2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * Implementation of the Transmission Control Protocol(TCP).
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
23 * Pedro Roque : Fast Retransmit/Recovery.
25 * Retransmit queue handled by TCP.
26 * Better retransmit timer handling.
27 * New congestion avoidance.
31 * Eric : Fast Retransmit.
32 * Randy Scott : MSS option defines.
33 * Eric Schenk : Fixes to slow start algorithm.
34 * Eric Schenk : Yet another double ACK bug.
35 * Eric Schenk : Delayed ACK bug fixes.
36 * Eric Schenk : Floyd style fast retrans war avoidance.
37 * David S. Miller : Don't allow zero congestion window.
38 * Eric Schenk : Fix retransmitter so that it sends
39 * next packet on ack of previous packet.
40 * Andi Kleen : Moved open_request checking here
41 * and process RSTs for open_requests.
42 * Andi Kleen : Better prune_queue, and other fixes.
43 * Andrey Savochkin: Fix RTT measurements in the presence of
45 * Andrey Savochkin: Check sequence numbers correctly when
46 * removing SACKs due to in sequence incoming
48 * Andi Kleen: Make sure we never ack data there is not
49 * enough room for. Also make this condition
50 * a fatal error if it might still happen.
51 * Andi Kleen: Add tcp_measure_rcv_mss to make
52 * connections with MSS<min(MTU,ann. MSS)
53 * work without delayed acks.
54 * Andi Kleen: Process packets with PSH set in the
56 * J Hadi Salim: ECN support
59 * Panu Kuhlberg: Experimental audit of TCP (re)transmission
60 * engine. Lots of bugs are found.
61 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs
64 #define pr_fmt(fmt) "TCP: " fmt
67 #include <linux/slab.h>
68 #include <linux/module.h>
69 #include <linux/sysctl.h>
70 #include <linux/kernel.h>
71 #include <linux/prefetch.h>
74 #include <net/inet_common.h>
75 #include <linux/ipsec.h>
76 #include <asm/unaligned.h>
77 #include <linux/errqueue.h>
79 int sysctl_tcp_timestamps __read_mostly = 1;
80 int sysctl_tcp_window_scaling __read_mostly = 1;
81 int sysctl_tcp_sack __read_mostly = 1;
82 int sysctl_tcp_fack __read_mostly = 1;
83 int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH;
84 int sysctl_tcp_max_reordering __read_mostly = 300;
85 EXPORT_SYMBOL(sysctl_tcp_reordering);
86 int sysctl_tcp_dsack __read_mostly = 1;
87 int sysctl_tcp_app_win __read_mostly = 31;
88 int sysctl_tcp_adv_win_scale __read_mostly = 1;
89 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale);
91 /* rfc5961 challenge ack rate limiting */
92 int sysctl_tcp_challenge_ack_limit = 100;
94 int sysctl_tcp_stdurg __read_mostly;
95 int sysctl_tcp_rfc1337 __read_mostly;
96 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
97 int sysctl_tcp_frto __read_mostly = 2;
98 int sysctl_tcp_min_rtt_wlen __read_mostly = 300;
100 int sysctl_tcp_thin_dupack __read_mostly;
102 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
103 int sysctl_tcp_early_retrans __read_mostly = 3;
104 int sysctl_tcp_invalid_ratelimit __read_mostly = HZ/2;
106 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
107 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
108 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
109 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
110 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
111 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
112 #define FLAG_ECE 0x40 /* ECE in this ACK */
113 #define FLAG_LOST_RETRANS 0x80 /* This ACK marks some retransmission lost */
114 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
115 #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
116 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
117 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
118 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
119 #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
121 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
122 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
123 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
124 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
126 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
127 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
129 #define REXMIT_NONE 0 /* no loss recovery to do */
130 #define REXMIT_LOST 1 /* retransmit packets marked lost */
131 #define REXMIT_NEW 2 /* FRTO-style transmit of unsent/new packets */
133 /* Adapt the MSS value used to make delayed ack decision to the
136 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
138 struct inet_connection_sock *icsk = inet_csk(sk);
139 const unsigned int lss = icsk->icsk_ack.last_seg_size;
142 icsk->icsk_ack.last_seg_size = 0;
144 /* skb->len may jitter because of SACKs, even if peer
145 * sends good full-sized frames.
147 len = skb_shinfo(skb)->gso_size ? : skb->len;
148 if (len >= icsk->icsk_ack.rcv_mss) {
149 icsk->icsk_ack.rcv_mss = len;
151 /* Otherwise, we make more careful check taking into account,
152 * that SACKs block is variable.
154 * "len" is invariant segment length, including TCP header.
156 len += skb->data - skb_transport_header(skb);
157 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
158 /* If PSH is not set, packet should be
159 * full sized, provided peer TCP is not badly broken.
160 * This observation (if it is correct 8)) allows
161 * to handle super-low mtu links fairly.
163 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
164 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
165 /* Subtract also invariant (if peer is RFC compliant),
166 * tcp header plus fixed timestamp option length.
167 * Resulting "len" is MSS free of SACK jitter.
169 len -= tcp_sk(sk)->tcp_header_len;
170 icsk->icsk_ack.last_seg_size = len;
172 icsk->icsk_ack.rcv_mss = len;
176 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
177 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
178 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
182 static void tcp_incr_quickack(struct sock *sk)
184 struct inet_connection_sock *icsk = inet_csk(sk);
185 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
189 if (quickacks > icsk->icsk_ack.quick)
190 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
193 static void tcp_enter_quickack_mode(struct sock *sk)
195 struct inet_connection_sock *icsk = inet_csk(sk);
196 tcp_incr_quickack(sk);
197 icsk->icsk_ack.pingpong = 0;
198 icsk->icsk_ack.ato = TCP_ATO_MIN;
201 /* Send ACKs quickly, if "quick" count is not exhausted
202 * and the session is not interactive.
205 static bool tcp_in_quickack_mode(struct sock *sk)
207 const struct inet_connection_sock *icsk = inet_csk(sk);
208 const struct dst_entry *dst = __sk_dst_get(sk);
210 return (dst && dst_metric(dst, RTAX_QUICKACK)) ||
211 (icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong);
214 static void tcp_ecn_queue_cwr(struct tcp_sock *tp)
216 if (tp->ecn_flags & TCP_ECN_OK)
217 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
220 static void tcp_ecn_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb)
222 if (tcp_hdr(skb)->cwr)
223 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
226 static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp)
228 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
231 static void __tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
233 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
234 case INET_ECN_NOT_ECT:
235 /* Funny extension: if ECT is not set on a segment,
236 * and we already seen ECT on a previous segment,
237 * it is probably a retransmit.
239 if (tp->ecn_flags & TCP_ECN_SEEN)
240 tcp_enter_quickack_mode((struct sock *)tp);
243 if (tcp_ca_needs_ecn((struct sock *)tp))
244 tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_IS_CE);
246 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
247 /* Better not delay acks, sender can have a very low cwnd */
248 tcp_enter_quickack_mode((struct sock *)tp);
249 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
251 tp->ecn_flags |= TCP_ECN_SEEN;
254 if (tcp_ca_needs_ecn((struct sock *)tp))
255 tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_NO_CE);
256 tp->ecn_flags |= TCP_ECN_SEEN;
261 static void tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
263 if (tp->ecn_flags & TCP_ECN_OK)
264 __tcp_ecn_check_ce(tp, skb);
267 static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
269 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
270 tp->ecn_flags &= ~TCP_ECN_OK;
273 static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
275 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
276 tp->ecn_flags &= ~TCP_ECN_OK;
279 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
281 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
286 /* Buffer size and advertised window tuning.
288 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
291 static void tcp_sndbuf_expand(struct sock *sk)
293 const struct tcp_sock *tp = tcp_sk(sk);
297 /* Worst case is non GSO/TSO : each frame consumes one skb
298 * and skb->head is kmalloced using power of two area of memory
300 per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
302 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
304 per_mss = roundup_pow_of_two(per_mss) +
305 SKB_DATA_ALIGN(sizeof(struct sk_buff));
307 nr_segs = max_t(u32, TCP_INIT_CWND, tp->snd_cwnd);
308 nr_segs = max_t(u32, nr_segs, tp->reordering + 1);
310 /* Fast Recovery (RFC 5681 3.2) :
311 * Cubic needs 1.7 factor, rounded to 2 to include
312 * extra cushion (application might react slowly to POLLOUT)
314 sndmem = 2 * nr_segs * per_mss;
316 if (sk->sk_sndbuf < sndmem)
317 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
320 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
322 * All tcp_full_space() is split to two parts: "network" buffer, allocated
323 * forward and advertised in receiver window (tp->rcv_wnd) and
324 * "application buffer", required to isolate scheduling/application
325 * latencies from network.
326 * window_clamp is maximal advertised window. It can be less than
327 * tcp_full_space(), in this case tcp_full_space() - window_clamp
328 * is reserved for "application" buffer. The less window_clamp is
329 * the smoother our behaviour from viewpoint of network, but the lower
330 * throughput and the higher sensitivity of the connection to losses. 8)
332 * rcv_ssthresh is more strict window_clamp used at "slow start"
333 * phase to predict further behaviour of this connection.
334 * It is used for two goals:
335 * - to enforce header prediction at sender, even when application
336 * requires some significant "application buffer". It is check #1.
337 * - to prevent pruning of receive queue because of misprediction
338 * of receiver window. Check #2.
340 * The scheme does not work when sender sends good segments opening
341 * window and then starts to feed us spaghetti. But it should work
342 * in common situations. Otherwise, we have to rely on queue collapsing.
345 /* Slow part of check#2. */
346 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
348 struct tcp_sock *tp = tcp_sk(sk);
350 int truesize = tcp_win_from_space(skb->truesize) >> 1;
351 int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1;
353 while (tp->rcv_ssthresh <= window) {
354 if (truesize <= skb->len)
355 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
363 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
365 struct tcp_sock *tp = tcp_sk(sk);
368 if (tp->rcv_ssthresh < tp->window_clamp &&
369 (int)tp->rcv_ssthresh < tcp_space(sk) &&
370 !tcp_under_memory_pressure(sk)) {
373 /* Check #2. Increase window, if skb with such overhead
374 * will fit to rcvbuf in future.
376 if (tcp_win_from_space(skb->truesize) <= skb->len)
377 incr = 2 * tp->advmss;
379 incr = __tcp_grow_window(sk, skb);
382 incr = max_t(int, incr, 2 * skb->len);
383 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
385 inet_csk(sk)->icsk_ack.quick |= 1;
390 /* 3. Tuning rcvbuf, when connection enters established state. */
391 static void tcp_fixup_rcvbuf(struct sock *sk)
393 u32 mss = tcp_sk(sk)->advmss;
396 rcvmem = 2 * SKB_TRUESIZE(mss + MAX_TCP_HEADER) *
397 tcp_default_init_rwnd(mss);
399 /* Dynamic Right Sizing (DRS) has 2 to 3 RTT latency
400 * Allow enough cushion so that sender is not limited by our window
402 if (sysctl_tcp_moderate_rcvbuf)
405 if (sk->sk_rcvbuf < rcvmem)
406 sk->sk_rcvbuf = min(rcvmem, sysctl_tcp_rmem[2]);
409 /* 4. Try to fixup all. It is made immediately after connection enters
412 void tcp_init_buffer_space(struct sock *sk)
414 struct tcp_sock *tp = tcp_sk(sk);
417 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
418 tcp_fixup_rcvbuf(sk);
419 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
420 tcp_sndbuf_expand(sk);
422 tp->rcvq_space.space = tp->rcv_wnd;
423 tp->rcvq_space.time = tcp_time_stamp;
424 tp->rcvq_space.seq = tp->copied_seq;
426 maxwin = tcp_full_space(sk);
428 if (tp->window_clamp >= maxwin) {
429 tp->window_clamp = maxwin;
431 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
432 tp->window_clamp = max(maxwin -
433 (maxwin >> sysctl_tcp_app_win),
437 /* Force reservation of one segment. */
438 if (sysctl_tcp_app_win &&
439 tp->window_clamp > 2 * tp->advmss &&
440 tp->window_clamp + tp->advmss > maxwin)
441 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
443 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
444 tp->snd_cwnd_stamp = tcp_time_stamp;
447 /* 5. Recalculate window clamp after socket hit its memory bounds. */
448 static void tcp_clamp_window(struct sock *sk)
450 struct tcp_sock *tp = tcp_sk(sk);
451 struct inet_connection_sock *icsk = inet_csk(sk);
453 icsk->icsk_ack.quick = 0;
455 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
456 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
457 !tcp_under_memory_pressure(sk) &&
458 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
459 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
462 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
463 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
466 /* Initialize RCV_MSS value.
467 * RCV_MSS is an our guess about MSS used by the peer.
468 * We haven't any direct information about the MSS.
469 * It's better to underestimate the RCV_MSS rather than overestimate.
470 * Overestimations make us ACKing less frequently than needed.
471 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
473 void tcp_initialize_rcv_mss(struct sock *sk)
475 const struct tcp_sock *tp = tcp_sk(sk);
476 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
478 hint = min(hint, tp->rcv_wnd / 2);
479 hint = min(hint, TCP_MSS_DEFAULT);
480 hint = max(hint, TCP_MIN_MSS);
482 inet_csk(sk)->icsk_ack.rcv_mss = hint;
484 EXPORT_SYMBOL(tcp_initialize_rcv_mss);
486 /* Receiver "autotuning" code.
488 * The algorithm for RTT estimation w/o timestamps is based on
489 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
490 * <http://public.lanl.gov/radiant/pubs.html#DRS>
492 * More detail on this code can be found at
493 * <http://staff.psc.edu/jheffner/>,
494 * though this reference is out of date. A new paper
497 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
499 u32 new_sample = tp->rcv_rtt_est.rtt;
505 if (new_sample != 0) {
506 /* If we sample in larger samples in the non-timestamp
507 * case, we could grossly overestimate the RTT especially
508 * with chatty applications or bulk transfer apps which
509 * are stalled on filesystem I/O.
511 * Also, since we are only going for a minimum in the
512 * non-timestamp case, we do not smooth things out
513 * else with timestamps disabled convergence takes too
517 m -= (new_sample >> 3);
525 /* No previous measure. */
529 if (tp->rcv_rtt_est.rtt != new_sample)
530 tp->rcv_rtt_est.rtt = new_sample;
533 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
535 if (tp->rcv_rtt_est.time == 0)
537 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
539 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rcv_rtt_est.time, 1);
542 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
543 tp->rcv_rtt_est.time = tcp_time_stamp;
546 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
547 const struct sk_buff *skb)
549 struct tcp_sock *tp = tcp_sk(sk);
550 if (tp->rx_opt.rcv_tsecr &&
551 (TCP_SKB_CB(skb)->end_seq -
552 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
553 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
557 * This function should be called every time data is copied to user space.
558 * It calculates the appropriate TCP receive buffer space.
560 void tcp_rcv_space_adjust(struct sock *sk)
562 struct tcp_sock *tp = tcp_sk(sk);
566 time = tcp_time_stamp - tp->rcvq_space.time;
567 if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0)
570 /* Number of bytes copied to user in last RTT */
571 copied = tp->copied_seq - tp->rcvq_space.seq;
572 if (copied <= tp->rcvq_space.space)
576 * copied = bytes received in previous RTT, our base window
577 * To cope with packet losses, we need a 2x factor
578 * To cope with slow start, and sender growing its cwin by 100 %
579 * every RTT, we need a 4x factor, because the ACK we are sending
580 * now is for the next RTT, not the current one :
581 * <prev RTT . ><current RTT .. ><next RTT .... >
584 if (sysctl_tcp_moderate_rcvbuf &&
585 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
586 int rcvwin, rcvmem, rcvbuf;
588 /* minimal window to cope with packet losses, assuming
589 * steady state. Add some cushion because of small variations.
591 rcvwin = (copied << 1) + 16 * tp->advmss;
593 /* If rate increased by 25%,
594 * assume slow start, rcvwin = 3 * copied
595 * If rate increased by 50%,
596 * assume sender can use 2x growth, rcvwin = 4 * copied
599 tp->rcvq_space.space + (tp->rcvq_space.space >> 2)) {
601 tp->rcvq_space.space + (tp->rcvq_space.space >> 1))
604 rcvwin += (rcvwin >> 1);
607 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
608 while (tcp_win_from_space(rcvmem) < tp->advmss)
611 rcvbuf = min(rcvwin / tp->advmss * rcvmem, sysctl_tcp_rmem[2]);
612 if (rcvbuf > sk->sk_rcvbuf) {
613 sk->sk_rcvbuf = rcvbuf;
615 /* Make the window clamp follow along. */
616 tp->window_clamp = rcvwin;
619 tp->rcvq_space.space = copied;
622 tp->rcvq_space.seq = tp->copied_seq;
623 tp->rcvq_space.time = tcp_time_stamp;
626 /* There is something which you must keep in mind when you analyze the
627 * behavior of the tp->ato delayed ack timeout interval. When a
628 * connection starts up, we want to ack as quickly as possible. The
629 * problem is that "good" TCP's do slow start at the beginning of data
630 * transmission. The means that until we send the first few ACK's the
631 * sender will sit on his end and only queue most of his data, because
632 * he can only send snd_cwnd unacked packets at any given time. For
633 * each ACK we send, he increments snd_cwnd and transmits more of his
636 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
638 struct tcp_sock *tp = tcp_sk(sk);
639 struct inet_connection_sock *icsk = inet_csk(sk);
642 inet_csk_schedule_ack(sk);
644 tcp_measure_rcv_mss(sk, skb);
646 tcp_rcv_rtt_measure(tp);
648 now = tcp_time_stamp;
650 if (!icsk->icsk_ack.ato) {
651 /* The _first_ data packet received, initialize
652 * delayed ACK engine.
654 tcp_incr_quickack(sk);
655 icsk->icsk_ack.ato = TCP_ATO_MIN;
657 int m = now - icsk->icsk_ack.lrcvtime;
659 if (m <= TCP_ATO_MIN / 2) {
660 /* The fastest case is the first. */
661 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
662 } else if (m < icsk->icsk_ack.ato) {
663 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
664 if (icsk->icsk_ack.ato > icsk->icsk_rto)
665 icsk->icsk_ack.ato = icsk->icsk_rto;
666 } else if (m > icsk->icsk_rto) {
667 /* Too long gap. Apparently sender failed to
668 * restart window, so that we send ACKs quickly.
670 tcp_incr_quickack(sk);
674 icsk->icsk_ack.lrcvtime = now;
676 tcp_ecn_check_ce(tp, skb);
679 tcp_grow_window(sk, skb);
682 /* Called to compute a smoothed rtt estimate. The data fed to this
683 * routine either comes from timestamps, or from segments that were
684 * known _not_ to have been retransmitted [see Karn/Partridge
685 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
686 * piece by Van Jacobson.
687 * NOTE: the next three routines used to be one big routine.
688 * To save cycles in the RFC 1323 implementation it was better to break
689 * it up into three procedures. -- erics
691 static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
693 struct tcp_sock *tp = tcp_sk(sk);
694 long m = mrtt_us; /* RTT */
695 u32 srtt = tp->srtt_us;
697 /* The following amusing code comes from Jacobson's
698 * article in SIGCOMM '88. Note that rtt and mdev
699 * are scaled versions of rtt and mean deviation.
700 * This is designed to be as fast as possible
701 * m stands for "measurement".
703 * On a 1990 paper the rto value is changed to:
704 * RTO = rtt + 4 * mdev
706 * Funny. This algorithm seems to be very broken.
707 * These formulae increase RTO, when it should be decreased, increase
708 * too slowly, when it should be increased quickly, decrease too quickly
709 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
710 * does not matter how to _calculate_ it. Seems, it was trap
711 * that VJ failed to avoid. 8)
714 m -= (srtt >> 3); /* m is now error in rtt est */
715 srtt += m; /* rtt = 7/8 rtt + 1/8 new */
717 m = -m; /* m is now abs(error) */
718 m -= (tp->mdev_us >> 2); /* similar update on mdev */
719 /* This is similar to one of Eifel findings.
720 * Eifel blocks mdev updates when rtt decreases.
721 * This solution is a bit different: we use finer gain
722 * for mdev in this case (alpha*beta).
723 * Like Eifel it also prevents growth of rto,
724 * but also it limits too fast rto decreases,
725 * happening in pure Eifel.
730 m -= (tp->mdev_us >> 2); /* similar update on mdev */
732 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */
733 if (tp->mdev_us > tp->mdev_max_us) {
734 tp->mdev_max_us = tp->mdev_us;
735 if (tp->mdev_max_us > tp->rttvar_us)
736 tp->rttvar_us = tp->mdev_max_us;
738 if (after(tp->snd_una, tp->rtt_seq)) {
739 if (tp->mdev_max_us < tp->rttvar_us)
740 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
741 tp->rtt_seq = tp->snd_nxt;
742 tp->mdev_max_us = tcp_rto_min_us(sk);
745 /* no previous measure. */
746 srtt = m << 3; /* take the measured time to be rtt */
747 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */
748 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
749 tp->mdev_max_us = tp->rttvar_us;
750 tp->rtt_seq = tp->snd_nxt;
752 tp->srtt_us = max(1U, srtt);
755 /* Set the sk_pacing_rate to allow proper sizing of TSO packets.
756 * Note: TCP stack does not yet implement pacing.
757 * FQ packet scheduler can be used to implement cheap but effective
758 * TCP pacing, to smooth the burst on large writes when packets
759 * in flight is significantly lower than cwnd (or rwin)
761 int sysctl_tcp_pacing_ss_ratio __read_mostly = 200;
762 int sysctl_tcp_pacing_ca_ratio __read_mostly = 120;
764 static void tcp_update_pacing_rate(struct sock *sk)
766 const struct tcp_sock *tp = tcp_sk(sk);
769 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
770 rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3);
772 /* current rate is (cwnd * mss) / srtt
773 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
774 * In Congestion Avoidance phase, set it to 120 % the current rate.
776 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
777 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
778 * end of slow start and should slow down.
780 if (tp->snd_cwnd < tp->snd_ssthresh / 2)
781 rate *= sysctl_tcp_pacing_ss_ratio;
783 rate *= sysctl_tcp_pacing_ca_ratio;
785 rate *= max(tp->snd_cwnd, tp->packets_out);
787 if (likely(tp->srtt_us))
788 do_div(rate, tp->srtt_us);
790 /* ACCESS_ONCE() is needed because sch_fq fetches sk_pacing_rate
791 * without any lock. We want to make sure compiler wont store
792 * intermediate values in this location.
794 ACCESS_ONCE(sk->sk_pacing_rate) = min_t(u64, rate,
795 sk->sk_max_pacing_rate);
798 /* Calculate rto without backoff. This is the second half of Van Jacobson's
799 * routine referred to above.
801 static void tcp_set_rto(struct sock *sk)
803 const struct tcp_sock *tp = tcp_sk(sk);
804 /* Old crap is replaced with new one. 8)
807 * 1. If rtt variance happened to be less 50msec, it is hallucination.
808 * It cannot be less due to utterly erratic ACK generation made
809 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
810 * to do with delayed acks, because at cwnd>2 true delack timeout
811 * is invisible. Actually, Linux-2.4 also generates erratic
812 * ACKs in some circumstances.
814 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
816 /* 2. Fixups made earlier cannot be right.
817 * If we do not estimate RTO correctly without them,
818 * all the algo is pure shit and should be replaced
819 * with correct one. It is exactly, which we pretend to do.
822 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
823 * guarantees that rto is higher.
828 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
830 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
833 cwnd = TCP_INIT_CWND;
834 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
838 * Packet counting of FACK is based on in-order assumptions, therefore TCP
839 * disables it when reordering is detected
841 void tcp_disable_fack(struct tcp_sock *tp)
843 /* RFC3517 uses different metric in lost marker => reset on change */
845 tp->lost_skb_hint = NULL;
846 tp->rx_opt.sack_ok &= ~TCP_FACK_ENABLED;
849 /* Take a notice that peer is sending D-SACKs */
850 static void tcp_dsack_seen(struct tcp_sock *tp)
852 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
855 static void tcp_update_reordering(struct sock *sk, const int metric,
858 struct tcp_sock *tp = tcp_sk(sk);
859 if (metric > tp->reordering) {
862 tp->reordering = min(sysctl_tcp_max_reordering, metric);
864 /* This exciting event is worth to be remembered. 8) */
866 mib_idx = LINUX_MIB_TCPTSREORDER;
867 else if (tcp_is_reno(tp))
868 mib_idx = LINUX_MIB_TCPRENOREORDER;
869 else if (tcp_is_fack(tp))
870 mib_idx = LINUX_MIB_TCPFACKREORDER;
872 mib_idx = LINUX_MIB_TCPSACKREORDER;
874 NET_INC_STATS_BH(sock_net(sk), mib_idx);
875 #if FASTRETRANS_DEBUG > 1
876 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
877 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
881 tp->undo_marker ? tp->undo_retrans : 0);
883 tcp_disable_fack(tp);
887 tcp_disable_early_retrans(tp);
891 /* This must be called before lost_out is incremented */
892 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
894 if (!tp->retransmit_skb_hint ||
895 before(TCP_SKB_CB(skb)->seq,
896 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
897 tp->retransmit_skb_hint = skb;
900 after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high))
901 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
904 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
906 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
907 tcp_verify_retransmit_hint(tp, skb);
909 tp->lost_out += tcp_skb_pcount(skb);
910 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
914 void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp, struct sk_buff *skb)
916 tcp_verify_retransmit_hint(tp, skb);
918 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
919 tp->lost_out += tcp_skb_pcount(skb);
920 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
924 /* This procedure tags the retransmission queue when SACKs arrive.
926 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
927 * Packets in queue with these bits set are counted in variables
928 * sacked_out, retrans_out and lost_out, correspondingly.
930 * Valid combinations are:
931 * Tag InFlight Description
932 * 0 1 - orig segment is in flight.
933 * S 0 - nothing flies, orig reached receiver.
934 * L 0 - nothing flies, orig lost by net.
935 * R 2 - both orig and retransmit are in flight.
936 * L|R 1 - orig is lost, retransmit is in flight.
937 * S|R 1 - orig reached receiver, retrans is still in flight.
938 * (L|S|R is logically valid, it could occur when L|R is sacked,
939 * but it is equivalent to plain S and code short-curcuits it to S.
940 * L|S is logically invalid, it would mean -1 packet in flight 8))
942 * These 6 states form finite state machine, controlled by the following events:
943 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
944 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
945 * 3. Loss detection event of two flavors:
946 * A. Scoreboard estimator decided the packet is lost.
947 * A'. Reno "three dupacks" marks head of queue lost.
948 * A''. Its FACK modification, head until snd.fack is lost.
949 * B. SACK arrives sacking SND.NXT at the moment, when the
950 * segment was retransmitted.
951 * 4. D-SACK added new rule: D-SACK changes any tag to S.
953 * It is pleasant to note, that state diagram turns out to be commutative,
954 * so that we are allowed not to be bothered by order of our actions,
955 * when multiple events arrive simultaneously. (see the function below).
957 * Reordering detection.
958 * --------------------
959 * Reordering metric is maximal distance, which a packet can be displaced
960 * in packet stream. With SACKs we can estimate it:
962 * 1. SACK fills old hole and the corresponding segment was not
963 * ever retransmitted -> reordering. Alas, we cannot use it
964 * when segment was retransmitted.
965 * 2. The last flaw is solved with D-SACK. D-SACK arrives
966 * for retransmitted and already SACKed segment -> reordering..
967 * Both of these heuristics are not used in Loss state, when we cannot
968 * account for retransmits accurately.
970 * SACK block validation.
971 * ----------------------
973 * SACK block range validation checks that the received SACK block fits to
974 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
975 * Note that SND.UNA is not included to the range though being valid because
976 * it means that the receiver is rather inconsistent with itself reporting
977 * SACK reneging when it should advance SND.UNA. Such SACK block this is
978 * perfectly valid, however, in light of RFC2018 which explicitly states
979 * that "SACK block MUST reflect the newest segment. Even if the newest
980 * segment is going to be discarded ...", not that it looks very clever
981 * in case of head skb. Due to potentional receiver driven attacks, we
982 * choose to avoid immediate execution of a walk in write queue due to
983 * reneging and defer head skb's loss recovery to standard loss recovery
984 * procedure that will eventually trigger (nothing forbids us doing this).
986 * Implements also blockage to start_seq wrap-around. Problem lies in the
987 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
988 * there's no guarantee that it will be before snd_nxt (n). The problem
989 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
992 * <- outs wnd -> <- wrapzone ->
993 * u e n u_w e_w s n_w
995 * |<------------+------+----- TCP seqno space --------------+---------->|
996 * ...-- <2^31 ->| |<--------...
997 * ...---- >2^31 ------>| |<--------...
999 * Current code wouldn't be vulnerable but it's better still to discard such
1000 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1001 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1002 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1003 * equal to the ideal case (infinite seqno space without wrap caused issues).
1005 * With D-SACK the lower bound is extended to cover sequence space below
1006 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1007 * again, D-SACK block must not to go across snd_una (for the same reason as
1008 * for the normal SACK blocks, explained above). But there all simplicity
1009 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1010 * fully below undo_marker they do not affect behavior in anyway and can
1011 * therefore be safely ignored. In rare cases (which are more or less
1012 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1013 * fragmentation and packet reordering past skb's retransmission. To consider
1014 * them correctly, the acceptable range must be extended even more though
1015 * the exact amount is rather hard to quantify. However, tp->max_window can
1016 * be used as an exaggerated estimate.
1018 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
1019 u32 start_seq, u32 end_seq)
1021 /* Too far in future, or reversed (interpretation is ambiguous) */
1022 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1025 /* Nasty start_seq wrap-around check (see comments above) */
1026 if (!before(start_seq, tp->snd_nxt))
1029 /* In outstanding window? ...This is valid exit for D-SACKs too.
1030 * start_seq == snd_una is non-sensical (see comments above)
1032 if (after(start_seq, tp->snd_una))
1035 if (!is_dsack || !tp->undo_marker)
1038 /* ...Then it's D-SACK, and must reside below snd_una completely */
1039 if (after(end_seq, tp->snd_una))
1042 if (!before(start_seq, tp->undo_marker))
1046 if (!after(end_seq, tp->undo_marker))
1049 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1050 * start_seq < undo_marker and end_seq >= undo_marker.
1052 return !before(start_seq, end_seq - tp->max_window);
1055 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1056 struct tcp_sack_block_wire *sp, int num_sacks,
1059 struct tcp_sock *tp = tcp_sk(sk);
1060 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1061 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1062 bool dup_sack = false;
1064 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1067 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1068 } else if (num_sacks > 1) {
1069 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1070 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1072 if (!after(end_seq_0, end_seq_1) &&
1073 !before(start_seq_0, start_seq_1)) {
1076 NET_INC_STATS_BH(sock_net(sk),
1077 LINUX_MIB_TCPDSACKOFORECV);
1081 /* D-SACK for already forgotten data... Do dumb counting. */
1082 if (dup_sack && tp->undo_marker && tp->undo_retrans > 0 &&
1083 !after(end_seq_0, prior_snd_una) &&
1084 after(end_seq_0, tp->undo_marker))
1090 struct tcp_sacktag_state {
1093 /* Timestamps for earliest and latest never-retransmitted segment
1094 * that was SACKed. RTO needs the earliest RTT to stay conservative,
1095 * but congestion control should still get an accurate delay signal.
1097 struct skb_mstamp first_sackt;
1098 struct skb_mstamp last_sackt;
1102 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1103 * the incoming SACK may not exactly match but we can find smaller MSS
1104 * aligned portion of it that matches. Therefore we might need to fragment
1105 * which may fail and creates some hassle (caller must handle error case
1108 * FIXME: this could be merged to shift decision code
1110 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1111 u32 start_seq, u32 end_seq)
1115 unsigned int pkt_len;
1118 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1119 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1121 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1122 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1123 mss = tcp_skb_mss(skb);
1124 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1127 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1131 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1136 /* Round if necessary so that SACKs cover only full MSSes
1137 * and/or the remaining small portion (if present)
1139 if (pkt_len > mss) {
1140 unsigned int new_len = (pkt_len / mss) * mss;
1141 if (!in_sack && new_len < pkt_len) {
1143 if (new_len >= skb->len)
1148 err = tcp_fragment(sk, skb, pkt_len, mss, GFP_ATOMIC);
1156 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1157 static u8 tcp_sacktag_one(struct sock *sk,
1158 struct tcp_sacktag_state *state, u8 sacked,
1159 u32 start_seq, u32 end_seq,
1160 int dup_sack, int pcount,
1161 const struct skb_mstamp *xmit_time)
1163 struct tcp_sock *tp = tcp_sk(sk);
1164 int fack_count = state->fack_count;
1166 /* Account D-SACK for retransmitted packet. */
1167 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1168 if (tp->undo_marker && tp->undo_retrans > 0 &&
1169 after(end_seq, tp->undo_marker))
1171 if (sacked & TCPCB_SACKED_ACKED)
1172 state->reord = min(fack_count, state->reord);
1175 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1176 if (!after(end_seq, tp->snd_una))
1179 if (!(sacked & TCPCB_SACKED_ACKED)) {
1180 tcp_rack_advance(tp, xmit_time, sacked);
1182 if (sacked & TCPCB_SACKED_RETRANS) {
1183 /* If the segment is not tagged as lost,
1184 * we do not clear RETRANS, believing
1185 * that retransmission is still in flight.
1187 if (sacked & TCPCB_LOST) {
1188 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1189 tp->lost_out -= pcount;
1190 tp->retrans_out -= pcount;
1193 if (!(sacked & TCPCB_RETRANS)) {
1194 /* New sack for not retransmitted frame,
1195 * which was in hole. It is reordering.
1197 if (before(start_seq,
1198 tcp_highest_sack_seq(tp)))
1199 state->reord = min(fack_count,
1201 if (!after(end_seq, tp->high_seq))
1202 state->flag |= FLAG_ORIG_SACK_ACKED;
1203 if (state->first_sackt.v64 == 0)
1204 state->first_sackt = *xmit_time;
1205 state->last_sackt = *xmit_time;
1208 if (sacked & TCPCB_LOST) {
1209 sacked &= ~TCPCB_LOST;
1210 tp->lost_out -= pcount;
1214 sacked |= TCPCB_SACKED_ACKED;
1215 state->flag |= FLAG_DATA_SACKED;
1216 tp->sacked_out += pcount;
1218 fack_count += pcount;
1220 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1221 if (!tcp_is_fack(tp) && tp->lost_skb_hint &&
1222 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1223 tp->lost_cnt_hint += pcount;
1225 if (fack_count > tp->fackets_out)
1226 tp->fackets_out = fack_count;
1229 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1230 * frames and clear it. undo_retrans is decreased above, L|R frames
1231 * are accounted above as well.
1233 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1234 sacked &= ~TCPCB_SACKED_RETRANS;
1235 tp->retrans_out -= pcount;
1241 /* Shift newly-SACKed bytes from this skb to the immediately previous
1242 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1244 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
1245 struct tcp_sacktag_state *state,
1246 unsigned int pcount, int shifted, int mss,
1249 struct tcp_sock *tp = tcp_sk(sk);
1250 struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
1251 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1252 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
1256 /* Adjust counters and hints for the newly sacked sequence
1257 * range but discard the return value since prev is already
1258 * marked. We must tag the range first because the seq
1259 * advancement below implicitly advances
1260 * tcp_highest_sack_seq() when skb is highest_sack.
1262 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1263 start_seq, end_seq, dup_sack, pcount,
1266 if (skb == tp->lost_skb_hint)
1267 tp->lost_cnt_hint += pcount;
1269 TCP_SKB_CB(prev)->end_seq += shifted;
1270 TCP_SKB_CB(skb)->seq += shifted;
1272 tcp_skb_pcount_add(prev, pcount);
1273 BUG_ON(tcp_skb_pcount(skb) < pcount);
1274 tcp_skb_pcount_add(skb, -pcount);
1276 /* When we're adding to gso_segs == 1, gso_size will be zero,
1277 * in theory this shouldn't be necessary but as long as DSACK
1278 * code can come after this skb later on it's better to keep
1279 * setting gso_size to something.
1281 if (!TCP_SKB_CB(prev)->tcp_gso_size)
1282 TCP_SKB_CB(prev)->tcp_gso_size = mss;
1284 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1285 if (tcp_skb_pcount(skb) <= 1)
1286 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1288 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1289 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1292 BUG_ON(!tcp_skb_pcount(skb));
1293 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1297 /* Whole SKB was eaten :-) */
1299 if (skb == tp->retransmit_skb_hint)
1300 tp->retransmit_skb_hint = prev;
1301 if (skb == tp->lost_skb_hint) {
1302 tp->lost_skb_hint = prev;
1303 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1306 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1307 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1308 TCP_SKB_CB(prev)->end_seq++;
1310 if (skb == tcp_highest_sack(sk))
1311 tcp_advance_highest_sack(sk, skb);
1313 tcp_unlink_write_queue(skb, sk);
1314 sk_wmem_free_skb(sk, skb);
1316 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED);
1321 /* I wish gso_size would have a bit more sane initialization than
1322 * something-or-zero which complicates things
1324 static int tcp_skb_seglen(const struct sk_buff *skb)
1326 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1329 /* Shifting pages past head area doesn't work */
1330 static int skb_can_shift(const struct sk_buff *skb)
1332 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1335 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1338 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1339 struct tcp_sacktag_state *state,
1340 u32 start_seq, u32 end_seq,
1343 struct tcp_sock *tp = tcp_sk(sk);
1344 struct sk_buff *prev;
1350 if (!sk_can_gso(sk))
1353 /* Normally R but no L won't result in plain S */
1355 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1357 if (!skb_can_shift(skb))
1359 /* This frame is about to be dropped (was ACKed). */
1360 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1363 /* Can only happen with delayed DSACK + discard craziness */
1364 if (unlikely(skb == tcp_write_queue_head(sk)))
1366 prev = tcp_write_queue_prev(sk, skb);
1368 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1371 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1372 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1376 pcount = tcp_skb_pcount(skb);
1377 mss = tcp_skb_seglen(skb);
1379 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1380 * drop this restriction as unnecessary
1382 if (mss != tcp_skb_seglen(prev))
1385 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1387 /* CHECKME: This is non-MSS split case only?, this will
1388 * cause skipped skbs due to advancing loop btw, original
1389 * has that feature too
1391 if (tcp_skb_pcount(skb) <= 1)
1394 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1396 /* TODO: head merge to next could be attempted here
1397 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1398 * though it might not be worth of the additional hassle
1400 * ...we can probably just fallback to what was done
1401 * previously. We could try merging non-SACKed ones
1402 * as well but it probably isn't going to buy off
1403 * because later SACKs might again split them, and
1404 * it would make skb timestamp tracking considerably
1410 len = end_seq - TCP_SKB_CB(skb)->seq;
1412 BUG_ON(len > skb->len);
1414 /* MSS boundaries should be honoured or else pcount will
1415 * severely break even though it makes things bit trickier.
1416 * Optimize common case to avoid most of the divides
1418 mss = tcp_skb_mss(skb);
1420 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1421 * drop this restriction as unnecessary
1423 if (mss != tcp_skb_seglen(prev))
1428 } else if (len < mss) {
1436 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1437 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1440 if (!skb_shift(prev, skb, len))
1442 if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack))
1445 /* Hole filled allows collapsing with the next as well, this is very
1446 * useful when hole on every nth skb pattern happens
1448 if (prev == tcp_write_queue_tail(sk))
1450 skb = tcp_write_queue_next(sk, prev);
1452 if (!skb_can_shift(skb) ||
1453 (skb == tcp_send_head(sk)) ||
1454 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1455 (mss != tcp_skb_seglen(skb)))
1459 if (skb_shift(prev, skb, len)) {
1460 pcount += tcp_skb_pcount(skb);
1461 tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0);
1465 state->fack_count += pcount;
1472 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1476 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1477 struct tcp_sack_block *next_dup,
1478 struct tcp_sacktag_state *state,
1479 u32 start_seq, u32 end_seq,
1482 struct tcp_sock *tp = tcp_sk(sk);
1483 struct sk_buff *tmp;
1485 tcp_for_write_queue_from(skb, sk) {
1487 bool dup_sack = dup_sack_in;
1489 if (skb == tcp_send_head(sk))
1492 /* queue is in-order => we can short-circuit the walk early */
1493 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1497 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1498 in_sack = tcp_match_skb_to_sack(sk, skb,
1499 next_dup->start_seq,
1505 /* skb reference here is a bit tricky to get right, since
1506 * shifting can eat and free both this skb and the next,
1507 * so not even _safe variant of the loop is enough.
1510 tmp = tcp_shift_skb_data(sk, skb, state,
1511 start_seq, end_seq, dup_sack);
1520 in_sack = tcp_match_skb_to_sack(sk, skb,
1526 if (unlikely(in_sack < 0))
1530 TCP_SKB_CB(skb)->sacked =
1533 TCP_SKB_CB(skb)->sacked,
1534 TCP_SKB_CB(skb)->seq,
1535 TCP_SKB_CB(skb)->end_seq,
1537 tcp_skb_pcount(skb),
1540 if (!before(TCP_SKB_CB(skb)->seq,
1541 tcp_highest_sack_seq(tp)))
1542 tcp_advance_highest_sack(sk, skb);
1545 state->fack_count += tcp_skb_pcount(skb);
1550 /* Avoid all extra work that is being done by sacktag while walking in
1553 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1554 struct tcp_sacktag_state *state,
1557 tcp_for_write_queue_from(skb, sk) {
1558 if (skb == tcp_send_head(sk))
1561 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1564 state->fack_count += tcp_skb_pcount(skb);
1569 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1571 struct tcp_sack_block *next_dup,
1572 struct tcp_sacktag_state *state,
1578 if (before(next_dup->start_seq, skip_to_seq)) {
1579 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1580 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1581 next_dup->start_seq, next_dup->end_seq,
1588 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1590 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1594 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1595 u32 prior_snd_una, struct tcp_sacktag_state *state)
1597 struct tcp_sock *tp = tcp_sk(sk);
1598 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1599 TCP_SKB_CB(ack_skb)->sacked);
1600 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1601 struct tcp_sack_block sp[TCP_NUM_SACKS];
1602 struct tcp_sack_block *cache;
1603 struct sk_buff *skb;
1604 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1606 bool found_dup_sack = false;
1608 int first_sack_index;
1611 state->reord = tp->packets_out;
1613 if (!tp->sacked_out) {
1614 if (WARN_ON(tp->fackets_out))
1615 tp->fackets_out = 0;
1616 tcp_highest_sack_reset(sk);
1619 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1620 num_sacks, prior_snd_una);
1622 state->flag |= FLAG_DSACKING_ACK;
1624 /* Eliminate too old ACKs, but take into
1625 * account more or less fresh ones, they can
1626 * contain valid SACK info.
1628 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1631 if (!tp->packets_out)
1635 first_sack_index = 0;
1636 for (i = 0; i < num_sacks; i++) {
1637 bool dup_sack = !i && found_dup_sack;
1639 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1640 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1642 if (!tcp_is_sackblock_valid(tp, dup_sack,
1643 sp[used_sacks].start_seq,
1644 sp[used_sacks].end_seq)) {
1648 if (!tp->undo_marker)
1649 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1651 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1653 /* Don't count olds caused by ACK reordering */
1654 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1655 !after(sp[used_sacks].end_seq, tp->snd_una))
1657 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1660 NET_INC_STATS_BH(sock_net(sk), mib_idx);
1662 first_sack_index = -1;
1666 /* Ignore very old stuff early */
1667 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1673 /* order SACK blocks to allow in order walk of the retrans queue */
1674 for (i = used_sacks - 1; i > 0; i--) {
1675 for (j = 0; j < i; j++) {
1676 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1677 swap(sp[j], sp[j + 1]);
1679 /* Track where the first SACK block goes to */
1680 if (j == first_sack_index)
1681 first_sack_index = j + 1;
1686 skb = tcp_write_queue_head(sk);
1687 state->fack_count = 0;
1690 if (!tp->sacked_out) {
1691 /* It's already past, so skip checking against it */
1692 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1694 cache = tp->recv_sack_cache;
1695 /* Skip empty blocks in at head of the cache */
1696 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1701 while (i < used_sacks) {
1702 u32 start_seq = sp[i].start_seq;
1703 u32 end_seq = sp[i].end_seq;
1704 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1705 struct tcp_sack_block *next_dup = NULL;
1707 if (found_dup_sack && ((i + 1) == first_sack_index))
1708 next_dup = &sp[i + 1];
1710 /* Skip too early cached blocks */
1711 while (tcp_sack_cache_ok(tp, cache) &&
1712 !before(start_seq, cache->end_seq))
1715 /* Can skip some work by looking recv_sack_cache? */
1716 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1717 after(end_seq, cache->start_seq)) {
1720 if (before(start_seq, cache->start_seq)) {
1721 skb = tcp_sacktag_skip(skb, sk, state,
1723 skb = tcp_sacktag_walk(skb, sk, next_dup,
1730 /* Rest of the block already fully processed? */
1731 if (!after(end_seq, cache->end_seq))
1734 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1738 /* ...tail remains todo... */
1739 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1740 /* ...but better entrypoint exists! */
1741 skb = tcp_highest_sack(sk);
1744 state->fack_count = tp->fackets_out;
1749 skb = tcp_sacktag_skip(skb, sk, state, cache->end_seq);
1750 /* Check overlap against next cached too (past this one already) */
1755 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1756 skb = tcp_highest_sack(sk);
1759 state->fack_count = tp->fackets_out;
1761 skb = tcp_sacktag_skip(skb, sk, state, start_seq);
1764 skb = tcp_sacktag_walk(skb, sk, next_dup, state,
1765 start_seq, end_seq, dup_sack);
1771 /* Clear the head of the cache sack blocks so we can skip it next time */
1772 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1773 tp->recv_sack_cache[i].start_seq = 0;
1774 tp->recv_sack_cache[i].end_seq = 0;
1776 for (j = 0; j < used_sacks; j++)
1777 tp->recv_sack_cache[i++] = sp[j];
1779 if ((state->reord < tp->fackets_out) &&
1780 ((inet_csk(sk)->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker))
1781 tcp_update_reordering(sk, tp->fackets_out - state->reord, 0);
1783 tcp_verify_left_out(tp);
1786 #if FASTRETRANS_DEBUG > 0
1787 WARN_ON((int)tp->sacked_out < 0);
1788 WARN_ON((int)tp->lost_out < 0);
1789 WARN_ON((int)tp->retrans_out < 0);
1790 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1795 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1796 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1798 static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
1802 holes = max(tp->lost_out, 1U);
1803 holes = min(holes, tp->packets_out);
1805 if ((tp->sacked_out + holes) > tp->packets_out) {
1806 tp->sacked_out = tp->packets_out - holes;
1812 /* If we receive more dupacks than we expected counting segments
1813 * in assumption of absent reordering, interpret this as reordering.
1814 * The only another reason could be bug in receiver TCP.
1816 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1818 struct tcp_sock *tp = tcp_sk(sk);
1819 if (tcp_limit_reno_sacked(tp))
1820 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1823 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1825 static void tcp_add_reno_sack(struct sock *sk)
1827 struct tcp_sock *tp = tcp_sk(sk);
1829 tcp_check_reno_reordering(sk, 0);
1830 tcp_verify_left_out(tp);
1833 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1835 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1837 struct tcp_sock *tp = tcp_sk(sk);
1840 /* One ACK acked hole. The rest eat duplicate ACKs. */
1841 if (acked - 1 >= tp->sacked_out)
1844 tp->sacked_out -= acked - 1;
1846 tcp_check_reno_reordering(sk, acked);
1847 tcp_verify_left_out(tp);
1850 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1855 void tcp_clear_retrans(struct tcp_sock *tp)
1857 tp->retrans_out = 0;
1859 tp->undo_marker = 0;
1860 tp->undo_retrans = -1;
1861 tp->fackets_out = 0;
1865 static inline void tcp_init_undo(struct tcp_sock *tp)
1867 tp->undo_marker = tp->snd_una;
1868 /* Retransmission still in flight may cause DSACKs later. */
1869 tp->undo_retrans = tp->retrans_out ? : -1;
1872 /* Enter Loss state. If we detect SACK reneging, forget all SACK information
1873 * and reset tags completely, otherwise preserve SACKs. If receiver
1874 * dropped its ofo queue, we will know this due to reneging detection.
1876 void tcp_enter_loss(struct sock *sk)
1878 const struct inet_connection_sock *icsk = inet_csk(sk);
1879 struct tcp_sock *tp = tcp_sk(sk);
1880 struct sk_buff *skb;
1881 bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery;
1882 bool is_reneg; /* is receiver reneging on SACKs? */
1884 /* Reduce ssthresh if it has not yet been made inside this window. */
1885 if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
1886 !after(tp->high_seq, tp->snd_una) ||
1887 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1888 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1889 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1890 tcp_ca_event(sk, CA_EVENT_LOSS);
1894 tp->snd_cwnd_cnt = 0;
1895 tp->snd_cwnd_stamp = tcp_time_stamp;
1897 tp->retrans_out = 0;
1900 if (tcp_is_reno(tp))
1901 tcp_reset_reno_sack(tp);
1903 skb = tcp_write_queue_head(sk);
1904 is_reneg = skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED);
1906 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
1908 tp->fackets_out = 0;
1910 tcp_clear_all_retrans_hints(tp);
1912 tcp_for_write_queue(skb, sk) {
1913 if (skb == tcp_send_head(sk))
1916 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1917 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || is_reneg) {
1918 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1919 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1920 tp->lost_out += tcp_skb_pcount(skb);
1921 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
1924 tcp_verify_left_out(tp);
1926 /* Timeout in disordered state after receiving substantial DUPACKs
1927 * suggests that the degree of reordering is over-estimated.
1929 if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
1930 tp->sacked_out >= sysctl_tcp_reordering)
1931 tp->reordering = min_t(unsigned int, tp->reordering,
1932 sysctl_tcp_reordering);
1933 tcp_set_ca_state(sk, TCP_CA_Loss);
1934 tp->high_seq = tp->snd_nxt;
1935 tcp_ecn_queue_cwr(tp);
1937 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
1938 * loss recovery is underway except recurring timeout(s) on
1939 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
1941 tp->frto = sysctl_tcp_frto &&
1942 (new_recovery || icsk->icsk_retransmits) &&
1943 !inet_csk(sk)->icsk_mtup.probe_size;
1946 /* If ACK arrived pointing to a remembered SACK, it means that our
1947 * remembered SACKs do not reflect real state of receiver i.e.
1948 * receiver _host_ is heavily congested (or buggy).
1950 * To avoid big spurious retransmission bursts due to transient SACK
1951 * scoreboard oddities that look like reneging, we give the receiver a
1952 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
1953 * restore sanity to the SACK scoreboard. If the apparent reneging
1954 * persists until this RTO then we'll clear the SACK scoreboard.
1956 static bool tcp_check_sack_reneging(struct sock *sk, int flag)
1958 if (flag & FLAG_SACK_RENEGING) {
1959 struct tcp_sock *tp = tcp_sk(sk);
1960 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4),
1961 msecs_to_jiffies(10));
1963 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1964 delay, TCP_RTO_MAX);
1970 static inline int tcp_fackets_out(const struct tcp_sock *tp)
1972 return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
1975 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
1976 * counter when SACK is enabled (without SACK, sacked_out is used for
1979 * Instead, with FACK TCP uses fackets_out that includes both SACKed
1980 * segments up to the highest received SACK block so far and holes in
1983 * With reordering, holes may still be in flight, so RFC3517 recovery
1984 * uses pure sacked_out (total number of SACKed segments) even though
1985 * it violates the RFC that uses duplicate ACKs, often these are equal
1986 * but when e.g. out-of-window ACKs or packet duplication occurs,
1987 * they differ. Since neither occurs due to loss, TCP should really
1990 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
1992 return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
1995 static bool tcp_pause_early_retransmit(struct sock *sk, int flag)
1997 struct tcp_sock *tp = tcp_sk(sk);
1998 unsigned long delay;
2000 /* Delay early retransmit and entering fast recovery for
2001 * max(RTT/4, 2msec) unless ack has ECE mark, no RTT samples
2002 * available, or RTO is scheduled to fire first.
2004 if (sysctl_tcp_early_retrans < 2 || sysctl_tcp_early_retrans > 3 ||
2005 (flag & FLAG_ECE) || !tp->srtt_us)
2008 delay = max(usecs_to_jiffies(tp->srtt_us >> 5),
2009 msecs_to_jiffies(2));
2011 if (!time_after(inet_csk(sk)->icsk_timeout, (jiffies + delay)))
2014 inet_csk_reset_xmit_timer(sk, ICSK_TIME_EARLY_RETRANS, delay,
2019 /* Linux NewReno/SACK/FACK/ECN state machine.
2020 * --------------------------------------
2022 * "Open" Normal state, no dubious events, fast path.
2023 * "Disorder" In all the respects it is "Open",
2024 * but requires a bit more attention. It is entered when
2025 * we see some SACKs or dupacks. It is split of "Open"
2026 * mainly to move some processing from fast path to slow one.
2027 * "CWR" CWND was reduced due to some Congestion Notification event.
2028 * It can be ECN, ICMP source quench, local device congestion.
2029 * "Recovery" CWND was reduced, we are fast-retransmitting.
2030 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2032 * tcp_fastretrans_alert() is entered:
2033 * - each incoming ACK, if state is not "Open"
2034 * - when arrived ACK is unusual, namely:
2039 * Counting packets in flight is pretty simple.
2041 * in_flight = packets_out - left_out + retrans_out
2043 * packets_out is SND.NXT-SND.UNA counted in packets.
2045 * retrans_out is number of retransmitted segments.
2047 * left_out is number of segments left network, but not ACKed yet.
2049 * left_out = sacked_out + lost_out
2051 * sacked_out: Packets, which arrived to receiver out of order
2052 * and hence not ACKed. With SACKs this number is simply
2053 * amount of SACKed data. Even without SACKs
2054 * it is easy to give pretty reliable estimate of this number,
2055 * counting duplicate ACKs.
2057 * lost_out: Packets lost by network. TCP has no explicit
2058 * "loss notification" feedback from network (for now).
2059 * It means that this number can be only _guessed_.
2060 * Actually, it is the heuristics to predict lossage that
2061 * distinguishes different algorithms.
2063 * F.e. after RTO, when all the queue is considered as lost,
2064 * lost_out = packets_out and in_flight = retrans_out.
2066 * Essentially, we have now two algorithms counting
2069 * FACK: It is the simplest heuristics. As soon as we decided
2070 * that something is lost, we decide that _all_ not SACKed
2071 * packets until the most forward SACK are lost. I.e.
2072 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2073 * It is absolutely correct estimate, if network does not reorder
2074 * packets. And it loses any connection to reality when reordering
2075 * takes place. We use FACK by default until reordering
2076 * is suspected on the path to this destination.
2078 * NewReno: when Recovery is entered, we assume that one segment
2079 * is lost (classic Reno). While we are in Recovery and
2080 * a partial ACK arrives, we assume that one more packet
2081 * is lost (NewReno). This heuristics are the same in NewReno
2084 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2085 * deflation etc. CWND is real congestion window, never inflated, changes
2086 * only according to classic VJ rules.
2088 * Really tricky (and requiring careful tuning) part of algorithm
2089 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2090 * The first determines the moment _when_ we should reduce CWND and,
2091 * hence, slow down forward transmission. In fact, it determines the moment
2092 * when we decide that hole is caused by loss, rather than by a reorder.
2094 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2095 * holes, caused by lost packets.
2097 * And the most logically complicated part of algorithm is undo
2098 * heuristics. We detect false retransmits due to both too early
2099 * fast retransmit (reordering) and underestimated RTO, analyzing
2100 * timestamps and D-SACKs. When we detect that some segments were
2101 * retransmitted by mistake and CWND reduction was wrong, we undo
2102 * window reduction and abort recovery phase. This logic is hidden
2103 * inside several functions named tcp_try_undo_<something>.
2106 /* This function decides, when we should leave Disordered state
2107 * and enter Recovery phase, reducing congestion window.
2109 * Main question: may we further continue forward transmission
2110 * with the same cwnd?
2112 static bool tcp_time_to_recover(struct sock *sk, int flag)
2114 struct tcp_sock *tp = tcp_sk(sk);
2117 /* Trick#1: The loss is proven. */
2121 /* Not-A-Trick#2 : Classic rule... */
2122 if (tcp_dupack_heuristics(tp) > tp->reordering)
2125 /* Trick#4: It is still not OK... But will it be useful to delay
2128 packets_out = tp->packets_out;
2129 if (packets_out <= tp->reordering &&
2130 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
2131 !tcp_may_send_now(sk)) {
2132 /* We have nothing to send. This connection is limited
2133 * either by receiver window or by application.
2138 /* If a thin stream is detected, retransmit after first
2139 * received dupack. Employ only if SACK is supported in order
2140 * to avoid possible corner-case series of spurious retransmissions
2141 * Use only if there are no unsent data.
2143 if ((tp->thin_dupack || sysctl_tcp_thin_dupack) &&
2144 tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 &&
2145 tcp_is_sack(tp) && !tcp_send_head(sk))
2148 /* Trick#6: TCP early retransmit, per RFC5827. To avoid spurious
2149 * retransmissions due to small network reorderings, we implement
2150 * Mitigation A.3 in the RFC and delay the retransmission for a short
2151 * interval if appropriate.
2153 if (tp->do_early_retrans && !tp->retrans_out && tp->sacked_out &&
2154 (tp->packets_out >= (tp->sacked_out + 1) && tp->packets_out < 4) &&
2155 !tcp_may_send_now(sk))
2156 return !tcp_pause_early_retransmit(sk, flag);
2161 /* Detect loss in event "A" above by marking head of queue up as lost.
2162 * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2163 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2164 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2165 * the maximum SACKed segments to pass before reaching this limit.
2167 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2169 struct tcp_sock *tp = tcp_sk(sk);
2170 struct sk_buff *skb;
2171 int cnt, oldcnt, lost;
2173 /* Use SACK to deduce losses of new sequences sent during recovery */
2174 const u32 loss_high = tcp_is_sack(tp) ? tp->snd_nxt : tp->high_seq;
2176 WARN_ON(packets > tp->packets_out);
2177 if (tp->lost_skb_hint) {
2178 skb = tp->lost_skb_hint;
2179 cnt = tp->lost_cnt_hint;
2180 /* Head already handled? */
2181 if (mark_head && skb != tcp_write_queue_head(sk))
2184 skb = tcp_write_queue_head(sk);
2188 tcp_for_write_queue_from(skb, sk) {
2189 if (skb == tcp_send_head(sk))
2191 /* TODO: do this better */
2192 /* this is not the most efficient way to do this... */
2193 tp->lost_skb_hint = skb;
2194 tp->lost_cnt_hint = cnt;
2196 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2200 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2201 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2202 cnt += tcp_skb_pcount(skb);
2204 if (cnt > packets) {
2205 if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) ||
2206 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
2207 (oldcnt >= packets))
2210 mss = tcp_skb_mss(skb);
2211 /* If needed, chop off the prefix to mark as lost. */
2212 lost = (packets - oldcnt) * mss;
2213 if (lost < skb->len &&
2214 tcp_fragment(sk, skb, lost, mss, GFP_ATOMIC) < 0)
2219 tcp_skb_mark_lost(tp, skb);
2224 tcp_verify_left_out(tp);
2227 /* Account newly detected lost packet(s) */
2229 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2231 struct tcp_sock *tp = tcp_sk(sk);
2233 if (tcp_is_reno(tp)) {
2234 tcp_mark_head_lost(sk, 1, 1);
2235 } else if (tcp_is_fack(tp)) {
2236 int lost = tp->fackets_out - tp->reordering;
2239 tcp_mark_head_lost(sk, lost, 0);
2241 int sacked_upto = tp->sacked_out - tp->reordering;
2242 if (sacked_upto >= 0)
2243 tcp_mark_head_lost(sk, sacked_upto, 0);
2244 else if (fast_rexmit)
2245 tcp_mark_head_lost(sk, 1, 1);
2249 /* CWND moderation, preventing bursts due to too big ACKs
2250 * in dubious situations.
2252 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
2254 tp->snd_cwnd = min(tp->snd_cwnd,
2255 tcp_packets_in_flight(tp) + tcp_max_burst(tp));
2256 tp->snd_cwnd_stamp = tcp_time_stamp;
2259 static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when)
2261 return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2262 before(tp->rx_opt.rcv_tsecr, when);
2265 /* skb is spurious retransmitted if the returned timestamp echo
2266 * reply is prior to the skb transmission time
2268 static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp,
2269 const struct sk_buff *skb)
2271 return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) &&
2272 tcp_tsopt_ecr_before(tp, tcp_skb_timestamp(skb));
2275 /* Nothing was retransmitted or returned timestamp is less
2276 * than timestamp of the first retransmission.
2278 static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2280 return !tp->retrans_stamp ||
2281 tcp_tsopt_ecr_before(tp, tp->retrans_stamp);
2284 /* Undo procedures. */
2286 /* We can clear retrans_stamp when there are no retransmissions in the
2287 * window. It would seem that it is trivially available for us in
2288 * tp->retrans_out, however, that kind of assumptions doesn't consider
2289 * what will happen if errors occur when sending retransmission for the
2290 * second time. ...It could the that such segment has only
2291 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2292 * the head skb is enough except for some reneging corner cases that
2293 * are not worth the effort.
2295 * Main reason for all this complexity is the fact that connection dying
2296 * time now depends on the validity of the retrans_stamp, in particular,
2297 * that successive retransmissions of a segment must not advance
2298 * retrans_stamp under any conditions.
2300 static bool tcp_any_retrans_done(const struct sock *sk)
2302 const struct tcp_sock *tp = tcp_sk(sk);
2303 struct sk_buff *skb;
2305 if (tp->retrans_out)
2308 skb = tcp_write_queue_head(sk);
2309 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2315 #if FASTRETRANS_DEBUG > 1
2316 static void DBGUNDO(struct sock *sk, const char *msg)
2318 struct tcp_sock *tp = tcp_sk(sk);
2319 struct inet_sock *inet = inet_sk(sk);
2321 if (sk->sk_family == AF_INET) {
2322 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2324 &inet->inet_daddr, ntohs(inet->inet_dport),
2325 tp->snd_cwnd, tcp_left_out(tp),
2326 tp->snd_ssthresh, tp->prior_ssthresh,
2329 #if IS_ENABLED(CONFIG_IPV6)
2330 else if (sk->sk_family == AF_INET6) {
2331 struct ipv6_pinfo *np = inet6_sk(sk);
2332 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2334 &np->daddr, ntohs(inet->inet_dport),
2335 tp->snd_cwnd, tcp_left_out(tp),
2336 tp->snd_ssthresh, tp->prior_ssthresh,
2342 #define DBGUNDO(x...) do { } while (0)
2345 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
2347 struct tcp_sock *tp = tcp_sk(sk);
2350 struct sk_buff *skb;
2352 tcp_for_write_queue(skb, sk) {
2353 if (skb == tcp_send_head(sk))
2355 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2358 tcp_clear_all_retrans_hints(tp);
2361 if (tp->prior_ssthresh) {
2362 const struct inet_connection_sock *icsk = inet_csk(sk);
2364 if (icsk->icsk_ca_ops->undo_cwnd)
2365 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2367 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
2369 if (tp->prior_ssthresh > tp->snd_ssthresh) {
2370 tp->snd_ssthresh = tp->prior_ssthresh;
2371 tcp_ecn_withdraw_cwr(tp);
2374 tp->snd_cwnd_stamp = tcp_time_stamp;
2375 tp->undo_marker = 0;
2378 static inline bool tcp_may_undo(const struct tcp_sock *tp)
2380 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2383 /* People celebrate: "We love our President!" */
2384 static bool tcp_try_undo_recovery(struct sock *sk)
2386 struct tcp_sock *tp = tcp_sk(sk);
2388 if (tcp_may_undo(tp)) {
2391 /* Happy end! We did not retransmit anything
2392 * or our original transmission succeeded.
2394 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2395 tcp_undo_cwnd_reduction(sk, false);
2396 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2397 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2399 mib_idx = LINUX_MIB_TCPFULLUNDO;
2401 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2403 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2404 /* Hold old state until something *above* high_seq
2405 * is ACKed. For Reno it is MUST to prevent false
2406 * fast retransmits (RFC2582). SACK TCP is safe. */
2407 tcp_moderate_cwnd(tp);
2408 if (!tcp_any_retrans_done(sk))
2409 tp->retrans_stamp = 0;
2412 tcp_set_ca_state(sk, TCP_CA_Open);
2416 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2417 static bool tcp_try_undo_dsack(struct sock *sk)
2419 struct tcp_sock *tp = tcp_sk(sk);
2421 if (tp->undo_marker && !tp->undo_retrans) {
2422 DBGUNDO(sk, "D-SACK");
2423 tcp_undo_cwnd_reduction(sk, false);
2424 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2430 /* Undo during loss recovery after partial ACK or using F-RTO. */
2431 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2433 struct tcp_sock *tp = tcp_sk(sk);
2435 if (frto_undo || tcp_may_undo(tp)) {
2436 tcp_undo_cwnd_reduction(sk, true);
2438 DBGUNDO(sk, "partial loss");
2439 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2441 NET_INC_STATS_BH(sock_net(sk),
2442 LINUX_MIB_TCPSPURIOUSRTOS);
2443 inet_csk(sk)->icsk_retransmits = 0;
2444 if (frto_undo || tcp_is_sack(tp))
2445 tcp_set_ca_state(sk, TCP_CA_Open);
2451 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
2452 * It computes the number of packets to send (sndcnt) based on packets newly
2454 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2455 * cwnd reductions across a full RTT.
2456 * 2) Otherwise PRR uses packet conservation to send as much as delivered.
2457 * But when the retransmits are acked without further losses, PRR
2458 * slow starts cwnd up to ssthresh to speed up the recovery.
2460 static void tcp_init_cwnd_reduction(struct sock *sk)
2462 struct tcp_sock *tp = tcp_sk(sk);
2464 tp->high_seq = tp->snd_nxt;
2465 tp->tlp_high_seq = 0;
2466 tp->snd_cwnd_cnt = 0;
2467 tp->prior_cwnd = tp->snd_cwnd;
2468 tp->prr_delivered = 0;
2470 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2471 tcp_ecn_queue_cwr(tp);
2474 static void tcp_cwnd_reduction(struct sock *sk, const int prior_unsacked,
2475 int fast_rexmit, int flag)
2477 struct tcp_sock *tp = tcp_sk(sk);
2479 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2480 int newly_acked_sacked = prior_unsacked -
2481 (tp->packets_out - tp->sacked_out);
2483 if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd))
2486 tp->prr_delivered += newly_acked_sacked;
2488 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2490 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2491 } else if ((flag & FLAG_RETRANS_DATA_ACKED) &&
2492 !(flag & FLAG_LOST_RETRANS)) {
2493 sndcnt = min_t(int, delta,
2494 max_t(int, tp->prr_delivered - tp->prr_out,
2495 newly_acked_sacked) + 1);
2497 sndcnt = min(delta, newly_acked_sacked);
2499 sndcnt = max(sndcnt, (fast_rexmit ? 1 : 0));
2500 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2503 static inline void tcp_end_cwnd_reduction(struct sock *sk)
2505 struct tcp_sock *tp = tcp_sk(sk);
2507 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2508 if (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR ||
2509 (tp->undo_marker && tp->snd_ssthresh < TCP_INFINITE_SSTHRESH)) {
2510 tp->snd_cwnd = tp->snd_ssthresh;
2511 tp->snd_cwnd_stamp = tcp_time_stamp;
2513 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2516 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2517 void tcp_enter_cwr(struct sock *sk)
2519 struct tcp_sock *tp = tcp_sk(sk);
2521 tp->prior_ssthresh = 0;
2522 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2523 tp->undo_marker = 0;
2524 tcp_init_cwnd_reduction(sk);
2525 tcp_set_ca_state(sk, TCP_CA_CWR);
2528 EXPORT_SYMBOL(tcp_enter_cwr);
2530 static void tcp_try_keep_open(struct sock *sk)
2532 struct tcp_sock *tp = tcp_sk(sk);
2533 int state = TCP_CA_Open;
2535 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2536 state = TCP_CA_Disorder;
2538 if (inet_csk(sk)->icsk_ca_state != state) {
2539 tcp_set_ca_state(sk, state);
2540 tp->high_seq = tp->snd_nxt;
2544 static void tcp_try_to_open(struct sock *sk, int flag, const int prior_unsacked)
2546 struct tcp_sock *tp = tcp_sk(sk);
2548 tcp_verify_left_out(tp);
2550 if (!tcp_any_retrans_done(sk))
2551 tp->retrans_stamp = 0;
2553 if (flag & FLAG_ECE)
2556 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2557 tcp_try_keep_open(sk);
2559 tcp_cwnd_reduction(sk, prior_unsacked, 0, flag);
2563 static void tcp_mtup_probe_failed(struct sock *sk)
2565 struct inet_connection_sock *icsk = inet_csk(sk);
2567 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2568 icsk->icsk_mtup.probe_size = 0;
2569 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPMTUPFAIL);
2572 static void tcp_mtup_probe_success(struct sock *sk)
2574 struct tcp_sock *tp = tcp_sk(sk);
2575 struct inet_connection_sock *icsk = inet_csk(sk);
2577 /* FIXME: breaks with very large cwnd */
2578 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2579 tp->snd_cwnd = tp->snd_cwnd *
2580 tcp_mss_to_mtu(sk, tp->mss_cache) /
2581 icsk->icsk_mtup.probe_size;
2582 tp->snd_cwnd_cnt = 0;
2583 tp->snd_cwnd_stamp = tcp_time_stamp;
2584 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2586 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2587 icsk->icsk_mtup.probe_size = 0;
2588 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2589 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS);
2592 /* Do a simple retransmit without using the backoff mechanisms in
2593 * tcp_timer. This is used for path mtu discovery.
2594 * The socket is already locked here.
2596 void tcp_simple_retransmit(struct sock *sk)
2598 const struct inet_connection_sock *icsk = inet_csk(sk);
2599 struct tcp_sock *tp = tcp_sk(sk);
2600 struct sk_buff *skb;
2601 unsigned int mss = tcp_current_mss(sk);
2602 u32 prior_lost = tp->lost_out;
2604 tcp_for_write_queue(skb, sk) {
2605 if (skb == tcp_send_head(sk))
2607 if (tcp_skb_seglen(skb) > mss &&
2608 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2609 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2610 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2611 tp->retrans_out -= tcp_skb_pcount(skb);
2613 tcp_skb_mark_lost_uncond_verify(tp, skb);
2617 tcp_clear_retrans_hints_partial(tp);
2619 if (prior_lost == tp->lost_out)
2622 if (tcp_is_reno(tp))
2623 tcp_limit_reno_sacked(tp);
2625 tcp_verify_left_out(tp);
2627 /* Don't muck with the congestion window here.
2628 * Reason is that we do not increase amount of _data_
2629 * in network, but units changed and effective
2630 * cwnd/ssthresh really reduced now.
2632 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2633 tp->high_seq = tp->snd_nxt;
2634 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2635 tp->prior_ssthresh = 0;
2636 tp->undo_marker = 0;
2637 tcp_set_ca_state(sk, TCP_CA_Loss);
2639 tcp_xmit_retransmit_queue(sk);
2641 EXPORT_SYMBOL(tcp_simple_retransmit);
2643 static void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2645 struct tcp_sock *tp = tcp_sk(sk);
2648 if (tcp_is_reno(tp))
2649 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2651 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2653 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2655 tp->prior_ssthresh = 0;
2658 if (!tcp_in_cwnd_reduction(sk)) {
2660 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2661 tcp_init_cwnd_reduction(sk);
2663 tcp_set_ca_state(sk, TCP_CA_Recovery);
2666 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2667 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2669 static void tcp_process_loss(struct sock *sk, int flag, bool is_dupack,
2672 struct tcp_sock *tp = tcp_sk(sk);
2673 bool recovered = !before(tp->snd_una, tp->high_seq);
2675 if ((flag & FLAG_SND_UNA_ADVANCED) &&
2676 tcp_try_undo_loss(sk, false))
2679 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2680 /* Step 3.b. A timeout is spurious if not all data are
2681 * lost, i.e., never-retransmitted data are (s)acked.
2683 if ((flag & FLAG_ORIG_SACK_ACKED) &&
2684 tcp_try_undo_loss(sk, true))
2687 if (after(tp->snd_nxt, tp->high_seq)) {
2688 if (flag & FLAG_DATA_SACKED || is_dupack)
2689 tp->frto = 0; /* Step 3.a. loss was real */
2690 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2691 tp->high_seq = tp->snd_nxt;
2692 /* Step 2.b. Try send new data (but deferred until cwnd
2693 * is updated in tcp_ack()). Otherwise fall back to
2694 * the conventional recovery.
2696 if (tcp_send_head(sk) &&
2697 after(tcp_wnd_end(tp), tp->snd_nxt)) {
2698 *rexmit = REXMIT_NEW;
2706 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2707 tcp_try_undo_recovery(sk);
2710 if (tcp_is_reno(tp)) {
2711 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2712 * delivered. Lower inflight to clock out (re)tranmissions.
2714 if (after(tp->snd_nxt, tp->high_seq) && is_dupack)
2715 tcp_add_reno_sack(sk);
2716 else if (flag & FLAG_SND_UNA_ADVANCED)
2717 tcp_reset_reno_sack(tp);
2719 *rexmit = REXMIT_LOST;
2722 /* Undo during fast recovery after partial ACK. */
2723 static bool tcp_try_undo_partial(struct sock *sk, const int acked,
2724 const int prior_unsacked, int flag)
2726 struct tcp_sock *tp = tcp_sk(sk);
2728 if (tp->undo_marker && tcp_packet_delayed(tp)) {
2729 /* Plain luck! Hole if filled with delayed
2730 * packet, rather than with a retransmit.
2732 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2734 /* We are getting evidence that the reordering degree is higher
2735 * than we realized. If there are no retransmits out then we
2736 * can undo. Otherwise we clock out new packets but do not
2737 * mark more packets lost or retransmit more.
2739 if (tp->retrans_out) {
2740 tcp_cwnd_reduction(sk, prior_unsacked, 0, flag);
2744 if (!tcp_any_retrans_done(sk))
2745 tp->retrans_stamp = 0;
2747 DBGUNDO(sk, "partial recovery");
2748 tcp_undo_cwnd_reduction(sk, true);
2749 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2750 tcp_try_keep_open(sk);
2756 /* Process an event, which can update packets-in-flight not trivially.
2757 * Main goal of this function is to calculate new estimate for left_out,
2758 * taking into account both packets sitting in receiver's buffer and
2759 * packets lost by network.
2761 * Besides that it does CWND reduction, when packet loss is detected
2762 * and changes state of machine.
2764 * It does _not_ decide what to send, it is made in function
2765 * tcp_xmit_retransmit_queue().
2767 static void tcp_fastretrans_alert(struct sock *sk, const int acked,
2768 const int prior_unsacked,
2769 bool is_dupack, int flag, int *rexmit)
2771 struct inet_connection_sock *icsk = inet_csk(sk);
2772 struct tcp_sock *tp = tcp_sk(sk);
2773 bool do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2774 (tcp_fackets_out(tp) > tp->reordering));
2775 int fast_rexmit = 0;
2777 if (WARN_ON(!tp->packets_out && tp->sacked_out))
2779 if (WARN_ON(!tp->sacked_out && tp->fackets_out))
2780 tp->fackets_out = 0;
2782 /* Now state machine starts.
2783 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2784 if (flag & FLAG_ECE)
2785 tp->prior_ssthresh = 0;
2787 /* B. In all the states check for reneging SACKs. */
2788 if (tcp_check_sack_reneging(sk, flag))
2791 /* C. Check consistency of the current state. */
2792 tcp_verify_left_out(tp);
2794 /* D. Check state exit conditions. State can be terminated
2795 * when high_seq is ACKed. */
2796 if (icsk->icsk_ca_state == TCP_CA_Open) {
2797 WARN_ON(tp->retrans_out != 0);
2798 tp->retrans_stamp = 0;
2799 } else if (!before(tp->snd_una, tp->high_seq)) {
2800 switch (icsk->icsk_ca_state) {
2802 /* CWR is to be held something *above* high_seq
2803 * is ACKed for CWR bit to reach receiver. */
2804 if (tp->snd_una != tp->high_seq) {
2805 tcp_end_cwnd_reduction(sk);
2806 tcp_set_ca_state(sk, TCP_CA_Open);
2810 case TCP_CA_Recovery:
2811 if (tcp_is_reno(tp))
2812 tcp_reset_reno_sack(tp);
2813 if (tcp_try_undo_recovery(sk))
2815 tcp_end_cwnd_reduction(sk);
2820 /* Use RACK to detect loss */
2821 if (sysctl_tcp_recovery & TCP_RACK_LOST_RETRANS &&
2822 tcp_rack_mark_lost(sk))
2823 flag |= FLAG_LOST_RETRANS;
2825 /* E. Process state. */
2826 switch (icsk->icsk_ca_state) {
2827 case TCP_CA_Recovery:
2828 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2829 if (tcp_is_reno(tp) && is_dupack)
2830 tcp_add_reno_sack(sk);
2832 if (tcp_try_undo_partial(sk, acked, prior_unsacked, flag))
2834 /* Partial ACK arrived. Force fast retransmit. */
2835 do_lost = tcp_is_reno(tp) ||
2836 tcp_fackets_out(tp) > tp->reordering;
2838 if (tcp_try_undo_dsack(sk)) {
2839 tcp_try_keep_open(sk);
2844 tcp_process_loss(sk, flag, is_dupack, rexmit);
2845 if (icsk->icsk_ca_state != TCP_CA_Open &&
2846 !(flag & FLAG_LOST_RETRANS))
2848 /* Change state if cwnd is undone or retransmits are lost */
2850 if (tcp_is_reno(tp)) {
2851 if (flag & FLAG_SND_UNA_ADVANCED)
2852 tcp_reset_reno_sack(tp);
2854 tcp_add_reno_sack(sk);
2857 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
2858 tcp_try_undo_dsack(sk);
2860 if (!tcp_time_to_recover(sk, flag)) {
2861 tcp_try_to_open(sk, flag, prior_unsacked);
2865 /* MTU probe failure: don't reduce cwnd */
2866 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2867 icsk->icsk_mtup.probe_size &&
2868 tp->snd_una == tp->mtu_probe.probe_seq_start) {
2869 tcp_mtup_probe_failed(sk);
2870 /* Restores the reduction we did in tcp_mtup_probe() */
2872 tcp_simple_retransmit(sk);
2876 /* Otherwise enter Recovery state */
2877 tcp_enter_recovery(sk, (flag & FLAG_ECE));
2882 tcp_update_scoreboard(sk, fast_rexmit);
2883 tcp_cwnd_reduction(sk, prior_unsacked, fast_rexmit, flag);
2884 *rexmit = REXMIT_LOST;
2887 /* Kathleen Nichols' algorithm for tracking the minimum value of
2888 * a data stream over some fixed time interval. (E.g., the minimum
2889 * RTT over the past five minutes.) It uses constant space and constant
2890 * time per update yet almost always delivers the same minimum as an
2891 * implementation that has to keep all the data in the window.
2893 * The algorithm keeps track of the best, 2nd best & 3rd best min
2894 * values, maintaining an invariant that the measurement time of the
2895 * n'th best >= n-1'th best. It also makes sure that the three values
2896 * are widely separated in the time window since that bounds the worse
2897 * case error when that data is monotonically increasing over the window.
2899 * Upon getting a new min, we can forget everything earlier because it
2900 * has no value - the new min is <= everything else in the window by
2901 * definition and it's the most recent. So we restart fresh on every new min
2902 * and overwrites 2nd & 3rd choices. The same property holds for 2nd & 3rd
2905 static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us)
2907 const u32 now = tcp_time_stamp, wlen = sysctl_tcp_min_rtt_wlen * HZ;
2908 struct rtt_meas *m = tcp_sk(sk)->rtt_min;
2909 struct rtt_meas rttm = { .rtt = (rtt_us ? : 1), .ts = now };
2912 /* Check if the new measurement updates the 1st, 2nd, or 3rd choices */
2913 if (unlikely(rttm.rtt <= m[0].rtt))
2914 m[0] = m[1] = m[2] = rttm;
2915 else if (rttm.rtt <= m[1].rtt)
2917 else if (rttm.rtt <= m[2].rtt)
2920 elapsed = now - m[0].ts;
2921 if (unlikely(elapsed > wlen)) {
2922 /* Passed entire window without a new min so make 2nd choice
2923 * the new min & 3rd choice the new 2nd. So forth and so on.
2928 if (now - m[0].ts > wlen) {
2931 if (now - m[0].ts > wlen)
2934 } else if (m[1].ts == m[0].ts && elapsed > wlen / 4) {
2935 /* Passed a quarter of the window without a new min so
2936 * take 2nd choice from the 2nd quarter of the window.
2939 } else if (m[2].ts == m[1].ts && elapsed > wlen / 2) {
2940 /* Passed half the window without a new min so take the 3rd
2941 * choice from the last half of the window.
2947 static inline bool tcp_ack_update_rtt(struct sock *sk, const int flag,
2948 long seq_rtt_us, long sack_rtt_us,
2951 const struct tcp_sock *tp = tcp_sk(sk);
2953 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2954 * broken middle-boxes or peers may corrupt TS-ECR fields. But
2955 * Karn's algorithm forbids taking RTT if some retransmitted data
2956 * is acked (RFC6298).
2959 seq_rtt_us = sack_rtt_us;
2961 /* RTTM Rule: A TSecr value received in a segment is used to
2962 * update the averaged RTT measurement only if the segment
2963 * acknowledges some new data, i.e., only if it advances the
2964 * left edge of the send window.
2965 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2967 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2969 seq_rtt_us = ca_rtt_us = jiffies_to_usecs(tcp_time_stamp -
2970 tp->rx_opt.rcv_tsecr);
2974 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
2975 * always taken together with ACK, SACK, or TS-opts. Any negative
2976 * values will be skipped with the seq_rtt_us < 0 check above.
2978 tcp_update_rtt_min(sk, ca_rtt_us);
2979 tcp_rtt_estimator(sk, seq_rtt_us);
2982 /* RFC6298: only reset backoff on valid RTT measurement. */
2983 inet_csk(sk)->icsk_backoff = 0;
2987 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
2988 void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req)
2992 if (req && !req->num_retrans && tcp_rsk(req)->snt_synack.v64) {
2993 struct skb_mstamp now;
2995 skb_mstamp_get(&now);
2996 rtt_us = skb_mstamp_us_delta(&now, &tcp_rsk(req)->snt_synack);
2999 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us);
3003 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
3005 const struct inet_connection_sock *icsk = inet_csk(sk);
3007 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
3008 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
3011 /* Restart timer after forward progress on connection.
3012 * RFC2988 recommends to restart timer to now+rto.
3014 void tcp_rearm_rto(struct sock *sk)
3016 const struct inet_connection_sock *icsk = inet_csk(sk);
3017 struct tcp_sock *tp = tcp_sk(sk);
3019 /* If the retrans timer is currently being used by Fast Open
3020 * for SYN-ACK retrans purpose, stay put.
3022 if (tp->fastopen_rsk)
3025 if (!tp->packets_out) {
3026 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
3028 u32 rto = inet_csk(sk)->icsk_rto;
3029 /* Offset the time elapsed after installing regular RTO */
3030 if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
3031 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
3032 struct sk_buff *skb = tcp_write_queue_head(sk);
3033 const u32 rto_time_stamp =
3034 tcp_skb_timestamp(skb) + rto;
3035 s32 delta = (s32)(rto_time_stamp - tcp_time_stamp);
3036 /* delta may not be positive if the socket is locked
3037 * when the retrans timer fires and is rescheduled.
3042 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
3047 /* This function is called when the delayed ER timer fires. TCP enters
3048 * fast recovery and performs fast-retransmit.
3050 void tcp_resume_early_retransmit(struct sock *sk)
3052 struct tcp_sock *tp = tcp_sk(sk);
3056 /* Stop if ER is disabled after the delayed ER timer is scheduled */
3057 if (!tp->do_early_retrans)
3060 tcp_enter_recovery(sk, false);
3061 tcp_update_scoreboard(sk, 1);
3062 tcp_xmit_retransmit_queue(sk);
3065 /* If we get here, the whole TSO packet has not been acked. */
3066 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3068 struct tcp_sock *tp = tcp_sk(sk);
3071 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3073 packets_acked = tcp_skb_pcount(skb);
3074 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3076 packets_acked -= tcp_skb_pcount(skb);
3078 if (packets_acked) {
3079 BUG_ON(tcp_skb_pcount(skb) == 0);
3080 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3083 return packets_acked;
3086 static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
3089 const struct skb_shared_info *shinfo;
3091 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3092 if (likely(!(sk->sk_tsflags & SOF_TIMESTAMPING_TX_ACK)))
3095 shinfo = skb_shinfo(skb);
3096 if ((shinfo->tx_flags & SKBTX_ACK_TSTAMP) &&
3097 between(shinfo->tskey, prior_snd_una, tcp_sk(sk)->snd_una - 1))
3098 __skb_tstamp_tx(skb, NULL, sk, SCM_TSTAMP_ACK);
3101 /* Remove acknowledged frames from the retransmission queue. If our packet
3102 * is before the ack sequence we can discard it as it's confirmed to have
3103 * arrived at the other end.
3105 static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
3107 struct tcp_sacktag_state *sack)
3109 const struct inet_connection_sock *icsk = inet_csk(sk);
3110 struct skb_mstamp first_ackt, last_ackt, now;
3111 struct tcp_sock *tp = tcp_sk(sk);
3112 u32 prior_sacked = tp->sacked_out;
3113 u32 reord = tp->packets_out;
3114 bool fully_acked = true;
3115 long sack_rtt_us = -1L;
3116 long seq_rtt_us = -1L;
3117 long ca_rtt_us = -1L;
3118 struct sk_buff *skb;
3125 while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
3126 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3127 u8 sacked = scb->sacked;
3130 tcp_ack_tstamp(sk, skb, prior_snd_una);
3132 /* Determine how many packets and what bytes were acked, tso and else */
3133 if (after(scb->end_seq, tp->snd_una)) {
3134 if (tcp_skb_pcount(skb) == 1 ||
3135 !after(tp->snd_una, scb->seq))
3138 acked_pcount = tcp_tso_acked(sk, skb);
3142 fully_acked = false;
3144 /* Speedup tcp_unlink_write_queue() and next loop */
3145 prefetchw(skb->next);
3146 acked_pcount = tcp_skb_pcount(skb);
3149 if (unlikely(sacked & TCPCB_RETRANS)) {
3150 if (sacked & TCPCB_SACKED_RETRANS)
3151 tp->retrans_out -= acked_pcount;
3152 flag |= FLAG_RETRANS_DATA_ACKED;
3153 } else if (!(sacked & TCPCB_SACKED_ACKED)) {
3154 last_ackt = skb->skb_mstamp;
3155 WARN_ON_ONCE(last_ackt.v64 == 0);
3156 if (!first_ackt.v64)
3157 first_ackt = last_ackt;
3159 reord = min(pkts_acked, reord);
3160 if (!after(scb->end_seq, tp->high_seq))
3161 flag |= FLAG_ORIG_SACK_ACKED;
3164 if (sacked & TCPCB_SACKED_ACKED)
3165 tp->sacked_out -= acked_pcount;
3166 else if (tcp_is_sack(tp) && !tcp_skb_spurious_retrans(tp, skb))
3167 tcp_rack_advance(tp, &skb->skb_mstamp, sacked);
3168 if (sacked & TCPCB_LOST)
3169 tp->lost_out -= acked_pcount;
3171 tp->packets_out -= acked_pcount;
3172 pkts_acked += acked_pcount;
3174 /* Initial outgoing SYN's get put onto the write_queue
3175 * just like anything else we transmit. It is not
3176 * true data, and if we misinform our callers that
3177 * this ACK acks real data, we will erroneously exit
3178 * connection startup slow start one packet too
3179 * quickly. This is severely frowned upon behavior.
3181 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
3182 flag |= FLAG_DATA_ACKED;
3184 flag |= FLAG_SYN_ACKED;
3185 tp->retrans_stamp = 0;
3191 tcp_unlink_write_queue(skb, sk);
3192 sk_wmem_free_skb(sk, skb);
3193 if (unlikely(skb == tp->retransmit_skb_hint))
3194 tp->retransmit_skb_hint = NULL;
3195 if (unlikely(skb == tp->lost_skb_hint))
3196 tp->lost_skb_hint = NULL;
3199 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3200 tp->snd_up = tp->snd_una;
3202 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3203 flag |= FLAG_SACK_RENEGING;
3205 skb_mstamp_get(&now);
3206 if (likely(first_ackt.v64) && !(flag & FLAG_RETRANS_DATA_ACKED)) {
3207 seq_rtt_us = skb_mstamp_us_delta(&now, &first_ackt);
3208 ca_rtt_us = skb_mstamp_us_delta(&now, &last_ackt);
3210 if (sack->first_sackt.v64) {
3211 sack_rtt_us = skb_mstamp_us_delta(&now, &sack->first_sackt);
3212 ca_rtt_us = skb_mstamp_us_delta(&now, &sack->last_sackt);
3215 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us,
3218 if (flag & FLAG_ACKED) {
3220 if (unlikely(icsk->icsk_mtup.probe_size &&
3221 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3222 tcp_mtup_probe_success(sk);
3225 if (tcp_is_reno(tp)) {
3226 tcp_remove_reno_sacks(sk, pkts_acked);
3230 /* Non-retransmitted hole got filled? That's reordering */
3231 if (reord < prior_fackets)
3232 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
3234 delta = tcp_is_fack(tp) ? pkts_acked :
3235 prior_sacked - tp->sacked_out;
3236 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3239 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
3241 } else if (skb && rtt_update && sack_rtt_us >= 0 &&
3242 sack_rtt_us > skb_mstamp_us_delta(&now, &skb->skb_mstamp)) {
3243 /* Do not re-arm RTO if the sack RTT is measured from data sent
3244 * after when the head was last (re)transmitted. Otherwise the
3245 * timeout may continue to extend in loss recovery.
3250 if (icsk->icsk_ca_ops->pkts_acked)
3251 icsk->icsk_ca_ops->pkts_acked(sk, pkts_acked, ca_rtt_us);
3253 #if FASTRETRANS_DEBUG > 0
3254 WARN_ON((int)tp->sacked_out < 0);
3255 WARN_ON((int)tp->lost_out < 0);
3256 WARN_ON((int)tp->retrans_out < 0);
3257 if (!tp->packets_out && tcp_is_sack(tp)) {
3258 icsk = inet_csk(sk);
3260 pr_debug("Leak l=%u %d\n",
3261 tp->lost_out, icsk->icsk_ca_state);
3264 if (tp->sacked_out) {
3265 pr_debug("Leak s=%u %d\n",
3266 tp->sacked_out, icsk->icsk_ca_state);
3269 if (tp->retrans_out) {
3270 pr_debug("Leak r=%u %d\n",
3271 tp->retrans_out, icsk->icsk_ca_state);
3272 tp->retrans_out = 0;
3279 static void tcp_ack_probe(struct sock *sk)
3281 const struct tcp_sock *tp = tcp_sk(sk);
3282 struct inet_connection_sock *icsk = inet_csk(sk);
3284 /* Was it a usable window open? */
3286 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
3287 icsk->icsk_backoff = 0;
3288 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3289 /* Socket must be waked up by subsequent tcp_data_snd_check().
3290 * This function is not for random using!
3293 unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX);
3295 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3300 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3302 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3303 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3306 /* Decide wheather to run the increase function of congestion control. */
3307 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3309 if (tcp_in_cwnd_reduction(sk))
3312 /* If reordering is high then always grow cwnd whenever data is
3313 * delivered regardless of its ordering. Otherwise stay conservative
3314 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3315 * new SACK or ECE mark may first advance cwnd here and later reduce
3316 * cwnd in tcp_fastretrans_alert() based on more states.
3318 if (tcp_sk(sk)->reordering > sysctl_tcp_reordering)
3319 return flag & FLAG_FORWARD_PROGRESS;
3321 return flag & FLAG_DATA_ACKED;
3324 /* Check that window update is acceptable.
3325 * The function assumes that snd_una<=ack<=snd_next.
3327 static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3328 const u32 ack, const u32 ack_seq,
3331 return after(ack, tp->snd_una) ||
3332 after(ack_seq, tp->snd_wl1) ||
3333 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3336 /* If we update tp->snd_una, also update tp->bytes_acked */
3337 static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack)
3339 u32 delta = ack - tp->snd_una;
3341 u64_stats_update_begin(&tp->syncp);
3342 tp->bytes_acked += delta;
3343 u64_stats_update_end(&tp->syncp);
3347 /* If we update tp->rcv_nxt, also update tp->bytes_received */
3348 static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq)
3350 u32 delta = seq - tp->rcv_nxt;
3352 u64_stats_update_begin(&tp->syncp);
3353 tp->bytes_received += delta;
3354 u64_stats_update_end(&tp->syncp);
3358 /* Update our send window.
3360 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3361 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3363 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3366 struct tcp_sock *tp = tcp_sk(sk);
3368 u32 nwin = ntohs(tcp_hdr(skb)->window);
3370 if (likely(!tcp_hdr(skb)->syn))
3371 nwin <<= tp->rx_opt.snd_wscale;
3373 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3374 flag |= FLAG_WIN_UPDATE;
3375 tcp_update_wl(tp, ack_seq);
3377 if (tp->snd_wnd != nwin) {
3380 /* Note, it is the only place, where
3381 * fast path is recovered for sending TCP.
3384 tcp_fast_path_check(sk);
3386 if (tcp_send_head(sk))
3387 tcp_slow_start_after_idle_check(sk);
3389 if (nwin > tp->max_window) {
3390 tp->max_window = nwin;
3391 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3396 tcp_snd_una_update(tp, ack);
3401 /* Return true if we're currently rate-limiting out-of-window ACKs and
3402 * thus shouldn't send a dupack right now. We rate-limit dupacks in
3403 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3404 * attacks that send repeated SYNs or ACKs for the same connection. To
3405 * do this, we do not send a duplicate SYNACK or ACK if the remote
3406 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3408 bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb,
3409 int mib_idx, u32 *last_oow_ack_time)
3411 /* Data packets without SYNs are not likely part of an ACK loop. */
3412 if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) &&
3414 goto not_rate_limited;
3416 if (*last_oow_ack_time) {
3417 s32 elapsed = (s32)(tcp_time_stamp - *last_oow_ack_time);
3419 if (0 <= elapsed && elapsed < sysctl_tcp_invalid_ratelimit) {
3420 NET_INC_STATS_BH(net, mib_idx);
3421 return true; /* rate-limited: don't send yet! */
3425 *last_oow_ack_time = tcp_time_stamp;
3428 return false; /* not rate-limited: go ahead, send dupack now! */
3431 /* RFC 5961 7 [ACK Throttling] */
3432 static void tcp_send_challenge_ack(struct sock *sk, const struct sk_buff *skb)
3434 /* unprotected vars, we dont care of overwrites */
3435 static u32 challenge_timestamp;
3436 static unsigned int challenge_count;
3437 struct tcp_sock *tp = tcp_sk(sk);
3440 /* First check our per-socket dupack rate limit. */
3441 if (tcp_oow_rate_limited(sock_net(sk), skb,
3442 LINUX_MIB_TCPACKSKIPPEDCHALLENGE,
3443 &tp->last_oow_ack_time))
3446 /* Then check the check host-wide RFC 5961 rate limit. */
3448 if (now != challenge_timestamp) {
3449 challenge_timestamp = now;
3450 challenge_count = 0;
3452 if (++challenge_count <= sysctl_tcp_challenge_ack_limit) {
3453 NET_INC_STATS_BH(sock_net(sk), 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_BH(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 /* This routine deals with incoming acks, but not outgoing ones. */
3541 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3543 struct inet_connection_sock *icsk = inet_csk(sk);
3544 struct tcp_sock *tp = tcp_sk(sk);
3545 struct tcp_sacktag_state sack_state;
3546 u32 prior_snd_una = tp->snd_una;
3547 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3548 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3549 bool is_dupack = false;
3551 int prior_packets = tp->packets_out;
3552 const int prior_unsacked = tp->packets_out - tp->sacked_out;
3553 int acked = 0; /* Number of packets newly acked */
3554 int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */
3556 sack_state.first_sackt.v64 = 0;
3558 /* We very likely will need to access write queue head. */
3559 prefetchw(sk->sk_write_queue.next);
3561 /* If the ack is older than previous acks
3562 * then we can probably ignore it.
3564 if (before(ack, prior_snd_una)) {
3565 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3566 if (before(ack, prior_snd_una - tp->max_window)) {
3567 tcp_send_challenge_ack(sk, skb);
3573 /* If the ack includes data we haven't sent yet, discard
3574 * this segment (RFC793 Section 3.9).
3576 if (after(ack, tp->snd_nxt))
3579 if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
3580 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
3583 if (after(ack, prior_snd_una)) {
3584 flag |= FLAG_SND_UNA_ADVANCED;
3585 icsk->icsk_retransmits = 0;
3588 prior_fackets = tp->fackets_out;
3590 /* ts_recent update must be made after we are sure that the packet
3593 if (flag & FLAG_UPDATE_TS_RECENT)
3594 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3596 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3597 /* Window is constant, pure forward advance.
3598 * No more checks are required.
3599 * Note, we use the fact that SND.UNA>=SND.WL2.
3601 tcp_update_wl(tp, ack_seq);
3602 tcp_snd_una_update(tp, ack);
3603 flag |= FLAG_WIN_UPDATE;
3605 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
3607 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS);
3609 u32 ack_ev_flags = CA_ACK_SLOWPATH;
3611 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3614 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3616 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3618 if (TCP_SKB_CB(skb)->sacked)
3619 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3622 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) {
3624 ack_ev_flags |= CA_ACK_ECE;
3627 if (flag & FLAG_WIN_UPDATE)
3628 ack_ev_flags |= CA_ACK_WIN_UPDATE;
3630 tcp_in_ack_event(sk, ack_ev_flags);
3633 /* We passed data and got it acked, remove any soft error
3634 * log. Something worked...
3636 sk->sk_err_soft = 0;
3637 icsk->icsk_probes_out = 0;
3638 tp->rcv_tstamp = tcp_time_stamp;
3642 /* See if we can take anything off of the retransmit queue. */
3643 acked = tp->packets_out;
3644 flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una,
3646 acked -= tp->packets_out;
3648 if (tcp_ack_is_dubious(sk, flag)) {
3649 is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
3650 tcp_fastretrans_alert(sk, acked, prior_unsacked,
3651 is_dupack, flag, &rexmit);
3653 if (tp->tlp_high_seq)
3654 tcp_process_tlp_ack(sk, ack, flag);
3656 /* Advance cwnd if state allows */
3657 if (tcp_may_raise_cwnd(sk, flag))
3658 tcp_cong_avoid(sk, ack, acked);
3660 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) {
3661 struct dst_entry *dst = __sk_dst_get(sk);
3666 if (icsk->icsk_pending == ICSK_TIME_RETRANS)
3667 tcp_schedule_loss_probe(sk);
3668 tcp_update_pacing_rate(sk);
3669 tcp_xmit_recovery(sk, rexmit);
3673 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3674 if (flag & FLAG_DSACKING_ACK)
3675 tcp_fastretrans_alert(sk, acked, prior_unsacked,
3676 is_dupack, flag, &rexmit);
3677 /* If this ack opens up a zero window, clear backoff. It was
3678 * being used to time the probes, and is probably far higher than
3679 * it needs to be for normal retransmission.
3681 if (tcp_send_head(sk))
3684 if (tp->tlp_high_seq)
3685 tcp_process_tlp_ack(sk, ack, flag);
3689 SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3693 /* If data was SACKed, tag it and see if we should send more data.
3694 * If data was DSACKed, see if we can undo a cwnd reduction.
3696 if (TCP_SKB_CB(skb)->sacked) {
3697 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3699 tcp_fastretrans_alert(sk, acked, prior_unsacked,
3700 is_dupack, flag, &rexmit);
3701 tcp_xmit_recovery(sk, rexmit);
3704 SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3708 static void tcp_parse_fastopen_option(int len, const unsigned char *cookie,
3709 bool syn, struct tcp_fastopen_cookie *foc,
3712 /* Valid only in SYN or SYN-ACK with an even length. */
3713 if (!foc || !syn || len < 0 || (len & 1))
3716 if (len >= TCP_FASTOPEN_COOKIE_MIN &&
3717 len <= TCP_FASTOPEN_COOKIE_MAX)
3718 memcpy(foc->val, cookie, len);
3725 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3726 * But, this can also be called on packets in the established flow when
3727 * the fast version below fails.
3729 void tcp_parse_options(const struct sk_buff *skb,
3730 struct tcp_options_received *opt_rx, int estab,
3731 struct tcp_fastopen_cookie *foc)
3733 const unsigned char *ptr;
3734 const struct tcphdr *th = tcp_hdr(skb);
3735 int length = (th->doff * 4) - sizeof(struct tcphdr);
3737 ptr = (const unsigned char *)(th + 1);
3738 opt_rx->saw_tstamp = 0;
3740 while (length > 0) {
3741 int opcode = *ptr++;
3747 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3752 if (opsize < 2) /* "silly options" */
3754 if (opsize > length)
3755 return; /* don't parse partial options */
3758 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3759 u16 in_mss = get_unaligned_be16(ptr);
3761 if (opt_rx->user_mss &&
3762 opt_rx->user_mss < in_mss)
3763 in_mss = opt_rx->user_mss;
3764 opt_rx->mss_clamp = in_mss;
3769 if (opsize == TCPOLEN_WINDOW && th->syn &&
3770 !estab && sysctl_tcp_window_scaling) {
3771 __u8 snd_wscale = *(__u8 *)ptr;
3772 opt_rx->wscale_ok = 1;
3773 if (snd_wscale > 14) {
3774 net_info_ratelimited("%s: Illegal window scaling value %d >14 received\n",
3779 opt_rx->snd_wscale = snd_wscale;
3782 case TCPOPT_TIMESTAMP:
3783 if ((opsize == TCPOLEN_TIMESTAMP) &&
3784 ((estab && opt_rx->tstamp_ok) ||
3785 (!estab && sysctl_tcp_timestamps))) {
3786 opt_rx->saw_tstamp = 1;
3787 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3788 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3791 case TCPOPT_SACK_PERM:
3792 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3793 !estab && sysctl_tcp_sack) {
3794 opt_rx->sack_ok = TCP_SACK_SEEN;
3795 tcp_sack_reset(opt_rx);
3800 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3801 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3803 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3806 #ifdef CONFIG_TCP_MD5SIG
3809 * The MD5 Hash has already been
3810 * checked (see tcp_v{4,6}_do_rcv()).
3814 case TCPOPT_FASTOPEN:
3815 tcp_parse_fastopen_option(
3816 opsize - TCPOLEN_FASTOPEN_BASE,
3817 ptr, th->syn, foc, false);
3821 /* Fast Open option shares code 254 using a
3822 * 16 bits magic number.
3824 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE &&
3825 get_unaligned_be16(ptr) ==
3826 TCPOPT_FASTOPEN_MAGIC)
3827 tcp_parse_fastopen_option(opsize -
3828 TCPOLEN_EXP_FASTOPEN_BASE,
3829 ptr + 2, th->syn, foc, true);
3838 EXPORT_SYMBOL(tcp_parse_options);
3840 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
3842 const __be32 *ptr = (const __be32 *)(th + 1);
3844 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3845 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3846 tp->rx_opt.saw_tstamp = 1;
3848 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3851 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
3853 tp->rx_opt.rcv_tsecr = 0;
3859 /* Fast parse options. This hopes to only see timestamps.
3860 * If it is wrong it falls back on tcp_parse_options().
3862 static bool tcp_fast_parse_options(const struct sk_buff *skb,
3863 const struct tcphdr *th, struct tcp_sock *tp)
3865 /* In the spirit of fast parsing, compare doff directly to constant
3866 * values. Because equality is used, short doff can be ignored here.
3868 if (th->doff == (sizeof(*th) / 4)) {
3869 tp->rx_opt.saw_tstamp = 0;
3871 } else if (tp->rx_opt.tstamp_ok &&
3872 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3873 if (tcp_parse_aligned_timestamp(tp, th))
3877 tcp_parse_options(skb, &tp->rx_opt, 1, NULL);
3878 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3879 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
3884 #ifdef CONFIG_TCP_MD5SIG
3886 * Parse MD5 Signature option
3888 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
3890 int length = (th->doff << 2) - sizeof(*th);
3891 const u8 *ptr = (const u8 *)(th + 1);
3893 /* If the TCP option is too short, we can short cut */
3894 if (length < TCPOLEN_MD5SIG)
3897 while (length > 0) {
3898 int opcode = *ptr++;
3909 if (opsize < 2 || opsize > length)
3911 if (opcode == TCPOPT_MD5SIG)
3912 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
3919 EXPORT_SYMBOL(tcp_parse_md5sig_option);
3922 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3924 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3925 * it can pass through stack. So, the following predicate verifies that
3926 * this segment is not used for anything but congestion avoidance or
3927 * fast retransmit. Moreover, we even are able to eliminate most of such
3928 * second order effects, if we apply some small "replay" window (~RTO)
3929 * to timestamp space.
3931 * All these measures still do not guarantee that we reject wrapped ACKs
3932 * on networks with high bandwidth, when sequence space is recycled fastly,
3933 * but it guarantees that such events will be very rare and do not affect
3934 * connection seriously. This doesn't look nice, but alas, PAWS is really
3937 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3938 * states that events when retransmit arrives after original data are rare.
3939 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3940 * the biggest problem on large power networks even with minor reordering.
3941 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3942 * up to bandwidth of 18Gigabit/sec. 8) ]
3945 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3947 const struct tcp_sock *tp = tcp_sk(sk);
3948 const struct tcphdr *th = tcp_hdr(skb);
3949 u32 seq = TCP_SKB_CB(skb)->seq;
3950 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3952 return (/* 1. Pure ACK with correct sequence number. */
3953 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3955 /* 2. ... and duplicate ACK. */
3956 ack == tp->snd_una &&
3958 /* 3. ... and does not update window. */
3959 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3961 /* 4. ... and sits in replay window. */
3962 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3965 static inline bool tcp_paws_discard(const struct sock *sk,
3966 const struct sk_buff *skb)
3968 const struct tcp_sock *tp = tcp_sk(sk);
3970 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
3971 !tcp_disordered_ack(sk, skb);
3974 /* Check segment sequence number for validity.
3976 * Segment controls are considered valid, if the segment
3977 * fits to the window after truncation to the window. Acceptability
3978 * of data (and SYN, FIN, of course) is checked separately.
3979 * See tcp_data_queue(), for example.
3981 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3982 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3983 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3984 * (borrowed from freebsd)
3987 static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
3989 return !before(end_seq, tp->rcv_wup) &&
3990 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3993 /* When we get a reset we do this. */
3994 void tcp_reset(struct sock *sk)
3996 /* We want the right error as BSD sees it (and indeed as we do). */
3997 switch (sk->sk_state) {
3999 sk->sk_err = ECONNREFUSED;
4001 case TCP_CLOSE_WAIT:
4007 sk->sk_err = ECONNRESET;
4009 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4012 if (!sock_flag(sk, SOCK_DEAD))
4013 sk->sk_error_report(sk);
4019 * Process the FIN bit. This now behaves as it is supposed to work
4020 * and the FIN takes effect when it is validly part of sequence
4021 * space. Not before when we get holes.
4023 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4024 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4027 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4028 * close and we go into CLOSING (and later onto TIME-WAIT)
4030 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4032 void tcp_fin(struct sock *sk)
4034 struct tcp_sock *tp = tcp_sk(sk);
4036 inet_csk_schedule_ack(sk);
4038 sk->sk_shutdown |= RCV_SHUTDOWN;
4039 sock_set_flag(sk, SOCK_DONE);
4041 switch (sk->sk_state) {
4043 case TCP_ESTABLISHED:
4044 /* Move to CLOSE_WAIT */
4045 tcp_set_state(sk, TCP_CLOSE_WAIT);
4046 inet_csk(sk)->icsk_ack.pingpong = 1;
4049 case TCP_CLOSE_WAIT:
4051 /* Received a retransmission of the FIN, do
4056 /* RFC793: Remain in the LAST-ACK state. */
4060 /* This case occurs when a simultaneous close
4061 * happens, we must ack the received FIN and
4062 * enter the CLOSING state.
4065 tcp_set_state(sk, TCP_CLOSING);
4068 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4070 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4073 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4074 * cases we should never reach this piece of code.
4076 pr_err("%s: Impossible, sk->sk_state=%d\n",
4077 __func__, sk->sk_state);
4081 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4082 * Probably, we should reset in this case. For now drop them.
4084 __skb_queue_purge(&tp->out_of_order_queue);
4085 if (tcp_is_sack(tp))
4086 tcp_sack_reset(&tp->rx_opt);
4089 if (!sock_flag(sk, SOCK_DEAD)) {
4090 sk->sk_state_change(sk);
4092 /* Do not send POLL_HUP for half duplex close. */
4093 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4094 sk->sk_state == TCP_CLOSE)
4095 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4097 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4101 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4104 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4105 if (before(seq, sp->start_seq))
4106 sp->start_seq = seq;
4107 if (after(end_seq, sp->end_seq))
4108 sp->end_seq = end_seq;
4114 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4116 struct tcp_sock *tp = tcp_sk(sk);
4118 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4121 if (before(seq, tp->rcv_nxt))
4122 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4124 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4126 NET_INC_STATS_BH(sock_net(sk), mib_idx);
4128 tp->rx_opt.dsack = 1;
4129 tp->duplicate_sack[0].start_seq = seq;
4130 tp->duplicate_sack[0].end_seq = end_seq;
4134 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4136 struct tcp_sock *tp = tcp_sk(sk);
4138 if (!tp->rx_opt.dsack)
4139 tcp_dsack_set(sk, seq, end_seq);
4141 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4144 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4146 struct tcp_sock *tp = tcp_sk(sk);
4148 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4149 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4150 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4151 tcp_enter_quickack_mode(sk);
4153 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4154 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4156 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4157 end_seq = tp->rcv_nxt;
4158 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4165 /* These routines update the SACK block as out-of-order packets arrive or
4166 * in-order packets close up the sequence space.
4168 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4171 struct tcp_sack_block *sp = &tp->selective_acks[0];
4172 struct tcp_sack_block *swalk = sp + 1;
4174 /* See if the recent change to the first SACK eats into
4175 * or hits the sequence space of other SACK blocks, if so coalesce.
4177 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4178 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4181 /* Zap SWALK, by moving every further SACK up by one slot.
4182 * Decrease num_sacks.
4184 tp->rx_opt.num_sacks--;
4185 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4189 this_sack++, swalk++;
4193 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4195 struct tcp_sock *tp = tcp_sk(sk);
4196 struct tcp_sack_block *sp = &tp->selective_acks[0];
4197 int cur_sacks = tp->rx_opt.num_sacks;
4203 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4204 if (tcp_sack_extend(sp, seq, end_seq)) {
4205 /* Rotate this_sack to the first one. */
4206 for (; this_sack > 0; this_sack--, sp--)
4207 swap(*sp, *(sp - 1));
4209 tcp_sack_maybe_coalesce(tp);
4214 /* Could not find an adjacent existing SACK, build a new one,
4215 * put it at the front, and shift everyone else down. We
4216 * always know there is at least one SACK present already here.
4218 * If the sack array is full, forget about the last one.
4220 if (this_sack >= TCP_NUM_SACKS) {
4222 tp->rx_opt.num_sacks--;
4225 for (; this_sack > 0; this_sack--, sp--)
4229 /* Build the new head SACK, and we're done. */
4230 sp->start_seq = seq;
4231 sp->end_seq = end_seq;
4232 tp->rx_opt.num_sacks++;
4235 /* RCV.NXT advances, some SACKs should be eaten. */
4237 static void tcp_sack_remove(struct tcp_sock *tp)
4239 struct tcp_sack_block *sp = &tp->selective_acks[0];
4240 int num_sacks = tp->rx_opt.num_sacks;
4243 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4244 if (skb_queue_empty(&tp->out_of_order_queue)) {
4245 tp->rx_opt.num_sacks = 0;
4249 for (this_sack = 0; this_sack < num_sacks;) {
4250 /* Check if the start of the sack is covered by RCV.NXT. */
4251 if (!before(tp->rcv_nxt, sp->start_seq)) {
4254 /* RCV.NXT must cover all the block! */
4255 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4257 /* Zap this SACK, by moving forward any other SACKS. */
4258 for (i = this_sack+1; i < num_sacks; i++)
4259 tp->selective_acks[i-1] = tp->selective_acks[i];
4266 tp->rx_opt.num_sacks = num_sacks;
4270 * tcp_try_coalesce - try to merge skb to prior one
4273 * @from: buffer to add in queue
4274 * @fragstolen: pointer to boolean
4276 * Before queueing skb @from after @to, try to merge them
4277 * to reduce overall memory use and queue lengths, if cost is small.
4278 * Packets in ofo or receive queues can stay a long time.
4279 * Better try to coalesce them right now to avoid future collapses.
4280 * Returns true if caller should free @from instead of queueing it
4282 static bool tcp_try_coalesce(struct sock *sk,
4284 struct sk_buff *from,
4289 *fragstolen = false;
4291 /* Its possible this segment overlaps with prior segment in queue */
4292 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4295 if (!skb_try_coalesce(to, from, fragstolen, &delta))
4298 atomic_add(delta, &sk->sk_rmem_alloc);
4299 sk_mem_charge(sk, delta);
4300 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4301 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4302 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4303 TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
4307 /* This one checks to see if we can put data from the
4308 * out_of_order queue into the receive_queue.
4310 static void tcp_ofo_queue(struct sock *sk)
4312 struct tcp_sock *tp = tcp_sk(sk);
4313 __u32 dsack_high = tp->rcv_nxt;
4314 struct sk_buff *skb, *tail;
4315 bool fragstolen, eaten;
4317 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
4318 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4321 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4322 __u32 dsack = dsack_high;
4323 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4324 dsack_high = TCP_SKB_CB(skb)->end_seq;
4325 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4328 __skb_unlink(skb, &tp->out_of_order_queue);
4329 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4330 SOCK_DEBUG(sk, "ofo packet was already received\n");
4334 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4335 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4336 TCP_SKB_CB(skb)->end_seq);
4338 tail = skb_peek_tail(&sk->sk_receive_queue);
4339 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen);
4340 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4342 __skb_queue_tail(&sk->sk_receive_queue, skb);
4343 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
4346 kfree_skb_partial(skb, fragstolen);
4350 static bool tcp_prune_ofo_queue(struct sock *sk);
4351 static int tcp_prune_queue(struct sock *sk);
4353 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4356 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4357 !sk_rmem_schedule(sk, skb, size)) {
4359 if (tcp_prune_queue(sk) < 0)
4362 if (!sk_rmem_schedule(sk, skb, size)) {
4363 if (!tcp_prune_ofo_queue(sk))
4366 if (!sk_rmem_schedule(sk, skb, size))
4373 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4375 struct tcp_sock *tp = tcp_sk(sk);
4376 struct sk_buff *skb1;
4379 tcp_ecn_check_ce(tp, skb);
4381 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4382 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFODROP);
4387 /* Disable header prediction. */
4389 inet_csk_schedule_ack(sk);
4391 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4392 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4393 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4395 skb1 = skb_peek_tail(&tp->out_of_order_queue);
4397 /* Initial out of order segment, build 1 SACK. */
4398 if (tcp_is_sack(tp)) {
4399 tp->rx_opt.num_sacks = 1;
4400 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
4401 tp->selective_acks[0].end_seq =
4402 TCP_SKB_CB(skb)->end_seq;
4404 __skb_queue_head(&tp->out_of_order_queue, skb);
4408 seq = TCP_SKB_CB(skb)->seq;
4409 end_seq = TCP_SKB_CB(skb)->end_seq;
4411 if (seq == TCP_SKB_CB(skb1)->end_seq) {
4414 if (!tcp_try_coalesce(sk, skb1, skb, &fragstolen)) {
4415 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4417 tcp_grow_window(sk, skb);
4418 kfree_skb_partial(skb, fragstolen);
4422 if (!tp->rx_opt.num_sacks ||
4423 tp->selective_acks[0].end_seq != seq)
4426 /* Common case: data arrive in order after hole. */
4427 tp->selective_acks[0].end_seq = end_seq;
4431 /* Find place to insert this segment. */
4433 if (!after(TCP_SKB_CB(skb1)->seq, seq))
4435 if (skb_queue_is_first(&tp->out_of_order_queue, skb1)) {
4439 skb1 = skb_queue_prev(&tp->out_of_order_queue, skb1);
4442 /* Do skb overlap to previous one? */
4443 if (skb1 && before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4444 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4445 /* All the bits are present. Drop. */
4446 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4449 tcp_dsack_set(sk, seq, end_seq);
4452 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4453 /* Partial overlap. */
4454 tcp_dsack_set(sk, seq,
4455 TCP_SKB_CB(skb1)->end_seq);
4457 if (skb_queue_is_first(&tp->out_of_order_queue,
4461 skb1 = skb_queue_prev(
4462 &tp->out_of_order_queue,
4467 __skb_queue_head(&tp->out_of_order_queue, skb);
4469 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4471 /* And clean segments covered by new one as whole. */
4472 while (!skb_queue_is_last(&tp->out_of_order_queue, skb)) {
4473 skb1 = skb_queue_next(&tp->out_of_order_queue, skb);
4475 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4477 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4478 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4482 __skb_unlink(skb1, &tp->out_of_order_queue);
4483 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4484 TCP_SKB_CB(skb1)->end_seq);
4485 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4490 if (tcp_is_sack(tp))
4491 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4494 tcp_grow_window(sk, skb);
4495 skb_set_owner_r(skb, sk);
4499 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, int hdrlen,
4503 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4505 __skb_pull(skb, hdrlen);
4507 tcp_try_coalesce(sk, tail, skb, fragstolen)) ? 1 : 0;
4508 tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq);
4510 __skb_queue_tail(&sk->sk_receive_queue, skb);
4511 skb_set_owner_r(skb, sk);
4516 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4518 struct sk_buff *skb;
4526 if (size > PAGE_SIZE) {
4527 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS);
4529 data_len = npages << PAGE_SHIFT;
4530 size = data_len + (size & ~PAGE_MASK);
4532 skb = alloc_skb_with_frags(size - data_len, data_len,
4533 PAGE_ALLOC_COSTLY_ORDER,
4534 &err, sk->sk_allocation);
4538 skb_put(skb, size - data_len);
4539 skb->data_len = data_len;
4542 if (tcp_try_rmem_schedule(sk, skb, skb->truesize))
4545 err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size);
4549 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4550 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4551 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4553 if (tcp_queue_rcv(sk, skb, 0, &fragstolen)) {
4554 WARN_ON_ONCE(fragstolen); /* should not happen */
4566 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4568 struct tcp_sock *tp = tcp_sk(sk);
4570 bool fragstolen = false;
4572 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
4576 __skb_pull(skb, tcp_hdr(skb)->doff * 4);
4578 tcp_ecn_accept_cwr(tp, skb);
4580 tp->rx_opt.dsack = 0;
4582 /* Queue data for delivery to the user.
4583 * Packets in sequence go to the receive queue.
4584 * Out of sequence packets to the out_of_order_queue.
4586 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4587 if (tcp_receive_window(tp) == 0)
4590 /* Ok. In sequence. In window. */
4591 if (tp->ucopy.task == current &&
4592 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4593 sock_owned_by_user(sk) && !tp->urg_data) {
4594 int chunk = min_t(unsigned int, skb->len,
4597 __set_current_state(TASK_RUNNING);
4600 if (!skb_copy_datagram_msg(skb, 0, tp->ucopy.msg, chunk)) {
4601 tp->ucopy.len -= chunk;
4602 tp->copied_seq += chunk;
4603 eaten = (chunk == skb->len);
4604 tcp_rcv_space_adjust(sk);
4612 if (skb_queue_len(&sk->sk_receive_queue) == 0)
4613 sk_forced_mem_schedule(sk, skb->truesize);
4614 else if (tcp_try_rmem_schedule(sk, skb, skb->truesize))
4617 eaten = tcp_queue_rcv(sk, skb, 0, &fragstolen);
4619 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4621 tcp_event_data_recv(sk, skb);
4622 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
4625 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4628 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4629 * gap in queue is filled.
4631 if (skb_queue_empty(&tp->out_of_order_queue))
4632 inet_csk(sk)->icsk_ack.pingpong = 0;
4635 if (tp->rx_opt.num_sacks)
4636 tcp_sack_remove(tp);
4638 tcp_fast_path_check(sk);
4641 kfree_skb_partial(skb, fragstolen);
4642 if (!sock_flag(sk, SOCK_DEAD))
4643 sk->sk_data_ready(sk);
4647 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4648 /* A retransmit, 2nd most common case. Force an immediate ack. */
4649 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4650 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4653 tcp_enter_quickack_mode(sk);
4654 inet_csk_schedule_ack(sk);
4660 /* Out of window. F.e. zero window probe. */
4661 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4664 tcp_enter_quickack_mode(sk);
4666 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4667 /* Partial packet, seq < rcv_next < end_seq */
4668 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4669 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4670 TCP_SKB_CB(skb)->end_seq);
4672 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4674 /* If window is closed, drop tail of packet. But after
4675 * remembering D-SACK for its head made in previous line.
4677 if (!tcp_receive_window(tp))
4682 tcp_data_queue_ofo(sk, skb);
4685 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4686 struct sk_buff_head *list)
4688 struct sk_buff *next = NULL;
4690 if (!skb_queue_is_last(list, skb))
4691 next = skb_queue_next(list, skb);
4693 __skb_unlink(skb, list);
4695 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4700 /* Collapse contiguous sequence of skbs head..tail with
4701 * sequence numbers start..end.
4703 * If tail is NULL, this means until the end of the list.
4705 * Segments with FIN/SYN are not collapsed (only because this
4709 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4710 struct sk_buff *head, struct sk_buff *tail,
4713 struct sk_buff *skb, *n;
4716 /* First, check that queue is collapsible and find
4717 * the point where collapsing can be useful. */
4721 skb_queue_walk_from_safe(list, skb, n) {
4724 /* No new bits? It is possible on ofo queue. */
4725 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4726 skb = tcp_collapse_one(sk, skb, list);
4732 /* The first skb to collapse is:
4734 * - bloated or contains data before "start" or
4735 * overlaps to the next one.
4737 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
4738 (tcp_win_from_space(skb->truesize) > skb->len ||
4739 before(TCP_SKB_CB(skb)->seq, start))) {
4740 end_of_skbs = false;
4744 if (!skb_queue_is_last(list, skb)) {
4745 struct sk_buff *next = skb_queue_next(list, skb);
4747 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(next)->seq) {
4748 end_of_skbs = false;
4753 /* Decided to skip this, advance start seq. */
4754 start = TCP_SKB_CB(skb)->end_seq;
4757 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4760 while (before(start, end)) {
4761 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
4762 struct sk_buff *nskb;
4764 nskb = alloc_skb(copy, GFP_ATOMIC);
4768 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4769 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4770 __skb_queue_before(list, skb, nskb);
4771 skb_set_owner_r(nskb, sk);
4773 /* Copy data, releasing collapsed skbs. */
4775 int offset = start - TCP_SKB_CB(skb)->seq;
4776 int size = TCP_SKB_CB(skb)->end_seq - start;
4780 size = min(copy, size);
4781 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4783 TCP_SKB_CB(nskb)->end_seq += size;
4787 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4788 skb = tcp_collapse_one(sk, skb, list);
4791 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4798 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4799 * and tcp_collapse() them until all the queue is collapsed.
4801 static void tcp_collapse_ofo_queue(struct sock *sk)
4803 struct tcp_sock *tp = tcp_sk(sk);
4804 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4805 struct sk_buff *head;
4811 start = TCP_SKB_CB(skb)->seq;
4812 end = TCP_SKB_CB(skb)->end_seq;
4816 struct sk_buff *next = NULL;
4818 if (!skb_queue_is_last(&tp->out_of_order_queue, skb))
4819 next = skb_queue_next(&tp->out_of_order_queue, skb);
4822 /* Segment is terminated when we see gap or when
4823 * we are at the end of all the queue. */
4825 after(TCP_SKB_CB(skb)->seq, end) ||
4826 before(TCP_SKB_CB(skb)->end_seq, start)) {
4827 tcp_collapse(sk, &tp->out_of_order_queue,
4828 head, skb, start, end);
4832 /* Start new segment */
4833 start = TCP_SKB_CB(skb)->seq;
4834 end = TCP_SKB_CB(skb)->end_seq;
4836 if (before(TCP_SKB_CB(skb)->seq, start))
4837 start = TCP_SKB_CB(skb)->seq;
4838 if (after(TCP_SKB_CB(skb)->end_seq, end))
4839 end = TCP_SKB_CB(skb)->end_seq;
4845 * Purge the out-of-order queue.
4846 * Return true if queue was pruned.
4848 static bool tcp_prune_ofo_queue(struct sock *sk)
4850 struct tcp_sock *tp = tcp_sk(sk);
4853 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4854 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED);
4855 __skb_queue_purge(&tp->out_of_order_queue);
4857 /* Reset SACK state. A conforming SACK implementation will
4858 * do the same at a timeout based retransmit. When a connection
4859 * is in a sad state like this, we care only about integrity
4860 * of the connection not performance.
4862 if (tp->rx_opt.sack_ok)
4863 tcp_sack_reset(&tp->rx_opt);
4870 /* Reduce allocated memory if we can, trying to get
4871 * the socket within its memory limits again.
4873 * Return less than zero if we should start dropping frames
4874 * until the socket owning process reads some of the data
4875 * to stabilize the situation.
4877 static int tcp_prune_queue(struct sock *sk)
4879 struct tcp_sock *tp = tcp_sk(sk);
4881 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4883 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED);
4885 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4886 tcp_clamp_window(sk);
4887 else if (tcp_under_memory_pressure(sk))
4888 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4890 tcp_collapse_ofo_queue(sk);
4891 if (!skb_queue_empty(&sk->sk_receive_queue))
4892 tcp_collapse(sk, &sk->sk_receive_queue,
4893 skb_peek(&sk->sk_receive_queue),
4895 tp->copied_seq, tp->rcv_nxt);
4898 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4901 /* Collapsing did not help, destructive actions follow.
4902 * This must not ever occur. */
4904 tcp_prune_ofo_queue(sk);
4906 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4909 /* If we are really being abused, tell the caller to silently
4910 * drop receive data on the floor. It will get retransmitted
4911 * and hopefully then we'll have sufficient space.
4913 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED);
4915 /* Massive buffer overcommit. */
4920 static bool tcp_should_expand_sndbuf(const struct sock *sk)
4922 const struct tcp_sock *tp = tcp_sk(sk);
4924 /* If the user specified a specific send buffer setting, do
4927 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4930 /* If we are under global TCP memory pressure, do not expand. */
4931 if (tcp_under_memory_pressure(sk))
4934 /* If we are under soft global TCP memory pressure, do not expand. */
4935 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
4938 /* If we filled the congestion window, do not expand. */
4939 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
4945 /* When incoming ACK allowed to free some skb from write_queue,
4946 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4947 * on the exit from tcp input handler.
4949 * PROBLEM: sndbuf expansion does not work well with largesend.
4951 static void tcp_new_space(struct sock *sk)
4953 struct tcp_sock *tp = tcp_sk(sk);
4955 if (tcp_should_expand_sndbuf(sk)) {
4956 tcp_sndbuf_expand(sk);
4957 tp->snd_cwnd_stamp = tcp_time_stamp;
4960 sk->sk_write_space(sk);
4963 static void tcp_check_space(struct sock *sk)
4965 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
4966 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
4967 /* pairs with tcp_poll() */
4968 smp_mb__after_atomic();
4969 if (sk->sk_socket &&
4970 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
4975 static inline void tcp_data_snd_check(struct sock *sk)
4977 tcp_push_pending_frames(sk);
4978 tcp_check_space(sk);
4982 * Check if sending an ack is needed.
4984 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
4986 struct tcp_sock *tp = tcp_sk(sk);
4988 /* More than one full frame received... */
4989 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
4990 /* ... and right edge of window advances far enough.
4991 * (tcp_recvmsg() will send ACK otherwise). Or...
4993 __tcp_select_window(sk) >= tp->rcv_wnd) ||
4994 /* We ACK each frame or... */
4995 tcp_in_quickack_mode(sk) ||
4996 /* We have out of order data. */
4997 (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
4998 /* Then ack it now */
5001 /* Else, send delayed ack. */
5002 tcp_send_delayed_ack(sk);
5006 static inline void tcp_ack_snd_check(struct sock *sk)
5008 if (!inet_csk_ack_scheduled(sk)) {
5009 /* We sent a data segment already. */
5012 __tcp_ack_snd_check(sk, 1);
5016 * This routine is only called when we have urgent data
5017 * signaled. Its the 'slow' part of tcp_urg. It could be
5018 * moved inline now as tcp_urg is only called from one
5019 * place. We handle URGent data wrong. We have to - as
5020 * BSD still doesn't use the correction from RFC961.
5021 * For 1003.1g we should support a new option TCP_STDURG to permit
5022 * either form (or just set the sysctl tcp_stdurg).
5025 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
5027 struct tcp_sock *tp = tcp_sk(sk);
5028 u32 ptr = ntohs(th->urg_ptr);
5030 if (ptr && !sysctl_tcp_stdurg)
5032 ptr += ntohl(th->seq);
5034 /* Ignore urgent data that we've already seen and read. */
5035 if (after(tp->copied_seq, ptr))
5038 /* Do not replay urg ptr.
5040 * NOTE: interesting situation not covered by specs.
5041 * Misbehaving sender may send urg ptr, pointing to segment,
5042 * which we already have in ofo queue. We are not able to fetch
5043 * such data and will stay in TCP_URG_NOTYET until will be eaten
5044 * by recvmsg(). Seems, we are not obliged to handle such wicked
5045 * situations. But it is worth to think about possibility of some
5046 * DoSes using some hypothetical application level deadlock.
5048 if (before(ptr, tp->rcv_nxt))
5051 /* Do we already have a newer (or duplicate) urgent pointer? */
5052 if (tp->urg_data && !after(ptr, tp->urg_seq))
5055 /* Tell the world about our new urgent pointer. */
5058 /* We may be adding urgent data when the last byte read was
5059 * urgent. To do this requires some care. We cannot just ignore
5060 * tp->copied_seq since we would read the last urgent byte again
5061 * as data, nor can we alter copied_seq until this data arrives
5062 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5064 * NOTE. Double Dutch. Rendering to plain English: author of comment
5065 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5066 * and expect that both A and B disappear from stream. This is _wrong_.
5067 * Though this happens in BSD with high probability, this is occasional.
5068 * Any application relying on this is buggy. Note also, that fix "works"
5069 * only in this artificial test. Insert some normal data between A and B and we will
5070 * decline of BSD again. Verdict: it is better to remove to trap
5073 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5074 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5075 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5077 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5078 __skb_unlink(skb, &sk->sk_receive_queue);
5083 tp->urg_data = TCP_URG_NOTYET;
5086 /* Disable header prediction. */
5090 /* This is the 'fast' part of urgent handling. */
5091 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
5093 struct tcp_sock *tp = tcp_sk(sk);
5095 /* Check if we get a new urgent pointer - normally not. */
5097 tcp_check_urg(sk, th);
5099 /* Do we wait for any urgent data? - normally not... */
5100 if (tp->urg_data == TCP_URG_NOTYET) {
5101 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5104 /* Is the urgent pointer pointing into this packet? */
5105 if (ptr < skb->len) {
5107 if (skb_copy_bits(skb, ptr, &tmp, 1))
5109 tp->urg_data = TCP_URG_VALID | tmp;
5110 if (!sock_flag(sk, SOCK_DEAD))
5111 sk->sk_data_ready(sk);
5116 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
5118 struct tcp_sock *tp = tcp_sk(sk);
5119 int chunk = skb->len - hlen;
5123 if (skb_csum_unnecessary(skb))
5124 err = skb_copy_datagram_msg(skb, hlen, tp->ucopy.msg, chunk);
5126 err = skb_copy_and_csum_datagram_msg(skb, hlen, tp->ucopy.msg);
5129 tp->ucopy.len -= chunk;
5130 tp->copied_seq += chunk;
5131 tcp_rcv_space_adjust(sk);
5138 static __sum16 __tcp_checksum_complete_user(struct sock *sk,
5139 struct sk_buff *skb)
5143 if (sock_owned_by_user(sk)) {
5145 result = __tcp_checksum_complete(skb);
5148 result = __tcp_checksum_complete(skb);
5153 static inline bool tcp_checksum_complete_user(struct sock *sk,
5154 struct sk_buff *skb)
5156 return !skb_csum_unnecessary(skb) &&
5157 __tcp_checksum_complete_user(sk, skb);
5160 /* Does PAWS and seqno based validation of an incoming segment, flags will
5161 * play significant role here.
5163 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5164 const struct tcphdr *th, int syn_inerr)
5166 struct tcp_sock *tp = tcp_sk(sk);
5168 /* RFC1323: H1. Apply PAWS check first. */
5169 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
5170 tcp_paws_discard(sk, skb)) {
5172 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5173 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5174 LINUX_MIB_TCPACKSKIPPEDPAWS,
5175 &tp->last_oow_ack_time))
5176 tcp_send_dupack(sk, skb);
5179 /* Reset is accepted even if it did not pass PAWS. */
5182 /* Step 1: check sequence number */
5183 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5184 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5185 * (RST) segments are validated by checking their SEQ-fields."
5186 * And page 69: "If an incoming segment is not acceptable,
5187 * an acknowledgment should be sent in reply (unless the RST
5188 * bit is set, if so drop the segment and return)".
5193 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5194 LINUX_MIB_TCPACKSKIPPEDSEQ,
5195 &tp->last_oow_ack_time))
5196 tcp_send_dupack(sk, skb);
5201 /* Step 2: check RST bit */
5204 * If sequence number exactly matches RCV.NXT, then
5205 * RESET the connection
5207 * Send a challenge ACK
5209 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt)
5212 tcp_send_challenge_ack(sk, skb);
5216 /* step 3: check security and precedence [ignored] */
5218 /* step 4: Check for a SYN
5219 * RFC 5961 4.2 : Send a challenge ack
5224 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5225 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5226 tcp_send_challenge_ack(sk, skb);
5238 * TCP receive function for the ESTABLISHED state.
5240 * It is split into a fast path and a slow path. The fast path is
5242 * - A zero window was announced from us - zero window probing
5243 * is only handled properly in the slow path.
5244 * - Out of order segments arrived.
5245 * - Urgent data is expected.
5246 * - There is no buffer space left
5247 * - Unexpected TCP flags/window values/header lengths are received
5248 * (detected by checking the TCP header against pred_flags)
5249 * - Data is sent in both directions. Fast path only supports pure senders
5250 * or pure receivers (this means either the sequence number or the ack
5251 * value must stay constant)
5252 * - Unexpected TCP option.
5254 * When these conditions are not satisfied it drops into a standard
5255 * receive procedure patterned after RFC793 to handle all cases.
5256 * The first three cases are guaranteed by proper pred_flags setting,
5257 * the rest is checked inline. Fast processing is turned on in
5258 * tcp_data_queue when everything is OK.
5260 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5261 const struct tcphdr *th, unsigned int len)
5263 struct tcp_sock *tp = tcp_sk(sk);
5265 if (unlikely(!sk->sk_rx_dst))
5266 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
5268 * Header prediction.
5269 * The code loosely follows the one in the famous
5270 * "30 instruction TCP receive" Van Jacobson mail.
5272 * Van's trick is to deposit buffers into socket queue
5273 * on a device interrupt, to call tcp_recv function
5274 * on the receive process context and checksum and copy
5275 * the buffer to user space. smart...
5277 * Our current scheme is not silly either but we take the
5278 * extra cost of the net_bh soft interrupt processing...
5279 * We do checksum and copy also but from device to kernel.
5282 tp->rx_opt.saw_tstamp = 0;
5284 /* pred_flags is 0xS?10 << 16 + snd_wnd
5285 * if header_prediction is to be made
5286 * 'S' will always be tp->tcp_header_len >> 2
5287 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5288 * turn it off (when there are holes in the receive
5289 * space for instance)
5290 * PSH flag is ignored.
5293 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5294 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5295 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5296 int tcp_header_len = tp->tcp_header_len;
5298 /* Timestamp header prediction: tcp_header_len
5299 * is automatically equal to th->doff*4 due to pred_flags
5303 /* Check timestamp */
5304 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5305 /* No? Slow path! */
5306 if (!tcp_parse_aligned_timestamp(tp, th))
5309 /* If PAWS failed, check it more carefully in slow path */
5310 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5313 /* DO NOT update ts_recent here, if checksum fails
5314 * and timestamp was corrupted part, it will result
5315 * in a hung connection since we will drop all
5316 * future packets due to the PAWS test.
5320 if (len <= tcp_header_len) {
5321 /* Bulk data transfer: sender */
5322 if (len == tcp_header_len) {
5323 /* Predicted packet is in window by definition.
5324 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5325 * Hence, check seq<=rcv_wup reduces to:
5327 if (tcp_header_len ==
5328 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5329 tp->rcv_nxt == tp->rcv_wup)
5330 tcp_store_ts_recent(tp);
5332 /* We know that such packets are checksummed
5335 tcp_ack(sk, skb, 0);
5337 tcp_data_snd_check(sk);
5339 } else { /* Header too small */
5340 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5345 bool fragstolen = false;
5347 if (tp->ucopy.task == current &&
5348 tp->copied_seq == tp->rcv_nxt &&
5349 len - tcp_header_len <= tp->ucopy.len &&
5350 sock_owned_by_user(sk)) {
5351 __set_current_state(TASK_RUNNING);
5353 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len)) {
5354 /* Predicted packet is in window by definition.
5355 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5356 * Hence, check seq<=rcv_wup reduces to:
5358 if (tcp_header_len ==
5359 (sizeof(struct tcphdr) +
5360 TCPOLEN_TSTAMP_ALIGNED) &&
5361 tp->rcv_nxt == tp->rcv_wup)
5362 tcp_store_ts_recent(tp);
5364 tcp_rcv_rtt_measure_ts(sk, skb);
5366 __skb_pull(skb, tcp_header_len);
5367 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
5368 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER);
5373 if (tcp_checksum_complete_user(sk, skb))
5376 if ((int)skb->truesize > sk->sk_forward_alloc)
5379 /* Predicted packet is in window by definition.
5380 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5381 * Hence, check seq<=rcv_wup reduces to:
5383 if (tcp_header_len ==
5384 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5385 tp->rcv_nxt == tp->rcv_wup)
5386 tcp_store_ts_recent(tp);
5388 tcp_rcv_rtt_measure_ts(sk, skb);
5390 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS);
5392 /* Bulk data transfer: receiver */
5393 eaten = tcp_queue_rcv(sk, skb, tcp_header_len,
5397 tcp_event_data_recv(sk, skb);
5399 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5400 /* Well, only one small jumplet in fast path... */
5401 tcp_ack(sk, skb, FLAG_DATA);
5402 tcp_data_snd_check(sk);
5403 if (!inet_csk_ack_scheduled(sk))
5407 __tcp_ack_snd_check(sk, 0);
5410 kfree_skb_partial(skb, fragstolen);
5411 sk->sk_data_ready(sk);
5417 if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb))
5420 if (!th->ack && !th->rst && !th->syn)
5424 * Standard slow path.
5427 if (!tcp_validate_incoming(sk, skb, th, 1))
5431 if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0)
5434 tcp_rcv_rtt_measure_ts(sk, skb);
5436 /* Process urgent data. */
5437 tcp_urg(sk, skb, th);
5439 /* step 7: process the segment text */
5440 tcp_data_queue(sk, skb);
5442 tcp_data_snd_check(sk);
5443 tcp_ack_snd_check(sk);
5447 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_CSUMERRORS);
5448 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5453 EXPORT_SYMBOL(tcp_rcv_established);
5455 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5457 struct tcp_sock *tp = tcp_sk(sk);
5458 struct inet_connection_sock *icsk = inet_csk(sk);
5460 tcp_set_state(sk, TCP_ESTABLISHED);
5463 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
5464 security_inet_conn_established(sk, skb);
5467 /* Make sure socket is routed, for correct metrics. */
5468 icsk->icsk_af_ops->rebuild_header(sk);
5470 tcp_init_metrics(sk);
5472 tcp_init_congestion_control(sk);
5474 /* Prevent spurious tcp_cwnd_restart() on first data
5477 tp->lsndtime = tcp_time_stamp;
5479 tcp_init_buffer_space(sk);
5481 if (sock_flag(sk, SOCK_KEEPOPEN))
5482 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5484 if (!tp->rx_opt.snd_wscale)
5485 __tcp_fast_path_on(tp, tp->snd_wnd);
5489 if (!sock_flag(sk, SOCK_DEAD)) {
5490 sk->sk_state_change(sk);
5491 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5495 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
5496 struct tcp_fastopen_cookie *cookie)
5498 struct tcp_sock *tp = tcp_sk(sk);
5499 struct sk_buff *data = tp->syn_data ? tcp_write_queue_head(sk) : NULL;
5500 u16 mss = tp->rx_opt.mss_clamp, try_exp = 0;
5501 bool syn_drop = false;
5503 if (mss == tp->rx_opt.user_mss) {
5504 struct tcp_options_received opt;
5506 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5507 tcp_clear_options(&opt);
5508 opt.user_mss = opt.mss_clamp = 0;
5509 tcp_parse_options(synack, &opt, 0, NULL);
5510 mss = opt.mss_clamp;
5513 if (!tp->syn_fastopen) {
5514 /* Ignore an unsolicited cookie */
5516 } else if (tp->total_retrans) {
5517 /* SYN timed out and the SYN-ACK neither has a cookie nor
5518 * acknowledges data. Presumably the remote received only
5519 * the retransmitted (regular) SYNs: either the original
5520 * SYN-data or the corresponding SYN-ACK was dropped.
5522 syn_drop = (cookie->len < 0 && data);
5523 } else if (cookie->len < 0 && !tp->syn_data) {
5524 /* We requested a cookie but didn't get it. If we did not use
5525 * the (old) exp opt format then try so next time (try_exp=1).
5526 * Otherwise we go back to use the RFC7413 opt (try_exp=2).
5528 try_exp = tp->syn_fastopen_exp ? 2 : 1;
5531 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp);
5533 if (data) { /* Retransmit unacked data in SYN */
5534 tcp_for_write_queue_from(data, sk) {
5535 if (data == tcp_send_head(sk) ||
5536 __tcp_retransmit_skb(sk, data))
5540 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVEFAIL);
5543 tp->syn_data_acked = tp->syn_data;
5544 if (tp->syn_data_acked)
5545 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
5547 tcp_fastopen_add_skb(sk, synack);
5552 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5553 const struct tcphdr *th)
5555 struct inet_connection_sock *icsk = inet_csk(sk);
5556 struct tcp_sock *tp = tcp_sk(sk);
5557 struct tcp_fastopen_cookie foc = { .len = -1 };
5558 int saved_clamp = tp->rx_opt.mss_clamp;
5560 tcp_parse_options(skb, &tp->rx_opt, 0, &foc);
5561 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
5562 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
5566 * "If the state is SYN-SENT then
5567 * first check the ACK bit
5568 * If the ACK bit is set
5569 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5570 * a reset (unless the RST bit is set, if so drop
5571 * the segment and return)"
5573 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
5574 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt))
5575 goto reset_and_undo;
5577 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5578 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5580 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED);
5581 goto reset_and_undo;
5584 /* Now ACK is acceptable.
5586 * "If the RST bit is set
5587 * If the ACK was acceptable then signal the user "error:
5588 * connection reset", drop the segment, enter CLOSED state,
5589 * delete TCB, and return."
5598 * "fifth, if neither of the SYN or RST bits is set then
5599 * drop the segment and return."
5605 goto discard_and_undo;
5608 * "If the SYN bit is on ...
5609 * are acceptable then ...
5610 * (our SYN has been ACKed), change the connection
5611 * state to ESTABLISHED..."
5614 tcp_ecn_rcv_synack(tp, th);
5616 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5617 tcp_ack(sk, skb, FLAG_SLOWPATH);
5619 /* Ok.. it's good. Set up sequence numbers and
5620 * move to established.
5622 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5623 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5625 /* RFC1323: The window in SYN & SYN/ACK segments is
5628 tp->snd_wnd = ntohs(th->window);
5630 if (!tp->rx_opt.wscale_ok) {
5631 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5632 tp->window_clamp = min(tp->window_clamp, 65535U);
5635 if (tp->rx_opt.saw_tstamp) {
5636 tp->rx_opt.tstamp_ok = 1;
5637 tp->tcp_header_len =
5638 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5639 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5640 tcp_store_ts_recent(tp);
5642 tp->tcp_header_len = sizeof(struct tcphdr);
5645 if (tcp_is_sack(tp) && sysctl_tcp_fack)
5646 tcp_enable_fack(tp);
5649 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5650 tcp_initialize_rcv_mss(sk);
5652 /* Remember, tcp_poll() does not lock socket!
5653 * Change state from SYN-SENT only after copied_seq
5654 * is initialized. */
5655 tp->copied_seq = tp->rcv_nxt;
5659 tcp_finish_connect(sk, skb);
5661 if ((tp->syn_fastopen || tp->syn_data) &&
5662 tcp_rcv_fastopen_synack(sk, skb, &foc))
5665 if (sk->sk_write_pending ||
5666 icsk->icsk_accept_queue.rskq_defer_accept ||
5667 icsk->icsk_ack.pingpong) {
5668 /* Save one ACK. Data will be ready after
5669 * several ticks, if write_pending is set.
5671 * It may be deleted, but with this feature tcpdumps
5672 * look so _wonderfully_ clever, that I was not able
5673 * to stand against the temptation 8) --ANK
5675 inet_csk_schedule_ack(sk);
5676 icsk->icsk_ack.lrcvtime = tcp_time_stamp;
5677 tcp_enter_quickack_mode(sk);
5678 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5679 TCP_DELACK_MAX, TCP_RTO_MAX);
5690 /* No ACK in the segment */
5694 * "If the RST bit is set
5696 * Otherwise (no ACK) drop the segment and return."
5699 goto discard_and_undo;
5703 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5704 tcp_paws_reject(&tp->rx_opt, 0))
5705 goto discard_and_undo;
5708 /* We see SYN without ACK. It is attempt of
5709 * simultaneous connect with crossed SYNs.
5710 * Particularly, it can be connect to self.
5712 tcp_set_state(sk, TCP_SYN_RECV);
5714 if (tp->rx_opt.saw_tstamp) {
5715 tp->rx_opt.tstamp_ok = 1;
5716 tcp_store_ts_recent(tp);
5717 tp->tcp_header_len =
5718 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5720 tp->tcp_header_len = sizeof(struct tcphdr);
5723 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5724 tp->copied_seq = tp->rcv_nxt;
5725 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5727 /* RFC1323: The window in SYN & SYN/ACK segments is
5730 tp->snd_wnd = ntohs(th->window);
5731 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5732 tp->max_window = tp->snd_wnd;
5734 tcp_ecn_rcv_syn(tp, th);
5737 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5738 tcp_initialize_rcv_mss(sk);
5740 tcp_send_synack(sk);
5742 /* Note, we could accept data and URG from this segment.
5743 * There are no obstacles to make this (except that we must
5744 * either change tcp_recvmsg() to prevent it from returning data
5745 * before 3WHS completes per RFC793, or employ TCP Fast Open).
5747 * However, if we ignore data in ACKless segments sometimes,
5748 * we have no reasons to accept it sometimes.
5749 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5750 * is not flawless. So, discard packet for sanity.
5751 * Uncomment this return to process the data.
5758 /* "fifth, if neither of the SYN or RST bits is set then
5759 * drop the segment and return."
5763 tcp_clear_options(&tp->rx_opt);
5764 tp->rx_opt.mss_clamp = saved_clamp;
5768 tcp_clear_options(&tp->rx_opt);
5769 tp->rx_opt.mss_clamp = saved_clamp;
5774 * This function implements the receiving procedure of RFC 793 for
5775 * all states except ESTABLISHED and TIME_WAIT.
5776 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5777 * address independent.
5780 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb)
5782 struct tcp_sock *tp = tcp_sk(sk);
5783 struct inet_connection_sock *icsk = inet_csk(sk);
5784 const struct tcphdr *th = tcp_hdr(skb);
5785 struct request_sock *req;
5789 tp->rx_opt.saw_tstamp = 0;
5791 switch (sk->sk_state) {
5805 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
5808 /* Now we have several options: In theory there is
5809 * nothing else in the frame. KA9Q has an option to
5810 * send data with the syn, BSD accepts data with the
5811 * syn up to the [to be] advertised window and
5812 * Solaris 2.1 gives you a protocol error. For now
5813 * we just ignore it, that fits the spec precisely
5814 * and avoids incompatibilities. It would be nice in
5815 * future to drop through and process the data.
5817 * Now that TTCP is starting to be used we ought to
5819 * But, this leaves one open to an easy denial of
5820 * service attack, and SYN cookies can't defend
5821 * against this problem. So, we drop the data
5822 * in the interest of security over speed unless
5823 * it's still in use.
5831 queued = tcp_rcv_synsent_state_process(sk, skb, th);
5835 /* Do step6 onward by hand. */
5836 tcp_urg(sk, skb, th);
5838 tcp_data_snd_check(sk);
5842 req = tp->fastopen_rsk;
5844 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
5845 sk->sk_state != TCP_FIN_WAIT1);
5847 if (!tcp_check_req(sk, skb, req, true))
5851 if (!th->ack && !th->rst && !th->syn)
5854 if (!tcp_validate_incoming(sk, skb, th, 0))
5857 /* step 5: check the ACK field */
5858 acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
5859 FLAG_UPDATE_TS_RECENT) > 0;
5861 switch (sk->sk_state) {
5867 tcp_synack_rtt_meas(sk, req);
5869 /* Once we leave TCP_SYN_RECV, we no longer need req
5873 tp->total_retrans = req->num_retrans;
5874 reqsk_fastopen_remove(sk, req, false);
5876 /* Make sure socket is routed, for correct metrics. */
5877 icsk->icsk_af_ops->rebuild_header(sk);
5878 tcp_init_congestion_control(sk);
5881 tp->copied_seq = tp->rcv_nxt;
5882 tcp_init_buffer_space(sk);
5885 tcp_set_state(sk, TCP_ESTABLISHED);
5886 sk->sk_state_change(sk);
5888 /* Note, that this wakeup is only for marginal crossed SYN case.
5889 * Passively open sockets are not waked up, because
5890 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
5893 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5895 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5896 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
5897 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5899 if (tp->rx_opt.tstamp_ok)
5900 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5903 /* Re-arm the timer because data may have been sent out.
5904 * This is similar to the regular data transmission case
5905 * when new data has just been ack'ed.
5907 * (TFO) - we could try to be more aggressive and
5908 * retransmitting any data sooner based on when they
5913 tcp_init_metrics(sk);
5915 tcp_update_pacing_rate(sk);
5917 /* Prevent spurious tcp_cwnd_restart() on first data packet */
5918 tp->lsndtime = tcp_time_stamp;
5920 tcp_initialize_rcv_mss(sk);
5921 tcp_fast_path_on(tp);
5924 case TCP_FIN_WAIT1: {
5925 struct dst_entry *dst;
5928 /* If we enter the TCP_FIN_WAIT1 state and we are a
5929 * Fast Open socket and this is the first acceptable
5930 * ACK we have received, this would have acknowledged
5931 * our SYNACK so stop the SYNACK timer.
5934 /* Return RST if ack_seq is invalid.
5935 * Note that RFC793 only says to generate a
5936 * DUPACK for it but for TCP Fast Open it seems
5937 * better to treat this case like TCP_SYN_RECV
5942 /* We no longer need the request sock. */
5943 reqsk_fastopen_remove(sk, req, false);
5946 if (tp->snd_una != tp->write_seq)
5949 tcp_set_state(sk, TCP_FIN_WAIT2);
5950 sk->sk_shutdown |= SEND_SHUTDOWN;
5952 dst = __sk_dst_get(sk);
5956 if (!sock_flag(sk, SOCK_DEAD)) {
5957 /* Wake up lingering close() */
5958 sk->sk_state_change(sk);
5962 if (tp->linger2 < 0 ||
5963 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5964 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
5966 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5970 tmo = tcp_fin_time(sk);
5971 if (tmo > TCP_TIMEWAIT_LEN) {
5972 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5973 } else if (th->fin || sock_owned_by_user(sk)) {
5974 /* Bad case. We could lose such FIN otherwise.
5975 * It is not a big problem, but it looks confusing
5976 * and not so rare event. We still can lose it now,
5977 * if it spins in bh_lock_sock(), but it is really
5980 inet_csk_reset_keepalive_timer(sk, tmo);
5982 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5989 if (tp->snd_una == tp->write_seq) {
5990 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
5996 if (tp->snd_una == tp->write_seq) {
5997 tcp_update_metrics(sk);
6004 /* step 6: check the URG bit */
6005 tcp_urg(sk, skb, th);
6007 /* step 7: process the segment text */
6008 switch (sk->sk_state) {
6009 case TCP_CLOSE_WAIT:
6012 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
6016 /* RFC 793 says to queue data in these states,
6017 * RFC 1122 says we MUST send a reset.
6018 * BSD 4.4 also does reset.
6020 if (sk->sk_shutdown & RCV_SHUTDOWN) {
6021 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6022 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6023 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6029 case TCP_ESTABLISHED:
6030 tcp_data_queue(sk, skb);
6035 /* tcp_data could move socket to TIME-WAIT */
6036 if (sk->sk_state != TCP_CLOSE) {
6037 tcp_data_snd_check(sk);
6038 tcp_ack_snd_check(sk);
6047 EXPORT_SYMBOL(tcp_rcv_state_process);
6049 static inline void pr_drop_req(struct request_sock *req, __u16 port, int family)
6051 struct inet_request_sock *ireq = inet_rsk(req);
6053 if (family == AF_INET)
6054 net_dbg_ratelimited("drop open request from %pI4/%u\n",
6055 &ireq->ir_rmt_addr, port);
6056 #if IS_ENABLED(CONFIG_IPV6)
6057 else if (family == AF_INET6)
6058 net_dbg_ratelimited("drop open request from %pI6/%u\n",
6059 &ireq->ir_v6_rmt_addr, port);
6063 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
6065 * If we receive a SYN packet with these bits set, it means a
6066 * network is playing bad games with TOS bits. In order to
6067 * avoid possible false congestion notifications, we disable
6068 * TCP ECN negotiation.
6070 * Exception: tcp_ca wants ECN. This is required for DCTCP
6071 * congestion control: Linux DCTCP asserts ECT on all packets,
6072 * including SYN, which is most optimal solution; however,
6073 * others, such as FreeBSD do not.
6075 static void tcp_ecn_create_request(struct request_sock *req,
6076 const struct sk_buff *skb,
6077 const struct sock *listen_sk,
6078 const struct dst_entry *dst)
6080 const struct tcphdr *th = tcp_hdr(skb);
6081 const struct net *net = sock_net(listen_sk);
6082 bool th_ecn = th->ece && th->cwr;
6089 ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield);
6090 ecn_ok_dst = dst_feature(dst, DST_FEATURE_ECN_MASK);
6091 ecn_ok = net->ipv4.sysctl_tcp_ecn || ecn_ok_dst;
6093 if ((!ect && ecn_ok) || tcp_ca_needs_ecn(listen_sk) ||
6094 (ecn_ok_dst & DST_FEATURE_ECN_CA))
6095 inet_rsk(req)->ecn_ok = 1;
6098 static void tcp_openreq_init(struct request_sock *req,
6099 const struct tcp_options_received *rx_opt,
6100 struct sk_buff *skb, const struct sock *sk)
6102 struct inet_request_sock *ireq = inet_rsk(req);
6104 req->rsk_rcv_wnd = 0; /* So that tcp_send_synack() knows! */
6106 tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq;
6107 tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
6108 skb_mstamp_get(&tcp_rsk(req)->snt_synack);
6109 tcp_rsk(req)->last_oow_ack_time = 0;
6110 req->mss = rx_opt->mss_clamp;
6111 req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0;
6112 ireq->tstamp_ok = rx_opt->tstamp_ok;
6113 ireq->sack_ok = rx_opt->sack_ok;
6114 ireq->snd_wscale = rx_opt->snd_wscale;
6115 ireq->wscale_ok = rx_opt->wscale_ok;
6118 ireq->ir_rmt_port = tcp_hdr(skb)->source;
6119 ireq->ir_num = ntohs(tcp_hdr(skb)->dest);
6120 ireq->ir_mark = inet_request_mark(sk, skb);
6123 struct request_sock *inet_reqsk_alloc(const struct request_sock_ops *ops,
6124 struct sock *sk_listener,
6125 bool attach_listener)
6127 struct request_sock *req = reqsk_alloc(ops, sk_listener,
6131 struct inet_request_sock *ireq = inet_rsk(req);
6133 kmemcheck_annotate_bitfield(ireq, flags);
6135 atomic64_set(&ireq->ir_cookie, 0);
6136 ireq->ireq_state = TCP_NEW_SYN_RECV;
6137 write_pnet(&ireq->ireq_net, sock_net(sk_listener));
6138 ireq->ireq_family = sk_listener->sk_family;
6143 EXPORT_SYMBOL(inet_reqsk_alloc);
6146 * Return true if a syncookie should be sent
6148 static bool tcp_syn_flood_action(const struct sock *sk,
6149 const struct sk_buff *skb,
6152 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
6153 const char *msg = "Dropping request";
6154 bool want_cookie = false;
6156 #ifdef CONFIG_SYN_COOKIES
6157 if (sysctl_tcp_syncookies) {
6158 msg = "Sending cookies";
6160 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES);
6163 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP);
6165 if (!queue->synflood_warned &&
6166 sysctl_tcp_syncookies != 2 &&
6167 xchg(&queue->synflood_warned, 1) == 0)
6168 pr_info("%s: Possible SYN flooding on port %d. %s. Check SNMP counters.\n",
6169 proto, ntohs(tcp_hdr(skb)->dest), msg);
6174 static void tcp_reqsk_record_syn(const struct sock *sk,
6175 struct request_sock *req,
6176 const struct sk_buff *skb)
6178 if (tcp_sk(sk)->save_syn) {
6179 u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb);
6182 copy = kmalloc(len + sizeof(u32), GFP_ATOMIC);
6185 memcpy(©[1], skb_network_header(skb), len);
6186 req->saved_syn = copy;
6191 int tcp_conn_request(struct request_sock_ops *rsk_ops,
6192 const struct tcp_request_sock_ops *af_ops,
6193 struct sock *sk, struct sk_buff *skb)
6195 struct tcp_fastopen_cookie foc = { .len = -1 };
6196 __u32 isn = TCP_SKB_CB(skb)->tcp_tw_isn;
6197 struct tcp_options_received tmp_opt;
6198 struct tcp_sock *tp = tcp_sk(sk);
6199 struct sock *fastopen_sk = NULL;
6200 struct dst_entry *dst = NULL;
6201 struct request_sock *req;
6202 bool want_cookie = false;
6205 /* TW buckets are converted to open requests without
6206 * limitations, they conserve resources and peer is
6207 * evidently real one.
6209 if ((sysctl_tcp_syncookies == 2 ||
6210 inet_csk_reqsk_queue_is_full(sk)) && !isn) {
6211 want_cookie = tcp_syn_flood_action(sk, skb, rsk_ops->slab_name);
6217 /* Accept backlog is full. If we have already queued enough
6218 * of warm entries in syn queue, drop request. It is better than
6219 * clogging syn queue with openreqs with exponentially increasing
6222 if (sk_acceptq_is_full(sk) && inet_csk_reqsk_queue_young(sk) > 1) {
6223 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
6227 req = inet_reqsk_alloc(rsk_ops, sk, !want_cookie);
6231 tcp_rsk(req)->af_specific = af_ops;
6233 tcp_clear_options(&tmp_opt);
6234 tmp_opt.mss_clamp = af_ops->mss_clamp;
6235 tmp_opt.user_mss = tp->rx_opt.user_mss;
6236 tcp_parse_options(skb, &tmp_opt, 0, want_cookie ? NULL : &foc);
6238 if (want_cookie && !tmp_opt.saw_tstamp)
6239 tcp_clear_options(&tmp_opt);
6241 tmp_opt.tstamp_ok = tmp_opt.saw_tstamp;
6242 tcp_openreq_init(req, &tmp_opt, skb, sk);
6244 /* Note: tcp_v6_init_req() might override ir_iif for link locals */
6245 inet_rsk(req)->ir_iif = inet_request_bound_dev_if(sk, skb);
6247 af_ops->init_req(req, sk, skb);
6249 if (security_inet_conn_request(sk, skb, req))
6252 if (!want_cookie && !isn) {
6253 /* VJ's idea. We save last timestamp seen
6254 * from the destination in peer table, when entering
6255 * state TIME-WAIT, and check against it before
6256 * accepting new connection request.
6258 * If "isn" is not zero, this request hit alive
6259 * timewait bucket, so that all the necessary checks
6260 * are made in the function processing timewait state.
6262 if (tcp_death_row.sysctl_tw_recycle) {
6265 dst = af_ops->route_req(sk, &fl, req, &strict);
6267 if (dst && strict &&
6268 !tcp_peer_is_proven(req, dst, true,
6269 tmp_opt.saw_tstamp)) {
6270 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSPASSIVEREJECTED);
6271 goto drop_and_release;
6274 /* Kill the following clause, if you dislike this way. */
6275 else if (!sysctl_tcp_syncookies &&
6276 (sysctl_max_syn_backlog - inet_csk_reqsk_queue_len(sk) <
6277 (sysctl_max_syn_backlog >> 2)) &&
6278 !tcp_peer_is_proven(req, dst, false,
6279 tmp_opt.saw_tstamp)) {
6280 /* Without syncookies last quarter of
6281 * backlog is filled with destinations,
6282 * proven to be alive.
6283 * It means that we continue to communicate
6284 * to destinations, already remembered
6285 * to the moment of synflood.
6287 pr_drop_req(req, ntohs(tcp_hdr(skb)->source),
6289 goto drop_and_release;
6292 isn = af_ops->init_seq(skb);
6295 dst = af_ops->route_req(sk, &fl, req, NULL);
6300 tcp_ecn_create_request(req, skb, sk, dst);
6303 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss);
6304 req->cookie_ts = tmp_opt.tstamp_ok;
6305 if (!tmp_opt.tstamp_ok)
6306 inet_rsk(req)->ecn_ok = 0;
6309 tcp_rsk(req)->snt_isn = isn;
6310 tcp_rsk(req)->txhash = net_tx_rndhash();
6311 tcp_openreq_init_rwin(req, sk, dst);
6313 tcp_reqsk_record_syn(sk, req, skb);
6314 fastopen_sk = tcp_try_fastopen(sk, skb, req, &foc, dst);
6317 af_ops->send_synack(fastopen_sk, dst, &fl, req,
6319 /* Add the child socket directly into the accept queue */
6320 inet_csk_reqsk_queue_add(sk, req, fastopen_sk);
6321 sk->sk_data_ready(sk);
6322 bh_unlock_sock(fastopen_sk);
6323 sock_put(fastopen_sk);
6325 tcp_rsk(req)->tfo_listener = false;
6327 inet_csk_reqsk_queue_hash_add(sk, req, TCP_TIMEOUT_INIT);
6328 af_ops->send_synack(sk, dst, &fl, req,
6329 &foc, !want_cookie);
6341 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_LISTENDROPS);
6344 EXPORT_SYMBOL(tcp_conn_request);