xref: /netbsd-src/sys/netinet/tcp_output.c (revision 17dd36da8292193180754d5047c0926dbb56818c)
1 /*	$NetBSD: tcp_output.c,v 1.65 2001/04/03 06:14:31 itojun Exp $	*/
2 
3 /*
4 %%% portions-copyright-nrl-95
5 Portions of this software are Copyright 1995-1998 by Randall Atkinson,
6 Ronald Lee, Daniel McDonald, Bao Phan, and Chris Winters. All Rights
7 Reserved. All rights under this copyright have been assigned to the US
8 Naval Research Laboratory (NRL). The NRL Copyright Notice and License
9 Agreement Version 1.1 (January 17, 1995) applies to these portions of the
10 software.
11 You should have received a copy of the license with this software. If you
12 didn't get a copy, you may request one from <license@ipv6.nrl.navy.mil>.
13 
14 */
15 
16 /*
17  * Copyright (C) 1995, 1996, 1997, and 1998 WIDE Project.
18  * All rights reserved.
19  *
20  * Redistribution and use in source and binary forms, with or without
21  * modification, are permitted provided that the following conditions
22  * are met:
23  * 1. Redistributions of source code must retain the above copyright
24  *    notice, this list of conditions and the following disclaimer.
25  * 2. Redistributions in binary form must reproduce the above copyright
26  *    notice, this list of conditions and the following disclaimer in the
27  *    documentation and/or other materials provided with the distribution.
28  * 3. Neither the name of the project nor the names of its contributors
29  *    may be used to endorse or promote products derived from this software
30  *    without specific prior written permission.
31  *
32  * THIS SOFTWARE IS PROVIDED BY THE PROJECT AND CONTRIBUTORS ``AS IS'' AND
33  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
34  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
35  * ARE DISCLAIMED.  IN NO EVENT SHALL THE PROJECT OR CONTRIBUTORS BE LIABLE
36  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
37  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
38  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
39  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
40  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
41  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
42  * SUCH DAMAGE.
43  */
44 
45 /*-
46  * Copyright (c) 1997, 1998, 2001 The NetBSD Foundation, Inc.
47  * All rights reserved.
48  *
49  * This code is derived from software contributed to The NetBSD Foundation
50  * by Jason R. Thorpe and Kevin M. Lahey of the Numerical Aerospace Simulation
51  * Facility, NASA Ames Research Center.
52  *
53  * Redistribution and use in source and binary forms, with or without
54  * modification, are permitted provided that the following conditions
55  * are met:
56  * 1. Redistributions of source code must retain the above copyright
57  *    notice, this list of conditions and the following disclaimer.
58  * 2. Redistributions in binary form must reproduce the above copyright
59  *    notice, this list of conditions and the following disclaimer in the
60  *    documentation and/or other materials provided with the distribution.
61  * 3. All advertising materials mentioning features or use of this software
62  *    must display the following acknowledgement:
63  *	This product includes software developed by the NetBSD
64  *	Foundation, Inc. and its contributors.
65  * 4. Neither the name of The NetBSD Foundation nor the names of its
66  *    contributors may be used to endorse or promote products derived
67  *    from this software without specific prior written permission.
68  *
69  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
70  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
71  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
72  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
73  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
74  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
75  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
76  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
77  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
78  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
79  * POSSIBILITY OF SUCH DAMAGE.
80  */
81 
82 /*
83  * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995
84  *	The Regents of the University of California.  All rights reserved.
85  *
86  * Redistribution and use in source and binary forms, with or without
87  * modification, are permitted provided that the following conditions
88  * are met:
89  * 1. Redistributions of source code must retain the above copyright
90  *    notice, this list of conditions and the following disclaimer.
91  * 2. Redistributions in binary form must reproduce the above copyright
92  *    notice, this list of conditions and the following disclaimer in the
93  *    documentation and/or other materials provided with the distribution.
94  * 3. All advertising materials mentioning features or use of this software
95  *    must display the following acknowledgement:
96  *	This product includes software developed by the University of
97  *	California, Berkeley and its contributors.
98  * 4. Neither the name of the University nor the names of its contributors
99  *    may be used to endorse or promote products derived from this software
100  *    without specific prior written permission.
101  *
102  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
103  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
104  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
105  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
106  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
107  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
108  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
109  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
110  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
111  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
112  * SUCH DAMAGE.
113  *
114  *	@(#)tcp_output.c	8.4 (Berkeley) 5/24/95
115  */
116 
117 #include "opt_inet.h"
118 #include "opt_ipsec.h"
119 
120 #include <sys/param.h>
121 #include <sys/systm.h>
122 #include <sys/malloc.h>
123 #include <sys/mbuf.h>
124 #include <sys/protosw.h>
125 #include <sys/socket.h>
126 #include <sys/socketvar.h>
127 #include <sys/errno.h>
128 #include <sys/domain.h>
129 
130 #include <net/if.h>
131 #include <net/route.h>
132 
133 #include <netinet/in.h>
134 #include <netinet/in_systm.h>
135 #include <netinet/ip.h>
136 #include <netinet/in_pcb.h>
137 #include <netinet/ip_var.h>
138 
139 #ifdef INET6
140 #ifndef INET
141 #include <netinet/in.h>
142 #endif
143 #include <netinet/ip6.h>
144 #include <netinet6/in6_pcb.h>
145 #include <netinet6/ip6_var.h>
146 #endif
147 
148 #include <netinet/tcp.h>
149 #define	TCPOUTFLAGS
150 #include <netinet/tcp_fsm.h>
151 #include <netinet/tcp_seq.h>
152 #include <netinet/tcp_timer.h>
153 #include <netinet/tcp_var.h>
154 #include <netinet/tcpip.h>
155 #include <netinet/tcp_debug.h>
156 
157 #ifdef notyet
158 extern struct mbuf *m_copypack();
159 #endif
160 
161 #define MAX_TCPOPTLEN	32	/* max # bytes that go in options */
162 
163 /*
164  * Knob to enable Congestion Window Monitoring, and control the
165  * the burst size it allows.  Default burst is 4 packets, per
166  * the Internet draft.
167  */
168 int	tcp_cwm = 0;
169 int	tcp_cwm_burstsize = 4;
170 
171 static __inline void tcp_segsize __P((struct tcpcb *, int *, int *));
172 static __inline void
173 tcp_segsize(tp, txsegsizep, rxsegsizep)
174 	struct tcpcb *tp;
175 	int *txsegsizep, *rxsegsizep;
176 {
177 #ifdef INET
178 	struct inpcb *inp = tp->t_inpcb;
179 #endif
180 #ifdef INET6
181 	struct in6pcb *in6p = tp->t_in6pcb;
182 #endif
183 	struct rtentry *rt;
184 	struct ifnet *ifp;
185 	int size;
186 	int iphlen;
187 
188 #ifdef DIAGNOSTIC
189 	if (tp->t_inpcb && tp->t_in6pcb)
190 		panic("tcp_segsize: both t_inpcb and t_in6pcb are set");
191 #endif
192 	switch (tp->t_family) {
193 #ifdef INET
194 	case AF_INET:
195 		iphlen = sizeof(struct ip);
196 		break;
197 #endif
198 #ifdef INET6
199 	case AF_INET6:
200 		iphlen = sizeof(struct ip6_hdr);
201 		break;
202 #endif
203 	default:
204 		size = tcp_mssdflt;
205 		goto out;
206 	}
207 
208 	rt = NULL;
209 #ifdef INET
210 	if (inp)
211 		rt = in_pcbrtentry(inp);
212 #endif
213 #ifdef INET6
214 	if (in6p)
215 		rt = in6_pcbrtentry(in6p);
216 #endif
217 	if (rt == NULL) {
218 		size = tcp_mssdflt;
219 		goto out;
220 	}
221 
222 	ifp = rt->rt_ifp;
223 
224 	size = tcp_mssdflt;
225 	if (rt->rt_rmx.rmx_mtu != 0)
226 		size = rt->rt_rmx.rmx_mtu - iphlen - sizeof(struct tcphdr);
227 	else if (ifp->if_flags & IFF_LOOPBACK)
228 		size = ifp->if_mtu - iphlen - sizeof(struct tcphdr);
229 #ifdef INET
230 	else if (inp && ip_mtudisc)
231 		size = ifp->if_mtu - iphlen - sizeof(struct tcphdr);
232 	else if (inp && in_localaddr(inp->inp_faddr))
233 		size = ifp->if_mtu - iphlen - sizeof(struct tcphdr);
234 #endif
235 #ifdef INET6
236 	else if (in6p) {
237 #ifdef INET
238 		if (IN6_IS_ADDR_V4MAPPED(&in6p->in6p_faddr)) {
239 			/* mapped addr case */
240 			struct in_addr d;
241 			bcopy(&in6p->in6p_faddr.s6_addr32[3], &d, sizeof(d));
242 			if (ip_mtudisc || in_localaddr(d))
243 				size = ifp->if_mtu - iphlen - sizeof(struct tcphdr);
244 		} else
245 #endif
246 		{
247 			/*
248 			 * for IPv6, path MTU discovery is always turned on,
249 			 * or the node must use packet size <= 1280.
250 			 */
251 			size = ifp->if_mtu - iphlen - sizeof(struct tcphdr);
252 		}
253 	}
254 #endif
255 	size -= tcp_optlen(tp);
256 	/*
257 	 * XXX tp->t_ourmss should have the right size, but without this code
258 	 * fragmentation will occur... need more investigation
259 	 */
260 #ifdef INET
261 	if (inp) {
262 #ifdef IPSEC
263 		size -= ipsec4_hdrsiz_tcp(tp);
264 #endif
265 		size -= ip_optlen(inp);
266 	}
267 #endif
268 #ifdef INET6
269 #ifdef INET
270 	if (in6p && tp->t_family == AF_INET) {
271 #ifdef IPSEC
272 		size -= ipsec4_hdrsiz_tcp(tp);
273 #endif
274 		/* XXX size -= ip_optlen(in6p); */
275 	} else
276 #endif
277 	if (in6p && tp->t_family == AF_INET6) {
278 #ifdef IPSEC
279 		size -= ipsec6_hdrsiz_tcp(tp);
280 #endif
281 		size -= ip6_optlen(in6p);
282 	}
283 #endif
284 
285  out:
286 	/*
287 	 * *rxsegsizep holds *estimated* inbound segment size (estimation
288 	 * assumes that path MTU is the same for both ways).  this is only
289 	 * for silly window avoidance, do not use the value for other purposes.
290 	 *
291 	 * ipseclen is subtracted from both sides, this may not be right.
292 	 * I'm not quite sure about this (could someone comment).
293 	 */
294 	*txsegsizep = min(tp->t_peermss, size);
295 	*rxsegsizep = min(tp->t_ourmss, size);
296 
297 	if (*txsegsizep != tp->t_segsz) {
298 		/*
299 		 * If the new segment size is larger, we don't want to
300 		 * mess up the congestion window, but if it is smaller
301 		 * we'll have to reduce the congestion window to ensure
302 		 * that we don't get into trouble with initial windows
303 		 * and the rest.  In any case, if the segment size
304 		 * has changed, chances are the path has, too, and
305 		 * our congestion window will be different.
306 		 */
307 		if (*txsegsizep < tp->t_segsz) {
308 			tp->snd_cwnd = max((tp->snd_cwnd / tp->t_segsz)
309 					   * *txsegsizep, *txsegsizep);
310 			tp->snd_ssthresh = max((tp->snd_ssthresh / tp->t_segsz)
311 						* *txsegsizep, *txsegsizep);
312 		}
313 		tp->t_segsz = *txsegsizep;
314 	}
315 }
316 
317 /*
318  * Tcp output routine: figure out what should be sent and send it.
319  */
320 int
321 tcp_output(tp)
322 	struct tcpcb *tp;
323 {
324 	struct socket *so;
325 	struct route *ro;
326 	long len, win;
327 	int off, flags, error;
328 	struct mbuf *m;
329 	struct ip *ip;
330 #ifdef INET6
331 	struct ip6_hdr *ip6;
332 #endif
333 	struct tcphdr *th;
334 	u_char opt[MAX_TCPOPTLEN];
335 	unsigned optlen, hdrlen;
336 	int idle, sendalot, txsegsize, rxsegsize;
337 	int maxburst = TCP_MAXBURST;
338 	int af;		/* address family on the wire */
339 	int iphdrlen;
340 
341 #ifdef DIAGNOSTIC
342 	if (tp->t_inpcb && tp->t_in6pcb)
343 		panic("tcp_output: both t_inpcb and t_in6pcb are set");
344 #endif
345 	so = NULL;
346 	ro = NULL;
347 	if (tp->t_inpcb) {
348 		so = tp->t_inpcb->inp_socket;
349 		ro = &tp->t_inpcb->inp_route;
350 	}
351 #ifdef INET6
352 	else if (tp->t_in6pcb) {
353 		so = tp->t_in6pcb->in6p_socket;
354 		ro = (struct route *)&tp->t_in6pcb->in6p_route;
355 	}
356 #endif
357 
358 	switch (af = tp->t_family) {
359 #ifdef INET
360 	case AF_INET:
361 		if (tp->t_inpcb)
362 			break;
363 #ifdef INET6
364 		/* mapped addr case */
365 		if (tp->t_in6pcb)
366 			break;
367 #endif
368 		return EINVAL;
369 #endif
370 #ifdef INET6
371 	case AF_INET6:
372 		if (tp->t_in6pcb)
373 			break;
374 		return EINVAL;
375 #endif
376 	default:
377 		return EAFNOSUPPORT;
378 	}
379 
380 	tcp_segsize(tp, &txsegsize, &rxsegsize);
381 
382 	idle = (tp->snd_max == tp->snd_una);
383 
384 	/*
385 	 * Restart Window computation.  From draft-floyd-incr-init-win-03:
386 	 *
387 	 *	Optionally, a TCP MAY set the restart window to the
388 	 *	minimum of the value used for the initial window and
389 	 *	the current value of cwnd (in other words, using a
390 	 *	larger value for the restart window should never increase
391 	 *	the size of cwnd).
392 	 */
393 	if (tcp_cwm) {
394 		/*
395 		 * Hughes/Touch/Heidemann Congestion Window Monitoring.
396 		 * Count the number of packets currently pending
397 		 * acknowledgement, and limit our congestion window
398 		 * to a pre-determined allowed burst size plus that count.
399 		 * This prevents bursting once all pending packets have
400 		 * been acknowledged (i.e. transmission is idle).
401 		 *
402 		 * XXX Link this to Initial Window?
403 		 */
404 		tp->snd_cwnd = min(tp->snd_cwnd,
405 		    (tcp_cwm_burstsize * txsegsize) +
406 		    (tp->snd_nxt - tp->snd_una));
407 	} else {
408 		if (idle && tp->t_idle >= tp->t_rxtcur) {
409 			/*
410 			 * We have been idle for "a while" and no acks are
411 			 * expected to clock out any data we send --
412 			 * slow start to get ack "clock" running again.
413 			 */
414 			tp->snd_cwnd = min(tp->snd_cwnd,
415 			    TCP_INITIAL_WINDOW(tcp_init_win, txsegsize));
416 		}
417 	}
418 
419 again:
420 	/*
421 	 * Determine length of data that should be transmitted, and
422 	 * flags that should be used.  If there is some data or critical
423 	 * controls (SYN, RST) to send, then transmit; otherwise,
424 	 * investigate further.
425 	 */
426 	sendalot = 0;
427 	off = tp->snd_nxt - tp->snd_una;
428 	win = min(tp->snd_wnd, tp->snd_cwnd);
429 
430 	flags = tcp_outflags[tp->t_state];
431 	/*
432 	 * If in persist timeout with window of 0, send 1 byte.
433 	 * Otherwise, if window is small but nonzero
434 	 * and timer expired, we will send what we can
435 	 * and go to transmit state.
436 	 */
437 	if (tp->t_force) {
438 		if (win == 0) {
439 			/*
440 			 * If we still have some data to send, then
441 			 * clear the FIN bit.  Usually this would
442 			 * happen below when it realizes that we
443 			 * aren't sending all the data.  However,
444 			 * if we have exactly 1 byte of unset data,
445 			 * then it won't clear the FIN bit below,
446 			 * and if we are in persist state, we wind
447 			 * up sending the packet without recording
448 			 * that we sent the FIN bit.
449 			 *
450 			 * We can't just blindly clear the FIN bit,
451 			 * because if we don't have any more data
452 			 * to send then the probe will be the FIN
453 			 * itself.
454 			 */
455 			if (off < so->so_snd.sb_cc)
456 				flags &= ~TH_FIN;
457 			win = 1;
458 		} else {
459 			TCP_TIMER_DISARM(tp, TCPT_PERSIST);
460 			tp->t_rxtshift = 0;
461 		}
462 	}
463 
464 	if (win < so->so_snd.sb_cc) {
465 		len = win - off;
466 		flags &= ~TH_FIN;
467 	} else
468 		len = so->so_snd.sb_cc - off;
469 
470 	if (len < 0) {
471 		/*
472 		 * If FIN has been sent but not acked,
473 		 * but we haven't been called to retransmit,
474 		 * len will be -1.  Otherwise, window shrank
475 		 * after we sent into it.  If window shrank to 0,
476 		 * cancel pending retransmit, pull snd_nxt back
477 		 * to (closed) window, and set the persist timer
478 		 * if it isn't already going.  If the window didn't
479 		 * close completely, just wait for an ACK.
480 		 *
481 		 * If we have a pending FIN, either it has already been
482 		 * transmitted or it is outside the window, so drop it.
483 		 * If the FIN has been transmitted, but this is not a
484 		 * retransmission, then len must be -1.  Therefore we also
485 		 * prevent here the sending of `gratuitous FINs'.  This
486 		 * eliminates the need to check for that case below (e.g.
487 		 * to back up snd_nxt before the FIN so that the sequence
488 		 * number is correct).
489 		 */
490 		len = 0;
491 		flags &= ~TH_FIN;
492 		if (win == 0) {
493 			TCP_TIMER_DISARM(tp, TCPT_REXMT);
494 			tp->t_rxtshift = 0;
495 			tp->snd_nxt = tp->snd_una;
496 			if (TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0)
497 				tcp_setpersist(tp);
498 		}
499 	}
500 	if (len > txsegsize) {
501 		len = txsegsize;
502 		flags &= ~TH_FIN;
503 		sendalot = 1;
504 	}
505 
506 	win = sbspace(&so->so_rcv);
507 
508 	/*
509 	 * Sender silly window avoidance.  If connection is idle
510 	 * and can send all data, a maximum segment,
511 	 * at least a maximum default-size segment do it,
512 	 * or are forced, do it; otherwise don't bother.
513 	 * If peer's buffer is tiny, then send
514 	 * when window is at least half open.
515 	 * If retransmitting (possibly after persist timer forced us
516 	 * to send into a small window), then must resend.
517 	 */
518 	if (len) {
519 		if (len == txsegsize)
520 			goto send;
521 		if ((so->so_state & SS_MORETOCOME) == 0 &&
522 		    ((idle || tp->t_flags & TF_NODELAY) &&
523 		     len + off >= so->so_snd.sb_cc))
524 			goto send;
525 		if (tp->t_force)
526 			goto send;
527 		if (len >= tp->max_sndwnd / 2)
528 			goto send;
529 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))
530 			goto send;
531 	}
532 
533 	/*
534 	 * Compare available window to amount of window known to peer
535 	 * (as advertised window less next expected input).  If the
536 	 * difference is at least twice the size of the largest segment
537 	 * we expect to receive (i.e. two segments) or at least 50% of
538 	 * the maximum possible window, then want to send a window update
539 	 * to peer.
540 	 */
541 	if (win > 0) {
542 		/*
543 		 * "adv" is the amount we can increase the window,
544 		 * taking into account that we are limited by
545 		 * TCP_MAXWIN << tp->rcv_scale.
546 		 */
547 		long adv = min(win, (long)TCP_MAXWIN << tp->rcv_scale) -
548 			(tp->rcv_adv - tp->rcv_nxt);
549 
550 		if (adv >= (long) (2 * rxsegsize))
551 			goto send;
552 		if (2 * adv >= (long) so->so_rcv.sb_hiwat)
553 			goto send;
554 	}
555 
556 	/*
557 	 * Send if we owe peer an ACK.
558 	 */
559 	if (tp->t_flags & TF_ACKNOW)
560 		goto send;
561 	if (flags & (TH_SYN|TH_FIN|TH_RST))
562 		goto send;
563 	if (SEQ_GT(tp->snd_up, tp->snd_una))
564 		goto send;
565 
566 	/*
567 	 * TCP window updates are not reliable, rather a polling protocol
568 	 * using ``persist'' packets is used to insure receipt of window
569 	 * updates.  The three ``states'' for the output side are:
570 	 *	idle			not doing retransmits or persists
571 	 *	persisting		to move a small or zero window
572 	 *	(re)transmitting	and thereby not persisting
573 	 *
574 	 * tp->t_timer[TCPT_PERSIST]
575 	 *	is set when we are in persist state.
576 	 * tp->t_force
577 	 *	is set when we are called to send a persist packet.
578 	 * tp->t_timer[TCPT_REXMT]
579 	 *	is set when we are retransmitting
580 	 * The output side is idle when both timers are zero.
581 	 *
582 	 * If send window is too small, there is data to transmit, and no
583 	 * retransmit or persist is pending, then go to persist state.
584 	 * If nothing happens soon, send when timer expires:
585 	 * if window is nonzero, transmit what we can,
586 	 * otherwise force out a byte.
587 	 */
588 	if (so->so_snd.sb_cc && TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 &&
589 	    TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) {
590 		tp->t_rxtshift = 0;
591 		tcp_setpersist(tp);
592 	}
593 
594 	/*
595 	 * No reason to send a segment, just return.
596 	 */
597 	return (0);
598 
599 send:
600 	/*
601 	 * Before ESTABLISHED, force sending of initial options
602 	 * unless TCP set not to do any options.
603 	 * NOTE: we assume that the IP/TCP header plus TCP options
604 	 * always fit in a single mbuf, leaving room for a maximum
605 	 * link header, i.e.
606 	 *	max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
607 	 */
608 	optlen = 0;
609 	switch (af) {
610 #ifdef INET
611 	case AF_INET:
612 		iphdrlen = sizeof(struct ip) + sizeof(struct tcphdr);
613 		break;
614 #endif
615 #ifdef INET6
616 	case AF_INET6:
617 		iphdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
618 		break;
619 #endif
620 	default:	/*pacify gcc*/
621 		iphdrlen = 0;
622 		break;
623 	}
624 	hdrlen = iphdrlen;
625 	if (flags & TH_SYN) {
626 		struct rtentry *rt;
627 
628 		rt = NULL;
629 #ifdef INET
630 		if (tp->t_inpcb)
631 			rt = in_pcbrtentry(tp->t_inpcb);
632 #endif
633 #ifdef INET6
634 		if (tp->t_in6pcb)
635 			rt = in6_pcbrtentry(tp->t_in6pcb);
636 #endif
637 
638 		tp->snd_nxt = tp->iss;
639 		tp->t_ourmss = tcp_mss_to_advertise(rt != NULL ?
640 						    rt->rt_ifp : NULL, af);
641 		if ((tp->t_flags & TF_NOOPT) == 0) {
642 			opt[0] = TCPOPT_MAXSEG;
643 			opt[1] = 4;
644 			opt[2] = (tp->t_ourmss >> 8) & 0xff;
645 			opt[3] = tp->t_ourmss & 0xff;
646 			optlen = 4;
647 
648 			if ((tp->t_flags & TF_REQ_SCALE) &&
649 			    ((flags & TH_ACK) == 0 ||
650 			    (tp->t_flags & TF_RCVD_SCALE))) {
651 				*((u_int32_t *) (opt + optlen)) = htonl(
652 					TCPOPT_NOP << 24 |
653 					TCPOPT_WINDOW << 16 |
654 					TCPOLEN_WINDOW << 8 |
655 					tp->request_r_scale);
656 				optlen += 4;
657 			}
658 		}
659  	}
660 
661  	/*
662 	 * Send a timestamp and echo-reply if this is a SYN and our side
663 	 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side
664 	 * and our peer have sent timestamps in our SYN's.
665  	 */
666  	if ((tp->t_flags & (TF_REQ_TSTMP|TF_NOOPT)) == TF_REQ_TSTMP &&
667  	     (flags & TH_RST) == 0 &&
668  	    ((flags & (TH_SYN|TH_ACK)) == TH_SYN ||
669 	     (tp->t_flags & TF_RCVD_TSTMP))) {
670 		u_int32_t *lp = (u_int32_t *)(opt + optlen);
671 
672  		/* Form timestamp option as shown in appendix A of RFC 1323. */
673  		*lp++ = htonl(TCPOPT_TSTAMP_HDR);
674  		*lp++ = htonl(TCP_TIMESTAMP(tp));
675  		*lp   = htonl(tp->ts_recent);
676  		optlen += TCPOLEN_TSTAMP_APPA;
677  	}
678 
679  	hdrlen += optlen;
680 
681 #ifdef DIAGNOSTIC
682 	if (len > txsegsize)
683 		panic("tcp data to be sent is larger than segment");
684  	if (max_linkhdr + hdrlen > MCLBYTES)
685 		panic("tcphdr too big");
686 #endif
687 
688 	/*
689 	 * Grab a header mbuf, attaching a copy of data to
690 	 * be transmitted, and initialize the header from
691 	 * the template for sends on this connection.
692 	 */
693 	if (len) {
694 		if (tp->t_force && len == 1)
695 			tcpstat.tcps_sndprobe++;
696 		else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
697 			tcpstat.tcps_sndrexmitpack++;
698 			tcpstat.tcps_sndrexmitbyte += len;
699 		} else {
700 			tcpstat.tcps_sndpack++;
701 			tcpstat.tcps_sndbyte += len;
702 		}
703 #ifdef notyet
704 		if ((m = m_copypack(so->so_snd.sb_mb, off,
705 		    (int)len, max_linkhdr + hdrlen)) == 0) {
706 			error = ENOBUFS;
707 			goto out;
708 		}
709 		/*
710 		 * m_copypack left space for our hdr; use it.
711 		 */
712 		m->m_len += hdrlen;
713 		m->m_data -= hdrlen;
714 #else
715 		MGETHDR(m, M_DONTWAIT, MT_HEADER);
716 		if (m != NULL &&
717 		    (max_linkhdr + hdrlen > MHLEN ||
718 		     max_linkhdr + hdrlen + len <= MCLBYTES)) {
719 			MCLGET(m, M_DONTWAIT);
720 			if ((m->m_flags & M_EXT) == 0) {
721 				m_freem(m);
722 				m = NULL;
723 			}
724 		}
725 		if (m == NULL) {
726 			error = ENOBUFS;
727 			goto out;
728 		}
729 		m->m_data += max_linkhdr;
730 		m->m_len = hdrlen;
731 		if (len <= M_TRAILINGSPACE(m)) {
732 			m_copydata(so->so_snd.sb_mb, off, (int) len,
733 			    mtod(m, caddr_t) + hdrlen);
734 			m->m_len += len;
735 		} else {
736 			m->m_next = m_copy(so->so_snd.sb_mb, off, (int) len);
737 			if (m->m_next == NULL) {
738 				m_freem(m);
739 				error = ENOBUFS;
740 				goto out;
741 			}
742 		}
743 #endif
744 		/*
745 		 * If we're sending everything we've got, set PUSH.
746 		 * (This will keep happy those implementations which only
747 		 * give data to the user when a buffer fills or
748 		 * a PUSH comes in.)
749 		 */
750 		if (off + len == so->so_snd.sb_cc)
751 			flags |= TH_PUSH;
752 	} else {
753 		if (tp->t_flags & TF_ACKNOW)
754 			tcpstat.tcps_sndacks++;
755 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
756 			tcpstat.tcps_sndctrl++;
757 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
758 			tcpstat.tcps_sndurg++;
759 		else
760 			tcpstat.tcps_sndwinup++;
761 
762 		MGETHDR(m, M_DONTWAIT, MT_HEADER);
763 		if (m != NULL && max_linkhdr + hdrlen > MHLEN) {
764 			MCLGET(m, M_DONTWAIT);
765 			if ((m->m_flags & M_EXT) == 0) {
766 				m_freem(m);
767 				m = NULL;
768 			}
769 		}
770 		if (m == NULL) {
771 			error = ENOBUFS;
772 			goto out;
773 		}
774 		m->m_data += max_linkhdr;
775 		m->m_len = hdrlen;
776 	}
777 	m->m_pkthdr.rcvif = (struct ifnet *)0;
778 	switch (af) {
779 #ifdef INET
780 	case AF_INET:
781 		ip = mtod(m, struct ip *);
782 #ifdef INET6
783 		ip6 = NULL;
784 #endif
785 		th = (struct tcphdr *)(ip + 1);
786 		break;
787 #endif
788 #ifdef INET6
789 	case AF_INET6:
790 		ip = NULL;
791 		ip6 = mtod(m, struct ip6_hdr *);
792 		th = (struct tcphdr *)(ip6 + 1);
793 		break;
794 #endif
795 	default:	/*pacify gcc*/
796 		ip = NULL;
797 #ifdef INET6
798 		ip6 = NULL;
799 #endif
800 		th = NULL;
801 		break;
802 	}
803 	if (tp->t_template == 0)
804 		panic("tcp_output");
805 	if (tp->t_template->m_len < iphdrlen)
806 		panic("tcp_output");
807 	bcopy(mtod(tp->t_template, caddr_t), mtod(m, caddr_t), iphdrlen);
808 
809 	/*
810 	 * If we are doing retransmissions, then snd_nxt will
811 	 * not reflect the first unsent octet.  For ACK only
812 	 * packets, we do not want the sequence number of the
813 	 * retransmitted packet, we want the sequence number
814 	 * of the next unsent octet.  So, if there is no data
815 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
816 	 * when filling in ti_seq.  But if we are in persist
817 	 * state, snd_max might reflect one byte beyond the
818 	 * right edge of the window, so use snd_nxt in that
819 	 * case, since we know we aren't doing a retransmission.
820 	 * (retransmit and persist are mutually exclusive...)
821 	 */
822 	if (len || (flags & (TH_SYN|TH_FIN)) ||
823 	    TCP_TIMER_ISARMED(tp, TCPT_PERSIST))
824 		th->th_seq = htonl(tp->snd_nxt);
825 	else
826 		th->th_seq = htonl(tp->snd_max);
827 	th->th_ack = htonl(tp->rcv_nxt);
828 	if (optlen) {
829 		bcopy((caddr_t)opt, (caddr_t)(th + 1), optlen);
830 		th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
831 	}
832 	th->th_flags = flags;
833 	/*
834 	 * Calculate receive window.  Don't shrink window,
835 	 * but avoid silly window syndrome.
836 	 */
837 	if (win < (long)(so->so_rcv.sb_hiwat / 4) && win < (long)rxsegsize)
838 		win = 0;
839 	if (win > (long)TCP_MAXWIN << tp->rcv_scale)
840 		win = (long)TCP_MAXWIN << tp->rcv_scale;
841 	if (win < (long)(tp->rcv_adv - tp->rcv_nxt))
842 		win = (long)(tp->rcv_adv - tp->rcv_nxt);
843 	th->th_win = htons((u_int16_t) (win>>tp->rcv_scale));
844 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
845 		u_int32_t urp = tp->snd_up - tp->snd_nxt;
846 		if (urp > IP_MAXPACKET)
847 			urp = IP_MAXPACKET;
848 		th->th_urp = htons((u_int16_t)urp);
849 		th->th_flags |= TH_URG;
850 	} else
851 		/*
852 		 * If no urgent pointer to send, then we pull
853 		 * the urgent pointer to the left edge of the send window
854 		 * so that it doesn't drift into the send window on sequence
855 		 * number wraparound.
856 		 */
857 		tp->snd_up = tp->snd_una;		/* drag it along */
858 
859 	/*
860 	 * Put TCP length in extended header, and then
861 	 * checksum extended header and data.
862 	 */
863 	switch (af) {
864 #ifdef INET
865 	case AF_INET:
866 	    {
867 		struct ipovly *ipov = (struct ipovly *)ip;
868 		if (len + optlen)
869 			ipov->ih_len = htons((u_int16_t)(sizeof(struct tcphdr) +
870 			    optlen + len));
871 		bzero(ipov->ih_x1, sizeof(ipov->ih_x1));
872 		th->th_sum = 0;
873 		th->th_sum = in_cksum(m, (int)(hdrlen + len));
874 		break;
875 	    }
876 #endif
877 #ifdef INET6
878 	case AF_INET6:
879 		/* equals to hdrlen + len */
880 		m->m_pkthdr.len = sizeof(struct ip6_hdr)
881 			+ sizeof(struct tcphdr) + optlen + len;
882 		th->th_sum = 0;
883 		th->th_sum = in6_cksum(m, IPPROTO_TCP,
884 				sizeof(struct ip6_hdr),
885 				sizeof(struct tcphdr) + optlen + len);
886 		break;
887 #endif
888 	}
889 
890 	/*
891 	 * In transmit state, time the transmission and arrange for
892 	 * the retransmit.  In persist state, just set snd_max.
893 	 */
894 	if (tp->t_force == 0 || TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) {
895 		tcp_seq startseq = tp->snd_nxt;
896 
897 		/*
898 		 * Advance snd_nxt over sequence space of this segment.
899 		 * There are no states in which we send both a SYN and a FIN,
900 		 * so we collapse the tests for these flags.
901 		 */
902 		if (flags & (TH_SYN|TH_FIN))
903 			tp->snd_nxt++;
904 		tp->snd_nxt += len;
905 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
906 			tp->snd_max = tp->snd_nxt;
907 			/*
908 			 * Time this transmission if not a retransmission and
909 			 * not currently timing anything.
910 			 */
911 			if (tp->t_rtt == 0) {
912 				tp->t_rtt = 1;
913 				tp->t_rtseq = startseq;
914 				tcpstat.tcps_segstimed++;
915 			}
916 		}
917 
918 		/*
919 		 * Set retransmit timer if not currently set,
920 		 * and not doing an ack or a keep-alive probe.
921 		 * Initial value for retransmit timer is smoothed
922 		 * round-trip time + 2 * round-trip time variance.
923 		 * Initialize shift counter which is used for backoff
924 		 * of retransmit time.
925 		 */
926 		if (TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 &&
927 		    tp->snd_nxt != tp->snd_una) {
928 			TCP_TIMER_ARM(tp, TCPT_REXMT, tp->t_rxtcur);
929 			if (TCP_TIMER_ISARMED(tp, TCPT_PERSIST)) {
930 				TCP_TIMER_DISARM(tp, TCPT_PERSIST);
931 				tp->t_rxtshift = 0;
932 			}
933 		}
934 	} else
935 		if (SEQ_GT(tp->snd_nxt + len, tp->snd_max))
936 			tp->snd_max = tp->snd_nxt + len;
937 
938 	/*
939 	 * Trace.
940 	 */
941 	if (so->so_options & SO_DEBUG) {
942 		/*
943 		 * need to recover version # field, which was overwritten
944 		 * on ip_cksum computation.
945 		 */
946 		struct ip *sip;
947 		sip = mtod(m, struct ip *);
948 		switch (af) {
949 #ifdef INET
950 		case AF_INET:
951 			sip->ip_v = 4;
952 			break;
953 #endif
954 #ifdef INET6
955 		case AF_INET6:
956 			sip->ip_v = 6;
957 			break;
958 #endif
959 		}
960 		tcp_trace(TA_OUTPUT, tp->t_state, tp, m, 0);
961 	}
962 
963 	/*
964 	 * Fill in IP length and desired time to live and
965 	 * send to IP level.  There should be a better way
966 	 * to handle ttl and tos; we could keep them in
967 	 * the template, but need a way to checksum without them.
968 	 */
969 	m->m_pkthdr.len = hdrlen + len;
970 
971 	switch (af) {
972 #ifdef INET
973 	case AF_INET:
974 		ip->ip_len = m->m_pkthdr.len;
975 		if (tp->t_inpcb) {
976 			ip->ip_ttl = tp->t_inpcb->inp_ip.ip_ttl;
977 			ip->ip_tos = tp->t_inpcb->inp_ip.ip_tos;
978 		}
979 #ifdef INET6
980 		else if (tp->t_in6pcb) {
981 			ip->ip_ttl = in6_selecthlim(tp->t_in6pcb, NULL); /*XXX*/
982 			ip->ip_tos = 0;	/*XXX*/
983 		}
984 #endif
985 		break;
986 #endif
987 #ifdef INET6
988 	case AF_INET6:
989 		ip6->ip6_nxt = IPPROTO_TCP;
990 		if (tp->t_in6pcb) {
991 			/*
992 			 * we separately set hoplimit for every segment, since
993 			 * the user might want to change the value via
994 			 * setsockopt. Also, desired default hop limit might
995 			 * be changed via Neighbor Discovery.
996 			 */
997 			ip6->ip6_hlim = in6_selecthlim(tp->t_in6pcb,
998 				ro->ro_rt ? ro->ro_rt->rt_ifp : NULL);
999 		}
1000 		/* ip6->ip6_flow = ??? */
1001 		/* ip6_plen will be filled in ip6_output(). */
1002 		break;
1003 #endif
1004 	}
1005 
1006 #ifdef IPSEC
1007 	if (ipsec_setsocket(m, so) != 0) {
1008 		m_freem(m);
1009 		error = ENOBUFS;
1010 		goto out;
1011 	}
1012 #endif /*IPSEC*/
1013 
1014 	switch (af) {
1015 #ifdef INET
1016 	case AF_INET:
1017 	    {
1018 		struct mbuf *opts;
1019 
1020 		if (tp->t_inpcb)
1021 			opts = tp->t_inpcb->inp_options;
1022 		else
1023 			opts = NULL;
1024 		error = ip_output(m, opts, ro,
1025 			(ip_mtudisc ? IP_MTUDISC : 0) |
1026 			(so->so_options & SO_DONTROUTE),
1027 			0);
1028 		break;
1029 	    }
1030 #endif
1031 #ifdef INET6
1032 	case AF_INET6:
1033 	    {
1034 		struct ip6_pktopts *opts;
1035 
1036 		if (tp->t_in6pcb)
1037 			opts = tp->t_in6pcb->in6p_outputopts;
1038 		else
1039 			opts = NULL;
1040 		error = ip6_output(m, opts, (struct route_in6 *)ro,
1041 			so->so_options & SO_DONTROUTE, 0, NULL);
1042 		break;
1043 	    }
1044 #endif
1045 	default:
1046 		error = EAFNOSUPPORT;
1047 		break;
1048 	}
1049 	if (error) {
1050 out:
1051 		if (error == ENOBUFS) {
1052 #ifdef INET
1053 			if (tp->t_inpcb)
1054 				tcp_quench(tp->t_inpcb, 0);
1055 #endif
1056 #ifdef INET6
1057 			if (tp->t_in6pcb)
1058 				tcp6_quench(tp->t_in6pcb, 0);
1059 #endif
1060 			return (0);
1061 		}
1062 		if ((error == EHOSTUNREACH || error == ENETDOWN)
1063 		    && TCPS_HAVERCVDSYN(tp->t_state)) {
1064 			tp->t_softerror = error;
1065 			return (0);
1066 		}
1067 		return (error);
1068 	}
1069 	tcpstat.tcps_sndtotal++;
1070 	if (tp->t_flags & TF_DELACK)
1071 		tcpstat.tcps_delack++;
1072 
1073 	/*
1074 	 * Data sent (as far as we can tell).
1075 	 * If this advertises a larger window than any other segment,
1076 	 * then remember the size of the advertised window.
1077 	 * Any pending ACK has now been sent.
1078 	 */
1079 	if (win > 0 && SEQ_GT(tp->rcv_nxt+win, tp->rcv_adv))
1080 		tp->rcv_adv = tp->rcv_nxt + win;
1081 	tp->last_ack_sent = tp->rcv_nxt;
1082 	tp->t_flags &= ~TF_ACKNOW;
1083 	TCP_CLEAR_DELACK(tp);
1084 #ifdef DIAGNOSTIC
1085 	if (maxburst < 0)
1086 		printf("tcp_output: maxburst exceeded by %d\n", -maxburst);
1087 #endif
1088 	if (sendalot && (!tcp_do_newreno || --maxburst))
1089 		goto again;
1090 	return (0);
1091 }
1092 
1093 void
1094 tcp_setpersist(tp)
1095 	struct tcpcb *tp;
1096 {
1097 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> (1 + 2);
1098 	int nticks;
1099 
1100 	if (TCP_TIMER_ISARMED(tp, TCPT_REXMT))
1101 		panic("tcp_output REXMT");
1102 	/*
1103 	 * Start/restart persistance timer.
1104 	 */
1105 	if (t < tp->t_rttmin)
1106 		t = tp->t_rttmin;
1107 	TCPT_RANGESET(nticks, t * tcp_backoff[tp->t_rxtshift],
1108 	    TCPTV_PERSMIN, TCPTV_PERSMAX);
1109 	TCP_TIMER_ARM(tp, TCPT_PERSIST, nticks);
1110 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1111 		tp->t_rxtshift++;
1112 }
1113