tcp_timer.c revision 161226
1/*-
2 * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995
3 *	The Regents of the University of California.  All rights reserved.
4 *
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
7 * are met:
8 * 1. Redistributions of source code must retain the above copyright
9 *    notice, this list of conditions and the following disclaimer.
10 * 2. Redistributions in binary form must reproduce the above copyright
11 *    notice, this list of conditions and the following disclaimer in the
12 *    documentation and/or other materials provided with the distribution.
13 * 4. Neither the name of the University nor the names of its contributors
14 *    may be used to endorse or promote products derived from this software
15 *    without specific prior written permission.
16 *
17 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
18 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20 * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
21 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27 * SUCH DAMAGE.
28 *
29 *	@(#)tcp_timer.c	8.2 (Berkeley) 5/24/95
30 * $FreeBSD: head/sys/netinet/tcp_timer.c 161226 2006-08-11 21:15:23Z mohans $
31 */
32
33#include "opt_inet6.h"
34#include "opt_tcpdebug.h"
35#include "opt_tcp_sack.h"
36
37#include <sys/param.h>
38#include <sys/kernel.h>
39#include <sys/lock.h>
40#include <sys/mbuf.h>
41#include <sys/mutex.h>
42#include <sys/protosw.h>
43#include <sys/socket.h>
44#include <sys/socketvar.h>
45#include <sys/sysctl.h>
46#include <sys/systm.h>
47
48#include <net/route.h>
49
50#include <netinet/in.h>
51#include <netinet/in_pcb.h>
52#include <netinet/in_systm.h>
53#ifdef INET6
54#include <netinet6/in6_pcb.h>
55#endif
56#include <netinet/ip_var.h>
57#include <netinet/tcp.h>
58#include <netinet/tcp_fsm.h>
59#include <netinet/tcp_timer.h>
60#include <netinet/tcp_var.h>
61#include <netinet/tcpip.h>
62#ifdef TCPDEBUG
63#include <netinet/tcp_debug.h>
64#endif
65
66int	tcp_keepinit;
67SYSCTL_PROC(_net_inet_tcp, TCPCTL_KEEPINIT, keepinit, CTLTYPE_INT|CTLFLAG_RW,
68    &tcp_keepinit, 0, sysctl_msec_to_ticks, "I", "");
69
70int	tcp_keepidle;
71SYSCTL_PROC(_net_inet_tcp, TCPCTL_KEEPIDLE, keepidle, CTLTYPE_INT|CTLFLAG_RW,
72    &tcp_keepidle, 0, sysctl_msec_to_ticks, "I", "");
73
74int	tcp_keepintvl;
75SYSCTL_PROC(_net_inet_tcp, TCPCTL_KEEPINTVL, keepintvl, CTLTYPE_INT|CTLFLAG_RW,
76    &tcp_keepintvl, 0, sysctl_msec_to_ticks, "I", "");
77
78int	tcp_delacktime;
79SYSCTL_PROC(_net_inet_tcp, TCPCTL_DELACKTIME, delacktime,
80    CTLTYPE_INT|CTLFLAG_RW, &tcp_delacktime, 0, sysctl_msec_to_ticks, "I",
81    "Time before a delayed ACK is sent");
82
83int	tcp_msl;
84SYSCTL_PROC(_net_inet_tcp, OID_AUTO, msl, CTLTYPE_INT|CTLFLAG_RW,
85    &tcp_msl, 0, sysctl_msec_to_ticks, "I", "Maximum segment lifetime");
86
87int	tcp_rexmit_min;
88SYSCTL_PROC(_net_inet_tcp, OID_AUTO, rexmit_min, CTLTYPE_INT|CTLFLAG_RW,
89    &tcp_rexmit_min, 0, sysctl_msec_to_ticks, "I", "Minimum Retransmission Timeout");
90
91int	tcp_rexmit_slop;
92SYSCTL_PROC(_net_inet_tcp, OID_AUTO, rexmit_slop, CTLTYPE_INT|CTLFLAG_RW,
93    &tcp_rexmit_slop, 0, sysctl_msec_to_ticks, "I", "Retransmission Timer Slop");
94
95static int	always_keepalive = 1;
96SYSCTL_INT(_net_inet_tcp, OID_AUTO, always_keepalive, CTLFLAG_RW,
97    &always_keepalive , 0, "Assume SO_KEEPALIVE on all TCP connections");
98
99static int	tcp_keepcnt = TCPTV_KEEPCNT;
100	/* max idle probes */
101int	tcp_maxpersistidle;
102	/* max idle time in persist */
103int	tcp_maxidle;
104
105/*
106 * Tcp protocol timeout routine called every 500 ms.
107 * Updates timestamps used for TCP
108 * causes finite state machine actions if timers expire.
109 */
110void
111tcp_slowtimo()
112{
113
114	tcp_maxidle = tcp_keepcnt * tcp_keepintvl;
115	INP_INFO_WLOCK(&tcbinfo);
116	(void) tcp_timer_2msl_tw(0);
117	INP_INFO_WUNLOCK(&tcbinfo);
118}
119
120int	tcp_syn_backoff[TCP_MAXRXTSHIFT + 1] =
121    { 1, 1, 1, 1, 1, 2, 4, 8, 16, 32, 64, 64, 64 };
122
123int	tcp_backoff[TCP_MAXRXTSHIFT + 1] =
124    { 1, 2, 4, 8, 16, 32, 64, 128, 256, 512, 512, 512, 512 };
125
126static int tcp_totbackoff = 2559;	/* sum of tcp_backoff[] */
127
128static int tcp_timer_race;
129SYSCTL_INT(_net_inet_tcp, OID_AUTO, timer_race, CTLFLAG_RD, &tcp_timer_race,
130    0, "Count of t_inpcb races on tcp_discardcb");
131
132/*
133 * TCP timer processing.
134 */
135
136void
137tcp_timer_delack(xtp)
138	void *xtp;
139{
140	struct tcpcb *tp = xtp;
141	struct inpcb *inp;
142
143	INP_INFO_RLOCK(&tcbinfo);
144	inp = tp->t_inpcb;
145	/*
146	 * XXXRW: While this assert is in fact correct, bugs in the tcpcb
147	 * tear-down mean we need it as a work-around for races between
148	 * timers and tcp_discardcb().
149	 *
150	 * KASSERT(inp != NULL, ("tcp_timer_delack: inp == NULL"));
151	 */
152	if (inp == NULL) {
153		tcp_timer_race++;
154		INP_INFO_RUNLOCK(&tcbinfo);
155		return;
156	}
157	INP_LOCK(inp);
158	INP_INFO_RUNLOCK(&tcbinfo);
159	if ((inp->inp_vflag & INP_DROPPED) || callout_pending(tp->tt_delack)
160	    || !callout_active(tp->tt_delack)) {
161		INP_UNLOCK(inp);
162		return;
163	}
164	callout_deactivate(tp->tt_delack);
165
166	tp->t_flags |= TF_ACKNOW;
167	tcpstat.tcps_delack++;
168	(void) tcp_output(tp);
169	INP_UNLOCK(inp);
170}
171
172void
173tcp_timer_2msl(xtp)
174	void *xtp;
175{
176	struct tcpcb *tp = xtp;
177	struct inpcb *inp;
178#ifdef TCPDEBUG
179	int ostate;
180
181	ostate = tp->t_state;
182#endif
183	/*
184	 * XXXRW: Does this actually happen?
185	 */
186	INP_INFO_WLOCK(&tcbinfo);
187	inp = tp->t_inpcb;
188	/*
189	 * XXXRW: While this assert is in fact correct, bugs in the tcpcb
190	 * tear-down mean we need it as a work-around for races between
191	 * timers and tcp_discardcb().
192	 *
193	 * KASSERT(inp != NULL, ("tcp_timer_2msl: inp == NULL"));
194	 */
195	if (inp == NULL) {
196		tcp_timer_race++;
197		INP_INFO_WUNLOCK(&tcbinfo);
198		return;
199	}
200	INP_LOCK(inp);
201	tcp_free_sackholes(tp);
202	if ((inp->inp_vflag & INP_DROPPED) || callout_pending(tp->tt_2msl) ||
203	    !callout_active(tp->tt_2msl)) {
204		INP_UNLOCK(tp->t_inpcb);
205		INP_INFO_WUNLOCK(&tcbinfo);
206		return;
207	}
208	callout_deactivate(tp->tt_2msl);
209	/*
210	 * 2 MSL timeout in shutdown went off.  If we're closed but
211	 * still waiting for peer to close and connection has been idle
212	 * too long, or if 2MSL time is up from TIME_WAIT, delete connection
213	 * control block.  Otherwise, check again in a bit.
214	 */
215	if (tp->t_state != TCPS_TIME_WAIT &&
216	    (ticks - tp->t_rcvtime) <= tcp_maxidle)
217		callout_reset(tp->tt_2msl, tcp_keepintvl,
218			      tcp_timer_2msl, tp);
219	else
220		tp = tcp_close(tp);
221
222#ifdef TCPDEBUG
223	if (tp != NULL && (tp->t_inpcb->inp_socket->so_options & SO_DEBUG))
224		tcp_trace(TA_USER, ostate, tp, (void *)0, (struct tcphdr *)0,
225			  PRU_SLOWTIMO);
226#endif
227	if (tp != NULL)
228		INP_UNLOCK(inp);
229	INP_INFO_WUNLOCK(&tcbinfo);
230}
231
232/*
233 * The timed wait lists contain references to each of the TCP sessions
234 * currently TIME_WAIT state.  The list pointers, including the list pointers
235 * in each tcptw structure, are protected using the global tcbinfo lock,
236 * which must be held over list iteration and modification.
237 */
238struct twlist {
239	LIST_HEAD(, tcptw)	tw_list;
240	struct tcptw	tw_tail;
241};
242#define TWLIST_NLISTS	2
243static struct twlist twl_2msl[TWLIST_NLISTS];
244static struct twlist *tw_2msl_list[] = { &twl_2msl[0], &twl_2msl[1], NULL };
245
246void
247tcp_timer_init(void)
248{
249	int i;
250	struct twlist *twl;
251
252	for (i = 0; i < TWLIST_NLISTS; i++) {
253		twl = &twl_2msl[i];
254		LIST_INIT(&twl->tw_list);
255		LIST_INSERT_HEAD(&twl->tw_list, &twl->tw_tail, tw_2msl);
256	}
257}
258
259void
260tcp_timer_2msl_reset(struct tcptw *tw, int timeo, int rearm)
261{
262	int i;
263	struct tcptw *tw_tail;
264
265	INP_INFO_WLOCK_ASSERT(&tcbinfo);
266	INP_LOCK_ASSERT(tw->tw_inpcb);
267	if (rearm)
268		LIST_REMOVE(tw, tw_2msl);
269	tw->tw_time = timeo + ticks;
270	i = timeo > tcp_msl ? 1 : 0;
271	tw_tail = &twl_2msl[i].tw_tail;
272	LIST_INSERT_BEFORE(tw_tail, tw, tw_2msl);
273}
274
275void
276tcp_timer_2msl_stop(struct tcptw *tw)
277{
278
279	INP_INFO_WLOCK_ASSERT(&tcbinfo);
280	LIST_REMOVE(tw, tw_2msl);
281}
282
283struct tcptw *
284tcp_timer_2msl_tw(int reuse)
285{
286	struct tcptw *tw, *tw_tail;
287	struct twlist *twl;
288	int i;
289
290	INP_INFO_WLOCK_ASSERT(&tcbinfo);
291	for (i = 0; i < 2; i++) {
292		twl = tw_2msl_list[i];
293		tw_tail = &twl->tw_tail;
294		tw = LIST_FIRST(&twl->tw_list);
295		if (tw == tw_tail || (!reuse && tw->tw_time > ticks))
296			continue;
297		INP_LOCK(tw->tw_inpcb);
298		tcp_twclose(tw, reuse);
299		return (reuse ? tw : NULL);
300	}
301	return (NULL);
302}
303
304void
305tcp_timer_keep(xtp)
306	void *xtp;
307{
308	struct tcpcb *tp = xtp;
309	struct tcptemp *t_template;
310	struct inpcb *inp;
311#ifdef TCPDEBUG
312	int ostate;
313
314	ostate = tp->t_state;
315#endif
316	INP_INFO_WLOCK(&tcbinfo);
317	inp = tp->t_inpcb;
318	/*
319	 * XXXRW: While this assert is in fact correct, bugs in the tcpcb
320	 * tear-down mean we need it as a work-around for races between
321	 * timers and tcp_discardcb().
322	 *
323	 * KASSERT(inp != NULL, ("tcp_timer_keep: inp == NULL"));
324	 */
325	if (inp == NULL) {
326		tcp_timer_race++;
327		INP_INFO_WUNLOCK(&tcbinfo);
328		return;
329	}
330	INP_LOCK(inp);
331	if ((inp->inp_vflag & INP_DROPPED) || callout_pending(tp->tt_keep)
332	    || !callout_active(tp->tt_keep)) {
333		INP_UNLOCK(inp);
334		INP_INFO_WUNLOCK(&tcbinfo);
335		return;
336	}
337	callout_deactivate(tp->tt_keep);
338	/*
339	 * Keep-alive timer went off; send something
340	 * or drop connection if idle for too long.
341	 */
342	tcpstat.tcps_keeptimeo++;
343	if (tp->t_state < TCPS_ESTABLISHED)
344		goto dropit;
345	if ((always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
346	    tp->t_state <= TCPS_CLOSING) {
347		if ((ticks - tp->t_rcvtime) >= tcp_keepidle + tcp_maxidle)
348			goto dropit;
349		/*
350		 * Send a packet designed to force a response
351		 * if the peer is up and reachable:
352		 * either an ACK if the connection is still alive,
353		 * or an RST if the peer has closed the connection
354		 * due to timeout or reboot.
355		 * Using sequence number tp->snd_una-1
356		 * causes the transmitted zero-length segment
357		 * to lie outside the receive window;
358		 * by the protocol spec, this requires the
359		 * correspondent TCP to respond.
360		 */
361		tcpstat.tcps_keepprobe++;
362		t_template = tcpip_maketemplate(inp);
363		if (t_template) {
364			tcp_respond(tp, t_template->tt_ipgen,
365				    &t_template->tt_t, (struct mbuf *)NULL,
366				    tp->rcv_nxt, tp->snd_una - 1, 0);
367			(void) m_free(dtom(t_template));
368		}
369		callout_reset(tp->tt_keep, tcp_keepintvl, tcp_timer_keep, tp);
370	} else
371		callout_reset(tp->tt_keep, tcp_keepidle, tcp_timer_keep, tp);
372
373#ifdef TCPDEBUG
374	if (inp->inp_socket->so_options & SO_DEBUG)
375		tcp_trace(TA_USER, ostate, tp, (void *)0, (struct tcphdr *)0,
376			  PRU_SLOWTIMO);
377#endif
378	INP_UNLOCK(inp);
379	INP_INFO_WUNLOCK(&tcbinfo);
380	return;
381
382dropit:
383	tcpstat.tcps_keepdrops++;
384	tp = tcp_drop(tp, ETIMEDOUT);
385
386#ifdef TCPDEBUG
387	if (tp != NULL && (tp->t_inpcb->inp_socket->so_options & SO_DEBUG))
388		tcp_trace(TA_USER, ostate, tp, (void *)0, (struct tcphdr *)0,
389			  PRU_SLOWTIMO);
390#endif
391	if (tp != NULL)
392		INP_UNLOCK(tp->t_inpcb);
393	INP_INFO_WUNLOCK(&tcbinfo);
394}
395
396void
397tcp_timer_persist(xtp)
398	void *xtp;
399{
400	struct tcpcb *tp = xtp;
401	struct inpcb *inp;
402#ifdef TCPDEBUG
403	int ostate;
404
405	ostate = tp->t_state;
406#endif
407	INP_INFO_WLOCK(&tcbinfo);
408	inp = tp->t_inpcb;
409	/*
410	 * XXXRW: While this assert is in fact correct, bugs in the tcpcb
411	 * tear-down mean we need it as a work-around for races between
412	 * timers and tcp_discardcb().
413	 *
414	 * KASSERT(inp != NULL, ("tcp_timer_persist: inp == NULL"));
415	 */
416	if (inp == NULL) {
417		tcp_timer_race++;
418		INP_INFO_WUNLOCK(&tcbinfo);
419		return;
420	}
421	INP_LOCK(inp);
422	if ((inp->inp_vflag & INP_DROPPED) || callout_pending(tp->tt_persist)
423	    || !callout_active(tp->tt_persist)) {
424		INP_UNLOCK(inp);
425		INP_INFO_WUNLOCK(&tcbinfo);
426		return;
427	}
428	callout_deactivate(tp->tt_persist);
429	/*
430	 * Persistance timer into zero window.
431	 * Force a byte to be output, if possible.
432	 */
433	tcpstat.tcps_persisttimeo++;
434	/*
435	 * Hack: if the peer is dead/unreachable, we do not
436	 * time out if the window is closed.  After a full
437	 * backoff, drop the connection if the idle time
438	 * (no responses to probes) reaches the maximum
439	 * backoff that we would use if retransmitting.
440	 */
441	if (tp->t_rxtshift == TCP_MAXRXTSHIFT &&
442	    ((ticks - tp->t_rcvtime) >= tcp_maxpersistidle ||
443	     (ticks - tp->t_rcvtime) >= TCP_REXMTVAL(tp) * tcp_totbackoff)) {
444		tcpstat.tcps_persistdrop++;
445		tp = tcp_drop(tp, ETIMEDOUT);
446		goto out;
447	}
448	tcp_setpersist(tp);
449	tp->t_flags |= TF_FORCEDATA;
450	(void) tcp_output(tp);
451	tp->t_flags &= ~TF_FORCEDATA;
452
453out:
454#ifdef TCPDEBUG
455	if (tp != NULL && tp->t_inpcb->inp_socket->so_options & SO_DEBUG)
456		tcp_trace(TA_USER, ostate, tp, NULL, NULL, PRU_SLOWTIMO);
457#endif
458	if (tp != NULL)
459		INP_UNLOCK(inp);
460	INP_INFO_WUNLOCK(&tcbinfo);
461}
462
463void
464tcp_timer_rexmt(xtp)
465	void *xtp;
466{
467	struct tcpcb *tp = xtp;
468	int rexmt;
469	int headlocked;
470	struct inpcb *inp;
471#ifdef TCPDEBUG
472	int ostate;
473
474	ostate = tp->t_state;
475#endif
476	INP_INFO_WLOCK(&tcbinfo);
477	headlocked = 1;
478	inp = tp->t_inpcb;
479	/*
480	 * XXXRW: While this assert is in fact correct, bugs in the tcpcb
481	 * tear-down mean we need it as a work-around for races between
482	 * timers and tcp_discardcb().
483	 *
484	 * KASSERT(inp != NULL, ("tcp_timer_rexmt: inp == NULL"));
485	 */
486	if (inp == NULL) {
487		tcp_timer_race++;
488		INP_INFO_WUNLOCK(&tcbinfo);
489		return;
490	}
491	INP_LOCK(inp);
492	if ((inp->inp_vflag & INP_DROPPED) || callout_pending(tp->tt_rexmt)
493	    || !callout_active(tp->tt_rexmt)) {
494		INP_UNLOCK(inp);
495		INP_INFO_WUNLOCK(&tcbinfo);
496		return;
497	}
498	callout_deactivate(tp->tt_rexmt);
499	tcp_free_sackholes(tp);
500	/*
501	 * Retransmission timer went off.  Message has not
502	 * been acked within retransmit interval.  Back off
503	 * to a longer retransmit interval and retransmit one segment.
504	 */
505	if (++tp->t_rxtshift > TCP_MAXRXTSHIFT) {
506		tp->t_rxtshift = TCP_MAXRXTSHIFT;
507		tcpstat.tcps_timeoutdrop++;
508		tp = tcp_drop(tp, tp->t_softerror ?
509			      tp->t_softerror : ETIMEDOUT);
510		goto out;
511	}
512	INP_INFO_WUNLOCK(&tcbinfo);
513	headlocked = 0;
514	if (tp->t_rxtshift == 1) {
515		/*
516		 * first retransmit; record ssthresh and cwnd so they can
517		 * be recovered if this turns out to be a "bad" retransmit.
518		 * A retransmit is considered "bad" if an ACK for this
519		 * segment is received within RTT/2 interval; the assumption
520		 * here is that the ACK was already in flight.  See
521		 * "On Estimating End-to-End Network Path Properties" by
522		 * Allman and Paxson for more details.
523		 */
524		tp->snd_cwnd_prev = tp->snd_cwnd;
525		tp->snd_ssthresh_prev = tp->snd_ssthresh;
526		tp->snd_recover_prev = tp->snd_recover;
527		if (IN_FASTRECOVERY(tp))
528		  tp->t_flags |= TF_WASFRECOVERY;
529		else
530		  tp->t_flags &= ~TF_WASFRECOVERY;
531		tp->t_badrxtwin = ticks + (tp->t_srtt >> (TCP_RTT_SHIFT + 1));
532	}
533	tcpstat.tcps_rexmttimeo++;
534	if (tp->t_state == TCPS_SYN_SENT)
535		rexmt = TCP_REXMTVAL(tp) * tcp_syn_backoff[tp->t_rxtshift];
536	else
537		rexmt = TCP_REXMTVAL(tp) * tcp_backoff[tp->t_rxtshift];
538	TCPT_RANGESET(tp->t_rxtcur, rexmt,
539		      tp->t_rttmin, TCPTV_REXMTMAX);
540	/*
541	 * Disable rfc1323 if we havn't got any response to
542	 * our third SYN to work-around some broken terminal servers
543	 * (most of which have hopefully been retired) that have bad VJ
544	 * header compression code which trashes TCP segments containing
545	 * unknown-to-them TCP options.
546	 */
547	if ((tp->t_state == TCPS_SYN_SENT) && (tp->t_rxtshift == 3))
548		tp->t_flags &= ~(TF_REQ_SCALE|TF_REQ_TSTMP);
549	/*
550	 * If we backed off this far, our srtt estimate is probably bogus.
551	 * Clobber it so we'll take the next rtt measurement as our srtt;
552	 * move the current srtt into rttvar to keep the current
553	 * retransmit times until then.
554	 */
555	if (tp->t_rxtshift > TCP_MAXRXTSHIFT / 4) {
556#ifdef INET6
557		if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0)
558			in6_losing(tp->t_inpcb);
559		else
560#endif
561		tp->t_rttvar += (tp->t_srtt >> TCP_RTT_SHIFT);
562		tp->t_srtt = 0;
563	}
564	tp->snd_nxt = tp->snd_una;
565	tp->snd_recover = tp->snd_max;
566	/*
567	 * Force a segment to be sent.
568	 */
569	tp->t_flags |= TF_ACKNOW;
570	/*
571	 * If timing a segment in this window, stop the timer.
572	 */
573	tp->t_rtttime = 0;
574	/*
575	 * Close the congestion window down to one segment
576	 * (we'll open it by one segment for each ack we get).
577	 * Since we probably have a window's worth of unacked
578	 * data accumulated, this "slow start" keeps us from
579	 * dumping all that data as back-to-back packets (which
580	 * might overwhelm an intermediate gateway).
581	 *
582	 * There are two phases to the opening: Initially we
583	 * open by one mss on each ack.  This makes the window
584	 * size increase exponentially with time.  If the
585	 * window is larger than the path can handle, this
586	 * exponential growth results in dropped packet(s)
587	 * almost immediately.  To get more time between
588	 * drops but still "push" the network to take advantage
589	 * of improving conditions, we switch from exponential
590	 * to linear window opening at some threshhold size.
591	 * For a threshhold, we use half the current window
592	 * size, truncated to a multiple of the mss.
593	 *
594	 * (the minimum cwnd that will give us exponential
595	 * growth is 2 mss.  We don't allow the threshhold
596	 * to go below this.)
597	 */
598	{
599		u_int win = min(tp->snd_wnd, tp->snd_cwnd) / 2 / tp->t_maxseg;
600		if (win < 2)
601			win = 2;
602		tp->snd_cwnd = tp->t_maxseg;
603		tp->snd_ssthresh = win * tp->t_maxseg;
604		tp->t_dupacks = 0;
605	}
606	EXIT_FASTRECOVERY(tp);
607	(void) tcp_output(tp);
608
609out:
610#ifdef TCPDEBUG
611	if (tp != NULL && (tp->t_inpcb->inp_socket->so_options & SO_DEBUG))
612		tcp_trace(TA_USER, ostate, tp, (void *)0, (struct tcphdr *)0,
613			  PRU_SLOWTIMO);
614#endif
615	if (tp != NULL)
616		INP_UNLOCK(inp);
617	if (headlocked)
618		INP_INFO_WUNLOCK(&tcbinfo);
619}
620