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, 2);
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, 2);
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 &&
586 (TCP_SKB_CB(skb)->end_seq -
587 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss)) {
588 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr;
593 delta_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
594 tcp_rcv_rtt_update(tp, delta_us, 0);
599 * This function should be called every time data is copied to user space.
600 * It calculates the appropriate TCP receive buffer space.
602 void tcp_rcv_space_adjust(struct sock *sk)
604 struct tcp_sock *tp = tcp_sk(sk);
608 trace_tcp_rcv_space_adjust(sk);
610 tcp_mstamp_refresh(tp);
611 time = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcvq_space.time);
612 if (time < (tp->rcv_rtt_est.rtt_us >> 3) || tp->rcv_rtt_est.rtt_us == 0)
615 /* Number of bytes copied to user in last RTT */
616 copied = tp->copied_seq - tp->rcvq_space.seq;
617 if (copied <= tp->rcvq_space.space)
621 * copied = bytes received in previous RTT, our base window
622 * To cope with packet losses, we need a 2x factor
623 * To cope with slow start, and sender growing its cwin by 100 %
624 * every RTT, we need a 4x factor, because the ACK we are sending
625 * now is for the next RTT, not the current one :
626 * <prev RTT . ><current RTT .. ><next RTT .... >
629 if (sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf &&
630 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
634 /* minimal window to cope with packet losses, assuming
635 * steady state. Add some cushion because of small variations.
637 rcvwin = ((u64)copied << 1) + 16 * tp->advmss;
639 /* Accommodate for sender rate increase (eg. slow start) */
640 grow = rcvwin * (copied - tp->rcvq_space.space);
641 do_div(grow, tp->rcvq_space.space);
642 rcvwin += (grow << 1);
644 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
645 while (tcp_win_from_space(sk, rcvmem) < tp->advmss)
648 do_div(rcvwin, tp->advmss);
649 rcvbuf = min_t(u64, rcvwin * rcvmem,
650 sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
651 if (rcvbuf > sk->sk_rcvbuf) {
652 sk->sk_rcvbuf = rcvbuf;
654 /* Make the window clamp follow along. */
655 tp->window_clamp = tcp_win_from_space(sk, rcvbuf);
658 tp->rcvq_space.space = copied;
661 tp->rcvq_space.seq = tp->copied_seq;
662 tp->rcvq_space.time = tp->tcp_mstamp;
665 /* There is something which you must keep in mind when you analyze the
666 * behavior of the tp->ato delayed ack timeout interval. When a
667 * connection starts up, we want to ack as quickly as possible. The
668 * problem is that "good" TCP's do slow start at the beginning of data
669 * transmission. The means that until we send the first few ACK's the
670 * sender will sit on his end and only queue most of his data, because
671 * he can only send snd_cwnd unacked packets at any given time. For
672 * each ACK we send, he increments snd_cwnd and transmits more of his
675 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
677 struct tcp_sock *tp = tcp_sk(sk);
678 struct inet_connection_sock *icsk = inet_csk(sk);
681 inet_csk_schedule_ack(sk);
683 tcp_measure_rcv_mss(sk, skb);
685 tcp_rcv_rtt_measure(tp);
689 if (!icsk->icsk_ack.ato) {
690 /* The _first_ data packet received, initialize
691 * delayed ACK engine.
693 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
694 icsk->icsk_ack.ato = TCP_ATO_MIN;
696 int m = now - icsk->icsk_ack.lrcvtime;
698 if (m <= TCP_ATO_MIN / 2) {
699 /* The fastest case is the first. */
700 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
701 } else if (m < icsk->icsk_ack.ato) {
702 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
703 if (icsk->icsk_ack.ato > icsk->icsk_rto)
704 icsk->icsk_ack.ato = icsk->icsk_rto;
705 } else if (m > icsk->icsk_rto) {
706 /* Too long gap. Apparently sender failed to
707 * restart window, so that we send ACKs quickly.
709 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS);
713 icsk->icsk_ack.lrcvtime = now;
715 tcp_ecn_check_ce(sk, skb);
718 tcp_grow_window(sk, skb);
721 /* Called to compute a smoothed rtt estimate. The data fed to this
722 * routine either comes from timestamps, or from segments that were
723 * known _not_ to have been retransmitted [see Karn/Partridge
724 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
725 * piece by Van Jacobson.
726 * NOTE: the next three routines used to be one big routine.
727 * To save cycles in the RFC 1323 implementation it was better to break
728 * it up into three procedures. -- erics
730 static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
732 struct tcp_sock *tp = tcp_sk(sk);
733 long m = mrtt_us; /* RTT */
734 u32 srtt = tp->srtt_us;
736 /* The following amusing code comes from Jacobson's
737 * article in SIGCOMM '88. Note that rtt and mdev
738 * are scaled versions of rtt and mean deviation.
739 * This is designed to be as fast as possible
740 * m stands for "measurement".
742 * On a 1990 paper the rto value is changed to:
743 * RTO = rtt + 4 * mdev
745 * Funny. This algorithm seems to be very broken.
746 * These formulae increase RTO, when it should be decreased, increase
747 * too slowly, when it should be increased quickly, decrease too quickly
748 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
749 * does not matter how to _calculate_ it. Seems, it was trap
750 * that VJ failed to avoid. 8)
753 m -= (srtt >> 3); /* m is now error in rtt est */
754 srtt += m; /* rtt = 7/8 rtt + 1/8 new */
756 m = -m; /* m is now abs(error) */
757 m -= (tp->mdev_us >> 2); /* similar update on mdev */
758 /* This is similar to one of Eifel findings.
759 * Eifel blocks mdev updates when rtt decreases.
760 * This solution is a bit different: we use finer gain
761 * for mdev in this case (alpha*beta).
762 * Like Eifel it also prevents growth of rto,
763 * but also it limits too fast rto decreases,
764 * happening in pure Eifel.
769 m -= (tp->mdev_us >> 2); /* similar update on mdev */
771 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */
772 if (tp->mdev_us > tp->mdev_max_us) {
773 tp->mdev_max_us = tp->mdev_us;
774 if (tp->mdev_max_us > tp->rttvar_us)
775 tp->rttvar_us = tp->mdev_max_us;
777 if (after(tp->snd_una, tp->rtt_seq)) {
778 if (tp->mdev_max_us < tp->rttvar_us)
779 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
780 tp->rtt_seq = tp->snd_nxt;
781 tp->mdev_max_us = tcp_rto_min_us(sk);
784 /* no previous measure. */
785 srtt = m << 3; /* take the measured time to be rtt */
786 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */
787 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
788 tp->mdev_max_us = tp->rttvar_us;
789 tp->rtt_seq = tp->snd_nxt;
791 tp->srtt_us = max(1U, srtt);
794 static void tcp_update_pacing_rate(struct sock *sk)
796 const struct tcp_sock *tp = tcp_sk(sk);
799 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
800 rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3);
802 /* current rate is (cwnd * mss) / srtt
803 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
804 * In Congestion Avoidance phase, set it to 120 % the current rate.
806 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
807 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
808 * end of slow start and should slow down.
810 if (tp->snd_cwnd < tp->snd_ssthresh / 2)
811 rate *= sock_net(sk)->ipv4.sysctl_tcp_pacing_ss_ratio;
813 rate *= sock_net(sk)->ipv4.sysctl_tcp_pacing_ca_ratio;
815 rate *= max(tp->snd_cwnd, tp->packets_out);
817 if (likely(tp->srtt_us))
818 do_div(rate, tp->srtt_us);
820 /* WRITE_ONCE() is needed because sch_fq fetches sk_pacing_rate
821 * without any lock. We want to make sure compiler wont store
822 * intermediate values in this location.
824 WRITE_ONCE(sk->sk_pacing_rate, min_t(u64, rate,
825 sk->sk_max_pacing_rate));
828 /* Calculate rto without backoff. This is the second half of Van Jacobson's
829 * routine referred to above.
831 static void tcp_set_rto(struct sock *sk)
833 const struct tcp_sock *tp = tcp_sk(sk);
834 /* Old crap is replaced with new one. 8)
837 * 1. If rtt variance happened to be less 50msec, it is hallucination.
838 * It cannot be less due to utterly erratic ACK generation made
839 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
840 * to do with delayed acks, because at cwnd>2 true delack timeout
841 * is invisible. Actually, Linux-2.4 also generates erratic
842 * ACKs in some circumstances.
844 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
846 /* 2. Fixups made earlier cannot be right.
847 * If we do not estimate RTO correctly without them,
848 * all the algo is pure shit and should be replaced
849 * with correct one. It is exactly, which we pretend to do.
852 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
853 * guarantees that rto is higher.
858 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
860 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
863 cwnd = TCP_INIT_CWND;
864 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
867 /* Take a notice that peer is sending D-SACKs */
868 static void tcp_dsack_seen(struct tcp_sock *tp)
870 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
871 tp->rack.dsack_seen = 1;
874 /* It's reordering when higher sequence was delivered (i.e. sacked) before
875 * some lower never-retransmitted sequence ("low_seq"). The maximum reordering
876 * distance is approximated in full-mss packet distance ("reordering").
878 static void tcp_check_sack_reordering(struct sock *sk, const u32 low_seq,
881 struct tcp_sock *tp = tcp_sk(sk);
882 const u32 mss = tp->mss_cache;
885 fack = tcp_highest_sack_seq(tp);
886 if (!before(low_seq, fack))
889 metric = fack - low_seq;
890 if ((metric > tp->reordering * mss) && mss) {
891 #if FASTRETRANS_DEBUG > 1
892 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
893 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
897 tp->undo_marker ? tp->undo_retrans : 0);
899 tp->reordering = min_t(u32, (metric + mss - 1) / mss,
900 sock_net(sk)->ipv4.sysctl_tcp_max_reordering);
904 /* This exciting event is worth to be remembered. 8) */
905 NET_INC_STATS(sock_net(sk),
906 ts ? LINUX_MIB_TCPTSREORDER : LINUX_MIB_TCPSACKREORDER);
909 /* This must be called before lost_out is incremented */
910 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
912 if (!tp->retransmit_skb_hint ||
913 before(TCP_SKB_CB(skb)->seq,
914 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
915 tp->retransmit_skb_hint = skb;
918 /* Sum the number of packets on the wire we have marked as lost.
919 * There are two cases we care about here:
920 * a) Packet hasn't been marked lost (nor retransmitted),
921 * and this is the first loss.
922 * b) Packet has been marked both lost and retransmitted,
923 * and this means we think it was lost again.
925 static void tcp_sum_lost(struct tcp_sock *tp, struct sk_buff *skb)
927 __u8 sacked = TCP_SKB_CB(skb)->sacked;
929 if (!(sacked & TCPCB_LOST) ||
930 ((sacked & TCPCB_LOST) && (sacked & TCPCB_SACKED_RETRANS)))
931 tp->lost += tcp_skb_pcount(skb);
934 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
936 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
937 tcp_verify_retransmit_hint(tp, skb);
939 tp->lost_out += tcp_skb_pcount(skb);
940 tcp_sum_lost(tp, skb);
941 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
945 void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp, struct sk_buff *skb)
947 tcp_verify_retransmit_hint(tp, skb);
949 tcp_sum_lost(tp, skb);
950 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
951 tp->lost_out += tcp_skb_pcount(skb);
952 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
956 /* This procedure tags the retransmission queue when SACKs arrive.
958 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
959 * Packets in queue with these bits set are counted in variables
960 * sacked_out, retrans_out and lost_out, correspondingly.
962 * Valid combinations are:
963 * Tag InFlight Description
964 * 0 1 - orig segment is in flight.
965 * S 0 - nothing flies, orig reached receiver.
966 * L 0 - nothing flies, orig lost by net.
967 * R 2 - both orig and retransmit are in flight.
968 * L|R 1 - orig is lost, retransmit is in flight.
969 * S|R 1 - orig reached receiver, retrans is still in flight.
970 * (L|S|R is logically valid, it could occur when L|R is sacked,
971 * but it is equivalent to plain S and code short-curcuits it to S.
972 * L|S is logically invalid, it would mean -1 packet in flight 8))
974 * These 6 states form finite state machine, controlled by the following events:
975 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
976 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
977 * 3. Loss detection event of two flavors:
978 * A. Scoreboard estimator decided the packet is lost.
979 * A'. Reno "three dupacks" marks head of queue lost.
980 * B. SACK arrives sacking SND.NXT at the moment, when the
981 * segment was retransmitted.
982 * 4. D-SACK added new rule: D-SACK changes any tag to S.
984 * It is pleasant to note, that state diagram turns out to be commutative,
985 * so that we are allowed not to be bothered by order of our actions,
986 * when multiple events arrive simultaneously. (see the function below).
988 * Reordering detection.
989 * --------------------
990 * Reordering metric is maximal distance, which a packet can be displaced
991 * in packet stream. With SACKs we can estimate it:
993 * 1. SACK fills old hole and the corresponding segment was not
994 * ever retransmitted -> reordering. Alas, we cannot use it
995 * when segment was retransmitted.
996 * 2. The last flaw is solved with D-SACK. D-SACK arrives
997 * for retransmitted and already SACKed segment -> reordering..
998 * Both of these heuristics are not used in Loss state, when we cannot
999 * account for retransmits accurately.
1001 * SACK block validation.
1002 * ----------------------
1004 * SACK block range validation checks that the received SACK block fits to
1005 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
1006 * Note that SND.UNA is not included to the range though being valid because
1007 * it means that the receiver is rather inconsistent with itself reporting
1008 * SACK reneging when it should advance SND.UNA. Such SACK block this is
1009 * perfectly valid, however, in light of RFC2018 which explicitly states
1010 * that "SACK block MUST reflect the newest segment. Even if the newest
1011 * segment is going to be discarded ...", not that it looks very clever
1012 * in case of head skb. Due to potentional receiver driven attacks, we
1013 * choose to avoid immediate execution of a walk in write queue due to
1014 * reneging and defer head skb's loss recovery to standard loss recovery
1015 * procedure that will eventually trigger (nothing forbids us doing this).
1017 * Implements also blockage to start_seq wrap-around. Problem lies in the
1018 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1019 * there's no guarantee that it will be before snd_nxt (n). The problem
1020 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1023 * <- outs wnd -> <- wrapzone ->
1024 * u e n u_w e_w s n_w
1026 * |<------------+------+----- TCP seqno space --------------+---------->|
1027 * ...-- <2^31 ->| |<--------...
1028 * ...---- >2^31 ------>| |<--------...
1030 * Current code wouldn't be vulnerable but it's better still to discard such
1031 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1032 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1033 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1034 * equal to the ideal case (infinite seqno space without wrap caused issues).
1036 * With D-SACK the lower bound is extended to cover sequence space below
1037 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1038 * again, D-SACK block must not to go across snd_una (for the same reason as
1039 * for the normal SACK blocks, explained above). But there all simplicity
1040 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1041 * fully below undo_marker they do not affect behavior in anyway and can
1042 * therefore be safely ignored. In rare cases (which are more or less
1043 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1044 * fragmentation and packet reordering past skb's retransmission. To consider
1045 * them correctly, the acceptable range must be extended even more though
1046 * the exact amount is rather hard to quantify. However, tp->max_window can
1047 * be used as an exaggerated estimate.
1049 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
1050 u32 start_seq, u32 end_seq)
1052 /* Too far in future, or reversed (interpretation is ambiguous) */
1053 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1056 /* Nasty start_seq wrap-around check (see comments above) */
1057 if (!before(start_seq, tp->snd_nxt))
1060 /* In outstanding window? ...This is valid exit for D-SACKs too.
1061 * start_seq == snd_una is non-sensical (see comments above)
1063 if (after(start_seq, tp->snd_una))
1066 if (!is_dsack || !tp->undo_marker)
1069 /* ...Then it's D-SACK, and must reside below snd_una completely */
1070 if (after(end_seq, tp->snd_una))
1073 if (!before(start_seq, tp->undo_marker))
1077 if (!after(end_seq, tp->undo_marker))
1080 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1081 * start_seq < undo_marker and end_seq >= undo_marker.
1083 return !before(start_seq, end_seq - tp->max_window);
1086 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1087 struct tcp_sack_block_wire *sp, int num_sacks,
1090 struct tcp_sock *tp = tcp_sk(sk);
1091 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1092 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1093 bool dup_sack = false;
1095 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1098 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1099 } else if (num_sacks > 1) {
1100 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1101 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1103 if (!after(end_seq_0, end_seq_1) &&
1104 !before(start_seq_0, start_seq_1)) {
1107 NET_INC_STATS(sock_net(sk),
1108 LINUX_MIB_TCPDSACKOFORECV);
1112 /* D-SACK for already forgotten data... Do dumb counting. */
1113 if (dup_sack && tp->undo_marker && tp->undo_retrans > 0 &&
1114 !after(end_seq_0, prior_snd_una) &&
1115 after(end_seq_0, tp->undo_marker))
1121 struct tcp_sacktag_state {
1123 /* Timestamps for earliest and latest never-retransmitted segment
1124 * that was SACKed. RTO needs the earliest RTT to stay conservative,
1125 * but congestion control should still get an accurate delay signal.
1129 struct rate_sample *rate;
1131 unsigned int mss_now;
1134 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1135 * the incoming SACK may not exactly match but we can find smaller MSS
1136 * aligned portion of it that matches. Therefore we might need to fragment
1137 * which may fail and creates some hassle (caller must handle error case
1140 * FIXME: this could be merged to shift decision code
1142 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1143 u32 start_seq, u32 end_seq)
1147 unsigned int pkt_len;
1150 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1151 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1153 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1154 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1155 mss = tcp_skb_mss(skb);
1156 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1159 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1163 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1168 /* Round if necessary so that SACKs cover only full MSSes
1169 * and/or the remaining small portion (if present)
1171 if (pkt_len > mss) {
1172 unsigned int new_len = (pkt_len / mss) * mss;
1173 if (!in_sack && new_len < pkt_len)
1178 if (pkt_len >= skb->len && !in_sack)
1181 err = tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
1182 pkt_len, mss, GFP_ATOMIC);
1190 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1191 static u8 tcp_sacktag_one(struct sock *sk,
1192 struct tcp_sacktag_state *state, u8 sacked,
1193 u32 start_seq, u32 end_seq,
1194 int dup_sack, int pcount,
1197 struct tcp_sock *tp = tcp_sk(sk);
1199 /* Account D-SACK for retransmitted packet. */
1200 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1201 if (tp->undo_marker && tp->undo_retrans > 0 &&
1202 after(end_seq, tp->undo_marker))
1204 if ((sacked & TCPCB_SACKED_ACKED) &&
1205 before(start_seq, state->reord))
1206 state->reord = start_seq;
1209 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1210 if (!after(end_seq, tp->snd_una))
1213 if (!(sacked & TCPCB_SACKED_ACKED)) {
1214 tcp_rack_advance(tp, sacked, end_seq, xmit_time);
1216 if (sacked & TCPCB_SACKED_RETRANS) {
1217 /* If the segment is not tagged as lost,
1218 * we do not clear RETRANS, believing
1219 * that retransmission is still in flight.
1221 if (sacked & TCPCB_LOST) {
1222 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1223 tp->lost_out -= pcount;
1224 tp->retrans_out -= pcount;
1227 if (!(sacked & TCPCB_RETRANS)) {
1228 /* New sack for not retransmitted frame,
1229 * which was in hole. It is reordering.
1231 if (before(start_seq,
1232 tcp_highest_sack_seq(tp)) &&
1233 before(start_seq, state->reord))
1234 state->reord = start_seq;
1236 if (!after(end_seq, tp->high_seq))
1237 state->flag |= FLAG_ORIG_SACK_ACKED;
1238 if (state->first_sackt == 0)
1239 state->first_sackt = xmit_time;
1240 state->last_sackt = xmit_time;
1243 if (sacked & TCPCB_LOST) {
1244 sacked &= ~TCPCB_LOST;
1245 tp->lost_out -= pcount;
1249 sacked |= TCPCB_SACKED_ACKED;
1250 state->flag |= FLAG_DATA_SACKED;
1251 tp->sacked_out += pcount;
1252 tp->delivered += pcount; /* Out-of-order packets delivered */
1254 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1255 if (tp->lost_skb_hint &&
1256 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1257 tp->lost_cnt_hint += pcount;
1260 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1261 * frames and clear it. undo_retrans is decreased above, L|R frames
1262 * are accounted above as well.
1264 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1265 sacked &= ~TCPCB_SACKED_RETRANS;
1266 tp->retrans_out -= pcount;
1272 /* Shift newly-SACKed bytes from this skb to the immediately previous
1273 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1275 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *prev,
1276 struct sk_buff *skb,
1277 struct tcp_sacktag_state *state,
1278 unsigned int pcount, int shifted, int mss,
1281 struct tcp_sock *tp = tcp_sk(sk);
1282 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1283 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
1287 /* Adjust counters and hints for the newly sacked sequence
1288 * range but discard the return value since prev is already
1289 * marked. We must tag the range first because the seq
1290 * advancement below implicitly advances
1291 * tcp_highest_sack_seq() when skb is highest_sack.
1293 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1294 start_seq, end_seq, dup_sack, pcount,
1296 tcp_rate_skb_delivered(sk, skb, state->rate);
1298 if (skb == tp->lost_skb_hint)
1299 tp->lost_cnt_hint += pcount;
1301 TCP_SKB_CB(prev)->end_seq += shifted;
1302 TCP_SKB_CB(skb)->seq += shifted;
1304 tcp_skb_pcount_add(prev, pcount);
1305 BUG_ON(tcp_skb_pcount(skb) < pcount);
1306 tcp_skb_pcount_add(skb, -pcount);
1308 /* When we're adding to gso_segs == 1, gso_size will be zero,
1309 * in theory this shouldn't be necessary but as long as DSACK
1310 * code can come after this skb later on it's better to keep
1311 * setting gso_size to something.
1313 if (!TCP_SKB_CB(prev)->tcp_gso_size)
1314 TCP_SKB_CB(prev)->tcp_gso_size = mss;
1316 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1317 if (tcp_skb_pcount(skb) <= 1)
1318 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1320 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1321 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1324 BUG_ON(!tcp_skb_pcount(skb));
1325 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1329 /* Whole SKB was eaten :-) */
1331 if (skb == tp->retransmit_skb_hint)
1332 tp->retransmit_skb_hint = prev;
1333 if (skb == tp->lost_skb_hint) {
1334 tp->lost_skb_hint = prev;
1335 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1338 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1339 TCP_SKB_CB(prev)->eor = TCP_SKB_CB(skb)->eor;
1340 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1341 TCP_SKB_CB(prev)->end_seq++;
1343 if (skb == tcp_highest_sack(sk))
1344 tcp_advance_highest_sack(sk, skb);
1346 tcp_skb_collapse_tstamp(prev, skb);
1347 if (unlikely(TCP_SKB_CB(prev)->tx.delivered_mstamp))
1348 TCP_SKB_CB(prev)->tx.delivered_mstamp = 0;
1350 tcp_rtx_queue_unlink_and_free(skb, sk);
1352 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKMERGED);
1357 /* I wish gso_size would have a bit more sane initialization than
1358 * something-or-zero which complicates things
1360 static int tcp_skb_seglen(const struct sk_buff *skb)
1362 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1365 /* Shifting pages past head area doesn't work */
1366 static int skb_can_shift(const struct sk_buff *skb)
1368 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1371 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1374 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1375 struct tcp_sacktag_state *state,
1376 u32 start_seq, u32 end_seq,
1379 struct tcp_sock *tp = tcp_sk(sk);
1380 struct sk_buff *prev;
1386 /* Normally R but no L won't result in plain S */
1388 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1390 if (!skb_can_shift(skb))
1392 /* This frame is about to be dropped (was ACKed). */
1393 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1396 /* Can only happen with delayed DSACK + discard craziness */
1397 prev = skb_rb_prev(skb);
1401 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1404 if (!tcp_skb_can_collapse_to(prev))
1407 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1408 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1412 pcount = tcp_skb_pcount(skb);
1413 mss = tcp_skb_seglen(skb);
1415 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1416 * drop this restriction as unnecessary
1418 if (mss != tcp_skb_seglen(prev))
1421 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1423 /* CHECKME: This is non-MSS split case only?, this will
1424 * cause skipped skbs due to advancing loop btw, original
1425 * has that feature too
1427 if (tcp_skb_pcount(skb) <= 1)
1430 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1432 /* TODO: head merge to next could be attempted here
1433 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1434 * though it might not be worth of the additional hassle
1436 * ...we can probably just fallback to what was done
1437 * previously. We could try merging non-SACKed ones
1438 * as well but it probably isn't going to buy off
1439 * because later SACKs might again split them, and
1440 * it would make skb timestamp tracking considerably
1446 len = end_seq - TCP_SKB_CB(skb)->seq;
1448 BUG_ON(len > skb->len);
1450 /* MSS boundaries should be honoured or else pcount will
1451 * severely break even though it makes things bit trickier.
1452 * Optimize common case to avoid most of the divides
1454 mss = tcp_skb_mss(skb);
1456 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1457 * drop this restriction as unnecessary
1459 if (mss != tcp_skb_seglen(prev))
1464 } else if (len < mss) {
1472 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1473 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1476 if (!skb_shift(prev, skb, len))
1478 if (!tcp_shifted_skb(sk, prev, skb, state, pcount, len, mss, dup_sack))
1481 /* Hole filled allows collapsing with the next as well, this is very
1482 * useful when hole on every nth skb pattern happens
1484 skb = skb_rb_next(prev);
1488 if (!skb_can_shift(skb) ||
1489 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1490 (mss != tcp_skb_seglen(skb)))
1494 if (skb_shift(prev, skb, len)) {
1495 pcount += tcp_skb_pcount(skb);
1496 tcp_shifted_skb(sk, prev, skb, state, tcp_skb_pcount(skb),
1507 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1511 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1512 struct tcp_sack_block *next_dup,
1513 struct tcp_sacktag_state *state,
1514 u32 start_seq, u32 end_seq,
1517 struct tcp_sock *tp = tcp_sk(sk);
1518 struct sk_buff *tmp;
1520 skb_rbtree_walk_from(skb) {
1522 bool dup_sack = dup_sack_in;
1524 /* queue is in-order => we can short-circuit the walk early */
1525 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1529 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1530 in_sack = tcp_match_skb_to_sack(sk, skb,
1531 next_dup->start_seq,
1537 /* skb reference here is a bit tricky to get right, since
1538 * shifting can eat and free both this skb and the next,
1539 * so not even _safe variant of the loop is enough.
1542 tmp = tcp_shift_skb_data(sk, skb, state,
1543 start_seq, end_seq, dup_sack);
1552 in_sack = tcp_match_skb_to_sack(sk, skb,
1558 if (unlikely(in_sack < 0))
1562 TCP_SKB_CB(skb)->sacked =
1565 TCP_SKB_CB(skb)->sacked,
1566 TCP_SKB_CB(skb)->seq,
1567 TCP_SKB_CB(skb)->end_seq,
1569 tcp_skb_pcount(skb),
1571 tcp_rate_skb_delivered(sk, skb, state->rate);
1572 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1573 list_del_init(&skb->tcp_tsorted_anchor);
1575 if (!before(TCP_SKB_CB(skb)->seq,
1576 tcp_highest_sack_seq(tp)))
1577 tcp_advance_highest_sack(sk, skb);
1583 static struct sk_buff *tcp_sacktag_bsearch(struct sock *sk,
1584 struct tcp_sacktag_state *state,
1587 struct rb_node *parent, **p = &sk->tcp_rtx_queue.rb_node;
1588 struct sk_buff *skb;
1592 skb = rb_to_skb(parent);
1593 if (before(seq, TCP_SKB_CB(skb)->seq)) {
1594 p = &parent->rb_left;
1597 if (!before(seq, TCP_SKB_CB(skb)->end_seq)) {
1598 p = &parent->rb_right;
1606 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1607 struct tcp_sacktag_state *state,
1610 if (skb && after(TCP_SKB_CB(skb)->seq, skip_to_seq))
1613 return tcp_sacktag_bsearch(sk, state, skip_to_seq);
1616 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1618 struct tcp_sack_block *next_dup,
1619 struct tcp_sacktag_state *state,
1625 if (before(next_dup->start_seq, skip_to_seq)) {
1626 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1627 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1628 next_dup->start_seq, next_dup->end_seq,
1635 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1637 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1641 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1642 u32 prior_snd_una, struct tcp_sacktag_state *state)
1644 struct tcp_sock *tp = tcp_sk(sk);
1645 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1646 TCP_SKB_CB(ack_skb)->sacked);
1647 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1648 struct tcp_sack_block sp[TCP_NUM_SACKS];
1649 struct tcp_sack_block *cache;
1650 struct sk_buff *skb;
1651 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1653 bool found_dup_sack = false;
1655 int first_sack_index;
1658 state->reord = tp->snd_nxt;
1660 if (!tp->sacked_out)
1661 tcp_highest_sack_reset(sk);
1663 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1664 num_sacks, prior_snd_una);
1665 if (found_dup_sack) {
1666 state->flag |= FLAG_DSACKING_ACK;
1667 tp->delivered++; /* A spurious retransmission is delivered */
1670 /* Eliminate too old ACKs, but take into
1671 * account more or less fresh ones, they can
1672 * contain valid SACK info.
1674 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1677 if (!tp->packets_out)
1681 first_sack_index = 0;
1682 for (i = 0; i < num_sacks; i++) {
1683 bool dup_sack = !i && found_dup_sack;
1685 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1686 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1688 if (!tcp_is_sackblock_valid(tp, dup_sack,
1689 sp[used_sacks].start_seq,
1690 sp[used_sacks].end_seq)) {
1694 if (!tp->undo_marker)
1695 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1697 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1699 /* Don't count olds caused by ACK reordering */
1700 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1701 !after(sp[used_sacks].end_seq, tp->snd_una))
1703 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1706 NET_INC_STATS(sock_net(sk), mib_idx);
1708 first_sack_index = -1;
1712 /* Ignore very old stuff early */
1713 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1719 /* order SACK blocks to allow in order walk of the retrans queue */
1720 for (i = used_sacks - 1; i > 0; i--) {
1721 for (j = 0; j < i; j++) {
1722 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1723 swap(sp[j], sp[j + 1]);
1725 /* Track where the first SACK block goes to */
1726 if (j == first_sack_index)
1727 first_sack_index = j + 1;
1732 state->mss_now = tcp_current_mss(sk);
1736 if (!tp->sacked_out) {
1737 /* It's already past, so skip checking against it */
1738 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1740 cache = tp->recv_sack_cache;
1741 /* Skip empty blocks in at head of the cache */
1742 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1747 while (i < used_sacks) {
1748 u32 start_seq = sp[i].start_seq;
1749 u32 end_seq = sp[i].end_seq;
1750 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1751 struct tcp_sack_block *next_dup = NULL;
1753 if (found_dup_sack && ((i + 1) == first_sack_index))
1754 next_dup = &sp[i + 1];
1756 /* Skip too early cached blocks */
1757 while (tcp_sack_cache_ok(tp, cache) &&
1758 !before(start_seq, cache->end_seq))
1761 /* Can skip some work by looking recv_sack_cache? */
1762 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1763 after(end_seq, cache->start_seq)) {
1766 if (before(start_seq, cache->start_seq)) {
1767 skb = tcp_sacktag_skip(skb, sk, state,
1769 skb = tcp_sacktag_walk(skb, sk, next_dup,
1776 /* Rest of the block already fully processed? */
1777 if (!after(end_seq, cache->end_seq))
1780 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1784 /* ...tail remains todo... */
1785 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1786 /* ...but better entrypoint exists! */
1787 skb = tcp_highest_sack(sk);
1794 skb = tcp_sacktag_skip(skb, sk, state, cache->end_seq);
1795 /* Check overlap against next cached too (past this one already) */
1800 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1801 skb = tcp_highest_sack(sk);
1805 skb = tcp_sacktag_skip(skb, sk, state, start_seq);
1808 skb = tcp_sacktag_walk(skb, sk, next_dup, state,
1809 start_seq, end_seq, dup_sack);
1815 /* Clear the head of the cache sack blocks so we can skip it next time */
1816 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1817 tp->recv_sack_cache[i].start_seq = 0;
1818 tp->recv_sack_cache[i].end_seq = 0;
1820 for (j = 0; j < used_sacks; j++)
1821 tp->recv_sack_cache[i++] = sp[j];
1823 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Loss || tp->undo_marker)
1824 tcp_check_sack_reordering(sk, state->reord, 0);
1826 tcp_verify_left_out(tp);
1829 #if FASTRETRANS_DEBUG > 0
1830 WARN_ON((int)tp->sacked_out < 0);
1831 WARN_ON((int)tp->lost_out < 0);
1832 WARN_ON((int)tp->retrans_out < 0);
1833 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1838 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1839 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1841 static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
1845 holes = max(tp->lost_out, 1U);
1846 holes = min(holes, tp->packets_out);
1848 if ((tp->sacked_out + holes) > tp->packets_out) {
1849 tp->sacked_out = tp->packets_out - holes;
1855 /* If we receive more dupacks than we expected counting segments
1856 * in assumption of absent reordering, interpret this as reordering.
1857 * The only another reason could be bug in receiver TCP.
1859 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1861 struct tcp_sock *tp = tcp_sk(sk);
1863 if (!tcp_limit_reno_sacked(tp))
1866 tp->reordering = min_t(u32, tp->packets_out + addend,
1867 sock_net(sk)->ipv4.sysctl_tcp_max_reordering);
1868 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRENOREORDER);
1871 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1873 static void tcp_add_reno_sack(struct sock *sk)
1875 struct tcp_sock *tp = tcp_sk(sk);
1876 u32 prior_sacked = tp->sacked_out;
1879 tcp_check_reno_reordering(sk, 0);
1880 if (tp->sacked_out > prior_sacked)
1881 tp->delivered++; /* Some out-of-order packet is delivered */
1882 tcp_verify_left_out(tp);
1885 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1887 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1889 struct tcp_sock *tp = tcp_sk(sk);
1892 /* One ACK acked hole. The rest eat duplicate ACKs. */
1893 tp->delivered += max_t(int, acked - tp->sacked_out, 1);
1894 if (acked - 1 >= tp->sacked_out)
1897 tp->sacked_out -= acked - 1;
1899 tcp_check_reno_reordering(sk, acked);
1900 tcp_verify_left_out(tp);
1903 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1908 void tcp_clear_retrans(struct tcp_sock *tp)
1910 tp->retrans_out = 0;
1912 tp->undo_marker = 0;
1913 tp->undo_retrans = -1;
1917 static inline void tcp_init_undo(struct tcp_sock *tp)
1919 tp->undo_marker = tp->snd_una;
1920 /* Retransmission still in flight may cause DSACKs later. */
1921 tp->undo_retrans = tp->retrans_out ? : -1;
1924 static bool tcp_is_rack(const struct sock *sk)
1926 return sock_net(sk)->ipv4.sysctl_tcp_recovery & TCP_RACK_LOSS_DETECTION;
1929 /* If we detect SACK reneging, forget all SACK information
1930 * and reset tags completely, otherwise preserve SACKs. If receiver
1931 * dropped its ofo queue, we will know this due to reneging detection.
1933 static void tcp_timeout_mark_lost(struct sock *sk)
1935 struct tcp_sock *tp = tcp_sk(sk);
1936 struct sk_buff *skb, *head;
1937 bool is_reneg; /* is receiver reneging on SACKs? */
1939 head = tcp_rtx_queue_head(sk);
1940 is_reneg = head && (TCP_SKB_CB(head)->sacked & TCPCB_SACKED_ACKED);
1942 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
1944 /* Mark SACK reneging until we recover from this loss event. */
1945 tp->is_sack_reneg = 1;
1946 } else if (tcp_is_reno(tp)) {
1947 tcp_reset_reno_sack(tp);
1951 skb_rbtree_walk_from(skb) {
1953 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1954 else if (tcp_is_rack(sk) && skb != head &&
1955 tcp_rack_skb_timeout(tp, skb, 0) > 0)
1956 continue; /* Don't mark recently sent ones lost yet */
1957 tcp_mark_skb_lost(sk, skb);
1959 tcp_verify_left_out(tp);
1960 tcp_clear_all_retrans_hints(tp);
1963 /* Enter Loss state. */
1964 void tcp_enter_loss(struct sock *sk)
1966 const struct inet_connection_sock *icsk = inet_csk(sk);
1967 struct tcp_sock *tp = tcp_sk(sk);
1968 struct net *net = sock_net(sk);
1969 bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery;
1971 tcp_timeout_mark_lost(sk);
1973 /* Reduce ssthresh if it has not yet been made inside this window. */
1974 if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
1975 !after(tp->high_seq, tp->snd_una) ||
1976 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1977 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1978 tp->prior_cwnd = tp->snd_cwnd;
1979 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1980 tcp_ca_event(sk, CA_EVENT_LOSS);
1983 tp->snd_cwnd = tcp_packets_in_flight(tp) + 1;
1984 tp->snd_cwnd_cnt = 0;
1985 tp->snd_cwnd_stamp = tcp_jiffies32;
1987 /* Timeout in disordered state after receiving substantial DUPACKs
1988 * suggests that the degree of reordering is over-estimated.
1990 if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
1991 tp->sacked_out >= net->ipv4.sysctl_tcp_reordering)
1992 tp->reordering = min_t(unsigned int, tp->reordering,
1993 net->ipv4.sysctl_tcp_reordering);
1994 tcp_set_ca_state(sk, TCP_CA_Loss);
1995 tp->high_seq = tp->snd_nxt;
1996 tcp_ecn_queue_cwr(tp);
1998 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
1999 * loss recovery is underway except recurring timeout(s) on
2000 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
2002 tp->frto = net->ipv4.sysctl_tcp_frto &&
2003 (new_recovery || icsk->icsk_retransmits) &&
2004 !inet_csk(sk)->icsk_mtup.probe_size;
2007 /* If ACK arrived pointing to a remembered SACK, it means that our
2008 * remembered SACKs do not reflect real state of receiver i.e.
2009 * receiver _host_ is heavily congested (or buggy).
2011 * To avoid big spurious retransmission bursts due to transient SACK
2012 * scoreboard oddities that look like reneging, we give the receiver a
2013 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
2014 * restore sanity to the SACK scoreboard. If the apparent reneging
2015 * persists until this RTO then we'll clear the SACK scoreboard.
2017 static bool tcp_check_sack_reneging(struct sock *sk, int flag)
2019 if (flag & FLAG_SACK_RENEGING) {
2020 struct tcp_sock *tp = tcp_sk(sk);
2021 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4),
2022 msecs_to_jiffies(10));
2024 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2025 delay, TCP_RTO_MAX);
2031 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2032 * counter when SACK is enabled (without SACK, sacked_out is used for
2035 * With reordering, holes may still be in flight, so RFC3517 recovery
2036 * uses pure sacked_out (total number of SACKed segments) even though
2037 * it violates the RFC that uses duplicate ACKs, often these are equal
2038 * but when e.g. out-of-window ACKs or packet duplication occurs,
2039 * they differ. Since neither occurs due to loss, TCP should really
2042 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2044 return tp->sacked_out + 1;
2047 /* Linux NewReno/SACK/ECN state machine.
2048 * --------------------------------------
2050 * "Open" Normal state, no dubious events, fast path.
2051 * "Disorder" In all the respects it is "Open",
2052 * but requires a bit more attention. It is entered when
2053 * we see some SACKs or dupacks. It is split of "Open"
2054 * mainly to move some processing from fast path to slow one.
2055 * "CWR" CWND was reduced due to some Congestion Notification event.
2056 * It can be ECN, ICMP source quench, local device congestion.
2057 * "Recovery" CWND was reduced, we are fast-retransmitting.
2058 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2060 * tcp_fastretrans_alert() is entered:
2061 * - each incoming ACK, if state is not "Open"
2062 * - when arrived ACK is unusual, namely:
2067 * Counting packets in flight is pretty simple.
2069 * in_flight = packets_out - left_out + retrans_out
2071 * packets_out is SND.NXT-SND.UNA counted in packets.
2073 * retrans_out is number of retransmitted segments.
2075 * left_out is number of segments left network, but not ACKed yet.
2077 * left_out = sacked_out + lost_out
2079 * sacked_out: Packets, which arrived to receiver out of order
2080 * and hence not ACKed. With SACKs this number is simply
2081 * amount of SACKed data. Even without SACKs
2082 * it is easy to give pretty reliable estimate of this number,
2083 * counting duplicate ACKs.
2085 * lost_out: Packets lost by network. TCP has no explicit
2086 * "loss notification" feedback from network (for now).
2087 * It means that this number can be only _guessed_.
2088 * Actually, it is the heuristics to predict lossage that
2089 * distinguishes different algorithms.
2091 * F.e. after RTO, when all the queue is considered as lost,
2092 * lost_out = packets_out and in_flight = retrans_out.
2094 * Essentially, we have now a few algorithms detecting
2097 * If the receiver supports SACK:
2099 * RFC6675/3517: It is the conventional algorithm. A packet is
2100 * considered lost if the number of higher sequence packets
2101 * SACKed is greater than or equal the DUPACK thoreshold
2102 * (reordering). This is implemented in tcp_mark_head_lost and
2103 * tcp_update_scoreboard.
2105 * RACK (draft-ietf-tcpm-rack-01): it is a newer algorithm
2106 * (2017-) that checks timing instead of counting DUPACKs.
2107 * Essentially a packet is considered lost if it's not S/ACKed
2108 * after RTT + reordering_window, where both metrics are
2109 * dynamically measured and adjusted. This is implemented in
2110 * tcp_rack_mark_lost.
2112 * If the receiver does not support SACK:
2114 * NewReno (RFC6582): in Recovery we assume that one segment
2115 * is lost (classic Reno). While we are in Recovery and
2116 * a partial ACK arrives, we assume that one more packet
2117 * is lost (NewReno). This heuristics are the same in NewReno
2120 * Really tricky (and requiring careful tuning) part of algorithm
2121 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2122 * The first determines the moment _when_ we should reduce CWND and,
2123 * hence, slow down forward transmission. In fact, it determines the moment
2124 * when we decide that hole is caused by loss, rather than by a reorder.
2126 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2127 * holes, caused by lost packets.
2129 * And the most logically complicated part of algorithm is undo
2130 * heuristics. We detect false retransmits due to both too early
2131 * fast retransmit (reordering) and underestimated RTO, analyzing
2132 * timestamps and D-SACKs. When we detect that some segments were
2133 * retransmitted by mistake and CWND reduction was wrong, we undo
2134 * window reduction and abort recovery phase. This logic is hidden
2135 * inside several functions named tcp_try_undo_<something>.
2138 /* This function decides, when we should leave Disordered state
2139 * and enter Recovery phase, reducing congestion window.
2141 * Main question: may we further continue forward transmission
2142 * with the same cwnd?
2144 static bool tcp_time_to_recover(struct sock *sk, int flag)
2146 struct tcp_sock *tp = tcp_sk(sk);
2148 /* Trick#1: The loss is proven. */
2152 /* Not-A-Trick#2 : Classic rule... */
2153 if (!tcp_is_rack(sk) && tcp_dupack_heuristics(tp) > tp->reordering)
2159 /* Detect loss in event "A" above by marking head of queue up as lost.
2160 * For non-SACK(Reno) senders, the first "packets" number of segments
2161 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2162 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2163 * the maximum SACKed segments to pass before reaching this limit.
2165 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2167 struct tcp_sock *tp = tcp_sk(sk);
2168 struct sk_buff *skb;
2169 int cnt, oldcnt, lost;
2171 /* Use SACK to deduce losses of new sequences sent during recovery */
2172 const u32 loss_high = tcp_is_sack(tp) ? tp->snd_nxt : tp->high_seq;
2174 WARN_ON(packets > tp->packets_out);
2175 skb = tp->lost_skb_hint;
2177 /* Head already handled? */
2178 if (mark_head && after(TCP_SKB_CB(skb)->seq, tp->snd_una))
2180 cnt = tp->lost_cnt_hint;
2182 skb = tcp_rtx_queue_head(sk);
2186 skb_rbtree_walk_from(skb) {
2187 /* TODO: do this better */
2188 /* this is not the most efficient way to do this... */
2189 tp->lost_skb_hint = skb;
2190 tp->lost_cnt_hint = cnt;
2192 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2196 if (tcp_is_reno(tp) ||
2197 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2198 cnt += tcp_skb_pcount(skb);
2200 if (cnt > packets) {
2201 if (tcp_is_sack(tp) ||
2202 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
2203 (oldcnt >= packets))
2206 mss = tcp_skb_mss(skb);
2207 /* If needed, chop off the prefix to mark as lost. */
2208 lost = (packets - oldcnt) * mss;
2209 if (lost < skb->len &&
2210 tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
2211 lost, mss, GFP_ATOMIC) < 0)
2216 tcp_skb_mark_lost(tp, skb);
2221 tcp_verify_left_out(tp);
2224 /* Account newly detected lost packet(s) */
2226 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2228 struct tcp_sock *tp = tcp_sk(sk);
2230 if (tcp_is_sack(tp)) {
2231 int sacked_upto = tp->sacked_out - tp->reordering;
2232 if (sacked_upto >= 0)
2233 tcp_mark_head_lost(sk, sacked_upto, 0);
2234 else if (fast_rexmit)
2235 tcp_mark_head_lost(sk, 1, 1);
2239 static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when)
2241 return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2242 before(tp->rx_opt.rcv_tsecr, when);
2245 /* skb is spurious retransmitted if the returned timestamp echo
2246 * reply is prior to the skb transmission time
2248 static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp,
2249 const struct sk_buff *skb)
2251 return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) &&
2252 tcp_tsopt_ecr_before(tp, tcp_skb_timestamp(skb));
2255 /* Nothing was retransmitted or returned timestamp is less
2256 * than timestamp of the first retransmission.
2258 static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2260 return !tp->retrans_stamp ||
2261 tcp_tsopt_ecr_before(tp, tp->retrans_stamp);
2264 /* Undo procedures. */
2266 /* We can clear retrans_stamp when there are no retransmissions in the
2267 * window. It would seem that it is trivially available for us in
2268 * tp->retrans_out, however, that kind of assumptions doesn't consider
2269 * what will happen if errors occur when sending retransmission for the
2270 * second time. ...It could the that such segment has only
2271 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2272 * the head skb is enough except for some reneging corner cases that
2273 * are not worth the effort.
2275 * Main reason for all this complexity is the fact that connection dying
2276 * time now depends on the validity of the retrans_stamp, in particular,
2277 * that successive retransmissions of a segment must not advance
2278 * retrans_stamp under any conditions.
2280 static bool tcp_any_retrans_done(const struct sock *sk)
2282 const struct tcp_sock *tp = tcp_sk(sk);
2283 struct sk_buff *skb;
2285 if (tp->retrans_out)
2288 skb = tcp_rtx_queue_head(sk);
2289 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2295 static void DBGUNDO(struct sock *sk, const char *msg)
2297 #if FASTRETRANS_DEBUG > 1
2298 struct tcp_sock *tp = tcp_sk(sk);
2299 struct inet_sock *inet = inet_sk(sk);
2301 if (sk->sk_family == AF_INET) {
2302 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2304 &inet->inet_daddr, ntohs(inet->inet_dport),
2305 tp->snd_cwnd, tcp_left_out(tp),
2306 tp->snd_ssthresh, tp->prior_ssthresh,
2309 #if IS_ENABLED(CONFIG_IPV6)
2310 else if (sk->sk_family == AF_INET6) {
2311 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2313 &sk->sk_v6_daddr, ntohs(inet->inet_dport),
2314 tp->snd_cwnd, tcp_left_out(tp),
2315 tp->snd_ssthresh, tp->prior_ssthresh,
2322 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
2324 struct tcp_sock *tp = tcp_sk(sk);
2327 struct sk_buff *skb;
2329 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2330 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2333 tcp_clear_all_retrans_hints(tp);
2336 if (tp->prior_ssthresh) {
2337 const struct inet_connection_sock *icsk = inet_csk(sk);
2339 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2341 if (tp->prior_ssthresh > tp->snd_ssthresh) {
2342 tp->snd_ssthresh = tp->prior_ssthresh;
2343 tcp_ecn_withdraw_cwr(tp);
2346 tp->snd_cwnd_stamp = tcp_jiffies32;
2347 tp->undo_marker = 0;
2348 tp->rack.advanced = 1; /* Force RACK to re-exam losses */
2351 static inline bool tcp_may_undo(const struct tcp_sock *tp)
2353 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2356 /* People celebrate: "We love our President!" */
2357 static bool tcp_try_undo_recovery(struct sock *sk)
2359 struct tcp_sock *tp = tcp_sk(sk);
2361 if (tcp_may_undo(tp)) {
2364 /* Happy end! We did not retransmit anything
2365 * or our original transmission succeeded.
2367 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2368 tcp_undo_cwnd_reduction(sk, false);
2369 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2370 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2372 mib_idx = LINUX_MIB_TCPFULLUNDO;
2374 NET_INC_STATS(sock_net(sk), mib_idx);
2375 } else if (tp->rack.reo_wnd_persist) {
2376 tp->rack.reo_wnd_persist--;
2378 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2379 /* Hold old state until something *above* high_seq
2380 * is ACKed. For Reno it is MUST to prevent false
2381 * fast retransmits (RFC2582). SACK TCP is safe. */
2382 if (!tcp_any_retrans_done(sk))
2383 tp->retrans_stamp = 0;
2386 tcp_set_ca_state(sk, TCP_CA_Open);
2387 tp->is_sack_reneg = 0;
2391 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2392 static bool tcp_try_undo_dsack(struct sock *sk)
2394 struct tcp_sock *tp = tcp_sk(sk);
2396 if (tp->undo_marker && !tp->undo_retrans) {
2397 tp->rack.reo_wnd_persist = min(TCP_RACK_RECOVERY_THRESH,
2398 tp->rack.reo_wnd_persist + 1);
2399 DBGUNDO(sk, "D-SACK");
2400 tcp_undo_cwnd_reduction(sk, false);
2401 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2407 /* Undo during loss recovery after partial ACK or using F-RTO. */
2408 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2410 struct tcp_sock *tp = tcp_sk(sk);
2412 if (frto_undo || tcp_may_undo(tp)) {
2413 tcp_undo_cwnd_reduction(sk, true);
2415 DBGUNDO(sk, "partial loss");
2416 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2418 NET_INC_STATS(sock_net(sk),
2419 LINUX_MIB_TCPSPURIOUSRTOS);
2420 inet_csk(sk)->icsk_retransmits = 0;
2421 if (frto_undo || tcp_is_sack(tp)) {
2422 tcp_set_ca_state(sk, TCP_CA_Open);
2423 tp->is_sack_reneg = 0;
2430 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
2431 * It computes the number of packets to send (sndcnt) based on packets newly
2433 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2434 * cwnd reductions across a full RTT.
2435 * 2) Otherwise PRR uses packet conservation to send as much as delivered.
2436 * But when the retransmits are acked without further losses, PRR
2437 * slow starts cwnd up to ssthresh to speed up the recovery.
2439 static void tcp_init_cwnd_reduction(struct sock *sk)
2441 struct tcp_sock *tp = tcp_sk(sk);
2443 tp->high_seq = tp->snd_nxt;
2444 tp->tlp_high_seq = 0;
2445 tp->snd_cwnd_cnt = 0;
2446 tp->prior_cwnd = tp->snd_cwnd;
2447 tp->prr_delivered = 0;
2449 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2450 tcp_ecn_queue_cwr(tp);
2453 void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked, int flag)
2455 struct tcp_sock *tp = tcp_sk(sk);
2457 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2459 if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd))
2462 tp->prr_delivered += newly_acked_sacked;
2464 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2466 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2467 } else if ((flag & FLAG_RETRANS_DATA_ACKED) &&
2468 !(flag & FLAG_LOST_RETRANS)) {
2469 sndcnt = min_t(int, delta,
2470 max_t(int, tp->prr_delivered - tp->prr_out,
2471 newly_acked_sacked) + 1);
2473 sndcnt = min(delta, newly_acked_sacked);
2475 /* Force a fast retransmit upon entering fast recovery */
2476 sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1));
2477 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2480 static inline void tcp_end_cwnd_reduction(struct sock *sk)
2482 struct tcp_sock *tp = tcp_sk(sk);
2484 if (inet_csk(sk)->icsk_ca_ops->cong_control)
2487 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2488 if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH &&
2489 (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR || tp->undo_marker)) {
2490 tp->snd_cwnd = tp->snd_ssthresh;
2491 tp->snd_cwnd_stamp = tcp_jiffies32;
2493 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2496 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2497 void tcp_enter_cwr(struct sock *sk)
2499 struct tcp_sock *tp = tcp_sk(sk);
2501 tp->prior_ssthresh = 0;
2502 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2503 tp->undo_marker = 0;
2504 tcp_init_cwnd_reduction(sk);
2505 tcp_set_ca_state(sk, TCP_CA_CWR);
2508 EXPORT_SYMBOL(tcp_enter_cwr);
2510 static void tcp_try_keep_open(struct sock *sk)
2512 struct tcp_sock *tp = tcp_sk(sk);
2513 int state = TCP_CA_Open;
2515 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2516 state = TCP_CA_Disorder;
2518 if (inet_csk(sk)->icsk_ca_state != state) {
2519 tcp_set_ca_state(sk, state);
2520 tp->high_seq = tp->snd_nxt;
2524 static void tcp_try_to_open(struct sock *sk, int flag)
2526 struct tcp_sock *tp = tcp_sk(sk);
2528 tcp_verify_left_out(tp);
2530 if (!tcp_any_retrans_done(sk))
2531 tp->retrans_stamp = 0;
2533 if (flag & FLAG_ECE)
2536 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2537 tcp_try_keep_open(sk);
2541 static void tcp_mtup_probe_failed(struct sock *sk)
2543 struct inet_connection_sock *icsk = inet_csk(sk);
2545 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2546 icsk->icsk_mtup.probe_size = 0;
2547 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPFAIL);
2550 static void tcp_mtup_probe_success(struct sock *sk)
2552 struct tcp_sock *tp = tcp_sk(sk);
2553 struct inet_connection_sock *icsk = inet_csk(sk);
2555 /* FIXME: breaks with very large cwnd */
2556 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2557 tp->snd_cwnd = tp->snd_cwnd *
2558 tcp_mss_to_mtu(sk, tp->mss_cache) /
2559 icsk->icsk_mtup.probe_size;
2560 tp->snd_cwnd_cnt = 0;
2561 tp->snd_cwnd_stamp = tcp_jiffies32;
2562 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2564 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2565 icsk->icsk_mtup.probe_size = 0;
2566 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2567 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS);
2570 /* Do a simple retransmit without using the backoff mechanisms in
2571 * tcp_timer. This is used for path mtu discovery.
2572 * The socket is already locked here.
2574 void tcp_simple_retransmit(struct sock *sk)
2576 const struct inet_connection_sock *icsk = inet_csk(sk);
2577 struct tcp_sock *tp = tcp_sk(sk);
2578 struct sk_buff *skb;
2579 unsigned int mss = tcp_current_mss(sk);
2581 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
2582 if (tcp_skb_seglen(skb) > mss &&
2583 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2584 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2585 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2586 tp->retrans_out -= tcp_skb_pcount(skb);
2588 tcp_skb_mark_lost_uncond_verify(tp, skb);
2592 tcp_clear_retrans_hints_partial(tp);
2597 if (tcp_is_reno(tp))
2598 tcp_limit_reno_sacked(tp);
2600 tcp_verify_left_out(tp);
2602 /* Don't muck with the congestion window here.
2603 * Reason is that we do not increase amount of _data_
2604 * in network, but units changed and effective
2605 * cwnd/ssthresh really reduced now.
2607 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2608 tp->high_seq = tp->snd_nxt;
2609 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2610 tp->prior_ssthresh = 0;
2611 tp->undo_marker = 0;
2612 tcp_set_ca_state(sk, TCP_CA_Loss);
2614 tcp_xmit_retransmit_queue(sk);
2616 EXPORT_SYMBOL(tcp_simple_retransmit);
2618 void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2620 struct tcp_sock *tp = tcp_sk(sk);
2623 if (tcp_is_reno(tp))
2624 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2626 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2628 NET_INC_STATS(sock_net(sk), mib_idx);
2630 tp->prior_ssthresh = 0;
2633 if (!tcp_in_cwnd_reduction(sk)) {
2635 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2636 tcp_init_cwnd_reduction(sk);
2638 tcp_set_ca_state(sk, TCP_CA_Recovery);
2641 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2642 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2644 static void tcp_process_loss(struct sock *sk, int flag, bool is_dupack,
2647 struct tcp_sock *tp = tcp_sk(sk);
2648 bool recovered = !before(tp->snd_una, tp->high_seq);
2650 if ((flag & FLAG_SND_UNA_ADVANCED) &&
2651 tcp_try_undo_loss(sk, false))
2654 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2655 /* Step 3.b. A timeout is spurious if not all data are
2656 * lost, i.e., never-retransmitted data are (s)acked.
2658 if ((flag & FLAG_ORIG_SACK_ACKED) &&
2659 tcp_try_undo_loss(sk, true))
2662 if (after(tp->snd_nxt, tp->high_seq)) {
2663 if (flag & FLAG_DATA_SACKED || is_dupack)
2664 tp->frto = 0; /* Step 3.a. loss was real */
2665 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2666 tp->high_seq = tp->snd_nxt;
2667 /* Step 2.b. Try send new data (but deferred until cwnd
2668 * is updated in tcp_ack()). Otherwise fall back to
2669 * the conventional recovery.
2671 if (!tcp_write_queue_empty(sk) &&
2672 after(tcp_wnd_end(tp), tp->snd_nxt)) {
2673 *rexmit = REXMIT_NEW;
2681 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2682 tcp_try_undo_recovery(sk);
2685 if (tcp_is_reno(tp)) {
2686 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2687 * delivered. Lower inflight to clock out (re)tranmissions.
2689 if (after(tp->snd_nxt, tp->high_seq) && is_dupack)
2690 tcp_add_reno_sack(sk);
2691 else if (flag & FLAG_SND_UNA_ADVANCED)
2692 tcp_reset_reno_sack(tp);
2694 *rexmit = REXMIT_LOST;
2697 /* Undo during fast recovery after partial ACK. */
2698 static bool tcp_try_undo_partial(struct sock *sk, u32 prior_snd_una)
2700 struct tcp_sock *tp = tcp_sk(sk);
2702 if (tp->undo_marker && tcp_packet_delayed(tp)) {
2703 /* Plain luck! Hole if filled with delayed
2704 * packet, rather than with a retransmit. Check reordering.
2706 tcp_check_sack_reordering(sk, prior_snd_una, 1);
2708 /* We are getting evidence that the reordering degree is higher
2709 * than we realized. If there are no retransmits out then we
2710 * can undo. Otherwise we clock out new packets but do not
2711 * mark more packets lost or retransmit more.
2713 if (tp->retrans_out)
2716 if (!tcp_any_retrans_done(sk))
2717 tp->retrans_stamp = 0;
2719 DBGUNDO(sk, "partial recovery");
2720 tcp_undo_cwnd_reduction(sk, true);
2721 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2722 tcp_try_keep_open(sk);
2728 static void tcp_identify_packet_loss(struct sock *sk, int *ack_flag)
2730 struct tcp_sock *tp = tcp_sk(sk);
2732 if (tcp_rtx_queue_empty(sk))
2735 if (unlikely(tcp_is_reno(tp))) {
2736 tcp_newreno_mark_lost(sk, *ack_flag & FLAG_SND_UNA_ADVANCED);
2737 } else if (tcp_is_rack(sk)) {
2738 u32 prior_retrans = tp->retrans_out;
2740 tcp_rack_mark_lost(sk);
2741 if (prior_retrans > tp->retrans_out)
2742 *ack_flag |= FLAG_LOST_RETRANS;
2746 static bool tcp_force_fast_retransmit(struct sock *sk)
2748 struct tcp_sock *tp = tcp_sk(sk);
2750 return after(tcp_highest_sack_seq(tp),
2751 tp->snd_una + tp->reordering * tp->mss_cache);
2754 /* Process an event, which can update packets-in-flight not trivially.
2755 * Main goal of this function is to calculate new estimate for left_out,
2756 * taking into account both packets sitting in receiver's buffer and
2757 * packets lost by network.
2759 * Besides that it updates the congestion state when packet loss or ECN
2760 * is detected. But it does not reduce the cwnd, it is done by the
2761 * congestion control later.
2763 * It does _not_ decide what to send, it is made in function
2764 * tcp_xmit_retransmit_queue().
2766 static void tcp_fastretrans_alert(struct sock *sk, const u32 prior_snd_una,
2767 bool is_dupack, int *ack_flag, int *rexmit)
2769 struct inet_connection_sock *icsk = inet_csk(sk);
2770 struct tcp_sock *tp = tcp_sk(sk);
2771 int fast_rexmit = 0, flag = *ack_flag;
2772 bool do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2773 tcp_force_fast_retransmit(sk));
2775 if (!tp->packets_out && tp->sacked_out)
2778 /* Now state machine starts.
2779 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2780 if (flag & FLAG_ECE)
2781 tp->prior_ssthresh = 0;
2783 /* B. In all the states check for reneging SACKs. */
2784 if (tcp_check_sack_reneging(sk, flag))
2787 /* C. Check consistency of the current state. */
2788 tcp_verify_left_out(tp);
2790 /* D. Check state exit conditions. State can be terminated
2791 * when high_seq is ACKed. */
2792 if (icsk->icsk_ca_state == TCP_CA_Open) {
2793 WARN_ON(tp->retrans_out != 0);
2794 tp->retrans_stamp = 0;
2795 } else if (!before(tp->snd_una, tp->high_seq)) {
2796 switch (icsk->icsk_ca_state) {
2798 /* CWR is to be held something *above* high_seq
2799 * is ACKed for CWR bit to reach receiver. */
2800 if (tp->snd_una != tp->high_seq) {
2801 tcp_end_cwnd_reduction(sk);
2802 tcp_set_ca_state(sk, TCP_CA_Open);
2806 case TCP_CA_Recovery:
2807 if (tcp_is_reno(tp))
2808 tcp_reset_reno_sack(tp);
2809 if (tcp_try_undo_recovery(sk))
2811 tcp_end_cwnd_reduction(sk);
2816 /* E. Process state. */
2817 switch (icsk->icsk_ca_state) {
2818 case TCP_CA_Recovery:
2819 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2820 if (tcp_is_reno(tp) && is_dupack)
2821 tcp_add_reno_sack(sk);
2823 if (tcp_try_undo_partial(sk, prior_snd_una))
2825 /* Partial ACK arrived. Force fast retransmit. */
2826 do_lost = tcp_is_reno(tp) ||
2827 tcp_force_fast_retransmit(sk);
2829 if (tcp_try_undo_dsack(sk)) {
2830 tcp_try_keep_open(sk);
2833 tcp_identify_packet_loss(sk, ack_flag);
2836 tcp_process_loss(sk, flag, is_dupack, rexmit);
2837 tcp_identify_packet_loss(sk, ack_flag);
2838 if (!(icsk->icsk_ca_state == TCP_CA_Open ||
2839 (*ack_flag & FLAG_LOST_RETRANS)))
2841 /* Change state if cwnd is undone or retransmits are lost */
2844 if (tcp_is_reno(tp)) {
2845 if (flag & FLAG_SND_UNA_ADVANCED)
2846 tcp_reset_reno_sack(tp);
2848 tcp_add_reno_sack(sk);
2851 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
2852 tcp_try_undo_dsack(sk);
2854 tcp_identify_packet_loss(sk, ack_flag);
2855 if (!tcp_time_to_recover(sk, flag)) {
2856 tcp_try_to_open(sk, flag);
2860 /* MTU probe failure: don't reduce cwnd */
2861 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2862 icsk->icsk_mtup.probe_size &&
2863 tp->snd_una == tp->mtu_probe.probe_seq_start) {
2864 tcp_mtup_probe_failed(sk);
2865 /* Restores the reduction we did in tcp_mtup_probe() */
2867 tcp_simple_retransmit(sk);
2871 /* Otherwise enter Recovery state */
2872 tcp_enter_recovery(sk, (flag & FLAG_ECE));
2876 if (!tcp_is_rack(sk) && do_lost)
2877 tcp_update_scoreboard(sk, fast_rexmit);
2878 *rexmit = REXMIT_LOST;
2881 static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us, const int flag)
2883 u32 wlen = sock_net(sk)->ipv4.sysctl_tcp_min_rtt_wlen * HZ;
2884 struct tcp_sock *tp = tcp_sk(sk);
2886 if ((flag & FLAG_ACK_MAYBE_DELAYED) && rtt_us > tcp_min_rtt(tp)) {
2887 /* If the remote keeps returning delayed ACKs, eventually
2888 * the min filter would pick it up and overestimate the
2889 * prop. delay when it expires. Skip suspected delayed ACKs.
2893 minmax_running_min(&tp->rtt_min, wlen, tcp_jiffies32,
2894 rtt_us ? : jiffies_to_usecs(1));
2897 static bool tcp_ack_update_rtt(struct sock *sk, const int flag,
2898 long seq_rtt_us, long sack_rtt_us,
2899 long ca_rtt_us, struct rate_sample *rs)
2901 const struct tcp_sock *tp = tcp_sk(sk);
2903 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2904 * broken middle-boxes or peers may corrupt TS-ECR fields. But
2905 * Karn's algorithm forbids taking RTT if some retransmitted data
2906 * is acked (RFC6298).
2909 seq_rtt_us = sack_rtt_us;
2911 /* RTTM Rule: A TSecr value received in a segment is used to
2912 * update the averaged RTT measurement only if the segment
2913 * acknowledges some new data, i.e., only if it advances the
2914 * left edge of the send window.
2915 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2917 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2918 flag & FLAG_ACKED) {
2919 u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr;
2920 u32 delta_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
2922 seq_rtt_us = ca_rtt_us = delta_us;
2924 rs->rtt_us = ca_rtt_us; /* RTT of last (S)ACKed packet (or -1) */
2928 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
2929 * always taken together with ACK, SACK, or TS-opts. Any negative
2930 * values will be skipped with the seq_rtt_us < 0 check above.
2932 tcp_update_rtt_min(sk, ca_rtt_us, flag);
2933 tcp_rtt_estimator(sk, seq_rtt_us);
2936 /* RFC6298: only reset backoff on valid RTT measurement. */
2937 inet_csk(sk)->icsk_backoff = 0;
2941 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
2942 void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req)
2944 struct rate_sample rs;
2947 if (req && !req->num_retrans && tcp_rsk(req)->snt_synack)
2948 rtt_us = tcp_stamp_us_delta(tcp_clock_us(), tcp_rsk(req)->snt_synack);
2950 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us, &rs);
2954 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
2956 const struct inet_connection_sock *icsk = inet_csk(sk);
2958 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
2959 tcp_sk(sk)->snd_cwnd_stamp = tcp_jiffies32;
2962 /* Restart timer after forward progress on connection.
2963 * RFC2988 recommends to restart timer to now+rto.
2965 void tcp_rearm_rto(struct sock *sk)
2967 const struct inet_connection_sock *icsk = inet_csk(sk);
2968 struct tcp_sock *tp = tcp_sk(sk);
2970 /* If the retrans timer is currently being used by Fast Open
2971 * for SYN-ACK retrans purpose, stay put.
2973 if (tp->fastopen_rsk)
2976 if (!tp->packets_out) {
2977 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
2979 u32 rto = inet_csk(sk)->icsk_rto;
2980 /* Offset the time elapsed after installing regular RTO */
2981 if (icsk->icsk_pending == ICSK_TIME_REO_TIMEOUT ||
2982 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
2983 s64 delta_us = tcp_rto_delta_us(sk);
2984 /* delta_us may not be positive if the socket is locked
2985 * when the retrans timer fires and is rescheduled.
2987 rto = usecs_to_jiffies(max_t(int, delta_us, 1));
2989 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
2994 /* Try to schedule a loss probe; if that doesn't work, then schedule an RTO. */
2995 static void tcp_set_xmit_timer(struct sock *sk)
2997 if (!tcp_schedule_loss_probe(sk, true))
3001 /* If we get here, the whole TSO packet has not been acked. */
3002 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3004 struct tcp_sock *tp = tcp_sk(sk);
3007 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3009 packets_acked = tcp_skb_pcount(skb);
3010 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3012 packets_acked -= tcp_skb_pcount(skb);
3014 if (packets_acked) {
3015 BUG_ON(tcp_skb_pcount(skb) == 0);
3016 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3019 return packets_acked;
3022 static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
3025 const struct skb_shared_info *shinfo;
3027 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3028 if (likely(!TCP_SKB_CB(skb)->txstamp_ack))
3031 shinfo = skb_shinfo(skb);
3032 if (!before(shinfo->tskey, prior_snd_una) &&
3033 before(shinfo->tskey, tcp_sk(sk)->snd_una)) {
3034 tcp_skb_tsorted_save(skb) {
3035 __skb_tstamp_tx(skb, NULL, sk, SCM_TSTAMP_ACK);
3036 } tcp_skb_tsorted_restore(skb);
3040 /* Remove acknowledged frames from the retransmission queue. If our packet
3041 * is before the ack sequence we can discard it as it's confirmed to have
3042 * arrived at the other end.
3044 static int tcp_clean_rtx_queue(struct sock *sk, u32 prior_fack,
3046 struct tcp_sacktag_state *sack)
3048 const struct inet_connection_sock *icsk = inet_csk(sk);
3049 u64 first_ackt, last_ackt;
3050 struct tcp_sock *tp = tcp_sk(sk);
3051 u32 prior_sacked = tp->sacked_out;
3052 u32 reord = tp->snd_nxt; /* lowest acked un-retx un-sacked seq */
3053 struct sk_buff *skb, *next;
3054 bool fully_acked = true;
3055 long sack_rtt_us = -1L;
3056 long seq_rtt_us = -1L;
3057 long ca_rtt_us = -1L;
3059 u32 last_in_flight = 0;
3065 for (skb = skb_rb_first(&sk->tcp_rtx_queue); skb; skb = next) {
3066 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3067 const u32 start_seq = scb->seq;
3068 u8 sacked = scb->sacked;
3071 tcp_ack_tstamp(sk, skb, prior_snd_una);
3073 /* Determine how many packets and what bytes were acked, tso and else */
3074 if (after(scb->end_seq, tp->snd_una)) {
3075 if (tcp_skb_pcount(skb) == 1 ||
3076 !after(tp->snd_una, scb->seq))
3079 acked_pcount = tcp_tso_acked(sk, skb);
3082 fully_acked = false;
3084 acked_pcount = tcp_skb_pcount(skb);
3087 if (unlikely(sacked & TCPCB_RETRANS)) {
3088 if (sacked & TCPCB_SACKED_RETRANS)
3089 tp->retrans_out -= acked_pcount;
3090 flag |= FLAG_RETRANS_DATA_ACKED;
3091 } else if (!(sacked & TCPCB_SACKED_ACKED)) {
3092 last_ackt = skb->skb_mstamp;
3093 WARN_ON_ONCE(last_ackt == 0);
3095 first_ackt = last_ackt;
3097 last_in_flight = TCP_SKB_CB(skb)->tx.in_flight;
3098 if (before(start_seq, reord))
3100 if (!after(scb->end_seq, tp->high_seq))
3101 flag |= FLAG_ORIG_SACK_ACKED;
3104 if (sacked & TCPCB_SACKED_ACKED) {
3105 tp->sacked_out -= acked_pcount;
3106 } else if (tcp_is_sack(tp)) {
3107 tp->delivered += acked_pcount;
3108 if (!tcp_skb_spurious_retrans(tp, skb))
3109 tcp_rack_advance(tp, sacked, scb->end_seq,
3112 if (sacked & TCPCB_LOST)
3113 tp->lost_out -= acked_pcount;
3115 tp->packets_out -= acked_pcount;
3116 pkts_acked += acked_pcount;
3117 tcp_rate_skb_delivered(sk, skb, sack->rate);
3119 /* Initial outgoing SYN's get put onto the write_queue
3120 * just like anything else we transmit. It is not
3121 * true data, and if we misinform our callers that
3122 * this ACK acks real data, we will erroneously exit
3123 * connection startup slow start one packet too
3124 * quickly. This is severely frowned upon behavior.
3126 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
3127 flag |= FLAG_DATA_ACKED;
3129 flag |= FLAG_SYN_ACKED;
3130 tp->retrans_stamp = 0;
3136 next = skb_rb_next(skb);
3137 if (unlikely(skb == tp->retransmit_skb_hint))
3138 tp->retransmit_skb_hint = NULL;
3139 if (unlikely(skb == tp->lost_skb_hint))
3140 tp->lost_skb_hint = NULL;
3141 tcp_rtx_queue_unlink_and_free(skb, sk);
3145 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
3147 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3148 tp->snd_up = tp->snd_una;
3150 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3151 flag |= FLAG_SACK_RENEGING;
3153 if (likely(first_ackt) && !(flag & FLAG_RETRANS_DATA_ACKED)) {
3154 seq_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, first_ackt);
3155 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, last_ackt);
3157 if (pkts_acked == 1 && last_in_flight < tp->mss_cache &&
3158 last_in_flight && !prior_sacked && fully_acked &&
3159 sack->rate->prior_delivered + 1 == tp->delivered &&
3160 !(flag & (FLAG_CA_ALERT | FLAG_SYN_ACKED))) {
3161 /* Conservatively mark a delayed ACK. It's typically
3162 * from a lone runt packet over the round trip to
3163 * a receiver w/o out-of-order or CE events.
3165 flag |= FLAG_ACK_MAYBE_DELAYED;
3168 if (sack->first_sackt) {
3169 sack_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->first_sackt);
3170 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->last_sackt);
3172 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us,
3173 ca_rtt_us, sack->rate);
3175 if (flag & FLAG_ACKED) {
3176 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3177 if (unlikely(icsk->icsk_mtup.probe_size &&
3178 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3179 tcp_mtup_probe_success(sk);
3182 if (tcp_is_reno(tp)) {
3183 tcp_remove_reno_sacks(sk, pkts_acked);
3185 /* If any of the cumulatively ACKed segments was
3186 * retransmitted, non-SACK case cannot confirm that
3187 * progress was due to original transmission due to
3188 * lack of TCPCB_SACKED_ACKED bits even if some of
3189 * the packets may have been never retransmitted.
3191 if (flag & FLAG_RETRANS_DATA_ACKED)
3192 flag &= ~FLAG_ORIG_SACK_ACKED;
3196 /* Non-retransmitted hole got filled? That's reordering */
3197 if (before(reord, prior_fack))
3198 tcp_check_sack_reordering(sk, reord, 0);
3200 delta = prior_sacked - tp->sacked_out;
3201 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3203 } else if (skb && rtt_update && sack_rtt_us >= 0 &&
3204 sack_rtt_us > tcp_stamp_us_delta(tp->tcp_mstamp, skb->skb_mstamp)) {
3205 /* Do not re-arm RTO if the sack RTT is measured from data sent
3206 * after when the head was last (re)transmitted. Otherwise the
3207 * timeout may continue to extend in loss recovery.
3209 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */
3212 if (icsk->icsk_ca_ops->pkts_acked) {
3213 struct ack_sample sample = { .pkts_acked = pkts_acked,
3214 .rtt_us = sack->rate->rtt_us,
3215 .in_flight = last_in_flight };
3217 icsk->icsk_ca_ops->pkts_acked(sk, &sample);
3220 #if FASTRETRANS_DEBUG > 0
3221 WARN_ON((int)tp->sacked_out < 0);
3222 WARN_ON((int)tp->lost_out < 0);
3223 WARN_ON((int)tp->retrans_out < 0);
3224 if (!tp->packets_out && tcp_is_sack(tp)) {
3225 icsk = inet_csk(sk);
3227 pr_debug("Leak l=%u %d\n",
3228 tp->lost_out, icsk->icsk_ca_state);
3231 if (tp->sacked_out) {
3232 pr_debug("Leak s=%u %d\n",
3233 tp->sacked_out, icsk->icsk_ca_state);
3236 if (tp->retrans_out) {
3237 pr_debug("Leak r=%u %d\n",
3238 tp->retrans_out, icsk->icsk_ca_state);
3239 tp->retrans_out = 0;
3246 static void tcp_ack_probe(struct sock *sk)
3248 struct inet_connection_sock *icsk = inet_csk(sk);
3249 struct sk_buff *head = tcp_send_head(sk);
3250 const struct tcp_sock *tp = tcp_sk(sk);
3252 /* Was it a usable window open? */
3255 if (!after(TCP_SKB_CB(head)->end_seq, tcp_wnd_end(tp))) {
3256 icsk->icsk_backoff = 0;
3257 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3258 /* Socket must be waked up by subsequent tcp_data_snd_check().
3259 * This function is not for random using!
3262 unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX);
3264 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3269 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3271 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3272 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3275 /* Decide wheather to run the increase function of congestion control. */
3276 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3278 /* If reordering is high then always grow cwnd whenever data is
3279 * delivered regardless of its ordering. Otherwise stay conservative
3280 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3281 * new SACK or ECE mark may first advance cwnd here and later reduce
3282 * cwnd in tcp_fastretrans_alert() based on more states.
3284 if (tcp_sk(sk)->reordering > sock_net(sk)->ipv4.sysctl_tcp_reordering)
3285 return flag & FLAG_FORWARD_PROGRESS;
3287 return flag & FLAG_DATA_ACKED;
3290 /* The "ultimate" congestion control function that aims to replace the rigid
3291 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction).
3292 * It's called toward the end of processing an ACK with precise rate
3293 * information. All transmission or retransmission are delayed afterwards.
3295 static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked,
3296 int flag, const struct rate_sample *rs)
3298 const struct inet_connection_sock *icsk = inet_csk(sk);
3300 if (icsk->icsk_ca_ops->cong_control) {
3301 icsk->icsk_ca_ops->cong_control(sk, rs);
3305 if (tcp_in_cwnd_reduction(sk)) {
3306 /* Reduce cwnd if state mandates */
3307 tcp_cwnd_reduction(sk, acked_sacked, flag);
3308 } else if (tcp_may_raise_cwnd(sk, flag)) {
3309 /* Advance cwnd if state allows */
3310 tcp_cong_avoid(sk, ack, acked_sacked);
3312 tcp_update_pacing_rate(sk);
3315 /* Check that window update is acceptable.
3316 * The function assumes that snd_una<=ack<=snd_next.
3318 static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3319 const u32 ack, const u32 ack_seq,
3322 return after(ack, tp->snd_una) ||
3323 after(ack_seq, tp->snd_wl1) ||
3324 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3327 /* If we update tp->snd_una, also update tp->bytes_acked */
3328 static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack)
3330 u32 delta = ack - tp->snd_una;
3332 sock_owned_by_me((struct sock *)tp);
3333 tp->bytes_acked += delta;
3337 /* If we update tp->rcv_nxt, also update tp->bytes_received */
3338 static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq)
3340 u32 delta = seq - tp->rcv_nxt;
3342 sock_owned_by_me((struct sock *)tp);
3343 tp->bytes_received += delta;
3347 /* Update our send window.
3349 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3350 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3352 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3355 struct tcp_sock *tp = tcp_sk(sk);
3357 u32 nwin = ntohs(tcp_hdr(skb)->window);
3359 if (likely(!tcp_hdr(skb)->syn))
3360 nwin <<= tp->rx_opt.snd_wscale;
3362 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3363 flag |= FLAG_WIN_UPDATE;
3364 tcp_update_wl(tp, ack_seq);
3366 if (tp->snd_wnd != nwin) {
3369 /* Note, it is the only place, where
3370 * fast path is recovered for sending TCP.
3373 tcp_fast_path_check(sk);
3375 if (!tcp_write_queue_empty(sk))
3376 tcp_slow_start_after_idle_check(sk);
3378 if (nwin > tp->max_window) {
3379 tp->max_window = nwin;
3380 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3385 tcp_snd_una_update(tp, ack);
3390 static bool __tcp_oow_rate_limited(struct net *net, int mib_idx,
3391 u32 *last_oow_ack_time)
3393 if (*last_oow_ack_time) {
3394 s32 elapsed = (s32)(tcp_jiffies32 - *last_oow_ack_time);
3396 if (0 <= elapsed && elapsed < net->ipv4.sysctl_tcp_invalid_ratelimit) {
3397 NET_INC_STATS(net, mib_idx);
3398 return true; /* rate-limited: don't send yet! */
3402 *last_oow_ack_time = tcp_jiffies32;
3404 return false; /* not rate-limited: go ahead, send dupack now! */
3407 /* Return true if we're currently rate-limiting out-of-window ACKs and
3408 * thus shouldn't send a dupack right now. We rate-limit dupacks in
3409 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3410 * attacks that send repeated SYNs or ACKs for the same connection. To
3411 * do this, we do not send a duplicate SYNACK or ACK if the remote
3412 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3414 bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb,
3415 int mib_idx, u32 *last_oow_ack_time)
3417 /* Data packets without SYNs are not likely part of an ACK loop. */
3418 if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) &&
3422 return __tcp_oow_rate_limited(net, mib_idx, last_oow_ack_time);
3425 /* RFC 5961 7 [ACK Throttling] */
3426 static void tcp_send_challenge_ack(struct sock *sk, const struct sk_buff *skb)
3428 /* unprotected vars, we dont care of overwrites */
3429 static u32 challenge_timestamp;
3430 static unsigned int challenge_count;
3431 struct tcp_sock *tp = tcp_sk(sk);
3432 struct net *net = sock_net(sk);
3435 /* First check our per-socket dupack rate limit. */
3436 if (__tcp_oow_rate_limited(net,
3437 LINUX_MIB_TCPACKSKIPPEDCHALLENGE,
3438 &tp->last_oow_ack_time))
3441 /* Then check host-wide RFC 5961 rate limit. */
3443 if (now != challenge_timestamp) {
3444 u32 ack_limit = net->ipv4.sysctl_tcp_challenge_ack_limit;
3445 u32 half = (ack_limit + 1) >> 1;
3447 challenge_timestamp = now;
3448 WRITE_ONCE(challenge_count, half + prandom_u32_max(ack_limit));
3450 count = READ_ONCE(challenge_count);
3452 WRITE_ONCE(challenge_count, count - 1);
3453 NET_INC_STATS(net, LINUX_MIB_TCPCHALLENGEACK);
3458 static void tcp_store_ts_recent(struct tcp_sock *tp)
3460 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3461 tp->rx_opt.ts_recent_stamp = get_seconds();
3464 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3466 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3467 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3468 * extra check below makes sure this can only happen
3469 * for pure ACK frames. -DaveM
3471 * Not only, also it occurs for expired timestamps.
3474 if (tcp_paws_check(&tp->rx_opt, 0))
3475 tcp_store_ts_recent(tp);
3479 /* This routine deals with acks during a TLP episode.
3480 * We mark the end of a TLP episode on receiving TLP dupack or when
3481 * ack is after tlp_high_seq.
3482 * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3484 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3486 struct tcp_sock *tp = tcp_sk(sk);
3488 if (before(ack, tp->tlp_high_seq))
3491 if (flag & FLAG_DSACKING_ACK) {
3492 /* This DSACK means original and TLP probe arrived; no loss */
3493 tp->tlp_high_seq = 0;
3494 } else if (after(ack, tp->tlp_high_seq)) {
3495 /* ACK advances: there was a loss, so reduce cwnd. Reset
3496 * tlp_high_seq in tcp_init_cwnd_reduction()
3498 tcp_init_cwnd_reduction(sk);
3499 tcp_set_ca_state(sk, TCP_CA_CWR);
3500 tcp_end_cwnd_reduction(sk);
3501 tcp_try_keep_open(sk);
3502 NET_INC_STATS(sock_net(sk),
3503 LINUX_MIB_TCPLOSSPROBERECOVERY);
3504 } else if (!(flag & (FLAG_SND_UNA_ADVANCED |
3505 FLAG_NOT_DUP | FLAG_DATA_SACKED))) {
3506 /* Pure dupack: original and TLP probe arrived; no loss */
3507 tp->tlp_high_seq = 0;
3511 static inline void tcp_in_ack_event(struct sock *sk, u32 flags)
3513 const struct inet_connection_sock *icsk = inet_csk(sk);
3515 if (icsk->icsk_ca_ops->in_ack_event)
3516 icsk->icsk_ca_ops->in_ack_event(sk, flags);
3519 /* Congestion control has updated the cwnd already. So if we're in
3520 * loss recovery then now we do any new sends (for FRTO) or
3521 * retransmits (for CA_Loss or CA_recovery) that make sense.
3523 static void tcp_xmit_recovery(struct sock *sk, int rexmit)
3525 struct tcp_sock *tp = tcp_sk(sk);
3527 if (rexmit == REXMIT_NONE)
3530 if (unlikely(rexmit == 2)) {
3531 __tcp_push_pending_frames(sk, tcp_current_mss(sk),
3533 if (after(tp->snd_nxt, tp->high_seq))
3537 tcp_xmit_retransmit_queue(sk);
3540 /* Returns the number of packets newly acked or sacked by the current ACK */
3541 static u32 tcp_newly_delivered(struct sock *sk, u32 prior_delivered, int flag)
3543 const struct net *net = sock_net(sk);
3544 struct tcp_sock *tp = tcp_sk(sk);
3547 delivered = tp->delivered - prior_delivered;
3548 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVERED, delivered);
3549 if (flag & FLAG_ECE) {
3550 tp->delivered_ce += delivered;
3551 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVEREDCE, delivered);
3556 /* This routine deals with incoming acks, but not outgoing ones. */
3557 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3559 struct inet_connection_sock *icsk = inet_csk(sk);
3560 struct tcp_sock *tp = tcp_sk(sk);
3561 struct tcp_sacktag_state sack_state;
3562 struct rate_sample rs = { .prior_delivered = 0 };
3563 u32 prior_snd_una = tp->snd_una;
3564 bool is_sack_reneg = tp->is_sack_reneg;
3565 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3566 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3567 bool is_dupack = false;
3568 int prior_packets = tp->packets_out;
3569 u32 delivered = tp->delivered;
3570 u32 lost = tp->lost;
3571 int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */
3574 sack_state.first_sackt = 0;
3575 sack_state.rate = &rs;
3577 /* We very likely will need to access rtx queue. */
3578 prefetch(sk->tcp_rtx_queue.rb_node);
3580 /* If the ack is older than previous acks
3581 * then we can probably ignore it.
3583 if (before(ack, prior_snd_una)) {
3584 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3585 if (before(ack, prior_snd_una - tp->max_window)) {
3586 if (!(flag & FLAG_NO_CHALLENGE_ACK))
3587 tcp_send_challenge_ack(sk, skb);
3593 /* If the ack includes data we haven't sent yet, discard
3594 * this segment (RFC793 Section 3.9).
3596 if (after(ack, tp->snd_nxt))
3599 if (after(ack, prior_snd_una)) {
3600 flag |= FLAG_SND_UNA_ADVANCED;
3601 icsk->icsk_retransmits = 0;
3603 #if IS_ENABLED(CONFIG_TLS_DEVICE)
3604 if (static_branch_unlikely(&clean_acked_data_enabled))
3605 if (icsk->icsk_clean_acked)
3606 icsk->icsk_clean_acked(sk, ack);
3610 prior_fack = tcp_is_sack(tp) ? tcp_highest_sack_seq(tp) : tp->snd_una;
3611 rs.prior_in_flight = tcp_packets_in_flight(tp);
3613 /* ts_recent update must be made after we are sure that the packet
3616 if (flag & FLAG_UPDATE_TS_RECENT)
3617 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3619 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3620 /* Window is constant, pure forward advance.
3621 * No more checks are required.
3622 * Note, we use the fact that SND.UNA>=SND.WL2.
3624 tcp_update_wl(tp, ack_seq);
3625 tcp_snd_una_update(tp, ack);
3626 flag |= FLAG_WIN_UPDATE;
3628 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
3630 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPACKS);
3632 u32 ack_ev_flags = CA_ACK_SLOWPATH;
3634 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3637 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3639 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3641 if (TCP_SKB_CB(skb)->sacked)
3642 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3645 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) {
3647 ack_ev_flags |= CA_ACK_ECE;
3650 if (flag & FLAG_WIN_UPDATE)
3651 ack_ev_flags |= CA_ACK_WIN_UPDATE;
3653 tcp_in_ack_event(sk, ack_ev_flags);
3656 /* We passed data and got it acked, remove any soft error
3657 * log. Something worked...
3659 sk->sk_err_soft = 0;
3660 icsk->icsk_probes_out = 0;
3661 tp->rcv_tstamp = tcp_jiffies32;
3665 /* See if we can take anything off of the retransmit queue. */
3666 flag |= tcp_clean_rtx_queue(sk, prior_fack, prior_snd_una, &sack_state);
3668 tcp_rack_update_reo_wnd(sk, &rs);
3670 if (tp->tlp_high_seq)
3671 tcp_process_tlp_ack(sk, ack, flag);
3672 /* If needed, reset TLP/RTO timer; RACK may later override this. */
3673 if (flag & FLAG_SET_XMIT_TIMER)
3674 tcp_set_xmit_timer(sk);
3676 if (tcp_ack_is_dubious(sk, flag)) {
3677 is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
3678 tcp_fastretrans_alert(sk, prior_snd_una, is_dupack, &flag,
3682 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP))
3685 delivered = tcp_newly_delivered(sk, delivered, flag);
3686 lost = tp->lost - lost; /* freshly marked lost */
3687 rs.is_ack_delayed = !!(flag & FLAG_ACK_MAYBE_DELAYED);
3688 tcp_rate_gen(sk, delivered, lost, is_sack_reneg, sack_state.rate);
3689 tcp_cong_control(sk, ack, delivered, flag, sack_state.rate);
3690 tcp_xmit_recovery(sk, rexmit);
3694 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3695 if (flag & FLAG_DSACKING_ACK) {
3696 tcp_fastretrans_alert(sk, prior_snd_una, is_dupack, &flag,
3698 tcp_newly_delivered(sk, delivered, flag);
3700 /* If this ack opens up a zero window, clear backoff. It was
3701 * being used to time the probes, and is probably far higher than
3702 * it needs to be for normal retransmission.
3706 if (tp->tlp_high_seq)
3707 tcp_process_tlp_ack(sk, ack, flag);
3711 SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3715 /* If data was SACKed, tag it and see if we should send more data.
3716 * If data was DSACKed, see if we can undo a cwnd reduction.
3718 if (TCP_SKB_CB(skb)->sacked) {
3719 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3721 tcp_fastretrans_alert(sk, prior_snd_una, is_dupack, &flag,
3723 tcp_newly_delivered(sk, delivered, flag);
3724 tcp_xmit_recovery(sk, rexmit);
3727 SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3731 static void tcp_parse_fastopen_option(int len, const unsigned char *cookie,
3732 bool syn, struct tcp_fastopen_cookie *foc,
3735 /* Valid only in SYN or SYN-ACK with an even length. */
3736 if (!foc || !syn || len < 0 || (len & 1))
3739 if (len >= TCP_FASTOPEN_COOKIE_MIN &&
3740 len <= TCP_FASTOPEN_COOKIE_MAX)
3741 memcpy(foc->val, cookie, len);
3748 static void smc_parse_options(const struct tcphdr *th,
3749 struct tcp_options_received *opt_rx,
3750 const unsigned char *ptr,
3753 #if IS_ENABLED(CONFIG_SMC)
3754 if (static_branch_unlikely(&tcp_have_smc)) {
3755 if (th->syn && !(opsize & 1) &&
3756 opsize >= TCPOLEN_EXP_SMC_BASE &&
3757 get_unaligned_be32(ptr) == TCPOPT_SMC_MAGIC)
3763 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3764 * But, this can also be called on packets in the established flow when
3765 * the fast version below fails.
3767 void tcp_parse_options(const struct net *net,
3768 const struct sk_buff *skb,
3769 struct tcp_options_received *opt_rx, int estab,
3770 struct tcp_fastopen_cookie *foc)
3772 const unsigned char *ptr;
3773 const struct tcphdr *th = tcp_hdr(skb);
3774 int length = (th->doff * 4) - sizeof(struct tcphdr);
3776 ptr = (const unsigned char *)(th + 1);
3777 opt_rx->saw_tstamp = 0;
3779 while (length > 0) {
3780 int opcode = *ptr++;
3786 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3791 if (opsize < 2) /* "silly options" */
3793 if (opsize > length)
3794 return; /* don't parse partial options */
3797 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3798 u16 in_mss = get_unaligned_be16(ptr);
3800 if (opt_rx->user_mss &&
3801 opt_rx->user_mss < in_mss)
3802 in_mss = opt_rx->user_mss;
3803 opt_rx->mss_clamp = in_mss;
3808 if (opsize == TCPOLEN_WINDOW && th->syn &&
3809 !estab && net->ipv4.sysctl_tcp_window_scaling) {
3810 __u8 snd_wscale = *(__u8 *)ptr;
3811 opt_rx->wscale_ok = 1;
3812 if (snd_wscale > TCP_MAX_WSCALE) {
3813 net_info_ratelimited("%s: Illegal window scaling value %d > %u received\n",
3817 snd_wscale = TCP_MAX_WSCALE;
3819 opt_rx->snd_wscale = snd_wscale;
3822 case TCPOPT_TIMESTAMP:
3823 if ((opsize == TCPOLEN_TIMESTAMP) &&
3824 ((estab && opt_rx->tstamp_ok) ||
3825 (!estab && net->ipv4.sysctl_tcp_timestamps))) {
3826 opt_rx->saw_tstamp = 1;
3827 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3828 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3831 case TCPOPT_SACK_PERM:
3832 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3833 !estab && net->ipv4.sysctl_tcp_sack) {
3834 opt_rx->sack_ok = TCP_SACK_SEEN;
3835 tcp_sack_reset(opt_rx);
3840 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3841 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3843 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3846 #ifdef CONFIG_TCP_MD5SIG
3849 * The MD5 Hash has already been
3850 * checked (see tcp_v{4,6}_do_rcv()).
3854 case TCPOPT_FASTOPEN:
3855 tcp_parse_fastopen_option(
3856 opsize - TCPOLEN_FASTOPEN_BASE,
3857 ptr, th->syn, foc, false);
3861 /* Fast Open option shares code 254 using a
3862 * 16 bits magic number.
3864 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE &&
3865 get_unaligned_be16(ptr) ==
3866 TCPOPT_FASTOPEN_MAGIC)
3867 tcp_parse_fastopen_option(opsize -
3868 TCPOLEN_EXP_FASTOPEN_BASE,
3869 ptr + 2, th->syn, foc, true);
3871 smc_parse_options(th, opt_rx, ptr,
3881 EXPORT_SYMBOL(tcp_parse_options);
3883 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
3885 const __be32 *ptr = (const __be32 *)(th + 1);
3887 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3888 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3889 tp->rx_opt.saw_tstamp = 1;
3891 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3894 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
3896 tp->rx_opt.rcv_tsecr = 0;
3902 /* Fast parse options. This hopes to only see timestamps.
3903 * If it is wrong it falls back on tcp_parse_options().
3905 static bool tcp_fast_parse_options(const struct net *net,
3906 const struct sk_buff *skb,
3907 const struct tcphdr *th, struct tcp_sock *tp)
3909 /* In the spirit of fast parsing, compare doff directly to constant
3910 * values. Because equality is used, short doff can be ignored here.
3912 if (th->doff == (sizeof(*th) / 4)) {
3913 tp->rx_opt.saw_tstamp = 0;
3915 } else if (tp->rx_opt.tstamp_ok &&
3916 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3917 if (tcp_parse_aligned_timestamp(tp, th))
3921 tcp_parse_options(net, skb, &tp->rx_opt, 1, NULL);
3922 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3923 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
3928 #ifdef CONFIG_TCP_MD5SIG
3930 * Parse MD5 Signature option
3932 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
3934 int length = (th->doff << 2) - sizeof(*th);
3935 const u8 *ptr = (const u8 *)(th + 1);
3937 /* If not enough data remaining, we can short cut */
3938 while (length >= TCPOLEN_MD5SIG) {
3939 int opcode = *ptr++;
3950 if (opsize < 2 || opsize > length)
3952 if (opcode == TCPOPT_MD5SIG)
3953 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
3960 EXPORT_SYMBOL(tcp_parse_md5sig_option);
3963 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3965 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3966 * it can pass through stack. So, the following predicate verifies that
3967 * this segment is not used for anything but congestion avoidance or
3968 * fast retransmit. Moreover, we even are able to eliminate most of such
3969 * second order effects, if we apply some small "replay" window (~RTO)
3970 * to timestamp space.
3972 * All these measures still do not guarantee that we reject wrapped ACKs
3973 * on networks with high bandwidth, when sequence space is recycled fastly,
3974 * but it guarantees that such events will be very rare and do not affect
3975 * connection seriously. This doesn't look nice, but alas, PAWS is really
3978 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3979 * states that events when retransmit arrives after original data are rare.
3980 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3981 * the biggest problem on large power networks even with minor reordering.
3982 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3983 * up to bandwidth of 18Gigabit/sec. 8) ]
3986 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3988 const struct tcp_sock *tp = tcp_sk(sk);
3989 const struct tcphdr *th = tcp_hdr(skb);
3990 u32 seq = TCP_SKB_CB(skb)->seq;
3991 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3993 return (/* 1. Pure ACK with correct sequence number. */
3994 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3996 /* 2. ... and duplicate ACK. */
3997 ack == tp->snd_una &&
3999 /* 3. ... and does not update window. */
4000 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
4002 /* 4. ... and sits in replay window. */
4003 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
4006 static inline bool tcp_paws_discard(const struct sock *sk,
4007 const struct sk_buff *skb)
4009 const struct tcp_sock *tp = tcp_sk(sk);
4011 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
4012 !tcp_disordered_ack(sk, skb);
4015 /* Check segment sequence number for validity.
4017 * Segment controls are considered valid, if the segment
4018 * fits to the window after truncation to the window. Acceptability
4019 * of data (and SYN, FIN, of course) is checked separately.
4020 * See tcp_data_queue(), for example.
4022 * Also, controls (RST is main one) are accepted using RCV.WUP instead
4023 * of RCV.NXT. Peer still did not advance his SND.UNA when we
4024 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
4025 * (borrowed from freebsd)
4028 static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
4030 return !before(end_seq, tp->rcv_wup) &&
4031 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
4034 /* When we get a reset we do this. */
4035 void tcp_reset(struct sock *sk)
4037 trace_tcp_receive_reset(sk);
4039 /* We want the right error as BSD sees it (and indeed as we do). */
4040 switch (sk->sk_state) {
4042 sk->sk_err = ECONNREFUSED;
4044 case TCP_CLOSE_WAIT:
4050 sk->sk_err = ECONNRESET;
4052 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4055 tcp_write_queue_purge(sk);
4058 if (!sock_flag(sk, SOCK_DEAD))
4059 sk->sk_error_report(sk);
4063 * Process the FIN bit. This now behaves as it is supposed to work
4064 * and the FIN takes effect when it is validly part of sequence
4065 * space. Not before when we get holes.
4067 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4068 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4071 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4072 * close and we go into CLOSING (and later onto TIME-WAIT)
4074 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4076 void tcp_fin(struct sock *sk)
4078 struct tcp_sock *tp = tcp_sk(sk);
4080 inet_csk_schedule_ack(sk);
4082 sk->sk_shutdown |= RCV_SHUTDOWN;
4083 sock_set_flag(sk, SOCK_DONE);
4085 switch (sk->sk_state) {
4087 case TCP_ESTABLISHED:
4088 /* Move to CLOSE_WAIT */
4089 tcp_set_state(sk, TCP_CLOSE_WAIT);
4090 inet_csk(sk)->icsk_ack.pingpong = 1;
4093 case TCP_CLOSE_WAIT:
4095 /* Received a retransmission of the FIN, do
4100 /* RFC793: Remain in the LAST-ACK state. */
4104 /* This case occurs when a simultaneous close
4105 * happens, we must ack the received FIN and
4106 * enter the CLOSING state.
4109 tcp_set_state(sk, TCP_CLOSING);
4112 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4114 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4117 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4118 * cases we should never reach this piece of code.
4120 pr_err("%s: Impossible, sk->sk_state=%d\n",
4121 __func__, sk->sk_state);
4125 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4126 * Probably, we should reset in this case. For now drop them.
4128 skb_rbtree_purge(&tp->out_of_order_queue);
4129 if (tcp_is_sack(tp))
4130 tcp_sack_reset(&tp->rx_opt);
4133 if (!sock_flag(sk, SOCK_DEAD)) {
4134 sk->sk_state_change(sk);
4136 /* Do not send POLL_HUP for half duplex close. */
4137 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4138 sk->sk_state == TCP_CLOSE)
4139 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4141 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4145 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4148 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4149 if (before(seq, sp->start_seq))
4150 sp->start_seq = seq;
4151 if (after(end_seq, sp->end_seq))
4152 sp->end_seq = end_seq;
4158 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4160 struct tcp_sock *tp = tcp_sk(sk);
4162 if (tcp_is_sack(tp) && sock_net(sk)->ipv4.sysctl_tcp_dsack) {
4165 if (before(seq, tp->rcv_nxt))
4166 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4168 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4170 NET_INC_STATS(sock_net(sk), mib_idx);
4172 tp->rx_opt.dsack = 1;
4173 tp->duplicate_sack[0].start_seq = seq;
4174 tp->duplicate_sack[0].end_seq = end_seq;
4178 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4180 struct tcp_sock *tp = tcp_sk(sk);
4182 if (!tp->rx_opt.dsack)
4183 tcp_dsack_set(sk, seq, end_seq);
4185 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4188 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4190 struct tcp_sock *tp = tcp_sk(sk);
4192 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4193 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4194 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4195 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
4197 if (tcp_is_sack(tp) && sock_net(sk)->ipv4.sysctl_tcp_dsack) {
4198 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4200 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4201 end_seq = tp->rcv_nxt;
4202 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4209 /* These routines update the SACK block as out-of-order packets arrive or
4210 * in-order packets close up the sequence space.
4212 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4215 struct tcp_sack_block *sp = &tp->selective_acks[0];
4216 struct tcp_sack_block *swalk = sp + 1;
4218 /* See if the recent change to the first SACK eats into
4219 * or hits the sequence space of other SACK blocks, if so coalesce.
4221 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4222 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4225 /* Zap SWALK, by moving every further SACK up by one slot.
4226 * Decrease num_sacks.
4228 tp->rx_opt.num_sacks--;
4229 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4233 this_sack++, swalk++;
4237 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4239 struct tcp_sock *tp = tcp_sk(sk);
4240 struct tcp_sack_block *sp = &tp->selective_acks[0];
4241 int cur_sacks = tp->rx_opt.num_sacks;
4247 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4248 if (tcp_sack_extend(sp, seq, end_seq)) {
4249 /* Rotate this_sack to the first one. */
4250 for (; this_sack > 0; this_sack--, sp--)
4251 swap(*sp, *(sp - 1));
4253 tcp_sack_maybe_coalesce(tp);
4258 /* Could not find an adjacent existing SACK, build a new one,
4259 * put it at the front, and shift everyone else down. We
4260 * always know there is at least one SACK present already here.
4262 * If the sack array is full, forget about the last one.
4264 if (this_sack >= TCP_NUM_SACKS) {
4265 if (tp->compressed_ack)
4268 tp->rx_opt.num_sacks--;
4271 for (; this_sack > 0; this_sack--, sp--)
4275 /* Build the new head SACK, and we're done. */
4276 sp->start_seq = seq;
4277 sp->end_seq = end_seq;
4278 tp->rx_opt.num_sacks++;
4281 /* RCV.NXT advances, some SACKs should be eaten. */
4283 static void tcp_sack_remove(struct tcp_sock *tp)
4285 struct tcp_sack_block *sp = &tp->selective_acks[0];
4286 int num_sacks = tp->rx_opt.num_sacks;
4289 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4290 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4291 tp->rx_opt.num_sacks = 0;
4295 for (this_sack = 0; this_sack < num_sacks;) {
4296 /* Check if the start of the sack is covered by RCV.NXT. */
4297 if (!before(tp->rcv_nxt, sp->start_seq)) {
4300 /* RCV.NXT must cover all the block! */
4301 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4303 /* Zap this SACK, by moving forward any other SACKS. */
4304 for (i = this_sack+1; i < num_sacks; i++)
4305 tp->selective_acks[i-1] = tp->selective_acks[i];
4312 tp->rx_opt.num_sacks = num_sacks;
4316 * tcp_try_coalesce - try to merge skb to prior one
4318 * @dest: destination queue
4320 * @from: buffer to add in queue
4321 * @fragstolen: pointer to boolean
4323 * Before queueing skb @from after @to, try to merge them
4324 * to reduce overall memory use and queue lengths, if cost is small.
4325 * Packets in ofo or receive queues can stay a long time.
4326 * Better try to coalesce them right now to avoid future collapses.
4327 * Returns true if caller should free @from instead of queueing it
4329 static bool tcp_try_coalesce(struct sock *sk,
4331 struct sk_buff *from,
4336 *fragstolen = false;
4338 /* Its possible this segment overlaps with prior segment in queue */
4339 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4342 if (!skb_try_coalesce(to, from, fragstolen, &delta))
4345 atomic_add(delta, &sk->sk_rmem_alloc);
4346 sk_mem_charge(sk, delta);
4347 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4348 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4349 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4350 TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
4352 if (TCP_SKB_CB(from)->has_rxtstamp) {
4353 TCP_SKB_CB(to)->has_rxtstamp = true;
4354 to->tstamp = from->tstamp;
4360 static void tcp_drop(struct sock *sk, struct sk_buff *skb)
4362 sk_drops_add(sk, skb);
4366 /* This one checks to see if we can put data from the
4367 * out_of_order queue into the receive_queue.
4369 static void tcp_ofo_queue(struct sock *sk)
4371 struct tcp_sock *tp = tcp_sk(sk);
4372 __u32 dsack_high = tp->rcv_nxt;
4373 bool fin, fragstolen, eaten;
4374 struct sk_buff *skb, *tail;
4377 p = rb_first(&tp->out_of_order_queue);
4380 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4383 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4384 __u32 dsack = dsack_high;
4385 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4386 dsack_high = TCP_SKB_CB(skb)->end_seq;
4387 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4390 rb_erase(&skb->rbnode, &tp->out_of_order_queue);
4392 if (unlikely(!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))) {
4393 SOCK_DEBUG(sk, "ofo packet was already received\n");
4397 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4398 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4399 TCP_SKB_CB(skb)->end_seq);
4401 tail = skb_peek_tail(&sk->sk_receive_queue);
4402 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen);
4403 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4404 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
4406 __skb_queue_tail(&sk->sk_receive_queue, skb);
4408 kfree_skb_partial(skb, fragstolen);
4410 if (unlikely(fin)) {
4412 /* tcp_fin() purges tp->out_of_order_queue,
4413 * so we must end this loop right now.
4420 static bool tcp_prune_ofo_queue(struct sock *sk);
4421 static int tcp_prune_queue(struct sock *sk);
4423 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4426 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4427 !sk_rmem_schedule(sk, skb, size)) {
4429 if (tcp_prune_queue(sk) < 0)
4432 while (!sk_rmem_schedule(sk, skb, size)) {
4433 if (!tcp_prune_ofo_queue(sk))
4440 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4442 struct tcp_sock *tp = tcp_sk(sk);
4443 struct rb_node **p, *parent;
4444 struct sk_buff *skb1;
4448 tcp_ecn_check_ce(sk, skb);
4450 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4451 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFODROP);
4456 /* Disable header prediction. */
4458 inet_csk_schedule_ack(sk);
4460 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4461 seq = TCP_SKB_CB(skb)->seq;
4462 end_seq = TCP_SKB_CB(skb)->end_seq;
4463 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4464 tp->rcv_nxt, seq, end_seq);
4466 p = &tp->out_of_order_queue.rb_node;
4467 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4468 /* Initial out of order segment, build 1 SACK. */
4469 if (tcp_is_sack(tp)) {
4470 tp->rx_opt.num_sacks = 1;
4471 tp->selective_acks[0].start_seq = seq;
4472 tp->selective_acks[0].end_seq = end_seq;
4474 rb_link_node(&skb->rbnode, NULL, p);
4475 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4476 tp->ooo_last_skb = skb;
4480 /* In the typical case, we are adding an skb to the end of the list.
4481 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
4483 if (tcp_try_coalesce(sk, tp->ooo_last_skb,
4484 skb, &fragstolen)) {
4486 tcp_grow_window(sk, skb);
4487 kfree_skb_partial(skb, fragstolen);
4491 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
4492 if (!before(seq, TCP_SKB_CB(tp->ooo_last_skb)->end_seq)) {
4493 parent = &tp->ooo_last_skb->rbnode;
4494 p = &parent->rb_right;
4498 /* Find place to insert this segment. Handle overlaps on the way. */
4502 skb1 = rb_to_skb(parent);
4503 if (before(seq, TCP_SKB_CB(skb1)->seq)) {
4504 p = &parent->rb_left;
4507 if (before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4508 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4509 /* All the bits are present. Drop. */
4510 NET_INC_STATS(sock_net(sk),
4511 LINUX_MIB_TCPOFOMERGE);
4514 tcp_dsack_set(sk, seq, end_seq);
4517 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4518 /* Partial overlap. */
4519 tcp_dsack_set(sk, seq, TCP_SKB_CB(skb1)->end_seq);
4521 /* skb's seq == skb1's seq and skb covers skb1.
4522 * Replace skb1 with skb.
4524 rb_replace_node(&skb1->rbnode, &skb->rbnode,
4525 &tp->out_of_order_queue);
4526 tcp_dsack_extend(sk,
4527 TCP_SKB_CB(skb1)->seq,
4528 TCP_SKB_CB(skb1)->end_seq);
4529 NET_INC_STATS(sock_net(sk),
4530 LINUX_MIB_TCPOFOMERGE);
4534 } else if (tcp_try_coalesce(sk, skb1,
4535 skb, &fragstolen)) {
4538 p = &parent->rb_right;
4541 /* Insert segment into RB tree. */
4542 rb_link_node(&skb->rbnode, parent, p);
4543 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4546 /* Remove other segments covered by skb. */
4547 while ((skb1 = skb_rb_next(skb)) != NULL) {
4548 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4550 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4551 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4555 rb_erase(&skb1->rbnode, &tp->out_of_order_queue);
4556 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4557 TCP_SKB_CB(skb1)->end_seq);
4558 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4561 /* If there is no skb after us, we are the last_skb ! */
4563 tp->ooo_last_skb = skb;
4566 if (tcp_is_sack(tp))
4567 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4570 tcp_grow_window(sk, skb);
4572 skb_set_owner_r(skb, sk);
4576 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, int hdrlen,
4580 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4582 __skb_pull(skb, hdrlen);
4584 tcp_try_coalesce(sk, tail,
4585 skb, fragstolen)) ? 1 : 0;
4586 tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq);
4588 __skb_queue_tail(&sk->sk_receive_queue, skb);
4589 skb_set_owner_r(skb, sk);
4594 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4596 struct sk_buff *skb;
4604 if (size > PAGE_SIZE) {
4605 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS);
4607 data_len = npages << PAGE_SHIFT;
4608 size = data_len + (size & ~PAGE_MASK);
4610 skb = alloc_skb_with_frags(size - data_len, data_len,
4611 PAGE_ALLOC_COSTLY_ORDER,
4612 &err, sk->sk_allocation);
4616 skb_put(skb, size - data_len);
4617 skb->data_len = data_len;
4620 if (tcp_try_rmem_schedule(sk, skb, skb->truesize))
4623 err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size);
4627 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4628 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4629 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4631 if (tcp_queue_rcv(sk, skb, 0, &fragstolen)) {
4632 WARN_ON_ONCE(fragstolen); /* should not happen */
4644 void tcp_data_ready(struct sock *sk)
4646 const struct tcp_sock *tp = tcp_sk(sk);
4647 int avail = tp->rcv_nxt - tp->copied_seq;
4649 if (avail < sk->sk_rcvlowat && !sock_flag(sk, SOCK_DONE))
4652 sk->sk_data_ready(sk);
4655 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4657 struct tcp_sock *tp = tcp_sk(sk);
4661 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
4666 __skb_pull(skb, tcp_hdr(skb)->doff * 4);
4668 tcp_ecn_accept_cwr(tp, skb);
4670 tp->rx_opt.dsack = 0;
4672 /* Queue data for delivery to the user.
4673 * Packets in sequence go to the receive queue.
4674 * Out of sequence packets to the out_of_order_queue.
4676 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4677 if (tcp_receive_window(tp) == 0)
4680 /* Ok. In sequence. In window. */
4682 if (skb_queue_len(&sk->sk_receive_queue) == 0)
4683 sk_forced_mem_schedule(sk, skb->truesize);
4684 else if (tcp_try_rmem_schedule(sk, skb, skb->truesize))
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))
4749 tcp_data_queue_ofo(sk, skb);
4752 static struct sk_buff *tcp_skb_next(struct sk_buff *skb, struct sk_buff_head *list)
4755 return !skb_queue_is_last(list, skb) ? skb->next : NULL;
4757 return skb_rb_next(skb);
4760 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4761 struct sk_buff_head *list,
4762 struct rb_root *root)
4764 struct sk_buff *next = tcp_skb_next(skb, list);
4767 __skb_unlink(skb, list);
4769 rb_erase(&skb->rbnode, root);
4772 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4777 /* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */
4778 void tcp_rbtree_insert(struct rb_root *root, struct sk_buff *skb)
4780 struct rb_node **p = &root->rb_node;
4781 struct rb_node *parent = NULL;
4782 struct sk_buff *skb1;
4786 skb1 = rb_to_skb(parent);
4787 if (before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb1)->seq))
4788 p = &parent->rb_left;
4790 p = &parent->rb_right;
4792 rb_link_node(&skb->rbnode, parent, p);
4793 rb_insert_color(&skb->rbnode, root);
4796 /* Collapse contiguous sequence of skbs head..tail with
4797 * sequence numbers start..end.
4799 * If tail is NULL, this means until the end of the queue.
4801 * Segments with FIN/SYN are not collapsed (only because this
4805 tcp_collapse(struct sock *sk, struct sk_buff_head *list, struct rb_root *root,
4806 struct sk_buff *head, struct sk_buff *tail, u32 start, u32 end)
4808 struct sk_buff *skb = head, *n;
4809 struct sk_buff_head tmp;
4812 /* First, check that queue is collapsible and find
4813 * the point where collapsing can be useful.
4816 for (end_of_skbs = true; skb != NULL && skb != tail; skb = n) {
4817 n = tcp_skb_next(skb, list);
4819 /* No new bits? It is possible on ofo queue. */
4820 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4821 skb = tcp_collapse_one(sk, skb, list, root);
4827 /* The first skb to collapse is:
4829 * - bloated or contains data before "start" or
4830 * overlaps to the next one.
4832 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
4833 (tcp_win_from_space(sk, skb->truesize) > skb->len ||
4834 before(TCP_SKB_CB(skb)->seq, start))) {
4835 end_of_skbs = false;
4839 if (n && n != tail &&
4840 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(n)->seq) {
4841 end_of_skbs = false;
4845 /* Decided to skip this, advance start seq. */
4846 start = TCP_SKB_CB(skb)->end_seq;
4849 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4852 __skb_queue_head_init(&tmp);
4854 while (before(start, end)) {
4855 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
4856 struct sk_buff *nskb;
4858 nskb = alloc_skb(copy, GFP_ATOMIC);
4862 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4863 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4865 __skb_queue_before(list, skb, nskb);
4867 __skb_queue_tail(&tmp, nskb); /* defer rbtree insertion */
4868 skb_set_owner_r(nskb, sk);
4870 /* Copy data, releasing collapsed skbs. */
4872 int offset = start - TCP_SKB_CB(skb)->seq;
4873 int size = TCP_SKB_CB(skb)->end_seq - start;
4877 size = min(copy, size);
4878 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4880 TCP_SKB_CB(nskb)->end_seq += size;
4884 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4885 skb = tcp_collapse_one(sk, skb, list, root);
4888 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4894 skb_queue_walk_safe(&tmp, skb, n)
4895 tcp_rbtree_insert(root, skb);
4898 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4899 * and tcp_collapse() them until all the queue is collapsed.
4901 static void tcp_collapse_ofo_queue(struct sock *sk)
4903 struct tcp_sock *tp = tcp_sk(sk);
4904 struct sk_buff *skb, *head;
4907 skb = skb_rb_first(&tp->out_of_order_queue);
4910 tp->ooo_last_skb = skb_rb_last(&tp->out_of_order_queue);
4913 start = TCP_SKB_CB(skb)->seq;
4914 end = TCP_SKB_CB(skb)->end_seq;
4916 for (head = skb;;) {
4917 skb = skb_rb_next(skb);
4919 /* Range is terminated when we see a gap or when
4920 * we are at the queue end.
4923 after(TCP_SKB_CB(skb)->seq, end) ||
4924 before(TCP_SKB_CB(skb)->end_seq, start)) {
4925 tcp_collapse(sk, NULL, &tp->out_of_order_queue,
4926 head, skb, start, end);
4930 if (unlikely(before(TCP_SKB_CB(skb)->seq, start)))
4931 start = TCP_SKB_CB(skb)->seq;
4932 if (after(TCP_SKB_CB(skb)->end_seq, end))
4933 end = TCP_SKB_CB(skb)->end_seq;
4938 * Clean the out-of-order queue to make room.
4939 * We drop high sequences packets to :
4940 * 1) Let a chance for holes to be filled.
4941 * 2) not add too big latencies if thousands of packets sit there.
4942 * (But if application shrinks SO_RCVBUF, we could still end up
4943 * freeing whole queue here)
4945 * Return true if queue has shrunk.
4947 static bool tcp_prune_ofo_queue(struct sock *sk)
4949 struct tcp_sock *tp = tcp_sk(sk);
4950 struct rb_node *node, *prev;
4952 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
4955 NET_INC_STATS(sock_net(sk), LINUX_MIB_OFOPRUNED);
4956 node = &tp->ooo_last_skb->rbnode;
4958 prev = rb_prev(node);
4959 rb_erase(node, &tp->out_of_order_queue);
4960 tcp_drop(sk, rb_to_skb(node));
4962 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf &&
4963 !tcp_under_memory_pressure(sk))
4967 tp->ooo_last_skb = rb_to_skb(prev);
4969 /* Reset SACK state. A conforming SACK implementation will
4970 * do the same at a timeout based retransmit. When a connection
4971 * is in a sad state like this, we care only about integrity
4972 * of the connection not performance.
4974 if (tp->rx_opt.sack_ok)
4975 tcp_sack_reset(&tp->rx_opt);
4979 /* Reduce allocated memory if we can, trying to get
4980 * the socket within its memory limits again.
4982 * Return less than zero if we should start dropping frames
4983 * until the socket owning process reads some of the data
4984 * to stabilize the situation.
4986 static int tcp_prune_queue(struct sock *sk)
4988 struct tcp_sock *tp = tcp_sk(sk);
4990 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4992 NET_INC_STATS(sock_net(sk), LINUX_MIB_PRUNECALLED);
4994 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4995 tcp_clamp_window(sk);
4996 else if (tcp_under_memory_pressure(sk))
4997 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4999 tcp_collapse_ofo_queue(sk);
5000 if (!skb_queue_empty(&sk->sk_receive_queue))
5001 tcp_collapse(sk, &sk->sk_receive_queue, NULL,
5002 skb_peek(&sk->sk_receive_queue),
5004 tp->copied_seq, tp->rcv_nxt);
5007 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5010 /* Collapsing did not help, destructive actions follow.
5011 * This must not ever occur. */
5013 tcp_prune_ofo_queue(sk);
5015 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
5018 /* If we are really being abused, tell the caller to silently
5019 * drop receive data on the floor. It will get retransmitted
5020 * and hopefully then we'll have sufficient space.
5022 NET_INC_STATS(sock_net(sk), LINUX_MIB_RCVPRUNED);
5024 /* Massive buffer overcommit. */
5029 static bool tcp_should_expand_sndbuf(const struct sock *sk)
5031 const struct tcp_sock *tp = tcp_sk(sk);
5033 /* If the user specified a specific send buffer setting, do
5036 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
5039 /* If we are under global TCP memory pressure, do not expand. */
5040 if (tcp_under_memory_pressure(sk))
5043 /* If we are under soft global TCP memory pressure, do not expand. */
5044 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
5047 /* If we filled the congestion window, do not expand. */
5048 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
5054 /* When incoming ACK allowed to free some skb from write_queue,
5055 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
5056 * on the exit from tcp input handler.
5058 * PROBLEM: sndbuf expansion does not work well with largesend.
5060 static void tcp_new_space(struct sock *sk)
5062 struct tcp_sock *tp = tcp_sk(sk);
5064 if (tcp_should_expand_sndbuf(sk)) {
5065 tcp_sndbuf_expand(sk);
5066 tp->snd_cwnd_stamp = tcp_jiffies32;
5069 sk->sk_write_space(sk);
5072 static void tcp_check_space(struct sock *sk)
5074 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
5075 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
5076 /* pairs with tcp_poll() */
5078 if (sk->sk_socket &&
5079 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
5081 if (!test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
5082 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
5087 static inline void tcp_data_snd_check(struct sock *sk)
5089 tcp_push_pending_frames(sk);
5090 tcp_check_space(sk);
5094 * Check if sending an ack is needed.
5096 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
5098 struct tcp_sock *tp = tcp_sk(sk);
5099 unsigned long rtt, delay;
5101 /* More than one full frame received... */
5102 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
5103 /* ... and right edge of window advances far enough.
5104 * (tcp_recvmsg() will send ACK otherwise).
5105 * If application uses SO_RCVLOWAT, we want send ack now if
5106 * we have not received enough bytes to satisfy the condition.
5108 (tp->rcv_nxt - tp->copied_seq < sk->sk_rcvlowat ||
5109 __tcp_select_window(sk) >= tp->rcv_wnd)) ||
5110 /* We ACK each frame or... */
5111 tcp_in_quickack_mode(sk)) {
5117 if (!ofo_possible || RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
5118 tcp_send_delayed_ack(sk);
5122 if (!tcp_is_sack(tp) ||
5123 tp->compressed_ack >= sock_net(sk)->ipv4.sysctl_tcp_comp_sack_nr)
5125 tp->compressed_ack++;
5127 if (hrtimer_is_queued(&tp->compressed_ack_timer))
5130 /* compress ack timer : 5 % of rtt, but no more than tcp_comp_sack_delay_ns */
5132 rtt = tp->rcv_rtt_est.rtt_us;
5133 if (tp->srtt_us && tp->srtt_us < rtt)
5136 delay = min_t(unsigned long, sock_net(sk)->ipv4.sysctl_tcp_comp_sack_delay_ns,
5137 rtt * (NSEC_PER_USEC >> 3)/20);
5139 hrtimer_start(&tp->compressed_ack_timer, ns_to_ktime(delay),
5140 HRTIMER_MODE_REL_PINNED_SOFT);
5143 static inline void tcp_ack_snd_check(struct sock *sk)
5145 if (!inet_csk_ack_scheduled(sk)) {
5146 /* We sent a data segment already. */
5149 __tcp_ack_snd_check(sk, 1);
5153 * This routine is only called when we have urgent data
5154 * signaled. Its the 'slow' part of tcp_urg. It could be
5155 * moved inline now as tcp_urg is only called from one
5156 * place. We handle URGent data wrong. We have to - as
5157 * BSD still doesn't use the correction from RFC961.
5158 * For 1003.1g we should support a new option TCP_STDURG to permit
5159 * either form (or just set the sysctl tcp_stdurg).
5162 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
5164 struct tcp_sock *tp = tcp_sk(sk);
5165 u32 ptr = ntohs(th->urg_ptr);
5167 if (ptr && !sock_net(sk)->ipv4.sysctl_tcp_stdurg)
5169 ptr += ntohl(th->seq);
5171 /* Ignore urgent data that we've already seen and read. */
5172 if (after(tp->copied_seq, ptr))
5175 /* Do not replay urg ptr.
5177 * NOTE: interesting situation not covered by specs.
5178 * Misbehaving sender may send urg ptr, pointing to segment,
5179 * which we already have in ofo queue. We are not able to fetch
5180 * such data and will stay in TCP_URG_NOTYET until will be eaten
5181 * by recvmsg(). Seems, we are not obliged to handle such wicked
5182 * situations. But it is worth to think about possibility of some
5183 * DoSes using some hypothetical application level deadlock.
5185 if (before(ptr, tp->rcv_nxt))
5188 /* Do we already have a newer (or duplicate) urgent pointer? */
5189 if (tp->urg_data && !after(ptr, tp->urg_seq))
5192 /* Tell the world about our new urgent pointer. */
5195 /* We may be adding urgent data when the last byte read was
5196 * urgent. To do this requires some care. We cannot just ignore
5197 * tp->copied_seq since we would read the last urgent byte again
5198 * as data, nor can we alter copied_seq until this data arrives
5199 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5201 * NOTE. Double Dutch. Rendering to plain English: author of comment
5202 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5203 * and expect that both A and B disappear from stream. This is _wrong_.
5204 * Though this happens in BSD with high probability, this is occasional.
5205 * Any application relying on this is buggy. Note also, that fix "works"
5206 * only in this artificial test. Insert some normal data between A and B and we will
5207 * decline of BSD again. Verdict: it is better to remove to trap
5210 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5211 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5212 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5214 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5215 __skb_unlink(skb, &sk->sk_receive_queue);
5220 tp->urg_data = TCP_URG_NOTYET;
5223 /* Disable header prediction. */
5227 /* This is the 'fast' part of urgent handling. */
5228 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
5230 struct tcp_sock *tp = tcp_sk(sk);
5232 /* Check if we get a new urgent pointer - normally not. */
5234 tcp_check_urg(sk, th);
5236 /* Do we wait for any urgent data? - normally not... */
5237 if (tp->urg_data == TCP_URG_NOTYET) {
5238 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5241 /* Is the urgent pointer pointing into this packet? */
5242 if (ptr < skb->len) {
5244 if (skb_copy_bits(skb, ptr, &tmp, 1))
5246 tp->urg_data = TCP_URG_VALID | tmp;
5247 if (!sock_flag(sk, SOCK_DEAD))
5248 sk->sk_data_ready(sk);
5253 /* Accept RST for rcv_nxt - 1 after a FIN.
5254 * When tcp connections are abruptly terminated from Mac OSX (via ^C), a
5255 * FIN is sent followed by a RST packet. The RST is sent with the same
5256 * sequence number as the FIN, and thus according to RFC 5961 a challenge
5257 * ACK should be sent. However, Mac OSX rate limits replies to challenge
5258 * ACKs on the closed socket. In addition middleboxes can drop either the
5259 * challenge ACK or a subsequent RST.
5261 static bool tcp_reset_check(const struct sock *sk, const struct sk_buff *skb)
5263 struct tcp_sock *tp = tcp_sk(sk);
5265 return unlikely(TCP_SKB_CB(skb)->seq == (tp->rcv_nxt - 1) &&
5266 (1 << sk->sk_state) & (TCPF_CLOSE_WAIT | TCPF_LAST_ACK |
5270 /* Does PAWS and seqno based validation of an incoming segment, flags will
5271 * play significant role here.
5273 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5274 const struct tcphdr *th, int syn_inerr)
5276 struct tcp_sock *tp = tcp_sk(sk);
5277 bool rst_seq_match = false;
5279 /* RFC1323: H1. Apply PAWS check first. */
5280 if (tcp_fast_parse_options(sock_net(sk), skb, th, tp) &&
5281 tp->rx_opt.saw_tstamp &&
5282 tcp_paws_discard(sk, skb)) {
5284 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5285 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5286 LINUX_MIB_TCPACKSKIPPEDPAWS,
5287 &tp->last_oow_ack_time))
5288 tcp_send_dupack(sk, skb);
5291 /* Reset is accepted even if it did not pass PAWS. */
5294 /* Step 1: check sequence number */
5295 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5296 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5297 * (RST) segments are validated by checking their SEQ-fields."
5298 * And page 69: "If an incoming segment is not acceptable,
5299 * an acknowledgment should be sent in reply (unless the RST
5300 * bit is set, if so drop the segment and return)".
5305 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5306 LINUX_MIB_TCPACKSKIPPEDSEQ,
5307 &tp->last_oow_ack_time))
5308 tcp_send_dupack(sk, skb);
5309 } else if (tcp_reset_check(sk, skb)) {
5315 /* Step 2: check RST bit */
5317 /* RFC 5961 3.2 (extend to match against (RCV.NXT - 1) after a
5318 * FIN and SACK too if available):
5319 * If seq num matches RCV.NXT or (RCV.NXT - 1) after a FIN, or
5320 * the right-most SACK block,
5322 * RESET the connection
5324 * Send a challenge ACK
5326 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt ||
5327 tcp_reset_check(sk, skb)) {
5328 rst_seq_match = true;
5329 } else if (tcp_is_sack(tp) && tp->rx_opt.num_sacks > 0) {
5330 struct tcp_sack_block *sp = &tp->selective_acks[0];
5331 int max_sack = sp[0].end_seq;
5334 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;
5336 max_sack = after(sp[this_sack].end_seq,
5338 sp[this_sack].end_seq : max_sack;
5341 if (TCP_SKB_CB(skb)->seq == max_sack)
5342 rst_seq_match = true;
5348 /* Disable TFO if RST is out-of-order
5349 * and no data has been received
5350 * for current active TFO socket
5352 if (tp->syn_fastopen && !tp->data_segs_in &&
5353 sk->sk_state == TCP_ESTABLISHED)
5354 tcp_fastopen_active_disable(sk);
5355 tcp_send_challenge_ack(sk, skb);
5360 /* step 3: check security and precedence [ignored] */
5362 /* step 4: Check for a SYN
5363 * RFC 5961 4.2 : Send a challenge ack
5368 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5369 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5370 tcp_send_challenge_ack(sk, skb);
5382 * TCP receive function for the ESTABLISHED state.
5384 * It is split into a fast path and a slow path. The fast path is
5386 * - A zero window was announced from us - zero window probing
5387 * is only handled properly in the slow path.
5388 * - Out of order segments arrived.
5389 * - Urgent data is expected.
5390 * - There is no buffer space left
5391 * - Unexpected TCP flags/window values/header lengths are received
5392 * (detected by checking the TCP header against pred_flags)
5393 * - Data is sent in both directions. Fast path only supports pure senders
5394 * or pure receivers (this means either the sequence number or the ack
5395 * value must stay constant)
5396 * - Unexpected TCP option.
5398 * When these conditions are not satisfied it drops into a standard
5399 * receive procedure patterned after RFC793 to handle all cases.
5400 * The first three cases are guaranteed by proper pred_flags setting,
5401 * the rest is checked inline. Fast processing is turned on in
5402 * tcp_data_queue when everything is OK.
5404 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb)
5406 const struct tcphdr *th = (const struct tcphdr *)skb->data;
5407 struct tcp_sock *tp = tcp_sk(sk);
5408 unsigned int len = skb->len;
5410 /* TCP congestion window tracking */
5411 trace_tcp_probe(sk, skb);
5413 tcp_mstamp_refresh(tp);
5414 if (unlikely(!sk->sk_rx_dst))
5415 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
5417 * Header prediction.
5418 * The code loosely follows the one in the famous
5419 * "30 instruction TCP receive" Van Jacobson mail.
5421 * Van's trick is to deposit buffers into socket queue
5422 * on a device interrupt, to call tcp_recv function
5423 * on the receive process context and checksum and copy
5424 * the buffer to user space. smart...
5426 * Our current scheme is not silly either but we take the
5427 * extra cost of the net_bh soft interrupt processing...
5428 * We do checksum and copy also but from device to kernel.
5431 tp->rx_opt.saw_tstamp = 0;
5433 /* pred_flags is 0xS?10 << 16 + snd_wnd
5434 * if header_prediction is to be made
5435 * 'S' will always be tp->tcp_header_len >> 2
5436 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5437 * turn it off (when there are holes in the receive
5438 * space for instance)
5439 * PSH flag is ignored.
5442 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5443 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5444 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5445 int tcp_header_len = tp->tcp_header_len;
5447 /* Timestamp header prediction: tcp_header_len
5448 * is automatically equal to th->doff*4 due to pred_flags
5452 /* Check timestamp */
5453 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5454 /* No? Slow path! */
5455 if (!tcp_parse_aligned_timestamp(tp, th))
5458 /* If PAWS failed, check it more carefully in slow path */
5459 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5462 /* DO NOT update ts_recent here, if checksum fails
5463 * and timestamp was corrupted part, it will result
5464 * in a hung connection since we will drop all
5465 * future packets due to the PAWS test.
5469 if (len <= tcp_header_len) {
5470 /* Bulk data transfer: sender */
5471 if (len == tcp_header_len) {
5472 /* Predicted packet is in window by definition.
5473 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5474 * Hence, check seq<=rcv_wup reduces to:
5476 if (tcp_header_len ==
5477 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5478 tp->rcv_nxt == tp->rcv_wup)
5479 tcp_store_ts_recent(tp);
5481 /* We know that such packets are checksummed
5484 tcp_ack(sk, skb, 0);
5486 tcp_data_snd_check(sk);
5488 } else { /* Header too small */
5489 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5494 bool fragstolen = false;
5496 if (tcp_checksum_complete(skb))
5499 if ((int)skb->truesize > sk->sk_forward_alloc)
5502 /* Predicted packet is in window by definition.
5503 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5504 * Hence, check seq<=rcv_wup reduces to:
5506 if (tcp_header_len ==
5507 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5508 tp->rcv_nxt == tp->rcv_wup)
5509 tcp_store_ts_recent(tp);
5511 tcp_rcv_rtt_measure_ts(sk, skb);
5513 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPHITS);
5515 /* Bulk data transfer: receiver */
5516 eaten = tcp_queue_rcv(sk, skb, tcp_header_len,
5519 tcp_event_data_recv(sk, skb);
5521 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5522 /* Well, only one small jumplet in fast path... */
5523 tcp_ack(sk, skb, FLAG_DATA);
5524 tcp_data_snd_check(sk);
5525 if (!inet_csk_ack_scheduled(sk))
5529 __tcp_ack_snd_check(sk, 0);
5532 kfree_skb_partial(skb, fragstolen);
5539 if (len < (th->doff << 2) || tcp_checksum_complete(skb))
5542 if (!th->ack && !th->rst && !th->syn)
5546 * Standard slow path.
5549 if (!tcp_validate_incoming(sk, skb, th, 1))
5553 if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0)
5556 tcp_rcv_rtt_measure_ts(sk, skb);
5558 /* Process urgent data. */
5559 tcp_urg(sk, skb, th);
5561 /* step 7: process the segment text */
5562 tcp_data_queue(sk, skb);
5564 tcp_data_snd_check(sk);
5565 tcp_ack_snd_check(sk);
5569 TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS);
5570 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5575 EXPORT_SYMBOL(tcp_rcv_established);
5577 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5579 struct tcp_sock *tp = tcp_sk(sk);
5580 struct inet_connection_sock *icsk = inet_csk(sk);
5582 tcp_set_state(sk, TCP_ESTABLISHED);
5583 icsk->icsk_ack.lrcvtime = tcp_jiffies32;
5586 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
5587 security_inet_conn_established(sk, skb);
5590 tcp_init_transfer(sk, BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB);
5592 /* Prevent spurious tcp_cwnd_restart() on first data
5595 tp->lsndtime = tcp_jiffies32;
5597 if (sock_flag(sk, SOCK_KEEPOPEN))
5598 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5600 if (!tp->rx_opt.snd_wscale)
5601 __tcp_fast_path_on(tp, tp->snd_wnd);
5606 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
5607 struct tcp_fastopen_cookie *cookie)
5609 struct tcp_sock *tp = tcp_sk(sk);
5610 struct sk_buff *data = tp->syn_data ? tcp_rtx_queue_head(sk) : NULL;
5611 u16 mss = tp->rx_opt.mss_clamp, try_exp = 0;
5612 bool syn_drop = false;
5614 if (mss == tp->rx_opt.user_mss) {
5615 struct tcp_options_received opt;
5617 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5618 tcp_clear_options(&opt);
5619 opt.user_mss = opt.mss_clamp = 0;
5620 tcp_parse_options(sock_net(sk), synack, &opt, 0, NULL);
5621 mss = opt.mss_clamp;
5624 if (!tp->syn_fastopen) {
5625 /* Ignore an unsolicited cookie */
5627 } else if (tp->total_retrans) {
5628 /* SYN timed out and the SYN-ACK neither has a cookie nor
5629 * acknowledges data. Presumably the remote received only
5630 * the retransmitted (regular) SYNs: either the original
5631 * SYN-data or the corresponding SYN-ACK was dropped.
5633 syn_drop = (cookie->len < 0 && data);
5634 } else if (cookie->len < 0 && !tp->syn_data) {
5635 /* We requested a cookie but didn't get it. If we did not use
5636 * the (old) exp opt format then try so next time (try_exp=1).
5637 * Otherwise we go back to use the RFC7413 opt (try_exp=2).
5639 try_exp = tp->syn_fastopen_exp ? 2 : 1;
5642 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp);
5644 if (data) { /* Retransmit unacked data in SYN */
5645 skb_rbtree_walk_from(data) {
5646 if (__tcp_retransmit_skb(sk, data, 1))
5650 NET_INC_STATS(sock_net(sk),
5651 LINUX_MIB_TCPFASTOPENACTIVEFAIL);
5654 tp->syn_data_acked = tp->syn_data;
5655 if (tp->syn_data_acked) {
5656 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
5657 /* SYN-data is counted as two separate packets in tcp_ack() */
5658 if (tp->delivered > 1)
5662 tcp_fastopen_add_skb(sk, synack);
5667 static void smc_check_reset_syn(struct tcp_sock *tp)
5669 #if IS_ENABLED(CONFIG_SMC)
5670 if (static_branch_unlikely(&tcp_have_smc)) {
5671 if (tp->syn_smc && !tp->rx_opt.smc_ok)
5677 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5678 const struct tcphdr *th)
5680 struct inet_connection_sock *icsk = inet_csk(sk);
5681 struct tcp_sock *tp = tcp_sk(sk);
5682 struct tcp_fastopen_cookie foc = { .len = -1 };
5683 int saved_clamp = tp->rx_opt.mss_clamp;
5686 tcp_parse_options(sock_net(sk), skb, &tp->rx_opt, 0, &foc);
5687 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
5688 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
5692 * "If the state is SYN-SENT then
5693 * first check the ACK bit
5694 * If the ACK bit is set
5695 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5696 * a reset (unless the RST bit is set, if so drop
5697 * the segment and return)"
5699 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
5700 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt))
5701 goto reset_and_undo;
5703 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5704 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5705 tcp_time_stamp(tp))) {
5706 NET_INC_STATS(sock_net(sk),
5707 LINUX_MIB_PAWSACTIVEREJECTED);
5708 goto reset_and_undo;
5711 /* Now ACK is acceptable.
5713 * "If the RST bit is set
5714 * If the ACK was acceptable then signal the user "error:
5715 * connection reset", drop the segment, enter CLOSED state,
5716 * delete TCB, and return."
5725 * "fifth, if neither of the SYN or RST bits is set then
5726 * drop the segment and return."
5732 goto discard_and_undo;
5735 * "If the SYN bit is on ...
5736 * are acceptable then ...
5737 * (our SYN has been ACKed), change the connection
5738 * state to ESTABLISHED..."
5741 tcp_ecn_rcv_synack(tp, th);
5743 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5744 tcp_ack(sk, skb, FLAG_SLOWPATH);
5746 /* Ok.. it's good. Set up sequence numbers and
5747 * move to established.
5749 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5750 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5752 /* RFC1323: The window in SYN & SYN/ACK segments is
5755 tp->snd_wnd = ntohs(th->window);
5757 if (!tp->rx_opt.wscale_ok) {
5758 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5759 tp->window_clamp = min(tp->window_clamp, 65535U);
5762 if (tp->rx_opt.saw_tstamp) {
5763 tp->rx_opt.tstamp_ok = 1;
5764 tp->tcp_header_len =
5765 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5766 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5767 tcp_store_ts_recent(tp);
5769 tp->tcp_header_len = sizeof(struct tcphdr);
5772 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5773 tcp_initialize_rcv_mss(sk);
5775 /* Remember, tcp_poll() does not lock socket!
5776 * Change state from SYN-SENT only after copied_seq
5777 * is initialized. */
5778 tp->copied_seq = tp->rcv_nxt;
5780 smc_check_reset_syn(tp);
5784 tcp_finish_connect(sk, skb);
5786 fastopen_fail = (tp->syn_fastopen || tp->syn_data) &&
5787 tcp_rcv_fastopen_synack(sk, skb, &foc);
5789 if (!sock_flag(sk, SOCK_DEAD)) {
5790 sk->sk_state_change(sk);
5791 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5795 if (sk->sk_write_pending ||
5796 icsk->icsk_accept_queue.rskq_defer_accept ||
5797 icsk->icsk_ack.pingpong) {
5798 /* Save one ACK. Data will be ready after
5799 * several ticks, if write_pending is set.
5801 * It may be deleted, but with this feature tcpdumps
5802 * look so _wonderfully_ clever, that I was not able
5803 * to stand against the temptation 8) --ANK
5805 inet_csk_schedule_ack(sk);
5806 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS);
5807 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5808 TCP_DELACK_MAX, TCP_RTO_MAX);
5819 /* No ACK in the segment */
5823 * "If the RST bit is set
5825 * Otherwise (no ACK) drop the segment and return."
5828 goto discard_and_undo;
5832 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5833 tcp_paws_reject(&tp->rx_opt, 0))
5834 goto discard_and_undo;
5837 /* We see SYN without ACK. It is attempt of
5838 * simultaneous connect with crossed SYNs.
5839 * Particularly, it can be connect to self.
5841 tcp_set_state(sk, TCP_SYN_RECV);
5843 if (tp->rx_opt.saw_tstamp) {
5844 tp->rx_opt.tstamp_ok = 1;
5845 tcp_store_ts_recent(tp);
5846 tp->tcp_header_len =
5847 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5849 tp->tcp_header_len = sizeof(struct tcphdr);
5852 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5853 tp->copied_seq = tp->rcv_nxt;
5854 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5856 /* RFC1323: The window in SYN & SYN/ACK segments is
5859 tp->snd_wnd = ntohs(th->window);
5860 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5861 tp->max_window = tp->snd_wnd;
5863 tcp_ecn_rcv_syn(tp, th);
5866 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5867 tcp_initialize_rcv_mss(sk);
5869 tcp_send_synack(sk);
5871 /* Note, we could accept data and URG from this segment.
5872 * There are no obstacles to make this (except that we must
5873 * either change tcp_recvmsg() to prevent it from returning data
5874 * before 3WHS completes per RFC793, or employ TCP Fast Open).
5876 * However, if we ignore data in ACKless segments sometimes,
5877 * we have no reasons to accept it sometimes.
5878 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5879 * is not flawless. So, discard packet for sanity.
5880 * Uncomment this return to process the data.
5887 /* "fifth, if neither of the SYN or RST bits is set then
5888 * drop the segment and return."
5892 tcp_clear_options(&tp->rx_opt);
5893 tp->rx_opt.mss_clamp = saved_clamp;
5897 tcp_clear_options(&tp->rx_opt);
5898 tp->rx_opt.mss_clamp = saved_clamp;
5903 * This function implements the receiving procedure of RFC 793 for
5904 * all states except ESTABLISHED and TIME_WAIT.
5905 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5906 * address independent.
5909 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb)
5911 struct tcp_sock *tp = tcp_sk(sk);
5912 struct inet_connection_sock *icsk = inet_csk(sk);
5913 const struct tcphdr *th = tcp_hdr(skb);
5914 struct request_sock *req;
5918 switch (sk->sk_state) {
5932 /* It is possible that we process SYN packets from backlog,
5933 * so we need to make sure to disable BH right there.
5936 acceptable = icsk->icsk_af_ops->conn_request(sk, skb) >= 0;
5947 tp->rx_opt.saw_tstamp = 0;
5948 tcp_mstamp_refresh(tp);
5949 queued = tcp_rcv_synsent_state_process(sk, skb, th);
5953 /* Do step6 onward by hand. */
5954 tcp_urg(sk, skb, th);
5956 tcp_data_snd_check(sk);
5960 tcp_mstamp_refresh(tp);
5961 tp->rx_opt.saw_tstamp = 0;
5962 req = tp->fastopen_rsk;
5966 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
5967 sk->sk_state != TCP_FIN_WAIT1);
5969 if (!tcp_check_req(sk, skb, req, true, &req_stolen))
5973 if (!th->ack && !th->rst && !th->syn)
5976 if (!tcp_validate_incoming(sk, skb, th, 0))
5979 /* step 5: check the ACK field */
5980 acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
5981 FLAG_UPDATE_TS_RECENT |
5982 FLAG_NO_CHALLENGE_ACK) > 0;
5985 if (sk->sk_state == TCP_SYN_RECV)
5986 return 1; /* send one RST */
5987 tcp_send_challenge_ack(sk, skb);
5990 switch (sk->sk_state) {
5992 tp->delivered++; /* SYN-ACK delivery isn't tracked in tcp_ack */
5994 tcp_synack_rtt_meas(sk, req);
5996 /* Once we leave TCP_SYN_RECV, we no longer need req
6000 inet_csk(sk)->icsk_retransmits = 0;
6001 reqsk_fastopen_remove(sk, req, false);
6002 /* Re-arm the timer because data may have been sent out.
6003 * This is similar to the regular data transmission case
6004 * when new data has just been ack'ed.
6006 * (TFO) - we could try to be more aggressive and
6007 * retransmitting any data sooner based on when they
6012 tcp_init_transfer(sk, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB);
6013 tp->copied_seq = tp->rcv_nxt;
6016 tcp_set_state(sk, TCP_ESTABLISHED);
6017 sk->sk_state_change(sk);
6019 /* Note, that this wakeup is only for marginal crossed SYN case.
6020 * Passively open sockets are not waked up, because
6021 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
6024 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
6026 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
6027 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
6028 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
6030 if (tp->rx_opt.tstamp_ok)
6031 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
6033 if (!inet_csk(sk)->icsk_ca_ops->cong_control)
6034 tcp_update_pacing_rate(sk);
6036 /* Prevent spurious tcp_cwnd_restart() on first data packet */
6037 tp->lsndtime = tcp_jiffies32;
6039 tcp_initialize_rcv_mss(sk);
6040 tcp_fast_path_on(tp);
6043 case TCP_FIN_WAIT1: {
6046 /* If we enter the TCP_FIN_WAIT1 state and we are a
6047 * Fast Open socket and this is the first acceptable
6048 * ACK we have received, this would have acknowledged
6049 * our SYNACK so stop the SYNACK timer.
6052 /* We no longer need the request sock. */
6053 reqsk_fastopen_remove(sk, req, false);
6056 if (tp->snd_una != tp->write_seq)
6059 tcp_set_state(sk, TCP_FIN_WAIT2);
6060 sk->sk_shutdown |= SEND_SHUTDOWN;
6064 if (!sock_flag(sk, SOCK_DEAD)) {
6065 /* Wake up lingering close() */
6066 sk->sk_state_change(sk);
6070 if (tp->linger2 < 0) {
6072 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6075 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6076 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6077 /* Receive out of order FIN after close() */
6078 if (tp->syn_fastopen && th->fin)
6079 tcp_fastopen_active_disable(sk);
6081 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6085 tmo = tcp_fin_time(sk);
6086 if (tmo > TCP_TIMEWAIT_LEN) {
6087 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
6088 } else if (th->fin || sock_owned_by_user(sk)) {
6089 /* Bad case. We could lose such FIN otherwise.
6090 * It is not a big problem, but it looks confusing
6091 * and not so rare event. We still can lose it now,
6092 * if it spins in bh_lock_sock(), but it is really
6095 inet_csk_reset_keepalive_timer(sk, tmo);
6097 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
6104 if (tp->snd_una == tp->write_seq) {
6105 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
6111 if (tp->snd_una == tp->write_seq) {
6112 tcp_update_metrics(sk);
6119 /* step 6: check the URG bit */
6120 tcp_urg(sk, skb, th);
6122 /* step 7: process the segment text */
6123 switch (sk->sk_state) {
6124 case TCP_CLOSE_WAIT:
6127 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
6132 /* RFC 793 says to queue data in these states,
6133 * RFC 1122 says we MUST send a reset.
6134 * BSD 4.4 also does reset.
6136 if (sk->sk_shutdown & RCV_SHUTDOWN) {
6137 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6138 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6139 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6145 case TCP_ESTABLISHED:
6146 tcp_data_queue(sk, skb);
6151 /* tcp_data could move socket to TIME-WAIT */
6152 if (sk->sk_state != TCP_CLOSE) {
6153 tcp_data_snd_check(sk);
6154 tcp_ack_snd_check(sk);
6163 EXPORT_SYMBOL(tcp_rcv_state_process);
6165 static inline void pr_drop_req(struct request_sock *req, __u16 port, int family)
6167 struct inet_request_sock *ireq = inet_rsk(req);
6169 if (family == AF_INET)
6170 net_dbg_ratelimited("drop open request from %pI4/%u\n",
6171 &ireq->ir_rmt_addr, port);
6172 #if IS_ENABLED(CONFIG_IPV6)
6173 else if (family == AF_INET6)
6174 net_dbg_ratelimited("drop open request from %pI6/%u\n",
6175 &ireq->ir_v6_rmt_addr, port);
6179 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
6181 * If we receive a SYN packet with these bits set, it means a
6182 * network is playing bad games with TOS bits. In order to
6183 * avoid possible false congestion notifications, we disable
6184 * TCP ECN negotiation.
6186 * Exception: tcp_ca wants ECN. This is required for DCTCP
6187 * congestion control: Linux DCTCP asserts ECT on all packets,
6188 * including SYN, which is most optimal solution; however,
6189 * others, such as FreeBSD do not.
6191 static void tcp_ecn_create_request(struct request_sock *req,
6192 const struct sk_buff *skb,
6193 const struct sock *listen_sk,
6194 const struct dst_entry *dst)
6196 const struct tcphdr *th = tcp_hdr(skb);
6197 const struct net *net = sock_net(listen_sk);
6198 bool th_ecn = th->ece && th->cwr;
6205 ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield);
6206 ecn_ok_dst = dst_feature(dst, DST_FEATURE_ECN_MASK);
6207 ecn_ok = net->ipv4.sysctl_tcp_ecn || ecn_ok_dst;
6209 if ((!ect && ecn_ok) || tcp_ca_needs_ecn(listen_sk) ||
6210 (ecn_ok_dst & DST_FEATURE_ECN_CA) ||
6211 tcp_bpf_ca_needs_ecn((struct sock *)req))
6212 inet_rsk(req)->ecn_ok = 1;
6215 static void tcp_openreq_init(struct request_sock *req,
6216 const struct tcp_options_received *rx_opt,
6217 struct sk_buff *skb, const struct sock *sk)
6219 struct inet_request_sock *ireq = inet_rsk(req);
6221 req->rsk_rcv_wnd = 0; /* So that tcp_send_synack() knows! */
6223 tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq;
6224 tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
6225 tcp_rsk(req)->snt_synack = tcp_clock_us();
6226 tcp_rsk(req)->last_oow_ack_time = 0;
6227 req->mss = rx_opt->mss_clamp;
6228 req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0;
6229 ireq->tstamp_ok = rx_opt->tstamp_ok;
6230 ireq->sack_ok = rx_opt->sack_ok;
6231 ireq->snd_wscale = rx_opt->snd_wscale;
6232 ireq->wscale_ok = rx_opt->wscale_ok;
6235 ireq->ir_rmt_port = tcp_hdr(skb)->source;
6236 ireq->ir_num = ntohs(tcp_hdr(skb)->dest);
6237 ireq->ir_mark = inet_request_mark(sk, skb);
6238 #if IS_ENABLED(CONFIG_SMC)
6239 ireq->smc_ok = rx_opt->smc_ok;
6243 struct request_sock *inet_reqsk_alloc(const struct request_sock_ops *ops,
6244 struct sock *sk_listener,
6245 bool attach_listener)
6247 struct request_sock *req = reqsk_alloc(ops, sk_listener,
6251 struct inet_request_sock *ireq = inet_rsk(req);
6253 ireq->ireq_opt = NULL;
6254 #if IS_ENABLED(CONFIG_IPV6)
6255 ireq->pktopts = NULL;
6257 atomic64_set(&ireq->ir_cookie, 0);
6258 ireq->ireq_state = TCP_NEW_SYN_RECV;
6259 write_pnet(&ireq->ireq_net, sock_net(sk_listener));
6260 ireq->ireq_family = sk_listener->sk_family;
6265 EXPORT_SYMBOL(inet_reqsk_alloc);
6268 * Return true if a syncookie should be sent
6270 static bool tcp_syn_flood_action(const struct sock *sk,
6271 const struct sk_buff *skb,
6274 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
6275 const char *msg = "Dropping request";
6276 bool want_cookie = false;
6277 struct net *net = sock_net(sk);
6279 #ifdef CONFIG_SYN_COOKIES
6280 if (net->ipv4.sysctl_tcp_syncookies) {
6281 msg = "Sending cookies";
6283 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES);
6286 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP);
6288 if (!queue->synflood_warned &&
6289 net->ipv4.sysctl_tcp_syncookies != 2 &&
6290 xchg(&queue->synflood_warned, 1) == 0)
6291 pr_info("%s: Possible SYN flooding on port %d. %s. Check SNMP counters.\n",
6292 proto, ntohs(tcp_hdr(skb)->dest), msg);
6297 static void tcp_reqsk_record_syn(const struct sock *sk,
6298 struct request_sock *req,
6299 const struct sk_buff *skb)
6301 if (tcp_sk(sk)->save_syn) {
6302 u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb);
6305 copy = kmalloc(len + sizeof(u32), GFP_ATOMIC);
6308 memcpy(©[1], skb_network_header(skb), len);
6309 req->saved_syn = copy;
6314 int tcp_conn_request(struct request_sock_ops *rsk_ops,
6315 const struct tcp_request_sock_ops *af_ops,
6316 struct sock *sk, struct sk_buff *skb)
6318 struct tcp_fastopen_cookie foc = { .len = -1 };
6319 __u32 isn = TCP_SKB_CB(skb)->tcp_tw_isn;
6320 struct tcp_options_received tmp_opt;
6321 struct tcp_sock *tp = tcp_sk(sk);
6322 struct net *net = sock_net(sk);
6323 struct sock *fastopen_sk = NULL;
6324 struct request_sock *req;
6325 bool want_cookie = false;
6326 struct dst_entry *dst;
6329 /* TW buckets are converted to open requests without
6330 * limitations, they conserve resources and peer is
6331 * evidently real one.
6333 if ((net->ipv4.sysctl_tcp_syncookies == 2 ||
6334 inet_csk_reqsk_queue_is_full(sk)) && !isn) {
6335 want_cookie = tcp_syn_flood_action(sk, skb, rsk_ops->slab_name);
6340 if (sk_acceptq_is_full(sk)) {
6341 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
6345 req = inet_reqsk_alloc(rsk_ops, sk, !want_cookie);
6349 tcp_rsk(req)->af_specific = af_ops;
6350 tcp_rsk(req)->ts_off = 0;
6352 tcp_clear_options(&tmp_opt);
6353 tmp_opt.mss_clamp = af_ops->mss_clamp;
6354 tmp_opt.user_mss = tp->rx_opt.user_mss;
6355 tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0,
6356 want_cookie ? NULL : &foc);
6358 if (want_cookie && !tmp_opt.saw_tstamp)
6359 tcp_clear_options(&tmp_opt);
6361 if (IS_ENABLED(CONFIG_SMC) && want_cookie)
6364 tmp_opt.tstamp_ok = tmp_opt.saw_tstamp;
6365 tcp_openreq_init(req, &tmp_opt, skb, sk);
6366 inet_rsk(req)->no_srccheck = inet_sk(sk)->transparent;
6368 /* Note: tcp_v6_init_req() might override ir_iif for link locals */
6369 inet_rsk(req)->ir_iif = inet_request_bound_dev_if(sk, skb);
6371 af_ops->init_req(req, sk, skb);
6373 if (security_inet_conn_request(sk, skb, req))
6376 if (tmp_opt.tstamp_ok)
6377 tcp_rsk(req)->ts_off = af_ops->init_ts_off(net, skb);
6379 dst = af_ops->route_req(sk, &fl, req);
6383 if (!want_cookie && !isn) {
6384 /* Kill the following clause, if you dislike this way. */
6385 if (!net->ipv4.sysctl_tcp_syncookies &&
6386 (net->ipv4.sysctl_max_syn_backlog - inet_csk_reqsk_queue_len(sk) <
6387 (net->ipv4.sysctl_max_syn_backlog >> 2)) &&
6388 !tcp_peer_is_proven(req, dst)) {
6389 /* Without syncookies last quarter of
6390 * backlog is filled with destinations,
6391 * proven to be alive.
6392 * It means that we continue to communicate
6393 * to destinations, already remembered
6394 * to the moment of synflood.
6396 pr_drop_req(req, ntohs(tcp_hdr(skb)->source),
6398 goto drop_and_release;
6401 isn = af_ops->init_seq(skb);
6404 tcp_ecn_create_request(req, skb, sk, dst);
6407 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss);
6408 req->cookie_ts = tmp_opt.tstamp_ok;
6409 if (!tmp_opt.tstamp_ok)
6410 inet_rsk(req)->ecn_ok = 0;
6413 tcp_rsk(req)->snt_isn = isn;
6414 tcp_rsk(req)->txhash = net_tx_rndhash();
6415 tcp_openreq_init_rwin(req, sk, dst);
6417 tcp_reqsk_record_syn(sk, req, skb);
6418 fastopen_sk = tcp_try_fastopen(sk, skb, req, &foc, dst);
6421 af_ops->send_synack(fastopen_sk, dst, &fl, req,
6422 &foc, TCP_SYNACK_FASTOPEN);
6423 /* Add the child socket directly into the accept queue */
6424 inet_csk_reqsk_queue_add(sk, req, fastopen_sk);
6425 sk->sk_data_ready(sk);
6426 bh_unlock_sock(fastopen_sk);
6427 sock_put(fastopen_sk);
6429 tcp_rsk(req)->tfo_listener = false;
6431 inet_csk_reqsk_queue_hash_add(sk, req,
6432 tcp_timeout_init((struct sock *)req));
6433 af_ops->send_synack(sk, dst, &fl, req, &foc,
6434 !want_cookie ? TCP_SYNACK_NORMAL :
6452 EXPORT_SYMBOL(tcp_conn_request);