xref: /netbsd-src/sys/netinet/tcp_output.c (revision 21e37cc72a480a47828990a439cde7ac9ffaf0c6)
1 /*	$NetBSD: tcp_output.c,v 1.114 2004/05/20 22:59:02 jonathan Exp $	*/
2 
3 /*
4  * Copyright (C) 1995, 1996, 1997, and 1998 WIDE Project.
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  * 3. Neither the name of the project nor the names of its contributors
16  *    may be used to endorse or promote products derived from this software
17  *    without specific prior written permission.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE PROJECT AND CONTRIBUTORS ``AS IS'' AND
20  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
21  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
22  * ARE DISCLAIMED.  IN NO EVENT SHALL THE PROJECT OR CONTRIBUTORS BE LIABLE
23  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
24  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
25  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
26  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
28  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
29  * SUCH DAMAGE.
30  */
31 
32 /*
33  *      @(#)COPYRIGHT   1.1 (NRL) 17 January 1995
34  *
35  * NRL grants permission for redistribution and use in source and binary
36  * forms, with or without modification, of the software and documentation
37  * created at NRL provided that the following conditions are met:
38  *
39  * 1. Redistributions of source code must retain the above copyright
40  *    notice, this list of conditions and the following disclaimer.
41  * 2. Redistributions in binary form must reproduce the above copyright
42  *    notice, this list of conditions and the following disclaimer in the
43  *    documentation and/or other materials provided with the distribution.
44  * 3. All advertising materials mentioning features or use of this software
45  *    must display the following acknowledgements:
46  *      This product includes software developed by the University of
47  *      California, Berkeley and its contributors.
48  *      This product includes software developed at the Information
49  *      Technology Division, US Naval Research Laboratory.
50  * 4. Neither the name of the NRL nor the names of its contributors
51  *    may be used to endorse or promote products derived from this software
52  *    without specific prior written permission.
53  *
54  * THE SOFTWARE PROVIDED BY NRL IS PROVIDED BY NRL AND CONTRIBUTORS ``AS
55  * IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
56  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A
57  * PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL NRL OR
58  * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
59  * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
60  * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
61  * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
62  * LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
63  * NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
64  * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
65  *
66  * The views and conclusions contained in the software and documentation
67  * are those of the authors and should not be interpreted as representing
68  * official policies, either expressed or implied, of the US Naval
69  * Research Laboratory (NRL).
70  */
71 
72 /*-
73  * Copyright (c) 1997, 1998, 2001 The NetBSD Foundation, Inc.
74  * All rights reserved.
75  *
76  * This code is derived from software contributed to The NetBSD Foundation
77  * by Jason R. Thorpe and Kevin M. Lahey of the Numerical Aerospace Simulation
78  * Facility, NASA Ames Research Center.
79  *
80  * Redistribution and use in source and binary forms, with or without
81  * modification, are permitted provided that the following conditions
82  * are met:
83  * 1. Redistributions of source code must retain the above copyright
84  *    notice, this list of conditions and the following disclaimer.
85  * 2. Redistributions in binary form must reproduce the above copyright
86  *    notice, this list of conditions and the following disclaimer in the
87  *    documentation and/or other materials provided with the distribution.
88  * 3. All advertising materials mentioning features or use of this software
89  *    must display the following acknowledgement:
90  *	This product includes software developed by the NetBSD
91  *	Foundation, Inc. and its contributors.
92  * 4. Neither the name of The NetBSD Foundation nor the names of its
93  *    contributors may be used to endorse or promote products derived
94  *    from this software without specific prior written permission.
95  *
96  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
97  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
98  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
99  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
100  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
101  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
102  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
103  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
104  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
105  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
106  * POSSIBILITY OF SUCH DAMAGE.
107  */
108 
109 /*
110  * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995
111  *	The Regents of the University of California.  All rights reserved.
112  *
113  * Redistribution and use in source and binary forms, with or without
114  * modification, are permitted provided that the following conditions
115  * are met:
116  * 1. Redistributions of source code must retain the above copyright
117  *    notice, this list of conditions and the following disclaimer.
118  * 2. Redistributions in binary form must reproduce the above copyright
119  *    notice, this list of conditions and the following disclaimer in the
120  *    documentation and/or other materials provided with the distribution.
121  * 3. Neither the name of the University nor the names of its contributors
122  *    may be used to endorse or promote products derived from this software
123  *    without specific prior written permission.
124  *
125  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
126  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
127  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
128  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
129  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
130  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
131  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
132  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
133  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
134  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
135  * SUCH DAMAGE.
136  *
137  *	@(#)tcp_output.c	8.4 (Berkeley) 5/24/95
138  */
139 
140 #include <sys/cdefs.h>
141 __KERNEL_RCSID(0, "$NetBSD: tcp_output.c,v 1.114 2004/05/20 22:59:02 jonathan Exp $");
142 
143 #include "opt_inet.h"
144 #include "opt_ipsec.h"
145 #include "opt_tcp_debug.h"
146 
147 #include <sys/param.h>
148 #include <sys/systm.h>
149 #include <sys/malloc.h>
150 #include <sys/mbuf.h>
151 #include <sys/protosw.h>
152 #include <sys/socket.h>
153 #include <sys/socketvar.h>
154 #include <sys/errno.h>
155 #include <sys/domain.h>
156 #include <sys/kernel.h>
157 #ifdef TCP_SIGNATURE
158 #include <sys/md5.h>
159 #endif
160 
161 #include <net/if.h>
162 #include <net/route.h>
163 
164 #include <netinet/in.h>
165 #include <netinet/in_systm.h>
166 #include <netinet/ip.h>
167 #include <netinet/in_pcb.h>
168 #include <netinet/ip_var.h>
169 
170 #ifdef INET6
171 #ifndef INET
172 #include <netinet/in.h>
173 #endif
174 #include <netinet/ip6.h>
175 #include <netinet6/in6_var.h>
176 #include <netinet6/ip6_var.h>
177 #include <netinet6/in6_pcb.h>
178 #include <netinet6/nd6.h>
179 #endif
180 
181 #ifdef FAST_IPSEC
182 #include <netipsec/ipsec.h>
183 #include <netipsec/key.h>
184 #endif	/* FAST_IPSEC*/
185 #ifdef IPSEC
186 #include <netinet6/ipsec.h>
187 #endif
188 
189 #include <netinet/tcp.h>
190 #define	TCPOUTFLAGS
191 #include <netinet/tcp_fsm.h>
192 #include <netinet/tcp_seq.h>
193 #include <netinet/tcp_timer.h>
194 #include <netinet/tcp_var.h>
195 #include <netinet/tcpip.h>
196 #include <netinet/tcp_debug.h>
197 
198 #ifdef IPSEC
199 #include <netkey/key.h>
200 #endif
201 
202 #ifdef notyet
203 extern struct mbuf *m_copypack();
204 #endif
205 
206 #define MAX_TCPOPTLEN	32	/* max # bytes that go in options */
207 
208 /*
209  * Knob to enable Congestion Window Monitoring, and control the
210  * the burst size it allows.  Default burst is 4 packets, per
211  * the Internet draft.
212  */
213 int	tcp_cwm = 0;
214 int	tcp_cwm_burstsize = 4;
215 
216 #ifdef TCP_OUTPUT_COUNTERS
217 #include <sys/device.h>
218 
219 extern struct evcnt tcp_output_bigheader;
220 extern struct evcnt tcp_output_predict_hit;
221 extern struct evcnt tcp_output_predict_miss;
222 extern struct evcnt tcp_output_copysmall;
223 extern struct evcnt tcp_output_copybig;
224 extern struct evcnt tcp_output_refbig;
225 
226 #define	TCP_OUTPUT_COUNTER_INCR(ev)	(ev)->ev_count++
227 #else
228 
229 #define	TCP_OUTPUT_COUNTER_INCR(ev)	/* nothing */
230 
231 #endif /* TCP_OUTPUT_COUNTERS */
232 
233 static
234 #ifndef GPROF
235 __inline
236 #endif
237 int
238 tcp_segsize(struct tcpcb *tp, int *txsegsizep, int *rxsegsizep)
239 {
240 #ifdef INET
241 	struct inpcb *inp = tp->t_inpcb;
242 #endif
243 #ifdef INET6
244 	struct in6pcb *in6p = tp->t_in6pcb;
245 #endif
246 	struct socket *so = NULL;
247 	struct rtentry *rt;
248 	struct ifnet *ifp;
249 	int size;
250 	int iphlen;
251 	int optlen;
252 
253 #ifdef DIAGNOSTIC
254 	if (tp->t_inpcb && tp->t_in6pcb)
255 		panic("tcp_segsize: both t_inpcb and t_in6pcb are set");
256 #endif
257 	switch (tp->t_family) {
258 #ifdef INET
259 	case AF_INET:
260 		iphlen = sizeof(struct ip);
261 		break;
262 #endif
263 #ifdef INET6
264 	case AF_INET6:
265 		iphlen = sizeof(struct ip6_hdr);
266 		break;
267 #endif
268 	default:
269 		size = tcp_mssdflt;
270 		goto out;
271 	}
272 
273 	rt = NULL;
274 #ifdef INET
275 	if (inp) {
276 		rt = in_pcbrtentry(inp);
277 		so = inp->inp_socket;
278 	}
279 #endif
280 #ifdef INET6
281 	if (in6p) {
282 		rt = in6_pcbrtentry(in6p);
283 		so = in6p->in6p_socket;
284 	}
285 #endif
286 	if (rt == NULL) {
287 		size = tcp_mssdflt;
288 		goto out;
289 	}
290 
291 	ifp = rt->rt_ifp;
292 
293 	size = tcp_mssdflt;
294 	if (tp->t_mtudisc && rt->rt_rmx.rmx_mtu != 0) {
295 #ifdef INET6
296 		if (in6p && rt->rt_rmx.rmx_mtu < IPV6_MMTU) {
297 			/*
298 			 * RFC2460 section 5, last paragraph: if path MTU is
299 			 * smaller than 1280, use 1280 as packet size and
300 			 * attach fragment header.
301 			 */
302 			size = IPV6_MMTU - iphlen - sizeof(struct ip6_frag) -
303 			    sizeof(struct tcphdr);
304 		} else
305 			size = rt->rt_rmx.rmx_mtu - iphlen -
306 			    sizeof(struct tcphdr);
307 #else
308 		size = rt->rt_rmx.rmx_mtu - iphlen - sizeof(struct tcphdr);
309 #endif
310 	} else if (ifp->if_flags & IFF_LOOPBACK)
311 		size = ifp->if_mtu - iphlen - sizeof(struct tcphdr);
312 #ifdef INET
313 	else if (inp && tp->t_mtudisc)
314 		size = ifp->if_mtu - iphlen - sizeof(struct tcphdr);
315 	else if (inp && in_localaddr(inp->inp_faddr))
316 		size = ifp->if_mtu - iphlen - sizeof(struct tcphdr);
317 #endif
318 #ifdef INET6
319 	else if (in6p) {
320 #ifdef INET
321 		if (IN6_IS_ADDR_V4MAPPED(&in6p->in6p_faddr)) {
322 			/* mapped addr case */
323 			struct in_addr d;
324 			bcopy(&in6p->in6p_faddr.s6_addr32[3], &d, sizeof(d));
325 			if (tp->t_mtudisc || in_localaddr(d))
326 				size = ifp->if_mtu - iphlen - sizeof(struct tcphdr);
327 		} else
328 #endif
329 		{
330 			/*
331 			 * for IPv6, path MTU discovery is always turned on,
332 			 * or the node must use packet size <= 1280.
333 			 */
334 			size = tp->t_mtudisc ? IN6_LINKMTU(ifp) : IPV6_MMTU;
335 			size -= (iphlen + sizeof(struct tcphdr));
336 		}
337 	}
338 #endif
339  out:
340 	/*
341 	 * Now we must make room for whatever extra TCP/IP options are in
342 	 * the packet.
343 	 */
344 	optlen = tcp_optlen(tp);
345 
346 	/*
347 	 * XXX tp->t_ourmss should have the right size, but without this code
348 	 * fragmentation will occur... need more investigation
349 	 */
350 #ifdef INET
351 	if (inp) {
352 #if defined(IPSEC) || defined(FAST_IPSEC)
353 		if (! IPSEC_PCB_SKIP_IPSEC(inp->inp_sp, IPSEC_DIR_OUTBOUND))
354 			optlen += ipsec4_hdrsiz_tcp(tp);
355 #endif
356 		optlen += ip_optlen(inp);
357 	}
358 #endif
359 #ifdef INET6
360 #ifdef INET
361 	if (in6p && tp->t_family == AF_INET) {
362 #if defined(IPSEC) || defined(FAST_IPSEC)
363 		if (! IPSEC_PCB_SKIP_IPSEC(in6p->in6p_sp, IPSEC_DIR_OUTBOUND))
364 			optlen += ipsec4_hdrsiz_tcp(tp);
365 #endif
366 		/* XXX size -= ip_optlen(in6p); */
367 	} else
368 #endif
369 	if (in6p && tp->t_family == AF_INET6) {
370 #ifdef IPSEC
371 		if (! IPSEC_PCB_SKIP_IPSEC(in6p->in6p_sp, IPSEC_DIR_OUTBOUND))
372 			optlen += ipsec6_hdrsiz_tcp(tp);
373 #endif
374 		optlen += ip6_optlen(in6p);
375 	}
376 #endif
377 	size -= optlen;
378 
379 	/* there may not be any room for data if mtu is too small */
380 	if (size < 0)
381 		return (EMSGSIZE);
382 
383 	/*
384 	 * *rxsegsizep holds *estimated* inbound segment size (estimation
385 	 * assumes that path MTU is the same for both ways).  this is only
386 	 * for silly window avoidance, do not use the value for other purposes.
387 	 *
388 	 * ipseclen is subtracted from both sides, this may not be right.
389 	 * I'm not quite sure about this (could someone comment).
390 	 */
391 	*txsegsizep = min(tp->t_peermss - optlen, size);
392 	/*
393 	 * Never send more than half a buffer full.  This insures that we can
394 	 * always keep 2 packets on the wire, no matter what SO_SNDBUF is, and
395 	 * therefore acks will never be delayed unless we run out of data to
396 	 * transmit.
397 	 */
398 	if (so)
399 		*txsegsizep = min(so->so_snd.sb_hiwat >> 1, *txsegsizep);
400 	*rxsegsizep = min(tp->t_ourmss - optlen, size);
401 
402 	if (*txsegsizep != tp->t_segsz) {
403 		/*
404 		 * If the new segment size is larger, we don't want to
405 		 * mess up the congestion window, but if it is smaller
406 		 * we'll have to reduce the congestion window to ensure
407 		 * that we don't get into trouble with initial windows
408 		 * and the rest.  In any case, if the segment size
409 		 * has changed, chances are the path has, too, and
410 		 * our congestion window will be different.
411 		 */
412 		if (*txsegsizep < tp->t_segsz) {
413 			tp->snd_cwnd = max((tp->snd_cwnd / tp->t_segsz)
414 					   * *txsegsizep, *txsegsizep);
415 			tp->snd_ssthresh = max((tp->snd_ssthresh / tp->t_segsz)
416 						* *txsegsizep, *txsegsizep);
417 		}
418 		tp->t_segsz = *txsegsizep;
419 	}
420 
421 	return (0);
422 }
423 
424 static
425 #ifndef GPROF
426 __inline
427 #endif
428 int
429 tcp_build_datapkt(struct tcpcb *tp, struct socket *so, int off,
430     long len, int hdrlen, struct mbuf **mp)
431 {
432 	struct mbuf *m, *m0;
433 
434 	if (tp->t_force && len == 1)
435 		tcpstat.tcps_sndprobe++;
436 	else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
437 		tcpstat.tcps_sndrexmitpack++;
438 		tcpstat.tcps_sndrexmitbyte += len;
439 	} else {
440 		tcpstat.tcps_sndpack++;
441 		tcpstat.tcps_sndbyte += len;
442 	}
443 #ifdef notyet
444 	if ((m = m_copypack(so->so_snd.sb_mb, off,
445 	    (int)len, max_linkhdr + hdrlen)) == 0)
446 		return (ENOBUFS);
447 	/*
448 	 * m_copypack left space for our hdr; use it.
449 	 */
450 	m->m_len += hdrlen;
451 	m->m_data -= hdrlen;
452 #else
453 	MGETHDR(m, M_DONTWAIT, MT_HEADER);
454 	if (__predict_false(m == NULL))
455 		return (ENOBUFS);
456 	MCLAIM(m, &tcp_tx_mowner);
457 
458 	/*
459 	 * XXX Because other code assumes headers will fit in
460 	 * XXX one header mbuf.
461 	 *
462 	 * (This code should almost *never* be run.)
463 	 */
464 	if (__predict_false((max_linkhdr + hdrlen) > MHLEN)) {
465 		TCP_OUTPUT_COUNTER_INCR(&tcp_output_bigheader);
466 		MCLGET(m, M_DONTWAIT);
467 		if ((m->m_flags & M_EXT) == 0) {
468 			m_freem(m);
469 			return (ENOBUFS);
470 		}
471 	}
472 
473 	m->m_data += max_linkhdr;
474 	m->m_len = hdrlen;
475 
476 	/*
477 	 * To avoid traversing the whole sb_mb chain for correct
478 	 * data to send, remember last sent mbuf, its offset and
479 	 * the sent size.  When called the next time, see if the
480 	 * data to send is directly following the previous transfer.
481 	 * This is important for large TCP windows.
482 	 */
483 	if (off == 0 || tp->t_lastm == NULL ||
484 	    (tp->t_lastoff + tp->t_lastlen) != off) {
485 		TCP_OUTPUT_COUNTER_INCR(&tcp_output_predict_miss);
486 		/*
487 		 * Either a new packet or a retransmit.
488 		 * Start from the beginning.
489 		 */
490 		tp->t_lastm = so->so_snd.sb_mb;
491 		tp->t_inoff = off;
492 	} else {
493 		TCP_OUTPUT_COUNTER_INCR(&tcp_output_predict_hit);
494 		tp->t_inoff += tp->t_lastlen;
495 	}
496 
497 	/* Traverse forward to next packet */
498 	while (tp->t_inoff > 0) {
499 		if (tp->t_lastm == NULL)
500 			panic("tp->t_lastm == NULL");
501 		if (tp->t_inoff < tp->t_lastm->m_len)
502 			break;
503 		tp->t_inoff -= tp->t_lastm->m_len;
504 		tp->t_lastm = tp->t_lastm->m_next;
505 	}
506 
507 	tp->t_lastoff = off;
508 	tp->t_lastlen = len;
509 	m0 = tp->t_lastm;
510 	off = tp->t_inoff;
511 
512 	if (len <= M_TRAILINGSPACE(m)) {
513 		m_copydata(m0, off, (int) len, mtod(m, caddr_t) + hdrlen);
514 		m->m_len += len;
515 		TCP_OUTPUT_COUNTER_INCR(&tcp_output_copysmall);
516 	} else {
517 		m->m_next = m_copy(m0, off, (int) len);
518 		if (m->m_next == NULL) {
519 			m_freem(m);
520 			return (ENOBUFS);
521 		}
522 #ifdef TCP_OUTPUT_COUNTERS
523 		if (m->m_next->m_flags & M_EXT)
524 			TCP_OUTPUT_COUNTER_INCR(&tcp_output_refbig);
525 		else
526 			TCP_OUTPUT_COUNTER_INCR(&tcp_output_copybig);
527 #endif /* TCP_OUTPUT_COUNTERS */
528 	}
529 #endif
530 
531 	*mp = m;
532 	return (0);
533 }
534 
535 /*
536  * Tcp output routine: figure out what should be sent and send it.
537  */
538 int
539 tcp_output(tp)
540 	struct tcpcb *tp;
541 {
542 	struct socket *so;
543 	struct route *ro;
544 	long len, win;
545 	int off, flags, error;
546 	struct mbuf *m;
547 	struct ip *ip;
548 #ifdef INET6
549 	struct ip6_hdr *ip6;
550 #endif
551 	struct tcphdr *th;
552 	u_char opt[MAX_TCPOPTLEN];
553 	unsigned optlen, hdrlen;
554 	int idle, sendalot, txsegsize, rxsegsize;
555 	int maxburst = TCP_MAXBURST;
556 	int af;		/* address family on the wire */
557 	int iphdrlen;
558 #ifdef TCP_SIGNATURE
559 	int sigoff = 0;
560 #endif
561 
562 #ifdef DIAGNOSTIC
563 	if (tp->t_inpcb && tp->t_in6pcb)
564 		panic("tcp_output: both t_inpcb and t_in6pcb are set");
565 #endif
566 	so = NULL;
567 	ro = NULL;
568 	if (tp->t_inpcb) {
569 		so = tp->t_inpcb->inp_socket;
570 		ro = &tp->t_inpcb->inp_route;
571 	}
572 #ifdef INET6
573 	else if (tp->t_in6pcb) {
574 		so = tp->t_in6pcb->in6p_socket;
575 		ro = (struct route *)&tp->t_in6pcb->in6p_route;
576 	}
577 #endif
578 
579 	switch (af = tp->t_family) {
580 #ifdef INET
581 	case AF_INET:
582 		if (tp->t_inpcb)
583 			break;
584 #ifdef INET6
585 		/* mapped addr case */
586 		if (tp->t_in6pcb)
587 			break;
588 #endif
589 		return (EINVAL);
590 #endif
591 #ifdef INET6
592 	case AF_INET6:
593 		if (tp->t_in6pcb)
594 			break;
595 		return (EINVAL);
596 #endif
597 	default:
598 		return (EAFNOSUPPORT);
599 	}
600 
601 	if (tcp_segsize(tp, &txsegsize, &rxsegsize))
602 		return (EMSGSIZE);
603 
604 	idle = (tp->snd_max == tp->snd_una);
605 
606 	/*
607 	 * Restart Window computation.  From draft-floyd-incr-init-win-03:
608 	 *
609 	 *	Optionally, a TCP MAY set the restart window to the
610 	 *	minimum of the value used for the initial window and
611 	 *	the current value of cwnd (in other words, using a
612 	 *	larger value for the restart window should never increase
613 	 *	the size of cwnd).
614 	 */
615 	if (tcp_cwm) {
616 		/*
617 		 * Hughes/Touch/Heidemann Congestion Window Monitoring.
618 		 * Count the number of packets currently pending
619 		 * acknowledgement, and limit our congestion window
620 		 * to a pre-determined allowed burst size plus that count.
621 		 * This prevents bursting once all pending packets have
622 		 * been acknowledged (i.e. transmission is idle).
623 		 *
624 		 * XXX Link this to Initial Window?
625 		 */
626 		tp->snd_cwnd = min(tp->snd_cwnd,
627 		    (tcp_cwm_burstsize * txsegsize) +
628 		    (tp->snd_nxt - tp->snd_una));
629 	} else {
630 		if (idle && (tcp_now - tp->t_rcvtime) >= tp->t_rxtcur) {
631 			/*
632 			 * We have been idle for "a while" and no acks are
633 			 * expected to clock out any data we send --
634 			 * slow start to get ack "clock" running again.
635 			 */
636 			int ss = tcp_init_win;
637 #ifdef INET
638 			if (tp->t_inpcb &&
639 			    in_localaddr(tp->t_inpcb->inp_faddr))
640 				ss = tcp_init_win_local;
641 #endif
642 #ifdef INET6
643 			if (tp->t_in6pcb &&
644 			    in6_localaddr(&tp->t_in6pcb->in6p_faddr))
645 				ss = tcp_init_win_local;
646 #endif
647 			tp->snd_cwnd = min(tp->snd_cwnd,
648 			    TCP_INITIAL_WINDOW(ss, txsegsize));
649 		}
650 	}
651 
652 again:
653 	/*
654 	 * Determine length of data that should be transmitted, and
655 	 * flags that should be used.  If there is some data or critical
656 	 * controls (SYN, RST) to send, then transmit; otherwise,
657 	 * investigate further.
658 	 */
659 	sendalot = 0;
660 	off = tp->snd_nxt - tp->snd_una;
661 	win = min(tp->snd_wnd, tp->snd_cwnd);
662 
663 	flags = tcp_outflags[tp->t_state];
664 	/*
665 	 * If in persist timeout with window of 0, send 1 byte.
666 	 * Otherwise, if window is small but nonzero
667 	 * and timer expired, we will send what we can
668 	 * and go to transmit state.
669 	 */
670 	if (tp->t_force) {
671 		if (win == 0) {
672 			/*
673 			 * If we still have some data to send, then
674 			 * clear the FIN bit.  Usually this would
675 			 * happen below when it realizes that we
676 			 * aren't sending all the data.  However,
677 			 * if we have exactly 1 byte of unset data,
678 			 * then it won't clear the FIN bit below,
679 			 * and if we are in persist state, we wind
680 			 * up sending the packet without recording
681 			 * that we sent the FIN bit.
682 			 *
683 			 * We can't just blindly clear the FIN bit,
684 			 * because if we don't have any more data
685 			 * to send then the probe will be the FIN
686 			 * itself.
687 			 */
688 			if (off < so->so_snd.sb_cc)
689 				flags &= ~TH_FIN;
690 			win = 1;
691 		} else {
692 			TCP_TIMER_DISARM(tp, TCPT_PERSIST);
693 			tp->t_rxtshift = 0;
694 		}
695 	}
696 
697 	if (win < so->so_snd.sb_cc) {
698 		len = win - off;
699 		flags &= ~TH_FIN;
700 	} else
701 		len = so->so_snd.sb_cc - off;
702 
703 	if (len < 0) {
704 		/*
705 		 * If FIN has been sent but not acked,
706 		 * but we haven't been called to retransmit,
707 		 * len will be -1.  Otherwise, window shrank
708 		 * after we sent into it.  If window shrank to 0,
709 		 * cancel pending retransmit, pull snd_nxt back
710 		 * to (closed) window, and set the persist timer
711 		 * if it isn't already going.  If the window didn't
712 		 * close completely, just wait for an ACK.
713 		 *
714 		 * If we have a pending FIN, either it has already been
715 		 * transmitted or it is outside the window, so drop it.
716 		 * If the FIN has been transmitted, but this is not a
717 		 * retransmission, then len must be -1.  Therefore we also
718 		 * prevent here the sending of `gratuitous FINs'.  This
719 		 * eliminates the need to check for that case below (e.g.
720 		 * to back up snd_nxt before the FIN so that the sequence
721 		 * number is correct).
722 		 */
723 		len = 0;
724 		flags &= ~TH_FIN;
725 		if (win == 0) {
726 			TCP_TIMER_DISARM(tp, TCPT_REXMT);
727 			tp->t_rxtshift = 0;
728 			tp->snd_nxt = tp->snd_una;
729 			if (TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0)
730 				tcp_setpersist(tp);
731 		}
732 	}
733 	if (len > txsegsize) {
734 		len = txsegsize;
735 		flags &= ~TH_FIN;
736 		sendalot = 1;
737 	}
738 
739 	win = sbspace(&so->so_rcv);
740 
741 	/*
742 	 * Sender silly window avoidance.  If connection is idle
743 	 * and can send all data, a maximum segment,
744 	 * at least a maximum default-size segment do it,
745 	 * or are forced, do it; otherwise don't bother.
746 	 * If peer's buffer is tiny, then send
747 	 * when window is at least half open.
748 	 * If retransmitting (possibly after persist timer forced us
749 	 * to send into a small window), then must resend.
750 	 */
751 	if (len) {
752 		if (len == txsegsize)
753 			goto send;
754 		if ((so->so_state & SS_MORETOCOME) == 0 &&
755 		    ((idle || tp->t_flags & TF_NODELAY) &&
756 		     len + off >= so->so_snd.sb_cc))
757 			goto send;
758 		if (tp->t_force)
759 			goto send;
760 		if (len >= tp->max_sndwnd / 2)
761 			goto send;
762 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))
763 			goto send;
764 	}
765 
766 	/*
767 	 * Compare available window to amount of window known to peer
768 	 * (as advertised window less next expected input).  If the
769 	 * difference is at least twice the size of the largest segment
770 	 * we expect to receive (i.e. two segments) or at least 50% of
771 	 * the maximum possible window, then want to send a window update
772 	 * to peer.
773 	 */
774 	if (win > 0) {
775 		/*
776 		 * "adv" is the amount we can increase the window,
777 		 * taking into account that we are limited by
778 		 * TCP_MAXWIN << tp->rcv_scale.
779 		 */
780 		long adv = min(win, (long)TCP_MAXWIN << tp->rcv_scale) -
781 			(tp->rcv_adv - tp->rcv_nxt);
782 
783 		if (adv >= (long) (2 * rxsegsize))
784 			goto send;
785 		if (2 * adv >= (long) so->so_rcv.sb_hiwat)
786 			goto send;
787 	}
788 
789 	/*
790 	 * Send if we owe peer an ACK.
791 	 */
792 	if (tp->t_flags & TF_ACKNOW)
793 		goto send;
794 	if (flags & (TH_SYN|TH_FIN|TH_RST))
795 		goto send;
796 	if (SEQ_GT(tp->snd_up, tp->snd_una))
797 		goto send;
798 
799 	/*
800 	 * TCP window updates are not reliable, rather a polling protocol
801 	 * using ``persist'' packets is used to insure receipt of window
802 	 * updates.  The three ``states'' for the output side are:
803 	 *	idle			not doing retransmits or persists
804 	 *	persisting		to move a small or zero window
805 	 *	(re)transmitting	and thereby not persisting
806 	 *
807 	 * tp->t_timer[TCPT_PERSIST]
808 	 *	is set when we are in persist state.
809 	 * tp->t_force
810 	 *	is set when we are called to send a persist packet.
811 	 * tp->t_timer[TCPT_REXMT]
812 	 *	is set when we are retransmitting
813 	 * The output side is idle when both timers are zero.
814 	 *
815 	 * If send window is too small, there is data to transmit, and no
816 	 * retransmit or persist is pending, then go to persist state.
817 	 * If nothing happens soon, send when timer expires:
818 	 * if window is nonzero, transmit what we can,
819 	 * otherwise force out a byte.
820 	 */
821 	if (so->so_snd.sb_cc && TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 &&
822 	    TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) {
823 		tp->t_rxtshift = 0;
824 		tcp_setpersist(tp);
825 	}
826 
827 	/*
828 	 * No reason to send a segment, just return.
829 	 */
830 	return (0);
831 
832 send:
833 	/*
834 	 * Before ESTABLISHED, force sending of initial options
835 	 * unless TCP set not to do any options.
836 	 * NOTE: we assume that the IP/TCP header plus TCP options
837 	 * always fit in a single mbuf, leaving room for a maximum
838 	 * link header, i.e.
839 	 *	max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
840 	 */
841 	optlen = 0;
842 	switch (af) {
843 #ifdef INET
844 	case AF_INET:
845 		iphdrlen = sizeof(struct ip) + sizeof(struct tcphdr);
846 		break;
847 #endif
848 #ifdef INET6
849 	case AF_INET6:
850 		iphdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
851 		break;
852 #endif
853 	default:	/*pacify gcc*/
854 		iphdrlen = 0;
855 		break;
856 	}
857 	hdrlen = iphdrlen;
858 	if (flags & TH_SYN) {
859 		struct rtentry *rt;
860 
861 		rt = NULL;
862 #ifdef INET
863 		if (tp->t_inpcb)
864 			rt = in_pcbrtentry(tp->t_inpcb);
865 #endif
866 #ifdef INET6
867 		if (tp->t_in6pcb)
868 			rt = in6_pcbrtentry(tp->t_in6pcb);
869 #endif
870 
871 		tp->snd_nxt = tp->iss;
872 		tp->t_ourmss = tcp_mss_to_advertise(rt != NULL ?
873 						    rt->rt_ifp : NULL, af);
874 		if ((tp->t_flags & TF_NOOPT) == 0) {
875 			opt[0] = TCPOPT_MAXSEG;
876 			opt[1] = 4;
877 			opt[2] = (tp->t_ourmss >> 8) & 0xff;
878 			opt[3] = tp->t_ourmss & 0xff;
879 			optlen = 4;
880 
881 			if ((tp->t_flags & TF_REQ_SCALE) &&
882 			    ((flags & TH_ACK) == 0 ||
883 			    (tp->t_flags & TF_RCVD_SCALE))) {
884 				*((u_int32_t *) (opt + optlen)) = htonl(
885 					TCPOPT_NOP << 24 |
886 					TCPOPT_WINDOW << 16 |
887 					TCPOLEN_WINDOW << 8 |
888 					tp->request_r_scale);
889 				optlen += 4;
890 			}
891 		}
892 	}
893 
894 	/*
895 	 * Send a timestamp and echo-reply if this is a SYN and our side
896 	 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side
897 	 * and our peer have sent timestamps in our SYN's.
898 	 */
899 	if ((tp->t_flags & (TF_REQ_TSTMP|TF_NOOPT)) == TF_REQ_TSTMP &&
900 	     (flags & TH_RST) == 0 &&
901 	    ((flags & (TH_SYN|TH_ACK)) == TH_SYN ||
902 	     (tp->t_flags & TF_RCVD_TSTMP))) {
903 		u_int32_t *lp = (u_int32_t *)(opt + optlen);
904 
905 		/* Form timestamp option as shown in appendix A of RFC 1323. */
906 		*lp++ = htonl(TCPOPT_TSTAMP_HDR);
907 		*lp++ = htonl(TCP_TIMESTAMP(tp));
908 		*lp   = htonl(tp->ts_recent);
909 		optlen += TCPOLEN_TSTAMP_APPA;
910 	}
911 
912 #ifdef TCP_SIGNATURE
913 #if defined(INET6) && defined(FAST_IPSEC)
914 	if (tp->t_family == AF_INET)
915 #endif
916 	if (tp->t_flags & TF_SIGNATURE) {
917 		u_char *bp;
918 		/*
919 		 * Initialize TCP-MD5 option (RFC2385)
920 		 */
921 		bp = (u_char *)opt + optlen;
922 		*bp++ = TCPOPT_SIGNATURE;
923 		*bp++ = TCPOLEN_SIGNATURE;
924 		sigoff = optlen + 2;
925 		bzero(bp, TCP_SIGLEN);
926 		bp += TCP_SIGLEN;
927 		optlen += TCPOLEN_SIGNATURE;
928 		/*
929 		 * Terminate options list and maintain 32-bit alignment.
930  		 */
931 		*bp++ = TCPOPT_NOP;
932 		*bp++ = TCPOPT_EOL;
933  		optlen += 2;
934  	}
935 #endif /* TCP_SIGNATURE */
936 
937 	hdrlen += optlen;
938 
939 #ifdef DIAGNOSTIC
940 	if (len > txsegsize)
941 		panic("tcp data to be sent is larger than segment");
942 	if (max_linkhdr + hdrlen > MCLBYTES)
943 		panic("tcphdr too big");
944 #endif
945 
946 	/*
947 	 * Grab a header mbuf, attaching a copy of data to
948 	 * be transmitted, and initialize the header from
949 	 * the template for sends on this connection.
950 	 */
951 	if (len) {
952 		error = tcp_build_datapkt(tp, so, off, len, hdrlen, &m);
953 		if (error)
954 			goto out;
955 		/*
956 		 * If we're sending everything we've got, set PUSH.
957 		 * (This will keep happy those implementations which only
958 		 * give data to the user when a buffer fills or
959 		 * a PUSH comes in.)
960 		 */
961 		if (off + len == so->so_snd.sb_cc)
962 			flags |= TH_PUSH;
963 	} else {
964 		if (tp->t_flags & TF_ACKNOW)
965 			tcpstat.tcps_sndacks++;
966 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
967 			tcpstat.tcps_sndctrl++;
968 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
969 			tcpstat.tcps_sndurg++;
970 		else
971 			tcpstat.tcps_sndwinup++;
972 
973 		MGETHDR(m, M_DONTWAIT, MT_HEADER);
974 		if (m != NULL && max_linkhdr + hdrlen > MHLEN) {
975 			MCLGET(m, M_DONTWAIT);
976 			if ((m->m_flags & M_EXT) == 0) {
977 				m_freem(m);
978 				m = NULL;
979 			}
980 		}
981 		if (m == NULL) {
982 			error = ENOBUFS;
983 			goto out;
984 		}
985 		MCLAIM(m, &tcp_tx_mowner);
986 		m->m_data += max_linkhdr;
987 		m->m_len = hdrlen;
988 	}
989 	m->m_pkthdr.rcvif = (struct ifnet *)0;
990 	switch (af) {
991 #ifdef INET
992 	case AF_INET:
993 		ip = mtod(m, struct ip *);
994 #ifdef INET6
995 		ip6 = NULL;
996 #endif
997 		th = (struct tcphdr *)(ip + 1);
998 		break;
999 #endif
1000 #ifdef INET6
1001 	case AF_INET6:
1002 		ip = NULL;
1003 		ip6 = mtod(m, struct ip6_hdr *);
1004 		th = (struct tcphdr *)(ip6 + 1);
1005 		break;
1006 #endif
1007 	default:	/*pacify gcc*/
1008 		ip = NULL;
1009 #ifdef INET6
1010 		ip6 = NULL;
1011 #endif
1012 		th = NULL;
1013 		break;
1014 	}
1015 	if (tp->t_template == 0)
1016 		panic("tcp_output");
1017 	if (tp->t_template->m_len < iphdrlen)
1018 		panic("tcp_output");
1019 	bcopy(mtod(tp->t_template, caddr_t), mtod(m, caddr_t), iphdrlen);
1020 
1021 	/*
1022 	 * If we are doing retransmissions, then snd_nxt will
1023 	 * not reflect the first unsent octet.  For ACK only
1024 	 * packets, we do not want the sequence number of the
1025 	 * retransmitted packet, we want the sequence number
1026 	 * of the next unsent octet.  So, if there is no data
1027 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
1028 	 * when filling in ti_seq.  But if we are in persist
1029 	 * state, snd_max might reflect one byte beyond the
1030 	 * right edge of the window, so use snd_nxt in that
1031 	 * case, since we know we aren't doing a retransmission.
1032 	 * (retransmit and persist are mutually exclusive...)
1033 	 */
1034 	if (len || (flags & (TH_SYN|TH_FIN)) ||
1035 	    TCP_TIMER_ISARMED(tp, TCPT_PERSIST))
1036 		th->th_seq = htonl(tp->snd_nxt);
1037 	else
1038 		th->th_seq = htonl(tp->snd_max);
1039 	th->th_ack = htonl(tp->rcv_nxt);
1040 	if (optlen) {
1041 		bcopy((caddr_t)opt, (caddr_t)(th + 1), optlen);
1042 		th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
1043 	}
1044 	th->th_flags = flags;
1045 	/*
1046 	 * Calculate receive window.  Don't shrink window,
1047 	 * but avoid silly window syndrome.
1048 	 */
1049 	if (win < (long)(so->so_rcv.sb_hiwat / 4) && win < (long)rxsegsize)
1050 		win = 0;
1051 	if (win > (long)TCP_MAXWIN << tp->rcv_scale)
1052 		win = (long)TCP_MAXWIN << tp->rcv_scale;
1053 	if (win < (long)(int32_t)(tp->rcv_adv - tp->rcv_nxt))
1054 		win = (long)(int32_t)(tp->rcv_adv - tp->rcv_nxt);
1055 	th->th_win = htons((u_int16_t) (win>>tp->rcv_scale));
1056 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1057 		u_int32_t urp = tp->snd_up - tp->snd_nxt;
1058 		if (urp > IP_MAXPACKET)
1059 			urp = IP_MAXPACKET;
1060 		th->th_urp = htons((u_int16_t)urp);
1061 		th->th_flags |= TH_URG;
1062 	} else
1063 		/*
1064 		 * If no urgent pointer to send, then we pull
1065 		 * the urgent pointer to the left edge of the send window
1066 		 * so that it doesn't drift into the send window on sequence
1067 		 * number wraparound.
1068 		 */
1069 		tp->snd_up = tp->snd_una;		/* drag it along */
1070 
1071 #ifdef TCP_SIGNATURE
1072 #if defined(INET6) && defined(FAST_IPSEC)
1073 	if (tp->t_family == AF_INET) /* XXX */
1074 #endif
1075 	if (sigoff && (tp->t_flags & TF_SIGNATURE)) {
1076 		struct secasvar *sav;
1077 		u_int8_t *sigp;
1078 
1079 		sav = tcp_signature_getsav(m, th);
1080 
1081 		if (sav == NULL) {
1082 			if (m)
1083 				m_freem(m);
1084 			return (EPERM);
1085 		}
1086 
1087 		m->m_pkthdr.len = hdrlen + len;
1088 		sigp = (caddr_t)th + sizeof(*th) + sigoff;
1089 		tcp_signature(m, th, (caddr_t)th - mtod(m, caddr_t), sav, sigp);
1090 
1091 		key_sa_recordxfer(sav, m);
1092 #ifdef FAST_IPSEC
1093 		KEY_FREESAV(&sav);
1094 #else
1095 		key_freesav(sav);
1096 #endif
1097 	}
1098 #endif
1099 
1100 	/*
1101 	 * Set ourselves up to be checksummed just before the packet
1102 	 * hits the wire.
1103 	 */
1104 	switch (af) {
1105 #ifdef INET
1106 	case AF_INET:
1107 		m->m_pkthdr.csum_flags = M_CSUM_TCPv4;
1108 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1109 		if (len + optlen) {
1110 			/* Fixup the pseudo-header checksum. */
1111 			/* XXXJRT Not IP Jumbogram safe. */
1112 			th->th_sum = in_cksum_addword(th->th_sum,
1113 			    htons((u_int16_t) (len + optlen)));
1114 		}
1115 		break;
1116 #endif
1117 #ifdef INET6
1118 	case AF_INET6:
1119 		/*
1120 		 * XXX Actually delaying the checksum is Hard
1121 		 * XXX (well, maybe not for Itojun, but it is
1122 		 * XXX for me), but we can still take advantage
1123 		 * XXX of the cached pseudo-header checksum.
1124 		 */
1125 		/* equals to hdrlen + len */
1126 		m->m_pkthdr.len = sizeof(struct ip6_hdr)
1127 			+ sizeof(struct tcphdr) + optlen + len;
1128 #ifdef notyet
1129 		m->m_pkthdr.csum_flags = M_CSUM_TCPv6;
1130 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1131 #endif
1132 		if (len + optlen) {
1133 			/* Fixup the pseudo-header checksum. */
1134 			/* XXXJRT: Not IPv6 Jumbogram safe. */
1135 			th->th_sum = in_cksum_addword(th->th_sum,
1136 			    htons((u_int16_t) (len + optlen)));
1137 		}
1138 #ifndef notyet
1139 		th->th_sum = in6_cksum(m, 0, sizeof(struct ip6_hdr),
1140 		    sizeof(struct tcphdr) + optlen + len);
1141 #endif
1142 		break;
1143 #endif
1144 	}
1145 
1146 	/*
1147 	 * In transmit state, time the transmission and arrange for
1148 	 * the retransmit.  In persist state, just set snd_max.
1149 	 */
1150 	if (tp->t_force == 0 || TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) {
1151 		tcp_seq startseq = tp->snd_nxt;
1152 
1153 		/*
1154 		 * Advance snd_nxt over sequence space of this segment.
1155 		 * There are no states in which we send both a SYN and a FIN,
1156 		 * so we collapse the tests for these flags.
1157 		 */
1158 		if (flags & (TH_SYN|TH_FIN))
1159 			tp->snd_nxt++;
1160 		tp->snd_nxt += len;
1161 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1162 			tp->snd_max = tp->snd_nxt;
1163 			/*
1164 			 * Time this transmission if not a retransmission and
1165 			 * not currently timing anything.
1166 			 */
1167 			if (tp->t_rtttime == 0) {
1168 				tp->t_rtttime = tcp_now;
1169 				tp->t_rtseq = startseq;
1170 				tcpstat.tcps_segstimed++;
1171 			}
1172 		}
1173 
1174 		/*
1175 		 * Set retransmit timer if not currently set,
1176 		 * and not doing an ack or a keep-alive probe.
1177 		 * Initial value for retransmit timer is smoothed
1178 		 * round-trip time + 2 * round-trip time variance.
1179 		 * Initialize shift counter which is used for backoff
1180 		 * of retransmit time.
1181 		 */
1182 		if (TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 &&
1183 		    tp->snd_nxt != tp->snd_una) {
1184 			if (TCP_TIMER_ISARMED(tp, TCPT_PERSIST)) {
1185 				TCP_TIMER_DISARM(tp, TCPT_PERSIST);
1186 				tp->t_rxtshift = 0;
1187 			}
1188 			TCP_TIMER_ARM(tp, TCPT_REXMT, tp->t_rxtcur);
1189 		}
1190 	} else
1191 		if (SEQ_GT(tp->snd_nxt + len, tp->snd_max))
1192 			tp->snd_max = tp->snd_nxt + len;
1193 
1194 #ifdef TCP_DEBUG
1195 	/*
1196 	 * Trace.
1197 	 */
1198 	if (so->so_options & SO_DEBUG)
1199 		tcp_trace(TA_OUTPUT, tp->t_state, tp, m, 0);
1200 #endif
1201 
1202 	/*
1203 	 * Fill in IP length and desired time to live and
1204 	 * send to IP level.  There should be a better way
1205 	 * to handle ttl and tos; we could keep them in
1206 	 * the template, but need a way to checksum without them.
1207 	 */
1208 	m->m_pkthdr.len = hdrlen + len;
1209 
1210 	switch (af) {
1211 #ifdef INET
1212 	case AF_INET:
1213 		ip->ip_len = htons(m->m_pkthdr.len);
1214 		if (tp->t_inpcb) {
1215 			ip->ip_ttl = tp->t_inpcb->inp_ip.ip_ttl;
1216 			ip->ip_tos = tp->t_inpcb->inp_ip.ip_tos;
1217 		}
1218 #ifdef INET6
1219 		else if (tp->t_in6pcb) {
1220 			ip->ip_ttl = in6_selecthlim(tp->t_in6pcb, NULL); /*XXX*/
1221 			ip->ip_tos = 0;	/*XXX*/
1222 		}
1223 #endif
1224 		break;
1225 #endif
1226 #ifdef INET6
1227 	case AF_INET6:
1228 		ip6->ip6_nxt = IPPROTO_TCP;
1229 		if (tp->t_in6pcb) {
1230 			/*
1231 			 * we separately set hoplimit for every segment, since
1232 			 * the user might want to change the value via
1233 			 * setsockopt. Also, desired default hop limit might
1234 			 * be changed via Neighbor Discovery.
1235 			 */
1236 			ip6->ip6_hlim = in6_selecthlim(tp->t_in6pcb,
1237 				ro->ro_rt ? ro->ro_rt->rt_ifp : NULL);
1238 		}
1239 		/* ip6->ip6_flow = ??? */
1240 		/* ip6_plen will be filled in ip6_output(). */
1241 		break;
1242 #endif
1243 	}
1244 
1245 	switch (af) {
1246 #ifdef INET
1247 	case AF_INET:
1248 	    {
1249 		struct mbuf *opts;
1250 
1251 		if (tp->t_inpcb)
1252 			opts = tp->t_inpcb->inp_options;
1253 		else
1254 			opts = NULL;
1255 		error = ip_output(m, opts, ro,
1256 			(tp->t_mtudisc ? IP_MTUDISC : 0) |
1257 			(so->so_options & SO_DONTROUTE),
1258 			(struct ip_moptions *)0, so);
1259 		break;
1260 	    }
1261 #endif
1262 #ifdef INET6
1263 	case AF_INET6:
1264 	    {
1265 		struct ip6_pktopts *opts;
1266 
1267 		if (tp->t_in6pcb)
1268 			opts = tp->t_in6pcb->in6p_outputopts;
1269 		else
1270 			opts = NULL;
1271 		error = ip6_output(m, opts, (struct route_in6 *)ro,
1272 			so->so_options & SO_DONTROUTE,
1273 			(struct ip6_moptions *)0, so, NULL);
1274 		break;
1275 	    }
1276 #endif
1277 	default:
1278 		error = EAFNOSUPPORT;
1279 		break;
1280 	}
1281 	if (error) {
1282 out:
1283 		if (error == ENOBUFS) {
1284 			tcpstat.tcps_selfquench++;
1285 #ifdef INET
1286 			if (tp->t_inpcb)
1287 				tcp_quench(tp->t_inpcb, 0);
1288 #endif
1289 #ifdef INET6
1290 			if (tp->t_in6pcb)
1291 				tcp6_quench(tp->t_in6pcb, 0);
1292 #endif
1293 			error = 0;
1294 		} else if ((error == EHOSTUNREACH || error == ENETDOWN) &&
1295 		    TCPS_HAVERCVDSYN(tp->t_state)) {
1296 			tp->t_softerror = error;
1297 			error = 0;
1298 		}
1299 
1300 		/* Restart the delayed ACK timer, if necessary. */
1301 		if (tp->t_flags & TF_DELACK)
1302 			TCP_RESTART_DELACK(tp);
1303 
1304 		return (error);
1305 	}
1306 	tcpstat.tcps_sndtotal++;
1307 	if (tp->t_flags & TF_DELACK)
1308 		tcpstat.tcps_delack++;
1309 
1310 	/*
1311 	 * Data sent (as far as we can tell).
1312 	 * If this advertises a larger window than any other segment,
1313 	 * then remember the size of the advertised window.
1314 	 * Any pending ACK has now been sent.
1315 	 */
1316 	if (win > 0 && SEQ_GT(tp->rcv_nxt+win, tp->rcv_adv))
1317 		tp->rcv_adv = tp->rcv_nxt + win;
1318 	tp->last_ack_sent = tp->rcv_nxt;
1319 	tp->t_flags &= ~TF_ACKNOW;
1320 	TCP_CLEAR_DELACK(tp);
1321 #ifdef DIAGNOSTIC
1322 	if (maxburst < 0)
1323 		printf("tcp_output: maxburst exceeded by %d\n", -maxburst);
1324 #endif
1325 	if (sendalot && (!tcp_do_newreno || --maxburst))
1326 		goto again;
1327 	return (0);
1328 }
1329 
1330 void
1331 tcp_setpersist(tp)
1332 	struct tcpcb *tp;
1333 {
1334 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> (1 + 2);
1335 	int nticks;
1336 
1337 	if (TCP_TIMER_ISARMED(tp, TCPT_REXMT))
1338 		panic("tcp_output REXMT");
1339 	/*
1340 	 * Start/restart persistance timer.
1341 	 */
1342 	if (t < tp->t_rttmin)
1343 		t = tp->t_rttmin;
1344 	TCPT_RANGESET(nticks, t * tcp_backoff[tp->t_rxtshift],
1345 	    TCPTV_PERSMIN, TCPTV_PERSMAX);
1346 	TCP_TIMER_ARM(tp, TCPT_PERSIST, nticks);
1347 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1348 		tp->t_rxtshift++;
1349 }
1350