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 /* 5. Recalculate window clamp after socket hit its memory bounds. */
334 static void tcp_clamp_window(struct sock *sk, struct tcp_opt *tp)
337 unsigned int app_win = tp->rcv_nxt - tp->copied_seq;
342 skb_queue_walk(&tp->out_of_order_queue, skb) {
346 /* If overcommit is due to out of order segments,
347 * do not clamp window. Try to expand rcvbuf instead.
350 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
351 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
352 !tcp_memory_pressure &&
353 atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0])
354 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
357 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf) {
359 if (atomic_read(&sk->sk_rmem_alloc) >= 2 * sk->sk_rcvbuf)
361 if (app_win > tp->ack.rcv_mss)
362 app_win -= tp->ack.rcv_mss;
363 app_win = max(app_win, 2U*tp->advmss);
366 tp->window_clamp = min(tp->window_clamp, app_win);
367 tp->rcv_ssthresh = min(tp->window_clamp, 2U*tp->advmss);
371 /* Receiver "autotuning" code.
373 * The algorithm for RTT estimation w/o timestamps is based on
374 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
375 * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps>
377 * More detail on this code can be found at
378 * <http://www.psc.edu/~jheffner/senior_thesis.ps>,
379 * though this reference is out of date. A new paper
382 static void tcp_rcv_rtt_update(struct tcp_opt *tp, u32 sample, int win_dep)
384 u32 new_sample = tp->rcv_rtt_est.rtt;
390 if (new_sample != 0) {
391 /* If we sample in larger samples in the non-timestamp
392 * case, we could grossly overestimate the RTT especially
393 * with chatty applications or bulk transfer apps which
394 * are stalled on filesystem I/O.
396 * Also, since we are only going for a minimum in the
397 * non-timestamp case, we do not smoothe things out
398 * else with timestamps disabled convergance takes too
402 m -= (new_sample >> 3);
404 } else if (m < new_sample)
407 /* No previous mesaure. */
411 if (tp->rcv_rtt_est.rtt != new_sample)
412 tp->rcv_rtt_est.rtt = new_sample;
415 static inline void tcp_rcv_rtt_measure(struct tcp_opt *tp)
417 if (tp->rcv_rtt_est.time == 0)
419 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
421 tcp_rcv_rtt_update(tp,
422 jiffies - tp->rcv_rtt_est.time,
426 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
427 tp->rcv_rtt_est.time = tcp_time_stamp;
430 static inline void tcp_rcv_rtt_measure_ts(struct tcp_opt *tp, struct sk_buff *skb)
433 (TCP_SKB_CB(skb)->end_seq -
434 TCP_SKB_CB(skb)->seq >= tp->ack.rcv_mss))
435 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rcv_tsecr, 0);
439 * This function should be called every time data is copied to user space.
440 * It calculates the appropriate TCP receive buffer space.
442 void tcp_rcv_space_adjust(struct sock *sk)
444 struct tcp_opt *tp = tcp_sk(sk);
448 if (tp->rcvq_space.time == 0)
451 time = tcp_time_stamp - tp->rcvq_space.time;
452 if (time < (tp->rcv_rtt_est.rtt >> 3) ||
453 tp->rcv_rtt_est.rtt == 0)
456 space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
458 space = max(tp->rcvq_space.space, space);
460 if (tp->rcvq_space.space != space) {
463 tp->rcvq_space.space = space;
465 if (sysctl_tcp_moderate_rcvbuf) {
466 int new_clamp = space;
468 /* Receive space grows, normalize in order to
469 * take into account packet headers and sk_buff
470 * structure overhead.
475 rcvmem = (tp->advmss + MAX_TCP_HEADER +
476 16 + sizeof(struct sk_buff));
477 while (tcp_win_from_space(rcvmem) < tp->advmss)
480 space = min(space, sysctl_tcp_rmem[2]);
481 if (space > sk->sk_rcvbuf) {
482 sk->sk_rcvbuf = space;
484 /* Make the window clamp follow along. */
485 tp->window_clamp = new_clamp;
491 tp->rcvq_space.seq = tp->copied_seq;
492 tp->rcvq_space.time = tcp_time_stamp;
495 /* There is something which you must keep in mind when you analyze the
496 * behavior of the tp->ato delayed ack timeout interval. When a
497 * connection starts up, we want to ack as quickly as possible. The
498 * problem is that "good" TCP's do slow start at the beginning of data
499 * transmission. The means that until we send the first few ACK's the
500 * sender will sit on his end and only queue most of his data, because
501 * he can only send snd_cwnd unacked packets at any given time. For
502 * each ACK we send, he increments snd_cwnd and transmits more of his
505 static void tcp_event_data_recv(struct sock *sk, struct tcp_opt *tp, struct sk_buff *skb)
509 tcp_schedule_ack(tp);
511 tcp_measure_rcv_mss(tp, skb);
513 tcp_rcv_rtt_measure(tp);
515 now = tcp_time_stamp;
518 /* The _first_ data packet received, initialize
519 * delayed ACK engine.
521 tcp_incr_quickack(tp);
522 tp->ack.ato = TCP_ATO_MIN;
524 int m = now - tp->ack.lrcvtime;
526 if (m <= TCP_ATO_MIN/2) {
527 /* The fastest case is the first. */
528 tp->ack.ato = (tp->ack.ato>>1) + TCP_ATO_MIN/2;
529 } else if (m < tp->ack.ato) {
530 tp->ack.ato = (tp->ack.ato>>1) + m;
531 if (tp->ack.ato > tp->rto)
532 tp->ack.ato = tp->rto;
533 } else if (m > tp->rto) {
534 /* Too long gap. Apparently sender falled to
535 * restart window, so that we send ACKs quickly.
537 tcp_incr_quickack(tp);
541 tp->ack.lrcvtime = now;
543 TCP_ECN_check_ce(tp, skb);
546 tcp_grow_window(sk, tp, skb);
549 /* Set up a new TCP connection, depending on whether it should be
550 * using Vegas or not.
552 void tcp_vegas_init(struct tcp_opt *tp)
554 if (sysctl_tcp_vegas_cong_avoid) {
555 tp->vegas.do_vegas = 1;
556 tp->vegas.baseRTT = 0x7fffffff;
557 tcp_vegas_enable(tp);
559 tcp_vegas_disable(tp);
562 /* Do RTT sampling needed for Vegas.
564 * o min-filter RTT samples from within an RTT to get the current
565 * propagation delay + queuing delay (we are min-filtering to try to
566 * avoid the effects of delayed ACKs)
567 * o min-filter RTT samples from a much longer window (forever for now)
568 * to find the propagation delay (baseRTT)
570 static inline void vegas_rtt_calc(struct tcp_opt *tp, __u32 rtt)
572 __u32 vrtt = rtt + 1; /* Never allow zero rtt or baseRTT */
574 /* Filter to find propagation delay: */
575 if (vrtt < tp->vegas.baseRTT)
576 tp->vegas.baseRTT = vrtt;
578 /* Find the min RTT during the last RTT to find
579 * the current prop. delay + queuing delay:
581 tp->vegas.minRTT = min(tp->vegas.minRTT, vrtt);
585 /* Called to compute a smoothed rtt estimate. The data fed to this
586 * routine either comes from timestamps, or from segments that were
587 * known _not_ to have been retransmitted [see Karn/Partridge
588 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
589 * piece by Van Jacobson.
590 * NOTE: the next three routines used to be one big routine.
591 * To save cycles in the RFC 1323 implementation it was better to break
592 * it up into three procedures. -- erics
594 static void tcp_rtt_estimator(struct tcp_opt *tp, __u32 mrtt)
596 long m = mrtt; /* RTT */
598 if (tcp_vegas_enabled(tp))
599 vegas_rtt_calc(tp, mrtt);
601 /* The following amusing code comes from Jacobson's
602 * article in SIGCOMM '88. Note that rtt and mdev
603 * are scaled versions of rtt and mean deviation.
604 * This is designed to be as fast as possible
605 * m stands for "measurement".
607 * On a 1990 paper the rto value is changed to:
608 * RTO = rtt + 4 * mdev
610 * Funny. This algorithm seems to be very broken.
611 * These formulae increase RTO, when it should be decreased, increase
612 * too slowly, when it should be incresed fastly, decrease too fastly
613 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
614 * does not matter how to _calculate_ it. Seems, it was trap
615 * that VJ failed to avoid. 8)
620 m -= (tp->srtt >> 3); /* m is now error in rtt est */
621 tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */
623 m = -m; /* m is now abs(error) */
624 m -= (tp->mdev >> 2); /* similar update on mdev */
625 /* This is similar to one of Eifel findings.
626 * Eifel blocks mdev updates when rtt decreases.
627 * This solution is a bit different: we use finer gain
628 * for mdev in this case (alpha*beta).
629 * Like Eifel it also prevents growth of rto,
630 * but also it limits too fast rto decreases,
631 * happening in pure Eifel.
636 m -= (tp->mdev >> 2); /* similar update on mdev */
638 tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */
639 if (tp->mdev > tp->mdev_max) {
640 tp->mdev_max = tp->mdev;
641 if (tp->mdev_max > tp->rttvar)
642 tp->rttvar = tp->mdev_max;
644 if (after(tp->snd_una, tp->rtt_seq)) {
645 if (tp->mdev_max < tp->rttvar)
646 tp->rttvar -= (tp->rttvar-tp->mdev_max)>>2;
647 tp->rtt_seq = tp->snd_nxt;
648 tp->mdev_max = TCP_RTO_MIN;
651 /* no previous measure. */
652 tp->srtt = m<<3; /* take the measured time to be rtt */
653 tp->mdev = m<<1; /* make sure rto = 3*rtt */
654 tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN);
655 tp->rtt_seq = tp->snd_nxt;
658 tcp_westwood_update_rtt(tp, tp->srtt >> 3);
661 /* Calculate rto without backoff. This is the second half of Van Jacobson's
662 * routine referred to above.
664 static __inline__ void tcp_set_rto(struct tcp_opt *tp)
666 /* Old crap is replaced with new one. 8)
669 * 1. If rtt variance happened to be less 50msec, it is hallucination.
670 * It cannot be less due to utterly erratic ACK generation made
671 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
672 * to do with delayed acks, because at cwnd>2 true delack timeout
673 * is invisible. Actually, Linux-2.4 also generates erratic
674 * ACKs in some curcumstances.
676 tp->rto = (tp->srtt >> 3) + tp->rttvar;
678 /* 2. Fixups made earlier cannot be right.
679 * If we do not estimate RTO correctly without them,
680 * all the algo is pure shit and should be replaced
681 * with correct one. It is exaclty, which we pretend to do.
685 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
686 * guarantees that rto is higher.
688 static __inline__ void tcp_bound_rto(struct tcp_opt *tp)
690 if (tp->rto > TCP_RTO_MAX)
691 tp->rto = TCP_RTO_MAX;
694 /* Save metrics learned by this TCP session.
695 This function is called only, when TCP finishes successfully
696 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
698 void tcp_update_metrics(struct sock *sk)
700 struct tcp_opt *tp = tcp_sk(sk);
701 struct dst_entry *dst = __sk_dst_get(sk);
703 if (sysctl_tcp_nometrics_save)
708 if (dst && (dst->flags&DST_HOST)) {
711 if (tp->backoff || !tp->srtt) {
712 /* This session failed to estimate rtt. Why?
713 * Probably, no packets returned in time.
716 if (!(dst_metric_locked(dst, RTAX_RTT)))
717 dst->metrics[RTAX_RTT-1] = 0;
721 m = dst_metric(dst, RTAX_RTT) - tp->srtt;
723 /* If newly calculated rtt larger than stored one,
724 * store new one. Otherwise, use EWMA. Remember,
725 * rtt overestimation is always better than underestimation.
727 if (!(dst_metric_locked(dst, RTAX_RTT))) {
729 dst->metrics[RTAX_RTT-1] = tp->srtt;
731 dst->metrics[RTAX_RTT-1] -= (m>>3);
734 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) {
738 /* Scale deviation to rttvar fixed point */
743 if (m >= dst_metric(dst, RTAX_RTTVAR))
744 dst->metrics[RTAX_RTTVAR-1] = m;
746 dst->metrics[RTAX_RTTVAR-1] -=
747 (dst->metrics[RTAX_RTTVAR-1] - m)>>2;
750 if (tp->snd_ssthresh >= 0xFFFF) {
751 /* Slow start still did not finish. */
752 if (dst_metric(dst, RTAX_SSTHRESH) &&
753 !dst_metric_locked(dst, RTAX_SSTHRESH) &&
754 (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH))
755 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1;
756 if (!dst_metric_locked(dst, RTAX_CWND) &&
757 tp->snd_cwnd > dst_metric(dst, RTAX_CWND))
758 dst->metrics[RTAX_CWND-1] = tp->snd_cwnd;
759 } else if (tp->snd_cwnd > tp->snd_ssthresh &&
760 tp->ca_state == TCP_CA_Open) {
761 /* Cong. avoidance phase, cwnd is reliable. */
762 if (!dst_metric_locked(dst, RTAX_SSTHRESH))
763 dst->metrics[RTAX_SSTHRESH-1] =
764 max(tp->snd_cwnd >> 1, tp->snd_ssthresh);
765 if (!dst_metric_locked(dst, RTAX_CWND))
766 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_cwnd) >> 1;
768 /* Else slow start did not finish, cwnd is non-sense,
769 ssthresh may be also invalid.
771 if (!dst_metric_locked(dst, RTAX_CWND))
772 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_ssthresh) >> 1;
773 if (dst->metrics[RTAX_SSTHRESH-1] &&
774 !dst_metric_locked(dst, RTAX_SSTHRESH) &&
775 tp->snd_ssthresh > dst->metrics[RTAX_SSTHRESH-1])
776 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh;
779 if (!dst_metric_locked(dst, RTAX_REORDERING)) {
780 if (dst->metrics[RTAX_REORDERING-1] < tp->reordering &&
781 tp->reordering != sysctl_tcp_reordering)
782 dst->metrics[RTAX_REORDERING-1] = tp->reordering;
787 /* Numbers are taken from RFC2414. */
788 __u32 tcp_init_cwnd(struct tcp_opt *tp, struct dst_entry *dst)
790 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
793 if (tp->mss_cache > 1460)
796 cwnd = (tp->mss_cache > 1095) ? 3 : 4;
798 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
801 /* Initialize metrics on socket. */
803 static void tcp_init_metrics(struct sock *sk)
805 struct tcp_opt *tp = tcp_sk(sk);
806 struct dst_entry *dst = __sk_dst_get(sk);
813 if (dst_metric_locked(dst, RTAX_CWND))
814 tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND);
815 if (dst_metric(dst, RTAX_SSTHRESH)) {
816 tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH);
817 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
818 tp->snd_ssthresh = tp->snd_cwnd_clamp;
820 if (dst_metric(dst, RTAX_REORDERING) &&
821 tp->reordering != dst_metric(dst, RTAX_REORDERING)) {
823 tp->reordering = dst_metric(dst, RTAX_REORDERING);
826 if (dst_metric(dst, RTAX_RTT) == 0)
829 if (!tp->srtt && dst_metric(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3))
832 /* Initial rtt is determined from SYN,SYN-ACK.
833 * The segment is small and rtt may appear much
834 * less than real one. Use per-dst memory
835 * to make it more realistic.
837 * A bit of theory. RTT is time passed after "normal" sized packet
838 * is sent until it is ACKed. In normal curcumstances sending small
839 * packets force peer to delay ACKs and calculation is correct too.
840 * The algorithm is adaptive and, provided we follow specs, it
841 * NEVER underestimate RTT. BUT! If peer tries to make some clever
842 * tricks sort of "quick acks" for time long enough to decrease RTT
843 * to low value, and then abruptly stops to do it and starts to delay
844 * ACKs, wait for troubles.
846 if (dst_metric(dst, RTAX_RTT) > tp->srtt)
847 tp->srtt = dst_metric(dst, RTAX_RTT);
848 if (dst_metric(dst, RTAX_RTTVAR) > tp->mdev) {
849 tp->mdev = dst_metric(dst, RTAX_RTTVAR);
850 tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN);
854 if (tp->rto < TCP_TIMEOUT_INIT && !tp->saw_tstamp)
856 tp->snd_cwnd = tcp_init_cwnd(tp, dst);
857 tp->snd_cwnd_stamp = tcp_time_stamp;
861 /* Play conservative. If timestamps are not
862 * supported, TCP will fail to recalculate correct
863 * rtt, if initial rto is too small. FORGET ALL AND RESET!
865 if (!tp->saw_tstamp && tp->srtt) {
867 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT;
868 tp->rto = TCP_TIMEOUT_INIT;
872 static void tcp_update_reordering(struct tcp_opt *tp, int metric, int ts)
874 if (metric > tp->reordering) {
875 tp->reordering = min(TCP_MAX_REORDERING, metric);
877 /* This exciting event is worth to be remembered. 8) */
879 NET_INC_STATS_BH(TCPTSReorder);
881 NET_INC_STATS_BH(TCPRenoReorder);
883 NET_INC_STATS_BH(TCPFACKReorder);
885 NET_INC_STATS_BH(TCPSACKReorder);
886 #if FASTRETRANS_DEBUG > 1
887 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n",
888 tp->sack_ok, tp->ca_state,
889 tp->reordering, tp->fackets_out, tp->sacked_out,
890 tp->undo_marker ? tp->undo_retrans : 0);
892 /* Disable FACK yet. */
897 /* This procedure tags the retransmission queue when SACKs arrive.
899 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
900 * Packets in queue with these bits set are counted in variables
901 * sacked_out, retrans_out and lost_out, correspondingly.
903 * Valid combinations are:
904 * Tag InFlight Description
905 * 0 1 - orig segment is in flight.
906 * S 0 - nothing flies, orig reached receiver.
907 * L 0 - nothing flies, orig lost by net.
908 * R 2 - both orig and retransmit are in flight.
909 * L|R 1 - orig is lost, retransmit is in flight.
910 * S|R 1 - orig reached receiver, retrans is still in flight.
911 * (L|S|R is logically valid, it could occur when L|R is sacked,
912 * but it is equivalent to plain S and code short-curcuits it to S.
913 * L|S is logically invalid, it would mean -1 packet in flight 8))
915 * These 6 states form finite state machine, controlled by the following events:
916 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
917 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
918 * 3. Loss detection event of one of three flavors:
919 * A. Scoreboard estimator decided the packet is lost.
920 * A'. Reno "three dupacks" marks head of queue lost.
921 * A''. Its FACK modfication, head until snd.fack is lost.
922 * B. SACK arrives sacking data transmitted after never retransmitted
924 * C. SACK arrives sacking SND.NXT at the moment, when the
925 * segment was retransmitted.
926 * 4. D-SACK added new rule: D-SACK changes any tag to S.
928 * It is pleasant to note, that state diagram turns out to be commutative,
929 * so that we are allowed not to be bothered by order of our actions,
930 * when multiple events arrive simultaneously. (see the function below).
932 * Reordering detection.
933 * --------------------
934 * Reordering metric is maximal distance, which a packet can be displaced
935 * in packet stream. With SACKs we can estimate it:
937 * 1. SACK fills old hole and the corresponding segment was not
938 * ever retransmitted -> reordering. Alas, we cannot use it
939 * when segment was retransmitted.
940 * 2. The last flaw is solved with D-SACK. D-SACK arrives
941 * for retransmitted and already SACKed segment -> reordering..
942 * Both of these heuristics are not used in Loss state, when we cannot
943 * account for retransmits accurately.
946 tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, u32 prior_snd_una)
948 struct tcp_opt *tp = tcp_sk(sk);
949 unsigned char *ptr = ack_skb->h.raw + TCP_SKB_CB(ack_skb)->sacked;
950 struct tcp_sack_block *sp = (struct tcp_sack_block *)(ptr+2);
951 int num_sacks = (ptr[1] - TCPOLEN_SACK_BASE)>>3;
952 int reord = tp->packets_out;
954 u32 lost_retrans = 0;
958 /* So, SACKs for already sent large segments will be lost.
959 * Not good, but alternative is to resegment the queue. */
960 if (sk->sk_route_caps & NETIF_F_TSO) {
961 sk->sk_route_caps &= ~NETIF_F_TSO;
962 sk->sk_no_largesend = 1;
963 tp->mss_cache = tp->mss_cache_std;
968 prior_fackets = tp->fackets_out;
970 for (i=0; i<num_sacks; i++, sp++) {
972 __u32 start_seq = ntohl(sp->start_seq);
973 __u32 end_seq = ntohl(sp->end_seq);
977 /* Check for D-SACK. */
979 u32 ack = TCP_SKB_CB(ack_skb)->ack_seq;
981 if (before(start_seq, ack)) {
984 NET_INC_STATS_BH(TCPDSACKRecv);
985 } else if (num_sacks > 1 &&
986 !after(end_seq, ntohl(sp[1].end_seq)) &&
987 !before(start_seq, ntohl(sp[1].start_seq))) {
990 NET_INC_STATS_BH(TCPDSACKOfoRecv);
993 /* D-SACK for already forgotten data...
994 * Do dumb counting. */
996 !after(end_seq, prior_snd_una) &&
997 after(end_seq, tp->undo_marker))
1000 /* Eliminate too old ACKs, but take into
1001 * account more or less fresh ones, they can
1002 * contain valid SACK info.
1004 if (before(ack, prior_snd_una - tp->max_window))
1008 /* Event "B" in the comment above. */
1009 if (after(end_seq, tp->high_seq))
1010 flag |= FLAG_DATA_LOST;
1012 for_retrans_queue(skb, sk, tp) {
1013 u8 sacked = TCP_SKB_CB(skb)->sacked;
1016 /* The retransmission queue is always in order, so
1017 * we can short-circuit the walk early.
1019 if(!before(TCP_SKB_CB(skb)->seq, end_seq))
1024 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1025 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1027 /* Account D-SACK for retransmitted packet. */
1028 if ((dup_sack && in_sack) &&
1029 (sacked & TCPCB_RETRANS) &&
1030 after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker))
1033 /* The frame is ACKed. */
1034 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) {
1035 if (sacked&TCPCB_RETRANS) {
1036 if ((dup_sack && in_sack) &&
1037 (sacked&TCPCB_SACKED_ACKED))
1038 reord = min(fack_count, reord);
1040 /* If it was in a hole, we detected reordering. */
1041 if (fack_count < prior_fackets &&
1042 !(sacked&TCPCB_SACKED_ACKED))
1043 reord = min(fack_count, reord);
1046 /* Nothing to do; acked frame is about to be dropped. */
1050 if ((sacked&TCPCB_SACKED_RETRANS) &&
1051 after(end_seq, TCP_SKB_CB(skb)->ack_seq) &&
1052 (!lost_retrans || after(end_seq, lost_retrans)))
1053 lost_retrans = end_seq;
1058 if (!(sacked&TCPCB_SACKED_ACKED)) {
1059 if (sacked & TCPCB_SACKED_RETRANS) {
1060 /* If the segment is not tagged as lost,
1061 * we do not clear RETRANS, believing
1062 * that retransmission is still in flight.
1064 if (sacked & TCPCB_LOST) {
1065 TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1070 /* New sack for not retransmitted frame,
1071 * which was in hole. It is reordering.
1073 if (!(sacked & TCPCB_RETRANS) &&
1074 fack_count < prior_fackets)
1075 reord = min(fack_count, reord);
1077 if (sacked & TCPCB_LOST) {
1078 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1083 TCP_SKB_CB(skb)->sacked |= TCPCB_SACKED_ACKED;
1084 flag |= FLAG_DATA_SACKED;
1087 if (fack_count > tp->fackets_out)
1088 tp->fackets_out = fack_count;
1090 if (dup_sack && (sacked&TCPCB_RETRANS))
1091 reord = min(fack_count, reord);
1094 /* D-SACK. We can detect redundant retransmission
1095 * in S|R and plain R frames and clear it.
1096 * undo_retrans is decreased above, L|R frames
1097 * are accounted above as well.
1100 (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS)) {
1101 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1107 /* Check for lost retransmit. This superb idea is
1108 * borrowed from "ratehalving". Event "C".
1109 * Later note: FACK people cheated me again 8),
1110 * we have to account for reordering! Ugly,
1113 if (lost_retrans && tp->ca_state == TCP_CA_Recovery) {
1114 struct sk_buff *skb;
1116 for_retrans_queue(skb, sk, tp) {
1117 if (after(TCP_SKB_CB(skb)->seq, lost_retrans))
1119 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1121 if ((TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) &&
1122 after(lost_retrans, TCP_SKB_CB(skb)->ack_seq) &&
1124 !before(lost_retrans,
1125 TCP_SKB_CB(skb)->ack_seq + tp->reordering *
1127 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1130 if (!(TCP_SKB_CB(skb)->sacked&(TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1132 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1133 flag |= FLAG_DATA_SACKED;
1134 NET_INC_STATS_BH(TCPLostRetransmit);
1140 tp->left_out = tp->sacked_out + tp->lost_out;
1142 if (reord < tp->fackets_out && tp->ca_state != TCP_CA_Loss)
1143 tcp_update_reordering(tp, (tp->fackets_out + 1) - reord, 0);
1145 #if FASTRETRANS_DEBUG > 0
1146 BUG_TRAP((int)tp->sacked_out >= 0);
1147 BUG_TRAP((int)tp->lost_out >= 0);
1148 BUG_TRAP((int)tp->retrans_out >= 0);
1149 BUG_TRAP((int)tcp_packets_in_flight(tp) >= 0);
1154 /* RTO occurred, but do not yet enter loss state. Instead, transmit two new
1155 * segments to see from the next ACKs whether any data was really missing.
1156 * If the RTO was spurious, new ACKs should arrive.
1158 void tcp_enter_frto(struct sock *sk)
1160 struct tcp_opt *tp = tcp_sk(sk);
1161 struct sk_buff *skb;
1163 tp->frto_counter = 1;
1165 if (tp->ca_state <= TCP_CA_Disorder ||
1166 tp->snd_una == tp->high_seq ||
1167 (tp->ca_state == TCP_CA_Loss && !tp->retransmits)) {
1168 tp->prior_ssthresh = tcp_current_ssthresh(tp);
1169 if (!tcp_westwood_ssthresh(tp))
1170 tp->snd_ssthresh = tcp_recalc_ssthresh(tp);
1173 /* Have to clear retransmission markers here to keep the bookkeeping
1174 * in shape, even though we are not yet in Loss state.
1175 * If something was really lost, it is eventually caught up
1176 * in tcp_enter_frto_loss.
1178 tp->retrans_out = 0;
1179 tp->undo_marker = tp->snd_una;
1180 tp->undo_retrans = 0;
1182 for_retrans_queue(skb, sk, tp) {
1183 TCP_SKB_CB(skb)->sacked &= ~TCPCB_RETRANS;
1185 tcp_sync_left_out(tp);
1187 tcp_set_ca_state(tp, TCP_CA_Open);
1188 tp->frto_highmark = tp->snd_nxt;
1191 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1192 * which indicates that we should follow the traditional RTO recovery,
1193 * i.e. mark everything lost and do go-back-N retransmission.
1195 static void tcp_enter_frto_loss(struct sock *sk)
1197 struct tcp_opt *tp = tcp_sk(sk);
1198 struct sk_buff *skb;
1203 tp->fackets_out = 0;
1205 for_retrans_queue(skb, sk, tp) {
1207 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1208 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) {
1210 /* Do not mark those segments lost that were
1211 * forward transmitted after RTO
1213 if(!after(TCP_SKB_CB(skb)->end_seq,
1214 tp->frto_highmark)) {
1215 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1220 tp->fackets_out = cnt;
1223 tcp_sync_left_out(tp);
1225 tp->snd_cwnd = tp->frto_counter + tcp_packets_in_flight(tp)+1;
1226 tp->snd_cwnd_cnt = 0;
1227 tp->snd_cwnd_stamp = tcp_time_stamp;
1228 tp->undo_marker = 0;
1229 tp->frto_counter = 0;
1231 tp->reordering = min_t(unsigned int, tp->reordering,
1232 sysctl_tcp_reordering);
1233 tcp_set_ca_state(tp, TCP_CA_Loss);
1234 tp->high_seq = tp->frto_highmark;
1235 TCP_ECN_queue_cwr(tp);
1238 void tcp_clear_retrans(struct tcp_opt *tp)
1241 tp->retrans_out = 0;
1243 tp->fackets_out = 0;
1247 tp->undo_marker = 0;
1248 tp->undo_retrans = 0;
1251 /* Enter Loss state. If "how" is not zero, forget all SACK information
1252 * and reset tags completely, otherwise preserve SACKs. If receiver
1253 * dropped its ofo queue, we will know this due to reneging detection.
1255 void tcp_enter_loss(struct sock *sk, int how)
1257 struct tcp_opt *tp = tcp_sk(sk);
1258 struct sk_buff *skb;
1261 /* Reduce ssthresh if it has not yet been made inside this window. */
1262 if (tp->ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
1263 (tp->ca_state == TCP_CA_Loss && !tp->retransmits)) {
1264 tp->prior_ssthresh = tcp_current_ssthresh(tp);
1265 tp->snd_ssthresh = tcp_recalc_ssthresh(tp);
1268 tp->snd_cwnd_cnt = 0;
1269 tp->snd_cwnd_stamp = tcp_time_stamp;
1271 tcp_clear_retrans(tp);
1273 /* Push undo marker, if it was plain RTO and nothing
1274 * was retransmitted. */
1276 tp->undo_marker = tp->snd_una;
1278 for_retrans_queue(skb, sk, tp) {
1280 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS)
1281 tp->undo_marker = 0;
1282 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1283 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
1284 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1285 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1289 tp->fackets_out = cnt;
1292 tcp_sync_left_out(tp);
1294 tp->reordering = min_t(unsigned int, tp->reordering,
1295 sysctl_tcp_reordering);
1296 tcp_set_ca_state(tp, TCP_CA_Loss);
1297 tp->high_seq = tp->snd_nxt;
1298 TCP_ECN_queue_cwr(tp);
1301 static int tcp_check_sack_reneging(struct sock *sk, struct tcp_opt *tp)
1303 struct sk_buff *skb;
1305 /* If ACK arrived pointing to a remembered SACK,
1306 * it means that our remembered SACKs do not reflect
1307 * real state of receiver i.e.
1308 * receiver _host_ is heavily congested (or buggy).
1309 * Do processing similar to RTO timeout.
1311 if ((skb = skb_peek(&sk->sk_write_queue)) != NULL &&
1312 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
1313 NET_INC_STATS_BH(TCPSACKReneging);
1315 tcp_enter_loss(sk, 1);
1317 tcp_retransmit_skb(sk, skb_peek(&sk->sk_write_queue));
1318 tcp_reset_xmit_timer(sk, TCP_TIME_RETRANS, tp->rto);
1324 static inline int tcp_fackets_out(struct tcp_opt *tp)
1326 return IsReno(tp) ? tp->sacked_out+1 : tp->fackets_out;
1329 static inline int tcp_skb_timedout(struct tcp_opt *tp, struct sk_buff *skb)
1331 return (tcp_time_stamp - TCP_SKB_CB(skb)->when > tp->rto);
1334 static inline int tcp_head_timedout(struct sock *sk, struct tcp_opt *tp)
1336 return tp->packets_out &&
1337 tcp_skb_timedout(tp, skb_peek(&sk->sk_write_queue));
1340 /* Linux NewReno/SACK/FACK/ECN state machine.
1341 * --------------------------------------
1343 * "Open" Normal state, no dubious events, fast path.
1344 * "Disorder" In all the respects it is "Open",
1345 * but requires a bit more attention. It is entered when
1346 * we see some SACKs or dupacks. It is split of "Open"
1347 * mainly to move some processing from fast path to slow one.
1348 * "CWR" CWND was reduced due to some Congestion Notification event.
1349 * It can be ECN, ICMP source quench, local device congestion.
1350 * "Recovery" CWND was reduced, we are fast-retransmitting.
1351 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
1353 * tcp_fastretrans_alert() is entered:
1354 * - each incoming ACK, if state is not "Open"
1355 * - when arrived ACK is unusual, namely:
1360 * Counting packets in flight is pretty simple.
1362 * in_flight = packets_out - left_out + retrans_out
1364 * packets_out is SND.NXT-SND.UNA counted in packets.
1366 * retrans_out is number of retransmitted segments.
1368 * left_out is number of segments left network, but not ACKed yet.
1370 * left_out = sacked_out + lost_out
1372 * sacked_out: Packets, which arrived to receiver out of order
1373 * and hence not ACKed. With SACKs this number is simply
1374 * amount of SACKed data. Even without SACKs
1375 * it is easy to give pretty reliable estimate of this number,
1376 * counting duplicate ACKs.
1378 * lost_out: Packets lost by network. TCP has no explicit
1379 * "loss notification" feedback from network (for now).
1380 * It means that this number can be only _guessed_.
1381 * Actually, it is the heuristics to predict lossage that
1382 * distinguishes different algorithms.
1384 * F.e. after RTO, when all the queue is considered as lost,
1385 * lost_out = packets_out and in_flight = retrans_out.
1387 * Essentially, we have now two algorithms counting
1390 * FACK: It is the simplest heuristics. As soon as we decided
1391 * that something is lost, we decide that _all_ not SACKed
1392 * packets until the most forward SACK are lost. I.e.
1393 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
1394 * It is absolutely correct estimate, if network does not reorder
1395 * packets. And it loses any connection to reality when reordering
1396 * takes place. We use FACK by default until reordering
1397 * is suspected on the path to this destination.
1399 * NewReno: when Recovery is entered, we assume that one segment
1400 * is lost (classic Reno). While we are in Recovery and
1401 * a partial ACK arrives, we assume that one more packet
1402 * is lost (NewReno). This heuristics are the same in NewReno
1405 * Imagine, that's all! Forget about all this shamanism about CWND inflation
1406 * deflation etc. CWND is real congestion window, never inflated, changes
1407 * only according to classic VJ rules.
1409 * Really tricky (and requiring careful tuning) part of algorithm
1410 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
1411 * The first determines the moment _when_ we should reduce CWND and,
1412 * hence, slow down forward transmission. In fact, it determines the moment
1413 * when we decide that hole is caused by loss, rather than by a reorder.
1415 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
1416 * holes, caused by lost packets.
1418 * And the most logically complicated part of algorithm is undo
1419 * heuristics. We detect false retransmits due to both too early
1420 * fast retransmit (reordering) and underestimated RTO, analyzing
1421 * timestamps and D-SACKs. When we detect that some segments were
1422 * retransmitted by mistake and CWND reduction was wrong, we undo
1423 * window reduction and abort recovery phase. This logic is hidden
1424 * inside several functions named tcp_try_undo_<something>.
1427 /* This function decides, when we should leave Disordered state
1428 * and enter Recovery phase, reducing congestion window.
1430 * Main question: may we further continue forward transmission
1431 * with the same cwnd?
1434 tcp_time_to_recover(struct sock *sk, struct tcp_opt *tp)
1436 /* Trick#1: The loss is proven. */
1440 /* Not-A-Trick#2 : Classic rule... */
1441 if (tcp_fackets_out(tp) > tp->reordering)
1444 /* Trick#3 : when we use RFC2988 timer restart, fast
1445 * retransmit can be triggered by timeout of queue head.
1447 if (tcp_head_timedout(sk, tp))
1450 /* Trick#4: It is still not OK... But will it be useful to delay
1453 if (tp->packets_out <= tp->reordering &&
1454 tp->sacked_out >= max_t(__u32, tp->packets_out/2, sysctl_tcp_reordering) &&
1455 !tcp_may_send_now(sk, tp)) {
1456 /* We have nothing to send. This connection is limited
1457 * either by receiver window or by application.
1465 /* If we receive more dupacks than we expected counting segments
1466 * in assumption of absent reordering, interpret this as reordering.
1467 * The only another reason could be bug in receiver TCP.
1469 static void tcp_check_reno_reordering(struct tcp_opt *tp, int addend)
1473 holes = max(tp->lost_out, 1U);
1474 holes = min(holes, tp->packets_out);
1476 if (tp->sacked_out + holes > tp->packets_out) {
1477 tp->sacked_out = tp->packets_out - holes;
1478 tcp_update_reordering(tp, tp->packets_out+addend, 0);
1482 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1484 static void tcp_add_reno_sack(struct tcp_opt *tp)
1487 tcp_check_reno_reordering(tp, 0);
1488 tcp_sync_left_out(tp);
1491 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1493 static void tcp_remove_reno_sacks(struct sock *sk, struct tcp_opt *tp, int acked)
1496 /* One ACK acked hole. The rest eat duplicate ACKs. */
1497 if (acked-1 >= tp->sacked_out)
1500 tp->sacked_out -= acked-1;
1502 tcp_check_reno_reordering(tp, acked);
1503 tcp_sync_left_out(tp);
1506 static inline void tcp_reset_reno_sack(struct tcp_opt *tp)
1509 tp->left_out = tp->lost_out;
1512 /* Mark head of queue up as lost. */
1514 tcp_mark_head_lost(struct sock *sk, struct tcp_opt *tp, int packets, u32 high_seq)
1516 struct sk_buff *skb;
1519 BUG_TRAP(cnt <= tp->packets_out);
1521 for_retrans_queue(skb, sk, tp) {
1522 if (--cnt < 0 || after(TCP_SKB_CB(skb)->end_seq, high_seq))
1524 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
1525 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1529 tcp_sync_left_out(tp);
1532 /* Account newly detected lost packet(s) */
1534 static void tcp_update_scoreboard(struct sock *sk, struct tcp_opt *tp)
1537 int lost = tp->fackets_out - tp->reordering;
1540 tcp_mark_head_lost(sk, tp, lost, tp->high_seq);
1542 tcp_mark_head_lost(sk, tp, 1, tp->high_seq);
1545 /* New heuristics: it is possible only after we switched
1546 * to restart timer each time when something is ACKed.
1547 * Hence, we can detect timed out packets during fast
1548 * retransmit without falling to slow start.
1550 if (tcp_head_timedout(sk, tp)) {
1551 struct sk_buff *skb;
1553 for_retrans_queue(skb, sk, tp) {
1554 if (tcp_skb_timedout(tp, skb) &&
1555 !(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
1556 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1560 tcp_sync_left_out(tp);
1564 /* CWND moderation, preventing bursts due to too big ACKs
1565 * in dubious situations.
1567 static __inline__ void tcp_moderate_cwnd(struct tcp_opt *tp)
1569 tp->snd_cwnd = min(tp->snd_cwnd,
1570 tcp_packets_in_flight(tp)+tcp_max_burst(tp));
1571 tp->snd_cwnd_stamp = tcp_time_stamp;
1574 /* Decrease cwnd each second ack. */
1576 static void tcp_cwnd_down(struct tcp_opt *tp)
1578 int decr = tp->snd_cwnd_cnt + 1;
1583 * Here limit is evaluated as BWestimation*RTTmin (for obtaining it
1584 * in packets we use mss_cache). If sysctl_tcp_westwood is off
1585 * tcp_westwood_bw_rttmin() returns 0. In such case snd_ssthresh is
1586 * still used as usual. It prevents other strange cases in which
1587 * BWE*RTTmin could assume value 0. It should not happen but...
1590 if (!(limit = tcp_westwood_bw_rttmin(tp)))
1591 limit = tp->snd_ssthresh/2;
1593 tp->snd_cwnd_cnt = decr&1;
1596 if (decr && tp->snd_cwnd > limit)
1597 tp->snd_cwnd -= decr;
1599 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp)+1);
1600 tp->snd_cwnd_stamp = tcp_time_stamp;
1603 /* Nothing was retransmitted or returned timestamp is less
1604 * than timestamp of the first retransmission.
1606 static __inline__ int tcp_packet_delayed(struct tcp_opt *tp)
1608 return !tp->retrans_stamp ||
1609 (tp->saw_tstamp && tp->rcv_tsecr &&
1610 (__s32)(tp->rcv_tsecr - tp->retrans_stamp) < 0);
1613 /* Undo procedures. */
1615 #if FASTRETRANS_DEBUG > 1
1616 static void DBGUNDO(struct sock *sk, struct tcp_opt *tp, const char *msg)
1618 struct inet_opt *inet = inet_sk(sk);
1619 printk(KERN_DEBUG "Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n",
1621 NIPQUAD(inet->daddr), ntohs(inet->dport),
1622 tp->snd_cwnd, tp->left_out,
1623 tp->snd_ssthresh, tp->prior_ssthresh, tp->packets_out);
1626 #define DBGUNDO(x...) do { } while (0)
1629 static void tcp_undo_cwr(struct tcp_opt *tp, int undo)
1631 if (tp->prior_ssthresh) {
1632 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh<<1);
1634 if (undo && tp->prior_ssthresh > tp->snd_ssthresh) {
1635 tp->snd_ssthresh = tp->prior_ssthresh;
1636 TCP_ECN_withdraw_cwr(tp);
1639 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
1641 tcp_moderate_cwnd(tp);
1642 tp->snd_cwnd_stamp = tcp_time_stamp;
1645 static inline int tcp_may_undo(struct tcp_opt *tp)
1647 return tp->undo_marker &&
1648 (!tp->undo_retrans || tcp_packet_delayed(tp));
1651 /* People celebrate: "We love our President!" */
1652 static int tcp_try_undo_recovery(struct sock *sk, struct tcp_opt *tp)
1654 if (tcp_may_undo(tp)) {
1655 /* Happy end! We did not retransmit anything
1656 * or our original transmission succeeded.
1658 DBGUNDO(sk, tp, tp->ca_state == TCP_CA_Loss ? "loss" : "retrans");
1659 tcp_undo_cwr(tp, 1);
1660 if (tp->ca_state == TCP_CA_Loss)
1661 NET_INC_STATS_BH(TCPLossUndo);
1663 NET_INC_STATS_BH(TCPFullUndo);
1664 tp->undo_marker = 0;
1666 if (tp->snd_una == tp->high_seq && IsReno(tp)) {
1667 /* Hold old state until something *above* high_seq
1668 * is ACKed. For Reno it is MUST to prevent false
1669 * fast retransmits (RFC2582). SACK TCP is safe. */
1670 tcp_moderate_cwnd(tp);
1673 tcp_set_ca_state(tp, TCP_CA_Open);
1677 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
1678 static void tcp_try_undo_dsack(struct sock *sk, struct tcp_opt *tp)
1680 if (tp->undo_marker && !tp->undo_retrans) {
1681 DBGUNDO(sk, tp, "D-SACK");
1682 tcp_undo_cwr(tp, 1);
1683 tp->undo_marker = 0;
1684 NET_INC_STATS_BH(TCPDSACKUndo);
1688 /* Undo during fast recovery after partial ACK. */
1690 static int tcp_try_undo_partial(struct sock *sk, struct tcp_opt *tp, int acked)
1692 /* Partial ACK arrived. Force Hoe's retransmit. */
1693 int failed = IsReno(tp) || tp->fackets_out>tp->reordering;
1695 if (tcp_may_undo(tp)) {
1696 /* Plain luck! Hole if filled with delayed
1697 * packet, rather than with a retransmit.
1699 if (tp->retrans_out == 0)
1700 tp->retrans_stamp = 0;
1702 tcp_update_reordering(tp, tcp_fackets_out(tp)+acked, 1);
1704 DBGUNDO(sk, tp, "Hoe");
1705 tcp_undo_cwr(tp, 0);
1706 NET_INC_STATS_BH(TCPPartialUndo);
1708 /* So... Do not make Hoe's retransmit yet.
1709 * If the first packet was delayed, the rest
1710 * ones are most probably delayed as well.
1717 /* Undo during loss recovery after partial ACK. */
1718 static int tcp_try_undo_loss(struct sock *sk, struct tcp_opt *tp)
1720 if (tcp_may_undo(tp)) {
1721 struct sk_buff *skb;
1722 for_retrans_queue(skb, sk, tp) {
1723 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1725 DBGUNDO(sk, tp, "partial loss");
1727 tp->left_out = tp->sacked_out;
1728 tcp_undo_cwr(tp, 1);
1729 NET_INC_STATS_BH(TCPLossUndo);
1730 tp->retransmits = 0;
1731 tp->undo_marker = 0;
1733 tcp_set_ca_state(tp, TCP_CA_Open);
1739 static __inline__ void tcp_complete_cwr(struct tcp_opt *tp)
1741 if (tcp_westwood_cwnd(tp))
1742 tp->snd_ssthresh = tp->snd_cwnd;
1744 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
1745 tp->snd_cwnd_stamp = tcp_time_stamp;
1748 static void tcp_try_to_open(struct sock *sk, struct tcp_opt *tp, int flag)
1750 tp->left_out = tp->sacked_out;
1752 if (tp->retrans_out == 0)
1753 tp->retrans_stamp = 0;
1758 if (tp->ca_state != TCP_CA_CWR) {
1759 int state = TCP_CA_Open;
1764 state = TCP_CA_Disorder;
1766 if (tp->ca_state != state) {
1767 tcp_set_ca_state(tp, state);
1768 tp->high_seq = tp->snd_nxt;
1770 tcp_moderate_cwnd(tp);
1776 /* Process an event, which can update packets-in-flight not trivially.
1777 * Main goal of this function is to calculate new estimate for left_out,
1778 * taking into account both packets sitting in receiver's buffer and
1779 * packets lost by network.
1781 * Besides that it does CWND reduction, when packet loss is detected
1782 * and changes state of machine.
1784 * It does _not_ decide what to send, it is made in function
1785 * tcp_xmit_retransmit_queue().
1788 tcp_fastretrans_alert(struct sock *sk, u32 prior_snd_una,
1789 int prior_packets, int flag)
1791 struct tcp_opt *tp = tcp_sk(sk);
1792 int is_dupack = (tp->snd_una == prior_snd_una && !(flag&FLAG_NOT_DUP));
1794 /* Some technical things:
1795 * 1. Reno does not count dupacks (sacked_out) automatically. */
1796 if (!tp->packets_out)
1798 /* 2. SACK counts snd_fack in packets inaccurately. */
1799 if (tp->sacked_out == 0)
1800 tp->fackets_out = 0;
1802 /* Now state machine starts.
1803 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
1805 tp->prior_ssthresh = 0;
1807 /* B. In all the states check for reneging SACKs. */
1808 if (tp->sacked_out && tcp_check_sack_reneging(sk, tp))
1811 /* C. Process data loss notification, provided it is valid. */
1812 if ((flag&FLAG_DATA_LOST) &&
1813 before(tp->snd_una, tp->high_seq) &&
1814 tp->ca_state != TCP_CA_Open &&
1815 tp->fackets_out > tp->reordering) {
1816 tcp_mark_head_lost(sk, tp, tp->fackets_out-tp->reordering, tp->high_seq);
1817 NET_INC_STATS_BH(TCPLoss);
1820 /* D. Synchronize left_out to current state. */
1821 tcp_sync_left_out(tp);
1823 /* E. Check state exit conditions. State can be terminated
1824 * when high_seq is ACKed. */
1825 if (tp->ca_state == TCP_CA_Open) {
1826 if (!sysctl_tcp_frto)
1827 BUG_TRAP(tp->retrans_out == 0);
1828 tp->retrans_stamp = 0;
1829 } else if (!before(tp->snd_una, tp->high_seq)) {
1830 switch (tp->ca_state) {
1832 tp->retransmits = 0;
1833 if (tcp_try_undo_recovery(sk, tp))
1838 /* CWR is to be held something *above* high_seq
1839 * is ACKed for CWR bit to reach receiver. */
1840 if (tp->snd_una != tp->high_seq) {
1841 tcp_complete_cwr(tp);
1842 tcp_set_ca_state(tp, TCP_CA_Open);
1846 case TCP_CA_Disorder:
1847 tcp_try_undo_dsack(sk, tp);
1848 if (!tp->undo_marker ||
1849 /* For SACK case do not Open to allow to undo
1850 * catching for all duplicate ACKs. */
1851 IsReno(tp) || tp->snd_una != tp->high_seq) {
1852 tp->undo_marker = 0;
1853 tcp_set_ca_state(tp, TCP_CA_Open);
1857 case TCP_CA_Recovery:
1859 tcp_reset_reno_sack(tp);
1860 if (tcp_try_undo_recovery(sk, tp))
1862 tcp_complete_cwr(tp);
1867 /* F. Process state. */
1868 switch (tp->ca_state) {
1869 case TCP_CA_Recovery:
1870 if (prior_snd_una == tp->snd_una) {
1871 if (IsReno(tp) && is_dupack)
1872 tcp_add_reno_sack(tp);
1874 int acked = prior_packets - tp->packets_out;
1876 tcp_remove_reno_sacks(sk, tp, acked);
1877 is_dupack = tcp_try_undo_partial(sk, tp, acked);
1881 if (flag&FLAG_DATA_ACKED)
1882 tp->retransmits = 0;
1883 if (!tcp_try_undo_loss(sk, tp)) {
1884 tcp_moderate_cwnd(tp);
1885 tcp_xmit_retransmit_queue(sk);
1888 if (tp->ca_state != TCP_CA_Open)
1890 /* Loss is undone; fall through to processing in Open state. */
1893 if (tp->snd_una != prior_snd_una)
1894 tcp_reset_reno_sack(tp);
1896 tcp_add_reno_sack(tp);
1899 if (tp->ca_state == TCP_CA_Disorder)
1900 tcp_try_undo_dsack(sk, tp);
1902 if (!tcp_time_to_recover(sk, tp)) {
1903 tcp_try_to_open(sk, tp, flag);
1907 /* Otherwise enter Recovery state */
1910 NET_INC_STATS_BH(TCPRenoRecovery);
1912 NET_INC_STATS_BH(TCPSackRecovery);
1914 tp->high_seq = tp->snd_nxt;
1915 tp->prior_ssthresh = 0;
1916 tp->undo_marker = tp->snd_una;
1917 tp->undo_retrans = tp->retrans_out;
1919 if (tp->ca_state < TCP_CA_CWR) {
1920 if (!(flag&FLAG_ECE))
1921 tp->prior_ssthresh = tcp_current_ssthresh(tp);
1922 tp->snd_ssthresh = tcp_recalc_ssthresh(tp);
1923 TCP_ECN_queue_cwr(tp);
1926 tp->snd_cwnd_cnt = 0;
1927 tcp_set_ca_state(tp, TCP_CA_Recovery);
1930 if (is_dupack || tcp_head_timedout(sk, tp))
1931 tcp_update_scoreboard(sk, tp);
1933 tcp_xmit_retransmit_queue(sk);
1936 /* Read draft-ietf-tcplw-high-performance before mucking
1937 * with this code. (Superceeds RFC1323)
1939 static void tcp_ack_saw_tstamp(struct tcp_opt *tp, int flag)
1943 /* RTTM Rule: A TSecr value received in a segment is used to
1944 * update the averaged RTT measurement only if the segment
1945 * acknowledges some new data, i.e., only if it advances the
1946 * left edge of the send window.
1948 * See draft-ietf-tcplw-high-performance-00, section 3.3.
1949 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
1951 * Changed: reset backoff as soon as we see the first valid sample.
1952 * If we do not, we get strongly overstimated rto. With timestamps
1953 * samples are accepted even from very old segments: f.e., when rtt=1
1954 * increases to 8, we retransmit 5 times and after 8 seconds delayed
1955 * answer arrives rto becomes 120 seconds! If at least one of segments
1956 * in window is lost... Voila. --ANK (010210)
1958 seq_rtt = tcp_time_stamp - tp->rcv_tsecr;
1959 tcp_rtt_estimator(tp, seq_rtt);
1965 static void tcp_ack_no_tstamp(struct tcp_opt *tp, u32 seq_rtt, int flag)
1967 /* We don't have a timestamp. Can only use
1968 * packets that are not retransmitted to determine
1969 * rtt estimates. Also, we must not reset the
1970 * backoff for rto until we get a non-retransmitted
1971 * packet. This allows us to deal with a situation
1972 * where the network delay has increased suddenly.
1973 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
1976 if (flag & FLAG_RETRANS_DATA_ACKED)
1979 tcp_rtt_estimator(tp, seq_rtt);
1985 static __inline__ void
1986 tcp_ack_update_rtt(struct tcp_opt *tp, int flag, s32 seq_rtt)
1988 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
1989 if (tp->saw_tstamp && tp->rcv_tsecr)
1990 tcp_ack_saw_tstamp(tp, flag);
1991 else if (seq_rtt >= 0)
1992 tcp_ack_no_tstamp(tp, seq_rtt, flag);
1996 * Compute congestion window to use.
1998 * This is from the implementation of BICTCP in
1999 * Lison-Xu, Kahaled Harfoush, and Injog Rhee.
2000 * "Binary Increase Congestion Control for Fast, Long Distance
2001 * Networks" in InfoComm 2004
2003 * http://www.csc.ncsu.edu/faculty/rhee/export/bitcp.pdf
2005 * Unless BIC is enabled and congestion window is large
2006 * this behaves the same as the original Reno.
2008 static inline __u32 bictcp_cwnd(struct tcp_opt *tp)
2010 /* orignal Reno behaviour */
2011 if (!sysctl_tcp_bic)
2012 return tp->snd_cwnd;
2014 if (tp->bictcp.last_cwnd == tp->snd_cwnd)
2015 return tp->bictcp.cnt; /* same cwnd, no update */
2017 tp->bictcp.last_cwnd = tp->snd_cwnd;
2019 /* start off normal */
2020 if (tp->snd_cwnd <= sysctl_tcp_bic_low_window)
2021 tp->bictcp.cnt = tp->snd_cwnd;
2023 /* binary increase */
2024 else if (tp->snd_cwnd < tp->bictcp.last_max_cwnd) {
2025 __u32 dist = (tp->bictcp.last_max_cwnd - tp->snd_cwnd)
2028 if (dist > BICTCP_MAX_INCREMENT)
2029 /* linear increase */
2030 tp->bictcp.cnt = tp->snd_cwnd / BICTCP_MAX_INCREMENT;
2031 else if (dist <= 1U)
2032 /* binary search increase */
2033 tp->bictcp.cnt = tp->snd_cwnd * BICTCP_FUNC_OF_MIN_INCR
2036 /* binary search increase */
2037 tp->bictcp.cnt = tp->snd_cwnd / dist;
2039 /* slow start amd linear increase */
2040 if (tp->snd_cwnd < tp->bictcp.last_max_cwnd + BICTCP_B)
2042 tp->bictcp.cnt = tp->snd_cwnd * BICTCP_FUNC_OF_MIN_INCR
2044 else if (tp->snd_cwnd < tp->bictcp.last_max_cwnd
2045 + BICTCP_MAX_INCREMENT*(BICTCP_B-1))
2047 tp->bictcp.cnt = tp->snd_cwnd * (BICTCP_B-1)
2048 / (tp->snd_cwnd-tp->bictcp.last_max_cwnd);
2050 /* linear increase */
2051 tp->bictcp.cnt = tp->snd_cwnd / BICTCP_MAX_INCREMENT;
2053 return tp->bictcp.cnt;
2056 /* This is Jacobson's slow start and congestion avoidance.
2057 * SIGCOMM '88, p. 328.
2059 static __inline__ void reno_cong_avoid(struct tcp_opt *tp)
2061 if (tp->snd_cwnd <= tp->snd_ssthresh) {
2062 /* In "safe" area, increase. */
2063 if (tp->snd_cwnd < tp->snd_cwnd_clamp)
2066 /* In dangerous area, increase slowly.
2067 * In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd
2069 if (tp->snd_cwnd_cnt >= bictcp_cwnd(tp)) {
2070 if (tp->snd_cwnd < tp->snd_cwnd_clamp)
2076 tp->snd_cwnd_stamp = tcp_time_stamp;
2079 /* This is based on the congestion detection/avoidance scheme described in
2080 * Lawrence S. Brakmo and Larry L. Peterson.
2081 * "TCP Vegas: End to end congestion avoidance on a global internet."
2082 * IEEE Journal on Selected Areas in Communication, 13(8):1465--1480,
2083 * October 1995. Available from:
2084 * ftp://ftp.cs.arizona.edu/xkernel/Papers/jsac.ps
2086 * See http://www.cs.arizona.edu/xkernel/ for their implementation.
2087 * The main aspects that distinguish this implementation from the
2088 * Arizona Vegas implementation are:
2089 * o We do not change the loss detection or recovery mechanisms of
2090 * Linux in any way. Linux already recovers from losses quite well,
2091 * using fine-grained timers, NewReno, and FACK.
2092 * o To avoid the performance penalty imposed by increasing cwnd
2093 * only every-other RTT during slow start, we increase during
2094 * every RTT during slow start, just like Reno.
2095 * o Largely to allow continuous cwnd growth during slow start,
2096 * we use the rate at which ACKs come back as the "actual"
2097 * rate, rather than the rate at which data is sent.
2098 * o To speed convergence to the right rate, we set the cwnd
2099 * to achieve the right ("actual") rate when we exit slow start.
2100 * o To filter out the noise caused by delayed ACKs, we use the
2101 * minimum RTT sample observed during the last RTT to calculate
2103 * o When the sender re-starts from idle, it waits until it has
2104 * received ACKs for an entire flight of new data before making
2105 * a cwnd adjustment decision. The original Vegas implementation
2106 * assumed senders never went idle.
2108 static void vegas_cong_avoid(struct tcp_opt *tp, u32 ack, u32 seq_rtt)
2110 /* The key players are v_beg_snd_una and v_beg_snd_nxt.
2112 * These are so named because they represent the approximate values
2113 * of snd_una and snd_nxt at the beginning of the current RTT. More
2114 * precisely, they represent the amount of data sent during the RTT.
2115 * At the end of the RTT, when we receive an ACK for v_beg_snd_nxt,
2116 * we will calculate that (v_beg_snd_nxt - v_beg_snd_una) outstanding
2117 * bytes of data have been ACKed during the course of the RTT, giving
2118 * an "actual" rate of:
2120 * (v_beg_snd_nxt - v_beg_snd_una) / (rtt duration)
2122 * Unfortunately, v_beg_snd_una is not exactly equal to snd_una,
2123 * because delayed ACKs can cover more than one segment, so they
2124 * don't line up nicely with the boundaries of RTTs.
2126 * Another unfortunate fact of life is that delayed ACKs delay the
2127 * advance of the left edge of our send window, so that the number
2128 * of bytes we send in an RTT is often less than our cwnd will allow.
2129 * So we keep track of our cwnd separately, in v_beg_snd_cwnd.
2132 if (after(ack, tp->vegas.beg_snd_nxt)) {
2133 /* Do the Vegas once-per-RTT cwnd adjustment. */
2134 u32 old_wnd, old_snd_cwnd;
2137 /* Here old_wnd is essentially the window of data that was
2138 * sent during the previous RTT, and has all
2139 * been acknowledged in the course of the RTT that ended
2140 * with the ACK we just received. Likewise, old_snd_cwnd
2141 * is the cwnd during the previous RTT.
2143 old_wnd = (tp->vegas.beg_snd_nxt - tp->vegas.beg_snd_una) /
2145 old_snd_cwnd = tp->vegas.beg_snd_cwnd;
2147 /* Save the extent of the current window so we can use this
2148 * at the end of the next RTT.
2150 tp->vegas.beg_snd_una = tp->vegas.beg_snd_nxt;
2151 tp->vegas.beg_snd_nxt = tp->snd_nxt;
2152 tp->vegas.beg_snd_cwnd = tp->snd_cwnd;
2154 /* Take into account the current RTT sample too, to
2155 * decrease the impact of delayed acks. This double counts
2156 * this sample since we count it for the next window as well,
2157 * but that's not too awful, since we're taking the min,
2158 * rather than averaging.
2160 vegas_rtt_calc(tp, seq_rtt);
2162 /* We do the Vegas calculations only if we got enough RTT
2163 * samples that we can be reasonably sure that we got
2164 * at least one RTT sample that wasn't from a delayed ACK.
2165 * If we only had 2 samples total,
2166 * then that means we're getting only 1 ACK per RTT, which
2167 * means they're almost certainly delayed ACKs.
2168 * If we have 3 samples, we should be OK.
2171 if (tp->vegas.cntRTT <= 2) {
2172 /* We don't have enough RTT samples to do the Vegas
2173 * calculation, so we'll behave like Reno.
2175 if (tp->snd_cwnd > tp->snd_ssthresh)
2178 u32 rtt, target_cwnd, diff;
2180 /* We have enough RTT samples, so, using the Vegas
2181 * algorithm, we determine if we should increase or
2182 * decrease cwnd, and by how much.
2185 /* Pluck out the RTT we are using for the Vegas
2186 * calculations. This is the min RTT seen during the
2187 * last RTT. Taking the min filters out the effects
2188 * of delayed ACKs, at the cost of noticing congestion
2191 rtt = tp->vegas.minRTT;
2193 /* Calculate the cwnd we should have, if we weren't
2197 * (actual rate in segments) * baseRTT
2198 * We keep it as a fixed point number with
2199 * V_PARAM_SHIFT bits to the right of the binary point.
2201 target_cwnd = ((old_wnd * tp->vegas.baseRTT)
2202 << V_PARAM_SHIFT) / rtt;
2204 /* Calculate the difference between the window we had,
2205 * and the window we would like to have. This quantity
2206 * is the "Diff" from the Arizona Vegas papers.
2208 * Again, this is a fixed point number with
2209 * V_PARAM_SHIFT bits to the right of the binary
2212 diff = (old_wnd << V_PARAM_SHIFT) - target_cwnd;
2214 if (tp->snd_cwnd < tp->snd_ssthresh) {
2216 if (diff > sysctl_tcp_vegas_gamma) {
2217 /* Going too fast. Time to slow down
2218 * and switch to congestion avoidance.
2220 tp->snd_ssthresh = 2;
2222 /* Set cwnd to match the actual rate
2224 * cwnd = (actual rate) * baseRTT
2225 * Then we add 1 because the integer
2226 * truncation robs us of full link
2229 tp->snd_cwnd = min(tp->snd_cwnd,
2235 /* Congestion avoidance. */
2238 /* Figure out where we would like cwnd
2241 if (diff > sysctl_tcp_vegas_beta) {
2242 /* The old window was too fast, so
2245 next_snd_cwnd = old_snd_cwnd - 1;
2246 } else if (diff < sysctl_tcp_vegas_alpha) {
2247 /* We don't have enough extra packets
2248 * in the network, so speed up.
2250 next_snd_cwnd = old_snd_cwnd + 1;
2252 /* Sending just as fast as we
2255 next_snd_cwnd = old_snd_cwnd;
2258 /* Adjust cwnd upward or downward, toward the
2261 if (next_snd_cwnd > tp->snd_cwnd)
2263 else if (next_snd_cwnd < tp->snd_cwnd)
2268 /* Wipe the slate clean for the next RTT. */
2269 tp->vegas.cntRTT = 0;
2270 tp->vegas.minRTT = 0x7fffffff;
2273 /* The following code is executed for every ack we receive,
2274 * except for conditions checked in should_advance_cwnd()
2275 * before the call to tcp_cong_avoid(). Mainly this means that
2276 * we only execute this code if the ack actually acked some
2280 /* If we are in slow start, increase our cwnd in response to this ACK.
2281 * (If we are not in slow start then we are in congestion avoidance,
2282 * and adjust our congestion window only once per RTT. See the code
2285 if (tp->snd_cwnd <= tp->snd_ssthresh)
2288 /* to keep cwnd from growing without bound */
2289 tp->snd_cwnd = min_t(u32, tp->snd_cwnd, tp->snd_cwnd_clamp);
2291 /* Make sure that we are never so timid as to reduce our cwnd below
2294 * Going below 2 MSS would risk huge delayed ACKs from our receiver.
2296 tp->snd_cwnd = max(tp->snd_cwnd, 2U);
2298 tp->snd_cwnd_stamp = tcp_time_stamp;
2301 static inline void tcp_cong_avoid(struct tcp_opt *tp, u32 ack, u32 seq_rtt)
2303 if (tcp_vegas_enabled(tp))
2304 vegas_cong_avoid(tp, ack, seq_rtt);
2306 reno_cong_avoid(tp);
2309 /* Restart timer after forward progress on connection.
2310 * RFC2988 recommends to restart timer to now+rto.
2313 static __inline__ void tcp_ack_packets_out(struct sock *sk, struct tcp_opt *tp)
2315 if (tp->packets_out==0) {
2316 tcp_clear_xmit_timer(sk, TCP_TIME_RETRANS);
2318 tcp_reset_xmit_timer(sk, TCP_TIME_RETRANS, tp->rto);
2322 /* Remove acknowledged frames from the retransmission queue. */
2323 static int tcp_clean_rtx_queue(struct sock *sk, __s32 *seq_rtt_p)
2325 struct tcp_opt *tp = tcp_sk(sk);
2326 struct sk_buff *skb;
2327 __u32 now = tcp_time_stamp;
2331 while ((skb = skb_peek(&sk->sk_write_queue)) && skb != tp->send_head) {
2332 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
2333 __u8 sacked = scb->sacked;
2335 /* If our packet is before the ack sequence we can
2336 * discard it as it's confirmed to have arrived at
2339 if (after(scb->end_seq, tp->snd_una))
2342 /* Initial outgoing SYN's get put onto the write_queue
2343 * just like anything else we transmit. It is not
2344 * true data, and if we misinform our callers that
2345 * this ACK acks real data, we will erroneously exit
2346 * connection startup slow start one packet too
2347 * quickly. This is severely frowned upon behavior.
2349 if(!(scb->flags & TCPCB_FLAG_SYN)) {
2350 acked |= FLAG_DATA_ACKED;
2352 acked |= FLAG_SYN_ACKED;
2353 tp->retrans_stamp = 0;
2357 if(sacked & TCPCB_RETRANS) {
2358 if(sacked & TCPCB_SACKED_RETRANS)
2360 acked |= FLAG_RETRANS_DATA_ACKED;
2362 } else if (seq_rtt < 0)
2363 seq_rtt = now - scb->when;
2364 if(sacked & TCPCB_SACKED_ACKED)
2366 if(sacked & TCPCB_LOST)
2368 if(sacked & TCPCB_URG) {
2370 !before(scb->end_seq, tp->snd_up))
2373 } else if (seq_rtt < 0)
2374 seq_rtt = now - scb->when;
2375 if (tp->fackets_out)
2378 __skb_unlink(skb, skb->list);
2379 tcp_free_skb(sk, skb);
2382 if (acked&FLAG_ACKED) {
2383 tcp_ack_update_rtt(tp, acked, seq_rtt);
2384 tcp_ack_packets_out(sk, tp);
2387 #if FASTRETRANS_DEBUG > 0
2388 BUG_TRAP((int)tp->sacked_out >= 0);
2389 BUG_TRAP((int)tp->lost_out >= 0);
2390 BUG_TRAP((int)tp->retrans_out >= 0);
2391 if (!tp->packets_out && tp->sack_ok) {
2393 printk(KERN_DEBUG "Leak l=%u %d\n", tp->lost_out,
2397 if (tp->sacked_out) {
2398 printk(KERN_DEBUG "Leak s=%u %d\n", tp->sacked_out,
2402 if (tp->retrans_out) {
2403 printk(KERN_DEBUG "Leak r=%u %d\n", tp->retrans_out,
2405 tp->retrans_out = 0;
2409 *seq_rtt_p = seq_rtt;
2413 static void tcp_ack_probe(struct sock *sk)
2415 struct tcp_opt *tp = tcp_sk(sk);
2417 /* Was it a usable window open? */
2419 if (!after(TCP_SKB_CB(tp->send_head)->end_seq,
2420 tp->snd_una + tp->snd_wnd)) {
2422 tcp_clear_xmit_timer(sk, TCP_TIME_PROBE0);
2423 /* Socket must be waked up by subsequent tcp_data_snd_check().
2424 * This function is not for random using!
2427 tcp_reset_xmit_timer(sk, TCP_TIME_PROBE0,
2428 min(tp->rto << tp->backoff, TCP_RTO_MAX));
2432 static __inline__ int tcp_ack_is_dubious(struct tcp_opt *tp, int flag)
2434 return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
2435 tp->ca_state != TCP_CA_Open);
2438 static __inline__ int tcp_may_raise_cwnd(struct tcp_opt *tp, int flag)
2440 return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
2441 !((1<<tp->ca_state)&(TCPF_CA_Recovery|TCPF_CA_CWR));
2444 /* Check that window update is acceptable.
2445 * The function assumes that snd_una<=ack<=snd_next.
2447 static __inline__ int
2448 tcp_may_update_window(struct tcp_opt *tp, u32 ack, u32 ack_seq, u32 nwin)
2450 return (after(ack, tp->snd_una) ||
2451 after(ack_seq, tp->snd_wl1) ||
2452 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd));
2455 /* Update our send window.
2457 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
2458 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
2460 static int tcp_ack_update_window(struct sock *sk, struct tcp_opt *tp,
2461 struct sk_buff *skb, u32 ack, u32 ack_seq)
2464 u32 nwin = ntohs(skb->h.th->window);
2466 if (likely(!skb->h.th->syn))
2467 nwin <<= tp->snd_wscale;
2469 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
2470 flag |= FLAG_WIN_UPDATE;
2471 tcp_update_wl(tp, ack, ack_seq);
2473 if (tp->snd_wnd != nwin) {
2476 /* Note, it is the only place, where
2477 * fast path is recovered for sending TCP.
2479 tcp_fast_path_check(sk, tp);
2481 if (nwin > tp->max_window) {
2482 tp->max_window = nwin;
2483 tcp_sync_mss(sk, tp->pmtu_cookie);
2493 static void tcp_process_frto(struct sock *sk, u32 prior_snd_una)
2495 struct tcp_opt *tp = tcp_sk(sk);
2497 tcp_sync_left_out(tp);
2499 if (tp->snd_una == prior_snd_una ||
2500 !before(tp->snd_una, tp->frto_highmark)) {
2501 /* RTO was caused by loss, start retransmitting in
2502 * go-back-N slow start
2504 tcp_enter_frto_loss(sk);
2508 if (tp->frto_counter == 1) {
2509 /* First ACK after RTO advances the window: allow two new
2512 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
2514 /* Also the second ACK after RTO advances the window.
2515 * The RTO was likely spurious. Reduce cwnd and continue
2516 * in congestion avoidance
2518 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2519 tcp_moderate_cwnd(tp);
2522 /* F-RTO affects on two new ACKs following RTO.
2523 * At latest on third ACK the TCP behavor is back to normal.
2525 tp->frto_counter = (tp->frto_counter + 1) % 3;
2534 * This function initializes fields used in TCP Westwood+. We can't
2535 * get no information about RTTmin at this time so we simply set it to
2536 * TCP_WESTWOOD_INIT_RTT. This value was chosen to be too conservative
2537 * since in this way we're sure it will be updated in a consistent
2538 * way as soon as possible. It will reasonably happen within the first
2539 * RTT period of the connection lifetime.
2542 static void init_westwood(struct sock *sk)
2544 struct tcp_opt *tp = tcp_sk(sk);
2546 tp->westwood.bw_ns_est = 0;
2547 tp->westwood.bw_est = 0;
2548 tp->westwood.accounted = 0;
2549 tp->westwood.cumul_ack = 0;
2550 tp->westwood.rtt_win_sx = tcp_time_stamp;
2551 tp->westwood.rtt = TCP_WESTWOOD_INIT_RTT;
2552 tp->westwood.rtt_min = TCP_WESTWOOD_INIT_RTT;
2553 tp->westwood.snd_una = tp->snd_una;
2557 * @westwood_do_filter
2558 * Low-pass filter. Implemented using constant coeffients.
2561 static inline __u32 westwood_do_filter(__u32 a, __u32 b)
2563 return (((7 * a) + b) >> 3);
2566 static void westwood_filter(struct sock *sk, __u32 delta)
2568 struct tcp_opt *tp = tcp_sk(sk);
2570 tp->westwood.bw_ns_est =
2571 westwood_do_filter(tp->westwood.bw_ns_est,
2572 tp->westwood.bk / delta);
2573 tp->westwood.bw_est =
2574 westwood_do_filter(tp->westwood.bw_est,
2575 tp->westwood.bw_ns_est);
2579 * @westwood_update_rttmin
2580 * It is used to update RTTmin. In this case we MUST NOT use
2581 * WESTWOOD_RTT_MIN minimum bound since we could be on a LAN!
2584 static inline __u32 westwood_update_rttmin(struct sock *sk)
2586 struct tcp_opt *tp = tcp_sk(sk);
2587 __u32 rttmin = tp->westwood.rtt_min;
2589 if (tp->westwood.rtt == 0)
2592 if (tp->westwood.rtt < tp->westwood.rtt_min || !rttmin)
2593 rttmin = tp->westwood.rtt;
2600 * Evaluate increases for dk.
2603 static inline __u32 westwood_acked(struct sock *sk)
2605 struct tcp_opt *tp = tcp_sk(sk);
2607 return ((tp->snd_una) - (tp->westwood.snd_una));
2611 * @westwood_new_window
2612 * It evaluates if we are receiving data inside the same RTT window as
2615 * It returns 0 if we are still evaluating samples in the same RTT
2616 * window, 1 if the sample has to be considered in the next window.
2619 static int westwood_new_window(struct sock *sk)
2621 struct tcp_opt *tp = tcp_sk(sk);
2626 left_bound = tp->westwood.rtt_win_sx;
2627 rtt = max(tp->westwood.rtt, (u32) TCP_WESTWOOD_RTT_MIN);
2630 * A RTT-window has passed. Be careful since if RTT is less than
2631 * 50ms we don't filter but we continue 'building the sample'.
2632 * This minimum limit was choosen since an estimation on small
2633 * time intervals is better to avoid...
2634 * Obvioulsy on a LAN we reasonably will always have
2635 * right_bound = left_bound + WESTWOOD_RTT_MIN
2638 if ((left_bound + rtt) < tcp_time_stamp)
2645 * @westwood_update_window
2646 * It updates RTT evaluation window if it is the right moment to do
2647 * it. If so it calls filter for evaluating bandwidth.
2650 static void __westwood_update_window(struct sock *sk, __u32 now)
2652 struct tcp_opt *tp = tcp_sk(sk);
2653 __u32 delta = now - tp->westwood.rtt_win_sx;
2658 if (tp->westwood.rtt)
2659 westwood_filter(sk, delta);
2661 tp->westwood.bk = 0;
2662 tp->westwood.rtt_win_sx = tcp_time_stamp;
2666 static void westwood_update_window(struct sock *sk, __u32 now)
2668 if (westwood_new_window(sk))
2669 __westwood_update_window(sk, now);
2673 * @__tcp_westwood_fast_bw
2674 * It is called when we are in fast path. In particular it is called when
2675 * header prediction is successfull. In such case infact update is
2676 * straight forward and doesn't need any particular care.
2679 void __tcp_westwood_fast_bw(struct sock *sk, struct sk_buff *skb)
2681 struct tcp_opt *tp = tcp_sk(sk);
2683 westwood_update_window(sk, tcp_time_stamp);
2685 tp->westwood.bk += westwood_acked(sk);
2686 tp->westwood.snd_una = tp->snd_una;
2687 tp->westwood.rtt_min = westwood_update_rttmin(sk);
2692 * @westwood_dupack_update
2693 * It updates accounted and cumul_ack when receiving a dupack.
2696 static void westwood_dupack_update(struct sock *sk)
2698 struct tcp_opt *tp = tcp_sk(sk);
2700 tp->westwood.accounted += tp->mss_cache;
2701 tp->westwood.cumul_ack = tp->mss_cache;
2704 static inline int westwood_may_change_cumul(struct tcp_opt *tp)
2706 return ((tp->westwood.cumul_ack) > tp->mss_cache);
2709 static inline void westwood_partial_update(struct tcp_opt *tp)
2711 tp->westwood.accounted -= tp->westwood.cumul_ack;
2712 tp->westwood.cumul_ack = tp->mss_cache;
2715 static inline void westwood_complete_update(struct tcp_opt *tp)
2717 tp->westwood.cumul_ack -= tp->westwood.accounted;
2718 tp->westwood.accounted = 0;
2722 * @westwood_acked_count
2723 * This function evaluates cumul_ack for evaluating dk in case of
2724 * delayed or partial acks.
2727 static __u32 westwood_acked_count(struct sock *sk)
2729 struct tcp_opt *tp = tcp_sk(sk);
2731 tp->westwood.cumul_ack = westwood_acked(sk);
2733 /* If cumul_ack is 0 this is a dupack since it's not moving
2736 if (!(tp->westwood.cumul_ack))
2737 westwood_dupack_update(sk);
2739 if (westwood_may_change_cumul(tp)) {
2740 /* Partial or delayed ack */
2741 if ((tp->westwood.accounted) >= (tp->westwood.cumul_ack))
2742 westwood_partial_update(tp);
2744 westwood_complete_update(tp);
2747 tp->westwood.snd_una = tp->snd_una;
2749 return tp->westwood.cumul_ack;
2754 * @__tcp_westwood_slow_bw
2755 * It is called when something is going wrong..even if there could
2756 * be no problems! Infact a simple delayed packet may trigger a
2757 * dupack. But we need to be careful in such case.
2760 void __tcp_westwood_slow_bw(struct sock *sk, struct sk_buff *skb)
2762 struct tcp_opt *tp = tcp_sk(sk);
2764 westwood_update_window(sk, tcp_time_stamp);
2766 tp->westwood.bk += westwood_acked_count(sk);
2767 tp->westwood.rtt_min = westwood_update_rttmin(sk);
2770 /* This routine deals with incoming acks, but not outgoing ones. */
2771 static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag)
2773 struct tcp_opt *tp = tcp_sk(sk);
2774 u32 prior_snd_una = tp->snd_una;
2775 u32 ack_seq = TCP_SKB_CB(skb)->seq;
2776 u32 ack = TCP_SKB_CB(skb)->ack_seq;
2777 u32 prior_in_flight;
2781 /* If the ack is newer than sent or older than previous acks
2782 * then we can probably ignore it.
2784 if (after(ack, tp->snd_nxt))
2785 goto uninteresting_ack;
2787 if (before(ack, prior_snd_una))
2790 if (!(flag&FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
2791 /* Window is constant, pure forward advance.
2792 * No more checks are required.
2793 * Note, we use the fact that SND.UNA>=SND.WL2.
2795 tcp_update_wl(tp, ack, ack_seq);
2797 tcp_westwood_fast_bw(sk, skb);
2798 flag |= FLAG_WIN_UPDATE;
2800 NET_INC_STATS_BH(TCPHPAcks);
2802 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
2805 NET_INC_STATS_BH(TCPPureAcks);
2807 flag |= tcp_ack_update_window(sk, tp, skb, ack, ack_seq);
2809 if (TCP_SKB_CB(skb)->sacked)
2810 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
2812 if (TCP_ECN_rcv_ecn_echo(tp, skb->h.th))
2815 tcp_westwood_slow_bw(sk,skb);
2818 /* We passed data and got it acked, remove any soft error
2819 * log. Something worked...
2821 sk->sk_err_soft = 0;
2822 tp->rcv_tstamp = tcp_time_stamp;
2823 prior_packets = tp->packets_out;
2827 prior_in_flight = tcp_packets_in_flight(tp);
2829 /* See if we can take anything off of the retransmit queue. */
2830 flag |= tcp_clean_rtx_queue(sk, &seq_rtt);
2832 if (tp->frto_counter)
2833 tcp_process_frto(sk, prior_snd_una);
2835 if (tcp_ack_is_dubious(tp, flag)) {
2836 /* Advanve CWND, if state allows this. */
2837 if ((flag & FLAG_DATA_ACKED) &&
2838 (tcp_vegas_enabled(tp) || prior_in_flight >= tp->snd_cwnd) &&
2839 tcp_may_raise_cwnd(tp, flag))
2840 tcp_cong_avoid(tp, ack, seq_rtt);
2841 tcp_fastretrans_alert(sk, prior_snd_una, prior_packets, flag);
2843 if ((flag & FLAG_DATA_ACKED) &&
2844 (tcp_vegas_enabled(tp) || prior_in_flight >= tp->snd_cwnd))
2845 tcp_cong_avoid(tp, ack, seq_rtt);
2848 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag&FLAG_NOT_DUP))
2849 dst_confirm(sk->sk_dst_cache);
2856 /* If this ack opens up a zero window, clear backoff. It was
2857 * being used to time the probes, and is probably far higher than
2858 * it needs to be for normal retransmission.
2865 if (TCP_SKB_CB(skb)->sacked)
2866 tcp_sacktag_write_queue(sk, skb, prior_snd_una);
2869 SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
2874 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
2875 * But, this can also be called on packets in the established flow when
2876 * the fast version below fails.
2878 void tcp_parse_options(struct sk_buff *skb, struct tcp_opt *tp, int estab)
2881 struct tcphdr *th = skb->h.th;
2882 int length=(th->doff*4)-sizeof(struct tcphdr);
2884 ptr = (unsigned char *)(th + 1);
2894 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
2899 if (opsize < 2) /* "silly options" */
2901 if (opsize > length)
2902 return; /* don't parse partial options */
2905 if(opsize==TCPOLEN_MSS && th->syn && !estab) {
2906 u16 in_mss = ntohs(*(__u16 *)ptr);
2908 if (tp->user_mss && tp->user_mss < in_mss)
2909 in_mss = tp->user_mss;
2910 tp->mss_clamp = in_mss;
2915 if(opsize==TCPOLEN_WINDOW && th->syn && !estab)
2916 if (sysctl_tcp_window_scaling) {
2918 tp->snd_wscale = *(__u8 *)ptr;
2919 if(tp->snd_wscale > 14) {
2921 printk("tcp_parse_options: Illegal window "
2922 "scaling value %d >14 received.",
2924 tp->snd_wscale = 14;
2928 case TCPOPT_TIMESTAMP:
2929 if(opsize==TCPOLEN_TIMESTAMP) {
2930 if ((estab && tp->tstamp_ok) ||
2931 (!estab && sysctl_tcp_timestamps)) {
2933 tp->rcv_tsval = ntohl(*(__u32 *)ptr);
2934 tp->rcv_tsecr = ntohl(*(__u32 *)(ptr+4));
2938 case TCPOPT_SACK_PERM:
2939 if(opsize==TCPOLEN_SACK_PERM && th->syn && !estab) {
2940 if (sysctl_tcp_sack) {
2948 if((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
2949 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
2951 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
2960 /* Fast parse options. This hopes to only see timestamps.
2961 * If it is wrong it falls back on tcp_parse_options().
2963 static __inline__ int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th, struct tcp_opt *tp)
2965 if (th->doff == sizeof(struct tcphdr)>>2) {
2968 } else if (tp->tstamp_ok &&
2969 th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) {
2970 __u32 *ptr = (__u32 *)(th + 1);
2971 if (*ptr == ntohl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
2972 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
2975 tp->rcv_tsval = ntohl(*ptr);
2977 tp->rcv_tsecr = ntohl(*ptr);
2981 tcp_parse_options(skb, tp, 1);
2985 static __inline__ void
2986 tcp_store_ts_recent(struct tcp_opt *tp)
2988 tp->ts_recent = tp->rcv_tsval;
2989 tp->ts_recent_stamp = xtime.tv_sec;
2992 static __inline__ void
2993 tcp_replace_ts_recent(struct tcp_opt *tp, u32 seq)
2995 if (tp->saw_tstamp && !after(seq, tp->rcv_wup)) {
2996 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
2997 * extra check below makes sure this can only happen
2998 * for pure ACK frames. -DaveM
3000 * Not only, also it occurs for expired timestamps.
3003 if((s32)(tp->rcv_tsval - tp->ts_recent) >= 0 ||
3004 xtime.tv_sec >= tp->ts_recent_stamp + TCP_PAWS_24DAYS)
3005 tcp_store_ts_recent(tp);
3009 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3011 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3012 * it can pass through stack. So, the following predicate verifies that
3013 * this segment is not used for anything but congestion avoidance or
3014 * fast retransmit. Moreover, we even are able to eliminate most of such
3015 * second order effects, if we apply some small "replay" window (~RTO)
3016 * to timestamp space.
3018 * All these measures still do not guarantee that we reject wrapped ACKs
3019 * on networks with high bandwidth, when sequence space is recycled fastly,
3020 * but it guarantees that such events will be very rare and do not affect
3021 * connection seriously. This doesn't look nice, but alas, PAWS is really
3024 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3025 * states that events when retransmit arrives after original data are rare.
3026 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3027 * the biggest problem on large power networks even with minor reordering.
3028 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3029 * up to bandwidth of 18Gigabit/sec. 8) ]
3032 static int tcp_disordered_ack(struct tcp_opt *tp, struct sk_buff *skb)
3034 struct tcphdr *th = skb->h.th;
3035 u32 seq = TCP_SKB_CB(skb)->seq;
3036 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3038 return (/* 1. Pure ACK with correct sequence number. */
3039 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3041 /* 2. ... and duplicate ACK. */
3042 ack == tp->snd_una &&
3044 /* 3. ... and does not update window. */
3045 !tcp_may_update_window(tp, ack, seq, ntohs(th->window)<<tp->snd_wscale) &&
3047 /* 4. ... and sits in replay window. */
3048 (s32)(tp->ts_recent - tp->rcv_tsval) <= (tp->rto*1024)/HZ);
3051 static __inline__ int tcp_paws_discard(struct tcp_opt *tp, struct sk_buff *skb)
3053 return ((s32)(tp->ts_recent - tp->rcv_tsval) > TCP_PAWS_WINDOW &&
3054 xtime.tv_sec < tp->ts_recent_stamp + TCP_PAWS_24DAYS &&
3055 !tcp_disordered_ack(tp, skb));
3058 /* Check segment sequence number for validity.
3060 * Segment controls are considered valid, if the segment
3061 * fits to the window after truncation to the window. Acceptability
3062 * of data (and SYN, FIN, of course) is checked separately.
3063 * See tcp_data_queue(), for example.
3065 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3066 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3067 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3068 * (borrowed from freebsd)
3071 static inline int tcp_sequence(struct tcp_opt *tp, u32 seq, u32 end_seq)
3073 return !before(end_seq, tp->rcv_wup) &&
3074 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3077 /* When we get a reset we do this. */
3078 static void tcp_reset(struct sock *sk)
3080 /* We want the right error as BSD sees it (and indeed as we do). */
3081 switch (sk->sk_state) {
3083 sk->sk_err = ECONNREFUSED;
3085 case TCP_CLOSE_WAIT:
3091 sk->sk_err = ECONNRESET;
3094 if (!sock_flag(sk, SOCK_DEAD))
3095 sk->sk_error_report(sk);
3101 * Process the FIN bit. This now behaves as it is supposed to work
3102 * and the FIN takes effect when it is validly part of sequence
3103 * space. Not before when we get holes.
3105 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3106 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3109 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3110 * close and we go into CLOSING (and later onto TIME-WAIT)
3112 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3114 static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th)
3116 struct tcp_opt *tp = tcp_sk(sk);
3118 tcp_schedule_ack(tp);
3120 sk->sk_shutdown |= RCV_SHUTDOWN;
3121 sock_set_flag(sk, SOCK_DONE);
3123 switch (sk->sk_state) {
3125 case TCP_ESTABLISHED:
3126 /* Move to CLOSE_WAIT */
3127 tcp_set_state(sk, TCP_CLOSE_WAIT);
3128 tp->ack.pingpong = 1;
3131 case TCP_CLOSE_WAIT:
3133 /* Received a retransmission of the FIN, do
3138 /* RFC793: Remain in the LAST-ACK state. */
3142 /* This case occurs when a simultaneous close
3143 * happens, we must ack the received FIN and
3144 * enter the CLOSING state.
3147 tcp_set_state(sk, TCP_CLOSING);
3150 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3152 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
3155 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3156 * cases we should never reach this piece of code.
3158 printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n",
3159 __FUNCTION__, sk->sk_state);
3163 /* It _is_ possible, that we have something out-of-order _after_ FIN.
3164 * Probably, we should reset in this case. For now drop them.
3166 __skb_queue_purge(&tp->out_of_order_queue);
3169 tcp_mem_reclaim(sk);
3171 if (!sock_flag(sk, SOCK_DEAD)) {
3172 sk->sk_state_change(sk);
3174 /* Do not send POLL_HUP for half duplex close. */
3175 if (sk->sk_shutdown == SHUTDOWN_MASK ||
3176 sk->sk_state == TCP_CLOSE)
3177 sk_wake_async(sk, 1, POLL_HUP);
3179 sk_wake_async(sk, 1, POLL_IN);
3183 static __inline__ int
3184 tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, u32 end_seq)
3186 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
3187 if (before(seq, sp->start_seq))
3188 sp->start_seq = seq;
3189 if (after(end_seq, sp->end_seq))
3190 sp->end_seq = end_seq;
3196 static __inline__ void tcp_dsack_set(struct tcp_opt *tp, u32 seq, u32 end_seq)
3198 if (tp->sack_ok && sysctl_tcp_dsack) {
3199 if (before(seq, tp->rcv_nxt))
3200 NET_INC_STATS_BH(TCPDSACKOldSent);
3202 NET_INC_STATS_BH(TCPDSACKOfoSent);
3205 tp->duplicate_sack[0].start_seq = seq;
3206 tp->duplicate_sack[0].end_seq = end_seq;
3207 tp->eff_sacks = min(tp->num_sacks+1, 4-tp->tstamp_ok);
3211 static __inline__ void tcp_dsack_extend(struct tcp_opt *tp, u32 seq, u32 end_seq)
3214 tcp_dsack_set(tp, seq, end_seq);
3216 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
3219 static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb)
3221 struct tcp_opt *tp = tcp_sk(sk);
3223 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
3224 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3225 NET_INC_STATS_BH(DelayedACKLost);
3226 tcp_enter_quickack_mode(tp);
3228 if (tp->sack_ok && sysctl_tcp_dsack) {
3229 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3231 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
3232 end_seq = tp->rcv_nxt;
3233 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, end_seq);
3240 /* These routines update the SACK block as out-of-order packets arrive or
3241 * in-order packets close up the sequence space.
3243 static void tcp_sack_maybe_coalesce(struct tcp_opt *tp)
3246 struct tcp_sack_block *sp = &tp->selective_acks[0];
3247 struct tcp_sack_block *swalk = sp+1;
3249 /* See if the recent change to the first SACK eats into
3250 * or hits the sequence space of other SACK blocks, if so coalesce.
3252 for (this_sack = 1; this_sack < tp->num_sacks; ) {
3253 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
3256 /* Zap SWALK, by moving every further SACK up by one slot.
3257 * Decrease num_sacks.
3260 tp->eff_sacks = min(tp->num_sacks+tp->dsack, 4-tp->tstamp_ok);
3261 for(i=this_sack; i < tp->num_sacks; i++)
3265 this_sack++, swalk++;
3269 static __inline__ void tcp_sack_swap(struct tcp_sack_block *sack1, struct tcp_sack_block *sack2)
3273 tmp = sack1->start_seq;
3274 sack1->start_seq = sack2->start_seq;
3275 sack2->start_seq = tmp;
3277 tmp = sack1->end_seq;
3278 sack1->end_seq = sack2->end_seq;
3279 sack2->end_seq = tmp;
3282 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
3284 struct tcp_opt *tp = tcp_sk(sk);
3285 struct tcp_sack_block *sp = &tp->selective_acks[0];
3286 int cur_sacks = tp->num_sacks;
3292 for (this_sack=0; this_sack<cur_sacks; this_sack++, sp++) {
3293 if (tcp_sack_extend(sp, seq, end_seq)) {
3294 /* Rotate this_sack to the first one. */
3295 for (; this_sack>0; this_sack--, sp--)
3296 tcp_sack_swap(sp, sp-1);
3298 tcp_sack_maybe_coalesce(tp);
3303 /* Could not find an adjacent existing SACK, build a new one,
3304 * put it at the front, and shift everyone else down. We
3305 * always know there is at least one SACK present already here.
3307 * If the sack array is full, forget about the last one.
3309 if (this_sack >= 4) {
3314 for(; this_sack > 0; this_sack--, sp--)
3318 /* Build the new head SACK, and we're done. */
3319 sp->start_seq = seq;
3320 sp->end_seq = end_seq;
3322 tp->eff_sacks = min(tp->num_sacks + tp->dsack, 4 - tp->tstamp_ok);
3325 /* RCV.NXT advances, some SACKs should be eaten. */
3327 static void tcp_sack_remove(struct tcp_opt *tp)
3329 struct tcp_sack_block *sp = &tp->selective_acks[0];
3330 int num_sacks = tp->num_sacks;
3333 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
3334 if (skb_queue_len(&tp->out_of_order_queue) == 0) {
3336 tp->eff_sacks = tp->dsack;
3340 for(this_sack = 0; this_sack < num_sacks; ) {
3341 /* Check if the start of the sack is covered by RCV.NXT. */
3342 if (!before(tp->rcv_nxt, sp->start_seq)) {
3345 /* RCV.NXT must cover all the block! */
3346 BUG_TRAP(!before(tp->rcv_nxt, sp->end_seq));
3348 /* Zap this SACK, by moving forward any other SACKS. */
3349 for (i=this_sack+1; i < num_sacks; i++)
3350 tp->selective_acks[i-1] = tp->selective_acks[i];
3357 if (num_sacks != tp->num_sacks) {
3358 tp->num_sacks = num_sacks;
3359 tp->eff_sacks = min(tp->num_sacks+tp->dsack, 4-tp->tstamp_ok);
3363 /* This one checks to see if we can put data from the
3364 * out_of_order queue into the receive_queue.
3366 static void tcp_ofo_queue(struct sock *sk)
3368 struct tcp_opt *tp = tcp_sk(sk);
3369 __u32 dsack_high = tp->rcv_nxt;
3370 struct sk_buff *skb;
3372 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
3373 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
3376 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
3377 __u32 dsack = dsack_high;
3378 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
3379 dsack_high = TCP_SKB_CB(skb)->end_seq;
3380 tcp_dsack_extend(tp, TCP_SKB_CB(skb)->seq, dsack);
3383 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3384 SOCK_DEBUG(sk, "ofo packet was already received \n");
3385 __skb_unlink(skb, skb->list);
3389 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
3390 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3391 TCP_SKB_CB(skb)->end_seq);
3393 __skb_unlink(skb, skb->list);
3394 __skb_queue_tail(&sk->sk_receive_queue, skb);
3395 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3397 tcp_fin(skb, sk, skb->h.th);
3401 static inline int tcp_rmem_schedule(struct sock *sk, struct sk_buff *skb)
3403 return (int)skb->truesize <= sk->sk_forward_alloc ||
3404 tcp_mem_schedule(sk, skb->truesize, 1);
3407 static int tcp_prune_queue(struct sock *sk);
3409 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
3411 struct tcphdr *th = skb->h.th;
3412 struct tcp_opt *tp = tcp_sk(sk);
3415 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
3419 __skb_pull(skb, th->doff*4);
3421 TCP_ECN_accept_cwr(tp, skb);
3425 tp->eff_sacks = min_t(unsigned int, tp->num_sacks,
3429 /* Queue data for delivery to the user.
3430 * Packets in sequence go to the receive queue.
3431 * Out of sequence packets to the out_of_order_queue.
3433 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
3434 if (tcp_receive_window(tp) == 0)
3437 /* Ok. In sequence. In window. */
3438 if (tp->ucopy.task == current &&
3439 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
3440 sock_owned_by_user(sk) && !tp->urg_data) {
3441 int chunk = min_t(unsigned int, skb->len,
3444 __set_current_state(TASK_RUNNING);
3447 if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
3448 tp->ucopy.len -= chunk;
3449 tp->copied_seq += chunk;
3450 eaten = (chunk == skb->len && !th->fin);
3451 tcp_rcv_space_adjust(sk);
3459 (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3460 !tcp_rmem_schedule(sk, skb))) {
3461 if (tcp_prune_queue(sk) < 0 || !tcp_rmem_schedule(sk, skb))
3464 tcp_set_owner_r(skb, sk);
3465 __skb_queue_tail(&sk->sk_receive_queue, skb);
3467 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3469 tcp_event_data_recv(sk, tp, skb);
3471 tcp_fin(skb, sk, th);
3473 if (skb_queue_len(&tp->out_of_order_queue)) {
3476 /* RFC2581. 4.2. SHOULD send immediate ACK, when
3477 * gap in queue is filled.
3479 if (!skb_queue_len(&tp->out_of_order_queue))
3480 tp->ack.pingpong = 0;
3484 tcp_sack_remove(tp);
3486 tcp_fast_path_check(sk, tp);
3490 else if (!sock_flag(sk, SOCK_DEAD))
3491 sk->sk_data_ready(sk, 0);
3495 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3496 /* A retransmit, 2nd most common case. Force an immediate ack. */
3497 NET_INC_STATS_BH(DelayedACKLost);
3498 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3501 tcp_enter_quickack_mode(tp);
3502 tcp_schedule_ack(tp);
3508 /* Out of window. F.e. zero window probe. */
3509 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
3512 tcp_enter_quickack_mode(tp);
3514 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3515 /* Partial packet, seq < rcv_next < end_seq */
3516 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
3517 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3518 TCP_SKB_CB(skb)->end_seq);
3520 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
3522 /* If window is closed, drop tail of packet. But after
3523 * remembering D-SACK for its head made in previous line.
3525 if (!tcp_receive_window(tp))
3530 TCP_ECN_check_ce(tp, skb);
3532 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3533 !tcp_rmem_schedule(sk, skb)) {
3534 if (tcp_prune_queue(sk) < 0 || !tcp_rmem_schedule(sk, skb))
3538 /* Disable header prediction. */
3540 tcp_schedule_ack(tp);
3542 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
3543 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3545 tcp_set_owner_r(skb, sk);
3547 if (!skb_peek(&tp->out_of_order_queue)) {
3548 /* Initial out of order segment, build 1 SACK. */
3553 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
3554 tp->selective_acks[0].end_seq =
3555 TCP_SKB_CB(skb)->end_seq;
3557 __skb_queue_head(&tp->out_of_order_queue,skb);
3559 struct sk_buff *skb1 = tp->out_of_order_queue.prev;
3560 u32 seq = TCP_SKB_CB(skb)->seq;
3561 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3563 if (seq == TCP_SKB_CB(skb1)->end_seq) {
3564 __skb_append(skb1, skb);
3566 if (!tp->num_sacks ||
3567 tp->selective_acks[0].end_seq != seq)
3570 /* Common case: data arrive in order after hole. */
3571 tp->selective_acks[0].end_seq = end_seq;
3575 /* Find place to insert this segment. */
3577 if (!after(TCP_SKB_CB(skb1)->seq, seq))
3579 } while ((skb1 = skb1->prev) !=
3580 (struct sk_buff*)&tp->out_of_order_queue);
3582 /* Do skb overlap to previous one? */
3583 if (skb1 != (struct sk_buff*)&tp->out_of_order_queue &&
3584 before(seq, TCP_SKB_CB(skb1)->end_seq)) {
3585 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
3586 /* All the bits are present. Drop. */
3588 tcp_dsack_set(tp, seq, end_seq);
3591 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
3592 /* Partial overlap. */
3593 tcp_dsack_set(tp, seq, TCP_SKB_CB(skb1)->end_seq);
3598 __skb_insert(skb, skb1, skb1->next, &tp->out_of_order_queue);
3600 /* And clean segments covered by new one as whole. */
3601 while ((skb1 = skb->next) !=
3602 (struct sk_buff*)&tp->out_of_order_queue &&
3603 after(end_seq, TCP_SKB_CB(skb1)->seq)) {
3604 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
3605 tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, end_seq);
3608 __skb_unlink(skb1, skb1->list);
3609 tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, TCP_SKB_CB(skb1)->end_seq);
3615 tcp_sack_new_ofo_skb(sk, seq, end_seq);
3619 /* Collapse contiguous sequence of skbs head..tail with
3620 * sequence numbers start..end.
3621 * Segments with FIN/SYN are not collapsed (only because this
3625 tcp_collapse(struct sock *sk, struct sk_buff *head,
3626 struct sk_buff *tail, u32 start, u32 end)
3628 struct sk_buff *skb;
3630 /* First, check that queue is collapsable and find
3631 * the point where collapsing can be useful. */
3632 for (skb = head; skb != tail; ) {
3633 /* No new bits? It is possible on ofo queue. */
3634 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
3635 struct sk_buff *next = skb->next;
3636 __skb_unlink(skb, skb->list);
3638 NET_INC_STATS_BH(TCPRcvCollapsed);
3643 /* The first skb to collapse is:
3645 * - bloated or contains data before "start" or
3646 * overlaps to the next one.
3648 if (!skb->h.th->syn && !skb->h.th->fin &&
3649 (tcp_win_from_space(skb->truesize) > skb->len ||
3650 before(TCP_SKB_CB(skb)->seq, start) ||
3651 (skb->next != tail &&
3652 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq)))
3655 /* Decided to skip this, advance start seq. */
3656 start = TCP_SKB_CB(skb)->end_seq;
3659 if (skb == tail || skb->h.th->syn || skb->h.th->fin)
3662 while (before(start, end)) {
3663 struct sk_buff *nskb;
3664 int header = skb_headroom(skb);
3665 int copy = (PAGE_SIZE - sizeof(struct sk_buff) -
3666 sizeof(struct skb_shared_info) - header - 31)&~15;
3668 /* Too big header? This can happen with IPv6. */
3671 if (end-start < copy)
3673 nskb = alloc_skb(copy+header, GFP_ATOMIC);
3676 skb_reserve(nskb, header);
3677 memcpy(nskb->head, skb->head, header);
3678 nskb->nh.raw = nskb->head + (skb->nh.raw-skb->head);
3679 nskb->h.raw = nskb->head + (skb->h.raw-skb->head);
3680 nskb->mac.raw = nskb->head + (skb->mac.raw-skb->head);
3681 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
3682 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
3683 __skb_insert(nskb, skb->prev, skb, skb->list);
3684 tcp_set_owner_r(nskb, sk);
3686 /* Copy data, releasing collapsed skbs. */
3688 int offset = start - TCP_SKB_CB(skb)->seq;
3689 int size = TCP_SKB_CB(skb)->end_seq - start;
3691 if (offset < 0) BUG();
3693 size = min(copy, size);
3694 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
3696 TCP_SKB_CB(nskb)->end_seq += size;
3700 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
3701 struct sk_buff *next = skb->next;
3702 __skb_unlink(skb, skb->list);
3704 NET_INC_STATS_BH(TCPRcvCollapsed);
3706 if (skb == tail || skb->h.th->syn || skb->h.th->fin)
3713 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
3714 * and tcp_collapse() them until all the queue is collapsed.
3716 static void tcp_collapse_ofo_queue(struct sock *sk)
3718 struct tcp_opt *tp = tcp_sk(sk);
3719 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
3720 struct sk_buff *head;
3726 start = TCP_SKB_CB(skb)->seq;
3727 end = TCP_SKB_CB(skb)->end_seq;
3733 /* Segment is terminated when we see gap or when
3734 * we are at the end of all the queue. */
3735 if (skb == (struct sk_buff *)&tp->out_of_order_queue ||
3736 after(TCP_SKB_CB(skb)->seq, end) ||
3737 before(TCP_SKB_CB(skb)->end_seq, start)) {
3738 tcp_collapse(sk, head, skb, start, end);
3740 if (skb == (struct sk_buff *)&tp->out_of_order_queue)
3742 /* Start new segment */
3743 start = TCP_SKB_CB(skb)->seq;
3744 end = TCP_SKB_CB(skb)->end_seq;
3746 if (before(TCP_SKB_CB(skb)->seq, start))
3747 start = TCP_SKB_CB(skb)->seq;
3748 if (after(TCP_SKB_CB(skb)->end_seq, end))
3749 end = TCP_SKB_CB(skb)->end_seq;
3754 /* Reduce allocated memory if we can, trying to get
3755 * the socket within its memory limits again.
3757 * Return less than zero if we should start dropping frames
3758 * until the socket owning process reads some of the data
3759 * to stabilize the situation.
3761 static int tcp_prune_queue(struct sock *sk)
3763 struct tcp_opt *tp = tcp_sk(sk);
3765 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
3767 NET_INC_STATS_BH(PruneCalled);
3769 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
3770 tcp_clamp_window(sk, tp);
3771 else if (tcp_memory_pressure)
3772 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
3774 tcp_collapse_ofo_queue(sk);
3775 tcp_collapse(sk, sk->sk_receive_queue.next,
3776 (struct sk_buff*)&sk->sk_receive_queue,
3777 tp->copied_seq, tp->rcv_nxt);
3778 tcp_mem_reclaim(sk);
3780 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
3783 /* Collapsing did not help, destructive actions follow.
3784 * This must not ever occur. */
3786 /* First, purge the out_of_order queue. */
3787 if (skb_queue_len(&tp->out_of_order_queue)) {
3788 NET_ADD_STATS_BH(OfoPruned,
3789 skb_queue_len(&tp->out_of_order_queue));
3790 __skb_queue_purge(&tp->out_of_order_queue);
3792 /* Reset SACK state. A conforming SACK implementation will
3793 * do the same at a timeout based retransmit. When a connection
3794 * is in a sad state like this, we care only about integrity
3795 * of the connection not performance.
3799 tcp_mem_reclaim(sk);
3802 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
3805 /* If we are really being abused, tell the caller to silently
3806 * drop receive data on the floor. It will get retransmitted
3807 * and hopefully then we'll have sufficient space.
3809 NET_INC_STATS_BH(RcvPruned);
3811 /* Massive buffer overcommit. */
3817 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
3818 * As additional protections, we do not touch cwnd in retransmission phases,
3819 * and if application hit its sndbuf limit recently.
3821 void tcp_cwnd_application_limited(struct sock *sk)
3823 struct tcp_opt *tp = tcp_sk(sk);
3825 if (tp->ca_state == TCP_CA_Open &&
3826 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
3827 /* Limited by application or receiver window. */
3828 u32 win_used = max(tp->snd_cwnd_used, 2U);
3829 if (win_used < tp->snd_cwnd) {
3830 tp->snd_ssthresh = tcp_current_ssthresh(tp);
3831 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
3833 tp->snd_cwnd_used = 0;
3835 tp->snd_cwnd_stamp = tcp_time_stamp;
3839 /* When incoming ACK allowed to free some skb from write_queue,
3840 * we remember this event in flag tp->queue_shrunk and wake up socket
3841 * on the exit from tcp input handler.
3843 * PROBLEM: sndbuf expansion does not work well with largesend.
3845 static void tcp_new_space(struct sock *sk)
3847 struct tcp_opt *tp = tcp_sk(sk);
3849 if (tp->packets_out < tp->snd_cwnd &&
3850 !(sk->sk_userlocks & SOCK_SNDBUF_LOCK) &&
3851 !tcp_memory_pressure &&
3852 atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) {
3853 int sndmem = max_t(u32, tp->mss_clamp, tp->mss_cache) +
3854 MAX_TCP_HEADER + 16 + sizeof(struct sk_buff),
3855 demanded = max_t(unsigned int, tp->snd_cwnd,
3856 tp->reordering + 1);
3857 sndmem *= 2*demanded;
3858 if (sndmem > sk->sk_sndbuf)
3859 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
3860 tp->snd_cwnd_stamp = tcp_time_stamp;
3863 sk->sk_write_space(sk);
3866 static inline void tcp_check_space(struct sock *sk)
3868 struct tcp_opt *tp = tcp_sk(sk);
3870 if (tp->queue_shrunk) {
3871 tp->queue_shrunk = 0;
3872 if (sk->sk_socket &&
3873 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
3878 static void __tcp_data_snd_check(struct sock *sk, struct sk_buff *skb)
3880 struct tcp_opt *tp = tcp_sk(sk);
3882 if (after(TCP_SKB_CB(skb)->end_seq, tp->snd_una + tp->snd_wnd) ||
3883 tcp_packets_in_flight(tp) >= tp->snd_cwnd ||
3884 tcp_write_xmit(sk, tp->nonagle))
3885 tcp_check_probe_timer(sk, tp);
3888 static __inline__ void tcp_data_snd_check(struct sock *sk)
3890 struct tcp_opt *tp = tcp_sk(sk);
3891 struct sk_buff *skb = tp->send_head;
3894 __tcp_data_snd_check(sk, skb);
3895 tcp_check_space(sk);
3899 * Check if sending an ack is needed.
3901 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
3903 struct tcp_opt *tp = tcp_sk(sk);
3905 /* More than one full frame received... */
3906 if (((tp->rcv_nxt - tp->rcv_wup) > tp->ack.rcv_mss
3907 /* ... and right edge of window advances far enough.
3908 * (tcp_recvmsg() will send ACK otherwise). Or...
3910 && __tcp_select_window(sk) >= tp->rcv_wnd) ||
3911 /* We ACK each frame or... */
3912 tcp_in_quickack_mode(tp) ||
3913 /* We have out of order data. */
3915 skb_peek(&tp->out_of_order_queue))) {
3916 /* Then ack it now */
3919 /* Else, send delayed ack. */
3920 tcp_send_delayed_ack(sk);
3924 static __inline__ void tcp_ack_snd_check(struct sock *sk)
3926 struct tcp_opt *tp = tcp_sk(sk);
3927 if (!tcp_ack_scheduled(tp)) {
3928 /* We sent a data segment already. */
3931 __tcp_ack_snd_check(sk, 1);
3935 * This routine is only called when we have urgent data
3936 * signalled. Its the 'slow' part of tcp_urg. It could be
3937 * moved inline now as tcp_urg is only called from one
3938 * place. We handle URGent data wrong. We have to - as
3939 * BSD still doesn't use the correction from RFC961.
3940 * For 1003.1g we should support a new option TCP_STDURG to permit
3941 * either form (or just set the sysctl tcp_stdurg).
3944 static void tcp_check_urg(struct sock * sk, struct tcphdr * th)
3946 struct tcp_opt *tp = tcp_sk(sk);
3947 u32 ptr = ntohs(th->urg_ptr);
3949 if (ptr && !sysctl_tcp_stdurg)
3951 ptr += ntohl(th->seq);
3953 /* Ignore urgent data that we've already seen and read. */
3954 if (after(tp->copied_seq, ptr))
3957 /* Do not replay urg ptr.
3959 * NOTE: interesting situation not covered by specs.
3960 * Misbehaving sender may send urg ptr, pointing to segment,
3961 * which we already have in ofo queue. We are not able to fetch
3962 * such data and will stay in TCP_URG_NOTYET until will be eaten
3963 * by recvmsg(). Seems, we are not obliged to handle such wicked
3964 * situations. But it is worth to think about possibility of some
3965 * DoSes using some hypothetical application level deadlock.
3967 if (before(ptr, tp->rcv_nxt))
3970 /* Do we already have a newer (or duplicate) urgent pointer? */
3971 if (tp->urg_data && !after(ptr, tp->urg_seq))
3974 /* Tell the world about our new urgent pointer. */
3977 /* We may be adding urgent data when the last byte read was
3978 * urgent. To do this requires some care. We cannot just ignore
3979 * tp->copied_seq since we would read the last urgent byte again
3980 * as data, nor can we alter copied_seq until this data arrives
3981 * or we break the sematics of SIOCATMARK (and thus sockatmark())
3983 * NOTE. Double Dutch. Rendering to plain English: author of comment
3984 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
3985 * and expect that both A and B disappear from stream. This is _wrong_.
3986 * Though this happens in BSD with high probability, this is occasional.
3987 * Any application relying on this is buggy. Note also, that fix "works"
3988 * only in this artificial test. Insert some normal data between A and B and we will
3989 * decline of BSD again. Verdict: it is better to remove to trap
3992 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
3993 !sock_flag(sk, SOCK_URGINLINE) &&
3994 tp->copied_seq != tp->rcv_nxt) {
3995 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
3997 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
3998 __skb_unlink(skb, skb->list);
4003 tp->urg_data = TCP_URG_NOTYET;
4006 /* Disable header prediction. */
4010 /* This is the 'fast' part of urgent handling. */
4011 static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th)
4013 struct tcp_opt *tp = tcp_sk(sk);
4015 /* Check if we get a new urgent pointer - normally not. */
4017 tcp_check_urg(sk,th);
4019 /* Do we wait for any urgent data? - normally not... */
4020 if (tp->urg_data == TCP_URG_NOTYET) {
4021 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
4024 /* Is the urgent pointer pointing into this packet? */
4025 if (ptr < skb->len) {
4027 if (skb_copy_bits(skb, ptr, &tmp, 1))
4029 tp->urg_data = TCP_URG_VALID | tmp;
4030 if (!sock_flag(sk, SOCK_DEAD))
4031 sk->sk_data_ready(sk, 0);
4036 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
4038 struct tcp_opt *tp = tcp_sk(sk);
4039 int chunk = skb->len - hlen;
4043 if (skb->ip_summed==CHECKSUM_UNNECESSARY)
4044 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
4046 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
4050 tp->ucopy.len -= chunk;
4051 tp->copied_seq += chunk;
4052 tcp_rcv_space_adjust(sk);
4059 static int __tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
4063 if (sock_owned_by_user(sk)) {
4065 result = __tcp_checksum_complete(skb);
4068 result = __tcp_checksum_complete(skb);
4073 static __inline__ int
4074 tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
4076 return skb->ip_summed != CHECKSUM_UNNECESSARY &&
4077 __tcp_checksum_complete_user(sk, skb);
4081 * TCP receive function for the ESTABLISHED state.
4083 * It is split into a fast path and a slow path. The fast path is
4085 * - A zero window was announced from us - zero window probing
4086 * is only handled properly in the slow path.
4087 * - Out of order segments arrived.
4088 * - Urgent data is expected.
4089 * - There is no buffer space left
4090 * - Unexpected TCP flags/window values/header lengths are received
4091 * (detected by checking the TCP header against pred_flags)
4092 * - Data is sent in both directions. Fast path only supports pure senders
4093 * or pure receivers (this means either the sequence number or the ack
4094 * value must stay constant)
4095 * - Unexpected TCP option.
4097 * When these conditions are not satisfied it drops into a standard
4098 * receive procedure patterned after RFC793 to handle all cases.
4099 * The first three cases are guaranteed by proper pred_flags setting,
4100 * the rest is checked inline. Fast processing is turned on in
4101 * tcp_data_queue when everything is OK.
4103 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
4104 struct tcphdr *th, unsigned len)
4106 struct tcp_opt *tp = tcp_sk(sk);
4109 * Header prediction.
4110 * The code loosely follows the one in the famous
4111 * "30 instruction TCP receive" Van Jacobson mail.
4113 * Van's trick is to deposit buffers into socket queue
4114 * on a device interrupt, to call tcp_recv function
4115 * on the receive process context and checksum and copy
4116 * the buffer to user space. smart...
4118 * Our current scheme is not silly either but we take the
4119 * extra cost of the net_bh soft interrupt processing...
4120 * We do checksum and copy also but from device to kernel.
4125 /* pred_flags is 0xS?10 << 16 + snd_wnd
4126 * if header_predition is to be made
4127 * 'S' will always be tp->tcp_header_len >> 2
4128 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
4129 * turn it off (when there are holes in the receive
4130 * space for instance)
4131 * PSH flag is ignored.
4134 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
4135 TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4136 int tcp_header_len = tp->tcp_header_len;
4138 /* Timestamp header prediction: tcp_header_len
4139 * is automatically equal to th->doff*4 due to pred_flags
4143 /* Check timestamp */
4144 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
4145 __u32 *ptr = (__u32 *)(th + 1);
4147 /* No? Slow path! */
4148 if (*ptr != ntohl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
4149 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP))
4154 tp->rcv_tsval = ntohl(*ptr);
4156 tp->rcv_tsecr = ntohl(*ptr);
4158 /* If PAWS failed, check it more carefully in slow path */
4159 if ((s32)(tp->rcv_tsval - tp->ts_recent) < 0)
4162 /* DO NOT update ts_recent here, if checksum fails
4163 * and timestamp was corrupted part, it will result
4164 * in a hung connection since we will drop all
4165 * future packets due to the PAWS test.
4169 if (len <= tcp_header_len) {
4170 /* Bulk data transfer: sender */
4171 if (len == tcp_header_len) {
4172 /* Predicted packet is in window by definition.
4173 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4174 * Hence, check seq<=rcv_wup reduces to:
4176 if (tcp_header_len ==
4177 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
4178 tp->rcv_nxt == tp->rcv_wup)
4179 tcp_store_ts_recent(tp);
4181 tcp_rcv_rtt_measure_ts(tp, skb);
4183 /* We know that such packets are checksummed
4186 tcp_ack(sk, skb, 0);
4188 tcp_data_snd_check(sk);
4190 } else { /* Header too small */
4191 TCP_INC_STATS_BH(TcpInErrs);
4197 if (tp->ucopy.task == current &&
4198 tp->copied_seq == tp->rcv_nxt &&
4199 len - tcp_header_len <= tp->ucopy.len &&
4200 sock_owned_by_user(sk)) {
4201 __set_current_state(TASK_RUNNING);
4203 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len)) {
4204 /* Predicted packet is in window by definition.
4205 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4206 * Hence, check seq<=rcv_wup reduces to:
4208 if (tcp_header_len ==
4209 (sizeof(struct tcphdr) +
4210 TCPOLEN_TSTAMP_ALIGNED) &&
4211 tp->rcv_nxt == tp->rcv_wup)
4212 tcp_store_ts_recent(tp);
4214 tcp_rcv_rtt_measure_ts(tp, skb);
4216 __skb_pull(skb, tcp_header_len);
4217 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4218 NET_INC_STATS_BH(TCPHPHitsToUser);
4223 if (tcp_checksum_complete_user(sk, skb))
4226 /* Predicted packet is in window by definition.
4227 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4228 * Hence, check seq<=rcv_wup reduces to:
4230 if (tcp_header_len ==
4231 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
4232 tp->rcv_nxt == tp->rcv_wup)
4233 tcp_store_ts_recent(tp);
4235 tcp_rcv_rtt_measure_ts(tp, skb);
4237 if ((int)skb->truesize > sk->sk_forward_alloc)
4240 NET_INC_STATS_BH(TCPHPHits);
4242 /* Bulk data transfer: receiver */
4243 __skb_pull(skb,tcp_header_len);
4244 __skb_queue_tail(&sk->sk_receive_queue, skb);
4245 tcp_set_owner_r(skb, sk);
4246 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4249 tcp_event_data_recv(sk, tp, skb);
4251 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
4252 /* Well, only one small jumplet in fast path... */
4253 tcp_ack(sk, skb, FLAG_DATA);
4254 tcp_data_snd_check(sk);
4255 if (!tcp_ack_scheduled(tp))
4260 if (tcp_in_quickack_mode(tp)) {
4263 tcp_send_delayed_ack(sk);
4266 __tcp_ack_snd_check(sk, 0);
4273 sk->sk_data_ready(sk, 0);
4279 if (len < (th->doff<<2) || tcp_checksum_complete_user(sk, skb))
4283 * RFC1323: H1. Apply PAWS check first.
4285 if (tcp_fast_parse_options(skb, th, tp) && tp->saw_tstamp &&
4286 tcp_paws_discard(tp, skb)) {
4288 NET_INC_STATS_BH(PAWSEstabRejected);
4289 tcp_send_dupack(sk, skb);
4292 /* Resets are accepted even if PAWS failed.
4294 ts_recent update must be made after we are sure
4295 that the packet is in window.
4300 * Standard slow path.
4303 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
4304 /* RFC793, page 37: "In all states except SYN-SENT, all reset
4305 * (RST) segments are validated by checking their SEQ-fields."
4306 * And page 69: "If an incoming segment is not acceptable,
4307 * an acknowledgment should be sent in reply (unless the RST bit
4308 * is set, if so drop the segment and return)".
4311 tcp_send_dupack(sk, skb);
4320 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
4322 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4323 TCP_INC_STATS_BH(TcpInErrs);
4324 NET_INC_STATS_BH(TCPAbortOnSyn);
4331 tcp_ack(sk, skb, FLAG_SLOWPATH);
4333 tcp_rcv_rtt_measure_ts(tp, skb);
4335 /* Process urgent data. */
4336 tcp_urg(sk, skb, th);
4338 /* step 7: process the segment text */
4339 tcp_data_queue(sk, skb);
4341 tcp_data_snd_check(sk);
4342 tcp_ack_snd_check(sk);
4346 TCP_INC_STATS_BH(TcpInErrs);
4353 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
4354 struct tcphdr *th, unsigned len)
4356 struct tcp_opt *tp = tcp_sk(sk);
4357 int saved_clamp = tp->mss_clamp;
4359 tcp_parse_options(skb, tp, 0);
4363 * "If the state is SYN-SENT then
4364 * first check the ACK bit
4365 * If the ACK bit is set
4366 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
4367 * a reset (unless the RST bit is set, if so drop
4368 * the segment and return)"
4370 * We do not send data with SYN, so that RFC-correct
4373 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
4374 goto reset_and_undo;
4376 if (tp->saw_tstamp && tp->rcv_tsecr &&
4377 !between(tp->rcv_tsecr, tp->retrans_stamp,
4379 NET_INC_STATS_BH(PAWSActiveRejected);
4380 goto reset_and_undo;
4383 /* Now ACK is acceptable.
4385 * "If the RST bit is set
4386 * If the ACK was acceptable then signal the user "error:
4387 * connection reset", drop the segment, enter CLOSED state,
4388 * delete TCB, and return."
4397 * "fifth, if neither of the SYN or RST bits is set then
4398 * drop the segment and return."
4404 goto discard_and_undo;
4407 * "If the SYN bit is on ...
4408 * are acceptable then ...
4409 * (our SYN has been ACKed), change the connection
4410 * state to ESTABLISHED..."
4413 TCP_ECN_rcv_synack(tp, th);
4414 if (tp->ecn_flags&TCP_ECN_OK)
4415 sk->sk_no_largesend = 1;
4417 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
4418 tcp_ack(sk, skb, FLAG_SLOWPATH);
4420 /* Ok.. it's good. Set up sequence numbers and
4421 * move to established.
4423 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
4424 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
4426 /* RFC1323: The window in SYN & SYN/ACK segments is
4429 tp->snd_wnd = ntohs(th->window);
4430 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(skb)->seq);
4432 if (!tp->wscale_ok) {
4433 tp->snd_wscale = tp->rcv_wscale = 0;
4434 tp->window_clamp = min(tp->window_clamp, 65535U);
4437 if (tp->saw_tstamp) {
4439 tp->tcp_header_len =
4440 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
4441 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
4442 tcp_store_ts_recent(tp);
4444 tp->tcp_header_len = sizeof(struct tcphdr);
4447 if (tp->sack_ok && sysctl_tcp_fack)
4450 tcp_sync_mss(sk, tp->pmtu_cookie);
4451 tcp_initialize_rcv_mss(sk);
4453 /* Remember, tcp_poll() does not lock socket!
4454 * Change state from SYN-SENT only after copied_seq
4455 * is initialized. */
4456 tp->copied_seq = tp->rcv_nxt;
4458 tcp_set_state(sk, TCP_ESTABLISHED);
4460 /* Make sure socket is routed, for correct metrics. */
4461 tp->af_specific->rebuild_header(sk);
4463 tcp_init_metrics(sk);
4465 /* Prevent spurious tcp_cwnd_restart() on first data
4468 tp->lsndtime = tcp_time_stamp;
4470 tcp_init_buffer_space(sk);
4472 if (sock_flag(sk, SOCK_KEEPOPEN))
4473 tcp_reset_keepalive_timer(sk, keepalive_time_when(tp));
4475 if (!tp->snd_wscale)
4476 __tcp_fast_path_on(tp, tp->snd_wnd);
4480 if (!sock_flag(sk, SOCK_DEAD)) {
4481 sk->sk_state_change(sk);
4482 sk_wake_async(sk, 0, POLL_OUT);
4485 if (tp->write_pending || tp->defer_accept || tp->ack.pingpong) {
4486 /* Save one ACK. Data will be ready after
4487 * several ticks, if write_pending is set.
4489 * It may be deleted, but with this feature tcpdumps
4490 * look so _wonderfully_ clever, that I was not able
4491 * to stand against the temptation 8) --ANK
4493 tcp_schedule_ack(tp);
4494 tp->ack.lrcvtime = tcp_time_stamp;
4495 tp->ack.ato = TCP_ATO_MIN;
4496 tcp_incr_quickack(tp);
4497 tcp_enter_quickack_mode(tp);
4498 tcp_reset_xmit_timer(sk, TCP_TIME_DACK, TCP_DELACK_MAX);
4509 /* No ACK in the segment */
4513 * "If the RST bit is set
4515 * Otherwise (no ACK) drop the segment and return."
4518 goto discard_and_undo;
4522 if (tp->ts_recent_stamp && tp->saw_tstamp && tcp_paws_check(tp, 0))
4523 goto discard_and_undo;
4526 /* We see SYN without ACK. It is attempt of
4527 * simultaneous connect with crossed SYNs.
4528 * Particularly, it can be connect to self.
4530 tcp_set_state(sk, TCP_SYN_RECV);
4532 if (tp->saw_tstamp) {
4534 tcp_store_ts_recent(tp);
4535 tp->tcp_header_len =
4536 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
4538 tp->tcp_header_len = sizeof(struct tcphdr);
4541 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
4542 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
4544 /* RFC1323: The window in SYN & SYN/ACK segments is
4547 tp->snd_wnd = ntohs(th->window);
4548 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
4549 tp->max_window = tp->snd_wnd;
4551 TCP_ECN_rcv_syn(tp, th);
4552 if (tp->ecn_flags&TCP_ECN_OK)
4553 sk->sk_no_largesend = 1;
4555 tcp_sync_mss(sk, tp->pmtu_cookie);
4556 tcp_initialize_rcv_mss(sk);
4559 tcp_send_synack(sk);
4561 /* Note, we could accept data and URG from this segment.
4562 * There are no obstacles to make this.
4564 * However, if we ignore data in ACKless segments sometimes,
4565 * we have no reasons to accept it sometimes.
4566 * Also, seems the code doing it in step6 of tcp_rcv_state_process
4567 * is not flawless. So, discard packet for sanity.
4568 * Uncomment this return to process the data.
4575 /* "fifth, if neither of the SYN or RST bits is set then
4576 * drop the segment and return."
4580 tcp_clear_options(tp);
4581 tp->mss_clamp = saved_clamp;
4585 tcp_clear_options(tp);
4586 tp->mss_clamp = saved_clamp;
4592 * This function implements the receiving procedure of RFC 793 for
4593 * all states except ESTABLISHED and TIME_WAIT.
4594 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
4595 * address independent.
4598 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
4599 struct tcphdr *th, unsigned len)
4601 struct tcp_opt *tp = tcp_sk(sk);
4606 switch (sk->sk_state) {
4618 if(tp->af_specific->conn_request(sk, skb) < 0)
4623 /* Now we have several options: In theory there is
4624 * nothing else in the frame. KA9Q has an option to
4625 * send data with the syn, BSD accepts data with the
4626 * syn up to the [to be] advertised window and
4627 * Solaris 2.1 gives you a protocol error. For now
4628 * we just ignore it, that fits the spec precisely
4629 * and avoids incompatibilities. It would be nice in
4630 * future to drop through and process the data.
4632 * Now that TTCP is starting to be used we ought to
4634 * But, this leaves one open to an easy denial of
4635 * service attack, and SYN cookies can't defend
4636 * against this problem. So, we drop the data
4637 * in the interest of security over speed.
4646 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
4650 /* Do step6 onward by hand. */
4651 tcp_urg(sk, skb, th);
4653 tcp_data_snd_check(sk);
4657 if (tcp_fast_parse_options(skb, th, tp) && tp->saw_tstamp &&
4658 tcp_paws_discard(tp, skb)) {
4660 NET_INC_STATS_BH(PAWSEstabRejected);
4661 tcp_send_dupack(sk, skb);
4664 /* Reset is accepted even if it did not pass PAWS. */
4667 /* step 1: check sequence number */
4668 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
4670 tcp_send_dupack(sk, skb);
4674 /* step 2: check RST bit */
4680 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
4682 /* step 3: check security and precedence [ignored] */
4686 * Check for a SYN in window.
4688 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4689 NET_INC_STATS_BH(TCPAbortOnSyn);
4694 /* step 5: check the ACK field */
4696 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH);
4698 switch(sk->sk_state) {
4701 tp->copied_seq = tp->rcv_nxt;
4703 tcp_set_state(sk, TCP_ESTABLISHED);
4704 sk->sk_state_change(sk);
4706 /* Note, that this wakeup is only for marginal
4707 * crossed SYN case. Passively open sockets
4708 * are not waked up, because sk->sk_sleep ==
4709 * NULL and sk->sk_socket == NULL.
4711 if (sk->sk_socket) {
4712 sk_wake_async(sk,0,POLL_OUT);
4715 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
4716 tp->snd_wnd = ntohs(th->window) <<
4718 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq,
4719 TCP_SKB_CB(skb)->seq);
4721 /* tcp_ack considers this ACK as duplicate
4722 * and does not calculate rtt.
4723 * Fix it at least with timestamps.
4725 if (tp->saw_tstamp && tp->rcv_tsecr &&
4727 tcp_ack_saw_tstamp(tp, 0);
4730 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
4732 /* Make sure socket is routed, for
4735 tp->af_specific->rebuild_header(sk);
4737 tcp_init_metrics(sk);
4739 /* Prevent spurious tcp_cwnd_restart() on
4740 * first data packet.
4742 tp->lsndtime = tcp_time_stamp;
4744 tcp_initialize_rcv_mss(sk);
4745 tcp_init_buffer_space(sk);
4746 tcp_fast_path_on(tp);
4753 if (tp->snd_una == tp->write_seq) {
4754 tcp_set_state(sk, TCP_FIN_WAIT2);
4755 sk->sk_shutdown |= SEND_SHUTDOWN;
4756 dst_confirm(sk->sk_dst_cache);
4758 if (!sock_flag(sk, SOCK_DEAD))
4759 /* Wake up lingering close() */
4760 sk->sk_state_change(sk);
4764 if (tp->linger2 < 0 ||
4765 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4766 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
4768 NET_INC_STATS_BH(TCPAbortOnData);
4772 tmo = tcp_fin_time(tp);
4773 if (tmo > TCP_TIMEWAIT_LEN) {
4774 tcp_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
4775 } else if (th->fin || sock_owned_by_user(sk)) {
4776 /* Bad case. We could lose such FIN otherwise.
4777 * It is not a big problem, but it looks confusing
4778 * and not so rare event. We still can lose it now,
4779 * if it spins in bh_lock_sock(), but it is really
4782 tcp_reset_keepalive_timer(sk, tmo);
4784 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
4792 if (tp->snd_una == tp->write_seq) {
4793 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4799 if (tp->snd_una == tp->write_seq) {
4800 tcp_update_metrics(sk);
4809 /* step 6: check the URG bit */
4810 tcp_urg(sk, skb, th);
4812 /* step 7: process the segment text */
4813 switch (sk->sk_state) {
4814 case TCP_CLOSE_WAIT:
4817 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4821 /* RFC 793 says to queue data in these states,
4822 * RFC 1122 says we MUST send a reset.
4823 * BSD 4.4 also does reset.
4825 if (sk->sk_shutdown & RCV_SHUTDOWN) {
4826 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4827 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
4828 NET_INC_STATS_BH(TCPAbortOnData);
4834 case TCP_ESTABLISHED:
4835 tcp_data_queue(sk, skb);
4840 /* tcp_data could move socket to TIME-WAIT */
4841 if (sk->sk_state != TCP_CLOSE) {
4842 tcp_data_snd_check(sk);
4843 tcp_ack_snd_check(sk);
4853 EXPORT_SYMBOL(sysctl_tcp_ecn);
4854 EXPORT_SYMBOL(sysctl_tcp_reordering);
4855 EXPORT_SYMBOL(tcp_cwnd_application_limited);
4856 EXPORT_SYMBOL(tcp_parse_options);
4857 EXPORT_SYMBOL(tcp_rcv_established);
4858 EXPORT_SYMBOL(tcp_rcv_state_process);