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 $
10 * Authors: Ross Biro, <bir7@leland.Stanford.Edu>
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 presnce 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
64 * Angelo Dell'Aera: TCP Westwood+ support
67 #include <linux/config.h>
69 #include <linux/module.h>
70 #include <linux/sysctl.h>
72 #include <net/inet_common.h>
73 #include <linux/ipsec.h>
75 int sysctl_tcp_timestamps = 1;
76 int sysctl_tcp_window_scaling = 1;
77 int sysctl_tcp_sack = 1;
78 int sysctl_tcp_fack = 1;
79 int sysctl_tcp_reordering = TCP_FASTRETRANS_THRESH;
81 int sysctl_tcp_dsack = 1;
82 int sysctl_tcp_app_win = 31;
83 int sysctl_tcp_adv_win_scale = 2;
85 int sysctl_tcp_stdurg;
86 int sysctl_tcp_rfc1337;
87 int sysctl_tcp_max_orphans = NR_FILE;
89 int sysctl_tcp_nometrics_save;
90 int sysctl_tcp_westwood;
91 int sysctl_tcp_vegas_cong_avoid;
93 int sysctl_tcp_moderate_rcvbuf = 1;
95 /* Default values of the Vegas variables, in fixed-point representation
96 * with V_PARAM_SHIFT bits to the right of the binary point.
98 #define V_PARAM_SHIFT 1
99 int sysctl_tcp_vegas_alpha = 1<<V_PARAM_SHIFT;
100 int sysctl_tcp_vegas_beta = 3<<V_PARAM_SHIFT;
101 int sysctl_tcp_vegas_gamma = 1<<V_PARAM_SHIFT;
102 int sysctl_tcp_bic = 1;
103 int sysctl_tcp_bic_fast_convergence = 1;
104 int sysctl_tcp_bic_low_window = 14;
106 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
107 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
108 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
109 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
110 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
111 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
112 #define FLAG_ECE 0x40 /* ECE in this ACK */
113 #define FLAG_DATA_LOST 0x80 /* SACK detected data lossage. */
114 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
116 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
117 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
118 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
119 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
121 #define IsReno(tp) ((tp)->sack_ok == 0)
122 #define IsFack(tp) ((tp)->sack_ok & 2)
123 #define IsDSack(tp) ((tp)->sack_ok & 4)
125 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
127 /* Adapt the MSS value used to make delayed ack decision to the
130 static __inline__ void tcp_measure_rcv_mss(struct tcp_opt *tp, struct sk_buff *skb)
132 unsigned int len, lss;
134 lss = tp->ack.last_seg_size;
135 tp->ack.last_seg_size = 0;
137 /* skb->len may jitter because of SACKs, even if peer
138 * sends good full-sized frames.
141 if (len >= tp->ack.rcv_mss) {
142 tp->ack.rcv_mss = len;
144 /* Otherwise, we make more careful check taking into account,
145 * that SACKs block is variable.
147 * "len" is invariant segment length, including TCP header.
149 len += skb->data - skb->h.raw;
150 if (len >= TCP_MIN_RCVMSS + sizeof(struct tcphdr) ||
151 /* If PSH is not set, packet should be
152 * full sized, provided peer TCP is not badly broken.
153 * This observation (if it is correct 8)) allows
154 * to handle super-low mtu links fairly.
156 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
157 !(tcp_flag_word(skb->h.th)&TCP_REMNANT))) {
158 /* Subtract also invariant (if peer is RFC compliant),
159 * tcp header plus fixed timestamp option length.
160 * Resulting "len" is MSS free of SACK jitter.
162 len -= tp->tcp_header_len;
163 tp->ack.last_seg_size = len;
165 tp->ack.rcv_mss = len;
169 tp->ack.pending |= TCP_ACK_PUSHED;
173 static void tcp_incr_quickack(struct tcp_opt *tp)
175 unsigned quickacks = tp->rcv_wnd/(2*tp->ack.rcv_mss);
179 if (quickacks > tp->ack.quick)
180 tp->ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
183 void tcp_enter_quickack_mode(struct tcp_opt *tp)
185 tcp_incr_quickack(tp);
186 tp->ack.pingpong = 0;
187 tp->ack.ato = TCP_ATO_MIN;
190 /* Send ACKs quickly, if "quick" count is not exhausted
191 * and the session is not interactive.
194 static __inline__ int tcp_in_quickack_mode(struct tcp_opt *tp)
196 return (tp->ack.quick && !tp->ack.pingpong);
199 /* Buffer size and advertised window tuning.
201 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
204 static void tcp_fixup_sndbuf(struct sock *sk)
206 int sndmem = tcp_sk(sk)->mss_clamp + MAX_TCP_HEADER + 16 +
207 sizeof(struct sk_buff);
209 if (sk->sk_sndbuf < 3 * sndmem)
210 sk->sk_sndbuf = min(3 * sndmem, sysctl_tcp_wmem[2]);
213 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
215 * All tcp_full_space() is split to two parts: "network" buffer, allocated
216 * forward and advertised in receiver window (tp->rcv_wnd) and
217 * "application buffer", required to isolate scheduling/application
218 * latencies from network.
219 * window_clamp is maximal advertised window. It can be less than
220 * tcp_full_space(), in this case tcp_full_space() - window_clamp
221 * is reserved for "application" buffer. The less window_clamp is
222 * the smoother our behaviour from viewpoint of network, but the lower
223 * throughput and the higher sensitivity of the connection to losses. 8)
225 * rcv_ssthresh is more strict window_clamp used at "slow start"
226 * phase to predict further behaviour of this connection.
227 * It is used for two goals:
228 * - to enforce header prediction at sender, even when application
229 * requires some significant "application buffer". It is check #1.
230 * - to prevent pruning of receive queue because of misprediction
231 * of receiver window. Check #2.
233 * The scheme does not work when sender sends good segments opening
234 * window and then starts to feed us spagetti. But it should work
235 * in common situations. Otherwise, we have to rely on queue collapsing.
238 /* Slow part of check#2. */
240 __tcp_grow_window(struct sock *sk, struct tcp_opt *tp, struct sk_buff *skb)
243 int truesize = tcp_win_from_space(skb->truesize)/2;
244 int window = tcp_full_space(sk)/2;
246 while (tp->rcv_ssthresh <= window) {
247 if (truesize <= skb->len)
248 return 2*tp->ack.rcv_mss;
256 static __inline__ void
257 tcp_grow_window(struct sock *sk, struct tcp_opt *tp, struct sk_buff *skb)
260 if (tp->rcv_ssthresh < tp->window_clamp &&
261 (int)tp->rcv_ssthresh < tcp_space(sk) &&
262 !tcp_memory_pressure) {
265 /* Check #2. Increase window, if skb with such overhead
266 * will fit to rcvbuf in future.
268 if (tcp_win_from_space(skb->truesize) <= skb->len)
271 incr = __tcp_grow_window(sk, tp, skb);
274 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr, tp->window_clamp);
280 /* 3. Tuning rcvbuf, when connection enters established state. */
282 static void tcp_fixup_rcvbuf(struct sock *sk)
284 struct tcp_opt *tp = tcp_sk(sk);
285 int rcvmem = tp->advmss + MAX_TCP_HEADER + 16 + sizeof(struct sk_buff);
287 /* Try to select rcvbuf so that 4 mss-sized segments
288 * will fit to window and correspoding skbs will fit to our rcvbuf.
289 * (was 3; 4 is minimum to allow fast retransmit to work.)
291 while (tcp_win_from_space(rcvmem) < tp->advmss)
293 if (sk->sk_rcvbuf < 4 * rcvmem)
294 sk->sk_rcvbuf = min(4 * rcvmem, sysctl_tcp_rmem[2]);
297 /* 4. Try to fixup all. It is made iimediately after connection enters
300 static void tcp_init_buffer_space(struct sock *sk)
302 struct tcp_opt *tp = tcp_sk(sk);
305 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
306 tcp_fixup_rcvbuf(sk);
307 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
308 tcp_fixup_sndbuf(sk);
310 tp->rcvq_space.space = tp->rcv_wnd;
312 maxwin = tcp_full_space(sk);
314 if (tp->window_clamp >= maxwin) {
315 tp->window_clamp = maxwin;
317 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
318 tp->window_clamp = max(maxwin -
319 (maxwin >> sysctl_tcp_app_win),
323 /* Force reservation of one segment. */
324 if (sysctl_tcp_app_win &&
325 tp->window_clamp > 2 * tp->advmss &&
326 tp->window_clamp + tp->advmss > maxwin)
327 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
329 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
330 tp->snd_cwnd_stamp = tcp_time_stamp;
333 static void init_bictcp(struct tcp_opt *tp)
337 tp->bictcp.last_max_cwnd = 0;
338 tp->bictcp.last_cwnd = 0;
339 tp->bictcp.last_stamp = 0;
342 /* 5. Recalculate window clamp after socket hit its memory bounds. */
343 static void tcp_clamp_window(struct sock *sk, struct tcp_opt *tp)
346 unsigned int app_win = tp->rcv_nxt - tp->copied_seq;
351 skb_queue_walk(&tp->out_of_order_queue, skb) {
355 /* If overcommit is due to out of order segments,
356 * do not clamp window. Try to expand rcvbuf instead.
359 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
360 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
361 !tcp_memory_pressure &&
362 atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0])
363 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
366 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf) {
368 if (atomic_read(&sk->sk_rmem_alloc) >= 2 * sk->sk_rcvbuf)
370 if (app_win > tp->ack.rcv_mss)
371 app_win -= tp->ack.rcv_mss;
372 app_win = max(app_win, 2U*tp->advmss);
375 tp->window_clamp = min(tp->window_clamp, app_win);
376 tp->rcv_ssthresh = min(tp->window_clamp, 2U*tp->advmss);
380 /* Receiver "autotuning" code.
382 * The algorithm for RTT estimation w/o timestamps is based on
383 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
384 * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps>
386 * More detail on this code can be found at
387 * <http://www.psc.edu/~jheffner/senior_thesis.ps>,
388 * though this reference is out of date. A new paper
391 static void tcp_rcv_rtt_update(struct tcp_opt *tp, u32 sample, int win_dep)
393 u32 new_sample = tp->rcv_rtt_est.rtt;
399 if (new_sample != 0) {
400 /* If we sample in larger samples in the non-timestamp
401 * case, we could grossly overestimate the RTT especially
402 * with chatty applications or bulk transfer apps which
403 * are stalled on filesystem I/O.
405 * Also, since we are only going for a minimum in the
406 * non-timestamp case, we do not smoothe things out
407 * else with timestamps disabled convergance takes too
411 m -= (new_sample >> 3);
413 } else if (m < new_sample)
416 /* No previous mesaure. */
420 if (tp->rcv_rtt_est.rtt != new_sample)
421 tp->rcv_rtt_est.rtt = new_sample;
424 static inline void tcp_rcv_rtt_measure(struct tcp_opt *tp)
426 if (tp->rcv_rtt_est.time == 0)
428 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
430 tcp_rcv_rtt_update(tp,
431 jiffies - tp->rcv_rtt_est.time,
435 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
436 tp->rcv_rtt_est.time = tcp_time_stamp;
439 static inline void tcp_rcv_rtt_measure_ts(struct tcp_opt *tp, struct sk_buff *skb)
442 (TCP_SKB_CB(skb)->end_seq -
443 TCP_SKB_CB(skb)->seq >= tp->ack.rcv_mss))
444 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rcv_tsecr, 0);
448 * This function should be called every time data is copied to user space.
449 * It calculates the appropriate TCP receive buffer space.
451 void tcp_rcv_space_adjust(struct sock *sk)
453 struct tcp_opt *tp = tcp_sk(sk);
457 if (tp->rcvq_space.time == 0)
460 time = tcp_time_stamp - tp->rcvq_space.time;
461 if (time < (tp->rcv_rtt_est.rtt >> 3) ||
462 tp->rcv_rtt_est.rtt == 0)
465 space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
467 space = max(tp->rcvq_space.space, space);
469 if (tp->rcvq_space.space != space) {
472 tp->rcvq_space.space = space;
474 if (sysctl_tcp_moderate_rcvbuf) {
475 int new_clamp = space;
477 /* Receive space grows, normalize in order to
478 * take into account packet headers and sk_buff
479 * structure overhead.
484 rcvmem = (tp->advmss + MAX_TCP_HEADER +
485 16 + sizeof(struct sk_buff));
486 while (tcp_win_from_space(rcvmem) < tp->advmss)
489 space = min(space, sysctl_tcp_rmem[2]);
490 if (space > sk->sk_rcvbuf) {
491 sk->sk_rcvbuf = space;
493 /* Make the window clamp follow along. */
494 tp->window_clamp = new_clamp;
500 tp->rcvq_space.seq = tp->copied_seq;
501 tp->rcvq_space.time = tcp_time_stamp;
504 /* There is something which you must keep in mind when you analyze the
505 * behavior of the tp->ato delayed ack timeout interval. When a
506 * connection starts up, we want to ack as quickly as possible. The
507 * problem is that "good" TCP's do slow start at the beginning of data
508 * transmission. The means that until we send the first few ACK's the
509 * sender will sit on his end and only queue most of his data, because
510 * he can only send snd_cwnd unacked packets at any given time. For
511 * each ACK we send, he increments snd_cwnd and transmits more of his
514 static void tcp_event_data_recv(struct sock *sk, struct tcp_opt *tp, struct sk_buff *skb)
518 tcp_schedule_ack(tp);
520 tcp_measure_rcv_mss(tp, skb);
522 tcp_rcv_rtt_measure(tp);
524 now = tcp_time_stamp;
527 /* The _first_ data packet received, initialize
528 * delayed ACK engine.
530 tcp_incr_quickack(tp);
531 tp->ack.ato = TCP_ATO_MIN;
533 int m = now - tp->ack.lrcvtime;
535 if (m <= TCP_ATO_MIN/2) {
536 /* The fastest case is the first. */
537 tp->ack.ato = (tp->ack.ato>>1) + TCP_ATO_MIN/2;
538 } else if (m < tp->ack.ato) {
539 tp->ack.ato = (tp->ack.ato>>1) + m;
540 if (tp->ack.ato > tp->rto)
541 tp->ack.ato = tp->rto;
542 } else if (m > tp->rto) {
543 /* Too long gap. Apparently sender falled to
544 * restart window, so that we send ACKs quickly.
546 tcp_incr_quickack(tp);
547 sk_stream_mem_reclaim(sk);
550 tp->ack.lrcvtime = now;
552 TCP_ECN_check_ce(tp, skb);
555 tcp_grow_window(sk, tp, skb);
558 /* Set up a new TCP connection, depending on whether it should be
559 * using Vegas or not.
561 void tcp_vegas_init(struct tcp_opt *tp)
563 if (sysctl_tcp_vegas_cong_avoid) {
564 tp->vegas.do_vegas = 1;
565 tp->vegas.baseRTT = 0x7fffffff;
566 tcp_vegas_enable(tp);
568 tcp_vegas_disable(tp);
571 /* Do RTT sampling needed for Vegas.
573 * o min-filter RTT samples from within an RTT to get the current
574 * propagation delay + queuing delay (we are min-filtering to try to
575 * avoid the effects of delayed ACKs)
576 * o min-filter RTT samples from a much longer window (forever for now)
577 * to find the propagation delay (baseRTT)
579 static inline void vegas_rtt_calc(struct tcp_opt *tp, __u32 rtt)
581 __u32 vrtt = rtt + 1; /* Never allow zero rtt or baseRTT */
583 /* Filter to find propagation delay: */
584 if (vrtt < tp->vegas.baseRTT)
585 tp->vegas.baseRTT = vrtt;
587 /* Find the min RTT during the last RTT to find
588 * the current prop. delay + queuing delay:
590 tp->vegas.minRTT = min(tp->vegas.minRTT, vrtt);
594 /* Called to compute a smoothed rtt estimate. The data fed to this
595 * routine either comes from timestamps, or from segments that were
596 * known _not_ to have been retransmitted [see Karn/Partridge
597 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
598 * piece by Van Jacobson.
599 * NOTE: the next three routines used to be one big routine.
600 * To save cycles in the RFC 1323 implementation it was better to break
601 * it up into three procedures. -- erics
603 static void tcp_rtt_estimator(struct tcp_opt *tp, __u32 mrtt)
605 long m = mrtt; /* RTT */
607 if (tcp_vegas_enabled(tp))
608 vegas_rtt_calc(tp, mrtt);
610 /* The following amusing code comes from Jacobson's
611 * article in SIGCOMM '88. Note that rtt and mdev
612 * are scaled versions of rtt and mean deviation.
613 * This is designed to be as fast as possible
614 * m stands for "measurement".
616 * On a 1990 paper the rto value is changed to:
617 * RTO = rtt + 4 * mdev
619 * Funny. This algorithm seems to be very broken.
620 * These formulae increase RTO, when it should be decreased, increase
621 * too slowly, when it should be incresed fastly, decrease too fastly
622 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
623 * does not matter how to _calculate_ it. Seems, it was trap
624 * that VJ failed to avoid. 8)
629 m -= (tp->srtt >> 3); /* m is now error in rtt est */
630 tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */
632 m = -m; /* m is now abs(error) */
633 m -= (tp->mdev >> 2); /* similar update on mdev */
634 /* This is similar to one of Eifel findings.
635 * Eifel blocks mdev updates when rtt decreases.
636 * This solution is a bit different: we use finer gain
637 * for mdev in this case (alpha*beta).
638 * Like Eifel it also prevents growth of rto,
639 * but also it limits too fast rto decreases,
640 * happening in pure Eifel.
645 m -= (tp->mdev >> 2); /* similar update on mdev */
647 tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */
648 if (tp->mdev > tp->mdev_max) {
649 tp->mdev_max = tp->mdev;
650 if (tp->mdev_max > tp->rttvar)
651 tp->rttvar = tp->mdev_max;
653 if (after(tp->snd_una, tp->rtt_seq)) {
654 if (tp->mdev_max < tp->rttvar)
655 tp->rttvar -= (tp->rttvar-tp->mdev_max)>>2;
656 tp->rtt_seq = tp->snd_nxt;
657 tp->mdev_max = TCP_RTO_MIN;
660 /* no previous measure. */
661 tp->srtt = m<<3; /* take the measured time to be rtt */
662 tp->mdev = m<<1; /* make sure rto = 3*rtt */
663 tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN);
664 tp->rtt_seq = tp->snd_nxt;
667 tcp_westwood_update_rtt(tp, tp->srtt >> 3);
670 /* Calculate rto without backoff. This is the second half of Van Jacobson's
671 * routine referred to above.
673 static __inline__ void tcp_set_rto(struct tcp_opt *tp)
675 /* Old crap is replaced with new one. 8)
678 * 1. If rtt variance happened to be less 50msec, it is hallucination.
679 * It cannot be less due to utterly erratic ACK generation made
680 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
681 * to do with delayed acks, because at cwnd>2 true delack timeout
682 * is invisible. Actually, Linux-2.4 also generates erratic
683 * ACKs in some curcumstances.
685 tp->rto = (tp->srtt >> 3) + tp->rttvar;
687 /* 2. Fixups made earlier cannot be right.
688 * If we do not estimate RTO correctly without them,
689 * all the algo is pure shit and should be replaced
690 * with correct one. It is exaclty, which we pretend to do.
694 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
695 * guarantees that rto is higher.
697 static __inline__ void tcp_bound_rto(struct tcp_opt *tp)
699 if (tp->rto > TCP_RTO_MAX)
700 tp->rto = TCP_RTO_MAX;
703 /* Save metrics learned by this TCP session.
704 This function is called only, when TCP finishes successfully
705 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
707 void tcp_update_metrics(struct sock *sk)
709 struct tcp_opt *tp = tcp_sk(sk);
710 struct dst_entry *dst = __sk_dst_get(sk);
712 if (sysctl_tcp_nometrics_save)
717 if (dst && (dst->flags&DST_HOST)) {
720 if (tp->backoff || !tp->srtt) {
721 /* This session failed to estimate rtt. Why?
722 * Probably, no packets returned in time.
725 if (!(dst_metric_locked(dst, RTAX_RTT)))
726 dst->metrics[RTAX_RTT-1] = 0;
730 m = dst_metric(dst, RTAX_RTT) - tp->srtt;
732 /* If newly calculated rtt larger than stored one,
733 * store new one. Otherwise, use EWMA. Remember,
734 * rtt overestimation is always better than underestimation.
736 if (!(dst_metric_locked(dst, RTAX_RTT))) {
738 dst->metrics[RTAX_RTT-1] = tp->srtt;
740 dst->metrics[RTAX_RTT-1] -= (m>>3);
743 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) {
747 /* Scale deviation to rttvar fixed point */
752 if (m >= dst_metric(dst, RTAX_RTTVAR))
753 dst->metrics[RTAX_RTTVAR-1] = m;
755 dst->metrics[RTAX_RTTVAR-1] -=
756 (dst->metrics[RTAX_RTTVAR-1] - m)>>2;
759 if (tp->snd_ssthresh >= 0xFFFF) {
760 /* Slow start still did not finish. */
761 if (dst_metric(dst, RTAX_SSTHRESH) &&
762 !dst_metric_locked(dst, RTAX_SSTHRESH) &&
763 (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH))
764 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1;
765 if (!dst_metric_locked(dst, RTAX_CWND) &&
766 tp->snd_cwnd > dst_metric(dst, RTAX_CWND))
767 dst->metrics[RTAX_CWND-1] = tp->snd_cwnd;
768 } else if (tp->snd_cwnd > tp->snd_ssthresh &&
769 tp->ca_state == TCP_CA_Open) {
770 /* Cong. avoidance phase, cwnd is reliable. */
771 if (!dst_metric_locked(dst, RTAX_SSTHRESH))
772 dst->metrics[RTAX_SSTHRESH-1] =
773 max(tp->snd_cwnd >> 1, tp->snd_ssthresh);
774 if (!dst_metric_locked(dst, RTAX_CWND))
775 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_cwnd) >> 1;
777 /* Else slow start did not finish, cwnd is non-sense,
778 ssthresh may be also invalid.
780 if (!dst_metric_locked(dst, RTAX_CWND))
781 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_ssthresh) >> 1;
782 if (dst->metrics[RTAX_SSTHRESH-1] &&
783 !dst_metric_locked(dst, RTAX_SSTHRESH) &&
784 tp->snd_ssthresh > dst->metrics[RTAX_SSTHRESH-1])
785 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh;
788 if (!dst_metric_locked(dst, RTAX_REORDERING)) {
789 if (dst->metrics[RTAX_REORDERING-1] < tp->reordering &&
790 tp->reordering != sysctl_tcp_reordering)
791 dst->metrics[RTAX_REORDERING-1] = tp->reordering;
796 /* Numbers are taken from RFC2414. */
797 __u32 tcp_init_cwnd(struct tcp_opt *tp, struct dst_entry *dst)
799 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
802 if (tp->mss_cache > 1460)
805 cwnd = (tp->mss_cache > 1095) ? 3 : 4;
807 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
810 /* Initialize metrics on socket. */
812 static void tcp_init_metrics(struct sock *sk)
814 struct tcp_opt *tp = tcp_sk(sk);
815 struct dst_entry *dst = __sk_dst_get(sk);
822 if (dst_metric_locked(dst, RTAX_CWND))
823 tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND);
824 if (dst_metric(dst, RTAX_SSTHRESH)) {
825 tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH);
826 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
827 tp->snd_ssthresh = tp->snd_cwnd_clamp;
829 if (dst_metric(dst, RTAX_REORDERING) &&
830 tp->reordering != dst_metric(dst, RTAX_REORDERING)) {
832 tp->reordering = dst_metric(dst, RTAX_REORDERING);
835 if (dst_metric(dst, RTAX_RTT) == 0)
838 if (!tp->srtt && dst_metric(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3))
841 /* Initial rtt is determined from SYN,SYN-ACK.
842 * The segment is small and rtt may appear much
843 * less than real one. Use per-dst memory
844 * to make it more realistic.
846 * A bit of theory. RTT is time passed after "normal" sized packet
847 * is sent until it is ACKed. In normal curcumstances sending small
848 * packets force peer to delay ACKs and calculation is correct too.
849 * The algorithm is adaptive and, provided we follow specs, it
850 * NEVER underestimate RTT. BUT! If peer tries to make some clever
851 * tricks sort of "quick acks" for time long enough to decrease RTT
852 * to low value, and then abruptly stops to do it and starts to delay
853 * ACKs, wait for troubles.
855 if (dst_metric(dst, RTAX_RTT) > tp->srtt)
856 tp->srtt = dst_metric(dst, RTAX_RTT);
857 if (dst_metric(dst, RTAX_RTTVAR) > tp->mdev) {
858 tp->mdev = dst_metric(dst, RTAX_RTTVAR);
859 tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN);
863 if (tp->rto < TCP_TIMEOUT_INIT && !tp->saw_tstamp)
865 tp->snd_cwnd = tcp_init_cwnd(tp, dst);
866 tp->snd_cwnd_stamp = tcp_time_stamp;
870 /* Play conservative. If timestamps are not
871 * supported, TCP will fail to recalculate correct
872 * rtt, if initial rto is too small. FORGET ALL AND RESET!
874 if (!tp->saw_tstamp && tp->srtt) {
876 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT;
877 tp->rto = TCP_TIMEOUT_INIT;
881 static void tcp_update_reordering(struct tcp_opt *tp, int metric, int ts)
883 if (metric > tp->reordering) {
884 tp->reordering = min(TCP_MAX_REORDERING, metric);
886 /* This exciting event is worth to be remembered. 8) */
888 NET_INC_STATS_BH(LINUX_MIB_TCPTSREORDER);
890 NET_INC_STATS_BH(LINUX_MIB_TCPRENOREORDER);
892 NET_INC_STATS_BH(LINUX_MIB_TCPFACKREORDER);
894 NET_INC_STATS_BH(LINUX_MIB_TCPSACKREORDER);
895 #if FASTRETRANS_DEBUG > 1
896 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n",
897 tp->sack_ok, tp->ca_state,
898 tp->reordering, tp->fackets_out, tp->sacked_out,
899 tp->undo_marker ? tp->undo_retrans : 0);
901 /* Disable FACK yet. */
906 /* This procedure tags the retransmission queue when SACKs arrive.
908 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
909 * Packets in queue with these bits set are counted in variables
910 * sacked_out, retrans_out and lost_out, correspondingly.
912 * Valid combinations are:
913 * Tag InFlight Description
914 * 0 1 - orig segment is in flight.
915 * S 0 - nothing flies, orig reached receiver.
916 * L 0 - nothing flies, orig lost by net.
917 * R 2 - both orig and retransmit are in flight.
918 * L|R 1 - orig is lost, retransmit is in flight.
919 * S|R 1 - orig reached receiver, retrans is still in flight.
920 * (L|S|R is logically valid, it could occur when L|R is sacked,
921 * but it is equivalent to plain S and code short-curcuits it to S.
922 * L|S is logically invalid, it would mean -1 packet in flight 8))
924 * These 6 states form finite state machine, controlled by the following events:
925 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
926 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
927 * 3. Loss detection event of one of three flavors:
928 * A. Scoreboard estimator decided the packet is lost.
929 * A'. Reno "three dupacks" marks head of queue lost.
930 * A''. Its FACK modfication, head until snd.fack is lost.
931 * B. SACK arrives sacking data transmitted after never retransmitted
933 * C. SACK arrives sacking SND.NXT at the moment, when the
934 * segment was retransmitted.
935 * 4. D-SACK added new rule: D-SACK changes any tag to S.
937 * It is pleasant to note, that state diagram turns out to be commutative,
938 * so that we are allowed not to be bothered by order of our actions,
939 * when multiple events arrive simultaneously. (see the function below).
941 * Reordering detection.
942 * --------------------
943 * Reordering metric is maximal distance, which a packet can be displaced
944 * in packet stream. With SACKs we can estimate it:
946 * 1. SACK fills old hole and the corresponding segment was not
947 * ever retransmitted -> reordering. Alas, we cannot use it
948 * when segment was retransmitted.
949 * 2. The last flaw is solved with D-SACK. D-SACK arrives
950 * for retransmitted and already SACKed segment -> reordering..
951 * Both of these heuristics are not used in Loss state, when we cannot
952 * account for retransmits accurately.
955 tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, u32 prior_snd_una)
957 struct tcp_opt *tp = tcp_sk(sk);
958 unsigned char *ptr = ack_skb->h.raw + TCP_SKB_CB(ack_skb)->sacked;
959 struct tcp_sack_block *sp = (struct tcp_sack_block *)(ptr+2);
960 int num_sacks = (ptr[1] - TCPOLEN_SACK_BASE)>>3;
961 int reord = tp->packets_out;
963 u32 lost_retrans = 0;
967 /* So, SACKs for already sent large segments will be lost.
968 * Not good, but alternative is to resegment the queue. */
969 if (sk->sk_route_caps & NETIF_F_TSO) {
970 sk->sk_route_caps &= ~NETIF_F_TSO;
971 sk->sk_no_largesend = 1;
972 tp->mss_cache = tp->mss_cache_std;
977 prior_fackets = tp->fackets_out;
979 for (i=0; i<num_sacks; i++, sp++) {
981 __u32 start_seq = ntohl(sp->start_seq);
982 __u32 end_seq = ntohl(sp->end_seq);
986 /* Check for D-SACK. */
988 u32 ack = TCP_SKB_CB(ack_skb)->ack_seq;
990 if (before(start_seq, ack)) {
993 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV);
994 } else if (num_sacks > 1 &&
995 !after(end_seq, ntohl(sp[1].end_seq)) &&
996 !before(start_seq, ntohl(sp[1].start_seq))) {
999 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV);
1002 /* D-SACK for already forgotten data...
1003 * Do dumb counting. */
1005 !after(end_seq, prior_snd_una) &&
1006 after(end_seq, tp->undo_marker))
1009 /* Eliminate too old ACKs, but take into
1010 * account more or less fresh ones, they can
1011 * contain valid SACK info.
1013 if (before(ack, prior_snd_una - tp->max_window))
1017 /* Event "B" in the comment above. */
1018 if (after(end_seq, tp->high_seq))
1019 flag |= FLAG_DATA_LOST;
1021 sk_stream_for_retrans_queue(skb, sk) {
1022 u8 sacked = TCP_SKB_CB(skb)->sacked;
1025 /* The retransmission queue is always in order, so
1026 * we can short-circuit the walk early.
1028 if(!before(TCP_SKB_CB(skb)->seq, end_seq))
1033 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1034 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1036 /* Account D-SACK for retransmitted packet. */
1037 if ((dup_sack && in_sack) &&
1038 (sacked & TCPCB_RETRANS) &&
1039 after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker))
1042 /* The frame is ACKed. */
1043 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) {
1044 if (sacked&TCPCB_RETRANS) {
1045 if ((dup_sack && in_sack) &&
1046 (sacked&TCPCB_SACKED_ACKED))
1047 reord = min(fack_count, reord);
1049 /* If it was in a hole, we detected reordering. */
1050 if (fack_count < prior_fackets &&
1051 !(sacked&TCPCB_SACKED_ACKED))
1052 reord = min(fack_count, reord);
1055 /* Nothing to do; acked frame is about to be dropped. */
1059 if ((sacked&TCPCB_SACKED_RETRANS) &&
1060 after(end_seq, TCP_SKB_CB(skb)->ack_seq) &&
1061 (!lost_retrans || after(end_seq, lost_retrans)))
1062 lost_retrans = end_seq;
1067 if (!(sacked&TCPCB_SACKED_ACKED)) {
1068 if (sacked & TCPCB_SACKED_RETRANS) {
1069 /* If the segment is not tagged as lost,
1070 * we do not clear RETRANS, believing
1071 * that retransmission is still in flight.
1073 if (sacked & TCPCB_LOST) {
1074 TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1079 /* New sack for not retransmitted frame,
1080 * which was in hole. It is reordering.
1082 if (!(sacked & TCPCB_RETRANS) &&
1083 fack_count < prior_fackets)
1084 reord = min(fack_count, reord);
1086 if (sacked & TCPCB_LOST) {
1087 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1092 TCP_SKB_CB(skb)->sacked |= TCPCB_SACKED_ACKED;
1093 flag |= FLAG_DATA_SACKED;
1096 if (fack_count > tp->fackets_out)
1097 tp->fackets_out = fack_count;
1099 if (dup_sack && (sacked&TCPCB_RETRANS))
1100 reord = min(fack_count, reord);
1103 /* D-SACK. We can detect redundant retransmission
1104 * in S|R and plain R frames and clear it.
1105 * undo_retrans is decreased above, L|R frames
1106 * are accounted above as well.
1109 (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS)) {
1110 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1116 /* Check for lost retransmit. This superb idea is
1117 * borrowed from "ratehalving". Event "C".
1118 * Later note: FACK people cheated me again 8),
1119 * we have to account for reordering! Ugly,
1122 if (lost_retrans && tp->ca_state == TCP_CA_Recovery) {
1123 struct sk_buff *skb;
1125 sk_stream_for_retrans_queue(skb, sk) {
1126 if (after(TCP_SKB_CB(skb)->seq, lost_retrans))
1128 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1130 if ((TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) &&
1131 after(lost_retrans, TCP_SKB_CB(skb)->ack_seq) &&
1133 !before(lost_retrans,
1134 TCP_SKB_CB(skb)->ack_seq + tp->reordering *
1136 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1139 if (!(TCP_SKB_CB(skb)->sacked&(TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1141 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1142 flag |= FLAG_DATA_SACKED;
1143 NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT);
1149 tp->left_out = tp->sacked_out + tp->lost_out;
1151 if (reord < tp->fackets_out && tp->ca_state != TCP_CA_Loss)
1152 tcp_update_reordering(tp, (tp->fackets_out + 1) - reord, 0);
1154 #if FASTRETRANS_DEBUG > 0
1155 BUG_TRAP((int)tp->sacked_out >= 0);
1156 BUG_TRAP((int)tp->lost_out >= 0);
1157 BUG_TRAP((int)tp->retrans_out >= 0);
1158 BUG_TRAP((int)tcp_packets_in_flight(tp) >= 0);
1163 /* RTO occurred, but do not yet enter loss state. Instead, transmit two new
1164 * segments to see from the next ACKs whether any data was really missing.
1165 * If the RTO was spurious, new ACKs should arrive.
1167 void tcp_enter_frto(struct sock *sk)
1169 struct tcp_opt *tp = tcp_sk(sk);
1170 struct sk_buff *skb;
1172 tp->frto_counter = 1;
1174 if (tp->ca_state <= TCP_CA_Disorder ||
1175 tp->snd_una == tp->high_seq ||
1176 (tp->ca_state == TCP_CA_Loss && !tp->retransmits)) {
1177 tp->prior_ssthresh = tcp_current_ssthresh(tp);
1178 if (!tcp_westwood_ssthresh(tp))
1179 tp->snd_ssthresh = tcp_recalc_ssthresh(tp);
1182 /* Have to clear retransmission markers here to keep the bookkeeping
1183 * in shape, even though we are not yet in Loss state.
1184 * If something was really lost, it is eventually caught up
1185 * in tcp_enter_frto_loss.
1187 tp->retrans_out = 0;
1188 tp->undo_marker = tp->snd_una;
1189 tp->undo_retrans = 0;
1191 sk_stream_for_retrans_queue(skb, sk) {
1192 TCP_SKB_CB(skb)->sacked &= ~TCPCB_RETRANS;
1194 tcp_sync_left_out(tp);
1196 tcp_set_ca_state(tp, TCP_CA_Open);
1197 tp->frto_highmark = tp->snd_nxt;
1200 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1201 * which indicates that we should follow the traditional RTO recovery,
1202 * i.e. mark everything lost and do go-back-N retransmission.
1204 static void tcp_enter_frto_loss(struct sock *sk)
1206 struct tcp_opt *tp = tcp_sk(sk);
1207 struct sk_buff *skb;
1212 tp->fackets_out = 0;
1214 sk_stream_for_retrans_queue(skb, sk) {
1216 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1217 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) {
1219 /* Do not mark those segments lost that were
1220 * forward transmitted after RTO
1222 if(!after(TCP_SKB_CB(skb)->end_seq,
1223 tp->frto_highmark)) {
1224 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1229 tp->fackets_out = cnt;
1232 tcp_sync_left_out(tp);
1234 tp->snd_cwnd = tp->frto_counter + tcp_packets_in_flight(tp)+1;
1235 tp->snd_cwnd_cnt = 0;
1236 tp->snd_cwnd_stamp = tcp_time_stamp;
1237 tp->undo_marker = 0;
1238 tp->frto_counter = 0;
1240 tp->reordering = min_t(unsigned int, tp->reordering,
1241 sysctl_tcp_reordering);
1242 tcp_set_ca_state(tp, TCP_CA_Loss);
1243 tp->high_seq = tp->frto_highmark;
1244 TCP_ECN_queue_cwr(tp);
1249 void tcp_clear_retrans(struct tcp_opt *tp)
1252 tp->retrans_out = 0;
1254 tp->fackets_out = 0;
1258 tp->undo_marker = 0;
1259 tp->undo_retrans = 0;
1262 /* Enter Loss state. If "how" is not zero, forget all SACK information
1263 * and reset tags completely, otherwise preserve SACKs. If receiver
1264 * dropped its ofo queue, we will know this due to reneging detection.
1266 void tcp_enter_loss(struct sock *sk, int how)
1268 struct tcp_opt *tp = tcp_sk(sk);
1269 struct sk_buff *skb;
1272 /* Reduce ssthresh if it has not yet been made inside this window. */
1273 if (tp->ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
1274 (tp->ca_state == TCP_CA_Loss && !tp->retransmits)) {
1275 tp->prior_ssthresh = tcp_current_ssthresh(tp);
1276 tp->snd_ssthresh = tcp_recalc_ssthresh(tp);
1279 tp->snd_cwnd_cnt = 0;
1280 tp->snd_cwnd_stamp = tcp_time_stamp;
1282 tcp_clear_retrans(tp);
1284 /* Push undo marker, if it was plain RTO and nothing
1285 * was retransmitted. */
1287 tp->undo_marker = tp->snd_una;
1289 sk_stream_for_retrans_queue(skb, sk) {
1291 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS)
1292 tp->undo_marker = 0;
1293 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1294 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
1295 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1296 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1300 tp->fackets_out = cnt;
1303 tcp_sync_left_out(tp);
1305 tp->reordering = min_t(unsigned int, tp->reordering,
1306 sysctl_tcp_reordering);
1307 tcp_set_ca_state(tp, TCP_CA_Loss);
1308 tp->high_seq = tp->snd_nxt;
1309 TCP_ECN_queue_cwr(tp);
1312 static int tcp_check_sack_reneging(struct sock *sk, struct tcp_opt *tp)
1314 struct sk_buff *skb;
1316 /* If ACK arrived pointing to a remembered SACK,
1317 * it means that our remembered SACKs do not reflect
1318 * real state of receiver i.e.
1319 * receiver _host_ is heavily congested (or buggy).
1320 * Do processing similar to RTO timeout.
1322 if ((skb = skb_peek(&sk->sk_write_queue)) != NULL &&
1323 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
1324 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING);
1326 tcp_enter_loss(sk, 1);
1328 tcp_retransmit_skb(sk, skb_peek(&sk->sk_write_queue));
1329 tcp_reset_xmit_timer(sk, TCP_TIME_RETRANS, tp->rto);
1335 static inline int tcp_fackets_out(struct tcp_opt *tp)
1337 return IsReno(tp) ? tp->sacked_out+1 : tp->fackets_out;
1340 static inline int tcp_skb_timedout(struct tcp_opt *tp, struct sk_buff *skb)
1342 return (tcp_time_stamp - TCP_SKB_CB(skb)->when > tp->rto);
1345 static inline int tcp_head_timedout(struct sock *sk, struct tcp_opt *tp)
1347 return tp->packets_out &&
1348 tcp_skb_timedout(tp, skb_peek(&sk->sk_write_queue));
1351 /* Linux NewReno/SACK/FACK/ECN state machine.
1352 * --------------------------------------
1354 * "Open" Normal state, no dubious events, fast path.
1355 * "Disorder" In all the respects it is "Open",
1356 * but requires a bit more attention. It is entered when
1357 * we see some SACKs or dupacks. It is split of "Open"
1358 * mainly to move some processing from fast path to slow one.
1359 * "CWR" CWND was reduced due to some Congestion Notification event.
1360 * It can be ECN, ICMP source quench, local device congestion.
1361 * "Recovery" CWND was reduced, we are fast-retransmitting.
1362 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
1364 * tcp_fastretrans_alert() is entered:
1365 * - each incoming ACK, if state is not "Open"
1366 * - when arrived ACK is unusual, namely:
1371 * Counting packets in flight is pretty simple.
1373 * in_flight = packets_out - left_out + retrans_out
1375 * packets_out is SND.NXT-SND.UNA counted in packets.
1377 * retrans_out is number of retransmitted segments.
1379 * left_out is number of segments left network, but not ACKed yet.
1381 * left_out = sacked_out + lost_out
1383 * sacked_out: Packets, which arrived to receiver out of order
1384 * and hence not ACKed. With SACKs this number is simply
1385 * amount of SACKed data. Even without SACKs
1386 * it is easy to give pretty reliable estimate of this number,
1387 * counting duplicate ACKs.
1389 * lost_out: Packets lost by network. TCP has no explicit
1390 * "loss notification" feedback from network (for now).
1391 * It means that this number can be only _guessed_.
1392 * Actually, it is the heuristics to predict lossage that
1393 * distinguishes different algorithms.
1395 * F.e. after RTO, when all the queue is considered as lost,
1396 * lost_out = packets_out and in_flight = retrans_out.
1398 * Essentially, we have now two algorithms counting
1401 * FACK: It is the simplest heuristics. As soon as we decided
1402 * that something is lost, we decide that _all_ not SACKed
1403 * packets until the most forward SACK are lost. I.e.
1404 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
1405 * It is absolutely correct estimate, if network does not reorder
1406 * packets. And it loses any connection to reality when reordering
1407 * takes place. We use FACK by default until reordering
1408 * is suspected on the path to this destination.
1410 * NewReno: when Recovery is entered, we assume that one segment
1411 * is lost (classic Reno). While we are in Recovery and
1412 * a partial ACK arrives, we assume that one more packet
1413 * is lost (NewReno). This heuristics are the same in NewReno
1416 * Imagine, that's all! Forget about all this shamanism about CWND inflation
1417 * deflation etc. CWND is real congestion window, never inflated, changes
1418 * only according to classic VJ rules.
1420 * Really tricky (and requiring careful tuning) part of algorithm
1421 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
1422 * The first determines the moment _when_ we should reduce CWND and,
1423 * hence, slow down forward transmission. In fact, it determines the moment
1424 * when we decide that hole is caused by loss, rather than by a reorder.
1426 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
1427 * holes, caused by lost packets.
1429 * And the most logically complicated part of algorithm is undo
1430 * heuristics. We detect false retransmits due to both too early
1431 * fast retransmit (reordering) and underestimated RTO, analyzing
1432 * timestamps and D-SACKs. When we detect that some segments were
1433 * retransmitted by mistake and CWND reduction was wrong, we undo
1434 * window reduction and abort recovery phase. This logic is hidden
1435 * inside several functions named tcp_try_undo_<something>.
1438 /* This function decides, when we should leave Disordered state
1439 * and enter Recovery phase, reducing congestion window.
1441 * Main question: may we further continue forward transmission
1442 * with the same cwnd?
1445 tcp_time_to_recover(struct sock *sk, struct tcp_opt *tp)
1447 /* Trick#1: The loss is proven. */
1451 /* Not-A-Trick#2 : Classic rule... */
1452 if (tcp_fackets_out(tp) > tp->reordering)
1455 /* Trick#3 : when we use RFC2988 timer restart, fast
1456 * retransmit can be triggered by timeout of queue head.
1458 if (tcp_head_timedout(sk, tp))
1461 /* Trick#4: It is still not OK... But will it be useful to delay
1464 if (tp->packets_out <= tp->reordering &&
1465 tp->sacked_out >= max_t(__u32, tp->packets_out/2, sysctl_tcp_reordering) &&
1466 !tcp_may_send_now(sk, tp)) {
1467 /* We have nothing to send. This connection is limited
1468 * either by receiver window or by application.
1476 /* If we receive more dupacks than we expected counting segments
1477 * in assumption of absent reordering, interpret this as reordering.
1478 * The only another reason could be bug in receiver TCP.
1480 static void tcp_check_reno_reordering(struct tcp_opt *tp, int addend)
1484 holes = max(tp->lost_out, 1U);
1485 holes = min(holes, tp->packets_out);
1487 if (tp->sacked_out + holes > tp->packets_out) {
1488 tp->sacked_out = tp->packets_out - holes;
1489 tcp_update_reordering(tp, tp->packets_out+addend, 0);
1493 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1495 static void tcp_add_reno_sack(struct tcp_opt *tp)
1498 tcp_check_reno_reordering(tp, 0);
1499 tcp_sync_left_out(tp);
1502 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1504 static void tcp_remove_reno_sacks(struct sock *sk, struct tcp_opt *tp, int acked)
1507 /* One ACK acked hole. The rest eat duplicate ACKs. */
1508 if (acked-1 >= tp->sacked_out)
1511 tp->sacked_out -= acked-1;
1513 tcp_check_reno_reordering(tp, acked);
1514 tcp_sync_left_out(tp);
1517 static inline void tcp_reset_reno_sack(struct tcp_opt *tp)
1520 tp->left_out = tp->lost_out;
1523 /* Mark head of queue up as lost. */
1525 tcp_mark_head_lost(struct sock *sk, struct tcp_opt *tp, int packets, u32 high_seq)
1527 struct sk_buff *skb;
1530 BUG_TRAP(cnt <= tp->packets_out);
1532 sk_stream_for_retrans_queue(skb, sk) {
1533 if (--cnt < 0 || after(TCP_SKB_CB(skb)->end_seq, high_seq))
1535 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
1536 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1540 tcp_sync_left_out(tp);
1543 /* Account newly detected lost packet(s) */
1545 static void tcp_update_scoreboard(struct sock *sk, struct tcp_opt *tp)
1548 int lost = tp->fackets_out - tp->reordering;
1551 tcp_mark_head_lost(sk, tp, lost, tp->high_seq);
1553 tcp_mark_head_lost(sk, tp, 1, tp->high_seq);
1556 /* New heuristics: it is possible only after we switched
1557 * to restart timer each time when something is ACKed.
1558 * Hence, we can detect timed out packets during fast
1559 * retransmit without falling to slow start.
1561 if (tcp_head_timedout(sk, tp)) {
1562 struct sk_buff *skb;
1564 sk_stream_for_retrans_queue(skb, sk) {
1565 if (tcp_skb_timedout(tp, skb) &&
1566 !(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
1567 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1571 tcp_sync_left_out(tp);
1575 /* CWND moderation, preventing bursts due to too big ACKs
1576 * in dubious situations.
1578 static __inline__ void tcp_moderate_cwnd(struct tcp_opt *tp)
1580 tp->snd_cwnd = min(tp->snd_cwnd,
1581 tcp_packets_in_flight(tp)+tcp_max_burst(tp));
1582 tp->snd_cwnd_stamp = tcp_time_stamp;
1585 /* Decrease cwnd each second ack. */
1587 static void tcp_cwnd_down(struct tcp_opt *tp)
1589 int decr = tp->snd_cwnd_cnt + 1;
1594 * Here limit is evaluated as BWestimation*RTTmin (for obtaining it
1595 * in packets we use mss_cache). If sysctl_tcp_westwood is off
1596 * tcp_westwood_bw_rttmin() returns 0. In such case snd_ssthresh is
1597 * still used as usual. It prevents other strange cases in which
1598 * BWE*RTTmin could assume value 0. It should not happen but...
1601 if (!(limit = tcp_westwood_bw_rttmin(tp)))
1602 limit = tp->snd_ssthresh/2;
1604 tp->snd_cwnd_cnt = decr&1;
1607 if (decr && tp->snd_cwnd > limit)
1608 tp->snd_cwnd -= decr;
1610 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp)+1);
1611 tp->snd_cwnd_stamp = tcp_time_stamp;
1614 /* Nothing was retransmitted or returned timestamp is less
1615 * than timestamp of the first retransmission.
1617 static __inline__ int tcp_packet_delayed(struct tcp_opt *tp)
1619 return !tp->retrans_stamp ||
1620 (tp->saw_tstamp && tp->rcv_tsecr &&
1621 (__s32)(tp->rcv_tsecr - tp->retrans_stamp) < 0);
1624 /* Undo procedures. */
1626 #if FASTRETRANS_DEBUG > 1
1627 static void DBGUNDO(struct sock *sk, struct tcp_opt *tp, const char *msg)
1629 struct inet_opt *inet = inet_sk(sk);
1630 printk(KERN_DEBUG "Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n",
1632 NIPQUAD(inet->daddr), ntohs(inet->dport),
1633 tp->snd_cwnd, tp->left_out,
1634 tp->snd_ssthresh, tp->prior_ssthresh, tp->packets_out);
1637 #define DBGUNDO(x...) do { } while (0)
1640 static void tcp_undo_cwr(struct tcp_opt *tp, int undo)
1642 if (tp->prior_ssthresh) {
1643 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh<<1);
1645 if (undo && tp->prior_ssthresh > tp->snd_ssthresh) {
1646 tp->snd_ssthresh = tp->prior_ssthresh;
1647 TCP_ECN_withdraw_cwr(tp);
1650 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
1652 tcp_moderate_cwnd(tp);
1653 tp->snd_cwnd_stamp = tcp_time_stamp;
1656 static inline int tcp_may_undo(struct tcp_opt *tp)
1658 return tp->undo_marker &&
1659 (!tp->undo_retrans || tcp_packet_delayed(tp));
1662 /* People celebrate: "We love our President!" */
1663 static int tcp_try_undo_recovery(struct sock *sk, struct tcp_opt *tp)
1665 if (tcp_may_undo(tp)) {
1666 /* Happy end! We did not retransmit anything
1667 * or our original transmission succeeded.
1669 DBGUNDO(sk, tp, tp->ca_state == TCP_CA_Loss ? "loss" : "retrans");
1670 tcp_undo_cwr(tp, 1);
1671 if (tp->ca_state == TCP_CA_Loss)
1672 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
1674 NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO);
1675 tp->undo_marker = 0;
1677 if (tp->snd_una == tp->high_seq && IsReno(tp)) {
1678 /* Hold old state until something *above* high_seq
1679 * is ACKed. For Reno it is MUST to prevent false
1680 * fast retransmits (RFC2582). SACK TCP is safe. */
1681 tcp_moderate_cwnd(tp);
1684 tcp_set_ca_state(tp, TCP_CA_Open);
1688 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
1689 static void tcp_try_undo_dsack(struct sock *sk, struct tcp_opt *tp)
1691 if (tp->undo_marker && !tp->undo_retrans) {
1692 DBGUNDO(sk, tp, "D-SACK");
1693 tcp_undo_cwr(tp, 1);
1694 tp->undo_marker = 0;
1695 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO);
1699 /* Undo during fast recovery after partial ACK. */
1701 static int tcp_try_undo_partial(struct sock *sk, struct tcp_opt *tp, int acked)
1703 /* Partial ACK arrived. Force Hoe's retransmit. */
1704 int failed = IsReno(tp) || tp->fackets_out>tp->reordering;
1706 if (tcp_may_undo(tp)) {
1707 /* Plain luck! Hole if filled with delayed
1708 * packet, rather than with a retransmit.
1710 if (tp->retrans_out == 0)
1711 tp->retrans_stamp = 0;
1713 tcp_update_reordering(tp, tcp_fackets_out(tp)+acked, 1);
1715 DBGUNDO(sk, tp, "Hoe");
1716 tcp_undo_cwr(tp, 0);
1717 NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO);
1719 /* So... Do not make Hoe's retransmit yet.
1720 * If the first packet was delayed, the rest
1721 * ones are most probably delayed as well.
1728 /* Undo during loss recovery after partial ACK. */
1729 static int tcp_try_undo_loss(struct sock *sk, struct tcp_opt *tp)
1731 if (tcp_may_undo(tp)) {
1732 struct sk_buff *skb;
1733 sk_stream_for_retrans_queue(skb, sk) {
1734 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1736 DBGUNDO(sk, tp, "partial loss");
1738 tp->left_out = tp->sacked_out;
1739 tcp_undo_cwr(tp, 1);
1740 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
1741 tp->retransmits = 0;
1742 tp->undo_marker = 0;
1744 tcp_set_ca_state(tp, TCP_CA_Open);
1750 static __inline__ void tcp_complete_cwr(struct tcp_opt *tp)
1752 if (tcp_westwood_cwnd(tp))
1753 tp->snd_ssthresh = tp->snd_cwnd;
1755 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
1756 tp->snd_cwnd_stamp = tcp_time_stamp;
1759 static void tcp_try_to_open(struct sock *sk, struct tcp_opt *tp, int flag)
1761 tp->left_out = tp->sacked_out;
1763 if (tp->retrans_out == 0)
1764 tp->retrans_stamp = 0;
1769 if (tp->ca_state != TCP_CA_CWR) {
1770 int state = TCP_CA_Open;
1775 state = TCP_CA_Disorder;
1777 if (tp->ca_state != state) {
1778 tcp_set_ca_state(tp, state);
1779 tp->high_seq = tp->snd_nxt;
1781 tcp_moderate_cwnd(tp);
1787 /* Process an event, which can update packets-in-flight not trivially.
1788 * Main goal of this function is to calculate new estimate for left_out,
1789 * taking into account both packets sitting in receiver's buffer and
1790 * packets lost by network.
1792 * Besides that it does CWND reduction, when packet loss is detected
1793 * and changes state of machine.
1795 * It does _not_ decide what to send, it is made in function
1796 * tcp_xmit_retransmit_queue().
1799 tcp_fastretrans_alert(struct sock *sk, u32 prior_snd_una,
1800 int prior_packets, int flag)
1802 struct tcp_opt *tp = tcp_sk(sk);
1803 int is_dupack = (tp->snd_una == prior_snd_una && !(flag&FLAG_NOT_DUP));
1805 /* Some technical things:
1806 * 1. Reno does not count dupacks (sacked_out) automatically. */
1807 if (!tp->packets_out)
1809 /* 2. SACK counts snd_fack in packets inaccurately. */
1810 if (tp->sacked_out == 0)
1811 tp->fackets_out = 0;
1813 /* Now state machine starts.
1814 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
1816 tp->prior_ssthresh = 0;
1818 /* B. In all the states check for reneging SACKs. */
1819 if (tp->sacked_out && tcp_check_sack_reneging(sk, tp))
1822 /* C. Process data loss notification, provided it is valid. */
1823 if ((flag&FLAG_DATA_LOST) &&
1824 before(tp->snd_una, tp->high_seq) &&
1825 tp->ca_state != TCP_CA_Open &&
1826 tp->fackets_out > tp->reordering) {
1827 tcp_mark_head_lost(sk, tp, tp->fackets_out-tp->reordering, tp->high_seq);
1828 NET_INC_STATS_BH(LINUX_MIB_TCPLOSS);
1831 /* D. Synchronize left_out to current state. */
1832 tcp_sync_left_out(tp);
1834 /* E. Check state exit conditions. State can be terminated
1835 * when high_seq is ACKed. */
1836 if (tp->ca_state == TCP_CA_Open) {
1837 if (!sysctl_tcp_frto)
1838 BUG_TRAP(tp->retrans_out == 0);
1839 tp->retrans_stamp = 0;
1840 } else if (!before(tp->snd_una, tp->high_seq)) {
1841 switch (tp->ca_state) {
1843 tp->retransmits = 0;
1844 if (tcp_try_undo_recovery(sk, tp))
1849 /* CWR is to be held something *above* high_seq
1850 * is ACKed for CWR bit to reach receiver. */
1851 if (tp->snd_una != tp->high_seq) {
1852 tcp_complete_cwr(tp);
1853 tcp_set_ca_state(tp, TCP_CA_Open);
1857 case TCP_CA_Disorder:
1858 tcp_try_undo_dsack(sk, tp);
1859 if (!tp->undo_marker ||
1860 /* For SACK case do not Open to allow to undo
1861 * catching for all duplicate ACKs. */
1862 IsReno(tp) || tp->snd_una != tp->high_seq) {
1863 tp->undo_marker = 0;
1864 tcp_set_ca_state(tp, TCP_CA_Open);
1868 case TCP_CA_Recovery:
1870 tcp_reset_reno_sack(tp);
1871 if (tcp_try_undo_recovery(sk, tp))
1873 tcp_complete_cwr(tp);
1878 /* F. Process state. */
1879 switch (tp->ca_state) {
1880 case TCP_CA_Recovery:
1881 if (prior_snd_una == tp->snd_una) {
1882 if (IsReno(tp) && is_dupack)
1883 tcp_add_reno_sack(tp);
1885 int acked = prior_packets - tp->packets_out;
1887 tcp_remove_reno_sacks(sk, tp, acked);
1888 is_dupack = tcp_try_undo_partial(sk, tp, acked);
1892 if (flag&FLAG_DATA_ACKED)
1893 tp->retransmits = 0;
1894 if (!tcp_try_undo_loss(sk, tp)) {
1895 tcp_moderate_cwnd(tp);
1896 tcp_xmit_retransmit_queue(sk);
1899 if (tp->ca_state != TCP_CA_Open)
1901 /* Loss is undone; fall through to processing in Open state. */
1904 if (tp->snd_una != prior_snd_una)
1905 tcp_reset_reno_sack(tp);
1907 tcp_add_reno_sack(tp);
1910 if (tp->ca_state == TCP_CA_Disorder)
1911 tcp_try_undo_dsack(sk, tp);
1913 if (!tcp_time_to_recover(sk, tp)) {
1914 tcp_try_to_open(sk, tp, flag);
1918 /* Otherwise enter Recovery state */
1921 NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY);
1923 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY);
1925 tp->high_seq = tp->snd_nxt;
1926 tp->prior_ssthresh = 0;
1927 tp->undo_marker = tp->snd_una;
1928 tp->undo_retrans = tp->retrans_out;
1930 if (tp->ca_state < TCP_CA_CWR) {
1931 if (!(flag&FLAG_ECE))
1932 tp->prior_ssthresh = tcp_current_ssthresh(tp);
1933 tp->snd_ssthresh = tcp_recalc_ssthresh(tp);
1934 TCP_ECN_queue_cwr(tp);
1937 tp->snd_cwnd_cnt = 0;
1938 tcp_set_ca_state(tp, TCP_CA_Recovery);
1941 if (is_dupack || tcp_head_timedout(sk, tp))
1942 tcp_update_scoreboard(sk, tp);
1944 tcp_xmit_retransmit_queue(sk);
1947 /* Read draft-ietf-tcplw-high-performance before mucking
1948 * with this code. (Superceeds RFC1323)
1950 static void tcp_ack_saw_tstamp(struct tcp_opt *tp, int flag)
1954 /* RTTM Rule: A TSecr value received in a segment is used to
1955 * update the averaged RTT measurement only if the segment
1956 * acknowledges some new data, i.e., only if it advances the
1957 * left edge of the send window.
1959 * See draft-ietf-tcplw-high-performance-00, section 3.3.
1960 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
1962 * Changed: reset backoff as soon as we see the first valid sample.
1963 * If we do not, we get strongly overstimated rto. With timestamps
1964 * samples are accepted even from very old segments: f.e., when rtt=1
1965 * increases to 8, we retransmit 5 times and after 8 seconds delayed
1966 * answer arrives rto becomes 120 seconds! If at least one of segments
1967 * in window is lost... Voila. --ANK (010210)
1969 seq_rtt = tcp_time_stamp - tp->rcv_tsecr;
1970 tcp_rtt_estimator(tp, seq_rtt);
1976 static void tcp_ack_no_tstamp(struct tcp_opt *tp, u32 seq_rtt, int flag)
1978 /* We don't have a timestamp. Can only use
1979 * packets that are not retransmitted to determine
1980 * rtt estimates. Also, we must not reset the
1981 * backoff for rto until we get a non-retransmitted
1982 * packet. This allows us to deal with a situation
1983 * where the network delay has increased suddenly.
1984 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
1987 if (flag & FLAG_RETRANS_DATA_ACKED)
1990 tcp_rtt_estimator(tp, seq_rtt);
1996 static __inline__ void
1997 tcp_ack_update_rtt(struct tcp_opt *tp, int flag, s32 seq_rtt)
1999 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
2000 if (tp->saw_tstamp && tp->rcv_tsecr)
2001 tcp_ack_saw_tstamp(tp, flag);
2002 else if (seq_rtt >= 0)
2003 tcp_ack_no_tstamp(tp, seq_rtt, flag);
2007 * Compute congestion window to use.
2009 * This is from the implementation of BICTCP in
2010 * Lison-Xu, Kahaled Harfoush, and Injog Rhee.
2011 * "Binary Increase Congestion Control for Fast, Long Distance
2012 * Networks" in InfoComm 2004
2014 * http://www.csc.ncsu.edu/faculty/rhee/export/bitcp.pdf
2016 * Unless BIC is enabled and congestion window is large
2017 * this behaves the same as the original Reno.
2019 static inline __u32 bictcp_cwnd(struct tcp_opt *tp)
2021 /* orignal Reno behaviour */
2022 if (!sysctl_tcp_bic)
2023 return tp->snd_cwnd;
2025 if (tp->bictcp.last_cwnd == tp->snd_cwnd &&
2026 (s32)(tcp_time_stamp - tp->bictcp.last_stamp) <= (HZ>>5))
2027 return tp->bictcp.cnt;
2029 tp->bictcp.last_cwnd = tp->snd_cwnd;
2030 tp->bictcp.last_stamp = tcp_time_stamp;
2032 /* start off normal */
2033 if (tp->snd_cwnd <= sysctl_tcp_bic_low_window)
2034 tp->bictcp.cnt = tp->snd_cwnd;
2036 /* binary increase */
2037 else if (tp->snd_cwnd < tp->bictcp.last_max_cwnd) {
2038 __u32 dist = (tp->bictcp.last_max_cwnd - tp->snd_cwnd)
2041 if (dist > BICTCP_MAX_INCREMENT)
2042 /* linear increase */
2043 tp->bictcp.cnt = tp->snd_cwnd / BICTCP_MAX_INCREMENT;
2044 else if (dist <= 1U)
2045 /* binary search increase */
2046 tp->bictcp.cnt = tp->snd_cwnd * BICTCP_FUNC_OF_MIN_INCR
2049 /* binary search increase */
2050 tp->bictcp.cnt = tp->snd_cwnd / dist;
2052 /* slow start amd linear increase */
2053 if (tp->snd_cwnd < tp->bictcp.last_max_cwnd + BICTCP_B)
2055 tp->bictcp.cnt = tp->snd_cwnd * BICTCP_FUNC_OF_MIN_INCR
2057 else if (tp->snd_cwnd < tp->bictcp.last_max_cwnd
2058 + BICTCP_MAX_INCREMENT*(BICTCP_B-1))
2060 tp->bictcp.cnt = tp->snd_cwnd * (BICTCP_B-1)
2061 / (tp->snd_cwnd-tp->bictcp.last_max_cwnd);
2063 /* linear increase */
2064 tp->bictcp.cnt = tp->snd_cwnd / BICTCP_MAX_INCREMENT;
2066 return tp->bictcp.cnt;
2069 /* This is Jacobson's slow start and congestion avoidance.
2070 * SIGCOMM '88, p. 328.
2072 static __inline__ void reno_cong_avoid(struct tcp_opt *tp)
2074 if (tp->snd_cwnd <= tp->snd_ssthresh) {
2075 /* In "safe" area, increase. */
2076 if (tp->snd_cwnd < tp->snd_cwnd_clamp)
2079 /* In dangerous area, increase slowly.
2080 * In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd
2082 if (tp->snd_cwnd_cnt >= bictcp_cwnd(tp)) {
2083 if (tp->snd_cwnd < tp->snd_cwnd_clamp)
2089 tp->snd_cwnd_stamp = tcp_time_stamp;
2092 /* This is based on the congestion detection/avoidance scheme described in
2093 * Lawrence S. Brakmo and Larry L. Peterson.
2094 * "TCP Vegas: End to end congestion avoidance on a global internet."
2095 * IEEE Journal on Selected Areas in Communication, 13(8):1465--1480,
2096 * October 1995. Available from:
2097 * ftp://ftp.cs.arizona.edu/xkernel/Papers/jsac.ps
2099 * See http://www.cs.arizona.edu/xkernel/ for their implementation.
2100 * The main aspects that distinguish this implementation from the
2101 * Arizona Vegas implementation are:
2102 * o We do not change the loss detection or recovery mechanisms of
2103 * Linux in any way. Linux already recovers from losses quite well,
2104 * using fine-grained timers, NewReno, and FACK.
2105 * o To avoid the performance penalty imposed by increasing cwnd
2106 * only every-other RTT during slow start, we increase during
2107 * every RTT during slow start, just like Reno.
2108 * o Largely to allow continuous cwnd growth during slow start,
2109 * we use the rate at which ACKs come back as the "actual"
2110 * rate, rather than the rate at which data is sent.
2111 * o To speed convergence to the right rate, we set the cwnd
2112 * to achieve the right ("actual") rate when we exit slow start.
2113 * o To filter out the noise caused by delayed ACKs, we use the
2114 * minimum RTT sample observed during the last RTT to calculate
2116 * o When the sender re-starts from idle, it waits until it has
2117 * received ACKs for an entire flight of new data before making
2118 * a cwnd adjustment decision. The original Vegas implementation
2119 * assumed senders never went idle.
2121 static void vegas_cong_avoid(struct tcp_opt *tp, u32 ack, u32 seq_rtt)
2123 /* The key players are v_beg_snd_una and v_beg_snd_nxt.
2125 * These are so named because they represent the approximate values
2126 * of snd_una and snd_nxt at the beginning of the current RTT. More
2127 * precisely, they represent the amount of data sent during the RTT.
2128 * At the end of the RTT, when we receive an ACK for v_beg_snd_nxt,
2129 * we will calculate that (v_beg_snd_nxt - v_beg_snd_una) outstanding
2130 * bytes of data have been ACKed during the course of the RTT, giving
2131 * an "actual" rate of:
2133 * (v_beg_snd_nxt - v_beg_snd_una) / (rtt duration)
2135 * Unfortunately, v_beg_snd_una is not exactly equal to snd_una,
2136 * because delayed ACKs can cover more than one segment, so they
2137 * don't line up nicely with the boundaries of RTTs.
2139 * Another unfortunate fact of life is that delayed ACKs delay the
2140 * advance of the left edge of our send window, so that the number
2141 * of bytes we send in an RTT is often less than our cwnd will allow.
2142 * So we keep track of our cwnd separately, in v_beg_snd_cwnd.
2145 if (after(ack, tp->vegas.beg_snd_nxt)) {
2146 /* Do the Vegas once-per-RTT cwnd adjustment. */
2147 u32 old_wnd, old_snd_cwnd;
2150 /* Here old_wnd is essentially the window of data that was
2151 * sent during the previous RTT, and has all
2152 * been acknowledged in the course of the RTT that ended
2153 * with the ACK we just received. Likewise, old_snd_cwnd
2154 * is the cwnd during the previous RTT.
2156 old_wnd = (tp->vegas.beg_snd_nxt - tp->vegas.beg_snd_una) /
2158 old_snd_cwnd = tp->vegas.beg_snd_cwnd;
2160 /* Save the extent of the current window so we can use this
2161 * at the end of the next RTT.
2163 tp->vegas.beg_snd_una = tp->vegas.beg_snd_nxt;
2164 tp->vegas.beg_snd_nxt = tp->snd_nxt;
2165 tp->vegas.beg_snd_cwnd = tp->snd_cwnd;
2167 /* Take into account the current RTT sample too, to
2168 * decrease the impact of delayed acks. This double counts
2169 * this sample since we count it for the next window as well,
2170 * but that's not too awful, since we're taking the min,
2171 * rather than averaging.
2173 vegas_rtt_calc(tp, seq_rtt);
2175 /* We do the Vegas calculations only if we got enough RTT
2176 * samples that we can be reasonably sure that we got
2177 * at least one RTT sample that wasn't from a delayed ACK.
2178 * If we only had 2 samples total,
2179 * then that means we're getting only 1 ACK per RTT, which
2180 * means they're almost certainly delayed ACKs.
2181 * If we have 3 samples, we should be OK.
2184 if (tp->vegas.cntRTT <= 2) {
2185 /* We don't have enough RTT samples to do the Vegas
2186 * calculation, so we'll behave like Reno.
2188 if (tp->snd_cwnd > tp->snd_ssthresh)
2191 u32 rtt, target_cwnd, diff;
2193 /* We have enough RTT samples, so, using the Vegas
2194 * algorithm, we determine if we should increase or
2195 * decrease cwnd, and by how much.
2198 /* Pluck out the RTT we are using for the Vegas
2199 * calculations. This is the min RTT seen during the
2200 * last RTT. Taking the min filters out the effects
2201 * of delayed ACKs, at the cost of noticing congestion
2204 rtt = tp->vegas.minRTT;
2206 /* Calculate the cwnd we should have, if we weren't
2210 * (actual rate in segments) * baseRTT
2211 * We keep it as a fixed point number with
2212 * V_PARAM_SHIFT bits to the right of the binary point.
2214 target_cwnd = ((old_wnd * tp->vegas.baseRTT)
2215 << V_PARAM_SHIFT) / rtt;
2217 /* Calculate the difference between the window we had,
2218 * and the window we would like to have. This quantity
2219 * is the "Diff" from the Arizona Vegas papers.
2221 * Again, this is a fixed point number with
2222 * V_PARAM_SHIFT bits to the right of the binary
2225 diff = (old_wnd << V_PARAM_SHIFT) - target_cwnd;
2227 if (tp->snd_cwnd < tp->snd_ssthresh) {
2229 if (diff > sysctl_tcp_vegas_gamma) {
2230 /* Going too fast. Time to slow down
2231 * and switch to congestion avoidance.
2233 tp->snd_ssthresh = 2;
2235 /* Set cwnd to match the actual rate
2237 * cwnd = (actual rate) * baseRTT
2238 * Then we add 1 because the integer
2239 * truncation robs us of full link
2242 tp->snd_cwnd = min(tp->snd_cwnd,
2248 /* Congestion avoidance. */
2251 /* Figure out where we would like cwnd
2254 if (diff > sysctl_tcp_vegas_beta) {
2255 /* The old window was too fast, so
2258 next_snd_cwnd = old_snd_cwnd - 1;
2259 } else if (diff < sysctl_tcp_vegas_alpha) {
2260 /* We don't have enough extra packets
2261 * in the network, so speed up.
2263 next_snd_cwnd = old_snd_cwnd + 1;
2265 /* Sending just as fast as we
2268 next_snd_cwnd = old_snd_cwnd;
2271 /* Adjust cwnd upward or downward, toward the
2274 if (next_snd_cwnd > tp->snd_cwnd)
2276 else if (next_snd_cwnd < tp->snd_cwnd)
2281 /* Wipe the slate clean for the next RTT. */
2282 tp->vegas.cntRTT = 0;
2283 tp->vegas.minRTT = 0x7fffffff;
2286 /* The following code is executed for every ack we receive,
2287 * except for conditions checked in should_advance_cwnd()
2288 * before the call to tcp_cong_avoid(). Mainly this means that
2289 * we only execute this code if the ack actually acked some
2293 /* If we are in slow start, increase our cwnd in response to this ACK.
2294 * (If we are not in slow start then we are in congestion avoidance,
2295 * and adjust our congestion window only once per RTT. See the code
2298 if (tp->snd_cwnd <= tp->snd_ssthresh)
2301 /* to keep cwnd from growing without bound */
2302 tp->snd_cwnd = min_t(u32, tp->snd_cwnd, tp->snd_cwnd_clamp);
2304 /* Make sure that we are never so timid as to reduce our cwnd below
2307 * Going below 2 MSS would risk huge delayed ACKs from our receiver.
2309 tp->snd_cwnd = max(tp->snd_cwnd, 2U);
2311 tp->snd_cwnd_stamp = tcp_time_stamp;
2314 static inline void tcp_cong_avoid(struct tcp_opt *tp, u32 ack, u32 seq_rtt)
2316 if (tcp_vegas_enabled(tp))
2317 vegas_cong_avoid(tp, ack, seq_rtt);
2319 reno_cong_avoid(tp);
2322 /* Restart timer after forward progress on connection.
2323 * RFC2988 recommends to restart timer to now+rto.
2326 static __inline__ void tcp_ack_packets_out(struct sock *sk, struct tcp_opt *tp)
2328 if (tp->packets_out==0) {
2329 tcp_clear_xmit_timer(sk, TCP_TIME_RETRANS);
2331 tcp_reset_xmit_timer(sk, TCP_TIME_RETRANS, tp->rto);
2335 /* Remove acknowledged frames from the retransmission queue. */
2336 static int tcp_clean_rtx_queue(struct sock *sk, __s32 *seq_rtt_p)
2338 struct tcp_opt *tp = tcp_sk(sk);
2339 struct sk_buff *skb;
2340 __u32 now = tcp_time_stamp;
2344 while ((skb = skb_peek(&sk->sk_write_queue)) && skb != sk->sk_send_head) {
2345 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
2346 __u8 sacked = scb->sacked;
2348 /* If our packet is before the ack sequence we can
2349 * discard it as it's confirmed to have arrived at
2352 if (after(scb->end_seq, tp->snd_una))
2355 /* Initial outgoing SYN's get put onto the write_queue
2356 * just like anything else we transmit. It is not
2357 * true data, and if we misinform our callers that
2358 * this ACK acks real data, we will erroneously exit
2359 * connection startup slow start one packet too
2360 * quickly. This is severely frowned upon behavior.
2362 if(!(scb->flags & TCPCB_FLAG_SYN)) {
2363 acked |= FLAG_DATA_ACKED;
2365 acked |= FLAG_SYN_ACKED;
2366 tp->retrans_stamp = 0;
2370 if(sacked & TCPCB_RETRANS) {
2371 if(sacked & TCPCB_SACKED_RETRANS)
2373 acked |= FLAG_RETRANS_DATA_ACKED;
2375 } else if (seq_rtt < 0)
2376 seq_rtt = now - scb->when;
2377 if(sacked & TCPCB_SACKED_ACKED)
2379 if(sacked & TCPCB_LOST)
2381 if(sacked & TCPCB_URG) {
2383 !before(scb->end_seq, tp->snd_up))
2386 } else if (seq_rtt < 0)
2387 seq_rtt = now - scb->when;
2388 if (tp->fackets_out)
2391 __skb_unlink(skb, skb->list);
2392 sk_stream_free_skb(sk, skb);
2395 if (acked&FLAG_ACKED) {
2396 tcp_ack_update_rtt(tp, acked, seq_rtt);
2397 tcp_ack_packets_out(sk, tp);
2400 #if FASTRETRANS_DEBUG > 0
2401 BUG_TRAP((int)tp->sacked_out >= 0);
2402 BUG_TRAP((int)tp->lost_out >= 0);
2403 BUG_TRAP((int)tp->retrans_out >= 0);
2404 if (!tp->packets_out && tp->sack_ok) {
2406 printk(KERN_DEBUG "Leak l=%u %d\n", tp->lost_out,
2410 if (tp->sacked_out) {
2411 printk(KERN_DEBUG "Leak s=%u %d\n", tp->sacked_out,
2415 if (tp->retrans_out) {
2416 printk(KERN_DEBUG "Leak r=%u %d\n", tp->retrans_out,
2418 tp->retrans_out = 0;
2422 *seq_rtt_p = seq_rtt;
2426 static void tcp_ack_probe(struct sock *sk)
2428 struct tcp_opt *tp = tcp_sk(sk);
2430 /* Was it a usable window open? */
2432 if (!after(TCP_SKB_CB(sk->sk_send_head)->end_seq,
2433 tp->snd_una + tp->snd_wnd)) {
2435 tcp_clear_xmit_timer(sk, TCP_TIME_PROBE0);
2436 /* Socket must be waked up by subsequent tcp_data_snd_check().
2437 * This function is not for random using!
2440 tcp_reset_xmit_timer(sk, TCP_TIME_PROBE0,
2441 min(tp->rto << tp->backoff, TCP_RTO_MAX));
2445 static __inline__ int tcp_ack_is_dubious(struct tcp_opt *tp, int flag)
2447 return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
2448 tp->ca_state != TCP_CA_Open);
2451 static __inline__ int tcp_may_raise_cwnd(struct tcp_opt *tp, int flag)
2453 return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
2454 !((1<<tp->ca_state)&(TCPF_CA_Recovery|TCPF_CA_CWR));
2457 /* Check that window update is acceptable.
2458 * The function assumes that snd_una<=ack<=snd_next.
2460 static __inline__ int
2461 tcp_may_update_window(struct tcp_opt *tp, u32 ack, u32 ack_seq, u32 nwin)
2463 return (after(ack, tp->snd_una) ||
2464 after(ack_seq, tp->snd_wl1) ||
2465 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd));
2468 /* Update our send window.
2470 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
2471 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
2473 static int tcp_ack_update_window(struct sock *sk, struct tcp_opt *tp,
2474 struct sk_buff *skb, u32 ack, u32 ack_seq)
2477 u32 nwin = ntohs(skb->h.th->window);
2479 if (likely(!skb->h.th->syn))
2480 nwin <<= tp->snd_wscale;
2482 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
2483 flag |= FLAG_WIN_UPDATE;
2484 tcp_update_wl(tp, ack, ack_seq);
2486 if (tp->snd_wnd != nwin) {
2489 /* Note, it is the only place, where
2490 * fast path is recovered for sending TCP.
2492 tcp_fast_path_check(sk, tp);
2494 if (nwin > tp->max_window) {
2495 tp->max_window = nwin;
2496 tcp_sync_mss(sk, tp->pmtu_cookie);
2506 static void tcp_process_frto(struct sock *sk, u32 prior_snd_una)
2508 struct tcp_opt *tp = tcp_sk(sk);
2510 tcp_sync_left_out(tp);
2512 if (tp->snd_una == prior_snd_una ||
2513 !before(tp->snd_una, tp->frto_highmark)) {
2514 /* RTO was caused by loss, start retransmitting in
2515 * go-back-N slow start
2517 tcp_enter_frto_loss(sk);
2521 if (tp->frto_counter == 1) {
2522 /* First ACK after RTO advances the window: allow two new
2525 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
2527 /* Also the second ACK after RTO advances the window.
2528 * The RTO was likely spurious. Reduce cwnd and continue
2529 * in congestion avoidance
2531 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2532 tcp_moderate_cwnd(tp);
2535 /* F-RTO affects on two new ACKs following RTO.
2536 * At latest on third ACK the TCP behavor is back to normal.
2538 tp->frto_counter = (tp->frto_counter + 1) % 3;
2547 * This function initializes fields used in TCP Westwood+. We can't
2548 * get no information about RTTmin at this time so we simply set it to
2549 * TCP_WESTWOOD_INIT_RTT. This value was chosen to be too conservative
2550 * since in this way we're sure it will be updated in a consistent
2551 * way as soon as possible. It will reasonably happen within the first
2552 * RTT period of the connection lifetime.
2555 static void init_westwood(struct sock *sk)
2557 struct tcp_opt *tp = tcp_sk(sk);
2559 tp->westwood.bw_ns_est = 0;
2560 tp->westwood.bw_est = 0;
2561 tp->westwood.accounted = 0;
2562 tp->westwood.cumul_ack = 0;
2563 tp->westwood.rtt_win_sx = tcp_time_stamp;
2564 tp->westwood.rtt = TCP_WESTWOOD_INIT_RTT;
2565 tp->westwood.rtt_min = TCP_WESTWOOD_INIT_RTT;
2566 tp->westwood.snd_una = tp->snd_una;
2570 * @westwood_do_filter
2571 * Low-pass filter. Implemented using constant coeffients.
2574 static inline __u32 westwood_do_filter(__u32 a, __u32 b)
2576 return (((7 * a) + b) >> 3);
2579 static void westwood_filter(struct sock *sk, __u32 delta)
2581 struct tcp_opt *tp = tcp_sk(sk);
2583 tp->westwood.bw_ns_est =
2584 westwood_do_filter(tp->westwood.bw_ns_est,
2585 tp->westwood.bk / delta);
2586 tp->westwood.bw_est =
2587 westwood_do_filter(tp->westwood.bw_est,
2588 tp->westwood.bw_ns_est);
2592 * @westwood_update_rttmin
2593 * It is used to update RTTmin. In this case we MUST NOT use
2594 * WESTWOOD_RTT_MIN minimum bound since we could be on a LAN!
2597 static inline __u32 westwood_update_rttmin(struct sock *sk)
2599 struct tcp_opt *tp = tcp_sk(sk);
2600 __u32 rttmin = tp->westwood.rtt_min;
2602 if (tp->westwood.rtt == 0)
2605 if (tp->westwood.rtt < tp->westwood.rtt_min || !rttmin)
2606 rttmin = tp->westwood.rtt;
2613 * Evaluate increases for dk.
2616 static inline __u32 westwood_acked(struct sock *sk)
2618 struct tcp_opt *tp = tcp_sk(sk);
2620 return ((tp->snd_una) - (tp->westwood.snd_una));
2624 * @westwood_new_window
2625 * It evaluates if we are receiving data inside the same RTT window as
2628 * It returns 0 if we are still evaluating samples in the same RTT
2629 * window, 1 if the sample has to be considered in the next window.
2632 static int westwood_new_window(struct sock *sk)
2634 struct tcp_opt *tp = tcp_sk(sk);
2639 left_bound = tp->westwood.rtt_win_sx;
2640 rtt = max(tp->westwood.rtt, (u32) TCP_WESTWOOD_RTT_MIN);
2643 * A RTT-window has passed. Be careful since if RTT is less than
2644 * 50ms we don't filter but we continue 'building the sample'.
2645 * This minimum limit was choosen since an estimation on small
2646 * time intervals is better to avoid...
2647 * Obvioulsy on a LAN we reasonably will always have
2648 * right_bound = left_bound + WESTWOOD_RTT_MIN
2651 if ((left_bound + rtt) < tcp_time_stamp)
2658 * @westwood_update_window
2659 * It updates RTT evaluation window if it is the right moment to do
2660 * it. If so it calls filter for evaluating bandwidth.
2663 static void __westwood_update_window(struct sock *sk, __u32 now)
2665 struct tcp_opt *tp = tcp_sk(sk);
2666 __u32 delta = now - tp->westwood.rtt_win_sx;
2671 if (tp->westwood.rtt)
2672 westwood_filter(sk, delta);
2674 tp->westwood.bk = 0;
2675 tp->westwood.rtt_win_sx = tcp_time_stamp;
2679 static void westwood_update_window(struct sock *sk, __u32 now)
2681 if (westwood_new_window(sk))
2682 __westwood_update_window(sk, now);
2686 * @__tcp_westwood_fast_bw
2687 * It is called when we are in fast path. In particular it is called when
2688 * header prediction is successfull. In such case infact update is
2689 * straight forward and doesn't need any particular care.
2692 void __tcp_westwood_fast_bw(struct sock *sk, struct sk_buff *skb)
2694 struct tcp_opt *tp = tcp_sk(sk);
2696 westwood_update_window(sk, tcp_time_stamp);
2698 tp->westwood.bk += westwood_acked(sk);
2699 tp->westwood.snd_una = tp->snd_una;
2700 tp->westwood.rtt_min = westwood_update_rttmin(sk);
2705 * @westwood_dupack_update
2706 * It updates accounted and cumul_ack when receiving a dupack.
2709 static void westwood_dupack_update(struct sock *sk)
2711 struct tcp_opt *tp = tcp_sk(sk);
2713 tp->westwood.accounted += tp->mss_cache;
2714 tp->westwood.cumul_ack = tp->mss_cache;
2717 static inline int westwood_may_change_cumul(struct tcp_opt *tp)
2719 return ((tp->westwood.cumul_ack) > tp->mss_cache);
2722 static inline void westwood_partial_update(struct tcp_opt *tp)
2724 tp->westwood.accounted -= tp->westwood.cumul_ack;
2725 tp->westwood.cumul_ack = tp->mss_cache;
2728 static inline void westwood_complete_update(struct tcp_opt *tp)
2730 tp->westwood.cumul_ack -= tp->westwood.accounted;
2731 tp->westwood.accounted = 0;
2735 * @westwood_acked_count
2736 * This function evaluates cumul_ack for evaluating dk in case of
2737 * delayed or partial acks.
2740 static __u32 westwood_acked_count(struct sock *sk)
2742 struct tcp_opt *tp = tcp_sk(sk);
2744 tp->westwood.cumul_ack = westwood_acked(sk);
2746 /* If cumul_ack is 0 this is a dupack since it's not moving
2749 if (!(tp->westwood.cumul_ack))
2750 westwood_dupack_update(sk);
2752 if (westwood_may_change_cumul(tp)) {
2753 /* Partial or delayed ack */
2754 if ((tp->westwood.accounted) >= (tp->westwood.cumul_ack))
2755 westwood_partial_update(tp);
2757 westwood_complete_update(tp);
2760 tp->westwood.snd_una = tp->snd_una;
2762 return tp->westwood.cumul_ack;
2767 * @__tcp_westwood_slow_bw
2768 * It is called when something is going wrong..even if there could
2769 * be no problems! Infact a simple delayed packet may trigger a
2770 * dupack. But we need to be careful in such case.
2773 void __tcp_westwood_slow_bw(struct sock *sk, struct sk_buff *skb)
2775 struct tcp_opt *tp = tcp_sk(sk);
2777 westwood_update_window(sk, tcp_time_stamp);
2779 tp->westwood.bk += westwood_acked_count(sk);
2780 tp->westwood.rtt_min = westwood_update_rttmin(sk);
2783 /* This routine deals with incoming acks, but not outgoing ones. */
2784 static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag)
2786 struct tcp_opt *tp = tcp_sk(sk);
2787 u32 prior_snd_una = tp->snd_una;
2788 u32 ack_seq = TCP_SKB_CB(skb)->seq;
2789 u32 ack = TCP_SKB_CB(skb)->ack_seq;
2790 u32 prior_in_flight;
2794 /* If the ack is newer than sent or older than previous acks
2795 * then we can probably ignore it.
2797 if (after(ack, tp->snd_nxt))
2798 goto uninteresting_ack;
2800 if (before(ack, prior_snd_una))
2803 if (!(flag&FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
2804 /* Window is constant, pure forward advance.
2805 * No more checks are required.
2806 * Note, we use the fact that SND.UNA>=SND.WL2.
2808 tcp_update_wl(tp, ack, ack_seq);
2810 tcp_westwood_fast_bw(sk, skb);
2811 flag |= FLAG_WIN_UPDATE;
2813 NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS);
2815 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
2818 NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS);
2820 flag |= tcp_ack_update_window(sk, tp, skb, ack, ack_seq);
2822 if (TCP_SKB_CB(skb)->sacked)
2823 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
2825 if (TCP_ECN_rcv_ecn_echo(tp, skb->h.th))
2828 tcp_westwood_slow_bw(sk,skb);
2831 /* We passed data and got it acked, remove any soft error
2832 * log. Something worked...
2834 sk->sk_err_soft = 0;
2835 tp->rcv_tstamp = tcp_time_stamp;
2836 prior_packets = tp->packets_out;
2840 prior_in_flight = tcp_packets_in_flight(tp);
2842 /* See if we can take anything off of the retransmit queue. */
2843 flag |= tcp_clean_rtx_queue(sk, &seq_rtt);
2845 if (tp->frto_counter)
2846 tcp_process_frto(sk, prior_snd_una);
2848 if (tcp_ack_is_dubious(tp, flag)) {
2849 /* Advanve CWND, if state allows this. */
2850 if ((flag & FLAG_DATA_ACKED) &&
2851 (tcp_vegas_enabled(tp) || prior_in_flight >= tp->snd_cwnd) &&
2852 tcp_may_raise_cwnd(tp, flag))
2853 tcp_cong_avoid(tp, ack, seq_rtt);
2854 tcp_fastretrans_alert(sk, prior_snd_una, prior_packets, flag);
2856 if ((flag & FLAG_DATA_ACKED) &&
2857 (tcp_vegas_enabled(tp) || prior_in_flight >= tp->snd_cwnd))
2858 tcp_cong_avoid(tp, ack, seq_rtt);
2861 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag&FLAG_NOT_DUP))
2862 dst_confirm(sk->sk_dst_cache);
2869 /* If this ack opens up a zero window, clear backoff. It was
2870 * being used to time the probes, and is probably far higher than
2871 * it needs to be for normal retransmission.
2873 if (sk->sk_send_head)
2878 if (TCP_SKB_CB(skb)->sacked)
2879 tcp_sacktag_write_queue(sk, skb, prior_snd_una);
2882 SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
2887 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
2888 * But, this can also be called on packets in the established flow when
2889 * the fast version below fails.
2891 void tcp_parse_options(struct sk_buff *skb, struct tcp_opt *tp, int estab)
2894 struct tcphdr *th = skb->h.th;
2895 int length=(th->doff*4)-sizeof(struct tcphdr);
2897 ptr = (unsigned char *)(th + 1);
2907 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
2912 if (opsize < 2) /* "silly options" */
2914 if (opsize > length)
2915 return; /* don't parse partial options */
2918 if(opsize==TCPOLEN_MSS && th->syn && !estab) {
2919 u16 in_mss = ntohs(*(__u16 *)ptr);
2921 if (tp->user_mss && tp->user_mss < in_mss)
2922 in_mss = tp->user_mss;
2923 tp->mss_clamp = in_mss;
2928 if(opsize==TCPOLEN_WINDOW && th->syn && !estab)
2929 if (sysctl_tcp_window_scaling) {
2931 tp->snd_wscale = *(__u8 *)ptr;
2932 if(tp->snd_wscale > 14) {
2934 printk("tcp_parse_options: Illegal window "
2935 "scaling value %d >14 received.",
2937 tp->snd_wscale = 14;
2941 case TCPOPT_TIMESTAMP:
2942 if(opsize==TCPOLEN_TIMESTAMP) {
2943 if ((estab && tp->tstamp_ok) ||
2944 (!estab && sysctl_tcp_timestamps)) {
2946 tp->rcv_tsval = ntohl(*(__u32 *)ptr);
2947 tp->rcv_tsecr = ntohl(*(__u32 *)(ptr+4));
2951 case TCPOPT_SACK_PERM:
2952 if(opsize==TCPOLEN_SACK_PERM && th->syn && !estab) {
2953 if (sysctl_tcp_sack) {
2961 if((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
2962 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
2964 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
2973 /* Fast parse options. This hopes to only see timestamps.
2974 * If it is wrong it falls back on tcp_parse_options().
2976 static __inline__ int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th, struct tcp_opt *tp)
2978 if (th->doff == sizeof(struct tcphdr)>>2) {
2981 } else if (tp->tstamp_ok &&
2982 th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) {
2983 __u32 *ptr = (__u32 *)(th + 1);
2984 if (*ptr == ntohl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
2985 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
2988 tp->rcv_tsval = ntohl(*ptr);
2990 tp->rcv_tsecr = ntohl(*ptr);
2994 tcp_parse_options(skb, tp, 1);
2998 static __inline__ void
2999 tcp_store_ts_recent(struct tcp_opt *tp)
3001 tp->ts_recent = tp->rcv_tsval;
3002 tp->ts_recent_stamp = xtime.tv_sec;
3005 static __inline__ void
3006 tcp_replace_ts_recent(struct tcp_opt *tp, u32 seq)
3008 if (tp->saw_tstamp && !after(seq, tp->rcv_wup)) {
3009 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3010 * extra check below makes sure this can only happen
3011 * for pure ACK frames. -DaveM
3013 * Not only, also it occurs for expired timestamps.
3016 if((s32)(tp->rcv_tsval - tp->ts_recent) >= 0 ||
3017 xtime.tv_sec >= tp->ts_recent_stamp + TCP_PAWS_24DAYS)
3018 tcp_store_ts_recent(tp);
3022 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3024 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3025 * it can pass through stack. So, the following predicate verifies that
3026 * this segment is not used for anything but congestion avoidance or
3027 * fast retransmit. Moreover, we even are able to eliminate most of such
3028 * second order effects, if we apply some small "replay" window (~RTO)
3029 * to timestamp space.
3031 * All these measures still do not guarantee that we reject wrapped ACKs
3032 * on networks with high bandwidth, when sequence space is recycled fastly,
3033 * but it guarantees that such events will be very rare and do not affect
3034 * connection seriously. This doesn't look nice, but alas, PAWS is really
3037 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3038 * states that events when retransmit arrives after original data are rare.
3039 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3040 * the biggest problem on large power networks even with minor reordering.
3041 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3042 * up to bandwidth of 18Gigabit/sec. 8) ]
3045 static int tcp_disordered_ack(struct tcp_opt *tp, struct sk_buff *skb)
3047 struct tcphdr *th = skb->h.th;
3048 u32 seq = TCP_SKB_CB(skb)->seq;
3049 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3051 return (/* 1. Pure ACK with correct sequence number. */
3052 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3054 /* 2. ... and duplicate ACK. */
3055 ack == tp->snd_una &&
3057 /* 3. ... and does not update window. */
3058 !tcp_may_update_window(tp, ack, seq, ntohs(th->window)<<tp->snd_wscale) &&
3060 /* 4. ... and sits in replay window. */
3061 (s32)(tp->ts_recent - tp->rcv_tsval) <= (tp->rto*1024)/HZ);
3064 static __inline__ int tcp_paws_discard(struct tcp_opt *tp, struct sk_buff *skb)
3066 return ((s32)(tp->ts_recent - tp->rcv_tsval) > TCP_PAWS_WINDOW &&
3067 xtime.tv_sec < tp->ts_recent_stamp + TCP_PAWS_24DAYS &&
3068 !tcp_disordered_ack(tp, skb));
3071 /* Check segment sequence number for validity.
3073 * Segment controls are considered valid, if the segment
3074 * fits to the window after truncation to the window. Acceptability
3075 * of data (and SYN, FIN, of course) is checked separately.
3076 * See tcp_data_queue(), for example.
3078 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3079 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3080 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3081 * (borrowed from freebsd)
3084 static inline int tcp_sequence(struct tcp_opt *tp, u32 seq, u32 end_seq)
3086 return !before(end_seq, tp->rcv_wup) &&
3087 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3090 /* When we get a reset we do this. */
3091 static void tcp_reset(struct sock *sk)
3093 /* We want the right error as BSD sees it (and indeed as we do). */
3094 switch (sk->sk_state) {
3096 sk->sk_err = ECONNREFUSED;
3098 case TCP_CLOSE_WAIT:
3104 sk->sk_err = ECONNRESET;
3107 if (!sock_flag(sk, SOCK_DEAD))
3108 sk->sk_error_report(sk);
3114 * Process the FIN bit. This now behaves as it is supposed to work
3115 * and the FIN takes effect when it is validly part of sequence
3116 * space. Not before when we get holes.
3118 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3119 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3122 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3123 * close and we go into CLOSING (and later onto TIME-WAIT)
3125 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3127 static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th)
3129 struct tcp_opt *tp = tcp_sk(sk);
3131 tcp_schedule_ack(tp);
3133 sk->sk_shutdown |= RCV_SHUTDOWN;
3134 sock_set_flag(sk, SOCK_DONE);
3136 switch (sk->sk_state) {
3138 case TCP_ESTABLISHED:
3139 /* Move to CLOSE_WAIT */
3140 tcp_set_state(sk, TCP_CLOSE_WAIT);
3141 tp->ack.pingpong = 1;
3144 case TCP_CLOSE_WAIT:
3146 /* Received a retransmission of the FIN, do
3151 /* RFC793: Remain in the LAST-ACK state. */
3155 /* This case occurs when a simultaneous close
3156 * happens, we must ack the received FIN and
3157 * enter the CLOSING state.
3160 tcp_set_state(sk, TCP_CLOSING);
3163 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3165 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
3168 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3169 * cases we should never reach this piece of code.
3171 printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n",
3172 __FUNCTION__, sk->sk_state);
3176 /* It _is_ possible, that we have something out-of-order _after_ FIN.
3177 * Probably, we should reset in this case. For now drop them.
3179 __skb_queue_purge(&tp->out_of_order_queue);
3182 sk_stream_mem_reclaim(sk);
3184 if (!sock_flag(sk, SOCK_DEAD)) {
3185 sk->sk_state_change(sk);
3187 /* Do not send POLL_HUP for half duplex close. */
3188 if (sk->sk_shutdown == SHUTDOWN_MASK ||
3189 sk->sk_state == TCP_CLOSE)
3190 sk_wake_async(sk, 1, POLL_HUP);
3192 sk_wake_async(sk, 1, POLL_IN);
3196 static __inline__ int
3197 tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, u32 end_seq)
3199 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
3200 if (before(seq, sp->start_seq))
3201 sp->start_seq = seq;
3202 if (after(end_seq, sp->end_seq))
3203 sp->end_seq = end_seq;
3209 static __inline__ void tcp_dsack_set(struct tcp_opt *tp, u32 seq, u32 end_seq)
3211 if (tp->sack_ok && sysctl_tcp_dsack) {
3212 if (before(seq, tp->rcv_nxt))
3213 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT);
3215 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT);
3218 tp->duplicate_sack[0].start_seq = seq;
3219 tp->duplicate_sack[0].end_seq = end_seq;
3220 tp->eff_sacks = min(tp->num_sacks+1, 4-tp->tstamp_ok);
3224 static __inline__ void tcp_dsack_extend(struct tcp_opt *tp, u32 seq, u32 end_seq)
3227 tcp_dsack_set(tp, seq, end_seq);
3229 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
3232 static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb)
3234 struct tcp_opt *tp = tcp_sk(sk);
3236 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
3237 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3238 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
3239 tcp_enter_quickack_mode(tp);
3241 if (tp->sack_ok && sysctl_tcp_dsack) {
3242 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3244 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
3245 end_seq = tp->rcv_nxt;
3246 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, end_seq);
3253 /* These routines update the SACK block as out-of-order packets arrive or
3254 * in-order packets close up the sequence space.
3256 static void tcp_sack_maybe_coalesce(struct tcp_opt *tp)
3259 struct tcp_sack_block *sp = &tp->selective_acks[0];
3260 struct tcp_sack_block *swalk = sp+1;
3262 /* See if the recent change to the first SACK eats into
3263 * or hits the sequence space of other SACK blocks, if so coalesce.
3265 for (this_sack = 1; this_sack < tp->num_sacks; ) {
3266 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
3269 /* Zap SWALK, by moving every further SACK up by one slot.
3270 * Decrease num_sacks.
3273 tp->eff_sacks = min(tp->num_sacks+tp->dsack, 4-tp->tstamp_ok);
3274 for(i=this_sack; i < tp->num_sacks; i++)
3278 this_sack++, swalk++;
3282 static __inline__ void tcp_sack_swap(struct tcp_sack_block *sack1, struct tcp_sack_block *sack2)
3286 tmp = sack1->start_seq;
3287 sack1->start_seq = sack2->start_seq;
3288 sack2->start_seq = tmp;
3290 tmp = sack1->end_seq;
3291 sack1->end_seq = sack2->end_seq;
3292 sack2->end_seq = tmp;
3295 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
3297 struct tcp_opt *tp = tcp_sk(sk);
3298 struct tcp_sack_block *sp = &tp->selective_acks[0];
3299 int cur_sacks = tp->num_sacks;
3305 for (this_sack=0; this_sack<cur_sacks; this_sack++, sp++) {
3306 if (tcp_sack_extend(sp, seq, end_seq)) {
3307 /* Rotate this_sack to the first one. */
3308 for (; this_sack>0; this_sack--, sp--)
3309 tcp_sack_swap(sp, sp-1);
3311 tcp_sack_maybe_coalesce(tp);
3316 /* Could not find an adjacent existing SACK, build a new one,
3317 * put it at the front, and shift everyone else down. We
3318 * always know there is at least one SACK present already here.
3320 * If the sack array is full, forget about the last one.
3322 if (this_sack >= 4) {
3327 for(; this_sack > 0; this_sack--, sp--)
3331 /* Build the new head SACK, and we're done. */
3332 sp->start_seq = seq;
3333 sp->end_seq = end_seq;
3335 tp->eff_sacks = min(tp->num_sacks + tp->dsack, 4 - tp->tstamp_ok);
3338 /* RCV.NXT advances, some SACKs should be eaten. */
3340 static void tcp_sack_remove(struct tcp_opt *tp)
3342 struct tcp_sack_block *sp = &tp->selective_acks[0];
3343 int num_sacks = tp->num_sacks;
3346 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
3347 if (skb_queue_len(&tp->out_of_order_queue) == 0) {
3349 tp->eff_sacks = tp->dsack;
3353 for(this_sack = 0; this_sack < num_sacks; ) {
3354 /* Check if the start of the sack is covered by RCV.NXT. */
3355 if (!before(tp->rcv_nxt, sp->start_seq)) {
3358 /* RCV.NXT must cover all the block! */
3359 BUG_TRAP(!before(tp->rcv_nxt, sp->end_seq));
3361 /* Zap this SACK, by moving forward any other SACKS. */
3362 for (i=this_sack+1; i < num_sacks; i++)
3363 tp->selective_acks[i-1] = tp->selective_acks[i];
3370 if (num_sacks != tp->num_sacks) {
3371 tp->num_sacks = num_sacks;
3372 tp->eff_sacks = min(tp->num_sacks+tp->dsack, 4-tp->tstamp_ok);
3376 /* This one checks to see if we can put data from the
3377 * out_of_order queue into the receive_queue.
3379 static void tcp_ofo_queue(struct sock *sk)
3381 struct tcp_opt *tp = tcp_sk(sk);
3382 __u32 dsack_high = tp->rcv_nxt;
3383 struct sk_buff *skb;
3385 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
3386 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
3389 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
3390 __u32 dsack = dsack_high;
3391 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
3392 dsack_high = TCP_SKB_CB(skb)->end_seq;
3393 tcp_dsack_extend(tp, TCP_SKB_CB(skb)->seq, dsack);
3396 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3397 SOCK_DEBUG(sk, "ofo packet was already received \n");
3398 __skb_unlink(skb, skb->list);
3402 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
3403 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3404 TCP_SKB_CB(skb)->end_seq);
3406 __skb_unlink(skb, skb->list);
3407 __skb_queue_tail(&sk->sk_receive_queue, skb);
3408 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3410 tcp_fin(skb, sk, skb->h.th);
3414 static int tcp_prune_queue(struct sock *sk);
3416 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
3418 struct tcphdr *th = skb->h.th;
3419 struct tcp_opt *tp = tcp_sk(sk);
3422 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
3426 __skb_pull(skb, th->doff*4);
3428 TCP_ECN_accept_cwr(tp, skb);
3432 tp->eff_sacks = min_t(unsigned int, tp->num_sacks,
3436 /* Queue data for delivery to the user.
3437 * Packets in sequence go to the receive queue.
3438 * Out of sequence packets to the out_of_order_queue.
3440 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
3441 if (tcp_receive_window(tp) == 0)
3444 /* Ok. In sequence. In window. */
3445 if (tp->ucopy.task == current &&
3446 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
3447 sock_owned_by_user(sk) && !tp->urg_data) {
3448 int chunk = min_t(unsigned int, skb->len,
3451 __set_current_state(TASK_RUNNING);
3454 if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
3455 tp->ucopy.len -= chunk;
3456 tp->copied_seq += chunk;
3457 eaten = (chunk == skb->len && !th->fin);
3458 tcp_rcv_space_adjust(sk);
3466 (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3467 !sk_stream_rmem_schedule(sk, skb))) {
3468 if (tcp_prune_queue(sk) < 0 ||
3469 !sk_stream_rmem_schedule(sk, skb))
3472 sk_stream_set_owner_r(skb, sk);
3473 __skb_queue_tail(&sk->sk_receive_queue, skb);
3475 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3477 tcp_event_data_recv(sk, tp, skb);
3479 tcp_fin(skb, sk, th);
3481 if (skb_queue_len(&tp->out_of_order_queue)) {
3484 /* RFC2581. 4.2. SHOULD send immediate ACK, when
3485 * gap in queue is filled.
3487 if (!skb_queue_len(&tp->out_of_order_queue))
3488 tp->ack.pingpong = 0;
3492 tcp_sack_remove(tp);
3494 tcp_fast_path_check(sk, tp);
3498 else if (!sock_flag(sk, SOCK_DEAD))
3499 sk->sk_data_ready(sk, 0);
3503 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3504 /* A retransmit, 2nd most common case. Force an immediate ack. */
3505 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
3506 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3509 tcp_enter_quickack_mode(tp);
3510 tcp_schedule_ack(tp);
3516 /* Out of window. F.e. zero window probe. */
3517 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
3520 tcp_enter_quickack_mode(tp);
3522 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3523 /* Partial packet, seq < rcv_next < end_seq */
3524 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
3525 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3526 TCP_SKB_CB(skb)->end_seq);
3528 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
3530 /* If window is closed, drop tail of packet. But after
3531 * remembering D-SACK for its head made in previous line.
3533 if (!tcp_receive_window(tp))
3538 TCP_ECN_check_ce(tp, skb);
3540 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3541 !sk_stream_rmem_schedule(sk, skb)) {
3542 if (tcp_prune_queue(sk) < 0 ||
3543 !sk_stream_rmem_schedule(sk, skb))
3547 /* Disable header prediction. */
3549 tcp_schedule_ack(tp);
3551 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
3552 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3554 sk_stream_set_owner_r(skb, sk);
3556 if (!skb_peek(&tp->out_of_order_queue)) {
3557 /* Initial out of order segment, build 1 SACK. */
3562 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
3563 tp->selective_acks[0].end_seq =
3564 TCP_SKB_CB(skb)->end_seq;
3566 __skb_queue_head(&tp->out_of_order_queue,skb);
3568 struct sk_buff *skb1 = tp->out_of_order_queue.prev;
3569 u32 seq = TCP_SKB_CB(skb)->seq;
3570 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3572 if (seq == TCP_SKB_CB(skb1)->end_seq) {
3573 __skb_append(skb1, skb);
3575 if (!tp->num_sacks ||
3576 tp->selective_acks[0].end_seq != seq)
3579 /* Common case: data arrive in order after hole. */
3580 tp->selective_acks[0].end_seq = end_seq;
3584 /* Find place to insert this segment. */
3586 if (!after(TCP_SKB_CB(skb1)->seq, seq))
3588 } while ((skb1 = skb1->prev) !=
3589 (struct sk_buff*)&tp->out_of_order_queue);
3591 /* Do skb overlap to previous one? */
3592 if (skb1 != (struct sk_buff*)&tp->out_of_order_queue &&
3593 before(seq, TCP_SKB_CB(skb1)->end_seq)) {
3594 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
3595 /* All the bits are present. Drop. */
3597 tcp_dsack_set(tp, seq, end_seq);
3600 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
3601 /* Partial overlap. */
3602 tcp_dsack_set(tp, seq, TCP_SKB_CB(skb1)->end_seq);
3607 __skb_insert(skb, skb1, skb1->next, &tp->out_of_order_queue);
3609 /* And clean segments covered by new one as whole. */
3610 while ((skb1 = skb->next) !=
3611 (struct sk_buff*)&tp->out_of_order_queue &&
3612 after(end_seq, TCP_SKB_CB(skb1)->seq)) {
3613 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
3614 tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, end_seq);
3617 __skb_unlink(skb1, skb1->list);
3618 tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, TCP_SKB_CB(skb1)->end_seq);
3624 tcp_sack_new_ofo_skb(sk, seq, end_seq);
3628 /* Collapse contiguous sequence of skbs head..tail with
3629 * sequence numbers start..end.
3630 * Segments with FIN/SYN are not collapsed (only because this
3634 tcp_collapse(struct sock *sk, struct sk_buff *head,
3635 struct sk_buff *tail, u32 start, u32 end)
3637 struct sk_buff *skb;
3639 /* First, check that queue is collapsable and find
3640 * the point where collapsing can be useful. */
3641 for (skb = head; skb != tail; ) {
3642 /* No new bits? It is possible on ofo queue. */
3643 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
3644 struct sk_buff *next = skb->next;
3645 __skb_unlink(skb, skb->list);
3647 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
3652 /* The first skb to collapse is:
3654 * - bloated or contains data before "start" or
3655 * overlaps to the next one.
3657 if (!skb->h.th->syn && !skb->h.th->fin &&
3658 (tcp_win_from_space(skb->truesize) > skb->len ||
3659 before(TCP_SKB_CB(skb)->seq, start) ||
3660 (skb->next != tail &&
3661 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq)))
3664 /* Decided to skip this, advance start seq. */
3665 start = TCP_SKB_CB(skb)->end_seq;
3668 if (skb == tail || skb->h.th->syn || skb->h.th->fin)
3671 while (before(start, end)) {
3672 struct sk_buff *nskb;
3673 int header = skb_headroom(skb);
3674 int copy = (PAGE_SIZE - sizeof(struct sk_buff) -
3675 sizeof(struct skb_shared_info) - header - 31)&~15;
3677 /* Too big header? This can happen with IPv6. */
3680 if (end-start < copy)
3682 nskb = alloc_skb(copy+header, GFP_ATOMIC);
3685 skb_reserve(nskb, header);
3686 memcpy(nskb->head, skb->head, header);
3687 nskb->nh.raw = nskb->head + (skb->nh.raw-skb->head);
3688 nskb->h.raw = nskb->head + (skb->h.raw-skb->head);
3689 nskb->mac.raw = nskb->head + (skb->mac.raw-skb->head);
3690 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
3691 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
3692 __skb_insert(nskb, skb->prev, skb, skb->list);
3693 sk_stream_set_owner_r(nskb, sk);
3695 /* Copy data, releasing collapsed skbs. */
3697 int offset = start - TCP_SKB_CB(skb)->seq;
3698 int size = TCP_SKB_CB(skb)->end_seq - start;
3700 if (offset < 0) BUG();
3702 size = min(copy, size);
3703 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
3705 TCP_SKB_CB(nskb)->end_seq += size;
3709 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
3710 struct sk_buff *next = skb->next;
3711 __skb_unlink(skb, skb->list);
3713 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
3715 if (skb == tail || skb->h.th->syn || skb->h.th->fin)
3722 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
3723 * and tcp_collapse() them until all the queue is collapsed.
3725 static void tcp_collapse_ofo_queue(struct sock *sk)
3727 struct tcp_opt *tp = tcp_sk(sk);
3728 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
3729 struct sk_buff *head;
3735 start = TCP_SKB_CB(skb)->seq;
3736 end = TCP_SKB_CB(skb)->end_seq;
3742 /* Segment is terminated when we see gap or when
3743 * we are at the end of all the queue. */
3744 if (skb == (struct sk_buff *)&tp->out_of_order_queue ||
3745 after(TCP_SKB_CB(skb)->seq, end) ||
3746 before(TCP_SKB_CB(skb)->end_seq, start)) {
3747 tcp_collapse(sk, head, skb, start, end);
3749 if (skb == (struct sk_buff *)&tp->out_of_order_queue)
3751 /* Start new segment */
3752 start = TCP_SKB_CB(skb)->seq;
3753 end = TCP_SKB_CB(skb)->end_seq;
3755 if (before(TCP_SKB_CB(skb)->seq, start))
3756 start = TCP_SKB_CB(skb)->seq;
3757 if (after(TCP_SKB_CB(skb)->end_seq, end))
3758 end = TCP_SKB_CB(skb)->end_seq;
3763 /* Reduce allocated memory if we can, trying to get
3764 * the socket within its memory limits again.
3766 * Return less than zero if we should start dropping frames
3767 * until the socket owning process reads some of the data
3768 * to stabilize the situation.
3770 static int tcp_prune_queue(struct sock *sk)
3772 struct tcp_opt *tp = tcp_sk(sk);
3774 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
3776 NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED);
3778 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
3779 tcp_clamp_window(sk, tp);
3780 else if (tcp_memory_pressure)
3781 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
3783 tcp_collapse_ofo_queue(sk);
3784 tcp_collapse(sk, sk->sk_receive_queue.next,
3785 (struct sk_buff*)&sk->sk_receive_queue,
3786 tp->copied_seq, tp->rcv_nxt);
3787 sk_stream_mem_reclaim(sk);
3789 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
3792 /* Collapsing did not help, destructive actions follow.
3793 * This must not ever occur. */
3795 /* First, purge the out_of_order queue. */
3796 if (skb_queue_len(&tp->out_of_order_queue)) {
3797 NET_ADD_STATS_BH(LINUX_MIB_OFOPRUNED,
3798 skb_queue_len(&tp->out_of_order_queue));
3799 __skb_queue_purge(&tp->out_of_order_queue);
3801 /* Reset SACK state. A conforming SACK implementation will
3802 * do the same at a timeout based retransmit. When a connection
3803 * is in a sad state like this, we care only about integrity
3804 * of the connection not performance.
3808 sk_stream_mem_reclaim(sk);
3811 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
3814 /* If we are really being abused, tell the caller to silently
3815 * drop receive data on the floor. It will get retransmitted
3816 * and hopefully then we'll have sufficient space.
3818 NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED);
3820 /* Massive buffer overcommit. */
3826 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
3827 * As additional protections, we do not touch cwnd in retransmission phases,
3828 * and if application hit its sndbuf limit recently.
3830 void tcp_cwnd_application_limited(struct sock *sk)
3832 struct tcp_opt *tp = tcp_sk(sk);
3834 if (tp->ca_state == TCP_CA_Open &&
3835 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
3836 /* Limited by application or receiver window. */
3837 u32 win_used = max(tp->snd_cwnd_used, 2U);
3838 if (win_used < tp->snd_cwnd) {
3839 tp->snd_ssthresh = tcp_current_ssthresh(tp);
3840 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
3842 tp->snd_cwnd_used = 0;
3844 tp->snd_cwnd_stamp = tcp_time_stamp;
3848 /* When incoming ACK allowed to free some skb from write_queue,
3849 * we remember this event in flag sk->sk_queue_shrunk and wake up socket
3850 * on the exit from tcp input handler.
3852 * PROBLEM: sndbuf expansion does not work well with largesend.
3854 static void tcp_new_space(struct sock *sk)
3856 struct tcp_opt *tp = tcp_sk(sk);
3858 if (tp->packets_out < tp->snd_cwnd &&
3859 !(sk->sk_userlocks & SOCK_SNDBUF_LOCK) &&
3860 !tcp_memory_pressure &&
3861 atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) {
3862 int sndmem = max_t(u32, tp->mss_clamp, tp->mss_cache) +
3863 MAX_TCP_HEADER + 16 + sizeof(struct sk_buff),
3864 demanded = max_t(unsigned int, tp->snd_cwnd,
3865 tp->reordering + 1);
3866 sndmem *= 2*demanded;
3867 if (sndmem > sk->sk_sndbuf)
3868 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
3869 tp->snd_cwnd_stamp = tcp_time_stamp;
3872 sk->sk_write_space(sk);
3875 static inline void tcp_check_space(struct sock *sk)
3877 if (sk->sk_queue_shrunk) {
3878 sk->sk_queue_shrunk = 0;
3879 if (sk->sk_socket &&
3880 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
3885 static void __tcp_data_snd_check(struct sock *sk, struct sk_buff *skb)
3887 struct tcp_opt *tp = tcp_sk(sk);
3889 if (after(TCP_SKB_CB(skb)->end_seq, tp->snd_una + tp->snd_wnd) ||
3890 tcp_packets_in_flight(tp) >= tp->snd_cwnd ||
3891 tcp_write_xmit(sk, tp->nonagle))
3892 tcp_check_probe_timer(sk, tp);
3895 static __inline__ void tcp_data_snd_check(struct sock *sk)
3897 struct sk_buff *skb = sk->sk_send_head;
3900 __tcp_data_snd_check(sk, skb);
3901 tcp_check_space(sk);
3905 * Check if sending an ack is needed.
3907 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
3909 struct tcp_opt *tp = tcp_sk(sk);
3911 /* More than one full frame received... */
3912 if (((tp->rcv_nxt - tp->rcv_wup) > tp->ack.rcv_mss
3913 /* ... and right edge of window advances far enough.
3914 * (tcp_recvmsg() will send ACK otherwise). Or...
3916 && __tcp_select_window(sk) >= tp->rcv_wnd) ||
3917 /* We ACK each frame or... */
3918 tcp_in_quickack_mode(tp) ||
3919 /* We have out of order data. */
3921 skb_peek(&tp->out_of_order_queue))) {
3922 /* Then ack it now */
3925 /* Else, send delayed ack. */
3926 tcp_send_delayed_ack(sk);
3930 static __inline__ void tcp_ack_snd_check(struct sock *sk)
3932 struct tcp_opt *tp = tcp_sk(sk);
3933 if (!tcp_ack_scheduled(tp)) {
3934 /* We sent a data segment already. */
3937 __tcp_ack_snd_check(sk, 1);
3941 * This routine is only called when we have urgent data
3942 * signalled. Its the 'slow' part of tcp_urg. It could be
3943 * moved inline now as tcp_urg is only called from one
3944 * place. We handle URGent data wrong. We have to - as
3945 * BSD still doesn't use the correction from RFC961.
3946 * For 1003.1g we should support a new option TCP_STDURG to permit
3947 * either form (or just set the sysctl tcp_stdurg).
3950 static void tcp_check_urg(struct sock * sk, struct tcphdr * th)
3952 struct tcp_opt *tp = tcp_sk(sk);
3953 u32 ptr = ntohs(th->urg_ptr);
3955 if (ptr && !sysctl_tcp_stdurg)
3957 ptr += ntohl(th->seq);
3959 /* Ignore urgent data that we've already seen and read. */
3960 if (after(tp->copied_seq, ptr))
3963 /* Do not replay urg ptr.
3965 * NOTE: interesting situation not covered by specs.
3966 * Misbehaving sender may send urg ptr, pointing to segment,
3967 * which we already have in ofo queue. We are not able to fetch
3968 * such data and will stay in TCP_URG_NOTYET until will be eaten
3969 * by recvmsg(). Seems, we are not obliged to handle such wicked
3970 * situations. But it is worth to think about possibility of some
3971 * DoSes using some hypothetical application level deadlock.
3973 if (before(ptr, tp->rcv_nxt))
3976 /* Do we already have a newer (or duplicate) urgent pointer? */
3977 if (tp->urg_data && !after(ptr, tp->urg_seq))
3980 /* Tell the world about our new urgent pointer. */
3983 /* We may be adding urgent data when the last byte read was
3984 * urgent. To do this requires some care. We cannot just ignore
3985 * tp->copied_seq since we would read the last urgent byte again
3986 * as data, nor can we alter copied_seq until this data arrives
3987 * or we break the sematics of SIOCATMARK (and thus sockatmark())
3989 * NOTE. Double Dutch. Rendering to plain English: author of comment
3990 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
3991 * and expect that both A and B disappear from stream. This is _wrong_.
3992 * Though this happens in BSD with high probability, this is occasional.
3993 * Any application relying on this is buggy. Note also, that fix "works"
3994 * only in this artificial test. Insert some normal data between A and B and we will
3995 * decline of BSD again. Verdict: it is better to remove to trap
3998 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
3999 !sock_flag(sk, SOCK_URGINLINE) &&
4000 tp->copied_seq != tp->rcv_nxt) {
4001 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
4003 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
4004 __skb_unlink(skb, skb->list);
4009 tp->urg_data = TCP_URG_NOTYET;
4012 /* Disable header prediction. */
4016 /* This is the 'fast' part of urgent handling. */
4017 static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th)
4019 struct tcp_opt *tp = tcp_sk(sk);
4021 /* Check if we get a new urgent pointer - normally not. */
4023 tcp_check_urg(sk,th);
4025 /* Do we wait for any urgent data? - normally not... */
4026 if (tp->urg_data == TCP_URG_NOTYET) {
4027 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
4030 /* Is the urgent pointer pointing into this packet? */
4031 if (ptr < skb->len) {
4033 if (skb_copy_bits(skb, ptr, &tmp, 1))
4035 tp->urg_data = TCP_URG_VALID | tmp;
4036 if (!sock_flag(sk, SOCK_DEAD))
4037 sk->sk_data_ready(sk, 0);
4042 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
4044 struct tcp_opt *tp = tcp_sk(sk);
4045 int chunk = skb->len - hlen;
4049 if (skb->ip_summed==CHECKSUM_UNNECESSARY)
4050 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
4052 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
4056 tp->ucopy.len -= chunk;
4057 tp->copied_seq += chunk;
4058 tcp_rcv_space_adjust(sk);
4065 static int __tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
4069 if (sock_owned_by_user(sk)) {
4071 result = __tcp_checksum_complete(skb);
4074 result = __tcp_checksum_complete(skb);
4079 static __inline__ int
4080 tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
4082 return skb->ip_summed != CHECKSUM_UNNECESSARY &&
4083 __tcp_checksum_complete_user(sk, skb);
4087 * TCP receive function for the ESTABLISHED state.
4089 * It is split into a fast path and a slow path. The fast path is
4091 * - A zero window was announced from us - zero window probing
4092 * is only handled properly in the slow path.
4093 * - Out of order segments arrived.
4094 * - Urgent data is expected.
4095 * - There is no buffer space left
4096 * - Unexpected TCP flags/window values/header lengths are received
4097 * (detected by checking the TCP header against pred_flags)
4098 * - Data is sent in both directions. Fast path only supports pure senders
4099 * or pure receivers (this means either the sequence number or the ack
4100 * value must stay constant)
4101 * - Unexpected TCP option.
4103 * When these conditions are not satisfied it drops into a standard
4104 * receive procedure patterned after RFC793 to handle all cases.
4105 * The first three cases are guaranteed by proper pred_flags setting,
4106 * the rest is checked inline. Fast processing is turned on in
4107 * tcp_data_queue when everything is OK.
4109 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
4110 struct tcphdr *th, unsigned len)
4112 struct tcp_opt *tp = tcp_sk(sk);
4115 * Header prediction.
4116 * The code loosely follows the one in the famous
4117 * "30 instruction TCP receive" Van Jacobson mail.
4119 * Van's trick is to deposit buffers into socket queue
4120 * on a device interrupt, to call tcp_recv function
4121 * on the receive process context and checksum and copy
4122 * the buffer to user space. smart...
4124 * Our current scheme is not silly either but we take the
4125 * extra cost of the net_bh soft interrupt processing...
4126 * We do checksum and copy also but from device to kernel.
4131 /* pred_flags is 0xS?10 << 16 + snd_wnd
4132 * if header_predition is to be made
4133 * 'S' will always be tp->tcp_header_len >> 2
4134 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
4135 * turn it off (when there are holes in the receive
4136 * space for instance)
4137 * PSH flag is ignored.
4140 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
4141 TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4142 int tcp_header_len = tp->tcp_header_len;
4144 /* Timestamp header prediction: tcp_header_len
4145 * is automatically equal to th->doff*4 due to pred_flags
4149 /* Check timestamp */
4150 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
4151 __u32 *ptr = (__u32 *)(th + 1);
4153 /* No? Slow path! */
4154 if (*ptr != ntohl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
4155 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP))
4160 tp->rcv_tsval = ntohl(*ptr);
4162 tp->rcv_tsecr = ntohl(*ptr);
4164 /* If PAWS failed, check it more carefully in slow path */
4165 if ((s32)(tp->rcv_tsval - tp->ts_recent) < 0)
4168 /* DO NOT update ts_recent here, if checksum fails
4169 * and timestamp was corrupted part, it will result
4170 * in a hung connection since we will drop all
4171 * future packets due to the PAWS test.
4175 if (len <= tcp_header_len) {
4176 /* Bulk data transfer: sender */
4177 if (len == tcp_header_len) {
4178 /* Predicted packet is in window by definition.
4179 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4180 * Hence, check seq<=rcv_wup reduces to:
4182 if (tcp_header_len ==
4183 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
4184 tp->rcv_nxt == tp->rcv_wup)
4185 tcp_store_ts_recent(tp);
4187 tcp_rcv_rtt_measure_ts(tp, skb);
4189 /* We know that such packets are checksummed
4192 tcp_ack(sk, skb, 0);
4194 tcp_data_snd_check(sk);
4196 } else { /* Header too small */
4197 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4203 if (tp->ucopy.task == current &&
4204 tp->copied_seq == tp->rcv_nxt &&
4205 len - tcp_header_len <= tp->ucopy.len &&
4206 sock_owned_by_user(sk)) {
4207 __set_current_state(TASK_RUNNING);
4209 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len)) {
4210 /* Predicted packet is in window by definition.
4211 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4212 * Hence, check seq<=rcv_wup reduces to:
4214 if (tcp_header_len ==
4215 (sizeof(struct tcphdr) +
4216 TCPOLEN_TSTAMP_ALIGNED) &&
4217 tp->rcv_nxt == tp->rcv_wup)
4218 tcp_store_ts_recent(tp);
4220 tcp_rcv_rtt_measure_ts(tp, skb);
4222 __skb_pull(skb, tcp_header_len);
4223 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4224 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER);
4229 if (tcp_checksum_complete_user(sk, skb))
4232 /* Predicted packet is in window by definition.
4233 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4234 * Hence, check seq<=rcv_wup reduces to:
4236 if (tcp_header_len ==
4237 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
4238 tp->rcv_nxt == tp->rcv_wup)
4239 tcp_store_ts_recent(tp);
4241 tcp_rcv_rtt_measure_ts(tp, skb);
4243 if ((int)skb->truesize > sk->sk_forward_alloc)
4246 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS);
4248 /* Bulk data transfer: receiver */
4249 __skb_pull(skb,tcp_header_len);
4250 __skb_queue_tail(&sk->sk_receive_queue, skb);
4251 sk_stream_set_owner_r(skb, sk);
4252 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4255 tcp_event_data_recv(sk, tp, skb);
4257 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
4258 /* Well, only one small jumplet in fast path... */
4259 tcp_ack(sk, skb, FLAG_DATA);
4260 tcp_data_snd_check(sk);
4261 if (!tcp_ack_scheduled(tp))
4266 if (tcp_in_quickack_mode(tp)) {
4269 tcp_send_delayed_ack(sk);
4272 __tcp_ack_snd_check(sk, 0);
4279 sk->sk_data_ready(sk, 0);
4285 if (len < (th->doff<<2) || tcp_checksum_complete_user(sk, skb))
4289 * RFC1323: H1. Apply PAWS check first.
4291 if (tcp_fast_parse_options(skb, th, tp) && tp->saw_tstamp &&
4292 tcp_paws_discard(tp, skb)) {
4294 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
4295 tcp_send_dupack(sk, skb);
4298 /* Resets are accepted even if PAWS failed.
4300 ts_recent update must be made after we are sure
4301 that the packet is in window.
4306 * Standard slow path.
4309 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
4310 /* RFC793, page 37: "In all states except SYN-SENT, all reset
4311 * (RST) segments are validated by checking their SEQ-fields."
4312 * And page 69: "If an incoming segment is not acceptable,
4313 * an acknowledgment should be sent in reply (unless the RST bit
4314 * is set, if so drop the segment and return)".
4317 tcp_send_dupack(sk, skb);
4326 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
4328 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4329 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4330 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
4337 tcp_ack(sk, skb, FLAG_SLOWPATH);
4339 tcp_rcv_rtt_measure_ts(tp, skb);
4341 /* Process urgent data. */
4342 tcp_urg(sk, skb, th);
4344 /* step 7: process the segment text */
4345 tcp_data_queue(sk, skb);
4347 tcp_data_snd_check(sk);
4348 tcp_ack_snd_check(sk);
4352 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4359 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
4360 struct tcphdr *th, unsigned len)
4362 struct tcp_opt *tp = tcp_sk(sk);
4363 int saved_clamp = tp->mss_clamp;
4365 tcp_parse_options(skb, tp, 0);
4369 * "If the state is SYN-SENT then
4370 * first check the ACK bit
4371 * If the ACK bit is set
4372 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
4373 * a reset (unless the RST bit is set, if so drop
4374 * the segment and return)"
4376 * We do not send data with SYN, so that RFC-correct
4379 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
4380 goto reset_and_undo;
4382 if (tp->saw_tstamp && tp->rcv_tsecr &&
4383 !between(tp->rcv_tsecr, tp->retrans_stamp,
4385 NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED);
4386 goto reset_and_undo;
4389 /* Now ACK is acceptable.
4391 * "If the RST bit is set
4392 * If the ACK was acceptable then signal the user "error:
4393 * connection reset", drop the segment, enter CLOSED state,
4394 * delete TCB, and return."
4403 * "fifth, if neither of the SYN or RST bits is set then
4404 * drop the segment and return."
4410 goto discard_and_undo;
4413 * "If the SYN bit is on ...
4414 * are acceptable then ...
4415 * (our SYN has been ACKed), change the connection
4416 * state to ESTABLISHED..."
4419 TCP_ECN_rcv_synack(tp, th);
4420 if (tp->ecn_flags&TCP_ECN_OK)
4421 sk->sk_no_largesend = 1;
4423 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
4424 tcp_ack(sk, skb, FLAG_SLOWPATH);
4426 /* Ok.. it's good. Set up sequence numbers and
4427 * move to established.
4429 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
4430 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
4432 /* RFC1323: The window in SYN & SYN/ACK segments is
4435 tp->snd_wnd = ntohs(th->window);
4436 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(skb)->seq);
4438 if (!tp->wscale_ok) {
4439 tp->snd_wscale = tp->rcv_wscale = 0;
4440 tp->window_clamp = min(tp->window_clamp, 65535U);
4443 if (tp->saw_tstamp) {
4445 tp->tcp_header_len =
4446 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
4447 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
4448 tcp_store_ts_recent(tp);
4450 tp->tcp_header_len = sizeof(struct tcphdr);
4453 if (tp->sack_ok && sysctl_tcp_fack)
4456 tcp_sync_mss(sk, tp->pmtu_cookie);
4457 tcp_initialize_rcv_mss(sk);
4459 /* Remember, tcp_poll() does not lock socket!
4460 * Change state from SYN-SENT only after copied_seq
4461 * is initialized. */
4462 tp->copied_seq = tp->rcv_nxt;
4464 tcp_set_state(sk, TCP_ESTABLISHED);
4466 /* Make sure socket is routed, for correct metrics. */
4467 tp->af_specific->rebuild_header(sk);
4469 tcp_init_metrics(sk);
4471 /* Prevent spurious tcp_cwnd_restart() on first data
4474 tp->lsndtime = tcp_time_stamp;
4476 tcp_init_buffer_space(sk);
4478 if (sock_flag(sk, SOCK_KEEPOPEN))
4479 tcp_reset_keepalive_timer(sk, keepalive_time_when(tp));
4481 if (!tp->snd_wscale)
4482 __tcp_fast_path_on(tp, tp->snd_wnd);
4486 if (!sock_flag(sk, SOCK_DEAD)) {
4487 sk->sk_state_change(sk);
4488 sk_wake_async(sk, 0, POLL_OUT);
4491 if (sk->sk_write_pending || tp->defer_accept || tp->ack.pingpong) {
4492 /* Save one ACK. Data will be ready after
4493 * several ticks, if write_pending is set.
4495 * It may be deleted, but with this feature tcpdumps
4496 * look so _wonderfully_ clever, that I was not able
4497 * to stand against the temptation 8) --ANK
4499 tcp_schedule_ack(tp);
4500 tp->ack.lrcvtime = tcp_time_stamp;
4501 tp->ack.ato = TCP_ATO_MIN;
4502 tcp_incr_quickack(tp);
4503 tcp_enter_quickack_mode(tp);
4504 tcp_reset_xmit_timer(sk, TCP_TIME_DACK, TCP_DELACK_MAX);
4515 /* No ACK in the segment */
4519 * "If the RST bit is set
4521 * Otherwise (no ACK) drop the segment and return."
4524 goto discard_and_undo;
4528 if (tp->ts_recent_stamp && tp->saw_tstamp && tcp_paws_check(tp, 0))
4529 goto discard_and_undo;
4532 /* We see SYN without ACK. It is attempt of
4533 * simultaneous connect with crossed SYNs.
4534 * Particularly, it can be connect to self.
4536 tcp_set_state(sk, TCP_SYN_RECV);
4538 if (tp->saw_tstamp) {
4540 tcp_store_ts_recent(tp);
4541 tp->tcp_header_len =
4542 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
4544 tp->tcp_header_len = sizeof(struct tcphdr);
4547 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
4548 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
4550 /* RFC1323: The window in SYN & SYN/ACK segments is
4553 tp->snd_wnd = ntohs(th->window);
4554 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
4555 tp->max_window = tp->snd_wnd;
4557 TCP_ECN_rcv_syn(tp, th);
4558 if (tp->ecn_flags&TCP_ECN_OK)
4559 sk->sk_no_largesend = 1;
4561 tcp_sync_mss(sk, tp->pmtu_cookie);
4562 tcp_initialize_rcv_mss(sk);
4565 tcp_send_synack(sk);
4567 /* Note, we could accept data and URG from this segment.
4568 * There are no obstacles to make this.
4570 * However, if we ignore data in ACKless segments sometimes,
4571 * we have no reasons to accept it sometimes.
4572 * Also, seems the code doing it in step6 of tcp_rcv_state_process
4573 * is not flawless. So, discard packet for sanity.
4574 * Uncomment this return to process the data.
4581 /* "fifth, if neither of the SYN or RST bits is set then
4582 * drop the segment and return."
4586 tcp_clear_options(tp);
4587 tp->mss_clamp = saved_clamp;
4591 tcp_clear_options(tp);
4592 tp->mss_clamp = saved_clamp;
4598 * This function implements the receiving procedure of RFC 793 for
4599 * all states except ESTABLISHED and TIME_WAIT.
4600 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
4601 * address independent.
4604 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
4605 struct tcphdr *th, unsigned len)
4607 struct tcp_opt *tp = tcp_sk(sk);
4612 switch (sk->sk_state) {
4624 if(tp->af_specific->conn_request(sk, skb) < 0)
4630 /* Now we have several options: In theory there is
4631 * nothing else in the frame. KA9Q has an option to
4632 * send data with the syn, BSD accepts data with the
4633 * syn up to the [to be] advertised window and
4634 * Solaris 2.1 gives you a protocol error. For now
4635 * we just ignore it, that fits the spec precisely
4636 * and avoids incompatibilities. It would be nice in
4637 * future to drop through and process the data.
4639 * Now that TTCP is starting to be used we ought to
4641 * But, this leaves one open to an easy denial of
4642 * service attack, and SYN cookies can't defend
4643 * against this problem. So, we drop the data
4644 * in the interest of security over speed.
4654 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
4658 /* Do step6 onward by hand. */
4659 tcp_urg(sk, skb, th);
4661 tcp_data_snd_check(sk);
4665 if (tcp_fast_parse_options(skb, th, tp) && tp->saw_tstamp &&
4666 tcp_paws_discard(tp, skb)) {
4668 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
4669 tcp_send_dupack(sk, skb);
4672 /* Reset is accepted even if it did not pass PAWS. */
4675 /* step 1: check sequence number */
4676 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
4678 tcp_send_dupack(sk, skb);
4682 /* step 2: check RST bit */
4688 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
4690 /* step 3: check security and precedence [ignored] */
4694 * Check for a SYN in window.
4696 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4697 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
4702 /* step 5: check the ACK field */
4704 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH);
4706 switch(sk->sk_state) {
4709 tp->copied_seq = tp->rcv_nxt;
4711 tcp_set_state(sk, TCP_ESTABLISHED);
4712 sk->sk_state_change(sk);
4714 /* Note, that this wakeup is only for marginal
4715 * crossed SYN case. Passively open sockets
4716 * are not waked up, because sk->sk_sleep ==
4717 * NULL and sk->sk_socket == NULL.
4719 if (sk->sk_socket) {
4720 sk_wake_async(sk,0,POLL_OUT);
4723 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
4724 tp->snd_wnd = ntohs(th->window) <<
4726 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq,
4727 TCP_SKB_CB(skb)->seq);
4729 /* tcp_ack considers this ACK as duplicate
4730 * and does not calculate rtt.
4731 * Fix it at least with timestamps.
4733 if (tp->saw_tstamp && tp->rcv_tsecr &&
4735 tcp_ack_saw_tstamp(tp, 0);
4738 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
4740 /* Make sure socket is routed, for
4743 tp->af_specific->rebuild_header(sk);
4745 tcp_init_metrics(sk);
4747 /* Prevent spurious tcp_cwnd_restart() on
4748 * first data packet.
4750 tp->lsndtime = tcp_time_stamp;
4752 tcp_initialize_rcv_mss(sk);
4753 tcp_init_buffer_space(sk);
4754 tcp_fast_path_on(tp);
4761 if (tp->snd_una == tp->write_seq) {
4762 tcp_set_state(sk, TCP_FIN_WAIT2);
4763 sk->sk_shutdown |= SEND_SHUTDOWN;
4764 dst_confirm(sk->sk_dst_cache);
4766 if (!sock_flag(sk, SOCK_DEAD))
4767 /* Wake up lingering close() */
4768 sk->sk_state_change(sk);
4772 if (tp->linger2 < 0 ||
4773 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4774 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
4776 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
4780 tmo = tcp_fin_time(tp);
4781 if (tmo > TCP_TIMEWAIT_LEN) {
4782 tcp_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
4783 } else if (th->fin || sock_owned_by_user(sk)) {
4784 /* Bad case. We could lose such FIN otherwise.
4785 * It is not a big problem, but it looks confusing
4786 * and not so rare event. We still can lose it now,
4787 * if it spins in bh_lock_sock(), but it is really
4790 tcp_reset_keepalive_timer(sk, tmo);
4792 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
4800 if (tp->snd_una == tp->write_seq) {
4801 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4807 if (tp->snd_una == tp->write_seq) {
4808 tcp_update_metrics(sk);
4817 /* step 6: check the URG bit */
4818 tcp_urg(sk, skb, th);
4820 /* step 7: process the segment text */
4821 switch (sk->sk_state) {
4822 case TCP_CLOSE_WAIT:
4825 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4829 /* RFC 793 says to queue data in these states,
4830 * RFC 1122 says we MUST send a reset.
4831 * BSD 4.4 also does reset.
4833 if (sk->sk_shutdown & RCV_SHUTDOWN) {
4834 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4835 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
4836 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
4842 case TCP_ESTABLISHED:
4843 tcp_data_queue(sk, skb);
4848 /* tcp_data could move socket to TIME-WAIT */
4849 if (sk->sk_state != TCP_CLOSE) {
4850 tcp_data_snd_check(sk);
4851 tcp_ack_snd_check(sk);
4861 EXPORT_SYMBOL(sysctl_tcp_ecn);
4862 EXPORT_SYMBOL(sysctl_tcp_reordering);
4863 EXPORT_SYMBOL(tcp_cwnd_application_limited);
4864 EXPORT_SYMBOL(tcp_parse_options);
4865 EXPORT_SYMBOL(tcp_rcv_established);
4866 EXPORT_SYMBOL(tcp_rcv_state_process);