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).
8 * Version: $Id: tcp_input.c,v 1.243 2002/02/01 22:01:04 davem Exp $
11 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
12 * Mark Evans, <evansmp@uhura.aston.ac.uk>
13 * Corey Minyard <wf-rch!minyard@relay.EU.net>
14 * Florian La Roche, <flla@stud.uni-sb.de>
15 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
16 * Linus Torvalds, <torvalds@cs.helsinki.fi>
17 * Alan Cox, <gw4pts@gw4pts.ampr.org>
18 * Matthew Dillon, <dillon@apollo.west.oic.com>
19 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
20 * Jorge Cwik, <jorge@laser.satlink.net>
25 * Pedro Roque : Fast Retransmit/Recovery.
27 * Retransmit queue handled by TCP.
28 * Better retransmit timer handling.
29 * New congestion avoidance.
33 * Eric : Fast Retransmit.
34 * Randy Scott : MSS option defines.
35 * Eric Schenk : Fixes to slow start algorithm.
36 * Eric Schenk : Yet another double ACK bug.
37 * Eric Schenk : Delayed ACK bug fixes.
38 * Eric Schenk : Floyd style fast retrans war avoidance.
39 * David S. Miller : Don't allow zero congestion window.
40 * Eric Schenk : Fix retransmitter so that it sends
41 * next packet on ack of previous packet.
42 * Andi Kleen : Moved open_request checking here
43 * and process RSTs for open_requests.
44 * Andi Kleen : Better prune_queue, and other fixes.
45 * Andrey Savochkin: Fix RTT measurements in the presence of
47 * Andrey Savochkin: Check sequence numbers correctly when
48 * removing SACKs due to in sequence incoming
50 * Andi Kleen: Make sure we never ack data there is not
51 * enough room for. Also make this condition
52 * a fatal error if it might still happen.
53 * Andi Kleen: Add tcp_measure_rcv_mss to make
54 * connections with MSS<min(MTU,ann. MSS)
55 * work without delayed acks.
56 * Andi Kleen: Process packets with PSH set in the
58 * J Hadi Salim: ECN support
61 * Panu Kuhlberg: Experimental audit of TCP (re)transmission
62 * engine. Lots of bugs are found.
63 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs
67 #include <linux/module.h>
68 #include <linux/sysctl.h>
70 #include <net/inet_common.h>
71 #include <linux/ipsec.h>
72 #include <asm/unaligned.h>
73 #include <net/netdma.h>
75 int sysctl_tcp_timestamps __read_mostly = 1;
76 int sysctl_tcp_window_scaling __read_mostly = 1;
77 int sysctl_tcp_sack __read_mostly = 1;
78 int sysctl_tcp_fack __read_mostly = 1;
79 int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH;
80 int sysctl_tcp_ecn __read_mostly;
81 int sysctl_tcp_dsack __read_mostly = 1;
82 int sysctl_tcp_app_win __read_mostly = 31;
83 int sysctl_tcp_adv_win_scale __read_mostly = 2;
85 int sysctl_tcp_stdurg __read_mostly;
86 int sysctl_tcp_rfc1337 __read_mostly;
87 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
88 int sysctl_tcp_frto __read_mostly;
89 int sysctl_tcp_nometrics_save __read_mostly;
91 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
92 int sysctl_tcp_abc __read_mostly;
94 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
95 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
96 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
97 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
98 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
99 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
100 #define FLAG_ECE 0x40 /* ECE in this ACK */
101 #define FLAG_DATA_LOST 0x80 /* SACK detected data lossage. */
102 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
104 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
105 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
106 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
107 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
109 #define IsReno(tp) ((tp)->rx_opt.sack_ok == 0)
110 #define IsFack(tp) ((tp)->rx_opt.sack_ok & 2)
111 #define IsDSack(tp) ((tp)->rx_opt.sack_ok & 4)
113 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
115 /* Adapt the MSS value used to make delayed ack decision to the
118 static void tcp_measure_rcv_mss(struct sock *sk,
119 const struct sk_buff *skb)
121 struct inet_connection_sock *icsk = inet_csk(sk);
122 const unsigned int lss = icsk->icsk_ack.last_seg_size;
125 icsk->icsk_ack.last_seg_size = 0;
127 /* skb->len may jitter because of SACKs, even if peer
128 * sends good full-sized frames.
130 len = skb_shinfo(skb)->gso_size ?: skb->len;
131 if (len >= icsk->icsk_ack.rcv_mss) {
132 icsk->icsk_ack.rcv_mss = len;
134 /* Otherwise, we make more careful check taking into account,
135 * that SACKs block is variable.
137 * "len" is invariant segment length, including TCP header.
139 len += skb->data - skb->h.raw;
140 if (len >= TCP_MIN_RCVMSS + sizeof(struct tcphdr) ||
141 /* If PSH is not set, packet should be
142 * full sized, provided peer TCP is not badly broken.
143 * This observation (if it is correct 8)) allows
144 * to handle super-low mtu links fairly.
146 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
147 !(tcp_flag_word(skb->h.th)&TCP_REMNANT))) {
148 /* Subtract also invariant (if peer is RFC compliant),
149 * tcp header plus fixed timestamp option length.
150 * Resulting "len" is MSS free of SACK jitter.
152 len -= tcp_sk(sk)->tcp_header_len;
153 icsk->icsk_ack.last_seg_size = len;
155 icsk->icsk_ack.rcv_mss = len;
159 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
160 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
161 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
165 static void tcp_incr_quickack(struct sock *sk)
167 struct inet_connection_sock *icsk = inet_csk(sk);
168 unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
172 if (quickacks > icsk->icsk_ack.quick)
173 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
176 void tcp_enter_quickack_mode(struct sock *sk)
178 struct inet_connection_sock *icsk = inet_csk(sk);
179 tcp_incr_quickack(sk);
180 icsk->icsk_ack.pingpong = 0;
181 icsk->icsk_ack.ato = TCP_ATO_MIN;
184 /* Send ACKs quickly, if "quick" count is not exhausted
185 * and the session is not interactive.
188 static inline int tcp_in_quickack_mode(const struct sock *sk)
190 const struct inet_connection_sock *icsk = inet_csk(sk);
191 return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
194 /* Buffer size and advertised window tuning.
196 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
199 static void tcp_fixup_sndbuf(struct sock *sk)
201 int sndmem = tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER + 16 +
202 sizeof(struct sk_buff);
204 if (sk->sk_sndbuf < 3 * sndmem)
205 sk->sk_sndbuf = min(3 * sndmem, sysctl_tcp_wmem[2]);
208 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
210 * All tcp_full_space() is split to two parts: "network" buffer, allocated
211 * forward and advertised in receiver window (tp->rcv_wnd) and
212 * "application buffer", required to isolate scheduling/application
213 * latencies from network.
214 * window_clamp is maximal advertised window. It can be less than
215 * tcp_full_space(), in this case tcp_full_space() - window_clamp
216 * is reserved for "application" buffer. The less window_clamp is
217 * the smoother our behaviour from viewpoint of network, but the lower
218 * throughput and the higher sensitivity of the connection to losses. 8)
220 * rcv_ssthresh is more strict window_clamp used at "slow start"
221 * phase to predict further behaviour of this connection.
222 * It is used for two goals:
223 * - to enforce header prediction at sender, even when application
224 * requires some significant "application buffer". It is check #1.
225 * - to prevent pruning of receive queue because of misprediction
226 * of receiver window. Check #2.
228 * The scheme does not work when sender sends good segments opening
229 * window and then starts to feed us spaghetti. But it should work
230 * in common situations. Otherwise, we have to rely on queue collapsing.
233 /* Slow part of check#2. */
234 static int __tcp_grow_window(const struct sock *sk, struct tcp_sock *tp,
235 const struct sk_buff *skb)
238 int truesize = tcp_win_from_space(skb->truesize)/2;
239 int window = tcp_win_from_space(sysctl_tcp_rmem[2])/2;
241 while (tp->rcv_ssthresh <= window) {
242 if (truesize <= skb->len)
243 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
251 static void tcp_grow_window(struct sock *sk, struct tcp_sock *tp,
255 if (tp->rcv_ssthresh < tp->window_clamp &&
256 (int)tp->rcv_ssthresh < tcp_space(sk) &&
257 !tcp_memory_pressure) {
260 /* Check #2. Increase window, if skb with such overhead
261 * will fit to rcvbuf in future.
263 if (tcp_win_from_space(skb->truesize) <= skb->len)
266 incr = __tcp_grow_window(sk, tp, skb);
269 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr, tp->window_clamp);
270 inet_csk(sk)->icsk_ack.quick |= 1;
275 /* 3. Tuning rcvbuf, when connection enters established state. */
277 static void tcp_fixup_rcvbuf(struct sock *sk)
279 struct tcp_sock *tp = tcp_sk(sk);
280 int rcvmem = tp->advmss + MAX_TCP_HEADER + 16 + sizeof(struct sk_buff);
282 /* Try to select rcvbuf so that 4 mss-sized segments
283 * will fit to window and corresponding skbs will fit to our rcvbuf.
284 * (was 3; 4 is minimum to allow fast retransmit to work.)
286 while (tcp_win_from_space(rcvmem) < tp->advmss)
288 if (sk->sk_rcvbuf < 4 * rcvmem)
289 sk->sk_rcvbuf = min(4 * rcvmem, sysctl_tcp_rmem[2]);
292 /* 4. Try to fixup all. It is made immediately after connection enters
295 static void tcp_init_buffer_space(struct sock *sk)
297 struct tcp_sock *tp = tcp_sk(sk);
300 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
301 tcp_fixup_rcvbuf(sk);
302 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
303 tcp_fixup_sndbuf(sk);
305 tp->rcvq_space.space = tp->rcv_wnd;
307 maxwin = tcp_full_space(sk);
309 if (tp->window_clamp >= maxwin) {
310 tp->window_clamp = maxwin;
312 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
313 tp->window_clamp = max(maxwin -
314 (maxwin >> sysctl_tcp_app_win),
318 /* Force reservation of one segment. */
319 if (sysctl_tcp_app_win &&
320 tp->window_clamp > 2 * tp->advmss &&
321 tp->window_clamp + tp->advmss > maxwin)
322 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
324 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
325 tp->snd_cwnd_stamp = tcp_time_stamp;
328 /* 5. Recalculate window clamp after socket hit its memory bounds. */
329 static void tcp_clamp_window(struct sock *sk, struct tcp_sock *tp)
331 struct inet_connection_sock *icsk = inet_csk(sk);
333 icsk->icsk_ack.quick = 0;
335 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
336 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
337 !tcp_memory_pressure &&
338 atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) {
339 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
342 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
343 tp->rcv_ssthresh = min(tp->window_clamp, 2U*tp->advmss);
347 /* Initialize RCV_MSS value.
348 * RCV_MSS is an our guess about MSS used by the peer.
349 * We haven't any direct information about the MSS.
350 * It's better to underestimate the RCV_MSS rather than overestimate.
351 * Overestimations make us ACKing less frequently than needed.
352 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
354 void tcp_initialize_rcv_mss(struct sock *sk)
356 struct tcp_sock *tp = tcp_sk(sk);
357 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
359 hint = min(hint, tp->rcv_wnd/2);
360 hint = min(hint, TCP_MIN_RCVMSS);
361 hint = max(hint, TCP_MIN_MSS);
363 inet_csk(sk)->icsk_ack.rcv_mss = hint;
366 /* Receiver "autotuning" code.
368 * The algorithm for RTT estimation w/o timestamps is based on
369 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
370 * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps>
372 * More detail on this code can be found at
373 * <http://www.psc.edu/~jheffner/senior_thesis.ps>,
374 * though this reference is out of date. A new paper
377 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
379 u32 new_sample = tp->rcv_rtt_est.rtt;
385 if (new_sample != 0) {
386 /* If we sample in larger samples in the non-timestamp
387 * case, we could grossly overestimate the RTT especially
388 * with chatty applications or bulk transfer apps which
389 * are stalled on filesystem I/O.
391 * Also, since we are only going for a minimum in the
392 * non-timestamp case, we do not smooth things out
393 * else with timestamps disabled convergence takes too
397 m -= (new_sample >> 3);
399 } else if (m < new_sample)
402 /* No previous measure. */
406 if (tp->rcv_rtt_est.rtt != new_sample)
407 tp->rcv_rtt_est.rtt = new_sample;
410 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
412 if (tp->rcv_rtt_est.time == 0)
414 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
416 tcp_rcv_rtt_update(tp,
417 jiffies - tp->rcv_rtt_est.time,
421 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
422 tp->rcv_rtt_est.time = tcp_time_stamp;
425 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, const struct sk_buff *skb)
427 struct tcp_sock *tp = tcp_sk(sk);
428 if (tp->rx_opt.rcv_tsecr &&
429 (TCP_SKB_CB(skb)->end_seq -
430 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
431 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
435 * This function should be called every time data is copied to user space.
436 * It calculates the appropriate TCP receive buffer space.
438 void tcp_rcv_space_adjust(struct sock *sk)
440 struct tcp_sock *tp = tcp_sk(sk);
444 if (tp->rcvq_space.time == 0)
447 time = tcp_time_stamp - tp->rcvq_space.time;
448 if (time < (tp->rcv_rtt_est.rtt >> 3) ||
449 tp->rcv_rtt_est.rtt == 0)
452 space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
454 space = max(tp->rcvq_space.space, space);
456 if (tp->rcvq_space.space != space) {
459 tp->rcvq_space.space = space;
461 if (sysctl_tcp_moderate_rcvbuf &&
462 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
463 int new_clamp = space;
465 /* Receive space grows, normalize in order to
466 * take into account packet headers and sk_buff
467 * structure overhead.
472 rcvmem = (tp->advmss + MAX_TCP_HEADER +
473 16 + sizeof(struct sk_buff));
474 while (tcp_win_from_space(rcvmem) < tp->advmss)
477 space = min(space, sysctl_tcp_rmem[2]);
478 if (space > sk->sk_rcvbuf) {
479 sk->sk_rcvbuf = space;
481 /* Make the window clamp follow along. */
482 tp->window_clamp = new_clamp;
488 tp->rcvq_space.seq = tp->copied_seq;
489 tp->rcvq_space.time = tcp_time_stamp;
492 /* There is something which you must keep in mind when you analyze the
493 * behavior of the tp->ato delayed ack timeout interval. When a
494 * connection starts up, we want to ack as quickly as possible. The
495 * problem is that "good" TCP's do slow start at the beginning of data
496 * transmission. The means that until we send the first few ACK's the
497 * sender will sit on his end and only queue most of his data, because
498 * he can only send snd_cwnd unacked packets at any given time. For
499 * each ACK we send, he increments snd_cwnd and transmits more of his
502 static void tcp_event_data_recv(struct sock *sk, struct tcp_sock *tp, struct sk_buff *skb)
504 struct inet_connection_sock *icsk = inet_csk(sk);
507 inet_csk_schedule_ack(sk);
509 tcp_measure_rcv_mss(sk, skb);
511 tcp_rcv_rtt_measure(tp);
513 now = tcp_time_stamp;
515 if (!icsk->icsk_ack.ato) {
516 /* The _first_ data packet received, initialize
517 * delayed ACK engine.
519 tcp_incr_quickack(sk);
520 icsk->icsk_ack.ato = TCP_ATO_MIN;
522 int m = now - icsk->icsk_ack.lrcvtime;
524 if (m <= TCP_ATO_MIN/2) {
525 /* The fastest case is the first. */
526 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
527 } else if (m < icsk->icsk_ack.ato) {
528 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
529 if (icsk->icsk_ack.ato > icsk->icsk_rto)
530 icsk->icsk_ack.ato = icsk->icsk_rto;
531 } else if (m > icsk->icsk_rto) {
532 /* Too long gap. Apparently sender failed to
533 * restart window, so that we send ACKs quickly.
535 tcp_incr_quickack(sk);
536 sk_stream_mem_reclaim(sk);
539 icsk->icsk_ack.lrcvtime = now;
541 TCP_ECN_check_ce(tp, skb);
544 tcp_grow_window(sk, tp, skb);
547 /* Called to compute a smoothed rtt estimate. The data fed to this
548 * routine either comes from timestamps, or from segments that were
549 * known _not_ to have been retransmitted [see Karn/Partridge
550 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
551 * piece by Van Jacobson.
552 * NOTE: the next three routines used to be one big routine.
553 * To save cycles in the RFC 1323 implementation it was better to break
554 * it up into three procedures. -- erics
556 static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
558 struct tcp_sock *tp = tcp_sk(sk);
559 long m = mrtt; /* RTT */
561 /* The following amusing code comes from Jacobson's
562 * article in SIGCOMM '88. Note that rtt and mdev
563 * are scaled versions of rtt and mean deviation.
564 * This is designed to be as fast as possible
565 * m stands for "measurement".
567 * On a 1990 paper the rto value is changed to:
568 * RTO = rtt + 4 * mdev
570 * Funny. This algorithm seems to be very broken.
571 * These formulae increase RTO, when it should be decreased, increase
572 * too slowly, when it should be increased quickly, decrease too quickly
573 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
574 * does not matter how to _calculate_ it. Seems, it was trap
575 * that VJ failed to avoid. 8)
580 m -= (tp->srtt >> 3); /* m is now error in rtt est */
581 tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */
583 m = -m; /* m is now abs(error) */
584 m -= (tp->mdev >> 2); /* similar update on mdev */
585 /* This is similar to one of Eifel findings.
586 * Eifel blocks mdev updates when rtt decreases.
587 * This solution is a bit different: we use finer gain
588 * for mdev in this case (alpha*beta).
589 * Like Eifel it also prevents growth of rto,
590 * but also it limits too fast rto decreases,
591 * happening in pure Eifel.
596 m -= (tp->mdev >> 2); /* similar update on mdev */
598 tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */
599 if (tp->mdev > tp->mdev_max) {
600 tp->mdev_max = tp->mdev;
601 if (tp->mdev_max > tp->rttvar)
602 tp->rttvar = tp->mdev_max;
604 if (after(tp->snd_una, tp->rtt_seq)) {
605 if (tp->mdev_max < tp->rttvar)
606 tp->rttvar -= (tp->rttvar-tp->mdev_max)>>2;
607 tp->rtt_seq = tp->snd_nxt;
608 tp->mdev_max = TCP_RTO_MIN;
611 /* no previous measure. */
612 tp->srtt = m<<3; /* take the measured time to be rtt */
613 tp->mdev = m<<1; /* make sure rto = 3*rtt */
614 tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN);
615 tp->rtt_seq = tp->snd_nxt;
619 /* Calculate rto without backoff. This is the second half of Van Jacobson's
620 * routine referred to above.
622 static inline void tcp_set_rto(struct sock *sk)
624 const struct tcp_sock *tp = tcp_sk(sk);
625 /* Old crap is replaced with new one. 8)
628 * 1. If rtt variance happened to be less 50msec, it is hallucination.
629 * It cannot be less due to utterly erratic ACK generation made
630 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
631 * to do with delayed acks, because at cwnd>2 true delack timeout
632 * is invisible. Actually, Linux-2.4 also generates erratic
633 * ACKs in some circumstances.
635 inet_csk(sk)->icsk_rto = (tp->srtt >> 3) + tp->rttvar;
637 /* 2. Fixups made earlier cannot be right.
638 * If we do not estimate RTO correctly without them,
639 * all the algo is pure shit and should be replaced
640 * with correct one. It is exactly, which we pretend to do.
644 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
645 * guarantees that rto is higher.
647 static inline void tcp_bound_rto(struct sock *sk)
649 if (inet_csk(sk)->icsk_rto > TCP_RTO_MAX)
650 inet_csk(sk)->icsk_rto = TCP_RTO_MAX;
653 /* Save metrics learned by this TCP session.
654 This function is called only, when TCP finishes successfully
655 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
657 void tcp_update_metrics(struct sock *sk)
659 struct tcp_sock *tp = tcp_sk(sk);
660 struct dst_entry *dst = __sk_dst_get(sk);
662 if (sysctl_tcp_nometrics_save)
667 if (dst && (dst->flags&DST_HOST)) {
668 const struct inet_connection_sock *icsk = inet_csk(sk);
671 if (icsk->icsk_backoff || !tp->srtt) {
672 /* This session failed to estimate rtt. Why?
673 * Probably, no packets returned in time.
676 if (!(dst_metric_locked(dst, RTAX_RTT)))
677 dst->metrics[RTAX_RTT-1] = 0;
681 m = dst_metric(dst, RTAX_RTT) - tp->srtt;
683 /* If newly calculated rtt larger than stored one,
684 * store new one. Otherwise, use EWMA. Remember,
685 * rtt overestimation is always better than underestimation.
687 if (!(dst_metric_locked(dst, RTAX_RTT))) {
689 dst->metrics[RTAX_RTT-1] = tp->srtt;
691 dst->metrics[RTAX_RTT-1] -= (m>>3);
694 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) {
698 /* Scale deviation to rttvar fixed point */
703 if (m >= dst_metric(dst, RTAX_RTTVAR))
704 dst->metrics[RTAX_RTTVAR-1] = m;
706 dst->metrics[RTAX_RTTVAR-1] -=
707 (dst->metrics[RTAX_RTTVAR-1] - m)>>2;
710 if (tp->snd_ssthresh >= 0xFFFF) {
711 /* Slow start still did not finish. */
712 if (dst_metric(dst, RTAX_SSTHRESH) &&
713 !dst_metric_locked(dst, RTAX_SSTHRESH) &&
714 (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH))
715 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1;
716 if (!dst_metric_locked(dst, RTAX_CWND) &&
717 tp->snd_cwnd > dst_metric(dst, RTAX_CWND))
718 dst->metrics[RTAX_CWND-1] = tp->snd_cwnd;
719 } else if (tp->snd_cwnd > tp->snd_ssthresh &&
720 icsk->icsk_ca_state == TCP_CA_Open) {
721 /* Cong. avoidance phase, cwnd is reliable. */
722 if (!dst_metric_locked(dst, RTAX_SSTHRESH))
723 dst->metrics[RTAX_SSTHRESH-1] =
724 max(tp->snd_cwnd >> 1, tp->snd_ssthresh);
725 if (!dst_metric_locked(dst, RTAX_CWND))
726 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_cwnd) >> 1;
728 /* Else slow start did not finish, cwnd is non-sense,
729 ssthresh may be also invalid.
731 if (!dst_metric_locked(dst, RTAX_CWND))
732 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_ssthresh) >> 1;
733 if (dst->metrics[RTAX_SSTHRESH-1] &&
734 !dst_metric_locked(dst, RTAX_SSTHRESH) &&
735 tp->snd_ssthresh > dst->metrics[RTAX_SSTHRESH-1])
736 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh;
739 if (!dst_metric_locked(dst, RTAX_REORDERING)) {
740 if (dst->metrics[RTAX_REORDERING-1] < tp->reordering &&
741 tp->reordering != sysctl_tcp_reordering)
742 dst->metrics[RTAX_REORDERING-1] = tp->reordering;
747 /* Numbers are taken from RFC2414. */
748 __u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst)
750 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
753 if (tp->mss_cache > 1460)
756 cwnd = (tp->mss_cache > 1095) ? 3 : 4;
758 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
761 /* Set slow start threshold and cwnd not falling to slow start */
762 void tcp_enter_cwr(struct sock *sk)
764 struct tcp_sock *tp = tcp_sk(sk);
766 tp->prior_ssthresh = 0;
768 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
770 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
771 tp->snd_cwnd = min(tp->snd_cwnd,
772 tcp_packets_in_flight(tp) + 1U);
773 tp->snd_cwnd_cnt = 0;
774 tp->high_seq = tp->snd_nxt;
775 tp->snd_cwnd_stamp = tcp_time_stamp;
776 TCP_ECN_queue_cwr(tp);
778 tcp_set_ca_state(sk, TCP_CA_CWR);
782 /* Initialize metrics on socket. */
784 static void tcp_init_metrics(struct sock *sk)
786 struct tcp_sock *tp = tcp_sk(sk);
787 struct dst_entry *dst = __sk_dst_get(sk);
794 if (dst_metric_locked(dst, RTAX_CWND))
795 tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND);
796 if (dst_metric(dst, RTAX_SSTHRESH)) {
797 tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH);
798 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
799 tp->snd_ssthresh = tp->snd_cwnd_clamp;
801 if (dst_metric(dst, RTAX_REORDERING) &&
802 tp->reordering != dst_metric(dst, RTAX_REORDERING)) {
803 tp->rx_opt.sack_ok &= ~2;
804 tp->reordering = dst_metric(dst, RTAX_REORDERING);
807 if (dst_metric(dst, RTAX_RTT) == 0)
810 if (!tp->srtt && dst_metric(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3))
813 /* Initial rtt is determined from SYN,SYN-ACK.
814 * The segment is small and rtt may appear much
815 * less than real one. Use per-dst memory
816 * to make it more realistic.
818 * A bit of theory. RTT is time passed after "normal" sized packet
819 * is sent until it is ACKed. In normal circumstances sending small
820 * packets force peer to delay ACKs and calculation is correct too.
821 * The algorithm is adaptive and, provided we follow specs, it
822 * NEVER underestimate RTT. BUT! If peer tries to make some clever
823 * tricks sort of "quick acks" for time long enough to decrease RTT
824 * to low value, and then abruptly stops to do it and starts to delay
825 * ACKs, wait for troubles.
827 if (dst_metric(dst, RTAX_RTT) > tp->srtt) {
828 tp->srtt = dst_metric(dst, RTAX_RTT);
829 tp->rtt_seq = tp->snd_nxt;
831 if (dst_metric(dst, RTAX_RTTVAR) > tp->mdev) {
832 tp->mdev = dst_metric(dst, RTAX_RTTVAR);
833 tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN);
837 if (inet_csk(sk)->icsk_rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp)
839 tp->snd_cwnd = tcp_init_cwnd(tp, dst);
840 tp->snd_cwnd_stamp = tcp_time_stamp;
844 /* Play conservative. If timestamps are not
845 * supported, TCP will fail to recalculate correct
846 * rtt, if initial rto is too small. FORGET ALL AND RESET!
848 if (!tp->rx_opt.saw_tstamp && tp->srtt) {
850 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT;
851 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
855 static void tcp_update_reordering(struct sock *sk, const int metric,
858 struct tcp_sock *tp = tcp_sk(sk);
859 if (metric > tp->reordering) {
860 tp->reordering = min(TCP_MAX_REORDERING, metric);
862 /* This exciting event is worth to be remembered. 8) */
864 NET_INC_STATS_BH(LINUX_MIB_TCPTSREORDER);
866 NET_INC_STATS_BH(LINUX_MIB_TCPRENOREORDER);
868 NET_INC_STATS_BH(LINUX_MIB_TCPFACKREORDER);
870 NET_INC_STATS_BH(LINUX_MIB_TCPSACKREORDER);
871 #if FASTRETRANS_DEBUG > 1
872 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n",
873 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
877 tp->undo_marker ? tp->undo_retrans : 0);
879 /* Disable FACK yet. */
880 tp->rx_opt.sack_ok &= ~2;
884 /* This procedure tags the retransmission queue when SACKs arrive.
886 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
887 * Packets in queue with these bits set are counted in variables
888 * sacked_out, retrans_out and lost_out, correspondingly.
890 * Valid combinations are:
891 * Tag InFlight Description
892 * 0 1 - orig segment is in flight.
893 * S 0 - nothing flies, orig reached receiver.
894 * L 0 - nothing flies, orig lost by net.
895 * R 2 - both orig and retransmit are in flight.
896 * L|R 1 - orig is lost, retransmit is in flight.
897 * S|R 1 - orig reached receiver, retrans is still in flight.
898 * (L|S|R is logically valid, it could occur when L|R is sacked,
899 * but it is equivalent to plain S and code short-curcuits it to S.
900 * L|S is logically invalid, it would mean -1 packet in flight 8))
902 * These 6 states form finite state machine, controlled by the following events:
903 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
904 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
905 * 3. Loss detection event of one of three flavors:
906 * A. Scoreboard estimator decided the packet is lost.
907 * A'. Reno "three dupacks" marks head of queue lost.
908 * A''. Its FACK modfication, head until snd.fack is lost.
909 * B. SACK arrives sacking data transmitted after never retransmitted
911 * C. SACK arrives sacking SND.NXT at the moment, when the
912 * segment was retransmitted.
913 * 4. D-SACK added new rule: D-SACK changes any tag to S.
915 * It is pleasant to note, that state diagram turns out to be commutative,
916 * so that we are allowed not to be bothered by order of our actions,
917 * when multiple events arrive simultaneously. (see the function below).
919 * Reordering detection.
920 * --------------------
921 * Reordering metric is maximal distance, which a packet can be displaced
922 * in packet stream. With SACKs we can estimate it:
924 * 1. SACK fills old hole and the corresponding segment was not
925 * ever retransmitted -> reordering. Alas, we cannot use it
926 * when segment was retransmitted.
927 * 2. The last flaw is solved with D-SACK. D-SACK arrives
928 * for retransmitted and already SACKed segment -> reordering..
929 * Both of these heuristics are not used in Loss state, when we cannot
930 * account for retransmits accurately.
933 tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, u32 prior_snd_una)
935 const struct inet_connection_sock *icsk = inet_csk(sk);
936 struct tcp_sock *tp = tcp_sk(sk);
937 unsigned char *ptr = ack_skb->h.raw + TCP_SKB_CB(ack_skb)->sacked;
938 struct tcp_sack_block_wire *sp = (struct tcp_sack_block_wire *)(ptr+2);
939 struct sk_buff *cached_skb;
940 int num_sacks = (ptr[1] - TCPOLEN_SACK_BASE)>>3;
941 int reord = tp->packets_out;
943 u32 lost_retrans = 0;
946 int cached_fack_count;
948 int first_sack_index;
952 prior_fackets = tp->fackets_out;
954 /* Check for D-SACK. */
955 if (before(ntohl(sp[0].start_seq), TCP_SKB_CB(ack_skb)->ack_seq)) {
957 tp->rx_opt.sack_ok |= 4;
958 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV);
959 } else if (num_sacks > 1 &&
960 !after(ntohl(sp[0].end_seq), ntohl(sp[1].end_seq)) &&
961 !before(ntohl(sp[0].start_seq), ntohl(sp[1].start_seq))) {
963 tp->rx_opt.sack_ok |= 4;
964 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV);
967 /* D-SACK for already forgotten data...
968 * Do dumb counting. */
970 !after(ntohl(sp[0].end_seq), prior_snd_una) &&
971 after(ntohl(sp[0].end_seq), tp->undo_marker))
974 /* Eliminate too old ACKs, but take into
975 * account more or less fresh ones, they can
976 * contain valid SACK info.
978 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
982 * if the only SACK change is the increase of the end_seq of
983 * the first block then only apply that SACK block
984 * and use retrans queue hinting otherwise slowpath */
986 for (i = 0; i < num_sacks; i++) {
987 __be32 start_seq = sp[i].start_seq;
988 __be32 end_seq = sp[i].end_seq;
991 if (tp->recv_sack_cache[i].start_seq != start_seq)
994 if ((tp->recv_sack_cache[i].start_seq != start_seq) ||
995 (tp->recv_sack_cache[i].end_seq != end_seq))
998 tp->recv_sack_cache[i].start_seq = start_seq;
999 tp->recv_sack_cache[i].end_seq = end_seq;
1001 /* Clear the rest of the cache sack blocks so they won't match mistakenly. */
1002 for (; i < ARRAY_SIZE(tp->recv_sack_cache); i++) {
1003 tp->recv_sack_cache[i].start_seq = 0;
1004 tp->recv_sack_cache[i].end_seq = 0;
1007 first_sack_index = 0;
1012 tp->fastpath_skb_hint = NULL;
1014 /* order SACK blocks to allow in order walk of the retrans queue */
1015 for (i = num_sacks-1; i > 0; i--) {
1016 for (j = 0; j < i; j++){
1017 if (after(ntohl(sp[j].start_seq),
1018 ntohl(sp[j+1].start_seq))){
1019 struct tcp_sack_block_wire tmp;
1025 /* Track where the first SACK block goes to */
1026 if (j == first_sack_index)
1027 first_sack_index = j+1;
1034 /* clear flag as used for different purpose in following code */
1037 /* Use SACK fastpath hint if valid */
1038 cached_skb = tp->fastpath_skb_hint;
1039 cached_fack_count = tp->fastpath_cnt_hint;
1041 cached_skb = sk->sk_write_queue.next;
1042 cached_fack_count = 0;
1045 for (i=0; i<num_sacks; i++, sp++) {
1046 struct sk_buff *skb;
1047 __u32 start_seq = ntohl(sp->start_seq);
1048 __u32 end_seq = ntohl(sp->end_seq);
1052 fack_count = cached_fack_count;
1054 /* Event "B" in the comment above. */
1055 if (after(end_seq, tp->high_seq))
1056 flag |= FLAG_DATA_LOST;
1058 sk_stream_for_retrans_queue_from(skb, sk) {
1059 int in_sack, pcount;
1063 cached_fack_count = fack_count;
1064 if (i == first_sack_index) {
1065 tp->fastpath_skb_hint = skb;
1066 tp->fastpath_cnt_hint = fack_count;
1069 /* The retransmission queue is always in order, so
1070 * we can short-circuit the walk early.
1072 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1075 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1076 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1078 pcount = tcp_skb_pcount(skb);
1080 if (pcount > 1 && !in_sack &&
1081 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1082 unsigned int pkt_len;
1084 in_sack = !after(start_seq,
1085 TCP_SKB_CB(skb)->seq);
1088 pkt_len = (start_seq -
1089 TCP_SKB_CB(skb)->seq);
1091 pkt_len = (end_seq -
1092 TCP_SKB_CB(skb)->seq);
1093 if (tcp_fragment(sk, skb, pkt_len, skb_shinfo(skb)->gso_size))
1095 pcount = tcp_skb_pcount(skb);
1098 fack_count += pcount;
1100 sacked = TCP_SKB_CB(skb)->sacked;
1102 /* Account D-SACK for retransmitted packet. */
1103 if ((dup_sack && in_sack) &&
1104 (sacked & TCPCB_RETRANS) &&
1105 after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker))
1108 /* The frame is ACKed. */
1109 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) {
1110 if (sacked&TCPCB_RETRANS) {
1111 if ((dup_sack && in_sack) &&
1112 (sacked&TCPCB_SACKED_ACKED))
1113 reord = min(fack_count, reord);
1115 /* If it was in a hole, we detected reordering. */
1116 if (fack_count < prior_fackets &&
1117 !(sacked&TCPCB_SACKED_ACKED))
1118 reord = min(fack_count, reord);
1121 /* Nothing to do; acked frame is about to be dropped. */
1125 if ((sacked&TCPCB_SACKED_RETRANS) &&
1126 after(end_seq, TCP_SKB_CB(skb)->ack_seq) &&
1127 (!lost_retrans || after(end_seq, lost_retrans)))
1128 lost_retrans = end_seq;
1133 if (!(sacked&TCPCB_SACKED_ACKED)) {
1134 if (sacked & TCPCB_SACKED_RETRANS) {
1135 /* If the segment is not tagged as lost,
1136 * we do not clear RETRANS, believing
1137 * that retransmission is still in flight.
1139 if (sacked & TCPCB_LOST) {
1140 TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1141 tp->lost_out -= tcp_skb_pcount(skb);
1142 tp->retrans_out -= tcp_skb_pcount(skb);
1144 /* clear lost hint */
1145 tp->retransmit_skb_hint = NULL;
1148 /* New sack for not retransmitted frame,
1149 * which was in hole. It is reordering.
1151 if (!(sacked & TCPCB_RETRANS) &&
1152 fack_count < prior_fackets)
1153 reord = min(fack_count, reord);
1155 if (sacked & TCPCB_LOST) {
1156 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1157 tp->lost_out -= tcp_skb_pcount(skb);
1159 /* clear lost hint */
1160 tp->retransmit_skb_hint = NULL;
1164 TCP_SKB_CB(skb)->sacked |= TCPCB_SACKED_ACKED;
1165 flag |= FLAG_DATA_SACKED;
1166 tp->sacked_out += tcp_skb_pcount(skb);
1168 if (fack_count > tp->fackets_out)
1169 tp->fackets_out = fack_count;
1171 if (dup_sack && (sacked&TCPCB_RETRANS))
1172 reord = min(fack_count, reord);
1175 /* D-SACK. We can detect redundant retransmission
1176 * in S|R and plain R frames and clear it.
1177 * undo_retrans is decreased above, L|R frames
1178 * are accounted above as well.
1181 (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS)) {
1182 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1183 tp->retrans_out -= tcp_skb_pcount(skb);
1184 tp->retransmit_skb_hint = NULL;
1189 /* Check for lost retransmit. This superb idea is
1190 * borrowed from "ratehalving". Event "C".
1191 * Later note: FACK people cheated me again 8),
1192 * we have to account for reordering! Ugly,
1195 if (lost_retrans && icsk->icsk_ca_state == TCP_CA_Recovery) {
1196 struct sk_buff *skb;
1198 sk_stream_for_retrans_queue(skb, sk) {
1199 if (after(TCP_SKB_CB(skb)->seq, lost_retrans))
1201 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1203 if ((TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) &&
1204 after(lost_retrans, TCP_SKB_CB(skb)->ack_seq) &&
1206 !before(lost_retrans,
1207 TCP_SKB_CB(skb)->ack_seq + tp->reordering *
1209 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1210 tp->retrans_out -= tcp_skb_pcount(skb);
1212 /* clear lost hint */
1213 tp->retransmit_skb_hint = NULL;
1215 if (!(TCP_SKB_CB(skb)->sacked&(TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1216 tp->lost_out += tcp_skb_pcount(skb);
1217 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1218 flag |= FLAG_DATA_SACKED;
1219 NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT);
1225 tp->left_out = tp->sacked_out + tp->lost_out;
1227 if ((reord < tp->fackets_out) && icsk->icsk_ca_state != TCP_CA_Loss)
1228 tcp_update_reordering(sk, ((tp->fackets_out + 1) - reord), 0);
1230 #if FASTRETRANS_DEBUG > 0
1231 BUG_TRAP((int)tp->sacked_out >= 0);
1232 BUG_TRAP((int)tp->lost_out >= 0);
1233 BUG_TRAP((int)tp->retrans_out >= 0);
1234 BUG_TRAP((int)tcp_packets_in_flight(tp) >= 0);
1239 /* F-RTO can only be used if these conditions are satisfied:
1240 * - there must be some unsent new data
1241 * - the advertised window should allow sending it
1243 int tcp_use_frto(const struct sock *sk)
1245 const struct tcp_sock *tp = tcp_sk(sk);
1247 return (sysctl_tcp_frto && sk->sk_send_head &&
1248 !after(TCP_SKB_CB(sk->sk_send_head)->end_seq,
1249 tp->snd_una + tp->snd_wnd));
1252 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
1253 * recovery a bit and use heuristics in tcp_process_frto() to detect if
1254 * the RTO was spurious.
1256 * Do like tcp_enter_loss() would; when RTO expires the second time it
1258 * "Reduce ssthresh if it has not yet been made inside this window."
1260 void tcp_enter_frto(struct sock *sk)
1262 const struct inet_connection_sock *icsk = inet_csk(sk);
1263 struct tcp_sock *tp = tcp_sk(sk);
1264 struct sk_buff *skb;
1266 if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) ||
1267 tp->snd_una == tp->high_seq ||
1268 ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) &&
1269 !icsk->icsk_retransmits)) {
1270 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1271 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1272 tcp_ca_event(sk, CA_EVENT_FRTO);
1275 /* Have to clear retransmission markers here to keep the bookkeeping
1276 * in shape, even though we are not yet in Loss state.
1277 * If something was really lost, it is eventually caught up
1278 * in tcp_enter_frto_loss.
1280 tp->retrans_out = 0;
1281 tp->undo_marker = tp->snd_una;
1282 tp->undo_retrans = 0;
1284 sk_stream_for_retrans_queue(skb, sk) {
1285 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1287 tcp_sync_left_out(tp);
1289 tcp_set_ca_state(sk, TCP_CA_Disorder);
1290 tp->high_seq = tp->snd_nxt;
1291 tp->frto_highmark = tp->snd_nxt;
1292 tp->frto_counter = 1;
1295 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1296 * which indicates that we should follow the traditional RTO recovery,
1297 * i.e. mark everything lost and do go-back-N retransmission.
1299 static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments)
1301 struct tcp_sock *tp = tcp_sk(sk);
1302 struct sk_buff *skb;
1307 tp->fackets_out = 0;
1309 sk_stream_for_retrans_queue(skb, sk) {
1310 cnt += tcp_skb_pcount(skb);
1311 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1312 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) {
1314 /* Do not mark those segments lost that were
1315 * forward transmitted after RTO
1317 if (!after(TCP_SKB_CB(skb)->end_seq,
1318 tp->frto_highmark)) {
1319 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1320 tp->lost_out += tcp_skb_pcount(skb);
1323 tp->sacked_out += tcp_skb_pcount(skb);
1324 tp->fackets_out = cnt;
1327 tcp_sync_left_out(tp);
1329 tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments;
1330 tp->snd_cwnd_cnt = 0;
1331 tp->snd_cwnd_stamp = tcp_time_stamp;
1332 tp->undo_marker = 0;
1333 tp->frto_counter = 0;
1335 tp->reordering = min_t(unsigned int, tp->reordering,
1336 sysctl_tcp_reordering);
1337 tcp_set_ca_state(sk, TCP_CA_Loss);
1338 tp->high_seq = tp->frto_highmark;
1339 TCP_ECN_queue_cwr(tp);
1341 clear_all_retrans_hints(tp);
1344 void tcp_clear_retrans(struct tcp_sock *tp)
1347 tp->retrans_out = 0;
1349 tp->fackets_out = 0;
1353 tp->undo_marker = 0;
1354 tp->undo_retrans = 0;
1357 /* Enter Loss state. If "how" is not zero, forget all SACK information
1358 * and reset tags completely, otherwise preserve SACKs. If receiver
1359 * dropped its ofo queue, we will know this due to reneging detection.
1361 void tcp_enter_loss(struct sock *sk, int how)
1363 const struct inet_connection_sock *icsk = inet_csk(sk);
1364 struct tcp_sock *tp = tcp_sk(sk);
1365 struct sk_buff *skb;
1368 /* Reduce ssthresh if it has not yet been made inside this window. */
1369 if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
1370 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1371 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1372 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1373 tcp_ca_event(sk, CA_EVENT_LOSS);
1376 tp->snd_cwnd_cnt = 0;
1377 tp->snd_cwnd_stamp = tcp_time_stamp;
1379 tp->bytes_acked = 0;
1380 tcp_clear_retrans(tp);
1382 /* Push undo marker, if it was plain RTO and nothing
1383 * was retransmitted. */
1385 tp->undo_marker = tp->snd_una;
1387 sk_stream_for_retrans_queue(skb, sk) {
1388 cnt += tcp_skb_pcount(skb);
1389 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS)
1390 tp->undo_marker = 0;
1391 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1392 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
1393 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1394 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1395 tp->lost_out += tcp_skb_pcount(skb);
1397 tp->sacked_out += tcp_skb_pcount(skb);
1398 tp->fackets_out = cnt;
1401 tcp_sync_left_out(tp);
1403 tp->reordering = min_t(unsigned int, tp->reordering,
1404 sysctl_tcp_reordering);
1405 tcp_set_ca_state(sk, TCP_CA_Loss);
1406 tp->high_seq = tp->snd_nxt;
1407 TCP_ECN_queue_cwr(tp);
1409 clear_all_retrans_hints(tp);
1412 static int tcp_check_sack_reneging(struct sock *sk)
1414 struct sk_buff *skb;
1416 /* If ACK arrived pointing to a remembered SACK,
1417 * it means that our remembered SACKs do not reflect
1418 * real state of receiver i.e.
1419 * receiver _host_ is heavily congested (or buggy).
1420 * Do processing similar to RTO timeout.
1422 if ((skb = skb_peek(&sk->sk_write_queue)) != NULL &&
1423 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
1424 struct inet_connection_sock *icsk = inet_csk(sk);
1425 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING);
1427 tcp_enter_loss(sk, 1);
1428 icsk->icsk_retransmits++;
1429 tcp_retransmit_skb(sk, skb_peek(&sk->sk_write_queue));
1430 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1431 icsk->icsk_rto, TCP_RTO_MAX);
1437 static inline int tcp_fackets_out(struct tcp_sock *tp)
1439 return IsReno(tp) ? tp->sacked_out+1 : tp->fackets_out;
1442 static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb)
1444 return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto);
1447 static inline int tcp_head_timedout(struct sock *sk, struct tcp_sock *tp)
1449 return tp->packets_out &&
1450 tcp_skb_timedout(sk, skb_peek(&sk->sk_write_queue));
1453 /* Linux NewReno/SACK/FACK/ECN state machine.
1454 * --------------------------------------
1456 * "Open" Normal state, no dubious events, fast path.
1457 * "Disorder" In all the respects it is "Open",
1458 * but requires a bit more attention. It is entered when
1459 * we see some SACKs or dupacks. It is split of "Open"
1460 * mainly to move some processing from fast path to slow one.
1461 * "CWR" CWND was reduced due to some Congestion Notification event.
1462 * It can be ECN, ICMP source quench, local device congestion.
1463 * "Recovery" CWND was reduced, we are fast-retransmitting.
1464 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
1466 * tcp_fastretrans_alert() is entered:
1467 * - each incoming ACK, if state is not "Open"
1468 * - when arrived ACK is unusual, namely:
1473 * Counting packets in flight is pretty simple.
1475 * in_flight = packets_out - left_out + retrans_out
1477 * packets_out is SND.NXT-SND.UNA counted in packets.
1479 * retrans_out is number of retransmitted segments.
1481 * left_out is number of segments left network, but not ACKed yet.
1483 * left_out = sacked_out + lost_out
1485 * sacked_out: Packets, which arrived to receiver out of order
1486 * and hence not ACKed. With SACKs this number is simply
1487 * amount of SACKed data. Even without SACKs
1488 * it is easy to give pretty reliable estimate of this number,
1489 * counting duplicate ACKs.
1491 * lost_out: Packets lost by network. TCP has no explicit
1492 * "loss notification" feedback from network (for now).
1493 * It means that this number can be only _guessed_.
1494 * Actually, it is the heuristics to predict lossage that
1495 * distinguishes different algorithms.
1497 * F.e. after RTO, when all the queue is considered as lost,
1498 * lost_out = packets_out and in_flight = retrans_out.
1500 * Essentially, we have now two algorithms counting
1503 * FACK: It is the simplest heuristics. As soon as we decided
1504 * that something is lost, we decide that _all_ not SACKed
1505 * packets until the most forward SACK are lost. I.e.
1506 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
1507 * It is absolutely correct estimate, if network does not reorder
1508 * packets. And it loses any connection to reality when reordering
1509 * takes place. We use FACK by default until reordering
1510 * is suspected on the path to this destination.
1512 * NewReno: when Recovery is entered, we assume that one segment
1513 * is lost (classic Reno). While we are in Recovery and
1514 * a partial ACK arrives, we assume that one more packet
1515 * is lost (NewReno). This heuristics are the same in NewReno
1518 * Imagine, that's all! Forget about all this shamanism about CWND inflation
1519 * deflation etc. CWND is real congestion window, never inflated, changes
1520 * only according to classic VJ rules.
1522 * Really tricky (and requiring careful tuning) part of algorithm
1523 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
1524 * The first determines the moment _when_ we should reduce CWND and,
1525 * hence, slow down forward transmission. In fact, it determines the moment
1526 * when we decide that hole is caused by loss, rather than by a reorder.
1528 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
1529 * holes, caused by lost packets.
1531 * And the most logically complicated part of algorithm is undo
1532 * heuristics. We detect false retransmits due to both too early
1533 * fast retransmit (reordering) and underestimated RTO, analyzing
1534 * timestamps and D-SACKs. When we detect that some segments were
1535 * retransmitted by mistake and CWND reduction was wrong, we undo
1536 * window reduction and abort recovery phase. This logic is hidden
1537 * inside several functions named tcp_try_undo_<something>.
1540 /* This function decides, when we should leave Disordered state
1541 * and enter Recovery phase, reducing congestion window.
1543 * Main question: may we further continue forward transmission
1544 * with the same cwnd?
1546 static int tcp_time_to_recover(struct sock *sk, struct tcp_sock *tp)
1550 /* Do not perform any recovery during FRTO algorithm */
1551 if (tp->frto_counter)
1554 /* Trick#1: The loss is proven. */
1558 /* Not-A-Trick#2 : Classic rule... */
1559 if (tcp_fackets_out(tp) > tp->reordering)
1562 /* Trick#3 : when we use RFC2988 timer restart, fast
1563 * retransmit can be triggered by timeout of queue head.
1565 if (tcp_head_timedout(sk, tp))
1568 /* Trick#4: It is still not OK... But will it be useful to delay
1571 packets_out = tp->packets_out;
1572 if (packets_out <= tp->reordering &&
1573 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
1574 !tcp_may_send_now(sk, tp)) {
1575 /* We have nothing to send. This connection is limited
1576 * either by receiver window or by application.
1584 /* If we receive more dupacks than we expected counting segments
1585 * in assumption of absent reordering, interpret this as reordering.
1586 * The only another reason could be bug in receiver TCP.
1588 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1590 struct tcp_sock *tp = tcp_sk(sk);
1593 holes = max(tp->lost_out, 1U);
1594 holes = min(holes, tp->packets_out);
1596 if ((tp->sacked_out + holes) > tp->packets_out) {
1597 tp->sacked_out = tp->packets_out - holes;
1598 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1602 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1604 static void tcp_add_reno_sack(struct sock *sk)
1606 struct tcp_sock *tp = tcp_sk(sk);
1608 tcp_check_reno_reordering(sk, 0);
1609 tcp_sync_left_out(tp);
1612 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1614 static void tcp_remove_reno_sacks(struct sock *sk, struct tcp_sock *tp, int acked)
1617 /* One ACK acked hole. The rest eat duplicate ACKs. */
1618 if (acked-1 >= tp->sacked_out)
1621 tp->sacked_out -= acked-1;
1623 tcp_check_reno_reordering(sk, acked);
1624 tcp_sync_left_out(tp);
1627 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1630 tp->left_out = tp->lost_out;
1633 /* Mark head of queue up as lost. */
1634 static void tcp_mark_head_lost(struct sock *sk, struct tcp_sock *tp,
1635 int packets, u32 high_seq)
1637 struct sk_buff *skb;
1640 BUG_TRAP(packets <= tp->packets_out);
1641 if (tp->lost_skb_hint) {
1642 skb = tp->lost_skb_hint;
1643 cnt = tp->lost_cnt_hint;
1645 skb = sk->sk_write_queue.next;
1649 sk_stream_for_retrans_queue_from(skb, sk) {
1650 /* TODO: do this better */
1651 /* this is not the most efficient way to do this... */
1652 tp->lost_skb_hint = skb;
1653 tp->lost_cnt_hint = cnt;
1654 cnt += tcp_skb_pcount(skb);
1655 if (cnt > packets || after(TCP_SKB_CB(skb)->end_seq, high_seq))
1657 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
1658 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1659 tp->lost_out += tcp_skb_pcount(skb);
1661 /* clear xmit_retransmit_queue hints
1662 * if this is beyond hint */
1663 if(tp->retransmit_skb_hint != NULL &&
1664 before(TCP_SKB_CB(skb)->seq,
1665 TCP_SKB_CB(tp->retransmit_skb_hint)->seq)) {
1667 tp->retransmit_skb_hint = NULL;
1671 tcp_sync_left_out(tp);
1674 /* Account newly detected lost packet(s) */
1676 static void tcp_update_scoreboard(struct sock *sk, struct tcp_sock *tp)
1679 int lost = tp->fackets_out - tp->reordering;
1682 tcp_mark_head_lost(sk, tp, lost, tp->high_seq);
1684 tcp_mark_head_lost(sk, tp, 1, tp->high_seq);
1687 /* New heuristics: it is possible only after we switched
1688 * to restart timer each time when something is ACKed.
1689 * Hence, we can detect timed out packets during fast
1690 * retransmit without falling to slow start.
1692 if (!IsReno(tp) && tcp_head_timedout(sk, tp)) {
1693 struct sk_buff *skb;
1695 skb = tp->scoreboard_skb_hint ? tp->scoreboard_skb_hint
1696 : sk->sk_write_queue.next;
1698 sk_stream_for_retrans_queue_from(skb, sk) {
1699 if (!tcp_skb_timedout(sk, skb))
1702 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
1703 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1704 tp->lost_out += tcp_skb_pcount(skb);
1706 /* clear xmit_retrans hint */
1707 if (tp->retransmit_skb_hint &&
1708 before(TCP_SKB_CB(skb)->seq,
1709 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
1711 tp->retransmit_skb_hint = NULL;
1715 tp->scoreboard_skb_hint = skb;
1717 tcp_sync_left_out(tp);
1721 /* CWND moderation, preventing bursts due to too big ACKs
1722 * in dubious situations.
1724 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
1726 tp->snd_cwnd = min(tp->snd_cwnd,
1727 tcp_packets_in_flight(tp)+tcp_max_burst(tp));
1728 tp->snd_cwnd_stamp = tcp_time_stamp;
1731 /* Lower bound on congestion window is slow start threshold
1732 * unless congestion avoidance choice decides to overide it.
1734 static inline u32 tcp_cwnd_min(const struct sock *sk)
1736 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1738 return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh;
1741 /* Decrease cwnd each second ack. */
1742 static void tcp_cwnd_down(struct sock *sk)
1744 struct tcp_sock *tp = tcp_sk(sk);
1745 int decr = tp->snd_cwnd_cnt + 1;
1747 tp->snd_cwnd_cnt = decr&1;
1750 if (decr && tp->snd_cwnd > tcp_cwnd_min(sk))
1751 tp->snd_cwnd -= decr;
1753 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp)+1);
1754 tp->snd_cwnd_stamp = tcp_time_stamp;
1757 /* Nothing was retransmitted or returned timestamp is less
1758 * than timestamp of the first retransmission.
1760 static inline int tcp_packet_delayed(struct tcp_sock *tp)
1762 return !tp->retrans_stamp ||
1763 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
1764 (__s32)(tp->rx_opt.rcv_tsecr - tp->retrans_stamp) < 0);
1767 /* Undo procedures. */
1769 #if FASTRETRANS_DEBUG > 1
1770 static void DBGUNDO(struct sock *sk, struct tcp_sock *tp, const char *msg)
1772 struct inet_sock *inet = inet_sk(sk);
1773 printk(KERN_DEBUG "Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n",
1775 NIPQUAD(inet->daddr), ntohs(inet->dport),
1776 tp->snd_cwnd, tp->left_out,
1777 tp->snd_ssthresh, tp->prior_ssthresh,
1781 #define DBGUNDO(x...) do { } while (0)
1784 static void tcp_undo_cwr(struct sock *sk, const int undo)
1786 struct tcp_sock *tp = tcp_sk(sk);
1788 if (tp->prior_ssthresh) {
1789 const struct inet_connection_sock *icsk = inet_csk(sk);
1791 if (icsk->icsk_ca_ops->undo_cwnd)
1792 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
1794 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh<<1);
1796 if (undo && tp->prior_ssthresh > tp->snd_ssthresh) {
1797 tp->snd_ssthresh = tp->prior_ssthresh;
1798 TCP_ECN_withdraw_cwr(tp);
1801 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
1803 tcp_moderate_cwnd(tp);
1804 tp->snd_cwnd_stamp = tcp_time_stamp;
1806 /* There is something screwy going on with the retrans hints after
1808 clear_all_retrans_hints(tp);
1811 static inline int tcp_may_undo(struct tcp_sock *tp)
1813 return tp->undo_marker &&
1814 (!tp->undo_retrans || tcp_packet_delayed(tp));
1817 /* People celebrate: "We love our President!" */
1818 static int tcp_try_undo_recovery(struct sock *sk, struct tcp_sock *tp)
1820 if (tcp_may_undo(tp)) {
1821 /* Happy end! We did not retransmit anything
1822 * or our original transmission succeeded.
1824 DBGUNDO(sk, tp, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
1825 tcp_undo_cwr(sk, 1);
1826 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
1827 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
1829 NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO);
1830 tp->undo_marker = 0;
1832 if (tp->snd_una == tp->high_seq && IsReno(tp)) {
1833 /* Hold old state until something *above* high_seq
1834 * is ACKed. For Reno it is MUST to prevent false
1835 * fast retransmits (RFC2582). SACK TCP is safe. */
1836 tcp_moderate_cwnd(tp);
1839 tcp_set_ca_state(sk, TCP_CA_Open);
1843 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
1844 static void tcp_try_undo_dsack(struct sock *sk, struct tcp_sock *tp)
1846 if (tp->undo_marker && !tp->undo_retrans) {
1847 DBGUNDO(sk, tp, "D-SACK");
1848 tcp_undo_cwr(sk, 1);
1849 tp->undo_marker = 0;
1850 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO);
1854 /* Undo during fast recovery after partial ACK. */
1856 static int tcp_try_undo_partial(struct sock *sk, struct tcp_sock *tp,
1859 /* Partial ACK arrived. Force Hoe's retransmit. */
1860 int failed = IsReno(tp) || tp->fackets_out>tp->reordering;
1862 if (tcp_may_undo(tp)) {
1863 /* Plain luck! Hole if filled with delayed
1864 * packet, rather than with a retransmit.
1866 if (tp->retrans_out == 0)
1867 tp->retrans_stamp = 0;
1869 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
1871 DBGUNDO(sk, tp, "Hoe");
1872 tcp_undo_cwr(sk, 0);
1873 NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO);
1875 /* So... Do not make Hoe's retransmit yet.
1876 * If the first packet was delayed, the rest
1877 * ones are most probably delayed as well.
1884 /* Undo during loss recovery after partial ACK. */
1885 static int tcp_try_undo_loss(struct sock *sk, struct tcp_sock *tp)
1887 if (tcp_may_undo(tp)) {
1888 struct sk_buff *skb;
1889 sk_stream_for_retrans_queue(skb, sk) {
1890 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1893 clear_all_retrans_hints(tp);
1895 DBGUNDO(sk, tp, "partial loss");
1897 tp->left_out = tp->sacked_out;
1898 tcp_undo_cwr(sk, 1);
1899 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
1900 inet_csk(sk)->icsk_retransmits = 0;
1901 tp->undo_marker = 0;
1903 tcp_set_ca_state(sk, TCP_CA_Open);
1909 static inline void tcp_complete_cwr(struct sock *sk)
1911 struct tcp_sock *tp = tcp_sk(sk);
1912 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
1913 tp->snd_cwnd_stamp = tcp_time_stamp;
1914 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
1917 static void tcp_try_to_open(struct sock *sk, struct tcp_sock *tp, int flag)
1919 tp->left_out = tp->sacked_out;
1921 if (tp->retrans_out == 0)
1922 tp->retrans_stamp = 0;
1927 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
1928 int state = TCP_CA_Open;
1930 if (tp->left_out || tp->retrans_out || tp->undo_marker)
1931 state = TCP_CA_Disorder;
1933 if (inet_csk(sk)->icsk_ca_state != state) {
1934 tcp_set_ca_state(sk, state);
1935 tp->high_seq = tp->snd_nxt;
1937 tcp_moderate_cwnd(tp);
1943 static void tcp_mtup_probe_failed(struct sock *sk)
1945 struct inet_connection_sock *icsk = inet_csk(sk);
1947 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
1948 icsk->icsk_mtup.probe_size = 0;
1951 static void tcp_mtup_probe_success(struct sock *sk, struct sk_buff *skb)
1953 struct tcp_sock *tp = tcp_sk(sk);
1954 struct inet_connection_sock *icsk = inet_csk(sk);
1956 /* FIXME: breaks with very large cwnd */
1957 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1958 tp->snd_cwnd = tp->snd_cwnd *
1959 tcp_mss_to_mtu(sk, tp->mss_cache) /
1960 icsk->icsk_mtup.probe_size;
1961 tp->snd_cwnd_cnt = 0;
1962 tp->snd_cwnd_stamp = tcp_time_stamp;
1963 tp->rcv_ssthresh = tcp_current_ssthresh(sk);
1965 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
1966 icsk->icsk_mtup.probe_size = 0;
1967 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
1971 /* Process an event, which can update packets-in-flight not trivially.
1972 * Main goal of this function is to calculate new estimate for left_out,
1973 * taking into account both packets sitting in receiver's buffer and
1974 * packets lost by network.
1976 * Besides that it does CWND reduction, when packet loss is detected
1977 * and changes state of machine.
1979 * It does _not_ decide what to send, it is made in function
1980 * tcp_xmit_retransmit_queue().
1983 tcp_fastretrans_alert(struct sock *sk, u32 prior_snd_una,
1984 int prior_packets, int flag)
1986 struct inet_connection_sock *icsk = inet_csk(sk);
1987 struct tcp_sock *tp = tcp_sk(sk);
1988 int is_dupack = (tp->snd_una == prior_snd_una && !(flag&FLAG_NOT_DUP));
1990 /* Some technical things:
1991 * 1. Reno does not count dupacks (sacked_out) automatically. */
1992 if (!tp->packets_out)
1994 /* 2. SACK counts snd_fack in packets inaccurately. */
1995 if (tp->sacked_out == 0)
1996 tp->fackets_out = 0;
1998 /* Now state machine starts.
1999 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2001 tp->prior_ssthresh = 0;
2003 /* B. In all the states check for reneging SACKs. */
2004 if (tp->sacked_out && tcp_check_sack_reneging(sk))
2007 /* C. Process data loss notification, provided it is valid. */
2008 if ((flag&FLAG_DATA_LOST) &&
2009 before(tp->snd_una, tp->high_seq) &&
2010 icsk->icsk_ca_state != TCP_CA_Open &&
2011 tp->fackets_out > tp->reordering) {
2012 tcp_mark_head_lost(sk, tp, tp->fackets_out-tp->reordering, tp->high_seq);
2013 NET_INC_STATS_BH(LINUX_MIB_TCPLOSS);
2016 /* D. Synchronize left_out to current state. */
2017 tcp_sync_left_out(tp);
2019 /* E. Check state exit conditions. State can be terminated
2020 * when high_seq is ACKed. */
2021 if (icsk->icsk_ca_state == TCP_CA_Open) {
2022 BUG_TRAP(tp->retrans_out == 0);
2023 tp->retrans_stamp = 0;
2024 } else if (!before(tp->snd_una, tp->high_seq)) {
2025 switch (icsk->icsk_ca_state) {
2027 icsk->icsk_retransmits = 0;
2028 if (tcp_try_undo_recovery(sk, tp))
2033 /* CWR is to be held something *above* high_seq
2034 * is ACKed for CWR bit to reach receiver. */
2035 if (tp->snd_una != tp->high_seq) {
2036 tcp_complete_cwr(sk);
2037 tcp_set_ca_state(sk, TCP_CA_Open);
2041 case TCP_CA_Disorder:
2042 tcp_try_undo_dsack(sk, tp);
2043 if (!tp->undo_marker ||
2044 /* For SACK case do not Open to allow to undo
2045 * catching for all duplicate ACKs. */
2046 IsReno(tp) || tp->snd_una != tp->high_seq) {
2047 tp->undo_marker = 0;
2048 tcp_set_ca_state(sk, TCP_CA_Open);
2052 case TCP_CA_Recovery:
2054 tcp_reset_reno_sack(tp);
2055 if (tcp_try_undo_recovery(sk, tp))
2057 tcp_complete_cwr(sk);
2062 /* F. Process state. */
2063 switch (icsk->icsk_ca_state) {
2064 case TCP_CA_Recovery:
2065 if (prior_snd_una == tp->snd_una) {
2066 if (IsReno(tp) && is_dupack)
2067 tcp_add_reno_sack(sk);
2069 int acked = prior_packets - tp->packets_out;
2071 tcp_remove_reno_sacks(sk, tp, acked);
2072 is_dupack = tcp_try_undo_partial(sk, tp, acked);
2076 if (flag&FLAG_DATA_ACKED)
2077 icsk->icsk_retransmits = 0;
2078 if (!tcp_try_undo_loss(sk, tp)) {
2079 tcp_moderate_cwnd(tp);
2080 tcp_xmit_retransmit_queue(sk);
2083 if (icsk->icsk_ca_state != TCP_CA_Open)
2085 /* Loss is undone; fall through to processing in Open state. */
2088 if (tp->snd_una != prior_snd_una)
2089 tcp_reset_reno_sack(tp);
2091 tcp_add_reno_sack(sk);
2094 if (icsk->icsk_ca_state == TCP_CA_Disorder)
2095 tcp_try_undo_dsack(sk, tp);
2097 if (!tcp_time_to_recover(sk, tp)) {
2098 tcp_try_to_open(sk, tp, flag);
2102 /* MTU probe failure: don't reduce cwnd */
2103 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2104 icsk->icsk_mtup.probe_size &&
2105 tp->snd_una == tp->mtu_probe.probe_seq_start) {
2106 tcp_mtup_probe_failed(sk);
2107 /* Restores the reduction we did in tcp_mtup_probe() */
2109 tcp_simple_retransmit(sk);
2113 /* Otherwise enter Recovery state */
2116 NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY);
2118 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY);
2120 tp->high_seq = tp->snd_nxt;
2121 tp->prior_ssthresh = 0;
2122 tp->undo_marker = tp->snd_una;
2123 tp->undo_retrans = tp->retrans_out;
2125 if (icsk->icsk_ca_state < TCP_CA_CWR) {
2126 if (!(flag&FLAG_ECE))
2127 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2128 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2129 TCP_ECN_queue_cwr(tp);
2132 tp->bytes_acked = 0;
2133 tp->snd_cwnd_cnt = 0;
2134 tcp_set_ca_state(sk, TCP_CA_Recovery);
2137 if (is_dupack || tcp_head_timedout(sk, tp))
2138 tcp_update_scoreboard(sk, tp);
2140 tcp_xmit_retransmit_queue(sk);
2143 /* Read draft-ietf-tcplw-high-performance before mucking
2144 * with this code. (Supersedes RFC1323)
2146 static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
2148 /* RTTM Rule: A TSecr value received in a segment is used to
2149 * update the averaged RTT measurement only if the segment
2150 * acknowledges some new data, i.e., only if it advances the
2151 * left edge of the send window.
2153 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2154 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
2156 * Changed: reset backoff as soon as we see the first valid sample.
2157 * If we do not, we get strongly overestimated rto. With timestamps
2158 * samples are accepted even from very old segments: f.e., when rtt=1
2159 * increases to 8, we retransmit 5 times and after 8 seconds delayed
2160 * answer arrives rto becomes 120 seconds! If at least one of segments
2161 * in window is lost... Voila. --ANK (010210)
2163 struct tcp_sock *tp = tcp_sk(sk);
2164 const __u32 seq_rtt = tcp_time_stamp - tp->rx_opt.rcv_tsecr;
2165 tcp_rtt_estimator(sk, seq_rtt);
2167 inet_csk(sk)->icsk_backoff = 0;
2171 static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
2173 /* We don't have a timestamp. Can only use
2174 * packets that are not retransmitted to determine
2175 * rtt estimates. Also, we must not reset the
2176 * backoff for rto until we get a non-retransmitted
2177 * packet. This allows us to deal with a situation
2178 * where the network delay has increased suddenly.
2179 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
2182 if (flag & FLAG_RETRANS_DATA_ACKED)
2185 tcp_rtt_estimator(sk, seq_rtt);
2187 inet_csk(sk)->icsk_backoff = 0;
2191 static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
2194 const struct tcp_sock *tp = tcp_sk(sk);
2195 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
2196 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
2197 tcp_ack_saw_tstamp(sk, flag);
2198 else if (seq_rtt >= 0)
2199 tcp_ack_no_tstamp(sk, seq_rtt, flag);
2202 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 rtt,
2203 u32 in_flight, int good)
2205 const struct inet_connection_sock *icsk = inet_csk(sk);
2206 icsk->icsk_ca_ops->cong_avoid(sk, ack, rtt, in_flight, good);
2207 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
2210 /* Restart timer after forward progress on connection.
2211 * RFC2988 recommends to restart timer to now+rto.
2214 static void tcp_ack_packets_out(struct sock *sk, struct tcp_sock *tp)
2216 if (!tp->packets_out) {
2217 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
2219 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2223 static int tcp_tso_acked(struct sock *sk, struct sk_buff *skb,
2224 __u32 now, __s32 *seq_rtt)
2226 struct tcp_sock *tp = tcp_sk(sk);
2227 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
2228 __u32 seq = tp->snd_una;
2229 __u32 packets_acked;
2232 /* If we get here, the whole TSO packet has not been
2235 BUG_ON(!after(scb->end_seq, seq));
2237 packets_acked = tcp_skb_pcount(skb);
2238 if (tcp_trim_head(sk, skb, seq - scb->seq))
2240 packets_acked -= tcp_skb_pcount(skb);
2242 if (packets_acked) {
2243 __u8 sacked = scb->sacked;
2245 acked |= FLAG_DATA_ACKED;
2247 if (sacked & TCPCB_RETRANS) {
2248 if (sacked & TCPCB_SACKED_RETRANS)
2249 tp->retrans_out -= packets_acked;
2250 acked |= FLAG_RETRANS_DATA_ACKED;
2252 } else if (*seq_rtt < 0)
2253 *seq_rtt = now - scb->when;
2254 if (sacked & TCPCB_SACKED_ACKED)
2255 tp->sacked_out -= packets_acked;
2256 if (sacked & TCPCB_LOST)
2257 tp->lost_out -= packets_acked;
2258 if (sacked & TCPCB_URG) {
2260 !before(seq, tp->snd_up))
2263 } else if (*seq_rtt < 0)
2264 *seq_rtt = now - scb->when;
2266 if (tp->fackets_out) {
2267 __u32 dval = min(tp->fackets_out, packets_acked);
2268 tp->fackets_out -= dval;
2270 tp->packets_out -= packets_acked;
2272 BUG_ON(tcp_skb_pcount(skb) == 0);
2273 BUG_ON(!before(scb->seq, scb->end_seq));
2279 static u32 tcp_usrtt(struct timeval *tv)
2283 do_gettimeofday(&now);
2284 return (now.tv_sec - tv->tv_sec) * 1000000 + (now.tv_usec - tv->tv_usec);
2287 /* Remove acknowledged frames from the retransmission queue. */
2288 static int tcp_clean_rtx_queue(struct sock *sk, __s32 *seq_rtt_p)
2290 struct tcp_sock *tp = tcp_sk(sk);
2291 const struct inet_connection_sock *icsk = inet_csk(sk);
2292 struct sk_buff *skb;
2293 __u32 now = tcp_time_stamp;
2297 void (*rtt_sample)(struct sock *sk, u32 usrtt)
2298 = icsk->icsk_ca_ops->rtt_sample;
2299 struct timeval tv = { .tv_sec = 0, .tv_usec = 0 };
2301 while ((skb = skb_peek(&sk->sk_write_queue)) &&
2302 skb != sk->sk_send_head) {
2303 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
2304 __u8 sacked = scb->sacked;
2306 /* If our packet is before the ack sequence we can
2307 * discard it as it's confirmed to have arrived at
2310 if (after(scb->end_seq, tp->snd_una)) {
2311 if (tcp_skb_pcount(skb) > 1 &&
2312 after(tp->snd_una, scb->seq))
2313 acked |= tcp_tso_acked(sk, skb,
2318 /* Initial outgoing SYN's get put onto the write_queue
2319 * just like anything else we transmit. It is not
2320 * true data, and if we misinform our callers that
2321 * this ACK acks real data, we will erroneously exit
2322 * connection startup slow start one packet too
2323 * quickly. This is severely frowned upon behavior.
2325 if (!(scb->flags & TCPCB_FLAG_SYN)) {
2326 acked |= FLAG_DATA_ACKED;
2329 acked |= FLAG_SYN_ACKED;
2330 tp->retrans_stamp = 0;
2333 /* MTU probing checks */
2334 if (icsk->icsk_mtup.probe_size) {
2335 if (!after(tp->mtu_probe.probe_seq_end, TCP_SKB_CB(skb)->end_seq)) {
2336 tcp_mtup_probe_success(sk, skb);
2341 if (sacked & TCPCB_RETRANS) {
2342 if(sacked & TCPCB_SACKED_RETRANS)
2343 tp->retrans_out -= tcp_skb_pcount(skb);
2344 acked |= FLAG_RETRANS_DATA_ACKED;
2346 } else if (seq_rtt < 0) {
2347 seq_rtt = now - scb->when;
2348 skb_get_timestamp(skb, &tv);
2350 if (sacked & TCPCB_SACKED_ACKED)
2351 tp->sacked_out -= tcp_skb_pcount(skb);
2352 if (sacked & TCPCB_LOST)
2353 tp->lost_out -= tcp_skb_pcount(skb);
2354 if (sacked & TCPCB_URG) {
2356 !before(scb->end_seq, tp->snd_up))
2359 } else if (seq_rtt < 0) {
2360 seq_rtt = now - scb->when;
2361 skb_get_timestamp(skb, &tv);
2363 tcp_dec_pcount_approx(&tp->fackets_out, skb);
2364 tcp_packets_out_dec(tp, skb);
2365 __skb_unlink(skb, &sk->sk_write_queue);
2366 sk_stream_free_skb(sk, skb);
2367 clear_all_retrans_hints(tp);
2370 if (acked&FLAG_ACKED) {
2371 tcp_ack_update_rtt(sk, acked, seq_rtt);
2372 tcp_ack_packets_out(sk, tp);
2373 if (rtt_sample && !(acked & FLAG_RETRANS_DATA_ACKED))
2374 (*rtt_sample)(sk, tcp_usrtt(&tv));
2376 if (icsk->icsk_ca_ops->pkts_acked)
2377 icsk->icsk_ca_ops->pkts_acked(sk, pkts_acked);
2380 #if FASTRETRANS_DEBUG > 0
2381 BUG_TRAP((int)tp->sacked_out >= 0);
2382 BUG_TRAP((int)tp->lost_out >= 0);
2383 BUG_TRAP((int)tp->retrans_out >= 0);
2384 if (!tp->packets_out && tp->rx_opt.sack_ok) {
2385 const struct inet_connection_sock *icsk = inet_csk(sk);
2387 printk(KERN_DEBUG "Leak l=%u %d\n",
2388 tp->lost_out, icsk->icsk_ca_state);
2391 if (tp->sacked_out) {
2392 printk(KERN_DEBUG "Leak s=%u %d\n",
2393 tp->sacked_out, icsk->icsk_ca_state);
2396 if (tp->retrans_out) {
2397 printk(KERN_DEBUG "Leak r=%u %d\n",
2398 tp->retrans_out, icsk->icsk_ca_state);
2399 tp->retrans_out = 0;
2403 *seq_rtt_p = seq_rtt;
2407 static void tcp_ack_probe(struct sock *sk)
2409 const struct tcp_sock *tp = tcp_sk(sk);
2410 struct inet_connection_sock *icsk = inet_csk(sk);
2412 /* Was it a usable window open? */
2414 if (!after(TCP_SKB_CB(sk->sk_send_head)->end_seq,
2415 tp->snd_una + tp->snd_wnd)) {
2416 icsk->icsk_backoff = 0;
2417 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
2418 /* Socket must be waked up by subsequent tcp_data_snd_check().
2419 * This function is not for random using!
2422 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2423 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2428 static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag)
2430 return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
2431 inet_csk(sk)->icsk_ca_state != TCP_CA_Open);
2434 static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag)
2436 const struct tcp_sock *tp = tcp_sk(sk);
2437 return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
2438 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
2441 /* Check that window update is acceptable.
2442 * The function assumes that snd_una<=ack<=snd_next.
2444 static inline int tcp_may_update_window(const struct tcp_sock *tp, const u32 ack,
2445 const u32 ack_seq, const u32 nwin)
2447 return (after(ack, tp->snd_una) ||
2448 after(ack_seq, tp->snd_wl1) ||
2449 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd));
2452 /* Update our send window.
2454 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
2455 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
2457 static int tcp_ack_update_window(struct sock *sk, struct tcp_sock *tp,
2458 struct sk_buff *skb, u32 ack, u32 ack_seq)
2461 u32 nwin = ntohs(skb->h.th->window);
2463 if (likely(!skb->h.th->syn))
2464 nwin <<= tp->rx_opt.snd_wscale;
2466 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
2467 flag |= FLAG_WIN_UPDATE;
2468 tcp_update_wl(tp, ack, ack_seq);
2470 if (tp->snd_wnd != nwin) {
2473 /* Note, it is the only place, where
2474 * fast path is recovered for sending TCP.
2477 tcp_fast_path_check(sk, tp);
2479 if (nwin > tp->max_window) {
2480 tp->max_window = nwin;
2481 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
2491 /* A very conservative spurious RTO response algorithm: reduce cwnd and
2492 * continue in congestion avoidance.
2494 static void tcp_conservative_spur_to_response(struct tcp_sock *tp)
2496 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2497 tp->snd_cwnd_cnt = 0;
2498 tcp_moderate_cwnd(tp);
2501 /* F-RTO spurious RTO detection algorithm (RFC4138)
2503 * F-RTO affects during two new ACKs following RTO (well, almost, see inline
2504 * comments). State (ACK number) is kept in frto_counter. When ACK advances
2505 * window (but not to or beyond highest sequence sent before RTO):
2506 * On First ACK, send two new segments out.
2507 * On Second ACK, RTO was likely spurious. Do spurious response (response
2508 * algorithm is not part of the F-RTO detection algorithm
2509 * given in RFC4138 but can be selected separately).
2510 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
2511 * and TCP falls back to conventional RTO recovery.
2513 * Rationale: if the RTO was spurious, new ACKs should arrive from the
2514 * original window even after we transmit two new data segments.
2516 * F-RTO is implemented (mainly) in four functions:
2517 * - tcp_use_frto() is used to determine if TCP is can use F-RTO
2518 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
2519 * called when tcp_use_frto() showed green light
2520 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
2521 * - tcp_enter_frto_loss() is called if there is not enough evidence
2522 * to prove that the RTO is indeed spurious. It transfers the control
2523 * from F-RTO to the conventional RTO recovery
2525 static void tcp_process_frto(struct sock *sk, u32 prior_snd_una, int flag)
2527 struct tcp_sock *tp = tcp_sk(sk);
2529 tcp_sync_left_out(tp);
2531 /* Duplicate the behavior from Loss state (fastretrans_alert) */
2532 if (flag&FLAG_DATA_ACKED)
2533 inet_csk(sk)->icsk_retransmits = 0;
2535 if (!before(tp->snd_una, tp->frto_highmark)) {
2536 tcp_enter_frto_loss(sk, tp->frto_counter + 1);
2540 /* RFC4138 shortcoming in step 2; should also have case c): ACK isn't
2541 * duplicate nor advances window, e.g., opposite dir data, winupdate
2543 if ((tp->snd_una == prior_snd_una) && (flag&FLAG_NOT_DUP) &&
2544 !(flag&FLAG_FORWARD_PROGRESS))
2547 if (!(flag&FLAG_DATA_ACKED)) {
2548 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3));
2552 if (tp->frto_counter == 1) {
2553 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
2554 tp->frto_counter = 2;
2555 } else /* frto_counter == 2 */ {
2556 tcp_conservative_spur_to_response(tp);
2557 tp->frto_counter = 0;
2561 /* This routine deals with incoming acks, but not outgoing ones. */
2562 static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag)
2564 struct inet_connection_sock *icsk = inet_csk(sk);
2565 struct tcp_sock *tp = tcp_sk(sk);
2566 u32 prior_snd_una = tp->snd_una;
2567 u32 ack_seq = TCP_SKB_CB(skb)->seq;
2568 u32 ack = TCP_SKB_CB(skb)->ack_seq;
2569 u32 prior_in_flight;
2573 /* If the ack is newer than sent or older than previous acks
2574 * then we can probably ignore it.
2576 if (after(ack, tp->snd_nxt))
2577 goto uninteresting_ack;
2579 if (before(ack, prior_snd_una))
2582 if (sysctl_tcp_abc) {
2583 if (icsk->icsk_ca_state < TCP_CA_CWR)
2584 tp->bytes_acked += ack - prior_snd_una;
2585 else if (icsk->icsk_ca_state == TCP_CA_Loss)
2586 /* we assume just one segment left network */
2587 tp->bytes_acked += min(ack - prior_snd_una, tp->mss_cache);
2590 if (!(flag&FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
2591 /* Window is constant, pure forward advance.
2592 * No more checks are required.
2593 * Note, we use the fact that SND.UNA>=SND.WL2.
2595 tcp_update_wl(tp, ack, ack_seq);
2597 flag |= FLAG_WIN_UPDATE;
2599 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
2601 NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS);
2603 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
2606 NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS);
2608 flag |= tcp_ack_update_window(sk, tp, skb, ack, ack_seq);
2610 if (TCP_SKB_CB(skb)->sacked)
2611 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
2613 if (TCP_ECN_rcv_ecn_echo(tp, skb->h.th))
2616 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
2619 /* We passed data and got it acked, remove any soft error
2620 * log. Something worked...
2622 sk->sk_err_soft = 0;
2623 tp->rcv_tstamp = tcp_time_stamp;
2624 prior_packets = tp->packets_out;
2628 prior_in_flight = tcp_packets_in_flight(tp);
2630 /* See if we can take anything off of the retransmit queue. */
2631 flag |= tcp_clean_rtx_queue(sk, &seq_rtt);
2633 if (tp->frto_counter)
2634 tcp_process_frto(sk, prior_snd_una, flag);
2636 if (tcp_ack_is_dubious(sk, flag)) {
2637 /* Advance CWND, if state allows this. */
2638 if ((flag & FLAG_DATA_ACKED) && tcp_may_raise_cwnd(sk, flag))
2639 tcp_cong_avoid(sk, ack, seq_rtt, prior_in_flight, 0);
2640 tcp_fastretrans_alert(sk, prior_snd_una, prior_packets, flag);
2642 if ((flag & FLAG_DATA_ACKED))
2643 tcp_cong_avoid(sk, ack, seq_rtt, prior_in_flight, 1);
2646 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag&FLAG_NOT_DUP))
2647 dst_confirm(sk->sk_dst_cache);
2652 icsk->icsk_probes_out = 0;
2654 /* If this ack opens up a zero window, clear backoff. It was
2655 * being used to time the probes, and is probably far higher than
2656 * it needs to be for normal retransmission.
2658 if (sk->sk_send_head)
2663 if (TCP_SKB_CB(skb)->sacked)
2664 tcp_sacktag_write_queue(sk, skb, prior_snd_una);
2667 SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
2672 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
2673 * But, this can also be called on packets in the established flow when
2674 * the fast version below fails.
2676 void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx, int estab)
2679 struct tcphdr *th = skb->h.th;
2680 int length=(th->doff*4)-sizeof(struct tcphdr);
2682 ptr = (unsigned char *)(th + 1);
2683 opt_rx->saw_tstamp = 0;
2692 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
2697 if (opsize < 2) /* "silly options" */
2699 if (opsize > length)
2700 return; /* don't parse partial options */
2703 if(opsize==TCPOLEN_MSS && th->syn && !estab) {
2704 u16 in_mss = ntohs(get_unaligned((__be16 *)ptr));
2706 if (opt_rx->user_mss && opt_rx->user_mss < in_mss)
2707 in_mss = opt_rx->user_mss;
2708 opt_rx->mss_clamp = in_mss;
2713 if(opsize==TCPOLEN_WINDOW && th->syn && !estab)
2714 if (sysctl_tcp_window_scaling) {
2715 __u8 snd_wscale = *(__u8 *) ptr;
2716 opt_rx->wscale_ok = 1;
2717 if (snd_wscale > 14) {
2719 printk(KERN_INFO "tcp_parse_options: Illegal window "
2720 "scaling value %d >14 received.\n",
2724 opt_rx->snd_wscale = snd_wscale;
2727 case TCPOPT_TIMESTAMP:
2728 if(opsize==TCPOLEN_TIMESTAMP) {
2729 if ((estab && opt_rx->tstamp_ok) ||
2730 (!estab && sysctl_tcp_timestamps)) {
2731 opt_rx->saw_tstamp = 1;
2732 opt_rx->rcv_tsval = ntohl(get_unaligned((__be32 *)ptr));
2733 opt_rx->rcv_tsecr = ntohl(get_unaligned((__be32 *)(ptr+4)));
2737 case TCPOPT_SACK_PERM:
2738 if(opsize==TCPOLEN_SACK_PERM && th->syn && !estab) {
2739 if (sysctl_tcp_sack) {
2740 opt_rx->sack_ok = 1;
2741 tcp_sack_reset(opt_rx);
2747 if((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
2748 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
2750 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
2752 #ifdef CONFIG_TCP_MD5SIG
2755 * The MD5 Hash has already been
2756 * checked (see tcp_v{4,6}_do_rcv()).
2767 /* Fast parse options. This hopes to only see timestamps.
2768 * If it is wrong it falls back on tcp_parse_options().
2770 static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th,
2771 struct tcp_sock *tp)
2773 if (th->doff == sizeof(struct tcphdr)>>2) {
2774 tp->rx_opt.saw_tstamp = 0;
2776 } else if (tp->rx_opt.tstamp_ok &&
2777 th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) {
2778 __be32 *ptr = (__be32 *)(th + 1);
2779 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
2780 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
2781 tp->rx_opt.saw_tstamp = 1;
2783 tp->rx_opt.rcv_tsval = ntohl(*ptr);
2785 tp->rx_opt.rcv_tsecr = ntohl(*ptr);
2789 tcp_parse_options(skb, &tp->rx_opt, 1);
2793 static inline void tcp_store_ts_recent(struct tcp_sock *tp)
2795 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
2796 tp->rx_opt.ts_recent_stamp = xtime.tv_sec;
2799 static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
2801 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
2802 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
2803 * extra check below makes sure this can only happen
2804 * for pure ACK frames. -DaveM
2806 * Not only, also it occurs for expired timestamps.
2809 if((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) >= 0 ||
2810 xtime.tv_sec >= tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS)
2811 tcp_store_ts_recent(tp);
2815 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
2817 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
2818 * it can pass through stack. So, the following predicate verifies that
2819 * this segment is not used for anything but congestion avoidance or
2820 * fast retransmit. Moreover, we even are able to eliminate most of such
2821 * second order effects, if we apply some small "replay" window (~RTO)
2822 * to timestamp space.
2824 * All these measures still do not guarantee that we reject wrapped ACKs
2825 * on networks with high bandwidth, when sequence space is recycled fastly,
2826 * but it guarantees that such events will be very rare and do not affect
2827 * connection seriously. This doesn't look nice, but alas, PAWS is really
2830 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
2831 * states that events when retransmit arrives after original data are rare.
2832 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
2833 * the biggest problem on large power networks even with minor reordering.
2834 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
2835 * up to bandwidth of 18Gigabit/sec. 8) ]
2838 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
2840 struct tcp_sock *tp = tcp_sk(sk);
2841 struct tcphdr *th = skb->h.th;
2842 u32 seq = TCP_SKB_CB(skb)->seq;
2843 u32 ack = TCP_SKB_CB(skb)->ack_seq;
2845 return (/* 1. Pure ACK with correct sequence number. */
2846 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
2848 /* 2. ... and duplicate ACK. */
2849 ack == tp->snd_una &&
2851 /* 3. ... and does not update window. */
2852 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
2854 /* 4. ... and sits in replay window. */
2855 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
2858 static inline int tcp_paws_discard(const struct sock *sk, const struct sk_buff *skb)
2860 const struct tcp_sock *tp = tcp_sk(sk);
2861 return ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > TCP_PAWS_WINDOW &&
2862 xtime.tv_sec < tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS &&
2863 !tcp_disordered_ack(sk, skb));
2866 /* Check segment sequence number for validity.
2868 * Segment controls are considered valid, if the segment
2869 * fits to the window after truncation to the window. Acceptability
2870 * of data (and SYN, FIN, of course) is checked separately.
2871 * See tcp_data_queue(), for example.
2873 * Also, controls (RST is main one) are accepted using RCV.WUP instead
2874 * of RCV.NXT. Peer still did not advance his SND.UNA when we
2875 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
2876 * (borrowed from freebsd)
2879 static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq)
2881 return !before(end_seq, tp->rcv_wup) &&
2882 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
2885 /* When we get a reset we do this. */
2886 static void tcp_reset(struct sock *sk)
2888 /* We want the right error as BSD sees it (and indeed as we do). */
2889 switch (sk->sk_state) {
2891 sk->sk_err = ECONNREFUSED;
2893 case TCP_CLOSE_WAIT:
2899 sk->sk_err = ECONNRESET;
2902 if (!sock_flag(sk, SOCK_DEAD))
2903 sk->sk_error_report(sk);
2909 * Process the FIN bit. This now behaves as it is supposed to work
2910 * and the FIN takes effect when it is validly part of sequence
2911 * space. Not before when we get holes.
2913 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
2914 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
2917 * If we are in FINWAIT-1, a received FIN indicates simultaneous
2918 * close and we go into CLOSING (and later onto TIME-WAIT)
2920 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
2922 static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th)
2924 struct tcp_sock *tp = tcp_sk(sk);
2926 inet_csk_schedule_ack(sk);
2928 sk->sk_shutdown |= RCV_SHUTDOWN;
2929 sock_set_flag(sk, SOCK_DONE);
2931 switch (sk->sk_state) {
2933 case TCP_ESTABLISHED:
2934 /* Move to CLOSE_WAIT */
2935 tcp_set_state(sk, TCP_CLOSE_WAIT);
2936 inet_csk(sk)->icsk_ack.pingpong = 1;
2939 case TCP_CLOSE_WAIT:
2941 /* Received a retransmission of the FIN, do
2946 /* RFC793: Remain in the LAST-ACK state. */
2950 /* This case occurs when a simultaneous close
2951 * happens, we must ack the received FIN and
2952 * enter the CLOSING state.
2955 tcp_set_state(sk, TCP_CLOSING);
2958 /* Received a FIN -- send ACK and enter TIME_WAIT. */
2960 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
2963 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
2964 * cases we should never reach this piece of code.
2966 printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n",
2967 __FUNCTION__, sk->sk_state);
2971 /* It _is_ possible, that we have something out-of-order _after_ FIN.
2972 * Probably, we should reset in this case. For now drop them.
2974 __skb_queue_purge(&tp->out_of_order_queue);
2975 if (tp->rx_opt.sack_ok)
2976 tcp_sack_reset(&tp->rx_opt);
2977 sk_stream_mem_reclaim(sk);
2979 if (!sock_flag(sk, SOCK_DEAD)) {
2980 sk->sk_state_change(sk);
2982 /* Do not send POLL_HUP for half duplex close. */
2983 if (sk->sk_shutdown == SHUTDOWN_MASK ||
2984 sk->sk_state == TCP_CLOSE)
2985 sk_wake_async(sk, 1, POLL_HUP);
2987 sk_wake_async(sk, 1, POLL_IN);
2991 static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, u32 end_seq)
2993 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
2994 if (before(seq, sp->start_seq))
2995 sp->start_seq = seq;
2996 if (after(end_seq, sp->end_seq))
2997 sp->end_seq = end_seq;
3003 static void tcp_dsack_set(struct tcp_sock *tp, u32 seq, u32 end_seq)
3005 if (tp->rx_opt.sack_ok && sysctl_tcp_dsack) {
3006 if (before(seq, tp->rcv_nxt))
3007 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT);
3009 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT);
3011 tp->rx_opt.dsack = 1;
3012 tp->duplicate_sack[0].start_seq = seq;
3013 tp->duplicate_sack[0].end_seq = end_seq;
3014 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + 1, 4 - tp->rx_opt.tstamp_ok);
3018 static void tcp_dsack_extend(struct tcp_sock *tp, u32 seq, u32 end_seq)
3020 if (!tp->rx_opt.dsack)
3021 tcp_dsack_set(tp, seq, end_seq);
3023 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
3026 static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb)
3028 struct tcp_sock *tp = tcp_sk(sk);
3030 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
3031 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3032 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
3033 tcp_enter_quickack_mode(sk);
3035 if (tp->rx_opt.sack_ok && sysctl_tcp_dsack) {
3036 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3038 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
3039 end_seq = tp->rcv_nxt;
3040 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, end_seq);
3047 /* These routines update the SACK block as out-of-order packets arrive or
3048 * in-order packets close up the sequence space.
3050 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
3053 struct tcp_sack_block *sp = &tp->selective_acks[0];
3054 struct tcp_sack_block *swalk = sp+1;
3056 /* See if the recent change to the first SACK eats into
3057 * or hits the sequence space of other SACK blocks, if so coalesce.
3059 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks; ) {
3060 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
3063 /* Zap SWALK, by moving every further SACK up by one slot.
3064 * Decrease num_sacks.
3066 tp->rx_opt.num_sacks--;
3067 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
3068 for(i=this_sack; i < tp->rx_opt.num_sacks; i++)
3072 this_sack++, swalk++;
3076 static inline void tcp_sack_swap(struct tcp_sack_block *sack1, struct tcp_sack_block *sack2)
3080 tmp = sack1->start_seq;
3081 sack1->start_seq = sack2->start_seq;
3082 sack2->start_seq = tmp;
3084 tmp = sack1->end_seq;
3085 sack1->end_seq = sack2->end_seq;
3086 sack2->end_seq = tmp;
3089 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
3091 struct tcp_sock *tp = tcp_sk(sk);
3092 struct tcp_sack_block *sp = &tp->selective_acks[0];
3093 int cur_sacks = tp->rx_opt.num_sacks;
3099 for (this_sack=0; this_sack<cur_sacks; this_sack++, sp++) {
3100 if (tcp_sack_extend(sp, seq, end_seq)) {
3101 /* Rotate this_sack to the first one. */
3102 for (; this_sack>0; this_sack--, sp--)
3103 tcp_sack_swap(sp, sp-1);
3105 tcp_sack_maybe_coalesce(tp);
3110 /* Could not find an adjacent existing SACK, build a new one,
3111 * put it at the front, and shift everyone else down. We
3112 * always know there is at least one SACK present already here.
3114 * If the sack array is full, forget about the last one.
3116 if (this_sack >= 4) {
3118 tp->rx_opt.num_sacks--;
3121 for(; this_sack > 0; this_sack--, sp--)
3125 /* Build the new head SACK, and we're done. */
3126 sp->start_seq = seq;
3127 sp->end_seq = end_seq;
3128 tp->rx_opt.num_sacks++;
3129 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
3132 /* RCV.NXT advances, some SACKs should be eaten. */
3134 static void tcp_sack_remove(struct tcp_sock *tp)
3136 struct tcp_sack_block *sp = &tp->selective_acks[0];
3137 int num_sacks = tp->rx_opt.num_sacks;
3140 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
3141 if (skb_queue_empty(&tp->out_of_order_queue)) {
3142 tp->rx_opt.num_sacks = 0;
3143 tp->rx_opt.eff_sacks = tp->rx_opt.dsack;
3147 for(this_sack = 0; this_sack < num_sacks; ) {
3148 /* Check if the start of the sack is covered by RCV.NXT. */
3149 if (!before(tp->rcv_nxt, sp->start_seq)) {
3152 /* RCV.NXT must cover all the block! */
3153 BUG_TRAP(!before(tp->rcv_nxt, sp->end_seq));
3155 /* Zap this SACK, by moving forward any other SACKS. */
3156 for (i=this_sack+1; i < num_sacks; i++)
3157 tp->selective_acks[i-1] = tp->selective_acks[i];
3164 if (num_sacks != tp->rx_opt.num_sacks) {
3165 tp->rx_opt.num_sacks = num_sacks;
3166 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
3170 /* This one checks to see if we can put data from the
3171 * out_of_order queue into the receive_queue.
3173 static void tcp_ofo_queue(struct sock *sk)
3175 struct tcp_sock *tp = tcp_sk(sk);
3176 __u32 dsack_high = tp->rcv_nxt;
3177 struct sk_buff *skb;
3179 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
3180 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
3183 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
3184 __u32 dsack = dsack_high;
3185 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
3186 dsack_high = TCP_SKB_CB(skb)->end_seq;
3187 tcp_dsack_extend(tp, TCP_SKB_CB(skb)->seq, dsack);
3190 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3191 SOCK_DEBUG(sk, "ofo packet was already received \n");
3192 __skb_unlink(skb, &tp->out_of_order_queue);
3196 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
3197 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3198 TCP_SKB_CB(skb)->end_seq);
3200 __skb_unlink(skb, &tp->out_of_order_queue);
3201 __skb_queue_tail(&sk->sk_receive_queue, skb);
3202 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3204 tcp_fin(skb, sk, skb->h.th);
3208 static int tcp_prune_queue(struct sock *sk);
3210 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
3212 struct tcphdr *th = skb->h.th;
3213 struct tcp_sock *tp = tcp_sk(sk);
3216 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
3219 __skb_pull(skb, th->doff*4);
3221 TCP_ECN_accept_cwr(tp, skb);
3223 if (tp->rx_opt.dsack) {
3224 tp->rx_opt.dsack = 0;
3225 tp->rx_opt.eff_sacks = min_t(unsigned int, tp->rx_opt.num_sacks,
3226 4 - tp->rx_opt.tstamp_ok);
3229 /* Queue data for delivery to the user.
3230 * Packets in sequence go to the receive queue.
3231 * Out of sequence packets to the out_of_order_queue.
3233 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
3234 if (tcp_receive_window(tp) == 0)
3237 /* Ok. In sequence. In window. */
3238 if (tp->ucopy.task == current &&
3239 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
3240 sock_owned_by_user(sk) && !tp->urg_data) {
3241 int chunk = min_t(unsigned int, skb->len,
3244 __set_current_state(TASK_RUNNING);
3247 if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
3248 tp->ucopy.len -= chunk;
3249 tp->copied_seq += chunk;
3250 eaten = (chunk == skb->len && !th->fin);
3251 tcp_rcv_space_adjust(sk);
3259 (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3260 !sk_stream_rmem_schedule(sk, skb))) {
3261 if (tcp_prune_queue(sk) < 0 ||
3262 !sk_stream_rmem_schedule(sk, skb))
3265 sk_stream_set_owner_r(skb, sk);
3266 __skb_queue_tail(&sk->sk_receive_queue, skb);
3268 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3270 tcp_event_data_recv(sk, tp, skb);
3272 tcp_fin(skb, sk, th);
3274 if (!skb_queue_empty(&tp->out_of_order_queue)) {
3277 /* RFC2581. 4.2. SHOULD send immediate ACK, when
3278 * gap in queue is filled.
3280 if (skb_queue_empty(&tp->out_of_order_queue))
3281 inet_csk(sk)->icsk_ack.pingpong = 0;
3284 if (tp->rx_opt.num_sacks)
3285 tcp_sack_remove(tp);
3287 tcp_fast_path_check(sk, tp);
3291 else if (!sock_flag(sk, SOCK_DEAD))
3292 sk->sk_data_ready(sk, 0);
3296 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3297 /* A retransmit, 2nd most common case. Force an immediate ack. */
3298 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
3299 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3302 tcp_enter_quickack_mode(sk);
3303 inet_csk_schedule_ack(sk);
3309 /* Out of window. F.e. zero window probe. */
3310 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
3313 tcp_enter_quickack_mode(sk);
3315 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3316 /* Partial packet, seq < rcv_next < end_seq */
3317 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
3318 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3319 TCP_SKB_CB(skb)->end_seq);
3321 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
3323 /* If window is closed, drop tail of packet. But after
3324 * remembering D-SACK for its head made in previous line.
3326 if (!tcp_receive_window(tp))
3331 TCP_ECN_check_ce(tp, skb);
3333 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3334 !sk_stream_rmem_schedule(sk, skb)) {
3335 if (tcp_prune_queue(sk) < 0 ||
3336 !sk_stream_rmem_schedule(sk, skb))
3340 /* Disable header prediction. */
3342 inet_csk_schedule_ack(sk);
3344 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
3345 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3347 sk_stream_set_owner_r(skb, sk);
3349 if (!skb_peek(&tp->out_of_order_queue)) {
3350 /* Initial out of order segment, build 1 SACK. */
3351 if (tp->rx_opt.sack_ok) {
3352 tp->rx_opt.num_sacks = 1;
3353 tp->rx_opt.dsack = 0;
3354 tp->rx_opt.eff_sacks = 1;
3355 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
3356 tp->selective_acks[0].end_seq =
3357 TCP_SKB_CB(skb)->end_seq;
3359 __skb_queue_head(&tp->out_of_order_queue,skb);
3361 struct sk_buff *skb1 = tp->out_of_order_queue.prev;
3362 u32 seq = TCP_SKB_CB(skb)->seq;
3363 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3365 if (seq == TCP_SKB_CB(skb1)->end_seq) {
3366 __skb_append(skb1, skb, &tp->out_of_order_queue);
3368 if (!tp->rx_opt.num_sacks ||
3369 tp->selective_acks[0].end_seq != seq)
3372 /* Common case: data arrive in order after hole. */
3373 tp->selective_acks[0].end_seq = end_seq;
3377 /* Find place to insert this segment. */
3379 if (!after(TCP_SKB_CB(skb1)->seq, seq))
3381 } while ((skb1 = skb1->prev) !=
3382 (struct sk_buff*)&tp->out_of_order_queue);
3384 /* Do skb overlap to previous one? */
3385 if (skb1 != (struct sk_buff*)&tp->out_of_order_queue &&
3386 before(seq, TCP_SKB_CB(skb1)->end_seq)) {
3387 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
3388 /* All the bits are present. Drop. */
3390 tcp_dsack_set(tp, seq, end_seq);
3393 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
3394 /* Partial overlap. */
3395 tcp_dsack_set(tp, seq, TCP_SKB_CB(skb1)->end_seq);
3400 __skb_insert(skb, skb1, skb1->next, &tp->out_of_order_queue);
3402 /* And clean segments covered by new one as whole. */
3403 while ((skb1 = skb->next) !=
3404 (struct sk_buff*)&tp->out_of_order_queue &&
3405 after(end_seq, TCP_SKB_CB(skb1)->seq)) {
3406 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
3407 tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, end_seq);
3410 __skb_unlink(skb1, &tp->out_of_order_queue);
3411 tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, TCP_SKB_CB(skb1)->end_seq);
3416 if (tp->rx_opt.sack_ok)
3417 tcp_sack_new_ofo_skb(sk, seq, end_seq);
3421 /* Collapse contiguous sequence of skbs head..tail with
3422 * sequence numbers start..end.
3423 * Segments with FIN/SYN are not collapsed (only because this
3427 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
3428 struct sk_buff *head, struct sk_buff *tail,
3431 struct sk_buff *skb;
3433 /* First, check that queue is collapsible and find
3434 * the point where collapsing can be useful. */
3435 for (skb = head; skb != tail; ) {
3436 /* No new bits? It is possible on ofo queue. */
3437 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
3438 struct sk_buff *next = skb->next;
3439 __skb_unlink(skb, list);
3441 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
3446 /* The first skb to collapse is:
3448 * - bloated or contains data before "start" or
3449 * overlaps to the next one.
3451 if (!skb->h.th->syn && !skb->h.th->fin &&
3452 (tcp_win_from_space(skb->truesize) > skb->len ||
3453 before(TCP_SKB_CB(skb)->seq, start) ||
3454 (skb->next != tail &&
3455 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq)))
3458 /* Decided to skip this, advance start seq. */
3459 start = TCP_SKB_CB(skb)->end_seq;
3462 if (skb == tail || skb->h.th->syn || skb->h.th->fin)
3465 while (before(start, end)) {
3466 struct sk_buff *nskb;
3467 int header = skb_headroom(skb);
3468 int copy = SKB_MAX_ORDER(header, 0);
3470 /* Too big header? This can happen with IPv6. */
3473 if (end-start < copy)
3475 nskb = alloc_skb(copy+header, GFP_ATOMIC);
3478 skb_reserve(nskb, header);
3479 memcpy(nskb->head, skb->head, header);
3480 nskb->nh.raw = nskb->head + (skb->nh.raw-skb->head);
3481 nskb->h.raw = nskb->head + (skb->h.raw-skb->head);
3482 nskb->mac.raw = nskb->head + (skb->mac.raw-skb->head);
3483 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
3484 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
3485 __skb_insert(nskb, skb->prev, skb, list);
3486 sk_stream_set_owner_r(nskb, sk);
3488 /* Copy data, releasing collapsed skbs. */
3490 int offset = start - TCP_SKB_CB(skb)->seq;
3491 int size = TCP_SKB_CB(skb)->end_seq - start;
3495 size = min(copy, size);
3496 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
3498 TCP_SKB_CB(nskb)->end_seq += size;
3502 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
3503 struct sk_buff *next = skb->next;
3504 __skb_unlink(skb, list);
3506 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
3508 if (skb == tail || skb->h.th->syn || skb->h.th->fin)
3515 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
3516 * and tcp_collapse() them until all the queue is collapsed.
3518 static void tcp_collapse_ofo_queue(struct sock *sk)
3520 struct tcp_sock *tp = tcp_sk(sk);
3521 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
3522 struct sk_buff *head;
3528 start = TCP_SKB_CB(skb)->seq;
3529 end = TCP_SKB_CB(skb)->end_seq;
3535 /* Segment is terminated when we see gap or when
3536 * we are at the end of all the queue. */
3537 if (skb == (struct sk_buff *)&tp->out_of_order_queue ||
3538 after(TCP_SKB_CB(skb)->seq, end) ||
3539 before(TCP_SKB_CB(skb)->end_seq, start)) {
3540 tcp_collapse(sk, &tp->out_of_order_queue,
3541 head, skb, start, end);
3543 if (skb == (struct sk_buff *)&tp->out_of_order_queue)
3545 /* Start new segment */
3546 start = TCP_SKB_CB(skb)->seq;
3547 end = TCP_SKB_CB(skb)->end_seq;
3549 if (before(TCP_SKB_CB(skb)->seq, start))
3550 start = TCP_SKB_CB(skb)->seq;
3551 if (after(TCP_SKB_CB(skb)->end_seq, end))
3552 end = TCP_SKB_CB(skb)->end_seq;
3557 /* Reduce allocated memory if we can, trying to get
3558 * the socket within its memory limits again.
3560 * Return less than zero if we should start dropping frames
3561 * until the socket owning process reads some of the data
3562 * to stabilize the situation.
3564 static int tcp_prune_queue(struct sock *sk)
3566 struct tcp_sock *tp = tcp_sk(sk);
3568 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
3570 NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED);
3572 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
3573 tcp_clamp_window(sk, tp);
3574 else if (tcp_memory_pressure)
3575 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
3577 tcp_collapse_ofo_queue(sk);
3578 tcp_collapse(sk, &sk->sk_receive_queue,
3579 sk->sk_receive_queue.next,
3580 (struct sk_buff*)&sk->sk_receive_queue,
3581 tp->copied_seq, tp->rcv_nxt);
3582 sk_stream_mem_reclaim(sk);
3584 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
3587 /* Collapsing did not help, destructive actions follow.
3588 * This must not ever occur. */
3590 /* First, purge the out_of_order queue. */
3591 if (!skb_queue_empty(&tp->out_of_order_queue)) {
3592 NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED);
3593 __skb_queue_purge(&tp->out_of_order_queue);
3595 /* Reset SACK state. A conforming SACK implementation will
3596 * do the same at a timeout based retransmit. When a connection
3597 * is in a sad state like this, we care only about integrity
3598 * of the connection not performance.
3600 if (tp->rx_opt.sack_ok)
3601 tcp_sack_reset(&tp->rx_opt);
3602 sk_stream_mem_reclaim(sk);
3605 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
3608 /* If we are really being abused, tell the caller to silently
3609 * drop receive data on the floor. It will get retransmitted
3610 * and hopefully then we'll have sufficient space.
3612 NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED);
3614 /* Massive buffer overcommit. */
3620 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
3621 * As additional protections, we do not touch cwnd in retransmission phases,
3622 * and if application hit its sndbuf limit recently.
3624 void tcp_cwnd_application_limited(struct sock *sk)
3626 struct tcp_sock *tp = tcp_sk(sk);
3628 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
3629 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
3630 /* Limited by application or receiver window. */
3631 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
3632 u32 win_used = max(tp->snd_cwnd_used, init_win);
3633 if (win_used < tp->snd_cwnd) {
3634 tp->snd_ssthresh = tcp_current_ssthresh(sk);
3635 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
3637 tp->snd_cwnd_used = 0;
3639 tp->snd_cwnd_stamp = tcp_time_stamp;
3642 static int tcp_should_expand_sndbuf(struct sock *sk, struct tcp_sock *tp)
3644 /* If the user specified a specific send buffer setting, do
3647 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
3650 /* If we are under global TCP memory pressure, do not expand. */
3651 if (tcp_memory_pressure)
3654 /* If we are under soft global TCP memory pressure, do not expand. */
3655 if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0])
3658 /* If we filled the congestion window, do not expand. */
3659 if (tp->packets_out >= tp->snd_cwnd)
3665 /* When incoming ACK allowed to free some skb from write_queue,
3666 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
3667 * on the exit from tcp input handler.
3669 * PROBLEM: sndbuf expansion does not work well with largesend.
3671 static void tcp_new_space(struct sock *sk)
3673 struct tcp_sock *tp = tcp_sk(sk);
3675 if (tcp_should_expand_sndbuf(sk, tp)) {
3676 int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
3677 MAX_TCP_HEADER + 16 + sizeof(struct sk_buff),
3678 demanded = max_t(unsigned int, tp->snd_cwnd,
3679 tp->reordering + 1);
3680 sndmem *= 2*demanded;
3681 if (sndmem > sk->sk_sndbuf)
3682 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
3683 tp->snd_cwnd_stamp = tcp_time_stamp;
3686 sk->sk_write_space(sk);
3689 static void tcp_check_space(struct sock *sk)
3691 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
3692 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
3693 if (sk->sk_socket &&
3694 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
3699 static inline void tcp_data_snd_check(struct sock *sk, struct tcp_sock *tp)
3701 tcp_push_pending_frames(sk, tp);
3702 tcp_check_space(sk);
3706 * Check if sending an ack is needed.
3708 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
3710 struct tcp_sock *tp = tcp_sk(sk);
3712 /* More than one full frame received... */
3713 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss
3714 /* ... and right edge of window advances far enough.
3715 * (tcp_recvmsg() will send ACK otherwise). Or...
3717 && __tcp_select_window(sk) >= tp->rcv_wnd) ||
3718 /* We ACK each frame or... */
3719 tcp_in_quickack_mode(sk) ||
3720 /* We have out of order data. */
3722 skb_peek(&tp->out_of_order_queue))) {
3723 /* Then ack it now */
3726 /* Else, send delayed ack. */
3727 tcp_send_delayed_ack(sk);
3731 static inline void tcp_ack_snd_check(struct sock *sk)
3733 if (!inet_csk_ack_scheduled(sk)) {
3734 /* We sent a data segment already. */
3737 __tcp_ack_snd_check(sk, 1);
3741 * This routine is only called when we have urgent data
3742 * signaled. Its the 'slow' part of tcp_urg. It could be
3743 * moved inline now as tcp_urg is only called from one
3744 * place. We handle URGent data wrong. We have to - as
3745 * BSD still doesn't use the correction from RFC961.
3746 * For 1003.1g we should support a new option TCP_STDURG to permit
3747 * either form (or just set the sysctl tcp_stdurg).
3750 static void tcp_check_urg(struct sock * sk, struct tcphdr * th)
3752 struct tcp_sock *tp = tcp_sk(sk);
3753 u32 ptr = ntohs(th->urg_ptr);
3755 if (ptr && !sysctl_tcp_stdurg)
3757 ptr += ntohl(th->seq);
3759 /* Ignore urgent data that we've already seen and read. */
3760 if (after(tp->copied_seq, ptr))
3763 /* Do not replay urg ptr.
3765 * NOTE: interesting situation not covered by specs.
3766 * Misbehaving sender may send urg ptr, pointing to segment,
3767 * which we already have in ofo queue. We are not able to fetch
3768 * such data and will stay in TCP_URG_NOTYET until will be eaten
3769 * by recvmsg(). Seems, we are not obliged to handle such wicked
3770 * situations. But it is worth to think about possibility of some
3771 * DoSes using some hypothetical application level deadlock.
3773 if (before(ptr, tp->rcv_nxt))
3776 /* Do we already have a newer (or duplicate) urgent pointer? */
3777 if (tp->urg_data && !after(ptr, tp->urg_seq))
3780 /* Tell the world about our new urgent pointer. */
3783 /* We may be adding urgent data when the last byte read was
3784 * urgent. To do this requires some care. We cannot just ignore
3785 * tp->copied_seq since we would read the last urgent byte again
3786 * as data, nor can we alter copied_seq until this data arrives
3787 * or we break the semantics of SIOCATMARK (and thus sockatmark())
3789 * NOTE. Double Dutch. Rendering to plain English: author of comment
3790 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
3791 * and expect that both A and B disappear from stream. This is _wrong_.
3792 * Though this happens in BSD with high probability, this is occasional.
3793 * Any application relying on this is buggy. Note also, that fix "works"
3794 * only in this artificial test. Insert some normal data between A and B and we will
3795 * decline of BSD again. Verdict: it is better to remove to trap
3798 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
3799 !sock_flag(sk, SOCK_URGINLINE) &&
3800 tp->copied_seq != tp->rcv_nxt) {
3801 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
3803 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
3804 __skb_unlink(skb, &sk->sk_receive_queue);
3809 tp->urg_data = TCP_URG_NOTYET;
3812 /* Disable header prediction. */
3816 /* This is the 'fast' part of urgent handling. */
3817 static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th)
3819 struct tcp_sock *tp = tcp_sk(sk);
3821 /* Check if we get a new urgent pointer - normally not. */
3823 tcp_check_urg(sk,th);
3825 /* Do we wait for any urgent data? - normally not... */
3826 if (tp->urg_data == TCP_URG_NOTYET) {
3827 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
3830 /* Is the urgent pointer pointing into this packet? */
3831 if (ptr < skb->len) {
3833 if (skb_copy_bits(skb, ptr, &tmp, 1))
3835 tp->urg_data = TCP_URG_VALID | tmp;
3836 if (!sock_flag(sk, SOCK_DEAD))
3837 sk->sk_data_ready(sk, 0);
3842 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
3844 struct tcp_sock *tp = tcp_sk(sk);
3845 int chunk = skb->len - hlen;
3849 if (skb->ip_summed==CHECKSUM_UNNECESSARY)
3850 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
3852 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
3856 tp->ucopy.len -= chunk;
3857 tp->copied_seq += chunk;
3858 tcp_rcv_space_adjust(sk);
3865 static __sum16 __tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
3869 if (sock_owned_by_user(sk)) {
3871 result = __tcp_checksum_complete(skb);
3874 result = __tcp_checksum_complete(skb);
3879 static inline int tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
3881 return skb->ip_summed != CHECKSUM_UNNECESSARY &&
3882 __tcp_checksum_complete_user(sk, skb);
3885 #ifdef CONFIG_NET_DMA
3886 static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb, int hlen)
3888 struct tcp_sock *tp = tcp_sk(sk);
3889 int chunk = skb->len - hlen;
3891 int copied_early = 0;
3893 if (tp->ucopy.wakeup)
3896 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
3897 tp->ucopy.dma_chan = get_softnet_dma();
3899 if (tp->ucopy.dma_chan && skb->ip_summed == CHECKSUM_UNNECESSARY) {
3901 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
3902 skb, hlen, tp->ucopy.iov, chunk, tp->ucopy.pinned_list);
3907 tp->ucopy.dma_cookie = dma_cookie;
3910 tp->ucopy.len -= chunk;
3911 tp->copied_seq += chunk;
3912 tcp_rcv_space_adjust(sk);
3914 if ((tp->ucopy.len == 0) ||
3915 (tcp_flag_word(skb->h.th) & TCP_FLAG_PSH) ||
3916 (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
3917 tp->ucopy.wakeup = 1;
3918 sk->sk_data_ready(sk, 0);
3920 } else if (chunk > 0) {
3921 tp->ucopy.wakeup = 1;
3922 sk->sk_data_ready(sk, 0);
3925 return copied_early;
3927 #endif /* CONFIG_NET_DMA */
3930 * TCP receive function for the ESTABLISHED state.
3932 * It is split into a fast path and a slow path. The fast path is
3934 * - A zero window was announced from us - zero window probing
3935 * is only handled properly in the slow path.
3936 * - Out of order segments arrived.
3937 * - Urgent data is expected.
3938 * - There is no buffer space left
3939 * - Unexpected TCP flags/window values/header lengths are received
3940 * (detected by checking the TCP header against pred_flags)
3941 * - Data is sent in both directions. Fast path only supports pure senders
3942 * or pure receivers (this means either the sequence number or the ack
3943 * value must stay constant)
3944 * - Unexpected TCP option.
3946 * When these conditions are not satisfied it drops into a standard
3947 * receive procedure patterned after RFC793 to handle all cases.
3948 * The first three cases are guaranteed by proper pred_flags setting,
3949 * the rest is checked inline. Fast processing is turned on in
3950 * tcp_data_queue when everything is OK.
3952 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
3953 struct tcphdr *th, unsigned len)
3955 struct tcp_sock *tp = tcp_sk(sk);
3958 * Header prediction.
3959 * The code loosely follows the one in the famous
3960 * "30 instruction TCP receive" Van Jacobson mail.
3962 * Van's trick is to deposit buffers into socket queue
3963 * on a device interrupt, to call tcp_recv function
3964 * on the receive process context and checksum and copy
3965 * the buffer to user space. smart...
3967 * Our current scheme is not silly either but we take the
3968 * extra cost of the net_bh soft interrupt processing...
3969 * We do checksum and copy also but from device to kernel.
3972 tp->rx_opt.saw_tstamp = 0;
3974 /* pred_flags is 0xS?10 << 16 + snd_wnd
3975 * if header_prediction is to be made
3976 * 'S' will always be tp->tcp_header_len >> 2
3977 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
3978 * turn it off (when there are holes in the receive
3979 * space for instance)
3980 * PSH flag is ignored.
3983 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
3984 TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
3985 int tcp_header_len = tp->tcp_header_len;
3987 /* Timestamp header prediction: tcp_header_len
3988 * is automatically equal to th->doff*4 due to pred_flags
3992 /* Check timestamp */
3993 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
3994 __be32 *ptr = (__be32 *)(th + 1);
3996 /* No? Slow path! */
3997 if (*ptr != htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3998 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP))
4001 tp->rx_opt.saw_tstamp = 1;
4003 tp->rx_opt.rcv_tsval = ntohl(*ptr);
4005 tp->rx_opt.rcv_tsecr = ntohl(*ptr);
4007 /* If PAWS failed, check it more carefully in slow path */
4008 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
4011 /* DO NOT update ts_recent here, if checksum fails
4012 * and timestamp was corrupted part, it will result
4013 * in a hung connection since we will drop all
4014 * future packets due to the PAWS test.
4018 if (len <= tcp_header_len) {
4019 /* Bulk data transfer: sender */
4020 if (len == tcp_header_len) {
4021 /* Predicted packet is in window by definition.
4022 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4023 * Hence, check seq<=rcv_wup reduces to:
4025 if (tcp_header_len ==
4026 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
4027 tp->rcv_nxt == tp->rcv_wup)
4028 tcp_store_ts_recent(tp);
4030 /* We know that such packets are checksummed
4033 tcp_ack(sk, skb, 0);
4035 tcp_data_snd_check(sk, tp);
4037 } else { /* Header too small */
4038 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4043 int copied_early = 0;
4045 if (tp->copied_seq == tp->rcv_nxt &&
4046 len - tcp_header_len <= tp->ucopy.len) {
4047 #ifdef CONFIG_NET_DMA
4048 if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
4053 if (tp->ucopy.task == current && sock_owned_by_user(sk) && !copied_early) {
4054 __set_current_state(TASK_RUNNING);
4056 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
4060 /* Predicted packet is in window by definition.
4061 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4062 * Hence, check seq<=rcv_wup reduces to:
4064 if (tcp_header_len ==
4065 (sizeof(struct tcphdr) +
4066 TCPOLEN_TSTAMP_ALIGNED) &&
4067 tp->rcv_nxt == tp->rcv_wup)
4068 tcp_store_ts_recent(tp);
4070 tcp_rcv_rtt_measure_ts(sk, skb);
4072 __skb_pull(skb, tcp_header_len);
4073 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4074 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER);
4077 tcp_cleanup_rbuf(sk, skb->len);
4080 if (tcp_checksum_complete_user(sk, skb))
4083 /* Predicted packet is in window by definition.
4084 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4085 * Hence, check seq<=rcv_wup reduces to:
4087 if (tcp_header_len ==
4088 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
4089 tp->rcv_nxt == tp->rcv_wup)
4090 tcp_store_ts_recent(tp);
4092 tcp_rcv_rtt_measure_ts(sk, skb);
4094 if ((int)skb->truesize > sk->sk_forward_alloc)
4097 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS);
4099 /* Bulk data transfer: receiver */
4100 __skb_pull(skb,tcp_header_len);
4101 __skb_queue_tail(&sk->sk_receive_queue, skb);
4102 sk_stream_set_owner_r(skb, sk);
4103 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4106 tcp_event_data_recv(sk, tp, skb);
4108 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
4109 /* Well, only one small jumplet in fast path... */
4110 tcp_ack(sk, skb, FLAG_DATA);
4111 tcp_data_snd_check(sk, tp);
4112 if (!inet_csk_ack_scheduled(sk))
4116 __tcp_ack_snd_check(sk, 0);
4118 #ifdef CONFIG_NET_DMA
4120 __skb_queue_tail(&sk->sk_async_wait_queue, skb);
4126 sk->sk_data_ready(sk, 0);
4132 if (len < (th->doff<<2) || tcp_checksum_complete_user(sk, skb))
4136 * RFC1323: H1. Apply PAWS check first.
4138 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
4139 tcp_paws_discard(sk, skb)) {
4141 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
4142 tcp_send_dupack(sk, skb);
4145 /* Resets are accepted even if PAWS failed.
4147 ts_recent update must be made after we are sure
4148 that the packet is in window.
4153 * Standard slow path.
4156 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
4157 /* RFC793, page 37: "In all states except SYN-SENT, all reset
4158 * (RST) segments are validated by checking their SEQ-fields."
4159 * And page 69: "If an incoming segment is not acceptable,
4160 * an acknowledgment should be sent in reply (unless the RST bit
4161 * is set, if so drop the segment and return)".
4164 tcp_send_dupack(sk, skb);
4173 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
4175 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4176 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4177 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
4184 tcp_ack(sk, skb, FLAG_SLOWPATH);
4186 tcp_rcv_rtt_measure_ts(sk, skb);
4188 /* Process urgent data. */
4189 tcp_urg(sk, skb, th);
4191 /* step 7: process the segment text */
4192 tcp_data_queue(sk, skb);
4194 tcp_data_snd_check(sk, tp);
4195 tcp_ack_snd_check(sk);
4199 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4206 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
4207 struct tcphdr *th, unsigned len)
4209 struct tcp_sock *tp = tcp_sk(sk);
4210 struct inet_connection_sock *icsk = inet_csk(sk);
4211 int saved_clamp = tp->rx_opt.mss_clamp;
4213 tcp_parse_options(skb, &tp->rx_opt, 0);
4217 * "If the state is SYN-SENT then
4218 * first check the ACK bit
4219 * If the ACK bit is set
4220 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
4221 * a reset (unless the RST bit is set, if so drop
4222 * the segment and return)"
4224 * We do not send data with SYN, so that RFC-correct
4227 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
4228 goto reset_and_undo;
4230 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
4231 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
4233 NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED);
4234 goto reset_and_undo;
4237 /* Now ACK is acceptable.
4239 * "If the RST bit is set
4240 * If the ACK was acceptable then signal the user "error:
4241 * connection reset", drop the segment, enter CLOSED state,
4242 * delete TCB, and return."
4251 * "fifth, if neither of the SYN or RST bits is set then
4252 * drop the segment and return."
4258 goto discard_and_undo;
4261 * "If the SYN bit is on ...
4262 * are acceptable then ...
4263 * (our SYN has been ACKed), change the connection
4264 * state to ESTABLISHED..."
4267 TCP_ECN_rcv_synack(tp, th);
4269 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
4270 tcp_ack(sk, skb, FLAG_SLOWPATH);
4272 /* Ok.. it's good. Set up sequence numbers and
4273 * move to established.
4275 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
4276 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
4278 /* RFC1323: The window in SYN & SYN/ACK segments is
4281 tp->snd_wnd = ntohs(th->window);
4282 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(skb)->seq);
4284 if (!tp->rx_opt.wscale_ok) {
4285 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
4286 tp->window_clamp = min(tp->window_clamp, 65535U);
4289 if (tp->rx_opt.saw_tstamp) {
4290 tp->rx_opt.tstamp_ok = 1;
4291 tp->tcp_header_len =
4292 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
4293 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
4294 tcp_store_ts_recent(tp);
4296 tp->tcp_header_len = sizeof(struct tcphdr);
4299 if (tp->rx_opt.sack_ok && sysctl_tcp_fack)
4300 tp->rx_opt.sack_ok |= 2;
4303 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
4304 tcp_initialize_rcv_mss(sk);
4306 /* Remember, tcp_poll() does not lock socket!
4307 * Change state from SYN-SENT only after copied_seq
4308 * is initialized. */
4309 tp->copied_seq = tp->rcv_nxt;
4311 tcp_set_state(sk, TCP_ESTABLISHED);
4313 security_inet_conn_established(sk, skb);
4315 /* Make sure socket is routed, for correct metrics. */
4316 icsk->icsk_af_ops->rebuild_header(sk);
4318 tcp_init_metrics(sk);
4320 tcp_init_congestion_control(sk);
4322 /* Prevent spurious tcp_cwnd_restart() on first data
4325 tp->lsndtime = tcp_time_stamp;
4327 tcp_init_buffer_space(sk);
4329 if (sock_flag(sk, SOCK_KEEPOPEN))
4330 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
4332 if (!tp->rx_opt.snd_wscale)
4333 __tcp_fast_path_on(tp, tp->snd_wnd);
4337 if (!sock_flag(sk, SOCK_DEAD)) {
4338 sk->sk_state_change(sk);
4339 sk_wake_async(sk, 0, POLL_OUT);
4342 if (sk->sk_write_pending ||
4343 icsk->icsk_accept_queue.rskq_defer_accept ||
4344 icsk->icsk_ack.pingpong) {
4345 /* Save one ACK. Data will be ready after
4346 * several ticks, if write_pending is set.
4348 * It may be deleted, but with this feature tcpdumps
4349 * look so _wonderfully_ clever, that I was not able
4350 * to stand against the temptation 8) --ANK
4352 inet_csk_schedule_ack(sk);
4353 icsk->icsk_ack.lrcvtime = tcp_time_stamp;
4354 icsk->icsk_ack.ato = TCP_ATO_MIN;
4355 tcp_incr_quickack(sk);
4356 tcp_enter_quickack_mode(sk);
4357 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
4358 TCP_DELACK_MAX, TCP_RTO_MAX);
4369 /* No ACK in the segment */
4373 * "If the RST bit is set
4375 * Otherwise (no ACK) drop the segment and return."
4378 goto discard_and_undo;
4382 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && tcp_paws_check(&tp->rx_opt, 0))
4383 goto discard_and_undo;
4386 /* We see SYN without ACK. It is attempt of
4387 * simultaneous connect with crossed SYNs.
4388 * Particularly, it can be connect to self.
4390 tcp_set_state(sk, TCP_SYN_RECV);
4392 if (tp->rx_opt.saw_tstamp) {
4393 tp->rx_opt.tstamp_ok = 1;
4394 tcp_store_ts_recent(tp);
4395 tp->tcp_header_len =
4396 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
4398 tp->tcp_header_len = sizeof(struct tcphdr);
4401 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
4402 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
4404 /* RFC1323: The window in SYN & SYN/ACK segments is
4407 tp->snd_wnd = ntohs(th->window);
4408 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
4409 tp->max_window = tp->snd_wnd;
4411 TCP_ECN_rcv_syn(tp, th);
4414 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
4415 tcp_initialize_rcv_mss(sk);
4418 tcp_send_synack(sk);
4420 /* Note, we could accept data and URG from this segment.
4421 * There are no obstacles to make this.
4423 * However, if we ignore data in ACKless segments sometimes,
4424 * we have no reasons to accept it sometimes.
4425 * Also, seems the code doing it in step6 of tcp_rcv_state_process
4426 * is not flawless. So, discard packet for sanity.
4427 * Uncomment this return to process the data.
4434 /* "fifth, if neither of the SYN or RST bits is set then
4435 * drop the segment and return."
4439 tcp_clear_options(&tp->rx_opt);
4440 tp->rx_opt.mss_clamp = saved_clamp;
4444 tcp_clear_options(&tp->rx_opt);
4445 tp->rx_opt.mss_clamp = saved_clamp;
4451 * This function implements the receiving procedure of RFC 793 for
4452 * all states except ESTABLISHED and TIME_WAIT.
4453 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
4454 * address independent.
4457 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
4458 struct tcphdr *th, unsigned len)
4460 struct tcp_sock *tp = tcp_sk(sk);
4461 struct inet_connection_sock *icsk = inet_csk(sk);
4464 tp->rx_opt.saw_tstamp = 0;
4466 switch (sk->sk_state) {
4478 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
4481 /* Now we have several options: In theory there is
4482 * nothing else in the frame. KA9Q has an option to
4483 * send data with the syn, BSD accepts data with the
4484 * syn up to the [to be] advertised window and
4485 * Solaris 2.1 gives you a protocol error. For now
4486 * we just ignore it, that fits the spec precisely
4487 * and avoids incompatibilities. It would be nice in
4488 * future to drop through and process the data.
4490 * Now that TTCP is starting to be used we ought to
4492 * But, this leaves one open to an easy denial of
4493 * service attack, and SYN cookies can't defend
4494 * against this problem. So, we drop the data
4495 * in the interest of security over speed unless
4496 * it's still in use.
4504 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
4508 /* Do step6 onward by hand. */
4509 tcp_urg(sk, skb, th);
4511 tcp_data_snd_check(sk, tp);
4515 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
4516 tcp_paws_discard(sk, skb)) {
4518 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
4519 tcp_send_dupack(sk, skb);
4522 /* Reset is accepted even if it did not pass PAWS. */
4525 /* step 1: check sequence number */
4526 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
4528 tcp_send_dupack(sk, skb);
4532 /* step 2: check RST bit */
4538 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
4540 /* step 3: check security and precedence [ignored] */
4544 * Check for a SYN in window.
4546 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4547 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
4552 /* step 5: check the ACK field */
4554 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH);
4556 switch(sk->sk_state) {
4559 tp->copied_seq = tp->rcv_nxt;
4561 tcp_set_state(sk, TCP_ESTABLISHED);
4562 sk->sk_state_change(sk);
4564 /* Note, that this wakeup is only for marginal
4565 * crossed SYN case. Passively open sockets
4566 * are not waked up, because sk->sk_sleep ==
4567 * NULL and sk->sk_socket == NULL.
4569 if (sk->sk_socket) {
4570 sk_wake_async(sk,0,POLL_OUT);
4573 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
4574 tp->snd_wnd = ntohs(th->window) <<
4575 tp->rx_opt.snd_wscale;
4576 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq,
4577 TCP_SKB_CB(skb)->seq);
4579 /* tcp_ack considers this ACK as duplicate
4580 * and does not calculate rtt.
4581 * Fix it at least with timestamps.
4583 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
4585 tcp_ack_saw_tstamp(sk, 0);
4587 if (tp->rx_opt.tstamp_ok)
4588 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
4590 /* Make sure socket is routed, for
4593 icsk->icsk_af_ops->rebuild_header(sk);
4595 tcp_init_metrics(sk);
4597 tcp_init_congestion_control(sk);
4599 /* Prevent spurious tcp_cwnd_restart() on
4600 * first data packet.
4602 tp->lsndtime = tcp_time_stamp;
4605 tcp_initialize_rcv_mss(sk);
4606 tcp_init_buffer_space(sk);
4607 tcp_fast_path_on(tp);
4614 if (tp->snd_una == tp->write_seq) {
4615 tcp_set_state(sk, TCP_FIN_WAIT2);
4616 sk->sk_shutdown |= SEND_SHUTDOWN;
4617 dst_confirm(sk->sk_dst_cache);
4619 if (!sock_flag(sk, SOCK_DEAD))
4620 /* Wake up lingering close() */
4621 sk->sk_state_change(sk);
4625 if (tp->linger2 < 0 ||
4626 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4627 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
4629 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
4633 tmo = tcp_fin_time(sk);
4634 if (tmo > TCP_TIMEWAIT_LEN) {
4635 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
4636 } else if (th->fin || sock_owned_by_user(sk)) {
4637 /* Bad case. We could lose such FIN otherwise.
4638 * It is not a big problem, but it looks confusing
4639 * and not so rare event. We still can lose it now,
4640 * if it spins in bh_lock_sock(), but it is really
4643 inet_csk_reset_keepalive_timer(sk, tmo);
4645 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
4653 if (tp->snd_una == tp->write_seq) {
4654 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4660 if (tp->snd_una == tp->write_seq) {
4661 tcp_update_metrics(sk);
4670 /* step 6: check the URG bit */
4671 tcp_urg(sk, skb, th);
4673 /* step 7: process the segment text */
4674 switch (sk->sk_state) {
4675 case TCP_CLOSE_WAIT:
4678 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4682 /* RFC 793 says to queue data in these states,
4683 * RFC 1122 says we MUST send a reset.
4684 * BSD 4.4 also does reset.
4686 if (sk->sk_shutdown & RCV_SHUTDOWN) {
4687 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4688 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
4689 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
4695 case TCP_ESTABLISHED:
4696 tcp_data_queue(sk, skb);
4701 /* tcp_data could move socket to TIME-WAIT */
4702 if (sk->sk_state != TCP_CLOSE) {
4703 tcp_data_snd_check(sk, tp);
4704 tcp_ack_snd_check(sk);
4714 EXPORT_SYMBOL(sysctl_tcp_ecn);
4715 EXPORT_SYMBOL(sysctl_tcp_reordering);
4716 EXPORT_SYMBOL(tcp_parse_options);
4717 EXPORT_SYMBOL(tcp_rcv_established);
4718 EXPORT_SYMBOL(tcp_rcv_state_process);
4719 EXPORT_SYMBOL(tcp_initialize_rcv_mss);