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 */
105 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
106 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained DSACK info */
108 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
109 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
110 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
111 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
112 #define FLAG_ANY_PROGRESS (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED)
114 #define IsReno(tp) ((tp)->rx_opt.sack_ok == 0)
115 #define IsFack(tp) ((tp)->rx_opt.sack_ok & 2)
116 #define IsDSack(tp) ((tp)->rx_opt.sack_ok & 4)
118 #define IsSackFrto() (sysctl_tcp_frto == 0x2)
120 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
122 /* Adapt the MSS value used to make delayed ack decision to the
125 static void tcp_measure_rcv_mss(struct sock *sk,
126 const struct sk_buff *skb)
128 struct inet_connection_sock *icsk = inet_csk(sk);
129 const unsigned int lss = icsk->icsk_ack.last_seg_size;
132 icsk->icsk_ack.last_seg_size = 0;
134 /* skb->len may jitter because of SACKs, even if peer
135 * sends good full-sized frames.
137 len = skb_shinfo(skb)->gso_size ?: skb->len;
138 if (len >= icsk->icsk_ack.rcv_mss) {
139 icsk->icsk_ack.rcv_mss = len;
141 /* Otherwise, we make more careful check taking into account,
142 * that SACKs block is variable.
144 * "len" is invariant segment length, including TCP header.
146 len += skb->data - skb_transport_header(skb);
147 if (len >= TCP_MIN_RCVMSS + sizeof(struct tcphdr) ||
148 /* If PSH is not set, packet should be
149 * full sized, provided peer TCP is not badly broken.
150 * This observation (if it is correct 8)) allows
151 * to handle super-low mtu links fairly.
153 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
154 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
155 /* Subtract also invariant (if peer is RFC compliant),
156 * tcp header plus fixed timestamp option length.
157 * Resulting "len" is MSS free of SACK jitter.
159 len -= tcp_sk(sk)->tcp_header_len;
160 icsk->icsk_ack.last_seg_size = len;
162 icsk->icsk_ack.rcv_mss = len;
166 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
167 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
168 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
172 static void tcp_incr_quickack(struct sock *sk)
174 struct inet_connection_sock *icsk = inet_csk(sk);
175 unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
179 if (quickacks > icsk->icsk_ack.quick)
180 icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
183 void tcp_enter_quickack_mode(struct sock *sk)
185 struct inet_connection_sock *icsk = inet_csk(sk);
186 tcp_incr_quickack(sk);
187 icsk->icsk_ack.pingpong = 0;
188 icsk->icsk_ack.ato = TCP_ATO_MIN;
191 /* Send ACKs quickly, if "quick" count is not exhausted
192 * and the session is not interactive.
195 static inline int tcp_in_quickack_mode(const struct sock *sk)
197 const struct inet_connection_sock *icsk = inet_csk(sk);
198 return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
201 /* Buffer size and advertised window tuning.
203 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
206 static void tcp_fixup_sndbuf(struct sock *sk)
208 int sndmem = tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER + 16 +
209 sizeof(struct sk_buff);
211 if (sk->sk_sndbuf < 3 * sndmem)
212 sk->sk_sndbuf = min(3 * sndmem, sysctl_tcp_wmem[2]);
215 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
217 * All tcp_full_space() is split to two parts: "network" buffer, allocated
218 * forward and advertised in receiver window (tp->rcv_wnd) and
219 * "application buffer", required to isolate scheduling/application
220 * latencies from network.
221 * window_clamp is maximal advertised window. It can be less than
222 * tcp_full_space(), in this case tcp_full_space() - window_clamp
223 * is reserved for "application" buffer. The less window_clamp is
224 * the smoother our behaviour from viewpoint of network, but the lower
225 * throughput and the higher sensitivity of the connection to losses. 8)
227 * rcv_ssthresh is more strict window_clamp used at "slow start"
228 * phase to predict further behaviour of this connection.
229 * It is used for two goals:
230 * - to enforce header prediction at sender, even when application
231 * requires some significant "application buffer". It is check #1.
232 * - to prevent pruning of receive queue because of misprediction
233 * of receiver window. Check #2.
235 * The scheme does not work when sender sends good segments opening
236 * window and then starts to feed us spaghetti. But it should work
237 * in common situations. Otherwise, we have to rely on queue collapsing.
240 /* Slow part of check#2. */
241 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
243 struct tcp_sock *tp = tcp_sk(sk);
245 int truesize = tcp_win_from_space(skb->truesize)/2;
246 int window = tcp_win_from_space(sysctl_tcp_rmem[2])/2;
248 while (tp->rcv_ssthresh <= window) {
249 if (truesize <= skb->len)
250 return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
258 static void tcp_grow_window(struct sock *sk,
261 struct tcp_sock *tp = tcp_sk(sk);
264 if (tp->rcv_ssthresh < tp->window_clamp &&
265 (int)tp->rcv_ssthresh < tcp_space(sk) &&
266 !tcp_memory_pressure) {
269 /* Check #2. Increase window, if skb with such overhead
270 * will fit to rcvbuf in future.
272 if (tcp_win_from_space(skb->truesize) <= skb->len)
275 incr = __tcp_grow_window(sk, skb);
278 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr, tp->window_clamp);
279 inet_csk(sk)->icsk_ack.quick |= 1;
284 /* 3. Tuning rcvbuf, when connection enters established state. */
286 static void tcp_fixup_rcvbuf(struct sock *sk)
288 struct tcp_sock *tp = tcp_sk(sk);
289 int rcvmem = tp->advmss + MAX_TCP_HEADER + 16 + sizeof(struct sk_buff);
291 /* Try to select rcvbuf so that 4 mss-sized segments
292 * will fit to window and corresponding skbs will fit to our rcvbuf.
293 * (was 3; 4 is minimum to allow fast retransmit to work.)
295 while (tcp_win_from_space(rcvmem) < tp->advmss)
297 if (sk->sk_rcvbuf < 4 * rcvmem)
298 sk->sk_rcvbuf = min(4 * rcvmem, sysctl_tcp_rmem[2]);
301 /* 4. Try to fixup all. It is made immediately after connection enters
304 static void tcp_init_buffer_space(struct sock *sk)
306 struct tcp_sock *tp = tcp_sk(sk);
309 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
310 tcp_fixup_rcvbuf(sk);
311 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
312 tcp_fixup_sndbuf(sk);
314 tp->rcvq_space.space = tp->rcv_wnd;
316 maxwin = tcp_full_space(sk);
318 if (tp->window_clamp >= maxwin) {
319 tp->window_clamp = maxwin;
321 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
322 tp->window_clamp = max(maxwin -
323 (maxwin >> sysctl_tcp_app_win),
327 /* Force reservation of one segment. */
328 if (sysctl_tcp_app_win &&
329 tp->window_clamp > 2 * tp->advmss &&
330 tp->window_clamp + tp->advmss > maxwin)
331 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
333 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
334 tp->snd_cwnd_stamp = tcp_time_stamp;
337 /* 5. Recalculate window clamp after socket hit its memory bounds. */
338 static void tcp_clamp_window(struct sock *sk)
340 struct tcp_sock *tp = tcp_sk(sk);
341 struct inet_connection_sock *icsk = inet_csk(sk);
343 icsk->icsk_ack.quick = 0;
345 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
346 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
347 !tcp_memory_pressure &&
348 atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) {
349 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
352 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
353 tp->rcv_ssthresh = min(tp->window_clamp, 2U*tp->advmss);
357 /* Initialize RCV_MSS value.
358 * RCV_MSS is an our guess about MSS used by the peer.
359 * We haven't any direct information about the MSS.
360 * It's better to underestimate the RCV_MSS rather than overestimate.
361 * Overestimations make us ACKing less frequently than needed.
362 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
364 void tcp_initialize_rcv_mss(struct sock *sk)
366 struct tcp_sock *tp = tcp_sk(sk);
367 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
369 hint = min(hint, tp->rcv_wnd/2);
370 hint = min(hint, TCP_MIN_RCVMSS);
371 hint = max(hint, TCP_MIN_MSS);
373 inet_csk(sk)->icsk_ack.rcv_mss = hint;
376 /* Receiver "autotuning" code.
378 * The algorithm for RTT estimation w/o timestamps is based on
379 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
380 * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps>
382 * More detail on this code can be found at
383 * <http://www.psc.edu/~jheffner/senior_thesis.ps>,
384 * though this reference is out of date. A new paper
387 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
389 u32 new_sample = tp->rcv_rtt_est.rtt;
395 if (new_sample != 0) {
396 /* If we sample in larger samples in the non-timestamp
397 * case, we could grossly overestimate the RTT especially
398 * with chatty applications or bulk transfer apps which
399 * are stalled on filesystem I/O.
401 * Also, since we are only going for a minimum in the
402 * non-timestamp case, we do not smooth things out
403 * else with timestamps disabled convergence takes too
407 m -= (new_sample >> 3);
409 } else if (m < new_sample)
412 /* No previous measure. */
416 if (tp->rcv_rtt_est.rtt != new_sample)
417 tp->rcv_rtt_est.rtt = new_sample;
420 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
422 if (tp->rcv_rtt_est.time == 0)
424 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
426 tcp_rcv_rtt_update(tp,
427 jiffies - tp->rcv_rtt_est.time,
431 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
432 tp->rcv_rtt_est.time = tcp_time_stamp;
435 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, const struct sk_buff *skb)
437 struct tcp_sock *tp = tcp_sk(sk);
438 if (tp->rx_opt.rcv_tsecr &&
439 (TCP_SKB_CB(skb)->end_seq -
440 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
441 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
445 * This function should be called every time data is copied to user space.
446 * It calculates the appropriate TCP receive buffer space.
448 void tcp_rcv_space_adjust(struct sock *sk)
450 struct tcp_sock *tp = tcp_sk(sk);
454 if (tp->rcvq_space.time == 0)
457 time = tcp_time_stamp - tp->rcvq_space.time;
458 if (time < (tp->rcv_rtt_est.rtt >> 3) ||
459 tp->rcv_rtt_est.rtt == 0)
462 space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
464 space = max(tp->rcvq_space.space, space);
466 if (tp->rcvq_space.space != space) {
469 tp->rcvq_space.space = space;
471 if (sysctl_tcp_moderate_rcvbuf &&
472 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
473 int new_clamp = space;
475 /* Receive space grows, normalize in order to
476 * take into account packet headers and sk_buff
477 * structure overhead.
482 rcvmem = (tp->advmss + MAX_TCP_HEADER +
483 16 + sizeof(struct sk_buff));
484 while (tcp_win_from_space(rcvmem) < tp->advmss)
487 space = min(space, sysctl_tcp_rmem[2]);
488 if (space > sk->sk_rcvbuf) {
489 sk->sk_rcvbuf = space;
491 /* Make the window clamp follow along. */
492 tp->window_clamp = new_clamp;
498 tp->rcvq_space.seq = tp->copied_seq;
499 tp->rcvq_space.time = tcp_time_stamp;
502 /* There is something which you must keep in mind when you analyze the
503 * behavior of the tp->ato delayed ack timeout interval. When a
504 * connection starts up, we want to ack as quickly as possible. The
505 * problem is that "good" TCP's do slow start at the beginning of data
506 * transmission. The means that until we send the first few ACK's the
507 * sender will sit on his end and only queue most of his data, because
508 * he can only send snd_cwnd unacked packets at any given time. For
509 * each ACK we send, he increments snd_cwnd and transmits more of his
512 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
514 struct tcp_sock *tp = tcp_sk(sk);
515 struct inet_connection_sock *icsk = inet_csk(sk);
518 inet_csk_schedule_ack(sk);
520 tcp_measure_rcv_mss(sk, skb);
522 tcp_rcv_rtt_measure(tp);
524 now = tcp_time_stamp;
526 if (!icsk->icsk_ack.ato) {
527 /* The _first_ data packet received, initialize
528 * delayed ACK engine.
530 tcp_incr_quickack(sk);
531 icsk->icsk_ack.ato = TCP_ATO_MIN;
533 int m = now - icsk->icsk_ack.lrcvtime;
535 if (m <= TCP_ATO_MIN/2) {
536 /* The fastest case is the first. */
537 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
538 } else if (m < icsk->icsk_ack.ato) {
539 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
540 if (icsk->icsk_ack.ato > icsk->icsk_rto)
541 icsk->icsk_ack.ato = icsk->icsk_rto;
542 } else if (m > icsk->icsk_rto) {
543 /* Too long gap. Apparently sender failed to
544 * restart window, so that we send ACKs quickly.
546 tcp_incr_quickack(sk);
547 sk_stream_mem_reclaim(sk);
550 icsk->icsk_ack.lrcvtime = now;
552 TCP_ECN_check_ce(tp, skb);
555 tcp_grow_window(sk, skb);
558 static u32 tcp_rto_min(struct sock *sk)
560 struct dst_entry *dst = __sk_dst_get(sk);
561 u32 rto_min = TCP_RTO_MIN;
563 if (dst && dst_metric_locked(dst, RTAX_RTO_MIN))
564 rto_min = dst->metrics[RTAX_RTO_MIN-1];
568 /* Called to compute a smoothed rtt estimate. The data fed to this
569 * routine either comes from timestamps, or from segments that were
570 * known _not_ to have been retransmitted [see Karn/Partridge
571 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
572 * piece by Van Jacobson.
573 * NOTE: the next three routines used to be one big routine.
574 * To save cycles in the RFC 1323 implementation it was better to break
575 * it up into three procedures. -- erics
577 static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
579 struct tcp_sock *tp = tcp_sk(sk);
580 long m = mrtt; /* RTT */
582 /* The following amusing code comes from Jacobson's
583 * article in SIGCOMM '88. Note that rtt and mdev
584 * are scaled versions of rtt and mean deviation.
585 * This is designed to be as fast as possible
586 * m stands for "measurement".
588 * On a 1990 paper the rto value is changed to:
589 * RTO = rtt + 4 * mdev
591 * Funny. This algorithm seems to be very broken.
592 * These formulae increase RTO, when it should be decreased, increase
593 * too slowly, when it should be increased quickly, decrease too quickly
594 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
595 * does not matter how to _calculate_ it. Seems, it was trap
596 * that VJ failed to avoid. 8)
601 m -= (tp->srtt >> 3); /* m is now error in rtt est */
602 tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */
604 m = -m; /* m is now abs(error) */
605 m -= (tp->mdev >> 2); /* similar update on mdev */
606 /* This is similar to one of Eifel findings.
607 * Eifel blocks mdev updates when rtt decreases.
608 * This solution is a bit different: we use finer gain
609 * for mdev in this case (alpha*beta).
610 * Like Eifel it also prevents growth of rto,
611 * but also it limits too fast rto decreases,
612 * happening in pure Eifel.
617 m -= (tp->mdev >> 2); /* similar update on mdev */
619 tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */
620 if (tp->mdev > tp->mdev_max) {
621 tp->mdev_max = tp->mdev;
622 if (tp->mdev_max > tp->rttvar)
623 tp->rttvar = tp->mdev_max;
625 if (after(tp->snd_una, tp->rtt_seq)) {
626 if (tp->mdev_max < tp->rttvar)
627 tp->rttvar -= (tp->rttvar-tp->mdev_max)>>2;
628 tp->rtt_seq = tp->snd_nxt;
629 tp->mdev_max = tcp_rto_min(sk);
632 /* no previous measure. */
633 tp->srtt = m<<3; /* take the measured time to be rtt */
634 tp->mdev = m<<1; /* make sure rto = 3*rtt */
635 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
636 tp->rtt_seq = tp->snd_nxt;
640 /* Calculate rto without backoff. This is the second half of Van Jacobson's
641 * routine referred to above.
643 static inline void tcp_set_rto(struct sock *sk)
645 const struct tcp_sock *tp = tcp_sk(sk);
646 /* Old crap is replaced with new one. 8)
649 * 1. If rtt variance happened to be less 50msec, it is hallucination.
650 * It cannot be less due to utterly erratic ACK generation made
651 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
652 * to do with delayed acks, because at cwnd>2 true delack timeout
653 * is invisible. Actually, Linux-2.4 also generates erratic
654 * ACKs in some circumstances.
656 inet_csk(sk)->icsk_rto = (tp->srtt >> 3) + tp->rttvar;
658 /* 2. Fixups made earlier cannot be right.
659 * If we do not estimate RTO correctly without them,
660 * all the algo is pure shit and should be replaced
661 * with correct one. It is exactly, which we pretend to do.
665 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
666 * guarantees that rto is higher.
668 static inline void tcp_bound_rto(struct sock *sk)
670 if (inet_csk(sk)->icsk_rto > TCP_RTO_MAX)
671 inet_csk(sk)->icsk_rto = TCP_RTO_MAX;
674 /* Save metrics learned by this TCP session.
675 This function is called only, when TCP finishes successfully
676 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
678 void tcp_update_metrics(struct sock *sk)
680 struct tcp_sock *tp = tcp_sk(sk);
681 struct dst_entry *dst = __sk_dst_get(sk);
683 if (sysctl_tcp_nometrics_save)
688 if (dst && (dst->flags&DST_HOST)) {
689 const struct inet_connection_sock *icsk = inet_csk(sk);
692 if (icsk->icsk_backoff || !tp->srtt) {
693 /* This session failed to estimate rtt. Why?
694 * Probably, no packets returned in time.
697 if (!(dst_metric_locked(dst, RTAX_RTT)))
698 dst->metrics[RTAX_RTT-1] = 0;
702 m = dst_metric(dst, RTAX_RTT) - tp->srtt;
704 /* If newly calculated rtt larger than stored one,
705 * store new one. Otherwise, use EWMA. Remember,
706 * rtt overestimation is always better than underestimation.
708 if (!(dst_metric_locked(dst, RTAX_RTT))) {
710 dst->metrics[RTAX_RTT-1] = tp->srtt;
712 dst->metrics[RTAX_RTT-1] -= (m>>3);
715 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) {
719 /* Scale deviation to rttvar fixed point */
724 if (m >= dst_metric(dst, RTAX_RTTVAR))
725 dst->metrics[RTAX_RTTVAR-1] = m;
727 dst->metrics[RTAX_RTTVAR-1] -=
728 (dst->metrics[RTAX_RTTVAR-1] - m)>>2;
731 if (tp->snd_ssthresh >= 0xFFFF) {
732 /* Slow start still did not finish. */
733 if (dst_metric(dst, RTAX_SSTHRESH) &&
734 !dst_metric_locked(dst, RTAX_SSTHRESH) &&
735 (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH))
736 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1;
737 if (!dst_metric_locked(dst, RTAX_CWND) &&
738 tp->snd_cwnd > dst_metric(dst, RTAX_CWND))
739 dst->metrics[RTAX_CWND-1] = tp->snd_cwnd;
740 } else if (tp->snd_cwnd > tp->snd_ssthresh &&
741 icsk->icsk_ca_state == TCP_CA_Open) {
742 /* Cong. avoidance phase, cwnd is reliable. */
743 if (!dst_metric_locked(dst, RTAX_SSTHRESH))
744 dst->metrics[RTAX_SSTHRESH-1] =
745 max(tp->snd_cwnd >> 1, tp->snd_ssthresh);
746 if (!dst_metric_locked(dst, RTAX_CWND))
747 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_cwnd) >> 1;
749 /* Else slow start did not finish, cwnd is non-sense,
750 ssthresh may be also invalid.
752 if (!dst_metric_locked(dst, RTAX_CWND))
753 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_ssthresh) >> 1;
754 if (dst->metrics[RTAX_SSTHRESH-1] &&
755 !dst_metric_locked(dst, RTAX_SSTHRESH) &&
756 tp->snd_ssthresh > dst->metrics[RTAX_SSTHRESH-1])
757 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh;
760 if (!dst_metric_locked(dst, RTAX_REORDERING)) {
761 if (dst->metrics[RTAX_REORDERING-1] < tp->reordering &&
762 tp->reordering != sysctl_tcp_reordering)
763 dst->metrics[RTAX_REORDERING-1] = tp->reordering;
768 /* Numbers are taken from RFC3390.
770 * John Heffner states:
772 * The RFC specifies a window of no more than 4380 bytes
773 * unless 2*MSS > 4380. Reading the pseudocode in the RFC
774 * is a bit misleading because they use a clamp at 4380 bytes
775 * rather than use a multiplier in the relevant range.
777 __u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst)
779 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
782 if (tp->mss_cache > 1460)
785 cwnd = (tp->mss_cache > 1095) ? 3 : 4;
787 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
790 /* Set slow start threshold and cwnd not falling to slow start */
791 void tcp_enter_cwr(struct sock *sk, const int set_ssthresh)
793 struct tcp_sock *tp = tcp_sk(sk);
794 const struct inet_connection_sock *icsk = inet_csk(sk);
796 tp->prior_ssthresh = 0;
798 if (icsk->icsk_ca_state < TCP_CA_CWR) {
801 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
802 tp->snd_cwnd = min(tp->snd_cwnd,
803 tcp_packets_in_flight(tp) + 1U);
804 tp->snd_cwnd_cnt = 0;
805 tp->high_seq = tp->snd_nxt;
806 tp->snd_cwnd_stamp = tcp_time_stamp;
807 TCP_ECN_queue_cwr(tp);
809 tcp_set_ca_state(sk, TCP_CA_CWR);
813 /* Initialize metrics on socket. */
815 static void tcp_init_metrics(struct sock *sk)
817 struct tcp_sock *tp = tcp_sk(sk);
818 struct dst_entry *dst = __sk_dst_get(sk);
825 if (dst_metric_locked(dst, RTAX_CWND))
826 tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND);
827 if (dst_metric(dst, RTAX_SSTHRESH)) {
828 tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH);
829 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
830 tp->snd_ssthresh = tp->snd_cwnd_clamp;
832 if (dst_metric(dst, RTAX_REORDERING) &&
833 tp->reordering != dst_metric(dst, RTAX_REORDERING)) {
834 tp->rx_opt.sack_ok &= ~2;
835 tp->reordering = dst_metric(dst, RTAX_REORDERING);
838 if (dst_metric(dst, RTAX_RTT) == 0)
841 if (!tp->srtt && dst_metric(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3))
844 /* Initial rtt is determined from SYN,SYN-ACK.
845 * The segment is small and rtt may appear much
846 * less than real one. Use per-dst memory
847 * to make it more realistic.
849 * A bit of theory. RTT is time passed after "normal" sized packet
850 * is sent until it is ACKed. In normal circumstances sending small
851 * packets force peer to delay ACKs and calculation is correct too.
852 * The algorithm is adaptive and, provided we follow specs, it
853 * NEVER underestimate RTT. BUT! If peer tries to make some clever
854 * tricks sort of "quick acks" for time long enough to decrease RTT
855 * to low value, and then abruptly stops to do it and starts to delay
856 * ACKs, wait for troubles.
858 if (dst_metric(dst, RTAX_RTT) > tp->srtt) {
859 tp->srtt = dst_metric(dst, RTAX_RTT);
860 tp->rtt_seq = tp->snd_nxt;
862 if (dst_metric(dst, RTAX_RTTVAR) > tp->mdev) {
863 tp->mdev = dst_metric(dst, RTAX_RTTVAR);
864 tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN);
868 if (inet_csk(sk)->icsk_rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp)
870 tp->snd_cwnd = tcp_init_cwnd(tp, dst);
871 tp->snd_cwnd_stamp = tcp_time_stamp;
875 /* Play conservative. If timestamps are not
876 * supported, TCP will fail to recalculate correct
877 * rtt, if initial rto is too small. FORGET ALL AND RESET!
879 if (!tp->rx_opt.saw_tstamp && tp->srtt) {
881 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT;
882 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
886 static void tcp_update_reordering(struct sock *sk, const int metric,
889 struct tcp_sock *tp = tcp_sk(sk);
890 if (metric > tp->reordering) {
891 tp->reordering = min(TCP_MAX_REORDERING, metric);
893 /* This exciting event is worth to be remembered. 8) */
895 NET_INC_STATS_BH(LINUX_MIB_TCPTSREORDER);
897 NET_INC_STATS_BH(LINUX_MIB_TCPRENOREORDER);
899 NET_INC_STATS_BH(LINUX_MIB_TCPFACKREORDER);
901 NET_INC_STATS_BH(LINUX_MIB_TCPSACKREORDER);
902 #if FASTRETRANS_DEBUG > 1
903 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n",
904 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
908 tp->undo_marker ? tp->undo_retrans : 0);
910 /* Disable FACK yet. */
911 tp->rx_opt.sack_ok &= ~2;
915 /* This procedure tags the retransmission queue when SACKs arrive.
917 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
918 * Packets in queue with these bits set are counted in variables
919 * sacked_out, retrans_out and lost_out, correspondingly.
921 * Valid combinations are:
922 * Tag InFlight Description
923 * 0 1 - orig segment is in flight.
924 * S 0 - nothing flies, orig reached receiver.
925 * L 0 - nothing flies, orig lost by net.
926 * R 2 - both orig and retransmit are in flight.
927 * L|R 1 - orig is lost, retransmit is in flight.
928 * S|R 1 - orig reached receiver, retrans is still in flight.
929 * (L|S|R is logically valid, it could occur when L|R is sacked,
930 * but it is equivalent to plain S and code short-curcuits it to S.
931 * L|S is logically invalid, it would mean -1 packet in flight 8))
933 * These 6 states form finite state machine, controlled by the following events:
934 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
935 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
936 * 3. Loss detection event of one of three flavors:
937 * A. Scoreboard estimator decided the packet is lost.
938 * A'. Reno "three dupacks" marks head of queue lost.
939 * A''. Its FACK modfication, head until snd.fack is lost.
940 * B. SACK arrives sacking data transmitted after never retransmitted
942 * C. SACK arrives sacking SND.NXT at the moment, when the
943 * segment was retransmitted.
944 * 4. D-SACK added new rule: D-SACK changes any tag to S.
946 * It is pleasant to note, that state diagram turns out to be commutative,
947 * so that we are allowed not to be bothered by order of our actions,
948 * when multiple events arrive simultaneously. (see the function below).
950 * Reordering detection.
951 * --------------------
952 * Reordering metric is maximal distance, which a packet can be displaced
953 * in packet stream. With SACKs we can estimate it:
955 * 1. SACK fills old hole and the corresponding segment was not
956 * ever retransmitted -> reordering. Alas, we cannot use it
957 * when segment was retransmitted.
958 * 2. The last flaw is solved with D-SACK. D-SACK arrives
959 * for retransmitted and already SACKed segment -> reordering..
960 * Both of these heuristics are not used in Loss state, when we cannot
961 * account for retransmits accurately.
963 static int tcp_check_dsack(struct tcp_sock *tp, struct sk_buff *ack_skb,
964 struct tcp_sack_block_wire *sp, int num_sacks,
967 u32 start_seq_0 = ntohl(get_unaligned(&sp[0].start_seq));
968 u32 end_seq_0 = ntohl(get_unaligned(&sp[0].end_seq));
971 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
973 tp->rx_opt.sack_ok |= 4;
974 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV);
975 } else if (num_sacks > 1) {
976 u32 end_seq_1 = ntohl(get_unaligned(&sp[1].end_seq));
977 u32 start_seq_1 = ntohl(get_unaligned(&sp[1].start_seq));
979 if (!after(end_seq_0, end_seq_1) &&
980 !before(start_seq_0, start_seq_1)) {
982 tp->rx_opt.sack_ok |= 4;
983 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV);
987 /* D-SACK for already forgotten data... Do dumb counting. */
989 !after(end_seq_0, prior_snd_una) &&
990 after(end_seq_0, tp->undo_marker))
997 tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, u32 prior_snd_una)
999 const struct inet_connection_sock *icsk = inet_csk(sk);
1000 struct tcp_sock *tp = tcp_sk(sk);
1001 unsigned char *ptr = (skb_transport_header(ack_skb) +
1002 TCP_SKB_CB(ack_skb)->sacked);
1003 struct tcp_sack_block_wire *sp = (struct tcp_sack_block_wire *)(ptr+2);
1004 struct sk_buff *cached_skb;
1005 int num_sacks = (ptr[1] - TCPOLEN_SACK_BASE)>>3;
1006 int reord = tp->packets_out;
1008 u32 lost_retrans = 0;
1010 int found_dup_sack = 0;
1011 int cached_fack_count;
1013 int first_sack_index;
1015 if (!tp->sacked_out) {
1016 tp->fackets_out = 0;
1017 tp->highest_sack = tp->snd_una;
1019 prior_fackets = tp->fackets_out;
1021 found_dup_sack = tcp_check_dsack(tp, ack_skb, sp,
1022 num_sacks, prior_snd_una);
1024 flag |= FLAG_DSACKING_ACK;
1026 /* Eliminate too old ACKs, but take into
1027 * account more or less fresh ones, they can
1028 * contain valid SACK info.
1030 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1034 * if the only SACK change is the increase of the end_seq of
1035 * the first block then only apply that SACK block
1036 * and use retrans queue hinting otherwise slowpath */
1038 for (i = 0; i < num_sacks; i++) {
1039 __be32 start_seq = sp[i].start_seq;
1040 __be32 end_seq = sp[i].end_seq;
1043 if (tp->recv_sack_cache[i].start_seq != start_seq)
1046 if ((tp->recv_sack_cache[i].start_seq != start_seq) ||
1047 (tp->recv_sack_cache[i].end_seq != end_seq))
1050 tp->recv_sack_cache[i].start_seq = start_seq;
1051 tp->recv_sack_cache[i].end_seq = end_seq;
1053 /* Clear the rest of the cache sack blocks so they won't match mistakenly. */
1054 for (; i < ARRAY_SIZE(tp->recv_sack_cache); i++) {
1055 tp->recv_sack_cache[i].start_seq = 0;
1056 tp->recv_sack_cache[i].end_seq = 0;
1059 first_sack_index = 0;
1064 tp->fastpath_skb_hint = NULL;
1066 /* order SACK blocks to allow in order walk of the retrans queue */
1067 for (i = num_sacks-1; i > 0; i--) {
1068 for (j = 0; j < i; j++){
1069 if (after(ntohl(sp[j].start_seq),
1070 ntohl(sp[j+1].start_seq))){
1071 struct tcp_sack_block_wire tmp;
1077 /* Track where the first SACK block goes to */
1078 if (j == first_sack_index)
1079 first_sack_index = j+1;
1086 /* clear flag as used for different purpose in following code */
1089 /* Use SACK fastpath hint if valid */
1090 cached_skb = tp->fastpath_skb_hint;
1091 cached_fack_count = tp->fastpath_cnt_hint;
1093 cached_skb = tcp_write_queue_head(sk);
1094 cached_fack_count = 0;
1097 for (i=0; i<num_sacks; i++, sp++) {
1098 struct sk_buff *skb;
1099 __u32 start_seq = ntohl(sp->start_seq);
1100 __u32 end_seq = ntohl(sp->end_seq);
1102 int dup_sack = (found_dup_sack && (i == first_sack_index));
1105 fack_count = cached_fack_count;
1107 /* Event "B" in the comment above. */
1108 if (after(end_seq, tp->high_seq))
1109 flag |= FLAG_DATA_LOST;
1111 tcp_for_write_queue_from(skb, sk) {
1112 int in_sack, pcount;
1115 if (skb == tcp_send_head(sk))
1119 cached_fack_count = fack_count;
1120 if (i == first_sack_index) {
1121 tp->fastpath_skb_hint = skb;
1122 tp->fastpath_cnt_hint = fack_count;
1125 /* The retransmission queue is always in order, so
1126 * we can short-circuit the walk early.
1128 if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1131 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1132 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1134 pcount = tcp_skb_pcount(skb);
1136 if (pcount > 1 && !in_sack &&
1137 after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1138 unsigned int pkt_len;
1140 in_sack = !after(start_seq,
1141 TCP_SKB_CB(skb)->seq);
1144 pkt_len = (start_seq -
1145 TCP_SKB_CB(skb)->seq);
1147 pkt_len = (end_seq -
1148 TCP_SKB_CB(skb)->seq);
1149 if (tcp_fragment(sk, skb, pkt_len, skb_shinfo(skb)->gso_size))
1151 pcount = tcp_skb_pcount(skb);
1154 fack_count += pcount;
1156 sacked = TCP_SKB_CB(skb)->sacked;
1158 /* Account D-SACK for retransmitted packet. */
1159 if ((dup_sack && in_sack) &&
1160 (sacked & TCPCB_RETRANS) &&
1161 after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker))
1164 /* The frame is ACKed. */
1165 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) {
1166 if (sacked&TCPCB_RETRANS) {
1167 if ((dup_sack && in_sack) &&
1168 (sacked&TCPCB_SACKED_ACKED))
1169 reord = min(fack_count, reord);
1171 /* If it was in a hole, we detected reordering. */
1172 if (fack_count < prior_fackets &&
1173 !(sacked&TCPCB_SACKED_ACKED))
1174 reord = min(fack_count, reord);
1177 /* Nothing to do; acked frame is about to be dropped. */
1181 if ((sacked&TCPCB_SACKED_RETRANS) &&
1182 after(end_seq, TCP_SKB_CB(skb)->ack_seq) &&
1183 (!lost_retrans || after(end_seq, lost_retrans)))
1184 lost_retrans = end_seq;
1189 if (!(sacked&TCPCB_SACKED_ACKED)) {
1190 if (sacked & TCPCB_SACKED_RETRANS) {
1191 /* If the segment is not tagged as lost,
1192 * we do not clear RETRANS, believing
1193 * that retransmission is still in flight.
1195 if (sacked & TCPCB_LOST) {
1196 TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1197 tp->lost_out -= tcp_skb_pcount(skb);
1198 tp->retrans_out -= tcp_skb_pcount(skb);
1200 /* clear lost hint */
1201 tp->retransmit_skb_hint = NULL;
1204 /* New sack for not retransmitted frame,
1205 * which was in hole. It is reordering.
1207 if (!(sacked & TCPCB_RETRANS) &&
1208 fack_count < prior_fackets)
1209 reord = min(fack_count, reord);
1211 if (sacked & TCPCB_LOST) {
1212 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1213 tp->lost_out -= tcp_skb_pcount(skb);
1215 /* clear lost hint */
1216 tp->retransmit_skb_hint = NULL;
1218 /* SACK enhanced F-RTO detection.
1219 * Set flag if and only if non-rexmitted
1220 * segments below frto_highmark are
1221 * SACKed (RFC4138; Appendix B).
1222 * Clearing correct due to in-order walk
1224 if (after(end_seq, tp->frto_highmark)) {
1225 flag &= ~FLAG_ONLY_ORIG_SACKED;
1227 if (!(sacked & TCPCB_RETRANS))
1228 flag |= FLAG_ONLY_ORIG_SACKED;
1232 TCP_SKB_CB(skb)->sacked |= TCPCB_SACKED_ACKED;
1233 flag |= FLAG_DATA_SACKED;
1234 tp->sacked_out += tcp_skb_pcount(skb);
1236 if (fack_count > tp->fackets_out)
1237 tp->fackets_out = fack_count;
1239 if (after(TCP_SKB_CB(skb)->seq,
1241 tp->highest_sack = TCP_SKB_CB(skb)->seq;
1243 if (dup_sack && (sacked&TCPCB_RETRANS))
1244 reord = min(fack_count, reord);
1247 /* D-SACK. We can detect redundant retransmission
1248 * in S|R and plain R frames and clear it.
1249 * undo_retrans is decreased above, L|R frames
1250 * are accounted above as well.
1253 (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS)) {
1254 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1255 tp->retrans_out -= tcp_skb_pcount(skb);
1256 tp->retransmit_skb_hint = NULL;
1261 /* Check for lost retransmit. This superb idea is
1262 * borrowed from "ratehalving". Event "C".
1263 * Later note: FACK people cheated me again 8),
1264 * we have to account for reordering! Ugly,
1267 if (lost_retrans && icsk->icsk_ca_state == TCP_CA_Recovery) {
1268 struct sk_buff *skb;
1270 tcp_for_write_queue(skb, sk) {
1271 if (skb == tcp_send_head(sk))
1273 if (after(TCP_SKB_CB(skb)->seq, lost_retrans))
1275 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1277 if ((TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) &&
1278 after(lost_retrans, TCP_SKB_CB(skb)->ack_seq) &&
1280 !before(lost_retrans,
1281 TCP_SKB_CB(skb)->ack_seq + tp->reordering *
1283 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1284 tp->retrans_out -= tcp_skb_pcount(skb);
1286 /* clear lost hint */
1287 tp->retransmit_skb_hint = NULL;
1289 if (!(TCP_SKB_CB(skb)->sacked&(TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1290 tp->lost_out += tcp_skb_pcount(skb);
1291 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1292 flag |= FLAG_DATA_SACKED;
1293 NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT);
1299 tp->left_out = tp->sacked_out + tp->lost_out;
1301 if ((reord < tp->fackets_out) && icsk->icsk_ca_state != TCP_CA_Loss &&
1302 (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark)))
1303 tcp_update_reordering(sk, ((tp->fackets_out + 1) - reord), 0);
1305 #if FASTRETRANS_DEBUG > 0
1306 BUG_TRAP((int)tp->sacked_out >= 0);
1307 BUG_TRAP((int)tp->lost_out >= 0);
1308 BUG_TRAP((int)tp->retrans_out >= 0);
1309 BUG_TRAP((int)tcp_packets_in_flight(tp) >= 0);
1314 /* F-RTO can only be used if TCP has never retransmitted anything other than
1315 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
1317 static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1319 struct tcp_sock *tp = tcp_sk(sk);
1322 holes = max(tp->lost_out, 1U);
1323 holes = min(holes, tp->packets_out);
1325 if ((tp->sacked_out + holes) > tp->packets_out) {
1326 tp->sacked_out = tp->packets_out - holes;
1327 tcp_update_reordering(sk, tp->packets_out + addend, 0);
1331 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1333 static void tcp_add_reno_sack(struct sock *sk)
1335 struct tcp_sock *tp = tcp_sk(sk);
1337 tcp_check_reno_reordering(sk, 0);
1338 tcp_sync_left_out(tp);
1341 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1343 static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1345 struct tcp_sock *tp = tcp_sk(sk);
1348 /* One ACK acked hole. The rest eat duplicate ACKs. */
1349 if (acked-1 >= tp->sacked_out)
1352 tp->sacked_out -= acked-1;
1354 tcp_check_reno_reordering(sk, acked);
1355 tcp_sync_left_out(tp);
1358 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1361 tp->left_out = tp->lost_out;
1364 int tcp_use_frto(struct sock *sk)
1366 const struct tcp_sock *tp = tcp_sk(sk);
1367 struct sk_buff *skb;
1369 if (!sysctl_tcp_frto)
1375 /* Avoid expensive walking of rexmit queue if possible */
1376 if (tp->retrans_out > 1)
1379 skb = tcp_write_queue_head(sk);
1380 skb = tcp_write_queue_next(sk, skb); /* Skips head */
1381 tcp_for_write_queue_from(skb, sk) {
1382 if (skb == tcp_send_head(sk))
1384 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS)
1386 /* Short-circuit when first non-SACKed skb has been checked */
1387 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED))
1393 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
1394 * recovery a bit and use heuristics in tcp_process_frto() to detect if
1395 * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
1396 * keep retrans_out counting accurate (with SACK F-RTO, other than head
1397 * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
1398 * bits are handled if the Loss state is really to be entered (in
1399 * tcp_enter_frto_loss).
1401 * Do like tcp_enter_loss() would; when RTO expires the second time it
1403 * "Reduce ssthresh if it has not yet been made inside this window."
1405 void tcp_enter_frto(struct sock *sk)
1407 const struct inet_connection_sock *icsk = inet_csk(sk);
1408 struct tcp_sock *tp = tcp_sk(sk);
1409 struct sk_buff *skb;
1411 if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) ||
1412 tp->snd_una == tp->high_seq ||
1413 ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) &&
1414 !icsk->icsk_retransmits)) {
1415 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1416 /* Our state is too optimistic in ssthresh() call because cwnd
1417 * is not reduced until tcp_enter_frto_loss() when previous FRTO
1418 * recovery has not yet completed. Pattern would be this: RTO,
1419 * Cumulative ACK, RTO (2xRTO for the same segment does not end
1421 * RFC4138 should be more specific on what to do, even though
1422 * RTO is quite unlikely to occur after the first Cumulative ACK
1423 * due to back-off and complexity of triggering events ...
1425 if (tp->frto_counter) {
1427 stored_cwnd = tp->snd_cwnd;
1429 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1430 tp->snd_cwnd = stored_cwnd;
1432 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1434 /* ... in theory, cong.control module could do "any tricks" in
1435 * ssthresh(), which means that ca_state, lost bits and lost_out
1436 * counter would have to be faked before the call occurs. We
1437 * consider that too expensive, unlikely and hacky, so modules
1438 * using these in ssthresh() must deal these incompatibility
1439 * issues if they receives CA_EVENT_FRTO and frto_counter != 0
1441 tcp_ca_event(sk, CA_EVENT_FRTO);
1444 tp->undo_marker = tp->snd_una;
1445 tp->undo_retrans = 0;
1447 skb = tcp_write_queue_head(sk);
1448 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
1449 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1450 tp->retrans_out -= tcp_skb_pcount(skb);
1452 tcp_sync_left_out(tp);
1454 /* Earlier loss recovery underway (see RFC4138; Appendix B).
1455 * The last condition is necessary at least in tp->frto_counter case.
1457 if (IsSackFrto() && (tp->frto_counter ||
1458 ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) &&
1459 after(tp->high_seq, tp->snd_una)) {
1460 tp->frto_highmark = tp->high_seq;
1462 tp->frto_highmark = tp->snd_nxt;
1464 tcp_set_ca_state(sk, TCP_CA_Disorder);
1465 tp->high_seq = tp->snd_nxt;
1466 tp->frto_counter = 1;
1469 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1470 * which indicates that we should follow the traditional RTO recovery,
1471 * i.e. mark everything lost and do go-back-N retransmission.
1473 static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag)
1475 struct tcp_sock *tp = tcp_sk(sk);
1476 struct sk_buff *skb;
1479 tp->retrans_out = 0;
1481 tcp_reset_reno_sack(tp);
1483 tcp_for_write_queue(skb, sk) {
1484 if (skb == tcp_send_head(sk))
1487 * Count the retransmission made on RTO correctly (only when
1488 * waiting for the first ACK and did not get it)...
1490 if ((tp->frto_counter == 1) && !(flag&FLAG_DATA_ACKED)) {
1491 /* For some reason this R-bit might get cleared? */
1492 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1493 tp->retrans_out += tcp_skb_pcount(skb);
1494 /* ...enter this if branch just for the first segment */
1495 flag |= FLAG_DATA_ACKED;
1497 TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1500 /* Don't lost mark skbs that were fwd transmitted after RTO */
1501 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) &&
1502 !after(TCP_SKB_CB(skb)->end_seq, tp->frto_highmark)) {
1503 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1504 tp->lost_out += tcp_skb_pcount(skb);
1507 tcp_sync_left_out(tp);
1509 tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments;
1510 tp->snd_cwnd_cnt = 0;
1511 tp->snd_cwnd_stamp = tcp_time_stamp;
1512 tp->undo_marker = 0;
1513 tp->frto_counter = 0;
1515 tp->reordering = min_t(unsigned int, tp->reordering,
1516 sysctl_tcp_reordering);
1517 tcp_set_ca_state(sk, TCP_CA_Loss);
1518 tp->high_seq = tp->frto_highmark;
1519 TCP_ECN_queue_cwr(tp);
1521 clear_all_retrans_hints(tp);
1524 void tcp_clear_retrans(struct tcp_sock *tp)
1527 tp->retrans_out = 0;
1529 tp->fackets_out = 0;
1533 tp->undo_marker = 0;
1534 tp->undo_retrans = 0;
1537 /* Enter Loss state. If "how" is not zero, forget all SACK information
1538 * and reset tags completely, otherwise preserve SACKs. If receiver
1539 * dropped its ofo queue, we will know this due to reneging detection.
1541 void tcp_enter_loss(struct sock *sk, int how)
1543 const struct inet_connection_sock *icsk = inet_csk(sk);
1544 struct tcp_sock *tp = tcp_sk(sk);
1545 struct sk_buff *skb;
1548 /* Reduce ssthresh if it has not yet been made inside this window. */
1549 if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
1550 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1551 tp->prior_ssthresh = tcp_current_ssthresh(sk);
1552 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1553 tcp_ca_event(sk, CA_EVENT_LOSS);
1556 tp->snd_cwnd_cnt = 0;
1557 tp->snd_cwnd_stamp = tcp_time_stamp;
1559 tp->bytes_acked = 0;
1560 tcp_clear_retrans(tp);
1562 /* Push undo marker, if it was plain RTO and nothing
1563 * was retransmitted. */
1565 tp->undo_marker = tp->snd_una;
1567 tcp_for_write_queue(skb, sk) {
1568 if (skb == tcp_send_head(sk))
1570 cnt += tcp_skb_pcount(skb);
1571 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS)
1572 tp->undo_marker = 0;
1573 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1574 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
1575 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1576 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1577 tp->lost_out += tcp_skb_pcount(skb);
1579 tp->sacked_out += tcp_skb_pcount(skb);
1580 tp->fackets_out = cnt;
1583 tcp_sync_left_out(tp);
1585 tp->reordering = min_t(unsigned int, tp->reordering,
1586 sysctl_tcp_reordering);
1587 tcp_set_ca_state(sk, TCP_CA_Loss);
1588 tp->high_seq = tp->snd_nxt;
1589 TCP_ECN_queue_cwr(tp);
1590 /* Abort FRTO algorithm if one is in progress */
1591 tp->frto_counter = 0;
1593 clear_all_retrans_hints(tp);
1596 static int tcp_check_sack_reneging(struct sock *sk)
1598 struct sk_buff *skb;
1600 /* If ACK arrived pointing to a remembered SACK,
1601 * it means that our remembered SACKs do not reflect
1602 * real state of receiver i.e.
1603 * receiver _host_ is heavily congested (or buggy).
1604 * Do processing similar to RTO timeout.
1606 if ((skb = tcp_write_queue_head(sk)) != NULL &&
1607 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
1608 struct inet_connection_sock *icsk = inet_csk(sk);
1609 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING);
1611 tcp_enter_loss(sk, 1);
1612 icsk->icsk_retransmits++;
1613 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
1614 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1615 icsk->icsk_rto, TCP_RTO_MAX);
1621 static inline int tcp_fackets_out(struct tcp_sock *tp)
1623 return IsReno(tp) ? tp->sacked_out+1 : tp->fackets_out;
1626 static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb)
1628 return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto);
1631 static inline int tcp_head_timedout(struct sock *sk)
1633 struct tcp_sock *tp = tcp_sk(sk);
1635 return tp->packets_out &&
1636 tcp_skb_timedout(sk, tcp_write_queue_head(sk));
1639 /* Linux NewReno/SACK/FACK/ECN state machine.
1640 * --------------------------------------
1642 * "Open" Normal state, no dubious events, fast path.
1643 * "Disorder" In all the respects it is "Open",
1644 * but requires a bit more attention. It is entered when
1645 * we see some SACKs or dupacks. It is split of "Open"
1646 * mainly to move some processing from fast path to slow one.
1647 * "CWR" CWND was reduced due to some Congestion Notification event.
1648 * It can be ECN, ICMP source quench, local device congestion.
1649 * "Recovery" CWND was reduced, we are fast-retransmitting.
1650 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
1652 * tcp_fastretrans_alert() is entered:
1653 * - each incoming ACK, if state is not "Open"
1654 * - when arrived ACK is unusual, namely:
1659 * Counting packets in flight is pretty simple.
1661 * in_flight = packets_out - left_out + retrans_out
1663 * packets_out is SND.NXT-SND.UNA counted in packets.
1665 * retrans_out is number of retransmitted segments.
1667 * left_out is number of segments left network, but not ACKed yet.
1669 * left_out = sacked_out + lost_out
1671 * sacked_out: Packets, which arrived to receiver out of order
1672 * and hence not ACKed. With SACKs this number is simply
1673 * amount of SACKed data. Even without SACKs
1674 * it is easy to give pretty reliable estimate of this number,
1675 * counting duplicate ACKs.
1677 * lost_out: Packets lost by network. TCP has no explicit
1678 * "loss notification" feedback from network (for now).
1679 * It means that this number can be only _guessed_.
1680 * Actually, it is the heuristics to predict lossage that
1681 * distinguishes different algorithms.
1683 * F.e. after RTO, when all the queue is considered as lost,
1684 * lost_out = packets_out and in_flight = retrans_out.
1686 * Essentially, we have now two algorithms counting
1689 * FACK: It is the simplest heuristics. As soon as we decided
1690 * that something is lost, we decide that _all_ not SACKed
1691 * packets until the most forward SACK are lost. I.e.
1692 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
1693 * It is absolutely correct estimate, if network does not reorder
1694 * packets. And it loses any connection to reality when reordering
1695 * takes place. We use FACK by default until reordering
1696 * is suspected on the path to this destination.
1698 * NewReno: when Recovery is entered, we assume that one segment
1699 * is lost (classic Reno). While we are in Recovery and
1700 * a partial ACK arrives, we assume that one more packet
1701 * is lost (NewReno). This heuristics are the same in NewReno
1704 * Imagine, that's all! Forget about all this shamanism about CWND inflation
1705 * deflation etc. CWND is real congestion window, never inflated, changes
1706 * only according to classic VJ rules.
1708 * Really tricky (and requiring careful tuning) part of algorithm
1709 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
1710 * The first determines the moment _when_ we should reduce CWND and,
1711 * hence, slow down forward transmission. In fact, it determines the moment
1712 * when we decide that hole is caused by loss, rather than by a reorder.
1714 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
1715 * holes, caused by lost packets.
1717 * And the most logically complicated part of algorithm is undo
1718 * heuristics. We detect false retransmits due to both too early
1719 * fast retransmit (reordering) and underestimated RTO, analyzing
1720 * timestamps and D-SACKs. When we detect that some segments were
1721 * retransmitted by mistake and CWND reduction was wrong, we undo
1722 * window reduction and abort recovery phase. This logic is hidden
1723 * inside several functions named tcp_try_undo_<something>.
1726 /* This function decides, when we should leave Disordered state
1727 * and enter Recovery phase, reducing congestion window.
1729 * Main question: may we further continue forward transmission
1730 * with the same cwnd?
1732 static int tcp_time_to_recover(struct sock *sk)
1734 struct tcp_sock *tp = tcp_sk(sk);
1737 /* Do not perform any recovery during FRTO algorithm */
1738 if (tp->frto_counter)
1741 /* Trick#1: The loss is proven. */
1745 /* Not-A-Trick#2 : Classic rule... */
1746 if (tcp_fackets_out(tp) > tp->reordering)
1749 /* Trick#3 : when we use RFC2988 timer restart, fast
1750 * retransmit can be triggered by timeout of queue head.
1752 if (tcp_head_timedout(sk))
1755 /* Trick#4: It is still not OK... But will it be useful to delay
1758 packets_out = tp->packets_out;
1759 if (packets_out <= tp->reordering &&
1760 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
1761 !tcp_may_send_now(sk)) {
1762 /* We have nothing to send. This connection is limited
1763 * either by receiver window or by application.
1771 /* RFC: This is from the original, I doubt that this is necessary at all:
1772 * clear xmit_retrans hint if seq of this skb is beyond hint. How could we
1773 * retransmitted past LOST markings in the first place? I'm not fully sure
1774 * about undo and end of connection cases, which can cause R without L?
1776 static void tcp_verify_retransmit_hint(struct tcp_sock *tp,
1777 struct sk_buff *skb)
1779 if ((tp->retransmit_skb_hint != NULL) &&
1780 before(TCP_SKB_CB(skb)->seq,
1781 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
1782 tp->retransmit_skb_hint = NULL;
1785 /* Mark head of queue up as lost. */
1786 static void tcp_mark_head_lost(struct sock *sk,
1787 int packets, u32 high_seq)
1789 struct tcp_sock *tp = tcp_sk(sk);
1790 struct sk_buff *skb;
1793 BUG_TRAP(packets <= tp->packets_out);
1794 if (tp->lost_skb_hint) {
1795 skb = tp->lost_skb_hint;
1796 cnt = tp->lost_cnt_hint;
1798 skb = tcp_write_queue_head(sk);
1802 tcp_for_write_queue_from(skb, sk) {
1803 if (skb == tcp_send_head(sk))
1805 /* TODO: do this better */
1806 /* this is not the most efficient way to do this... */
1807 tp->lost_skb_hint = skb;
1808 tp->lost_cnt_hint = cnt;
1809 cnt += tcp_skb_pcount(skb);
1810 if (cnt > packets || after(TCP_SKB_CB(skb)->end_seq, high_seq))
1812 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
1813 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1814 tp->lost_out += tcp_skb_pcount(skb);
1815 tcp_verify_retransmit_hint(tp, skb);
1818 tcp_sync_left_out(tp);
1821 /* Account newly detected lost packet(s) */
1823 static void tcp_update_scoreboard(struct sock *sk)
1825 struct tcp_sock *tp = tcp_sk(sk);
1828 int lost = tp->fackets_out - tp->reordering;
1831 tcp_mark_head_lost(sk, lost, tp->high_seq);
1833 tcp_mark_head_lost(sk, 1, tp->high_seq);
1836 /* New heuristics: it is possible only after we switched
1837 * to restart timer each time when something is ACKed.
1838 * Hence, we can detect timed out packets during fast
1839 * retransmit without falling to slow start.
1841 if (!IsReno(tp) && tcp_head_timedout(sk)) {
1842 struct sk_buff *skb;
1844 skb = tp->scoreboard_skb_hint ? tp->scoreboard_skb_hint
1845 : tcp_write_queue_head(sk);
1847 tcp_for_write_queue_from(skb, sk) {
1848 if (skb == tcp_send_head(sk))
1850 if (!tcp_skb_timedout(sk, skb))
1853 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
1854 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1855 tp->lost_out += tcp_skb_pcount(skb);
1856 tcp_verify_retransmit_hint(tp, skb);
1860 tp->scoreboard_skb_hint = skb;
1862 tcp_sync_left_out(tp);
1866 /* CWND moderation, preventing bursts due to too big ACKs
1867 * in dubious situations.
1869 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
1871 tp->snd_cwnd = min(tp->snd_cwnd,
1872 tcp_packets_in_flight(tp)+tcp_max_burst(tp));
1873 tp->snd_cwnd_stamp = tcp_time_stamp;
1876 /* Lower bound on congestion window is slow start threshold
1877 * unless congestion avoidance choice decides to overide it.
1879 static inline u32 tcp_cwnd_min(const struct sock *sk)
1881 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1883 return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh;
1886 /* Decrease cwnd each second ack. */
1887 static void tcp_cwnd_down(struct sock *sk, int flag)
1889 struct tcp_sock *tp = tcp_sk(sk);
1890 int decr = tp->snd_cwnd_cnt + 1;
1892 if ((flag&(FLAG_ANY_PROGRESS|FLAG_DSACKING_ACK)) ||
1893 (IsReno(tp) && !(flag&FLAG_NOT_DUP))) {
1894 tp->snd_cwnd_cnt = decr&1;
1897 if (decr && tp->snd_cwnd > tcp_cwnd_min(sk))
1898 tp->snd_cwnd -= decr;
1900 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp)+1);
1901 tp->snd_cwnd_stamp = tcp_time_stamp;
1905 /* Nothing was retransmitted or returned timestamp is less
1906 * than timestamp of the first retransmission.
1908 static inline int tcp_packet_delayed(struct tcp_sock *tp)
1910 return !tp->retrans_stamp ||
1911 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
1912 (__s32)(tp->rx_opt.rcv_tsecr - tp->retrans_stamp) < 0);
1915 /* Undo procedures. */
1917 #if FASTRETRANS_DEBUG > 1
1918 static void DBGUNDO(struct sock *sk, const char *msg)
1920 struct tcp_sock *tp = tcp_sk(sk);
1921 struct inet_sock *inet = inet_sk(sk);
1923 printk(KERN_DEBUG "Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n",
1925 NIPQUAD(inet->daddr), ntohs(inet->dport),
1926 tp->snd_cwnd, tp->left_out,
1927 tp->snd_ssthresh, tp->prior_ssthresh,
1931 #define DBGUNDO(x...) do { } while (0)
1934 static void tcp_undo_cwr(struct sock *sk, const int undo)
1936 struct tcp_sock *tp = tcp_sk(sk);
1938 if (tp->prior_ssthresh) {
1939 const struct inet_connection_sock *icsk = inet_csk(sk);
1941 if (icsk->icsk_ca_ops->undo_cwnd)
1942 tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
1944 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh<<1);
1946 if (undo && tp->prior_ssthresh > tp->snd_ssthresh) {
1947 tp->snd_ssthresh = tp->prior_ssthresh;
1948 TCP_ECN_withdraw_cwr(tp);
1951 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
1953 tcp_moderate_cwnd(tp);
1954 tp->snd_cwnd_stamp = tcp_time_stamp;
1956 /* There is something screwy going on with the retrans hints after
1958 clear_all_retrans_hints(tp);
1961 static inline int tcp_may_undo(struct tcp_sock *tp)
1963 return tp->undo_marker &&
1964 (!tp->undo_retrans || tcp_packet_delayed(tp));
1967 /* People celebrate: "We love our President!" */
1968 static int tcp_try_undo_recovery(struct sock *sk)
1970 struct tcp_sock *tp = tcp_sk(sk);
1972 if (tcp_may_undo(tp)) {
1973 /* Happy end! We did not retransmit anything
1974 * or our original transmission succeeded.
1976 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
1977 tcp_undo_cwr(sk, 1);
1978 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
1979 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
1981 NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO);
1982 tp->undo_marker = 0;
1984 if (tp->snd_una == tp->high_seq && IsReno(tp)) {
1985 /* Hold old state until something *above* high_seq
1986 * is ACKed. For Reno it is MUST to prevent false
1987 * fast retransmits (RFC2582). SACK TCP is safe. */
1988 tcp_moderate_cwnd(tp);
1991 tcp_set_ca_state(sk, TCP_CA_Open);
1995 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
1996 static void tcp_try_undo_dsack(struct sock *sk)
1998 struct tcp_sock *tp = tcp_sk(sk);
2000 if (tp->undo_marker && !tp->undo_retrans) {
2001 DBGUNDO(sk, "D-SACK");
2002 tcp_undo_cwr(sk, 1);
2003 tp->undo_marker = 0;
2004 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO);
2008 /* Undo during fast recovery after partial ACK. */
2010 static int tcp_try_undo_partial(struct sock *sk, int acked)
2012 struct tcp_sock *tp = tcp_sk(sk);
2013 /* Partial ACK arrived. Force Hoe's retransmit. */
2014 int failed = IsReno(tp) || tp->fackets_out>tp->reordering;
2016 if (tcp_may_undo(tp)) {
2017 /* Plain luck! Hole if filled with delayed
2018 * packet, rather than with a retransmit.
2020 if (tp->retrans_out == 0)
2021 tp->retrans_stamp = 0;
2023 tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2026 tcp_undo_cwr(sk, 0);
2027 NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO);
2029 /* So... Do not make Hoe's retransmit yet.
2030 * If the first packet was delayed, the rest
2031 * ones are most probably delayed as well.
2038 /* Undo during loss recovery after partial ACK. */
2039 static int tcp_try_undo_loss(struct sock *sk)
2041 struct tcp_sock *tp = tcp_sk(sk);
2043 if (tcp_may_undo(tp)) {
2044 struct sk_buff *skb;
2045 tcp_for_write_queue(skb, sk) {
2046 if (skb == tcp_send_head(sk))
2048 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2051 clear_all_retrans_hints(tp);
2053 DBGUNDO(sk, "partial loss");
2055 tp->left_out = tp->sacked_out;
2056 tcp_undo_cwr(sk, 1);
2057 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
2058 inet_csk(sk)->icsk_retransmits = 0;
2059 tp->undo_marker = 0;
2061 tcp_set_ca_state(sk, TCP_CA_Open);
2067 static inline void tcp_complete_cwr(struct sock *sk)
2069 struct tcp_sock *tp = tcp_sk(sk);
2070 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2071 tp->snd_cwnd_stamp = tcp_time_stamp;
2072 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2075 static void tcp_try_to_open(struct sock *sk, int flag)
2077 struct tcp_sock *tp = tcp_sk(sk);
2079 tcp_sync_left_out(tp);
2081 if (tp->retrans_out == 0)
2082 tp->retrans_stamp = 0;
2085 tcp_enter_cwr(sk, 1);
2087 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2088 int state = TCP_CA_Open;
2090 if (tp->left_out || tp->retrans_out || tp->undo_marker)
2091 state = TCP_CA_Disorder;
2093 if (inet_csk(sk)->icsk_ca_state != state) {
2094 tcp_set_ca_state(sk, state);
2095 tp->high_seq = tp->snd_nxt;
2097 tcp_moderate_cwnd(tp);
2099 tcp_cwnd_down(sk, flag);
2103 static void tcp_mtup_probe_failed(struct sock *sk)
2105 struct inet_connection_sock *icsk = inet_csk(sk);
2107 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2108 icsk->icsk_mtup.probe_size = 0;
2111 static void tcp_mtup_probe_success(struct sock *sk, struct sk_buff *skb)
2113 struct tcp_sock *tp = tcp_sk(sk);
2114 struct inet_connection_sock *icsk = inet_csk(sk);
2116 /* FIXME: breaks with very large cwnd */
2117 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2118 tp->snd_cwnd = tp->snd_cwnd *
2119 tcp_mss_to_mtu(sk, tp->mss_cache) /
2120 icsk->icsk_mtup.probe_size;
2121 tp->snd_cwnd_cnt = 0;
2122 tp->snd_cwnd_stamp = tcp_time_stamp;
2123 tp->rcv_ssthresh = tcp_current_ssthresh(sk);
2125 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2126 icsk->icsk_mtup.probe_size = 0;
2127 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
2131 /* Process an event, which can update packets-in-flight not trivially.
2132 * Main goal of this function is to calculate new estimate for left_out,
2133 * taking into account both packets sitting in receiver's buffer and
2134 * packets lost by network.
2136 * Besides that it does CWND reduction, when packet loss is detected
2137 * and changes state of machine.
2139 * It does _not_ decide what to send, it is made in function
2140 * tcp_xmit_retransmit_queue().
2143 tcp_fastretrans_alert(struct sock *sk, int prior_packets, int flag)
2145 struct inet_connection_sock *icsk = inet_csk(sk);
2146 struct tcp_sock *tp = tcp_sk(sk);
2147 int is_dupack = !(flag&(FLAG_SND_UNA_ADVANCED|FLAG_NOT_DUP));
2148 int do_lost = is_dupack || ((flag&FLAG_DATA_SACKED) &&
2149 (tp->fackets_out > tp->reordering));
2151 /* Some technical things:
2152 * 1. Reno does not count dupacks (sacked_out) automatically. */
2153 if (!tp->packets_out)
2155 /* 2. SACK counts snd_fack in packets inaccurately. */
2156 if (tp->sacked_out == 0)
2157 tp->fackets_out = 0;
2159 /* Now state machine starts.
2160 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2162 tp->prior_ssthresh = 0;
2164 /* B. In all the states check for reneging SACKs. */
2165 if (tp->sacked_out && tcp_check_sack_reneging(sk))
2168 /* C. Process data loss notification, provided it is valid. */
2169 if ((flag&FLAG_DATA_LOST) &&
2170 before(tp->snd_una, tp->high_seq) &&
2171 icsk->icsk_ca_state != TCP_CA_Open &&
2172 tp->fackets_out > tp->reordering) {
2173 tcp_mark_head_lost(sk, tp->fackets_out-tp->reordering, tp->high_seq);
2174 NET_INC_STATS_BH(LINUX_MIB_TCPLOSS);
2177 /* D. Synchronize left_out to current state. */
2178 tcp_sync_left_out(tp);
2180 /* E. Check state exit conditions. State can be terminated
2181 * when high_seq is ACKed. */
2182 if (icsk->icsk_ca_state == TCP_CA_Open) {
2183 BUG_TRAP(tp->retrans_out == 0);
2184 tp->retrans_stamp = 0;
2185 } else if (!before(tp->snd_una, tp->high_seq)) {
2186 switch (icsk->icsk_ca_state) {
2188 icsk->icsk_retransmits = 0;
2189 if (tcp_try_undo_recovery(sk))
2194 /* CWR is to be held something *above* high_seq
2195 * is ACKed for CWR bit to reach receiver. */
2196 if (tp->snd_una != tp->high_seq) {
2197 tcp_complete_cwr(sk);
2198 tcp_set_ca_state(sk, TCP_CA_Open);
2202 case TCP_CA_Disorder:
2203 tcp_try_undo_dsack(sk);
2204 if (!tp->undo_marker ||
2205 /* For SACK case do not Open to allow to undo
2206 * catching for all duplicate ACKs. */
2207 IsReno(tp) || tp->snd_una != tp->high_seq) {
2208 tp->undo_marker = 0;
2209 tcp_set_ca_state(sk, TCP_CA_Open);
2213 case TCP_CA_Recovery:
2215 tcp_reset_reno_sack(tp);
2216 if (tcp_try_undo_recovery(sk))
2218 tcp_complete_cwr(sk);
2223 /* F. Process state. */
2224 switch (icsk->icsk_ca_state) {
2225 case TCP_CA_Recovery:
2226 if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2227 if (IsReno(tp) && is_dupack)
2228 tcp_add_reno_sack(sk);
2230 int acked = prior_packets - tp->packets_out;
2232 tcp_remove_reno_sacks(sk, acked);
2233 do_lost = tcp_try_undo_partial(sk, acked);
2237 if (flag&FLAG_DATA_ACKED)
2238 icsk->icsk_retransmits = 0;
2239 if (!tcp_try_undo_loss(sk)) {
2240 tcp_moderate_cwnd(tp);
2241 tcp_xmit_retransmit_queue(sk);
2244 if (icsk->icsk_ca_state != TCP_CA_Open)
2246 /* Loss is undone; fall through to processing in Open state. */
2249 if (flag & FLAG_SND_UNA_ADVANCED)
2250 tcp_reset_reno_sack(tp);
2252 tcp_add_reno_sack(sk);
2255 if (icsk->icsk_ca_state == TCP_CA_Disorder)
2256 tcp_try_undo_dsack(sk);
2258 if (!tcp_time_to_recover(sk)) {
2259 tcp_try_to_open(sk, flag);
2263 /* MTU probe failure: don't reduce cwnd */
2264 if (icsk->icsk_ca_state < TCP_CA_CWR &&
2265 icsk->icsk_mtup.probe_size &&
2266 tp->snd_una == tp->mtu_probe.probe_seq_start) {
2267 tcp_mtup_probe_failed(sk);
2268 /* Restores the reduction we did in tcp_mtup_probe() */
2270 tcp_simple_retransmit(sk);
2274 /* Otherwise enter Recovery state */
2277 NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY);
2279 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY);
2281 tp->high_seq = tp->snd_nxt;
2282 tp->prior_ssthresh = 0;
2283 tp->undo_marker = tp->snd_una;
2284 tp->undo_retrans = tp->retrans_out;
2286 if (icsk->icsk_ca_state < TCP_CA_CWR) {
2287 if (!(flag&FLAG_ECE))
2288 tp->prior_ssthresh = tcp_current_ssthresh(sk);
2289 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
2290 TCP_ECN_queue_cwr(tp);
2293 tp->bytes_acked = 0;
2294 tp->snd_cwnd_cnt = 0;
2295 tcp_set_ca_state(sk, TCP_CA_Recovery);
2298 if (do_lost || tcp_head_timedout(sk))
2299 tcp_update_scoreboard(sk);
2300 tcp_cwnd_down(sk, flag);
2301 tcp_xmit_retransmit_queue(sk);
2304 /* Read draft-ietf-tcplw-high-performance before mucking
2305 * with this code. (Supersedes RFC1323)
2307 static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
2309 /* RTTM Rule: A TSecr value received in a segment is used to
2310 * update the averaged RTT measurement only if the segment
2311 * acknowledges some new data, i.e., only if it advances the
2312 * left edge of the send window.
2314 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2315 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
2317 * Changed: reset backoff as soon as we see the first valid sample.
2318 * If we do not, we get strongly overestimated rto. With timestamps
2319 * samples are accepted even from very old segments: f.e., when rtt=1
2320 * increases to 8, we retransmit 5 times and after 8 seconds delayed
2321 * answer arrives rto becomes 120 seconds! If at least one of segments
2322 * in window is lost... Voila. --ANK (010210)
2324 struct tcp_sock *tp = tcp_sk(sk);
2325 const __u32 seq_rtt = tcp_time_stamp - tp->rx_opt.rcv_tsecr;
2326 tcp_rtt_estimator(sk, seq_rtt);
2328 inet_csk(sk)->icsk_backoff = 0;
2332 static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
2334 /* We don't have a timestamp. Can only use
2335 * packets that are not retransmitted to determine
2336 * rtt estimates. Also, we must not reset the
2337 * backoff for rto until we get a non-retransmitted
2338 * packet. This allows us to deal with a situation
2339 * where the network delay has increased suddenly.
2340 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
2343 if (flag & FLAG_RETRANS_DATA_ACKED)
2346 tcp_rtt_estimator(sk, seq_rtt);
2348 inet_csk(sk)->icsk_backoff = 0;
2352 static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
2355 const struct tcp_sock *tp = tcp_sk(sk);
2356 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
2357 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
2358 tcp_ack_saw_tstamp(sk, flag);
2359 else if (seq_rtt >= 0)
2360 tcp_ack_no_tstamp(sk, seq_rtt, flag);
2363 static void tcp_cong_avoid(struct sock *sk, u32 ack,
2364 u32 in_flight, int good)
2366 const struct inet_connection_sock *icsk = inet_csk(sk);
2367 icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight, good);
2368 tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
2371 /* Restart timer after forward progress on connection.
2372 * RFC2988 recommends to restart timer to now+rto.
2375 static void tcp_ack_packets_out(struct sock *sk)
2377 struct tcp_sock *tp = tcp_sk(sk);
2379 if (!tp->packets_out) {
2380 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
2382 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2386 static int tcp_tso_acked(struct sock *sk, struct sk_buff *skb,
2387 __u32 now, __s32 *seq_rtt)
2389 struct tcp_sock *tp = tcp_sk(sk);
2390 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
2391 __u32 seq = tp->snd_una;
2392 __u32 packets_acked;
2395 /* If we get here, the whole TSO packet has not been
2398 BUG_ON(!after(scb->end_seq, seq));
2400 packets_acked = tcp_skb_pcount(skb);
2401 if (tcp_trim_head(sk, skb, seq - scb->seq))
2403 packets_acked -= tcp_skb_pcount(skb);
2405 if (packets_acked) {
2406 __u8 sacked = scb->sacked;
2408 acked |= FLAG_DATA_ACKED;
2410 if (sacked & TCPCB_RETRANS) {
2411 if (sacked & TCPCB_SACKED_RETRANS)
2412 tp->retrans_out -= packets_acked;
2413 acked |= FLAG_RETRANS_DATA_ACKED;
2415 } else if (*seq_rtt < 0)
2416 *seq_rtt = now - scb->when;
2417 if (sacked & TCPCB_SACKED_ACKED)
2418 tp->sacked_out -= packets_acked;
2419 if (sacked & TCPCB_LOST)
2420 tp->lost_out -= packets_acked;
2421 if (sacked & TCPCB_URG) {
2423 !before(seq, tp->snd_up))
2426 } else if (*seq_rtt < 0)
2427 *seq_rtt = now - scb->when;
2429 if (tp->fackets_out) {
2430 __u32 dval = min(tp->fackets_out, packets_acked);
2431 tp->fackets_out -= dval;
2433 /* hint's skb might be NULL but we don't need to care */
2434 tp->fastpath_cnt_hint -= min_t(u32, packets_acked,
2435 tp->fastpath_cnt_hint);
2436 tp->packets_out -= packets_acked;
2438 BUG_ON(tcp_skb_pcount(skb) == 0);
2439 BUG_ON(!before(scb->seq, scb->end_seq));
2445 /* Remove acknowledged frames from the retransmission queue. */
2446 static int tcp_clean_rtx_queue(struct sock *sk, __s32 *seq_rtt_p)
2448 struct tcp_sock *tp = tcp_sk(sk);
2449 const struct inet_connection_sock *icsk = inet_csk(sk);
2450 struct sk_buff *skb;
2451 __u32 now = tcp_time_stamp;
2453 int prior_packets = tp->packets_out;
2455 ktime_t last_ackt = net_invalid_timestamp();
2457 while ((skb = tcp_write_queue_head(sk)) &&
2458 skb != tcp_send_head(sk)) {
2459 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
2460 __u8 sacked = scb->sacked;
2462 /* If our packet is before the ack sequence we can
2463 * discard it as it's confirmed to have arrived at
2466 if (after(scb->end_seq, tp->snd_una)) {
2467 if (tcp_skb_pcount(skb) > 1 &&
2468 after(tp->snd_una, scb->seq))
2469 acked |= tcp_tso_acked(sk, skb,
2474 /* Initial outgoing SYN's get put onto the write_queue
2475 * just like anything else we transmit. It is not
2476 * true data, and if we misinform our callers that
2477 * this ACK acks real data, we will erroneously exit
2478 * connection startup slow start one packet too
2479 * quickly. This is severely frowned upon behavior.
2481 if (!(scb->flags & TCPCB_FLAG_SYN)) {
2482 acked |= FLAG_DATA_ACKED;
2484 acked |= FLAG_SYN_ACKED;
2485 tp->retrans_stamp = 0;
2488 /* MTU probing checks */
2489 if (icsk->icsk_mtup.probe_size) {
2490 if (!after(tp->mtu_probe.probe_seq_end, TCP_SKB_CB(skb)->end_seq)) {
2491 tcp_mtup_probe_success(sk, skb);
2496 if (sacked & TCPCB_RETRANS) {
2497 if (sacked & TCPCB_SACKED_RETRANS)
2498 tp->retrans_out -= tcp_skb_pcount(skb);
2499 acked |= FLAG_RETRANS_DATA_ACKED;
2501 } else if (seq_rtt < 0) {
2502 seq_rtt = now - scb->when;
2503 last_ackt = skb->tstamp;
2505 if (sacked & TCPCB_SACKED_ACKED)
2506 tp->sacked_out -= tcp_skb_pcount(skb);
2507 if (sacked & TCPCB_LOST)
2508 tp->lost_out -= tcp_skb_pcount(skb);
2509 if (sacked & TCPCB_URG) {
2511 !before(scb->end_seq, tp->snd_up))
2514 } else if (seq_rtt < 0) {
2515 seq_rtt = now - scb->when;
2516 last_ackt = skb->tstamp;
2518 tcp_dec_pcount_approx(&tp->fackets_out, skb);
2519 tcp_packets_out_dec(tp, skb);
2520 tcp_unlink_write_queue(skb, sk);
2521 sk_stream_free_skb(sk, skb);
2522 clear_all_retrans_hints(tp);
2525 if (acked&FLAG_ACKED) {
2526 u32 pkts_acked = prior_packets - tp->packets_out;
2527 const struct tcp_congestion_ops *ca_ops
2528 = inet_csk(sk)->icsk_ca_ops;
2530 tcp_ack_update_rtt(sk, acked, seq_rtt);
2531 tcp_ack_packets_out(sk);
2533 if (ca_ops->pkts_acked) {
2536 /* Is the ACK triggering packet unambiguous? */
2537 if (!(acked & FLAG_RETRANS_DATA_ACKED)) {
2538 /* High resolution needed and available? */
2539 if (ca_ops->flags & TCP_CONG_RTT_STAMP &&
2540 !ktime_equal(last_ackt,
2541 net_invalid_timestamp()))
2542 rtt_us = ktime_us_delta(ktime_get_real(),
2544 else if (seq_rtt > 0)
2545 rtt_us = jiffies_to_usecs(seq_rtt);
2548 ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
2552 #if FASTRETRANS_DEBUG > 0
2553 BUG_TRAP((int)tp->sacked_out >= 0);
2554 BUG_TRAP((int)tp->lost_out >= 0);
2555 BUG_TRAP((int)tp->retrans_out >= 0);
2556 if (!tp->packets_out && tp->rx_opt.sack_ok) {
2557 const struct inet_connection_sock *icsk = inet_csk(sk);
2559 printk(KERN_DEBUG "Leak l=%u %d\n",
2560 tp->lost_out, icsk->icsk_ca_state);
2563 if (tp->sacked_out) {
2564 printk(KERN_DEBUG "Leak s=%u %d\n",
2565 tp->sacked_out, icsk->icsk_ca_state);
2568 if (tp->retrans_out) {
2569 printk(KERN_DEBUG "Leak r=%u %d\n",
2570 tp->retrans_out, icsk->icsk_ca_state);
2571 tp->retrans_out = 0;
2575 *seq_rtt_p = seq_rtt;
2579 static void tcp_ack_probe(struct sock *sk)
2581 const struct tcp_sock *tp = tcp_sk(sk);
2582 struct inet_connection_sock *icsk = inet_csk(sk);
2584 /* Was it a usable window open? */
2586 if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq,
2587 tp->snd_una + tp->snd_wnd)) {
2588 icsk->icsk_backoff = 0;
2589 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
2590 /* Socket must be waked up by subsequent tcp_data_snd_check().
2591 * This function is not for random using!
2594 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2595 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2600 static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag)
2602 return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
2603 inet_csk(sk)->icsk_ca_state != TCP_CA_Open);
2606 static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag)
2608 const struct tcp_sock *tp = tcp_sk(sk);
2609 return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
2610 !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
2613 /* Check that window update is acceptable.
2614 * The function assumes that snd_una<=ack<=snd_next.
2616 static inline int tcp_may_update_window(const struct tcp_sock *tp, const u32 ack,
2617 const u32 ack_seq, const u32 nwin)
2619 return (after(ack, tp->snd_una) ||
2620 after(ack_seq, tp->snd_wl1) ||
2621 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd));
2624 /* Update our send window.
2626 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
2627 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
2629 static int tcp_ack_update_window(struct sock *sk, struct sk_buff *skb, u32 ack,
2632 struct tcp_sock *tp = tcp_sk(sk);
2634 u32 nwin = ntohs(tcp_hdr(skb)->window);
2636 if (likely(!tcp_hdr(skb)->syn))
2637 nwin <<= tp->rx_opt.snd_wscale;
2639 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
2640 flag |= FLAG_WIN_UPDATE;
2641 tcp_update_wl(tp, ack, ack_seq);
2643 if (tp->snd_wnd != nwin) {
2646 /* Note, it is the only place, where
2647 * fast path is recovered for sending TCP.
2650 tcp_fast_path_check(sk);
2652 if (nwin > tp->max_window) {
2653 tp->max_window = nwin;
2654 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
2664 /* A very conservative spurious RTO response algorithm: reduce cwnd and
2665 * continue in congestion avoidance.
2667 static void tcp_conservative_spur_to_response(struct tcp_sock *tp)
2669 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2670 tp->snd_cwnd_cnt = 0;
2671 TCP_ECN_queue_cwr(tp);
2672 tcp_moderate_cwnd(tp);
2675 /* A conservative spurious RTO response algorithm: reduce cwnd using
2676 * rate halving and continue in congestion avoidance.
2678 static void tcp_ratehalving_spur_to_response(struct sock *sk)
2680 tcp_enter_cwr(sk, 0);
2683 static void tcp_undo_spur_to_response(struct sock *sk, int flag)
2686 tcp_ratehalving_spur_to_response(sk);
2688 tcp_undo_cwr(sk, 1);
2691 /* F-RTO spurious RTO detection algorithm (RFC4138)
2693 * F-RTO affects during two new ACKs following RTO (well, almost, see inline
2694 * comments). State (ACK number) is kept in frto_counter. When ACK advances
2695 * window (but not to or beyond highest sequence sent before RTO):
2696 * On First ACK, send two new segments out.
2697 * On Second ACK, RTO was likely spurious. Do spurious response (response
2698 * algorithm is not part of the F-RTO detection algorithm
2699 * given in RFC4138 but can be selected separately).
2700 * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
2701 * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
2702 * of Nagle, this is done using frto_counter states 2 and 3, when a new data
2703 * segment of any size sent during F-RTO, state 2 is upgraded to 3.
2705 * Rationale: if the RTO was spurious, new ACKs should arrive from the
2706 * original window even after we transmit two new data segments.
2709 * on first step, wait until first cumulative ACK arrives, then move to
2710 * the second step. In second step, the next ACK decides.
2712 * F-RTO is implemented (mainly) in four functions:
2713 * - tcp_use_frto() is used to determine if TCP is can use F-RTO
2714 * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
2715 * called when tcp_use_frto() showed green light
2716 * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
2717 * - tcp_enter_frto_loss() is called if there is not enough evidence
2718 * to prove that the RTO is indeed spurious. It transfers the control
2719 * from F-RTO to the conventional RTO recovery
2721 static int tcp_process_frto(struct sock *sk, int flag)
2723 struct tcp_sock *tp = tcp_sk(sk);
2725 tcp_sync_left_out(tp);
2727 /* Duplicate the behavior from Loss state (fastretrans_alert) */
2728 if (flag&FLAG_DATA_ACKED)
2729 inet_csk(sk)->icsk_retransmits = 0;
2731 if (!before(tp->snd_una, tp->frto_highmark)) {
2732 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag);
2736 if (!IsSackFrto() || IsReno(tp)) {
2737 /* RFC4138 shortcoming in step 2; should also have case c):
2738 * ACK isn't duplicate nor advances window, e.g., opposite dir
2741 if (!(flag&FLAG_ANY_PROGRESS) && (flag&FLAG_NOT_DUP))
2744 if (!(flag&FLAG_DATA_ACKED)) {
2745 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3),
2750 if (!(flag&FLAG_DATA_ACKED) && (tp->frto_counter == 1)) {
2751 /* Prevent sending of new data. */
2752 tp->snd_cwnd = min(tp->snd_cwnd,
2753 tcp_packets_in_flight(tp));
2757 if ((tp->frto_counter >= 2) &&
2758 (!(flag&FLAG_FORWARD_PROGRESS) ||
2759 ((flag&FLAG_DATA_SACKED) && !(flag&FLAG_ONLY_ORIG_SACKED)))) {
2760 /* RFC4138 shortcoming (see comment above) */
2761 if (!(flag&FLAG_FORWARD_PROGRESS) && (flag&FLAG_NOT_DUP))
2764 tcp_enter_frto_loss(sk, 3, flag);
2769 if (tp->frto_counter == 1) {
2770 /* Sending of the next skb must be allowed or no FRTO */
2771 if (!tcp_send_head(sk) ||
2772 after(TCP_SKB_CB(tcp_send_head(sk))->end_seq,
2773 tp->snd_una + tp->snd_wnd)) {
2774 tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3),
2779 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
2780 tp->frto_counter = 2;
2783 switch (sysctl_tcp_frto_response) {
2785 tcp_undo_spur_to_response(sk, flag);
2788 tcp_conservative_spur_to_response(tp);
2791 tcp_ratehalving_spur_to_response(sk);
2794 tp->frto_counter = 0;
2799 /* This routine deals with incoming acks, but not outgoing ones. */
2800 static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag)
2802 struct inet_connection_sock *icsk = inet_csk(sk);
2803 struct tcp_sock *tp = tcp_sk(sk);
2804 u32 prior_snd_una = tp->snd_una;
2805 u32 ack_seq = TCP_SKB_CB(skb)->seq;
2806 u32 ack = TCP_SKB_CB(skb)->ack_seq;
2807 u32 prior_in_flight;
2812 /* If the ack is newer than sent or older than previous acks
2813 * then we can probably ignore it.
2815 if (after(ack, tp->snd_nxt))
2816 goto uninteresting_ack;
2818 if (before(ack, prior_snd_una))
2821 if (after(ack, prior_snd_una))
2822 flag |= FLAG_SND_UNA_ADVANCED;
2824 if (sysctl_tcp_abc) {
2825 if (icsk->icsk_ca_state < TCP_CA_CWR)
2826 tp->bytes_acked += ack - prior_snd_una;
2827 else if (icsk->icsk_ca_state == TCP_CA_Loss)
2828 /* we assume just one segment left network */
2829 tp->bytes_acked += min(ack - prior_snd_una, tp->mss_cache);
2832 if (!(flag&FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
2833 /* Window is constant, pure forward advance.
2834 * No more checks are required.
2835 * Note, we use the fact that SND.UNA>=SND.WL2.
2837 tcp_update_wl(tp, ack, ack_seq);
2839 flag |= FLAG_WIN_UPDATE;
2841 tcp_ca_event(sk, CA_EVENT_FAST_ACK);
2843 NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS);
2845 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
2848 NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS);
2850 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
2852 if (TCP_SKB_CB(skb)->sacked)
2853 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
2855 if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
2858 tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
2861 /* We passed data and got it acked, remove any soft error
2862 * log. Something worked...
2864 sk->sk_err_soft = 0;
2865 tp->rcv_tstamp = tcp_time_stamp;
2866 prior_packets = tp->packets_out;
2870 prior_in_flight = tcp_packets_in_flight(tp);
2872 /* See if we can take anything off of the retransmit queue. */
2873 flag |= tcp_clean_rtx_queue(sk, &seq_rtt);
2875 if (tp->frto_counter)
2876 frto_cwnd = tcp_process_frto(sk, flag);
2878 if (tcp_ack_is_dubious(sk, flag)) {
2879 /* Advance CWND, if state allows this. */
2880 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd &&
2881 tcp_may_raise_cwnd(sk, flag))
2882 tcp_cong_avoid(sk, ack, prior_in_flight, 0);
2883 tcp_fastretrans_alert(sk, prior_packets, flag);
2885 if ((flag & FLAG_DATA_ACKED) && !frto_cwnd)
2886 tcp_cong_avoid(sk, ack, prior_in_flight, 1);
2889 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag&FLAG_NOT_DUP))
2890 dst_confirm(sk->sk_dst_cache);
2895 icsk->icsk_probes_out = 0;
2897 /* If this ack opens up a zero window, clear backoff. It was
2898 * being used to time the probes, and is probably far higher than
2899 * it needs to be for normal retransmission.
2901 if (tcp_send_head(sk))
2906 if (TCP_SKB_CB(skb)->sacked)
2907 tcp_sacktag_write_queue(sk, skb, prior_snd_una);
2910 SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
2915 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
2916 * But, this can also be called on packets in the established flow when
2917 * the fast version below fails.
2919 void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx, int estab)
2922 struct tcphdr *th = tcp_hdr(skb);
2923 int length=(th->doff*4)-sizeof(struct tcphdr);
2925 ptr = (unsigned char *)(th + 1);
2926 opt_rx->saw_tstamp = 0;
2928 while (length > 0) {
2935 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
2940 if (opsize < 2) /* "silly options" */
2942 if (opsize > length)
2943 return; /* don't parse partial options */
2946 if (opsize==TCPOLEN_MSS && th->syn && !estab) {
2947 u16 in_mss = ntohs(get_unaligned((__be16 *)ptr));
2949 if (opt_rx->user_mss && opt_rx->user_mss < in_mss)
2950 in_mss = opt_rx->user_mss;
2951 opt_rx->mss_clamp = in_mss;
2956 if (opsize==TCPOLEN_WINDOW && th->syn && !estab)
2957 if (sysctl_tcp_window_scaling) {
2958 __u8 snd_wscale = *(__u8 *) ptr;
2959 opt_rx->wscale_ok = 1;
2960 if (snd_wscale > 14) {
2961 if (net_ratelimit())
2962 printk(KERN_INFO "tcp_parse_options: Illegal window "
2963 "scaling value %d >14 received.\n",
2967 opt_rx->snd_wscale = snd_wscale;
2970 case TCPOPT_TIMESTAMP:
2971 if (opsize==TCPOLEN_TIMESTAMP) {
2972 if ((estab && opt_rx->tstamp_ok) ||
2973 (!estab && sysctl_tcp_timestamps)) {
2974 opt_rx->saw_tstamp = 1;
2975 opt_rx->rcv_tsval = ntohl(get_unaligned((__be32 *)ptr));
2976 opt_rx->rcv_tsecr = ntohl(get_unaligned((__be32 *)(ptr+4)));
2980 case TCPOPT_SACK_PERM:
2981 if (opsize==TCPOLEN_SACK_PERM && th->syn && !estab) {
2982 if (sysctl_tcp_sack) {
2983 opt_rx->sack_ok = 1;
2984 tcp_sack_reset(opt_rx);
2990 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
2991 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
2993 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
2996 #ifdef CONFIG_TCP_MD5SIG
2999 * The MD5 Hash has already been
3000 * checked (see tcp_v{4,6}_do_rcv()).
3012 /* Fast parse options. This hopes to only see timestamps.
3013 * If it is wrong it falls back on tcp_parse_options().
3015 static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th,
3016 struct tcp_sock *tp)
3018 if (th->doff == sizeof(struct tcphdr)>>2) {
3019 tp->rx_opt.saw_tstamp = 0;
3021 } else if (tp->rx_opt.tstamp_ok &&
3022 th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) {
3023 __be32 *ptr = (__be32 *)(th + 1);
3024 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3025 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3026 tp->rx_opt.saw_tstamp = 1;
3028 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3030 tp->rx_opt.rcv_tsecr = ntohl(*ptr);
3034 tcp_parse_options(skb, &tp->rx_opt, 1);
3038 static inline void tcp_store_ts_recent(struct tcp_sock *tp)
3040 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3041 tp->rx_opt.ts_recent_stamp = get_seconds();
3044 static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3046 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3047 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3048 * extra check below makes sure this can only happen
3049 * for pure ACK frames. -DaveM
3051 * Not only, also it occurs for expired timestamps.
3054 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) >= 0 ||
3055 get_seconds() >= tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS)
3056 tcp_store_ts_recent(tp);
3060 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3062 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3063 * it can pass through stack. So, the following predicate verifies that
3064 * this segment is not used for anything but congestion avoidance or
3065 * fast retransmit. Moreover, we even are able to eliminate most of such
3066 * second order effects, if we apply some small "replay" window (~RTO)
3067 * to timestamp space.
3069 * All these measures still do not guarantee that we reject wrapped ACKs
3070 * on networks with high bandwidth, when sequence space is recycled fastly,
3071 * but it guarantees that such events will be very rare and do not affect
3072 * connection seriously. This doesn't look nice, but alas, PAWS is really
3075 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3076 * states that events when retransmit arrives after original data are rare.
3077 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3078 * the biggest problem on large power networks even with minor reordering.
3079 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3080 * up to bandwidth of 18Gigabit/sec. 8) ]
3083 static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3085 struct tcp_sock *tp = tcp_sk(sk);
3086 struct tcphdr *th = tcp_hdr(skb);
3087 u32 seq = TCP_SKB_CB(skb)->seq;
3088 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3090 return (/* 1. Pure ACK with correct sequence number. */
3091 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3093 /* 2. ... and duplicate ACK. */
3094 ack == tp->snd_una &&
3096 /* 3. ... and does not update window. */
3097 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3099 /* 4. ... and sits in replay window. */
3100 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
3103 static inline int tcp_paws_discard(const struct sock *sk, const struct sk_buff *skb)
3105 const struct tcp_sock *tp = tcp_sk(sk);
3106 return ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > TCP_PAWS_WINDOW &&
3107 get_seconds() < tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS &&
3108 !tcp_disordered_ack(sk, skb));
3111 /* Check segment sequence number for validity.
3113 * Segment controls are considered valid, if the segment
3114 * fits to the window after truncation to the window. Acceptability
3115 * of data (and SYN, FIN, of course) is checked separately.
3116 * See tcp_data_queue(), for example.
3118 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3119 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3120 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3121 * (borrowed from freebsd)
3124 static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq)
3126 return !before(end_seq, tp->rcv_wup) &&
3127 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3130 /* When we get a reset we do this. */
3131 static void tcp_reset(struct sock *sk)
3133 /* We want the right error as BSD sees it (and indeed as we do). */
3134 switch (sk->sk_state) {
3136 sk->sk_err = ECONNREFUSED;
3138 case TCP_CLOSE_WAIT:
3144 sk->sk_err = ECONNRESET;
3147 if (!sock_flag(sk, SOCK_DEAD))
3148 sk->sk_error_report(sk);
3154 * Process the FIN bit. This now behaves as it is supposed to work
3155 * and the FIN takes effect when it is validly part of sequence
3156 * space. Not before when we get holes.
3158 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3159 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3162 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3163 * close and we go into CLOSING (and later onto TIME-WAIT)
3165 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3167 static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th)
3169 struct tcp_sock *tp = tcp_sk(sk);
3171 inet_csk_schedule_ack(sk);
3173 sk->sk_shutdown |= RCV_SHUTDOWN;
3174 sock_set_flag(sk, SOCK_DONE);
3176 switch (sk->sk_state) {
3178 case TCP_ESTABLISHED:
3179 /* Move to CLOSE_WAIT */
3180 tcp_set_state(sk, TCP_CLOSE_WAIT);
3181 inet_csk(sk)->icsk_ack.pingpong = 1;
3184 case TCP_CLOSE_WAIT:
3186 /* Received a retransmission of the FIN, do
3191 /* RFC793: Remain in the LAST-ACK state. */
3195 /* This case occurs when a simultaneous close
3196 * happens, we must ack the received FIN and
3197 * enter the CLOSING state.
3200 tcp_set_state(sk, TCP_CLOSING);
3203 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3205 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
3208 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3209 * cases we should never reach this piece of code.
3211 printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n",
3212 __FUNCTION__, sk->sk_state);
3216 /* It _is_ possible, that we have something out-of-order _after_ FIN.
3217 * Probably, we should reset in this case. For now drop them.
3219 __skb_queue_purge(&tp->out_of_order_queue);
3220 if (tp->rx_opt.sack_ok)
3221 tcp_sack_reset(&tp->rx_opt);
3222 sk_stream_mem_reclaim(sk);
3224 if (!sock_flag(sk, SOCK_DEAD)) {
3225 sk->sk_state_change(sk);
3227 /* Do not send POLL_HUP for half duplex close. */
3228 if (sk->sk_shutdown == SHUTDOWN_MASK ||
3229 sk->sk_state == TCP_CLOSE)
3230 sk_wake_async(sk, 1, POLL_HUP);
3232 sk_wake_async(sk, 1, POLL_IN);
3236 static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, u32 end_seq)
3238 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
3239 if (before(seq, sp->start_seq))
3240 sp->start_seq = seq;
3241 if (after(end_seq, sp->end_seq))
3242 sp->end_seq = end_seq;
3248 static void tcp_dsack_set(struct tcp_sock *tp, u32 seq, u32 end_seq)
3250 if (tp->rx_opt.sack_ok && sysctl_tcp_dsack) {
3251 if (before(seq, tp->rcv_nxt))
3252 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT);
3254 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT);
3256 tp->rx_opt.dsack = 1;
3257 tp->duplicate_sack[0].start_seq = seq;
3258 tp->duplicate_sack[0].end_seq = end_seq;
3259 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + 1, 4 - tp->rx_opt.tstamp_ok);
3263 static void tcp_dsack_extend(struct tcp_sock *tp, u32 seq, u32 end_seq)
3265 if (!tp->rx_opt.dsack)
3266 tcp_dsack_set(tp, seq, end_seq);
3268 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
3271 static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb)
3273 struct tcp_sock *tp = tcp_sk(sk);
3275 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
3276 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3277 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
3278 tcp_enter_quickack_mode(sk);
3280 if (tp->rx_opt.sack_ok && sysctl_tcp_dsack) {
3281 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3283 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
3284 end_seq = tp->rcv_nxt;
3285 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, end_seq);
3292 /* These routines update the SACK block as out-of-order packets arrive or
3293 * in-order packets close up the sequence space.
3295 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
3298 struct tcp_sack_block *sp = &tp->selective_acks[0];
3299 struct tcp_sack_block *swalk = sp+1;
3301 /* See if the recent change to the first SACK eats into
3302 * or hits the sequence space of other SACK blocks, if so coalesce.
3304 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks; ) {
3305 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
3308 /* Zap SWALK, by moving every further SACK up by one slot.
3309 * Decrease num_sacks.
3311 tp->rx_opt.num_sacks--;
3312 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
3313 for (i=this_sack; i < tp->rx_opt.num_sacks; i++)
3317 this_sack++, swalk++;
3321 static inline void tcp_sack_swap(struct tcp_sack_block *sack1, struct tcp_sack_block *sack2)
3325 tmp = sack1->start_seq;
3326 sack1->start_seq = sack2->start_seq;
3327 sack2->start_seq = tmp;
3329 tmp = sack1->end_seq;
3330 sack1->end_seq = sack2->end_seq;
3331 sack2->end_seq = tmp;
3334 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
3336 struct tcp_sock *tp = tcp_sk(sk);
3337 struct tcp_sack_block *sp = &tp->selective_acks[0];
3338 int cur_sacks = tp->rx_opt.num_sacks;
3344 for (this_sack=0; this_sack<cur_sacks; this_sack++, sp++) {
3345 if (tcp_sack_extend(sp, seq, end_seq)) {
3346 /* Rotate this_sack to the first one. */
3347 for (; this_sack>0; this_sack--, sp--)
3348 tcp_sack_swap(sp, sp-1);
3350 tcp_sack_maybe_coalesce(tp);
3355 /* Could not find an adjacent existing SACK, build a new one,
3356 * put it at the front, and shift everyone else down. We
3357 * always know there is at least one SACK present already here.
3359 * If the sack array is full, forget about the last one.
3361 if (this_sack >= 4) {
3363 tp->rx_opt.num_sacks--;
3366 for (; this_sack > 0; this_sack--, sp--)
3370 /* Build the new head SACK, and we're done. */
3371 sp->start_seq = seq;
3372 sp->end_seq = end_seq;
3373 tp->rx_opt.num_sacks++;
3374 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
3377 /* RCV.NXT advances, some SACKs should be eaten. */
3379 static void tcp_sack_remove(struct tcp_sock *tp)
3381 struct tcp_sack_block *sp = &tp->selective_acks[0];
3382 int num_sacks = tp->rx_opt.num_sacks;
3385 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
3386 if (skb_queue_empty(&tp->out_of_order_queue)) {
3387 tp->rx_opt.num_sacks = 0;
3388 tp->rx_opt.eff_sacks = tp->rx_opt.dsack;
3392 for (this_sack = 0; this_sack < num_sacks; ) {
3393 /* Check if the start of the sack is covered by RCV.NXT. */
3394 if (!before(tp->rcv_nxt, sp->start_seq)) {
3397 /* RCV.NXT must cover all the block! */
3398 BUG_TRAP(!before(tp->rcv_nxt, sp->end_seq));
3400 /* Zap this SACK, by moving forward any other SACKS. */
3401 for (i=this_sack+1; i < num_sacks; i++)
3402 tp->selective_acks[i-1] = tp->selective_acks[i];
3409 if (num_sacks != tp->rx_opt.num_sacks) {
3410 tp->rx_opt.num_sacks = num_sacks;
3411 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
3415 /* This one checks to see if we can put data from the
3416 * out_of_order queue into the receive_queue.
3418 static void tcp_ofo_queue(struct sock *sk)
3420 struct tcp_sock *tp = tcp_sk(sk);
3421 __u32 dsack_high = tp->rcv_nxt;
3422 struct sk_buff *skb;
3424 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
3425 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
3428 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
3429 __u32 dsack = dsack_high;
3430 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
3431 dsack_high = TCP_SKB_CB(skb)->end_seq;
3432 tcp_dsack_extend(tp, TCP_SKB_CB(skb)->seq, dsack);
3435 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3436 SOCK_DEBUG(sk, "ofo packet was already received \n");
3437 __skb_unlink(skb, &tp->out_of_order_queue);
3441 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
3442 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3443 TCP_SKB_CB(skb)->end_seq);
3445 __skb_unlink(skb, &tp->out_of_order_queue);
3446 __skb_queue_tail(&sk->sk_receive_queue, skb);
3447 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3448 if (tcp_hdr(skb)->fin)
3449 tcp_fin(skb, sk, tcp_hdr(skb));
3453 static int tcp_prune_queue(struct sock *sk);
3455 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
3457 struct tcphdr *th = tcp_hdr(skb);
3458 struct tcp_sock *tp = tcp_sk(sk);
3461 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
3464 __skb_pull(skb, th->doff*4);
3466 TCP_ECN_accept_cwr(tp, skb);
3468 if (tp->rx_opt.dsack) {
3469 tp->rx_opt.dsack = 0;
3470 tp->rx_opt.eff_sacks = min_t(unsigned int, tp->rx_opt.num_sacks,
3471 4 - tp->rx_opt.tstamp_ok);
3474 /* Queue data for delivery to the user.
3475 * Packets in sequence go to the receive queue.
3476 * Out of sequence packets to the out_of_order_queue.
3478 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
3479 if (tcp_receive_window(tp) == 0)
3482 /* Ok. In sequence. In window. */
3483 if (tp->ucopy.task == current &&
3484 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
3485 sock_owned_by_user(sk) && !tp->urg_data) {
3486 int chunk = min_t(unsigned int, skb->len,
3489 __set_current_state(TASK_RUNNING);
3492 if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
3493 tp->ucopy.len -= chunk;
3494 tp->copied_seq += chunk;
3495 eaten = (chunk == skb->len && !th->fin);
3496 tcp_rcv_space_adjust(sk);
3504 (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3505 !sk_stream_rmem_schedule(sk, skb))) {
3506 if (tcp_prune_queue(sk) < 0 ||
3507 !sk_stream_rmem_schedule(sk, skb))
3510 sk_stream_set_owner_r(skb, sk);
3511 __skb_queue_tail(&sk->sk_receive_queue, skb);
3513 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3515 tcp_event_data_recv(sk, skb);
3517 tcp_fin(skb, sk, th);
3519 if (!skb_queue_empty(&tp->out_of_order_queue)) {
3522 /* RFC2581. 4.2. SHOULD send immediate ACK, when
3523 * gap in queue is filled.
3525 if (skb_queue_empty(&tp->out_of_order_queue))
3526 inet_csk(sk)->icsk_ack.pingpong = 0;
3529 if (tp->rx_opt.num_sacks)
3530 tcp_sack_remove(tp);
3532 tcp_fast_path_check(sk);
3536 else if (!sock_flag(sk, SOCK_DEAD))
3537 sk->sk_data_ready(sk, 0);
3541 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3542 /* A retransmit, 2nd most common case. Force an immediate ack. */
3543 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
3544 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3547 tcp_enter_quickack_mode(sk);
3548 inet_csk_schedule_ack(sk);
3554 /* Out of window. F.e. zero window probe. */
3555 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
3558 tcp_enter_quickack_mode(sk);
3560 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3561 /* Partial packet, seq < rcv_next < end_seq */
3562 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
3563 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3564 TCP_SKB_CB(skb)->end_seq);
3566 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
3568 /* If window is closed, drop tail of packet. But after
3569 * remembering D-SACK for its head made in previous line.
3571 if (!tcp_receive_window(tp))
3576 TCP_ECN_check_ce(tp, skb);
3578 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3579 !sk_stream_rmem_schedule(sk, skb)) {
3580 if (tcp_prune_queue(sk) < 0 ||
3581 !sk_stream_rmem_schedule(sk, skb))
3585 /* Disable header prediction. */
3587 inet_csk_schedule_ack(sk);
3589 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
3590 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3592 sk_stream_set_owner_r(skb, sk);
3594 if (!skb_peek(&tp->out_of_order_queue)) {
3595 /* Initial out of order segment, build 1 SACK. */
3596 if (tp->rx_opt.sack_ok) {
3597 tp->rx_opt.num_sacks = 1;
3598 tp->rx_opt.dsack = 0;
3599 tp->rx_opt.eff_sacks = 1;
3600 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
3601 tp->selective_acks[0].end_seq =
3602 TCP_SKB_CB(skb)->end_seq;
3604 __skb_queue_head(&tp->out_of_order_queue,skb);
3606 struct sk_buff *skb1 = tp->out_of_order_queue.prev;
3607 u32 seq = TCP_SKB_CB(skb)->seq;
3608 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3610 if (seq == TCP_SKB_CB(skb1)->end_seq) {
3611 __skb_append(skb1, skb, &tp->out_of_order_queue);
3613 if (!tp->rx_opt.num_sacks ||
3614 tp->selective_acks[0].end_seq != seq)
3617 /* Common case: data arrive in order after hole. */
3618 tp->selective_acks[0].end_seq = end_seq;
3622 /* Find place to insert this segment. */
3624 if (!after(TCP_SKB_CB(skb1)->seq, seq))
3626 } while ((skb1 = skb1->prev) !=
3627 (struct sk_buff*)&tp->out_of_order_queue);
3629 /* Do skb overlap to previous one? */
3630 if (skb1 != (struct sk_buff*)&tp->out_of_order_queue &&
3631 before(seq, TCP_SKB_CB(skb1)->end_seq)) {
3632 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
3633 /* All the bits are present. Drop. */
3635 tcp_dsack_set(tp, seq, end_seq);
3638 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
3639 /* Partial overlap. */
3640 tcp_dsack_set(tp, seq, TCP_SKB_CB(skb1)->end_seq);
3645 __skb_insert(skb, skb1, skb1->next, &tp->out_of_order_queue);
3647 /* And clean segments covered by new one as whole. */
3648 while ((skb1 = skb->next) !=
3649 (struct sk_buff*)&tp->out_of_order_queue &&
3650 after(end_seq, TCP_SKB_CB(skb1)->seq)) {
3651 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
3652 tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, end_seq);
3655 __skb_unlink(skb1, &tp->out_of_order_queue);
3656 tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, TCP_SKB_CB(skb1)->end_seq);
3661 if (tp->rx_opt.sack_ok)
3662 tcp_sack_new_ofo_skb(sk, seq, end_seq);
3666 /* Collapse contiguous sequence of skbs head..tail with
3667 * sequence numbers start..end.
3668 * Segments with FIN/SYN are not collapsed (only because this
3672 tcp_collapse(struct sock *sk, struct sk_buff_head *list,
3673 struct sk_buff *head, struct sk_buff *tail,
3676 struct sk_buff *skb;
3678 /* First, check that queue is collapsible and find
3679 * the point where collapsing can be useful. */
3680 for (skb = head; skb != tail; ) {
3681 /* No new bits? It is possible on ofo queue. */
3682 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
3683 struct sk_buff *next = skb->next;
3684 __skb_unlink(skb, list);
3686 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
3691 /* The first skb to collapse is:
3693 * - bloated or contains data before "start" or
3694 * overlaps to the next one.
3696 if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
3697 (tcp_win_from_space(skb->truesize) > skb->len ||
3698 before(TCP_SKB_CB(skb)->seq, start) ||
3699 (skb->next != tail &&
3700 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq)))
3703 /* Decided to skip this, advance start seq. */
3704 start = TCP_SKB_CB(skb)->end_seq;
3707 if (skb == tail || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
3710 while (before(start, end)) {
3711 struct sk_buff *nskb;
3712 int header = skb_headroom(skb);
3713 int copy = SKB_MAX_ORDER(header, 0);
3715 /* Too big header? This can happen with IPv6. */
3718 if (end-start < copy)
3720 nskb = alloc_skb(copy+header, GFP_ATOMIC);
3724 skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
3725 skb_set_network_header(nskb, (skb_network_header(skb) -
3727 skb_set_transport_header(nskb, (skb_transport_header(skb) -
3729 skb_reserve(nskb, header);
3730 memcpy(nskb->head, skb->head, header);
3731 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
3732 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
3733 __skb_insert(nskb, skb->prev, skb, list);
3734 sk_stream_set_owner_r(nskb, sk);
3736 /* Copy data, releasing collapsed skbs. */
3738 int offset = start - TCP_SKB_CB(skb)->seq;
3739 int size = TCP_SKB_CB(skb)->end_seq - start;
3743 size = min(copy, size);
3744 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
3746 TCP_SKB_CB(nskb)->end_seq += size;
3750 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
3751 struct sk_buff *next = skb->next;
3752 __skb_unlink(skb, list);
3754 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
3757 tcp_hdr(skb)->syn ||
3765 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
3766 * and tcp_collapse() them until all the queue is collapsed.
3768 static void tcp_collapse_ofo_queue(struct sock *sk)
3770 struct tcp_sock *tp = tcp_sk(sk);
3771 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
3772 struct sk_buff *head;
3778 start = TCP_SKB_CB(skb)->seq;
3779 end = TCP_SKB_CB(skb)->end_seq;
3785 /* Segment is terminated when we see gap or when
3786 * we are at the end of all the queue. */
3787 if (skb == (struct sk_buff *)&tp->out_of_order_queue ||
3788 after(TCP_SKB_CB(skb)->seq, end) ||
3789 before(TCP_SKB_CB(skb)->end_seq, start)) {
3790 tcp_collapse(sk, &tp->out_of_order_queue,
3791 head, skb, start, end);
3793 if (skb == (struct sk_buff *)&tp->out_of_order_queue)
3795 /* Start new segment */
3796 start = TCP_SKB_CB(skb)->seq;
3797 end = TCP_SKB_CB(skb)->end_seq;
3799 if (before(TCP_SKB_CB(skb)->seq, start))
3800 start = TCP_SKB_CB(skb)->seq;
3801 if (after(TCP_SKB_CB(skb)->end_seq, end))
3802 end = TCP_SKB_CB(skb)->end_seq;
3807 /* Reduce allocated memory if we can, trying to get
3808 * the socket within its memory limits again.
3810 * Return less than zero if we should start dropping frames
3811 * until the socket owning process reads some of the data
3812 * to stabilize the situation.
3814 static int tcp_prune_queue(struct sock *sk)
3816 struct tcp_sock *tp = tcp_sk(sk);
3818 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
3820 NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED);
3822 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
3823 tcp_clamp_window(sk);
3824 else if (tcp_memory_pressure)
3825 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
3827 tcp_collapse_ofo_queue(sk);
3828 tcp_collapse(sk, &sk->sk_receive_queue,
3829 sk->sk_receive_queue.next,
3830 (struct sk_buff*)&sk->sk_receive_queue,
3831 tp->copied_seq, tp->rcv_nxt);
3832 sk_stream_mem_reclaim(sk);
3834 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
3837 /* Collapsing did not help, destructive actions follow.
3838 * This must not ever occur. */
3840 /* First, purge the out_of_order queue. */
3841 if (!skb_queue_empty(&tp->out_of_order_queue)) {
3842 NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED);
3843 __skb_queue_purge(&tp->out_of_order_queue);
3845 /* Reset SACK state. A conforming SACK implementation will
3846 * do the same at a timeout based retransmit. When a connection
3847 * is in a sad state like this, we care only about integrity
3848 * of the connection not performance.
3850 if (tp->rx_opt.sack_ok)
3851 tcp_sack_reset(&tp->rx_opt);
3852 sk_stream_mem_reclaim(sk);
3855 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
3858 /* If we are really being abused, tell the caller to silently
3859 * drop receive data on the floor. It will get retransmitted
3860 * and hopefully then we'll have sufficient space.
3862 NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED);
3864 /* Massive buffer overcommit. */
3870 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
3871 * As additional protections, we do not touch cwnd in retransmission phases,
3872 * and if application hit its sndbuf limit recently.
3874 void tcp_cwnd_application_limited(struct sock *sk)
3876 struct tcp_sock *tp = tcp_sk(sk);
3878 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
3879 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
3880 /* Limited by application or receiver window. */
3881 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
3882 u32 win_used = max(tp->snd_cwnd_used, init_win);
3883 if (win_used < tp->snd_cwnd) {
3884 tp->snd_ssthresh = tcp_current_ssthresh(sk);
3885 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
3887 tp->snd_cwnd_used = 0;
3889 tp->snd_cwnd_stamp = tcp_time_stamp;
3892 static int tcp_should_expand_sndbuf(struct sock *sk)
3894 struct tcp_sock *tp = tcp_sk(sk);
3896 /* If the user specified a specific send buffer setting, do
3899 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
3902 /* If we are under global TCP memory pressure, do not expand. */
3903 if (tcp_memory_pressure)
3906 /* If we are under soft global TCP memory pressure, do not expand. */
3907 if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0])
3910 /* If we filled the congestion window, do not expand. */
3911 if (tp->packets_out >= tp->snd_cwnd)
3917 /* When incoming ACK allowed to free some skb from write_queue,
3918 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
3919 * on the exit from tcp input handler.
3921 * PROBLEM: sndbuf expansion does not work well with largesend.
3923 static void tcp_new_space(struct sock *sk)
3925 struct tcp_sock *tp = tcp_sk(sk);
3927 if (tcp_should_expand_sndbuf(sk)) {
3928 int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
3929 MAX_TCP_HEADER + 16 + sizeof(struct sk_buff),
3930 demanded = max_t(unsigned int, tp->snd_cwnd,
3931 tp->reordering + 1);
3932 sndmem *= 2*demanded;
3933 if (sndmem > sk->sk_sndbuf)
3934 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
3935 tp->snd_cwnd_stamp = tcp_time_stamp;
3938 sk->sk_write_space(sk);
3941 static void tcp_check_space(struct sock *sk)
3943 if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
3944 sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
3945 if (sk->sk_socket &&
3946 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
3951 static inline void tcp_data_snd_check(struct sock *sk)
3953 tcp_push_pending_frames(sk);
3954 tcp_check_space(sk);
3958 * Check if sending an ack is needed.
3960 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
3962 struct tcp_sock *tp = tcp_sk(sk);
3964 /* More than one full frame received... */
3965 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss
3966 /* ... and right edge of window advances far enough.
3967 * (tcp_recvmsg() will send ACK otherwise). Or...
3969 && __tcp_select_window(sk) >= tp->rcv_wnd) ||
3970 /* We ACK each frame or... */
3971 tcp_in_quickack_mode(sk) ||
3972 /* We have out of order data. */
3974 skb_peek(&tp->out_of_order_queue))) {
3975 /* Then ack it now */
3978 /* Else, send delayed ack. */
3979 tcp_send_delayed_ack(sk);
3983 static inline void tcp_ack_snd_check(struct sock *sk)
3985 if (!inet_csk_ack_scheduled(sk)) {
3986 /* We sent a data segment already. */
3989 __tcp_ack_snd_check(sk, 1);
3993 * This routine is only called when we have urgent data
3994 * signaled. Its the 'slow' part of tcp_urg. It could be
3995 * moved inline now as tcp_urg is only called from one
3996 * place. We handle URGent data wrong. We have to - as
3997 * BSD still doesn't use the correction from RFC961.
3998 * For 1003.1g we should support a new option TCP_STDURG to permit
3999 * either form (or just set the sysctl tcp_stdurg).
4002 static void tcp_check_urg(struct sock * sk, struct tcphdr * th)
4004 struct tcp_sock *tp = tcp_sk(sk);
4005 u32 ptr = ntohs(th->urg_ptr);
4007 if (ptr && !sysctl_tcp_stdurg)
4009 ptr += ntohl(th->seq);
4011 /* Ignore urgent data that we've already seen and read. */
4012 if (after(tp->copied_seq, ptr))
4015 /* Do not replay urg ptr.
4017 * NOTE: interesting situation not covered by specs.
4018 * Misbehaving sender may send urg ptr, pointing to segment,
4019 * which we already have in ofo queue. We are not able to fetch
4020 * such data and will stay in TCP_URG_NOTYET until will be eaten
4021 * by recvmsg(). Seems, we are not obliged to handle such wicked
4022 * situations. But it is worth to think about possibility of some
4023 * DoSes using some hypothetical application level deadlock.
4025 if (before(ptr, tp->rcv_nxt))
4028 /* Do we already have a newer (or duplicate) urgent pointer? */
4029 if (tp->urg_data && !after(ptr, tp->urg_seq))
4032 /* Tell the world about our new urgent pointer. */
4035 /* We may be adding urgent data when the last byte read was
4036 * urgent. To do this requires some care. We cannot just ignore
4037 * tp->copied_seq since we would read the last urgent byte again
4038 * as data, nor can we alter copied_seq until this data arrives
4039 * or we break the semantics of SIOCATMARK (and thus sockatmark())
4041 * NOTE. Double Dutch. Rendering to plain English: author of comment
4042 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
4043 * and expect that both A and B disappear from stream. This is _wrong_.
4044 * Though this happens in BSD with high probability, this is occasional.
4045 * Any application relying on this is buggy. Note also, that fix "works"
4046 * only in this artificial test. Insert some normal data between A and B and we will
4047 * decline of BSD again. Verdict: it is better to remove to trap
4050 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
4051 !sock_flag(sk, SOCK_URGINLINE) &&
4052 tp->copied_seq != tp->rcv_nxt) {
4053 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
4055 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
4056 __skb_unlink(skb, &sk->sk_receive_queue);
4061 tp->urg_data = TCP_URG_NOTYET;
4064 /* Disable header prediction. */
4068 /* This is the 'fast' part of urgent handling. */
4069 static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th)
4071 struct tcp_sock *tp = tcp_sk(sk);
4073 /* Check if we get a new urgent pointer - normally not. */
4075 tcp_check_urg(sk,th);
4077 /* Do we wait for any urgent data? - normally not... */
4078 if (tp->urg_data == TCP_URG_NOTYET) {
4079 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
4082 /* Is the urgent pointer pointing into this packet? */
4083 if (ptr < skb->len) {
4085 if (skb_copy_bits(skb, ptr, &tmp, 1))
4087 tp->urg_data = TCP_URG_VALID | tmp;
4088 if (!sock_flag(sk, SOCK_DEAD))
4089 sk->sk_data_ready(sk, 0);
4094 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
4096 struct tcp_sock *tp = tcp_sk(sk);
4097 int chunk = skb->len - hlen;
4101 if (skb_csum_unnecessary(skb))
4102 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
4104 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
4108 tp->ucopy.len -= chunk;
4109 tp->copied_seq += chunk;
4110 tcp_rcv_space_adjust(sk);
4117 static __sum16 __tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
4121 if (sock_owned_by_user(sk)) {
4123 result = __tcp_checksum_complete(skb);
4126 result = __tcp_checksum_complete(skb);
4131 static inline int tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
4133 return !skb_csum_unnecessary(skb) &&
4134 __tcp_checksum_complete_user(sk, skb);
4137 #ifdef CONFIG_NET_DMA
4138 static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb, int hlen)
4140 struct tcp_sock *tp = tcp_sk(sk);
4141 int chunk = skb->len - hlen;
4143 int copied_early = 0;
4145 if (tp->ucopy.wakeup)
4148 if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
4149 tp->ucopy.dma_chan = get_softnet_dma();
4151 if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) {
4153 dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
4154 skb, hlen, tp->ucopy.iov, chunk, tp->ucopy.pinned_list);
4159 tp->ucopy.dma_cookie = dma_cookie;
4162 tp->ucopy.len -= chunk;
4163 tp->copied_seq += chunk;
4164 tcp_rcv_space_adjust(sk);
4166 if ((tp->ucopy.len == 0) ||
4167 (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) ||
4168 (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
4169 tp->ucopy.wakeup = 1;
4170 sk->sk_data_ready(sk, 0);
4172 } else if (chunk > 0) {
4173 tp->ucopy.wakeup = 1;
4174 sk->sk_data_ready(sk, 0);
4177 return copied_early;
4179 #endif /* CONFIG_NET_DMA */
4182 * TCP receive function for the ESTABLISHED state.
4184 * It is split into a fast path and a slow path. The fast path is
4186 * - A zero window was announced from us - zero window probing
4187 * is only handled properly in the slow path.
4188 * - Out of order segments arrived.
4189 * - Urgent data is expected.
4190 * - There is no buffer space left
4191 * - Unexpected TCP flags/window values/header lengths are received
4192 * (detected by checking the TCP header against pred_flags)
4193 * - Data is sent in both directions. Fast path only supports pure senders
4194 * or pure receivers (this means either the sequence number or the ack
4195 * value must stay constant)
4196 * - Unexpected TCP option.
4198 * When these conditions are not satisfied it drops into a standard
4199 * receive procedure patterned after RFC793 to handle all cases.
4200 * The first three cases are guaranteed by proper pred_flags setting,
4201 * the rest is checked inline. Fast processing is turned on in
4202 * tcp_data_queue when everything is OK.
4204 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
4205 struct tcphdr *th, unsigned len)
4207 struct tcp_sock *tp = tcp_sk(sk);
4210 * Header prediction.
4211 * The code loosely follows the one in the famous
4212 * "30 instruction TCP receive" Van Jacobson mail.
4214 * Van's trick is to deposit buffers into socket queue
4215 * on a device interrupt, to call tcp_recv function
4216 * on the receive process context and checksum and copy
4217 * the buffer to user space. smart...
4219 * Our current scheme is not silly either but we take the
4220 * extra cost of the net_bh soft interrupt processing...
4221 * We do checksum and copy also but from device to kernel.
4224 tp->rx_opt.saw_tstamp = 0;
4226 /* pred_flags is 0xS?10 << 16 + snd_wnd
4227 * if header_prediction is to be made
4228 * 'S' will always be tp->tcp_header_len >> 2
4229 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
4230 * turn it off (when there are holes in the receive
4231 * space for instance)
4232 * PSH flag is ignored.
4235 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
4236 TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4237 int tcp_header_len = tp->tcp_header_len;
4239 /* Timestamp header prediction: tcp_header_len
4240 * is automatically equal to th->doff*4 due to pred_flags
4244 /* Check timestamp */
4245 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
4246 __be32 *ptr = (__be32 *)(th + 1);
4248 /* No? Slow path! */
4249 if (*ptr != htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
4250 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP))
4253 tp->rx_opt.saw_tstamp = 1;
4255 tp->rx_opt.rcv_tsval = ntohl(*ptr);
4257 tp->rx_opt.rcv_tsecr = ntohl(*ptr);
4259 /* If PAWS failed, check it more carefully in slow path */
4260 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
4263 /* DO NOT update ts_recent here, if checksum fails
4264 * and timestamp was corrupted part, it will result
4265 * in a hung connection since we will drop all
4266 * future packets due to the PAWS test.
4270 if (len <= tcp_header_len) {
4271 /* Bulk data transfer: sender */
4272 if (len == tcp_header_len) {
4273 /* Predicted packet is in window by definition.
4274 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4275 * Hence, check seq<=rcv_wup reduces to:
4277 if (tcp_header_len ==
4278 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
4279 tp->rcv_nxt == tp->rcv_wup)
4280 tcp_store_ts_recent(tp);
4282 /* We know that such packets are checksummed
4285 tcp_ack(sk, skb, 0);
4287 tcp_data_snd_check(sk);
4289 } else { /* Header too small */
4290 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4295 int copied_early = 0;
4297 if (tp->copied_seq == tp->rcv_nxt &&
4298 len - tcp_header_len <= tp->ucopy.len) {
4299 #ifdef CONFIG_NET_DMA
4300 if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
4305 if (tp->ucopy.task == current && sock_owned_by_user(sk) && !copied_early) {
4306 __set_current_state(TASK_RUNNING);
4308 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
4312 /* Predicted packet is in window by definition.
4313 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4314 * Hence, check seq<=rcv_wup reduces to:
4316 if (tcp_header_len ==
4317 (sizeof(struct tcphdr) +
4318 TCPOLEN_TSTAMP_ALIGNED) &&
4319 tp->rcv_nxt == tp->rcv_wup)
4320 tcp_store_ts_recent(tp);
4322 tcp_rcv_rtt_measure_ts(sk, skb);
4324 __skb_pull(skb, tcp_header_len);
4325 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4326 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER);
4329 tcp_cleanup_rbuf(sk, skb->len);
4332 if (tcp_checksum_complete_user(sk, skb))
4335 /* Predicted packet is in window by definition.
4336 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4337 * Hence, check seq<=rcv_wup reduces to:
4339 if (tcp_header_len ==
4340 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
4341 tp->rcv_nxt == tp->rcv_wup)
4342 tcp_store_ts_recent(tp);
4344 tcp_rcv_rtt_measure_ts(sk, skb);
4346 if ((int)skb->truesize > sk->sk_forward_alloc)
4349 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS);
4351 /* Bulk data transfer: receiver */
4352 __skb_pull(skb,tcp_header_len);
4353 __skb_queue_tail(&sk->sk_receive_queue, skb);
4354 sk_stream_set_owner_r(skb, sk);
4355 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4358 tcp_event_data_recv(sk, skb);
4360 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
4361 /* Well, only one small jumplet in fast path... */
4362 tcp_ack(sk, skb, FLAG_DATA);
4363 tcp_data_snd_check(sk);
4364 if (!inet_csk_ack_scheduled(sk))
4368 __tcp_ack_snd_check(sk, 0);
4370 #ifdef CONFIG_NET_DMA
4372 __skb_queue_tail(&sk->sk_async_wait_queue, skb);
4378 sk->sk_data_ready(sk, 0);
4384 if (len < (th->doff<<2) || tcp_checksum_complete_user(sk, skb))
4388 * RFC1323: H1. Apply PAWS check first.
4390 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
4391 tcp_paws_discard(sk, skb)) {
4393 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
4394 tcp_send_dupack(sk, skb);
4397 /* Resets are accepted even if PAWS failed.
4399 ts_recent update must be made after we are sure
4400 that the packet is in window.
4405 * Standard slow path.
4408 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
4409 /* RFC793, page 37: "In all states except SYN-SENT, all reset
4410 * (RST) segments are validated by checking their SEQ-fields."
4411 * And page 69: "If an incoming segment is not acceptable,
4412 * an acknowledgment should be sent in reply (unless the RST bit
4413 * is set, if so drop the segment and return)".
4416 tcp_send_dupack(sk, skb);
4425 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
4427 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4428 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4429 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
4436 tcp_ack(sk, skb, FLAG_SLOWPATH);
4438 tcp_rcv_rtt_measure_ts(sk, skb);
4440 /* Process urgent data. */
4441 tcp_urg(sk, skb, th);
4443 /* step 7: process the segment text */
4444 tcp_data_queue(sk, skb);
4446 tcp_data_snd_check(sk);
4447 tcp_ack_snd_check(sk);
4451 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4458 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
4459 struct tcphdr *th, unsigned len)
4461 struct tcp_sock *tp = tcp_sk(sk);
4462 struct inet_connection_sock *icsk = inet_csk(sk);
4463 int saved_clamp = tp->rx_opt.mss_clamp;
4465 tcp_parse_options(skb, &tp->rx_opt, 0);
4469 * "If the state is SYN-SENT then
4470 * first check the ACK bit
4471 * If the ACK bit is set
4472 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
4473 * a reset (unless the RST bit is set, if so drop
4474 * the segment and return)"
4476 * We do not send data with SYN, so that RFC-correct
4479 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
4480 goto reset_and_undo;
4482 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
4483 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
4485 NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED);
4486 goto reset_and_undo;
4489 /* Now ACK is acceptable.
4491 * "If the RST bit is set
4492 * If the ACK was acceptable then signal the user "error:
4493 * connection reset", drop the segment, enter CLOSED state,
4494 * delete TCB, and return."
4503 * "fifth, if neither of the SYN or RST bits is set then
4504 * drop the segment and return."
4510 goto discard_and_undo;
4513 * "If the SYN bit is on ...
4514 * are acceptable then ...
4515 * (our SYN has been ACKed), change the connection
4516 * state to ESTABLISHED..."
4519 TCP_ECN_rcv_synack(tp, th);
4521 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
4522 tcp_ack(sk, skb, FLAG_SLOWPATH);
4524 /* Ok.. it's good. Set up sequence numbers and
4525 * move to established.
4527 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
4528 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
4530 /* RFC1323: The window in SYN & SYN/ACK segments is
4533 tp->snd_wnd = ntohs(th->window);
4534 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(skb)->seq);
4536 if (!tp->rx_opt.wscale_ok) {
4537 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
4538 tp->window_clamp = min(tp->window_clamp, 65535U);
4541 if (tp->rx_opt.saw_tstamp) {
4542 tp->rx_opt.tstamp_ok = 1;
4543 tp->tcp_header_len =
4544 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
4545 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
4546 tcp_store_ts_recent(tp);
4548 tp->tcp_header_len = sizeof(struct tcphdr);
4551 if (tp->rx_opt.sack_ok && sysctl_tcp_fack)
4552 tp->rx_opt.sack_ok |= 2;
4555 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
4556 tcp_initialize_rcv_mss(sk);
4558 /* Remember, tcp_poll() does not lock socket!
4559 * Change state from SYN-SENT only after copied_seq
4560 * is initialized. */
4561 tp->copied_seq = tp->rcv_nxt;
4563 tcp_set_state(sk, TCP_ESTABLISHED);
4565 security_inet_conn_established(sk, skb);
4567 /* Make sure socket is routed, for correct metrics. */
4568 icsk->icsk_af_ops->rebuild_header(sk);
4570 tcp_init_metrics(sk);
4572 tcp_init_congestion_control(sk);
4574 /* Prevent spurious tcp_cwnd_restart() on first data
4577 tp->lsndtime = tcp_time_stamp;
4579 tcp_init_buffer_space(sk);
4581 if (sock_flag(sk, SOCK_KEEPOPEN))
4582 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
4584 if (!tp->rx_opt.snd_wscale)
4585 __tcp_fast_path_on(tp, tp->snd_wnd);
4589 if (!sock_flag(sk, SOCK_DEAD)) {
4590 sk->sk_state_change(sk);
4591 sk_wake_async(sk, 0, POLL_OUT);
4594 if (sk->sk_write_pending ||
4595 icsk->icsk_accept_queue.rskq_defer_accept ||
4596 icsk->icsk_ack.pingpong) {
4597 /* Save one ACK. Data will be ready after
4598 * several ticks, if write_pending is set.
4600 * It may be deleted, but with this feature tcpdumps
4601 * look so _wonderfully_ clever, that I was not able
4602 * to stand against the temptation 8) --ANK
4604 inet_csk_schedule_ack(sk);
4605 icsk->icsk_ack.lrcvtime = tcp_time_stamp;
4606 icsk->icsk_ack.ato = TCP_ATO_MIN;
4607 tcp_incr_quickack(sk);
4608 tcp_enter_quickack_mode(sk);
4609 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
4610 TCP_DELACK_MAX, TCP_RTO_MAX);
4621 /* No ACK in the segment */
4625 * "If the RST bit is set
4627 * Otherwise (no ACK) drop the segment and return."
4630 goto discard_and_undo;
4634 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && tcp_paws_check(&tp->rx_opt, 0))
4635 goto discard_and_undo;
4638 /* We see SYN without ACK. It is attempt of
4639 * simultaneous connect with crossed SYNs.
4640 * Particularly, it can be connect to self.
4642 tcp_set_state(sk, TCP_SYN_RECV);
4644 if (tp->rx_opt.saw_tstamp) {
4645 tp->rx_opt.tstamp_ok = 1;
4646 tcp_store_ts_recent(tp);
4647 tp->tcp_header_len =
4648 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
4650 tp->tcp_header_len = sizeof(struct tcphdr);
4653 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
4654 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
4656 /* RFC1323: The window in SYN & SYN/ACK segments is
4659 tp->snd_wnd = ntohs(th->window);
4660 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
4661 tp->max_window = tp->snd_wnd;
4663 TCP_ECN_rcv_syn(tp, th);
4666 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
4667 tcp_initialize_rcv_mss(sk);
4670 tcp_send_synack(sk);
4672 /* Note, we could accept data and URG from this segment.
4673 * There are no obstacles to make this.
4675 * However, if we ignore data in ACKless segments sometimes,
4676 * we have no reasons to accept it sometimes.
4677 * Also, seems the code doing it in step6 of tcp_rcv_state_process
4678 * is not flawless. So, discard packet for sanity.
4679 * Uncomment this return to process the data.
4686 /* "fifth, if neither of the SYN or RST bits is set then
4687 * drop the segment and return."
4691 tcp_clear_options(&tp->rx_opt);
4692 tp->rx_opt.mss_clamp = saved_clamp;
4696 tcp_clear_options(&tp->rx_opt);
4697 tp->rx_opt.mss_clamp = saved_clamp;
4703 * This function implements the receiving procedure of RFC 793 for
4704 * all states except ESTABLISHED and TIME_WAIT.
4705 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
4706 * address independent.
4709 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
4710 struct tcphdr *th, unsigned len)
4712 struct tcp_sock *tp = tcp_sk(sk);
4713 struct inet_connection_sock *icsk = inet_csk(sk);
4716 tp->rx_opt.saw_tstamp = 0;
4718 switch (sk->sk_state) {
4730 if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
4733 /* Now we have several options: In theory there is
4734 * nothing else in the frame. KA9Q has an option to
4735 * send data with the syn, BSD accepts data with the
4736 * syn up to the [to be] advertised window and
4737 * Solaris 2.1 gives you a protocol error. For now
4738 * we just ignore it, that fits the spec precisely
4739 * and avoids incompatibilities. It would be nice in
4740 * future to drop through and process the data.
4742 * Now that TTCP is starting to be used we ought to
4744 * But, this leaves one open to an easy denial of
4745 * service attack, and SYN cookies can't defend
4746 * against this problem. So, we drop the data
4747 * in the interest of security over speed unless
4748 * it's still in use.
4756 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
4760 /* Do step6 onward by hand. */
4761 tcp_urg(sk, skb, th);
4763 tcp_data_snd_check(sk);
4767 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
4768 tcp_paws_discard(sk, skb)) {
4770 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
4771 tcp_send_dupack(sk, skb);
4774 /* Reset is accepted even if it did not pass PAWS. */
4777 /* step 1: check sequence number */
4778 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
4780 tcp_send_dupack(sk, skb);
4784 /* step 2: check RST bit */
4790 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
4792 /* step 3: check security and precedence [ignored] */
4796 * Check for a SYN in window.
4798 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4799 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
4804 /* step 5: check the ACK field */
4806 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH);
4808 switch (sk->sk_state) {
4811 tp->copied_seq = tp->rcv_nxt;
4813 tcp_set_state(sk, TCP_ESTABLISHED);
4814 sk->sk_state_change(sk);
4816 /* Note, that this wakeup is only for marginal
4817 * crossed SYN case. Passively open sockets
4818 * are not waked up, because sk->sk_sleep ==
4819 * NULL and sk->sk_socket == NULL.
4821 if (sk->sk_socket) {
4822 sk_wake_async(sk,0,POLL_OUT);
4825 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
4826 tp->snd_wnd = ntohs(th->window) <<
4827 tp->rx_opt.snd_wscale;
4828 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq,
4829 TCP_SKB_CB(skb)->seq);
4831 /* tcp_ack considers this ACK as duplicate
4832 * and does not calculate rtt.
4833 * Fix it at least with timestamps.
4835 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
4837 tcp_ack_saw_tstamp(sk, 0);
4839 if (tp->rx_opt.tstamp_ok)
4840 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
4842 /* Make sure socket is routed, for
4845 icsk->icsk_af_ops->rebuild_header(sk);
4847 tcp_init_metrics(sk);
4849 tcp_init_congestion_control(sk);
4851 /* Prevent spurious tcp_cwnd_restart() on
4852 * first data packet.
4854 tp->lsndtime = tcp_time_stamp;
4857 tcp_initialize_rcv_mss(sk);
4858 tcp_init_buffer_space(sk);
4859 tcp_fast_path_on(tp);
4866 if (tp->snd_una == tp->write_seq) {
4867 tcp_set_state(sk, TCP_FIN_WAIT2);
4868 sk->sk_shutdown |= SEND_SHUTDOWN;
4869 dst_confirm(sk->sk_dst_cache);
4871 if (!sock_flag(sk, SOCK_DEAD))
4872 /* Wake up lingering close() */
4873 sk->sk_state_change(sk);
4877 if (tp->linger2 < 0 ||
4878 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4879 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
4881 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
4885 tmo = tcp_fin_time(sk);
4886 if (tmo > TCP_TIMEWAIT_LEN) {
4887 inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
4888 } else if (th->fin || sock_owned_by_user(sk)) {
4889 /* Bad case. We could lose such FIN otherwise.
4890 * It is not a big problem, but it looks confusing
4891 * and not so rare event. We still can lose it now,
4892 * if it spins in bh_lock_sock(), but it is really
4895 inet_csk_reset_keepalive_timer(sk, tmo);
4897 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
4905 if (tp->snd_una == tp->write_seq) {
4906 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4912 if (tp->snd_una == tp->write_seq) {
4913 tcp_update_metrics(sk);
4922 /* step 6: check the URG bit */
4923 tcp_urg(sk, skb, th);
4925 /* step 7: process the segment text */
4926 switch (sk->sk_state) {
4927 case TCP_CLOSE_WAIT:
4930 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4934 /* RFC 793 says to queue data in these states,
4935 * RFC 1122 says we MUST send a reset.
4936 * BSD 4.4 also does reset.
4938 if (sk->sk_shutdown & RCV_SHUTDOWN) {
4939 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4940 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
4941 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
4947 case TCP_ESTABLISHED:
4948 tcp_data_queue(sk, skb);
4953 /* tcp_data could move socket to TIME-WAIT */
4954 if (sk->sk_state != TCP_CLOSE) {
4955 tcp_data_snd_check(sk);
4956 tcp_ack_snd_check(sk);
4966 EXPORT_SYMBOL(sysctl_tcp_ecn);
4967 EXPORT_SYMBOL(sysctl_tcp_reordering);
4968 EXPORT_SYMBOL(tcp_parse_options);
4969 EXPORT_SYMBOL(tcp_rcv_established);
4970 EXPORT_SYMBOL(tcp_rcv_state_process);
4971 EXPORT_SYMBOL(tcp_initialize_rcv_mss);