xref: /netbsd-src/sys/netinet/tcp_output.c (revision d708163f7f0abab4c3b88f79c8cf7ec8bc530426)
1 /*	$NetBSD: tcp_output.c,v 1.105 2003/10/24 10:25:40 ragge 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 1.105 2003/10/24 10:25:40 ragge 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 
158 #include <net/if.h>
159 #include <net/route.h>
160 
161 #include <netinet/in.h>
162 #include <netinet/in_systm.h>
163 #include <netinet/ip.h>
164 #include <netinet/in_pcb.h>
165 #include <netinet/ip_var.h>
166 
167 #ifdef INET6
168 #ifndef INET
169 #include <netinet/in.h>
170 #endif
171 #include <netinet/ip6.h>
172 #include <netinet6/in6_var.h>
173 #include <netinet6/ip6_var.h>
174 #include <netinet6/in6_pcb.h>
175 #include <netinet6/nd6.h>
176 #endif
177 
178 #ifdef FAST_IPSEC
179 #include <netipsec/ipsec.h>
180 #endif	/* FAST_IPSEC*/
181 #ifdef IPSEC
182 #include <netinet6/ipsec.h>
183 #endif
184 
185 #include <netinet/tcp.h>
186 #define	TCPOUTFLAGS
187 #include <netinet/tcp_fsm.h>
188 #include <netinet/tcp_seq.h>
189 #include <netinet/tcp_timer.h>
190 #include <netinet/tcp_var.h>
191 #include <netinet/tcpip.h>
192 #include <netinet/tcp_debug.h>
193 
194 #ifdef notyet
195 extern struct mbuf *m_copypack();
196 #endif
197 
198 #define MAX_TCPOPTLEN	32	/* max # bytes that go in options */
199 
200 /*
201  * Knob to enable Congestion Window Monitoring, and control the
202  * the burst size it allows.  Default burst is 4 packets, per
203  * the Internet draft.
204  */
205 int	tcp_cwm = 0;
206 int	tcp_cwm_burstsize = 4;
207 
208 #ifdef TCP_OUTPUT_COUNTERS
209 #include <sys/device.h>
210 
211 extern struct evcnt tcp_output_bigheader;
212 extern struct evcnt tcp_output_predict_hit;
213 extern struct evcnt tcp_output_predict_miss;
214 extern struct evcnt tcp_output_copysmall;
215 extern struct evcnt tcp_output_copybig;
216 extern struct evcnt tcp_output_refbig;
217 
218 #define	TCP_OUTPUT_COUNTER_INCR(ev)	(ev)->ev_count++
219 #else
220 
221 #define	TCP_OUTPUT_COUNTER_INCR(ev)	/* nothing */
222 
223 #endif /* TCP_OUTPUT_COUNTERS */
224 
225 static
226 #ifndef GPROF
227 __inline
228 #endif
229 void
230 tcp_segsize(struct tcpcb *tp, int *txsegsizep, int *rxsegsizep)
231 {
232 #ifdef INET
233 	struct inpcb *inp = tp->t_inpcb;
234 #endif
235 #ifdef INET6
236 	struct in6pcb *in6p = tp->t_in6pcb;
237 #endif
238 	struct socket *so = NULL;
239 	struct rtentry *rt;
240 	struct ifnet *ifp;
241 	int size;
242 	int iphlen;
243 	int optlen;
244 
245 #ifdef DIAGNOSTIC
246 	if (tp->t_inpcb && tp->t_in6pcb)
247 		panic("tcp_segsize: both t_inpcb and t_in6pcb are set");
248 #endif
249 	switch (tp->t_family) {
250 #ifdef INET
251 	case AF_INET:
252 		iphlen = sizeof(struct ip);
253 		break;
254 #endif
255 #ifdef INET6
256 	case AF_INET6:
257 		iphlen = sizeof(struct ip6_hdr);
258 		break;
259 #endif
260 	default:
261 		size = tcp_mssdflt;
262 		goto out;
263 	}
264 
265 	rt = NULL;
266 #ifdef INET
267 	if (inp) {
268 		rt = in_pcbrtentry(inp);
269 		so = inp->inp_socket;
270 	}
271 #endif
272 #ifdef INET6
273 	if (in6p) {
274 		rt = in6_pcbrtentry(in6p);
275 		so = in6p->in6p_socket;
276 	}
277 #endif
278 	if (rt == NULL) {
279 		size = tcp_mssdflt;
280 		goto out;
281 	}
282 
283 	ifp = rt->rt_ifp;
284 
285 	size = tcp_mssdflt;
286 	if (tp->t_mtudisc && rt->rt_rmx.rmx_mtu != 0)
287 		size = rt->rt_rmx.rmx_mtu - iphlen - sizeof(struct tcphdr);
288 	else if (ifp->if_flags & IFF_LOOPBACK)
289 		size = ifp->if_mtu - iphlen - sizeof(struct tcphdr);
290 #ifdef INET
291 	else if (inp && tp->t_mtudisc)
292 		size = ifp->if_mtu - iphlen - sizeof(struct tcphdr);
293 	else if (inp && in_localaddr(inp->inp_faddr))
294 		size = ifp->if_mtu - iphlen - sizeof(struct tcphdr);
295 #endif
296 #ifdef INET6
297 	else if (in6p) {
298 #ifdef INET
299 		if (IN6_IS_ADDR_V4MAPPED(&in6p->in6p_faddr)) {
300 			/* mapped addr case */
301 			struct in_addr d;
302 			bcopy(&in6p->in6p_faddr.s6_addr32[3], &d, sizeof(d));
303 			if (tp->t_mtudisc || in_localaddr(d))
304 				size = ifp->if_mtu - iphlen - sizeof(struct tcphdr);
305 		} else
306 #endif
307 		{
308 			/*
309 			 * for IPv6, path MTU discovery is always turned on,
310 			 * or the node must use packet size <= 1280.
311 			 */
312 			size = tp->t_mtudisc ? IN6_LINKMTU(ifp) : IPV6_MMTU;
313 			size -= (iphlen + sizeof(struct tcphdr));
314 		}
315 	}
316 #endif
317  out:
318 	/*
319 	 * Now we must make room for whatever extra TCP/IP options are in
320 	 * the packet.
321 	 */
322 	optlen = tcp_optlen(tp);
323 
324 	/*
325 	 * XXX tp->t_ourmss should have the right size, but without this code
326 	 * fragmentation will occur... need more investigation
327 	 */
328 #ifdef INET
329 	if (inp) {
330 #if defined(IPSEC) || defined(FAST_IPSEC)
331 		optlen += ipsec4_hdrsiz_tcp(tp);
332 #endif
333 		optlen += ip_optlen(inp);
334 	}
335 #endif
336 #ifdef INET6
337 #ifdef INET
338 	if (in6p && tp->t_family == AF_INET) {
339 #if defined(IPSEC) || defined(FAST_IPSEC)
340 		optlen += ipsec4_hdrsiz_tcp(tp);
341 #endif
342 		/* XXX size -= ip_optlen(in6p); */
343 	} else
344 #endif
345 	if (in6p && tp->t_family == AF_INET6) {
346 #ifdef IPSEC
347 		optlen += ipsec6_hdrsiz_tcp(tp);
348 #endif
349 		optlen += ip6_optlen(in6p);
350 	}
351 #endif
352 	size -= optlen;
353 
354 	/*
355 	 * *rxsegsizep holds *estimated* inbound segment size (estimation
356 	 * assumes that path MTU is the same for both ways).  this is only
357 	 * for silly window avoidance, do not use the value for other purposes.
358 	 *
359 	 * ipseclen is subtracted from both sides, this may not be right.
360 	 * I'm not quite sure about this (could someone comment).
361 	 */
362 	*txsegsizep = min(tp->t_peermss - optlen, size);
363 	/*
364 	 * Never send more than half a buffer full.  This insures that we can
365 	 * always keep 2 packets on the wire, no matter what SO_SNDBUF is, and
366 	 * therefore acks will never be delayed unless we run out of data to
367 	 * transmit.
368 	 */
369 	if (so)
370 		*txsegsizep = min(so->so_snd.sb_hiwat >> 1, *txsegsizep);
371 	*rxsegsizep = min(tp->t_ourmss - optlen, size);
372 
373 	if (*txsegsizep != tp->t_segsz) {
374 		/*
375 		 * If the new segment size is larger, we don't want to
376 		 * mess up the congestion window, but if it is smaller
377 		 * we'll have to reduce the congestion window to ensure
378 		 * that we don't get into trouble with initial windows
379 		 * and the rest.  In any case, if the segment size
380 		 * has changed, chances are the path has, too, and
381 		 * our congestion window will be different.
382 		 */
383 		if (*txsegsizep < tp->t_segsz) {
384 			tp->snd_cwnd = max((tp->snd_cwnd / tp->t_segsz)
385 					   * *txsegsizep, *txsegsizep);
386 			tp->snd_ssthresh = max((tp->snd_ssthresh / tp->t_segsz)
387 						* *txsegsizep, *txsegsizep);
388 		}
389 		tp->t_segsz = *txsegsizep;
390 	}
391 }
392 
393 static
394 #ifndef GPROF
395 __inline
396 #endif
397 int
398 tcp_build_datapkt(struct tcpcb *tp, struct socket *so, int off,
399     long len, int hdrlen, struct mbuf **mp)
400 {
401 	struct mbuf *m, *m0;
402 
403 	if (tp->t_force && len == 1)
404 		tcpstat.tcps_sndprobe++;
405 	else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
406 		tcpstat.tcps_sndrexmitpack++;
407 		tcpstat.tcps_sndrexmitbyte += len;
408 	} else {
409 		tcpstat.tcps_sndpack++;
410 		tcpstat.tcps_sndbyte += len;
411 	}
412 #ifdef notyet
413 	if ((m = m_copypack(so->so_snd.sb_mb, off,
414 	    (int)len, max_linkhdr + hdrlen)) == 0)
415 		return (ENOBUFS);
416 	/*
417 	 * m_copypack left space for our hdr; use it.
418 	 */
419 	m->m_len += hdrlen;
420 	m->m_data -= hdrlen;
421 #else
422 	MGETHDR(m, M_DONTWAIT, MT_HEADER);
423 	if (__predict_false(m == NULL))
424 		return (ENOBUFS);
425 	MCLAIM(m, &tcp_tx_mowner);
426 
427 	/*
428 	 * XXX Because other code assumes headers will fit in
429 	 * XXX one header mbuf.
430 	 *
431 	 * (This code should almost *never* be run.)
432 	 */
433 	if (__predict_false((max_linkhdr + hdrlen) > MHLEN)) {
434 		TCP_OUTPUT_COUNTER_INCR(&tcp_output_bigheader);
435 		MCLGET(m, M_DONTWAIT);
436 		if ((m->m_flags & M_EXT) == 0) {
437 			m_freem(m);
438 			return (ENOBUFS);
439 		}
440 	}
441 
442 	m->m_data += max_linkhdr;
443 	m->m_len = hdrlen;
444 
445 	/*
446 	 * To avoid traversing the whole sb_mb chain for correct
447 	 * data to send, remember last sent mbuf, its offset and
448 	 * the sent size.  When called the next time, see if the
449 	 * data to send is directly following the previous transfer.
450 	 * This is important for large TCP windows.
451 	 */
452 	if (
453 #ifdef FAST_MBSEARCH
454 	    off == 0 || tp->t_lastm == NULL ||
455 		(tp->t_lastoff + tp->t_lastlen) != off
456 #else
457 	    1
458 #endif
459 	   )
460 	{
461 #ifdef FAST_MBSEARCH
462 		TCP_OUTPUT_COUNTER_INCR(&tcp_output_predict_miss);
463 #endif
464 		/*
465 		 * Either a new packet or a retransmit.
466 		 * Start from the beginning.
467 		 */
468 		tp->t_lastm = so->so_snd.sb_mb;
469 		tp->t_inoff = off;
470 	} else {
471 #ifdef FAST_MBSEARCH
472 		TCP_OUTPUT_COUNTER_INCR(&tcp_output_predict_hit);
473 #endif
474 		tp->t_inoff += tp->t_lastlen;
475 	}
476 
477 	/* Traverse forward to next packet */
478 	while (tp->t_inoff > 0) {
479 		if (tp->t_lastm == NULL)
480 			panic("tp->t_lastm == NULL");
481 		if (tp->t_inoff < tp->t_lastm->m_len)
482 			break;
483 		tp->t_inoff -= tp->t_lastm->m_len;
484 		tp->t_lastm = tp->t_lastm->m_next;
485 	}
486 
487 	tp->t_lastoff = off;
488 	tp->t_lastlen = len;
489 	m0 = tp->t_lastm;
490 	off = tp->t_inoff;
491 
492 	if (len <= M_TRAILINGSPACE(m)) {
493 		m_copydata(m0, off, (int) len, mtod(m, caddr_t) + hdrlen);
494 		m->m_len += len;
495 		TCP_OUTPUT_COUNTER_INCR(&tcp_output_copysmall);
496 	} else {
497 		m->m_next = m_copy(m0, off, (int) len);
498 		if (m->m_next == NULL) {
499 			m_freem(m);
500 			return (ENOBUFS);
501 		}
502 #ifdef TCP_OUTPUT_COUNTERS
503 		if (m->m_next->m_flags & M_EXT)
504 			TCP_OUTPUT_COUNTER_INCR(&tcp_output_refbig);
505 		else
506 			TCP_OUTPUT_COUNTER_INCR(&tcp_output_copybig);
507 #endif /* TCP_OUTPUT_COUNTERS */
508 	}
509 #endif
510 
511 	*mp = m;
512 	return (0);
513 }
514 
515 /*
516  * Tcp output routine: figure out what should be sent and send it.
517  */
518 int
519 tcp_output(tp)
520 	struct tcpcb *tp;
521 {
522 	struct socket *so;
523 	struct route *ro;
524 	long len, win;
525 	int off, flags, error;
526 	struct mbuf *m;
527 	struct ip *ip;
528 #ifdef INET6
529 	struct ip6_hdr *ip6;
530 #endif
531 	struct tcphdr *th;
532 	u_char opt[MAX_TCPOPTLEN];
533 	unsigned optlen, hdrlen;
534 	int idle, sendalot, txsegsize, rxsegsize;
535 	int maxburst = TCP_MAXBURST;
536 	int af;		/* address family on the wire */
537 	int iphdrlen;
538 
539 #ifdef DIAGNOSTIC
540 	if (tp->t_inpcb && tp->t_in6pcb)
541 		panic("tcp_output: both t_inpcb and t_in6pcb are set");
542 #endif
543 	so = NULL;
544 	ro = NULL;
545 	if (tp->t_inpcb) {
546 		so = tp->t_inpcb->inp_socket;
547 		ro = &tp->t_inpcb->inp_route;
548 	}
549 #ifdef INET6
550 	else if (tp->t_in6pcb) {
551 		so = tp->t_in6pcb->in6p_socket;
552 		ro = (struct route *)&tp->t_in6pcb->in6p_route;
553 	}
554 #endif
555 
556 	switch (af = tp->t_family) {
557 #ifdef INET
558 	case AF_INET:
559 		if (tp->t_inpcb)
560 			break;
561 #ifdef INET6
562 		/* mapped addr case */
563 		if (tp->t_in6pcb)
564 			break;
565 #endif
566 		return EINVAL;
567 #endif
568 #ifdef INET6
569 	case AF_INET6:
570 		if (tp->t_in6pcb)
571 			break;
572 		return EINVAL;
573 #endif
574 	default:
575 		return EAFNOSUPPORT;
576 	}
577 
578 	tcp_segsize(tp, &txsegsize, &rxsegsize);
579 
580 	idle = (tp->snd_max == tp->snd_una);
581 
582 	/*
583 	 * Restart Window computation.  From draft-floyd-incr-init-win-03:
584 	 *
585 	 *	Optionally, a TCP MAY set the restart window to the
586 	 *	minimum of the value used for the initial window and
587 	 *	the current value of cwnd (in other words, using a
588 	 *	larger value for the restart window should never increase
589 	 *	the size of cwnd).
590 	 */
591 	if (tcp_cwm) {
592 		/*
593 		 * Hughes/Touch/Heidemann Congestion Window Monitoring.
594 		 * Count the number of packets currently pending
595 		 * acknowledgement, and limit our congestion window
596 		 * to a pre-determined allowed burst size plus that count.
597 		 * This prevents bursting once all pending packets have
598 		 * been acknowledged (i.e. transmission is idle).
599 		 *
600 		 * XXX Link this to Initial Window?
601 		 */
602 		tp->snd_cwnd = min(tp->snd_cwnd,
603 		    (tcp_cwm_burstsize * txsegsize) +
604 		    (tp->snd_nxt - tp->snd_una));
605 	} else {
606 		if (idle && (tcp_now - tp->t_rcvtime) >= tp->t_rxtcur) {
607 			/*
608 			 * We have been idle for "a while" and no acks are
609 			 * expected to clock out any data we send --
610 			 * slow start to get ack "clock" running again.
611 			 */
612 			int ss = tcp_init_win;
613 #ifdef INET
614 			if (tp->t_inpcb &&
615 			    in_localaddr(tp->t_inpcb->inp_faddr))
616 				ss = tcp_init_win_local;
617 #endif
618 #ifdef INET6
619 			if (tp->t_in6pcb &&
620 			    in6_localaddr(&tp->t_in6pcb->in6p_faddr))
621 				ss = tcp_init_win_local;
622 #endif
623 			tp->snd_cwnd = min(tp->snd_cwnd,
624 			    TCP_INITIAL_WINDOW(ss, txsegsize));
625 		}
626 	}
627 
628 again:
629 	/*
630 	 * Determine length of data that should be transmitted, and
631 	 * flags that should be used.  If there is some data or critical
632 	 * controls (SYN, RST) to send, then transmit; otherwise,
633 	 * investigate further.
634 	 */
635 	sendalot = 0;
636 	off = tp->snd_nxt - tp->snd_una;
637 	win = min(tp->snd_wnd, tp->snd_cwnd);
638 
639 	flags = tcp_outflags[tp->t_state];
640 	/*
641 	 * If in persist timeout with window of 0, send 1 byte.
642 	 * Otherwise, if window is small but nonzero
643 	 * and timer expired, we will send what we can
644 	 * and go to transmit state.
645 	 */
646 	if (tp->t_force) {
647 		if (win == 0) {
648 			/*
649 			 * If we still have some data to send, then
650 			 * clear the FIN bit.  Usually this would
651 			 * happen below when it realizes that we
652 			 * aren't sending all the data.  However,
653 			 * if we have exactly 1 byte of unset data,
654 			 * then it won't clear the FIN bit below,
655 			 * and if we are in persist state, we wind
656 			 * up sending the packet without recording
657 			 * that we sent the FIN bit.
658 			 *
659 			 * We can't just blindly clear the FIN bit,
660 			 * because if we don't have any more data
661 			 * to send then the probe will be the FIN
662 			 * itself.
663 			 */
664 			if (off < so->so_snd.sb_cc)
665 				flags &= ~TH_FIN;
666 			win = 1;
667 		} else {
668 			TCP_TIMER_DISARM(tp, TCPT_PERSIST);
669 			tp->t_rxtshift = 0;
670 		}
671 	}
672 
673 	if (win < so->so_snd.sb_cc) {
674 		len = win - off;
675 		flags &= ~TH_FIN;
676 	} else
677 		len = so->so_snd.sb_cc - off;
678 
679 	if (len < 0) {
680 		/*
681 		 * If FIN has been sent but not acked,
682 		 * but we haven't been called to retransmit,
683 		 * len will be -1.  Otherwise, window shrank
684 		 * after we sent into it.  If window shrank to 0,
685 		 * cancel pending retransmit, pull snd_nxt back
686 		 * to (closed) window, and set the persist timer
687 		 * if it isn't already going.  If the window didn't
688 		 * close completely, just wait for an ACK.
689 		 *
690 		 * If we have a pending FIN, either it has already been
691 		 * transmitted or it is outside the window, so drop it.
692 		 * If the FIN has been transmitted, but this is not a
693 		 * retransmission, then len must be -1.  Therefore we also
694 		 * prevent here the sending of `gratuitous FINs'.  This
695 		 * eliminates the need to check for that case below (e.g.
696 		 * to back up snd_nxt before the FIN so that the sequence
697 		 * number is correct).
698 		 */
699 		len = 0;
700 		flags &= ~TH_FIN;
701 		if (win == 0) {
702 			TCP_TIMER_DISARM(tp, TCPT_REXMT);
703 			tp->t_rxtshift = 0;
704 			tp->snd_nxt = tp->snd_una;
705 			if (TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0)
706 				tcp_setpersist(tp);
707 		}
708 	}
709 	if (len > txsegsize) {
710 		len = txsegsize;
711 		flags &= ~TH_FIN;
712 		sendalot = 1;
713 	}
714 
715 	win = sbspace(&so->so_rcv);
716 
717 	/*
718 	 * Sender silly window avoidance.  If connection is idle
719 	 * and can send all data, a maximum segment,
720 	 * at least a maximum default-size segment do it,
721 	 * or are forced, do it; otherwise don't bother.
722 	 * If peer's buffer is tiny, then send
723 	 * when window is at least half open.
724 	 * If retransmitting (possibly after persist timer forced us
725 	 * to send into a small window), then must resend.
726 	 */
727 	if (len) {
728 		if (len == txsegsize)
729 			goto send;
730 		if ((so->so_state & SS_MORETOCOME) == 0 &&
731 		    ((idle || tp->t_flags & TF_NODELAY) &&
732 		     len + off >= so->so_snd.sb_cc))
733 			goto send;
734 		if (tp->t_force)
735 			goto send;
736 		if (len >= tp->max_sndwnd / 2)
737 			goto send;
738 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))
739 			goto send;
740 	}
741 
742 	/*
743 	 * Compare available window to amount of window known to peer
744 	 * (as advertised window less next expected input).  If the
745 	 * difference is at least twice the size of the largest segment
746 	 * we expect to receive (i.e. two segments) or at least 50% of
747 	 * the maximum possible window, then want to send a window update
748 	 * to peer.
749 	 */
750 	if (win > 0) {
751 		/*
752 		 * "adv" is the amount we can increase the window,
753 		 * taking into account that we are limited by
754 		 * TCP_MAXWIN << tp->rcv_scale.
755 		 */
756 		long adv = min(win, (long)TCP_MAXWIN << tp->rcv_scale) -
757 			(tp->rcv_adv - tp->rcv_nxt);
758 
759 		if (adv >= (long) (2 * rxsegsize))
760 			goto send;
761 		if (2 * adv >= (long) so->so_rcv.sb_hiwat)
762 			goto send;
763 	}
764 
765 	/*
766 	 * Send if we owe peer an ACK.
767 	 */
768 	if (tp->t_flags & TF_ACKNOW)
769 		goto send;
770 	if (flags & (TH_SYN|TH_FIN|TH_RST))
771 		goto send;
772 	if (SEQ_GT(tp->snd_up, tp->snd_una))
773 		goto send;
774 
775 	/*
776 	 * TCP window updates are not reliable, rather a polling protocol
777 	 * using ``persist'' packets is used to insure receipt of window
778 	 * updates.  The three ``states'' for the output side are:
779 	 *	idle			not doing retransmits or persists
780 	 *	persisting		to move a small or zero window
781 	 *	(re)transmitting	and thereby not persisting
782 	 *
783 	 * tp->t_timer[TCPT_PERSIST]
784 	 *	is set when we are in persist state.
785 	 * tp->t_force
786 	 *	is set when we are called to send a persist packet.
787 	 * tp->t_timer[TCPT_REXMT]
788 	 *	is set when we are retransmitting
789 	 * The output side is idle when both timers are zero.
790 	 *
791 	 * If send window is too small, there is data to transmit, and no
792 	 * retransmit or persist is pending, then go to persist state.
793 	 * If nothing happens soon, send when timer expires:
794 	 * if window is nonzero, transmit what we can,
795 	 * otherwise force out a byte.
796 	 */
797 	if (so->so_snd.sb_cc && TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 &&
798 	    TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) {
799 		tp->t_rxtshift = 0;
800 		tcp_setpersist(tp);
801 	}
802 
803 	/*
804 	 * No reason to send a segment, just return.
805 	 */
806 	return (0);
807 
808 send:
809 	/*
810 	 * Before ESTABLISHED, force sending of initial options
811 	 * unless TCP set not to do any options.
812 	 * NOTE: we assume that the IP/TCP header plus TCP options
813 	 * always fit in a single mbuf, leaving room for a maximum
814 	 * link header, i.e.
815 	 *	max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
816 	 */
817 	optlen = 0;
818 	switch (af) {
819 #ifdef INET
820 	case AF_INET:
821 		iphdrlen = sizeof(struct ip) + sizeof(struct tcphdr);
822 		break;
823 #endif
824 #ifdef INET6
825 	case AF_INET6:
826 		iphdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
827 		break;
828 #endif
829 	default:	/*pacify gcc*/
830 		iphdrlen = 0;
831 		break;
832 	}
833 	hdrlen = iphdrlen;
834 	if (flags & TH_SYN) {
835 		struct rtentry *rt;
836 
837 		rt = NULL;
838 #ifdef INET
839 		if (tp->t_inpcb)
840 			rt = in_pcbrtentry(tp->t_inpcb);
841 #endif
842 #ifdef INET6
843 		if (tp->t_in6pcb)
844 			rt = in6_pcbrtentry(tp->t_in6pcb);
845 #endif
846 
847 		tp->snd_nxt = tp->iss;
848 		tp->t_ourmss = tcp_mss_to_advertise(rt != NULL ?
849 						    rt->rt_ifp : NULL, af);
850 		if ((tp->t_flags & TF_NOOPT) == 0) {
851 			opt[0] = TCPOPT_MAXSEG;
852 			opt[1] = 4;
853 			opt[2] = (tp->t_ourmss >> 8) & 0xff;
854 			opt[3] = tp->t_ourmss & 0xff;
855 			optlen = 4;
856 
857 			if ((tp->t_flags & TF_REQ_SCALE) &&
858 			    ((flags & TH_ACK) == 0 ||
859 			    (tp->t_flags & TF_RCVD_SCALE))) {
860 				*((u_int32_t *) (opt + optlen)) = htonl(
861 					TCPOPT_NOP << 24 |
862 					TCPOPT_WINDOW << 16 |
863 					TCPOLEN_WINDOW << 8 |
864 					tp->request_r_scale);
865 				optlen += 4;
866 			}
867 		}
868 	}
869 
870 	/*
871 	 * Send a timestamp and echo-reply if this is a SYN and our side
872 	 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side
873 	 * and our peer have sent timestamps in our SYN's.
874 	 */
875 	if ((tp->t_flags & (TF_REQ_TSTMP|TF_NOOPT)) == TF_REQ_TSTMP &&
876 	     (flags & TH_RST) == 0 &&
877 	    ((flags & (TH_SYN|TH_ACK)) == TH_SYN ||
878 	     (tp->t_flags & TF_RCVD_TSTMP))) {
879 		u_int32_t *lp = (u_int32_t *)(opt + optlen);
880 
881 		/* Form timestamp option as shown in appendix A of RFC 1323. */
882 		*lp++ = htonl(TCPOPT_TSTAMP_HDR);
883 		*lp++ = htonl(TCP_TIMESTAMP(tp));
884 		*lp   = htonl(tp->ts_recent);
885 		optlen += TCPOLEN_TSTAMP_APPA;
886 	}
887 
888 	hdrlen += optlen;
889 
890 #ifdef DIAGNOSTIC
891 	if (len > txsegsize)
892 		panic("tcp data to be sent is larger than segment");
893 	if (max_linkhdr + hdrlen > MCLBYTES)
894 		panic("tcphdr too big");
895 #endif
896 
897 	/*
898 	 * Grab a header mbuf, attaching a copy of data to
899 	 * be transmitted, and initialize the header from
900 	 * the template for sends on this connection.
901 	 */
902 	if (len) {
903 		error = tcp_build_datapkt(tp, so, off, len, hdrlen, &m);
904 		if (error)
905 			goto out;
906 		/*
907 		 * If we're sending everything we've got, set PUSH.
908 		 * (This will keep happy those implementations which only
909 		 * give data to the user when a buffer fills or
910 		 * a PUSH comes in.)
911 		 */
912 		if (off + len == so->so_snd.sb_cc)
913 			flags |= TH_PUSH;
914 	} else {
915 		if (tp->t_flags & TF_ACKNOW)
916 			tcpstat.tcps_sndacks++;
917 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
918 			tcpstat.tcps_sndctrl++;
919 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
920 			tcpstat.tcps_sndurg++;
921 		else
922 			tcpstat.tcps_sndwinup++;
923 
924 		MGETHDR(m, M_DONTWAIT, MT_HEADER);
925 		if (m != NULL && max_linkhdr + hdrlen > MHLEN) {
926 			MCLGET(m, M_DONTWAIT);
927 			if ((m->m_flags & M_EXT) == 0) {
928 				m_freem(m);
929 				m = NULL;
930 			}
931 		}
932 		if (m == NULL) {
933 			error = ENOBUFS;
934 			goto out;
935 		}
936 		MCLAIM(m, &tcp_tx_mowner);
937 		m->m_data += max_linkhdr;
938 		m->m_len = hdrlen;
939 	}
940 	m->m_pkthdr.rcvif = (struct ifnet *)0;
941 	switch (af) {
942 #ifdef INET
943 	case AF_INET:
944 		ip = mtod(m, struct ip *);
945 #ifdef INET6
946 		ip6 = NULL;
947 #endif
948 		th = (struct tcphdr *)(ip + 1);
949 		break;
950 #endif
951 #ifdef INET6
952 	case AF_INET6:
953 		ip = NULL;
954 		ip6 = mtod(m, struct ip6_hdr *);
955 		th = (struct tcphdr *)(ip6 + 1);
956 		break;
957 #endif
958 	default:	/*pacify gcc*/
959 		ip = NULL;
960 #ifdef INET6
961 		ip6 = NULL;
962 #endif
963 		th = NULL;
964 		break;
965 	}
966 	if (tp->t_template == 0)
967 		panic("tcp_output");
968 	if (tp->t_template->m_len < iphdrlen)
969 		panic("tcp_output");
970 	bcopy(mtod(tp->t_template, caddr_t), mtod(m, caddr_t), iphdrlen);
971 
972 	/*
973 	 * If we are doing retransmissions, then snd_nxt will
974 	 * not reflect the first unsent octet.  For ACK only
975 	 * packets, we do not want the sequence number of the
976 	 * retransmitted packet, we want the sequence number
977 	 * of the next unsent octet.  So, if there is no data
978 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
979 	 * when filling in ti_seq.  But if we are in persist
980 	 * state, snd_max might reflect one byte beyond the
981 	 * right edge of the window, so use snd_nxt in that
982 	 * case, since we know we aren't doing a retransmission.
983 	 * (retransmit and persist are mutually exclusive...)
984 	 */
985 	if (len || (flags & (TH_SYN|TH_FIN)) ||
986 	    TCP_TIMER_ISARMED(tp, TCPT_PERSIST))
987 		th->th_seq = htonl(tp->snd_nxt);
988 	else
989 		th->th_seq = htonl(tp->snd_max);
990 	th->th_ack = htonl(tp->rcv_nxt);
991 	if (optlen) {
992 		bcopy((caddr_t)opt, (caddr_t)(th + 1), optlen);
993 		th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
994 	}
995 	th->th_flags = flags;
996 	/*
997 	 * Calculate receive window.  Don't shrink window,
998 	 * but avoid silly window syndrome.
999 	 */
1000 	if (win < (long)(so->so_rcv.sb_hiwat / 4) && win < (long)rxsegsize)
1001 		win = 0;
1002 	if (win > (long)TCP_MAXWIN << tp->rcv_scale)
1003 		win = (long)TCP_MAXWIN << tp->rcv_scale;
1004 	if (win < (long)(tp->rcv_adv - tp->rcv_nxt))
1005 		win = (long)(tp->rcv_adv - tp->rcv_nxt);
1006 	th->th_win = htons((u_int16_t) (win>>tp->rcv_scale));
1007 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1008 		u_int32_t urp = tp->snd_up - tp->snd_nxt;
1009 		if (urp > IP_MAXPACKET)
1010 			urp = IP_MAXPACKET;
1011 		th->th_urp = htons((u_int16_t)urp);
1012 		th->th_flags |= TH_URG;
1013 	} else
1014 		/*
1015 		 * If no urgent pointer to send, then we pull
1016 		 * the urgent pointer to the left edge of the send window
1017 		 * so that it doesn't drift into the send window on sequence
1018 		 * number wraparound.
1019 		 */
1020 		tp->snd_up = tp->snd_una;		/* drag it along */
1021 
1022 	/*
1023 	 * Set ourselves up to be checksummed just before the packet
1024 	 * hits the wire.
1025 	 */
1026 	switch (af) {
1027 #ifdef INET
1028 	case AF_INET:
1029 		m->m_pkthdr.csum_flags = M_CSUM_TCPv4;
1030 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1031 		if (len + optlen) {
1032 			/* Fixup the pseudo-header checksum. */
1033 			/* XXXJRT Not IP Jumbogram safe. */
1034 			th->th_sum = in_cksum_addword(th->th_sum,
1035 			    htons((u_int16_t) (len + optlen)));
1036 		}
1037 		break;
1038 #endif
1039 #ifdef INET6
1040 	case AF_INET6:
1041 		/*
1042 		 * XXX Actually delaying the checksum is Hard
1043 		 * XXX (well, maybe not for Itojun, but it is
1044 		 * XXX for me), but we can still take advantage
1045 		 * XXX of the cached pseudo-header checksum.
1046 		 */
1047 		/* equals to hdrlen + len */
1048 		m->m_pkthdr.len = sizeof(struct ip6_hdr)
1049 			+ sizeof(struct tcphdr) + optlen + len;
1050 #ifdef notyet
1051 		m->m_pkthdr.csum_flags = M_CSUM_TCPv6;
1052 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1053 #endif
1054 		if (len + optlen) {
1055 			/* Fixup the pseudo-header checksum. */
1056 			/* XXXJRT: Not IPv6 Jumbogram safe. */
1057 			th->th_sum = in_cksum_addword(th->th_sum,
1058 			    htons((u_int16_t) (len + optlen)));
1059 		}
1060 #ifndef notyet
1061 		th->th_sum = in6_cksum(m, 0, sizeof(struct ip6_hdr),
1062 		    sizeof(struct tcphdr) + optlen + len);
1063 #endif
1064 		break;
1065 #endif
1066 	}
1067 
1068 	/*
1069 	 * In transmit state, time the transmission and arrange for
1070 	 * the retransmit.  In persist state, just set snd_max.
1071 	 */
1072 	if (tp->t_force == 0 || TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) {
1073 		tcp_seq startseq = tp->snd_nxt;
1074 
1075 		/*
1076 		 * Advance snd_nxt over sequence space of this segment.
1077 		 * There are no states in which we send both a SYN and a FIN,
1078 		 * so we collapse the tests for these flags.
1079 		 */
1080 		if (flags & (TH_SYN|TH_FIN))
1081 			tp->snd_nxt++;
1082 		tp->snd_nxt += len;
1083 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1084 			tp->snd_max = tp->snd_nxt;
1085 			/*
1086 			 * Time this transmission if not a retransmission and
1087 			 * not currently timing anything.
1088 			 */
1089 			if (tp->t_rtttime == 0) {
1090 				tp->t_rtttime = tcp_now;
1091 				tp->t_rtseq = startseq;
1092 				tcpstat.tcps_segstimed++;
1093 			}
1094 		}
1095 
1096 		/*
1097 		 * Set retransmit timer if not currently set,
1098 		 * and not doing an ack or a keep-alive probe.
1099 		 * Initial value for retransmit timer is smoothed
1100 		 * round-trip time + 2 * round-trip time variance.
1101 		 * Initialize shift counter which is used for backoff
1102 		 * of retransmit time.
1103 		 */
1104 		if (TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 &&
1105 		    tp->snd_nxt != tp->snd_una) {
1106 			TCP_TIMER_ARM(tp, TCPT_REXMT, tp->t_rxtcur);
1107 			if (TCP_TIMER_ISARMED(tp, TCPT_PERSIST)) {
1108 				TCP_TIMER_DISARM(tp, TCPT_PERSIST);
1109 				tp->t_rxtshift = 0;
1110 			}
1111 		}
1112 	} else
1113 		if (SEQ_GT(tp->snd_nxt + len, tp->snd_max))
1114 			tp->snd_max = tp->snd_nxt + len;
1115 
1116 #ifdef TCP_DEBUG
1117 	/*
1118 	 * Trace.
1119 	 */
1120 	if (so->so_options & SO_DEBUG)
1121 		tcp_trace(TA_OUTPUT, tp->t_state, tp, m, 0);
1122 #endif
1123 
1124 	/*
1125 	 * Fill in IP length and desired time to live and
1126 	 * send to IP level.  There should be a better way
1127 	 * to handle ttl and tos; we could keep them in
1128 	 * the template, but need a way to checksum without them.
1129 	 */
1130 	m->m_pkthdr.len = hdrlen + len;
1131 
1132 	switch (af) {
1133 #ifdef INET
1134 	case AF_INET:
1135 		ip->ip_len = htons(m->m_pkthdr.len);
1136 		if (tp->t_inpcb) {
1137 			ip->ip_ttl = tp->t_inpcb->inp_ip.ip_ttl;
1138 			ip->ip_tos = tp->t_inpcb->inp_ip.ip_tos;
1139 		}
1140 #ifdef INET6
1141 		else if (tp->t_in6pcb) {
1142 			ip->ip_ttl = in6_selecthlim(tp->t_in6pcb, NULL); /*XXX*/
1143 			ip->ip_tos = 0;	/*XXX*/
1144 		}
1145 #endif
1146 		break;
1147 #endif
1148 #ifdef INET6
1149 	case AF_INET6:
1150 		ip6->ip6_nxt = IPPROTO_TCP;
1151 		if (tp->t_in6pcb) {
1152 			/*
1153 			 * we separately set hoplimit for every segment, since
1154 			 * the user might want to change the value via
1155 			 * setsockopt. Also, desired default hop limit might
1156 			 * be changed via Neighbor Discovery.
1157 			 */
1158 			ip6->ip6_hlim = in6_selecthlim(tp->t_in6pcb,
1159 				ro->ro_rt ? ro->ro_rt->rt_ifp : NULL);
1160 		}
1161 		/* ip6->ip6_flow = ??? */
1162 		/* ip6_plen will be filled in ip6_output(). */
1163 		break;
1164 #endif
1165 	}
1166 
1167 	switch (af) {
1168 #ifdef INET
1169 	case AF_INET:
1170 	    {
1171 		struct mbuf *opts;
1172 
1173 		if (tp->t_inpcb)
1174 			opts = tp->t_inpcb->inp_options;
1175 		else
1176 			opts = NULL;
1177 		error = ip_output(m, opts, ro,
1178 			(tp->t_mtudisc ? IP_MTUDISC : 0) |
1179 			(so->so_options & SO_DONTROUTE),
1180 			(struct ip_moptions *)0, so);
1181 		break;
1182 	    }
1183 #endif
1184 #ifdef INET6
1185 	case AF_INET6:
1186 	    {
1187 		struct ip6_pktopts *opts;
1188 
1189 		if (tp->t_in6pcb)
1190 			opts = tp->t_in6pcb->in6p_outputopts;
1191 		else
1192 			opts = NULL;
1193 		error = ip6_output(m, opts, (struct route_in6 *)ro,
1194 			so->so_options & SO_DONTROUTE,
1195 			(struct ip6_moptions *)0, so, NULL);
1196 		break;
1197 	    }
1198 #endif
1199 	default:
1200 		error = EAFNOSUPPORT;
1201 		break;
1202 	}
1203 	if (error) {
1204 out:
1205 		if (error == ENOBUFS) {
1206 			tcpstat.tcps_selfquench++;
1207 #ifdef INET
1208 			if (tp->t_inpcb)
1209 				tcp_quench(tp->t_inpcb, 0);
1210 #endif
1211 #ifdef INET6
1212 			if (tp->t_in6pcb)
1213 				tcp6_quench(tp->t_in6pcb, 0);
1214 #endif
1215 			error = 0;
1216 		} else if ((error == EHOSTUNREACH || error == ENETDOWN) &&
1217 		    TCPS_HAVERCVDSYN(tp->t_state)) {
1218 			tp->t_softerror = error;
1219 			error = 0;
1220 		}
1221 
1222 		/* Restart the delayed ACK timer, if necessary. */
1223 		if (tp->t_flags & TF_DELACK)
1224 			TCP_RESTART_DELACK(tp);
1225 
1226 		return (error);
1227 	}
1228 	tcpstat.tcps_sndtotal++;
1229 	if (tp->t_flags & TF_DELACK)
1230 		tcpstat.tcps_delack++;
1231 
1232 	/*
1233 	 * Data sent (as far as we can tell).
1234 	 * If this advertises a larger window than any other segment,
1235 	 * then remember the size of the advertised window.
1236 	 * Any pending ACK has now been sent.
1237 	 */
1238 	if (win > 0 && SEQ_GT(tp->rcv_nxt+win, tp->rcv_adv))
1239 		tp->rcv_adv = tp->rcv_nxt + win;
1240 	tp->last_ack_sent = tp->rcv_nxt;
1241 	tp->t_flags &= ~TF_ACKNOW;
1242 	TCP_CLEAR_DELACK(tp);
1243 #ifdef DIAGNOSTIC
1244 	if (maxburst < 0)
1245 		printf("tcp_output: maxburst exceeded by %d\n", -maxburst);
1246 #endif
1247 	if (sendalot && (!tcp_do_newreno || --maxburst))
1248 		goto again;
1249 	return (0);
1250 }
1251 
1252 void
1253 tcp_setpersist(tp)
1254 	struct tcpcb *tp;
1255 {
1256 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> (1 + 2);
1257 	int nticks;
1258 
1259 	if (TCP_TIMER_ISARMED(tp, TCPT_REXMT))
1260 		panic("tcp_output REXMT");
1261 	/*
1262 	 * Start/restart persistance timer.
1263 	 */
1264 	if (t < tp->t_rttmin)
1265 		t = tp->t_rttmin;
1266 	TCPT_RANGESET(nticks, t * tcp_backoff[tp->t_rxtshift],
1267 	    TCPTV_PERSMIN, TCPTV_PERSMAX);
1268 	TCP_TIMER_ARM(tp, TCPT_PERSIST, nticks);
1269 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1270 		tp->t_rxtshift++;
1271 }
1272