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_max_reordering __read_mostly = 300;
84 int sysctl_tcp_dsack __read_mostly = 1;
85 int sysctl_tcp_app_win __read_mostly = 31;
86 int sysctl_tcp_adv_win_scale __read_mostly = 1;
87 EXPORT_SYMBOL(sysctl_tcp_adv_win_scale);
89 /* rfc5961 challenge ack rate limiting */
90 int sysctl_tcp_challenge_ack_limit = 100;
92 int sysctl_tcp_stdurg __read_mostly;
93 int sysctl_tcp_rfc1337 __read_mostly;
94 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
95 int sysctl_tcp_frto __read_mostly = 2;
96 int sysctl_tcp_min_rtt_wlen __read_mostly = 300;
98 int sysctl_tcp_thin_dupack __read_mostly;
100 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
101 int sysctl_tcp_early_retrans __read_mostly = 3;
102 int sysctl_tcp_invalid_ratelimit __read_mostly = HZ/2;
104 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
105 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
106 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
107 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
108 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
109 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
110 #define FLAG_ECE 0x40 /* ECE in this ACK */
111 #define FLAG_LOST_RETRANS 0x80 /* This ACK marks some retransmission lost */
112 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
113 #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
114 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
115 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
116 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
117 #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
119 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
120 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
121 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
122 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
124 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
125 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
127 #define REXMIT_NONE 0 /* no loss recovery to do */
128 #define REXMIT_LOST 1 /* retransmit packets marked lost */
129 #define REXMIT_NEW 2 /* FRTO-style transmit of unsent/new packets */
131 /* Adapt the MSS value used to make delayed ack decision to the
134 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
136 struct inet_connection_sock *icsk = inet_csk(sk);
137 const unsigned int lss = icsk->icsk_ack.last_seg_size;
140 icsk->icsk_ack.last_seg_size = 0;
142 /* skb->len may jitter because of SACKs, even if peer
143 * sends good full-sized frames.
145 len = skb_shinfo(skb)->gso_size ? : skb->len;
146 if (len >= icsk->icsk_ack.rcv_mss) {
147 icsk->icsk_ack.rcv_mss = len;
149 /* Otherwise, we make more careful check taking into account,
150 * that SACKs block is variable.
152 * "len" is invariant segment length, including TCP header.
154 len += skb->data - skb_transport_header(skb);
155 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
156 /* If PSH is not set, packet should be
157 * full sized, provided peer TCP is not badly broken.
158 * This observation (if it is correct 8)) allows
159 * to handle super-low mtu links fairly.
161 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
162 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
163 /* Subtract also invariant (if peer is RFC compliant),
164 * tcp header plus fixed timestamp option length.
165 * Resulting "len" is MSS free of SACK jitter.
167 len -= tcp_sk(sk)->tcp_header_len;
168 icsk->icsk_ack.last_seg_size = len;
170 icsk->icsk_ack.rcv_mss = len;
174 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
175 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
176 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
180 static void tcp_incr_quickack(struct sock *sk)
182 struct inet_connection_sock *icsk = inet_csk(sk);
183 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
187 if (quickacks > icsk->icsk_ack.quick)
188 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
191 static void tcp_enter_quickack_mode(struct sock *sk)
193 struct inet_connection_sock *icsk = inet_csk(sk);
194 tcp_incr_quickack(sk);
195 icsk->icsk_ack.pingpong = 0;
196 icsk->icsk_ack.ato = TCP_ATO_MIN;
199 /* Send ACKs quickly, if "quick" count is not exhausted
200 * and the session is not interactive.
203 static bool tcp_in_quickack_mode(struct sock *sk)
205 const struct inet_connection_sock *icsk = inet_csk(sk);
206 const struct dst_entry *dst = __sk_dst_get(sk);
208 return (dst && dst_metric(dst, RTAX_QUICKACK)) ||
209 (icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong);
212 static void tcp_ecn_queue_cwr(struct tcp_sock *tp)
214 if (tp->ecn_flags & TCP_ECN_OK)
215 tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
218 static void tcp_ecn_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb)
220 if (tcp_hdr(skb)->cwr)
221 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
224 static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp)
226 tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
229 static void __tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
231 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
232 case INET_ECN_NOT_ECT:
233 /* Funny extension: if ECT is not set on a segment,
234 * and we already seen ECT on a previous segment,
235 * it is probably a retransmit.
237 if (tp->ecn_flags & TCP_ECN_SEEN)
238 tcp_enter_quickack_mode((struct sock *)tp);
241 if (tcp_ca_needs_ecn((struct sock *)tp))
242 tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_IS_CE);
244 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
245 /* Better not delay acks, sender can have a very low cwnd */
246 tcp_enter_quickack_mode((struct sock *)tp);
247 tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
249 tp->ecn_flags |= TCP_ECN_SEEN;
252 if (tcp_ca_needs_ecn((struct sock *)tp))
253 tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_NO_CE);
254 tp->ecn_flags |= TCP_ECN_SEEN;
259 static void tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
261 if (tp->ecn_flags & TCP_ECN_OK)
262 __tcp_ecn_check_ce(tp, skb);
265 static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
267 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
268 tp->ecn_flags &= ~TCP_ECN_OK;
271 static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
273 if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
274 tp->ecn_flags &= ~TCP_ECN_OK;
277 static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
279 if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
284 /* Buffer size and advertised window tuning.
286 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
289 static void tcp_sndbuf_expand(struct sock *sk)
291 const struct tcp_sock *tp = tcp_sk(sk);
295 /* Worst case is non GSO/TSO : each frame consumes one skb
296 * and skb->head is kmalloced using power of two area of memory
298 per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
300 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
302 per_mss = roundup_pow_of_two(per_mss) +
303 SKB_DATA_ALIGN(sizeof(struct sk_buff));
305 nr_segs = max_t(u32, TCP_INIT_CWND, tp->snd_cwnd);
306 nr_segs = max_t(u32, nr_segs, tp->reordering + 1);
308 /* Fast Recovery (RFC 5681 3.2) :
309 * Cubic needs 1.7 factor, rounded to 2 to include
310 * extra cushion (application might react slowly to POLLOUT)
312 sndmem = 2 * nr_segs * per_mss;
314 if (sk->sk_sndbuf < sndmem)
315 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
318 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
320 * All tcp_full_space() is split to two parts: "network" buffer, allocated
321 * forward and advertised in receiver window (tp->rcv_wnd) and
322 * "application buffer", required to isolate scheduling/application
323 * latencies from network.
324 * window_clamp is maximal advertised window. It can be less than
325 * tcp_full_space(), in this case tcp_full_space() - window_clamp
326 * is reserved for "application" buffer. The less window_clamp is
327 * the smoother our behaviour from viewpoint of network, but the lower
328 * throughput and the higher sensitivity of the connection to losses. 8)
330 * rcv_ssthresh is more strict window_clamp used at "slow start"
331 * phase to predict further behaviour of this connection.
332 * It is used for two goals:
333 * - to enforce header prediction at sender, even when application
334 * requires some significant "application buffer". It is check #1.
335 * - to prevent pruning of receive queue because of misprediction
336 * of receiver window. Check #2.
338 * The scheme does not work when sender sends good segments opening
339 * window and then starts to feed us spaghetti. But it should work
340 * in common situations. Otherwise, we have to rely on queue collapsing.
343 /* Slow part of check#2. */
344 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
346 struct tcp_sock *tp = tcp_sk(sk);
348 int truesize = tcp_win_from_space(skb->truesize) >> 1;
349 int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1;
351 while (tp->rcv_ssthresh <= window) {
352 if (truesize <= skb->len)
353 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
361 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
363 struct tcp_sock *tp = tcp_sk(sk);
366 if (tp->rcv_ssthresh < tp->window_clamp &&
367 (int)tp->rcv_ssthresh < tcp_space(sk) &&
368 !tcp_under_memory_pressure(sk)) {
371 /* Check #2. Increase window, if skb with such overhead
372 * will fit to rcvbuf in future.
374 if (tcp_win_from_space(skb->truesize) <= skb->len)
375 incr = 2 * tp->advmss;
377 incr = __tcp_grow_window(sk, skb);
380 incr = max_t(int, incr, 2 * skb->len);
381 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
383 inet_csk(sk)->icsk_ack.quick |= 1;
388 /* 3. Tuning rcvbuf, when connection enters established state. */
389 static void tcp_fixup_rcvbuf(struct sock *sk)
391 u32 mss = tcp_sk(sk)->advmss;
394 rcvmem = 2 * SKB_TRUESIZE(mss + MAX_TCP_HEADER) *
395 tcp_default_init_rwnd(mss);
397 /* Dynamic Right Sizing (DRS) has 2 to 3 RTT latency
398 * Allow enough cushion so that sender is not limited by our window
400 if (sysctl_tcp_moderate_rcvbuf)
403 if (sk->sk_rcvbuf < rcvmem)
404 sk->sk_rcvbuf = min(rcvmem, sysctl_tcp_rmem[2]);
407 /* 4. Try to fixup all. It is made immediately after connection enters
410 void tcp_init_buffer_space(struct sock *sk)
412 struct tcp_sock *tp = tcp_sk(sk);
415 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
416 tcp_fixup_rcvbuf(sk);
417 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
418 tcp_sndbuf_expand(sk);
420 tp->rcvq_space.space = tp->rcv_wnd;
421 tp->rcvq_space.time = tcp_time_stamp;
422 tp->rcvq_space.seq = tp->copied_seq;
424 maxwin = tcp_full_space(sk);
426 if (tp->window_clamp >= maxwin) {
427 tp->window_clamp = maxwin;
429 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
430 tp->window_clamp = max(maxwin -
431 (maxwin >> sysctl_tcp_app_win),
435 /* Force reservation of one segment. */
436 if (sysctl_tcp_app_win &&
437 tp->window_clamp > 2 * tp->advmss &&
438 tp->window_clamp + tp->advmss > maxwin)
439 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
441 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
442 tp->snd_cwnd_stamp = tcp_time_stamp;
445 /* 5. Recalculate window clamp after socket hit its memory bounds. */
446 static void tcp_clamp_window(struct sock *sk)
448 struct tcp_sock *tp = tcp_sk(sk);
449 struct inet_connection_sock *icsk = inet_csk(sk);
451 icsk->icsk_ack.quick = 0;
453 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
454 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
455 !tcp_under_memory_pressure(sk) &&
456 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
457 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
460 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
461 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
464 /* Initialize RCV_MSS value.
465 * RCV_MSS is an our guess about MSS used by the peer.
466 * We haven't any direct information about the MSS.
467 * It's better to underestimate the RCV_MSS rather than overestimate.
468 * Overestimations make us ACKing less frequently than needed.
469 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
471 void tcp_initialize_rcv_mss(struct sock *sk)
473 const struct tcp_sock *tp = tcp_sk(sk);
474 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
476 hint = min(hint, tp->rcv_wnd / 2);
477 hint = min(hint, TCP_MSS_DEFAULT);
478 hint = max(hint, TCP_MIN_MSS);
480 inet_csk(sk)->icsk_ack.rcv_mss = hint;
482 EXPORT_SYMBOL(tcp_initialize_rcv_mss);
484 /* Receiver "autotuning" code.
486 * The algorithm for RTT estimation w/o timestamps is based on
487 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
488 * <http://public.lanl.gov/radiant/pubs.html#DRS>
490 * More detail on this code can be found at
491 * <http://staff.psc.edu/jheffner/>,
492 * though this reference is out of date. A new paper
495 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
497 u32 new_sample = tp->rcv_rtt_est.rtt;
503 if (new_sample != 0) {
504 /* If we sample in larger samples in the non-timestamp
505 * case, we could grossly overestimate the RTT especially
506 * with chatty applications or bulk transfer apps which
507 * are stalled on filesystem I/O.
509 * Also, since we are only going for a minimum in the
510 * non-timestamp case, we do not smooth things out
511 * else with timestamps disabled convergence takes too
515 m -= (new_sample >> 3);
523 /* No previous measure. */
527 if (tp->rcv_rtt_est.rtt != new_sample)
528 tp->rcv_rtt_est.rtt = new_sample;
531 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
533 if (tp->rcv_rtt_est.time == 0)
535 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
537 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rcv_rtt_est.time, 1);
540 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
541 tp->rcv_rtt_est.time = tcp_time_stamp;
544 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
545 const struct sk_buff *skb)
547 struct tcp_sock *tp = tcp_sk(sk);
548 if (tp->rx_opt.rcv_tsecr &&
549 (TCP_SKB_CB(skb)->end_seq -
550 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
551 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
555 * This function should be called every time data is copied to user space.
556 * It calculates the appropriate TCP receive buffer space.
558 void tcp_rcv_space_adjust(struct sock *sk)
560 struct tcp_sock *tp = tcp_sk(sk);
564 time = tcp_time_stamp - tp->rcvq_space.time;
565 if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0)
568 /* Number of bytes copied to user in last RTT */
569 copied = tp->copied_seq - tp->rcvq_space.seq;
570 if (copied <= tp->rcvq_space.space)
574 * copied = bytes received in previous RTT, our base window
575 * To cope with packet losses, we need a 2x factor
576 * To cope with slow start, and sender growing its cwin by 100 %
577 * every RTT, we need a 4x factor, because the ACK we are sending
578 * now is for the next RTT, not the current one :
579 * <prev RTT . ><current RTT .. ><next RTT .... >
582 if (sysctl_tcp_moderate_rcvbuf &&
583 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
584 int rcvwin, rcvmem, rcvbuf;
586 /* minimal window to cope with packet losses, assuming
587 * steady state. Add some cushion because of small variations.
589 rcvwin = (copied << 1) + 16 * tp->advmss;
591 /* If rate increased by 25%,
592 * assume slow start, rcvwin = 3 * copied
593 * If rate increased by 50%,
594 * assume sender can use 2x growth, rcvwin = 4 * copied
597 tp->rcvq_space.space + (tp->rcvq_space.space >> 2)) {
599 tp->rcvq_space.space + (tp->rcvq_space.space >> 1))
602 rcvwin += (rcvwin >> 1);
605 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
606 while (tcp_win_from_space(rcvmem) < tp->advmss)
609 rcvbuf = min(rcvwin / tp->advmss * rcvmem, sysctl_tcp_rmem[2]);
610 if (rcvbuf > sk->sk_rcvbuf) {
611 sk->sk_rcvbuf = rcvbuf;
613 /* Make the window clamp follow along. */
614 tp->window_clamp = rcvwin;
617 tp->rcvq_space.space = copied;
620 tp->rcvq_space.seq = tp->copied_seq;
621 tp->rcvq_space.time = tcp_time_stamp;
624 /* There is something which you must keep in mind when you analyze the
625 * behavior of the tp->ato delayed ack timeout interval. When a
626 * connection starts up, we want to ack as quickly as possible. The
627 * problem is that "good" TCP's do slow start at the beginning of data
628 * transmission. The means that until we send the first few ACK's the
629 * sender will sit on his end and only queue most of his data, because
630 * he can only send snd_cwnd unacked packets at any given time. For
631 * each ACK we send, he increments snd_cwnd and transmits more of his
634 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
636 struct tcp_sock *tp = tcp_sk(sk);
637 struct inet_connection_sock *icsk = inet_csk(sk);
640 inet_csk_schedule_ack(sk);
642 tcp_measure_rcv_mss(sk, skb);
644 tcp_rcv_rtt_measure(tp);
646 now = tcp_time_stamp;
648 if (!icsk->icsk_ack.ato) {
649 /* The _first_ data packet received, initialize
650 * delayed ACK engine.
652 tcp_incr_quickack(sk);
653 icsk->icsk_ack.ato = TCP_ATO_MIN;
655 int m = now - icsk->icsk_ack.lrcvtime;
657 if (m <= TCP_ATO_MIN / 2) {
658 /* The fastest case is the first. */
659 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
660 } else if (m < icsk->icsk_ack.ato) {
661 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
662 if (icsk->icsk_ack.ato > icsk->icsk_rto)
663 icsk->icsk_ack.ato = icsk->icsk_rto;
664 } else if (m > icsk->icsk_rto) {
665 /* Too long gap. Apparently sender failed to
666 * restart window, so that we send ACKs quickly.
668 tcp_incr_quickack(sk);
672 icsk->icsk_ack.lrcvtime = now;
674 tcp_ecn_check_ce(tp, skb);
677 tcp_grow_window(sk, skb);
680 /* Called to compute a smoothed rtt estimate. The data fed to this
681 * routine either comes from timestamps, or from segments that were
682 * known _not_ to have been retransmitted [see Karn/Partridge
683 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
684 * piece by Van Jacobson.
685 * NOTE: the next three routines used to be one big routine.
686 * To save cycles in the RFC 1323 implementation it was better to break
687 * it up into three procedures. -- erics
689 static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
691 struct tcp_sock *tp = tcp_sk(sk);
692 long m = mrtt_us; /* RTT */
693 u32 srtt = tp->srtt_us;
695 /* The following amusing code comes from Jacobson's
696 * article in SIGCOMM '88. Note that rtt and mdev
697 * are scaled versions of rtt and mean deviation.
698 * This is designed to be as fast as possible
699 * m stands for "measurement".
701 * On a 1990 paper the rto value is changed to:
702 * RTO = rtt + 4 * mdev
704 * Funny. This algorithm seems to be very broken.
705 * These formulae increase RTO, when it should be decreased, increase
706 * too slowly, when it should be increased quickly, decrease too quickly
707 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
708 * does not matter how to _calculate_ it. Seems, it was trap
709 * that VJ failed to avoid. 8)
712 m -= (srtt >> 3); /* m is now error in rtt est */
713 srtt += m; /* rtt = 7/8 rtt + 1/8 new */
715 m = -m; /* m is now abs(error) */
716 m -= (tp->mdev_us >> 2); /* similar update on mdev */
717 /* This is similar to one of Eifel findings.
718 * Eifel blocks mdev updates when rtt decreases.
719 * This solution is a bit different: we use finer gain
720 * for mdev in this case (alpha*beta).
721 * Like Eifel it also prevents growth of rto,
722 * but also it limits too fast rto decreases,
723 * happening in pure Eifel.
728 m -= (tp->mdev_us >> 2); /* similar update on mdev */
730 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */
731 if (tp->mdev_us > tp->mdev_max_us) {
732 tp->mdev_max_us = tp->mdev_us;
733 if (tp->mdev_max_us > tp->rttvar_us)
734 tp->rttvar_us = tp->mdev_max_us;
736 if (after(tp->snd_una, tp->rtt_seq)) {
737 if (tp->mdev_max_us < tp->rttvar_us)
738 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
739 tp->rtt_seq = tp->snd_nxt;
740 tp->mdev_max_us = tcp_rto_min_us(sk);
743 /* no previous measure. */
744 srtt = m << 3; /* take the measured time to be rtt */
745 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */
746 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
747 tp->mdev_max_us = tp->rttvar_us;
748 tp->rtt_seq = tp->snd_nxt;
750 tp->srtt_us = max(1U, srtt);
753 /* Set the sk_pacing_rate to allow proper sizing of TSO packets.
754 * Note: TCP stack does not yet implement pacing.
755 * FQ packet scheduler can be used to implement cheap but effective
756 * TCP pacing, to smooth the burst on large writes when packets
757 * in flight is significantly lower than cwnd (or rwin)
759 int sysctl_tcp_pacing_ss_ratio __read_mostly = 200;
760 int sysctl_tcp_pacing_ca_ratio __read_mostly = 120;
762 static void tcp_update_pacing_rate(struct sock *sk)
764 const struct tcp_sock *tp = tcp_sk(sk);
767 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
768 rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3);
770 /* current rate is (cwnd * mss) / srtt
771 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate.
772 * In Congestion Avoidance phase, set it to 120 % the current rate.
774 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh)
775 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching
776 * end of slow start and should slow down.
778 if (tp->snd_cwnd < tp->snd_ssthresh / 2)
779 rate *= sysctl_tcp_pacing_ss_ratio;
781 rate *= sysctl_tcp_pacing_ca_ratio;
783 rate *= max(tp->snd_cwnd, tp->packets_out);
785 if (likely(tp->srtt_us))
786 do_div(rate, tp->srtt_us);
788 /* ACCESS_ONCE() is needed because sch_fq fetches sk_pacing_rate
789 * without any lock. We want to make sure compiler wont store
790 * intermediate values in this location.
792 ACCESS_ONCE(sk->sk_pacing_rate) = min_t(u64, rate,
793 sk->sk_max_pacing_rate);
796 /* Calculate rto without backoff. This is the second half of Van Jacobson's
797 * routine referred to above.
799 static void tcp_set_rto(struct sock *sk)
801 const struct tcp_sock *tp = tcp_sk(sk);
802 /* Old crap is replaced with new one. 8)
805 * 1. If rtt variance happened to be less 50msec, it is hallucination.
806 * It cannot be less due to utterly erratic ACK generation made
807 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
808 * to do with delayed acks, because at cwnd>2 true delack timeout
809 * is invisible. Actually, Linux-2.4 also generates erratic
810 * ACKs in some circumstances.
812 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
814 /* 2. Fixups made earlier cannot be right.
815 * If we do not estimate RTO correctly without them,
816 * all the algo is pure shit and should be replaced
817 * with correct one. It is exactly, which we pretend to do.
820 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
821 * guarantees that rto is higher.
826 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
828 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
831 cwnd = TCP_INIT_CWND;
832 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
836 * Packet counting of FACK is based on in-order assumptions, therefore TCP
837 * disables it when reordering is detected
839 void tcp_disable_fack(struct tcp_sock *tp)
841 /* RFC3517 uses different metric in lost marker => reset on change */
843 tp->lost_skb_hint = NULL;
844 tp->rx_opt.sack_ok &= ~TCP_FACK_ENABLED;
847 /* Take a notice that peer is sending D-SACKs */
848 static void tcp_dsack_seen(struct tcp_sock *tp)
850 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
853 static void tcp_update_reordering(struct sock *sk, const int metric,
856 struct tcp_sock *tp = tcp_sk(sk);
857 if (metric > tp->reordering) {
860 tp->reordering = min(sysctl_tcp_max_reordering, metric);
862 /* This exciting event is worth to be remembered. 8) */
864 mib_idx = LINUX_MIB_TCPTSREORDER;
865 else if (tcp_is_reno(tp))
866 mib_idx = LINUX_MIB_TCPRENOREORDER;
867 else if (tcp_is_fack(tp))
868 mib_idx = LINUX_MIB_TCPFACKREORDER;
870 mib_idx = LINUX_MIB_TCPSACKREORDER;
872 NET_INC_STATS_BH(sock_net(sk), mib_idx);
873 #if FASTRETRANS_DEBUG > 1
874 pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
875 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
879 tp->undo_marker ? tp->undo_retrans : 0);
881 tcp_disable_fack(tp);
885 tcp_disable_early_retrans(tp);
889 /* This must be called before lost_out is incremented */
890 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
892 if (!tp->retransmit_skb_hint ||
893 before(TCP_SKB_CB(skb)->seq,
894 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
895 tp->retransmit_skb_hint = skb;
898 after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high))
899 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
902 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
904 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
905 tcp_verify_retransmit_hint(tp, skb);
907 tp->lost_out += tcp_skb_pcount(skb);
908 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
912 void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp, struct sk_buff *skb)
914 tcp_verify_retransmit_hint(tp, skb);
916 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
917 tp->lost_out += tcp_skb_pcount(skb);
918 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
922 /* This procedure tags the retransmission queue when SACKs arrive.
924 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
925 * Packets in queue with these bits set are counted in variables
926 * sacked_out, retrans_out and lost_out, correspondingly.
928 * Valid combinations are:
929 * Tag InFlight Description
930 * 0 1 - orig segment is in flight.
931 * S 0 - nothing flies, orig reached receiver.
932 * L 0 - nothing flies, orig lost by net.
933 * R 2 - both orig and retransmit are in flight.
934 * L|R 1 - orig is lost, retransmit is in flight.
935 * S|R 1 - orig reached receiver, retrans is still in flight.
936 * (L|S|R is logically valid, it could occur when L|R is sacked,
937 * but it is equivalent to plain S and code short-curcuits it to S.
938 * L|S is logically invalid, it would mean -1 packet in flight 8))
940 * These 6 states form finite state machine, controlled by the following events:
941 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
942 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
943 * 3. Loss detection event of two flavors:
944 * A. Scoreboard estimator decided the packet is lost.
945 * A'. Reno "three dupacks" marks head of queue lost.
946 * A''. Its FACK modification, head until snd.fack is lost.
947 * B. SACK arrives sacking SND.NXT at the moment, when the
948 * segment was retransmitted.
949 * 4. D-SACK added new rule: D-SACK changes any tag to S.
951 * It is pleasant to note, that state diagram turns out to be commutative,
952 * so that we are allowed not to be bothered by order of our actions,
953 * when multiple events arrive simultaneously. (see the function below).
955 * Reordering detection.
956 * --------------------
957 * Reordering metric is maximal distance, which a packet can be displaced
958 * in packet stream. With SACKs we can estimate it:
960 * 1. SACK fills old hole and the corresponding segment was not
961 * ever retransmitted -> reordering. Alas, we cannot use it
962 * when segment was retransmitted.
963 * 2. The last flaw is solved with D-SACK. D-SACK arrives
964 * for retransmitted and already SACKed segment -> reordering..
965 * Both of these heuristics are not used in Loss state, when we cannot
966 * account for retransmits accurately.
968 * SACK block validation.
969 * ----------------------
971 * SACK block range validation checks that the received SACK block fits to
972 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
973 * Note that SND.UNA is not included to the range though being valid because
974 * it means that the receiver is rather inconsistent with itself reporting
975 * SACK reneging when it should advance SND.UNA. Such SACK block this is
976 * perfectly valid, however, in light of RFC2018 which explicitly states
977 * that "SACK block MUST reflect the newest segment. Even if the newest
978 * segment is going to be discarded ...", not that it looks very clever
979 * in case of head skb. Due to potentional receiver driven attacks, we
980 * choose to avoid immediate execution of a walk in write queue due to
981 * reneging and defer head skb's loss recovery to standard loss recovery
982 * procedure that will eventually trigger (nothing forbids us doing this).
984 * Implements also blockage to start_seq wrap-around. Problem lies in the
985 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
986 * there's no guarantee that it will be before snd_nxt (n). The problem
987 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
990 * <- outs wnd -> <- wrapzone ->
991 * u e n u_w e_w s n_w
993 * |<------------+------+----- TCP seqno space --------------+---------->|
994 * ...-- <2^31 ->| |<--------...
995 * ...---- >2^31 ------>| |<--------...
997 * Current code wouldn't be vulnerable but it's better still to discard such
998 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
999 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1000 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1001 * equal to the ideal case (infinite seqno space without wrap caused issues).
1003 * With D-SACK the lower bound is extended to cover sequence space below
1004 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1005 * again, D-SACK block must not to go across snd_una (for the same reason as
1006 * for the normal SACK blocks, explained above). But there all simplicity
1007 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1008 * fully below undo_marker they do not affect behavior in anyway and can
1009 * therefore be safely ignored. In rare cases (which are more or less
1010 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1011 * fragmentation and packet reordering past skb's retransmission. To consider
1012 * them correctly, the acceptable range must be extended even more though
1013 * the exact amount is rather hard to quantify. However, tp->max_window can
1014 * be used as an exaggerated estimate.
1016 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
1017 u32 start_seq, u32 end_seq)
1019 /* Too far in future, or reversed (interpretation is ambiguous) */
1020 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1023 /* Nasty start_seq wrap-around check (see comments above) */
1024 if (!before(start_seq, tp->snd_nxt))
1027 /* In outstanding window? ...This is valid exit for D-SACKs too.
1028 * start_seq == snd_una is non-sensical (see comments above)
1030 if (after(start_seq, tp->snd_una))
1033 if (!is_dsack || !tp->undo_marker)
1036 /* ...Then it's D-SACK, and must reside below snd_una completely */
1037 if (after(end_seq, tp->snd_una))
1040 if (!before(start_seq, tp->undo_marker))
1044 if (!after(end_seq, tp->undo_marker))
1047 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1048 * start_seq < undo_marker and end_seq >= undo_marker.
1050 return !before(start_seq, end_seq - tp->max_window);
1053 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1054 struct tcp_sack_block_wire *sp, int num_sacks,
1057 struct tcp_sock *tp = tcp_sk(sk);
1058 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1059 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1060 bool dup_sack = false;
1062 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1065 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1066 } else if (num_sacks > 1) {
1067 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1068 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1070 if (!after(end_seq_0, end_seq_1) &&
1071 !before(start_seq_0, start_seq_1)) {
1074 NET_INC_STATS_BH(sock_net(sk),
1075 LINUX_MIB_TCPDSACKOFORECV);
1079 /* D-SACK for already forgotten data... Do dumb counting. */
1080 if (dup_sack && tp->undo_marker && tp->undo_retrans > 0 &&
1081 !after(end_seq_0, prior_snd_una) &&
1082 after(end_seq_0, tp->undo_marker))
1088 struct tcp_sacktag_state {
1091 /* Timestamps for earliest and latest never-retransmitted segment
1092 * that was SACKed. RTO needs the earliest RTT to stay conservative,
1093 * but congestion control should still get an accurate delay signal.
1095 struct skb_mstamp first_sackt;
1096 struct skb_mstamp last_sackt;
1100 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1101 * the incoming SACK may not exactly match but we can find smaller MSS
1102 * aligned portion of it that matches. Therefore we might need to fragment
1103 * which may fail and creates some hassle (caller must handle error case
1106 * FIXME: this could be merged to shift decision code
1108 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1109 u32 start_seq, u32 end_seq)
1113 unsigned int pkt_len;
1116 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1117 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1119 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1120 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1121 mss = tcp_skb_mss(skb);
1122 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1125 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1129 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1134 /* Round if necessary so that SACKs cover only full MSSes
1135 * and/or the remaining small portion (if present)
1137 if (pkt_len > mss) {
1138 unsigned int new_len = (pkt_len / mss) * mss;
1139 if (!in_sack && new_len < pkt_len) {
1141 if (new_len >= skb->len)
1146 err = tcp_fragment(sk, skb, pkt_len, mss, GFP_ATOMIC);
1154 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1155 static u8 tcp_sacktag_one(struct sock *sk,
1156 struct tcp_sacktag_state *state, u8 sacked,
1157 u32 start_seq, u32 end_seq,
1158 int dup_sack, int pcount,
1159 const struct skb_mstamp *xmit_time)
1161 struct tcp_sock *tp = tcp_sk(sk);
1162 int fack_count = state->fack_count;
1164 /* Account D-SACK for retransmitted packet. */
1165 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1166 if (tp->undo_marker && tp->undo_retrans > 0 &&
1167 after(end_seq, tp->undo_marker))
1169 if (sacked & TCPCB_SACKED_ACKED)
1170 state->reord = min(fack_count, state->reord);
1173 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1174 if (!after(end_seq, tp->snd_una))
1177 if (!(sacked & TCPCB_SACKED_ACKED)) {
1178 tcp_rack_advance(tp, xmit_time, sacked);
1180 if (sacked & TCPCB_SACKED_RETRANS) {
1181 /* If the segment is not tagged as lost,
1182 * we do not clear RETRANS, believing
1183 * that retransmission is still in flight.
1185 if (sacked & TCPCB_LOST) {
1186 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1187 tp->lost_out -= pcount;
1188 tp->retrans_out -= pcount;
1191 if (!(sacked & TCPCB_RETRANS)) {
1192 /* New sack for not retransmitted frame,
1193 * which was in hole. It is reordering.
1195 if (before(start_seq,
1196 tcp_highest_sack_seq(tp)))
1197 state->reord = min(fack_count,
1199 if (!after(end_seq, tp->high_seq))
1200 state->flag |= FLAG_ORIG_SACK_ACKED;
1201 if (state->first_sackt.v64 == 0)
1202 state->first_sackt = *xmit_time;
1203 state->last_sackt = *xmit_time;
1206 if (sacked & TCPCB_LOST) {
1207 sacked &= ~TCPCB_LOST;
1208 tp->lost_out -= pcount;
1212 sacked |= TCPCB_SACKED_ACKED;
1213 state->flag |= FLAG_DATA_SACKED;
1214 tp->sacked_out += pcount;
1215 tp->delivered += pcount; /* Out-of-order packets delivered */
1217 fack_count += pcount;
1219 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1220 if (!tcp_is_fack(tp) && tp->lost_skb_hint &&
1221 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1222 tp->lost_cnt_hint += pcount;
1224 if (fack_count > tp->fackets_out)
1225 tp->fackets_out = fack_count;
1228 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1229 * frames and clear it. undo_retrans is decreased above, L|R frames
1230 * are accounted above as well.
1232 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1233 sacked &= ~TCPCB_SACKED_RETRANS;
1234 tp->retrans_out -= pcount;
1240 /* Shift newly-SACKed bytes from this skb to the immediately previous
1241 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1243 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
1244 struct tcp_sacktag_state *state,
1245 unsigned int pcount, int shifted, int mss,
1248 struct tcp_sock *tp = tcp_sk(sk);
1249 struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
1250 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1251 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
1255 /* Adjust counters and hints for the newly sacked sequence
1256 * range but discard the return value since prev is already
1257 * marked. We must tag the range first because the seq
1258 * advancement below implicitly advances
1259 * tcp_highest_sack_seq() when skb is highest_sack.
1261 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1262 start_seq, end_seq, dup_sack, pcount,
1265 if (skb == tp->lost_skb_hint)
1266 tp->lost_cnt_hint += pcount;
1268 TCP_SKB_CB(prev)->end_seq += shifted;
1269 TCP_SKB_CB(skb)->seq += shifted;
1271 tcp_skb_pcount_add(prev, pcount);
1272 BUG_ON(tcp_skb_pcount(skb) < pcount);
1273 tcp_skb_pcount_add(skb, -pcount);
1275 /* When we're adding to gso_segs == 1, gso_size will be zero,
1276 * in theory this shouldn't be necessary but as long as DSACK
1277 * code can come after this skb later on it's better to keep
1278 * setting gso_size to something.
1280 if (!TCP_SKB_CB(prev)->tcp_gso_size)
1281 TCP_SKB_CB(prev)->tcp_gso_size = mss;
1283 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1284 if (tcp_skb_pcount(skb) <= 1)
1285 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1287 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1288 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1291 BUG_ON(!tcp_skb_pcount(skb));
1292 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1296 /* Whole SKB was eaten :-) */
1298 if (skb == tp->retransmit_skb_hint)
1299 tp->retransmit_skb_hint = prev;
1300 if (skb == tp->lost_skb_hint) {
1301 tp->lost_skb_hint = prev;
1302 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1305 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1306 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1307 TCP_SKB_CB(prev)->end_seq++;
1309 if (skb == tcp_highest_sack(sk))
1310 tcp_advance_highest_sack(sk, skb);
1312 tcp_skb_collapse_tstamp(prev, 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);
1828 u32 prior_sacked = tp->sacked_out;
1831 tcp_check_reno_reordering(sk, 0);
1832 if (tp->sacked_out > prior_sacked)
1833 tp->delivered++; /* Some out-of-order packet is delivered */
1834 tcp_verify_left_out(tp);
1837 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1839 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1841 struct tcp_sock *tp = tcp_sk(sk);
1844 /* One ACK acked hole. The rest eat duplicate ACKs. */
1845 tp->delivered += max_t(int, acked - tp->sacked_out, 1);
1846 if (acked - 1 >= tp->sacked_out)
1849 tp->sacked_out -= acked - 1;
1851 tcp_check_reno_reordering(sk, acked);
1852 tcp_verify_left_out(tp);
1855 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1860 void tcp_clear_retrans(struct tcp_sock *tp)
1862 tp->retrans_out = 0;
1864 tp->undo_marker = 0;
1865 tp->undo_retrans = -1;
1866 tp->fackets_out = 0;
1870 static inline void tcp_init_undo(struct tcp_sock *tp)
1872 tp->undo_marker = tp->snd_una;
1873 /* Retransmission still in flight may cause DSACKs later. */
1874 tp->undo_retrans = tp->retrans_out ? : -1;
1877 /* Enter Loss state. If we detect SACK reneging, forget all SACK information
1878 * and reset tags completely, otherwise preserve SACKs. If receiver
1879 * dropped its ofo queue, we will know this due to reneging detection.
1881 void tcp_enter_loss(struct sock *sk)
1883 const struct inet_connection_sock *icsk = inet_csk(sk);
1884 struct tcp_sock *tp = tcp_sk(sk);
1885 struct net *net = sock_net(sk);
1886 struct sk_buff *skb;
1887 bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery;
1888 bool is_reneg; /* is receiver reneging on SACKs? */
1890 /* Reduce ssthresh if it has not yet been made inside this window. */
1891 if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
1892 !after(tp->high_seq, tp->snd_una) ||
1893 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1894 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1895 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1896 tcp_ca_event(sk, CA_EVENT_LOSS);
1900 tp->snd_cwnd_cnt = 0;
1901 tp->snd_cwnd_stamp = tcp_time_stamp;
1903 tp->retrans_out = 0;
1906 if (tcp_is_reno(tp))
1907 tcp_reset_reno_sack(tp);
1909 skb = tcp_write_queue_head(sk);
1910 is_reneg = skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED);
1912 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
1914 tp->fackets_out = 0;
1916 tcp_clear_all_retrans_hints(tp);
1918 tcp_for_write_queue(skb, sk) {
1919 if (skb == tcp_send_head(sk))
1922 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1923 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || is_reneg) {
1924 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1925 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1926 tp->lost_out += tcp_skb_pcount(skb);
1927 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
1930 tcp_verify_left_out(tp);
1932 /* Timeout in disordered state after receiving substantial DUPACKs
1933 * suggests that the degree of reordering is over-estimated.
1935 if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
1936 tp->sacked_out >= net->ipv4.sysctl_tcp_reordering)
1937 tp->reordering = min_t(unsigned int, tp->reordering,
1938 net->ipv4.sysctl_tcp_reordering);
1939 tcp_set_ca_state(sk, TCP_CA_Loss);
1940 tp->high_seq = tp->snd_nxt;
1941 tcp_ecn_queue_cwr(tp);
1943 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
1944 * loss recovery is underway except recurring timeout(s) on
1945 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
1947 tp->frto = sysctl_tcp_frto &&
1948 (new_recovery || icsk->icsk_retransmits) &&
1949 !inet_csk(sk)->icsk_mtup.probe_size;
1952 /* If ACK arrived pointing to a remembered SACK, it means that our
1953 * remembered SACKs do not reflect real state of receiver i.e.
1954 * receiver _host_ is heavily congested (or buggy).
1956 * To avoid big spurious retransmission bursts due to transient SACK
1957 * scoreboard oddities that look like reneging, we give the receiver a
1958 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
1959 * restore sanity to the SACK scoreboard. If the apparent reneging
1960 * persists until this RTO then we'll clear the SACK scoreboard.
1962 static bool tcp_check_sack_reneging(struct sock *sk, int flag)
1964 if (flag & FLAG_SACK_RENEGING) {
1965 struct tcp_sock *tp = tcp_sk(sk);
1966 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4),
1967 msecs_to_jiffies(10));
1969 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1970 delay, TCP_RTO_MAX);
1976 static inline int tcp_fackets_out(const struct tcp_sock *tp)
1978 return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
1981 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
1982 * counter when SACK is enabled (without SACK, sacked_out is used for
1985 * Instead, with FACK TCP uses fackets_out that includes both SACKed
1986 * segments up to the highest received SACK block so far and holes in
1989 * With reordering, holes may still be in flight, so RFC3517 recovery
1990 * uses pure sacked_out (total number of SACKed segments) even though
1991 * it violates the RFC that uses duplicate ACKs, often these are equal
1992 * but when e.g. out-of-window ACKs or packet duplication occurs,
1993 * they differ. Since neither occurs due to loss, TCP should really
1996 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
1998 return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
2001 static bool tcp_pause_early_retransmit(struct sock *sk, int flag)
2003 struct tcp_sock *tp = tcp_sk(sk);
2004 unsigned long delay;
2006 /* Delay early retransmit and entering fast recovery for
2007 * max(RTT/4, 2msec) unless ack has ECE mark, no RTT samples
2008 * available, or RTO is scheduled to fire first.
2010 if (sysctl_tcp_early_retrans < 2 || sysctl_tcp_early_retrans > 3 ||
2011 (flag & FLAG_ECE) || !tp->srtt_us)
2014 delay = max(usecs_to_jiffies(tp->srtt_us >> 5),
2015 msecs_to_jiffies(2));
2017 if (!time_after(inet_csk(sk)->icsk_timeout, (jiffies + delay)))
2020 inet_csk_reset_xmit_timer(sk, ICSK_TIME_EARLY_RETRANS, delay,
2025 /* Linux NewReno/SACK/FACK/ECN state machine.
2026 * --------------------------------------
2028 * "Open" Normal state, no dubious events, fast path.
2029 * "Disorder" In all the respects it is "Open",
2030 * but requires a bit more attention. It is entered when
2031 * we see some SACKs or dupacks. It is split of "Open"
2032 * mainly to move some processing from fast path to slow one.
2033 * "CWR" CWND was reduced due to some Congestion Notification event.
2034 * It can be ECN, ICMP source quench, local device congestion.
2035 * "Recovery" CWND was reduced, we are fast-retransmitting.
2036 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2038 * tcp_fastretrans_alert() is entered:
2039 * - each incoming ACK, if state is not "Open"
2040 * - when arrived ACK is unusual, namely:
2045 * Counting packets in flight is pretty simple.
2047 * in_flight = packets_out - left_out + retrans_out
2049 * packets_out is SND.NXT-SND.UNA counted in packets.
2051 * retrans_out is number of retransmitted segments.
2053 * left_out is number of segments left network, but not ACKed yet.
2055 * left_out = sacked_out + lost_out
2057 * sacked_out: Packets, which arrived to receiver out of order
2058 * and hence not ACKed. With SACKs this number is simply
2059 * amount of SACKed data. Even without SACKs
2060 * it is easy to give pretty reliable estimate of this number,
2061 * counting duplicate ACKs.
2063 * lost_out: Packets lost by network. TCP has no explicit
2064 * "loss notification" feedback from network (for now).
2065 * It means that this number can be only _guessed_.
2066 * Actually, it is the heuristics to predict lossage that
2067 * distinguishes different algorithms.
2069 * F.e. after RTO, when all the queue is considered as lost,
2070 * lost_out = packets_out and in_flight = retrans_out.
2072 * Essentially, we have now two algorithms counting
2075 * FACK: It is the simplest heuristics. As soon as we decided
2076 * that something is lost, we decide that _all_ not SACKed
2077 * packets until the most forward SACK are lost. I.e.
2078 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2079 * It is absolutely correct estimate, if network does not reorder
2080 * packets. And it loses any connection to reality when reordering
2081 * takes place. We use FACK by default until reordering
2082 * is suspected on the path to this destination.
2084 * NewReno: when Recovery is entered, we assume that one segment
2085 * is lost (classic Reno). While we are in Recovery and
2086 * a partial ACK arrives, we assume that one more packet
2087 * is lost (NewReno). This heuristics are the same in NewReno
2090 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2091 * deflation etc. CWND is real congestion window, never inflated, changes
2092 * only according to classic VJ rules.
2094 * Really tricky (and requiring careful tuning) part of algorithm
2095 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2096 * The first determines the moment _when_ we should reduce CWND and,
2097 * hence, slow down forward transmission. In fact, it determines the moment
2098 * when we decide that hole is caused by loss, rather than by a reorder.
2100 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2101 * holes, caused by lost packets.
2103 * And the most logically complicated part of algorithm is undo
2104 * heuristics. We detect false retransmits due to both too early
2105 * fast retransmit (reordering) and underestimated RTO, analyzing
2106 * timestamps and D-SACKs. When we detect that some segments were
2107 * retransmitted by mistake and CWND reduction was wrong, we undo
2108 * window reduction and abort recovery phase. This logic is hidden
2109 * inside several functions named tcp_try_undo_<something>.
2112 /* This function decides, when we should leave Disordered state
2113 * and enter Recovery phase, reducing congestion window.
2115 * Main question: may we further continue forward transmission
2116 * with the same cwnd?
2118 static bool tcp_time_to_recover(struct sock *sk, int flag)
2120 struct tcp_sock *tp = tcp_sk(sk);
2122 int tcp_reordering = sock_net(sk)->ipv4.sysctl_tcp_reordering;
2124 /* Trick#1: The loss is proven. */
2128 /* Not-A-Trick#2 : Classic rule... */
2129 if (tcp_dupack_heuristics(tp) > tp->reordering)
2132 /* Trick#4: It is still not OK... But will it be useful to delay
2135 packets_out = tp->packets_out;
2136 if (packets_out <= tp->reordering &&
2137 tp->sacked_out >= max_t(__u32, packets_out/2, tcp_reordering) &&
2138 !tcp_may_send_now(sk)) {
2139 /* We have nothing to send. This connection is limited
2140 * either by receiver window or by application.
2145 /* If a thin stream is detected, retransmit after first
2146 * received dupack. Employ only if SACK is supported in order
2147 * to avoid possible corner-case series of spurious retransmissions
2148 * Use only if there are no unsent data.
2150 if ((tp->thin_dupack || sysctl_tcp_thin_dupack) &&
2151 tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 &&
2152 tcp_is_sack(tp) && !tcp_send_head(sk))
2155 /* Trick#6: TCP early retransmit, per RFC5827. To avoid spurious
2156 * retransmissions due to small network reorderings, we implement
2157 * Mitigation A.3 in the RFC and delay the retransmission for a short
2158 * interval if appropriate.
2160 if (tp->do_early_retrans && !tp->retrans_out && tp->sacked_out &&
2161 (tp->packets_out >= (tp->sacked_out + 1) && tp->packets_out < 4) &&
2162 !tcp_may_send_now(sk))
2163 return !tcp_pause_early_retransmit(sk, flag);
2168 /* Detect loss in event "A" above by marking head of queue up as lost.
2169 * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2170 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2171 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2172 * the maximum SACKed segments to pass before reaching this limit.
2174 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2176 struct tcp_sock *tp = tcp_sk(sk);
2177 struct sk_buff *skb;
2178 int cnt, oldcnt, lost;
2180 /* Use SACK to deduce losses of new sequences sent during recovery */
2181 const u32 loss_high = tcp_is_sack(tp) ? tp->snd_nxt : tp->high_seq;
2183 WARN_ON(packets > tp->packets_out);
2184 if (tp->lost_skb_hint) {
2185 skb = tp->lost_skb_hint;
2186 cnt = tp->lost_cnt_hint;
2187 /* Head already handled? */
2188 if (mark_head && skb != tcp_write_queue_head(sk))
2191 skb = tcp_write_queue_head(sk);
2195 tcp_for_write_queue_from(skb, sk) {
2196 if (skb == tcp_send_head(sk))
2198 /* TODO: do this better */
2199 /* this is not the most efficient way to do this... */
2200 tp->lost_skb_hint = skb;
2201 tp->lost_cnt_hint = cnt;
2203 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2207 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2208 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2209 cnt += tcp_skb_pcount(skb);
2211 if (cnt > packets) {
2212 if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) ||
2213 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
2214 (oldcnt >= packets))
2217 mss = tcp_skb_mss(skb);
2218 /* If needed, chop off the prefix to mark as lost. */
2219 lost = (packets - oldcnt) * mss;
2220 if (lost < skb->len &&
2221 tcp_fragment(sk, skb, lost, mss, GFP_ATOMIC) < 0)
2226 tcp_skb_mark_lost(tp, skb);
2231 tcp_verify_left_out(tp);
2234 /* Account newly detected lost packet(s) */
2236 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2238 struct tcp_sock *tp = tcp_sk(sk);
2240 if (tcp_is_reno(tp)) {
2241 tcp_mark_head_lost(sk, 1, 1);
2242 } else if (tcp_is_fack(tp)) {
2243 int lost = tp->fackets_out - tp->reordering;
2246 tcp_mark_head_lost(sk, lost, 0);
2248 int sacked_upto = tp->sacked_out - tp->reordering;
2249 if (sacked_upto >= 0)
2250 tcp_mark_head_lost(sk, sacked_upto, 0);
2251 else if (fast_rexmit)
2252 tcp_mark_head_lost(sk, 1, 1);
2256 static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when)
2258 return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2259 before(tp->rx_opt.rcv_tsecr, when);
2262 /* skb is spurious retransmitted if the returned timestamp echo
2263 * reply is prior to the skb transmission time
2265 static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp,
2266 const struct sk_buff *skb)
2268 return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) &&
2269 tcp_tsopt_ecr_before(tp, tcp_skb_timestamp(skb));
2272 /* Nothing was retransmitted or returned timestamp is less
2273 * than timestamp of the first retransmission.
2275 static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2277 return !tp->retrans_stamp ||
2278 tcp_tsopt_ecr_before(tp, tp->retrans_stamp);
2281 /* Undo procedures. */
2283 /* We can clear retrans_stamp when there are no retransmissions in the
2284 * window. It would seem that it is trivially available for us in
2285 * tp->retrans_out, however, that kind of assumptions doesn't consider
2286 * what will happen if errors occur when sending retransmission for the
2287 * second time. ...It could the that such segment has only
2288 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2289 * the head skb is enough except for some reneging corner cases that
2290 * are not worth the effort.
2292 * Main reason for all this complexity is the fact that connection dying
2293 * time now depends on the validity of the retrans_stamp, in particular,
2294 * that successive retransmissions of a segment must not advance
2295 * retrans_stamp under any conditions.
2297 static bool tcp_any_retrans_done(const struct sock *sk)
2299 const struct tcp_sock *tp = tcp_sk(sk);
2300 struct sk_buff *skb;
2302 if (tp->retrans_out)
2305 skb = tcp_write_queue_head(sk);
2306 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2312 #if FASTRETRANS_DEBUG > 1
2313 static void DBGUNDO(struct sock *sk, const char *msg)
2315 struct tcp_sock *tp = tcp_sk(sk);
2316 struct inet_sock *inet = inet_sk(sk);
2318 if (sk->sk_family == AF_INET) {
2319 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2321 &inet->inet_daddr, ntohs(inet->inet_dport),
2322 tp->snd_cwnd, tcp_left_out(tp),
2323 tp->snd_ssthresh, tp->prior_ssthresh,
2326 #if IS_ENABLED(CONFIG_IPV6)
2327 else if (sk->sk_family == AF_INET6) {
2328 struct ipv6_pinfo *np = inet6_sk(sk);
2329 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2331 &np->daddr, ntohs(inet->inet_dport),
2332 tp->snd_cwnd, tcp_left_out(tp),
2333 tp->snd_ssthresh, tp->prior_ssthresh,
2339 #define DBGUNDO(x...) do { } while (0)
2342 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
2344 struct tcp_sock *tp = tcp_sk(sk);
2347 struct sk_buff *skb;
2349 tcp_for_write_queue(skb, sk) {
2350 if (skb == tcp_send_head(sk))
2352 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2355 tcp_clear_all_retrans_hints(tp);
2358 if (tp->prior_ssthresh) {
2359 const struct inet_connection_sock *icsk = inet_csk(sk);
2361 if (icsk->icsk_ca_ops->undo_cwnd)
2362 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2364 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
2366 if (tp->prior_ssthresh > tp->snd_ssthresh) {
2367 tp->snd_ssthresh = tp->prior_ssthresh;
2368 tcp_ecn_withdraw_cwr(tp);
2371 tp->snd_cwnd_stamp = tcp_time_stamp;
2372 tp->undo_marker = 0;
2375 static inline bool tcp_may_undo(const struct tcp_sock *tp)
2377 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2380 /* People celebrate: "We love our President!" */
2381 static bool tcp_try_undo_recovery(struct sock *sk)
2383 struct tcp_sock *tp = tcp_sk(sk);
2385 if (tcp_may_undo(tp)) {
2388 /* Happy end! We did not retransmit anything
2389 * or our original transmission succeeded.
2391 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2392 tcp_undo_cwnd_reduction(sk, false);
2393 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2394 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2396 mib_idx = LINUX_MIB_TCPFULLUNDO;
2398 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2400 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2401 /* Hold old state until something *above* high_seq
2402 * is ACKed. For Reno it is MUST to prevent false
2403 * fast retransmits (RFC2582). SACK TCP is safe. */
2404 if (!tcp_any_retrans_done(sk))
2405 tp->retrans_stamp = 0;
2408 tcp_set_ca_state(sk, TCP_CA_Open);
2412 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2413 static bool tcp_try_undo_dsack(struct sock *sk)
2415 struct tcp_sock *tp = tcp_sk(sk);
2417 if (tp->undo_marker && !tp->undo_retrans) {
2418 DBGUNDO(sk, "D-SACK");
2419 tcp_undo_cwnd_reduction(sk, false);
2420 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2426 /* Undo during loss recovery after partial ACK or using F-RTO. */
2427 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2429 struct tcp_sock *tp = tcp_sk(sk);
2431 if (frto_undo || tcp_may_undo(tp)) {
2432 tcp_undo_cwnd_reduction(sk, true);
2434 DBGUNDO(sk, "partial loss");
2435 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2437 NET_INC_STATS_BH(sock_net(sk),
2438 LINUX_MIB_TCPSPURIOUSRTOS);
2439 inet_csk(sk)->icsk_retransmits = 0;
2440 if (frto_undo || tcp_is_sack(tp))
2441 tcp_set_ca_state(sk, TCP_CA_Open);
2447 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
2448 * It computes the number of packets to send (sndcnt) based on packets newly
2450 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2451 * cwnd reductions across a full RTT.
2452 * 2) Otherwise PRR uses packet conservation to send as much as delivered.
2453 * But when the retransmits are acked without further losses, PRR
2454 * slow starts cwnd up to ssthresh to speed up the recovery.
2456 static void tcp_init_cwnd_reduction(struct sock *sk)
2458 struct tcp_sock *tp = tcp_sk(sk);
2460 tp->high_seq = tp->snd_nxt;
2461 tp->tlp_high_seq = 0;
2462 tp->snd_cwnd_cnt = 0;
2463 tp->prior_cwnd = tp->snd_cwnd;
2464 tp->prr_delivered = 0;
2466 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2467 tcp_ecn_queue_cwr(tp);
2470 static void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked,
2473 struct tcp_sock *tp = tcp_sk(sk);
2475 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2477 if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd))
2480 tp->prr_delivered += newly_acked_sacked;
2482 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2484 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2485 } else if ((flag & FLAG_RETRANS_DATA_ACKED) &&
2486 !(flag & FLAG_LOST_RETRANS)) {
2487 sndcnt = min_t(int, delta,
2488 max_t(int, tp->prr_delivered - tp->prr_out,
2489 newly_acked_sacked) + 1);
2491 sndcnt = min(delta, newly_acked_sacked);
2493 /* Force a fast retransmit upon entering fast recovery */
2494 sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1));
2495 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2498 static inline void tcp_end_cwnd_reduction(struct sock *sk)
2500 struct tcp_sock *tp = tcp_sk(sk);
2502 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2503 if (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR ||
2504 (tp->undo_marker && tp->snd_ssthresh < TCP_INFINITE_SSTHRESH)) {
2505 tp->snd_cwnd = tp->snd_ssthresh;
2506 tp->snd_cwnd_stamp = tcp_time_stamp;
2508 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2511 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2512 void tcp_enter_cwr(struct sock *sk)
2514 struct tcp_sock *tp = tcp_sk(sk);
2516 tp->prior_ssthresh = 0;
2517 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2518 tp->undo_marker = 0;
2519 tcp_init_cwnd_reduction(sk);
2520 tcp_set_ca_state(sk, TCP_CA_CWR);
2523 EXPORT_SYMBOL(tcp_enter_cwr);
2525 static void tcp_try_keep_open(struct sock *sk)
2527 struct tcp_sock *tp = tcp_sk(sk);
2528 int state = TCP_CA_Open;
2530 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2531 state = TCP_CA_Disorder;
2533 if (inet_csk(sk)->icsk_ca_state != state) {
2534 tcp_set_ca_state(sk, state);
2535 tp->high_seq = tp->snd_nxt;
2539 static void tcp_try_to_open(struct sock *sk, int flag)
2541 struct tcp_sock *tp = tcp_sk(sk);
2543 tcp_verify_left_out(tp);
2545 if (!tcp_any_retrans_done(sk))
2546 tp->retrans_stamp = 0;
2548 if (flag & FLAG_ECE)
2551 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2552 tcp_try_keep_open(sk);
2556 static void tcp_mtup_probe_failed(struct sock *sk)
2558 struct inet_connection_sock *icsk = inet_csk(sk);
2560 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2561 icsk->icsk_mtup.probe_size = 0;
2562 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPMTUPFAIL);
2565 static void tcp_mtup_probe_success(struct sock *sk)
2567 struct tcp_sock *tp = tcp_sk(sk);
2568 struct inet_connection_sock *icsk = inet_csk(sk);
2570 /* FIXME: breaks with very large cwnd */
2571 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2572 tp->snd_cwnd = tp->snd_cwnd *
2573 tcp_mss_to_mtu(sk, tp->mss_cache) /
2574 icsk->icsk_mtup.probe_size;
2575 tp->snd_cwnd_cnt = 0;
2576 tp->snd_cwnd_stamp = tcp_time_stamp;
2577 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2579 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2580 icsk->icsk_mtup.probe_size = 0;
2581 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2582 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS);
2585 /* Do a simple retransmit without using the backoff mechanisms in
2586 * tcp_timer. This is used for path mtu discovery.
2587 * The socket is already locked here.
2589 void tcp_simple_retransmit(struct sock *sk)
2591 const struct inet_connection_sock *icsk = inet_csk(sk);
2592 struct tcp_sock *tp = tcp_sk(sk);
2593 struct sk_buff *skb;
2594 unsigned int mss = tcp_current_mss(sk);
2595 u32 prior_lost = tp->lost_out;
2597 tcp_for_write_queue(skb, sk) {
2598 if (skb == tcp_send_head(sk))
2600 if (tcp_skb_seglen(skb) > mss &&
2601 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2602 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2603 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2604 tp->retrans_out -= tcp_skb_pcount(skb);
2606 tcp_skb_mark_lost_uncond_verify(tp, skb);
2610 tcp_clear_retrans_hints_partial(tp);
2612 if (prior_lost == tp->lost_out)
2615 if (tcp_is_reno(tp))
2616 tcp_limit_reno_sacked(tp);
2618 tcp_verify_left_out(tp);
2620 /* Don't muck with the congestion window here.
2621 * Reason is that we do not increase amount of _data_
2622 * in network, but units changed and effective
2623 * cwnd/ssthresh really reduced now.
2625 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2626 tp->high_seq = tp->snd_nxt;
2627 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2628 tp->prior_ssthresh = 0;
2629 tp->undo_marker = 0;
2630 tcp_set_ca_state(sk, TCP_CA_Loss);
2632 tcp_xmit_retransmit_queue(sk);
2634 EXPORT_SYMBOL(tcp_simple_retransmit);
2636 static void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2638 struct tcp_sock *tp = tcp_sk(sk);
2641 if (tcp_is_reno(tp))
2642 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2644 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2646 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2648 tp->prior_ssthresh = 0;
2651 if (!tcp_in_cwnd_reduction(sk)) {
2653 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2654 tcp_init_cwnd_reduction(sk);
2656 tcp_set_ca_state(sk, TCP_CA_Recovery);
2659 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2660 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2662 static void tcp_process_loss(struct sock *sk, int flag, bool is_dupack,
2665 struct tcp_sock *tp = tcp_sk(sk);
2666 bool recovered = !before(tp->snd_una, tp->high_seq);
2668 if ((flag & FLAG_SND_UNA_ADVANCED) &&
2669 tcp_try_undo_loss(sk, false))
2672 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2673 /* Step 3.b. A timeout is spurious if not all data are
2674 * lost, i.e., never-retransmitted data are (s)acked.
2676 if ((flag & FLAG_ORIG_SACK_ACKED) &&
2677 tcp_try_undo_loss(sk, true))
2680 if (after(tp->snd_nxt, tp->high_seq)) {
2681 if (flag & FLAG_DATA_SACKED || is_dupack)
2682 tp->frto = 0; /* Step 3.a. loss was real */
2683 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2684 tp->high_seq = tp->snd_nxt;
2685 /* Step 2.b. Try send new data (but deferred until cwnd
2686 * is updated in tcp_ack()). Otherwise fall back to
2687 * the conventional recovery.
2689 if (tcp_send_head(sk) &&
2690 after(tcp_wnd_end(tp), tp->snd_nxt)) {
2691 *rexmit = REXMIT_NEW;
2699 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2700 tcp_try_undo_recovery(sk);
2703 if (tcp_is_reno(tp)) {
2704 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2705 * delivered. Lower inflight to clock out (re)tranmissions.
2707 if (after(tp->snd_nxt, tp->high_seq) && is_dupack)
2708 tcp_add_reno_sack(sk);
2709 else if (flag & FLAG_SND_UNA_ADVANCED)
2710 tcp_reset_reno_sack(tp);
2712 *rexmit = REXMIT_LOST;
2715 /* Undo during fast recovery after partial ACK. */
2716 static bool tcp_try_undo_partial(struct sock *sk, const int acked)
2718 struct tcp_sock *tp = tcp_sk(sk);
2720 if (tp->undo_marker && tcp_packet_delayed(tp)) {
2721 /* Plain luck! Hole if filled with delayed
2722 * packet, rather than with a retransmit.
2724 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2726 /* We are getting evidence that the reordering degree is higher
2727 * than we realized. If there are no retransmits out then we
2728 * can undo. Otherwise we clock out new packets but do not
2729 * mark more packets lost or retransmit more.
2731 if (tp->retrans_out)
2734 if (!tcp_any_retrans_done(sk))
2735 tp->retrans_stamp = 0;
2737 DBGUNDO(sk, "partial recovery");
2738 tcp_undo_cwnd_reduction(sk, true);
2739 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2740 tcp_try_keep_open(sk);
2746 /* Process an event, which can update packets-in-flight not trivially.
2747 * Main goal of this function is to calculate new estimate for left_out,
2748 * taking into account both packets sitting in receiver's buffer and
2749 * packets lost by network.
2751 * Besides that it updates the congestion state when packet loss or ECN
2752 * is detected. But it does not reduce the cwnd, it is done by the
2753 * congestion control later.
2755 * It does _not_ decide what to send, it is made in function
2756 * tcp_xmit_retransmit_queue().
2758 static void tcp_fastretrans_alert(struct sock *sk, const int acked,
2759 bool is_dupack, int *ack_flag, int *rexmit)
2761 struct inet_connection_sock *icsk = inet_csk(sk);
2762 struct tcp_sock *tp = tcp_sk(sk);
2763 int fast_rexmit = 0, flag = *ack_flag;
2764 bool do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2765 (tcp_fackets_out(tp) > tp->reordering));
2767 if (WARN_ON(!tp->packets_out && tp->sacked_out))
2769 if (WARN_ON(!tp->sacked_out && tp->fackets_out))
2770 tp->fackets_out = 0;
2772 /* Now state machine starts.
2773 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2774 if (flag & FLAG_ECE)
2775 tp->prior_ssthresh = 0;
2777 /* B. In all the states check for reneging SACKs. */
2778 if (tcp_check_sack_reneging(sk, flag))
2781 /* C. Check consistency of the current state. */
2782 tcp_verify_left_out(tp);
2784 /* D. Check state exit conditions. State can be terminated
2785 * when high_seq is ACKed. */
2786 if (icsk->icsk_ca_state == TCP_CA_Open) {
2787 WARN_ON(tp->retrans_out != 0);
2788 tp->retrans_stamp = 0;
2789 } else if (!before(tp->snd_una, tp->high_seq)) {
2790 switch (icsk->icsk_ca_state) {
2792 /* CWR is to be held something *above* high_seq
2793 * is ACKed for CWR bit to reach receiver. */
2794 if (tp->snd_una != tp->high_seq) {
2795 tcp_end_cwnd_reduction(sk);
2796 tcp_set_ca_state(sk, TCP_CA_Open);
2800 case TCP_CA_Recovery:
2801 if (tcp_is_reno(tp))
2802 tcp_reset_reno_sack(tp);
2803 if (tcp_try_undo_recovery(sk))
2805 tcp_end_cwnd_reduction(sk);
2810 /* Use RACK to detect loss */
2811 if (sysctl_tcp_recovery & TCP_RACK_LOST_RETRANS &&
2812 tcp_rack_mark_lost(sk)) {
2813 flag |= FLAG_LOST_RETRANS;
2814 *ack_flag |= FLAG_LOST_RETRANS;
2817 /* E. Process state. */
2818 switch (icsk->icsk_ca_state) {
2819 case TCP_CA_Recovery:
2820 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2821 if (tcp_is_reno(tp) && is_dupack)
2822 tcp_add_reno_sack(sk);
2824 if (tcp_try_undo_partial(sk, acked))
2826 /* Partial ACK arrived. Force fast retransmit. */
2827 do_lost = tcp_is_reno(tp) ||
2828 tcp_fackets_out(tp) > tp->reordering;
2830 if (tcp_try_undo_dsack(sk)) {
2831 tcp_try_keep_open(sk);
2836 tcp_process_loss(sk, flag, is_dupack, rexmit);
2837 if (icsk->icsk_ca_state != TCP_CA_Open &&
2838 !(flag & FLAG_LOST_RETRANS))
2840 /* Change state if cwnd is undone or retransmits are lost */
2842 if (tcp_is_reno(tp)) {
2843 if (flag & FLAG_SND_UNA_ADVANCED)
2844 tcp_reset_reno_sack(tp);
2846 tcp_add_reno_sack(sk);
2849 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
2850 tcp_try_undo_dsack(sk);
2852 if (!tcp_time_to_recover(sk, flag)) {
2853 tcp_try_to_open(sk, flag);
2857 /* MTU probe failure: don't reduce cwnd */
2858 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2859 icsk->icsk_mtup.probe_size &&
2860 tp->snd_una == tp->mtu_probe.probe_seq_start) {
2861 tcp_mtup_probe_failed(sk);
2862 /* Restores the reduction we did in tcp_mtup_probe() */
2864 tcp_simple_retransmit(sk);
2868 /* Otherwise enter Recovery state */
2869 tcp_enter_recovery(sk, (flag & FLAG_ECE));
2874 tcp_update_scoreboard(sk, fast_rexmit);
2875 *rexmit = REXMIT_LOST;
2878 /* Kathleen Nichols' algorithm for tracking the minimum value of
2879 * a data stream over some fixed time interval. (E.g., the minimum
2880 * RTT over the past five minutes.) It uses constant space and constant
2881 * time per update yet almost always delivers the same minimum as an
2882 * implementation that has to keep all the data in the window.
2884 * The algorithm keeps track of the best, 2nd best & 3rd best min
2885 * values, maintaining an invariant that the measurement time of the
2886 * n'th best >= n-1'th best. It also makes sure that the three values
2887 * are widely separated in the time window since that bounds the worse
2888 * case error when that data is monotonically increasing over the window.
2890 * Upon getting a new min, we can forget everything earlier because it
2891 * has no value - the new min is <= everything else in the window by
2892 * definition and it's the most recent. So we restart fresh on every new min
2893 * and overwrites 2nd & 3rd choices. The same property holds for 2nd & 3rd
2896 static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us)
2898 const u32 now = tcp_time_stamp, wlen = sysctl_tcp_min_rtt_wlen * HZ;
2899 struct rtt_meas *m = tcp_sk(sk)->rtt_min;
2900 struct rtt_meas rttm = {
2901 .rtt = likely(rtt_us) ? rtt_us : jiffies_to_usecs(1),
2906 /* Check if the new measurement updates the 1st, 2nd, or 3rd choices */
2907 if (unlikely(rttm.rtt <= m[0].rtt))
2908 m[0] = m[1] = m[2] = rttm;
2909 else if (rttm.rtt <= m[1].rtt)
2911 else if (rttm.rtt <= m[2].rtt)
2914 elapsed = now - m[0].ts;
2915 if (unlikely(elapsed > wlen)) {
2916 /* Passed entire window without a new min so make 2nd choice
2917 * the new min & 3rd choice the new 2nd. So forth and so on.
2922 if (now - m[0].ts > wlen) {
2925 if (now - m[0].ts > wlen)
2928 } else if (m[1].ts == m[0].ts && elapsed > wlen / 4) {
2929 /* Passed a quarter of the window without a new min so
2930 * take 2nd choice from the 2nd quarter of the window.
2933 } else if (m[2].ts == m[1].ts && elapsed > wlen / 2) {
2934 /* Passed half the window without a new min so take the 3rd
2935 * choice from the last half of the window.
2941 static inline bool tcp_ack_update_rtt(struct sock *sk, const int flag,
2942 long seq_rtt_us, long sack_rtt_us,
2945 const struct tcp_sock *tp = tcp_sk(sk);
2947 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2948 * broken middle-boxes or peers may corrupt TS-ECR fields. But
2949 * Karn's algorithm forbids taking RTT if some retransmitted data
2950 * is acked (RFC6298).
2953 seq_rtt_us = sack_rtt_us;
2955 /* RTTM Rule: A TSecr value received in a segment is used to
2956 * update the averaged RTT measurement only if the segment
2957 * acknowledges some new data, i.e., only if it advances the
2958 * left edge of the send window.
2959 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2961 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2963 seq_rtt_us = ca_rtt_us = jiffies_to_usecs(tcp_time_stamp -
2964 tp->rx_opt.rcv_tsecr);
2968 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
2969 * always taken together with ACK, SACK, or TS-opts. Any negative
2970 * values will be skipped with the seq_rtt_us < 0 check above.
2972 tcp_update_rtt_min(sk, ca_rtt_us);
2973 tcp_rtt_estimator(sk, seq_rtt_us);
2976 /* RFC6298: only reset backoff on valid RTT measurement. */
2977 inet_csk(sk)->icsk_backoff = 0;
2981 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
2982 void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req)
2986 if (req && !req->num_retrans && tcp_rsk(req)->snt_synack.v64) {
2987 struct skb_mstamp now;
2989 skb_mstamp_get(&now);
2990 rtt_us = skb_mstamp_us_delta(&now, &tcp_rsk(req)->snt_synack);
2993 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us);
2997 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
2999 const struct inet_connection_sock *icsk = inet_csk(sk);
3001 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
3002 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
3005 /* Restart timer after forward progress on connection.
3006 * RFC2988 recommends to restart timer to now+rto.
3008 void tcp_rearm_rto(struct sock *sk)
3010 const struct inet_connection_sock *icsk = inet_csk(sk);
3011 struct tcp_sock *tp = tcp_sk(sk);
3013 /* If the retrans timer is currently being used by Fast Open
3014 * for SYN-ACK retrans purpose, stay put.
3016 if (tp->fastopen_rsk)
3019 if (!tp->packets_out) {
3020 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
3022 u32 rto = inet_csk(sk)->icsk_rto;
3023 /* Offset the time elapsed after installing regular RTO */
3024 if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
3025 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
3026 struct sk_buff *skb = tcp_write_queue_head(sk);
3027 const u32 rto_time_stamp =
3028 tcp_skb_timestamp(skb) + rto;
3029 s32 delta = (s32)(rto_time_stamp - tcp_time_stamp);
3030 /* delta may not be positive if the socket is locked
3031 * when the retrans timer fires and is rescheduled.
3036 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
3041 /* This function is called when the delayed ER timer fires. TCP enters
3042 * fast recovery and performs fast-retransmit.
3044 void tcp_resume_early_retransmit(struct sock *sk)
3046 struct tcp_sock *tp = tcp_sk(sk);
3050 /* Stop if ER is disabled after the delayed ER timer is scheduled */
3051 if (!tp->do_early_retrans)
3054 tcp_enter_recovery(sk, false);
3055 tcp_update_scoreboard(sk, 1);
3056 tcp_xmit_retransmit_queue(sk);
3059 /* If we get here, the whole TSO packet has not been acked. */
3060 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3062 struct tcp_sock *tp = tcp_sk(sk);
3065 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3067 packets_acked = tcp_skb_pcount(skb);
3068 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3070 packets_acked -= tcp_skb_pcount(skb);
3072 if (packets_acked) {
3073 BUG_ON(tcp_skb_pcount(skb) == 0);
3074 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3077 return packets_acked;
3080 static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
3083 const struct skb_shared_info *shinfo;
3085 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3086 if (likely(!TCP_SKB_CB(skb)->txstamp_ack))
3089 shinfo = skb_shinfo(skb);
3090 if ((shinfo->tx_flags & SKBTX_ACK_TSTAMP) &&
3091 !before(shinfo->tskey, prior_snd_una) &&
3092 before(shinfo->tskey, tcp_sk(sk)->snd_una))
3093 __skb_tstamp_tx(skb, NULL, sk, SCM_TSTAMP_ACK);
3096 /* Remove acknowledged frames from the retransmission queue. If our packet
3097 * is before the ack sequence we can discard it as it's confirmed to have
3098 * arrived at the other end.
3100 static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
3101 u32 prior_snd_una, int *acked,
3102 struct tcp_sacktag_state *sack)
3104 const struct inet_connection_sock *icsk = inet_csk(sk);
3105 struct skb_mstamp first_ackt, last_ackt, now;
3106 struct tcp_sock *tp = tcp_sk(sk);
3107 u32 prior_sacked = tp->sacked_out;
3108 u32 reord = tp->packets_out;
3109 bool fully_acked = true;
3110 long sack_rtt_us = -1L;
3111 long seq_rtt_us = -1L;
3112 long ca_rtt_us = -1L;
3113 struct sk_buff *skb;
3120 while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
3121 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3122 u8 sacked = scb->sacked;
3125 tcp_ack_tstamp(sk, skb, prior_snd_una);
3127 /* Determine how many packets and what bytes were acked, tso and else */
3128 if (after(scb->end_seq, tp->snd_una)) {
3129 if (tcp_skb_pcount(skb) == 1 ||
3130 !after(tp->snd_una, scb->seq))
3133 acked_pcount = tcp_tso_acked(sk, skb);
3137 fully_acked = false;
3139 /* Speedup tcp_unlink_write_queue() and next loop */
3140 prefetchw(skb->next);
3141 acked_pcount = tcp_skb_pcount(skb);
3144 if (unlikely(sacked & TCPCB_RETRANS)) {
3145 if (sacked & TCPCB_SACKED_RETRANS)
3146 tp->retrans_out -= acked_pcount;
3147 flag |= FLAG_RETRANS_DATA_ACKED;
3148 } else if (!(sacked & TCPCB_SACKED_ACKED)) {
3149 last_ackt = skb->skb_mstamp;
3150 WARN_ON_ONCE(last_ackt.v64 == 0);
3151 if (!first_ackt.v64)
3152 first_ackt = last_ackt;
3154 reord = min(pkts_acked, reord);
3155 if (!after(scb->end_seq, tp->high_seq))
3156 flag |= FLAG_ORIG_SACK_ACKED;
3159 if (sacked & TCPCB_SACKED_ACKED) {
3160 tp->sacked_out -= acked_pcount;
3161 } else if (tcp_is_sack(tp)) {
3162 tp->delivered += acked_pcount;
3163 if (!tcp_skb_spurious_retrans(tp, skb))
3164 tcp_rack_advance(tp, &skb->skb_mstamp, sacked);
3166 if (sacked & TCPCB_LOST)
3167 tp->lost_out -= acked_pcount;
3169 tp->packets_out -= acked_pcount;
3170 pkts_acked += acked_pcount;
3172 /* Initial outgoing SYN's get put onto the write_queue
3173 * just like anything else we transmit. It is not
3174 * true data, and if we misinform our callers that
3175 * this ACK acks real data, we will erroneously exit
3176 * connection startup slow start one packet too
3177 * quickly. This is severely frowned upon behavior.
3179 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
3180 flag |= FLAG_DATA_ACKED;
3182 flag |= FLAG_SYN_ACKED;
3183 tp->retrans_stamp = 0;
3189 tcp_unlink_write_queue(skb, sk);
3190 sk_wmem_free_skb(sk, skb);
3191 if (unlikely(skb == tp->retransmit_skb_hint))
3192 tp->retransmit_skb_hint = NULL;
3193 if (unlikely(skb == tp->lost_skb_hint))
3194 tp->lost_skb_hint = NULL;
3197 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3198 tp->snd_up = tp->snd_una;
3200 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3201 flag |= FLAG_SACK_RENEGING;
3203 skb_mstamp_get(&now);
3204 if (likely(first_ackt.v64) && !(flag & FLAG_RETRANS_DATA_ACKED)) {
3205 seq_rtt_us = skb_mstamp_us_delta(&now, &first_ackt);
3206 ca_rtt_us = skb_mstamp_us_delta(&now, &last_ackt);
3208 if (sack->first_sackt.v64) {
3209 sack_rtt_us = skb_mstamp_us_delta(&now, &sack->first_sackt);
3210 ca_rtt_us = skb_mstamp_us_delta(&now, &sack->last_sackt);
3213 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us,
3216 if (flag & FLAG_ACKED) {
3218 if (unlikely(icsk->icsk_mtup.probe_size &&
3219 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3220 tcp_mtup_probe_success(sk);
3223 if (tcp_is_reno(tp)) {
3224 tcp_remove_reno_sacks(sk, pkts_acked);
3228 /* Non-retransmitted hole got filled? That's reordering */
3229 if (reord < prior_fackets)
3230 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
3232 delta = tcp_is_fack(tp) ? pkts_acked :
3233 prior_sacked - tp->sacked_out;
3234 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3237 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
3239 } else if (skb && rtt_update && sack_rtt_us >= 0 &&
3240 sack_rtt_us > skb_mstamp_us_delta(&now, &skb->skb_mstamp)) {
3241 /* Do not re-arm RTO if the sack RTT is measured from data sent
3242 * after when the head was last (re)transmitted. Otherwise the
3243 * timeout may continue to extend in loss recovery.
3248 if (icsk->icsk_ca_ops->pkts_acked)
3249 icsk->icsk_ca_ops->pkts_acked(sk, pkts_acked, ca_rtt_us);
3251 #if FASTRETRANS_DEBUG > 0
3252 WARN_ON((int)tp->sacked_out < 0);
3253 WARN_ON((int)tp->lost_out < 0);
3254 WARN_ON((int)tp->retrans_out < 0);
3255 if (!tp->packets_out && tcp_is_sack(tp)) {
3256 icsk = inet_csk(sk);
3258 pr_debug("Leak l=%u %d\n",
3259 tp->lost_out, icsk->icsk_ca_state);
3262 if (tp->sacked_out) {
3263 pr_debug("Leak s=%u %d\n",
3264 tp->sacked_out, icsk->icsk_ca_state);
3267 if (tp->retrans_out) {
3268 pr_debug("Leak r=%u %d\n",
3269 tp->retrans_out, icsk->icsk_ca_state);
3270 tp->retrans_out = 0;
3274 *acked = pkts_acked;
3278 static void tcp_ack_probe(struct sock *sk)
3280 const struct tcp_sock *tp = tcp_sk(sk);
3281 struct inet_connection_sock *icsk = inet_csk(sk);
3283 /* Was it a usable window open? */
3285 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
3286 icsk->icsk_backoff = 0;
3287 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3288 /* Socket must be waked up by subsequent tcp_data_snd_check().
3289 * This function is not for random using!
3292 unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX);
3294 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3299 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3301 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3302 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3305 /* Decide wheather to run the increase function of congestion control. */
3306 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3308 /* If reordering is high then always grow cwnd whenever data is
3309 * delivered regardless of its ordering. Otherwise stay conservative
3310 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3311 * new SACK or ECE mark may first advance cwnd here and later reduce
3312 * cwnd in tcp_fastretrans_alert() based on more states.
3314 if (tcp_sk(sk)->reordering > sock_net(sk)->ipv4.sysctl_tcp_reordering)
3315 return flag & FLAG_FORWARD_PROGRESS;
3317 return flag & FLAG_DATA_ACKED;
3320 /* The "ultimate" congestion control function that aims to replace the rigid
3321 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction).
3322 * It's called toward the end of processing an ACK with precise rate
3323 * information. All transmission or retransmission are delayed afterwards.
3325 static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked,
3328 if (tcp_in_cwnd_reduction(sk)) {
3329 /* Reduce cwnd if state mandates */
3330 tcp_cwnd_reduction(sk, acked_sacked, flag);
3331 } else if (tcp_may_raise_cwnd(sk, flag)) {
3332 /* Advance cwnd if state allows */
3333 tcp_cong_avoid(sk, ack, acked_sacked);
3335 tcp_update_pacing_rate(sk);
3338 /* Check that window update is acceptable.
3339 * The function assumes that snd_una<=ack<=snd_next.
3341 static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3342 const u32 ack, const u32 ack_seq,
3345 return after(ack, tp->snd_una) ||
3346 after(ack_seq, tp->snd_wl1) ||
3347 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3350 /* If we update tp->snd_una, also update tp->bytes_acked */
3351 static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack)
3353 u32 delta = ack - tp->snd_una;
3355 u64_stats_update_begin(&tp->syncp);
3356 tp->bytes_acked += delta;
3357 u64_stats_update_end(&tp->syncp);
3361 /* If we update tp->rcv_nxt, also update tp->bytes_received */
3362 static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq)
3364 u32 delta = seq - tp->rcv_nxt;
3366 u64_stats_update_begin(&tp->syncp);
3367 tp->bytes_received += delta;
3368 u64_stats_update_end(&tp->syncp);
3372 /* Update our send window.
3374 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3375 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3377 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3380 struct tcp_sock *tp = tcp_sk(sk);
3382 u32 nwin = ntohs(tcp_hdr(skb)->window);
3384 if (likely(!tcp_hdr(skb)->syn))
3385 nwin <<= tp->rx_opt.snd_wscale;
3387 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3388 flag |= FLAG_WIN_UPDATE;
3389 tcp_update_wl(tp, ack_seq);
3391 if (tp->snd_wnd != nwin) {
3394 /* Note, it is the only place, where
3395 * fast path is recovered for sending TCP.
3398 tcp_fast_path_check(sk);
3400 if (tcp_send_head(sk))
3401 tcp_slow_start_after_idle_check(sk);
3403 if (nwin > tp->max_window) {
3404 tp->max_window = nwin;
3405 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3410 tcp_snd_una_update(tp, ack);
3415 /* Return true if we're currently rate-limiting out-of-window ACKs and
3416 * thus shouldn't send a dupack right now. We rate-limit dupacks in
3417 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3418 * attacks that send repeated SYNs or ACKs for the same connection. To
3419 * do this, we do not send a duplicate SYNACK or ACK if the remote
3420 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3422 bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb,
3423 int mib_idx, u32 *last_oow_ack_time)
3425 /* Data packets without SYNs are not likely part of an ACK loop. */
3426 if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) &&
3428 goto not_rate_limited;
3430 if (*last_oow_ack_time) {
3431 s32 elapsed = (s32)(tcp_time_stamp - *last_oow_ack_time);
3433 if (0 <= elapsed && elapsed < sysctl_tcp_invalid_ratelimit) {
3434 NET_INC_STATS_BH(net, mib_idx);
3435 return true; /* rate-limited: don't send yet! */
3439 *last_oow_ack_time = tcp_time_stamp;
3442 return false; /* not rate-limited: go ahead, send dupack now! */
3445 /* RFC 5961 7 [ACK Throttling] */
3446 static void tcp_send_challenge_ack(struct sock *sk, const struct sk_buff *skb)
3448 /* unprotected vars, we dont care of overwrites */
3449 static u32 challenge_timestamp;
3450 static unsigned int challenge_count;
3451 struct tcp_sock *tp = tcp_sk(sk);
3454 /* First check our per-socket dupack rate limit. */
3455 if (tcp_oow_rate_limited(sock_net(sk), skb,
3456 LINUX_MIB_TCPACKSKIPPEDCHALLENGE,
3457 &tp->last_oow_ack_time))
3460 /* Then check the check host-wide RFC 5961 rate limit. */
3462 if (now != challenge_timestamp) {
3463 challenge_timestamp = now;
3464 challenge_count = 0;
3466 if (++challenge_count <= sysctl_tcp_challenge_ack_limit) {
3467 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPCHALLENGEACK);
3472 static void tcp_store_ts_recent(struct tcp_sock *tp)
3474 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3475 tp->rx_opt.ts_recent_stamp = get_seconds();
3478 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3480 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3481 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3482 * extra check below makes sure this can only happen
3483 * for pure ACK frames. -DaveM
3485 * Not only, also it occurs for expired timestamps.
3488 if (tcp_paws_check(&tp->rx_opt, 0))
3489 tcp_store_ts_recent(tp);
3493 /* This routine deals with acks during a TLP episode.
3494 * We mark the end of a TLP episode on receiving TLP dupack or when
3495 * ack is after tlp_high_seq.
3496 * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3498 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3500 struct tcp_sock *tp = tcp_sk(sk);
3502 if (before(ack, tp->tlp_high_seq))
3505 if (flag & FLAG_DSACKING_ACK) {
3506 /* This DSACK means original and TLP probe arrived; no loss */
3507 tp->tlp_high_seq = 0;
3508 } else if (after(ack, tp->tlp_high_seq)) {
3509 /* ACK advances: there was a loss, so reduce cwnd. Reset
3510 * tlp_high_seq in tcp_init_cwnd_reduction()
3512 tcp_init_cwnd_reduction(sk);
3513 tcp_set_ca_state(sk, TCP_CA_CWR);
3514 tcp_end_cwnd_reduction(sk);
3515 tcp_try_keep_open(sk);
3516 NET_INC_STATS_BH(sock_net(sk),
3517 LINUX_MIB_TCPLOSSPROBERECOVERY);
3518 } else if (!(flag & (FLAG_SND_UNA_ADVANCED |
3519 FLAG_NOT_DUP | FLAG_DATA_SACKED))) {
3520 /* Pure dupack: original and TLP probe arrived; no loss */
3521 tp->tlp_high_seq = 0;
3525 static inline void tcp_in_ack_event(struct sock *sk, u32 flags)
3527 const struct inet_connection_sock *icsk = inet_csk(sk);
3529 if (icsk->icsk_ca_ops->in_ack_event)
3530 icsk->icsk_ca_ops->in_ack_event(sk, flags);
3533 /* Congestion control has updated the cwnd already. So if we're in
3534 * loss recovery then now we do any new sends (for FRTO) or
3535 * retransmits (for CA_Loss or CA_recovery) that make sense.
3537 static void tcp_xmit_recovery(struct sock *sk, int rexmit)
3539 struct tcp_sock *tp = tcp_sk(sk);
3541 if (rexmit == REXMIT_NONE)
3544 if (unlikely(rexmit == 2)) {
3545 __tcp_push_pending_frames(sk, tcp_current_mss(sk),
3547 if (after(tp->snd_nxt, tp->high_seq))
3551 tcp_xmit_retransmit_queue(sk);
3554 /* This routine deals with incoming acks, but not outgoing ones. */
3555 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3557 struct inet_connection_sock *icsk = inet_csk(sk);
3558 struct tcp_sock *tp = tcp_sk(sk);
3559 struct tcp_sacktag_state sack_state;
3560 u32 prior_snd_una = tp->snd_una;
3561 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3562 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3563 bool is_dupack = false;
3565 int prior_packets = tp->packets_out;
3566 u32 prior_delivered = tp->delivered;
3567 int acked = 0; /* Number of packets newly acked */
3568 int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */
3570 sack_state.first_sackt.v64 = 0;
3572 /* We very likely will need to access write queue head. */
3573 prefetchw(sk->sk_write_queue.next);
3575 /* If the ack is older than previous acks
3576 * then we can probably ignore it.
3578 if (before(ack, prior_snd_una)) {
3579 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3580 if (before(ack, prior_snd_una - tp->max_window)) {
3581 tcp_send_challenge_ack(sk, skb);
3587 /* If the ack includes data we haven't sent yet, discard
3588 * this segment (RFC793 Section 3.9).
3590 if (after(ack, tp->snd_nxt))
3593 if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
3594 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
3597 if (after(ack, prior_snd_una)) {
3598 flag |= FLAG_SND_UNA_ADVANCED;
3599 icsk->icsk_retransmits = 0;
3602 prior_fackets = tp->fackets_out;
3604 /* ts_recent update must be made after we are sure that the packet
3607 if (flag & FLAG_UPDATE_TS_RECENT)
3608 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3610 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3611 /* Window is constant, pure forward advance.
3612 * No more checks are required.
3613 * Note, we use the fact that SND.UNA>=SND.WL2.
3615 tcp_update_wl(tp, ack_seq);
3616 tcp_snd_una_update(tp, ack);
3617 flag |= FLAG_WIN_UPDATE;
3619 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
3621 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS);
3623 u32 ack_ev_flags = CA_ACK_SLOWPATH;
3625 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3628 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3630 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3632 if (TCP_SKB_CB(skb)->sacked)
3633 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3636 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) {
3638 ack_ev_flags |= CA_ACK_ECE;
3641 if (flag & FLAG_WIN_UPDATE)
3642 ack_ev_flags |= CA_ACK_WIN_UPDATE;
3644 tcp_in_ack_event(sk, ack_ev_flags);
3647 /* We passed data and got it acked, remove any soft error
3648 * log. Something worked...
3650 sk->sk_err_soft = 0;
3651 icsk->icsk_probes_out = 0;
3652 tp->rcv_tstamp = tcp_time_stamp;
3656 /* See if we can take anything off of the retransmit queue. */
3657 flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una, &acked,
3660 if (tcp_ack_is_dubious(sk, flag)) {
3661 is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
3662 tcp_fastretrans_alert(sk, acked, is_dupack, &flag, &rexmit);
3664 if (tp->tlp_high_seq)
3665 tcp_process_tlp_ack(sk, ack, flag);
3667 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) {
3668 struct dst_entry *dst = __sk_dst_get(sk);
3673 if (icsk->icsk_pending == ICSK_TIME_RETRANS)
3674 tcp_schedule_loss_probe(sk);
3675 tcp_cong_control(sk, ack, tp->delivered - prior_delivered, flag);
3676 tcp_xmit_recovery(sk, rexmit);
3680 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3681 if (flag & FLAG_DSACKING_ACK)
3682 tcp_fastretrans_alert(sk, acked, is_dupack, &flag, &rexmit);
3683 /* If this ack opens up a zero window, clear backoff. It was
3684 * being used to time the probes, and is probably far higher than
3685 * it needs to be for normal retransmission.
3687 if (tcp_send_head(sk))
3690 if (tp->tlp_high_seq)
3691 tcp_process_tlp_ack(sk, ack, flag);
3695 SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3699 /* If data was SACKed, tag it and see if we should send more data.
3700 * If data was DSACKed, see if we can undo a cwnd reduction.
3702 if (TCP_SKB_CB(skb)->sacked) {
3703 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3705 tcp_fastretrans_alert(sk, acked, is_dupack, &flag, &rexmit);
3706 tcp_xmit_recovery(sk, rexmit);
3709 SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3713 static void tcp_parse_fastopen_option(int len, const unsigned char *cookie,
3714 bool syn, struct tcp_fastopen_cookie *foc,
3717 /* Valid only in SYN or SYN-ACK with an even length. */
3718 if (!foc || !syn || len < 0 || (len & 1))
3721 if (len >= TCP_FASTOPEN_COOKIE_MIN &&
3722 len <= TCP_FASTOPEN_COOKIE_MAX)
3723 memcpy(foc->val, cookie, len);
3730 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3731 * But, this can also be called on packets in the established flow when
3732 * the fast version below fails.
3734 void tcp_parse_options(const struct sk_buff *skb,
3735 struct tcp_options_received *opt_rx, int estab,
3736 struct tcp_fastopen_cookie *foc)
3738 const unsigned char *ptr;
3739 const struct tcphdr *th = tcp_hdr(skb);
3740 int length = (th->doff * 4) - sizeof(struct tcphdr);
3742 ptr = (const unsigned char *)(th + 1);
3743 opt_rx->saw_tstamp = 0;
3745 while (length > 0) {
3746 int opcode = *ptr++;
3752 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3757 if (opsize < 2) /* "silly options" */
3759 if (opsize > length)
3760 return; /* don't parse partial options */
3763 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3764 u16 in_mss = get_unaligned_be16(ptr);
3766 if (opt_rx->user_mss &&
3767 opt_rx->user_mss < in_mss)
3768 in_mss = opt_rx->user_mss;
3769 opt_rx->mss_clamp = in_mss;
3774 if (opsize == TCPOLEN_WINDOW && th->syn &&
3775 !estab && sysctl_tcp_window_scaling) {
3776 __u8 snd_wscale = *(__u8 *)ptr;
3777 opt_rx->wscale_ok = 1;
3778 if (snd_wscale > 14) {
3779 net_info_ratelimited("%s: Illegal window scaling value %d >14 received\n",
3784 opt_rx->snd_wscale = snd_wscale;
3787 case TCPOPT_TIMESTAMP:
3788 if ((opsize == TCPOLEN_TIMESTAMP) &&
3789 ((estab && opt_rx->tstamp_ok) ||
3790 (!estab && sysctl_tcp_timestamps))) {
3791 opt_rx->saw_tstamp = 1;
3792 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3793 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3796 case TCPOPT_SACK_PERM:
3797 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3798 !estab && sysctl_tcp_sack) {
3799 opt_rx->sack_ok = TCP_SACK_SEEN;
3800 tcp_sack_reset(opt_rx);
3805 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3806 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3808 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3811 #ifdef CONFIG_TCP_MD5SIG
3814 * The MD5 Hash has already been
3815 * checked (see tcp_v{4,6}_do_rcv()).
3819 case TCPOPT_FASTOPEN:
3820 tcp_parse_fastopen_option(
3821 opsize - TCPOLEN_FASTOPEN_BASE,
3822 ptr, th->syn, foc, false);
3826 /* Fast Open option shares code 254 using a
3827 * 16 bits magic number.
3829 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE &&
3830 get_unaligned_be16(ptr) ==
3831 TCPOPT_FASTOPEN_MAGIC)
3832 tcp_parse_fastopen_option(opsize -
3833 TCPOLEN_EXP_FASTOPEN_BASE,
3834 ptr + 2, th->syn, foc, true);
3843 EXPORT_SYMBOL(tcp_parse_options);
3845 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
3847 const __be32 *ptr = (const __be32 *)(th + 1);
3849 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3850 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3851 tp->rx_opt.saw_tstamp = 1;
3853 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3856 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
3858 tp->rx_opt.rcv_tsecr = 0;
3864 /* Fast parse options. This hopes to only see timestamps.
3865 * If it is wrong it falls back on tcp_parse_options().
3867 static bool tcp_fast_parse_options(const struct sk_buff *skb,
3868 const struct tcphdr *th, struct tcp_sock *tp)
3870 /* In the spirit of fast parsing, compare doff directly to constant
3871 * values. Because equality is used, short doff can be ignored here.
3873 if (th->doff == (sizeof(*th) / 4)) {
3874 tp->rx_opt.saw_tstamp = 0;
3876 } else if (tp->rx_opt.tstamp_ok &&
3877 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3878 if (tcp_parse_aligned_timestamp(tp, th))
3882 tcp_parse_options(skb, &tp->rx_opt, 1, NULL);
3883 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3884 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
3889 #ifdef CONFIG_TCP_MD5SIG
3891 * Parse MD5 Signature option
3893 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
3895 int length = (th->doff << 2) - sizeof(*th);
3896 const u8 *ptr = (const u8 *)(th + 1);
3898 /* If the TCP option is too short, we can short cut */
3899 if (length < TCPOLEN_MD5SIG)
3902 while (length > 0) {
3903 int opcode = *ptr++;
3914 if (opsize < 2 || opsize > length)
3916 if (opcode == TCPOPT_MD5SIG)
3917 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
3924 EXPORT_SYMBOL(tcp_parse_md5sig_option);
3927 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3929 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3930 * it can pass through stack. So, the following predicate verifies that
3931 * this segment is not used for anything but congestion avoidance or
3932 * fast retransmit. Moreover, we even are able to eliminate most of such
3933 * second order effects, if we apply some small "replay" window (~RTO)
3934 * to timestamp space.
3936 * All these measures still do not guarantee that we reject wrapped ACKs
3937 * on networks with high bandwidth, when sequence space is recycled fastly,
3938 * but it guarantees that such events will be very rare and do not affect
3939 * connection seriously. This doesn't look nice, but alas, PAWS is really
3942 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3943 * states that events when retransmit arrives after original data are rare.
3944 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3945 * the biggest problem on large power networks even with minor reordering.
3946 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3947 * up to bandwidth of 18Gigabit/sec. 8) ]
3950 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3952 const struct tcp_sock *tp = tcp_sk(sk);
3953 const struct tcphdr *th = tcp_hdr(skb);
3954 u32 seq = TCP_SKB_CB(skb)->seq;
3955 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3957 return (/* 1. Pure ACK with correct sequence number. */
3958 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3960 /* 2. ... and duplicate ACK. */
3961 ack == tp->snd_una &&
3963 /* 3. ... and does not update window. */
3964 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3966 /* 4. ... and sits in replay window. */
3967 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3970 static inline bool tcp_paws_discard(const struct sock *sk,
3971 const struct sk_buff *skb)
3973 const struct tcp_sock *tp = tcp_sk(sk);
3975 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
3976 !tcp_disordered_ack(sk, skb);
3979 /* Check segment sequence number for validity.
3981 * Segment controls are considered valid, if the segment
3982 * fits to the window after truncation to the window. Acceptability
3983 * of data (and SYN, FIN, of course) is checked separately.
3984 * See tcp_data_queue(), for example.
3986 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3987 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3988 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3989 * (borrowed from freebsd)
3992 static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
3994 return !before(end_seq, tp->rcv_wup) &&
3995 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3998 /* When we get a reset we do this. */
3999 void tcp_reset(struct sock *sk)
4001 /* We want the right error as BSD sees it (and indeed as we do). */
4002 switch (sk->sk_state) {
4004 sk->sk_err = ECONNREFUSED;
4006 case TCP_CLOSE_WAIT:
4012 sk->sk_err = ECONNRESET;
4014 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4017 if (!sock_flag(sk, SOCK_DEAD))
4018 sk->sk_error_report(sk);
4024 * Process the FIN bit. This now behaves as it is supposed to work
4025 * and the FIN takes effect when it is validly part of sequence
4026 * space. Not before when we get holes.
4028 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4029 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4032 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4033 * close and we go into CLOSING (and later onto TIME-WAIT)
4035 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4037 void tcp_fin(struct sock *sk)
4039 struct tcp_sock *tp = tcp_sk(sk);
4041 inet_csk_schedule_ack(sk);
4043 sk->sk_shutdown |= RCV_SHUTDOWN;
4044 sock_set_flag(sk, SOCK_DONE);
4046 switch (sk->sk_state) {
4048 case TCP_ESTABLISHED:
4049 /* Move to CLOSE_WAIT */
4050 tcp_set_state(sk, TCP_CLOSE_WAIT);
4051 inet_csk(sk)->icsk_ack.pingpong = 1;
4054 case TCP_CLOSE_WAIT:
4056 /* Received a retransmission of the FIN, do
4061 /* RFC793: Remain in the LAST-ACK state. */
4065 /* This case occurs when a simultaneous close
4066 * happens, we must ack the received FIN and
4067 * enter the CLOSING state.
4070 tcp_set_state(sk, TCP_CLOSING);
4073 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4075 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4078 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4079 * cases we should never reach this piece of code.
4081 pr_err("%s: Impossible, sk->sk_state=%d\n",
4082 __func__, sk->sk_state);
4086 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4087 * Probably, we should reset in this case. For now drop them.
4089 __skb_queue_purge(&tp->out_of_order_queue);
4090 if (tcp_is_sack(tp))
4091 tcp_sack_reset(&tp->rx_opt);
4094 if (!sock_flag(sk, SOCK_DEAD)) {
4095 sk->sk_state_change(sk);
4097 /* Do not send POLL_HUP for half duplex close. */
4098 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4099 sk->sk_state == TCP_CLOSE)
4100 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4102 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4106 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4109 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4110 if (before(seq, sp->start_seq))
4111 sp->start_seq = seq;
4112 if (after(end_seq, sp->end_seq))
4113 sp->end_seq = end_seq;
4119 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4121 struct tcp_sock *tp = tcp_sk(sk);
4123 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4126 if (before(seq, tp->rcv_nxt))
4127 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4129 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4131 NET_INC_STATS_BH(sock_net(sk), mib_idx);
4133 tp->rx_opt.dsack = 1;
4134 tp->duplicate_sack[0].start_seq = seq;
4135 tp->duplicate_sack[0].end_seq = end_seq;
4139 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4141 struct tcp_sock *tp = tcp_sk(sk);
4143 if (!tp->rx_opt.dsack)
4144 tcp_dsack_set(sk, seq, end_seq);
4146 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4149 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4151 struct tcp_sock *tp = tcp_sk(sk);
4153 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4154 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4155 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4156 tcp_enter_quickack_mode(sk);
4158 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4159 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4161 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4162 end_seq = tp->rcv_nxt;
4163 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4170 /* These routines update the SACK block as out-of-order packets arrive or
4171 * in-order packets close up the sequence space.
4173 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4176 struct tcp_sack_block *sp = &tp->selective_acks[0];
4177 struct tcp_sack_block *swalk = sp + 1;
4179 /* See if the recent change to the first SACK eats into
4180 * or hits the sequence space of other SACK blocks, if so coalesce.
4182 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4183 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4186 /* Zap SWALK, by moving every further SACK up by one slot.
4187 * Decrease num_sacks.
4189 tp->rx_opt.num_sacks--;
4190 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4194 this_sack++, swalk++;
4198 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4200 struct tcp_sock *tp = tcp_sk(sk);
4201 struct tcp_sack_block *sp = &tp->selective_acks[0];
4202 int cur_sacks = tp->rx_opt.num_sacks;
4208 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4209 if (tcp_sack_extend(sp, seq, end_seq)) {
4210 /* Rotate this_sack to the first one. */
4211 for (; this_sack > 0; this_sack--, sp--)
4212 swap(*sp, *(sp - 1));
4214 tcp_sack_maybe_coalesce(tp);
4219 /* Could not find an adjacent existing SACK, build a new one,
4220 * put it at the front, and shift everyone else down. We
4221 * always know there is at least one SACK present already here.
4223 * If the sack array is full, forget about the last one.
4225 if (this_sack >= TCP_NUM_SACKS) {
4227 tp->rx_opt.num_sacks--;
4230 for (; this_sack > 0; this_sack--, sp--)
4234 /* Build the new head SACK, and we're done. */
4235 sp->start_seq = seq;
4236 sp->end_seq = end_seq;
4237 tp->rx_opt.num_sacks++;
4240 /* RCV.NXT advances, some SACKs should be eaten. */
4242 static void tcp_sack_remove(struct tcp_sock *tp)
4244 struct tcp_sack_block *sp = &tp->selective_acks[0];
4245 int num_sacks = tp->rx_opt.num_sacks;
4248 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4249 if (skb_queue_empty(&tp->out_of_order_queue)) {
4250 tp->rx_opt.num_sacks = 0;
4254 for (this_sack = 0; this_sack < num_sacks;) {
4255 /* Check if the start of the sack is covered by RCV.NXT. */
4256 if (!before(tp->rcv_nxt, sp->start_seq)) {
4259 /* RCV.NXT must cover all the block! */
4260 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4262 /* Zap this SACK, by moving forward any other SACKS. */
4263 for (i = this_sack+1; i < num_sacks; i++)
4264 tp->selective_acks[i-1] = tp->selective_acks[i];
4271 tp->rx_opt.num_sacks = num_sacks;
4275 * tcp_try_coalesce - try to merge skb to prior one
4278 * @from: buffer to add in queue
4279 * @fragstolen: pointer to boolean
4281 * Before queueing skb @from after @to, try to merge them
4282 * to reduce overall memory use and queue lengths, if cost is small.
4283 * Packets in ofo or receive queues can stay a long time.
4284 * Better try to coalesce them right now to avoid future collapses.
4285 * Returns true if caller should free @from instead of queueing it
4287 static bool tcp_try_coalesce(struct sock *sk,
4289 struct sk_buff *from,
4294 *fragstolen = false;
4296 /* Its possible this segment overlaps with prior segment in queue */
4297 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4300 if (!skb_try_coalesce(to, from, fragstolen, &delta))
4303 atomic_add(delta, &sk->sk_rmem_alloc);
4304 sk_mem_charge(sk, delta);
4305 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4306 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4307 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4308 TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
4312 static void tcp_drop(struct sock *sk, struct sk_buff *skb)
4314 sk_drops_add(sk, skb);
4318 /* This one checks to see if we can put data from the
4319 * out_of_order queue into the receive_queue.
4321 static void tcp_ofo_queue(struct sock *sk)
4323 struct tcp_sock *tp = tcp_sk(sk);
4324 __u32 dsack_high = tp->rcv_nxt;
4325 struct sk_buff *skb, *tail;
4326 bool fragstolen, eaten;
4328 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
4329 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4332 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4333 __u32 dsack = dsack_high;
4334 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4335 dsack_high = TCP_SKB_CB(skb)->end_seq;
4336 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4339 __skb_unlink(skb, &tp->out_of_order_queue);
4340 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4341 SOCK_DEBUG(sk, "ofo packet was already received\n");
4345 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4346 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4347 TCP_SKB_CB(skb)->end_seq);
4349 tail = skb_peek_tail(&sk->sk_receive_queue);
4350 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen);
4351 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4353 __skb_queue_tail(&sk->sk_receive_queue, skb);
4354 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
4357 kfree_skb_partial(skb, fragstolen);
4361 static bool tcp_prune_ofo_queue(struct sock *sk);
4362 static int tcp_prune_queue(struct sock *sk);
4364 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4367 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4368 !sk_rmem_schedule(sk, skb, size)) {
4370 if (tcp_prune_queue(sk) < 0)
4373 if (!sk_rmem_schedule(sk, skb, size)) {
4374 if (!tcp_prune_ofo_queue(sk))
4377 if (!sk_rmem_schedule(sk, skb, size))
4384 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4386 struct tcp_sock *tp = tcp_sk(sk);
4387 struct sk_buff *skb1;
4390 tcp_ecn_check_ce(tp, skb);
4392 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4393 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFODROP);
4398 /* Disable header prediction. */
4400 inet_csk_schedule_ack(sk);
4402 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4403 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4404 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4406 skb1 = skb_peek_tail(&tp->out_of_order_queue);
4408 /* Initial out of order segment, build 1 SACK. */
4409 if (tcp_is_sack(tp)) {
4410 tp->rx_opt.num_sacks = 1;
4411 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
4412 tp->selective_acks[0].end_seq =
4413 TCP_SKB_CB(skb)->end_seq;
4415 __skb_queue_head(&tp->out_of_order_queue, skb);
4419 seq = TCP_SKB_CB(skb)->seq;
4420 end_seq = TCP_SKB_CB(skb)->end_seq;
4422 if (seq == TCP_SKB_CB(skb1)->end_seq) {
4425 if (!tcp_try_coalesce(sk, skb1, skb, &fragstolen)) {
4426 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4428 tcp_grow_window(sk, skb);
4429 kfree_skb_partial(skb, fragstolen);
4433 if (!tp->rx_opt.num_sacks ||
4434 tp->selective_acks[0].end_seq != seq)
4437 /* Common case: data arrive in order after hole. */
4438 tp->selective_acks[0].end_seq = end_seq;
4442 /* Find place to insert this segment. */
4444 if (!after(TCP_SKB_CB(skb1)->seq, seq))
4446 if (skb_queue_is_first(&tp->out_of_order_queue, skb1)) {
4450 skb1 = skb_queue_prev(&tp->out_of_order_queue, skb1);
4453 /* Do skb overlap to previous one? */
4454 if (skb1 && before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4455 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4456 /* All the bits are present. Drop. */
4457 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4460 tcp_dsack_set(sk, seq, end_seq);
4463 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4464 /* Partial overlap. */
4465 tcp_dsack_set(sk, seq,
4466 TCP_SKB_CB(skb1)->end_seq);
4468 if (skb_queue_is_first(&tp->out_of_order_queue,
4472 skb1 = skb_queue_prev(
4473 &tp->out_of_order_queue,
4478 __skb_queue_head(&tp->out_of_order_queue, skb);
4480 __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4482 /* And clean segments covered by new one as whole. */
4483 while (!skb_queue_is_last(&tp->out_of_order_queue, skb)) {
4484 skb1 = skb_queue_next(&tp->out_of_order_queue, skb);
4486 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4488 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4489 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4493 __skb_unlink(skb1, &tp->out_of_order_queue);
4494 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4495 TCP_SKB_CB(skb1)->end_seq);
4496 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4501 if (tcp_is_sack(tp))
4502 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4505 tcp_grow_window(sk, skb);
4506 skb_set_owner_r(skb, sk);
4510 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, int hdrlen,
4514 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4516 __skb_pull(skb, hdrlen);
4518 tcp_try_coalesce(sk, tail, skb, fragstolen)) ? 1 : 0;
4519 tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq);
4521 __skb_queue_tail(&sk->sk_receive_queue, skb);
4522 skb_set_owner_r(skb, sk);
4527 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4529 struct sk_buff *skb;
4537 if (size > PAGE_SIZE) {
4538 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS);
4540 data_len = npages << PAGE_SHIFT;
4541 size = data_len + (size & ~PAGE_MASK);
4543 skb = alloc_skb_with_frags(size - data_len, data_len,
4544 PAGE_ALLOC_COSTLY_ORDER,
4545 &err, sk->sk_allocation);
4549 skb_put(skb, size - data_len);
4550 skb->data_len = data_len;
4553 if (tcp_try_rmem_schedule(sk, skb, skb->truesize))
4556 err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size);
4560 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4561 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4562 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4564 if (tcp_queue_rcv(sk, skb, 0, &fragstolen)) {
4565 WARN_ON_ONCE(fragstolen); /* should not happen */
4577 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4579 struct tcp_sock *tp = tcp_sk(sk);
4580 bool fragstolen = false;
4583 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
4588 __skb_pull(skb, tcp_hdr(skb)->doff * 4);
4590 tcp_ecn_accept_cwr(tp, skb);
4592 tp->rx_opt.dsack = 0;
4594 /* Queue data for delivery to the user.
4595 * Packets in sequence go to the receive queue.
4596 * Out of sequence packets to the out_of_order_queue.
4598 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4599 if (tcp_receive_window(tp) == 0)
4602 /* Ok. In sequence. In window. */
4603 if (tp->ucopy.task == current &&
4604 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4605 sock_owned_by_user(sk) && !tp->urg_data) {
4606 int chunk = min_t(unsigned int, skb->len,
4609 __set_current_state(TASK_RUNNING);
4612 if (!skb_copy_datagram_msg(skb, 0, tp->ucopy.msg, chunk)) {
4613 tp->ucopy.len -= chunk;
4614 tp->copied_seq += chunk;
4615 eaten = (chunk == skb->len);
4616 tcp_rcv_space_adjust(sk);
4624 if (skb_queue_len(&sk->sk_receive_queue) == 0)
4625 sk_forced_mem_schedule(sk, skb->truesize);
4626 else if (tcp_try_rmem_schedule(sk, skb, skb->truesize))
4629 eaten = tcp_queue_rcv(sk, skb, 0, &fragstolen);
4631 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4633 tcp_event_data_recv(sk, skb);
4634 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
4637 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4640 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4641 * gap in queue is filled.
4643 if (skb_queue_empty(&tp->out_of_order_queue))
4644 inet_csk(sk)->icsk_ack.pingpong = 0;
4647 if (tp->rx_opt.num_sacks)
4648 tcp_sack_remove(tp);
4650 tcp_fast_path_check(sk);
4653 kfree_skb_partial(skb, fragstolen);
4654 if (!sock_flag(sk, SOCK_DEAD))
4655 sk->sk_data_ready(sk);
4659 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4660 /* A retransmit, 2nd most common case. Force an immediate ack. */
4661 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4662 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4665 tcp_enter_quickack_mode(sk);
4666 inet_csk_schedule_ack(sk);
4672 /* Out of window. F.e. zero window probe. */
4673 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4676 tcp_enter_quickack_mode(sk);
4678 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4679 /* Partial packet, seq < rcv_next < end_seq */
4680 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4681 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4682 TCP_SKB_CB(skb)->end_seq);
4684 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4686 /* If window is closed, drop tail of packet. But after
4687 * remembering D-SACK for its head made in previous line.
4689 if (!tcp_receive_window(tp))
4694 tcp_data_queue_ofo(sk, skb);
4697 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4698 struct sk_buff_head *list)
4700 struct sk_buff *next = NULL;
4702 if (!skb_queue_is_last(list, skb))
4703 next = skb_queue_next(list, skb);
4705 __skb_unlink(skb, list);
4707 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4712 /* Collapse contiguous sequence of skbs head..tail with
4713 * sequence numbers start..end.
4715 * If tail is NULL, this means until the end of the list.
4717 * Segments with FIN/SYN are not collapsed (only because this
4721 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4722 struct sk_buff *head, struct sk_buff *tail,
4725 struct sk_buff *skb, *n;
4728 /* First, check that queue is collapsible and find
4729 * the point where collapsing can be useful. */
4733 skb_queue_walk_from_safe(list, skb, n) {
4736 /* No new bits? It is possible on ofo queue. */
4737 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4738 skb = tcp_collapse_one(sk, skb, list);
4744 /* The first skb to collapse is:
4746 * - bloated or contains data before "start" or
4747 * overlaps to the next one.
4749 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
4750 (tcp_win_from_space(skb->truesize) > skb->len ||
4751 before(TCP_SKB_CB(skb)->seq, start))) {
4752 end_of_skbs = false;
4756 if (!skb_queue_is_last(list, skb)) {
4757 struct sk_buff *next = skb_queue_next(list, skb);
4759 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(next)->seq) {
4760 end_of_skbs = false;
4765 /* Decided to skip this, advance start seq. */
4766 start = TCP_SKB_CB(skb)->end_seq;
4769 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4772 while (before(start, end)) {
4773 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
4774 struct sk_buff *nskb;
4776 nskb = alloc_skb(copy, GFP_ATOMIC);
4780 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4781 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4782 __skb_queue_before(list, skb, nskb);
4783 skb_set_owner_r(nskb, sk);
4785 /* Copy data, releasing collapsed skbs. */
4787 int offset = start - TCP_SKB_CB(skb)->seq;
4788 int size = TCP_SKB_CB(skb)->end_seq - start;
4792 size = min(copy, size);
4793 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4795 TCP_SKB_CB(nskb)->end_seq += size;
4799 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4800 skb = tcp_collapse_one(sk, skb, list);
4803 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4810 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4811 * and tcp_collapse() them until all the queue is collapsed.
4813 static void tcp_collapse_ofo_queue(struct sock *sk)
4815 struct tcp_sock *tp = tcp_sk(sk);
4816 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4817 struct sk_buff *head;
4823 start = TCP_SKB_CB(skb)->seq;
4824 end = TCP_SKB_CB(skb)->end_seq;
4828 struct sk_buff *next = NULL;
4830 if (!skb_queue_is_last(&tp->out_of_order_queue, skb))
4831 next = skb_queue_next(&tp->out_of_order_queue, skb);
4834 /* Segment is terminated when we see gap or when
4835 * we are at the end of all the queue. */
4837 after(TCP_SKB_CB(skb)->seq, end) ||
4838 before(TCP_SKB_CB(skb)->end_seq, start)) {
4839 tcp_collapse(sk, &tp->out_of_order_queue,
4840 head, skb, start, end);
4844 /* Start new segment */
4845 start = TCP_SKB_CB(skb)->seq;
4846 end = TCP_SKB_CB(skb)->end_seq;
4848 if (before(TCP_SKB_CB(skb)->seq, start))
4849 start = TCP_SKB_CB(skb)->seq;
4850 if (after(TCP_SKB_CB(skb)->end_seq, end))
4851 end = TCP_SKB_CB(skb)->end_seq;
4857 * Purge the out-of-order queue.
4858 * Return true if queue was pruned.
4860 static bool tcp_prune_ofo_queue(struct sock *sk)
4862 struct tcp_sock *tp = tcp_sk(sk);
4865 if (!skb_queue_empty(&tp->out_of_order_queue)) {
4866 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED);
4867 __skb_queue_purge(&tp->out_of_order_queue);
4869 /* Reset SACK state. A conforming SACK implementation will
4870 * do the same at a timeout based retransmit. When a connection
4871 * is in a sad state like this, we care only about integrity
4872 * of the connection not performance.
4874 if (tp->rx_opt.sack_ok)
4875 tcp_sack_reset(&tp->rx_opt);
4882 /* Reduce allocated memory if we can, trying to get
4883 * the socket within its memory limits again.
4885 * Return less than zero if we should start dropping frames
4886 * until the socket owning process reads some of the data
4887 * to stabilize the situation.
4889 static int tcp_prune_queue(struct sock *sk)
4891 struct tcp_sock *tp = tcp_sk(sk);
4893 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4895 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED);
4897 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4898 tcp_clamp_window(sk);
4899 else if (tcp_under_memory_pressure(sk))
4900 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4902 tcp_collapse_ofo_queue(sk);
4903 if (!skb_queue_empty(&sk->sk_receive_queue))
4904 tcp_collapse(sk, &sk->sk_receive_queue,
4905 skb_peek(&sk->sk_receive_queue),
4907 tp->copied_seq, tp->rcv_nxt);
4910 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4913 /* Collapsing did not help, destructive actions follow.
4914 * This must not ever occur. */
4916 tcp_prune_ofo_queue(sk);
4918 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4921 /* If we are really being abused, tell the caller to silently
4922 * drop receive data on the floor. It will get retransmitted
4923 * and hopefully then we'll have sufficient space.
4925 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED);
4927 /* Massive buffer overcommit. */
4932 static bool tcp_should_expand_sndbuf(const struct sock *sk)
4934 const struct tcp_sock *tp = tcp_sk(sk);
4936 /* If the user specified a specific send buffer setting, do
4939 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4942 /* If we are under global TCP memory pressure, do not expand. */
4943 if (tcp_under_memory_pressure(sk))
4946 /* If we are under soft global TCP memory pressure, do not expand. */
4947 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
4950 /* If we filled the congestion window, do not expand. */
4951 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
4957 /* When incoming ACK allowed to free some skb from write_queue,
4958 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4959 * on the exit from tcp input handler.
4961 * PROBLEM: sndbuf expansion does not work well with largesend.
4963 static void tcp_new_space(struct sock *sk)
4965 struct tcp_sock *tp = tcp_sk(sk);
4967 if (tcp_should_expand_sndbuf(sk)) {
4968 tcp_sndbuf_expand(sk);
4969 tp->snd_cwnd_stamp = tcp_time_stamp;
4972 sk->sk_write_space(sk);
4975 static void tcp_check_space(struct sock *sk)
4977 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
4978 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
4979 /* pairs with tcp_poll() */
4980 smp_mb__after_atomic();
4981 if (sk->sk_socket &&
4982 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
4987 static inline void tcp_data_snd_check(struct sock *sk)
4989 tcp_push_pending_frames(sk);
4990 tcp_check_space(sk);
4994 * Check if sending an ack is needed.
4996 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
4998 struct tcp_sock *tp = tcp_sk(sk);
5000 /* More than one full frame received... */
5001 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
5002 /* ... and right edge of window advances far enough.
5003 * (tcp_recvmsg() will send ACK otherwise). Or...
5005 __tcp_select_window(sk) >= tp->rcv_wnd) ||
5006 /* We ACK each frame or... */
5007 tcp_in_quickack_mode(sk) ||
5008 /* We have out of order data. */
5009 (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
5010 /* Then ack it now */
5013 /* Else, send delayed ack. */
5014 tcp_send_delayed_ack(sk);
5018 static inline void tcp_ack_snd_check(struct sock *sk)
5020 if (!inet_csk_ack_scheduled(sk)) {
5021 /* We sent a data segment already. */
5024 __tcp_ack_snd_check(sk, 1);
5028 * This routine is only called when we have urgent data
5029 * signaled. Its the 'slow' part of tcp_urg. It could be
5030 * moved inline now as tcp_urg is only called from one
5031 * place. We handle URGent data wrong. We have to - as
5032 * BSD still doesn't use the correction from RFC961.
5033 * For 1003.1g we should support a new option TCP_STDURG to permit
5034 * either form (or just set the sysctl tcp_stdurg).
5037 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
5039 struct tcp_sock *tp = tcp_sk(sk);
5040 u32 ptr = ntohs(th->urg_ptr);
5042 if (ptr && !sysctl_tcp_stdurg)
5044 ptr += ntohl(th->seq);
5046 /* Ignore urgent data that we've already seen and read. */
5047 if (after(tp->copied_seq, ptr))
5050 /* Do not replay urg ptr.
5052 * NOTE: interesting situation not covered by specs.
5053 * Misbehaving sender may send urg ptr, pointing to segment,
5054 * which we already have in ofo queue. We are not able to fetch
5055 * such data and will stay in TCP_URG_NOTYET until will be eaten
5056 * by recvmsg(). Seems, we are not obliged to handle such wicked
5057 * situations. But it is worth to think about possibility of some
5058 * DoSes using some hypothetical application level deadlock.
5060 if (before(ptr, tp->rcv_nxt))
5063 /* Do we already have a newer (or duplicate) urgent pointer? */
5064 if (tp->urg_data && !after(ptr, tp->urg_seq))
5067 /* Tell the world about our new urgent pointer. */
5070 /* We may be adding urgent data when the last byte read was
5071 * urgent. To do this requires some care. We cannot just ignore
5072 * tp->copied_seq since we would read the last urgent byte again
5073 * as data, nor can we alter copied_seq until this data arrives
5074 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5076 * NOTE. Double Dutch. Rendering to plain English: author of comment
5077 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5078 * and expect that both A and B disappear from stream. This is _wrong_.
5079 * Though this happens in BSD with high probability, this is occasional.
5080 * Any application relying on this is buggy. Note also, that fix "works"
5081 * only in this artificial test. Insert some normal data between A and B and we will
5082 * decline of BSD again. Verdict: it is better to remove to trap
5085 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5086 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5087 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5089 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5090 __skb_unlink(skb, &sk->sk_receive_queue);
5095 tp->urg_data = TCP_URG_NOTYET;
5098 /* Disable header prediction. */
5102 /* This is the 'fast' part of urgent handling. */
5103 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
5105 struct tcp_sock *tp = tcp_sk(sk);
5107 /* Check if we get a new urgent pointer - normally not. */
5109 tcp_check_urg(sk, th);
5111 /* Do we wait for any urgent data? - normally not... */
5112 if (tp->urg_data == TCP_URG_NOTYET) {
5113 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5116 /* Is the urgent pointer pointing into this packet? */
5117 if (ptr < skb->len) {
5119 if (skb_copy_bits(skb, ptr, &tmp, 1))
5121 tp->urg_data = TCP_URG_VALID | tmp;
5122 if (!sock_flag(sk, SOCK_DEAD))
5123 sk->sk_data_ready(sk);
5128 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
5130 struct tcp_sock *tp = tcp_sk(sk);
5131 int chunk = skb->len - hlen;
5135 if (skb_csum_unnecessary(skb))
5136 err = skb_copy_datagram_msg(skb, hlen, tp->ucopy.msg, chunk);
5138 err = skb_copy_and_csum_datagram_msg(skb, hlen, tp->ucopy.msg);
5141 tp->ucopy.len -= chunk;
5142 tp->copied_seq += chunk;
5143 tcp_rcv_space_adjust(sk);
5150 static __sum16 __tcp_checksum_complete_user(struct sock *sk,
5151 struct sk_buff *skb)
5155 if (sock_owned_by_user(sk)) {
5157 result = __tcp_checksum_complete(skb);
5160 result = __tcp_checksum_complete(skb);
5165 static inline bool tcp_checksum_complete_user(struct sock *sk,
5166 struct sk_buff *skb)
5168 return !skb_csum_unnecessary(skb) &&
5169 __tcp_checksum_complete_user(sk, skb);
5172 /* Does PAWS and seqno based validation of an incoming segment, flags will
5173 * play significant role here.
5175 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5176 const struct tcphdr *th, int syn_inerr)
5178 struct tcp_sock *tp = tcp_sk(sk);
5180 /* RFC1323: H1. Apply PAWS check first. */
5181 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
5182 tcp_paws_discard(sk, skb)) {
5184 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5185 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5186 LINUX_MIB_TCPACKSKIPPEDPAWS,
5187 &tp->last_oow_ack_time))
5188 tcp_send_dupack(sk, skb);
5191 /* Reset is accepted even if it did not pass PAWS. */
5194 /* Step 1: check sequence number */
5195 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5196 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5197 * (RST) segments are validated by checking their SEQ-fields."
5198 * And page 69: "If an incoming segment is not acceptable,
5199 * an acknowledgment should be sent in reply (unless the RST
5200 * bit is set, if so drop the segment and return)".
5205 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5206 LINUX_MIB_TCPACKSKIPPEDSEQ,
5207 &tp->last_oow_ack_time))
5208 tcp_send_dupack(sk, skb);
5213 /* Step 2: check RST bit */
5216 * If sequence number exactly matches RCV.NXT, then
5217 * RESET the connection
5219 * Send a challenge ACK
5221 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt)
5224 tcp_send_challenge_ack(sk, skb);
5228 /* step 3: check security and precedence [ignored] */
5230 /* step 4: Check for a SYN
5231 * RFC 5961 4.2 : Send a challenge ack
5236 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5237 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5238 tcp_send_challenge_ack(sk, skb);
5250 * TCP receive function for the ESTABLISHED state.
5252 * It is split into a fast path and a slow path. The fast path is
5254 * - A zero window was announced from us - zero window probing
5255 * is only handled properly in the slow path.
5256 * - Out of order segments arrived.
5257 * - Urgent data is expected.
5258 * - There is no buffer space left
5259 * - Unexpected TCP flags/window values/header lengths are received
5260 * (detected by checking the TCP header against pred_flags)
5261 * - Data is sent in both directions. Fast path only supports pure senders
5262 * or pure receivers (this means either the sequence number or the ack
5263 * value must stay constant)
5264 * - Unexpected TCP option.
5266 * When these conditions are not satisfied it drops into a standard
5267 * receive procedure patterned after RFC793 to handle all cases.
5268 * The first three cases are guaranteed by proper pred_flags setting,
5269 * the rest is checked inline. Fast processing is turned on in
5270 * tcp_data_queue when everything is OK.
5272 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5273 const struct tcphdr *th, unsigned int len)
5275 struct tcp_sock *tp = tcp_sk(sk);
5277 if (unlikely(!sk->sk_rx_dst))
5278 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
5280 * Header prediction.
5281 * The code loosely follows the one in the famous
5282 * "30 instruction TCP receive" Van Jacobson mail.
5284 * Van's trick is to deposit buffers into socket queue
5285 * on a device interrupt, to call tcp_recv function
5286 * on the receive process context and checksum and copy
5287 * the buffer to user space. smart...
5289 * Our current scheme is not silly either but we take the
5290 * extra cost of the net_bh soft interrupt processing...
5291 * We do checksum and copy also but from device to kernel.
5294 tp->rx_opt.saw_tstamp = 0;
5296 /* pred_flags is 0xS?10 << 16 + snd_wnd
5297 * if header_prediction is to be made
5298 * 'S' will always be tp->tcp_header_len >> 2
5299 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5300 * turn it off (when there are holes in the receive
5301 * space for instance)
5302 * PSH flag is ignored.
5305 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5306 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5307 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5308 int tcp_header_len = tp->tcp_header_len;
5310 /* Timestamp header prediction: tcp_header_len
5311 * is automatically equal to th->doff*4 due to pred_flags
5315 /* Check timestamp */
5316 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5317 /* No? Slow path! */
5318 if (!tcp_parse_aligned_timestamp(tp, th))
5321 /* If PAWS failed, check it more carefully in slow path */
5322 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5325 /* DO NOT update ts_recent here, if checksum fails
5326 * and timestamp was corrupted part, it will result
5327 * in a hung connection since we will drop all
5328 * future packets due to the PAWS test.
5332 if (len <= tcp_header_len) {
5333 /* Bulk data transfer: sender */
5334 if (len == tcp_header_len) {
5335 /* Predicted packet is in window by definition.
5336 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5337 * Hence, check seq<=rcv_wup reduces to:
5339 if (tcp_header_len ==
5340 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5341 tp->rcv_nxt == tp->rcv_wup)
5342 tcp_store_ts_recent(tp);
5344 /* We know that such packets are checksummed
5347 tcp_ack(sk, skb, 0);
5349 tcp_data_snd_check(sk);
5351 } else { /* Header too small */
5352 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5357 bool fragstolen = false;
5359 if (tp->ucopy.task == current &&
5360 tp->copied_seq == tp->rcv_nxt &&
5361 len - tcp_header_len <= tp->ucopy.len &&
5362 sock_owned_by_user(sk)) {
5363 __set_current_state(TASK_RUNNING);
5365 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len)) {
5366 /* Predicted packet is in window by definition.
5367 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5368 * Hence, check seq<=rcv_wup reduces to:
5370 if (tcp_header_len ==
5371 (sizeof(struct tcphdr) +
5372 TCPOLEN_TSTAMP_ALIGNED) &&
5373 tp->rcv_nxt == tp->rcv_wup)
5374 tcp_store_ts_recent(tp);
5376 tcp_rcv_rtt_measure_ts(sk, skb);
5378 __skb_pull(skb, tcp_header_len);
5379 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
5380 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER);
5385 if (tcp_checksum_complete_user(sk, skb))
5388 if ((int)skb->truesize > sk->sk_forward_alloc)
5391 /* Predicted packet is in window by definition.
5392 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5393 * Hence, check seq<=rcv_wup reduces to:
5395 if (tcp_header_len ==
5396 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5397 tp->rcv_nxt == tp->rcv_wup)
5398 tcp_store_ts_recent(tp);
5400 tcp_rcv_rtt_measure_ts(sk, skb);
5402 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS);
5404 /* Bulk data transfer: receiver */
5405 eaten = tcp_queue_rcv(sk, skb, tcp_header_len,
5409 tcp_event_data_recv(sk, skb);
5411 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5412 /* Well, only one small jumplet in fast path... */
5413 tcp_ack(sk, skb, FLAG_DATA);
5414 tcp_data_snd_check(sk);
5415 if (!inet_csk_ack_scheduled(sk))
5419 __tcp_ack_snd_check(sk, 0);
5422 kfree_skb_partial(skb, fragstolen);
5423 sk->sk_data_ready(sk);
5429 if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb))
5432 if (!th->ack && !th->rst && !th->syn)
5436 * Standard slow path.
5439 if (!tcp_validate_incoming(sk, skb, th, 1))
5443 if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0)
5446 tcp_rcv_rtt_measure_ts(sk, skb);
5448 /* Process urgent data. */
5449 tcp_urg(sk, skb, th);
5451 /* step 7: process the segment text */
5452 tcp_data_queue(sk, skb);
5454 tcp_data_snd_check(sk);
5455 tcp_ack_snd_check(sk);
5459 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_CSUMERRORS);
5460 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5465 EXPORT_SYMBOL(tcp_rcv_established);
5467 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5469 struct tcp_sock *tp = tcp_sk(sk);
5470 struct inet_connection_sock *icsk = inet_csk(sk);
5472 tcp_set_state(sk, TCP_ESTABLISHED);
5475 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
5476 security_inet_conn_established(sk, skb);
5479 /* Make sure socket is routed, for correct metrics. */
5480 icsk->icsk_af_ops->rebuild_header(sk);
5482 tcp_init_metrics(sk);
5484 tcp_init_congestion_control(sk);
5486 /* Prevent spurious tcp_cwnd_restart() on first data
5489 tp->lsndtime = tcp_time_stamp;
5491 tcp_init_buffer_space(sk);
5493 if (sock_flag(sk, SOCK_KEEPOPEN))
5494 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5496 if (!tp->rx_opt.snd_wscale)
5497 __tcp_fast_path_on(tp, tp->snd_wnd);
5501 if (!sock_flag(sk, SOCK_DEAD)) {
5502 sk->sk_state_change(sk);
5503 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5507 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
5508 struct tcp_fastopen_cookie *cookie)
5510 struct tcp_sock *tp = tcp_sk(sk);
5511 struct sk_buff *data = tp->syn_data ? tcp_write_queue_head(sk) : NULL;
5512 u16 mss = tp->rx_opt.mss_clamp, try_exp = 0;
5513 bool syn_drop = false;
5515 if (mss == tp->rx_opt.user_mss) {
5516 struct tcp_options_received opt;
5518 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5519 tcp_clear_options(&opt);
5520 opt.user_mss = opt.mss_clamp = 0;
5521 tcp_parse_options(synack, &opt, 0, NULL);
5522 mss = opt.mss_clamp;
5525 if (!tp->syn_fastopen) {
5526 /* Ignore an unsolicited cookie */
5528 } else if (tp->total_retrans) {
5529 /* SYN timed out and the SYN-ACK neither has a cookie nor
5530 * acknowledges data. Presumably the remote received only
5531 * the retransmitted (regular) SYNs: either the original
5532 * SYN-data or the corresponding SYN-ACK was dropped.
5534 syn_drop = (cookie->len < 0 && data);
5535 } else if (cookie->len < 0 && !tp->syn_data) {
5536 /* We requested a cookie but didn't get it. If we did not use
5537 * the (old) exp opt format then try so next time (try_exp=1).
5538 * Otherwise we go back to use the RFC7413 opt (try_exp=2).
5540 try_exp = tp->syn_fastopen_exp ? 2 : 1;
5543 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp);
5545 if (data) { /* Retransmit unacked data in SYN */
5546 tcp_for_write_queue_from(data, sk) {
5547 if (data == tcp_send_head(sk) ||
5548 __tcp_retransmit_skb(sk, data, 1))
5552 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVEFAIL);
5555 tp->syn_data_acked = tp->syn_data;
5556 if (tp->syn_data_acked)
5557 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
5559 tcp_fastopen_add_skb(sk, synack);
5564 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5565 const struct tcphdr *th)
5567 struct inet_connection_sock *icsk = inet_csk(sk);
5568 struct tcp_sock *tp = tcp_sk(sk);
5569 struct tcp_fastopen_cookie foc = { .len = -1 };
5570 int saved_clamp = tp->rx_opt.mss_clamp;
5572 tcp_parse_options(skb, &tp->rx_opt, 0, &foc);
5573 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
5574 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
5578 * "If the state is SYN-SENT then
5579 * first check the ACK bit
5580 * If the ACK bit is set
5581 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5582 * a reset (unless the RST bit is set, if so drop
5583 * the segment and return)"
5585 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
5586 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt))
5587 goto reset_and_undo;
5589 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5590 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5592 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED);
5593 goto reset_and_undo;
5596 /* Now ACK is acceptable.
5598 * "If the RST bit is set
5599 * If the ACK was acceptable then signal the user "error:
5600 * connection reset", drop the segment, enter CLOSED state,
5601 * delete TCB, and return."
5610 * "fifth, if neither of the SYN or RST bits is set then
5611 * drop the segment and return."
5617 goto discard_and_undo;
5620 * "If the SYN bit is on ...
5621 * are acceptable then ...
5622 * (our SYN has been ACKed), change the connection
5623 * state to ESTABLISHED..."
5626 tcp_ecn_rcv_synack(tp, th);
5628 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5629 tcp_ack(sk, skb, FLAG_SLOWPATH);
5631 /* Ok.. it's good. Set up sequence numbers and
5632 * move to established.
5634 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5635 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5637 /* RFC1323: The window in SYN & SYN/ACK segments is
5640 tp->snd_wnd = ntohs(th->window);
5642 if (!tp->rx_opt.wscale_ok) {
5643 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5644 tp->window_clamp = min(tp->window_clamp, 65535U);
5647 if (tp->rx_opt.saw_tstamp) {
5648 tp->rx_opt.tstamp_ok = 1;
5649 tp->tcp_header_len =
5650 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5651 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5652 tcp_store_ts_recent(tp);
5654 tp->tcp_header_len = sizeof(struct tcphdr);
5657 if (tcp_is_sack(tp) && sysctl_tcp_fack)
5658 tcp_enable_fack(tp);
5661 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5662 tcp_initialize_rcv_mss(sk);
5664 /* Remember, tcp_poll() does not lock socket!
5665 * Change state from SYN-SENT only after copied_seq
5666 * is initialized. */
5667 tp->copied_seq = tp->rcv_nxt;
5671 tcp_finish_connect(sk, skb);
5673 if ((tp->syn_fastopen || tp->syn_data) &&
5674 tcp_rcv_fastopen_synack(sk, skb, &foc))
5677 if (sk->sk_write_pending ||
5678 icsk->icsk_accept_queue.rskq_defer_accept ||
5679 icsk->icsk_ack.pingpong) {
5680 /* Save one ACK. Data will be ready after
5681 * several ticks, if write_pending is set.
5683 * It may be deleted, but with this feature tcpdumps
5684 * look so _wonderfully_ clever, that I was not able
5685 * to stand against the temptation 8) --ANK
5687 inet_csk_schedule_ack(sk);
5688 icsk->icsk_ack.lrcvtime = tcp_time_stamp;
5689 tcp_enter_quickack_mode(sk);
5690 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5691 TCP_DELACK_MAX, TCP_RTO_MAX);
5702 /* No ACK in the segment */
5706 * "If the RST bit is set
5708 * Otherwise (no ACK) drop the segment and return."
5711 goto discard_and_undo;
5715 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5716 tcp_paws_reject(&tp->rx_opt, 0))
5717 goto discard_and_undo;
5720 /* We see SYN without ACK. It is attempt of
5721 * simultaneous connect with crossed SYNs.
5722 * Particularly, it can be connect to self.
5724 tcp_set_state(sk, TCP_SYN_RECV);
5726 if (tp->rx_opt.saw_tstamp) {
5727 tp->rx_opt.tstamp_ok = 1;
5728 tcp_store_ts_recent(tp);
5729 tp->tcp_header_len =
5730 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5732 tp->tcp_header_len = sizeof(struct tcphdr);
5735 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5736 tp->copied_seq = tp->rcv_nxt;
5737 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5739 /* RFC1323: The window in SYN & SYN/ACK segments is
5742 tp->snd_wnd = ntohs(th->window);
5743 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5744 tp->max_window = tp->snd_wnd;
5746 tcp_ecn_rcv_syn(tp, th);
5749 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5750 tcp_initialize_rcv_mss(sk);
5752 tcp_send_synack(sk);
5754 /* Note, we could accept data and URG from this segment.
5755 * There are no obstacles to make this (except that we must
5756 * either change tcp_recvmsg() to prevent it from returning data
5757 * before 3WHS completes per RFC793, or employ TCP Fast Open).
5759 * However, if we ignore data in ACKless segments sometimes,
5760 * we have no reasons to accept it sometimes.
5761 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5762 * is not flawless. So, discard packet for sanity.
5763 * Uncomment this return to process the data.
5770 /* "fifth, if neither of the SYN or RST bits is set then
5771 * drop the segment and return."
5775 tcp_clear_options(&tp->rx_opt);
5776 tp->rx_opt.mss_clamp = saved_clamp;
5780 tcp_clear_options(&tp->rx_opt);
5781 tp->rx_opt.mss_clamp = saved_clamp;
5786 * This function implements the receiving procedure of RFC 793 for
5787 * all states except ESTABLISHED and TIME_WAIT.
5788 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5789 * address independent.
5792 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb)
5794 struct tcp_sock *tp = tcp_sk(sk);
5795 struct inet_connection_sock *icsk = inet_csk(sk);
5796 const struct tcphdr *th = tcp_hdr(skb);
5797 struct request_sock *req;
5801 switch (sk->sk_state) {
5815 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
5818 /* Now we have several options: In theory there is
5819 * nothing else in the frame. KA9Q has an option to
5820 * send data with the syn, BSD accepts data with the
5821 * syn up to the [to be] advertised window and
5822 * Solaris 2.1 gives you a protocol error. For now
5823 * we just ignore it, that fits the spec precisely
5824 * and avoids incompatibilities. It would be nice in
5825 * future to drop through and process the data.
5827 * Now that TTCP is starting to be used we ought to
5829 * But, this leaves one open to an easy denial of
5830 * service attack, and SYN cookies can't defend
5831 * against this problem. So, we drop the data
5832 * in the interest of security over speed unless
5833 * it's still in use.
5841 tp->rx_opt.saw_tstamp = 0;
5842 queued = tcp_rcv_synsent_state_process(sk, skb, th);
5846 /* Do step6 onward by hand. */
5847 tcp_urg(sk, skb, th);
5849 tcp_data_snd_check(sk);
5853 tp->rx_opt.saw_tstamp = 0;
5854 req = tp->fastopen_rsk;
5856 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
5857 sk->sk_state != TCP_FIN_WAIT1);
5859 if (!tcp_check_req(sk, skb, req, true))
5863 if (!th->ack && !th->rst && !th->syn)
5866 if (!tcp_validate_incoming(sk, skb, th, 0))
5869 /* step 5: check the ACK field */
5870 acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
5871 FLAG_UPDATE_TS_RECENT) > 0;
5873 switch (sk->sk_state) {
5879 tcp_synack_rtt_meas(sk, req);
5881 /* Once we leave TCP_SYN_RECV, we no longer need req
5885 tp->total_retrans = req->num_retrans;
5886 reqsk_fastopen_remove(sk, req, false);
5888 /* Make sure socket is routed, for correct metrics. */
5889 icsk->icsk_af_ops->rebuild_header(sk);
5890 tcp_init_congestion_control(sk);
5893 tp->copied_seq = tp->rcv_nxt;
5894 tcp_init_buffer_space(sk);
5897 tcp_set_state(sk, TCP_ESTABLISHED);
5898 sk->sk_state_change(sk);
5900 /* Note, that this wakeup is only for marginal crossed SYN case.
5901 * Passively open sockets are not waked up, because
5902 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
5905 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5907 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5908 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
5909 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5911 if (tp->rx_opt.tstamp_ok)
5912 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5915 /* Re-arm the timer because data may have been sent out.
5916 * This is similar to the regular data transmission case
5917 * when new data has just been ack'ed.
5919 * (TFO) - we could try to be more aggressive and
5920 * retransmitting any data sooner based on when they
5925 tcp_init_metrics(sk);
5927 tcp_update_pacing_rate(sk);
5929 /* Prevent spurious tcp_cwnd_restart() on first data packet */
5930 tp->lsndtime = tcp_time_stamp;
5932 tcp_initialize_rcv_mss(sk);
5933 tcp_fast_path_on(tp);
5936 case TCP_FIN_WAIT1: {
5937 struct dst_entry *dst;
5940 /* If we enter the TCP_FIN_WAIT1 state and we are a
5941 * Fast Open socket and this is the first acceptable
5942 * ACK we have received, this would have acknowledged
5943 * our SYNACK so stop the SYNACK timer.
5946 /* Return RST if ack_seq is invalid.
5947 * Note that RFC793 only says to generate a
5948 * DUPACK for it but for TCP Fast Open it seems
5949 * better to treat this case like TCP_SYN_RECV
5954 /* We no longer need the request sock. */
5955 reqsk_fastopen_remove(sk, req, false);
5958 if (tp->snd_una != tp->write_seq)
5961 tcp_set_state(sk, TCP_FIN_WAIT2);
5962 sk->sk_shutdown |= SEND_SHUTDOWN;
5964 dst = __sk_dst_get(sk);
5968 if (!sock_flag(sk, SOCK_DEAD)) {
5969 /* Wake up lingering close() */
5970 sk->sk_state_change(sk);
5974 if (tp->linger2 < 0 ||
5975 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5976 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
5978 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5982 tmo = tcp_fin_time(sk);
5983 if (tmo > TCP_TIMEWAIT_LEN) {
5984 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5985 } else if (th->fin || sock_owned_by_user(sk)) {
5986 /* Bad case. We could lose such FIN otherwise.
5987 * It is not a big problem, but it looks confusing
5988 * and not so rare event. We still can lose it now,
5989 * if it spins in bh_lock_sock(), but it is really
5992 inet_csk_reset_keepalive_timer(sk, tmo);
5994 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
6001 if (tp->snd_una == tp->write_seq) {
6002 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
6008 if (tp->snd_una == tp->write_seq) {
6009 tcp_update_metrics(sk);
6016 /* step 6: check the URG bit */
6017 tcp_urg(sk, skb, th);
6019 /* step 7: process the segment text */
6020 switch (sk->sk_state) {
6021 case TCP_CLOSE_WAIT:
6024 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
6028 /* RFC 793 says to queue data in these states,
6029 * RFC 1122 says we MUST send a reset.
6030 * BSD 4.4 also does reset.
6032 if (sk->sk_shutdown & RCV_SHUTDOWN) {
6033 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6034 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6035 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6041 case TCP_ESTABLISHED:
6042 tcp_data_queue(sk, skb);
6047 /* tcp_data could move socket to TIME-WAIT */
6048 if (sk->sk_state != TCP_CLOSE) {
6049 tcp_data_snd_check(sk);
6050 tcp_ack_snd_check(sk);
6059 EXPORT_SYMBOL(tcp_rcv_state_process);
6061 static inline void pr_drop_req(struct request_sock *req, __u16 port, int family)
6063 struct inet_request_sock *ireq = inet_rsk(req);
6065 if (family == AF_INET)
6066 net_dbg_ratelimited("drop open request from %pI4/%u\n",
6067 &ireq->ir_rmt_addr, port);
6068 #if IS_ENABLED(CONFIG_IPV6)
6069 else if (family == AF_INET6)
6070 net_dbg_ratelimited("drop open request from %pI6/%u\n",
6071 &ireq->ir_v6_rmt_addr, port);
6075 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
6077 * If we receive a SYN packet with these bits set, it means a
6078 * network is playing bad games with TOS bits. In order to
6079 * avoid possible false congestion notifications, we disable
6080 * TCP ECN negotiation.
6082 * Exception: tcp_ca wants ECN. This is required for DCTCP
6083 * congestion control: Linux DCTCP asserts ECT on all packets,
6084 * including SYN, which is most optimal solution; however,
6085 * others, such as FreeBSD do not.
6087 static void tcp_ecn_create_request(struct request_sock *req,
6088 const struct sk_buff *skb,
6089 const struct sock *listen_sk,
6090 const struct dst_entry *dst)
6092 const struct tcphdr *th = tcp_hdr(skb);
6093 const struct net *net = sock_net(listen_sk);
6094 bool th_ecn = th->ece && th->cwr;
6101 ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield);
6102 ecn_ok_dst = dst_feature(dst, DST_FEATURE_ECN_MASK);
6103 ecn_ok = net->ipv4.sysctl_tcp_ecn || ecn_ok_dst;
6105 if ((!ect && ecn_ok) || tcp_ca_needs_ecn(listen_sk) ||
6106 (ecn_ok_dst & DST_FEATURE_ECN_CA))
6107 inet_rsk(req)->ecn_ok = 1;
6110 static void tcp_openreq_init(struct request_sock *req,
6111 const struct tcp_options_received *rx_opt,
6112 struct sk_buff *skb, const struct sock *sk)
6114 struct inet_request_sock *ireq = inet_rsk(req);
6116 req->rsk_rcv_wnd = 0; /* So that tcp_send_synack() knows! */
6118 tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq;
6119 tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
6120 skb_mstamp_get(&tcp_rsk(req)->snt_synack);
6121 tcp_rsk(req)->last_oow_ack_time = 0;
6122 req->mss = rx_opt->mss_clamp;
6123 req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0;
6124 ireq->tstamp_ok = rx_opt->tstamp_ok;
6125 ireq->sack_ok = rx_opt->sack_ok;
6126 ireq->snd_wscale = rx_opt->snd_wscale;
6127 ireq->wscale_ok = rx_opt->wscale_ok;
6130 ireq->ir_rmt_port = tcp_hdr(skb)->source;
6131 ireq->ir_num = ntohs(tcp_hdr(skb)->dest);
6132 ireq->ir_mark = inet_request_mark(sk, skb);
6135 struct request_sock *inet_reqsk_alloc(const struct request_sock_ops *ops,
6136 struct sock *sk_listener,
6137 bool attach_listener)
6139 struct request_sock *req = reqsk_alloc(ops, sk_listener,
6143 struct inet_request_sock *ireq = inet_rsk(req);
6145 kmemcheck_annotate_bitfield(ireq, flags);
6147 atomic64_set(&ireq->ir_cookie, 0);
6148 ireq->ireq_state = TCP_NEW_SYN_RECV;
6149 write_pnet(&ireq->ireq_net, sock_net(sk_listener));
6150 ireq->ireq_family = sk_listener->sk_family;
6155 EXPORT_SYMBOL(inet_reqsk_alloc);
6158 * Return true if a syncookie should be sent
6160 static bool tcp_syn_flood_action(const struct sock *sk,
6161 const struct sk_buff *skb,
6164 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
6165 const char *msg = "Dropping request";
6166 bool want_cookie = false;
6167 struct net *net = sock_net(sk);
6169 #ifdef CONFIG_SYN_COOKIES
6170 if (net->ipv4.sysctl_tcp_syncookies) {
6171 msg = "Sending cookies";
6173 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES);
6176 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP);
6178 if (!queue->synflood_warned &&
6179 net->ipv4.sysctl_tcp_syncookies != 2 &&
6180 xchg(&queue->synflood_warned, 1) == 0)
6181 pr_info("%s: Possible SYN flooding on port %d. %s. Check SNMP counters.\n",
6182 proto, ntohs(tcp_hdr(skb)->dest), msg);
6187 static void tcp_reqsk_record_syn(const struct sock *sk,
6188 struct request_sock *req,
6189 const struct sk_buff *skb)
6191 if (tcp_sk(sk)->save_syn) {
6192 u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb);
6195 copy = kmalloc(len + sizeof(u32), GFP_ATOMIC);
6198 memcpy(©[1], skb_network_header(skb), len);
6199 req->saved_syn = copy;
6204 int tcp_conn_request(struct request_sock_ops *rsk_ops,
6205 const struct tcp_request_sock_ops *af_ops,
6206 struct sock *sk, struct sk_buff *skb)
6208 struct tcp_fastopen_cookie foc = { .len = -1 };
6209 __u32 isn = TCP_SKB_CB(skb)->tcp_tw_isn;
6210 struct tcp_options_received tmp_opt;
6211 struct tcp_sock *tp = tcp_sk(sk);
6212 struct net *net = sock_net(sk);
6213 struct sock *fastopen_sk = NULL;
6214 struct dst_entry *dst = NULL;
6215 struct request_sock *req;
6216 bool want_cookie = false;
6219 /* TW buckets are converted to open requests without
6220 * limitations, they conserve resources and peer is
6221 * evidently real one.
6223 if ((net->ipv4.sysctl_tcp_syncookies == 2 ||
6224 inet_csk_reqsk_queue_is_full(sk)) && !isn) {
6225 want_cookie = tcp_syn_flood_action(sk, skb, rsk_ops->slab_name);
6231 /* Accept backlog is full. If we have already queued enough
6232 * of warm entries in syn queue, drop request. It is better than
6233 * clogging syn queue with openreqs with exponentially increasing
6236 if (sk_acceptq_is_full(sk) && inet_csk_reqsk_queue_young(sk) > 1) {
6237 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
6241 req = inet_reqsk_alloc(rsk_ops, sk, !want_cookie);
6245 tcp_rsk(req)->af_specific = af_ops;
6247 tcp_clear_options(&tmp_opt);
6248 tmp_opt.mss_clamp = af_ops->mss_clamp;
6249 tmp_opt.user_mss = tp->rx_opt.user_mss;
6250 tcp_parse_options(skb, &tmp_opt, 0, want_cookie ? NULL : &foc);
6252 if (want_cookie && !tmp_opt.saw_tstamp)
6253 tcp_clear_options(&tmp_opt);
6255 tmp_opt.tstamp_ok = tmp_opt.saw_tstamp;
6256 tcp_openreq_init(req, &tmp_opt, skb, sk);
6258 /* Note: tcp_v6_init_req() might override ir_iif for link locals */
6259 inet_rsk(req)->ir_iif = inet_request_bound_dev_if(sk, skb);
6261 af_ops->init_req(req, sk, skb);
6263 if (security_inet_conn_request(sk, skb, req))
6266 if (!want_cookie && !isn) {
6267 /* VJ's idea. We save last timestamp seen
6268 * from the destination in peer table, when entering
6269 * state TIME-WAIT, and check against it before
6270 * accepting new connection request.
6272 * If "isn" is not zero, this request hit alive
6273 * timewait bucket, so that all the necessary checks
6274 * are made in the function processing timewait state.
6276 if (tcp_death_row.sysctl_tw_recycle) {
6279 dst = af_ops->route_req(sk, &fl, req, &strict);
6281 if (dst && strict &&
6282 !tcp_peer_is_proven(req, dst, true,
6283 tmp_opt.saw_tstamp)) {
6284 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSPASSIVEREJECTED);
6285 goto drop_and_release;
6288 /* Kill the following clause, if you dislike this way. */
6289 else if (!net->ipv4.sysctl_tcp_syncookies &&
6290 (sysctl_max_syn_backlog - inet_csk_reqsk_queue_len(sk) <
6291 (sysctl_max_syn_backlog >> 2)) &&
6292 !tcp_peer_is_proven(req, dst, false,
6293 tmp_opt.saw_tstamp)) {
6294 /* Without syncookies last quarter of
6295 * backlog is filled with destinations,
6296 * proven to be alive.
6297 * It means that we continue to communicate
6298 * to destinations, already remembered
6299 * to the moment of synflood.
6301 pr_drop_req(req, ntohs(tcp_hdr(skb)->source),
6303 goto drop_and_release;
6306 isn = af_ops->init_seq(skb);
6309 dst = af_ops->route_req(sk, &fl, req, NULL);
6314 tcp_ecn_create_request(req, skb, sk, dst);
6317 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss);
6318 req->cookie_ts = tmp_opt.tstamp_ok;
6319 if (!tmp_opt.tstamp_ok)
6320 inet_rsk(req)->ecn_ok = 0;
6323 tcp_rsk(req)->snt_isn = isn;
6324 tcp_rsk(req)->txhash = net_tx_rndhash();
6325 tcp_openreq_init_rwin(req, sk, dst);
6327 tcp_reqsk_record_syn(sk, req, skb);
6328 fastopen_sk = tcp_try_fastopen(sk, skb, req, &foc, dst);
6331 af_ops->send_synack(fastopen_sk, dst, &fl, req,
6332 &foc, TCP_SYNACK_FASTOPEN);
6333 /* Add the child socket directly into the accept queue */
6334 inet_csk_reqsk_queue_add(sk, req, fastopen_sk);
6335 sk->sk_data_ready(sk);
6336 bh_unlock_sock(fastopen_sk);
6337 sock_put(fastopen_sk);
6339 tcp_rsk(req)->tfo_listener = false;
6341 inet_csk_reqsk_queue_hash_add(sk, req, TCP_TIMEOUT_INIT);
6342 af_ops->send_synack(sk, dst, &fl, req, &foc,
6343 !want_cookie ? TCP_SYNACK_NORMAL :
6361 EXPORT_SYMBOL(tcp_conn_request);