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;
105 int sysctl_tcp_bic_beta = 819; /* = 819/1024 (BICTCP_BETA_SCALE) */
107 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
108 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
109 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
110 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
111 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
112 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
113 #define FLAG_ECE 0x40 /* ECE in this ACK */
114 #define FLAG_DATA_LOST 0x80 /* SACK detected data lossage. */
115 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
117 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
118 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
119 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
120 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
122 #define IsReno(tp) ((tp)->rx_opt.sack_ok == 0)
123 #define IsFack(tp) ((tp)->rx_opt.sack_ok & 2)
124 #define IsDSack(tp) ((tp)->rx_opt.sack_ok & 4)
126 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
128 /* Adapt the MSS value used to make delayed ack decision to the
131 static inline void tcp_measure_rcv_mss(struct tcp_sock *tp,
134 unsigned int len, lss;
136 lss = tp->ack.last_seg_size;
137 tp->ack.last_seg_size = 0;
139 /* skb->len may jitter because of SACKs, even if peer
140 * sends good full-sized frames.
143 if (len >= tp->ack.rcv_mss) {
144 tp->ack.rcv_mss = len;
146 /* Otherwise, we make more careful check taking into account,
147 * that SACKs block is variable.
149 * "len" is invariant segment length, including TCP header.
151 len += skb->data - skb->h.raw;
152 if (len >= TCP_MIN_RCVMSS + sizeof(struct tcphdr) ||
153 /* If PSH is not set, packet should be
154 * full sized, provided peer TCP is not badly broken.
155 * This observation (if it is correct 8)) allows
156 * to handle super-low mtu links fairly.
158 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
159 !(tcp_flag_word(skb->h.th)&TCP_REMNANT))) {
160 /* Subtract also invariant (if peer is RFC compliant),
161 * tcp header plus fixed timestamp option length.
162 * Resulting "len" is MSS free of SACK jitter.
164 len -= tp->tcp_header_len;
165 tp->ack.last_seg_size = len;
167 tp->ack.rcv_mss = len;
171 tp->ack.pending |= TCP_ACK_PUSHED;
175 static void tcp_incr_quickack(struct tcp_sock *tp)
177 unsigned quickacks = tp->rcv_wnd/(2*tp->ack.rcv_mss);
181 if (quickacks > tp->ack.quick)
182 tp->ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
185 void tcp_enter_quickack_mode(struct tcp_sock *tp)
187 tcp_incr_quickack(tp);
188 tp->ack.pingpong = 0;
189 tp->ack.ato = TCP_ATO_MIN;
192 /* Send ACKs quickly, if "quick" count is not exhausted
193 * and the session is not interactive.
196 static __inline__ int tcp_in_quickack_mode(struct tcp_sock *tp)
198 return (tp->ack.quick && !tp->ack.pingpong);
201 /* Buffer size and advertised window tuning.
203 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
206 static void tcp_fixup_sndbuf(struct sock *sk)
208 int sndmem = tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER + 16 +
209 sizeof(struct sk_buff);
211 if (sk->sk_sndbuf < 3 * sndmem)
212 sk->sk_sndbuf = min(3 * sndmem, sysctl_tcp_wmem[2]);
215 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
217 * All tcp_full_space() is split to two parts: "network" buffer, allocated
218 * forward and advertised in receiver window (tp->rcv_wnd) and
219 * "application buffer", required to isolate scheduling/application
220 * latencies from network.
221 * window_clamp is maximal advertised window. It can be less than
222 * tcp_full_space(), in this case tcp_full_space() - window_clamp
223 * is reserved for "application" buffer. The less window_clamp is
224 * the smoother our behaviour from viewpoint of network, but the lower
225 * throughput and the higher sensitivity of the connection to losses. 8)
227 * rcv_ssthresh is more strict window_clamp used at "slow start"
228 * phase to predict further behaviour of this connection.
229 * It is used for two goals:
230 * - to enforce header prediction at sender, even when application
231 * requires some significant "application buffer". It is check #1.
232 * - to prevent pruning of receive queue because of misprediction
233 * of receiver window. Check #2.
235 * The scheme does not work when sender sends good segments opening
236 * window and then starts to feed us spagetti. But it should work
237 * in common situations. Otherwise, we have to rely on queue collapsing.
240 /* Slow part of check#2. */
241 static int __tcp_grow_window(struct sock *sk, struct tcp_sock *tp,
245 int truesize = tcp_win_from_space(skb->truesize)/2;
246 int window = tcp_full_space(sk)/2;
248 while (tp->rcv_ssthresh <= window) {
249 if (truesize <= skb->len)
250 return 2*tp->ack.rcv_mss;
258 static inline void tcp_grow_window(struct sock *sk, struct tcp_sock *tp,
262 if (tp->rcv_ssthresh < tp->window_clamp &&
263 (int)tp->rcv_ssthresh < tcp_space(sk) &&
264 !tcp_memory_pressure) {
267 /* Check #2. Increase window, if skb with such overhead
268 * will fit to rcvbuf in future.
270 if (tcp_win_from_space(skb->truesize) <= skb->len)
273 incr = __tcp_grow_window(sk, tp, skb);
276 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr, tp->window_clamp);
282 /* 3. Tuning rcvbuf, when connection enters established state. */
284 static void tcp_fixup_rcvbuf(struct sock *sk)
286 struct tcp_sock *tp = tcp_sk(sk);
287 int rcvmem = tp->advmss + MAX_TCP_HEADER + 16 + sizeof(struct sk_buff);
289 /* Try to select rcvbuf so that 4 mss-sized segments
290 * will fit to window and correspoding skbs will fit to our rcvbuf.
291 * (was 3; 4 is minimum to allow fast retransmit to work.)
293 while (tcp_win_from_space(rcvmem) < tp->advmss)
295 if (sk->sk_rcvbuf < 4 * rcvmem)
296 sk->sk_rcvbuf = min(4 * rcvmem, sysctl_tcp_rmem[2]);
299 /* 4. Try to fixup all. It is made iimediately after connection enters
302 static void tcp_init_buffer_space(struct sock *sk)
304 struct tcp_sock *tp = tcp_sk(sk);
307 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
308 tcp_fixup_rcvbuf(sk);
309 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
310 tcp_fixup_sndbuf(sk);
312 tp->rcvq_space.space = tp->rcv_wnd;
314 maxwin = tcp_full_space(sk);
316 if (tp->window_clamp >= maxwin) {
317 tp->window_clamp = maxwin;
319 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
320 tp->window_clamp = max(maxwin -
321 (maxwin >> sysctl_tcp_app_win),
325 /* Force reservation of one segment. */
326 if (sysctl_tcp_app_win &&
327 tp->window_clamp > 2 * tp->advmss &&
328 tp->window_clamp + tp->advmss > maxwin)
329 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
331 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
332 tp->snd_cwnd_stamp = tcp_time_stamp;
335 static void init_bictcp(struct tcp_sock *tp)
339 tp->bictcp.last_max_cwnd = 0;
340 tp->bictcp.last_cwnd = 0;
341 tp->bictcp.last_stamp = 0;
344 /* 5. Recalculate window clamp after socket hit its memory bounds. */
345 static void tcp_clamp_window(struct sock *sk, struct tcp_sock *tp)
348 unsigned int app_win = tp->rcv_nxt - tp->copied_seq;
353 skb_queue_walk(&tp->out_of_order_queue, skb) {
357 /* If overcommit is due to out of order segments,
358 * do not clamp window. Try to expand rcvbuf instead.
361 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
362 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
363 !tcp_memory_pressure &&
364 atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0])
365 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
368 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf) {
370 if (atomic_read(&sk->sk_rmem_alloc) >= 2 * sk->sk_rcvbuf)
372 if (app_win > tp->ack.rcv_mss)
373 app_win -= tp->ack.rcv_mss;
374 app_win = max(app_win, 2U*tp->advmss);
377 tp->window_clamp = min(tp->window_clamp, app_win);
378 tp->rcv_ssthresh = min(tp->window_clamp, 2U*tp->advmss);
382 /* Receiver "autotuning" code.
384 * The algorithm for RTT estimation w/o timestamps is based on
385 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
386 * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps>
388 * More detail on this code can be found at
389 * <http://www.psc.edu/~jheffner/senior_thesis.ps>,
390 * though this reference is out of date. A new paper
393 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
395 u32 new_sample = tp->rcv_rtt_est.rtt;
401 if (new_sample != 0) {
402 /* If we sample in larger samples in the non-timestamp
403 * case, we could grossly overestimate the RTT especially
404 * with chatty applications or bulk transfer apps which
405 * are stalled on filesystem I/O.
407 * Also, since we are only going for a minimum in the
408 * non-timestamp case, we do not smoothe things out
409 * else with timestamps disabled convergance takes too
413 m -= (new_sample >> 3);
415 } else if (m < new_sample)
418 /* No previous mesaure. */
422 if (tp->rcv_rtt_est.rtt != new_sample)
423 tp->rcv_rtt_est.rtt = new_sample;
426 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
428 if (tp->rcv_rtt_est.time == 0)
430 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
432 tcp_rcv_rtt_update(tp,
433 jiffies - tp->rcv_rtt_est.time,
437 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
438 tp->rcv_rtt_est.time = tcp_time_stamp;
441 static inline void tcp_rcv_rtt_measure_ts(struct tcp_sock *tp, struct sk_buff *skb)
443 if (tp->rx_opt.rcv_tsecr &&
444 (TCP_SKB_CB(skb)->end_seq -
445 TCP_SKB_CB(skb)->seq >= tp->ack.rcv_mss))
446 tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
450 * This function should be called every time data is copied to user space.
451 * It calculates the appropriate TCP receive buffer space.
453 void tcp_rcv_space_adjust(struct sock *sk)
455 struct tcp_sock *tp = tcp_sk(sk);
459 if (tp->rcvq_space.time == 0)
462 time = tcp_time_stamp - tp->rcvq_space.time;
463 if (time < (tp->rcv_rtt_est.rtt >> 3) ||
464 tp->rcv_rtt_est.rtt == 0)
467 space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
469 space = max(tp->rcvq_space.space, space);
471 if (tp->rcvq_space.space != space) {
474 tp->rcvq_space.space = space;
476 if (sysctl_tcp_moderate_rcvbuf) {
477 int new_clamp = space;
479 /* Receive space grows, normalize in order to
480 * take into account packet headers and sk_buff
481 * structure overhead.
486 rcvmem = (tp->advmss + MAX_TCP_HEADER +
487 16 + sizeof(struct sk_buff));
488 while (tcp_win_from_space(rcvmem) < tp->advmss)
491 space = min(space, sysctl_tcp_rmem[2]);
492 if (space > sk->sk_rcvbuf) {
493 sk->sk_rcvbuf = space;
495 /* Make the window clamp follow along. */
496 tp->window_clamp = new_clamp;
502 tp->rcvq_space.seq = tp->copied_seq;
503 tp->rcvq_space.time = tcp_time_stamp;
506 /* There is something which you must keep in mind when you analyze the
507 * behavior of the tp->ato delayed ack timeout interval. When a
508 * connection starts up, we want to ack as quickly as possible. The
509 * problem is that "good" TCP's do slow start at the beginning of data
510 * transmission. The means that until we send the first few ACK's the
511 * sender will sit on his end and only queue most of his data, because
512 * he can only send snd_cwnd unacked packets at any given time. For
513 * each ACK we send, he increments snd_cwnd and transmits more of his
516 static void tcp_event_data_recv(struct sock *sk, struct tcp_sock *tp, struct sk_buff *skb)
520 tcp_schedule_ack(tp);
522 tcp_measure_rcv_mss(tp, skb);
524 tcp_rcv_rtt_measure(tp);
526 now = tcp_time_stamp;
529 /* The _first_ data packet received, initialize
530 * delayed ACK engine.
532 tcp_incr_quickack(tp);
533 tp->ack.ato = TCP_ATO_MIN;
535 int m = now - tp->ack.lrcvtime;
537 if (m <= TCP_ATO_MIN/2) {
538 /* The fastest case is the first. */
539 tp->ack.ato = (tp->ack.ato>>1) + TCP_ATO_MIN/2;
540 } else if (m < tp->ack.ato) {
541 tp->ack.ato = (tp->ack.ato>>1) + m;
542 if (tp->ack.ato > tp->rto)
543 tp->ack.ato = tp->rto;
544 } else if (m > tp->rto) {
545 /* Too long gap. Apparently sender falled to
546 * restart window, so that we send ACKs quickly.
548 tcp_incr_quickack(tp);
549 sk_stream_mem_reclaim(sk);
552 tp->ack.lrcvtime = now;
554 TCP_ECN_check_ce(tp, skb);
557 tcp_grow_window(sk, tp, skb);
560 /* When starting a new connection, pin down the current choice of
561 * congestion algorithm.
563 void tcp_ca_init(struct tcp_sock *tp)
565 if (sysctl_tcp_westwood)
566 tp->adv_cong = TCP_WESTWOOD;
567 else if (sysctl_tcp_bic)
568 tp->adv_cong = TCP_BIC;
569 else if (sysctl_tcp_vegas_cong_avoid) {
570 tp->adv_cong = TCP_VEGAS;
571 tp->vegas.baseRTT = 0x7fffffff;
572 tcp_vegas_enable(tp);
576 /* Do RTT sampling needed for Vegas.
578 * o min-filter RTT samples from within an RTT to get the current
579 * propagation delay + queuing delay (we are min-filtering to try to
580 * avoid the effects of delayed ACKs)
581 * o min-filter RTT samples from a much longer window (forever for now)
582 * to find the propagation delay (baseRTT)
584 static inline void vegas_rtt_calc(struct tcp_sock *tp, __u32 rtt)
586 __u32 vrtt = rtt + 1; /* Never allow zero rtt or baseRTT */
588 /* Filter to find propagation delay: */
589 if (vrtt < tp->vegas.baseRTT)
590 tp->vegas.baseRTT = vrtt;
592 /* Find the min RTT during the last RTT to find
593 * the current prop. delay + queuing delay:
595 tp->vegas.minRTT = min(tp->vegas.minRTT, vrtt);
599 /* Called to compute a smoothed rtt estimate. The data fed to this
600 * routine either comes from timestamps, or from segments that were
601 * known _not_ to have been retransmitted [see Karn/Partridge
602 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
603 * piece by Van Jacobson.
604 * NOTE: the next three routines used to be one big routine.
605 * To save cycles in the RFC 1323 implementation it was better to break
606 * it up into three procedures. -- erics
608 static void tcp_rtt_estimator(struct tcp_sock *tp, __u32 mrtt)
610 long m = mrtt; /* RTT */
612 if (tcp_vegas_enabled(tp))
613 vegas_rtt_calc(tp, mrtt);
615 /* The following amusing code comes from Jacobson's
616 * article in SIGCOMM '88. Note that rtt and mdev
617 * are scaled versions of rtt and mean deviation.
618 * This is designed to be as fast as possible
619 * m stands for "measurement".
621 * On a 1990 paper the rto value is changed to:
622 * RTO = rtt + 4 * mdev
624 * Funny. This algorithm seems to be very broken.
625 * These formulae increase RTO, when it should be decreased, increase
626 * too slowly, when it should be incresed fastly, decrease too fastly
627 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
628 * does not matter how to _calculate_ it. Seems, it was trap
629 * that VJ failed to avoid. 8)
634 m -= (tp->srtt >> 3); /* m is now error in rtt est */
635 tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */
637 m = -m; /* m is now abs(error) */
638 m -= (tp->mdev >> 2); /* similar update on mdev */
639 /* This is similar to one of Eifel findings.
640 * Eifel blocks mdev updates when rtt decreases.
641 * This solution is a bit different: we use finer gain
642 * for mdev in this case (alpha*beta).
643 * Like Eifel it also prevents growth of rto,
644 * but also it limits too fast rto decreases,
645 * happening in pure Eifel.
650 m -= (tp->mdev >> 2); /* similar update on mdev */
652 tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */
653 if (tp->mdev > tp->mdev_max) {
654 tp->mdev_max = tp->mdev;
655 if (tp->mdev_max > tp->rttvar)
656 tp->rttvar = tp->mdev_max;
658 if (after(tp->snd_una, tp->rtt_seq)) {
659 if (tp->mdev_max < tp->rttvar)
660 tp->rttvar -= (tp->rttvar-tp->mdev_max)>>2;
661 tp->rtt_seq = tp->snd_nxt;
662 tp->mdev_max = TCP_RTO_MIN;
665 /* no previous measure. */
666 tp->srtt = m<<3; /* take the measured time to be rtt */
667 tp->mdev = m<<1; /* make sure rto = 3*rtt */
668 tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN);
669 tp->rtt_seq = tp->snd_nxt;
672 tcp_westwood_update_rtt(tp, tp->srtt >> 3);
675 /* Calculate rto without backoff. This is the second half of Van Jacobson's
676 * routine referred to above.
678 static inline void tcp_set_rto(struct tcp_sock *tp)
680 /* Old crap is replaced with new one. 8)
683 * 1. If rtt variance happened to be less 50msec, it is hallucination.
684 * It cannot be less due to utterly erratic ACK generation made
685 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
686 * to do with delayed acks, because at cwnd>2 true delack timeout
687 * is invisible. Actually, Linux-2.4 also generates erratic
688 * ACKs in some curcumstances.
690 tp->rto = (tp->srtt >> 3) + tp->rttvar;
692 /* 2. Fixups made earlier cannot be right.
693 * If we do not estimate RTO correctly without them,
694 * all the algo is pure shit and should be replaced
695 * with correct one. It is exaclty, which we pretend to do.
699 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
700 * guarantees that rto is higher.
702 static inline void tcp_bound_rto(struct tcp_sock *tp)
704 if (tp->rto > TCP_RTO_MAX)
705 tp->rto = TCP_RTO_MAX;
708 /* Save metrics learned by this TCP session.
709 This function is called only, when TCP finishes successfully
710 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
712 void tcp_update_metrics(struct sock *sk)
714 struct tcp_sock *tp = tcp_sk(sk);
715 struct dst_entry *dst = __sk_dst_get(sk);
717 if (sysctl_tcp_nometrics_save)
722 if (dst && (dst->flags&DST_HOST)) {
725 if (tp->backoff || !tp->srtt) {
726 /* This session failed to estimate rtt. Why?
727 * Probably, no packets returned in time.
730 if (!(dst_metric_locked(dst, RTAX_RTT)))
731 dst->metrics[RTAX_RTT-1] = 0;
735 m = dst_metric(dst, RTAX_RTT) - tp->srtt;
737 /* If newly calculated rtt larger than stored one,
738 * store new one. Otherwise, use EWMA. Remember,
739 * rtt overestimation is always better than underestimation.
741 if (!(dst_metric_locked(dst, RTAX_RTT))) {
743 dst->metrics[RTAX_RTT-1] = tp->srtt;
745 dst->metrics[RTAX_RTT-1] -= (m>>3);
748 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) {
752 /* Scale deviation to rttvar fixed point */
757 if (m >= dst_metric(dst, RTAX_RTTVAR))
758 dst->metrics[RTAX_RTTVAR-1] = m;
760 dst->metrics[RTAX_RTTVAR-1] -=
761 (dst->metrics[RTAX_RTTVAR-1] - m)>>2;
764 if (tp->snd_ssthresh >= 0xFFFF) {
765 /* Slow start still did not finish. */
766 if (dst_metric(dst, RTAX_SSTHRESH) &&
767 !dst_metric_locked(dst, RTAX_SSTHRESH) &&
768 (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH))
769 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1;
770 if (!dst_metric_locked(dst, RTAX_CWND) &&
771 tp->snd_cwnd > dst_metric(dst, RTAX_CWND))
772 dst->metrics[RTAX_CWND-1] = tp->snd_cwnd;
773 } else if (tp->snd_cwnd > tp->snd_ssthresh &&
774 tp->ca_state == TCP_CA_Open) {
775 /* Cong. avoidance phase, cwnd is reliable. */
776 if (!dst_metric_locked(dst, RTAX_SSTHRESH))
777 dst->metrics[RTAX_SSTHRESH-1] =
778 max(tp->snd_cwnd >> 1, tp->snd_ssthresh);
779 if (!dst_metric_locked(dst, RTAX_CWND))
780 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_cwnd) >> 1;
782 /* Else slow start did not finish, cwnd is non-sense,
783 ssthresh may be also invalid.
785 if (!dst_metric_locked(dst, RTAX_CWND))
786 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_ssthresh) >> 1;
787 if (dst->metrics[RTAX_SSTHRESH-1] &&
788 !dst_metric_locked(dst, RTAX_SSTHRESH) &&
789 tp->snd_ssthresh > dst->metrics[RTAX_SSTHRESH-1])
790 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh;
793 if (!dst_metric_locked(dst, RTAX_REORDERING)) {
794 if (dst->metrics[RTAX_REORDERING-1] < tp->reordering &&
795 tp->reordering != sysctl_tcp_reordering)
796 dst->metrics[RTAX_REORDERING-1] = tp->reordering;
801 /* Numbers are taken from RFC2414. */
802 __u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst)
804 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
807 if (tp->mss_cache_std > 1460)
810 cwnd = (tp->mss_cache_std > 1095) ? 3 : 4;
812 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
815 /* Initialize metrics on socket. */
817 static void tcp_init_metrics(struct sock *sk)
819 struct tcp_sock *tp = tcp_sk(sk);
820 struct dst_entry *dst = __sk_dst_get(sk);
827 if (dst_metric_locked(dst, RTAX_CWND))
828 tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND);
829 if (dst_metric(dst, RTAX_SSTHRESH)) {
830 tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH);
831 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
832 tp->snd_ssthresh = tp->snd_cwnd_clamp;
834 if (dst_metric(dst, RTAX_REORDERING) &&
835 tp->reordering != dst_metric(dst, RTAX_REORDERING)) {
836 tp->rx_opt.sack_ok &= ~2;
837 tp->reordering = dst_metric(dst, RTAX_REORDERING);
840 if (dst_metric(dst, RTAX_RTT) == 0)
843 if (!tp->srtt && dst_metric(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3))
846 /* Initial rtt is determined from SYN,SYN-ACK.
847 * The segment is small and rtt may appear much
848 * less than real one. Use per-dst memory
849 * to make it more realistic.
851 * A bit of theory. RTT is time passed after "normal" sized packet
852 * is sent until it is ACKed. In normal curcumstances sending small
853 * packets force peer to delay ACKs and calculation is correct too.
854 * The algorithm is adaptive and, provided we follow specs, it
855 * NEVER underestimate RTT. BUT! If peer tries to make some clever
856 * tricks sort of "quick acks" for time long enough to decrease RTT
857 * to low value, and then abruptly stops to do it and starts to delay
858 * ACKs, wait for troubles.
860 if (dst_metric(dst, RTAX_RTT) > tp->srtt) {
861 tp->srtt = dst_metric(dst, RTAX_RTT);
862 tp->rtt_seq = tp->snd_nxt;
864 if (dst_metric(dst, RTAX_RTTVAR) > tp->mdev) {
865 tp->mdev = dst_metric(dst, RTAX_RTTVAR);
866 tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN);
870 if (tp->rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp)
872 tp->snd_cwnd = tcp_init_cwnd(tp, dst);
873 tp->snd_cwnd_stamp = tcp_time_stamp;
877 /* Play conservative. If timestamps are not
878 * supported, TCP will fail to recalculate correct
879 * rtt, if initial rto is too small. FORGET ALL AND RESET!
881 if (!tp->rx_opt.saw_tstamp && tp->srtt) {
883 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT;
884 tp->rto = TCP_TIMEOUT_INIT;
888 static void tcp_update_reordering(struct tcp_sock *tp, int metric, int ts)
890 if (metric > tp->reordering) {
891 tp->reordering = min(TCP_MAX_REORDERING, metric);
893 /* This exciting event is worth to be remembered. 8) */
895 NET_INC_STATS_BH(LINUX_MIB_TCPTSREORDER);
897 NET_INC_STATS_BH(LINUX_MIB_TCPRENOREORDER);
899 NET_INC_STATS_BH(LINUX_MIB_TCPFACKREORDER);
901 NET_INC_STATS_BH(LINUX_MIB_TCPSACKREORDER);
902 #if FASTRETRANS_DEBUG > 1
903 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n",
904 tp->rx_opt.sack_ok, tp->ca_state,
908 tp->undo_marker ? tp->undo_retrans : 0);
910 /* Disable FACK yet. */
911 tp->rx_opt.sack_ok &= ~2;
915 /* This procedure tags the retransmission queue when SACKs arrive.
917 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
918 * Packets in queue with these bits set are counted in variables
919 * sacked_out, retrans_out and lost_out, correspondingly.
921 * Valid combinations are:
922 * Tag InFlight Description
923 * 0 1 - orig segment is in flight.
924 * S 0 - nothing flies, orig reached receiver.
925 * L 0 - nothing flies, orig lost by net.
926 * R 2 - both orig and retransmit are in flight.
927 * L|R 1 - orig is lost, retransmit is in flight.
928 * S|R 1 - orig reached receiver, retrans is still in flight.
929 * (L|S|R is logically valid, it could occur when L|R is sacked,
930 * but it is equivalent to plain S and code short-curcuits it to S.
931 * L|S is logically invalid, it would mean -1 packet in flight 8))
933 * These 6 states form finite state machine, controlled by the following events:
934 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
935 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
936 * 3. Loss detection event of one of three flavors:
937 * A. Scoreboard estimator decided the packet is lost.
938 * A'. Reno "three dupacks" marks head of queue lost.
939 * A''. Its FACK modfication, head until snd.fack is lost.
940 * B. SACK arrives sacking data transmitted after never retransmitted
942 * C. SACK arrives sacking SND.NXT at the moment, when the
943 * segment was retransmitted.
944 * 4. D-SACK added new rule: D-SACK changes any tag to S.
946 * It is pleasant to note, that state diagram turns out to be commutative,
947 * so that we are allowed not to be bothered by order of our actions,
948 * when multiple events arrive simultaneously. (see the function below).
950 * Reordering detection.
951 * --------------------
952 * Reordering metric is maximal distance, which a packet can be displaced
953 * in packet stream. With SACKs we can estimate it:
955 * 1. SACK fills old hole and the corresponding segment was not
956 * ever retransmitted -> reordering. Alas, we cannot use it
957 * when segment was retransmitted.
958 * 2. The last flaw is solved with D-SACK. D-SACK arrives
959 * for retransmitted and already SACKed segment -> reordering..
960 * Both of these heuristics are not used in Loss state, when we cannot
961 * account for retransmits accurately.
964 tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, u32 prior_snd_una)
966 struct tcp_sock *tp = tcp_sk(sk);
967 unsigned char *ptr = ack_skb->h.raw + TCP_SKB_CB(ack_skb)->sacked;
968 struct tcp_sack_block *sp = (struct tcp_sack_block *)(ptr+2);
969 int num_sacks = (ptr[1] - TCPOLEN_SACK_BASE)>>3;
970 int reord = tp->packets_out;
972 u32 lost_retrans = 0;
976 /* So, SACKs for already sent large segments will be lost.
977 * Not good, but alternative is to resegment the queue. */
978 if (sk->sk_route_caps & NETIF_F_TSO) {
979 sk->sk_route_caps &= ~NETIF_F_TSO;
980 sk->sk_no_largesend = 1;
981 tp->mss_cache = tp->mss_cache_std;
986 prior_fackets = tp->fackets_out;
988 for (i=0; i<num_sacks; i++, sp++) {
990 __u32 start_seq = ntohl(sp->start_seq);
991 __u32 end_seq = ntohl(sp->end_seq);
995 /* Check for D-SACK. */
997 u32 ack = TCP_SKB_CB(ack_skb)->ack_seq;
999 if (before(start_seq, ack)) {
1001 tp->rx_opt.sack_ok |= 4;
1002 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV);
1003 } else if (num_sacks > 1 &&
1004 !after(end_seq, ntohl(sp[1].end_seq)) &&
1005 !before(start_seq, ntohl(sp[1].start_seq))) {
1007 tp->rx_opt.sack_ok |= 4;
1008 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV);
1011 /* D-SACK for already forgotten data...
1012 * Do dumb counting. */
1014 !after(end_seq, prior_snd_una) &&
1015 after(end_seq, tp->undo_marker))
1018 /* Eliminate too old ACKs, but take into
1019 * account more or less fresh ones, they can
1020 * contain valid SACK info.
1022 if (before(ack, prior_snd_una - tp->max_window))
1026 /* Event "B" in the comment above. */
1027 if (after(end_seq, tp->high_seq))
1028 flag |= FLAG_DATA_LOST;
1030 sk_stream_for_retrans_queue(skb, sk) {
1031 u8 sacked = TCP_SKB_CB(skb)->sacked;
1034 /* The retransmission queue is always in order, so
1035 * we can short-circuit the walk early.
1037 if(!before(TCP_SKB_CB(skb)->seq, end_seq))
1040 fack_count += tcp_skb_pcount(skb);
1042 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1043 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1045 /* Account D-SACK for retransmitted packet. */
1046 if ((dup_sack && in_sack) &&
1047 (sacked & TCPCB_RETRANS) &&
1048 after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker))
1051 /* The frame is ACKed. */
1052 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) {
1053 if (sacked&TCPCB_RETRANS) {
1054 if ((dup_sack && in_sack) &&
1055 (sacked&TCPCB_SACKED_ACKED))
1056 reord = min(fack_count, reord);
1058 /* If it was in a hole, we detected reordering. */
1059 if (fack_count < prior_fackets &&
1060 !(sacked&TCPCB_SACKED_ACKED))
1061 reord = min(fack_count, reord);
1064 /* Nothing to do; acked frame is about to be dropped. */
1068 if ((sacked&TCPCB_SACKED_RETRANS) &&
1069 after(end_seq, TCP_SKB_CB(skb)->ack_seq) &&
1070 (!lost_retrans || after(end_seq, lost_retrans)))
1071 lost_retrans = end_seq;
1076 if (!(sacked&TCPCB_SACKED_ACKED)) {
1077 if (sacked & TCPCB_SACKED_RETRANS) {
1078 /* If the segment is not tagged as lost,
1079 * we do not clear RETRANS, believing
1080 * that retransmission is still in flight.
1082 if (sacked & TCPCB_LOST) {
1083 TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1084 tp->lost_out -= tcp_skb_pcount(skb);
1085 tp->retrans_out -= tcp_skb_pcount(skb);
1088 /* New sack for not retransmitted frame,
1089 * which was in hole. It is reordering.
1091 if (!(sacked & TCPCB_RETRANS) &&
1092 fack_count < prior_fackets)
1093 reord = min(fack_count, reord);
1095 if (sacked & TCPCB_LOST) {
1096 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1097 tp->lost_out -= tcp_skb_pcount(skb);
1101 TCP_SKB_CB(skb)->sacked |= TCPCB_SACKED_ACKED;
1102 flag |= FLAG_DATA_SACKED;
1103 tp->sacked_out += tcp_skb_pcount(skb);
1105 if (fack_count > tp->fackets_out)
1106 tp->fackets_out = fack_count;
1108 if (dup_sack && (sacked&TCPCB_RETRANS))
1109 reord = min(fack_count, reord);
1112 /* D-SACK. We can detect redundant retransmission
1113 * in S|R and plain R frames and clear it.
1114 * undo_retrans is decreased above, L|R frames
1115 * are accounted above as well.
1118 (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS)) {
1119 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1120 tp->retrans_out -= tcp_skb_pcount(skb);
1125 /* Check for lost retransmit. This superb idea is
1126 * borrowed from "ratehalving". Event "C".
1127 * Later note: FACK people cheated me again 8),
1128 * we have to account for reordering! Ugly,
1131 if (lost_retrans && tp->ca_state == TCP_CA_Recovery) {
1132 struct sk_buff *skb;
1134 sk_stream_for_retrans_queue(skb, sk) {
1135 if (after(TCP_SKB_CB(skb)->seq, lost_retrans))
1137 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1139 if ((TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) &&
1140 after(lost_retrans, TCP_SKB_CB(skb)->ack_seq) &&
1142 !before(lost_retrans,
1143 TCP_SKB_CB(skb)->ack_seq + tp->reordering *
1144 tp->mss_cache_std))) {
1145 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1146 tp->retrans_out -= tcp_skb_pcount(skb);
1148 if (!(TCP_SKB_CB(skb)->sacked&(TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1149 tp->lost_out += tcp_skb_pcount(skb);
1150 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1151 flag |= FLAG_DATA_SACKED;
1152 NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT);
1158 tp->left_out = tp->sacked_out + tp->lost_out;
1160 if ((reord < tp->fackets_out) && tp->ca_state != TCP_CA_Loss)
1161 tcp_update_reordering(tp, ((tp->fackets_out + 1) - reord), 0);
1163 #if FASTRETRANS_DEBUG > 0
1164 BUG_TRAP((int)tp->sacked_out >= 0);
1165 BUG_TRAP((int)tp->lost_out >= 0);
1166 BUG_TRAP((int)tp->retrans_out >= 0);
1167 BUG_TRAP((int)tcp_packets_in_flight(tp) >= 0);
1172 /* RTO occurred, but do not yet enter loss state. Instead, transmit two new
1173 * segments to see from the next ACKs whether any data was really missing.
1174 * If the RTO was spurious, new ACKs should arrive.
1176 void tcp_enter_frto(struct sock *sk)
1178 struct tcp_sock *tp = tcp_sk(sk);
1179 struct sk_buff *skb;
1181 tp->frto_counter = 1;
1183 if (tp->ca_state <= TCP_CA_Disorder ||
1184 tp->snd_una == tp->high_seq ||
1185 (tp->ca_state == TCP_CA_Loss && !tp->retransmits)) {
1186 tp->prior_ssthresh = tcp_current_ssthresh(tp);
1187 if (!tcp_westwood_ssthresh(tp))
1188 tp->snd_ssthresh = tcp_recalc_ssthresh(tp);
1191 /* Have to clear retransmission markers here to keep the bookkeeping
1192 * in shape, even though we are not yet in Loss state.
1193 * If something was really lost, it is eventually caught up
1194 * in tcp_enter_frto_loss.
1196 tp->retrans_out = 0;
1197 tp->undo_marker = tp->snd_una;
1198 tp->undo_retrans = 0;
1200 sk_stream_for_retrans_queue(skb, sk) {
1201 TCP_SKB_CB(skb)->sacked &= ~TCPCB_RETRANS;
1203 tcp_sync_left_out(tp);
1205 tcp_set_ca_state(tp, TCP_CA_Open);
1206 tp->frto_highmark = tp->snd_nxt;
1209 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1210 * which indicates that we should follow the traditional RTO recovery,
1211 * i.e. mark everything lost and do go-back-N retransmission.
1213 static void tcp_enter_frto_loss(struct sock *sk)
1215 struct tcp_sock *tp = tcp_sk(sk);
1216 struct sk_buff *skb;
1221 tp->fackets_out = 0;
1223 sk_stream_for_retrans_queue(skb, sk) {
1224 cnt += tcp_skb_pcount(skb);
1225 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1226 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) {
1228 /* Do not mark those segments lost that were
1229 * forward transmitted after RTO
1231 if (!after(TCP_SKB_CB(skb)->end_seq,
1232 tp->frto_highmark)) {
1233 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1234 tp->lost_out += tcp_skb_pcount(skb);
1237 tp->sacked_out += tcp_skb_pcount(skb);
1238 tp->fackets_out = cnt;
1241 tcp_sync_left_out(tp);
1243 tp->snd_cwnd = tp->frto_counter + tcp_packets_in_flight(tp)+1;
1244 tp->snd_cwnd_cnt = 0;
1245 tp->snd_cwnd_stamp = tcp_time_stamp;
1246 tp->undo_marker = 0;
1247 tp->frto_counter = 0;
1249 tp->reordering = min_t(unsigned int, tp->reordering,
1250 sysctl_tcp_reordering);
1251 tcp_set_ca_state(tp, TCP_CA_Loss);
1252 tp->high_seq = tp->frto_highmark;
1253 TCP_ECN_queue_cwr(tp);
1258 void tcp_clear_retrans(struct tcp_sock *tp)
1261 tp->retrans_out = 0;
1263 tp->fackets_out = 0;
1267 tp->undo_marker = 0;
1268 tp->undo_retrans = 0;
1271 /* Enter Loss state. If "how" is not zero, forget all SACK information
1272 * and reset tags completely, otherwise preserve SACKs. If receiver
1273 * dropped its ofo queue, we will know this due to reneging detection.
1275 void tcp_enter_loss(struct sock *sk, int how)
1277 struct tcp_sock *tp = tcp_sk(sk);
1278 struct sk_buff *skb;
1281 /* Reduce ssthresh if it has not yet been made inside this window. */
1282 if (tp->ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
1283 (tp->ca_state == TCP_CA_Loss && !tp->retransmits)) {
1284 tp->prior_ssthresh = tcp_current_ssthresh(tp);
1285 tp->snd_ssthresh = tcp_recalc_ssthresh(tp);
1288 tp->snd_cwnd_cnt = 0;
1289 tp->snd_cwnd_stamp = tcp_time_stamp;
1291 tcp_clear_retrans(tp);
1293 /* Push undo marker, if it was plain RTO and nothing
1294 * was retransmitted. */
1296 tp->undo_marker = tp->snd_una;
1298 sk_stream_for_retrans_queue(skb, sk) {
1299 cnt += tcp_skb_pcount(skb);
1300 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS)
1301 tp->undo_marker = 0;
1302 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1303 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
1304 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1305 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1306 tp->lost_out += tcp_skb_pcount(skb);
1308 tp->sacked_out += tcp_skb_pcount(skb);
1309 tp->fackets_out = cnt;
1312 tcp_sync_left_out(tp);
1314 tp->reordering = min_t(unsigned int, tp->reordering,
1315 sysctl_tcp_reordering);
1316 tcp_set_ca_state(tp, TCP_CA_Loss);
1317 tp->high_seq = tp->snd_nxt;
1318 TCP_ECN_queue_cwr(tp);
1321 static int tcp_check_sack_reneging(struct sock *sk, struct tcp_sock *tp)
1323 struct sk_buff *skb;
1325 /* If ACK arrived pointing to a remembered SACK,
1326 * it means that our remembered SACKs do not reflect
1327 * real state of receiver i.e.
1328 * receiver _host_ is heavily congested (or buggy).
1329 * Do processing similar to RTO timeout.
1331 if ((skb = skb_peek(&sk->sk_write_queue)) != NULL &&
1332 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
1333 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING);
1335 tcp_enter_loss(sk, 1);
1337 tcp_retransmit_skb(sk, skb_peek(&sk->sk_write_queue));
1338 tcp_reset_xmit_timer(sk, TCP_TIME_RETRANS, tp->rto);
1344 static inline int tcp_fackets_out(struct tcp_sock *tp)
1346 return IsReno(tp) ? tp->sacked_out+1 : tp->fackets_out;
1349 static inline int tcp_skb_timedout(struct tcp_sock *tp, struct sk_buff *skb)
1351 return (tcp_time_stamp - TCP_SKB_CB(skb)->when > tp->rto);
1354 static inline int tcp_head_timedout(struct sock *sk, struct tcp_sock *tp)
1356 return tp->packets_out &&
1357 tcp_skb_timedout(tp, skb_peek(&sk->sk_write_queue));
1360 /* Linux NewReno/SACK/FACK/ECN state machine.
1361 * --------------------------------------
1363 * "Open" Normal state, no dubious events, fast path.
1364 * "Disorder" In all the respects it is "Open",
1365 * but requires a bit more attention. It is entered when
1366 * we see some SACKs or dupacks. It is split of "Open"
1367 * mainly to move some processing from fast path to slow one.
1368 * "CWR" CWND was reduced due to some Congestion Notification event.
1369 * It can be ECN, ICMP source quench, local device congestion.
1370 * "Recovery" CWND was reduced, we are fast-retransmitting.
1371 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
1373 * tcp_fastretrans_alert() is entered:
1374 * - each incoming ACK, if state is not "Open"
1375 * - when arrived ACK is unusual, namely:
1380 * Counting packets in flight is pretty simple.
1382 * in_flight = packets_out - left_out + retrans_out
1384 * packets_out is SND.NXT-SND.UNA counted in packets.
1386 * retrans_out is number of retransmitted segments.
1388 * left_out is number of segments left network, but not ACKed yet.
1390 * left_out = sacked_out + lost_out
1392 * sacked_out: Packets, which arrived to receiver out of order
1393 * and hence not ACKed. With SACKs this number is simply
1394 * amount of SACKed data. Even without SACKs
1395 * it is easy to give pretty reliable estimate of this number,
1396 * counting duplicate ACKs.
1398 * lost_out: Packets lost by network. TCP has no explicit
1399 * "loss notification" feedback from network (for now).
1400 * It means that this number can be only _guessed_.
1401 * Actually, it is the heuristics to predict lossage that
1402 * distinguishes different algorithms.
1404 * F.e. after RTO, when all the queue is considered as lost,
1405 * lost_out = packets_out and in_flight = retrans_out.
1407 * Essentially, we have now two algorithms counting
1410 * FACK: It is the simplest heuristics. As soon as we decided
1411 * that something is lost, we decide that _all_ not SACKed
1412 * packets until the most forward SACK are lost. I.e.
1413 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
1414 * It is absolutely correct estimate, if network does not reorder
1415 * packets. And it loses any connection to reality when reordering
1416 * takes place. We use FACK by default until reordering
1417 * is suspected on the path to this destination.
1419 * NewReno: when Recovery is entered, we assume that one segment
1420 * is lost (classic Reno). While we are in Recovery and
1421 * a partial ACK arrives, we assume that one more packet
1422 * is lost (NewReno). This heuristics are the same in NewReno
1425 * Imagine, that's all! Forget about all this shamanism about CWND inflation
1426 * deflation etc. CWND is real congestion window, never inflated, changes
1427 * only according to classic VJ rules.
1429 * Really tricky (and requiring careful tuning) part of algorithm
1430 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
1431 * The first determines the moment _when_ we should reduce CWND and,
1432 * hence, slow down forward transmission. In fact, it determines the moment
1433 * when we decide that hole is caused by loss, rather than by a reorder.
1435 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
1436 * holes, caused by lost packets.
1438 * And the most logically complicated part of algorithm is undo
1439 * heuristics. We detect false retransmits due to both too early
1440 * fast retransmit (reordering) and underestimated RTO, analyzing
1441 * timestamps and D-SACKs. When we detect that some segments were
1442 * retransmitted by mistake and CWND reduction was wrong, we undo
1443 * window reduction and abort recovery phase. This logic is hidden
1444 * inside several functions named tcp_try_undo_<something>.
1447 /* This function decides, when we should leave Disordered state
1448 * and enter Recovery phase, reducing congestion window.
1450 * Main question: may we further continue forward transmission
1451 * with the same cwnd?
1453 static int tcp_time_to_recover(struct sock *sk, struct tcp_sock *tp)
1457 /* Trick#1: The loss is proven. */
1461 /* Not-A-Trick#2 : Classic rule... */
1462 if (tcp_fackets_out(tp) > tp->reordering)
1465 /* Trick#3 : when we use RFC2988 timer restart, fast
1466 * retransmit can be triggered by timeout of queue head.
1468 if (tcp_head_timedout(sk, tp))
1471 /* Trick#4: It is still not OK... But will it be useful to delay
1474 packets_out = tp->packets_out;
1475 if (packets_out <= tp->reordering &&
1476 tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
1477 !tcp_may_send_now(sk, tp)) {
1478 /* We have nothing to send. This connection is limited
1479 * either by receiver window or by application.
1487 /* If we receive more dupacks than we expected counting segments
1488 * in assumption of absent reordering, interpret this as reordering.
1489 * The only another reason could be bug in receiver TCP.
1491 static void tcp_check_reno_reordering(struct tcp_sock *tp, int addend)
1495 holes = max(tp->lost_out, 1U);
1496 holes = min(holes, tp->packets_out);
1498 if ((tp->sacked_out + holes) > tp->packets_out) {
1499 tp->sacked_out = tp->packets_out - holes;
1500 tcp_update_reordering(tp, tp->packets_out+addend, 0);
1504 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1506 static void tcp_add_reno_sack(struct tcp_sock *tp)
1509 tcp_check_reno_reordering(tp, 0);
1510 tcp_sync_left_out(tp);
1513 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1515 static void tcp_remove_reno_sacks(struct sock *sk, struct tcp_sock *tp, int acked)
1518 /* One ACK acked hole. The rest eat duplicate ACKs. */
1519 if (acked-1 >= tp->sacked_out)
1522 tp->sacked_out -= acked-1;
1524 tcp_check_reno_reordering(tp, acked);
1525 tcp_sync_left_out(tp);
1528 static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1531 tp->left_out = tp->lost_out;
1534 /* Mark head of queue up as lost. */
1535 static void tcp_mark_head_lost(struct sock *sk, struct tcp_sock *tp,
1536 int packets, u32 high_seq)
1538 struct sk_buff *skb;
1541 BUG_TRAP(cnt <= tp->packets_out);
1543 sk_stream_for_retrans_queue(skb, sk) {
1544 cnt -= tcp_skb_pcount(skb);
1545 if (cnt < 0 || after(TCP_SKB_CB(skb)->end_seq, high_seq))
1547 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
1548 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1549 tp->lost_out += tcp_skb_pcount(skb);
1552 tcp_sync_left_out(tp);
1555 /* Account newly detected lost packet(s) */
1557 static void tcp_update_scoreboard(struct sock *sk, struct tcp_sock *tp)
1560 int lost = tp->fackets_out - tp->reordering;
1563 tcp_mark_head_lost(sk, tp, lost, tp->high_seq);
1565 tcp_mark_head_lost(sk, tp, 1, tp->high_seq);
1568 /* New heuristics: it is possible only after we switched
1569 * to restart timer each time when something is ACKed.
1570 * Hence, we can detect timed out packets during fast
1571 * retransmit without falling to slow start.
1573 if (tcp_head_timedout(sk, tp)) {
1574 struct sk_buff *skb;
1576 sk_stream_for_retrans_queue(skb, sk) {
1577 if (tcp_skb_timedout(tp, skb) &&
1578 !(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
1579 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1580 tp->lost_out += tcp_skb_pcount(skb);
1583 tcp_sync_left_out(tp);
1587 /* CWND moderation, preventing bursts due to too big ACKs
1588 * in dubious situations.
1590 static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
1592 tp->snd_cwnd = min(tp->snd_cwnd,
1593 tcp_packets_in_flight(tp)+tcp_max_burst(tp));
1594 tp->snd_cwnd_stamp = tcp_time_stamp;
1597 /* Decrease cwnd each second ack. */
1599 static void tcp_cwnd_down(struct tcp_sock *tp)
1601 int decr = tp->snd_cwnd_cnt + 1;
1606 * Here limit is evaluated as BWestimation*RTTmin (for obtaining it
1607 * in packets we use mss_cache). If sysctl_tcp_westwood is off
1608 * tcp_westwood_bw_rttmin() returns 0. In such case snd_ssthresh is
1609 * still used as usual. It prevents other strange cases in which
1610 * BWE*RTTmin could assume value 0. It should not happen but...
1613 if (!(limit = tcp_westwood_bw_rttmin(tp)))
1614 limit = tp->snd_ssthresh/2;
1616 tp->snd_cwnd_cnt = decr&1;
1619 if (decr && tp->snd_cwnd > limit)
1620 tp->snd_cwnd -= decr;
1622 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp)+1);
1623 tp->snd_cwnd_stamp = tcp_time_stamp;
1626 /* Nothing was retransmitted or returned timestamp is less
1627 * than timestamp of the first retransmission.
1629 static inline int tcp_packet_delayed(struct tcp_sock *tp)
1631 return !tp->retrans_stamp ||
1632 (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
1633 (__s32)(tp->rx_opt.rcv_tsecr - tp->retrans_stamp) < 0);
1636 /* Undo procedures. */
1638 #if FASTRETRANS_DEBUG > 1
1639 static void DBGUNDO(struct sock *sk, struct tcp_sock *tp, const char *msg)
1641 struct inet_sock *inet = inet_sk(sk);
1642 printk(KERN_DEBUG "Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n",
1644 NIPQUAD(inet->daddr), ntohs(inet->dport),
1645 tp->snd_cwnd, tp->left_out,
1646 tp->snd_ssthresh, tp->prior_ssthresh,
1650 #define DBGUNDO(x...) do { } while (0)
1653 static void tcp_undo_cwr(struct tcp_sock *tp, int undo)
1655 if (tp->prior_ssthresh) {
1657 tp->snd_cwnd = max(tp->snd_cwnd, tp->bictcp.last_max_cwnd);
1659 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh<<1);
1661 if (undo && tp->prior_ssthresh > tp->snd_ssthresh) {
1662 tp->snd_ssthresh = tp->prior_ssthresh;
1663 TCP_ECN_withdraw_cwr(tp);
1666 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
1668 tcp_moderate_cwnd(tp);
1669 tp->snd_cwnd_stamp = tcp_time_stamp;
1672 static inline int tcp_may_undo(struct tcp_sock *tp)
1674 return tp->undo_marker &&
1675 (!tp->undo_retrans || tcp_packet_delayed(tp));
1678 /* People celebrate: "We love our President!" */
1679 static int tcp_try_undo_recovery(struct sock *sk, struct tcp_sock *tp)
1681 if (tcp_may_undo(tp)) {
1682 /* Happy end! We did not retransmit anything
1683 * or our original transmission succeeded.
1685 DBGUNDO(sk, tp, tp->ca_state == TCP_CA_Loss ? "loss" : "retrans");
1686 tcp_undo_cwr(tp, 1);
1687 if (tp->ca_state == TCP_CA_Loss)
1688 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
1690 NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO);
1691 tp->undo_marker = 0;
1693 if (tp->snd_una == tp->high_seq && IsReno(tp)) {
1694 /* Hold old state until something *above* high_seq
1695 * is ACKed. For Reno it is MUST to prevent false
1696 * fast retransmits (RFC2582). SACK TCP is safe. */
1697 tcp_moderate_cwnd(tp);
1700 tcp_set_ca_state(tp, TCP_CA_Open);
1704 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
1705 static void tcp_try_undo_dsack(struct sock *sk, struct tcp_sock *tp)
1707 if (tp->undo_marker && !tp->undo_retrans) {
1708 DBGUNDO(sk, tp, "D-SACK");
1709 tcp_undo_cwr(tp, 1);
1710 tp->undo_marker = 0;
1711 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO);
1715 /* Undo during fast recovery after partial ACK. */
1717 static int tcp_try_undo_partial(struct sock *sk, struct tcp_sock *tp,
1720 /* Partial ACK arrived. Force Hoe's retransmit. */
1721 int failed = IsReno(tp) || tp->fackets_out>tp->reordering;
1723 if (tcp_may_undo(tp)) {
1724 /* Plain luck! Hole if filled with delayed
1725 * packet, rather than with a retransmit.
1727 if (tp->retrans_out == 0)
1728 tp->retrans_stamp = 0;
1730 tcp_update_reordering(tp, tcp_fackets_out(tp)+acked, 1);
1732 DBGUNDO(sk, tp, "Hoe");
1733 tcp_undo_cwr(tp, 0);
1734 NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO);
1736 /* So... Do not make Hoe's retransmit yet.
1737 * If the first packet was delayed, the rest
1738 * ones are most probably delayed as well.
1745 /* Undo during loss recovery after partial ACK. */
1746 static int tcp_try_undo_loss(struct sock *sk, struct tcp_sock *tp)
1748 if (tcp_may_undo(tp)) {
1749 struct sk_buff *skb;
1750 sk_stream_for_retrans_queue(skb, sk) {
1751 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1753 DBGUNDO(sk, tp, "partial loss");
1755 tp->left_out = tp->sacked_out;
1756 tcp_undo_cwr(tp, 1);
1757 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
1758 tp->retransmits = 0;
1759 tp->undo_marker = 0;
1761 tcp_set_ca_state(tp, TCP_CA_Open);
1767 static inline void tcp_complete_cwr(struct tcp_sock *tp)
1769 if (tcp_westwood_cwnd(tp))
1770 tp->snd_ssthresh = tp->snd_cwnd;
1772 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
1773 tp->snd_cwnd_stamp = tcp_time_stamp;
1776 static void tcp_try_to_open(struct sock *sk, struct tcp_sock *tp, int flag)
1778 tp->left_out = tp->sacked_out;
1780 if (tp->retrans_out == 0)
1781 tp->retrans_stamp = 0;
1786 if (tp->ca_state != TCP_CA_CWR) {
1787 int state = TCP_CA_Open;
1789 if (tp->left_out || tp->retrans_out || tp->undo_marker)
1790 state = TCP_CA_Disorder;
1792 if (tp->ca_state != state) {
1793 tcp_set_ca_state(tp, state);
1794 tp->high_seq = tp->snd_nxt;
1796 tcp_moderate_cwnd(tp);
1802 /* Process an event, which can update packets-in-flight not trivially.
1803 * Main goal of this function is to calculate new estimate for left_out,
1804 * taking into account both packets sitting in receiver's buffer and
1805 * packets lost by network.
1807 * Besides that it does CWND reduction, when packet loss is detected
1808 * and changes state of machine.
1810 * It does _not_ decide what to send, it is made in function
1811 * tcp_xmit_retransmit_queue().
1814 tcp_fastretrans_alert(struct sock *sk, u32 prior_snd_una,
1815 int prior_packets, int flag)
1817 struct tcp_sock *tp = tcp_sk(sk);
1818 int is_dupack = (tp->snd_una == prior_snd_una && !(flag&FLAG_NOT_DUP));
1820 /* Some technical things:
1821 * 1. Reno does not count dupacks (sacked_out) automatically. */
1822 if (!tp->packets_out)
1824 /* 2. SACK counts snd_fack in packets inaccurately. */
1825 if (tp->sacked_out == 0)
1826 tp->fackets_out = 0;
1828 /* Now state machine starts.
1829 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
1831 tp->prior_ssthresh = 0;
1833 /* B. In all the states check for reneging SACKs. */
1834 if (tp->sacked_out && tcp_check_sack_reneging(sk, tp))
1837 /* C. Process data loss notification, provided it is valid. */
1838 if ((flag&FLAG_DATA_LOST) &&
1839 before(tp->snd_una, tp->high_seq) &&
1840 tp->ca_state != TCP_CA_Open &&
1841 tp->fackets_out > tp->reordering) {
1842 tcp_mark_head_lost(sk, tp, tp->fackets_out-tp->reordering, tp->high_seq);
1843 NET_INC_STATS_BH(LINUX_MIB_TCPLOSS);
1846 /* D. Synchronize left_out to current state. */
1847 tcp_sync_left_out(tp);
1849 /* E. Check state exit conditions. State can be terminated
1850 * when high_seq is ACKed. */
1851 if (tp->ca_state == TCP_CA_Open) {
1852 if (!sysctl_tcp_frto)
1853 BUG_TRAP(tp->retrans_out == 0);
1854 tp->retrans_stamp = 0;
1855 } else if (!before(tp->snd_una, tp->high_seq)) {
1856 switch (tp->ca_state) {
1858 tp->retransmits = 0;
1859 if (tcp_try_undo_recovery(sk, tp))
1864 /* CWR is to be held something *above* high_seq
1865 * is ACKed for CWR bit to reach receiver. */
1866 if (tp->snd_una != tp->high_seq) {
1867 tcp_complete_cwr(tp);
1868 tcp_set_ca_state(tp, TCP_CA_Open);
1872 case TCP_CA_Disorder:
1873 tcp_try_undo_dsack(sk, tp);
1874 if (!tp->undo_marker ||
1875 /* For SACK case do not Open to allow to undo
1876 * catching for all duplicate ACKs. */
1877 IsReno(tp) || tp->snd_una != tp->high_seq) {
1878 tp->undo_marker = 0;
1879 tcp_set_ca_state(tp, TCP_CA_Open);
1883 case TCP_CA_Recovery:
1885 tcp_reset_reno_sack(tp);
1886 if (tcp_try_undo_recovery(sk, tp))
1888 tcp_complete_cwr(tp);
1893 /* F. Process state. */
1894 switch (tp->ca_state) {
1895 case TCP_CA_Recovery:
1896 if (prior_snd_una == tp->snd_una) {
1897 if (IsReno(tp) && is_dupack)
1898 tcp_add_reno_sack(tp);
1900 int acked = prior_packets - tp->packets_out;
1902 tcp_remove_reno_sacks(sk, tp, acked);
1903 is_dupack = tcp_try_undo_partial(sk, tp, acked);
1907 if (flag&FLAG_DATA_ACKED)
1908 tp->retransmits = 0;
1909 if (!tcp_try_undo_loss(sk, tp)) {
1910 tcp_moderate_cwnd(tp);
1911 tcp_xmit_retransmit_queue(sk);
1914 if (tp->ca_state != TCP_CA_Open)
1916 /* Loss is undone; fall through to processing in Open state. */
1919 if (tp->snd_una != prior_snd_una)
1920 tcp_reset_reno_sack(tp);
1922 tcp_add_reno_sack(tp);
1925 if (tp->ca_state == TCP_CA_Disorder)
1926 tcp_try_undo_dsack(sk, tp);
1928 if (!tcp_time_to_recover(sk, tp)) {
1929 tcp_try_to_open(sk, tp, flag);
1933 /* Otherwise enter Recovery state */
1936 NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY);
1938 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY);
1940 tp->high_seq = tp->snd_nxt;
1941 tp->prior_ssthresh = 0;
1942 tp->undo_marker = tp->snd_una;
1943 tp->undo_retrans = tp->retrans_out;
1945 if (tp->ca_state < TCP_CA_CWR) {
1946 if (!(flag&FLAG_ECE))
1947 tp->prior_ssthresh = tcp_current_ssthresh(tp);
1948 tp->snd_ssthresh = tcp_recalc_ssthresh(tp);
1949 TCP_ECN_queue_cwr(tp);
1952 tp->snd_cwnd_cnt = 0;
1953 tcp_set_ca_state(tp, TCP_CA_Recovery);
1956 if (is_dupack || tcp_head_timedout(sk, tp))
1957 tcp_update_scoreboard(sk, tp);
1959 tcp_xmit_retransmit_queue(sk);
1962 /* Read draft-ietf-tcplw-high-performance before mucking
1963 * with this code. (Superceeds RFC1323)
1965 static void tcp_ack_saw_tstamp(struct tcp_sock *tp, int flag)
1969 /* RTTM Rule: A TSecr value received in a segment is used to
1970 * update the averaged RTT measurement only if the segment
1971 * acknowledges some new data, i.e., only if it advances the
1972 * left edge of the send window.
1974 * See draft-ietf-tcplw-high-performance-00, section 3.3.
1975 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
1977 * Changed: reset backoff as soon as we see the first valid sample.
1978 * If we do not, we get strongly overstimated rto. With timestamps
1979 * samples are accepted even from very old segments: f.e., when rtt=1
1980 * increases to 8, we retransmit 5 times and after 8 seconds delayed
1981 * answer arrives rto becomes 120 seconds! If at least one of segments
1982 * in window is lost... Voila. --ANK (010210)
1984 seq_rtt = tcp_time_stamp - tp->rx_opt.rcv_tsecr;
1985 tcp_rtt_estimator(tp, seq_rtt);
1991 static void tcp_ack_no_tstamp(struct tcp_sock *tp, u32 seq_rtt, int flag)
1993 /* We don't have a timestamp. Can only use
1994 * packets that are not retransmitted to determine
1995 * rtt estimates. Also, we must not reset the
1996 * backoff for rto until we get a non-retransmitted
1997 * packet. This allows us to deal with a situation
1998 * where the network delay has increased suddenly.
1999 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
2002 if (flag & FLAG_RETRANS_DATA_ACKED)
2005 tcp_rtt_estimator(tp, seq_rtt);
2011 static inline void tcp_ack_update_rtt(struct tcp_sock *tp,
2012 int flag, s32 seq_rtt)
2014 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
2015 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
2016 tcp_ack_saw_tstamp(tp, flag);
2017 else if (seq_rtt >= 0)
2018 tcp_ack_no_tstamp(tp, seq_rtt, flag);
2022 * Compute congestion window to use.
2024 * This is from the implementation of BICTCP in
2025 * Lison-Xu, Kahaled Harfoush, and Injog Rhee.
2026 * "Binary Increase Congestion Control for Fast, Long Distance
2027 * Networks" in InfoComm 2004
2029 * http://www.csc.ncsu.edu/faculty/rhee/export/bitcp.pdf
2031 * Unless BIC is enabled and congestion window is large
2032 * this behaves the same as the original Reno.
2034 static inline __u32 bictcp_cwnd(struct tcp_sock *tp)
2036 /* orignal Reno behaviour */
2037 if (!tcp_is_bic(tp))
2038 return tp->snd_cwnd;
2040 if (tp->bictcp.last_cwnd == tp->snd_cwnd &&
2041 (s32)(tcp_time_stamp - tp->bictcp.last_stamp) <= (HZ>>5))
2042 return tp->bictcp.cnt;
2044 tp->bictcp.last_cwnd = tp->snd_cwnd;
2045 tp->bictcp.last_stamp = tcp_time_stamp;
2047 /* start off normal */
2048 if (tp->snd_cwnd <= sysctl_tcp_bic_low_window)
2049 tp->bictcp.cnt = tp->snd_cwnd;
2051 /* binary increase */
2052 else if (tp->snd_cwnd < tp->bictcp.last_max_cwnd) {
2053 __u32 dist = (tp->bictcp.last_max_cwnd - tp->snd_cwnd)
2056 if (dist > BICTCP_MAX_INCREMENT)
2057 /* linear increase */
2058 tp->bictcp.cnt = tp->snd_cwnd / BICTCP_MAX_INCREMENT;
2059 else if (dist <= 1U)
2060 /* binary search increase */
2061 tp->bictcp.cnt = tp->snd_cwnd * BICTCP_FUNC_OF_MIN_INCR
2064 /* binary search increase */
2065 tp->bictcp.cnt = tp->snd_cwnd / dist;
2067 /* slow start amd linear increase */
2068 if (tp->snd_cwnd < tp->bictcp.last_max_cwnd + BICTCP_B)
2070 tp->bictcp.cnt = tp->snd_cwnd * BICTCP_FUNC_OF_MIN_INCR
2072 else if (tp->snd_cwnd < tp->bictcp.last_max_cwnd
2073 + BICTCP_MAX_INCREMENT*(BICTCP_B-1))
2075 tp->bictcp.cnt = tp->snd_cwnd * (BICTCP_B-1)
2076 / (tp->snd_cwnd-tp->bictcp.last_max_cwnd);
2078 /* linear increase */
2079 tp->bictcp.cnt = tp->snd_cwnd / BICTCP_MAX_INCREMENT;
2081 return tp->bictcp.cnt;
2084 /* This is Jacobson's slow start and congestion avoidance.
2085 * SIGCOMM '88, p. 328.
2087 static inline void reno_cong_avoid(struct tcp_sock *tp)
2089 if (tp->snd_cwnd <= tp->snd_ssthresh) {
2090 /* In "safe" area, increase. */
2091 if (tp->snd_cwnd < tp->snd_cwnd_clamp)
2094 /* In dangerous area, increase slowly.
2095 * In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd
2097 if (tp->snd_cwnd_cnt >= bictcp_cwnd(tp)) {
2098 if (tp->snd_cwnd < tp->snd_cwnd_clamp)
2104 tp->snd_cwnd_stamp = tcp_time_stamp;
2107 /* This is based on the congestion detection/avoidance scheme described in
2108 * Lawrence S. Brakmo and Larry L. Peterson.
2109 * "TCP Vegas: End to end congestion avoidance on a global internet."
2110 * IEEE Journal on Selected Areas in Communication, 13(8):1465--1480,
2111 * October 1995. Available from:
2112 * ftp://ftp.cs.arizona.edu/xkernel/Papers/jsac.ps
2114 * See http://www.cs.arizona.edu/xkernel/ for their implementation.
2115 * The main aspects that distinguish this implementation from the
2116 * Arizona Vegas implementation are:
2117 * o We do not change the loss detection or recovery mechanisms of
2118 * Linux in any way. Linux already recovers from losses quite well,
2119 * using fine-grained timers, NewReno, and FACK.
2120 * o To avoid the performance penalty imposed by increasing cwnd
2121 * only every-other RTT during slow start, we increase during
2122 * every RTT during slow start, just like Reno.
2123 * o Largely to allow continuous cwnd growth during slow start,
2124 * we use the rate at which ACKs come back as the "actual"
2125 * rate, rather than the rate at which data is sent.
2126 * o To speed convergence to the right rate, we set the cwnd
2127 * to achieve the right ("actual") rate when we exit slow start.
2128 * o To filter out the noise caused by delayed ACKs, we use the
2129 * minimum RTT sample observed during the last RTT to calculate
2131 * o When the sender re-starts from idle, it waits until it has
2132 * received ACKs for an entire flight of new data before making
2133 * a cwnd adjustment decision. The original Vegas implementation
2134 * assumed senders never went idle.
2136 static void vegas_cong_avoid(struct tcp_sock *tp, u32 ack, u32 seq_rtt)
2138 /* The key players are v_beg_snd_una and v_beg_snd_nxt.
2140 * These are so named because they represent the approximate values
2141 * of snd_una and snd_nxt at the beginning of the current RTT. More
2142 * precisely, they represent the amount of data sent during the RTT.
2143 * At the end of the RTT, when we receive an ACK for v_beg_snd_nxt,
2144 * we will calculate that (v_beg_snd_nxt - v_beg_snd_una) outstanding
2145 * bytes of data have been ACKed during the course of the RTT, giving
2146 * an "actual" rate of:
2148 * (v_beg_snd_nxt - v_beg_snd_una) / (rtt duration)
2150 * Unfortunately, v_beg_snd_una is not exactly equal to snd_una,
2151 * because delayed ACKs can cover more than one segment, so they
2152 * don't line up nicely with the boundaries of RTTs.
2154 * Another unfortunate fact of life is that delayed ACKs delay the
2155 * advance of the left edge of our send window, so that the number
2156 * of bytes we send in an RTT is often less than our cwnd will allow.
2157 * So we keep track of our cwnd separately, in v_beg_snd_cwnd.
2160 if (after(ack, tp->vegas.beg_snd_nxt)) {
2161 /* Do the Vegas once-per-RTT cwnd adjustment. */
2162 u32 old_wnd, old_snd_cwnd;
2165 /* Here old_wnd is essentially the window of data that was
2166 * sent during the previous RTT, and has all
2167 * been acknowledged in the course of the RTT that ended
2168 * with the ACK we just received. Likewise, old_snd_cwnd
2169 * is the cwnd during the previous RTT.
2171 old_wnd = (tp->vegas.beg_snd_nxt - tp->vegas.beg_snd_una) /
2173 old_snd_cwnd = tp->vegas.beg_snd_cwnd;
2175 /* Save the extent of the current window so we can use this
2176 * at the end of the next RTT.
2178 tp->vegas.beg_snd_una = tp->vegas.beg_snd_nxt;
2179 tp->vegas.beg_snd_nxt = tp->snd_nxt;
2180 tp->vegas.beg_snd_cwnd = tp->snd_cwnd;
2182 /* Take into account the current RTT sample too, to
2183 * decrease the impact of delayed acks. This double counts
2184 * this sample since we count it for the next window as well,
2185 * but that's not too awful, since we're taking the min,
2186 * rather than averaging.
2188 vegas_rtt_calc(tp, seq_rtt);
2190 /* We do the Vegas calculations only if we got enough RTT
2191 * samples that we can be reasonably sure that we got
2192 * at least one RTT sample that wasn't from a delayed ACK.
2193 * If we only had 2 samples total,
2194 * then that means we're getting only 1 ACK per RTT, which
2195 * means they're almost certainly delayed ACKs.
2196 * If we have 3 samples, we should be OK.
2199 if (tp->vegas.cntRTT <= 2) {
2200 /* We don't have enough RTT samples to do the Vegas
2201 * calculation, so we'll behave like Reno.
2203 if (tp->snd_cwnd > tp->snd_ssthresh)
2206 u32 rtt, target_cwnd, diff;
2208 /* We have enough RTT samples, so, using the Vegas
2209 * algorithm, we determine if we should increase or
2210 * decrease cwnd, and by how much.
2213 /* Pluck out the RTT we are using for the Vegas
2214 * calculations. This is the min RTT seen during the
2215 * last RTT. Taking the min filters out the effects
2216 * of delayed ACKs, at the cost of noticing congestion
2219 rtt = tp->vegas.minRTT;
2221 /* Calculate the cwnd we should have, if we weren't
2225 * (actual rate in segments) * baseRTT
2226 * We keep it as a fixed point number with
2227 * V_PARAM_SHIFT bits to the right of the binary point.
2229 target_cwnd = ((old_wnd * tp->vegas.baseRTT)
2230 << V_PARAM_SHIFT) / rtt;
2232 /* Calculate the difference between the window we had,
2233 * and the window we would like to have. This quantity
2234 * is the "Diff" from the Arizona Vegas papers.
2236 * Again, this is a fixed point number with
2237 * V_PARAM_SHIFT bits to the right of the binary
2240 diff = (old_wnd << V_PARAM_SHIFT) - target_cwnd;
2242 if (tp->snd_cwnd < tp->snd_ssthresh) {
2244 if (diff > sysctl_tcp_vegas_gamma) {
2245 /* Going too fast. Time to slow down
2246 * and switch to congestion avoidance.
2248 tp->snd_ssthresh = 2;
2250 /* Set cwnd to match the actual rate
2252 * cwnd = (actual rate) * baseRTT
2253 * Then we add 1 because the integer
2254 * truncation robs us of full link
2257 tp->snd_cwnd = min(tp->snd_cwnd,
2263 /* Congestion avoidance. */
2266 /* Figure out where we would like cwnd
2269 if (diff > sysctl_tcp_vegas_beta) {
2270 /* The old window was too fast, so
2273 next_snd_cwnd = old_snd_cwnd - 1;
2274 } else if (diff < sysctl_tcp_vegas_alpha) {
2275 /* We don't have enough extra packets
2276 * in the network, so speed up.
2278 next_snd_cwnd = old_snd_cwnd + 1;
2280 /* Sending just as fast as we
2283 next_snd_cwnd = old_snd_cwnd;
2286 /* Adjust cwnd upward or downward, toward the
2289 if (next_snd_cwnd > tp->snd_cwnd)
2291 else if (next_snd_cwnd < tp->snd_cwnd)
2296 /* Wipe the slate clean for the next RTT. */
2297 tp->vegas.cntRTT = 0;
2298 tp->vegas.minRTT = 0x7fffffff;
2301 /* The following code is executed for every ack we receive,
2302 * except for conditions checked in should_advance_cwnd()
2303 * before the call to tcp_cong_avoid(). Mainly this means that
2304 * we only execute this code if the ack actually acked some
2308 /* If we are in slow start, increase our cwnd in response to this ACK.
2309 * (If we are not in slow start then we are in congestion avoidance,
2310 * and adjust our congestion window only once per RTT. See the code
2313 if (tp->snd_cwnd <= tp->snd_ssthresh)
2316 /* to keep cwnd from growing without bound */
2317 tp->snd_cwnd = min_t(u32, tp->snd_cwnd, tp->snd_cwnd_clamp);
2319 /* Make sure that we are never so timid as to reduce our cwnd below
2322 * Going below 2 MSS would risk huge delayed ACKs from our receiver.
2324 tp->snd_cwnd = max(tp->snd_cwnd, 2U);
2326 tp->snd_cwnd_stamp = tcp_time_stamp;
2329 static inline void tcp_cong_avoid(struct tcp_sock *tp, u32 ack, u32 seq_rtt)
2331 if (tcp_vegas_enabled(tp))
2332 vegas_cong_avoid(tp, ack, seq_rtt);
2334 reno_cong_avoid(tp);
2337 /* Restart timer after forward progress on connection.
2338 * RFC2988 recommends to restart timer to now+rto.
2341 static inline void tcp_ack_packets_out(struct sock *sk, struct tcp_sock *tp)
2343 if (!tp->packets_out) {
2344 tcp_clear_xmit_timer(sk, TCP_TIME_RETRANS);
2346 tcp_reset_xmit_timer(sk, TCP_TIME_RETRANS, tp->rto);
2350 /* There is one downside to this scheme. Although we keep the
2351 * ACK clock ticking, adjusting packet counters and advancing
2352 * congestion window, we do not liberate socket send buffer
2355 * Mucking with skb->truesize and sk->sk_wmem_alloc et al.
2356 * then making a write space wakeup callback is a possible
2357 * future enhancement. WARNING: it is not trivial to make.
2359 static int tcp_tso_acked(struct sock *sk, struct sk_buff *skb,
2360 __u32 now, __s32 *seq_rtt)
2362 struct tcp_sock *tp = tcp_sk(sk);
2363 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
2364 __u32 seq = tp->snd_una;
2365 __u32 packets_acked;
2368 /* If we get here, the whole TSO packet has not been
2371 BUG_ON(!after(scb->end_seq, seq));
2373 packets_acked = tcp_skb_pcount(skb);
2374 if (tcp_trim_head(sk, skb, seq - scb->seq))
2376 packets_acked -= tcp_skb_pcount(skb);
2378 if (packets_acked) {
2379 __u8 sacked = scb->sacked;
2381 acked |= FLAG_DATA_ACKED;
2383 if (sacked & TCPCB_RETRANS) {
2384 if (sacked & TCPCB_SACKED_RETRANS)
2385 tp->retrans_out -= packets_acked;
2386 acked |= FLAG_RETRANS_DATA_ACKED;
2388 } else if (*seq_rtt < 0)
2389 *seq_rtt = now - scb->when;
2390 if (sacked & TCPCB_SACKED_ACKED)
2391 tp->sacked_out -= packets_acked;
2392 if (sacked & TCPCB_LOST)
2393 tp->lost_out -= packets_acked;
2394 if (sacked & TCPCB_URG) {
2396 !before(seq, tp->snd_up))
2399 } else if (*seq_rtt < 0)
2400 *seq_rtt = now - scb->when;
2402 if (tp->fackets_out) {
2403 __u32 dval = min(tp->fackets_out, packets_acked);
2404 tp->fackets_out -= dval;
2406 tp->packets_out -= packets_acked;
2408 BUG_ON(tcp_skb_pcount(skb) == 0);
2409 BUG_ON(!before(scb->seq, scb->end_seq));
2416 /* Remove acknowledged frames from the retransmission queue. */
2417 static int tcp_clean_rtx_queue(struct sock *sk, __s32 *seq_rtt_p)
2419 struct tcp_sock *tp = tcp_sk(sk);
2420 struct sk_buff *skb;
2421 __u32 now = tcp_time_stamp;
2425 while ((skb = skb_peek(&sk->sk_write_queue)) &&
2426 skb != sk->sk_send_head) {
2427 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
2428 __u8 sacked = scb->sacked;
2430 /* If our packet is before the ack sequence we can
2431 * discard it as it's confirmed to have arrived at
2434 if (after(scb->end_seq, tp->snd_una)) {
2435 if (tcp_skb_pcount(skb) > 1)
2436 acked |= tcp_tso_acked(sk, skb,
2441 /* Initial outgoing SYN's get put onto the write_queue
2442 * just like anything else we transmit. It is not
2443 * true data, and if we misinform our callers that
2444 * this ACK acks real data, we will erroneously exit
2445 * connection startup slow start one packet too
2446 * quickly. This is severely frowned upon behavior.
2448 if (!(scb->flags & TCPCB_FLAG_SYN)) {
2449 acked |= FLAG_DATA_ACKED;
2451 acked |= FLAG_SYN_ACKED;
2452 tp->retrans_stamp = 0;
2456 if (sacked & TCPCB_RETRANS) {
2457 if(sacked & TCPCB_SACKED_RETRANS)
2458 tp->retrans_out -= tcp_skb_pcount(skb);
2459 acked |= FLAG_RETRANS_DATA_ACKED;
2461 } else if (seq_rtt < 0)
2462 seq_rtt = now - scb->when;
2463 if (sacked & TCPCB_SACKED_ACKED)
2464 tp->sacked_out -= tcp_skb_pcount(skb);
2465 if (sacked & TCPCB_LOST)
2466 tp->lost_out -= tcp_skb_pcount(skb);
2467 if (sacked & TCPCB_URG) {
2469 !before(scb->end_seq, tp->snd_up))
2472 } else if (seq_rtt < 0)
2473 seq_rtt = now - scb->when;
2474 tcp_dec_pcount_approx(&tp->fackets_out, skb);
2475 tcp_packets_out_dec(tp, skb);
2476 __skb_unlink(skb, skb->list);
2477 sk_stream_free_skb(sk, skb);
2480 if (acked&FLAG_ACKED) {
2481 tcp_ack_update_rtt(tp, acked, seq_rtt);
2482 tcp_ack_packets_out(sk, tp);
2485 #if FASTRETRANS_DEBUG > 0
2486 BUG_TRAP((int)tp->sacked_out >= 0);
2487 BUG_TRAP((int)tp->lost_out >= 0);
2488 BUG_TRAP((int)tp->retrans_out >= 0);
2489 if (!tp->packets_out && tp->rx_opt.sack_ok) {
2491 printk(KERN_DEBUG "Leak l=%u %d\n",
2492 tp->lost_out, tp->ca_state);
2495 if (tp->sacked_out) {
2496 printk(KERN_DEBUG "Leak s=%u %d\n",
2497 tp->sacked_out, tp->ca_state);
2500 if (tp->retrans_out) {
2501 printk(KERN_DEBUG "Leak r=%u %d\n",
2502 tp->retrans_out, tp->ca_state);
2503 tp->retrans_out = 0;
2507 *seq_rtt_p = seq_rtt;
2511 static void tcp_ack_probe(struct sock *sk)
2513 struct tcp_sock *tp = tcp_sk(sk);
2515 /* Was it a usable window open? */
2517 if (!after(TCP_SKB_CB(sk->sk_send_head)->end_seq,
2518 tp->snd_una + tp->snd_wnd)) {
2520 tcp_clear_xmit_timer(sk, TCP_TIME_PROBE0);
2521 /* Socket must be waked up by subsequent tcp_data_snd_check().
2522 * This function is not for random using!
2525 tcp_reset_xmit_timer(sk, TCP_TIME_PROBE0,
2526 min(tp->rto << tp->backoff, TCP_RTO_MAX));
2530 static inline int tcp_ack_is_dubious(struct tcp_sock *tp, int flag)
2532 return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
2533 tp->ca_state != TCP_CA_Open);
2536 static inline int tcp_may_raise_cwnd(struct tcp_sock *tp, int flag)
2538 return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
2539 !((1<<tp->ca_state)&(TCPF_CA_Recovery|TCPF_CA_CWR));
2542 /* Check that window update is acceptable.
2543 * The function assumes that snd_una<=ack<=snd_next.
2545 static inline int tcp_may_update_window(struct tcp_sock *tp, u32 ack,
2546 u32 ack_seq, u32 nwin)
2548 return (after(ack, tp->snd_una) ||
2549 after(ack_seq, tp->snd_wl1) ||
2550 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd));
2553 /* Update our send window.
2555 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
2556 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
2558 static int tcp_ack_update_window(struct sock *sk, struct tcp_sock *tp,
2559 struct sk_buff *skb, u32 ack, u32 ack_seq)
2562 u32 nwin = ntohs(skb->h.th->window);
2564 if (likely(!skb->h.th->syn))
2565 nwin <<= tp->rx_opt.snd_wscale;
2567 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
2568 flag |= FLAG_WIN_UPDATE;
2569 tcp_update_wl(tp, ack, ack_seq);
2571 if (tp->snd_wnd != nwin) {
2574 /* Note, it is the only place, where
2575 * fast path is recovered for sending TCP.
2577 tcp_fast_path_check(sk, tp);
2579 if (nwin > tp->max_window) {
2580 tp->max_window = nwin;
2581 tcp_sync_mss(sk, tp->pmtu_cookie);
2591 static void tcp_process_frto(struct sock *sk, u32 prior_snd_una)
2593 struct tcp_sock *tp = tcp_sk(sk);
2595 tcp_sync_left_out(tp);
2597 if (tp->snd_una == prior_snd_una ||
2598 !before(tp->snd_una, tp->frto_highmark)) {
2599 /* RTO was caused by loss, start retransmitting in
2600 * go-back-N slow start
2602 tcp_enter_frto_loss(sk);
2606 if (tp->frto_counter == 1) {
2607 /* First ACK after RTO advances the window: allow two new
2610 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
2612 /* Also the second ACK after RTO advances the window.
2613 * The RTO was likely spurious. Reduce cwnd and continue
2614 * in congestion avoidance
2616 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2617 tcp_moderate_cwnd(tp);
2620 /* F-RTO affects on two new ACKs following RTO.
2621 * At latest on third ACK the TCP behavor is back to normal.
2623 tp->frto_counter = (tp->frto_counter + 1) % 3;
2632 * This function initializes fields used in TCP Westwood+. We can't
2633 * get no information about RTTmin at this time so we simply set it to
2634 * TCP_WESTWOOD_INIT_RTT. This value was chosen to be too conservative
2635 * since in this way we're sure it will be updated in a consistent
2636 * way as soon as possible. It will reasonably happen within the first
2637 * RTT period of the connection lifetime.
2640 static void init_westwood(struct sock *sk)
2642 struct tcp_sock *tp = tcp_sk(sk);
2644 tp->westwood.bw_ns_est = 0;
2645 tp->westwood.bw_est = 0;
2646 tp->westwood.accounted = 0;
2647 tp->westwood.cumul_ack = 0;
2648 tp->westwood.rtt_win_sx = tcp_time_stamp;
2649 tp->westwood.rtt = TCP_WESTWOOD_INIT_RTT;
2650 tp->westwood.rtt_min = TCP_WESTWOOD_INIT_RTT;
2651 tp->westwood.snd_una = tp->snd_una;
2655 * @westwood_do_filter
2656 * Low-pass filter. Implemented using constant coeffients.
2659 static inline __u32 westwood_do_filter(__u32 a, __u32 b)
2661 return (((7 * a) + b) >> 3);
2664 static void westwood_filter(struct sock *sk, __u32 delta)
2666 struct tcp_sock *tp = tcp_sk(sk);
2668 tp->westwood.bw_ns_est =
2669 westwood_do_filter(tp->westwood.bw_ns_est,
2670 tp->westwood.bk / delta);
2671 tp->westwood.bw_est =
2672 westwood_do_filter(tp->westwood.bw_est,
2673 tp->westwood.bw_ns_est);
2677 * @westwood_update_rttmin
2678 * It is used to update RTTmin. In this case we MUST NOT use
2679 * WESTWOOD_RTT_MIN minimum bound since we could be on a LAN!
2682 static inline __u32 westwood_update_rttmin(const struct sock *sk)
2684 const struct tcp_sock *tp = tcp_sk(sk);
2685 __u32 rttmin = tp->westwood.rtt_min;
2687 if (tp->westwood.rtt != 0 &&
2688 (tp->westwood.rtt < tp->westwood.rtt_min || !rttmin))
2689 rttmin = tp->westwood.rtt;
2696 * Evaluate increases for dk.
2699 static inline __u32 westwood_acked(const struct sock *sk)
2701 const struct tcp_sock *tp = tcp_sk(sk);
2703 return tp->snd_una - tp->westwood.snd_una;
2707 * @westwood_new_window
2708 * It evaluates if we are receiving data inside the same RTT window as
2711 * It returns 0 if we are still evaluating samples in the same RTT
2712 * window, 1 if the sample has to be considered in the next window.
2715 static int westwood_new_window(const struct sock *sk)
2717 const struct tcp_sock *tp = tcp_sk(sk);
2722 left_bound = tp->westwood.rtt_win_sx;
2723 rtt = max(tp->westwood.rtt, (u32) TCP_WESTWOOD_RTT_MIN);
2726 * A RTT-window has passed. Be careful since if RTT is less than
2727 * 50ms we don't filter but we continue 'building the sample'.
2728 * This minimum limit was choosen since an estimation on small
2729 * time intervals is better to avoid...
2730 * Obvioulsy on a LAN we reasonably will always have
2731 * right_bound = left_bound + WESTWOOD_RTT_MIN
2734 if ((left_bound + rtt) < tcp_time_stamp)
2741 * @westwood_update_window
2742 * It updates RTT evaluation window if it is the right moment to do
2743 * it. If so it calls filter for evaluating bandwidth.
2746 static void __westwood_update_window(struct sock *sk, __u32 now)
2748 struct tcp_sock *tp = tcp_sk(sk);
2749 __u32 delta = now - tp->westwood.rtt_win_sx;
2752 if (tp->westwood.rtt)
2753 westwood_filter(sk, delta);
2755 tp->westwood.bk = 0;
2756 tp->westwood.rtt_win_sx = tcp_time_stamp;
2761 static void westwood_update_window(struct sock *sk, __u32 now)
2763 if (westwood_new_window(sk))
2764 __westwood_update_window(sk, now);
2768 * @__tcp_westwood_fast_bw
2769 * It is called when we are in fast path. In particular it is called when
2770 * header prediction is successfull. In such case infact update is
2771 * straight forward and doesn't need any particular care.
2774 static void __tcp_westwood_fast_bw(struct sock *sk, struct sk_buff *skb)
2776 struct tcp_sock *tp = tcp_sk(sk);
2778 westwood_update_window(sk, tcp_time_stamp);
2780 tp->westwood.bk += westwood_acked(sk);
2781 tp->westwood.snd_una = tp->snd_una;
2782 tp->westwood.rtt_min = westwood_update_rttmin(sk);
2785 static inline void tcp_westwood_fast_bw(struct sock *sk, struct sk_buff *skb)
2787 if (tcp_is_westwood(tcp_sk(sk)))
2788 __tcp_westwood_fast_bw(sk, skb);
2793 * @westwood_dupack_update
2794 * It updates accounted and cumul_ack when receiving a dupack.
2797 static void westwood_dupack_update(struct sock *sk)
2799 struct tcp_sock *tp = tcp_sk(sk);
2801 tp->westwood.accounted += tp->mss_cache_std;
2802 tp->westwood.cumul_ack = tp->mss_cache_std;
2805 static inline int westwood_may_change_cumul(struct tcp_sock *tp)
2807 return (tp->westwood.cumul_ack > tp->mss_cache_std);
2810 static inline void westwood_partial_update(struct tcp_sock *tp)
2812 tp->westwood.accounted -= tp->westwood.cumul_ack;
2813 tp->westwood.cumul_ack = tp->mss_cache_std;
2816 static inline void westwood_complete_update(struct tcp_sock *tp)
2818 tp->westwood.cumul_ack -= tp->westwood.accounted;
2819 tp->westwood.accounted = 0;
2823 * @westwood_acked_count
2824 * This function evaluates cumul_ack for evaluating dk in case of
2825 * delayed or partial acks.
2828 static inline __u32 westwood_acked_count(struct sock *sk)
2830 struct tcp_sock *tp = tcp_sk(sk);
2832 tp->westwood.cumul_ack = westwood_acked(sk);
2834 /* If cumul_ack is 0 this is a dupack since it's not moving
2837 if (!(tp->westwood.cumul_ack))
2838 westwood_dupack_update(sk);
2840 if (westwood_may_change_cumul(tp)) {
2841 /* Partial or delayed ack */
2842 if (tp->westwood.accounted >= tp->westwood.cumul_ack)
2843 westwood_partial_update(tp);
2845 westwood_complete_update(tp);
2848 tp->westwood.snd_una = tp->snd_una;
2850 return tp->westwood.cumul_ack;
2855 * @__tcp_westwood_slow_bw
2856 * It is called when something is going wrong..even if there could
2857 * be no problems! Infact a simple delayed packet may trigger a
2858 * dupack. But we need to be careful in such case.
2861 static void __tcp_westwood_slow_bw(struct sock *sk, struct sk_buff *skb)
2863 struct tcp_sock *tp = tcp_sk(sk);
2865 westwood_update_window(sk, tcp_time_stamp);
2867 tp->westwood.bk += westwood_acked_count(sk);
2868 tp->westwood.rtt_min = westwood_update_rttmin(sk);
2871 static inline void tcp_westwood_slow_bw(struct sock *sk, struct sk_buff *skb)
2873 if (tcp_is_westwood(tcp_sk(sk)))
2874 __tcp_westwood_slow_bw(sk, skb);
2877 /* This routine deals with incoming acks, but not outgoing ones. */
2878 static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag)
2880 struct tcp_sock *tp = tcp_sk(sk);
2881 u32 prior_snd_una = tp->snd_una;
2882 u32 ack_seq = TCP_SKB_CB(skb)->seq;
2883 u32 ack = TCP_SKB_CB(skb)->ack_seq;
2884 u32 prior_in_flight;
2888 /* If the ack is newer than sent or older than previous acks
2889 * then we can probably ignore it.
2891 if (after(ack, tp->snd_nxt))
2892 goto uninteresting_ack;
2894 if (before(ack, prior_snd_una))
2897 if (!(flag&FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
2898 /* Window is constant, pure forward advance.
2899 * No more checks are required.
2900 * Note, we use the fact that SND.UNA>=SND.WL2.
2902 tcp_update_wl(tp, ack, ack_seq);
2904 tcp_westwood_fast_bw(sk, skb);
2905 flag |= FLAG_WIN_UPDATE;
2907 NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS);
2909 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
2912 NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS);
2914 flag |= tcp_ack_update_window(sk, tp, skb, ack, ack_seq);
2916 if (TCP_SKB_CB(skb)->sacked)
2917 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
2919 if (TCP_ECN_rcv_ecn_echo(tp, skb->h.th))
2922 tcp_westwood_slow_bw(sk,skb);
2925 /* We passed data and got it acked, remove any soft error
2926 * log. Something worked...
2928 sk->sk_err_soft = 0;
2929 tp->rcv_tstamp = tcp_time_stamp;
2930 prior_packets = tp->packets_out;
2934 prior_in_flight = tcp_packets_in_flight(tp);
2936 /* See if we can take anything off of the retransmit queue. */
2937 flag |= tcp_clean_rtx_queue(sk, &seq_rtt);
2939 if (tp->frto_counter)
2940 tcp_process_frto(sk, prior_snd_una);
2942 if (tcp_ack_is_dubious(tp, flag)) {
2943 /* Advanve CWND, if state allows this. */
2944 if ((flag & FLAG_DATA_ACKED) &&
2945 (tcp_vegas_enabled(tp) || prior_in_flight >= tp->snd_cwnd) &&
2946 tcp_may_raise_cwnd(tp, flag))
2947 tcp_cong_avoid(tp, ack, seq_rtt);
2948 tcp_fastretrans_alert(sk, prior_snd_una, prior_packets, flag);
2950 if ((flag & FLAG_DATA_ACKED) &&
2951 (tcp_vegas_enabled(tp) || prior_in_flight >= tp->snd_cwnd))
2952 tcp_cong_avoid(tp, ack, seq_rtt);
2955 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag&FLAG_NOT_DUP))
2956 dst_confirm(sk->sk_dst_cache);
2963 /* If this ack opens up a zero window, clear backoff. It was
2964 * being used to time the probes, and is probably far higher than
2965 * it needs to be for normal retransmission.
2967 if (sk->sk_send_head)
2972 if (TCP_SKB_CB(skb)->sacked)
2973 tcp_sacktag_write_queue(sk, skb, prior_snd_una);
2976 SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
2981 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
2982 * But, this can also be called on packets in the established flow when
2983 * the fast version below fails.
2985 void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx, int estab)
2988 struct tcphdr *th = skb->h.th;
2989 int length=(th->doff*4)-sizeof(struct tcphdr);
2991 ptr = (unsigned char *)(th + 1);
2992 opt_rx->saw_tstamp = 0;
3001 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
3006 if (opsize < 2) /* "silly options" */
3008 if (opsize > length)
3009 return; /* don't parse partial options */
3012 if(opsize==TCPOLEN_MSS && th->syn && !estab) {
3013 u16 in_mss = ntohs(*(__u16 *)ptr);
3015 if (opt_rx->user_mss && opt_rx->user_mss < in_mss)
3016 in_mss = opt_rx->user_mss;
3017 opt_rx->mss_clamp = in_mss;
3022 if(opsize==TCPOLEN_WINDOW && th->syn && !estab)
3023 if (sysctl_tcp_window_scaling) {
3024 opt_rx->wscale_ok = 1;
3025 opt_rx->snd_wscale = *(__u8 *)ptr;
3026 if(opt_rx->snd_wscale > 14) {
3028 printk(KERN_INFO "tcp_parse_options: Illegal window "
3029 "scaling value %d >14 received.\n",
3030 opt_rx->snd_wscale);
3031 opt_rx->snd_wscale = 14;
3035 case TCPOPT_TIMESTAMP:
3036 if(opsize==TCPOLEN_TIMESTAMP) {
3037 if ((estab && opt_rx->tstamp_ok) ||
3038 (!estab && sysctl_tcp_timestamps)) {
3039 opt_rx->saw_tstamp = 1;
3040 opt_rx->rcv_tsval = ntohl(*(__u32 *)ptr);
3041 opt_rx->rcv_tsecr = ntohl(*(__u32 *)(ptr+4));
3045 case TCPOPT_SACK_PERM:
3046 if(opsize==TCPOLEN_SACK_PERM && th->syn && !estab) {
3047 if (sysctl_tcp_sack) {
3048 opt_rx->sack_ok = 1;
3049 tcp_sack_reset(opt_rx);
3055 if((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3056 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3058 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3067 /* Fast parse options. This hopes to only see timestamps.
3068 * If it is wrong it falls back on tcp_parse_options().
3070 static inline int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th,
3071 struct tcp_sock *tp)
3073 if (th->doff == sizeof(struct tcphdr)>>2) {
3074 tp->rx_opt.saw_tstamp = 0;
3076 } else if (tp->rx_opt.tstamp_ok &&
3077 th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) {
3078 __u32 *ptr = (__u32 *)(th + 1);
3079 if (*ptr == ntohl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3080 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3081 tp->rx_opt.saw_tstamp = 1;
3083 tp->rx_opt.rcv_tsval = ntohl(*ptr);
3085 tp->rx_opt.rcv_tsecr = ntohl(*ptr);
3089 tcp_parse_options(skb, &tp->rx_opt, 1);
3093 static inline void tcp_store_ts_recent(struct tcp_sock *tp)
3095 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3096 tp->rx_opt.ts_recent_stamp = xtime.tv_sec;
3099 static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3101 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3102 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
3103 * extra check below makes sure this can only happen
3104 * for pure ACK frames. -DaveM
3106 * Not only, also it occurs for expired timestamps.
3109 if((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) >= 0 ||
3110 xtime.tv_sec >= tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS)
3111 tcp_store_ts_recent(tp);
3115 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3117 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3118 * it can pass through stack. So, the following predicate verifies that
3119 * this segment is not used for anything but congestion avoidance or
3120 * fast retransmit. Moreover, we even are able to eliminate most of such
3121 * second order effects, if we apply some small "replay" window (~RTO)
3122 * to timestamp space.
3124 * All these measures still do not guarantee that we reject wrapped ACKs
3125 * on networks with high bandwidth, when sequence space is recycled fastly,
3126 * but it guarantees that such events will be very rare and do not affect
3127 * connection seriously. This doesn't look nice, but alas, PAWS is really
3130 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3131 * states that events when retransmit arrives after original data are rare.
3132 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3133 * the biggest problem on large power networks even with minor reordering.
3134 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3135 * up to bandwidth of 18Gigabit/sec. 8) ]
3138 static int tcp_disordered_ack(struct tcp_sock *tp, struct sk_buff *skb)
3140 struct tcphdr *th = skb->h.th;
3141 u32 seq = TCP_SKB_CB(skb)->seq;
3142 u32 ack = TCP_SKB_CB(skb)->ack_seq;
3144 return (/* 1. Pure ACK with correct sequence number. */
3145 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3147 /* 2. ... and duplicate ACK. */
3148 ack == tp->snd_una &&
3150 /* 3. ... and does not update window. */
3151 !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3153 /* 4. ... and sits in replay window. */
3154 (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (tp->rto*1024)/HZ);
3157 static inline int tcp_paws_discard(struct tcp_sock *tp, struct sk_buff *skb)
3159 return ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > TCP_PAWS_WINDOW &&
3160 xtime.tv_sec < tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS &&
3161 !tcp_disordered_ack(tp, skb));
3164 /* Check segment sequence number for validity.
3166 * Segment controls are considered valid, if the segment
3167 * fits to the window after truncation to the window. Acceptability
3168 * of data (and SYN, FIN, of course) is checked separately.
3169 * See tcp_data_queue(), for example.
3171 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3172 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3173 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3174 * (borrowed from freebsd)
3177 static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq)
3179 return !before(end_seq, tp->rcv_wup) &&
3180 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
3183 /* When we get a reset we do this. */
3184 static void tcp_reset(struct sock *sk)
3186 /* We want the right error as BSD sees it (and indeed as we do). */
3187 switch (sk->sk_state) {
3189 sk->sk_err = ECONNREFUSED;
3191 case TCP_CLOSE_WAIT:
3197 sk->sk_err = ECONNRESET;
3200 if (!sock_flag(sk, SOCK_DEAD))
3201 sk->sk_error_report(sk);
3207 * Process the FIN bit. This now behaves as it is supposed to work
3208 * and the FIN takes effect when it is validly part of sequence
3209 * space. Not before when we get holes.
3211 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3212 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
3215 * If we are in FINWAIT-1, a received FIN indicates simultaneous
3216 * close and we go into CLOSING (and later onto TIME-WAIT)
3218 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3220 static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th)
3222 struct tcp_sock *tp = tcp_sk(sk);
3224 tcp_schedule_ack(tp);
3226 sk->sk_shutdown |= RCV_SHUTDOWN;
3227 sock_set_flag(sk, SOCK_DONE);
3229 switch (sk->sk_state) {
3231 case TCP_ESTABLISHED:
3232 /* Move to CLOSE_WAIT */
3233 tcp_set_state(sk, TCP_CLOSE_WAIT);
3234 tp->ack.pingpong = 1;
3237 case TCP_CLOSE_WAIT:
3239 /* Received a retransmission of the FIN, do
3244 /* RFC793: Remain in the LAST-ACK state. */
3248 /* This case occurs when a simultaneous close
3249 * happens, we must ack the received FIN and
3250 * enter the CLOSING state.
3253 tcp_set_state(sk, TCP_CLOSING);
3256 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3258 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
3261 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3262 * cases we should never reach this piece of code.
3264 printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n",
3265 __FUNCTION__, sk->sk_state);
3269 /* It _is_ possible, that we have something out-of-order _after_ FIN.
3270 * Probably, we should reset in this case. For now drop them.
3272 __skb_queue_purge(&tp->out_of_order_queue);
3273 if (tp->rx_opt.sack_ok)
3274 tcp_sack_reset(&tp->rx_opt);
3275 sk_stream_mem_reclaim(sk);
3277 if (!sock_flag(sk, SOCK_DEAD)) {
3278 sk->sk_state_change(sk);
3280 /* Do not send POLL_HUP for half duplex close. */
3281 if (sk->sk_shutdown == SHUTDOWN_MASK ||
3282 sk->sk_state == TCP_CLOSE)
3283 sk_wake_async(sk, 1, POLL_HUP);
3285 sk_wake_async(sk, 1, POLL_IN);
3289 static __inline__ int
3290 tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, u32 end_seq)
3292 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
3293 if (before(seq, sp->start_seq))
3294 sp->start_seq = seq;
3295 if (after(end_seq, sp->end_seq))
3296 sp->end_seq = end_seq;
3302 static inline void tcp_dsack_set(struct tcp_sock *tp, u32 seq, u32 end_seq)
3304 if (tp->rx_opt.sack_ok && sysctl_tcp_dsack) {
3305 if (before(seq, tp->rcv_nxt))
3306 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT);
3308 NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT);
3310 tp->rx_opt.dsack = 1;
3311 tp->duplicate_sack[0].start_seq = seq;
3312 tp->duplicate_sack[0].end_seq = end_seq;
3313 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + 1, 4 - tp->rx_opt.tstamp_ok);
3317 static inline void tcp_dsack_extend(struct tcp_sock *tp, u32 seq, u32 end_seq)
3319 if (!tp->rx_opt.dsack)
3320 tcp_dsack_set(tp, seq, end_seq);
3322 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
3325 static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb)
3327 struct tcp_sock *tp = tcp_sk(sk);
3329 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
3330 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3331 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
3332 tcp_enter_quickack_mode(tp);
3334 if (tp->rx_opt.sack_ok && sysctl_tcp_dsack) {
3335 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3337 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
3338 end_seq = tp->rcv_nxt;
3339 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, end_seq);
3346 /* These routines update the SACK block as out-of-order packets arrive or
3347 * in-order packets close up the sequence space.
3349 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
3352 struct tcp_sack_block *sp = &tp->selective_acks[0];
3353 struct tcp_sack_block *swalk = sp+1;
3355 /* See if the recent change to the first SACK eats into
3356 * or hits the sequence space of other SACK blocks, if so coalesce.
3358 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks; ) {
3359 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
3362 /* Zap SWALK, by moving every further SACK up by one slot.
3363 * Decrease num_sacks.
3365 tp->rx_opt.num_sacks--;
3366 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
3367 for(i=this_sack; i < tp->rx_opt.num_sacks; i++)
3371 this_sack++, swalk++;
3375 static __inline__ void tcp_sack_swap(struct tcp_sack_block *sack1, struct tcp_sack_block *sack2)
3379 tmp = sack1->start_seq;
3380 sack1->start_seq = sack2->start_seq;
3381 sack2->start_seq = tmp;
3383 tmp = sack1->end_seq;
3384 sack1->end_seq = sack2->end_seq;
3385 sack2->end_seq = tmp;
3388 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
3390 struct tcp_sock *tp = tcp_sk(sk);
3391 struct tcp_sack_block *sp = &tp->selective_acks[0];
3392 int cur_sacks = tp->rx_opt.num_sacks;
3398 for (this_sack=0; this_sack<cur_sacks; this_sack++, sp++) {
3399 if (tcp_sack_extend(sp, seq, end_seq)) {
3400 /* Rotate this_sack to the first one. */
3401 for (; this_sack>0; this_sack--, sp--)
3402 tcp_sack_swap(sp, sp-1);
3404 tcp_sack_maybe_coalesce(tp);
3409 /* Could not find an adjacent existing SACK, build a new one,
3410 * put it at the front, and shift everyone else down. We
3411 * always know there is at least one SACK present already here.
3413 * If the sack array is full, forget about the last one.
3415 if (this_sack >= 4) {
3417 tp->rx_opt.num_sacks--;
3420 for(; this_sack > 0; this_sack--, sp--)
3424 /* Build the new head SACK, and we're done. */
3425 sp->start_seq = seq;
3426 sp->end_seq = end_seq;
3427 tp->rx_opt.num_sacks++;
3428 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
3431 /* RCV.NXT advances, some SACKs should be eaten. */
3433 static void tcp_sack_remove(struct tcp_sock *tp)
3435 struct tcp_sack_block *sp = &tp->selective_acks[0];
3436 int num_sacks = tp->rx_opt.num_sacks;
3439 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
3440 if (skb_queue_len(&tp->out_of_order_queue) == 0) {
3441 tp->rx_opt.num_sacks = 0;
3442 tp->rx_opt.eff_sacks = tp->rx_opt.dsack;
3446 for(this_sack = 0; this_sack < num_sacks; ) {
3447 /* Check if the start of the sack is covered by RCV.NXT. */
3448 if (!before(tp->rcv_nxt, sp->start_seq)) {
3451 /* RCV.NXT must cover all the block! */
3452 BUG_TRAP(!before(tp->rcv_nxt, sp->end_seq));
3454 /* Zap this SACK, by moving forward any other SACKS. */
3455 for (i=this_sack+1; i < num_sacks; i++)
3456 tp->selective_acks[i-1] = tp->selective_acks[i];
3463 if (num_sacks != tp->rx_opt.num_sacks) {
3464 tp->rx_opt.num_sacks = num_sacks;
3465 tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
3469 /* This one checks to see if we can put data from the
3470 * out_of_order queue into the receive_queue.
3472 static void tcp_ofo_queue(struct sock *sk)
3474 struct tcp_sock *tp = tcp_sk(sk);
3475 __u32 dsack_high = tp->rcv_nxt;
3476 struct sk_buff *skb;
3478 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
3479 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
3482 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
3483 __u32 dsack = dsack_high;
3484 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
3485 dsack_high = TCP_SKB_CB(skb)->end_seq;
3486 tcp_dsack_extend(tp, TCP_SKB_CB(skb)->seq, dsack);
3489 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3490 SOCK_DEBUG(sk, "ofo packet was already received \n");
3491 __skb_unlink(skb, skb->list);
3495 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
3496 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3497 TCP_SKB_CB(skb)->end_seq);
3499 __skb_unlink(skb, skb->list);
3500 __skb_queue_tail(&sk->sk_receive_queue, skb);
3501 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3503 tcp_fin(skb, sk, skb->h.th);
3507 static int tcp_prune_queue(struct sock *sk);
3509 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
3511 struct tcphdr *th = skb->h.th;
3512 struct tcp_sock *tp = tcp_sk(sk);
3515 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
3519 __skb_pull(skb, th->doff*4);
3521 TCP_ECN_accept_cwr(tp, skb);
3523 if (tp->rx_opt.dsack) {
3524 tp->rx_opt.dsack = 0;
3525 tp->rx_opt.eff_sacks = min_t(unsigned int, tp->rx_opt.num_sacks,
3526 4 - tp->rx_opt.tstamp_ok);
3529 /* Queue data for delivery to the user.
3530 * Packets in sequence go to the receive queue.
3531 * Out of sequence packets to the out_of_order_queue.
3533 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
3534 if (tcp_receive_window(tp) == 0)
3537 /* Ok. In sequence. In window. */
3538 if (tp->ucopy.task == current &&
3539 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
3540 sock_owned_by_user(sk) && !tp->urg_data) {
3541 int chunk = min_t(unsigned int, skb->len,
3544 __set_current_state(TASK_RUNNING);
3547 if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
3548 tp->ucopy.len -= chunk;
3549 tp->copied_seq += chunk;
3550 eaten = (chunk == skb->len && !th->fin);
3551 tcp_rcv_space_adjust(sk);
3559 (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3560 !sk_stream_rmem_schedule(sk, skb))) {
3561 if (tcp_prune_queue(sk) < 0 ||
3562 !sk_stream_rmem_schedule(sk, skb))
3565 sk_stream_set_owner_r(skb, sk);
3566 __skb_queue_tail(&sk->sk_receive_queue, skb);
3568 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3570 tcp_event_data_recv(sk, tp, skb);
3572 tcp_fin(skb, sk, th);
3574 if (skb_queue_len(&tp->out_of_order_queue)) {
3577 /* RFC2581. 4.2. SHOULD send immediate ACK, when
3578 * gap in queue is filled.
3580 if (!skb_queue_len(&tp->out_of_order_queue))
3581 tp->ack.pingpong = 0;
3584 if (tp->rx_opt.num_sacks)
3585 tcp_sack_remove(tp);
3587 tcp_fast_path_check(sk, tp);
3591 else if (!sock_flag(sk, SOCK_DEAD))
3592 sk->sk_data_ready(sk, 0);
3596 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3597 /* A retransmit, 2nd most common case. Force an immediate ack. */
3598 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
3599 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3602 tcp_enter_quickack_mode(tp);
3603 tcp_schedule_ack(tp);
3609 /* Out of window. F.e. zero window probe. */
3610 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
3613 tcp_enter_quickack_mode(tp);
3615 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3616 /* Partial packet, seq < rcv_next < end_seq */
3617 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
3618 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3619 TCP_SKB_CB(skb)->end_seq);
3621 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
3623 /* If window is closed, drop tail of packet. But after
3624 * remembering D-SACK for its head made in previous line.
3626 if (!tcp_receive_window(tp))
3631 TCP_ECN_check_ce(tp, skb);
3633 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3634 !sk_stream_rmem_schedule(sk, skb)) {
3635 if (tcp_prune_queue(sk) < 0 ||
3636 !sk_stream_rmem_schedule(sk, skb))
3640 /* Disable header prediction. */
3642 tcp_schedule_ack(tp);
3644 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
3645 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3647 sk_stream_set_owner_r(skb, sk);
3649 if (!skb_peek(&tp->out_of_order_queue)) {
3650 /* Initial out of order segment, build 1 SACK. */
3651 if (tp->rx_opt.sack_ok) {
3652 tp->rx_opt.num_sacks = 1;
3653 tp->rx_opt.dsack = 0;
3654 tp->rx_opt.eff_sacks = 1;
3655 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
3656 tp->selective_acks[0].end_seq =
3657 TCP_SKB_CB(skb)->end_seq;
3659 __skb_queue_head(&tp->out_of_order_queue,skb);
3661 struct sk_buff *skb1 = tp->out_of_order_queue.prev;
3662 u32 seq = TCP_SKB_CB(skb)->seq;
3663 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3665 if (seq == TCP_SKB_CB(skb1)->end_seq) {
3666 __skb_append(skb1, skb);
3668 if (!tp->rx_opt.num_sacks ||
3669 tp->selective_acks[0].end_seq != seq)
3672 /* Common case: data arrive in order after hole. */
3673 tp->selective_acks[0].end_seq = end_seq;
3677 /* Find place to insert this segment. */
3679 if (!after(TCP_SKB_CB(skb1)->seq, seq))
3681 } while ((skb1 = skb1->prev) !=
3682 (struct sk_buff*)&tp->out_of_order_queue);
3684 /* Do skb overlap to previous one? */
3685 if (skb1 != (struct sk_buff*)&tp->out_of_order_queue &&
3686 before(seq, TCP_SKB_CB(skb1)->end_seq)) {
3687 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
3688 /* All the bits are present. Drop. */
3690 tcp_dsack_set(tp, seq, end_seq);
3693 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
3694 /* Partial overlap. */
3695 tcp_dsack_set(tp, seq, TCP_SKB_CB(skb1)->end_seq);
3700 __skb_insert(skb, skb1, skb1->next, &tp->out_of_order_queue);
3702 /* And clean segments covered by new one as whole. */
3703 while ((skb1 = skb->next) !=
3704 (struct sk_buff*)&tp->out_of_order_queue &&
3705 after(end_seq, TCP_SKB_CB(skb1)->seq)) {
3706 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
3707 tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, end_seq);
3710 __skb_unlink(skb1, skb1->list);
3711 tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, TCP_SKB_CB(skb1)->end_seq);
3716 if (tp->rx_opt.sack_ok)
3717 tcp_sack_new_ofo_skb(sk, seq, end_seq);
3721 /* Collapse contiguous sequence of skbs head..tail with
3722 * sequence numbers start..end.
3723 * Segments with FIN/SYN are not collapsed (only because this
3727 tcp_collapse(struct sock *sk, struct sk_buff *head,
3728 struct sk_buff *tail, u32 start, u32 end)
3730 struct sk_buff *skb;
3732 /* First, check that queue is collapsable and find
3733 * the point where collapsing can be useful. */
3734 for (skb = head; skb != tail; ) {
3735 /* No new bits? It is possible on ofo queue. */
3736 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
3737 struct sk_buff *next = skb->next;
3738 __skb_unlink(skb, skb->list);
3740 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
3745 /* The first skb to collapse is:
3747 * - bloated or contains data before "start" or
3748 * overlaps to the next one.
3750 if (!skb->h.th->syn && !skb->h.th->fin &&
3751 (tcp_win_from_space(skb->truesize) > skb->len ||
3752 before(TCP_SKB_CB(skb)->seq, start) ||
3753 (skb->next != tail &&
3754 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq)))
3757 /* Decided to skip this, advance start seq. */
3758 start = TCP_SKB_CB(skb)->end_seq;
3761 if (skb == tail || skb->h.th->syn || skb->h.th->fin)
3764 while (before(start, end)) {
3765 struct sk_buff *nskb;
3766 int header = skb_headroom(skb);
3767 int copy = SKB_MAX_ORDER(header, 0);
3769 /* Too big header? This can happen with IPv6. */
3772 if (end-start < copy)
3774 nskb = alloc_skb(copy+header, GFP_ATOMIC);
3777 skb_reserve(nskb, header);
3778 memcpy(nskb->head, skb->head, header);
3779 nskb->nh.raw = nskb->head + (skb->nh.raw-skb->head);
3780 nskb->h.raw = nskb->head + (skb->h.raw-skb->head);
3781 nskb->mac.raw = nskb->head + (skb->mac.raw-skb->head);
3782 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
3783 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
3784 __skb_insert(nskb, skb->prev, skb, skb->list);
3785 sk_stream_set_owner_r(nskb, sk);
3787 /* Copy data, releasing collapsed skbs. */
3789 int offset = start - TCP_SKB_CB(skb)->seq;
3790 int size = TCP_SKB_CB(skb)->end_seq - start;
3792 if (offset < 0) BUG();
3794 size = min(copy, size);
3795 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
3797 TCP_SKB_CB(nskb)->end_seq += size;
3801 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
3802 struct sk_buff *next = skb->next;
3803 __skb_unlink(skb, skb->list);
3805 NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
3807 if (skb == tail || skb->h.th->syn || skb->h.th->fin)
3814 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
3815 * and tcp_collapse() them until all the queue is collapsed.
3817 static void tcp_collapse_ofo_queue(struct sock *sk)
3819 struct tcp_sock *tp = tcp_sk(sk);
3820 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
3821 struct sk_buff *head;
3827 start = TCP_SKB_CB(skb)->seq;
3828 end = TCP_SKB_CB(skb)->end_seq;
3834 /* Segment is terminated when we see gap or when
3835 * we are at the end of all the queue. */
3836 if (skb == (struct sk_buff *)&tp->out_of_order_queue ||
3837 after(TCP_SKB_CB(skb)->seq, end) ||
3838 before(TCP_SKB_CB(skb)->end_seq, start)) {
3839 tcp_collapse(sk, head, skb, start, end);
3841 if (skb == (struct sk_buff *)&tp->out_of_order_queue)
3843 /* Start new segment */
3844 start = TCP_SKB_CB(skb)->seq;
3845 end = TCP_SKB_CB(skb)->end_seq;
3847 if (before(TCP_SKB_CB(skb)->seq, start))
3848 start = TCP_SKB_CB(skb)->seq;
3849 if (after(TCP_SKB_CB(skb)->end_seq, end))
3850 end = TCP_SKB_CB(skb)->end_seq;
3855 /* Reduce allocated memory if we can, trying to get
3856 * the socket within its memory limits again.
3858 * Return less than zero if we should start dropping frames
3859 * until the socket owning process reads some of the data
3860 * to stabilize the situation.
3862 static int tcp_prune_queue(struct sock *sk)
3864 struct tcp_sock *tp = tcp_sk(sk);
3866 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
3868 NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED);
3870 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
3871 tcp_clamp_window(sk, tp);
3872 else if (tcp_memory_pressure)
3873 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
3875 tcp_collapse_ofo_queue(sk);
3876 tcp_collapse(sk, sk->sk_receive_queue.next,
3877 (struct sk_buff*)&sk->sk_receive_queue,
3878 tp->copied_seq, tp->rcv_nxt);
3879 sk_stream_mem_reclaim(sk);
3881 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
3884 /* Collapsing did not help, destructive actions follow.
3885 * This must not ever occur. */
3887 /* First, purge the out_of_order queue. */
3888 if (skb_queue_len(&tp->out_of_order_queue)) {
3889 NET_ADD_STATS_BH(LINUX_MIB_OFOPRUNED,
3890 skb_queue_len(&tp->out_of_order_queue));
3891 __skb_queue_purge(&tp->out_of_order_queue);
3893 /* Reset SACK state. A conforming SACK implementation will
3894 * do the same at a timeout based retransmit. When a connection
3895 * is in a sad state like this, we care only about integrity
3896 * of the connection not performance.
3898 if (tp->rx_opt.sack_ok)
3899 tcp_sack_reset(&tp->rx_opt);
3900 sk_stream_mem_reclaim(sk);
3903 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
3906 /* If we are really being abused, tell the caller to silently
3907 * drop receive data on the floor. It will get retransmitted
3908 * and hopefully then we'll have sufficient space.
3910 NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED);
3912 /* Massive buffer overcommit. */
3918 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
3919 * As additional protections, we do not touch cwnd in retransmission phases,
3920 * and if application hit its sndbuf limit recently.
3922 void tcp_cwnd_application_limited(struct sock *sk)
3924 struct tcp_sock *tp = tcp_sk(sk);
3926 if (tp->ca_state == TCP_CA_Open &&
3927 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
3928 /* Limited by application or receiver window. */
3929 u32 win_used = max(tp->snd_cwnd_used, 2U);
3930 if (win_used < tp->snd_cwnd) {
3931 tp->snd_ssthresh = tcp_current_ssthresh(tp);
3932 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
3934 tp->snd_cwnd_used = 0;
3936 tp->snd_cwnd_stamp = tcp_time_stamp;
3940 /* When incoming ACK allowed to free some skb from write_queue,
3941 * we remember this event in flag sk->sk_queue_shrunk and wake up socket
3942 * on the exit from tcp input handler.
3944 * PROBLEM: sndbuf expansion does not work well with largesend.
3946 static void tcp_new_space(struct sock *sk)
3948 struct tcp_sock *tp = tcp_sk(sk);
3950 if (tp->packets_out < tp->snd_cwnd &&
3951 !(sk->sk_userlocks & SOCK_SNDBUF_LOCK) &&
3952 !tcp_memory_pressure &&
3953 atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) {
3954 int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache_std) +
3955 MAX_TCP_HEADER + 16 + sizeof(struct sk_buff),
3956 demanded = max_t(unsigned int, tp->snd_cwnd,
3957 tp->reordering + 1);
3958 sndmem *= 2*demanded;
3959 if (sndmem > sk->sk_sndbuf)
3960 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
3961 tp->snd_cwnd_stamp = tcp_time_stamp;
3964 sk->sk_write_space(sk);
3967 static inline void tcp_check_space(struct sock *sk)
3969 if (sk->sk_queue_shrunk) {
3970 sk->sk_queue_shrunk = 0;
3971 if (sk->sk_socket &&
3972 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
3977 static void __tcp_data_snd_check(struct sock *sk, struct sk_buff *skb)
3979 struct tcp_sock *tp = tcp_sk(sk);
3981 if (after(TCP_SKB_CB(skb)->end_seq, tp->snd_una + tp->snd_wnd) ||
3982 tcp_packets_in_flight(tp) >= tp->snd_cwnd ||
3983 tcp_write_xmit(sk, tp->nonagle))
3984 tcp_check_probe_timer(sk, tp);
3987 static __inline__ void tcp_data_snd_check(struct sock *sk)
3989 struct sk_buff *skb = sk->sk_send_head;
3992 __tcp_data_snd_check(sk, skb);
3993 tcp_check_space(sk);
3997 * Check if sending an ack is needed.
3999 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
4001 struct tcp_sock *tp = tcp_sk(sk);
4003 /* More than one full frame received... */
4004 if (((tp->rcv_nxt - tp->rcv_wup) > tp->ack.rcv_mss
4005 /* ... and right edge of window advances far enough.
4006 * (tcp_recvmsg() will send ACK otherwise). Or...
4008 && __tcp_select_window(sk) >= tp->rcv_wnd) ||
4009 /* We ACK each frame or... */
4010 tcp_in_quickack_mode(tp) ||
4011 /* We have out of order data. */
4013 skb_peek(&tp->out_of_order_queue))) {
4014 /* Then ack it now */
4017 /* Else, send delayed ack. */
4018 tcp_send_delayed_ack(sk);
4022 static __inline__ void tcp_ack_snd_check(struct sock *sk)
4024 struct tcp_sock *tp = tcp_sk(sk);
4025 if (!tcp_ack_scheduled(tp)) {
4026 /* We sent a data segment already. */
4029 __tcp_ack_snd_check(sk, 1);
4033 * This routine is only called when we have urgent data
4034 * signalled. Its the 'slow' part of tcp_urg. It could be
4035 * moved inline now as tcp_urg is only called from one
4036 * place. We handle URGent data wrong. We have to - as
4037 * BSD still doesn't use the correction from RFC961.
4038 * For 1003.1g we should support a new option TCP_STDURG to permit
4039 * either form (or just set the sysctl tcp_stdurg).
4042 static void tcp_check_urg(struct sock * sk, struct tcphdr * th)
4044 struct tcp_sock *tp = tcp_sk(sk);
4045 u32 ptr = ntohs(th->urg_ptr);
4047 if (ptr && !sysctl_tcp_stdurg)
4049 ptr += ntohl(th->seq);
4051 /* Ignore urgent data that we've already seen and read. */
4052 if (after(tp->copied_seq, ptr))
4055 /* Do not replay urg ptr.
4057 * NOTE: interesting situation not covered by specs.
4058 * Misbehaving sender may send urg ptr, pointing to segment,
4059 * which we already have in ofo queue. We are not able to fetch
4060 * such data and will stay in TCP_URG_NOTYET until will be eaten
4061 * by recvmsg(). Seems, we are not obliged to handle such wicked
4062 * situations. But it is worth to think about possibility of some
4063 * DoSes using some hypothetical application level deadlock.
4065 if (before(ptr, tp->rcv_nxt))
4068 /* Do we already have a newer (or duplicate) urgent pointer? */
4069 if (tp->urg_data && !after(ptr, tp->urg_seq))
4072 /* Tell the world about our new urgent pointer. */
4075 /* We may be adding urgent data when the last byte read was
4076 * urgent. To do this requires some care. We cannot just ignore
4077 * tp->copied_seq since we would read the last urgent byte again
4078 * as data, nor can we alter copied_seq until this data arrives
4079 * or we break the sematics of SIOCATMARK (and thus sockatmark())
4081 * NOTE. Double Dutch. Rendering to plain English: author of comment
4082 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
4083 * and expect that both A and B disappear from stream. This is _wrong_.
4084 * Though this happens in BSD with high probability, this is occasional.
4085 * Any application relying on this is buggy. Note also, that fix "works"
4086 * only in this artificial test. Insert some normal data between A and B and we will
4087 * decline of BSD again. Verdict: it is better to remove to trap
4090 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
4091 !sock_flag(sk, SOCK_URGINLINE) &&
4092 tp->copied_seq != tp->rcv_nxt) {
4093 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
4095 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
4096 __skb_unlink(skb, skb->list);
4101 tp->urg_data = TCP_URG_NOTYET;
4104 /* Disable header prediction. */
4108 /* This is the 'fast' part of urgent handling. */
4109 static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th)
4111 struct tcp_sock *tp = tcp_sk(sk);
4113 /* Check if we get a new urgent pointer - normally not. */
4115 tcp_check_urg(sk,th);
4117 /* Do we wait for any urgent data? - normally not... */
4118 if (tp->urg_data == TCP_URG_NOTYET) {
4119 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
4122 /* Is the urgent pointer pointing into this packet? */
4123 if (ptr < skb->len) {
4125 if (skb_copy_bits(skb, ptr, &tmp, 1))
4127 tp->urg_data = TCP_URG_VALID | tmp;
4128 if (!sock_flag(sk, SOCK_DEAD))
4129 sk->sk_data_ready(sk, 0);
4134 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
4136 struct tcp_sock *tp = tcp_sk(sk);
4137 int chunk = skb->len - hlen;
4141 if (skb->ip_summed==CHECKSUM_UNNECESSARY)
4142 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
4144 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
4148 tp->ucopy.len -= chunk;
4149 tp->copied_seq += chunk;
4150 tcp_rcv_space_adjust(sk);
4157 static int __tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
4161 if (sock_owned_by_user(sk)) {
4163 result = __tcp_checksum_complete(skb);
4166 result = __tcp_checksum_complete(skb);
4171 static __inline__ int
4172 tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
4174 return skb->ip_summed != CHECKSUM_UNNECESSARY &&
4175 __tcp_checksum_complete_user(sk, skb);
4179 * TCP receive function for the ESTABLISHED state.
4181 * It is split into a fast path and a slow path. The fast path is
4183 * - A zero window was announced from us - zero window probing
4184 * is only handled properly in the slow path.
4185 * - Out of order segments arrived.
4186 * - Urgent data is expected.
4187 * - There is no buffer space left
4188 * - Unexpected TCP flags/window values/header lengths are received
4189 * (detected by checking the TCP header against pred_flags)
4190 * - Data is sent in both directions. Fast path only supports pure senders
4191 * or pure receivers (this means either the sequence number or the ack
4192 * value must stay constant)
4193 * - Unexpected TCP option.
4195 * When these conditions are not satisfied it drops into a standard
4196 * receive procedure patterned after RFC793 to handle all cases.
4197 * The first three cases are guaranteed by proper pred_flags setting,
4198 * the rest is checked inline. Fast processing is turned on in
4199 * tcp_data_queue when everything is OK.
4201 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
4202 struct tcphdr *th, unsigned len)
4204 struct tcp_sock *tp = tcp_sk(sk);
4207 * Header prediction.
4208 * The code loosely follows the one in the famous
4209 * "30 instruction TCP receive" Van Jacobson mail.
4211 * Van's trick is to deposit buffers into socket queue
4212 * on a device interrupt, to call tcp_recv function
4213 * on the receive process context and checksum and copy
4214 * the buffer to user space. smart...
4216 * Our current scheme is not silly either but we take the
4217 * extra cost of the net_bh soft interrupt processing...
4218 * We do checksum and copy also but from device to kernel.
4221 tp->rx_opt.saw_tstamp = 0;
4223 /* pred_flags is 0xS?10 << 16 + snd_wnd
4224 * if header_predition is to be made
4225 * 'S' will always be tp->tcp_header_len >> 2
4226 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
4227 * turn it off (when there are holes in the receive
4228 * space for instance)
4229 * PSH flag is ignored.
4232 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
4233 TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4234 int tcp_header_len = tp->tcp_header_len;
4236 /* Timestamp header prediction: tcp_header_len
4237 * is automatically equal to th->doff*4 due to pred_flags
4241 /* Check timestamp */
4242 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
4243 __u32 *ptr = (__u32 *)(th + 1);
4245 /* No? Slow path! */
4246 if (*ptr != ntohl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
4247 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP))
4250 tp->rx_opt.saw_tstamp = 1;
4252 tp->rx_opt.rcv_tsval = ntohl(*ptr);
4254 tp->rx_opt.rcv_tsecr = ntohl(*ptr);
4256 /* If PAWS failed, check it more carefully in slow path */
4257 if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
4260 /* DO NOT update ts_recent here, if checksum fails
4261 * and timestamp was corrupted part, it will result
4262 * in a hung connection since we will drop all
4263 * future packets due to the PAWS test.
4267 if (len <= tcp_header_len) {
4268 /* Bulk data transfer: sender */
4269 if (len == tcp_header_len) {
4270 /* Predicted packet is in window by definition.
4271 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4272 * Hence, check seq<=rcv_wup reduces to:
4274 if (tcp_header_len ==
4275 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
4276 tp->rcv_nxt == tp->rcv_wup)
4277 tcp_store_ts_recent(tp);
4279 tcp_rcv_rtt_measure_ts(tp, skb);
4281 /* We know that such packets are checksummed
4284 tcp_ack(sk, skb, 0);
4286 tcp_data_snd_check(sk);
4288 } else { /* Header too small */
4289 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4295 if (tp->ucopy.task == current &&
4296 tp->copied_seq == tp->rcv_nxt &&
4297 len - tcp_header_len <= tp->ucopy.len &&
4298 sock_owned_by_user(sk)) {
4299 __set_current_state(TASK_RUNNING);
4301 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len)) {
4302 /* Predicted packet is in window by definition.
4303 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4304 * Hence, check seq<=rcv_wup reduces to:
4306 if (tcp_header_len ==
4307 (sizeof(struct tcphdr) +
4308 TCPOLEN_TSTAMP_ALIGNED) &&
4309 tp->rcv_nxt == tp->rcv_wup)
4310 tcp_store_ts_recent(tp);
4312 tcp_rcv_rtt_measure_ts(tp, skb);
4314 __skb_pull(skb, tcp_header_len);
4315 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4316 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER);
4321 if (tcp_checksum_complete_user(sk, skb))
4324 /* Predicted packet is in window by definition.
4325 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4326 * Hence, check seq<=rcv_wup reduces to:
4328 if (tcp_header_len ==
4329 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
4330 tp->rcv_nxt == tp->rcv_wup)
4331 tcp_store_ts_recent(tp);
4333 tcp_rcv_rtt_measure_ts(tp, skb);
4335 if ((int)skb->truesize > sk->sk_forward_alloc)
4338 NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS);
4340 /* Bulk data transfer: receiver */
4341 __skb_pull(skb,tcp_header_len);
4342 __skb_queue_tail(&sk->sk_receive_queue, skb);
4343 sk_stream_set_owner_r(skb, sk);
4344 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4347 tcp_event_data_recv(sk, tp, skb);
4349 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
4350 /* Well, only one small jumplet in fast path... */
4351 tcp_ack(sk, skb, FLAG_DATA);
4352 tcp_data_snd_check(sk);
4353 if (!tcp_ack_scheduled(tp))
4358 if (tcp_in_quickack_mode(tp)) {
4361 tcp_send_delayed_ack(sk);
4364 __tcp_ack_snd_check(sk, 0);
4371 sk->sk_data_ready(sk, 0);
4377 if (len < (th->doff<<2) || tcp_checksum_complete_user(sk, skb))
4381 * RFC1323: H1. Apply PAWS check first.
4383 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
4384 tcp_paws_discard(tp, skb)) {
4386 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
4387 tcp_send_dupack(sk, skb);
4390 /* Resets are accepted even if PAWS failed.
4392 ts_recent update must be made after we are sure
4393 that the packet is in window.
4398 * Standard slow path.
4401 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
4402 /* RFC793, page 37: "In all states except SYN-SENT, all reset
4403 * (RST) segments are validated by checking their SEQ-fields."
4404 * And page 69: "If an incoming segment is not acceptable,
4405 * an acknowledgment should be sent in reply (unless the RST bit
4406 * is set, if so drop the segment and return)".
4409 tcp_send_dupack(sk, skb);
4418 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
4420 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4421 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4422 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
4429 tcp_ack(sk, skb, FLAG_SLOWPATH);
4431 tcp_rcv_rtt_measure_ts(tp, skb);
4433 /* Process urgent data. */
4434 tcp_urg(sk, skb, th);
4436 /* step 7: process the segment text */
4437 tcp_data_queue(sk, skb);
4439 tcp_data_snd_check(sk);
4440 tcp_ack_snd_check(sk);
4444 TCP_INC_STATS_BH(TCP_MIB_INERRS);
4451 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
4452 struct tcphdr *th, unsigned len)
4454 struct tcp_sock *tp = tcp_sk(sk);
4455 int saved_clamp = tp->rx_opt.mss_clamp;
4457 tcp_parse_options(skb, &tp->rx_opt, 0);
4461 * "If the state is SYN-SENT then
4462 * first check the ACK bit
4463 * If the ACK bit is set
4464 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
4465 * a reset (unless the RST bit is set, if so drop
4466 * the segment and return)"
4468 * We do not send data with SYN, so that RFC-correct
4471 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
4472 goto reset_and_undo;
4474 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
4475 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
4477 NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED);
4478 goto reset_and_undo;
4481 /* Now ACK is acceptable.
4483 * "If the RST bit is set
4484 * If the ACK was acceptable then signal the user "error:
4485 * connection reset", drop the segment, enter CLOSED state,
4486 * delete TCB, and return."
4495 * "fifth, if neither of the SYN or RST bits is set then
4496 * drop the segment and return."
4502 goto discard_and_undo;
4505 * "If the SYN bit is on ...
4506 * are acceptable then ...
4507 * (our SYN has been ACKed), change the connection
4508 * state to ESTABLISHED..."
4511 TCP_ECN_rcv_synack(tp, th);
4512 if (tp->ecn_flags&TCP_ECN_OK)
4513 sk->sk_no_largesend = 1;
4515 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
4516 tcp_ack(sk, skb, FLAG_SLOWPATH);
4518 /* Ok.. it's good. Set up sequence numbers and
4519 * move to established.
4521 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
4522 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
4524 /* RFC1323: The window in SYN & SYN/ACK segments is
4527 tp->snd_wnd = ntohs(th->window);
4528 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(skb)->seq);
4530 if (!tp->rx_opt.wscale_ok) {
4531 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
4532 tp->window_clamp = min(tp->window_clamp, 65535U);
4535 if (tp->rx_opt.saw_tstamp) {
4536 tp->rx_opt.tstamp_ok = 1;
4537 tp->tcp_header_len =
4538 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
4539 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
4540 tcp_store_ts_recent(tp);
4542 tp->tcp_header_len = sizeof(struct tcphdr);
4545 if (tp->rx_opt.sack_ok && sysctl_tcp_fack)
4546 tp->rx_opt.sack_ok |= 2;
4548 tcp_sync_mss(sk, tp->pmtu_cookie);
4549 tcp_initialize_rcv_mss(sk);
4551 /* Remember, tcp_poll() does not lock socket!
4552 * Change state from SYN-SENT only after copied_seq
4553 * is initialized. */
4554 tp->copied_seq = tp->rcv_nxt;
4556 tcp_set_state(sk, TCP_ESTABLISHED);
4558 /* Make sure socket is routed, for correct metrics. */
4559 tp->af_specific->rebuild_header(sk);
4561 tcp_init_metrics(sk);
4563 /* Prevent spurious tcp_cwnd_restart() on first data
4566 tp->lsndtime = tcp_time_stamp;
4568 tcp_init_buffer_space(sk);
4570 if (sock_flag(sk, SOCK_KEEPOPEN))
4571 tcp_reset_keepalive_timer(sk, keepalive_time_when(tp));
4573 if (!tp->rx_opt.snd_wscale)
4574 __tcp_fast_path_on(tp, tp->snd_wnd);
4578 if (!sock_flag(sk, SOCK_DEAD)) {
4579 sk->sk_state_change(sk);
4580 sk_wake_async(sk, 0, POLL_OUT);
4583 if (sk->sk_write_pending || tp->defer_accept || tp->ack.pingpong) {
4584 /* Save one ACK. Data will be ready after
4585 * several ticks, if write_pending is set.
4587 * It may be deleted, but with this feature tcpdumps
4588 * look so _wonderfully_ clever, that I was not able
4589 * to stand against the temptation 8) --ANK
4591 tcp_schedule_ack(tp);
4592 tp->ack.lrcvtime = tcp_time_stamp;
4593 tp->ack.ato = TCP_ATO_MIN;
4594 tcp_incr_quickack(tp);
4595 tcp_enter_quickack_mode(tp);
4596 tcp_reset_xmit_timer(sk, TCP_TIME_DACK, TCP_DELACK_MAX);
4607 /* No ACK in the segment */
4611 * "If the RST bit is set
4613 * Otherwise (no ACK) drop the segment and return."
4616 goto discard_and_undo;
4620 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && tcp_paws_check(&tp->rx_opt, 0))
4621 goto discard_and_undo;
4624 /* We see SYN without ACK. It is attempt of
4625 * simultaneous connect with crossed SYNs.
4626 * Particularly, it can be connect to self.
4628 tcp_set_state(sk, TCP_SYN_RECV);
4630 if (tp->rx_opt.saw_tstamp) {
4631 tp->rx_opt.tstamp_ok = 1;
4632 tcp_store_ts_recent(tp);
4633 tp->tcp_header_len =
4634 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
4636 tp->tcp_header_len = sizeof(struct tcphdr);
4639 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
4640 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
4642 /* RFC1323: The window in SYN & SYN/ACK segments is
4645 tp->snd_wnd = ntohs(th->window);
4646 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
4647 tp->max_window = tp->snd_wnd;
4649 TCP_ECN_rcv_syn(tp, th);
4650 if (tp->ecn_flags&TCP_ECN_OK)
4651 sk->sk_no_largesend = 1;
4653 tcp_sync_mss(sk, tp->pmtu_cookie);
4654 tcp_initialize_rcv_mss(sk);
4657 tcp_send_synack(sk);
4659 /* Note, we could accept data and URG from this segment.
4660 * There are no obstacles to make this.
4662 * However, if we ignore data in ACKless segments sometimes,
4663 * we have no reasons to accept it sometimes.
4664 * Also, seems the code doing it in step6 of tcp_rcv_state_process
4665 * is not flawless. So, discard packet for sanity.
4666 * Uncomment this return to process the data.
4673 /* "fifth, if neither of the SYN or RST bits is set then
4674 * drop the segment and return."
4678 tcp_clear_options(&tp->rx_opt);
4679 tp->rx_opt.mss_clamp = saved_clamp;
4683 tcp_clear_options(&tp->rx_opt);
4684 tp->rx_opt.mss_clamp = saved_clamp;
4690 * This function implements the receiving procedure of RFC 793 for
4691 * all states except ESTABLISHED and TIME_WAIT.
4692 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
4693 * address independent.
4696 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
4697 struct tcphdr *th, unsigned len)
4699 struct tcp_sock *tp = tcp_sk(sk);
4702 tp->rx_opt.saw_tstamp = 0;
4704 switch (sk->sk_state) {
4716 if(tp->af_specific->conn_request(sk, skb) < 0)
4722 /* Now we have several options: In theory there is
4723 * nothing else in the frame. KA9Q has an option to
4724 * send data with the syn, BSD accepts data with the
4725 * syn up to the [to be] advertised window and
4726 * Solaris 2.1 gives you a protocol error. For now
4727 * we just ignore it, that fits the spec precisely
4728 * and avoids incompatibilities. It would be nice in
4729 * future to drop through and process the data.
4731 * Now that TTCP is starting to be used we ought to
4733 * But, this leaves one open to an easy denial of
4734 * service attack, and SYN cookies can't defend
4735 * against this problem. So, we drop the data
4736 * in the interest of security over speed.
4746 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
4750 /* Do step6 onward by hand. */
4751 tcp_urg(sk, skb, th);
4753 tcp_data_snd_check(sk);
4757 if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
4758 tcp_paws_discard(tp, skb)) {
4760 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
4761 tcp_send_dupack(sk, skb);
4764 /* Reset is accepted even if it did not pass PAWS. */
4767 /* step 1: check sequence number */
4768 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
4770 tcp_send_dupack(sk, skb);
4774 /* step 2: check RST bit */
4780 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
4782 /* step 3: check security and precedence [ignored] */
4786 * Check for a SYN in window.
4788 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4789 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
4794 /* step 5: check the ACK field */
4796 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH);
4798 switch(sk->sk_state) {
4801 tp->copied_seq = tp->rcv_nxt;
4803 tcp_set_state(sk, TCP_ESTABLISHED);
4804 sk->sk_state_change(sk);
4806 /* Note, that this wakeup is only for marginal
4807 * crossed SYN case. Passively open sockets
4808 * are not waked up, because sk->sk_sleep ==
4809 * NULL and sk->sk_socket == NULL.
4811 if (sk->sk_socket) {
4812 sk_wake_async(sk,0,POLL_OUT);
4815 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
4816 tp->snd_wnd = ntohs(th->window) <<
4817 tp->rx_opt.snd_wscale;
4818 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq,
4819 TCP_SKB_CB(skb)->seq);
4821 /* tcp_ack considers this ACK as duplicate
4822 * and does not calculate rtt.
4823 * Fix it at least with timestamps.
4825 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
4827 tcp_ack_saw_tstamp(tp, 0);
4829 if (tp->rx_opt.tstamp_ok)
4830 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
4832 /* Make sure socket is routed, for
4835 tp->af_specific->rebuild_header(sk);
4837 tcp_init_metrics(sk);
4839 /* Prevent spurious tcp_cwnd_restart() on
4840 * first data packet.
4842 tp->lsndtime = tcp_time_stamp;
4844 tcp_initialize_rcv_mss(sk);
4845 tcp_init_buffer_space(sk);
4846 tcp_fast_path_on(tp);
4853 if (tp->snd_una == tp->write_seq) {
4854 tcp_set_state(sk, TCP_FIN_WAIT2);
4855 sk->sk_shutdown |= SEND_SHUTDOWN;
4856 dst_confirm(sk->sk_dst_cache);
4858 if (!sock_flag(sk, SOCK_DEAD))
4859 /* Wake up lingering close() */
4860 sk->sk_state_change(sk);
4864 if (tp->linger2 < 0 ||
4865 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4866 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
4868 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
4872 tmo = tcp_fin_time(tp);
4873 if (tmo > TCP_TIMEWAIT_LEN) {
4874 tcp_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
4875 } else if (th->fin || sock_owned_by_user(sk)) {
4876 /* Bad case. We could lose such FIN otherwise.
4877 * It is not a big problem, but it looks confusing
4878 * and not so rare event. We still can lose it now,
4879 * if it spins in bh_lock_sock(), but it is really
4882 tcp_reset_keepalive_timer(sk, tmo);
4884 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
4892 if (tp->snd_una == tp->write_seq) {
4893 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4899 if (tp->snd_una == tp->write_seq) {
4900 tcp_update_metrics(sk);
4909 /* step 6: check the URG bit */
4910 tcp_urg(sk, skb, th);
4912 /* step 7: process the segment text */
4913 switch (sk->sk_state) {
4914 case TCP_CLOSE_WAIT:
4917 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4921 /* RFC 793 says to queue data in these states,
4922 * RFC 1122 says we MUST send a reset.
4923 * BSD 4.4 also does reset.
4925 if (sk->sk_shutdown & RCV_SHUTDOWN) {
4926 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4927 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
4928 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
4934 case TCP_ESTABLISHED:
4935 tcp_data_queue(sk, skb);
4940 /* tcp_data could move socket to TIME-WAIT */
4941 if (sk->sk_state != TCP_CLOSE) {
4942 tcp_data_snd_check(sk);
4943 tcp_ack_snd_check(sk);
4953 EXPORT_SYMBOL(sysctl_tcp_ecn);
4954 EXPORT_SYMBOL(sysctl_tcp_reordering);
4955 EXPORT_SYMBOL(tcp_parse_options);
4956 EXPORT_SYMBOL(tcp_rcv_established);
4957 EXPORT_SYMBOL(tcp_rcv_state_process);