_fast_retransmit(sk);
3002 static void tcp_identify_packet_loss(struct sock *sk, int *ack_flag)
3004 struct tcp_sock *tp = tcp_sk(sk);
3006 if (tcp_rtx_queue_empty(sk))
3009 if (unlikely(tcp_is_reno(tp))) {
3010 tcp_newreno_mark_lost(sk, *ack_flag & FLAG_SND_UNA_ADVANCED);
3011 } else if (tcp_is_rack(sk)) {
3012 u32 prior_retrans = tp->retrans_out;
3014 if (tcp_rack_mark_lost(sk))
3015 *ack_flag &= ~FLAG_SET_XMIT_TIMER;
3016 if (prior_retrans > tp->retrans_out)
3017 *ack_flag |= FLAG_LOST_RETRANS;
3021 /* Process an event, which can update packets-in-flight not trivially.
3022 * Main goal of this function is to calculate new estimate for left_out,
3023 * taking into account both packets sitting in receiver's buffer and
3024 * packets lost by network.
3026 * Besides that it updates the congestion state when packet loss or ECN
3027 * is detected. But it does not reduce the cwnd, it is done by the
3028 * congestion control later.
3030 * It does _not_ decide what to send, it is made in function
3031 * tcp_xmit_retransmit_queue().
3033 static void tcp_fastretrans_alert(struct sock *sk, const u32 prior_snd_una,
3034 int num_dupack, int *ack_flag, int *rexmit)
3036 struct inet_connection_sock *icsk = inet_csk(sk);
3037 struct tcp_sock *tp = tcp_sk(sk);
3038 int fast_rexmit = 0, flag = *ack_flag;
3039 bool ece_ack = flag & FLAG_ECE;
3040 bool do_lost = num_dupack || ((flag & FLAG_DATA_SACKED) &&
3041 tcp_force_fast_retransmit(sk));
3043 if (!tp->packets_out && tp->sacked_out)
3046 /* Now state machine starts.
3047 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
3049 tp->prior_ssthresh = 0;
3051 /* B. In all the states check for reneging SACKs. */
3052 if (tcp_check_sack_reneging(sk, ack_flag))
3055 /* C. Check consistency of the current state. */
3056 tcp_verify_left_out(tp);
3058 /* D. Check state exit conditions. State can be terminated
3059 * when high_seq is ACKed. */
3060 if (icsk->icsk_ca_state == TCP_CA_Open) {
3061 WARN_ON(tp->retrans_out != 0 && !tp->syn_data);
3062 tp->retrans_stamp = 0;
3063 } else if (!before(tp->snd_una, tp->high_seq)) {
3064 switch (icsk->icsk_ca_state) {
3066 /* CWR is to be held something *above* high_seq
3067 * is ACKed for CWR bit to reach receiver. */
3068 if (tp->snd_una != tp->high_seq) {
3069 tcp_end_cwnd_reduction(sk);
3070 tcp_set_ca_state(sk, TCP_CA_Open);
3074 case TCP_CA_Recovery:
3075 if (tcp_is_reno(tp))
3076 tcp_reset_reno_sack(tp);
3077 if (tcp_try_undo_recovery(sk))
3079 tcp_end_cwnd_reduction(sk);
3084 /* E. Process state. */
3085 switch (icsk->icsk_ca_state) {
3086 case TCP_CA_Recovery:
3087 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
3088 if (tcp_is_reno(tp))
3089 tcp_add_reno_sack(sk, num_dupack, ece_ack);
3090 } else if (tcp_try_undo_partial(sk, prior_snd_una, &do_lost))
3093 if (tcp_try_undo_dsack(sk))
3094 tcp_try_to_open(sk, flag);
3096 tcp_identify_packet_loss(sk, ack_flag);
3097 if (icsk->icsk_ca_state != TCP_CA_Recovery) {
3098 if (!tcp_time_to_recover(sk, flag))
3100 /* Undo reverts the recovery state. If loss is evident,
3101 * starts a new recovery (e.g. reordering then loss);
3103 tcp_enter_recovery(sk, ece_ack);
3107 tcp_process_loss(sk, flag, num_dupack, rexmit);
3108 if (icsk->icsk_ca_state != TCP_CA_Loss)
3109 tcp_update_rto_time(tp);
3110 tcp_identify_packet_loss(sk, ack_flag);
3111 if (!(icsk->icsk_ca_state == TCP_CA_Open ||
3112 (*ack_flag & FLAG_LOST_RETRANS)))
3114 /* Change state if cwnd is undone or retransmits are lost */
3117 if (tcp_is_reno(tp)) {
3118 if (flag & FLAG_SND_UNA_ADVANCED)
3119 tcp_reset_reno_sack(tp);
3120 tcp_add_reno_sack(sk, num_dupack, ece_ack);
3123 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
3124 tcp_try_undo_dsack(sk);
3126 tcp_identify_packet_loss(sk, ack_flag);
3127 if (!tcp_time_to_recover(sk, flag)) {
3128 tcp_try_to_open(sk, flag);
3132 /* MTU probe failure: don't reduce cwnd */
3133 if (icsk->icsk_ca_state < TCP_CA_CWR &&
3134 icsk->icsk_mtup.probe_size &&
3135 tp->snd_una == tp->mtu_probe.probe_seq_start) {
3136 tcp_mtup_probe_failed(sk);
3137 /* Restores the reduction we did in tcp_mtup_probe() */
3138 tcp_snd_cwnd_set(tp, tcp_snd_cwnd(tp) + 1);
3139 tcp_simple_retransmit(sk);
3143 /* Otherwise enter Recovery state */
3144 tcp_enter_recovery(sk, ece_ack);
3148 if (!tcp_is_rack(sk) && do_lost)
3149 tcp_update_scoreboard(sk, fast_rexmit);
3150 *rexmit = REXMIT_LOST;
3153 static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us, const int flag)
3155 u32 wlen = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_rtt_wlen) * HZ;
3156 struct tcp_sock *tp = tcp_sk(sk);
3158 if ((flag & FLAG_ACK_MAYBE_DELAYED) && rtt_us > tcp_min_rtt(tp)) {
3159 /* If the remote keeps returning delayed ACKs, eventually
3160 * the min filter would pick it up and overestimate the
3161 * prop. delay when it expires. Skip suspected delayed ACKs.
3165 minmax_running_min(&tp->rtt_min, wlen, tcp_jiffies32,
3166 rtt_us ? : jiffies_to_usecs(1));
3169 static bool tcp_ack_update_rtt(struct sock *sk, const int flag,
3170 long seq_rtt_us, long sack_rtt_us,
3171 long ca_rtt_us, struct rate_sample *rs)
3173 const struct tcp_sock *tp = tcp_sk(sk);
3175 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
3176 * broken middle-boxes or peers may corrupt TS-ECR fields. But
3177 * Karn's algorithm forbids taking RTT if some retransmitted data
3178 * is acked (RFC6298).
3181 seq_rtt_us = sack_rtt_us;
3183 /* RTTM Rule: A TSecr value received in a segment is used to
3184 * update the averaged RTT measurement only if the segment
3185 * acknowledges some new data, i.e., only if it advances the
3186 * left edge of the send window.
3187 * See draft-ietf-tcplw-high-performance-00, section 3.3.
3189 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp &&
3190 tp->rx_opt.rcv_tsecr && flag & FLAG_ACKED)
3191 seq_rtt_us = ca_rtt_us = tcp_rtt_tsopt_us(tp);
3193 rs->rtt_us = ca_rtt_us; /* RTT of last (S)ACKed packet (or -1) */
3197 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
3198 * always taken together with ACK, SACK, or TS-opts. Any negative
3199 * values will be skipped with the seq_rtt_us < 0 check above.
3201 tcp_update_rtt_min(sk, ca_rtt_us, flag);
3202 tcp_rtt_estimator(sk, seq_rtt_us);
3205 /* RFC6298: only reset backoff on valid RTT measurement. */
3206 inet_csk(sk)->icsk_backoff = 0;
3210 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
3211 void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req)
3213 struct rate_sample rs;
3216 if (req && !req->num_retrans && tcp_rsk(req)->snt_synack)
3217 rtt_us = tcp_stamp_us_delta(tcp_clock_us(), tcp_rsk(req)->snt_synack);
3219 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us, &rs);
3223 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
3225 const struct inet_connection_sock *icsk = inet_csk(sk);
3227 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
3228 tcp_sk(sk)->snd_cwnd_stamp = tcp_jiffies32;
3231 /* Restart timer after forward progress on connection.
3232 * RFC2988 recommends to restart timer to now+rto.
3234 void tcp_rearm_rto(struct sock *sk)
3236 const struct inet_connection_sock *icsk = inet_csk(sk);
3237 struct tcp_sock *tp = tcp_sk(sk);
3239 /* If the retrans timer is currently being used by Fast Open
3240 * for SYN-ACK retrans purpose, stay put.
3242 if (rcu_access_pointer(tp->fastopen_rsk))
3245 if (!tp->packets_out) {
3246 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
3248 u32 rto = inet_csk(sk)->icsk_rto;
3249 /* Offset the time elapsed after installing regular RTO */
3250 if (icsk->icsk_pending == ICSK_TIME_REO_TIMEOUT ||
3251 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
3252 s64 delta_us = tcp_rto_delta_us(sk);
3253 /* delta_us may not be positive if the socket is locked
3254 * when the retrans timer fires and is rescheduled.
3256 rto = usecs_to_jiffies(max_t(int, delta_us, 1));
3258 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
3263 /* Try to schedule a loss probe; if that doesn't work, then schedule an RTO. */
3264 static void tcp_set_xmit_timer(struct sock *sk)
3266 if (!tcp_schedule_loss_probe(sk, true))
3270 /* If we get here, the whole TSO packet has not been acked. */
3271 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3273 struct tcp_sock *tp = tcp_sk(sk);
3276 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3278 packets_acked = tcp_skb_pcount(skb);
3279 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3281 packets_acked -= tcp_skb_pcount(skb);
3283 if (packets_acked) {
3284 BUG_ON(tcp_skb_pcount(skb) == 0);
3285 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3288 return packets_acked;
3291 static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
3292 const struct sk_buff *ack_skb, u32 prior_snd_una)
3294 const struct skb_shared_info *shinfo;
3296 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3297 if (likely(!TCP_SKB_CB(skb)->txstamp_ack))
3300 shinfo = skb_shinfo(skb);
3301 if (!before(shinfo->tskey, prior_snd_una) &&
3302 before(shinfo->tskey, tcp_sk(sk)->snd_una)) {
3303 tcp_skb_tsorted_save(skb) {
3304 __skb_tstamp_tx(skb, ack_skb, NULL, sk, SCM_TSTAMP_ACK);
3305 } tcp_skb_tsorted_restore(skb);
3309 /* Remove acknowledged frames from the retransmission queue. If our packet
3310 * is before the ack sequence we can discard it as it's confirmed to have
3311 * arrived at the other end.
3313 static int tcp_clean_rtx_queue(struct sock *sk, const struct sk_buff *ack_skb,
3314 u32 prior_fack, u32 prior_snd_una,
3315 struct tcp_sacktag_state *sack, bool ece_ack)
3317 const struct inet_connection_sock *icsk = inet_csk(sk);
3318 u64 first_ackt, last_ackt;
3319 struct tcp_sock *tp = tcp_sk(sk);
3320 u32 prior_sacked = tp->sacked_out;
3321 u32 reord = tp->snd_nxt; /* lowest acked un-retx un-sacked seq */
3322 struct sk_buff *skb, *next;
3323 bool fully_acked = true;
3324 long sack_rtt_us = -1L;
3325 long seq_rtt_us = -1L;
3326 long ca_rtt_us = -1L;
3333 for (skb = skb_rb_first(&sk->tcp_rtx_queue); skb; skb = next) {
3334 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3335 const u32 start_seq = scb->seq;
3336 u8 sacked = scb->sacked;
3339 /* Determine how many packets and what bytes were acked, tso and else */
3340 if (after(scb->end_seq, tp->snd_una)) {
3341 if (tcp_skb_pcount(skb) == 1 ||
3342 !after(tp->snd_una, scb->seq))
3345 acked_pcount = tcp_tso_acked(sk, skb);
3348 fully_acked = false;
3350 acked_pcount = tcp_skb_pcount(skb);
3353 if (unlikely(sacked & TCPCB_RETRANS)) {
3354 if (sacked & TCPCB_SACKED_RETRANS)
3355 tp->retrans_out -= acked_pcount;
3356 flag |= FLAG_RETRANS_DATA_ACKED;
3357 } else if (!(sacked & TCPCB_SACKED_ACKED)) {
3358 last_ackt = tcp_skb_timestamp_us(skb);
3359 WARN_ON_ONCE(last_ackt == 0);
3361 first_ackt = last_ackt;
3363 if (before(start_seq, reord))
3365 if (!after(scb->end_seq, tp->high_seq))
3366 flag |= FLAG_ORIG_SACK_ACKED;
3369 if (sacked & TCPCB_SACKED_ACKED) {
3370 tp->sacked_out -= acked_pcount;
3371 } else if (tcp_is_sack(tp)) {
3372 tcp_count_delivered(tp, acked_pcount, ece_ack);
3373 if (!tcp_skb_spurious_retrans(tp, skb))
3374 tcp_rack_advance(tp, sacked, scb->end_seq,
3375 tcp_skb_timestamp_us(skb));
3377 if (sacked & TCPCB_LOST)
3378 tp->lost_out -= acked_pcount;
3380 tp->packets_out -= acked_pcount;
3381 pkts_acked += acked_pcount;
3382 tcp_rate_skb_delivered(sk, skb, sack->rate);
3384 /* Initial outgoing SYN's get put onto the write_queue
3385 * just like anything else we transmit. It is not
3386 * true data, and if we misinform our callers that
3387 * this ACK acks real data, we will erroneously exit
3388 * connection startup slow start one packet too
3389 * quickly. This is severely frowned upon behavior.
3391 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
3392 flag |= FLAG_DATA_ACKED;
3394 flag |= FLAG_SYN_ACKED;
3395 tp->retrans_stamp = 0;
3401 tcp_ack_tstamp(sk, skb, ack_skb, prior_snd_una);
3403 next = skb_rb_next(skb);
3404 if (unlikely(skb == tp->retransmit_skb_hint))
3405 tp->retransmit_skb_hint = NULL;
3406 if (unlikely(skb == tp->lost_skb_hint))
3407 tp->lost_skb_hint = NULL;
3408 tcp_highest_sack_replace(sk, skb, next);
3409 tcp_rtx_queue_unlink_and_free(skb, sk);
3413 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
3415 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3416 tp->snd_up = tp->snd_una;
3419 tcp_ack_tstamp(sk, skb, ack_skb, prior_snd_una);
3420 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
3421 flag |= FLAG_SACK_RENEGING;
3424 if (likely(first_ackt) && !(flag & FLAG_RETRANS_DATA_ACKED)) {
3425 seq_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, first_ackt);
3426 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, last_ackt);
3428 if (pkts_acked == 1 && fully_acked && !prior_sacked &&
3429 (tp->snd_una - prior_snd_una) < tp->mss_cache &&
3430 sack->rate->prior_delivered + 1 == tp->delivered &&
3431 !(flag & (FLAG_CA_ALERT | FLAG_SYN_ACKED))) {
3432 /* Conservatively mark a delayed ACK. It's typically
3433 * from a lone runt packet over the round trip to
3434 * a receiver w/o out-of-order or CE events.
3436 flag |= FLAG_ACK_MAYBE_DELAYED;
3439 if (sack->first_sackt) {
3440 sack_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->first_sackt);
3441 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->last_sackt);
3443 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us,
3444 ca_rtt_us, sack->rate);
3446 if (flag & FLAG_ACKED) {
3447 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3448 if (unlikely(icsk->icsk_mtup.probe_size &&
3449 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3450 tcp_mtup_probe_success(sk);
3453 if (tcp_is_reno(tp)) {
3454 tcp_remove_reno_sacks(sk, pkts_acked, ece_ack);
3456 /* If any of the cumulatively ACKed segments was
3457 * retransmitted, non-SACK case cannot confirm that
3458 * progress was due to original transmission due to
3459 * lack of TCPCB_SACKED_ACKED bits even if some of
3460 * the packets may have been never retransmitted.
3462 if (flag & FLAG_RETRANS_DATA_ACKED)
3463 flag &= ~FLAG_ORIG_SACK_ACKED;
3467 /* Non-retransmitted hole got filled? That's reordering */
3468 if (before(reord, prior_fack))
3469 tcp_check_sack_reordering(sk, reord, 0);
3471 delta = prior_sacked - tp->sacked_out;
3472 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3474 } else if (skb && rtt_update && sack_rtt_us >= 0 &&
3475 sack_rtt_us > tcp_stamp_us_delta(tp->tcp_mstamp,
3476 tcp_skb_timestamp_us(skb))) {
3477 /* Do not re-arm RTO if the sack RTT is measured from data sent
3478 * after when the head was last (re)transmitted. Otherwise the
3479 * timeout may continue to extend in loss recovery.
3481 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3484 if (icsk->icsk_ca_ops->pkts_acked) {
3485 struct ack_sample sample = { .pkts_acked = pkts_acked,
3486 .rtt_us = sack->rate->rtt_us };
3488 sample.in_flight = tp->mss_cache *
3489 (tp->delivered - sack->rate->prior_delivered);
3490 icsk->icsk_ca_ops->pkts_acked(sk, &sample);
3493 #if FASTRETRANS_DEBUG > 0
3494 WARN_ON((int)tp->sacked_out < 0);
3495 WARN_ON((int)tp->lost_out < 0);
3496 WARN_ON((int)tp->retrans_out < 0);
3497 if (!tp->packets_out && tcp_is_sack(tp)) {
3498 icsk = inet_csk(sk);
3500 pr_debug("Leak l=%u %d\n",
3501 tp->lost_out, icsk->icsk_ca_state);
3504 if (tp->sacked_out) {
3505 pr_debug("Leak s=%u %d\n",
3506 tp->sacked_out, icsk->icsk_ca_state);
3509 if (tp->retrans_out) {
3510 pr_debug("Leak r=%u %d\n",
3511 tp->retrans_out, icsk->icsk_ca_state);
3512 tp->retrans_out = 0;
3519 static void tcp_ack_probe(struct sock *sk)
3521 struct inet_connection_sock *icsk = inet_csk(sk);
3522 struct sk_buff *head = tcp_send_head(sk);
3523 const struct tcp_sock *tp = tcp_sk(sk);
3525 /* Was it a usable window open? */
3528 if (!after(TCP_SKB_CB(head)->end_seq, tcp_wnd_end(tp))) {
3529 icsk->icsk_backoff = 0;
3530 icsk->icsk_probes_tstamp = 0;
3531 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3532 /* Socket must be waked up by subsequent tcp_data_snd_check().
3533 * This function is not for random using!
3536 unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX);
3538 when = tcp_clamp_probe0_to_user_timeout(sk, when);
3539 tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, when, TCP_RTO_MAX);
3543 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3545 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3546 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3549 /* Decide wheather to run the increase function of congestion control. */
3550 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3552 /* If reordering is high then always grow cwnd whenever data is
3553 * delivered regardless of its ordering. Otherwise stay conservative
3554 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3555 * new SACK or ECE mark may first advance cwnd here and later reduce
3556 * cwnd in tcp_fastretrans_alert() based on more states.
3558 if (tcp_sk(sk)->reordering >
3559 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_reordering))
3560 return flag & FLAG_FORWARD_PROGRESS;
3562 return flag & FLAG_DATA_ACKED;
3565 /* The "ultimate" congestion control function that aims to replace the rigid
3566 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction).
3567 * It's called toward the end of processing an ACK with precise rate
3568 * information. All transmission or retransmission are delayed afterwards.
3570 static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked,
3571 int flag, const struct rate_sample *rs)
3573 const struct inet_connection_sock *icsk = inet_csk(sk);
3575 if (icsk->icsk_ca_ops->cong_control) {
3576 icsk->icsk_ca_ops->cong_control(sk, ack, flag, rs);
3580 if (tcp_in_cwnd_reduction(sk)) {
3581 /* Reduce cwnd if state mandates */
3582 tcp_cwnd_reduction(sk, acked_sacked, rs->losses, flag);
3583 } else if (tcp_may_raise_cwnd(sk, flag)) {
3584 /* Advance cwnd if state allows */
3585 tcp_cong_avoid(sk, ack, acked_sacked);
3587 tcp_update_pacing_rate(sk);
3590 /* Check that window update is acceptable.
3591 * The function assumes that snd_una<=ack<=snd_next.
3593 static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3594 const u32 ack, const u32 ack_seq,
3597 return after(ack, tp->snd_una) ||
3598 after(ack_seq, tp->snd_wl1) ||
3599 (ack_seq == tp->snd_wl1 && (nwin > tp->snd_wnd || !nwin));
3602 static void tcp_snd_sne_update(struct tcp_sock *tp, u32 ack)
3604 #ifdef CONFIG_TCP_AO
3605 struct tcp_ao_info *ao;
3607 if (!static_branch_unlikely(&tcp_ao_needed.key))
3610 ao = rcu_dereference_protected(tp->ao_info,
3611 lockdep_sock_is_held((struct sock *)tp));
3612 if (ao && ack < tp->snd_una) {
3614 trace_tcp_ao_snd_sne_update((struct sock *)tp, ao->snd_sne);
3619 /* If we update tp->snd_una, also update tp->bytes_acked */
3620 static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack)
3622 u32 delta = ack - tp->snd_una;
3624 sock_owned_by_me((struct sock *)tp);
3625 tp->bytes_acked += delta;
3626 tcp_snd_sne_update(tp, ack);
3630 static void tcp_rcv_sne_update(struct tcp_sock *tp, u32 seq)
3632 #ifdef CONFIG_TCP_AO
3633 struct tcp_ao_info *ao;
3635 if (!static_branch_unlikely(&tcp_ao_needed.key))
3638 ao = rcu_dereference_protected(tp->ao_info,
3639 lockdep_sock_is_held((struct sock *)tp));
3640 if (ao && seq < tp->rcv_nxt) {
3642 trace_tcp_ao_rcv_sne_update((struct sock *)tp, ao->rcv_sne);
3647 /* If we update tp->rcv_nxt, also update tp->bytes_received */
3648 static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq)
3650 u32 delta = seq - tp->rcv_nxt;
3652 sock_owned_by_me((struct sock *)tp);
3653 tp->bytes_received += delta;
3654 tcp_rcv_sne_update(tp, seq);
3655 WRITE_ONCE(tp->rcv_nxt, seq);
3658 /* Update our send window.
3660 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3661 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3663 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3666 struct tcp_sock *tp = tcp_sk(sk);
3668 u32 nwin = ntohs(tcp_hdr(skb)->window);
3670 if (likely(!tcp_hdr(skb)->syn))
3671 nwin <<= tp->rx_opt.snd_wscale;
3673 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3674 flag |= FLAG_WIN_UPDATE;
3675 tcp_update_wl(tp, ack_seq);
3677 if (tp->snd_wnd != nwin) {
3680 /* Note, it is the only place, where
3681 * fast path is recovered for sending TCP.
3684 tcp_fast_path_check(sk);
3686 if (!tcp_write_queue_empty(sk))
3687 tcp_slow_start_after_idle_check(sk);
3689 if (nwin > tp->max_window) {
3690 tp->max_window = nwin;
3691 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3696 tcp_snd_una_update(tp, ack);
3701 static bool __tcp_oow_rate_limited(struct net *net, int mib_idx,
3702 u32 *last_oow_ack_time)
3704 /* Paired with the WRITE_ONCE() in this function. */
3705 u32 val = READ_ONCE(*last_oow_ack_time);
3708 s32 elapsed = (s32)(tcp_jiffies32 - val);
3711 elapsed < READ_ONCE(net->ipv4.sysctl_tcp_invalid_ratelimit)) {
3712 NET_INC_STATS(net, mib_idx);
3713 return true; /* rate-limited: don't send yet! */
3717 /* Paired with the prior READ_ONCE() and with itself,
3718 * as we might be lockless.
3720 WRITE_ONCE(*last_oow_ack_time, tcp_jiffies32);
3722 return false; /* not rate-limited: go ahead, send dupack now! */
3725 /* Return true if we're currently rate-limiting out-of-window ACKs and
3726 * thus shouldn't send a dupack right now. We rate-limit dupacks in
3727 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3728 * attacks that send repeated SYNs or ACKs for the same connection. To
3729 * do this, we do not send a duplicate SYNACK or ACK if the remote
3730 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3732 bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb,
3733 int mib_idx, u32 *last_oow_ack_time)
3735 /* Data packets without SYNs are not likely part of an ACK loop. */
3736 if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) &&
3740 return __tcp_oow_rate_limited(net, mib_idx, last_oow_ack_time);
3743 /* RFC 5961 7 [ACK Throttling] */
3744 static void tcp_send_challenge_ack(struct sock *sk)
3746 struct tcp_sock *tp = tcp_sk(sk);
3747 struct net *net = sock_net(sk);
3748 u32 count, now, ack_limit;
3750 /* First check our per-socket dupack rate limit. */
3751 if (__tcp_oow_rate_limited(net,
3752 LINUX_MIB_TCPACKSKIPPEDCHALLENGE,
3753 &tp->last_oow_ack_time))
3756 ack_limit = READ_ONCE(net->ipv4.sysctl_tcp_challenge_ack_limit);
3757 if (ack_limit == INT_MAX)
3760 /* Then check host-wide RFC 5961 rate limit. */
3762 if (now != READ_ONCE(net->ipv4.tcp_challenge_timestamp)) {
3763 u32 half = (ack_limit + 1) >> 1;
3765 WRITE_ONCE(net->ipv4.tcp_challenge_timestamp, now);
3766 WRITE_ONCE(net->ipv4.tcp_challenge_count,
3767 get_random_u32_inclusive(half, ack_limit + half - 1));
3769 count = READ_ONCE(net->ipv4.tcp_challenge_count);
3771 WRITE_ONCE(net->ipv4.tcp_challenge_count, count - 1);
3773 NET_INC_STATS(net, LINUX_MIB_TCPCHALLENGEACK);
3778 static void tcp_store_ts_recent(struct tcp_sock *tp)
3780 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3781 tp->rx_opt.ts_recent_stamp = ktime_get_seconds();
3784 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3786 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3787 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3788 * extra check below makes sure this can only happen
3789 * for pure ACK frames. -DaveM
3791 * Not only, also it occurs for expired timestamps.
3794 if (tcp_paws_check(&tp->rx_opt, 0))
3795 tcp_store_ts_recent(tp);
3799 /* This routine deals with acks during a TLP episode and ends an episode by
3800 * resetting tlp_high_seq. Ref: TLP algorithm in draft-ietf-tcpm-rack
3802 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3804 struct tcp_sock *tp = tcp_sk(sk);
3806 if (before(ack, tp->tlp_high_seq))
3809 if (!tp->tlp_retrans) {
3810 /* TLP of new data has been acknowledged */
3811 tp->tlp_high_seq = 0;
3812 } else if (flag & FLAG_DSACK_TLP) {
3813 /* This DSACK means original and TLP probe arrived; no loss */
3814 tp->tlp_high_seq = 0;
3815 } else if (after(ack, tp->tlp_high_seq)) {
3816 /* ACK advances: there was a loss, so reduce cwnd. Reset
3817 * tlp_high_seq in tcp_init_cwnd_reduction()
3819 tcp_init_cwnd_reduction(sk);
3820 tcp_set_ca_state(sk, TCP_CA_CWR);
3821 tcp_end_cwnd_reduction(sk);
3822 tcp_try_keep_open(sk);
3823 NET_INC_STATS(sock_net(sk),
3824 LINUX_MIB_TCPLOSSPROBERECOVERY);
3825 } else if (!(flag & (FLAG_SND_UNA_ADVANCED |
3826 FLAG_NOT_DUP | FLAG_DATA_SACKED))) {
3827 /* Pure dupack: original and TLP probe arrived; no loss */
3828 tp->tlp_high_seq = 0;
3832 static inline void tcp_in_ack_event(struct sock *sk, u32 flags)
3834 const struct inet_connection_sock *icsk = inet_csk(sk);
3836 if (icsk->icsk_ca_ops->in_ack_event)
3837 icsk->icsk_ca_ops->in_ack_event(sk, flags);
3840 /* Congestion control has updated the cwnd already. So if we're in
3841 * loss recovery then now we do any new sends (for FRTO) or
3842 * retransmits (for CA_Loss or CA_recovery) that make sense.
3844 static void tcp_xmit_recovery(struct sock *sk, int rexmit)
3846 struct tcp_sock *tp = tcp_sk(sk);
3848 if (rexmit == REXMIT_NONE || sk->sk_state == TCP_SYN_SENT)
3851 if (unlikely(rexmit == REXMIT_NEW)) {
3852 __tcp_push_pending_frames(sk, tcp_current_mss(sk),
3854 if (after(tp->snd_nxt, tp->high_seq))
3858 tcp_xmit_retransmit_queue(sk);
3861 /* Returns the number of packets newly acked or sacked by the current ACK */
3862 static u32 tcp_newly_delivered(struct sock *sk, u32 prior_delivered, int flag)
3864 const struct net *net = sock_net(sk);
3865 struct tcp_sock *tp = tcp_sk(sk);
3868 delivered = tp->delivered - prior_delivered;
3869 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVERED, delivered);
3870 if (flag & FLAG_ECE)
3871 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVEREDCE, delivered);
3876 /* This routine deals with incoming acks, but not outgoing ones. */
3877 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3879 struct inet_connection_sock *icsk = inet_csk(sk);
3880 struct tcp_sock *tp = tcp_sk(sk);
3881 struct tcp_sacktag_state sack_state;
3882 struct rate_sample rs = { .prior_delivered = 0 };
3883 u32 prior_snd_una = tp->snd_una;
3884 bool is_sack_reneg = tp->is_sack_reneg;
3885 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3886 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3888 int prior_packets = tp->packets_out;
3889 u32 delivered = tp->delivered;
3890 u32 lost = tp->lost;
3891 int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */
3894 sack_state.first_sackt = 0;
3895 sack_state.rate = &rs;
3896 sack_state.sack_delivered = 0;
3898 /* We very likely will need to access rtx queue. */
3899 prefetch(sk->tcp_rtx_queue.rb_node);
3901 /* If the ack is older than previous acks
3902 * then we can probably ignore it.
3904 if (before(ack, prior_snd_una)) {
3907 /* do not accept ACK for bytes we never sent. */
3908 max_window = min_t(u64, tp->max_window, tp->bytes_acked);
3909 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3910 if (before(ack, prior_snd_una - max_window)) {
3911 if (!(flag & FLAG_NO_CHALLENGE_ACK))
3912 tcp_send_challenge_ack(sk);
3913 return -SKB_DROP_REASON_TCP_TOO_OLD_ACK;
3918 /* If the ack includes data we haven't sent yet, discard
3919 * this segment (RFC793 Section 3.9).
3921 if (after(ack, tp->snd_nxt))
3922 return -SKB_DROP_REASON_TCP_ACK_UNSENT_DATA;
3924 if (after(ack, prior_snd_una)) {
3925 flag |= FLAG_SND_UNA_ADVANCED;
3926 icsk->icsk_retransmits = 0;
3928 #if IS_ENABLED(CONFIG_TLS_DEVICE)
3929 if (static_branch_unlikely(&clean_acked_data_enabled.key))
3930 if (icsk->icsk_clean_acked)
3931 icsk->icsk_clean_acked(sk, ack);
3935 prior_fack = tcp_is_sack(tp) ? tcp_highest_sack_seq(tp) : tp->snd_una;
3936 rs.prior_in_flight = tcp_packets_in_flight(tp);
3938 /* ts_recent update must be made after we are sure that the packet
3941 if (flag & FLAG_UPDATE_TS_RECENT)
3942 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3944 if ((flag & (FLAG_SLOWPATH | FLAG_SND_UNA_ADVANCED)) ==
3945 FLAG_SND_UNA_ADVANCED) {
3946 /* Window is constant, pure forward advance.
3947 * No more checks are required.
3948 * Note, we use the fact that SND.UNA>=SND.WL2.
3950 tcp_update_wl(tp, ack_seq);
3951 tcp_snd_una_update(tp, ack);
3952 flag |= FLAG_WIN_UPDATE;
3954 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
3956 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPACKS);
3958 u32 ack_ev_flags = CA_ACK_SLOWPATH;
3960 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3963 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3965 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3967 if (TCP_SKB_CB(skb)->sacked)
3968 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3971 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) {
3973 ack_ev_flags |= CA_ACK_ECE;
3976 if (sack_state.sack_delivered)
3977 tcp_count_delivered(tp, sack_state.sack_delivered,
3980 if (flag & FLAG_WIN_UPDATE)
3981 ack_ev_flags |= CA_ACK_WIN_UPDATE;
3983 tcp_in_ack_event(sk, ack_ev_flags);
3986 /* This is a deviation from RFC3168 since it states that:
3987 * "When the TCP data sender is ready to set the CWR bit after reducing
3988 * the congestion window, it SHOULD set the CWR bit only on the first
3989 * new data packet that it transmits."
3990 * We accept CWR on pure ACKs to be more robust
3991 * with widely-deployed TCP implementations that do this.
3993 tcp_ecn_accept_cwr(sk, skb);
3995 /* We passed data and got it acked, remove any soft error
3996 * log. Something worked...
3998 WRITE_ONCE(sk->sk_err_soft, 0);
3999 icsk->icsk_probes_out = 0;
4000 tp->rcv_tstamp = tcp_jiffies32;
4004 /* See if we can take anything off of the retransmit queue. */
4005 flag |= tcp_clean_rtx_queue(sk, skb, prior_fack, prior_snd_una,
4006 &sack_state, flag & FLAG_ECE);
4008 tcp_rack_update_reo_wnd(sk, &rs);
4010 if (tp->tlp_high_seq)
4011 tcp_process_tlp_ack(sk, ack, flag);
4013 if (tcp_ack_is_dubious(sk, flag)) {
4014 if (!(flag & (FLAG_SND_UNA_ADVANCED |
4015 FLAG_NOT_DUP | FLAG_DSACKING_ACK))) {
4017 /* Consider if pure acks were aggregated in tcp_add_backlog() */
4018 if (!(flag & FLAG_DATA))
4019 num_dupack = max_t(u16, 1, skb_shinfo(skb)->gso_segs);
4021 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
4025 /* If needed, reset TLP/RTO timer when RACK doesn't set. */
4026 if (flag & FLAG_SET_XMIT_TIMER)
4027 tcp_set_xmit_timer(sk);
4029 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
4032 delivered = tcp_newly_delivered(sk, delivered, flag);
4033 lost = tp->lost - lost; /* freshly marked lost */
4034 rs.is_ack_delayed = !!(flag & FLAG_ACK_MAYBE_DELAYED);
4035 tcp_rate_gen(sk, delivered, lost, is_sack_reneg, sack_state.rate);
4036 tcp_cong_control(sk, ack, delivered, flag, sack_state.rate);
4037 tcp_xmit_recovery(sk, rexmit);
4041 /* If data was DSACKed, see if we can undo a cwnd reduction. */
4042 if (flag & FLAG_DSACKING_ACK) {
4043 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
4045 tcp_newly_delivered(sk, delivered, flag);
4047 /* If this ack opens up a zero window, clear backoff. It was
4048 * being used to time the probes, and is probably far higher than
4049 * it needs to be for normal retransmission.
4053 if (tp->tlp_high_seq)
4054 tcp_process_tlp_ack(sk, ack, flag);
4058 /* If data was SACKed, tag it and see if we should send more data.
4059 * If data was DSACKed, see if we can undo a cwnd reduction.
4061 if (TCP_SKB_CB(skb)->sacked) {
4062 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
4064 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag,
4066 tcp_newly_delivered(sk, delivered, flag);
4067 tcp_xmit_recovery(sk, rexmit);
4073 static void tcp_parse_fastopen_option(int len, const unsigned char *cookie,
4074 bool syn, struct tcp_fastopen_cookie *foc,
4077 /* Valid only in SYN or SYN-ACK with an even length. */
4078 if (!foc || !syn || len < 0 || (len & 1))
4081 if (len >= TCP_FASTOPEN_COOKIE_MIN &&
4082 len <= TCP_FASTOPEN_COOKIE_MAX)
4083 memcpy(foc->val, cookie, len);
4090 static bool smc_parse_options(const struct tcphdr *th,
4091 struct tcp_options_received *opt_rx,
4092 const unsigned char *ptr,
4095 #if IS_ENABLED(CONFIG_SMC)
4096 if (static_branch_unlikely(&tcp_have_smc)) {
4097 if (th->syn && !(opsize & 1) &&
4098 opsize >= TCPOLEN_EXP_SMC_BASE &&
4099 get_unaligned_be32(ptr) == TCPOPT_SMC_MAGIC) {
4108 /* Try to parse the MSS option from the TCP header. Return 0 on failure, clamped
4111 u16 tcp_parse_mss_option(const struct tcphdr *th, u16 user_mss)
4113 const unsigned char *ptr = (const unsigned char *)(th + 1);
4114 int length = (th->doff * 4) - sizeof(struct tcphdr);
4117 while (length > 0) {
4118 int opcode = *ptr++;
4124 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
4131 if (opsize < 2) /* "silly options" */
4133 if (opsize > length)
4134 return mss; /* fail on partial options */
4135 if (opcode == TCPOPT_MSS && opsize == TCPOLEN_MSS) {
4136 u16 in_mss = get_unaligned_be16(ptr);
4139 if (user_mss && user_mss < in_mss)
4150 EXPORT_SYMBOL_GPL(tcp_parse_mss_option);
4152 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
4153 * But, this can also be called on packets in the established flow when
4154 * the fast version below fails.
4156 void tcp_parse_options(const struct net *net,
4157 const struct sk_buff *skb,
4158 struct tcp_options_received *opt_rx, int estab,
4159 struct tcp_fastopen_cookie *foc)
4161 const unsigned char *ptr;
4162 const struct tcphdr *th = tcp_hdr(skb);
4163 int length = (th->doff * 4) - sizeof(struct tcphdr);
4165 ptr = (const unsigned char *)(th + 1);
4166 opt_rx->saw_tstamp = 0;
4167 opt_rx->saw_unknown = 0;
4169 while (length > 0) {
4170 int opcode = *ptr++;
4176 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
4183 if (opsize < 2) /* "silly options" */
4185 if (opsize > length)
4186 return; /* don't parse partial options */
4189 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
4190 u16 in_mss = get_unaligned_be16(ptr);
4192 if (opt_rx->user_mss &&
4193 opt_rx->user_mss < in_mss)
4194 in_mss = opt_rx->user_mss;
4195 opt_rx->mss_clamp = in_mss;
4200 if (opsize == TCPOLEN_WINDOW && th->syn &&
4201 !estab && READ_ONCE(net->ipv4.sysctl_tcp_window_scaling)) {
4202 __u8 snd_wscale = *(__u8 *)ptr;
4203 opt_rx->wscale_ok = 1;
4204 if (snd_wscale > TCP_MAX_WSCALE) {
4205 net_info_ratelimited("%s: Illegal window scaling value %d > %u received\n",
4209 snd_wscale = TCP_MAX_WSCALE;
4211 opt_rx->snd_wscale = snd_wscale;
4214 case TCPOPT_TIMESTAMP:
4215 if ((opsize == TCPOLEN_TIMESTAMP) &&
4216 ((estab && opt_rx->tstamp_ok) ||
4217 (!estab && READ_ONCE(net->ipv4.sysctl_tcp_timestamps)))) {
4218 opt_rx->saw_tstamp = 1;
4219 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
4220 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
4223 case TCPOPT_SACK_PERM:
4224 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
4225 !estab && READ_ONCE(net->ipv4.sysctl_tcp_sack)) {
4226 opt_rx->sack_ok = TCP_SACK_SEEN;
4227 tcp_sack_reset(opt_rx);
4232 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
4233 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
4235 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
4238 #ifdef CONFIG_TCP_MD5SIG
4240 /* The MD5 Hash has already been
4241 * checked (see tcp_v{4,6}_rcv()).
4245 #ifdef CONFIG_TCP_AO
4247 /* TCP AO has already been checked
4248 * (see tcp_inbound_ao_hash()).
4252 case TCPOPT_FASTOPEN:
4253 tcp_parse_fastopen_option(
4254 opsize - TCPOLEN_FASTOPEN_BASE,
4255 ptr, th->syn, foc, false);
4259 /* Fast Open option shares code 254 using a
4260 * 16 bits magic number.
4262 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE &&
4263 get_unaligned_be16(ptr) ==
4264 TCPOPT_FASTOPEN_MAGIC) {
4265 tcp_parse_fastopen_option(opsize -
4266 TCPOLEN_EXP_FASTOPEN_BASE,
4267 ptr + 2, th->syn, foc, true);
4271 if (smc_parse_options(th, opt_rx, ptr, opsize))
4274 opt_rx->saw_unknown = 1;
4278 opt_rx->saw_unknown = 1;
4285 EXPORT_SYMBOL(tcp_parse_options);
4287 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
4289 const __be32 *ptr = (const __be32 *)(th + 1);
4291 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
4292 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
4293 tp->rx_opt.saw_tstamp = 1;
4295 tp->rx_opt.rcv_tsval = ntohl(*ptr);
4298 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
4300 tp->rx_opt.rcv_tsecr = 0;
4306 /* Fast parse options. This hopes to only see timestamps.
4307 * If it is wrong it falls back on tcp_parse_options().
4309 static bool tcp_fast_parse_options(const struct net *net,
4310 const struct sk_buff *skb,
4311 const struct tcphdr *th, struct tcp_sock *tp)
4313 /* In the spirit of fast parsing, compare doff directly to constant
4314 * values. Because equality is used, short doff can be ignored here.
4316 if (th->doff == (sizeof(*th) / 4)) {
4317 tp->rx_opt.saw_tstamp = 0;
4319 } else if (tp->rx_opt.tstamp_ok &&
4320 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
4321 if (tcp_parse_aligned_timestamp(tp, th))
4325 tcp_parse_options(net, skb, &tp->rx_opt, 1, NULL);
4326 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
4327 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
4332 #if defined(CONFIG_TCP_MD5SIG) || defined(CONFIG_TCP_AO)
4334 * Parse Signature options
4336 int tcp_do_parse_auth_options(const struct tcphdr *th,
4337 const u8 **md5_hash, const u8 **ao_hash)
4339 int length = (th->doff << 2) - sizeof(*th);
4340 const u8 *ptr = (const u8 *)(th + 1);
4341 unsigned int minlen = TCPOLEN_MD5SIG;
4343 if (IS_ENABLED(CONFIG_TCP_AO))
4344 minlen = sizeof(struct tcp_ao_hdr) + 1;
4349 /* If not enough data remaining, we can short cut */
4350 while (length >= minlen) {
4351 int opcode = *ptr++;
4362 if (opsize < 2 || opsize > length)
4364 if (opcode == TCPOPT_MD5SIG) {
4365 if (opsize != TCPOLEN_MD5SIG)
4367 if (unlikely(*md5_hash || *ao_hash))
4370 } else if (opcode == TCPOPT_AO) {
4371 if (opsize <= sizeof(struct tcp_ao_hdr))
4373 if (unlikely(*md5_hash || *ao_hash))
4383 EXPORT_SYMBOL(tcp_do_parse_auth_options);
4386 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
4388 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
4389 * it can pass through stack. So, the following predicate verifies that
4390 * this segment is not used for anything but congestion avoidance or
4391 * fast retransmit. Moreover, we even are able to eliminate most of such
4392 * second order effects, if we apply some small "replay" window (~RTO)
4393 * to timestamp space.
4395 * All these measures still do not guarantee that we reject wrapped ACKs
4396 * on networks with high bandwidth, when sequence space is recycled fastly,
4397 * but it guarantees that such events will be very rare and do not affect
4398 * connection seriously. This doesn't look nice, but alas, PAWS is really
4401 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
4402 * states that events when retransmit arrives after original data are rare.
4403 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
4404 * the biggest problem on large power networks even with minor reordering.
4405 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
4406 * up to bandwidth of 18Gigabit/sec. 8) ]
4409 /* Estimates max number of increments of remote peer TSval in
4410 * a replay window (based on our current RTO estimation).
4412 static u32 tcp_tsval_replay(const struct sock *sk)
4414 /* If we use usec TS resolution,
4415 * then expect the remote peer to use the same resolution.
4417 if (tcp_sk(sk)->tcp_usec_ts)
4418 return inet_csk(sk)->icsk_rto * (USEC_PER_SEC / HZ);
4420 /* RFC 7323 recommends a TSval clock between 1ms and 1sec.
4421 * We know that some OS (including old linux) can use 1200 Hz.
4423 return inet_csk(sk)->icsk_rto * 1200 / HZ;
4426 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
4428 const struct tcp_sock *tp = tcp_sk(sk);
4429 const struct tcphdr *th = tcp_hdr(skb);
4430 u32 seq = TCP_SKB_CB(skb)->seq;
4431 u32 ack = TCP_SKB_CB(skb)->ack_seq;
4433 return /* 1. Pure ACK with correct sequence number. */
4434 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
4436 /* 2. ... and duplicate ACK. */
4437 ack == tp->snd_una &&
4439 /* 3. ... and does not update window. */
4440 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
4442 /* 4. ... and sits in replay window. */
4443 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <=
4444 tcp_tsval_replay(sk);
4447 static inline bool tcp_paws_discard(const struct sock *sk,
4448 const struct sk_buff *skb)
4450 const struct tcp_sock *tp = tcp_sk(sk);
4452 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
4453 !tcp_disordered_ack(sk, skb);
4456 /* Check segment sequence number for validity.
4458 * Segment controls are considered valid, if the segment
4459 * fits to the window after truncation to the window. Acceptability
4460 * of data (and SYN, FIN, of course) is checked separately.
4461 * See tcp_data_queue(), for example.
4463 * Also, controls (RST is main one) are accepted using RCV.WUP instead
4464 * of RCV.NXT. Peer still did not advance his SND.UNA when we
4465 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4466 * (borrowed from freebsd)
4469 static enum skb_drop_reason tcp_sequence(const struct tcp_sock *tp,
4470 u32 seq, u32 end_seq)
4472 if (before(end_seq, tp->rcv_wup))
4473 return SKB_DROP_REASON_TCP_OLD_SEQUENCE;
4475 if (after(seq, tp->rcv_nxt + tcp_receive_window(tp)))
4476 return SKB_DROP_REASON_TCP_INVALID_SEQUENCE;
4478 return SKB_NOT_DROPPED_YET;
4482 void tcp_done_with_error(struct sock *sk, int err)
4484 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4485 WRITE_ONCE(sk->sk_err, err);
4488 tcp_write_queue_purge(sk);
4491 if (!sock_flag(sk, SOCK_DEAD))
4492 sk_error_report(sk);
4494 EXPORT_SYMBOL(tcp_done_with_error);
4496 /* When we get a reset we do this. */
4497 void tcp_reset(struct sock *sk, struct sk_buff *skb)
4501 trace_tcp_receive_reset(sk);
4503 /* mptcp can't tell us to ignore reset pkts,
4504 * so just ignore the return value of mptcp_incoming_options().
4506 if (sk_is_mptcp(sk))
4507 mptcp_incoming_options(sk, skb);
4509 /* We want the right error as BSD sees it (and indeed as we do). */
4510 switch (sk->sk_state) {
4514 case TCP_CLOSE_WAIT:
4522 tcp_done_with_error(sk, err);
4526 * Process the FIN bit. This now behaves as it is supposed to work
4527 * and the FIN takes effect when it is validly part of sequence
4528 * space. Not before when we get holes.
4530 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4531 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4534 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4535 * close and we go into CLOSING (and later onto TIME-WAIT)
4537 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4539 void tcp_fin(struct sock *sk)
4541 struct tcp_sock *tp = tcp_sk(sk);
4543 inet_csk_schedule_ack(sk);
4545 WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | RCV_SHUTDOWN);
4546 sock_set_flag(sk, SOCK_DONE);
4548 switch (sk->sk_state) {
4550 case TCP_ESTABLISHED:
4551 /* Move to CLOSE_WAIT */
4552 tcp_set_state(sk, TCP_CLOSE_WAIT);
4553 inet_csk_enter_pingpong_mode(sk);
4556 case TCP_CLOSE_WAIT:
4558 /* Received a retransmission of the FIN, do
4563 /* RFC793: Remain in the LAST-ACK state. */
4567 /* This case occurs when a simultaneous close
4568 * happens, we must ack the received FIN and
4569 * enter the CLOSING state.
4572 tcp_set_state(sk, TCP_CLOSING);
4575 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4577 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4580 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4581 * cases we should never reach this piece of code.
4583 pr_err("%s: Impossible, sk->sk_state=%d\n",
4584 __func__, sk->sk_state);
4588 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4589 * Probably, we should reset in this case. For now drop them.
4591 skb_rbtree_purge(&tp->out_of_order_queue);
4592 if (tcp_is_sack(tp))
4593 tcp_sack_reset(&tp->rx_opt);
4595 if (!sock_flag(sk, SOCK_DEAD)) {
4596 sk->sk_state_change(sk);
4598 /* Do not send POLL_HUP for half duplex close. */
4599 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4600 sk->sk_state == TCP_CLOSE)
4601 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4603 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4607 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4610 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4611 if (before(seq, sp->start_seq))
4612 sp->start_seq = seq;
4613 if (after(end_seq, sp->end_seq))
4614 sp->end_seq = end_seq;
4620 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4622 struct tcp_sock *tp = tcp_sk(sk);
4624 if (tcp_is_sack(tp) && READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_dsack)) {
4627 if (before(seq, tp->rcv_nxt))
4628 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4630 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4632 NET_INC_STATS(sock_net(sk), mib_idx);
4634 tp->rx_opt.dsack = 1;
4635 tp->duplicate_sack[0].start_seq = seq;
4636 tp->duplicate_sack[0].end_seq = end_seq;
4640 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4642 struct tcp_sock *tp = tcp_sk(sk);
4644 if (!tp->rx_opt.dsack)
4645 tcp_dsack_set(sk, seq, end_seq);
4647 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4650 static void tcp_rcv_spurious_retrans(struct sock *sk, const struct sk_buff *skb)
4652 /* When the ACK path fails or drops most ACKs, the sender would
4653 * timeout and spuriously retransmit the same segment repeatedly.
4654 * If it seems our ACKs are not reaching the other side,
4655 * based on receiving a duplicate data segment with new flowlabel
4656 * (suggesting the sender suffered an RTO), and we are not already
4657 * repathing due to our own RTO, then rehash the socket to repath our
4660 #if IS_ENABLED(CONFIG_IPV6)
4661 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Loss &&
4662 skb->protocol == htons(ETH_P_IPV6) &&
4663 (tcp_sk(sk)->inet_conn.icsk_ack.lrcv_flowlabel !=
4664 ntohl(ip6_flowlabel(ipv6_hdr(skb)))) &&
4665 sk_rethink_txhash(sk))
4666 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDUPLICATEDATAREHASH);
4668 /* Save last flowlabel after a spurious retrans. */
4669 tcp_save_lrcv_flowlabel(sk, skb);
4673 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4675 struct tcp_sock *tp = tcp_sk(sk);
4677 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4678 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4679 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4680 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
4682 if (tcp_is_sack(tp) && READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_dsack)) {
4683 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4685 tcp_rcv_spurious_retrans(sk, skb);
4686 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4687 end_seq = tp->rcv_nxt;
4688 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4695 /* These routines update the SACK block as out-of-order packets arrive or
4696 * in-order packets close up the sequence space.
4698 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4701 struct tcp_sack_block *sp = &tp->selective_acks[0];
4702 struct tcp_sack_block *swalk = sp + 1;
4704 /* See if the recent change to the first SACK eats into
4705 * or hits the sequence space of other SACK blocks, if so coalesce.
4707 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4708 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4711 /* Zap SWALK, by moving every further SACK up by one slot.
4712 * Decrease num_sacks.
4714 tp->rx_opt.num_sacks--;
4715 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4724 void tcp_sack_compress_send_ack(struct sock *sk)
4726 struct tcp_sock *tp = tcp_sk(sk);
4728 if (!tp->compressed_ack)
4731 if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1)
4734 /* Since we have to send one ack finally,
4735 * substract one from tp->compressed_ack to keep
4736 * LINUX_MIB_TCPACKCOMPRESSED accurate.
4738 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED,
4739 tp->compressed_ack - 1);
4741 tp->compressed_ack = 0;
4745 /* Reasonable amount of sack blocks included in TCP SACK option
4746 * The max is 4, but this becomes 3 if TCP timestamps are there.
4747 * Given that SACK packets might be lost, be conservative and use 2.
4749 #define TCP_SACK_BLOCKS_EXPECTED 2
4751 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4753 struct tcp_sock *tp = tcp_sk(sk);
4754 struct tcp_sack_block *sp = &tp->selective_acks[0];
4755 int cur_sacks = tp->rx_opt.num_sacks;
4761 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4762 if (tcp_sack_extend(sp, seq, end_seq)) {
4763 if (this_sack >= TCP_SACK_BLOCKS_EXPECTED)
4764 tcp_sack_compress_send_ack(sk);
4765 /* Rotate this_sack to the first one. */
4766 for (; this_sack > 0; this_sack--, sp--)
4767 swap(*sp, *(sp - 1));
4769 tcp_sack_maybe_coalesce(tp);
4774 if (this_sack >= TCP_SACK_BLOCKS_EXPECTED)
4775 tcp_sack_compress_send_ack(sk);
4777 /* Could not find an adjacent existing SACK, build a new one,
4778 * put it at the front, and shift everyone else down. We
4779 * always know there is at least one SACK present already here.
4781 * If the sack array is full, forget about the last one.
4783 if (this_sack >= TCP_NUM_SACKS) {
4785 tp->rx_opt.num_sacks--;
4788 for (; this_sack > 0; this_sack--, sp--)
4792 /* Build the new head SACK, and we're done. */
4793 sp->start_seq = seq;
4794 sp->end_seq = end_seq;
4795 tp->rx_opt.num_sacks++;
4798 /* RCV.NXT advances, some SACKs should be eaten. */
4800 static void tcp_sack_remove(struct tcp_sock *tp)
4802 struct tcp_sack_block *sp = &tp->selective_acks[0];
4803 int num_sacks = tp->rx_opt.num_sacks;
4806 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4807 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4808 tp->rx_opt.num_sacks = 0;
4812 for (this_sack = 0; this_sack < num_sacks;) {
4813 /* Check if the start of the sack is covered by RCV.NXT. */
4814 if (!before(tp->rcv_nxt, sp->start_seq)) {
4817 /* RCV.NXT must cover all the block! */
4818 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4820 /* Zap this SACK, by moving forward any other SACKS. */
4821 for (i = this_sack+1; i < num_sacks; i++)
4822 tp->selective_acks[i-1] = tp->selective_acks[i];
4829 tp->rx_opt.num_sacks = num_sacks;
4833 * tcp_try_coalesce - try to merge skb to prior one
4836 * @from: buffer to add in queue
4837 * @fragstolen: pointer to boolean
4839 * Before queueing skb @from after @to, try to merge them
4840 * to reduce overall memory use and queue lengths, if cost is small.
4841 * Packets in ofo or receive queues can stay a long time.
4842 * Better try to coalesce them right now to avoid future collapses.
4843 * Returns true if caller should free @from instead of queueing it
4845 static bool tcp_try_coalesce(struct sock *sk,
4847 struct sk_buff *from,
4852 *fragstolen = false;
4854 /* Its possible this segment overlaps with prior segment in queue */
4855 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4858 if (!tcp_skb_can_collapse_rx(to, from))
4861 if (!skb_try_coalesce(to, from, fragstolen, &delta))
4864 atomic_add(delta, &sk->sk_rmem_alloc);
4865 sk_mem_charge(sk, delta);
4866 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4867 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4868 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4869 TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
4871 if (TCP_SKB_CB(from)->has_rxtstamp) {
4872 TCP_SKB_CB(to)->has_rxtstamp = true;
4873 to->tstamp = from->tstamp;
4874 skb_hwtstamps(to)->hwtstamp = skb_hwtstamps(from)->hwtstamp;
4880 static bool tcp_ooo_try_coalesce(struct sock *sk,
4882 struct sk_buff *from,
4885 bool res = tcp_try_coalesce(sk, to, from, fragstolen);
4887 /* In case tcp_drop_reason() is called later, update to->gso_segs */
4889 u32 gso_segs = max_t(u16, 1, skb_shinfo(to)->gso_segs) +
4890 max_t(u16, 1, skb_shinfo(from)->gso_segs);
4892 skb_shinfo(to)->gso_segs = min_t(u32, gso_segs, 0xFFFF);
4897 static void tcp_drop_reason(struct sock *sk, struct sk_buff *skb,
4898 enum skb_drop_reason reason)
4900 sk_drops_add(sk, skb);
4901 sk_skb_reason_drop(sk, skb, reason);
4904 /* This one checks to see if we can put data from the
4905 * out_of_order queue into the receive_queue.
4907 static void tcp_ofo_queue(struct sock *sk)
4909 struct tcp_sock *tp = tcp_sk(sk);
4910 __u32 dsack_high = tp->rcv_nxt;
4911 bool fin, fragstolen, eaten;
4912 struct sk_buff *skb, *tail;
4915 p = rb_first(&tp->out_of_order_queue);
4918 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4921 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4922 __u32 dsack = dsack_high;
4923 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4924 dsack_high = TCP_SKB_CB(skb)->end_seq;
4925 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4928 rb_erase(&skb->rbnode, &tp->out_of_order_queue);
4930 if (unlikely(!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))) {
4931 tcp_drop_reason(sk, skb, SKB_DROP_REASON_TCP_OFO_DROP);
4935 tail = skb_peek_tail(&sk->sk_receive_queue);
4936 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen);
4937 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4938 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
4940 __skb_queue_tail(&sk->sk_receive_queue, skb);
4942 kfree_skb_partial(skb, fragstolen);
4944 if (unlikely(fin)) {
4946 /* tcp_fin() purges tp->out_of_order_queue,
4947 * so we must end this loop right now.
4954 static bool tcp_prune_ofo_queue(struct sock *sk, const struct sk_buff *in_skb);
4955 static int tcp_prune_queue(struct sock *sk, const struct sk_buff *in_skb);
4957 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4960 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4961 !sk_rmem_schedule(sk, skb, size)) {
4963 if (tcp_prune_queue(sk, skb) < 0)
4966 while (!sk_rmem_schedule(sk, skb, size)) {
4967 if (!tcp_prune_ofo_queue(sk, skb))
4974 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4976 struct tcp_sock *tp = tcp_sk(sk);
4977 struct rb_node **p, *parent;
4978 struct sk_buff *skb1;
4982 tcp_save_lrcv_flowlabel(sk, skb);
4983 tcp_ecn_check_ce(sk, skb);
4985 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4986 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFODROP);
4987 sk->sk_data_ready(sk);
4988 tcp_drop_reason(sk, skb, SKB_DROP_REASON_PROTO_MEM);
4992 /* Disable header prediction. */
4994 inet_csk_schedule_ack(sk);
4996 tp->rcv_ooopack += max_t(u16, 1, skb_shinfo(skb)->gso_segs);
4997 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4998 seq = TCP_SKB_CB(skb)->seq;
4999 end_seq = TCP_SKB_CB(skb)->end_seq;
5001 p = &tp->out_of_order_queue.rb_node;
5002 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
5003 /* Initial out of order segment, build 1 SACK. */
5004 if (tcp_is_sack(tp)) {
5005 tp->rx_opt.num_sacks = 1;
5006 tp->selective_acks[0].start_seq = seq;
5007 tp->selective_acks[0].end_seq = end_seq;
5009 rb_link_node(&skb->rbnode, NULL, p);
5010 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
5011 tp->ooo_last_skb = skb;
5015 /* In the typical case, we are adding an skb to the end of the list.
5016 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
5018 if (tcp_ooo_try_coalesce(sk, tp->ooo_last_skb,
5019 skb, &fragstolen)) {
5021 /* For non sack flows, do not grow window to force DUPACK
5022 * and trigger fast retransmit.
5024 if (tcp_is_sack(tp))
5025 tcp_grow_window(sk, skb, true);
5026 kfree_skb_partial(skb, fragstolen);
5030 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
5031 if (!before(seq, TCP_SKB_CB(tp->ooo_last_skb)->end_seq)) {
5032 parent = &tp->ooo_last_skb->rbnode;
5033 p = &parent->rb_right;
5037 /* Find place to insert this segment. Handle overlaps on the way. */
5041 skb1 = rb_to_skb(parent);
5042 if (before(seq, TCP_SKB_CB(skb1)->seq)) {
5043 p = &parent->rb_left;
5046 if (before(seq, TCP_SKB_CB(skb1)->end_seq)) {
5047 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
5048 /* All the bits are present. Drop. */
5049 NET_INC_STATS(sock_net(sk),
5050 LINUX_MIB_TCPOFOMERGE);
5051 tcp_drop_reason(sk, skb,
5052 SKB_DROP_REASON_TCP_OFOMERGE);
5054 tcp_dsack_set(sk, seq, end_seq);
5057 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
5058 /* Partial overlap. */
5059 tcp_dsack_set(sk, seq, TCP_SKB_CB(skb1)->end_seq);
5061 /* skb's seq == skb1's seq and skb covers skb1.
5062 * Replace skb1 with skb.
5064 rb_replace_node(&skb1->rbnode, &skb->rbnode,
5065 &tp->out_of_order_queue);
5066 tcp_dsack_extend(sk,
5067 TCP_SKB_CB(skb1)->seq,
5068 TCP_SKB_CB(skb1)->end_seq);
5069 NET_INC_STATS(sock_net(sk),
5070 LINUX_MIB_TCPOFOMERGE);
5071 tcp_drop_reason(sk, skb1,
5072 SKB_DROP_REASON_TCP_OFOMERGE);
5075 } else if (tcp_ooo_try_coalesce(sk, skb1,
5076 skb, &fragstolen)) {
5079 p = &parent->rb_right;
5082 /* Insert segment into RB tree. */
5083 rb_link_node(&skb->rbnode, parent, p);
5084 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
5087 /* Remove other segments covered by skb. */
5088 while ((skb1 = skb_rb_next(skb)) != NULL) {
5089 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
5091 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
5092 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
5096 rb_erase(&skb1->rbnode, &tp->out_of_order_queue);
5097 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
5098 TCP_SKB_CB(skb1)->end_seq);
5099 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
5100 tcp_drop_reason(sk, skb1, SKB_DROP_REASON_TCP_OFOMERGE);
5102 /* If there is no skb after us, we are the last_skb ! */
5104 tp->ooo_last_skb = skb;
5107 if (tcp_is_sack(tp))
5108 tcp_sack_new_ofo_skb(sk, seq, end_seq);
5111 /* For non sack flows, do not grow window to force DUPACK
5112 * and trigger fast retransmit.
5114 if (tcp_is_sack(tp))
5115 tcp_grow_window(sk, skb, false);
5117 skb_set_owner_r(skb, sk);
5121 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb,
5125 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
5128 tcp_try_coalesce(sk, tail,
5129 skb, fragstolen)) ? 1 : 0;
5130 tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq);
5132 __skb_queue_tail(&sk->sk_receive_queue, skb);
5133 skb_set_owner_r(skb, sk);
5138 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
5140 struct sk_buff *skb;
5148 if (size > PAGE_SIZE) {
5149 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS);
5151 data_len = npages << PAGE_SHIFT;
5152 size = data_len + (size & ~PAGE_MASK);
5154 skb = alloc_skb_with_frags(size - data_len, data_len,
5155 PAGE_ALLOC_COSTLY_ORDER,
5156 &err, sk->sk_allocation);
5160 skb_put(skb, size - data_len);
5161 skb->data_len = data_len;
5164 if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) {
5165 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP);
5169 err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size);
5173 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
5174 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
5175 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
5177 if (tcp_queue_rcv(sk, skb, &fragstolen)) {
5178 WARN_ON_ONCE(fragstolen); /* should not happen */
5190 void tcp_data_ready(struct sock *sk)
5192 if (tcp_epollin_ready(sk, sk->sk_rcvlowat) || sock_flag(sk, SOCK_DONE))
5193 sk->sk_data_ready(sk);
5196 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
5198 struct tcp_sock *tp = tcp_sk(sk);
5199 enum skb_drop_reason reason;
5203 /* If a subflow has been reset, the packet should not continue
5204 * to be processed, drop the packet.
5206 if (sk_is_mptcp(sk) && !mptcp_incoming_options(sk, skb)) {
5211 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
5216 __skb_pull(skb, tcp_hdr(skb)->doff * 4);
5218 reason = SKB_DROP_REASON_NOT_SPECIFIED;
5219 tp->rx_opt.dsack = 0;
5221 /* Queue data for delivery to the user.
5222 * Packets in sequence go to the receive queue.
5223 * Out of sequence packets to the out_of_order_queue.
5225 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
5226 if (tcp_receive_window(tp) == 0) {
5227 /* Some stacks are known to send bare FIN packets
5228 * in a loop even if we send RWIN 0 in our ACK.
5229 * Accepting this FIN does not hurt memory pressure
5230 * because the FIN flag will simply be merged to the
5231 * receive queue tail skb in most cases.
5234 (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
5237 reason = SKB_DROP_REASON_TCP_ZEROWINDOW;
5238 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP);
5242 /* Ok. In sequence. In window. */
5244 if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) {
5245 /* TODO: maybe ratelimit these WIN 0 ACK ? */
5246 inet_csk(sk)->icsk_ack.pending |=
5247 (ICSK_ACK_NOMEM | ICSK_ACK_NOW);
5248 inet_csk_schedule_ack(sk);
5249 sk->sk_data_ready(sk);
5251 if (skb_queue_len(&sk->sk_receive_queue) && skb->len) {
5252 reason = SKB_DROP_REASON_PROTO_MEM;
5253 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP);
5256 sk_forced_mem_schedule(sk, skb->truesize);
5259 eaten = tcp_queue_rcv(sk, skb, &fragstolen);
5261 tcp_event_data_recv(sk, skb);
5262 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
5265 if (!RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
5268 /* RFC5681. 4.2. SHOULD send immediate ACK, when
5269 * gap in queue is filled.
5271 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
5272 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW;
5275 if (tp->rx_opt.num_sacks)
5276 tcp_sack_remove(tp);
5278 tcp_fast_path_check(sk);
5281 kfree_skb_partial(skb, fragstolen);
5282 if (!sock_flag(sk, SOCK_DEAD))
5287 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
5288 tcp_rcv_spurious_retrans(sk, skb);
5289 /* A retransmit, 2nd most common case. Force an immediate ack. */
5290 reason = SKB_DROP_REASON_TCP_OLD_DATA;
5291 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
5292 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
5295 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
5296 inet_csk_schedule_ack(sk);
5298 tcp_drop_reason(sk, skb, reason);
5302 /* Out of window. F.e. zero window probe. */
5303 if (!before(TCP_SKB_CB(skb)->seq,
5304 tp->rcv_nxt + tcp_receive_window(tp))) {
5305 reason = SKB_DROP_REASON_TCP_OVERWINDOW;
5309 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
5310 /* Partial packet, seq < rcv_next < end_seq */
5311 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
5313 /* If window is closed, drop tail of packet. But after
5314 * remembering D-SACK for its head made in previous line.
5316 if (!tcp_receive_window(tp)) {
5317 reason = SKB_DROP_REASON_TCP_ZEROWINDOW;
5318 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP);
5324 tcp_data_queue_ofo(sk, skb);
5327 static struct sk_buff *tcp_skb_next(struct sk_buff *skb, struct sk_buff_head *list)
5330 return !skb_queue_is_last(list, skb) ? skb->next : NULL;
5332 return skb_rb_next(skb);
5335 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
5336 struct sk_buff_head *list,
5337 struct rb_root *root)
5339 struct sk_buff *next = tcp_skb_next(skb, list);
5342 __skb_unlink(skb, list);
5344 rb_erase(&skb->rbnode, root);
5347 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
5352 /* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */
5353 void tcp_rbtree_insert(struct rb_root *root, struct sk_buff *skb)
5355 struct rb_node **p = &root->rb_node;
5356 struct rb_node *parent = NULL;
5357 struct sk_buff *skb1;
5361 skb1 = rb_to_skb(parent);
5362 if (before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb1)->seq))
5363 p = &parent->rb_left;
5365 p = &parent->rb_right;
5367 rb_link_node(&skb->rbnode, parent, p);
5368 rb_insert_color(&skb->rbnode, root);
5371 /* Collapse contiguous sequence of skbs head..tail with
5372 * sequence numbers start..end.
5374 * If tail is NULL, this means until the end of the queue.
5376 * Segments with FIN/SYN are not collapsed (only because this
5380 tcp_collapse(struct sock *sk, struct sk_buff_head *list, struct rb_root *root,
5381 struct sk_buff *head, struct sk_buff *tail, u32 start, u32 end)
5383 struct sk_buff *skb = head, *n;
5384 struct sk_buff_head tmp;
5387 /* First, check that queue is collapsible and find
5388 * the point where collapsing can be useful.
5391 for (end_of_skbs = true; skb != NULL && skb != tail; skb = n) {
5392 n = tcp_skb_next(skb, list);
5394 /* No new bits? It is possible on ofo queue. */
5395 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
5396 skb = tcp_collapse_one(sk, skb, list, root);
5402 /* The first skb to collapse is:
5404 * - bloated or contains data before "start" or
5405 * overlaps to the next one and mptcp allow collapsing.
5407 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
5408 (tcp_win_from_space(sk, skb->truesize) > skb->len ||
5409 before(TCP_SKB_CB(skb)->seq, start))) {
5410 end_of_skbs = false;
5414 if (n && n != tail && tcp_skb_can_collapse_rx(skb, n) &&
5415 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(n)->seq) {
5416 end_of_skbs = false;
5420 /* Decided to skip this, advance start seq. */
5421 start = TCP_SKB_CB(skb)->end_seq;
5424 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
5427 __skb_queue_head_init(&tmp);
5429 while (before(start, end)) {
5430 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
5431 struct sk_buff *nskb;
5433 nskb = alloc_skb(copy, GFP_ATOMIC);
5437 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
5438 skb_copy_decrypted(nskb, skb);
5439 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
5441 __skb_queue_before(list, skb, nskb);
5443 __skb_queue_tail(&tmp, nskb); /* defer rbtree insertion */
5444 skb_set_owner_r(nskb, sk);
5445 mptcp_skb_ext_move(nskb, skb);
5447 /* Copy data, releasing collapsed skbs. */
5449 int offset = start - TCP_SKB_CB(skb)->seq;
5450 int size = TCP_SKB_CB(skb)->end_seq - start;
5454 size = min(copy, size);
5455 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
5457 TCP_SKB_CB(nskb)->end_seq += size;
5461 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
5462 skb = tcp_collapse_one(sk, skb, list, root);
5465 !tcp_skb_can_collapse_rx(nskb, skb) ||
5466 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
5472 skb_queue_walk_safe(&tmp, skb, n)
5473 tcp_rbtree_insert(root, skb);
5476 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
5477 * and tcp_collapse() them until all the queue is collapsed.
5479 static void tcp_collapse_ofo_queue(struct sock *sk)
5481 struct tcp_sock *tp = tcp_sk(sk);
5482 u32 range_truesize, sum_tiny = 0;
5483 struct sk_buff *skb, *head;
5486 skb = skb_rb_first(&tp->out_of_order_queue);
5489 tp->ooo_last_skb = skb_rb_last(&tp->out_of_order_queue);
5492 start = TCP_SKB_CB(skb)->seq;
5493 end = TCP_SKB_CB(skb)->end_seq;
5494 range_truesize = skb->truesize;
5496 for (head = skb;;) {
5497 skb = skb_rb_next(skb);
5499 /* Range is terminated when we see a gap or when
5500 * we are at the queue end.
5503 after(TCP_SKB_CB(skb)->seq, end) ||
5504 before(TCP_SKB_CB(skb)->end_seq, start)) {
5505 /* Do not attempt collapsing tiny skbs */
5506 if (range_truesize != head->truesize ||
5507 end - start >= SKB_WITH_OVERHEAD(PAGE_SIZE)) {
5508 tcp_collapse(sk, NULL, &tp->out_of_order_queue,
5509 head, skb, start, end);
5511 sum_tiny += range_truesize;
5512 if (sum_tiny > sk->sk_rcvbuf >> 3)
5518 range_truesize += skb->truesize;
5519 if (unlikely(before(TCP_SKB_CB(skb)->seq, start)))
5520 start = TCP_SKB_CB(skb)->seq;
5521 if (after(TCP_SKB_CB(skb)->end_seq, end))
5522 end = TCP_SKB_CB(skb)->end_seq;
5527 * Clean the out-of-order queue to make room.
5528 * We drop high sequences packets to :
5529 * 1) Let a chance for holes to be filled.
5530 * This means we do not drop packets from ooo queue if their sequence
5531 * is before incoming packet sequence.
5532 * 2) not add too big latencies if thousands of packets sit there.
5533 * (But if application shrinks SO_RCVBUF, we could still end up
5534 * freeing whole queue here)
5535 * 3) Drop at least 12.5 % of sk_rcvbuf to avoid malicious attacks.
5537 * Return true if queue has shrunk.
5539 static bool tcp_prune_ofo_queue(struct sock *sk, const struct sk_buff *in_skb)
5541 struct tcp_sock *tp = tcp_sk(sk);
5542 struct rb_node *node, *prev;
5543 bool pruned = false;
5546 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
5549 goal = sk->sk_rcvbuf >> 3;
5550 node = &tp->ooo_last_skb->rbnode;
5553 struct sk_buff *skb = rb_to_skb(node);
5555 /* If incoming skb would land last in ofo queue, stop pruning. */
5556 if (after(TCP_SKB_CB(in_skb)->seq, TCP_SKB_CB(skb)->seq))
5559 prev = rb_prev(node);
5560 rb_erase(node, &tp->out_of_order_queue);
5561 goal -= skb->truesize;
5562 tcp_drop_reason(sk, skb, SKB_DROP_REASON_TCP_OFO_QUEUE_PRUNE);
5563 tp->ooo_last_skb = rb_to_skb(prev);
5564 if (!prev || goal <= 0) {
5565 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf &&
5566 !tcp_under_memory_pressure(sk))
5568 goal = sk->sk_rcvbuf >> 3;
5574 NET_INC_STATS(sock_net(sk), LINUX_MIB_OFOPRUNED);
5575 /* Reset SACK state. A conforming SACK implementation will
5576 * do the same at a timeout based retransmit. When a connection
5577 * is in a sad state like this, we care only about integrity
5578 * of the connection not performance.
5580 if (tp->rx_opt.sack_ok)
5581 tcp_sack_reset(&tp->rx_opt);
5586 /* Reduce allocated memory if we can, trying to get
5587 * the socket within its memory limits again.
5589 * Return less than zero if we should start dropping frames
5590 * until the socket owning process reads some of the data
5591 * to stabilize the situation.
5593 static int tcp_prune_queue(struct sock *sk, const struct sk_buff *in_skb)
5595 struct tcp_sock *tp = tcp_sk(sk);
5597 NET_INC_STATS(sock_net(sk), LINUX_MIB_PRUNECALLED);
5599 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
5600 tcp_clamp_window(sk);
5601 else if (tcp_under_memory_pressure(sk))
5602 tcp_adjust_rcv_ssthresh(sk);
5604 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5607 tcp_collapse_ofo_queue(sk);
5608 if (!skb_queue_empty(&sk->sk_receive_queue))
5609 tcp_collapse(sk, &sk->sk_receive_queue, NULL,
5610 skb_peek(&sk->sk_receive_queue),
5612 tp->copied_seq, tp->rcv_nxt);
5614 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5617 /* Collapsing did not help, destructive actions follow.
5618 * This must not ever occur. */
5620 tcp_prune_ofo_queue(sk, in_skb);
5622 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5625 /* If we are really being abused, tell the caller to silently
5626 * drop receive data on the floor. It will get retransmitted
5627 * and hopefully then we'll have sufficient space.
5629 NET_INC_STATS(sock_net(sk), LINUX_MIB_RCVPRUNED);
5631 /* Massive buffer overcommit. */
5636 static bool tcp_should_expand_sndbuf(struct sock *sk)
5638 const struct tcp_sock *tp = tcp_sk(sk);
5640 /* If the user specified a specific send buffer setting, do
5643 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
5646 /* If we are under global TCP memory pressure, do not expand. */
5647 if (tcp_under_memory_pressure(sk)) {
5648 int unused_mem = sk_unused_reserved_mem(sk);
5650 /* Adjust sndbuf according to reserved mem. But make sure
5651 * it never goes below SOCK_MIN_SNDBUF.
5652 * See sk_stream_moderate_sndbuf() for more details.
5654 if (unused_mem > SOCK_MIN_SNDBUF)
5655 WRITE_ONCE(sk->sk_sndbuf, unused_mem);
5660 /* If we are under soft global TCP memory pressure, do not expand. */
5661 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
5664 /* If we filled the congestion window, do not expand. */
5665 if (tcp_packets_in_flight(tp) >= tcp_snd_cwnd(tp))
5671 static void tcp_new_space(struct sock *sk)
5673 struct tcp_sock *tp = tcp_sk(sk);
5675 if (tcp_should_expand_sndbuf(sk)) {
5676 tcp_sndbuf_expand(sk);
5677 tp->snd_cwnd_stamp = tcp_jiffies32;
5680 INDIRECT_CALL_1(sk->sk_write_space, sk_stream_write_space, sk);
5683 /* Caller made space either from:
5684 * 1) Freeing skbs in rtx queues (after tp->snd_una has advanced)
5685 * 2) Sent skbs from output queue (and thus advancing tp->snd_nxt)
5687 * We might be able to generate EPOLLOUT to the application if:
5688 * 1) Space consumed in output/rtx queues is below sk->sk_sndbuf/2
5689 * 2) notsent amount (tp->write_seq - tp->snd_nxt) became
5690 * small enough that tcp_stream_memory_free() decides it
5691 * is time to generate EPOLLOUT.
5693 void tcp_check_space(struct sock *sk)
5695 /* pairs with tcp_poll() */
5697 if (sk->sk_socket &&
5698 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
5700 if (!test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
5701 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
5705 static inline void tcp_data_snd_check(struct sock *sk)
5707 tcp_push_pending_frames(sk);
5708 tcp_check_space(sk);
5712 * Check if sending an ack is needed.
5714 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
5716 struct tcp_sock *tp = tcp_sk(sk);
5717 unsigned long rtt, delay;
5719 /* More than one full frame received... */
5720 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
5721 /* ... and right edge of window advances far enough.
5722 * (tcp_recvmsg() will send ACK otherwise).
5723 * If application uses SO_RCVLOWAT, we want send ack now if
5724 * we have not received enough bytes to satisfy the condition.
5726 (tp->rcv_nxt - tp->copied_seq < sk->sk_rcvlowat ||
5727 __tcp_select_window(sk) >= tp->rcv_wnd)) ||
5728 /* We ACK each frame or... */
5729 tcp_in_quickack_mode(sk) ||
5730 /* Protocol state mandates a one-time immediate ACK */
5731 inet_csk(sk)->icsk_ack.pending & ICSK_ACK_NOW) {
5732 /* If we are running from __release_sock() in user context,
5733 * Defer the ack until tcp_release_cb().
5735 if (sock_owned_by_user_nocheck(sk) &&
5736 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_backlog_ack_defer)) {
5737 set_bit(TCP_ACK_DEFERRED, &sk->sk_tsq_flags);
5745 if (!ofo_possible || RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
5746 tcp_send_delayed_ack(sk);
5750 if (!tcp_is_sack(tp) ||
5751 tp->compressed_ack >= READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_nr))
5754 if (tp->compressed_ack_rcv_nxt != tp->rcv_nxt) {
5755 tp->compressed_ack_rcv_nxt = tp->rcv_nxt;
5756 tp->dup_ack_counter = 0;
5758 if (tp->dup_ack_counter < TCP_FASTRETRANS_THRESH) {
5759 tp->dup_ack_counter++;
5762 tp->compressed_ack++;
5763 if (hrtimer_is_queued(&tp->compressed_ack_timer))
5766 /* compress ack timer : 5 % of rtt, but no more than tcp_comp_sack_delay_ns */
5768 rtt = tp->rcv_rtt_est.rtt_us;
5769 if (tp->srtt_us && tp->srtt_us < rtt)
5772 delay = min_t(unsigned long,
5773 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_delay_ns),
5774 rtt * (NSEC_PER_USEC >> 3)/20);
5776 hrtimer_start_range_ns(&tp->compressed_ack_timer, ns_to_ktime(delay),
5777 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_slack_ns),
5778 HRTIMER_MODE_REL_PINNED_SOFT);
5781 static inline void tcp_ack_snd_check(struct sock *sk)
5783 if (!inet_csk_ack_scheduled(sk)) {
5784 /* We sent a data segment already. */
5787 __tcp_ack_snd_check(sk, 1);
5791 * This routine is only called when we have urgent data
5792 * signaled. Its the 'slow' part of tcp_urg. It could be
5793 * moved inline now as tcp_urg is only called from one
5794 * place. We handle URGent data wrong. We have to - as
5795 * BSD still doesn't use the correction from RFC961.
5796 * For 1003.1g we should support a new option TCP_STDURG to permit
5797 * either form (or just set the sysctl tcp_stdurg).
5800 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
5802 struct tcp_sock *tp = tcp_sk(sk);
5803 u32 ptr = ntohs(th->urg_ptr);
5805 if (ptr && !READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_stdurg))
5807 ptr += ntohl(th->seq);
5809 /* Ignore urgent data that we've already seen and read. */
5810 if (after(tp->copied_seq, ptr))
5813 /* Do not replay urg ptr.
5815 * NOTE: interesting situation not covered by specs.
5816 * Misbehaving sender may send urg ptr, pointing to segment,
5817 * which we already have in ofo queue. We are not able to fetch
5818 * such data and will stay in TCP_URG_NOTYET until will be eaten
5819 * by recvmsg(). Seems, we are not obliged to handle such wicked
5820 * situations. But it is worth to think about possibility of some
5821 * DoSes using some hypothetical application level deadlock.
5823 if (before(ptr, tp->rcv_nxt))
5826 /* Do we already have a newer (or duplicate) urgent pointer? */
5827 if (tp->urg_data && !after(ptr, tp->urg_seq))
5830 /* Tell the world about our new urgent pointer. */
5833 /* We may be adding urgent data when the last byte read was
5834 * urgent. To do this requires some care. We cannot just ignore
5835 * tp->copied_seq since we would read the last urgent byte again
5836 * as data, nor can we alter copied_seq until this data arrives
5837 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5839 * NOTE. Double Dutch. Rendering to plain English: author of comment
5840 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5841 * and expect that both A and B disappear from stream. This is _wrong_.
5842 * Though this happens in BSD with high probability, this is occasional.
5843 * Any application relying on this is buggy. Note also, that fix "works"
5844 * only in this artificial test. Insert some normal data between A and B and we will
5845 * decline of BSD again. Verdict: it is better to remove to trap
5848 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5849 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5850 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5852 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5853 __skb_unlink(skb, &sk->sk_receive_queue);
5858 WRITE_ONCE(tp->urg_data, TCP_URG_NOTYET);
5859 WRITE_ONCE(tp->urg_seq, ptr);
5861 /* Disable header prediction. */
5865 /* This is the 'fast' part of urgent handling. */
5866 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
5868 struct tcp_sock *tp = tcp_sk(sk);
5870 /* Check if we get a new urgent pointer - normally not. */
5871 if (unlikely(th->urg))
5872 tcp_check_urg(sk, th);
5874 /* Do we wait for any urgent data? - normally not... */
5875 if (unlikely(tp->urg_data == TCP_URG_NOTYET)) {
5876 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5879 /* Is the urgent pointer pointing into this packet? */
5880 if (ptr < skb->len) {
5882 if (skb_copy_bits(skb, ptr, &tmp, 1))
5884 WRITE_ONCE(tp->urg_data, TCP_URG_VALID | tmp);
5885 if (!sock_flag(sk, SOCK_DEAD))
5886 sk->sk_data_ready(sk);
5891 /* Accept RST for rcv_nxt - 1 after a FIN.
5892 * When tcp connections are abruptly terminated from Mac OSX (via ^C), a
5893 * FIN is sent followed by a RST packet. The RST is sent with the same
5894 * sequence number as the FIN, and thus according to RFC 5961 a challenge
5895 * ACK should be sent. However, Mac OSX rate limits replies to challenge
5896 * ACKs on the closed socket. In addition middleboxes can drop either the
5897 * challenge ACK or a subsequent RST.
5899 static bool tcp_reset_check(const struct sock *sk, const struct sk_buff *skb)
5901 const struct tcp_sock *tp = tcp_sk(sk);
5903 return unlikely(TCP_SKB_CB(skb)->seq == (tp->rcv_nxt - 1) &&
5904 (1 << sk->sk_state) & (TCPF_CLOSE_WAIT | TCPF_LAST_ACK |
5908 /* Does PAWS and seqno based validation of an incoming segment, flags will
5909 * play significant role here.
5911 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5912 const struct tcphdr *th, int syn_inerr)
5914 struct tcp_sock *tp = tcp_sk(sk);
5917 /* RFC1323: H1. Apply PAWS check first. */
5918 if (tcp_fast_parse_options(sock_net(sk), skb, th, tp) &&
5919 tp->rx_opt.saw_tstamp &&
5920 tcp_paws_discard(sk, skb)) {
5922 if (unlikely(th->syn))
5924 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5925 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5926 LINUX_MIB_TCPACKSKIPPEDPAWS,
5927 &tp->last_oow_ack_time))
5928 tcp_send_dupack(sk, skb);
5929 SKB_DR_SET(reason, TCP_RFC7323_PAWS);
5932 /* Reset is accepted even if it did not pass PAWS. */
5935 /* Step 1: check sequence number */
5936 reason = tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
5938 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5939 * (RST) segments are validated by checking their SEQ-fields."
5940 * And page 69: "If an incoming segment is not acceptable,
5941 * an acknowledgment should be sent in reply (unless the RST
5942 * bit is set, if so drop the segment and return)".
5947 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5948 LINUX_MIB_TCPACKSKIPPEDSEQ,
5949 &tp->last_oow_ack_time))
5950 tcp_send_dupack(sk, skb);
5951 } else if (tcp_reset_check(sk, skb)) {
5957 /* Step 2: check RST bit */
5959 /* RFC 5961 3.2 (extend to match against (RCV.NXT - 1) after a
5960 * FIN and SACK too if available):
5961 * If seq num matches RCV.NXT or (RCV.NXT - 1) after a FIN, or
5962 * the right-most SACK block,
5964 * RESET the connection
5966 * Send a challenge ACK
5968 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt ||
5969 tcp_reset_check(sk, skb))
5972 if (tcp_is_sack(tp) && tp->rx_opt.num_sacks > 0) {
5973 struct tcp_sack_block *sp = &tp->selective_acks[0];
5974 int max_sack = sp[0].end_seq;
5977 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;
5979 max_sack = after(sp[this_sack].end_seq,
5981 sp[this_sack].end_seq : max_sack;
5984 if (TCP_SKB_CB(skb)->seq == max_sack)
5988 /* Disable TFO if RST is out-of-order
5989 * and no data has been received
5990 * for current active TFO socket
5992 if (tp->syn_fastopen && !tp->data_segs_in &&
5993 sk->sk_state == TCP_ESTABLISHED)
5994 tcp_fastopen_active_disable(sk);
5995 tcp_send_challenge_ack(sk);
5996 SKB_DR_SET(reason, TCP_RESET);
6000 /* step 3: check security and precedence [ignored] */
6002 /* step 4: Check for a SYN
6003 * RFC 5961 4.2 : Send a challenge ack
6006 if (sk->sk_state == TCP_SYN_RECV && sk->sk_socket && th->ack &&
6007 TCP_SKB_CB(skb)->seq + 1 == TCP_SKB_CB(skb)->end_seq &&
6008 TCP_SKB_CB(skb)->seq + 1 == tp->rcv_nxt &&
6009 TCP_SKB_CB(skb)->ack_seq == tp->snd_nxt)
6013 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
6014 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
6015 tcp_send_challenge_ack(sk);
6016 SKB_DR_SET(reason, TCP_INVALID_SYN);
6021 bpf_skops_parse_hdr(sk, skb);
6026 tcp_drop_reason(sk, skb, reason);
6036 * TCP receive function for the ESTABLISHED state.
6038 * It is split into a fast path and a slow path. The fast path is
6040 * - A zero window was announced from us - zero window probing
6041 * is only handled properly in the slow path.
6042 * - Out of order segments arrived.
6043 * - Urgent data is expected.
6044 * - There is no buffer space left
6045 * - Unexpected TCP flags/window values/header lengths are received
6046 * (detected by checking the TCP header against pred_flags)
6047 * - Data is sent in both directions. Fast path only supports pure senders
6048 * or pure receivers (this means either the sequence number or the ack
6049 * value must stay constant)
6050 * - Unexpected TCP option.
6052 * When these conditions are not satisfied it drops into a standard
6053 * receive procedure patterned after RFC793 to handle all cases.
6054 * The first three cases are guaranteed by proper pred_flags setting,
6055 * the rest is checked inline. Fast processing is turned on in
6056 * tcp_data_queue when everything is OK.
6058 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb)
6060 enum skb_drop_reason reason = SKB_DROP_REASON_NOT_SPECIFIED;
6061 const struct tcphdr *th = (const struct tcphdr *)skb->data;
6062 struct tcp_sock *tp = tcp_sk(sk);
6063 unsigned int len = skb->len;
6065 /* TCP congestion window tracking */
6066 trace_tcp_probe(sk, skb);
6068 tcp_mstamp_refresh(tp);
6069 if (unlikely(!rcu_access_pointer(sk->sk_rx_dst)))
6070 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
6072 * Header prediction.
6073 * The code loosely follows the one in the famous
6074 * "30 instruction TCP receive" Van Jacobson mail.
6076 * Van's trick is to deposit buffers into socket queue
6077 * on a device interrupt, to call tcp_recv function
6078 * on the receive process context and checksum and copy
6079 * the buffer to user space. smart...
6081 * Our current scheme is not silly either but we take the
6082 * extra cost of the net_bh soft interrupt processing...
6083 * We do checksum and copy also but from device to kernel.
6086 tp->rx_opt.saw_tstamp = 0;
6088 /* pred_flags is 0xS?10 << 16 + snd_wnd
6089 * if header_prediction is to be made
6090 * 'S' will always be tp->tcp_header_len >> 2
6091 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
6092 * turn it off (when there are holes in the receive
6093 * space for instance)
6094 * PSH flag is ignored.
6097 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
6098 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
6099 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
6100 int tcp_header_len = tp->tcp_header_len;
6102 /* Timestamp header prediction: tcp_header_len
6103 * is automatically equal to th->doff*4 due to pred_flags
6107 /* Check timestamp */
6108 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
6109 /* No? Slow path! */
6110 if (!tcp_parse_aligned_timestamp(tp, th))
6113 /* If PAWS failed, check it more carefully in slow path */
6114 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
6117 /* DO NOT update ts_recent here, if checksum fails
6118 * and timestamp was corrupted part, it will result
6119 * in a hung connection since we will drop all
6120 * future packets due to the PAWS test.
6124 if (len <= tcp_header_len) {
6125 /* Bulk data transfer: sender */
6126 if (len == tcp_header_len) {
6127 /* Predicted packet is in window by definition.
6128 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
6129 * Hence, check seq<=rcv_wup reduces to:
6131 if (tcp_header_len ==
6132 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
6133 tp->rcv_nxt == tp->rcv_wup)
6134 tcp_store_ts_recent(tp);
6136 /* We know that such packets are checksummed
6139 tcp_ack(sk, skb, 0);
6141 tcp_data_snd_check(sk);
6142 /* When receiving pure ack in fast path, update
6143 * last ts ecr directly instead of calling
6144 * tcp_rcv_rtt_measure_ts()
6146 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr;
6148 } else { /* Header too small */
6149 reason = SKB_DROP_REASON_PKT_TOO_SMALL;
6150 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
6155 bool fragstolen = false;
6157 if (tcp_checksum_complete(skb))
6160 if ((int)skb->truesize > sk->sk_forward_alloc)
6163 /* Predicted packet is in window by definition.
6164 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
6165 * Hence, check seq<=rcv_wup reduces to:
6167 if (tcp_header_len ==
6168 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
6169 tp->rcv_nxt == tp->rcv_wup)
6170 tcp_store_ts_recent(tp);
6172 tcp_rcv_rtt_measure_ts(sk, skb);
6174 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPHITS);
6176 /* Bulk data transfer: receiver */
6178 __skb_pull(skb, tcp_header_len);
6179 eaten = tcp_queue_rcv(sk, skb, &fragstolen);
6181 tcp_event_data_recv(sk, skb);
6183 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
6184 /* Well, only one small jumplet in fast path... */
6185 tcp_ack(sk, skb, FLAG_DATA);
6186 tcp_data_snd_check(sk);
6187 if (!inet_csk_ack_scheduled(sk))
6190 tcp_update_wl(tp, TCP_SKB_CB(skb)->seq);
6193 __tcp_ack_snd_check(sk, 0);
6196 kfree_skb_partial(skb, fragstolen);
6203 if (len < (th->doff << 2) || tcp_checksum_complete(skb))
6206 if (!th->ack && !th->rst && !th->syn) {
6207 reason = SKB_DROP_REASON_TCP_FLAGS;
6212 * Standard slow path.
6215 if (!tcp_validate_incoming(sk, skb, th, 1))
6219 reason = tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT);
6220 if ((int)reason < 0) {
6224 tcp_rcv_rtt_measure_ts(sk, skb);
6226 /* Process urgent data. */
6227 tcp_urg(sk, skb, th);
6229 /* step 7: process the segment text */
6230 tcp_data_queue(sk, skb);
6232 tcp_data_snd_check(sk);
6233 tcp_ack_snd_check(sk);
6237 reason = SKB_DROP_REASON_TCP_CSUM;
6238 trace_tcp_bad_csum(skb);
6239 TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS);
6240 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
6243 tcp_drop_reason(sk, skb, reason);
6245 EXPORT_SYMBOL(tcp_rcv_established);
6247 void tcp_init_transfer(struct sock *sk, int bpf_op, struct sk_buff *skb)
6249 struct inet_connection_sock *icsk = inet_csk(sk);
6250 struct tcp_sock *tp = tcp_sk(sk);
6253 icsk->icsk_af_ops->rebuild_header(sk);
6254 tcp_init_metrics(sk);
6256 /* Initialize the congestion window to start the transfer.
6257 * Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
6258 * retransmitted. In light of RFC6298 more aggressive 1sec
6259 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
6260 * retransmission has occurred.
6262 if (tp->total_retrans > 1 && tp->undo_marker)
6263 tcp_snd_cwnd_set(tp, 1);
6265 tcp_snd_cwnd_set(tp, tcp_init_cwnd(tp, __sk_dst_get(sk)));
6266 tp->snd_cwnd_stamp = tcp_jiffies32;
6268 bpf_skops_established(sk, bpf_op, skb);
6269 /* Initialize congestion control unless BPF initialized it already: */
6270 if (!icsk->icsk_ca_initialized)
6271 tcp_init_congestion_control(sk);
6272 tcp_init_buffer_space(sk);
6275 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
6277 struct tcp_sock *tp = tcp_sk(sk);
6278 struct inet_connection_sock *icsk = inet_csk(sk);
6280 tcp_ao_finish_connect(sk, skb);
6281 tcp_set_state(sk, TCP_ESTABLISHED);
6282 icsk->icsk_ack.lrcvtime = tcp_jiffies32;
6285 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
6286 security_inet_conn_established(sk, skb);
6287 sk_mark_napi_id(sk, skb);
6290 tcp_init_transfer(sk, BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB, skb);
6292 /* Prevent spurious tcp_cwnd_restart() on first data
6295 tp->lsndtime = tcp_jiffies32;
6297 if (sock_flag(sk, SOCK_KEEPOPEN))
6298 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
6300 if (!tp->rx_opt.snd_wscale)
6301 __tcp_fast_path_on(tp, tp->snd_wnd);
6306 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
6307 struct tcp_fastopen_cookie *cookie)
6309 struct tcp_sock *tp = tcp_sk(sk);
6310 struct sk_buff *data = tp->syn_data ? tcp_rtx_queue_head(sk) : NULL;
6311 u16 mss = tp->rx_opt.mss_clamp, try_exp = 0;
6312 bool syn_drop = false;
6314 if (mss == tp->rx_opt.user_mss) {
6315 struct tcp_options_received opt;
6317 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
6318 tcp_clear_options(&opt);
6319 opt.user_mss = opt.mss_clamp = 0;
6320 tcp_parse_options(sock_net(sk), synack, &opt, 0, NULL);
6321 mss = opt.mss_clamp;
6324 if (!tp->syn_fastopen) {
6325 /* Ignore an unsolicited cookie */
6327 } else if (tp->total_retrans) {
6328 /* SYN timed out and the SYN-ACK neither has a cookie nor
6329 * acknowledges data. Presumably the remote received only
6330 * the retransmitted (regular) SYNs: either the original
6331 * SYN-data or the corresponding SYN-ACK was dropped.
6333 syn_drop = (cookie->len < 0 && data);
6334 } else if (cookie->len < 0 && !tp->syn_data) {
6335 /* We requested a cookie but didn't get it. If we did not use
6336 * the (old) exp opt format then try so next time (try_exp=1).
6337 * Otherwise we go back to use the RFC7413 opt (try_exp=2).
6339 try_exp = tp->syn_fastopen_exp ? 2 : 1;
6342 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp);
6344 if (data) { /* Retransmit unacked data in SYN */
6345 if (tp->total_retrans)
6346 tp->fastopen_client_fail = TFO_SYN_RETRANSMITTED;
6348 tp->fastopen_client_fail = TFO_DATA_NOT_ACKED;
6349 skb_rbtree_walk_from(data)
6350 tcp_mark_skb_lost(sk, data);
6351 tcp_non_congestion_loss_retransmit(sk);
6352 NET_INC_STATS(sock_net(sk),
6353 LINUX_MIB_TCPFASTOPENACTIVEFAIL);
6356 tp->syn_data_acked = tp->syn_data;
6357 if (tp->syn_data_acked) {
6358 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
6359 /* SYN-data is counted as two separate packets in tcp_ack() */
6360 if (tp->delivered > 1)
6364 tcp_fastopen_add_skb(sk, synack);
6369 static void smc_check_reset_syn(struct tcp_sock *tp)
6371 #if IS_ENABLED(CONFIG_SMC)
6372 if (static_branch_unlikely(&tcp_have_smc)) {
6373 if (tp->syn_smc && !tp->rx_opt.smc_ok)
6379 static void tcp_try_undo_spurious_syn(struct sock *sk)
6381 struct tcp_sock *tp = tcp_sk(sk);
6384 /* undo_marker is set when SYN or SYNACK times out. The timeout is
6385 * spurious if the ACK's timestamp option echo value matches the
6386 * original SYN timestamp.
6388 syn_stamp = tp->retrans_stamp;
6389 if (tp->undo_marker && syn_stamp && tp->rx_opt.saw_tstamp &&
6390 syn_stamp == tp->rx_opt.rcv_tsecr)
6391 tp->undo_marker = 0;
6394 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
6395 const struct tcphdr *th)
6397 struct inet_connection_sock *icsk = inet_csk(sk);
6398 struct tcp_sock *tp = tcp_sk(sk);
6399 struct tcp_fastopen_cookie foc = { .len = -1 };
6400 int saved_clamp = tp->rx_opt.mss_clamp;
6404 tcp_parse_options(sock_net(sk), skb, &tp->rx_opt, 0, &foc);
6405 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
6406 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
6410 * "If the state is SYN-SENT then
6411 * first check the ACK bit
6412 * If the ACK bit is set
6413 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
6414 * a reset (unless the RST bit is set, if so drop
6415 * the segment and return)"
6417 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
6418 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
6419 /* Previous FIN/ACK or RST/ACK might be ignored. */
6420 if (icsk->icsk_retransmits == 0)
6421 inet_csk_reset_xmit_timer(sk,
6423 TCP_TIMEOUT_MIN, TCP_RTO_MAX);
6424 SKB_DR_SET(reason, TCP_INVALID_ACK_SEQUENCE);
6425 goto reset_and_undo;
6428 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
6429 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
6430 tcp_time_stamp_ts(tp))) {
6431 NET_INC_STATS(sock_net(sk),
6432 LINUX_MIB_PAWSACTIVEREJECTED);
6433 SKB_DR_SET(reason, TCP_RFC7323_PAWS);
6434 goto reset_and_undo;
6437 /* Now ACK is acceptable.
6439 * "If the RST bit is set
6440 * If the ACK was acceptable then signal the user "error:
6441 * connection reset", drop the segment, enter CLOSED state,
6442 * delete TCB, and return."
6453 * "fifth, if neither of the SYN or RST bits is set then
6454 * drop the segment and return."
6460 SKB_DR_SET(reason, TCP_FLAGS);
6461 goto discard_and_undo;
6464 * "If the SYN bit is on ...
6465 * are acceptable then ...
6466 * (our SYN has been ACKed), change the connection
6467 * state to ESTABLISHED..."
6470 tcp_ecn_rcv_synack(tp, th);
6472 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
6473 tcp_try_undo_spurious_syn(sk);
6474 tcp_ack(sk, skb, FLAG_SLOWPATH);
6476 /* Ok.. it's good. Set up sequence numbers and
6477 * move to established.
6479 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1);
6480 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
6482 /* RFC1323: The window in SYN & SYN/ACK segments is
6485 tp->snd_wnd = ntohs(th->window);
6487 if (!tp->rx_opt.wscale_ok) {
6488 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
6489 WRITE_ONCE(tp->window_clamp,
6490 min(tp->window_clamp, 65535U));
6493 if (tp->rx_opt.saw_tstamp) {
6494 tp->rx_opt.tstamp_ok = 1;
6495 tp->tcp_header_len =
6496 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
6497 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
6498 tcp_store_ts_recent(tp);
6500 tp->tcp_header_len = sizeof(struct tcphdr);
6503 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
6504 tcp_initialize_rcv_mss(sk);
6506 /* Remember, tcp_poll() does not lock socket!
6507 * Change state from SYN-SENT only after copied_seq
6508 * is initialized. */
6509 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6511 smc_check_reset_syn(tp);
6515 tcp_finish_connect(sk, skb);
6517 fastopen_fail = (tp->syn_fastopen || tp->syn_data) &&
6518 tcp_rcv_fastopen_synack(sk, skb, &foc);
6520 if (!sock_flag(sk, SOCK_DEAD)) {
6521 sk->sk_state_change(sk);
6522 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
6526 if (sk->sk_write_pending ||
6527 READ_ONCE(icsk->icsk_accept_queue.rskq_defer_accept) ||
6528 inet_csk_in_pingpong_mode(sk)) {
6529 /* Save one ACK. Data will be ready after
6530 * several ticks, if write_pending is set.
6532 * It may be deleted, but with this feature tcpdumps
6533 * look so _wonderfully_ clever, that I was not able
6534 * to stand against the temptation 8) --ANK
6536 inet_csk_schedule_ack(sk);
6537 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
6538 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
6539 TCP_DELACK_MAX, TCP_RTO_MAX);
6546 /* No ACK in the segment */
6550 * "If the RST bit is set
6552 * Otherwise (no ACK) drop the segment and return."
6554 SKB_DR_SET(reason, TCP_RESET);
6555 goto discard_and_undo;
6559 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
6560 tcp_paws_reject(&tp->rx_opt, 0)) {
6561 SKB_DR_SET(reason, TCP_RFC7323_PAWS);
6562 goto discard_and_undo;
6565 /* We see SYN without ACK. It is attempt of
6566 * simultaneous connect with crossed SYNs.
6567 * Particularly, it can be connect to self.
6569 #ifdef CONFIG_TCP_AO
6570 struct tcp_ao_info *ao;
6572 ao = rcu_dereference_protected(tp->ao_info,
6573 lockdep_sock_is_held(sk));
6575 WRITE_ONCE(ao->risn, th->seq);
6579 tcp_set_state(sk, TCP_SYN_RECV);
6581 if (tp->rx_opt.saw_tstamp) {
6582 tp->rx_opt.tstamp_ok = 1;
6583 tcp_store_ts_recent(tp);
6584 tp->tcp_header_len =
6585 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
6587 tp->tcp_header_len = sizeof(struct tcphdr);
6590 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1);
6591 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6592 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
6594 /* RFC1323: The window in SYN & SYN/ACK segments is
6597 tp->snd_wnd = ntohs(th->window);
6598 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
6599 tp->max_window = tp->snd_wnd;
6601 tcp_ecn_rcv_syn(tp, th);
6604 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
6605 tcp_initialize_rcv_mss(sk);
6607 tcp_send_synack(sk);
6609 /* Note, we could accept data and URG from this segment.
6610 * There are no obstacles to make this (except that we must
6611 * either change tcp_recvmsg() to prevent it from returning data
6612 * before 3WHS completes per RFC793, or employ TCP Fast Open).
6614 * However, if we ignore data in ACKless segments sometimes,
6615 * we have no reasons to accept it sometimes.
6616 * Also, seems the code doing it in step6 of tcp_rcv_state_process
6617 * is not flawless. So, discard packet for sanity.
6618 * Uncomment this return to process the data.
6625 /* "fifth, if neither of the SYN or RST bits is set then
6626 * drop the segment and return."
6630 tcp_clear_options(&tp->rx_opt);
6631 tp->rx_opt.mss_clamp = saved_clamp;
6632 tcp_drop_reason(sk, skb, reason);
6636 tcp_clear_options(&tp->rx_opt);
6637 tp->rx_opt.mss_clamp = saved_clamp;
6638 /* we can reuse/return @reason to its caller to handle the exception */
6642 static void tcp_rcv_synrecv_state_fastopen(struct sock *sk)
6644 struct tcp_sock *tp = tcp_sk(sk);
6645 struct request_sock *req;
6647 /* If we are still handling the SYNACK RTO, see if timestamp ECR allows
6648 * undo. If peer SACKs triggered fast recovery, we can't undo here.
6650 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss && !tp->packets_out)
6651 tcp_try_undo_recovery(sk);
6653 /* Reset rtx states to prevent spurious retransmits_timed_out() */
6654 tcp_update_rto_time(tp);
6655 tp->retrans_stamp = 0;
6656 inet_csk(sk)->icsk_retransmits = 0;
6658 /* Once we leave TCP_SYN_RECV or TCP_FIN_WAIT_1,
6659 * we no longer need req so release it.
6661 req = rcu_dereference_protected(tp->fastopen_rsk,
6662 lockdep_sock_is_held(sk));
6663 reqsk_fastopen_remove(sk, req, false);
6665 /* Re-arm the timer because data may have been sent out.
6666 * This is similar to the regular data transmission case
6667 * when new data has just been ack'ed.
6669 * (TFO) - we could try to be more aggressive and
6670 * retransmitting any data sooner based on when they
6677 * This function implements the receiving procedure of RFC 793 for
6678 * all states except ESTABLISHED and TIME_WAIT.
6679 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
6680 * address independent.
6683 enum skb_drop_reason
6684 tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb)
6686 struct tcp_sock *tp = tcp_sk(sk);
6687 struct inet_connection_sock *icsk = inet_csk(sk);
6688 const struct tcphdr *th = tcp_hdr(skb);
6689 struct request_sock *req;
6693 switch (sk->sk_state) {
6695 SKB_DR_SET(reason, TCP_CLOSE);
6700 return SKB_DROP_REASON_TCP_FLAGS;
6703 SKB_DR_SET(reason, TCP_RESET);
6708 SKB_DR_SET(reason, TCP_FLAGS);
6711 /* It is possible that we process SYN packets from backlog,
6712 * so we need to make sure to disable BH and RCU right there.
6716 icsk->icsk_af_ops->conn_request(sk, skb);
6723 SKB_DR_SET(reason, TCP_FLAGS);
6727 tp->rx_opt.saw_tstamp = 0;
6728 tcp_mstamp_refresh(tp);
6729 queued = tcp_rcv_synsent_state_process(sk, skb, th);
6733 /* Do step6 onward by hand. */
6734 tcp_urg(sk, skb, th);
6736 tcp_data_snd_check(sk);
6740 tcp_mstamp_refresh(tp);
6741 tp->rx_opt.saw_tstamp = 0;
6742 req = rcu_dereference_protected(tp->fastopen_rsk,
6743 lockdep_sock_is_held(sk));
6747 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
6748 sk->sk_state != TCP_FIN_WAIT1);
6750 if (!tcp_check_req(sk, skb, req, true, &req_stolen)) {
6751 SKB_DR_SET(reason, TCP_FASTOPEN);
6756 if (!th->ack && !th->rst && !th->syn) {
6757 SKB_DR_SET(reason, TCP_FLAGS);
6760 if (!tcp_validate_incoming(sk, skb, th, 0))
6763 /* step 5: check the ACK field */
6764 reason = tcp_ack(sk, skb, FLAG_SLOWPATH |
6765 FLAG_UPDATE_TS_RECENT |
6766 FLAG_NO_CHALLENGE_ACK);
6768 if ((int)reason <= 0) {
6769 if (sk->sk_state == TCP_SYN_RECV) {
6772 return SKB_DROP_REASON_TCP_OLD_ACK;
6775 /* accept old ack during closing */
6776 if ((int)reason < 0) {
6777 tcp_send_challenge_ack(sk);
6782 SKB_DR_SET(reason, NOT_SPECIFIED);
6783 switch (sk->sk_state) {
6785 tp->delivered++; /* SYN-ACK delivery isn't tracked in tcp_ack */
6787 tcp_synack_rtt_meas(sk, req);
6790 tcp_rcv_synrecv_state_fastopen(sk);
6792 tcp_try_undo_spurious_syn(sk);
6793 tp->retrans_stamp = 0;
6794 tcp_init_transfer(sk, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB,
6796 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
6798 tcp_ao_established(sk);
6800 tcp_set_state(sk, TCP_ESTABLISHED);
6801 sk->sk_state_change(sk);
6803 /* Note, that this wakeup is only for marginal crossed SYN case.
6804 * Passively open sockets are not waked up, because
6805 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
6808 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
6810 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
6811 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
6812 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
6814 if (tp->rx_opt.tstamp_ok)
6815 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
6817 if (!inet_csk(sk)->icsk_ca_ops->cong_control)
6818 tcp_update_pacing_rate(sk);
6820 /* Prevent spurious tcp_cwnd_restart() on first data packet */
6821 tp->lsndtime = tcp_jiffies32;
6823 tcp_initialize_rcv_mss(sk);
6824 tcp_fast_path_on(tp);
6825 if (sk->sk_shutdown & SEND_SHUTDOWN)
6826 tcp_shutdown(sk, SEND_SHUTDOWN);
6829 case TCP_FIN_WAIT1: {
6833 tcp_rcv_synrecv_state_fastopen(sk);
6835 if (tp->snd_una != tp->write_seq)
6838 tcp_set_state(sk, TCP_FIN_WAIT2);
6839 WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | SEND_SHUTDOWN);
6843 if (!sock_flag(sk, SOCK_DEAD)) {
6844 /* Wake up lingering close() */
6845 sk->sk_state_change(sk);
6849 if (READ_ONCE(tp->linger2) < 0) {
6851 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6852 return SKB_DROP_REASON_TCP_ABORT_ON_DATA;
6854 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6855 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6856 /* Receive out of order FIN after close() */
6857 if (tp->syn_fastopen && th->fin)
6858 tcp_fastopen_active_disable(sk);
6860 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6861 return SKB_DROP_REASON_TCP_ABORT_ON_DATA;
6864 tmo = tcp_fin_time(sk);
6865 if (tmo > TCP_TIMEWAIT_LEN) {
6866 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
6867 } else if (th->fin || sock_owned_by_user(sk)) {
6868 /* Bad case. We could lose such FIN otherwise.
6869 * It is not a big problem, but it looks confusing
6870 * and not so rare event. We still can lose it now,
6871 * if it spins in bh_lock_sock(), but it is really
6874 inet_csk_reset_keepalive_timer(sk, tmo);
6876 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
6883 if (tp->snd_una == tp->write_seq) {
6884 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
6890 if (tp->snd_una == tp->write_seq) {
6891 tcp_update_metrics(sk);
6898 /* step 6: check the URG bit */
6899 tcp_urg(sk, skb, th);
6901 /* step 7: process the segment text */
6902 switch (sk->sk_state) {
6903 case TCP_CLOSE_WAIT:
6906 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
6907 /* If a subflow has been reset, the packet should not
6908 * continue to be processed, drop the packet.
6910 if (sk_is_mptcp(sk) && !mptcp_incoming_options(sk, skb))
6917 /* RFC 793 says to queue data in these states,
6918 * RFC 1122 says we MUST send a reset.
6919 * BSD 4.4 also does reset.
6921 if (sk->sk_shutdown & RCV_SHUTDOWN) {
6922 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6923 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6924 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6926 return SKB_DROP_REASON_TCP_ABORT_ON_DATA;
6930 case TCP_ESTABLISHED:
6931 tcp_data_queue(sk, skb);
6936 /* tcp_data could move socket to TIME-WAIT */
6937 if (sk->sk_state != TCP_CLOSE) {
6938 tcp_data_snd_check(sk);
6939 tcp_ack_snd_check(sk);
6944 tcp_drop_reason(sk, skb, reason);
6952 EXPORT_SYMBOL(tcp_rcv_state_process);
6954 static inline void pr_drop_req(struct request_sock *req, __u16 port, int family)
6956 struct inet_request_sock *ireq = inet_rsk(req);
6958 if (family == AF_INET)
6959 net_dbg_ratelimited("drop open request from %pI4/%u\n",
6960 &ireq->ir_rmt_addr, port);
6961 #if IS_ENABLED(CONFIG_IPV6)
6962 else if (family == AF_INET6)
6963 net_dbg_ratelimited("drop open request from %pI6/%u\n",
6964 &ireq->ir_v6_rmt_addr, port);
6968 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
6970 * If we receive a SYN packet with these bits set, it means a
6971 * network is playing bad games with TOS bits. In order to
6972 * avoid possible false congestion notifications, we disable
6973 * TCP ECN negotiation.
6975 * Exception: tcp_ca wants ECN. This is required for DCTCP
6976 * congestion control: Linux DCTCP asserts ECT on all packets,
6977 * including SYN, which is most optimal solution; however,
6978 * others, such as FreeBSD do not.
6980 * Exception: At least one of the reserved bits of the TCP header (th->res1) is
6981 * set, indicating the use of a future TCP extension (such as AccECN). See
6982 * RFC8311 §4.3 which updates RFC3168 to allow the development of such
6985 static void tcp_ecn_create_request(struct request_sock *req,
6986 const struct sk_buff *skb,
6987 const struct sock *listen_sk,
6988 const struct dst_entry *dst)
6990 const struct tcphdr *th = tcp_hdr(skb);
6991 const struct net *net = sock_net(listen_sk);
6992 bool th_ecn = th->ece && th->cwr;
6999 ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield);
7000 ecn_ok_dst = dst_feature(dst, DST_FEATURE_ECN_MASK);
7001 ecn_ok = READ_ONCE(net->ipv4.sysctl_tcp_ecn) || ecn_ok_dst;
7003 if (((!ect || th->res1) && ecn_ok) || tcp_ca_needs_ecn(listen_sk) ||
7004 (ecn_ok_dst & DST_FEATURE_ECN_CA) ||
7005 tcp_bpf_ca_needs_ecn((struct sock *)req))
7006 inet_rsk(req)->ecn_ok = 1;
7009 static void tcp_openreq_init(struct request_sock *req,
7010 const struct tcp_options_received *rx_opt,
7011 struct sk_buff *skb, const struct sock *sk)
7013 struct inet_request_sock *ireq = inet_rsk(req);
7015 req->rsk_rcv_wnd = 0; /* So that tcp_send_synack() knows! */
7016 tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq;
7017 tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
7018 tcp_rsk(req)->snt_synack = 0;
7019 tcp_rsk(req)->last_oow_ack_time = 0;
7020 req->mss = rx_opt->mss_clamp;
7021 req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0;
7022 ireq->tstamp_ok = rx_opt->tstamp_ok;
7023 ireq->sack_ok = rx_opt->sack_ok;
7024 ireq->snd_wscale = rx_opt->snd_wscale;
7025 ireq->wscale_ok = rx_opt->wscale_ok;
7028 ireq->ir_rmt_port = tcp_hdr(skb)->source;
7029 ireq->ir_num = ntohs(tcp_hdr(skb)->dest);
7030 ireq->ir_mark = inet_request_mark(sk, skb);
7031 #if IS_ENABLED(CONFIG_SMC)
7032 ireq->smc_ok = rx_opt->smc_ok && !(tcp_sk(sk)->smc_hs_congested &&
7033 tcp_sk(sk)->smc_hs_congested(sk));
7038 * Return true if a syncookie should be sent
7040 static bool tcp_syn_flood_action(struct sock *sk, const char *proto)
7042 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
7043 const char *msg = "Dropping request";
7044 struct net *net = sock_net(sk);
7045 bool want_cookie = false;
7048 syncookies = READ_ONCE(net->ipv4.sysctl_tcp_syncookies);
7050 #ifdef CONFIG_SYN_COOKIES
7052 msg = "Sending cookies";
7054 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES);
7057 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP);
7059 if (!READ_ONCE(queue->synflood_warned) && syncookies != 2 &&
7060 xchg(&queue->synflood_warned, 1) == 0) {
7061 if (IS_ENABLED(CONFIG_IPV6) && sk->sk_family == AF_INET6) {
7062 net_info_ratelimited("%s: Possible SYN flooding on port [%pI6c]:%u. %s.\n",
7063 proto, inet6_rcv_saddr(sk),
7066 net_info_ratelimited("%s: Possible SYN flooding on port %pI4:%u. %s.\n",
7067 proto, &sk->sk_rcv_saddr,
7075 static void tcp_reqsk_record_syn(const struct sock *sk,
7076 struct request_sock *req,
7077 const struct sk_buff *skb)
7079 if (tcp_sk(sk)->save_syn) {
7080 u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb);
7081 struct saved_syn *saved_syn;
7085 if (tcp_sk(sk)->save_syn == 2) { /* Save full header. */
7086 base = skb_mac_header(skb);
7087 mac_hdrlen = skb_mac_header_len(skb);
7090 base = skb_network_header(skb);
7094 saved_syn = kmalloc(struct_size(saved_syn, data, len),
7097 saved_syn->mac_hdrlen = mac_hdrlen;
7098 saved_syn->network_hdrlen = skb_network_header_len(skb);
7099 saved_syn->tcp_hdrlen = tcp_hdrlen(skb);
7100 memcpy(saved_syn->data, base, len);
7101 req->saved_syn = saved_syn;
7106 /* If a SYN cookie is required and supported, returns a clamped MSS value to be
7107 * used for SYN cookie generation.
7109 u16 tcp_get_syncookie_mss(struct request_sock_ops *rsk_ops,
7110 const struct tcp_request_sock_ops *af_ops,
7111 struct sock *sk, struct tcphdr *th)
7113 struct tcp_sock *tp = tcp_sk(sk);
7116 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_syncookies) != 2 &&
7117 !inet_csk_reqsk_queue_is_full(sk))
7120 if (!tcp_syn_flood_action(sk, rsk_ops->slab_name))
7123 if (sk_acceptq_is_full(sk)) {
7124 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
7128 mss = tcp_parse_mss_option(th, tp->rx_opt.user_mss);
7130 mss = af_ops->mss_clamp;
7134 EXPORT_SYMBOL_GPL(tcp_get_syncookie_mss);
7136 int tcp_conn_request(struct request_sock_ops *rsk_ops,
7137 const struct tcp_request_sock_ops *af_ops,
7138 struct sock *sk, struct sk_buff *skb)
7140 struct tcp_fastopen_cookie foc = { .len = -1 };
7141 struct tcp_options_received tmp_opt;
7142 struct tcp_sock *tp = tcp_sk(sk);
7143 struct net *net = sock_net(sk);
7144 struct sock *fastopen_sk = NULL;
7145 struct request_sock *req;
7146 bool want_cookie = false;
7147 struct dst_entry *dst;
7152 #ifdef CONFIG_TCP_AO
7153 const struct tcp_ao_hdr *aoh;
7156 isn = __this_cpu_read(tcp_tw_isn);
7158 /* TW buckets are converted to open requests without
7159 * limitations, they conserve resources and peer is
7160 * evidently real one.
7162 __this_cpu_write(tcp_tw_isn, 0);
7164 syncookies = READ_ONCE(net->ipv4.sysctl_tcp_syncookies);
7166 if (syncookies == 2 || inet_csk_reqsk_queue_is_full(sk)) {
7167 want_cookie = tcp_syn_flood_action(sk,
7168 rsk_ops->slab_name);
7174 if (sk_acceptq_is_full(sk)) {
7175 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
7179 req = inet_reqsk_alloc(rsk_ops, sk, !want_cookie);
7183 req->syncookie = want_cookie;
7184 tcp_rsk(req)->af_specific = af_ops;
7185 tcp_rsk(req)->ts_off = 0;
7186 tcp_rsk(req)->req_usec_ts = false;
7187 #if IS_ENABLED(CONFIG_MPTCP)
7188 tcp_rsk(req)->is_mptcp = 0;
7191 tcp_clear_options(&tmp_opt);
7192 tmp_opt.mss_clamp = af_ops->mss_clamp;
7193 tmp_opt.user_mss = tp->rx_opt.user_mss;
7194 tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0,
7195 want_cookie ? NULL : &foc);
7197 if (want_cookie && !tmp_opt.saw_tstamp)
7198 tcp_clear_options(&tmp_opt);
7200 if (IS_ENABLED(CONFIG_SMC) && want_cookie)
7203 tmp_opt.tstamp_ok = tmp_opt.saw_tstamp;
7204 tcp_openreq_init(req, &tmp_opt, skb, sk);
7205 inet_rsk(req)->no_srccheck = inet_test_bit(TRANSPARENT, sk);
7207 /* Note: tcp_v6_init_req() might override ir_iif for link locals */
7208 inet_rsk(req)->ir_iif = inet_request_bound_dev_if(sk, skb);
7210 dst = af_ops->route_req(sk, skb, &fl, req, isn);
7214 if (tmp_opt.tstamp_ok) {
7215 tcp_rsk(req)->req_usec_ts = dst_tcp_usec_ts(dst);
7216 tcp_rsk(req)->ts_off = af_ops->init_ts_off(net, skb);
7218 if (!want_cookie && !isn) {
7219 int max_syn_backlog = READ_ONCE(net->ipv4.sysctl_max_syn_backlog);
7221 /* Kill the following clause, if you dislike this way. */
7223 (max_syn_backlog - inet_csk_reqsk_queue_len(sk) <
7224 (max_syn_backlog >> 2)) &&
7225 !tcp_peer_is_proven(req, dst)) {
7226 /* Without syncookies last quarter of
7227 * backlog is filled with destinations,
7228 * proven to be alive.
7229 * It means that we continue to communicate
7230 * to destinations, already remembered
7231 * to the moment of synflood.
7233 pr_drop_req(req, ntohs(tcp_hdr(skb)->source),
7235 goto drop_and_release;
7238 isn = af_ops->init_seq(skb);
7241 tcp_ecn_create_request(req, skb, sk, dst);
7244 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss);
7245 if (!tmp_opt.tstamp_ok)
7246 inet_rsk(req)->ecn_ok = 0;
7249 #ifdef CONFIG_TCP_AO
7250 if (tcp_parse_auth_options(tcp_hdr(skb), NULL, &aoh))
7251 goto drop_and_release; /* Invalid TCP options */
7253 tcp_rsk(req)->used_tcp_ao = true;
7254 tcp_rsk(req)->ao_rcv_next = aoh->keyid;
7255 tcp_rsk(req)->ao_keyid = aoh->rnext_keyid;
7258 tcp_rsk(req)->used_tcp_ao = false;
7261 tcp_rsk(req)->snt_isn = isn;
7262 tcp_rsk(req)->txhash = net_tx_rndhash();
7263 tcp_rsk(req)->syn_tos = TCP_SKB_CB(skb)->ip_dsfield;
7264 tcp_openreq_init_rwin(req, sk, dst);
7265 sk_rx_queue_set(req_to_sk(req), skb);
7267 tcp_reqsk_record_syn(sk, req, skb);
7268 fastopen_sk = tcp_try_fastopen(sk, skb, req, &foc, dst);
7271 af_ops->send_synack(fastopen_sk, dst, &fl, req,
7272 &foc, TCP_SYNACK_FASTOPEN, skb);
7273 /* Add the child socket directly into the accept queue */
7274 if (!inet_csk_reqsk_queue_add(sk, req, fastopen_sk)) {
7275 reqsk_fastopen_remove(fastopen_sk, req, false);
7276 bh_unlock_sock(fastopen_sk);
7277 sock_put(fastopen_sk);
7280 sk->sk_data_ready(sk);
7281 bh_unlock_sock(fastopen_sk);
7282 sock_put(fastopen_sk);
7284 tcp_rsk(req)->tfo_listener = false;
7286 req->timeout = tcp_timeout_init((struct sock *)req);
7287 if (unlikely(!inet_csk_reqsk_queue_hash_add(sk, req,
7294 af_ops->send_synack(sk, dst, &fl, req, &foc,
7295 !want_cookie ? TCP_SYNACK_NORMAL :
7314 EXPORT_SYMBOL(tcp_conn_request);