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 = 1000;
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(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 /* Sum the number of packets on the wire we have marked as lost.
903 * There are two cases we care about here:
904 * a) Packet hasn't been marked lost (nor retransmitted),
905 * and this is the first loss.
906 * b) Packet has been marked both lost and retransmitted,
907 * and this means we think it was lost again.
909 static void tcp_sum_lost(struct tcp_sock *tp, struct sk_buff *skb)
911 __u8 sacked = TCP_SKB_CB(skb)->sacked;
913 if (!(sacked & TCPCB_LOST) ||
914 ((sacked & TCPCB_LOST) && (sacked & TCPCB_SACKED_RETRANS)))
915 tp->lost += tcp_skb_pcount(skb);
918 static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
920 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
921 tcp_verify_retransmit_hint(tp, skb);
923 tp->lost_out += tcp_skb_pcount(skb);
924 tcp_sum_lost(tp, skb);
925 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
929 void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp, struct sk_buff *skb)
931 tcp_verify_retransmit_hint(tp, skb);
933 tcp_sum_lost(tp, skb);
934 if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
935 tp->lost_out += tcp_skb_pcount(skb);
936 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
940 /* This procedure tags the retransmission queue when SACKs arrive.
942 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
943 * Packets in queue with these bits set are counted in variables
944 * sacked_out, retrans_out and lost_out, correspondingly.
946 * Valid combinations are:
947 * Tag InFlight Description
948 * 0 1 - orig segment is in flight.
949 * S 0 - nothing flies, orig reached receiver.
950 * L 0 - nothing flies, orig lost by net.
951 * R 2 - both orig and retransmit are in flight.
952 * L|R 1 - orig is lost, retransmit is in flight.
953 * S|R 1 - orig reached receiver, retrans is still in flight.
954 * (L|S|R is logically valid, it could occur when L|R is sacked,
955 * but it is equivalent to plain S and code short-curcuits it to S.
956 * L|S is logically invalid, it would mean -1 packet in flight 8))
958 * These 6 states form finite state machine, controlled by the following events:
959 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
960 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
961 * 3. Loss detection event of two flavors:
962 * A. Scoreboard estimator decided the packet is lost.
963 * A'. Reno "three dupacks" marks head of queue lost.
964 * A''. Its FACK modification, head until snd.fack is lost.
965 * B. SACK arrives sacking SND.NXT at the moment, when the
966 * segment was retransmitted.
967 * 4. D-SACK added new rule: D-SACK changes any tag to S.
969 * It is pleasant to note, that state diagram turns out to be commutative,
970 * so that we are allowed not to be bothered by order of our actions,
971 * when multiple events arrive simultaneously. (see the function below).
973 * Reordering detection.
974 * --------------------
975 * Reordering metric is maximal distance, which a packet can be displaced
976 * in packet stream. With SACKs we can estimate it:
978 * 1. SACK fills old hole and the corresponding segment was not
979 * ever retransmitted -> reordering. Alas, we cannot use it
980 * when segment was retransmitted.
981 * 2. The last flaw is solved with D-SACK. D-SACK arrives
982 * for retransmitted and already SACKed segment -> reordering..
983 * Both of these heuristics are not used in Loss state, when we cannot
984 * account for retransmits accurately.
986 * SACK block validation.
987 * ----------------------
989 * SACK block range validation checks that the received SACK block fits to
990 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
991 * Note that SND.UNA is not included to the range though being valid because
992 * it means that the receiver is rather inconsistent with itself reporting
993 * SACK reneging when it should advance SND.UNA. Such SACK block this is
994 * perfectly valid, however, in light of RFC2018 which explicitly states
995 * that "SACK block MUST reflect the newest segment. Even if the newest
996 * segment is going to be discarded ...", not that it looks very clever
997 * in case of head skb. Due to potentional receiver driven attacks, we
998 * choose to avoid immediate execution of a walk in write queue due to
999 * reneging and defer head skb's loss recovery to standard loss recovery
1000 * procedure that will eventually trigger (nothing forbids us doing this).
1002 * Implements also blockage to start_seq wrap-around. Problem lies in the
1003 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
1004 * there's no guarantee that it will be before snd_nxt (n). The problem
1005 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
1008 * <- outs wnd -> <- wrapzone ->
1009 * u e n u_w e_w s n_w
1011 * |<------------+------+----- TCP seqno space --------------+---------->|
1012 * ...-- <2^31 ->| |<--------...
1013 * ...---- >2^31 ------>| |<--------...
1015 * Current code wouldn't be vulnerable but it's better still to discard such
1016 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
1017 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
1018 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
1019 * equal to the ideal case (infinite seqno space without wrap caused issues).
1021 * With D-SACK the lower bound is extended to cover sequence space below
1022 * SND.UNA down to undo_marker, which is the last point of interest. Yet
1023 * again, D-SACK block must not to go across snd_una (for the same reason as
1024 * for the normal SACK blocks, explained above). But there all simplicity
1025 * ends, TCP might receive valid D-SACKs below that. As long as they reside
1026 * fully below undo_marker they do not affect behavior in anyway and can
1027 * therefore be safely ignored. In rare cases (which are more or less
1028 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
1029 * fragmentation and packet reordering past skb's retransmission. To consider
1030 * them correctly, the acceptable range must be extended even more though
1031 * the exact amount is rather hard to quantify. However, tp->max_window can
1032 * be used as an exaggerated estimate.
1034 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
1035 u32 start_seq, u32 end_seq)
1037 /* Too far in future, or reversed (interpretation is ambiguous) */
1038 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
1041 /* Nasty start_seq wrap-around check (see comments above) */
1042 if (!before(start_seq, tp->snd_nxt))
1045 /* In outstanding window? ...This is valid exit for D-SACKs too.
1046 * start_seq == snd_una is non-sensical (see comments above)
1048 if (after(start_seq, tp->snd_una))
1051 if (!is_dsack || !tp->undo_marker)
1054 /* ...Then it's D-SACK, and must reside below snd_una completely */
1055 if (after(end_seq, tp->snd_una))
1058 if (!before(start_seq, tp->undo_marker))
1062 if (!after(end_seq, tp->undo_marker))
1065 /* Undo_marker boundary crossing (overestimates a lot). Known already:
1066 * start_seq < undo_marker and end_seq >= undo_marker.
1068 return !before(start_seq, end_seq - tp->max_window);
1071 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1072 struct tcp_sack_block_wire *sp, int num_sacks,
1075 struct tcp_sock *tp = tcp_sk(sk);
1076 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1077 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1078 bool dup_sack = false;
1080 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1083 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1084 } else if (num_sacks > 1) {
1085 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1086 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1088 if (!after(end_seq_0, end_seq_1) &&
1089 !before(start_seq_0, start_seq_1)) {
1092 NET_INC_STATS(sock_net(sk),
1093 LINUX_MIB_TCPDSACKOFORECV);
1097 /* D-SACK for already forgotten data... Do dumb counting. */
1098 if (dup_sack && tp->undo_marker && tp->undo_retrans > 0 &&
1099 !after(end_seq_0, prior_snd_una) &&
1100 after(end_seq_0, tp->undo_marker))
1106 struct tcp_sacktag_state {
1109 /* Timestamps for earliest and latest never-retransmitted segment
1110 * that was SACKed. RTO needs the earliest RTT to stay conservative,
1111 * but congestion control should still get an accurate delay signal.
1113 struct skb_mstamp first_sackt;
1114 struct skb_mstamp last_sackt;
1118 /* Check if skb is fully within the SACK block. In presence of GSO skbs,
1119 * the incoming SACK may not exactly match but we can find smaller MSS
1120 * aligned portion of it that matches. Therefore we might need to fragment
1121 * which may fail and creates some hassle (caller must handle error case
1124 * FIXME: this could be merged to shift decision code
1126 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1127 u32 start_seq, u32 end_seq)
1131 unsigned int pkt_len;
1134 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1135 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1137 if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1138 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1139 mss = tcp_skb_mss(skb);
1140 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1143 pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1147 pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1152 /* Round if necessary so that SACKs cover only full MSSes
1153 * and/or the remaining small portion (if present)
1155 if (pkt_len > mss) {
1156 unsigned int new_len = (pkt_len / mss) * mss;
1157 if (!in_sack && new_len < pkt_len) {
1159 if (new_len >= skb->len)
1164 err = tcp_fragment(sk, skb, pkt_len, mss, GFP_ATOMIC);
1172 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1173 static u8 tcp_sacktag_one(struct sock *sk,
1174 struct tcp_sacktag_state *state, u8 sacked,
1175 u32 start_seq, u32 end_seq,
1176 int dup_sack, int pcount,
1177 const struct skb_mstamp *xmit_time)
1179 struct tcp_sock *tp = tcp_sk(sk);
1180 int fack_count = state->fack_count;
1182 /* Account D-SACK for retransmitted packet. */
1183 if (dup_sack && (sacked & TCPCB_RETRANS)) {
1184 if (tp->undo_marker && tp->undo_retrans > 0 &&
1185 after(end_seq, tp->undo_marker))
1187 if (sacked & TCPCB_SACKED_ACKED)
1188 state->reord = min(fack_count, state->reord);
1191 /* Nothing to do; acked frame is about to be dropped (was ACKed). */
1192 if (!after(end_seq, tp->snd_una))
1195 if (!(sacked & TCPCB_SACKED_ACKED)) {
1196 tcp_rack_advance(tp, xmit_time, sacked);
1198 if (sacked & TCPCB_SACKED_RETRANS) {
1199 /* If the segment is not tagged as lost,
1200 * we do not clear RETRANS, believing
1201 * that retransmission is still in flight.
1203 if (sacked & TCPCB_LOST) {
1204 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1205 tp->lost_out -= pcount;
1206 tp->retrans_out -= pcount;
1209 if (!(sacked & TCPCB_RETRANS)) {
1210 /* New sack for not retransmitted frame,
1211 * which was in hole. It is reordering.
1213 if (before(start_seq,
1214 tcp_highest_sack_seq(tp)))
1215 state->reord = min(fack_count,
1217 if (!after(end_seq, tp->high_seq))
1218 state->flag |= FLAG_ORIG_SACK_ACKED;
1219 if (state->first_sackt.v64 == 0)
1220 state->first_sackt = *xmit_time;
1221 state->last_sackt = *xmit_time;
1224 if (sacked & TCPCB_LOST) {
1225 sacked &= ~TCPCB_LOST;
1226 tp->lost_out -= pcount;
1230 sacked |= TCPCB_SACKED_ACKED;
1231 state->flag |= FLAG_DATA_SACKED;
1232 tp->sacked_out += pcount;
1233 tp->delivered += pcount; /* Out-of-order packets delivered */
1235 fack_count += pcount;
1237 /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1238 if (!tcp_is_fack(tp) && tp->lost_skb_hint &&
1239 before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1240 tp->lost_cnt_hint += pcount;
1242 if (fack_count > tp->fackets_out)
1243 tp->fackets_out = fack_count;
1246 /* D-SACK. We can detect redundant retransmission in S|R and plain R
1247 * frames and clear it. undo_retrans is decreased above, L|R frames
1248 * are accounted above as well.
1250 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1251 sacked &= ~TCPCB_SACKED_RETRANS;
1252 tp->retrans_out -= pcount;
1258 /* Shift newly-SACKed bytes from this skb to the immediately previous
1259 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1261 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
1262 struct tcp_sacktag_state *state,
1263 unsigned int pcount, int shifted, int mss,
1266 struct tcp_sock *tp = tcp_sk(sk);
1267 struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
1268 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
1269 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
1273 /* Adjust counters and hints for the newly sacked sequence
1274 * range but discard the return value since prev is already
1275 * marked. We must tag the range first because the seq
1276 * advancement below implicitly advances
1277 * tcp_highest_sack_seq() when skb is highest_sack.
1279 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1280 start_seq, end_seq, dup_sack, pcount,
1283 if (skb == tp->lost_skb_hint)
1284 tp->lost_cnt_hint += pcount;
1286 TCP_SKB_CB(prev)->end_seq += shifted;
1287 TCP_SKB_CB(skb)->seq += shifted;
1289 tcp_skb_pcount_add(prev, pcount);
1290 BUG_ON(tcp_skb_pcount(skb) < pcount);
1291 tcp_skb_pcount_add(skb, -pcount);
1293 /* When we're adding to gso_segs == 1, gso_size will be zero,
1294 * in theory this shouldn't be necessary but as long as DSACK
1295 * code can come after this skb later on it's better to keep
1296 * setting gso_size to something.
1298 if (!TCP_SKB_CB(prev)->tcp_gso_size)
1299 TCP_SKB_CB(prev)->tcp_gso_size = mss;
1301 /* CHECKME: To clear or not to clear? Mimics normal skb currently */
1302 if (tcp_skb_pcount(skb) <= 1)
1303 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1305 /* Difference in this won't matter, both ACKed by the same cumul. ACK */
1306 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1309 BUG_ON(!tcp_skb_pcount(skb));
1310 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1314 /* Whole SKB was eaten :-) */
1316 if (skb == tp->retransmit_skb_hint)
1317 tp->retransmit_skb_hint = prev;
1318 if (skb == tp->lost_skb_hint) {
1319 tp->lost_skb_hint = prev;
1320 tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1323 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1324 TCP_SKB_CB(prev)->eor = TCP_SKB_CB(skb)->eor;
1325 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1326 TCP_SKB_CB(prev)->end_seq++;
1328 if (skb == tcp_highest_sack(sk))
1329 tcp_advance_highest_sack(sk, skb);
1331 tcp_skb_collapse_tstamp(prev, skb);
1332 tcp_unlink_write_queue(skb, sk);
1333 sk_wmem_free_skb(sk, skb);
1335 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKMERGED);
1340 /* I wish gso_size would have a bit more sane initialization than
1341 * something-or-zero which complicates things
1343 static int tcp_skb_seglen(const struct sk_buff *skb)
1345 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1348 /* Shifting pages past head area doesn't work */
1349 static int skb_can_shift(const struct sk_buff *skb)
1351 return !skb_headlen(skb) && skb_is_nonlinear(skb);
1354 /* Try collapsing SACK blocks spanning across multiple skbs to a single
1357 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1358 struct tcp_sacktag_state *state,
1359 u32 start_seq, u32 end_seq,
1362 struct tcp_sock *tp = tcp_sk(sk);
1363 struct sk_buff *prev;
1369 if (!sk_can_gso(sk))
1372 /* Normally R but no L won't result in plain S */
1374 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1376 if (!skb_can_shift(skb))
1378 /* This frame is about to be dropped (was ACKed). */
1379 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1382 /* Can only happen with delayed DSACK + discard craziness */
1383 if (unlikely(skb == tcp_write_queue_head(sk)))
1385 prev = tcp_write_queue_prev(sk, skb);
1387 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1390 if (!tcp_skb_can_collapse_to(prev))
1393 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1394 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1398 pcount = tcp_skb_pcount(skb);
1399 mss = tcp_skb_seglen(skb);
1401 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1402 * drop this restriction as unnecessary
1404 if (mss != tcp_skb_seglen(prev))
1407 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1409 /* CHECKME: This is non-MSS split case only?, this will
1410 * cause skipped skbs due to advancing loop btw, original
1411 * has that feature too
1413 if (tcp_skb_pcount(skb) <= 1)
1416 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1418 /* TODO: head merge to next could be attempted here
1419 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1420 * though it might not be worth of the additional hassle
1422 * ...we can probably just fallback to what was done
1423 * previously. We could try merging non-SACKed ones
1424 * as well but it probably isn't going to buy off
1425 * because later SACKs might again split them, and
1426 * it would make skb timestamp tracking considerably
1432 len = end_seq - TCP_SKB_CB(skb)->seq;
1434 BUG_ON(len > skb->len);
1436 /* MSS boundaries should be honoured or else pcount will
1437 * severely break even though it makes things bit trickier.
1438 * Optimize common case to avoid most of the divides
1440 mss = tcp_skb_mss(skb);
1442 /* TODO: Fix DSACKs to not fragment already SACKed and we can
1443 * drop this restriction as unnecessary
1445 if (mss != tcp_skb_seglen(prev))
1450 } else if (len < mss) {
1458 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1459 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1462 if (!skb_shift(prev, skb, len))
1464 if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack))
1467 /* Hole filled allows collapsing with the next as well, this is very
1468 * useful when hole on every nth skb pattern happens
1470 if (prev == tcp_write_queue_tail(sk))
1472 skb = tcp_write_queue_next(sk, prev);
1474 if (!skb_can_shift(skb) ||
1475 (skb == tcp_send_head(sk)) ||
1476 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1477 (mss != tcp_skb_seglen(skb)))
1481 if (skb_shift(prev, skb, len)) {
1482 pcount += tcp_skb_pcount(skb);
1483 tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0);
1487 state->fack_count += pcount;
1494 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1498 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1499 struct tcp_sack_block *next_dup,
1500 struct tcp_sacktag_state *state,
1501 u32 start_seq, u32 end_seq,
1504 struct tcp_sock *tp = tcp_sk(sk);
1505 struct sk_buff *tmp;
1507 tcp_for_write_queue_from(skb, sk) {
1509 bool dup_sack = dup_sack_in;
1511 if (skb == tcp_send_head(sk))
1514 /* queue is in-order => we can short-circuit the walk early */
1515 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1519 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1520 in_sack = tcp_match_skb_to_sack(sk, skb,
1521 next_dup->start_seq,
1527 /* skb reference here is a bit tricky to get right, since
1528 * shifting can eat and free both this skb and the next,
1529 * so not even _safe variant of the loop is enough.
1532 tmp = tcp_shift_skb_data(sk, skb, state,
1533 start_seq, end_seq, dup_sack);
1542 in_sack = tcp_match_skb_to_sack(sk, skb,
1548 if (unlikely(in_sack < 0))
1552 TCP_SKB_CB(skb)->sacked =
1555 TCP_SKB_CB(skb)->sacked,
1556 TCP_SKB_CB(skb)->seq,
1557 TCP_SKB_CB(skb)->end_seq,
1559 tcp_skb_pcount(skb),
1562 if (!before(TCP_SKB_CB(skb)->seq,
1563 tcp_highest_sack_seq(tp)))
1564 tcp_advance_highest_sack(sk, skb);
1567 state->fack_count += tcp_skb_pcount(skb);
1572 /* Avoid all extra work that is being done by sacktag while walking in
1575 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1576 struct tcp_sacktag_state *state,
1579 tcp_for_write_queue_from(skb, sk) {
1580 if (skb == tcp_send_head(sk))
1583 if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1586 state->fack_count += tcp_skb_pcount(skb);
1591 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1593 struct tcp_sack_block *next_dup,
1594 struct tcp_sacktag_state *state,
1600 if (before(next_dup->start_seq, skip_to_seq)) {
1601 skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1602 skb = tcp_sacktag_walk(skb, sk, NULL, state,
1603 next_dup->start_seq, next_dup->end_seq,
1610 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1612 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1616 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1617 u32 prior_snd_una, struct tcp_sacktag_state *state)
1619 struct tcp_sock *tp = tcp_sk(sk);
1620 const unsigned char *ptr = (skb_transport_header(ack_skb) +
1621 TCP_SKB_CB(ack_skb)->sacked);
1622 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1623 struct tcp_sack_block sp[TCP_NUM_SACKS];
1624 struct tcp_sack_block *cache;
1625 struct sk_buff *skb;
1626 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1628 bool found_dup_sack = false;
1630 int first_sack_index;
1633 state->reord = tp->packets_out;
1635 if (!tp->sacked_out) {
1636 if (WARN_ON(tp->fackets_out))
1637 tp->fackets_out = 0;
1638 tcp_highest_sack_reset(sk);
1641 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1642 num_sacks, prior_snd_una);
1644 state->flag |= FLAG_DSACKING_ACK;
1646 /* Eliminate too old ACKs, but take into
1647 * account more or less fresh ones, they can
1648 * contain valid SACK info.
1650 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1653 if (!tp->packets_out)
1657 first_sack_index = 0;
1658 for (i = 0; i < num_sacks; i++) {
1659 bool dup_sack = !i && found_dup_sack;
1661 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1662 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1664 if (!tcp_is_sackblock_valid(tp, dup_sack,
1665 sp[used_sacks].start_seq,
1666 sp[used_sacks].end_seq)) {
1670 if (!tp->undo_marker)
1671 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1673 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1675 /* Don't count olds caused by ACK reordering */
1676 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1677 !after(sp[used_sacks].end_seq, tp->snd_una))
1679 mib_idx = LINUX_MIB_TCPSACKDISCARD;
1682 NET_INC_STATS(sock_net(sk), mib_idx);
1684 first_sack_index = -1;
1688 /* Ignore very old stuff early */
1689 if (!after(sp[used_sacks].end_seq, prior_snd_una))
1695 /* order SACK blocks to allow in order walk of the retrans queue */
1696 for (i = used_sacks - 1; i > 0; i--) {
1697 for (j = 0; j < i; j++) {
1698 if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1699 swap(sp[j], sp[j + 1]);
1701 /* Track where the first SACK block goes to */
1702 if (j == first_sack_index)
1703 first_sack_index = j + 1;
1708 skb = tcp_write_queue_head(sk);
1709 state->fack_count = 0;
1712 if (!tp->sacked_out) {
1713 /* It's already past, so skip checking against it */
1714 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1716 cache = tp->recv_sack_cache;
1717 /* Skip empty blocks in at head of the cache */
1718 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1723 while (i < used_sacks) {
1724 u32 start_seq = sp[i].start_seq;
1725 u32 end_seq = sp[i].end_seq;
1726 bool dup_sack = (found_dup_sack && (i == first_sack_index));
1727 struct tcp_sack_block *next_dup = NULL;
1729 if (found_dup_sack && ((i + 1) == first_sack_index))
1730 next_dup = &sp[i + 1];
1732 /* Skip too early cached blocks */
1733 while (tcp_sack_cache_ok(tp, cache) &&
1734 !before(start_seq, cache->end_seq))
1737 /* Can skip some work by looking recv_sack_cache? */
1738 if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1739 after(end_seq, cache->start_seq)) {
1742 if (before(start_seq, cache->start_seq)) {
1743 skb = tcp_sacktag_skip(skb, sk, state,
1745 skb = tcp_sacktag_walk(skb, sk, next_dup,
1752 /* Rest of the block already fully processed? */
1753 if (!after(end_seq, cache->end_seq))
1756 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1760 /* ...tail remains todo... */
1761 if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1762 /* ...but better entrypoint exists! */
1763 skb = tcp_highest_sack(sk);
1766 state->fack_count = tp->fackets_out;
1771 skb = tcp_sacktag_skip(skb, sk, state, cache->end_seq);
1772 /* Check overlap against next cached too (past this one already) */
1777 if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1778 skb = tcp_highest_sack(sk);
1781 state->fack_count = tp->fackets_out;
1783 skb = tcp_sacktag_skip(skb, sk, state, start_seq);
1786 skb = tcp_sacktag_walk(skb, sk, next_dup, state,
1787 start_seq, end_seq, dup_sack);
1793 /* Clear the head of the cache sack blocks so we can skip it next time */
1794 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1795 tp->recv_sack_cache[i].start_seq = 0;
1796 tp->recv_sack_cache[i].end_seq = 0;
1798 for (j = 0; j < used_sacks; j++)
1799 tp->recv_sack_cache[i++] = sp[j];
1801 if ((state->reord < tp->fackets_out) &&
1802 ((inet_csk(sk)->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker))
1803 tcp_update_reordering(sk, tp->fackets_out - state->reord, 0);
1805 tcp_verify_left_out(tp);
1808 #if FASTRETRANS_DEBUG > 0
1809 WARN_ON((int)tp->sacked_out < 0);
1810 WARN_ON((int)tp->lost_out < 0);
1811 WARN_ON((int)tp->retrans_out < 0);
1812 WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1817 /* Limits sacked_out so that sum with lost_out isn't ever larger than
1818 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1820 static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
1824 holes = max(tp->lost_out, 1U);
1825 holes = min(holes, tp->packets_out);
1827 if ((tp->sacked_out + holes) > tp->packets_out) {
1828 tp->sacked_out = tp->packets_out - holes;
1834 /* If we receive more dupacks than we expected counting segments
1835 * in assumption of absent reordering, interpret this as reordering.
1836 * The only another reason could be bug in receiver TCP.
1838 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1840 struct tcp_sock *tp = tcp_sk(sk);
1841 if (tcp_limit_reno_sacked(tp))
1842 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1845 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1847 static void tcp_add_reno_sack(struct sock *sk)
1849 struct tcp_sock *tp = tcp_sk(sk);
1850 u32 prior_sacked = tp->sacked_out;
1853 tcp_check_reno_reordering(sk, 0);
1854 if (tp->sacked_out > prior_sacked)
1855 tp->delivered++; /* Some out-of-order packet is delivered */
1856 tcp_verify_left_out(tp);
1859 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1861 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1863 struct tcp_sock *tp = tcp_sk(sk);
1866 /* One ACK acked hole. The rest eat duplicate ACKs. */
1867 tp->delivered += max_t(int, acked - tp->sacked_out, 1);
1868 if (acked - 1 >= tp->sacked_out)
1871 tp->sacked_out -= acked - 1;
1873 tcp_check_reno_reordering(sk, acked);
1874 tcp_verify_left_out(tp);
1877 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1882 void tcp_clear_retrans(struct tcp_sock *tp)
1884 tp->retrans_out = 0;
1886 tp->undo_marker = 0;
1887 tp->undo_retrans = -1;
1888 tp->fackets_out = 0;
1892 static inline void tcp_init_undo(struct tcp_sock *tp)
1894 tp->undo_marker = tp->snd_una;
1895 /* Retransmission still in flight may cause DSACKs later. */
1896 tp->undo_retrans = tp->retrans_out ? : -1;
1899 /* Enter Loss state. If we detect SACK reneging, forget all SACK information
1900 * and reset tags completely, otherwise preserve SACKs. If receiver
1901 * dropped its ofo queue, we will know this due to reneging detection.
1903 void tcp_enter_loss(struct sock *sk)
1905 const struct inet_connection_sock *icsk = inet_csk(sk);
1906 struct tcp_sock *tp = tcp_sk(sk);
1907 struct net *net = sock_net(sk);
1908 struct sk_buff *skb;
1909 bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery;
1910 bool is_reneg; /* is receiver reneging on SACKs? */
1913 /* Reduce ssthresh if it has not yet been made inside this window. */
1914 if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
1915 !after(tp->high_seq, tp->snd_una) ||
1916 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1917 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1918 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1919 tcp_ca_event(sk, CA_EVENT_LOSS);
1923 tp->snd_cwnd_cnt = 0;
1924 tp->snd_cwnd_stamp = tcp_time_stamp;
1926 tp->retrans_out = 0;
1929 if (tcp_is_reno(tp))
1930 tcp_reset_reno_sack(tp);
1932 skb = tcp_write_queue_head(sk);
1933 is_reneg = skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED);
1935 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
1937 tp->fackets_out = 0;
1939 tcp_clear_all_retrans_hints(tp);
1941 tcp_for_write_queue(skb, sk) {
1942 if (skb == tcp_send_head(sk))
1945 mark_lost = (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
1948 tcp_sum_lost(tp, skb);
1949 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1951 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1952 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1953 tp->lost_out += tcp_skb_pcount(skb);
1954 tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
1957 tcp_verify_left_out(tp);
1959 /* Timeout in disordered state after receiving substantial DUPACKs
1960 * suggests that the degree of reordering is over-estimated.
1962 if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
1963 tp->sacked_out >= net->ipv4.sysctl_tcp_reordering)
1964 tp->reordering = min_t(unsigned int, tp->reordering,
1965 net->ipv4.sysctl_tcp_reordering);
1966 tcp_set_ca_state(sk, TCP_CA_Loss);
1967 tp->high_seq = tp->snd_nxt;
1968 tcp_ecn_queue_cwr(tp);
1970 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
1971 * loss recovery is underway except recurring timeout(s) on
1972 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
1974 tp->frto = sysctl_tcp_frto &&
1975 (new_recovery || icsk->icsk_retransmits) &&
1976 !inet_csk(sk)->icsk_mtup.probe_size;
1979 /* If ACK arrived pointing to a remembered SACK, it means that our
1980 * remembered SACKs do not reflect real state of receiver i.e.
1981 * receiver _host_ is heavily congested (or buggy).
1983 * To avoid big spurious retransmission bursts due to transient SACK
1984 * scoreboard oddities that look like reneging, we give the receiver a
1985 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will
1986 * restore sanity to the SACK scoreboard. If the apparent reneging
1987 * persists until this RTO then we'll clear the SACK scoreboard.
1989 static bool tcp_check_sack_reneging(struct sock *sk, int flag)
1991 if (flag & FLAG_SACK_RENEGING) {
1992 struct tcp_sock *tp = tcp_sk(sk);
1993 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4),
1994 msecs_to_jiffies(10));
1996 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1997 delay, TCP_RTO_MAX);
2003 static inline int tcp_fackets_out(const struct tcp_sock *tp)
2005 return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
2008 /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2009 * counter when SACK is enabled (without SACK, sacked_out is used for
2012 * Instead, with FACK TCP uses fackets_out that includes both SACKed
2013 * segments up to the highest received SACK block so far and holes in
2016 * With reordering, holes may still be in flight, so RFC3517 recovery
2017 * uses pure sacked_out (total number of SACKed segments) even though
2018 * it violates the RFC that uses duplicate ACKs, often these are equal
2019 * but when e.g. out-of-window ACKs or packet duplication occurs,
2020 * they differ. Since neither occurs due to loss, TCP should really
2023 static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2025 return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
2028 static bool tcp_pause_early_retransmit(struct sock *sk, int flag)
2030 struct tcp_sock *tp = tcp_sk(sk);
2031 unsigned long delay;
2033 /* Delay early retransmit and entering fast recovery for
2034 * max(RTT/4, 2msec) unless ack has ECE mark, no RTT samples
2035 * available, or RTO is scheduled to fire first.
2037 if (sysctl_tcp_early_retrans < 2 || sysctl_tcp_early_retrans > 3 ||
2038 (flag & FLAG_ECE) || !tp->srtt_us)
2041 delay = max(usecs_to_jiffies(tp->srtt_us >> 5),
2042 msecs_to_jiffies(2));
2044 if (!time_after(inet_csk(sk)->icsk_timeout, (jiffies + delay)))
2047 inet_csk_reset_xmit_timer(sk, ICSK_TIME_EARLY_RETRANS, delay,
2052 /* Linux NewReno/SACK/FACK/ECN state machine.
2053 * --------------------------------------
2055 * "Open" Normal state, no dubious events, fast path.
2056 * "Disorder" In all the respects it is "Open",
2057 * but requires a bit more attention. It is entered when
2058 * we see some SACKs or dupacks. It is split of "Open"
2059 * mainly to move some processing from fast path to slow one.
2060 * "CWR" CWND was reduced due to some Congestion Notification event.
2061 * It can be ECN, ICMP source quench, local device congestion.
2062 * "Recovery" CWND was reduced, we are fast-retransmitting.
2063 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
2065 * tcp_fastretrans_alert() is entered:
2066 * - each incoming ACK, if state is not "Open"
2067 * - when arrived ACK is unusual, namely:
2072 * Counting packets in flight is pretty simple.
2074 * in_flight = packets_out - left_out + retrans_out
2076 * packets_out is SND.NXT-SND.UNA counted in packets.
2078 * retrans_out is number of retransmitted segments.
2080 * left_out is number of segments left network, but not ACKed yet.
2082 * left_out = sacked_out + lost_out
2084 * sacked_out: Packets, which arrived to receiver out of order
2085 * and hence not ACKed. With SACKs this number is simply
2086 * amount of SACKed data. Even without SACKs
2087 * it is easy to give pretty reliable estimate of this number,
2088 * counting duplicate ACKs.
2090 * lost_out: Packets lost by network. TCP has no explicit
2091 * "loss notification" feedback from network (for now).
2092 * It means that this number can be only _guessed_.
2093 * Actually, it is the heuristics to predict lossage that
2094 * distinguishes different algorithms.
2096 * F.e. after RTO, when all the queue is considered as lost,
2097 * lost_out = packets_out and in_flight = retrans_out.
2099 * Essentially, we have now two algorithms counting
2102 * FACK: It is the simplest heuristics. As soon as we decided
2103 * that something is lost, we decide that _all_ not SACKed
2104 * packets until the most forward SACK are lost. I.e.
2105 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
2106 * It is absolutely correct estimate, if network does not reorder
2107 * packets. And it loses any connection to reality when reordering
2108 * takes place. We use FACK by default until reordering
2109 * is suspected on the path to this destination.
2111 * NewReno: when Recovery is entered, we assume that one segment
2112 * is lost (classic Reno). While we are in Recovery and
2113 * a partial ACK arrives, we assume that one more packet
2114 * is lost (NewReno). This heuristics are the same in NewReno
2117 * Imagine, that's all! Forget about all this shamanism about CWND inflation
2118 * deflation etc. CWND is real congestion window, never inflated, changes
2119 * only according to classic VJ rules.
2121 * Really tricky (and requiring careful tuning) part of algorithm
2122 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2123 * The first determines the moment _when_ we should reduce CWND and,
2124 * hence, slow down forward transmission. In fact, it determines the moment
2125 * when we decide that hole is caused by loss, rather than by a reorder.
2127 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2128 * holes, caused by lost packets.
2130 * And the most logically complicated part of algorithm is undo
2131 * heuristics. We detect false retransmits due to both too early
2132 * fast retransmit (reordering) and underestimated RTO, analyzing
2133 * timestamps and D-SACKs. When we detect that some segments were
2134 * retransmitted by mistake and CWND reduction was wrong, we undo
2135 * window reduction and abort recovery phase. This logic is hidden
2136 * inside several functions named tcp_try_undo_<something>.
2139 /* This function decides, when we should leave Disordered state
2140 * and enter Recovery phase, reducing congestion window.
2142 * Main question: may we further continue forward transmission
2143 * with the same cwnd?
2145 static bool tcp_time_to_recover(struct sock *sk, int flag)
2147 struct tcp_sock *tp = tcp_sk(sk);
2149 int tcp_reordering = sock_net(sk)->ipv4.sysctl_tcp_reordering;
2151 /* Trick#1: The loss is proven. */
2155 /* Not-A-Trick#2 : Classic rule... */
2156 if (tcp_dupack_heuristics(tp) > tp->reordering)
2159 /* Trick#4: It is still not OK... But will it be useful to delay
2162 packets_out = tp->packets_out;
2163 if (packets_out <= tp->reordering &&
2164 tp->sacked_out >= max_t(__u32, packets_out/2, tcp_reordering) &&
2165 !tcp_may_send_now(sk)) {
2166 /* We have nothing to send. This connection is limited
2167 * either by receiver window or by application.
2172 /* If a thin stream is detected, retransmit after first
2173 * received dupack. Employ only if SACK is supported in order
2174 * to avoid possible corner-case series of spurious retransmissions
2175 * Use only if there are no unsent data.
2177 if ((tp->thin_dupack || sysctl_tcp_thin_dupack) &&
2178 tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 &&
2179 tcp_is_sack(tp) && !tcp_send_head(sk))
2182 /* Trick#6: TCP early retransmit, per RFC5827. To avoid spurious
2183 * retransmissions due to small network reorderings, we implement
2184 * Mitigation A.3 in the RFC and delay the retransmission for a short
2185 * interval if appropriate.
2187 if (tp->do_early_retrans && !tp->retrans_out && tp->sacked_out &&
2188 (tp->packets_out >= (tp->sacked_out + 1) && tp->packets_out < 4) &&
2189 !tcp_may_send_now(sk))
2190 return !tcp_pause_early_retransmit(sk, flag);
2195 /* Detect loss in event "A" above by marking head of queue up as lost.
2196 * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2197 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2198 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2199 * the maximum SACKed segments to pass before reaching this limit.
2201 static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2203 struct tcp_sock *tp = tcp_sk(sk);
2204 struct sk_buff *skb;
2205 int cnt, oldcnt, lost;
2207 /* Use SACK to deduce losses of new sequences sent during recovery */
2208 const u32 loss_high = tcp_is_sack(tp) ? tp->snd_nxt : tp->high_seq;
2210 WARN_ON(packets > tp->packets_out);
2211 if (tp->lost_skb_hint) {
2212 skb = tp->lost_skb_hint;
2213 cnt = tp->lost_cnt_hint;
2214 /* Head already handled? */
2215 if (mark_head && skb != tcp_write_queue_head(sk))
2218 skb = tcp_write_queue_head(sk);
2222 tcp_for_write_queue_from(skb, sk) {
2223 if (skb == tcp_send_head(sk))
2225 /* TODO: do this better */
2226 /* this is not the most efficient way to do this... */
2227 tp->lost_skb_hint = skb;
2228 tp->lost_cnt_hint = cnt;
2230 if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2234 if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2235 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2236 cnt += tcp_skb_pcount(skb);
2238 if (cnt > packets) {
2239 if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) ||
2240 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
2241 (oldcnt >= packets))
2244 mss = tcp_skb_mss(skb);
2245 /* If needed, chop off the prefix to mark as lost. */
2246 lost = (packets - oldcnt) * mss;
2247 if (lost < skb->len &&
2248 tcp_fragment(sk, skb, lost, mss, GFP_ATOMIC) < 0)
2253 tcp_skb_mark_lost(tp, skb);
2258 tcp_verify_left_out(tp);
2261 /* Account newly detected lost packet(s) */
2263 static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2265 struct tcp_sock *tp = tcp_sk(sk);
2267 if (tcp_is_reno(tp)) {
2268 tcp_mark_head_lost(sk, 1, 1);
2269 } else if (tcp_is_fack(tp)) {
2270 int lost = tp->fackets_out - tp->reordering;
2273 tcp_mark_head_lost(sk, lost, 0);
2275 int sacked_upto = tp->sacked_out - tp->reordering;
2276 if (sacked_upto >= 0)
2277 tcp_mark_head_lost(sk, sacked_upto, 0);
2278 else if (fast_rexmit)
2279 tcp_mark_head_lost(sk, 1, 1);
2283 static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when)
2285 return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2286 before(tp->rx_opt.rcv_tsecr, when);
2289 /* skb is spurious retransmitted if the returned timestamp echo
2290 * reply is prior to the skb transmission time
2292 static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp,
2293 const struct sk_buff *skb)
2295 return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) &&
2296 tcp_tsopt_ecr_before(tp, tcp_skb_timestamp(skb));
2299 /* Nothing was retransmitted or returned timestamp is less
2300 * than timestamp of the first retransmission.
2302 static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2304 return !tp->retrans_stamp ||
2305 tcp_tsopt_ecr_before(tp, tp->retrans_stamp);
2308 /* Undo procedures. */
2310 /* We can clear retrans_stamp when there are no retransmissions in the
2311 * window. It would seem that it is trivially available for us in
2312 * tp->retrans_out, however, that kind of assumptions doesn't consider
2313 * what will happen if errors occur when sending retransmission for the
2314 * second time. ...It could the that such segment has only
2315 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2316 * the head skb is enough except for some reneging corner cases that
2317 * are not worth the effort.
2319 * Main reason for all this complexity is the fact that connection dying
2320 * time now depends on the validity of the retrans_stamp, in particular,
2321 * that successive retransmissions of a segment must not advance
2322 * retrans_stamp under any conditions.
2324 static bool tcp_any_retrans_done(const struct sock *sk)
2326 const struct tcp_sock *tp = tcp_sk(sk);
2327 struct sk_buff *skb;
2329 if (tp->retrans_out)
2332 skb = tcp_write_queue_head(sk);
2333 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2339 #if FASTRETRANS_DEBUG > 1
2340 static void DBGUNDO(struct sock *sk, const char *msg)
2342 struct tcp_sock *tp = tcp_sk(sk);
2343 struct inet_sock *inet = inet_sk(sk);
2345 if (sk->sk_family == AF_INET) {
2346 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2348 &inet->inet_daddr, ntohs(inet->inet_dport),
2349 tp->snd_cwnd, tcp_left_out(tp),
2350 tp->snd_ssthresh, tp->prior_ssthresh,
2353 #if IS_ENABLED(CONFIG_IPV6)
2354 else if (sk->sk_family == AF_INET6) {
2355 struct ipv6_pinfo *np = inet6_sk(sk);
2356 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2358 &np->daddr, ntohs(inet->inet_dport),
2359 tp->snd_cwnd, tcp_left_out(tp),
2360 tp->snd_ssthresh, tp->prior_ssthresh,
2366 #define DBGUNDO(x...) do { } while (0)
2369 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
2371 struct tcp_sock *tp = tcp_sk(sk);
2374 struct sk_buff *skb;
2376 tcp_for_write_queue(skb, sk) {
2377 if (skb == tcp_send_head(sk))
2379 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2382 tcp_clear_all_retrans_hints(tp);
2385 if (tp->prior_ssthresh) {
2386 const struct inet_connection_sock *icsk = inet_csk(sk);
2388 if (icsk->icsk_ca_ops->undo_cwnd)
2389 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2391 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
2393 if (tp->prior_ssthresh > tp->snd_ssthresh) {
2394 tp->snd_ssthresh = tp->prior_ssthresh;
2395 tcp_ecn_withdraw_cwr(tp);
2398 tp->snd_cwnd_stamp = tcp_time_stamp;
2399 tp->undo_marker = 0;
2402 static inline bool tcp_may_undo(const struct tcp_sock *tp)
2404 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2407 /* People celebrate: "We love our President!" */
2408 static bool tcp_try_undo_recovery(struct sock *sk)
2410 struct tcp_sock *tp = tcp_sk(sk);
2412 if (tcp_may_undo(tp)) {
2415 /* Happy end! We did not retransmit anything
2416 * or our original transmission succeeded.
2418 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2419 tcp_undo_cwnd_reduction(sk, false);
2420 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2421 mib_idx = LINUX_MIB_TCPLOSSUNDO;
2423 mib_idx = LINUX_MIB_TCPFULLUNDO;
2425 NET_INC_STATS(sock_net(sk), mib_idx);
2427 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2428 /* Hold old state until something *above* high_seq
2429 * is ACKed. For Reno it is MUST to prevent false
2430 * fast retransmits (RFC2582). SACK TCP is safe. */
2431 if (!tcp_any_retrans_done(sk))
2432 tp->retrans_stamp = 0;
2435 tcp_set_ca_state(sk, TCP_CA_Open);
2439 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2440 static bool tcp_try_undo_dsack(struct sock *sk)
2442 struct tcp_sock *tp = tcp_sk(sk);
2444 if (tp->undo_marker && !tp->undo_retrans) {
2445 DBGUNDO(sk, "D-SACK");
2446 tcp_undo_cwnd_reduction(sk, false);
2447 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2453 /* Undo during loss recovery after partial ACK or using F-RTO. */
2454 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2456 struct tcp_sock *tp = tcp_sk(sk);
2458 if (frto_undo || tcp_may_undo(tp)) {
2459 tcp_undo_cwnd_reduction(sk, true);
2461 DBGUNDO(sk, "partial loss");
2462 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2464 NET_INC_STATS(sock_net(sk),
2465 LINUX_MIB_TCPSPURIOUSRTOS);
2466 inet_csk(sk)->icsk_retransmits = 0;
2467 if (frto_undo || tcp_is_sack(tp))
2468 tcp_set_ca_state(sk, TCP_CA_Open);
2474 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937.
2475 * It computes the number of packets to send (sndcnt) based on packets newly
2477 * 1) If the packets in flight is larger than ssthresh, PRR spreads the
2478 * cwnd reductions across a full RTT.
2479 * 2) Otherwise PRR uses packet conservation to send as much as delivered.
2480 * But when the retransmits are acked without further losses, PRR
2481 * slow starts cwnd up to ssthresh to speed up the recovery.
2483 static void tcp_init_cwnd_reduction(struct sock *sk)
2485 struct tcp_sock *tp = tcp_sk(sk);
2487 tp->high_seq = tp->snd_nxt;
2488 tp->tlp_high_seq = 0;
2489 tp->snd_cwnd_cnt = 0;
2490 tp->prior_cwnd = tp->snd_cwnd;
2491 tp->prr_delivered = 0;
2493 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2494 tcp_ecn_queue_cwr(tp);
2497 static void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked,
2500 struct tcp_sock *tp = tcp_sk(sk);
2502 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2504 if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd))
2507 tp->prr_delivered += newly_acked_sacked;
2509 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2511 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2512 } else if ((flag & FLAG_RETRANS_DATA_ACKED) &&
2513 !(flag & FLAG_LOST_RETRANS)) {
2514 sndcnt = min_t(int, delta,
2515 max_t(int, tp->prr_delivered - tp->prr_out,
2516 newly_acked_sacked) + 1);
2518 sndcnt = min(delta, newly_acked_sacked);
2520 /* Force a fast retransmit upon entering fast recovery */
2521 sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1));
2522 tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
2525 static inline void tcp_end_cwnd_reduction(struct sock *sk)
2527 struct tcp_sock *tp = tcp_sk(sk);
2529 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2530 if (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR ||
2531 (tp->undo_marker && tp->snd_ssthresh < TCP_INFINITE_SSTHRESH)) {
2532 tp->snd_cwnd = tp->snd_ssthresh;
2533 tp->snd_cwnd_stamp = tcp_time_stamp;
2535 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2538 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2539 void tcp_enter_cwr(struct sock *sk)
2541 struct tcp_sock *tp = tcp_sk(sk);
2543 tp->prior_ssthresh = 0;
2544 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2545 tp->undo_marker = 0;
2546 tcp_init_cwnd_reduction(sk);
2547 tcp_set_ca_state(sk, TCP_CA_CWR);
2550 EXPORT_SYMBOL(tcp_enter_cwr);
2552 static void tcp_try_keep_open(struct sock *sk)
2554 struct tcp_sock *tp = tcp_sk(sk);
2555 int state = TCP_CA_Open;
2557 if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2558 state = TCP_CA_Disorder;
2560 if (inet_csk(sk)->icsk_ca_state != state) {
2561 tcp_set_ca_state(sk, state);
2562 tp->high_seq = tp->snd_nxt;
2566 static void tcp_try_to_open(struct sock *sk, int flag)
2568 struct tcp_sock *tp = tcp_sk(sk);
2570 tcp_verify_left_out(tp);
2572 if (!tcp_any_retrans_done(sk))
2573 tp->retrans_stamp = 0;
2575 if (flag & FLAG_ECE)
2578 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2579 tcp_try_keep_open(sk);
2583 static void tcp_mtup_probe_failed(struct sock *sk)
2585 struct inet_connection_sock *icsk = inet_csk(sk);
2587 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2588 icsk->icsk_mtup.probe_size = 0;
2589 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPFAIL);
2592 static void tcp_mtup_probe_success(struct sock *sk)
2594 struct tcp_sock *tp = tcp_sk(sk);
2595 struct inet_connection_sock *icsk = inet_csk(sk);
2597 /* FIXME: breaks with very large cwnd */
2598 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2599 tp->snd_cwnd = tp->snd_cwnd *
2600 tcp_mss_to_mtu(sk, tp->mss_cache) /
2601 icsk->icsk_mtup.probe_size;
2602 tp->snd_cwnd_cnt = 0;
2603 tp->snd_cwnd_stamp = tcp_time_stamp;
2604 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2606 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2607 icsk->icsk_mtup.probe_size = 0;
2608 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2609 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS);
2612 /* Do a simple retransmit without using the backoff mechanisms in
2613 * tcp_timer. This is used for path mtu discovery.
2614 * The socket is already locked here.
2616 void tcp_simple_retransmit(struct sock *sk)
2618 const struct inet_connection_sock *icsk = inet_csk(sk);
2619 struct tcp_sock *tp = tcp_sk(sk);
2620 struct sk_buff *skb;
2621 unsigned int mss = tcp_current_mss(sk);
2622 u32 prior_lost = tp->lost_out;
2624 tcp_for_write_queue(skb, sk) {
2625 if (skb == tcp_send_head(sk))
2627 if (tcp_skb_seglen(skb) > mss &&
2628 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2629 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2630 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2631 tp->retrans_out -= tcp_skb_pcount(skb);
2633 tcp_skb_mark_lost_uncond_verify(tp, skb);
2637 tcp_clear_retrans_hints_partial(tp);
2639 if (prior_lost == tp->lost_out)
2642 if (tcp_is_reno(tp))
2643 tcp_limit_reno_sacked(tp);
2645 tcp_verify_left_out(tp);
2647 /* Don't muck with the congestion window here.
2648 * Reason is that we do not increase amount of _data_
2649 * in network, but units changed and effective
2650 * cwnd/ssthresh really reduced now.
2652 if (icsk->icsk_ca_state != TCP_CA_Loss) {
2653 tp->high_seq = tp->snd_nxt;
2654 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2655 tp->prior_ssthresh = 0;
2656 tp->undo_marker = 0;
2657 tcp_set_ca_state(sk, TCP_CA_Loss);
2659 tcp_xmit_retransmit_queue(sk);
2661 EXPORT_SYMBOL(tcp_simple_retransmit);
2663 static void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2665 struct tcp_sock *tp = tcp_sk(sk);
2668 if (tcp_is_reno(tp))
2669 mib_idx = LINUX_MIB_TCPRENORECOVERY;
2671 mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2673 NET_INC_STATS(sock_net(sk), mib_idx);
2675 tp->prior_ssthresh = 0;
2678 if (!tcp_in_cwnd_reduction(sk)) {
2680 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2681 tcp_init_cwnd_reduction(sk);
2683 tcp_set_ca_state(sk, TCP_CA_Recovery);
2686 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2687 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2689 static void tcp_process_loss(struct sock *sk, int flag, bool is_dupack,
2692 struct tcp_sock *tp = tcp_sk(sk);
2693 bool recovered = !before(tp->snd_una, tp->high_seq);
2695 if ((flag & FLAG_SND_UNA_ADVANCED) &&
2696 tcp_try_undo_loss(sk, false))
2699 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2700 /* Step 3.b. A timeout is spurious if not all data are
2701 * lost, i.e., never-retransmitted data are (s)acked.
2703 if ((flag & FLAG_ORIG_SACK_ACKED) &&
2704 tcp_try_undo_loss(sk, true))
2707 if (after(tp->snd_nxt, tp->high_seq)) {
2708 if (flag & FLAG_DATA_SACKED || is_dupack)
2709 tp->frto = 0; /* Step 3.a. loss was real */
2710 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2711 tp->high_seq = tp->snd_nxt;
2712 /* Step 2.b. Try send new data (but deferred until cwnd
2713 * is updated in tcp_ack()). Otherwise fall back to
2714 * the conventional recovery.
2716 if (tcp_send_head(sk) &&
2717 after(tcp_wnd_end(tp), tp->snd_nxt)) {
2718 *rexmit = REXMIT_NEW;
2726 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2727 tcp_try_undo_recovery(sk);
2730 if (tcp_is_reno(tp)) {
2731 /* A Reno DUPACK means new data in F-RTO step 2.b above are
2732 * delivered. Lower inflight to clock out (re)tranmissions.
2734 if (after(tp->snd_nxt, tp->high_seq) && is_dupack)
2735 tcp_add_reno_sack(sk);
2736 else if (flag & FLAG_SND_UNA_ADVANCED)
2737 tcp_reset_reno_sack(tp);
2739 *rexmit = REXMIT_LOST;
2742 /* Undo during fast recovery after partial ACK. */
2743 static bool tcp_try_undo_partial(struct sock *sk, const int acked)
2745 struct tcp_sock *tp = tcp_sk(sk);
2747 if (tp->undo_marker && tcp_packet_delayed(tp)) {
2748 /* Plain luck! Hole if filled with delayed
2749 * packet, rather than with a retransmit.
2751 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2753 /* We are getting evidence that the reordering degree is higher
2754 * than we realized. If there are no retransmits out then we
2755 * can undo. Otherwise we clock out new packets but do not
2756 * mark more packets lost or retransmit more.
2758 if (tp->retrans_out)
2761 if (!tcp_any_retrans_done(sk))
2762 tp->retrans_stamp = 0;
2764 DBGUNDO(sk, "partial recovery");
2765 tcp_undo_cwnd_reduction(sk, true);
2766 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2767 tcp_try_keep_open(sk);
2773 /* Process an event, which can update packets-in-flight not trivially.
2774 * Main goal of this function is to calculate new estimate for left_out,
2775 * taking into account both packets sitting in receiver's buffer and
2776 * packets lost by network.
2778 * Besides that it updates the congestion state when packet loss or ECN
2779 * is detected. But it does not reduce the cwnd, it is done by the
2780 * congestion control later.
2782 * It does _not_ decide what to send, it is made in function
2783 * tcp_xmit_retransmit_queue().
2785 static void tcp_fastretrans_alert(struct sock *sk, const int acked,
2786 bool is_dupack, int *ack_flag, int *rexmit)
2788 struct inet_connection_sock *icsk = inet_csk(sk);
2789 struct tcp_sock *tp = tcp_sk(sk);
2790 int fast_rexmit = 0, flag = *ack_flag;
2791 bool do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2792 (tcp_fackets_out(tp) > tp->reordering));
2794 if (WARN_ON(!tp->packets_out && tp->sacked_out))
2796 if (WARN_ON(!tp->sacked_out && tp->fackets_out))
2797 tp->fackets_out = 0;
2799 /* Now state machine starts.
2800 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2801 if (flag & FLAG_ECE)
2802 tp->prior_ssthresh = 0;
2804 /* B. In all the states check for reneging SACKs. */
2805 if (tcp_check_sack_reneging(sk, flag))
2808 /* C. Check consistency of the current state. */
2809 tcp_verify_left_out(tp);
2811 /* D. Check state exit conditions. State can be terminated
2812 * when high_seq is ACKed. */
2813 if (icsk->icsk_ca_state == TCP_CA_Open) {
2814 WARN_ON(tp->retrans_out != 0);
2815 tp->retrans_stamp = 0;
2816 } else if (!before(tp->snd_una, tp->high_seq)) {
2817 switch (icsk->icsk_ca_state) {
2819 /* CWR is to be held something *above* high_seq
2820 * is ACKed for CWR bit to reach receiver. */
2821 if (tp->snd_una != tp->high_seq) {
2822 tcp_end_cwnd_reduction(sk);
2823 tcp_set_ca_state(sk, TCP_CA_Open);
2827 case TCP_CA_Recovery:
2828 if (tcp_is_reno(tp))
2829 tcp_reset_reno_sack(tp);
2830 if (tcp_try_undo_recovery(sk))
2832 tcp_end_cwnd_reduction(sk);
2837 /* Use RACK to detect loss */
2838 if (sysctl_tcp_recovery & TCP_RACK_LOST_RETRANS &&
2839 tcp_rack_mark_lost(sk)) {
2840 flag |= FLAG_LOST_RETRANS;
2841 *ack_flag |= FLAG_LOST_RETRANS;
2844 /* E. Process state. */
2845 switch (icsk->icsk_ca_state) {
2846 case TCP_CA_Recovery:
2847 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2848 if (tcp_is_reno(tp) && is_dupack)
2849 tcp_add_reno_sack(sk);
2851 if (tcp_try_undo_partial(sk, acked))
2853 /* Partial ACK arrived. Force fast retransmit. */
2854 do_lost = tcp_is_reno(tp) ||
2855 tcp_fackets_out(tp) > tp->reordering;
2857 if (tcp_try_undo_dsack(sk)) {
2858 tcp_try_keep_open(sk);
2863 tcp_process_loss(sk, flag, is_dupack, rexmit);
2864 if (icsk->icsk_ca_state != TCP_CA_Open &&
2865 !(flag & FLAG_LOST_RETRANS))
2867 /* Change state if cwnd is undone or retransmits are lost */
2869 if (tcp_is_reno(tp)) {
2870 if (flag & FLAG_SND_UNA_ADVANCED)
2871 tcp_reset_reno_sack(tp);
2873 tcp_add_reno_sack(sk);
2876 if (icsk->icsk_ca_state <= TCP_CA_Disorder)
2877 tcp_try_undo_dsack(sk);
2879 if (!tcp_time_to_recover(sk, flag)) {
2880 tcp_try_to_open(sk, flag);
2884 /* MTU probe failure: don't reduce cwnd */
2885 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2886 icsk->icsk_mtup.probe_size &&
2887 tp->snd_una == tp->mtu_probe.probe_seq_start) {
2888 tcp_mtup_probe_failed(sk);
2889 /* Restores the reduction we did in tcp_mtup_probe() */
2891 tcp_simple_retransmit(sk);
2895 /* Otherwise enter Recovery state */
2896 tcp_enter_recovery(sk, (flag & FLAG_ECE));
2901 tcp_update_scoreboard(sk, fast_rexmit);
2902 *rexmit = REXMIT_LOST;
2905 static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us)
2907 struct tcp_sock *tp = tcp_sk(sk);
2908 u32 wlen = sysctl_tcp_min_rtt_wlen * HZ;
2910 minmax_running_min(&tp->rtt_min, wlen, tcp_time_stamp,
2911 rtt_us ? : jiffies_to_usecs(1));
2914 static inline bool tcp_ack_update_rtt(struct sock *sk, const int flag,
2915 long seq_rtt_us, long sack_rtt_us,
2918 const struct tcp_sock *tp = tcp_sk(sk);
2920 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2921 * broken middle-boxes or peers may corrupt TS-ECR fields. But
2922 * Karn's algorithm forbids taking RTT if some retransmitted data
2923 * is acked (RFC6298).
2926 seq_rtt_us = sack_rtt_us;
2928 /* RTTM Rule: A TSecr value received in a segment is used to
2929 * update the averaged RTT measurement only if the segment
2930 * acknowledges some new data, i.e., only if it advances the
2931 * left edge of the send window.
2932 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2934 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2936 seq_rtt_us = ca_rtt_us = jiffies_to_usecs(tcp_time_stamp -
2937 tp->rx_opt.rcv_tsecr);
2941 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is
2942 * always taken together with ACK, SACK, or TS-opts. Any negative
2943 * values will be skipped with the seq_rtt_us < 0 check above.
2945 tcp_update_rtt_min(sk, ca_rtt_us);
2946 tcp_rtt_estimator(sk, seq_rtt_us);
2949 /* RFC6298: only reset backoff on valid RTT measurement. */
2950 inet_csk(sk)->icsk_backoff = 0;
2954 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
2955 void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req)
2959 if (req && !req->num_retrans && tcp_rsk(req)->snt_synack.v64) {
2960 struct skb_mstamp now;
2962 skb_mstamp_get(&now);
2963 rtt_us = skb_mstamp_us_delta(&now, &tcp_rsk(req)->snt_synack);
2966 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us);
2970 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
2972 const struct inet_connection_sock *icsk = inet_csk(sk);
2974 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);
2975 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
2978 /* Restart timer after forward progress on connection.
2979 * RFC2988 recommends to restart timer to now+rto.
2981 void tcp_rearm_rto(struct sock *sk)
2983 const struct inet_connection_sock *icsk = inet_csk(sk);
2984 struct tcp_sock *tp = tcp_sk(sk);
2986 /* If the retrans timer is currently being used by Fast Open
2987 * for SYN-ACK retrans purpose, stay put.
2989 if (tp->fastopen_rsk)
2992 if (!tp->packets_out) {
2993 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
2995 u32 rto = inet_csk(sk)->icsk_rto;
2996 /* Offset the time elapsed after installing regular RTO */
2997 if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
2998 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
2999 struct sk_buff *skb = tcp_write_queue_head(sk);
3000 const u32 rto_time_stamp =
3001 tcp_skb_timestamp(skb) + rto;
3002 s32 delta = (s32)(rto_time_stamp - tcp_time_stamp);
3003 /* delta may not be positive if the socket is locked
3004 * when the retrans timer fires and is rescheduled.
3009 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
3014 /* This function is called when the delayed ER timer fires. TCP enters
3015 * fast recovery and performs fast-retransmit.
3017 void tcp_resume_early_retransmit(struct sock *sk)
3019 struct tcp_sock *tp = tcp_sk(sk);
3023 /* Stop if ER is disabled after the delayed ER timer is scheduled */
3024 if (!tp->do_early_retrans)
3027 tcp_enter_recovery(sk, false);
3028 tcp_update_scoreboard(sk, 1);
3029 tcp_xmit_retransmit_queue(sk);
3032 /* If we get here, the whole TSO packet has not been acked. */
3033 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3035 struct tcp_sock *tp = tcp_sk(sk);
3038 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3040 packets_acked = tcp_skb_pcount(skb);
3041 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3043 packets_acked -= tcp_skb_pcount(skb);
3045 if (packets_acked) {
3046 BUG_ON(tcp_skb_pcount(skb) == 0);
3047 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3050 return packets_acked;
3053 static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
3056 const struct skb_shared_info *shinfo;
3058 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */
3059 if (likely(!TCP_SKB_CB(skb)->txstamp_ack))
3062 shinfo = skb_shinfo(skb);
3063 if (!before(shinfo->tskey, prior_snd_una) &&
3064 before(shinfo->tskey, tcp_sk(sk)->snd_una))
3065 __skb_tstamp_tx(skb, NULL, sk, SCM_TSTAMP_ACK);
3068 /* Remove acknowledged frames from the retransmission queue. If our packet
3069 * is before the ack sequence we can discard it as it's confirmed to have
3070 * arrived at the other end.
3072 static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
3073 u32 prior_snd_una, int *acked,
3074 struct tcp_sacktag_state *sack)
3076 const struct inet_connection_sock *icsk = inet_csk(sk);
3077 struct skb_mstamp first_ackt, last_ackt, now;
3078 struct tcp_sock *tp = tcp_sk(sk);
3079 u32 prior_sacked = tp->sacked_out;
3080 u32 reord = tp->packets_out;
3081 bool fully_acked = true;
3082 long sack_rtt_us = -1L;
3083 long seq_rtt_us = -1L;
3084 long ca_rtt_us = -1L;
3085 struct sk_buff *skb;
3087 u32 last_in_flight = 0;
3093 while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
3094 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
3095 u8 sacked = scb->sacked;
3098 tcp_ack_tstamp(sk, skb, prior_snd_una);
3100 /* Determine how many packets and what bytes were acked, tso and else */
3101 if (after(scb->end_seq, tp->snd_una)) {
3102 if (tcp_skb_pcount(skb) == 1 ||
3103 !after(tp->snd_una, scb->seq))
3106 acked_pcount = tcp_tso_acked(sk, skb);
3110 fully_acked = false;
3112 /* Speedup tcp_unlink_write_queue() and next loop */
3113 prefetchw(skb->next);
3114 acked_pcount = tcp_skb_pcount(skb);
3117 if (unlikely(sacked & TCPCB_RETRANS)) {
3118 if (sacked & TCPCB_SACKED_RETRANS)
3119 tp->retrans_out -= acked_pcount;
3120 flag |= FLAG_RETRANS_DATA_ACKED;
3121 } else if (!(sacked & TCPCB_SACKED_ACKED)) {
3122 last_ackt = skb->skb_mstamp;
3123 WARN_ON_ONCE(last_ackt.v64 == 0);
3124 if (!first_ackt.v64)
3125 first_ackt = last_ackt;
3127 last_in_flight = TCP_SKB_CB(skb)->tx.in_flight;
3128 reord = min(pkts_acked, reord);
3129 if (!after(scb->end_seq, tp->high_seq))
3130 flag |= FLAG_ORIG_SACK_ACKED;
3133 if (sacked & TCPCB_SACKED_ACKED) {
3134 tp->sacked_out -= acked_pcount;
3135 } else if (tcp_is_sack(tp)) {
3136 tp->delivered += acked_pcount;
3137 if (!tcp_skb_spurious_retrans(tp, skb))
3138 tcp_rack_advance(tp, &skb->skb_mstamp, sacked);
3140 if (sacked & TCPCB_LOST)
3141 tp->lost_out -= acked_pcount;
3143 tp->packets_out -= acked_pcount;
3144 pkts_acked += acked_pcount;
3146 /* Initial outgoing SYN's get put onto the write_queue
3147 * just like anything else we transmit. It is not
3148 * true data, and if we misinform our callers that
3149 * this ACK acks real data, we will erroneously exit
3150 * connection startup slow start one packet too
3151 * quickly. This is severely frowned upon behavior.
3153 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) {
3154 flag |= FLAG_DATA_ACKED;
3156 flag |= FLAG_SYN_ACKED;
3157 tp->retrans_stamp = 0;
3163 tcp_unlink_write_queue(skb, sk);
3164 sk_wmem_free_skb(sk, skb);
3165 if (unlikely(skb == tp->retransmit_skb_hint))
3166 tp->retransmit_skb_hint = NULL;
3167 if (unlikely(skb == tp->lost_skb_hint))
3168 tp->lost_skb_hint = NULL;
3171 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3172 tp->snd_up = tp->snd_una;
3174 if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3175 flag |= FLAG_SACK_RENEGING;
3177 skb_mstamp_get(&now);
3178 if (likely(first_ackt.v64) && !(flag & FLAG_RETRANS_DATA_ACKED)) {
3179 seq_rtt_us = skb_mstamp_us_delta(&now, &first_ackt);
3180 ca_rtt_us = skb_mstamp_us_delta(&now, &last_ackt);
3182 if (sack->first_sackt.v64) {
3183 sack_rtt_us = skb_mstamp_us_delta(&now, &sack->first_sackt);
3184 ca_rtt_us = skb_mstamp_us_delta(&now, &sack->last_sackt);
3187 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us,
3190 if (flag & FLAG_ACKED) {
3192 if (unlikely(icsk->icsk_mtup.probe_size &&
3193 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3194 tcp_mtup_probe_success(sk);
3197 if (tcp_is_reno(tp)) {
3198 tcp_remove_reno_sacks(sk, pkts_acked);
3202 /* Non-retransmitted hole got filled? That's reordering */
3203 if (reord < prior_fackets)
3204 tcp_update_reordering(sk, tp->fackets_out - reord, 0);
3206 delta = tcp_is_fack(tp) ? pkts_acked :
3207 prior_sacked - tp->sacked_out;
3208 tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3211 tp->fackets_out -= min(pkts_acked, tp->fackets_out);
3213 } else if (skb && rtt_update && sack_rtt_us >= 0 &&
3214 sack_rtt_us > skb_mstamp_us_delta(&now, &skb->skb_mstamp)) {
3215 /* Do not re-arm RTO if the sack RTT is measured from data sent
3216 * after when the head was last (re)transmitted. Otherwise the
3217 * timeout may continue to extend in loss recovery.
3222 if (icsk->icsk_ca_ops->pkts_acked) {
3223 struct ack_sample sample = { .pkts_acked = pkts_acked,
3224 .rtt_us = ca_rtt_us,
3225 .in_flight = last_in_flight };
3227 icsk->icsk_ca_ops->pkts_acked(sk, &sample);
3230 #if FASTRETRANS_DEBUG > 0
3231 WARN_ON((int)tp->sacked_out < 0);
3232 WARN_ON((int)tp->lost_out < 0);
3233 WARN_ON((int)tp->retrans_out < 0);
3234 if (!tp->packets_out && tcp_is_sack(tp)) {
3235 icsk = inet_csk(sk);
3237 pr_debug("Leak l=%u %d\n",
3238 tp->lost_out, icsk->icsk_ca_state);
3241 if (tp->sacked_out) {
3242 pr_debug("Leak s=%u %d\n",
3243 tp->sacked_out, icsk->icsk_ca_state);
3246 if (tp->retrans_out) {
3247 pr_debug("Leak r=%u %d\n",
3248 tp->retrans_out, icsk->icsk_ca_state);
3249 tp->retrans_out = 0;
3253 *acked = pkts_acked;
3257 static void tcp_ack_probe(struct sock *sk)
3259 const struct tcp_sock *tp = tcp_sk(sk);
3260 struct inet_connection_sock *icsk = inet_csk(sk);
3262 /* Was it a usable window open? */
3264 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
3265 icsk->icsk_backoff = 0;
3266 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3267 /* Socket must be waked up by subsequent tcp_data_snd_check().
3268 * This function is not for random using!
3271 unsigned long when = tcp_probe0_when(sk, TCP_RTO_MAX);
3273 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3278 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3280 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3281 inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3284 /* Decide wheather to run the increase function of congestion control. */
3285 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3287 /* If reordering is high then always grow cwnd whenever data is
3288 * delivered regardless of its ordering. Otherwise stay conservative
3289 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3290 * new SACK or ECE mark may first advance cwnd here and later reduce
3291 * cwnd in tcp_fastretrans_alert() based on more states.
3293 if (tcp_sk(sk)->reordering > sock_net(sk)->ipv4.sysctl_tcp_reordering)
3294 return flag & FLAG_FORWARD_PROGRESS;
3296 return flag & FLAG_DATA_ACKED;
3299 /* The "ultimate" congestion control function that aims to replace the rigid
3300 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction).
3301 * It's called toward the end of processing an ACK with precise rate
3302 * information. All transmission or retransmission are delayed afterwards.
3304 static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked,
3307 if (tcp_in_cwnd_reduction(sk)) {
3308 /* Reduce cwnd if state mandates */
3309 tcp_cwnd_reduction(sk, acked_sacked, flag);
3310 } else if (tcp_may_raise_cwnd(sk, flag)) {
3311 /* Advance cwnd if state allows */
3312 tcp_cong_avoid(sk, ack, acked_sacked);
3314 tcp_update_pacing_rate(sk);
3317 /* Check that window update is acceptable.
3318 * The function assumes that snd_una<=ack<=snd_next.
3320 static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3321 const u32 ack, const u32 ack_seq,
3324 return after(ack, tp->snd_una) ||
3325 after(ack_seq, tp->snd_wl1) ||
3326 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
3329 /* If we update tp->snd_una, also update tp->bytes_acked */
3330 static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack)
3332 u32 delta = ack - tp->snd_una;
3334 sock_owned_by_me((struct sock *)tp);
3335 u64_stats_update_begin_raw(&tp->syncp);
3336 tp->bytes_acked += delta;
3337 u64_stats_update_end_raw(&tp->syncp);
3341 /* If we update tp->rcv_nxt, also update tp->bytes_received */
3342 static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq)
3344 u32 delta = seq - tp->rcv_nxt;
3346 sock_owned_by_me((struct sock *)tp);
3347 u64_stats_update_begin_raw(&tp->syncp);
3348 tp->bytes_received += delta;
3349 u64_stats_update_end_raw(&tp->syncp);
3353 /* Update our send window.
3355 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3356 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3358 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3361 struct tcp_sock *tp = tcp_sk(sk);
3363 u32 nwin = ntohs(tcp_hdr(skb)->window);
3365 if (likely(!tcp_hdr(skb)->syn))
3366 nwin <<= tp->rx_opt.snd_wscale;
3368 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3369 flag |= FLAG_WIN_UPDATE;
3370 tcp_update_wl(tp, ack_seq);
3372 if (tp->snd_wnd != nwin) {
3375 /* Note, it is the only place, where
3376 * fast path is recovered for sending TCP.
3379 tcp_fast_path_check(sk);
3381 if (tcp_send_head(sk))
3382 tcp_slow_start_after_idle_check(sk);
3384 if (nwin > tp->max_window) {
3385 tp->max_window = nwin;
3386 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3391 tcp_snd_una_update(tp, ack);
3396 static bool __tcp_oow_rate_limited(struct net *net, int mib_idx,
3397 u32 *last_oow_ack_time)
3399 if (*last_oow_ack_time) {
3400 s32 elapsed = (s32)(tcp_time_stamp - *last_oow_ack_time);
3402 if (0 <= elapsed && elapsed < sysctl_tcp_invalid_ratelimit) {
3403 NET_INC_STATS(net, mib_idx);
3404 return true; /* rate-limited: don't send yet! */
3408 *last_oow_ack_time = tcp_time_stamp;
3410 return false; /* not rate-limited: go ahead, send dupack now! */
3413 /* Return true if we're currently rate-limiting out-of-window ACKs and
3414 * thus shouldn't send a dupack right now. We rate-limit dupacks in
3415 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS
3416 * attacks that send repeated SYNs or ACKs for the same connection. To
3417 * do this, we do not send a duplicate SYNACK or ACK if the remote
3418 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate.
3420 bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb,
3421 int mib_idx, u32 *last_oow_ack_time)
3423 /* Data packets without SYNs are not likely part of an ACK loop. */
3424 if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) &&
3428 return __tcp_oow_rate_limited(net, mib_idx, last_oow_ack_time);
3431 /* RFC 5961 7 [ACK Throttling] */
3432 static void tcp_send_challenge_ack(struct sock *sk, const struct sk_buff *skb)
3434 /* unprotected vars, we dont care of overwrites */
3435 static u32 challenge_timestamp;
3436 static unsigned int challenge_count;
3437 struct tcp_sock *tp = tcp_sk(sk);
3440 /* First check our per-socket dupack rate limit. */
3441 if (__tcp_oow_rate_limited(sock_net(sk),
3442 LINUX_MIB_TCPACKSKIPPEDCHALLENGE,
3443 &tp->last_oow_ack_time))
3446 /* Then check host-wide RFC 5961 rate limit. */
3448 if (now != challenge_timestamp) {
3449 u32 half = (sysctl_tcp_challenge_ack_limit + 1) >> 1;
3451 challenge_timestamp = now;
3452 WRITE_ONCE(challenge_count, half +
3453 prandom_u32_max(sysctl_tcp_challenge_ack_limit));
3455 count = READ_ONCE(challenge_count);
3457 WRITE_ONCE(challenge_count, count - 1);
3458 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPCHALLENGEACK);
3463 static void tcp_store_ts_recent(struct tcp_sock *tp)
3465 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3466 tp->rx_opt.ts_recent_stamp = get_seconds();
3469 static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3471 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3472 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3473 * extra check below makes sure this can only happen
3474 * for pure ACK frames. -DaveM
3476 * Not only, also it occurs for expired timestamps.
3479 if (tcp_paws_check(&tp->rx_opt, 0))
3480 tcp_store_ts_recent(tp);
3484 /* This routine deals with acks during a TLP episode.
3485 * We mark the end of a TLP episode on receiving TLP dupack or when
3486 * ack is after tlp_high_seq.
3487 * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3489 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3491 struct tcp_sock *tp = tcp_sk(sk);
3493 if (before(ack, tp->tlp_high_seq))
3496 if (flag & FLAG_DSACKING_ACK) {
3497 /* This DSACK means original and TLP probe arrived; no loss */
3498 tp->tlp_high_seq = 0;
3499 } else if (after(ack, tp->tlp_high_seq)) {
3500 /* ACK advances: there was a loss, so reduce cwnd. Reset
3501 * tlp_high_seq in tcp_init_cwnd_reduction()
3503 tcp_init_cwnd_reduction(sk);
3504 tcp_set_ca_state(sk, TCP_CA_CWR);
3505 tcp_end_cwnd_reduction(sk);
3506 tcp_try_keep_open(sk);
3507 NET_INC_STATS(sock_net(sk),
3508 LINUX_MIB_TCPLOSSPROBERECOVERY);
3509 } else if (!(flag & (FLAG_SND_UNA_ADVANCED |
3510 FLAG_NOT_DUP | FLAG_DATA_SACKED))) {
3511 /* Pure dupack: original and TLP probe arrived; no loss */
3512 tp->tlp_high_seq = 0;
3516 static inline void tcp_in_ack_event(struct sock *sk, u32 flags)
3518 const struct inet_connection_sock *icsk = inet_csk(sk);
3520 if (icsk->icsk_ca_ops->in_ack_event)
3521 icsk->icsk_ca_ops->in_ack_event(sk, flags);
3524 /* Congestion control has updated the cwnd already. So if we're in
3525 * loss recovery then now we do any new sends (for FRTO) or
3526 * retransmits (for CA_Loss or CA_recovery) that make sense.
3528 static void tcp_xmit_recovery(struct sock *sk, int rexmit)
3530 struct tcp_sock *tp = tcp_sk(sk);
3532 if (rexmit == REXMIT_NONE)
3535 if (unlikely(rexmit == 2)) {
3536 __tcp_push_pending_frames(sk, tcp_current_mss(sk),
3538 if (after(tp->snd_nxt, tp->high_seq))
3542 tcp_xmit_retransmit_queue(sk);
3545 /* This routine deals with incoming acks, but not outgoing ones. */
3546 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3548 struct inet_connection_sock *icsk = inet_csk(sk);
3549 struct tcp_sock *tp = tcp_sk(sk);
3550 struct tcp_sacktag_state sack_state;
3551 u32 prior_snd_una = tp->snd_una;
3552 u32 ack_seq = TCP_SKB_CB(skb)->seq;
3553 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3554 bool is_dupack = false;
3556 int prior_packets = tp->packets_out;
3557 u32 prior_delivered = tp->delivered;
3558 int acked = 0; /* Number of packets newly acked */
3559 int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */
3561 sack_state.first_sackt.v64 = 0;
3563 /* We very likely will need to access write queue head. */
3564 prefetchw(sk->sk_write_queue.next);
3566 /* If the ack is older than previous acks
3567 * then we can probably ignore it.
3569 if (before(ack, prior_snd_una)) {
3570 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3571 if (before(ack, prior_snd_una - tp->max_window)) {
3572 tcp_send_challenge_ack(sk, skb);
3578 /* If the ack includes data we haven't sent yet, discard
3579 * this segment (RFC793 Section 3.9).
3581 if (after(ack, tp->snd_nxt))
3584 if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
3585 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
3588 if (after(ack, prior_snd_una)) {
3589 flag |= FLAG_SND_UNA_ADVANCED;
3590 icsk->icsk_retransmits = 0;
3593 prior_fackets = tp->fackets_out;
3595 /* ts_recent update must be made after we are sure that the packet
3598 if (flag & FLAG_UPDATE_TS_RECENT)
3599 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3601 if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
3602 /* Window is constant, pure forward advance.
3603 * No more checks are required.
3604 * Note, we use the fact that SND.UNA>=SND.WL2.
3606 tcp_update_wl(tp, ack_seq);
3607 tcp_snd_una_update(tp, ack);
3608 flag |= FLAG_WIN_UPDATE;
3610 tcp_in_ack_event(sk, CA_ACK_WIN_UPDATE);
3612 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPACKS);
3614 u32 ack_ev_flags = CA_ACK_SLOWPATH;
3616 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3619 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3621 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3623 if (TCP_SKB_CB(skb)->sacked)
3624 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3627 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) {
3629 ack_ev_flags |= CA_ACK_ECE;
3632 if (flag & FLAG_WIN_UPDATE)
3633 ack_ev_flags |= CA_ACK_WIN_UPDATE;
3635 tcp_in_ack_event(sk, ack_ev_flags);
3638 /* We passed data and got it acked, remove any soft error
3639 * log. Something worked...
3641 sk->sk_err_soft = 0;
3642 icsk->icsk_probes_out = 0;
3643 tp->rcv_tstamp = tcp_time_stamp;
3647 /* See if we can take anything off of the retransmit queue. */
3648 flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una, &acked,
3651 if (tcp_ack_is_dubious(sk, flag)) {
3652 is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
3653 tcp_fastretrans_alert(sk, acked, is_dupack, &flag, &rexmit);
3655 if (tp->tlp_high_seq)
3656 tcp_process_tlp_ack(sk, ack, flag);
3658 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) {
3659 struct dst_entry *dst = __sk_dst_get(sk);
3664 if (icsk->icsk_pending == ICSK_TIME_RETRANS)
3665 tcp_schedule_loss_probe(sk);
3666 tcp_cong_control(sk, ack, tp->delivered - prior_delivered, flag);
3667 tcp_xmit_recovery(sk, rexmit);
3671 /* If data was DSACKed, see if we can undo a cwnd reduction. */
3672 if (flag & FLAG_DSACKING_ACK)
3673 tcp_fastretrans_alert(sk, acked, is_dupack, &flag, &rexmit);
3674 /* If this ack opens up a zero window, clear backoff. It was
3675 * being used to time the probes, and is probably far higher than
3676 * it needs to be for normal retransmission.
3678 if (tcp_send_head(sk))
3681 if (tp->tlp_high_seq)
3682 tcp_process_tlp_ack(sk, ack, flag);
3686 SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3690 /* If data was SACKed, tag it and see if we should send more data.
3691 * If data was DSACKed, see if we can undo a cwnd reduction.
3693 if (TCP_SKB_CB(skb)->sacked) {
3694 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3696 tcp_fastretrans_alert(sk, acked, is_dupack, &flag, &rexmit);
3697 tcp_xmit_recovery(sk, rexmit);
3700 SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3704 static void tcp_parse_fastopen_option(int len, const unsigned char *cookie,
3705 bool syn, struct tcp_fastopen_cookie *foc,
3708 /* Valid only in SYN or SYN-ACK with an even length. */
3709 if (!foc || !syn || len < 0 || (len & 1))
3712 if (len >= TCP_FASTOPEN_COOKIE_MIN &&
3713 len <= TCP_FASTOPEN_COOKIE_MAX)
3714 memcpy(foc->val, cookie, len);
3721 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
3722 * But, this can also be called on packets in the established flow when
3723 * the fast version below fails.
3725 void tcp_parse_options(const struct sk_buff *skb,
3726 struct tcp_options_received *opt_rx, int estab,
3727 struct tcp_fastopen_cookie *foc)
3729 const unsigned char *ptr;
3730 const struct tcphdr *th = tcp_hdr(skb);
3731 int length = (th->doff * 4) - sizeof(struct tcphdr);
3733 ptr = (const unsigned char *)(th + 1);
3734 opt_rx->saw_tstamp = 0;
3736 while (length > 0) {
3737 int opcode = *ptr++;
3743 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3748 if (opsize < 2) /* "silly options" */
3750 if (opsize > length)
3751 return; /* don't parse partial options */
3754 if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3755 u16 in_mss = get_unaligned_be16(ptr);
3757 if (opt_rx->user_mss &&
3758 opt_rx->user_mss < in_mss)
3759 in_mss = opt_rx->user_mss;
3760 opt_rx->mss_clamp = in_mss;
3765 if (opsize == TCPOLEN_WINDOW && th->syn &&
3766 !estab && sysctl_tcp_window_scaling) {
3767 __u8 snd_wscale = *(__u8 *)ptr;
3768 opt_rx->wscale_ok = 1;
3769 if (snd_wscale > 14) {
3770 net_info_ratelimited("%s: Illegal window scaling value %d >14 received\n",
3775 opt_rx->snd_wscale = snd_wscale;
3778 case TCPOPT_TIMESTAMP:
3779 if ((opsize == TCPOLEN_TIMESTAMP) &&
3780 ((estab && opt_rx->tstamp_ok) ||
3781 (!estab && sysctl_tcp_timestamps))) {
3782 opt_rx->saw_tstamp = 1;
3783 opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3784 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3787 case TCPOPT_SACK_PERM:
3788 if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3789 !estab && sysctl_tcp_sack) {
3790 opt_rx->sack_ok = TCP_SACK_SEEN;
3791 tcp_sack_reset(opt_rx);
3796 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3797 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3799 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3802 #ifdef CONFIG_TCP_MD5SIG
3805 * The MD5 Hash has already been
3806 * checked (see tcp_v{4,6}_do_rcv()).
3810 case TCPOPT_FASTOPEN:
3811 tcp_parse_fastopen_option(
3812 opsize - TCPOLEN_FASTOPEN_BASE,
3813 ptr, th->syn, foc, false);
3817 /* Fast Open option shares code 254 using a
3818 * 16 bits magic number.
3820 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE &&
3821 get_unaligned_be16(ptr) ==
3822 TCPOPT_FASTOPEN_MAGIC)
3823 tcp_parse_fastopen_option(opsize -
3824 TCPOLEN_EXP_FASTOPEN_BASE,
3825 ptr + 2, th->syn, foc, true);
3834 EXPORT_SYMBOL(tcp_parse_options);
3836 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
3838 const __be32 *ptr = (const __be32 *)(th + 1);
3840 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3841 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3842 tp->rx_opt.saw_tstamp = 1;
3844 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3847 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
3849 tp->rx_opt.rcv_tsecr = 0;
3855 /* Fast parse options. This hopes to only see timestamps.
3856 * If it is wrong it falls back on tcp_parse_options().
3858 static bool tcp_fast_parse_options(const struct sk_buff *skb,
3859 const struct tcphdr *th, struct tcp_sock *tp)
3861 /* In the spirit of fast parsing, compare doff directly to constant
3862 * values. Because equality is used, short doff can be ignored here.
3864 if (th->doff == (sizeof(*th) / 4)) {
3865 tp->rx_opt.saw_tstamp = 0;
3867 } else if (tp->rx_opt.tstamp_ok &&
3868 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3869 if (tcp_parse_aligned_timestamp(tp, th))
3873 tcp_parse_options(skb, &tp->rx_opt, 1, NULL);
3874 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3875 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
3880 #ifdef CONFIG_TCP_MD5SIG
3882 * Parse MD5 Signature option
3884 const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
3886 int length = (th->doff << 2) - sizeof(*th);
3887 const u8 *ptr = (const u8 *)(th + 1);
3889 /* If the TCP option is too short, we can short cut */
3890 if (length < TCPOLEN_MD5SIG)
3893 while (length > 0) {
3894 int opcode = *ptr++;
3905 if (opsize < 2 || opsize > length)
3907 if (opcode == TCPOPT_MD5SIG)
3908 return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
3915 EXPORT_SYMBOL(tcp_parse_md5sig_option);
3918 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3920 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3921 * it can pass through stack. So, the following predicate verifies that
3922 * this segment is not used for anything but congestion avoidance or
3923 * fast retransmit. Moreover, we even are able to eliminate most of such
3924 * second order effects, if we apply some small "replay" window (~RTO)
3925 * to timestamp space.
3927 * All these measures still do not guarantee that we reject wrapped ACKs
3928 * on networks with high bandwidth, when sequence space is recycled fastly,
3929 * but it guarantees that such events will be very rare and do not affect
3930 * connection seriously. This doesn't look nice, but alas, PAWS is really
3933 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3934 * states that events when retransmit arrives after original data are rare.
3935 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3936 * the biggest problem on large power networks even with minor reordering.
3937 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3938 * up to bandwidth of 18Gigabit/sec. 8) ]
3941 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3943 const struct tcp_sock *tp = tcp_sk(sk);
3944 const struct tcphdr *th = tcp_hdr(skb);
3945 u32 seq = TCP_SKB_CB(skb)->seq;
3946 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3948 return (/* 1. Pure ACK with correct sequence number. */
3949 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3951 /* 2. ... and duplicate ACK. */
3952 ack == tp->snd_una &&
3954 /* 3. ... and does not update window. */
3955 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3957 /* 4. ... and sits in replay window. */
3958 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3961 static inline bool tcp_paws_discard(const struct sock *sk,
3962 const struct sk_buff *skb)
3964 const struct tcp_sock *tp = tcp_sk(sk);
3966 return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
3967 !tcp_disordered_ack(sk, skb);
3970 /* Check segment sequence number for validity.
3972 * Segment controls are considered valid, if the segment
3973 * fits to the window after truncation to the window. Acceptability
3974 * of data (and SYN, FIN, of course) is checked separately.
3975 * See tcp_data_queue(), for example.
3977 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3978 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3979 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3980 * (borrowed from freebsd)
3983 static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
3985 return !before(end_seq, tp->rcv_wup) &&
3986 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3989 /* When we get a reset we do this. */
3990 void tcp_reset(struct sock *sk)
3992 /* We want the right error as BSD sees it (and indeed as we do). */
3993 switch (sk->sk_state) {
3995 sk->sk_err = ECONNREFUSED;
3997 case TCP_CLOSE_WAIT:
4003 sk->sk_err = ECONNRESET;
4005 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4008 if (!sock_flag(sk, SOCK_DEAD))
4009 sk->sk_error_report(sk);
4015 * Process the FIN bit. This now behaves as it is supposed to work
4016 * and the FIN takes effect when it is validly part of sequence
4017 * space. Not before when we get holes.
4019 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
4020 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
4023 * If we are in FINWAIT-1, a received FIN indicates simultaneous
4024 * close and we go into CLOSING (and later onto TIME-WAIT)
4026 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
4028 void tcp_fin(struct sock *sk)
4030 struct tcp_sock *tp = tcp_sk(sk);
4032 inet_csk_schedule_ack(sk);
4034 sk->sk_shutdown |= RCV_SHUTDOWN;
4035 sock_set_flag(sk, SOCK_DONE);
4037 switch (sk->sk_state) {
4039 case TCP_ESTABLISHED:
4040 /* Move to CLOSE_WAIT */
4041 tcp_set_state(sk, TCP_CLOSE_WAIT);
4042 inet_csk(sk)->icsk_ack.pingpong = 1;
4045 case TCP_CLOSE_WAIT:
4047 /* Received a retransmission of the FIN, do
4052 /* RFC793: Remain in the LAST-ACK state. */
4056 /* This case occurs when a simultaneous close
4057 * happens, we must ack the received FIN and
4058 * enter the CLOSING state.
4061 tcp_set_state(sk, TCP_CLOSING);
4064 /* Received a FIN -- send ACK and enter TIME_WAIT. */
4066 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4069 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
4070 * cases we should never reach this piece of code.
4072 pr_err("%s: Impossible, sk->sk_state=%d\n",
4073 __func__, sk->sk_state);
4077 /* It _is_ possible, that we have something out-of-order _after_ FIN.
4078 * Probably, we should reset in this case. For now drop them.
4080 skb_rbtree_purge(&tp->out_of_order_queue);
4081 if (tcp_is_sack(tp))
4082 tcp_sack_reset(&tp->rx_opt);
4085 if (!sock_flag(sk, SOCK_DEAD)) {
4086 sk->sk_state_change(sk);
4088 /* Do not send POLL_HUP for half duplex close. */
4089 if (sk->sk_shutdown == SHUTDOWN_MASK ||
4090 sk->sk_state == TCP_CLOSE)
4091 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
4093 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
4097 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
4100 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
4101 if (before(seq, sp->start_seq))
4102 sp->start_seq = seq;
4103 if (after(end_seq, sp->end_seq))
4104 sp->end_seq = end_seq;
4110 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
4112 struct tcp_sock *tp = tcp_sk(sk);
4114 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4117 if (before(seq, tp->rcv_nxt))
4118 mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
4120 mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
4122 NET_INC_STATS(sock_net(sk), mib_idx);
4124 tp->rx_opt.dsack = 1;
4125 tp->duplicate_sack[0].start_seq = seq;
4126 tp->duplicate_sack[0].end_seq = end_seq;
4130 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
4132 struct tcp_sock *tp = tcp_sk(sk);
4134 if (!tp->rx_opt.dsack)
4135 tcp_dsack_set(sk, seq, end_seq);
4137 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
4140 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
4142 struct tcp_sock *tp = tcp_sk(sk);
4144 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4145 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4146 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4147 tcp_enter_quickack_mode(sk);
4149 if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
4150 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
4152 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
4153 end_seq = tp->rcv_nxt;
4154 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
4161 /* These routines update the SACK block as out-of-order packets arrive or
4162 * in-order packets close up the sequence space.
4164 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
4167 struct tcp_sack_block *sp = &tp->selective_acks[0];
4168 struct tcp_sack_block *swalk = sp + 1;
4170 /* See if the recent change to the first SACK eats into
4171 * or hits the sequence space of other SACK blocks, if so coalesce.
4173 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
4174 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
4177 /* Zap SWALK, by moving every further SACK up by one slot.
4178 * Decrease num_sacks.
4180 tp->rx_opt.num_sacks--;
4181 for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
4185 this_sack++, swalk++;
4189 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
4191 struct tcp_sock *tp = tcp_sk(sk);
4192 struct tcp_sack_block *sp = &tp->selective_acks[0];
4193 int cur_sacks = tp->rx_opt.num_sacks;
4199 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
4200 if (tcp_sack_extend(sp, seq, end_seq)) {
4201 /* Rotate this_sack to the first one. */
4202 for (; this_sack > 0; this_sack--, sp--)
4203 swap(*sp, *(sp - 1));
4205 tcp_sack_maybe_coalesce(tp);
4210 /* Could not find an adjacent existing SACK, build a new one,
4211 * put it at the front, and shift everyone else down. We
4212 * always know there is at least one SACK present already here.
4214 * If the sack array is full, forget about the last one.
4216 if (this_sack >= TCP_NUM_SACKS) {
4218 tp->rx_opt.num_sacks--;
4221 for (; this_sack > 0; this_sack--, sp--)
4225 /* Build the new head SACK, and we're done. */
4226 sp->start_seq = seq;
4227 sp->end_seq = end_seq;
4228 tp->rx_opt.num_sacks++;
4231 /* RCV.NXT advances, some SACKs should be eaten. */
4233 static void tcp_sack_remove(struct tcp_sock *tp)
4235 struct tcp_sack_block *sp = &tp->selective_acks[0];
4236 int num_sacks = tp->rx_opt.num_sacks;
4239 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4240 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4241 tp->rx_opt.num_sacks = 0;
4245 for (this_sack = 0; this_sack < num_sacks;) {
4246 /* Check if the start of the sack is covered by RCV.NXT. */
4247 if (!before(tp->rcv_nxt, sp->start_seq)) {
4250 /* RCV.NXT must cover all the block! */
4251 WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4253 /* Zap this SACK, by moving forward any other SACKS. */
4254 for (i = this_sack+1; i < num_sacks; i++)
4255 tp->selective_acks[i-1] = tp->selective_acks[i];
4262 tp->rx_opt.num_sacks = num_sacks;
4266 * tcp_try_coalesce - try to merge skb to prior one
4269 * @from: buffer to add in queue
4270 * @fragstolen: pointer to boolean
4272 * Before queueing skb @from after @to, try to merge them
4273 * to reduce overall memory use and queue lengths, if cost is small.
4274 * Packets in ofo or receive queues can stay a long time.
4275 * Better try to coalesce them right now to avoid future collapses.
4276 * Returns true if caller should free @from instead of queueing it
4278 static bool tcp_try_coalesce(struct sock *sk,
4280 struct sk_buff *from,
4285 *fragstolen = false;
4287 /* Its possible this segment overlaps with prior segment in queue */
4288 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4291 if (!skb_try_coalesce(to, from, fragstolen, &delta))
4294 atomic_add(delta, &sk->sk_rmem_alloc);
4295 sk_mem_charge(sk, delta);
4296 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4297 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4298 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4299 TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags;
4303 static void tcp_drop(struct sock *sk, struct sk_buff *skb)
4305 sk_drops_add(sk, skb);
4309 /* This one checks to see if we can put data from the
4310 * out_of_order queue into the receive_queue.
4312 static void tcp_ofo_queue(struct sock *sk)
4314 struct tcp_sock *tp = tcp_sk(sk);
4315 __u32 dsack_high = tp->rcv_nxt;
4316 bool fin, fragstolen, eaten;
4317 struct sk_buff *skb, *tail;
4320 p = rb_first(&tp->out_of_order_queue);
4322 skb = rb_entry(p, struct sk_buff, rbnode);
4323 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4326 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4327 __u32 dsack = dsack_high;
4328 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4329 dsack_high = TCP_SKB_CB(skb)->end_seq;
4330 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4333 rb_erase(&skb->rbnode, &tp->out_of_order_queue);
4335 if (unlikely(!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))) {
4336 SOCK_DEBUG(sk, "ofo packet was already received\n");
4340 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4341 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4342 TCP_SKB_CB(skb)->end_seq);
4344 tail = skb_peek_tail(&sk->sk_receive_queue);
4345 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen);
4346 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4347 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
4349 __skb_queue_tail(&sk->sk_receive_queue, skb);
4351 kfree_skb_partial(skb, fragstolen);
4353 if (unlikely(fin)) {
4355 /* tcp_fin() purges tp->out_of_order_queue,
4356 * so we must end this loop right now.
4363 static bool tcp_prune_ofo_queue(struct sock *sk);
4364 static int tcp_prune_queue(struct sock *sk);
4366 static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4369 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4370 !sk_rmem_schedule(sk, skb, size)) {
4372 if (tcp_prune_queue(sk) < 0)
4375 while (!sk_rmem_schedule(sk, skb, size)) {
4376 if (!tcp_prune_ofo_queue(sk))
4383 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4385 struct tcp_sock *tp = tcp_sk(sk);
4386 struct rb_node **p, *q, *parent;
4387 struct sk_buff *skb1;
4391 tcp_ecn_check_ce(tp, skb);
4393 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4394 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFODROP);
4399 /* Disable header prediction. */
4401 inet_csk_schedule_ack(sk);
4403 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4404 seq = TCP_SKB_CB(skb)->seq;
4405 end_seq = TCP_SKB_CB(skb)->end_seq;
4406 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4407 tp->rcv_nxt, seq, end_seq);
4409 p = &tp->out_of_order_queue.rb_node;
4410 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4411 /* Initial out of order segment, build 1 SACK. */
4412 if (tcp_is_sack(tp)) {
4413 tp->rx_opt.num_sacks = 1;
4414 tp->selective_acks[0].start_seq = seq;
4415 tp->selective_acks[0].end_seq = end_seq;
4417 rb_link_node(&skb->rbnode, NULL, p);
4418 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4419 tp->ooo_last_skb = skb;
4423 /* In the typical case, we are adding an skb to the end of the list.
4424 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
4426 if (tcp_try_coalesce(sk, tp->ooo_last_skb, skb, &fragstolen)) {
4428 tcp_grow_window(sk, skb);
4429 kfree_skb_partial(skb, fragstolen);
4433 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
4434 if (!before(seq, TCP_SKB_CB(tp->ooo_last_skb)->end_seq)) {
4435 parent = &tp->ooo_last_skb->rbnode;
4436 p = &parent->rb_right;
4440 /* Find place to insert this segment. Handle overlaps on the way. */
4444 skb1 = rb_entry(parent, struct sk_buff, rbnode);
4445 if (before(seq, TCP_SKB_CB(skb1)->seq)) {
4446 p = &parent->rb_left;
4449 if (before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4450 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4451 /* All the bits are present. Drop. */
4452 NET_INC_STATS(sock_net(sk),
4453 LINUX_MIB_TCPOFOMERGE);
4456 tcp_dsack_set(sk, seq, end_seq);
4459 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4460 /* Partial overlap. */
4461 tcp_dsack_set(sk, seq, TCP_SKB_CB(skb1)->end_seq);
4463 /* skb's seq == skb1's seq and skb covers skb1.
4464 * Replace skb1 with skb.
4466 rb_replace_node(&skb1->rbnode, &skb->rbnode,
4467 &tp->out_of_order_queue);
4468 tcp_dsack_extend(sk,
4469 TCP_SKB_CB(skb1)->seq,
4470 TCP_SKB_CB(skb1)->end_seq);
4471 NET_INC_STATS(sock_net(sk),
4472 LINUX_MIB_TCPOFOMERGE);
4476 } else if (tcp_try_coalesce(sk, skb1, skb, &fragstolen)) {
4479 p = &parent->rb_right;
4482 /* Insert segment into RB tree. */
4483 rb_link_node(&skb->rbnode, parent, p);
4484 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue);
4487 /* Remove other segments covered by skb. */
4488 while ((q = rb_next(&skb->rbnode)) != NULL) {
4489 skb1 = rb_entry(q, struct sk_buff, rbnode);
4491 if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4493 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4494 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4498 rb_erase(&skb1->rbnode, &tp->out_of_order_queue);
4499 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4500 TCP_SKB_CB(skb1)->end_seq);
4501 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4504 /* If there is no skb after us, we are the last_skb ! */
4506 tp->ooo_last_skb = skb;
4509 if (tcp_is_sack(tp))
4510 tcp_sack_new_ofo_skb(sk, seq, end_seq);
4513 tcp_grow_window(sk, skb);
4514 skb_set_owner_r(skb, sk);
4518 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, int hdrlen,
4522 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4524 __skb_pull(skb, hdrlen);
4526 tcp_try_coalesce(sk, tail, skb, fragstolen)) ? 1 : 0;
4527 tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq);
4529 __skb_queue_tail(&sk->sk_receive_queue, skb);
4530 skb_set_owner_r(skb, sk);
4535 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4537 struct sk_buff *skb;
4545 if (size > PAGE_SIZE) {
4546 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS);
4548 data_len = npages << PAGE_SHIFT;
4549 size = data_len + (size & ~PAGE_MASK);
4551 skb = alloc_skb_with_frags(size - data_len, data_len,
4552 PAGE_ALLOC_COSTLY_ORDER,
4553 &err, sk->sk_allocation);
4557 skb_put(skb, size - data_len);
4558 skb->data_len = data_len;
4561 if (tcp_try_rmem_schedule(sk, skb, skb->truesize))
4564 err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size);
4568 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4569 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4570 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4572 if (tcp_queue_rcv(sk, skb, 0, &fragstolen)) {
4573 WARN_ON_ONCE(fragstolen); /* should not happen */
4585 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4587 struct tcp_sock *tp = tcp_sk(sk);
4588 bool fragstolen = false;
4591 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
4596 __skb_pull(skb, tcp_hdr(skb)->doff * 4);
4598 tcp_ecn_accept_cwr(tp, skb);
4600 tp->rx_opt.dsack = 0;
4602 /* Queue data for delivery to the user.
4603 * Packets in sequence go to the receive queue.
4604 * Out of sequence packets to the out_of_order_queue.
4606 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4607 if (tcp_receive_window(tp) == 0)
4610 /* Ok. In sequence. In window. */
4611 if (tp->ucopy.task == current &&
4612 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4613 sock_owned_by_user(sk) && !tp->urg_data) {
4614 int chunk = min_t(unsigned int, skb->len,
4617 __set_current_state(TASK_RUNNING);
4619 if (!skb_copy_datagram_msg(skb, 0, tp->ucopy.msg, chunk)) {
4620 tp->ucopy.len -= chunk;
4621 tp->copied_seq += chunk;
4622 eaten = (chunk == skb->len);
4623 tcp_rcv_space_adjust(sk);
4630 if (skb_queue_len(&sk->sk_receive_queue) == 0)
4631 sk_forced_mem_schedule(sk, skb->truesize);
4632 else if (tcp_try_rmem_schedule(sk, skb, skb->truesize))
4635 eaten = tcp_queue_rcv(sk, skb, 0, &fragstolen);
4637 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
4639 tcp_event_data_recv(sk, skb);
4640 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
4643 if (!RB_EMPTY_ROOT(&tp->out_of_order_queue)) {
4646 /* RFC2581. 4.2. SHOULD send immediate ACK, when
4647 * gap in queue is filled.
4649 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
4650 inet_csk(sk)->icsk_ack.pingpong = 0;
4653 if (tp->rx_opt.num_sacks)
4654 tcp_sack_remove(tp);
4656 tcp_fast_path_check(sk);
4659 kfree_skb_partial(skb, fragstolen);
4660 if (!sock_flag(sk, SOCK_DEAD))
4661 sk->sk_data_ready(sk);
4665 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4666 /* A retransmit, 2nd most common case. Force an immediate ack. */
4667 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
4668 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4671 tcp_enter_quickack_mode(sk);
4672 inet_csk_schedule_ack(sk);
4678 /* Out of window. F.e. zero window probe. */
4679 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
4682 tcp_enter_quickack_mode(sk);
4684 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4685 /* Partial packet, seq < rcv_next < end_seq */
4686 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4687 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4688 TCP_SKB_CB(skb)->end_seq);
4690 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4692 /* If window is closed, drop tail of packet. But after
4693 * remembering D-SACK for its head made in previous line.
4695 if (!tcp_receive_window(tp))
4700 tcp_data_queue_ofo(sk, skb);
4703 static struct sk_buff *tcp_skb_next(struct sk_buff *skb, struct sk_buff_head *list)
4706 return !skb_queue_is_last(list, skb) ? skb->next : NULL;
4708 return rb_entry_safe(rb_next(&skb->rbnode), struct sk_buff, rbnode);
4711 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4712 struct sk_buff_head *list,
4713 struct rb_root *root)
4715 struct sk_buff *next = tcp_skb_next(skb, list);
4718 __skb_unlink(skb, list);
4720 rb_erase(&skb->rbnode, root);
4723 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4728 /* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */
4729 static void tcp_rbtree_insert(struct rb_root *root, struct sk_buff *skb)
4731 struct rb_node **p = &root->rb_node;
4732 struct rb_node *parent = NULL;
4733 struct sk_buff *skb1;
4737 skb1 = rb_entry(parent, struct sk_buff, rbnode);
4738 if (before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb1)->seq))
4739 p = &parent->rb_left;
4741 p = &parent->rb_right;
4743 rb_link_node(&skb->rbnode, parent, p);
4744 rb_insert_color(&skb->rbnode, root);
4747 /* Collapse contiguous sequence of skbs head..tail with
4748 * sequence numbers start..end.
4750 * If tail is NULL, this means until the end of the queue.
4752 * Segments with FIN/SYN are not collapsed (only because this
4756 tcp_collapse(struct sock *sk, struct sk_buff_head *list, struct rb_root *root,
4757 struct sk_buff *head, struct sk_buff *tail, u32 start, u32 end)
4759 struct sk_buff *skb = head, *n;
4760 struct sk_buff_head tmp;
4763 /* First, check that queue is collapsible and find
4764 * the point where collapsing can be useful.
4767 for (end_of_skbs = true; skb != NULL && skb != tail; skb = n) {
4768 n = tcp_skb_next(skb, list);
4770 /* No new bits? It is possible on ofo queue. */
4771 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4772 skb = tcp_collapse_one(sk, skb, list, root);
4778 /* The first skb to collapse is:
4780 * - bloated or contains data before "start" or
4781 * overlaps to the next one.
4783 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) &&
4784 (tcp_win_from_space(skb->truesize) > skb->len ||
4785 before(TCP_SKB_CB(skb)->seq, start))) {
4786 end_of_skbs = false;
4790 if (n && n != tail &&
4791 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(n)->seq) {
4792 end_of_skbs = false;
4796 /* Decided to skip this, advance start seq. */
4797 start = TCP_SKB_CB(skb)->end_seq;
4800 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4803 __skb_queue_head_init(&tmp);
4805 while (before(start, end)) {
4806 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start);
4807 struct sk_buff *nskb;
4809 nskb = alloc_skb(copy, GFP_ATOMIC);
4813 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
4814 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4816 __skb_queue_before(list, skb, nskb);
4818 __skb_queue_tail(&tmp, nskb); /* defer rbtree insertion */
4819 skb_set_owner_r(nskb, sk);
4821 /* Copy data, releasing collapsed skbs. */
4823 int offset = start - TCP_SKB_CB(skb)->seq;
4824 int size = TCP_SKB_CB(skb)->end_seq - start;
4828 size = min(copy, size);
4829 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4831 TCP_SKB_CB(nskb)->end_seq += size;
4835 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4836 skb = tcp_collapse_one(sk, skb, list, root);
4839 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)))
4845 skb_queue_walk_safe(&tmp, skb, n)
4846 tcp_rbtree_insert(root, skb);
4849 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4850 * and tcp_collapse() them until all the queue is collapsed.
4852 static void tcp_collapse_ofo_queue(struct sock *sk)
4854 struct tcp_sock *tp = tcp_sk(sk);
4855 struct sk_buff *skb, *head;
4859 p = rb_first(&tp->out_of_order_queue);
4860 skb = rb_entry_safe(p, struct sk_buff, rbnode);
4863 p = rb_last(&tp->out_of_order_queue);
4864 /* Note: This is possible p is NULL here. We do not
4865 * use rb_entry_safe(), as ooo_last_skb is valid only
4866 * if rbtree is not empty.
4868 tp->ooo_last_skb = rb_entry(p, struct sk_buff, rbnode);
4871 start = TCP_SKB_CB(skb)->seq;
4872 end = TCP_SKB_CB(skb)->end_seq;
4874 for (head = skb;;) {
4875 skb = tcp_skb_next(skb, NULL);
4877 /* Range is terminated when we see a gap or when
4878 * we are at the queue end.
4881 after(TCP_SKB_CB(skb)->seq, end) ||
4882 before(TCP_SKB_CB(skb)->end_seq, start)) {
4883 tcp_collapse(sk, NULL, &tp->out_of_order_queue,
4884 head, skb, start, end);
4888 if (unlikely(before(TCP_SKB_CB(skb)->seq, start)))
4889 start = TCP_SKB_CB(skb)->seq;
4890 if (after(TCP_SKB_CB(skb)->end_seq, end))
4891 end = TCP_SKB_CB(skb)->end_seq;
4896 * Clean the out-of-order queue to make room.
4897 * We drop high sequences packets to :
4898 * 1) Let a chance for holes to be filled.
4899 * 2) not add too big latencies if thousands of packets sit there.
4900 * (But if application shrinks SO_RCVBUF, we could still end up
4901 * freeing whole queue here)
4903 * Return true if queue has shrunk.
4905 static bool tcp_prune_ofo_queue(struct sock *sk)
4907 struct tcp_sock *tp = tcp_sk(sk);
4908 struct rb_node *node, *prev;
4910 if (RB_EMPTY_ROOT(&tp->out_of_order_queue))
4913 NET_INC_STATS(sock_net(sk), LINUX_MIB_OFOPRUNED);
4914 node = &tp->ooo_last_skb->rbnode;
4916 prev = rb_prev(node);
4917 rb_erase(node, &tp->out_of_order_queue);
4918 tcp_drop(sk, rb_entry(node, struct sk_buff, rbnode));
4920 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf &&
4921 !tcp_under_memory_pressure(sk))
4925 tp->ooo_last_skb = rb_entry(prev, struct sk_buff, rbnode);
4927 /* Reset SACK state. A conforming SACK implementation will
4928 * do the same at a timeout based retransmit. When a connection
4929 * is in a sad state like this, we care only about integrity
4930 * of the connection not performance.
4932 if (tp->rx_opt.sack_ok)
4933 tcp_sack_reset(&tp->rx_opt);
4937 /* Reduce allocated memory if we can, trying to get
4938 * the socket within its memory limits again.
4940 * Return less than zero if we should start dropping frames
4941 * until the socket owning process reads some of the data
4942 * to stabilize the situation.
4944 static int tcp_prune_queue(struct sock *sk)
4946 struct tcp_sock *tp = tcp_sk(sk);
4948 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4950 NET_INC_STATS(sock_net(sk), LINUX_MIB_PRUNECALLED);
4952 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4953 tcp_clamp_window(sk);
4954 else if (tcp_under_memory_pressure(sk))
4955 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
4957 tcp_collapse_ofo_queue(sk);
4958 if (!skb_queue_empty(&sk->sk_receive_queue))
4959 tcp_collapse(sk, &sk->sk_receive_queue, NULL,
4960 skb_peek(&sk->sk_receive_queue),
4962 tp->copied_seq, tp->rcv_nxt);
4965 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4968 /* Collapsing did not help, destructive actions follow.
4969 * This must not ever occur. */
4971 tcp_prune_ofo_queue(sk);
4973 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4976 /* If we are really being abused, tell the caller to silently
4977 * drop receive data on the floor. It will get retransmitted
4978 * and hopefully then we'll have sufficient space.
4980 NET_INC_STATS(sock_net(sk), LINUX_MIB_RCVPRUNED);
4982 /* Massive buffer overcommit. */
4987 static bool tcp_should_expand_sndbuf(const struct sock *sk)
4989 const struct tcp_sock *tp = tcp_sk(sk);
4991 /* If the user specified a specific send buffer setting, do
4994 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4997 /* If we are under global TCP memory pressure, do not expand. */
4998 if (tcp_under_memory_pressure(sk))
5001 /* If we are under soft global TCP memory pressure, do not expand. */
5002 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
5005 /* If we filled the congestion window, do not expand. */
5006 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
5012 /* When incoming ACK allowed to free some skb from write_queue,
5013 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
5014 * on the exit from tcp input handler.
5016 * PROBLEM: sndbuf expansion does not work well with largesend.
5018 static void tcp_new_space(struct sock *sk)
5020 struct tcp_sock *tp = tcp_sk(sk);
5022 if (tcp_should_expand_sndbuf(sk)) {
5023 tcp_sndbuf_expand(sk);
5024 tp->snd_cwnd_stamp = tcp_time_stamp;
5027 sk->sk_write_space(sk);
5030 static void tcp_check_space(struct sock *sk)
5032 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
5033 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
5034 /* pairs with tcp_poll() */
5035 smp_mb__after_atomic();
5036 if (sk->sk_socket &&
5037 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
5042 static inline void tcp_data_snd_check(struct sock *sk)
5044 tcp_push_pending_frames(sk);
5045 tcp_check_space(sk);
5049 * Check if sending an ack is needed.
5051 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
5053 struct tcp_sock *tp = tcp_sk(sk);
5055 /* More than one full frame received... */
5056 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
5057 /* ... and right edge of window advances far enough.
5058 * (tcp_recvmsg() will send ACK otherwise). Or...
5060 __tcp_select_window(sk) >= tp->rcv_wnd) ||
5061 /* We ACK each frame or... */
5062 tcp_in_quickack_mode(sk) ||
5063 /* We have out of order data. */
5064 (ofo_possible && !RB_EMPTY_ROOT(&tp->out_of_order_queue))) {
5065 /* Then ack it now */
5068 /* Else, send delayed ack. */
5069 tcp_send_delayed_ack(sk);
5073 static inline void tcp_ack_snd_check(struct sock *sk)
5075 if (!inet_csk_ack_scheduled(sk)) {
5076 /* We sent a data segment already. */
5079 __tcp_ack_snd_check(sk, 1);
5083 * This routine is only called when we have urgent data
5084 * signaled. Its the 'slow' part of tcp_urg. It could be
5085 * moved inline now as tcp_urg is only called from one
5086 * place. We handle URGent data wrong. We have to - as
5087 * BSD still doesn't use the correction from RFC961.
5088 * For 1003.1g we should support a new option TCP_STDURG to permit
5089 * either form (or just set the sysctl tcp_stdurg).
5092 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
5094 struct tcp_sock *tp = tcp_sk(sk);
5095 u32 ptr = ntohs(th->urg_ptr);
5097 if (ptr && !sysctl_tcp_stdurg)
5099 ptr += ntohl(th->seq);
5101 /* Ignore urgent data that we've already seen and read. */
5102 if (after(tp->copied_seq, ptr))
5105 /* Do not replay urg ptr.
5107 * NOTE: interesting situation not covered by specs.
5108 * Misbehaving sender may send urg ptr, pointing to segment,
5109 * which we already have in ofo queue. We are not able to fetch
5110 * such data and will stay in TCP_URG_NOTYET until will be eaten
5111 * by recvmsg(). Seems, we are not obliged to handle such wicked
5112 * situations. But it is worth to think about possibility of some
5113 * DoSes using some hypothetical application level deadlock.
5115 if (before(ptr, tp->rcv_nxt))
5118 /* Do we already have a newer (or duplicate) urgent pointer? */
5119 if (tp->urg_data && !after(ptr, tp->urg_seq))
5122 /* Tell the world about our new urgent pointer. */
5125 /* We may be adding urgent data when the last byte read was
5126 * urgent. To do this requires some care. We cannot just ignore
5127 * tp->copied_seq since we would read the last urgent byte again
5128 * as data, nor can we alter copied_seq until this data arrives
5129 * or we break the semantics of SIOCATMARK (and thus sockatmark())
5131 * NOTE. Double Dutch. Rendering to plain English: author of comment
5132 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
5133 * and expect that both A and B disappear from stream. This is _wrong_.
5134 * Though this happens in BSD with high probability, this is occasional.
5135 * Any application relying on this is buggy. Note also, that fix "works"
5136 * only in this artificial test. Insert some normal data between A and B and we will
5137 * decline of BSD again. Verdict: it is better to remove to trap
5140 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
5141 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
5142 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
5144 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
5145 __skb_unlink(skb, &sk->sk_receive_queue);
5150 tp->urg_data = TCP_URG_NOTYET;
5153 /* Disable header prediction. */
5157 /* This is the 'fast' part of urgent handling. */
5158 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
5160 struct tcp_sock *tp = tcp_sk(sk);
5162 /* Check if we get a new urgent pointer - normally not. */
5164 tcp_check_urg(sk, th);
5166 /* Do we wait for any urgent data? - normally not... */
5167 if (tp->urg_data == TCP_URG_NOTYET) {
5168 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
5171 /* Is the urgent pointer pointing into this packet? */
5172 if (ptr < skb->len) {
5174 if (skb_copy_bits(skb, ptr, &tmp, 1))
5176 tp->urg_data = TCP_URG_VALID | tmp;
5177 if (!sock_flag(sk, SOCK_DEAD))
5178 sk->sk_data_ready(sk);
5183 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
5185 struct tcp_sock *tp = tcp_sk(sk);
5186 int chunk = skb->len - hlen;
5189 if (skb_csum_unnecessary(skb))
5190 err = skb_copy_datagram_msg(skb, hlen, tp->ucopy.msg, chunk);
5192 err = skb_copy_and_csum_datagram_msg(skb, hlen, tp->ucopy.msg);
5195 tp->ucopy.len -= chunk;
5196 tp->copied_seq += chunk;
5197 tcp_rcv_space_adjust(sk);
5203 /* Does PAWS and seqno based validation of an incoming segment, flags will
5204 * play significant role here.
5206 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5207 const struct tcphdr *th, int syn_inerr)
5209 struct tcp_sock *tp = tcp_sk(sk);
5210 bool rst_seq_match = false;
5212 /* RFC1323: H1. Apply PAWS check first. */
5213 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
5214 tcp_paws_discard(sk, skb)) {
5216 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5217 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5218 LINUX_MIB_TCPACKSKIPPEDPAWS,
5219 &tp->last_oow_ack_time))
5220 tcp_send_dupack(sk, skb);
5223 /* Reset is accepted even if it did not pass PAWS. */
5226 /* Step 1: check sequence number */
5227 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
5228 /* RFC793, page 37: "In all states except SYN-SENT, all reset
5229 * (RST) segments are validated by checking their SEQ-fields."
5230 * And page 69: "If an incoming segment is not acceptable,
5231 * an acknowledgment should be sent in reply (unless the RST
5232 * bit is set, if so drop the segment and return)".
5237 if (!tcp_oow_rate_limited(sock_net(sk), skb,
5238 LINUX_MIB_TCPACKSKIPPEDSEQ,
5239 &tp->last_oow_ack_time))
5240 tcp_send_dupack(sk, skb);
5245 /* Step 2: check RST bit */
5247 /* RFC 5961 3.2 (extend to match against SACK too if available):
5248 * If seq num matches RCV.NXT or the right-most SACK block,
5250 * RESET the connection
5252 * Send a challenge ACK
5254 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
5255 rst_seq_match = true;
5256 } else if (tcp_is_sack(tp) && tp->rx_opt.num_sacks > 0) {
5257 struct tcp_sack_block *sp = &tp->selective_acks[0];
5258 int max_sack = sp[0].end_seq;
5261 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;
5263 max_sack = after(sp[this_sack].end_seq,
5265 sp[this_sack].end_seq : max_sack;
5268 if (TCP_SKB_CB(skb)->seq == max_sack)
5269 rst_seq_match = true;
5275 tcp_send_challenge_ack(sk, skb);
5279 /* step 3: check security and precedence [ignored] */
5281 /* step 4: Check for a SYN
5282 * RFC 5961 4.2 : Send a challenge ack
5287 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5288 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5289 tcp_send_challenge_ack(sk, skb);
5301 * TCP receive function for the ESTABLISHED state.
5303 * It is split into a fast path and a slow path. The fast path is
5305 * - A zero window was announced from us - zero window probing
5306 * is only handled properly in the slow path.
5307 * - Out of order segments arrived.
5308 * - Urgent data is expected.
5309 * - There is no buffer space left
5310 * - Unexpected TCP flags/window values/header lengths are received
5311 * (detected by checking the TCP header against pred_flags)
5312 * - Data is sent in both directions. Fast path only supports pure senders
5313 * or pure receivers (this means either the sequence number or the ack
5314 * value must stay constant)
5315 * - Unexpected TCP option.
5317 * When these conditions are not satisfied it drops into a standard
5318 * receive procedure patterned after RFC793 to handle all cases.
5319 * The first three cases are guaranteed by proper pred_flags setting,
5320 * the rest is checked inline. Fast processing is turned on in
5321 * tcp_data_queue when everything is OK.
5323 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5324 const struct tcphdr *th, unsigned int len)
5326 struct tcp_sock *tp = tcp_sk(sk);
5328 if (unlikely(!sk->sk_rx_dst))
5329 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
5331 * Header prediction.
5332 * The code loosely follows the one in the famous
5333 * "30 instruction TCP receive" Van Jacobson mail.
5335 * Van's trick is to deposit buffers into socket queue
5336 * on a device interrupt, to call tcp_recv function
5337 * on the receive process context and checksum and copy
5338 * the buffer to user space. smart...
5340 * Our current scheme is not silly either but we take the
5341 * extra cost of the net_bh soft interrupt processing...
5342 * We do checksum and copy also but from device to kernel.
5345 tp->rx_opt.saw_tstamp = 0;
5347 /* pred_flags is 0xS?10 << 16 + snd_wnd
5348 * if header_prediction is to be made
5349 * 'S' will always be tp->tcp_header_len >> 2
5350 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
5351 * turn it off (when there are holes in the receive
5352 * space for instance)
5353 * PSH flag is ignored.
5356 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5357 TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5358 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5359 int tcp_header_len = tp->tcp_header_len;
5361 /* Timestamp header prediction: tcp_header_len
5362 * is automatically equal to th->doff*4 due to pred_flags
5366 /* Check timestamp */
5367 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5368 /* No? Slow path! */
5369 if (!tcp_parse_aligned_timestamp(tp, th))
5372 /* If PAWS failed, check it more carefully in slow path */
5373 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5376 /* DO NOT update ts_recent here, if checksum fails
5377 * and timestamp was corrupted part, it will result
5378 * in a hung connection since we will drop all
5379 * future packets due to the PAWS test.
5383 if (len <= tcp_header_len) {
5384 /* Bulk data transfer: sender */
5385 if (len == tcp_header_len) {
5386 /* Predicted packet is in window by definition.
5387 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5388 * Hence, check seq<=rcv_wup reduces to:
5390 if (tcp_header_len ==
5391 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5392 tp->rcv_nxt == tp->rcv_wup)
5393 tcp_store_ts_recent(tp);
5395 /* We know that such packets are checksummed
5398 tcp_ack(sk, skb, 0);
5400 tcp_data_snd_check(sk);
5402 } else { /* Header too small */
5403 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5408 bool fragstolen = false;
5410 if (tp->ucopy.task == current &&
5411 tp->copied_seq == tp->rcv_nxt &&
5412 len - tcp_header_len <= tp->ucopy.len &&
5413 sock_owned_by_user(sk)) {
5414 __set_current_state(TASK_RUNNING);
5416 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len)) {
5417 /* Predicted packet is in window by definition.
5418 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5419 * Hence, check seq<=rcv_wup reduces to:
5421 if (tcp_header_len ==
5422 (sizeof(struct tcphdr) +
5423 TCPOLEN_TSTAMP_ALIGNED) &&
5424 tp->rcv_nxt == tp->rcv_wup)
5425 tcp_store_ts_recent(tp);
5427 tcp_rcv_rtt_measure_ts(sk, skb);
5429 __skb_pull(skb, tcp_header_len);
5430 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
5431 NET_INC_STATS(sock_net(sk),
5432 LINUX_MIB_TCPHPHITSTOUSER);
5437 if (tcp_checksum_complete(skb))
5440 if ((int)skb->truesize > sk->sk_forward_alloc)
5443 /* Predicted packet is in window by definition.
5444 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5445 * Hence, check seq<=rcv_wup reduces to:
5447 if (tcp_header_len ==
5448 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5449 tp->rcv_nxt == tp->rcv_wup)
5450 tcp_store_ts_recent(tp);
5452 tcp_rcv_rtt_measure_ts(sk, skb);
5454 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPHITS);
5456 /* Bulk data transfer: receiver */
5457 eaten = tcp_queue_rcv(sk, skb, tcp_header_len,
5461 tcp_event_data_recv(sk, skb);
5463 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5464 /* Well, only one small jumplet in fast path... */
5465 tcp_ack(sk, skb, FLAG_DATA);
5466 tcp_data_snd_check(sk);
5467 if (!inet_csk_ack_scheduled(sk))
5471 __tcp_ack_snd_check(sk, 0);
5474 kfree_skb_partial(skb, fragstolen);
5475 sk->sk_data_ready(sk);
5481 if (len < (th->doff << 2) || tcp_checksum_complete(skb))
5484 if (!th->ack && !th->rst && !th->syn)
5488 * Standard slow path.
5491 if (!tcp_validate_incoming(sk, skb, th, 1))
5495 if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0)
5498 tcp_rcv_rtt_measure_ts(sk, skb);
5500 /* Process urgent data. */
5501 tcp_urg(sk, skb, th);
5503 /* step 7: process the segment text */
5504 tcp_data_queue(sk, skb);
5506 tcp_data_snd_check(sk);
5507 tcp_ack_snd_check(sk);
5511 TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS);
5512 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
5517 EXPORT_SYMBOL(tcp_rcv_established);
5519 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5521 struct tcp_sock *tp = tcp_sk(sk);
5522 struct inet_connection_sock *icsk = inet_csk(sk);
5524 tcp_set_state(sk, TCP_ESTABLISHED);
5527 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
5528 security_inet_conn_established(sk, skb);
5531 /* Make sure socket is routed, for correct metrics. */
5532 icsk->icsk_af_ops->rebuild_header(sk);
5534 tcp_init_metrics(sk);
5536 tcp_init_congestion_control(sk);
5538 /* Prevent spurious tcp_cwnd_restart() on first data
5541 tp->lsndtime = tcp_time_stamp;
5543 tcp_init_buffer_space(sk);
5545 if (sock_flag(sk, SOCK_KEEPOPEN))
5546 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5548 if (!tp->rx_opt.snd_wscale)
5549 __tcp_fast_path_on(tp, tp->snd_wnd);
5553 if (!sock_flag(sk, SOCK_DEAD)) {
5554 sk->sk_state_change(sk);
5555 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5559 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
5560 struct tcp_fastopen_cookie *cookie)
5562 struct tcp_sock *tp = tcp_sk(sk);
5563 struct sk_buff *data = tp->syn_data ? tcp_write_queue_head(sk) : NULL;
5564 u16 mss = tp->rx_opt.mss_clamp, try_exp = 0;
5565 bool syn_drop = false;
5567 if (mss == tp->rx_opt.user_mss) {
5568 struct tcp_options_received opt;
5570 /* Get original SYNACK MSS value if user MSS sets mss_clamp */
5571 tcp_clear_options(&opt);
5572 opt.user_mss = opt.mss_clamp = 0;
5573 tcp_parse_options(synack, &opt, 0, NULL);
5574 mss = opt.mss_clamp;
5577 if (!tp->syn_fastopen) {
5578 /* Ignore an unsolicited cookie */
5580 } else if (tp->total_retrans) {
5581 /* SYN timed out and the SYN-ACK neither has a cookie nor
5582 * acknowledges data. Presumably the remote received only
5583 * the retransmitted (regular) SYNs: either the original
5584 * SYN-data or the corresponding SYN-ACK was dropped.
5586 syn_drop = (cookie->len < 0 && data);
5587 } else if (cookie->len < 0 && !tp->syn_data) {
5588 /* We requested a cookie but didn't get it. If we did not use
5589 * the (old) exp opt format then try so next time (try_exp=1).
5590 * Otherwise we go back to use the RFC7413 opt (try_exp=2).
5592 try_exp = tp->syn_fastopen_exp ? 2 : 1;
5595 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp);
5597 if (data) { /* Retransmit unacked data in SYN */
5598 tcp_for_write_queue_from(data, sk) {
5599 if (data == tcp_send_head(sk) ||
5600 __tcp_retransmit_skb(sk, data, 1))
5604 NET_INC_STATS(sock_net(sk),
5605 LINUX_MIB_TCPFASTOPENACTIVEFAIL);
5608 tp->syn_data_acked = tp->syn_data;
5609 if (tp->syn_data_acked)
5610 NET_INC_STATS(sock_net(sk),
5611 LINUX_MIB_TCPFASTOPENACTIVE);
5613 tcp_fastopen_add_skb(sk, synack);
5618 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5619 const struct tcphdr *th)
5621 struct inet_connection_sock *icsk = inet_csk(sk);
5622 struct tcp_sock *tp = tcp_sk(sk);
5623 struct tcp_fastopen_cookie foc = { .len = -1 };
5624 int saved_clamp = tp->rx_opt.mss_clamp;
5626 tcp_parse_options(skb, &tp->rx_opt, 0, &foc);
5627 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
5628 tp->rx_opt.rcv_tsecr -= tp->tsoffset;
5632 * "If the state is SYN-SENT then
5633 * first check the ACK bit
5634 * If the ACK bit is set
5635 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5636 * a reset (unless the RST bit is set, if so drop
5637 * the segment and return)"
5639 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
5640 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt))
5641 goto reset_and_undo;
5643 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5644 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5646 NET_INC_STATS(sock_net(sk),
5647 LINUX_MIB_PAWSACTIVEREJECTED);
5648 goto reset_and_undo;
5651 /* Now ACK is acceptable.
5653 * "If the RST bit is set
5654 * If the ACK was acceptable then signal the user "error:
5655 * connection reset", drop the segment, enter CLOSED state,
5656 * delete TCB, and return."
5665 * "fifth, if neither of the SYN or RST bits is set then
5666 * drop the segment and return."
5672 goto discard_and_undo;
5675 * "If the SYN bit is on ...
5676 * are acceptable then ...
5677 * (our SYN has been ACKed), change the connection
5678 * state to ESTABLISHED..."
5681 tcp_ecn_rcv_synack(tp, th);
5683 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5684 tcp_ack(sk, skb, FLAG_SLOWPATH);
5686 /* Ok.. it's good. Set up sequence numbers and
5687 * move to established.
5689 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5690 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5692 /* RFC1323: The window in SYN & SYN/ACK segments is
5695 tp->snd_wnd = ntohs(th->window);
5697 if (!tp->rx_opt.wscale_ok) {
5698 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5699 tp->window_clamp = min(tp->window_clamp, 65535U);
5702 if (tp->rx_opt.saw_tstamp) {
5703 tp->rx_opt.tstamp_ok = 1;
5704 tp->tcp_header_len =
5705 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5706 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5707 tcp_store_ts_recent(tp);
5709 tp->tcp_header_len = sizeof(struct tcphdr);
5712 if (tcp_is_sack(tp) && sysctl_tcp_fack)
5713 tcp_enable_fack(tp);
5716 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5717 tcp_initialize_rcv_mss(sk);
5719 /* Remember, tcp_poll() does not lock socket!
5720 * Change state from SYN-SENT only after copied_seq
5721 * is initialized. */
5722 tp->copied_seq = tp->rcv_nxt;
5726 tcp_finish_connect(sk, skb);
5728 if ((tp->syn_fastopen || tp->syn_data) &&
5729 tcp_rcv_fastopen_synack(sk, skb, &foc))
5732 if (sk->sk_write_pending ||
5733 icsk->icsk_accept_queue.rskq_defer_accept ||
5734 icsk->icsk_ack.pingpong) {
5735 /* Save one ACK. Data will be ready after
5736 * several ticks, if write_pending is set.
5738 * It may be deleted, but with this feature tcpdumps
5739 * look so _wonderfully_ clever, that I was not able
5740 * to stand against the temptation 8) --ANK
5742 inet_csk_schedule_ack(sk);
5743 icsk->icsk_ack.lrcvtime = tcp_time_stamp;
5744 tcp_enter_quickack_mode(sk);
5745 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5746 TCP_DELACK_MAX, TCP_RTO_MAX);
5757 /* No ACK in the segment */
5761 * "If the RST bit is set
5763 * Otherwise (no ACK) drop the segment and return."
5766 goto discard_and_undo;
5770 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5771 tcp_paws_reject(&tp->rx_opt, 0))
5772 goto discard_and_undo;
5775 /* We see SYN without ACK. It is attempt of
5776 * simultaneous connect with crossed SYNs.
5777 * Particularly, it can be connect to self.
5779 tcp_set_state(sk, TCP_SYN_RECV);
5781 if (tp->rx_opt.saw_tstamp) {
5782 tp->rx_opt.tstamp_ok = 1;
5783 tcp_store_ts_recent(tp);
5784 tp->tcp_header_len =
5785 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5787 tp->tcp_header_len = sizeof(struct tcphdr);
5790 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5791 tp->copied_seq = tp->rcv_nxt;
5792 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5794 /* RFC1323: The window in SYN & SYN/ACK segments is
5797 tp->snd_wnd = ntohs(th->window);
5798 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
5799 tp->max_window = tp->snd_wnd;
5801 tcp_ecn_rcv_syn(tp, th);
5804 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5805 tcp_initialize_rcv_mss(sk);
5807 tcp_send_synack(sk);
5809 /* Note, we could accept data and URG from this segment.
5810 * There are no obstacles to make this (except that we must
5811 * either change tcp_recvmsg() to prevent it from returning data
5812 * before 3WHS completes per RFC793, or employ TCP Fast Open).
5814 * However, if we ignore data in ACKless segments sometimes,
5815 * we have no reasons to accept it sometimes.
5816 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5817 * is not flawless. So, discard packet for sanity.
5818 * Uncomment this return to process the data.
5825 /* "fifth, if neither of the SYN or RST bits is set then
5826 * drop the segment and return."
5830 tcp_clear_options(&tp->rx_opt);
5831 tp->rx_opt.mss_clamp = saved_clamp;
5835 tcp_clear_options(&tp->rx_opt);
5836 tp->rx_opt.mss_clamp = saved_clamp;
5841 * This function implements the receiving procedure of RFC 793 for
5842 * all states except ESTABLISHED and TIME_WAIT.
5843 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5844 * address independent.
5847 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb)
5849 struct tcp_sock *tp = tcp_sk(sk);
5850 struct inet_connection_sock *icsk = inet_csk(sk);
5851 const struct tcphdr *th = tcp_hdr(skb);
5852 struct request_sock *req;
5856 switch (sk->sk_state) {
5870 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
5879 tp->rx_opt.saw_tstamp = 0;
5880 queued = tcp_rcv_synsent_state_process(sk, skb, th);
5884 /* Do step6 onward by hand. */
5885 tcp_urg(sk, skb, th);
5887 tcp_data_snd_check(sk);
5891 tp->rx_opt.saw_tstamp = 0;
5892 req = tp->fastopen_rsk;
5894 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
5895 sk->sk_state != TCP_FIN_WAIT1);
5897 if (!tcp_check_req(sk, skb, req, true))
5901 if (!th->ack && !th->rst && !th->syn)
5904 if (!tcp_validate_incoming(sk, skb, th, 0))
5907 /* step 5: check the ACK field */
5908 acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
5909 FLAG_UPDATE_TS_RECENT) > 0;
5911 switch (sk->sk_state) {
5917 tcp_synack_rtt_meas(sk, req);
5919 /* Once we leave TCP_SYN_RECV, we no longer need req
5923 tp->total_retrans = req->num_retrans;
5924 reqsk_fastopen_remove(sk, req, false);
5926 /* Make sure socket is routed, for correct metrics. */
5927 icsk->icsk_af_ops->rebuild_header(sk);
5928 tcp_init_congestion_control(sk);
5931 tp->copied_seq = tp->rcv_nxt;
5932 tcp_init_buffer_space(sk);
5935 tcp_set_state(sk, TCP_ESTABLISHED);
5936 sk->sk_state_change(sk);
5938 /* Note, that this wakeup is only for marginal crossed SYN case.
5939 * Passively open sockets are not waked up, because
5940 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
5943 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5945 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5946 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
5947 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5949 if (tp->rx_opt.tstamp_ok)
5950 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5953 /* Re-arm the timer because data may have been sent out.
5954 * This is similar to the regular data transmission case
5955 * when new data has just been ack'ed.
5957 * (TFO) - we could try to be more aggressive and
5958 * retransmitting any data sooner based on when they
5963 tcp_init_metrics(sk);
5965 tcp_update_pacing_rate(sk);
5967 /* Prevent spurious tcp_cwnd_restart() on first data packet */
5968 tp->lsndtime = tcp_time_stamp;
5970 tcp_initialize_rcv_mss(sk);
5971 tcp_fast_path_on(tp);
5974 case TCP_FIN_WAIT1: {
5975 struct dst_entry *dst;
5978 /* If we enter the TCP_FIN_WAIT1 state and we are a
5979 * Fast Open socket and this is the first acceptable
5980 * ACK we have received, this would have acknowledged
5981 * our SYNACK so stop the SYNACK timer.
5984 /* Return RST if ack_seq is invalid.
5985 * Note that RFC793 only says to generate a
5986 * DUPACK for it but for TCP Fast Open it seems
5987 * better to treat this case like TCP_SYN_RECV
5992 /* We no longer need the request sock. */
5993 reqsk_fastopen_remove(sk, req, false);
5996 if (tp->snd_una != tp->write_seq)
5999 tcp_set_state(sk, TCP_FIN_WAIT2);
6000 sk->sk_shutdown |= SEND_SHUTDOWN;
6002 dst = __sk_dst_get(sk);
6006 if (!sock_flag(sk, SOCK_DEAD)) {
6007 /* Wake up lingering close() */
6008 sk->sk_state_change(sk);
6012 if (tp->linger2 < 0 ||
6013 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6014 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
6016 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6020 tmo = tcp_fin_time(sk);
6021 if (tmo > TCP_TIMEWAIT_LEN) {
6022 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
6023 } else if (th->fin || sock_owned_by_user(sk)) {
6024 /* Bad case. We could lose such FIN otherwise.
6025 * It is not a big problem, but it looks confusing
6026 * and not so rare event. We still can lose it now,
6027 * if it spins in bh_lock_sock(), but it is really
6030 inet_csk_reset_keepalive_timer(sk, tmo);
6032 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
6039 if (tp->snd_una == tp->write_seq) {
6040 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
6046 if (tp->snd_una == tp->write_seq) {
6047 tcp_update_metrics(sk);
6054 /* step 6: check the URG bit */
6055 tcp_urg(sk, skb, th);
6057 /* step 7: process the segment text */
6058 switch (sk->sk_state) {
6059 case TCP_CLOSE_WAIT:
6062 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
6066 /* RFC 793 says to queue data in these states,
6067 * RFC 1122 says we MUST send a reset.
6068 * BSD 4.4 also does reset.
6070 if (sk->sk_shutdown & RCV_SHUTDOWN) {
6071 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
6072 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
6073 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
6079 case TCP_ESTABLISHED:
6080 tcp_data_queue(sk, skb);
6085 /* tcp_data could move socket to TIME-WAIT */
6086 if (sk->sk_state != TCP_CLOSE) {
6087 tcp_data_snd_check(sk);
6088 tcp_ack_snd_check(sk);
6097 EXPORT_SYMBOL(tcp_rcv_state_process);
6099 static inline void pr_drop_req(struct request_sock *req, __u16 port, int family)
6101 struct inet_request_sock *ireq = inet_rsk(req);
6103 if (family == AF_INET)
6104 net_dbg_ratelimited("drop open request from %pI4/%u\n",
6105 &ireq->ir_rmt_addr, port);
6106 #if IS_ENABLED(CONFIG_IPV6)
6107 else if (family == AF_INET6)
6108 net_dbg_ratelimited("drop open request from %pI6/%u\n",
6109 &ireq->ir_v6_rmt_addr, port);
6113 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set
6115 * If we receive a SYN packet with these bits set, it means a
6116 * network is playing bad games with TOS bits. In order to
6117 * avoid possible false congestion notifications, we disable
6118 * TCP ECN negotiation.
6120 * Exception: tcp_ca wants ECN. This is required for DCTCP
6121 * congestion control: Linux DCTCP asserts ECT on all packets,
6122 * including SYN, which is most optimal solution; however,
6123 * others, such as FreeBSD do not.
6125 static void tcp_ecn_create_request(struct request_sock *req,
6126 const struct sk_buff *skb,
6127 const struct sock *listen_sk,
6128 const struct dst_entry *dst)
6130 const struct tcphdr *th = tcp_hdr(skb);
6131 const struct net *net = sock_net(listen_sk);
6132 bool th_ecn = th->ece && th->cwr;
6139 ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield);
6140 ecn_ok_dst = dst_feature(dst, DST_FEATURE_ECN_MASK);
6141 ecn_ok = net->ipv4.sysctl_tcp_ecn || ecn_ok_dst;
6143 if ((!ect && ecn_ok) || tcp_ca_needs_ecn(listen_sk) ||
6144 (ecn_ok_dst & DST_FEATURE_ECN_CA))
6145 inet_rsk(req)->ecn_ok = 1;
6148 static void tcp_openreq_init(struct request_sock *req,
6149 const struct tcp_options_received *rx_opt,
6150 struct sk_buff *skb, const struct sock *sk)
6152 struct inet_request_sock *ireq = inet_rsk(req);
6154 req->rsk_rcv_wnd = 0; /* So that tcp_send_synack() knows! */
6156 tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq;
6157 tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
6158 skb_mstamp_get(&tcp_rsk(req)->snt_synack);
6159 tcp_rsk(req)->last_oow_ack_time = 0;
6160 req->mss = rx_opt->mss_clamp;
6161 req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0;
6162 ireq->tstamp_ok = rx_opt->tstamp_ok;
6163 ireq->sack_ok = rx_opt->sack_ok;
6164 ireq->snd_wscale = rx_opt->snd_wscale;
6165 ireq->wscale_ok = rx_opt->wscale_ok;
6168 ireq->ir_rmt_port = tcp_hdr(skb)->source;
6169 ireq->ir_num = ntohs(tcp_hdr(skb)->dest);
6170 ireq->ir_mark = inet_request_mark(sk, skb);
6173 struct request_sock *inet_reqsk_alloc(const struct request_sock_ops *ops,
6174 struct sock *sk_listener,
6175 bool attach_listener)
6177 struct request_sock *req = reqsk_alloc(ops, sk_listener,
6181 struct inet_request_sock *ireq = inet_rsk(req);
6183 kmemcheck_annotate_bitfield(ireq, flags);
6185 #if IS_ENABLED(CONFIG_IPV6)
6186 ireq->pktopts = NULL;
6188 atomic64_set(&ireq->ir_cookie, 0);
6189 ireq->ireq_state = TCP_NEW_SYN_RECV;
6190 write_pnet(&ireq->ireq_net, sock_net(sk_listener));
6191 ireq->ireq_family = sk_listener->sk_family;
6196 EXPORT_SYMBOL(inet_reqsk_alloc);
6199 * Return true if a syncookie should be sent
6201 static bool tcp_syn_flood_action(const struct sock *sk,
6202 const struct sk_buff *skb,
6205 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
6206 const char *msg = "Dropping request";
6207 bool want_cookie = false;
6208 struct net *net = sock_net(sk);
6210 #ifdef CONFIG_SYN_COOKIES
6211 if (net->ipv4.sysctl_tcp_syncookies) {
6212 msg = "Sending cookies";
6214 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES);
6217 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP);
6219 if (!queue->synflood_warned &&
6220 net->ipv4.sysctl_tcp_syncookies != 2 &&
6221 xchg(&queue->synflood_warned, 1) == 0)
6222 pr_info("%s: Possible SYN flooding on port %d. %s. Check SNMP counters.\n",
6223 proto, ntohs(tcp_hdr(skb)->dest), msg);
6228 static void tcp_reqsk_record_syn(const struct sock *sk,
6229 struct request_sock *req,
6230 const struct sk_buff *skb)
6232 if (tcp_sk(sk)->save_syn) {
6233 u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb);
6236 copy = kmalloc(len + sizeof(u32), GFP_ATOMIC);
6239 memcpy(©[1], skb_network_header(skb), len);
6240 req->saved_syn = copy;
6245 int tcp_conn_request(struct request_sock_ops *rsk_ops,
6246 const struct tcp_request_sock_ops *af_ops,
6247 struct sock *sk, struct sk_buff *skb)
6249 struct tcp_fastopen_cookie foc = { .len = -1 };
6250 __u32 isn = TCP_SKB_CB(skb)->tcp_tw_isn;
6251 struct tcp_options_received tmp_opt;
6252 struct tcp_sock *tp = tcp_sk(sk);
6253 struct net *net = sock_net(sk);
6254 struct sock *fastopen_sk = NULL;
6255 struct dst_entry *dst = NULL;
6256 struct request_sock *req;
6257 bool want_cookie = false;
6260 /* TW buckets are converted to open requests without
6261 * limitations, they conserve resources and peer is
6262 * evidently real one.
6264 if ((net->ipv4.sysctl_tcp_syncookies == 2 ||
6265 inet_csk_reqsk_queue_is_full(sk)) && !isn) {
6266 want_cookie = tcp_syn_flood_action(sk, skb, rsk_ops->slab_name);
6272 /* Accept backlog is full. If we have already queued enough
6273 * of warm entries in syn queue, drop request. It is better than
6274 * clogging syn queue with openreqs with exponentially increasing
6277 if (sk_acceptq_is_full(sk) && inet_csk_reqsk_queue_young(sk) > 1) {
6278 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
6282 req = inet_reqsk_alloc(rsk_ops, sk, !want_cookie);
6286 tcp_rsk(req)->af_specific = af_ops;
6288 tcp_clear_options(&tmp_opt);
6289 tmp_opt.mss_clamp = af_ops->mss_clamp;
6290 tmp_opt.user_mss = tp->rx_opt.user_mss;
6291 tcp_parse_options(skb, &tmp_opt, 0, want_cookie ? NULL : &foc);
6293 if (want_cookie && !tmp_opt.saw_tstamp)
6294 tcp_clear_options(&tmp_opt);
6296 tmp_opt.tstamp_ok = tmp_opt.saw_tstamp;
6297 tcp_openreq_init(req, &tmp_opt, skb, sk);
6299 /* Note: tcp_v6_init_req() might override ir_iif for link locals */
6300 inet_rsk(req)->ir_iif = inet_request_bound_dev_if(sk, skb);
6302 af_ops->init_req(req, sk, skb);
6304 if (security_inet_conn_request(sk, skb, req))
6307 if (!want_cookie && !isn) {
6308 /* VJ's idea. We save last timestamp seen
6309 * from the destination in peer table, when entering
6310 * state TIME-WAIT, and check against it before
6311 * accepting new connection request.
6313 * If "isn" is not zero, this request hit alive
6314 * timewait bucket, so that all the necessary checks
6315 * are made in the function processing timewait state.
6317 if (tcp_death_row.sysctl_tw_recycle) {
6320 dst = af_ops->route_req(sk, &fl, req, &strict);
6322 if (dst && strict &&
6323 !tcp_peer_is_proven(req, dst, true,
6324 tmp_opt.saw_tstamp)) {
6325 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSPASSIVEREJECTED);
6326 goto drop_and_release;
6329 /* Kill the following clause, if you dislike this way. */
6330 else if (!net->ipv4.sysctl_tcp_syncookies &&
6331 (sysctl_max_syn_backlog - inet_csk_reqsk_queue_len(sk) <
6332 (sysctl_max_syn_backlog >> 2)) &&
6333 !tcp_peer_is_proven(req, dst, false,
6334 tmp_opt.saw_tstamp)) {
6335 /* Without syncookies last quarter of
6336 * backlog is filled with destinations,
6337 * proven to be alive.
6338 * It means that we continue to communicate
6339 * to destinations, already remembered
6340 * to the moment of synflood.
6342 pr_drop_req(req, ntohs(tcp_hdr(skb)->source),
6344 goto drop_and_release;
6347 isn = af_ops->init_seq(skb);
6350 dst = af_ops->route_req(sk, &fl, req, NULL);
6355 tcp_ecn_create_request(req, skb, sk, dst);
6358 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss);
6359 req->cookie_ts = tmp_opt.tstamp_ok;
6360 if (!tmp_opt.tstamp_ok)
6361 inet_rsk(req)->ecn_ok = 0;
6364 tcp_rsk(req)->snt_isn = isn;
6365 tcp_rsk(req)->txhash = net_tx_rndhash();
6366 tcp_openreq_init_rwin(req, sk, dst);
6368 tcp_reqsk_record_syn(sk, req, skb);
6369 fastopen_sk = tcp_try_fastopen(sk, skb, req, &foc, dst);
6372 af_ops->send_synack(fastopen_sk, dst, &fl, req,
6373 &foc, TCP_SYNACK_FASTOPEN);
6374 /* Add the child socket directly into the accept queue */
6375 inet_csk_reqsk_queue_add(sk, req, fastopen_sk);
6376 sk->sk_data_ready(sk);
6377 bh_unlock_sock(fastopen_sk);
6378 sock_put(fastopen_sk);
6380 tcp_rsk(req)->tfo_listener = false;
6382 inet_csk_reqsk_queue_hash_add(sk, req, TCP_TIMEOUT_INIT);
6383 af_ops->send_synack(sk, dst, &fl, req, &foc,
6384 !want_cookie ? TCP_SYNACK_NORMAL :
6402 EXPORT_SYMBOL(tcp_conn_request);