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_frto_response __read_mostly;
90 int sysctl_tcp_nometrics_save __read_mostly;
92 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
93 int sysctl_tcp_abc __read_mostly;
95 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
96 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
97 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
98 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
99 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
100 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
101 #define FLAG_ECE 0x40 /* ECE in this ACK */
102 #define FLAG_DATA_LOST 0x80 /* SACK detected data lossage. */
103 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
104 #define FLAG_ONLY_ORIG_SACKED 0x200 /* SACKs only non-rexmit sent before RTO */
106 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
107 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
108 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
109 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
111 #define IsReno(tp) ((tp)->rx_opt.sack_ok == 0)
112 #define IsFack(tp) ((tp)->rx_opt.sack_ok & 2)
113 #define IsDSack(tp) ((tp)->rx_opt.sack_ok & 4)
115 #define IsSackFrto() (sysctl_tcp_frto == 0x2)
117 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
119 /* Adapt the MSS value used to make delayed ack decision to the
122 static void tcp_measure_rcv_mss(struct sock *sk,
123 const struct sk_buff *skb)
125 struct inet_connection_sock *icsk = inet_csk(sk);
126 const unsigned int lss = icsk->icsk_ack.last_seg_size;
129 icsk->icsk_ack.last_seg_size = 0;
131 /* skb->len may jitter because of SACKs, even if peer
132 * sends good full-sized frames.
134 len = skb_shinfo(skb)->gso_size ?: skb->len;
135 if (len >= icsk->icsk_ack.rcv_mss) {
136 icsk->icsk_ack.rcv_mss = len;
138 /* Otherwise, we make more careful check taking into account,
139 * that SACKs block is variable.
141 * "len" is invariant segment length, including TCP header.
143 len += skb->data - skb->h.raw;
144 if (len >= TCP_MIN_RCVMSS + sizeof(struct tcphdr) ||
145 /* If PSH is not set, packet should be
146 * full sized, provided peer TCP is not badly broken.
147 * This observation (if it is correct 8)) allows
148 * to handle super-low mtu links fairly.
150 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
151 !(tcp_flag_word(skb->h.th)&TCP_REMNANT))) {
152 /* Subtract also invariant (if peer is RFC compliant),
153 * tcp header plus fixed timestamp option length.
154 * Resulting "len" is MSS free of SACK jitter.
156 len -= tcp_sk(sk)->tcp_header_len;
157 icsk->icsk_ack.last_seg_size = len;
159 icsk->icsk_ack.rcv_mss = len;
163 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
164 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
165 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
169 static void tcp_incr_quickack(struct sock *sk)
171 struct inet_connection_sock *icsk = inet_csk(sk);
172 unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
176 if (quickacks > icsk->icsk_ack.quick)
177 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
180 void tcp_enter_quickack_mode(struct sock *sk)
182 struct inet_connection_sock *icsk = inet_csk(sk);
183 tcp_incr_quickack(sk);
184 icsk->icsk_ack.pingpong = 0;
185 icsk->icsk_ack.ato = TCP_ATO_MIN;
188 /* Send ACKs quickly, if "quick" count is not exhausted
189 * and the session is not interactive.
192 static inline int tcp_in_quickack_mode(const struct sock *sk)
194 const struct inet_connection_sock *icsk = inet_csk(sk);
195 return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
198 /* Buffer size and advertised window tuning.
200 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
203 static void tcp_fixup_sndbuf(struct sock *sk)
205 int sndmem = tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER + 16 +
206 sizeof(struct sk_buff);
208 if (sk->sk_sndbuf < 3 * sndmem)
209 sk->sk_sndbuf = min(3 * sndmem, sysctl_tcp_wmem[2]);
212 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
214 * All tcp_full_space() is split to two parts: "network" buffer, allocated
215 * forward and advertised in receiver window (tp->rcv_wnd) and
216 * "application buffer", required to isolate scheduling/application
217 * latencies from network.
218 * window_clamp is maximal advertised window. It can be less than
219 * tcp_full_space(), in this case tcp_full_space() - window_clamp
220 * is reserved for "application" buffer. The less window_clamp is
221 * the smoother our behaviour from viewpoint of network, but the lower
222 * throughput and the higher sensitivity of the connection to losses. 8)
224 * rcv_ssthresh is more strict window_clamp used at "slow start"
225 * phase to predict further behaviour of this connection.
226 * It is used for two goals:
227 * - to enforce header prediction at sender, even when application
228 * requires some significant "application buffer". It is check #1.
229 * - to prevent pruning of receive queue because of misprediction
230 * of receiver window. Check #2.
232 * The scheme does not work when sender sends good segments opening
233 * window and then starts to feed us spaghetti. But it should work
234 * in common situations. Otherwise, we have to rely on queue collapsing.
237 /* Slow part of check#2. */
238 static int __tcp_grow_window(const struct sock *sk, struct tcp_sock *tp,
239 const struct sk_buff *skb)
242 int truesize = tcp_win_from_space(skb->truesize)/2;
243 int window = tcp_win_from_space(sysctl_tcp_rmem[2])/2;
245 while (tp->rcv_ssthresh <= window) {
246 if (truesize <= skb->len)
247 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
255 static void tcp_grow_window(struct sock *sk, struct tcp_sock *tp,
259 if (tp->rcv_ssthresh < tp->window_clamp &&
260 (int)tp->rcv_ssthresh < tcp_space(sk) &&
261 !tcp_memory_pressure) {
264 /* Check #2. Increase window, if skb with such overhead
265 * will fit to rcvbuf in future.
267 if (tcp_win_from_space(skb->truesize) <= skb->len)
270 incr = __tcp_grow_window(sk, tp, skb);
273 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr, tp->window_clamp);
274 inet_csk(sk)->icsk_ack.quick |= 1;
279 /* 3. Tuning rcvbuf, when connection enters established state. */
281 static void tcp_fixup_rcvbuf(struct sock *sk)
283 struct tcp_sock *tp = tcp_sk(sk);
284 int rcvmem = tp->advmss + MAX_TCP_HEADER + 16 + sizeof(struct sk_buff);
286 /* Try to select rcvbuf so that 4 mss-sized segments
287 * will fit to window and corresponding skbs will fit to our rcvbuf.
288 * (was 3; 4 is minimum to allow fast retransmit to work.)
290 while (tcp_win_from_space(rcvmem) < tp->advmss)
292 if (sk->sk_rcvbuf < 4 * rcvmem)
293 sk->sk_rcvbuf = min(4 * rcvmem, sysctl_tcp_rmem[2]);
296 /* 4. Try to fixup all. It is made immediately after connection enters
299 static void tcp_init_buffer_space(struct sock *sk)
301 struct tcp_sock *tp = tcp_sk(sk);
304 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
305 tcp_fixup_rcvbuf(sk);
306 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
307 tcp_fixup_sndbuf(sk);
309 tp->rcvq_space.space = tp->rcv_wnd;
311 maxwin = tcp_full_space(sk);
313 if (tp->window_clamp >= maxwin) {
314 tp->window_clamp = maxwin;
316 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
317 tp->window_clamp = max(maxwin -
318 (maxwin >> sysctl_tcp_app_win),
322 /* Force reservation of one segment. */
323 if (sysctl_tcp_app_win &&
324 tp->window_clamp > 2 * tp->advmss &&
325 tp->window_clamp + tp->advmss > maxwin)
326 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
328 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
329 tp->snd_cwnd_stamp = tcp_time_stamp;
332 /* 5. Recalculate window clamp after socket hit its memory bounds. */
333 static void tcp_clamp_window(struct sock *sk, struct tcp_sock *tp)
335 struct inet_connection_sock *icsk = inet_csk(sk);
337 icsk->icsk_ack.quick = 0;
339 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
340 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
341 !tcp_memory_pressure &&
342 atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) {
343 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
346 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
347 tp->rcv_ssthresh = min(tp->window_clamp, 2U*tp->advmss);
351 /* Initialize RCV_MSS value.
352 * RCV_MSS is an our guess about MSS used by the peer.
353 * We haven't any direct information about the MSS.
354 * It's better to underestimate the RCV_MSS rather than overestimate.
355 * Overestimations make us ACKing less frequently than needed.
356 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
358 void tcp_initialize_rcv_mss(struct sock *sk)
360 struct tcp_sock *tp = tcp_sk(sk);
361 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
363 hint = min(hint, tp->rcv_wnd/2);
364 hint = min(hint, TCP_MIN_RCVMSS);
365 hint = max(hint, TCP_MIN_MSS);
367 inet_csk(sk)->icsk_ack.rcv_mss = hint;
370 /* Receiver "autotuning" code.
372 * The algorithm for RTT estimation w/o timestamps is based on
373 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
374 * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps>
376 * More detail on this code can be found at
377 * <http://www.psc.edu/~jheffner/senior_thesis.ps>,
378 * though this reference is out of date. A new paper
381 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
383 u32 new_sample = tp->rcv_rtt_est.rtt;
389 if (new_sample != 0) {
390 /* If we sample in larger samples in the non-timestamp
391 * case, we could grossly overestimate the RTT especially
392 * with chatty applications or bulk transfer apps which
393 * are stalled on filesystem I/O.
395 * Also, since we are only going for a minimum in the
396 * non-timestamp case, we do not smooth things out
397 * else with timestamps disabled convergence takes too
401 m -= (new_sample >> 3);
403 } else if (m < new_sample)
406 /* No previous measure. */
410 if (tp->rcv_rtt_est.rtt != new_sample)
411 tp->rcv_rtt_est.rtt = new_sample;
414 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
416 if (tp->rcv_rtt_est.time == 0)
418 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
420 tcp_rcv_rtt_update(tp,
421 jiffies - tp->rcv_rtt_est.time,
425 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
426 tp->rcv_rtt_est.time = tcp_time_stamp;
429 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, const struct sk_buff *skb)
431 struct tcp_sock *tp = tcp_sk(sk);
432 if (tp->rx_opt.rcv_tsecr &&
433 (TCP_SKB_CB(skb)->end_seq -
434 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
435 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
439 * This function should be called every time data is copied to user space.
440 * It calculates the appropriate TCP receive buffer space.
442 void tcp_rcv_space_adjust(struct sock *sk)
444 struct tcp_sock *tp = tcp_sk(sk);
448 if (tp->rcvq_space.time == 0)
451 time = tcp_time_stamp - tp->rcvq_space.time;
452 if (time < (tp->rcv_rtt_est.rtt >> 3) ||
453 tp->rcv_rtt_est.rtt == 0)
456 space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
458 space = max(tp->rcvq_space.space, space);
460 if (tp->rcvq_space.space != space) {
463 tp->rcvq_space.space = space;
465 if (sysctl_tcp_moderate_rcvbuf &&
466 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
467 int new_clamp = space;
469 /* Receive space grows, normalize in order to
470 * take into account packet headers and sk_buff
471 * structure overhead.
476 rcvmem = (tp->advmss + MAX_TCP_HEADER +
477 16 + sizeof(struct sk_buff));
478 while (tcp_win_from_space(rcvmem) < tp->advmss)
481 space = min(space, sysctl_tcp_rmem[2]);
482 if (space > sk->sk_rcvbuf) {
483 sk->sk_rcvbuf = space;
485 /* Make the window clamp follow along. */
486 tp->window_clamp = new_clamp;
492 tp->rcvq_space.seq = tp->copied_seq;
493 tp->rcvq_space.time = tcp_time_stamp;
496 /* There is something which you must keep in mind when you analyze the
497 * behavior of the tp->ato delayed ack timeout interval. When a
498 * connection starts up, we want to ack as quickly as possible. The
499 * problem is that "good" TCP's do slow start at the beginning of data
500 * transmission. The means that until we send the first few ACK's the
501 * sender will sit on his end and only queue most of his data, because
502 * he can only send snd_cwnd unacked packets at any given time. For
503 * each ACK we send, he increments snd_cwnd and transmits more of his
506 static void tcp_event_data_recv(struct sock *sk, struct tcp_sock *tp, struct sk_buff *skb)
508 struct inet_connection_sock *icsk = inet_csk(sk);
511 inet_csk_schedule_ack(sk);
513 tcp_measure_rcv_mss(sk, skb);
515 tcp_rcv_rtt_measure(tp);
517 now = tcp_time_stamp;
519 if (!icsk->icsk_ack.ato) {
520 /* The _first_ data packet received, initialize
521 * delayed ACK engine.
523 tcp_incr_quickack(sk);
524 icsk->icsk_ack.ato = TCP_ATO_MIN;
526 int m = now - icsk->icsk_ack.lrcvtime;
528 if (m <= TCP_ATO_MIN/2) {
529 /* The fastest case is the first. */
530 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
531 } else if (m < icsk->icsk_ack.ato) {
532 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
533 if (icsk->icsk_ack.ato > icsk->icsk_rto)
534 icsk->icsk_ack.ato = icsk->icsk_rto;
535 } else if (m > icsk->icsk_rto) {
536 /* Too long gap. Apparently sender failed to
537 * restart window, so that we send ACKs quickly.
539 tcp_incr_quickack(sk);
540 sk_stream_mem_reclaim(sk);
543 icsk->icsk_ack.lrcvtime = now;
545 TCP_ECN_check_ce(tp, skb);
548 tcp_grow_window(sk, tp, skb);
551 /* Called to compute a smoothed rtt estimate. The data fed to this
552 * routine either comes from timestamps, or from segments that were
553 * known _not_ to have been retransmitted [see Karn/Partridge
554 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
555 * piece by Van Jacobson.
556 * NOTE: the next three routines used to be one big routine.
557 * To save cycles in the RFC 1323 implementation it was better to break
558 * it up into three procedures. -- erics
560 static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
562 struct tcp_sock *tp = tcp_sk(sk);
563 long m = mrtt; /* RTT */
565 /* The following amusing code comes from Jacobson's
566 * article in SIGCOMM '88. Note that rtt and mdev
567 * are scaled versions of rtt and mean deviation.
568 * This is designed to be as fast as possible
569 * m stands for "measurement".
571 * On a 1990 paper the rto value is changed to:
572 * RTO = rtt + 4 * mdev
574 * Funny. This algorithm seems to be very broken.
575 * These formulae increase RTO, when it should be decreased, increase
576 * too slowly, when it should be increased quickly, decrease too quickly
577 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
578 * does not matter how to _calculate_ it. Seems, it was trap
579 * that VJ failed to avoid. 8)
584 m -= (tp->srtt >> 3); /* m is now error in rtt est */
585 tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */
587 m = -m; /* m is now abs(error) */
588 m -= (tp->mdev >> 2); /* similar update on mdev */
589 /* This is similar to one of Eifel findings.
590 * Eifel blocks mdev updates when rtt decreases.
591 * This solution is a bit different: we use finer gain
592 * for mdev in this case (alpha*beta).
593 * Like Eifel it also prevents growth of rto,
594 * but also it limits too fast rto decreases,
595 * happening in pure Eifel.
600 m -= (tp->mdev >> 2); /* similar update on mdev */
602 tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */
603 if (tp->mdev > tp->mdev_max) {
604 tp->mdev_max = tp->mdev;
605 if (tp->mdev_max > tp->rttvar)
606 tp->rttvar = tp->mdev_max;
608 if (after(tp->snd_una, tp->rtt_seq)) {
609 if (tp->mdev_max < tp->rttvar)
610 tp->rttvar -= (tp->rttvar-tp->mdev_max)>>2;
611 tp->rtt_seq = tp->snd_nxt;
612 tp->mdev_max = TCP_RTO_MIN;
615 /* no previous measure. */
616 tp->srtt = m<<3; /* take the measured time to be rtt */
617 tp->mdev = m<<1; /* make sure rto = 3*rtt */
618 tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN);
619 tp->rtt_seq = tp->snd_nxt;
623 /* Calculate rto without backoff. This is the second half of Van Jacobson's
624 * routine referred to above.
626 static inline void tcp_set_rto(struct sock *sk)
628 const struct tcp_sock *tp = tcp_sk(sk);
629 /* Old crap is replaced with new one. 8)
632 * 1. If rtt variance happened to be less 50msec, it is hallucination.
633 * It cannot be less due to utterly erratic ACK generation made
634 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
635 * to do with delayed acks, because at cwnd>2 true delack timeout
636 * is invisible. Actually, Linux-2.4 also generates erratic
637 * ACKs in some circumstances.
639 inet_csk(sk)->icsk_rto = (tp->srtt >> 3) + tp->rttvar;
641 /* 2. Fixups made earlier cannot be right.
642 * If we do not estimate RTO correctly without them,
643 * all the algo is pure shit and should be replaced
644 * with correct one. It is exactly, which we pretend to do.
648 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
649 * guarantees that rto is higher.
651 static inline void tcp_bound_rto(struct sock *sk)
653 if (inet_csk(sk)->icsk_rto > TCP_RTO_MAX)
654 inet_csk(sk)->icsk_rto = TCP_RTO_MAX;
657 /* Save metrics learned by this TCP session.
658 This function is called only, when TCP finishes successfully
659 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
661 void tcp_update_metrics(struct sock *sk)
663 struct tcp_sock *tp = tcp_sk(sk);
664 struct dst_entry *dst = __sk_dst_get(sk);
666 if (sysctl_tcp_nometrics_save)
671 if (dst && (dst->flags&DST_HOST)) {
672 const struct inet_connection_sock *icsk = inet_csk(sk);
675 if (icsk->icsk_backoff || !tp->srtt) {
676 /* This session failed to estimate rtt. Why?
677 * Probably, no packets returned in time.
680 if (!(dst_metric_locked(dst, RTAX_RTT)))
681 dst->metrics[RTAX_RTT-1] = 0;
685 m = dst_metric(dst, RTAX_RTT) - tp->srtt;
687 /* If newly calculated rtt larger than stored one,
688 * store new one. Otherwise, use EWMA. Remember,
689 * rtt overestimation is always better than underestimation.
691 if (!(dst_metric_locked(dst, RTAX_RTT))) {
693 dst->metrics[RTAX_RTT-1] = tp->srtt;
695 dst->metrics[RTAX_RTT-1] -= (m>>3);
698 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) {
702 /* Scale deviation to rttvar fixed point */
707 if (m >= dst_metric(dst, RTAX_RTTVAR))
708 dst->metrics[RTAX_RTTVAR-1] = m;
710 dst->metrics[RTAX_RTTVAR-1] -=
711 (dst->metrics[RTAX_RTTVAR-1] - m)>>2;
714 if (tp->snd_ssthresh >= 0xFFFF) {
715 /* Slow start still did not finish. */
716 if (dst_metric(dst, RTAX_SSTHRESH) &&
717 !dst_metric_locked(dst, RTAX_SSTHRESH) &&
718 (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH))
719 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1;
720 if (!dst_metric_locked(dst, RTAX_CWND) &&
721 tp->snd_cwnd > dst_metric(dst, RTAX_CWND))
722 dst->metrics[RTAX_CWND-1] = tp->snd_cwnd;
723 } else if (tp->snd_cwnd > tp->snd_ssthresh &&
724 icsk->icsk_ca_state == TCP_CA_Open) {
725 /* Cong. avoidance phase, cwnd is reliable. */
726 if (!dst_metric_locked(dst, RTAX_SSTHRESH))
727 dst->metrics[RTAX_SSTHRESH-1] =
728 max(tp->snd_cwnd >> 1, tp->snd_ssthresh);
729 if (!dst_metric_locked(dst, RTAX_CWND))
730 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_cwnd) >> 1;
732 /* Else slow start did not finish, cwnd is non-sense,
733 ssthresh may be also invalid.
735 if (!dst_metric_locked(dst, RTAX_CWND))
736 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_ssthresh) >> 1;
737 if (dst->metrics[RTAX_SSTHRESH-1] &&
738 !dst_metric_locked(dst, RTAX_SSTHRESH) &&
739 tp->snd_ssthresh > dst->metrics[RTAX_SSTHRESH-1])
740 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh;
743 if (!dst_metric_locked(dst, RTAX_REORDERING)) {
744 if (dst->metrics[RTAX_REORDERING-1] < tp->reordering &&
745 tp->reordering != sysctl_tcp_reordering)
746 dst->metrics[RTAX_REORDERING-1] = tp->reordering;
751 /* Numbers are taken from RFC2414. */
752 __u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst)
754 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
757 if (tp->mss_cache > 1460)
760 cwnd = (tp->mss_cache > 1095) ? 3 : 4;
762 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
765 /* Set slow start threshold and cwnd not falling to slow start */
766 void tcp_enter_cwr(struct sock *sk, const int set_ssthresh)
768 struct tcp_sock *tp = tcp_sk(sk);
769 const struct inet_connection_sock *icsk = inet_csk(sk);
771 tp->prior_ssthresh = 0;
773 if (icsk->icsk_ca_state < TCP_CA_CWR) {
776 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
777 tp->snd_cwnd = min(tp->snd_cwnd,
778 tcp_packets_in_flight(tp) + 1U);
779 tp->snd_cwnd_cnt = 0;
780 tp->high_seq = tp->snd_nxt;
781 tp->snd_cwnd_stamp = tcp_time_stamp;
782 TCP_ECN_queue_cwr(tp);
784 tcp_set_ca_state(sk, TCP_CA_CWR);
788 /* Initialize metrics on socket. */
790 static void tcp_init_metrics(struct sock *sk)
792 struct tcp_sock *tp = tcp_sk(sk);
793 struct dst_entry *dst = __sk_dst_get(sk);
800 if (dst_metric_locked(dst, RTAX_CWND))
801 tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND);
802 if (dst_metric(dst, RTAX_SSTHRESH)) {
803 tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH);
804 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
805 tp->snd_ssthresh = tp->snd_cwnd_clamp;
807 if (dst_metric(dst, RTAX_REORDERING) &&
808 tp->reordering != dst_metric(dst, RTAX_REORDERING)) {
809 tp->rx_opt.sack_ok &= ~2;
810 tp->reordering = dst_metric(dst, RTAX_REORDERING);
813 if (dst_metric(dst, RTAX_RTT) == 0)
816 if (!tp->srtt && dst_metric(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3))
819 /* Initial rtt is determined from SYN,SYN-ACK.
820 * The segment is small and rtt may appear much
821 * less than real one. Use per-dst memory
822 * to make it more realistic.
824 * A bit of theory. RTT is time passed after "normal" sized packet
825 * is sent until it is ACKed. In normal circumstances sending small
826 * packets force peer to delay ACKs and calculation is correct too.
827 * The algorithm is adaptive and, provided we follow specs, it
828 * NEVER underestimate RTT. BUT! If peer tries to make some clever
829 * tricks sort of "quick acks" for time long enough to decrease RTT
830 * to low value, and then abruptly stops to do it and starts to delay
831 * ACKs, wait for troubles.
833 if (dst_metric(dst, RTAX_RTT) > tp->srtt) {
834 tp->srtt = dst_metric(dst, RTAX_RTT);
835 tp->rtt_seq = tp->snd_nxt;
837 if (dst_metric(dst, RTAX_RTTVAR) > tp->mdev) {
838 tp->mdev = dst_metric(dst, RTAX_RTTVAR);
839 tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN);
843 if (inet_csk(sk)->icsk_rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp)
845 tp->snd_cwnd = tcp_init_cwnd(tp, dst);
846 tp->snd_cwnd_stamp = tcp_time_stamp;
850 /* Play conservative. If timestamps are not
851 * supported, TCP will fail to recalculate correct
852 * rtt, if initial rto is too small. FORGET ALL AND RESET!
854 if (!tp->rx_opt.saw_tstamp && tp->srtt) {
856 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT;
857 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
861 static void tcp_update_reordering(struct sock *sk, const int metric,
864 struct tcp_sock *tp = tcp_sk(sk);
865 if (metric > tp->reordering) {
866 tp->reordering = min(TCP_MAX_REORDERING, metric);
868 /* This exciting event is worth to be remembered. 8) */
870 NET_INC_STATS_BH(LINUX_MIB_TCPTSREORDER);
872 NET_INC_STATS_BH(LINUX_MIB_TCPRENOREORDER);
874 NET_INC_STATS_BH(LINUX_MIB_TCPFACKREORDER);
876 NET_INC_STATS_BH(LINUX_MIB_TCPSACKREORDER);
877 #if FASTRETRANS_DEBUG > 1
878 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n",
879 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
883 tp->undo_marker ? tp->undo_retrans : 0);
885 /* Disable FACK yet. */
886 tp->rx_opt.sack_ok &= ~2;
890 /* This procedure tags the retransmission queue when SACKs arrive.
892 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
893 * Packets in queue with these bits set are counted in variables
894 * sacked_out, retrans_out and lost_out, correspondingly.
896 * Valid combinations are:
897 * Tag InFlight Description
898 * 0 1 - orig segment is in flight.
899 * S 0 - nothing flies, orig reached receiver.
900 * L 0 - nothing flies, orig lost by net.
901 * R 2 - both orig and retransmit are in flight.
902 * L|R 1 - orig is lost, retransmit is in flight.
903 * S|R 1 - orig reached receiver, retrans is still in flight.
904 * (L|S|R is logically valid, it could occur when L|R is sacked,
905 * but it is equivalent to plain S and code short-curcuits it to S.
906 * L|S is logically invalid, it would mean -1 packet in flight 8))
908 * These 6 states form finite state machine, controlled by the following events:
909 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
910 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
911 * 3. Loss detection event of one of three flavors:
912 * A. Scoreboard estimator decided the packet is lost.
913 * A'. Reno "three dupacks" marks head of queue lost.
914 * A''. Its FACK modfication, head until snd.fack is lost.
915 * B. SACK arrives sacking data transmitted after never retransmitted
917 * C. SACK arrives sacking SND.NXT at the moment, when the
918 * segment was retransmitted.
919 * 4. D-SACK added new rule: D-SACK changes any tag to S.
921 * It is pleasant to note, that state diagram turns out to be commutative,
922 * so that we are allowed not to be bothered by order of our actions,
923 * when multiple events arrive simultaneously. (see the function below).
925 * Reordering detection.
926 * --------------------
927 * Reordering metric is maximal distance, which a packet can be displaced
928 * in packet stream. With SACKs we can estimate it:
930 * 1. SACK fills old hole and the corresponding segment was not
931 * ever retransmitted -> reordering. Alas, we cannot use it
932 * when segment was retransmitted.
933 * 2. The last flaw is solved with D-SACK. D-SACK arrives
934 * for retransmitted and already SACKed segment -> reordering..
935 * Both of these heuristics are not used in Loss state, when we cannot
936 * account for retransmits accurately.
939 tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, u32 prior_snd_una)
941 const struct inet_connection_sock *icsk = inet_csk(sk);
942 struct tcp_sock *tp = tcp_sk(sk);
943 unsigned char *ptr = ack_skb->h.raw + TCP_SKB_CB(ack_skb)->sacked;
944 struct tcp_sack_block_wire *sp = (struct tcp_sack_block_wire *)(ptr+2);
945 struct sk_buff *cached_skb;
946 int num_sacks = (ptr[1] - TCPOLEN_SACK_BASE)>>3;
947 int reord = tp->packets_out;
949 u32 lost_retrans = 0;
952 int cached_fack_count;
954 int first_sack_index;
958 prior_fackets = tp->fackets_out;
960 /* Check for D-SACK. */
961 if (before(ntohl(sp[0].start_seq), TCP_SKB_CB(ack_skb)->ack_seq)) {
963 tp->rx_opt.sack_ok |= 4;
964 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV);
965 } else if (num_sacks > 1 &&
966 !after(ntohl(sp[0].end_seq), ntohl(sp[1].end_seq)) &&
967 !before(ntohl(sp[0].start_seq), ntohl(sp[1].start_seq))) {
969 tp->rx_opt.sack_ok |= 4;
970 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV);
973 /* D-SACK for already forgotten data...
974 * Do dumb counting. */
976 !after(ntohl(sp[0].end_seq), prior_snd_una) &&
977 after(ntohl(sp[0].end_seq), tp->undo_marker))
980 /* Eliminate too old ACKs, but take into
981 * account more or less fresh ones, they can
982 * contain valid SACK info.
984 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
988 * if the only SACK change is the increase of the end_seq of
989 * the first block then only apply that SACK block
990 * and use retrans queue hinting otherwise slowpath */
992 for (i = 0; i < num_sacks; i++) {
993 __be32 start_seq = sp[i].start_seq;
994 __be32 end_seq = sp[i].end_seq;
997 if (tp->recv_sack_cache[i].start_seq != start_seq)
1000 if ((tp->recv_sack_cache[i].start_seq != start_seq) ||
1001 (tp->recv_sack_cache[i].end_seq != end_seq))
1004 tp->recv_sack_cache[i].start_seq = start_seq;
1005 tp->recv_sack_cache[i].end_seq = end_seq;
1007 /* Clear the rest of the cache sack blocks so they won't match mistakenly. */
1008 for (; i < ARRAY_SIZE(tp->recv_sack_cache); i++) {
1009 tp->recv_sack_cache[i].start_seq = 0;
1010 tp->recv_sack_cache[i].end_seq = 0;
1013 first_sack_index = 0;
1018 tp->fastpath_skb_hint = NULL;
1020 /* order SACK blocks to allow in order walk of the retrans queue */
1021 for (i = num_sacks-1; i > 0; i--) {
1022 for (j = 0; j < i; j++){
1023 if (after(ntohl(sp[j].start_seq),
1024 ntohl(sp[j+1].start_seq))){
1025 struct tcp_sack_block_wire tmp;
1031 /* Track where the first SACK block goes to */
1032 if (j == first_sack_index)
1033 first_sack_index = j+1;
1040 /* clear flag as used for different purpose in following code */
1043 /* Use SACK fastpath hint if valid */
1044 cached_skb = tp->fastpath_skb_hint;
1045 cached_fack_count = tp->fastpath_cnt_hint;
1047 cached_skb = tcp_write_queue_head(sk);
1048 cached_fack_count = 0;
1051 for (i=0; i<num_sacks; i++, sp++) {
1052 struct sk_buff *skb;
1053 __u32 start_seq = ntohl(sp->start_seq);
1054 __u32 end_seq = ntohl(sp->end_seq);
1058 fack_count = cached_fack_count;
1060 /* Event "B" in the comment above. */
1061 if (after(end_seq, tp->high_seq))
1062 flag |= FLAG_DATA_LOST;
1064 tcp_for_write_queue_from(skb, sk) {
1065 int in_sack, pcount;
1068 if (skb == tcp_send_head(sk))
1072 cached_fack_count = fack_count;
1073 if (i == first_sack_index) {
1074 tp->fastpath_skb_hint = skb;
1075 tp->fastpath_cnt_hint = fack_count;
1078 /* The retransmission queue is always in order, so
1079 * we can short-circuit the walk early.
1081 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1084 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1085 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1087 pcount = tcp_skb_pcount(skb);
1089 if (pcount > 1 && !in_sack &&
1090 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1091 unsigned int pkt_len;
1093 in_sack = !after(start_seq,
1094 TCP_SKB_CB(skb)->seq);
1097 pkt_len = (start_seq -
1098 TCP_SKB_CB(skb)->seq);
1100 pkt_len = (end_seq -
1101 TCP_SKB_CB(skb)->seq);
1102 if (tcp_fragment(sk, skb, pkt_len, skb_shinfo(skb)->gso_size))
1104 pcount = tcp_skb_pcount(skb);
1107 fack_count += pcount;
1109 sacked = TCP_SKB_CB(skb)->sacked;
1111 /* Account D-SACK for retransmitted packet. */
1112 if ((dup_sack && in_sack) &&
1113 (sacked & TCPCB_RETRANS) &&
1114 after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker))
1117 /* The frame is ACKed. */
1118 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) {
1119 if (sacked&TCPCB_RETRANS) {
1120 if ((dup_sack && in_sack) &&
1121 (sacked&TCPCB_SACKED_ACKED))
1122 reord = min(fack_count, reord);
1124 /* If it was in a hole, we detected reordering. */
1125 if (fack_count < prior_fackets &&
1126 !(sacked&TCPCB_SACKED_ACKED))
1127 reord = min(fack_count, reord);
1130 /* Nothing to do; acked frame is about to be dropped. */
1134 if ((sacked&TCPCB_SACKED_RETRANS) &&
1135 after(end_seq, TCP_SKB_CB(skb)->ack_seq) &&
1136 (!lost_retrans || after(end_seq, lost_retrans)))
1137 lost_retrans = end_seq;
1142 if (!(sacked&TCPCB_SACKED_ACKED)) {
1143 if (sacked & TCPCB_SACKED_RETRANS) {
1144 /* If the segment is not tagged as lost,
1145 * we do not clear RETRANS, believing
1146 * that retransmission is still in flight.
1148 if (sacked & TCPCB_LOST) {
1149 TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1150 tp->lost_out -= tcp_skb_pcount(skb);
1151 tp->retrans_out -= tcp_skb_pcount(skb);
1153 /* clear lost hint */
1154 tp->retransmit_skb_hint = NULL;
1157 /* New sack for not retransmitted frame,
1158 * which was in hole. It is reordering.
1160 if (!(sacked & TCPCB_RETRANS) &&
1161 fack_count < prior_fackets)
1162 reord = min(fack_count, reord);
1164 if (sacked & TCPCB_LOST) {
1165 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1166 tp->lost_out -= tcp_skb_pcount(skb);
1168 /* clear lost hint */
1169 tp->retransmit_skb_hint = NULL;
1171 /* SACK enhanced F-RTO detection.
1172 * Set flag if and only if non-rexmitted
1173 * segments below frto_highmark are
1174 * SACKed (RFC4138; Appendix B).
1175 * Clearing correct due to in-order walk
1177 if (after(end_seq, tp->frto_highmark)) {
1178 flag &= ~FLAG_ONLY_ORIG_SACKED;
1180 if (!(sacked & TCPCB_RETRANS))
1181 flag |= FLAG_ONLY_ORIG_SACKED;
1185 TCP_SKB_CB(skb)->sacked |= TCPCB_SACKED_ACKED;
1186 flag |= FLAG_DATA_SACKED;
1187 tp->sacked_out += tcp_skb_pcount(skb);
1189 if (fack_count > tp->fackets_out)
1190 tp->fackets_out = fack_count;
1192 if (dup_sack && (sacked&TCPCB_RETRANS))
1193 reord = min(fack_count, reord);
1196 /* D-SACK. We can detect redundant retransmission
1197 * in S|R and plain R frames and clear it.
1198 * undo_retrans is decreased above, L|R frames
1199 * are accounted above as well.
1202 (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS)) {
1203 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1204 tp->retrans_out -= tcp_skb_pcount(skb);
1205 tp->retransmit_skb_hint = NULL;
1210 /* Check for lost retransmit. This superb idea is
1211 * borrowed from "ratehalving". Event "C".
1212 * Later note: FACK people cheated me again 8),
1213 * we have to account for reordering! Ugly,
1216 if (lost_retrans && icsk->icsk_ca_state == TCP_CA_Recovery) {
1217 struct sk_buff *skb;
1219 tcp_for_write_queue(skb, sk) {
1220 if (skb == tcp_send_head(sk))
1222 if (after(TCP_SKB_CB(skb)->seq, lost_retrans))
1224 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1226 if ((TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) &&
1227 after(lost_retrans, TCP_SKB_CB(skb)->ack_seq) &&
1229 !before(lost_retrans,
1230 TCP_SKB_CB(skb)->ack_seq + tp->reordering *
1232 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1233 tp->retrans_out -= tcp_skb_pcount(skb);
1235 /* clear lost hint */
1236 tp->retransmit_skb_hint = NULL;
1238 if (!(TCP_SKB_CB(skb)->sacked&(TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1239 tp->lost_out += tcp_skb_pcount(skb);
1240 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1241 flag |= FLAG_DATA_SACKED;
1242 NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT);
1248 tp->left_out = tp->sacked_out + tp->lost_out;
1250 if ((reord < tp->fackets_out) && icsk->icsk_ca_state != TCP_CA_Loss &&
1251 (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark)))
1252 tcp_update_reordering(sk, ((tp->fackets_out + 1) - reord), 0);
1254 #if FASTRETRANS_DEBUG > 0
1255 BUG_TRAP((int)tp->sacked_out >= 0);
1256 BUG_TRAP((int)tp->lost_out >= 0);
1257 BUG_TRAP((int)tp->retrans_out >= 0);
1258 BUG_TRAP((int)tcp_packets_in_flight(tp) >= 0);
1263 /* F-RTO can only be used if these conditions are satisfied:
1264 * - there must be some unsent new data
1265 * - the advertised window should allow sending it
1266 * - TCP has never retransmitted anything other than head (SACK enhanced
1267 * variant from Appendix B of RFC4138 is more robust here)
1269 int tcp_use_frto(struct sock *sk)
1271 const struct tcp_sock *tp = tcp_sk(sk);
1272 struct sk_buff *skb;
1274 if (!sysctl_tcp_frto || !tcp_send_head(sk) ||
1275 after(TCP_SKB_CB(tcp_send_head(sk))->end_seq,
1276 tp->snd_una + tp->snd_wnd))
1282 /* Avoid expensive walking of rexmit queue if possible */
1283 if (tp->retrans_out > 1)
1286 skb = tcp_write_queue_head(sk);
1287 skb = tcp_write_queue_next(sk, skb); /* Skips head */
1288 tcp_for_write_queue_from(skb, sk) {
1289 if (skb == tcp_send_head(sk))
1291 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS)
1293 /* Short-circuit when first non-SACKed skb has been checked */
1294 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED))
1300 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
1301 * recovery a bit and use heuristics in tcp_process_frto() to detect if
1302 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
1303 * keep retrans_out counting accurate (with SACK F-RTO, other than head
1304 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
1305 * bits are handled if the Loss state is really to be entered (in
1306 * tcp_enter_frto_loss).
1308 * Do like tcp_enter_loss() would; when RTO expires the second time it
1310 * "Reduce ssthresh if it has not yet been made inside this window."
1312 void tcp_enter_frto(struct sock *sk)
1314 const struct inet_connection_sock *icsk = inet_csk(sk);
1315 struct tcp_sock *tp = tcp_sk(sk);
1316 struct sk_buff *skb;
1318 if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) ||
1319 tp->snd_una == tp->high_seq ||
1320 ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) &&
1321 !icsk->icsk_retransmits)) {
1322 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1323 /* Our state is too optimistic in ssthresh() call because cwnd
1324 * is not reduced until tcp_enter_frto_loss() when previous FRTO
1325 * recovery has not yet completed. Pattern would be this: RTO,
1326 * Cumulative ACK, RTO (2xRTO for the same segment does not end
1328 * RFC4138 should be more specific on what to do, even though
1329 * RTO is quite unlikely to occur after the first Cumulative ACK
1330 * due to back-off and complexity of triggering events ...
1332 if (tp->frto_counter) {
1334 stored_cwnd = tp->snd_cwnd;
1336 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1337 tp->snd_cwnd = stored_cwnd;
1339 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1341 /* ... in theory, cong.control module could do "any tricks" in
1342 * ssthresh(), which means that ca_state, lost bits and lost_out
1343 * counter would have to be faked before the call occurs. We
1344 * consider that too expensive, unlikely and hacky, so modules
1345 * using these in ssthresh() must deal these incompatibility
1346 * issues if they receives CA_EVENT_FRTO and frto_counter != 0
1348 tcp_ca_event(sk, CA_EVENT_FRTO);
1351 tp->undo_marker = tp->snd_una;
1352 tp->undo_retrans = 0;
1354 skb = tcp_write_queue_head(sk);
1355 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
1356 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1357 tp->retrans_out -= tcp_skb_pcount(skb);
1359 tcp_sync_left_out(tp);
1361 /* Earlier loss recovery underway (see RFC4138; Appendix B).
1362 * The last condition is necessary at least in tp->frto_counter case.
1364 if (IsSackFrto() && (tp->frto_counter ||
1365 ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) &&
1366 after(tp->high_seq, tp->snd_una)) {
1367 tp->frto_highmark = tp->high_seq;
1369 tp->frto_highmark = tp->snd_nxt;
1371 tcp_set_ca_state(sk, TCP_CA_Disorder);
1372 tp->high_seq = tp->snd_nxt;
1373 tp->frto_counter = 1;
1376 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1377 * which indicates that we should follow the traditional RTO recovery,
1378 * i.e. mark everything lost and do go-back-N retransmission.
1380 static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag)
1382 struct tcp_sock *tp = tcp_sk(sk);
1383 struct sk_buff *skb;
1388 tp->fackets_out = 0;
1389 tp->retrans_out = 0;
1391 tcp_for_write_queue(skb, sk) {
1392 if (skb == tcp_send_head(sk))
1394 cnt += tcp_skb_pcount(skb);
1396 * Count the retransmission made on RTO correctly (only when
1397 * waiting for the first ACK and did not get it)...
1399 if ((tp->frto_counter == 1) && !(flag&FLAG_DATA_ACKED)) {
1400 tp->retrans_out += tcp_skb_pcount(skb);
1401 /* ...enter this if branch just for the first segment */
1402 flag |= FLAG_DATA_ACKED;
1404 TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1406 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) {
1408 /* Do not mark those segments lost that were
1409 * forward transmitted after RTO
1411 if (!after(TCP_SKB_CB(skb)->end_seq,
1412 tp->frto_highmark)) {
1413 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1414 tp->lost_out += tcp_skb_pcount(skb);
1417 tp->sacked_out += tcp_skb_pcount(skb);
1418 tp->fackets_out = cnt;
1421 tcp_sync_left_out(tp);
1423 tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments;
1424 tp->snd_cwnd_cnt = 0;
1425 tp->snd_cwnd_stamp = tcp_time_stamp;
1426 tp->undo_marker = 0;
1427 tp->frto_counter = 0;
1429 tp->reordering = min_t(unsigned int, tp->reordering,
1430 sysctl_tcp_reordering);
1431 tcp_set_ca_state(sk, TCP_CA_Loss);
1432 tp->high_seq = tp->frto_highmark;
1433 TCP_ECN_queue_cwr(tp);
1435 clear_all_retrans_hints(tp);
1438 void tcp_clear_retrans(struct tcp_sock *tp)
1441 tp->retrans_out = 0;
1443 tp->fackets_out = 0;
1447 tp->undo_marker = 0;
1448 tp->undo_retrans = 0;
1451 /* Enter Loss state. If "how" is not zero, forget all SACK information
1452 * and reset tags completely, otherwise preserve SACKs. If receiver
1453 * dropped its ofo queue, we will know this due to reneging detection.
1455 void tcp_enter_loss(struct sock *sk, int how)
1457 const struct inet_connection_sock *icsk = inet_csk(sk);
1458 struct tcp_sock *tp = tcp_sk(sk);
1459 struct sk_buff *skb;
1462 /* Reduce ssthresh if it has not yet been made inside this window. */
1463 if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
1464 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1465 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1466 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1467 tcp_ca_event(sk, CA_EVENT_LOSS);
1470 tp->snd_cwnd_cnt = 0;
1471 tp->snd_cwnd_stamp = tcp_time_stamp;
1473 tp->bytes_acked = 0;
1474 tcp_clear_retrans(tp);
1476 /* Push undo marker, if it was plain RTO and nothing
1477 * was retransmitted. */
1479 tp->undo_marker = tp->snd_una;
1481 tcp_for_write_queue(skb, sk) {
1482 if (skb == tcp_send_head(sk))
1484 cnt += tcp_skb_pcount(skb);
1485 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS)
1486 tp->undo_marker = 0;
1487 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1488 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
1489 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1490 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1491 tp->lost_out += tcp_skb_pcount(skb);
1493 tp->sacked_out += tcp_skb_pcount(skb);
1494 tp->fackets_out = cnt;
1497 tcp_sync_left_out(tp);
1499 tp->reordering = min_t(unsigned int, tp->reordering,
1500 sysctl_tcp_reordering);
1501 tcp_set_ca_state(sk, TCP_CA_Loss);
1502 tp->high_seq = tp->snd_nxt;
1503 TCP_ECN_queue_cwr(tp);
1505 clear_all_retrans_hints(tp);
1508 static int tcp_check_sack_reneging(struct sock *sk)
1510 struct sk_buff *skb;
1512 /* If ACK arrived pointing to a remembered SACK,
1513 * it means that our remembered SACKs do not reflect
1514 * real state of receiver i.e.
1515 * receiver _host_ is heavily congested (or buggy).
1516 * Do processing similar to RTO timeout.
1518 if ((skb = tcp_write_queue_head(sk)) != NULL &&
1519 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
1520 struct inet_connection_sock *icsk = inet_csk(sk);
1521 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING);
1523 tcp_enter_loss(sk, 1);
1524 icsk->icsk_retransmits++;
1525 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
1526 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1527 icsk->icsk_rto, TCP_RTO_MAX);
1533 static inline int tcp_fackets_out(struct tcp_sock *tp)
1535 return IsReno(tp) ? tp->sacked_out+1 : tp->fackets_out;
1538 static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb)
1540 return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto);
1543 static inline int tcp_head_timedout(struct sock *sk, struct tcp_sock *tp)
1545 return tp->packets_out &&
1546 tcp_skb_timedout(sk, tcp_write_queue_head(sk));
1549 /* Linux NewReno/SACK/FACK/ECN state machine.
1550 * --------------------------------------
1552 * "Open" Normal state, no dubious events, fast path.
1553 * "Disorder" In all the respects it is "Open",
1554 * but requires a bit more attention. It is entered when
1555 * we see some SACKs or dupacks. It is split of "Open"
1556 * mainly to move some processing from fast path to slow one.
1557 * "CWR" CWND was reduced due to some Congestion Notification event.
1558 * It can be ECN, ICMP source quench, local device congestion.
1559 * "Recovery" CWND was reduced, we are fast-retransmitting.
1560 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
1562 * tcp_fastretrans_alert() is entered:
1563 * - each incoming ACK, if state is not "Open"
1564 * - when arrived ACK is unusual, namely:
1569 * Counting packets in flight is pretty simple.
1571 * in_flight = packets_out - left_out + retrans_out
1573 * packets_out is SND.NXT-SND.UNA counted in packets.
1575 * retrans_out is number of retransmitted segments.
1577 * left_out is number of segments left network, but not ACKed yet.
1579 * left_out = sacked_out + lost_out
1581 * sacked_out: Packets, which arrived to receiver out of order
1582 * and hence not ACKed. With SACKs this number is simply
1583 * amount of SACKed data. Even without SACKs
1584 * it is easy to give pretty reliable estimate of this number,
1585 * counting duplicate ACKs.
1587 * lost_out: Packets lost by network. TCP has no explicit
1588 * "loss notification" feedback from network (for now).
1589 * It means that this number can be only _guessed_.
1590 * Actually, it is the heuristics to predict lossage that
1591 * distinguishes different algorithms.
1593 * F.e. after RTO, when all the queue is considered as lost,
1594 * lost_out = packets_out and in_flight = retrans_out.
1596 * Essentially, we have now two algorithms counting
1599 * FACK: It is the simplest heuristics. As soon as we decided
1600 * that something is lost, we decide that _all_ not SACKed
1601 * packets until the most forward SACK are lost. I.e.
1602 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
1603 * It is absolutely correct estimate, if network does not reorder
1604 * packets. And it loses any connection to reality when reordering
1605 * takes place. We use FACK by default until reordering
1606 * is suspected on the path to this destination.
1608 * NewReno: when Recovery is entered, we assume that one segment
1609 * is lost (classic Reno). While we are in Recovery and
1610 * a partial ACK arrives, we assume that one more packet
1611 * is lost (NewReno). This heuristics are the same in NewReno
1614 * Imagine, that's all! Forget about all this shamanism about CWND inflation
1615 * deflation etc. CWND is real congestion window, never inflated, changes
1616 * only according to classic VJ rules.
1618 * Really tricky (and requiring careful tuning) part of algorithm
1619 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
1620 * The first determines the moment _when_ we should reduce CWND and,
1621 * hence, slow down forward transmission. In fact, it determines the moment
1622 * when we decide that hole is caused by loss, rather than by a reorder.
1624 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
1625 * holes, caused by lost packets.
1627 * And the most logically complicated part of algorithm is undo
1628 * heuristics. We detect false retransmits due to both too early
1629 * fast retransmit (reordering) and underestimated RTO, analyzing
1630 * timestamps and D-SACKs. When we detect that some segments were
1631 * retransmitted by mistake and CWND reduction was wrong, we undo
1632 * window reduction and abort recovery phase. This logic is hidden
1633 * inside several functions named tcp_try_undo_<something>.
1636 /* This function decides, when we should leave Disordered state
1637 * and enter Recovery phase, reducing congestion window.
1639 * Main question: may we further continue forward transmission
1640 * with the same cwnd?
1642 static int tcp_time_to_recover(struct sock *sk, struct tcp_sock *tp)
1646 /* Do not perform any recovery during FRTO algorithm */
1647 if (tp->frto_counter)
1650 /* Trick#1: The loss is proven. */
1654 /* Not-A-Trick#2 : Classic rule... */
1655 if (tcp_fackets_out(tp) > tp->reordering)
1658 /* Trick#3 : when we use RFC2988 timer restart, fast
1659 * retransmit can be triggered by timeout of queue head.
1661 if (tcp_head_timedout(sk, tp))
1664 /* Trick#4: It is still not OK... But will it be useful to delay
1667 packets_out = tp->packets_out;
1668 if (packets_out <= tp->reordering &&
1669 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
1670 !tcp_may_send_now(sk, tp)) {
1671 /* We have nothing to send. This connection is limited
1672 * either by receiver window or by application.
1680 /* If we receive more dupacks than we expected counting segments
1681 * in assumption of absent reordering, interpret this as reordering.
1682 * The only another reason could be bug in receiver TCP.
1684 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1686 struct tcp_sock *tp = tcp_sk(sk);
1689 holes = max(tp->lost_out, 1U);
1690 holes = min(holes, tp->packets_out);
1692 if ((tp->sacked_out + holes) > tp->packets_out) {
1693 tp->sacked_out = tp->packets_out - holes;
1694 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1698 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1700 static void tcp_add_reno_sack(struct sock *sk)
1702 struct tcp_sock *tp = tcp_sk(sk);
1704 tcp_check_reno_reordering(sk, 0);
1705 tcp_sync_left_out(tp);
1708 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1710 static void tcp_remove_reno_sacks(struct sock *sk, struct tcp_sock *tp, int acked)
1713 /* One ACK acked hole. The rest eat duplicate ACKs. */
1714 if (acked-1 >= tp->sacked_out)
1717 tp->sacked_out -= acked-1;
1719 tcp_check_reno_reordering(sk, acked);
1720 tcp_sync_left_out(tp);
1723 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1726 tp->left_out = tp->lost_out;
1729 /* Mark head of queue up as lost. */
1730 static void tcp_mark_head_lost(struct sock *sk, struct tcp_sock *tp,
1731 int packets, u32 high_seq)
1733 struct sk_buff *skb;
1736 BUG_TRAP(packets <= tp->packets_out);
1737 if (tp->lost_skb_hint) {
1738 skb = tp->lost_skb_hint;
1739 cnt = tp->lost_cnt_hint;
1741 skb = tcp_write_queue_head(sk);
1745 tcp_for_write_queue_from(skb, sk) {
1746 if (skb == tcp_send_head(sk))
1748 /* TODO: do this better */
1749 /* this is not the most efficient way to do this... */
1750 tp->lost_skb_hint = skb;
1751 tp->lost_cnt_hint = cnt;
1752 cnt += tcp_skb_pcount(skb);
1753 if (cnt > packets || after(TCP_SKB_CB(skb)->end_seq, high_seq))
1755 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
1756 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1757 tp->lost_out += tcp_skb_pcount(skb);
1759 /* clear xmit_retransmit_queue hints
1760 * if this is beyond hint */
1761 if(tp->retransmit_skb_hint != NULL &&
1762 before(TCP_SKB_CB(skb)->seq,
1763 TCP_SKB_CB(tp->retransmit_skb_hint)->seq)) {
1765 tp->retransmit_skb_hint = NULL;
1769 tcp_sync_left_out(tp);
1772 /* Account newly detected lost packet(s) */
1774 static void tcp_update_scoreboard(struct sock *sk, struct tcp_sock *tp)
1777 int lost = tp->fackets_out - tp->reordering;
1780 tcp_mark_head_lost(sk, tp, lost, tp->high_seq);
1782 tcp_mark_head_lost(sk, tp, 1, tp->high_seq);
1785 /* New heuristics: it is possible only after we switched
1786 * to restart timer each time when something is ACKed.
1787 * Hence, we can detect timed out packets during fast
1788 * retransmit without falling to slow start.
1790 if (!IsReno(tp) && tcp_head_timedout(sk, tp)) {
1791 struct sk_buff *skb;
1793 skb = tp->scoreboard_skb_hint ? tp->scoreboard_skb_hint
1794 : tcp_write_queue_head(sk);
1796 tcp_for_write_queue_from(skb, sk) {
1797 if (skb == tcp_send_head(sk))
1799 if (!tcp_skb_timedout(sk, skb))
1802 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
1803 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1804 tp->lost_out += tcp_skb_pcount(skb);
1806 /* clear xmit_retrans hint */
1807 if (tp->retransmit_skb_hint &&
1808 before(TCP_SKB_CB(skb)->seq,
1809 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
1811 tp->retransmit_skb_hint = NULL;
1815 tp->scoreboard_skb_hint = skb;
1817 tcp_sync_left_out(tp);
1821 /* CWND moderation, preventing bursts due to too big ACKs
1822 * in dubious situations.
1824 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
1826 tp->snd_cwnd = min(tp->snd_cwnd,
1827 tcp_packets_in_flight(tp)+tcp_max_burst(tp));
1828 tp->snd_cwnd_stamp = tcp_time_stamp;
1831 /* Lower bound on congestion window is slow start threshold
1832 * unless congestion avoidance choice decides to overide it.
1834 static inline u32 tcp_cwnd_min(const struct sock *sk)
1836 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1838 return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh;
1841 /* Decrease cwnd each second ack. */
1842 static void tcp_cwnd_down(struct sock *sk)
1844 struct tcp_sock *tp = tcp_sk(sk);
1845 int decr = tp->snd_cwnd_cnt + 1;
1847 tp->snd_cwnd_cnt = decr&1;
1850 if (decr && tp->snd_cwnd > tcp_cwnd_min(sk))
1851 tp->snd_cwnd -= decr;
1853 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp)+1);
1854 tp->snd_cwnd_stamp = tcp_time_stamp;
1857 /* Nothing was retransmitted or returned timestamp is less
1858 * than timestamp of the first retransmission.
1860 static inline int tcp_packet_delayed(struct tcp_sock *tp)
1862 return !tp->retrans_stamp ||
1863 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
1864 (__s32)(tp->rx_opt.rcv_tsecr - tp->retrans_stamp) < 0);
1867 /* Undo procedures. */
1869 #if FASTRETRANS_DEBUG > 1
1870 static void DBGUNDO(struct sock *sk, struct tcp_sock *tp, const char *msg)
1872 struct inet_sock *inet = inet_sk(sk);
1873 printk(KERN_DEBUG "Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n",
1875 NIPQUAD(inet->daddr), ntohs(inet->dport),
1876 tp->snd_cwnd, tp->left_out,
1877 tp->snd_ssthresh, tp->prior_ssthresh,
1881 #define DBGUNDO(x...) do { } while (0)
1884 static void tcp_undo_cwr(struct sock *sk, const int undo)
1886 struct tcp_sock *tp = tcp_sk(sk);
1888 if (tp->prior_ssthresh) {
1889 const struct inet_connection_sock *icsk = inet_csk(sk);
1891 if (icsk->icsk_ca_ops->undo_cwnd)
1892 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
1894 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh<<1);
1896 if (undo && tp->prior_ssthresh > tp->snd_ssthresh) {
1897 tp->snd_ssthresh = tp->prior_ssthresh;
1898 TCP_ECN_withdraw_cwr(tp);
1901 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
1903 tcp_moderate_cwnd(tp);
1904 tp->snd_cwnd_stamp = tcp_time_stamp;
1906 /* There is something screwy going on with the retrans hints after
1908 clear_all_retrans_hints(tp);
1911 static inline int tcp_may_undo(struct tcp_sock *tp)
1913 return tp->undo_marker &&
1914 (!tp->undo_retrans || tcp_packet_delayed(tp));
1917 /* People celebrate: "We love our President!" */
1918 static int tcp_try_undo_recovery(struct sock *sk, struct tcp_sock *tp)
1920 if (tcp_may_undo(tp)) {
1921 /* Happy end! We did not retransmit anything
1922 * or our original transmission succeeded.
1924 DBGUNDO(sk, tp, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
1925 tcp_undo_cwr(sk, 1);
1926 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
1927 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
1929 NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO);
1930 tp->undo_marker = 0;
1932 if (tp->snd_una == tp->high_seq && IsReno(tp)) {
1933 /* Hold old state until something *above* high_seq
1934 * is ACKed. For Reno it is MUST to prevent false
1935 * fast retransmits (RFC2582). SACK TCP is safe. */
1936 tcp_moderate_cwnd(tp);
1939 tcp_set_ca_state(sk, TCP_CA_Open);
1943 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
1944 static void tcp_try_undo_dsack(struct sock *sk, struct tcp_sock *tp)
1946 if (tp->undo_marker && !tp->undo_retrans) {
1947 DBGUNDO(sk, tp, "D-SACK");
1948 tcp_undo_cwr(sk, 1);
1949 tp->undo_marker = 0;
1950 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO);
1954 /* Undo during fast recovery after partial ACK. */
1956 static int tcp_try_undo_partial(struct sock *sk, struct tcp_sock *tp,
1959 /* Partial ACK arrived. Force Hoe's retransmit. */
1960 int failed = IsReno(tp) || tp->fackets_out>tp->reordering;
1962 if (tcp_may_undo(tp)) {
1963 /* Plain luck! Hole if filled with delayed
1964 * packet, rather than with a retransmit.
1966 if (tp->retrans_out == 0)
1967 tp->retrans_stamp = 0;
1969 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
1971 DBGUNDO(sk, tp, "Hoe");
1972 tcp_undo_cwr(sk, 0);
1973 NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO);
1975 /* So... Do not make Hoe's retransmit yet.
1976 * If the first packet was delayed, the rest
1977 * ones are most probably delayed as well.
1984 /* Undo during loss recovery after partial ACK. */
1985 static int tcp_try_undo_loss(struct sock *sk, struct tcp_sock *tp)
1987 if (tcp_may_undo(tp)) {
1988 struct sk_buff *skb;
1989 tcp_for_write_queue(skb, sk) {
1990 if (skb == tcp_send_head(sk))
1992 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1995 clear_all_retrans_hints(tp);
1997 DBGUNDO(sk, tp, "partial loss");
1999 tp->left_out = tp->sacked_out;
2000 tcp_undo_cwr(sk, 1);
2001 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
2002 inet_csk(sk)->icsk_retransmits = 0;
2003 tp->undo_marker = 0;
2005 tcp_set_ca_state(sk, TCP_CA_Open);
2011 static inline void tcp_complete_cwr(struct sock *sk)
2013 struct tcp_sock *tp = tcp_sk(sk);
2014 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2015 tp->snd_cwnd_stamp = tcp_time_stamp;
2016 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2019 static void tcp_try_to_open(struct sock *sk, struct tcp_sock *tp, int flag)
2021 tp->left_out = tp->sacked_out;
2023 if (tp->retrans_out == 0)
2024 tp->retrans_stamp = 0;
2027 tcp_enter_cwr(sk, 1);
2029 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2030 int state = TCP_CA_Open;
2032 if (tp->left_out || tp->retrans_out || tp->undo_marker)
2033 state = TCP_CA_Disorder;
2035 if (inet_csk(sk)->icsk_ca_state != state) {
2036 tcp_set_ca_state(sk, state);
2037 tp->high_seq = tp->snd_nxt;
2039 tcp_moderate_cwnd(tp);
2045 static void tcp_mtup_probe_failed(struct sock *sk)
2047 struct inet_connection_sock *icsk = inet_csk(sk);
2049 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2050 icsk->icsk_mtup.probe_size = 0;
2053 static void tcp_mtup_probe_success(struct sock *sk, struct sk_buff *skb)
2055 struct tcp_sock *tp = tcp_sk(sk);
2056 struct inet_connection_sock *icsk = inet_csk(sk);
2058 /* FIXME: breaks with very large cwnd */
2059 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2060 tp->snd_cwnd = tp->snd_cwnd *
2061 tcp_mss_to_mtu(sk, tp->mss_cache) /
2062 icsk->icsk_mtup.probe_size;
2063 tp->snd_cwnd_cnt = 0;
2064 tp->snd_cwnd_stamp = tcp_time_stamp;
2065 tp->rcv_ssthresh = tcp_current_ssthresh(sk);
2067 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2068 icsk->icsk_mtup.probe_size = 0;
2069 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2073 /* Process an event, which can update packets-in-flight not trivially.
2074 * Main goal of this function is to calculate new estimate for left_out,
2075 * taking into account both packets sitting in receiver's buffer and
2076 * packets lost by network.
2078 * Besides that it does CWND reduction, when packet loss is detected
2079 * and changes state of machine.
2081 * It does _not_ decide what to send, it is made in function
2082 * tcp_xmit_retransmit_queue().
2085 tcp_fastretrans_alert(struct sock *sk, u32 prior_snd_una,
2086 int prior_packets, int flag)
2088 struct inet_connection_sock *icsk = inet_csk(sk);
2089 struct tcp_sock *tp = tcp_sk(sk);
2090 int is_dupack = (tp->snd_una == prior_snd_una && !(flag&FLAG_NOT_DUP));
2092 /* Some technical things:
2093 * 1. Reno does not count dupacks (sacked_out) automatically. */
2094 if (!tp->packets_out)
2096 /* 2. SACK counts snd_fack in packets inaccurately. */
2097 if (tp->sacked_out == 0)
2098 tp->fackets_out = 0;
2100 /* Now state machine starts.
2101 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2103 tp->prior_ssthresh = 0;
2105 /* B. In all the states check for reneging SACKs. */
2106 if (tp->sacked_out && tcp_check_sack_reneging(sk))
2109 /* C. Process data loss notification, provided it is valid. */
2110 if ((flag&FLAG_DATA_LOST) &&
2111 before(tp->snd_una, tp->high_seq) &&
2112 icsk->icsk_ca_state != TCP_CA_Open &&
2113 tp->fackets_out > tp->reordering) {
2114 tcp_mark_head_lost(sk, tp, tp->fackets_out-tp->reordering, tp->high_seq);
2115 NET_INC_STATS_BH(LINUX_MIB_TCPLOSS);
2118 /* D. Synchronize left_out to current state. */
2119 tcp_sync_left_out(tp);
2121 /* E. Check state exit conditions. State can be terminated
2122 * when high_seq is ACKed. */
2123 if (icsk->icsk_ca_state == TCP_CA_Open) {
2124 BUG_TRAP(tp->retrans_out == 0);
2125 tp->retrans_stamp = 0;
2126 } else if (!before(tp->snd_una, tp->high_seq)) {
2127 switch (icsk->icsk_ca_state) {
2129 icsk->icsk_retransmits = 0;
2130 if (tcp_try_undo_recovery(sk, tp))
2135 /* CWR is to be held something *above* high_seq
2136 * is ACKed for CWR bit to reach receiver. */
2137 if (tp->snd_una != tp->high_seq) {
2138 tcp_complete_cwr(sk);
2139 tcp_set_ca_state(sk, TCP_CA_Open);
2143 case TCP_CA_Disorder:
2144 tcp_try_undo_dsack(sk, tp);
2145 if (!tp->undo_marker ||
2146 /* For SACK case do not Open to allow to undo
2147 * catching for all duplicate ACKs. */
2148 IsReno(tp) || tp->snd_una != tp->high_seq) {
2149 tp->undo_marker = 0;
2150 tcp_set_ca_state(sk, TCP_CA_Open);
2154 case TCP_CA_Recovery:
2156 tcp_reset_reno_sack(tp);
2157 if (tcp_try_undo_recovery(sk, tp))
2159 tcp_complete_cwr(sk);
2164 /* F. Process state. */
2165 switch (icsk->icsk_ca_state) {
2166 case TCP_CA_Recovery:
2167 if (prior_snd_una == tp->snd_una) {
2168 if (IsReno(tp) && is_dupack)
2169 tcp_add_reno_sack(sk);
2171 int acked = prior_packets - tp->packets_out;
2173 tcp_remove_reno_sacks(sk, tp, acked);
2174 is_dupack = tcp_try_undo_partial(sk, tp, acked);
2178 if (flag&FLAG_DATA_ACKED)
2179 icsk->icsk_retransmits = 0;
2180 if (!tcp_try_undo_loss(sk, tp)) {
2181 tcp_moderate_cwnd(tp);
2182 tcp_xmit_retransmit_queue(sk);
2185 if (icsk->icsk_ca_state != TCP_CA_Open)
2187 /* Loss is undone; fall through to processing in Open state. */
2190 if (tp->snd_una != prior_snd_una)
2191 tcp_reset_reno_sack(tp);
2193 tcp_add_reno_sack(sk);
2196 if (icsk->icsk_ca_state == TCP_CA_Disorder)
2197 tcp_try_undo_dsack(sk, tp);
2199 if (!tcp_time_to_recover(sk, tp)) {
2200 tcp_try_to_open(sk, tp, flag);
2204 /* MTU probe failure: don't reduce cwnd */
2205 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2206 icsk->icsk_mtup.probe_size &&
2207 tp->snd_una == tp->mtu_probe.probe_seq_start) {
2208 tcp_mtup_probe_failed(sk);
2209 /* Restores the reduction we did in tcp_mtup_probe() */
2211 tcp_simple_retransmit(sk);
2215 /* Otherwise enter Recovery state */
2218 NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY);
2220 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY);
2222 tp->high_seq = tp->snd_nxt;
2223 tp->prior_ssthresh = 0;
2224 tp->undo_marker = tp->snd_una;
2225 tp->undo_retrans = tp->retrans_out;
2227 if (icsk->icsk_ca_state < TCP_CA_CWR) {
2228 if (!(flag&FLAG_ECE))
2229 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2230 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2231 TCP_ECN_queue_cwr(tp);
2234 tp->bytes_acked = 0;
2235 tp->snd_cwnd_cnt = 0;
2236 tcp_set_ca_state(sk, TCP_CA_Recovery);
2239 if (is_dupack || tcp_head_timedout(sk, tp))
2240 tcp_update_scoreboard(sk, tp);
2242 tcp_xmit_retransmit_queue(sk);
2245 /* Read draft-ietf-tcplw-high-performance before mucking
2246 * with this code. (Supersedes RFC1323)
2248 static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
2250 /* RTTM Rule: A TSecr value received in a segment is used to
2251 * update the averaged RTT measurement only if the segment
2252 * acknowledges some new data, i.e., only if it advances the
2253 * left edge of the send window.
2255 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2256 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
2258 * Changed: reset backoff as soon as we see the first valid sample.
2259 * If we do not, we get strongly overestimated rto. With timestamps
2260 * samples are accepted even from very old segments: f.e., when rtt=1
2261 * increases to 8, we retransmit 5 times and after 8 seconds delayed
2262 * answer arrives rto becomes 120 seconds! If at least one of segments
2263 * in window is lost... Voila. --ANK (010210)
2265 struct tcp_sock *tp = tcp_sk(sk);
2266 const __u32 seq_rtt = tcp_time_stamp - tp->rx_opt.rcv_tsecr;
2267 tcp_rtt_estimator(sk, seq_rtt);
2269 inet_csk(sk)->icsk_backoff = 0;
2273 static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
2275 /* We don't have a timestamp. Can only use
2276 * packets that are not retransmitted to determine
2277 * rtt estimates. Also, we must not reset the
2278 * backoff for rto until we get a non-retransmitted
2279 * packet. This allows us to deal with a situation
2280 * where the network delay has increased suddenly.
2281 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
2284 if (flag & FLAG_RETRANS_DATA_ACKED)
2287 tcp_rtt_estimator(sk, seq_rtt);
2289 inet_csk(sk)->icsk_backoff = 0;
2293 static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
2296 const struct tcp_sock *tp = tcp_sk(sk);
2297 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
2298 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
2299 tcp_ack_saw_tstamp(sk, flag);
2300 else if (seq_rtt >= 0)
2301 tcp_ack_no_tstamp(sk, seq_rtt, flag);
2304 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 rtt,
2305 u32 in_flight, int good)
2307 const struct inet_connection_sock *icsk = inet_csk(sk);
2308 icsk->icsk_ca_ops->cong_avoid(sk, ack, rtt, in_flight, good);
2309 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
2312 /* Restart timer after forward progress on connection.
2313 * RFC2988 recommends to restart timer to now+rto.
2316 static void tcp_ack_packets_out(struct sock *sk, struct tcp_sock *tp)
2318 if (!tp->packets_out) {
2319 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
2321 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2325 static int tcp_tso_acked(struct sock *sk, struct sk_buff *skb,
2326 __u32 now, __s32 *seq_rtt)
2328 struct tcp_sock *tp = tcp_sk(sk);
2329 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
2330 __u32 seq = tp->snd_una;
2331 __u32 packets_acked;
2334 /* If we get here, the whole TSO packet has not been
2337 BUG_ON(!after(scb->end_seq, seq));
2339 packets_acked = tcp_skb_pcount(skb);
2340 if (tcp_trim_head(sk, skb, seq - scb->seq))
2342 packets_acked -= tcp_skb_pcount(skb);
2344 if (packets_acked) {
2345 __u8 sacked = scb->sacked;
2347 acked |= FLAG_DATA_ACKED;
2349 if (sacked & TCPCB_RETRANS) {
2350 if (sacked & TCPCB_SACKED_RETRANS)
2351 tp->retrans_out -= packets_acked;
2352 acked |= FLAG_RETRANS_DATA_ACKED;
2354 } else if (*seq_rtt < 0)
2355 *seq_rtt = now - scb->when;
2356 if (sacked & TCPCB_SACKED_ACKED)
2357 tp->sacked_out -= packets_acked;
2358 if (sacked & TCPCB_LOST)
2359 tp->lost_out -= packets_acked;
2360 if (sacked & TCPCB_URG) {
2362 !before(seq, tp->snd_up))
2365 } else if (*seq_rtt < 0)
2366 *seq_rtt = now - scb->when;
2368 if (tp->fackets_out) {
2369 __u32 dval = min(tp->fackets_out, packets_acked);
2370 tp->fackets_out -= dval;
2372 tp->packets_out -= packets_acked;
2374 BUG_ON(tcp_skb_pcount(skb) == 0);
2375 BUG_ON(!before(scb->seq, scb->end_seq));
2381 static u32 tcp_usrtt(struct timeval *tv)
2385 do_gettimeofday(&now);
2386 return (now.tv_sec - tv->tv_sec) * 1000000 + (now.tv_usec - tv->tv_usec);
2389 /* Remove acknowledged frames from the retransmission queue. */
2390 static int tcp_clean_rtx_queue(struct sock *sk, __s32 *seq_rtt_p)
2392 struct tcp_sock *tp = tcp_sk(sk);
2393 const struct inet_connection_sock *icsk = inet_csk(sk);
2394 struct sk_buff *skb;
2395 __u32 now = tcp_time_stamp;
2399 void (*rtt_sample)(struct sock *sk, u32 usrtt)
2400 = icsk->icsk_ca_ops->rtt_sample;
2401 struct timeval tv = { .tv_sec = 0, .tv_usec = 0 };
2403 while ((skb = tcp_write_queue_head(sk)) &&
2404 skb != tcp_send_head(sk)) {
2405 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
2406 __u8 sacked = scb->sacked;
2408 /* If our packet is before the ack sequence we can
2409 * discard it as it's confirmed to have arrived at
2412 if (after(scb->end_seq, tp->snd_una)) {
2413 if (tcp_skb_pcount(skb) > 1 &&
2414 after(tp->snd_una, scb->seq))
2415 acked |= tcp_tso_acked(sk, skb,
2420 /* Initial outgoing SYN's get put onto the write_queue
2421 * just like anything else we transmit. It is not
2422 * true data, and if we misinform our callers that
2423 * this ACK acks real data, we will erroneously exit
2424 * connection startup slow start one packet too
2425 * quickly. This is severely frowned upon behavior.
2427 if (!(scb->flags & TCPCB_FLAG_SYN)) {
2428 acked |= FLAG_DATA_ACKED;
2431 acked |= FLAG_SYN_ACKED;
2432 tp->retrans_stamp = 0;
2435 /* MTU probing checks */
2436 if (icsk->icsk_mtup.probe_size) {
2437 if (!after(tp->mtu_probe.probe_seq_end, TCP_SKB_CB(skb)->end_seq)) {
2438 tcp_mtup_probe_success(sk, skb);
2443 if (sacked & TCPCB_RETRANS) {
2444 if(sacked & TCPCB_SACKED_RETRANS)
2445 tp->retrans_out -= tcp_skb_pcount(skb);
2446 acked |= FLAG_RETRANS_DATA_ACKED;
2448 } else if (seq_rtt < 0) {
2449 seq_rtt = now - scb->when;
2450 skb_get_timestamp(skb, &tv);
2452 if (sacked & TCPCB_SACKED_ACKED)
2453 tp->sacked_out -= tcp_skb_pcount(skb);
2454 if (sacked & TCPCB_LOST)
2455 tp->lost_out -= tcp_skb_pcount(skb);
2456 if (sacked & TCPCB_URG) {
2458 !before(scb->end_seq, tp->snd_up))
2461 } else if (seq_rtt < 0) {
2462 seq_rtt = now - scb->when;
2463 skb_get_timestamp(skb, &tv);
2465 tcp_dec_pcount_approx(&tp->fackets_out, skb);
2466 tcp_packets_out_dec(tp, skb);
2467 tcp_unlink_write_queue(skb, sk);
2468 sk_stream_free_skb(sk, skb);
2469 clear_all_retrans_hints(tp);
2472 if (acked&FLAG_ACKED) {
2473 tcp_ack_update_rtt(sk, acked, seq_rtt);
2474 tcp_ack_packets_out(sk, tp);
2475 if (rtt_sample && !(acked & FLAG_RETRANS_DATA_ACKED))
2476 (*rtt_sample)(sk, tcp_usrtt(&tv));
2478 if (icsk->icsk_ca_ops->pkts_acked)
2479 icsk->icsk_ca_ops->pkts_acked(sk, pkts_acked);
2482 #if FASTRETRANS_DEBUG > 0
2483 BUG_TRAP((int)tp->sacked_out >= 0);
2484 BUG_TRAP((int)tp->lost_out >= 0);
2485 BUG_TRAP((int)tp->retrans_out >= 0);
2486 if (!tp->packets_out && tp->rx_opt.sack_ok) {
2487 const struct inet_connection_sock *icsk = inet_csk(sk);
2489 printk(KERN_DEBUG "Leak l=%u %d\n",
2490 tp->lost_out, icsk->icsk_ca_state);
2493 if (tp->sacked_out) {
2494 printk(KERN_DEBUG "Leak s=%u %d\n",
2495 tp->sacked_out, icsk->icsk_ca_state);
2498 if (tp->retrans_out) {
2499 printk(KERN_DEBUG "Leak r=%u %d\n",
2500 tp->retrans_out, icsk->icsk_ca_state);
2501 tp->retrans_out = 0;
2505 *seq_rtt_p = seq_rtt;
2509 static void tcp_ack_probe(struct sock *sk)
2511 const struct tcp_sock *tp = tcp_sk(sk);
2512 struct inet_connection_sock *icsk = inet_csk(sk);
2514 /* Was it a usable window open? */
2516 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq,
2517 tp->snd_una + tp->snd_wnd)) {
2518 icsk->icsk_backoff = 0;
2519 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
2520 /* Socket must be waked up by subsequent tcp_data_snd_check().
2521 * This function is not for random using!
2524 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2525 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2530 static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag)
2532 return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
2533 inet_csk(sk)->icsk_ca_state != TCP_CA_Open);
2536 static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag)
2538 const struct tcp_sock *tp = tcp_sk(sk);
2539 return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
2540 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
2543 /* Check that window update is acceptable.
2544 * The function assumes that snd_una<=ack<=snd_next.
2546 static inline int tcp_may_update_window(const struct tcp_sock *tp, const u32 ack,
2547 const u32 ack_seq, const u32 nwin)
2549 return (after(ack, tp->snd_una) ||
2550 after(ack_seq, tp->snd_wl1) ||
2551 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd));
2554 /* Update our send window.
2556 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
2557 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
2559 static int tcp_ack_update_window(struct sock *sk, struct tcp_sock *tp,
2560 struct sk_buff *skb, u32 ack, u32 ack_seq)
2563 u32 nwin = ntohs(skb->h.th->window);
2565 if (likely(!skb->h.th->syn))
2566 nwin <<= tp->rx_opt.snd_wscale;
2568 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
2569 flag |= FLAG_WIN_UPDATE;
2570 tcp_update_wl(tp, ack, ack_seq);
2572 if (tp->snd_wnd != nwin) {
2575 /* Note, it is the only place, where
2576 * fast path is recovered for sending TCP.
2579 tcp_fast_path_check(sk, tp);
2581 if (nwin > tp->max_window) {
2582 tp->max_window = nwin;
2583 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
2593 /* A very conservative spurious RTO response algorithm: reduce cwnd and
2594 * continue in congestion avoidance.
2596 static void tcp_conservative_spur_to_response(struct tcp_sock *tp)
2598 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2599 tp->snd_cwnd_cnt = 0;
2600 tcp_moderate_cwnd(tp);
2603 /* A conservative spurious RTO response algorithm: reduce cwnd using
2604 * rate halving and continue in congestion avoidance.
2606 static void tcp_ratehalving_spur_to_response(struct sock *sk)
2608 tcp_enter_cwr(sk, 0);
2611 static void tcp_undo_spur_to_response(struct sock *sk, int flag)
2614 tcp_ratehalving_spur_to_response(sk);
2616 tcp_undo_cwr(sk, 1);
2619 /* F-RTO spurious RTO detection algorithm (RFC4138)
2621 * F-RTO affects during two new ACKs following RTO (well, almost, see inline
2622 * comments). State (ACK number) is kept in frto_counter. When ACK advances
2623 * window (but not to or beyond highest sequence sent before RTO):
2624 * On First ACK, send two new segments out.
2625 * On Second ACK, RTO was likely spurious. Do spurious response (response
2626 * algorithm is not part of the F-RTO detection algorithm
2627 * given in RFC4138 but can be selected separately).
2628 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
2629 * and TCP falls back to conventional RTO recovery.
2631 * Rationale: if the RTO was spurious, new ACKs should arrive from the
2632 * original window even after we transmit two new data segments.
2635 * on first step, wait until first cumulative ACK arrives, then move to
2636 * the second step. In second step, the next ACK decides.
2638 * F-RTO is implemented (mainly) in four functions:
2639 * - tcp_use_frto() is used to determine if TCP is can use F-RTO
2640 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
2641 * called when tcp_use_frto() showed green light
2642 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
2643 * - tcp_enter_frto_loss() is called if there is not enough evidence
2644 * to prove that the RTO is indeed spurious. It transfers the control
2645 * from F-RTO to the conventional RTO recovery
2647 static int tcp_process_frto(struct sock *sk, u32 prior_snd_una, int flag)
2649 struct tcp_sock *tp = tcp_sk(sk);
2651 tcp_sync_left_out(tp);
2653 /* Duplicate the behavior from Loss state (fastretrans_alert) */
2654 if (flag&FLAG_DATA_ACKED)
2655 inet_csk(sk)->icsk_retransmits = 0;
2657 if (!before(tp->snd_una, tp->frto_highmark)) {
2658 tcp_enter_frto_loss(sk, tp->frto_counter + 1, flag);
2662 if (!IsSackFrto() || IsReno(tp)) {
2663 /* RFC4138 shortcoming in step 2; should also have case c):
2664 * ACK isn't duplicate nor advances window, e.g., opposite dir
2667 if ((tp->snd_una == prior_snd_una) && (flag&FLAG_NOT_DUP) &&
2668 !(flag&FLAG_FORWARD_PROGRESS))
2671 if (!(flag&FLAG_DATA_ACKED)) {
2672 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3),
2677 if (!(flag&FLAG_DATA_ACKED) && (tp->frto_counter == 1)) {
2678 /* Prevent sending of new data. */
2679 tp->snd_cwnd = min(tp->snd_cwnd,
2680 tcp_packets_in_flight(tp));
2684 if ((tp->frto_counter == 2) &&
2685 (!(flag&FLAG_FORWARD_PROGRESS) ||
2686 ((flag&FLAG_DATA_SACKED) && !(flag&FLAG_ONLY_ORIG_SACKED)))) {
2687 /* RFC4138 shortcoming (see comment above) */
2688 if (!(flag&FLAG_FORWARD_PROGRESS) && (flag&FLAG_NOT_DUP))
2691 tcp_enter_frto_loss(sk, 3, flag);
2696 if (tp->frto_counter == 1) {
2697 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
2698 tp->frto_counter = 2;
2700 } else /* frto_counter == 2 */ {
2701 switch (sysctl_tcp_frto_response) {
2703 tcp_undo_spur_to_response(sk, flag);
2706 tcp_conservative_spur_to_response(tp);
2709 tcp_ratehalving_spur_to_response(sk);
2712 tp->frto_counter = 0;
2717 /* This routine deals with incoming acks, but not outgoing ones. */
2718 static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag)
2720 struct inet_connection_sock *icsk = inet_csk(sk);
2721 struct tcp_sock *tp = tcp_sk(sk);
2722 u32 prior_snd_una = tp->snd_una;
2723 u32 ack_seq = TCP_SKB_CB(skb)->seq;
2724 u32 ack = TCP_SKB_CB(skb)->ack_seq;
2725 u32 prior_in_flight;
2730 /* If the ack is newer than sent or older than previous acks
2731 * then we can probably ignore it.
2733 if (after(ack, tp->snd_nxt))
2734 goto uninteresting_ack;
2736 if (before(ack, prior_snd_una))
2739 if (sysctl_tcp_abc) {
2740 if (icsk->icsk_ca_state < TCP_CA_CWR)
2741 tp->bytes_acked += ack - prior_snd_una;
2742 else if (icsk->icsk_ca_state == TCP_CA_Loss)
2743 /* we assume just one segment left network */
2744 tp->bytes_acked += min(ack - prior_snd_una, tp->mss_cache);
2747 if (!(flag&FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
2748 /* Window is constant, pure forward advance.
2749 * No more checks are required.
2750 * Note, we use the fact that SND.UNA>=SND.WL2.
2752 tcp_update_wl(tp, ack, ack_seq);
2754 flag |= FLAG_WIN_UPDATE;
2756 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
2758 NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS);
2760 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
2763 NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS);
2765 flag |= tcp_ack_update_window(sk, tp, skb, ack, ack_seq);
2767 if (TCP_SKB_CB(skb)->sacked)
2768 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
2770 if (TCP_ECN_rcv_ecn_echo(tp, skb->h.th))
2773 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
2776 /* We passed data and got it acked, remove any soft error
2777 * log. Something worked...
2779 sk->sk_err_soft = 0;
2780 tp->rcv_tstamp = tcp_time_stamp;
2781 prior_packets = tp->packets_out;
2785 prior_in_flight = tcp_packets_in_flight(tp);
2787 /* See if we can take anything off of the retransmit queue. */
2788 flag |= tcp_clean_rtx_queue(sk, &seq_rtt);
2790 if (tp->frto_counter)
2791 frto_cwnd = tcp_process_frto(sk, prior_snd_una, flag);
2793 if (tcp_ack_is_dubious(sk, flag)) {
2794 /* Advance CWND, if state allows this. */
2795 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd &&
2796 tcp_may_raise_cwnd(sk, flag))
2797 tcp_cong_avoid(sk, ack, seq_rtt, prior_in_flight, 0);
2798 tcp_fastretrans_alert(sk, prior_snd_una, prior_packets, flag);
2800 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd)
2801 tcp_cong_avoid(sk, ack, seq_rtt, prior_in_flight, 1);
2804 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag&FLAG_NOT_DUP))
2805 dst_confirm(sk->sk_dst_cache);
2810 icsk->icsk_probes_out = 0;
2812 /* If this ack opens up a zero window, clear backoff. It was
2813 * being used to time the probes, and is probably far higher than
2814 * it needs to be for normal retransmission.
2816 if (tcp_send_head(sk))
2821 if (TCP_SKB_CB(skb)->sacked)
2822 tcp_sacktag_write_queue(sk, skb, prior_snd_una);
2825 SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
2830 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
2831 * But, this can also be called on packets in the established flow when
2832 * the fast version below fails.
2834 void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx, int estab)
2837 struct tcphdr *th = skb->h.th;
2838 int length=(th->doff*4)-sizeof(struct tcphdr);
2840 ptr = (unsigned char *)(th + 1);
2841 opt_rx->saw_tstamp = 0;
2850 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
2855 if (opsize < 2) /* "silly options" */
2857 if (opsize > length)
2858 return; /* don't parse partial options */
2861 if(opsize==TCPOLEN_MSS && th->syn && !estab) {
2862 u16 in_mss = ntohs(get_unaligned((__be16 *)ptr));
2864 if (opt_rx->user_mss && opt_rx->user_mss < in_mss)
2865 in_mss = opt_rx->user_mss;
2866 opt_rx->mss_clamp = in_mss;
2871 if(opsize==TCPOLEN_WINDOW && th->syn && !estab)
2872 if (sysctl_tcp_window_scaling) {
2873 __u8 snd_wscale = *(__u8 *) ptr;
2874 opt_rx->wscale_ok = 1;
2875 if (snd_wscale > 14) {
2877 printk(KERN_INFO "tcp_parse_options: Illegal window "
2878 "scaling value %d >14 received.\n",
2882 opt_rx->snd_wscale = snd_wscale;
2885 case TCPOPT_TIMESTAMP:
2886 if(opsize==TCPOLEN_TIMESTAMP) {
2887 if ((estab && opt_rx->tstamp_ok) ||
2888 (!estab && sysctl_tcp_timestamps)) {
2889 opt_rx->saw_tstamp = 1;
2890 opt_rx->rcv_tsval = ntohl(get_unaligned((__be32 *)ptr));
2891 opt_rx->rcv_tsecr = ntohl(get_unaligned((__be32 *)(ptr+4)));
2895 case TCPOPT_SACK_PERM:
2896 if(opsize==TCPOLEN_SACK_PERM && th->syn && !estab) {
2897 if (sysctl_tcp_sack) {
2898 opt_rx->sack_ok = 1;
2899 tcp_sack_reset(opt_rx);
2905 if((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
2906 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
2908 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
2910 #ifdef CONFIG_TCP_MD5SIG
2913 * The MD5 Hash has already been
2914 * checked (see tcp_v{4,6}_do_rcv()).
2925 /* Fast parse options. This hopes to only see timestamps.
2926 * If it is wrong it falls back on tcp_parse_options().
2928 static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th,
2929 struct tcp_sock *tp)
2931 if (th->doff == sizeof(struct tcphdr)>>2) {
2932 tp->rx_opt.saw_tstamp = 0;
2934 } else if (tp->rx_opt.tstamp_ok &&
2935 th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) {
2936 __be32 *ptr = (__be32 *)(th + 1);
2937 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
2938 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
2939 tp->rx_opt.saw_tstamp = 1;
2941 tp->rx_opt.rcv_tsval = ntohl(*ptr);
2943 tp->rx_opt.rcv_tsecr = ntohl(*ptr);
2947 tcp_parse_options(skb, &tp->rx_opt, 1);
2951 static inline void tcp_store_ts_recent(struct tcp_sock *tp)
2953 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
2954 tp->rx_opt.ts_recent_stamp = get_seconds();
2957 static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
2959 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
2960 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
2961 * extra check below makes sure this can only happen
2962 * for pure ACK frames. -DaveM
2964 * Not only, also it occurs for expired timestamps.
2967 if((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) >= 0 ||
2968 get_seconds() >= tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS)
2969 tcp_store_ts_recent(tp);
2973 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
2975 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
2976 * it can pass through stack. So, the following predicate verifies that
2977 * this segment is not used for anything but congestion avoidance or
2978 * fast retransmit. Moreover, we even are able to eliminate most of such
2979 * second order effects, if we apply some small "replay" window (~RTO)
2980 * to timestamp space.
2982 * All these measures still do not guarantee that we reject wrapped ACKs
2983 * on networks with high bandwidth, when sequence space is recycled fastly,
2984 * but it guarantees that such events will be very rare and do not affect
2985 * connection seriously. This doesn't look nice, but alas, PAWS is really
2988 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
2989 * states that events when retransmit arrives after original data are rare.
2990 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
2991 * the biggest problem on large power networks even with minor reordering.
2992 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
2993 * up to bandwidth of 18Gigabit/sec. 8) ]
2996 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
2998 struct tcp_sock *tp = tcp_sk(sk);
2999 struct tcphdr *th = skb->h.th;
3000 u32 seq = TCP_SKB_CB(skb)->seq;
3001 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3003 return (/* 1. Pure ACK with correct sequence number. */
3004 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3006 /* 2. ... and duplicate ACK. */
3007 ack == tp->snd_una &&
3009 /* 3. ... and does not update window. */
3010 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3012 /* 4. ... and sits in replay window. */
3013 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3016 static inline int tcp_paws_discard(const struct sock *sk, const struct sk_buff *skb)
3018 const struct tcp_sock *tp = tcp_sk(sk);
3019 return ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > TCP_PAWS_WINDOW &&
3020 get_seconds() < tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS &&
3021 !tcp_disordered_ack(sk, skb));
3024 /* Check segment sequence number for validity.
3026 * Segment controls are considered valid, if the segment
3027 * fits to the window after truncation to the window. Acceptability
3028 * of data (and SYN, FIN, of course) is checked separately.
3029 * See tcp_data_queue(), for example.
3031 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3032 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3033 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3034 * (borrowed from freebsd)
3037 static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq)
3039 return !before(end_seq, tp->rcv_wup) &&
3040 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3043 /* When we get a reset we do this. */
3044 static void tcp_reset(struct sock *sk)
3046 /* We want the right error as BSD sees it (and indeed as we do). */
3047 switch (sk->sk_state) {
3049 sk->sk_err = ECONNREFUSED;
3051 case TCP_CLOSE_WAIT:
3057 sk->sk_err = ECONNRESET;
3060 if (!sock_flag(sk, SOCK_DEAD))
3061 sk->sk_error_report(sk);
3067 * Process the FIN bit. This now behaves as it is supposed to work
3068 * and the FIN takes effect when it is validly part of sequence
3069 * space. Not before when we get holes.
3071 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3072 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3075 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3076 * close and we go into CLOSING (and later onto TIME-WAIT)
3078 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3080 static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th)
3082 struct tcp_sock *tp = tcp_sk(sk);
3084 inet_csk_schedule_ack(sk);
3086 sk->sk_shutdown |= RCV_SHUTDOWN;
3087 sock_set_flag(sk, SOCK_DONE);
3089 switch (sk->sk_state) {
3091 case TCP_ESTABLISHED:
3092 /* Move to CLOSE_WAIT */
3093 tcp_set_state(sk, TCP_CLOSE_WAIT);
3094 inet_csk(sk)->icsk_ack.pingpong = 1;
3097 case TCP_CLOSE_WAIT:
3099 /* Received a retransmission of the FIN, do
3104 /* RFC793: Remain in the LAST-ACK state. */
3108 /* This case occurs when a simultaneous close
3109 * happens, we must ack the received FIN and
3110 * enter the CLOSING state.
3113 tcp_set_state(sk, TCP_CLOSING);
3116 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3118 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
3121 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3122 * cases we should never reach this piece of code.
3124 printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n",
3125 __FUNCTION__, sk->sk_state);
3129 /* It _is_ possible, that we have something out-of-order _after_ FIN.
3130 * Probably, we should reset in this case. For now drop them.
3132 __skb_queue_purge(&tp->out_of_order_queue);
3133 if (tp->rx_opt.sack_ok)
3134 tcp_sack_reset(&tp->rx_opt);
3135 sk_stream_mem_reclaim(sk);
3137 if (!sock_flag(sk, SOCK_DEAD)) {
3138 sk->sk_state_change(sk);
3140 /* Do not send POLL_HUP for half duplex close. */
3141 if (sk->sk_shutdown == SHUTDOWN_MASK ||
3142 sk->sk_state == TCP_CLOSE)
3143 sk_wake_async(sk, 1, POLL_HUP);
3145 sk_wake_async(sk, 1, POLL_IN);
3149 static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, u32 end_seq)
3151 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
3152 if (before(seq, sp->start_seq))
3153 sp->start_seq = seq;
3154 if (after(end_seq, sp->end_seq))
3155 sp->end_seq = end_seq;
3161 static void tcp_dsack_set(struct tcp_sock *tp, u32 seq, u32 end_seq)
3163 if (tp->rx_opt.sack_ok && sysctl_tcp_dsack) {
3164 if (before(seq, tp->rcv_nxt))
3165 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT);
3167 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT);
3169 tp->rx_opt.dsack = 1;
3170 tp->duplicate_sack[0].start_seq = seq;
3171 tp->duplicate_sack[0].end_seq = end_seq;
3172 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + 1, 4 - tp->rx_opt.tstamp_ok);
3176 static void tcp_dsack_extend(struct tcp_sock *tp, u32 seq, u32 end_seq)
3178 if (!tp->rx_opt.dsack)
3179 tcp_dsack_set(tp, seq, end_seq);
3181 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
3184 static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb)
3186 struct tcp_sock *tp = tcp_sk(sk);
3188 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
3189 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3190 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
3191 tcp_enter_quickack_mode(sk);
3193 if (tp->rx_opt.sack_ok && sysctl_tcp_dsack) {
3194 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3196 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
3197 end_seq = tp->rcv_nxt;
3198 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, end_seq);
3205 /* These routines update the SACK block as out-of-order packets arrive or
3206 * in-order packets close up the sequence space.
3208 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
3211 struct tcp_sack_block *sp = &tp->selective_acks[0];
3212 struct tcp_sack_block *swalk = sp+1;
3214 /* See if the recent change to the first SACK eats into
3215 * or hits the sequence space of other SACK blocks, if so coalesce.
3217 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks; ) {
3218 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
3221 /* Zap SWALK, by moving every further SACK up by one slot.
3222 * Decrease num_sacks.
3224 tp->rx_opt.num_sacks--;
3225 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
3226 for(i=this_sack; i < tp->rx_opt.num_sacks; i++)
3230 this_sack++, swalk++;
3234 static inline void tcp_sack_swap(struct tcp_sack_block *sack1, struct tcp_sack_block *sack2)
3238 tmp = sack1->start_seq;
3239 sack1->start_seq = sack2->start_seq;
3240 sack2->start_seq = tmp;
3242 tmp = sack1->end_seq;
3243 sack1->end_seq = sack2->end_seq;
3244 sack2->end_seq = tmp;
3247 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
3249 struct tcp_sock *tp = tcp_sk(sk);
3250 struct tcp_sack_block *sp = &tp->selective_acks[0];
3251 int cur_sacks = tp->rx_opt.num_sacks;
3257 for (this_sack=0; this_sack<cur_sacks; this_sack++, sp++) {
3258 if (tcp_sack_extend(sp, seq, end_seq)) {
3259 /* Rotate this_sack to the first one. */
3260 for (; this_sack>0; this_sack--, sp--)
3261 tcp_sack_swap(sp, sp-1);
3263 tcp_sack_maybe_coalesce(tp);
3268 /* Could not find an adjacent existing SACK, build a new one,
3269 * put it at the front, and shift everyone else down. We
3270 * always know there is at least one SACK present already here.
3272 * If the sack array is full, forget about the last one.
3274 if (this_sack >= 4) {
3276 tp->rx_opt.num_sacks--;
3279 for(; this_sack > 0; this_sack--, sp--)
3283 /* Build the new head SACK, and we're done. */
3284 sp->start_seq = seq;
3285 sp->end_seq = end_seq;
3286 tp->rx_opt.num_sacks++;
3287 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
3290 /* RCV.NXT advances, some SACKs should be eaten. */
3292 static void tcp_sack_remove(struct tcp_sock *tp)
3294 struct tcp_sack_block *sp = &tp->selective_acks[0];
3295 int num_sacks = tp->rx_opt.num_sacks;
3298 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
3299 if (skb_queue_empty(&tp->out_of_order_queue)) {
3300 tp->rx_opt.num_sacks = 0;
3301 tp->rx_opt.eff_sacks = tp->rx_opt.dsack;
3305 for(this_sack = 0; this_sack < num_sacks; ) {
3306 /* Check if the start of the sack is covered by RCV.NXT. */
3307 if (!before(tp->rcv_nxt, sp->start_seq)) {
3310 /* RCV.NXT must cover all the block! */
3311 BUG_TRAP(!before(tp->rcv_nxt, sp->end_seq));
3313 /* Zap this SACK, by moving forward any other SACKS. */
3314 for (i=this_sack+1; i < num_sacks; i++)
3315 tp->selective_acks[i-1] = tp->selective_acks[i];
3322 if (num_sacks != tp->rx_opt.num_sacks) {
3323 tp->rx_opt.num_sacks = num_sacks;
3324 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
3328 /* This one checks to see if we can put data from the
3329 * out_of_order queue into the receive_queue.
3331 static void tcp_ofo_queue(struct sock *sk)
3333 struct tcp_sock *tp = tcp_sk(sk);
3334 __u32 dsack_high = tp->rcv_nxt;
3335 struct sk_buff *skb;
3337 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
3338 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
3341 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
3342 __u32 dsack = dsack_high;
3343 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
3344 dsack_high = TCP_SKB_CB(skb)->end_seq;
3345 tcp_dsack_extend(tp, TCP_SKB_CB(skb)->seq, dsack);
3348 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3349 SOCK_DEBUG(sk, "ofo packet was already received \n");
3350 __skb_unlink(skb, &tp->out_of_order_queue);
3354 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
3355 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3356 TCP_SKB_CB(skb)->end_seq);
3358 __skb_unlink(skb, &tp->out_of_order_queue);
3359 __skb_queue_tail(&sk->sk_receive_queue, skb);
3360 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3362 tcp_fin(skb, sk, skb->h.th);
3366 static int tcp_prune_queue(struct sock *sk);
3368 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
3370 struct tcphdr *th = skb->h.th;
3371 struct tcp_sock *tp = tcp_sk(sk);
3374 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
3377 __skb_pull(skb, th->doff*4);
3379 TCP_ECN_accept_cwr(tp, skb);
3381 if (tp->rx_opt.dsack) {
3382 tp->rx_opt.dsack = 0;
3383 tp->rx_opt.eff_sacks = min_t(unsigned int, tp->rx_opt.num_sacks,
3384 4 - tp->rx_opt.tstamp_ok);
3387 /* Queue data for delivery to the user.
3388 * Packets in sequence go to the receive queue.
3389 * Out of sequence packets to the out_of_order_queue.
3391 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
3392 if (tcp_receive_window(tp) == 0)
3395 /* Ok. In sequence. In window. */
3396 if (tp->ucopy.task == current &&
3397 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
3398 sock_owned_by_user(sk) && !tp->urg_data) {
3399 int chunk = min_t(unsigned int, skb->len,
3402 __set_current_state(TASK_RUNNING);
3405 if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
3406 tp->ucopy.len -= chunk;
3407 tp->copied_seq += chunk;
3408 eaten = (chunk == skb->len && !th->fin);
3409 tcp_rcv_space_adjust(sk);
3417 (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3418 !sk_stream_rmem_schedule(sk, skb))) {
3419 if (tcp_prune_queue(sk) < 0 ||
3420 !sk_stream_rmem_schedule(sk, skb))
3423 sk_stream_set_owner_r(skb, sk);
3424 __skb_queue_tail(&sk->sk_receive_queue, skb);
3426 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3428 tcp_event_data_recv(sk, tp, skb);
3430 tcp_fin(skb, sk, th);
3432 if (!skb_queue_empty(&tp->out_of_order_queue)) {
3435 /* RFC2581. 4.2. SHOULD send immediate ACK, when
3436 * gap in queue is filled.
3438 if (skb_queue_empty(&tp->out_of_order_queue))
3439 inet_csk(sk)->icsk_ack.pingpong = 0;
3442 if (tp->rx_opt.num_sacks)
3443 tcp_sack_remove(tp);
3445 tcp_fast_path_check(sk, tp);
3449 else if (!sock_flag(sk, SOCK_DEAD))
3450 sk->sk_data_ready(sk, 0);
3454 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3455 /* A retransmit, 2nd most common case. Force an immediate ack. */
3456 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
3457 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3460 tcp_enter_quickack_mode(sk);
3461 inet_csk_schedule_ack(sk);
3467 /* Out of window. F.e. zero window probe. */
3468 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
3471 tcp_enter_quickack_mode(sk);
3473 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3474 /* Partial packet, seq < rcv_next < end_seq */
3475 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
3476 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3477 TCP_SKB_CB(skb)->end_seq);
3479 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
3481 /* If window is closed, drop tail of packet. But after
3482 * remembering D-SACK for its head made in previous line.
3484 if (!tcp_receive_window(tp))
3489 TCP_ECN_check_ce(tp, skb);
3491 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3492 !sk_stream_rmem_schedule(sk, skb)) {
3493 if (tcp_prune_queue(sk) < 0 ||
3494 !sk_stream_rmem_schedule(sk, skb))
3498 /* Disable header prediction. */
3500 inet_csk_schedule_ack(sk);
3502 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
3503 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3505 sk_stream_set_owner_r(skb, sk);
3507 if (!skb_peek(&tp->out_of_order_queue)) {
3508 /* Initial out of order segment, build 1 SACK. */
3509 if (tp->rx_opt.sack_ok) {
3510 tp->rx_opt.num_sacks = 1;
3511 tp->rx_opt.dsack = 0;
3512 tp->rx_opt.eff_sacks = 1;
3513 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
3514 tp->selective_acks[0].end_seq =
3515 TCP_SKB_CB(skb)->end_seq;
3517 __skb_queue_head(&tp->out_of_order_queue,skb);
3519 struct sk_buff *skb1 = tp->out_of_order_queue.prev;
3520 u32 seq = TCP_SKB_CB(skb)->seq;
3521 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3523 if (seq == TCP_SKB_CB(skb1)->end_seq) {
3524 __skb_append(skb1, skb, &tp->out_of_order_queue);
3526 if (!tp->rx_opt.num_sacks ||
3527 tp->selective_acks[0].end_seq != seq)
3530 /* Common case: data arrive in order after hole. */
3531 tp->selective_acks[0].end_seq = end_seq;
3535 /* Find place to insert this segment. */
3537 if (!after(TCP_SKB_CB(skb1)->seq, seq))
3539 } while ((skb1 = skb1->prev) !=
3540 (struct sk_buff*)&tp->out_of_order_queue);
3542 /* Do skb overlap to previous one? */
3543 if (skb1 != (struct sk_buff*)&tp->out_of_order_queue &&
3544 before(seq, TCP_SKB_CB(skb1)->end_seq)) {
3545 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
3546 /* All the bits are present. Drop. */
3548 tcp_dsack_set(tp, seq, end_seq);
3551 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
3552 /* Partial overlap. */
3553 tcp_dsack_set(tp, seq, TCP_SKB_CB(skb1)->end_seq);
3558 __skb_insert(skb, skb1, skb1->next, &tp->out_of_order_queue);
3560 /* And clean segments covered by new one as whole. */
3561 while ((skb1 = skb->next) !=
3562 (struct sk_buff*)&tp->out_of_order_queue &&
3563 after(end_seq, TCP_SKB_CB(skb1)->seq)) {
3564 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
3565 tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, end_seq);
3568 __skb_unlink(skb1, &tp->out_of_order_queue);
3569 tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, TCP_SKB_CB(skb1)->end_seq);
3574 if (tp->rx_opt.sack_ok)
3575 tcp_sack_new_ofo_skb(sk, seq, end_seq);
3579 /* Collapse contiguous sequence of skbs head..tail with
3580 * sequence numbers start..end.
3581 * Segments with FIN/SYN are not collapsed (only because this
3585 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
3586 struct sk_buff *head, struct sk_buff *tail,
3589 struct sk_buff *skb;
3591 /* First, check that queue is collapsible and find
3592 * the point where collapsing can be useful. */
3593 for (skb = head; skb != tail; ) {
3594 /* No new bits? It is possible on ofo queue. */
3595 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
3596 struct sk_buff *next = skb->next;
3597 __skb_unlink(skb, list);
3599 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
3604 /* The first skb to collapse is:
3606 * - bloated or contains data before "start" or
3607 * overlaps to the next one.
3609 if (!skb->h.th->syn && !skb->h.th->fin &&
3610 (tcp_win_from_space(skb->truesize) > skb->len ||
3611 before(TCP_SKB_CB(skb)->seq, start) ||
3612 (skb->next != tail &&
3613 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq)))
3616 /* Decided to skip this, advance start seq. */
3617 start = TCP_SKB_CB(skb)->end_seq;
3620 if (skb == tail || skb->h.th->syn || skb->h.th->fin)
3623 while (before(start, end)) {
3624 struct sk_buff *nskb;
3625 int header = skb_headroom(skb);
3626 int copy = SKB_MAX_ORDER(header, 0);
3628 /* Too big header? This can happen with IPv6. */
3631 if (end-start < copy)
3633 nskb = alloc_skb(copy+header, GFP_ATOMIC);
3636 skb_reserve(nskb, header);
3637 memcpy(nskb->head, skb->head, header);
3638 nskb->nh.raw = nskb->head + (skb->nh.raw-skb->head);
3639 nskb->h.raw = nskb->head + (skb->h.raw-skb->head);
3640 nskb->mac.raw = nskb->head + (skb->mac.raw-skb->head);
3641 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
3642 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
3643 __skb_insert(nskb, skb->prev, skb, list);
3644 sk_stream_set_owner_r(nskb, sk);
3646 /* Copy data, releasing collapsed skbs. */
3648 int offset = start - TCP_SKB_CB(skb)->seq;
3649 int size = TCP_SKB_CB(skb)->end_seq - start;
3653 size = min(copy, size);
3654 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
3656 TCP_SKB_CB(nskb)->end_seq += size;
3660 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
3661 struct sk_buff *next = skb->next;
3662 __skb_unlink(skb, list);
3664 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
3666 if (skb == tail || skb->h.th->syn || skb->h.th->fin)
3673 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
3674 * and tcp_collapse() them until all the queue is collapsed.
3676 static void tcp_collapse_ofo_queue(struct sock *sk)
3678 struct tcp_sock *tp = tcp_sk(sk);
3679 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
3680 struct sk_buff *head;
3686 start = TCP_SKB_CB(skb)->seq;
3687 end = TCP_SKB_CB(skb)->end_seq;
3693 /* Segment is terminated when we see gap or when
3694 * we are at the end of all the queue. */
3695 if (skb == (struct sk_buff *)&tp->out_of_order_queue ||
3696 after(TCP_SKB_CB(skb)->seq, end) ||
3697 before(TCP_SKB_CB(skb)->end_seq, start)) {
3698 tcp_collapse(sk, &tp->out_of_order_queue,
3699 head, skb, start, end);
3701 if (skb == (struct sk_buff *)&tp->out_of_order_queue)
3703 /* Start new segment */
3704 start = TCP_SKB_CB(skb)->seq;
3705 end = TCP_SKB_CB(skb)->end_seq;
3707 if (before(TCP_SKB_CB(skb)->seq, start))
3708 start = TCP_SKB_CB(skb)->seq;
3709 if (after(TCP_SKB_CB(skb)->end_seq, end))
3710 end = TCP_SKB_CB(skb)->end_seq;
3715 /* Reduce allocated memory if we can, trying to get
3716 * the socket within its memory limits again.
3718 * Return less than zero if we should start dropping frames
3719 * until the socket owning process reads some of the data
3720 * to stabilize the situation.
3722 static int tcp_prune_queue(struct sock *sk)
3724 struct tcp_sock *tp = tcp_sk(sk);
3726 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
3728 NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED);
3730 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
3731 tcp_clamp_window(sk, tp);
3732 else if (tcp_memory_pressure)
3733 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
3735 tcp_collapse_ofo_queue(sk);
3736 tcp_collapse(sk, &sk->sk_receive_queue,
3737 sk->sk_receive_queue.next,
3738 (struct sk_buff*)&sk->sk_receive_queue,
3739 tp->copied_seq, tp->rcv_nxt);
3740 sk_stream_mem_reclaim(sk);
3742 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
3745 /* Collapsing did not help, destructive actions follow.
3746 * This must not ever occur. */
3748 /* First, purge the out_of_order queue. */
3749 if (!skb_queue_empty(&tp->out_of_order_queue)) {
3750 NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED);
3751 __skb_queue_purge(&tp->out_of_order_queue);
3753 /* Reset SACK state. A conforming SACK implementation will
3754 * do the same at a timeout based retransmit. When a connection
3755 * is in a sad state like this, we care only about integrity
3756 * of the connection not performance.
3758 if (tp->rx_opt.sack_ok)
3759 tcp_sack_reset(&tp->rx_opt);
3760 sk_stream_mem_reclaim(sk);
3763 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
3766 /* If we are really being abused, tell the caller to silently
3767 * drop receive data on the floor. It will get retransmitted
3768 * and hopefully then we'll have sufficient space.
3770 NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED);
3772 /* Massive buffer overcommit. */
3778 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
3779 * As additional protections, we do not touch cwnd in retransmission phases,
3780 * and if application hit its sndbuf limit recently.
3782 void tcp_cwnd_application_limited(struct sock *sk)
3784 struct tcp_sock *tp = tcp_sk(sk);
3786 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
3787 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
3788 /* Limited by application or receiver window. */
3789 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
3790 u32 win_used = max(tp->snd_cwnd_used, init_win);
3791 if (win_used < tp->snd_cwnd) {
3792 tp->snd_ssthresh = tcp_current_ssthresh(sk);
3793 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
3795 tp->snd_cwnd_used = 0;
3797 tp->snd_cwnd_stamp = tcp_time_stamp;
3800 static int tcp_should_expand_sndbuf(struct sock *sk, struct tcp_sock *tp)
3802 /* If the user specified a specific send buffer setting, do
3805 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
3808 /* If we are under global TCP memory pressure, do not expand. */
3809 if (tcp_memory_pressure)
3812 /* If we are under soft global TCP memory pressure, do not expand. */
3813 if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0])
3816 /* If we filled the congestion window, do not expand. */
3817 if (tp->packets_out >= tp->snd_cwnd)
3823 /* When incoming ACK allowed to free some skb from write_queue,
3824 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
3825 * on the exit from tcp input handler.
3827 * PROBLEM: sndbuf expansion does not work well with largesend.
3829 static void tcp_new_space(struct sock *sk)
3831 struct tcp_sock *tp = tcp_sk(sk);
3833 if (tcp_should_expand_sndbuf(sk, tp)) {
3834 int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
3835 MAX_TCP_HEADER + 16 + sizeof(struct sk_buff),
3836 demanded = max_t(unsigned int, tp->snd_cwnd,
3837 tp->reordering + 1);
3838 sndmem *= 2*demanded;
3839 if (sndmem > sk->sk_sndbuf)
3840 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
3841 tp->snd_cwnd_stamp = tcp_time_stamp;
3844 sk->sk_write_space(sk);
3847 static void tcp_check_space(struct sock *sk)
3849 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
3850 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
3851 if (sk->sk_socket &&
3852 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
3857 static inline void tcp_data_snd_check(struct sock *sk, struct tcp_sock *tp)
3859 tcp_push_pending_frames(sk, tp);
3860 tcp_check_space(sk);
3864 * Check if sending an ack is needed.
3866 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
3868 struct tcp_sock *tp = tcp_sk(sk);
3870 /* More than one full frame received... */
3871 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss
3872 /* ... and right edge of window advances far enough.
3873 * (tcp_recvmsg() will send ACK otherwise). Or...
3875 && __tcp_select_window(sk) >= tp->rcv_wnd) ||
3876 /* We ACK each frame or... */
3877 tcp_in_quickack_mode(sk) ||
3878 /* We have out of order data. */
3880 skb_peek(&tp->out_of_order_queue))) {
3881 /* Then ack it now */
3884 /* Else, send delayed ack. */
3885 tcp_send_delayed_ack(sk);
3889 static inline void tcp_ack_snd_check(struct sock *sk)
3891 if (!inet_csk_ack_scheduled(sk)) {
3892 /* We sent a data segment already. */
3895 __tcp_ack_snd_check(sk, 1);
3899 * This routine is only called when we have urgent data
3900 * signaled. Its the 'slow' part of tcp_urg. It could be
3901 * moved inline now as tcp_urg is only called from one
3902 * place. We handle URGent data wrong. We have to - as
3903 * BSD still doesn't use the correction from RFC961.
3904 * For 1003.1g we should support a new option TCP_STDURG to permit
3905 * either form (or just set the sysctl tcp_stdurg).
3908 static void tcp_check_urg(struct sock * sk, struct tcphdr * th)
3910 struct tcp_sock *tp = tcp_sk(sk);
3911 u32 ptr = ntohs(th->urg_ptr);
3913 if (ptr && !sysctl_tcp_stdurg)
3915 ptr += ntohl(th->seq);
3917 /* Ignore urgent data that we've already seen and read. */
3918 if (after(tp->copied_seq, ptr))
3921 /* Do not replay urg ptr.
3923 * NOTE: interesting situation not covered by specs.
3924 * Misbehaving sender may send urg ptr, pointing to segment,
3925 * which we already have in ofo queue. We are not able to fetch
3926 * such data and will stay in TCP_URG_NOTYET until will be eaten
3927 * by recvmsg(). Seems, we are not obliged to handle such wicked
3928 * situations. But it is worth to think about possibility of some
3929 * DoSes using some hypothetical application level deadlock.
3931 if (before(ptr, tp->rcv_nxt))
3934 /* Do we already have a newer (or duplicate) urgent pointer? */
3935 if (tp->urg_data && !after(ptr, tp->urg_seq))
3938 /* Tell the world about our new urgent pointer. */
3941 /* We may be adding urgent data when the last byte read was
3942 * urgent. To do this requires some care. We cannot just ignore
3943 * tp->copied_seq since we would read the last urgent byte again
3944 * as data, nor can we alter copied_seq until this data arrives
3945 * or we break the semantics of SIOCATMARK (and thus sockatmark())
3947 * NOTE. Double Dutch. Rendering to plain English: author of comment
3948 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
3949 * and expect that both A and B disappear from stream. This is _wrong_.
3950 * Though this happens in BSD with high probability, this is occasional.
3951 * Any application relying on this is buggy. Note also, that fix "works"
3952 * only in this artificial test. Insert some normal data between A and B and we will
3953 * decline of BSD again. Verdict: it is better to remove to trap
3956 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
3957 !sock_flag(sk, SOCK_URGINLINE) &&
3958 tp->copied_seq != tp->rcv_nxt) {
3959 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
3961 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
3962 __skb_unlink(skb, &sk->sk_receive_queue);
3967 tp->urg_data = TCP_URG_NOTYET;
3970 /* Disable header prediction. */
3974 /* This is the 'fast' part of urgent handling. */
3975 static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th)
3977 struct tcp_sock *tp = tcp_sk(sk);
3979 /* Check if we get a new urgent pointer - normally not. */
3981 tcp_check_urg(sk,th);
3983 /* Do we wait for any urgent data? - normally not... */
3984 if (tp->urg_data == TCP_URG_NOTYET) {
3985 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
3988 /* Is the urgent pointer pointing into this packet? */
3989 if (ptr < skb->len) {
3991 if (skb_copy_bits(skb, ptr, &tmp, 1))
3993 tp->urg_data = TCP_URG_VALID | tmp;
3994 if (!sock_flag(sk, SOCK_DEAD))
3995 sk->sk_data_ready(sk, 0);
4000 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
4002 struct tcp_sock *tp = tcp_sk(sk);
4003 int chunk = skb->len - hlen;
4007 if (skb->ip_summed==CHECKSUM_UNNECESSARY)
4008 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
4010 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
4014 tp->ucopy.len -= chunk;
4015 tp->copied_seq += chunk;
4016 tcp_rcv_space_adjust(sk);
4023 static __sum16 __tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
4027 if (sock_owned_by_user(sk)) {
4029 result = __tcp_checksum_complete(skb);
4032 result = __tcp_checksum_complete(skb);
4037 static inline int tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
4039 return skb->ip_summed != CHECKSUM_UNNECESSARY &&
4040 __tcp_checksum_complete_user(sk, skb);
4043 #ifdef CONFIG_NET_DMA
4044 static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb, int hlen)
4046 struct tcp_sock *tp = tcp_sk(sk);
4047 int chunk = skb->len - hlen;
4049 int copied_early = 0;
4051 if (tp->ucopy.wakeup)
4054 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
4055 tp->ucopy.dma_chan = get_softnet_dma();
4057 if (tp->ucopy.dma_chan && skb->ip_summed == CHECKSUM_UNNECESSARY) {
4059 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
4060 skb, hlen, tp->ucopy.iov, chunk, tp->ucopy.pinned_list);
4065 tp->ucopy.dma_cookie = dma_cookie;
4068 tp->ucopy.len -= chunk;
4069 tp->copied_seq += chunk;
4070 tcp_rcv_space_adjust(sk);
4072 if ((tp->ucopy.len == 0) ||
4073 (tcp_flag_word(skb->h.th) & TCP_FLAG_PSH) ||
4074 (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
4075 tp->ucopy.wakeup = 1;
4076 sk->sk_data_ready(sk, 0);
4078 } else if (chunk > 0) {
4079 tp->ucopy.wakeup = 1;
4080 sk->sk_data_ready(sk, 0);
4083 return copied_early;
4085 #endif /* CONFIG_NET_DMA */
4088 * TCP receive function for the ESTABLISHED state.
4090 * It is split into a fast path and a slow path. The fast path is
4092 * - A zero window was announced from us - zero window probing
4093 * is only handled properly in the slow path.
4094 * - Out of order segments arrived.
4095 * - Urgent data is expected.
4096 * - There is no buffer space left
4097 * - Unexpected TCP flags/window values/header lengths are received
4098 * (detected by checking the TCP header against pred_flags)
4099 * - Data is sent in both directions. Fast path only supports pure senders
4100 * or pure receivers (this means either the sequence number or the ack
4101 * value must stay constant)
4102 * - Unexpected TCP option.
4104 * When these conditions are not satisfied it drops into a standard
4105 * receive procedure patterned after RFC793 to handle all cases.
4106 * The first three cases are guaranteed by proper pred_flags setting,
4107 * the rest is checked inline. Fast processing is turned on in
4108 * tcp_data_queue when everything is OK.
4110 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
4111 struct tcphdr *th, unsigned len)
4113 struct tcp_sock *tp = tcp_sk(sk);
4116 * Header prediction.
4117 * The code loosely follows the one in the famous
4118 * "30 instruction TCP receive" Van Jacobson mail.
4120 * Van's trick is to deposit buffers into socket queue
4121 * on a device interrupt, to call tcp_recv function
4122 * on the receive process context and checksum and copy
4123 * the buffer to user space. smart...
4125 * Our current scheme is not silly either but we take the
4126 * extra cost of the net_bh soft interrupt processing...
4127 * We do checksum and copy also but from device to kernel.
4130 tp->rx_opt.saw_tstamp = 0;
4132 /* pred_flags is 0xS?10 << 16 + snd_wnd
4133 * if header_prediction is to be made
4134 * 'S' will always be tp->tcp_header_len >> 2
4135 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
4136 * turn it off (when there are holes in the receive
4137 * space for instance)
4138 * PSH flag is ignored.
4141 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
4142 TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4143 int tcp_header_len = tp->tcp_header_len;
4145 /* Timestamp header prediction: tcp_header_len
4146 * is automatically equal to th->doff*4 due to pred_flags
4150 /* Check timestamp */
4151 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
4152 __be32 *ptr = (__be32 *)(th + 1);
4154 /* No? Slow path! */
4155 if (*ptr != htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
4156 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP))
4159 tp->rx_opt.saw_tstamp = 1;
4161 tp->rx_opt.rcv_tsval = ntohl(*ptr);
4163 tp->rx_opt.rcv_tsecr = ntohl(*ptr);
4165 /* If PAWS failed, check it more carefully in slow path */
4166 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
4169 /* DO NOT update ts_recent here, if checksum fails
4170 * and timestamp was corrupted part, it will result
4171 * in a hung connection since we will drop all
4172 * future packets due to the PAWS test.
4176 if (len <= tcp_header_len) {
4177 /* Bulk data transfer: sender */
4178 if (len == tcp_header_len) {
4179 /* Predicted packet is in window by definition.
4180 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4181 * Hence, check seq<=rcv_wup reduces to:
4183 if (tcp_header_len ==
4184 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
4185 tp->rcv_nxt == tp->rcv_wup)
4186 tcp_store_ts_recent(tp);
4188 /* We know that such packets are checksummed
4191 tcp_ack(sk, skb, 0);
4193 tcp_data_snd_check(sk, tp);
4195 } else { /* Header too small */
4196 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4201 int copied_early = 0;
4203 if (tp->copied_seq == tp->rcv_nxt &&
4204 len - tcp_header_len <= tp->ucopy.len) {
4205 #ifdef CONFIG_NET_DMA
4206 if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
4211 if (tp->ucopy.task == current && sock_owned_by_user(sk) && !copied_early) {
4212 __set_current_state(TASK_RUNNING);
4214 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
4218 /* Predicted packet is in window by definition.
4219 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4220 * Hence, check seq<=rcv_wup reduces to:
4222 if (tcp_header_len ==
4223 (sizeof(struct tcphdr) +
4224 TCPOLEN_TSTAMP_ALIGNED) &&
4225 tp->rcv_nxt == tp->rcv_wup)
4226 tcp_store_ts_recent(tp);
4228 tcp_rcv_rtt_measure_ts(sk, skb);
4230 __skb_pull(skb, tcp_header_len);
4231 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4232 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER);
4235 tcp_cleanup_rbuf(sk, skb->len);
4238 if (tcp_checksum_complete_user(sk, skb))
4241 /* Predicted packet is in window by definition.
4242 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4243 * Hence, check seq<=rcv_wup reduces to:
4245 if (tcp_header_len ==
4246 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
4247 tp->rcv_nxt == tp->rcv_wup)
4248 tcp_store_ts_recent(tp);
4250 tcp_rcv_rtt_measure_ts(sk, skb);
4252 if ((int)skb->truesize > sk->sk_forward_alloc)
4255 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS);
4257 /* Bulk data transfer: receiver */
4258 __skb_pull(skb,tcp_header_len);
4259 __skb_queue_tail(&sk->sk_receive_queue, skb);
4260 sk_stream_set_owner_r(skb, sk);
4261 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4264 tcp_event_data_recv(sk, tp, skb);
4266 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
4267 /* Well, only one small jumplet in fast path... */
4268 tcp_ack(sk, skb, FLAG_DATA);
4269 tcp_data_snd_check(sk, tp);
4270 if (!inet_csk_ack_scheduled(sk))
4274 __tcp_ack_snd_check(sk, 0);
4276 #ifdef CONFIG_NET_DMA
4278 __skb_queue_tail(&sk->sk_async_wait_queue, skb);
4284 sk->sk_data_ready(sk, 0);
4290 if (len < (th->doff<<2) || tcp_checksum_complete_user(sk, skb))
4294 * RFC1323: H1. Apply PAWS check first.
4296 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
4297 tcp_paws_discard(sk, skb)) {
4299 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
4300 tcp_send_dupack(sk, skb);
4303 /* Resets are accepted even if PAWS failed.
4305 ts_recent update must be made after we are sure
4306 that the packet is in window.
4311 * Standard slow path.
4314 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
4315 /* RFC793, page 37: "In all states except SYN-SENT, all reset
4316 * (RST) segments are validated by checking their SEQ-fields."
4317 * And page 69: "If an incoming segment is not acceptable,
4318 * an acknowledgment should be sent in reply (unless the RST bit
4319 * is set, if so drop the segment and return)".
4322 tcp_send_dupack(sk, skb);
4331 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
4333 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4334 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4335 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
4342 tcp_ack(sk, skb, FLAG_SLOWPATH);
4344 tcp_rcv_rtt_measure_ts(sk, skb);
4346 /* Process urgent data. */
4347 tcp_urg(sk, skb, th);
4349 /* step 7: process the segment text */
4350 tcp_data_queue(sk, skb);
4352 tcp_data_snd_check(sk, tp);
4353 tcp_ack_snd_check(sk);
4357 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4364 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
4365 struct tcphdr *th, unsigned len)
4367 struct tcp_sock *tp = tcp_sk(sk);
4368 struct inet_connection_sock *icsk = inet_csk(sk);
4369 int saved_clamp = tp->rx_opt.mss_clamp;
4371 tcp_parse_options(skb, &tp->rx_opt, 0);
4375 * "If the state is SYN-SENT then
4376 * first check the ACK bit
4377 * If the ACK bit is set
4378 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
4379 * a reset (unless the RST bit is set, if so drop
4380 * the segment and return)"
4382 * We do not send data with SYN, so that RFC-correct
4385 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
4386 goto reset_and_undo;
4388 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
4389 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
4391 NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED);
4392 goto reset_and_undo;
4395 /* Now ACK is acceptable.
4397 * "If the RST bit is set
4398 * If the ACK was acceptable then signal the user "error:
4399 * connection reset", drop the segment, enter CLOSED state,
4400 * delete TCB, and return."
4409 * "fifth, if neither of the SYN or RST bits is set then
4410 * drop the segment and return."
4416 goto discard_and_undo;
4419 * "If the SYN bit is on ...
4420 * are acceptable then ...
4421 * (our SYN has been ACKed), change the connection
4422 * state to ESTABLISHED..."
4425 TCP_ECN_rcv_synack(tp, th);
4427 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
4428 tcp_ack(sk, skb, FLAG_SLOWPATH);
4430 /* Ok.. it's good. Set up sequence numbers and
4431 * move to established.
4433 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
4434 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
4436 /* RFC1323: The window in SYN & SYN/ACK segments is
4439 tp->snd_wnd = ntohs(th->window);
4440 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(skb)->seq);
4442 if (!tp->rx_opt.wscale_ok) {
4443 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
4444 tp->window_clamp = min(tp->window_clamp, 65535U);
4447 if (tp->rx_opt.saw_tstamp) {
4448 tp->rx_opt.tstamp_ok = 1;
4449 tp->tcp_header_len =
4450 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
4451 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
4452 tcp_store_ts_recent(tp);
4454 tp->tcp_header_len = sizeof(struct tcphdr);
4457 if (tp->rx_opt.sack_ok && sysctl_tcp_fack)
4458 tp->rx_opt.sack_ok |= 2;
4461 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
4462 tcp_initialize_rcv_mss(sk);
4464 /* Remember, tcp_poll() does not lock socket!
4465 * Change state from SYN-SENT only after copied_seq
4466 * is initialized. */
4467 tp->copied_seq = tp->rcv_nxt;
4469 tcp_set_state(sk, TCP_ESTABLISHED);
4471 security_inet_conn_established(sk, skb);
4473 /* Make sure socket is routed, for correct metrics. */
4474 icsk->icsk_af_ops->rebuild_header(sk);
4476 tcp_init_metrics(sk);
4478 tcp_init_congestion_control(sk);
4480 /* Prevent spurious tcp_cwnd_restart() on first data
4483 tp->lsndtime = tcp_time_stamp;
4485 tcp_init_buffer_space(sk);
4487 if (sock_flag(sk, SOCK_KEEPOPEN))
4488 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
4490 if (!tp->rx_opt.snd_wscale)
4491 __tcp_fast_path_on(tp, tp->snd_wnd);
4495 if (!sock_flag(sk, SOCK_DEAD)) {
4496 sk->sk_state_change(sk);
4497 sk_wake_async(sk, 0, POLL_OUT);
4500 if (sk->sk_write_pending ||
4501 icsk->icsk_accept_queue.rskq_defer_accept ||
4502 icsk->icsk_ack.pingpong) {
4503 /* Save one ACK. Data will be ready after
4504 * several ticks, if write_pending is set.
4506 * It may be deleted, but with this feature tcpdumps
4507 * look so _wonderfully_ clever, that I was not able
4508 * to stand against the temptation 8) --ANK
4510 inet_csk_schedule_ack(sk);
4511 icsk->icsk_ack.lrcvtime = tcp_time_stamp;
4512 icsk->icsk_ack.ato = TCP_ATO_MIN;
4513 tcp_incr_quickack(sk);
4514 tcp_enter_quickack_mode(sk);
4515 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
4516 TCP_DELACK_MAX, TCP_RTO_MAX);
4527 /* No ACK in the segment */
4531 * "If the RST bit is set
4533 * Otherwise (no ACK) drop the segment and return."
4536 goto discard_and_undo;
4540 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && tcp_paws_check(&tp->rx_opt, 0))
4541 goto discard_and_undo;
4544 /* We see SYN without ACK. It is attempt of
4545 * simultaneous connect with crossed SYNs.
4546 * Particularly, it can be connect to self.
4548 tcp_set_state(sk, TCP_SYN_RECV);
4550 if (tp->rx_opt.saw_tstamp) {
4551 tp->rx_opt.tstamp_ok = 1;
4552 tcp_store_ts_recent(tp);
4553 tp->tcp_header_len =
4554 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
4556 tp->tcp_header_len = sizeof(struct tcphdr);
4559 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
4560 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
4562 /* RFC1323: The window in SYN & SYN/ACK segments is
4565 tp->snd_wnd = ntohs(th->window);
4566 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
4567 tp->max_window = tp->snd_wnd;
4569 TCP_ECN_rcv_syn(tp, th);
4572 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
4573 tcp_initialize_rcv_mss(sk);
4576 tcp_send_synack(sk);
4578 /* Note, we could accept data and URG from this segment.
4579 * There are no obstacles to make this.
4581 * However, if we ignore data in ACKless segments sometimes,
4582 * we have no reasons to accept it sometimes.
4583 * Also, seems the code doing it in step6 of tcp_rcv_state_process
4584 * is not flawless. So, discard packet for sanity.
4585 * Uncomment this return to process the data.
4592 /* "fifth, if neither of the SYN or RST bits is set then
4593 * drop the segment and return."
4597 tcp_clear_options(&tp->rx_opt);
4598 tp->rx_opt.mss_clamp = saved_clamp;
4602 tcp_clear_options(&tp->rx_opt);
4603 tp->rx_opt.mss_clamp = saved_clamp;
4609 * This function implements the receiving procedure of RFC 793 for
4610 * all states except ESTABLISHED and TIME_WAIT.
4611 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
4612 * address independent.
4615 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
4616 struct tcphdr *th, unsigned len)
4618 struct tcp_sock *tp = tcp_sk(sk);
4619 struct inet_connection_sock *icsk = inet_csk(sk);
4622 tp->rx_opt.saw_tstamp = 0;
4624 switch (sk->sk_state) {
4636 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
4639 /* Now we have several options: In theory there is
4640 * nothing else in the frame. KA9Q has an option to
4641 * send data with the syn, BSD accepts data with the
4642 * syn up to the [to be] advertised window and
4643 * Solaris 2.1 gives you a protocol error. For now
4644 * we just ignore it, that fits the spec precisely
4645 * and avoids incompatibilities. It would be nice in
4646 * future to drop through and process the data.
4648 * Now that TTCP is starting to be used we ought to
4650 * But, this leaves one open to an easy denial of
4651 * service attack, and SYN cookies can't defend
4652 * against this problem. So, we drop the data
4653 * in the interest of security over speed unless
4654 * it's still in use.
4662 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
4666 /* Do step6 onward by hand. */
4667 tcp_urg(sk, skb, th);
4669 tcp_data_snd_check(sk, tp);
4673 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
4674 tcp_paws_discard(sk, skb)) {
4676 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
4677 tcp_send_dupack(sk, skb);
4680 /* Reset is accepted even if it did not pass PAWS. */
4683 /* step 1: check sequence number */
4684 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
4686 tcp_send_dupack(sk, skb);
4690 /* step 2: check RST bit */
4696 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
4698 /* step 3: check security and precedence [ignored] */
4702 * Check for a SYN in window.
4704 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4705 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
4710 /* step 5: check the ACK field */
4712 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH);
4714 switch(sk->sk_state) {
4717 tp->copied_seq = tp->rcv_nxt;
4719 tcp_set_state(sk, TCP_ESTABLISHED);
4720 sk->sk_state_change(sk);
4722 /* Note, that this wakeup is only for marginal
4723 * crossed SYN case. Passively open sockets
4724 * are not waked up, because sk->sk_sleep ==
4725 * NULL and sk->sk_socket == NULL.
4727 if (sk->sk_socket) {
4728 sk_wake_async(sk,0,POLL_OUT);
4731 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
4732 tp->snd_wnd = ntohs(th->window) <<
4733 tp->rx_opt.snd_wscale;
4734 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq,
4735 TCP_SKB_CB(skb)->seq);
4737 /* tcp_ack considers this ACK as duplicate
4738 * and does not calculate rtt.
4739 * Fix it at least with timestamps.
4741 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
4743 tcp_ack_saw_tstamp(sk, 0);
4745 if (tp->rx_opt.tstamp_ok)
4746 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
4748 /* Make sure socket is routed, for
4751 icsk->icsk_af_ops->rebuild_header(sk);
4753 tcp_init_metrics(sk);
4755 tcp_init_congestion_control(sk);
4757 /* Prevent spurious tcp_cwnd_restart() on
4758 * first data packet.
4760 tp->lsndtime = tcp_time_stamp;
4763 tcp_initialize_rcv_mss(sk);
4764 tcp_init_buffer_space(sk);
4765 tcp_fast_path_on(tp);
4772 if (tp->snd_una == tp->write_seq) {
4773 tcp_set_state(sk, TCP_FIN_WAIT2);
4774 sk->sk_shutdown |= SEND_SHUTDOWN;
4775 dst_confirm(sk->sk_dst_cache);
4777 if (!sock_flag(sk, SOCK_DEAD))
4778 /* Wake up lingering close() */
4779 sk->sk_state_change(sk);
4783 if (tp->linger2 < 0 ||
4784 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4785 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
4787 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
4791 tmo = tcp_fin_time(sk);
4792 if (tmo > TCP_TIMEWAIT_LEN) {
4793 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
4794 } else if (th->fin || sock_owned_by_user(sk)) {
4795 /* Bad case. We could lose such FIN otherwise.
4796 * It is not a big problem, but it looks confusing
4797 * and not so rare event. We still can lose it now,
4798 * if it spins in bh_lock_sock(), but it is really
4801 inet_csk_reset_keepalive_timer(sk, tmo);
4803 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
4811 if (tp->snd_una == tp->write_seq) {
4812 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4818 if (tp->snd_una == tp->write_seq) {
4819 tcp_update_metrics(sk);
4828 /* step 6: check the URG bit */
4829 tcp_urg(sk, skb, th);
4831 /* step 7: process the segment text */
4832 switch (sk->sk_state) {
4833 case TCP_CLOSE_WAIT:
4836 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4840 /* RFC 793 says to queue data in these states,
4841 * RFC 1122 says we MUST send a reset.
4842 * BSD 4.4 also does reset.
4844 if (sk->sk_shutdown & RCV_SHUTDOWN) {
4845 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4846 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
4847 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
4853 case TCP_ESTABLISHED:
4854 tcp_data_queue(sk, skb);
4859 /* tcp_data could move socket to TIME-WAIT */
4860 if (sk->sk_state != TCP_CLOSE) {
4861 tcp_data_snd_check(sk, tp);
4862 tcp_ack_snd_check(sk);
4872 EXPORT_SYMBOL(sysctl_tcp_ecn);
4873 EXPORT_SYMBOL(sysctl_tcp_reordering);
4874 EXPORT_SYMBOL(tcp_parse_options);
4875 EXPORT_SYMBOL(tcp_rcv_established);
4876 EXPORT_SYMBOL(tcp_rcv_state_process);
4877 EXPORT_SYMBOL(tcp_initialize_rcv_mss);