1 // SPDX-License-Identifier: GPL-2.0
3 * INET An implementation of the TCP/IP protocol suite for the LINUX
4 * operating system. INET is implemented using the BSD Socket
5 * interface as the means of communication with the user level.
7 * Implementation of the Transmission Control Protocol(TCP).
10 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
11 * Mark Evans, <evansmp@uhura.aston.ac.uk>
12 * Corey Minyard <wf-rch!minyard@relay.EU.net>
13 * Florian La Roche, <flla@stud.uni-sb.de>
14 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
15 * Linus Torvalds, <torvalds@cs.helsinki.fi>
16 * Alan Cox, <gw4pts@gw4pts.ampr.org>
17 * Matthew Dillon, <dillon@apollo.west.oic.com>
18 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
19 * Jorge Cwik, <jorge@laser.satlink.net>
24 * Pedro Roque : Fast Retransmit/Recovery.
26 * Retransmit queue handled by TCP.
27 * Better retransmit timer handling.
28 * New congestion avoidance.
32 * Eric : Fast Retransmit.
33 * Randy Scott : MSS option defines.
34 * Eric Schenk : Fixes to slow start algorithm.
35 * Eric Schenk : Yet another double ACK bug.
36 * Eric Schenk : Delayed ACK bug fixes.
37 * Eric Schenk : Floyd style fast retrans war avoidance.
38 * David S. Miller : Don't allow zero congestion window.
39 * Eric Schenk : Fix retransmitter so that it sends
40 * next packet on ack of previous packet.
41 * Andi Kleen : Moved open_request checking here
42 * and process RSTs for open_requests.
43 * Andi Kleen : Better prune_queue, and other fixes.
44 * Andrey Savochkin: Fix RTT measurements in the presence of
46 * Andrey Savochkin: Check sequence numbers correctly when
47 * removing SACKs due to in sequence incoming
49 * Andi Kleen: Make sure we never ack data there is not
50 * enough room for. Also make this condition
51 * a fatal error if it might still happen.
52 * Andi Kleen: Add tcp_measure_rcv_mss to make
53 * connections with MSS<min(MTU,ann. MSS)
54 * work without delayed acks.
55 * Andi Kleen: Process packets with PSH set in the
57 * J Hadi Salim: ECN support
60 * Panu Kuhlberg: Experimental audit of TCP (re)transmission
61 * engine. Lots of bugs are found.
62 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs
65 #define pr_fmt(fmt) "TCP: " fmt
68 #include <linux/slab.h>
69 #include <linux/module.h>
70 #include <linux/sysctl.h>
71 #include <linux/kernel.h>
72 #include <linux/prefetch.h>
75 #include <net/inet_common.h>
76 #include <linux/ipsec.h>
77 #include <asm/unaligned.h>
78 #include <linux/errqueue.h>
79 #include <trace/events/tcp.h>
80 #include <linux/static_key.h>
82 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
84 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
85 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
86 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
87 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
88 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
89 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
90 #define FLAG_ECE 0x40 /* ECE in this ACK */
91 #define FLAG_LOST_RETRANS 0x80 /* This ACK marks some retransmission lost */
92 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
93 #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
94 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
95 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
96 #define FLAG_SET_XMIT_TIMER 0x1000 /* Set TLP or RTO timer */
97 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
98 #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
99 #define FLAG_NO_CHALLENGE_ACK 0x8000 /* do not call tcp_send_challenge_ack() */
100 #define FLAG_ACK_MAYBE_DELAYED 0x10000 /* Likely a delayed ACK */
102 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
103 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
104 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE|FLAG_DSACKING_ACK)
105 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
107 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
108 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
110 #define REXMIT_NONE 0 /* no loss recovery to do */
111 #define REXMIT_LOST 1 /* retransmit packets marked lost */
112 #define REXMIT_NEW 2 /* FRTO-style transmit of unsent/new packets */
114 #if IS_ENABLED(CONFIG_TLS_DEVICE)
115 static DEFINE_STATIC_KEY_FALSE(clean_acked_data_enabled);
117 void clean_acked_data_enable(struct inet_connection_sock *icsk,
118 void (*cad)(struct sock *sk, u32 ack_seq))
120 icsk->icsk_clean_acked = cad;
121 static_branch_inc(&clean_acked_data_enabled);
123 EXPORT_SYMBOL_GPL(clean_acked_data_enable);
125 void clean_acked_data_disable(struct inet_connection_sock *icsk)
127 static_branch_dec(&clean_acked_data_enabled);
128 icsk->icsk_clean_acked = NULL;
130 EXPORT_SYMBOL_GPL(clean_acked_data_disable);
133 static void tcp_gro_dev_warn(struct sock *sk, const struct sk_buff *skb,
136 static bool __once __read_mostly;
139 struct net_device *dev;
144 dev = dev_get_by_index_rcu(sock_net(sk), skb->skb_iif);
145 if (!dev || len >= dev->mtu)
146 pr_warn("%s: Driver has suspect GRO implementation, TCP performance may be compromised.\n",
147 dev ? dev->name : "Unknown driver");
152 /* Adapt the MSS value used to make delayed ack decision to the
155 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
157 struct inet_connection_sock *icsk = inet_csk(sk);
158 const unsigned int lss = icsk->icsk_ack.last_seg_size;
161 icsk->icsk_ack.last_seg_size = 0;
163 /* skb->len may jitter because of SACKs, even if peer
164 * sends good full-sized frames.
166 len = skb_shinfo(skb)->gso_size ? : skb->len;
167 if (len >= icsk->icsk_ack.rcv_mss) {
168 icsk->icsk_ack.rcv_mss = min_t(unsigned int, len,
170 /* Account for possibly-removed options */
171 if (unlikely(len > icsk->icsk_ack.rcv_mss +
172 MAX_TCP_OPTION_SPACE))
173 tcp_gro_dev_warn(sk, skb, len);
175 /* Otherwise, we make more careful check taking into account,
176 * that SACKs block is variable.
178 * "len" is invariant segment length, including TCP header.
180 len += skb->data - skb_transport_header(skb);
181 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
182 /* If PSH is not set, packet should be
183 * full sized, provided peer TCP is not badly broken.
184 * This observation (if it is correct 8)) allows
185 * to handle super-low mtu links fairly.
187 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
188 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
189 /* Subtract also invariant (if peer is RFC compliant),
190 * tcp header plus fixed timestamp option length.
191 * Resulting "len" is MSS free of SACK jitter.
193 len -= tcp_sk(sk)->tcp_header_len;
194 icsk->icsk_ack.last_seg_size = len;
196 icsk->icsk_ack.rcv_mss = len;
200 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
201 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
202 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
206 static void tcp_incr_quickack(struct sock *sk, unsigned int max_quickacks)
208 struct inet_connection_sock *icsk = inet_csk(sk);
209 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
213 quickacks = min(quickacks, max_quickacks);
214 if (quickacks > icsk->icsk_ack.quick)
215 icsk->icsk_ack.quick = quickacks;
218 static void tcp_enter_quickack_mode(struct sock *sk, unsigned int max_quickacks)
220 struct inet_connection_sock *icsk = inet_csk(sk);
222 tcp_incr_quickack(sk, max_quickacks);
223 icsk->icsk_ack.pingpong = 0;
224 icsk->icsk_ack.ato = TCP_ATO_MIN;
227 /* Send ACKs quickly, if "quick" count is not exhausted
228 * and the session is not interactive.
231 static bool tcp_in_quickack_mode(struct sock *sk)
233 const struct inet_connection_sock *icsk = inet_csk(sk);
234 const struct dst_entry *dst = __sk_dst_get(sk);
236 return (dst && dst_metric(dst, RTAX_QUICKACK)) ||
237 (icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong);
240 static void tcp_ecn_queue_cwr(struct tcp_sock *tp)
242 if (tp->ecn_flags & TCP_ECN_OK)
243 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
246 static void tcp_ecn_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb)
248 if (tcp_hdr(skb)->cwr)
249 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
252 static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp)
254 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
257 static void __tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb)
259 struct tcp_sock *tp = tcp_sk(sk);
261 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
262 case INET_ECN_NOT_ECT:
263 /* Funny extension: if ECT is not set on a segment,
264 * and we already seen ECT on a previous segment,
265 * it is probably a retransmit.
267 if (tp->ecn_flags & TCP_ECN_SEEN)
268 tcp_enter_quickack_mode(sk, 1);
271 if (tcp_ca_needs_ecn(sk))
272 tcp_ca_event(sk, CA_EVENT_ECN_IS_CE);
274 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
275 /* Better not delay acks, sender can have a very low cwnd */
276 tcp_enter_quickack_mode(sk, 1);
277 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
279 tp->ecn_flags |= TCP_ECN_SEEN;
282 if (tcp_ca_needs_ecn(sk))
283 tcp_ca_event(sk, CA_EVENT_ECN_NO_CE);
284 tp->ecn_flags |= TCP_ECN_SEEN;
289 static void tcp_ecn_check_ce(struct sock *sk, const struct sk_buff *skb)
291 if (tcp_sk(sk)->ecn_flags & TCP_ECN_OK)
292 __tcp_ecn_check_ce(sk, skb);
295 static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
297 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
298 tp->ecn_flags &= ~TCP_ECN_OK;
301 static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
303 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
304 tp->ecn_flags &= ~TCP_ECN_OK;
307 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
309 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
314 /* Buffer size and advertised window tuning.
316 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
319 static void tcp_sndbuf_expand(struct sock *sk)
321 const struct tcp_sock *tp = tcp_sk(sk);
322 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
326 /* Worst case is non GSO/TSO : each frame consumes one skb
327 * and skb->head is kmalloced using power of two area of memory
329 per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
331 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
333 per_mss = roundup_pow_of_two(per_mss) +
334 SKB_DATA_ALIGN(sizeof(struct sk_buff));
336 nr_segs = max_t(u32, TCP_INIT_CWND, tp->snd_cwnd);
337 nr_segs = max_t(u32, nr_segs, tp->reordering + 1);
339 /* Fast Recovery (RFC 5681 3.2) :
340 * Cubic needs 1.7 factor, rounded to 2 to include
341 * extra cushion (application might react slowly to EPOLLOUT)
343 sndmem = ca_ops->sndbuf_expand ? ca_ops->sndbuf_expand(sk) : 2;
344 sndmem *= nr_segs * per_mss;
346 if (sk->sk_sndbuf < sndmem)
347 sk->sk_sndbuf = min(sndmem, sock_net(sk)->ipv4.sysctl_tcp_wmem[2]);
350 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
352 * All tcp_full_space() is split to two parts: "network" buffer, allocated
353 * forward and advertised in receiver window (tp->rcv_wnd) and
354 * "application buffer", required to isolate scheduling/application
355 * latencies from network.
356 * window_clamp is maximal advertised window. It can be less than
357 * tcp_full_space(), in this case tcp_full_space() - window_clamp
358 * is reserved for "application" buffer. The less window_clamp is
359 * the smoother our behaviour from viewpoint of network, but the lower
360 * throughput and the higher sensitivity of the connection to losses. 8)
362 * rcv_ssthresh is more strict window_clamp used at "slow start"
363 * phase to predict further behaviour of this connection.
364 * It is used for two goals:
365 * - to enforce header prediction at sender, even when application
366 * requires some significant "application buffer". It is check #1.
367 * - to prevent pruning of receive queue because of misprediction
368 * of receiver window. Check #2.
370 * The scheme does not work when sender sends good segments opening
371 * window and then starts to feed us spaghetti. But it should work
372 * in common situations. Otherwise, we have to rely on queue collapsing.
375 /* Slow part of check#2. */
376 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
378 struct tcp_sock *tp = tcp_sk(sk);
380 int truesize = tcp_win_from_space(sk, skb->truesize) >> 1;
381 int window = tcp_win_from_space(sk, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]) >> 1;
383 while (tp->rcv_ssthresh <= window) {
384 if (truesize <= skb->len)
385 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
393 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
395 struct tcp_sock *tp = tcp_sk(sk);
398 if (tp->rcv_ssthresh < tp->window_clamp &&
399 (int)tp->rcv_ssthresh < tcp_space(sk) &&
400 !tcp_under_memory_pressure(sk)) {
403 /* Check #2. Increase window, if skb with such overhead
404 * will fit to rcvbuf in future.
406 if (tcp_win_from_space(sk, skb->truesize) <= skb->len)
407 incr = 2 * tp->advmss;
409 incr = __tcp_grow_window(sk, skb);
412 incr = max_t(int, incr, 2 * skb->len);
413 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
415 inet_csk(sk)->icsk_ack.quick |= 1;
420 /* 3. Tuning rcvbuf, when connection enters established state. */
421 static void tcp_fixup_rcvbuf(struct sock *sk)
423 u32 mss = tcp_sk(sk)->advmss;
426 rcvmem = 2 * SKB_TRUESIZE(mss + MAX_TCP_HEADER) *
427 tcp_default_init_rwnd(mss);
429 /* Dynamic Right Sizing (DRS) has 2 to 3 RTT latency
430 * Allow enough cushion so that sender is not limited by our window
432 if (sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf)
435 if (sk->sk_rcvbuf < rcvmem)
436 sk->sk_rcvbuf = min(rcvmem, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
439 /* 4. Try to fixup all. It is made immediately after connection enters
442 void tcp_init_buffer_space(struct sock *sk)
444 int tcp_app_win = sock_net(sk)->ipv4.sysctl_tcp_app_win;
445 struct tcp_sock *tp = tcp_sk(sk);
448 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
449 tcp_fixup_rcvbuf(sk);
450 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
451 tcp_sndbuf_expand(sk);
453 tp->rcvq_space.space = tp->rcv_wnd;
454 tcp_mstamp_refresh(tp);
455 tp->rcvq_space.time = tp->tcp_mstamp;
456 tp->rcvq_space.seq = tp->copied_seq;
458 maxwin = tcp_full_space(sk);
460 if (tp->window_clamp >= maxwin) {
461 tp->window_clamp = maxwin;
463 if (tcp_app_win && maxwin > 4 * tp->advmss)
464 tp->window_clamp = max(maxwin -
465 (maxwin >> tcp_app_win),
469 /* Force reservation of one segment. */
471 tp->window_clamp > 2 * tp->advmss &&
472 tp->window_clamp + tp->advmss > maxwin)
473 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
475 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
476 tp->snd_cwnd_stamp = tcp_jiffies32;
479 /* 5. Recalculate window clamp after socket hit its memory bounds. */
480 static void tcp_clamp_window(struct sock *sk)
482 struct tcp_sock *tp = tcp_sk(sk);
483 struct inet_connection_sock *icsk = inet_csk(sk);
484 struct net *net = sock_net(sk);
486 icsk->icsk_ack.quick = 0;
488 if (sk->sk_rcvbuf < net->ipv4.sysctl_tcp_rmem[2] &&
489 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
490 !tcp_under_memory_pressure(sk) &&
491 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
492 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
493 net->ipv4.sysctl_tcp_rmem[2]);
495 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
496 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
499 /* Initialize RCV_MSS value.
500 * RCV_MSS is an our guess about MSS used by the peer.
501 * We haven't any direct information about the MSS.
502 * It's better to underestimate the RCV_MSS rather than overestimate.
503 * Overestimations make us ACKing less frequently than needed.
504 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
506 void tcp_initialize_rcv_mss(struct sock *sk)
508 const struct tcp_sock *tp = tcp_sk(sk);
509 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
511 hint = min(hint, tp->rcv_wnd / 2);
512 hint = min(hint, TCP_MSS_DEFAULT);
513 hint = max(hint, TCP_MIN_MSS);
515 inet_csk(sk)->icsk_ack.rcv_mss = hint;
517 EXPORT_SYMBOL(tcp_initialize_rcv_mss);
519 /* Receiver "autotuning" code.
521 * The algorithm for RTT estimation w/o timestamps is based on
522 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
523 * <http://public.lanl.gov/radiant/pubs.html#DRS>
525 * More detail on this code can be found at
526 * <http://staff.psc.edu/jheffner/>,
527 * though this reference is out of date. A new paper
530 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
532 u32 new_sample = tp->rcv_rtt_est.rtt_us;
535 if (new_sample != 0) {
536 /* If we sample in larger samples in the non-timestamp
537 * case, we could grossly overestimate the RTT especially
538 * with chatty applications or bulk transfer apps which
539 * are stalled on filesystem I/O.
541 * Also, since we are only going for a minimum in the
542 * non-timestamp case, we do not smooth things out
543 * else with timestamps disabled convergence takes too
547 m -= (new_sample >> 3);
555 /* No previous measure. */
559 tp->rcv_rtt_est.rtt_us = new_sample;
562 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
566 if (tp->rcv_rtt_est.time == 0)
568 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
570 delta_us = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcv_rtt_est.time);
573 tcp_rcv_rtt_update(tp, delta_us, 1);
576 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
577 tp->rcv_rtt_est.time = tp->tcp_mstamp;
580 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
581 const struct sk_buff *skb)
583 struct tcp_sock *tp = tcp_sk(sk);
585 if (tp->rx_opt.rcv_tsecr == tp->rcv_rtt_last_tsecr)
587 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr;
589 if (TCP_SKB_CB(skb)->end_seq -
590 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss) {
591 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr;
596 delta_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
597 tcp_rcv_rtt_update(tp, delta_us, 0);
602 * This function should be called every time data is copied to user space.
603 * It calculates the appropriate TCP receive buffer space.
605 void tcp_rcv_space_adjust(struct sock *sk)
607 struct tcp_sock *tp = tcp_sk(sk);
611 trace_tcp_rcv_space_adjust(sk);
613 tcp_mstamp_refresh(tp);
614 time = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcvq_space.time);
615 if (time < (tp->rcv_rtt_est.rtt_us >> 3) || tp->rcv_rtt_est.rtt_us == 0)
618 /* Number of bytes copied to user in last RTT */
619 copied = tp->copied_seq - tp->rcvq_space.seq;
620 if (copied <= tp->rcvq_space.space)
624 * copied = bytes received in previous RTT, our base window
625 * To cope with packet losses, we need a 2x factor
626 * To cope with slow start, and sender growing its cwin by 100 %
627 * every RTT, we need a 4x factor, because the ACK we are sending
628 * now is for the next RTT, not the current one :
629 * <prev RTT . ><current RTT .. ><next RTT .... >
632 if (sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf &&
633 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
637 /* minimal window to cope with packet losses, assuming
638 * steady state. Add some cushion because of small variations.
640 rcvwin = ((u64)copied << 1) + 16 * tp->advmss;
642 /* Accommodate for sender rate increase (eg. slow start) */
643 grow = rcvwin * (copied - tp->rcvq_space.space);
644 do_div(grow, tp->rcvq_space.space);
645 rcvwin += (grow << 1);
647 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
648 while (tcp_win_from_space(sk, rcvmem) < tp->advmss)
651 do_div(rcvwin, tp->advmss);
652 rcvbuf = min_t(u64, rcvwin * rcvmem,
653 sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
654 if (rcvbuf > sk->sk_rcvbuf) {
655 sk->sk_rcvbuf = rcvbuf;
657 /* Make the window clamp follow along. */
658 tp->window_clamp = tcp_win_from_space(sk, rcvbuf);
661 tp->rcvq_space.space = copied;
664 tp->rcvq_space.seq = tp->copied_seq;
665 tp->rcvq_space.time = tp->tcp_mstamp;
668 /* There is something which you must keep in mind when you analyze the
669 * behavior of the tp->ato delayed ack timeout interval. When a
670 * connection starts up, we want to ack as quickly as possible. The
671 * problem is that "good" TCP's do slow start at the beginning of data
672 * transmission. The means that until we send the first few ACK's the
673 * sender will sit on his end and only queue most of his data, because
674 * he can only send snd_cwnd unacked packets at any given time. For
675 * each ACK we send, he increments snd_cwnd and transmits more of his
678 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
680 struct tcp_sock *tp = tcp_sk(sk);
681 struct inet_connection_sock *icsk = inet_csk(sk);
684 inet_csk_schedule_ack(sk);
686 tcp_measure_rcv_mss(sk, skb);
688 tcp_rcv_rtt_measure(tp);
692 if (!icsk->icsk_ack.ato) {
693 /* The _first_ data packet received, initialize
694 * delayed ACK engine.
696 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
697 icsk->icsk_ack.ato = TCP_ATO_MIN;
699 int m = now - icsk->icsk_ack.lrcvtime;
701 if (m <= TCP_ATO_MIN / 2) {
702 /* The fastest case is the first. */
703 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
704 } else if (m < icsk->icsk_ack.ato) {
705 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
706 if (icsk->icsk_ack.ato > icsk->icsk_rto)
707 icsk->icsk_ack.ato = icsk->icsk_rto;
708 } else if (m > icsk->icsk_rto) {
709 /* Too long gap. Apparently sender failed to
710 * restart window, so that we send ACKs quickly.
712 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
716 icsk->icsk_ack.lrcvtime = now;
718 tcp_ecn_check_ce(sk, skb);
721 tcp_grow_window(sk, skb);
724 /* Called to compute a smoothed rtt estimate. The data fed to this
725 * routine either comes from timestamps, or from segments that were
726 * known _not_ to have been retransmitted [see Karn/Partridge
727 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
728 * piece by Van Jacobson.
729 * NOTE: the next three routines used to be one big routine.
730 * To save cycles in the RFC 1323 implementation it was better to break
731 * it up into three procedures. -- erics
733 static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
735 struct tcp_sock *tp = tcp_sk(sk);
736 long m = mrtt_us; /* RTT */
737 u32 srtt = tp->srtt_us;
739 /* The following amusing code comes from Jacobson's
740 * article in SIGCOMM '88. Note that rtt and mdev
741 * are scaled versions of rtt and mean deviation.
742 * This is designed to be as fast as possible
743 * m stands for "measurement".
745 * On a 1990 paper the rto value is changed to:
746 * RTO = rtt + 4 * mdev
748 * Funny. This algorithm seems to be very broken.
749 * These formulae increase RTO, when it should be decreased, increase
750 * too slowly, when it should be increased quickly, decrease too quickly
751 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
752 * does not matter how to _calculate_ it. Seems, it was trap
753 * that VJ failed to avoid. 8)
756 m -= (srtt >> 3); /* m is now error in rtt est */
757 srtt += m; /* rtt = 7/8 rtt + 1/8 new */
759 m = -m; /* m is now abs(error) */
760 m -= (tp->mdev_us >> 2); /* similar update on mdev */
761 /* This is similar to one of Eifel findings.
762 * Eifel blocks mdev updates when rtt decreases.
763 * This solution is a bit different: we use finer gain
764 * for mdev in this case (alpha*beta).
765 * Like Eifel it also prevents growth of rto,
766 * but also it limits too fast rto decreases,
767 * happening in pure Eifel.
772 m -= (tp->mdev_us >> 2); /* similar update on mdev */
774 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */
775 if (tp->mdev_us > tp->mdev_max_us) {
776 tp->mdev_max_us = tp->mdev_us;
777 if (tp->mdev_max_us > tp->rttvar_us)
778 tp->rttvar_us = tp->mdev_max_us;
780 if (after(tp->snd_una, tp->rtt_seq)) {
781 if (tp->mdev_max_us < tp->rttvar_us)
782 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
783 tp->rtt_seq = tp->snd_nxt;
784 tp->mdev_max_us = tcp_rto_min_us(sk);
787 /* no previous measure. */
788 srtt = m << 3; /* take the measured time to be rtt */
789 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */
790 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
791 tp->mdev_max_us = tp->rttvar_us;
792 tp->rtt_seq = tp->snd_nxt;
794 tp->srtt_us = max(1U, srtt);
797 static void tcp_update_pacing_rate(struct sock *sk)
799 const struct tcp_sock *tp = tcp_sk(sk);
802 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
803 rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3);
805 /* current rate is (cwnd * mss) / srtt
806 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
807 * In Congestion Avoidance phase, set it to 120 % the current rate.
809 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
810 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
811 * end of slow start and should slow down.
813 if (tp->snd_cwnd < tp->snd_ssthresh / 2)
814 rate *= sock_net(sk)->ipv4.sysctl_tcp_pacing_ss_ratio;
816 rate *= sock_net(sk)->ipv4.sysctl_tcp_pacing_ca_ratio;
818 rate *= max(tp->snd_cwnd, tp->packets_out);
820 if (likely(tp->srtt_us))
821 do_div(rate, tp->srtt_us);
823 /* WRITE_ONCE() is needed because sch_fq fetches sk_pacing_rate
824 * without any lock. We want to make sure compiler wont store
825 * intermediate values in this location.
827 WRITE_ONCE(sk->sk_pacing_rate, min_t(u64, rate,
828 sk->sk_max_pacing_rate));
831 /* Calculate rto without backoff. This is the second half of Van Jacobson's
832 * routine referred to above.
834 static void tcp_set_rto(struct sock *sk)
836 const struct tcp_sock *tp = tcp_sk(sk);
837 /* Old crap is replaced with new one. 8)
840 * 1. If rtt variance happened to be less 50msec, it is hallucination.
841 * It cannot be less due to utterly erratic ACK generation made
842 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
843 * to do with delayed acks, because at cwnd>2 true delack timeout
844 * is invisible. Actually, Linux-2.4 also generates erratic
845 * ACKs in some circumstances.
847 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
849 /* 2. Fixups made earlier cannot be right.
850 * If we do not estimate RTO correctly without them,
851 * all the algo is pure shit and should be replaced
852 * with correct one. It is exactly, which we pretend to do.
855 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
856 * guarantees that rto is higher.
861 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
863 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
866 cwnd = TCP_INIT_CWND;
867 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
870 /* Take a notice that peer is sending D-SACKs */
871 static void tcp_dsack_seen(struct tcp_sock *tp)
873 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
874 tp->rack.dsack_seen = 1;
877 /* It's reordering when higher sequence was delivered (i.e. sacked) before
878 * some lower never-retransmitted sequence ("low_seq"). The maximum reordering
879 * distance is approximated in full-mss packet distance ("reordering").
881 static void tcp_check_sack_reordering(struct sock *sk, const u32 low_seq,
884 struct tcp_sock *tp = tcp_sk(sk);
885 const u32 mss = tp->mss_cache;
888 fack = tcp_highest_sack_seq(tp);
889 if (!before(low_seq, fack))
892 metric = fack - low_seq;
893 if ((metric > tp->reordering * mss) && mss) {
894 #if FASTRETRANS_DEBUG > 1
895 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
896 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
900 tp->undo_marker ? tp->undo_retrans : 0);
902 tp->reordering = min_t(u32, (metric + mss - 1) / mss,
903 sock_net(sk)->ipv4.sysctl_tcp_max_reordering);
907 /* This exciting event is worth to be remembered. 8) */
908 NET_INC_STATS(sock_net(sk),
909 ts ? LINUX_MIB_TCPTSREORDER : LINUX_MIB_TCPSACKREORDER);
912 /* This must be called before lost_out is incremented */
913 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
915 if (!tp->retransmit_skb_hint ||
916 before(TCP_SKB_CB(skb)->seq,
917 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
918 tp->retransmit_skb_hint = skb;
921 /* Sum the number of packets on the wire we have marked as lost.
922 * There are two cases we care about here:
923 * a) Packet hasn't been marked lost (nor retransmitted),
924 * and this is the first loss.
925 * b) Packet has been marked both lost and retransmitted,
926 * and this means we think it was lost again.
928 static void tcp_sum_lost(struct tcp_sock *tp, struct sk_buff *skb)
930 __u8 sacked = TCP_SKB_CB(skb)->sacked;
932 if (!(sacked & TCPCB_LOST) ||
933 ((sacked & TCPCB_LOST) && (sacked & TCPCB_SACKED_RETRANS)))
934 tp->lost += tcp_skb_pcount(skb);
937 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
939 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
940 tcp_verify_retransmit_hint(tp, skb);
942 tp->lost_out += tcp_skb_pcount(skb);
943 tcp_sum_lost(tp, skb);
944 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
948 void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp, struct sk_buff *skb)
950 tcp_verify_retransmit_hint(tp, skb);
952 tcp_sum_lost(tp, skb);
953 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
954 tp->lost_out += tcp_skb_pcount(skb);
955 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
959 /* This procedure tags the retransmission queue when SACKs arrive.
961 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
962 * Packets in queue with these bits set are counted in variables
963 * sacked_out, retrans_out and lost_out, correspondingly.
965 * Valid combinations are:
966 * Tag InFlight Description
967 * 0 1 - orig segment is in flight.
968 * S 0 - nothing flies, orig reached receiver.
969 * L 0 - nothing flies, orig lost by net.
970 * R 2 - both orig and retransmit are in flight.
971 * L|R 1 - orig is lost, retransmit is in flight.
972 * S|R 1 - orig reached receiver, retrans is still in flight.
973 * (L|S|R is logically valid, it could occur when L|R is sacked,
974 * but it is equivalent to plain S and code short-curcuits it to S.
975 * L|S is logically invalid, it would mean -1 packet in flight 8))
977 * These 6 states form finite state machine, controlled by the following events:
978 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
979 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
980 * 3. Loss detection event of two flavors:
981 * A. Scoreboard estimator decided the packet is lost.
982 * A'. Reno "three dupacks" marks head of queue lost.
983 * B. SACK arrives sacking SND.NXT at the moment, when the
984 * segment was retransmitted.
985 * 4. D-SACK added new rule: D-SACK changes any tag to S.
987 * It is pleasant to note, that state diagram turns out to be commutative,
988 * so that we are allowed not to be bothered by order of our actions,
989 * when multiple events arrive simultaneously. (see the function below).
991 * Reordering detection.
992 * --------------------
993 * Reordering metric is maximal distance, which a packet can be displaced
994 * in packet stream. With SACKs we can estimate it:
996 * 1. SACK fills old hole and the corresponding segment was not
997 * ever retransmitted -> reordering. Alas, we cannot use it
998 * when segment was retransmitted.
999 * 2. The last flaw is solved with D-SACK. D-SACK arrives
1000 * for retransmitted and already SACKed segment -> reordering..
1001 * Both of these heuristics are not used in Loss state, when we cannot
1002 * account for retransmits accurately.
1004 * SACK block validation.
1005 * ----------------------
1007 * SACK block range validation checks that the received SACK block fits to
1008 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1009 * Note that SND.UNA is not included to the range though being valid because
1010 * it means that the receiver is rather inconsistent with itself reporting
1011 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1012 * perfectly valid, however, in light of RFC2018 which explicitly states
1013 * that "SACK block MUST reflect the newest segment. Even if the newest
1014 * segment is going to be discarded ...", not that it looks very clever
1015 * in case of head skb. Due to potentional receiver driven attacks, we
1016 * choose to avoid immediate execution of a walk in write queue due to
1017 * reneging and defer head skb's loss recovery to standard loss recovery
1018 * procedure that will eventually trigger (nothing forbids us doing this).
1020 * Implements also blockage to start_seq wrap-around. Problem lies in the
1021 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1022 * there's no guarantee that it will be before snd_nxt (n). The problem
1023 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1026 * <- outs wnd -> <- wrapzone ->
1027 * u e n u_w e_w s n_w
1029 * |<------------+------+----- TCP seqno space --------------+---------->|
1030 * ...-- <2^31 ->| |<--------...
1031 * ...---- >2^31 ------>| |<--------...
1033 * Current code wouldn't be vulnerable but it's better still to discard such
1034 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1035 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1036 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1037 * equal to the ideal case (infinite seqno space without wrap caused issues).
1039 * With D-SACK the lower bound is extended to cover sequence space below
1040 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1041 * again, D-SACK block must not to go across snd_una (for the same reason as
1042 * for the normal SACK blocks, explained above). But there all simplicity
1043 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1044 * fully below undo_marker they do not affect behavior in anyway and can
1045 * therefore be safely ignored. In rare cases (which are more or less
1046 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1047 * fragmentation and packet reordering past skb's retransmission. To consider
1048 * them correctly, the acceptable range must be extended even more though
1049 * the exact amount is rather hard to quantify. However, tp->max_window can
1050 * be used as an exaggerated estimate.
1052 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
1053 u32 start_seq, u32 end_seq)
1055 /* Too far in future, or reversed (interpretation is ambiguous) */
1056 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1059 /* Nasty start_seq wrap-around check (see comments above) */
1060 if (!before(start_seq, tp->snd_nxt))
1063 /* In outstanding window? ...This is valid exit for D-SACKs too.
1064 * start_seq == snd_una is non-sensical (see comments above)
1066 if (after(start_seq, tp->snd_una))
1069 if (!is_dsack || !tp->undo_marker)
1072 /* ...Then it's D-SACK, and must reside below snd_una completely */
1073 if (after(end_seq, tp->snd_una))
1076 if (!before(start_seq, tp->undo_marker))
1080 if (!after(end_seq, tp->undo_marker))
1083 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1084 * start_seq < undo_marker and end_seq >= undo_marker.
1086 return !before(start_seq, end_seq - tp->max_window);
1089 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1090 struct tcp_sack_block_wire *sp, int num_sacks,
1093 struct tcp_sock *tp = tcp_sk(sk);
1094 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1095 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1096 bool dup_sack = false;
1098 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1101 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1102 } else if (num_sacks > 1) {
1103 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1104 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1106 if (!after(end_seq_0, end_seq_1) &&
1107 !before(start_seq_0, start_seq_1)) {
1110 NET_INC_STATS(sock_net(sk),
1111 LINUX_MIB_TCPDSACKOFORECV);
1115 /* D-SACK for already forgotten data... Do dumb counting. */
1116 if (dup_sack && tp->undo_marker && tp->undo_retrans > 0 &&
1117 !after(end_seq_0, prior_snd_una) &&
1118 after(end_seq_0, tp->undo_marker))
1124 struct tcp_sacktag_state {
1126 /* Timestamps for earliest and latest never-retransmitted segment
1127 * that was SACKed. RTO needs the earliest RTT to stay conservative,
1128 * but congestion control should still get an accurate delay signal.
1132 struct rate_sample *rate;
1134 unsigned int mss_now;
1137 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1138 * the incoming SACK may not exactly match but we can find smaller MSS
1139 * aligned portion of it that matches. Therefore we might need to fragment
1140 * which may fail and creates some hassle (caller must handle error case
1143 * FIXME: this could be merged to shift decision code
1145 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1146 u32 start_seq, u32 end_seq)
1150 unsigned int pkt_len;
1153 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1154 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1156 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1157 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1158 mss = tcp_skb_mss(skb);
1159 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1162 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1166 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1171 /* Round if necessary so that SACKs cover only full MSSes
1172 * and/or the remaining small portion (if present)
1174 if (pkt_len > mss) {
1175 unsigned int new_len = (pkt_len / mss) * mss;
1176 if (!in_sack && new_len < pkt_len)
1181 if (pkt_len >= skb->len && !in_sack)
1184 err = tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
1185 pkt_len, mss, GFP_ATOMIC);
1193 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1194 static u8 tcp_sacktag_one(struct sock *sk,
1195 struct tcp_sacktag_state *state, u8 sacked,
1196 u32 start_seq, u32 end_seq,
1197 int dup_sack, int pcount,
1200 struct tcp_sock *tp = tcp_sk(sk);
1202 /* Account D-SACK for retransmitted packet. */
1203 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1204 if (tp->undo_marker && tp->undo_retrans > 0 &&
1205 after(end_seq, tp->undo_marker))
1207 if ((sacked & TCPCB_SACKED_ACKED) &&
1208 before(start_seq, state->reord))
1209 state->reord = start_seq;
1212 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1213 if (!after(end_seq, tp->snd_una))
1216 if (!(sacked & TCPCB_SACKED_ACKED)) {
1217 tcp_rack_advance(tp, sacked, end_seq, xmit_time);
1219 if (sacked & TCPCB_SACKED_RETRANS) {
1220 /* If the segment is not tagged as lost,
1221 * we do not clear RETRANS, believing
1222 * that retransmission is still in flight.
1224 if (sacked & TCPCB_LOST) {
1225 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1226 tp->lost_out -= pcount;
1227 tp->retrans_out -= pcount;
1230 if (!(sacked & TCPCB_RETRANS)) {
1231 /* New sack for not retransmitted frame,
1232 * which was in hole. It is reordering.
1234 if (before(start_seq,
1235 tcp_highest_sack_seq(tp)) &&
1236 before(start_seq, state->reord))
1237 state->reord = start_seq;
1239 if (!after(end_seq, tp->high_seq))
1240 state->flag |= FLAG_ORIG_SACK_ACKED;
1241 if (state->first_sackt == 0)
1242 state->first_sackt = xmit_time;
1243 state->last_sackt = xmit_time;
1246 if (sacked & TCPCB_LOST) {
1247 sacked &= ~TCPCB_LOST;
1248 tp->lost_out -= pcount;
1252 sacked |= TCPCB_SACKED_ACKED;
1253 state->flag |= FLAG_DATA_SACKED;
1254 tp->sacked_out += pcount;
1255 tp->delivered += pcount; /* Out-of-order packets delivered */
1257 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1258 if (tp->lost_skb_hint &&
1259 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1260 tp->lost_cnt_hint += pcount;
1263 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1264 * frames and clear it. undo_retrans is decreased above, L|R frames
1265 * are accounted above as well.
1267 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1268 sacked &= ~TCPCB_SACKED_RETRANS;
1269 tp->retrans_out -= pcount;
1275 /* Shift newly-SACKed bytes from this skb to the immediately previous
1276 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1278 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *prev,
1279 struct sk_buff *skb,
1280 struct tcp_sacktag_state *state,
1281 unsigned int pcount, int shifted, int mss,
1284 struct tcp_sock *tp = tcp_sk(sk);
1285 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1286 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
1290 /* Adjust counters and hints for the newly sacked sequence
1291 * range but discard the return value since prev is already
1292 * marked. We must tag the range first because the seq
1293 * advancement below implicitly advances
1294 * tcp_highest_sack_seq() when skb is highest_sack.
1296 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1297 start_seq, end_seq, dup_sack, pcount,
1299 tcp_rate_skb_delivered(sk, skb, state->rate);
1301 if (skb == tp->lost_skb_hint)
1302 tp->lost_cnt_hint += pcount;
1304 TCP_SKB_CB(prev)->end_seq += shifted;
1305 TCP_SKB_CB(skb)->seq += shifted;
1307 tcp_skb_pcount_add(prev, pcount);
1308 BUG_ON(tcp_skb_pcount(skb) < pcount);
1309 tcp_skb_pcount_add(skb, -pcount);
1311 /* When we're adding to gso_segs == 1, gso_size will be zero,
1312 * in theory this shouldn't be necessary but as long as DSACK
1313 * code can come after this skb later on it's better to keep
1314 * setting gso_size to something.
1316 if (!TCP_SKB_CB(prev)->tcp_gso_size)
1317 TCP_SKB_CB(prev)->tcp_gso_size = mss;
1319 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1320 if (tcp_skb_pcount(skb) <= 1)
1321 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1323 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1324 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1327 BUG_ON(!tcp_skb_pcount(skb));
1328 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1332 /* Whole SKB was eaten :-) */
1334 if (skb == tp->retransmit_skb_hint)
1335 tp->retransmit_skb_hint = prev;
1336 if (skb == tp->lost_skb_hint) {
1337 tp->lost_skb_hint = prev;
1338 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1341 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1342 TCP_SKB_CB(prev)->eor = TCP_SKB_CB(skb)->eor;
1343 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1344 TCP_SKB_CB(prev)->end_seq++;
1346 if (skb == tcp_highest_sack(sk))
1347 tcp_advance_highest_sack(sk, skb);
1349 tcp_skb_collapse_tstamp(prev, skb);
1350 if (unlikely(TCP_SKB_CB(prev)->tx.delivered_mstamp))
1351 TCP_SKB_CB(prev)->tx.delivered_mstamp = 0;
1353 tcp_rtx_queue_unlink_and_free(skb, sk);
1355 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKMERGED);
1360 /* I wish gso_size would have a bit more sane initialization than
1361 * something-or-zero which complicates things
1363 static int tcp_skb_seglen(const struct sk_buff *skb)
1365 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1368 /* Shifting pages past head area doesn't work */
1369 static int skb_can_shift(const struct sk_buff *skb)
1371 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1374 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1377 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1378 struct tcp_sacktag_state *state,
1379 u32 start_seq, u32 end_seq,
1382 struct tcp_sock *tp = tcp_sk(sk);
1383 struct sk_buff *prev;
1389 /* Normally R but no L won't result in plain S */
1391 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1393 if (!skb_can_shift(skb))
1395 /* This frame is about to be dropped (was ACKed). */
1396 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1399 /* Can only happen with delayed DSACK + discard craziness */
1400 prev = skb_rb_prev(skb);
1404 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1407 if (!tcp_skb_can_collapse_to(prev))
1410 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1411 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1415 pcount = tcp_skb_pcount(skb);
1416 mss = tcp_skb_seglen(skb);
1418 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1419 * drop this restriction as unnecessary
1421 if (mss != tcp_skb_seglen(prev))
1424 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1426 /* CHECKME: This is non-MSS split case only?, this will
1427 * cause skipped skbs due to advancing loop btw, original
1428 * has that feature too
1430 if (tcp_skb_pcount(skb) <= 1)
1433 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1435 /* TODO: head merge to next could be attempted here
1436 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1437 * though it might not be worth of the additional hassle
1439 * ...we can probably just fallback to what was done
1440 * previously. We could try merging non-SACKed ones
1441 * as well but it probably isn't going to buy off
1442 * because later SACKs might again split them, and
1443 * it would make skb timestamp tracking considerably
1449 len = end_seq - TCP_SKB_CB(skb)->seq;
1451 BUG_ON(len > skb->len);
1453 /* MSS boundaries should be honoured or else pcount will
1454 * severely break even though it makes things bit trickier.
1455 * Optimize common case to avoid most of the divides
1457 mss = tcp_skb_mss(skb);
1459 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1460 * drop this restriction as unnecessary
1462 if (mss != tcp_skb_seglen(prev))
1467 } else if (len < mss) {
1475 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1476 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1479 if (!skb_shift(prev, skb, len))
1481 if (!tcp_shifted_skb(sk, prev, skb, state, pcount, len, mss, dup_sack))
1484 /* Hole filled allows collapsing with the next as well, this is very
1485 * useful when hole on every nth skb pattern happens
1487 skb = skb_rb_next(prev);
1491 if (!skb_can_shift(skb) ||
1492 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1493 (mss != tcp_skb_seglen(skb)))
1497 if (skb_shift(prev, skb, len)) {
1498 pcount += tcp_skb_pcount(skb);
1499 tcp_shifted_skb(sk, prev, skb, state, tcp_skb_pcount(skb),
1510 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1514 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1515 struct tcp_sack_block *next_dup,
1516 struct tcp_sacktag_state *state,
1517 u32 start_seq, u32 end_seq,
1520 struct tcp_sock *tp = tcp_sk(sk);
1521 struct sk_buff *tmp;
1523 skb_rbtree_walk_from(skb) {
1525 bool dup_sack = dup_sack_in;
1527 /* queue is in-order => we can short-circuit the walk early */
1528 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1532 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1533 in_sack = tcp_match_skb_to_sack(sk, skb,
1534 next_dup->start_seq,
1540 /* skb reference here is a bit tricky to get right, since
1541 * shifting can eat and free both this skb and the next,
1542 * so not even _safe variant of the loop is enough.
1545 tmp = tcp_shift_skb_data(sk, skb, state,
1546 start_seq, end_seq, dup_sack);
1555 in_sack = tcp_match_skb_to_sack(sk, skb,
1561 if (unlikely(in_sack < 0))
1565 TCP_SKB_CB(skb)->sacked =
1568 TCP_SKB_CB(skb)->sacked,
1569 TCP_SKB_CB(skb)->seq,
1570 TCP_SKB_CB(skb)->end_seq,
1572 tcp_skb_pcount(skb),
1574 tcp_rate_skb_delivered(sk, skb, state->rate);
1575 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1576 list_del_init(&skb->tcp_tsorted_anchor);
1578 if (!before(TCP_SKB_CB(skb)->seq,
1579 tcp_highest_sack_seq(tp)))
1580 tcp_advance_highest_sack(sk, skb);
1586 static struct sk_buff *tcp_sacktag_bsearch(struct sock *sk,
1587 struct tcp_sacktag_state *state,
1590 struct rb_node *parent, **p = &sk->tcp_rtx_queue.rb_node;
1591 struct sk_buff *skb;
1595 skb = rb_to_skb(parent);
1596 if (before(seq, TCP_SKB_CB(skb)->seq)) {
1597 p = &parent->rb_left;
1600 if (!before(seq, TCP_SKB_CB(skb)->end_seq)) {
1601 p = &parent->rb_right;
1609 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1610 struct tcp_sacktag_state *state,
1613 if (skb && after(TCP_SKB_CB(skb)->seq, skip_to_seq))
1616 return tcp_sacktag_bsearch(sk, state, skip_to_seq);
1619 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1621 struct tcp_sack_block *next_dup,
1622 struct tcp_sacktag_state *state,
1628 if (before(next_dup->start_seq, skip_to_seq)) {
1629 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1630 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1631 next_dup->start_seq, next_dup->end_seq,
1638 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1640 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1644 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1645 u32 prior_snd_una, struct tcp_sacktag_state *state)
1647 struct tcp_sock *tp = tcp_sk(sk);
1648 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1649 TCP_SKB_CB(ack_skb)->sacked);
1650 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1651 struct tcp_sack_block sp[TCP_NUM_SACKS];
1652 struct tcp_sack_block *cache;
1653 struct sk_buff *skb;
1654 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1656 bool found_dup_sack = false;
1658 int first_sack_index;
1661 state->reord = tp->snd_nxt;
1663 if (!tp->sacked_out)
1664 tcp_highest_sack_reset(sk);
1666 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1667 num_sacks, prior_snd_una);
1668 if (found_dup_sack) {
1669 state->flag |= FLAG_DSACKING_ACK;
1670 tp->delivered++; /* A spurious retransmission is delivered */
1673 /* Eliminate too old ACKs, but take into
1674 * account more or less fresh ones, they can
1675 * contain valid SACK info.
1677 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1680 if (!tp->packets_out)
1684 first_sack_index = 0;
1685 for (i = 0; i < num_sacks; i++) {
1686 bool dup_sack = !i && found_dup_sack;
1688 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1689 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1691 if (!tcp_is_sackblock_valid(tp, dup_sack,
1692 sp[used_sacks].start_seq,
1693 sp[used_sacks].end_seq)) {
1697 if (!tp->undo_marker)
1698 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1700 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1702 /* Don't count olds caused by ACK reordering */
1703 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1704 !after(sp[used_sacks].end_seq, tp->snd_una))
1706 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1709 NET_INC_STATS(sock_net(sk), mib_idx);
1711 first_sack_index = -1;
1715 /* Ignore very old stuff early */
1716 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1722 /* order SACK blocks to allow in order walk of the retrans queue */
1723 for (i = used_sacks - 1; i > 0; i--) {
1724 for (j = 0; j < i; j++) {
1725 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1726 swap(sp[j], sp[j + 1]);
1728 /* Track where the first SACK block goes to */
1729 if (j == first_sack_index)
1730 first_sack_index = j + 1;
1735 state->mss_now = tcp_current_mss(sk);
1739 if (!tp->sacked_out) {
1740 /* It's already past, so skip checking against it */
1741 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1743 cache = tp->recv_sack_cache;
1744 /* Skip empty blocks in at head of the cache */
1745 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1750 while (i < used_sacks) {
1751 u32 start_seq = sp[i].start_seq;
1752 u32 end_seq = sp[i].end_seq;
1753 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1754 struct tcp_sack_block *next_dup = NULL;
1756 if (found_dup_sack && ((i + 1) == first_sack_index))
1757 next_dup = &sp[i + 1];
1759 /* Skip too early cached blocks */
1760 while (tcp_sack_cache_ok(tp, cache) &&
1761 !before(start_seq, cache->end_seq))
1764 /* Can skip some work by looking recv_sack_cache? */
1765 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1766 after(end_seq, cache->start_seq)) {
1769 if (before(start_seq, cache->start_seq)) {
1770 skb = tcp_sacktag_skip(skb, sk, state,
1772 skb = tcp_sacktag_walk(skb, sk, next_dup,
1779 /* Rest of the block already fully processed? */
1780 if (!after(end_seq, cache->end_seq))
1783 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1787 /* ...tail remains todo... */
1788 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1789 /* ...but better entrypoint exists! */
1790 skb = tcp_highest_sack(sk);
1797 skb = tcp_sacktag_skip(skb, sk, state, cache->end_seq);
1798 /* Check overlap against next cached too (past this one already) */
1803 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1804 skb = tcp_highest_sack(sk);
1808 skb = tcp_sacktag_skip(skb, sk, state, start_seq);
1811 skb = tcp_sacktag_walk(skb, sk, next_dup, state,
1812 start_seq, end_seq, dup_sack);
1818 /* Clear the head of the cache sack blocks so we can skip it next time */
1819 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1820 tp->recv_sack_cache[i].start_seq = 0;
1821 tp->recv_sack_cache[i].end_seq = 0;
1823 for (j = 0; j < used_sacks; j++)
1824 tp->recv_sack_cache[i++] = sp[j];
1826 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Loss || tp->undo_marker)
1827 tcp_check_sack_reordering(sk, state->reord, 0);
1829 tcp_verify_left_out(tp);
1832 #if FASTRETRANS_DEBUG > 0
1833 WARN_ON((int)tp->sacked_out < 0);
1834 WARN_ON((int)tp->lost_out < 0);
1835 WARN_ON((int)tp->retrans_out < 0);
1836 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1841 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1842 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1844 static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
1848 holes = max(tp->lost_out, 1U);
1849 holes = min(holes, tp->packets_out);
1851 if ((tp->sacked_out + holes) > tp->packets_out) {
1852 tp->sacked_out = tp->packets_out - holes;
1858 /* If we receive more dupacks than we expected counting segments
1859 * in assumption of absent reordering, interpret this as reordering.
1860 * The only another reason could be bug in receiver TCP.
1862 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1864 struct tcp_sock *tp = tcp_sk(sk);
1866 if (!tcp_limit_reno_sacked(tp))
1869 tp->reordering = min_t(u32, tp->packets_out + addend,
1870 sock_net(sk)->ipv4.sysctl_tcp_max_reordering);
1871 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRENOREORDER);
1874 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1876 static void tcp_add_reno_sack(struct sock *sk)
1878 struct tcp_sock *tp = tcp_sk(sk);
1879 u32 prior_sacked = tp->sacked_out;
1882 tcp_check_reno_reordering(sk, 0);
1883 if (tp->sacked_out > prior_sacked)
1884 tp->delivered++; /* Some out-of-order packet is delivered */
1885 tcp_verify_left_out(tp);
1888 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1890 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1892 struct tcp_sock *tp = tcp_sk(sk);
1895 /* One ACK acked hole. The rest eat duplicate ACKs. */
1896 tp->delivered += max_t(int, acked - tp->sacked_out, 1);
1897 if (acked - 1 >= tp->sacked_out)
1900 tp->sacked_out -= acked - 1;
1902 tcp_check_reno_reordering(sk, acked);
1903 tcp_verify_left_out(tp);
1906 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1911 void tcp_clear_retrans(struct tcp_sock *tp)
1913 tp->retrans_out = 0;
1915 tp->undo_marker = 0;
1916 tp->undo_retrans = -1;
1920 static inline void tcp_init_undo(struct tcp_sock *tp)
1922 tp->undo_marker = tp->snd_una;
1923 /* Retransmission still in flight may cause DSACKs later. */
1924 tp->undo_retrans = tp->retrans_out ? : -1;
1927 static bool tcp_is_rack(const struct sock *sk)
1929 return sock_net(sk)->ipv4.sysctl_tcp_recovery & TCP_RACK_LOSS_DETECTION;
1932 /* If we detect SACK reneging, forget all SACK information
1933 * and reset tags completely, otherwise preserve SACKs. If receiver
1934 * dropped its ofo queue, we will know this due to reneging detection.
1936 static void tcp_timeout_mark_lost(struct sock *sk)
1938 struct tcp_sock *tp = tcp_sk(sk);
1939 struct sk_buff *skb, *head;
1940 bool is_reneg; /* is receiver reneging on SACKs? */
1942 head = tcp_rtx_queue_head(sk);
1943 is_reneg = head && (TCP_SKB_CB(head)->sacked & TCPCB_SACKED_ACKED);
1945 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
1947 /* Mark SACK reneging until we recover from this loss event. */
1948 tp->is_sack_reneg = 1;
1949 } else if (tcp_is_reno(tp)) {
1950 tcp_reset_reno_sack(tp);
1954 skb_rbtree_walk_from(skb) {
1956 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1957 else if (tcp_is_rack(sk) && skb != head &&
1958 tcp_rack_skb_timeout(tp, skb, 0) > 0)
1959 continue; /* Don't mark recently sent ones lost yet */
1960 tcp_mark_skb_lost(sk, skb);
1962 tcp_verify_left_out(tp);
1963 tcp_clear_all_retrans_hints(tp);
1966 /* Enter Loss state. */
1967 void tcp_enter_loss(struct sock *sk)
1969 const struct inet_connection_sock *icsk = inet_csk(sk);
1970 struct tcp_sock *tp = tcp_sk(sk);
1971 struct net *net = sock_net(sk);
1972 bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery;
1974 tcp_timeout_mark_lost(sk);
1976 /* Reduce ssthresh if it has not yet been made inside this window. */
1977 if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
1978 !after(tp->high_seq, tp->snd_una) ||
1979 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1980 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1981 tp->prior_cwnd = tp->snd_cwnd;
1982 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1983 tcp_ca_event(sk, CA_EVENT_LOSS);
1986 tp->snd_cwnd = tcp_packets_in_flight(tp) + 1;
1987 tp->snd_cwnd_cnt = 0;
1988 tp->snd_cwnd_stamp = tcp_jiffies32;
1990 /* Timeout in disordered state after receiving substantial DUPACKs
1991 * suggests that the degree of reordering is over-estimated.
1993 if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
1994 tp->sacked_out >= net->ipv4.sysctl_tcp_reordering)
1995 tp->reordering = min_t(unsigned int, tp->reordering,
1996 net->ipv4.sysctl_tcp_reordering);
1997 tcp_set_ca_state(sk, TCP_CA_Loss);
1998 tp->high_seq = tp->snd_nxt;
1999 tcp_ecn_queue_cwr(tp);
2001 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
2002 * loss recovery is underway except recurring timeout(s) on
2003 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
2005 tp->frto = net->ipv4.sysctl_tcp_frto &&
2006 (new_recovery || icsk->icsk_retransmits) &&
2007 !inet_csk(sk)->icsk_mtup.probe_size;
2010 /* If ACK arrived pointing to a remembered SACK, it means that our
2011 * remembered SACKs do not reflect real state of receiver i.e.
2012 * receiver _host_ is heavily congested (or buggy).
2014 * To avoid big spurious retransmission bursts due to transient SACK
2015 * scoreboard oddities that look like reneging, we give the receiver a
2016 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
2017 * restore sanity to the SACK scoreboard. If the apparent reneging
2018 * persists until this RTO then we'll clear the SACK scoreboard.
2020 static bool tcp_check_sack_reneging(struct sock *sk, int flag)
2022 if (flag & FLAG_SACK_RENEGING) {
2023 struct tcp_sock *tp = tcp_sk(sk);
2024 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4),
2025 msecs_to_jiffies(10));
2027 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2028 delay, TCP_RTO_MAX);
2034 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2035 * counter when SACK is enabled (without SACK, sacked_out is used for
2038 * With reordering, holes may still be in flight, so RFC3517 recovery
2039 * uses pure sacked_out (total number of SACKed segments) even though
2040 * it violates the RFC that uses duplicate ACKs, often these are equal
2041 * but when e.g. out-of-window ACKs or packet duplication occurs,
2042 * they differ. Since neither occurs due to loss, TCP should really
2045 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2047 return tp->sacked_out + 1;
2050 /* Linux NewReno/SACK/ECN state machine.
2051 * --------------------------------------
2053 * "Open" Normal state, no dubious events, fast path.
2054 * "Disorder" In all the respects it is "Open",
2055 * but requires a bit more attention. It is entered when
2056 * we see some SACKs or dupacks. It is split of "Open"
2057 * mainly to move some processing from fast path to slow one.
2058 * "CWR" CWND was reduced due to some Congestion Notification event.
2059 * It can be ECN, ICMP source quench, local device congestion.
2060 * "Recovery" CWND was reduced, we are fast-retransmitting.
2061 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2063 * tcp_fastretrans_alert() is entered:
2064 * - each incoming ACK, if state is not "Open"
2065 * - when arrived ACK is unusual, namely:
2070 * Counting packets in flight is pretty simple.
2072 * in_flight = packets_out - left_out + retrans_out
2074 * packets_out is SND.NXT-SND.UNA counted in packets.
2076 * retrans_out is number of retransmitted segments.
2078 * left_out is number of segments left network, but not ACKed yet.
2080 * left_out = sacked_out + lost_out
2082 * sacked_out: Packets, which arrived to receiver out of order
2083 * and hence not ACKed. With SACKs this number is simply
2084 * amount of SACKed data. Even without SACKs
2085 * it is easy to give pretty reliable estimate of this number,
2086 * counting duplicate ACKs.
2088 * lost_out: Packets lost by network. TCP has no explicit
2089 * "loss notification" feedback from network (for now).
2090 * It means that this number can be only _guessed_.
2091 * Actually, it is the heuristics to predict lossage that
2092 * distinguishes different algorithms.
2094 * F.e. after RTO, when all the queue is considered as lost,
2095 * lost_out = packets_out and in_flight = retrans_out.
2097 * Essentially, we have now a few algorithms detecting
2100 * If the receiver supports SACK:
2102 * RFC6675/3517: It is the conventional algorithm. A packet is
2103 * considered lost if the number of higher sequence packets
2104 * SACKed is greater than or equal the DUPACK thoreshold
2105 * (reordering). This is implemented in tcp_mark_head_lost and
2106 * tcp_update_scoreboard.
2108 * RACK (draft-ietf-tcpm-rack-01): it is a newer algorithm
2109 * (2017-) that checks timing instead of counting DUPACKs.
2110 * Essentially a packet is considered lost if it's not S/ACKed
2111 * after RTT + reordering_window, where both metrics are
2112 * dynamically measured and adjusted. This is implemented in
2113 * tcp_rack_mark_lost.
2115 * If the receiver does not support SACK:
2117 * NewReno (RFC6582): in Recovery we assume that one segment
2118 * is lost (classic Reno). While we are in Recovery and
2119 * a partial ACK arrives, we assume that one more packet
2120 * is lost (NewReno). This heuristics are the same in NewReno
2123 * Really tricky (and requiring careful tuning) part of algorithm
2124 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2125 * The first determines the moment _when_ we should reduce CWND and,
2126 * hence, slow down forward transmission. In fact, it determines the moment
2127 * when we decide that hole is caused by loss, rather than by a reorder.
2129 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2130 * holes, caused by lost packets.
2132 * And the most logically complicated part of algorithm is undo
2133 * heuristics. We detect false retransmits due to both too early
2134 * fast retransmit (reordering) and underestimated RTO, analyzing
2135 * timestamps and D-SACKs. When we detect that some segments were
2136 * retransmitted by mistake and CWND reduction was wrong, we undo
2137 * window reduction and abort recovery phase. This logic is hidden
2138 * inside several functions named tcp_try_undo_<something>.
2141 /* This function decides, when we should leave Disordered state
2142 * and enter Recovery phase, reducing congestion window.
2144 * Main question: may we further continue forward transmission
2145 * with the same cwnd?
2147 static bool tcp_time_to_recover(struct sock *sk, int flag)
2149 struct tcp_sock *tp = tcp_sk(sk);
2151 /* Trick#1: The loss is proven. */
2155 /* Not-A-Trick#2 : Classic rule... */
2156 if (!tcp_is_rack(sk) && tcp_dupack_heuristics(tp) > tp->reordering)
2162 /* Detect loss in event "A" above by marking head of queue up as lost.
2163 * For non-SACK(Reno) senders, the first "packets" number of segments
2164 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2165 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2166 * the maximum SACKed segments to pass before reaching this limit.
2168 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2170 struct tcp_sock *tp = tcp_sk(sk);
2171 struct sk_buff *skb;
2172 int cnt, oldcnt, lost;
2174 /* Use SACK to deduce losses of new sequences sent during recovery */
2175 const u32 loss_high = tcp_is_sack(tp) ? tp->snd_nxt : tp->high_seq;
2177 WARN_ON(packets > tp->packets_out);
2178 skb = tp->lost_skb_hint;
2180 /* Head already handled? */
2181 if (mark_head && after(TCP_SKB_CB(skb)->seq, tp->snd_una))
2183 cnt = tp->lost_cnt_hint;
2185 skb = tcp_rtx_queue_head(sk);
2189 skb_rbtree_walk_from(skb) {
2190 /* TODO: do this better */
2191 /* this is not the most efficient way to do this... */
2192 tp->lost_skb_hint = skb;
2193 tp->lost_cnt_hint = cnt;
2195 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2199 if (tcp_is_reno(tp) ||
2200 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2201 cnt += tcp_skb_pcount(skb);
2203 if (cnt > packets) {
2204 if (tcp_is_sack(tp) ||
2205 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
2206 (oldcnt >= packets))
2209 mss = tcp_skb_mss(skb);
2210 /* If needed, chop off the prefix to mark as lost. */
2211 lost = (packets - oldcnt) * mss;
2212 if (lost < skb->len &&
2213 tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
2214 lost, mss, GFP_ATOMIC) < 0)
2219 tcp_skb_mark_lost(tp, skb);
2224 tcp_verify_left_out(tp);
2227 /* Account newly detected lost packet(s) */
2229 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2231 struct tcp_sock *tp = tcp_sk(sk);
2233 if (tcp_is_sack(tp)) {
2234 int sacked_upto = tp->sacked_out - tp->reordering;
2235 if (sacked_upto >= 0)
2236 tcp_mark_head_lost(sk, sacked_upto, 0);
2237 else if (fast_rexmit)
2238 tcp_mark_head_lost(sk, 1, 1);
2242 static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when)
2244 return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2245 before(tp->rx_opt.rcv_tsecr, when);
2248 /* skb is spurious retransmitted if the returned timestamp echo
2249 * reply is prior to the skb transmission time
2251 static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp,
2252 const struct sk_buff *skb)
2254 return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) &&
2255 tcp_tsopt_ecr_before(tp, tcp_skb_timestamp(skb));
2258 /* Nothing was retransmitted or returned timestamp is less
2259 * than timestamp of the first retransmission.
2261 static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2263 return !tp->retrans_stamp ||
2264 tcp_tsopt_ecr_before(tp, tp->retrans_stamp);
2267 /* Undo procedures. */
2269 /* We can clear retrans_stamp when there are no retransmissions in the
2270 * window. It would seem that it is trivially available for us in
2271 * tp->retrans_out, however, that kind of assumptions doesn't consider
2272 * what will happen if errors occur when sending retransmission for the
2273 * second time. ...It could the that such segment has only
2274 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2275 * the head skb is enough except for some reneging corner cases that
2276 * are not worth the effort.
2278 * Main reason for all this complexity is the fact that connection dying
2279 * time now depends on the validity of the retrans_stamp, in particular,
2280 * that successive retransmissions of a segment must not advance
2281 * retrans_stamp under any conditions.
2283 static bool tcp_any_retrans_done(const struct sock *sk)
2285 const struct tcp_sock *tp = tcp_sk(sk);
2286 struct sk_buff *skb;
2288 if (tp->retrans_out)
2291 skb = tcp_rtx_queue_head(sk);
2292 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2298 static void DBGUNDO(struct sock *sk, const char *msg)
2300 #if FASTRETRANS_DEBUG > 1
2301 struct tcp_sock *tp = tcp_sk(sk);
2302 struct inet_sock *inet = inet_sk(sk);
2304 if (sk->sk_family == AF_INET) {
2305 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2307 &inet->inet_daddr, ntohs(inet->inet_dport),
2308 tp->snd_cwnd, tcp_left_out(tp),
2309 tp->snd_ssthresh, tp->prior_ssthresh,
2312 #if IS_ENABLED(CONFIG_IPV6)
2313 else if (sk->sk_family == AF_INET6) {
2314 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2316 &sk->sk_v6_daddr, ntohs(inet->inet_dport),
2317 tp->snd_cwnd, tcp_left_out(tp),
2318 tp->snd_ssthresh, tp->prior_ssthresh,
2325 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
2327 struct tcp_sock *tp = tcp_sk(sk);
2330 struct sk_buff *skb;
2332 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2333 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2336 tcp_clear_all_retrans_hints(tp);
2339 if (tp->prior_ssthresh) {
2340 const struct inet_connection_sock *icsk = inet_csk(sk);
2342 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2344 if (tp->prior_ssthresh > tp->snd_ssthresh) {
2345 tp->snd_ssthresh = tp->prior_ssthresh;
2346 tcp_ecn_withdraw_cwr(tp);
2349 tp->snd_cwnd_stamp = tcp_jiffies32;
2350 tp->undo_marker = 0;
2351 tp->rack.advanced = 1; /* Force RACK to re-exam losses */
2354 static inline bool tcp_may_undo(const struct tcp_sock *tp)
2356 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2359 /* People celebrate: "We love our President!" */
2360 static bool tcp_try_undo_recovery(struct sock *sk)
2362 struct tcp_sock *tp = tcp_sk(sk);
2364 if (tcp_may_undo(tp)) {
2367 /* Happy end! We did not retransmit anything
2368 * or our original transmission succeeded.
2370 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2371 tcp_undo_cwnd_reduction(sk, false);
2372 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2373 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2375 mib_idx = LINUX_MIB_TCPFULLUNDO;
2377 NET_INC_STATS(sock_net(sk), mib_idx);
2378 } else if (tp->rack.reo_wnd_persist) {
2379 tp->rack.reo_wnd_persist--;
2381 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2382 /* Hold old state until something *above* high_seq
2383 * is ACKed. For Reno it is MUST to prevent false
2384 * fast retransmits (RFC2582). SACK TCP is safe. */
2385 if (!tcp_any_retrans_done(sk))
2386 tp->retrans_stamp = 0;
2389 tcp_set_ca_state(sk, TCP_CA_Open);
2390 tp->is_sack_reneg = 0;
2394 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2395 static bool tcp_try_undo_dsack(struct sock *sk)
2397 struct tcp_sock *tp = tcp_sk(sk);
2399 if (tp->undo_marker && !tp->undo_retrans) {
2400 tp->rack.reo_wnd_persist = min(TCP_RACK_RECOVERY_THRESH,
2401 tp->rack.reo_wnd_persist + 1);
2402 DBGUNDO(sk, "D-SACK");
2403 tcp_undo_cwnd_reduction(sk, false);
2404 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2410 /* Undo during loss recovery after partial ACK or using F-RTO. */
2411 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2413 struct tcp_sock *tp = tcp_sk(sk);
2415 if (frto_undo || tcp_may_undo(tp)) {
2416 tcp_undo_cwnd_reduction(sk, true);
2418 DBGUNDO(sk, "partial loss");
2419 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2421 NET_INC_STATS(sock_net(sk),
2422 LINUX_MIB_TCPSPURIOUSRTOS);
2423 inet_csk(sk)->icsk_retransmits = 0;
2424 if (frto_undo || tcp_is_sack(tp)) {
2425 tcp_set_ca_state(sk, TCP_CA_Open);
2426 tp->is_sack_reneg = 0;
2433 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
2434 * It computes the number of packets to send (sndcnt) based on packets newly
2436 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2437 * cwnd reductions across a full RTT.
2438 * 2) Otherwise PRR uses packet conservation to send as much as delivered.
2439 * But when the retransmits are acked without further losses, PRR
2440 * slow starts cwnd up to ssthresh to speed up the recovery.
2442 static void tcp_init_cwnd_reduction(struct sock *sk)
2444 struct tcp_sock *tp = tcp_sk(sk);
2446 tp->high_seq = tp->snd_nxt;
2447 tp->tlp_high_seq = 0;
2448 tp->snd_cwnd_cnt = 0;
2449 tp->prior_cwnd = tp->snd_cwnd;
2450 tp->prr_delivered = 0;
2452 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2453 tcp_ecn_queue_cwr(tp);
2456 void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked, int flag)
2458 struct tcp_sock *tp = tcp_sk(sk);
2460 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2462 if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd))
2465 tp->prr_delivered += newly_acked_sacked;
2467 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2469 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2470 } else if ((flag & FLAG_RETRANS_DATA_ACKED) &&
2471 !(flag & FLAG_LOST_RETRANS)) {
2472 sndcnt = min_t(int, delta,
2473 max_t(int, tp->prr_delivered - tp->prr_out,
2474 newly_acked_sacked) + 1);
2476 sndcnt = min(delta, newly_acked_sacked);
2478 /* Force a fast retransmit upon entering fast recovery */
2479 sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1));
2480 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2483 static inline void tcp_end_cwnd_reduction(struct sock *sk)
2485 struct tcp_sock *tp = tcp_sk(sk);
2487 if (inet_csk(sk)->icsk_ca_ops->cong_control)
2490 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2491 if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH &&
2492 (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR || tp->undo_marker)) {
2493 tp->snd_cwnd = tp->snd_ssthresh;
2494 tp->snd_cwnd_stamp = tcp_jiffies32;
2496 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2499 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2500 void tcp_enter_cwr(struct sock *sk)
2502 struct tcp_sock *tp = tcp_sk(sk);
2504 tp->prior_ssthresh = 0;
2505 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2506 tp->undo_marker = 0;
2507 tcp_init_cwnd_reduction(sk);
2508 tcp_set_ca_state(sk, TCP_CA_CWR);
2511 EXPORT_SYMBOL(tcp_enter_cwr);
2513 static void tcp_try_keep_open(struct sock *sk)
2515 struct tcp_sock *tp = tcp_sk(sk);
2516 int state = TCP_CA_Open;
2518 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2519 state = TCP_CA_Disorder;
2521 if (inet_csk(sk)->icsk_ca_state != state) {
2522 tcp_set_ca_state(sk, state);
2523 tp->high_seq = tp->snd_nxt;
2527 static void tcp_try_to_open(struct sock *sk, int flag)
2529 struct tcp_sock *tp = tcp_sk(sk);
2531 tcp_verify_left_out(tp);
2533 if (!tcp_any_retrans_done(sk))
2534 tp->retrans_stamp = 0;
2536 if (flag & FLAG_ECE)
2539 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2540 tcp_try_keep_open(sk);
2544 static void tcp_mtup_probe_failed(struct sock *sk)
2546 struct inet_connection_sock *icsk = inet_csk(sk);
2548 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2549 icsk->icsk_mtup.probe_size = 0;
2550 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPFAIL);
2553 static void tcp_mtup_probe_success(struct sock *sk)
2555 struct tcp_sock *tp = tcp_sk(sk);
2556 struct inet_connection_sock *icsk = inet_csk(sk);
2558 /* FIXME: breaks with very large cwnd */
2559 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2560 tp->snd_cwnd = tp->snd_cwnd *
2561 tcp_mss_to_mtu(sk, tp->mss_cache) /
2562 icsk->icsk_mtup.probe_size;
2563 tp->snd_cwnd_cnt = 0;
2564 tp->snd_cwnd_stamp = tcp_jiffies32;
2565 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2567 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2568 icsk->icsk_mtup.probe_size = 0;
2569 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2570 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS);
2573 /* Do a simple retransmit without using the backoff mechanisms in
2574 * tcp_timer. This is used for path mtu discovery.
2575 * The socket is already locked here.
2577 void tcp_simple_retransmit(struct sock *sk)
2579 const struct inet_connection_sock *icsk = inet_csk(sk);
2580 struct tcp_sock *tp = tcp_sk(sk);
2581 struct sk_buff *skb;
2582 unsigned int mss = tcp_current_mss(sk);
2584 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2585 if (tcp_skb_seglen(skb) > mss &&
2586 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2587 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2588 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2589 tp->retrans_out -= tcp_skb_pcount(skb);
2591 tcp_skb_mark_lost_uncond_verify(tp, skb);
2595 tcp_clear_retrans_hints_partial(tp);
2600 if (tcp_is_reno(tp))
2601 tcp_limit_reno_sacked(tp);
2603 tcp_verify_left_out(tp);
2605 /* Don't muck with the congestion window here.
2606 * Reason is that we do not increase amount of _data_
2607 * in network, but units changed and effective
2608 * cwnd/ssthresh really reduced now.
2610 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2611 tp->high_seq = tp->snd_nxt;
2612 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2613 tp->prior_ssthresh = 0;
2614 tp->undo_marker = 0;
2615 tcp_set_ca_state(sk, TCP_CA_Loss);
2617 tcp_xmit_retransmit_queue(sk);
2619 EXPORT_SYMBOL(tcp_simple_retransmit);
2621 void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2623 struct tcp_sock *tp = tcp_sk(sk);
2626 if (tcp_is_reno(tp))
2627 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2629 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2631 NET_INC_STATS(sock_net(sk), mib_idx);
2633 tp->prior_ssthresh = 0;
2636 if (!tcp_in_cwnd_reduction(sk)) {
2638 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2639 tcp_init_cwnd_reduction(sk);
2641 tcp_set_ca_state(sk, TCP_CA_Recovery);
2644 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2645 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2647 static void tcp_process_loss(struct sock *sk, int flag, bool is_dupack,
2650 struct tcp_sock *tp = tcp_sk(sk);
2651 bool recovered = !before(tp->snd_una, tp->high_seq);
2653 if ((flag & FLAG_SND_UNA_ADVANCED) &&
2654 tcp_try_undo_loss(sk, false))
2657 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2658 /* Step 3.b. A timeout is spurious if not all data are
2659 * lost, i.e., never-retransmitted data are (s)acked.
2661 if ((flag & FLAG_ORIG_SACK_ACKED) &&
2662 tcp_try_undo_loss(sk, true))
2665 if (after(tp->snd_nxt, tp->high_seq)) {
2666 if (flag & FLAG_DATA_SACKED || is_dupack)
2667 tp->frto = 0; /* Step 3.a. loss was real */
2668 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2669 tp->high_seq = tp->snd_nxt;
2670 /* Step 2.b. Try send new data (but deferred until cwnd
2671 * is updated in tcp_ack()). Otherwise fall back to
2672 * the conventional recovery.
2674 if (!tcp_write_queue_empty(sk) &&
2675 after(tcp_wnd_end(tp), tp->snd_nxt)) {
2676 *rexmit = REXMIT_NEW;
2684 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2685 tcp_try_undo_recovery(sk);
2688 if (tcp_is_reno(tp)) {
2689 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2690 * delivered. Lower inflight to clock out (re)tranmissions.
2692 if (after(tp->snd_nxt, tp->high_seq) && is_dupack)
2693 tcp_add_reno_sack(sk);
2694 else if (flag & FLAG_SND_UNA_ADVANCED)
2695 tcp_reset_reno_sack(tp);
2697 *rexmit = REXMIT_LOST;
2700 /* Undo during fast recovery after partial ACK. */
2701 static bool tcp_try_undo_partial(struct sock *sk, u32 prior_snd_una)
2703 struct tcp_sock *tp = tcp_sk(sk);
2705 if (tp->undo_marker && tcp_packet_delayed(tp)) {
2706 /* Plain luck! Hole if filled with delayed
2707 * packet, rather than with a retransmit. Check reordering.
2709 tcp_check_sack_reordering(sk, prior_snd_una, 1);
2711 /* We are getting evidence that the reordering degree is higher
2712 * than we realized. If there are no retransmits out then we
2713 * can undo. Otherwise we clock out new packets but do not
2714 * mark more packets lost or retransmit more.
2716 if (tp->retrans_out)
2719 if (!tcp_any_retrans_done(sk))
2720 tp->retrans_stamp = 0;
2722 DBGUNDO(sk, "partial recovery");
2723 tcp_undo_cwnd_reduction(sk, true);
2724 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2725 tcp_try_keep_open(sk);
2731 static void tcp_identify_packet_loss(struct sock *sk, int *ack_flag)
2733 struct tcp_sock *tp = tcp_sk(sk);
2735 if (tcp_rtx_queue_empty(sk))
2738 if (unlikely(tcp_is_reno(tp))) {
2739 tcp_newreno_mark_lost(sk, *ack_flag & FLAG_SND_UNA_ADVANCED);
2740 } else if (tcp_is_rack(sk)) {
2741 u32 prior_retrans = tp->retrans_out;
2743 tcp_rack_mark_lost(sk);
2744 if (prior_retrans > tp->retrans_out)
2745 *ack_flag |= FLAG_LOST_RETRANS;
2749 static bool tcp_force_fast_retransmit(struct sock *sk)
2751 struct tcp_sock *tp = tcp_sk(sk);
2753 return after(tcp_highest_sack_seq(tp),
2754 tp->snd_una + tp->reordering * tp->mss_cache);
2757 /* Process an event, which can update packets-in-flight not trivially.
2758 * Main goal of this function is to calculate new estimate for left_out,
2759 * taking into account both packets sitting in receiver's buffer and
2760 * packets lost by network.
2762 * Besides that it updates the congestion state when packet loss or ECN
2763 * is detected. But it does not reduce the cwnd, it is done by the
2764 * congestion control later.
2766 * It does _not_ decide what to send, it is made in function
2767 * tcp_xmit_retransmit_queue().
2769 static void tcp_fastretrans_alert(struct sock *sk, const u32 prior_snd_una,
2770 bool is_dupack, int *ack_flag, int *rexmit)
2772 struct inet_connection_sock *icsk = inet_csk(sk);
2773 struct tcp_sock *tp = tcp_sk(sk);
2774 int fast_rexmit = 0, flag = *ack_flag;
2775 bool do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2776 tcp_force_fast_retransmit(sk));
2778 if (!tp->packets_out && tp->sacked_out)
2781 /* Now state machine starts.
2782 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2783 if (flag & FLAG_ECE)
2784 tp->prior_ssthresh = 0;
2786 /* B. In all the states check for reneging SACKs. */
2787 if (tcp_check_sack_reneging(sk, flag))
2790 /* C. Check consistency of the current state. */
2791 tcp_verify_left_out(tp);
2793 /* D. Check state exit conditions. State can be terminated
2794 * when high_seq is ACKed. */
2795 if (icsk->icsk_ca_state == TCP_CA_Open) {
2796 WARN_ON(tp->retrans_out != 0);
2797 tp->retrans_stamp = 0;
2798 } else if (!before(tp->snd_una, tp->high_seq)) {
2799 switch (icsk->icsk_ca_state) {
2801 /* CWR is to be held something *above* high_seq
2802 * is ACKed for CWR bit to reach receiver. */
2803 if (tp->snd_una != tp->high_seq) {
2804 tcp_end_cwnd_reduction(sk);
2805 tcp_set_ca_state(sk, TCP_CA_Open);
2809 case TCP_CA_Recovery:
2810 if (tcp_is_reno(tp))
2811 tcp_reset_reno_sack(tp);
2812 if (tcp_try_undo_recovery(sk))
2814 tcp_end_cwnd_reduction(sk);
2819 /* E. Process state. */
2820 switch (icsk->icsk_ca_state) {
2821 case TCP_CA_Recovery:
2822 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2823 if (tcp_is_reno(tp) && is_dupack)
2824 tcp_add_reno_sack(sk);
2826 if (tcp_try_undo_partial(sk, prior_snd_una))
2828 /* Partial ACK arrived. Force fast retransmit. */
2829 do_lost = tcp_is_reno(tp) ||
2830 tcp_force_fast_retransmit(sk);
2832 if (tcp_try_undo_dsack(sk)) {
2833 tcp_try_keep_open(sk);
2836 tcp_identify_packet_loss(sk, ack_flag);
2839 tcp_process_loss(sk, flag, is_dupack, rexmit);
2840 tcp_identify_packet_loss(sk, ack_flag);
2841 if (!(icsk->icsk_ca_state == TCP_CA_Open ||
2842 (*ack_flag & FLAG_LOST_RETRANS)))
2844 /* Change state if cwnd is undone or retransmits are lost */
2847 if (tcp_is_reno(tp)) {
2848 if (flag & FLAG_SND_UNA_ADVANCED)
2849 tcp_reset_reno_sack(tp);
2851 tcp_add_reno_sack(sk);
2854 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
2855 tcp_try_undo_dsack(sk);
2857 tcp_identify_packet_loss(sk, ack_flag);
2858 if (!tcp_time_to_recover(sk, flag)) {
2859 tcp_try_to_open(sk, flag);
2863 /* MTU probe failure: don't reduce cwnd */
2864 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2865 icsk->icsk_mtup.probe_size &&
2866 tp->snd_una == tp->mtu_probe.probe_seq_start) {
2867 tcp_mtup_probe_failed(sk);
2868 /* Restores the reduction we did in tcp_mtup_probe() */
2870 tcp_simple_retransmit(sk);
2874 /* Otherwise enter Recovery state */
2875 tcp_enter_recovery(sk, (flag & FLAG_ECE));
2879 if (!tcp_is_rack(sk) && do_lost)
2880 tcp_update_scoreboard(sk, fast_rexmit);
2881 *rexmit = REXMIT_LOST;
2884 static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us, const int flag)
2886 u32 wlen = sock_net(sk)->ipv4.sysctl_tcp_min_rtt_wlen * HZ;
2887 struct tcp_sock *tp = tcp_sk(sk);
2889 if ((flag & FLAG_ACK_MAYBE_DELAYED) && rtt_us > tcp_min_rtt(tp)) {
2890 /* If the remote keeps returning delayed ACKs, eventually
2891 * the min filter would pick it up and overestimate the
2892 * prop. delay when it expires. Skip suspected delayed ACKs.
2896 minmax_running_min(&tp->rtt_min, wlen, tcp_jiffies32,
2897 rtt_us ? : jiffies_to_usecs(1));
2900 static bool tcp_ack_update_rtt(struct sock *sk, const int flag,
2901 long seq_rtt_us, long sack_rtt_us,
2902 long ca_rtt_us, struct rate_sample *rs)
2904 const struct tcp_sock *tp = tcp_sk(sk);
2906 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2907 * broken middle-boxes or peers may corrupt TS-ECR fields. But
2908 * Karn's algorithm forbids taking RTT if some retransmitted data
2909 * is acked (RFC6298).
2912 seq_rtt_us = sack_rtt_us;
2914 /* RTTM Rule: A TSecr value received in a segment is used to
2915 * update the averaged RTT measurement only if the segment
2916 * acknowledges some new data, i.e., only if it advances the
2917 * left edge of the send window.
2918 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2920 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2921 flag & FLAG_ACKED) {
2922 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr;
2923 u32 delta_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
2925 seq_rtt_us = ca_rtt_us = delta_us;
2927 rs->rtt_us = ca_rtt_us; /* RTT of last (S)ACKed packet (or -1) */
2931 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
2932 * always taken together with ACK, SACK, or TS-opts. Any negative
2933 * values will be skipped with the seq_rtt_us < 0 check above.
2935 tcp_update_rtt_min(sk, ca_rtt_us, flag);
2936 tcp_rtt_estimator(sk, seq_rtt_us);
2939 /* RFC6298: only reset backoff on valid RTT measurement. */
2940 inet_csk(sk)->icsk_backoff = 0;
2944 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
2945 void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req)
2947 struct rate_sample rs;
2950 if (req && !req->num_retrans && tcp_rsk(req)->snt_synack)
2951 rtt_us = tcp_stamp_us_delta(tcp_clock_us(), tcp_rsk(req)->snt_synack);
2953 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us, &rs);
2957 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
2959 const struct inet_connection_sock *icsk = inet_csk(sk);
2961 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
2962 tcp_sk(sk)->snd_cwnd_stamp = tcp_jiffies32;
2965 /* Restart timer after forward progress on connection.
2966 * RFC2988 recommends to restart timer to now+rto.
2968 void tcp_rearm_rto(struct sock *sk)
2970 const struct inet_connection_sock *icsk = inet_csk(sk);
2971 struct tcp_sock *tp = tcp_sk(sk);
2973 /* If the retrans timer is currently being used by Fast Open
2974 * for SYN-ACK retrans purpose, stay put.
2976 if (tp->fastopen_rsk)
2979 if (!tp->packets_out) {
2980 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
2982 u32 rto = inet_csk(sk)->icsk_rto;
2983 /* Offset the time elapsed after installing regular RTO */
2984 if (icsk->icsk_pending == ICSK_TIME_REO_TIMEOUT ||
2985 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
2986 s64 delta_us = tcp_rto_delta_us(sk);
2987 /* delta_us may not be positive if the socket is locked
2988 * when the retrans timer fires and is rescheduled.
2990 rto = usecs_to_jiffies(max_t(int, delta_us, 1));
2992 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
2997 /* Try to schedule a loss probe; if that doesn't work, then schedule an RTO. */
2998 static void tcp_set_xmit_timer(struct sock *sk)
3000 if (!tcp_schedule_loss_probe(sk, true))
3004 /* If we get here, the whole TSO packet has not been acked. */
3005 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3007 struct tcp_sock *tp = tcp_sk(sk);
3010 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3012 packets_acked = tcp_skb_pcount(skb);
3013 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3015 packets_acked -= tcp_skb_pcount(skb);
3017 if (packets_acked) {
3018 BUG_ON(tcp_skb_pcount(skb) == 0);
3019 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3022 return packets_acked;
3025 static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
3028 const struct skb_shared_info *shinfo;
3030 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3031 if (likely(!TCP_SKB_CB(skb)->txstamp_ack))
3034 shinfo = skb_shinfo(skb);
3035 if (!before(shinfo->tskey, prior_snd_una) &&
3036 before(shinfo->tskey, tcp_sk(sk)->snd_una)) {
3037 tcp_skb_tsorted_save(skb) {
3038 __skb_tstamp_tx(skb, NULL, sk, SCM_TSTAMP_ACK);
3039 } tcp_skb_tsorted_restore(skb);
3043 /* Remove acknowledged frames from the retransmission queue. If our packet
3044 * is before the ack sequence we can discard it as it's confirmed to have
3045 * arrived at the other end.
3047 static int tcp_clean_rtx_queue(struct sock *sk, u32 prior_fack,
3049 struct tcp_sacktag_state *sack)
3051 const struct inet_connection_sock *icsk = inet_csk(sk);
3052 u64 first_ackt, last_ackt;
3053 struct tcp_sock *tp = tcp_sk(sk);
3054 u32 prior_sacked = tp->sacked_out;
3055 u32 reord = tp->snd_nxt; /* lowest acked un-retx un-sacked seq */
3056 struct sk_buff *skb, *next;
3057 bool fully_acked = true;
3058 long sack_rtt_us = -1L;
3059 long seq_rtt_us = -1L;
3060 long ca_rtt_us = -1L;
3062 u32 last_in_flight = 0;
3068 for (skb = skb_rb_first(&sk->tcp_rtx_queue); skb; skb = next) {
3069 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3070 const u32 start_seq = scb->seq;
3071 u8 sacked = scb->sacked;
3074 tcp_ack_tstamp(sk, skb, prior_snd_una);
3076 /* Determine how many packets and what bytes were acked, tso and else */
3077 if (after(scb->end_seq, tp->snd_una)) {
3078 if (tcp_skb_pcount(skb) == 1 ||
3079 !after(tp->snd_una, scb->seq))
3082 acked_pcount = tcp_tso_acked(sk, skb);
3085 fully_acked = false;
3087 acked_pcount = tcp_skb_pcount(skb);
3090 if (unlikely(sacked & TCPCB_RETRANS)) {
3091 if (sacked & TCPCB_SACKED_RETRANS)
3092 tp->retrans_out -= acked_pcount;
3093 flag |= FLAG_RETRANS_DATA_ACKED;
3094 } else if (!(sacked & TCPCB_SACKED_ACKED)) {
3095 last_ackt = skb->skb_mstamp;
3096 WARN_ON_ONCE(last_ackt == 0);
3098 first_ackt = last_ackt;
3100 last_in_flight = TCP_SKB_CB(skb)->tx.in_flight;
3101 if (before(start_seq, reord))
3103 if (!after(scb->end_seq, tp->high_seq))
3104 flag |= FLAG_ORIG_SACK_ACKED;
3107 if (sacked & TCPCB_SACKED_ACKED) {
3108 tp->sacked_out -= acked_pcount;
3109 } else if (tcp_is_sack(tp)) {
3110 tp->delivered += acked_pcount;
3111 if (!tcp_skb_spurious_retrans(tp, skb))
3112 tcp_rack_advance(tp, sacked, scb->end_seq,
3115 if (sacked & TCPCB_LOST)
3116 tp->lost_out -= acked_pcount;
3118 tp->packets_out -= acked_pcount;
3119 pkts_acked += acked_pcount;
3120 tcp_rate_skb_delivered(sk, skb, sack->rate);
3122 /* Initial outgoing SYN's get put onto the write_queue
3123 * just like anything else we transmit. It is not
3124 * true data, and if we misinform our callers that
3125 * this ACK acks real data, we will erroneously exit
3126 * connection startup slow start one packet too
3127 * quickly. This is severely frowned upon behavior.
3129 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
3130 flag |= FLAG_DATA_ACKED;
3132 flag |= FLAG_SYN_ACKED;
3133 tp->retrans_stamp = 0;
3139 next = skb_rb_next(skb);
3140 if (unlikely(skb == tp->retransmit_skb_hint))
3141 tp->retransmit_skb_hint = NULL;
3142 if (unlikely(skb == tp->lost_skb_hint))
3143 tp->lost_skb_hint = NULL;
3144 tcp_rtx_queue_unlink_and_free(skb, sk);
3148 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
3150 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3151 tp->snd_up = tp->snd_una;
3153 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3154 flag |= FLAG_SACK_RENEGING;
3156 if (likely(first_ackt) && !(flag & FLAG_RETRANS_DATA_ACKED)) {
3157 seq_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, first_ackt);
3158 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, last_ackt);
3160 if (pkts_acked == 1 && last_in_flight < tp->mss_cache &&
3161 last_in_flight && !prior_sacked && fully_acked &&
3162 sack->rate->prior_delivered + 1 == tp->delivered &&
3163 !(flag & (FLAG_CA_ALERT | FLAG_SYN_ACKED))) {
3164 /* Conservatively mark a delayed ACK. It's typically
3165 * from a lone runt packet over the round trip to
3166 * a receiver w/o out-of-order or CE events.
3168 flag |= FLAG_ACK_MAYBE_DELAYED;
3171 if (sack->first_sackt) {
3172 sack_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->first_sackt);
3173 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->last_sackt);
3175 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us,
3176 ca_rtt_us, sack->rate);
3178 if (flag & FLAG_ACKED) {
3179 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3180 if (unlikely(icsk->icsk_mtup.probe_size &&
3181 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3182 tcp_mtup_probe_success(sk);
3185 if (tcp_is_reno(tp)) {
3186 tcp_remove_reno_sacks(sk, pkts_acked);
3190 /* Non-retransmitted hole got filled? That's reordering */
3191 if (before(reord, prior_fack))
3192 tcp_check_sack_reordering(sk, reord, 0);
3194 delta = prior_sacked - tp->sacked_out;
3195 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3197 } else if (skb && rtt_update && sack_rtt_us >= 0 &&
3198 sack_rtt_us > tcp_stamp_us_delta(tp->tcp_mstamp, skb->skb_mstamp)) {
3199 /* Do not re-arm RTO if the sack RTT is measured from data sent
3200 * after when the head was last (re)transmitted. Otherwise the
3201 * timeout may continue to extend in loss recovery.
3203 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3206 if (icsk->icsk_ca_ops->pkts_acked) {
3207 struct ack_sample sample = { .pkts_acked = pkts_acked,
3208 .rtt_us = sack->rate->rtt_us,
3209 .in_flight = last_in_flight };
3211 icsk->icsk_ca_ops->pkts_acked(sk, &sample);
3214 #if FASTRETRANS_DEBUG > 0
3215 WARN_ON((int)tp->sacked_out < 0);
3216 WARN_ON((int)tp->lost_out < 0);
3217 WARN_ON((int)tp->retrans_out < 0);
3218 if (!tp->packets_out && tcp_is_sack(tp)) {
3219 icsk = inet_csk(sk);
3221 pr_debug("Leak l=%u %d\n",
3222 tp->lost_out, icsk->icsk_ca_state);
3225 if (tp->sacked_out) {
3226 pr_debug("Leak s=%u %d\n",
3227 tp->sacked_out, icsk->icsk_ca_state);
3230 if (tp->retrans_out) {
3231 pr_debug("Leak r=%u %d\n",
3232 tp->retrans_out, icsk->icsk_ca_state);
3233 tp->retrans_out = 0;
3240 static void tcp_ack_probe(struct sock *sk)
3242 struct inet_connection_sock *icsk = inet_csk(sk);
3243 struct sk_buff *head = tcp_send_head(sk);
3244 const struct tcp_sock *tp = tcp_sk(sk);
3246 /* Was it a usable window open? */
3249 if (!after(TCP_SKB_CB(head)->end_seq, tcp_wnd_end(tp))) {
3250 icsk->icsk_backoff = 0;
3251 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3252 /* Socket must be waked up by subsequent tcp_data_snd_check().
3253 * This function is not for random using!
3256 unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX);
3258 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3263 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3265 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3266 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3269 /* Decide wheather to run the increase function of congestion control. */
3270 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3272 /* If reordering is high then always grow cwnd whenever data is
3273 * delivered regardless of its ordering. Otherwise stay conservative
3274 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3275 * new SACK or ECE mark may first advance cwnd here and later reduce
3276 * cwnd in tcp_fastretrans_alert() based on more states.
3278 if (tcp_sk(sk)->reordering > sock_net(sk)->ipv4.sysctl_tcp_reordering)
3279 return flag & FLAG_FORWARD_PROGRESS;
3281 return flag & FLAG_DATA_ACKED;
3284 /* The "ultimate" congestion control function that aims to replace the rigid
3285 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction).
3286 * It's called toward the end of processing an ACK with precise rate
3287 * information. All transmission or retransmission are delayed afterwards.
3289 static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked,
3290 int flag, const struct rate_sample *rs)
3292 const struct inet_connection_sock *icsk = inet_csk(sk);
3294 if (icsk->icsk_ca_ops->cong_control) {
3295 icsk->icsk_ca_ops->cong_control(sk, rs);
3299 if (tcp_in_cwnd_reduction(sk)) {
3300 /* Reduce cwnd if state mandates */
3301 tcp_cwnd_reduction(sk, acked_sacked, flag);
3302 } else if (tcp_may_raise_cwnd(sk, flag)) {
3303 /* Advance cwnd if state allows */
3304 tcp_cong_avoid(sk, ack, acked_sacked);
3306 tcp_update_pacing_rate(sk);
3309 /* Check that window update is acceptable.
3310 * The function assumes that snd_una<=ack<=snd_next.
3312 static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3313 const u32 ack, const u32 ack_seq,
3316 return after(ack, tp->snd_una) ||
3317 after(ack_seq, tp->snd_wl1) ||
3318 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3321 /* If we update tp->snd_una, also update tp->bytes_acked */
3322 static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack)
3324 u32 delta = ack - tp->snd_una;
3326 sock_owned_by_me((struct sock *)tp);
3327 tp->bytes_acked += delta;
3331 /* If we update tp->rcv_nxt, also update tp->bytes_received */
3332 static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq)
3334 u32 delta = seq - tp->rcv_nxt;
3336 sock_owned_by_me((struct sock *)tp);
3337 tp->bytes_received += delta;
3341 /* Update our send window.
3343 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3344 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3346 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3349 struct tcp_sock *tp = tcp_sk(sk);
3351 u32 nwin = ntohs(tcp_hdr(skb)->window);
3353 if (likely(!tcp_hdr(skb)->syn))
3354 nwin <<= tp->rx_opt.snd_wscale;
3356 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3357 flag |= FLAG_WIN_UPDATE;
3358 tcp_update_wl(tp, ack_seq);
3360 if (tp->snd_wnd != nwin) {
3363 /* Note, it is the only place, where
3364 * fast path is recovered for sending TCP.
3367 tcp_fast_path_check(sk);
3369 if (!tcp_write_queue_empty(sk))
3370 tcp_slow_start_after_idle_check(sk);
3372 if (nwin > tp->max_window) {
3373 tp->max_window = nwin;
3374 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3379 tcp_snd_una_update(tp, ack);
3384 static bool __tcp_oow_rate_limited(struct net *net, int mib_idx,
3385 u32 *last_oow_ack_time)
3387 if (*last_oow_ack_time) {
3388 s32 elapsed = (s32)(tcp_jiffies32 - *last_oow_ack_time);
3390 if (0 <= elapsed && elapsed < net->ipv4.sysctl_tcp_invalid_ratelimit) {
3391 NET_INC_STATS(net, mib_idx);
3392 return true; /* rate-limited: don't send yet! */
3396 *last_oow_ack_time = tcp_jiffies32;
3398 return false; /* not rate-limited: go ahead, send dupack now! */
3401 /* Return true if we're currently rate-limiting out-of-window ACKs and
3402 * thus shouldn't send a dupack right now. We rate-limit dupacks in
3403 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3404 * attacks that send repeated SYNs or ACKs for the same connection. To
3405 * do this, we do not send a duplicate SYNACK or ACK if the remote
3406 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3408 bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb,
3409 int mib_idx, u32 *last_oow_ack_time)
3411 /* Data packets without SYNs are not likely part of an ACK loop. */
3412 if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) &&
3416 return __tcp_oow_rate_limited(net, mib_idx, last_oow_ack_time);
3419 /* RFC 5961 7 [ACK Throttling] */
3420 static void tcp_send_challenge_ack(struct sock *sk, const struct sk_buff *skb)
3422 /* unprotected vars, we dont care of overwrites */
3423 static u32 challenge_timestamp;
3424 static unsigned int challenge_count;
3425 struct tcp_sock *tp = tcp_sk(sk);
3426 struct net *net = sock_net(sk);
3429 /* First check our per-socket dupack rate limit. */
3430 if (__tcp_oow_rate_limited(net,
3431 LINUX_MIB_TCPACKSKIPPEDCHALLENGE,
3432 &tp->last_oow_ack_time))
3435 /* Then check host-wide RFC 5961 rate limit. */
3437 if (now != challenge_timestamp) {
3438 u32 ack_limit = net->ipv4.sysctl_tcp_challenge_ack_limit;
3439 u32 half = (ack_limit + 1) >> 1;
3441 challenge_timestamp = now;
3442 WRITE_ONCE(challenge_count, half + prandom_u32_max(ack_limit));
3444 count = READ_ONCE(challenge_count);
3446 WRITE_ONCE(challenge_count, count - 1);
3447 NET_INC_STATS(net, LINUX_MIB_TCPCHALLENGEACK);
3452 static void tcp_store_ts_recent(struct tcp_sock *tp)
3454 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3455 tp->rx_opt.ts_recent_stamp = get_seconds();
3458 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3460 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3461 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3462 * extra check below makes sure this can only happen
3463 * for pure ACK frames. -DaveM
3465 * Not only, also it occurs for expired timestamps.
3468 if (tcp_paws_check(&tp->rx_opt, 0))
3469 tcp_store_ts_recent(tp);
3473 /* This routine deals with acks during a TLP episode.
3474 * We mark the end of a TLP episode on receiving TLP dupack or when
3475 * ack is after tlp_high_seq.
3476 * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3478 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3480 struct tcp_sock *tp = tcp_sk(sk);
3482 if (before(ack, tp->tlp_high_seq))
3485 if (flag & FLAG_DSACKING_ACK) {
3486 /* This DSACK means original and TLP probe arrived; no loss */
3487 tp->tlp_high_seq = 0;
3488 } else if (after(ack, tp->tlp_high_seq)) {
3489 /* ACK advances: there was a loss, so reduce cwnd. Reset
3490 * tlp_high_seq in tcp_init_cwnd_reduction()
3492 tcp_init_cwnd_reduction(sk);
3493 tcp_set_ca_state(sk, TCP_CA_CWR);
3494 tcp_end_cwnd_reduction(sk);
3495 tcp_try_keep_open(sk);
3496 NET_INC_STATS(sock_net(sk),
3497 LINUX_MIB_TCPLOSSPROBERECOVERY);
3498 } else if (!(flag & (FLAG_SND_UNA_ADVANCED |
3499 FLAG_NOT_DUP | FLAG_DATA_SACKED))) {
3500 /* Pure dupack: original and TLP probe arrived; no loss */
3501 tp->tlp_high_seq = 0;
3505 static inline void tcp_in_ack_event(struct sock *sk, u32 flags)
3507 const struct inet_connection_sock *icsk = inet_csk(sk);
3509 if (icsk->icsk_ca_ops->in_ack_event)
3510 icsk->icsk_ca_ops->in_ack_event(sk, flags);
3513 /* Congestion control has updated the cwnd already. So if we're in
3514 * loss recovery then now we do any new sends (for FRTO) or
3515 * retransmits (for CA_Loss or CA_recovery) that make sense.
3517 static void tcp_xmit_recovery(struct sock *sk, int rexmit)
3519 struct tcp_sock *tp = tcp_sk(sk);
3521 if (rexmit == REXMIT_NONE)
3524 if (unlikely(rexmit == 2)) {
3525 __tcp_push_pending_frames(sk, tcp_current_mss(sk),
3527 if (after(tp->snd_nxt, tp->high_seq))
3531 tcp_xmit_retransmit_queue(sk);
3534 /* Returns the number of packets newly acked or sacked by the current ACK */
3535 static u32 tcp_newly_delivered(struct sock *sk, u32 prior_delivered, int flag)
3537 const struct net *net = sock_net(sk);
3538 struct tcp_sock *tp = tcp_sk(sk);
3541 delivered = tp->delivered - prior_delivered;
3542 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVERED, delivered);
3543 if (flag & FLAG_ECE) {
3544 tp->delivered_ce += delivered;
3545 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVEREDCE, delivered);
3550 /* This routine deals with incoming acks, but not outgoing ones. */
3551 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3553 struct inet_connection_sock *icsk = inet_csk(sk);
3554 struct tcp_sock *tp = tcp_sk(sk);
3555 struct tcp_sacktag_state sack_state;
3556 struct rate_sample rs = { .prior_delivered = 0 };
3557 u32 prior_snd_una = tp->snd_una;
3558 bool is_sack_reneg = tp->is_sack_reneg;
3559 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3560 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3561 bool is_dupack = false;
3562 int prior_packets = tp->packets_out;
3563 u32 delivered = tp->delivered;
3564 u32 lost = tp->lost;
3565 int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */
3568 sack_state.first_sackt = 0;
3569 sack_state.rate = &rs;
3571 /* We very likely will need to access rtx queue. */
3572 prefetch(sk->tcp_rtx_queue.rb_node);
3574 /* If the ack is older than previous acks
3575 * then we can probably ignore it.
3577 if (before(ack, prior_snd_una)) {
3578 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3579 if (before(ack, prior_snd_una - tp->max_window)) {
3580 if (!(flag & FLAG_NO_CHALLENGE_ACK))
3581 tcp_send_challenge_ack(sk, skb);
3587 /* If the ack includes data we haven't sent yet, discard
3588 * this segment (RFC793 Section 3.9).
3590 if (after(ack, tp->snd_nxt))
3593 if (after(ack, prior_snd_una)) {
3594 flag |= FLAG_SND_UNA_ADVANCED;
3595 icsk->icsk_retransmits = 0;
3597 #if IS_ENABLED(CONFIG_TLS_DEVICE)
3598 if (static_branch_unlikely(&clean_acked_data_enabled))
3599 if (icsk->icsk_clean_acked)
3600 icsk->icsk_clean_acked(sk, ack);
3604 prior_fack = tcp_is_sack(tp) ? tcp_highest_sack_seq(tp) : tp->snd_una;
3605 rs.prior_in_flight = tcp_packets_in_flight(tp);
3607 /* ts_recent update must be made after we are sure that the packet
3610 if (flag & FLAG_UPDATE_TS_RECENT)
3611 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3613 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3614 /* Window is constant, pure forward advance.
3615 * No more checks are required.
3616 * Note, we use the fact that SND.UNA>=SND.WL2.
3618 tcp_update_wl(tp, ack_seq);
3619 tcp_snd_una_update(tp, ack);
3620 flag |= FLAG_WIN_UPDATE;
3622 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
3624 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPACKS);
3626 u32 ack_ev_flags = CA_ACK_SLOWPATH;
3628 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3631 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3633 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3635 if (TCP_SKB_CB(skb)->sacked)
3636 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3639 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) {
3641 ack_ev_flags |= CA_ACK_ECE;
3644 if (flag & FLAG_WIN_UPDATE)
3645 ack_ev_flags |= CA_ACK_WIN_UPDATE;
3647 tcp_in_ack_event(sk, ack_ev_flags);
3650 /* We passed data and got it acked, remove any soft error
3651 * log. Something worked...
3653 sk->sk_err_soft = 0;
3654 icsk->icsk_probes_out = 0;
3655 tp->rcv_tstamp = tcp_jiffies32;
3659 /* See if we can take anything off of the retransmit queue. */
3660 flag |= tcp_clean_rtx_queue(sk, prior_fack, prior_snd_una, &sack_state);
3662 tcp_rack_update_reo_wnd(sk, &rs);
3664 if (tp->tlp_high_seq)
3665 tcp_process_tlp_ack(sk, ack, flag);
3666 /* If needed, reset TLP/RTO timer; RACK may later override this. */
3667 if (flag & FLAG_SET_XMIT_TIMER)
3668 tcp_set_xmit_timer(sk);
3670 if (tcp_ack_is_dubious(sk, flag)) {
3671 is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
3672 tcp_fastretrans_alert(sk, prior_snd_una, is_dupack, &flag,
3676 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
3679 delivered = tcp_newly_delivered(sk, delivered, flag);
3680 lost = tp->lost - lost; /* freshly marked lost */
3681 rs.is_ack_delayed = !!(flag & FLAG_ACK_MAYBE_DELAYED);
3682 tcp_rate_gen(sk, delivered, lost, is_sack_reneg, sack_state.rate);
3683 tcp_cong_control(sk, ack, delivered, flag, sack_state.rate);
3684 tcp_xmit_recovery(sk, rexmit);
3688 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3689 if (flag & FLAG_DSACKING_ACK) {
3690 tcp_fastretrans_alert(sk, prior_snd_una, is_dupack, &flag,
3692 tcp_newly_delivered(sk, delivered, flag);
3694 /* If this ack opens up a zero window, clear backoff. It was
3695 * being used to time the probes, and is probably far higher than
3696 * it needs to be for normal retransmission.
3700 if (tp->tlp_high_seq)
3701 tcp_process_tlp_ack(sk, ack, flag);
3705 SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3709 /* If data was SACKed, tag it and see if we should send more data.
3710 * If data was DSACKed, see if we can undo a cwnd reduction.
3712 if (TCP_SKB_CB(skb)->sacked) {
3713 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3715 tcp_fastretrans_alert(sk, prior_snd_una, is_dupack, &flag,
3717 tcp_newly_delivered(sk, delivered, flag);
3718 tcp_xmit_recovery(sk, rexmit);
3721 SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3725 static void tcp_parse_fastopen_option(int len, const unsigned char *cookie,
3726 bool syn, struct tcp_fastopen_cookie *foc,
3729 /* Valid only in SYN or SYN-ACK with an even length. */
3730 if (!foc || !syn || len < 0 || (len & 1))
3733 if (len >= TCP_FASTOPEN_COOKIE_MIN &&
3734 len <= TCP_FASTOPEN_COOKIE_MAX)
3735 memcpy(foc->val, cookie, len);
3742 static void smc_parse_options(const struct tcphdr *th,
3743 struct tcp_options_received *opt_rx,
3744 const unsigned char *ptr,
3747 #if IS_ENABLED(CONFIG_SMC)
3748 if (static_branch_unlikely(&tcp_have_smc)) {
3749 if (th->syn && !(opsize & 1) &&
3750 opsize >= TCPOLEN_EXP_SMC_BASE &&
3751 get_unaligned_be32(ptr) == TCPOPT_SMC_MAGIC)
3757 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3758 * But, this can also be called on packets in the established flow when
3759 * the fast version below fails.
3761 void tcp_parse_options(const struct net *net,
3762 const struct sk_buff *skb,
3763 struct tcp_options_received *opt_rx, int estab,
3764 struct tcp_fastopen_cookie *foc)
3766 const unsigned char *ptr;
3767 const struct tcphdr *th = tcp_hdr(skb);
3768 int length = (th->doff * 4) - sizeof(struct tcphdr);
3770 ptr = (const unsigned char *)(th + 1);
3771 opt_rx->saw_tstamp = 0;
3773 while (length > 0) {
3774 int opcode = *ptr++;
3780 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3785 if (opsize < 2) /* "silly options" */
3787 if (opsize > length)
3788 return; /* don't parse partial options */
3791 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3792 u16 in_mss = get_unaligned_be16(ptr);
3794 if (opt_rx->user_mss &&
3795 opt_rx->user_mss < in_mss)
3796 in_mss = opt_rx->user_mss;
3797 opt_rx->mss_clamp = in_mss;
3802 if (opsize == TCPOLEN_WINDOW && th->syn &&
3803 !estab && net->ipv4.sysctl_tcp_window_scaling) {
3804 __u8 snd_wscale = *(__u8 *)ptr;
3805 opt_rx->wscale_ok = 1;
3806 if (snd_wscale > TCP_MAX_WSCALE) {
3807 net_info_ratelimited("%s: Illegal window scaling value %d > %u received\n",
3811 snd_wscale = TCP_MAX_WSCALE;
3813 opt_rx->snd_wscale = snd_wscale;
3816 case TCPOPT_TIMESTAMP:
3817 if ((opsize == TCPOLEN_TIMESTAMP) &&
3818 ((estab && opt_rx->tstamp_ok) ||
3819 (!estab && net->ipv4.sysctl_tcp_timestamps))) {
3820 opt_rx->saw_tstamp = 1;
3821 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3822 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3825 case TCPOPT_SACK_PERM:
3826 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3827 !estab && net->ipv4.sysctl_tcp_sack) {
3828 opt_rx->sack_ok = TCP_SACK_SEEN;
3829 tcp_sack_reset(opt_rx);
3834 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3835 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3837 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3840 #ifdef CONFIG_TCP_MD5SIG
3843 * The MD5 Hash has already been
3844 * checked (see tcp_v{4,6}_do_rcv()).
3848 case TCPOPT_FASTOPEN:
3849 tcp_parse_fastopen_option(
3850 opsize - TCPOLEN_FASTOPEN_BASE,
3851 ptr, th->syn, foc, false);
3855 /* Fast Open option shares code 254 using a
3856 * 16 bits magic number.
3858 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE &&
3859 get_unaligned_be16(ptr) ==
3860 TCPOPT_FASTOPEN_MAGIC)
3861 tcp_parse_fastopen_option(opsize -
3862 TCPOLEN_EXP_FASTOPEN_BASE,
3863 ptr + 2, th->syn, foc, true);
3865 smc_parse_options(th, opt_rx, ptr,
3875 EXPORT_SYMBOL(tcp_parse_options);
3877 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
3879 const __be32 *ptr = (const __be32 *)(th + 1);
3881 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3882 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3883 tp->rx_opt.saw_tstamp = 1;
3885 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3888 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
3890 tp->rx_opt.rcv_tsecr = 0;
3896 /* Fast parse options. This hopes to only see timestamps.
3897 * If it is wrong it falls back on tcp_parse_options().
3899 static bool tcp_fast_parse_options(const struct net *net,
3900 const struct sk_buff *skb,
3901 const struct tcphdr *th, struct tcp_sock *tp)
3903 /* In the spirit of fast parsing, compare doff directly to constant
3904 * values. Because equality is used, short doff can be ignored here.
3906 if (th->doff == (sizeof(*th) / 4)) {
3907 tp->rx_opt.saw_tstamp = 0;
3909 } else if (tp->rx_opt.tstamp_ok &&
3910 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3911 if (tcp_parse_aligned_timestamp(tp, th))
3915 tcp_parse_options(net, skb, &tp->rx_opt, 1, NULL);
3916 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3917 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
3922 #ifdef CONFIG_TCP_MD5SIG
3924 * Parse MD5 Signature option
3926 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
3928 int length = (th->doff << 2) - sizeof(*th);
3929 const u8 *ptr = (const u8 *)(th + 1);
3931 /* If not enough data remaining, we can short cut */
3932 while (length >= TCPOLEN_MD5SIG) {
3933 int opcode = *ptr++;
3944 if (opsize < 2 || opsize > length)
3946 if (opcode == TCPOPT_MD5SIG)
3947 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
3954 EXPORT_SYMBOL(tcp_parse_md5sig_option);
3957 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3959 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3960 * it can pass through stack. So, the following predicate verifies that
3961 * this segment is not used for anything but congestion avoidance or
3962 * fast retransmit. Moreover, we even are able to eliminate most of such
3963 * second order effects, if we apply some small "replay" window (~RTO)
3964 * to timestamp space.
3966 * All these measures still do not guarantee that we reject wrapped ACKs
3967 * on networks with high bandwidth, when sequence space is recycled fastly,
3968 * but it guarantees that such events will be very rare and do not affect
3969 * connection seriously. This doesn't look nice, but alas, PAWS is really
3972 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3973 * states that events when retransmit arrives after original data are rare.
3974 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3975 * the biggest problem on large power networks even with minor reordering.
3976 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3977 * up to bandwidth of 18Gigabit/sec. 8) ]
3980 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3982 const struct tcp_sock *tp = tcp_sk(sk);
3983 const struct tcphdr *th = tcp_hdr(skb);
3984 u32 seq = TCP_SKB_CB(skb)->seq;
3985 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3987 return (/* 1. Pure ACK with correct sequence number. */
3988 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3990 /* 2. ... and duplicate ACK. */
3991 ack == tp->snd_una &&
3993 /* 3. ... and does not update window. */
3994 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3996 /* 4. ... and sits in replay window. */
3997 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
4000 static inline bool tcp_paws_discard(const struct sock *sk,
4001 const struct sk_buff *skb)
4003 const struct tcp_sock *tp = tcp_sk(sk);
4005 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
4006 !tcp_disordered_ack(sk, skb);
4009 /* Check segment sequence number for validity.
4011 * Segment controls are considered valid, if the segment
4012 * fits to the window after truncation to the window. Acceptability
4013 * of data (and SYN, FIN, of course) is checked separately.
4014 * See tcp_data_queue(), for example.
4016 * Also, controls (RST is main one) are accepted using RCV.WUP instead
4017 * of RCV.NXT. Peer still did not advance his SND.UNA when we
4018 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4019 * (borrowed from freebsd)
4022 static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
4024 return !before(end_seq, tp->rcv_wup) &&
4025 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
4028 /* When we get a reset we do this. */
4029 void tcp_reset(struct sock *sk)
4031 trace_tcp_receive_reset(sk);
4033 /* We want the right error as BSD sees it (and indeed as we do). */
4034 switch (sk->sk_state) {
4036 sk->sk_err = ECONNREFUSED;
4038 case TCP_CLOSE_WAIT:
4044 sk->sk_err = ECONNRESET;
4046 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4049 tcp_write_queue_purge(sk);
4052 if (!sock_flag(sk, SOCK_DEAD))
4053 sk->sk_error_report(sk);
4057 * Process the FIN bit. This now behaves as it is supposed to work
4058 * and the FIN takes effect when it is validly part of sequence
4059 * space. Not before when we get holes.
4061 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4062 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4065 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4066 * close and we go into CLOSING (and later onto TIME-WAIT)
4068 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4070 void tcp_fin(struct sock *sk)
4072 struct tcp_sock *tp = tcp_sk(sk);
4074 inet_csk_schedule_ack(sk);
4076 sk->sk_shutdown |= RCV_SHUTDOWN;
4077 sock_set_flag(sk, SOCK_DONE);
4079 switch (sk->sk_state) {
4081 case TCP_ESTABLISHED:
4082 /* Move to CLOSE_WAIT */
4083 tcp_set_state(sk, TCP_CLOSE_WAIT);
4084 inet_csk(sk)->icsk_ack.pingpong = 1;
4087 case TCP_CLOSE_WAIT:
4089 /* Received a retransmission of the FIN, do
4094 /* RFC793: Remain in the LAST-ACK state. */
4098 /* This case occurs when a simultaneous close
4099 * happens, we must ack the received FIN and
4100 * enter the CLOSING state.
4103 tcp_set_state(sk, TCP_CLOSING);
4106 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4108 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4111 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4112 * cases we should never reach this piece of code.
4114 pr_err("%s: Impossible, sk->sk_state=%d\n",
4115 __func__, sk->sk_state);
4119 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4120 * Probably, we should reset in this case. For now drop them.
4122 skb_rbtree_purge(&tp->out_of_order_queue);
4123 if (tcp_is_sack(tp))
4124 tcp_sack_reset(&tp->rx_opt);
4127 if (!sock_flag(sk, SOCK_DEAD)) {
4128 sk->sk_state_change(sk);
4130 /* Do not send POLL_HUP for half duplex close. */
4131 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4132 sk->sk_state == TCP_CLOSE)
4133 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4135 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4139 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4142 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4143 if (before(seq, sp->start_seq))
4144 sp->start_seq = seq;
4145 if (after(end_seq, sp->end_seq))
4146 sp->end_seq = end_seq;
4152 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4154 struct tcp_sock *tp = tcp_sk(sk);
4156 if (tcp_is_sack(tp) && sock_net(sk)->ipv4.sysctl_tcp_dsack) {
4159 if (before(seq, tp->rcv_nxt))
4160 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4162 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4164 NET_INC_STATS(sock_net(sk), mib_idx);
4166 tp->rx_opt.dsack = 1;
4167 tp->duplicate_sack[0].start_seq = seq;
4168 tp->duplicate_sack[0].end_seq = end_seq;
4172 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4174 struct tcp_sock *tp = tcp_sk(sk);
4176 if (!tp->rx_opt.dsack)
4177 tcp_dsack_set(sk, seq, end_seq);
4179 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4182 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4184 struct tcp_sock *tp = tcp_sk(sk);
4186 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4187 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4188 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4189 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
4191 if (tcp_is_sack(tp) && sock_net(sk)->ipv4.sysctl_tcp_dsack) {
4192 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4194 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4195 end_seq = tp->rcv_nxt;
4196 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4203 /* These routines update the SACK block as out-of-order packets arrive or
4204 * in-order packets close up the sequence space.
4206 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4209 struct tcp_sack_block *sp = &tp->selective_acks[0];
4210 struct tcp_sack_block *swalk = sp + 1;
4212 /* See if the recent change to the first SACK eats into
4213 * or hits the sequence space of other SACK blocks, if so coalesce.
4215 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4216 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4219 /* Zap SWALK, by moving every further SACK up by one slot.
4220 * Decrease num_sacks.
4222 tp->rx_opt.num_sacks--;
4223 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4227 this_sack++, swalk++;
4231 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4233 struct tcp_sock *tp = tcp_sk(sk);
4234 struct tcp_sack_block *sp = &tp->selective_acks[0];
4235 int cur_sacks = tp->rx_opt.num_sacks;
4241 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4242 if (tcp_sack_extend(sp, seq, end_seq)) {
4243 /* Rotate this_sack to the first one. */
4244 for (; this_sack > 0; this_sack--, sp--)
4245 swap(*sp, *(sp - 1));
4247 tcp_sack_maybe_coalesce(tp);
4252 /* Could not find an adjacent existing SACK, build a new one,
4253 * put it at the front, and shift everyone else down. We
4254 * always know there is at least one SACK present already here.
4256 * If the sack array is full, forget about the last one.
4258 if (this_sack >= TCP_NUM_SACKS) {
4259 if (tp->compressed_ack)
4262 tp->rx_opt.num_sacks--;
4265 for (; this_sack > 0; this_sack--, sp--)
4269 /* Build the new head SACK, and we're done. */
4270 sp->start_seq = seq;
4271 sp->end_seq = end_seq;
4272 tp->rx_opt.num_sacks++;
4275 /* RCV.NXT advances, some SACKs should be eaten. */
4277 static void tcp_sack_remove(struct tcp_sock *tp)
4279 struct tcp_sack_block *sp = &tp->selective_acks[0];
4280 int num_sacks = tp->rx_opt.num_sacks;
4283 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4284 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4285 tp->rx_opt.num_sacks = 0;
4289 for (this_sack = 0; this_sack < num_sacks;) {
4290 /* Check if the start of the sack is covered by RCV.NXT. */
4291 if (!before(tp->rcv_nxt, sp->start_seq)) {
4294 /* RCV.NXT must cover all the block! */
4295 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4297 /* Zap this SACK, by moving forward any other SACKS. */
4298 for (i = this_sack+1; i < num_sacks; i++)
4299 tp->selective_acks[i-1] = tp->selective_acks[i];
4306 tp->rx_opt.num_sacks = num_sacks;
4310 * tcp_try_coalesce - try to merge skb to prior one
4312 * @dest: destination queue
4314 * @from: buffer to add in queue
4315 * @fragstolen: pointer to boolean
4317 * Before queueing skb @from after @to, try to merge them
4318 * to reduce overall memory use and queue lengths, if cost is small.
4319 * Packets in ofo or receive queues can stay a long time.
4320 * Better try to coalesce them right now to avoid future collapses.
4321 * Returns true if caller should free @from instead of queueing it
4323 static bool tcp_try_coalesce(struct sock *sk,
4325 struct sk_buff *from,
4330 *fragstolen = false;
4332 /* Its possible this segment overlaps with prior segment in queue */
4333 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4336 if (!skb_try_coalesce(to, from, fragstolen, &delta))
4339 atomic_add(delta, &sk->sk_rmem_alloc);
4340 sk_mem_charge(sk, delta);
4341 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4342 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4343 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4344 TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
4346 if (TCP_SKB_CB(from)->has_rxtstamp) {
4347 TCP_SKB_CB(to)->has_rxtstamp = true;
4348 to->tstamp = from->tstamp;
4354 static void tcp_drop(struct sock *sk, struct sk_buff *skb)
4356 sk_drops_add(sk, skb);
4360 /* This one checks to see if we can put data from the
4361 * out_of_order queue into the receive_queue.
4363 static void tcp_ofo_queue(struct sock *sk)
4365 struct tcp_sock *tp = tcp_sk(sk);
4366 __u32 dsack_high = tp->rcv_nxt;
4367 bool fin, fragstolen, eaten;
4368 struct sk_buff *skb, *tail;
4371 p = rb_first(&tp->out_of_order_queue);
4374 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4377 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4378 __u32 dsack = dsack_high;
4379 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4380 dsack_high = TCP_SKB_CB(skb)->end_seq;
4381 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4384 rb_erase(&skb->rbnode, &tp->out_of_order_queue);
4386 if (unlikely(!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))) {
4387 SOCK_DEBUG(sk, "ofo packet was already received\n");
4391 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4392 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4393 TCP_SKB_CB(skb)->end_seq);
4395 tail = skb_peek_tail(&sk->sk_receive_queue);
4396 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen);
4397 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4398 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
4400 __skb_queue_tail(&sk->sk_receive_queue, skb);
4402 kfree_skb_partial(skb, fragstolen);
4404 if (unlikely(fin)) {
4406 /* tcp_fin() purges tp->out_of_order_queue,
4407 * so we must end this loop right now.
4414 static bool tcp_prune_ofo_queue(struct sock *sk);
4415 static int tcp_prune_queue(struct sock *sk);
4417 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4420 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4421 !sk_rmem_schedule(sk, skb, size)) {
4423 if (tcp_prune_queue(sk) < 0)
4426 while (!sk_rmem_schedule(sk, skb, size)) {
4427 if (!tcp_prune_ofo_queue(sk))
4434 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4436 struct tcp_sock *tp = tcp_sk(sk);
4437 struct rb_node **p, *parent;
4438 struct sk_buff *skb1;
4442 tcp_ecn_check_ce(sk, skb);
4444 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4445 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFODROP);
4450 /* Disable header prediction. */
4452 inet_csk_schedule_ack(sk);
4454 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4455 seq = TCP_SKB_CB(skb)->seq;
4456 end_seq = TCP_SKB_CB(skb)->end_seq;
4457 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4458 tp->rcv_nxt, seq, end_seq);
4460 p = &tp->out_of_order_queue.rb_node;
4461 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4462 /* Initial out of order segment, build 1 SACK. */
4463 if (tcp_is_sack(tp)) {
4464 tp->rx_opt.num_sacks = 1;
4465 tp->selective_acks[0].start_seq = seq;
4466 tp->selective_acks[0].end_seq = end_seq;
4468 rb_link_node(&skb->rbnode, NULL, p);
4469 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4470 tp->ooo_last_skb = skb;
4474 /* In the typical case, we are adding an skb to the end of the list.
4475 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
4477 if (tcp_try_coalesce(sk, tp->ooo_last_skb,
4478 skb, &fragstolen)) {
4480 tcp_grow_window(sk, skb);
4481 kfree_skb_partial(skb, fragstolen);
4485 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
4486 if (!before(seq, TCP_SKB_CB(tp->ooo_last_skb)->end_seq)) {
4487 parent = &tp->ooo_last_skb->rbnode;
4488 p = &parent->rb_right;
4492 /* Find place to insert this segment. Handle overlaps on the way. */
4496 skb1 = rb_to_skb(parent);
4497 if (before(seq, TCP_SKB_CB(skb1)->seq)) {
4498 p = &parent->rb_left;
4501 if (before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4502 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4503 /* All the bits are present. Drop. */
4504 NET_INC_STATS(sock_net(sk),
4505 LINUX_MIB_TCPOFOMERGE);
4508 tcp_dsack_set(sk, seq, end_seq);
4511 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4512 /* Partial overlap. */
4513 tcp_dsack_set(sk, seq, TCP_SKB_CB(skb1)->end_seq);
4515 /* skb's seq == skb1's seq and skb covers skb1.
4516 * Replace skb1 with skb.
4518 rb_replace_node(&skb1->rbnode, &skb->rbnode,
4519 &tp->out_of_order_queue);
4520 tcp_dsack_extend(sk,
4521 TCP_SKB_CB(skb1)->seq,
4522 TCP_SKB_CB(skb1)->end_seq);
4523 NET_INC_STATS(sock_net(sk),
4524 LINUX_MIB_TCPOFOMERGE);
4528 } else if (tcp_try_coalesce(sk, skb1,
4529 skb, &fragstolen)) {
4532 p = &parent->rb_right;
4535 /* Insert segment into RB tree. */
4536 rb_link_node(&skb->rbnode, parent, p);
4537 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4540 /* Remove other segments covered by skb. */
4541 while ((skb1 = skb_rb_next(skb)) != NULL) {
4542 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4544 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4545 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4549 rb_erase(&skb1->rbnode, &tp->out_of_order_queue);
4550 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4551 TCP_SKB_CB(skb1)->end_seq);
4552 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4555 /* If there is no skb after us, we are the last_skb ! */
4557 tp->ooo_last_skb = skb;
4560 if (tcp_is_sack(tp))
4561 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4564 tcp_grow_window(sk, skb);
4566 skb_set_owner_r(skb, sk);
4570 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, int hdrlen,
4574 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4576 __skb_pull(skb, hdrlen);
4578 tcp_try_coalesce(sk, tail,
4579 skb, fragstolen)) ? 1 : 0;
4580 tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq);
4582 __skb_queue_tail(&sk->sk_receive_queue, skb);
4583 skb_set_owner_r(skb, sk);
4588 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4590 struct sk_buff *skb;
4598 if (size > PAGE_SIZE) {
4599 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS);
4601 data_len = npages << PAGE_SHIFT;
4602 size = data_len + (size & ~PAGE_MASK);
4604 skb = alloc_skb_with_frags(size - data_len, data_len,
4605 PAGE_ALLOC_COSTLY_ORDER,
4606 &err, sk->sk_allocation);
4610 skb_put(skb, size - data_len);
4611 skb->data_len = data_len;
4614 if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) {
4615 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP);
4619 err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size);
4623 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4624 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4625 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4627 if (tcp_queue_rcv(sk, skb, 0, &fragstolen)) {
4628 WARN_ON_ONCE(fragstolen); /* should not happen */
4640 void tcp_data_ready(struct sock *sk)
4642 const struct tcp_sock *tp = tcp_sk(sk);
4643 int avail = tp->rcv_nxt - tp->copied_seq;
4645 if (avail < sk->sk_rcvlowat && !sock_flag(sk, SOCK_DONE))
4648 sk->sk_data_ready(sk);
4651 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4653 struct tcp_sock *tp = tcp_sk(sk);
4657 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
4662 __skb_pull(skb, tcp_hdr(skb)->doff * 4);
4664 tcp_ecn_accept_cwr(tp, skb);
4666 tp->rx_opt.dsack = 0;
4668 /* Queue data for delivery to the user.
4669 * Packets in sequence go to the receive queue.
4670 * Out of sequence packets to the out_of_order_queue.
4672 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4673 if (tcp_receive_window(tp) == 0) {
4674 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP);
4678 /* Ok. In sequence. In window. */
4680 if (skb_queue_len(&sk->sk_receive_queue) == 0)
4681 sk_forced_mem_schedule(sk, skb->truesize);
4682 else if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) {
4683 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP);
4687 eaten = tcp_queue_rcv(sk, skb, 0, &fragstolen);
4688 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4690 tcp_event_data_recv(sk, skb);
4691 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
4694 if (!RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4697 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4698 * gap in queue is filled.
4700 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
4701 inet_csk(sk)->icsk_ack.pingpong = 0;
4704 if (tp->rx_opt.num_sacks)
4705 tcp_sack_remove(tp);
4707 tcp_fast_path_check(sk);
4710 kfree_skb_partial(skb, fragstolen);
4711 if (!sock_flag(sk, SOCK_DEAD))
4716 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4717 /* A retransmit, 2nd most common case. Force an immediate ack. */
4718 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4719 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4722 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
4723 inet_csk_schedule_ack(sk);
4729 /* Out of window. F.e. zero window probe. */
4730 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4733 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4734 /* Partial packet, seq < rcv_next < end_seq */
4735 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4736 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4737 TCP_SKB_CB(skb)->end_seq);
4739 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4741 /* If window is closed, drop tail of packet. But after
4742 * remembering D-SACK for its head made in previous line.
4744 if (!tcp_receive_window(tp)) {
4745 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP);
4751 tcp_data_queue_ofo(sk, skb);
4754 static struct sk_buff *tcp_skb_next(struct sk_buff *skb, struct sk_buff_head *list)
4757 return !skb_queue_is_last(list, skb) ? skb->next : NULL;
4759 return skb_rb_next(skb);
4762 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4763 struct sk_buff_head *list,
4764 struct rb_root *root)
4766 struct sk_buff *next = tcp_skb_next(skb, list);
4769 __skb_unlink(skb, list);
4771 rb_erase(&skb->rbnode, root);
4774 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4779 /* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */
4780 void tcp_rbtree_insert(struct rb_root *root, struct sk_buff *skb)
4782 struct rb_node **p = &root->rb_node;
4783 struct rb_node *parent = NULL;
4784 struct sk_buff *skb1;
4788 skb1 = rb_to_skb(parent);
4789 if (before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb1)->seq))
4790 p = &parent->rb_left;
4792 p = &parent->rb_right;
4794 rb_link_node(&skb->rbnode, parent, p);
4795 rb_insert_color(&skb->rbnode, root);
4798 /* Collapse contiguous sequence of skbs head..tail with
4799 * sequence numbers start..end.
4801 * If tail is NULL, this means until the end of the queue.
4803 * Segments with FIN/SYN are not collapsed (only because this
4807 tcp_collapse(struct sock *sk, struct sk_buff_head *list, struct rb_root *root,
4808 struct sk_buff *head, struct sk_buff *tail, u32 start, u32 end)
4810 struct sk_buff *skb = head, *n;
4811 struct sk_buff_head tmp;
4814 /* First, check that queue is collapsible and find
4815 * the point where collapsing can be useful.
4818 for (end_of_skbs = true; skb != NULL && skb != tail; skb = n) {
4819 n = tcp_skb_next(skb, list);
4821 /* No new bits? It is possible on ofo queue. */
4822 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4823 skb = tcp_collapse_one(sk, skb, list, root);
4829 /* The first skb to collapse is:
4831 * - bloated or contains data before "start" or
4832 * overlaps to the next one.
4834 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
4835 (tcp_win_from_space(sk, skb->truesize) > skb->len ||
4836 before(TCP_SKB_CB(skb)->seq, start))) {
4837 end_of_skbs = false;
4841 if (n && n != tail &&
4842 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(n)->seq) {
4843 end_of_skbs = false;
4847 /* Decided to skip this, advance start seq. */
4848 start = TCP_SKB_CB(skb)->end_seq;
4851 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4854 __skb_queue_head_init(&tmp);
4856 while (before(start, end)) {
4857 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
4858 struct sk_buff *nskb;
4860 nskb = alloc_skb(copy, GFP_ATOMIC);
4864 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4865 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4867 __skb_queue_before(list, skb, nskb);
4869 __skb_queue_tail(&tmp, nskb); /* defer rbtree insertion */
4870 skb_set_owner_r(nskb, sk);
4872 /* Copy data, releasing collapsed skbs. */
4874 int offset = start - TCP_SKB_CB(skb)->seq;
4875 int size = TCP_SKB_CB(skb)->end_seq - start;
4879 size = min(copy, size);
4880 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4882 TCP_SKB_CB(nskb)->end_seq += size;
4886 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4887 skb = tcp_collapse_one(sk, skb, list, root);
4890 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4896 skb_queue_walk_safe(&tmp, skb, n)
4897 tcp_rbtree_insert(root, skb);
4900 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4901 * and tcp_collapse() them until all the queue is collapsed.
4903 static void tcp_collapse_ofo_queue(struct sock *sk)
4905 struct tcp_sock *tp = tcp_sk(sk);
4906 struct sk_buff *skb, *head;
4909 skb = skb_rb_first(&tp->out_of_order_queue);
4912 tp->ooo_last_skb = skb_rb_last(&tp->out_of_order_queue);
4915 start = TCP_SKB_CB(skb)->seq;
4916 end = TCP_SKB_CB(skb)->end_seq;
4918 for (head = skb;;) {
4919 skb = skb_rb_next(skb);
4921 /* Range is terminated when we see a gap or when
4922 * we are at the queue end.
4925 after(TCP_SKB_CB(skb)->seq, end) ||
4926 before(TCP_SKB_CB(skb)->end_seq, start)) {
4927 tcp_collapse(sk, NULL, &tp->out_of_order_queue,
4928 head, skb, start, end);
4932 if (unlikely(before(TCP_SKB_CB(skb)->seq, start)))
4933 start = TCP_SKB_CB(skb)->seq;
4934 if (after(TCP_SKB_CB(skb)->end_seq, end))
4935 end = TCP_SKB_CB(skb)->end_seq;
4940 * Clean the out-of-order queue to make room.
4941 * We drop high sequences packets to :
4942 * 1) Let a chance for holes to be filled.
4943 * 2) not add too big latencies if thousands of packets sit there.
4944 * (But if application shrinks SO_RCVBUF, we could still end up
4945 * freeing whole queue here)
4947 * Return true if queue has shrunk.
4949 static bool tcp_prune_ofo_queue(struct sock *sk)
4951 struct tcp_sock *tp = tcp_sk(sk);
4952 struct rb_node *node, *prev;
4954 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
4957 NET_INC_STATS(sock_net(sk), LINUX_MIB_OFOPRUNED);
4958 node = &tp->ooo_last_skb->rbnode;
4960 prev = rb_prev(node);
4961 rb_erase(node, &tp->out_of_order_queue);
4962 tcp_drop(sk, rb_to_skb(node));
4964 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf &&
4965 !tcp_under_memory_pressure(sk))
4969 tp->ooo_last_skb = rb_to_skb(prev);
4971 /* Reset SACK state. A conforming SACK implementation will
4972 * do the same at a timeout based retransmit. When a connection
4973 * is in a sad state like this, we care only about integrity
4974 * of the connection not performance.
4976 if (tp->rx_opt.sack_ok)
4977 tcp_sack_reset(&tp->rx_opt);
4981 /* Reduce allocated memory if we can, trying to get
4982 * the socket within its memory limits again.
4984 * Return less than zero if we should start dropping frames
4985 * until the socket owning process reads some of the data
4986 * to stabilize the situation.
4988 static int tcp_prune_queue(struct sock *sk)
4990 struct tcp_sock *tp = tcp_sk(sk);
4992 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4994 NET_INC_STATS(sock_net(sk), LINUX_MIB_PRUNECALLED);
4996 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4997 tcp_clamp_window(sk);
4998 else if (tcp_under_memory_pressure(sk))
4999 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
5001 tcp_collapse_ofo_queue(sk);
5002 if (!skb_queue_empty(&sk->sk_receive_queue))
5003 tcp_collapse(sk, &sk->sk_receive_queue, NULL,
5004 skb_peek(&sk->sk_receive_queue),
5006 tp->copied_seq, tp->rcv_nxt);
5009 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5012 /* Collapsing did not help, destructive actions follow.
5013 * This must not ever occur. */
5015 tcp_prune_ofo_queue(sk);
5017 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5020 /* If we are really being abused, tell the caller to silently
5021 * drop receive data on the floor. It will get retransmitted
5022 * and hopefully then we'll have sufficient space.
5024 NET_INC_STATS(sock_net(sk), LINUX_MIB_RCVPRUNED);
5026 /* Massive buffer overcommit. */
5031 static bool tcp_should_expand_sndbuf(const struct sock *sk)
5033 const struct tcp_sock *tp = tcp_sk(sk);
5035 /* If the user specified a specific send buffer setting, do
5038 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
5041 /* If we are under global TCP memory pressure, do not expand. */
5042 if (tcp_under_memory_pressure(sk))
5045 /* If we are under soft global TCP memory pressure, do not expand. */
5046 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
5049 /* If we filled the congestion window, do not expand. */
5050 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
5056 /* When incoming ACK allowed to free some skb from write_queue,
5057 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
5058 * on the exit from tcp input handler.
5060 * PROBLEM: sndbuf expansion does not work well with largesend.
5062 static void tcp_new_space(struct sock *sk)
5064 struct tcp_sock *tp = tcp_sk(sk);
5066 if (tcp_should_expand_sndbuf(sk)) {
5067 tcp_sndbuf_expand(sk);
5068 tp->snd_cwnd_stamp = tcp_jiffies32;
5071 sk->sk_write_space(sk);
5074 static void tcp_check_space(struct sock *sk)
5076 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
5077 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
5078 /* pairs with tcp_poll() */
5080 if (sk->sk_socket &&
5081 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
5083 if (!test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
5084 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
5089 static inline void tcp_data_snd_check(struct sock *sk)
5091 tcp_push_pending_frames(sk);
5092 tcp_check_space(sk);
5096 * Check if sending an ack is needed.
5098 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
5100 struct tcp_sock *tp = tcp_sk(sk);
5101 unsigned long rtt, delay;
5103 /* More than one full frame received... */
5104 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
5105 /* ... and right edge of window advances far enough.
5106 * (tcp_recvmsg() will send ACK otherwise).
5107 * If application uses SO_RCVLOWAT, we want send ack now if
5108 * we have not received enough bytes to satisfy the condition.
5110 (tp->rcv_nxt - tp->copied_seq < sk->sk_rcvlowat ||
5111 __tcp_select_window(sk) >= tp->rcv_wnd)) ||
5112 /* We ACK each frame or... */
5113 tcp_in_quickack_mode(sk)) {
5119 if (!ofo_possible || RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
5120 tcp_send_delayed_ack(sk);
5124 if (!tcp_is_sack(tp) ||
5125 tp->compressed_ack >= sock_net(sk)->ipv4.sysctl_tcp_comp_sack_nr)
5127 tp->compressed_ack++;
5129 if (hrtimer_is_queued(&tp->compressed_ack_timer))
5132 /* compress ack timer : 5 % of rtt, but no more than tcp_comp_sack_delay_ns */
5134 rtt = tp->rcv_rtt_est.rtt_us;
5135 if (tp->srtt_us && tp->srtt_us < rtt)
5138 delay = min_t(unsigned long, sock_net(sk)->ipv4.sysctl_tcp_comp_sack_delay_ns,
5139 rtt * (NSEC_PER_USEC >> 3)/20);
5141 hrtimer_start(&tp->compressed_ack_timer, ns_to_ktime(delay),
5142 HRTIMER_MODE_REL_PINNED_SOFT);
5145 static inline void tcp_ack_snd_check(struct sock *sk)
5147 if (!inet_csk_ack_scheduled(sk)) {
5148 /* We sent a data segment already. */
5151 __tcp_ack_snd_check(sk, 1);
5155 * This routine is only called when we have urgent data
5156 * signaled. Its the 'slow' part of tcp_urg. It could be
5157 * moved inline now as tcp_urg is only called from one
5158 * place. We handle URGent data wrong. We have to - as
5159 * BSD still doesn't use the correction from RFC961.
5160 * For 1003.1g we should support a new option TCP_STDURG to permit
5161 * either form (or just set the sysctl tcp_stdurg).
5164 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
5166 struct tcp_sock *tp = tcp_sk(sk);
5167 u32 ptr = ntohs(th->urg_ptr);
5169 if (ptr && !sock_net(sk)->ipv4.sysctl_tcp_stdurg)
5171 ptr += ntohl(th->seq);
5173 /* Ignore urgent data that we've already seen and read. */
5174 if (after(tp->copied_seq, ptr))
5177 /* Do not replay urg ptr.
5179 * NOTE: interesting situation not covered by specs.
5180 * Misbehaving sender may send urg ptr, pointing to segment,
5181 * which we already have in ofo queue. We are not able to fetch
5182 * such data and will stay in TCP_URG_NOTYET until will be eaten
5183 * by recvmsg(). Seems, we are not obliged to handle such wicked
5184 * situations. But it is worth to think about possibility of some
5185 * DoSes using some hypothetical application level deadlock.
5187 if (before(ptr, tp->rcv_nxt))
5190 /* Do we already have a newer (or duplicate) urgent pointer? */
5191 if (tp->urg_data && !after(ptr, tp->urg_seq))
5194 /* Tell the world about our new urgent pointer. */
5197 /* We may be adding urgent data when the last byte read was
5198 * urgent. To do this requires some care. We cannot just ignore
5199 * tp->copied_seq since we would read the last urgent byte again
5200 * as data, nor can we alter copied_seq until this data arrives
5201 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5203 * NOTE. Double Dutch. Rendering to plain English: author of comment
5204 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5205 * and expect that both A and B disappear from stream. This is _wrong_.
5206 * Though this happens in BSD with high probability, this is occasional.
5207 * Any application relying on this is buggy. Note also, that fix "works"
5208 * only in this artificial test. Insert some normal data between A and B and we will
5209 * decline of BSD again. Verdict: it is better to remove to trap
5212 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5213 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5214 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5216 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5217 __skb_unlink(skb, &sk->sk_receive_queue);
5222 tp->urg_data = TCP_URG_NOTYET;
5225 /* Disable header prediction. */
5229 /* This is the 'fast' part of urgent handling. */
5230 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
5232 struct tcp_sock *tp = tcp_sk(sk);
5234 /* Check if we get a new urgent pointer - normally not. */
5236 tcp_check_urg(sk, th);
5238 /* Do we wait for any urgent data? - normally not... */
5239 if (tp->urg_data == TCP_URG_NOTYET) {
5240 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5243 /* Is the urgent pointer pointing into this packet? */
5244 if (ptr < skb->len) {
5246 if (skb_copy_bits(skb, ptr, &tmp, 1))
5248 tp->urg_data = TCP_URG_VALID | tmp;
5249 if (!sock_flag(sk, SOCK_DEAD))
5250 sk->sk_data_ready(sk);
5255 /* Accept RST for rcv_nxt - 1 after a FIN.
5256 * When tcp connections are abruptly terminated from Mac OSX (via ^C), a
5257 * FIN is sent followed by a RST packet. The RST is sent with the same
5258 * sequence number as the FIN, and thus according to RFC 5961 a challenge
5259 * ACK should be sent. However, Mac OSX rate limits replies to challenge
5260 * ACKs on the closed socket. In addition middleboxes can drop either the
5261 * challenge ACK or a subsequent RST.
5263 static bool tcp_reset_check(const struct sock *sk, const struct sk_buff *skb)
5265 struct tcp_sock *tp = tcp_sk(sk);
5267 return unlikely(TCP_SKB_CB(skb)->seq == (tp->rcv_nxt - 1) &&
5268 (1 << sk->sk_state) & (TCPF_CLOSE_WAIT | TCPF_LAST_ACK |
5272 /* Does PAWS and seqno based validation of an incoming segment, flags will
5273 * play significant role here.
5275 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5276 const struct tcphdr *th, int syn_inerr)
5278 struct tcp_sock *tp = tcp_sk(sk);
5279 bool rst_seq_match = false;
5281 /* RFC1323: H1. Apply PAWS check first. */
5282 if (tcp_fast_parse_options(sock_net(sk), skb, th, tp) &&
5283 tp->rx_opt.saw_tstamp &&
5284 tcp_paws_discard(sk, skb)) {
5286 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5287 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5288 LINUX_MIB_TCPACKSKIPPEDPAWS,
5289 &tp->last_oow_ack_time))
5290 tcp_send_dupack(sk, skb);
5293 /* Reset is accepted even if it did not pass PAWS. */
5296 /* Step 1: check sequence number */
5297 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5298 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5299 * (RST) segments are validated by checking their SEQ-fields."
5300 * And page 69: "If an incoming segment is not acceptable,
5301 * an acknowledgment should be sent in reply (unless the RST
5302 * bit is set, if so drop the segment and return)".
5307 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5308 LINUX_MIB_TCPACKSKIPPEDSEQ,
5309 &tp->last_oow_ack_time))
5310 tcp_send_dupack(sk, skb);
5311 } else if (tcp_reset_check(sk, skb)) {
5317 /* Step 2: check RST bit */
5319 /* RFC 5961 3.2 (extend to match against (RCV.NXT - 1) after a
5320 * FIN and SACK too if available):
5321 * If seq num matches RCV.NXT or (RCV.NXT - 1) after a FIN, or
5322 * the right-most SACK block,
5324 * RESET the connection
5326 * Send a challenge ACK
5328 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt ||
5329 tcp_reset_check(sk, skb)) {
5330 rst_seq_match = true;
5331 } else if (tcp_is_sack(tp) && tp->rx_opt.num_sacks > 0) {
5332 struct tcp_sack_block *sp = &tp->selective_acks[0];
5333 int max_sack = sp[0].end_seq;
5336 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;
5338 max_sack = after(sp[this_sack].end_seq,
5340 sp[this_sack].end_seq : max_sack;
5343 if (TCP_SKB_CB(skb)->seq == max_sack)
5344 rst_seq_match = true;
5350 /* Disable TFO if RST is out-of-order
5351 * and no data has been received
5352 * for current active TFO socket
5354 if (tp->syn_fastopen && !tp->data_segs_in &&
5355 sk->sk_state == TCP_ESTABLISHED)
5356 tcp_fastopen_active_disable(sk);
5357 tcp_send_challenge_ack(sk, skb);
5362 /* step 3: check security and precedence [ignored] */
5364 /* step 4: Check for a SYN
5365 * RFC 5961 4.2 : Send a challenge ack
5370 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5371 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5372 tcp_send_challenge_ack(sk, skb);
5384 * TCP receive function for the ESTABLISHED state.
5386 * It is split into a fast path and a slow path. The fast path is
5388 * - A zero window was announced from us - zero window probing
5389 * is only handled properly in the slow path.
5390 * - Out of order segments arrived.
5391 * - Urgent data is expected.
5392 * - There is no buffer space left
5393 * - Unexpected TCP flags/window values/header lengths are received
5394 * (detected by checking the TCP header against pred_flags)
5395 * - Data is sent in both directions. Fast path only supports pure senders
5396 * or pure receivers (this means either the sequence number or the ack
5397 * value must stay constant)
5398 * - Unexpected TCP option.
5400 * When these conditions are not satisfied it drops into a standard
5401 * receive procedure patterned after RFC793 to handle all cases.
5402 * The first three cases are guaranteed by proper pred_flags setting,
5403 * the rest is checked inline. Fast processing is turned on in
5404 * tcp_data_queue when everything is OK.
5406 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb)
5408 const struct tcphdr *th = (const struct tcphdr *)skb->data;
5409 struct tcp_sock *tp = tcp_sk(sk);
5410 unsigned int len = skb->len;
5412 /* TCP congestion window tracking */
5413 trace_tcp_probe(sk, skb);
5415 tcp_mstamp_refresh(tp);
5416 if (unlikely(!sk->sk_rx_dst))
5417 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
5419 * Header prediction.
5420 * The code loosely follows the one in the famous
5421 * "30 instruction TCP receive" Van Jacobson mail.
5423 * Van's trick is to deposit buffers into socket queue
5424 * on a device interrupt, to call tcp_recv function
5425 * on the receive process context and checksum and copy
5426 * the buffer to user space. smart...
5428 * Our current scheme is not silly either but we take the
5429 * extra cost of the net_bh soft interrupt processing...
5430 * We do checksum and copy also but from device to kernel.
5433 tp->rx_opt.saw_tstamp = 0;
5435 /* pred_flags is 0xS?10 << 16 + snd_wnd
5436 * if header_prediction is to be made
5437 * 'S' will always be tp->tcp_header_len >> 2
5438 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5439 * turn it off (when there are holes in the receive
5440 * space for instance)
5441 * PSH flag is ignored.
5444 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5445 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5446 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5447 int tcp_header_len = tp->tcp_header_len;
5449 /* Timestamp header prediction: tcp_header_len
5450 * is automatically equal to th->doff*4 due to pred_flags
5454 /* Check timestamp */
5455 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5456 /* No? Slow path! */
5457 if (!tcp_parse_aligned_timestamp(tp, th))
5460 /* If PAWS failed, check it more carefully in slow path */
5461 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5464 /* DO NOT update ts_recent here, if checksum fails
5465 * and timestamp was corrupted part, it will result
5466 * in a hung connection since we will drop all
5467 * future packets due to the PAWS test.
5471 if (len <= tcp_header_len) {
5472 /* Bulk data transfer: sender */
5473 if (len == tcp_header_len) {
5474 /* Predicted packet is in window by definition.
5475 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5476 * Hence, check seq<=rcv_wup reduces to:
5478 if (tcp_header_len ==
5479 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5480 tp->rcv_nxt == tp->rcv_wup)
5481 tcp_store_ts_recent(tp);
5483 /* We know that such packets are checksummed
5486 tcp_ack(sk, skb, 0);
5488 tcp_data_snd_check(sk);
5489 /* When receiving pure ack in fast path, update
5490 * last ts ecr directly instead of calling
5491 * tcp_rcv_rtt_measure_ts()
5493 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr;
5495 } else { /* Header too small */
5496 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5501 bool fragstolen = false;
5503 if (tcp_checksum_complete(skb))
5506 if ((int)skb->truesize > sk->sk_forward_alloc)
5509 /* Predicted packet is in window by definition.
5510 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5511 * Hence, check seq<=rcv_wup reduces to:
5513 if (tcp_header_len ==
5514 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5515 tp->rcv_nxt == tp->rcv_wup)
5516 tcp_store_ts_recent(tp);
5518 tcp_rcv_rtt_measure_ts(sk, skb);
5520 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPHITS);
5522 /* Bulk data transfer: receiver */
5523 eaten = tcp_queue_rcv(sk, skb, tcp_header_len,
5526 tcp_event_data_recv(sk, skb);
5528 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5529 /* Well, only one small jumplet in fast path... */
5530 tcp_ack(sk, skb, FLAG_DATA);
5531 tcp_data_snd_check(sk);
5532 if (!inet_csk_ack_scheduled(sk))
5536 __tcp_ack_snd_check(sk, 0);
5539 kfree_skb_partial(skb, fragstolen);
5546 if (len < (th->doff << 2) || tcp_checksum_complete(skb))
5549 if (!th->ack && !th->rst && !th->syn)
5553 * Standard slow path.
5556 if (!tcp_validate_incoming(sk, skb, th, 1))
5560 if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0)
5563 tcp_rcv_rtt_measure_ts(sk, skb);
5565 /* Process urgent data. */
5566 tcp_urg(sk, skb, th);
5568 /* step 7: process the segment text */
5569 tcp_data_queue(sk, skb);
5571 tcp_data_snd_check(sk);
5572 tcp_ack_snd_check(sk);
5576 TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS);
5577 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5582 EXPORT_SYMBOL(tcp_rcv_established);
5584 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5586 struct tcp_sock *tp = tcp_sk(sk);
5587 struct inet_connection_sock *icsk = inet_csk(sk);
5589 tcp_set_state(sk, TCP_ESTABLISHED);
5590 icsk->icsk_ack.lrcvtime = tcp_jiffies32;
5593 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
5594 security_inet_conn_established(sk, skb);
5597 tcp_init_transfer(sk, BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB);
5599 /* Prevent spurious tcp_cwnd_restart() on first data
5602 tp->lsndtime = tcp_jiffies32;
5604 if (sock_flag(sk, SOCK_KEEPOPEN))
5605 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5607 if (!tp->rx_opt.snd_wscale)
5608 __tcp_fast_path_on(tp, tp->snd_wnd);
5613 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
5614 struct tcp_fastopen_cookie *cookie)
5616 struct tcp_sock *tp = tcp_sk(sk);
5617 struct sk_buff *data = tp->syn_data ? tcp_rtx_queue_head(sk) : NULL;
5618 u16 mss = tp->rx_opt.mss_clamp, try_exp = 0;
5619 bool syn_drop = false;
5621 if (mss == tp->rx_opt.user_mss) {
5622 struct tcp_options_received opt;
5624 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5625 tcp_clear_options(&opt);
5626 opt.user_mss = opt.mss_clamp = 0;
5627 tcp_parse_options(sock_net(sk), synack, &opt, 0, NULL);
5628 mss = opt.mss_clamp;
5631 if (!tp->syn_fastopen) {
5632 /* Ignore an unsolicited cookie */
5634 } else if (tp->total_retrans) {
5635 /* SYN timed out and the SYN-ACK neither has a cookie nor
5636 * acknowledges data. Presumably the remote received only
5637 * the retransmitted (regular) SYNs: either the original
5638 * SYN-data or the corresponding SYN-ACK was dropped.
5640 syn_drop = (cookie->len < 0 && data);
5641 } else if (cookie->len < 0 && !tp->syn_data) {
5642 /* We requested a cookie but didn't get it. If we did not use
5643 * the (old) exp opt format then try so next time (try_exp=1).
5644 * Otherwise we go back to use the RFC7413 opt (try_exp=2).
5646 try_exp = tp->syn_fastopen_exp ? 2 : 1;
5649 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp);
5651 if (data) { /* Retransmit unacked data in SYN */
5652 skb_rbtree_walk_from(data) {
5653 if (__tcp_retransmit_skb(sk, data, 1))
5657 NET_INC_STATS(sock_net(sk),
5658 LINUX_MIB_TCPFASTOPENACTIVEFAIL);
5661 tp->syn_data_acked = tp->syn_data;
5662 if (tp->syn_data_acked) {
5663 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
5664 /* SYN-data is counted as two separate packets in tcp_ack() */
5665 if (tp->delivered > 1)
5669 tcp_fastopen_add_skb(sk, synack);
5674 static void smc_check_reset_syn(struct tcp_sock *tp)
5676 #if IS_ENABLED(CONFIG_SMC)
5677 if (static_branch_unlikely(&tcp_have_smc)) {
5678 if (tp->syn_smc && !tp->rx_opt.smc_ok)
5684 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5685 const struct tcphdr *th)
5687 struct inet_connection_sock *icsk = inet_csk(sk);
5688 struct tcp_sock *tp = tcp_sk(sk);
5689 struct tcp_fastopen_cookie foc = { .len = -1 };
5690 int saved_clamp = tp->rx_opt.mss_clamp;
5693 tcp_parse_options(sock_net(sk), skb, &tp->rx_opt, 0, &foc);
5694 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
5695 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
5699 * "If the state is SYN-SENT then
5700 * first check the ACK bit
5701 * If the ACK bit is set
5702 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5703 * a reset (unless the RST bit is set, if so drop
5704 * the segment and return)"
5706 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
5707 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt))
5708 goto reset_and_undo;
5710 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5711 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5712 tcp_time_stamp(tp))) {
5713 NET_INC_STATS(sock_net(sk),
5714 LINUX_MIB_PAWSACTIVEREJECTED);
5715 goto reset_and_undo;
5718 /* Now ACK is acceptable.
5720 * "If the RST bit is set
5721 * If the ACK was acceptable then signal the user "error:
5722 * connection reset", drop the segment, enter CLOSED state,
5723 * delete TCB, and return."
5732 * "fifth, if neither of the SYN or RST bits is set then
5733 * drop the segment and return."
5739 goto discard_and_undo;
5742 * "If the SYN bit is on ...
5743 * are acceptable then ...
5744 * (our SYN has been ACKed), change the connection
5745 * state to ESTABLISHED..."
5748 tcp_ecn_rcv_synack(tp, th);
5750 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5751 tcp_ack(sk, skb, FLAG_SLOWPATH);
5753 /* Ok.. it's good. Set up sequence numbers and
5754 * move to established.
5756 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5757 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5759 /* RFC1323: The window in SYN & SYN/ACK segments is
5762 tp->snd_wnd = ntohs(th->window);
5764 if (!tp->rx_opt.wscale_ok) {
5765 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5766 tp->window_clamp = min(tp->window_clamp, 65535U);
5769 if (tp->rx_opt.saw_tstamp) {
5770 tp->rx_opt.tstamp_ok = 1;
5771 tp->tcp_header_len =
5772 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5773 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5774 tcp_store_ts_recent(tp);
5776 tp->tcp_header_len = sizeof(struct tcphdr);
5779 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5780 tcp_initialize_rcv_mss(sk);
5782 /* Remember, tcp_poll() does not lock socket!
5783 * Change state from SYN-SENT only after copied_seq
5784 * is initialized. */
5785 tp->copied_seq = tp->rcv_nxt;
5787 smc_check_reset_syn(tp);
5791 tcp_finish_connect(sk, skb);
5793 fastopen_fail = (tp->syn_fastopen || tp->syn_data) &&
5794 tcp_rcv_fastopen_synack(sk, skb, &foc);
5796 if (!sock_flag(sk, SOCK_DEAD)) {
5797 sk->sk_state_change(sk);
5798 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5802 if (sk->sk_write_pending ||
5803 icsk->icsk_accept_queue.rskq_defer_accept ||
5804 icsk->icsk_ack.pingpong) {
5805 /* Save one ACK. Data will be ready after
5806 * several ticks, if write_pending is set.
5808 * It may be deleted, but with this feature tcpdumps
5809 * look so _wonderfully_ clever, that I was not able
5810 * to stand against the temptation 8) --ANK
5812 inet_csk_schedule_ack(sk);
5813 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
5814 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5815 TCP_DELACK_MAX, TCP_RTO_MAX);
5826 /* No ACK in the segment */
5830 * "If the RST bit is set
5832 * Otherwise (no ACK) drop the segment and return."
5835 goto discard_and_undo;
5839 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5840 tcp_paws_reject(&tp->rx_opt, 0))
5841 goto discard_and_undo;
5844 /* We see SYN without ACK. It is attempt of
5845 * simultaneous connect with crossed SYNs.
5846 * Particularly, it can be connect to self.
5848 tcp_set_state(sk, TCP_SYN_RECV);
5850 if (tp->rx_opt.saw_tstamp) {
5851 tp->rx_opt.tstamp_ok = 1;
5852 tcp_store_ts_recent(tp);
5853 tp->tcp_header_len =
5854 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5856 tp->tcp_header_len = sizeof(struct tcphdr);
5859 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5860 tp->copied_seq = tp->rcv_nxt;
5861 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5863 /* RFC1323: The window in SYN & SYN/ACK segments is
5866 tp->snd_wnd = ntohs(th->window);
5867 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5868 tp->max_window = tp->snd_wnd;
5870 tcp_ecn_rcv_syn(tp, th);
5873 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5874 tcp_initialize_rcv_mss(sk);
5876 tcp_send_synack(sk);
5878 /* Note, we could accept data and URG from this segment.
5879 * There are no obstacles to make this (except that we must
5880 * either change tcp_recvmsg() to prevent it from returning data
5881 * before 3WHS completes per RFC793, or employ TCP Fast Open).
5883 * However, if we ignore data in ACKless segments sometimes,
5884 * we have no reasons to accept it sometimes.
5885 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5886 * is not flawless. So, discard packet for sanity.
5887 * Uncomment this return to process the data.
5894 /* "fifth, if neither of the SYN or RST bits is set then
5895 * drop the segment and return."
5899 tcp_clear_options(&tp->rx_opt);
5900 tp->rx_opt.mss_clamp = saved_clamp;
5904 tcp_clear_options(&tp->rx_opt);
5905 tp->rx_opt.mss_clamp = saved_clamp;
5910 * This function implements the receiving procedure of RFC 793 for
5911 * all states except ESTABLISHED and TIME_WAIT.
5912 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5913 * address independent.
5916 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb)
5918 struct tcp_sock *tp = tcp_sk(sk);
5919 struct inet_connection_sock *icsk = inet_csk(sk);
5920 const struct tcphdr *th = tcp_hdr(skb);
5921 struct request_sock *req;
5925 switch (sk->sk_state) {
5939 /* It is possible that we process SYN packets from backlog,
5940 * so we need to make sure to disable BH right there.
5943 acceptable = icsk->icsk_af_ops->conn_request(sk, skb) >= 0;
5954 tp->rx_opt.saw_tstamp = 0;
5955 tcp_mstamp_refresh(tp);
5956 queued = tcp_rcv_synsent_state_process(sk, skb, th);
5960 /* Do step6 onward by hand. */
5961 tcp_urg(sk, skb, th);
5963 tcp_data_snd_check(sk);
5967 tcp_mstamp_refresh(tp);
5968 tp->rx_opt.saw_tstamp = 0;
5969 req = tp->fastopen_rsk;
5973 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
5974 sk->sk_state != TCP_FIN_WAIT1);
5976 if (!tcp_check_req(sk, skb, req, true, &req_stolen))
5980 if (!th->ack && !th->rst && !th->syn)
5983 if (!tcp_validate_incoming(sk, skb, th, 0))
5986 /* step 5: check the ACK field */
5987 acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
5988 FLAG_UPDATE_TS_RECENT |
5989 FLAG_NO_CHALLENGE_ACK) > 0;
5992 if (sk->sk_state == TCP_SYN_RECV)
5993 return 1; /* send one RST */
5994 tcp_send_challenge_ack(sk, skb);
5997 switch (sk->sk_state) {
5999 tp->delivered++; /* SYN-ACK delivery isn't tracked in tcp_ack */
6001 tcp_synack_rtt_meas(sk, req);
6003 /* Once we leave TCP_SYN_RECV, we no longer need req
6007 inet_csk(sk)->icsk_retransmits = 0;
6008 reqsk_fastopen_remove(sk, req, false);
6009 /* Re-arm the timer because data may have been sent out.
6010 * This is similar to the regular data transmission case
6011 * when new data has just been ack'ed.
6013 * (TFO) - we could try to be more aggressive and
6014 * retransmitting any data sooner based on when they
6019 tcp_init_transfer(sk, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB);
6020 tp->copied_seq = tp->rcv_nxt;
6023 tcp_set_state(sk, TCP_ESTABLISHED);
6024 sk->sk_state_change(sk);
6026 /* Note, that this wakeup is only for marginal crossed SYN case.
6027 * Passively open sockets are not waked up, because
6028 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
6031 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
6033 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
6034 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
6035 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
6037 if (tp->rx_opt.tstamp_ok)
6038 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
6040 if (!inet_csk(sk)->icsk_ca_ops->cong_control)
6041 tcp_update_pacing_rate(sk);
6043 /* Prevent spurious tcp_cwnd_restart() on first data packet */
6044 tp->lsndtime = tcp_jiffies32;
6046 tcp_initialize_rcv_mss(sk);
6047 tcp_fast_path_on(tp);
6050 case TCP_FIN_WAIT1: {
6053 /* If we enter the TCP_FIN_WAIT1 state and we are a
6054 * Fast Open socket and this is the first acceptable
6055 * ACK we have received, this would have acknowledged
6056 * our SYNACK so stop the SYNACK timer.
6059 /* We no longer need the request sock. */
6060 reqsk_fastopen_remove(sk, req, false);
6063 if (tp->snd_una != tp->write_seq)
6066 tcp_set_state(sk, TCP_FIN_WAIT2);
6067 sk->sk_shutdown |= SEND_SHUTDOWN;
6071 if (!sock_flag(sk, SOCK_DEAD)) {
6072 /* Wake up lingering close() */
6073 sk->sk_state_change(sk);
6077 if (tp->linger2 < 0) {
6079 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6082 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6083 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6084 /* Receive out of order FIN after close() */
6085 if (tp->syn_fastopen && th->fin)
6086 tcp_fastopen_active_disable(sk);
6088 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6092 tmo = tcp_fin_time(sk);
6093 if (tmo > TCP_TIMEWAIT_LEN) {
6094 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
6095 } else if (th->fin || sock_owned_by_user(sk)) {
6096 /* Bad case. We could lose such FIN otherwise.
6097 * It is not a big problem, but it looks confusing
6098 * and not so rare event. We still can lose it now,
6099 * if it spins in bh_lock_sock(), but it is really
6102 inet_csk_reset_keepalive_timer(sk, tmo);
6104 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
6111 if (tp->snd_una == tp->write_seq) {
6112 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
6118 if (tp->snd_una == tp->write_seq) {
6119 tcp_update_metrics(sk);
6126 /* step 6: check the URG bit */
6127 tcp_urg(sk, skb, th);
6129 /* step 7: process the segment text */
6130 switch (sk->sk_state) {
6131 case TCP_CLOSE_WAIT:
6134 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
6139 /* RFC 793 says to queue data in these states,
6140 * RFC 1122 says we MUST send a reset.
6141 * BSD 4.4 also does reset.
6143 if (sk->sk_shutdown & RCV_SHUTDOWN) {
6144 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6145 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6146 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6152 case TCP_ESTABLISHED:
6153 tcp_data_queue(sk, skb);
6158 /* tcp_data could move socket to TIME-WAIT */
6159 if (sk->sk_state != TCP_CLOSE) {
6160 tcp_data_snd_check(sk);
6161 tcp_ack_snd_check(sk);
6170 EXPORT_SYMBOL(tcp_rcv_state_process);
6172 static inline void pr_drop_req(struct request_sock *req, __u16 port, int family)
6174 struct inet_request_sock *ireq = inet_rsk(req);
6176 if (family == AF_INET)
6177 net_dbg_ratelimited("drop open request from %pI4/%u\n",
6178 &ireq->ir_rmt_addr, port);
6179 #if IS_ENABLED(CONFIG_IPV6)
6180 else if (family == AF_INET6)
6181 net_dbg_ratelimited("drop open request from %pI6/%u\n",
6182 &ireq->ir_v6_rmt_addr, port);
6186 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
6188 * If we receive a SYN packet with these bits set, it means a
6189 * network is playing bad games with TOS bits. In order to
6190 * avoid possible false congestion notifications, we disable
6191 * TCP ECN negotiation.
6193 * Exception: tcp_ca wants ECN. This is required for DCTCP
6194 * congestion control: Linux DCTCP asserts ECT on all packets,
6195 * including SYN, which is most optimal solution; however,
6196 * others, such as FreeBSD do not.
6198 static void tcp_ecn_create_request(struct request_sock *req,
6199 const struct sk_buff *skb,
6200 const struct sock *listen_sk,
6201 const struct dst_entry *dst)
6203 const struct tcphdr *th = tcp_hdr(skb);
6204 const struct net *net = sock_net(listen_sk);
6205 bool th_ecn = th->ece && th->cwr;
6212 ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield);
6213 ecn_ok_dst = dst_feature(dst, DST_FEATURE_ECN_MASK);
6214 ecn_ok = net->ipv4.sysctl_tcp_ecn || ecn_ok_dst;
6216 if ((!ect && ecn_ok) || tcp_ca_needs_ecn(listen_sk) ||
6217 (ecn_ok_dst & DST_FEATURE_ECN_CA) ||
6218 tcp_bpf_ca_needs_ecn((struct sock *)req))
6219 inet_rsk(req)->ecn_ok = 1;
6222 static void tcp_openreq_init(struct request_sock *req,
6223 const struct tcp_options_received *rx_opt,
6224 struct sk_buff *skb, const struct sock *sk)
6226 struct inet_request_sock *ireq = inet_rsk(req);
6228 req->rsk_rcv_wnd = 0; /* So that tcp_send_synack() knows! */
6230 tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq;
6231 tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
6232 tcp_rsk(req)->snt_synack = tcp_clock_us();
6233 tcp_rsk(req)->last_oow_ack_time = 0;
6234 req->mss = rx_opt->mss_clamp;
6235 req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0;
6236 ireq->tstamp_ok = rx_opt->tstamp_ok;
6237 ireq->sack_ok = rx_opt->sack_ok;
6238 ireq->snd_wscale = rx_opt->snd_wscale;
6239 ireq->wscale_ok = rx_opt->wscale_ok;
6242 ireq->ir_rmt_port = tcp_hdr(skb)->source;
6243 ireq->ir_num = ntohs(tcp_hdr(skb)->dest);
6244 ireq->ir_mark = inet_request_mark(sk, skb);
6245 #if IS_ENABLED(CONFIG_SMC)
6246 ireq->smc_ok = rx_opt->smc_ok;
6250 struct request_sock *inet_reqsk_alloc(const struct request_sock_ops *ops,
6251 struct sock *sk_listener,
6252 bool attach_listener)
6254 struct request_sock *req = reqsk_alloc(ops, sk_listener,
6258 struct inet_request_sock *ireq = inet_rsk(req);
6260 ireq->ireq_opt = NULL;
6261 #if IS_ENABLED(CONFIG_IPV6)
6262 ireq->pktopts = NULL;
6264 atomic64_set(&ireq->ir_cookie, 0);
6265 ireq->ireq_state = TCP_NEW_SYN_RECV;
6266 write_pnet(&ireq->ireq_net, sock_net(sk_listener));
6267 ireq->ireq_family = sk_listener->sk_family;
6272 EXPORT_SYMBOL(inet_reqsk_alloc);
6275 * Return true if a syncookie should be sent
6277 static bool tcp_syn_flood_action(const struct sock *sk,
6278 const struct sk_buff *skb,
6281 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
6282 const char *msg = "Dropping request";
6283 bool want_cookie = false;
6284 struct net *net = sock_net(sk);
6286 #ifdef CONFIG_SYN_COOKIES
6287 if (net->ipv4.sysctl_tcp_syncookies) {
6288 msg = "Sending cookies";
6290 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES);
6293 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP);
6295 if (!queue->synflood_warned &&
6296 net->ipv4.sysctl_tcp_syncookies != 2 &&
6297 xchg(&queue->synflood_warned, 1) == 0)
6298 pr_info("%s: Possible SYN flooding on port %d. %s. Check SNMP counters.\n",
6299 proto, ntohs(tcp_hdr(skb)->dest), msg);
6304 static void tcp_reqsk_record_syn(const struct sock *sk,
6305 struct request_sock *req,
6306 const struct sk_buff *skb)
6308 if (tcp_sk(sk)->save_syn) {
6309 u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb);
6312 copy = kmalloc(len + sizeof(u32), GFP_ATOMIC);
6315 memcpy(©[1], skb_network_header(skb), len);
6316 req->saved_syn = copy;
6321 int tcp_conn_request(struct request_sock_ops *rsk_ops,
6322 const struct tcp_request_sock_ops *af_ops,
6323 struct sock *sk, struct sk_buff *skb)
6325 struct tcp_fastopen_cookie foc = { .len = -1 };
6326 __u32 isn = TCP_SKB_CB(skb)->tcp_tw_isn;
6327 struct tcp_options_received tmp_opt;
6328 struct tcp_sock *tp = tcp_sk(sk);
6329 struct net *net = sock_net(sk);
6330 struct sock *fastopen_sk = NULL;
6331 struct request_sock *req;
6332 bool want_cookie = false;
6333 struct dst_entry *dst;
6336 /* TW buckets are converted to open requests without
6337 * limitations, they conserve resources and peer is
6338 * evidently real one.
6340 if ((net->ipv4.sysctl_tcp_syncookies == 2 ||
6341 inet_csk_reqsk_queue_is_full(sk)) && !isn) {
6342 want_cookie = tcp_syn_flood_action(sk, skb, rsk_ops->slab_name);
6347 if (sk_acceptq_is_full(sk)) {
6348 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
6352 req = inet_reqsk_alloc(rsk_ops, sk, !want_cookie);
6356 tcp_rsk(req)->af_specific = af_ops;
6357 tcp_rsk(req)->ts_off = 0;
6359 tcp_clear_options(&tmp_opt);
6360 tmp_opt.mss_clamp = af_ops->mss_clamp;
6361 tmp_opt.user_mss = tp->rx_opt.user_mss;
6362 tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0,
6363 want_cookie ? NULL : &foc);
6365 if (want_cookie && !tmp_opt.saw_tstamp)
6366 tcp_clear_options(&tmp_opt);
6368 if (IS_ENABLED(CONFIG_SMC) && want_cookie)
6371 tmp_opt.tstamp_ok = tmp_opt.saw_tstamp;
6372 tcp_openreq_init(req, &tmp_opt, skb, sk);
6373 inet_rsk(req)->no_srccheck = inet_sk(sk)->transparent;
6375 /* Note: tcp_v6_init_req() might override ir_iif for link locals */
6376 inet_rsk(req)->ir_iif = inet_request_bound_dev_if(sk, skb);
6378 af_ops->init_req(req, sk, skb);
6380 if (security_inet_conn_request(sk, skb, req))
6383 if (tmp_opt.tstamp_ok)
6384 tcp_rsk(req)->ts_off = af_ops->init_ts_off(net, skb);
6386 dst = af_ops->route_req(sk, &fl, req);
6390 if (!want_cookie && !isn) {
6391 /* Kill the following clause, if you dislike this way. */
6392 if (!net->ipv4.sysctl_tcp_syncookies &&
6393 (net->ipv4.sysctl_max_syn_backlog - inet_csk_reqsk_queue_len(sk) <
6394 (net->ipv4.sysctl_max_syn_backlog >> 2)) &&
6395 !tcp_peer_is_proven(req, dst)) {
6396 /* Without syncookies last quarter of
6397 * backlog is filled with destinations,
6398 * proven to be alive.
6399 * It means that we continue to communicate
6400 * to destinations, already remembered
6401 * to the moment of synflood.
6403 pr_drop_req(req, ntohs(tcp_hdr(skb)->source),
6405 goto drop_and_release;
6408 isn = af_ops->init_seq(skb);
6411 tcp_ecn_create_request(req, skb, sk, dst);
6414 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss);
6415 req->cookie_ts = tmp_opt.tstamp_ok;
6416 if (!tmp_opt.tstamp_ok)
6417 inet_rsk(req)->ecn_ok = 0;
6420 tcp_rsk(req)->snt_isn = isn;
6421 tcp_rsk(req)->txhash = net_tx_rndhash();
6422 tcp_openreq_init_rwin(req, sk, dst);
6424 tcp_reqsk_record_syn(sk, req, skb);
6425 fastopen_sk = tcp_try_fastopen(sk, skb, req, &foc, dst);
6428 af_ops->send_synack(fastopen_sk, dst, &fl, req,
6429 &foc, TCP_SYNACK_FASTOPEN);
6430 /* Add the child socket directly into the accept queue */
6431 inet_csk_reqsk_queue_add(sk, req, fastopen_sk);
6432 sk->sk_data_ready(sk);
6433 bh_unlock_sock(fastopen_sk);
6434 sock_put(fastopen_sk);
6436 tcp_rsk(req)->tfo_listener = false;
6438 inet_csk_reqsk_queue_hash_add(sk, req,
6439 tcp_timeout_init((struct sock *)req));
6440 af_ops->send_synack(sk, dst, &fl, req, &foc,
6441 !want_cookie ? TCP_SYNACK_NORMAL :
6459 EXPORT_SYMBOL(tcp_conn_request);