xref: /netbsd-src/sys/netinet/tcp_output.c (revision 92570d0dfc3fce0c370aa0a225edd325598ed3b1)
1 /*	$NetBSD: tcp_output.c,v 1.131 2005/04/18 21:55:06 yamt 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, 2005 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  * This code is derived from software contributed to The NetBSD Foundation
80  * by Charles M. Hannum.
81  *
82  * Redistribution and use in source and binary forms, with or without
83  * modification, are permitted provided that the following conditions
84  * are met:
85  * 1. Redistributions of source code must retain the above copyright
86  *    notice, this list of conditions and the following disclaimer.
87  * 2. Redistributions in binary form must reproduce the above copyright
88  *    notice, this list of conditions and the following disclaimer in the
89  *    documentation and/or other materials provided with the distribution.
90  * 3. All advertising materials mentioning features or use of this software
91  *    must display the following acknowledgement:
92  *	This product includes software developed by the NetBSD
93  *	Foundation, Inc. and its contributors.
94  * 4. Neither the name of The NetBSD Foundation nor the names of its
95  *    contributors may be used to endorse or promote products derived
96  *    from this software without specific prior written permission.
97  *
98  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
99  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
100  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
101  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
102  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
103  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
104  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
105  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
106  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
107  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
108  * POSSIBILITY OF SUCH DAMAGE.
109  */
110 
111 /*
112  * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995
113  *	The Regents of the University of California.  All rights reserved.
114  *
115  * Redistribution and use in source and binary forms, with or without
116  * modification, are permitted provided that the following conditions
117  * are met:
118  * 1. Redistributions of source code must retain the above copyright
119  *    notice, this list of conditions and the following disclaimer.
120  * 2. Redistributions in binary form must reproduce the above copyright
121  *    notice, this list of conditions and the following disclaimer in the
122  *    documentation and/or other materials provided with the distribution.
123  * 3. Neither the name of the University nor the names of its contributors
124  *    may be used to endorse or promote products derived from this software
125  *    without specific prior written permission.
126  *
127  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
128  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
129  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
130  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
131  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
132  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
133  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
134  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
135  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
136  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
137  * SUCH DAMAGE.
138  *
139  *	@(#)tcp_output.c	8.4 (Berkeley) 5/24/95
140  */
141 
142 #include <sys/cdefs.h>
143 __KERNEL_RCSID(0, "$NetBSD: tcp_output.c,v 1.131 2005/04/18 21:55:06 yamt Exp $");
144 
145 #include "opt_inet.h"
146 #include "opt_ipsec.h"
147 #include "opt_tcp_debug.h"
148 
149 #include <sys/param.h>
150 #include <sys/systm.h>
151 #include <sys/malloc.h>
152 #include <sys/mbuf.h>
153 #include <sys/protosw.h>
154 #include <sys/socket.h>
155 #include <sys/socketvar.h>
156 #include <sys/errno.h>
157 #include <sys/domain.h>
158 #include <sys/kernel.h>
159 #ifdef TCP_SIGNATURE
160 #include <sys/md5.h>
161 #endif
162 
163 #include <net/if.h>
164 #include <net/route.h>
165 
166 #include <netinet/in.h>
167 #include <netinet/in_systm.h>
168 #include <netinet/ip.h>
169 #include <netinet/in_pcb.h>
170 #include <netinet/ip_var.h>
171 
172 #ifdef INET6
173 #ifndef INET
174 #include <netinet/in.h>
175 #endif
176 #include <netinet/ip6.h>
177 #include <netinet6/in6_var.h>
178 #include <netinet6/ip6_var.h>
179 #include <netinet6/in6_pcb.h>
180 #include <netinet6/nd6.h>
181 #endif
182 
183 #ifdef FAST_IPSEC
184 #include <netipsec/ipsec.h>
185 #include <netipsec/key.h>
186 #endif	/* FAST_IPSEC*/
187 #ifdef IPSEC
188 #include <netinet6/ipsec.h>
189 #endif
190 
191 #include <netinet/tcp.h>
192 #define	TCPOUTFLAGS
193 #include <netinet/tcp_fsm.h>
194 #include <netinet/tcp_seq.h>
195 #include <netinet/tcp_timer.h>
196 #include <netinet/tcp_var.h>
197 #include <netinet/tcpip.h>
198 #include <netinet/tcp_debug.h>
199 #include <netinet/in_offload.h>
200 
201 #ifdef IPSEC
202 #include <netkey/key.h>
203 #endif
204 
205 #ifdef notyet
206 extern struct mbuf *m_copypack();
207 #endif
208 
209 /*
210  * Knob to enable Congestion Window Monitoring, and control the
211  * the burst size it allows.  Default burst is 4 packets, per
212  * the Internet draft.
213  */
214 int	tcp_cwm = 0;
215 int	tcp_cwm_burstsize = 4;
216 
217 #ifdef TCP_OUTPUT_COUNTERS
218 #include <sys/device.h>
219 
220 extern struct evcnt tcp_output_bigheader;
221 extern struct evcnt tcp_output_predict_hit;
222 extern struct evcnt tcp_output_predict_miss;
223 extern struct evcnt tcp_output_copysmall;
224 extern struct evcnt tcp_output_copybig;
225 extern struct evcnt tcp_output_refbig;
226 
227 #define	TCP_OUTPUT_COUNTER_INCR(ev)	(ev)->ev_count++
228 #else
229 
230 #define	TCP_OUTPUT_COUNTER_INCR(ev)	/* nothing */
231 
232 #endif /* TCP_OUTPUT_COUNTERS */
233 
234 static
235 #ifndef GPROF
236 __inline
237 #endif
238 int
239 tcp_segsize(struct tcpcb *tp, int *txsegsizep, int *rxsegsizep)
240 {
241 #ifdef INET
242 	struct inpcb *inp = tp->t_inpcb;
243 #endif
244 #ifdef INET6
245 	struct in6pcb *in6p = tp->t_in6pcb;
246 #endif
247 	struct socket *so = NULL;
248 	struct rtentry *rt;
249 	struct ifnet *ifp;
250 	int size;
251 	int iphlen;
252 	int optlen;
253 
254 #ifdef DIAGNOSTIC
255 	if (tp->t_inpcb && tp->t_in6pcb)
256 		panic("tcp_segsize: both t_inpcb and t_in6pcb are set");
257 #endif
258 	switch (tp->t_family) {
259 #ifdef INET
260 	case AF_INET:
261 		iphlen = sizeof(struct ip);
262 		break;
263 #endif
264 #ifdef INET6
265 	case AF_INET6:
266 		iphlen = sizeof(struct ip6_hdr);
267 		break;
268 #endif
269 	default:
270 		size = tcp_mssdflt;
271 		goto out;
272 	}
273 
274 	rt = NULL;
275 #ifdef INET
276 	if (inp) {
277 		rt = in_pcbrtentry(inp);
278 		so = inp->inp_socket;
279 	}
280 #endif
281 #ifdef INET6
282 	if (in6p) {
283 		rt = in6_pcbrtentry(in6p);
284 		so = in6p->in6p_socket;
285 	}
286 #endif
287 	if (rt == NULL) {
288 		size = tcp_mssdflt;
289 		goto out;
290 	}
291 
292 	ifp = rt->rt_ifp;
293 
294 	size = tcp_mssdflt;
295 	if (tp->t_mtudisc && rt->rt_rmx.rmx_mtu != 0) {
296 #ifdef INET6
297 		if (in6p && rt->rt_rmx.rmx_mtu < IPV6_MMTU) {
298 			/*
299 			 * RFC2460 section 5, last paragraph: if path MTU is
300 			 * smaller than 1280, use 1280 as packet size and
301 			 * attach fragment header.
302 			 */
303 			size = IPV6_MMTU - iphlen - sizeof(struct ip6_frag) -
304 			    sizeof(struct tcphdr);
305 		} else
306 			size = rt->rt_rmx.rmx_mtu - iphlen -
307 			    sizeof(struct tcphdr);
308 #else
309 		size = rt->rt_rmx.rmx_mtu - iphlen - sizeof(struct tcphdr);
310 #endif
311 	} else if (ifp->if_flags & IFF_LOOPBACK)
312 		size = ifp->if_mtu - iphlen - sizeof(struct tcphdr);
313 #ifdef INET
314 	else if (inp && tp->t_mtudisc)
315 		size = ifp->if_mtu - iphlen - sizeof(struct tcphdr);
316 	else if (inp && in_localaddr(inp->inp_faddr))
317 		size = ifp->if_mtu - iphlen - sizeof(struct tcphdr);
318 #endif
319 #ifdef INET6
320 	else if (in6p) {
321 #ifdef INET
322 		if (IN6_IS_ADDR_V4MAPPED(&in6p->in6p_faddr)) {
323 			/* mapped addr case */
324 			struct in_addr d;
325 			bcopy(&in6p->in6p_faddr.s6_addr32[3], &d, sizeof(d));
326 			if (tp->t_mtudisc || in_localaddr(d))
327 				size = ifp->if_mtu - iphlen - sizeof(struct tcphdr);
328 		} else
329 #endif
330 		{
331 			/*
332 			 * for IPv6, path MTU discovery is always turned on,
333 			 * or the node must use packet size <= 1280.
334 			 */
335 			size = tp->t_mtudisc ? IN6_LINKMTU(ifp) : IPV6_MMTU;
336 			size -= (iphlen + sizeof(struct tcphdr));
337 		}
338 	}
339 #endif
340  out:
341 	/*
342 	 * Now we must make room for whatever extra TCP/IP options are in
343 	 * the packet.
344 	 */
345 	optlen = tcp_optlen(tp);
346 
347 	/*
348 	 * XXX tp->t_ourmss should have the right size, but without this code
349 	 * fragmentation will occur... need more investigation
350 	 */
351 #ifdef INET
352 	if (inp) {
353 #if defined(IPSEC) || defined(FAST_IPSEC)
354 		if (! IPSEC_PCB_SKIP_IPSEC(inp->inp_sp, IPSEC_DIR_OUTBOUND))
355 			optlen += ipsec4_hdrsiz_tcp(tp);
356 #endif
357 		optlen += ip_optlen(inp);
358 	}
359 #endif
360 #ifdef INET6
361 #ifdef INET
362 	if (in6p && tp->t_family == AF_INET) {
363 #if defined(IPSEC) || defined(FAST_IPSEC)
364 		if (! IPSEC_PCB_SKIP_IPSEC(in6p->in6p_sp, IPSEC_DIR_OUTBOUND))
365 			optlen += ipsec4_hdrsiz_tcp(tp);
366 #endif
367 		/* XXX size -= ip_optlen(in6p); */
368 	} else
369 #endif
370 	if (in6p && tp->t_family == AF_INET6) {
371 #ifdef IPSEC
372 		if (! IPSEC_PCB_SKIP_IPSEC(in6p->in6p_sp, IPSEC_DIR_OUTBOUND))
373 			optlen += ipsec6_hdrsiz_tcp(tp);
374 #endif
375 		optlen += ip6_optlen(in6p);
376 	}
377 #endif
378 	size -= optlen;
379 
380 	/* there may not be any room for data if mtu is too small */
381 	if (size < 0)
382 		return (EMSGSIZE);
383 
384 	/*
385 	 * *rxsegsizep holds *estimated* inbound segment size (estimation
386 	 * assumes that path MTU is the same for both ways).  this is only
387 	 * for silly window avoidance, do not use the value for other purposes.
388 	 *
389 	 * ipseclen is subtracted from both sides, this may not be right.
390 	 * I'm not quite sure about this (could someone comment).
391 	 */
392 	*txsegsizep = min(tp->t_peermss - optlen, size);
393 	/*
394 	 * Never send more than half a buffer full.  This insures that we can
395 	 * always keep 2 packets on the wire, no matter what SO_SNDBUF is, and
396 	 * therefore acks will never be delayed unless we run out of data to
397 	 * transmit.
398 	 */
399 	if (so)
400 		*txsegsizep = min(so->so_snd.sb_hiwat >> 1, *txsegsizep);
401 	*rxsegsizep = min(tp->t_ourmss - optlen, size);
402 
403 	if (*txsegsizep != tp->t_segsz) {
404 		/*
405 		 * If the new segment size is larger, we don't want to
406 		 * mess up the congestion window, but if it is smaller
407 		 * we'll have to reduce the congestion window to ensure
408 		 * that we don't get into trouble with initial windows
409 		 * and the rest.  In any case, if the segment size
410 		 * has changed, chances are the path has, too, and
411 		 * our congestion window will be different.
412 		 */
413 		if (*txsegsizep < tp->t_segsz) {
414 			tp->snd_cwnd = max((tp->snd_cwnd / tp->t_segsz)
415 					   * *txsegsizep, *txsegsizep);
416 			tp->snd_ssthresh = max((tp->snd_ssthresh / tp->t_segsz)
417 						* *txsegsizep, *txsegsizep);
418 		}
419 		tp->t_segsz = *txsegsizep;
420 	}
421 
422 	return (0);
423 }
424 
425 static
426 #ifndef GPROF
427 __inline
428 #endif
429 int
430 tcp_build_datapkt(struct tcpcb *tp, struct socket *so, int off,
431     long len, int hdrlen, struct mbuf **mp)
432 {
433 	struct mbuf *m, *m0;
434 
435 	if (tp->t_force && len == 1)
436 		tcpstat.tcps_sndprobe++;
437 	else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
438 		tcpstat.tcps_sndrexmitpack++;
439 		tcpstat.tcps_sndrexmitbyte += len;
440 	} else {
441 		tcpstat.tcps_sndpack++;
442 		tcpstat.tcps_sndbyte += len;
443 	}
444 #ifdef notyet
445 	if ((m = m_copypack(so->so_snd.sb_mb, off,
446 	    (int)len, max_linkhdr + hdrlen)) == 0)
447 		return (ENOBUFS);
448 	/*
449 	 * m_copypack left space for our hdr; use it.
450 	 */
451 	m->m_len += hdrlen;
452 	m->m_data -= hdrlen;
453 #else
454 	MGETHDR(m, M_DONTWAIT, MT_HEADER);
455 	if (__predict_false(m == NULL))
456 		return (ENOBUFS);
457 	MCLAIM(m, &tcp_tx_mowner);
458 
459 	/*
460 	 * XXX Because other code assumes headers will fit in
461 	 * XXX one header mbuf.
462 	 *
463 	 * (This code should almost *never* be run.)
464 	 */
465 	if (__predict_false((max_linkhdr + hdrlen) > MHLEN)) {
466 		TCP_OUTPUT_COUNTER_INCR(&tcp_output_bigheader);
467 		MCLGET(m, M_DONTWAIT);
468 		if ((m->m_flags & M_EXT) == 0) {
469 			m_freem(m);
470 			return (ENOBUFS);
471 		}
472 	}
473 
474 	m->m_data += max_linkhdr;
475 	m->m_len = hdrlen;
476 
477 	/*
478 	 * To avoid traversing the whole sb_mb chain for correct
479 	 * data to send, remember last sent mbuf, its offset and
480 	 * the sent size.  When called the next time, see if the
481 	 * data to send is directly following the previous transfer.
482 	 * This is important for large TCP windows.
483 	 */
484 	if (off == 0 || tp->t_lastm == NULL ||
485 	    (tp->t_lastoff + tp->t_lastlen) != off) {
486 		TCP_OUTPUT_COUNTER_INCR(&tcp_output_predict_miss);
487 		/*
488 		 * Either a new packet or a retransmit.
489 		 * Start from the beginning.
490 		 */
491 		tp->t_lastm = so->so_snd.sb_mb;
492 		tp->t_inoff = off;
493 	} else {
494 		TCP_OUTPUT_COUNTER_INCR(&tcp_output_predict_hit);
495 		tp->t_inoff += tp->t_lastlen;
496 	}
497 
498 	/* Traverse forward to next packet */
499 	while (tp->t_inoff > 0) {
500 		if (tp->t_lastm == NULL)
501 			panic("tp->t_lastm == NULL");
502 		if (tp->t_inoff < tp->t_lastm->m_len)
503 			break;
504 		tp->t_inoff -= tp->t_lastm->m_len;
505 		tp->t_lastm = tp->t_lastm->m_next;
506 	}
507 
508 	tp->t_lastoff = off;
509 	tp->t_lastlen = len;
510 	m0 = tp->t_lastm;
511 	off = tp->t_inoff;
512 
513 	if (len <= M_TRAILINGSPACE(m)) {
514 		m_copydata(m0, off, (int) len, mtod(m, caddr_t) + hdrlen);
515 		m->m_len += len;
516 		TCP_OUTPUT_COUNTER_INCR(&tcp_output_copysmall);
517 	} else {
518 		m->m_next = m_copy(m0, off, (int) len);
519 		if (m->m_next == NULL) {
520 			m_freem(m);
521 			return (ENOBUFS);
522 		}
523 #ifdef TCP_OUTPUT_COUNTERS
524 		if (m->m_next->m_flags & M_EXT)
525 			TCP_OUTPUT_COUNTER_INCR(&tcp_output_refbig);
526 		else
527 			TCP_OUTPUT_COUNTER_INCR(&tcp_output_copybig);
528 #endif /* TCP_OUTPUT_COUNTERS */
529 	}
530 #endif
531 
532 	*mp = m;
533 	return (0);
534 }
535 
536 /*
537  * Tcp output routine: figure out what should be sent and send it.
538  */
539 int
540 tcp_output(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 	unsigned int sack_numblks;
555 	int idle, sendalot, txsegsize, rxsegsize;
556 	int txsegsize_nosack;
557 	int maxburst = TCP_MAXBURST;
558 	int af;		/* address family on the wire */
559 	int iphdrlen;
560 	int has_tso, use_tso;
561 	int sack_rxmit;
562 	int sack_bytes_rxmt;
563 	struct sackhole *p;
564 #ifdef TCP_SIGNATURE
565 	int sigoff = 0;
566 #endif
567 
568 #ifdef DIAGNOSTIC
569 	if (tp->t_inpcb && tp->t_in6pcb)
570 		panic("tcp_output: both t_inpcb and t_in6pcb are set");
571 #endif
572 	so = NULL;
573 	ro = NULL;
574 	if (tp->t_inpcb) {
575 		so = tp->t_inpcb->inp_socket;
576 		ro = &tp->t_inpcb->inp_route;
577 	}
578 #ifdef INET6
579 	else if (tp->t_in6pcb) {
580 		so = tp->t_in6pcb->in6p_socket;
581 		ro = (struct route *)&tp->t_in6pcb->in6p_route;
582 	}
583 #endif
584 
585 	switch (af = tp->t_family) {
586 #ifdef INET
587 	case AF_INET:
588 		if (tp->t_inpcb)
589 			break;
590 #ifdef INET6
591 		/* mapped addr case */
592 		if (tp->t_in6pcb)
593 			break;
594 #endif
595 		return (EINVAL);
596 #endif
597 #ifdef INET6
598 	case AF_INET6:
599 		if (tp->t_in6pcb)
600 			break;
601 		return (EINVAL);
602 #endif
603 	default:
604 		return (EAFNOSUPPORT);
605 	}
606 
607 	if (tcp_segsize(tp, &txsegsize, &rxsegsize))
608 		return (EMSGSIZE);
609 
610 	idle = (tp->snd_max == tp->snd_una);
611 
612 	/*
613 	 * Determine if we can use TCP segmentation offload:
614 	 * - If we're using IPv4
615 	 * - If there is not an IPsec policy that prevents it
616 	 * - If the interface can do it
617 	 */
618 	has_tso = tp->t_inpcb != NULL &&
619 #if defined(IPSEC) || defined(FAST_IPSEC)
620 		  IPSEC_PCB_SKIP_IPSEC(tp->t_inpcb->inp_sp,
621 		  		       IPSEC_DIR_OUTBOUND) &&
622 #endif
623 		  tp->t_inpcb->inp_route.ro_rt != NULL &&
624 		  (tp->t_inpcb->inp_route.ro_rt->rt_ifp->if_capenable &
625 		   IFCAP_TSOv4) != 0;
626 
627 	/*
628 	 * Restart Window computation.  From draft-floyd-incr-init-win-03:
629 	 *
630 	 *	Optionally, a TCP MAY set the restart window to the
631 	 *	minimum of the value used for the initial window and
632 	 *	the current value of cwnd (in other words, using a
633 	 *	larger value for the restart window should never increase
634 	 *	the size of cwnd).
635 	 */
636 	if (tcp_cwm) {
637 		/*
638 		 * Hughes/Touch/Heidemann Congestion Window Monitoring.
639 		 * Count the number of packets currently pending
640 		 * acknowledgement, and limit our congestion window
641 		 * to a pre-determined allowed burst size plus that count.
642 		 * This prevents bursting once all pending packets have
643 		 * been acknowledged (i.e. transmission is idle).
644 		 *
645 		 * XXX Link this to Initial Window?
646 		 */
647 		tp->snd_cwnd = min(tp->snd_cwnd,
648 		    (tcp_cwm_burstsize * txsegsize) +
649 		    (tp->snd_nxt - tp->snd_una));
650 	} else {
651 		if (idle && (tcp_now - tp->t_rcvtime) >= tp->t_rxtcur) {
652 			/*
653 			 * We have been idle for "a while" and no acks are
654 			 * expected to clock out any data we send --
655 			 * slow start to get ack "clock" running again.
656 			 */
657 			int ss = tcp_init_win;
658 #ifdef INET
659 			if (tp->t_inpcb &&
660 			    in_localaddr(tp->t_inpcb->inp_faddr))
661 				ss = tcp_init_win_local;
662 #endif
663 #ifdef INET6
664 			if (tp->t_in6pcb &&
665 			    in6_localaddr(&tp->t_in6pcb->in6p_faddr))
666 				ss = tcp_init_win_local;
667 #endif
668 			tp->snd_cwnd = min(tp->snd_cwnd,
669 			    TCP_INITIAL_WINDOW(ss, txsegsize));
670 		}
671 	}
672 
673 	txsegsize_nosack = txsegsize;
674 again:
675 	use_tso = has_tso;
676 	TCP_REASS_LOCK(tp);
677 	sack_numblks = tcp_sack_numblks(tp);
678 	if (sack_numblks) {
679 		if ((tp->rcv_sack_flags & TCPSACK_HAVED) != 0) {
680 			/* don't duplicate D-SACK. */
681 			use_tso = 0;
682 		}
683 		txsegsize = txsegsize_nosack - TCP_SACK_OPTLEN(sack_numblks);
684 	} else {
685 		txsegsize = txsegsize_nosack;
686 	}
687 
688 	/*
689 	 * Determine length of data that should be transmitted, and
690 	 * flags that should be used.  If there is some data or critical
691 	 * controls (SYN, RST) to send, then transmit; otherwise,
692 	 * investigate further.
693 	 *
694 	 * Readjust SACK information to avoid resending duplicate data.
695 	 */
696 	if (TCP_SACK_ENABLED(tp) && SEQ_LT(tp->snd_nxt, tp->snd_max))
697 		tcp_sack_adjust(tp);
698 	sendalot = 0;
699 	off = tp->snd_nxt - tp->snd_una;
700 	win = min(tp->snd_wnd, tp->snd_cwnd);
701 
702 	flags = tcp_outflags[tp->t_state];
703 
704 	/*
705 	 * Send any SACK-generated retransmissions.  If we're explicitly trying
706 	 * to send out new data (when sendalot is 1), bypass this function.
707 	 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
708 	 * we're replacing a (future) new transmission with a retransmission
709 	 * now, and we previously incremented snd_cwnd in tcp_input().
710 	 */
711 	/*
712 	 * Still in sack recovery , reset rxmit flag to zero.
713 	 */
714 	sack_rxmit = 0;
715 	sack_bytes_rxmt = 0;
716 	len = 0;
717 	p = NULL;
718 	do {
719 		long cwin;
720 		if (!TCP_SACK_ENABLED(tp))
721 			break;
722 		if (tp->t_partialacks < 0)
723 			break;
724 		p = tcp_sack_output(tp, &sack_bytes_rxmt);
725 		if (p == NULL)
726 			break;
727 
728 		cwin = min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt;
729 		if (cwin < 0)
730 			cwin = 0;
731 		/* Do not retransmit SACK segments beyond snd_recover */
732 		if (SEQ_GT(p->end, tp->snd_recover)) {
733 			/*
734 			 * (At least) part of sack hole extends beyond
735 			 * snd_recover. Check to see if we can rexmit data
736 			 * for this hole.
737 			 */
738 			if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
739 				/*
740 				 * Can't rexmit any more data for this hole.
741 				 * That data will be rexmitted in the next
742 				 * sack recovery episode, when snd_recover
743 				 * moves past p->rxmit.
744 				 */
745 				p = NULL;
746 				break;
747 			}
748 			/* Can rexmit part of the current hole */
749 			len = ((long)ulmin(cwin, tp->snd_recover - p->rxmit));
750 		} else
751 			len = ((long)ulmin(cwin, p->end - p->rxmit));
752 		off = p->rxmit - tp->snd_una;
753 		if (len > 0) {
754 			sack_rxmit = 1;
755 			sendalot = 1;
756 		}
757 	} while (/*CONSTCOND*/0);
758 
759 	/*
760 	 * If in persist timeout with window of 0, send 1 byte.
761 	 * Otherwise, if window is small but nonzero
762 	 * and timer expired, we will send what we can
763 	 * and go to transmit state.
764 	 */
765 	if (tp->t_force) {
766 		if (win == 0) {
767 			/*
768 			 * If we still have some data to send, then
769 			 * clear the FIN bit.  Usually this would
770 			 * happen below when it realizes that we
771 			 * aren't sending all the data.  However,
772 			 * if we have exactly 1 byte of unset data,
773 			 * then it won't clear the FIN bit below,
774 			 * and if we are in persist state, we wind
775 			 * up sending the packet without recording
776 			 * that we sent the FIN bit.
777 			 *
778 			 * We can't just blindly clear the FIN bit,
779 			 * because if we don't have any more data
780 			 * to send then the probe will be the FIN
781 			 * itself.
782 			 */
783 			if (off < so->so_snd.sb_cc)
784 				flags &= ~TH_FIN;
785 			win = 1;
786 		} else {
787 			TCP_TIMER_DISARM(tp, TCPT_PERSIST);
788 			tp->t_rxtshift = 0;
789 		}
790 	}
791 
792 	if (!TCP_SACK_ENABLED(tp)) {
793 		if (win < so->so_snd.sb_cc) {
794 			len = win - off;
795 			flags &= ~TH_FIN;
796 		} else
797 			len = so->so_snd.sb_cc - off;
798 	} else if (sack_rxmit == 0) {
799 		if (sack_bytes_rxmt != 0) {
800 			long cwin;
801 
802 			/*
803 			 * We are inside of a SACK recovery episode and are
804 			 * sending new data, having retransmitted all the
805 			 * data possible in the scoreboard.
806 			 */
807 			len = ((long)ulmin(so->so_snd.sb_cc, tp->snd_wnd)
808 				       - off);
809 			/*
810 			 * From FreeBSD:
811 			 *  Don't remove this (len > 0) check !
812 			 *  We explicitly check for len > 0 here (although it
813 			 *  isn't really necessary), to work around a gcc
814 			 *  optimization issue - to force gcc to compute
815 			 *  len above. Without this check, the computation
816 			 *  of len is bungled by the optimizer.
817 			 */
818 			if (len > 0) {
819 				cwin = tp->snd_cwnd -
820 						(tp->snd_nxt - tp->sack_newdata) -
821 						sack_bytes_rxmt;
822 				if (cwin < 0)
823 					cwin = 0;
824 				len = lmin(len, cwin);
825 			}
826 		} else if (win < so->so_snd.sb_cc) {
827 			len = win - off;
828 			flags &= ~TH_FIN;
829 		} else
830 			len = so->so_snd.sb_cc - off;
831 	}
832 
833 	if (len < 0) {
834 		/*
835 		 * If FIN has been sent but not acked,
836 		 * but we haven't been called to retransmit,
837 		 * len will be -1.  Otherwise, window shrank
838 		 * after we sent into it.  If window shrank to 0,
839 		 * cancel pending retransmit, pull snd_nxt back
840 		 * to (closed) window, and set the persist timer
841 		 * if it isn't already going.  If the window didn't
842 		 * close completely, just wait for an ACK.
843 		 *
844 		 * If we have a pending FIN, either it has already been
845 		 * transmitted or it is outside the window, so drop it.
846 		 * If the FIN has been transmitted, but this is not a
847 		 * retransmission, then len must be -1.  Therefore we also
848 		 * prevent here the sending of `gratuitous FINs'.  This
849 		 * eliminates the need to check for that case below (e.g.
850 		 * to back up snd_nxt before the FIN so that the sequence
851 		 * number is correct).
852 		 */
853 		len = 0;
854 		flags &= ~TH_FIN;
855 		if (win == 0) {
856 			TCP_TIMER_DISARM(tp, TCPT_REXMT);
857 			tp->t_rxtshift = 0;
858 			tp->snd_nxt = tp->snd_una;
859 			if (TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0)
860 				tcp_setpersist(tp);
861 		}
862 	}
863 	if (len > txsegsize) {
864 		if (use_tso) {
865 			/*
866 			 * Truncate TSO transfers to IP_MAXPACKET, and make
867 			 * sure that we send equal size transfers down the
868 			 * stack (rather than big-small-big-small-...).
869 			 */
870 			len = (min(len, IP_MAXPACKET) / txsegsize) * txsegsize;
871 			if (len <= txsegsize) {
872 				use_tso = 0;
873 			}
874 		} else
875 			len = txsegsize;
876 		flags &= ~TH_FIN;
877 		sendalot = 1;
878 	} else
879 		use_tso = 0;
880 	if (sack_rxmit) {
881 		if (SEQ_LT(p->rxmit + len, tp->snd_una + so->so_snd.sb_cc))
882 			flags &= ~TH_FIN;
883 	}
884 
885 	win = sbspace(&so->so_rcv);
886 
887 	/*
888 	 * Sender silly window avoidance.  If connection is idle
889 	 * and can send all data, a maximum segment,
890 	 * at least a maximum default-size segment do it,
891 	 * or are forced, do it; otherwise don't bother.
892 	 * If peer's buffer is tiny, then send
893 	 * when window is at least half open.
894 	 * If retransmitting (possibly after persist timer forced us
895 	 * to send into a small window), then must resend.
896 	 */
897 	if (len) {
898 		if (len >= txsegsize)
899 			goto send;
900 		if ((so->so_state & SS_MORETOCOME) == 0 &&
901 		    ((idle || tp->t_flags & TF_NODELAY) &&
902 		     len + off >= so->so_snd.sb_cc))
903 			goto send;
904 		if (tp->t_force)
905 			goto send;
906 		if (len >= tp->max_sndwnd / 2)
907 			goto send;
908 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))
909 			goto send;
910 		if (sack_rxmit)
911 			goto send;
912 	}
913 
914 	/*
915 	 * Compare available window to amount of window known to peer
916 	 * (as advertised window less next expected input).  If the
917 	 * difference is at least twice the size of the largest segment
918 	 * we expect to receive (i.e. two segments) or at least 50% of
919 	 * the maximum possible window, then want to send a window update
920 	 * to peer.
921 	 */
922 	if (win > 0) {
923 		/*
924 		 * "adv" is the amount we can increase the window,
925 		 * taking into account that we are limited by
926 		 * TCP_MAXWIN << tp->rcv_scale.
927 		 */
928 		long adv = min(win, (long)TCP_MAXWIN << tp->rcv_scale) -
929 			(tp->rcv_adv - tp->rcv_nxt);
930 
931 		if (adv >= (long) (2 * rxsegsize))
932 			goto send;
933 		if (2 * adv >= (long) so->so_rcv.sb_hiwat)
934 			goto send;
935 	}
936 
937 	/*
938 	 * Send if we owe peer an ACK.
939 	 */
940 	if (tp->t_flags & TF_ACKNOW)
941 		goto send;
942 	if (flags & (TH_SYN|TH_FIN|TH_RST))
943 		goto send;
944 	if (SEQ_GT(tp->snd_up, tp->snd_una))
945 		goto send;
946 	/*
947 	 * In SACK, it is possible for tcp_output to fail to send a segment
948 	 * after the retransmission timer has been turned off.  Make sure
949 	 * that the retransmission timer is set.
950 	 */
951 	if (TCP_SACK_ENABLED(tp) && SEQ_GT(tp->snd_max, tp->snd_una) &&
952 	    !TCP_TIMER_ISARMED(tp, TCPT_REXMT) &&
953 	    !TCP_TIMER_ISARMED(tp, TCPT_PERSIST)) {
954 		TCP_TIMER_ARM(tp, TCPT_REXMT, tp->t_rxtcur);
955 		goto just_return;
956 	}
957 
958 	/*
959 	 * TCP window updates are not reliable, rather a polling protocol
960 	 * using ``persist'' packets is used to insure receipt of window
961 	 * updates.  The three ``states'' for the output side are:
962 	 *	idle			not doing retransmits or persists
963 	 *	persisting		to move a small or zero window
964 	 *	(re)transmitting	and thereby not persisting
965 	 *
966 	 * tp->t_timer[TCPT_PERSIST]
967 	 *	is set when we are in persist state.
968 	 * tp->t_force
969 	 *	is set when we are called to send a persist packet.
970 	 * tp->t_timer[TCPT_REXMT]
971 	 *	is set when we are retransmitting
972 	 * The output side is idle when both timers are zero.
973 	 *
974 	 * If send window is too small, there is data to transmit, and no
975 	 * retransmit or persist is pending, then go to persist state.
976 	 * If nothing happens soon, send when timer expires:
977 	 * if window is nonzero, transmit what we can,
978 	 * otherwise force out a byte.
979 	 */
980 	if (so->so_snd.sb_cc && TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 &&
981 	    TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) {
982 		tp->t_rxtshift = 0;
983 		tcp_setpersist(tp);
984 	}
985 
986 	/*
987 	 * No reason to send a segment, just return.
988 	 */
989 just_return:
990 	TCP_REASS_UNLOCK(tp);
991 	return (0);
992 
993 send:
994 	/*
995 	 * Before ESTABLISHED, force sending of initial options
996 	 * unless TCP set not to do any options.
997 	 * NOTE: we assume that the IP/TCP header plus TCP options
998 	 * always fit in a single mbuf, leaving room for a maximum
999 	 * link header, i.e.
1000 	 *	max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
1001 	 */
1002 	optlen = 0;
1003 	switch (af) {
1004 #ifdef INET
1005 	case AF_INET:
1006 		iphdrlen = sizeof(struct ip) + sizeof(struct tcphdr);
1007 		break;
1008 #endif
1009 #ifdef INET6
1010 	case AF_INET6:
1011 		iphdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
1012 		break;
1013 #endif
1014 	default:	/*pacify gcc*/
1015 		iphdrlen = 0;
1016 		break;
1017 	}
1018 	hdrlen = iphdrlen;
1019 	if (flags & TH_SYN) {
1020 		struct rtentry *rt;
1021 
1022 		rt = NULL;
1023 #ifdef INET
1024 		if (tp->t_inpcb)
1025 			rt = in_pcbrtentry(tp->t_inpcb);
1026 #endif
1027 #ifdef INET6
1028 		if (tp->t_in6pcb)
1029 			rt = in6_pcbrtentry(tp->t_in6pcb);
1030 #endif
1031 
1032 		tp->snd_nxt = tp->iss;
1033 		tp->t_ourmss = tcp_mss_to_advertise(rt != NULL ?
1034 						    rt->rt_ifp : NULL, af);
1035 		if ((tp->t_flags & TF_NOOPT) == 0) {
1036 			opt[0] = TCPOPT_MAXSEG;
1037 			opt[1] = 4;
1038 			opt[2] = (tp->t_ourmss >> 8) & 0xff;
1039 			opt[3] = tp->t_ourmss & 0xff;
1040 			optlen = 4;
1041 
1042 			if ((tp->t_flags & TF_REQ_SCALE) &&
1043 			    ((flags & TH_ACK) == 0 ||
1044 			    (tp->t_flags & TF_RCVD_SCALE))) {
1045 				*((u_int32_t *) (opt + optlen)) = htonl(
1046 					TCPOPT_NOP << 24 |
1047 					TCPOPT_WINDOW << 16 |
1048 					TCPOLEN_WINDOW << 8 |
1049 					tp->request_r_scale);
1050 				optlen += 4;
1051 			}
1052 			if (tcp_do_sack) {
1053 				u_int8_t *p = (u_int8_t *)(opt + optlen);
1054 
1055 				p[0] = TCPOPT_SACK_PERMITTED;
1056 				p[1] = 2;
1057 				p[2] = TCPOPT_NOP;
1058 				p[3] = TCPOPT_NOP;
1059 				optlen += 4;
1060 			}
1061 		}
1062 	}
1063 
1064 	/*
1065 	 * Send a timestamp and echo-reply if this is a SYN and our side
1066 	 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side
1067 	 * and our peer have sent timestamps in our SYN's.
1068 	 */
1069 	if ((tp->t_flags & (TF_REQ_TSTMP|TF_NOOPT)) == TF_REQ_TSTMP &&
1070 	     (flags & TH_RST) == 0 &&
1071 	    ((flags & (TH_SYN|TH_ACK)) == TH_SYN ||
1072 	     (tp->t_flags & TF_RCVD_TSTMP))) {
1073 		u_int32_t *lp = (u_int32_t *)(opt + optlen);
1074 
1075 		/* Form timestamp option as shown in appendix A of RFC 1323. */
1076 		*lp++ = htonl(TCPOPT_TSTAMP_HDR);
1077 		*lp++ = htonl(TCP_TIMESTAMP(tp));
1078 		*lp   = htonl(tp->ts_recent);
1079 		optlen += TCPOLEN_TSTAMP_APPA;
1080 	}
1081 
1082 	/*
1083 	 * Tack on the SACK block if it is necessary.
1084 	 */
1085 	if (sack_numblks) {
1086 		int sack_len;
1087 		u_char *bp = (u_char *)(opt + optlen);
1088 		u_int32_t *lp = (u_int32_t *)(bp + 4);
1089 		struct ipqent *tiqe;
1090 
1091 		sack_len = sack_numblks * 8 + 2;
1092 		bp[0] = TCPOPT_NOP;
1093 		bp[1] = TCPOPT_NOP;
1094 		bp[2] = TCPOPT_SACK;
1095 		bp[3] = sack_len;
1096 		if ((tp->rcv_sack_flags & TCPSACK_HAVED) != 0) {
1097 			sack_numblks--;
1098 			*lp++ = htonl(tp->rcv_dsack_block.left);
1099 			*lp++ = htonl(tp->rcv_dsack_block.right);
1100 			tp->rcv_sack_flags &= ~TCPSACK_HAVED;
1101 		}
1102 		for (tiqe = TAILQ_FIRST(&tp->timeq);
1103 		    sack_numblks > 0; tiqe = TAILQ_NEXT(tiqe, ipqe_timeq)) {
1104 			KASSERT(tiqe != NULL);
1105 			sack_numblks--;
1106 			*lp++ = htonl(tiqe->ipqe_seq);
1107 			*lp++ = htonl(tiqe->ipqe_seq + tiqe->ipqe_len);
1108 		}
1109 		optlen += sack_len + 2;
1110 	}
1111 	TCP_REASS_UNLOCK(tp);
1112 
1113 #ifdef TCP_SIGNATURE
1114 #if defined(INET6) && defined(FAST_IPSEC)
1115 	if (tp->t_family == AF_INET)
1116 #endif
1117 	if (tp->t_flags & TF_SIGNATURE) {
1118 		u_char *bp;
1119 		/*
1120 		 * Initialize TCP-MD5 option (RFC2385)
1121 		 */
1122 		bp = (u_char *)opt + optlen;
1123 		*bp++ = TCPOPT_SIGNATURE;
1124 		*bp++ = TCPOLEN_SIGNATURE;
1125 		sigoff = optlen + 2;
1126 		bzero(bp, TCP_SIGLEN);
1127 		bp += TCP_SIGLEN;
1128 		optlen += TCPOLEN_SIGNATURE;
1129 		/*
1130 		 * Terminate options list and maintain 32-bit alignment.
1131  		 */
1132 		*bp++ = TCPOPT_NOP;
1133 		*bp++ = TCPOPT_EOL;
1134  		optlen += 2;
1135  	}
1136 #endif /* TCP_SIGNATURE */
1137 
1138 	hdrlen += optlen;
1139 
1140 #ifdef DIAGNOSTIC
1141 	if (!use_tso && len > txsegsize)
1142 		panic("tcp data to be sent is larger than segment");
1143 	else if (use_tso && len > IP_MAXPACKET)
1144 		panic("tcp data to be sent is larger than max TSO size");
1145 	if (max_linkhdr + hdrlen > MCLBYTES)
1146 		panic("tcphdr too big");
1147 #endif
1148 
1149 	/*
1150 	 * Grab a header mbuf, attaching a copy of data to
1151 	 * be transmitted, and initialize the header from
1152 	 * the template for sends on this connection.
1153 	 */
1154 	if (len) {
1155 		error = tcp_build_datapkt(tp, so, off, len, hdrlen, &m);
1156 		if (error)
1157 			goto out;
1158 		/*
1159 		 * If we're sending everything we've got, set PUSH.
1160 		 * (This will keep happy those implementations which only
1161 		 * give data to the user when a buffer fills or
1162 		 * a PUSH comes in.)
1163 		 */
1164 		if (off + len == so->so_snd.sb_cc)
1165 			flags |= TH_PUSH;
1166 	} else {
1167 		if (tp->t_flags & TF_ACKNOW)
1168 			tcpstat.tcps_sndacks++;
1169 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
1170 			tcpstat.tcps_sndctrl++;
1171 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
1172 			tcpstat.tcps_sndurg++;
1173 		else
1174 			tcpstat.tcps_sndwinup++;
1175 
1176 		MGETHDR(m, M_DONTWAIT, MT_HEADER);
1177 		if (m != NULL && max_linkhdr + hdrlen > MHLEN) {
1178 			MCLGET(m, M_DONTWAIT);
1179 			if ((m->m_flags & M_EXT) == 0) {
1180 				m_freem(m);
1181 				m = NULL;
1182 			}
1183 		}
1184 		if (m == NULL) {
1185 			error = ENOBUFS;
1186 			goto out;
1187 		}
1188 		MCLAIM(m, &tcp_tx_mowner);
1189 		m->m_data += max_linkhdr;
1190 		m->m_len = hdrlen;
1191 	}
1192 	m->m_pkthdr.rcvif = (struct ifnet *)0;
1193 	switch (af) {
1194 #ifdef INET
1195 	case AF_INET:
1196 		ip = mtod(m, struct ip *);
1197 #ifdef INET6
1198 		ip6 = NULL;
1199 #endif
1200 		th = (struct tcphdr *)(ip + 1);
1201 		break;
1202 #endif
1203 #ifdef INET6
1204 	case AF_INET6:
1205 		ip = NULL;
1206 		ip6 = mtod(m, struct ip6_hdr *);
1207 		th = (struct tcphdr *)(ip6 + 1);
1208 		break;
1209 #endif
1210 	default:	/*pacify gcc*/
1211 		ip = NULL;
1212 #ifdef INET6
1213 		ip6 = NULL;
1214 #endif
1215 		th = NULL;
1216 		break;
1217 	}
1218 	if (tp->t_template == 0)
1219 		panic("tcp_output");
1220 	if (tp->t_template->m_len < iphdrlen)
1221 		panic("tcp_output");
1222 	bcopy(mtod(tp->t_template, caddr_t), mtod(m, caddr_t), iphdrlen);
1223 
1224 	/*
1225 	 * If we are doing retransmissions, then snd_nxt will
1226 	 * not reflect the first unsent octet.  For ACK only
1227 	 * packets, we do not want the sequence number of the
1228 	 * retransmitted packet, we want the sequence number
1229 	 * of the next unsent octet.  So, if there is no data
1230 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
1231 	 * when filling in ti_seq.  But if we are in persist
1232 	 * state, snd_max might reflect one byte beyond the
1233 	 * right edge of the window, so use snd_nxt in that
1234 	 * case, since we know we aren't doing a retransmission.
1235 	 * (retransmit and persist are mutually exclusive...)
1236 	 */
1237 	if (TCP_SACK_ENABLED(tp) && sack_rxmit) {
1238 		th->th_seq = htonl(p->rxmit);
1239 		p->rxmit += len;
1240 	} else {
1241 		if (len || (flags & (TH_SYN|TH_FIN)) ||
1242 		    TCP_TIMER_ISARMED(tp, TCPT_PERSIST))
1243 			th->th_seq = htonl(tp->snd_nxt);
1244 		else
1245 			th->th_seq = htonl(tp->snd_max);
1246 	}
1247 	th->th_ack = htonl(tp->rcv_nxt);
1248 	if (optlen) {
1249 		bcopy((caddr_t)opt, (caddr_t)(th + 1), optlen);
1250 		th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
1251 	}
1252 	th->th_flags = flags;
1253 	/*
1254 	 * Calculate receive window.  Don't shrink window,
1255 	 * but avoid silly window syndrome.
1256 	 */
1257 	if (win < (long)(so->so_rcv.sb_hiwat / 4) && win < (long)rxsegsize)
1258 		win = 0;
1259 	if (win > (long)TCP_MAXWIN << tp->rcv_scale)
1260 		win = (long)TCP_MAXWIN << tp->rcv_scale;
1261 	if (win < (long)(int32_t)(tp->rcv_adv - tp->rcv_nxt))
1262 		win = (long)(int32_t)(tp->rcv_adv - tp->rcv_nxt);
1263 	th->th_win = htons((u_int16_t) (win>>tp->rcv_scale));
1264 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1265 		u_int32_t urp = tp->snd_up - tp->snd_nxt;
1266 		if (urp > IP_MAXPACKET)
1267 			urp = IP_MAXPACKET;
1268 		th->th_urp = htons((u_int16_t)urp);
1269 		th->th_flags |= TH_URG;
1270 	} else
1271 		/*
1272 		 * If no urgent pointer to send, then we pull
1273 		 * the urgent pointer to the left edge of the send window
1274 		 * so that it doesn't drift into the send window on sequence
1275 		 * number wraparound.
1276 		 */
1277 		tp->snd_up = tp->snd_una;		/* drag it along */
1278 
1279 #ifdef TCP_SIGNATURE
1280 #if defined(INET6) && defined(FAST_IPSEC)
1281 	if (tp->t_family == AF_INET) /* XXX */
1282 #endif
1283 	if (sigoff && (tp->t_flags & TF_SIGNATURE)) {
1284 		struct secasvar *sav;
1285 		u_int8_t *sigp;
1286 
1287 		sav = tcp_signature_getsav(m, th);
1288 
1289 		if (sav == NULL) {
1290 			if (m)
1291 				m_freem(m);
1292 			return (EPERM);
1293 		}
1294 
1295 		m->m_pkthdr.len = hdrlen + len;
1296 		sigp = (caddr_t)th + sizeof(*th) + sigoff;
1297 		tcp_signature(m, th, (caddr_t)th - mtod(m, caddr_t), sav, sigp);
1298 
1299 		key_sa_recordxfer(sav, m);
1300 #ifdef FAST_IPSEC
1301 		KEY_FREESAV(&sav);
1302 #else
1303 		key_freesav(sav);
1304 #endif
1305 	}
1306 #endif
1307 
1308 	/*
1309 	 * Set ourselves up to be checksummed just before the packet
1310 	 * hits the wire.
1311 	 */
1312 	switch (af) {
1313 #ifdef INET
1314 	case AF_INET:
1315 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1316 		if (use_tso) {
1317 			m->m_pkthdr.segsz = txsegsize;
1318 			m->m_pkthdr.csum_flags = M_CSUM_TSOv4;
1319 		} else {
1320 			m->m_pkthdr.csum_flags = M_CSUM_TCPv4;
1321 			if (len + optlen) {
1322 				/* Fixup the pseudo-header checksum. */
1323 				/* XXXJRT Not IP Jumbogram safe. */
1324 				th->th_sum = in_cksum_addword(th->th_sum,
1325 				    htons((u_int16_t) (len + optlen)));
1326 			}
1327 		}
1328 		break;
1329 #endif
1330 #ifdef INET6
1331 	case AF_INET6:
1332 		/*
1333 		 * XXX Actually delaying the checksum is Hard
1334 		 * XXX (well, maybe not for Itojun, but it is
1335 		 * XXX for me), but we can still take advantage
1336 		 * XXX of the cached pseudo-header checksum.
1337 		 */
1338 		/* equals to hdrlen + len */
1339 		m->m_pkthdr.len = sizeof(struct ip6_hdr)
1340 			+ sizeof(struct tcphdr) + optlen + len;
1341 #ifdef notyet
1342 		m->m_pkthdr.csum_flags = M_CSUM_TCPv6;
1343 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1344 #endif
1345 		if (len + optlen) {
1346 			/* Fixup the pseudo-header checksum. */
1347 			/* XXXJRT: Not IPv6 Jumbogram safe. */
1348 			th->th_sum = in_cksum_addword(th->th_sum,
1349 			    htons((u_int16_t) (len + optlen)));
1350 		}
1351 		th->th_sum = in6_cksum(m, 0, sizeof(struct ip6_hdr),
1352 		    sizeof(struct tcphdr) + optlen + len);
1353 		break;
1354 #endif
1355 	}
1356 
1357 	/*
1358 	 * In transmit state, time the transmission and arrange for
1359 	 * the retransmit.  In persist state, just set snd_max.
1360 	 */
1361 	if (tp->t_force == 0 || TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) {
1362 		tcp_seq startseq = tp->snd_nxt;
1363 
1364 		/*
1365 		 * Advance snd_nxt over sequence space of this segment.
1366 		 * There are no states in which we send both a SYN and a FIN,
1367 		 * so we collapse the tests for these flags.
1368 		 */
1369 		if (flags & (TH_SYN|TH_FIN))
1370 			tp->snd_nxt++;
1371 		if (sack_rxmit)
1372 			goto timer;
1373 		tp->snd_nxt += len;
1374 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1375 			tp->snd_max = tp->snd_nxt;
1376 			/*
1377 			 * Time this transmission if not a retransmission and
1378 			 * not currently timing anything.
1379 			 */
1380 			if (tp->t_rtttime == 0) {
1381 				tp->t_rtttime = tcp_now;
1382 				tp->t_rtseq = startseq;
1383 				tcpstat.tcps_segstimed++;
1384 			}
1385 		}
1386 
1387 		/*
1388 		 * Set retransmit timer if not currently set,
1389 		 * and not doing an ack or a keep-alive probe.
1390 		 * Initial value for retransmit timer is smoothed
1391 		 * round-trip time + 2 * round-trip time variance.
1392 		 * Initialize shift counter which is used for backoff
1393 		 * of retransmit time.
1394 		 */
1395 timer:
1396 		if (TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 &&
1397 			((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1398 		    tp->snd_nxt != tp->snd_una)) {
1399 			if (TCP_TIMER_ISARMED(tp, TCPT_PERSIST)) {
1400 				TCP_TIMER_DISARM(tp, TCPT_PERSIST);
1401 				tp->t_rxtshift = 0;
1402 			}
1403 			TCP_TIMER_ARM(tp, TCPT_REXMT, tp->t_rxtcur);
1404 		}
1405 	} else
1406 		if (SEQ_GT(tp->snd_nxt + len, tp->snd_max))
1407 			tp->snd_max = tp->snd_nxt + len;
1408 
1409 #ifdef TCP_DEBUG
1410 	/*
1411 	 * Trace.
1412 	 */
1413 	if (so->so_options & SO_DEBUG)
1414 		tcp_trace(TA_OUTPUT, tp->t_state, tp, m, 0);
1415 #endif
1416 
1417 	/*
1418 	 * Fill in IP length and desired time to live and
1419 	 * send to IP level.  There should be a better way
1420 	 * to handle ttl and tos; we could keep them in
1421 	 * the template, but need a way to checksum without them.
1422 	 */
1423 	m->m_pkthdr.len = hdrlen + len;
1424 
1425 	switch (af) {
1426 #ifdef INET
1427 	case AF_INET:
1428 		ip->ip_len = htons(m->m_pkthdr.len);
1429 		if (tp->t_inpcb) {
1430 			ip->ip_ttl = tp->t_inpcb->inp_ip.ip_ttl;
1431 			ip->ip_tos = tp->t_inpcb->inp_ip.ip_tos;
1432 		}
1433 #ifdef INET6
1434 		else if (tp->t_in6pcb) {
1435 			ip->ip_ttl = in6_selecthlim(tp->t_in6pcb, NULL); /*XXX*/
1436 			ip->ip_tos = 0;	/*XXX*/
1437 		}
1438 #endif
1439 		break;
1440 #endif
1441 #ifdef INET6
1442 	case AF_INET6:
1443 		ip6->ip6_nxt = IPPROTO_TCP;
1444 		if (tp->t_in6pcb) {
1445 			/*
1446 			 * we separately set hoplimit for every segment, since
1447 			 * the user might want to change the value via
1448 			 * setsockopt. Also, desired default hop limit might
1449 			 * be changed via Neighbor Discovery.
1450 			 */
1451 			ip6->ip6_hlim = in6_selecthlim(tp->t_in6pcb,
1452 				ro->ro_rt ? ro->ro_rt->rt_ifp : NULL);
1453 		}
1454 		/* ip6->ip6_flow = ??? */
1455 		/* ip6_plen will be filled in ip6_output(). */
1456 		break;
1457 #endif
1458 	}
1459 
1460 	switch (af) {
1461 #ifdef INET
1462 	case AF_INET:
1463 	    {
1464 		struct mbuf *opts;
1465 
1466 		if (tp->t_inpcb)
1467 			opts = tp->t_inpcb->inp_options;
1468 		else
1469 			opts = NULL;
1470 		error = ip_output(m, opts, ro,
1471 			(tp->t_mtudisc ? IP_MTUDISC : 0) |
1472 			(so->so_options & SO_DONTROUTE),
1473 			(struct ip_moptions *)0, so);
1474 		break;
1475 	    }
1476 #endif
1477 #ifdef INET6
1478 	case AF_INET6:
1479 	    {
1480 		struct ip6_pktopts *opts;
1481 
1482 		if (tp->t_in6pcb)
1483 			opts = tp->t_in6pcb->in6p_outputopts;
1484 		else
1485 			opts = NULL;
1486 		error = ip6_output(m, opts, (struct route_in6 *)ro,
1487 			so->so_options & SO_DONTROUTE,
1488 			(struct ip6_moptions *)0, so, NULL);
1489 		break;
1490 	    }
1491 #endif
1492 	default:
1493 		error = EAFNOSUPPORT;
1494 		break;
1495 	}
1496 	if (error) {
1497 out:
1498 		if (error == ENOBUFS) {
1499 			tcpstat.tcps_selfquench++;
1500 #ifdef INET
1501 			if (tp->t_inpcb)
1502 				tcp_quench(tp->t_inpcb, 0);
1503 #endif
1504 #ifdef INET6
1505 			if (tp->t_in6pcb)
1506 				tcp6_quench(tp->t_in6pcb, 0);
1507 #endif
1508 			error = 0;
1509 		} else if ((error == EHOSTUNREACH || error == ENETDOWN) &&
1510 		    TCPS_HAVERCVDSYN(tp->t_state)) {
1511 			tp->t_softerror = error;
1512 			error = 0;
1513 		}
1514 
1515 		/* Back out the seqence number advance. */
1516 		if (sack_rxmit)
1517 			p->rxmit -= len;
1518 
1519 		/* Restart the delayed ACK timer, if necessary. */
1520 		if (tp->t_flags & TF_DELACK)
1521 			TCP_RESTART_DELACK(tp);
1522 
1523 		return (error);
1524 	}
1525 	tcpstat.tcps_sndtotal++;
1526 	if (tp->t_flags & TF_DELACK)
1527 		tcpstat.tcps_delack++;
1528 
1529 	/*
1530 	 * Data sent (as far as we can tell).
1531 	 * If this advertises a larger window than any other segment,
1532 	 * then remember the size of the advertised window.
1533 	 * Any pending ACK has now been sent.
1534 	 */
1535 	if (win > 0 && SEQ_GT(tp->rcv_nxt+win, tp->rcv_adv))
1536 		tp->rcv_adv = tp->rcv_nxt + win;
1537 	tp->last_ack_sent = tp->rcv_nxt;
1538 	tp->t_flags &= ~TF_ACKNOW;
1539 	TCP_CLEAR_DELACK(tp);
1540 #ifdef DIAGNOSTIC
1541 	if (maxburst < 0)
1542 		printf("tcp_output: maxburst exceeded by %d\n", -maxburst);
1543 #endif
1544 	if (sendalot && (!tcp_do_newreno || --maxburst))
1545 		goto again;
1546 	return (0);
1547 }
1548 
1549 void
1550 tcp_setpersist(struct tcpcb *tp)
1551 {
1552 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> (1 + 2);
1553 	int nticks;
1554 
1555 	if (TCP_TIMER_ISARMED(tp, TCPT_REXMT))
1556 		panic("tcp_output REXMT");
1557 	/*
1558 	 * Start/restart persistance timer.
1559 	 */
1560 	if (t < tp->t_rttmin)
1561 		t = tp->t_rttmin;
1562 	TCPT_RANGESET(nticks, t * tcp_backoff[tp->t_rxtshift],
1563 	    TCPTV_PERSMIN, TCPTV_PERSMAX);
1564 	TCP_TIMER_ARM(tp, TCPT_PERSIST, nticks);
1565 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1566 		tp->t_rxtshift++;
1567 }
1568 
1569 /*
1570  * tcp4_segment: handle M_CSUM_TSOv4 by software.
1571  *
1572  * => always consume m.
1573  * => call output_func with output_arg for each segments.
1574  */
1575 
1576 int
1577 tcp4_segment(struct mbuf *m, int (*output_func)(void *, struct mbuf *),
1578     void *output_arg)
1579 {
1580 	int mss;
1581 	int iphlen;
1582 	int thlen;
1583 	int hlen;
1584 	int len;
1585 	struct ip *iph;
1586 	struct tcphdr *th;
1587 	uint16_t ipid;
1588 	uint32_t tcpseq;
1589 	struct mbuf *hdr = NULL;
1590 	struct mbuf *t;
1591 	int error = 0;
1592 
1593 	KASSERT((m->m_flags & M_PKTHDR) != 0);
1594 	KASSERT((m->m_pkthdr.csum_flags & M_CSUM_TSOv4) != 0);
1595 
1596 	m->m_pkthdr.csum_flags = 0;
1597 
1598 	len = m->m_pkthdr.len;
1599 	KASSERT(len >= sizeof(*iph) + sizeof(*th));
1600 
1601 	if (m->m_len < sizeof(*iph)) {
1602 		m = m_pullup(m, sizeof(*iph));
1603 		if (m == NULL) {
1604 			error = ENOMEM;
1605 			goto quit;
1606 		}
1607 	}
1608 	iph = mtod(m, struct ip *);
1609 	iphlen = iph->ip_hl * 4;
1610 	KASSERT(iph->ip_v == IPVERSION);
1611 	KASSERT(iphlen >= sizeof(*iph));
1612 	KASSERT(iph->ip_p == IPPROTO_TCP);
1613 	ipid = ntohs(iph->ip_id);
1614 
1615 	hlen = iphlen + sizeof(*th);
1616 	if (m->m_len < hlen) {
1617 		m = m_pullup(m, hlen);
1618 		if (m == NULL) {
1619 			error = ENOMEM;
1620 			goto quit;
1621 		}
1622 	}
1623 	th = (void *)(mtod(m, char *) + iphlen);
1624 	tcpseq = ntohl(th->th_seq);
1625 	thlen = th->th_off * 4;
1626 	hlen = iphlen + thlen;
1627 
1628 	mss = m->m_pkthdr.segsz;
1629 	KASSERT(mss != 0);
1630 	KASSERT(len > hlen);
1631 
1632 	t = m_split(m, hlen, M_NOWAIT);
1633 	if (t == NULL) {
1634 		error = ENOMEM;
1635 		goto quit;
1636 	}
1637 	hdr = m;
1638 	m = t;
1639 	len -= hlen;
1640 	KASSERT(len % mss == 0);
1641 	while (len > 0) {
1642 		struct mbuf *n;
1643 
1644 		n = m_dup(hdr, 0, hlen, M_NOWAIT);
1645 		if (n == NULL) {
1646 			error = ENOMEM;
1647 			goto quit;
1648 		}
1649 		KASSERT(n->m_len == hlen); /* XXX */
1650 
1651 		t = m_split(m, mss, M_NOWAIT);
1652 		if (t == NULL) {
1653 			m_freem(n);
1654 			error = ENOMEM;
1655 			goto quit;
1656 		}
1657 		m_cat(n, m);
1658 		m = t;
1659 
1660 		KASSERT(n->m_len >= hlen); /* XXX */
1661 
1662 		n->m_pkthdr.len = hlen + mss;
1663 		iph = mtod(n, struct ip *);
1664 		KASSERT(iph->ip_v == IPVERSION);
1665 		iph->ip_len = htons(n->m_pkthdr.len);
1666 		iph->ip_id = htons(ipid);
1667 		th = (void *)(mtod(n, char *) + iphlen);
1668 		th->th_seq = htonl(tcpseq);
1669 		iph->ip_sum = 0;
1670 		iph->ip_sum = in_cksum(n, iphlen);
1671 		th->th_sum = 0;
1672 		th->th_sum = in4_cksum(n, IPPROTO_TCP, iphlen, thlen + mss);
1673 
1674 		error = (*output_func)(output_arg, n);
1675 		if (error) {
1676 			goto quit;
1677 		}
1678 
1679 		tcpseq += mss;
1680 		ipid++;
1681 		len -= mss;
1682 	}
1683 
1684 quit:
1685 	if (hdr != NULL) {
1686 		m_freem(hdr);
1687 	}
1688 	if (m != NULL) {
1689 		m_freem(m);
1690 	}
1691 
1692 	return error;
1693 }
1694