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 /* Default values of the Vegas variables, in fixed-point representation
94 * with V_PARAM_SHIFT bits to the right of the binary point.
96 #define V_PARAM_SHIFT 1
97 int sysctl_tcp_vegas_alpha = 1<<V_PARAM_SHIFT;
98 int sysctl_tcp_vegas_beta = 3<<V_PARAM_SHIFT;
99 int sysctl_tcp_vegas_gamma = 1<<V_PARAM_SHIFT;
101 int sysctl_tcp_bic_fast_convergence = 1;
102 int sysctl_tcp_bic_low_window = 14;
104 #define FLAG_DATA 0x01 /* Incoming frame contained data. */
105 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
106 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
107 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
108 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
109 #define FLAG_DATA_SACKED 0x20 /* New SACK. */
110 #define FLAG_ECE 0x40 /* ECE in this ACK */
111 #define FLAG_DATA_LOST 0x80 /* SACK detected data lossage. */
112 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
114 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
115 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
116 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
117 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
119 #define IsReno(tp) ((tp)->sack_ok == 0)
120 #define IsFack(tp) ((tp)->sack_ok & 2)
121 #define IsDSack(tp) ((tp)->sack_ok & 4)
123 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
125 /* Adapt the MSS value used to make delayed ack decision to the
128 static __inline__ void tcp_measure_rcv_mss(struct tcp_opt *tp, struct sk_buff *skb)
130 unsigned int len, lss;
132 lss = tp->ack.last_seg_size;
133 tp->ack.last_seg_size = 0;
135 /* skb->len may jitter because of SACKs, even if peer
136 * sends good full-sized frames.
139 if (len >= tp->ack.rcv_mss) {
140 tp->ack.rcv_mss = len;
142 /* Otherwise, we make more careful check taking into account,
143 * that SACKs block is variable.
145 * "len" is invariant segment length, including TCP header.
147 len += skb->data - skb->h.raw;
148 if (len >= TCP_MIN_RCVMSS + sizeof(struct tcphdr) ||
149 /* If PSH is not set, packet should be
150 * full sized, provided peer TCP is not badly broken.
151 * This observation (if it is correct 8)) allows
152 * to handle super-low mtu links fairly.
154 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
155 !(tcp_flag_word(skb->h.th)&TCP_REMNANT))) {
156 /* Subtract also invariant (if peer is RFC compliant),
157 * tcp header plus fixed timestamp option length.
158 * Resulting "len" is MSS free of SACK jitter.
160 len -= tp->tcp_header_len;
161 tp->ack.last_seg_size = len;
163 tp->ack.rcv_mss = len;
167 tp->ack.pending |= TCP_ACK_PUSHED;
171 static void tcp_incr_quickack(struct tcp_opt *tp)
173 unsigned quickacks = tp->rcv_wnd/(2*tp->ack.rcv_mss);
177 if (quickacks > tp->ack.quick)
178 tp->ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
181 void tcp_enter_quickack_mode(struct tcp_opt *tp)
183 tcp_incr_quickack(tp);
184 tp->ack.pingpong = 0;
185 tp->ack.ato = TCP_ATO_MIN;
188 /* Send ACKs quickly, if "quick" count is not exhausted
189 * and the session is not interactive.
192 static __inline__ int tcp_in_quickack_mode(struct tcp_opt *tp)
194 return (tp->ack.quick && !tp->ack.pingpong);
197 /* Buffer size and advertised window tuning.
199 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
202 static void tcp_fixup_sndbuf(struct sock *sk)
204 int sndmem = tcp_sk(sk)->mss_clamp + MAX_TCP_HEADER + 16 +
205 sizeof(struct sk_buff);
207 if (sk->sk_sndbuf < 3 * sndmem)
208 sk->sk_sndbuf = min(3 * sndmem, sysctl_tcp_wmem[2]);
211 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
213 * All tcp_full_space() is split to two parts: "network" buffer, allocated
214 * forward and advertised in receiver window (tp->rcv_wnd) and
215 * "application buffer", required to isolate scheduling/application
216 * latencies from network.
217 * window_clamp is maximal advertised window. It can be less than
218 * tcp_full_space(), in this case tcp_full_space() - window_clamp
219 * is reserved for "application" buffer. The less window_clamp is
220 * the smoother our behaviour from viewpoint of network, but the lower
221 * throughput and the higher sensitivity of the connection to losses. 8)
223 * rcv_ssthresh is more strict window_clamp used at "slow start"
224 * phase to predict further behaviour of this connection.
225 * It is used for two goals:
226 * - to enforce header prediction at sender, even when application
227 * requires some significant "application buffer". It is check #1.
228 * - to prevent pruning of receive queue because of misprediction
229 * of receiver window. Check #2.
231 * The scheme does not work when sender sends good segments opening
232 * window and then starts to feed us spagetti. But it should work
233 * in common situations. Otherwise, we have to rely on queue collapsing.
236 /* Slow part of check#2. */
238 __tcp_grow_window(struct sock *sk, struct tcp_opt *tp, struct sk_buff *skb)
241 int truesize = tcp_win_from_space(skb->truesize)/2;
242 int window = tcp_full_space(sk)/2;
244 while (tp->rcv_ssthresh <= window) {
245 if (truesize <= skb->len)
246 return 2*tp->ack.rcv_mss;
254 static __inline__ void
255 tcp_grow_window(struct sock *sk, struct tcp_opt *tp, struct sk_buff *skb)
258 if (tp->rcv_ssthresh < tp->window_clamp &&
259 (int)tp->rcv_ssthresh < tcp_space(sk) &&
260 !tcp_memory_pressure) {
263 /* Check #2. Increase window, if skb with such overhead
264 * will fit to rcvbuf in future.
266 if (tcp_win_from_space(skb->truesize) <= skb->len)
269 incr = __tcp_grow_window(sk, tp, skb);
272 tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr, tp->window_clamp);
278 /* 3. Tuning rcvbuf, when connection enters established state. */
280 static void tcp_fixup_rcvbuf(struct sock *sk)
282 struct tcp_opt *tp = tcp_sk(sk);
283 int rcvmem = tp->advmss + MAX_TCP_HEADER + 16 + sizeof(struct sk_buff);
285 /* Try to select rcvbuf so that 4 mss-sized segments
286 * will fit to window and correspoding skbs will fit to our rcvbuf.
287 * (was 3; 4 is minimum to allow fast retransmit to work.)
289 while (tcp_win_from_space(rcvmem) < tp->advmss)
291 if (sk->sk_rcvbuf < 4 * rcvmem)
292 sk->sk_rcvbuf = min(4 * rcvmem, sysctl_tcp_rmem[2]);
295 /* 4. Try to fixup all. It is made iimediately after connection enters
298 static void tcp_init_buffer_space(struct sock *sk)
300 struct tcp_opt *tp = tcp_sk(sk);
303 if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
304 tcp_fixup_rcvbuf(sk);
305 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
306 tcp_fixup_sndbuf(sk);
308 maxwin = tcp_full_space(sk);
310 if (tp->window_clamp >= maxwin) {
311 tp->window_clamp = maxwin;
313 if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
314 tp->window_clamp = max(maxwin -
315 (maxwin >> sysctl_tcp_app_win),
319 /* Force reservation of one segment. */
320 if (sysctl_tcp_app_win &&
321 tp->window_clamp > 2 * tp->advmss &&
322 tp->window_clamp + tp->advmss > maxwin)
323 tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
325 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
326 tp->snd_cwnd_stamp = tcp_time_stamp;
329 /* 5. Recalculate window clamp after socket hit its memory bounds. */
330 static void tcp_clamp_window(struct sock *sk, struct tcp_opt *tp)
333 unsigned int app_win = tp->rcv_nxt - tp->copied_seq;
338 skb_queue_walk(&tp->out_of_order_queue, skb) {
342 /* If overcommit is due to out of order segments,
343 * do not clamp window. Try to expand rcvbuf instead.
346 if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
347 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
348 !tcp_memory_pressure &&
349 atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0])
350 sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
353 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf) {
355 if (atomic_read(&sk->sk_rmem_alloc) >= 2 * sk->sk_rcvbuf)
357 if (app_win > tp->ack.rcv_mss)
358 app_win -= tp->ack.rcv_mss;
359 app_win = max(app_win, 2U*tp->advmss);
362 tp->window_clamp = min(tp->window_clamp, app_win);
363 tp->rcv_ssthresh = min(tp->window_clamp, 2U*tp->advmss);
367 /* There is something which you must keep in mind when you analyze the
368 * behavior of the tp->ato delayed ack timeout interval. When a
369 * connection starts up, we want to ack as quickly as possible. The
370 * problem is that "good" TCP's do slow start at the beginning of data
371 * transmission. The means that until we send the first few ACK's the
372 * sender will sit on his end and only queue most of his data, because
373 * he can only send snd_cwnd unacked packets at any given time. For
374 * each ACK we send, he increments snd_cwnd and transmits more of his
377 static void tcp_event_data_recv(struct sock *sk, struct tcp_opt *tp, struct sk_buff *skb)
381 tcp_schedule_ack(tp);
383 tcp_measure_rcv_mss(tp, skb);
385 now = tcp_time_stamp;
388 /* The _first_ data packet received, initialize
389 * delayed ACK engine.
391 tcp_incr_quickack(tp);
392 tp->ack.ato = TCP_ATO_MIN;
394 int m = now - tp->ack.lrcvtime;
396 if (m <= TCP_ATO_MIN/2) {
397 /* The fastest case is the first. */
398 tp->ack.ato = (tp->ack.ato>>1) + TCP_ATO_MIN/2;
399 } else if (m < tp->ack.ato) {
400 tp->ack.ato = (tp->ack.ato>>1) + m;
401 if (tp->ack.ato > tp->rto)
402 tp->ack.ato = tp->rto;
403 } else if (m > tp->rto) {
404 /* Too long gap. Apparently sender falled to
405 * restart window, so that we send ACKs quickly.
407 tcp_incr_quickack(tp);
411 tp->ack.lrcvtime = now;
413 TCP_ECN_check_ce(tp, skb);
416 tcp_grow_window(sk, tp, skb);
419 /* Set up a new TCP connection, depending on whether it should be
420 * using Vegas or not.
422 void tcp_vegas_init(struct tcp_opt *tp)
424 if (sysctl_tcp_vegas_cong_avoid) {
425 tp->vegas.do_vegas = 1;
426 tp->vegas.baseRTT = 0x7fffffff;
427 tcp_vegas_enable(tp);
429 tcp_vegas_disable(tp);
432 /* Do RTT sampling needed for Vegas.
434 * o min-filter RTT samples from within an RTT to get the current
435 * propagation delay + queuing delay (we are min-filtering to try to
436 * avoid the effects of delayed ACKs)
437 * o min-filter RTT samples from a much longer window (forever for now)
438 * to find the propagation delay (baseRTT)
440 static inline void vegas_rtt_calc(struct tcp_opt *tp, __u32 rtt)
442 __u32 vrtt = rtt + 1; /* Never allow zero rtt or baseRTT */
444 /* Filter to find propagation delay: */
445 if (vrtt < tp->vegas.baseRTT)
446 tp->vegas.baseRTT = vrtt;
448 /* Find the min RTT during the last RTT to find
449 * the current prop. delay + queuing delay:
451 tp->vegas.minRTT = min(tp->vegas.minRTT, vrtt);
455 /* Called to compute a smoothed rtt estimate. The data fed to this
456 * routine either comes from timestamps, or from segments that were
457 * known _not_ to have been retransmitted [see Karn/Partridge
458 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
459 * piece by Van Jacobson.
460 * NOTE: the next three routines used to be one big routine.
461 * To save cycles in the RFC 1323 implementation it was better to break
462 * it up into three procedures. -- erics
464 static void tcp_rtt_estimator(struct tcp_opt *tp, __u32 mrtt)
466 long m = mrtt; /* RTT */
468 if (tcp_vegas_enabled(tp))
469 vegas_rtt_calc(tp, mrtt);
471 /* The following amusing code comes from Jacobson's
472 * article in SIGCOMM '88. Note that rtt and mdev
473 * are scaled versions of rtt and mean deviation.
474 * This is designed to be as fast as possible
475 * m stands for "measurement".
477 * On a 1990 paper the rto value is changed to:
478 * RTO = rtt + 4 * mdev
480 * Funny. This algorithm seems to be very broken.
481 * These formulae increase RTO, when it should be decreased, increase
482 * too slowly, when it should be incresed fastly, decrease too fastly
483 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
484 * does not matter how to _calculate_ it. Seems, it was trap
485 * that VJ failed to avoid. 8)
490 m -= (tp->srtt >> 3); /* m is now error in rtt est */
491 tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */
493 m = -m; /* m is now abs(error) */
494 m -= (tp->mdev >> 2); /* similar update on mdev */
495 /* This is similar to one of Eifel findings.
496 * Eifel blocks mdev updates when rtt decreases.
497 * This solution is a bit different: we use finer gain
498 * for mdev in this case (alpha*beta).
499 * Like Eifel it also prevents growth of rto,
500 * but also it limits too fast rto decreases,
501 * happening in pure Eifel.
506 m -= (tp->mdev >> 2); /* similar update on mdev */
508 tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */
509 if (tp->mdev > tp->mdev_max) {
510 tp->mdev_max = tp->mdev;
511 if (tp->mdev_max > tp->rttvar)
512 tp->rttvar = tp->mdev_max;
514 if (after(tp->snd_una, tp->rtt_seq)) {
515 if (tp->mdev_max < tp->rttvar)
516 tp->rttvar -= (tp->rttvar-tp->mdev_max)>>2;
517 tp->rtt_seq = tp->snd_nxt;
518 tp->mdev_max = TCP_RTO_MIN;
521 /* no previous measure. */
522 tp->srtt = m<<3; /* take the measured time to be rtt */
523 tp->mdev = m<<1; /* make sure rto = 3*rtt */
524 tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN);
525 tp->rtt_seq = tp->snd_nxt;
528 tcp_westwood_update_rtt(tp, tp->srtt >> 3);
531 /* Calculate rto without backoff. This is the second half of Van Jacobson's
532 * routine referred to above.
534 static __inline__ void tcp_set_rto(struct tcp_opt *tp)
536 /* Old crap is replaced with new one. 8)
539 * 1. If rtt variance happened to be less 50msec, it is hallucination.
540 * It cannot be less due to utterly erratic ACK generation made
541 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
542 * to do with delayed acks, because at cwnd>2 true delack timeout
543 * is invisible. Actually, Linux-2.4 also generates erratic
544 * ACKs in some curcumstances.
546 tp->rto = (tp->srtt >> 3) + tp->rttvar;
548 /* 2. Fixups made earlier cannot be right.
549 * If we do not estimate RTO correctly without them,
550 * all the algo is pure shit and should be replaced
551 * with correct one. It is exaclty, which we pretend to do.
555 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
556 * guarantees that rto is higher.
558 static __inline__ void tcp_bound_rto(struct tcp_opt *tp)
560 if (tp->rto > TCP_RTO_MAX)
561 tp->rto = TCP_RTO_MAX;
564 /* Save metrics learned by this TCP session.
565 This function is called only, when TCP finishes successfully
566 i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
568 void tcp_update_metrics(struct sock *sk)
570 struct tcp_opt *tp = tcp_sk(sk);
571 struct dst_entry *dst = __sk_dst_get(sk);
573 if (sysctl_tcp_nometrics_save)
578 if (dst && (dst->flags&DST_HOST)) {
581 if (tp->backoff || !tp->srtt) {
582 /* This session failed to estimate rtt. Why?
583 * Probably, no packets returned in time.
586 if (!(dst_metric_locked(dst, RTAX_RTT)))
587 dst->metrics[RTAX_RTT-1] = 0;
591 m = dst_metric(dst, RTAX_RTT) - tp->srtt;
593 /* If newly calculated rtt larger than stored one,
594 * store new one. Otherwise, use EWMA. Remember,
595 * rtt overestimation is always better than underestimation.
597 if (!(dst_metric_locked(dst, RTAX_RTT))) {
599 dst->metrics[RTAX_RTT-1] = tp->srtt;
601 dst->metrics[RTAX_RTT-1] -= (m>>3);
604 if (!(dst_metric_locked(dst, RTAX_RTTVAR))) {
608 /* Scale deviation to rttvar fixed point */
613 if (m >= dst_metric(dst, RTAX_RTTVAR))
614 dst->metrics[RTAX_RTTVAR-1] = m;
616 dst->metrics[RTAX_RTTVAR-1] -=
617 (dst->metrics[RTAX_RTTVAR-1] - m)>>2;
620 if (tp->snd_ssthresh >= 0xFFFF) {
621 /* Slow start still did not finish. */
622 if (dst_metric(dst, RTAX_SSTHRESH) &&
623 !dst_metric_locked(dst, RTAX_SSTHRESH) &&
624 (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH))
625 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1;
626 if (!dst_metric_locked(dst, RTAX_CWND) &&
627 tp->snd_cwnd > dst_metric(dst, RTAX_CWND))
628 dst->metrics[RTAX_CWND-1] = tp->snd_cwnd;
629 } else if (tp->snd_cwnd > tp->snd_ssthresh &&
630 tp->ca_state == TCP_CA_Open) {
631 /* Cong. avoidance phase, cwnd is reliable. */
632 if (!dst_metric_locked(dst, RTAX_SSTHRESH))
633 dst->metrics[RTAX_SSTHRESH-1] =
634 max(tp->snd_cwnd >> 1, tp->snd_ssthresh);
635 if (!dst_metric_locked(dst, RTAX_CWND))
636 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_cwnd) >> 1;
638 /* Else slow start did not finish, cwnd is non-sense,
639 ssthresh may be also invalid.
641 if (!dst_metric_locked(dst, RTAX_CWND))
642 dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_ssthresh) >> 1;
643 if (dst->metrics[RTAX_SSTHRESH-1] &&
644 !dst_metric_locked(dst, RTAX_SSTHRESH) &&
645 tp->snd_ssthresh > dst->metrics[RTAX_SSTHRESH-1])
646 dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh;
649 if (!dst_metric_locked(dst, RTAX_REORDERING)) {
650 if (dst->metrics[RTAX_REORDERING-1] < tp->reordering &&
651 tp->reordering != sysctl_tcp_reordering)
652 dst->metrics[RTAX_REORDERING-1] = tp->reordering;
657 /* Numbers are taken from RFC2414. */
658 __u32 tcp_init_cwnd(struct tcp_opt *tp, struct dst_entry *dst)
660 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
663 if (tp->mss_cache > 1460)
666 cwnd = (tp->mss_cache > 1095) ? 3 : 4;
668 return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
671 /* Initialize metrics on socket. */
673 static void tcp_init_metrics(struct sock *sk)
675 struct tcp_opt *tp = tcp_sk(sk);
676 struct dst_entry *dst = __sk_dst_get(sk);
683 if (dst_metric_locked(dst, RTAX_CWND))
684 tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND);
685 if (dst_metric(dst, RTAX_SSTHRESH)) {
686 tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH);
687 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
688 tp->snd_ssthresh = tp->snd_cwnd_clamp;
690 if (dst_metric(dst, RTAX_REORDERING) &&
691 tp->reordering != dst_metric(dst, RTAX_REORDERING)) {
693 tp->reordering = dst_metric(dst, RTAX_REORDERING);
696 if (dst_metric(dst, RTAX_RTT) == 0)
699 if (!tp->srtt && dst_metric(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3))
702 /* Initial rtt is determined from SYN,SYN-ACK.
703 * The segment is small and rtt may appear much
704 * less than real one. Use per-dst memory
705 * to make it more realistic.
707 * A bit of theory. RTT is time passed after "normal" sized packet
708 * is sent until it is ACKed. In normal curcumstances sending small
709 * packets force peer to delay ACKs and calculation is correct too.
710 * The algorithm is adaptive and, provided we follow specs, it
711 * NEVER underestimate RTT. BUT! If peer tries to make some clever
712 * tricks sort of "quick acks" for time long enough to decrease RTT
713 * to low value, and then abruptly stops to do it and starts to delay
714 * ACKs, wait for troubles.
716 if (dst_metric(dst, RTAX_RTT) > tp->srtt)
717 tp->srtt = dst_metric(dst, RTAX_RTT);
718 if (dst_metric(dst, RTAX_RTTVAR) > tp->mdev) {
719 tp->mdev = dst_metric(dst, RTAX_RTTVAR);
720 tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN);
724 if (tp->rto < TCP_TIMEOUT_INIT && !tp->saw_tstamp)
726 tp->snd_cwnd = tcp_init_cwnd(tp, dst);
727 tp->snd_cwnd_stamp = tcp_time_stamp;
731 /* Play conservative. If timestamps are not
732 * supported, TCP will fail to recalculate correct
733 * rtt, if initial rto is too small. FORGET ALL AND RESET!
735 if (!tp->saw_tstamp && tp->srtt) {
737 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT;
738 tp->rto = TCP_TIMEOUT_INIT;
742 static void tcp_update_reordering(struct tcp_opt *tp, int metric, int ts)
744 if (metric > tp->reordering) {
745 tp->reordering = min(TCP_MAX_REORDERING, metric);
747 /* This exciting event is worth to be remembered. 8) */
749 NET_INC_STATS_BH(TCPTSReorder);
751 NET_INC_STATS_BH(TCPRenoReorder);
753 NET_INC_STATS_BH(TCPFACKReorder);
755 NET_INC_STATS_BH(TCPSACKReorder);
756 #if FASTRETRANS_DEBUG > 1
757 printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n",
758 tp->sack_ok, tp->ca_state,
759 tp->reordering, tp->fackets_out, tp->sacked_out,
760 tp->undo_marker ? tp->undo_retrans : 0);
762 /* Disable FACK yet. */
767 /* This procedure tags the retransmission queue when SACKs arrive.
769 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
770 * Packets in queue with these bits set are counted in variables
771 * sacked_out, retrans_out and lost_out, correspondingly.
773 * Valid combinations are:
774 * Tag InFlight Description
775 * 0 1 - orig segment is in flight.
776 * S 0 - nothing flies, orig reached receiver.
777 * L 0 - nothing flies, orig lost by net.
778 * R 2 - both orig and retransmit are in flight.
779 * L|R 1 - orig is lost, retransmit is in flight.
780 * S|R 1 - orig reached receiver, retrans is still in flight.
781 * (L|S|R is logically valid, it could occur when L|R is sacked,
782 * but it is equivalent to plain S and code short-curcuits it to S.
783 * L|S is logically invalid, it would mean -1 packet in flight 8))
785 * These 6 states form finite state machine, controlled by the following events:
786 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
787 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
788 * 3. Loss detection event of one of three flavors:
789 * A. Scoreboard estimator decided the packet is lost.
790 * A'. Reno "three dupacks" marks head of queue lost.
791 * A''. Its FACK modfication, head until snd.fack is lost.
792 * B. SACK arrives sacking data transmitted after never retransmitted
794 * C. SACK arrives sacking SND.NXT at the moment, when the
795 * segment was retransmitted.
796 * 4. D-SACK added new rule: D-SACK changes any tag to S.
798 * It is pleasant to note, that state diagram turns out to be commutative,
799 * so that we are allowed not to be bothered by order of our actions,
800 * when multiple events arrive simultaneously. (see the function below).
802 * Reordering detection.
803 * --------------------
804 * Reordering metric is maximal distance, which a packet can be displaced
805 * in packet stream. With SACKs we can estimate it:
807 * 1. SACK fills old hole and the corresponding segment was not
808 * ever retransmitted -> reordering. Alas, we cannot use it
809 * when segment was retransmitted.
810 * 2. The last flaw is solved with D-SACK. D-SACK arrives
811 * for retransmitted and already SACKed segment -> reordering..
812 * Both of these heuristics are not used in Loss state, when we cannot
813 * account for retransmits accurately.
816 tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, u32 prior_snd_una)
818 struct tcp_opt *tp = tcp_sk(sk);
819 unsigned char *ptr = ack_skb->h.raw + TCP_SKB_CB(ack_skb)->sacked;
820 struct tcp_sack_block *sp = (struct tcp_sack_block *)(ptr+2);
821 int num_sacks = (ptr[1] - TCPOLEN_SACK_BASE)>>3;
822 int reord = tp->packets_out;
824 u32 lost_retrans = 0;
828 /* So, SACKs for already sent large segments will be lost.
829 * Not good, but alternative is to resegment the queue. */
830 if (sk->sk_route_caps & NETIF_F_TSO) {
831 sk->sk_route_caps &= ~NETIF_F_TSO;
832 sk->sk_no_largesend = 1;
833 tp->mss_cache = tp->mss_cache_std;
838 prior_fackets = tp->fackets_out;
840 for (i=0; i<num_sacks; i++, sp++) {
842 __u32 start_seq = ntohl(sp->start_seq);
843 __u32 end_seq = ntohl(sp->end_seq);
847 /* Check for D-SACK. */
849 u32 ack = TCP_SKB_CB(ack_skb)->ack_seq;
851 if (before(start_seq, ack)) {
854 NET_INC_STATS_BH(TCPDSACKRecv);
855 } else if (num_sacks > 1 &&
856 !after(end_seq, ntohl(sp[1].end_seq)) &&
857 !before(start_seq, ntohl(sp[1].start_seq))) {
860 NET_INC_STATS_BH(TCPDSACKOfoRecv);
863 /* D-SACK for already forgotten data...
864 * Do dumb counting. */
866 !after(end_seq, prior_snd_una) &&
867 after(end_seq, tp->undo_marker))
870 /* Eliminate too old ACKs, but take into
871 * account more or less fresh ones, they can
872 * contain valid SACK info.
874 if (before(ack, prior_snd_una - tp->max_window))
878 /* Event "B" in the comment above. */
879 if (after(end_seq, tp->high_seq))
880 flag |= FLAG_DATA_LOST;
882 for_retrans_queue(skb, sk, tp) {
883 u8 sacked = TCP_SKB_CB(skb)->sacked;
886 /* The retransmission queue is always in order, so
887 * we can short-circuit the walk early.
889 if(!before(TCP_SKB_CB(skb)->seq, end_seq))
894 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
895 !before(end_seq, TCP_SKB_CB(skb)->end_seq);
897 /* Account D-SACK for retransmitted packet. */
898 if ((dup_sack && in_sack) &&
899 (sacked & TCPCB_RETRANS) &&
900 after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker))
903 /* The frame is ACKed. */
904 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) {
905 if (sacked&TCPCB_RETRANS) {
906 if ((dup_sack && in_sack) &&
907 (sacked&TCPCB_SACKED_ACKED))
908 reord = min(fack_count, reord);
910 /* If it was in a hole, we detected reordering. */
911 if (fack_count < prior_fackets &&
912 !(sacked&TCPCB_SACKED_ACKED))
913 reord = min(fack_count, reord);
916 /* Nothing to do; acked frame is about to be dropped. */
920 if ((sacked&TCPCB_SACKED_RETRANS) &&
921 after(end_seq, TCP_SKB_CB(skb)->ack_seq) &&
922 (!lost_retrans || after(end_seq, lost_retrans)))
923 lost_retrans = end_seq;
928 if (!(sacked&TCPCB_SACKED_ACKED)) {
929 if (sacked & TCPCB_SACKED_RETRANS) {
930 /* If the segment is not tagged as lost,
931 * we do not clear RETRANS, believing
932 * that retransmission is still in flight.
934 if (sacked & TCPCB_LOST) {
935 TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
940 /* New sack for not retransmitted frame,
941 * which was in hole. It is reordering.
943 if (!(sacked & TCPCB_RETRANS) &&
944 fack_count < prior_fackets)
945 reord = min(fack_count, reord);
947 if (sacked & TCPCB_LOST) {
948 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
953 TCP_SKB_CB(skb)->sacked |= TCPCB_SACKED_ACKED;
954 flag |= FLAG_DATA_SACKED;
957 if (fack_count > tp->fackets_out)
958 tp->fackets_out = fack_count;
960 if (dup_sack && (sacked&TCPCB_RETRANS))
961 reord = min(fack_count, reord);
964 /* D-SACK. We can detect redundant retransmission
965 * in S|R and plain R frames and clear it.
966 * undo_retrans is decreased above, L|R frames
967 * are accounted above as well.
970 (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS)) {
971 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
977 /* Check for lost retransmit. This superb idea is
978 * borrowed from "ratehalving". Event "C".
979 * Later note: FACK people cheated me again 8),
980 * we have to account for reordering! Ugly,
983 if (lost_retrans && tp->ca_state == TCP_CA_Recovery) {
986 for_retrans_queue(skb, sk, tp) {
987 if (after(TCP_SKB_CB(skb)->seq, lost_retrans))
989 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
991 if ((TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) &&
992 after(lost_retrans, TCP_SKB_CB(skb)->ack_seq) &&
994 !before(lost_retrans,
995 TCP_SKB_CB(skb)->ack_seq + tp->reordering *
997 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1000 if (!(TCP_SKB_CB(skb)->sacked&(TCPCB_LOST|TCPCB_SACKED_ACKED))) {
1002 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1003 flag |= FLAG_DATA_SACKED;
1004 NET_INC_STATS_BH(TCPLostRetransmit);
1010 tp->left_out = tp->sacked_out + tp->lost_out;
1012 if (reord < tp->fackets_out && tp->ca_state != TCP_CA_Loss)
1013 tcp_update_reordering(tp, (tp->fackets_out + 1) - reord, 0);
1015 #if FASTRETRANS_DEBUG > 0
1016 BUG_TRAP((int)tp->sacked_out >= 0);
1017 BUG_TRAP((int)tp->lost_out >= 0);
1018 BUG_TRAP((int)tp->retrans_out >= 0);
1019 BUG_TRAP((int)tcp_packets_in_flight(tp) >= 0);
1024 /* RTO occurred, but do not yet enter loss state. Instead, transmit two new
1025 * segments to see from the next ACKs whether any data was really missing.
1026 * If the RTO was spurious, new ACKs should arrive.
1028 void tcp_enter_frto(struct sock *sk)
1030 struct tcp_opt *tp = tcp_sk(sk);
1031 struct sk_buff *skb;
1033 tp->frto_counter = 1;
1035 if (tp->ca_state <= TCP_CA_Disorder ||
1036 tp->snd_una == tp->high_seq ||
1037 (tp->ca_state == TCP_CA_Loss && !tp->retransmits)) {
1038 tp->prior_ssthresh = tcp_current_ssthresh(tp);
1039 if (!tcp_westwood_ssthresh(tp))
1040 tp->snd_ssthresh = tcp_recalc_ssthresh(tp);
1043 /* Have to clear retransmission markers here to keep the bookkeeping
1044 * in shape, even though we are not yet in Loss state.
1045 * If something was really lost, it is eventually caught up
1046 * in tcp_enter_frto_loss.
1048 tp->retrans_out = 0;
1049 tp->undo_marker = tp->snd_una;
1050 tp->undo_retrans = 0;
1052 for_retrans_queue(skb, sk, tp) {
1053 TCP_SKB_CB(skb)->sacked &= ~TCPCB_RETRANS;
1055 tcp_sync_left_out(tp);
1057 tcp_set_ca_state(tp, TCP_CA_Open);
1058 tp->frto_highmark = tp->snd_nxt;
1061 /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
1062 * which indicates that we should follow the traditional RTO recovery,
1063 * i.e. mark everything lost and do go-back-N retransmission.
1065 static void tcp_enter_frto_loss(struct sock *sk)
1067 struct tcp_opt *tp = tcp_sk(sk);
1068 struct sk_buff *skb;
1073 tp->fackets_out = 0;
1075 for_retrans_queue(skb, sk, tp) {
1077 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1078 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) {
1080 /* Do not mark those segments lost that were
1081 * forward transmitted after RTO
1083 if(!after(TCP_SKB_CB(skb)->end_seq,
1084 tp->frto_highmark)) {
1085 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1090 tp->fackets_out = cnt;
1093 tcp_sync_left_out(tp);
1095 tp->snd_cwnd = tp->frto_counter + tcp_packets_in_flight(tp)+1;
1096 tp->snd_cwnd_cnt = 0;
1097 tp->snd_cwnd_stamp = tcp_time_stamp;
1098 tp->undo_marker = 0;
1099 tp->frto_counter = 0;
1101 tp->reordering = min_t(unsigned int, tp->reordering,
1102 sysctl_tcp_reordering);
1103 tcp_set_ca_state(tp, TCP_CA_Loss);
1104 tp->high_seq = tp->frto_highmark;
1105 TCP_ECN_queue_cwr(tp);
1108 void tcp_clear_retrans(struct tcp_opt *tp)
1111 tp->retrans_out = 0;
1113 tp->fackets_out = 0;
1117 tp->undo_marker = 0;
1118 tp->undo_retrans = 0;
1121 /* Enter Loss state. If "how" is not zero, forget all SACK information
1122 * and reset tags completely, otherwise preserve SACKs. If receiver
1123 * dropped its ofo queue, we will know this due to reneging detection.
1125 void tcp_enter_loss(struct sock *sk, int how)
1127 struct tcp_opt *tp = tcp_sk(sk);
1128 struct sk_buff *skb;
1131 /* Reduce ssthresh if it has not yet been made inside this window. */
1132 if (tp->ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
1133 (tp->ca_state == TCP_CA_Loss && !tp->retransmits)) {
1134 tp->prior_ssthresh = tcp_current_ssthresh(tp);
1135 tp->snd_ssthresh = tcp_recalc_ssthresh(tp);
1138 tp->snd_cwnd_cnt = 0;
1139 tp->snd_cwnd_stamp = tcp_time_stamp;
1141 tcp_clear_retrans(tp);
1143 /* Push undo marker, if it was plain RTO and nothing
1144 * was retransmitted. */
1146 tp->undo_marker = tp->snd_una;
1148 for_retrans_queue(skb, sk, tp) {
1150 if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS)
1151 tp->undo_marker = 0;
1152 TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1153 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
1154 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1155 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1159 tp->fackets_out = cnt;
1162 tcp_sync_left_out(tp);
1164 tp->reordering = min_t(unsigned int, tp->reordering,
1165 sysctl_tcp_reordering);
1166 tcp_set_ca_state(tp, TCP_CA_Loss);
1167 tp->high_seq = tp->snd_nxt;
1168 TCP_ECN_queue_cwr(tp);
1171 static int tcp_check_sack_reneging(struct sock *sk, struct tcp_opt *tp)
1173 struct sk_buff *skb;
1175 /* If ACK arrived pointing to a remembered SACK,
1176 * it means that our remembered SACKs do not reflect
1177 * real state of receiver i.e.
1178 * receiver _host_ is heavily congested (or buggy).
1179 * Do processing similar to RTO timeout.
1181 if ((skb = skb_peek(&sk->sk_write_queue)) != NULL &&
1182 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
1183 NET_INC_STATS_BH(TCPSACKReneging);
1185 tcp_enter_loss(sk, 1);
1187 tcp_retransmit_skb(sk, skb_peek(&sk->sk_write_queue));
1188 tcp_reset_xmit_timer(sk, TCP_TIME_RETRANS, tp->rto);
1194 static inline int tcp_fackets_out(struct tcp_opt *tp)
1196 return IsReno(tp) ? tp->sacked_out+1 : tp->fackets_out;
1199 static inline int tcp_skb_timedout(struct tcp_opt *tp, struct sk_buff *skb)
1201 return (tcp_time_stamp - TCP_SKB_CB(skb)->when > tp->rto);
1204 static inline int tcp_head_timedout(struct sock *sk, struct tcp_opt *tp)
1206 return tp->packets_out &&
1207 tcp_skb_timedout(tp, skb_peek(&sk->sk_write_queue));
1210 /* Linux NewReno/SACK/FACK/ECN state machine.
1211 * --------------------------------------
1213 * "Open" Normal state, no dubious events, fast path.
1214 * "Disorder" In all the respects it is "Open",
1215 * but requires a bit more attention. It is entered when
1216 * we see some SACKs or dupacks. It is split of "Open"
1217 * mainly to move some processing from fast path to slow one.
1218 * "CWR" CWND was reduced due to some Congestion Notification event.
1219 * It can be ECN, ICMP source quench, local device congestion.
1220 * "Recovery" CWND was reduced, we are fast-retransmitting.
1221 * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
1223 * tcp_fastretrans_alert() is entered:
1224 * - each incoming ACK, if state is not "Open"
1225 * - when arrived ACK is unusual, namely:
1230 * Counting packets in flight is pretty simple.
1232 * in_flight = packets_out - left_out + retrans_out
1234 * packets_out is SND.NXT-SND.UNA counted in packets.
1236 * retrans_out is number of retransmitted segments.
1238 * left_out is number of segments left network, but not ACKed yet.
1240 * left_out = sacked_out + lost_out
1242 * sacked_out: Packets, which arrived to receiver out of order
1243 * and hence not ACKed. With SACKs this number is simply
1244 * amount of SACKed data. Even without SACKs
1245 * it is easy to give pretty reliable estimate of this number,
1246 * counting duplicate ACKs.
1248 * lost_out: Packets lost by network. TCP has no explicit
1249 * "loss notification" feedback from network (for now).
1250 * It means that this number can be only _guessed_.
1251 * Actually, it is the heuristics to predict lossage that
1252 * distinguishes different algorithms.
1254 * F.e. after RTO, when all the queue is considered as lost,
1255 * lost_out = packets_out and in_flight = retrans_out.
1257 * Essentially, we have now two algorithms counting
1260 * FACK: It is the simplest heuristics. As soon as we decided
1261 * that something is lost, we decide that _all_ not SACKed
1262 * packets until the most forward SACK are lost. I.e.
1263 * lost_out = fackets_out - sacked_out and left_out = fackets_out.
1264 * It is absolutely correct estimate, if network does not reorder
1265 * packets. And it loses any connection to reality when reordering
1266 * takes place. We use FACK by default until reordering
1267 * is suspected on the path to this destination.
1269 * NewReno: when Recovery is entered, we assume that one segment
1270 * is lost (classic Reno). While we are in Recovery and
1271 * a partial ACK arrives, we assume that one more packet
1272 * is lost (NewReno). This heuristics are the same in NewReno
1275 * Imagine, that's all! Forget about all this shamanism about CWND inflation
1276 * deflation etc. CWND is real congestion window, never inflated, changes
1277 * only according to classic VJ rules.
1279 * Really tricky (and requiring careful tuning) part of algorithm
1280 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
1281 * The first determines the moment _when_ we should reduce CWND and,
1282 * hence, slow down forward transmission. In fact, it determines the moment
1283 * when we decide that hole is caused by loss, rather than by a reorder.
1285 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
1286 * holes, caused by lost packets.
1288 * And the most logically complicated part of algorithm is undo
1289 * heuristics. We detect false retransmits due to both too early
1290 * fast retransmit (reordering) and underestimated RTO, analyzing
1291 * timestamps and D-SACKs. When we detect that some segments were
1292 * retransmitted by mistake and CWND reduction was wrong, we undo
1293 * window reduction and abort recovery phase. This logic is hidden
1294 * inside several functions named tcp_try_undo_<something>.
1297 /* This function decides, when we should leave Disordered state
1298 * and enter Recovery phase, reducing congestion window.
1300 * Main question: may we further continue forward transmission
1301 * with the same cwnd?
1304 tcp_time_to_recover(struct sock *sk, struct tcp_opt *tp)
1306 /* Trick#1: The loss is proven. */
1310 /* Not-A-Trick#2 : Classic rule... */
1311 if (tcp_fackets_out(tp) > tp->reordering)
1314 /* Trick#3 : when we use RFC2988 timer restart, fast
1315 * retransmit can be triggered by timeout of queue head.
1317 if (tcp_head_timedout(sk, tp))
1320 /* Trick#4: It is still not OK... But will it be useful to delay
1323 if (tp->packets_out <= tp->reordering &&
1324 tp->sacked_out >= max_t(__u32, tp->packets_out/2, sysctl_tcp_reordering) &&
1325 !tcp_may_send_now(sk, tp)) {
1326 /* We have nothing to send. This connection is limited
1327 * either by receiver window or by application.
1335 /* If we receive more dupacks than we expected counting segments
1336 * in assumption of absent reordering, interpret this as reordering.
1337 * The only another reason could be bug in receiver TCP.
1339 static void tcp_check_reno_reordering(struct tcp_opt *tp, int addend)
1343 holes = max(tp->lost_out, 1U);
1344 holes = min(holes, tp->packets_out);
1346 if (tp->sacked_out + holes > tp->packets_out) {
1347 tp->sacked_out = tp->packets_out - holes;
1348 tcp_update_reordering(tp, tp->packets_out+addend, 0);
1352 /* Emulate SACKs for SACKless connection: account for a new dupack. */
1354 static void tcp_add_reno_sack(struct tcp_opt *tp)
1357 tcp_check_reno_reordering(tp, 0);
1358 tcp_sync_left_out(tp);
1361 /* Account for ACK, ACKing some data in Reno Recovery phase. */
1363 static void tcp_remove_reno_sacks(struct sock *sk, struct tcp_opt *tp, int acked)
1366 /* One ACK acked hole. The rest eat duplicate ACKs. */
1367 if (acked-1 >= tp->sacked_out)
1370 tp->sacked_out -= acked-1;
1372 tcp_check_reno_reordering(tp, acked);
1373 tcp_sync_left_out(tp);
1376 static inline void tcp_reset_reno_sack(struct tcp_opt *tp)
1379 tp->left_out = tp->lost_out;
1382 /* Mark head of queue up as lost. */
1384 tcp_mark_head_lost(struct sock *sk, struct tcp_opt *tp, int packets, u32 high_seq)
1386 struct sk_buff *skb;
1389 BUG_TRAP(cnt <= tp->packets_out);
1391 for_retrans_queue(skb, sk, tp) {
1392 if (--cnt < 0 || after(TCP_SKB_CB(skb)->end_seq, high_seq))
1394 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
1395 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1399 tcp_sync_left_out(tp);
1402 /* Account newly detected lost packet(s) */
1404 static void tcp_update_scoreboard(struct sock *sk, struct tcp_opt *tp)
1407 int lost = tp->fackets_out - tp->reordering;
1410 tcp_mark_head_lost(sk, tp, lost, tp->high_seq);
1412 tcp_mark_head_lost(sk, tp, 1, tp->high_seq);
1415 /* New heuristics: it is possible only after we switched
1416 * to restart timer each time when something is ACKed.
1417 * Hence, we can detect timed out packets during fast
1418 * retransmit without falling to slow start.
1420 if (tcp_head_timedout(sk, tp)) {
1421 struct sk_buff *skb;
1423 for_retrans_queue(skb, sk, tp) {
1424 if (tcp_skb_timedout(tp, skb) &&
1425 !(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
1426 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1430 tcp_sync_left_out(tp);
1434 /* CWND moderation, preventing bursts due to too big ACKs
1435 * in dubious situations.
1437 static __inline__ void tcp_moderate_cwnd(struct tcp_opt *tp)
1439 tp->snd_cwnd = min(tp->snd_cwnd,
1440 tcp_packets_in_flight(tp)+tcp_max_burst(tp));
1441 tp->snd_cwnd_stamp = tcp_time_stamp;
1444 /* Decrease cwnd each second ack. */
1446 static void tcp_cwnd_down(struct tcp_opt *tp)
1448 int decr = tp->snd_cwnd_cnt + 1;
1453 * Here limit is evaluated as BWestimation*RTTmin (for obtaining it
1454 * in packets we use mss_cache). If sysctl_tcp_westwood is off
1455 * tcp_westwood_bw_rttmin() returns 0. In such case snd_ssthresh is
1456 * still used as usual. It prevents other strange cases in which
1457 * BWE*RTTmin could assume value 0. It should not happen but...
1460 if (!(limit = tcp_westwood_bw_rttmin(tp)))
1461 limit = tp->snd_ssthresh/2;
1463 tp->snd_cwnd_cnt = decr&1;
1466 if (decr && tp->snd_cwnd > limit)
1467 tp->snd_cwnd -= decr;
1469 tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp)+1);
1470 tp->snd_cwnd_stamp = tcp_time_stamp;
1473 /* Nothing was retransmitted or returned timestamp is less
1474 * than timestamp of the first retransmission.
1476 static __inline__ int tcp_packet_delayed(struct tcp_opt *tp)
1478 return !tp->retrans_stamp ||
1479 (tp->saw_tstamp && tp->rcv_tsecr &&
1480 (__s32)(tp->rcv_tsecr - tp->retrans_stamp) < 0);
1483 /* Undo procedures. */
1485 #if FASTRETRANS_DEBUG > 1
1486 static void DBGUNDO(struct sock *sk, struct tcp_opt *tp, const char *msg)
1488 struct inet_opt *inet = inet_sk(sk);
1489 printk(KERN_DEBUG "Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n",
1491 NIPQUAD(inet->daddr), ntohs(inet->dport),
1492 tp->snd_cwnd, tp->left_out,
1493 tp->snd_ssthresh, tp->prior_ssthresh, tp->packets_out);
1496 #define DBGUNDO(x...) do { } while (0)
1499 static void tcp_undo_cwr(struct tcp_opt *tp, int undo)
1501 if (tp->prior_ssthresh) {
1502 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh<<1);
1504 if (undo && tp->prior_ssthresh > tp->snd_ssthresh) {
1505 tp->snd_ssthresh = tp->prior_ssthresh;
1506 TCP_ECN_withdraw_cwr(tp);
1509 tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
1511 tcp_moderate_cwnd(tp);
1512 tp->snd_cwnd_stamp = tcp_time_stamp;
1515 static inline int tcp_may_undo(struct tcp_opt *tp)
1517 return tp->undo_marker &&
1518 (!tp->undo_retrans || tcp_packet_delayed(tp));
1521 /* People celebrate: "We love our President!" */
1522 static int tcp_try_undo_recovery(struct sock *sk, struct tcp_opt *tp)
1524 if (tcp_may_undo(tp)) {
1525 /* Happy end! We did not retransmit anything
1526 * or our original transmission succeeded.
1528 DBGUNDO(sk, tp, tp->ca_state == TCP_CA_Loss ? "loss" : "retrans");
1529 tcp_undo_cwr(tp, 1);
1530 if (tp->ca_state == TCP_CA_Loss)
1531 NET_INC_STATS_BH(TCPLossUndo);
1533 NET_INC_STATS_BH(TCPFullUndo);
1534 tp->undo_marker = 0;
1536 if (tp->snd_una == tp->high_seq && IsReno(tp)) {
1537 /* Hold old state until something *above* high_seq
1538 * is ACKed. For Reno it is MUST to prevent false
1539 * fast retransmits (RFC2582). SACK TCP is safe. */
1540 tcp_moderate_cwnd(tp);
1543 tcp_set_ca_state(tp, TCP_CA_Open);
1547 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
1548 static void tcp_try_undo_dsack(struct sock *sk, struct tcp_opt *tp)
1550 if (tp->undo_marker && !tp->undo_retrans) {
1551 DBGUNDO(sk, tp, "D-SACK");
1552 tcp_undo_cwr(tp, 1);
1553 tp->undo_marker = 0;
1554 NET_INC_STATS_BH(TCPDSACKUndo);
1558 /* Undo during fast recovery after partial ACK. */
1560 static int tcp_try_undo_partial(struct sock *sk, struct tcp_opt *tp, int acked)
1562 /* Partial ACK arrived. Force Hoe's retransmit. */
1563 int failed = IsReno(tp) || tp->fackets_out>tp->reordering;
1565 if (tcp_may_undo(tp)) {
1566 /* Plain luck! Hole if filled with delayed
1567 * packet, rather than with a retransmit.
1569 if (tp->retrans_out == 0)
1570 tp->retrans_stamp = 0;
1572 tcp_update_reordering(tp, tcp_fackets_out(tp)+acked, 1);
1574 DBGUNDO(sk, tp, "Hoe");
1575 tcp_undo_cwr(tp, 0);
1576 NET_INC_STATS_BH(TCPPartialUndo);
1578 /* So... Do not make Hoe's retransmit yet.
1579 * If the first packet was delayed, the rest
1580 * ones are most probably delayed as well.
1587 /* Undo during loss recovery after partial ACK. */
1588 static int tcp_try_undo_loss(struct sock *sk, struct tcp_opt *tp)
1590 if (tcp_may_undo(tp)) {
1591 struct sk_buff *skb;
1592 for_retrans_queue(skb, sk, tp) {
1593 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
1595 DBGUNDO(sk, tp, "partial loss");
1597 tp->left_out = tp->sacked_out;
1598 tcp_undo_cwr(tp, 1);
1599 NET_INC_STATS_BH(TCPLossUndo);
1600 tp->retransmits = 0;
1601 tp->undo_marker = 0;
1603 tcp_set_ca_state(tp, TCP_CA_Open);
1609 static __inline__ void tcp_complete_cwr(struct tcp_opt *tp)
1611 if (tcp_westwood_cwnd(tp))
1612 tp->snd_ssthresh = tp->snd_cwnd;
1614 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
1615 tp->snd_cwnd_stamp = tcp_time_stamp;
1618 static void tcp_try_to_open(struct sock *sk, struct tcp_opt *tp, int flag)
1620 tp->left_out = tp->sacked_out;
1622 if (tp->retrans_out == 0)
1623 tp->retrans_stamp = 0;
1628 if (tp->ca_state != TCP_CA_CWR) {
1629 int state = TCP_CA_Open;
1634 state = TCP_CA_Disorder;
1636 if (tp->ca_state != state) {
1637 tcp_set_ca_state(tp, state);
1638 tp->high_seq = tp->snd_nxt;
1640 tcp_moderate_cwnd(tp);
1646 /* Process an event, which can update packets-in-flight not trivially.
1647 * Main goal of this function is to calculate new estimate for left_out,
1648 * taking into account both packets sitting in receiver's buffer and
1649 * packets lost by network.
1651 * Besides that it does CWND reduction, when packet loss is detected
1652 * and changes state of machine.
1654 * It does _not_ decide what to send, it is made in function
1655 * tcp_xmit_retransmit_queue().
1658 tcp_fastretrans_alert(struct sock *sk, u32 prior_snd_una,
1659 int prior_packets, int flag)
1661 struct tcp_opt *tp = tcp_sk(sk);
1662 int is_dupack = (tp->snd_una == prior_snd_una && !(flag&FLAG_NOT_DUP));
1664 /* Some technical things:
1665 * 1. Reno does not count dupacks (sacked_out) automatically. */
1666 if (!tp->packets_out)
1668 /* 2. SACK counts snd_fack in packets inaccurately. */
1669 if (tp->sacked_out == 0)
1670 tp->fackets_out = 0;
1672 /* Now state machine starts.
1673 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
1675 tp->prior_ssthresh = 0;
1677 /* B. In all the states check for reneging SACKs. */
1678 if (tp->sacked_out && tcp_check_sack_reneging(sk, tp))
1681 /* C. Process data loss notification, provided it is valid. */
1682 if ((flag&FLAG_DATA_LOST) &&
1683 before(tp->snd_una, tp->high_seq) &&
1684 tp->ca_state != TCP_CA_Open &&
1685 tp->fackets_out > tp->reordering) {
1686 tcp_mark_head_lost(sk, tp, tp->fackets_out-tp->reordering, tp->high_seq);
1687 NET_INC_STATS_BH(TCPLoss);
1690 /* D. Synchronize left_out to current state. */
1691 tcp_sync_left_out(tp);
1693 /* E. Check state exit conditions. State can be terminated
1694 * when high_seq is ACKed. */
1695 if (tp->ca_state == TCP_CA_Open) {
1696 if (!sysctl_tcp_frto)
1697 BUG_TRAP(tp->retrans_out == 0);
1698 tp->retrans_stamp = 0;
1699 } else if (!before(tp->snd_una, tp->high_seq)) {
1700 switch (tp->ca_state) {
1702 tp->retransmits = 0;
1703 if (tcp_try_undo_recovery(sk, tp))
1708 /* CWR is to be held something *above* high_seq
1709 * is ACKed for CWR bit to reach receiver. */
1710 if (tp->snd_una != tp->high_seq) {
1711 tcp_complete_cwr(tp);
1712 tcp_set_ca_state(tp, TCP_CA_Open);
1716 case TCP_CA_Disorder:
1717 tcp_try_undo_dsack(sk, tp);
1718 if (!tp->undo_marker ||
1719 /* For SACK case do not Open to allow to undo
1720 * catching for all duplicate ACKs. */
1721 IsReno(tp) || tp->snd_una != tp->high_seq) {
1722 tp->undo_marker = 0;
1723 tcp_set_ca_state(tp, TCP_CA_Open);
1727 case TCP_CA_Recovery:
1729 tcp_reset_reno_sack(tp);
1730 if (tcp_try_undo_recovery(sk, tp))
1732 tcp_complete_cwr(tp);
1737 /* F. Process state. */
1738 switch (tp->ca_state) {
1739 case TCP_CA_Recovery:
1740 if (prior_snd_una == tp->snd_una) {
1741 if (IsReno(tp) && is_dupack)
1742 tcp_add_reno_sack(tp);
1744 int acked = prior_packets - tp->packets_out;
1746 tcp_remove_reno_sacks(sk, tp, acked);
1747 is_dupack = tcp_try_undo_partial(sk, tp, acked);
1751 if (flag&FLAG_DATA_ACKED)
1752 tp->retransmits = 0;
1753 if (!tcp_try_undo_loss(sk, tp)) {
1754 tcp_moderate_cwnd(tp);
1755 tcp_xmit_retransmit_queue(sk);
1758 if (tp->ca_state != TCP_CA_Open)
1760 /* Loss is undone; fall through to processing in Open state. */
1763 if (tp->snd_una != prior_snd_una)
1764 tcp_reset_reno_sack(tp);
1766 tcp_add_reno_sack(tp);
1769 if (tp->ca_state == TCP_CA_Disorder)
1770 tcp_try_undo_dsack(sk, tp);
1772 if (!tcp_time_to_recover(sk, tp)) {
1773 tcp_try_to_open(sk, tp, flag);
1777 /* Otherwise enter Recovery state */
1780 NET_INC_STATS_BH(TCPRenoRecovery);
1782 NET_INC_STATS_BH(TCPSackRecovery);
1784 tp->high_seq = tp->snd_nxt;
1785 tp->prior_ssthresh = 0;
1786 tp->undo_marker = tp->snd_una;
1787 tp->undo_retrans = tp->retrans_out;
1789 if (tp->ca_state < TCP_CA_CWR) {
1790 if (!(flag&FLAG_ECE))
1791 tp->prior_ssthresh = tcp_current_ssthresh(tp);
1792 tp->snd_ssthresh = tcp_recalc_ssthresh(tp);
1793 TCP_ECN_queue_cwr(tp);
1796 tp->snd_cwnd_cnt = 0;
1797 tcp_set_ca_state(tp, TCP_CA_Recovery);
1800 if (is_dupack || tcp_head_timedout(sk, tp))
1801 tcp_update_scoreboard(sk, tp);
1803 tcp_xmit_retransmit_queue(sk);
1806 /* Read draft-ietf-tcplw-high-performance before mucking
1807 * with this code. (Superceeds RFC1323)
1809 static void tcp_ack_saw_tstamp(struct tcp_opt *tp, int flag)
1813 /* RTTM Rule: A TSecr value received in a segment is used to
1814 * update the averaged RTT measurement only if the segment
1815 * acknowledges some new data, i.e., only if it advances the
1816 * left edge of the send window.
1818 * See draft-ietf-tcplw-high-performance-00, section 3.3.
1819 * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
1821 * Changed: reset backoff as soon as we see the first valid sample.
1822 * If we do not, we get strongly overstimated rto. With timestamps
1823 * samples are accepted even from very old segments: f.e., when rtt=1
1824 * increases to 8, we retransmit 5 times and after 8 seconds delayed
1825 * answer arrives rto becomes 120 seconds! If at least one of segments
1826 * in window is lost... Voila. --ANK (010210)
1828 seq_rtt = tcp_time_stamp - tp->rcv_tsecr;
1829 tcp_rtt_estimator(tp, seq_rtt);
1835 static void tcp_ack_no_tstamp(struct tcp_opt *tp, u32 seq_rtt, int flag)
1837 /* We don't have a timestamp. Can only use
1838 * packets that are not retransmitted to determine
1839 * rtt estimates. Also, we must not reset the
1840 * backoff for rto until we get a non-retransmitted
1841 * packet. This allows us to deal with a situation
1842 * where the network delay has increased suddenly.
1843 * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
1846 if (flag & FLAG_RETRANS_DATA_ACKED)
1849 tcp_rtt_estimator(tp, seq_rtt);
1855 static __inline__ void
1856 tcp_ack_update_rtt(struct tcp_opt *tp, int flag, s32 seq_rtt)
1858 /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
1859 if (tp->saw_tstamp && tp->rcv_tsecr)
1860 tcp_ack_saw_tstamp(tp, flag);
1861 else if (seq_rtt >= 0)
1862 tcp_ack_no_tstamp(tp, seq_rtt, flag);
1866 * Compute congestion window to use.
1868 * This is from the implementation of BICTCP in
1869 * Lison-Xu, Kahaled Harfoush, and Injog Rhee.
1870 * "Binary Increase Congestion Control for Fast, Long Distance
1871 * Networks" in InfoComm 2004
1873 * http://www.csc.ncsu.edu/faculty/rhee/export/bitcp.pdf
1875 * Unless BIC is enabled and congestion window is large
1876 * this behaves the same as the original Reno.
1878 static inline __u32 bictcp_cwnd(struct tcp_opt *tp)
1880 /* orignal Reno behaviour */
1881 if (!sysctl_tcp_bic)
1882 return tp->snd_cwnd;
1884 if (tp->bictcp.last_cwnd == tp->snd_cwnd)
1885 return tp->bictcp.cnt; /* same cwnd, no update */
1887 tp->bictcp.last_cwnd = tp->snd_cwnd;
1889 /* start off normal */
1890 if (tp->snd_cwnd <= sysctl_tcp_bic_low_window)
1891 tp->bictcp.cnt = tp->snd_cwnd;
1893 /* binary increase */
1894 else if (tp->snd_cwnd < tp->bictcp.last_max_cwnd) {
1895 __u32 dist = (tp->bictcp.last_max_cwnd - tp->snd_cwnd)
1898 if (dist > BICTCP_MAX_INCREMENT)
1899 /* linear increase */
1900 tp->bictcp.cnt = tp->snd_cwnd / BICTCP_MAX_INCREMENT;
1901 else if (dist <= 1U)
1902 /* binary search increase */
1903 tp->bictcp.cnt = tp->snd_cwnd * BICTCP_FUNC_OF_MIN_INCR
1906 /* binary search increase */
1907 tp->bictcp.cnt = tp->snd_cwnd / dist;
1909 /* slow start amd linear increase */
1910 if (tp->snd_cwnd < tp->bictcp.last_max_cwnd + BICTCP_B)
1912 tp->bictcp.cnt = tp->snd_cwnd * BICTCP_FUNC_OF_MIN_INCR
1914 else if (tp->snd_cwnd < tp->bictcp.last_max_cwnd
1915 + BICTCP_MAX_INCREMENT*(BICTCP_B-1))
1917 tp->bictcp.cnt = tp->snd_cwnd * (BICTCP_B-1)
1918 / (tp->snd_cwnd-tp->bictcp.last_max_cwnd);
1920 /* linear increase */
1921 tp->bictcp.cnt = tp->snd_cwnd / BICTCP_MAX_INCREMENT;
1923 return tp->bictcp.cnt;
1926 /* This is Jacobson's slow start and congestion avoidance.
1927 * SIGCOMM '88, p. 328.
1929 static __inline__ void reno_cong_avoid(struct tcp_opt *tp)
1931 if (tp->snd_cwnd <= tp->snd_ssthresh) {
1932 /* In "safe" area, increase. */
1933 if (tp->snd_cwnd < tp->snd_cwnd_clamp)
1936 /* In dangerous area, increase slowly.
1937 * In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd
1939 if (tp->snd_cwnd_cnt >= bictcp_cwnd(tp)) {
1940 if (tp->snd_cwnd < tp->snd_cwnd_clamp)
1946 tp->snd_cwnd_stamp = tcp_time_stamp;
1949 /* This is based on the congestion detection/avoidance scheme described in
1950 * Lawrence S. Brakmo and Larry L. Peterson.
1951 * "TCP Vegas: End to end congestion avoidance on a global internet."
1952 * IEEE Journal on Selected Areas in Communication, 13(8):1465--1480,
1953 * October 1995. Available from:
1954 * ftp://ftp.cs.arizona.edu/xkernel/Papers/jsac.ps
1956 * See http://www.cs.arizona.edu/xkernel/ for their implementation.
1957 * The main aspects that distinguish this implementation from the
1958 * Arizona Vegas implementation are:
1959 * o We do not change the loss detection or recovery mechanisms of
1960 * Linux in any way. Linux already recovers from losses quite well,
1961 * using fine-grained timers, NewReno, and FACK.
1962 * o To avoid the performance penalty imposed by increasing cwnd
1963 * only every-other RTT during slow start, we increase during
1964 * every RTT during slow start, just like Reno.
1965 * o Largely to allow continuous cwnd growth during slow start,
1966 * we use the rate at which ACKs come back as the "actual"
1967 * rate, rather than the rate at which data is sent.
1968 * o To speed convergence to the right rate, we set the cwnd
1969 * to achieve the right ("actual") rate when we exit slow start.
1970 * o To filter out the noise caused by delayed ACKs, we use the
1971 * minimum RTT sample observed during the last RTT to calculate
1973 * o When the sender re-starts from idle, it waits until it has
1974 * received ACKs for an entire flight of new data before making
1975 * a cwnd adjustment decision. The original Vegas implementation
1976 * assumed senders never went idle.
1978 static void vegas_cong_avoid(struct tcp_opt *tp, u32 ack, u32 seq_rtt)
1980 /* The key players are v_beg_snd_una and v_beg_snd_nxt.
1982 * These are so named because they represent the approximate values
1983 * of snd_una and snd_nxt at the beginning of the current RTT. More
1984 * precisely, they represent the amount of data sent during the RTT.
1985 * At the end of the RTT, when we receive an ACK for v_beg_snd_nxt,
1986 * we will calculate that (v_beg_snd_nxt - v_beg_snd_una) outstanding
1987 * bytes of data have been ACKed during the course of the RTT, giving
1988 * an "actual" rate of:
1990 * (v_beg_snd_nxt - v_beg_snd_una) / (rtt duration)
1992 * Unfortunately, v_beg_snd_una is not exactly equal to snd_una,
1993 * because delayed ACKs can cover more than one segment, so they
1994 * don't line up nicely with the boundaries of RTTs.
1996 * Another unfortunate fact of life is that delayed ACKs delay the
1997 * advance of the left edge of our send window, so that the number
1998 * of bytes we send in an RTT is often less than our cwnd will allow.
1999 * So we keep track of our cwnd separately, in v_beg_snd_cwnd.
2002 if (after(ack, tp->vegas.beg_snd_nxt)) {
2003 /* Do the Vegas once-per-RTT cwnd adjustment. */
2004 u32 old_wnd, old_snd_cwnd;
2007 /* Here old_wnd is essentially the window of data that was
2008 * sent during the previous RTT, and has all
2009 * been acknowledged in the course of the RTT that ended
2010 * with the ACK we just received. Likewise, old_snd_cwnd
2011 * is the cwnd during the previous RTT.
2013 old_wnd = (tp->vegas.beg_snd_nxt - tp->vegas.beg_snd_una) /
2015 old_snd_cwnd = tp->vegas.beg_snd_cwnd;
2017 /* Save the extent of the current window so we can use this
2018 * at the end of the next RTT.
2020 tp->vegas.beg_snd_una = tp->vegas.beg_snd_nxt;
2021 tp->vegas.beg_snd_nxt = tp->snd_nxt;
2022 tp->vegas.beg_snd_cwnd = tp->snd_cwnd;
2024 /* Take into account the current RTT sample too, to
2025 * decrease the impact of delayed acks. This double counts
2026 * this sample since we count it for the next window as well,
2027 * but that's not too awful, since we're taking the min,
2028 * rather than averaging.
2030 vegas_rtt_calc(tp, seq_rtt);
2032 /* We do the Vegas calculations only if we got enough RTT
2033 * samples that we can be reasonably sure that we got
2034 * at least one RTT sample that wasn't from a delayed ACK.
2035 * If we only had 2 samples total,
2036 * then that means we're getting only 1 ACK per RTT, which
2037 * means they're almost certainly delayed ACKs.
2038 * If we have 3 samples, we should be OK.
2041 if (tp->vegas.cntRTT <= 2) {
2042 /* We don't have enough RTT samples to do the Vegas
2043 * calculation, so we'll behave like Reno.
2045 if (tp->snd_cwnd > tp->snd_ssthresh)
2048 u32 rtt, target_cwnd, diff;
2050 /* We have enough RTT samples, so, using the Vegas
2051 * algorithm, we determine if we should increase or
2052 * decrease cwnd, and by how much.
2055 /* Pluck out the RTT we are using for the Vegas
2056 * calculations. This is the min RTT seen during the
2057 * last RTT. Taking the min filters out the effects
2058 * of delayed ACKs, at the cost of noticing congestion
2061 rtt = tp->vegas.minRTT;
2063 /* Calculate the cwnd we should have, if we weren't
2067 * (actual rate in segments) * baseRTT
2068 * We keep it as a fixed point number with
2069 * V_PARAM_SHIFT bits to the right of the binary point.
2071 target_cwnd = ((old_wnd * tp->vegas.baseRTT)
2072 << V_PARAM_SHIFT) / rtt;
2074 /* Calculate the difference between the window we had,
2075 * and the window we would like to have. This quantity
2076 * is the "Diff" from the Arizona Vegas papers.
2078 * Again, this is a fixed point number with
2079 * V_PARAM_SHIFT bits to the right of the binary
2082 diff = (old_wnd << V_PARAM_SHIFT) - target_cwnd;
2084 if (tp->snd_cwnd < tp->snd_ssthresh) {
2086 if (diff > sysctl_tcp_vegas_gamma) {
2087 /* Going too fast. Time to slow down
2088 * and switch to congestion avoidance.
2090 tp->snd_ssthresh = 2;
2092 /* Set cwnd to match the actual rate
2094 * cwnd = (actual rate) * baseRTT
2095 * Then we add 1 because the integer
2096 * truncation robs us of full link
2099 tp->snd_cwnd = min(tp->snd_cwnd,
2105 /* Congestion avoidance. */
2108 /* Figure out where we would like cwnd
2111 if (diff > sysctl_tcp_vegas_beta) {
2112 /* The old window was too fast, so
2115 next_snd_cwnd = old_snd_cwnd - 1;
2116 } else if (diff < sysctl_tcp_vegas_alpha) {
2117 /* We don't have enough extra packets
2118 * in the network, so speed up.
2120 next_snd_cwnd = old_snd_cwnd + 1;
2122 /* Sending just as fast as we
2125 next_snd_cwnd = old_snd_cwnd;
2128 /* Adjust cwnd upward or downward, toward the
2131 if (next_snd_cwnd > tp->snd_cwnd)
2133 else if (next_snd_cwnd < tp->snd_cwnd)
2138 /* Wipe the slate clean for the next RTT. */
2139 tp->vegas.cntRTT = 0;
2140 tp->vegas.minRTT = 0x7fffffff;
2143 /* The following code is executed for every ack we receive,
2144 * except for conditions checked in should_advance_cwnd()
2145 * before the call to tcp_cong_avoid(). Mainly this means that
2146 * we only execute this code if the ack actually acked some
2150 /* If we are in slow start, increase our cwnd in response to this ACK.
2151 * (If we are not in slow start then we are in congestion avoidance,
2152 * and adjust our congestion window only once per RTT. See the code
2155 if (tp->snd_cwnd <= tp->snd_ssthresh)
2158 /* to keep cwnd from growing without bound */
2159 tp->snd_cwnd = min_t(u32, tp->snd_cwnd, tp->snd_cwnd_clamp);
2161 /* Make sure that we are never so timid as to reduce our cwnd below
2164 * Going below 2 MSS would risk huge delayed ACKs from our receiver.
2166 tp->snd_cwnd = max(tp->snd_cwnd, 2U);
2168 tp->snd_cwnd_stamp = tcp_time_stamp;
2171 static inline void tcp_cong_avoid(struct tcp_opt *tp, u32 ack, u32 seq_rtt)
2173 if (tcp_vegas_enabled(tp))
2174 vegas_cong_avoid(tp, ack, seq_rtt);
2176 reno_cong_avoid(tp);
2179 /* Restart timer after forward progress on connection.
2180 * RFC2988 recommends to restart timer to now+rto.
2183 static __inline__ void tcp_ack_packets_out(struct sock *sk, struct tcp_opt *tp)
2185 if (tp->packets_out==0) {
2186 tcp_clear_xmit_timer(sk, TCP_TIME_RETRANS);
2188 tcp_reset_xmit_timer(sk, TCP_TIME_RETRANS, tp->rto);
2192 /* Remove acknowledged frames from the retransmission queue. */
2193 static int tcp_clean_rtx_queue(struct sock *sk, __s32 *seq_rtt_p)
2195 struct tcp_opt *tp = tcp_sk(sk);
2196 struct sk_buff *skb;
2197 __u32 now = tcp_time_stamp;
2201 while ((skb = skb_peek(&sk->sk_write_queue)) && skb != tp->send_head) {
2202 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
2203 __u8 sacked = scb->sacked;
2205 /* If our packet is before the ack sequence we can
2206 * discard it as it's confirmed to have arrived at
2209 if (after(scb->end_seq, tp->snd_una))
2212 /* Initial outgoing SYN's get put onto the write_queue
2213 * just like anything else we transmit. It is not
2214 * true data, and if we misinform our callers that
2215 * this ACK acks real data, we will erroneously exit
2216 * connection startup slow start one packet too
2217 * quickly. This is severely frowned upon behavior.
2219 if(!(scb->flags & TCPCB_FLAG_SYN)) {
2220 acked |= FLAG_DATA_ACKED;
2222 acked |= FLAG_SYN_ACKED;
2223 tp->retrans_stamp = 0;
2227 if(sacked & TCPCB_RETRANS) {
2228 if(sacked & TCPCB_SACKED_RETRANS)
2230 acked |= FLAG_RETRANS_DATA_ACKED;
2232 } else if (seq_rtt < 0)
2233 seq_rtt = now - scb->when;
2234 if(sacked & TCPCB_SACKED_ACKED)
2236 if(sacked & TCPCB_LOST)
2238 if(sacked & TCPCB_URG) {
2240 !before(scb->end_seq, tp->snd_up))
2243 } else if (seq_rtt < 0)
2244 seq_rtt = now - scb->when;
2245 if (tp->fackets_out)
2248 __skb_unlink(skb, skb->list);
2249 tcp_free_skb(sk, skb);
2252 if (acked&FLAG_ACKED) {
2253 tcp_ack_update_rtt(tp, acked, seq_rtt);
2254 tcp_ack_packets_out(sk, tp);
2257 #if FASTRETRANS_DEBUG > 0
2258 BUG_TRAP((int)tp->sacked_out >= 0);
2259 BUG_TRAP((int)tp->lost_out >= 0);
2260 BUG_TRAP((int)tp->retrans_out >= 0);
2261 if (!tp->packets_out && tp->sack_ok) {
2263 printk(KERN_DEBUG "Leak l=%u %d\n", tp->lost_out,
2267 if (tp->sacked_out) {
2268 printk(KERN_DEBUG "Leak s=%u %d\n", tp->sacked_out,
2272 if (tp->retrans_out) {
2273 printk(KERN_DEBUG "Leak r=%u %d\n", tp->retrans_out,
2275 tp->retrans_out = 0;
2279 *seq_rtt_p = seq_rtt;
2283 static void tcp_ack_probe(struct sock *sk)
2285 struct tcp_opt *tp = tcp_sk(sk);
2287 /* Was it a usable window open? */
2289 if (!after(TCP_SKB_CB(tp->send_head)->end_seq,
2290 tp->snd_una + tp->snd_wnd)) {
2292 tcp_clear_xmit_timer(sk, TCP_TIME_PROBE0);
2293 /* Socket must be waked up by subsequent tcp_data_snd_check().
2294 * This function is not for random using!
2297 tcp_reset_xmit_timer(sk, TCP_TIME_PROBE0,
2298 min(tp->rto << tp->backoff, TCP_RTO_MAX));
2302 static __inline__ int tcp_ack_is_dubious(struct tcp_opt *tp, int flag)
2304 return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
2305 tp->ca_state != TCP_CA_Open);
2308 static __inline__ int tcp_may_raise_cwnd(struct tcp_opt *tp, int flag)
2310 return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
2311 !((1<<tp->ca_state)&(TCPF_CA_Recovery|TCPF_CA_CWR));
2314 /* Check that window update is acceptable.
2315 * The function assumes that snd_una<=ack<=snd_next.
2317 static __inline__ int
2318 tcp_may_update_window(struct tcp_opt *tp, u32 ack, u32 ack_seq, u32 nwin)
2320 return (after(ack, tp->snd_una) ||
2321 after(ack_seq, tp->snd_wl1) ||
2322 (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd));
2325 /* Update our send window.
2327 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
2328 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
2330 static int tcp_ack_update_window(struct sock *sk, struct tcp_opt *tp,
2331 struct sk_buff *skb, u32 ack, u32 ack_seq)
2334 u32 nwin = ntohs(skb->h.th->window);
2336 if (likely(!skb->h.th->syn))
2337 nwin <<= tp->snd_wscale;
2339 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
2340 flag |= FLAG_WIN_UPDATE;
2341 tcp_update_wl(tp, ack, ack_seq);
2343 if (tp->snd_wnd != nwin) {
2346 /* Note, it is the only place, where
2347 * fast path is recovered for sending TCP.
2349 tcp_fast_path_check(sk, tp);
2351 if (nwin > tp->max_window) {
2352 tp->max_window = nwin;
2353 tcp_sync_mss(sk, tp->pmtu_cookie);
2363 static void tcp_process_frto(struct sock *sk, u32 prior_snd_una)
2365 struct tcp_opt *tp = tcp_sk(sk);
2367 tcp_sync_left_out(tp);
2369 if (tp->snd_una == prior_snd_una ||
2370 !before(tp->snd_una, tp->frto_highmark)) {
2371 /* RTO was caused by loss, start retransmitting in
2372 * go-back-N slow start
2374 tcp_enter_frto_loss(sk);
2378 if (tp->frto_counter == 1) {
2379 /* First ACK after RTO advances the window: allow two new
2382 tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
2384 /* Also the second ACK after RTO advances the window.
2385 * The RTO was likely spurious. Reduce cwnd and continue
2386 * in congestion avoidance
2388 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
2389 tcp_moderate_cwnd(tp);
2392 /* F-RTO affects on two new ACKs following RTO.
2393 * At latest on third ACK the TCP behavor is back to normal.
2395 tp->frto_counter = (tp->frto_counter + 1) % 3;
2404 * This function initializes fields used in TCP Westwood+. We can't
2405 * get no information about RTTmin at this time so we simply set it to
2406 * TCP_WESTWOOD_INIT_RTT. This value was chosen to be too conservative
2407 * since in this way we're sure it will be updated in a consistent
2408 * way as soon as possible. It will reasonably happen within the first
2409 * RTT period of the connection lifetime.
2412 static void init_westwood(struct sock *sk)
2414 struct tcp_opt *tp = tcp_sk(sk);
2416 tp->westwood.bw_ns_est = 0;
2417 tp->westwood.bw_est = 0;
2418 tp->westwood.accounted = 0;
2419 tp->westwood.cumul_ack = 0;
2420 tp->westwood.rtt_win_sx = tcp_time_stamp;
2421 tp->westwood.rtt = TCP_WESTWOOD_INIT_RTT;
2422 tp->westwood.rtt_min = TCP_WESTWOOD_INIT_RTT;
2423 tp->westwood.snd_una = tp->snd_una;
2427 * @westwood_do_filter
2428 * Low-pass filter. Implemented using constant coeffients.
2431 static inline __u32 westwood_do_filter(__u32 a, __u32 b)
2433 return (((7 * a) + b) >> 3);
2436 static void westwood_filter(struct sock *sk, __u32 delta)
2438 struct tcp_opt *tp = tcp_sk(sk);
2440 tp->westwood.bw_ns_est =
2441 westwood_do_filter(tp->westwood.bw_ns_est,
2442 tp->westwood.bk / delta);
2443 tp->westwood.bw_est =
2444 westwood_do_filter(tp->westwood.bw_est,
2445 tp->westwood.bw_ns_est);
2449 * @westwood_update_rttmin
2450 * It is used to update RTTmin. In this case we MUST NOT use
2451 * WESTWOOD_RTT_MIN minimum bound since we could be on a LAN!
2454 static inline __u32 westwood_update_rttmin(struct sock *sk)
2456 struct tcp_opt *tp = tcp_sk(sk);
2457 __u32 rttmin = tp->westwood.rtt_min;
2459 if (tp->westwood.rtt == 0)
2462 if (tp->westwood.rtt < tp->westwood.rtt_min || !rttmin)
2463 rttmin = tp->westwood.rtt;
2470 * Evaluate increases for dk.
2473 static inline __u32 westwood_acked(struct sock *sk)
2475 struct tcp_opt *tp = tcp_sk(sk);
2477 return ((tp->snd_una) - (tp->westwood.snd_una));
2481 * @westwood_new_window
2482 * It evaluates if we are receiving data inside the same RTT window as
2485 * It returns 0 if we are still evaluating samples in the same RTT
2486 * window, 1 if the sample has to be considered in the next window.
2489 static int westwood_new_window(struct sock *sk)
2491 struct tcp_opt *tp = tcp_sk(sk);
2496 left_bound = tp->westwood.rtt_win_sx;
2497 rtt = max(tp->westwood.rtt, (u32) TCP_WESTWOOD_RTT_MIN);
2500 * A RTT-window has passed. Be careful since if RTT is less than
2501 * 50ms we don't filter but we continue 'building the sample'.
2502 * This minimum limit was choosen since an estimation on small
2503 * time intervals is better to avoid...
2504 * Obvioulsy on a LAN we reasonably will always have
2505 * right_bound = left_bound + WESTWOOD_RTT_MIN
2508 if ((left_bound + rtt) < tcp_time_stamp)
2515 * @westwood_update_window
2516 * It updates RTT evaluation window if it is the right moment to do
2517 * it. If so it calls filter for evaluating bandwidth.
2520 static void __westwood_update_window(struct sock *sk, __u32 now)
2522 struct tcp_opt *tp = tcp_sk(sk);
2523 __u32 delta = now - tp->westwood.rtt_win_sx;
2528 if (tp->westwood.rtt)
2529 westwood_filter(sk, delta);
2531 tp->westwood.bk = 0;
2532 tp->westwood.rtt_win_sx = tcp_time_stamp;
2536 static void westwood_update_window(struct sock *sk, __u32 now)
2538 if (westwood_new_window(sk))
2539 __westwood_update_window(sk, now);
2543 * @__tcp_westwood_fast_bw
2544 * It is called when we are in fast path. In particular it is called when
2545 * header prediction is successfull. In such case infact update is
2546 * straight forward and doesn't need any particular care.
2549 void __tcp_westwood_fast_bw(struct sock *sk, struct sk_buff *skb)
2551 struct tcp_opt *tp = tcp_sk(sk);
2553 westwood_update_window(sk, tcp_time_stamp);
2555 tp->westwood.bk += westwood_acked(sk);
2556 tp->westwood.snd_una = tp->snd_una;
2557 tp->westwood.rtt_min = westwood_update_rttmin(sk);
2562 * @westwood_dupack_update
2563 * It updates accounted and cumul_ack when receiving a dupack.
2566 static void westwood_dupack_update(struct sock *sk)
2568 struct tcp_opt *tp = tcp_sk(sk);
2570 tp->westwood.accounted += tp->mss_cache;
2571 tp->westwood.cumul_ack = tp->mss_cache;
2574 static inline int westwood_may_change_cumul(struct tcp_opt *tp)
2576 return ((tp->westwood.cumul_ack) > tp->mss_cache);
2579 static inline void westwood_partial_update(struct tcp_opt *tp)
2581 tp->westwood.accounted -= tp->westwood.cumul_ack;
2582 tp->westwood.cumul_ack = tp->mss_cache;
2585 static inline void westwood_complete_update(struct tcp_opt *tp)
2587 tp->westwood.cumul_ack -= tp->westwood.accounted;
2588 tp->westwood.accounted = 0;
2592 * @westwood_acked_count
2593 * This function evaluates cumul_ack for evaluating dk in case of
2594 * delayed or partial acks.
2597 static __u32 westwood_acked_count(struct sock *sk)
2599 struct tcp_opt *tp = tcp_sk(sk);
2601 tp->westwood.cumul_ack = westwood_acked(sk);
2603 /* If cumul_ack is 0 this is a dupack since it's not moving
2606 if (!(tp->westwood.cumul_ack))
2607 westwood_dupack_update(sk);
2609 if (westwood_may_change_cumul(tp)) {
2610 /* Partial or delayed ack */
2611 if ((tp->westwood.accounted) >= (tp->westwood.cumul_ack))
2612 westwood_partial_update(tp);
2614 westwood_complete_update(tp);
2617 tp->westwood.snd_una = tp->snd_una;
2619 return tp->westwood.cumul_ack;
2624 * @__tcp_westwood_slow_bw
2625 * It is called when something is going wrong..even if there could
2626 * be no problems! Infact a simple delayed packet may trigger a
2627 * dupack. But we need to be careful in such case.
2630 void __tcp_westwood_slow_bw(struct sock *sk, struct sk_buff *skb)
2632 struct tcp_opt *tp = tcp_sk(sk);
2634 westwood_update_window(sk, tcp_time_stamp);
2636 tp->westwood.bk += westwood_acked_count(sk);
2637 tp->westwood.rtt_min = westwood_update_rttmin(sk);
2640 /* This routine deals with incoming acks, but not outgoing ones. */
2641 static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag)
2643 struct tcp_opt *tp = tcp_sk(sk);
2644 u32 prior_snd_una = tp->snd_una;
2645 u32 ack_seq = TCP_SKB_CB(skb)->seq;
2646 u32 ack = TCP_SKB_CB(skb)->ack_seq;
2647 u32 prior_in_flight;
2651 /* If the ack is newer than sent or older than previous acks
2652 * then we can probably ignore it.
2654 if (after(ack, tp->snd_nxt))
2655 goto uninteresting_ack;
2657 if (before(ack, prior_snd_una))
2660 if (!(flag&FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
2661 /* Window is constant, pure forward advance.
2662 * No more checks are required.
2663 * Note, we use the fact that SND.UNA>=SND.WL2.
2665 tcp_update_wl(tp, ack, ack_seq);
2667 tcp_westwood_fast_bw(sk, skb);
2668 flag |= FLAG_WIN_UPDATE;
2670 NET_INC_STATS_BH(TCPHPAcks);
2672 if (ack_seq != TCP_SKB_CB(skb)->end_seq)
2675 NET_INC_STATS_BH(TCPPureAcks);
2677 flag |= tcp_ack_update_window(sk, tp, skb, ack, ack_seq);
2679 if (TCP_SKB_CB(skb)->sacked)
2680 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
2682 if (TCP_ECN_rcv_ecn_echo(tp, skb->h.th))
2685 tcp_westwood_slow_bw(sk,skb);
2688 /* We passed data and got it acked, remove any soft error
2689 * log. Something worked...
2691 sk->sk_err_soft = 0;
2692 tp->rcv_tstamp = tcp_time_stamp;
2693 prior_packets = tp->packets_out;
2697 prior_in_flight = tcp_packets_in_flight(tp);
2699 /* See if we can take anything off of the retransmit queue. */
2700 flag |= tcp_clean_rtx_queue(sk, &seq_rtt);
2702 if (tp->frto_counter)
2703 tcp_process_frto(sk, prior_snd_una);
2705 if (tcp_ack_is_dubious(tp, flag)) {
2706 /* Advanve CWND, if state allows this. */
2707 if ((flag & FLAG_DATA_ACKED) &&
2708 (tcp_vegas_enabled(tp) || prior_in_flight >= tp->snd_cwnd) &&
2709 tcp_may_raise_cwnd(tp, flag))
2710 tcp_cong_avoid(tp, ack, seq_rtt);
2711 tcp_fastretrans_alert(sk, prior_snd_una, prior_packets, flag);
2713 if ((flag & FLAG_DATA_ACKED) &&
2714 (tcp_vegas_enabled(tp) || prior_in_flight >= tp->snd_cwnd))
2715 tcp_cong_avoid(tp, ack, seq_rtt);
2718 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag&FLAG_NOT_DUP))
2719 dst_confirm(sk->sk_dst_cache);
2726 /* If this ack opens up a zero window, clear backoff. It was
2727 * being used to time the probes, and is probably far higher than
2728 * it needs to be for normal retransmission.
2735 if (TCP_SKB_CB(skb)->sacked)
2736 tcp_sacktag_write_queue(sk, skb, prior_snd_una);
2739 SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
2744 /* Look for tcp options. Normally only called on SYN and SYNACK packets.
2745 * But, this can also be called on packets in the established flow when
2746 * the fast version below fails.
2748 void tcp_parse_options(struct sk_buff *skb, struct tcp_opt *tp, int estab)
2751 struct tcphdr *th = skb->h.th;
2752 int length=(th->doff*4)-sizeof(struct tcphdr);
2754 ptr = (unsigned char *)(th + 1);
2764 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
2769 if (opsize < 2) /* "silly options" */
2771 if (opsize > length)
2772 return; /* don't parse partial options */
2775 if(opsize==TCPOLEN_MSS && th->syn && !estab) {
2776 u16 in_mss = ntohs(*(__u16 *)ptr);
2778 if (tp->user_mss && tp->user_mss < in_mss)
2779 in_mss = tp->user_mss;
2780 tp->mss_clamp = in_mss;
2785 if(opsize==TCPOLEN_WINDOW && th->syn && !estab)
2786 if (sysctl_tcp_window_scaling) {
2788 tp->snd_wscale = *(__u8 *)ptr;
2789 if(tp->snd_wscale > 14) {
2791 printk("tcp_parse_options: Illegal window "
2792 "scaling value %d >14 received.",
2794 tp->snd_wscale = 14;
2798 case TCPOPT_TIMESTAMP:
2799 if(opsize==TCPOLEN_TIMESTAMP) {
2800 if ((estab && tp->tstamp_ok) ||
2801 (!estab && sysctl_tcp_timestamps)) {
2803 tp->rcv_tsval = ntohl(*(__u32 *)ptr);
2804 tp->rcv_tsecr = ntohl(*(__u32 *)(ptr+4));
2808 case TCPOPT_SACK_PERM:
2809 if(opsize==TCPOLEN_SACK_PERM && th->syn && !estab) {
2810 if (sysctl_tcp_sack) {
2818 if((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
2819 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
2821 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
2830 /* Fast parse options. This hopes to only see timestamps.
2831 * If it is wrong it falls back on tcp_parse_options().
2833 static __inline__ int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th, struct tcp_opt *tp)
2835 if (th->doff == sizeof(struct tcphdr)>>2) {
2838 } else if (tp->tstamp_ok &&
2839 th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) {
2840 __u32 *ptr = (__u32 *)(th + 1);
2841 if (*ptr == ntohl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
2842 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
2845 tp->rcv_tsval = ntohl(*ptr);
2847 tp->rcv_tsecr = ntohl(*ptr);
2851 tcp_parse_options(skb, tp, 1);
2855 static __inline__ void
2856 tcp_store_ts_recent(struct tcp_opt *tp)
2858 tp->ts_recent = tp->rcv_tsval;
2859 tp->ts_recent_stamp = xtime.tv_sec;
2862 static __inline__ void
2863 tcp_replace_ts_recent(struct tcp_opt *tp, u32 seq)
2865 if (tp->saw_tstamp && !after(seq, tp->rcv_wup)) {
2866 /* PAWS bug workaround wrt. ACK frames, the PAWS discard
2867 * extra check below makes sure this can only happen
2868 * for pure ACK frames. -DaveM
2870 * Not only, also it occurs for expired timestamps.
2873 if((s32)(tp->rcv_tsval - tp->ts_recent) >= 0 ||
2874 xtime.tv_sec >= tp->ts_recent_stamp + TCP_PAWS_24DAYS)
2875 tcp_store_ts_recent(tp);
2879 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
2881 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
2882 * it can pass through stack. So, the following predicate verifies that
2883 * this segment is not used for anything but congestion avoidance or
2884 * fast retransmit. Moreover, we even are able to eliminate most of such
2885 * second order effects, if we apply some small "replay" window (~RTO)
2886 * to timestamp space.
2888 * All these measures still do not guarantee that we reject wrapped ACKs
2889 * on networks with high bandwidth, when sequence space is recycled fastly,
2890 * but it guarantees that such events will be very rare and do not affect
2891 * connection seriously. This doesn't look nice, but alas, PAWS is really
2894 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
2895 * states that events when retransmit arrives after original data are rare.
2896 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
2897 * the biggest problem on large power networks even with minor reordering.
2898 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
2899 * up to bandwidth of 18Gigabit/sec. 8) ]
2902 static int tcp_disordered_ack(struct tcp_opt *tp, struct sk_buff *skb)
2904 struct tcphdr *th = skb->h.th;
2905 u32 seq = TCP_SKB_CB(skb)->seq;
2906 u32 ack = TCP_SKB_CB(skb)->ack_seq;
2908 return (/* 1. Pure ACK with correct sequence number. */
2909 (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
2911 /* 2. ... and duplicate ACK. */
2912 ack == tp->snd_una &&
2914 /* 3. ... and does not update window. */
2915 !tcp_may_update_window(tp, ack, seq, ntohs(th->window)<<tp->snd_wscale) &&
2917 /* 4. ... and sits in replay window. */
2918 (s32)(tp->ts_recent - tp->rcv_tsval) <= (tp->rto*1024)/HZ);
2921 static __inline__ int tcp_paws_discard(struct tcp_opt *tp, struct sk_buff *skb)
2923 return ((s32)(tp->ts_recent - tp->rcv_tsval) > TCP_PAWS_WINDOW &&
2924 xtime.tv_sec < tp->ts_recent_stamp + TCP_PAWS_24DAYS &&
2925 !tcp_disordered_ack(tp, skb));
2928 /* Check segment sequence number for validity.
2930 * Segment controls are considered valid, if the segment
2931 * fits to the window after truncation to the window. Acceptability
2932 * of data (and SYN, FIN, of course) is checked separately.
2933 * See tcp_data_queue(), for example.
2935 * Also, controls (RST is main one) are accepted using RCV.WUP instead
2936 * of RCV.NXT. Peer still did not advance his SND.UNA when we
2937 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
2938 * (borrowed from freebsd)
2941 static inline int tcp_sequence(struct tcp_opt *tp, u32 seq, u32 end_seq)
2943 return !before(end_seq, tp->rcv_wup) &&
2944 !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
2947 /* When we get a reset we do this. */
2948 static void tcp_reset(struct sock *sk)
2950 /* We want the right error as BSD sees it (and indeed as we do). */
2951 switch (sk->sk_state) {
2953 sk->sk_err = ECONNREFUSED;
2955 case TCP_CLOSE_WAIT:
2961 sk->sk_err = ECONNRESET;
2964 if (!sock_flag(sk, SOCK_DEAD))
2965 sk->sk_error_report(sk);
2971 * Process the FIN bit. This now behaves as it is supposed to work
2972 * and the FIN takes effect when it is validly part of sequence
2973 * space. Not before when we get holes.
2975 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
2976 * (and thence onto LAST-ACK and finally, CLOSE, we never enter
2979 * If we are in FINWAIT-1, a received FIN indicates simultaneous
2980 * close and we go into CLOSING (and later onto TIME-WAIT)
2982 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
2984 static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th)
2986 struct tcp_opt *tp = tcp_sk(sk);
2988 tcp_schedule_ack(tp);
2990 sk->sk_shutdown |= RCV_SHUTDOWN;
2991 sock_set_flag(sk, SOCK_DONE);
2993 switch (sk->sk_state) {
2995 case TCP_ESTABLISHED:
2996 /* Move to CLOSE_WAIT */
2997 tcp_set_state(sk, TCP_CLOSE_WAIT);
2998 tp->ack.pingpong = 1;
3001 case TCP_CLOSE_WAIT:
3003 /* Received a retransmission of the FIN, do
3008 /* RFC793: Remain in the LAST-ACK state. */
3012 /* This case occurs when a simultaneous close
3013 * happens, we must ack the received FIN and
3014 * enter the CLOSING state.
3017 tcp_set_state(sk, TCP_CLOSING);
3020 /* Received a FIN -- send ACK and enter TIME_WAIT. */
3022 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
3025 /* Only TCP_LISTEN and TCP_CLOSE are left, in these
3026 * cases we should never reach this piece of code.
3028 printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n",
3029 __FUNCTION__, sk->sk_state);
3033 /* It _is_ possible, that we have something out-of-order _after_ FIN.
3034 * Probably, we should reset in this case. For now drop them.
3036 __skb_queue_purge(&tp->out_of_order_queue);
3039 tcp_mem_reclaim(sk);
3041 if (!sock_flag(sk, SOCK_DEAD)) {
3042 sk->sk_state_change(sk);
3044 /* Do not send POLL_HUP for half duplex close. */
3045 if (sk->sk_shutdown == SHUTDOWN_MASK ||
3046 sk->sk_state == TCP_CLOSE)
3047 sk_wake_async(sk, 1, POLL_HUP);
3049 sk_wake_async(sk, 1, POLL_IN);
3053 static __inline__ int
3054 tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, u32 end_seq)
3056 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
3057 if (before(seq, sp->start_seq))
3058 sp->start_seq = seq;
3059 if (after(end_seq, sp->end_seq))
3060 sp->end_seq = end_seq;
3066 static __inline__ void tcp_dsack_set(struct tcp_opt *tp, u32 seq, u32 end_seq)
3068 if (tp->sack_ok && sysctl_tcp_dsack) {
3069 if (before(seq, tp->rcv_nxt))
3070 NET_INC_STATS_BH(TCPDSACKOldSent);
3072 NET_INC_STATS_BH(TCPDSACKOfoSent);
3075 tp->duplicate_sack[0].start_seq = seq;
3076 tp->duplicate_sack[0].end_seq = end_seq;
3077 tp->eff_sacks = min(tp->num_sacks+1, 4-tp->tstamp_ok);
3081 static __inline__ void tcp_dsack_extend(struct tcp_opt *tp, u32 seq, u32 end_seq)
3084 tcp_dsack_set(tp, seq, end_seq);
3086 tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
3089 static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb)
3091 struct tcp_opt *tp = tcp_sk(sk);
3093 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
3094 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3095 NET_INC_STATS_BH(DelayedACKLost);
3096 tcp_enter_quickack_mode(tp);
3098 if (tp->sack_ok && sysctl_tcp_dsack) {
3099 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3101 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
3102 end_seq = tp->rcv_nxt;
3103 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, end_seq);
3110 /* These routines update the SACK block as out-of-order packets arrive or
3111 * in-order packets close up the sequence space.
3113 static void tcp_sack_maybe_coalesce(struct tcp_opt *tp)
3116 struct tcp_sack_block *sp = &tp->selective_acks[0];
3117 struct tcp_sack_block *swalk = sp+1;
3119 /* See if the recent change to the first SACK eats into
3120 * or hits the sequence space of other SACK blocks, if so coalesce.
3122 for (this_sack = 1; this_sack < tp->num_sacks; ) {
3123 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
3126 /* Zap SWALK, by moving every further SACK up by one slot.
3127 * Decrease num_sacks.
3130 tp->eff_sacks = min(tp->num_sacks+tp->dsack, 4-tp->tstamp_ok);
3131 for(i=this_sack; i < tp->num_sacks; i++)
3135 this_sack++, swalk++;
3139 static __inline__ void tcp_sack_swap(struct tcp_sack_block *sack1, struct tcp_sack_block *sack2)
3143 tmp = sack1->start_seq;
3144 sack1->start_seq = sack2->start_seq;
3145 sack2->start_seq = tmp;
3147 tmp = sack1->end_seq;
3148 sack1->end_seq = sack2->end_seq;
3149 sack2->end_seq = tmp;
3152 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
3154 struct tcp_opt *tp = tcp_sk(sk);
3155 struct tcp_sack_block *sp = &tp->selective_acks[0];
3156 int cur_sacks = tp->num_sacks;
3162 for (this_sack=0; this_sack<cur_sacks; this_sack++, sp++) {
3163 if (tcp_sack_extend(sp, seq, end_seq)) {
3164 /* Rotate this_sack to the first one. */
3165 for (; this_sack>0; this_sack--, sp--)
3166 tcp_sack_swap(sp, sp-1);
3168 tcp_sack_maybe_coalesce(tp);
3173 /* Could not find an adjacent existing SACK, build a new one,
3174 * put it at the front, and shift everyone else down. We
3175 * always know there is at least one SACK present already here.
3177 * If the sack array is full, forget about the last one.
3179 if (this_sack >= 4) {
3184 for(; this_sack > 0; this_sack--, sp--)
3188 /* Build the new head SACK, and we're done. */
3189 sp->start_seq = seq;
3190 sp->end_seq = end_seq;
3192 tp->eff_sacks = min(tp->num_sacks + tp->dsack, 4 - tp->tstamp_ok);
3195 /* RCV.NXT advances, some SACKs should be eaten. */
3197 static void tcp_sack_remove(struct tcp_opt *tp)
3199 struct tcp_sack_block *sp = &tp->selective_acks[0];
3200 int num_sacks = tp->num_sacks;
3203 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
3204 if (skb_queue_len(&tp->out_of_order_queue) == 0) {
3206 tp->eff_sacks = tp->dsack;
3210 for(this_sack = 0; this_sack < num_sacks; ) {
3211 /* Check if the start of the sack is covered by RCV.NXT. */
3212 if (!before(tp->rcv_nxt, sp->start_seq)) {
3215 /* RCV.NXT must cover all the block! */
3216 BUG_TRAP(!before(tp->rcv_nxt, sp->end_seq));
3218 /* Zap this SACK, by moving forward any other SACKS. */
3219 for (i=this_sack+1; i < num_sacks; i++)
3220 tp->selective_acks[i-1] = tp->selective_acks[i];
3227 if (num_sacks != tp->num_sacks) {
3228 tp->num_sacks = num_sacks;
3229 tp->eff_sacks = min(tp->num_sacks+tp->dsack, 4-tp->tstamp_ok);
3233 /* This one checks to see if we can put data from the
3234 * out_of_order queue into the receive_queue.
3236 static void tcp_ofo_queue(struct sock *sk)
3238 struct tcp_opt *tp = tcp_sk(sk);
3239 __u32 dsack_high = tp->rcv_nxt;
3240 struct sk_buff *skb;
3242 while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
3243 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
3246 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
3247 __u32 dsack = dsack_high;
3248 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
3249 dsack_high = TCP_SKB_CB(skb)->end_seq;
3250 tcp_dsack_extend(tp, TCP_SKB_CB(skb)->seq, dsack);
3253 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3254 SOCK_DEBUG(sk, "ofo packet was already received \n");
3255 __skb_unlink(skb, skb->list);
3259 SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
3260 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3261 TCP_SKB_CB(skb)->end_seq);
3263 __skb_unlink(skb, skb->list);
3264 __skb_queue_tail(&sk->sk_receive_queue, skb);
3265 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3267 tcp_fin(skb, sk, skb->h.th);
3271 static inline int tcp_rmem_schedule(struct sock *sk, struct sk_buff *skb)
3273 return (int)skb->truesize <= sk->sk_forward_alloc ||
3274 tcp_mem_schedule(sk, skb->truesize, 1);
3277 static int tcp_prune_queue(struct sock *sk);
3279 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
3281 struct tcphdr *th = skb->h.th;
3282 struct tcp_opt *tp = tcp_sk(sk);
3285 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
3289 __skb_pull(skb, th->doff*4);
3291 TCP_ECN_accept_cwr(tp, skb);
3295 tp->eff_sacks = min_t(unsigned int, tp->num_sacks,
3299 /* Queue data for delivery to the user.
3300 * Packets in sequence go to the receive queue.
3301 * Out of sequence packets to the out_of_order_queue.
3303 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
3304 if (tcp_receive_window(tp) == 0)
3307 /* Ok. In sequence. In window. */
3308 if (tp->ucopy.task == current &&
3309 tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
3310 sock_owned_by_user(sk) && !tp->urg_data) {
3311 int chunk = min_t(unsigned int, skb->len,
3314 __set_current_state(TASK_RUNNING);
3317 if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
3318 tp->ucopy.len -= chunk;
3319 tp->copied_seq += chunk;
3320 eaten = (chunk == skb->len && !th->fin);
3328 (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3329 !tcp_rmem_schedule(sk, skb))) {
3330 if (tcp_prune_queue(sk) < 0 || !tcp_rmem_schedule(sk, skb))
3333 tcp_set_owner_r(skb, sk);
3334 __skb_queue_tail(&sk->sk_receive_queue, skb);
3336 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
3338 tcp_event_data_recv(sk, tp, skb);
3340 tcp_fin(skb, sk, th);
3342 if (skb_queue_len(&tp->out_of_order_queue)) {
3345 /* RFC2581. 4.2. SHOULD send immediate ACK, when
3346 * gap in queue is filled.
3348 if (!skb_queue_len(&tp->out_of_order_queue))
3349 tp->ack.pingpong = 0;
3353 tcp_sack_remove(tp);
3355 tcp_fast_path_check(sk, tp);
3359 else if (!sock_flag(sk, SOCK_DEAD))
3360 sk->sk_data_ready(sk, 0);
3364 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
3365 /* A retransmit, 2nd most common case. Force an immediate ack. */
3366 NET_INC_STATS_BH(DelayedACKLost);
3367 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3370 tcp_enter_quickack_mode(tp);
3371 tcp_schedule_ack(tp);
3377 /* Out of window. F.e. zero window probe. */
3378 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
3381 tcp_enter_quickack_mode(tp);
3383 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3384 /* Partial packet, seq < rcv_next < end_seq */
3385 SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
3386 tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
3387 TCP_SKB_CB(skb)->end_seq);
3389 tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
3391 /* If window is closed, drop tail of packet. But after
3392 * remembering D-SACK for its head made in previous line.
3394 if (!tcp_receive_window(tp))
3399 TCP_ECN_check_ce(tp, skb);
3401 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
3402 !tcp_rmem_schedule(sk, skb)) {
3403 if (tcp_prune_queue(sk) < 0 || !tcp_rmem_schedule(sk, skb))
3407 /* Disable header prediction. */
3409 tcp_schedule_ack(tp);
3411 SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
3412 tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
3414 tcp_set_owner_r(skb, sk);
3416 if (!skb_peek(&tp->out_of_order_queue)) {
3417 /* Initial out of order segment, build 1 SACK. */
3422 tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
3423 tp->selective_acks[0].end_seq =
3424 TCP_SKB_CB(skb)->end_seq;
3426 __skb_queue_head(&tp->out_of_order_queue,skb);
3428 struct sk_buff *skb1 = tp->out_of_order_queue.prev;
3429 u32 seq = TCP_SKB_CB(skb)->seq;
3430 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3432 if (seq == TCP_SKB_CB(skb1)->end_seq) {
3433 __skb_append(skb1, skb);
3435 if (!tp->num_sacks ||
3436 tp->selective_acks[0].end_seq != seq)
3439 /* Common case: data arrive in order after hole. */
3440 tp->selective_acks[0].end_seq = end_seq;
3444 /* Find place to insert this segment. */
3446 if (!after(TCP_SKB_CB(skb1)->seq, seq))
3448 } while ((skb1 = skb1->prev) !=
3449 (struct sk_buff*)&tp->out_of_order_queue);
3451 /* Do skb overlap to previous one? */
3452 if (skb1 != (struct sk_buff*)&tp->out_of_order_queue &&
3453 before(seq, TCP_SKB_CB(skb1)->end_seq)) {
3454 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
3455 /* All the bits are present. Drop. */
3457 tcp_dsack_set(tp, seq, end_seq);
3460 if (after(seq, TCP_SKB_CB(skb1)->seq)) {
3461 /* Partial overlap. */
3462 tcp_dsack_set(tp, seq, TCP_SKB_CB(skb1)->end_seq);
3467 __skb_insert(skb, skb1, skb1->next, &tp->out_of_order_queue);
3469 /* And clean segments covered by new one as whole. */
3470 while ((skb1 = skb->next) !=
3471 (struct sk_buff*)&tp->out_of_order_queue &&
3472 after(end_seq, TCP_SKB_CB(skb1)->seq)) {
3473 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
3474 tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, end_seq);
3477 __skb_unlink(skb1, skb1->list);
3478 tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, TCP_SKB_CB(skb1)->end_seq);
3484 tcp_sack_new_ofo_skb(sk, seq, end_seq);
3488 /* Collapse contiguous sequence of skbs head..tail with
3489 * sequence numbers start..end.
3490 * Segments with FIN/SYN are not collapsed (only because this
3494 tcp_collapse(struct sock *sk, struct sk_buff *head,
3495 struct sk_buff *tail, u32 start, u32 end)
3497 struct sk_buff *skb;
3499 /* First, check that queue is collapsable and find
3500 * the point where collapsing can be useful. */
3501 for (skb = head; skb != tail; ) {
3502 /* No new bits? It is possible on ofo queue. */
3503 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
3504 struct sk_buff *next = skb->next;
3505 __skb_unlink(skb, skb->list);
3507 NET_INC_STATS_BH(TCPRcvCollapsed);
3512 /* The first skb to collapse is:
3514 * - bloated or contains data before "start" or
3515 * overlaps to the next one.
3517 if (!skb->h.th->syn && !skb->h.th->fin &&
3518 (tcp_win_from_space(skb->truesize) > skb->len ||
3519 before(TCP_SKB_CB(skb)->seq, start) ||
3520 (skb->next != tail &&
3521 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq)))
3524 /* Decided to skip this, advance start seq. */
3525 start = TCP_SKB_CB(skb)->end_seq;
3528 if (skb == tail || skb->h.th->syn || skb->h.th->fin)
3531 while (before(start, end)) {
3532 struct sk_buff *nskb;
3533 int header = skb_headroom(skb);
3534 int copy = (PAGE_SIZE - sizeof(struct sk_buff) -
3535 sizeof(struct skb_shared_info) - header - 31)&~15;
3537 /* Too big header? This can happen with IPv6. */
3540 if (end-start < copy)
3542 nskb = alloc_skb(copy+header, GFP_ATOMIC);
3545 skb_reserve(nskb, header);
3546 memcpy(nskb->head, skb->head, header);
3547 nskb->nh.raw = nskb->head + (skb->nh.raw-skb->head);
3548 nskb->h.raw = nskb->head + (skb->h.raw-skb->head);
3549 nskb->mac.raw = nskb->head + (skb->mac.raw-skb->head);
3550 memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
3551 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
3552 __skb_insert(nskb, skb->prev, skb, skb->list);
3553 tcp_set_owner_r(nskb, sk);
3555 /* Copy data, releasing collapsed skbs. */
3557 int offset = start - TCP_SKB_CB(skb)->seq;
3558 int size = TCP_SKB_CB(skb)->end_seq - start;
3560 if (offset < 0) BUG();
3562 size = min(copy, size);
3563 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
3565 TCP_SKB_CB(nskb)->end_seq += size;
3569 if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
3570 struct sk_buff *next = skb->next;
3571 __skb_unlink(skb, skb->list);
3573 NET_INC_STATS_BH(TCPRcvCollapsed);
3575 if (skb == tail || skb->h.th->syn || skb->h.th->fin)
3582 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
3583 * and tcp_collapse() them until all the queue is collapsed.
3585 static void tcp_collapse_ofo_queue(struct sock *sk)
3587 struct tcp_opt *tp = tcp_sk(sk);
3588 struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
3589 struct sk_buff *head;
3595 start = TCP_SKB_CB(skb)->seq;
3596 end = TCP_SKB_CB(skb)->end_seq;
3602 /* Segment is terminated when we see gap or when
3603 * we are at the end of all the queue. */
3604 if (skb == (struct sk_buff *)&tp->out_of_order_queue ||
3605 after(TCP_SKB_CB(skb)->seq, end) ||
3606 before(TCP_SKB_CB(skb)->end_seq, start)) {
3607 tcp_collapse(sk, head, skb, start, end);
3609 if (skb == (struct sk_buff *)&tp->out_of_order_queue)
3611 /* Start new segment */
3612 start = TCP_SKB_CB(skb)->seq;
3613 end = TCP_SKB_CB(skb)->end_seq;
3615 if (before(TCP_SKB_CB(skb)->seq, start))
3616 start = TCP_SKB_CB(skb)->seq;
3617 if (after(TCP_SKB_CB(skb)->end_seq, end))
3618 end = TCP_SKB_CB(skb)->end_seq;
3623 /* Reduce allocated memory if we can, trying to get
3624 * the socket within its memory limits again.
3626 * Return less than zero if we should start dropping frames
3627 * until the socket owning process reads some of the data
3628 * to stabilize the situation.
3630 static int tcp_prune_queue(struct sock *sk)
3632 struct tcp_opt *tp = tcp_sk(sk);
3634 SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
3636 NET_INC_STATS_BH(PruneCalled);
3638 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
3639 tcp_clamp_window(sk, tp);
3640 else if (tcp_memory_pressure)
3641 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
3643 tcp_collapse_ofo_queue(sk);
3644 tcp_collapse(sk, sk->sk_receive_queue.next,
3645 (struct sk_buff*)&sk->sk_receive_queue,
3646 tp->copied_seq, tp->rcv_nxt);
3647 tcp_mem_reclaim(sk);
3649 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
3652 /* Collapsing did not help, destructive actions follow.
3653 * This must not ever occur. */
3655 /* First, purge the out_of_order queue. */
3656 if (skb_queue_len(&tp->out_of_order_queue)) {
3657 NET_ADD_STATS_BH(OfoPruned,
3658 skb_queue_len(&tp->out_of_order_queue));
3659 __skb_queue_purge(&tp->out_of_order_queue);
3661 /* Reset SACK state. A conforming SACK implementation will
3662 * do the same at a timeout based retransmit. When a connection
3663 * is in a sad state like this, we care only about integrity
3664 * of the connection not performance.
3668 tcp_mem_reclaim(sk);
3671 if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
3674 /* If we are really being abused, tell the caller to silently
3675 * drop receive data on the floor. It will get retransmitted
3676 * and hopefully then we'll have sufficient space.
3678 NET_INC_STATS_BH(RcvPruned);
3680 /* Massive buffer overcommit. */
3686 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
3687 * As additional protections, we do not touch cwnd in retransmission phases,
3688 * and if application hit its sndbuf limit recently.
3690 void tcp_cwnd_application_limited(struct sock *sk)
3692 struct tcp_opt *tp = tcp_sk(sk);
3694 if (tp->ca_state == TCP_CA_Open &&
3695 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
3696 /* Limited by application or receiver window. */
3697 u32 win_used = max(tp->snd_cwnd_used, 2U);
3698 if (win_used < tp->snd_cwnd) {
3699 tp->snd_ssthresh = tcp_current_ssthresh(tp);
3700 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
3702 tp->snd_cwnd_used = 0;
3704 tp->snd_cwnd_stamp = tcp_time_stamp;
3708 /* When incoming ACK allowed to free some skb from write_queue,
3709 * we remember this event in flag tp->queue_shrunk and wake up socket
3710 * on the exit from tcp input handler.
3712 * PROBLEM: sndbuf expansion does not work well with largesend.
3714 static void tcp_new_space(struct sock *sk)
3716 struct tcp_opt *tp = tcp_sk(sk);
3718 if (tp->packets_out < tp->snd_cwnd &&
3719 !(sk->sk_userlocks & SOCK_SNDBUF_LOCK) &&
3720 !tcp_memory_pressure &&
3721 atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) {
3722 int sndmem = max_t(u32, tp->mss_clamp, tp->mss_cache) +
3723 MAX_TCP_HEADER + 16 + sizeof(struct sk_buff),
3724 demanded = max_t(unsigned int, tp->snd_cwnd,
3725 tp->reordering + 1);
3726 sndmem *= 2*demanded;
3727 if (sndmem > sk->sk_sndbuf)
3728 sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
3729 tp->snd_cwnd_stamp = tcp_time_stamp;
3732 sk->sk_write_space(sk);
3735 static inline void tcp_check_space(struct sock *sk)
3737 struct tcp_opt *tp = tcp_sk(sk);
3739 if (tp->queue_shrunk) {
3740 tp->queue_shrunk = 0;
3741 if (sk->sk_socket &&
3742 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
3747 static void __tcp_data_snd_check(struct sock *sk, struct sk_buff *skb)
3749 struct tcp_opt *tp = tcp_sk(sk);
3751 if (after(TCP_SKB_CB(skb)->end_seq, tp->snd_una + tp->snd_wnd) ||
3752 tcp_packets_in_flight(tp) >= tp->snd_cwnd ||
3753 tcp_write_xmit(sk, tp->nonagle))
3754 tcp_check_probe_timer(sk, tp);
3757 static __inline__ void tcp_data_snd_check(struct sock *sk)
3759 struct tcp_opt *tp = tcp_sk(sk);
3760 struct sk_buff *skb = tp->send_head;
3763 __tcp_data_snd_check(sk, skb);
3764 tcp_check_space(sk);
3768 * Check if sending an ack is needed.
3770 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
3772 struct tcp_opt *tp = tcp_sk(sk);
3774 /* More than one full frame received... */
3775 if (((tp->rcv_nxt - tp->rcv_wup) > tp->ack.rcv_mss
3776 /* ... and right edge of window advances far enough.
3777 * (tcp_recvmsg() will send ACK otherwise). Or...
3779 && __tcp_select_window(sk) >= tp->rcv_wnd) ||
3780 /* We ACK each frame or... */
3781 tcp_in_quickack_mode(tp) ||
3782 /* We have out of order data. */
3784 skb_peek(&tp->out_of_order_queue))) {
3785 /* Then ack it now */
3788 /* Else, send delayed ack. */
3789 tcp_send_delayed_ack(sk);
3793 static __inline__ void tcp_ack_snd_check(struct sock *sk)
3795 struct tcp_opt *tp = tcp_sk(sk);
3796 if (!tcp_ack_scheduled(tp)) {
3797 /* We sent a data segment already. */
3800 __tcp_ack_snd_check(sk, 1);
3804 * This routine is only called when we have urgent data
3805 * signalled. Its the 'slow' part of tcp_urg. It could be
3806 * moved inline now as tcp_urg is only called from one
3807 * place. We handle URGent data wrong. We have to - as
3808 * BSD still doesn't use the correction from RFC961.
3809 * For 1003.1g we should support a new option TCP_STDURG to permit
3810 * either form (or just set the sysctl tcp_stdurg).
3813 static void tcp_check_urg(struct sock * sk, struct tcphdr * th)
3815 struct tcp_opt *tp = tcp_sk(sk);
3816 u32 ptr = ntohs(th->urg_ptr);
3818 if (ptr && !sysctl_tcp_stdurg)
3820 ptr += ntohl(th->seq);
3822 /* Ignore urgent data that we've already seen and read. */
3823 if (after(tp->copied_seq, ptr))
3826 /* Do not replay urg ptr.
3828 * NOTE: interesting situation not covered by specs.
3829 * Misbehaving sender may send urg ptr, pointing to segment,
3830 * which we already have in ofo queue. We are not able to fetch
3831 * such data and will stay in TCP_URG_NOTYET until will be eaten
3832 * by recvmsg(). Seems, we are not obliged to handle such wicked
3833 * situations. But it is worth to think about possibility of some
3834 * DoSes using some hypothetical application level deadlock.
3836 if (before(ptr, tp->rcv_nxt))
3839 /* Do we already have a newer (or duplicate) urgent pointer? */
3840 if (tp->urg_data && !after(ptr, tp->urg_seq))
3843 /* Tell the world about our new urgent pointer. */
3846 /* We may be adding urgent data when the last byte read was
3847 * urgent. To do this requires some care. We cannot just ignore
3848 * tp->copied_seq since we would read the last urgent byte again
3849 * as data, nor can we alter copied_seq until this data arrives
3850 * or we break the sematics of SIOCATMARK (and thus sockatmark())
3852 * NOTE. Double Dutch. Rendering to plain English: author of comment
3853 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
3854 * and expect that both A and B disappear from stream. This is _wrong_.
3855 * Though this happens in BSD with high probability, this is occasional.
3856 * Any application relying on this is buggy. Note also, that fix "works"
3857 * only in this artificial test. Insert some normal data between A and B and we will
3858 * decline of BSD again. Verdict: it is better to remove to trap
3861 if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
3862 !sock_flag(sk, SOCK_URGINLINE) &&
3863 tp->copied_seq != tp->rcv_nxt) {
3864 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
3866 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
3867 __skb_unlink(skb, skb->list);
3872 tp->urg_data = TCP_URG_NOTYET;
3875 /* Disable header prediction. */
3879 /* This is the 'fast' part of urgent handling. */
3880 static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th)
3882 struct tcp_opt *tp = tcp_sk(sk);
3884 /* Check if we get a new urgent pointer - normally not. */
3886 tcp_check_urg(sk,th);
3888 /* Do we wait for any urgent data? - normally not... */
3889 if (tp->urg_data == TCP_URG_NOTYET) {
3890 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
3893 /* Is the urgent pointer pointing into this packet? */
3894 if (ptr < skb->len) {
3896 if (skb_copy_bits(skb, ptr, &tmp, 1))
3898 tp->urg_data = TCP_URG_VALID | tmp;
3899 if (!sock_flag(sk, SOCK_DEAD))
3900 sk->sk_data_ready(sk, 0);
3905 static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
3907 struct tcp_opt *tp = tcp_sk(sk);
3908 int chunk = skb->len - hlen;
3912 if (skb->ip_summed==CHECKSUM_UNNECESSARY)
3913 err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
3915 err = skb_copy_and_csum_datagram_iovec(skb, hlen,
3919 tp->ucopy.len -= chunk;
3920 tp->copied_seq += chunk;
3927 static int __tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
3931 if (sock_owned_by_user(sk)) {
3933 result = __tcp_checksum_complete(skb);
3936 result = __tcp_checksum_complete(skb);
3941 static __inline__ int
3942 tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
3944 return skb->ip_summed != CHECKSUM_UNNECESSARY &&
3945 __tcp_checksum_complete_user(sk, skb);
3949 * TCP receive function for the ESTABLISHED state.
3951 * It is split into a fast path and a slow path. The fast path is
3953 * - A zero window was announced from us - zero window probing
3954 * is only handled properly in the slow path.
3955 * - Out of order segments arrived.
3956 * - Urgent data is expected.
3957 * - There is no buffer space left
3958 * - Unexpected TCP flags/window values/header lengths are received
3959 * (detected by checking the TCP header against pred_flags)
3960 * - Data is sent in both directions. Fast path only supports pure senders
3961 * or pure receivers (this means either the sequence number or the ack
3962 * value must stay constant)
3963 * - Unexpected TCP option.
3965 * When these conditions are not satisfied it drops into a standard
3966 * receive procedure patterned after RFC793 to handle all cases.
3967 * The first three cases are guaranteed by proper pred_flags setting,
3968 * the rest is checked inline. Fast processing is turned on in
3969 * tcp_data_queue when everything is OK.
3971 int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
3972 struct tcphdr *th, unsigned len)
3974 struct tcp_opt *tp = tcp_sk(sk);
3977 * Header prediction.
3978 * The code loosely follows the one in the famous
3979 * "30 instruction TCP receive" Van Jacobson mail.
3981 * Van's trick is to deposit buffers into socket queue
3982 * on a device interrupt, to call tcp_recv function
3983 * on the receive process context and checksum and copy
3984 * the buffer to user space. smart...
3986 * Our current scheme is not silly either but we take the
3987 * extra cost of the net_bh soft interrupt processing...
3988 * We do checksum and copy also but from device to kernel.
3993 /* pred_flags is 0xS?10 << 16 + snd_wnd
3994 * if header_predition is to be made
3995 * 'S' will always be tp->tcp_header_len >> 2
3996 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
3997 * turn it off (when there are holes in the receive
3998 * space for instance)
3999 * PSH flag is ignored.
4002 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
4003 TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4004 int tcp_header_len = tp->tcp_header_len;
4006 /* Timestamp header prediction: tcp_header_len
4007 * is automatically equal to th->doff*4 due to pred_flags
4011 /* Check timestamp */
4012 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
4013 __u32 *ptr = (__u32 *)(th + 1);
4015 /* No? Slow path! */
4016 if (*ptr != ntohl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
4017 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP))
4022 tp->rcv_tsval = ntohl(*ptr);
4024 tp->rcv_tsecr = ntohl(*ptr);
4026 /* If PAWS failed, check it more carefully in slow path */
4027 if ((s32)(tp->rcv_tsval - tp->ts_recent) < 0)
4030 /* DO NOT update ts_recent here, if checksum fails
4031 * and timestamp was corrupted part, it will result
4032 * in a hung connection since we will drop all
4033 * future packets due to the PAWS test.
4037 if (len <= tcp_header_len) {
4038 /* Bulk data transfer: sender */
4039 if (len == tcp_header_len) {
4040 /* Predicted packet is in window by definition.
4041 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4042 * Hence, check seq<=rcv_wup reduces to:
4044 if (tcp_header_len ==
4045 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
4046 tp->rcv_nxt == tp->rcv_wup)
4047 tcp_store_ts_recent(tp);
4048 /* We know that such packets are checksummed
4051 tcp_ack(sk, skb, 0);
4053 tcp_data_snd_check(sk);
4055 } else { /* Header too small */
4056 TCP_INC_STATS_BH(TcpInErrs);
4062 if (tp->ucopy.task == current &&
4063 tp->copied_seq == tp->rcv_nxt &&
4064 len - tcp_header_len <= tp->ucopy.len &&
4065 sock_owned_by_user(sk)) {
4066 __set_current_state(TASK_RUNNING);
4068 if (!tcp_copy_to_iovec(sk, skb, tcp_header_len)) {
4069 /* Predicted packet is in window by definition.
4070 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4071 * Hence, check seq<=rcv_wup reduces to:
4073 if (tcp_header_len ==
4074 (sizeof(struct tcphdr) +
4075 TCPOLEN_TSTAMP_ALIGNED) &&
4076 tp->rcv_nxt == tp->rcv_wup)
4077 tcp_store_ts_recent(tp);
4079 __skb_pull(skb, tcp_header_len);
4080 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4081 NET_INC_STATS_BH(TCPHPHitsToUser);
4086 if (tcp_checksum_complete_user(sk, skb))
4089 /* Predicted packet is in window by definition.
4090 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
4091 * Hence, check seq<=rcv_wup reduces to:
4093 if (tcp_header_len ==
4094 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
4095 tp->rcv_nxt == tp->rcv_wup)
4096 tcp_store_ts_recent(tp);
4098 if ((int)skb->truesize > sk->sk_forward_alloc)
4101 NET_INC_STATS_BH(TCPHPHits);
4103 /* Bulk data transfer: receiver */
4104 __skb_pull(skb,tcp_header_len);
4105 __skb_queue_tail(&sk->sk_receive_queue, skb);
4106 tcp_set_owner_r(skb, sk);
4107 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4110 tcp_event_data_recv(sk, tp, skb);
4112 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
4113 /* Well, only one small jumplet in fast path... */
4114 tcp_ack(sk, skb, FLAG_DATA);
4115 tcp_data_snd_check(sk);
4116 if (!tcp_ack_scheduled(tp))
4121 if (tcp_in_quickack_mode(tp)) {
4124 tcp_send_delayed_ack(sk);
4127 __tcp_ack_snd_check(sk, 0);
4134 sk->sk_data_ready(sk, 0);
4140 if (len < (th->doff<<2) || tcp_checksum_complete_user(sk, skb))
4144 * RFC1323: H1. Apply PAWS check first.
4146 if (tcp_fast_parse_options(skb, th, tp) && tp->saw_tstamp &&
4147 tcp_paws_discard(tp, skb)) {
4149 NET_INC_STATS_BH(PAWSEstabRejected);
4150 tcp_send_dupack(sk, skb);
4153 /* Resets are accepted even if PAWS failed.
4155 ts_recent update must be made after we are sure
4156 that the packet is in window.
4161 * Standard slow path.
4164 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
4165 /* RFC793, page 37: "In all states except SYN-SENT, all reset
4166 * (RST) segments are validated by checking their SEQ-fields."
4167 * And page 69: "If an incoming segment is not acceptable,
4168 * an acknowledgment should be sent in reply (unless the RST bit
4169 * is set, if so drop the segment and return)".
4172 tcp_send_dupack(sk, skb);
4181 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
4183 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4184 TCP_INC_STATS_BH(TcpInErrs);
4185 NET_INC_STATS_BH(TCPAbortOnSyn);
4192 tcp_ack(sk, skb, FLAG_SLOWPATH);
4194 /* Process urgent data. */
4195 tcp_urg(sk, skb, th);
4197 /* step 7: process the segment text */
4198 tcp_data_queue(sk, skb);
4200 tcp_data_snd_check(sk);
4201 tcp_ack_snd_check(sk);
4205 TCP_INC_STATS_BH(TcpInErrs);
4212 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
4213 struct tcphdr *th, unsigned len)
4215 struct tcp_opt *tp = tcp_sk(sk);
4216 int saved_clamp = tp->mss_clamp;
4218 tcp_parse_options(skb, tp, 0);
4222 * "If the state is SYN-SENT then
4223 * first check the ACK bit
4224 * If the ACK bit is set
4225 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
4226 * a reset (unless the RST bit is set, if so drop
4227 * the segment and return)"
4229 * We do not send data with SYN, so that RFC-correct
4232 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
4233 goto reset_and_undo;
4235 if (tp->saw_tstamp && tp->rcv_tsecr &&
4236 !between(tp->rcv_tsecr, tp->retrans_stamp,
4238 NET_INC_STATS_BH(PAWSActiveRejected);
4239 goto reset_and_undo;
4242 /* Now ACK is acceptable.
4244 * "If the RST bit is set
4245 * If the ACK was acceptable then signal the user "error:
4246 * connection reset", drop the segment, enter CLOSED state,
4247 * delete TCB, and return."
4256 * "fifth, if neither of the SYN or RST bits is set then
4257 * drop the segment and return."
4263 goto discard_and_undo;
4266 * "If the SYN bit is on ...
4267 * are acceptable then ...
4268 * (our SYN has been ACKed), change the connection
4269 * state to ESTABLISHED..."
4272 TCP_ECN_rcv_synack(tp, th);
4273 if (tp->ecn_flags&TCP_ECN_OK)
4274 sk->sk_no_largesend = 1;
4276 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
4277 tcp_ack(sk, skb, FLAG_SLOWPATH);
4279 /* Ok.. it's good. Set up sequence numbers and
4280 * move to established.
4282 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
4283 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
4285 /* RFC1323: The window in SYN & SYN/ACK segments is
4288 tp->snd_wnd = ntohs(th->window);
4289 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(skb)->seq);
4291 if (!tp->wscale_ok) {
4292 tp->snd_wscale = tp->rcv_wscale = 0;
4293 tp->window_clamp = min(tp->window_clamp, 65535U);
4296 if (tp->saw_tstamp) {
4298 tp->tcp_header_len =
4299 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
4300 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
4301 tcp_store_ts_recent(tp);
4303 tp->tcp_header_len = sizeof(struct tcphdr);
4306 if (tp->sack_ok && sysctl_tcp_fack)
4309 tcp_sync_mss(sk, tp->pmtu_cookie);
4310 tcp_initialize_rcv_mss(sk);
4312 /* Remember, tcp_poll() does not lock socket!
4313 * Change state from SYN-SENT only after copied_seq
4314 * is initialized. */
4315 tp->copied_seq = tp->rcv_nxt;
4317 tcp_set_state(sk, TCP_ESTABLISHED);
4319 /* Make sure socket is routed, for correct metrics. */
4320 tp->af_specific->rebuild_header(sk);
4322 tcp_init_metrics(sk);
4324 /* Prevent spurious tcp_cwnd_restart() on first data
4327 tp->lsndtime = tcp_time_stamp;
4329 tcp_init_buffer_space(sk);
4331 if (sock_flag(sk, SOCK_KEEPOPEN))
4332 tcp_reset_keepalive_timer(sk, keepalive_time_when(tp));
4334 if (!tp->snd_wscale)
4335 __tcp_fast_path_on(tp, tp->snd_wnd);
4339 if (!sock_flag(sk, SOCK_DEAD)) {
4340 sk->sk_state_change(sk);
4341 sk_wake_async(sk, 0, POLL_OUT);
4344 if (tp->write_pending || tp->defer_accept || tp->ack.pingpong) {
4345 /* Save one ACK. Data will be ready after
4346 * several ticks, if write_pending is set.
4348 * It may be deleted, but with this feature tcpdumps
4349 * look so _wonderfully_ clever, that I was not able
4350 * to stand against the temptation 8) --ANK
4352 tcp_schedule_ack(tp);
4353 tp->ack.lrcvtime = tcp_time_stamp;
4354 tp->ack.ato = TCP_ATO_MIN;
4355 tcp_incr_quickack(tp);
4356 tcp_enter_quickack_mode(tp);
4357 tcp_reset_xmit_timer(sk, TCP_TIME_DACK, TCP_DELACK_MAX);
4368 /* No ACK in the segment */
4372 * "If the RST bit is set
4374 * Otherwise (no ACK) drop the segment and return."
4377 goto discard_and_undo;
4381 if (tp->ts_recent_stamp && tp->saw_tstamp && tcp_paws_check(tp, 0))
4382 goto discard_and_undo;
4385 /* We see SYN without ACK. It is attempt of
4386 * simultaneous connect with crossed SYNs.
4387 * Particularly, it can be connect to self.
4389 tcp_set_state(sk, TCP_SYN_RECV);
4391 if (tp->saw_tstamp) {
4393 tcp_store_ts_recent(tp);
4394 tp->tcp_header_len =
4395 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
4397 tp->tcp_header_len = sizeof(struct tcphdr);
4400 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
4401 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
4403 /* RFC1323: The window in SYN & SYN/ACK segments is
4406 tp->snd_wnd = ntohs(th->window);
4407 tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
4408 tp->max_window = tp->snd_wnd;
4410 TCP_ECN_rcv_syn(tp, th);
4411 if (tp->ecn_flags&TCP_ECN_OK)
4412 sk->sk_no_largesend = 1;
4414 tcp_sync_mss(sk, tp->pmtu_cookie);
4415 tcp_initialize_rcv_mss(sk);
4418 tcp_send_synack(sk);
4420 /* Note, we could accept data and URG from this segment.
4421 * There are no obstacles to make this.
4423 * However, if we ignore data in ACKless segments sometimes,
4424 * we have no reasons to accept it sometimes.
4425 * Also, seems the code doing it in step6 of tcp_rcv_state_process
4426 * is not flawless. So, discard packet for sanity.
4427 * Uncomment this return to process the data.
4434 /* "fifth, if neither of the SYN or RST bits is set then
4435 * drop the segment and return."
4439 tcp_clear_options(tp);
4440 tp->mss_clamp = saved_clamp;
4444 tcp_clear_options(tp);
4445 tp->mss_clamp = saved_clamp;
4451 * This function implements the receiving procedure of RFC 793 for
4452 * all states except ESTABLISHED and TIME_WAIT.
4453 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
4454 * address independent.
4457 int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
4458 struct tcphdr *th, unsigned len)
4460 struct tcp_opt *tp = tcp_sk(sk);
4465 switch (sk->sk_state) {
4477 if(tp->af_specific->conn_request(sk, skb) < 0)
4482 /* Now we have several options: In theory there is
4483 * nothing else in the frame. KA9Q has an option to
4484 * send data with the syn, BSD accepts data with the
4485 * syn up to the [to be] advertised window and
4486 * Solaris 2.1 gives you a protocol error. For now
4487 * we just ignore it, that fits the spec precisely
4488 * and avoids incompatibilities. It would be nice in
4489 * future to drop through and process the data.
4491 * Now that TTCP is starting to be used we ought to
4493 * But, this leaves one open to an easy denial of
4494 * service attack, and SYN cookies can't defend
4495 * against this problem. So, we drop the data
4496 * in the interest of security over speed.
4505 queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
4509 /* Do step6 onward by hand. */
4510 tcp_urg(sk, skb, th);
4512 tcp_data_snd_check(sk);
4516 if (tcp_fast_parse_options(skb, th, tp) && tp->saw_tstamp &&
4517 tcp_paws_discard(tp, skb)) {
4519 NET_INC_STATS_BH(PAWSEstabRejected);
4520 tcp_send_dupack(sk, skb);
4523 /* Reset is accepted even if it did not pass PAWS. */
4526 /* step 1: check sequence number */
4527 if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
4529 tcp_send_dupack(sk, skb);
4533 /* step 2: check RST bit */
4539 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
4541 /* step 3: check security and precedence [ignored] */
4545 * Check for a SYN in window.
4547 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4548 NET_INC_STATS_BH(TCPAbortOnSyn);
4553 /* step 5: check the ACK field */
4555 int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH);
4557 switch(sk->sk_state) {
4560 tp->copied_seq = tp->rcv_nxt;
4562 tcp_set_state(sk, TCP_ESTABLISHED);
4563 sk->sk_state_change(sk);
4565 /* Note, that this wakeup is only for marginal
4566 * crossed SYN case. Passively open sockets
4567 * are not waked up, because sk->sk_sleep ==
4568 * NULL and sk->sk_socket == NULL.
4570 if (sk->sk_socket) {
4571 sk_wake_async(sk,0,POLL_OUT);
4574 tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
4575 tp->snd_wnd = ntohs(th->window) <<
4577 tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq,
4578 TCP_SKB_CB(skb)->seq);
4580 /* tcp_ack considers this ACK as duplicate
4581 * and does not calculate rtt.
4582 * Fix it at least with timestamps.
4584 if (tp->saw_tstamp && tp->rcv_tsecr &&
4586 tcp_ack_saw_tstamp(tp, 0);
4589 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
4591 /* Make sure socket is routed, for
4594 tp->af_specific->rebuild_header(sk);
4596 tcp_init_metrics(sk);
4598 /* Prevent spurious tcp_cwnd_restart() on
4599 * first data packet.
4601 tp->lsndtime = tcp_time_stamp;
4603 tcp_initialize_rcv_mss(sk);
4604 tcp_init_buffer_space(sk);
4605 tcp_fast_path_on(tp);
4612 if (tp->snd_una == tp->write_seq) {
4613 tcp_set_state(sk, TCP_FIN_WAIT2);
4614 sk->sk_shutdown |= SEND_SHUTDOWN;
4615 dst_confirm(sk->sk_dst_cache);
4617 if (!sock_flag(sk, SOCK_DEAD))
4618 /* Wake up lingering close() */
4619 sk->sk_state_change(sk);
4623 if (tp->linger2 < 0 ||
4624 (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4625 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
4627 NET_INC_STATS_BH(TCPAbortOnData);
4631 tmo = tcp_fin_time(tp);
4632 if (tmo > TCP_TIMEWAIT_LEN) {
4633 tcp_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
4634 } else if (th->fin || sock_owned_by_user(sk)) {
4635 /* Bad case. We could lose such FIN otherwise.
4636 * It is not a big problem, but it looks confusing
4637 * and not so rare event. We still can lose it now,
4638 * if it spins in bh_lock_sock(), but it is really
4641 tcp_reset_keepalive_timer(sk, tmo);
4643 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
4651 if (tp->snd_una == tp->write_seq) {
4652 tcp_time_wait(sk, TCP_TIME_WAIT, 0);
4658 if (tp->snd_una == tp->write_seq) {
4659 tcp_update_metrics(sk);
4668 /* step 6: check the URG bit */
4669 tcp_urg(sk, skb, th);
4671 /* step 7: process the segment text */
4672 switch (sk->sk_state) {
4673 case TCP_CLOSE_WAIT:
4676 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4680 /* RFC 793 says to queue data in these states,
4681 * RFC 1122 says we MUST send a reset.
4682 * BSD 4.4 also does reset.
4684 if (sk->sk_shutdown & RCV_SHUTDOWN) {
4685 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
4686 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
4687 NET_INC_STATS_BH(TCPAbortOnData);
4693 case TCP_ESTABLISHED:
4694 tcp_data_queue(sk, skb);
4699 /* tcp_data could move socket to TIME-WAIT */
4700 if (sk->sk_state != TCP_CLOSE) {
4701 tcp_data_snd_check(sk);
4702 tcp_ack_snd_check(sk);
4712 EXPORT_SYMBOL(sysctl_tcp_ecn);
4713 EXPORT_SYMBOL(sysctl_tcp_reordering);
4714 EXPORT_SYMBOL(tcp_cwnd_application_limited);
4715 EXPORT_SYMBOL(tcp_parse_options);
4716 EXPORT_SYMBOL(tcp_rcv_established);
4717 EXPORT_SYMBOL(tcp_rcv_state_process);