xref: /netbsd-src/sys/netinet/tcp_output.c (revision c0179c282a5968435315a82f4128c61372c68fc3)
1 /*	$NetBSD: tcp_output.c,v 1.152 2006/11/23 23:12:59 martin 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, 2006 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  * This code is derived from software contributed to The NetBSD Foundation
82  * by Rui Paulo.
83  *
84  * Redistribution and use in source and binary forms, with or without
85  * modification, are permitted provided that the following conditions
86  * are met:
87  * 1. Redistributions of source code must retain the above copyright
88  *    notice, this list of conditions and the following disclaimer.
89  * 2. Redistributions in binary form must reproduce the above copyright
90  *    notice, this list of conditions and the following disclaimer in the
91  *    documentation and/or other materials provided with the distribution.
92  * 3. All advertising materials mentioning features or use of this software
93  *    must display the following acknowledgement:
94  *	This product includes software developed by the NetBSD
95  *	Foundation, Inc. and its contributors.
96  * 4. Neither the name of The NetBSD Foundation nor the names of its
97  *    contributors may be used to endorse or promote products derived
98  *    from this software without specific prior written permission.
99  *
100  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
101  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
102  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
103  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
104  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
105  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
106  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
107  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
108  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
109  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
110  * POSSIBILITY OF SUCH DAMAGE.
111  */
112 
113 /*
114  * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995
115  *	The Regents of the University of California.  All rights reserved.
116  *
117  * Redistribution and use in source and binary forms, with or without
118  * modification, are permitted provided that the following conditions
119  * are met:
120  * 1. Redistributions of source code must retain the above copyright
121  *    notice, this list of conditions and the following disclaimer.
122  * 2. Redistributions in binary form must reproduce the above copyright
123  *    notice, this list of conditions and the following disclaimer in the
124  *    documentation and/or other materials provided with the distribution.
125  * 3. Neither the name of the University nor the names of its contributors
126  *    may be used to endorse or promote products derived from this software
127  *    without specific prior written permission.
128  *
129  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
130  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
131  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
132  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
133  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
134  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
135  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
136  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
137  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
138  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
139  * SUCH DAMAGE.
140  *
141  *	@(#)tcp_output.c	8.4 (Berkeley) 5/24/95
142  */
143 
144 #include <sys/cdefs.h>
145 __KERNEL_RCSID(0, "$NetBSD: tcp_output.c,v 1.152 2006/11/23 23:12:59 martin Exp $");
146 
147 #include "opt_inet.h"
148 #include "opt_ipsec.h"
149 #include "opt_tcp_debug.h"
150 
151 #include <sys/param.h>
152 #include <sys/systm.h>
153 #include <sys/malloc.h>
154 #include <sys/mbuf.h>
155 #include <sys/protosw.h>
156 #include <sys/socket.h>
157 #include <sys/socketvar.h>
158 #include <sys/errno.h>
159 #include <sys/domain.h>
160 #include <sys/kernel.h>
161 #ifdef TCP_SIGNATURE
162 #include <sys/md5.h>
163 #endif
164 
165 #include <net/if.h>
166 #include <net/route.h>
167 
168 #include <netinet/in.h>
169 #include <netinet/in_systm.h>
170 #include <netinet/ip.h>
171 #include <netinet/in_pcb.h>
172 #include <netinet/ip_var.h>
173 
174 #ifdef INET6
175 #ifndef INET
176 #include <netinet/in.h>
177 #endif
178 #include <netinet/ip6.h>
179 #include <netinet6/in6_var.h>
180 #include <netinet6/ip6_var.h>
181 #include <netinet6/in6_pcb.h>
182 #include <netinet6/nd6.h>
183 #endif
184 
185 #ifdef FAST_IPSEC
186 #include <netipsec/ipsec.h>
187 #include <netipsec/key.h>
188 #endif	/* FAST_IPSEC*/
189 #ifdef IPSEC
190 #include <netinet6/ipsec.h>
191 #endif
192 
193 #include <netinet/tcp.h>
194 #define	TCPOUTFLAGS
195 #include <netinet/tcp_fsm.h>
196 #include <netinet/tcp_seq.h>
197 #include <netinet/tcp_timer.h>
198 #include <netinet/tcp_var.h>
199 #include <netinet/tcp_congctl.h>
200 #include <netinet/tcpip.h>
201 #include <netinet/tcp_debug.h>
202 #include <netinet/in_offload.h>
203 #include <netinet6/in6_offload.h>
204 
205 #ifdef IPSEC
206 #include <netkey/key.h>
207 #endif
208 
209 #ifdef notyet
210 extern struct mbuf *m_copypack();
211 #endif
212 
213 /*
214  * Knob to enable Congestion Window Monitoring, and control
215  * the burst size it allows.  Default burst is 4 packets, per
216  * the Internet draft.
217  */
218 int	tcp_cwm = 0;
219 int	tcp_cwm_burstsize = 4;
220 
221 #ifdef TCP_OUTPUT_COUNTERS
222 #include <sys/device.h>
223 
224 extern struct evcnt tcp_output_bigheader;
225 extern struct evcnt tcp_output_predict_hit;
226 extern struct evcnt tcp_output_predict_miss;
227 extern struct evcnt tcp_output_copysmall;
228 extern struct evcnt tcp_output_copybig;
229 extern struct evcnt tcp_output_refbig;
230 
231 #define	TCP_OUTPUT_COUNTER_INCR(ev)	(ev)->ev_count++
232 #else
233 
234 #define	TCP_OUTPUT_COUNTER_INCR(ev)	/* nothing */
235 
236 #endif /* TCP_OUTPUT_COUNTERS */
237 
238 static
239 #ifndef GPROF
240 inline
241 #endif
242 int
243 tcp_segsize(struct tcpcb *tp, int *txsegsizep, int *rxsegsizep,
244     boolean_t *alwaysfragp)
245 {
246 #ifdef INET
247 	struct inpcb *inp = tp->t_inpcb;
248 #endif
249 #ifdef INET6
250 	struct in6pcb *in6p = tp->t_in6pcb;
251 #endif
252 	struct socket *so = NULL;
253 	struct rtentry *rt;
254 	struct ifnet *ifp;
255 	int size;
256 	int hdrlen;
257 	int optlen;
258 
259 	*alwaysfragp = FALSE;
260 
261 #ifdef DIAGNOSTIC
262 	if (tp->t_inpcb && tp->t_in6pcb)
263 		panic("tcp_segsize: both t_inpcb and t_in6pcb are set");
264 #endif
265 	switch (tp->t_family) {
266 #ifdef INET
267 	case AF_INET:
268 		hdrlen = sizeof(struct ip) + sizeof(struct tcphdr);
269 		break;
270 #endif
271 #ifdef INET6
272 	case AF_INET6:
273 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
274 		break;
275 #endif
276 	default:
277 		size = tcp_mssdflt;
278 		goto out;
279 	}
280 
281 	rt = NULL;
282 #ifdef INET
283 	if (inp) {
284 		rt = in_pcbrtentry(inp);
285 		so = inp->inp_socket;
286 	}
287 #endif
288 #ifdef INET6
289 	if (in6p) {
290 		rt = in6_pcbrtentry(in6p);
291 		so = in6p->in6p_socket;
292 	}
293 #endif
294 	if (rt == NULL) {
295 		size = tcp_mssdflt;
296 		goto out;
297 	}
298 
299 	ifp = rt->rt_ifp;
300 
301 	size = tcp_mssdflt;
302 	if (tp->t_mtudisc && rt->rt_rmx.rmx_mtu != 0) {
303 #ifdef INET6
304 		if (in6p && rt->rt_rmx.rmx_mtu < IPV6_MMTU) {
305 			/*
306 			 * RFC2460 section 5, last paragraph: if path MTU is
307 			 * smaller than 1280, use 1280 as packet size and
308 			 * attach fragment header.
309 			 */
310 			size = IPV6_MMTU - hdrlen - sizeof(struct ip6_frag);
311 			*alwaysfragp = TRUE;
312 		} else
313 			size = rt->rt_rmx.rmx_mtu - hdrlen;
314 #else
315 		size = rt->rt_rmx.rmx_mtu - hdrlen;
316 #endif
317 	} else if (ifp->if_flags & IFF_LOOPBACK)
318 		size = ifp->if_mtu - hdrlen;
319 #ifdef INET
320 	else if (inp && tp->t_mtudisc)
321 		size = ifp->if_mtu - hdrlen;
322 	else if (inp && in_localaddr(inp->inp_faddr))
323 		size = ifp->if_mtu - hdrlen;
324 #endif
325 #ifdef INET6
326 	else if (in6p) {
327 #ifdef INET
328 		if (IN6_IS_ADDR_V4MAPPED(&in6p->in6p_faddr)) {
329 			/* mapped addr case */
330 			struct in_addr d;
331 			bcopy(&in6p->in6p_faddr.s6_addr32[3], &d, sizeof(d));
332 			if (tp->t_mtudisc || in_localaddr(d))
333 				size = ifp->if_mtu - hdrlen;
334 		} else
335 #endif
336 		{
337 			/*
338 			 * for IPv6, path MTU discovery is always turned on,
339 			 * or the node must use packet size <= 1280.
340 			 */
341 			size = tp->t_mtudisc ? IN6_LINKMTU(ifp) : IPV6_MMTU;
342 			size -= hdrlen;
343 		}
344 	}
345 #endif
346  out:
347 	/*
348 	 * Now we must make room for whatever extra TCP/IP options are in
349 	 * the packet.
350 	 */
351 	optlen = tcp_optlen(tp);
352 
353 	/*
354 	 * XXX tp->t_ourmss should have the right size, but without this code
355 	 * fragmentation will occur... need more investigation
356 	 */
357 #ifdef INET
358 	if (inp) {
359 #if defined(IPSEC) || defined(FAST_IPSEC)
360 		if (! IPSEC_PCB_SKIP_IPSEC(inp->inp_sp, IPSEC_DIR_OUTBOUND))
361 			optlen += ipsec4_hdrsiz_tcp(tp);
362 #endif
363 		optlen += ip_optlen(inp);
364 	}
365 #endif
366 #ifdef INET6
367 #ifdef INET
368 	if (in6p && tp->t_family == AF_INET) {
369 #if defined(IPSEC) || defined(FAST_IPSEC)
370 		if (! IPSEC_PCB_SKIP_IPSEC(in6p->in6p_sp, IPSEC_DIR_OUTBOUND))
371 			optlen += ipsec4_hdrsiz_tcp(tp);
372 #endif
373 		/* XXX size -= ip_optlen(in6p); */
374 	} else
375 #endif
376 	if (in6p && tp->t_family == AF_INET6) {
377 #ifdef IPSEC
378 		if (! IPSEC_PCB_SKIP_IPSEC(in6p->in6p_sp, IPSEC_DIR_OUTBOUND))
379 			optlen += ipsec6_hdrsiz_tcp(tp);
380 #endif
381 		optlen += ip6_optlen(in6p);
382 	}
383 #endif
384 	size -= optlen;
385 
386 	/* there may not be any room for data if mtu is too small */
387 	if (size < 0)
388 		return (EMSGSIZE);
389 
390 	/*
391 	 * *rxsegsizep holds *estimated* inbound segment size (estimation
392 	 * assumes that path MTU is the same for both ways).  this is only
393 	 * for silly window avoidance, do not use the value for other purposes.
394 	 *
395 	 * ipseclen is subtracted from both sides, this may not be right.
396 	 * I'm not quite sure about this (could someone comment).
397 	 */
398 	*txsegsizep = min(tp->t_peermss - optlen, size);
399 	/*
400 	 * Never send more than half a buffer full.  This insures that we can
401 	 * always keep 2 packets on the wire, no matter what SO_SNDBUF is, and
402 	 * therefore acks will never be delayed unless we run out of data to
403 	 * transmit.
404 	 */
405 	if (so)
406 		*txsegsizep = min(so->so_snd.sb_hiwat >> 1, *txsegsizep);
407 	*rxsegsizep = min(tp->t_ourmss - optlen, size);
408 
409 	if (*txsegsizep != tp->t_segsz) {
410 		/*
411 		 * If the new segment size is larger, we don't want to
412 		 * mess up the congestion window, but if it is smaller
413 		 * we'll have to reduce the congestion window to ensure
414 		 * that we don't get into trouble with initial windows
415 		 * and the rest.  In any case, if the segment size
416 		 * has changed, chances are the path has, too, and
417 		 * our congestion window will be different.
418 		 */
419 		if (*txsegsizep < tp->t_segsz) {
420 			tp->snd_cwnd = max((tp->snd_cwnd / tp->t_segsz)
421 					   * *txsegsizep, *txsegsizep);
422 			tp->snd_ssthresh = max((tp->snd_ssthresh / tp->t_segsz)
423 						* *txsegsizep, *txsegsizep);
424 		}
425 		tp->t_segsz = *txsegsizep;
426 	}
427 
428 	return (0);
429 }
430 
431 static
432 #ifndef GPROF
433 inline
434 #endif
435 int
436 tcp_build_datapkt(struct tcpcb *tp, struct socket *so, int off,
437     long len, int hdrlen, struct mbuf **mp)
438 {
439 	struct mbuf *m, *m0;
440 
441 	if (tp->t_force && len == 1)
442 		tcpstat.tcps_sndprobe++;
443 	else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
444 		tcpstat.tcps_sndrexmitpack++;
445 		tcpstat.tcps_sndrexmitbyte += len;
446 	} else {
447 		tcpstat.tcps_sndpack++;
448 		tcpstat.tcps_sndbyte += len;
449 	}
450 #ifdef notyet
451 	if ((m = m_copypack(so->so_snd.sb_mb, off,
452 	    (int)len, max_linkhdr + hdrlen)) == 0)
453 		return (ENOBUFS);
454 	/*
455 	 * m_copypack left space for our hdr; use it.
456 	 */
457 	m->m_len += hdrlen;
458 	m->m_data -= hdrlen;
459 #else
460 	MGETHDR(m, M_DONTWAIT, MT_HEADER);
461 	if (__predict_false(m == NULL))
462 		return (ENOBUFS);
463 	MCLAIM(m, &tcp_tx_mowner);
464 
465 	/*
466 	 * XXX Because other code assumes headers will fit in
467 	 * XXX one header mbuf.
468 	 *
469 	 * (This code should almost *never* be run.)
470 	 */
471 	if (__predict_false((max_linkhdr + hdrlen) > MHLEN)) {
472 		TCP_OUTPUT_COUNTER_INCR(&tcp_output_bigheader);
473 		MCLGET(m, M_DONTWAIT);
474 		if ((m->m_flags & M_EXT) == 0) {
475 			m_freem(m);
476 			return (ENOBUFS);
477 		}
478 	}
479 
480 	m->m_data += max_linkhdr;
481 	m->m_len = hdrlen;
482 
483 	/*
484 	 * To avoid traversing the whole sb_mb chain for correct
485 	 * data to send, remember last sent mbuf, its offset and
486 	 * the sent size.  When called the next time, see if the
487 	 * data to send is directly following the previous transfer.
488 	 * This is important for large TCP windows.
489 	 */
490 	if (off == 0 || tp->t_lastm == NULL ||
491 	    (tp->t_lastoff + tp->t_lastlen) != off) {
492 		TCP_OUTPUT_COUNTER_INCR(&tcp_output_predict_miss);
493 		/*
494 		 * Either a new packet or a retransmit.
495 		 * Start from the beginning.
496 		 */
497 		tp->t_lastm = so->so_snd.sb_mb;
498 		tp->t_inoff = off;
499 	} else {
500 		TCP_OUTPUT_COUNTER_INCR(&tcp_output_predict_hit);
501 		tp->t_inoff += tp->t_lastlen;
502 	}
503 
504 	/* Traverse forward to next packet */
505 	while (tp->t_inoff > 0) {
506 		if (tp->t_lastm == NULL)
507 			panic("tp->t_lastm == NULL");
508 		if (tp->t_inoff < tp->t_lastm->m_len)
509 			break;
510 		tp->t_inoff -= tp->t_lastm->m_len;
511 		tp->t_lastm = tp->t_lastm->m_next;
512 	}
513 
514 	tp->t_lastoff = off;
515 	tp->t_lastlen = len;
516 	m0 = tp->t_lastm;
517 	off = tp->t_inoff;
518 
519 	if (len <= M_TRAILINGSPACE(m)) {
520 		m_copydata(m0, off, (int) len, mtod(m, caddr_t) + hdrlen);
521 		m->m_len += len;
522 		TCP_OUTPUT_COUNTER_INCR(&tcp_output_copysmall);
523 	} else {
524 		m->m_next = m_copy(m0, off, (int) len);
525 		if (m->m_next == NULL) {
526 			m_freem(m);
527 			return (ENOBUFS);
528 		}
529 #ifdef TCP_OUTPUT_COUNTERS
530 		if (m->m_next->m_flags & M_EXT)
531 			TCP_OUTPUT_COUNTER_INCR(&tcp_output_refbig);
532 		else
533 			TCP_OUTPUT_COUNTER_INCR(&tcp_output_copybig);
534 #endif /* TCP_OUTPUT_COUNTERS */
535 	}
536 #endif
537 
538 	*mp = m;
539 	return (0);
540 }
541 
542 /*
543  * Tcp output routine: figure out what should be sent and send it.
544  */
545 int
546 tcp_output(struct tcpcb *tp)
547 {
548 	struct socket *so;
549 	struct route *ro;
550 	long len, win;
551 	int off, flags, error;
552 	struct mbuf *m;
553 	struct ip *ip;
554 #ifdef INET6
555 	struct ip6_hdr *ip6;
556 #endif
557 	struct tcphdr *th;
558 	u_char opt[MAX_TCPOPTLEN];
559 	unsigned optlen, hdrlen, packetlen;
560 	unsigned int sack_numblks;
561 	int idle, sendalot, txsegsize, rxsegsize;
562 	int txsegsize_nosack;
563 	int maxburst = TCP_MAXBURST;
564 	int af;		/* address family on the wire */
565 	int iphdrlen;
566 	int has_tso4, has_tso6;
567 	int has_tso, use_tso;
568 	boolean_t alwaysfrag;
569 	int sack_rxmit;
570 	int sack_bytes_rxmt;
571 	struct sackhole *p;
572 #ifdef TCP_SIGNATURE
573 	int sigoff = 0;
574 #endif
575 
576 #ifdef DIAGNOSTIC
577 	if (tp->t_inpcb && tp->t_in6pcb)
578 		panic("tcp_output: both t_inpcb and t_in6pcb are set");
579 #endif
580 	so = NULL;
581 	ro = NULL;
582 	if (tp->t_inpcb) {
583 		so = tp->t_inpcb->inp_socket;
584 		ro = &tp->t_inpcb->inp_route;
585 	}
586 #ifdef INET6
587 	else if (tp->t_in6pcb) {
588 		so = tp->t_in6pcb->in6p_socket;
589 		ro = (struct route *)&tp->t_in6pcb->in6p_route;
590 	}
591 #endif
592 
593 	switch (af = tp->t_family) {
594 #ifdef INET
595 	case AF_INET:
596 		if (tp->t_inpcb)
597 			break;
598 #ifdef INET6
599 		/* mapped addr case */
600 		if (tp->t_in6pcb)
601 			break;
602 #endif
603 		return (EINVAL);
604 #endif
605 #ifdef INET6
606 	case AF_INET6:
607 		if (tp->t_in6pcb)
608 			break;
609 		return (EINVAL);
610 #endif
611 	default:
612 		return (EAFNOSUPPORT);
613 	}
614 
615 	if (tcp_segsize(tp, &txsegsize, &rxsegsize, &alwaysfrag))
616 		return (EMSGSIZE);
617 
618 	idle = (tp->snd_max == tp->snd_una);
619 
620 	/*
621 	 * Determine if we can use TCP segmentation offload:
622 	 * - If we're using IPv4
623 	 * - If there is not an IPsec policy that prevents it
624 	 * - If the interface can do it
625 	 */
626 	has_tso4 = has_tso6 = FALSE;
627 #if defined(INET)
628 	has_tso4 = tp->t_inpcb != NULL &&
629 #if defined(IPSEC) || defined(FAST_IPSEC)
630 		  IPSEC_PCB_SKIP_IPSEC(tp->t_inpcb->inp_sp,
631 		  		       IPSEC_DIR_OUTBOUND) &&
632 #endif
633 		  tp->t_inpcb->inp_route.ro_rt != NULL &&
634 		  (tp->t_inpcb->inp_route.ro_rt->rt_ifp->if_capenable &
635 		   IFCAP_TSOv4) != 0;
636 #endif /* defined(INET) */
637 #if defined(INET6)
638 	has_tso6 = tp->t_in6pcb != NULL &&
639 #if defined(IPSEC) || defined(FAST_IPSEC)
640 		  IPSEC_PCB_SKIP_IPSEC(tp->t_in6pcb->in6p_sp,
641 		  		       IPSEC_DIR_OUTBOUND) &&
642 #endif
643 		  tp->t_in6pcb->in6p_route.ro_rt != NULL &&
644 		  (tp->t_in6pcb->in6p_route.ro_rt->rt_ifp->if_capenable &
645 		   IFCAP_TSOv6) != 0;
646 #endif /* defined(INET6) */
647 	has_tso = (has_tso4 || has_tso6) && !alwaysfrag;
648 
649 	/*
650 	 * Restart Window computation.  From draft-floyd-incr-init-win-03:
651 	 *
652 	 *	Optionally, a TCP MAY set the restart window to the
653 	 *	minimum of the value used for the initial window and
654 	 *	the current value of cwnd (in other words, using a
655 	 *	larger value for the restart window should never increase
656 	 *	the size of cwnd).
657 	 */
658 	if (tcp_cwm) {
659 		/*
660 		 * Hughes/Touch/Heidemann Congestion Window Monitoring.
661 		 * Count the number of packets currently pending
662 		 * acknowledgement, and limit our congestion window
663 		 * to a pre-determined allowed burst size plus that count.
664 		 * This prevents bursting once all pending packets have
665 		 * been acknowledged (i.e. transmission is idle).
666 		 *
667 		 * XXX Link this to Initial Window?
668 		 */
669 		tp->snd_cwnd = min(tp->snd_cwnd,
670 		    (tcp_cwm_burstsize * txsegsize) +
671 		    (tp->snd_nxt - tp->snd_una));
672 	} else {
673 		if (idle && (tcp_now - tp->t_rcvtime) >= tp->t_rxtcur) {
674 			/*
675 			 * We have been idle for "a while" and no acks are
676 			 * expected to clock out any data we send --
677 			 * slow start to get ack "clock" running again.
678 			 */
679 			int ss = tcp_init_win;
680 #ifdef INET
681 			if (tp->t_inpcb &&
682 			    in_localaddr(tp->t_inpcb->inp_faddr))
683 				ss = tcp_init_win_local;
684 #endif
685 #ifdef INET6
686 			if (tp->t_in6pcb &&
687 			    in6_localaddr(&tp->t_in6pcb->in6p_faddr))
688 				ss = tcp_init_win_local;
689 #endif
690 			tp->snd_cwnd = min(tp->snd_cwnd,
691 			    TCP_INITIAL_WINDOW(ss, txsegsize));
692 		}
693 	}
694 
695 	txsegsize_nosack = txsegsize;
696 again:
697 	use_tso = has_tso;
698 	if ((tp->t_flags & (TF_ECN_SND_CWR|TF_ECN_SND_ECE)) != 0) {
699 		/* don't duplicate CWR/ECE. */
700 		use_tso = 0;
701 	}
702 	TCP_REASS_LOCK(tp);
703 	sack_numblks = tcp_sack_numblks(tp);
704 	if (sack_numblks) {
705 		int sackoptlen;
706 
707 		sackoptlen = TCP_SACK_OPTLEN(sack_numblks);
708 		if (sackoptlen > txsegsize_nosack) {
709 			sack_numblks = 0; /* give up SACK */
710 			txsegsize = txsegsize_nosack;
711 		} else {
712 			if ((tp->rcv_sack_flags & TCPSACK_HAVED) != 0) {
713 				/* don't duplicate D-SACK. */
714 				use_tso = 0;
715 			}
716 			txsegsize = txsegsize_nosack - sackoptlen;
717 		}
718 	} else {
719 		txsegsize = txsegsize_nosack;
720 	}
721 
722 	/*
723 	 * Determine length of data that should be transmitted, and
724 	 * flags that should be used.  If there is some data or critical
725 	 * controls (SYN, RST) to send, then transmit; otherwise,
726 	 * investigate further.
727 	 *
728 	 * Readjust SACK information to avoid resending duplicate data.
729 	 */
730 	if (TCP_SACK_ENABLED(tp) && SEQ_LT(tp->snd_nxt, tp->snd_max))
731 		tcp_sack_adjust(tp);
732 	sendalot = 0;
733 	off = tp->snd_nxt - tp->snd_una;
734 	win = min(tp->snd_wnd, tp->snd_cwnd);
735 
736 	flags = tcp_outflags[tp->t_state];
737 
738 	/*
739 	 * Send any SACK-generated retransmissions.  If we're explicitly trying
740 	 * to send out new data (when sendalot is 1), bypass this function.
741 	 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
742 	 * we're replacing a (future) new transmission with a retransmission
743 	 * now, and we previously incremented snd_cwnd in tcp_input().
744 	 */
745 	/*
746 	 * Still in sack recovery , reset rxmit flag to zero.
747 	 */
748 	sack_rxmit = 0;
749 	sack_bytes_rxmt = 0;
750 	len = 0;
751 	p = NULL;
752 	do {
753 		long cwin;
754 		if (!TCP_SACK_ENABLED(tp))
755 			break;
756 		if (tp->t_partialacks < 0)
757 			break;
758 		p = tcp_sack_output(tp, &sack_bytes_rxmt);
759 		if (p == NULL)
760 			break;
761 
762 		cwin = min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt;
763 		if (cwin < 0)
764 			cwin = 0;
765 		/* Do not retransmit SACK segments beyond snd_recover */
766 		if (SEQ_GT(p->end, tp->snd_recover)) {
767 			/*
768 			 * (At least) part of sack hole extends beyond
769 			 * snd_recover. Check to see if we can rexmit data
770 			 * for this hole.
771 			 */
772 			if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
773 				/*
774 				 * Can't rexmit any more data for this hole.
775 				 * That data will be rexmitted in the next
776 				 * sack recovery episode, when snd_recover
777 				 * moves past p->rxmit.
778 				 */
779 				p = NULL;
780 				break;
781 			}
782 			/* Can rexmit part of the current hole */
783 			len = ((long)ulmin(cwin, tp->snd_recover - p->rxmit));
784 		} else
785 			len = ((long)ulmin(cwin, p->end - p->rxmit));
786 		off = p->rxmit - tp->snd_una;
787 		if (off + len > so->so_snd.sb_cc) {
788 			/* 1 for TH_FIN */
789 			KASSERT(off + len == so->so_snd.sb_cc + 1);
790 			KASSERT(p->rxmit + len == tp->snd_max);
791 			len = so->so_snd.sb_cc - off;
792 		}
793 		if (len > 0) {
794 			sack_rxmit = 1;
795 			sendalot = 1;
796 		}
797 	} while (/*CONSTCOND*/0);
798 
799 	/*
800 	 * If in persist timeout with window of 0, send 1 byte.
801 	 * Otherwise, if window is small but nonzero
802 	 * and timer expired, we will send what we can
803 	 * and go to transmit state.
804 	 */
805 	if (tp->t_force) {
806 		if (win == 0) {
807 			/*
808 			 * If we still have some data to send, then
809 			 * clear the FIN bit.  Usually this would
810 			 * happen below when it realizes that we
811 			 * aren't sending all the data.  However,
812 			 * if we have exactly 1 byte of unset data,
813 			 * then it won't clear the FIN bit below,
814 			 * and if we are in persist state, we wind
815 			 * up sending the packet without recording
816 			 * that we sent the FIN bit.
817 			 *
818 			 * We can't just blindly clear the FIN bit,
819 			 * because if we don't have any more data
820 			 * to send then the probe will be the FIN
821 			 * itself.
822 			 */
823 			if (off < so->so_snd.sb_cc)
824 				flags &= ~TH_FIN;
825 			win = 1;
826 		} else {
827 			TCP_TIMER_DISARM(tp, TCPT_PERSIST);
828 			tp->t_rxtshift = 0;
829 		}
830 	}
831 
832 	if (sack_rxmit == 0) {
833 		if (TCP_SACK_ENABLED(tp) && tp->t_partialacks >= 0) {
834 			long cwin;
835 
836 			/*
837 			 * We are inside of a SACK recovery episode and are
838 			 * sending new data, having retransmitted all the
839 			 * data possible in the scoreboard.
840 			 */
841 			if (tp->snd_wnd < so->so_snd.sb_cc) {
842 				len = tp->snd_wnd - off;
843 				flags &= ~TH_FIN;
844 			} else {
845 				len = so->so_snd.sb_cc - off;
846 			}
847 
848 			/*
849 			 * From FreeBSD:
850 			 *  Don't remove this (len > 0) check !
851 			 *  We explicitly check for len > 0 here (although it
852 			 *  isn't really necessary), to work around a gcc
853 			 *  optimization issue - to force gcc to compute
854 			 *  len above. Without this check, the computation
855 			 *  of len is bungled by the optimizer.
856 			 */
857 			if (len > 0) {
858 				cwin = tp->snd_cwnd -
859 				    (tp->snd_nxt - tp->sack_newdata) -
860 				    sack_bytes_rxmt;
861 				if (cwin < 0)
862 					cwin = 0;
863 				if (cwin < len) {
864 					len = cwin;
865 					flags &= ~TH_FIN;
866 				}
867 			}
868 		} else if (win < so->so_snd.sb_cc) {
869 			len = win - off;
870 			flags &= ~TH_FIN;
871 		} else {
872 			len = so->so_snd.sb_cc - off;
873 		}
874 	}
875 
876 	if (len < 0) {
877 		/*
878 		 * If FIN has been sent but not acked,
879 		 * but we haven't been called to retransmit,
880 		 * len will be -1.  Otherwise, window shrank
881 		 * after we sent into it.  If window shrank to 0,
882 		 * cancel pending retransmit, pull snd_nxt back
883 		 * to (closed) window, and set the persist timer
884 		 * if it isn't already going.  If the window didn't
885 		 * close completely, just wait for an ACK.
886 		 *
887 		 * If we have a pending FIN, either it has already been
888 		 * transmitted or it is outside the window, so drop it.
889 		 * If the FIN has been transmitted, but this is not a
890 		 * retransmission, then len must be -1.  Therefore we also
891 		 * prevent here the sending of `gratuitous FINs'.  This
892 		 * eliminates the need to check for that case below (e.g.
893 		 * to back up snd_nxt before the FIN so that the sequence
894 		 * number is correct).
895 		 */
896 		len = 0;
897 		flags &= ~TH_FIN;
898 		if (win == 0) {
899 			TCP_TIMER_DISARM(tp, TCPT_REXMT);
900 			tp->t_rxtshift = 0;
901 			tp->snd_nxt = tp->snd_una;
902 			if (TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0)
903 				tcp_setpersist(tp);
904 		}
905 	}
906 	if (len > txsegsize) {
907 		if (use_tso) {
908 			/*
909 			 * Truncate TSO transfers to IP_MAXPACKET, and make
910 			 * sure that we send equal size transfers down the
911 			 * stack (rather than big-small-big-small-...).
912 			 */
913 #ifdef INET6
914 #if IPV6_MAXPACKET != IP_MAXPACKET
915 #error IPV6_MAXPACKET != IP_MAXPACKET
916 #endif
917 #endif
918 			len = (min(len, IP_MAXPACKET) / txsegsize) * txsegsize;
919 			if (len <= txsegsize) {
920 				use_tso = 0;
921 			}
922 		} else
923 			len = txsegsize;
924 		flags &= ~TH_FIN;
925 		sendalot = 1;
926 	} else
927 		use_tso = 0;
928 	if (sack_rxmit) {
929 		if (SEQ_LT(p->rxmit + len, tp->snd_una + so->so_snd.sb_cc))
930 			flags &= ~TH_FIN;
931 	}
932 
933 	win = sbspace(&so->so_rcv);
934 
935 	/*
936 	 * Sender silly window avoidance.  If connection is idle
937 	 * and can send all data, a maximum segment,
938 	 * at least a maximum default-size segment do it,
939 	 * or are forced, do it; otherwise don't bother.
940 	 * If peer's buffer is tiny, then send
941 	 * when window is at least half open.
942 	 * If retransmitting (possibly after persist timer forced us
943 	 * to send into a small window), then must resend.
944 	 */
945 	if (len) {
946 		if (len >= txsegsize)
947 			goto send;
948 		if ((so->so_state & SS_MORETOCOME) == 0 &&
949 		    ((idle || tp->t_flags & TF_NODELAY) &&
950 		     len + off >= so->so_snd.sb_cc))
951 			goto send;
952 		if (tp->t_force)
953 			goto send;
954 		if (len >= tp->max_sndwnd / 2)
955 			goto send;
956 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))
957 			goto send;
958 		if (sack_rxmit)
959 			goto send;
960 	}
961 
962 	/*
963 	 * Compare available window to amount of window known to peer
964 	 * (as advertised window less next expected input).  If the
965 	 * difference is at least twice the size of the largest segment
966 	 * we expect to receive (i.e. two segments) or at least 50% of
967 	 * the maximum possible window, then want to send a window update
968 	 * to peer.
969 	 */
970 	if (win > 0) {
971 		/*
972 		 * "adv" is the amount we can increase the window,
973 		 * taking into account that we are limited by
974 		 * TCP_MAXWIN << tp->rcv_scale.
975 		 */
976 		long adv = min(win, (long)TCP_MAXWIN << tp->rcv_scale) -
977 			(tp->rcv_adv - tp->rcv_nxt);
978 
979 		if (adv >= (long) (2 * rxsegsize))
980 			goto send;
981 		if (2 * adv >= (long) so->so_rcv.sb_hiwat)
982 			goto send;
983 	}
984 
985 	/*
986 	 * Send if we owe peer an ACK.
987 	 */
988 	if (tp->t_flags & TF_ACKNOW)
989 		goto send;
990 	if (flags & (TH_SYN|TH_FIN|TH_RST))
991 		goto send;
992 	if (SEQ_GT(tp->snd_up, tp->snd_una))
993 		goto send;
994 	/*
995 	 * In SACK, it is possible for tcp_output to fail to send a segment
996 	 * after the retransmission timer has been turned off.  Make sure
997 	 * that the retransmission timer is set.
998 	 */
999 	if (TCP_SACK_ENABLED(tp) && SEQ_GT(tp->snd_max, tp->snd_una) &&
1000 	    !TCP_TIMER_ISARMED(tp, TCPT_REXMT) &&
1001 	    !TCP_TIMER_ISARMED(tp, TCPT_PERSIST)) {
1002 		TCP_TIMER_ARM(tp, TCPT_REXMT, tp->t_rxtcur);
1003 		goto just_return;
1004 	}
1005 
1006 	/*
1007 	 * TCP window updates are not reliable, rather a polling protocol
1008 	 * using ``persist'' packets is used to insure receipt of window
1009 	 * updates.  The three ``states'' for the output side are:
1010 	 *	idle			not doing retransmits or persists
1011 	 *	persisting		to move a small or zero window
1012 	 *	(re)transmitting	and thereby not persisting
1013 	 *
1014 	 * tp->t_timer[TCPT_PERSIST]
1015 	 *	is set when we are in persist state.
1016 	 * tp->t_force
1017 	 *	is set when we are called to send a persist packet.
1018 	 * tp->t_timer[TCPT_REXMT]
1019 	 *	is set when we are retransmitting
1020 	 * The output side is idle when both timers are zero.
1021 	 *
1022 	 * If send window is too small, there is data to transmit, and no
1023 	 * retransmit or persist is pending, then go to persist state.
1024 	 * If nothing happens soon, send when timer expires:
1025 	 * if window is nonzero, transmit what we can,
1026 	 * otherwise force out a byte.
1027 	 */
1028 	if (so->so_snd.sb_cc && TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 &&
1029 	    TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) {
1030 		tp->t_rxtshift = 0;
1031 		tcp_setpersist(tp);
1032 	}
1033 
1034 	/*
1035 	 * No reason to send a segment, just return.
1036 	 */
1037 just_return:
1038 	TCP_REASS_UNLOCK(tp);
1039 	return (0);
1040 
1041 send:
1042 	/*
1043 	 * Before ESTABLISHED, force sending of initial options
1044 	 * unless TCP set not to do any options.
1045 	 * NOTE: we assume that the IP/TCP header plus TCP options
1046 	 * always fit in a single mbuf, leaving room for a maximum
1047 	 * link header, i.e.
1048 	 *	max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
1049 	 */
1050 	optlen = 0;
1051 	switch (af) {
1052 #ifdef INET
1053 	case AF_INET:
1054 		iphdrlen = sizeof(struct ip) + sizeof(struct tcphdr);
1055 		break;
1056 #endif
1057 #ifdef INET6
1058 	case AF_INET6:
1059 		iphdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
1060 		break;
1061 #endif
1062 	default:	/*pacify gcc*/
1063 		iphdrlen = 0;
1064 		break;
1065 	}
1066 	hdrlen = iphdrlen;
1067 	if (flags & TH_SYN) {
1068 		struct rtentry *rt;
1069 
1070 		rt = NULL;
1071 #ifdef INET
1072 		if (tp->t_inpcb)
1073 			rt = in_pcbrtentry(tp->t_inpcb);
1074 #endif
1075 #ifdef INET6
1076 		if (tp->t_in6pcb)
1077 			rt = in6_pcbrtentry(tp->t_in6pcb);
1078 #endif
1079 
1080 		tp->snd_nxt = tp->iss;
1081 		tp->t_ourmss = tcp_mss_to_advertise(rt != NULL ?
1082 						    rt->rt_ifp : NULL, af);
1083 		if ((tp->t_flags & TF_NOOPT) == 0) {
1084 			opt[0] = TCPOPT_MAXSEG;
1085 			opt[1] = 4;
1086 			opt[2] = (tp->t_ourmss >> 8) & 0xff;
1087 			opt[3] = tp->t_ourmss & 0xff;
1088 			optlen = 4;
1089 
1090 			if ((tp->t_flags & TF_REQ_SCALE) &&
1091 			    ((flags & TH_ACK) == 0 ||
1092 			    (tp->t_flags & TF_RCVD_SCALE))) {
1093 				*((u_int32_t *) (opt + optlen)) = htonl(
1094 					TCPOPT_NOP << 24 |
1095 					TCPOPT_WINDOW << 16 |
1096 					TCPOLEN_WINDOW << 8 |
1097 					tp->request_r_scale);
1098 				optlen += 4;
1099 			}
1100 			if (tcp_do_sack) {
1101 				u_int8_t *cp = (u_int8_t *)(opt + optlen);
1102 
1103 				cp[0] = TCPOPT_SACK_PERMITTED;
1104 				cp[1] = 2;
1105 				cp[2] = TCPOPT_NOP;
1106 				cp[3] = TCPOPT_NOP;
1107 				optlen += 4;
1108 			}
1109 		}
1110 	}
1111 
1112 	/*
1113 	 * Send a timestamp and echo-reply if this is a SYN and our side
1114 	 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side
1115 	 * and our peer have sent timestamps in our SYN's.
1116 	 */
1117 	if ((tp->t_flags & (TF_REQ_TSTMP|TF_NOOPT)) == TF_REQ_TSTMP &&
1118 	     (flags & TH_RST) == 0 &&
1119 	    ((flags & (TH_SYN|TH_ACK)) == TH_SYN ||
1120 	     (tp->t_flags & TF_RCVD_TSTMP))) {
1121 		u_int32_t *lp = (u_int32_t *)(opt + optlen);
1122 
1123 		/* Form timestamp option as shown in appendix A of RFC 1323. */
1124 		*lp++ = htonl(TCPOPT_TSTAMP_HDR);
1125 		*lp++ = htonl(TCP_TIMESTAMP(tp));
1126 		*lp   = htonl(tp->ts_recent);
1127 		optlen += TCPOLEN_TSTAMP_APPA;
1128 	}
1129 
1130 	/*
1131 	 * Tack on the SACK block if it is necessary.
1132 	 */
1133 	if (sack_numblks) {
1134 		int sack_len;
1135 		u_char *bp = (u_char *)(opt + optlen);
1136 		u_int32_t *lp = (u_int32_t *)(bp + 4);
1137 		struct ipqent *tiqe;
1138 
1139 		sack_len = sack_numblks * 8 + 2;
1140 		bp[0] = TCPOPT_NOP;
1141 		bp[1] = TCPOPT_NOP;
1142 		bp[2] = TCPOPT_SACK;
1143 		bp[3] = sack_len;
1144 		if ((tp->rcv_sack_flags & TCPSACK_HAVED) != 0) {
1145 			sack_numblks--;
1146 			*lp++ = htonl(tp->rcv_dsack_block.left);
1147 			*lp++ = htonl(tp->rcv_dsack_block.right);
1148 			tp->rcv_sack_flags &= ~TCPSACK_HAVED;
1149 		}
1150 		for (tiqe = TAILQ_FIRST(&tp->timeq);
1151 		    sack_numblks > 0; tiqe = TAILQ_NEXT(tiqe, ipqe_timeq)) {
1152 			KASSERT(tiqe != NULL);
1153 			sack_numblks--;
1154 			*lp++ = htonl(tiqe->ipqe_seq);
1155 			*lp++ = htonl(tiqe->ipqe_seq + tiqe->ipqe_len +
1156 			    ((tiqe->ipqe_flags & TH_FIN) != 0 ? 1 : 0));
1157 		}
1158 		optlen += sack_len + 2;
1159 	}
1160 	TCP_REASS_UNLOCK(tp);
1161 
1162 #ifdef TCP_SIGNATURE
1163 #if defined(INET6) && defined(FAST_IPSEC)
1164 	if (tp->t_family == AF_INET)
1165 #endif
1166 	if (tp->t_flags & TF_SIGNATURE) {
1167 		u_char *bp;
1168 		/*
1169 		 * Initialize TCP-MD5 option (RFC2385)
1170 		 */
1171 		bp = (u_char *)opt + optlen;
1172 		*bp++ = TCPOPT_SIGNATURE;
1173 		*bp++ = TCPOLEN_SIGNATURE;
1174 		sigoff = optlen + 2;
1175 		bzero(bp, TCP_SIGLEN);
1176 		bp += TCP_SIGLEN;
1177 		optlen += TCPOLEN_SIGNATURE;
1178 		/*
1179 		 * Terminate options list and maintain 32-bit alignment.
1180  		 */
1181 		*bp++ = TCPOPT_NOP;
1182 		*bp++ = TCPOPT_EOL;
1183  		optlen += 2;
1184  	}
1185 #endif /* TCP_SIGNATURE */
1186 
1187 	hdrlen += optlen;
1188 
1189 #ifdef DIAGNOSTIC
1190 	if (!use_tso && len > txsegsize)
1191 		panic("tcp data to be sent is larger than segment");
1192 	else if (use_tso && len > IP_MAXPACKET)
1193 		panic("tcp data to be sent is larger than max TSO size");
1194 	if (max_linkhdr + hdrlen > MCLBYTES)
1195 		panic("tcphdr too big");
1196 #endif
1197 
1198 	/*
1199 	 * Grab a header mbuf, attaching a copy of data to
1200 	 * be transmitted, and initialize the header from
1201 	 * the template for sends on this connection.
1202 	 */
1203 	if (len) {
1204 		error = tcp_build_datapkt(tp, so, off, len, hdrlen, &m);
1205 		if (error)
1206 			goto out;
1207 		/*
1208 		 * If we're sending everything we've got, set PUSH.
1209 		 * (This will keep happy those implementations which only
1210 		 * give data to the user when a buffer fills or
1211 		 * a PUSH comes in.)
1212 		 */
1213 		if (off + len == so->so_snd.sb_cc)
1214 			flags |= TH_PUSH;
1215 	} else {
1216 		if (tp->t_flags & TF_ACKNOW)
1217 			tcpstat.tcps_sndacks++;
1218 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
1219 			tcpstat.tcps_sndctrl++;
1220 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
1221 			tcpstat.tcps_sndurg++;
1222 		else
1223 			tcpstat.tcps_sndwinup++;
1224 
1225 		MGETHDR(m, M_DONTWAIT, MT_HEADER);
1226 		if (m != NULL && max_linkhdr + hdrlen > MHLEN) {
1227 			MCLGET(m, M_DONTWAIT);
1228 			if ((m->m_flags & M_EXT) == 0) {
1229 				m_freem(m);
1230 				m = NULL;
1231 			}
1232 		}
1233 		if (m == NULL) {
1234 			error = ENOBUFS;
1235 			goto out;
1236 		}
1237 		MCLAIM(m, &tcp_tx_mowner);
1238 		m->m_data += max_linkhdr;
1239 		m->m_len = hdrlen;
1240 	}
1241 	m->m_pkthdr.rcvif = (struct ifnet *)0;
1242 	switch (af) {
1243 #ifdef INET
1244 	case AF_INET:
1245 		ip = mtod(m, struct ip *);
1246 #ifdef INET6
1247 		ip6 = NULL;
1248 #endif
1249 		th = (struct tcphdr *)(ip + 1);
1250 		break;
1251 #endif
1252 #ifdef INET6
1253 	case AF_INET6:
1254 		ip = NULL;
1255 		ip6 = mtod(m, struct ip6_hdr *);
1256 		th = (struct tcphdr *)(ip6 + 1);
1257 		break;
1258 #endif
1259 	default:	/*pacify gcc*/
1260 		ip = NULL;
1261 #ifdef INET6
1262 		ip6 = NULL;
1263 #endif
1264 		th = NULL;
1265 		break;
1266 	}
1267 	if (tp->t_template == 0)
1268 		panic("tcp_output");
1269 	if (tp->t_template->m_len < iphdrlen)
1270 		panic("tcp_output");
1271 	bcopy(mtod(tp->t_template, caddr_t), mtod(m, caddr_t), iphdrlen);
1272 
1273 	/*
1274 	 * If we are starting a connection, send ECN setup
1275 	 * SYN packet. If we are on a retransmit, we may
1276 	 * resend those bits a number of times as per
1277 	 * RFC 3168.
1278 	 */
1279 	if (tp->t_state == TCPS_SYN_SENT && tcp_do_ecn) {
1280 		if (tp->t_flags & TF_SYN_REXMT) {
1281 			if (tp->t_ecn_retries--)
1282 				flags |= TH_ECE|TH_CWR;
1283 		} else {
1284 			flags |= TH_ECE|TH_CWR;
1285 			tp->t_ecn_retries = tcp_ecn_maxretries;
1286 		}
1287 	}
1288 
1289 	if (TCP_ECN_ALLOWED(tp)) {
1290 		/*
1291 		 * If the peer has ECN, mark data packets
1292 		 * ECN capable. Ignore pure ack packets, retransmissions
1293 		 * and window probes.
1294 		 */
1295 		if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
1296 		    !(tp->t_force && len == 1)) {
1297 			switch (af) {
1298 #ifdef INET
1299 			case AF_INET:
1300 				tp->t_inpcb->inp_ip.ip_tos |= IPTOS_ECN_ECT0;
1301 				break;
1302 #endif
1303 #ifdef INET6
1304 			case AF_INET6:
1305 				ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
1306 				break;
1307 #endif
1308 			}
1309 			tcpstat.tcps_ecn_ect++;
1310 		}
1311 
1312 		/*
1313 		 * Reply with proper ECN notifications.
1314 		 */
1315 		if (tp->t_flags & TF_ECN_SND_CWR) {
1316 			flags |= TH_CWR;
1317 			tp->t_flags &= ~TF_ECN_SND_CWR;
1318 		}
1319 		if (tp->t_flags & TF_ECN_SND_ECE) {
1320 			flags |= TH_ECE;
1321 		}
1322 	}
1323 
1324 
1325 	/*
1326 	 * If we are doing retransmissions, then snd_nxt will
1327 	 * not reflect the first unsent octet.  For ACK only
1328 	 * packets, we do not want the sequence number of the
1329 	 * retransmitted packet, we want the sequence number
1330 	 * of the next unsent octet.  So, if there is no data
1331 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
1332 	 * when filling in ti_seq.  But if we are in persist
1333 	 * state, snd_max might reflect one byte beyond the
1334 	 * right edge of the window, so use snd_nxt in that
1335 	 * case, since we know we aren't doing a retransmission.
1336 	 * (retransmit and persist are mutually exclusive...)
1337 	 */
1338 	if (TCP_SACK_ENABLED(tp) && sack_rxmit) {
1339 		th->th_seq = htonl(p->rxmit);
1340 		p->rxmit += len;
1341 	} else {
1342 		if (len || (flags & (TH_SYN|TH_FIN)) ||
1343 		    TCP_TIMER_ISARMED(tp, TCPT_PERSIST))
1344 			th->th_seq = htonl(tp->snd_nxt);
1345 		else
1346 			th->th_seq = htonl(tp->snd_max);
1347 	}
1348 	th->th_ack = htonl(tp->rcv_nxt);
1349 	if (optlen) {
1350 		bcopy((caddr_t)opt, (caddr_t)(th + 1), optlen);
1351 		th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
1352 	}
1353 	th->th_flags = flags;
1354 	/*
1355 	 * Calculate receive window.  Don't shrink window,
1356 	 * but avoid silly window syndrome.
1357 	 */
1358 	if (win < (long)(so->so_rcv.sb_hiwat / 4) && win < (long)rxsegsize)
1359 		win = 0;
1360 	if (win > (long)TCP_MAXWIN << tp->rcv_scale)
1361 		win = (long)TCP_MAXWIN << tp->rcv_scale;
1362 	if (win < (long)(int32_t)(tp->rcv_adv - tp->rcv_nxt))
1363 		win = (long)(int32_t)(tp->rcv_adv - tp->rcv_nxt);
1364 	th->th_win = htons((u_int16_t) (win>>tp->rcv_scale));
1365 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1366 		u_int32_t urp = tp->snd_up - tp->snd_nxt;
1367 		if (urp > IP_MAXPACKET)
1368 			urp = IP_MAXPACKET;
1369 		th->th_urp = htons((u_int16_t)urp);
1370 		th->th_flags |= TH_URG;
1371 	} else
1372 		/*
1373 		 * If no urgent pointer to send, then we pull
1374 		 * the urgent pointer to the left edge of the send window
1375 		 * so that it doesn't drift into the send window on sequence
1376 		 * number wraparound.
1377 		 */
1378 		tp->snd_up = tp->snd_una;		/* drag it along */
1379 
1380 #ifdef TCP_SIGNATURE
1381 #if defined(INET6) && defined(FAST_IPSEC)
1382 	if (tp->t_family == AF_INET) /* XXX */
1383 #endif
1384 	if (sigoff && (tp->t_flags & TF_SIGNATURE)) {
1385 		struct secasvar *sav;
1386 		u_int8_t *sigp;
1387 
1388 		sav = tcp_signature_getsav(m, th);
1389 
1390 		if (sav == NULL) {
1391 			if (m)
1392 				m_freem(m);
1393 			return (EPERM);
1394 		}
1395 
1396 		m->m_pkthdr.len = hdrlen + len;
1397 		sigp = (caddr_t)th + sizeof(*th) + sigoff;
1398 		tcp_signature(m, th, (caddr_t)th - mtod(m, caddr_t), sav, sigp);
1399 
1400 		key_sa_recordxfer(sav, m);
1401 #ifdef FAST_IPSEC
1402 		KEY_FREESAV(&sav);
1403 #else
1404 		key_freesav(sav);
1405 #endif
1406 	}
1407 #endif
1408 
1409 	/*
1410 	 * Set ourselves up to be checksummed just before the packet
1411 	 * hits the wire.
1412 	 */
1413 	switch (af) {
1414 #ifdef INET
1415 	case AF_INET:
1416 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1417 		if (use_tso) {
1418 			m->m_pkthdr.segsz = txsegsize;
1419 			m->m_pkthdr.csum_flags = M_CSUM_TSOv4;
1420 		} else {
1421 			m->m_pkthdr.csum_flags = M_CSUM_TCPv4;
1422 			if (len + optlen) {
1423 				/* Fixup the pseudo-header checksum. */
1424 				/* XXXJRT Not IP Jumbogram safe. */
1425 				th->th_sum = in_cksum_addword(th->th_sum,
1426 				    htons((u_int16_t) (len + optlen)));
1427 			}
1428 		}
1429 		break;
1430 #endif
1431 #ifdef INET6
1432 	case AF_INET6:
1433 		m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1434 		if (use_tso) {
1435 			m->m_pkthdr.segsz = txsegsize;
1436 			m->m_pkthdr.csum_flags = M_CSUM_TSOv6;
1437 		} else {
1438 			m->m_pkthdr.csum_flags = M_CSUM_TCPv6;
1439 			if (len + optlen) {
1440 				/* Fixup the pseudo-header checksum. */
1441 				/* XXXJRT: Not IPv6 Jumbogram safe. */
1442 				th->th_sum = in_cksum_addword(th->th_sum,
1443 				    htons((u_int16_t) (len + optlen)));
1444 			}
1445 		}
1446 		break;
1447 #endif
1448 	}
1449 
1450 	/*
1451 	 * In transmit state, time the transmission and arrange for
1452 	 * the retransmit.  In persist state, just set snd_max.
1453 	 */
1454 	if (tp->t_force == 0 || TCP_TIMER_ISARMED(tp, TCPT_PERSIST) == 0) {
1455 		tcp_seq startseq = tp->snd_nxt;
1456 
1457 		/*
1458 		 * Advance snd_nxt over sequence space of this segment.
1459 		 * There are no states in which we send both a SYN and a FIN,
1460 		 * so we collapse the tests for these flags.
1461 		 */
1462 		if (flags & (TH_SYN|TH_FIN))
1463 			tp->snd_nxt++;
1464 		if (sack_rxmit)
1465 			goto timer;
1466 		tp->snd_nxt += len;
1467 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1468 			tp->snd_max = tp->snd_nxt;
1469 			/*
1470 			 * Time this transmission if not a retransmission and
1471 			 * not currently timing anything.
1472 			 */
1473 			if (tp->t_rtttime == 0) {
1474 				tp->t_rtttime = tcp_now;
1475 				tp->t_rtseq = startseq;
1476 				tcpstat.tcps_segstimed++;
1477 			}
1478 		}
1479 
1480 		/*
1481 		 * Set retransmit timer if not currently set,
1482 		 * and not doing an ack or a keep-alive probe.
1483 		 * Initial value for retransmit timer is smoothed
1484 		 * round-trip time + 2 * round-trip time variance.
1485 		 * Initialize shift counter which is used for backoff
1486 		 * of retransmit time.
1487 		 */
1488 timer:
1489 		if (TCP_TIMER_ISARMED(tp, TCPT_REXMT) == 0 &&
1490 			((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1491 		    tp->snd_nxt != tp->snd_una)) {
1492 			if (TCP_TIMER_ISARMED(tp, TCPT_PERSIST)) {
1493 				TCP_TIMER_DISARM(tp, TCPT_PERSIST);
1494 				tp->t_rxtshift = 0;
1495 			}
1496 			TCP_TIMER_ARM(tp, TCPT_REXMT, tp->t_rxtcur);
1497 		}
1498 	} else
1499 		if (SEQ_GT(tp->snd_nxt + len, tp->snd_max))
1500 			tp->snd_max = tp->snd_nxt + len;
1501 
1502 #ifdef TCP_DEBUG
1503 	/*
1504 	 * Trace.
1505 	 */
1506 	if (so->so_options & SO_DEBUG)
1507 		tcp_trace(TA_OUTPUT, tp->t_state, tp, m, 0);
1508 #endif
1509 
1510 	/*
1511 	 * Fill in IP length and desired time to live and
1512 	 * send to IP level.  There should be a better way
1513 	 * to handle ttl and tos; we could keep them in
1514 	 * the template, but need a way to checksum without them.
1515 	 */
1516 	m->m_pkthdr.len = hdrlen + len;
1517 
1518 	switch (af) {
1519 #ifdef INET
1520 	case AF_INET:
1521 		ip->ip_len = htons(m->m_pkthdr.len);
1522 		packetlen = m->m_pkthdr.len;
1523 		if (tp->t_inpcb) {
1524 			ip->ip_ttl = tp->t_inpcb->inp_ip.ip_ttl;
1525 			ip->ip_tos = tp->t_inpcb->inp_ip.ip_tos;
1526 		}
1527 #ifdef INET6
1528 		else if (tp->t_in6pcb) {
1529 			ip->ip_ttl = in6_selecthlim(tp->t_in6pcb, NULL); /*XXX*/
1530 			ip->ip_tos = 0;	/*XXX*/
1531 		}
1532 #endif
1533 		break;
1534 #endif
1535 #ifdef INET6
1536 	case AF_INET6:
1537 		packetlen = m->m_pkthdr.len;
1538 		ip6->ip6_nxt = IPPROTO_TCP;
1539 		if (tp->t_in6pcb) {
1540 			/*
1541 			 * we separately set hoplimit for every segment, since
1542 			 * the user might want to change the value via
1543 			 * setsockopt. Also, desired default hop limit might
1544 			 * be changed via Neighbor Discovery.
1545 			 */
1546 			ip6->ip6_hlim = in6_selecthlim(tp->t_in6pcb,
1547 				ro->ro_rt ? ro->ro_rt->rt_ifp : NULL);
1548 		}
1549 		/* ip6->ip6_flow = ??? */
1550 		/* ip6_plen will be filled in ip6_output(). */
1551 		break;
1552 #endif
1553 	default:	/*pacify gcc*/
1554 		packetlen = 0;
1555 		break;
1556 	}
1557 
1558 	switch (af) {
1559 #ifdef INET
1560 	case AF_INET:
1561 	    {
1562 		struct mbuf *opts;
1563 
1564 		if (tp->t_inpcb)
1565 			opts = tp->t_inpcb->inp_options;
1566 		else
1567 			opts = NULL;
1568 		error = ip_output(m, opts, ro,
1569 			(tp->t_mtudisc ? IP_MTUDISC : 0) |
1570 			(so->so_options & SO_DONTROUTE),
1571 			(struct ip_moptions *)0, so);
1572 		break;
1573 	    }
1574 #endif
1575 #ifdef INET6
1576 	case AF_INET6:
1577 	    {
1578 		struct ip6_pktopts *opts;
1579 
1580 		if (tp->t_in6pcb)
1581 			opts = tp->t_in6pcb->in6p_outputopts;
1582 		else
1583 			opts = NULL;
1584 		error = ip6_output(m, opts, (struct route_in6 *)ro,
1585 			so->so_options & SO_DONTROUTE,
1586 			(struct ip6_moptions *)0, so, NULL);
1587 		break;
1588 	    }
1589 #endif
1590 	default:
1591 		error = EAFNOSUPPORT;
1592 		break;
1593 	}
1594 	if (error) {
1595 out:
1596 		if (error == ENOBUFS) {
1597 			tcpstat.tcps_selfquench++;
1598 #ifdef INET
1599 			if (tp->t_inpcb)
1600 				tcp_quench(tp->t_inpcb, 0);
1601 #endif
1602 #ifdef INET6
1603 			if (tp->t_in6pcb)
1604 				tcp6_quench(tp->t_in6pcb, 0);
1605 #endif
1606 			error = 0;
1607 		} else if ((error == EHOSTUNREACH || error == ENETDOWN) &&
1608 		    TCPS_HAVERCVDSYN(tp->t_state)) {
1609 			tp->t_softerror = error;
1610 			error = 0;
1611 		}
1612 
1613 		/* Back out the seqence number advance. */
1614 		if (sack_rxmit)
1615 			p->rxmit -= len;
1616 
1617 		/* Restart the delayed ACK timer, if necessary. */
1618 		if (tp->t_flags & TF_DELACK)
1619 			TCP_RESTART_DELACK(tp);
1620 
1621 		return (error);
1622 	}
1623 
1624 	if (packetlen > tp->t_pmtud_mtu_sent)
1625 		tp->t_pmtud_mtu_sent = packetlen;
1626 
1627 	tcpstat.tcps_sndtotal++;
1628 	if (tp->t_flags & TF_DELACK)
1629 		tcpstat.tcps_delack++;
1630 
1631 	/*
1632 	 * Data sent (as far as we can tell).
1633 	 * If this advertises a larger window than any other segment,
1634 	 * then remember the size of the advertised window.
1635 	 * Any pending ACK has now been sent.
1636 	 */
1637 	if (win > 0 && SEQ_GT(tp->rcv_nxt+win, tp->rcv_adv))
1638 		tp->rcv_adv = tp->rcv_nxt + win;
1639 	tp->last_ack_sent = tp->rcv_nxt;
1640 	tp->t_flags &= ~TF_ACKNOW;
1641 	TCP_CLEAR_DELACK(tp);
1642 #ifdef DIAGNOSTIC
1643 	if (maxburst < 0)
1644 		printf("tcp_output: maxburst exceeded by %d\n", -maxburst);
1645 #endif
1646 	if (sendalot && (tp->t_congctl == &tcp_reno_ctl || --maxburst))
1647 		goto again;
1648 	return (0);
1649 }
1650 
1651 void
1652 tcp_setpersist(struct tcpcb *tp)
1653 {
1654 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> (1 + 2);
1655 	int nticks;
1656 
1657 	if (TCP_TIMER_ISARMED(tp, TCPT_REXMT))
1658 		panic("tcp_output REXMT");
1659 	/*
1660 	 * Start/restart persistance timer.
1661 	 */
1662 	if (t < tp->t_rttmin)
1663 		t = tp->t_rttmin;
1664 	TCPT_RANGESET(nticks, t * tcp_backoff[tp->t_rxtshift],
1665 	    TCPTV_PERSMIN, TCPTV_PERSMAX);
1666 	TCP_TIMER_ARM(tp, TCPT_PERSIST, nticks);
1667 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1668 		tp->t_rxtshift++;
1669 }
1670 
1671 #if defined(INET)
1672 /*
1673  * tcp4_segment: handle M_CSUM_TSOv4 by software.
1674  *
1675  * => always consume m.
1676  * => call output_func with output_arg for each segments.
1677  */
1678 
1679 int
1680 tcp4_segment(struct mbuf *m, int (*output_func)(void *, struct mbuf *),
1681     void *output_arg)
1682 {
1683 	int mss;
1684 	int iphlen;
1685 	int thlen;
1686 	int hlen;
1687 	int len;
1688 	struct ip *iph;
1689 	struct tcphdr *th;
1690 	uint16_t ipid;
1691 	uint32_t tcpseq;
1692 	struct mbuf *hdr = NULL;
1693 	struct mbuf *t;
1694 	int error = 0;
1695 
1696 	KASSERT((m->m_flags & M_PKTHDR) != 0);
1697 	KASSERT((m->m_pkthdr.csum_flags & M_CSUM_TSOv4) != 0);
1698 
1699 	m->m_pkthdr.csum_flags = 0;
1700 
1701 	len = m->m_pkthdr.len;
1702 	KASSERT(len >= sizeof(*iph) + sizeof(*th));
1703 
1704 	if (m->m_len < sizeof(*iph)) {
1705 		m = m_pullup(m, sizeof(*iph));
1706 		if (m == NULL) {
1707 			error = ENOMEM;
1708 			goto quit;
1709 		}
1710 	}
1711 	iph = mtod(m, struct ip *);
1712 	iphlen = iph->ip_hl * 4;
1713 	KASSERT(iph->ip_v == IPVERSION);
1714 	KASSERT(iphlen >= sizeof(*iph));
1715 	KASSERT(iph->ip_p == IPPROTO_TCP);
1716 	ipid = ntohs(iph->ip_id);
1717 
1718 	hlen = iphlen + sizeof(*th);
1719 	if (m->m_len < hlen) {
1720 		m = m_pullup(m, hlen);
1721 		if (m == NULL) {
1722 			error = ENOMEM;
1723 			goto quit;
1724 		}
1725 	}
1726 	th = (void *)(mtod(m, char *) + iphlen);
1727 	tcpseq = ntohl(th->th_seq);
1728 	thlen = th->th_off * 4;
1729 	hlen = iphlen + thlen;
1730 
1731 	mss = m->m_pkthdr.segsz;
1732 	KASSERT(mss != 0);
1733 	KASSERT(len > hlen);
1734 
1735 	t = m_split(m, hlen, M_NOWAIT);
1736 	if (t == NULL) {
1737 		error = ENOMEM;
1738 		goto quit;
1739 	}
1740 	hdr = m;
1741 	m = t;
1742 	len -= hlen;
1743 	KASSERT(len % mss == 0);
1744 	while (len > 0) {
1745 		struct mbuf *n;
1746 
1747 		n = m_dup(hdr, 0, hlen, M_NOWAIT);
1748 		if (n == NULL) {
1749 			error = ENOMEM;
1750 			goto quit;
1751 		}
1752 		KASSERT(n->m_len == hlen); /* XXX */
1753 
1754 		t = m_split(m, mss, M_NOWAIT);
1755 		if (t == NULL) {
1756 			m_freem(n);
1757 			error = ENOMEM;
1758 			goto quit;
1759 		}
1760 		m_cat(n, m);
1761 		m = t;
1762 
1763 		KASSERT(n->m_len >= hlen); /* XXX */
1764 
1765 		n->m_pkthdr.len = hlen + mss;
1766 		iph = mtod(n, struct ip *);
1767 		KASSERT(iph->ip_v == IPVERSION);
1768 		iph->ip_len = htons(n->m_pkthdr.len);
1769 		iph->ip_id = htons(ipid);
1770 		th = (void *)(mtod(n, char *) + iphlen);
1771 		th->th_seq = htonl(tcpseq);
1772 		iph->ip_sum = 0;
1773 		iph->ip_sum = in_cksum(n, iphlen);
1774 		th->th_sum = 0;
1775 		th->th_sum = in4_cksum(n, IPPROTO_TCP, iphlen, thlen + mss);
1776 
1777 		error = (*output_func)(output_arg, n);
1778 		if (error) {
1779 			goto quit;
1780 		}
1781 
1782 		tcpseq += mss;
1783 		ipid++;
1784 		len -= mss;
1785 	}
1786 
1787 quit:
1788 	if (hdr != NULL) {
1789 		m_freem(hdr);
1790 	}
1791 	if (m != NULL) {
1792 		m_freem(m);
1793 	}
1794 
1795 	return error;
1796 }
1797 #endif /* defined(INET) */
1798 
1799 #if defined(INET6)
1800 /*
1801  * tcp6_segment: handle M_CSUM_TSOv6 by software.
1802  *
1803  * => always consume m.
1804  * => call output_func with output_arg for each segments.
1805  */
1806 
1807 int
1808 tcp6_segment(struct mbuf *m, int (*output_func)(void *, struct mbuf *),
1809     void *output_arg)
1810 {
1811 	int mss;
1812 	int iphlen;
1813 	int thlen;
1814 	int hlen;
1815 	int len;
1816 	struct ip6_hdr *iph;
1817 	struct tcphdr *th;
1818 	uint32_t tcpseq;
1819 	struct mbuf *hdr = NULL;
1820 	struct mbuf *t;
1821 	int error = 0;
1822 
1823 	KASSERT((m->m_flags & M_PKTHDR) != 0);
1824 	KASSERT((m->m_pkthdr.csum_flags & M_CSUM_TSOv6) != 0);
1825 
1826 	m->m_pkthdr.csum_flags = 0;
1827 
1828 	len = m->m_pkthdr.len;
1829 	KASSERT(len >= sizeof(*iph) + sizeof(*th));
1830 
1831 	if (m->m_len < sizeof(*iph)) {
1832 		m = m_pullup(m, sizeof(*iph));
1833 		if (m == NULL) {
1834 			error = ENOMEM;
1835 			goto quit;
1836 		}
1837 	}
1838 	iph = mtod(m, struct ip6_hdr *);
1839 	iphlen = sizeof(*iph);
1840 	KASSERT((iph->ip6_vfc & IPV6_VERSION_MASK) == IPV6_VERSION);
1841 	KASSERT(iph->ip6_nxt == IPPROTO_TCP);
1842 
1843 	hlen = iphlen + sizeof(*th);
1844 	if (m->m_len < hlen) {
1845 		m = m_pullup(m, hlen);
1846 		if (m == NULL) {
1847 			error = ENOMEM;
1848 			goto quit;
1849 		}
1850 	}
1851 	th = (void *)(mtod(m, char *) + iphlen);
1852 	tcpseq = ntohl(th->th_seq);
1853 	thlen = th->th_off * 4;
1854 	hlen = iphlen + thlen;
1855 
1856 	mss = m->m_pkthdr.segsz;
1857 	KASSERT(mss != 0);
1858 	KASSERT(len > hlen);
1859 
1860 	t = m_split(m, hlen, M_NOWAIT);
1861 	if (t == NULL) {
1862 		error = ENOMEM;
1863 		goto quit;
1864 	}
1865 	hdr = m;
1866 	m = t;
1867 	len -= hlen;
1868 	KASSERT(len % mss == 0);
1869 	while (len > 0) {
1870 		struct mbuf *n;
1871 
1872 		n = m_dup(hdr, 0, hlen, M_NOWAIT);
1873 		if (n == NULL) {
1874 			error = ENOMEM;
1875 			goto quit;
1876 		}
1877 		KASSERT(n->m_len == hlen); /* XXX */
1878 
1879 		t = m_split(m, mss, M_NOWAIT);
1880 		if (t == NULL) {
1881 			m_freem(n);
1882 			error = ENOMEM;
1883 			goto quit;
1884 		}
1885 		m_cat(n, m);
1886 		m = t;
1887 
1888 		KASSERT(n->m_len >= hlen); /* XXX */
1889 
1890 		n->m_pkthdr.len = hlen + mss;
1891 		iph = mtod(n, struct ip6_hdr *);
1892 		KASSERT((iph->ip6_vfc & IPV6_VERSION_MASK) == IPV6_VERSION);
1893 		iph->ip6_plen = htons(thlen + mss);
1894 		th = (void *)(mtod(n, char *) + iphlen);
1895 		th->th_seq = htonl(tcpseq);
1896 		th->th_sum = 0;
1897 		th->th_sum = in6_cksum(n, IPPROTO_TCP, iphlen, thlen + mss);
1898 
1899 		error = (*output_func)(output_arg, n);
1900 		if (error) {
1901 			goto quit;
1902 		}
1903 
1904 		tcpseq += mss;
1905 		len -= mss;
1906 	}
1907 
1908 quit:
1909 	if (hdr != NULL) {
1910 		m_freem(hdr);
1911 	}
1912 	if (m != NULL) {
1913 		m_freem(m);
1914 	}
1915 
1916 	return error;
1917 }
1918 #endif /* defined(INET6) */
1919