2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * Implementation of the Transmission Control Protocol(TCP).
8 * Version: $Id: tcp_input.c,v 1.243 2002/02/01 22:01:04 davem Exp $
10 * Authors: Ross Biro, <bir7@leland.Stanford.Edu>
11 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
12 * Mark Evans, <evansmp@uhura.aston.ac.uk>
13 * Corey Minyard <wf-rch!minyard@relay.EU.net>
14 * Florian La Roche, <flla@stud.uni-sb.de>
15 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
16 * Linus Torvalds, <torvalds@cs.helsinki.fi>
17 * Alan Cox, <gw4pts@gw4pts.ampr.org>
18 * Matthew Dillon, <dillon@apollo.west.oic.com>
19 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
20 * Jorge Cwik, <jorge@laser.satlink.net>
25 * Pedro Roque : Fast Retransmit/Recovery.
27 * Retransmit queue handled by TCP.
28 * Better retransmit timer handling.
29 * New congestion avoidance.
33 * Eric : Fast Retransmit.
34 * Randy Scott : MSS option defines.
35 * Eric Schenk : Fixes to slow start algorithm.
36 * Eric Schenk : Yet another double ACK bug.
37 * Eric Schenk : Delayed ACK bug fixes.
38 * Eric Schenk : Floyd style fast retrans war avoidance.
39 * David S. Miller : Don't allow zero congestion window.
40 * Eric Schenk : Fix retransmitter so that it sends
41 * next packet on ack of previous packet.
42 * Andi Kleen : Moved open_request checking here
43 * and process RSTs for open_requests.
44 * Andi Kleen : Better prune_queue, and other fixes.
45 * Andrey Savochkin: Fix RTT measurements in the presnce of
47 * Andrey Savochkin: Check sequence numbers correctly when
48 * removing SACKs due to in sequence incoming
50 * Andi Kleen: Make sure we never ack data there is not
51 * enough room for. Also make this condition
52 * a fatal error if it might still happen.
53 * Andi Kleen: Add tcp_measure_rcv_mss to make
54 * connections with MSS<min(MTU,ann. MSS)
55 * work without delayed acks.
56 * Andi Kleen: Process packets with PSH set in the
58 * J Hadi Salim: ECN support
61 * Panu Kuhlberg: Experimental audit of TCP (re)transmission
62 * engine. Lots of bugs are found.
63 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs
64 * Angelo Dell'Aera: TCP Westwood+ support
67 #include <linux/config.h>
69 #include <linux/module.h>
70 #include <linux/sysctl.h>
72 #include <net/inet_common.h>
73 #include <linux/ipsec.h>
75 int sysctl_tcp_timestamps = 1;
76 int sysctl_tcp_window_scaling = 1;
77 int sysctl_tcp_sack = 1;
78 int sysctl_tcp_fack = 1;
79 int sysctl_tcp_reordering = TCP_FASTRETRANS_THRESH;
81 int sysctl_tcp_dsack = 1;
82 int sysctl_tcp_app_win = 31;
83 int sysctl_tcp_adv_win_scale = 2;
85 int sysctl_tcp_stdurg;
86 int sysctl_tcp_rfc1337;
87 int sysctl_tcp_max_orphans = NR_FILE;
89 int sysctl_tcp_nometrics_save;
90 int sysctl_tcp_westwood;
91 int sysctl_tcp_vegas_cong_avoid;
93 int sysctl_tcp_moderate_rcvbuf = 1;
95 /* Default values of the Vegas variables, in fixed-point representation
96 * with V_PARAM_SHIFT bits to the right of the binary point.
98 #define V_PARAM_SHIFT 1
99 int sysctl_tcp_vegas_alpha = 1<<V_PARAM_SHIFT;
100 int sysctl_tcp_vegas_beta = 3<<V_PARAM_SHIFT;
101 int sysctl_tcp_vegas_gamma = 1<<V_PARAM_SHIFT;
102 int sysctl_tcp_bic = 1;
103 int sysctl_tcp_bic_fast_convergence = 1;
104 int sysctl_tcp_bic_low_window = 14;
106 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
107 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
108 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
109 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
110 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
111 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
112 #define FLAG_ECE 0x40 /* ECE in this ACK */
113 #define FLAG_DATA_LOST 0x80 /* SACK detected data lossage. */
114 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
116 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
117 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
118 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
119 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
121 #define IsReno(tp) ((tp)->sack_ok == 0)
122 #define IsFack(tp) ((tp)->sack_ok & 2)
123 #define IsDSack(tp) ((tp)->sack_ok & 4)
125 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
127 /* Adapt the MSS value used to make delayed ack decision to the
130 static __inline__ void tcp_measure_rcv_mss(struct tcp_opt *tp, struct sk_buff *skb)
132 unsigned int len, lss;
134 lss = tp->ack.last_seg_size;
135 tp->ack.last_seg_size = 0;
137 /* skb->len may jitter because of SACKs, even if peer
138 * sends good full-sized frames.
141 if (len >= tp->ack.rcv_mss) {
142 tp->ack.rcv_mss = len;
144 /* Otherwise, we make more careful check taking into account,
145 * that SACKs block is variable.
147 * "len" is invariant segment length, including TCP header.
149 len += skb->data - skb->h.raw;
150 if (len >= TCP_MIN_RCVMSS + sizeof(struct tcphdr) ||
151 /* If PSH is not set, packet should be
152 * full sized, provided peer TCP is not badly broken.
153 * This observation (if it is correct 8)) allows
154 * to handle super-low mtu links fairly.
156 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
157 !(tcp_flag_word(skb->h.th)&TCP_REMNANT))) {
158 /* Subtract also invariant (if peer is RFC compliant),
159 * tcp header plus fixed timestamp option length.
160 * Resulting "len" is MSS free of SACK jitter.
162 len -= tp->tcp_header_len;
163 tp->ack.last_seg_size = len;
165 tp->ack.rcv_mss = len;
169 tp->ack.pending |= TCP_ACK_PUSHED;
173 static void tcp_incr_quickack(struct tcp_opt *tp)
175 unsigned quickacks = tp->rcv_wnd/(2*tp->ack.rcv_mss);
179 if (quickacks > tp->ack.quick)
180 tp->ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
183 void tcp_enter_quickack_mode(struct tcp_opt *tp)
185 tcp_incr_quickack(tp);
186 tp->ack.pingpong = 0;
187 tp->ack.ato = TCP_ATO_MIN;
190 /* Send ACKs quickly, if "quick" count is not exhausted
191 * and the session is not interactive.
194 static __inline__ int tcp_in_quickack_mode(struct tcp_opt *tp)
196 return (tp->ack.quick && !tp->ack.pingpong);
199 /* Buffer size and advertised window tuning.
201 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
204 static void tcp_fixup_sndbuf(struct sock *sk)
206 int sndmem = tcp_sk(sk)->mss_clamp + MAX_TCP_HEADER + 16 +
207 sizeof(struct sk_buff);
209 if (sk->sk_sndbuf < 3 * sndmem)
210 sk->sk_sndbuf = min(3 * sndmem, sysctl_tcp_wmem[2]);
213 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
215 * All tcp_full_space() is split to two parts: "network" buffer, allocated
216 * forward and advertised in receiver window (tp->rcv_wnd) and
217 * "application buffer", required to isolate scheduling/application
218 * latencies from network.
219 * window_clamp is maximal advertised window. It can be less than
220 * tcp_full_space(), in this case tcp_full_space() - window_clamp
221 * is reserved for "application" buffer. The less window_clamp is
222 * the smoother our behaviour from viewpoint of network, but the lower
223 * throughput and the higher sensitivity of the connection to losses. 8)
225 * rcv_ssthresh is more strict window_clamp used at "slow start"
226 * phase to predict further behaviour of this connection.
227 * It is used for two goals:
228 * - to enforce header prediction at sender, even when application
229 * requires some significant "application buffer". It is check #1.
230 * - to prevent pruning of receive queue because of misprediction
231 * of receiver window. Check #2.
233 * The scheme does not work when sender sends good segments opening
234 * window and then starts to feed us spagetti. But it should work
235 * in common situations. Otherwise, we have to rely on queue collapsing.
238 /* Slow part of check#2. */
240 __tcp_grow_window(struct sock *sk, struct tcp_opt *tp, struct sk_buff *skb)
243 int truesize = tcp_win_from_space(skb->truesize)/2;
244 int window = tcp_full_space(sk)/2;
246 while (tp->rcv_ssthresh <= window) {
247 if (truesize <= skb->len)
248 return 2*tp->ack.rcv_mss;
256 static __inline__ void
257 tcp_grow_window(struct sock *sk, struct tcp_opt *tp, struct sk_buff *skb)
260 if (tp->rcv_ssthresh < tp->window_clamp &&
261 (int)tp->rcv_ssthresh < tcp_space(sk) &&
262 !tcp_memory_pressure) {
265 /* Check #2. Increase window, if skb with such overhead
266 * will fit to rcvbuf in future.
268 if (tcp_win_from_space(skb->truesize) <= skb->len)
271 incr = __tcp_grow_window(sk, tp, skb);
274 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr, tp->window_clamp);
280 /* 3. Tuning rcvbuf, when connection enters established state. */
282 static void tcp_fixup_rcvbuf(struct sock *sk)
284 struct tcp_opt *tp = tcp_sk(sk);
285 int rcvmem = tp->advmss + MAX_TCP_HEADER + 16 + sizeof(struct sk_buff);
287 /* Try to select rcvbuf so that 4 mss-sized segments
288 * will fit to window and correspoding skbs will fit to our rcvbuf.
289 * (was 3; 4 is minimum to allow fast retransmit to work.)
291 while (tcp_win_from_space(rcvmem) < tp->advmss)
293 if (sk->sk_rcvbuf < 4 * rcvmem)
294 sk->sk_rcvbuf = min(4 * rcvmem, sysctl_tcp_rmem[2]);
297 /* 4. Try to fixup all. It is made iimediately after connection enters
300 static void tcp_init_buffer_space(struct sock *sk)
302 struct tcp_opt *tp = tcp_sk(sk);
305 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
306 tcp_fixup_rcvbuf(sk);
307 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
308 tcp_fixup_sndbuf(sk);
310 tp->rcvq_space.space = tp->rcv_wnd;
312 maxwin = tcp_full_space(sk);
314 if (tp->window_clamp >= maxwin) {
315 tp->window_clamp = maxwin;
317 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
318 tp->window_clamp = max(maxwin -
319 (maxwin >> sysctl_tcp_app_win),
323 /* Force reservation of one segment. */
324 if (sysctl_tcp_app_win &&
325 tp->window_clamp > 2 * tp->advmss &&
326 tp->window_clamp + tp->advmss > maxwin)
327 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
329 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
330 tp->snd_cwnd_stamp = tcp_time_stamp;
333 static void init_bictcp(struct tcp_opt *tp)
337 tp->bictcp.last_max_cwnd = 0;
338 tp->bictcp.last_cwnd = 0;
339 tp->bictcp.last_stamp = 0;
342 /* 5. Recalculate window clamp after socket hit its memory bounds. */
343 static void tcp_clamp_window(struct sock *sk, struct tcp_opt *tp)
346 unsigned int app_win = tp->rcv_nxt - tp->copied_seq;
351 skb_queue_walk(&tp->out_of_order_queue, skb) {
355 /* If overcommit is due to out of order segments,
356 * do not clamp window. Try to expand rcvbuf instead.
359 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
360 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
361 !tcp_memory_pressure &&
362 atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0])
363 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
366 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf) {
368 if (atomic_read(&sk->sk_rmem_alloc) >= 2 * sk->sk_rcvbuf)
370 if (app_win > tp->ack.rcv_mss)
371 app_win -= tp->ack.rcv_mss;
372 app_win = max(app_win, 2U*tp->advmss);
375 tp->window_clamp = min(tp->window_clamp, app_win);
376 tp->rcv_ssthresh = min(tp->window_clamp, 2U*tp->advmss);
380 /* Receiver "autotuning" code.
382 * The algorithm for RTT estimation w/o timestamps is based on
383 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
384 * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps>
386 * More detail on this code can be found at
387 * <http://www.psc.edu/~jheffner/senior_thesis.ps>,
388 * though this reference is out of date. A new paper
391 static void tcp_rcv_rtt_update(struct tcp_opt *tp, u32 sample, int win_dep)
393 u32 new_sample = tp->rcv_rtt_est.rtt;
399 if (new_sample != 0) {
400 /* If we sample in larger samples in the non-timestamp
401 * case, we could grossly overestimate the RTT especially
402 * with chatty applications or bulk transfer apps which
403 * are stalled on filesystem I/O.
405 * Also, since we are only going for a minimum in the
406 * non-timestamp case, we do not smoothe things out
407 * else with timestamps disabled convergance takes too
411 m -= (new_sample >> 3);
413 } else if (m < new_sample)
416 /* No previous mesaure. */
420 if (tp->rcv_rtt_est.rtt != new_sample)
421 tp->rcv_rtt_est.rtt = new_sample;
424 static inline void tcp_rcv_rtt_measure(struct tcp_opt *tp)
426 if (tp->rcv_rtt_est.time == 0)
428 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
430 tcp_rcv_rtt_update(tp,
431 jiffies - tp->rcv_rtt_est.time,
435 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
436 tp->rcv_rtt_est.time = tcp_time_stamp;
439 static inline void tcp_rcv_rtt_measure_ts(struct tcp_opt *tp, struct sk_buff *skb)
442 (TCP_SKB_CB(skb)->end_seq -
443 TCP_SKB_CB(skb)->seq >= tp->ack.rcv_mss))
444 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rcv_tsecr, 0);
448 * This function should be called every time data is copied to user space.
449 * It calculates the appropriate TCP receive buffer space.
451 void tcp_rcv_space_adjust(struct sock *sk)
453 struct tcp_opt *tp = tcp_sk(sk);
457 if (tp->rcvq_space.time == 0)
460 time = tcp_time_stamp - tp->rcvq_space.time;
461 if (time < (tp->rcv_rtt_est.rtt >> 3) ||
462 tp->rcv_rtt_est.rtt == 0)
465 space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
467 space = max(tp->rcvq_space.space, space);
469 if (tp->rcvq_space.space != space) {
472 tp->rcvq_space.space = space;
474 if (sysctl_tcp_moderate_rcvbuf) {
475 int new_clamp = space;
477 /* Receive space grows, normalize in order to
478 * take into account packet headers and sk_buff
479 * structure overhead.
484 rcvmem = (tp->advmss + MAX_TCP_HEADER +
485 16 + sizeof(struct sk_buff));
486 while (tcp_win_from_space(rcvmem) < tp->advmss)
489 space = min(space, sysctl_tcp_rmem[2]);
490 if (space > sk->sk_rcvbuf) {
491 sk->sk_rcvbuf = space;
493 /* Make the window clamp follow along. */
494 tp->window_clamp = new_clamp;
500 tp->rcvq_space.seq = tp->copied_seq;
501 tp->rcvq_space.time = tcp_time_stamp;
504 /* There is something which you must keep in mind when you analyze the
505 * behavior of the tp->ato delayed ack timeout interval. When a
506 * connection starts up, we want to ack as quickly as possible. The
507 * problem is that "good" TCP's do slow start at the beginning of data
508 * transmission. The means that until we send the first few ACK's the
509 * sender will sit on his end and only queue most of his data, because
510 * he can only send snd_cwnd unacked packets at any given time. For
511 * each ACK we send, he increments snd_cwnd and transmits more of his
514 static void tcp_event_data_recv(struct sock *sk, struct tcp_opt *tp, struct sk_buff *skb)
518 tcp_schedule_ack(tp);
520 tcp_measure_rcv_mss(tp, skb);
522 tcp_rcv_rtt_measure(tp);
524 now = tcp_time_stamp;
527 /* The _first_ data packet received, initialize
528 * delayed ACK engine.
530 tcp_incr_quickack(tp);
531 tp->ack.ato = TCP_ATO_MIN;
533 int m = now - tp->ack.lrcvtime;
535 if (m <= TCP_ATO_MIN/2) {
536 /* The fastest case is the first. */
537 tp->ack.ato = (tp->ack.ato>>1) + TCP_ATO_MIN/2;
538 } else if (m < tp->ack.ato) {
539 tp->ack.ato = (tp->ack.ato>>1) + m;
540 if (tp->ack.ato > tp->rto)
541 tp->ack.ato = tp->rto;
542 } else if (m > tp->rto) {
543 /* Too long gap. Apparently sender falled to
544 * restart window, so that we send ACKs quickly.
546 tcp_incr_quickack(tp);
547 sk_stream_mem_reclaim(sk);
550 tp->ack.lrcvtime = now;
552 TCP_ECN_check_ce(tp, skb);
555 tcp_grow_window(sk, tp, skb);
558 /* When starting a new connection, pin down the current choice of
559 * congestion algorithm.
561 void tcp_ca_init(struct tcp_opt *tp)
563 if (sysctl_tcp_westwood)
564 tp->adv_cong = TCP_WESTWOOD;
565 else if (sysctl_tcp_bic)
566 tp->adv_cong = TCP_BIC;
567 else if (sysctl_tcp_vegas_cong_avoid) {
568 tp->adv_cong = TCP_VEGAS;
569 tp->vegas.baseRTT = 0x7fffffff;
570 tcp_vegas_enable(tp);
574 /* Do RTT sampling needed for Vegas.
576 * o min-filter RTT samples from within an RTT to get the current
577 * propagation delay + queuing delay (we are min-filtering to try to
578 * avoid the effects of delayed ACKs)
579 * o min-filter RTT samples from a much longer window (forever for now)
580 * to find the propagation delay (baseRTT)
582 static inline void vegas_rtt_calc(struct tcp_opt *tp, __u32 rtt)
584 __u32 vrtt = rtt + 1; /* Never allow zero rtt or baseRTT */
586 /* Filter to find propagation delay: */
587 if (vrtt < tp->vegas.baseRTT)
588 tp->vegas.baseRTT = vrtt;
590 /* Find the min RTT during the last RTT to find
591 * the current prop. delay + queuing delay:
593 tp->vegas.minRTT = min(tp->vegas.minRTT, vrtt);
597 /* Called to compute a smoothed rtt estimate. The data fed to this
598 * routine either comes from timestamps, or from segments that were
599 * known _not_ to have been retransmitted [see Karn/Partridge
600 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
601 * piece by Van Jacobson.
602 * NOTE: the next three routines used to be one big routine.
603 * To save cycles in the RFC 1323 implementation it was better to break
604 * it up into three procedures. -- erics
606 static void tcp_rtt_estimator(struct tcp_opt *tp, __u32 mrtt)
608 long m = mrtt; /* RTT */
610 if (tcp_vegas_enabled(tp))
611 vegas_rtt_calc(tp, mrtt);
613 /* The following amusing code comes from Jacobson's
614 * article in SIGCOMM '88. Note that rtt and mdev
615 * are scaled versions of rtt and mean deviation.
616 * This is designed to be as fast as possible
617 * m stands for "measurement".
619 * On a 1990 paper the rto value is changed to:
620 * RTO = rtt + 4 * mdev
622 * Funny. This algorithm seems to be very broken.
623 * These formulae increase RTO, when it should be decreased, increase
624 * too slowly, when it should be incresed fastly, decrease too fastly
625 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
626 * does not matter how to _calculate_ it. Seems, it was trap
627 * that VJ failed to avoid. 8)
632 m -= (tp->srtt >> 3); /* m is now error in rtt est */
633 tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */
635 m = -m; /* m is now abs(error) */
636 m -= (tp->mdev >> 2); /* similar update on mdev */
637 /* This is similar to one of Eifel findings.
638 * Eifel blocks mdev updates when rtt decreases.
639 * This solution is a bit different: we use finer gain
640 * for mdev in this case (alpha*beta).
641 * Like Eifel it also prevents growth of rto,
642 * but also it limits too fast rto decreases,
643 * happening in pure Eifel.
648 m -= (tp->mdev >> 2); /* similar update on mdev */
650 tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */
651 if (tp->mdev > tp->mdev_max) {
652 tp->mdev_max = tp->mdev;
653 if (tp->mdev_max > tp->rttvar)
654 tp->rttvar = tp->mdev_max;
656 if (after(tp->snd_una, tp->rtt_seq)) {
657 if (tp->mdev_max < tp->rttvar)
658 tp->rttvar -= (tp->rttvar-tp->mdev_max)>>2;
659 tp->rtt_seq = tp->snd_nxt;
660 tp->mdev_max = TCP_RTO_MIN;
663 /* no previous measure. */
664 tp->srtt = m<<3; /* take the measured time to be rtt */
665 tp->mdev = m<<1; /* make sure rto = 3*rtt */
666 tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN);
667 tp->rtt_seq = tp->snd_nxt;
670 tcp_westwood_update_rtt(tp, tp->srtt >> 3);
673 /* Calculate rto without backoff. This is the second half of Van Jacobson's
674 * routine referred to above.
676 static __inline__ void tcp_set_rto(struct tcp_opt *tp)
678 /* Old crap is replaced with new one. 8)
681 * 1. If rtt variance happened to be less 50msec, it is hallucination.
682 * It cannot be less due to utterly erratic ACK generation made
683 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
684 * to do with delayed acks, because at cwnd>2 true delack timeout
685 * is invisible. Actually, Linux-2.4 also generates erratic
686 * ACKs in some curcumstances.
688 tp->rto = (tp->srtt >> 3) + tp->rttvar;
690 /* 2. Fixups made earlier cannot be right.
691 * If we do not estimate RTO correctly without them,
692 * all the algo is pure shit and should be replaced
693 * with correct one. It is exaclty, which we pretend to do.
697 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
698 * guarantees that rto is higher.
700 static __inline__ void tcp_bound_rto(struct tcp_opt *tp)
702 if (tp->rto > TCP_RTO_MAX)
703 tp->rto = TCP_RTO_MAX;
706 /* Save metrics learned by this TCP session.
707 This function is called only, when TCP finishes successfully
708 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
710 void tcp_update_metrics(struct sock *sk)
712 struct tcp_opt *tp = tcp_sk(sk);
713 struct dst_entry *dst = __sk_dst_get(sk);
715 if (sysctl_tcp_nometrics_save)
720 if (dst && (dst->flags&DST_HOST)) {
723 if (tp->backoff || !tp->srtt) {
724 /* This session failed to estimate rtt. Why?
725 * Probably, no packets returned in time.
728 if (!(dst_metric_locked(dst, RTAX_RTT)))
729 dst->metrics[RTAX_RTT-1] = 0;
733 m = dst_metric(dst, RTAX_RTT) - tp->srtt;
735 /* If newly calculated rtt larger than stored one,
736 * store new one. Otherwise, use EWMA. Remember,
737 * rtt overestimation is always better than underestimation.
739 if (!(dst_metric_locked(dst, RTAX_RTT))) {
741 dst->metrics[RTAX_RTT-1] = tp->srtt;
743 dst->metrics[RTAX_RTT-1] -= (m>>3);
746 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) {
750 /* Scale deviation to rttvar fixed point */
755 if (m >= dst_metric(dst, RTAX_RTTVAR))
756 dst->metrics[RTAX_RTTVAR-1] = m;
758 dst->metrics[RTAX_RTTVAR-1] -=
759 (dst->metrics[RTAX_RTTVAR-1] - m)>>2;
762 if (tp->snd_ssthresh >= 0xFFFF) {
763 /* Slow start still did not finish. */
764 if (dst_metric(dst, RTAX_SSTHRESH) &&
765 !dst_metric_locked(dst, RTAX_SSTHRESH) &&
766 (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH))
767 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1;
768 if (!dst_metric_locked(dst, RTAX_CWND) &&
769 tp->snd_cwnd > dst_metric(dst, RTAX_CWND))
770 dst->metrics[RTAX_CWND-1] = tp->snd_cwnd;
771 } else if (tp->snd_cwnd > tp->snd_ssthresh &&
772 tp->ca_state == TCP_CA_Open) {
773 /* Cong. avoidance phase, cwnd is reliable. */
774 if (!dst_metric_locked(dst, RTAX_SSTHRESH))
775 dst->metrics[RTAX_SSTHRESH-1] =
776 max(tp->snd_cwnd >> 1, tp->snd_ssthresh);
777 if (!dst_metric_locked(dst, RTAX_CWND))
778 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_cwnd) >> 1;
780 /* Else slow start did not finish, cwnd is non-sense,
781 ssthresh may be also invalid.
783 if (!dst_metric_locked(dst, RTAX_CWND))
784 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_ssthresh) >> 1;
785 if (dst->metrics[RTAX_SSTHRESH-1] &&
786 !dst_metric_locked(dst, RTAX_SSTHRESH) &&
787 tp->snd_ssthresh > dst->metrics[RTAX_SSTHRESH-1])
788 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh;
791 if (!dst_metric_locked(dst, RTAX_REORDERING)) {
792 if (dst->metrics[RTAX_REORDERING-1] < tp->reordering &&
793 tp->reordering != sysctl_tcp_reordering)
794 dst->metrics[RTAX_REORDERING-1] = tp->reordering;
799 /* Numbers are taken from RFC2414. */
800 __u32 tcp_init_cwnd(struct tcp_opt *tp, struct dst_entry *dst)
802 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
805 if (tp->mss_cache_std > 1460)
808 cwnd = (tp->mss_cache_std > 1095) ? 3 : 4;
810 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
813 /* Initialize metrics on socket. */
815 static void tcp_init_metrics(struct sock *sk)
817 struct tcp_opt *tp = tcp_sk(sk);
818 struct dst_entry *dst = __sk_dst_get(sk);
825 if (dst_metric_locked(dst, RTAX_CWND))
826 tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND);
827 if (dst_metric(dst, RTAX_SSTHRESH)) {
828 tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH);
829 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
830 tp->snd_ssthresh = tp->snd_cwnd_clamp;
832 if (dst_metric(dst, RTAX_REORDERING) &&
833 tp->reordering != dst_metric(dst, RTAX_REORDERING)) {
835 tp->reordering = dst_metric(dst, RTAX_REORDERING);
838 if (dst_metric(dst, RTAX_RTT) == 0)
841 if (!tp->srtt && dst_metric(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3))
844 /* Initial rtt is determined from SYN,SYN-ACK.
845 * The segment is small and rtt may appear much
846 * less than real one. Use per-dst memory
847 * to make it more realistic.
849 * A bit of theory. RTT is time passed after "normal" sized packet
850 * is sent until it is ACKed. In normal curcumstances sending small
851 * packets force peer to delay ACKs and calculation is correct too.
852 * The algorithm is adaptive and, provided we follow specs, it
853 * NEVER underestimate RTT. BUT! If peer tries to make some clever
854 * tricks sort of "quick acks" for time long enough to decrease RTT
855 * to low value, and then abruptly stops to do it and starts to delay
856 * ACKs, wait for troubles.
858 if (dst_metric(dst, RTAX_RTT) > tp->srtt) {
859 tp->srtt = dst_metric(dst, RTAX_RTT);
860 tp->rtt_seq = tp->snd_nxt;
862 if (dst_metric(dst, RTAX_RTTVAR) > tp->mdev) {
863 tp->mdev = dst_metric(dst, RTAX_RTTVAR);
864 tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN);
868 if (tp->rto < TCP_TIMEOUT_INIT && !tp->saw_tstamp)
870 tp->snd_cwnd = tcp_init_cwnd(tp, dst);
871 tp->snd_cwnd_stamp = tcp_time_stamp;
875 /* Play conservative. If timestamps are not
876 * supported, TCP will fail to recalculate correct
877 * rtt, if initial rto is too small. FORGET ALL AND RESET!
879 if (!tp->saw_tstamp && tp->srtt) {
881 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT;
882 tp->rto = TCP_TIMEOUT_INIT;
886 static void tcp_update_reordering(struct tcp_opt *tp, int metric, int ts)
888 if (metric > tp->reordering) {
889 tp->reordering = min(TCP_MAX_REORDERING, metric);
891 /* This exciting event is worth to be remembered. 8) */
893 NET_INC_STATS_BH(LINUX_MIB_TCPTSREORDER);
895 NET_INC_STATS_BH(LINUX_MIB_TCPRENOREORDER);
897 NET_INC_STATS_BH(LINUX_MIB_TCPFACKREORDER);
899 NET_INC_STATS_BH(LINUX_MIB_TCPSACKREORDER);
900 #if FASTRETRANS_DEBUG > 1
901 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n",
902 tp->sack_ok, tp->ca_state,
904 tcp_get_pcount(&tp->fackets_out),
905 tcp_get_pcount(&tp->sacked_out),
906 tp->undo_marker ? tp->undo_retrans : 0);
908 /* Disable FACK yet. */
913 /* This procedure tags the retransmission queue when SACKs arrive.
915 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
916 * Packets in queue with these bits set are counted in variables
917 * sacked_out, retrans_out and lost_out, correspondingly.
919 * Valid combinations are:
920 * Tag InFlight Description
921 * 0 1 - orig segment is in flight.
922 * S 0 - nothing flies, orig reached receiver.
923 * L 0 - nothing flies, orig lost by net.
924 * R 2 - both orig and retransmit are in flight.
925 * L|R 1 - orig is lost, retransmit is in flight.
926 * S|R 1 - orig reached receiver, retrans is still in flight.
927 * (L|S|R is logically valid, it could occur when L|R is sacked,
928 * but it is equivalent to plain S and code short-curcuits it to S.
929 * L|S is logically invalid, it would mean -1 packet in flight 8))
931 * These 6 states form finite state machine, controlled by the following events:
932 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
933 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
934 * 3. Loss detection event of one of three flavors:
935 * A. Scoreboard estimator decided the packet is lost.
936 * A'. Reno "three dupacks" marks head of queue lost.
937 * A''. Its FACK modfication, head until snd.fack is lost.
938 * B. SACK arrives sacking data transmitted after never retransmitted
940 * C. SACK arrives sacking SND.NXT at the moment, when the
941 * segment was retransmitted.
942 * 4. D-SACK added new rule: D-SACK changes any tag to S.
944 * It is pleasant to note, that state diagram turns out to be commutative,
945 * so that we are allowed not to be bothered by order of our actions,
946 * when multiple events arrive simultaneously. (see the function below).
948 * Reordering detection.
949 * --------------------
950 * Reordering metric is maximal distance, which a packet can be displaced
951 * in packet stream. With SACKs we can estimate it:
953 * 1. SACK fills old hole and the corresponding segment was not
954 * ever retransmitted -> reordering. Alas, we cannot use it
955 * when segment was retransmitted.
956 * 2. The last flaw is solved with D-SACK. D-SACK arrives
957 * for retransmitted and already SACKed segment -> reordering..
958 * Both of these heuristics are not used in Loss state, when we cannot
959 * account for retransmits accurately.
962 tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, u32 prior_snd_una)
964 struct tcp_opt *tp = tcp_sk(sk);
965 unsigned char *ptr = ack_skb->h.raw + TCP_SKB_CB(ack_skb)->sacked;
966 struct tcp_sack_block *sp = (struct tcp_sack_block *)(ptr+2);
967 int num_sacks = (ptr[1] - TCPOLEN_SACK_BASE)>>3;
968 int reord = tcp_get_pcount(&tp->packets_out);
970 u32 lost_retrans = 0;
974 /* So, SACKs for already sent large segments will be lost.
975 * Not good, but alternative is to resegment the queue. */
976 if (sk->sk_route_caps & NETIF_F_TSO) {
977 sk->sk_route_caps &= ~NETIF_F_TSO;
978 sk->sk_no_largesend = 1;
979 tp->mss_cache = tp->mss_cache_std;
982 if (!tcp_get_pcount(&tp->sacked_out))
983 tcp_set_pcount(&tp->fackets_out, 0);
984 prior_fackets = tcp_get_pcount(&tp->fackets_out);
986 for (i=0; i<num_sacks; i++, sp++) {
988 __u32 start_seq = ntohl(sp->start_seq);
989 __u32 end_seq = ntohl(sp->end_seq);
993 /* Check for D-SACK. */
995 u32 ack = TCP_SKB_CB(ack_skb)->ack_seq;
997 if (before(start_seq, ack)) {
1000 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV);
1001 } else if (num_sacks > 1 &&
1002 !after(end_seq, ntohl(sp[1].end_seq)) &&
1003 !before(start_seq, ntohl(sp[1].start_seq))) {
1006 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV);
1009 /* D-SACK for already forgotten data...
1010 * Do dumb counting. */
1012 !after(end_seq, prior_snd_una) &&
1013 after(end_seq, tp->undo_marker))
1016 /* Eliminate too old ACKs, but take into
1017 * account more or less fresh ones, they can
1018 * contain valid SACK info.
1020 if (before(ack, prior_snd_una - tp->max_window))
1024 /* Event "B" in the comment above. */
1025 if (after(end_seq, tp->high_seq))
1026 flag |= FLAG_DATA_LOST;
1028 sk_stream_for_retrans_queue(skb, sk) {
1029 u8 sacked = TCP_SKB_CB(skb)->sacked;
1032 /* The retransmission queue is always in order, so
1033 * we can short-circuit the walk early.
1035 if(!before(TCP_SKB_CB(skb)->seq, end_seq))
1038 fack_count += tcp_skb_pcount(skb);
1040 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1041 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1043 /* Account D-SACK for retransmitted packet. */
1044 if ((dup_sack && in_sack) &&
1045 (sacked & TCPCB_RETRANS) &&
1046 after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker))
1049 /* The frame is ACKed. */
1050 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) {
1051 if (sacked&TCPCB_RETRANS) {
1052 if ((dup_sack && in_sack) &&
1053 (sacked&TCPCB_SACKED_ACKED))
1054 reord = min(fack_count, reord);
1056 /* If it was in a hole, we detected reordering. */
1057 if (fack_count < prior_fackets &&
1058 !(sacked&TCPCB_SACKED_ACKED))
1059 reord = min(fack_count, reord);
1062 /* Nothing to do; acked frame is about to be dropped. */
1066 if ((sacked&TCPCB_SACKED_RETRANS) &&
1067 after(end_seq, TCP_SKB_CB(skb)->ack_seq) &&
1068 (!lost_retrans || after(end_seq, lost_retrans)))
1069 lost_retrans = end_seq;
1074 if (!(sacked&TCPCB_SACKED_ACKED)) {
1075 if (sacked & TCPCB_SACKED_RETRANS) {
1076 /* If the segment is not tagged as lost,
1077 * we do not clear RETRANS, believing
1078 * that retransmission is still in flight.
1080 if (sacked & TCPCB_LOST) {
1081 TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1082 tcp_dec_pcount(&tp->lost_out, skb);
1083 tcp_dec_pcount(&tp->retrans_out, skb);
1086 /* New sack for not retransmitted frame,
1087 * which was in hole. It is reordering.
1089 if (!(sacked & TCPCB_RETRANS) &&
1090 fack_count < prior_fackets)
1091 reord = min(fack_count, reord);
1093 if (sacked & TCPCB_LOST) {
1094 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1095 tcp_dec_pcount(&tp->lost_out, skb);
1099 TCP_SKB_CB(skb)->sacked |= TCPCB_SACKED_ACKED;
1100 flag |= FLAG_DATA_SACKED;
1101 tcp_inc_pcount(&tp->sacked_out, skb);
1103 if (fack_count > tcp_get_pcount(&tp->fackets_out))
1104 tcp_set_pcount(&tp->fackets_out, fack_count);
1106 if (dup_sack && (sacked&TCPCB_RETRANS))
1107 reord = min(fack_count, reord);
1110 /* D-SACK. We can detect redundant retransmission
1111 * in S|R and plain R frames and clear it.
1112 * undo_retrans is decreased above, L|R frames
1113 * are accounted above as well.
1116 (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS)) {
1117 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1118 tcp_dec_pcount(&tp->retrans_out, skb);
1123 /* Check for lost retransmit. This superb idea is
1124 * borrowed from "ratehalving". Event "C".
1125 * Later note: FACK people cheated me again 8),
1126 * we have to account for reordering! Ugly,
1129 if (lost_retrans && tp->ca_state == TCP_CA_Recovery) {
1130 struct sk_buff *skb;
1132 sk_stream_for_retrans_queue(skb, sk) {
1133 if (after(TCP_SKB_CB(skb)->seq, lost_retrans))
1135 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1137 if ((TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) &&
1138 after(lost_retrans, TCP_SKB_CB(skb)->ack_seq) &&
1140 !before(lost_retrans,
1141 TCP_SKB_CB(skb)->ack_seq + tp->reordering *
1142 tp->mss_cache_std))) {
1143 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1144 tcp_dec_pcount(&tp->retrans_out, skb);
1146 if (!(TCP_SKB_CB(skb)->sacked&(TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1147 tcp_inc_pcount(&tp->lost_out, skb);
1148 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1149 flag |= FLAG_DATA_SACKED;
1150 NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT);
1156 tcp_set_pcount(&tp->left_out,
1157 (tcp_get_pcount(&tp->sacked_out) +
1158 tcp_get_pcount(&tp->lost_out)));
1160 if ((reord < tcp_get_pcount(&tp->fackets_out)) &&
1161 tp->ca_state != TCP_CA_Loss)
1162 tcp_update_reordering(tp,
1163 ((tcp_get_pcount(&tp->fackets_out) + 1) -
1166 #if FASTRETRANS_DEBUG > 0
1167 BUG_TRAP((int)tcp_get_pcount(&tp->sacked_out) >= 0);
1168 BUG_TRAP((int)tcp_get_pcount(&tp->lost_out) >= 0);
1169 BUG_TRAP((int)tcp_get_pcount(&tp->retrans_out) >= 0);
1170 BUG_TRAP((int)tcp_packets_in_flight(tp) >= 0);
1175 /* RTO occurred, but do not yet enter loss state. Instead, transmit two new
1176 * segments to see from the next ACKs whether any data was really missing.
1177 * If the RTO was spurious, new ACKs should arrive.
1179 void tcp_enter_frto(struct sock *sk)
1181 struct tcp_opt *tp = tcp_sk(sk);
1182 struct sk_buff *skb;
1184 tp->frto_counter = 1;
1186 if (tp->ca_state <= TCP_CA_Disorder ||
1187 tp->snd_una == tp->high_seq ||
1188 (tp->ca_state == TCP_CA_Loss && !tp->retransmits)) {
1189 tp->prior_ssthresh = tcp_current_ssthresh(tp);
1190 if (!tcp_westwood_ssthresh(tp))
1191 tp->snd_ssthresh = tcp_recalc_ssthresh(tp);
1194 /* Have to clear retransmission markers here to keep the bookkeeping
1195 * in shape, even though we are not yet in Loss state.
1196 * If something was really lost, it is eventually caught up
1197 * in tcp_enter_frto_loss.
1199 tcp_set_pcount(&tp->retrans_out, 0);
1200 tp->undo_marker = tp->snd_una;
1201 tp->undo_retrans = 0;
1203 sk_stream_for_retrans_queue(skb, sk) {
1204 TCP_SKB_CB(skb)->sacked &= ~TCPCB_RETRANS;
1206 tcp_sync_left_out(tp);
1208 tcp_set_ca_state(tp, TCP_CA_Open);
1209 tp->frto_highmark = tp->snd_nxt;
1212 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1213 * which indicates that we should follow the traditional RTO recovery,
1214 * i.e. mark everything lost and do go-back-N retransmission.
1216 static void tcp_enter_frto_loss(struct sock *sk)
1218 struct tcp_opt *tp = tcp_sk(sk);
1219 struct sk_buff *skb;
1222 tcp_set_pcount(&tp->sacked_out, 0);
1223 tcp_set_pcount(&tp->lost_out, 0);
1224 tcp_set_pcount(&tp->fackets_out, 0);
1226 sk_stream_for_retrans_queue(skb, sk) {
1227 cnt += tcp_skb_pcount(skb);
1228 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1229 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) {
1231 /* Do not mark those segments lost that were
1232 * forward transmitted after RTO
1234 if (!after(TCP_SKB_CB(skb)->end_seq,
1235 tp->frto_highmark)) {
1236 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1237 tcp_inc_pcount(&tp->lost_out, skb);
1240 tcp_inc_pcount(&tp->sacked_out, skb);
1241 tcp_set_pcount(&tp->fackets_out, cnt);
1244 tcp_sync_left_out(tp);
1246 tp->snd_cwnd = tp->frto_counter + tcp_packets_in_flight(tp)+1;
1247 tp->snd_cwnd_cnt = 0;
1248 tp->snd_cwnd_stamp = tcp_time_stamp;
1249 tp->undo_marker = 0;
1250 tp->frto_counter = 0;
1252 tp->reordering = min_t(unsigned int, tp->reordering,
1253 sysctl_tcp_reordering);
1254 tcp_set_ca_state(tp, TCP_CA_Loss);
1255 tp->high_seq = tp->frto_highmark;
1256 TCP_ECN_queue_cwr(tp);
1261 void tcp_clear_retrans(struct tcp_opt *tp)
1263 tcp_set_pcount(&tp->left_out, 0);
1264 tcp_set_pcount(&tp->retrans_out, 0);
1266 tcp_set_pcount(&tp->fackets_out, 0);
1267 tcp_set_pcount(&tp->sacked_out, 0);
1268 tcp_set_pcount(&tp->lost_out, 0);
1270 tp->undo_marker = 0;
1271 tp->undo_retrans = 0;
1274 /* Enter Loss state. If "how" is not zero, forget all SACK information
1275 * and reset tags completely, otherwise preserve SACKs. If receiver
1276 * dropped its ofo queue, we will know this due to reneging detection.
1278 void tcp_enter_loss(struct sock *sk, int how)
1280 struct tcp_opt *tp = tcp_sk(sk);
1281 struct sk_buff *skb;
1284 /* Reduce ssthresh if it has not yet been made inside this window. */
1285 if (tp->ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
1286 (tp->ca_state == TCP_CA_Loss && !tp->retransmits)) {
1287 tp->prior_ssthresh = tcp_current_ssthresh(tp);
1288 tp->snd_ssthresh = tcp_recalc_ssthresh(tp);
1291 tp->snd_cwnd_cnt = 0;
1292 tp->snd_cwnd_stamp = tcp_time_stamp;
1294 tcp_clear_retrans(tp);
1296 /* Push undo marker, if it was plain RTO and nothing
1297 * was retransmitted. */
1299 tp->undo_marker = tp->snd_una;
1301 sk_stream_for_retrans_queue(skb, sk) {
1302 cnt += tcp_skb_pcount(skb);
1303 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS)
1304 tp->undo_marker = 0;
1305 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1306 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
1307 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1308 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1309 tcp_inc_pcount(&tp->lost_out, skb);
1311 tcp_inc_pcount(&tp->sacked_out, skb);
1312 tcp_set_pcount(&tp->fackets_out, cnt);
1315 tcp_sync_left_out(tp);
1317 tp->reordering = min_t(unsigned int, tp->reordering,
1318 sysctl_tcp_reordering);
1319 tcp_set_ca_state(tp, TCP_CA_Loss);
1320 tp->high_seq = tp->snd_nxt;
1321 TCP_ECN_queue_cwr(tp);
1324 static int tcp_check_sack_reneging(struct sock *sk, struct tcp_opt *tp)
1326 struct sk_buff *skb;
1328 /* If ACK arrived pointing to a remembered SACK,
1329 * it means that our remembered SACKs do not reflect
1330 * real state of receiver i.e.
1331 * receiver _host_ is heavily congested (or buggy).
1332 * Do processing similar to RTO timeout.
1334 if ((skb = skb_peek(&sk->sk_write_queue)) != NULL &&
1335 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
1336 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING);
1338 tcp_enter_loss(sk, 1);
1340 tcp_retransmit_skb(sk, skb_peek(&sk->sk_write_queue));
1341 tcp_reset_xmit_timer(sk, TCP_TIME_RETRANS, tp->rto);
1347 static inline int tcp_fackets_out(struct tcp_opt *tp)
1349 return IsReno(tp) ? tcp_get_pcount(&tp->sacked_out)+1 :
1350 tcp_get_pcount(&tp->fackets_out);
1353 static inline int tcp_skb_timedout(struct tcp_opt *tp, struct sk_buff *skb)
1355 return (tcp_time_stamp - TCP_SKB_CB(skb)->when > tp->rto);
1358 static inline int tcp_head_timedout(struct sock *sk, struct tcp_opt *tp)
1360 return tcp_get_pcount(&tp->packets_out) &&
1361 tcp_skb_timedout(tp, skb_peek(&sk->sk_write_queue));
1364 /* Linux NewReno/SACK/FACK/ECN state machine.
1365 * --------------------------------------
1367 * "Open" Normal state, no dubious events, fast path.
1368 * "Disorder" In all the respects it is "Open",
1369 * but requires a bit more attention. It is entered when
1370 * we see some SACKs or dupacks. It is split of "Open"
1371 * mainly to move some processing from fast path to slow one.
1372 * "CWR" CWND was reduced due to some Congestion Notification event.
1373 * It can be ECN, ICMP source quench, local device congestion.
1374 * "Recovery" CWND was reduced, we are fast-retransmitting.
1375 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
1377 * tcp_fastretrans_alert() is entered:
1378 * - each incoming ACK, if state is not "Open"
1379 * - when arrived ACK is unusual, namely:
1384 * Counting packets in flight is pretty simple.
1386 * in_flight = packets_out - left_out + retrans_out
1388 * packets_out is SND.NXT-SND.UNA counted in packets.
1390 * retrans_out is number of retransmitted segments.
1392 * left_out is number of segments left network, but not ACKed yet.
1394 * left_out = sacked_out + lost_out
1396 * sacked_out: Packets, which arrived to receiver out of order
1397 * and hence not ACKed. With SACKs this number is simply
1398 * amount of SACKed data. Even without SACKs
1399 * it is easy to give pretty reliable estimate of this number,
1400 * counting duplicate ACKs.
1402 * lost_out: Packets lost by network. TCP has no explicit
1403 * "loss notification" feedback from network (for now).
1404 * It means that this number can be only _guessed_.
1405 * Actually, it is the heuristics to predict lossage that
1406 * distinguishes different algorithms.
1408 * F.e. after RTO, when all the queue is considered as lost,
1409 * lost_out = packets_out and in_flight = retrans_out.
1411 * Essentially, we have now two algorithms counting
1414 * FACK: It is the simplest heuristics. As soon as we decided
1415 * that something is lost, we decide that _all_ not SACKed
1416 * packets until the most forward SACK are lost. I.e.
1417 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
1418 * It is absolutely correct estimate, if network does not reorder
1419 * packets. And it loses any connection to reality when reordering
1420 * takes place. We use FACK by default until reordering
1421 * is suspected on the path to this destination.
1423 * NewReno: when Recovery is entered, we assume that one segment
1424 * is lost (classic Reno). While we are in Recovery and
1425 * a partial ACK arrives, we assume that one more packet
1426 * is lost (NewReno). This heuristics are the same in NewReno
1429 * Imagine, that's all! Forget about all this shamanism about CWND inflation
1430 * deflation etc. CWND is real congestion window, never inflated, changes
1431 * only according to classic VJ rules.
1433 * Really tricky (and requiring careful tuning) part of algorithm
1434 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
1435 * The first determines the moment _when_ we should reduce CWND and,
1436 * hence, slow down forward transmission. In fact, it determines the moment
1437 * when we decide that hole is caused by loss, rather than by a reorder.
1439 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
1440 * holes, caused by lost packets.
1442 * And the most logically complicated part of algorithm is undo
1443 * heuristics. We detect false retransmits due to both too early
1444 * fast retransmit (reordering) and underestimated RTO, analyzing
1445 * timestamps and D-SACKs. When we detect that some segments were
1446 * retransmitted by mistake and CWND reduction was wrong, we undo
1447 * window reduction and abort recovery phase. This logic is hidden
1448 * inside several functions named tcp_try_undo_<something>.
1451 /* This function decides, when we should leave Disordered state
1452 * and enter Recovery phase, reducing congestion window.
1454 * Main question: may we further continue forward transmission
1455 * with the same cwnd?
1458 tcp_time_to_recover(struct sock *sk, struct tcp_opt *tp)
1462 /* Trick#1: The loss is proven. */
1463 if (tcp_get_pcount(&tp->lost_out))
1466 /* Not-A-Trick#2 : Classic rule... */
1467 if (tcp_fackets_out(tp) > tp->reordering)
1470 /* Trick#3 : when we use RFC2988 timer restart, fast
1471 * retransmit can be triggered by timeout of queue head.
1473 if (tcp_head_timedout(sk, tp))
1476 /* Trick#4: It is still not OK... But will it be useful to delay
1479 packets_out = tcp_get_pcount(&tp->packets_out);
1480 if (packets_out <= tp->reordering &&
1481 tcp_get_pcount(&tp->sacked_out) >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
1482 !tcp_may_send_now(sk, tp)) {
1483 /* We have nothing to send. This connection is limited
1484 * either by receiver window or by application.
1492 /* If we receive more dupacks than we expected counting segments
1493 * in assumption of absent reordering, interpret this as reordering.
1494 * The only another reason could be bug in receiver TCP.
1496 static void tcp_check_reno_reordering(struct tcp_opt *tp, int addend)
1500 holes = max(tcp_get_pcount(&tp->lost_out), 1U);
1501 holes = min(holes, tcp_get_pcount(&tp->packets_out));
1503 if ((tcp_get_pcount(&tp->sacked_out) + holes) >
1504 tcp_get_pcount(&tp->packets_out)) {
1505 tcp_set_pcount(&tp->sacked_out,
1506 (tcp_get_pcount(&tp->packets_out) - holes));
1507 tcp_update_reordering(tp,
1508 tcp_get_pcount(&tp->packets_out)+addend,
1513 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1515 static void tcp_add_reno_sack(struct tcp_opt *tp)
1517 tcp_inc_pcount_explicit(&tp->sacked_out, 1);
1518 tcp_check_reno_reordering(tp, 0);
1519 tcp_sync_left_out(tp);
1522 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1524 static void tcp_remove_reno_sacks(struct sock *sk, struct tcp_opt *tp, int acked)
1527 /* One ACK acked hole. The rest eat duplicate ACKs. */
1528 if (acked-1 >= tcp_get_pcount(&tp->sacked_out))
1529 tcp_set_pcount(&tp->sacked_out, 0);
1531 tcp_dec_pcount_explicit(&tp->sacked_out, acked-1);
1533 tcp_check_reno_reordering(tp, acked);
1534 tcp_sync_left_out(tp);
1537 static inline void tcp_reset_reno_sack(struct tcp_opt *tp)
1539 tcp_set_pcount(&tp->sacked_out, 0);
1540 tcp_set_pcount(&tp->left_out, tcp_get_pcount(&tp->lost_out));
1543 /* Mark head of queue up as lost. */
1545 tcp_mark_head_lost(struct sock *sk, struct tcp_opt *tp, int packets, u32 high_seq)
1547 struct sk_buff *skb;
1550 BUG_TRAP(cnt <= tcp_get_pcount(&tp->packets_out));
1552 sk_stream_for_retrans_queue(skb, sk) {
1553 cnt -= tcp_skb_pcount(skb);
1554 if (cnt < 0 || after(TCP_SKB_CB(skb)->end_seq, high_seq))
1556 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
1557 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1558 tcp_inc_pcount(&tp->lost_out, skb);
1561 tcp_sync_left_out(tp);
1564 /* Account newly detected lost packet(s) */
1566 static void tcp_update_scoreboard(struct sock *sk, struct tcp_opt *tp)
1569 int lost = tcp_get_pcount(&tp->fackets_out) - tp->reordering;
1572 tcp_mark_head_lost(sk, tp, lost, tp->high_seq);
1574 tcp_mark_head_lost(sk, tp, 1, tp->high_seq);
1577 /* New heuristics: it is possible only after we switched
1578 * to restart timer each time when something is ACKed.
1579 * Hence, we can detect timed out packets during fast
1580 * retransmit without falling to slow start.
1582 if (tcp_head_timedout(sk, tp)) {
1583 struct sk_buff *skb;
1585 sk_stream_for_retrans_queue(skb, sk) {
1586 if (tcp_skb_timedout(tp, skb) &&
1587 !(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
1588 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1589 tcp_inc_pcount(&tp->lost_out, skb);
1592 tcp_sync_left_out(tp);
1596 /* CWND moderation, preventing bursts due to too big ACKs
1597 * in dubious situations.
1599 static __inline__ void tcp_moderate_cwnd(struct tcp_opt *tp)
1601 tp->snd_cwnd = min(tp->snd_cwnd,
1602 tcp_packets_in_flight(tp)+tcp_max_burst(tp));
1603 tp->snd_cwnd_stamp = tcp_time_stamp;
1606 /* Decrease cwnd each second ack. */
1608 static void tcp_cwnd_down(struct tcp_opt *tp)
1610 int decr = tp->snd_cwnd_cnt + 1;
1615 * Here limit is evaluated as BWestimation*RTTmin (for obtaining it
1616 * in packets we use mss_cache). If sysctl_tcp_westwood is off
1617 * tcp_westwood_bw_rttmin() returns 0. In such case snd_ssthresh is
1618 * still used as usual. It prevents other strange cases in which
1619 * BWE*RTTmin could assume value 0. It should not happen but...
1622 if (!(limit = tcp_westwood_bw_rttmin(tp)))
1623 limit = tp->snd_ssthresh/2;
1625 tp->snd_cwnd_cnt = decr&1;
1628 if (decr && tp->snd_cwnd > limit)
1629 tp->snd_cwnd -= decr;
1631 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp)+1);
1632 tp->snd_cwnd_stamp = tcp_time_stamp;
1635 /* Nothing was retransmitted or returned timestamp is less
1636 * than timestamp of the first retransmission.
1638 static __inline__ int tcp_packet_delayed(struct tcp_opt *tp)
1640 return !tp->retrans_stamp ||
1641 (tp->saw_tstamp && tp->rcv_tsecr &&
1642 (__s32)(tp->rcv_tsecr - tp->retrans_stamp) < 0);
1645 /* Undo procedures. */
1647 #if FASTRETRANS_DEBUG > 1
1648 static void DBGUNDO(struct sock *sk, struct tcp_opt *tp, const char *msg)
1650 struct inet_opt *inet = inet_sk(sk);
1651 printk(KERN_DEBUG "Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n",
1653 NIPQUAD(inet->daddr), ntohs(inet->dport),
1654 tp->snd_cwnd, tcp_get_pcount(&tp->left_out),
1655 tp->snd_ssthresh, tp->prior_ssthresh,
1656 tcp_get_pcount(&tp->packets_out));
1659 #define DBGUNDO(x...) do { } while (0)
1662 static void tcp_undo_cwr(struct tcp_opt *tp, int undo)
1664 if (tp->prior_ssthresh) {
1665 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh<<1);
1667 if (undo && tp->prior_ssthresh > tp->snd_ssthresh) {
1668 tp->snd_ssthresh = tp->prior_ssthresh;
1669 TCP_ECN_withdraw_cwr(tp);
1672 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
1674 tcp_moderate_cwnd(tp);
1675 tp->snd_cwnd_stamp = tcp_time_stamp;
1678 static inline int tcp_may_undo(struct tcp_opt *tp)
1680 return tp->undo_marker &&
1681 (!tp->undo_retrans || tcp_packet_delayed(tp));
1684 /* People celebrate: "We love our President!" */
1685 static int tcp_try_undo_recovery(struct sock *sk, struct tcp_opt *tp)
1687 if (tcp_may_undo(tp)) {
1688 /* Happy end! We did not retransmit anything
1689 * or our original transmission succeeded.
1691 DBGUNDO(sk, tp, tp->ca_state == TCP_CA_Loss ? "loss" : "retrans");
1692 tcp_undo_cwr(tp, 1);
1693 if (tp->ca_state == TCP_CA_Loss)
1694 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
1696 NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO);
1697 tp->undo_marker = 0;
1699 if (tp->snd_una == tp->high_seq && IsReno(tp)) {
1700 /* Hold old state until something *above* high_seq
1701 * is ACKed. For Reno it is MUST to prevent false
1702 * fast retransmits (RFC2582). SACK TCP is safe. */
1703 tcp_moderate_cwnd(tp);
1706 tcp_set_ca_state(tp, TCP_CA_Open);
1710 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
1711 static void tcp_try_undo_dsack(struct sock *sk, struct tcp_opt *tp)
1713 if (tp->undo_marker && !tp->undo_retrans) {
1714 DBGUNDO(sk, tp, "D-SACK");
1715 tcp_undo_cwr(tp, 1);
1716 tp->undo_marker = 0;
1717 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO);
1721 /* Undo during fast recovery after partial ACK. */
1723 static int tcp_try_undo_partial(struct sock *sk, struct tcp_opt *tp, int acked)
1725 /* Partial ACK arrived. Force Hoe's retransmit. */
1726 int failed = IsReno(tp) || tcp_get_pcount(&tp->fackets_out)>tp->reordering;
1728 if (tcp_may_undo(tp)) {
1729 /* Plain luck! Hole if filled with delayed
1730 * packet, rather than with a retransmit.
1732 if (tcp_get_pcount(&tp->retrans_out) == 0)
1733 tp->retrans_stamp = 0;
1735 tcp_update_reordering(tp, tcp_fackets_out(tp)+acked, 1);
1737 DBGUNDO(sk, tp, "Hoe");
1738 tcp_undo_cwr(tp, 0);
1739 NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO);
1741 /* So... Do not make Hoe's retransmit yet.
1742 * If the first packet was delayed, the rest
1743 * ones are most probably delayed as well.
1750 /* Undo during loss recovery after partial ACK. */
1751 static int tcp_try_undo_loss(struct sock *sk, struct tcp_opt *tp)
1753 if (tcp_may_undo(tp)) {
1754 struct sk_buff *skb;
1755 sk_stream_for_retrans_queue(skb, sk) {
1756 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1758 DBGUNDO(sk, tp, "partial loss");
1759 tcp_set_pcount(&tp->lost_out, 0);
1760 tcp_set_pcount(&tp->left_out, tcp_get_pcount(&tp->sacked_out));
1761 tcp_undo_cwr(tp, 1);
1762 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
1763 tp->retransmits = 0;
1764 tp->undo_marker = 0;
1766 tcp_set_ca_state(tp, TCP_CA_Open);
1772 static __inline__ void tcp_complete_cwr(struct tcp_opt *tp)
1774 if (tcp_westwood_cwnd(tp))
1775 tp->snd_ssthresh = tp->snd_cwnd;
1777 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
1778 tp->snd_cwnd_stamp = tcp_time_stamp;
1781 static void tcp_try_to_open(struct sock *sk, struct tcp_opt *tp, int flag)
1783 tcp_set_pcount(&tp->left_out, tcp_get_pcount(&tp->sacked_out));
1785 if (tcp_get_pcount(&tp->retrans_out) == 0)
1786 tp->retrans_stamp = 0;
1791 if (tp->ca_state != TCP_CA_CWR) {
1792 int state = TCP_CA_Open;
1794 if (tcp_get_pcount(&tp->left_out) ||
1795 tcp_get_pcount(&tp->retrans_out) ||
1797 state = TCP_CA_Disorder;
1799 if (tp->ca_state != state) {
1800 tcp_set_ca_state(tp, state);
1801 tp->high_seq = tp->snd_nxt;
1803 tcp_moderate_cwnd(tp);
1809 /* Process an event, which can update packets-in-flight not trivially.
1810 * Main goal of this function is to calculate new estimate for left_out,
1811 * taking into account both packets sitting in receiver's buffer and
1812 * packets lost by network.
1814 * Besides that it does CWND reduction, when packet loss is detected
1815 * and changes state of machine.
1817 * It does _not_ decide what to send, it is made in function
1818 * tcp_xmit_retransmit_queue().
1821 tcp_fastretrans_alert(struct sock *sk, u32 prior_snd_una,
1822 int prior_packets, int flag)
1824 struct tcp_opt *tp = tcp_sk(sk);
1825 int is_dupack = (tp->snd_una == prior_snd_una && !(flag&FLAG_NOT_DUP));
1827 /* Some technical things:
1828 * 1. Reno does not count dupacks (sacked_out) automatically. */
1829 if (!tcp_get_pcount(&tp->packets_out))
1830 tcp_set_pcount(&tp->sacked_out, 0);
1831 /* 2. SACK counts snd_fack in packets inaccurately. */
1832 if (tcp_get_pcount(&tp->sacked_out) == 0)
1833 tcp_set_pcount(&tp->fackets_out, 0);
1835 /* Now state machine starts.
1836 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
1838 tp->prior_ssthresh = 0;
1840 /* B. In all the states check for reneging SACKs. */
1841 if (tcp_get_pcount(&tp->sacked_out) && tcp_check_sack_reneging(sk, tp))
1844 /* C. Process data loss notification, provided it is valid. */
1845 if ((flag&FLAG_DATA_LOST) &&
1846 before(tp->snd_una, tp->high_seq) &&
1847 tp->ca_state != TCP_CA_Open &&
1848 tcp_get_pcount(&tp->fackets_out) > tp->reordering) {
1849 tcp_mark_head_lost(sk, tp, tcp_get_pcount(&tp->fackets_out)-tp->reordering, tp->high_seq);
1850 NET_INC_STATS_BH(LINUX_MIB_TCPLOSS);
1853 /* D. Synchronize left_out to current state. */
1854 tcp_sync_left_out(tp);
1856 /* E. Check state exit conditions. State can be terminated
1857 * when high_seq is ACKed. */
1858 if (tp->ca_state == TCP_CA_Open) {
1859 if (!sysctl_tcp_frto)
1860 BUG_TRAP(tcp_get_pcount(&tp->retrans_out) == 0);
1861 tp->retrans_stamp = 0;
1862 } else if (!before(tp->snd_una, tp->high_seq)) {
1863 switch (tp->ca_state) {
1865 tp->retransmits = 0;
1866 if (tcp_try_undo_recovery(sk, tp))
1871 /* CWR is to be held something *above* high_seq
1872 * is ACKed for CWR bit to reach receiver. */
1873 if (tp->snd_una != tp->high_seq) {
1874 tcp_complete_cwr(tp);
1875 tcp_set_ca_state(tp, TCP_CA_Open);
1879 case TCP_CA_Disorder:
1880 tcp_try_undo_dsack(sk, tp);
1881 if (!tp->undo_marker ||
1882 /* For SACK case do not Open to allow to undo
1883 * catching for all duplicate ACKs. */
1884 IsReno(tp) || tp->snd_una != tp->high_seq) {
1885 tp->undo_marker = 0;
1886 tcp_set_ca_state(tp, TCP_CA_Open);
1890 case TCP_CA_Recovery:
1892 tcp_reset_reno_sack(tp);
1893 if (tcp_try_undo_recovery(sk, tp))
1895 tcp_complete_cwr(tp);
1900 /* F. Process state. */
1901 switch (tp->ca_state) {
1902 case TCP_CA_Recovery:
1903 if (prior_snd_una == tp->snd_una) {
1904 if (IsReno(tp) && is_dupack)
1905 tcp_add_reno_sack(tp);
1907 int acked = prior_packets -
1908 tcp_get_pcount(&tp->packets_out);
1910 tcp_remove_reno_sacks(sk, tp, acked);
1911 is_dupack = tcp_try_undo_partial(sk, tp, acked);
1915 if (flag&FLAG_DATA_ACKED)
1916 tp->retransmits = 0;
1917 if (!tcp_try_undo_loss(sk, tp)) {
1918 tcp_moderate_cwnd(tp);
1919 tcp_xmit_retransmit_queue(sk);
1922 if (tp->ca_state != TCP_CA_Open)
1924 /* Loss is undone; fall through to processing in Open state. */
1927 if (tp->snd_una != prior_snd_una)
1928 tcp_reset_reno_sack(tp);
1930 tcp_add_reno_sack(tp);
1933 if (tp->ca_state == TCP_CA_Disorder)
1934 tcp_try_undo_dsack(sk, tp);
1936 if (!tcp_time_to_recover(sk, tp)) {
1937 tcp_try_to_open(sk, tp, flag);
1941 /* Otherwise enter Recovery state */
1944 NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY);
1946 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY);
1948 tp->high_seq = tp->snd_nxt;
1949 tp->prior_ssthresh = 0;
1950 tp->undo_marker = tp->snd_una;
1951 tp->undo_retrans = tcp_get_pcount(&tp->retrans_out);
1953 if (tp->ca_state < TCP_CA_CWR) {
1954 if (!(flag&FLAG_ECE))
1955 tp->prior_ssthresh = tcp_current_ssthresh(tp);
1956 tp->snd_ssthresh = tcp_recalc_ssthresh(tp);
1957 TCP_ECN_queue_cwr(tp);
1960 tp->snd_cwnd_cnt = 0;
1961 tcp_set_ca_state(tp, TCP_CA_Recovery);
1964 if (is_dupack || tcp_head_timedout(sk, tp))
1965 tcp_update_scoreboard(sk, tp);
1967 tcp_xmit_retransmit_queue(sk);
1970 /* Read draft-ietf-tcplw-high-performance before mucking
1971 * with this code. (Superceeds RFC1323)
1973 static void tcp_ack_saw_tstamp(struct tcp_opt *tp, int flag)
1977 /* RTTM Rule: A TSecr value received in a segment is used to
1978 * update the averaged RTT measurement only if the segment
1979 * acknowledges some new data, i.e., only if it advances the
1980 * left edge of the send window.
1982 * See draft-ietf-tcplw-high-performance-00, section 3.3.
1983 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
1985 * Changed: reset backoff as soon as we see the first valid sample.
1986 * If we do not, we get strongly overstimated rto. With timestamps
1987 * samples are accepted even from very old segments: f.e., when rtt=1
1988 * increases to 8, we retransmit 5 times and after 8 seconds delayed
1989 * answer arrives rto becomes 120 seconds! If at least one of segments
1990 * in window is lost... Voila. --ANK (010210)
1992 seq_rtt = tcp_time_stamp - tp->rcv_tsecr;
1993 tcp_rtt_estimator(tp, seq_rtt);
1999 static void tcp_ack_no_tstamp(struct tcp_opt *tp, u32 seq_rtt, int flag)
2001 /* We don't have a timestamp. Can only use
2002 * packets that are not retransmitted to determine
2003 * rtt estimates. Also, we must not reset the
2004 * backoff for rto until we get a non-retransmitted
2005 * packet. This allows us to deal with a situation
2006 * where the network delay has increased suddenly.
2007 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
2010 if (flag & FLAG_RETRANS_DATA_ACKED)
2013 tcp_rtt_estimator(tp, seq_rtt);
2019 static __inline__ void
2020 tcp_ack_update_rtt(struct tcp_opt *tp, int flag, s32 seq_rtt)
2022 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
2023 if (tp->saw_tstamp && tp->rcv_tsecr)
2024 tcp_ack_saw_tstamp(tp, flag);
2025 else if (seq_rtt >= 0)
2026 tcp_ack_no_tstamp(tp, seq_rtt, flag);
2030 * Compute congestion window to use.
2032 * This is from the implementation of BICTCP in
2033 * Lison-Xu, Kahaled Harfoush, and Injog Rhee.
2034 * "Binary Increase Congestion Control for Fast, Long Distance
2035 * Networks" in InfoComm 2004
2037 * http://www.csc.ncsu.edu/faculty/rhee/export/bitcp.pdf
2039 * Unless BIC is enabled and congestion window is large
2040 * this behaves the same as the original Reno.
2042 static inline __u32 bictcp_cwnd(struct tcp_opt *tp)
2044 /* orignal Reno behaviour */
2045 if (!tcp_is_bic(tp))
2046 return tp->snd_cwnd;
2048 if (tp->bictcp.last_cwnd == tp->snd_cwnd &&
2049 (s32)(tcp_time_stamp - tp->bictcp.last_stamp) <= (HZ>>5))
2050 return tp->bictcp.cnt;
2052 tp->bictcp.last_cwnd = tp->snd_cwnd;
2053 tp->bictcp.last_stamp = tcp_time_stamp;
2055 /* start off normal */
2056 if (tp->snd_cwnd <= sysctl_tcp_bic_low_window)
2057 tp->bictcp.cnt = tp->snd_cwnd;
2059 /* binary increase */
2060 else if (tp->snd_cwnd < tp->bictcp.last_max_cwnd) {
2061 __u32 dist = (tp->bictcp.last_max_cwnd - tp->snd_cwnd)
2064 if (dist > BICTCP_MAX_INCREMENT)
2065 /* linear increase */
2066 tp->bictcp.cnt = tp->snd_cwnd / BICTCP_MAX_INCREMENT;
2067 else if (dist <= 1U)
2068 /* binary search increase */
2069 tp->bictcp.cnt = tp->snd_cwnd * BICTCP_FUNC_OF_MIN_INCR
2072 /* binary search increase */
2073 tp->bictcp.cnt = tp->snd_cwnd / dist;
2075 /* slow start amd linear increase */
2076 if (tp->snd_cwnd < tp->bictcp.last_max_cwnd + BICTCP_B)
2078 tp->bictcp.cnt = tp->snd_cwnd * BICTCP_FUNC_OF_MIN_INCR
2080 else if (tp->snd_cwnd < tp->bictcp.last_max_cwnd
2081 + BICTCP_MAX_INCREMENT*(BICTCP_B-1))
2083 tp->bictcp.cnt = tp->snd_cwnd * (BICTCP_B-1)
2084 / (tp->snd_cwnd-tp->bictcp.last_max_cwnd);
2086 /* linear increase */
2087 tp->bictcp.cnt = tp->snd_cwnd / BICTCP_MAX_INCREMENT;
2089 return tp->bictcp.cnt;
2092 /* This is Jacobson's slow start and congestion avoidance.
2093 * SIGCOMM '88, p. 328.
2095 static __inline__ void reno_cong_avoid(struct tcp_opt *tp)
2097 if (tp->snd_cwnd <= tp->snd_ssthresh) {
2098 /* In "safe" area, increase. */
2099 if (tp->snd_cwnd < tp->snd_cwnd_clamp)
2102 /* In dangerous area, increase slowly.
2103 * In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd
2105 if (tp->snd_cwnd_cnt >= bictcp_cwnd(tp)) {
2106 if (tp->snd_cwnd < tp->snd_cwnd_clamp)
2112 tp->snd_cwnd_stamp = tcp_time_stamp;
2115 /* This is based on the congestion detection/avoidance scheme described in
2116 * Lawrence S. Brakmo and Larry L. Peterson.
2117 * "TCP Vegas: End to end congestion avoidance on a global internet."
2118 * IEEE Journal on Selected Areas in Communication, 13(8):1465--1480,
2119 * October 1995. Available from:
2120 * ftp://ftp.cs.arizona.edu/xkernel/Papers/jsac.ps
2122 * See http://www.cs.arizona.edu/xkernel/ for their implementation.
2123 * The main aspects that distinguish this implementation from the
2124 * Arizona Vegas implementation are:
2125 * o We do not change the loss detection or recovery mechanisms of
2126 * Linux in any way. Linux already recovers from losses quite well,
2127 * using fine-grained timers, NewReno, and FACK.
2128 * o To avoid the performance penalty imposed by increasing cwnd
2129 * only every-other RTT during slow start, we increase during
2130 * every RTT during slow start, just like Reno.
2131 * o Largely to allow continuous cwnd growth during slow start,
2132 * we use the rate at which ACKs come back as the "actual"
2133 * rate, rather than the rate at which data is sent.
2134 * o To speed convergence to the right rate, we set the cwnd
2135 * to achieve the right ("actual") rate when we exit slow start.
2136 * o To filter out the noise caused by delayed ACKs, we use the
2137 * minimum RTT sample observed during the last RTT to calculate
2139 * o When the sender re-starts from idle, it waits until it has
2140 * received ACKs for an entire flight of new data before making
2141 * a cwnd adjustment decision. The original Vegas implementation
2142 * assumed senders never went idle.
2144 static void vegas_cong_avoid(struct tcp_opt *tp, u32 ack, u32 seq_rtt)
2146 /* The key players are v_beg_snd_una and v_beg_snd_nxt.
2148 * These are so named because they represent the approximate values
2149 * of snd_una and snd_nxt at the beginning of the current RTT. More
2150 * precisely, they represent the amount of data sent during the RTT.
2151 * At the end of the RTT, when we receive an ACK for v_beg_snd_nxt,
2152 * we will calculate that (v_beg_snd_nxt - v_beg_snd_una) outstanding
2153 * bytes of data have been ACKed during the course of the RTT, giving
2154 * an "actual" rate of:
2156 * (v_beg_snd_nxt - v_beg_snd_una) / (rtt duration)
2158 * Unfortunately, v_beg_snd_una is not exactly equal to snd_una,
2159 * because delayed ACKs can cover more than one segment, so they
2160 * don't line up nicely with the boundaries of RTTs.
2162 * Another unfortunate fact of life is that delayed ACKs delay the
2163 * advance of the left edge of our send window, so that the number
2164 * of bytes we send in an RTT is often less than our cwnd will allow.
2165 * So we keep track of our cwnd separately, in v_beg_snd_cwnd.
2168 if (after(ack, tp->vegas.beg_snd_nxt)) {
2169 /* Do the Vegas once-per-RTT cwnd adjustment. */
2170 u32 old_wnd, old_snd_cwnd;
2173 /* Here old_wnd is essentially the window of data that was
2174 * sent during the previous RTT, and has all
2175 * been acknowledged in the course of the RTT that ended
2176 * with the ACK we just received. Likewise, old_snd_cwnd
2177 * is the cwnd during the previous RTT.
2179 old_wnd = (tp->vegas.beg_snd_nxt - tp->vegas.beg_snd_una) /
2181 old_snd_cwnd = tp->vegas.beg_snd_cwnd;
2183 /* Save the extent of the current window so we can use this
2184 * at the end of the next RTT.
2186 tp->vegas.beg_snd_una = tp->vegas.beg_snd_nxt;
2187 tp->vegas.beg_snd_nxt = tp->snd_nxt;
2188 tp->vegas.beg_snd_cwnd = tp->snd_cwnd;
2190 /* Take into account the current RTT sample too, to
2191 * decrease the impact of delayed acks. This double counts
2192 * this sample since we count it for the next window as well,
2193 * but that's not too awful, since we're taking the min,
2194 * rather than averaging.
2196 vegas_rtt_calc(tp, seq_rtt);
2198 /* We do the Vegas calculations only if we got enough RTT
2199 * samples that we can be reasonably sure that we got
2200 * at least one RTT sample that wasn't from a delayed ACK.
2201 * If we only had 2 samples total,
2202 * then that means we're getting only 1 ACK per RTT, which
2203 * means they're almost certainly delayed ACKs.
2204 * If we have 3 samples, we should be OK.
2207 if (tp->vegas.cntRTT <= 2) {
2208 /* We don't have enough RTT samples to do the Vegas
2209 * calculation, so we'll behave like Reno.
2211 if (tp->snd_cwnd > tp->snd_ssthresh)
2214 u32 rtt, target_cwnd, diff;
2216 /* We have enough RTT samples, so, using the Vegas
2217 * algorithm, we determine if we should increase or
2218 * decrease cwnd, and by how much.
2221 /* Pluck out the RTT we are using for the Vegas
2222 * calculations. This is the min RTT seen during the
2223 * last RTT. Taking the min filters out the effects
2224 * of delayed ACKs, at the cost of noticing congestion
2227 rtt = tp->vegas.minRTT;
2229 /* Calculate the cwnd we should have, if we weren't
2233 * (actual rate in segments) * baseRTT
2234 * We keep it as a fixed point number with
2235 * V_PARAM_SHIFT bits to the right of the binary point.
2237 target_cwnd = ((old_wnd * tp->vegas.baseRTT)
2238 << V_PARAM_SHIFT) / rtt;
2240 /* Calculate the difference between the window we had,
2241 * and the window we would like to have. This quantity
2242 * is the "Diff" from the Arizona Vegas papers.
2244 * Again, this is a fixed point number with
2245 * V_PARAM_SHIFT bits to the right of the binary
2248 diff = (old_wnd << V_PARAM_SHIFT) - target_cwnd;
2250 if (tp->snd_cwnd < tp->snd_ssthresh) {
2252 if (diff > sysctl_tcp_vegas_gamma) {
2253 /* Going too fast. Time to slow down
2254 * and switch to congestion avoidance.
2256 tp->snd_ssthresh = 2;
2258 /* Set cwnd to match the actual rate
2260 * cwnd = (actual rate) * baseRTT
2261 * Then we add 1 because the integer
2262 * truncation robs us of full link
2265 tp->snd_cwnd = min(tp->snd_cwnd,
2271 /* Congestion avoidance. */
2274 /* Figure out where we would like cwnd
2277 if (diff > sysctl_tcp_vegas_beta) {
2278 /* The old window was too fast, so
2281 next_snd_cwnd = old_snd_cwnd - 1;
2282 } else if (diff < sysctl_tcp_vegas_alpha) {
2283 /* We don't have enough extra packets
2284 * in the network, so speed up.
2286 next_snd_cwnd = old_snd_cwnd + 1;
2288 /* Sending just as fast as we
2291 next_snd_cwnd = old_snd_cwnd;
2294 /* Adjust cwnd upward or downward, toward the
2297 if (next_snd_cwnd > tp->snd_cwnd)
2299 else if (next_snd_cwnd < tp->snd_cwnd)
2304 /* Wipe the slate clean for the next RTT. */
2305 tp->vegas.cntRTT = 0;
2306 tp->vegas.minRTT = 0x7fffffff;
2309 /* The following code is executed for every ack we receive,
2310 * except for conditions checked in should_advance_cwnd()
2311 * before the call to tcp_cong_avoid(). Mainly this means that
2312 * we only execute this code if the ack actually acked some
2316 /* If we are in slow start, increase our cwnd in response to this ACK.
2317 * (If we are not in slow start then we are in congestion avoidance,
2318 * and adjust our congestion window only once per RTT. See the code
2321 if (tp->snd_cwnd <= tp->snd_ssthresh)
2324 /* to keep cwnd from growing without bound */
2325 tp->snd_cwnd = min_t(u32, tp->snd_cwnd, tp->snd_cwnd_clamp);
2327 /* Make sure that we are never so timid as to reduce our cwnd below
2330 * Going below 2 MSS would risk huge delayed ACKs from our receiver.
2332 tp->snd_cwnd = max(tp->snd_cwnd, 2U);
2334 tp->snd_cwnd_stamp = tcp_time_stamp;
2337 static inline void tcp_cong_avoid(struct tcp_opt *tp, u32 ack, u32 seq_rtt)
2339 if (tcp_vegas_enabled(tp))
2340 vegas_cong_avoid(tp, ack, seq_rtt);
2342 reno_cong_avoid(tp);
2345 /* Restart timer after forward progress on connection.
2346 * RFC2988 recommends to restart timer to now+rto.
2349 static __inline__ void tcp_ack_packets_out(struct sock *sk, struct tcp_opt *tp)
2351 if (!tcp_get_pcount(&tp->packets_out)) {
2352 tcp_clear_xmit_timer(sk, TCP_TIME_RETRANS);
2354 tcp_reset_xmit_timer(sk, TCP_TIME_RETRANS, tp->rto);
2358 /* There is one downside to this scheme. Although we keep the
2359 * ACK clock ticking, adjusting packet counters and advancing
2360 * congestion window, we do not liberate socket send buffer
2363 * Mucking with skb->truesize and sk->sk_wmem_alloc et al.
2364 * then making a write space wakeup callback is a possible
2365 * future enhancement. WARNING: it is not trivial to make.
2367 static int tcp_tso_acked(struct sock *sk, struct sk_buff *skb,
2368 __u32 now, __s32 *seq_rtt)
2370 struct tcp_opt *tp = tcp_sk(sk);
2371 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
2372 __u32 seq = tp->snd_una;
2373 __u32 packets_acked;
2376 /* If we get here, the whole TSO packet has not been
2379 BUG_ON(!after(scb->end_seq, seq));
2381 packets_acked = tcp_skb_pcount(skb);
2382 if (tcp_trim_head(sk, skb, seq - scb->seq))
2384 packets_acked -= tcp_skb_pcount(skb);
2386 if (packets_acked) {
2387 __u8 sacked = scb->sacked;
2389 acked |= FLAG_DATA_ACKED;
2391 if (sacked & TCPCB_RETRANS) {
2392 if (sacked & TCPCB_SACKED_RETRANS)
2393 tcp_dec_pcount_explicit(&tp->retrans_out,
2395 acked |= FLAG_RETRANS_DATA_ACKED;
2397 } else if (*seq_rtt < 0)
2398 *seq_rtt = now - scb->when;
2399 if (sacked & TCPCB_SACKED_ACKED)
2400 tcp_dec_pcount_explicit(&tp->sacked_out,
2402 if (sacked & TCPCB_LOST)
2403 tcp_dec_pcount_explicit(&tp->lost_out,
2405 if (sacked & TCPCB_URG) {
2407 !before(seq, tp->snd_up))
2410 } else if (*seq_rtt < 0)
2411 *seq_rtt = now - scb->when;
2413 if (tcp_get_pcount(&tp->fackets_out)) {
2414 __u32 dval = min(tcp_get_pcount(&tp->fackets_out),
2416 tcp_dec_pcount_explicit(&tp->fackets_out, dval);
2418 tcp_dec_pcount_explicit(&tp->packets_out, packets_acked);
2420 BUG_ON(tcp_skb_pcount(skb) == 0);
2421 BUG_ON(!before(scb->seq, scb->end_seq));
2428 /* Remove acknowledged frames from the retransmission queue. */
2429 static int tcp_clean_rtx_queue(struct sock *sk, __s32 *seq_rtt_p)
2431 struct tcp_opt *tp = tcp_sk(sk);
2432 struct sk_buff *skb;
2433 __u32 now = tcp_time_stamp;
2437 while ((skb = skb_peek(&sk->sk_write_queue)) &&
2438 skb != sk->sk_send_head) {
2439 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
2440 __u8 sacked = scb->sacked;
2442 /* If our packet is before the ack sequence we can
2443 * discard it as it's confirmed to have arrived at
2446 if (after(scb->end_seq, tp->snd_una)) {
2447 if (tcp_skb_pcount(skb) > 1)
2448 acked |= tcp_tso_acked(sk, skb,
2453 /* Initial outgoing SYN's get put onto the write_queue
2454 * just like anything else we transmit. It is not
2455 * true data, and if we misinform our callers that
2456 * this ACK acks real data, we will erroneously exit
2457 * connection startup slow start one packet too
2458 * quickly. This is severely frowned upon behavior.
2460 if (!(scb->flags & TCPCB_FLAG_SYN)) {
2461 acked |= FLAG_DATA_ACKED;
2463 acked |= FLAG_SYN_ACKED;
2464 tp->retrans_stamp = 0;
2468 if (sacked & TCPCB_RETRANS) {
2469 if(sacked & TCPCB_SACKED_RETRANS)
2470 tcp_dec_pcount(&tp->retrans_out, skb);
2471 acked |= FLAG_RETRANS_DATA_ACKED;
2473 } else if (seq_rtt < 0)
2474 seq_rtt = now - scb->when;
2475 if (sacked & TCPCB_SACKED_ACKED)
2476 tcp_dec_pcount(&tp->sacked_out, skb);
2477 if (sacked & TCPCB_LOST)
2478 tcp_dec_pcount(&tp->lost_out, skb);
2479 if (sacked & TCPCB_URG) {
2481 !before(scb->end_seq, tp->snd_up))
2484 } else if (seq_rtt < 0)
2485 seq_rtt = now - scb->when;
2486 tcp_dec_pcount_approx(&tp->fackets_out, skb);
2487 tcp_packets_out_dec(tp, skb);
2488 __skb_unlink(skb, skb->list);
2489 sk_stream_free_skb(sk, skb);
2492 if (acked&FLAG_ACKED) {
2493 tcp_ack_update_rtt(tp, acked, seq_rtt);
2494 tcp_ack_packets_out(sk, tp);
2497 #if FASTRETRANS_DEBUG > 0
2498 BUG_TRAP((int)tcp_get_pcount(&tp->sacked_out) >= 0);
2499 BUG_TRAP((int)tcp_get_pcount(&tp->lost_out) >= 0);
2500 BUG_TRAP((int)tcp_get_pcount(&tp->retrans_out) >= 0);
2501 if (!tcp_get_pcount(&tp->packets_out) && tp->sack_ok) {
2502 if (tcp_get_pcount(&tp->lost_out)) {
2503 printk(KERN_DEBUG "Leak l=%u %d\n",
2504 tcp_get_pcount(&tp->lost_out),
2506 tcp_set_pcount(&tp->lost_out, 0);
2508 if (tcp_get_pcount(&tp->sacked_out)) {
2509 printk(KERN_DEBUG "Leak s=%u %d\n",
2510 tcp_get_pcount(&tp->sacked_out),
2512 tcp_set_pcount(&tp->sacked_out, 0);
2514 if (tcp_get_pcount(&tp->retrans_out)) {
2515 printk(KERN_DEBUG "Leak r=%u %d\n",
2516 tcp_get_pcount(&tp->retrans_out),
2518 tcp_set_pcount(&tp->retrans_out, 0);
2522 *seq_rtt_p = seq_rtt;
2526 static void tcp_ack_probe(struct sock *sk)
2528 struct tcp_opt *tp = tcp_sk(sk);
2530 /* Was it a usable window open? */
2532 if (!after(TCP_SKB_CB(sk->sk_send_head)->end_seq,
2533 tp->snd_una + tp->snd_wnd)) {
2535 tcp_clear_xmit_timer(sk, TCP_TIME_PROBE0);
2536 /* Socket must be waked up by subsequent tcp_data_snd_check().
2537 * This function is not for random using!
2540 tcp_reset_xmit_timer(sk, TCP_TIME_PROBE0,
2541 min(tp->rto << tp->backoff, TCP_RTO_MAX));
2545 static __inline__ int tcp_ack_is_dubious(struct tcp_opt *tp, int flag)
2547 return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
2548 tp->ca_state != TCP_CA_Open);
2551 static __inline__ int tcp_may_raise_cwnd(struct tcp_opt *tp, int flag)
2553 return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
2554 !((1<<tp->ca_state)&(TCPF_CA_Recovery|TCPF_CA_CWR));
2557 /* Check that window update is acceptable.
2558 * The function assumes that snd_una<=ack<=snd_next.
2560 static __inline__ int
2561 tcp_may_update_window(struct tcp_opt *tp, u32 ack, u32 ack_seq, u32 nwin)
2563 return (after(ack, tp->snd_una) ||
2564 after(ack_seq, tp->snd_wl1) ||
2565 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd));
2568 /* Update our send window.
2570 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
2571 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
2573 static int tcp_ack_update_window(struct sock *sk, struct tcp_opt *tp,
2574 struct sk_buff *skb, u32 ack, u32 ack_seq)
2577 u32 nwin = ntohs(skb->h.th->window);
2579 if (likely(!skb->h.th->syn))
2580 nwin <<= tp->snd_wscale;
2582 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
2583 flag |= FLAG_WIN_UPDATE;
2584 tcp_update_wl(tp, ack, ack_seq);
2586 if (tp->snd_wnd != nwin) {
2589 /* Note, it is the only place, where
2590 * fast path is recovered for sending TCP.
2592 tcp_fast_path_check(sk, tp);
2594 if (nwin > tp->max_window) {
2595 tp->max_window = nwin;
2596 tcp_sync_mss(sk, tp->pmtu_cookie);
2606 static void tcp_process_frto(struct sock *sk, u32 prior_snd_una)
2608 struct tcp_opt *tp = tcp_sk(sk);
2610 tcp_sync_left_out(tp);
2612 if (tp->snd_una == prior_snd_una ||
2613 !before(tp->snd_una, tp->frto_highmark)) {
2614 /* RTO was caused by loss, start retransmitting in
2615 * go-back-N slow start
2617 tcp_enter_frto_loss(sk);
2621 if (tp->frto_counter == 1) {
2622 /* First ACK after RTO advances the window: allow two new
2625 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
2627 /* Also the second ACK after RTO advances the window.
2628 * The RTO was likely spurious. Reduce cwnd and continue
2629 * in congestion avoidance
2631 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2632 tcp_moderate_cwnd(tp);
2635 /* F-RTO affects on two new ACKs following RTO.
2636 * At latest on third ACK the TCP behavor is back to normal.
2638 tp->frto_counter = (tp->frto_counter + 1) % 3;
2647 * This function initializes fields used in TCP Westwood+. We can't
2648 * get no information about RTTmin at this time so we simply set it to
2649 * TCP_WESTWOOD_INIT_RTT. This value was chosen to be too conservative
2650 * since in this way we're sure it will be updated in a consistent
2651 * way as soon as possible. It will reasonably happen within the first
2652 * RTT period of the connection lifetime.
2655 static void init_westwood(struct sock *sk)
2657 struct tcp_opt *tp = tcp_sk(sk);
2659 tp->westwood.bw_ns_est = 0;
2660 tp->westwood.bw_est = 0;
2661 tp->westwood.accounted = 0;
2662 tp->westwood.cumul_ack = 0;
2663 tp->westwood.rtt_win_sx = tcp_time_stamp;
2664 tp->westwood.rtt = TCP_WESTWOOD_INIT_RTT;
2665 tp->westwood.rtt_min = TCP_WESTWOOD_INIT_RTT;
2666 tp->westwood.snd_una = tp->snd_una;
2670 * @westwood_do_filter
2671 * Low-pass filter. Implemented using constant coeffients.
2674 static inline __u32 westwood_do_filter(__u32 a, __u32 b)
2676 return (((7 * a) + b) >> 3);
2679 static void westwood_filter(struct sock *sk, __u32 delta)
2681 struct tcp_opt *tp = tcp_sk(sk);
2683 tp->westwood.bw_ns_est =
2684 westwood_do_filter(tp->westwood.bw_ns_est,
2685 tp->westwood.bk / delta);
2686 tp->westwood.bw_est =
2687 westwood_do_filter(tp->westwood.bw_est,
2688 tp->westwood.bw_ns_est);
2692 * @westwood_update_rttmin
2693 * It is used to update RTTmin. In this case we MUST NOT use
2694 * WESTWOOD_RTT_MIN minimum bound since we could be on a LAN!
2697 static inline __u32 westwood_update_rttmin(const struct sock *sk)
2699 const struct tcp_opt *tp = tcp_sk(sk);
2700 __u32 rttmin = tp->westwood.rtt_min;
2702 if (tp->westwood.rtt != 0 &&
2703 (tp->westwood.rtt < tp->westwood.rtt_min || !rttmin))
2704 rttmin = tp->westwood.rtt;
2711 * Evaluate increases for dk.
2714 static inline __u32 westwood_acked(const struct sock *sk)
2716 const struct tcp_opt *tp = tcp_sk(sk);
2718 return tp->snd_una - tp->westwood.snd_una;
2722 * @westwood_new_window
2723 * It evaluates if we are receiving data inside the same RTT window as
2726 * It returns 0 if we are still evaluating samples in the same RTT
2727 * window, 1 if the sample has to be considered in the next window.
2730 static int westwood_new_window(const struct sock *sk)
2732 const struct tcp_opt *tp = tcp_sk(sk);
2737 left_bound = tp->westwood.rtt_win_sx;
2738 rtt = max(tp->westwood.rtt, (u32) TCP_WESTWOOD_RTT_MIN);
2741 * A RTT-window has passed. Be careful since if RTT is less than
2742 * 50ms we don't filter but we continue 'building the sample'.
2743 * This minimum limit was choosen since an estimation on small
2744 * time intervals is better to avoid...
2745 * Obvioulsy on a LAN we reasonably will always have
2746 * right_bound = left_bound + WESTWOOD_RTT_MIN
2749 if ((left_bound + rtt) < tcp_time_stamp)
2756 * @westwood_update_window
2757 * It updates RTT evaluation window if it is the right moment to do
2758 * it. If so it calls filter for evaluating bandwidth.
2761 static void __westwood_update_window(struct sock *sk, __u32 now)
2763 struct tcp_opt *tp = tcp_sk(sk);
2764 __u32 delta = now - tp->westwood.rtt_win_sx;
2767 if (tp->westwood.rtt)
2768 westwood_filter(sk, delta);
2770 tp->westwood.bk = 0;
2771 tp->westwood.rtt_win_sx = tcp_time_stamp;
2776 static void westwood_update_window(struct sock *sk, __u32 now)
2778 if (westwood_new_window(sk))
2779 __westwood_update_window(sk, now);
2783 * @__tcp_westwood_fast_bw
2784 * It is called when we are in fast path. In particular it is called when
2785 * header prediction is successfull. In such case infact update is
2786 * straight forward and doesn't need any particular care.
2789 void __tcp_westwood_fast_bw(struct sock *sk, struct sk_buff *skb)
2791 struct tcp_opt *tp = tcp_sk(sk);
2793 westwood_update_window(sk, tcp_time_stamp);
2795 tp->westwood.bk += westwood_acked(sk);
2796 tp->westwood.snd_una = tp->snd_una;
2797 tp->westwood.rtt_min = westwood_update_rttmin(sk);
2802 * @westwood_dupack_update
2803 * It updates accounted and cumul_ack when receiving a dupack.
2806 static void westwood_dupack_update(struct sock *sk)
2808 struct tcp_opt *tp = tcp_sk(sk);
2810 tp->westwood.accounted += tp->mss_cache_std;
2811 tp->westwood.cumul_ack = tp->mss_cache_std;
2814 static inline int westwood_may_change_cumul(struct tcp_opt *tp)
2816 return (tp->westwood.cumul_ack > tp->mss_cache_std);
2819 static inline void westwood_partial_update(struct tcp_opt *tp)
2821 tp->westwood.accounted -= tp->westwood.cumul_ack;
2822 tp->westwood.cumul_ack = tp->mss_cache_std;
2825 static inline void westwood_complete_update(struct tcp_opt *tp)
2827 tp->westwood.cumul_ack -= tp->westwood.accounted;
2828 tp->westwood.accounted = 0;
2832 * @westwood_acked_count
2833 * This function evaluates cumul_ack for evaluating dk in case of
2834 * delayed or partial acks.
2837 static inline __u32 westwood_acked_count(struct sock *sk)
2839 struct tcp_opt *tp = tcp_sk(sk);
2841 tp->westwood.cumul_ack = westwood_acked(sk);
2843 /* If cumul_ack is 0 this is a dupack since it's not moving
2846 if (!(tp->westwood.cumul_ack))
2847 westwood_dupack_update(sk);
2849 if (westwood_may_change_cumul(tp)) {
2850 /* Partial or delayed ack */
2851 if (tp->westwood.accounted >= tp->westwood.cumul_ack)
2852 westwood_partial_update(tp);
2854 westwood_complete_update(tp);
2857 tp->westwood.snd_una = tp->snd_una;
2859 return tp->westwood.cumul_ack;
2864 * @__tcp_westwood_slow_bw
2865 * It is called when something is going wrong..even if there could
2866 * be no problems! Infact a simple delayed packet may trigger a
2867 * dupack. But we need to be careful in such case.
2870 void __tcp_westwood_slow_bw(struct sock *sk, struct sk_buff *skb)
2872 struct tcp_opt *tp = tcp_sk(sk);
2874 westwood_update_window(sk, tcp_time_stamp);
2876 tp->westwood.bk += westwood_acked_count(sk);
2877 tp->westwood.rtt_min = westwood_update_rttmin(sk);
2880 /* This routine deals with incoming acks, but not outgoing ones. */
2881 static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag)
2883 struct tcp_opt *tp = tcp_sk(sk);
2884 u32 prior_snd_una = tp->snd_una;
2885 u32 ack_seq = TCP_SKB_CB(skb)->seq;
2886 u32 ack = TCP_SKB_CB(skb)->ack_seq;
2887 u32 prior_in_flight;
2891 /* If the ack is newer than sent or older than previous acks
2892 * then we can probably ignore it.
2894 if (after(ack, tp->snd_nxt))
2895 goto uninteresting_ack;
2897 if (before(ack, prior_snd_una))
2900 if (!(flag&FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
2901 /* Window is constant, pure forward advance.
2902 * No more checks are required.
2903 * Note, we use the fact that SND.UNA>=SND.WL2.
2905 tcp_update_wl(tp, ack, ack_seq);
2907 tcp_westwood_fast_bw(sk, skb);
2908 flag |= FLAG_WIN_UPDATE;
2910 NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS);
2912 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
2915 NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS);
2917 flag |= tcp_ack_update_window(sk, tp, skb, ack, ack_seq);
2919 if (TCP_SKB_CB(skb)->sacked)
2920 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
2922 if (TCP_ECN_rcv_ecn_echo(tp, skb->h.th))
2925 tcp_westwood_slow_bw(sk,skb);
2928 /* We passed data and got it acked, remove any soft error
2929 * log. Something worked...
2931 sk->sk_err_soft = 0;
2932 tp->rcv_tstamp = tcp_time_stamp;
2933 prior_packets = tcp_get_pcount(&tp->packets_out);
2937 prior_in_flight = tcp_packets_in_flight(tp);
2939 /* See if we can take anything off of the retransmit queue. */
2940 flag |= tcp_clean_rtx_queue(sk, &seq_rtt);
2942 if (tp->frto_counter)
2943 tcp_process_frto(sk, prior_snd_una);
2945 if (tcp_ack_is_dubious(tp, flag)) {
2946 /* Advanve CWND, if state allows this. */
2947 if ((flag & FLAG_DATA_ACKED) &&
2948 (tcp_vegas_enabled(tp) || prior_in_flight >= tp->snd_cwnd) &&
2949 tcp_may_raise_cwnd(tp, flag))
2950 tcp_cong_avoid(tp, ack, seq_rtt);
2951 tcp_fastretrans_alert(sk, prior_snd_una, prior_packets, flag);
2953 if ((flag & FLAG_DATA_ACKED) &&
2954 (tcp_vegas_enabled(tp) || prior_in_flight >= tp->snd_cwnd))
2955 tcp_cong_avoid(tp, ack, seq_rtt);
2958 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag&FLAG_NOT_DUP))
2959 dst_confirm(sk->sk_dst_cache);
2966 /* If this ack opens up a zero window, clear backoff. It was
2967 * being used to time the probes, and is probably far higher than
2968 * it needs to be for normal retransmission.
2970 if (sk->sk_send_head)
2975 if (TCP_SKB_CB(skb)->sacked)
2976 tcp_sacktag_write_queue(sk, skb, prior_snd_una);
2979 SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
2984 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
2985 * But, this can also be called on packets in the established flow when
2986 * the fast version below fails.
2988 void tcp_parse_options(struct sk_buff *skb, struct tcp_opt *tp, int estab)
2991 struct tcphdr *th = skb->h.th;
2992 int length=(th->doff*4)-sizeof(struct tcphdr);
2994 ptr = (unsigned char *)(th + 1);
3004 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3009 if (opsize < 2) /* "silly options" */
3011 if (opsize > length)
3012 return; /* don't parse partial options */
3015 if(opsize==TCPOLEN_MSS && th->syn && !estab) {
3016 u16 in_mss = ntohs(*(__u16 *)ptr);
3018 if (tp->user_mss && tp->user_mss < in_mss)
3019 in_mss = tp->user_mss;
3020 tp->mss_clamp = in_mss;
3025 if(opsize==TCPOLEN_WINDOW && th->syn && !estab)
3026 if (sysctl_tcp_window_scaling) {
3028 tp->snd_wscale = *(__u8 *)ptr;
3029 if(tp->snd_wscale > 14) {
3031 printk(KERN_INFO "tcp_parse_options: Illegal window "
3032 "scaling value %d >14 received.\n",
3034 tp->snd_wscale = 14;
3038 case TCPOPT_TIMESTAMP:
3039 if(opsize==TCPOLEN_TIMESTAMP) {
3040 if ((estab && tp->tstamp_ok) ||
3041 (!estab && sysctl_tcp_timestamps)) {
3043 tp->rcv_tsval = ntohl(*(__u32 *)ptr);
3044 tp->rcv_tsecr = ntohl(*(__u32 *)(ptr+4));
3048 case TCPOPT_SACK_PERM:
3049 if(opsize==TCPOLEN_SACK_PERM && th->syn && !estab) {
3050 if (sysctl_tcp_sack) {
3058 if((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3059 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3061 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3070 /* Fast parse options. This hopes to only see timestamps.
3071 * If it is wrong it falls back on tcp_parse_options().
3073 static __inline__ int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th, struct tcp_opt *tp)
3075 if (th->doff == sizeof(struct tcphdr)>>2) {
3078 } else if (tp->tstamp_ok &&
3079 th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) {
3080 __u32 *ptr = (__u32 *)(th + 1);
3081 if (*ptr == ntohl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3082 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3085 tp->rcv_tsval = ntohl(*ptr);
3087 tp->rcv_tsecr = ntohl(*ptr);
3091 tcp_parse_options(skb, tp, 1);
3095 static __inline__ void
3096 tcp_store_ts_recent(struct tcp_opt *tp)
3098 tp->ts_recent = tp->rcv_tsval;
3099 tp->ts_recent_stamp = xtime.tv_sec;
3102 static __inline__ void
3103 tcp_replace_ts_recent(struct tcp_opt *tp, u32 seq)
3105 if (tp->saw_tstamp && !after(seq, tp->rcv_wup)) {
3106 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3107 * extra check below makes sure this can only happen
3108 * for pure ACK frames. -DaveM
3110 * Not only, also it occurs for expired timestamps.
3113 if((s32)(tp->rcv_tsval - tp->ts_recent) >= 0 ||
3114 xtime.tv_sec >= tp->ts_recent_stamp + TCP_PAWS_24DAYS)
3115 tcp_store_ts_recent(tp);
3119 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3121 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3122 * it can pass through stack. So, the following predicate verifies that
3123 * this segment is not used for anything but congestion avoidance or
3124 * fast retransmit. Moreover, we even are able to eliminate most of such
3125 * second order effects, if we apply some small "replay" window (~RTO)
3126 * to timestamp space.
3128 * All these measures still do not guarantee that we reject wrapped ACKs
3129 * on networks with high bandwidth, when sequence space is recycled fastly,
3130 * but it guarantees that such events will be very rare and do not affect
3131 * connection seriously. This doesn't look nice, but alas, PAWS is really
3134 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3135 * states that events when retransmit arrives after original data are rare.
3136 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3137 * the biggest problem on large power networks even with minor reordering.
3138 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3139 * up to bandwidth of 18Gigabit/sec. 8) ]
3142 static int tcp_disordered_ack(struct tcp_opt *tp, struct sk_buff *skb)
3144 struct tcphdr *th = skb->h.th;
3145 u32 seq = TCP_SKB_CB(skb)->seq;
3146 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3148 return (/* 1. Pure ACK with correct sequence number. */
3149 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3151 /* 2. ... and duplicate ACK. */
3152 ack == tp->snd_una &&
3154 /* 3. ... and does not update window. */
3155 !tcp_may_update_window(tp, ack, seq, ntohs(th->window)<<tp->snd_wscale) &&
3157 /* 4. ... and sits in replay window. */
3158 (s32)(tp->ts_recent - tp->rcv_tsval) <= (tp->rto*1024)/HZ);
3161 static __inline__ int tcp_paws_discard(struct tcp_opt *tp, struct sk_buff *skb)
3163 return ((s32)(tp->ts_recent - tp->rcv_tsval) > TCP_PAWS_WINDOW &&
3164 xtime.tv_sec < tp->ts_recent_stamp + TCP_PAWS_24DAYS &&
3165 !tcp_disordered_ack(tp, skb));
3168 /* Check segment sequence number for validity.
3170 * Segment controls are considered valid, if the segment
3171 * fits to the window after truncation to the window. Acceptability
3172 * of data (and SYN, FIN, of course) is checked separately.
3173 * See tcp_data_queue(), for example.
3175 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3176 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3177 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3178 * (borrowed from freebsd)
3181 static inline int tcp_sequence(struct tcp_opt *tp, u32 seq, u32 end_seq)
3183 return !before(end_seq, tp->rcv_wup) &&
3184 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3187 /* When we get a reset we do this. */
3188 static void tcp_reset(struct sock *sk)
3190 /* We want the right error as BSD sees it (and indeed as we do). */
3191 switch (sk->sk_state) {
3193 sk->sk_err = ECONNREFUSED;
3195 case TCP_CLOSE_WAIT:
3201 sk->sk_err = ECONNRESET;
3204 if (!sock_flag(sk, SOCK_DEAD))
3205 sk->sk_error_report(sk);
3211 * Process the FIN bit. This now behaves as it is supposed to work
3212 * and the FIN takes effect when it is validly part of sequence
3213 * space. Not before when we get holes.
3215 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3216 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3219 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3220 * close and we go into CLOSING (and later onto TIME-WAIT)
3222 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3224 static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th)
3226 struct tcp_opt *tp = tcp_sk(sk);
3228 tcp_schedule_ack(tp);
3230 sk->sk_shutdown |= RCV_SHUTDOWN;
3231 sock_set_flag(sk, SOCK_DONE);
3233 switch (sk->sk_state) {
3235 case TCP_ESTABLISHED:
3236 /* Move to CLOSE_WAIT */
3237 tcp_set_state(sk, TCP_CLOSE_WAIT);
3238 tp->ack.pingpong = 1;
3241 case TCP_CLOSE_WAIT:
3243 /* Received a retransmission of the FIN, do
3248 /* RFC793: Remain in the LAST-ACK state. */
3252 /* This case occurs when a simultaneous close
3253 * happens, we must ack the received FIN and
3254 * enter the CLOSING state.
3257 tcp_set_state(sk, TCP_CLOSING);
3260 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3262 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
3265 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3266 * cases we should never reach this piece of code.
3268 printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n",
3269 __FUNCTION__, sk->sk_state);
3273 /* It _is_ possible, that we have something out-of-order _after_ FIN.
3274 * Probably, we should reset in this case. For now drop them.
3276 __skb_queue_purge(&tp->out_of_order_queue);
3279 sk_stream_mem_reclaim(sk);
3281 if (!sock_flag(sk, SOCK_DEAD)) {
3282 sk->sk_state_change(sk);
3284 /* Do not send POLL_HUP for half duplex close. */
3285 if (sk->sk_shutdown == SHUTDOWN_MASK ||
3286 sk->sk_state == TCP_CLOSE)
3287 sk_wake_async(sk, 1, POLL_HUP);
3289 sk_wake_async(sk, 1, POLL_IN);
3293 static __inline__ int
3294 tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, u32 end_seq)
3296 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
3297 if (before(seq, sp->start_seq))
3298 sp->start_seq = seq;
3299 if (after(end_seq, sp->end_seq))
3300 sp->end_seq = end_seq;
3306 static __inline__ void tcp_dsack_set(struct tcp_opt *tp, u32 seq, u32 end_seq)
3308 if (tp->sack_ok && sysctl_tcp_dsack) {
3309 if (before(seq, tp->rcv_nxt))
3310 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT);
3312 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT);
3315 tp->duplicate_sack[0].start_seq = seq;
3316 tp->duplicate_sack[0].end_seq = end_seq;
3317 tp->eff_sacks = min(tp->num_sacks+1, 4-tp->tstamp_ok);
3321 static __inline__ void tcp_dsack_extend(struct tcp_opt *tp, u32 seq, u32 end_seq)
3324 tcp_dsack_set(tp, seq, end_seq);
3326 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
3329 static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb)
3331 struct tcp_opt *tp = tcp_sk(sk);
3333 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
3334 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3335 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
3336 tcp_enter_quickack_mode(tp);
3338 if (tp->sack_ok && sysctl_tcp_dsack) {
3339 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3341 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
3342 end_seq = tp->rcv_nxt;
3343 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, end_seq);
3350 /* These routines update the SACK block as out-of-order packets arrive or
3351 * in-order packets close up the sequence space.
3353 static void tcp_sack_maybe_coalesce(struct tcp_opt *tp)
3356 struct tcp_sack_block *sp = &tp->selective_acks[0];
3357 struct tcp_sack_block *swalk = sp+1;
3359 /* See if the recent change to the first SACK eats into
3360 * or hits the sequence space of other SACK blocks, if so coalesce.
3362 for (this_sack = 1; this_sack < tp->num_sacks; ) {
3363 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
3366 /* Zap SWALK, by moving every further SACK up by one slot.
3367 * Decrease num_sacks.
3370 tp->eff_sacks = min(tp->num_sacks+tp->dsack, 4-tp->tstamp_ok);
3371 for(i=this_sack; i < tp->num_sacks; i++)
3375 this_sack++, swalk++;
3379 static __inline__ void tcp_sack_swap(struct tcp_sack_block *sack1, struct tcp_sack_block *sack2)
3383 tmp = sack1->start_seq;
3384 sack1->start_seq = sack2->start_seq;
3385 sack2->start_seq = tmp;
3387 tmp = sack1->end_seq;
3388 sack1->end_seq = sack2->end_seq;
3389 sack2->end_seq = tmp;
3392 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
3394 struct tcp_opt *tp = tcp_sk(sk);
3395 struct tcp_sack_block *sp = &tp->selective_acks[0];
3396 int cur_sacks = tp->num_sacks;
3402 for (this_sack=0; this_sack<cur_sacks; this_sack++, sp++) {
3403 if (tcp_sack_extend(sp, seq, end_seq)) {
3404 /* Rotate this_sack to the first one. */
3405 for (; this_sack>0; this_sack--, sp--)
3406 tcp_sack_swap(sp, sp-1);
3408 tcp_sack_maybe_coalesce(tp);
3413 /* Could not find an adjacent existing SACK, build a new one,
3414 * put it at the front, and shift everyone else down. We
3415 * always know there is at least one SACK present already here.
3417 * If the sack array is full, forget about the last one.
3419 if (this_sack >= 4) {
3424 for(; this_sack > 0; this_sack--, sp--)
3428 /* Build the new head SACK, and we're done. */
3429 sp->start_seq = seq;
3430 sp->end_seq = end_seq;
3432 tp->eff_sacks = min(tp->num_sacks + tp->dsack, 4 - tp->tstamp_ok);
3435 /* RCV.NXT advances, some SACKs should be eaten. */
3437 static void tcp_sack_remove(struct tcp_opt *tp)
3439 struct tcp_sack_block *sp = &tp->selective_acks[0];
3440 int num_sacks = tp->num_sacks;
3443 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
3444 if (skb_queue_len(&tp->out_of_order_queue) == 0) {
3446 tp->eff_sacks = tp->dsack;
3450 for(this_sack = 0; this_sack < num_sacks; ) {
3451 /* Check if the start of the sack is covered by RCV.NXT. */
3452 if (!before(tp->rcv_nxt, sp->start_seq)) {
3455 /* RCV.NXT must cover all the block! */
3456 BUG_TRAP(!before(tp->rcv_nxt, sp->end_seq));
3458 /* Zap this SACK, by moving forward any other SACKS. */
3459 for (i=this_sack+1; i < num_sacks; i++)
3460 tp->selective_acks[i-1] = tp->selective_acks[i];
3467 if (num_sacks != tp->num_sacks) {
3468 tp->num_sacks = num_sacks;
3469 tp->eff_sacks = min(tp->num_sacks+tp->dsack, 4-tp->tstamp_ok);
3473 /* This one checks to see if we can put data from the
3474 * out_of_order queue into the receive_queue.
3476 static void tcp_ofo_queue(struct sock *sk)
3478 struct tcp_opt *tp = tcp_sk(sk);
3479 __u32 dsack_high = tp->rcv_nxt;
3480 struct sk_buff *skb;
3482 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
3483 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
3486 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
3487 __u32 dsack = dsack_high;
3488 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
3489 dsack_high = TCP_SKB_CB(skb)->end_seq;
3490 tcp_dsack_extend(tp, TCP_SKB_CB(skb)->seq, dsack);
3493 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3494 SOCK_DEBUG(sk, "ofo packet was already received \n");
3495 __skb_unlink(skb, skb->list);
3499 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
3500 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3501 TCP_SKB_CB(skb)->end_seq);
3503 __skb_unlink(skb, skb->list);
3504 __skb_queue_tail(&sk->sk_receive_queue, skb);
3505 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3507 tcp_fin(skb, sk, skb->h.th);
3511 static int tcp_prune_queue(struct sock *sk);
3513 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
3515 struct tcphdr *th = skb->h.th;
3516 struct tcp_opt *tp = tcp_sk(sk);
3519 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
3523 __skb_pull(skb, th->doff*4);
3525 TCP_ECN_accept_cwr(tp, skb);
3529 tp->eff_sacks = min_t(unsigned int, tp->num_sacks,
3533 /* Queue data for delivery to the user.
3534 * Packets in sequence go to the receive queue.
3535 * Out of sequence packets to the out_of_order_queue.
3537 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
3538 if (tcp_receive_window(tp) == 0)
3541 /* Ok. In sequence. In window. */
3542 if (tp->ucopy.task == current &&
3543 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
3544 sock_owned_by_user(sk) && !tp->urg_data) {
3545 int chunk = min_t(unsigned int, skb->len,
3548 __set_current_state(TASK_RUNNING);
3551 if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
3552 tp->ucopy.len -= chunk;
3553 tp->copied_seq += chunk;
3554 eaten = (chunk == skb->len && !th->fin);
3555 tcp_rcv_space_adjust(sk);
3563 (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3564 !sk_stream_rmem_schedule(sk, skb))) {
3565 if (tcp_prune_queue(sk) < 0 ||
3566 !sk_stream_rmem_schedule(sk, skb))
3569 sk_stream_set_owner_r(skb, sk);
3570 __skb_queue_tail(&sk->sk_receive_queue, skb);
3572 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3574 tcp_event_data_recv(sk, tp, skb);
3576 tcp_fin(skb, sk, th);
3578 if (skb_queue_len(&tp->out_of_order_queue)) {
3581 /* RFC2581. 4.2. SHOULD send immediate ACK, when
3582 * gap in queue is filled.
3584 if (!skb_queue_len(&tp->out_of_order_queue))
3585 tp->ack.pingpong = 0;
3589 tcp_sack_remove(tp);
3591 tcp_fast_path_check(sk, tp);
3595 else if (!sock_flag(sk, SOCK_DEAD))
3596 sk->sk_data_ready(sk, 0);
3600 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3601 /* A retransmit, 2nd most common case. Force an immediate ack. */
3602 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
3603 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3606 tcp_enter_quickack_mode(tp);
3607 tcp_schedule_ack(tp);
3613 /* Out of window. F.e. zero window probe. */
3614 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
3617 tcp_enter_quickack_mode(tp);
3619 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3620 /* Partial packet, seq < rcv_next < end_seq */
3621 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
3622 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3623 TCP_SKB_CB(skb)->end_seq);
3625 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
3627 /* If window is closed, drop tail of packet. But after
3628 * remembering D-SACK for its head made in previous line.
3630 if (!tcp_receive_window(tp))
3635 TCP_ECN_check_ce(tp, skb);
3637 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3638 !sk_stream_rmem_schedule(sk, skb)) {
3639 if (tcp_prune_queue(sk) < 0 ||
3640 !sk_stream_rmem_schedule(sk, skb))
3644 /* Disable header prediction. */
3646 tcp_schedule_ack(tp);
3648 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
3649 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3651 sk_stream_set_owner_r(skb, sk);
3653 if (!skb_peek(&tp->out_of_order_queue)) {
3654 /* Initial out of order segment, build 1 SACK. */
3659 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
3660 tp->selective_acks[0].end_seq =
3661 TCP_SKB_CB(skb)->end_seq;
3663 __skb_queue_head(&tp->out_of_order_queue,skb);
3665 struct sk_buff *skb1 = tp->out_of_order_queue.prev;
3666 u32 seq = TCP_SKB_CB(skb)->seq;
3667 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3669 if (seq == TCP_SKB_CB(skb1)->end_seq) {
3670 __skb_append(skb1, skb);
3672 if (!tp->num_sacks ||
3673 tp->selective_acks[0].end_seq != seq)
3676 /* Common case: data arrive in order after hole. */
3677 tp->selective_acks[0].end_seq = end_seq;
3681 /* Find place to insert this segment. */
3683 if (!after(TCP_SKB_CB(skb1)->seq, seq))
3685 } while ((skb1 = skb1->prev) !=
3686 (struct sk_buff*)&tp->out_of_order_queue);
3688 /* Do skb overlap to previous one? */
3689 if (skb1 != (struct sk_buff*)&tp->out_of_order_queue &&
3690 before(seq, TCP_SKB_CB(skb1)->end_seq)) {
3691 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
3692 /* All the bits are present. Drop. */
3694 tcp_dsack_set(tp, seq, end_seq);
3697 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
3698 /* Partial overlap. */
3699 tcp_dsack_set(tp, seq, TCP_SKB_CB(skb1)->end_seq);
3704 __skb_insert(skb, skb1, skb1->next, &tp->out_of_order_queue);
3706 /* And clean segments covered by new one as whole. */
3707 while ((skb1 = skb->next) !=
3708 (struct sk_buff*)&tp->out_of_order_queue &&
3709 after(end_seq, TCP_SKB_CB(skb1)->seq)) {
3710 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
3711 tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, end_seq);
3714 __skb_unlink(skb1, skb1->list);
3715 tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, TCP_SKB_CB(skb1)->end_seq);
3721 tcp_sack_new_ofo_skb(sk, seq, end_seq);
3725 /* Collapse contiguous sequence of skbs head..tail with
3726 * sequence numbers start..end.
3727 * Segments with FIN/SYN are not collapsed (only because this
3731 tcp_collapse(struct sock *sk, struct sk_buff *head,
3732 struct sk_buff *tail, u32 start, u32 end)
3734 struct sk_buff *skb;
3736 /* First, check that queue is collapsable and find
3737 * the point where collapsing can be useful. */
3738 for (skb = head; skb != tail; ) {
3739 /* No new bits? It is possible on ofo queue. */
3740 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
3741 struct sk_buff *next = skb->next;
3742 __skb_unlink(skb, skb->list);
3744 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
3749 /* The first skb to collapse is:
3751 * - bloated or contains data before "start" or
3752 * overlaps to the next one.
3754 if (!skb->h.th->syn && !skb->h.th->fin &&
3755 (tcp_win_from_space(skb->truesize) > skb->len ||
3756 before(TCP_SKB_CB(skb)->seq, start) ||
3757 (skb->next != tail &&
3758 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq)))
3761 /* Decided to skip this, advance start seq. */
3762 start = TCP_SKB_CB(skb)->end_seq;
3765 if (skb == tail || skb->h.th->syn || skb->h.th->fin)
3768 while (before(start, end)) {
3769 struct sk_buff *nskb;
3770 int header = skb_headroom(skb);
3771 int copy = (PAGE_SIZE - sizeof(struct sk_buff) -
3772 sizeof(struct skb_shared_info) - header - 31)&~15;
3774 /* Too big header? This can happen with IPv6. */
3777 if (end-start < copy)
3779 nskb = alloc_skb(copy+header, GFP_ATOMIC);
3782 skb_reserve(nskb, header);
3783 memcpy(nskb->head, skb->head, header);
3784 nskb->nh.raw = nskb->head + (skb->nh.raw-skb->head);
3785 nskb->h.raw = nskb->head + (skb->h.raw-skb->head);
3786 nskb->mac.raw = nskb->head + (skb->mac.raw-skb->head);
3787 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
3788 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
3789 __skb_insert(nskb, skb->prev, skb, skb->list);
3790 sk_stream_set_owner_r(nskb, sk);
3792 /* Copy data, releasing collapsed skbs. */
3794 int offset = start - TCP_SKB_CB(skb)->seq;
3795 int size = TCP_SKB_CB(skb)->end_seq - start;
3797 if (offset < 0) BUG();
3799 size = min(copy, size);
3800 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
3802 TCP_SKB_CB(nskb)->end_seq += size;
3806 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
3807 struct sk_buff *next = skb->next;
3808 __skb_unlink(skb, skb->list);
3810 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
3812 if (skb == tail || skb->h.th->syn || skb->h.th->fin)
3819 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
3820 * and tcp_collapse() them until all the queue is collapsed.
3822 static void tcp_collapse_ofo_queue(struct sock *sk)
3824 struct tcp_opt *tp = tcp_sk(sk);
3825 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
3826 struct sk_buff *head;
3832 start = TCP_SKB_CB(skb)->seq;
3833 end = TCP_SKB_CB(skb)->end_seq;
3839 /* Segment is terminated when we see gap or when
3840 * we are at the end of all the queue. */
3841 if (skb == (struct sk_buff *)&tp->out_of_order_queue ||
3842 after(TCP_SKB_CB(skb)->seq, end) ||
3843 before(TCP_SKB_CB(skb)->end_seq, start)) {
3844 tcp_collapse(sk, head, skb, start, end);
3846 if (skb == (struct sk_buff *)&tp->out_of_order_queue)
3848 /* Start new segment */
3849 start = TCP_SKB_CB(skb)->seq;
3850 end = TCP_SKB_CB(skb)->end_seq;
3852 if (before(TCP_SKB_CB(skb)->seq, start))
3853 start = TCP_SKB_CB(skb)->seq;
3854 if (after(TCP_SKB_CB(skb)->end_seq, end))
3855 end = TCP_SKB_CB(skb)->end_seq;
3860 /* Reduce allocated memory if we can, trying to get
3861 * the socket within its memory limits again.
3863 * Return less than zero if we should start dropping frames
3864 * until the socket owning process reads some of the data
3865 * to stabilize the situation.
3867 static int tcp_prune_queue(struct sock *sk)
3869 struct tcp_opt *tp = tcp_sk(sk);
3871 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
3873 NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED);
3875 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
3876 tcp_clamp_window(sk, tp);
3877 else if (tcp_memory_pressure)
3878 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
3880 tcp_collapse_ofo_queue(sk);
3881 tcp_collapse(sk, sk->sk_receive_queue.next,
3882 (struct sk_buff*)&sk->sk_receive_queue,
3883 tp->copied_seq, tp->rcv_nxt);
3884 sk_stream_mem_reclaim(sk);
3886 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
3889 /* Collapsing did not help, destructive actions follow.
3890 * This must not ever occur. */
3892 /* First, purge the out_of_order queue. */
3893 if (skb_queue_len(&tp->out_of_order_queue)) {
3894 NET_ADD_STATS_BH(LINUX_MIB_OFOPRUNED,
3895 skb_queue_len(&tp->out_of_order_queue));
3896 __skb_queue_purge(&tp->out_of_order_queue);
3898 /* Reset SACK state. A conforming SACK implementation will
3899 * do the same at a timeout based retransmit. When a connection
3900 * is in a sad state like this, we care only about integrity
3901 * of the connection not performance.
3905 sk_stream_mem_reclaim(sk);
3908 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
3911 /* If we are really being abused, tell the caller to silently
3912 * drop receive data on the floor. It will get retransmitted
3913 * and hopefully then we'll have sufficient space.
3915 NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED);
3917 /* Massive buffer overcommit. */
3923 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
3924 * As additional protections, we do not touch cwnd in retransmission phases,
3925 * and if application hit its sndbuf limit recently.
3927 void tcp_cwnd_application_limited(struct sock *sk)
3929 struct tcp_opt *tp = tcp_sk(sk);
3931 if (tp->ca_state == TCP_CA_Open &&
3932 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
3933 /* Limited by application or receiver window. */
3934 u32 win_used = max(tp->snd_cwnd_used, 2U);
3935 if (win_used < tp->snd_cwnd) {
3936 tp->snd_ssthresh = tcp_current_ssthresh(tp);
3937 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
3939 tp->snd_cwnd_used = 0;
3941 tp->snd_cwnd_stamp = tcp_time_stamp;
3944 EXPORT_SYMBOL_GPL(tcp_cwnd_application_limited);
3946 /* When incoming ACK allowed to free some skb from write_queue,
3947 * we remember this event in flag sk->sk_queue_shrunk and wake up socket
3948 * on the exit from tcp input handler.
3950 * PROBLEM: sndbuf expansion does not work well with largesend.
3952 static void tcp_new_space(struct sock *sk)
3954 struct tcp_opt *tp = tcp_sk(sk);
3956 if (tcp_get_pcount(&tp->packets_out) < tp->snd_cwnd &&
3957 !(sk->sk_userlocks & SOCK_SNDBUF_LOCK) &&
3958 !tcp_memory_pressure &&
3959 atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) {
3960 int sndmem = max_t(u32, tp->mss_clamp, tp->mss_cache_std) +
3961 MAX_TCP_HEADER + 16 + sizeof(struct sk_buff),
3962 demanded = max_t(unsigned int, tp->snd_cwnd,
3963 tp->reordering + 1);
3964 sndmem *= 2*demanded;
3965 if (sndmem > sk->sk_sndbuf)
3966 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
3967 tp->snd_cwnd_stamp = tcp_time_stamp;
3970 sk->sk_write_space(sk);
3973 static inline void tcp_check_space(struct sock *sk)
3975 if (sk->sk_queue_shrunk) {
3976 sk->sk_queue_shrunk = 0;
3977 if (sk->sk_socket &&
3978 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
3983 static void __tcp_data_snd_check(struct sock *sk, struct sk_buff *skb)
3985 struct tcp_opt *tp = tcp_sk(sk);
3987 if (after(TCP_SKB_CB(skb)->end_seq, tp->snd_una + tp->snd_wnd) ||
3988 tcp_packets_in_flight(tp) >= tp->snd_cwnd ||
3989 tcp_write_xmit(sk, tp->nonagle))
3990 tcp_check_probe_timer(sk, tp);
3993 static __inline__ void tcp_data_snd_check(struct sock *sk)
3995 struct sk_buff *skb = sk->sk_send_head;
3998 __tcp_data_snd_check(sk, skb);
3999 tcp_check_space(sk);
4003 * Check if sending an ack is needed.
4005 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
4007 struct tcp_opt *tp = tcp_sk(sk);
4009 /* More than one full frame received... */
4010 if (((tp->rcv_nxt - tp->rcv_wup) > tp->ack.rcv_mss
4011 /* ... and right edge of window advances far enough.
4012 * (tcp_recvmsg() will send ACK otherwise). Or...
4014 && __tcp_select_window(sk) >= tp->rcv_wnd) ||
4015 /* We ACK each frame or... */
4016 tcp_in_quickack_mode(tp) ||
4017 /* We have out of order data. */
4019 skb_peek(&tp->out_of_order_queue))) {
4020 /* Then ack it now */
4023 /* Else, send delayed ack. */
4024 tcp_send_delayed_ack(sk);
4028 static __inline__ void tcp_ack_snd_check(struct sock *sk)
4030 struct tcp_opt *tp = tcp_sk(sk);
4031 if (!tcp_ack_scheduled(tp)) {
4032 /* We sent a data segment already. */
4035 __tcp_ack_snd_check(sk, 1);
4039 * This routine is only called when we have urgent data
4040 * signalled. Its the 'slow' part of tcp_urg. It could be
4041 * moved inline now as tcp_urg is only called from one
4042 * place. We handle URGent data wrong. We have to - as
4043 * BSD still doesn't use the correction from RFC961.
4044 * For 1003.1g we should support a new option TCP_STDURG to permit
4045 * either form (or just set the sysctl tcp_stdurg).
4048 static void tcp_check_urg(struct sock * sk, struct tcphdr * th)
4050 struct tcp_opt *tp = tcp_sk(sk);
4051 u32 ptr = ntohs(th->urg_ptr);
4053 if (ptr && !sysctl_tcp_stdurg)
4055 ptr += ntohl(th->seq);
4057 /* Ignore urgent data that we've already seen and read. */
4058 if (after(tp->copied_seq, ptr))
4061 /* Do not replay urg ptr.
4063 * NOTE: interesting situation not covered by specs.
4064 * Misbehaving sender may send urg ptr, pointing to segment,
4065 * which we already have in ofo queue. We are not able to fetch
4066 * such data and will stay in TCP_URG_NOTYET until will be eaten
4067 * by recvmsg(). Seems, we are not obliged to handle such wicked
4068 * situations. But it is worth to think about possibility of some
4069 * DoSes using some hypothetical application level deadlock.
4071 if (before(ptr, tp->rcv_nxt))
4074 /* Do we already have a newer (or duplicate) urgent pointer? */
4075 if (tp->urg_data && !after(ptr, tp->urg_seq))
4078 /* Tell the world about our new urgent pointer. */
4081 /* We may be adding urgent data when the last byte read was
4082 * urgent. To do this requires some care. We cannot just ignore
4083 * tp->copied_seq since we would read the last urgent byte again
4084 * as data, nor can we alter copied_seq until this data arrives
4085 * or we break the sematics of SIOCATMARK (and thus sockatmark())
4087 * NOTE. Double Dutch. Rendering to plain English: author of comment
4088 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
4089 * and expect that both A and B disappear from stream. This is _wrong_.
4090 * Though this happens in BSD with high probability, this is occasional.
4091 * Any application relying on this is buggy. Note also, that fix "works"
4092 * only in this artificial test. Insert some normal data between A and B and we will
4093 * decline of BSD again. Verdict: it is better to remove to trap
4096 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
4097 !sock_flag(sk, SOCK_URGINLINE) &&
4098 tp->copied_seq != tp->rcv_nxt) {
4099 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
4101 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
4102 __skb_unlink(skb, skb->list);
4107 tp->urg_data = TCP_URG_NOTYET;
4110 /* Disable header prediction. */
4114 /* This is the 'fast' part of urgent handling. */
4115 static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th)
4117 struct tcp_opt *tp = tcp_sk(sk);
4119 /* Check if we get a new urgent pointer - normally not. */
4121 tcp_check_urg(sk,th);
4123 /* Do we wait for any urgent data? - normally not... */
4124 if (tp->urg_data == TCP_URG_NOTYET) {
4125 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
4128 /* Is the urgent pointer pointing into this packet? */
4129 if (ptr < skb->len) {
4131 if (skb_copy_bits(skb, ptr, &tmp, 1))
4133 tp->urg_data = TCP_URG_VALID | tmp;
4134 if (!sock_flag(sk, SOCK_DEAD))
4135 sk->sk_data_ready(sk, 0);
4140 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
4142 struct tcp_opt *tp = tcp_sk(sk);
4143 int chunk = skb->len - hlen;
4147 if (skb->ip_summed==CHECKSUM_UNNECESSARY)
4148 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
4150 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
4154 tp->ucopy.len -= chunk;
4155 tp->copied_seq += chunk;
4156 tcp_rcv_space_adjust(sk);
4163 static int __tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
4167 if (sock_owned_by_user(sk)) {
4169 result = __tcp_checksum_complete(skb);
4172 result = __tcp_checksum_complete(skb);
4177 static __inline__ int
4178 tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
4180 return skb->ip_summed != CHECKSUM_UNNECESSARY &&
4181 __tcp_checksum_complete_user(sk, skb);
4185 * TCP receive function for the ESTABLISHED state.
4187 * It is split into a fast path and a slow path. The fast path is
4189 * - A zero window was announced from us - zero window probing
4190 * is only handled properly in the slow path.
4191 * - Out of order segments arrived.
4192 * - Urgent data is expected.
4193 * - There is no buffer space left
4194 * - Unexpected TCP flags/window values/header lengths are received
4195 * (detected by checking the TCP header against pred_flags)
4196 * - Data is sent in both directions. Fast path only supports pure senders
4197 * or pure receivers (this means either the sequence number or the ack
4198 * value must stay constant)
4199 * - Unexpected TCP option.
4201 * When these conditions are not satisfied it drops into a standard
4202 * receive procedure patterned after RFC793 to handle all cases.
4203 * The first three cases are guaranteed by proper pred_flags setting,
4204 * the rest is checked inline. Fast processing is turned on in
4205 * tcp_data_queue when everything is OK.
4207 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
4208 struct tcphdr *th, unsigned len)
4210 struct tcp_opt *tp = tcp_sk(sk);
4213 * Header prediction.
4214 * The code loosely follows the one in the famous
4215 * "30 instruction TCP receive" Van Jacobson mail.
4217 * Van's trick is to deposit buffers into socket queue
4218 * on a device interrupt, to call tcp_recv function
4219 * on the receive process context and checksum and copy
4220 * the buffer to user space. smart...
4222 * Our current scheme is not silly either but we take the
4223 * extra cost of the net_bh soft interrupt processing...
4224 * We do checksum and copy also but from device to kernel.
4229 /* pred_flags is 0xS?10 << 16 + snd_wnd
4230 * if header_predition is to be made
4231 * 'S' will always be tp->tcp_header_len >> 2
4232 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
4233 * turn it off (when there are holes in the receive
4234 * space for instance)
4235 * PSH flag is ignored.
4238 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
4239 TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4240 int tcp_header_len = tp->tcp_header_len;
4242 /* Timestamp header prediction: tcp_header_len
4243 * is automatically equal to th->doff*4 due to pred_flags
4247 /* Check timestamp */
4248 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
4249 __u32 *ptr = (__u32 *)(th + 1);
4251 /* No? Slow path! */
4252 if (*ptr != ntohl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
4253 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP))
4258 tp->rcv_tsval = ntohl(*ptr);
4260 tp->rcv_tsecr = ntohl(*ptr);
4262 /* If PAWS failed, check it more carefully in slow path */
4263 if ((s32)(tp->rcv_tsval - tp->ts_recent) < 0)
4266 /* DO NOT update ts_recent here, if checksum fails
4267 * and timestamp was corrupted part, it will result
4268 * in a hung connection since we will drop all
4269 * future packets due to the PAWS test.
4273 if (len <= tcp_header_len) {
4274 /* Bulk data transfer: sender */
4275 if (len == tcp_header_len) {
4276 /* Predicted packet is in window by definition.
4277 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4278 * Hence, check seq<=rcv_wup reduces to:
4280 if (tcp_header_len ==
4281 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
4282 tp->rcv_nxt == tp->rcv_wup)
4283 tcp_store_ts_recent(tp);
4285 tcp_rcv_rtt_measure_ts(tp, skb);
4287 /* We know that such packets are checksummed
4290 tcp_ack(sk, skb, 0);
4292 tcp_data_snd_check(sk);
4294 } else { /* Header too small */
4295 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4301 if (tp->ucopy.task == current &&
4302 tp->copied_seq == tp->rcv_nxt &&
4303 len - tcp_header_len <= tp->ucopy.len &&
4304 sock_owned_by_user(sk)) {
4305 __set_current_state(TASK_RUNNING);
4307 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len)) {
4308 /* Predicted packet is in window by definition.
4309 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4310 * Hence, check seq<=rcv_wup reduces to:
4312 if (tcp_header_len ==
4313 (sizeof(struct tcphdr) +
4314 TCPOLEN_TSTAMP_ALIGNED) &&
4315 tp->rcv_nxt == tp->rcv_wup)
4316 tcp_store_ts_recent(tp);
4318 tcp_rcv_rtt_measure_ts(tp, skb);
4320 __skb_pull(skb, tcp_header_len);
4321 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4322 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER);
4327 if (tcp_checksum_complete_user(sk, skb))
4330 /* Predicted packet is in window by definition.
4331 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4332 * Hence, check seq<=rcv_wup reduces to:
4334 if (tcp_header_len ==
4335 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
4336 tp->rcv_nxt == tp->rcv_wup)
4337 tcp_store_ts_recent(tp);
4339 tcp_rcv_rtt_measure_ts(tp, skb);
4341 if ((int)skb->truesize > sk->sk_forward_alloc)
4344 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS);
4346 /* Bulk data transfer: receiver */
4347 __skb_pull(skb,tcp_header_len);
4348 __skb_queue_tail(&sk->sk_receive_queue, skb);
4349 sk_stream_set_owner_r(skb, sk);
4350 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4353 tcp_event_data_recv(sk, tp, skb);
4355 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
4356 /* Well, only one small jumplet in fast path... */
4357 tcp_ack(sk, skb, FLAG_DATA);
4358 tcp_data_snd_check(sk);
4359 if (!tcp_ack_scheduled(tp))
4364 if (tcp_in_quickack_mode(tp)) {
4367 tcp_send_delayed_ack(sk);
4370 __tcp_ack_snd_check(sk, 0);
4377 sk->sk_data_ready(sk, 0);
4383 if (len < (th->doff<<2) || tcp_checksum_complete_user(sk, skb))
4387 * RFC1323: H1. Apply PAWS check first.
4389 if (tcp_fast_parse_options(skb, th, tp) && tp->saw_tstamp &&
4390 tcp_paws_discard(tp, skb)) {
4392 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
4393 tcp_send_dupack(sk, skb);
4396 /* Resets are accepted even if PAWS failed.
4398 ts_recent update must be made after we are sure
4399 that the packet is in window.
4404 * Standard slow path.
4407 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
4408 /* RFC793, page 37: "In all states except SYN-SENT, all reset
4409 * (RST) segments are validated by checking their SEQ-fields."
4410 * And page 69: "If an incoming segment is not acceptable,
4411 * an acknowledgment should be sent in reply (unless the RST bit
4412 * is set, if so drop the segment and return)".
4415 tcp_send_dupack(sk, skb);
4424 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
4426 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4427 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4428 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
4435 tcp_ack(sk, skb, FLAG_SLOWPATH);
4437 tcp_rcv_rtt_measure_ts(tp, skb);
4439 /* Process urgent data. */
4440 tcp_urg(sk, skb, th);
4442 /* step 7: process the segment text */
4443 tcp_data_queue(sk, skb);
4445 tcp_data_snd_check(sk);
4446 tcp_ack_snd_check(sk);
4450 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4457 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
4458 struct tcphdr *th, unsigned len)
4460 struct tcp_opt *tp = tcp_sk(sk);
4461 int saved_clamp = tp->mss_clamp;
4463 tcp_parse_options(skb, tp, 0);
4467 * "If the state is SYN-SENT then
4468 * first check the ACK bit
4469 * If the ACK bit is set
4470 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
4471 * a reset (unless the RST bit is set, if so drop
4472 * the segment and return)"
4474 * We do not send data with SYN, so that RFC-correct
4477 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
4478 goto reset_and_undo;
4480 if (tp->saw_tstamp && tp->rcv_tsecr &&
4481 !between(tp->rcv_tsecr, tp->retrans_stamp,
4483 NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED);
4484 goto reset_and_undo;
4487 /* Now ACK is acceptable.
4489 * "If the RST bit is set
4490 * If the ACK was acceptable then signal the user "error:
4491 * connection reset", drop the segment, enter CLOSED state,
4492 * delete TCB, and return."
4501 * "fifth, if neither of the SYN or RST bits is set then
4502 * drop the segment and return."
4508 goto discard_and_undo;
4511 * "If the SYN bit is on ...
4512 * are acceptable then ...
4513 * (our SYN has been ACKed), change the connection
4514 * state to ESTABLISHED..."
4517 TCP_ECN_rcv_synack(tp, th);
4518 if (tp->ecn_flags&TCP_ECN_OK)
4519 sk->sk_no_largesend = 1;
4521 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
4522 tcp_ack(sk, skb, FLAG_SLOWPATH);
4524 /* Ok.. it's good. Set up sequence numbers and
4525 * move to established.
4527 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
4528 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
4530 /* RFC1323: The window in SYN & SYN/ACK segments is
4533 tp->snd_wnd = ntohs(th->window);
4534 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(skb)->seq);
4536 if (!tp->wscale_ok) {
4537 tp->snd_wscale = tp->rcv_wscale = 0;
4538 tp->window_clamp = min(tp->window_clamp, 65535U);
4541 if (tp->saw_tstamp) {
4543 tp->tcp_header_len =
4544 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
4545 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
4546 tcp_store_ts_recent(tp);
4548 tp->tcp_header_len = sizeof(struct tcphdr);
4551 if (tp->sack_ok && sysctl_tcp_fack)
4554 tcp_sync_mss(sk, tp->pmtu_cookie);
4555 tcp_initialize_rcv_mss(sk);
4557 /* Remember, tcp_poll() does not lock socket!
4558 * Change state from SYN-SENT only after copied_seq
4559 * is initialized. */
4560 tp->copied_seq = tp->rcv_nxt;
4562 tcp_set_state(sk, TCP_ESTABLISHED);
4564 /* Make sure socket is routed, for correct metrics. */
4565 tp->af_specific->rebuild_header(sk);
4567 tcp_init_metrics(sk);
4569 /* Prevent spurious tcp_cwnd_restart() on first data
4572 tp->lsndtime = tcp_time_stamp;
4574 tcp_init_buffer_space(sk);
4576 if (sock_flag(sk, SOCK_KEEPOPEN))
4577 tcp_reset_keepalive_timer(sk, keepalive_time_when(tp));
4579 if (!tp->snd_wscale)
4580 __tcp_fast_path_on(tp, tp->snd_wnd);
4584 if (!sock_flag(sk, SOCK_DEAD)) {
4585 sk->sk_state_change(sk);
4586 sk_wake_async(sk, 0, POLL_OUT);
4589 if (sk->sk_write_pending || tp->defer_accept || tp->ack.pingpong) {
4590 /* Save one ACK. Data will be ready after
4591 * several ticks, if write_pending is set.
4593 * It may be deleted, but with this feature tcpdumps
4594 * look so _wonderfully_ clever, that I was not able
4595 * to stand against the temptation 8) --ANK
4597 tcp_schedule_ack(tp);
4598 tp->ack.lrcvtime = tcp_time_stamp;
4599 tp->ack.ato = TCP_ATO_MIN;
4600 tcp_incr_quickack(tp);
4601 tcp_enter_quickack_mode(tp);
4602 tcp_reset_xmit_timer(sk, TCP_TIME_DACK, TCP_DELACK_MAX);
4613 /* No ACK in the segment */
4617 * "If the RST bit is set
4619 * Otherwise (no ACK) drop the segment and return."
4622 goto discard_and_undo;
4626 if (tp->ts_recent_stamp && tp->saw_tstamp && tcp_paws_check(tp, 0))
4627 goto discard_and_undo;
4630 /* We see SYN without ACK. It is attempt of
4631 * simultaneous connect with crossed SYNs.
4632 * Particularly, it can be connect to self.
4634 tcp_set_state(sk, TCP_SYN_RECV);
4636 if (tp->saw_tstamp) {
4638 tcp_store_ts_recent(tp);
4639 tp->tcp_header_len =
4640 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
4642 tp->tcp_header_len = sizeof(struct tcphdr);
4645 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
4646 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
4648 /* RFC1323: The window in SYN & SYN/ACK segments is
4651 tp->snd_wnd = ntohs(th->window);
4652 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
4653 tp->max_window = tp->snd_wnd;
4655 TCP_ECN_rcv_syn(tp, th);
4656 if (tp->ecn_flags&TCP_ECN_OK)
4657 sk->sk_no_largesend = 1;
4659 tcp_sync_mss(sk, tp->pmtu_cookie);
4660 tcp_initialize_rcv_mss(sk);
4663 tcp_send_synack(sk);
4665 /* Note, we could accept data and URG from this segment.
4666 * There are no obstacles to make this.
4668 * However, if we ignore data in ACKless segments sometimes,
4669 * we have no reasons to accept it sometimes.
4670 * Also, seems the code doing it in step6 of tcp_rcv_state_process
4671 * is not flawless. So, discard packet for sanity.
4672 * Uncomment this return to process the data.
4679 /* "fifth, if neither of the SYN or RST bits is set then
4680 * drop the segment and return."
4684 tcp_clear_options(tp);
4685 tp->mss_clamp = saved_clamp;
4689 tcp_clear_options(tp);
4690 tp->mss_clamp = saved_clamp;
4696 * This function implements the receiving procedure of RFC 793 for
4697 * all states except ESTABLISHED and TIME_WAIT.
4698 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
4699 * address independent.
4702 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
4703 struct tcphdr *th, unsigned len)
4705 struct tcp_opt *tp = tcp_sk(sk);
4710 switch (sk->sk_state) {
4722 if(tp->af_specific->conn_request(sk, skb) < 0)
4728 /* Now we have several options: In theory there is
4729 * nothing else in the frame. KA9Q has an option to
4730 * send data with the syn, BSD accepts data with the
4731 * syn up to the [to be] advertised window and
4732 * Solaris 2.1 gives you a protocol error. For now
4733 * we just ignore it, that fits the spec precisely
4734 * and avoids incompatibilities. It would be nice in
4735 * future to drop through and process the data.
4737 * Now that TTCP is starting to be used we ought to
4739 * But, this leaves one open to an easy denial of
4740 * service attack, and SYN cookies can't defend
4741 * against this problem. So, we drop the data
4742 * in the interest of security over speed.
4752 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
4756 /* Do step6 onward by hand. */
4757 tcp_urg(sk, skb, th);
4759 tcp_data_snd_check(sk);
4763 if (tcp_fast_parse_options(skb, th, tp) && tp->saw_tstamp &&
4764 tcp_paws_discard(tp, skb)) {
4766 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
4767 tcp_send_dupack(sk, skb);
4770 /* Reset is accepted even if it did not pass PAWS. */
4773 /* step 1: check sequence number */
4774 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
4776 tcp_send_dupack(sk, skb);
4780 /* step 2: check RST bit */
4786 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
4788 /* step 3: check security and precedence [ignored] */
4792 * Check for a SYN in window.
4794 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4795 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
4800 /* step 5: check the ACK field */
4802 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH);
4804 switch(sk->sk_state) {
4807 tp->copied_seq = tp->rcv_nxt;
4809 tcp_set_state(sk, TCP_ESTABLISHED);
4810 sk->sk_state_change(sk);
4812 /* Note, that this wakeup is only for marginal
4813 * crossed SYN case. Passively open sockets
4814 * are not waked up, because sk->sk_sleep ==
4815 * NULL and sk->sk_socket == NULL.
4817 if (sk->sk_socket) {
4818 sk_wake_async(sk,0,POLL_OUT);
4821 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
4822 tp->snd_wnd = ntohs(th->window) <<
4824 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq,
4825 TCP_SKB_CB(skb)->seq);
4827 /* tcp_ack considers this ACK as duplicate
4828 * and does not calculate rtt.
4829 * Fix it at least with timestamps.
4831 if (tp->saw_tstamp && tp->rcv_tsecr &&
4833 tcp_ack_saw_tstamp(tp, 0);
4836 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
4838 /* Make sure socket is routed, for
4841 tp->af_specific->rebuild_header(sk);
4843 tcp_init_metrics(sk);
4845 /* Prevent spurious tcp_cwnd_restart() on
4846 * first data packet.
4848 tp->lsndtime = tcp_time_stamp;
4850 tcp_initialize_rcv_mss(sk);
4851 tcp_init_buffer_space(sk);
4852 tcp_fast_path_on(tp);
4859 if (tp->snd_una == tp->write_seq) {
4860 tcp_set_state(sk, TCP_FIN_WAIT2);
4861 sk->sk_shutdown |= SEND_SHUTDOWN;
4862 dst_confirm(sk->sk_dst_cache);
4864 if (!sock_flag(sk, SOCK_DEAD))
4865 /* Wake up lingering close() */
4866 sk->sk_state_change(sk);
4870 if (tp->linger2 < 0 ||
4871 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4872 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
4874 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
4878 tmo = tcp_fin_time(tp);
4879 if (tmo > TCP_TIMEWAIT_LEN) {
4880 tcp_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
4881 } else if (th->fin || sock_owned_by_user(sk)) {
4882 /* Bad case. We could lose such FIN otherwise.
4883 * It is not a big problem, but it looks confusing
4884 * and not so rare event. We still can lose it now,
4885 * if it spins in bh_lock_sock(), but it is really
4888 tcp_reset_keepalive_timer(sk, tmo);
4890 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
4898 if (tp->snd_una == tp->write_seq) {
4899 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4905 if (tp->snd_una == tp->write_seq) {
4906 tcp_update_metrics(sk);
4915 /* step 6: check the URG bit */
4916 tcp_urg(sk, skb, th);
4918 /* step 7: process the segment text */
4919 switch (sk->sk_state) {
4920 case TCP_CLOSE_WAIT:
4923 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4927 /* RFC 793 says to queue data in these states,
4928 * RFC 1122 says we MUST send a reset.
4929 * BSD 4.4 also does reset.
4931 if (sk->sk_shutdown & RCV_SHUTDOWN) {
4932 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4933 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
4934 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
4940 case TCP_ESTABLISHED:
4941 tcp_data_queue(sk, skb);
4946 /* tcp_data could move socket to TIME-WAIT */
4947 if (sk->sk_state != TCP_CLOSE) {
4948 tcp_data_snd_check(sk);
4949 tcp_ack_snd_check(sk);
4959 EXPORT_SYMBOL(sysctl_tcp_ecn);
4960 EXPORT_SYMBOL(sysctl_tcp_reordering);
4961 EXPORT_SYMBOL(tcp_parse_options);
4962 EXPORT_SYMBOL(tcp_rcv_established);
4963 EXPORT_SYMBOL(tcp_rcv_state_process);