xref: /netbsd-src/sys/netinet/tcp_output.c (revision aa73cae19608873cc4d1f712c4a0f8f8435f1ffa)
1 /*	$NetBSD: tcp_output.c,v 1.117 2005/02/26 22:45:12 perry 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 1.117 2005/02/26 22:45:12 perry 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(struct tcpcb *tp)
540 {
541 	struct socket *so;
542 	struct route *ro;
543 	long len, win;
544 	int off, flags, error;
545 	struct mbuf *m;
546 	struct ip *ip;
547 #ifdef INET6
548 	struct ip6_hdr *ip6;
549 #endif
550 	struct tcphdr *th;
551 	u_char opt[MAX_TCPOPTLEN];
552 	unsigned optlen, hdrlen;
553 	int idle, sendalot, txsegsize, rxsegsize;
554 	int maxburst = TCP_MAXBURST;
555 	int af;		/* address family on the wire */
556 	int iphdrlen;
557 #ifdef TCP_SIGNATURE
558 	int sigoff = 0;
559 #endif
560 
561 #ifdef DIAGNOSTIC
562 	if (tp->t_inpcb && tp->t_in6pcb)
563 		panic("tcp_output: both t_inpcb and t_in6pcb are set");
564 #endif
565 	so = NULL;
566 	ro = NULL;
567 	if (tp->t_inpcb) {
568 		so = tp->t_inpcb->inp_socket;
569 		ro = &tp->t_inpcb->inp_route;
570 	}
571 #ifdef INET6
572 	else if (tp->t_in6pcb) {
573 		so = tp->t_in6pcb->in6p_socket;
574 		ro = (struct route *)&tp->t_in6pcb->in6p_route;
575 	}
576 #endif
577 
578 	switch (af = tp->t_family) {
579 #ifdef INET
580 	case AF_INET:
581 		if (tp->t_inpcb)
582 			break;
583 #ifdef INET6
584 		/* mapped addr case */
585 		if (tp->t_in6pcb)
586 			break;
587 #endif
588 		return (EINVAL);
589 #endif
590 #ifdef INET6
591 	case AF_INET6:
592 		if (tp->t_in6pcb)
593 			break;
594 		return (EINVAL);
595 #endif
596 	default:
597 		return (EAFNOSUPPORT);
598 	}
599 
600 	if (tcp_segsize(tp, &txsegsize, &rxsegsize))
601 		return (EMSGSIZE);
602 
603 	idle = (tp->snd_max == tp->snd_una);
604 
605 	/*
606 	 * Restart Window computation.  From draft-floyd-incr-init-win-03:
607 	 *
608 	 *	Optionally, a TCP MAY set the restart window to the
609 	 *	minimum of the value used for the initial window and
610 	 *	the current value of cwnd (in other words, using a
611 	 *	larger value for the restart window should never increase
612 	 *	the size of cwnd).
613 	 */
614 	if (tcp_cwm) {
615 		/*
616 		 * Hughes/Touch/Heidemann Congestion Window Monitoring.
617 		 * Count the number of packets currently pending
618 		 * acknowledgement, and limit our congestion window
619 		 * to a pre-determined allowed burst size plus that count.
620 		 * This prevents bursting once all pending packets have
621 		 * been acknowledged (i.e. transmission is idle).
622 		 *
623 		 * XXX Link this to Initial Window?
624 		 */
625 		tp->snd_cwnd = min(tp->snd_cwnd,
626 		    (tcp_cwm_burstsize * txsegsize) +
627 		    (tp->snd_nxt - tp->snd_una));
628 	} else {
629 		if (idle && (tcp_now - tp->t_rcvtime) >= tp->t_rxtcur) {
630 			/*
631 			 * We have been idle for "a while" and no acks are
632 			 * expected to clock out any data we send --
633 			 * slow start to get ack "clock" running again.
634 			 */
635 			int ss = tcp_init_win;
636 #ifdef INET
637 			if (tp->t_inpcb &&
638 			    in_localaddr(tp->t_inpcb->inp_faddr))
639 				ss = tcp_init_win_local;
640 #endif
641 #ifdef INET6
642 			if (tp->t_in6pcb &&
643 			    in6_localaddr(&tp->t_in6pcb->in6p_faddr))
644 				ss = tcp_init_win_local;
645 #endif
646 			tp->snd_cwnd = min(tp->snd_cwnd,
647 			    TCP_INITIAL_WINDOW(ss, txsegsize));
648 		}
649 	}
650 
651 again:
652 	/*
653 	 * Determine length of data that should be transmitted, and
654 	 * flags that should be used.  If there is some data or critical
655 	 * controls (SYN, RST) to send, then transmit; otherwise,
656 	 * investigate further.
657 	 */
658 	sendalot = 0;
659 	off = tp->snd_nxt - tp->snd_una;
660 	win = min(tp->snd_wnd, tp->snd_cwnd);
661 
662 	flags = tcp_outflags[tp->t_state];
663 	/*
664 	 * If in persist timeout with window of 0, send 1 byte.
665 	 * Otherwise, if window is small but nonzero
666 	 * and timer expired, we will send what we can
667 	 * and go to transmit state.
668 	 */
669 	if (tp->t_force) {
670 		if (win == 0) {
671 			/*
672 			 * If we still have some data to send, then
673 			 * clear the FIN bit.  Usually this would
674 			 * happen below when it realizes that we
675 			 * aren't sending all the data.  However,
676 			 * if we have exactly 1 byte of unset data,
677 			 * then it won't clear the FIN bit below,
678 			 * and if we are in persist state, we wind
679 			 * up sending the packet without recording
680 			 * that we sent the FIN bit.
681 			 *
682 			 * We can't just blindly clear the FIN bit,
683 			 * because if we don't have any more data
684 			 * to send then the probe will be the FIN
685 			 * itself.
686 			 */
687 			if (off < so->so_snd.sb_cc)
688 				flags &= ~TH_FIN;
689 			win = 1;
690 		} else {
691 			TCP_TIMER_DISARM(tp, TCPT_PERSIST);
692 			tp->t_rxtshift = 0;
693 		}
694 	}
695 
696 	if (win < so->so_snd.sb_cc) {
697 		len = win - off;
698 		flags &= ~TH_FIN;
699 	} else
700 		len = so->so_snd.sb_cc - off;
701 
702 	if (len < 0) {
703 		/*
704 		 * If FIN has been sent but not acked,
705 		 * but we haven't been called to retransmit,
706 		 * len will be -1.  Otherwise, window shrank
707 		 * after we sent into it.  If window shrank to 0,
708 		 * cancel pending retransmit, pull snd_nxt back
709 		 * to (closed) window, and set the persist timer
710 		 * if it isn't already going.  If the window didn't
711 		 * close completely, just wait for an ACK.
712 		 *
713 		 * If we have a pending FIN, either it has already been
714 		 * transmitted or it is outside the window, so drop it.
715 		 * If the FIN has been transmitted, but this is not a
716 		 * retransmission, then len must be -1.  Therefore we also
717 		 * prevent here the sending of `gratuitous FINs'.  This
718 		 * eliminates the need to check for that case below (e.g.
719 		 * to back up snd_nxt before the FIN so that the sequence
720 		 * number is correct).
721 		 */
722 		len = 0;
723 		flags &= ~TH_FIN;
724 		if (win == 0) {
725 			TCP_TIMER_DISARM(tp, TCPT_REXMT);
726 			tp->t_rxtshift = 0;
727 			tp->snd_nxt = tp->snd_una;
728 			if (TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0)
729 				tcp_setpersist(tp);
730 		}
731 	}
732 	if (len > txsegsize) {
733 		len = txsegsize;
734 		flags &= ~TH_FIN;
735 		sendalot = 1;
736 	}
737 
738 	win = sbspace(&so->so_rcv);
739 
740 	/*
741 	 * Sender silly window avoidance.  If connection is idle
742 	 * and can send all data, a maximum segment,
743 	 * at least a maximum default-size segment do it,
744 	 * or are forced, do it; otherwise don't bother.
745 	 * If peer's buffer is tiny, then send
746 	 * when window is at least half open.
747 	 * If retransmitting (possibly after persist timer forced us
748 	 * to send into a small window), then must resend.
749 	 */
750 	if (len) {
751 		if (len == txsegsize)
752 			goto send;
753 		if ((so->so_state & SS_MORETOCOME) == 0 &&
754 		    ((idle || tp->t_flags & TF_NODELAY) &&
755 		     len + off >= so->so_snd.sb_cc))
756 			goto send;
757 		if (tp->t_force)
758 			goto send;
759 		if (len >= tp->max_sndwnd / 2)
760 			goto send;
761 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))
762 			goto send;
763 	}
764 
765 	/*
766 	 * Compare available window to amount of window known to peer
767 	 * (as advertised window less next expected input).  If the
768 	 * difference is at least twice the size of the largest segment
769 	 * we expect to receive (i.e. two segments) or at least 50% of
770 	 * the maximum possible window, then want to send a window update
771 	 * to peer.
772 	 */
773 	if (win > 0) {
774 		/*
775 		 * "adv" is the amount we can increase the window,
776 		 * taking into account that we are limited by
777 		 * TCP_MAXWIN << tp->rcv_scale.
778 		 */
779 		long adv = min(win, (long)TCP_MAXWIN << tp->rcv_scale) -
780 			(tp->rcv_adv - tp->rcv_nxt);
781 
782 		if (adv >= (long) (2 * rxsegsize))
783 			goto send;
784 		if (2 * adv >= (long) so->so_rcv.sb_hiwat)
785 			goto send;
786 	}
787 
788 	/*
789 	 * Send if we owe peer an ACK.
790 	 */
791 	if (tp->t_flags & TF_ACKNOW)
792 		goto send;
793 	if (flags & (TH_SYN|TH_FIN|TH_RST))
794 		goto send;
795 	if (SEQ_GT(tp->snd_up, tp->snd_una))
796 		goto send;
797 
798 	/*
799 	 * TCP window updates are not reliable, rather a polling protocol
800 	 * using ``persist'' packets is used to insure receipt of window
801 	 * updates.  The three ``states'' for the output side are:
802 	 *	idle			not doing retransmits or persists
803 	 *	persisting		to move a small or zero window
804 	 *	(re)transmitting	and thereby not persisting
805 	 *
806 	 * tp->t_timer[TCPT_PERSIST]
807 	 *	is set when we are in persist state.
808 	 * tp->t_force
809 	 *	is set when we are called to send a persist packet.
810 	 * tp->t_timer[TCPT_REXMT]
811 	 *	is set when we are retransmitting
812 	 * The output side is idle when both timers are zero.
813 	 *
814 	 * If send window is too small, there is data to transmit, and no
815 	 * retransmit or persist is pending, then go to persist state.
816 	 * If nothing happens soon, send when timer expires:
817 	 * if window is nonzero, transmit what we can,
818 	 * otherwise force out a byte.
819 	 */
820 	if (so->so_snd.sb_cc && TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 &&
821 	    TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) {
822 		tp->t_rxtshift = 0;
823 		tcp_setpersist(tp);
824 	}
825 
826 	/*
827 	 * No reason to send a segment, just return.
828 	 */
829 	return (0);
830 
831 send:
832 	/*
833 	 * Before ESTABLISHED, force sending of initial options
834 	 * unless TCP set not to do any options.
835 	 * NOTE: we assume that the IP/TCP header plus TCP options
836 	 * always fit in a single mbuf, leaving room for a maximum
837 	 * link header, i.e.
838 	 *	max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
839 	 */
840 	optlen = 0;
841 	switch (af) {
842 #ifdef INET
843 	case AF_INET:
844 		iphdrlen = sizeof(struct ip) + sizeof(struct tcphdr);
845 		break;
846 #endif
847 #ifdef INET6
848 	case AF_INET6:
849 		iphdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
850 		break;
851 #endif
852 	default:	/*pacify gcc*/
853 		iphdrlen = 0;
854 		break;
855 	}
856 	hdrlen = iphdrlen;
857 	if (flags & TH_SYN) {
858 		struct rtentry *rt;
859 
860 		rt = NULL;
861 #ifdef INET
862 		if (tp->t_inpcb)
863 			rt = in_pcbrtentry(tp->t_inpcb);
864 #endif
865 #ifdef INET6
866 		if (tp->t_in6pcb)
867 			rt = in6_pcbrtentry(tp->t_in6pcb);
868 #endif
869 
870 		tp->snd_nxt = tp->iss;
871 		tp->t_ourmss = tcp_mss_to_advertise(rt != NULL ?
872 						    rt->rt_ifp : NULL, af);
873 		if ((tp->t_flags & TF_NOOPT) == 0) {
874 			opt[0] = TCPOPT_MAXSEG;
875 			opt[1] = 4;
876 			opt[2] = (tp->t_ourmss >> 8) & 0xff;
877 			opt[3] = tp->t_ourmss & 0xff;
878 			optlen = 4;
879 
880 			if ((tp->t_flags & TF_REQ_SCALE) &&
881 			    ((flags & TH_ACK) == 0 ||
882 			    (tp->t_flags & TF_RCVD_SCALE))) {
883 				*((u_int32_t *) (opt + optlen)) = htonl(
884 					TCPOPT_NOP << 24 |
885 					TCPOPT_WINDOW << 16 |
886 					TCPOLEN_WINDOW << 8 |
887 					tp->request_r_scale);
888 				optlen += 4;
889 			}
890 		}
891 	}
892 
893 	/*
894 	 * Send a timestamp and echo-reply if this is a SYN and our side
895 	 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side
896 	 * and our peer have sent timestamps in our SYN's.
897 	 */
898 	if ((tp->t_flags & (TF_REQ_TSTMP|TF_NOOPT)) == TF_REQ_TSTMP &&
899 	     (flags & TH_RST) == 0 &&
900 	    ((flags & (TH_SYN|TH_ACK)) == TH_SYN ||
901 	     (tp->t_flags & TF_RCVD_TSTMP))) {
902 		u_int32_t *lp = (u_int32_t *)(opt + optlen);
903 
904 		/* Form timestamp option as shown in appendix A of RFC 1323. */
905 		*lp++ = htonl(TCPOPT_TSTAMP_HDR);
906 		*lp++ = htonl(TCP_TIMESTAMP(tp));
907 		*lp   = htonl(tp->ts_recent);
908 		optlen += TCPOLEN_TSTAMP_APPA;
909 	}
910 
911 #ifdef TCP_SIGNATURE
912 #if defined(INET6) && defined(FAST_IPSEC)
913 	if (tp->t_family == AF_INET)
914 #endif
915 	if (tp->t_flags & TF_SIGNATURE) {
916 		u_char *bp;
917 		/*
918 		 * Initialize TCP-MD5 option (RFC2385)
919 		 */
920 		bp = (u_char *)opt + optlen;
921 		*bp++ = TCPOPT_SIGNATURE;
922 		*bp++ = TCPOLEN_SIGNATURE;
923 		sigoff = optlen + 2;
924 		bzero(bp, TCP_SIGLEN);
925 		bp += TCP_SIGLEN;
926 		optlen += TCPOLEN_SIGNATURE;
927 		/*
928 		 * Terminate options list and maintain 32-bit alignment.
929  		 */
930 		*bp++ = TCPOPT_NOP;
931 		*bp++ = TCPOPT_EOL;
932  		optlen += 2;
933  	}
934 #endif /* TCP_SIGNATURE */
935 
936 	hdrlen += optlen;
937 
938 #ifdef DIAGNOSTIC
939 	if (len > txsegsize)
940 		panic("tcp data to be sent is larger than segment");
941 	if (max_linkhdr + hdrlen > MCLBYTES)
942 		panic("tcphdr too big");
943 #endif
944 
945 	/*
946 	 * Grab a header mbuf, attaching a copy of data to
947 	 * be transmitted, and initialize the header from
948 	 * the template for sends on this connection.
949 	 */
950 	if (len) {
951 		error = tcp_build_datapkt(tp, so, off, len, hdrlen, &m);
952 		if (error)
953 			goto out;
954 		/*
955 		 * If we're sending everything we've got, set PUSH.
956 		 * (This will keep happy those implementations which only
957 		 * give data to the user when a buffer fills or
958 		 * a PUSH comes in.)
959 		 */
960 		if (off + len == so->so_snd.sb_cc)
961 			flags |= TH_PUSH;
962 	} else {
963 		if (tp->t_flags & TF_ACKNOW)
964 			tcpstat.tcps_sndacks++;
965 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
966 			tcpstat.tcps_sndctrl++;
967 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
968 			tcpstat.tcps_sndurg++;
969 		else
970 			tcpstat.tcps_sndwinup++;
971 
972 		MGETHDR(m, M_DONTWAIT, MT_HEADER);
973 		if (m != NULL && max_linkhdr + hdrlen > MHLEN) {
974 			MCLGET(m, M_DONTWAIT);
975 			if ((m->m_flags & M_EXT) == 0) {
976 				m_freem(m);
977 				m = NULL;
978 			}
979 		}
980 		if (m == NULL) {
981 			error = ENOBUFS;
982 			goto out;
983 		}
984 		MCLAIM(m, &tcp_tx_mowner);
985 		m->m_data += max_linkhdr;
986 		m->m_len = hdrlen;
987 	}
988 	m->m_pkthdr.rcvif = (struct ifnet *)0;
989 	switch (af) {
990 #ifdef INET
991 	case AF_INET:
992 		ip = mtod(m, struct ip *);
993 #ifdef INET6
994 		ip6 = NULL;
995 #endif
996 		th = (struct tcphdr *)(ip + 1);
997 		break;
998 #endif
999 #ifdef INET6
1000 	case AF_INET6:
1001 		ip = NULL;
1002 		ip6 = mtod(m, struct ip6_hdr *);
1003 		th = (struct tcphdr *)(ip6 + 1);
1004 		break;
1005 #endif
1006 	default:	/*pacify gcc*/
1007 		ip = NULL;
1008 #ifdef INET6
1009 		ip6 = NULL;
1010 #endif
1011 		th = NULL;
1012 		break;
1013 	}
1014 	if (tp->t_template == 0)
1015 		panic("tcp_output");
1016 	if (tp->t_template->m_len < iphdrlen)
1017 		panic("tcp_output");
1018 	bcopy(mtod(tp->t_template, caddr_t), mtod(m, caddr_t), iphdrlen);
1019 
1020 	/*
1021 	 * If we are doing retransmissions, then snd_nxt will
1022 	 * not reflect the first unsent octet.  For ACK only
1023 	 * packets, we do not want the sequence number of the
1024 	 * retransmitted packet, we want the sequence number
1025 	 * of the next unsent octet.  So, if there is no data
1026 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
1027 	 * when filling in ti_seq.  But if we are in persist
1028 	 * state, snd_max might reflect one byte beyond the
1029 	 * right edge of the window, so use snd_nxt in that
1030 	 * case, since we know we aren't doing a retransmission.
1031 	 * (retransmit and persist are mutually exclusive...)
1032 	 */
1033 	if (len || (flags & (TH_SYN|TH_FIN)) ||
1034 	    TCP_TIMER_ISARMED(tp, TCPT_PERSIST))
1035 		th->th_seq = htonl(tp->snd_nxt);
1036 	else
1037 		th->th_seq = htonl(tp->snd_max);
1038 	th->th_ack = htonl(tp->rcv_nxt);
1039 	if (optlen) {
1040 		bcopy((caddr_t)opt, (caddr_t)(th + 1), optlen);
1041 		th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
1042 	}
1043 	th->th_flags = flags;
1044 	/*
1045 	 * Calculate receive window.  Don't shrink window,
1046 	 * but avoid silly window syndrome.
1047 	 */
1048 	if (win < (long)(so->so_rcv.sb_hiwat / 4) && win < (long)rxsegsize)
1049 		win = 0;
1050 	if (win > (long)TCP_MAXWIN << tp->rcv_scale)
1051 		win = (long)TCP_MAXWIN << tp->rcv_scale;
1052 	if (win < (long)(int32_t)(tp->rcv_adv - tp->rcv_nxt))
1053 		win = (long)(int32_t)(tp->rcv_adv - tp->rcv_nxt);
1054 	th->th_win = htons((u_int16_t) (win>>tp->rcv_scale));
1055 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1056 		u_int32_t urp = tp->snd_up - tp->snd_nxt;
1057 		if (urp > IP_MAXPACKET)
1058 			urp = IP_MAXPACKET;
1059 		th->th_urp = htons((u_int16_t)urp);
1060 		th->th_flags |= TH_URG;
1061 	} else
1062 		/*
1063 		 * If no urgent pointer to send, then we pull
1064 		 * the urgent pointer to the left edge of the send window
1065 		 * so that it doesn't drift into the send window on sequence
1066 		 * number wraparound.
1067 		 */
1068 		tp->snd_up = tp->snd_una;		/* drag it along */
1069 
1070 #ifdef TCP_SIGNATURE
1071 #if defined(INET6) && defined(FAST_IPSEC)
1072 	if (tp->t_family == AF_INET) /* XXX */
1073 #endif
1074 	if (sigoff && (tp->t_flags & TF_SIGNATURE)) {
1075 		struct secasvar *sav;
1076 		u_int8_t *sigp;
1077 
1078 		sav = tcp_signature_getsav(m, th);
1079 
1080 		if (sav == NULL) {
1081 			if (m)
1082 				m_freem(m);
1083 			return (EPERM);
1084 		}
1085 
1086 		m->m_pkthdr.len = hdrlen + len;
1087 		sigp = (caddr_t)th + sizeof(*th) + sigoff;
1088 		tcp_signature(m, th, (caddr_t)th - mtod(m, caddr_t), sav, sigp);
1089 
1090 		key_sa_recordxfer(sav, m);
1091 #ifdef FAST_IPSEC
1092 		KEY_FREESAV(&sav);
1093 #else
1094 		key_freesav(sav);
1095 #endif
1096 	}
1097 #endif
1098 
1099 	/*
1100 	 * Set ourselves up to be checksummed just before the packet
1101 	 * hits the wire.  Maybe skip checksums on loopback interfaces.
1102 	 */
1103 	switch (af) {
1104 #ifdef INET
1105 	case AF_INET:
1106 		if (__predict_true(ro->ro_rt == NULL ||
1107 				   !(ro->ro_rt->rt_ifp->if_flags &
1108 				     IFF_LOOPBACK) ||
1109 				   tcp_do_loopback_cksum))
1110 			m->m_pkthdr.csum_flags = M_CSUM_TCPv4;
1111 		else
1112 			m->m_pkthdr.csum_flags = 0;
1113 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1114 		if (len + optlen) {
1115 			/* Fixup the pseudo-header checksum. */
1116 			/* XXXJRT Not IP Jumbogram safe. */
1117 			th->th_sum = in_cksum_addword(th->th_sum,
1118 			    htons((u_int16_t) (len + optlen)));
1119 		}
1120 		break;
1121 #endif
1122 #ifdef INET6
1123 	case AF_INET6:
1124 		/*
1125 		 * XXX Actually delaying the checksum is Hard
1126 		 * XXX (well, maybe not for Itojun, but it is
1127 		 * XXX for me), but we can still take advantage
1128 		 * XXX of the cached pseudo-header checksum.
1129 		 */
1130 		/* equals to hdrlen + len */
1131 		m->m_pkthdr.len = sizeof(struct ip6_hdr)
1132 			+ sizeof(struct tcphdr) + optlen + len;
1133 #ifdef notyet
1134 		if (__predict_true(ro->ro_rt == NULL ||
1135 				   !(ro->ro_rt->rt_ifp->if_flags &
1136 				     IFF_LOOPBACK) ||
1137 				   tcp_do_loopback_cksum))
1138 			m->m_pkthdr.csum_flags = M_CSUM_TCPv6;
1139 		else
1140 			m->m_pkthdr.csum_flags = 0;
1141 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1142 #endif
1143 		if (len + optlen) {
1144 			/* Fixup the pseudo-header checksum. */
1145 			/* XXXJRT: Not IPv6 Jumbogram safe. */
1146 			th->th_sum = in_cksum_addword(th->th_sum,
1147 			    htons((u_int16_t) (len + optlen)));
1148 		}
1149 #ifndef notyet
1150 		if (__predict_true(ro->ro_rt == NULL ||
1151 				   !(ro->ro_rt->rt_ifp->if_flags &
1152 				     IFF_LOOPBACK) ||
1153 				   tcp_do_loopback_cksum))
1154 			th->th_sum = in6_cksum(m, 0, sizeof(struct ip6_hdr),
1155 			    sizeof(struct tcphdr) + optlen + len);
1156 #endif
1157 		break;
1158 #endif
1159 	}
1160 
1161 	/*
1162 	 * In transmit state, time the transmission and arrange for
1163 	 * the retransmit.  In persist state, just set snd_max.
1164 	 */
1165 	if (tp->t_force == 0 || TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) {
1166 		tcp_seq startseq = tp->snd_nxt;
1167 
1168 		/*
1169 		 * Advance snd_nxt over sequence space of this segment.
1170 		 * There are no states in which we send both a SYN and a FIN,
1171 		 * so we collapse the tests for these flags.
1172 		 */
1173 		if (flags & (TH_SYN|TH_FIN))
1174 			tp->snd_nxt++;
1175 		tp->snd_nxt += len;
1176 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1177 			tp->snd_max = tp->snd_nxt;
1178 			/*
1179 			 * Time this transmission if not a retransmission and
1180 			 * not currently timing anything.
1181 			 */
1182 			if (tp->t_rtttime == 0) {
1183 				tp->t_rtttime = tcp_now;
1184 				tp->t_rtseq = startseq;
1185 				tcpstat.tcps_segstimed++;
1186 			}
1187 		}
1188 
1189 		/*
1190 		 * Set retransmit timer if not currently set,
1191 		 * and not doing an ack or a keep-alive probe.
1192 		 * Initial value for retransmit timer is smoothed
1193 		 * round-trip time + 2 * round-trip time variance.
1194 		 * Initialize shift counter which is used for backoff
1195 		 * of retransmit time.
1196 		 */
1197 		if (TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 &&
1198 		    tp->snd_nxt != tp->snd_una) {
1199 			if (TCP_TIMER_ISARMED(tp, TCPT_PERSIST)) {
1200 				TCP_TIMER_DISARM(tp, TCPT_PERSIST);
1201 				tp->t_rxtshift = 0;
1202 			}
1203 			TCP_TIMER_ARM(tp, TCPT_REXMT, tp->t_rxtcur);
1204 		}
1205 	} else
1206 		if (SEQ_GT(tp->snd_nxt + len, tp->snd_max))
1207 			tp->snd_max = tp->snd_nxt + len;
1208 
1209 #ifdef TCP_DEBUG
1210 	/*
1211 	 * Trace.
1212 	 */
1213 	if (so->so_options & SO_DEBUG)
1214 		tcp_trace(TA_OUTPUT, tp->t_state, tp, m, 0);
1215 #endif
1216 
1217 	/*
1218 	 * Fill in IP length and desired time to live and
1219 	 * send to IP level.  There should be a better way
1220 	 * to handle ttl and tos; we could keep them in
1221 	 * the template, but need a way to checksum without them.
1222 	 */
1223 	m->m_pkthdr.len = hdrlen + len;
1224 
1225 	switch (af) {
1226 #ifdef INET
1227 	case AF_INET:
1228 		ip->ip_len = htons(m->m_pkthdr.len);
1229 		if (tp->t_inpcb) {
1230 			ip->ip_ttl = tp->t_inpcb->inp_ip.ip_ttl;
1231 			ip->ip_tos = tp->t_inpcb->inp_ip.ip_tos;
1232 		}
1233 #ifdef INET6
1234 		else if (tp->t_in6pcb) {
1235 			ip->ip_ttl = in6_selecthlim(tp->t_in6pcb, NULL); /*XXX*/
1236 			ip->ip_tos = 0;	/*XXX*/
1237 		}
1238 #endif
1239 		break;
1240 #endif
1241 #ifdef INET6
1242 	case AF_INET6:
1243 		ip6->ip6_nxt = IPPROTO_TCP;
1244 		if (tp->t_in6pcb) {
1245 			/*
1246 			 * we separately set hoplimit for every segment, since
1247 			 * the user might want to change the value via
1248 			 * setsockopt. Also, desired default hop limit might
1249 			 * be changed via Neighbor Discovery.
1250 			 */
1251 			ip6->ip6_hlim = in6_selecthlim(tp->t_in6pcb,
1252 				ro->ro_rt ? ro->ro_rt->rt_ifp : NULL);
1253 		}
1254 		/* ip6->ip6_flow = ??? */
1255 		/* ip6_plen will be filled in ip6_output(). */
1256 		break;
1257 #endif
1258 	}
1259 
1260 	switch (af) {
1261 #ifdef INET
1262 	case AF_INET:
1263 	    {
1264 		struct mbuf *opts;
1265 
1266 		if (tp->t_inpcb)
1267 			opts = tp->t_inpcb->inp_options;
1268 		else
1269 			opts = NULL;
1270 		error = ip_output(m, opts, ro,
1271 			(tp->t_mtudisc ? IP_MTUDISC : 0) |
1272 			(so->so_options & SO_DONTROUTE),
1273 			(struct ip_moptions *)0, so);
1274 		break;
1275 	    }
1276 #endif
1277 #ifdef INET6
1278 	case AF_INET6:
1279 	    {
1280 		struct ip6_pktopts *opts;
1281 
1282 		if (tp->t_in6pcb)
1283 			opts = tp->t_in6pcb->in6p_outputopts;
1284 		else
1285 			opts = NULL;
1286 		error = ip6_output(m, opts, (struct route_in6 *)ro,
1287 			so->so_options & SO_DONTROUTE,
1288 			(struct ip6_moptions *)0, so, NULL);
1289 		break;
1290 	    }
1291 #endif
1292 	default:
1293 		error = EAFNOSUPPORT;
1294 		break;
1295 	}
1296 	if (error) {
1297 out:
1298 		if (error == ENOBUFS) {
1299 			tcpstat.tcps_selfquench++;
1300 #ifdef INET
1301 			if (tp->t_inpcb)
1302 				tcp_quench(tp->t_inpcb, 0);
1303 #endif
1304 #ifdef INET6
1305 			if (tp->t_in6pcb)
1306 				tcp6_quench(tp->t_in6pcb, 0);
1307 #endif
1308 			error = 0;
1309 		} else if ((error == EHOSTUNREACH || error == ENETDOWN) &&
1310 		    TCPS_HAVERCVDSYN(tp->t_state)) {
1311 			tp->t_softerror = error;
1312 			error = 0;
1313 		}
1314 
1315 		/* Restart the delayed ACK timer, if necessary. */
1316 		if (tp->t_flags & TF_DELACK)
1317 			TCP_RESTART_DELACK(tp);
1318 
1319 		return (error);
1320 	}
1321 	tcpstat.tcps_sndtotal++;
1322 	if (tp->t_flags & TF_DELACK)
1323 		tcpstat.tcps_delack++;
1324 
1325 	/*
1326 	 * Data sent (as far as we can tell).
1327 	 * If this advertises a larger window than any other segment,
1328 	 * then remember the size of the advertised window.
1329 	 * Any pending ACK has now been sent.
1330 	 */
1331 	if (win > 0 && SEQ_GT(tp->rcv_nxt+win, tp->rcv_adv))
1332 		tp->rcv_adv = tp->rcv_nxt + win;
1333 	tp->last_ack_sent = tp->rcv_nxt;
1334 	tp->t_flags &= ~TF_ACKNOW;
1335 	TCP_CLEAR_DELACK(tp);
1336 #ifdef DIAGNOSTIC
1337 	if (maxburst < 0)
1338 		printf("tcp_output: maxburst exceeded by %d\n", -maxburst);
1339 #endif
1340 	if (sendalot && (!tcp_do_newreno || --maxburst))
1341 		goto again;
1342 	return (0);
1343 }
1344 
1345 void
1346 tcp_setpersist(struct tcpcb *tp)
1347 {
1348 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> (1 + 2);
1349 	int nticks;
1350 
1351 	if (TCP_TIMER_ISARMED(tp, TCPT_REXMT))
1352 		panic("tcp_output REXMT");
1353 	/*
1354 	 * Start/restart persistance timer.
1355 	 */
1356 	if (t < tp->t_rttmin)
1357 		t = tp->t_rttmin;
1358 	TCPT_RANGESET(nticks, t * tcp_backoff[tp->t_rxtshift],
1359 	    TCPTV_PERSMIN, TCPTV_PERSMAX);
1360 	TCP_TIMER_ARM(tp, TCPT_PERSIST, nticks);
1361 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1362 		tp->t_rxtshift++;
1363 }
1364